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/LoopPassManager.h"
50 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
51 #include "llvm/Analysis/ScalarEvolutionExpander.h"
52 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
53 #include "llvm/Analysis/TargetLibraryInfo.h"
54 #include "llvm/Analysis/TargetTransformInfo.h"
55 #include "llvm/Analysis/ValueTracking.h"
56 #include "llvm/IR/DataLayout.h"
57 #include "llvm/IR/Dominators.h"
58 #include "llvm/IR/IRBuilder.h"
59 #include "llvm/IR/IntrinsicInst.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Transforms/Scalar.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);
151 class LoopIdiomRecognizeLegacyPass : public LoopPass {
154 explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) {
155 initializeLoopIdiomRecognizeLegacyPassPass(
156 *PassRegistry::getPassRegistry());
159 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
163 AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
164 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
165 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
166 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
167 TargetLibraryInfo *TLI =
168 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
169 const TargetTransformInfo *TTI =
170 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
171 *L->getHeader()->getParent());
172 const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout();
174 LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, DL);
175 return LIR.runOnLoop(L);
178 /// This transformation requires natural loop information & requires that
179 /// loop preheaders be inserted into the CFG.
181 void getAnalysisUsage(AnalysisUsage &AU) const override {
182 AU.addRequired<TargetLibraryInfoWrapperPass>();
183 AU.addRequired<TargetTransformInfoWrapperPass>();
184 getLoopAnalysisUsage(AU);
187 } // End anonymous namespace.
189 PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L,
190 LoopAnalysisManager &AM) {
192 AM.getResult<FunctionAnalysisManagerLoopProxy>(L).getManager();
193 Function *F = L.getHeader()->getParent();
195 // Use getCachedResult because Loop pass cannot trigger a function analysis.
196 auto *AA = FAM.getCachedResult<AAManager>(*F);
197 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(*F);
198 auto *LI = FAM.getCachedResult<LoopAnalysis>(*F);
199 auto *SE = FAM.getCachedResult<ScalarEvolutionAnalysis>(*F);
200 auto *TLI = FAM.getCachedResult<TargetLibraryAnalysis>(*F);
201 const auto *TTI = FAM.getCachedResult<TargetIRAnalysis>(*F);
202 const auto *DL = &L.getHeader()->getModule()->getDataLayout();
203 assert((AA && DT && LI && SE && TLI && TTI && DL) &&
204 "Analyses for Loop Idiom Recognition not available");
206 LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, DL);
207 if (!LIR.runOnLoop(&L))
208 return PreservedAnalyses::all();
210 return getLoopPassPreservedAnalyses();
213 char LoopIdiomRecognizeLegacyPass::ID = 0;
214 INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",
215 "Recognize loop idioms", false, false)
216 INITIALIZE_PASS_DEPENDENCY(LoopPass)
217 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
218 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
219 INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",
220 "Recognize loop idioms", false, false)
222 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); }
224 static void deleteDeadInstruction(Instruction *I) {
225 I->replaceAllUsesWith(UndefValue::get(I->getType()));
226 I->eraseFromParent();
229 //===----------------------------------------------------------------------===//
231 // Implementation of LoopIdiomRecognize
233 //===----------------------------------------------------------------------===//
235 bool LoopIdiomRecognize::runOnLoop(Loop *L) {
237 // If the loop could not be converted to canonical form, it must have an
238 // indirectbr in it, just give up.
239 if (!L->getLoopPreheader())
242 // Disable loop idiom recognition if the function's name is a common idiom.
243 StringRef Name = L->getHeader()->getParent()->getName();
244 if (Name == "memset" || Name == "memcpy")
247 // Determine if code size heuristics need to be applied.
248 ApplyCodeSizeHeuristics =
249 L->getHeader()->getParent()->optForSize() && UseLIRCodeSizeHeurs;
251 HasMemset = TLI->has(LibFunc::memset);
252 HasMemsetPattern = TLI->has(LibFunc::memset_pattern16);
253 HasMemcpy = TLI->has(LibFunc::memcpy);
255 if (HasMemset || HasMemsetPattern || HasMemcpy)
256 if (SE->hasLoopInvariantBackedgeTakenCount(L))
257 return runOnCountableLoop();
259 return runOnNoncountableLoop();
262 bool LoopIdiomRecognize::runOnCountableLoop() {
263 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
264 assert(!isa<SCEVCouldNotCompute>(BECount) &&
265 "runOnCountableLoop() called on a loop without a predictable"
266 "backedge-taken count");
268 // If this loop executes exactly one time, then it should be peeled, not
269 // optimized by this pass.
270 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
271 if (BECst->getAPInt() == 0)
274 SmallVector<BasicBlock *, 8> ExitBlocks;
275 CurLoop->getUniqueExitBlocks(ExitBlocks);
277 DEBUG(dbgs() << "loop-idiom Scanning: F["
278 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
279 << CurLoop->getHeader()->getName() << "\n");
281 bool MadeChange = false;
283 // The following transforms hoist stores/memsets into the loop pre-header.
284 // Give up if the loop has instructions may throw.
285 LoopSafetyInfo SafetyInfo;
286 computeLoopSafetyInfo(&SafetyInfo, CurLoop);
287 if (SafetyInfo.MayThrow)
290 // Scan all the blocks in the loop that are not in subloops.
291 for (auto *BB : CurLoop->getBlocks()) {
292 // Ignore blocks in subloops.
293 if (LI->getLoopFor(BB) != CurLoop)
296 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
301 static unsigned getStoreSizeInBytes(StoreInst *SI, const DataLayout *DL) {
302 uint64_t SizeInBits = DL->getTypeSizeInBits(SI->getValueOperand()->getType());
303 assert(((SizeInBits & 7) || (SizeInBits >> 32) == 0) &&
304 "Don't overflow unsigned.");
305 return (unsigned)SizeInBits >> 3;
308 static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) {
309 const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
310 return ConstStride->getAPInt();
313 /// getMemSetPatternValue - If a strided store of the specified value is safe to
314 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
315 /// be passed in. Otherwise, return null.
317 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
318 /// just replicate their input array and then pass on to memset_pattern16.
319 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
320 // If the value isn't a constant, we can't promote it to being in a constant
321 // array. We could theoretically do a store to an alloca or something, but
322 // that doesn't seem worthwhile.
323 Constant *C = dyn_cast<Constant>(V);
327 // Only handle simple values that are a power of two bytes in size.
328 uint64_t Size = DL->getTypeSizeInBits(V->getType());
329 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
332 // Don't care enough about darwin/ppc to implement this.
333 if (DL->isBigEndian())
336 // Convert to size in bytes.
339 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
340 // if the top and bottom are the same (e.g. for vectors and large integers).
344 // If the constant is exactly 16 bytes, just use it.
348 // Otherwise, we'll use an array of the constants.
349 unsigned ArraySize = 16 / Size;
350 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
351 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
354 bool LoopIdiomRecognize::isLegalStore(StoreInst *SI, bool &ForMemset,
355 bool &ForMemsetPattern, bool &ForMemcpy) {
356 // Don't touch volatile stores.
360 // Avoid merging nontemporal stores.
361 if (SI->getMetadata(LLVMContext::MD_nontemporal))
364 Value *StoredVal = SI->getValueOperand();
365 Value *StorePtr = SI->getPointerOperand();
367 // Reject stores that are so large that they overflow an unsigned.
368 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
369 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
372 // See if the pointer expression is an AddRec like {base,+,1} on the current
373 // loop, which indicates a strided store. If we have something else, it's a
374 // random store we can't handle.
375 const SCEVAddRecExpr *StoreEv =
376 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
377 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
380 // Check to see if we have a constant stride.
381 if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
384 // See if the store can be turned into a memset.
386 // If the stored value is a byte-wise value (like i32 -1), then it may be
387 // turned into a memset of i8 -1, assuming that all the consecutive bytes
388 // are stored. A store of i32 0x01020304 can never be turned into a memset,
389 // but it can be turned into memset_pattern if the target supports it.
390 Value *SplatValue = isBytewiseValue(StoredVal);
391 Constant *PatternValue = nullptr;
393 // If we're allowed to form a memset, and the stored value would be
394 // acceptable for memset, use it.
395 if (HasMemset && SplatValue &&
396 // Verify that the stored value is loop invariant. If not, we can't
397 // promote the memset.
398 CurLoop->isLoopInvariant(SplatValue)) {
399 // It looks like we can use SplatValue.
402 } else if (HasMemsetPattern &&
403 // Don't create memset_pattern16s with address spaces.
404 StorePtr->getType()->getPointerAddressSpace() == 0 &&
405 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
406 // It looks like we can use PatternValue!
407 ForMemsetPattern = true;
411 // Otherwise, see if the store can be turned into a memcpy.
413 // Check to see if the stride matches the size of the store. If so, then we
414 // know that every byte is touched in the loop.
415 APInt Stride = getStoreStride(StoreEv);
416 unsigned StoreSize = getStoreSizeInBytes(SI, DL);
417 if (StoreSize != Stride && StoreSize != -Stride)
420 // The store must be feeding a non-volatile load.
421 LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
422 if (!LI || !LI->isSimple())
425 // See if the pointer expression is an AddRec like {base,+,1} on the current
426 // loop, which indicates a strided load. If we have something else, it's a
427 // random load we can't handle.
428 const SCEVAddRecExpr *LoadEv =
429 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
430 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
433 // The store and load must share the same stride.
434 if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
437 // Success. This store can be converted into a memcpy.
441 // This store can't be transformed into a memset/memcpy.
445 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
446 StoreRefsForMemset.clear();
447 StoreRefsForMemsetPattern.clear();
448 StoreRefsForMemcpy.clear();
449 for (Instruction &I : *BB) {
450 StoreInst *SI = dyn_cast<StoreInst>(&I);
454 bool ForMemset = false;
455 bool ForMemsetPattern = false;
456 bool ForMemcpy = false;
457 // Make sure this is a strided store with a constant stride.
458 if (!isLegalStore(SI, ForMemset, ForMemsetPattern, ForMemcpy))
461 // Save the store locations.
463 // Find the base pointer.
464 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), *DL);
465 StoreRefsForMemset[Ptr].push_back(SI);
466 } else if (ForMemsetPattern) {
467 // Find the base pointer.
468 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), *DL);
469 StoreRefsForMemsetPattern[Ptr].push_back(SI);
470 } else if (ForMemcpy)
471 StoreRefsForMemcpy.push_back(SI);
475 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
476 /// with the specified backedge count. This block is known to be in the current
477 /// loop and not in any subloops.
478 bool LoopIdiomRecognize::runOnLoopBlock(
479 BasicBlock *BB, const SCEV *BECount,
480 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
481 // We can only promote stores in this block if they are unconditionally
482 // executed in the loop. For a block to be unconditionally executed, it has
483 // to dominate all the exit blocks of the loop. Verify this now.
484 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
485 if (!DT->dominates(BB, ExitBlocks[i]))
488 bool MadeChange = false;
489 // Look for store instructions, which may be optimized to memset/memcpy.
492 // Look for a single store or sets of stores with a common base, which can be
493 // optimized into a memset (memset_pattern). The latter most commonly happens
494 // with structs and handunrolled loops.
495 for (auto &SL : StoreRefsForMemset)
496 MadeChange |= processLoopStores(SL.second, BECount, true);
498 for (auto &SL : StoreRefsForMemsetPattern)
499 MadeChange |= processLoopStores(SL.second, BECount, false);
501 // Optimize the store into a memcpy, if it feeds an similarly strided load.
502 for (auto &SI : StoreRefsForMemcpy)
503 MadeChange |= processLoopStoreOfLoopLoad(SI, BECount);
505 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
506 Instruction *Inst = &*I++;
507 // Look for memset instructions, which may be optimized to a larger memset.
508 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
510 if (!processLoopMemSet(MSI, BECount))
514 // If processing the memset invalidated our iterator, start over from the
525 /// processLoopStores - See if this store(s) can be promoted to a memset.
526 bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL,
529 // Try to find consecutive stores that can be transformed into memsets.
530 SetVector<StoreInst *> Heads, Tails;
531 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
533 // Do a quadratic search on all of the given stores and find
534 // all of the pairs of stores that follow each other.
535 SmallVector<unsigned, 16> IndexQueue;
536 for (unsigned i = 0, e = SL.size(); i < e; ++i) {
537 assert(SL[i]->isSimple() && "Expected only non-volatile stores.");
539 Value *FirstStoredVal = SL[i]->getValueOperand();
540 Value *FirstStorePtr = SL[i]->getPointerOperand();
541 const SCEVAddRecExpr *FirstStoreEv =
542 cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr));
543 APInt FirstStride = getStoreStride(FirstStoreEv);
544 unsigned FirstStoreSize = getStoreSizeInBytes(SL[i], DL);
546 // See if we can optimize just this store in isolation.
547 if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) {
552 Value *FirstSplatValue = nullptr;
553 Constant *FirstPatternValue = nullptr;
556 FirstSplatValue = isBytewiseValue(FirstStoredVal);
558 FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL);
560 assert((FirstSplatValue || FirstPatternValue) &&
561 "Expected either splat value or pattern value.");
564 // If a store has multiple consecutive store candidates, search Stores
565 // array according to the sequence: from i+1 to e, then from i-1 to 0.
566 // This is because usually pairing with immediate succeeding or preceding
567 // candidate create the best chance to find memset opportunity.
569 for (j = i + 1; j < e; ++j)
570 IndexQueue.push_back(j);
571 for (j = i; j > 0; --j)
572 IndexQueue.push_back(j - 1);
574 for (auto &k : IndexQueue) {
575 assert(SL[k]->isSimple() && "Expected only non-volatile stores.");
576 Value *SecondStorePtr = SL[k]->getPointerOperand();
577 const SCEVAddRecExpr *SecondStoreEv =
578 cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr));
579 APInt SecondStride = getStoreStride(SecondStoreEv);
581 if (FirstStride != SecondStride)
584 Value *SecondStoredVal = SL[k]->getValueOperand();
585 Value *SecondSplatValue = nullptr;
586 Constant *SecondPatternValue = nullptr;
589 SecondSplatValue = isBytewiseValue(SecondStoredVal);
591 SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL);
593 assert((SecondSplatValue || SecondPatternValue) &&
594 "Expected either splat value or pattern value.");
596 if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) {
598 if (FirstSplatValue != SecondSplatValue)
601 if (FirstPatternValue != SecondPatternValue)
606 ConsecutiveChain[SL[i]] = SL[k];
612 // We may run into multiple chains that merge into a single chain. We mark the
613 // stores that we transformed so that we don't visit the same store twice.
614 SmallPtrSet<Value *, 16> TransformedStores;
615 bool Changed = false;
617 // For stores that start but don't end a link in the chain:
618 for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end();
620 if (Tails.count(*it))
623 // We found a store instr that starts a chain. Now follow the chain and try
625 SmallPtrSet<Instruction *, 8> AdjacentStores;
628 StoreInst *HeadStore = I;
629 unsigned StoreSize = 0;
631 // Collect the chain into a list.
632 while (Tails.count(I) || Heads.count(I)) {
633 if (TransformedStores.count(I))
635 AdjacentStores.insert(I);
637 StoreSize += getStoreSizeInBytes(I, DL);
638 // Move to the next value in the chain.
639 I = ConsecutiveChain[I];
642 Value *StoredVal = HeadStore->getValueOperand();
643 Value *StorePtr = HeadStore->getPointerOperand();
644 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
645 APInt Stride = getStoreStride(StoreEv);
647 // Check to see if the stride matches the size of the stores. If so, then
648 // we know that every byte is touched in the loop.
649 if (StoreSize != Stride && StoreSize != -Stride)
652 bool NegStride = StoreSize == -Stride;
654 if (processLoopStridedStore(StorePtr, StoreSize, HeadStore->getAlignment(),
655 StoredVal, HeadStore, AdjacentStores, StoreEv,
656 BECount, NegStride)) {
657 TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end());
665 /// processLoopMemSet - See if this memset can be promoted to a large memset.
666 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
667 const SCEV *BECount) {
668 // We can only handle non-volatile memsets with a constant size.
669 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
672 // If we're not allowed to hack on memset, we fail.
676 Value *Pointer = MSI->getDest();
678 // See if the pointer expression is an AddRec like {base,+,1} on the current
679 // loop, which indicates a strided store. If we have something else, it's a
680 // random store we can't handle.
681 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
682 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
685 // Reject memsets that are so large that they overflow an unsigned.
686 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
687 if ((SizeInBytes >> 32) != 0)
690 // Check to see if the stride matches the size of the memset. If so, then we
691 // know that every byte is touched in the loop.
692 const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
696 APInt Stride = ConstStride->getAPInt();
697 if (SizeInBytes != Stride && SizeInBytes != -Stride)
700 // Verify that the memset value is loop invariant. If not, we can't promote
702 Value *SplatValue = MSI->getValue();
703 if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue))
706 SmallPtrSet<Instruction *, 1> MSIs;
708 bool NegStride = SizeInBytes == -Stride;
709 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
710 MSI->getAlignment(), SplatValue, MSI, MSIs, Ev,
711 BECount, NegStride, /*IsLoopMemset=*/true);
714 /// mayLoopAccessLocation - Return true if the specified loop might access the
715 /// specified pointer location, which is a loop-strided access. The 'Access'
716 /// argument specifies what the verboten forms of access are (read or write).
718 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
719 const SCEV *BECount, unsigned StoreSize,
721 SmallPtrSetImpl<Instruction *> &IgnoredStores) {
722 // Get the location that may be stored across the loop. Since the access is
723 // strided positively through memory, we say that the modified location starts
724 // at the pointer and has infinite size.
725 uint64_t AccessSize = MemoryLocation::UnknownSize;
727 // If the loop iterates a fixed number of times, we can refine the access size
728 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
729 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
730 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
732 // TODO: For this to be really effective, we have to dive into the pointer
733 // operand in the store. Store to &A[i] of 100 will always return may alias
734 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
735 // which will then no-alias a store to &A[100].
736 MemoryLocation StoreLoc(Ptr, AccessSize);
738 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
740 for (Instruction &I : **BI)
741 if (IgnoredStores.count(&I) == 0 &&
742 (AA.getModRefInfo(&I, StoreLoc) & Access))
748 // If we have a negative stride, Start refers to the end of the memory location
749 // we're trying to memset. Therefore, we need to recompute the base pointer,
750 // which is just Start - BECount*Size.
751 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
752 Type *IntPtr, unsigned StoreSize,
753 ScalarEvolution *SE) {
754 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
756 Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize),
758 return SE->getMinusSCEV(Start, Index);
761 /// processLoopStridedStore - We see a strided store of some value. If we can
762 /// transform this into a memset or memset_pattern in the loop preheader, do so.
763 bool LoopIdiomRecognize::processLoopStridedStore(
764 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
765 Value *StoredVal, Instruction *TheStore,
766 SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev,
767 const SCEV *BECount, bool NegStride, bool IsLoopMemset) {
768 Value *SplatValue = isBytewiseValue(StoredVal);
769 Constant *PatternValue = nullptr;
772 PatternValue = getMemSetPatternValue(StoredVal, DL);
774 assert((SplatValue || PatternValue) &&
775 "Expected either splat value or pattern value.");
777 // The trip count of the loop and the base pointer of the addrec SCEV is
778 // guaranteed to be loop invariant, which means that it should dominate the
779 // header. This allows us to insert code for it in the preheader.
780 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
781 BasicBlock *Preheader = CurLoop->getLoopPreheader();
782 IRBuilder<> Builder(Preheader->getTerminator());
783 SCEVExpander Expander(*SE, *DL, "loop-idiom");
785 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
786 Type *IntPtr = Builder.getIntPtrTy(*DL, DestAS);
788 const SCEV *Start = Ev->getStart();
789 // Handle negative strided loops.
791 Start = getStartForNegStride(Start, BECount, IntPtr, StoreSize, SE);
793 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
794 // this into a memset in the loop preheader now if we want. However, this
795 // would be unsafe to do if there is anything else in the loop that may read
796 // or write to the aliased location. Check for any overlap by generating the
797 // base pointer and checking the region.
799 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
800 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
803 // If we generated new code for the base pointer, clean up.
804 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
808 if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset))
811 // Okay, everything looks good, insert the memset.
813 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
814 // pointer size if it isn't already.
815 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
817 const SCEV *NumBytesS =
818 SE->getAddExpr(BECount, SE->getOne(IntPtr), SCEV::FlagNUW);
819 if (StoreSize != 1) {
820 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
825 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
830 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
832 // Everything is emitted in default address space
833 Type *Int8PtrTy = DestInt8PtrTy;
835 Module *M = TheStore->getModule();
837 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
838 Int8PtrTy, Int8PtrTy, IntPtr, (void *)nullptr);
839 inferLibFuncAttributes(*M->getFunction("memset_pattern16"), *TLI);
841 // Otherwise we should form a memset_pattern16. PatternValue is known to be
842 // an constant array of 16-bytes. Plop the value into a mergable global.
843 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
844 GlobalValue::PrivateLinkage,
845 PatternValue, ".memset_pattern");
846 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these.
847 GV->setAlignment(16);
848 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
849 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
852 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
853 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
854 NewCall->setDebugLoc(TheStore->getDebugLoc());
856 // Okay, the memset has been formed. Zap the original store and anything that
858 for (auto *I : Stores)
859 deleteDeadInstruction(I);
864 /// If the stored value is a strided load in the same loop with the same stride
865 /// this may be transformable into a memcpy. This kicks in for stuff like
866 /// for (i) A[i] = B[i];
867 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI,
868 const SCEV *BECount) {
869 assert(SI->isSimple() && "Expected only non-volatile stores.");
871 Value *StorePtr = SI->getPointerOperand();
872 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
873 APInt Stride = getStoreStride(StoreEv);
874 unsigned StoreSize = getStoreSizeInBytes(SI, DL);
875 bool NegStride = StoreSize == -Stride;
877 // The store must be feeding a non-volatile load.
878 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
879 assert(LI->isSimple() && "Expected only non-volatile stores.");
881 // See if the pointer expression is an AddRec like {base,+,1} on the current
882 // loop, which indicates a strided load. If we have something else, it's a
883 // random load we can't handle.
884 const SCEVAddRecExpr *LoadEv =
885 cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
887 // The trip count of the loop and the base pointer of the addrec SCEV is
888 // guaranteed to be loop invariant, which means that it should dominate the
889 // header. This allows us to insert code for it in the preheader.
890 BasicBlock *Preheader = CurLoop->getLoopPreheader();
891 IRBuilder<> Builder(Preheader->getTerminator());
892 SCEVExpander Expander(*SE, *DL, "loop-idiom");
894 const SCEV *StrStart = StoreEv->getStart();
895 unsigned StrAS = SI->getPointerAddressSpace();
896 Type *IntPtrTy = Builder.getIntPtrTy(*DL, StrAS);
898 // Handle negative strided loops.
900 StrStart = getStartForNegStride(StrStart, BECount, IntPtrTy, StoreSize, SE);
902 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
903 // this into a memcpy in the loop preheader now if we want. However, this
904 // would be unsafe to do if there is anything else in the loop that may read
905 // or write the memory region we're storing to. This includes the load that
906 // feeds the stores. Check for an alias by generating the base address and
907 // checking everything.
908 Value *StoreBasePtr = Expander.expandCodeFor(
909 StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator());
911 SmallPtrSet<Instruction *, 1> Stores;
913 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
914 StoreSize, *AA, Stores)) {
916 // If we generated new code for the base pointer, clean up.
917 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
921 const SCEV *LdStart = LoadEv->getStart();
922 unsigned LdAS = LI->getPointerAddressSpace();
924 // Handle negative strided loops.
926 LdStart = getStartForNegStride(LdStart, BECount, IntPtrTy, StoreSize, SE);
928 // For a memcpy, we have to make sure that the input array is not being
929 // mutated by the loop.
930 Value *LoadBasePtr = Expander.expandCodeFor(
931 LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator());
933 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
936 // If we generated new code for the base pointer, clean up.
937 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
938 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
942 if (avoidLIRForMultiBlockLoop())
945 // Okay, everything is safe, we can transform this!
947 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
948 // pointer size if it isn't already.
949 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
951 const SCEV *NumBytesS =
952 SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW);
954 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
958 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
961 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
962 std::min(SI->getAlignment(), LI->getAlignment()));
963 NewCall->setDebugLoc(SI->getDebugLoc());
965 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
966 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
967 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
969 // Okay, the memcpy has been formed. Zap the original store and anything that
971 deleteDeadInstruction(SI);
976 // When compiling for codesize we avoid idiom recognition for a multi-block loop
977 // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop.
979 bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
981 if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) {
982 if (!CurLoop->getParentLoop() && (!IsMemset || !IsLoopMemset)) {
983 DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName()
984 << " : LIR " << (IsMemset ? "Memset" : "Memcpy")
985 << " avoided: multi-block top-level loop\n");
993 bool LoopIdiomRecognize::runOnNoncountableLoop() {
994 return recognizePopcount();
997 /// Check if the given conditional branch is based on the comparison between
998 /// a variable and zero, and if the variable is non-zero, the control yields to
999 /// the loop entry. If the branch matches the behavior, the variable involved
1000 /// in the comparison is returned. This function will be called to see if the
1001 /// precondition and postcondition of the loop are in desirable form.
1002 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
1003 if (!BI || !BI->isConditional())
1006 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1010 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
1011 if (!CmpZero || !CmpZero->isZero())
1014 ICmpInst::Predicate Pred = Cond->getPredicate();
1015 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
1016 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
1017 return Cond->getOperand(0);
1022 /// Return true iff the idiom is detected in the loop.
1025 /// 1) \p CntInst is set to the instruction counting the population bit.
1026 /// 2) \p CntPhi is set to the corresponding phi node.
1027 /// 3) \p Var is set to the value whose population bits are being counted.
1029 /// The core idiom we are trying to detect is:
1032 /// goto loop-exit // the precondition of the loop
1033 /// cnt0 = init-val;
1035 /// x1 = phi (x0, x2);
1036 /// cnt1 = phi(cnt0, cnt2);
1038 /// cnt2 = cnt1 + 1;
1040 /// x2 = x1 & (x1 - 1);
1042 /// } while(x != 0);
1046 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
1047 Instruction *&CntInst, PHINode *&CntPhi,
1049 // step 1: Check to see if the look-back branch match this pattern:
1050 // "if (a!=0) goto loop-entry".
1051 BasicBlock *LoopEntry;
1052 Instruction *DefX2, *CountInst;
1053 Value *VarX1, *VarX0;
1054 PHINode *PhiX, *CountPhi;
1056 DefX2 = CountInst = nullptr;
1057 VarX1 = VarX0 = nullptr;
1058 PhiX = CountPhi = nullptr;
1059 LoopEntry = *(CurLoop->block_begin());
1061 // step 1: Check if the loop-back branch is in desirable form.
1063 if (Value *T = matchCondition(
1064 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1065 DefX2 = dyn_cast<Instruction>(T);
1070 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
1072 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
1075 BinaryOperator *SubOneOp;
1077 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
1078 VarX1 = DefX2->getOperand(1);
1080 VarX1 = DefX2->getOperand(0);
1081 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
1086 Instruction *SubInst = cast<Instruction>(SubOneOp);
1087 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
1089 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
1090 (SubInst->getOpcode() == Instruction::Add &&
1091 Dec->isAllOnesValue()))) {
1096 // step 3: Check the recurrence of variable X
1098 PhiX = dyn_cast<PHINode>(VarX1);
1100 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
1105 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
1107 CountInst = nullptr;
1108 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1109 IterE = LoopEntry->end();
1110 Iter != IterE; Iter++) {
1111 Instruction *Inst = &*Iter;
1112 if (Inst->getOpcode() != Instruction::Add)
1115 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1116 if (!Inc || !Inc->isOne())
1119 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
1120 if (!Phi || Phi->getParent() != LoopEntry)
1123 // Check if the result of the instruction is live of the loop.
1124 bool LiveOutLoop = false;
1125 for (User *U : Inst->users()) {
1126 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
1143 // step 5: check if the precondition is in this form:
1144 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
1146 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1147 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
1148 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
1151 CntInst = CountInst;
1159 /// Recognizes a population count idiom in a non-countable loop.
1161 /// If detected, transforms the relevant code to issue the popcount intrinsic
1162 /// function call, and returns true; otherwise, returns false.
1163 bool LoopIdiomRecognize::recognizePopcount() {
1164 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
1167 // Counting population are usually conducted by few arithmetic instructions.
1168 // Such instructions can be easily "absorbed" by vacant slots in a
1169 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
1170 // in a compact loop.
1172 // Give up if the loop has multiple blocks or multiple backedges.
1173 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1176 BasicBlock *LoopBody = *(CurLoop->block_begin());
1177 if (LoopBody->size() >= 20) {
1178 // The loop is too big, bail out.
1182 // It should have a preheader containing nothing but an unconditional branch.
1183 BasicBlock *PH = CurLoop->getLoopPreheader();
1184 if (!PH || &PH->front() != PH->getTerminator())
1186 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
1187 if (!EntryBI || EntryBI->isConditional())
1190 // It should have a precondition block where the generated popcount instrinsic
1191 // function can be inserted.
1192 auto *PreCondBB = PH->getSinglePredecessor();
1195 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1196 if (!PreCondBI || PreCondBI->isUnconditional())
1199 Instruction *CntInst;
1202 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
1205 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
1209 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1210 const DebugLoc &DL) {
1211 Value *Ops[] = {Val};
1212 Type *Tys[] = {Val->getType()};
1214 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1215 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
1216 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1217 CI->setDebugLoc(DL);
1222 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
1223 Instruction *CntInst,
1224 PHINode *CntPhi, Value *Var) {
1225 BasicBlock *PreHead = CurLoop->getLoopPreheader();
1226 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1227 const DebugLoc DL = CntInst->getDebugLoc();
1229 // Assuming before transformation, the loop is following:
1230 // if (x) // the precondition
1231 // do { cnt++; x &= x - 1; } while(x);
1233 // Step 1: Insert the ctpop instruction at the end of the precondition block
1234 IRBuilder<> Builder(PreCondBr);
1235 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
1237 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
1238 NewCount = PopCntZext =
1239 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
1241 if (NewCount != PopCnt)
1242 (cast<Instruction>(NewCount))->setDebugLoc(DL);
1244 // TripCnt is exactly the number of iterations the loop has
1247 // If the population counter's initial value is not zero, insert Add Inst.
1248 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
1249 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
1250 if (!InitConst || !InitConst->isZero()) {
1251 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
1252 (cast<Instruction>(NewCount))->setDebugLoc(DL);
1256 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
1257 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
1258 // function would be partial dead code, and downstream passes will drag
1259 // it back from the precondition block to the preheader.
1261 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
1263 Value *Opnd0 = PopCntZext;
1264 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
1265 if (PreCond->getOperand(0) != Var)
1266 std::swap(Opnd0, Opnd1);
1268 ICmpInst *NewPreCond = cast<ICmpInst>(
1269 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
1270 PreCondBr->setCondition(NewPreCond);
1272 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
1275 // Step 3: Note that the population count is exactly the trip count of the
1276 // loop in question, which enable us to to convert the loop from noncountable
1277 // loop into a countable one. The benefit is twofold:
1279 // - If the loop only counts population, the entire loop becomes dead after
1280 // the transformation. It is a lot easier to prove a countable loop dead
1281 // than to prove a noncountable one. (In some C dialects, an infinite loop
1282 // isn't dead even if it computes nothing useful. In general, DCE needs
1283 // to prove a noncountable loop finite before safely delete it.)
1285 // - If the loop also performs something else, it remains alive.
1286 // Since it is transformed to countable form, it can be aggressively
1287 // optimized by some optimizations which are in general not applicable
1288 // to a noncountable loop.
1290 // After this step, this loop (conceptually) would look like following:
1291 // newcnt = __builtin_ctpop(x);
1294 // do { cnt++; x &= x-1; t--) } while (t > 0);
1295 BasicBlock *Body = *(CurLoop->block_begin());
1297 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
1298 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1299 Type *Ty = TripCnt->getType();
1301 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1303 Builder.SetInsertPoint(LbCond);
1304 Instruction *TcDec = cast<Instruction>(
1305 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1306 "tcdec", false, true));
1308 TcPhi->addIncoming(TripCnt, PreHead);
1309 TcPhi->addIncoming(TcDec, Body);
1311 CmpInst::Predicate Pred =
1312 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
1313 LbCond->setPredicate(Pred);
1314 LbCond->setOperand(0, TcDec);
1315 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1318 // Step 4: All the references to the original population counter outside
1319 // the loop are replaced with the NewCount -- the value returned from
1320 // __builtin_ctpop().
1321 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1323 // step 5: Forget the "non-computable" trip-count SCEV associated with the
1324 // loop. The loop would otherwise not be deleted even if it becomes empty.
1325 SE->forgetLoop(CurLoop);