1 //===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements an idiom recognizer that transforms simple loops into a
11 // non-loop form. In cases that this kicks in, it can be a significant
14 //===----------------------------------------------------------------------===//
18 // Future loop memory idioms to recognize:
19 // memcmp, memmove, strlen, etc.
20 // Future floating point idioms to recognize in -ffast-math mode:
22 // Future integer operation idioms to recognize:
25 // Beware that isel's default lowering for ctpop is highly inefficient for
26 // i64 and larger types when i64 is legal and the value has few bits set. It
27 // would be good to enhance isel to emit a loop for ctpop in this case.
29 // We should enhance the memset/memcpy recognition to handle multiple stores in
30 // the loop. This would handle things like:
31 // void foo(_Complex float *P)
32 // for (i) { __real__(*P) = 0; __imag__(*P) = 0; }
34 // This could recognize common matrix multiplies and dot product idioms and
35 // replace them with calls to BLAS (if linked in??).
37 //===----------------------------------------------------------------------===//
39 #include "llvm/Transforms/Scalar.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/Analysis/AliasAnalysis.h"
42 #include "llvm/Analysis/BasicAliasAnalysis.h"
43 #include "llvm/Analysis/GlobalsModRef.h"
44 #include "llvm/Analysis/LoopPass.h"
45 #include "llvm/Analysis/ScalarEvolutionExpander.h"
46 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
47 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
48 #include "llvm/Analysis/TargetLibraryInfo.h"
49 #include "llvm/Analysis/TargetTransformInfo.h"
50 #include "llvm/Analysis/ValueTracking.h"
51 #include "llvm/IR/DataLayout.h"
52 #include "llvm/IR/Dominators.h"
53 #include "llvm/IR/IRBuilder.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/Module.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Transforms/Utils/Local.h"
61 #define DEBUG_TYPE "loop-idiom"
63 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
64 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
68 class LoopIdiomRecognize : public LoopPass {
74 TargetLibraryInfo *TLI;
75 const TargetTransformInfo *TTI;
80 explicit LoopIdiomRecognize() : LoopPass(ID) {
81 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
84 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
86 /// This transformation requires natural loop information & requires that
87 /// loop preheaders be inserted into the CFG.
89 void getAnalysisUsage(AnalysisUsage &AU) const override {
90 AU.addRequired<LoopInfoWrapperPass>();
91 AU.addPreserved<LoopInfoWrapperPass>();
92 AU.addRequiredID(LoopSimplifyID);
93 AU.addPreservedID(LoopSimplifyID);
94 AU.addRequiredID(LCSSAID);
95 AU.addPreservedID(LCSSAID);
96 AU.addRequired<AAResultsWrapperPass>();
97 AU.addPreserved<AAResultsWrapperPass>();
98 AU.addRequired<ScalarEvolutionWrapperPass>();
99 AU.addPreserved<ScalarEvolutionWrapperPass>();
100 AU.addPreserved<SCEVAAWrapperPass>();
101 AU.addRequired<DominatorTreeWrapperPass>();
102 AU.addPreserved<DominatorTreeWrapperPass>();
103 AU.addRequired<TargetLibraryInfoWrapperPass>();
104 AU.addRequired<TargetTransformInfoWrapperPass>();
105 AU.addPreserved<BasicAAWrapperPass>();
106 AU.addPreserved<GlobalsAAWrapperPass>();
110 typedef SmallVector<StoreInst *, 8> StoreList;
113 /// \name Countable Loop Idiom Handling
116 bool runOnCountableLoop();
117 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
118 SmallVectorImpl<BasicBlock *> &ExitBlocks);
120 void collectStores(BasicBlock *BB);
121 bool isLegalStore(StoreInst *SI);
122 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
123 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
125 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
126 unsigned StoreAlignment, Value *SplatValue,
127 Instruction *TheStore, const SCEVAddRecExpr *Ev,
128 const SCEV *BECount, bool NegStride);
129 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
130 const SCEVAddRecExpr *StoreEv,
131 const SCEVAddRecExpr *LoadEv,
132 const SCEV *BECount);
135 /// \name Noncountable Loop Idiom Handling
138 bool runOnNoncountableLoop();
140 bool recognizePopcount();
141 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
142 PHINode *CntPhi, Value *Var);
147 } // End anonymous namespace.
149 char LoopIdiomRecognize::ID = 0;
150 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
152 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
153 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
154 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
155 INITIALIZE_PASS_DEPENDENCY(LCSSA)
156 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
157 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
158 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
159 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
160 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
161 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
162 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
163 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
166 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
168 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
169 /// and zero out all the operands of this instruction. If any of them become
170 /// dead, delete them and the computation tree that feeds them.
172 static void deleteDeadInstruction(Instruction *I,
173 const TargetLibraryInfo *TLI) {
174 SmallVector<Value *, 16> Operands(I->value_op_begin(), I->value_op_end());
175 I->replaceAllUsesWith(UndefValue::get(I->getType()));
176 I->eraseFromParent();
177 for (Value *Op : Operands)
178 RecursivelyDeleteTriviallyDeadInstructions(Op, TLI);
181 //===----------------------------------------------------------------------===//
183 // Implementation of LoopIdiomRecognize
185 //===----------------------------------------------------------------------===//
187 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
188 if (skipOptnoneFunction(L))
192 // If the loop could not be converted to canonical form, it must have an
193 // indirectbr in it, just give up.
194 if (!L->getLoopPreheader())
197 // Disable loop idiom recognition if the function's name is a common idiom.
198 StringRef Name = L->getHeader()->getParent()->getName();
199 if (Name == "memset" || Name == "memcpy")
202 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
203 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
204 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
205 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
206 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
207 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
208 *CurLoop->getHeader()->getParent());
209 DL = &CurLoop->getHeader()->getModule()->getDataLayout();
211 if (SE->hasLoopInvariantBackedgeTakenCount(L))
212 return runOnCountableLoop();
214 return runOnNoncountableLoop();
217 bool LoopIdiomRecognize::runOnCountableLoop() {
218 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
219 assert(!isa<SCEVCouldNotCompute>(BECount) &&
220 "runOnCountableLoop() called on a loop without a predictable"
221 "backedge-taken count");
223 // If this loop executes exactly one time, then it should be peeled, not
224 // optimized by this pass.
225 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
226 if (BECst->getValue()->getValue() == 0)
229 SmallVector<BasicBlock *, 8> ExitBlocks;
230 CurLoop->getUniqueExitBlocks(ExitBlocks);
232 DEBUG(dbgs() << "loop-idiom Scanning: F["
233 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
234 << CurLoop->getHeader()->getName() << "\n");
236 bool MadeChange = false;
237 // Scan all the blocks in the loop that are not in subloops.
238 for (auto *BB : CurLoop->getBlocks()) {
239 // Ignore blocks in subloops.
240 if (LI->getLoopFor(BB) != CurLoop)
243 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
248 static unsigned getStoreSizeInBytes(StoreInst *SI, const DataLayout *DL) {
249 uint64_t SizeInBits = DL->getTypeSizeInBits(SI->getValueOperand()->getType());
250 assert(((SizeInBits & 7) || (SizeInBits >> 32) == 0) &&
251 "Don't overflow unsigned.");
252 return (unsigned)SizeInBits >> 3;
255 static unsigned getStoreStride(const SCEVAddRecExpr *StoreEv) {
256 const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
257 return ConstStride->getValue()->getValue().getZExtValue();
260 bool LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
261 Value *StoredVal = SI->getValueOperand();
262 Value *StorePtr = SI->getPointerOperand();
264 // Reject stores that are so large that they overflow an unsigned.
265 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
266 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
269 // See if the pointer expression is an AddRec like {base,+,1} on the current
270 // loop, which indicates a strided store. If we have something else, it's a
271 // random store we can't handle.
272 const SCEVAddRecExpr *StoreEv =
273 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
274 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
277 // Check to see if we have a constant stride.
278 if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
284 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
286 for (Instruction &I : *BB) {
287 StoreInst *SI = dyn_cast<StoreInst>(&I);
291 // Don't touch volatile stores.
295 // Make sure this is a strided store with a constant stride.
296 if (!isLegalStore(SI))
299 // Save the store locations.
300 StoreRefs.push_back(SI);
304 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
305 /// with the specified backedge count. This block is known to be in the current
306 /// loop and not in any subloops.
307 bool LoopIdiomRecognize::runOnLoopBlock(
308 BasicBlock *BB, const SCEV *BECount,
309 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
310 // We can only promote stores in this block if they are unconditionally
311 // executed in the loop. For a block to be unconditionally executed, it has
312 // to dominate all the exit blocks of the loop. Verify this now.
313 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
314 if (!DT->dominates(BB, ExitBlocks[i]))
317 bool MadeChange = false;
318 // Look for store instructions, which may be optimized to memset/memcpy.
320 for (auto &SI : StoreRefs)
321 MadeChange |= processLoopStore(SI, BECount);
323 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
324 Instruction *Inst = &*I++;
325 // Look for memset instructions, which may be optimized to a larger memset.
326 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
328 if (!processLoopMemSet(MSI, BECount))
332 // If processing the memset invalidated our iterator, start over from the
343 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
344 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
345 assert(SI->isSimple() && "Expected only non-volatile stores.");
347 Value *StoredVal = SI->getValueOperand();
348 Value *StorePtr = SI->getPointerOperand();
350 // Check to see if the stride matches the size of the store. If so, then we
351 // know that every byte is touched in the loop.
352 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
353 unsigned Stride = getStoreStride(StoreEv);
354 unsigned StoreSize = getStoreSizeInBytes(SI, DL);
355 if (StoreSize != Stride && StoreSize != -Stride)
358 bool NegStride = StoreSize == -Stride;
360 // See if we can optimize just this store in isolation.
361 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
362 StoredVal, SI, StoreEv, BECount, NegStride))
365 // TODO: We don't handle negative stride memcpys.
369 // If the stored value is a strided load in the same loop with the same stride
370 // this may be transformable into a memcpy. This kicks in for stuff like
371 // for (i) A[i] = B[i];
372 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
373 const SCEVAddRecExpr *LoadEv =
374 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
375 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
376 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
377 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
380 // errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
385 /// processLoopMemSet - See if this memset can be promoted to a large memset.
386 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
387 const SCEV *BECount) {
388 // We can only handle non-volatile memsets with a constant size.
389 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
392 // If we're not allowed to hack on memset, we fail.
393 if (!TLI->has(LibFunc::memset))
396 Value *Pointer = MSI->getDest();
398 // See if the pointer expression is an AddRec like {base,+,1} on the current
399 // loop, which indicates a strided store. If we have something else, it's a
400 // random store we can't handle.
401 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
402 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
405 // Reject memsets that are so large that they overflow an unsigned.
406 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
407 if ((SizeInBytes >> 32) != 0)
410 // Check to see if the stride matches the size of the memset. If so, then we
411 // know that every byte is touched in the loop.
412 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
414 // TODO: Could also handle negative stride here someday, that will require the
415 // validity check in mayLoopAccessLocation to be updated though.
416 if (!Stride || MSI->getLength() != Stride->getValue())
419 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
420 MSI->getAlignment(), MSI->getValue(), MSI, Ev,
421 BECount, /*NegStride=*/false);
424 /// mayLoopAccessLocation - Return true if the specified loop might access the
425 /// specified pointer location, which is a loop-strided access. The 'Access'
426 /// argument specifies what the verboten forms of access are (read or write).
427 static bool mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
428 const SCEV *BECount, unsigned StoreSize,
430 Instruction *IgnoredStore) {
431 // Get the location that may be stored across the loop. Since the access is
432 // strided positively through memory, we say that the modified location starts
433 // at the pointer and has infinite size.
434 uint64_t AccessSize = MemoryLocation::UnknownSize;
436 // If the loop iterates a fixed number of times, we can refine the access size
437 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
438 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
439 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
441 // TODO: For this to be really effective, we have to dive into the pointer
442 // operand in the store. Store to &A[i] of 100 will always return may alias
443 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
444 // which will then no-alias a store to &A[100].
445 MemoryLocation StoreLoc(Ptr, AccessSize);
447 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
449 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
450 if (&*I != IgnoredStore && (AA.getModRefInfo(&*I, StoreLoc) & Access))
456 /// getMemSetPatternValue - If a strided store of the specified value is safe to
457 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
458 /// be passed in. Otherwise, return null.
460 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
461 /// just replicate their input array and then pass on to memset_pattern16.
462 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
463 // If the value isn't a constant, we can't promote it to being in a constant
464 // array. We could theoretically do a store to an alloca or something, but
465 // that doesn't seem worthwhile.
466 Constant *C = dyn_cast<Constant>(V);
470 // Only handle simple values that are a power of two bytes in size.
471 uint64_t Size = DL->getTypeSizeInBits(V->getType());
472 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
475 // Don't care enough about darwin/ppc to implement this.
476 if (DL->isBigEndian())
479 // Convert to size in bytes.
482 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
483 // if the top and bottom are the same (e.g. for vectors and large integers).
487 // If the constant is exactly 16 bytes, just use it.
491 // Otherwise, we'll use an array of the constants.
492 unsigned ArraySize = 16 / Size;
493 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
494 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
497 // If we have a negative stride, Start refers to the end of the memory location
498 // we're trying to memset. Therefore, we need to recompute the base pointer,
499 // which is just Start - BECount*Size.
500 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
501 Type *IntPtr, unsigned StoreSize,
502 ScalarEvolution *SE) {
503 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
505 Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize),
507 return SE->getMinusSCEV(Start, Index);
510 /// processLoopStridedStore - We see a strided store of some value. If we can
511 /// transform this into a memset or memset_pattern in the loop preheader, do so.
512 bool LoopIdiomRecognize::processLoopStridedStore(
513 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
514 Value *StoredVal, Instruction *TheStore, const SCEVAddRecExpr *Ev,
515 const SCEV *BECount, bool NegStride) {
517 // If the stored value is a byte-wise value (like i32 -1), then it may be
518 // turned into a memset of i8 -1, assuming that all the consecutive bytes
519 // are stored. A store of i32 0x01020304 can never be turned into a memset,
520 // but it can be turned into memset_pattern if the target supports it.
521 Value *SplatValue = isBytewiseValue(StoredVal);
522 Constant *PatternValue = nullptr;
523 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
525 // If we're allowed to form a memset, and the stored value would be acceptable
526 // for memset, use it.
527 if (SplatValue && TLI->has(LibFunc::memset) &&
528 // Verify that the stored value is loop invariant. If not, we can't
529 // promote the memset.
530 CurLoop->isLoopInvariant(SplatValue)) {
531 // Keep and use SplatValue.
532 PatternValue = nullptr;
533 } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) &&
534 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
535 // Don't create memset_pattern16s with address spaces.
536 // It looks like we can use PatternValue!
537 SplatValue = nullptr;
539 // Otherwise, this isn't an idiom we can transform. For example, we can't
540 // do anything with a 3-byte store.
544 // The trip count of the loop and the base pointer of the addrec SCEV is
545 // guaranteed to be loop invariant, which means that it should dominate the
546 // header. This allows us to insert code for it in the preheader.
547 BasicBlock *Preheader = CurLoop->getLoopPreheader();
548 IRBuilder<> Builder(Preheader->getTerminator());
549 SCEVExpander Expander(*SE, *DL, "loop-idiom");
551 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
552 Type *IntPtr = Builder.getIntPtrTy(*DL, DestAS);
554 const SCEV *Start = Ev->getStart();
555 // Handle negative strided loops.
557 Start = getStartForNegStride(Start, BECount, IntPtr, StoreSize, SE);
559 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
560 // this into a memset in the loop preheader now if we want. However, this
561 // would be unsafe to do if there is anything else in the loop that may read
562 // or write to the aliased location. Check for any overlap by generating the
563 // base pointer and checking the region.
565 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
566 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
569 // If we generated new code for the base pointer, clean up.
570 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
574 // Okay, everything looks good, insert the memset.
576 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
577 // pointer size if it isn't already.
578 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
580 const SCEV *NumBytesS =
581 SE->getAddExpr(BECount, SE->getOne(IntPtr), SCEV::FlagNUW);
582 if (StoreSize != 1) {
583 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
588 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
593 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
595 // Everything is emitted in default address space
596 Type *Int8PtrTy = DestInt8PtrTy;
598 Module *M = TheStore->getParent()->getParent()->getParent();
600 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
601 Int8PtrTy, Int8PtrTy, IntPtr, (void *)nullptr);
603 // Otherwise we should form a memset_pattern16. PatternValue is known to be
604 // an constant array of 16-bytes. Plop the value into a mergable global.
605 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
606 GlobalValue::PrivateLinkage,
607 PatternValue, ".memset_pattern");
608 GV->setUnnamedAddr(true); // Ok to merge these.
609 GV->setAlignment(16);
610 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
611 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
614 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
615 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
616 NewCall->setDebugLoc(TheStore->getDebugLoc());
618 // Okay, the memset has been formed. Zap the original store and anything that
620 deleteDeadInstruction(TheStore, TLI);
625 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
626 /// same-strided load.
627 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
628 StoreInst *SI, unsigned StoreSize, const SCEVAddRecExpr *StoreEv,
629 const SCEVAddRecExpr *LoadEv, const SCEV *BECount) {
630 // If we're not allowed to form memcpy, we fail.
631 if (!TLI->has(LibFunc::memcpy))
634 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
636 // The trip count of the loop and the base pointer of the addrec SCEV is
637 // guaranteed to be loop invariant, which means that it should dominate the
638 // header. This allows us to insert code for it in the preheader.
639 BasicBlock *Preheader = CurLoop->getLoopPreheader();
640 IRBuilder<> Builder(Preheader->getTerminator());
641 SCEVExpander Expander(*SE, *DL, "loop-idiom");
643 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
644 // this into a memcpy in the loop preheader now if we want. However, this
645 // would be unsafe to do if there is anything else in the loop that may read
646 // or write the memory region we're storing to. This includes the load that
647 // feeds the stores. Check for an alias by generating the base address and
648 // checking everything.
649 Value *StoreBasePtr = Expander.expandCodeFor(
650 StoreEv->getStart(), Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
651 Preheader->getTerminator());
653 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
654 StoreSize, *AA, SI)) {
656 // If we generated new code for the base pointer, clean up.
657 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
661 // For a memcpy, we have to make sure that the input array is not being
662 // mutated by the loop.
663 Value *LoadBasePtr = Expander.expandCodeFor(
664 LoadEv->getStart(), Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
665 Preheader->getTerminator());
667 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
670 // If we generated new code for the base pointer, clean up.
671 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
672 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
676 // Okay, everything is safe, we can transform this!
678 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
679 // pointer size if it isn't already.
680 Type *IntPtrTy = Builder.getIntPtrTy(*DL, SI->getPointerAddressSpace());
681 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
683 const SCEV *NumBytesS =
684 SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW);
686 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
690 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
693 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
694 std::min(SI->getAlignment(), LI->getAlignment()));
695 NewCall->setDebugLoc(SI->getDebugLoc());
697 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
698 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
699 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
701 // Okay, the memcpy has been formed. Zap the original store and anything that
703 deleteDeadInstruction(SI, TLI);
708 bool LoopIdiomRecognize::runOnNoncountableLoop() {
709 return recognizePopcount();
712 /// Check if the given conditional branch is based on the comparison between
713 /// a variable and zero, and if the variable is non-zero, the control yields to
714 /// the loop entry. If the branch matches the behavior, the variable involved
715 /// in the comparion is returned. This function will be called to see if the
716 /// precondition and postcondition of the loop are in desirable form.
717 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
718 if (!BI || !BI->isConditional())
721 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
725 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
726 if (!CmpZero || !CmpZero->isZero())
729 ICmpInst::Predicate Pred = Cond->getPredicate();
730 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
731 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
732 return Cond->getOperand(0);
737 /// Return true iff the idiom is detected in the loop.
740 /// 1) \p CntInst is set to the instruction counting the population bit.
741 /// 2) \p CntPhi is set to the corresponding phi node.
742 /// 3) \p Var is set to the value whose population bits are being counted.
744 /// The core idiom we are trying to detect is:
747 /// goto loop-exit // the precondition of the loop
750 /// x1 = phi (x0, x2);
751 /// cnt1 = phi(cnt0, cnt2);
755 /// x2 = x1 & (x1 - 1);
761 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
762 Instruction *&CntInst, PHINode *&CntPhi,
764 // step 1: Check to see if the look-back branch match this pattern:
765 // "if (a!=0) goto loop-entry".
766 BasicBlock *LoopEntry;
767 Instruction *DefX2, *CountInst;
768 Value *VarX1, *VarX0;
769 PHINode *PhiX, *CountPhi;
771 DefX2 = CountInst = nullptr;
772 VarX1 = VarX0 = nullptr;
773 PhiX = CountPhi = nullptr;
774 LoopEntry = *(CurLoop->block_begin());
776 // step 1: Check if the loop-back branch is in desirable form.
778 if (Value *T = matchCondition(
779 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
780 DefX2 = dyn_cast<Instruction>(T);
785 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
787 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
790 BinaryOperator *SubOneOp;
792 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
793 VarX1 = DefX2->getOperand(1);
795 VarX1 = DefX2->getOperand(0);
796 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
801 Instruction *SubInst = cast<Instruction>(SubOneOp);
802 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
804 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
805 (SubInst->getOpcode() == Instruction::Add &&
806 Dec->isAllOnesValue()))) {
811 // step 3: Check the recurrence of variable X
813 PhiX = dyn_cast<PHINode>(VarX1);
815 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
820 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
823 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
824 IterE = LoopEntry->end();
825 Iter != IterE; Iter++) {
826 Instruction *Inst = &*Iter;
827 if (Inst->getOpcode() != Instruction::Add)
830 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
831 if (!Inc || !Inc->isOne())
834 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
835 if (!Phi || Phi->getParent() != LoopEntry)
838 // Check if the result of the instruction is live of the loop.
839 bool LiveOutLoop = false;
840 for (User *U : Inst->users()) {
841 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
858 // step 5: check if the precondition is in this form:
859 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
861 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
862 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
863 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
874 /// Recognizes a population count idiom in a non-countable loop.
876 /// If detected, transforms the relevant code to issue the popcount intrinsic
877 /// function call, and returns true; otherwise, returns false.
878 bool LoopIdiomRecognize::recognizePopcount() {
879 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
882 // Counting population are usually conducted by few arithmetic instructions.
883 // Such instructions can be easily "absorbed" by vacant slots in a
884 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
885 // in a compact loop.
887 // Give up if the loop has multiple blocks or multiple backedges.
888 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
891 BasicBlock *LoopBody = *(CurLoop->block_begin());
892 if (LoopBody->size() >= 20) {
893 // The loop is too big, bail out.
897 // It should have a preheader containing nothing but an unconditional branch.
898 BasicBlock *PH = CurLoop->getLoopPreheader();
901 if (&PH->front() != PH->getTerminator())
903 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
904 if (!EntryBI || EntryBI->isConditional())
907 // It should have a precondition block where the generated popcount instrinsic
908 // function can be inserted.
909 auto *PreCondBB = PH->getSinglePredecessor();
912 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
913 if (!PreCondBI || PreCondBI->isUnconditional())
916 Instruction *CntInst;
919 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
922 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
926 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
928 Value *Ops[] = {Val};
929 Type *Tys[] = {Val->getType()};
931 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
932 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
933 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
939 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
940 Instruction *CntInst,
941 PHINode *CntPhi, Value *Var) {
942 BasicBlock *PreHead = CurLoop->getLoopPreheader();
943 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
944 const DebugLoc DL = CntInst->getDebugLoc();
946 // Assuming before transformation, the loop is following:
947 // if (x) // the precondition
948 // do { cnt++; x &= x - 1; } while(x);
950 // Step 1: Insert the ctpop instruction at the end of the precondition block
951 IRBuilder<> Builder(PreCondBr);
952 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
954 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
955 NewCount = PopCntZext =
956 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
958 if (NewCount != PopCnt)
959 (cast<Instruction>(NewCount))->setDebugLoc(DL);
961 // TripCnt is exactly the number of iterations the loop has
964 // If the population counter's initial value is not zero, insert Add Inst.
965 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
966 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
967 if (!InitConst || !InitConst->isZero()) {
968 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
969 (cast<Instruction>(NewCount))->setDebugLoc(DL);
973 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
974 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
975 // function would be partial dead code, and downstream passes will drag
976 // it back from the precondition block to the preheader.
978 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
980 Value *Opnd0 = PopCntZext;
981 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
982 if (PreCond->getOperand(0) != Var)
983 std::swap(Opnd0, Opnd1);
985 ICmpInst *NewPreCond = cast<ICmpInst>(
986 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
987 PreCondBr->setCondition(NewPreCond);
989 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
992 // Step 3: Note that the population count is exactly the trip count of the
993 // loop in question, which enable us to to convert the loop from noncountable
994 // loop into a countable one. The benefit is twofold:
996 // - If the loop only counts population, the entire loop becomes dead after
997 // the transformation. It is a lot easier to prove a countable loop dead
998 // than to prove a noncountable one. (In some C dialects, an infinite loop
999 // isn't dead even if it computes nothing useful. In general, DCE needs
1000 // to prove a noncountable loop finite before safely delete it.)
1002 // - If the loop also performs something else, it remains alive.
1003 // Since it is transformed to countable form, it can be aggressively
1004 // optimized by some optimizations which are in general not applicable
1005 // to a noncountable loop.
1007 // After this step, this loop (conceptually) would look like following:
1008 // newcnt = __builtin_ctpop(x);
1011 // do { cnt++; x &= x-1; t--) } while (t > 0);
1012 BasicBlock *Body = *(CurLoop->block_begin());
1014 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
1015 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1016 Type *Ty = TripCnt->getType();
1018 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1020 Builder.SetInsertPoint(LbCond);
1021 Instruction *TcDec = cast<Instruction>(
1022 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1023 "tcdec", false, true));
1025 TcPhi->addIncoming(TripCnt, PreHead);
1026 TcPhi->addIncoming(TcDec, Body);
1028 CmpInst::Predicate Pred =
1029 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
1030 LbCond->setPredicate(Pred);
1031 LbCond->setOperand(0, TcDec);
1032 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1035 // Step 4: All the references to the original population counter outside
1036 // the loop are replaced with the NewCount -- the value returned from
1037 // __builtin_ctpop().
1038 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1040 // step 5: Forget the "non-computable" trip-count SCEV associated with the
1041 // loop. The loop would otherwise not be deleted even if it becomes empty.
1042 SE->forgetLoop(CurLoop);