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 /// \name Countable Loop Idiom Handling
113 bool runOnCountableLoop();
114 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
115 SmallVectorImpl<BasicBlock *> &ExitBlocks);
117 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
118 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
120 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
121 unsigned StoreAlignment, Value *SplatValue,
122 Instruction *TheStore, const SCEVAddRecExpr *Ev,
123 const SCEV *BECount, bool NegStride);
124 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
125 const SCEVAddRecExpr *StoreEv,
126 const SCEVAddRecExpr *LoadEv,
127 const SCEV *BECount);
130 /// \name Noncountable Loop Idiom Handling
133 bool runOnNoncountableLoop();
135 bool recognizePopcount();
136 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
137 PHINode *CntPhi, Value *Var);
142 } // End anonymous namespace.
144 char LoopIdiomRecognize::ID = 0;
145 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
147 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
148 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
149 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
150 INITIALIZE_PASS_DEPENDENCY(LCSSA)
151 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
152 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
153 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
154 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
155 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
156 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
157 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
158 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
161 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
163 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
164 /// and zero out all the operands of this instruction. If any of them become
165 /// dead, delete them and the computation tree that feeds them.
167 static void deleteDeadInstruction(Instruction *I,
168 const TargetLibraryInfo *TLI) {
169 SmallVector<Value *, 16> Operands(I->value_op_begin(), I->value_op_end());
170 I->replaceAllUsesWith(UndefValue::get(I->getType()));
171 I->eraseFromParent();
172 for (Value *Op : Operands)
173 RecursivelyDeleteTriviallyDeadInstructions(Op, TLI);
176 //===----------------------------------------------------------------------===//
178 // Implementation of LoopIdiomRecognize
180 //===----------------------------------------------------------------------===//
182 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
183 if (skipOptnoneFunction(L))
187 // If the loop could not be converted to canonical form, it must have an
188 // indirectbr in it, just give up.
189 if (!L->getLoopPreheader())
192 // Disable loop idiom recognition if the function's name is a common idiom.
193 StringRef Name = L->getHeader()->getParent()->getName();
194 if (Name == "memset" || Name == "memcpy")
197 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
198 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
199 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
200 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
201 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
202 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
203 *CurLoop->getHeader()->getParent());
204 DL = &CurLoop->getHeader()->getModule()->getDataLayout();
206 if (SE->hasLoopInvariantBackedgeTakenCount(L))
207 return runOnCountableLoop();
209 return runOnNoncountableLoop();
212 bool LoopIdiomRecognize::runOnCountableLoop() {
213 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
214 assert(!isa<SCEVCouldNotCompute>(BECount) &&
215 "runOnCountableLoop() called on a loop without a predictable"
216 "backedge-taken count");
218 // If this loop executes exactly one time, then it should be peeled, not
219 // optimized by this pass.
220 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
221 if (BECst->getValue()->getValue() == 0)
224 SmallVector<BasicBlock *, 8> ExitBlocks;
225 CurLoop->getUniqueExitBlocks(ExitBlocks);
227 DEBUG(dbgs() << "loop-idiom Scanning: F["
228 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
229 << CurLoop->getHeader()->getName() << "\n");
231 bool MadeChange = false;
232 // Scan all the blocks in the loop that are not in subloops.
233 for (auto *BB : CurLoop->getBlocks()) {
234 // Ignore blocks in subloops.
235 if (LI->getLoopFor(BB) != CurLoop)
238 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
243 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
244 /// with the specified backedge count. This block is known to be in the current
245 /// loop and not in any subloops.
246 bool LoopIdiomRecognize::runOnLoopBlock(
247 BasicBlock *BB, const SCEV *BECount,
248 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
249 // We can only promote stores in this block if they are unconditionally
250 // executed in the loop. For a block to be unconditionally executed, it has
251 // to dominate all the exit blocks of the loop. Verify this now.
252 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
253 if (!DT->dominates(BB, ExitBlocks[i]))
256 bool MadeChange = false;
257 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
258 Instruction *Inst = &*I++;
259 // Look for store instructions, which may be optimized to memset/memcpy.
260 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
262 if (!processLoopStore(SI, BECount))
266 // If processing the store invalidated our iterator, start over from the
273 // Look for memset instructions, which may be optimized to a larger memset.
274 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
276 if (!processLoopMemSet(MSI, BECount))
280 // If processing the memset invalidated our iterator, start over from the
291 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
292 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
296 Value *StoredVal = SI->getValueOperand();
297 Value *StorePtr = SI->getPointerOperand();
299 // Reject stores that are so large that they overflow an unsigned.
300 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
301 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
304 // See if the pointer expression is an AddRec like {base,+,1} on the current
305 // loop, which indicates a strided store. If we have something else, it's a
306 // random store we can't handle.
307 const SCEVAddRecExpr *StoreEv =
308 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
309 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
312 // Check to see if the stride matches the size of the store. If so, then we
313 // know that every byte is touched in the loop.
314 unsigned StoreSize = (unsigned)SizeInBits >> 3;
316 const SCEVConstant *ConstStride =
317 dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
321 APInt Stride = ConstStride->getValue()->getValue();
322 if (StoreSize != Stride && StoreSize != -Stride)
325 bool NegStride = StoreSize == -Stride;
327 // See if we can optimize just this store in isolation.
328 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
329 StoredVal, SI, StoreEv, BECount, NegStride))
332 // TODO: We don't handle negative stride memcpys.
336 // If the stored value is a strided load in the same loop with the same stride
337 // this may be transformable into a memcpy. This kicks in for stuff like
338 // for (i) A[i] = B[i];
339 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
340 const SCEVAddRecExpr *LoadEv =
341 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
342 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
343 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
344 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
347 // errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
352 /// processLoopMemSet - See if this memset can be promoted to a large memset.
353 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
354 const SCEV *BECount) {
355 // We can only handle non-volatile memsets with a constant size.
356 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
359 // If we're not allowed to hack on memset, we fail.
360 if (!TLI->has(LibFunc::memset))
363 Value *Pointer = MSI->getDest();
365 // See if the pointer expression is an AddRec like {base,+,1} on the current
366 // loop, which indicates a strided store. If we have something else, it's a
367 // random store we can't handle.
368 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
369 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
372 // Reject memsets that are so large that they overflow an unsigned.
373 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
374 if ((SizeInBytes >> 32) != 0)
377 // Check to see if the stride matches the size of the memset. If so, then we
378 // know that every byte is touched in the loop.
379 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
381 // TODO: Could also handle negative stride here someday, that will require the
382 // validity check in mayLoopAccessLocation to be updated though.
383 if (!Stride || MSI->getLength() != Stride->getValue())
386 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
387 MSI->getAlignment(), MSI->getValue(), MSI, Ev,
388 BECount, /*NegStride=*/false);
391 /// mayLoopAccessLocation - Return true if the specified loop might access the
392 /// specified pointer location, which is a loop-strided access. The 'Access'
393 /// argument specifies what the verboten forms of access are (read or write).
394 static bool mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
395 const SCEV *BECount, unsigned StoreSize,
397 Instruction *IgnoredStore) {
398 // Get the location that may be stored across the loop. Since the access is
399 // strided positively through memory, we say that the modified location starts
400 // at the pointer and has infinite size.
401 uint64_t AccessSize = MemoryLocation::UnknownSize;
403 // If the loop iterates a fixed number of times, we can refine the access size
404 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
405 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
406 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
408 // TODO: For this to be really effective, we have to dive into the pointer
409 // operand in the store. Store to &A[i] of 100 will always return may alias
410 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
411 // which will then no-alias a store to &A[100].
412 MemoryLocation StoreLoc(Ptr, AccessSize);
414 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
416 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
417 if (&*I != IgnoredStore && (AA.getModRefInfo(&*I, StoreLoc) & Access))
423 /// getMemSetPatternValue - If a strided store of the specified value is safe to
424 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
425 /// be passed in. Otherwise, return null.
427 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
428 /// just replicate their input array and then pass on to memset_pattern16.
429 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
430 // If the value isn't a constant, we can't promote it to being in a constant
431 // array. We could theoretically do a store to an alloca or something, but
432 // that doesn't seem worthwhile.
433 Constant *C = dyn_cast<Constant>(V);
437 // Only handle simple values that are a power of two bytes in size.
438 uint64_t Size = DL->getTypeSizeInBits(V->getType());
439 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
442 // Don't care enough about darwin/ppc to implement this.
443 if (DL->isBigEndian())
446 // Convert to size in bytes.
449 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
450 // if the top and bottom are the same (e.g. for vectors and large integers).
454 // If the constant is exactly 16 bytes, just use it.
458 // Otherwise, we'll use an array of the constants.
459 unsigned ArraySize = 16 / Size;
460 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
461 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
464 /// processLoopStridedStore - We see a strided store of some value. If we can
465 /// transform this into a memset or memset_pattern in the loop preheader, do so.
466 bool LoopIdiomRecognize::processLoopStridedStore(
467 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
468 Value *StoredVal, Instruction *TheStore, const SCEVAddRecExpr *Ev,
469 const SCEV *BECount, bool NegStride) {
471 // If the stored value is a byte-wise value (like i32 -1), then it may be
472 // turned into a memset of i8 -1, assuming that all the consecutive bytes
473 // are stored. A store of i32 0x01020304 can never be turned into a memset,
474 // but it can be turned into memset_pattern if the target supports it.
475 Value *SplatValue = isBytewiseValue(StoredVal);
476 Constant *PatternValue = nullptr;
477 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
479 // If we're allowed to form a memset, and the stored value would be acceptable
480 // for memset, use it.
481 if (SplatValue && TLI->has(LibFunc::memset) &&
482 // Verify that the stored value is loop invariant. If not, we can't
483 // promote the memset.
484 CurLoop->isLoopInvariant(SplatValue)) {
485 // Keep and use SplatValue.
486 PatternValue = nullptr;
487 } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) &&
488 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
489 // Don't create memset_pattern16s with address spaces.
490 // It looks like we can use PatternValue!
491 SplatValue = nullptr;
493 // Otherwise, this isn't an idiom we can transform. For example, we can't
494 // do anything with a 3-byte store.
498 // The trip count of the loop and the base pointer of the addrec SCEV is
499 // guaranteed to be loop invariant, which means that it should dominate the
500 // header. This allows us to insert code for it in the preheader.
501 BasicBlock *Preheader = CurLoop->getLoopPreheader();
502 IRBuilder<> Builder(Preheader->getTerminator());
503 SCEVExpander Expander(*SE, *DL, "loop-idiom");
505 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
506 Type *IntPtr = Builder.getIntPtrTy(*DL, DestAS);
508 const SCEV *Start = Ev->getStart();
509 // If we have a negative stride, Start refers to the end of the memory
510 // location we're trying to memset. Therefore, we need to recompute the start
511 // point, which is just Start - BECount*Size.
513 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
515 Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize),
517 Start = SE->getMinusSCEV(Ev->getStart(), Index);
520 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
521 // this into a memset in the loop preheader now if we want. However, this
522 // would be unsafe to do if there is anything else in the loop that may read
523 // or write to the aliased location. Check for any overlap by generating the
524 // base pointer and checking the region.
526 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
527 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
530 // If we generated new code for the base pointer, clean up.
531 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
535 // Okay, everything looks good, insert the memset.
537 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
538 // pointer size if it isn't already.
539 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
541 const SCEV *NumBytesS =
542 SE->getAddExpr(BECount, SE->getOne(IntPtr), SCEV::FlagNUW);
543 if (StoreSize != 1) {
544 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
549 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
554 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
556 // Everything is emitted in default address space
557 Type *Int8PtrTy = DestInt8PtrTy;
559 Module *M = TheStore->getParent()->getParent()->getParent();
561 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
562 Int8PtrTy, Int8PtrTy, IntPtr, (void *)nullptr);
564 // Otherwise we should form a memset_pattern16. PatternValue is known to be
565 // an constant array of 16-bytes. Plop the value into a mergable global.
566 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
567 GlobalValue::PrivateLinkage,
568 PatternValue, ".memset_pattern");
569 GV->setUnnamedAddr(true); // Ok to merge these.
570 GV->setAlignment(16);
571 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
572 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
575 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
576 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
577 NewCall->setDebugLoc(TheStore->getDebugLoc());
579 // Okay, the memset has been formed. Zap the original store and anything that
581 deleteDeadInstruction(TheStore, TLI);
586 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
587 /// same-strided load.
588 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
589 StoreInst *SI, unsigned StoreSize, const SCEVAddRecExpr *StoreEv,
590 const SCEVAddRecExpr *LoadEv, const SCEV *BECount) {
591 // If we're not allowed to form memcpy, we fail.
592 if (!TLI->has(LibFunc::memcpy))
595 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
597 // The trip count of the loop and the base pointer of the addrec SCEV is
598 // guaranteed to be loop invariant, which means that it should dominate the
599 // header. This allows us to insert code for it in the preheader.
600 BasicBlock *Preheader = CurLoop->getLoopPreheader();
601 IRBuilder<> Builder(Preheader->getTerminator());
602 SCEVExpander Expander(*SE, *DL, "loop-idiom");
604 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
605 // this into a memcpy in the loop preheader now if we want. However, this
606 // would be unsafe to do if there is anything else in the loop that may read
607 // or write the memory region we're storing to. This includes the load that
608 // feeds the stores. Check for an alias by generating the base address and
609 // checking everything.
610 Value *StoreBasePtr = Expander.expandCodeFor(
611 StoreEv->getStart(), Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
612 Preheader->getTerminator());
614 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
615 StoreSize, *AA, SI)) {
617 // If we generated new code for the base pointer, clean up.
618 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
622 // For a memcpy, we have to make sure that the input array is not being
623 // mutated by the loop.
624 Value *LoadBasePtr = Expander.expandCodeFor(
625 LoadEv->getStart(), Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
626 Preheader->getTerminator());
628 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
631 // If we generated new code for the base pointer, clean up.
632 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
633 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
637 // Okay, everything is safe, we can transform this!
639 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
640 // pointer size if it isn't already.
641 Type *IntPtrTy = Builder.getIntPtrTy(*DL, SI->getPointerAddressSpace());
642 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
644 const SCEV *NumBytesS =
645 SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW);
647 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
651 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
654 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
655 std::min(SI->getAlignment(), LI->getAlignment()));
656 NewCall->setDebugLoc(SI->getDebugLoc());
658 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
659 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
660 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
662 // Okay, the memcpy has been formed. Zap the original store and anything that
664 deleteDeadInstruction(SI, TLI);
669 bool LoopIdiomRecognize::runOnNoncountableLoop() {
670 if (recognizePopcount())
676 /// Check if the given conditional branch is based on the comparison between
677 /// a variable and zero, and if the variable is non-zero, the control yields to
678 /// the loop entry. If the branch matches the behavior, the variable involved
679 /// in the comparion is returned. This function will be called to see if the
680 /// precondition and postcondition of the loop are in desirable form.
681 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
682 if (!BI || !BI->isConditional())
685 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
689 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
690 if (!CmpZero || !CmpZero->isZero())
693 ICmpInst::Predicate Pred = Cond->getPredicate();
694 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
695 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
696 return Cond->getOperand(0);
701 /// Return true iff the idiom is detected in the loop.
704 /// 1) \p CntInst is set to the instruction counting the population bit.
705 /// 2) \p CntPhi is set to the corresponding phi node.
706 /// 3) \p Var is set to the value whose population bits are being counted.
708 /// The core idiom we are trying to detect is:
711 /// goto loop-exit // the precondition of the loop
714 /// x1 = phi (x0, x2);
715 /// cnt1 = phi(cnt0, cnt2);
719 /// x2 = x1 & (x1 - 1);
725 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
726 Instruction *&CntInst, PHINode *&CntPhi,
728 // step 1: Check to see if the look-back branch match this pattern:
729 // "if (a!=0) goto loop-entry".
730 BasicBlock *LoopEntry;
731 Instruction *DefX2, *CountInst;
732 Value *VarX1, *VarX0;
733 PHINode *PhiX, *CountPhi;
735 DefX2 = CountInst = nullptr;
736 VarX1 = VarX0 = nullptr;
737 PhiX = CountPhi = nullptr;
738 LoopEntry = *(CurLoop->block_begin());
740 // step 1: Check if the loop-back branch is in desirable form.
742 if (Value *T = matchCondition(
743 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
744 DefX2 = dyn_cast<Instruction>(T);
749 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
751 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
754 BinaryOperator *SubOneOp;
756 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
757 VarX1 = DefX2->getOperand(1);
759 VarX1 = DefX2->getOperand(0);
760 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
765 Instruction *SubInst = cast<Instruction>(SubOneOp);
766 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
768 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
769 (SubInst->getOpcode() == Instruction::Add &&
770 Dec->isAllOnesValue()))) {
775 // step 3: Check the recurrence of variable X
777 PhiX = dyn_cast<PHINode>(VarX1);
779 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
784 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
787 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
788 IterE = LoopEntry->end();
789 Iter != IterE; Iter++) {
790 Instruction *Inst = &*Iter;
791 if (Inst->getOpcode() != Instruction::Add)
794 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
795 if (!Inc || !Inc->isOne())
798 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
799 if (!Phi || Phi->getParent() != LoopEntry)
802 // Check if the result of the instruction is live of the loop.
803 bool LiveOutLoop = false;
804 for (User *U : Inst->users()) {
805 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
822 // step 5: check if the precondition is in this form:
823 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
825 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
826 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
827 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
838 /// Recognizes a population count idiom in a non-countable loop.
840 /// If detected, transforms the relevant code to issue the popcount intrinsic
841 /// function call, and returns true; otherwise, returns false.
842 bool LoopIdiomRecognize::recognizePopcount() {
843 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
846 // Counting population are usually conducted by few arithmetic instructions.
847 // Such instructions can be easily "absorbed" by vacant slots in a
848 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
849 // in a compact loop.
851 // Give up if the loop has multiple blocks or multiple backedges.
852 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
855 BasicBlock *LoopBody = *(CurLoop->block_begin());
856 if (LoopBody->size() >= 20) {
857 // The loop is too big, bail out.
861 // It should have a preheader containing nothing but an unconditional branch.
862 BasicBlock *PH = CurLoop->getLoopPreheader();
865 if (&PH->front() != PH->getTerminator())
867 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
868 if (!EntryBI || EntryBI->isConditional())
871 // It should have a precondition block where the generated popcount instrinsic
872 // function can be inserted.
873 auto *PreCondBB = PH->getSinglePredecessor();
876 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
877 if (!PreCondBI || PreCondBI->isUnconditional())
880 Instruction *CntInst;
883 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
886 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
890 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
892 Value *Ops[] = {Val};
893 Type *Tys[] = {Val->getType()};
895 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
896 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
897 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
903 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
904 Instruction *CntInst,
905 PHINode *CntPhi, Value *Var) {
906 BasicBlock *PreHead = CurLoop->getLoopPreheader();
907 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
908 const DebugLoc DL = CntInst->getDebugLoc();
910 // Assuming before transformation, the loop is following:
911 // if (x) // the precondition
912 // do { cnt++; x &= x - 1; } while(x);
914 // Step 1: Insert the ctpop instruction at the end of the precondition block
915 IRBuilder<> Builder(PreCondBr);
916 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
918 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
919 NewCount = PopCntZext =
920 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
922 if (NewCount != PopCnt)
923 (cast<Instruction>(NewCount))->setDebugLoc(DL);
925 // TripCnt is exactly the number of iterations the loop has
928 // If the population counter's initial value is not zero, insert Add Inst.
929 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
930 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
931 if (!InitConst || !InitConst->isZero()) {
932 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
933 (cast<Instruction>(NewCount))->setDebugLoc(DL);
937 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
938 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
939 // function would be partial dead code, and downstream passes will drag
940 // it back from the precondition block to the preheader.
942 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
944 Value *Opnd0 = PopCntZext;
945 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
946 if (PreCond->getOperand(0) != Var)
947 std::swap(Opnd0, Opnd1);
949 ICmpInst *NewPreCond = cast<ICmpInst>(
950 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
951 PreCondBr->setCondition(NewPreCond);
953 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
956 // Step 3: Note that the population count is exactly the trip count of the
957 // loop in question, which enable us to to convert the loop from noncountable
958 // loop into a countable one. The benefit is twofold:
960 // - If the loop only counts population, the entire loop becomes dead after
961 // the transformation. It is a lot easier to prove a countable loop dead
962 // than to prove a noncountable one. (In some C dialects, an infinite loop
963 // isn't dead even if it computes nothing useful. In general, DCE needs
964 // to prove a noncountable loop finite before safely delete it.)
966 // - If the loop also performs something else, it remains alive.
967 // Since it is transformed to countable form, it can be aggressively
968 // optimized by some optimizations which are in general not applicable
969 // to a noncountable loop.
971 // After this step, this loop (conceptually) would look like following:
972 // newcnt = __builtin_ctpop(x);
975 // do { cnt++; x &= x-1; t--) } while (t > 0);
976 BasicBlock *Body = *(CurLoop->block_begin());
978 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
979 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
980 Type *Ty = TripCnt->getType();
982 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
984 Builder.SetInsertPoint(LbCond);
985 Instruction *TcDec = cast<Instruction>(
986 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
987 "tcdec", false, true));
989 TcPhi->addIncoming(TripCnt, PreHead);
990 TcPhi->addIncoming(TcDec, Body);
992 CmpInst::Predicate Pred =
993 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
994 LbCond->setPredicate(Pred);
995 LbCond->setOperand(0, TcDec);
996 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
999 // Step 4: All the references to the original population counter outside
1000 // the loop are replaced with the NewCount -- the value returned from
1001 // __builtin_ctpop().
1002 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1004 // step 5: Forget the "non-computable" trip-count SCEV associated with the
1005 // loop. The loop would otherwise not be deleted even if it becomes empty.
1006 SE->forgetLoop(CurLoop);