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 // We should enhance this to handle negative strides through memory.
35 // Alternatively (and perhaps better) we could rely on an earlier pass to force
36 // forward iteration through memory, which is generally better for cache
37 // behavior. Negative strides *do* happen for memset/memcpy loops.
39 // This could recognize common matrix multiplies and dot product idioms and
40 // replace them with calls to BLAS (if linked in??).
42 //===----------------------------------------------------------------------===//
44 #include "llvm/Transforms/Scalar.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/Analysis/AliasAnalysis.h"
47 #include "llvm/Analysis/BasicAliasAnalysis.h"
48 #include "llvm/Analysis/GlobalsModRef.h"
49 #include "llvm/Analysis/LoopPass.h"
50 #include "llvm/Analysis/ScalarEvolutionExpander.h"
51 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.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/Utils/Local.h"
66 #define DEBUG_TYPE "loop-idiom"
68 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
69 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
73 class LoopIdiomRecognize : public LoopPass {
79 TargetLibraryInfo *TLI;
80 const TargetTransformInfo *TTI;
84 explicit LoopIdiomRecognize() : LoopPass(ID) {
85 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
88 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
90 /// This transformation requires natural loop information & requires that
91 /// loop preheaders be inserted into the CFG.
93 void getAnalysisUsage(AnalysisUsage &AU) const override {
94 AU.addRequired<LoopInfoWrapperPass>();
95 AU.addPreserved<LoopInfoWrapperPass>();
96 AU.addRequiredID(LoopSimplifyID);
97 AU.addPreservedID(LoopSimplifyID);
98 AU.addRequiredID(LCSSAID);
99 AU.addPreservedID(LCSSAID);
100 AU.addRequired<AAResultsWrapperPass>();
101 AU.addPreserved<AAResultsWrapperPass>();
102 AU.addRequired<ScalarEvolutionWrapperPass>();
103 AU.addPreserved<ScalarEvolutionWrapperPass>();
104 AU.addPreserved<SCEVAAWrapperPass>();
105 AU.addRequired<DominatorTreeWrapperPass>();
106 AU.addPreserved<DominatorTreeWrapperPass>();
107 AU.addRequired<TargetLibraryInfoWrapperPass>();
108 AU.addRequired<TargetTransformInfoWrapperPass>();
109 AU.addPreserved<BasicAAWrapperPass>();
110 AU.addPreserved<GlobalsAAWrapperPass>();
114 /// \name Countable Loop Idiom Handling
117 bool runOnCountableLoop();
118 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
119 SmallVectorImpl<BasicBlock *> &ExitBlocks);
121 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
122 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
124 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
125 unsigned StoreAlignment, Value *SplatValue,
126 Instruction *TheStore, const SCEVAddRecExpr *Ev,
127 const SCEV *BECount);
128 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
129 const SCEVAddRecExpr *StoreEv,
130 const SCEVAddRecExpr *LoadEv,
131 const SCEV *BECount);
134 /// \name Noncountable Loop Idiom Handling
137 bool runOnNoncountableLoop();
139 bool recognizePopcount();
140 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
141 PHINode *CntPhi, Value *Var);
146 } // End anonymous namespace.
148 char LoopIdiomRecognize::ID = 0;
149 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
151 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
152 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
153 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
154 INITIALIZE_PASS_DEPENDENCY(LCSSA)
155 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
156 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
157 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
158 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
159 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
160 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
161 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
162 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
165 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
167 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
168 /// and zero out all the operands of this instruction. If any of them become
169 /// dead, delete them and the computation tree that feeds them.
171 static void deleteDeadInstruction(Instruction *I,
172 const TargetLibraryInfo *TLI) {
173 SmallVector<Value *, 16> Operands(I->value_op_begin(), I->value_op_end());
174 I->replaceAllUsesWith(UndefValue::get(I->getType()));
175 I->eraseFromParent();
176 for (Value *Op : Operands)
177 RecursivelyDeleteTriviallyDeadInstructions(Op, TLI);
180 //===----------------------------------------------------------------------===//
182 // Implementation of LoopIdiomRecognize
184 //===----------------------------------------------------------------------===//
186 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
187 if (skipOptnoneFunction(L))
191 // If the loop could not be converted to canonical form, it must have an
192 // indirectbr in it, just give up.
193 if (!L->getLoopPreheader())
196 // Disable loop idiom recognition if the function's name is a common idiom.
197 StringRef Name = L->getHeader()->getParent()->getName();
198 if (Name == "memset" || Name == "memcpy")
201 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
202 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
203 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
204 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
205 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
206 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
207 *CurLoop->getHeader()->getParent());
209 if (SE->hasLoopInvariantBackedgeTakenCount(L))
210 return runOnCountableLoop();
212 return runOnNoncountableLoop();
215 bool LoopIdiomRecognize::runOnCountableLoop() {
216 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
217 assert(!isa<SCEVCouldNotCompute>(BECount) &&
218 "runOnCountableLoop() called on a loop without a predictable"
219 "backedge-taken count");
221 // If this loop executes exactly one time, then it should be peeled, not
222 // optimized by this pass.
223 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
224 if (BECst->getValue()->getValue() == 0)
227 SmallVector<BasicBlock *, 8> ExitBlocks;
228 CurLoop->getUniqueExitBlocks(ExitBlocks);
230 DEBUG(dbgs() << "loop-idiom Scanning: F["
231 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
232 << CurLoop->getHeader()->getName() << "\n");
234 bool MadeChange = false;
235 // Scan all the blocks in the loop that are not in subloops.
236 for (auto *BB : CurLoop->getBlocks()) {
237 // Ignore blocks in subloops.
238 if (LI->getLoopFor(BB) != CurLoop)
241 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
246 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
247 /// with the specified backedge count. This block is known to be in the current
248 /// loop and not in any subloops.
249 bool LoopIdiomRecognize::runOnLoopBlock(
250 BasicBlock *BB, const SCEV *BECount,
251 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
252 // We can only promote stores in this block if they are unconditionally
253 // executed in the loop. For a block to be unconditionally executed, it has
254 // to dominate all the exit blocks of the loop. Verify this now.
255 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
256 if (!DT->dominates(BB, ExitBlocks[i]))
259 bool MadeChange = false;
260 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
261 Instruction *Inst = I++;
262 // Look for store instructions, which may be optimized to memset/memcpy.
263 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
265 if (!processLoopStore(SI, BECount))
269 // If processing the store invalidated our iterator, start over from the
276 // Look for memset instructions, which may be optimized to a larger memset.
277 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
279 if (!processLoopMemSet(MSI, BECount))
283 // If processing the memset invalidated our iterator, start over from the
294 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
295 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
299 Value *StoredVal = SI->getValueOperand();
300 Value *StorePtr = SI->getPointerOperand();
302 // Reject stores that are so large that they overflow an unsigned.
303 auto &DL = CurLoop->getHeader()->getModule()->getDataLayout();
304 uint64_t SizeInBits = DL.getTypeSizeInBits(StoredVal->getType());
305 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
308 // See if the pointer expression is an AddRec like {base,+,1} on the current
309 // loop, which indicates a strided store. If we have something else, it's a
310 // random store we can't handle.
311 const SCEVAddRecExpr *StoreEv =
312 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
313 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
316 // Check to see if the stride matches the size of the store. If so, then we
317 // know that every byte is touched in the loop.
318 unsigned StoreSize = (unsigned)SizeInBits >> 3;
319 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
321 if (!Stride || StoreSize != Stride->getValue()->getValue()) {
322 // TODO: Could also handle negative stride here someday, that will require
323 // the validity check in mayLoopAccessLocation to be updated though.
324 // Enable this to print exact negative strides.
325 if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
326 dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
327 dbgs() << "BB: " << *SI->getParent();
333 // See if we can optimize just this store in isolation.
334 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
335 StoredVal, SI, StoreEv, BECount))
338 // If the stored value is a strided load in the same loop with the same stride
339 // this this may be transformable into a memcpy. This kicks in for stuff like
340 // for (i) A[i] = B[i];
341 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
342 const SCEVAddRecExpr *LoadEv =
343 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
344 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
345 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
346 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
349 // errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
354 /// processLoopMemSet - See if this memset can be promoted to a large memset.
355 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
356 const SCEV *BECount) {
357 // We can only handle non-volatile memsets with a constant size.
358 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
361 // If we're not allowed to hack on memset, we fail.
362 if (!TLI->has(LibFunc::memset))
365 Value *Pointer = MSI->getDest();
367 // See if the pointer expression is an AddRec like {base,+,1} on the current
368 // loop, which indicates a strided store. If we have something else, it's a
369 // random store we can't handle.
370 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
371 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
374 // Reject memsets that are so large that they overflow an unsigned.
375 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
376 if ((SizeInBytes >> 32) != 0)
379 // Check to see if the stride matches the size of the memset. If so, then we
380 // know that every byte is touched in the loop.
381 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
383 // TODO: Could also handle negative stride here someday, that will require the
384 // validity check in mayLoopAccessLocation to be updated though.
385 if (!Stride || MSI->getLength() != Stride->getValue())
388 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
389 MSI->getAlignment(), MSI->getValue(), MSI, Ev,
393 /// mayLoopAccessLocation - Return true if the specified loop might access the
394 /// specified pointer location, which is a loop-strided access. The 'Access'
395 /// argument specifies what the verboten forms of access are (read or write).
396 static bool mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
397 const SCEV *BECount, unsigned StoreSize,
399 Instruction *IgnoredStore) {
400 // Get the location that may be stored across the loop. Since the access is
401 // strided positively through memory, we say that the modified location starts
402 // at the pointer and has infinite size.
403 uint64_t AccessSize = MemoryLocation::UnknownSize;
405 // If the loop iterates a fixed number of times, we can refine the access size
406 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
407 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
408 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
410 // TODO: For this to be really effective, we have to dive into the pointer
411 // operand in the store. Store to &A[i] of 100 will always return may alias
412 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
413 // which will then no-alias a store to &A[100].
414 MemoryLocation StoreLoc(Ptr, AccessSize);
416 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
418 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
419 if (&*I != IgnoredStore && (AA.getModRefInfo(I, StoreLoc) & Access))
425 /// getMemSetPatternValue - If a strided store of the specified value is safe to
426 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
427 /// be passed in. Otherwise, return null.
429 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
430 /// just replicate their input array and then pass on to memset_pattern16.
431 static Constant *getMemSetPatternValue(Value *V, const DataLayout &DL) {
432 // If the value isn't a constant, we can't promote it to being in a constant
433 // array. We could theoretically do a store to an alloca or something, but
434 // that doesn't seem worthwhile.
435 Constant *C = dyn_cast<Constant>(V);
439 // Only handle simple values that are a power of two bytes in size.
440 uint64_t Size = DL.getTypeSizeInBits(V->getType());
441 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
444 // Don't care enough about darwin/ppc to implement this.
445 if (DL.isBigEndian())
448 // Convert to size in bytes.
451 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
452 // if the top and bottom are the same (e.g. for vectors and large integers).
456 // If the constant is exactly 16 bytes, just use it.
460 // Otherwise, we'll use an array of the constants.
461 unsigned ArraySize = 16 / Size;
462 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
463 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
466 /// processLoopStridedStore - We see a strided store of some value. If we can
467 /// transform this into a memset or memset_pattern in the loop preheader, do so.
468 bool LoopIdiomRecognize::processLoopStridedStore(
469 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
470 Value *StoredVal, Instruction *TheStore, const SCEVAddRecExpr *Ev,
471 const SCEV *BECount) {
473 // If the stored value is a byte-wise value (like i32 -1), then it may be
474 // turned into a memset of i8 -1, assuming that all the consecutive bytes
475 // are stored. A store of i32 0x01020304 can never be turned into a memset,
476 // but it can be turned into memset_pattern if the target supports it.
477 Value *SplatValue = isBytewiseValue(StoredVal);
478 Constant *PatternValue = nullptr;
479 auto &DL = CurLoop->getHeader()->getModule()->getDataLayout();
480 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
482 // If we're allowed to form a memset, and the stored value would be acceptable
483 // for memset, use it.
484 if (SplatValue && TLI->has(LibFunc::memset) &&
485 // Verify that the stored value is loop invariant. If not, we can't
486 // promote the memset.
487 CurLoop->isLoopInvariant(SplatValue)) {
488 // Keep and use SplatValue.
489 PatternValue = nullptr;
490 } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) &&
491 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
492 // Don't create memset_pattern16s with address spaces.
493 // It looks like we can use PatternValue!
494 SplatValue = nullptr;
496 // Otherwise, this isn't an idiom we can transform. For example, we can't
497 // do anything with a 3-byte store.
501 // The trip count of the loop and the base pointer of the addrec SCEV is
502 // guaranteed to be loop invariant, which means that it should dominate the
503 // header. This allows us to insert code for it in the preheader.
504 BasicBlock *Preheader = CurLoop->getLoopPreheader();
505 IRBuilder<> Builder(Preheader->getTerminator());
506 SCEVExpander Expander(*SE, DL, "loop-idiom");
508 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
510 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
511 // this into a memset in the loop preheader now if we want. However, this
512 // would be unsafe to do if there is anything else in the loop that may read
513 // or write to the aliased location. Check for any overlap by generating the
514 // base pointer and checking the region.
515 Value *BasePtr = Expander.expandCodeFor(Ev->getStart(), DestInt8PtrTy,
516 Preheader->getTerminator());
518 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
521 // If we generated new code for the base pointer, clean up.
522 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
526 // Okay, everything looks good, insert the memset.
528 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
529 // pointer size if it isn't already.
530 Type *IntPtr = Builder.getIntPtrTy(DL, DestAS);
531 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
533 const SCEV *NumBytesS =
534 SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1), SCEV::FlagNUW);
535 if (StoreSize != 1) {
536 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
541 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
546 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
548 // Everything is emitted in default address space
549 Type *Int8PtrTy = DestInt8PtrTy;
551 Module *M = TheStore->getParent()->getParent()->getParent();
553 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
554 Int8PtrTy, Int8PtrTy, IntPtr, (void *)nullptr);
556 // Otherwise we should form a memset_pattern16. PatternValue is known to be
557 // an constant array of 16-bytes. Plop the value into a mergable global.
558 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
559 GlobalValue::PrivateLinkage,
560 PatternValue, ".memset_pattern");
561 GV->setUnnamedAddr(true); // Ok to merge these.
562 GV->setAlignment(16);
563 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
564 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
567 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
568 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
569 NewCall->setDebugLoc(TheStore->getDebugLoc());
571 // Okay, the memset has been formed. Zap the original store and anything that
573 deleteDeadInstruction(TheStore, TLI);
578 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
579 /// same-strided load.
580 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
581 StoreInst *SI, unsigned StoreSize, const SCEVAddRecExpr *StoreEv,
582 const SCEVAddRecExpr *LoadEv, const SCEV *BECount) {
583 // If we're not allowed to form memcpy, we fail.
584 if (!TLI->has(LibFunc::memcpy))
587 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
589 // The trip count of the loop and the base pointer of the addrec SCEV is
590 // guaranteed to be loop invariant, which means that it should dominate the
591 // header. This allows us to insert code for it in the preheader.
592 BasicBlock *Preheader = CurLoop->getLoopPreheader();
593 IRBuilder<> Builder(Preheader->getTerminator());
594 const DataLayout &DL = Preheader->getModule()->getDataLayout();
595 SCEVExpander Expander(*SE, DL, "loop-idiom");
597 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
598 // this into a memcpy in the loop preheader now if we want. However, this
599 // would be unsafe to do if there is anything else in the loop that may read
600 // or write the memory region we're storing to. This includes the load that
601 // feeds the stores. Check for an alias by generating the base address and
602 // checking everything.
603 Value *StoreBasePtr = Expander.expandCodeFor(
604 StoreEv->getStart(), Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
605 Preheader->getTerminator());
607 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
608 StoreSize, *AA, SI)) {
610 // If we generated new code for the base pointer, clean up.
611 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
615 // For a memcpy, we have to make sure that the input array is not being
616 // mutated by the loop.
617 Value *LoadBasePtr = Expander.expandCodeFor(
618 LoadEv->getStart(), Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
619 Preheader->getTerminator());
621 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
624 // If we generated new code for the base pointer, clean up.
625 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
626 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
630 // Okay, everything is safe, we can transform this!
632 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
633 // pointer size if it isn't already.
634 Type *IntPtrTy = Builder.getIntPtrTy(DL, SI->getPointerAddressSpace());
635 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
637 const SCEV *NumBytesS =
638 SE->getAddExpr(BECount, SE->getConstant(IntPtrTy, 1), SCEV::FlagNUW);
640 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
644 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
647 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
648 std::min(SI->getAlignment(), LI->getAlignment()));
649 NewCall->setDebugLoc(SI->getDebugLoc());
651 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
652 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
653 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
655 // Okay, the memset has been formed. Zap the original store and anything that
657 deleteDeadInstruction(SI, TLI);
662 bool LoopIdiomRecognize::runOnNoncountableLoop() {
663 if (recognizePopcount())
669 /// Check if the given conditional branch is based on the comparison between
670 /// a variable and zero, and if the variable is non-zero, the control yields to
671 /// the loop entry. If the branch matches the behavior, the variable involved
672 /// in the comparion is returned. This function will be called to see if the
673 /// precondition and postcondition of the loop are in desirable form.
674 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
675 if (!BI || !BI->isConditional())
678 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
682 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
683 if (!CmpZero || !CmpZero->isZero())
686 ICmpInst::Predicate Pred = Cond->getPredicate();
687 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
688 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
689 return Cond->getOperand(0);
694 /// Return true iff the idiom is detected in the loop.
697 /// 1) \p CntInst is set to the instruction counting the population bit.
698 /// 2) \p CntPhi is set to the corresponding phi node.
699 /// 3) \p Var is set to the value whose population bits are being counted.
701 /// The core idiom we are trying to detect is:
704 /// goto loop-exit // the precondition of the loop
707 /// x1 = phi (x0, x2);
708 /// cnt1 = phi(cnt0, cnt2);
712 /// x2 = x1 & (x1 - 1);
718 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
719 Instruction *&CntInst, PHINode *&CntPhi,
721 // step 1: Check to see if the look-back branch match this pattern:
722 // "if (a!=0) goto loop-entry".
723 BasicBlock *LoopEntry;
724 Instruction *DefX2, *CountInst;
725 Value *VarX1, *VarX0;
726 PHINode *PhiX, *CountPhi;
728 DefX2 = CountInst = nullptr;
729 VarX1 = VarX0 = nullptr;
730 PhiX = CountPhi = nullptr;
731 LoopEntry = *(CurLoop->block_begin());
733 // step 1: Check if the loop-back branch is in desirable form.
735 if (Value *T = matchCondition(
736 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
737 DefX2 = dyn_cast<Instruction>(T);
742 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
744 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
747 BinaryOperator *SubOneOp;
749 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
750 VarX1 = DefX2->getOperand(1);
752 VarX1 = DefX2->getOperand(0);
753 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
758 Instruction *SubInst = cast<Instruction>(SubOneOp);
759 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
761 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
762 (SubInst->getOpcode() == Instruction::Add &&
763 Dec->isAllOnesValue()))) {
768 // step 3: Check the recurrence of variable X
770 PhiX = dyn_cast<PHINode>(VarX1);
772 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
777 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
780 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(),
781 IterE = LoopEntry->end();
782 Iter != IterE; Iter++) {
783 Instruction *Inst = Iter;
784 if (Inst->getOpcode() != Instruction::Add)
787 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
788 if (!Inc || !Inc->isOne())
791 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
792 if (!Phi || Phi->getParent() != LoopEntry)
795 // Check if the result of the instruction is live of the loop.
796 bool LiveOutLoop = false;
797 for (User *U : Inst->users()) {
798 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
815 // step 5: check if the precondition is in this form:
816 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
818 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
819 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
820 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
831 /// Recognizes a population count idiom in a non-countable loop.
833 /// If detected, transforms the relevant code to issue the popcount intrinsic
834 /// function call, and returns true; otherwise, returns false.
835 bool LoopIdiomRecognize::recognizePopcount() {
836 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
839 // Counting population are usually conducted by few arithmetic instructions.
840 // Such instructions can be easily "absorbed" by vacant slots in a
841 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
842 // in a compact loop.
844 // Give up if the loop has multiple blocks or multiple backedges.
845 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
848 BasicBlock *LoopBody = *(CurLoop->block_begin());
849 if (LoopBody->size() >= 20) {
850 // The loop is too big, bail out.
854 // It should have a preheader containing nothing but an unconditional branch.
855 BasicBlock *PH = CurLoop->getLoopPreheader();
858 if (&PH->front() != PH->getTerminator())
860 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
861 if (!EntryBI || EntryBI->isConditional())
864 // It should have a precondition block where the generated popcount instrinsic
865 // function can be inserted.
866 auto *PreCondBB = PH->getSinglePredecessor();
869 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
870 if (!PreCondBI || PreCondBI->isUnconditional())
873 Instruction *CntInst;
876 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
879 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
883 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
885 Value *Ops[] = {Val};
886 Type *Tys[] = {Val->getType()};
888 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
889 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
890 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
896 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
897 Instruction *CntInst,
898 PHINode *CntPhi, Value *Var) {
899 BasicBlock *PreHead = CurLoop->getLoopPreheader();
900 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
901 const DebugLoc DL = CntInst->getDebugLoc();
903 // Assuming before transformation, the loop is following:
904 // if (x) // the precondition
905 // do { cnt++; x &= x - 1; } while(x);
907 // Step 1: Insert the ctpop instruction at the end of the precondition block
908 IRBuilder<> Builder(PreCondBr);
909 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
911 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
912 NewCount = PopCntZext =
913 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
915 if (NewCount != PopCnt)
916 (cast<Instruction>(NewCount))->setDebugLoc(DL);
918 // TripCnt is exactly the number of iterations the loop has
921 // If the population counter's initial value is not zero, insert Add Inst.
922 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
923 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
924 if (!InitConst || !InitConst->isZero()) {
925 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
926 (cast<Instruction>(NewCount))->setDebugLoc(DL);
930 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
931 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
932 // function would be partial dead code, and downstream passes will drag
933 // it back from the precondition block to the preheader.
935 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
937 Value *Opnd0 = PopCntZext;
938 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
939 if (PreCond->getOperand(0) != Var)
940 std::swap(Opnd0, Opnd1);
942 ICmpInst *NewPreCond = cast<ICmpInst>(
943 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
944 PreCondBr->setCondition(NewPreCond);
946 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
949 // Step 3: Note that the population count is exactly the trip count of the
950 // loop in question, which enable us to to convert the loop from noncountable
951 // loop into a countable one. The benefit is twofold:
953 // - If the loop only counts population, the entire loop becomes dead after
954 // the transformation. It is a lot easier to prove a countable loop dead
955 // than to prove a noncountable one. (In some C dialects, an infinite loop
956 // isn't dead even if it computes nothing useful. In general, DCE needs
957 // to prove a noncountable loop finite before safely delete it.)
959 // - If the loop also performs something else, it remains alive.
960 // Since it is transformed to countable form, it can be aggressively
961 // optimized by some optimizations which are in general not applicable
962 // to a noncountable loop.
964 // After this step, this loop (conceptually) would look like following:
965 // newcnt = __builtin_ctpop(x);
968 // do { cnt++; x &= x-1; t--) } while (t > 0);
969 BasicBlock *Body = *(CurLoop->block_begin());
971 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
972 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
973 Type *Ty = TripCnt->getType();
975 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin());
977 Builder.SetInsertPoint(LbCond);
978 Instruction *TcDec = cast<Instruction>(
979 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
980 "tcdec", false, true));
982 TcPhi->addIncoming(TripCnt, PreHead);
983 TcPhi->addIncoming(TcDec, Body);
985 CmpInst::Predicate Pred =
986 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
987 LbCond->setPredicate(Pred);
988 LbCond->setOperand(0, TcDec);
989 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
992 // Step 4: All the references to the original population counter outside
993 // the loop are replaced with the NewCount -- the value returned from
994 // __builtin_ctpop().
995 CntInst->replaceUsesOutsideBlock(NewCount, Body);
997 // step 5: Forget the "non-computable" trip-count SCEV associated with the
998 // loop. The loop would otherwise not be deleted even if it becomes empty.
999 SE->forgetLoop(CurLoop);