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/LoopPass.h"
48 #include "llvm/Analysis/ScalarEvolutionExpander.h"
49 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
50 #include "llvm/Analysis/TargetLibraryInfo.h"
51 #include "llvm/Analysis/TargetTransformInfo.h"
52 #include "llvm/Analysis/ValueTracking.h"
53 #include "llvm/IR/DataLayout.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/IRBuilder.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Module.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/raw_ostream.h"
60 #include "llvm/Transforms/Utils/Local.h"
63 #define DEBUG_TYPE "loop-idiom"
65 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
66 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
70 class LoopIdiomRecognize : public LoopPass {
75 TargetLibraryInfo *TLI;
76 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<AliasAnalysis>();
97 AU.addPreserved<AliasAnalysis>();
98 AU.addRequired<ScalarEvolution>();
99 AU.addPreserved<ScalarEvolution>();
100 AU.addPreserved<DominatorTreeWrapperPass>();
101 AU.addRequired<DominatorTreeWrapperPass>();
102 AU.addRequired<TargetLibraryInfoWrapperPass>();
103 AU.addRequired<TargetTransformInfoWrapperPass>();
107 /// \name Countable Loop Idiom Handling
110 bool runOnCountableLoop();
111 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
112 SmallVectorImpl<BasicBlock *> &ExitBlocks);
114 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
115 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
117 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
118 unsigned StoreAlignment, Value *SplatValue,
119 Instruction *TheStore, const SCEVAddRecExpr *Ev,
120 const SCEV *BECount);
121 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
122 const SCEVAddRecExpr *StoreEv,
123 const SCEVAddRecExpr *LoadEv,
124 const SCEV *BECount);
127 /// \name Noncountable Loop Idiom Handling
130 bool runOnNoncountableLoop();
132 bool recognizePopcount();
133 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
134 PHINode *CntPhi, Value *Var);
139 } // End anonymous namespace.
141 char LoopIdiomRecognize::ID = 0;
142 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
144 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
145 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
146 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
147 INITIALIZE_PASS_DEPENDENCY(LCSSA)
148 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
149 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
150 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
151 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
152 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
155 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
157 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
158 /// and zero out all the operands of this instruction. If any of them become
159 /// dead, delete them and the computation tree that feeds them.
161 static void deleteDeadInstruction(Instruction *I,
162 const TargetLibraryInfo *TLI) {
163 SmallVector<Value *, 16> Operands(I->value_op_begin(), I->value_op_end());
164 I->replaceAllUsesWith(UndefValue::get(I->getType()));
165 I->eraseFromParent();
166 for (Value *Op : Operands)
167 RecursivelyDeleteTriviallyDeadInstructions(Op, TLI);
170 //===----------------------------------------------------------------------===//
172 // Implementation of LoopIdiomRecognize
174 //===----------------------------------------------------------------------===//
176 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
177 if (skipOptnoneFunction(L))
181 // If the loop could not be converted to canonical form, it must have an
182 // indirectbr in it, just give up.
183 if (!L->getLoopPreheader())
186 // Disable loop idiom recognition if the function's name is a common idiom.
187 StringRef Name = L->getHeader()->getParent()->getName();
188 if (Name == "memset" || Name == "memcpy")
191 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
192 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
193 SE = &getAnalysis<ScalarEvolution>();
194 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
195 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
196 *CurLoop->getHeader()->getParent());
198 if (SE->hasLoopInvariantBackedgeTakenCount(L))
199 return runOnCountableLoop();
201 return runOnNoncountableLoop();
204 bool LoopIdiomRecognize::runOnCountableLoop() {
205 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
206 assert(!isa<SCEVCouldNotCompute>(BECount) &&
207 "runOnCountableLoop() called on a loop without a predictable"
208 "backedge-taken count");
210 // If this loop executes exactly one time, then it should be peeled, not
211 // optimized by this pass.
212 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
213 if (BECst->getValue()->getValue() == 0)
216 SmallVector<BasicBlock *, 8> ExitBlocks;
217 CurLoop->getUniqueExitBlocks(ExitBlocks);
219 DEBUG(dbgs() << "loop-idiom Scanning: F["
220 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
221 << CurLoop->getHeader()->getName() << "\n");
223 bool MadeChange = false;
224 // Scan all the blocks in the loop that are not in subloops.
225 for (auto *BB : CurLoop->getBlocks()) {
226 // Ignore blocks in subloops.
227 if (LI->getLoopFor(BB) != CurLoop)
230 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
235 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
236 /// with the specified backedge count. This block is known to be in the current
237 /// loop and not in any subloops.
238 bool LoopIdiomRecognize::runOnLoopBlock(
239 BasicBlock *BB, const SCEV *BECount,
240 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
241 // We can only promote stores in this block if they are unconditionally
242 // executed in the loop. For a block to be unconditionally executed, it has
243 // to dominate all the exit blocks of the loop. Verify this now.
244 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
245 if (!DT->dominates(BB, ExitBlocks[i]))
248 bool MadeChange = false;
249 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
250 Instruction *Inst = I++;
251 // Look for store instructions, which may be optimized to memset/memcpy.
252 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
254 if (!processLoopStore(SI, BECount))
258 // If processing the store invalidated our iterator, start over from the
265 // Look for memset instructions, which may be optimized to a larger memset.
266 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
268 if (!processLoopMemSet(MSI, BECount))
272 // If processing the memset invalidated our iterator, start over from the
283 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
284 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
288 Value *StoredVal = SI->getValueOperand();
289 Value *StorePtr = SI->getPointerOperand();
291 // Reject stores that are so large that they overflow an unsigned.
292 auto &DL = CurLoop->getHeader()->getModule()->getDataLayout();
293 uint64_t SizeInBits = DL.getTypeSizeInBits(StoredVal->getType());
294 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
297 // See if the pointer expression is an AddRec like {base,+,1} on the current
298 // loop, which indicates a strided store. If we have something else, it's a
299 // random store we can't handle.
300 const SCEVAddRecExpr *StoreEv =
301 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
302 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
305 // Check to see if the stride matches the size of the store. If so, then we
306 // know that every byte is touched in the loop.
307 unsigned StoreSize = (unsigned)SizeInBits >> 3;
308 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
310 if (!Stride || StoreSize != Stride->getValue()->getValue()) {
311 // TODO: Could also handle negative stride here someday, that will require
312 // the validity check in mayLoopAccessLocation to be updated though.
313 // Enable this to print exact negative strides.
314 if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
315 dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
316 dbgs() << "BB: " << *SI->getParent();
322 // See if we can optimize just this store in isolation.
323 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
324 StoredVal, SI, StoreEv, BECount))
327 // If the stored value is a strided load in the same loop with the same stride
328 // this this may be transformable into a memcpy. This kicks in for stuff like
329 // for (i) A[i] = B[i];
330 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
331 const SCEVAddRecExpr *LoadEv =
332 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
333 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
334 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
335 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
338 // errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
343 /// processLoopMemSet - See if this memset can be promoted to a large memset.
344 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
345 const SCEV *BECount) {
346 // We can only handle non-volatile memsets with a constant size.
347 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
350 // If we're not allowed to hack on memset, we fail.
351 if (!TLI->has(LibFunc::memset))
354 Value *Pointer = MSI->getDest();
356 // See if the pointer expression is an AddRec like {base,+,1} on the current
357 // loop, which indicates a strided store. If we have something else, it's a
358 // random store we can't handle.
359 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
360 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
363 // Reject memsets that are so large that they overflow an unsigned.
364 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
365 if ((SizeInBytes >> 32) != 0)
368 // Check to see if the stride matches the size of the memset. If so, then we
369 // know that every byte is touched in the loop.
370 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
372 // TODO: Could also handle negative stride here someday, that will require the
373 // validity check in mayLoopAccessLocation to be updated though.
374 if (!Stride || MSI->getLength() != Stride->getValue())
377 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
378 MSI->getAlignment(), MSI->getValue(), MSI, Ev,
382 /// mayLoopAccessLocation - Return true if the specified loop might access the
383 /// specified pointer location, which is a loop-strided access. The 'Access'
384 /// argument specifies what the verboten forms of access are (read or write).
385 static bool mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
386 const SCEV *BECount, unsigned StoreSize,
388 Instruction *IgnoredStore) {
389 // Get the location that may be stored across the loop. Since the access is
390 // strided positively through memory, we say that the modified location starts
391 // at the pointer and has infinite size.
392 uint64_t AccessSize = MemoryLocation::UnknownSize;
394 // If the loop iterates a fixed number of times, we can refine the access size
395 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
396 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
397 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
399 // TODO: For this to be really effective, we have to dive into the pointer
400 // operand in the store. Store to &A[i] of 100 will always return may alias
401 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
402 // which will then no-alias a store to &A[100].
403 MemoryLocation StoreLoc(Ptr, AccessSize);
405 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
407 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
408 if (&*I != IgnoredStore && (AA.getModRefInfo(I, StoreLoc) & Access))
414 /// getMemSetPatternValue - If a strided store of the specified value is safe to
415 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
416 /// be passed in. Otherwise, return null.
418 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
419 /// just replicate their input array and then pass on to memset_pattern16.
420 static Constant *getMemSetPatternValue(Value *V, const DataLayout &DL) {
421 // If the value isn't a constant, we can't promote it to being in a constant
422 // array. We could theoretically do a store to an alloca or something, but
423 // that doesn't seem worthwhile.
424 Constant *C = dyn_cast<Constant>(V);
428 // Only handle simple values that are a power of two bytes in size.
429 uint64_t Size = DL.getTypeSizeInBits(V->getType());
430 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
433 // Don't care enough about darwin/ppc to implement this.
434 if (DL.isBigEndian())
437 // Convert to size in bytes.
440 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
441 // if the top and bottom are the same (e.g. for vectors and large integers).
445 // If the constant is exactly 16 bytes, just use it.
449 // Otherwise, we'll use an array of the constants.
450 unsigned ArraySize = 16 / Size;
451 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
452 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
455 /// processLoopStridedStore - We see a strided store of some value. If we can
456 /// transform this into a memset or memset_pattern in the loop preheader, do so.
457 bool LoopIdiomRecognize::processLoopStridedStore(
458 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
459 Value *StoredVal, Instruction *TheStore, const SCEVAddRecExpr *Ev,
460 const SCEV *BECount) {
462 // If the stored value is a byte-wise value (like i32 -1), then it may be
463 // turned into a memset of i8 -1, assuming that all the consecutive bytes
464 // are stored. A store of i32 0x01020304 can never be turned into a memset,
465 // but it can be turned into memset_pattern if the target supports it.
466 Value *SplatValue = isBytewiseValue(StoredVal);
467 Constant *PatternValue = nullptr;
468 auto &DL = CurLoop->getHeader()->getModule()->getDataLayout();
469 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
471 // If we're allowed to form a memset, and the stored value would be acceptable
472 // for memset, use it.
473 if (SplatValue && TLI->has(LibFunc::memset) &&
474 // Verify that the stored value is loop invariant. If not, we can't
475 // promote the memset.
476 CurLoop->isLoopInvariant(SplatValue)) {
477 // Keep and use SplatValue.
478 PatternValue = nullptr;
479 } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) &&
480 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
481 // Don't create memset_pattern16s with address spaces.
482 // It looks like we can use PatternValue!
483 SplatValue = nullptr;
485 // Otherwise, this isn't an idiom we can transform. For example, we can't
486 // do anything with a 3-byte store.
490 // The trip count of the loop and the base pointer of the addrec SCEV is
491 // guaranteed to be loop invariant, which means that it should dominate the
492 // header. This allows us to insert code for it in the preheader.
493 BasicBlock *Preheader = CurLoop->getLoopPreheader();
494 IRBuilder<> Builder(Preheader->getTerminator());
495 SCEVExpander Expander(*SE, DL, "loop-idiom");
497 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
499 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
500 // this into a memset in the loop preheader now if we want. However, this
501 // would be unsafe to do if there is anything else in the loop that may read
502 // or write to the aliased location. Check for any overlap by generating the
503 // base pointer and checking the region.
504 Value *BasePtr = Expander.expandCodeFor(Ev->getStart(), DestInt8PtrTy,
505 Preheader->getTerminator());
507 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
508 getAnalysis<AliasAnalysis>(), TheStore)) {
510 // If we generated new code for the base pointer, clean up.
511 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
515 // Okay, everything looks good, insert the memset.
517 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
518 // pointer size if it isn't already.
519 Type *IntPtr = Builder.getIntPtrTy(DL, DestAS);
520 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
522 const SCEV *NumBytesS =
523 SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1), SCEV::FlagNUW);
524 if (StoreSize != 1) {
525 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
530 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
535 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
537 // Everything is emitted in default address space
538 Type *Int8PtrTy = DestInt8PtrTy;
540 Module *M = TheStore->getParent()->getParent()->getParent();
542 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
543 Int8PtrTy, Int8PtrTy, IntPtr, (void *)nullptr);
545 // Otherwise we should form a memset_pattern16. PatternValue is known to be
546 // an constant array of 16-bytes. Plop the value into a mergable global.
547 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
548 GlobalValue::PrivateLinkage,
549 PatternValue, ".memset_pattern");
550 GV->setUnnamedAddr(true); // Ok to merge these.
551 GV->setAlignment(16);
552 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
553 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
556 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
557 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
558 NewCall->setDebugLoc(TheStore->getDebugLoc());
560 // Okay, the memset has been formed. Zap the original store and anything that
562 deleteDeadInstruction(TheStore, TLI);
567 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
568 /// same-strided load.
569 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
570 StoreInst *SI, unsigned StoreSize, const SCEVAddRecExpr *StoreEv,
571 const SCEVAddRecExpr *LoadEv, const SCEV *BECount) {
572 // If we're not allowed to form memcpy, we fail.
573 if (!TLI->has(LibFunc::memcpy))
576 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
578 // The trip count of the loop and the base pointer of the addrec SCEV is
579 // guaranteed to be loop invariant, which means that it should dominate the
580 // header. This allows us to insert code for it in the preheader.
581 BasicBlock *Preheader = CurLoop->getLoopPreheader();
582 IRBuilder<> Builder(Preheader->getTerminator());
583 const DataLayout &DL = Preheader->getModule()->getDataLayout();
584 SCEVExpander Expander(*SE, DL, "loop-idiom");
586 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
587 // this into a memcpy in the loop preheader now if we want. However, this
588 // would be unsafe to do if there is anything else in the loop that may read
589 // or write the memory region we're storing to. This includes the load that
590 // feeds the stores. Check for an alias by generating the base address and
591 // checking everything.
592 Value *StoreBasePtr = Expander.expandCodeFor(
593 StoreEv->getStart(), Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
594 Preheader->getTerminator());
596 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
597 StoreSize, getAnalysis<AliasAnalysis>(), SI)) {
599 // If we generated new code for the base pointer, clean up.
600 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
604 // For a memcpy, we have to make sure that the input array is not being
605 // mutated by the loop.
606 Value *LoadBasePtr = Expander.expandCodeFor(
607 LoadEv->getStart(), Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
608 Preheader->getTerminator());
610 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
611 getAnalysis<AliasAnalysis>(), SI)) {
613 // If we generated new code for the base pointer, clean up.
614 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
615 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
619 // Okay, everything is safe, we can transform this!
621 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
622 // pointer size if it isn't already.
623 Type *IntPtrTy = Builder.getIntPtrTy(DL, SI->getPointerAddressSpace());
624 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
626 const SCEV *NumBytesS =
627 SE->getAddExpr(BECount, SE->getConstant(IntPtrTy, 1), SCEV::FlagNUW);
629 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
633 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
636 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
637 std::min(SI->getAlignment(), LI->getAlignment()));
638 NewCall->setDebugLoc(SI->getDebugLoc());
640 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
641 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
642 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
644 // Okay, the memset has been formed. Zap the original store and anything that
646 deleteDeadInstruction(SI, TLI);
651 bool LoopIdiomRecognize::runOnNoncountableLoop() {
652 if (recognizePopcount())
658 /// Check if the given conditional branch is based on the comparison between
659 /// a variable and zero, and if the variable is non-zero, the control yields to
660 /// the loop entry. If the branch matches the behavior, the variable involved
661 /// in the comparion is returned. This function will be called to see if the
662 /// precondition and postcondition of the loop are in desirable form.
663 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
664 if (!BI || !BI->isConditional())
667 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
671 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
672 if (!CmpZero || !CmpZero->isZero())
675 ICmpInst::Predicate Pred = Cond->getPredicate();
676 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
677 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
678 return Cond->getOperand(0);
683 /// Return true iff the idiom is detected in the loop.
686 /// 1) \p CntInst is set to the instruction counting the population bit.
687 /// 2) \p CntPhi is set to the corresponding phi node.
688 /// 3) \p Var is set to the value whose population bits are being counted.
690 /// The core idiom we are trying to detect is:
693 /// goto loop-exit // the precondition of the loop
696 /// x1 = phi (x0, x2);
697 /// cnt1 = phi(cnt0, cnt2);
701 /// x2 = x1 & (x1 - 1);
707 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
708 Instruction *&CntInst, PHINode *&CntPhi,
710 // step 1: Check to see if the look-back branch match this pattern:
711 // "if (a!=0) goto loop-entry".
712 BasicBlock *LoopEntry;
713 Instruction *DefX2, *CountInst;
714 Value *VarX1, *VarX0;
715 PHINode *PhiX, *CountPhi;
717 DefX2 = CountInst = nullptr;
718 VarX1 = VarX0 = nullptr;
719 PhiX = CountPhi = nullptr;
720 LoopEntry = *(CurLoop->block_begin());
722 // step 1: Check if the loop-back branch is in desirable form.
724 if (Value *T = matchCondition(
725 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
726 DefX2 = dyn_cast<Instruction>(T);
731 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
733 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
736 BinaryOperator *SubOneOp;
738 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
739 VarX1 = DefX2->getOperand(1);
741 VarX1 = DefX2->getOperand(0);
742 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
747 Instruction *SubInst = cast<Instruction>(SubOneOp);
748 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
750 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
751 (SubInst->getOpcode() == Instruction::Add &&
752 Dec->isAllOnesValue()))) {
757 // step 3: Check the recurrence of variable X
759 PhiX = dyn_cast<PHINode>(VarX1);
761 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
766 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
769 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(),
770 IterE = LoopEntry->end();
771 Iter != IterE; Iter++) {
772 Instruction *Inst = Iter;
773 if (Inst->getOpcode() != Instruction::Add)
776 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
777 if (!Inc || !Inc->isOne())
780 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
781 if (!Phi || Phi->getParent() != LoopEntry)
784 // Check if the result of the instruction is live of the loop.
785 bool LiveOutLoop = false;
786 for (User *U : Inst->users()) {
787 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
804 // step 5: check if the precondition is in this form:
805 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
807 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
808 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
809 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
820 /// Recognizes a population count idiom in a non-countable loop.
822 /// If detected, transforms the relevant code to issue the popcount intrinsic
823 /// function call, and returns true; otherwise, returns false.
824 bool LoopIdiomRecognize::recognizePopcount() {
825 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
828 // Counting population are usually conducted by few arithmetic instructions.
829 // Such instructions can be easilly "absorbed" by vacant slots in a
830 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
831 // in a compact loop.
833 // Give up if the loop has multiple blocks or multiple backedges.
834 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
837 BasicBlock *LoopBody = *(CurLoop->block_begin());
838 if (LoopBody->size() >= 20) {
839 // The loop is too big, bail out.
843 // It should have a preheader containing nothing but an unconditional branch.
844 BasicBlock *PH = CurLoop->getLoopPreheader();
847 if (&PH->front() != PH->getTerminator())
849 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
850 if (!EntryBI || EntryBI->isConditional())
853 // It should have a precondition block where the generated popcount instrinsic
854 // function can be inserted.
855 auto *PreCondBB = PH->getSinglePredecessor();
858 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
859 if (!PreCondBI || PreCondBI->isUnconditional())
862 Instruction *CntInst;
865 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
868 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
872 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
874 Value *Ops[] = {Val};
875 Type *Tys[] = {Val->getType()};
877 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
878 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
879 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
885 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
886 Instruction *CntInst,
887 PHINode *CntPhi, Value *Var) {
888 BasicBlock *PreHead = CurLoop->getLoopPreheader();
889 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
890 const DebugLoc DL = CntInst->getDebugLoc();
892 // Assuming before transformation, the loop is following:
893 // if (x) // the precondition
894 // do { cnt++; x &= x - 1; } while(x);
896 // Step 1: Insert the ctpop instruction at the end of the precondition block
897 IRBuilder<> Builder(PreCondBr);
898 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
900 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
901 NewCount = PopCntZext =
902 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
904 if (NewCount != PopCnt)
905 (cast<Instruction>(NewCount))->setDebugLoc(DL);
907 // TripCnt is exactly the number of iterations the loop has
910 // If the population counter's initial value is not zero, insert Add Inst.
911 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
912 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
913 if (!InitConst || !InitConst->isZero()) {
914 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
915 (cast<Instruction>(NewCount))->setDebugLoc(DL);
919 // Step 2: Replace the precondition from "if(x == 0) goto loop-exit" to
920 // "if(NewCount == 0) loop-exit". Withtout this change, the intrinsic
921 // function would be partial dead code, and downstream passes will drag
922 // it back from the precondition block to the preheader.
924 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
926 Value *Opnd0 = PopCntZext;
927 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
928 if (PreCond->getOperand(0) != Var)
929 std::swap(Opnd0, Opnd1);
931 ICmpInst *NewPreCond = cast<ICmpInst>(
932 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
933 PreCondBr->setCondition(NewPreCond);
935 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
938 // Step 3: Note that the population count is exactly the trip count of the
939 // loop in question, which enble us to to convert the loop from noncountable
940 // loop into a countable one. The benefit is twofold:
942 // - If the loop only counts population, the entire loop become dead after
943 // the transformation. It is lots easier to prove a countable loop dead
944 // than to prove a noncountable one. (In some C dialects, a infite loop
945 // isn't dead even if it computes nothing useful. In general, DCE needs
946 // to prove a noncountable loop finite before safely delete it.)
948 // - If the loop also performs something else, it remains alive.
949 // Since it is transformed to countable form, it can be aggressively
950 // optimized by some optimizations which are in general not applicable
951 // to a noncountable loop.
953 // After this step, this loop (conceptually) would look like following:
954 // newcnt = __builtin_ctpop(x);
957 // do { cnt++; x &= x-1; t--) } while (t > 0);
958 BasicBlock *Body = *(CurLoop->block_begin());
960 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
961 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
962 Type *Ty = TripCnt->getType();
964 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin());
966 Builder.SetInsertPoint(LbCond);
967 Value *Opnd1 = cast<Value>(TcPhi);
968 Value *Opnd2 = cast<Value>(ConstantInt::get(Ty, 1));
969 Instruction *TcDec = cast<Instruction>(
970 Builder.CreateSub(Opnd1, Opnd2, "tcdec", false, true));
972 TcPhi->addIncoming(TripCnt, PreHead);
973 TcPhi->addIncoming(TcDec, Body);
975 CmpInst::Predicate Pred =
976 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
977 LbCond->setPredicate(Pred);
978 LbCond->setOperand(0, TcDec);
979 LbCond->setOperand(1, cast<Value>(ConstantInt::get(Ty, 0)));
982 // Step 4: All the references to the original population counter outside
983 // the loop are replaced with the NewCount -- the value returned from
984 // __builtin_ctpop().
985 CntInst->replaceUsesOutsideBlock(NewCount, Body);
987 // step 5: Forget the "non-computable" trip-count SCEV associated with the
988 // loop. The loop would otherwise not be deleted even if it becomes empty.
989 SE->forgetLoop(CurLoop);