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, 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 #define DEBUG_TYPE "loop-idiom"
45 #include "llvm/Transforms/Scalar.h"
46 #include "llvm/IRBuilder.h"
47 #include "llvm/IntrinsicInst.h"
48 #include "llvm/Module.h"
49 #include "llvm/ADT/Statistic.h"
50 #include "llvm/Analysis/AliasAnalysis.h"
51 #include "llvm/Analysis/DependenceAnalysis.h"
52 #include "llvm/Analysis/LoopPass.h"
53 #include "llvm/Analysis/ScalarEvolutionExpander.h"
54 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
55 #include "llvm/Analysis/ValueTracking.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/DataLayout.h"
59 #include "llvm/Target/TargetLibraryInfo.h"
60 #include "llvm/Transforms/Utils/Local.h"
63 STATISTIC(NumMemSet, "Number of memsets formed from loop stores");
64 STATISTIC(NumMemCpy, "Number of memcpys formed from loop load+stores");
65 STATISTIC(NumMemMove, "Number of memmoves formed from loop load+stores");
68 class LoopIdiomRecognize : public LoopPass {
73 TargetLibraryInfo *TLI;
76 explicit LoopIdiomRecognize() : LoopPass(ID) {
77 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
80 bool runOnLoop(Loop *L, LPPassManager &LPM);
81 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
82 SmallVectorImpl<BasicBlock*> &ExitBlocks);
84 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
85 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
87 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
88 unsigned StoreAlignment,
89 Value *SplatValue, Instruction *TheStore,
90 const SCEVAddRecExpr *Ev,
92 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
93 const SCEVAddRecExpr *StoreEv,
94 const SCEVAddRecExpr *LoadEv,
97 /// This transformation requires natural loop information & requires that
98 /// loop preheaders be inserted into the CFG.
100 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
101 AU.addRequired<LoopInfo>();
102 AU.addPreserved<LoopInfo>();
103 AU.addRequiredID(LoopSimplifyID);
104 AU.addPreservedID(LoopSimplifyID);
105 AU.addRequiredID(LCSSAID);
106 AU.addPreservedID(LCSSAID);
107 AU.addRequired<AliasAnalysis>();
108 AU.addPreserved<AliasAnalysis>();
109 AU.addRequired<ScalarEvolution>();
110 AU.addPreserved<ScalarEvolution>();
111 AU.addRequired<DependenceAnalysis>();
112 AU.addPreserved<DependenceAnalysis>();
113 AU.addPreserved<DominatorTree>();
114 AU.addRequired<DominatorTree>();
115 AU.addRequired<TargetLibraryInfo>();
120 char LoopIdiomRecognize::ID = 0;
121 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
123 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
124 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
125 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
126 INITIALIZE_PASS_DEPENDENCY(LCSSA)
127 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
128 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
129 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysis)
130 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
131 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
134 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
136 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
137 /// and zero out all the operands of this instruction. If any of them become
138 /// dead, delete them and the computation tree that feeds them.
140 static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE,
141 const TargetLibraryInfo *TLI) {
142 SmallVector<Instruction*, 32> NowDeadInsts;
144 NowDeadInsts.push_back(I);
146 // Before we touch this instruction, remove it from SE!
148 Instruction *DeadInst = NowDeadInsts.pop_back_val();
150 // This instruction is dead, zap it, in stages. Start by removing it from
152 SE.forgetValue(DeadInst);
154 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
155 Value *Op = DeadInst->getOperand(op);
156 DeadInst->setOperand(op, 0);
158 // If this operand just became dead, add it to the NowDeadInsts list.
159 if (!Op->use_empty()) continue;
161 if (Instruction *OpI = dyn_cast<Instruction>(Op))
162 if (isInstructionTriviallyDead(OpI, TLI))
163 NowDeadInsts.push_back(OpI);
166 DeadInst->eraseFromParent();
168 } while (!NowDeadInsts.empty());
171 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
174 // If the loop could not be converted to canonical form, it must have an
175 // indirectbr in it, just give up.
176 if (!L->getLoopPreheader())
179 // Disable loop idiom recognition if the function's name is a common idiom.
180 StringRef Name = L->getHeader()->getParent()->getName();
181 if (Name == "memset" || Name == "memcpy" || Name == "memmove")
184 // The trip count of the loop must be analyzable.
185 SE = &getAnalysis<ScalarEvolution>();
186 if (!SE->hasLoopInvariantBackedgeTakenCount(L))
188 const SCEV *BECount = SE->getBackedgeTakenCount(L);
189 if (isa<SCEVCouldNotCompute>(BECount)) return false;
191 // If this loop executes exactly one time, then it should be peeled, not
192 // optimized by this pass.
193 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
194 if (BECst->getValue()->getValue() == 0)
197 // We require target data for now.
198 TD = getAnalysisIfAvailable<DataLayout>();
199 if (TD == 0) return false;
201 DT = &getAnalysis<DominatorTree>();
202 LoopInfo &LI = getAnalysis<LoopInfo>();
203 TLI = &getAnalysis<TargetLibraryInfo>();
205 SmallVector<BasicBlock*, 8> ExitBlocks;
206 CurLoop->getUniqueExitBlocks(ExitBlocks);
208 DEBUG(dbgs() << "loop-idiom Scanning: F["
209 << L->getHeader()->getParent()->getName()
210 << "] Loop %" << L->getHeader()->getName() << "\n");
212 bool MadeChange = false;
213 // Scan all the blocks in the loop that are not in subloops.
214 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
216 // Ignore blocks in subloops.
217 if (LI.getLoopFor(*BI) != CurLoop)
220 MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks);
225 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
226 /// with the specified backedge count. This block is known to be in the current
227 /// loop and not in any subloops.
228 bool LoopIdiomRecognize::runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
229 SmallVectorImpl<BasicBlock*> &ExitBlocks) {
230 // We can only promote stores in this block if they are unconditionally
231 // executed in the loop. For a block to be unconditionally executed, it has
232 // to dominate all the exit blocks of the loop. Verify this now.
233 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
234 if (!DT->dominates(BB, ExitBlocks[i]))
237 bool MadeChange = false;
238 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
239 Instruction *Inst = I++;
240 // Look for store instructions, which may be optimized to memset/memcpy.
241 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
243 if (!processLoopStore(SI, BECount)) continue;
246 // If processing the store invalidated our iterator, start over from the
253 // Look for memset instructions, which may be optimized to a larger memset.
254 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
256 if (!processLoopMemSet(MSI, BECount)) continue;
259 // If processing the memset invalidated our iterator, start over from the
271 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
272 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
273 if (!SI->isSimple()) return false;
275 Value *StoredVal = SI->getValueOperand();
276 Value *StorePtr = SI->getPointerOperand();
278 // Reject stores that are so large that they overflow an unsigned.
279 uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType());
280 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
283 // See if the pointer expression is an AddRec like {base,+,1} on the current
284 // loop, which indicates a strided store. If we have something else, it's a
285 // random store we can't handle.
286 const SCEVAddRecExpr *StoreEv =
287 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
288 if (StoreEv == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
291 // Check to see if the stride matches the size of the store. If so, then we
292 // know that every byte is touched in the loop.
293 unsigned StoreSize = (unsigned)SizeInBits >> 3;
294 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
296 if (Stride == 0 || StoreSize != Stride->getValue()->getValue()) {
297 // TODO: Could also handle negative stride here someday, that will require
298 // the validity check in mayLoopAccessLocation to be updated though.
299 // Enable this to print exact negative strides.
300 if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
301 dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
302 dbgs() << "BB: " << *SI->getParent();
308 // See if we can optimize just this store in isolation.
309 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
310 StoredVal, SI, StoreEv, BECount))
313 // If the stored value is a strided load in the same loop with the same stride
314 // this this may be transformable into a memcpy. This kicks in for stuff like
315 // for (i) A[i] = B[i];
316 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
317 const SCEVAddRecExpr *LoadEv =
318 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
319 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
320 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
321 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
324 //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
329 /// processLoopMemSet - See if this memset can be promoted to a large memset.
330 bool LoopIdiomRecognize::
331 processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) {
332 // We can only handle non-volatile memsets with a constant size.
333 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) return false;
335 // If we're not allowed to hack on memset, we fail.
336 if (!TLI->has(LibFunc::memset))
339 Value *Pointer = MSI->getDest();
341 // See if the pointer expression is an AddRec like {base,+,1} on the current
342 // loop, which indicates a strided store. If we have something else, it's a
343 // random store we can't handle.
344 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
345 if (Ev == 0 || Ev->getLoop() != CurLoop || !Ev->isAffine())
348 // Reject memsets that are so large that they overflow an unsigned.
349 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
350 if ((SizeInBytes >> 32) != 0)
353 // Check to see if the stride matches the size of the memset. If so, then we
354 // know that every byte is touched in the loop.
355 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
357 // TODO: Could also handle negative stride here someday, that will require the
358 // validity check in mayLoopAccessLocation to be updated though.
359 if (Stride == 0 || MSI->getLength() != Stride->getValue())
362 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
363 MSI->getAlignment(), MSI->getValue(),
367 /// getMemSetPatternValue - If a strided store of the specified value is safe to
368 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
369 /// be passed in. Otherwise, return null.
371 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
372 /// just replicate their input array and then pass on to memset_pattern16.
373 static Constant *getMemSetPatternValue(Value *V, const DataLayout &TD) {
374 // If the value isn't a constant, we can't promote it to being in a constant
375 // array. We could theoretically do a store to an alloca or something, but
376 // that doesn't seem worthwhile.
377 Constant *C = dyn_cast<Constant>(V);
378 if (C == 0) return 0;
380 // Only handle simple values that are a power of two bytes in size.
381 uint64_t Size = TD.getTypeSizeInBits(V->getType());
382 if (Size == 0 || (Size & 7) || (Size & (Size-1)))
385 // Don't care enough about darwin/ppc to implement this.
386 if (TD.isBigEndian())
389 // Convert to size in bytes.
392 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
393 // if the top and bottom are the same (e.g. for vectors and large integers).
394 if (Size > 16) return 0;
396 // If the constant is exactly 16 bytes, just use it.
397 if (Size == 16) return C;
399 // Otherwise, we'll use an array of the constants.
400 unsigned ArraySize = 16/Size;
401 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
402 return ConstantArray::get(AT, std::vector<Constant*>(ArraySize, C));
406 /// processLoopStridedStore - We see a strided store of some value. If we can
407 /// transform this into a memset or memset_pattern in the loop preheader, do so.
408 bool LoopIdiomRecognize::
409 processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
410 unsigned StoreAlignment, Value *StoredVal,
411 Instruction *TheStore, const SCEVAddRecExpr *Ev,
412 const SCEV *BECount) {
414 // If the stored value is a byte-wise value (like i32 -1), then it may be
415 // turned into a memset of i8 -1, assuming that all the consecutive bytes
416 // are stored. A store of i32 0x01020304 can never be turned into a memset,
417 // but it can be turned into memset_pattern if the target supports it.
418 Value *SplatValue = isBytewiseValue(StoredVal);
419 Constant *PatternValue = 0;
421 // If we're allowed to form a memset, and the stored value would be acceptable
422 // for memset, use it.
423 if (SplatValue && TLI->has(LibFunc::memset) &&
424 // Verify that the stored value is loop invariant. If not, we can't
425 // promote the memset.
426 CurLoop->isLoopInvariant(SplatValue)) {
427 // Keep and use SplatValue.
429 } else if (TLI->has(LibFunc::memset_pattern16) &&
430 (PatternValue = getMemSetPatternValue(StoredVal, *TD))) {
431 // It looks like we can use PatternValue!
434 // Otherwise, this isn't an idiom we can transform. For example, we can't
435 // do anything with a 3-byte store.
439 // Make sure the store has no dependencies (i.e. other loads and stores) in
441 DependenceAnalysis &DA = getAnalysis<DependenceAnalysis>();
442 for (Loop::block_iterator BI = CurLoop->block_begin(),
443 BE = CurLoop->block_end(); BI != BE; ++BI)
444 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
445 if (&*I != TheStore && I->mayReadOrWriteMemory()) {
446 OwningPtr<Dependence> D(DA.depends(TheStore, I, true));
451 // The trip count of the loop and the base pointer of the addrec SCEV is
452 // guaranteed to be loop invariant, which means that it should dominate the
453 // header. This allows us to insert code for it in the preheader.
454 BasicBlock *Preheader = CurLoop->getLoopPreheader();
455 IRBuilder<> Builder(Preheader->getTerminator());
456 SCEVExpander Expander(*SE, "loop-idiom");
458 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
459 // this into a memset in the loop preheader now if we want. However, this
460 // would be unsafe to do if there is anything else in the loop that may read
461 // or write to the aliased location. Check for any overlap by generating the
462 // base pointer and checking the region.
463 unsigned AddrSpace = cast<PointerType>(DestPtr->getType())->getAddressSpace();
465 Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace),
466 Preheader->getTerminator());
469 // Okay, everything looks good, insert the memset.
471 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
472 // pointer size if it isn't already.
473 Type *IntPtr = TD->getIntPtrType(DestPtr->getContext());
474 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
476 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
479 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
483 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
487 NewCall = Builder.CreateMemSet(BasePtr, SplatValue,NumBytes,StoreAlignment);
489 Module *M = TheStore->getParent()->getParent()->getParent();
490 Value *MSP = M->getOrInsertFunction("memset_pattern16",
492 Builder.getInt8PtrTy(),
493 Builder.getInt8PtrTy(), IntPtr,
496 // Otherwise we should form a memset_pattern16. PatternValue is known to be
497 // an constant array of 16-bytes. Plop the value into a mergable global.
498 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
499 GlobalValue::InternalLinkage,
500 PatternValue, ".memset_pattern");
501 GV->setUnnamedAddr(true); // Ok to merge these.
502 GV->setAlignment(16);
503 Value *PatternPtr = ConstantExpr::getBitCast(GV, Builder.getInt8PtrTy());
504 NewCall = Builder.CreateCall3(MSP, BasePtr, PatternPtr, NumBytes);
507 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
508 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
509 NewCall->setDebugLoc(TheStore->getDebugLoc());
511 // Okay, the memset has been formed. Zap the original store and anything that
513 deleteDeadInstruction(TheStore, *SE, TLI);
518 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
519 /// same-strided load.
520 bool LoopIdiomRecognize::
521 processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
522 const SCEVAddRecExpr *StoreEv,
523 const SCEVAddRecExpr *LoadEv,
524 const SCEV *BECount) {
525 // If we're not allowed to form memcpy, we fail.
526 if (!TLI->has(LibFunc::memcpy) || !TLI->has(LibFunc::memmove))
529 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
531 // Make sure the load and the store have no dependencies (i.e. other loads and
532 // stores) in the loop. We ignore the direct dependency between SI and LI here
533 // and check it later.
534 DependenceAnalysis &DA = getAnalysis<DependenceAnalysis>();
535 bool isMemcpySafe = true;
536 for (Loop::block_iterator BI = CurLoop->block_begin(),
537 BE = CurLoop->block_end(); BI != BE; ++BI)
538 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
539 if (&*I != SI && &*I != LI && I->mayReadOrWriteMemory()) {
540 // First, check if there is a dependence of the store.
541 OwningPtr<Dependence> DS(DA.depends(SI, I, true));
544 // If the scanned instructon may modify memory then we also have to
545 // check for dependencys on the load.
546 if (I->mayWriteToMemory()) {
547 OwningPtr<Dependence> DL(DA.depends(I, LI, true));
553 // Now check the dependency between SI and LI. If there is no dependency we
554 // can safely emit a memcpy.
555 OwningPtr<Dependence> Dep(DA.depends(SI, LI, true));
557 // If there is a dependence but the direction is positive (or none) we can
558 // still safely turn this into memmove.
559 unsigned Direction = Dep->getDirection(Dep->getLevels());
560 if (Direction != Dependence::DVEntry::NONE &&
561 Direction != Dependence::DVEntry::GT)
563 isMemcpySafe = false;
566 // The trip count of the loop and the base pointer of the addrec SCEV is
567 // guaranteed to be loop invariant, which means that it should dominate the
568 // header. This allows us to insert code for it in the preheader.
569 BasicBlock *Preheader = CurLoop->getLoopPreheader();
570 IRBuilder<> Builder(Preheader->getTerminator());
571 SCEVExpander Expander(*SE, "loop-idiom");
573 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
574 // this into a memcpy in the loop preheader now if we want.
575 Value *StoreBasePtr =
576 Expander.expandCodeFor(StoreEv->getStart(),
577 Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
578 Preheader->getTerminator());
580 Expander.expandCodeFor(LoadEv->getStart(),
581 Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
582 Preheader->getTerminator());
584 // Okay, everything is safe, we can transform this!
587 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
588 // pointer size if it isn't already.
589 Type *IntPtr = TD->getIntPtrType(SI->getContext());
590 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
592 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
595 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
599 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
602 unsigned Align = std::min(SI->getAlignment(), LI->getAlignment());
604 NewCall = Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes, Align);
607 NewCall = Builder.CreateMemMove(StoreBasePtr, LoadBasePtr, NumBytes, Align);
610 NewCall->setDebugLoc(SI->getDebugLoc());
612 DEBUG(dbgs() << " Formed " << (isMemcpySafe ? "memcpy: " : "memmove: ")
614 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
615 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
618 // Okay, the memset has been formed. Zap the original store and anything that
620 deleteDeadInstruction(SI, *SE, TLI);