1 //===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===//
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 performs various transformations related to eliminating memcpy
11 // calls, or transforming sets of stores into memset's.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "memcpyopt"
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/Dominators.h"
21 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/GlobalVariable.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetLibraryInfo.h"
32 #include "llvm/Transforms/Utils/Local.h"
36 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
37 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
38 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
39 STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
41 static int64_t GetOffsetFromIndex(const GEPOperator *GEP, unsigned Idx,
42 bool &VariableIdxFound, const DataLayout &TD){
43 // Skip over the first indices.
44 gep_type_iterator GTI = gep_type_begin(GEP);
45 for (unsigned i = 1; i != Idx; ++i, ++GTI)
48 // Compute the offset implied by the rest of the indices.
50 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
51 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
53 return VariableIdxFound = true;
54 if (OpC->isZero()) continue; // No offset.
56 // Handle struct indices, which add their field offset to the pointer.
57 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
58 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
62 // Otherwise, we have a sequential type like an array or vector. Multiply
63 // the index by the ElementSize.
64 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
65 Offset += Size*OpC->getSExtValue();
71 /// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a
72 /// constant offset, and return that constant offset. For example, Ptr1 might
73 /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8.
74 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
75 const DataLayout &TD) {
76 Ptr1 = Ptr1->stripPointerCasts();
77 Ptr2 = Ptr2->stripPointerCasts();
78 GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1);
79 GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2);
81 bool VariableIdxFound = false;
83 // If one pointer is a GEP and the other isn't, then see if the GEP is a
84 // constant offset from the base, as in "P" and "gep P, 1".
85 if (GEP1 && GEP2 == 0 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
86 Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, TD);
87 return !VariableIdxFound;
90 if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
91 Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD);
92 return !VariableIdxFound;
95 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
96 // base. After that base, they may have some number of common (and
97 // potentially variable) indices. After that they handle some constant
98 // offset, which determines their offset from each other. At this point, we
99 // handle no other case.
100 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
103 // Skip any common indices and track the GEP types.
105 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
106 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
109 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
110 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
111 if (VariableIdxFound) return false;
113 Offset = Offset2-Offset1;
118 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
119 /// This allows us to analyze stores like:
124 /// which sometimes happens with stores to arrays of structs etc. When we see
125 /// the first store, we make a range [1, 2). The second store extends the range
126 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
127 /// two ranges into [0, 3) which is memset'able.
130 // Start/End - A semi range that describes the span that this range covers.
131 // The range is closed at the start and open at the end: [Start, End).
134 /// StartPtr - The getelementptr instruction that points to the start of the
138 /// Alignment - The known alignment of the first store.
141 /// TheStores - The actual stores that make up this range.
142 SmallVector<Instruction*, 16> TheStores;
144 bool isProfitableToUseMemset(const DataLayout &TD) const;
147 } // end anon namespace
149 bool MemsetRange::isProfitableToUseMemset(const DataLayout &TD) const {
150 // If we found more than 4 stores to merge or 16 bytes, use memset.
151 if (TheStores.size() >= 4 || End-Start >= 16) return true;
153 // If there is nothing to merge, don't do anything.
154 if (TheStores.size() < 2) return false;
156 // If any of the stores are a memset, then it is always good to extend the
158 for (unsigned i = 0, e = TheStores.size(); i != e; ++i)
159 if (!isa<StoreInst>(TheStores[i]))
162 // Assume that the code generator is capable of merging pairs of stores
163 // together if it wants to.
164 if (TheStores.size() == 2) return false;
166 // If we have fewer than 8 stores, it can still be worthwhile to do this.
167 // For example, merging 4 i8 stores into an i32 store is useful almost always.
168 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
169 // memset will be split into 2 32-bit stores anyway) and doing so can
170 // pessimize the llvm optimizer.
172 // Since we don't have perfect knowledge here, make some assumptions: assume
173 // the maximum GPR width is the same size as the largest legal integer
174 // size. If so, check to see whether we will end up actually reducing the
175 // number of stores used.
176 unsigned Bytes = unsigned(End-Start);
177 unsigned MaxIntSize = TD.getLargestLegalIntTypeSize();
180 unsigned NumPointerStores = Bytes / MaxIntSize;
182 // Assume the remaining bytes if any are done a byte at a time.
183 unsigned NumByteStores = Bytes - NumPointerStores * MaxIntSize;
185 // If we will reduce the # stores (according to this heuristic), do the
186 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
188 return TheStores.size() > NumPointerStores+NumByteStores;
194 /// Ranges - A sorted list of the memset ranges. We use std::list here
195 /// because each element is relatively large and expensive to copy.
196 std::list<MemsetRange> Ranges;
197 typedef std::list<MemsetRange>::iterator range_iterator;
198 const DataLayout &TD;
200 MemsetRanges(const DataLayout &td) : TD(td) {}
202 typedef std::list<MemsetRange>::const_iterator const_iterator;
203 const_iterator begin() const { return Ranges.begin(); }
204 const_iterator end() const { return Ranges.end(); }
205 bool empty() const { return Ranges.empty(); }
207 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
208 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
209 addStore(OffsetFromFirst, SI);
211 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
214 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
215 int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType());
217 addRange(OffsetFromFirst, StoreSize,
218 SI->getPointerOperand(), SI->getAlignment(), SI);
221 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
222 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
223 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI);
226 void addRange(int64_t Start, int64_t Size, Value *Ptr,
227 unsigned Alignment, Instruction *Inst);
231 } // end anon namespace
234 /// addRange - Add a new store to the MemsetRanges data structure. This adds a
235 /// new range for the specified store at the specified offset, merging into
236 /// existing ranges as appropriate.
238 /// Do a linear search of the ranges to see if this can be joined and/or to
239 /// find the insertion point in the list. We keep the ranges sorted for
240 /// simplicity here. This is a linear search of a linked list, which is ugly,
241 /// however the number of ranges is limited, so this won't get crazy slow.
242 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
243 unsigned Alignment, Instruction *Inst) {
244 int64_t End = Start+Size;
245 range_iterator I = Ranges.begin(), E = Ranges.end();
247 while (I != E && Start > I->End)
250 // We now know that I == E, in which case we didn't find anything to merge
251 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
252 // to insert a new range. Handle this now.
253 if (I == E || End < I->Start) {
254 MemsetRange &R = *Ranges.insert(I, MemsetRange());
258 R.Alignment = Alignment;
259 R.TheStores.push_back(Inst);
263 // This store overlaps with I, add it.
264 I->TheStores.push_back(Inst);
266 // At this point, we may have an interval that completely contains our store.
267 // If so, just add it to the interval and return.
268 if (I->Start <= Start && I->End >= End)
271 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
272 // but is not entirely contained within the range.
274 // See if the range extends the start of the range. In this case, it couldn't
275 // possibly cause it to join the prior range, because otherwise we would have
277 if (Start < I->Start) {
280 I->Alignment = Alignment;
283 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
284 // is in or right at the end of I), and that End >= I->Start. Extend I out to
288 range_iterator NextI = I;
289 while (++NextI != E && End >= NextI->Start) {
290 // Merge the range in.
291 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
292 if (NextI->End > I->End)
300 //===----------------------------------------------------------------------===//
302 //===----------------------------------------------------------------------===//
305 class MemCpyOpt : public FunctionPass {
306 MemoryDependenceAnalysis *MD;
307 TargetLibraryInfo *TLI;
308 const DataLayout *TD;
310 static char ID; // Pass identification, replacement for typeid
311 MemCpyOpt() : FunctionPass(ID) {
312 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
318 bool runOnFunction(Function &F);
321 // This transformation requires dominator postdominator info
322 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
323 AU.setPreservesCFG();
324 AU.addRequired<DominatorTree>();
325 AU.addRequired<MemoryDependenceAnalysis>();
326 AU.addRequired<AliasAnalysis>();
327 AU.addRequired<TargetLibraryInfo>();
328 AU.addPreserved<AliasAnalysis>();
329 AU.addPreserved<MemoryDependenceAnalysis>();
333 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
334 bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI);
335 bool processMemCpy(MemCpyInst *M);
336 bool processMemMove(MemMoveInst *M);
337 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
338 uint64_t cpyLen, unsigned cpyAlign, CallInst *C);
339 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
341 bool processByValArgument(CallSite CS, unsigned ArgNo);
342 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr,
345 bool iterateOnFunction(Function &F);
348 char MemCpyOpt::ID = 0;
351 // createMemCpyOptPass - The public interface to this file...
352 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
354 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
356 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
357 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
358 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
359 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
360 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
363 /// tryMergingIntoMemset - When scanning forward over instructions, we look for
364 /// some other patterns to fold away. In particular, this looks for stores to
365 /// neighboring locations of memory. If it sees enough consecutive ones, it
366 /// attempts to merge them together into a memcpy/memset.
367 Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
368 Value *StartPtr, Value *ByteVal) {
369 if (TD == 0) return 0;
371 // Okay, so we now have a single store that can be splatable. Scan to find
372 // all subsequent stores of the same value to offset from the same pointer.
373 // Join these together into ranges, so we can decide whether contiguous blocks
375 MemsetRanges Ranges(*TD);
377 BasicBlock::iterator BI = StartInst;
378 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
379 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
380 // If the instruction is readnone, ignore it, otherwise bail out. We
381 // don't even allow readonly here because we don't want something like:
382 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
383 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
388 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
389 // If this is a store, see if we can merge it in.
390 if (!NextStore->isSimple()) break;
392 // Check to see if this stored value is of the same byte-splattable value.
393 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
396 // Check to see if this store is to a constant offset from the start ptr.
398 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(),
402 Ranges.addStore(Offset, NextStore);
404 MemSetInst *MSI = cast<MemSetInst>(BI);
406 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
407 !isa<ConstantInt>(MSI->getLength()))
410 // Check to see if this store is to a constant offset from the start ptr.
412 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD))
415 Ranges.addMemSet(Offset, MSI);
419 // If we have no ranges, then we just had a single store with nothing that
420 // could be merged in. This is a very common case of course.
424 // If we had at least one store that could be merged in, add the starting
425 // store as well. We try to avoid this unless there is at least something
426 // interesting as a small compile-time optimization.
427 Ranges.addInst(0, StartInst);
429 // If we create any memsets, we put it right before the first instruction that
430 // isn't part of the memset block. This ensure that the memset is dominated
431 // by any addressing instruction needed by the start of the block.
432 IRBuilder<> Builder(BI);
434 // Now that we have full information about ranges, loop over the ranges and
435 // emit memset's for anything big enough to be worthwhile.
436 Instruction *AMemSet = 0;
437 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
439 const MemsetRange &Range = *I;
441 if (Range.TheStores.size() == 1) continue;
443 // If it is profitable to lower this range to memset, do so now.
444 if (!Range.isProfitableToUseMemset(*TD))
447 // Otherwise, we do want to transform this! Create a new memset.
448 // Get the starting pointer of the block.
449 StartPtr = Range.StartPtr;
451 // Determine alignment
452 unsigned Alignment = Range.Alignment;
453 if (Alignment == 0) {
455 cast<PointerType>(StartPtr->getType())->getElementType();
456 Alignment = TD->getABITypeAlignment(EltType);
460 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
462 DEBUG(dbgs() << "Replace stores:\n";
463 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
464 dbgs() << *Range.TheStores[i] << '\n';
465 dbgs() << "With: " << *AMemSet << '\n');
467 if (!Range.TheStores.empty())
468 AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
470 // Zap all the stores.
471 for (SmallVectorImpl<Instruction *>::const_iterator
472 SI = Range.TheStores.begin(),
473 SE = Range.TheStores.end(); SI != SE; ++SI) {
474 MD->removeInstruction(*SI);
475 (*SI)->eraseFromParent();
484 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
485 if (!SI->isSimple()) return false;
487 if (TD == 0) return false;
489 // Detect cases where we're performing call slot forwarding, but
490 // happen to be using a load-store pair to implement it, rather than
492 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
493 if (LI->isSimple() && LI->hasOneUse() &&
494 LI->getParent() == SI->getParent()) {
495 MemDepResult ldep = MD->getDependency(LI);
497 if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
498 C = dyn_cast<CallInst>(ldep.getInst());
501 // Check that nothing touches the dest of the "copy" between
502 // the call and the store.
503 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
504 AliasAnalysis::Location StoreLoc = AA.getLocation(SI);
505 for (BasicBlock::iterator I = --BasicBlock::iterator(SI),
506 E = C; I != E; --I) {
507 if (AA.getModRefInfo(&*I, StoreLoc) != AliasAnalysis::NoModRef) {
515 unsigned storeAlign = SI->getAlignment();
517 storeAlign = TD->getABITypeAlignment(SI->getOperand(0)->getType());
518 unsigned loadAlign = LI->getAlignment();
520 loadAlign = TD->getABITypeAlignment(LI->getType());
522 bool changed = performCallSlotOptzn(LI,
523 SI->getPointerOperand()->stripPointerCasts(),
524 LI->getPointerOperand()->stripPointerCasts(),
525 TD->getTypeStoreSize(SI->getOperand(0)->getType()),
526 std::min(storeAlign, loadAlign), C);
528 MD->removeInstruction(SI);
529 SI->eraseFromParent();
530 MD->removeInstruction(LI);
531 LI->eraseFromParent();
539 // There are two cases that are interesting for this code to handle: memcpy
540 // and memset. Right now we only handle memset.
542 // Ensure that the value being stored is something that can be memset'able a
543 // byte at a time like "0" or "-1" or any width, as well as things like
544 // 0xA0A0A0A0 and 0.0.
545 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
546 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
548 BBI = I; // Don't invalidate iterator.
555 bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
556 // See if there is another memset or store neighboring this memset which
557 // allows us to widen out the memset to do a single larger store.
558 if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
559 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
561 BBI = I; // Don't invalidate iterator.
568 /// performCallSlotOptzn - takes a memcpy and a call that it depends on,
569 /// and checks for the possibility of a call slot optimization by having
570 /// the call write its result directly into the destination of the memcpy.
571 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
572 Value *cpyDest, Value *cpySrc,
573 uint64_t cpyLen, unsigned cpyAlign,
575 // The general transformation to keep in mind is
577 // call @func(..., src, ...)
578 // memcpy(dest, src, ...)
582 // memcpy(dest, src, ...)
583 // call @func(..., dest, ...)
585 // Since moving the memcpy is technically awkward, we additionally check that
586 // src only holds uninitialized values at the moment of the call, meaning that
587 // the memcpy can be discarded rather than moved.
589 // Deliberately get the source and destination with bitcasts stripped away,
590 // because we'll need to do type comparisons based on the underlying type.
593 // Require that src be an alloca. This simplifies the reasoning considerably.
594 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
598 // Check that all of src is copied to dest.
599 if (TD == 0) return false;
601 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
605 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
606 srcArraySize->getZExtValue();
608 if (cpyLen < srcSize)
611 // Check that accessing the first srcSize bytes of dest will not cause a
612 // trap. Otherwise the transform is invalid since it might cause a trap
613 // to occur earlier than it otherwise would.
614 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
615 // The destination is an alloca. Check it is larger than srcSize.
616 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
620 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
621 destArraySize->getZExtValue();
623 if (destSize < srcSize)
625 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
626 // If the destination is an sret parameter then only accesses that are
627 // outside of the returned struct type can trap.
628 if (!A->hasStructRetAttr())
631 Type *StructTy = cast<PointerType>(A->getType())->getElementType();
632 if (!StructTy->isSized()) {
633 // The call may never return and hence the copy-instruction may never
634 // be executed, and therefore it's not safe to say "the destination
635 // has at least <cpyLen> bytes, as implied by the copy-instruction",
639 uint64_t destSize = TD->getTypeAllocSize(StructTy);
640 if (destSize < srcSize)
646 // Check that dest points to memory that is at least as aligned as src.
647 unsigned srcAlign = srcAlloca->getAlignment();
649 srcAlign = TD->getABITypeAlignment(srcAlloca->getAllocatedType());
650 bool isDestSufficientlyAligned = srcAlign <= cpyAlign;
651 // If dest is not aligned enough and we can't increase its alignment then
653 if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest))
656 // Check that src is not accessed except via the call and the memcpy. This
657 // guarantees that it holds only undefined values when passed in (so the final
658 // memcpy can be dropped), that it is not read or written between the call and
659 // the memcpy, and that writing beyond the end of it is undefined.
660 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
661 srcAlloca->use_end());
662 while (!srcUseList.empty()) {
663 User *UI = srcUseList.pop_back_val();
665 if (isa<BitCastInst>(UI)) {
666 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
668 srcUseList.push_back(*I);
669 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
670 if (G->hasAllZeroIndices())
671 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
673 srcUseList.push_back(*I);
676 } else if (UI != C && UI != cpy) {
681 // Since we're changing the parameter to the callsite, we need to make sure
682 // that what would be the new parameter dominates the callsite.
683 DominatorTree &DT = getAnalysis<DominatorTree>();
684 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
685 if (!DT.dominates(cpyDestInst, C))
688 // In addition to knowing that the call does not access src in some
689 // unexpected manner, for example via a global, which we deduce from
690 // the use analysis, we also need to know that it does not sneakily
691 // access dest. We rely on AA to figure this out for us.
692 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
693 AliasAnalysis::ModRefResult MR = AA.getModRefInfo(C, cpyDest, srcSize);
694 // If necessary, perform additional analysis.
695 if (MR != AliasAnalysis::NoModRef)
696 MR = AA.callCapturesBefore(C, cpyDest, srcSize, &DT);
697 if (MR != AliasAnalysis::NoModRef)
700 // All the checks have passed, so do the transformation.
701 bool changedArgument = false;
702 for (unsigned i = 0; i < CS.arg_size(); ++i)
703 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
704 Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest
705 : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
706 cpyDest->getName(), C);
707 changedArgument = true;
708 if (CS.getArgument(i)->getType() == Dest->getType())
709 CS.setArgument(i, Dest);
711 CS.setArgument(i, CastInst::CreatePointerCast(Dest,
712 CS.getArgument(i)->getType(), Dest->getName(), C));
715 if (!changedArgument)
718 // If the destination wasn't sufficiently aligned then increase its alignment.
719 if (!isDestSufficientlyAligned) {
720 assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!");
721 cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
724 // Drop any cached information about the call, because we may have changed
725 // its dependence information by changing its parameter.
726 MD->removeInstruction(C);
728 // Remove the memcpy.
729 MD->removeInstruction(cpy);
735 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)
736 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
737 /// copy from MDep's input if we can. MSize is the size of M's copy.
739 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
741 // We can only transforms memcpy's where the dest of one is the source of the
743 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
746 // If dep instruction is reading from our current input, then it is a noop
747 // transfer and substituting the input won't change this instruction. Just
748 // ignore the input and let someone else zap MDep. This handles cases like:
751 if (M->getSource() == MDep->getSource())
754 // Second, the length of the memcpy's must be the same, or the preceding one
755 // must be larger than the following one.
756 ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
757 ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength());
758 if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
761 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
763 // Verify that the copied-from memory doesn't change in between the two
764 // transfers. For example, in:
768 // It would be invalid to transform the second memcpy into memcpy(c <- b).
770 // TODO: If the code between M and MDep is transparent to the destination "c",
771 // then we could still perform the xform by moving M up to the first memcpy.
773 // NOTE: This is conservative, it will stop on any read from the source loc,
774 // not just the defining memcpy.
775 MemDepResult SourceDep =
776 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
777 false, M, M->getParent());
778 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
781 // If the dest of the second might alias the source of the first, then the
782 // source and dest might overlap. We still want to eliminate the intermediate
783 // value, but we have to generate a memmove instead of memcpy.
784 bool UseMemMove = false;
785 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)))
788 // If all checks passed, then we can transform M.
790 // Make sure to use the lesser of the alignment of the source and the dest
791 // since we're changing where we're reading from, but don't want to increase
792 // the alignment past what can be read from or written to.
793 // TODO: Is this worth it if we're creating a less aligned memcpy? For
794 // example we could be moving from movaps -> movq on x86.
795 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
797 IRBuilder<> Builder(M);
799 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
800 Align, M->isVolatile());
802 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
803 Align, M->isVolatile());
805 // Remove the instruction we're replacing.
806 MD->removeInstruction(M);
807 M->eraseFromParent();
813 /// processMemCpy - perform simplification of memcpy's. If we have memcpy A
814 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
815 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
816 /// circumstances). This allows later passes to remove the first memcpy
818 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
819 // We can only optimize statically-sized memcpy's that are non-volatile.
820 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
821 if (CopySize == 0 || M->isVolatile()) return false;
823 // If the source and destination of the memcpy are the same, then zap it.
824 if (M->getSource() == M->getDest()) {
825 MD->removeInstruction(M);
826 M->eraseFromParent();
830 // If copying from a constant, try to turn the memcpy into a memset.
831 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
832 if (GV->isConstant() && GV->hasDefinitiveInitializer())
833 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
834 IRBuilder<> Builder(M);
835 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize,
836 M->getAlignment(), false);
837 MD->removeInstruction(M);
838 M->eraseFromParent();
843 // The are two possible optimizations we can do for memcpy:
844 // a) memcpy-memcpy xform which exposes redundance for DSE.
845 // b) call-memcpy xform for return slot optimization.
846 MemDepResult DepInfo = MD->getDependency(M);
847 if (DepInfo.isClobber()) {
848 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
849 if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
850 CopySize->getZExtValue(), M->getAlignment(),
852 MD->removeInstruction(M);
853 M->eraseFromParent();
859 AliasAnalysis::Location SrcLoc = AliasAnalysis::getLocationForSource(M);
860 MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(SrcLoc, true,
862 if (SrcDepInfo.isClobber()) {
863 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))
864 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
870 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
871 /// are guaranteed not to alias.
872 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
873 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
875 if (!TLI->has(LibFunc::memmove))
878 // See if the pointers alias.
879 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M)))
882 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
884 // If not, then we know we can transform this.
885 Module *Mod = M->getParent()->getParent()->getParent();
886 Type *ArgTys[3] = { M->getRawDest()->getType(),
887 M->getRawSource()->getType(),
888 M->getLength()->getType() };
889 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
892 // MemDep may have over conservative information about this instruction, just
893 // conservatively flush it from the cache.
894 MD->removeInstruction(M);
900 /// processByValArgument - This is called on every byval argument in call sites.
901 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
902 if (TD == 0) return false;
904 // Find out what feeds this byval argument.
905 Value *ByValArg = CS.getArgument(ArgNo);
906 Type *ByValTy = cast<PointerType>(ByValArg->getType())->getElementType();
907 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
908 MemDepResult DepInfo =
909 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
910 true, CS.getInstruction(),
911 CS.getInstruction()->getParent());
912 if (!DepInfo.isClobber())
915 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
916 // a memcpy, see if we can byval from the source of the memcpy instead of the
918 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
919 if (MDep == 0 || MDep->isVolatile() ||
920 ByValArg->stripPointerCasts() != MDep->getDest())
923 // The length of the memcpy must be larger or equal to the size of the byval.
924 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
925 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
928 // Get the alignment of the byval. If the call doesn't specify the alignment,
929 // then it is some target specific value that we can't know.
930 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
931 if (ByValAlign == 0) return false;
933 // If it is greater than the memcpy, then we check to see if we can force the
934 // source of the memcpy to the alignment we need. If we fail, we bail out.
935 if (MDep->getAlignment() < ByValAlign &&
936 getOrEnforceKnownAlignment(MDep->getSource(),ByValAlign, TD) < ByValAlign)
939 // Verify that the copied-from memory doesn't change in between the memcpy and
944 // It would be invalid to transform the second memcpy into foo(*b).
946 // NOTE: This is conservative, it will stop on any read from the source loc,
947 // not just the defining memcpy.
948 MemDepResult SourceDep =
949 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
950 false, CS.getInstruction(), MDep->getParent());
951 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
954 Value *TmpCast = MDep->getSource();
955 if (MDep->getSource()->getType() != ByValArg->getType())
956 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
957 "tmpcast", CS.getInstruction());
959 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
960 << " " << *MDep << "\n"
961 << " " << *CS.getInstruction() << "\n");
963 // Otherwise we're good! Update the byval argument.
964 CS.setArgument(ArgNo, TmpCast);
969 /// iterateOnFunction - Executes one iteration of MemCpyOpt.
970 bool MemCpyOpt::iterateOnFunction(Function &F) {
971 bool MadeChange = false;
973 // Walk all instruction in the function.
974 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
975 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
976 // Avoid invalidating the iterator.
977 Instruction *I = BI++;
979 bool RepeatInstruction = false;
981 if (StoreInst *SI = dyn_cast<StoreInst>(I))
982 MadeChange |= processStore(SI, BI);
983 else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
984 RepeatInstruction = processMemSet(M, BI);
985 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
986 RepeatInstruction = processMemCpy(M);
987 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I))
988 RepeatInstruction = processMemMove(M);
989 else if (CallSite CS = (Value*)I) {
990 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
991 if (CS.isByValArgument(i))
992 MadeChange |= processByValArgument(CS, i);
995 // Reprocess the instruction if desired.
996 if (RepeatInstruction) {
997 if (BI != BB->begin()) --BI;
1006 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
1009 bool MemCpyOpt::runOnFunction(Function &F) {
1010 bool MadeChange = false;
1011 MD = &getAnalysis<MemoryDependenceAnalysis>();
1012 TD = getAnalysisIfAvailable<DataLayout>();
1013 TLI = &getAnalysis<TargetLibraryInfo>();
1015 // If we don't have at least memset and memcpy, there is little point of doing
1016 // anything here. These are required by a freestanding implementation, so if
1017 // even they are disabled, there is no point in trying hard.
1018 if (!TLI->has(LibFunc::memset) || !TLI->has(LibFunc::memcpy))
1022 if (!iterateOnFunction(F))