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/GlobalVariable.h"
18 #include "llvm/IntrinsicInst.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/GetElementPtrTypeIterator.h"
28 #include "llvm/Support/IRBuilder.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/Target/TargetData.h"
34 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
35 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
36 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
37 STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
39 static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx,
40 bool &VariableIdxFound, const TargetData &TD){
41 // Skip over the first indices.
42 gep_type_iterator GTI = gep_type_begin(GEP);
43 for (unsigned i = 1; i != Idx; ++i, ++GTI)
46 // Compute the offset implied by the rest of the indices.
48 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
49 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
51 return VariableIdxFound = true;
52 if (OpC->isZero()) continue; // No offset.
54 // Handle struct indices, which add their field offset to the pointer.
55 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
56 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
60 // Otherwise, we have a sequential type like an array or vector. Multiply
61 // the index by the ElementSize.
62 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
63 Offset += Size*OpC->getSExtValue();
69 /// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a
70 /// constant offset, and return that constant offset. For example, Ptr1 might
71 /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8.
72 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
73 const TargetData &TD) {
74 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
75 // base. After that base, they may have some number of common (and
76 // potentially variable) indices. After that they handle some constant
77 // offset, which determines their offset from each other. At this point, we
78 // handle no other case.
79 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
80 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
81 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
84 // Skip any common indices and track the GEP types.
86 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
87 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
90 bool VariableIdxFound = false;
91 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
92 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
93 if (VariableIdxFound) return false;
95 Offset = Offset2-Offset1;
100 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
101 /// This allows us to analyze stores like:
106 /// which sometimes happens with stores to arrays of structs etc. When we see
107 /// the first store, we make a range [1, 2). The second store extends the range
108 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
109 /// two ranges into [0, 3) which is memset'able.
112 // Start/End - A semi range that describes the span that this range covers.
113 // The range is closed at the start and open at the end: [Start, End).
116 /// StartPtr - The getelementptr instruction that points to the start of the
120 /// Alignment - The known alignment of the first store.
123 /// TheStores - The actual stores that make up this range.
124 SmallVector<Instruction*, 16> TheStores;
126 bool isProfitableToUseMemset(const TargetData &TD) const;
129 } // end anon namespace
131 bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const {
132 // If we found more than 8 stores to merge or 64 bytes, use memset.
133 if (TheStores.size() >= 8 || End-Start >= 64) return true;
135 // If there is nothing to merge, don't do anything.
136 if (TheStores.size() < 2) return false;
138 // If any of the stores are a memset, then it is always good to extend the
140 for (unsigned i = 0, e = TheStores.size(); i != e; ++i)
141 if (!isa<StoreInst>(TheStores[i]))
144 // Assume that the code generator is capable of merging pairs of stores
145 // together if it wants to.
146 if (TheStores.size() == 2) return false;
148 // If we have fewer than 8 stores, it can still be worthwhile to do this.
149 // For example, merging 4 i8 stores into an i32 store is useful almost always.
150 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
151 // memset will be split into 2 32-bit stores anyway) and doing so can
152 // pessimize the llvm optimizer.
154 // Since we don't have perfect knowledge here, make some assumptions: assume
155 // the maximum GPR width is the same size as the pointer size and assume that
156 // this width can be stored. If so, check to see whether we will end up
157 // actually reducing the number of stores used.
158 unsigned Bytes = unsigned(End-Start);
159 unsigned NumPointerStores = Bytes/TD.getPointerSize();
161 // Assume the remaining bytes if any are done a byte at a time.
162 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
164 // If we will reduce the # stores (according to this heuristic), do the
165 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
167 return TheStores.size() > NumPointerStores+NumByteStores;
173 /// Ranges - A sorted list of the memset ranges. We use std::list here
174 /// because each element is relatively large and expensive to copy.
175 std::list<MemsetRange> Ranges;
176 typedef std::list<MemsetRange>::iterator range_iterator;
177 const TargetData &TD;
179 MemsetRanges(const TargetData &td) : TD(td) {}
181 typedef std::list<MemsetRange>::const_iterator const_iterator;
182 const_iterator begin() const { return Ranges.begin(); }
183 const_iterator end() const { return Ranges.end(); }
184 bool empty() const { return Ranges.empty(); }
186 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
187 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
188 addStore(OffsetFromFirst, SI);
190 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
193 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
194 int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType());
196 addRange(OffsetFromFirst, StoreSize,
197 SI->getPointerOperand(), SI->getAlignment(), SI);
200 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
201 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
202 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI);
205 void addRange(int64_t Start, int64_t Size, Value *Ptr,
206 unsigned Alignment, Instruction *Inst);
210 } // end anon namespace
213 /// addRange - Add a new store to the MemsetRanges data structure. This adds a
214 /// new range for the specified store at the specified offset, merging into
215 /// existing ranges as appropriate.
217 /// Do a linear search of the ranges to see if this can be joined and/or to
218 /// find the insertion point in the list. We keep the ranges sorted for
219 /// simplicity here. This is a linear search of a linked list, which is ugly,
220 /// however the number of ranges is limited, so this won't get crazy slow.
221 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
222 unsigned Alignment, Instruction *Inst) {
223 int64_t End = Start+Size;
224 range_iterator I = Ranges.begin(), E = Ranges.end();
226 while (I != E && Start > I->End)
229 // We now know that I == E, in which case we didn't find anything to merge
230 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
231 // to insert a new range. Handle this now.
232 if (I == E || End < I->Start) {
233 MemsetRange &R = *Ranges.insert(I, MemsetRange());
237 R.Alignment = Alignment;
238 R.TheStores.push_back(Inst);
242 // This store overlaps with I, add it.
243 I->TheStores.push_back(Inst);
245 // At this point, we may have an interval that completely contains our store.
246 // If so, just add it to the interval and return.
247 if (I->Start <= Start && I->End >= End)
250 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
251 // but is not entirely contained within the range.
253 // See if the range extends the start of the range. In this case, it couldn't
254 // possibly cause it to join the prior range, because otherwise we would have
256 if (Start < I->Start) {
259 I->Alignment = Alignment;
262 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
263 // is in or right at the end of I), and that End >= I->Start. Extend I out to
267 range_iterator NextI = I;
268 while (++NextI != E && End >= NextI->Start) {
269 // Merge the range in.
270 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
271 if (NextI->End > I->End)
279 //===----------------------------------------------------------------------===//
281 //===----------------------------------------------------------------------===//
284 class MemCpyOpt : public FunctionPass {
285 MemoryDependenceAnalysis *MD;
286 const TargetData *TD;
288 static char ID; // Pass identification, replacement for typeid
289 MemCpyOpt() : FunctionPass(ID) {
290 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
294 bool runOnFunction(Function &F);
297 // This transformation requires dominator postdominator info
298 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
299 AU.setPreservesCFG();
300 AU.addRequired<DominatorTree>();
301 AU.addRequired<MemoryDependenceAnalysis>();
302 AU.addRequired<AliasAnalysis>();
303 AU.addPreserved<AliasAnalysis>();
304 AU.addPreserved<MemoryDependenceAnalysis>();
308 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
309 bool processMemCpy(MemCpyInst *M);
310 bool processMemMove(MemMoveInst *M);
311 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
312 uint64_t cpyLen, CallInst *C);
313 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
315 bool processByValArgument(CallSite CS, unsigned ArgNo);
316 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr,
319 bool iterateOnFunction(Function &F);
322 char MemCpyOpt::ID = 0;
325 // createMemCpyOptPass - The public interface to this file...
326 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
328 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
330 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
331 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
332 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
333 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
336 /// tryMergingIntoMemset - When scanning forward over instructions, we look for
337 /// some other patterns to fold away. In particular, this looks for stores to
338 /// neighboring locations of memory. If it sees enough consequtive ones, it
339 /// attempts to merge them together into a memcpy/memset.
340 Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
341 Value *StartPtr, Value *ByteVal) {
342 if (TD == 0) return 0;
344 // Okay, so we now have a single store that can be splatable. Scan to find
345 // all subsequent stores of the same value to offset from the same pointer.
346 // Join these together into ranges, so we can decide whether contiguous blocks
348 MemsetRanges Ranges(*TD);
350 BasicBlock::iterator BI = StartInst;
351 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
352 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
353 // If the instruction is readnone, ignore it, otherwise bail out. We
354 // don't even allow readonly here because we don't want something like:
355 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
356 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
361 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
362 // If this is a store, see if we can merge it in.
363 if (NextStore->isVolatile()) break;
365 // Check to see if this stored value is of the same byte-splattable value.
366 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
369 // Check to see if this store is to a constant offset from the start ptr.
371 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
374 Ranges.addStore(Offset, NextStore);
376 MemSetInst *MSI = cast<MemSetInst>(BI);
378 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
379 !isa<ConstantInt>(MSI->getLength()))
382 // Check to see if this store is to a constant offset from the start ptr.
384 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD))
387 Ranges.addMemSet(Offset, MSI);
391 // If we have no ranges, then we just had a single store with nothing that
392 // could be merged in. This is a very common case of course.
396 // If we had at least one store that could be merged in, add the starting
397 // store as well. We try to avoid this unless there is at least something
398 // interesting as a small compile-time optimization.
399 Ranges.addInst(0, StartInst);
401 // If we create any memsets, we put it right before the first instruction that
402 // isn't part of the memset block. This ensure that the memset is dominated
403 // by any addressing instruction needed by the start of the block.
404 IRBuilder<> Builder(BI);
406 // Now that we have full information about ranges, loop over the ranges and
407 // emit memset's for anything big enough to be worthwhile.
408 Instruction *AMemSet = 0;
409 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
411 const MemsetRange &Range = *I;
413 if (Range.TheStores.size() == 1) continue;
415 // If it is profitable to lower this range to memset, do so now.
416 if (!Range.isProfitableToUseMemset(*TD))
419 // Otherwise, we do want to transform this! Create a new memset.
420 // Get the starting pointer of the block.
421 StartPtr = Range.StartPtr;
423 // Determine alignment
424 unsigned Alignment = Range.Alignment;
425 if (Alignment == 0) {
426 const Type *EltType =
427 cast<PointerType>(StartPtr->getType())->getElementType();
428 Alignment = TD->getABITypeAlignment(EltType);
432 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
434 DEBUG(dbgs() << "Replace stores:\n";
435 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
436 dbgs() << *Range.TheStores[i] << '\n';
437 dbgs() << "With: " << *AMemSet << '\n');
439 // Zap all the stores.
440 for (SmallVector<Instruction*, 16>::const_iterator
441 SI = Range.TheStores.begin(),
442 SE = Range.TheStores.end(); SI != SE; ++SI)
443 (*SI)->eraseFromParent();
451 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
452 if (SI->isVolatile()) return false;
454 if (TD == 0) return false;
456 // Detect cases where we're performing call slot forwarding, but
457 // happen to be using a load-store pair to implement it, rather than
459 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
460 if (!LI->isVolatile() && LI->hasOneUse()) {
461 MemDepResult dep = MD->getDependency(LI);
463 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
464 C = dyn_cast<CallInst>(dep.getInst());
467 bool changed = performCallSlotOptzn(LI,
468 SI->getPointerOperand()->stripPointerCasts(),
469 LI->getPointerOperand()->stripPointerCasts(),
470 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
472 MD->removeInstruction(SI);
473 SI->eraseFromParent();
474 LI->eraseFromParent();
482 // There are two cases that are interesting for this code to handle: memcpy
483 // and memset. Right now we only handle memset.
485 // Ensure that the value being stored is something that can be memset'able a
486 // byte at a time like "0" or "-1" or any width, as well as things like
487 // 0xA0A0A0A0 and 0.0.
488 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
489 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
491 BBI = I; // Don't invalidate iterator.
499 /// performCallSlotOptzn - takes a memcpy and a call that it depends on,
500 /// and checks for the possibility of a call slot optimization by having
501 /// the call write its result directly into the destination of the memcpy.
502 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
503 Value *cpyDest, Value *cpySrc,
504 uint64_t cpyLen, CallInst *C) {
505 // The general transformation to keep in mind is
507 // call @func(..., src, ...)
508 // memcpy(dest, src, ...)
512 // memcpy(dest, src, ...)
513 // call @func(..., dest, ...)
515 // Since moving the memcpy is technically awkward, we additionally check that
516 // src only holds uninitialized values at the moment of the call, meaning that
517 // the memcpy can be discarded rather than moved.
519 // Deliberately get the source and destination with bitcasts stripped away,
520 // because we'll need to do type comparisons based on the underlying type.
523 // Require that src be an alloca. This simplifies the reasoning considerably.
524 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
528 // Check that all of src is copied to dest.
529 if (TD == 0) return false;
531 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
535 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
536 srcArraySize->getZExtValue();
538 if (cpyLen < srcSize)
541 // Check that accessing the first srcSize bytes of dest will not cause a
542 // trap. Otherwise the transform is invalid since it might cause a trap
543 // to occur earlier than it otherwise would.
544 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
545 // The destination is an alloca. Check it is larger than srcSize.
546 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
550 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
551 destArraySize->getZExtValue();
553 if (destSize < srcSize)
555 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
556 // If the destination is an sret parameter then only accesses that are
557 // outside of the returned struct type can trap.
558 if (!A->hasStructRetAttr())
561 const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
562 uint64_t destSize = TD->getTypeAllocSize(StructTy);
564 if (destSize < srcSize)
570 // Check that src is not accessed except via the call and the memcpy. This
571 // guarantees that it holds only undefined values when passed in (so the final
572 // memcpy can be dropped), that it is not read or written between the call and
573 // the memcpy, and that writing beyond the end of it is undefined.
574 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
575 srcAlloca->use_end());
576 while (!srcUseList.empty()) {
577 User *UI = srcUseList.pop_back_val();
579 if (isa<BitCastInst>(UI)) {
580 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
582 srcUseList.push_back(*I);
583 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
584 if (G->hasAllZeroIndices())
585 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
587 srcUseList.push_back(*I);
590 } else if (UI != C && UI != cpy) {
595 // Since we're changing the parameter to the callsite, we need to make sure
596 // that what would be the new parameter dominates the callsite.
597 DominatorTree &DT = getAnalysis<DominatorTree>();
598 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
599 if (!DT.dominates(cpyDestInst, C))
602 // In addition to knowing that the call does not access src in some
603 // unexpected manner, for example via a global, which we deduce from
604 // the use analysis, we also need to know that it does not sneakily
605 // access dest. We rely on AA to figure this out for us.
606 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
607 if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef)
610 // All the checks have passed, so do the transformation.
611 bool changedArgument = false;
612 for (unsigned i = 0; i < CS.arg_size(); ++i)
613 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
614 if (cpySrc->getType() != cpyDest->getType())
615 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
616 cpyDest->getName(), C);
617 changedArgument = true;
618 if (CS.getArgument(i)->getType() == cpyDest->getType())
619 CS.setArgument(i, cpyDest);
621 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
622 CS.getArgument(i)->getType(), cpyDest->getName(), C));
625 if (!changedArgument)
628 // Drop any cached information about the call, because we may have changed
629 // its dependence information by changing its parameter.
630 MD->removeInstruction(C);
632 // Remove the memcpy.
633 MD->removeInstruction(cpy);
639 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)
640 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
641 /// copy from MDep's input if we can. MSize is the size of M's copy.
643 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
645 // We can only transforms memcpy's where the dest of one is the source of the
647 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
650 // If dep instruction is reading from our current input, then it is a noop
651 // transfer and substituting the input won't change this instruction. Just
652 // ignore the input and let someone else zap MDep. This handles cases like:
655 if (M->getSource() == MDep->getSource())
658 // Second, the length of the memcpy's must be the same, or the preceeding one
659 // must be larger than the following one.
660 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
661 if (!C1) return false;
663 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
665 // Verify that the copied-from memory doesn't change in between the two
666 // transfers. For example, in:
670 // It would be invalid to transform the second memcpy into memcpy(c <- b).
672 // TODO: If the code between M and MDep is transparent to the destination "c",
673 // then we could still perform the xform by moving M up to the first memcpy.
675 // NOTE: This is conservative, it will stop on any read from the source loc,
676 // not just the defining memcpy.
677 MemDepResult SourceDep =
678 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
679 false, M, M->getParent());
680 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
683 // If the dest of the second might alias the source of the first, then the
684 // source and dest might overlap. We still want to eliminate the intermediate
685 // value, but we have to generate a memmove instead of memcpy.
686 bool UseMemMove = false;
687 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)))
690 // If all checks passed, then we can transform M.
692 // Make sure to use the lesser of the alignment of the source and the dest
693 // since we're changing where we're reading from, but don't want to increase
694 // the alignment past what can be read from or written to.
695 // TODO: Is this worth it if we're creating a less aligned memcpy? For
696 // example we could be moving from movaps -> movq on x86.
697 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
699 IRBuilder<> Builder(M);
701 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
702 Align, M->isVolatile());
704 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
705 Align, M->isVolatile());
707 // Remove the instruction we're replacing.
708 MD->removeInstruction(M);
709 M->eraseFromParent();
715 /// processMemCpy - perform simplification of memcpy's. If we have memcpy A
716 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
717 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
718 /// circumstances). This allows later passes to remove the first memcpy
720 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
721 // We can only optimize statically-sized memcpy's that are non-volatile.
722 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
723 if (CopySize == 0 || M->isVolatile()) return false;
725 // If the source and destination of the memcpy are the same, then zap it.
726 if (M->getSource() == M->getDest()) {
727 MD->removeInstruction(M);
728 M->eraseFromParent();
732 // If copying from a constant, try to turn the memcpy into a memset.
733 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
734 if (GV->isConstant() && GV->hasDefinitiveInitializer())
735 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
736 IRBuilder<> Builder(M);
737 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize,
738 M->getAlignment(), false);
739 MD->removeInstruction(M);
740 M->eraseFromParent();
745 // The are two possible optimizations we can do for memcpy:
746 // a) memcpy-memcpy xform which exposes redundance for DSE.
747 // b) call-memcpy xform for return slot optimization.
748 MemDepResult DepInfo = MD->getDependency(M);
749 if (!DepInfo.isClobber())
752 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
753 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
755 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
756 if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
757 CopySize->getZExtValue(), C)) {
758 M->eraseFromParent();
765 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
766 /// are guaranteed not to alias.
767 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
768 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
770 // See if the pointers alias.
771 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M)))
774 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
776 // If not, then we know we can transform this.
777 Module *Mod = M->getParent()->getParent()->getParent();
778 const Type *ArgTys[3] = { M->getRawDest()->getType(),
779 M->getRawSource()->getType(),
780 M->getLength()->getType() };
781 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
784 // MemDep may have over conservative information about this instruction, just
785 // conservatively flush it from the cache.
786 MD->removeInstruction(M);
792 /// processByValArgument - This is called on every byval argument in call sites.
793 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
794 if (TD == 0) return false;
796 // Find out what feeds this byval argument.
797 Value *ByValArg = CS.getArgument(ArgNo);
798 const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType();
799 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
800 MemDepResult DepInfo =
801 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
802 true, CS.getInstruction(),
803 CS.getInstruction()->getParent());
804 if (!DepInfo.isClobber())
807 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
808 // a memcpy, see if we can byval from the source of the memcpy instead of the
810 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
811 if (MDep == 0 || MDep->isVolatile() ||
812 ByValArg->stripPointerCasts() != MDep->getDest())
815 // The length of the memcpy must be larger or equal to the size of the byval.
816 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
817 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
820 // Get the alignment of the byval. If it is greater than the memcpy, then we
821 // can't do the substitution. If the call doesn't specify the alignment, then
822 // it is some target specific value that we can't know.
823 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
824 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign)
827 // Verify that the copied-from memory doesn't change in between the memcpy and
832 // It would be invalid to transform the second memcpy into foo(*b).
834 // NOTE: This is conservative, it will stop on any read from the source loc,
835 // not just the defining memcpy.
836 MemDepResult SourceDep =
837 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
838 false, CS.getInstruction(), MDep->getParent());
839 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
842 Value *TmpCast = MDep->getSource();
843 if (MDep->getSource()->getType() != ByValArg->getType())
844 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
845 "tmpcast", CS.getInstruction());
847 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
848 << " " << *MDep << "\n"
849 << " " << *CS.getInstruction() << "\n");
851 // Otherwise we're good! Update the byval argument.
852 CS.setArgument(ArgNo, TmpCast);
857 /// iterateOnFunction - Executes one iteration of MemCpyOpt.
858 bool MemCpyOpt::iterateOnFunction(Function &F) {
859 bool MadeChange = false;
861 // Walk all instruction in the function.
862 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
863 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
864 // Avoid invalidating the iterator.
865 Instruction *I = BI++;
867 bool RepeatInstruction = false;
869 if (StoreInst *SI = dyn_cast<StoreInst>(I))
870 MadeChange |= processStore(SI, BI);
871 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) {
872 RepeatInstruction = processMemCpy(M);
873 } else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) {
874 RepeatInstruction = processMemMove(M);
875 } else if (CallSite CS = (Value*)I) {
876 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
877 if (CS.paramHasAttr(i+1, Attribute::ByVal))
878 MadeChange |= processByValArgument(CS, i);
881 // Reprocess the instruction if desired.
882 if (RepeatInstruction) {
892 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
895 bool MemCpyOpt::runOnFunction(Function &F) {
896 bool MadeChange = false;
897 MD = &getAnalysis<MemoryDependenceAnalysis>();
898 TD = getAnalysisIfAvailable<TargetData>();
900 if (!iterateOnFunction(F))