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, 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,
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<StoreInst*, 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 // Assume that the code generator is capable of merging pairs of stores
136 // together if it wants to.
137 if (TheStores.size() <= 2) return false;
139 // If we have fewer than 8 stores, it can still be worthwhile to do this.
140 // For example, merging 4 i8 stores into an i32 store is useful almost always.
141 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
142 // memset will be split into 2 32-bit stores anyway) and doing so can
143 // pessimize the llvm optimizer.
145 // Since we don't have perfect knowledge here, make some assumptions: assume
146 // the maximum GPR width is the same size as the pointer size and assume that
147 // this width can be stored. If so, check to see whether we will end up
148 // actually reducing the number of stores used.
149 unsigned Bytes = unsigned(End-Start);
150 unsigned NumPointerStores = Bytes/TD.getPointerSize();
152 // Assume the remaining bytes if any are done a byte at a time.
153 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
155 // If we will reduce the # stores (according to this heuristic), do the
156 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
158 return TheStores.size() > NumPointerStores+NumByteStores;
164 /// Ranges - A sorted list of the memset ranges. We use std::list here
165 /// because each element is relatively large and expensive to copy.
166 std::list<MemsetRange> Ranges;
167 typedef std::list<MemsetRange>::iterator range_iterator;
170 MemsetRanges(TargetData &td) : TD(td) {}
172 typedef std::list<MemsetRange>::const_iterator const_iterator;
173 const_iterator begin() const { return Ranges.begin(); }
174 const_iterator end() const { return Ranges.end(); }
175 bool empty() const { return Ranges.empty(); }
177 void addStore(int64_t OffsetFromFirst, StoreInst *SI);
180 } // end anon namespace
183 /// addStore - Add a new store to the MemsetRanges data structure. This adds a
184 /// new range for the specified store at the specified offset, merging into
185 /// existing ranges as appropriate.
186 void MemsetRanges::addStore(int64_t Start, StoreInst *SI) {
187 int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType());
189 // Do a linear search of the ranges to see if this can be joined and/or to
190 // find the insertion point in the list. We keep the ranges sorted for
191 // simplicity here. This is a linear search of a linked list, which is ugly,
192 // however the number of ranges is limited, so this won't get crazy slow.
193 range_iterator I = Ranges.begin(), E = Ranges.end();
195 while (I != E && Start > I->End)
198 // We now know that I == E, in which case we didn't find anything to merge
199 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
200 // to insert a new range. Handle this now.
201 if (I == E || End < I->Start) {
202 MemsetRange &R = *Ranges.insert(I, MemsetRange());
205 R.StartPtr = SI->getPointerOperand();
206 R.Alignment = SI->getAlignment();
207 R.TheStores.push_back(SI);
211 // This store overlaps with I, add it.
212 I->TheStores.push_back(SI);
214 // At this point, we may have an interval that completely contains our store.
215 // If so, just add it to the interval and return.
216 if (I->Start <= Start && I->End >= End)
219 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
220 // but is not entirely contained within the range.
222 // See if the range extends the start of the range. In this case, it couldn't
223 // possibly cause it to join the prior range, because otherwise we would have
225 if (Start < I->Start) {
227 I->StartPtr = SI->getPointerOperand();
228 I->Alignment = SI->getAlignment();
231 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
232 // is in or right at the end of I), and that End >= I->Start. Extend I out to
236 range_iterator NextI = I;
237 while (++NextI != E && End >= NextI->Start) {
238 // Merge the range in.
239 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
240 if (NextI->End > I->End)
248 //===----------------------------------------------------------------------===//
250 //===----------------------------------------------------------------------===//
253 class MemCpyOpt : public FunctionPass {
254 MemoryDependenceAnalysis *MD;
255 bool runOnFunction(Function &F);
257 static char ID; // Pass identification, replacement for typeid
258 MemCpyOpt() : FunctionPass(ID) {
259 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
264 // This transformation requires dominator postdominator info
265 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
266 AU.setPreservesCFG();
267 AU.addRequired<DominatorTree>();
268 AU.addRequired<MemoryDependenceAnalysis>();
269 AU.addRequired<AliasAnalysis>();
270 AU.addPreserved<AliasAnalysis>();
271 AU.addPreserved<MemoryDependenceAnalysis>();
275 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
276 bool processMemCpy(MemCpyInst *M);
277 bool processMemMove(MemMoveInst *M);
278 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
279 uint64_t cpyLen, CallInst *C);
280 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
282 bool processByValArgument(CallSite CS, unsigned ArgNo);
283 bool iterateOnFunction(Function &F);
286 char MemCpyOpt::ID = 0;
289 // createMemCpyOptPass - The public interface to this file...
290 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
292 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
294 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
295 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
296 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
297 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
300 /// processStore - When GVN is scanning forward over instructions, we look for
301 /// some other patterns to fold away. In particular, this looks for stores to
302 /// neighboring locations of memory. If it sees enough consequtive ones
303 /// (currently 4) it attempts to merge them together into a memcpy/memset.
304 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
305 if (SI->isVolatile()) return false;
307 TargetData *TD = getAnalysisIfAvailable<TargetData>();
308 if (!TD) return false;
310 // Detect cases where we're performing call slot forwarding, but
311 // happen to be using a load-store pair to implement it, rather than
313 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
314 if (!LI->isVolatile() && LI->hasOneUse()) {
315 MemDepResult dep = MD->getDependency(LI);
317 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
318 C = dyn_cast<CallInst>(dep.getInst());
321 bool changed = performCallSlotOptzn(LI,
322 SI->getPointerOperand()->stripPointerCasts(),
323 LI->getPointerOperand()->stripPointerCasts(),
324 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
326 MD->removeInstruction(SI);
327 SI->eraseFromParent();
328 LI->eraseFromParent();
336 // There are two cases that are interesting for this code to handle: memcpy
337 // and memset. Right now we only handle memset.
339 // Ensure that the value being stored is something that can be memset'able a
340 // byte at a time like "0" or "-1" or any width, as well as things like
341 // 0xA0A0A0A0 and 0.0.
342 Value *ByteVal = isBytewiseValue(SI->getOperand(0));
346 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
348 // Okay, so we now have a single store that can be splatable. Scan to find
349 // all subsequent stores of the same value to offset from the same pointer.
350 // Join these together into ranges, so we can decide whether contiguous blocks
352 MemsetRanges Ranges(*TD);
354 Value *StartPtr = SI->getPointerOperand();
356 BasicBlock::iterator BI = SI;
357 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
358 if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
359 // If the call is readnone, ignore it, otherwise bail out. We don't even
360 // allow readonly here because we don't want something like:
361 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
362 if (AA.getModRefBehavior(CallSite(BI)) ==
363 AliasAnalysis::DoesNotAccessMemory)
366 // TODO: If this is a memset, try to join it in.
369 } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI))
372 // If this is a non-store instruction it is fine, ignore it.
373 StoreInst *NextStore = dyn_cast<StoreInst>(BI);
374 if (NextStore == 0) continue;
376 // If this is a store, see if we can merge it in.
377 if (NextStore->isVolatile()) break;
379 // Check to see if this stored value is of the same byte-splattable value.
380 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
383 // Check to see if this store is to a constant offset from the start ptr.
385 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
388 Ranges.addStore(Offset, NextStore);
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.addStore(0, SI);
402 // Now that we have full information about ranges, loop over the ranges and
403 // emit memset's for anything big enough to be worthwhile.
404 bool MadeChange = false;
405 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
407 const MemsetRange &Range = *I;
409 if (Range.TheStores.size() == 1) continue;
411 // If it is profitable to lower this range to memset, do so now.
412 if (!Range.isProfitableToUseMemset(*TD))
415 // Otherwise, we do want to transform this! Create a new memset. We put
416 // the memset right before the first instruction that isn't part of this
417 // memset block. This ensure that the memset is dominated by any addressing
418 // instruction needed by the start of the block.
419 BasicBlock::iterator InsertPt = BI;
421 // Get the starting pointer of the block.
422 StartPtr = Range.StartPtr;
424 // Determine alignment
425 unsigned Alignment = Range.Alignment;
426 if (Alignment == 0) {
427 const Type *EltType =
428 cast<PointerType>(StartPtr->getType())->getElementType();
429 Alignment = TD->getABITypeAlignment(EltType);
432 IRBuilder<> Builder(InsertPt);
434 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
436 DEBUG(dbgs() << "Replace stores:\n";
437 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
438 dbgs() << *Range.TheStores[i] << '\n';
439 dbgs() << "With: " << *C << '\n'); (void)C;
441 // Don't invalidate the iterator
444 // Zap all the stores.
445 for (SmallVector<StoreInst*, 16>::const_iterator
446 SI = Range.TheStores.begin(),
447 SE = Range.TheStores.end(); SI != SE; ++SI)
448 (*SI)->eraseFromParent();
457 /// performCallSlotOptzn - takes a memcpy and a call that it depends on,
458 /// and checks for the possibility of a call slot optimization by having
459 /// the call write its result directly into the destination of the memcpy.
460 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
461 Value *cpyDest, Value *cpySrc,
462 uint64_t cpyLen, CallInst *C) {
463 // The general transformation to keep in mind is
465 // call @func(..., src, ...)
466 // memcpy(dest, src, ...)
470 // memcpy(dest, src, ...)
471 // call @func(..., dest, ...)
473 // Since moving the memcpy is technically awkward, we additionally check that
474 // src only holds uninitialized values at the moment of the call, meaning that
475 // the memcpy can be discarded rather than moved.
477 // Deliberately get the source and destination with bitcasts stripped away,
478 // because we'll need to do type comparisons based on the underlying type.
481 // Require that src be an alloca. This simplifies the reasoning considerably.
482 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
486 // Check that all of src is copied to dest.
487 TargetData *TD = getAnalysisIfAvailable<TargetData>();
488 if (!TD) return false;
490 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
494 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
495 srcArraySize->getZExtValue();
497 if (cpyLen < srcSize)
500 // Check that accessing the first srcSize bytes of dest will not cause a
501 // trap. Otherwise the transform is invalid since it might cause a trap
502 // to occur earlier than it otherwise would.
503 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
504 // The destination is an alloca. Check it is larger than srcSize.
505 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
509 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
510 destArraySize->getZExtValue();
512 if (destSize < srcSize)
514 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
515 // If the destination is an sret parameter then only accesses that are
516 // outside of the returned struct type can trap.
517 if (!A->hasStructRetAttr())
520 const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
521 uint64_t destSize = TD->getTypeAllocSize(StructTy);
523 if (destSize < srcSize)
529 // Check that src is not accessed except via the call and the memcpy. This
530 // guarantees that it holds only undefined values when passed in (so the final
531 // memcpy can be dropped), that it is not read or written between the call and
532 // the memcpy, and that writing beyond the end of it is undefined.
533 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
534 srcAlloca->use_end());
535 while (!srcUseList.empty()) {
536 User *UI = srcUseList.pop_back_val();
538 if (isa<BitCastInst>(UI)) {
539 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
541 srcUseList.push_back(*I);
542 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
543 if (G->hasAllZeroIndices())
544 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
546 srcUseList.push_back(*I);
549 } else if (UI != C && UI != cpy) {
554 // Since we're changing the parameter to the callsite, we need to make sure
555 // that what would be the new parameter dominates the callsite.
556 DominatorTree &DT = getAnalysis<DominatorTree>();
557 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
558 if (!DT.dominates(cpyDestInst, C))
561 // In addition to knowing that the call does not access src in some
562 // unexpected manner, for example via a global, which we deduce from
563 // the use analysis, we also need to know that it does not sneakily
564 // access dest. We rely on AA to figure this out for us.
565 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
566 if (AA.getModRefInfo(C, cpyDest, srcSize) !=
567 AliasAnalysis::NoModRef)
570 // All the checks have passed, so do the transformation.
571 bool changedArgument = false;
572 for (unsigned i = 0; i < CS.arg_size(); ++i)
573 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
574 if (cpySrc->getType() != cpyDest->getType())
575 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
576 cpyDest->getName(), C);
577 changedArgument = true;
578 if (CS.getArgument(i)->getType() == cpyDest->getType())
579 CS.setArgument(i, cpyDest);
581 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
582 CS.getArgument(i)->getType(), cpyDest->getName(), C));
585 if (!changedArgument)
588 // Drop any cached information about the call, because we may have changed
589 // its dependence information by changing its parameter.
590 MD->removeInstruction(C);
592 // Remove the memcpy.
593 MD->removeInstruction(cpy);
599 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)
600 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
601 /// copy from MDep's input if we can. MSize is the size of M's copy.
603 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
605 // We can only transforms memcpy's where the dest of one is the source of the
607 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
610 // If dep instruction is reading from our current input, then it is a noop
611 // transfer and substituting the input won't change this instruction. Just
612 // ignore the input and let someone else zap MDep. This handles cases like:
615 if (M->getSource() == MDep->getSource())
618 // Second, the length of the memcpy's must be the same, or the preceeding one
619 // must be larger than the following one.
620 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
621 if (!C1) return false;
623 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
625 // Verify that the copied-from memory doesn't change in between the two
626 // transfers. For example, in:
630 // It would be invalid to transform the second memcpy into memcpy(c <- b).
632 // TODO: If the code between M and MDep is transparent to the destination "c",
633 // then we could still perform the xform by moving M up to the first memcpy.
635 // NOTE: This is conservative, it will stop on any read from the source loc,
636 // not just the defining memcpy.
637 MemDepResult SourceDep =
638 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
639 false, M, M->getParent());
640 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
643 // If the dest of the second might alias the source of the first, then the
644 // source and dest might overlap. We still want to eliminate the intermediate
645 // value, but we have to generate a memmove instead of memcpy.
646 bool UseMemMove = false;
647 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)))
650 // If all checks passed, then we can transform M.
652 // Make sure to use the lesser of the alignment of the source and the dest
653 // since we're changing where we're reading from, but don't want to increase
654 // the alignment past what can be read from or written to.
655 // TODO: Is this worth it if we're creating a less aligned memcpy? For
656 // example we could be moving from movaps -> movq on x86.
657 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
659 IRBuilder<> Builder(M);
661 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
662 Align, M->isVolatile());
664 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
665 Align, M->isVolatile());
667 // Remove the instruction we're replacing.
668 MD->removeInstruction(M);
669 M->eraseFromParent();
675 /// processMemCpy - perform simplification of memcpy's. If we have memcpy A
676 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
677 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
678 /// circumstances). This allows later passes to remove the first memcpy
680 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
681 // We can only optimize statically-sized memcpy's that are non-volatile.
682 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
683 if (CopySize == 0 || M->isVolatile()) return false;
685 // If the source and destination of the memcpy are the same, then zap it.
686 if (M->getSource() == M->getDest()) {
687 MD->removeInstruction(M);
688 M->eraseFromParent();
692 // If copying from a constant, try to turn the memcpy into a memset.
693 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
694 if (GV->isConstant() && GV->hasDefinitiveInitializer())
695 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
696 IRBuilder<> Builder(M);
697 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize,
698 M->getAlignment(), false);
699 MD->removeInstruction(M);
700 M->eraseFromParent();
705 // The are two possible optimizations we can do for memcpy:
706 // a) memcpy-memcpy xform which exposes redundance for DSE.
707 // b) call-memcpy xform for return slot optimization.
708 MemDepResult DepInfo = MD->getDependency(M);
709 if (!DepInfo.isClobber())
712 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
713 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
715 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
716 if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
717 CopySize->getZExtValue(), C)) {
718 M->eraseFromParent();
725 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
726 /// are guaranteed not to alias.
727 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
728 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
730 // See if the pointers alias.
731 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M)))
734 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
736 // If not, then we know we can transform this.
737 Module *Mod = M->getParent()->getParent()->getParent();
738 const Type *ArgTys[3] = { M->getRawDest()->getType(),
739 M->getRawSource()->getType(),
740 M->getLength()->getType() };
741 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
744 // MemDep may have over conservative information about this instruction, just
745 // conservatively flush it from the cache.
746 MD->removeInstruction(M);
752 /// processByValArgument - This is called on every byval argument in call sites.
753 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
754 TargetData *TD = getAnalysisIfAvailable<TargetData>();
755 if (!TD) return false;
757 // Find out what feeds this byval argument.
758 Value *ByValArg = CS.getArgument(ArgNo);
759 const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType();
760 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
761 MemDepResult DepInfo =
762 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
763 true, CS.getInstruction(),
764 CS.getInstruction()->getParent());
765 if (!DepInfo.isClobber())
768 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
769 // a memcpy, see if we can byval from the source of the memcpy instead of the
771 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
772 if (MDep == 0 || MDep->isVolatile() ||
773 ByValArg->stripPointerCasts() != MDep->getDest())
776 // The length of the memcpy must be larger or equal to the size of the byval.
777 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
778 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
781 // Get the alignment of the byval. If it is greater than the memcpy, then we
782 // can't do the substitution. If the call doesn't specify the alignment, then
783 // it is some target specific value that we can't know.
784 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
785 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign)
788 // Verify that the copied-from memory doesn't change in between the memcpy and
793 // It would be invalid to transform the second memcpy into foo(*b).
795 // NOTE: This is conservative, it will stop on any read from the source loc,
796 // not just the defining memcpy.
797 MemDepResult SourceDep =
798 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
799 false, CS.getInstruction(), MDep->getParent());
800 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
803 Value *TmpCast = MDep->getSource();
804 if (MDep->getSource()->getType() != ByValArg->getType())
805 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
806 "tmpcast", CS.getInstruction());
808 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
809 << " " << *MDep << "\n"
810 << " " << *CS.getInstruction() << "\n");
812 // Otherwise we're good! Update the byval argument.
813 CS.setArgument(ArgNo, TmpCast);
818 /// iterateOnFunction - Executes one iteration of MemCpyOpt.
819 bool MemCpyOpt::iterateOnFunction(Function &F) {
820 bool MadeChange = false;
822 // Walk all instruction in the function.
823 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
824 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
825 // Avoid invalidating the iterator.
826 Instruction *I = BI++;
828 bool RepeatInstruction = false;
830 if (StoreInst *SI = dyn_cast<StoreInst>(I))
831 MadeChange |= processStore(SI, BI);
832 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) {
833 RepeatInstruction = processMemCpy(M);
834 } else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) {
835 RepeatInstruction = processMemMove(M);
836 } else if (CallSite CS = (Value*)I) {
837 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
838 if (CS.paramHasAttr(i+1, Attribute::ByVal))
839 MadeChange |= processByValArgument(CS, i);
842 // Reprocess the instruction if desired.
843 if (RepeatInstruction) {
853 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
856 bool MemCpyOpt::runOnFunction(Function &F) {
857 bool MadeChange = false;
858 MD = &getAnalysis<MemoryDependenceAnalysis>();
860 if (!iterateOnFunction(F))