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/IntrinsicInst.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/ParameterAttributes.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/Support/Debug.h"
26 #include "llvm/Support/GetElementPtrTypeIterator.h"
27 #include "llvm/Target/TargetData.h"
31 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
32 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
34 /// isBytewiseValue - If the specified value can be set by repeating the same
35 /// byte in memory, return the i8 value that it is represented with. This is
36 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
37 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
38 /// byte store (e.g. i16 0x1234), return null.
39 static Value *isBytewiseValue(Value *V) {
40 // All byte-wide stores are splatable, even of arbitrary variables.
41 if (V->getType() == Type::Int8Ty) return V;
43 // Constant float and double values can be handled as integer values if the
44 // corresponding integer value is "byteable". An important case is 0.0.
45 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
46 if (CFP->getType() == Type::FloatTy)
47 V = ConstantExpr::getBitCast(CFP, Type::Int32Ty);
48 if (CFP->getType() == Type::DoubleTy)
49 V = ConstantExpr::getBitCast(CFP, Type::Int64Ty);
50 // Don't handle long double formats, which have strange constraints.
53 // We can handle constant integers that are power of two in size and a
54 // multiple of 8 bits.
55 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
56 unsigned Width = CI->getBitWidth();
57 if (isPowerOf2_32(Width) && Width > 8) {
58 // We can handle this value if the recursive binary decomposition is the
59 // same at all levels.
60 APInt Val = CI->getValue();
62 while (Val.getBitWidth() != 8) {
63 unsigned NextWidth = Val.getBitWidth()/2;
64 Val2 = Val.lshr(NextWidth);
65 Val2.trunc(Val.getBitWidth()/2);
66 Val.trunc(Val.getBitWidth()/2);
68 // If the top/bottom halves aren't the same, reject it.
72 return ConstantInt::get(Val);
76 // Conceptually, we could handle things like:
77 // %a = zext i8 %X to i16
80 // but until there is an example that actually needs this, it doesn't seem
81 // worth worrying about.
85 static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx,
86 bool &VariableIdxFound, TargetData &TD) {
87 // Skip over the first indices.
88 gep_type_iterator GTI = gep_type_begin(GEP);
89 for (unsigned i = 1; i != Idx; ++i, ++GTI)
92 // Compute the offset implied by the rest of the indices.
94 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
95 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
97 return VariableIdxFound = true;
98 if (OpC->isZero()) continue; // No offset.
100 // Handle struct indices, which add their field offset to the pointer.
101 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
102 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
106 // Otherwise, we have a sequential type like an array or vector. Multiply
107 // the index by the ElementSize.
108 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
109 Offset += Size*OpC->getSExtValue();
115 /// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a
116 /// constant offset, and return that constant offset. For example, Ptr1 might
117 /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8.
118 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
120 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
121 // base. After that base, they may have some number of common (and
122 // potentially variable) indices. After that they handle some constant
123 // offset, which determines their offset from each other. At this point, we
124 // handle no other case.
125 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
126 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
127 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
130 // Skip any common indices and track the GEP types.
132 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
133 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
136 bool VariableIdxFound = false;
137 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
138 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
139 if (VariableIdxFound) return false;
141 Offset = Offset2-Offset1;
146 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
147 /// This allows us to analyze stores like:
152 /// which sometimes happens with stores to arrays of structs etc. When we see
153 /// the first store, we make a range [1, 2). The second store extends the range
154 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
155 /// two ranges into [0, 3) which is memset'able.
158 // Start/End - A semi range that describes the span that this range covers.
159 // The range is closed at the start and open at the end: [Start, End).
162 /// StartPtr - The getelementptr instruction that points to the start of the
166 /// Alignment - The known alignment of the first store.
169 /// TheStores - The actual stores that make up this range.
170 SmallVector<StoreInst*, 16> TheStores;
172 bool isProfitableToUseMemset(const TargetData &TD) const;
175 } // end anon namespace
177 bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const {
178 // If we found more than 8 stores to merge or 64 bytes, use memset.
179 if (TheStores.size() >= 8 || End-Start >= 64) return true;
181 // Assume that the code generator is capable of merging pairs of stores
182 // together if it wants to.
183 if (TheStores.size() <= 2) return false;
185 // If we have fewer than 8 stores, it can still be worthwhile to do this.
186 // For example, merging 4 i8 stores into an i32 store is useful almost always.
187 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
188 // memset will be split into 2 32-bit stores anyway) and doing so can
189 // pessimize the llvm optimizer.
191 // Since we don't have perfect knowledge here, make some assumptions: assume
192 // the maximum GPR width is the same size as the pointer size and assume that
193 // this width can be stored. If so, check to see whether we will end up
194 // actually reducing the number of stores used.
195 unsigned Bytes = unsigned(End-Start);
196 unsigned NumPointerStores = Bytes/TD.getPointerSize();
198 // Assume the remaining bytes if any are done a byte at a time.
199 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
201 // If we will reduce the # stores (according to this heuristic), do the
202 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
204 return TheStores.size() > NumPointerStores+NumByteStores;
210 /// Ranges - A sorted list of the memset ranges. We use std::list here
211 /// because each element is relatively large and expensive to copy.
212 std::list<MemsetRange> Ranges;
213 typedef std::list<MemsetRange>::iterator range_iterator;
216 MemsetRanges(TargetData &td) : TD(td) {}
218 typedef std::list<MemsetRange>::const_iterator const_iterator;
219 const_iterator begin() const { return Ranges.begin(); }
220 const_iterator end() const { return Ranges.end(); }
221 bool empty() const { return Ranges.empty(); }
223 void addStore(int64_t OffsetFromFirst, StoreInst *SI);
226 } // end anon namespace
229 /// addStore - Add a new store to the MemsetRanges data structure. This adds a
230 /// new range for the specified store at the specified offset, merging into
231 /// existing ranges as appropriate.
232 void MemsetRanges::addStore(int64_t Start, StoreInst *SI) {
233 int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType());
235 // Do a linear search of the ranges to see if this can be joined and/or to
236 // find the insertion point in the list. We keep the ranges sorted for
237 // simplicity here. This is a linear search of a linked list, which is ugly,
238 // however the number of ranges is limited, so this won't get crazy slow.
239 range_iterator I = Ranges.begin(), E = Ranges.end();
241 while (I != E && Start > I->End)
244 // We now know that I == E, in which case we didn't find anything to merge
245 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
246 // to insert a new range. Handle this now.
247 if (I == E || End < I->Start) {
248 MemsetRange &R = *Ranges.insert(I, MemsetRange());
251 R.StartPtr = SI->getPointerOperand();
252 R.Alignment = SI->getAlignment();
253 R.TheStores.push_back(SI);
257 // This store overlaps with I, add it.
258 I->TheStores.push_back(SI);
260 // At this point, we may have an interval that completely contains our store.
261 // If so, just add it to the interval and return.
262 if (I->Start <= Start && I->End >= End)
265 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
266 // but is not entirely contained within the range.
268 // See if the range extends the start of the range. In this case, it couldn't
269 // possibly cause it to join the prior range, because otherwise we would have
271 if (Start < I->Start) {
273 I->StartPtr = SI->getPointerOperand();
276 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
277 // is in or right at the end of I), and that End >= I->Start. Extend I out to
281 range_iterator NextI = I;;
282 while (++NextI != E && End >= NextI->Start) {
283 // Merge the range in.
284 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
285 if (NextI->End > I->End)
293 //===----------------------------------------------------------------------===//
295 //===----------------------------------------------------------------------===//
299 class VISIBILITY_HIDDEN MemCpyOpt : public FunctionPass {
300 bool runOnFunction(Function &F);
302 static char ID; // Pass identification, replacement for typeid
303 MemCpyOpt() : FunctionPass((intptr_t)&ID) { }
306 // This transformation requires dominator postdominator info
307 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
308 AU.setPreservesCFG();
309 AU.addRequired<DominatorTree>();
310 AU.addRequired<MemoryDependenceAnalysis>();
311 AU.addRequired<AliasAnalysis>();
312 AU.addRequired<TargetData>();
313 AU.addPreserved<AliasAnalysis>();
314 AU.addPreserved<MemoryDependenceAnalysis>();
315 AU.addPreserved<TargetData>();
319 bool processStore(StoreInst *SI, BasicBlock::iterator& BBI);
320 bool processMemCpy(MemCpyInst* M);
321 bool performCallSlotOptzn(MemCpyInst* cpy, CallInst* C);
322 bool iterateOnFunction(Function &F);
325 char MemCpyOpt::ID = 0;
328 // createMemCpyOptPass - The public interface to this file...
329 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
331 static RegisterPass<MemCpyOpt> X("memcpyopt",
332 "MemCpy Optimization");
336 /// processStore - When GVN is 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
339 /// (currently 4) it attempts to merge them together into a memcpy/memset.
340 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator& BBI) {
341 if (SI->isVolatile()) return false;
343 // There are two cases that are interesting for this code to handle: memcpy
344 // and memset. Right now we only handle memset.
346 // Ensure that the value being stored is something that can be memset'able a
347 // byte at a time like "0" or "-1" or any width, as well as things like
348 // 0xA0A0A0A0 and 0.0.
349 Value *ByteVal = isBytewiseValue(SI->getOperand(0));
353 TargetData &TD = getAnalysis<TargetData>();
354 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
356 // Okay, so we now have a single store that can be splatable. Scan to find
357 // all subsequent stores of the same value to offset from the same pointer.
358 // Join these together into ranges, so we can decide whether contiguous blocks
360 MemsetRanges Ranges(TD);
362 Value *StartPtr = SI->getPointerOperand();
364 BasicBlock::iterator BI = SI;
365 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
366 if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
367 // If the call is readnone, ignore it, otherwise bail out. We don't even
368 // allow readonly here because we don't want something like:
369 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
370 if (AA.getModRefBehavior(CallSite::get(BI)) ==
371 AliasAnalysis::DoesNotAccessMemory)
374 // TODO: If this is a memset, try to join it in.
377 } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI))
380 // If this is a non-store instruction it is fine, ignore it.
381 StoreInst *NextStore = dyn_cast<StoreInst>(BI);
382 if (NextStore == 0) continue;
384 // If this is a store, see if we can merge it in.
385 if (NextStore->isVolatile()) break;
387 // Check to see if this stored value is of the same byte-splattable value.
388 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
391 // Check to see if this store is to a constant offset from the start ptr.
393 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, TD))
396 Ranges.addStore(Offset, NextStore);
399 // If we have no ranges, then we just had a single store with nothing that
400 // could be merged in. This is a very common case of course.
404 // If we had at least one store that could be merged in, add the starting
405 // store as well. We try to avoid this unless there is at least something
406 // interesting as a small compile-time optimization.
407 Ranges.addStore(0, SI);
410 Function *MemSetF = 0;
412 // Now that we have full information about ranges, loop over the ranges and
413 // emit memset's for anything big enough to be worthwhile.
414 bool MadeChange = false;
415 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
417 const MemsetRange &Range = *I;
419 if (Range.TheStores.size() == 1) continue;
421 // If it is profitable to lower this range to memset, do so now.
422 if (!Range.isProfitableToUseMemset(TD))
425 // Otherwise, we do want to transform this! Create a new memset. We put
426 // the memset right before the first instruction that isn't part of this
427 // memset block. This ensure that the memset is dominated by any addressing
428 // instruction needed by the start of the block.
429 BasicBlock::iterator InsertPt = BI;
432 MemSetF = Intrinsic::getDeclaration(SI->getParent()->getParent()
433 ->getParent(), Intrinsic::memset_i64);
435 // Get the starting pointer of the block.
436 StartPtr = Range.StartPtr;
438 // Cast the start ptr to be i8* as memset requires.
439 const Type *i8Ptr = PointerType::getUnqual(Type::Int8Ty);
440 if (StartPtr->getType() != i8Ptr)
441 StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getNameStart(),
445 StartPtr, ByteVal, // Start, value
446 ConstantInt::get(Type::Int64Ty, Range.End-Range.Start), // size
447 ConstantInt::get(Type::Int32Ty, Range.Alignment) // align
449 Value *C = CallInst::Create(MemSetF, Ops, Ops+4, "", InsertPt);
450 DEBUG(cerr << "Replace stores:\n";
451 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
452 cerr << *Range.TheStores[i];
453 cerr << "With: " << *C); C=C;
455 // Don't invalidate the iterator
458 // Zap all the stores.
459 for (SmallVector<StoreInst*, 16>::const_iterator SI = Range.TheStores.begin(),
460 SE = Range.TheStores.end(); SI != SE; ++SI)
461 (*SI)->eraseFromParent();
470 /// performCallSlotOptzn - takes a memcpy and a call that it depends on,
471 /// and checks for the possibility of a call slot optimization by having
472 /// the call write its result directly into the destination of the memcpy.
473 bool MemCpyOpt::performCallSlotOptzn(MemCpyInst *cpy, CallInst *C) {
474 // The general transformation to keep in mind is
476 // call @func(..., src, ...)
477 // memcpy(dest, src, ...)
481 // memcpy(dest, src, ...)
482 // call @func(..., dest, ...)
484 // Since moving the memcpy is technically awkward, we additionally check that
485 // src only holds uninitialized values at the moment of the call, meaning that
486 // the memcpy can be discarded rather than moved.
488 // Deliberately get the source and destination with bitcasts stripped away,
489 // because we'll need to do type comparisons based on the underlying type.
490 Value* cpyDest = cpy->getDest();
491 Value* cpySrc = cpy->getSource();
492 CallSite CS = CallSite::get(C);
494 // We need to be able to reason about the size of the memcpy, so we require
495 // that it be a constant.
496 ConstantInt* cpyLength = dyn_cast<ConstantInt>(cpy->getLength());
500 // Require that src be an alloca. This simplifies the reasoning considerably.
501 AllocaInst* srcAlloca = dyn_cast<AllocaInst>(cpySrc);
505 // Check that all of src is copied to dest.
506 TargetData& TD = getAnalysis<TargetData>();
508 ConstantInt* srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
512 uint64_t srcSize = TD.getABITypeSize(srcAlloca->getAllocatedType()) *
513 srcArraySize->getZExtValue();
515 if (cpyLength->getZExtValue() < srcSize)
518 // Check that accessing the first srcSize bytes of dest will not cause a
519 // trap. Otherwise the transform is invalid since it might cause a trap
520 // to occur earlier than it otherwise would.
521 if (AllocaInst* A = dyn_cast<AllocaInst>(cpyDest)) {
522 // The destination is an alloca. Check it is larger than srcSize.
523 ConstantInt* destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
527 uint64_t destSize = TD.getABITypeSize(A->getAllocatedType()) *
528 destArraySize->getZExtValue();
530 if (destSize < srcSize)
532 } else if (Argument* A = dyn_cast<Argument>(cpyDest)) {
533 // If the destination is an sret parameter then only accesses that are
534 // outside of the returned struct type can trap.
535 if (!A->hasStructRetAttr())
538 const Type* StructTy = cast<PointerType>(A->getType())->getElementType();
539 uint64_t destSize = TD.getABITypeSize(StructTy);
541 if (destSize < srcSize)
547 // Check that src is not accessed except via the call and the memcpy. This
548 // guarantees that it holds only undefined values when passed in (so the final
549 // memcpy can be dropped), that it is not read or written between the call and
550 // the memcpy, and that writing beyond the end of it is undefined.
551 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
552 srcAlloca->use_end());
553 while (!srcUseList.empty()) {
554 User* UI = srcUseList.back();
555 srcUseList.pop_back();
557 if (isa<BitCastInst>(UI)) {
558 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
560 srcUseList.push_back(*I);
561 } else if (GetElementPtrInst* G = dyn_cast<GetElementPtrInst>(UI)) {
562 if (G->hasAllZeroIndices())
563 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
565 srcUseList.push_back(*I);
568 } else if (UI != C && UI != cpy) {
573 // Since we're changing the parameter to the callsite, we need to make sure
574 // that what would be the new parameter dominates the callsite.
575 DominatorTree& DT = getAnalysis<DominatorTree>();
576 if (Instruction* cpyDestInst = dyn_cast<Instruction>(cpyDest))
577 if (!DT.dominates(cpyDestInst, C))
580 // In addition to knowing that the call does not access src in some
581 // unexpected manner, for example via a global, which we deduce from
582 // the use analysis, we also need to know that it does not sneakily
583 // access dest. We rely on AA to figure this out for us.
584 AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
585 if (AA.getModRefInfo(C, cpy->getRawDest(), srcSize) !=
586 AliasAnalysis::NoModRef)
589 // All the checks have passed, so do the transformation.
590 bool changedArgument = false;
591 for (unsigned i = 0; i < CS.arg_size(); ++i)
592 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
593 if (cpySrc->getType() != cpyDest->getType())
594 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
595 cpyDest->getName(), C);
596 changedArgument = true;
597 if (CS.getArgument(i)->getType() != cpyDest->getType())
598 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
599 CS.getArgument(i)->getType(), cpyDest->getName(), C));
601 CS.setArgument(i, cpyDest);
604 if (!changedArgument)
607 // Drop any cached information about the call, because we may have changed
608 // its dependence information by changing its parameter.
609 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
610 MD.dropInstruction(C);
613 MD.removeInstruction(cpy);
614 cpy->eraseFromParent();
620 /// processMemCpy - perform simplication of memcpy's. If we have memcpy A which
621 /// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be
622 /// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
623 /// This allows later passes to remove the first memcpy altogether.
624 bool MemCpyOpt::processMemCpy(MemCpyInst* M) {
625 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
627 // The are two possible optimizations we can do for memcpy:
628 // a) memcpy-memcpy xform which exposes redundance for DSE
629 // b) call-memcpy xform for return slot optimization
630 Instruction* dep = MD.getDependency(M);
631 if (dep == MemoryDependenceAnalysis::None ||
632 dep == MemoryDependenceAnalysis::NonLocal)
634 else if (!isa<MemCpyInst>(dep)) {
635 if (CallInst* C = dyn_cast<CallInst>(dep))
636 return performCallSlotOptzn(M, C);
641 MemCpyInst* MDep = cast<MemCpyInst>(dep);
643 // We can only transforms memcpy's where the dest of one is the source of the
645 if (M->getSource() != MDep->getDest())
648 // Second, the length of the memcpy's must be the same, or the preceeding one
649 // must be larger than the following one.
650 ConstantInt* C1 = dyn_cast<ConstantInt>(MDep->getLength());
651 ConstantInt* C2 = dyn_cast<ConstantInt>(M->getLength());
655 uint64_t DepSize = C1->getValue().getZExtValue();
656 uint64_t CpySize = C2->getValue().getZExtValue();
658 if (DepSize < CpySize)
661 // Finally, we have to make sure that the dest of the second does not
662 // alias the source of the first
663 AliasAnalysis& AA = getAnalysis<AliasAnalysis>();
664 if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
665 AliasAnalysis::NoAlias)
667 else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) !=
668 AliasAnalysis::NoAlias)
670 else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize)
671 != AliasAnalysis::NoAlias)
674 // If all checks passed, then we can transform these memcpy's
675 Function* MemCpyFun = Intrinsic::getDeclaration(
676 M->getParent()->getParent()->getParent(),
677 M->getIntrinsicID());
679 std::vector<Value*> args;
680 args.push_back(M->getRawDest());
681 args.push_back(MDep->getRawSource());
682 args.push_back(M->getLength());
683 args.push_back(M->getAlignment());
685 CallInst* C = CallInst::Create(MemCpyFun, args.begin(), args.end(), "", M);
688 // If C and M don't interfere, then this is a valid transformation. If they
689 // did, this would mean that the two sources overlap, which would be bad.
690 if (MD.getDependency(C) == MDep) {
691 MD.dropInstruction(M);
692 M->eraseFromParent();
699 // Otherwise, there was no point in doing this, so we remove the call we
700 // inserted and act like nothing happened.
701 MD.removeInstruction(C);
702 C->eraseFromParent();
707 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
710 bool MemCpyOpt::runOnFunction(Function& F) {
712 bool changed = false;
713 bool shouldContinue = true;
715 while (shouldContinue) {
716 shouldContinue = iterateOnFunction(F);
717 changed |= shouldContinue;
724 // MemCpyOpt::iterateOnFunction - Executes one iteration of GVN
725 bool MemCpyOpt::iterateOnFunction(Function &F) {
726 bool changed_function = false;
728 // Walk all instruction in the function
729 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
730 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
732 // Avoid invalidating the iterator
733 Instruction* I = BI++;
735 if (StoreInst *SI = dyn_cast<StoreInst>(I))
736 changed_function |= processStore(SI, BI);
737 else if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) {
738 changed_function |= processMemCpy(M);
743 return changed_function;