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/LLVMContext.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/Support/raw_ostream.h"
28 #include "llvm/Target/TargetData.h"
32 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
33 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
34 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
36 /// isBytewiseValue - If the specified value can be set by repeating the same
37 /// byte in memory, return the i8 value that it is represented with. This is
38 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
39 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
40 /// byte store (e.g. i16 0x1234), return null.
41 static Value *isBytewiseValue(Value *V) {
42 LLVMContext &Context = V->getContext();
44 // All byte-wide stores are splatable, even of arbitrary variables.
45 if (V->getType()->isIntegerTy(8)) return V;
47 // Constant float and double values can be handled as integer values if the
48 // corresponding integer value is "byteable". An important case is 0.0.
49 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
50 if (CFP->getType()->isFloatTy())
51 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(Context));
52 if (CFP->getType()->isDoubleTy())
53 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(Context));
54 // Don't handle long double formats, which have strange constraints.
57 // We can handle constant integers that are power of two in size and a
58 // multiple of 8 bits.
59 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
60 unsigned Width = CI->getBitWidth();
61 if (isPowerOf2_32(Width) && Width > 8) {
62 // We can handle this value if the recursive binary decomposition is the
63 // same at all levels.
64 APInt Val = CI->getValue();
66 while (Val.getBitWidth() != 8) {
67 unsigned NextWidth = Val.getBitWidth()/2;
68 Val2 = Val.lshr(NextWidth);
69 Val2.trunc(Val.getBitWidth()/2);
70 Val.trunc(Val.getBitWidth()/2);
72 // If the top/bottom halves aren't the same, reject it.
76 return ConstantInt::get(Context, Val);
80 // Conceptually, we could handle things like:
81 // %a = zext i8 %X to i16
84 // but until there is an example that actually needs this, it doesn't seem
85 // worth worrying about.
89 static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx,
90 bool &VariableIdxFound, TargetData &TD) {
91 // Skip over the first indices.
92 gep_type_iterator GTI = gep_type_begin(GEP);
93 for (unsigned i = 1; i != Idx; ++i, ++GTI)
96 // Compute the offset implied by the rest of the indices.
98 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
99 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
101 return VariableIdxFound = true;
102 if (OpC->isZero()) continue; // No offset.
104 // Handle struct indices, which add their field offset to the pointer.
105 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
106 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
110 // Otherwise, we have a sequential type like an array or vector. Multiply
111 // the index by the ElementSize.
112 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
113 Offset += Size*OpC->getSExtValue();
119 /// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a
120 /// constant offset, and return that constant offset. For example, Ptr1 might
121 /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8.
122 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
124 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
125 // base. After that base, they may have some number of common (and
126 // potentially variable) indices. After that they handle some constant
127 // offset, which determines their offset from each other. At this point, we
128 // handle no other case.
129 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
130 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
131 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
134 // Skip any common indices and track the GEP types.
136 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
137 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
140 bool VariableIdxFound = false;
141 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
142 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
143 if (VariableIdxFound) return false;
145 Offset = Offset2-Offset1;
150 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
151 /// This allows us to analyze stores like:
156 /// which sometimes happens with stores to arrays of structs etc. When we see
157 /// the first store, we make a range [1, 2). The second store extends the range
158 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
159 /// two ranges into [0, 3) which is memset'able.
162 // Start/End - A semi range that describes the span that this range covers.
163 // The range is closed at the start and open at the end: [Start, End).
166 /// StartPtr - The getelementptr instruction that points to the start of the
170 /// Alignment - The known alignment of the first store.
173 /// TheStores - The actual stores that make up this range.
174 SmallVector<StoreInst*, 16> TheStores;
176 bool isProfitableToUseMemset(const TargetData &TD) const;
179 } // end anon namespace
181 bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const {
182 // If we found more than 8 stores to merge or 64 bytes, use memset.
183 if (TheStores.size() >= 8 || End-Start >= 64) return true;
185 // Assume that the code generator is capable of merging pairs of stores
186 // together if it wants to.
187 if (TheStores.size() <= 2) return false;
189 // If we have fewer than 8 stores, it can still be worthwhile to do this.
190 // For example, merging 4 i8 stores into an i32 store is useful almost always.
191 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
192 // memset will be split into 2 32-bit stores anyway) and doing so can
193 // pessimize the llvm optimizer.
195 // Since we don't have perfect knowledge here, make some assumptions: assume
196 // the maximum GPR width is the same size as the pointer size and assume that
197 // this width can be stored. If so, check to see whether we will end up
198 // actually reducing the number of stores used.
199 unsigned Bytes = unsigned(End-Start);
200 unsigned NumPointerStores = Bytes/TD.getPointerSize();
202 // Assume the remaining bytes if any are done a byte at a time.
203 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
205 // If we will reduce the # stores (according to this heuristic), do the
206 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
208 return TheStores.size() > NumPointerStores+NumByteStores;
214 /// Ranges - A sorted list of the memset ranges. We use std::list here
215 /// because each element is relatively large and expensive to copy.
216 std::list<MemsetRange> Ranges;
217 typedef std::list<MemsetRange>::iterator range_iterator;
220 MemsetRanges(TargetData &td) : TD(td) {}
222 typedef std::list<MemsetRange>::const_iterator const_iterator;
223 const_iterator begin() const { return Ranges.begin(); }
224 const_iterator end() const { return Ranges.end(); }
225 bool empty() const { return Ranges.empty(); }
227 void addStore(int64_t OffsetFromFirst, StoreInst *SI);
230 } // end anon namespace
233 /// addStore - Add a new store to the MemsetRanges data structure. This adds a
234 /// new range for the specified store at the specified offset, merging into
235 /// existing ranges as appropriate.
236 void MemsetRanges::addStore(int64_t Start, StoreInst *SI) {
237 int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType());
239 // Do a linear search of the ranges to see if this can be joined and/or to
240 // find the insertion point in the list. We keep the ranges sorted for
241 // simplicity here. This is a linear search of a linked list, which is ugly,
242 // however the number of ranges is limited, so this won't get crazy slow.
243 range_iterator I = Ranges.begin(), E = Ranges.end();
245 while (I != E && Start > I->End)
248 // We now know that I == E, in which case we didn't find anything to merge
249 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
250 // to insert a new range. Handle this now.
251 if (I == E || End < I->Start) {
252 MemsetRange &R = *Ranges.insert(I, MemsetRange());
255 R.StartPtr = SI->getPointerOperand();
256 R.Alignment = SI->getAlignment();
257 R.TheStores.push_back(SI);
261 // This store overlaps with I, add it.
262 I->TheStores.push_back(SI);
264 // At this point, we may have an interval that completely contains our store.
265 // If so, just add it to the interval and return.
266 if (I->Start <= Start && I->End >= End)
269 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
270 // but is not entirely contained within the range.
272 // See if the range extends the start of the range. In this case, it couldn't
273 // possibly cause it to join the prior range, because otherwise we would have
275 if (Start < I->Start) {
277 I->StartPtr = SI->getPointerOperand();
278 I->Alignment = SI->getAlignment();
281 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
282 // is in or right at the end of I), and that End >= I->Start. Extend I out to
286 range_iterator NextI = I;
287 while (++NextI != E && End >= NextI->Start) {
288 // Merge the range in.
289 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
290 if (NextI->End > I->End)
298 //===----------------------------------------------------------------------===//
300 //===----------------------------------------------------------------------===//
303 class MemCpyOpt : public FunctionPass {
304 bool runOnFunction(Function &F);
306 static char ID; // Pass identification, replacement for typeid
307 MemCpyOpt() : FunctionPass(ID) {
308 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
312 // This transformation requires dominator postdominator info
313 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
314 AU.setPreservesCFG();
315 AU.addRequired<DominatorTree>();
316 AU.addRequired<MemoryDependenceAnalysis>();
317 AU.addRequired<AliasAnalysis>();
318 AU.addPreserved<AliasAnalysis>();
319 AU.addPreserved<MemoryDependenceAnalysis>();
323 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
324 bool processMemCpy(MemCpyInst *M);
325 bool processMemMove(MemMoveInst *M);
326 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
327 uint64_t cpyLen, CallInst *C);
328 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
330 bool iterateOnFunction(Function &F);
333 char MemCpyOpt::ID = 0;
336 // createMemCpyOptPass - The public interface to this file...
337 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
339 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
341 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
342 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
343 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
344 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
347 /// processStore - When GVN is scanning forward over instructions, we look for
348 /// some other patterns to fold away. In particular, this looks for stores to
349 /// neighboring locations of memory. If it sees enough consequtive ones
350 /// (currently 4) it attempts to merge them together into a memcpy/memset.
351 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
352 if (SI->isVolatile()) return false;
354 TargetData *TD = getAnalysisIfAvailable<TargetData>();
355 if (!TD) return false;
357 // Detect cases where we're performing call slot forwarding, but
358 // happen to be using a load-store pair to implement it, rather than
360 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
361 if (!LI->isVolatile() && LI->hasOneUse()) {
362 MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
364 MemDepResult dep = MD.getDependency(LI);
366 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
367 C = dyn_cast<CallInst>(dep.getInst());
370 bool changed = performCallSlotOptzn(LI,
371 SI->getPointerOperand()->stripPointerCasts(),
372 LI->getPointerOperand()->stripPointerCasts(),
373 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
375 MD.removeInstruction(SI);
376 SI->eraseFromParent();
377 LI->eraseFromParent();
385 LLVMContext &Context = SI->getContext();
387 // There are two cases that are interesting for this code to handle: memcpy
388 // and memset. Right now we only handle memset.
390 // Ensure that the value being stored is something that can be memset'able a
391 // byte at a time like "0" or "-1" or any width, as well as things like
392 // 0xA0A0A0A0 and 0.0.
393 Value *ByteVal = isBytewiseValue(SI->getOperand(0));
397 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
398 Module *M = SI->getParent()->getParent()->getParent();
400 // Okay, so we now have a single store that can be splatable. Scan to find
401 // all subsequent stores of the same value to offset from the same pointer.
402 // Join these together into ranges, so we can decide whether contiguous blocks
404 MemsetRanges Ranges(*TD);
406 Value *StartPtr = SI->getPointerOperand();
408 BasicBlock::iterator BI = SI;
409 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
410 if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
411 // If the call is readnone, ignore it, otherwise bail out. We don't even
412 // allow readonly here because we don't want something like:
413 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
414 if (AA.getModRefBehavior(CallSite(BI)) ==
415 AliasAnalysis::DoesNotAccessMemory)
418 // TODO: If this is a memset, try to join it in.
421 } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI))
424 // If this is a non-store instruction it is fine, ignore it.
425 StoreInst *NextStore = dyn_cast<StoreInst>(BI);
426 if (NextStore == 0) continue;
428 // If this is a store, see if we can merge it in.
429 if (NextStore->isVolatile()) break;
431 // Check to see if this stored value is of the same byte-splattable value.
432 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
435 // Check to see if this store is to a constant offset from the start ptr.
437 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
440 Ranges.addStore(Offset, NextStore);
443 // If we have no ranges, then we just had a single store with nothing that
444 // could be merged in. This is a very common case of course.
448 // If we had at least one store that could be merged in, add the starting
449 // store as well. We try to avoid this unless there is at least something
450 // interesting as a small compile-time optimization.
451 Ranges.addStore(0, SI);
454 // Now that we have full information about ranges, loop over the ranges and
455 // emit memset's for anything big enough to be worthwhile.
456 bool MadeChange = false;
457 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
459 const MemsetRange &Range = *I;
461 if (Range.TheStores.size() == 1) continue;
463 // If it is profitable to lower this range to memset, do so now.
464 if (!Range.isProfitableToUseMemset(*TD))
467 // Otherwise, we do want to transform this! Create a new memset. We put
468 // the memset right before the first instruction that isn't part of this
469 // memset block. This ensure that the memset is dominated by any addressing
470 // instruction needed by the start of the block.
471 BasicBlock::iterator InsertPt = BI;
473 // Get the starting pointer of the block.
474 StartPtr = Range.StartPtr;
476 // Determine alignment
477 unsigned Alignment = Range.Alignment;
478 if (Alignment == 0) {
479 const Type *EltType =
480 cast<PointerType>(StartPtr->getType())->getElementType();
481 Alignment = TD->getABITypeAlignment(EltType);
484 // Cast the start ptr to be i8* as memset requires.
485 const PointerType* StartPTy = cast<PointerType>(StartPtr->getType());
486 const PointerType *i8Ptr = Type::getInt8PtrTy(Context,
487 StartPTy->getAddressSpace());
488 if (StartPTy!= i8Ptr)
489 StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getName(),
493 StartPtr, ByteVal, // Start, value
495 ConstantInt::get(Type::getInt64Ty(Context), Range.End-Range.Start),
497 ConstantInt::get(Type::getInt32Ty(Context), Alignment),
499 ConstantInt::get(Type::getInt1Ty(Context), 0),
501 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
503 Function *MemSetF = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2);
505 Value *C = CallInst::Create(MemSetF, Ops, Ops+5, "", InsertPt);
506 DEBUG(dbgs() << "Replace stores:\n";
507 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
508 dbgs() << *Range.TheStores[i];
509 dbgs() << "With: " << *C); C=C;
511 // Don't invalidate the iterator
514 // Zap all the stores.
515 for (SmallVector<StoreInst*, 16>::const_iterator
516 SI = Range.TheStores.begin(),
517 SE = Range.TheStores.end(); SI != SE; ++SI)
518 (*SI)->eraseFromParent();
527 /// performCallSlotOptzn - takes a memcpy and a call that it depends on,
528 /// and checks for the possibility of a call slot optimization by having
529 /// the call write its result directly into the destination of the memcpy.
530 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
531 Value *cpyDest, Value *cpySrc,
532 uint64_t cpyLen, CallInst *C) {
533 // The general transformation to keep in mind is
535 // call @func(..., src, ...)
536 // memcpy(dest, src, ...)
540 // memcpy(dest, src, ...)
541 // call @func(..., dest, ...)
543 // Since moving the memcpy is technically awkward, we additionally check that
544 // src only holds uninitialized values at the moment of the call, meaning that
545 // the memcpy can be discarded rather than moved.
547 // Deliberately get the source and destination with bitcasts stripped away,
548 // because we'll need to do type comparisons based on the underlying type.
551 // Require that src be an alloca. This simplifies the reasoning considerably.
552 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
556 // Check that all of src is copied to dest.
557 TargetData *TD = getAnalysisIfAvailable<TargetData>();
558 if (!TD) return false;
560 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
564 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
565 srcArraySize->getZExtValue();
567 if (cpyLen < srcSize)
570 // Check that accessing the first srcSize bytes of dest will not cause a
571 // trap. Otherwise the transform is invalid since it might cause a trap
572 // to occur earlier than it otherwise would.
573 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
574 // The destination is an alloca. Check it is larger than srcSize.
575 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
579 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
580 destArraySize->getZExtValue();
582 if (destSize < srcSize)
584 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
585 // If the destination is an sret parameter then only accesses that are
586 // outside of the returned struct type can trap.
587 if (!A->hasStructRetAttr())
590 const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
591 uint64_t destSize = TD->getTypeAllocSize(StructTy);
593 if (destSize < srcSize)
599 // Check that src is not accessed except via the call and the memcpy. This
600 // guarantees that it holds only undefined values when passed in (so the final
601 // memcpy can be dropped), that it is not read or written between the call and
602 // the memcpy, and that writing beyond the end of it is undefined.
603 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
604 srcAlloca->use_end());
605 while (!srcUseList.empty()) {
606 User *UI = srcUseList.pop_back_val();
608 if (isa<BitCastInst>(UI)) {
609 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
611 srcUseList.push_back(*I);
612 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
613 if (G->hasAllZeroIndices())
614 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
616 srcUseList.push_back(*I);
619 } else if (UI != C && UI != cpy) {
624 // Since we're changing the parameter to the callsite, we need to make sure
625 // that what would be the new parameter dominates the callsite.
626 DominatorTree &DT = getAnalysis<DominatorTree>();
627 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
628 if (!DT.dominates(cpyDestInst, C))
631 // In addition to knowing that the call does not access src in some
632 // unexpected manner, for example via a global, which we deduce from
633 // the use analysis, we also need to know that it does not sneakily
634 // access dest. We rely on AA to figure this out for us.
635 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
636 if (AA.getModRefInfo(C, cpyDest, srcSize) !=
637 AliasAnalysis::NoModRef)
640 // All the checks have passed, so do the transformation.
641 bool changedArgument = false;
642 for (unsigned i = 0; i < CS.arg_size(); ++i)
643 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
644 if (cpySrc->getType() != cpyDest->getType())
645 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
646 cpyDest->getName(), C);
647 changedArgument = true;
648 if (CS.getArgument(i)->getType() == cpyDest->getType())
649 CS.setArgument(i, cpyDest);
651 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
652 CS.getArgument(i)->getType(), cpyDest->getName(), C));
655 if (!changedArgument)
658 // Drop any cached information about the call, because we may have changed
659 // its dependence information by changing its parameter.
660 MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
661 MD.removeInstruction(C);
664 MD.removeInstruction(cpy);
670 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)
671 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
672 /// copy from MDep's input if we can. MSize is the size of M's copy.
674 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
676 // We can only transforms memcpy's where the dest of one is the source of the
678 if (M->getSource() != MDep->getDest())
681 // Second, the length of the memcpy's must be the same, or the preceeding one
682 // must be larger than the following one.
683 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
684 if (!C1) return false;
686 uint64_t DepSize = C1->getValue().getZExtValue();
691 // Finally, we have to make sure that the dest of the second does not
692 // alias the source of the first.
693 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
694 if (!AA.isNoAlias(M->getRawDest(), MSize, MDep->getRawSource(), DepSize))
697 // If all checks passed, then we can transform these memcpy's
698 const Type *ArgTys[3] = {
699 M->getRawDest()->getType(),
700 MDep->getRawSource()->getType(),
701 M->getLength()->getType()
703 Function *MemCpyFun =
704 Intrinsic::getDeclaration(M->getParent()->getParent()->getParent(),
705 M->getIntrinsicID(), ArgTys, 3);
707 // Make sure to use the lesser of the alignment of the source and the dest
708 // since we're changing where we're reading from, but don't want to increase
709 // the alignment past what can be read from or written to.
710 // TODO: Is this worth it if we're creating a less aligned memcpy? For
711 // example we could be moving from movaps -> movq on x86.
712 unsigned Align = std::min(MDep->getAlignmentCst()->getZExtValue(),
713 M->getAlignmentCst()->getZExtValue());
714 LLVMContext &Context = M->getContext();
715 ConstantInt *AlignCI = ConstantInt::get(Type::getInt32Ty(Context), Align);
717 M->getRawDest(), MDep->getRawSource(), M->getLength(),
718 AlignCI, M->getVolatileCst()
720 CallInst *C = CallInst::Create(MemCpyFun, Args, Args+5, "", M);
723 MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
725 // If C and M don't interfere, then this is a valid transformation. If they
726 // did, this would mean that the two sources overlap, which would be bad.
727 MemDepResult dep = MD.getDependency(C);
728 if (dep.isClobber() && dep.getInst() == MDep) {
729 MD.removeInstruction(M);
730 M->eraseFromParent();
735 // Otherwise, there was no point in doing this, so we remove the call we
736 // inserted and act like nothing happened.
737 MD.removeInstruction(C);
738 C->eraseFromParent();
743 /// processMemCpy - perform simplification of memcpy's. If we have memcpy A
744 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
745 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
746 /// circumstances). This allows later passes to remove the first memcpy
748 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
749 MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
751 // We can only optimize statically-sized memcpy's.
752 ConstantInt *cpyLen = dyn_cast<ConstantInt>(M->getLength());
753 if (!cpyLen) return false;
755 // The are two possible optimizations we can do for memcpy:
756 // a) memcpy-memcpy xform which exposes redundance for DSE.
757 // b) call-memcpy xform for return slot optimization.
758 MemDepResult dep = MD.getDependency(M);
759 if (!dep.isClobber())
762 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(dep.getInst()))
763 return processMemCpyMemCpyDependence(M, MDep, cpyLen->getZExtValue());
765 if (CallInst *C = dyn_cast<CallInst>(dep.getInst())) {
766 bool changed = performCallSlotOptzn(M, M->getDest(), M->getSource(),
767 cpyLen->getZExtValue(), C);
768 if (changed) M->eraseFromParent();
774 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
775 /// are guaranteed not to alias.
776 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
777 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
779 // If the memmove is a constant size, use it for the alias query, this allows
780 // us to optimize things like: memmove(P, P+64, 64);
781 uint64_t MemMoveSize = AliasAnalysis::UnknownSize;
782 if (ConstantInt *Len = dyn_cast<ConstantInt>(M->getLength()))
783 MemMoveSize = Len->getZExtValue();
785 // See if the pointers alias.
786 if (AA.alias(M->getRawDest(), MemMoveSize, M->getRawSource(), MemMoveSize) !=
787 AliasAnalysis::NoAlias)
790 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
792 // If not, then we know we can transform this.
793 Module *Mod = M->getParent()->getParent()->getParent();
794 const Type *ArgTys[3] = { M->getRawDest()->getType(),
795 M->getRawSource()->getType(),
796 M->getLength()->getType() };
797 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
800 // MemDep may have over conservative information about this instruction, just
801 // conservatively flush it from the cache.
802 getAnalysis<MemoryDependenceAnalysis>().removeInstruction(M);
809 // MemCpyOpt::iterateOnFunction - Executes one iteration of GVN.
810 bool MemCpyOpt::iterateOnFunction(Function &F) {
811 bool MadeChange = false;
813 // Walk all instruction in the function.
814 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
815 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
817 // Avoid invalidating the iterator.
818 Instruction *I = BI++;
820 if (StoreInst *SI = dyn_cast<StoreInst>(I))
821 MadeChange |= processStore(SI, BI);
822 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
823 MadeChange |= processMemCpy(M);
824 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) {
825 if (processMemMove(M)) {
826 --BI; // Reprocess the new memcpy.
836 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
839 bool MemCpyOpt::runOnFunction(Function &F) {
840 bool MadeChange = false;
842 if (!iterateOnFunction(F))