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 //Ptr1 = Ptr1->stripPointerCasts();
75 //Ptr2 = Ptr2->stripPointerCasts();
76 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
77 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
79 bool VariableIdxFound = false;
82 // If one pointer is a GEP and the other isn't, then see if the GEP is a
83 // constant offset from the base, as in "P" and "gep P, 1".
84 if (GEP1 && GEP2 == 0 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
85 Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, TD);
86 return !VariableIdxFound;
89 if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
90 Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD);
91 return !VariableIdxFound;
95 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
96 // base. After that base, they may have some number of common (and
97 // potentially variable) indices. After that they handle some constant
98 // offset, which determines their offset from each other. At this point, we
99 // handle no other case.
100 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
103 // Skip any common indices and track the GEP types.
105 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
106 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
109 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
110 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
111 if (VariableIdxFound) return false;
113 Offset = Offset2-Offset1;
118 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
119 /// This allows us to analyze stores like:
124 /// which sometimes happens with stores to arrays of structs etc. When we see
125 /// the first store, we make a range [1, 2). The second store extends the range
126 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
127 /// two ranges into [0, 3) which is memset'able.
130 // Start/End - A semi range that describes the span that this range covers.
131 // The range is closed at the start and open at the end: [Start, End).
134 /// StartPtr - The getelementptr instruction that points to the start of the
138 /// Alignment - The known alignment of the first store.
141 /// TheStores - The actual stores that make up this range.
142 SmallVector<Instruction*, 16> TheStores;
144 bool isProfitableToUseMemset(const TargetData &TD) const;
147 } // end anon namespace
149 bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const {
150 // If we found more than 8 stores to merge or 64 bytes, use memset.
151 if (TheStores.size() >= 8 || End-Start >= 64) return true;
153 // If there is nothing to merge, don't do anything.
154 if (TheStores.size() < 2) return false;
156 // If any of the stores are a memset, then it is always good to extend the
158 for (unsigned i = 0, e = TheStores.size(); i != e; ++i)
159 if (!isa<StoreInst>(TheStores[i]))
162 // Assume that the code generator is capable of merging pairs of stores
163 // together if it wants to.
164 if (TheStores.size() == 2) return false;
166 // If we have fewer than 8 stores, it can still be worthwhile to do this.
167 // For example, merging 4 i8 stores into an i32 store is useful almost always.
168 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
169 // memset will be split into 2 32-bit stores anyway) and doing so can
170 // pessimize the llvm optimizer.
172 // Since we don't have perfect knowledge here, make some assumptions: assume
173 // the maximum GPR width is the same size as the pointer size and assume that
174 // this width can be stored. If so, check to see whether we will end up
175 // actually reducing the number of stores used.
176 unsigned Bytes = unsigned(End-Start);
177 unsigned NumPointerStores = Bytes/TD.getPointerSize();
179 // Assume the remaining bytes if any are done a byte at a time.
180 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
182 // If we will reduce the # stores (according to this heuristic), do the
183 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
185 return TheStores.size() > NumPointerStores+NumByteStores;
191 /// Ranges - A sorted list of the memset ranges. We use std::list here
192 /// because each element is relatively large and expensive to copy.
193 std::list<MemsetRange> Ranges;
194 typedef std::list<MemsetRange>::iterator range_iterator;
195 const TargetData &TD;
197 MemsetRanges(const TargetData &td) : TD(td) {}
199 typedef std::list<MemsetRange>::const_iterator const_iterator;
200 const_iterator begin() const { return Ranges.begin(); }
201 const_iterator end() const { return Ranges.end(); }
202 bool empty() const { return Ranges.empty(); }
204 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
205 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
206 addStore(OffsetFromFirst, SI);
208 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
211 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
212 int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType());
214 addRange(OffsetFromFirst, StoreSize,
215 SI->getPointerOperand(), SI->getAlignment(), SI);
218 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
219 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
220 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI);
223 void addRange(int64_t Start, int64_t Size, Value *Ptr,
224 unsigned Alignment, Instruction *Inst);
228 } // end anon namespace
231 /// addRange - Add a new store to the MemsetRanges data structure. This adds a
232 /// new range for the specified store at the specified offset, merging into
233 /// existing ranges as appropriate.
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 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
240 unsigned Alignment, Instruction *Inst) {
241 int64_t End = Start+Size;
242 range_iterator I = Ranges.begin(), E = Ranges.end();
244 while (I != E && Start > I->End)
247 // We now know that I == E, in which case we didn't find anything to merge
248 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
249 // to insert a new range. Handle this now.
250 if (I == E || End < I->Start) {
251 MemsetRange &R = *Ranges.insert(I, MemsetRange());
255 R.Alignment = Alignment;
256 R.TheStores.push_back(Inst);
260 // This store overlaps with I, add it.
261 I->TheStores.push_back(Inst);
263 // At this point, we may have an interval that completely contains our store.
264 // If so, just add it to the interval and return.
265 if (I->Start <= Start && I->End >= End)
268 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
269 // but is not entirely contained within the range.
271 // See if the range extends the start of the range. In this case, it couldn't
272 // possibly cause it to join the prior range, because otherwise we would have
274 if (Start < I->Start) {
277 I->Alignment = Alignment;
280 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
281 // is in or right at the end of I), and that End >= I->Start. Extend I out to
285 range_iterator NextI = I;
286 while (++NextI != E && End >= NextI->Start) {
287 // Merge the range in.
288 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
289 if (NextI->End > I->End)
297 //===----------------------------------------------------------------------===//
299 //===----------------------------------------------------------------------===//
302 class MemCpyOpt : public FunctionPass {
303 MemoryDependenceAnalysis *MD;
304 const TargetData *TD;
306 static char ID; // Pass identification, replacement for typeid
307 MemCpyOpt() : FunctionPass(ID) {
308 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
312 bool runOnFunction(Function &F);
315 // This transformation requires dominator postdominator info
316 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
317 AU.setPreservesCFG();
318 AU.addRequired<DominatorTree>();
319 AU.addRequired<MemoryDependenceAnalysis>();
320 AU.addRequired<AliasAnalysis>();
321 AU.addPreserved<AliasAnalysis>();
322 AU.addPreserved<MemoryDependenceAnalysis>();
326 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
327 bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI);
328 bool processMemCpy(MemCpyInst *M);
329 bool processMemMove(MemMoveInst *M);
330 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
331 uint64_t cpyLen, CallInst *C);
332 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
334 bool processByValArgument(CallSite CS, unsigned ArgNo);
335 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr,
338 bool iterateOnFunction(Function &F);
341 char MemCpyOpt::ID = 0;
344 // createMemCpyOptPass - The public interface to this file...
345 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
347 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
349 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
350 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
351 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
352 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
355 /// tryMergingIntoMemset - When scanning forward over instructions, we look for
356 /// some other patterns to fold away. In particular, this looks for stores to
357 /// neighboring locations of memory. If it sees enough consequtive ones, it
358 /// attempts to merge them together into a memcpy/memset.
359 Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
360 Value *StartPtr, Value *ByteVal) {
361 if (TD == 0) return 0;
363 // Okay, so we now have a single store that can be splatable. Scan to find
364 // all subsequent stores of the same value to offset from the same pointer.
365 // Join these together into ranges, so we can decide whether contiguous blocks
367 MemsetRanges Ranges(*TD);
369 BasicBlock::iterator BI = StartInst;
370 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
371 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
372 // If the instruction is readnone, ignore it, otherwise bail out. We
373 // don't even allow readonly here because we don't want something like:
374 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
375 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
380 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
381 // If this is a store, see if we can merge it in.
382 if (NextStore->isVolatile()) break;
384 // Check to see if this stored value is of the same byte-splattable value.
385 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
388 // Check to see if this store is to a constant offset from the start ptr.
390 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(),
394 Ranges.addStore(Offset, NextStore);
398 MemSetInst *MSI = cast<MemSetInst>(BI);
400 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
401 !isa<ConstantInt>(MSI->getLength()))
404 // Check to see if this store is to a constant offset from the start ptr.
406 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD))
409 Ranges.addMemSet(Offset, MSI);
413 // If we have no ranges, then we just had a single store with nothing that
414 // could be merged in. This is a very common case of course.
418 // If we had at least one store that could be merged in, add the starting
419 // store as well. We try to avoid this unless there is at least something
420 // interesting as a small compile-time optimization.
421 Ranges.addInst(0, StartInst);
423 // If we create any memsets, we put it right before the first instruction that
424 // isn't part of the memset block. This ensure that the memset is dominated
425 // by any addressing instruction needed by the start of the block.
426 IRBuilder<> Builder(BI);
428 // Now that we have full information about ranges, loop over the ranges and
429 // emit memset's for anything big enough to be worthwhile.
430 Instruction *AMemSet = 0;
431 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
433 const MemsetRange &Range = *I;
435 if (Range.TheStores.size() == 1) continue;
437 // If it is profitable to lower this range to memset, do so now.
438 if (!Range.isProfitableToUseMemset(*TD))
441 // Otherwise, we do want to transform this! Create a new memset.
442 // Get the starting pointer of the block.
443 StartPtr = Range.StartPtr;
445 // Determine alignment
446 unsigned Alignment = Range.Alignment;
447 if (Alignment == 0) {
448 const Type *EltType =
449 cast<PointerType>(StartPtr->getType())->getElementType();
450 Alignment = TD->getABITypeAlignment(EltType);
454 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
456 DEBUG(dbgs() << "Replace stores:\n";
457 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
458 dbgs() << *Range.TheStores[i] << '\n';
459 dbgs() << "With: " << *AMemSet << '\n');
461 // Zap all the stores.
462 for (SmallVector<Instruction*, 16>::const_iterator
463 SI = Range.TheStores.begin(),
464 SE = Range.TheStores.end(); SI != SE; ++SI) {
465 MD->removeInstruction(*SI);
466 (*SI)->eraseFromParent();
475 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
476 if (SI->isVolatile()) return false;
478 if (TD == 0) return false;
480 // Detect cases where we're performing call slot forwarding, but
481 // happen to be using a load-store pair to implement it, rather than
483 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
484 if (!LI->isVolatile() && LI->hasOneUse()) {
485 MemDepResult dep = MD->getDependency(LI);
487 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
488 C = dyn_cast<CallInst>(dep.getInst());
491 bool changed = performCallSlotOptzn(LI,
492 SI->getPointerOperand()->stripPointerCasts(),
493 LI->getPointerOperand()->stripPointerCasts(),
494 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
496 MD->removeInstruction(SI);
497 SI->eraseFromParent();
498 MD->removeInstruction(LI);
499 LI->eraseFromParent();
507 // There are two cases that are interesting for this code to handle: memcpy
508 // and memset. Right now we only handle memset.
510 // Ensure that the value being stored is something that can be memset'able a
511 // byte at a time like "0" or "-1" or any width, as well as things like
512 // 0xA0A0A0A0 and 0.0.
513 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
514 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
516 BBI = I; // Don't invalidate iterator.
523 bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
524 // Temporarily disable this.
527 // See if there is another memset or store neighboring this memset which
528 // allows us to widen out the memset to do a single larger store.
529 if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
530 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
532 BBI = I; // Don't invalidate iterator.
539 /// performCallSlotOptzn - takes a memcpy and a call that it depends on,
540 /// and checks for the possibility of a call slot optimization by having
541 /// the call write its result directly into the destination of the memcpy.
542 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
543 Value *cpyDest, Value *cpySrc,
544 uint64_t cpyLen, CallInst *C) {
545 // The general transformation to keep in mind is
547 // call @func(..., src, ...)
548 // memcpy(dest, src, ...)
552 // memcpy(dest, src, ...)
553 // call @func(..., dest, ...)
555 // Since moving the memcpy is technically awkward, we additionally check that
556 // src only holds uninitialized values at the moment of the call, meaning that
557 // the memcpy can be discarded rather than moved.
559 // Deliberately get the source and destination with bitcasts stripped away,
560 // because we'll need to do type comparisons based on the underlying type.
563 // Require that src be an alloca. This simplifies the reasoning considerably.
564 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
568 // Check that all of src is copied to dest.
569 if (TD == 0) return false;
571 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
575 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
576 srcArraySize->getZExtValue();
578 if (cpyLen < srcSize)
581 // Check that accessing the first srcSize bytes of dest will not cause a
582 // trap. Otherwise the transform is invalid since it might cause a trap
583 // to occur earlier than it otherwise would.
584 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
585 // The destination is an alloca. Check it is larger than srcSize.
586 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
590 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
591 destArraySize->getZExtValue();
593 if (destSize < srcSize)
595 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
596 // If the destination is an sret parameter then only accesses that are
597 // outside of the returned struct type can trap.
598 if (!A->hasStructRetAttr())
601 const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
602 uint64_t destSize = TD->getTypeAllocSize(StructTy);
604 if (destSize < srcSize)
610 // Check that src is not accessed except via the call and the memcpy. This
611 // guarantees that it holds only undefined values when passed in (so the final
612 // memcpy can be dropped), that it is not read or written between the call and
613 // the memcpy, and that writing beyond the end of it is undefined.
614 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
615 srcAlloca->use_end());
616 while (!srcUseList.empty()) {
617 User *UI = srcUseList.pop_back_val();
619 if (isa<BitCastInst>(UI)) {
620 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
622 srcUseList.push_back(*I);
623 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
624 if (G->hasAllZeroIndices())
625 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
627 srcUseList.push_back(*I);
630 } else if (UI != C && UI != cpy) {
635 // Since we're changing the parameter to the callsite, we need to make sure
636 // that what would be the new parameter dominates the callsite.
637 DominatorTree &DT = getAnalysis<DominatorTree>();
638 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
639 if (!DT.dominates(cpyDestInst, C))
642 // In addition to knowing that the call does not access src in some
643 // unexpected manner, for example via a global, which we deduce from
644 // the use analysis, we also need to know that it does not sneakily
645 // access dest. We rely on AA to figure this out for us.
646 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
647 if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef)
650 // All the checks have passed, so do the transformation.
651 bool changedArgument = false;
652 for (unsigned i = 0; i < CS.arg_size(); ++i)
653 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
654 if (cpySrc->getType() != cpyDest->getType())
655 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
656 cpyDest->getName(), C);
657 changedArgument = true;
658 if (CS.getArgument(i)->getType() == cpyDest->getType())
659 CS.setArgument(i, cpyDest);
661 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
662 CS.getArgument(i)->getType(), cpyDest->getName(), C));
665 if (!changedArgument)
668 // Drop any cached information about the call, because we may have changed
669 // its dependence information by changing its parameter.
670 MD->removeInstruction(C);
672 // Remove the memcpy.
673 MD->removeInstruction(cpy);
679 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)
680 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
681 /// copy from MDep's input if we can. MSize is the size of M's copy.
683 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
685 // We can only transforms memcpy's where the dest of one is the source of the
687 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
690 // If dep instruction is reading from our current input, then it is a noop
691 // transfer and substituting the input won't change this instruction. Just
692 // ignore the input and let someone else zap MDep. This handles cases like:
695 if (M->getSource() == MDep->getSource())
698 // Second, the length of the memcpy's must be the same, or the preceeding one
699 // must be larger than the following one.
700 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
701 if (!C1) return false;
703 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
705 // Verify that the copied-from memory doesn't change in between the two
706 // transfers. For example, in:
710 // It would be invalid to transform the second memcpy into memcpy(c <- b).
712 // TODO: If the code between M and MDep is transparent to the destination "c",
713 // then we could still perform the xform by moving M up to the first memcpy.
715 // NOTE: This is conservative, it will stop on any read from the source loc,
716 // not just the defining memcpy.
717 MemDepResult SourceDep =
718 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
719 false, M, M->getParent());
720 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
723 // If the dest of the second might alias the source of the first, then the
724 // source and dest might overlap. We still want to eliminate the intermediate
725 // value, but we have to generate a memmove instead of memcpy.
726 bool UseMemMove = false;
727 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)))
730 // If all checks passed, then we can transform M.
732 // Make sure to use the lesser of the alignment of the source and the dest
733 // since we're changing where we're reading from, but don't want to increase
734 // the alignment past what can be read from or written to.
735 // TODO: Is this worth it if we're creating a less aligned memcpy? For
736 // example we could be moving from movaps -> movq on x86.
737 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
739 IRBuilder<> Builder(M);
741 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
742 Align, M->isVolatile());
744 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
745 Align, M->isVolatile());
747 // Remove the instruction we're replacing.
748 MD->removeInstruction(M);
749 M->eraseFromParent();
755 /// processMemCpy - perform simplification of memcpy's. If we have memcpy A
756 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
757 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
758 /// circumstances). This allows later passes to remove the first memcpy
760 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
761 // We can only optimize statically-sized memcpy's that are non-volatile.
762 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
763 if (CopySize == 0 || M->isVolatile()) return false;
765 // If the source and destination of the memcpy are the same, then zap it.
766 if (M->getSource() == M->getDest()) {
767 MD->removeInstruction(M);
768 M->eraseFromParent();
772 // If copying from a constant, try to turn the memcpy into a memset.
773 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
774 if (GV->isConstant() && GV->hasDefinitiveInitializer())
775 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
776 IRBuilder<> Builder(M);
777 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize,
778 M->getAlignment(), false);
779 MD->removeInstruction(M);
780 M->eraseFromParent();
785 // The are two possible optimizations we can do for memcpy:
786 // a) memcpy-memcpy xform which exposes redundance for DSE.
787 // b) call-memcpy xform for return slot optimization.
788 MemDepResult DepInfo = MD->getDependency(M);
789 if (!DepInfo.isClobber())
792 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
793 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
795 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
796 if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
797 CopySize->getZExtValue(), C)) {
798 MD->removeInstruction(M);
799 M->eraseFromParent();
807 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
808 /// are guaranteed not to alias.
809 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
810 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
812 // See if the pointers alias.
813 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M)))
816 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
818 // If not, then we know we can transform this.
819 Module *Mod = M->getParent()->getParent()->getParent();
820 const Type *ArgTys[3] = { M->getRawDest()->getType(),
821 M->getRawSource()->getType(),
822 M->getLength()->getType() };
823 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
826 // MemDep may have over conservative information about this instruction, just
827 // conservatively flush it from the cache.
828 MD->removeInstruction(M);
834 /// processByValArgument - This is called on every byval argument in call sites.
835 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
836 if (TD == 0) return false;
838 // Find out what feeds this byval argument.
839 Value *ByValArg = CS.getArgument(ArgNo);
840 const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType();
841 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
842 MemDepResult DepInfo =
843 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
844 true, CS.getInstruction(),
845 CS.getInstruction()->getParent());
846 if (!DepInfo.isClobber())
849 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
850 // a memcpy, see if we can byval from the source of the memcpy instead of the
852 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
853 if (MDep == 0 || MDep->isVolatile() ||
854 ByValArg->stripPointerCasts() != MDep->getDest())
857 // The length of the memcpy must be larger or equal to the size of the byval.
858 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
859 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
862 // Get the alignment of the byval. If it is greater than the memcpy, then we
863 // can't do the substitution. If the call doesn't specify the alignment, then
864 // it is some target specific value that we can't know.
865 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
866 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign)
869 // Verify that the copied-from memory doesn't change in between the memcpy and
874 // It would be invalid to transform the second memcpy into foo(*b).
876 // NOTE: This is conservative, it will stop on any read from the source loc,
877 // not just the defining memcpy.
878 MemDepResult SourceDep =
879 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
880 false, CS.getInstruction(), MDep->getParent());
881 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
884 Value *TmpCast = MDep->getSource();
885 if (MDep->getSource()->getType() != ByValArg->getType())
886 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
887 "tmpcast", CS.getInstruction());
889 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
890 << " " << *MDep << "\n"
891 << " " << *CS.getInstruction() << "\n");
893 // Otherwise we're good! Update the byval argument.
894 CS.setArgument(ArgNo, TmpCast);
899 /// iterateOnFunction - Executes one iteration of MemCpyOpt.
900 bool MemCpyOpt::iterateOnFunction(Function &F) {
901 bool MadeChange = false;
903 // Walk all instruction in the function.
904 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
905 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
906 // Avoid invalidating the iterator.
907 Instruction *I = BI++;
909 bool RepeatInstruction = false;
911 if (StoreInst *SI = dyn_cast<StoreInst>(I))
912 MadeChange |= processStore(SI, BI);
913 else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
914 RepeatInstruction = processMemSet(M, BI);
915 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
916 RepeatInstruction = processMemCpy(M);
917 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I))
918 RepeatInstruction = processMemMove(M);
919 else if (CallSite CS = (Value*)I) {
920 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
921 if (CS.paramHasAttr(i+1, Attribute::ByVal))
922 MadeChange |= processByValArgument(CS, i);
925 // Reprocess the instruction if desired.
926 if (RepeatInstruction) {
927 if (BI != BB->begin()) --BI;
936 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
939 bool MemCpyOpt::runOnFunction(Function &F) {
940 bool MadeChange = false;
941 MD = &getAnalysis<MemoryDependenceAnalysis>();
942 TD = getAnalysisIfAvailable<TargetData>();
944 if (!iterateOnFunction(F))