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"
31 #include "llvm/Target/TargetLibraryInfo.h"
35 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
36 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
37 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
38 STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
40 static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx,
41 bool &VariableIdxFound, const TargetData &TD){
42 // Skip over the first indices.
43 gep_type_iterator GTI = gep_type_begin(GEP);
44 for (unsigned i = 1; i != Idx; ++i, ++GTI)
47 // Compute the offset implied by the rest of the indices.
49 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
50 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
52 return VariableIdxFound = true;
53 if (OpC->isZero()) continue; // No offset.
55 // Handle struct indices, which add their field offset to the pointer.
56 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
57 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
61 // Otherwise, we have a sequential type like an array or vector. Multiply
62 // the index by the ElementSize.
63 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
64 Offset += Size*OpC->getSExtValue();
70 /// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a
71 /// constant offset, and return that constant offset. For example, Ptr1 might
72 /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8.
73 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
74 const TargetData &TD) {
75 Ptr1 = Ptr1->stripPointerCasts();
76 Ptr2 = Ptr2->stripPointerCasts();
77 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
78 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
80 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;
94 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
95 // base. After that base, they may have some number of common (and
96 // potentially variable) indices. After that they handle some constant
97 // offset, which determines their offset from each other. At this point, we
98 // handle no other case.
99 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
102 // Skip any common indices and track the GEP types.
104 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
105 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
108 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
109 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
110 if (VariableIdxFound) return false;
112 Offset = Offset2-Offset1;
117 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
118 /// This allows us to analyze stores like:
123 /// which sometimes happens with stores to arrays of structs etc. When we see
124 /// the first store, we make a range [1, 2). The second store extends the range
125 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
126 /// two ranges into [0, 3) which is memset'able.
129 // Start/End - A semi range that describes the span that this range covers.
130 // The range is closed at the start and open at the end: [Start, End).
133 /// StartPtr - The getelementptr instruction that points to the start of the
137 /// Alignment - The known alignment of the first store.
140 /// TheStores - The actual stores that make up this range.
141 SmallVector<Instruction*, 16> TheStores;
143 bool isProfitableToUseMemset(const TargetData &TD) const;
146 } // end anon namespace
148 bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const {
149 // If we found more than 8 stores to merge or 64 bytes, use memset.
150 if (TheStores.size() >= 8 || End-Start >= 64) return true;
152 // If there is nothing to merge, don't do anything.
153 if (TheStores.size() < 2) return false;
155 // If any of the stores are a memset, then it is always good to extend the
157 for (unsigned i = 0, e = TheStores.size(); i != e; ++i)
158 if (!isa<StoreInst>(TheStores[i]))
161 // Assume that the code generator is capable of merging pairs of stores
162 // together if it wants to.
163 if (TheStores.size() == 2) return false;
165 // If we have fewer than 8 stores, it can still be worthwhile to do this.
166 // For example, merging 4 i8 stores into an i32 store is useful almost always.
167 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
168 // memset will be split into 2 32-bit stores anyway) and doing so can
169 // pessimize the llvm optimizer.
171 // Since we don't have perfect knowledge here, make some assumptions: assume
172 // the maximum GPR width is the same size as the pointer size and assume that
173 // this width can be stored. If so, check to see whether we will end up
174 // actually reducing the number of stores used.
175 unsigned Bytes = unsigned(End-Start);
176 unsigned NumPointerStores = Bytes/TD.getPointerSize();
178 // Assume the remaining bytes if any are done a byte at a time.
179 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
181 // If we will reduce the # stores (according to this heuristic), do the
182 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
184 return TheStores.size() > NumPointerStores+NumByteStores;
190 /// Ranges - A sorted list of the memset ranges. We use std::list here
191 /// because each element is relatively large and expensive to copy.
192 std::list<MemsetRange> Ranges;
193 typedef std::list<MemsetRange>::iterator range_iterator;
194 const TargetData &TD;
196 MemsetRanges(const TargetData &td) : TD(td) {}
198 typedef std::list<MemsetRange>::const_iterator const_iterator;
199 const_iterator begin() const { return Ranges.begin(); }
200 const_iterator end() const { return Ranges.end(); }
201 bool empty() const { return Ranges.empty(); }
203 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
204 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
205 addStore(OffsetFromFirst, SI);
207 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
210 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
211 int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType());
213 addRange(OffsetFromFirst, StoreSize,
214 SI->getPointerOperand(), SI->getAlignment(), SI);
217 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
218 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
219 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI);
222 void addRange(int64_t Start, int64_t Size, Value *Ptr,
223 unsigned Alignment, Instruction *Inst);
227 } // end anon namespace
230 /// addRange - Add a new store to the MemsetRanges data structure. This adds a
231 /// new range for the specified store at the specified offset, merging into
232 /// existing ranges as appropriate.
234 /// Do a linear search of the ranges to see if this can be joined and/or to
235 /// find the insertion point in the list. We keep the ranges sorted for
236 /// simplicity here. This is a linear search of a linked list, which is ugly,
237 /// however the number of ranges is limited, so this won't get crazy slow.
238 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
239 unsigned Alignment, Instruction *Inst) {
240 int64_t End = Start+Size;
241 range_iterator I = Ranges.begin(), E = Ranges.end();
243 while (I != E && Start > I->End)
246 // We now know that I == E, in which case we didn't find anything to merge
247 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
248 // to insert a new range. Handle this now.
249 if (I == E || End < I->Start) {
250 MemsetRange &R = *Ranges.insert(I, MemsetRange());
254 R.Alignment = Alignment;
255 R.TheStores.push_back(Inst);
259 // This store overlaps with I, add it.
260 I->TheStores.push_back(Inst);
262 // At this point, we may have an interval that completely contains our store.
263 // If so, just add it to the interval and return.
264 if (I->Start <= Start && I->End >= End)
267 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
268 // but is not entirely contained within the range.
270 // See if the range extends the start of the range. In this case, it couldn't
271 // possibly cause it to join the prior range, because otherwise we would have
273 if (Start < I->Start) {
276 I->Alignment = Alignment;
279 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
280 // is in or right at the end of I), and that End >= I->Start. Extend I out to
284 range_iterator NextI = I;
285 while (++NextI != E && End >= NextI->Start) {
286 // Merge the range in.
287 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
288 if (NextI->End > I->End)
296 //===----------------------------------------------------------------------===//
298 //===----------------------------------------------------------------------===//
301 class MemCpyOpt : public FunctionPass {
302 MemoryDependenceAnalysis *MD;
303 TargetLibraryInfo *TLI;
304 const TargetData *TD;
306 static char ID; // Pass identification, replacement for typeid
307 MemCpyOpt() : FunctionPass(ID) {
308 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
314 bool runOnFunction(Function &F);
317 // This transformation requires dominator postdominator info
318 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
319 AU.setPreservesCFG();
320 AU.addRequired<DominatorTree>();
321 AU.addRequired<MemoryDependenceAnalysis>();
322 AU.addRequired<AliasAnalysis>();
323 AU.addRequired<TargetLibraryInfo>();
324 AU.addPreserved<AliasAnalysis>();
325 AU.addPreserved<MemoryDependenceAnalysis>();
329 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
330 bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI);
331 bool processMemCpy(MemCpyInst *M);
332 bool processMemMove(MemMoveInst *M);
333 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
334 uint64_t cpyLen, CallInst *C);
335 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
337 bool processByValArgument(CallSite CS, unsigned ArgNo);
338 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr,
341 bool iterateOnFunction(Function &F);
344 char MemCpyOpt::ID = 0;
347 // createMemCpyOptPass - The public interface to this file...
348 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
350 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
352 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
353 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
354 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
355 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
356 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
359 /// tryMergingIntoMemset - When scanning forward over instructions, we look for
360 /// some other patterns to fold away. In particular, this looks for stores to
361 /// neighboring locations of memory. If it sees enough consecutive ones, it
362 /// attempts to merge them together into a memcpy/memset.
363 Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
364 Value *StartPtr, Value *ByteVal) {
365 if (TD == 0) return 0;
367 // Okay, so we now have a single store that can be splatable. Scan to find
368 // all subsequent stores of the same value to offset from the same pointer.
369 // Join these together into ranges, so we can decide whether contiguous blocks
371 MemsetRanges Ranges(*TD);
373 BasicBlock::iterator BI = StartInst;
374 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
375 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
376 // If the instruction is readnone, ignore it, otherwise bail out. We
377 // don't even allow readonly here because we don't want something like:
378 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
379 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
384 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
385 // If this is a store, see if we can merge it in.
386 if (NextStore->isVolatile()) break;
388 // Check to see if this stored value is of the same byte-splattable value.
389 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
392 // Check to see if this store is to a constant offset from the start ptr.
394 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(),
398 Ranges.addStore(Offset, NextStore);
400 MemSetInst *MSI = cast<MemSetInst>(BI);
402 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
403 !isa<ConstantInt>(MSI->getLength()))
406 // Check to see if this store is to a constant offset from the start ptr.
408 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD))
411 Ranges.addMemSet(Offset, MSI);
415 // If we have no ranges, then we just had a single store with nothing that
416 // could be merged in. This is a very common case of course.
420 // If we had at least one store that could be merged in, add the starting
421 // store as well. We try to avoid this unless there is at least something
422 // interesting as a small compile-time optimization.
423 Ranges.addInst(0, StartInst);
425 // If we create any memsets, we put it right before the first instruction that
426 // isn't part of the memset block. This ensure that the memset is dominated
427 // by any addressing instruction needed by the start of the block.
428 IRBuilder<> Builder(BI);
430 // Now that we have full information about ranges, loop over the ranges and
431 // emit memset's for anything big enough to be worthwhile.
432 Instruction *AMemSet = 0;
433 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
435 const MemsetRange &Range = *I;
437 if (Range.TheStores.size() == 1) continue;
439 // If it is profitable to lower this range to memset, do so now.
440 if (!Range.isProfitableToUseMemset(*TD))
443 // Otherwise, we do want to transform this! Create a new memset.
444 // Get the starting pointer of the block.
445 StartPtr = Range.StartPtr;
447 // Determine alignment
448 unsigned Alignment = Range.Alignment;
449 if (Alignment == 0) {
450 const Type *EltType =
451 cast<PointerType>(StartPtr->getType())->getElementType();
452 Alignment = TD->getABITypeAlignment(EltType);
456 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
458 DEBUG(dbgs() << "Replace stores:\n";
459 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
460 dbgs() << *Range.TheStores[i] << '\n';
461 dbgs() << "With: " << *AMemSet << '\n');
463 if (!Range.TheStores.empty())
464 AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
466 // Zap all the stores.
467 for (SmallVector<Instruction*, 16>::const_iterator
468 SI = Range.TheStores.begin(),
469 SE = Range.TheStores.end(); SI != SE; ++SI) {
470 MD->removeInstruction(*SI);
471 (*SI)->eraseFromParent();
480 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
481 if (SI->isVolatile()) return false;
483 if (TD == 0) return false;
485 // Detect cases where we're performing call slot forwarding, but
486 // happen to be using a load-store pair to implement it, rather than
488 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
489 if (!LI->isVolatile() && LI->hasOneUse()) {
490 MemDepResult dep = MD->getDependency(LI);
492 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
493 C = dyn_cast<CallInst>(dep.getInst());
496 bool changed = performCallSlotOptzn(LI,
497 SI->getPointerOperand()->stripPointerCasts(),
498 LI->getPointerOperand()->stripPointerCasts(),
499 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
501 MD->removeInstruction(SI);
502 SI->eraseFromParent();
503 MD->removeInstruction(LI);
504 LI->eraseFromParent();
512 // There are two cases that are interesting for this code to handle: memcpy
513 // and memset. Right now we only handle memset.
515 // Ensure that the value being stored is something that can be memset'able a
516 // byte at a time like "0" or "-1" or any width, as well as things like
517 // 0xA0A0A0A0 and 0.0.
518 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
519 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
521 BBI = I; // Don't invalidate iterator.
528 bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
529 // See if there is another memset or store neighboring this memset which
530 // allows us to widen out the memset to do a single larger store.
531 if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
532 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
534 BBI = I; // Don't invalidate iterator.
541 /// performCallSlotOptzn - takes a memcpy and a call that it depends on,
542 /// and checks for the possibility of a call slot optimization by having
543 /// the call write its result directly into the destination of the memcpy.
544 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
545 Value *cpyDest, Value *cpySrc,
546 uint64_t cpyLen, CallInst *C) {
547 // The general transformation to keep in mind is
549 // call @func(..., src, ...)
550 // memcpy(dest, src, ...)
554 // memcpy(dest, src, ...)
555 // call @func(..., dest, ...)
557 // Since moving the memcpy is technically awkward, we additionally check that
558 // src only holds uninitialized values at the moment of the call, meaning that
559 // the memcpy can be discarded rather than moved.
561 // Deliberately get the source and destination with bitcasts stripped away,
562 // because we'll need to do type comparisons based on the underlying type.
565 // Require that src be an alloca. This simplifies the reasoning considerably.
566 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
570 // Check that all of src is copied to dest.
571 if (TD == 0) return false;
573 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
577 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
578 srcArraySize->getZExtValue();
580 if (cpyLen < srcSize)
583 // Check that accessing the first srcSize bytes of dest will not cause a
584 // trap. Otherwise the transform is invalid since it might cause a trap
585 // to occur earlier than it otherwise would.
586 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
587 // The destination is an alloca. Check it is larger than srcSize.
588 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
592 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
593 destArraySize->getZExtValue();
595 if (destSize < srcSize)
597 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
598 // If the destination is an sret parameter then only accesses that are
599 // outside of the returned struct type can trap.
600 if (!A->hasStructRetAttr())
603 const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
604 uint64_t destSize = TD->getTypeAllocSize(StructTy);
606 if (destSize < srcSize)
612 // Check that src is not accessed except via the call and the memcpy. This
613 // guarantees that it holds only undefined values when passed in (so the final
614 // memcpy can be dropped), that it is not read or written between the call and
615 // the memcpy, and that writing beyond the end of it is undefined.
616 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
617 srcAlloca->use_end());
618 while (!srcUseList.empty()) {
619 User *UI = srcUseList.pop_back_val();
621 if (isa<BitCastInst>(UI)) {
622 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
624 srcUseList.push_back(*I);
625 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
626 if (G->hasAllZeroIndices())
627 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
629 srcUseList.push_back(*I);
632 } else if (UI != C && UI != cpy) {
637 // Since we're changing the parameter to the callsite, we need to make sure
638 // that what would be the new parameter dominates the callsite.
639 DominatorTree &DT = getAnalysis<DominatorTree>();
640 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
641 if (!DT.dominates(cpyDestInst, C))
644 // In addition to knowing that the call does not access src in some
645 // unexpected manner, for example via a global, which we deduce from
646 // the use analysis, we also need to know that it does not sneakily
647 // access dest. We rely on AA to figure this out for us.
648 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
649 if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef)
652 // All the checks have passed, so do the transformation.
653 bool changedArgument = false;
654 for (unsigned i = 0; i < CS.arg_size(); ++i)
655 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
656 if (cpySrc->getType() != cpyDest->getType())
657 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
658 cpyDest->getName(), C);
659 changedArgument = true;
660 if (CS.getArgument(i)->getType() == cpyDest->getType())
661 CS.setArgument(i, cpyDest);
663 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
664 CS.getArgument(i)->getType(), cpyDest->getName(), C));
667 if (!changedArgument)
670 // Drop any cached information about the call, because we may have changed
671 // its dependence information by changing its parameter.
672 MD->removeInstruction(C);
674 // Remove the memcpy.
675 MD->removeInstruction(cpy);
681 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)
682 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
683 /// copy from MDep's input if we can. MSize is the size of M's copy.
685 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
687 // We can only transforms memcpy's where the dest of one is the source of the
689 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
692 // If dep instruction is reading from our current input, then it is a noop
693 // transfer and substituting the input won't change this instruction. Just
694 // ignore the input and let someone else zap MDep. This handles cases like:
697 if (M->getSource() == MDep->getSource())
700 // Second, the length of the memcpy's must be the same, or the preceding one
701 // must be larger than the following one.
702 ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
703 ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength());
704 if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
707 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
709 // Verify that the copied-from memory doesn't change in between the two
710 // transfers. For example, in:
714 // It would be invalid to transform the second memcpy into memcpy(c <- b).
716 // TODO: If the code between M and MDep is transparent to the destination "c",
717 // then we could still perform the xform by moving M up to the first memcpy.
719 // NOTE: This is conservative, it will stop on any read from the source loc,
720 // not just the defining memcpy.
721 MemDepResult SourceDep =
722 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
723 false, M, M->getParent());
724 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
727 // If the dest of the second might alias the source of the first, then the
728 // source and dest might overlap. We still want to eliminate the intermediate
729 // value, but we have to generate a memmove instead of memcpy.
730 bool UseMemMove = false;
731 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)))
734 // If all checks passed, then we can transform M.
736 // Make sure to use the lesser of the alignment of the source and the dest
737 // since we're changing where we're reading from, but don't want to increase
738 // the alignment past what can be read from or written to.
739 // TODO: Is this worth it if we're creating a less aligned memcpy? For
740 // example we could be moving from movaps -> movq on x86.
741 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
743 IRBuilder<> Builder(M);
745 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
746 Align, M->isVolatile());
748 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
749 Align, M->isVolatile());
751 // Remove the instruction we're replacing.
752 MD->removeInstruction(M);
753 M->eraseFromParent();
759 /// processMemCpy - perform simplification of memcpy's. If we have memcpy A
760 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
761 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
762 /// circumstances). This allows later passes to remove the first memcpy
764 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
765 // We can only optimize statically-sized memcpy's that are non-volatile.
766 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
767 if (CopySize == 0 || M->isVolatile()) return false;
769 // If the source and destination of the memcpy are the same, then zap it.
770 if (M->getSource() == M->getDest()) {
771 MD->removeInstruction(M);
772 M->eraseFromParent();
776 // If copying from a constant, try to turn the memcpy into a memset.
777 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
778 if (GV->isConstant() && GV->hasDefinitiveInitializer())
779 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
780 IRBuilder<> Builder(M);
781 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize,
782 M->getAlignment(), false);
783 MD->removeInstruction(M);
784 M->eraseFromParent();
789 // The are two possible optimizations we can do for memcpy:
790 // a) memcpy-memcpy xform which exposes redundance for DSE.
791 // b) call-memcpy xform for return slot optimization.
792 MemDepResult DepInfo = MD->getDependency(M);
793 if (!DepInfo.isClobber())
796 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
797 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
799 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
800 if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
801 CopySize->getZExtValue(), C)) {
802 MD->removeInstruction(M);
803 M->eraseFromParent();
811 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
812 /// are guaranteed not to alias.
813 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
814 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
816 if (!TLI->has(LibFunc::memmove))
819 // See if the pointers alias.
820 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M)))
823 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
825 // If not, then we know we can transform this.
826 Module *Mod = M->getParent()->getParent()->getParent();
827 const Type *ArgTys[3] = { M->getRawDest()->getType(),
828 M->getRawSource()->getType(),
829 M->getLength()->getType() };
830 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
833 // MemDep may have over conservative information about this instruction, just
834 // conservatively flush it from the cache.
835 MD->removeInstruction(M);
841 /// processByValArgument - This is called on every byval argument in call sites.
842 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
843 if (TD == 0) return false;
845 // Find out what feeds this byval argument.
846 Value *ByValArg = CS.getArgument(ArgNo);
847 const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType();
848 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
849 MemDepResult DepInfo =
850 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
851 true, CS.getInstruction(),
852 CS.getInstruction()->getParent());
853 if (!DepInfo.isClobber())
856 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
857 // a memcpy, see if we can byval from the source of the memcpy instead of the
859 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
860 if (MDep == 0 || MDep->isVolatile() ||
861 ByValArg->stripPointerCasts() != MDep->getDest())
864 // The length of the memcpy must be larger or equal to the size of the byval.
865 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
866 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
869 // Get the alignment of the byval. If it is greater than the memcpy, then we
870 // can't do the substitution. If the call doesn't specify the alignment, then
871 // it is some target specific value that we can't know.
872 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
873 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign)
876 // Verify that the copied-from memory doesn't change in between the memcpy and
881 // It would be invalid to transform the second memcpy into foo(*b).
883 // NOTE: This is conservative, it will stop on any read from the source loc,
884 // not just the defining memcpy.
885 MemDepResult SourceDep =
886 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
887 false, CS.getInstruction(), MDep->getParent());
888 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
891 Value *TmpCast = MDep->getSource();
892 if (MDep->getSource()->getType() != ByValArg->getType())
893 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
894 "tmpcast", CS.getInstruction());
896 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
897 << " " << *MDep << "\n"
898 << " " << *CS.getInstruction() << "\n");
900 // Otherwise we're good! Update the byval argument.
901 CS.setArgument(ArgNo, TmpCast);
906 /// iterateOnFunction - Executes one iteration of MemCpyOpt.
907 bool MemCpyOpt::iterateOnFunction(Function &F) {
908 bool MadeChange = false;
910 // Walk all instruction in the function.
911 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
912 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
913 // Avoid invalidating the iterator.
914 Instruction *I = BI++;
916 bool RepeatInstruction = false;
918 if (StoreInst *SI = dyn_cast<StoreInst>(I))
919 MadeChange |= processStore(SI, BI);
920 else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
921 RepeatInstruction = processMemSet(M, BI);
922 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
923 RepeatInstruction = processMemCpy(M);
924 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I))
925 RepeatInstruction = processMemMove(M);
926 else if (CallSite CS = (Value*)I) {
927 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
928 if (CS.paramHasAttr(i+1, Attribute::ByVal))
929 MadeChange |= processByValArgument(CS, i);
932 // Reprocess the instruction if desired.
933 if (RepeatInstruction) {
934 if (BI != BB->begin()) --BI;
943 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
946 bool MemCpyOpt::runOnFunction(Function &F) {
947 bool MadeChange = false;
948 MD = &getAnalysis<MemoryDependenceAnalysis>();
949 TD = getAnalysisIfAvailable<TargetData>();
950 TLI = &getAnalysis<TargetLibraryInfo>();
952 // If we don't have at least memset and memcpy, there is little point of doing
953 // anything here. These are required by a freestanding implementation, so if
954 // even they are disabled, there is no point in trying hard.
955 if (!TLI->has(LibFunc::memset) || !TLI->has(LibFunc::memcpy))
959 if (!iterateOnFunction(F))