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;
81 // If one pointer is a GEP and the other isn't, then see if the GEP is a
82 // constant offset from the base, as in "P" and "gep P, 1".
83 if (GEP1 && GEP2 == 0 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
84 Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, TD);
85 return !VariableIdxFound;
88 if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
89 Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD);
90 return !VariableIdxFound;
93 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
94 // base. After that base, they may have some number of common (and
95 // potentially variable) indices. After that they handle some constant
96 // offset, which determines their offset from each other. At this point, we
97 // handle no other case.
98 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
101 // Skip any common indices and track the GEP types.
103 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
104 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
107 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
108 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
109 if (VariableIdxFound) return false;
111 Offset = Offset2-Offset1;
116 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
117 /// This allows us to analyze stores like:
122 /// which sometimes happens with stores to arrays of structs etc. When we see
123 /// the first store, we make a range [1, 2). The second store extends the range
124 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
125 /// two ranges into [0, 3) which is memset'able.
128 // Start/End - A semi range that describes the span that this range covers.
129 // The range is closed at the start and open at the end: [Start, End).
132 /// StartPtr - The getelementptr instruction that points to the start of the
136 /// Alignment - The known alignment of the first store.
139 /// TheStores - The actual stores that make up this range.
140 SmallVector<Instruction*, 16> TheStores;
142 bool isProfitableToUseMemset(const TargetData &TD) const;
145 } // end anon namespace
147 bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const {
148 // If we found more than 8 stores to merge or 64 bytes, use memset.
149 if (TheStores.size() >= 8 || End-Start >= 64) return true;
151 // If there is nothing to merge, don't do anything.
152 if (TheStores.size() < 2) return false;
154 // If any of the stores are a memset, then it is always good to extend the
156 for (unsigned i = 0, e = TheStores.size(); i != e; ++i)
157 if (!isa<StoreInst>(TheStores[i]))
160 // Assume that the code generator is capable of merging pairs of stores
161 // together if it wants to.
162 if (TheStores.size() == 2) return false;
164 // If we have fewer than 8 stores, it can still be worthwhile to do this.
165 // For example, merging 4 i8 stores into an i32 store is useful almost always.
166 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
167 // memset will be split into 2 32-bit stores anyway) and doing so can
168 // pessimize the llvm optimizer.
170 // Since we don't have perfect knowledge here, make some assumptions: assume
171 // the maximum GPR width is the same size as the pointer size and assume that
172 // this width can be stored. If so, check to see whether we will end up
173 // actually reducing the number of stores used.
174 unsigned Bytes = unsigned(End-Start);
175 unsigned NumPointerStores = Bytes/TD.getPointerSize();
177 // Assume the remaining bytes if any are done a byte at a time.
178 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
180 // If we will reduce the # stores (according to this heuristic), do the
181 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
183 return TheStores.size() > NumPointerStores+NumByteStores;
189 /// Ranges - A sorted list of the memset ranges. We use std::list here
190 /// because each element is relatively large and expensive to copy.
191 std::list<MemsetRange> Ranges;
192 typedef std::list<MemsetRange>::iterator range_iterator;
193 const TargetData &TD;
195 MemsetRanges(const TargetData &td) : TD(td) {}
197 typedef std::list<MemsetRange>::const_iterator const_iterator;
198 const_iterator begin() const { return Ranges.begin(); }
199 const_iterator end() const { return Ranges.end(); }
200 bool empty() const { return Ranges.empty(); }
202 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
203 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
204 addStore(OffsetFromFirst, SI);
206 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
209 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
210 int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType());
212 addRange(OffsetFromFirst, StoreSize,
213 SI->getPointerOperand(), SI->getAlignment(), SI);
216 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
217 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
218 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI);
221 void addRange(int64_t Start, int64_t Size, Value *Ptr,
222 unsigned Alignment, Instruction *Inst);
226 } // end anon namespace
229 /// addRange - Add a new store to the MemsetRanges data structure. This adds a
230 /// new range for the specified store at the specified offset, merging into
231 /// existing ranges as appropriate.
233 /// Do a linear search of the ranges to see if this can be joined and/or to
234 /// find the insertion point in the list. We keep the ranges sorted for
235 /// simplicity here. This is a linear search of a linked list, which is ugly,
236 /// however the number of ranges is limited, so this won't get crazy slow.
237 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
238 unsigned Alignment, Instruction *Inst) {
239 int64_t End = Start+Size;
240 range_iterator I = Ranges.begin(), E = Ranges.end();
242 while (I != E && Start > I->End)
245 // We now know that I == E, in which case we didn't find anything to merge
246 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
247 // to insert a new range. Handle this now.
248 if (I == E || End < I->Start) {
249 MemsetRange &R = *Ranges.insert(I, MemsetRange());
253 R.Alignment = Alignment;
254 R.TheStores.push_back(Inst);
258 // This store overlaps with I, add it.
259 I->TheStores.push_back(Inst);
261 // At this point, we may have an interval that completely contains our store.
262 // If so, just add it to the interval and return.
263 if (I->Start <= Start && I->End >= End)
266 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
267 // but is not entirely contained within the range.
269 // See if the range extends the start of the range. In this case, it couldn't
270 // possibly cause it to join the prior range, because otherwise we would have
272 if (Start < I->Start) {
275 I->Alignment = Alignment;
278 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
279 // is in or right at the end of I), and that End >= I->Start. Extend I out to
283 range_iterator NextI = I;
284 while (++NextI != E && End >= NextI->Start) {
285 // Merge the range in.
286 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
287 if (NextI->End > I->End)
295 //===----------------------------------------------------------------------===//
297 //===----------------------------------------------------------------------===//
300 class MemCpyOpt : public FunctionPass {
301 MemoryDependenceAnalysis *MD;
302 const TargetData *TD;
304 static char ID; // Pass identification, replacement for typeid
305 MemCpyOpt() : FunctionPass(ID) {
306 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
310 bool runOnFunction(Function &F);
313 // This transformation requires dominator postdominator info
314 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
315 AU.setPreservesCFG();
316 AU.addRequired<DominatorTree>();
317 AU.addRequired<MemoryDependenceAnalysis>();
318 AU.addRequired<AliasAnalysis>();
319 AU.addPreserved<AliasAnalysis>();
320 AU.addPreserved<MemoryDependenceAnalysis>();
324 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
325 bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI);
326 bool processMemCpy(MemCpyInst *M);
327 bool processMemMove(MemMoveInst *M);
328 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
329 uint64_t cpyLen, CallInst *C);
330 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
332 bool processByValArgument(CallSite CS, unsigned ArgNo);
333 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr,
336 bool iterateOnFunction(Function &F);
339 char MemCpyOpt::ID = 0;
342 // createMemCpyOptPass - The public interface to this file...
343 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
345 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
347 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
348 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
349 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
350 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
353 /// tryMergingIntoMemset - When scanning forward over instructions, we look for
354 /// some other patterns to fold away. In particular, this looks for stores to
355 /// neighboring locations of memory. If it sees enough consequtive ones, it
356 /// attempts to merge them together into a memcpy/memset.
357 Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
358 Value *StartPtr, Value *ByteVal) {
359 if (TD == 0) return 0;
361 // Okay, so we now have a single store that can be splatable. Scan to find
362 // all subsequent stores of the same value to offset from the same pointer.
363 // Join these together into ranges, so we can decide whether contiguous blocks
365 MemsetRanges Ranges(*TD);
367 BasicBlock::iterator BI = StartInst;
368 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
369 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
370 // If the instruction is readnone, ignore it, otherwise bail out. We
371 // don't even allow readonly here because we don't want something like:
372 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
373 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
378 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
379 // If this is a store, see if we can merge it in.
380 if (NextStore->isVolatile()) break;
382 // Check to see if this stored value is of the same byte-splattable value.
383 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
386 // Check to see if this store is to a constant offset from the start ptr.
388 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
391 Ranges.addStore(Offset, NextStore);
393 MemSetInst *MSI = cast<MemSetInst>(BI);
395 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
396 !isa<ConstantInt>(MSI->getLength()))
399 // Check to see if this store is to a constant offset from the start ptr.
401 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD))
404 Ranges.addMemSet(Offset, MSI);
408 // If we have no ranges, then we just had a single store with nothing that
409 // could be merged in. This is a very common case of course.
413 // If we had at least one store that could be merged in, add the starting
414 // store as well. We try to avoid this unless there is at least something
415 // interesting as a small compile-time optimization.
416 Ranges.addInst(0, StartInst);
418 // If we create any memsets, we put it right before the first instruction that
419 // isn't part of the memset block. This ensure that the memset is dominated
420 // by any addressing instruction needed by the start of the block.
421 IRBuilder<> Builder(BI);
423 // Now that we have full information about ranges, loop over the ranges and
424 // emit memset's for anything big enough to be worthwhile.
425 Instruction *AMemSet = 0;
426 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
428 const MemsetRange &Range = *I;
430 if (Range.TheStores.size() == 1) continue;
432 // If it is profitable to lower this range to memset, do so now.
433 if (!Range.isProfitableToUseMemset(*TD))
436 // Otherwise, we do want to transform this! Create a new memset.
437 // Get the starting pointer of the block.
438 StartPtr = Range.StartPtr;
440 // Determine alignment
441 unsigned Alignment = Range.Alignment;
442 if (Alignment == 0) {
443 const Type *EltType =
444 cast<PointerType>(StartPtr->getType())->getElementType();
445 Alignment = TD->getABITypeAlignment(EltType);
449 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
451 DEBUG(dbgs() << "Replace stores:\n";
452 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
453 dbgs() << *Range.TheStores[i] << '\n';
454 dbgs() << "With: " << *AMemSet << '\n');
456 // Zap all the stores.
457 for (SmallVector<Instruction*, 16>::const_iterator
458 SI = Range.TheStores.begin(),
459 SE = Range.TheStores.end(); SI != SE; ++SI)
460 (*SI)->eraseFromParent();
468 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
469 if (SI->isVolatile()) return false;
471 if (TD == 0) return false;
473 // Detect cases where we're performing call slot forwarding, but
474 // happen to be using a load-store pair to implement it, rather than
476 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
477 if (!LI->isVolatile() && LI->hasOneUse()) {
478 MemDepResult dep = MD->getDependency(LI);
480 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
481 C = dyn_cast<CallInst>(dep.getInst());
484 bool changed = performCallSlotOptzn(LI,
485 SI->getPointerOperand()->stripPointerCasts(),
486 LI->getPointerOperand()->stripPointerCasts(),
487 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
489 MD->removeInstruction(SI);
490 SI->eraseFromParent();
491 LI->eraseFromParent();
499 // There are two cases that are interesting for this code to handle: memcpy
500 // and memset. Right now we only handle memset.
502 // Ensure that the value being stored is something that can be memset'able a
503 // byte at a time like "0" or "-1" or any width, as well as things like
504 // 0xA0A0A0A0 and 0.0.
505 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
506 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
508 BBI = I; // Don't invalidate iterator.
515 bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
516 // See if there is another memset or store neighboring this memset which
517 // allows us to widen out the memset to do a single larger store.
518 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
520 BBI = I; // Don't invalidate iterator.
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 if (TD == 0) return false;
559 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
563 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
564 srcArraySize->getZExtValue();
566 if (cpyLen < srcSize)
569 // Check that accessing the first srcSize bytes of dest will not cause a
570 // trap. Otherwise the transform is invalid since it might cause a trap
571 // to occur earlier than it otherwise would.
572 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
573 // The destination is an alloca. Check it is larger than srcSize.
574 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
578 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
579 destArraySize->getZExtValue();
581 if (destSize < srcSize)
583 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
584 // If the destination is an sret parameter then only accesses that are
585 // outside of the returned struct type can trap.
586 if (!A->hasStructRetAttr())
589 const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
590 uint64_t destSize = TD->getTypeAllocSize(StructTy);
592 if (destSize < srcSize)
598 // Check that src is not accessed except via the call and the memcpy. This
599 // guarantees that it holds only undefined values when passed in (so the final
600 // memcpy can be dropped), that it is not read or written between the call and
601 // the memcpy, and that writing beyond the end of it is undefined.
602 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
603 srcAlloca->use_end());
604 while (!srcUseList.empty()) {
605 User *UI = srcUseList.pop_back_val();
607 if (isa<BitCastInst>(UI)) {
608 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
610 srcUseList.push_back(*I);
611 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
612 if (G->hasAllZeroIndices())
613 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
615 srcUseList.push_back(*I);
618 } else if (UI != C && UI != cpy) {
623 // Since we're changing the parameter to the callsite, we need to make sure
624 // that what would be the new parameter dominates the callsite.
625 DominatorTree &DT = getAnalysis<DominatorTree>();
626 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
627 if (!DT.dominates(cpyDestInst, C))
630 // In addition to knowing that the call does not access src in some
631 // unexpected manner, for example via a global, which we deduce from
632 // the use analysis, we also need to know that it does not sneakily
633 // access dest. We rely on AA to figure this out for us.
634 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
635 if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef)
638 // All the checks have passed, so do the transformation.
639 bool changedArgument = false;
640 for (unsigned i = 0; i < CS.arg_size(); ++i)
641 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
642 if (cpySrc->getType() != cpyDest->getType())
643 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
644 cpyDest->getName(), C);
645 changedArgument = true;
646 if (CS.getArgument(i)->getType() == cpyDest->getType())
647 CS.setArgument(i, cpyDest);
649 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
650 CS.getArgument(i)->getType(), cpyDest->getName(), C));
653 if (!changedArgument)
656 // Drop any cached information about the call, because we may have changed
657 // its dependence information by changing its parameter.
658 MD->removeInstruction(C);
660 // Remove the memcpy.
661 MD->removeInstruction(cpy);
667 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)
668 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
669 /// copy from MDep's input if we can. MSize is the size of M's copy.
671 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
673 // We can only transforms memcpy's where the dest of one is the source of the
675 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
678 // If dep instruction is reading from our current input, then it is a noop
679 // transfer and substituting the input won't change this instruction. Just
680 // ignore the input and let someone else zap MDep. This handles cases like:
683 if (M->getSource() == MDep->getSource())
686 // Second, the length of the memcpy's must be the same, or the preceeding one
687 // must be larger than the following one.
688 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
689 if (!C1) return false;
691 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
693 // Verify that the copied-from memory doesn't change in between the two
694 // transfers. For example, in:
698 // It would be invalid to transform the second memcpy into memcpy(c <- b).
700 // TODO: If the code between M and MDep is transparent to the destination "c",
701 // then we could still perform the xform by moving M up to the first memcpy.
703 // NOTE: This is conservative, it will stop on any read from the source loc,
704 // not just the defining memcpy.
705 MemDepResult SourceDep =
706 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
707 false, M, M->getParent());
708 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
711 // If the dest of the second might alias the source of the first, then the
712 // source and dest might overlap. We still want to eliminate the intermediate
713 // value, but we have to generate a memmove instead of memcpy.
714 bool UseMemMove = false;
715 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)))
718 // If all checks passed, then we can transform M.
720 // Make sure to use the lesser of the alignment of the source and the dest
721 // since we're changing where we're reading from, but don't want to increase
722 // the alignment past what can be read from or written to.
723 // TODO: Is this worth it if we're creating a less aligned memcpy? For
724 // example we could be moving from movaps -> movq on x86.
725 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
727 IRBuilder<> Builder(M);
729 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
730 Align, M->isVolatile());
732 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
733 Align, M->isVolatile());
735 // Remove the instruction we're replacing.
736 MD->removeInstruction(M);
737 M->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 // We can only optimize statically-sized memcpy's that are non-volatile.
750 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
751 if (CopySize == 0 || M->isVolatile()) return false;
753 // If the source and destination of the memcpy are the same, then zap it.
754 if (M->getSource() == M->getDest()) {
755 MD->removeInstruction(M);
756 M->eraseFromParent();
760 // If copying from a constant, try to turn the memcpy into a memset.
761 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
762 if (GV->isConstant() && GV->hasDefinitiveInitializer())
763 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
764 IRBuilder<> Builder(M);
765 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize,
766 M->getAlignment(), false);
767 MD->removeInstruction(M);
768 M->eraseFromParent();
773 // The are two possible optimizations we can do for memcpy:
774 // a) memcpy-memcpy xform which exposes redundance for DSE.
775 // b) call-memcpy xform for return slot optimization.
776 MemDepResult DepInfo = MD->getDependency(M);
777 if (!DepInfo.isClobber())
780 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
781 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
783 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
784 if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
785 CopySize->getZExtValue(), C)) {
786 M->eraseFromParent();
794 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
795 /// are guaranteed not to alias.
796 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
797 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
799 // See if the pointers alias.
800 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M)))
803 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
805 // If not, then we know we can transform this.
806 Module *Mod = M->getParent()->getParent()->getParent();
807 const Type *ArgTys[3] = { M->getRawDest()->getType(),
808 M->getRawSource()->getType(),
809 M->getLength()->getType() };
810 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
813 // MemDep may have over conservative information about this instruction, just
814 // conservatively flush it from the cache.
815 MD->removeInstruction(M);
821 /// processByValArgument - This is called on every byval argument in call sites.
822 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
823 if (TD == 0) return false;
825 // Find out what feeds this byval argument.
826 Value *ByValArg = CS.getArgument(ArgNo);
827 const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType();
828 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
829 MemDepResult DepInfo =
830 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
831 true, CS.getInstruction(),
832 CS.getInstruction()->getParent());
833 if (!DepInfo.isClobber())
836 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
837 // a memcpy, see if we can byval from the source of the memcpy instead of the
839 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
840 if (MDep == 0 || MDep->isVolatile() ||
841 ByValArg->stripPointerCasts() != MDep->getDest())
844 // The length of the memcpy must be larger or equal to the size of the byval.
845 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
846 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
849 // Get the alignment of the byval. If it is greater than the memcpy, then we
850 // can't do the substitution. If the call doesn't specify the alignment, then
851 // it is some target specific value that we can't know.
852 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
853 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign)
856 // Verify that the copied-from memory doesn't change in between the memcpy and
861 // It would be invalid to transform the second memcpy into foo(*b).
863 // NOTE: This is conservative, it will stop on any read from the source loc,
864 // not just the defining memcpy.
865 MemDepResult SourceDep =
866 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
867 false, CS.getInstruction(), MDep->getParent());
868 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
871 Value *TmpCast = MDep->getSource();
872 if (MDep->getSource()->getType() != ByValArg->getType())
873 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
874 "tmpcast", CS.getInstruction());
876 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
877 << " " << *MDep << "\n"
878 << " " << *CS.getInstruction() << "\n");
880 // Otherwise we're good! Update the byval argument.
881 CS.setArgument(ArgNo, TmpCast);
886 /// iterateOnFunction - Executes one iteration of MemCpyOpt.
887 bool MemCpyOpt::iterateOnFunction(Function &F) {
888 bool MadeChange = false;
890 // Walk all instruction in the function.
891 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
892 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
893 // Avoid invalidating the iterator.
894 Instruction *I = BI++;
896 bool RepeatInstruction = false;
898 if (StoreInst *SI = dyn_cast<StoreInst>(I))
899 MadeChange |= processStore(SI, BI);
900 else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
901 RepeatInstruction = processMemSet(M, BI);
902 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
903 RepeatInstruction = processMemCpy(M);
904 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I))
905 RepeatInstruction = processMemMove(M);
906 else if (CallSite CS = (Value*)I) {
907 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
908 if (CS.paramHasAttr(i+1, Attribute::ByVal))
909 MadeChange |= processByValArgument(CS, i);
912 // Reprocess the instruction if desired.
913 if (RepeatInstruction) {
923 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
926 bool MemCpyOpt::runOnFunction(Function &F) {
927 bool MadeChange = false;
928 MD = &getAnalysis<MemoryDependenceAnalysis>();
929 TD = getAnalysisIfAvailable<TargetData>();
931 if (!iterateOnFunction(F))