1 //===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs various transformations related to eliminating memcpy
11 // calls, or transforming sets of stores into memset's.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "memcpyopt"
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/IntrinsicInst.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/LLVMContext.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/GetElementPtrTypeIterator.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Target/TargetData.h"
32 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
33 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
34 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
36 /// isBytewiseValue - If the specified value can be set by repeating the same
37 /// byte in memory, return the i8 value that it is represented with. This is
38 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
39 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
40 /// byte store (e.g. i16 0x1234), return null.
41 static Value *isBytewiseValue(Value *V) {
42 LLVMContext &Context = V->getContext();
44 // All byte-wide stores are splatable, even of arbitrary variables.
45 if (V->getType()->isIntegerTy(8)) return V;
47 // Constant float and double values can be handled as integer values if the
48 // corresponding integer value is "byteable". An important case is 0.0.
49 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
50 if (CFP->getType()->isFloatTy())
51 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(Context));
52 if (CFP->getType()->isDoubleTy())
53 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(Context));
54 // Don't handle long double formats, which have strange constraints.
57 // We can handle constant integers that are power of two in size and a
58 // multiple of 8 bits.
59 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
60 unsigned Width = CI->getBitWidth();
61 if (isPowerOf2_32(Width) && Width > 8) {
62 // We can handle this value if the recursive binary decomposition is the
63 // same at all levels.
64 APInt Val = CI->getValue();
66 while (Val.getBitWidth() != 8) {
67 unsigned NextWidth = Val.getBitWidth()/2;
68 Val2 = Val.lshr(NextWidth);
69 Val2.trunc(Val.getBitWidth()/2);
70 Val.trunc(Val.getBitWidth()/2);
72 // If the top/bottom halves aren't the same, reject it.
76 return ConstantInt::get(Context, Val);
80 // Conceptually, we could handle things like:
81 // %a = zext i8 %X to i16
84 // but until there is an example that actually needs this, it doesn't seem
85 // worth worrying about.
89 static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx,
90 bool &VariableIdxFound, TargetData &TD) {
91 // Skip over the first indices.
92 gep_type_iterator GTI = gep_type_begin(GEP);
93 for (unsigned i = 1; i != Idx; ++i, ++GTI)
96 // Compute the offset implied by the rest of the indices.
98 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
99 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
101 return VariableIdxFound = true;
102 if (OpC->isZero()) continue; // No offset.
104 // Handle struct indices, which add their field offset to the pointer.
105 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
106 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
110 // Otherwise, we have a sequential type like an array or vector. Multiply
111 // the index by the ElementSize.
112 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
113 Offset += Size*OpC->getSExtValue();
119 /// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a
120 /// constant offset, and return that constant offset. For example, Ptr1 might
121 /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8.
122 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
124 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
125 // base. After that base, they may have some number of common (and
126 // potentially variable) indices. After that they handle some constant
127 // offset, which determines their offset from each other. At this point, we
128 // handle no other case.
129 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
130 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
131 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
134 // Skip any common indices and track the GEP types.
136 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
137 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
140 bool VariableIdxFound = false;
141 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
142 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
143 if (VariableIdxFound) return false;
145 Offset = Offset2-Offset1;
150 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
151 /// This allows us to analyze stores like:
156 /// which sometimes happens with stores to arrays of structs etc. When we see
157 /// the first store, we make a range [1, 2). The second store extends the range
158 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
159 /// two ranges into [0, 3) which is memset'able.
162 // Start/End - A semi range that describes the span that this range covers.
163 // The range is closed at the start and open at the end: [Start, End).
166 /// StartPtr - The getelementptr instruction that points to the start of the
170 /// Alignment - The known alignment of the first store.
173 /// TheStores - The actual stores that make up this range.
174 SmallVector<StoreInst*, 16> TheStores;
176 bool isProfitableToUseMemset(const TargetData &TD) const;
179 } // end anon namespace
181 bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const {
182 // If we found more than 8 stores to merge or 64 bytes, use memset.
183 if (TheStores.size() >= 8 || End-Start >= 64) return true;
185 // Assume that the code generator is capable of merging pairs of stores
186 // together if it wants to.
187 if (TheStores.size() <= 2) return false;
189 // If we have fewer than 8 stores, it can still be worthwhile to do this.
190 // For example, merging 4 i8 stores into an i32 store is useful almost always.
191 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
192 // memset will be split into 2 32-bit stores anyway) and doing so can
193 // pessimize the llvm optimizer.
195 // Since we don't have perfect knowledge here, make some assumptions: assume
196 // the maximum GPR width is the same size as the pointer size and assume that
197 // this width can be stored. If so, check to see whether we will end up
198 // actually reducing the number of stores used.
199 unsigned Bytes = unsigned(End-Start);
200 unsigned NumPointerStores = Bytes/TD.getPointerSize();
202 // Assume the remaining bytes if any are done a byte at a time.
203 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
205 // If we will reduce the # stores (according to this heuristic), do the
206 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
208 return TheStores.size() > NumPointerStores+NumByteStores;
214 /// Ranges - A sorted list of the memset ranges. We use std::list here
215 /// because each element is relatively large and expensive to copy.
216 std::list<MemsetRange> Ranges;
217 typedef std::list<MemsetRange>::iterator range_iterator;
220 MemsetRanges(TargetData &td) : TD(td) {}
222 typedef std::list<MemsetRange>::const_iterator const_iterator;
223 const_iterator begin() const { return Ranges.begin(); }
224 const_iterator end() const { return Ranges.end(); }
225 bool empty() const { return Ranges.empty(); }
227 void addStore(int64_t OffsetFromFirst, StoreInst *SI);
230 } // end anon namespace
233 /// addStore - Add a new store to the MemsetRanges data structure. This adds a
234 /// new range for the specified store at the specified offset, merging into
235 /// existing ranges as appropriate.
236 void MemsetRanges::addStore(int64_t Start, StoreInst *SI) {
237 int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType());
239 // Do a linear search of the ranges to see if this can be joined and/or to
240 // find the insertion point in the list. We keep the ranges sorted for
241 // simplicity here. This is a linear search of a linked list, which is ugly,
242 // however the number of ranges is limited, so this won't get crazy slow.
243 range_iterator I = Ranges.begin(), E = Ranges.end();
245 while (I != E && Start > I->End)
248 // We now know that I == E, in which case we didn't find anything to merge
249 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
250 // to insert a new range. Handle this now.
251 if (I == E || End < I->Start) {
252 MemsetRange &R = *Ranges.insert(I, MemsetRange());
255 R.StartPtr = SI->getPointerOperand();
256 R.Alignment = SI->getAlignment();
257 R.TheStores.push_back(SI);
261 // This store overlaps with I, add it.
262 I->TheStores.push_back(SI);
264 // At this point, we may have an interval that completely contains our store.
265 // If so, just add it to the interval and return.
266 if (I->Start <= Start && I->End >= End)
269 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
270 // but is not entirely contained within the range.
272 // See if the range extends the start of the range. In this case, it couldn't
273 // possibly cause it to join the prior range, because otherwise we would have
275 if (Start < I->Start) {
277 I->StartPtr = SI->getPointerOperand();
278 I->Alignment = SI->getAlignment();
281 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
282 // is in or right at the end of I), and that End >= I->Start. Extend I out to
286 range_iterator NextI = I;
287 while (++NextI != E && End >= NextI->Start) {
288 // Merge the range in.
289 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
290 if (NextI->End > I->End)
298 //===----------------------------------------------------------------------===//
300 //===----------------------------------------------------------------------===//
303 class MemCpyOpt : public FunctionPass {
304 MemoryDependenceAnalysis *MD;
305 bool runOnFunction(Function &F);
307 static char ID; // Pass identification, replacement for typeid
308 MemCpyOpt() : FunctionPass(ID) {
309 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
314 // This transformation requires dominator postdominator info
315 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
316 AU.setPreservesCFG();
317 AU.addRequired<DominatorTree>();
318 AU.addRequired<MemoryDependenceAnalysis>();
319 AU.addRequired<AliasAnalysis>();
320 AU.addPreserved<AliasAnalysis>();
321 AU.addPreserved<MemoryDependenceAnalysis>();
325 bool processStore(StoreInst *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 bool iterateOnFunction(Function &F);
336 char MemCpyOpt::ID = 0;
339 // createMemCpyOptPass - The public interface to this file...
340 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
342 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
344 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
345 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
346 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
347 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
350 /// processStore - When GVN is scanning forward over instructions, we look for
351 /// some other patterns to fold away. In particular, this looks for stores to
352 /// neighboring locations of memory. If it sees enough consequtive ones
353 /// (currently 4) it attempts to merge them together into a memcpy/memset.
354 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
355 if (SI->isVolatile()) return false;
357 TargetData *TD = getAnalysisIfAvailable<TargetData>();
358 if (!TD) return false;
360 // Detect cases where we're performing call slot forwarding, but
361 // happen to be using a load-store pair to implement it, rather than
363 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
364 if (!LI->isVolatile() && LI->hasOneUse()) {
365 MemDepResult dep = MD->getDependency(LI);
367 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
368 C = dyn_cast<CallInst>(dep.getInst());
371 bool changed = performCallSlotOptzn(LI,
372 SI->getPointerOperand()->stripPointerCasts(),
373 LI->getPointerOperand()->stripPointerCasts(),
374 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
376 MD->removeInstruction(SI);
377 SI->eraseFromParent();
378 LI->eraseFromParent();
386 LLVMContext &Context = SI->getContext();
388 // There are two cases that are interesting for this code to handle: memcpy
389 // and memset. Right now we only handle memset.
391 // Ensure that the value being stored is something that can be memset'able a
392 // byte at a time like "0" or "-1" or any width, as well as things like
393 // 0xA0A0A0A0 and 0.0.
394 Value *ByteVal = isBytewiseValue(SI->getOperand(0));
398 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
399 Module *M = SI->getParent()->getParent()->getParent();
401 // Okay, so we now have a single store that can be splatable. Scan to find
402 // all subsequent stores of the same value to offset from the same pointer.
403 // Join these together into ranges, so we can decide whether contiguous blocks
405 MemsetRanges Ranges(*TD);
407 Value *StartPtr = SI->getPointerOperand();
409 BasicBlock::iterator BI = SI;
410 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
411 if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
412 // If the call is readnone, ignore it, otherwise bail out. We don't even
413 // allow readonly here because we don't want something like:
414 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
415 if (AA.getModRefBehavior(CallSite(BI)) ==
416 AliasAnalysis::DoesNotAccessMemory)
419 // TODO: If this is a memset, try to join it in.
422 } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI))
425 // If this is a non-store instruction it is fine, ignore it.
426 StoreInst *NextStore = dyn_cast<StoreInst>(BI);
427 if (NextStore == 0) continue;
429 // If this is a store, see if we can merge it in.
430 if (NextStore->isVolatile()) break;
432 // Check to see if this stored value is of the same byte-splattable value.
433 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
436 // Check to see if this store is to a constant offset from the start ptr.
438 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
441 Ranges.addStore(Offset, NextStore);
444 // If we have no ranges, then we just had a single store with nothing that
445 // could be merged in. This is a very common case of course.
449 // If we had at least one store that could be merged in, add the starting
450 // store as well. We try to avoid this unless there is at least something
451 // interesting as a small compile-time optimization.
452 Ranges.addStore(0, SI);
455 // Now that we have full information about ranges, loop over the ranges and
456 // emit memset's for anything big enough to be worthwhile.
457 bool MadeChange = false;
458 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
460 const MemsetRange &Range = *I;
462 if (Range.TheStores.size() == 1) continue;
464 // If it is profitable to lower this range to memset, do so now.
465 if (!Range.isProfitableToUseMemset(*TD))
468 // Otherwise, we do want to transform this! Create a new memset. We put
469 // the memset right before the first instruction that isn't part of this
470 // memset block. This ensure that the memset is dominated by any addressing
471 // instruction needed by the start of the block.
472 BasicBlock::iterator InsertPt = BI;
474 // Get the starting pointer of the block.
475 StartPtr = Range.StartPtr;
477 // Determine alignment
478 unsigned Alignment = Range.Alignment;
479 if (Alignment == 0) {
480 const Type *EltType =
481 cast<PointerType>(StartPtr->getType())->getElementType();
482 Alignment = TD->getABITypeAlignment(EltType);
485 // Cast the start ptr to be i8* as memset requires.
486 const PointerType* StartPTy = cast<PointerType>(StartPtr->getType());
487 const PointerType *i8Ptr = Type::getInt8PtrTy(Context,
488 StartPTy->getAddressSpace());
489 if (StartPTy!= i8Ptr)
490 StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getName(),
494 StartPtr, ByteVal, // Start, value
496 ConstantInt::get(Type::getInt64Ty(Context), Range.End-Range.Start),
498 ConstantInt::get(Type::getInt32Ty(Context), Alignment),
500 ConstantInt::getFalse(Context),
502 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
504 Function *MemSetF = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2);
506 Value *C = CallInst::Create(MemSetF, Ops, Ops+5, "", InsertPt);
507 DEBUG(dbgs() << "Replace stores:\n";
508 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
509 dbgs() << *Range.TheStores[i] << '\n';
510 dbgs() << "With: " << *C << '\n'); C=C;
512 // Don't invalidate the iterator
515 // Zap all the stores.
516 for (SmallVector<StoreInst*, 16>::const_iterator
517 SI = Range.TheStores.begin(),
518 SE = Range.TheStores.end(); SI != SE; ++SI)
519 (*SI)->eraseFromParent();
528 /// performCallSlotOptzn - takes a memcpy and a call that it depends on,
529 /// and checks for the possibility of a call slot optimization by having
530 /// the call write its result directly into the destination of the memcpy.
531 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
532 Value *cpyDest, Value *cpySrc,
533 uint64_t cpyLen, CallInst *C) {
534 // The general transformation to keep in mind is
536 // call @func(..., src, ...)
537 // memcpy(dest, src, ...)
541 // memcpy(dest, src, ...)
542 // call @func(..., dest, ...)
544 // Since moving the memcpy is technically awkward, we additionally check that
545 // src only holds uninitialized values at the moment of the call, meaning that
546 // the memcpy can be discarded rather than moved.
548 // Deliberately get the source and destination with bitcasts stripped away,
549 // because we'll need to do type comparisons based on the underlying type.
552 // Require that src be an alloca. This simplifies the reasoning considerably.
553 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
557 // Check that all of src is copied to dest.
558 TargetData *TD = getAnalysisIfAvailable<TargetData>();
559 if (!TD) return false;
561 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
565 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
566 srcArraySize->getZExtValue();
568 if (cpyLen < srcSize)
571 // Check that accessing the first srcSize bytes of dest will not cause a
572 // trap. Otherwise the transform is invalid since it might cause a trap
573 // to occur earlier than it otherwise would.
574 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
575 // The destination is an alloca. Check it is larger than srcSize.
576 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
580 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
581 destArraySize->getZExtValue();
583 if (destSize < srcSize)
585 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
586 // If the destination is an sret parameter then only accesses that are
587 // outside of the returned struct type can trap.
588 if (!A->hasStructRetAttr())
591 const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
592 uint64_t destSize = TD->getTypeAllocSize(StructTy);
594 if (destSize < srcSize)
600 // Check that src is not accessed except via the call and the memcpy. This
601 // guarantees that it holds only undefined values when passed in (so the final
602 // memcpy can be dropped), that it is not read or written between the call and
603 // the memcpy, and that writing beyond the end of it is undefined.
604 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
605 srcAlloca->use_end());
606 while (!srcUseList.empty()) {
607 User *UI = srcUseList.pop_back_val();
609 if (isa<BitCastInst>(UI)) {
610 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
612 srcUseList.push_back(*I);
613 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
614 if (G->hasAllZeroIndices())
615 for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
617 srcUseList.push_back(*I);
620 } else if (UI != C && UI != cpy) {
625 // Since we're changing the parameter to the callsite, we need to make sure
626 // that what would be the new parameter dominates the callsite.
627 DominatorTree &DT = getAnalysis<DominatorTree>();
628 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
629 if (!DT.dominates(cpyDestInst, C))
632 // In addition to knowing that the call does not access src in some
633 // unexpected manner, for example via a global, which we deduce from
634 // the use analysis, we also need to know that it does not sneakily
635 // access dest. We rely on AA to figure this out for us.
636 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
637 if (AA.getModRefInfo(C, cpyDest, srcSize) !=
638 AliasAnalysis::NoModRef)
641 // All the checks have passed, so do the transformation.
642 bool changedArgument = false;
643 for (unsigned i = 0; i < CS.arg_size(); ++i)
644 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
645 if (cpySrc->getType() != cpyDest->getType())
646 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
647 cpyDest->getName(), C);
648 changedArgument = true;
649 if (CS.getArgument(i)->getType() == cpyDest->getType())
650 CS.setArgument(i, cpyDest);
652 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
653 CS.getArgument(i)->getType(), cpyDest->getName(), C));
656 if (!changedArgument)
659 // Drop any cached information about the call, because we may have changed
660 // its dependence information by changing its parameter.
661 MD->removeInstruction(C);
663 // Remove the memcpy.
664 MD->removeInstruction(cpy);
670 /// processMemCpyMemCpyDependence - We've found that the (upward scanning)
671 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
672 /// copy from MDep's input if we can. MSize is the size of M's copy.
674 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
676 // We can only transforms memcpy's where the dest of one is the source of the
678 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
681 // Second, the length of the memcpy's must be the same, or the preceeding one
682 // must be larger than the following one.
683 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
684 if (!C1) return false;
686 uint64_t DepSize = C1->getValue().getZExtValue();
690 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
692 // Verify that the copied-from memory doesn't change in between the two
693 // transfers. For example, in:
697 // It would be invalid to transform the second memcpy into memcpy(c <- b).
699 // TODO: If the code between M and MDep is transparent to the destination "c",
700 // then we could still perform the xform by moving M up to the first memcpy.
702 // NOTE: This is conservative, it will stop on any read from the source loc,
703 // not just the defining memcpy.
704 MemDepResult SourceDep =
705 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
706 false, M, M->getParent());
707 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
710 // If the dest of the second might alias the source of the first, then the
711 // source and dest might overlap. We still want to eliminate the intermediate
712 // value, but we have to generate a memmove instead of memcpy.
713 Intrinsic::ID ResultFn = Intrinsic::memcpy;
714 if (!AA.isNoAlias(M->getRawDest(), MSize, MDep->getRawSource(), DepSize))
715 ResultFn = Intrinsic::memmove;
717 // If all checks passed, then we can transform M.
718 const Type *ArgTys[3] = {
719 M->getRawDest()->getType(),
720 MDep->getRawSource()->getType(),
721 M->getLength()->getType()
723 Function *MemCpyFun =
724 Intrinsic::getDeclaration(MDep->getParent()->getParent()->getParent(),
725 ResultFn, ArgTys, 3);
727 // Make sure to use the lesser of the alignment of the source and the dest
728 // since we're changing where we're reading from, but don't want to increase
729 // the alignment past what can be read from or written to.
730 // TODO: Is this worth it if we're creating a less aligned memcpy? For
731 // example we could be moving from movaps -> movq on x86.
732 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
735 MDep->getRawSource(),
737 ConstantInt::get(Type::getInt32Ty(MemCpyFun->getContext()), Align),
740 CallInst::Create(MemCpyFun, Args, Args+5, "", M);
742 // Remove the instruction we're replacing.
743 MD->removeInstruction(M);
744 M->eraseFromParent();
750 /// processMemCpy - perform simplification of memcpy's. If we have memcpy A
751 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
752 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
753 /// circumstances). This allows later passes to remove the first memcpy
755 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
756 // We can only optimize statically-sized memcpy's that are non-volatile.
757 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
758 if (CopySize == 0 || M->isVolatile()) return false;
760 // The are two possible optimizations we can do for memcpy:
761 // a) memcpy-memcpy xform which exposes redundance for DSE.
762 // b) call-memcpy xform for return slot optimization.
763 MemDepResult DepInfo = MD->getDependency(M);
764 if (!DepInfo.isClobber())
767 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
768 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
770 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
771 bool changed = performCallSlotOptzn(M, M->getDest(), M->getSource(),
772 CopySize->getZExtValue(), C);
773 if (changed) M->eraseFromParent();
779 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
780 /// are guaranteed not to alias.
781 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
782 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
784 // If the memmove is a constant size, use it for the alias query, this allows
785 // us to optimize things like: memmove(P, P+64, 64);
786 uint64_t MemMoveSize = AliasAnalysis::UnknownSize;
787 if (ConstantInt *Len = dyn_cast<ConstantInt>(M->getLength()))
788 MemMoveSize = Len->getZExtValue();
790 // See if the pointers alias.
791 if (AA.alias(M->getRawDest(), MemMoveSize, M->getRawSource(), MemMoveSize) !=
792 AliasAnalysis::NoAlias)
795 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
797 // If not, then we know we can transform this.
798 Module *Mod = M->getParent()->getParent()->getParent();
799 const Type *ArgTys[3] = { M->getRawDest()->getType(),
800 M->getRawSource()->getType(),
801 M->getLength()->getType() };
802 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
805 // MemDep may have over conservative information about this instruction, just
806 // conservatively flush it from the cache.
807 MD->removeInstruction(M);
813 /// processByValArgument - This is called on every byval argument in call sites.
814 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
815 TargetData *TD = getAnalysisIfAvailable<TargetData>();
816 if (!TD) return false;
818 // Find out what feeds this byval argument.
819 Value *ByValArg = CS.getArgument(ArgNo);
820 uint64_t ByValSize = TD->getTypeAllocSize(ByValArg->getType());
821 MemDepResult DepInfo =
822 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
823 true, CS.getInstruction(),
824 CS.getInstruction()->getParent());
825 if (!DepInfo.isClobber())
828 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
829 // a memcpy, see if we can byval from the source of the memcpy instead of the
831 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
832 if (MDep == 0 || MDep->isVolatile() ||
833 ByValArg->stripPointerCasts() != MDep->getDest())
836 // The length of the memcpy must be larger or equal to the size of the byval.
837 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
838 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
841 // Get the alignment of the byval. If it is greater than the memcpy, then we
842 // can't do the substitution. If the call doesn't specify the alignment, then
843 // it is some target specific value that we can't know.
844 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
845 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign)
848 // Verify that the copied-from memory doesn't change in between the memcpy and
853 // It would be invalid to transform the second memcpy into foo(*b).
855 // NOTE: This is conservative, it will stop on any read from the source loc,
856 // not just the defining memcpy.
857 MemDepResult SourceDep =
858 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
859 false, CS.getInstruction(), MDep->getParent());
860 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
863 Value *TmpCast = MDep->getSource();
864 if (MDep->getSource()->getType() != ByValArg->getType())
865 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
866 "tmpcast", CS.getInstruction());
868 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
869 << " " << *MDep << "\n"
870 << " " << *CS.getInstruction() << "\n");
872 // Otherwise we're good! Update the byval argument.
873 CS.setArgument(ArgNo, TmpCast);
878 /// iterateOnFunction - Executes one iteration of MemCpyOpt.
879 bool MemCpyOpt::iterateOnFunction(Function &F) {
880 bool MadeChange = false;
882 // Walk all instruction in the function.
883 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
884 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
885 // Avoid invalidating the iterator.
886 Instruction *I = BI++;
888 bool RepeatInstruction = false;
890 if (StoreInst *SI = dyn_cast<StoreInst>(I))
891 MadeChange |= processStore(SI, BI);
892 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) {
893 RepeatInstruction = processMemCpy(M);
894 } else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) {
895 RepeatInstruction = processMemMove(M);
896 } else if (CallSite CS = (Value*)I) {
897 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
898 if (CS.paramHasAttr(i+1, Attribute::ByVal))
899 MadeChange |= processByValArgument(CS, i);
902 // Reprocess the instruction if desired.
903 if (RepeatInstruction) {
913 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a
916 bool MemCpyOpt::runOnFunction(Function &F) {
917 bool MadeChange = false;
918 MD = &getAnalysis<MemoryDependenceAnalysis>();
920 if (!iterateOnFunction(F))