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 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Analysis/AssumptionCache.h"
20 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalVariable.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Transforms/Utils/Local.h"
36 #define DEBUG_TYPE "memcpyopt"
38 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
39 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
40 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
41 STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
43 static int64_t GetOffsetFromIndex(const GEPOperator *GEP, unsigned Idx,
44 bool &VariableIdxFound,
45 const DataLayout &DL) {
46 // Skip over the first indices.
47 gep_type_iterator GTI = gep_type_begin(GEP);
48 for (unsigned i = 1; i != Idx; ++i, ++GTI)
51 // Compute the offset implied by the rest of the indices.
53 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
54 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
56 return VariableIdxFound = true;
57 if (OpC->isZero()) continue; // No offset.
59 // Handle struct indices, which add their field offset to the pointer.
60 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
61 Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
65 // Otherwise, we have a sequential type like an array or vector. Multiply
66 // the index by the ElementSize.
67 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
68 Offset += Size*OpC->getSExtValue();
74 /// Return true if Ptr1 is provably equal to Ptr2 plus a constant offset, and
75 /// return that constant offset. For example, Ptr1 might be &A[42], and Ptr2
76 /// might be &A[40]. In this case offset would be -8.
77 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
78 const DataLayout &DL) {
79 Ptr1 = Ptr1->stripPointerCasts();
80 Ptr2 = Ptr2->stripPointerCasts();
82 // Handle the trivial case first.
88 GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1);
89 GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2);
91 bool VariableIdxFound = false;
93 // If one pointer is a GEP and the other isn't, then see if the GEP is a
94 // constant offset from the base, as in "P" and "gep P, 1".
95 if (GEP1 && !GEP2 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
96 Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, DL);
97 return !VariableIdxFound;
100 if (GEP2 && !GEP1 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
101 Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, DL);
102 return !VariableIdxFound;
105 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
106 // base. After that base, they may have some number of common (and
107 // potentially variable) indices. After that they handle some constant
108 // offset, which determines their offset from each other. At this point, we
109 // handle no other case.
110 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
113 // Skip any common indices and track the GEP types.
115 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
116 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
119 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, DL);
120 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, DL);
121 if (VariableIdxFound) return false;
123 Offset = Offset2-Offset1;
128 /// Represents a range of memset'd bytes with the ByteVal value.
129 /// This allows us to analyze stores like:
134 /// which sometimes happens with stores to arrays of structs etc. When we see
135 /// the first store, we make a range [1, 2). The second store extends the range
136 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
137 /// two ranges into [0, 3) which is memset'able.
140 // Start/End - A semi range that describes the span that this range covers.
141 // The range is closed at the start and open at the end: [Start, End).
144 /// StartPtr - The getelementptr instruction that points to the start of the
148 /// Alignment - The known alignment of the first store.
151 /// TheStores - The actual stores that make up this range.
152 SmallVector<Instruction*, 16> TheStores;
154 bool isProfitableToUseMemset(const DataLayout &DL) const;
156 } // end anon namespace
158 bool MemsetRange::isProfitableToUseMemset(const DataLayout &DL) const {
159 // If we found more than 4 stores to merge or 16 bytes, use memset.
160 if (TheStores.size() >= 4 || End-Start >= 16) return true;
162 // If there is nothing to merge, don't do anything.
163 if (TheStores.size() < 2) return false;
165 // If any of the stores are a memset, then it is always good to extend the
167 for (unsigned i = 0, e = TheStores.size(); i != e; ++i)
168 if (!isa<StoreInst>(TheStores[i]))
171 // Assume that the code generator is capable of merging pairs of stores
172 // together if it wants to.
173 if (TheStores.size() == 2) return false;
175 // If we have fewer than 8 stores, it can still be worthwhile to do this.
176 // For example, merging 4 i8 stores into an i32 store is useful almost always.
177 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
178 // memset will be split into 2 32-bit stores anyway) and doing so can
179 // pessimize the llvm optimizer.
181 // Since we don't have perfect knowledge here, make some assumptions: assume
182 // the maximum GPR width is the same size as the largest legal integer
183 // size. If so, check to see whether we will end up actually reducing the
184 // number of stores used.
185 unsigned Bytes = unsigned(End-Start);
186 unsigned MaxIntSize = DL.getLargestLegalIntTypeSize();
189 unsigned NumPointerStores = Bytes / MaxIntSize;
191 // Assume the remaining bytes if any are done a byte at a time.
192 unsigned NumByteStores = Bytes - NumPointerStores * MaxIntSize;
194 // If we will reduce the # stores (according to this heuristic), do the
195 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
197 return TheStores.size() > NumPointerStores+NumByteStores;
203 /// A sorted list of the memset ranges.
204 SmallVector<MemsetRange, 8> Ranges;
205 typedef SmallVectorImpl<MemsetRange>::iterator range_iterator;
206 const DataLayout &DL;
208 MemsetRanges(const DataLayout &DL) : DL(DL) {}
210 typedef SmallVectorImpl<MemsetRange>::const_iterator const_iterator;
211 const_iterator begin() const { return Ranges.begin(); }
212 const_iterator end() const { return Ranges.end(); }
213 bool empty() const { return Ranges.empty(); }
215 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
216 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
217 addStore(OffsetFromFirst, SI);
219 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
222 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
223 int64_t StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType());
225 addRange(OffsetFromFirst, StoreSize,
226 SI->getPointerOperand(), SI->getAlignment(), SI);
229 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
230 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
231 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI);
234 void addRange(int64_t Start, int64_t Size, Value *Ptr,
235 unsigned Alignment, Instruction *Inst);
239 } // end anon namespace
242 /// Add a new store to the MemsetRanges data structure. This adds a
243 /// new range for the specified store at the specified offset, merging into
244 /// existing ranges as appropriate.
245 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
246 unsigned Alignment, Instruction *Inst) {
247 int64_t End = Start+Size;
249 range_iterator I = std::lower_bound(Ranges.begin(), Ranges.end(), Start,
250 [](const MemsetRange &LHS, int64_t RHS) { return LHS.End < RHS; });
252 // We now know that I == E, in which case we didn't find anything to merge
253 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
254 // to insert a new range. Handle this now.
255 if (I == Ranges.end() || End < I->Start) {
256 MemsetRange &R = *Ranges.insert(I, MemsetRange());
260 R.Alignment = Alignment;
261 R.TheStores.push_back(Inst);
265 // This store overlaps with I, add it.
266 I->TheStores.push_back(Inst);
268 // At this point, we may have an interval that completely contains our store.
269 // If so, just add it to the interval and return.
270 if (I->Start <= Start && I->End >= End)
273 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
274 // but is not entirely contained within the range.
276 // See if the range extends the start of the range. In this case, it couldn't
277 // possibly cause it to join the prior range, because otherwise we would have
279 if (Start < I->Start) {
282 I->Alignment = Alignment;
285 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
286 // is in or right at the end of I), and that End >= I->Start. Extend I out to
290 range_iterator NextI = I;
291 while (++NextI != Ranges.end() && End >= NextI->Start) {
292 // Merge the range in.
293 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
294 if (NextI->End > I->End)
302 //===----------------------------------------------------------------------===//
304 //===----------------------------------------------------------------------===//
307 class MemCpyOpt : public FunctionPass {
308 MemoryDependenceAnalysis *MD;
309 TargetLibraryInfo *TLI;
311 static char ID; // Pass identification, replacement for typeid
312 MemCpyOpt() : FunctionPass(ID) {
313 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
318 bool runOnFunction(Function &F) override;
321 // This transformation requires dominator postdominator info
322 void getAnalysisUsage(AnalysisUsage &AU) const override {
323 AU.setPreservesCFG();
324 AU.addRequired<AssumptionCacheTracker>();
325 AU.addRequired<DominatorTreeWrapperPass>();
326 AU.addRequired<MemoryDependenceAnalysis>();
327 AU.addRequired<AliasAnalysis>();
328 AU.addRequired<TargetLibraryInfoWrapperPass>();
329 AU.addPreserved<AliasAnalysis>();
330 AU.addPreserved<MemoryDependenceAnalysis>();
334 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
335 bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI);
336 bool processMemCpy(MemCpyInst *M);
337 bool processMemMove(MemMoveInst *M);
338 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
339 uint64_t cpyLen, unsigned cpyAlign, CallInst *C);
340 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep);
341 bool processMemSetMemCpyDependence(MemCpyInst *M, MemSetInst *MDep);
342 bool performMemCpyToMemSetOptzn(MemCpyInst *M, MemSetInst *MDep);
343 bool processByValArgument(CallSite CS, unsigned ArgNo);
344 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr,
347 bool iterateOnFunction(Function &F);
350 char MemCpyOpt::ID = 0;
353 /// The public interface to this file...
354 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
356 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
358 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
359 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
360 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
361 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
362 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
363 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
366 /// When scanning forward over instructions, we look for some other patterns to
367 /// fold away. In particular, this looks for stores to neighboring locations of
368 /// memory. If it sees enough consecutive ones, it attempts to merge them
369 /// together into a memcpy/memset.
370 Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
371 Value *StartPtr, Value *ByteVal) {
372 const DataLayout &DL = StartInst->getModule()->getDataLayout();
374 // Okay, so we now have a single store that can be splatable. Scan to find
375 // all subsequent stores of the same value to offset from the same pointer.
376 // Join these together into ranges, so we can decide whether contiguous blocks
378 MemsetRanges Ranges(DL);
380 BasicBlock::iterator BI = StartInst;
381 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
382 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
383 // If the instruction is readnone, ignore it, otherwise bail out. We
384 // don't even allow readonly here because we don't want something like:
385 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
386 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
391 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
392 // If this is a store, see if we can merge it in.
393 if (!NextStore->isSimple()) break;
395 // Check to see if this stored value is of the same byte-splattable value.
396 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
399 // Check to see if this store is to a constant offset from the start ptr.
401 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset,
405 Ranges.addStore(Offset, NextStore);
407 MemSetInst *MSI = cast<MemSetInst>(BI);
409 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
410 !isa<ConstantInt>(MSI->getLength()))
413 // Check to see if this store is to a constant offset from the start ptr.
415 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, DL))
418 Ranges.addMemSet(Offset, MSI);
422 // If we have no ranges, then we just had a single store with nothing that
423 // could be merged in. This is a very common case of course.
427 // If we had at least one store that could be merged in, add the starting
428 // store as well. We try to avoid this unless there is at least something
429 // interesting as a small compile-time optimization.
430 Ranges.addInst(0, StartInst);
432 // If we create any memsets, we put it right before the first instruction that
433 // isn't part of the memset block. This ensure that the memset is dominated
434 // by any addressing instruction needed by the start of the block.
435 IRBuilder<> Builder(BI);
437 // Now that we have full information about ranges, loop over the ranges and
438 // emit memset's for anything big enough to be worthwhile.
439 Instruction *AMemSet = nullptr;
440 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
442 const MemsetRange &Range = *I;
444 if (Range.TheStores.size() == 1) continue;
446 // If it is profitable to lower this range to memset, do so now.
447 if (!Range.isProfitableToUseMemset(DL))
450 // Otherwise, we do want to transform this! Create a new memset.
451 // Get the starting pointer of the block.
452 StartPtr = Range.StartPtr;
454 // Determine alignment
455 unsigned Alignment = Range.Alignment;
456 if (Alignment == 0) {
458 cast<PointerType>(StartPtr->getType())->getElementType();
459 Alignment = DL.getABITypeAlignment(EltType);
463 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
465 DEBUG(dbgs() << "Replace stores:\n";
466 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
467 dbgs() << *Range.TheStores[i] << '\n';
468 dbgs() << "With: " << *AMemSet << '\n');
470 if (!Range.TheStores.empty())
471 AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
473 // Zap all the stores.
474 for (SmallVectorImpl<Instruction *>::const_iterator
475 SI = Range.TheStores.begin(),
476 SE = Range.TheStores.end(); SI != SE; ++SI) {
477 MD->removeInstruction(*SI);
478 (*SI)->eraseFromParent();
487 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
488 if (!SI->isSimple()) return false;
489 const DataLayout &DL = SI->getModule()->getDataLayout();
491 // Detect cases where we're performing call slot forwarding, but
492 // happen to be using a load-store pair to implement it, rather than
494 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
495 if (LI->isSimple() && LI->hasOneUse() &&
496 LI->getParent() == SI->getParent()) {
497 MemDepResult ldep = MD->getDependency(LI);
498 CallInst *C = nullptr;
499 if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
500 C = dyn_cast<CallInst>(ldep.getInst());
503 // Check that nothing touches the dest of the "copy" between
504 // the call and the store.
505 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
506 MemoryLocation StoreLoc = MemoryLocation::get(SI);
507 for (BasicBlock::iterator I = --BasicBlock::iterator(SI),
508 E = C; I != E; --I) {
509 if (AA.getModRefInfo(&*I, StoreLoc) != MRI_NoModRef) {
517 unsigned storeAlign = SI->getAlignment();
519 storeAlign = DL.getABITypeAlignment(SI->getOperand(0)->getType());
520 unsigned loadAlign = LI->getAlignment();
522 loadAlign = DL.getABITypeAlignment(LI->getType());
524 bool changed = performCallSlotOptzn(
525 LI, SI->getPointerOperand()->stripPointerCasts(),
526 LI->getPointerOperand()->stripPointerCasts(),
527 DL.getTypeStoreSize(SI->getOperand(0)->getType()),
528 std::min(storeAlign, loadAlign), C);
530 MD->removeInstruction(SI);
531 SI->eraseFromParent();
532 MD->removeInstruction(LI);
533 LI->eraseFromParent();
541 // There are two cases that are interesting for this code to handle: memcpy
542 // and memset. Right now we only handle memset.
544 // Ensure that the value being stored is something that can be memset'able a
545 // byte at a time like "0" or "-1" or any width, as well as things like
546 // 0xA0A0A0A0 and 0.0.
547 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
548 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
550 BBI = I; // Don't invalidate iterator.
557 bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
558 // See if there is another memset or store neighboring this memset which
559 // allows us to widen out the memset to do a single larger store.
560 if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
561 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
563 BBI = I; // Don't invalidate iterator.
570 /// Takes a memcpy and a call that it depends on,
571 /// and checks for the possibility of a call slot optimization by having
572 /// the call write its result directly into the destination of the memcpy.
573 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
574 Value *cpyDest, Value *cpySrc,
575 uint64_t cpyLen, unsigned cpyAlign,
577 // The general transformation to keep in mind is
579 // call @func(..., src, ...)
580 // memcpy(dest, src, ...)
584 // memcpy(dest, src, ...)
585 // call @func(..., dest, ...)
587 // Since moving the memcpy is technically awkward, we additionally check that
588 // src only holds uninitialized values at the moment of the call, meaning that
589 // the memcpy can be discarded rather than moved.
591 // Deliberately get the source and destination with bitcasts stripped away,
592 // because we'll need to do type comparisons based on the underlying type.
595 // Require that src be an alloca. This simplifies the reasoning considerably.
596 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
600 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
604 const DataLayout &DL = cpy->getModule()->getDataLayout();
605 uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) *
606 srcArraySize->getZExtValue();
608 if (cpyLen < srcSize)
611 // Check that accessing the first srcSize bytes of dest will not cause a
612 // trap. Otherwise the transform is invalid since it might cause a trap
613 // to occur earlier than it otherwise would.
614 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
615 // The destination is an alloca. Check it is larger than srcSize.
616 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
620 uint64_t destSize = DL.getTypeAllocSize(A->getAllocatedType()) *
621 destArraySize->getZExtValue();
623 if (destSize < srcSize)
625 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
626 if (A->getDereferenceableBytes() < srcSize) {
627 // If the destination is an sret parameter then only accesses that are
628 // outside of the returned struct type can trap.
629 if (!A->hasStructRetAttr())
632 Type *StructTy = cast<PointerType>(A->getType())->getElementType();
633 if (!StructTy->isSized()) {
634 // The call may never return and hence the copy-instruction may never
635 // be executed, and therefore it's not safe to say "the destination
636 // has at least <cpyLen> bytes, as implied by the copy-instruction",
640 uint64_t destSize = DL.getTypeAllocSize(StructTy);
641 if (destSize < srcSize)
648 // Check that dest points to memory that is at least as aligned as src.
649 unsigned srcAlign = srcAlloca->getAlignment();
651 srcAlign = DL.getABITypeAlignment(srcAlloca->getAllocatedType());
652 bool isDestSufficientlyAligned = srcAlign <= cpyAlign;
653 // If dest is not aligned enough and we can't increase its alignment then
655 if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest))
658 // Check that src is not accessed except via the call and the memcpy. This
659 // guarantees that it holds only undefined values when passed in (so the final
660 // memcpy can be dropped), that it is not read or written between the call and
661 // the memcpy, and that writing beyond the end of it is undefined.
662 SmallVector<User*, 8> srcUseList(srcAlloca->user_begin(),
663 srcAlloca->user_end());
664 while (!srcUseList.empty()) {
665 User *U = srcUseList.pop_back_val();
667 if (isa<BitCastInst>(U) || isa<AddrSpaceCastInst>(U)) {
668 for (User *UU : U->users())
669 srcUseList.push_back(UU);
672 if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(U)) {
673 if (!G->hasAllZeroIndices())
676 for (User *UU : U->users())
677 srcUseList.push_back(UU);
680 if (const IntrinsicInst *IT = dyn_cast<IntrinsicInst>(U))
681 if (IT->getIntrinsicID() == Intrinsic::lifetime_start ||
682 IT->getIntrinsicID() == Intrinsic::lifetime_end)
685 if (U != C && U != cpy)
689 // Check that src isn't captured by the called function since the
690 // transformation can cause aliasing issues in that case.
691 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
692 if (CS.getArgument(i) == cpySrc && !CS.doesNotCapture(i))
695 // Since we're changing the parameter to the callsite, we need to make sure
696 // that what would be the new parameter dominates the callsite.
697 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
698 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
699 if (!DT.dominates(cpyDestInst, C))
702 // In addition to knowing that the call does not access src in some
703 // unexpected manner, for example via a global, which we deduce from
704 // the use analysis, we also need to know that it does not sneakily
705 // access dest. We rely on AA to figure this out for us.
706 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
707 ModRefInfo MR = AA.getModRefInfo(C, cpyDest, srcSize);
708 // If necessary, perform additional analysis.
709 if (MR != MRI_NoModRef)
710 MR = AA.callCapturesBefore(C, cpyDest, srcSize, &DT);
711 if (MR != MRI_NoModRef)
714 // All the checks have passed, so do the transformation.
715 bool changedArgument = false;
716 for (unsigned i = 0; i < CS.arg_size(); ++i)
717 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
718 Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest
719 : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
720 cpyDest->getName(), C);
721 changedArgument = true;
722 if (CS.getArgument(i)->getType() == Dest->getType())
723 CS.setArgument(i, Dest);
725 CS.setArgument(i, CastInst::CreatePointerCast(Dest,
726 CS.getArgument(i)->getType(), Dest->getName(), C));
729 if (!changedArgument)
732 // If the destination wasn't sufficiently aligned then increase its alignment.
733 if (!isDestSufficientlyAligned) {
734 assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!");
735 cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
738 // Drop any cached information about the call, because we may have changed
739 // its dependence information by changing its parameter.
740 MD->removeInstruction(C);
742 // Update AA metadata
743 // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
744 // handled here, but combineMetadata doesn't support them yet
745 unsigned KnownIDs[] = {
747 LLVMContext::MD_tbaa,
748 LLVMContext::MD_alias_scope,
749 LLVMContext::MD_noalias,
751 combineMetadata(C, cpy, KnownIDs);
753 // Remove the memcpy.
754 MD->removeInstruction(cpy);
760 /// We've found that the (upward scanning) memory dependence of memcpy 'M' is
761 /// the memcpy 'MDep'. Try to simplify M to copy from MDep's input if we can.
762 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep) {
763 // We can only transforms memcpy's where the dest of one is the source of the
765 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
768 // If dep instruction is reading from our current input, then it is a noop
769 // transfer and substituting the input won't change this instruction. Just
770 // ignore the input and let someone else zap MDep. This handles cases like:
773 if (M->getSource() == MDep->getSource())
776 // Second, the length of the memcpy's must be the same, or the preceding one
777 // must be larger than the following one.
778 ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
779 ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength());
780 if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
783 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
785 // Verify that the copied-from memory doesn't change in between the two
786 // transfers. For example, in:
790 // It would be invalid to transform the second memcpy into memcpy(c <- b).
792 // TODO: If the code between M and MDep is transparent to the destination "c",
793 // then we could still perform the xform by moving M up to the first memcpy.
795 // NOTE: This is conservative, it will stop on any read from the source loc,
796 // not just the defining memcpy.
797 MemDepResult SourceDep = MD->getPointerDependencyFrom(
798 MemoryLocation::getForSource(MDep), false, M, M->getParent());
799 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
802 // If the dest of the second might alias the source of the first, then the
803 // source and dest might overlap. We still want to eliminate the intermediate
804 // value, but we have to generate a memmove instead of memcpy.
805 bool UseMemMove = false;
806 if (!AA.isNoAlias(MemoryLocation::getForDest(M),
807 MemoryLocation::getForSource(MDep)))
810 // If all checks passed, then we can transform M.
812 // Make sure to use the lesser of the alignment of the source and the dest
813 // since we're changing where we're reading from, but don't want to increase
814 // the alignment past what can be read from or written to.
815 // TODO: Is this worth it if we're creating a less aligned memcpy? For
816 // example we could be moving from movaps -> movq on x86.
817 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
819 IRBuilder<> Builder(M);
821 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
822 Align, M->isVolatile());
824 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
825 Align, M->isVolatile());
827 // Remove the instruction we're replacing.
828 MD->removeInstruction(M);
829 M->eraseFromParent();
834 /// We've found that the (upward scanning) memory dependence of \p MemCpy is
835 /// \p MemSet. Try to simplify \p MemSet to only set the trailing bytes that
836 /// weren't copied over by \p MemCpy.
838 /// In other words, transform:
840 /// memset(dst, c, dst_size);
841 /// memcpy(dst, src, src_size);
845 /// memcpy(dst, src, src_size);
846 /// memset(dst + src_size, c, dst_size <= src_size ? 0 : dst_size - src_size);
848 bool MemCpyOpt::processMemSetMemCpyDependence(MemCpyInst *MemCpy,
849 MemSetInst *MemSet) {
850 // We can only transform memset/memcpy with the same destination.
851 if (MemSet->getDest() != MemCpy->getDest())
854 // Check that there are no other dependencies on the memset destination.
855 MemDepResult DstDepInfo = MD->getPointerDependencyFrom(
856 MemoryLocation::getForDest(MemSet), false, MemCpy, MemCpy->getParent());
857 if (DstDepInfo.getInst() != MemSet)
860 // Use the same i8* dest as the memcpy, killing the memset dest if different.
861 Value *Dest = MemCpy->getRawDest();
862 Value *DestSize = MemSet->getLength();
863 Value *SrcSize = MemCpy->getLength();
865 // By default, create an unaligned memset.
867 // If Dest is aligned, and SrcSize is constant, use the minimum alignment
869 const unsigned DestAlign =
870 std::max(MemSet->getAlignment(), MemCpy->getAlignment());
872 if (ConstantInt *SrcSizeC = dyn_cast<ConstantInt>(SrcSize))
873 Align = MinAlign(SrcSizeC->getZExtValue(), DestAlign);
875 IRBuilder<> Builder(MemCpy);
877 // If the sizes have different types, zext the smaller one.
878 if (DestSize->getType() != SrcSize->getType()) {
879 if (DestSize->getType()->getIntegerBitWidth() >
880 SrcSize->getType()->getIntegerBitWidth())
881 SrcSize = Builder.CreateZExt(SrcSize, DestSize->getType());
883 DestSize = Builder.CreateZExt(DestSize, SrcSize->getType());
887 Builder.CreateSelect(Builder.CreateICmpULE(DestSize, SrcSize),
888 ConstantInt::getNullValue(DestSize->getType()),
889 Builder.CreateSub(DestSize, SrcSize));
890 Builder.CreateMemSet(Builder.CreateGEP(Dest, SrcSize), MemSet->getOperand(1),
893 MD->removeInstruction(MemSet);
894 MemSet->eraseFromParent();
898 /// Transform memcpy to memset when its source was just memset.
899 /// In other words, turn:
901 /// memset(dst1, c, dst1_size);
902 /// memcpy(dst2, dst1, dst2_size);
906 /// memset(dst1, c, dst1_size);
907 /// memset(dst2, c, dst2_size);
909 /// When dst2_size <= dst1_size.
911 /// The \p MemCpy must have a Constant length.
912 bool MemCpyOpt::performMemCpyToMemSetOptzn(MemCpyInst *MemCpy,
913 MemSetInst *MemSet) {
914 // This only makes sense on memcpy(..., memset(...), ...).
915 if (MemSet->getRawDest() != MemCpy->getRawSource())
918 ConstantInt *CopySize = cast<ConstantInt>(MemCpy->getLength());
919 ConstantInt *MemSetSize = dyn_cast<ConstantInt>(MemSet->getLength());
920 // Make sure the memcpy doesn't read any more than what the memset wrote.
921 // Don't worry about sizes larger than i64.
922 if (!MemSetSize || CopySize->getZExtValue() > MemSetSize->getZExtValue())
925 IRBuilder<> Builder(MemCpy);
926 Builder.CreateMemSet(MemCpy->getRawDest(), MemSet->getOperand(1),
927 CopySize, MemCpy->getAlignment());
931 /// Perform simplification of memcpy's. If we have memcpy A
932 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
933 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
934 /// circumstances). This allows later passes to remove the first memcpy
936 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
937 // We can only optimize non-volatile memcpy's.
938 if (M->isVolatile()) return false;
940 // If the source and destination of the memcpy are the same, then zap it.
941 if (M->getSource() == M->getDest()) {
942 MD->removeInstruction(M);
943 M->eraseFromParent();
947 // If copying from a constant, try to turn the memcpy into a memset.
948 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
949 if (GV->isConstant() && GV->hasDefinitiveInitializer())
950 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
951 IRBuilder<> Builder(M);
952 Builder.CreateMemSet(M->getRawDest(), ByteVal, M->getLength(),
953 M->getAlignment(), false);
954 MD->removeInstruction(M);
955 M->eraseFromParent();
960 MemDepResult DepInfo = MD->getDependency(M);
962 // Try to turn a partially redundant memset + memcpy into
963 // memcpy + smaller memset. We don't need the memcpy size for this.
964 if (DepInfo.isClobber())
965 if (MemSetInst *MDep = dyn_cast<MemSetInst>(DepInfo.getInst()))
966 if (processMemSetMemCpyDependence(M, MDep))
969 // The optimizations after this point require the memcpy size.
970 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
971 if (!CopySize) return false;
973 // There are four possible optimizations we can do for memcpy:
974 // a) memcpy-memcpy xform which exposes redundance for DSE.
975 // b) call-memcpy xform for return slot optimization.
976 // c) memcpy from freshly alloca'd space or space that has just started its
977 // lifetime copies undefined data, and we can therefore eliminate the
978 // memcpy in favor of the data that was already at the destination.
979 // d) memcpy from a just-memset'd source can be turned into memset.
980 if (DepInfo.isClobber()) {
981 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
982 if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
983 CopySize->getZExtValue(), M->getAlignment(),
985 MD->removeInstruction(M);
986 M->eraseFromParent();
992 MemoryLocation SrcLoc = MemoryLocation::getForSource(M);
993 MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(SrcLoc, true,
996 if (SrcDepInfo.isClobber()) {
997 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))
998 return processMemCpyMemCpyDependence(M, MDep);
999 } else if (SrcDepInfo.isDef()) {
1000 Instruction *I = SrcDepInfo.getInst();
1001 bool hasUndefContents = false;
1003 if (isa<AllocaInst>(I)) {
1004 hasUndefContents = true;
1005 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1006 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1007 if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))
1008 if (LTSize->getZExtValue() >= CopySize->getZExtValue())
1009 hasUndefContents = true;
1012 if (hasUndefContents) {
1013 MD->removeInstruction(M);
1014 M->eraseFromParent();
1020 if (SrcDepInfo.isClobber())
1021 if (MemSetInst *MDep = dyn_cast<MemSetInst>(SrcDepInfo.getInst()))
1022 if (performMemCpyToMemSetOptzn(M, MDep)) {
1023 MD->removeInstruction(M);
1024 M->eraseFromParent();
1032 /// Transforms memmove calls to memcpy calls when the src/dst are guaranteed
1034 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
1035 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
1037 if (!TLI->has(LibFunc::memmove))
1040 // See if the pointers alias.
1041 if (!AA.isNoAlias(MemoryLocation::getForDest(M),
1042 MemoryLocation::getForSource(M)))
1045 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
1047 // If not, then we know we can transform this.
1048 Module *Mod = M->getParent()->getParent()->getParent();
1049 Type *ArgTys[3] = { M->getRawDest()->getType(),
1050 M->getRawSource()->getType(),
1051 M->getLength()->getType() };
1052 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
1055 // MemDep may have over conservative information about this instruction, just
1056 // conservatively flush it from the cache.
1057 MD->removeInstruction(M);
1063 /// This is called on every byval argument in call sites.
1064 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
1065 const DataLayout &DL = CS.getCaller()->getParent()->getDataLayout();
1066 // Find out what feeds this byval argument.
1067 Value *ByValArg = CS.getArgument(ArgNo);
1068 Type *ByValTy = cast<PointerType>(ByValArg->getType())->getElementType();
1069 uint64_t ByValSize = DL.getTypeAllocSize(ByValTy);
1070 MemDepResult DepInfo = MD->getPointerDependencyFrom(
1071 MemoryLocation(ByValArg, ByValSize), true, CS.getInstruction(),
1072 CS.getInstruction()->getParent());
1073 if (!DepInfo.isClobber())
1076 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
1077 // a memcpy, see if we can byval from the source of the memcpy instead of the
1079 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
1080 if (!MDep || MDep->isVolatile() ||
1081 ByValArg->stripPointerCasts() != MDep->getDest())
1084 // The length of the memcpy must be larger or equal to the size of the byval.
1085 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
1086 if (!C1 || C1->getValue().getZExtValue() < ByValSize)
1089 // Get the alignment of the byval. If the call doesn't specify the alignment,
1090 // then it is some target specific value that we can't know.
1091 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
1092 if (ByValAlign == 0) return false;
1094 // If it is greater than the memcpy, then we check to see if we can force the
1095 // source of the memcpy to the alignment we need. If we fail, we bail out.
1096 AssumptionCache &AC =
1097 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
1098 *CS->getParent()->getParent());
1099 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1100 if (MDep->getAlignment() < ByValAlign &&
1101 getOrEnforceKnownAlignment(MDep->getSource(), ByValAlign, DL,
1102 CS.getInstruction(), &AC, &DT) < ByValAlign)
1105 // Verify that the copied-from memory doesn't change in between the memcpy and
1110 // It would be invalid to transform the second memcpy into foo(*b).
1112 // NOTE: This is conservative, it will stop on any read from the source loc,
1113 // not just the defining memcpy.
1114 MemDepResult SourceDep =
1115 MD->getPointerDependencyFrom(MemoryLocation::getForSource(MDep), false,
1116 CS.getInstruction(), MDep->getParent());
1117 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
1120 Value *TmpCast = MDep->getSource();
1121 if (MDep->getSource()->getType() != ByValArg->getType())
1122 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
1123 "tmpcast", CS.getInstruction());
1125 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
1126 << " " << *MDep << "\n"
1127 << " " << *CS.getInstruction() << "\n");
1129 // Otherwise we're good! Update the byval argument.
1130 CS.setArgument(ArgNo, TmpCast);
1135 /// Executes one iteration of MemCpyOpt.
1136 bool MemCpyOpt::iterateOnFunction(Function &F) {
1137 bool MadeChange = false;
1139 // Walk all instruction in the function.
1140 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
1141 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
1142 // Avoid invalidating the iterator.
1143 Instruction *I = BI++;
1145 bool RepeatInstruction = false;
1147 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1148 MadeChange |= processStore(SI, BI);
1149 else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
1150 RepeatInstruction = processMemSet(M, BI);
1151 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
1152 RepeatInstruction = processMemCpy(M);
1153 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I))
1154 RepeatInstruction = processMemMove(M);
1155 else if (auto CS = CallSite(I)) {
1156 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
1157 if (CS.isByValArgument(i))
1158 MadeChange |= processByValArgument(CS, i);
1161 // Reprocess the instruction if desired.
1162 if (RepeatInstruction) {
1163 if (BI != BB->begin()) --BI;
1172 /// This is the main transformation entry point for a function.
1173 bool MemCpyOpt::runOnFunction(Function &F) {
1174 if (skipOptnoneFunction(F))
1177 bool MadeChange = false;
1178 MD = &getAnalysis<MemoryDependenceAnalysis>();
1179 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1181 // If we don't have at least memset and memcpy, there is little point of doing
1182 // anything here. These are required by a freestanding implementation, so if
1183 // even they are disabled, there is no point in trying hard.
1184 if (!TLI->has(LibFunc::memset) || !TLI->has(LibFunc::memcpy))
1188 if (!iterateOnFunction(F))