//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "memcpyopt"
#include "llvm/Transforms/Scalar.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/Instructions.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/GetElementPtrTypeIterator.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/Debug.h"
-#include "llvm/Support/GetElementPtrTypeIterator.h"
-#include "llvm/Support/IRBuilder.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/Target/TargetData.h"
-#include "llvm/Target/TargetLibraryInfo.h"
+#include "llvm/Transforms/Utils/Local.h"
#include <list>
using namespace llvm;
+#define DEBUG_TYPE "memcpyopt"
+
STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
STATISTIC(NumMemSetInfer, "Number of memsets inferred");
STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
-static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx,
- bool &VariableIdxFound, const TargetData &TD){
+static int64_t GetOffsetFromIndex(const GEPOperator *GEP, unsigned Idx,
+ bool &VariableIdxFound,
+ const DataLayout &DL) {
// Skip over the first indices.
gep_type_iterator GTI = gep_type_begin(GEP);
for (unsigned i = 1; i != Idx; ++i, ++GTI)
/*skip along*/;
-
+
// Compute the offset implied by the rest of the indices.
int64_t Offset = 0;
for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
- if (OpC == 0)
+ if (!OpC)
return VariableIdxFound = true;
if (OpC->isZero()) continue; // No offset.
// Handle struct indices, which add their field offset to the pointer.
- if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
- Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
+ if (StructType *STy = dyn_cast<StructType>(*GTI)) {
+ Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
continue;
}
-
+
// Otherwise, we have a sequential type like an array or vector. Multiply
// the index by the ElementSize.
- uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
+ uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
Offset += Size*OpC->getSExtValue();
}
/// constant offset, and return that constant offset. For example, Ptr1 might
/// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8.
static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
- const TargetData &TD) {
+ const DataLayout &DL) {
Ptr1 = Ptr1->stripPointerCasts();
Ptr2 = Ptr2->stripPointerCasts();
- GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
- GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
-
+
+ // Handle the trivial case first.
+ if (Ptr1 == Ptr2) {
+ Offset = 0;
+ return true;
+ }
+
+ GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1);
+ GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2);
+
bool VariableIdxFound = false;
// If one pointer is a GEP and the other isn't, then see if the GEP is a
// constant offset from the base, as in "P" and "gep P, 1".
- if (GEP1 && GEP2 == 0 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
- Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, TD);
+ if (GEP1 && !GEP2 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
+ Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, DL);
return !VariableIdxFound;
}
- if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
- Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD);
+ if (GEP2 && !GEP1 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
+ Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, DL);
return !VariableIdxFound;
}
-
+
// Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
// base. After that base, they may have some number of common (and
// potentially variable) indices. After that they handle some constant
// handle no other case.
if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
return false;
-
+
// Skip any common indices and track the GEP types.
unsigned Idx = 1;
for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
break;
- int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
- int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
+ int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, DL);
+ int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, DL);
if (VariableIdxFound) return false;
-
+
Offset = Offset2-Offset1;
return true;
}
namespace {
struct MemsetRange {
// Start/End - A semi range that describes the span that this range covers.
- // The range is closed at the start and open at the end: [Start, End).
+ // The range is closed at the start and open at the end: [Start, End).
int64_t Start, End;
/// StartPtr - The getelementptr instruction that points to the start of the
/// range.
Value *StartPtr;
-
+
/// Alignment - The known alignment of the first store.
unsigned Alignment;
-
+
/// TheStores - The actual stores that make up this range.
SmallVector<Instruction*, 16> TheStores;
-
- bool isProfitableToUseMemset(const TargetData &TD) const;
+ bool isProfitableToUseMemset(const DataLayout &DL) const;
};
} // end anon namespace
-bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const {
- // If we found more than 8 stores to merge or 64 bytes, use memset.
- if (TheStores.size() >= 8 || End-Start >= 64) return true;
+bool MemsetRange::isProfitableToUseMemset(const DataLayout &DL) const {
+ // If we found more than 4 stores to merge or 16 bytes, use memset.
+ if (TheStores.size() >= 4 || End-Start >= 16) return true;
// If there is nothing to merge, don't do anything.
if (TheStores.size() < 2) return false;
-
+
// If any of the stores are a memset, then it is always good to extend the
// memset.
for (unsigned i = 0, e = TheStores.size(); i != e; ++i)
if (!isa<StoreInst>(TheStores[i]))
return true;
-
+
// Assume that the code generator is capable of merging pairs of stores
// together if it wants to.
if (TheStores.size() == 2) return false;
-
+
// If we have fewer than 8 stores, it can still be worthwhile to do this.
// For example, merging 4 i8 stores into an i32 store is useful almost always.
// However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
// pessimize the llvm optimizer.
//
// Since we don't have perfect knowledge here, make some assumptions: assume
- // the maximum GPR width is the same size as the pointer size and assume that
- // this width can be stored. If so, check to see whether we will end up
- // actually reducing the number of stores used.
+ // the maximum GPR width is the same size as the largest legal integer
+ // size. If so, check to see whether we will end up actually reducing the
+ // number of stores used.
unsigned Bytes = unsigned(End-Start);
- unsigned NumPointerStores = Bytes/TD.getPointerSize();
-
+ unsigned MaxIntSize = DL.getLargestLegalIntTypeSize();
+ if (MaxIntSize == 0)
+ MaxIntSize = 1;
+ unsigned NumPointerStores = Bytes / MaxIntSize;
+
// Assume the remaining bytes if any are done a byte at a time.
- unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize();
-
+ unsigned NumByteStores = Bytes - NumPointerStores * MaxIntSize;
+
// If we will reduce the # stores (according to this heuristic), do the
// transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
// etc.
return TheStores.size() > NumPointerStores+NumByteStores;
-}
+}
namespace {
/// because each element is relatively large and expensive to copy.
std::list<MemsetRange> Ranges;
typedef std::list<MemsetRange>::iterator range_iterator;
- const TargetData &TD;
+ const DataLayout &DL;
public:
- MemsetRanges(const TargetData &td) : TD(td) {}
-
+ MemsetRanges(const DataLayout &DL) : DL(DL) {}
+
typedef std::list<MemsetRange>::const_iterator const_iterator;
const_iterator begin() const { return Ranges.begin(); }
const_iterator end() const { return Ranges.end(); }
bool empty() const { return Ranges.empty(); }
-
+
void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
addStore(OffsetFromFirst, SI);
}
void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
- int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType());
-
+ int64_t StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType());
+
addRange(OffsetFromFirst, StoreSize,
SI->getPointerOperand(), SI->getAlignment(), SI);
}
-
+
void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI);
}
-
+
void addRange(int64_t Start, int64_t Size, Value *Ptr,
unsigned Alignment, Instruction *Inst);
};
-
+
} // end anon namespace
unsigned Alignment, Instruction *Inst) {
int64_t End = Start+Size;
range_iterator I = Ranges.begin(), E = Ranges.end();
-
+
while (I != E && Start > I->End)
++I;
-
+
// We now know that I == E, in which case we didn't find anything to merge
// with, or that Start <= I->End. If End < I->Start or I == E, then we need
// to insert a new range. Handle this now.
R.TheStores.push_back(Inst);
return;
}
-
+
// This store overlaps with I, add it.
I->TheStores.push_back(Inst);
-
+
// At this point, we may have an interval that completely contains our store.
// If so, just add it to the interval and return.
if (I->Start <= Start && I->End >= End)
return;
-
+
// Now we know that Start <= I->End and End >= I->Start so the range overlaps
// but is not entirely contained within the range.
-
+
// See if the range extends the start of the range. In this case, it couldn't
// possibly cause it to join the prior range, because otherwise we would have
// stopped on *it*.
I->StartPtr = Ptr;
I->Alignment = Alignment;
}
-
+
// Now we know that Start <= I->End and Start >= I->Start (so the startpoint
// is in or right at the end of I), and that End >= I->Start. Extend I out to
// End.
class MemCpyOpt : public FunctionPass {
MemoryDependenceAnalysis *MD;
TargetLibraryInfo *TLI;
- const TargetData *TD;
public:
static char ID; // Pass identification, replacement for typeid
MemCpyOpt() : FunctionPass(ID) {
initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
- MD = 0;
- TLI = 0;
- TD = 0;
+ MD = nullptr;
+ TLI = nullptr;
}
- bool runOnFunction(Function &F);
+ bool runOnFunction(Function &F) override;
private:
// This transformation requires dominator postdominator info
- virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
- AU.addRequired<DominatorTree>();
+ AU.addRequired<AssumptionCacheTracker>();
+ AU.addRequired<DominatorTreeWrapperPass>();
AU.addRequired<MemoryDependenceAnalysis>();
AU.addRequired<AliasAnalysis>();
- AU.addRequired<TargetLibraryInfo>();
+ AU.addRequired<TargetLibraryInfoWrapperPass>();
AU.addPreserved<AliasAnalysis>();
AU.addPreserved<MemoryDependenceAnalysis>();
}
-
+
// Helper fuctions
bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI);
bool processMemCpy(MemCpyInst *M);
bool processMemMove(MemMoveInst *M);
bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
- uint64_t cpyLen, CallInst *C);
+ uint64_t cpyLen, unsigned cpyAlign, CallInst *C);
bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
uint64_t MSize);
bool processByValArgument(CallSite CS, unsigned ArgNo);
bool iterateOnFunction(Function &F);
};
-
+
char MemCpyOpt::ID = 0;
}
INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
false, false)
-INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
-INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
false, false)
/// some other patterns to fold away. In particular, this looks for stores to
/// neighboring locations of memory. If it sees enough consecutive ones, it
/// attempts to merge them together into a memcpy/memset.
-Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
+Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
Value *StartPtr, Value *ByteVal) {
- if (TD == 0) return 0;
-
+ const DataLayout &DL = StartInst->getModule()->getDataLayout();
+
// Okay, so we now have a single store that can be splatable. Scan to find
// all subsequent stores of the same value to offset from the same pointer.
// Join these together into ranges, so we can decide whether contiguous blocks
// are stored.
- MemsetRanges Ranges(*TD);
-
+ MemsetRanges Ranges(DL);
+
BasicBlock::iterator BI = StartInst;
for (++BI; !isa<TerminatorInst>(BI); ++BI) {
if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
break;
continue;
}
-
+
if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
// If this is a store, see if we can merge it in.
- if (NextStore->isVolatile()) break;
-
+ if (!NextStore->isSimple()) break;
+
// Check to see if this stored value is of the same byte-splattable value.
if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
break;
-
+
// Check to see if this store is to a constant offset from the start ptr.
int64_t Offset;
- if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(),
- Offset, *TD))
+ if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset,
+ DL))
break;
-
+
Ranges.addStore(Offset, NextStore);
} else {
MemSetInst *MSI = cast<MemSetInst>(BI);
-
+
if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
!isa<ConstantInt>(MSI->getLength()))
break;
-
+
// Check to see if this store is to a constant offset from the start ptr.
int64_t Offset;
- if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD))
+ if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, DL))
break;
-
+
Ranges.addMemSet(Offset, MSI);
}
}
-
+
// If we have no ranges, then we just had a single store with nothing that
// could be merged in. This is a very common case of course.
if (Ranges.empty())
- return 0;
-
+ return nullptr;
+
// If we had at least one store that could be merged in, add the starting
// store as well. We try to avoid this unless there is at least something
// interesting as a small compile-time optimization.
// Now that we have full information about ranges, loop over the ranges and
// emit memset's for anything big enough to be worthwhile.
- Instruction *AMemSet = 0;
+ Instruction *AMemSet = nullptr;
for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
I != E; ++I) {
const MemsetRange &Range = *I;
-
+
if (Range.TheStores.size() == 1) continue;
-
+
// If it is profitable to lower this range to memset, do so now.
- if (!Range.isProfitableToUseMemset(*TD))
+ if (!Range.isProfitableToUseMemset(DL))
continue;
-
+
// Otherwise, we do want to transform this! Create a new memset.
// Get the starting pointer of the block.
StartPtr = Range.StartPtr;
-
+
// Determine alignment
unsigned Alignment = Range.Alignment;
if (Alignment == 0) {
- const Type *EltType =
+ Type *EltType =
cast<PointerType>(StartPtr->getType())->getElementType();
- Alignment = TD->getABITypeAlignment(EltType);
+ Alignment = DL.getABITypeAlignment(EltType);
}
-
- AMemSet =
+
+ AMemSet =
Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
-
+
DEBUG(dbgs() << "Replace stores:\n";
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
dbgs() << *Range.TheStores[i] << '\n';
AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
// Zap all the stores.
- for (SmallVector<Instruction*, 16>::const_iterator
+ for (SmallVectorImpl<Instruction *>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
MD->removeInstruction(*SI);
}
++NumMemSetInfer;
}
-
+
return AMemSet;
}
bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
- if (SI->isVolatile()) return false;
-
- if (TD == 0) return false;
+ if (!SI->isSimple()) return false;
+ const DataLayout &DL = SI->getModule()->getDataLayout();
// Detect cases where we're performing call slot forwarding, but
// happen to be using a load-store pair to implement it, rather than
// a memcpy.
if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
- if (!LI->isVolatile() && LI->hasOneUse() &&
+ if (LI->isSimple() && LI->hasOneUse() &&
LI->getParent() == SI->getParent()) {
MemDepResult ldep = MD->getDependency(LI);
- CallInst *C = 0;
+ CallInst *C = nullptr;
if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
C = dyn_cast<CallInst>(ldep.getInst());
for (BasicBlock::iterator I = --BasicBlock::iterator(SI),
E = C; I != E; --I) {
if (AA.getModRefInfo(&*I, StoreLoc) != AliasAnalysis::NoModRef) {
- C = 0;
+ C = nullptr;
break;
}
}
}
if (C) {
- bool changed = performCallSlotOptzn(LI,
- SI->getPointerOperand()->stripPointerCasts(),
- LI->getPointerOperand()->stripPointerCasts(),
- TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
+ unsigned storeAlign = SI->getAlignment();
+ if (!storeAlign)
+ storeAlign = DL.getABITypeAlignment(SI->getOperand(0)->getType());
+ unsigned loadAlign = LI->getAlignment();
+ if (!loadAlign)
+ loadAlign = DL.getABITypeAlignment(LI->getType());
+
+ bool changed = performCallSlotOptzn(
+ LI, SI->getPointerOperand()->stripPointerCasts(),
+ LI->getPointerOperand()->stripPointerCasts(),
+ DL.getTypeStoreSize(SI->getOperand(0)->getType()),
+ std::min(storeAlign, loadAlign), C);
if (changed) {
MD->removeInstruction(SI);
SI->eraseFromParent();
}
}
}
-
+
// There are two cases that are interesting for this code to handle: memcpy
// and memset. Right now we only handle memset.
-
+
// Ensure that the value being stored is something that can be memset'able a
// byte at a time like "0" or "-1" or any width, as well as things like
// 0xA0A0A0A0 and 0.0.
BBI = I; // Don't invalidate iterator.
return true;
}
-
+
return false;
}
/// the call write its result directly into the destination of the memcpy.
bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
Value *cpyDest, Value *cpySrc,
- uint64_t cpyLen, CallInst *C) {
+ uint64_t cpyLen, unsigned cpyAlign,
+ CallInst *C) {
// The general transformation to keep in mind is
//
// call @func(..., src, ...)
if (!srcAlloca)
return false;
- // Check that all of src is copied to dest.
- if (TD == 0) return false;
-
ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
if (!srcArraySize)
return false;
- uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
- srcArraySize->getZExtValue();
+ const DataLayout &DL = cpy->getModule()->getDataLayout();
+ uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) *
+ srcArraySize->getZExtValue();
if (cpyLen < srcSize)
return false;
if (!destArraySize)
return false;
- uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
- destArraySize->getZExtValue();
+ uint64_t destSize = DL.getTypeAllocSize(A->getAllocatedType()) *
+ destArraySize->getZExtValue();
if (destSize < srcSize)
return false;
} else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
- // If the destination is an sret parameter then only accesses that are
- // outside of the returned struct type can trap.
- if (!A->hasStructRetAttr())
- return false;
+ if (A->getDereferenceableBytes() < srcSize) {
+ // If the destination is an sret parameter then only accesses that are
+ // outside of the returned struct type can trap.
+ if (!A->hasStructRetAttr())
+ return false;
- const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
- uint64_t destSize = TD->getTypeAllocSize(StructTy);
+ Type *StructTy = cast<PointerType>(A->getType())->getElementType();
+ if (!StructTy->isSized()) {
+ // The call may never return and hence the copy-instruction may never
+ // be executed, and therefore it's not safe to say "the destination
+ // has at least <cpyLen> bytes, as implied by the copy-instruction",
+ return false;
+ }
- if (destSize < srcSize)
- return false;
+ uint64_t destSize = DL.getTypeAllocSize(StructTy);
+ if (destSize < srcSize)
+ return false;
+ }
} else {
return false;
}
+ // Check that dest points to memory that is at least as aligned as src.
+ unsigned srcAlign = srcAlloca->getAlignment();
+ if (!srcAlign)
+ srcAlign = DL.getABITypeAlignment(srcAlloca->getAllocatedType());
+ bool isDestSufficientlyAligned = srcAlign <= cpyAlign;
+ // If dest is not aligned enough and we can't increase its alignment then
+ // bail out.
+ if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest))
+ return false;
+
// Check that src is not accessed except via the call and the memcpy. This
// guarantees that it holds only undefined values when passed in (so the final
// memcpy can be dropped), that it is not read or written between the call and
// the memcpy, and that writing beyond the end of it is undefined.
- SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
- srcAlloca->use_end());
+ SmallVector<User*, 8> srcUseList(srcAlloca->user_begin(),
+ srcAlloca->user_end());
while (!srcUseList.empty()) {
- User *UI = srcUseList.pop_back_val();
-
- if (isa<BitCastInst>(UI)) {
- for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
- I != E; ++I)
- srcUseList.push_back(*I);
- } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
- if (G->hasAllZeroIndices())
- for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
- I != E; ++I)
- srcUseList.push_back(*I);
- else
+ User *U = srcUseList.pop_back_val();
+
+ if (isa<BitCastInst>(U) || isa<AddrSpaceCastInst>(U)) {
+ for (User *UU : U->users())
+ srcUseList.push_back(UU);
+ continue;
+ }
+ if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(U)) {
+ if (!G->hasAllZeroIndices())
return false;
- } else if (UI != C && UI != cpy) {
- return false;
+
+ for (User *UU : U->users())
+ srcUseList.push_back(UU);
+ continue;
}
+ if (const IntrinsicInst *IT = dyn_cast<IntrinsicInst>(U))
+ if (IT->getIntrinsicID() == Intrinsic::lifetime_start ||
+ IT->getIntrinsicID() == Intrinsic::lifetime_end)
+ continue;
+
+ if (U != C && U != cpy)
+ return false;
}
+ // Check that src isn't captured by the called function since the
+ // transformation can cause aliasing issues in that case.
+ for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
+ if (CS.getArgument(i) == cpySrc && !CS.doesNotCapture(i))
+ return false;
+
// Since we're changing the parameter to the callsite, we need to make sure
// that what would be the new parameter dominates the callsite.
- DominatorTree &DT = getAnalysis<DominatorTree>();
+ DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
if (!DT.dominates(cpyDestInst, C))
return false;
// the use analysis, we also need to know that it does not sneakily
// access dest. We rely on AA to figure this out for us.
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
- if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef)
+ AliasAnalysis::ModRefResult MR = AA.getModRefInfo(C, cpyDest, srcSize);
+ // If necessary, perform additional analysis.
+ if (MR != AliasAnalysis::NoModRef)
+ MR = AA.callCapturesBefore(C, cpyDest, srcSize, &DT);
+ if (MR != AliasAnalysis::NoModRef)
return false;
// All the checks have passed, so do the transformation.
bool changedArgument = false;
for (unsigned i = 0; i < CS.arg_size(); ++i)
if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
- if (cpySrc->getType() != cpyDest->getType())
- cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
- cpyDest->getName(), C);
+ Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest
+ : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
+ cpyDest->getName(), C);
changedArgument = true;
- if (CS.getArgument(i)->getType() == cpyDest->getType())
- CS.setArgument(i, cpyDest);
+ if (CS.getArgument(i)->getType() == Dest->getType())
+ CS.setArgument(i, Dest);
else
- CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
- CS.getArgument(i)->getType(), cpyDest->getName(), C));
+ CS.setArgument(i, CastInst::CreatePointerCast(Dest,
+ CS.getArgument(i)->getType(), Dest->getName(), C));
}
if (!changedArgument)
return false;
+ // If the destination wasn't sufficiently aligned then increase its alignment.
+ if (!isDestSufficientlyAligned) {
+ assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!");
+ cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
+ }
+
// Drop any cached information about the call, because we may have changed
// its dependence information by changing its parameter.
MD->removeInstruction(C);
+ // Update AA metadata
+ // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
+ // handled here, but combineMetadata doesn't support them yet
+ unsigned KnownIDs[] = {
+ LLVMContext::MD_tbaa,
+ LLVMContext::MD_alias_scope,
+ LLVMContext::MD_noalias,
+ };
+ combineMetadata(C, cpy, KnownIDs);
+
// Remove the memcpy.
MD->removeInstruction(cpy);
++NumMemCpyInstr;
/// processMemCpyMemCpyDependence - We've found that the (upward scanning)
/// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
/// copy from MDep's input if we can. MSize is the size of M's copy.
-///
+///
bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
uint64_t MSize) {
// We can only transforms memcpy's where the dest of one is the source of the
// other.
if (M->getSource() != MDep->getDest() || MDep->isVolatile())
return false;
-
+
// If dep instruction is reading from our current input, then it is a noop
// transfer and substituting the input won't change this instruction. Just
// ignore the input and let someone else zap MDep. This handles cases like:
// memcpy(b <- a)
if (M->getSource() == MDep->getSource())
return false;
-
+
// Second, the length of the memcpy's must be the same, or the preceding one
// must be larger than the following one.
ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength());
if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
return false;
-
+
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
// Verify that the copied-from memory doesn't change in between the two
false, M, M->getParent());
if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
return false;
-
+
// If the dest of the second might alias the source of the first, then the
// source and dest might overlap. We still want to eliminate the intermediate
// value, but we have to generate a memmove instead of memcpy.
bool UseMemMove = false;
if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)))
UseMemMove = true;
-
+
// If all checks passed, then we can transform M.
-
+
// Make sure to use the lesser of the alignment of the source and the dest
// since we're changing where we're reading from, but don't want to increase
// the alignment past what can be read from or written to.
// TODO: Is this worth it if we're creating a less aligned memcpy? For
// example we could be moving from movaps -> movq on x86.
unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
-
+
IRBuilder<> Builder(M);
if (UseMemMove)
Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
/// circumstances). This allows later passes to remove the first memcpy
/// altogether.
bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
- // We can only optimize statically-sized memcpy's that are non-volatile.
- ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
- if (CopySize == 0 || M->isVolatile()) return false;
+ // We can only optimize non-volatile memcpy's.
+ if (M->isVolatile()) return false;
// If the source and destination of the memcpy are the same, then zap it.
if (M->getSource() == M->getDest()) {
if (GV->isConstant() && GV->hasDefinitiveInitializer())
if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
IRBuilder<> Builder(M);
- Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize,
+ Builder.CreateMemSet(M->getRawDest(), ByteVal, M->getLength(),
M->getAlignment(), false);
MD->removeInstruction(M);
M->eraseFromParent();
return true;
}
- // The are two possible optimizations we can do for memcpy:
+ // The optimizations after this point require the memcpy size.
+ ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
+ if (!CopySize) return false;
+
+ // The are three possible optimizations we can do for memcpy:
// a) memcpy-memcpy xform which exposes redundance for DSE.
// b) call-memcpy xform for return slot optimization.
+ // c) memcpy from freshly alloca'd space or space that has just started its
+ // lifetime copies undefined data, and we can therefore eliminate the
+ // memcpy in favor of the data that was already at the destination.
MemDepResult DepInfo = MD->getDependency(M);
- if (!DepInfo.isClobber())
- return false;
-
- if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
- return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
-
- if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
- if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
- CopySize->getZExtValue(), C)) {
+ if (DepInfo.isClobber()) {
+ if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
+ if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
+ CopySize->getZExtValue(), M->getAlignment(),
+ C)) {
+ MD->removeInstruction(M);
+ M->eraseFromParent();
+ return true;
+ }
+ }
+ }
+
+ AliasAnalysis::Location SrcLoc = AliasAnalysis::getLocationForSource(M);
+ MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(SrcLoc, true,
+ M, M->getParent());
+ if (SrcDepInfo.isClobber()) {
+ if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))
+ return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
+ } else if (SrcDepInfo.isDef()) {
+ Instruction *I = SrcDepInfo.getInst();
+ bool hasUndefContents = false;
+
+ if (isa<AllocaInst>(I)) {
+ hasUndefContents = true;
+ } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start)
+ if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))
+ if (LTSize->getZExtValue() >= CopySize->getZExtValue())
+ hasUndefContents = true;
+ }
+
+ if (hasUndefContents) {
MD->removeInstruction(M);
M->eraseFromParent();
+ ++NumMemCpyInstr;
return true;
}
}
-
+
return false;
}
if (!TLI->has(LibFunc::memmove))
return false;
-
+
// See if the pointers alias.
if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M)))
return false;
-
+
DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
-
+
// If not, then we know we can transform this.
Module *Mod = M->getParent()->getParent()->getParent();
Type *ArgTys[3] = { M->getRawDest()->getType(),
++NumMoveToCpy;
return true;
}
-
+
/// processByValArgument - This is called on every byval argument in call sites.
bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
- if (TD == 0) return false;
-
+ const DataLayout &DL = CS.getCaller()->getParent()->getDataLayout();
// Find out what feeds this byval argument.
Value *ByValArg = CS.getArgument(ArgNo);
- const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType();
- uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
+ Type *ByValTy = cast<PointerType>(ByValArg->getType())->getElementType();
+ uint64_t ByValSize = DL.getTypeAllocSize(ByValTy);
MemDepResult DepInfo =
MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
true, CS.getInstruction(),
// a memcpy, see if we can byval from the source of the memcpy instead of the
// result.
MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
- if (MDep == 0 || MDep->isVolatile() ||
+ if (!MDep || MDep->isVolatile() ||
ByValArg->stripPointerCasts() != MDep->getDest())
return false;
-
+
// The length of the memcpy must be larger or equal to the size of the byval.
ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
- if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
+ if (!C1 || C1->getValue().getZExtValue() < ByValSize)
return false;
// Get the alignment of the byval. If the call doesn't specify the alignment,
// then it is some target specific value that we can't know.
unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
if (ByValAlign == 0) return false;
-
+
// If it is greater than the memcpy, then we check to see if we can force the
// source of the memcpy to the alignment we need. If we fail, we bail out.
+ AssumptionCache &AC =
+ getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
+ *CS->getParent()->getParent());
+ DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
if (MDep->getAlignment() < ByValAlign &&
- getOrEnforceKnownAlignment(MDep->getSource(),ByValAlign, TD) < ByValAlign)
+ getOrEnforceKnownAlignment(MDep->getSource(), ByValAlign, DL,
+ CS.getInstruction(), &AC, &DT) < ByValAlign)
return false;
-
+
// Verify that the copied-from memory doesn't change in between the memcpy and
// the byval call.
// memcpy(a <- b)
false, CS.getInstruction(), MDep->getParent());
if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
return false;
-
+
Value *TmpCast = MDep->getSource();
if (MDep->getSource()->getType() != ByValArg->getType())
TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
"tmpcast", CS.getInstruction());
-
+
DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
<< " " << *MDep << "\n"
<< " " << *CS.getInstruction() << "\n");
-
+
// Otherwise we're good! Update the byval argument.
CS.setArgument(ArgNo, TmpCast);
++NumMemCpyInstr;
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
// Avoid invalidating the iterator.
Instruction *I = BI++;
-
+
bool RepeatInstruction = false;
-
+
if (StoreInst *SI = dyn_cast<StoreInst>(I))
MadeChange |= processStore(SI, BI);
else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
RepeatInstruction = processMemMove(M);
else if (CallSite CS = (Value*)I) {
for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
- if (CS.paramHasAttr(i+1, Attribute::ByVal))
+ if (CS.isByValArgument(i))
MadeChange |= processByValArgument(CS, i);
}
}
}
}
-
+
return MadeChange;
}
// function.
//
bool MemCpyOpt::runOnFunction(Function &F) {
+ if (skipOptnoneFunction(F))
+ return false;
+
bool MadeChange = false;
MD = &getAnalysis<MemoryDependenceAnalysis>();
- TD = getAnalysisIfAvailable<TargetData>();
- TLI = &getAnalysis<TargetLibraryInfo>();
-
+ TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
+
// If we don't have at least memset and memcpy, there is little point of doing
// anything here. These are required by a freestanding implementation, so if
// even they are disabled, there is no point in trying hard.
if (!TLI->has(LibFunc::memset) || !TLI->has(LibFunc::memcpy))
return false;
-
+
while (1) {
if (!iterateOnFunction(F))
break;
MadeChange = true;
}
-
- MD = 0;
+
+ MD = nullptr;
return MadeChange;
}