#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Transforms/Utils/Local.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 <list>
using namespace llvm;
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)) {
+ if (StructType *STy = dyn_cast<StructType>(*GTI)) {
Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
continue;
}
/// 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) {
+ Ptr1 = Ptr1->stripPointerCasts();
+ Ptr2 = Ptr2->stripPointerCasts();
+ GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
+ GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(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);
+ return !VariableIdxFound;
+ }
+
+ if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
+ Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD);
+ 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
// offset, which determines their offset from each other. At this point, we
// handle no other case.
- GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
- GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
return false;
if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
break;
- bool VariableIdxFound = false;
int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
if (VariableIdxFound) return false;
unsigned Alignment;
/// TheStores - The actual stores that make up this range.
- SmallVector<StoreInst*, 16> TheStores;
+ SmallVector<Instruction*, 16> TheStores;
bool isProfitableToUseMemset(const TargetData &TD) 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;
+ // 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 (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.
const_iterator end() const { return Ranges.end(); }
bool empty() const { return Ranges.empty(); }
- void addStore(int64_t OffsetFromFirst, StoreInst *SI);
-
void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
- addStore(OffsetFromFirst, cast<StoreInst>(Inst));
+ if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
+ addStore(OffsetFromFirst, SI);
+ else
+ addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
+ }
+
+ void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
+ int64_t StoreSize = TD.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
-/// addStore - Add a new store to the MemsetRanges data structure. This adds a
+/// addRange - Add a new store to the MemsetRanges data structure. This adds a
/// new range for the specified store at the specified offset, merging into
/// existing ranges as appropriate.
-void MemsetRanges::addStore(int64_t Start, StoreInst *SI) {
- int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType());
-
- // Do a linear search of the ranges to see if this can be joined and/or to
- // find the insertion point in the list. We keep the ranges sorted for
- // simplicity here. This is a linear search of a linked list, which is ugly,
- // however the number of ranges is limited, so this won't get crazy slow.
+///
+/// Do a linear search of the ranges to see if this can be joined and/or to
+/// find the insertion point in the list. We keep the ranges sorted for
+/// simplicity here. This is a linear search of a linked list, which is ugly,
+/// however the number of ranges is limited, so this won't get crazy slow.
+void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
+ unsigned Alignment, Instruction *Inst) {
+ int64_t End = Start+Size;
range_iterator I = Ranges.begin(), E = Ranges.end();
while (I != E && Start > I->End)
MemsetRange &R = *Ranges.insert(I, MemsetRange());
R.Start = Start;
R.End = End;
- R.StartPtr = SI->getPointerOperand();
- R.Alignment = SI->getAlignment();
- R.TheStores.push_back(SI);
+ R.StartPtr = Ptr;
+ R.Alignment = Alignment;
+ R.TheStores.push_back(Inst);
return;
}
-
+
// This store overlaps with I, add it.
- I->TheStores.push_back(SI);
+ 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.
// stopped on *it*.
if (Start < I->Start) {
I->Start = Start;
- I->StartPtr = SI->getPointerOperand();
- I->Alignment = SI->getAlignment();
+ I->StartPtr = Ptr;
+ I->Alignment = Alignment;
}
// Now we know that Start <= I->End and Start >= I->Start (so the startpoint
namespace {
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;
}
bool runOnFunction(Function &F);
AU.addRequired<DominatorTree>();
AU.addRequired<MemoryDependenceAnalysis>();
AU.addRequired<AliasAnalysis>();
+ AU.addRequired<TargetLibraryInfo>();
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,
false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTree)
INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
false, false)
/// tryMergingIntoMemset - When scanning forward over instructions, we look for
/// some other patterns to fold away. In particular, this looks for stores to
-/// neighboring locations of memory. If it sees enough consequtive ones, it
+/// 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,
Value *StartPtr, Value *ByteVal) {
if (TD == 0) return 0;
- AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
-
// 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
BasicBlock::iterator BI = StartInst;
for (++BI; !isa<TerminatorInst>(BI); ++BI) {
- if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
- // If the call is readnone, ignore it, otherwise bail out. We don't even
- // allow readonly here because we don't want something like:
+ if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
+ // If the instruction is readnone, ignore it, otherwise bail out. We
+ // don't even allow readonly here because we don't want something like:
// A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
- if (AA.getModRefBehavior(CallSite(BI)) ==
- AliasAnalysis::DoesNotAccessMemory)
- continue;
-
- // TODO: If this is a memset, try to join it in.
-
- break;
- } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI))
- break;
-
- // If this is a non-store instruction it is fine, ignore it.
- StoreInst *NextStore = dyn_cast<StoreInst>(BI);
- if (NextStore == 0) continue;
-
- // If this is a store, see if we can merge it in.
- if (NextStore->isVolatile()) break;
-
- // Check to see if this stored value is of the same byte-splattable value.
- if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
- break;
+ if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
+ break;
+ continue;
+ }
- // 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))
- break;
+ if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
+ // If this is a store, see if we can merge it in.
+ if (!NextStore->isSimple()) break;
- Ranges.addStore(Offset, NextStore);
+ // 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))
+ 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))
+ break;
+
+ Ranges.addMemSet(Offset, MSI);
+ }
}
// If we have no ranges, then we just had a single store with nothing that
// Determine alignment
unsigned Alignment = Range.Alignment;
if (Alignment == 0) {
- const Type *EltType =
+ Type *EltType =
cast<PointerType>(StartPtr->getType())->getElementType();
Alignment = TD->getABITypeAlignment(EltType);
}
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
dbgs() << *Range.TheStores[i] << '\n';
dbgs() << "With: " << *AMemSet << '\n');
-
+
+ if (!Range.TheStores.empty())
+ AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
+
// Zap all the stores.
- for (SmallVector<StoreInst*, 16>::const_iterator
+ for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
- SE = Range.TheStores.end(); SI != SE; ++SI)
+ SE = Range.TheStores.end(); SI != SE; ++SI) {
+ MD->removeInstruction(*SI);
(*SI)->eraseFromParent();
+ }
++NumMemSetInfer;
}
bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
- if (SI->isVolatile()) return false;
+ if (!SI->isSimple()) return false;
if (TD == 0) return false;
// 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()) {
- MemDepResult dep = MD->getDependency(LI);
+ if (LI->isSimple() && LI->hasOneUse() &&
+ LI->getParent() == SI->getParent()) {
+ MemDepResult ldep = MD->getDependency(LI);
CallInst *C = 0;
- if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
- C = dyn_cast<CallInst>(dep.getInst());
-
+ if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
+ C = dyn_cast<CallInst>(ldep.getInst());
+
+ if (C) {
+ // Check that nothing touches the dest of the "copy" between
+ // the call and the store.
+ AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+ AliasAnalysis::Location StoreLoc = AA.getLocation(SI);
+ for (BasicBlock::iterator I = --BasicBlock::iterator(SI),
+ E = C; I != E; --I) {
+ if (AA.getModRefInfo(&*I, StoreLoc) != AliasAnalysis::NoModRef) {
+ C = 0;
+ break;
+ }
+ }
+ }
+
if (C) {
bool changed = performCallSlotOptzn(LI,
SI->getPointerOperand()->stripPointerCasts(),
if (changed) {
MD->removeInstruction(SI);
SI->eraseFromParent();
+ MD->removeInstruction(LI);
LI->eraseFromParent();
++NumMemCpyInstr;
return true;
return false;
}
+bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
+ // See if there is another memset or store neighboring this memset which
+ // allows us to widen out the memset to do a single larger store.
+ if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
+ if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
+ MSI->getValue())) {
+ BBI = I; // Don't invalidate iterator.
+ return true;
+ }
+ return false;
+}
+
/// performCallSlotOptzn - takes a memcpy and a call that it depends on,
/// and checks for the possibility of a call slot optimization by having
if (!A->hasStructRetAttr())
return false;
- const Type *StructTy = cast<PointerType>(A->getType())->getElementType();
+ Type *StructTy = cast<PointerType>(A->getType())->getElementType();
uint64_t destSize = TD->getTypeAllocSize(StructTy);
if (destSize < srcSize)
// 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)
+ if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef)
return false;
// All the checks have passed, so do the transformation.
if (M->getSource() == MDep->getSource())
return false;
- // Second, the length of the memcpy's must be the same, or the preceeding one
+ // Second, the length of the memcpy's must be the same, or the preceding one
// must be larger than the following one.
- ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
- if (!C1) return false;
+ 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>();
// a) memcpy-memcpy xform which exposes redundance for DSE.
// b) call-memcpy xform for return slot optimization.
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)) {
- M->eraseFromParent();
- return true;
+ if (DepInfo.isClobber()) {
+ if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
+ if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
+ CopySize->getZExtValue(), 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());
+ }
+
return false;
}
bool MemCpyOpt::processMemMove(MemMoveInst *M) {
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+ if (!TLI->has(LibFunc::memmove))
+ return false;
+
// See if the pointers alias.
if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M)))
return false;
// If not, then we know we can transform this.
Module *Mod = M->getParent()->getParent()->getParent();
- const Type *ArgTys[3] = { M->getRawDest()->getType(),
- M->getRawSource()->getType(),
- M->getLength()->getType() };
+ Type *ArgTys[3] = { M->getRawDest()->getType(),
+ M->getRawSource()->getType(),
+ M->getLength()->getType() };
M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy,
- ArgTys, 3));
+ ArgTys));
// MemDep may have over conservative information about this instruction, just
// conservatively flush it from the cache.
// Find out what feeds this byval argument.
Value *ByValArg = CS.getArgument(ArgNo);
- const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType();
+ Type *ByValTy = cast<PointerType>(ByValArg->getType())->getElementType();
uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
MemDepResult DepInfo =
MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
return false;
- // Get the alignment of the byval. If it is greater than the memcpy, then we
- // can't do the substitution. If the call doesn't specify the alignment, then
- // it is some target specific value that we can't know.
+ // 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 || MDep->getAlignment() < ByValAlign)
- return false;
+ 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.
+ if (MDep->getAlignment() < ByValAlign &&
+ getOrEnforceKnownAlignment(MDep->getSource(),ByValAlign, TD) < ByValAlign)
+ return false;
// Verify that the copied-from memory doesn't change in between the memcpy and
// the byval call.
if (StoreInst *SI = dyn_cast<StoreInst>(I))
MadeChange |= processStore(SI, BI);
- else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) {
+ else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
+ RepeatInstruction = processMemSet(M, BI);
+ else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
RepeatInstruction = processMemCpy(M);
- } else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) {
+ else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I))
RepeatInstruction = processMemMove(M);
- } else if (CallSite CS = (Value*)I) {
+ 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);
}
// Reprocess the instruction if desired.
if (RepeatInstruction) {
- --BI;
+ if (BI != BB->begin()) --BI;
MadeChange = true;
}
}
bool MadeChange = false;
MD = &getAnalysis<MemoryDependenceAnalysis>();
TD = getAnalysisIfAvailable<TargetData>();
+ TLI = &getAnalysis<TargetLibraryInfo>();
+
+ // 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;