/// 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 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;
// 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);
BasicBlock::iterator BI = StartInst;
for (++BI; !isa<TerminatorInst>(BI); ++BI) {
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 (BI->mayWriteToMemory() || BI->mayReadFromMemory())
break;
continue;
}
if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
// If this is a store, see if we can merge it in.
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))
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
// could be merged in. This is a very common case of course.
if (Ranges.empty())
return 0;
// 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.
Ranges.addInst(0, StartInst);
// If we create any memsets, we put it right before the first instruction that
// isn't part of the memset block. This ensure that the memset is dominated
// by any addressing instruction needed by the start of the block.
IRBuilder<> Builder(BI);
// 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;
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))
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) {
Type *EltType =
cast<PointerType>(StartPtr->getType())->getElementType();
Alignment = TD->getABITypeAlignment(EltType);
}
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';
dbgs() << "With: " << *AMemSet << '\n');
if (!Range.TheStores.empty())
AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
// Zap all the stores.
for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
MD->removeInstruction(*SI);
(*SI)->eraseFromParent();
}
++NumMemSetInfer;
}
return AMemSet;
}
bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
SmallVector<LoadInst *, 4> aggrLoads;
SmallVector<MemTransferInst *, 4> aggrMemcpys;
SmallVector<MemSetInst *, 4> aggrMemsets;
DataLayout *TD = &getAnalysis<DataLayout>();
LLVMContext &Context = F.getParent()->getContext();
//
// Collect all the aggrLoads, aggrMemcpys and addrMemsets.
//
//const BasicBlock *firstBB = &F.front(); // first BB in F
for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
//BasicBlock *bb = BI;
for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
++II) {
if (LoadInst * load = dyn_cast<LoadInst>(II)) {
if (load->hasOneUse() == false) continue;
if (TD->getTypeStoreSize(load->getType()) < MaxAggrCopySize) continue;
User *use = *(load->use_begin());
if (StoreInst * store = dyn_cast<StoreInst>(use)) {
if (store->getOperand(0) != load) //getValueOperand
continue;
aggrLoads.push_back(load);
}
} else if (MemTransferInst * intr = dyn_cast<MemTransferInst>(II)) {
Value *len = intr->getLength();
// If the number of elements being copied is greater
// than MaxAggrCopySize, lower it to a loop
if (ConstantInt * len_int = dyn_cast < ConstantInt > (len)) {
if (len_int->getZExtValue() >= MaxAggrCopySize) {
aggrMemcpys.push_back(intr);
}
} else {
// turn variable length memcpy/memmov into loop
aggrMemcpys.push_back(intr);
}
} else if (MemSetInst * memsetintr = dyn_cast<MemSetInst>(II)) {
Value *len = memsetintr->getLength();
if (ConstantInt * len_int = dyn_cast<ConstantInt>(len)) {
if (len_int->getZExtValue() >= MaxAggrCopySize) {
aggrMemsets.push_back(memsetintr);
}
} else {
// turn variable length memset into loop
aggrMemsets.push_back(memsetintr);
}
}
}
}
if ((aggrLoads.size() == 0) && (aggrMemcpys.size() == 0)
&& (aggrMemsets.size() == 0)) return false;
//
// Do the transformation of an aggr load/copy/set to a loop
//
for (unsigned i = 0, e = aggrLoads.size(); i != e; ++i) {
LoadInst *load = aggrLoads[i];
StoreInst *store = dyn_cast<StoreInst>(*load->use_begin());
Value *srcAddr = load->getOperand(0);
Value *dstAddr = store->getOperand(1);
unsigned numLoads = TD->getTypeStoreSize(load->getType());
Value *len = ConstantInt::get(Type::getInt32Ty(Context), numLoads);
convertTransferToLoop(store, srcAddr, dstAddr, len, load->isVolatile(),
store->isVolatile(), Context, F);
store->eraseFromParent();
load->eraseFromParent();
}
for (unsigned i = 0, e = aggrMemcpys.size(); i != e; ++i) {
MemTransferInst *cpy = aggrMemcpys[i];
Value *len = cpy->getLength();
// llvm 2.7 version of memcpy does not have volatile
// operand yet. So always making it non-volatile
// optimistically, so that we don't see unnecessary
// st.volatile in ptx
convertTransferToLoop(cpy, cpy->getSource(), cpy->getDest(), len, false,
false, Context, F);
cpy->eraseFromParent();
}
for (unsigned i = 0, e = aggrMemsets.size(); i != e; ++i) {
MemSetInst *memsetinst = aggrMemsets[i];
Value *len = memsetinst->getLength();
Value *val = memsetinst->getValue();
convertMemSetToLoop(memsetinst, memsetinst->getDest(), len, val, Context,
F);
memsetinst->eraseFromParent();
}
return true;
}
void Lint::visitCallSite(CallSite CS) {
Instruction &I = *CS.getInstruction();
Value *Callee = CS.getCalledValue();
visitMemoryReference(I, Callee, MemoryLocation::UnknownSize, 0, nullptr,
MemRef::Callee);
if (Function *F = dyn_cast<Function>(findValue(Callee,
/*OffsetOk=*/false))) {
Assert(CS.getCallingConv() == F->getCallingConv(),
"Undefined behavior: Caller and callee calling convention differ",
&I);
FunctionType *FT = F->getFunctionType();
unsigned NumActualArgs = CS.arg_size();
Assert(FT->isVarArg() ? FT->getNumParams() <= NumActualArgs
: FT->getNumParams() == NumActualArgs,
"Undefined behavior: Call argument count mismatches callee "
"argument count",
&I);
Assert(FT->getReturnType() == I.getType(),
"Undefined behavior: Call return type mismatches "
"callee return type",
&I);
// Check argument types (in case the callee was casted) and attributes.
// TODO: Verify that caller and callee attributes are compatible.
Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
for (; AI != AE; ++AI) {
Value *Actual = *AI;
if (PI != PE) {
Argument *Formal = &*PI++;
Assert(Formal->getType() == Actual->getType(),
"Undefined behavior: Call argument type mismatches "
"callee parameter type",
&I);
// Check that noalias arguments don't alias other arguments. This is
// not fully precise because we don't know the sizes of the dereferenced
// memory regions.
if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy())
for (CallSite::arg_iterator BI = CS.arg_begin(); BI != AE; ++BI)
if (AI != BI && (*BI)->getType()->isPointerTy()) {
AliasResult Result = AA->alias(*AI, *BI);
Assert(Result != MustAlias && Result != PartialAlias,
"Unusual: noalias argument aliases another argument", &I);
}
// Check that an sret argument points to valid memory.
if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
Type *Ty =
cast<PointerType>(Formal->getType())->getElementType();
visitMemoryReference(I, Actual, DL->getTypeStoreSize(Ty),
DL->getABITypeAlignment(Ty), Ty,
MemRef::Read | MemRef::Write);
}
}
}
}
if (CS.isCall() && cast<CallInst>(CS.getInstruction())->isTailCall())
for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
AI != AE; ++AI) {
Value *Obj = findValue(*AI, /*OffsetOk=*/true);
Assert(!isa<AllocaInst>(Obj),
"Undefined behavior: Call with \"tail\" keyword references "
"alloca",
&I);
}
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
switch (II->getIntrinsicID()) {
default: break;
// TODO: Check more intrinsics
case Intrinsic::memcpy: {
MemCpyInst *MCI = cast<MemCpyInst>(&I);
// TODO: If the size is known, use it.
visitMemoryReference(I, MCI->getDest(), MemoryLocation::UnknownSize,
MCI->getAlignment(), nullptr, MemRef::Write);
visitMemoryReference(I, MCI->getSource(), MemoryLocation::UnknownSize,
MCI->getAlignment(), nullptr, MemRef::Read);
// Check that the memcpy arguments don't overlap. The AliasAnalysis API
// isn't expressive enough for what we really want to do. Known partial
// overlap is not distinguished from the case where nothing is known.
uint64_t Size = 0;
if (const ConstantInt *Len =
dyn_cast<ConstantInt>(findValue(MCI->getLength(),
/*OffsetOk=*/false)))
if (Len->getValue().isIntN(32))
Size = Len->getValue().getZExtValue();
Assert(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
MustAlias,
"Undefined behavior: memcpy source and destination overlap", &I);
break;
}
case Intrinsic::memmove: {
MemMoveInst *MMI = cast<MemMoveInst>(&I);
// TODO: If the size is known, use it.
visitMemoryReference(I, MMI->getDest(), MemoryLocation::UnknownSize,
MMI->getAlignment(), nullptr, MemRef::Write);
visitMemoryReference(I, MMI->getSource(), MemoryLocation::UnknownSize,
MMI->getAlignment(), nullptr, MemRef::Read);
break;
}
case Intrinsic::memset: {
MemSetInst *MSI = cast<MemSetInst>(&I);
// TODO: If the size is known, use it.
visitMemoryReference(I, MSI->getDest(), MemoryLocation::UnknownSize,
MSI->getAlignment(), nullptr, MemRef::Write);
break;
}
case Intrinsic::vastart:
Assert(I.getParent()->getParent()->isVarArg(),
"Undefined behavior: va_start called in a non-varargs function",
&I);
visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Read | MemRef::Write);
break;
case Intrinsic::vacopy:
visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Write);
visitMemoryReference(I, CS.getArgument(1), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Read);
break;
case Intrinsic::vaend:
visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Read | MemRef::Write);
break;
case Intrinsic::stackrestore:
// Stackrestore doesn't read or write memory, but it sets the
// stack pointer, which the compiler may read from or write to
// at any time, so check it for both readability and writeability.
visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
nullptr, MemRef::Read | MemRef::Write);
break;
}
}
void Lint::visitCallSite(CallSite CS) {
Instruction &I = *CS.getInstruction();
Value *Callee = CS.getCalledValue();
// TODO: Check function alignment?
visitMemoryReference(I, Callee, 0, 0);
if (Function *F = dyn_cast<Function>(Callee->stripPointerCasts())) {
Assert1(CS.getCallingConv() == F->getCallingConv(),
"Undefined behavior: Caller and callee calling convention differ",
&I);
const FunctionType *FT = F->getFunctionType();
unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
Assert1(FT->isVarArg() ?
FT->getNumParams() <= NumActualArgs :
FT->getNumParams() == NumActualArgs,
"Undefined behavior: Call argument count mismatches callee "
"argument count", &I);
// TODO: Check argument types (in case the callee was casted)
// TODO: Check ABI-significant attributes.
// TODO: Check noalias attribute.
// TODO: Check sret attribute.
}
// TODO: Check the "tail" keyword constraints.
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
switch (II->getIntrinsicID()) {
default: break;
// TODO: Check more intrinsics
case Intrinsic::memcpy: {
MemCpyInst *MCI = cast<MemCpyInst>(&I);
visitMemoryReference(I, MCI->getSource(), MCI->getAlignment(), 0);
visitMemoryReference(I, MCI->getDest(), MCI->getAlignment(), 0);
// Check that the memcpy arguments don't overlap. The AliasAnalysis API
// isn't expressive enough for what we really want to do. Known partial
// overlap is not distinguished from the case where nothing is known.
unsigned Size = 0;
if (const ConstantInt *Len =
dyn_cast<ConstantInt>(MCI->getLength()->stripPointerCasts()))
if (Len->getValue().isIntN(32))
Size = Len->getValue().getZExtValue();
Assert1(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
AliasAnalysis::MustAlias,
"Undefined behavior: memcpy source and destination overlap", &I);
break;
}
case Intrinsic::memmove: {
MemMoveInst *MMI = cast<MemMoveInst>(&I);
visitMemoryReference(I, MMI->getSource(), MMI->getAlignment(), 0);
visitMemoryReference(I, MMI->getDest(), MMI->getAlignment(), 0);
break;
}
case Intrinsic::memset: {
MemSetInst *MSI = cast<MemSetInst>(&I);
visitMemoryReference(I, MSI->getDest(), MSI->getAlignment(), 0);
break;
}
case Intrinsic::vastart:
Assert1(I.getParent()->getParent()->isVarArg(),
"Undefined behavior: va_start called in a non-varargs function",
&I);
visitMemoryReference(I, CS.getArgument(0), 0, 0);
break;
case Intrinsic::vacopy:
visitMemoryReference(I, CS.getArgument(0), 0, 0);
visitMemoryReference(I, CS.getArgument(1), 0, 0);
break;
case Intrinsic::vaend:
visitMemoryReference(I, CS.getArgument(0), 0, 0);
break;
case Intrinsic::stackrestore:
visitMemoryReference(I, CS.getArgument(0), 0, 0);
break;
}
}
