bool AMDGPURewriteOutArguments::isOutArgumentCandidate(Argument &Arg) const {
const unsigned MaxOutArgSizeBytes = 4 * MaxNumRetRegs;
PointerType *ArgTy = dyn_cast<PointerType>(Arg.getType());
// TODO: It might be useful for any out arguments, not just privates.
if (!ArgTy || (ArgTy->getAddressSpace() != DL->getAllocaAddrSpace() &&
!AnyAddressSpace) ||
Arg.hasByValAttr() || Arg.hasStructRetAttr() ||
DL->getTypeStoreSize(ArgTy->getPointerElementType()) > MaxOutArgSizeBytes) {
return false;
}
return checkArgumentUses(Arg);
}
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;
}
}
/// PromoteArguments - This method checks the specified function to see if there
/// are any promotable arguments and if it is safe to promote the function (for
/// example, all callers are direct). If safe to promote some arguments, it
/// calls the DoPromotion method.
///
CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
Function *F = CGN->getFunction();
// Make sure that it is local to this module.
if (!F || !F->hasLocalLinkage()) return nullptr;
// Don't promote arguments for variadic functions. Adding, removing, or
// changing non-pack parameters can change the classification of pack
// parameters. Frontends encode that classification at the call site in the
// IR, while in the callee the classification is determined dynamically based
// on the number of registers consumed so far.
if (F->isVarArg()) return nullptr;
// First check: see if there are any pointer arguments! If not, quick exit.
SmallVector<Argument*, 16> PointerArgs;
for (Argument &I : F->args())
if (I.getType()->isPointerTy())
PointerArgs.push_back(&I);
if (PointerArgs.empty()) return nullptr;
// Second check: make sure that all callers are direct callers. We can't
// transform functions that have indirect callers. Also see if the function
// is self-recursive.
bool isSelfRecursive = false;
for (Use &U : F->uses()) {
CallSite CS(U.getUser());
// Must be a direct call.
if (CS.getInstruction() == nullptr || !CS.isCallee(&U)) return nullptr;
if (CS.getInstruction()->getParent()->getParent() == F)
isSelfRecursive = true;
}
const DataLayout &DL = F->getParent()->getDataLayout();
// We need to manually construct BasicAA directly in order to disable its use
// of other function analyses.
BasicAAResult BAR(createLegacyPMBasicAAResult(*this, *F));
// Construct our own AA results for this function. We do this manually to
// work around the limitations of the legacy pass manager.
AAResults AAR(createLegacyPMAAResults(*this, *F, BAR));
// Check to see which arguments are promotable. If an argument is promotable,
// add it to ArgsToPromote.
SmallPtrSet<Argument*, 8> ArgsToPromote;
SmallPtrSet<Argument*, 8> ByValArgsToTransform;
for (unsigned i = 0, e = PointerArgs.size(); i != e; ++i) {
Argument *PtrArg = PointerArgs[i];
Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
// Replace sret attribute with noalias. This reduces register pressure by
// avoiding a register copy.
if (PtrArg->hasStructRetAttr()) {
unsigned ArgNo = PtrArg->getArgNo();
F->setAttributes(
F->getAttributes()
.removeAttribute(F->getContext(), ArgNo + 1, Attribute::StructRet)
.addAttribute(F->getContext(), ArgNo + 1, Attribute::NoAlias));
for (Use &U : F->uses()) {
CallSite CS(U.getUser());
CS.setAttributes(
CS.getAttributes()
.removeAttribute(F->getContext(), ArgNo + 1,
Attribute::StructRet)
.addAttribute(F->getContext(), ArgNo + 1, Attribute::NoAlias));
}
}
// If this is a byval argument, and if the aggregate type is small, just
// pass the elements, which is always safe, if the passed value is densely
// packed or if we can prove the padding bytes are never accessed. This does
// not apply to inalloca.
bool isSafeToPromote =
PtrArg->hasByValAttr() &&
(isDenselyPacked(AgTy, DL) || !canPaddingBeAccessed(PtrArg));
if (isSafeToPromote) {
if (StructType *STy = dyn_cast<StructType>(AgTy)) {
if (maxElements > 0 && STy->getNumElements() > maxElements) {
DEBUG(dbgs() << "argpromotion disable promoting argument '"
<< PtrArg->getName() << "' because it would require adding more"
<< " than " << maxElements << " arguments to the function.\n");
continue;
}
// If all the elements are single-value types, we can promote it.
bool AllSimple = true;
for (const auto *EltTy : STy->elements()) {
if (!EltTy->isSingleValueType()) {
AllSimple = false;
break;
}
}
// Safe to transform, don't even bother trying to "promote" it.
// Passing the elements as a scalar will allow sroa to hack on
// the new alloca we introduce.
if (AllSimple) {
ByValArgsToTransform.insert(PtrArg);
continue;
}
}
}
// If the argument is a recursive type and we're in a recursive
// function, we could end up infinitely peeling the function argument.
if (isSelfRecursive) {
if (StructType *STy = dyn_cast<StructType>(AgTy)) {
bool RecursiveType = false;
for (const auto *EltTy : STy->elements()) {
if (EltTy == PtrArg->getType()) {
RecursiveType = true;
break;
}
}
if (RecursiveType)
continue;
}
}
// Otherwise, see if we can promote the pointer to its value.
if (isSafeToPromoteArgument(PtrArg, PtrArg->hasByValOrInAllocaAttr(), AAR))
ArgsToPromote.insert(PtrArg);
}
// No promotable pointer arguments.
if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
return nullptr;
return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
}
