bool CallAnalyzer::visitAlloca(AllocaInst &I) {
// Check whether inlining will turn a dynamic alloca into a static
// alloca, and handle that case.
if (I.isArrayAllocation()) {
if (Constant *Size = SimplifiedValues.lookup(I.getArraySize())) {
ConstantInt *AllocSize = dyn_cast<ConstantInt>(Size);
assert(AllocSize && "Allocation size not a constant int?");
Type *Ty = I.getAllocatedType();
AllocatedSize += Ty->getPrimitiveSizeInBits() * AllocSize->getZExtValue();
return Base::visitAlloca(I);
}
}
// Accumulate the allocated size.
if (I.isStaticAlloca()) {
Type *Ty = I.getAllocatedType();
AllocatedSize += (DL ? DL->getTypeAllocSize(Ty) :
Ty->getPrimitiveSizeInBits());
}
// We will happily inline static alloca instructions.
if (I.isStaticAlloca())
return Base::visitAlloca(I);
// FIXME: This is overly conservative. Dynamic allocas are inefficient for
// a variety of reasons, and so we would like to not inline them into
// functions which don't currently have a dynamic alloca. This simply
// disables inlining altogether in the presence of a dynamic alloca.
HasDynamicAlloca = true;
return false;
}
SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) {
if (!I.getAllocatedType()->isSized())
return unknown();
// must be a VLA
assert(I.isArrayAllocation());
Value *ArraySize = I.getArraySize();
Value *Size = ConstantInt::get(ArraySize->getType(),
TD->getTypeAllocSize(I.getAllocatedType()));
Size = Builder.CreateMul(Size, ArraySize);
return std::make_pair(Size, Zero);
}
SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) {
if (!I.getAllocatedType()->isSized())
return unknown();
APInt Size(IntTyBits, TD->getTypeAllocSize(I.getAllocatedType()));
if (!I.isArrayAllocation())
return std::make_pair(align(Size, I.getAlignment()), Zero);
Value *ArraySize = I.getArraySize();
if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) {
Size *= C->getValue().zextOrSelf(IntTyBits);
return std::make_pair(align(Size, I.getAlignment()), Zero);
}
return unknown();
}
// -- handle unary instruction --
void UnsafeTypeCastingCheck::handleAllocInstruction (Instruction *inst) {
AllocaInst *ainst = dyn_cast<AllocaInst>(inst);
if (ainst == NULL)
utccAbort("handleAllocInstruction cannot process with a non-alloca instruction");
Type *atype = ainst->getAllocatedType();
setPointedType(ainst, llvmT2utccT(atype, ainst));
}
void ArrayIndexChecker::visitAllocaInst(AllocaInst& I) {
DEBUG(dbgs() << "ArrayIndexChecker: visiting alloca " << I << "\n");
if (isa<StructType>(I.getAllocatedType()) || isa<ArrayType>(I.getAllocatedType())) {
return;
}
auto pos = std::find(ptr_value_vec_.begin(), ptr_value_vec_.end(), &I);
assert(pos != ptr_value_vec_.end());
index_t varIdx = pos - ptr_value_vec_.begin();
assert(idx2addr_.find(varIdx) != idx2addr_.end());
if (addr2version_[idx2addr_[varIdx]] != 0)
throw ArrayIndexIsNotConstant;
DEBUG(dbgs() << "ArrayIndexChecker: visited alloca\n");
}
bool IRTranslator::translateStaticAlloca(const AllocaInst &AI) {
assert(AI.isStaticAlloca() && "only handle static allocas now");
MachineFunction &MF = MIRBuilder.getMF();
unsigned ElementSize = DL->getTypeStoreSize(AI.getAllocatedType());
unsigned Size =
ElementSize * cast<ConstantInt>(AI.getArraySize())->getZExtValue();
// Always allocate at least one byte.
Size = std::max(Size, 1u);
unsigned Alignment = AI.getAlignment();
if (!Alignment)
Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
unsigned Res = getOrCreateVReg(AI);
int FI = MF.getFrameInfo().CreateStackObject(Size, Alignment, false, &AI);
MIRBuilder.buildFrameIndex(LLT::pointer(0), Res, FI);
return true;
}
int IRTranslator::getOrCreateFrameIndex(const AllocaInst &AI) {
if (FrameIndices.find(&AI) != FrameIndices.end())
return FrameIndices[&AI];
MachineFunction &MF = MIRBuilder.getMF();
unsigned ElementSize = DL->getTypeStoreSize(AI.getAllocatedType());
unsigned Size =
ElementSize * cast<ConstantInt>(AI.getArraySize())->getZExtValue();
// Always allocate at least one byte.
Size = std::max(Size, 1u);
unsigned Alignment = AI.getAlignment();
if (!Alignment)
Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
int &FI = FrameIndices[&AI];
FI = MF.getFrameInfo().CreateStackObject(Size, Alignment, false, &AI);
return FI;
}
SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) {
if (!I.getAllocatedType()->isSized())
return unknown();
APInt Size(IntTyBits, DL.getTypeAllocSize(I.getAllocatedType()));
if (!I.isArrayAllocation())
return std::make_pair(align(Size, I.getAlignment()), Zero);
Value *ArraySize = I.getArraySize();
if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) {
APInt NumElems = C->getValue();
if (!CheckedZextOrTrunc(NumElems))
return unknown();
bool Overflow;
Size = Size.umul_ov(NumElems, Overflow);
return Overflow ? unknown() : std::make_pair(align(Size, I.getAlignment()),
Zero);
}
return unknown();
}
void LowerIntrinsics::InsertAutomaticGCRoots(Function &F, GCStrategy &S) {
if (!S.usesAutomaticRoots())
return;
LLVMContext &C = F.getParent()->getContext();
for (Function::iterator BBI = F.begin(), BBE = F.end(); BBI != BBE; ++BBI) {
for (BasicBlock::iterator II = BBI->begin(),
IE = BBI->end(); II != IE; ++II) {
AllocaInst *AI = dyn_cast<AllocaInst>(&*II);
if (!AI)
continue;
Value *Index = ConstantInt::get(Type::getInt32Ty(C), 0);
ArrayRef<Value *> Indices(&Index, 1);
AutomaticallyRootValue(*AI, AI->getAllocatedType(), Indices);
}
}
}
//
// Method: visitAllocaInst()
//
// Description:
// This method instruments an alloca instruction so that it is zero'ed out
// before any data is loaded from it.
//
void
InitAllocas::visitAllocaInst (AllocaInst & AI) {
//
// Scan for a place to insert the instruction to initialize the
// allocated memory.
//
Instruction * InsertPt = getInsertionPoint (AI);
//
// Zero the alloca with a memset. If this is done more efficiently with stores
// SelectionDAG will lower it appropriately based on target information.
//
TargetData & TD = getAnalysis<TargetData>();
//
// Get various types that we'll need.
//
Type * Int1Type = IntegerType::getInt1Ty(AI.getContext());
Type * Int8Type = IntegerType::getInt8Ty(AI.getContext());
Type * Int32Type = IntegerType::getInt32Ty(AI.getContext());
Type * VoidPtrType = getVoidPtrType (AI.getContext());
Type * AllocType = AI.getAllocatedType();
//
// Create a call to memset.
//
Module * M = AI.getParent()->getParent()->getParent();
Function * Memset = cast<Function>(M->getFunction ("llvm.memset.p0i8.i32"));
std::vector<Value *> args;
args.push_back (castTo (&AI, VoidPtrType, AI.getName().str(), InsertPt));
args.push_back (ConstantInt::get(Int8Type, 0));
args.push_back (ConstantInt::get(Int32Type,TD.getTypeAllocSize(AllocType)));
args.push_back (ConstantInt::get(Int32Type,
TD.getABITypeAlignment(AllocType)));
args.push_back (ConstantInt::get(Int1Type, 0));
CallInst::Create (Memset, args, "", InsertPt);
//
// Update statistics.
//
++InitedAllocas;
return;
}
bool CallAnalyzer::visitAlloca(AllocaInst &I) {
// FIXME: Check whether inlining will turn a dynamic alloca into a static
// alloca, and handle that case.
// Accumulate the allocated size.
if (I.isStaticAlloca()) {
Type *Ty = I.getAllocatedType();
AllocatedSize += (TD ? TD->getTypeAllocSize(Ty) :
Ty->getPrimitiveSizeInBits());
}
// We will happily inline static alloca instructions.
if (I.isStaticAlloca())
return Base::visitAlloca(I);
// FIXME: This is overly conservative. Dynamic allocas are inefficient for
// a variety of reasons, and so we would like to not inline them into
// functions which don't currently have a dynamic alloca. This simply
// disables inlining altogether in the presence of a dynamic alloca.
HasDynamicAlloca = true;
return false;
}
/// InlineCallIfPossible - If it is possible to inline the specified call site,
/// do so and update the CallGraph for this operation.
///
/// This function also does some basic book-keeping to update the IR. The
/// InlinedArrayAllocas map keeps track of any allocas that are already
/// available from other functions inlined into the caller. If we are able to
/// inline this call site we attempt to reuse already available allocas or add
/// any new allocas to the set if not possible.
static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
InlinedArrayAllocasTy &InlinedArrayAllocas) {
Function *Callee = CS.getCalledFunction();
Function *Caller = CS.getCaller();
// Try to inline the function. Get the list of static allocas that were
// inlined.
if (!InlineFunction(CS, IFI))
return false;
// If the inlined function had a higher stack protection level than the
// calling function, then bump up the caller's stack protection level.
if (Callee->hasFnAttr(Attribute::StackProtectReq))
Caller->addFnAttr(Attribute::StackProtectReq);
else if (Callee->hasFnAttr(Attribute::StackProtect) &&
!Caller->hasFnAttr(Attribute::StackProtectReq))
Caller->addFnAttr(Attribute::StackProtect);
// Look at all of the allocas that we inlined through this call site. If we
// have already inlined other allocas through other calls into this function,
// then we know that they have disjoint lifetimes and that we can merge them.
//
// There are many heuristics possible for merging these allocas, and the
// different options have different tradeoffs. One thing that we *really*
// don't want to hurt is SRoA: once inlining happens, often allocas are no
// longer address taken and so they can be promoted.
//
// Our "solution" for that is to only merge allocas whose outermost type is an
// array type. These are usually not promoted because someone is using a
// variable index into them. These are also often the most important ones to
// merge.
//
// A better solution would be to have real memory lifetime markers in the IR
// and not have the inliner do any merging of allocas at all. This would
// allow the backend to do proper stack slot coloring of all allocas that
// *actually make it to the backend*, which is really what we want.
//
// Because we don't have this information, we do this simple and useful hack.
//
SmallPtrSet<AllocaInst*, 16> UsedAllocas;
// Loop over all the allocas we have so far and see if they can be merged with
// a previously inlined alloca. If not, remember that we had it.
for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
AllocaNo != e; ++AllocaNo) {
AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
// Don't bother trying to merge array allocations (they will usually be
// canonicalized to be an allocation *of* an array), or allocations whose
// type is not itself an array (because we're afraid of pessimizing SRoA).
const ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
if (ATy == 0 || AI->isArrayAllocation())
continue;
// Get the list of all available allocas for this array type.
std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
// Loop over the allocas in AllocasForType to see if we can reuse one. Note
// that we have to be careful not to reuse the same "available" alloca for
// multiple different allocas that we just inlined, we use the 'UsedAllocas'
// set to keep track of which "available" allocas are being used by this
// function. Also, AllocasForType can be empty of course!
bool MergedAwayAlloca = false;
for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
AllocaInst *AvailableAlloca = AllocasForType[i];
// The available alloca has to be in the right function, not in some other
// function in this SCC.
if (AvailableAlloca->getParent() != AI->getParent())
continue;
// If the inlined function already uses this alloca then we can't reuse
// it.
if (!UsedAllocas.insert(AvailableAlloca))
continue;
// Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
// success!
DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI);
AI->replaceAllUsesWith(AvailableAlloca);
AI->eraseFromParent();
MergedAwayAlloca = true;
++NumMergedAllocas;
break;
}
// If we already nuked the alloca, we're done with it.
if (MergedAwayAlloca)
continue;
// If we were unable to merge away the alloca either because there are no
// allocas of the right type available or because we reused them all
// already, remember that this alloca came from an inlined function and mark
// it used so we don't reuse it for other allocas from this inline
// operation.
AllocasForType.push_back(AI);
UsedAllocas.insert(AI);
}
return true;
}
void AMDGPUPromoteAlloca::handleAlloca(AllocaInst &I) {
// Array allocations are probably not worth handling, since an allocation of
// the array type is the canonical form.
if (!I.isStaticAlloca() || I.isArrayAllocation())
return;
IRBuilder<> Builder(&I);
// First try to replace the alloca with a vector
Type *AllocaTy = I.getAllocatedType();
DEBUG(dbgs() << "Trying to promote " << I << '\n');
if (tryPromoteAllocaToVector(&I))
return;
DEBUG(dbgs() << " alloca is not a candidate for vectorization.\n");
const Function &ContainingFunction = *I.getParent()->getParent();
// FIXME: We should also try to get this value from the reqd_work_group_size
// function attribute if it is available.
unsigned WorkGroupSize = AMDGPU::getMaximumWorkGroupSize(ContainingFunction);
int AllocaSize =
WorkGroupSize * Mod->getDataLayout().getTypeAllocSize(AllocaTy);
if (AllocaSize > LocalMemAvailable) {
DEBUG(dbgs() << " Not enough local memory to promote alloca.\n");
return;
}
std::vector<Value*> WorkList;
if (!collectUsesWithPtrTypes(&I, WorkList)) {
DEBUG(dbgs() << " Do not know how to convert all uses\n");
return;
}
DEBUG(dbgs() << "Promoting alloca to local memory\n");
LocalMemAvailable -= AllocaSize;
Function *F = I.getParent()->getParent();
Type *GVTy = ArrayType::get(I.getAllocatedType(), WorkGroupSize);
GlobalVariable *GV = new GlobalVariable(
*Mod, GVTy, false, GlobalValue::InternalLinkage,
UndefValue::get(GVTy),
Twine(F->getName()) + Twine('.') + I.getName(),
nullptr,
GlobalVariable::NotThreadLocal,
AMDGPUAS::LOCAL_ADDRESS);
GV->setUnnamedAddr(true);
GV->setAlignment(I.getAlignment());
Value *TCntY, *TCntZ;
std::tie(TCntY, TCntZ) = getLocalSizeYZ(Builder);
Value *TIdX = getWorkitemID(Builder, 0);
Value *TIdY = getWorkitemID(Builder, 1);
Value *TIdZ = getWorkitemID(Builder, 2);
Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ, "", true, true);
Tmp0 = Builder.CreateMul(Tmp0, TIdX);
Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ, "", true, true);
Value *TID = Builder.CreateAdd(Tmp0, Tmp1);
TID = Builder.CreateAdd(TID, TIdZ);
Value *Indices[] = {
Constant::getNullValue(Type::getInt32Ty(Mod->getContext())),
TID
};
Value *Offset = Builder.CreateInBoundsGEP(GVTy, GV, Indices);
I.mutateType(Offset->getType());
I.replaceAllUsesWith(Offset);
I.eraseFromParent();
for (Value *V : WorkList) {
CallInst *Call = dyn_cast<CallInst>(V);
if (!Call) {
Type *EltTy = V->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
// The operand's value should be corrected on its own.
if (isa<AddrSpaceCastInst>(V))
continue;
// FIXME: It doesn't really make sense to try to do this for all
// instructions.
V->mutateType(NewTy);
continue;
}
IntrinsicInst *Intr = dyn_cast<IntrinsicInst>(Call);
if (!Intr) {
// FIXME: What is this for? It doesn't make sense to promote arbitrary
// function calls. If the call is to a defined function that can also be
// promoted, we should be able to do this once that function is also
// rewritten.
std::vector<Type*> ArgTypes;
for (unsigned ArgIdx = 0, ArgEnd = Call->getNumArgOperands();
ArgIdx != ArgEnd; ++ArgIdx) {
ArgTypes.push_back(Call->getArgOperand(ArgIdx)->getType());
}
Function *F = Call->getCalledFunction();
FunctionType *NewType = FunctionType::get(Call->getType(), ArgTypes,
F->isVarArg());
Constant *C = Mod->getOrInsertFunction((F->getName() + ".local").str(),
NewType, F->getAttributes());
Function *NewF = cast<Function>(C);
Call->setCalledFunction(NewF);
continue;
}
Builder.SetInsertPoint(Intr);
switch (Intr->getIntrinsicID()) {
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
// These intrinsics are for address space 0 only
Intr->eraseFromParent();
continue;
case Intrinsic::memcpy: {
MemCpyInst *MemCpy = cast<MemCpyInst>(Intr);
Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getRawSource(),
MemCpy->getLength(), MemCpy->getAlignment(),
MemCpy->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memmove: {
MemMoveInst *MemMove = cast<MemMoveInst>(Intr);
Builder.CreateMemMove(MemMove->getRawDest(), MemMove->getRawSource(),
MemMove->getLength(), MemMove->getAlignment(),
MemMove->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memset: {
MemSetInst *MemSet = cast<MemSetInst>(Intr);
Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(),
MemSet->getLength(), MemSet->getAlignment(),
MemSet->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::invariant_start:
case Intrinsic::invariant_end:
case Intrinsic::invariant_group_barrier:
Intr->eraseFromParent();
// FIXME: I think the invariant marker should still theoretically apply,
// but the intrinsics need to be changed to accept pointers with any
// address space.
continue;
case Intrinsic::objectsize: {
Value *Src = Intr->getOperand(0);
Type *SrcTy = Src->getType()->getPointerElementType();
Function *ObjectSize = Intrinsic::getDeclaration(Mod,
Intrinsic::objectsize,
{ Intr->getType(), PointerType::get(SrcTy, AMDGPUAS::LOCAL_ADDRESS) }
);
CallInst *NewCall
= Builder.CreateCall(ObjectSize, { Src, Intr->getOperand(1) });
Intr->replaceAllUsesWith(NewCall);
Intr->eraseFromParent();
continue;
}
default:
Intr->dump();
llvm_unreachable("Don't know how to promote alloca intrinsic use.");
}
}
}
Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
if (auto *I = simplifyAllocaArraySize(*this, AI))
return I;
if (AI.getAllocatedType()->isSized()) {
// If the alignment is 0 (unspecified), assign it the preferred alignment.
if (AI.getAlignment() == 0)
AI.setAlignment(DL.getPrefTypeAlignment(AI.getAllocatedType()));
// Move all alloca's of zero byte objects to the entry block and merge them
// together. Note that we only do this for alloca's, because malloc should
// allocate and return a unique pointer, even for a zero byte allocation.
if (DL.getTypeAllocSize(AI.getAllocatedType()) == 0) {
// For a zero sized alloca there is no point in doing an array allocation.
// This is helpful if the array size is a complicated expression not used
// elsewhere.
if (AI.isArrayAllocation()) {
AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
return &AI;
}
// Get the first instruction in the entry block.
BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
if (FirstInst != &AI) {
// If the entry block doesn't start with a zero-size alloca then move
// this one to the start of the entry block. There is no problem with
// dominance as the array size was forced to a constant earlier already.
AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
DL.getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
AI.moveBefore(FirstInst);
return &AI;
}
// If the alignment of the entry block alloca is 0 (unspecified),
// assign it the preferred alignment.
if (EntryAI->getAlignment() == 0)
EntryAI->setAlignment(
DL.getPrefTypeAlignment(EntryAI->getAllocatedType()));
// Replace this zero-sized alloca with the one at the start of the entry
// block after ensuring that the address will be aligned enough for both
// types.
unsigned MaxAlign = std::max(EntryAI->getAlignment(),
AI.getAlignment());
EntryAI->setAlignment(MaxAlign);
if (AI.getType() != EntryAI->getType())
return new BitCastInst(EntryAI, AI.getType());
return ReplaceInstUsesWith(AI, EntryAI);
}
}
}
if (AI.getAlignment()) {
// Check to see if this allocation is only modified by a memcpy/memmove from
// a constant global whose alignment is equal to or exceeds that of the
// allocation. If this is the case, we can change all users to use
// the constant global instead. This is commonly produced by the CFE by
// constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
// is only subsequently read.
SmallVector<Instruction *, 4> ToDelete;
if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
unsigned SourceAlign = getOrEnforceKnownAlignment(
Copy->getSource(), AI.getAlignment(), DL, &AI, AC, DT);
if (AI.getAlignment() <= SourceAlign) {
DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
EraseInstFromFunction(*ToDelete[i]);
Constant *TheSrc = cast<Constant>(Copy->getSource());
Constant *Cast
= ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, AI.getType());
Instruction *NewI = ReplaceInstUsesWith(AI, Cast);
EraseInstFromFunction(*Copy);
++NumGlobalCopies;
return NewI;
}
}
}
// At last, use the generic allocation site handler to aggressively remove
// unused allocas.
return visitAllocSite(AI);
}
// emit code to unpack a raw value from a box into registers or a stack slot
static Value *emit_unbox(jl_codectx_t &ctx, Type *to, const jl_cgval_t &x, jl_value_t *jt, Value *dest, bool volatile_store)
{
assert(to != T_void);
// TODO: fully validate that x.typ == jt?
if (x.isghost) {
// this can happen when a branch yielding a different type ends
// up being dead code, and type inference knows that the other
// branch's type is the only one that matters.
if (type_is_ghost(to)) {
return NULL;
}
//emit_error(ctx, "emit_unbox: a type mismatch error in occurred during codegen");
return UndefValue::get(to); // type mismatch error
}
Constant *c = x.constant ? julia_const_to_llvm(x.constant) : NULL;
if (!x.ispointer() || c) { // already unboxed, but sometimes need conversion
Value *unboxed = emit_unboxed_coercion(ctx, to, c ? c : x.V);
if (!dest)
return unboxed;
Type *dest_ty = unboxed->getType()->getPointerTo();
if (dest->getType() != dest_ty)
dest = emit_bitcast(ctx, dest, dest_ty);
ctx.builder.CreateStore(unboxed, dest, volatile_store);
return NULL;
}
// bools stored as int8, so an extra Trunc is needed to get an int1
Value *p = x.constant ? literal_pointer_val(ctx, x.constant) : x.V;
Type *ptype = (to == T_int1 ? T_pint8 : to->getPointerTo());
Value *unboxed = NULL;
if (to == T_int1)
unboxed = ctx.builder.CreateTrunc(tbaa_decorate(x.tbaa, ctx.builder.CreateLoad(maybe_bitcast(ctx, p, ptype))), T_int1);
else if (jt == (jl_value_t*)jl_bool_type)
unboxed = ctx.builder.CreateZExt(ctx.builder.CreateTrunc(tbaa_decorate(x.tbaa, ctx.builder.CreateLoad(maybe_bitcast(ctx, p, ptype))), T_int1), to);
if (unboxed) {
if (!dest)
return unboxed;
ctx.builder.CreateStore(unboxed, dest);
return NULL;
}
unsigned alignment = julia_alignment(jt, 0);
if (dest) {
MDNode *tbaa = x.tbaa;
// the memcpy intrinsic does not allow to specify different alias tags
// for the load part (x.tbaa) and the store part (tbaa_stack).
// since the tbaa lattice has to be a tree we have unfortunately
// x.tbaa ∪ tbaa_stack = tbaa_root if x.tbaa != tbaa_stack
if (tbaa != tbaa_stack)
tbaa = NULL;
emit_memcpy(ctx, dest, p, jl_datatype_size(jt), alignment, volatile_store, tbaa);
return NULL;
}
else {
if (p->getType() != ptype && isa<AllocaInst>(p)) {
// LLVM's mem2reg can't handle coercion if the load/store type does
// not match the type of the alloca. As such, it is better to
// perform the load using the alloca's type and then perform the
// appropriate coercion manually.
AllocaInst *AI = cast<AllocaInst>(p);
Type *AllocType = AI->getAllocatedType();
#if JL_LLVM_VERSION >= 40000
const DataLayout &DL = jl_data_layout;
#else
const DataLayout &DL = jl_ExecutionEngine->getDataLayout();
#endif
if (!AI->isArrayAllocation() &&
(AllocType->isFloatingPointTy() || AllocType->isIntegerTy() || AllocType->isPointerTy()) &&
(to->isFloatingPointTy() || to->isIntegerTy() || to->isPointerTy()) &&
DL.getTypeSizeInBits(AllocType) == DL.getTypeSizeInBits(to)) {
Instruction *load = ctx.builder.CreateAlignedLoad(p, alignment);
return emit_unboxed_coercion(ctx, to, tbaa_decorate(x.tbaa, load));
}
}
p = maybe_bitcast(ctx, p, ptype);
Instruction *load = ctx.builder.CreateAlignedLoad(p, alignment);
return tbaa_decorate(x.tbaa, load);
}
}
/// If it is possible to inline the specified call site,
/// do so and update the CallGraph for this operation.
///
/// This function also does some basic book-keeping to update the IR. The
/// InlinedArrayAllocas map keeps track of any allocas that are already
/// available from other functions inlined into the caller. If we are able to
/// inline this call site we attempt to reuse already available allocas or add
/// any new allocas to the set if not possible.
static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI,
InlinedArrayAllocasTy &InlinedArrayAllocas,
int InlineHistory, bool InsertLifetime) {
Function *Callee = CS.getCalledFunction();
Function *Caller = CS.getCaller();
// Try to inline the function. Get the list of static allocas that were
// inlined.
if (!InlineFunction(CS, IFI, InsertLifetime))
return false;
AdjustCallerSSPLevel(Caller, Callee);
// Look at all of the allocas that we inlined through this call site. If we
// have already inlined other allocas through other calls into this function,
// then we know that they have disjoint lifetimes and that we can merge them.
//
// There are many heuristics possible for merging these allocas, and the
// different options have different tradeoffs. One thing that we *really*
// don't want to hurt is SRoA: once inlining happens, often allocas are no
// longer address taken and so they can be promoted.
//
// Our "solution" for that is to only merge allocas whose outermost type is an
// array type. These are usually not promoted because someone is using a
// variable index into them. These are also often the most important ones to
// merge.
//
// A better solution would be to have real memory lifetime markers in the IR
// and not have the inliner do any merging of allocas at all. This would
// allow the backend to do proper stack slot coloring of all allocas that
// *actually make it to the backend*, which is really what we want.
//
// Because we don't have this information, we do this simple and useful hack.
//
SmallPtrSet<AllocaInst*, 16> UsedAllocas;
// When processing our SCC, check to see if CS was inlined from some other
// call site. For example, if we're processing "A" in this code:
// A() { B() }
// B() { x = alloca ... C() }
// C() { y = alloca ... }
// Assume that C was not inlined into B initially, and so we're processing A
// and decide to inline B into A. Doing this makes an alloca available for
// reuse and makes a callsite (C) available for inlining. When we process
// the C call site we don't want to do any alloca merging between X and Y
// because their scopes are not disjoint. We could make this smarter by
// keeping track of the inline history for each alloca in the
// InlinedArrayAllocas but this isn't likely to be a significant win.
if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
return true;
// Loop over all the allocas we have so far and see if they can be merged with
// a previously inlined alloca. If not, remember that we had it.
for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size();
AllocaNo != e; ++AllocaNo) {
AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
// Don't bother trying to merge array allocations (they will usually be
// canonicalized to be an allocation *of* an array), or allocations whose
// type is not itself an array (because we're afraid of pessimizing SRoA).
ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
if (!ATy || AI->isArrayAllocation())
continue;
// Get the list of all available allocas for this array type.
std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
// Loop over the allocas in AllocasForType to see if we can reuse one. Note
// that we have to be careful not to reuse the same "available" alloca for
// multiple different allocas that we just inlined, we use the 'UsedAllocas'
// set to keep track of which "available" allocas are being used by this
// function. Also, AllocasForType can be empty of course!
bool MergedAwayAlloca = false;
for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
AllocaInst *AvailableAlloca = AllocasForType[i];
unsigned Align1 = AI->getAlignment(),
Align2 = AvailableAlloca->getAlignment();
// The available alloca has to be in the right function, not in some other
// function in this SCC.
if (AvailableAlloca->getParent() != AI->getParent())
continue;
// If the inlined function already uses this alloca then we can't reuse
// it.
if (!UsedAllocas.insert(AvailableAlloca).second)
continue;
// Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
// success!
DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: "
<< *AvailableAlloca << '\n');
AI->replaceAllUsesWith(AvailableAlloca);
if (Align1 != Align2) {
if (!Align1 || !Align2) {
const DataLayout &DL = Caller->getParent()->getDataLayout();
unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType());
Align1 = Align1 ? Align1 : TypeAlign;
Align2 = Align2 ? Align2 : TypeAlign;
}
if (Align1 > Align2)
AvailableAlloca->setAlignment(AI->getAlignment());
}
AI->eraseFromParent();
MergedAwayAlloca = true;
++NumMergedAllocas;
IFI.StaticAllocas[AllocaNo] = nullptr;
break;
}
// If we already nuked the alloca, we're done with it.
if (MergedAwayAlloca)
continue;
// If we were unable to merge away the alloca either because there are no
// allocas of the right type available or because we reused them all
// already, remember that this alloca came from an inlined function and mark
// it used so we don't reuse it for other allocas from this inline
// operation.
AllocasForType.push_back(AI);
UsedAllocas.insert(AI);
}
return true;
}
bool AllocaArraysMerging::runOnFunction(Function& F)
{
class ArraysToMerge
{
private:
std::map<AllocaInst*, uint32_t> arraysToMerge;
uint32_t currentOffset;
public:
ArraysToMerge():currentOffset(0)
{
}
bool empty() const
{
return arraysToMerge.empty();
}
std::map<AllocaInst*, uint32_t>::iterator begin()
{
return arraysToMerge.begin();
}
std::map<AllocaInst*, uint32_t>::iterator end()
{
return arraysToMerge.end();
}
void add(AllocaInst* a)
{
arraysToMerge.insert(std::make_pair(a, currentOffset));
currentOffset+=cast<ArrayType>(a->getAllocatedType())->getNumElements();
}
uint32_t getNewSize() const
{
return currentOffset;
}
};
cheerp::PointerAnalyzer & PA = getAnalysis<cheerp::PointerAnalyzer>();
cheerp::Registerize & registerize = getAnalysis<cheerp::Registerize>();
cheerp::GlobalDepsAnalyzer & GDA = getAnalysis<cheerp::GlobalDepsAnalyzer>();
std::list<std::pair<AllocaInst*, cheerp::Registerize::LiveRange>> allocaInfos;
// Gather all the allocas
for(BasicBlock& BB: F)
analyzeBlock(registerize, BB, allocaInfos);
if (allocaInfos.size() < 2)
return false;
bool Changed = false;
// We can also try to merge arrays of the same type, if only pointers to values are passed around
while(!allocaInfos.empty())
{
// Build a map of array to be merged and their offseet into the new array
ArraysToMerge arraysToMerge;
auto targetCandidate = allocaInfos.begin();
AllocaInst* targetAlloca = targetCandidate->first;
if(!targetAlloca->getAllocatedType()->isArrayTy() ||
// Check target uses
!checkUsesForArrayMerging(targetAlloca))
{
allocaInfos.erase(targetCandidate);
continue;
}
Type* targetElementType = targetAlloca->getAllocatedType()->getSequentialElementType();
auto sourceCandidate=targetCandidate;
++sourceCandidate;
// Now that we have computed the sourceCandidate we can invalidate the targetCandidate
allocaInfos.erase(targetCandidate);
while(sourceCandidate!=allocaInfos.end())
{
AllocaInst* sourceAlloca = sourceCandidate->first;
// Check that allocas are arrays of the same type
if(!sourceAlloca->getAllocatedType()->isArrayTy())
{
++sourceCandidate;
continue;
}
// Both are arrays, check the types
if(targetElementType != sourceAlloca->getAllocatedType()->getSequentialElementType())
{
++sourceCandidate;
continue;
}
// Verify that the source candidate has supported uses
if(!checkUsesForArrayMerging(sourceAlloca))
{
++sourceCandidate;
continue;
}
// We can merge the source and the target
// If the set is empty add the target as well
if(arraysToMerge.empty())
arraysToMerge.add(targetAlloca);
arraysToMerge.add(sourceAlloca);
auto oldCandidate = sourceCandidate;
++sourceCandidate;
// Now that we have moved to the next candidate, we can invalidate the old one
allocaInfos.erase(oldCandidate);
}
// If we have a non-empty set of alloca merge them
if (arraysToMerge.empty())
continue;
if(!Changed)
registerize.invalidateLiveRangeForAllocas(F);
// Build new alloca
Type* newAllocaType = ArrayType::get(targetElementType, arraysToMerge.getNewSize());
// Add the new struct type to the GlobalDepsAnalyzer, it may need the createArray helper
GDA.visitType(newAllocaType, /*forceTypedArray*/ false);
AllocaInst* newAlloca = new AllocaInst(newAllocaType, "mergedArray", &(*F.getEntryBlock().begin()));
Type* indexType = IntegerType::get(newAllocaType->getContext(), 32);
// Change every use of every merged array with an appropiate GEP
for(auto it: arraysToMerge)
{
AllocaInst* allocaToMerge = it.first;
uint32_t baseOffset = it.second;
SmallVector<User*, 4> users(allocaToMerge->users());
for(User* u: users)
{
if(GetElementPtrInst* oldGep = dyn_cast<GetElementPtrInst>(u))
{
// Build 2 GEPs, one to reach the first element in the merged array
// and the other for the rest of the offsets
SmallVector<Value*, 4> indices;
// Dereference array
indices.push_back(ConstantInt::get(indexType, 0));
// Reach offset
indices.push_back(ConstantInt::get(indexType, baseOffset));
Value* gep1 = GetElementPtrInst::Create(newAlloca, indices, "", oldGep);
// Apply all the old offsets but the first one using a new GEP
indices.clear();
indices.insert(indices.begin(), oldGep->idx_begin()+1, oldGep->idx_end());
Value* gep2 = GetElementPtrInst::Create(gep1, indices, "", oldGep);
// Replace all uses with gep2
oldGep->replaceAllUsesWith(gep2);
PA.invalidate(oldGep);
oldGep->eraseFromParent();
}
else if(BitCastInst* BI=dyn_cast<BitCastInst>(u))
{
//Only used for lifetime intrinsics
Value* newBitCast=new BitCastInst(newAlloca, BI->getType(), "", BI);
BI->replaceAllUsesWith(newBitCast);
PA.invalidate(BI);
BI->eraseFromParent();
}
else
assert(false && "Unexpected use while merging arrays");
}
// Kill the alloca itself now
PA.invalidate(allocaToMerge);
allocaToMerge->eraseFromParent();
Changed = true;
}
}
if(Changed)
registerize.computeLiveRangeForAllocas(F);
return Changed;
}
// FIXME: Should try to pick the most likely to be profitable allocas first.
bool AMDGPUPromoteAlloca::handleAlloca(AllocaInst &I, bool SufficientLDS) {
// Array allocations are probably not worth handling, since an allocation of
// the array type is the canonical form.
if (!I.isStaticAlloca() || I.isArrayAllocation())
return false;
IRBuilder<> Builder(&I);
// First try to replace the alloca with a vector
Type *AllocaTy = I.getAllocatedType();
DEBUG(dbgs() << "Trying to promote " << I << '\n');
if (tryPromoteAllocaToVector(&I, AS))
return true; // Promoted to vector.
const Function &ContainingFunction = *I.getParent()->getParent();
CallingConv::ID CC = ContainingFunction.getCallingConv();
// Don't promote the alloca to LDS for shader calling conventions as the work
// item ID intrinsics are not supported for these calling conventions.
// Furthermore not all LDS is available for some of the stages.
switch (CC) {
case CallingConv::AMDGPU_KERNEL:
case CallingConv::SPIR_KERNEL:
break;
default:
DEBUG(dbgs() << " promote alloca to LDS not supported with calling convention.\n");
return false;
}
// Not likely to have sufficient local memory for promotion.
if (!SufficientLDS)
return false;
const AMDGPUSubtarget &ST =
TM->getSubtarget<AMDGPUSubtarget>(ContainingFunction);
unsigned WorkGroupSize = ST.getFlatWorkGroupSizes(ContainingFunction).second;
const DataLayout &DL = Mod->getDataLayout();
unsigned Align = I.getAlignment();
if (Align == 0)
Align = DL.getABITypeAlignment(I.getAllocatedType());
// FIXME: This computed padding is likely wrong since it depends on inverse
// usage order.
//
// FIXME: It is also possible that if we're allowed to use all of the memory
// could could end up using more than the maximum due to alignment padding.
uint32_t NewSize = alignTo(CurrentLocalMemUsage, Align);
uint32_t AllocSize = WorkGroupSize * DL.getTypeAllocSize(AllocaTy);
NewSize += AllocSize;
if (NewSize > LocalMemLimit) {
DEBUG(dbgs() << " " << AllocSize
<< " bytes of local memory not available to promote\n");
return false;
}
CurrentLocalMemUsage = NewSize;
std::vector<Value*> WorkList;
if (!collectUsesWithPtrTypes(&I, &I, WorkList)) {
DEBUG(dbgs() << " Do not know how to convert all uses\n");
return false;
}
DEBUG(dbgs() << "Promoting alloca to local memory\n");
Function *F = I.getParent()->getParent();
Type *GVTy = ArrayType::get(I.getAllocatedType(), WorkGroupSize);
GlobalVariable *GV = new GlobalVariable(
*Mod, GVTy, false, GlobalValue::InternalLinkage,
UndefValue::get(GVTy),
Twine(F->getName()) + Twine('.') + I.getName(),
nullptr,
GlobalVariable::NotThreadLocal,
AS.LOCAL_ADDRESS);
GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
GV->setAlignment(I.getAlignment());
Value *TCntY, *TCntZ;
std::tie(TCntY, TCntZ) = getLocalSizeYZ(Builder);
Value *TIdX = getWorkitemID(Builder, 0);
Value *TIdY = getWorkitemID(Builder, 1);
Value *TIdZ = getWorkitemID(Builder, 2);
Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ, "", true, true);
Tmp0 = Builder.CreateMul(Tmp0, TIdX);
Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ, "", true, true);
Value *TID = Builder.CreateAdd(Tmp0, Tmp1);
TID = Builder.CreateAdd(TID, TIdZ);
Value *Indices[] = {
Constant::getNullValue(Type::getInt32Ty(Mod->getContext())),
TID
};
Value *Offset = Builder.CreateInBoundsGEP(GVTy, GV, Indices);
I.mutateType(Offset->getType());
I.replaceAllUsesWith(Offset);
I.eraseFromParent();
for (Value *V : WorkList) {
CallInst *Call = dyn_cast<CallInst>(V);
if (!Call) {
if (ICmpInst *CI = dyn_cast<ICmpInst>(V)) {
Value *Src0 = CI->getOperand(0);
Type *EltTy = Src0->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AS.LOCAL_ADDRESS);
if (isa<ConstantPointerNull>(CI->getOperand(0)))
CI->setOperand(0, ConstantPointerNull::get(NewTy));
if (isa<ConstantPointerNull>(CI->getOperand(1)))
CI->setOperand(1, ConstantPointerNull::get(NewTy));
continue;
}
// The operand's value should be corrected on its own and we don't want to
// touch the users.
if (isa<AddrSpaceCastInst>(V))
continue;
Type *EltTy = V->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AS.LOCAL_ADDRESS);
// FIXME: It doesn't really make sense to try to do this for all
// instructions.
V->mutateType(NewTy);
// Adjust the types of any constant operands.
if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
if (isa<ConstantPointerNull>(SI->getOperand(1)))
SI->setOperand(1, ConstantPointerNull::get(NewTy));
if (isa<ConstantPointerNull>(SI->getOperand(2)))
SI->setOperand(2, ConstantPointerNull::get(NewTy));
} else if (PHINode *Phi = dyn_cast<PHINode>(V)) {
for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
if (isa<ConstantPointerNull>(Phi->getIncomingValue(I)))
Phi->setIncomingValue(I, ConstantPointerNull::get(NewTy));
}
}
continue;
}
IntrinsicInst *Intr = cast<IntrinsicInst>(Call);
Builder.SetInsertPoint(Intr);
switch (Intr->getIntrinsicID()) {
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
// These intrinsics are for address space 0 only
Intr->eraseFromParent();
continue;
case Intrinsic::memcpy: {
MemCpyInst *MemCpy = cast<MemCpyInst>(Intr);
Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getDestAlignment(),
MemCpy->getRawSource(), MemCpy->getSourceAlignment(),
MemCpy->getLength(), MemCpy->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memmove: {
MemMoveInst *MemMove = cast<MemMoveInst>(Intr);
Builder.CreateMemMove(MemMove->getRawDest(), MemMove->getDestAlignment(),
MemMove->getRawSource(), MemMove->getSourceAlignment(),
MemMove->getLength(), MemMove->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memset: {
MemSetInst *MemSet = cast<MemSetInst>(Intr);
Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(),
MemSet->getLength(), MemSet->getDestAlignment(),
MemSet->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::invariant_start:
case Intrinsic::invariant_end:
case Intrinsic::invariant_group_barrier:
Intr->eraseFromParent();
// FIXME: I think the invariant marker should still theoretically apply,
// but the intrinsics need to be changed to accept pointers with any
// address space.
continue;
case Intrinsic::objectsize: {
Value *Src = Intr->getOperand(0);
Type *SrcTy = Src->getType()->getPointerElementType();
Function *ObjectSize = Intrinsic::getDeclaration(Mod,
Intrinsic::objectsize,
{ Intr->getType(), PointerType::get(SrcTy, AS.LOCAL_ADDRESS) }
);
CallInst *NewCall = Builder.CreateCall(
ObjectSize, {Src, Intr->getOperand(1), Intr->getOperand(2)});
Intr->replaceAllUsesWith(NewCall);
Intr->eraseFromParent();
continue;
}
default:
Intr->print(errs());
llvm_unreachable("Don't know how to promote alloca intrinsic use.");
}
}
return true;
}
/// performCallSlotOptzn - takes a memcpy and a call that it depends on,
/// and checks for the possibility of a call slot optimization by having
/// the call write its result directly into the destination of the memcpy.
bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
Value *cpyDest, Value *cpySrc,
uint64_t cpyLen, unsigned cpyAlign,
CallInst *C) {
// The general transformation to keep in mind is
//
// call @func(..., src, ...)
// memcpy(dest, src, ...)
//
// ->
//
// memcpy(dest, src, ...)
// call @func(..., dest, ...)
//
// Since moving the memcpy is technically awkward, we additionally check that
// src only holds uninitialized values at the moment of the call, meaning that
// the memcpy can be discarded rather than moved.
// Deliberately get the source and destination with bitcasts stripped away,
// because we'll need to do type comparisons based on the underlying type.
CallSite CS(C);
// Require that src be an alloca. This simplifies the reasoning considerably.
AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
if (!srcAlloca)
return false;
ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
if (!srcArraySize)
return false;
const DataLayout &DL = cpy->getModule()->getDataLayout();
uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) *
srcArraySize->getZExtValue();
if (cpyLen < srcSize)
return false;
// Check that accessing the first srcSize bytes of dest will not cause a
// trap. Otherwise the transform is invalid since it might cause a trap
// to occur earlier than it otherwise would.
if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
// The destination is an alloca. Check it is larger than srcSize.
ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
if (!destArraySize)
return false;
uint64_t destSize = DL.getTypeAllocSize(A->getAllocatedType()) *
destArraySize->getZExtValue();
if (destSize < srcSize)
return false;
} else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
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;
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;
}
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->user_begin(),
srcAlloca->user_end());
while (!srcUseList.empty()) {
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;
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<DominatorTreeWrapperPass>().getDomTree();
if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
if (!DT.dominates(cpyDestInst, C))
return false;
// In addition to knowing that the call does not access src in some
// unexpected manner, for example via a global, which we deduce from
// 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>();
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) {
Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest
: CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
cpyDest->getName(), C);
changedArgument = true;
if (CS.getArgument(i)->getType() == Dest->getType())
CS.setArgument(i, Dest);
else
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;
return true;
}
//
// Method: insertBadAllocationSizes()
//
// Description:
// This method will look for allocations and change their size to be
// incorrect. It does the following:
// o) Changes the number of array elements allocated by alloca and malloc.
//
// Return value:
// true - The module was modified.
// false - The module was left unmodified.
//
bool
FaultInjector::insertBadAllocationSizes (Function & F) {
// Worklist of allocation sites to rewrite
std::vector<AllocaInst * > WorkList;
for (Function::iterator fI = F.begin(), fE = F.end(); fI != fE; ++fI) {
BasicBlock & BB = *fI;
for (BasicBlock::iterator I = BB.begin(), bE = BB.end(); I != bE; ++I) {
if (AllocaInst * AI = dyn_cast<AllocaInst>(I)) {
if (AI->isArrayAllocation()) {
// Skip if we should not insert a fault.
if (!doFault()) continue;
WorkList.push_back(AI);
}
}
}
}
while (WorkList.size()) {
AllocaInst * AI = WorkList.back();
WorkList.pop_back();
//
// Print information about where the fault is being inserted.
//
printSourceInfo ("Bad allocation size", AI);
Instruction * NewAlloc = 0;
NewAlloc = new AllocaInst (AI->getAllocatedType(),
ConstantInt::get(Int32Type,0),
AI->getAlignment(),
AI->getName(),
AI);
AI->replaceAllUsesWith (NewAlloc);
AI->eraseFromParent();
++BadSizes;
}
//
// Try harder to make bad allocation sizes.
//
WorkList.clear();
for (Function::iterator fI = F.begin(), fE = F.end(); fI != fE; ++fI) {
BasicBlock & BB = *fI;
for (BasicBlock::iterator I = BB.begin(), bE = BB.end(); I != bE; ++I) {
if (AllocaInst * AI = dyn_cast<AllocaInst>(I)) {
//
// Determine if this is a data type that we can make smaller.
//
if (((TD->getTypeAllocSize(AI->getAllocatedType())) > 4) && doFault()) {
WorkList.push_back(AI);
}
}
}
}
//
// Replace these allocations with an allocation of an integer and cast the
// result back into the appropriate type.
//
while (WorkList.size()) {
AllocaInst * AI = WorkList.back();
WorkList.pop_back();
Instruction * NewAlloc = 0;
NewAlloc = new AllocaInst (Int32Type,
AI->getArraySize(),
AI->getAlignment(),
AI->getName(),
AI);
NewAlloc = castTo (NewAlloc, AI->getType(), "", AI);
AI->replaceAllUsesWith (NewAlloc);
AI->eraseFromParent();
++BadSizes;
}
return (BadSizes > 0);
}
bool AllocaMerging::runOnFunction(Function& F)
{
cheerp::PointerAnalyzer & PA = getAnalysis<cheerp::PointerAnalyzer>();
cheerp::Registerize & registerize = getAnalysis<cheerp::Registerize>();
cheerp::TypeSupport types(*F.getParent());
AllocaInfos allocaInfos;
// Gather all the allocas
for(BasicBlock& BB: F)
analyzeBlock(registerize, BB, allocaInfos);
if (allocaInfos.size() < 2)
return false;
bool Changed = false;
BasicBlock& entryBlock=F.getEntryBlock();
// Look if we can merge allocas of the same type
for(auto targetCandidate=allocaInfos.begin();targetCandidate!=allocaInfos.end();++targetCandidate)
{
AllocaInst* targetAlloca = targetCandidate->first;
Type* targetType = targetAlloca->getAllocatedType();
// The range storing the sum of all ranges merged into target
cheerp::Registerize::LiveRange targetRange(targetCandidate->second);
// If the range is empty, we have an alloca that we can't analyze
if (targetRange.empty())
continue;
std::vector<AllocaInfos::iterator> mergeSet;
auto sourceCandidate=targetCandidate;
++sourceCandidate;
for(;sourceCandidate!=allocaInfos.end();++sourceCandidate)
{
AllocaInst* sourceAlloca = sourceCandidate->first;
Type* sourceType = sourceAlloca->getAllocatedType();
// Bail out for non compatible types
if(!areTypesEquivalent(types, PA, targetType, sourceType))
continue;
const cheerp::Registerize::LiveRange& sourceRange = sourceCandidate->second;
// Bail out if this source candidate is not analyzable
if(sourceRange.empty())
continue;
// Bail out if the allocas interfere
if(targetRange.doesInterfere(sourceRange))
continue;
// Add the range to the target range and the source alloca to the mergeSet
mergeSet.push_back(sourceCandidate);
PA.invalidate(sourceAlloca);
targetRange.merge(sourceRange);
}
// If the merge set is empty try another target
if(mergeSet.empty())
continue;
PA.invalidate(targetAlloca);
if(!Changed)
registerize.invalidateLiveRangeForAllocas(F);
// Make sure that this alloca is in the entry block
if(targetAlloca->getParent()!=&entryBlock)
targetAlloca->moveBefore(entryBlock.begin());
// We can merge the allocas
for(const AllocaInfos::iterator& it: mergeSet)
{
AllocaInst* allocaToMerge = it->first;
Instruction* targetVal=targetAlloca;
if(targetVal->getType()!=allocaToMerge->getType())
{
targetVal=new BitCastInst(targetVal, allocaToMerge->getType());
targetVal->insertAfter(targetAlloca);
}
allocaToMerge->replaceAllUsesWith(targetVal);
allocaToMerge->eraseFromParent();
if(targetVal != targetAlloca)
PA.getPointerKind(targetVal);
allocaInfos.erase(it);
NumAllocaMerged++;
}
PA.getPointerKind(targetAlloca);
Changed = true;
}
if(Changed)
registerize.computeLiveRangeForAllocas(F);
return Changed;
}
void AMDGPUPromoteAlloca::visitAlloca(AllocaInst &I) {
IRBuilder<> Builder(&I);
// First try to replace the alloca with a vector
Type *AllocaTy = I.getAllocatedType();
DEBUG(dbgs() << "Trying to promote " << I << '\n');
if (tryPromoteAllocaToVector(&I))
return;
DEBUG(dbgs() << " alloca is not a candidate for vectorization.\n");
// FIXME: This is the maximum work group size. We should try to get
// value from the reqd_work_group_size function attribute if it is
// available.
unsigned WorkGroupSize = 256;
int AllocaSize = WorkGroupSize *
Mod->getDataLayout()->getTypeAllocSize(AllocaTy);
if (AllocaSize > LocalMemAvailable) {
DEBUG(dbgs() << " Not enough local memory to promote alloca.\n");
return;
}
std::vector<Value*> WorkList;
if (!collectUsesWithPtrTypes(&I, WorkList)) {
DEBUG(dbgs() << " Do not know how to convert all uses\n");
return;
}
DEBUG(dbgs() << "Promoting alloca to local memory\n");
LocalMemAvailable -= AllocaSize;
GlobalVariable *GV = new GlobalVariable(
*Mod, ArrayType::get(I.getAllocatedType(), 256), false,
GlobalValue::ExternalLinkage, 0, I.getName(), 0,
GlobalVariable::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS);
FunctionType *FTy = FunctionType::get(
Type::getInt32Ty(Mod->getContext()), false);
AttributeSet AttrSet;
AttrSet.addAttribute(Mod->getContext(), 0, Attribute::ReadNone);
Value *ReadLocalSizeY = Mod->getOrInsertFunction(
"llvm.r600.read.local.size.y", FTy, AttrSet);
Value *ReadLocalSizeZ = Mod->getOrInsertFunction(
"llvm.r600.read.local.size.z", FTy, AttrSet);
Value *ReadTIDIGX = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.x", FTy, AttrSet);
Value *ReadTIDIGY = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.y", FTy, AttrSet);
Value *ReadTIDIGZ = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.z", FTy, AttrSet);
Value *TCntY = Builder.CreateCall(ReadLocalSizeY);
Value *TCntZ = Builder.CreateCall(ReadLocalSizeZ);
Value *TIdX = Builder.CreateCall(ReadTIDIGX);
Value *TIdY = Builder.CreateCall(ReadTIDIGY);
Value *TIdZ = Builder.CreateCall(ReadTIDIGZ);
Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ);
Tmp0 = Builder.CreateMul(Tmp0, TIdX);
Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ);
Value *TID = Builder.CreateAdd(Tmp0, Tmp1);
TID = Builder.CreateAdd(TID, TIdZ);
std::vector<Value*> Indices;
Indices.push_back(Constant::getNullValue(Type::getInt32Ty(Mod->getContext())));
Indices.push_back(TID);
Value *Offset = Builder.CreateGEP(GV, Indices);
I.mutateType(Offset->getType());
I.replaceAllUsesWith(Offset);
I.eraseFromParent();
for (std::vector<Value*>::iterator i = WorkList.begin(),
e = WorkList.end(); i != e; ++i) {
Value *V = *i;
CallInst *Call = dyn_cast<CallInst>(V);
if (!Call) {
Type *EltTy = V->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
// The operand's value should be corrected on its own.
if (isa<AddrSpaceCastInst>(V))
continue;
// FIXME: It doesn't really make sense to try to do this for all
// instructions.
V->mutateType(NewTy);
continue;
}
IntrinsicInst *Intr = dyn_cast<IntrinsicInst>(Call);
if (!Intr) {
std::vector<Type*> ArgTypes;
for (unsigned ArgIdx = 0, ArgEnd = Call->getNumArgOperands();
ArgIdx != ArgEnd; ++ArgIdx) {
ArgTypes.push_back(Call->getArgOperand(ArgIdx)->getType());
}
Function *F = Call->getCalledFunction();
FunctionType *NewType = FunctionType::get(Call->getType(), ArgTypes,
F->isVarArg());
Constant *C = Mod->getOrInsertFunction(StringRef(F->getName().str() + ".local"), NewType,
F->getAttributes());
Function *NewF = cast<Function>(C);
Call->setCalledFunction(NewF);
continue;
}
Builder.SetInsertPoint(Intr);
switch (Intr->getIntrinsicID()) {
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
// These intrinsics are for address space 0 only
Intr->eraseFromParent();
continue;
case Intrinsic::memcpy: {
MemCpyInst *MemCpy = cast<MemCpyInst>(Intr);
Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getRawSource(),
MemCpy->getLength(), MemCpy->getAlignment(),
MemCpy->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memset: {
MemSetInst *MemSet = cast<MemSetInst>(Intr);
Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(),
MemSet->getLength(), MemSet->getAlignment(),
MemSet->isVolatile());
Intr->eraseFromParent();
continue;
}
default:
Intr->dump();
llvm_unreachable("Don't know how to promote alloca intrinsic use.");
}
}
}
/// Look at all of the allocas that we inlined through this call site. If we
/// have already inlined other allocas through other calls into this function,
/// then we know that they have disjoint lifetimes and that we can merge them.
///
/// There are many heuristics possible for merging these allocas, and the
/// different options have different tradeoffs. One thing that we *really*
/// don't want to hurt is SRoA: once inlining happens, often allocas are no
/// longer address taken and so they can be promoted.
///
/// Our "solution" for that is to only merge allocas whose outermost type is an
/// array type. These are usually not promoted because someone is using a
/// variable index into them. These are also often the most important ones to
/// merge.
///
/// A better solution would be to have real memory lifetime markers in the IR
/// and not have the inliner do any merging of allocas at all. This would
/// allow the backend to do proper stack slot coloring of all allocas that
/// *actually make it to the backend*, which is really what we want.
///
/// Because we don't have this information, we do this simple and useful hack.
static void mergeInlinedArrayAllocas(
Function *Caller, InlineFunctionInfo &IFI,
InlinedArrayAllocasTy &InlinedArrayAllocas, int InlineHistory) {
SmallPtrSet<AllocaInst *, 16> UsedAllocas;
// When processing our SCC, check to see if CS was inlined from some other
// call site. For example, if we're processing "A" in this code:
// A() { B() }
// B() { x = alloca ... C() }
// C() { y = alloca ... }
// Assume that C was not inlined into B initially, and so we're processing A
// and decide to inline B into A. Doing this makes an alloca available for
// reuse and makes a callsite (C) available for inlining. When we process
// the C call site we don't want to do any alloca merging between X and Y
// because their scopes are not disjoint. We could make this smarter by
// keeping track of the inline history for each alloca in the
// InlinedArrayAllocas but this isn't likely to be a significant win.
if (InlineHistory != -1) // Only do merging for top-level call sites in SCC.
return;
// Loop over all the allocas we have so far and see if they can be merged with
// a previously inlined alloca. If not, remember that we had it.
for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size(); AllocaNo != e;
++AllocaNo) {
AllocaInst *AI = IFI.StaticAllocas[AllocaNo];
// Don't bother trying to merge array allocations (they will usually be
// canonicalized to be an allocation *of* an array), or allocations whose
// type is not itself an array (because we're afraid of pessimizing SRoA).
ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
if (!ATy || AI->isArrayAllocation())
continue;
// Get the list of all available allocas for this array type.
std::vector<AllocaInst *> &AllocasForType = InlinedArrayAllocas[ATy];
// Loop over the allocas in AllocasForType to see if we can reuse one. Note
// that we have to be careful not to reuse the same "available" alloca for
// multiple different allocas that we just inlined, we use the 'UsedAllocas'
// set to keep track of which "available" allocas are being used by this
// function. Also, AllocasForType can be empty of course!
bool MergedAwayAlloca = false;
for (AllocaInst *AvailableAlloca : AllocasForType) {
unsigned Align1 = AI->getAlignment(),
Align2 = AvailableAlloca->getAlignment();
// The available alloca has to be in the right function, not in some other
// function in this SCC.
if (AvailableAlloca->getParent() != AI->getParent())
continue;
// If the inlined function already uses this alloca then we can't reuse
// it.
if (!UsedAllocas.insert(AvailableAlloca).second)
continue;
// Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
// success!
LLVM_DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI
<< "\n\t\tINTO: " << *AvailableAlloca << '\n');
// Move affected dbg.declare calls immediately after the new alloca to
// avoid the situation when a dbg.declare precedes its alloca.
if (auto *L = LocalAsMetadata::getIfExists(AI))
if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L))
for (User *U : MDV->users())
if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
DDI->moveBefore(AvailableAlloca->getNextNode());
AI->replaceAllUsesWith(AvailableAlloca);
if (Align1 != Align2) {
if (!Align1 || !Align2) {
const DataLayout &DL = Caller->getParent()->getDataLayout();
unsigned TypeAlign = DL.getABITypeAlignment(AI->getAllocatedType());
Align1 = Align1 ? Align1 : TypeAlign;
Align2 = Align2 ? Align2 : TypeAlign;
}
if (Align1 > Align2)
AvailableAlloca->setAlignment(AI->getAlignment());
}
AI->eraseFromParent();
MergedAwayAlloca = true;
++NumMergedAllocas;
IFI.StaticAllocas[AllocaNo] = nullptr;
break;
}
// If we already nuked the alloca, we're done with it.
if (MergedAwayAlloca)
continue;
// If we were unable to merge away the alloca either because there are no
// allocas of the right type available or because we reused them all
// already, remember that this alloca came from an inlined function and mark
// it used so we don't reuse it for other allocas from this inline
// operation.
AllocasForType.push_back(AI);
UsedAllocas.insert(AI);
}
}
/// InlineFunction - This function inlines the called function into the basic
/// block of the caller. This returns false if it is not possible to inline
/// this call. The program is still in a well defined state if this occurs
/// though.
///
/// Note that this only does one level of inlining. For example, if the
/// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
/// exists in the instruction stream. Similarly this will inline a recursive
/// function by one level.
bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
bool InsertLifetime) {
Instruction *TheCall = CS.getInstruction();
assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
"Instruction not in function!");
// If IFI has any state in it, zap it before we fill it in.
IFI.reset();
const Function *CalledFunc = CS.getCalledFunction();
if (CalledFunc == 0 || // Can't inline external function or indirect
CalledFunc->isDeclaration() || // call, or call to a vararg function!
CalledFunc->getFunctionType()->isVarArg()) return false;
// If the call to the callee is not a tail call, we must clear the 'tail'
// flags on any calls that we inline.
bool MustClearTailCallFlags =
!(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall());
// If the call to the callee cannot throw, set the 'nounwind' flag on any
// calls that we inline.
bool MarkNoUnwind = CS.doesNotThrow();
BasicBlock *OrigBB = TheCall->getParent();
Function *Caller = OrigBB->getParent();
// GC poses two hazards to inlining, which only occur when the callee has GC:
// 1. If the caller has no GC, then the callee's GC must be propagated to the
// caller.
// 2. If the caller has a differing GC, it is invalid to inline.
if (CalledFunc->hasGC()) {
if (!Caller->hasGC())
Caller->setGC(CalledFunc->getGC());
else if (CalledFunc->getGC() != Caller->getGC())
return false;
}
// Get the personality function from the callee if it contains a landing pad.
Value *CalleePersonality = 0;
for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end();
I != E; ++I)
if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
const BasicBlock *BB = II->getUnwindDest();
const LandingPadInst *LP = BB->getLandingPadInst();
CalleePersonality = LP->getPersonalityFn();
break;
}
// Find the personality function used by the landing pads of the caller. If it
// exists, then check to see that it matches the personality function used in
// the callee.
if (CalleePersonality) {
for (Function::const_iterator I = Caller->begin(), E = Caller->end();
I != E; ++I)
if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) {
const BasicBlock *BB = II->getUnwindDest();
const LandingPadInst *LP = BB->getLandingPadInst();
// If the personality functions match, then we can perform the
// inlining. Otherwise, we can't inline.
// TODO: This isn't 100% true. Some personality functions are proper
// supersets of others and can be used in place of the other.
if (LP->getPersonalityFn() != CalleePersonality)
return false;
break;
}
}
// Get an iterator to the last basic block in the function, which will have
// the new function inlined after it.
Function::iterator LastBlock = &Caller->back();
// Make sure to capture all of the return instructions from the cloned
// function.
SmallVector<ReturnInst*, 8> Returns;
ClonedCodeInfo InlinedFunctionInfo;
Function::iterator FirstNewBlock;
{ // Scope to destroy VMap after cloning.
ValueToValueMapTy VMap;
assert(CalledFunc->arg_size() == CS.arg_size() &&
"No varargs calls can be inlined!");
// Calculate the vector of arguments to pass into the function cloner, which
// matches up the formal to the actual argument values.
CallSite::arg_iterator AI = CS.arg_begin();
unsigned ArgNo = 0;
for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
Value *ActualArg = *AI;
const Argument *Arg = I;
// When byval arguments actually inlined, we need to make the copy implied
// by them explicit. However, we don't do this if the callee is readonly
// or readnone, because the copy would be unneeded: the callee doesn't
// modify the struct.
if (CS.isByValArgument(ArgNo)) {
ActualArg = HandleByValArgument(ActualArg, Arg, TheCall, CalledFunc, IFI,
CalledFunc->getParamAlignment(ArgNo+1));
// Calls that we inline may use the new alloca, so we need to clear
// their 'tail' flags if HandleByValArgument introduced a new alloca and
// the callee has calls.
MustClearTailCallFlags |= ActualArg != *AI;
}
VMap[I] = ActualArg;
}
// We want the inliner to prune the code as it copies. We would LOVE to
// have no dead or constant instructions leftover after inlining occurs
// (which can happen, e.g., because an argument was constant), but we'll be
// happy with whatever the cloner can do.
CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
/*ModuleLevelChanges=*/false, Returns, ".i",
&InlinedFunctionInfo, IFI.TD, TheCall);
// Remember the first block that is newly cloned over.
FirstNewBlock = LastBlock; ++FirstNewBlock;
// Update the callgraph if requested.
if (IFI.CG)
UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
// Update inlined instructions' line number information.
fixupLineNumbers(Caller, FirstNewBlock, TheCall);
}
// If there are any alloca instructions in the block that used to be the entry
// block for the callee, move them to the entry block of the caller. First
// calculate which instruction they should be inserted before. We insert the
// instructions at the end of the current alloca list.
{
BasicBlock::iterator InsertPoint = Caller->begin()->begin();
for (BasicBlock::iterator I = FirstNewBlock->begin(),
E = FirstNewBlock->end(); I != E; ) {
AllocaInst *AI = dyn_cast<AllocaInst>(I++);
if (AI == 0) continue;
// If the alloca is now dead, remove it. This often occurs due to code
// specialization.
if (AI->use_empty()) {
AI->eraseFromParent();
continue;
}
if (!isa<Constant>(AI->getArraySize()))
continue;
// Keep track of the static allocas that we inline into the caller.
IFI.StaticAllocas.push_back(AI);
// Scan for the block of allocas that we can move over, and move them
// all at once.
while (isa<AllocaInst>(I) &&
isa<Constant>(cast<AllocaInst>(I)->getArraySize())) {
IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
++I;
}
// Transfer all of the allocas over in a block. Using splice means
// that the instructions aren't removed from the symbol table, then
// reinserted.
Caller->getEntryBlock().getInstList().splice(InsertPoint,
FirstNewBlock->getInstList(),
AI, I);
}
}
// Leave lifetime markers for the static alloca's, scoping them to the
// function we just inlined.
if (InsertLifetime && !IFI.StaticAllocas.empty()) {
IRBuilder<> builder(FirstNewBlock->begin());
for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
AllocaInst *AI = IFI.StaticAllocas[ai];
// If the alloca is already scoped to something smaller than the whole
// function then there's no need to add redundant, less accurate markers.
if (hasLifetimeMarkers(AI))
continue;
// Try to determine the size of the allocation.
ConstantInt *AllocaSize = 0;
if (ConstantInt *AIArraySize =
dyn_cast<ConstantInt>(AI->getArraySize())) {
if (IFI.TD) {
Type *AllocaType = AI->getAllocatedType();
uint64_t AllocaTypeSize = IFI.TD->getTypeAllocSize(AllocaType);
uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
assert(AllocaArraySize > 0 && "array size of AllocaInst is zero");
// Check that array size doesn't saturate uint64_t and doesn't
// overflow when it's multiplied by type size.
if (AllocaArraySize != ~0ULL &&
UINT64_MAX / AllocaArraySize >= AllocaTypeSize) {
AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
AllocaArraySize * AllocaTypeSize);
}
}
}
builder.CreateLifetimeStart(AI, AllocaSize);
for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) {
IRBuilder<> builder(Returns[ri]);
builder.CreateLifetimeEnd(AI, AllocaSize);
}
}
}
// If the inlined code contained dynamic alloca instructions, wrap the inlined
// code with llvm.stacksave/llvm.stackrestore intrinsics.
if (InlinedFunctionInfo.ContainsDynamicAllocas) {
Module *M = Caller->getParent();
// Get the two intrinsics we care about.
Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
// Insert the llvm.stacksave.
CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin())
.CreateCall(StackSave, "savedstack");
// Insert a call to llvm.stackrestore before any return instructions in the
// inlined function.
for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr);
}
}
// If we are inlining tail call instruction through a call site that isn't
// marked 'tail', we must remove the tail marker for any calls in the inlined
// code. Also, calls inlined through a 'nounwind' call site should be marked
// 'nounwind'.
if (InlinedFunctionInfo.ContainsCalls &&
(MustClearTailCallFlags || MarkNoUnwind)) {
for (Function::iterator BB = FirstNewBlock, E = Caller->end();
BB != E; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
if (CallInst *CI = dyn_cast<CallInst>(I)) {
if (MustClearTailCallFlags)
CI->setTailCall(false);
if (MarkNoUnwind)
CI->setDoesNotThrow();
}
}
// If we are inlining for an invoke instruction, we must make sure to rewrite
// any call instructions into invoke instructions.
if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
// If we cloned in _exactly one_ basic block, and if that block ends in a
// return instruction, we splice the body of the inlined callee directly into
// the calling basic block.
if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
// Move all of the instructions right before the call.
OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
FirstNewBlock->begin(), FirstNewBlock->end());
// Remove the cloned basic block.
Caller->getBasicBlockList().pop_back();
// If the call site was an invoke instruction, add a branch to the normal
// destination.
if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
NewBr->setDebugLoc(Returns[0]->getDebugLoc());
}
// If the return instruction returned a value, replace uses of the call with
// uses of the returned value.
if (!TheCall->use_empty()) {
ReturnInst *R = Returns[0];
if (TheCall == R->getReturnValue())
TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
else
TheCall->replaceAllUsesWith(R->getReturnValue());
}
// Since we are now done with the Call/Invoke, we can delete it.
TheCall->eraseFromParent();
// Since we are now done with the return instruction, delete it also.
Returns[0]->eraseFromParent();
// We are now done with the inlining.
return true;
}
// Otherwise, we have the normal case, of more than one block to inline or
// multiple return sites.
// We want to clone the entire callee function into the hole between the
// "starter" and "ender" blocks. How we accomplish this depends on whether
// this is an invoke instruction or a call instruction.
BasicBlock *AfterCallBB;
BranchInst *CreatedBranchToNormalDest = NULL;
if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
// Add an unconditional branch to make this look like the CallInst case...
CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
// Split the basic block. This guarantees that no PHI nodes will have to be
// updated due to new incoming edges, and make the invoke case more
// symmetric to the call case.
AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest,
CalledFunc->getName()+".exit");
} else { // It's a call
// If this is a call instruction, we need to split the basic block that
// the call lives in.
//
AfterCallBB = OrigBB->splitBasicBlock(TheCall,
CalledFunc->getName()+".exit");
}
// Change the branch that used to go to AfterCallBB to branch to the first
// basic block of the inlined function.
//
TerminatorInst *Br = OrigBB->getTerminator();
assert(Br && Br->getOpcode() == Instruction::Br &&
"splitBasicBlock broken!");
Br->setOperand(0, FirstNewBlock);
// Now that the function is correct, make it a little bit nicer. In
// particular, move the basic blocks inserted from the end of the function
// into the space made by splitting the source basic block.
Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
FirstNewBlock, Caller->end());
// Handle all of the return instructions that we just cloned in, and eliminate
// any users of the original call/invoke instruction.
Type *RTy = CalledFunc->getReturnType();
PHINode *PHI = 0;
if (Returns.size() > 1) {
// The PHI node should go at the front of the new basic block to merge all
// possible incoming values.
if (!TheCall->use_empty()) {
PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
AfterCallBB->begin());
// Anything that used the result of the function call should now use the
// PHI node as their operand.
TheCall->replaceAllUsesWith(PHI);
}
// Loop over all of the return instructions adding entries to the PHI node
// as appropriate.
if (PHI) {
for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
ReturnInst *RI = Returns[i];
assert(RI->getReturnValue()->getType() == PHI->getType() &&
"Ret value not consistent in function!");
PHI->addIncoming(RI->getReturnValue(), RI->getParent());
}
}
// Add a branch to the merge points and remove return instructions.
DebugLoc Loc;
for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
ReturnInst *RI = Returns[i];
BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
Loc = RI->getDebugLoc();
BI->setDebugLoc(Loc);
RI->eraseFromParent();
}
// We need to set the debug location to *somewhere* inside the
// inlined function. The line number may be nonsensical, but the
// instruction will at least be associated with the right
// function.
if (CreatedBranchToNormalDest)
CreatedBranchToNormalDest->setDebugLoc(Loc);
} else if (!Returns.empty()) {
// Otherwise, if there is exactly one return value, just replace anything
// using the return value of the call with the computed value.
if (!TheCall->use_empty()) {
if (TheCall == Returns[0]->getReturnValue())
TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
else
TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
}
// Update PHI nodes that use the ReturnBB to use the AfterCallBB.
BasicBlock *ReturnBB = Returns[0]->getParent();
ReturnBB->replaceAllUsesWith(AfterCallBB);
// Splice the code from the return block into the block that it will return
// to, which contains the code that was after the call.
AfterCallBB->getInstList().splice(AfterCallBB->begin(),
ReturnBB->getInstList());
if (CreatedBranchToNormalDest)
CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
// Delete the return instruction now and empty ReturnBB now.
Returns[0]->eraseFromParent();
ReturnBB->eraseFromParent();
} else if (!TheCall->use_empty()) {
// No returns, but something is using the return value of the call. Just
// nuke the result.
TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
}
// Since we are now done with the Call/Invoke, we can delete it.
TheCall->eraseFromParent();
// We should always be able to fold the entry block of the function into the
// single predecessor of the block...
assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
// Splice the code entry block into calling block, right before the
// unconditional branch.
CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
// Remove the unconditional branch.
OrigBB->getInstList().erase(Br);
// Now we can remove the CalleeEntry block, which is now empty.
Caller->getBasicBlockList().erase(CalleeEntry);
// If we inserted a phi node, check to see if it has a single value (e.g. all
// the entries are the same or undef). If so, remove the PHI so it doesn't
// block other optimizations.
if (PHI) {
if (Value *V = SimplifyInstruction(PHI, IFI.TD)) {
PHI->replaceAllUsesWith(V);
PHI->eraseFromParent();
}
}
return true;
}
/// performCallSlotOptzn - takes a memcpy and a call that it depends on,
/// and checks for the possibility of a call slot optimization by having
/// the call write its result directly into the destination of the memcpy.
bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
Value *cpyDest, Value *cpySrc,
uint64_t cpyLen, unsigned cpyAlign,
CallInst *C) {
// The general transformation to keep in mind is
//
// call @func(..., src, ...)
// memcpy(dest, src, ...)
//
// ->
//
// memcpy(dest, src, ...)
// call @func(..., dest, ...)
//
// Since moving the memcpy is technically awkward, we additionally check that
// src only holds uninitialized values at the moment of the call, meaning that
// the memcpy can be discarded rather than moved.
// Deliberately get the source and destination with bitcasts stripped away,
// because we'll need to do type comparisons based on the underlying type.
CallSite CS(C);
// Require that src be an alloca. This simplifies the reasoning considerably.
AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
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();
if (cpyLen < srcSize)
return false;
// Check that dest points to memory that is at least as aligned as src.
unsigned srcAlign = srcAlloca->getAlignment();
if (!srcAlign)
srcAlign = TD->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 accessing the first srcSize bytes of dest will not cause a
// trap. Otherwise the transform is invalid since it might cause a trap
// to occur earlier than it otherwise would.
if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
// The destination is an alloca. Check it is larger than srcSize.
ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
if (!destArraySize)
return false;
uint64_t destSize = TD->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;
Type *StructTy = cast<PointerType>(A->getType())->getElementType();
uint64_t destSize = TD->getTypeAllocSize(StructTy);
if (destSize < srcSize)
return false;
} else {
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());
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
return false;
} else if (UI != C && UI != cpy) {
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>();
if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
if (!DT.dominates(cpyDestInst, C))
return false;
// In addition to knowing that the call does not access src in some
// unexpected manner, for example via a global, which we deduce from
// 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>();
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) {
Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest
: CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
cpyDest->getName(), C);
changedArgument = true;
if (CS.getArgument(i)->getType() == Dest->getType())
CS.setArgument(i, Dest);
else
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);
// Remove the memcpy.
MD->removeInstruction(cpy);
++NumMemCpyInstr;
return true;
}
Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
// Ensure that the alloca array size argument has type intptr_t, so that
// any casting is exposed early.
if (DL) {
Type *IntPtrTy = DL->getIntPtrType(AI.getType());
if (AI.getArraySize()->getType() != IntPtrTy) {
Value *V = Builder->CreateIntCast(AI.getArraySize(),
IntPtrTy, false);
AI.setOperand(0, V);
return &AI;
}
}
// Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
if (AI.isArrayAllocation()) { // Check C != 1
if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
Type *NewTy =
ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
AllocaInst *New = Builder->CreateAlloca(NewTy, nullptr, AI.getName());
New->setAlignment(AI.getAlignment());
// Scan to the end of the allocation instructions, to skip over a block of
// allocas if possible...also skip interleaved debug info
//
BasicBlock::iterator It = New;
while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
// Now that I is pointing to the first non-allocation-inst in the block,
// insert our getelementptr instruction...
//
Type *IdxTy = DL
? DL->getIntPtrType(AI.getType())
: Type::getInt64Ty(AI.getContext());
Value *NullIdx = Constant::getNullValue(IdxTy);
Value *Idx[2] = { NullIdx, NullIdx };
Instruction *GEP =
GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub");
InsertNewInstBefore(GEP, *It);
// Now make everything use the getelementptr instead of the original
// allocation.
return ReplaceInstUsesWith(AI, GEP);
} else if (isa<UndefValue>(AI.getArraySize())) {
return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
}
}
if (DL && AI.getAllocatedType()->isSized()) {
// If the alignment is 0 (unspecified), assign it the preferred alignment.
if (AI.getAlignment() == 0)
AI.setAlignment(DL->getPrefTypeAlignment(AI.getAllocatedType()));
// Move all alloca's of zero byte objects to the entry block and merge them
// together. Note that we only do this for alloca's, because malloc should
// allocate and return a unique pointer, even for a zero byte allocation.
if (DL->getTypeAllocSize(AI.getAllocatedType()) == 0) {
// For a zero sized alloca there is no point in doing an array allocation.
// This is helpful if the array size is a complicated expression not used
// elsewhere.
if (AI.isArrayAllocation()) {
AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
return &AI;
}
// Get the first instruction in the entry block.
BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
if (FirstInst != &AI) {
// If the entry block doesn't start with a zero-size alloca then move
// this one to the start of the entry block. There is no problem with
// dominance as the array size was forced to a constant earlier already.
AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
DL->getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
AI.moveBefore(FirstInst);
return &AI;
}
// If the alignment of the entry block alloca is 0 (unspecified),
// assign it the preferred alignment.
if (EntryAI->getAlignment() == 0)
EntryAI->setAlignment(
DL->getPrefTypeAlignment(EntryAI->getAllocatedType()));
// Replace this zero-sized alloca with the one at the start of the entry
// block after ensuring that the address will be aligned enough for both
// types.
unsigned MaxAlign = std::max(EntryAI->getAlignment(),
AI.getAlignment());
EntryAI->setAlignment(MaxAlign);
if (AI.getType() != EntryAI->getType())
return new BitCastInst(EntryAI, AI.getType());
return ReplaceInstUsesWith(AI, EntryAI);
}
}
}
if (AI.getAlignment()) {
// Check to see if this allocation is only modified by a memcpy/memmove from
// a constant global whose alignment is equal to or exceeds that of the
// allocation. If this is the case, we can change all users to use
// the constant global instead. This is commonly produced by the CFE by
// constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
// is only subsequently read.
SmallVector<Instruction *, 4> ToDelete;
if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
unsigned SourceAlign = getOrEnforceKnownAlignment(Copy->getSource(),
AI.getAlignment(), DL);
if (AI.getAlignment() <= SourceAlign) {
DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
EraseInstFromFunction(*ToDelete[i]);
Constant *TheSrc = cast<Constant>(Copy->getSource());
Constant *Cast
= ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, AI.getType());
Instruction *NewI = ReplaceInstUsesWith(AI, Cast);
EraseInstFromFunction(*Copy);
++NumGlobalCopies;
return NewI;
}
}
}
// At last, use the generic allocation site handler to aggressively remove
// unused allocas.
return visitAllocSite(AI);
}
