void Closure::unpack_struct(Scope<Value *> &dst,
Value *src,
IRBuilder<> *builder) {
// src should be a pointer to a struct of the type returned by build_type
int idx = 0;
LLVMContext &context = builder->getContext();
vector<string> nm = names();
for (size_t i = 0; i < nm.size(); i++) {
Value *ptr = builder->CreateConstInBoundsGEP2_32(src, 0, idx++);
LoadInst *load = builder->CreateLoad(ptr);
if (load->getType()->isPointerTy()) {
// Give it a unique type so that tbaa tells llvm that this can't alias anything
load->setMetadata("tbaa", MDNode::get(context,
vec<Value *>(MDString::get(context, nm[i]))));
}
dst.push(nm[i], load);
load->setName(nm[i]);
}
}
/// check if value in memory at `memAddr` was changed when accessing `val`, store result into flag
void RedoBBBuilder::insertCheck(Value *val, Value *memAddr, Value* flag)
{
for (auto it = val->use_begin(), ite = val->use_end(); it != ite; it++) {
if (StoreInst *SI = dyn_cast<StoreInst>(*it)) {
// skip those not in current top loop
if (!isCurrentTopLoop(*SI)) { continue; }
// skip instruction we don't interested in
if (!shouldCheck(*SI)) { continue; }
// if we have checked this store for this memAddr
CmpInst *chkRes = 0;
if (StoreToCheckMap.count(SI)) {
for (auto pair : StoreToCheckMap[SI]) {
if (pair.first == memAddr) {
chkRes = pair.second;
DEBUG(dbgs() << " Found existing check '" << chkRes->getName()
<< "' for (" << *SI << " ) in "
<< SI->getParent()->getName() << "\n");
}
}
}
if (!chkRes) {
// check if before and after the store, the memore address changed
//
// %orig = load %memAddr
// the STORE we checking
// %modified = load %memAddr
// %cmp = icmp ne, %orig, %modified
LoadInst *orig = new LoadInst(memAddr, "", SI);
Instruction *next = SI->getNextNode();
LoadInst *modified = new LoadInst(memAddr, "", next);
chkRes = CmpInst::Create(Instruction::ICmp,
CmpInst::ICMP_NE,
orig, modified,
"chk", next);
CheckingInstrs.insert(orig);
CheckingInstrs.insert(modified);
CheckingInstrs.insert(chkRes);
// set consitant name
orig->setName(chkRes->getName() + ".orig");
modified->setName(chkRes->getName() + ".mod");
StoreToCheckMap[SI].push_back({memAddr, chkRes});
DEBUG(dbgs() << " Inserted check '" << chkRes->getName()
<< "' for (" << *SI << " ) in "
<< SI->getParent()->getName() << "\n");
}
// if we have stored the check result to the flag
Check pair = {memAddr, chkRes};
for (auto f : CheckToFlagMap[pair]) {
if (f == flag) {
DEBUG(dbgs() << " Existing flag store found\n");
return;
}
}
DEBUG(dbgs() << " Check result stored to " << flag->getName() << "\n");
// merge old value and new value with or
//
// %oldflgval = load %flag
// %newflgval = or %oldflgval, %chkRes
// store %newflgval, %flag
// the next instr after STORE we checking
Instruction *next = chkRes->getNextNode();
LoadInst *oldflgval = new LoadInst(flag, flag->getName() + ".oldval", next);
auto *newflgval = BinaryOperator::Create(Instruction::Or,
oldflgval, chkRes,
flag->getName() + ".newval",
next);
StoreInst *st = new StoreInst(newflgval, flag, next);
CheckingInstrs.insert(oldflgval);
CheckingInstrs.insert(newflgval);
CheckingInstrs.insert(st);
CheckToFlagMap[pair].push_back(flag);
}
}
}
bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
CallingConv::ID CC = F.getCallingConv();
if (CC != CallingConv::AMDGPU_KERNEL || F.arg_empty())
return false;
auto &TPC = getAnalysis<TargetPassConfig>();
const TargetMachine &TM = TPC.getTM<TargetMachine>();
const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
LLVMContext &Ctx = F.getParent()->getContext();
const DataLayout &DL = F.getParent()->getDataLayout();
BasicBlock &EntryBlock = *F.begin();
IRBuilder<> Builder(&*EntryBlock.begin());
const unsigned KernArgBaseAlign = 16; // FIXME: Increase if necessary
const uint64_t BaseOffset = ST.getExplicitKernelArgOffset(F);
unsigned MaxAlign;
// FIXME: Alignment is broken broken with explicit arg offset.;
const uint64_t TotalKernArgSize = ST.getKernArgSegmentSize(F, MaxAlign);
if (TotalKernArgSize == 0)
return false;
CallInst *KernArgSegment =
Builder.CreateIntrinsic(Intrinsic::amdgcn_kernarg_segment_ptr, {}, {},
nullptr, F.getName() + ".kernarg.segment");
KernArgSegment->addAttribute(AttributeList::ReturnIndex, Attribute::NonNull);
KernArgSegment->addAttribute(AttributeList::ReturnIndex,
Attribute::getWithDereferenceableBytes(Ctx, TotalKernArgSize));
unsigned AS = KernArgSegment->getType()->getPointerAddressSpace();
uint64_t ExplicitArgOffset = 0;
for (Argument &Arg : F.args()) {
Type *ArgTy = Arg.getType();
unsigned Align = DL.getABITypeAlignment(ArgTy);
unsigned Size = DL.getTypeSizeInBits(ArgTy);
unsigned AllocSize = DL.getTypeAllocSize(ArgTy);
uint64_t EltOffset = alignTo(ExplicitArgOffset, Align) + BaseOffset;
ExplicitArgOffset = alignTo(ExplicitArgOffset, Align) + AllocSize;
if (Arg.use_empty())
continue;
if (PointerType *PT = dyn_cast<PointerType>(ArgTy)) {
// FIXME: Hack. We rely on AssertZext to be able to fold DS addressing
// modes on SI to know the high bits are 0 so pointer adds don't wrap. We
// can't represent this with range metadata because it's only allowed for
// integer types.
if ((PT->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
PT->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) &&
ST.getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS)
continue;
// FIXME: We can replace this with equivalent alias.scope/noalias
// metadata, but this appears to be a lot of work.
if (Arg.hasNoAliasAttr())
continue;
}
VectorType *VT = dyn_cast<VectorType>(ArgTy);
bool IsV3 = VT && VT->getNumElements() == 3;
bool DoShiftOpt = Size < 32 && !ArgTy->isAggregateType();
VectorType *V4Ty = nullptr;
int64_t AlignDownOffset = alignDown(EltOffset, 4);
int64_t OffsetDiff = EltOffset - AlignDownOffset;
unsigned AdjustedAlign = MinAlign(DoShiftOpt ? AlignDownOffset : EltOffset,
KernArgBaseAlign);
Value *ArgPtr;
Type *AdjustedArgTy;
if (DoShiftOpt) { // FIXME: Handle aggregate types
// Since we don't have sub-dword scalar loads, avoid doing an extload by
// loading earlier than the argument address, and extracting the relevant
// bits.
//
// Additionally widen any sub-dword load to i32 even if suitably aligned,
// so that CSE between different argument loads works easily.
ArgPtr = Builder.CreateConstInBoundsGEP1_64(
Builder.getInt8Ty(), KernArgSegment, AlignDownOffset,
Arg.getName() + ".kernarg.offset.align.down");
AdjustedArgTy = Builder.getInt32Ty();
} else {
ArgPtr = Builder.CreateConstInBoundsGEP1_64(
Builder.getInt8Ty(), KernArgSegment, EltOffset,
Arg.getName() + ".kernarg.offset");
AdjustedArgTy = ArgTy;
}
if (IsV3 && Size >= 32) {
V4Ty = VectorType::get(VT->getVectorElementType(), 4);
// Use the hack that clang uses to avoid SelectionDAG ruining v3 loads
AdjustedArgTy = V4Ty;
}
ArgPtr = Builder.CreateBitCast(ArgPtr, AdjustedArgTy->getPointerTo(AS),
ArgPtr->getName() + ".cast");
LoadInst *Load =
Builder.CreateAlignedLoad(AdjustedArgTy, ArgPtr, AdjustedAlign);
Load->setMetadata(LLVMContext::MD_invariant_load, MDNode::get(Ctx, {}));
MDBuilder MDB(Ctx);
if (isa<PointerType>(ArgTy)) {
if (Arg.hasNonNullAttr())
Load->setMetadata(LLVMContext::MD_nonnull, MDNode::get(Ctx, {}));
uint64_t DerefBytes = Arg.getDereferenceableBytes();
if (DerefBytes != 0) {
Load->setMetadata(
LLVMContext::MD_dereferenceable,
MDNode::get(Ctx,
MDB.createConstant(
ConstantInt::get(Builder.getInt64Ty(), DerefBytes))));
}
uint64_t DerefOrNullBytes = Arg.getDereferenceableOrNullBytes();
if (DerefOrNullBytes != 0) {
Load->setMetadata(
LLVMContext::MD_dereferenceable_or_null,
MDNode::get(Ctx,
MDB.createConstant(ConstantInt::get(Builder.getInt64Ty(),
DerefOrNullBytes))));
}
unsigned ParamAlign = Arg.getParamAlignment();
if (ParamAlign != 0) {
Load->setMetadata(
LLVMContext::MD_align,
MDNode::get(Ctx,
MDB.createConstant(ConstantInt::get(Builder.getInt64Ty(),
ParamAlign))));
}
}
// TODO: Convert noalias arg to !noalias
if (DoShiftOpt) {
Value *ExtractBits = OffsetDiff == 0 ?
Load : Builder.CreateLShr(Load, OffsetDiff * 8);
IntegerType *ArgIntTy = Builder.getIntNTy(Size);
Value *Trunc = Builder.CreateTrunc(ExtractBits, ArgIntTy);
Value *NewVal = Builder.CreateBitCast(Trunc, ArgTy,
Arg.getName() + ".load");
Arg.replaceAllUsesWith(NewVal);
} else if (IsV3) {
Value *Shuf = Builder.CreateShuffleVector(Load, UndefValue::get(V4Ty),
{0, 1, 2},
Arg.getName() + ".load");
Arg.replaceAllUsesWith(Shuf);
} else {
Load->setName(Arg.getName() + ".load");
Arg.replaceAllUsesWith(Load);
}
}
KernArgSegment->addAttribute(
AttributeList::ReturnIndex,
Attribute::getWithAlignment(Ctx, std::max(KernArgBaseAlign, MaxAlign)));
return true;
}
void Closure::unpack_struct(Scope<Value *> &dst,
llvm::Type *
#if LLVM_VERSION >= 37
type
#endif
,
Value *src,
IRBuilder<> *builder) {
// type, type of src should be a pointer to a struct of the type returned by build_type
int idx = 0;
LLVMContext &context = builder->getContext();
vector<string> nm = names();
for (size_t i = 0; i < nm.size(); i++) {
#if LLVM_VERSION >= 37
Value *ptr = builder->CreateConstInBoundsGEP2_32(type, src, 0, idx++);
#else
Value *ptr = builder->CreateConstInBoundsGEP2_32(src, 0, idx++);
#endif
LoadInst *load = builder->CreateLoad(ptr);
if (load->getType()->isPointerTy()) {
// Give it a unique type so that tbaa tells llvm that this can't alias anything
LLVMMDNodeArgumentType md_args[] = {MDString::get(context, nm[i])};
load->setMetadata("tbaa", MDNode::get(context, md_args));
}
dst.push(nm[i], load);
load->setName(nm[i]);
}
}
