Skip to content

[DirectX] Implement memcpy in DXIL CBuffer Access pass #144436

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 10 commits into from
Jun 30, 2025
306 changes: 223 additions & 83 deletions llvm/lib/Target/DirectX/DXILCBufferAccess.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -11,9 +11,13 @@
#include "llvm/Frontend/HLSL/CBuffer.h"
#include "llvm/Frontend/HLSL/HLSLResource.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/IntrinsicsDirectX.h"
#include "llvm/IR/Operator.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/FormatVariadic.h"
#include "llvm/Transforms/Utils/Local.h"

#define DEBUG_TYPE "dxil-cbuffer-access"
Expand Down Expand Up @@ -54,114 +58,249 @@ struct CBufferRowIntrin {
}
}
};
} // namespace

static size_t getOffsetForCBufferGEP(GEPOperator *GEP, GlobalVariable *Global,
const DataLayout &DL) {
// Since we should always have a constant offset, we should only ever have a
// single GEP of indirection from the Global.
assert(GEP->getPointerOperand() == Global &&
"Indirect access to resource handle");
// Helper for creating CBuffer handles and loading data from them
struct CBufferResource {
GlobalVariable *GVHandle;
GlobalVariable *Member;
size_t MemberOffset;

APInt ConstantOffset(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
bool Success = GEP->accumulateConstantOffset(DL, ConstantOffset);
(void)Success;
assert(Success && "Offsets into cbuffer globals must be constant");
LoadInst *Handle;

if (auto *ATy = dyn_cast<ArrayType>(Global->getValueType()))
ConstantOffset = hlsl::translateCBufArrayOffset(DL, ConstantOffset, ATy);
CBufferResource(GlobalVariable *GVHandle, GlobalVariable *Member,
size_t MemberOffset)
: GVHandle(GVHandle), Member(Member), MemberOffset(MemberOffset) {}

return ConstantOffset.getZExtValue();
}
const DataLayout &getDataLayout() { return GVHandle->getDataLayout(); }
Type *getValueType() { return Member->getValueType(); }
iterator_range<ConstantDataSequential::user_iterator> users() {
return Member->users();
}

/// Replace access via cbuffer global with a load from the cbuffer handle
/// itself.
static void replaceAccess(LoadInst *LI, GlobalVariable *Global,
GlobalVariable *HandleGV, size_t BaseOffset,
SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
const DataLayout &DL = HandleGV->getDataLayout();
/// Get the byte offset of a Pointer-typed Value * `Val` relative to Member.
/// `Val` can either be Member itself, or a GEP of a constant offset from
/// Member
size_t getOffsetForCBufferGEP(Value *Val) {
assert(isa<PointerType>(Val->getType()) &&
"Expected a pointer-typed value");

if (Val == Member)
return 0;

if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
// Since we should always have a constant offset, we should only ever have
// a single GEP of indirection from the Global.
assert(GEP->getPointerOperand() == Member &&
"Indirect access to resource handle");

const DataLayout &DL = getDataLayout();
APInt ConstantOffset(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
bool Success = GEP->accumulateConstantOffset(DL, ConstantOffset);
(void)Success;
assert(Success && "Offsets into cbuffer globals must be constant");

if (auto *ATy = dyn_cast<ArrayType>(Member->getValueType()))
ConstantOffset =
hlsl::translateCBufArrayOffset(DL, ConstantOffset, ATy);

return ConstantOffset.getZExtValue();
}

size_t Offset = BaseOffset;
if (auto *GEP = dyn_cast<GEPOperator>(LI->getPointerOperand()))
Offset += getOffsetForCBufferGEP(GEP, Global, DL);
else if (LI->getPointerOperand() != Global)
llvm_unreachable("Load instruction doesn't reference cbuffer global");
llvm_unreachable("Expected Val to be a GlobalVariable or GEP");
}

IRBuilder<> Builder(LI);
auto *Handle = Builder.CreateLoad(HandleGV->getValueType(), HandleGV,
HandleGV->getName());

Type *Ty = LI->getType();
CBufferRowIntrin Intrin(DL, Ty->getScalarType());
// The cbuffer consists of some number of 16-byte rows.
unsigned int CurrentRow = Offset / hlsl::CBufferRowSizeInBytes;
unsigned int CurrentIndex =
(Offset % hlsl::CBufferRowSizeInBytes) / Intrin.EltSize;

auto *CBufLoad = Builder.CreateIntrinsic(
Intrin.RetTy, Intrin.IID,
{Handle, ConstantInt::get(Builder.getInt32Ty(), CurrentRow)}, nullptr,
LI->getName());
auto *Elt =
Builder.CreateExtractValue(CBufLoad, {CurrentIndex++}, LI->getName());

Value *Result = nullptr;
unsigned int Remaining =
((DL.getTypeSizeInBits(Ty) / 8) / Intrin.EltSize) - 1;
if (Remaining == 0) {
// We only have a single element, so we're done.
Result = Elt;

// However, if we loaded a <1 x T>, then we need to adjust the type here.
if (auto *VT = dyn_cast<FixedVectorType>(LI->getType())) {
assert(VT->getNumElements() == 1 && "Can't have multiple elements here");
Result = Builder.CreateInsertElement(PoisonValue::get(VT), Result,
Builder.getInt32(0));
}
} else {
// Walk each element and extract it, wrapping to new rows as needed.
SmallVector<Value *> Extracts{Elt};
while (Remaining--) {
CurrentIndex %= Intrin.NumElts;

if (CurrentIndex == 0)
CBufLoad = Builder.CreateIntrinsic(
Intrin.RetTy, Intrin.IID,
{Handle, ConstantInt::get(Builder.getInt32Ty(), ++CurrentRow)},
nullptr, LI->getName());

Extracts.push_back(Builder.CreateExtractValue(CBufLoad, {CurrentIndex++},
LI->getName()));
/// Create a handle for this cbuffer resource using the IRBuilder `Builder`
/// and sets the handle as the current one to use for subsequent calls to
/// `loadValue`
void createAndSetCurrentHandle(IRBuilder<> &Builder) {
Handle = Builder.CreateLoad(GVHandle->getValueType(), GVHandle,
GVHandle->getName());
}

/// Load a value of type `Ty` at offset `Offset` using the handle from the
/// last call to `createAndSetCurrentHandle`
Value *loadValue(IRBuilder<> &Builder, Type *Ty, size_t Offset,
const Twine &Name = "") {
assert(Handle &&
"Expected a handle for this cbuffer global resource to be created "
"before loading a value from it");
const DataLayout &DL = getDataLayout();

size_t TargetOffset = MemberOffset + Offset;
CBufferRowIntrin Intrin(DL, Ty->getScalarType());
// The cbuffer consists of some number of 16-byte rows.
unsigned int CurrentRow = TargetOffset / hlsl::CBufferRowSizeInBytes;
unsigned int CurrentIndex =
(TargetOffset % hlsl::CBufferRowSizeInBytes) / Intrin.EltSize;

auto *CBufLoad = Builder.CreateIntrinsic(
Intrin.RetTy, Intrin.IID,
{Handle, ConstantInt::get(Builder.getInt32Ty(), CurrentRow)}, nullptr,
Name + ".load");
auto *Elt = Builder.CreateExtractValue(CBufLoad, {CurrentIndex++},
Name + ".extract");

Value *Result = nullptr;
unsigned int Remaining =
((DL.getTypeSizeInBits(Ty) / 8) / Intrin.EltSize) - 1;
if (Remaining == 0) {
// We only have a single element, so we're done.
Result = Elt;

// However, if we loaded a <1 x T>, then we need to adjust the type here.
if (auto *VT = dyn_cast<FixedVectorType>(Ty)) {
assert(VT->getNumElements() == 1 &&
"Can't have multiple elements here");
Result = Builder.CreateInsertElement(PoisonValue::get(VT), Result,
Builder.getInt32(0), Name);
}
} else {
// Walk each element and extract it, wrapping to new rows as needed.
SmallVector<Value *> Extracts{Elt};
while (Remaining--) {
CurrentIndex %= Intrin.NumElts;

if (CurrentIndex == 0)
CBufLoad = Builder.CreateIntrinsic(
Intrin.RetTy, Intrin.IID,
{Handle, ConstantInt::get(Builder.getInt32Ty(), ++CurrentRow)},
nullptr, Name + ".load");

Extracts.push_back(Builder.CreateExtractValue(
CBufLoad, {CurrentIndex++}, Name + ".extract"));
}

// Finally, we build up the original loaded value.
Result = PoisonValue::get(Ty);
for (int I = 0, E = Extracts.size(); I < E; ++I)
Result = Builder.CreateInsertElement(Result, Extracts[I],
Builder.getInt32(I),
Name + formatv(".upto{}", I));
}

// Finally, we build up the original loaded value.
Result = PoisonValue::get(Ty);
for (int I = 0, E = Extracts.size(); I < E; ++I)
Result =
Builder.CreateInsertElement(Result, Extracts[I], Builder.getInt32(I));
return Result;
}
};

} // namespace

/// Replace load via cbuffer global with a load from the cbuffer handle itself.
static void replaceLoad(LoadInst *LI, CBufferResource &CBR,
SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
size_t Offset = CBR.getOffsetForCBufferGEP(LI->getPointerOperand());
IRBuilder<> Builder(LI);
CBR.createAndSetCurrentHandle(Builder);
Value *Result = CBR.loadValue(Builder, LI->getType(), Offset, LI->getName());
LI->replaceAllUsesWith(Result);
DeadInsts.push_back(LI);
}

static void replaceAccessesWithHandle(GlobalVariable *Global,
GlobalVariable *HandleGV,
size_t BaseOffset) {
/// Replace memcpy from a cbuffer global with a memcpy from the cbuffer handle
/// itself. Assumes the cbuffer global is an array, and the length of bytes to
/// copy is divisible by array element allocation size.
/// The memcpy source must also be a direct cbuffer global reference, not a GEP.
static void replaceMemCpy(MemCpyInst *MCI, CBufferResource &CBR,
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

would we even need to do this replacement if we did memcpy legalization before this pass?

Copy link
Contributor Author

@Icohedron Icohedron Jun 20, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, we still need this replacement because memcpy from a cbuffer array is very different from a memcpy from a normal array due to differences in the data layout.

SmallVectorImpl<WeakTrackingVH> &DeadInsts) {

ArrayType *ArrTy = dyn_cast<ArrayType>(CBR.getValueType());
assert(ArrTy && "MemCpy lowering is only supported for array types");

// This assumption vastly simplifies the implementation
if (MCI->getSource() != CBR.Member)
reportFatalUsageError(
"Expected MemCpy source to be a cbuffer global variable");

const std::string Name = ("memcpy." + MCI->getDest()->getName() + "." +
MCI->getSource()->getName())
.str();

ConstantInt *Length = dyn_cast<ConstantInt>(MCI->getLength());
uint64_t ByteLength = Length->getZExtValue();

// If length to copy is zero, no memcpy is needed
if (ByteLength == 0) {
DeadInsts.push_back(MCI);
return;
}

const DataLayout &DL = CBR.getDataLayout();

Type *ElemTy = ArrTy->getElementType();
size_t ElemSize = DL.getTypeAllocSize(ElemTy);
assert(ByteLength % ElemSize == 0 &&
"Length of bytes to MemCpy must be divisible by allocation size of "
"source/destination array elements");
size_t ElemsToCpy = ByteLength / ElemSize;

IRBuilder<> Builder(MCI);
CBR.createAndSetCurrentHandle(Builder);

// This function recursively copies N array elements from the CBuffer Resource
// to the MemCpy Destination. Recursion is used to unravel multidimensional
// arrays into a sequence of scalar/vector extracts and stores.
auto CopyElemsImpl = [&Builder, &MCI, &Name, &CBR,
&DL](const auto &Self, ArrayType *ArrTy,
size_t ArrOffset, size_t N) -> void {
Type *ElemTy = ArrTy->getElementType();
size_t ElemTySize = DL.getTypeAllocSize(ElemTy);
for (unsigned I = 0; I < N; ++I) {
size_t Offset = ArrOffset + I * ElemTySize;

// Recursively copy nested arrays
if (ArrayType *ElemArrTy = dyn_cast<ArrayType>(ElemTy)) {
Self(Self, ElemArrTy, Offset, ElemArrTy->getNumElements());
continue;
}

// Load CBuffer value and store it in Dest
APInt CBufArrayOffset(
DL.getIndexTypeSizeInBits(MCI->getSource()->getType()), Offset);
CBufArrayOffset =
hlsl::translateCBufArrayOffset(DL, CBufArrayOffset, ArrTy);
Value *CBufferVal =
CBR.loadValue(Builder, ElemTy, CBufArrayOffset.getZExtValue(), Name);
Value *GEP =
Builder.CreateInBoundsGEP(Builder.getInt8Ty(), MCI->getDest(),
{Builder.getInt32(Offset)}, Name + ".dest");
Builder.CreateStore(CBufferVal, GEP, MCI->isVolatile());
}
};
auto CopyElems = [&CopyElemsImpl](ArrayType *ArrTy, size_t N) -> void {
CopyElemsImpl(CopyElemsImpl, ArrTy, 0, N);
};

CopyElems(ArrTy, ElemsToCpy);

MCI->eraseFromParent();
}

static void replaceAccessesWithHandle(CBufferResource &CBR) {
SmallVector<WeakTrackingVH> DeadInsts;

SmallVector<User *> ToProcess{Global->users()};
SmallVector<User *> ToProcess{CBR.users()};
while (!ToProcess.empty()) {
User *Cur = ToProcess.pop_back_val();

// If we have a load instruction, replace the access.
if (auto *LI = dyn_cast<LoadInst>(Cur)) {
replaceAccess(LI, Global, HandleGV, BaseOffset, DeadInsts);
replaceLoad(LI, CBR, DeadInsts);
continue;
}

// If we have a memcpy instruction, replace it with multiple accesses and
// subsequent stores to the destination
if (auto *MCI = dyn_cast<MemCpyInst>(Cur)) {
replaceMemCpy(MCI, CBR, DeadInsts);
continue;
}

// Otherwise, walk users looking for a load...
ToProcess.append(Cur->user_begin(), Cur->user_end());
if (isa<GetElementPtrInst>(Cur) || isa<GEPOperator>(Cur)) {
ToProcess.append(Cur->user_begin(), Cur->user_end());
continue;
}

reportFatalInternalError("Unexpected user of Global");
}
RecursivelyDeleteTriviallyDeadInstructions(DeadInsts);
}
Expand All @@ -173,7 +312,8 @@ static bool replaceCBufferAccesses(Module &M) {

for (const hlsl::CBufferMapping &Mapping : *CBufMD)
for (const hlsl::CBufferMember &Member : Mapping.Members) {
replaceAccessesWithHandle(Member.GV, Mapping.Handle, Member.Offset);
CBufferResource CBR(Mapping.Handle, Member.GV, Member.Offset);
replaceAccessesWithHandle(CBR);
Member.GV->removeFromParent();
}

Expand Down
Loading
Loading