Skip to content

[DirectX] Clean up extra vectors when lowering to buffer store #116721

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.

Sign up for GitHub

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 2 commits into from
Dec 2, 2024
Merged

[DirectX] Clean up extra vectors when lowering to buffer store #116721

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ec3c4f4
[DirectX] Clean up extra vectors when lowering to buffer store
bogner Nov 15, 2024
3ad9863
Improve comment
bogner Nov 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
46 changes: 35 additions & 11 deletions llvm/lib/Target/DirectX/DXILOpLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -531,23 +531,47 @@ class OpLowerer {
return make_error<StringError>(
"typedBufferStore data must be a vector of 4 elements",
inconvertibleErrorCode());
Value *Data0 =
IRB.CreateExtractElement(Data, ConstantInt::get(Int32Ty, 0));
Value *Data1 =
IRB.CreateExtractElement(Data, ConstantInt::get(Int32Ty, 1));
Value *Data2 =
IRB.CreateExtractElement(Data, ConstantInt::get(Int32Ty, 2));
Value *Data3 =
IRB.CreateExtractElement(Data, ConstantInt::get(Int32Ty, 3));

std::array<Value *, 8> Args{Handle, Index0, Index1, Data0,
Data1, Data2, Data3, Mask};

// Since we're post-scalarizer, we likely have a vector that's constructed
// solely for the argument of the store. If so, just use the scalar values
// from before they're inserted into the temporary.
std::array<Value *, 4> DataElements{nullptr, nullptr, nullptr, nullptr};
auto *IEI = dyn_cast<InsertElementInst>(Data);
while (IEI) {
auto *IndexOp = dyn_cast<ConstantInt>(IEI->getOperand(2));
if (!IndexOp)
break;
size_t IndexVal = IndexOp->getZExtValue();
assert(IndexVal < 4 && "Too many elements for buffer store");
Copy link
Contributor

Choose a reason for hiding this comment

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

If we start support > 4 element vectors, are asserts like this going to bite us (ie hard to debug errors that don't show up in production builds?)

Copy link
Contributor Author

@bogner bogner Nov 20, 2024 •
edited
Loading

Choose a reason for hiding this comment

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

I don't think so, this is fairly likely to crash in a non-asserts build if it comes up anyways, and this is a fairly obvious place we'd need to update if we were to make such a change. Note that we'd have to expand into multiple calls to dx.op.BufferStore here if we were to support larger vectors in HLSL's typed buffers, so this isn't just some implementation detail we'd be likely to forget about.

DataElements[IndexVal] = IEI->getOperand(1);
IEI = dyn_cast<InsertElementInst>(IEI->getOperand(0));
}

// If for some reason we weren't able to forward the arguments from the
// scalarizer artifact, then we need to actually extract elements from the
// vector.
for (int I = 0, E = 4; I != E; ++I)
if (DataElements[I] == nullptr)
DataElements[I] =
IRB.CreateExtractElement(Data, ConstantInt::get(Int32Ty, I));

std::array<Value *, 8> Args{
Handle, Index0, Index1, DataElements[0],
DataElements[1], DataElements[2], DataElements[3], Mask};
Expected<CallInst *> OpCall =
OpBuilder.tryCreateOp(OpCode::BufferStore, Args, CI->getName());
if (Error E = OpCall.takeError())
return E;

CI->eraseFromParent();
// Clean up any leftover `insertelement`s
IEI = dyn_cast<InsertElementInst>(Data);
while (IEI && IEI->use_empty()) {
InsertElementInst *Tmp = IEI;
IEI = dyn_cast<InsertElementInst>(IEI->getOperand(0));
Tmp->eraseFromParent();
}

return Error::success();
});
}
Expand Down
24 changes: 24 additions & 0 deletions llvm/test/CodeGen/DirectX/BufferStore.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,3 +90,27 @@ define void @storei16(<4 x i16> %data, i32 %index) {

ret void
}

define void @store_scalarized_floats(float %data0, float %data1, float %data2, float %data3, i32 %index) {

; CHECK: [[BIND:%.*]] = call %dx.types.Handle @dx.op.createHandleFromBinding(i32 217,
; CHECK: [[HANDLE:%.*]] = call %dx.types.Handle @dx.op.annotateHandle(i32 216, %dx.types.Handle [[BIND]]
%buffer = call target("dx.TypedBuffer", <4 x float>, 1, 0, 0)
@llvm.dx.handle.fromBinding.tdx.TypedBuffer_v4f32_1_0_0(
i32 0, i32 0, i32 1, i32 0, i1 false)

; We shouldn't end up with any inserts/extracts.
; CHECK-NOT: insertelement
; CHECK-NOT: extractelement

; CHECK: call void @dx.op.bufferStore.f32(i32 69, %dx.types.Handle [[HANDLE]], i32 %index, i32 undef, float %data0, float %data1, float %data2, float %data3, i8 15)
%vec.upto0 = insertelement <4 x float> poison, float %data0, i64 0
%vec.upto1 = insertelement <4 x float> %vec.upto0, float %data1, i64 1
%vec.upto2 = insertelement <4 x float> %vec.upto1, float %data2, i64 2
%vec = insertelement <4 x float> %vec.upto2, float %data3, i64 3
call void @llvm.dx.typedBufferStore(
target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %buffer,
i32 %index, <4 x float> %vec)

ret void
}
Loading