| Issue |
169430
|
| Summary |
[DirectX] DXILResourceAccess does not handle writes to vector elements correctly
|
| Labels |
backend:DirectX
|
| Assignees |
|
| Reporter |
hekota
|
DXILResourceAccess does not seem to calculate vector element offset correctly when translating `llvm.dx.resource.getpointer` & `getelementptr` & `store` for a single vector element store.
HLSL:
```
RWStructuredBuffer<uint4> Out : register(u1);
[numthreads(4, 4, 1)]
void main(uint3 ThreadID : SV_GroupThreadID) {
Out[0][ThreadID.y] = 111;
}
```
IR before DXILResourceAccess pass:
```
; RUN: opt -S -dxil-resource-access -mtriple dxilv1.5-unknown-shadermodel6.5-compute %s
target triple = "dxilv1.5-unknown-shadermodel6.5-compute"
define void @main() local_unnamed_addr {
entry:
%0 = tail call target("dx.RawBuffer", <4 x i32>, 1, 0) @llvm.dx.resource.handlefrombinding.tdx.RawBuffer_v4i32_1_0t(i32 0, i32 1, i32 1, i32 0, ptr null)
%1 = tail call i32 @llvm.dx.thread.id.in.group(i32 1)
%2 = tail call noundef nonnull align 16 dereferenceable(16) ptr @llvm.dx.resource.getpointer.p0.tdx.RawBuffer_v4i32_1_0t(target("dx.RawBuffer", <4 x i32>, 1, 0) %0, i32 0)
%3 = getelementptr i32, ptr %2, i32 %1
store i32 111, ptr %3, align 4
ret void
}
```
IR after DXILResourceAccess pass:
```
define void @main() local_unnamed_addr {
entry:
%0 = tail call target("dx.RawBuffer", <4 x i32>, 1, 0) @llvm.dx.resource.handlefrombinding.tdx.RawBuffer_v4i32_1_0t(i32 0, i32 1, i32 1, i32 0, ptr null)
%1 = tail call i32 @llvm.dx.thread.id.in.group(i32 1)
call void @llvm.dx.resource.store.rawbuffer.tdx.RawBuffer_v4i32_1_0t.i32(target("dx.RawBuffer", <4 x i32>, 1, 0) %0, i32 0, i32 %1, i32 111)
ret void
}
```
The 3rd argument of `llvm.dx.resource.store.rawbuffer` is element offset and it should be multiplied by size of `i32`.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
