[HLSL] Attribute vectors have different cast behavior

[V-FEXrt]

https://godbolt.org/z/j1WPxaGKG

uint32_t4 can be implicitly casted to uint64_t4 but __attribute__((__vector_size__(4 * sizeof(unsigned int)))) unsigned int raises an error instead.

This causes issues with intrinsics such as firstbitlow which use the attribute type

DXC is perfectly fine with the implict cast for firstbitlow, while clang raises a compiler error

[V-FEXrt]

I need to poke at this further and pin down more details

[V-FEXrt]

Alright so the core issue here is that the firstbit* intrinsics were using Vector instead of ExtVector. This kind of error is generally expected in that case. The reason it wasn't detected during lowering or testing is that both vectors lower to the same IR.

Thanks @llvm-beanz for the context!

Fix will be up in a second

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Ashley Coleman (V-FEXrt)

https://godbolt.org/z/j1WPxaGKG

uint32_t4 can be implicitly casted to uint64_t4 but __attribute__((__vector_size__(4 * sizeof(unsigned int)))) unsigned int raises an error instead.

This causes issues with intrinsics such as firstbitlow which use the attribute type

DXC is perfectly fine with the implict cast for firstbitlow, while clang raises a compiler error