[llvm] [SPIRV] Legalize Cast Instructions for Matrix Types (PR #212619)

[Farzon Lotfi via llvm-commits]

farzonl wrote:

> LGTM. Is the issue with an i1 making it into a transpose worth a TODO comment?

No not really. Essentially my fix is to legalize so the transpose intrinsic never get an i1 vector. I don't think there is another way to do this via spirv shaders and therefore via HLSL. There could be an issue if say opencl which doesn't represent bools as i32s but as i1s were to start using the llvm matrix intrinsics and further that they support some type of long vector larger than 16 elements. None of these cases are currently possible and are further out of scope for us so it seems better for a future implementer to revisit this.

https://github.com/llvm/llvm-project/pull/212619