[Mlir-commits] [mlir] [mlir][gpu] Add pass for emulating unsupported types. (PR #138087)

[Md Abdullah Shahneous Bari]

mshahneo wrote:

> (and removing that code later - or making it conditional based on what extensions exist - is a non-issue in my book. I had to delete this exact sort of bf16-is-i16 thing for AMD LLVM lowering because the backend learned about bfloat - see #108409 for the deletion
> 
> I still claim that this is a better approach for the SPIR-V lowering than this bitcast-heavy rewrite on GPU)

I respectfully disagree with your claim. I believe this is better hands off approach to handle unsupported types. 
And since at this point there is not technical issue with this approach (I believe I addressed those issues), how do you suggest we resolve this?

https://github.com/llvm/llvm-project/pull/138087