[llvm] [AMDGPU][True16][CodeGen] support v_mov_b16 and v_swap_b16 in true16 format (PR #102198)

[Brox Chen via llvm-commits]

broxigarchen wrote:

> Hi there, I bisected to this change in a downstream compiler (https://github.com/iree-org/iree) crash on gfx1100 target. The error is `error: Illegal instruction detected: Operand has incorrect register class.`.
> 
> I'm going to revert it locally in our project. I don't know how to generate a repro for you at this moment, but I'll try to do that. I'm asking questions about generating a repro in [iree-org/iree#18177 (comment)](https://github.com/iree-org/iree/pull/18177#issuecomment-2278892368).

Hi @hanhanW Thanks for the notifiication for this error. Just want to double check with you that you are not actively testing anything related with true16 instructions and you did not use +real-true16 in the compile command, is that right?

The error is caused by t16 format instruction being selected while the 128-256 vgpr are mistakenly used. These t16 instructions should not be selected in default/fake16 flow.

https://github.com/llvm/llvm-project/pull/102198