[llvm] [AMDGPU] Do not allow M0 as v_readfirstlane_b32 dst (PR #128851)

[Jay Foad via llvm-commits]

https://github.com/jayfoad approved this pull request.

LGTM.

It's annoying that so many bits of C++ code duplicate the name of the regclass, which is already available from the instruction definition.

Should also do v_readlane_b32 as Matt said.

https://github.com/llvm/llvm-project/pull/128851