[llvm] [AMDGPU] Ensure all WMMA instructions are marked as convergent (PR #178314)

[Frederik Harwath via llvm-commits]

https://github.com/frederik-h requested changes to this pull request.

I have also been dealing with a bug where an WMMA instruction is being sunk because it is not marked as convergent. In my case, this is `V_MFMA_SCALE_F32_32X32X64_F8F6F4_f4_f4_mac_vgprcd_e64`. This is not covered by your PR. I have opened a PR to this PR's branch on your fork which marks this kind of instructions as convergent as well. In order to do justice to this PR's name, I think my change could be included here. I can also open a separate PR, but in this case I would suggest to change the PR description to make it clear which kinds of WMMA instructions are covered by it. Otherwise, this PR looks good to me.

https://github.com/llvm/llvm-project/pull/178314