[llvm-branch-commits] [llvm] [AMDGPU] Add machine-level inliner pass (PR #169476)

[Aiden Grossman via llvm-branch-commits]

https://github.com/boomanaiden154 commented:

> Why do we need this? I think this incurs quite a lot of complexity and infrastructure maintenance cost. Can we avoid this completely?

Agreed. It is unclear to me why this cannot be done at the IR level. Even if we can only know what we need to inline really late in the pipeline, it would still be possible to e.g., add `always_inline` where necessary and run the AlwaysInliner pass again. We should already know the calling convention during the IR stages of codegen.

> If we go forward with this, I think the new pass manager patch needs to be already implemented with the stack, and submitted at the same time. We cannot accept new blockers to enabling the new pass manager in codegen.

Also agreed. I don't think this would be terribly hard to implement as a module pass.

https://github.com/llvm/llvm-project/pull/169476