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

[Gang Chen via llvm-branch-commits]

cmc-rep wrote:

> We currently have WWM (whole-wave mode) intrinsics to expand the exec mask and do operations that need the whole wave active. However, WWM intrinsics are rather limited, especially once one tries to use it for more complex cases.
> As far as I know, trying to use control-flow with WWM intrinsics is rather complicated, if it works at all.
> Extending WWM intrinsics to better support control-flow and more IR in general is complex and maybe not something desirable (the way WWM intrinsics work is rather brittle).

I just quote Sebastian's previous reply here:

"We currently have WWM (whole-wave mode) intrinsics to expand the exec mask and do operations that need the whole wave active. However, WWM intrinsics are rather limited, especially once one tries to use it for more complex cases.
As far as I know, trying to use control-flow with WWM intrinsics is rather complicated, if it works at all.
Extending WWM intrinsics to better support control-flow and more IR in general is complex and maybe not something desirable (the way WWM intrinsics work is rather brittle)."



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