[llvm] [AMDGPU] Allow combining uniform OR/AND to V_PERM (PR #220048)

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

Does this include tests where the user of the v_perm is uniform alu?

> I can kind of understand this, but note that it is the exact opposite of our normal codegen strategy.

We have seen some other cases recently where we go against the isel optimization strategy to facilitate better scheduling. I'm concerned that these ad-hoc rules are muddying up the general isel strategy, and we may be better off by having these rules be packaged with a certain scheduler, or something similar.

https://github.com/llvm/llvm-project/pull/220048