[llvm] Late temporal divergence lowering for SDAG (PR #67033)

[Nicolai Hähnle via llvm-commits]

nhaehnle wrote:

We'd still be lying in the IR, which is never a good idea.

I suspect the only reason why the sinking happens today is that we lie somewhere about the impact of the implicit EXEC use that should be on VALU instructions.

Let's please fix this properly, even if it means adding some AMDGPU-specific (or really: vectorizing-backend-specific) callbacks.

https://github.com/llvm/llvm-project/pull/67033