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

[via llvm-commits]

petar-avramovic wrote:

I was trying to understand better what was the issue and I think it is not target specific but order of generic IR passes specific.
LICM was moved later in the pipeline https://github.com/llvm/llvm-project/commit/5b657f50b8e8dc5836fb80e566ca7569fd04c26f.
In old days LICM would sink all these add instructions out of the loop but now there are some lcssa phi in the way.
This happens before target specific control flow intrinsics are introduced.
We probably should not depend on optimization passes for correctness.

OK, I will add target hook for this.

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