[llvm] [SimplifyCFG] Avoid branch threading of divergent conditionals (PR #141867)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> Thanks for the explanation. I see now why this change as written won't be possible to land. That being said, doing this does show valuable performance benefits in some real cases. Is there anything that can be done to improve handling of this case for GPUs?

I think we need a separate "undo jump threading" optimization pass 

https://github.com/llvm/llvm-project/pull/141867