[llvm] [SimplifyCFG] Skip threading if the target may have divergent branches (PR #100185)

[Nikita Popov via llvm-commits]

nikic wrote:

The usual answer to "why does SimplifyCFG perform this optimization that another pass does more comprehensively" is "phase ordering". There are a lot of SimplifyCFG runs in the optimization pipeline, while we cannot afford to run dedicated passes (with expensive dependencies like LVI) so many times.

It's possible that this logic in SimplifyCFG is indeed no longer necessary, and you're welcome to give removing it a try, but I think the most likely outcome is that dropping it does cause regressions...

https://github.com/llvm/llvm-project/pull/100185