[llvm] [LoopUnroll] Penalize interior control flow (PR #67137)

[via llvm-commits]

Andarwinux wrote:

Any progress?

> I believe that fully unrolling loops with interior control flow tends to be less profitable, because we don't get straight-line (or extended-BB) code out of it, so it is less amenable to further optimization. It also dilutes branch predictors.

https://github.com/llvm/llvm-project/issues/203776

I found an example of this.

https://github.com/llvm/llvm-project/pull/67137