[llvm] [Uniformity] Fixed control-div early stop (PR #139667)

[Junjie Gu via llvm-commits]

jgu222 wrote:

I agree that IPD is the last join and the algo should stop at the IPD.

I did experiment disabling the early stop, and I got some regressions on irreducible lit tests. I am not sure what is wrong and think maybe irreducible cycles are specially handled and rely on particular propagation.

As you think the IPD should be good for irreducible cycles too, I will redo the patch to use dominator analysis.

https://github.com/llvm/llvm-project/pull/139667