LGTM -- based on stats this should be low impact, so let's try it. This very runs the pass two times (before and after unrolling), which is probably more than we need longer term but for now should hopefully avoid regressions.
The test regressions in this diff are just an artifact of inference moving into a different pass (the tests run instcombine as part of the pipeline, while is no longer responsible for alignment inference). If you see end-to-end regressions in alignment information, that would be a problem though. Do you have an example?
Ah, I see.
No, didn't see end-to-end regressions. I was looking somewhere in the middle of the pipeline only.
But this may impact alias analysis, depending on how often and there this new pass runs. Instcombine runs more often than this pass. We need to monitor for regressions.