fhahn added a comment. Yep I think we could preserve all CFG analysis. But wouldn't that potentially lead to worse results, if a preserved analysis could benefit from the propagated constants? https://reviews.llvm.org/D47149