spatel added a comment. Here's a partial fix that I think should be universally good (ie, can't cause regressions for any other analysis/pass and has no obvious compile-time impact): https://reviews.llvm.org/rL336293 https://reviews.llvm.org/D48930