https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85726
Hans-Peter Nilsson <hp at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|ASSIGNED |NEW Assignee|hp at gcc dot gnu.org |unassigned at gcc dot gnu.org --- Comment #5 from Hans-Peter Nilsson <hp at gcc dot gnu.org> --- (In reply to Jakub Jelinek from comment #4) > Regressed with r218211 aka PR15346. Hans-Peter, are you still working on > this or where the thread died? Oops, no I'm not actively working on this; de-assigning now. (I did mean to pick up the thread, sooner or later. For example, I checked that my match.pd patch for div (without the mult parts) had only benevolent effects AFAIU for x86_64, but got distracted before reporting on that. And, changes were requested in the this-is-not-a-review-replies (i.e. a side-condition of mod also matching or similar) that I had to figure out how to implement; maybe that's what you did.) So... > Do we want to do just something directed at this case, i.e. don't combine > two divisions if the lhs of the first one is also used in a modulo with the > same last operand as the second division and both are e.g. in the same bb? I don't know. If your patch handles the test-case, as I guess it does from a quick look, then I'm happy. The original code was manually unrolled loops and I *think* all the expressions were in the same bb. Thanks!