https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79191
--- Comment #5 from Marc Glisse <glisse at gcc dot gnu.org> --- (In reply to Eric Gallager from comment #4) > Where exactly in the compiler is this optimization supposed to be done and Eric, that's part of the problem. The optimization already exists in VRP, but it is defeated by another optimization in match.pd. We could restrict the match.pd optimization with single_use, although this might cause us to miss other optimizations. If we replace Y=(unsigned long)(unsigned)X with Y=X&4294967295, we could try and replace (unsigned)X<3 with Y<3 at the same time, but it isn't clear how/where/when to do that (restrict the transformation to single_use in match.pd and add the super special code in tree-ssa-forwprop.c?). We could try and introduce some magic in VRP (without value numbering, it could be, when processing X & 4294967295, checking the other uses of X for a conversion to a 32-bit int and using its interval if there is a suitable domination relation between the relevant statements), but that seems hard and ugly. Basically the PR seems to be asking for a better idea on where to perform this optimization ;-)
