https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79095
--- Comment #5 from Marc Glisse <glisse at gcc dot gnu.org> --- (In reply to Jeffrey A. Law from comment #4) > So the simplified tests are interesting, but ultimately too simplified. The > VRP prototype which works on the simplified testcases doesn't work on the > real testcase for this BZ. > > The key sequences look like: > > _13 = (long unsigned int) _12; > _121 = ADD_OVERFLOW (_13, 18446744073709551615); > _1 = REALPART_EXPR <_121>; > _39 = IMAGPART_EXPR <_121>; > if (_1 > _13) > goto <bb 4>; [29.56%] > else > goto <bb 18>; [70.44%] > > > Which defeat the simplistic prototype for VRP. It's essentially still the > same construct, but we'd have to extract stuff out of the COMPLEX_EXPR and > recognize the ADD_OVERFLOW. I thought ADD_OVERFLOW would appear in the widening_mul pass, i.e. after all the VRP passes are done? On the other hand, teaching VRP about *_OVERFLOW sounds like a good idea.
