https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65752
--- Comment #54 from rguenther at suse dot de <rguenther at suse dot de> --- On Fri, 21 Oct 2016, law at redhat dot com wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65752 > > --- Comment #53 from Jeffrey A. Law <law at redhat dot com> --- > And presumably walking forward/backward from the equivalency point to > determine > if an object is derived from or is used to derive a pointer is insufficient > because the equivalency point and casting to/from pointer types might be in > different functions? Yes. > Is this the final straw and do we just stop recording these conditional > equivalences when both sides are SSA_NAMEs and declare the fallout > unavoidable? I haven't made up my mind on a possible good solution here. But yes, it would mean to not propagate any equivalences derived from conditionals, not even symbolic constants such as &a btw. > And note I believe we do similar stuff in the RTL optimizers too. RTL has it's own share of issues, see the linked PR about its very optimistic treating of "base value"s.
