https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103168
--- Comment #16 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Jan Hubicka from comment #13) > Concerning comment #10, the problem was that the loop walking all accesses > was missing loads->every_base check. This is used to represent that we > track no useful info about loads performed at all. > > Anyway if I read the code correctly, it does nothing useful if the access > tree contains any access for which we can not construct ref and thus one can > simply check global_memory_access and do not care about > every_base/every_ref/every_access since these must be all false. > > I simplified the walk a bit and added code pre-computing number of accesses > in the tree into the summary. > > What we can also do is, when hitting access for which we can not construct > ref, or when hitting every_ref/every_acccess, is to construct ref with > base_alias_set/ref_alias_set as given by the access tree but with base=NULL, > offset=0 and size=max_size=-1. This should still let the basic TBAA oracle > to disambiguate. I don't think we support base = NULL in the oracle (ref = NULL is supported though). Can you attach your patch instead of cut&pasting so I can take it from there? Thanks.
