> On Tue, 11 Jun 2019, Jan Hubicka wrote:
>
> > > So the important thing is that during matching the paths
> > > aptr->a and *bptr we return false for
> > > same_type_for_tbaa (int *, struct a *), correct? But if
> > > we can do that then the alias sets should not conflict in
> > > the first place and we shouldn't need any path-based disambiguation
> > > here.
> > >
> > > So to me this is the wrong place to fix.
> >
> > The alias sets of int * conflict with alias set of struct a *
> > since that one is same as void * and we make void * to conflict
> > with all pointers.
> >
> > However we also do not want to return -1 for things like
> > same_type_for_tbaa (int *, struct a)
> > which we do currently, since int * is TYPE_STRUCTURAL_EQUALITY
> > and struct a is not.
>
> Yes, but that basically boils down to handling !AGGREGATE_TYPE_P
> vs AGGREGATE_TYPE_P in a better way, not looking into pointed-to
> types.
OK, so your preferred vairant would be to make same_type_for_tbaa
to return 0 instead of -1 if one type is AGGREGATE_TYPE_P and other
!AGGREGATE_TYPE_P and extend alias set calculation to somewhat account
for the non-transitive rules for pointers TBAA?
What about arrays?
Honza
>
> > Here you may declare struct a as
> > struct a {int *a;};
> > and alias sets will be in conflict even w/o void * globing.
> >
> > Honza
> >
>
> --
> Richard Biener <[email protected]>
> SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany;
> GF: Felix Imendörffer, Mary Higgins, Sri Rasiah; HRB 21284 (AG Nürnberg)
