This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH][java] Fix type problems in the Java frontend
Richard Guenther writes:
> On Thu, 12 Jul 2007, Andrew Haley wrote:
> > Tom Tromey writes:
> > > >>>>> "Richard" == Richard Guenther <firstname.lastname@example.org> writes:
> > >
> > > Richard> (It looks like there is funny stuff going on in the FE - it does
> > > Richard> the equivalent of *(T **)p with p of type (T *). This may ask for
> > > Richard> trouble with aliasing.)
> > >
> > > Yeah, there are some oddities with types. I don't recall this one
> > > offhand. Could you file it? Or update the existing gcj "types are
> > > weird" PR with more info?
> > My understanding is we are perfectly entitled to do this in the front
> > end. If C's aliasing rules (which, of course, only apply to C and
> > C++) are to be enforced with GIMPLE then I need a clear statement of
> > the fact, along with rules about exactly what I am allowed to do in
> > the front-end.
> It is the front-ends business to make sure to tell the middle end
> that for the above case (T *) and (T **) alias (if the pointer p
> above is ever dereferenced without casting it to (T **) before - as
> only mismatched accesses are troublesome, the pointer type itself
> is of course not). This is done through the usual way with the
> get_alias_set langhook and recording proper conflicts.
Right, that's what I thought. So, the alias sets are the
responsibility of the FE, and if we cast between pointers to different
alias sets that's a bug in the FE.