> On Mon, 17 Jun 2019, Jan Hubicka wrote:
>
> > >
> > > get_alias_set already handles VCEs properly. Btw I've said
> >
> > I see - i keep thinking of get_alias_set as a simple accessor to type's
> > alias set which it is not. It may make sense to separate these two
> > later.
> >
> > > BIT_FIELD_REF has the same issue but you didn't include that
> > > case below.
> >
> > Sorry, missed that in your reply. I am testing the fix.
> >
> > What about aliasing_component_refs. When I have VCE in the the path,
> > I suppose I do not want to consider the stuff after VCE as part of path
> > and also while testing whether one path can be continuation from other
> > I want to start from outermost VCE rather than ref_type, right?
>
> Yeah, when searching for base1/2 you want to initialize ref1/2p
> like
>
> while (handled_component_p (base1))
> {
> if (TREE_CODE (base1) == VIEW_CONVERT_EXPR
> || TREE_CODE (base1) == BIT_FIELD_REF)
> ref1p = &TREE_OPERAND (base1, 0);
> base1 = TREE_OPERAND (base1, 0);
> }
>
> and then instead of refp = &ref1/2 do refp = ref1/2p for the searches.
>
> So not use the type of the innermost(!) VIEW_CONVERT_EXPR but its
> base (same for BIT_FIELD_REF).
Yep, that is what I had in mind. I will test the patch.
Also will remove the ref pointers - we always just read them.
Thanks,
Honza
