On Thu, May 25, 2017 at 10:51:56AM +0200, Jakub Jelinek wrote:
> On Tue, May 09, 2017 at 04:37:16PM -0400, Jason Merrill wrote:
> > For C++17 aggregate bases, we have started adding base fields for
> > empty bases. The code for calculating whether a class is standard
> > layout needs to ignore these.
> >
> > The C++17 mode diagnostic for direct-enum-init1.C was incorrect.
> >
> > Tested x86_64-pc-linux-gnu, applying to trunk.
>
> > commit 9a612cc30d4b3ef905ce45304545d8b99a3cf5b9
> > Author: Jason Merrill <[email protected]>
> > Date: Tue May 9 14:15:38 2017 -0400
> >
> > * class.c (check_bases): Ignore empty bases.
>
> This should have referenced PR c++/80605 (and is also a 7 regression).
>
> > diff --git a/gcc/cp/class.c b/gcc/cp/class.c
> > index fc71766..085dbc3 100644
> > --- a/gcc/cp/class.c
> > +++ b/gcc/cp/class.c
> > @@ -1860,7 +1860,9 @@ check_bases (tree t,
> > members */
> > for (basefield = TYPE_FIELDS (basetype); basefield;
> > basefield = DECL_CHAIN (basefield))
> > - if (TREE_CODE (basefield) == FIELD_DECL)
> > + if (TREE_CODE (basefield) == FIELD_DECL
> > + && DECL_SIZE (basefield)
> > + && !integer_zerop (DECL_SIZE (basefield)))
>
> Is that what we really want? I mean, shouldn't we at least also
> check that the basefield we want to ignore is DECL_ARTIFICIAL,
> or that it doesn't have DECL_NAME or something similar, to avoid
> considering user fields with zero size the same?
> I believe your change changes e.g.:
> struct S { int a[0]; };
> struct T : public S { int b[0]; int c; };
> bool q = __is_standard_layout (T);
> which previously e.g. with -std=gnu++14 emitted q = false, but
> now emits q = true.
We even have DECL_FIELD_IS_BASE macro, so can't the above be
if (TREE_CODE (basefield) == FIELD_DECL
&& !DECL_FIELD_IS_BASE (basefield))
or
if (TREE_CODE (basefield) == FIELD_DECL
&& (!DECL_FIELD_IS_BASE (basefield)
|| (DECL_SIZE (basefield)
&& !integer_zerop (DECL_SIZE (basefield)))))
or something similar?
Jakub
