[Bug c++/63579] New: New attribute for empty member optimization

["glisse at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63579

            Bug ID: 63579
           Summary: New attribute for empty member optimization
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glisse at gcc dot gnu.org

Hello,

I am getting a bit sick of all the artificial base classes invented just so we
can benefit from the empty base optimization when we actually want a member. I
believe an attribute would be a nice way to work around this standard
misfeature:

template <class T1, class T2>
struct compressed_pair
{
  T1 t1 [[gnu::empty]];
  T2 t2 [[gnu::empty]];
// ... all the member functions
};

struct Empty { };
static_assert (sizeof (compressed_pair<int, Empty>) == sizeof (int), "too
long");


To be more specific, I was thinking of having the attribute mean: if the type
is a class type, for layout purposes pretend it is appended at the end of the
list of base classes (it also has an impact on non-empty members). But any
variant would do, as long as it stops being interesting to put a class as a
base instead of a member. The layout code handles bases differently enough that
it doesn't look as easy to implement as I had hoped. I'll attach a hack that I
am playing with.

We will also hit the ABI misfeature that means that in layout_empty_base we
only test offset 0 before jumping to eoc without testing offsets alignment,
2*alignment, etc. But that's another fight...