[Bug middle-end/58585] [4.9 Regression] ICE in ipa with virtual inheritance

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58585

--- Comment #4 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
This is the code I am playing with:
  /* Walk bases.  */
  for (i = 0; BINFO_BASE_ITERATE (binfo, i, base_binfo); i++)
    /* Walking bases that have no virtual method is pointless excercise.  */
    if (polymorphic_type_binfo_p (base_binfo))
      record_binfo (nodes, base_binfo, otr_type,
                    /* In the case of single inheritance, the virtual table
                       is shared with the outer type.  */
                    BINFO_OFFSET_ZEROP (base_binfo)
                    ? type_binfo : base_binfo,
                    otr_token, inserted,
                    matched_vtables, anonymous);

(i.e. using BINFO_OFFSET_ZEROP check instead of BINFO_VTABLE != NULL).

I think I understand the problem now.  The binfos of C looks like

  BINFO of C, BINFO vtable is C's vtable
  BASE:
    BINFO of B, BINFO vtable is non-NULL, offset is 8
    BASE:
      BINFO of A, BINFO vtable is non-NULL, offset is 0.

i.e. by the virtual inheritance the A is outside layout of B. So while B's
VTABLE is set, A's VTABLE should be VTABLE of C, my code however considers
VTABLE of B.

get_binfo_at_offset won't get confused, since type of C contains both A and B
as fields, so it gets to BINFO of A directly without walking BINFO of B.

Is there way how to keep track of the vtables w/o doing the walk based on
fields instead of BINFO_BASEs? If not, I will simply switch record_binfos to it
- it seems quite straighforward, but it will end up with more walking, since
there are more fileds than bases.