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

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

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

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jason at redhat dot com
             Blocks|                            |58252

--- Comment #3 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Thanks for nice testcase.

The problem here is that record_binfos and get_binfo_at_offset do not agree.
I think it may be the same problem as in PR58252.

Here we have class C with one (virtual) base B that has one (virtual) base A.

record_binfos starts walk in binfo of C and it looks at base B.  It passes two
binfos around.  type_binfo is binfo pointing to a virtual table to use, while
binfo is the BINFo being walked.  For non-virtual inheritance the first base
share virtual table with outer class and its BINFO has no pointer to any
vtable. So there is a conditional setting type_binfo to the base only if its
vtable is non-NULL.

Now for virtual inheritance, B have its own table defined, so we switch to B
and eventually we look into vtable assigned to A.

get_binfo_at_offset works slightly differently.  It first walks fields of the
structure identifying the field at proper offset.  If it is abstract it is
considered to represent base type.  Then if offset is non-0 it uses BINFO of
the base, while otherwise it stays on BINFO of outer type.

For the testcase get_binfo_at_offsets is called with offset 0 and it makes
lookup in the vtable pointed by BINFO of C that contains a thunk, while
record_binfos looks into vtable pointer by BINFO representing base A within C
that has directly method foo.

Obviously both functions must be made to agree.

C::C seems to be compiled as:
_ZN1CC1Ev:
.LFB12: 
        .cfi_startproc
        .cfi_personality 0x3,__gxx_personality_v0
        movq    _ZTT1C+24(%rip), %rax
        movq    _ZTT1C+32(%rip), %rdx
        movq    -32(%rax), %rax
        movq    %rdx, 8(%rdi,%rax)
        movq    $_ZTV1C+88, 8(%rdi)
        movq    $_ZTV1C+40, (%rdi)
        movq    _ZTV1C+40(%rip), %rax
        jmp     *%rax

and
_ZTV1C: 
        .quad   0
        .quad   8
        .quad   8
        .quad   0
        .quad   _ZTI1C
        .quad   _ZTv0_n24_N1B3fooEv

so we really call thunk here and probably record_binfos is incorrect.  I can
fix it by also waking fields and handling fields at offset 0 specially, but I
would hope there is direct way how to do so via BASE_BINFOs walk....

I tried to change conditional in record_binfos to only switch type_binfo if 
BINFO_OFFSET != 0, but in this case it is 8. This also seems incorect to me...

Jason, any ideas about the proper fix?

Honza