[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

--- Comment #20 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Looking even deeper, there are two problems.  The first is that we miss C in
type hiearchy graph.

C however may be defined in other unit. We do mistake already while walking B. 
There are two variants of B::foo.  One when base A is non-virtual and other
when it is virtual.

B's vtable is:
_ZTV1C: 
        .quad   0
        .quad   8
        .quad   8
        .quad   0
        .quad   _ZTI1C
        .quad   _ZTv0_n24_N1B3fooEv
        .quad   -8
        .quad   0
        .quad   -8
        .quad   _ZTI1C
        .quad   _ZN1B3fooEv

Our walking logic misses the fact that descendands of B may use the thunk
instead of FOO based on presence of virtual inheritance.

I suppose for complete correctness I would have to add code associating methods
with their virtual thunks and lookup the thunks when walking B and knowng that
I am interested in the derived types.

Simpler solution solving both problems practically is probably to simply walk
known vtables and insert their types to the type hiearchy: all internal
devirtualizations will be correct and all the missed calls to thunk from vtable
of type defined externally are harmless: the visibility code already makes all
the assumption that the call can happen.