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

[hubicka at ucw dot cz]

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

--- Comment #8 from Jan Hubicka <hubicka at ucw dot cz> ---
> Ah, right.  We don't have an A binfo in C's base list.
> 
> I believe you can distinguish this case by looking at the
> BINFO_INHERITANCE_CHAIN of the A binfo; in this case, it should point back to C
> rather than to B because A is a primary base of C and not of B-in-C.

BINFO_INHERITANCE_CHAIN is cleared by free_lang_data, so it is not available to
me.
It would not be hard to change this and add streaming of this pointer.  

I have bit problem to interpret the comment:

/* The BINFO_INHERITANCE_CHAIN points at the binfo for the base
   inheriting this base for non-virtual bases. For virtual bases it
   points either to the binfo for which this is a primary binfo, or to
   the binfo of the most derived type.  */

So basically given a BINFO I walk I can look at BINFO_INHERITANCE_CHAIN and if
my offset is the same as offset of BINFO_INHERITANCE_CHAIN I will take its
virtual table and if not, I will take virutal table of BINFO that will always
be non-NULL?

Will this work in a case of non-virtual multiple inheritance? I.e

       A
      / \
     B  C
       / \
      D   E
where I think A+B share vtable and C+D share vtable?
I need D BINFO_INHERITANCE_CHAIN to point to C instead of A, but in the case of
virtual inheritance I need it to point to A instead of C.

It is not big deal to change record_binfos to walk the fields
instead of base list and perhaps it would end up being easier than the above. 
However I am not sure about get_binfo_at_offset implementation either.  it does

1) walk fields and find corresponding filed
2)  a) if field is not artificial  then it switches to new binfo
   otherwise:
    b) if offset is non-0 it walks bases to look up BINFO corresponding to the
field.

What will hapen in the case of multiple virtual inheritance? Assume the example
above has all bases virtual.  Here I think D will be field of A with non-0
offset
but it will not be in list of bases of A. So I think get_binfo_at_offset will
fail
in b) and will return NULL.

I think we do not hit this problem only because devirt machinery do not
understand virtual inheritance lookup - it basically never gets into the outer
type.

Honza