Re: [PATCH] Fix incorrect devirtualization (PR middle-end/48661)

[Jakub Jelinek <jakub at redhat dot com>]

On Mon, Apr 18, 2011 at 03:33:18PM -0700, Jason Merrill wrote:
> On 04/18/2011 02:40 PM, Jakub Jelinek wrote:
> >If TREE_BINFO has BV_VCALL_INDEX set, this needs to be dynamically
> >adjusted, but none of the callers are prepared to handle that.
> 
> Well, it means that we do dynamic adjustment at runtime.  If we're
> able to do devirtualization, we should be able to figure out the
> right offset as well, just not in 4.6.

Sure, but how exactly?  If BV_VCALL_INDEX is non-NULL,
I guess we need to find for the given binfo corresponding rtti_binfo
that was (or would be if vtable isn't emitted in the current TU)
used in build_rtti_vtbl_entries to compute the difference between
BINFO_OFFSETs:
  offset = size_diffop_loc (input_location,
                        BINFO_OFFSET (vid->rtti_binfo), BINFO_OFFSET (b));
That is the value actually added on top of the fixed delta, right?
Or do you think it should just add to the delta the value read from the
vcall_offset from the vtable?

But first of all PR48674 needs solving on the trunk.
I think cgraph thunk nodes have fixed_offset field so it is easy to figure
out that adjustment, but for vcall_offset adjustment the same applies here,
is there a way to find out what value will be added, or does it need to be
dynamic?  And what kind of info do we need to find that out.

	Jakub