Re: PROPOSED PATCH: Robustify cp_fold_obj_type_ref() for ObjC++ use

[Andrew Pinski <pinskia at physics dot uc dot edu>]

On Sep 20, 2004, at 10:40 PM, Mark Mitchell wrote:

> Ziemowit Laski wrote:
>
> > For some reason, RTH introduced OBJ_TYPE_REF nodes for dispatching 
> > ObjC/ObjC++ method
> > calls.  In ObjC, this just falls through to the surrounding 
> > CALL_EXPR, but in ObjC++
> > it trips over cp_fold_obj_type_ref() since ObjC classes do not 
> > actually have any
> > virtual member functions.  The change below is the easiest way to get 
> > things going
> > again.  Would that be OK (with ChangeLog entry, of course), or should 
> > redirect this
> > to an ObjC++-specific lang hook?  Or perhaps revert the use of 
> > OBJ_TYPE_REF for
> > ObjC/ObjC++ methods altogether (as I honestly don't know what it is 
> > trying to accomplish)?
> >
> > Please advise,
>
> It looks like you're not only getting OBJ_TYPE_REF nodes -- you're 
> getting OBJ_TYPE_REF nodes that indicate that you're using a virtual 
> function.  (That's the OBJ_TYPE_REF_TOKEN.)  Either the Objective-C 
> and C++ front ends disagree about what OBJ_TYPE_REF_TOKEN means, or 
> you shouldn't be building an OBJ_TYPE_REF at all in objc-act.c.  My 
> guess is the latter.

No, the Objective-C front-end is correct in creating a OBJ_TYPE_REF
node as it is a virtual call (not like C++ but still a virtual call).
If we make OBJ_TYPE_REF_TOKEN a null then we get a crash in
cp_fold_obj_type_ref, anyways.

Basically in the C++ front-end we are assuming that BINFO_VIRTUALS
is non-null which is not true for Objective-C types which can
be located in OBJ_TYPE_REF so the bug is really in
cp_fold_obj_type_ref and should be fixed in a way which Zem proposed
either by the patch or by a hook.

Thanks,
Andrew Pinski
Someone who would like to see the Objective-C++ patches just approved
instead of being brutalized unlike optimization patches which you
(Mark) basically approved without any or little comment.