On 06/13/2012 08:02 AM, Paolo Bonzini wrote:
> Il 13/06/2012 14:59, Andreas Färber ha scritto:
>> Ouch! One can argue that's still not recursive, but what matters more
>> this borks Anthony's in-place object_initialize() concept.
>>
>> Two solutions come to mind:
>> * allocate the interfaces as part of object_new() beyond instance_size
>
> That won't work if you initialize in place, because you cannot allocate
> the room for the interface.  It is possible to put Interface objects
> explicitly in the class, and pass an offset when registering the type so
> that they can be initialized in place.
>
> But I still think we're fighting windmills...

There's no problem in my mind with allocating interfaces on the heap.  in-place 
initialization is a readability thing, it's not a memory management thing.  It's 
so you can see that:

struct PIIX3 {
   ...

   RTCState rtc;
   APICState *apic;
};

'rtc' is a child and 'apic' is a link.

Anyway, I don't like the idea of making interfaces concrete.  That means that a 
user could directly instantiate an interface which doesn't make a lot of sense.

Here's a different solution.  This has not been even compile tested but I think 
the concept is sound.

Regards,

Anthony Liguori

>
> Paolo
>
>