Aren't discussions via blogs fun? :)
Federico: two points:
-
Using the new GObject instance private data API, rather than the private data scheme that GNOME hackers have always used, is a small optimization in itself - the private data is allocated in the same chunk as the object itself.
As with all optimizations, you weigh up the benefit against the code obfuscation. In this case, I don't think the GObject scheme makes the code more difficult to understand ... especially since it's likely to become as much of a second-nature idiom as the old scheme. It's also one less opportunity to leak memory.
-
Using the GObject scheme, instance private data could have been added without the runtime hit you were seeing ... even where the object structure couldn't be extended without breaking ABI.
To do so, you'd just go back to something similar to Owen's initial suggestion for how the API should be used - in a static variable, you'd store the offset between the address of the object structure and the private data and use that offset for efficient access to the data.
add_private()originally returned this offset, but it no longer does, so you'd need to calculate the offset ininstance_init()- but it is guaranteed to be constant.Granted, that's a nasty hack which would genuinely obfuscate the code ... but at least it would actually be possible to add private data, whereas it wouldn't be possible with the old scheme.
Update: Tim points out that the offset to the private data isn't constant across all instances - the offset will be larger for subtypes since the private data is allocated after all object data.