[Development] qt_add/removeObject hooks

[Kevin Funk]

On Wednesday 16 April 2014 14:46:56 Branislav Katreniak wrote:
> > As mentioned on https://codereview.qt-project.org/#change,83272 Kevin Funk
> > measured the difference between the current approach and using atomics
> > using Olivier's initial benchmark, it's about 4-5x slower for checking if
> > there's a callback set. Since this is affecting every Qt application on a
> > not uncommon hot path, I'd rather trade in theoretical tool stability for
> > general performance.
> 
> The register / unregister methods can do the actual work in the
> QApplication thread to avoid locking.

That's too simple-minded.

The problem is that not just register/unregister (cf. QInternal::
{registerCallback,unregisterCallback}) would need to be thread-safe, but also 
the code path where the callback is triggered (cf. 
QInternal::activateCallbacks). Remember, assignments to pointer-sized 
variables are not atomic on all platforms, hence QInternal::activateCallbacks 
could be triggered while some callback function pointer is still garbage.

And we cannot just call that one in the main thread as that is a hot path 
(esp. for addObject/removeObject events).

> To support more callbacks, the pointer can be pointer to linked list
> of function pointer + context instead of pure function pointer.

Yep. That's what QInternal already does.

Greets.

-- 
Kevin Funk