Re: Fast handling of class-closure-only signals

[James Henstridge <james daa com au>]

On 28/08/03 03:48, Soeren Sandmann wrote:

> Recently, Tim applied a patch to gsignal.c that avoids emitting
> signals that won't have any effect.  Such signals are common during
> TreeView validation, so people who have performance problems here may
> want to upgrade to GLib CVS HEAD.
>
> Anyway, I got an idea from reading his patch. You can view what the
> noop patch does as:
>
>        - Is the class closure for this signal NULL?
>
>        - if yes, does anything else than the class closure need to run?
>
>        - if no, omit the emission.
>
> So my idea is to export API that does the check in the second step. By
> using that, gtk+ and other users of GObject can do
>
>    if (g_signal_check_class_closure_only (object, signal_id, detail))
>        FOO_GET_CLASS (object)->the_handler (...);
>    else
>        g_signal_emit (object, signal_id, detail);
>
> ie. directly calling the virtual function if nothing else needs to be
> done.
>
> For signals that only rarely have handlers, like "expose" or
> "size_allocate", but always a class closure, this is a big speed-up.
>
> I am attaching two patches, one for glib and one for gtk+. The glib
> patch exports the function g_signal_check_class_closure_only(), and
> the gtk+ patch makes use of this call to check if the event, size
> allocate, and the check_resize signals can be skipped.
>
> With these patches applied I measured with Gnumeric and Nautilus how
> often the signals were skipped by this during opaque resizing. The
> results are that
>
>   for Gnumeric 96.9% of the checked signals were skipped
>   for Nautilus 94.6% of the checked signals were skipped
>
> I also noted that the time spent inside g_signal_emit() (as measured
> with SpeedProf) during resize of Gnumeric dropped to around 2%.
>
> The main drawback is that for signals that actually *do* need to be
> emitted rather than directly called, the "class_closure_only" code
> will run twice. But as the numbers show, that is rare.
>  

Will this code affect objects where g_signal_override_class_closure() 
has been called?  In this case, the class closure may be the only 
callback for the signal, but it wouldn't be correct to short circuit to 
the C function pointer.

James.

--
Email: james daa com au
WWW:   http://www.daa.com.au/~james/