Re: Make'em suffer!

[Eero Tamminen <ext-eero tamminen nokia com>]

Hi,

ext Federico Mena Quintero wrote:
> > Note: Deep widget hierarchies add noticable overhead because of size 
> > request / allocate propagation.
>
> Yeah, and that's going to be rather tricky to test...  Some widgets do
> take a good amount of time in finding out their requisition
> (GtkTextView); some others should be trivial (GtkImage).  Then you have
> the overhead from the generic sizing mechanism.
>
> Do you have profiles where the generic sizing code shows up as a
> culprit?  I'd expect the problems to be in fat widgets like GtkTextView
> and GtkTreeView...

This is very old information (2-3 years), for Gtk 2.4 (I think)
and I don't have the profiles for this.  Nowadays we avoid deep
widget hierarchies, so we don't see this. :-)

However, I think the problem was in an application that used a lot of
containers to position things correctly.  I think it was "fixed" back 
then by using just GtkFixed instead...

So, I would suggest having in the torturer one test with just a lot
of different containers inside each other.


> > Btw. With large themes (hundreds of styles like in Maemo), binding the
> > styles to a widget shows up clearly in profiler if you create for
> > example a treeview with hundreds of items.
>
> Oooh, interesting.  Do you have a profile?

This is also quite old information (1 year) and I don't have the
profiles, but the bottleneck was in finding (matching) the correct
style for the corresponding object. The Maemo theme files were optimized
a bit after that (there were some styles that used more than one * for
matching) and split to separate files, so it might be quite a bit less 
problematic nowadays.  I could cachegrind this a bit in few weeks.

I think only Maemo has such a large number of styles in the themes that
this is a problem.  Or are there other themes with hundreds of styles? :-)


> [GtkCellRendererText does some stupid things like causing text to be
> measured two or three times; see my old profiles for the file chooser
> for the details.]

I would suggest also a torture test for each CellRenderer type
in which hundreds of them would be put into a list.


	- Eero