Armin's personal Blog

I was curious where most time in libinfinity, Gobby's infinote implementation, is spent. Not because it was too slow or something, but just because I wanted to know. So I used valgrind's callgrind tool to gain some profile data while running libinfinity's main test for the concurrncy control algorithm. When I displayed the result in kcachegrind (there is still no GNOME equivalent, no?) everything looked as I would have expected, until I noticed this line:

This means that 14.5% of the overall program time is spent in g_type_instance_get_private(). This is because libinfinity uses g_type_class_add_private for every class to store its members into, so it is easy to extend functionality later without breaking ABI. However, this also means that for every function that needs to access a member, g_type_instance_get_private() is called. I didn't expect such a huge impact, though.

To fix this, I added a simple gpointer to the public instance struct, and let it point to the private field in the instance_init function. So this means one call to g_type_instance_get_private() per instance, instead of roughly one per call to a function operating on an instance of the class in question. After doing this for the five classes suffering most from this, g_type_instance_get_private does no longer show up anywhere near relevant in kcachegrind.

I'd like it if improving performance was always as easy as this.