Re: shortcuts for g_type_register_static_simple

[Stefan Kost <ensonic hora-obscura de>]

hi,

I just converted one of my classes to use G_DEFINE_TYPE to benefit  
from less typing and from the less relocations. When trying it I was  
suprised to see that it actally adds two more relocations:

$ LD_DEBUG=statistics ./test-before
     30264:                      number of relocations: 2566
     30264:           number of relocations from cache: 2403
     30264:             number of relative relocations: 23747
     30264:
     30264:     runtime linker statistics:
     30264:                final number of relocations: 7783
     30264:     final number of relocations from cache: 2407
$ LD_DEBUG=statistics ./test-after
     31894:                      number of relocations: 2566
     31894:           number of relocations from cache: 2403
     31894:             number of relative relocations: 23745
     31894:
     31894:     runtime linker statistics:
     31894:                final number of relocations: 7785
     31894:     final number of relocations from cache: 2407

Anyone an idea whats causing this?  Matthis do you remember how you  
verified that this helps?

Stefan

Quoting Tim Janik <timj gtk org>:

> On Tue, 8 May 2007, Stefan Kost wrote:
>
> > Hi,
> >
> > I was reffering to the idea Matthis Clasen mentioned in this thread:
> > http://mail.gnome.org/archives/gtk-devel-list/2006-April/msg00169.html
> >
> > I look at my own application. Refdbg tells me that I use 162   
> > different types. So the overhead seems to be acceptable. Running   
> > oprofile shows these two top in the report for libgobject which   
> > itself has 7 % of the total time.
> > 168      11.2299  g_type_check_instance_is_a
> > 118       7.8877  g_type_is_a
>
> yeah, known bottlenecks. the issue with those two functions is really
> that they
> are being called from an insane amount of places. we've tried to squeeze best
> performance out of those already (normal type is_a checks are O(1), interface
> type is_a checks are O(ld(n_ifaces_per_type))). there don't seem to be
> any easy
> fruits left in here.
> we've even implemented obvious call reductions here already, e.g. see the
> __r = TRUE branch in the implementation of G_TYPE_CHECK_INSTANCE_TYPE():
>
> #  define _G_TYPE_CIT(ip, gt)             (G_GNUC_EXTENSION ({ \
>   GTypeInstance *__inst = (GTypeInstance*) ip; GType __t = gt;   
> gboolean __r; \
>   if (__inst && __inst->g_class && __inst->g_class->g_type == __t) \
>     __r = TRUE; \
>   else \
>     __r = g_type_check_instance_is_a (__inst, __t); \
>   __r; \
> }))
>
> > Nothing big to gain here, so it seems.
> >
> > Stefan
>
> ---
> ciaoTJ