profiling XEmacs, esp. on the C level

[Jan Rychter]

Ben Wing wrote:
> > So far, it seems that everyone's profiling on the Lisp level.  But C 
> > level profiling could give much better results.  Has anyone successfully 
> > used Gprof or the VC++ tools for doing C-level profiling?  When i tried 
> > in the past, i got stuck; things crashed and i couldn't get them to 
> > work.  Any success stories and instructions telling what exactly to do?

then, three weeks ago I wrote:
> Under Linux you could use oprofile
> (http://oprofile.sourceforge.net/download/). It isn't difficult to use
> and does not require profiler instrumentation when compiling (debugging
> information is useful, though). It also profiles the entire system, so
> it can show you how much time you spend running X11 functions.

Well, Ben had the right idea, so I actually went ahead and did some
profiling. It turned out that the major CPU-eater in modern builds was
mc-alloc and gc.

So, I went ahead and tried to see how I could make it faster.

The summary: relatively minor reworking of mc-alloc.c allows for about a
5% total performance increase as measured by my personal "gnus build
benchmark". Percentages for get_mark_bit, mc_alloc_1, sweep_page were
roughly halved.

Some of my tweaks are specific to 32-bit architectures at the moment,
especially one that replaces an expensive divide with a multiply, two
lookups and a shift.

I'll probably send the patch to Marcus Crestani, as it isn't
production-ready yet, it's closer to a proof-of-concept.

Does anybody have a good GC benchmark that I could use?

--J.