Hi Armin!
On Sat, Sep 24, 2005 at 10:10 +0200, Armin Rigo wrote:
> On Sat, Sep 24, 2005 at 08:21:47AM +0200, holger krekel wrote:
> > did anyone try to profile the translation process
> > itself recently? If so what where the results?
>
> I've been spending days on that, with the results that the standard
> library profile.py is very much broken -- it explodes in mysterious
> ways or just stop profiling after a few seconds.
>
> Michael and me have been polishing a new profiler posted on the CPython
> SF patch tracker, and I've got preliminary results on translate_pypy.
> I'll need to run it again now that a few last bugs in the profiler have
> been fixed. (Also, the profiler doesn't support dumping data to a file
> at the moment.)
Yes, i saw your works, good stuff. Sidenote: i guess it's
sometimes overlooked that PyPy developers have sometimes been
quite involved with CPython improvements.
> The place where translate_pypy spends the largest amount of time is
> objspace.flow.model.traverse(), which subsequently we again reduced and
> un-generalized. I guess we could at some point get rid of it completely
> and replace it with two methods on the graph object, to iterate over
> blocks or links.
Yip, the generality (i introduced IIRC at that time) is not really needed.
Refactoring to use specific methods on the graph/node object sounds
worthwhile.
cheers,
holger
_______________________________________________
[email protected]
http://codespeak.net/mailman/listinfo/pypy-dev
