I'm talking about PyPy at KiwiPyCon in a few weeks: http://nz.pycon.org/talks/talk/3/ My basic plan is to give a similar talk to the one I gave at OSDC last year (http://codespeak.net/svn/pypy/extradoc/talk/osdc2008/osdc08.pdf, which in turn is pretty similar to talks I gave back in 2006...). I've not been following recent progress super closely, so if someone who's been more involved could read the OSDC talk and point out things that are no longer true or not really relevant any more, that would be great :) I guess I should say a bit about the current approach to the JIT... Cheers, mwh
As Jacob pointed, we don't have half-maintained backends any more. You can have a bit more info about the JIT if you like, since this is what we're currently focused on (look at pypy blog for details). The GCs are rather good, there are no plans to improve them in the near future I guess. Impressive demo is to run the gameboy interpreter and play mario (targetgbstandalone I suppose). Cheers, fijal On Sat, Sep 26, 2009 at 4:29 PM, Michael Hudson-Doyle <micahel@gmail.com> wrote:
I'm talking about PyPy at KiwiPyCon in a few weeks: http://nz.pycon.org/talks/talk/3/
My basic plan is to give a similar talk to the one I gave at OSDC last year (http://codespeak.net/svn/pypy/extradoc/talk/osdc2008/osdc08.pdf, which in turn is pretty similar to talks I gave back in 2006...). I've not been following recent progress super closely, so if someone who's been more involved could read the OSDC talk and point out things that are no longer true or not really relevant any more, that would be great :) I guess I should say a bit about the current approach to the JIT...
Cheers, mwh _______________________________________________ pypy-dev@codespeak.net http://codespeak.net/mailman/listinfo/pypy-dev
Hi Michael, I don't know if it's a good idea to continue saying that our JIT is "Psyco-style". Mostly it's not :-) It's a tracing JIT now. "The prototype JIT compiler ran programs X times faster than CPython"... Assuming you are now talking about the new JIT compiler, I suppose that such impressive numbers could still be true, but more importantly it now starts to actually run larger benchmarks with "a good speed-up". I would say that it's around the speed of Psyco, even a bit better in general, except that producing the assembler from the PyPy JIT still takes far too much time for now. Other random notes: * I would kill "logic programming". * As Maciej pointed out, having "high performance GC" was moved to the background for now as our existing GCs, although admittedly simple, perform well enough. A bientot, Armin.
So I've finally got around to lightly updating my talk; you can see my latest version at: http://people.canonical.com/~mwh/pypy-kiwipycon09.pdf Any comments you have would be appreciated but as I'm travelling to the conference tomorrow I'm not going to have very long to act on them :-) (all my own fault of course). Cheers, mwh 2009/9/30 Armin Rigo <arigo@tunes.org>:
Hi Michael,
Michael Hudson-Doyle wrote:
So I've finally got around to lightly updating my talk; you can see my latest version at:
http://people.canonical.com/~mwh/pypy-kiwipycon09.pdf
Any comments you have would be appreciated but as I'm travelling to the conference tomorrow I'm not going to have very long to act on them :-) (all my own fault of course).
I don't think that V8 implements a tracing JIT, but just a "standard" JIT with polymorphic inline caches and "hidden classes" (i.e., the equivalent of SELF's map or pypy sharing dicts). ciao, Anto
participants (4)
-
Antonio Cuni -
Armin Rigo -
Maciej Fijalkowski -
Michael Hudson-Doyle