Mailman 3 March 2007 - pypy-dev

Google Summer of Code
by Armin Rigo March 18, 2007

March 18, 2007

Hi all, A few PyPy developers have signed up as mentors for the Google Summer of Code. Students interested in a summer project about PyPy (or any Python-related project) should have a look at: http://wiki.python.org/moin/SummerOfCode A bientot, Armin

8 9

daily #pypy-sync for 1.0 5pm
by holger krekel March 18, 2007

March 18, 2007

Hi folks, As you hopefully know by now, we are in the final run up to our long-awaited 1.0 release. Therefore we suggest to do daily #pypy-sync "meetings" until we have the release out and suggest as time frame 5pm-5:30pm (GMT+1) starting from 19th March (Monday). The meetings should focus on tracking, discussing and assigning pending issues, to be found and updated here: extradoc/planning/1.0/TODO.txt Please note that the dense timeplan at the top of the file aims at a release at … [View More]

1 0

Re: [pypy-dev] PyPy + LLVM
by Chris Lattner March 18, 2007

March 18, 2007

On Wed, 14 Mar 2007, Armin Rigo wrote: > Hi Chris, Hey armin, I'm cc'ing pypy-dev, because this is suitable for a broader audience. I'm sorry for the delay in my reply, I have been travelling recently. > After reading your slides from yesterday's presentation, it occurred to > me that you might have a wrong idea about what PyPy is. I'm Sorry about that. The third section of my talk was only loosely based on pypy. Some of it was "inspired" by pypy, some of it was "unification" … [View More]ideas that I have been thinking for a while, some is based on my own experience with static analysis, and some is just wishful thinking. I'm sorry if it came across as "here is what pypy is doing", that is now what I intended. > double-guessing here, so of course I may be completely off - please bear > with me. What PyPy is definitely not, is a Python compiler. In order > to get to a common ground of discussion, may I suggest the following > references: Right. If I had to do a layman's summary, I would say that pypy provides infrastructure for building interpreters in [r]python. This infrastructure makes it much easier than starting from scratch, e.g. by providing reusable components for language runtimes (like GC's). My talk (available here http://llvm.org/pubs/2007-03-12-BossaLLVMIntro.html ) clearly doesn't describe pypy as it exists today, but I think it describes a place that pypy could get to if the community desired it (In other words, I think the strengths of pypy and of its community both play well into this). The members of the pypy project clearly have experience with type inference, clearly understand dynamic language issues, and clearly understand that the semantics of these languages differ widely, but there are also a lot of commonality. If the pypy community isn't interested, I will approach others, eventually falling back to doing it directly in LLVM if needed. For the record, I have read Brett Cannon's thesis. I don't think his results are applicable for a number of reasons. First, his work was in the context of an interpreter that used type information for optimization. The approach I'm describing uses type information to give substantially more information to a static compiler. The result of this is that the codegen would be dramatically more efficient, and without the overhead of an interpreter loop, this is far more significant than in his evaluation. Also, significantly more type information would be available to the type propagator if you used more aggressive techniques than he did. To me, the much more significant issue with python (and ruby, and others) for integer operations is that they automatically promote to big integers on demand. In practice, this means that the code generated by a static compiler would be optimized for the small case, but then would fallback gracefully in the large case. If you're familiar with X86 asm, I'd expect code for a series of adds to look like this: add EAX, EBX jo recover_1 add EAX, 17 jo recover_2 add EAX, ECX jo recover_3 The idea of this is that (in the common case where the app deals with small integers) you have *exteremely* fast codegen with easily predicatable branches. The recovery code could, for example, package up the registers into real "integer" objects, and then callback into the interpreter to handle the hard case (note that this is very similar to the fast paths in the standard python interpreter, which eventually falls back to calling PyNumber_Add). Floating point code, if type inference is successful, does not need this sort of recovery code (unlike integer code). Other languages (like Javascript) don't have this issue, but they have others (e.g. there are no integers :), only floats. Of course I'm biased, but I think that this sort of code could easily and naturally be generated and optimized by LLVM, if you chose to target it. This would let LLVM do the range analysis required to eliminate the branches (when possible), handle other low level optimization, codegen, etc. You would get extremely high performance code and a very retargettable compiler. Has python even been executed with a sustained performance of two instructions per add? :) -Chris -- http://nondot.org/sabre/ http://llvm.org/ [View Less]

2 1

Re: [pypy-dev] page needs update
by Michael Hudson March 16, 2007

March 16, 2007

Meant to reply to all, sorry! On 16/03/07, Michael Hudson <micahel(a)gmail.com> wrote: > On 16/03/07, Laura Creighton <lac(a)openend.se> wrote: > > http://codespeak.net/pypy/dist/pypy/doc/dev_method.html > > add post limerick sprints and sprint participants. Having missed > > most of these, I cannot do it myself, sorry. > > All the information is on: > > http://codespeak.net/pypy/dist/pypy/doc/sprint-reports.html > > though the list of … [View More]

1 0

page needs update
by Laura Creighton March 16, 2007

March 16, 2007

http://codespeak.net/pypy/dist/pypy/doc/dev_method.html add post limerick sprints and sprint participants. Having missed most of these, I cannot do it myself, sorry.

1 0

exerpt from mail from John Gilmore. How close are we to doing this?
by Laura Creighton March 14, 2007

March 14, 2007

We're discussing OLPC. <Begin exerpt> > The PyPy project just got its first OLPC. (We're doing optimisation > work for them.) What kind of optimization? One thing I've been looking at for them recently is about dirty pages. A dirty page is a piece of virtual memory that has been written to since the last time it was loaded from mass storage (flash in their case). Since they have no disk drive to swap to, every dirty page will stay resident in memory forever, until its … [View More]

2 1

proofreading
by Laura Creighton March 14, 2007

March 14, 2007

I'm available to do 'check-the-English'ing. What I need is a list of files in the order that you would like me to read them, preferably when they are done with the massive-update-for-content phase. I also need to know how to check out a pypy src tree that has the papers - what I have now is missing that part. Then -- do you just want me to fix as I go, or do you want me to mark suggested changes? When I am not sure of the sense of the thing, should that happen, I will just mark up the … [View More]

2 1

Re: [pypy-dev] [pypy-svn] r40417 - in pypy/dist/pypy/module/cclp: . constraint
by Aurélien Campéas March 13, 2007

March 13, 2007

On Tue, Mar 13, 2007 at 10:48:16AM +0100, ale(a)codespeak.net wrote: > Author: ale > Date: Tue Mar 13 10:48:15 2007 > New Revision: 40417 > > Modified: > pypy/dist/pypy/module/cclp/constraint/constraint.py > pypy/dist/pypy/module/cclp/constraint/domain.py > pypy/dist/pypy/module/cclp/constraint/variable.py > pypy/dist/pypy/module/cclp/variable.py > Log: > Fixes for not demoting methods - still not translatable with --faasen Thanks for the fixes. … [View More]

1 0

GC_REDIRECT_TO_LOCALS breaks Win32
by Scott Dial March 13, 2007

March 13, 2007

I spoke to several people on IRC, but I didn't want to let this fall by the wayside. On revision 39896, mwh changed the pre_pre_gc_code generated for all platforms and made the remark in his log entry, "someone should do a windows build, i guess..." So, that someone would be me. My test runner has been broken all week.. I'll save the commentary on the wisdom of such a checkin. If someone really needs an explanation as to why GC_REDIRECT_TO_LOCALS breaks everything on Win32 as it is a … [View More]

1 0

sprint pictures
by Carl Friedrich Bolz March 9, 2007

March 9, 2007

Hi all! Some of the pictures I took during the Trillke sprint can be found here: http://codespeak.net/~cfbolz/hildesheim3-sprint-pictures/ Cheers, Carl Friedrich

3 2