Pyston Python JIT Talk | Hacker News Reader

Pyston Python JIT Talk

slideshare.net

74 pointsluu11 comments

Similar effort by Microsoft: https://github.com/Microsoft/Pyjion

There is a similarity, but there is also a difference. The general difference (among PyPy/Pyjion/Pyston) is "what to do with CPython C API". As far as I understand:

* Pyjion says "we gonna add a JIT API to CPython" - so no touching the C API at all (or the object model)

* Pyston says "we gonna implement our own object model supporting CPython C API", but they seem to have copied large portions of CPython. This is a semi-rigid approach compared to Pyjion, because it's not runtime-swappable

* PyPy says "we gonna implement a fake layer on top of our own object model" which means that while C API is feasible (and mostly implemented), it's gonna be always a compatibility layer with slower approach, but the underlaying data types/GC can change at will

Hope that helps

There are also a few compilers out there -- I've used: http://nuitka.net/ a bit.

I kinda wish we could get more wood behind a single arrow -- like I understand they have 80k lines in C extensions, but that has to be something like 8k lines if rewritten in Python -- and then that gives you access to pypy and now you can contribute to making that better instead of doing a new thing.

Based on the slides discussing Pypy performance, it looks to me that they're not confident that rewriting those 80k lines of C would provide sufficient performance. It might, but if you're not sure that's a lot of risk to take on.

>2.7

Welp, I guess I'm one of the complainers. Is there something in particular that prevents creating a JIT for py3?

It's pretty well documented Dropbox has a very large 2.7 codebase.

The talk is from Dec 2015, has already been on hacker news I believe (but can't find the discussion)

I'm curious what the best performance gain they think they can get by sticking to their current approach (idealistically).

They think (from what I understood) that the limit is "above pypy performance"

Yet another sibling in python family :)

Reply on news.ycombinator.com