per-function jit compiler

Luis M. González luismgz at gmail.com
Tue Apr 6 12:23:19 CEST 2010


On 6 abr, 03:40, Chris Rebert <c... at rebertia.com> wrote:
> 2010/4/5 Luis M. González <luis... at gmail.com>:
>
>
>
>
>
> > This post gave me an idea:http://groups.google.com/group/comp.lang.python/msg/5d75080707104b76
>
> > What if I write a simple decorator to figure out the types of every
> > function, and then we use it as a base for a simple method-jit
> > compiler for python?
>
> > example:
>
> > def typer(f):
> >        def wrap(*args):
> >            a = f.func_code.co_varnames
> >            b = [type(i) for i in args]
> >            return dict(zip(a,b))
> >        return wrap
>
> > @typer
> > def spam(a, b, c=3, d=4):
> >        pass
>
> >>>> spam(8,'hello',9.9, 10)
> > {'a': <type 'int'>, 'c': <type 'float'>, 'b': <type 'str'>, 'd':<type
> > 'int'>}
>
> > So by using this information, we record all the argument types used
> > the first time each function/method is executed, and then we generate
> > optimized code for them.
> > >From this point on, a guard should check if all arguments remain the
> > same and, if so, the optimized code is run.
> > Otherwise, just fall back to the interpreter.
>
> > He! I have no idea how to implement it...
>
> Guido's been lending out his time machine again:http://psyco.sourceforge.net/introduction.html
>
> Cheers,
> Chris
> --http://blog.rebertia.com

Well, psyco is a complex beast. Actually it's a Just-In-Time
Specializer, not really a compiler at all (Guido's machine told me
so).
It's been superseded by pypy, which is a way more complex beast, and
it's a tracing compiler (and many things more).
What I mean is, why not using python's simple introspection
capabilities (instead of partial evaluation, type inference or other
complex mechanism) to gather type information?
I can imagine that this, coupled with pyrex, cython or shedskin could
give some results perhaps... by selectively compiling some function or
all of them.

But probably someone will throw Guido's time machine again at me. If
it was that easier, someone would have done it already...
I just wonder why this simple approach is not feasible.

Luis



More information about the Python-list mailing list