[Python-Dev] Re: Code Generation Idea Was: Bytecode idea
Terry Reedy
tjreedy@udel.edu
Wed, 26 Feb 2003 13:06:44 -0500
"Skip Montanaro" <skip@pobox.com> wrote in message
news:15964.59840.342254.176499@montanaro.dyndns.org...
> For code that wants to cleanly cross the boundary between Python
with no
> boolean type and Python with a boolean type, you sometimes see
>
> True = 1==1
> False = 1==0
>
> You get True and False if it didn't exist before and have the added
benefit
> that if it does exist, it gets found in globals() or locals()
instead of in
> __builtins__ and has the right type.
This suggests a more radical optimization that might do more speedup
than all the byte-code fiddling currently proposed: automatically
localize, at definition time, builtins used within a function. This
would be like giving each function a customized implicit set of
default args: True=True, None=None, len=len, ...etc.
If a program alters __builtins__, def statement placement would
determine which version gets used.
A non-code-breaking alternative: add a localize() builtin that does
the fixup after the fact, something like staticmethod() and
classmethod(). If made recursively applicable to classes and modules
as well as functions, one could end the setup phase of a long running
program with localize(__import__(__name__)) to localize everything A
speedup of fraction x at cost of time t would pay if the program runs
longer than t/x. (Example: .05 reduction with 1 minute fixup pays
after 1/.05 = 20 minutes.)
I don't really know, but I imagine that psyco, for instance, would
benefit by having builtins localized and detectibly constant (by the
lack of any redefinition within the function). As it is now, neither
the compiler nor psyco can assume that True, for instance, is not
rebound between function invocations. Turning builtins into default
locals isolates a function much more than it now is.
Well, just some crazy thoughts...
Terry J. Reedy