[Python-Dev] Python is faster than C

[Dave Cole]

On Tue, 2004-04-06 at 00:36, Tom Emerson wrote:
> Moore, Paul writes:
> > Hmm. I'm torn. On the one hand, it's really cool that Python+Psyco can
> > out-perform C (another argument against the knee-jerk "recode bits in
> > C" reaction). But my gut feel is that not enough people use Psyco to
> > make this a good answer yet. On the third hand, if we recode chunks of
> > the stdlib in C, do we kill too many chances for Psyco to work its
> > magic?
> 
> If I understand correctly, the issue is that when stdlib functions are
> rewritten in C, they end up being faster but are still slower than
> Psyco optimized code because they are still polymorphic while the
> Psyco'ized generates multiple paths.
> 
> If that is the case then one could certainly hand-hack the C
> implementations to expand common code paths in an optimal way.

I have not looked at Psyco so could be operating from a misunderstanding
of what it does...

Would it make sense to build specialisation into the base Python
interpreter so that it builds specialised versions of bytecode for
functions.  I am assuming that there are only a few types that can be
targeted by specialisation (int, float).  Bytecodes then could be added
to the interpreter which operate on those native machine types rather
than the Python objects.

This has the advantage of allowing specialisation on all currently
supported platforms and does not preclude the further optimisation of
translation to machine code.

If specialisation was in the base interpreter then I suppose it should
possible for the compiler to know when a native machine type can be
compiled directly.  For example:

def foo():
    i = 0
    while i < 100:
        bar()
        i += 1

While noone writes loops like that, it does illustrate the point.  It is
clear that the local i is always an integer and is not used outside the
function.

- Dave

-- 
http://www.object-craft.com.au