[Python-Dev] Psyco update

[Armin Rigo]

Hello everybody,

Just a quick update on the Python Specializing Compiler.

I started to work on a quite different prototype one month ago. This is
all very experimental, so I will delay the technical explanations for some
more time. The motivations are as follows: the previous implementation
loosed WAY too much time in compilation proper: around 0.2 seconds for a
10-15 lines function! I am not sure that writing Psyco in very optimized C
would have been enough to be of much help in anything but the most
particular cases.

Because of this I am re-thinking the way it could work. While I did not
abandon the base ideas, I switched to a slightly different approach which,
I think, could produce slightly less efficient machine code but incredibly
more quickly. This work is based on research articles describing
techniques to produce code at run-time [XXX references]. With the
execution of only 4 to 6 processor instructions, they output one
instruction of the dynamic code. Compare it with a typical C compiler,
which takes thousands times more clock cycles to run than the number of
instructions it emits! Or course, I do not question that it outputs more
optimized code than what can be done with 4 to 6 instructions; however,
this has the great advantage that even if some dynamically produced code
is only used a few times, even only once, the overhead of having produced
it in the first place is very low -- maybe even lower than the time that
would have been required to load it from a disk file if it had been
precompiled, and in the worst cases similar to the current Python VM
interpreter.


A bientot,

Armin.