???? Hardware-accellerated python ????

[Alex Martelli]

"Armin Steinhoff" <Armin at Steinhoff_de> wrote in message
news:8ms9du$c3b at drn.newsguy.com...
    [snip]
> >> You are wrong ... tokens processed as machine code ist just faster.
    [snip]
> Right so far ... but the term 'general purpose CPU' has been changed since
we
> have the TRANSMETA CPUs.
>
> Could not a TRANSMETA CPU a candidat for implementing a  'special purpose
CPU'
> for processing interpreter tokens ??

??? The Transmeta Crusoe chips are just VLIW (an interesting approach to CPU
design that's been around for a few years but hasn't managed to score any
real
success so far) embellished by patented 'code-morphing' software that lets
them (allegedly) execute Pentium instruction streams pretty fast (claiming a
better power-consumption/performance ratio than Intel's and other Intel
competitors).  In other words, far from being "special-purpose", they're
quite
general-purpose CPU's with a (possibly) fast software approach to emulating
Intel's widespread architecture.

It is, of course, quite possible that, if you are running on a Crusoe that
is
mostly emulating a Pentium (no matter how well it emulates it), then, if you
also have a way to code directly to the Crusoe level, then your software can
get better performance than when compiled/assembled to Pentium machine
code.  As to whether Transmeta's patented technology would let you code a
better/faster interpreter/JITter than other well-known algorithms, I guess
that is possible -- I have not read their patents -- but I have my doubts.
The
problem they're addressing is interpreting/JITting an encoding that was
never
designed to be (the Pentium machine-code encoding) and it would be truly
a lucky accident if those presumably-clever algorithms happened to help on
the very different problem of optimally interpreting/JITting an encoding
that
WAS purposefully designed for best possible interpretation with classical
techniques.  In any case, optimal theoretical performance would be obtained
by hand-coding (in their native VLIW machine-code -- which may bear some
family-resemblance to horizontal-microcoding, I guess) whatever algorithms
work best for the bytecode/hardware combination.  Whether any performance
gains thus obtained would be worth the huge bother is a different issue.

In any case, by no stretch of the imagination could the resulting
interpreter
be called a 'special-purpose CPU'; I would hope that any marketeer trying to
come up with that particular spin would be laughed out of court by his peers
before he can make a laughing-stock of his company by going public with it.

Of course, judging from past performance, anything can happen if you let a
creative marketeer loose with some half-chewed concepts and an inexhaustible
appetite for appealing buzzwords.  But, at least between us techies, I hope
we
can avoid such excesses...!


Alex