[Python-Dev] Reordering opcodes (PEP 203 Augmented Assignment)

[Vladimir Marangozov]

Guido van Rossum wrote:
> 
> > Separating out LOAD_FAST from the switch shows a nice effect.
> > SET_LINENO is removed by -OO anyway, so there's really no
> > use in optimizing this one.
> 
> I tried this and found about three percent speed increase on pystone,
> for what that's worth.  This is with python -OO on Linux x86.  Note
> that removing the (now redundant) case from the switch seemed to make
> a small difference too.
>
> [patch deleted]

Some speedup confirmed on my Linux PIII too. But you saved a round-trip
around the ticker test. If this is acceptable, then let's do it for
the opcodes involving only stack operations (probably not the JUMPs)
_and_ which do not contain DECREFs which may trigger an external call.

Here's the picture on my machine (I get the same picture with -OO):

cvs           - the original ceval.c in CVS
load_fast     - ceval.c with your patch
top5          - ceval.c with my patch at SF moving 5 opcodes to the top
top5-loadfast - my patch and your patch


~/python/dev>./python-cvs Lib/test/pystone.py
Pystone(1.1) time for 120000 passes = 19.85
This machine benchmarks at 6045.34 pystones/second

~/python/dev>./python-load_fast Lib/test/pystone.py
Pystone(1.1) time for 120000 passes = 19.61
This machine benchmarks at 6119.33 pystones/second

~/python/dev>./python-top5 Lib/test/pystone.py
Pystone(1.1) time for 120000 passes = 18.87
This machine benchmarks at 6359.3 pystones/second

~/python/dev>./python-top5-load_fast Lib/test/pystone.py
Pystone(1.1) time for 120000 passes = 19.08
This machine benchmarks at 6289.31 pystones/second

Which shows, among others, that important cache effects are still here,
bacause "python-top5-load_fast" is slower than "python-top5" alone...

no-more-time-for-it-from-me-either'ly y'rs
-- 
       Vladimir MARANGOZOV          | Vladimir.Marangozov@inrialpes.fr
http://sirac.inrialpes.fr/~marangoz | tel:(+33-4)76615277 fax:76615252