"Andrew M. Kuchling" wrote:
Christian Tismer writes:
Instead, If I can manage to create something like static binding snapshots, then I could resolve many of the lookups and internal method indirections, for time critical applications.
The usual assumption is that the lookups are what takes time. Now, are we sure of that assumption? I'd expect the lookup code to be like:
1) Get hash of name 2) Retrieve object from dictionary 3) Do something with the object.
... Right, but when you become more restrictive, you can do lots, lots more. If I also allow to fix the type of an object, I can go under the hood of the Python API, so you could add some points 4) 5) 6) here. But I shouldn't have talked before I can show something. And at the time, I don't restrict myself to write clean code, but try very machine specific things which I don't believe should go into Python at all.
I also wonder about the costs of all the Py_BuildValue and Py_ParseTuple calls going on under the hood. A performance improvement project would definitely be a good idea for 1.6, and a good sub-topic for python-dev.
I wasn't sure, so I first wrote a module which does statistics on that, and I found a lot of these calls. Some are even involved in ceval's code, where things like "(OOOO)" is parsed all over again and again. Now I think this is ok in most cases. But in some places these are very bad: If your builtin function soes so very few work. Take len(): This uses ParseTuple, which calls a function to find out its object, then calls a function to get the object, then takes the len, and builds the result. But since everything has a length, checking for no NULL pointer and grabbing the len produces even shorter code. This applies to many places as well. But right now, I don't change the core, but build hacky little extension modules which squeak such things out at runtime. No big deal yet, but maybe it could pay off to introduce format strings which make them more suitableto be used in macro substitutions. sorry if I was too elaborate, this is so much fun :-) ciao - chris -- Christian Tismer :^) mailto:tismer@appliedbiometrics.com Applied Biometrics GmbH : Have a break! Take a ride on Python's Kaiserin-Augusta-Allee 101 : *Starship* http://starship.python.net 10553 Berlin : PGP key -> http://wwwkeys.pgp.net PGP Fingerprint E182 71C7 1A9D 66E9 9D15 D3CC D4D7 93E2 1FAE F6DF we're tired of banana software - shipped green, ripens at home