On Sun, 24 Mar 2002 codepage@... wrote:
> My codes run 10x regular python on FreeBSD.
Oh yes, I forgot to mention that I discovered that the Psyco-emitted
code was quite *slower* with debugging turned *off*, which I tracked to be
caused by poor code locality (numerous long-distance jumps), which let me
suspect the heap fragmentation problems. (In debugging mode Psyco would
allocate less memory at a time, thereby reducing fragmentation.)
So another side-effect of the allocator rewrite is a speed-up of up to
Still, the psyco.jit() does not work as well as psyco.bind(). I guess we
should consider a more serious selection algorithm for the JIT-rebinder.
Maybe interfacing with the high-performance profiler, hotshot, of Python
2.2, and rebinding only those functions which are worth the trouble. Does
anyone volunteer to look more deeply into it?