On May 6, 2009, at 8:56 AM, Carl Witty wrote: > On Wed, May 6, 2009 at 6:23 AM, Stefan Behnel <[email protected]> > wrote: >> What we tend to call PY_NEW here is a direct call to a type's >> tp_new(), >> which handles both the allocation and the basic initialisation of an >> instance. Cython could do a tiny bit better here, since we know >> exactly >> what function a cdef class has in its tp_new slot. However, that >> function >> will always recursively run into the base type's tp_new and at the >> end of >> the chain into the one of Python's 'type' type, and there is not >> much we >> can do to improve this situation in general. > > Well, you could optimize tp_new. You could have it call the base > type's tp_new directly, instead of through a vtable (at least if the > base type is a Cython type);
I already implemented this optimization, though only for cdef classes within the same module. (If they're small enough, they may get even inlined by the C compiler.) I'm not sure how possible this would be to do across modules, but one difficulty is that the contents of tp_new can change based on the .pyx file, not just the .pxd file. That being said, one does have control over allocation via the tp_alloc method. Allocating an array of objects as a big chuck is messy though--how would one handle deallocating them one by one? Or would one try to reference-count them as a group? I think it could be done, but it would be a non-trivial interface. - Robert _______________________________________________ Cython-dev mailing list [email protected] http://codespeak.net/mailman/listinfo/cython-dev
