[Python-Dev] Possible low-hanging optimization

[Guido van Rossum]

> In Boost.Python, extension class instance dictionaries are created
> only on-demand.  It's an obvious win for lots of wrapped C++ class
> instances, which will never have Python attributes tacked on.  This
> also seems like a potential big win for some kinds of Python
> instances, like those which use __slots__ but may also have a
> dictionary (IIRC that is a possible configuration).  The code to do
> this in Boost.Python is small and "just works" transparently, so I
> thought there might be some interest here for the Python core:
> 
> I've "C-ified" the code here, sort of on-the-fly, just to give an
> idea, but it's so simple that any errors can't obscure the intent
> /too/ much ;-)
> 
>       static PyObject* instance_get_dict(PyObject* op, void*)
>       {
>           bpl_instance* inst = (bpl_instance*)op;
>           if (inst->dict == 0)
>               inst->dict = PyDict_New();
>           Py_XINCREF(inst->dict);
>           return inst->dict;
>       }
>     
>       static int instance_set_dict(PyObject* op, PyObject* dict, void*)
>       {
>           (bpl_instance)* inst = (bpl_instance*)op;
>           Py_XDECREF(inst->dict);
>           Py_XINCREF(dict);
>           inst->dict = dict
>           return 0;
>       }
> 
>   static PyGetSetDef instance_getsets[] = { /* used as tp_getset */
>       {"__dict__",  instance_get_dict,  instance_set_dict, NULL},
>       {0}
>   };
> 
> Useful?

AFAICT, this is already done for new-style classes.  For classic
classes, it would be too much of a break with history.

--Guido van Rossum (home page: http://www.python.org/~guido/)