[Python-Dev] making weakref.ref objects subclassable

[Fred L. Drake, Jr.]

On Thursday 01 July 2004 10:43 am, Michael Hudson wrote:
 > Would "why?" be an apposite question?  I'm sure there is a reason, but
 > I haven't seen it yet :-)

Absolutely!  I hinted at the reason in my original email, but it was not made 
explicit that it was the reason.

The change allows the WeakValueDictionary to be implemented in a much more 
memory-efficient manner.  The current implementation causes each entry to 
create and retain the following objects in addition to the actual key:

- a weakref to the value, with a callback

- a function object used as the callback; a new function is created for every 
entry, since it contains a reference to the key in a default parameter value

- a weakref to the WeakValueDictionary itself, to avoid creating a cycle.

The last object could be removed without any change to the C implementation; 
all that's really needed is a single weakref to the WeakValueDictionary that 
can be shared by all such callbacks.

Using a subclass of the weakref.ref type (KeyedRef in the patch), the key can 
be stored as an attribute of the weakref, allowing a single object to be 
created and retained as overhead (with the allocation for that object being 
at most 8 bytes larger than a stock weakref).


  -Fred

-- 
Fred L. Drake, Jr.  <fdrake at acm.org>