The various caches in abc.py should be turned into weak sets so that the
caches don't keep the subclass objects alive forever.

Here's a working version of that idea, with a WeakSet implementation I
had lying around (but never really used.) It seems to work, and fixes
the refcount issues, but the WeakSet could do with some extra tests ;-)

On 8/30/07, Thomas Wouters <report@bugs.python.org> wrote:
>
> Thomas Wouters added the comment:
>
> Here's a working version of that idea, with a WeakSet implementation I
> had lying around (but never really used.) It seems to work, and fixes
> the refcount issues, but the WeakSet could do with some extra tests ;-)

I was torturing the WeakSet implementation, but didn't get very far:

Python 3.0x (py3k, Aug 30 2007, 09:27:35)
[GCC 4.0.3 (Ubuntu 4.0.3-1ubuntu5)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> from weakref import WeakSet as WS
[40407 refs]
>>> a = WS([1, 2, 3])
Fatal Python error: Cannot recover from stack overflow.
Aborted

Guido van Rossum schrieb:
> Guido van Rossum added the comment:
> 
> On 8/30/07, Thomas Wouters <report@bugs.python.org> wrote:
>>
>> Thomas Wouters added the comment:
>>
>> Here's a working version of that idea, with a WeakSet implementation I
>> had lying around (but never really used.) It seems to work, and fixes
>> the refcount issues, but the WeakSet could do with some extra tests ;-)
> 
> I was torturing the WeakSet implementation, but didn't get very far:
> 
> Python 3.0x (py3k, Aug 30 2007, 09:27:35)
> [GCC 4.0.3 (Ubuntu 4.0.3-1ubuntu5)] on linux2
> Type "help", "copyright", "credits" or "license" for more information.
>>>> from weakref import WeakSet as WS
> [40407 refs]
>>>> a = WS([1, 2, 3])
> Fatal Python error: Cannot recover from stack overflow.
> Aborted

The update() method can not be implemented in terms of converting the
argument to a WeakSet, since that leads to infinite recursion on creation
(as to why you get this fatal error, not a RuntimeError, no idea).

Also, the fact that the operator-versions of set operations only support
other sets as their second operands, while the method-versions support any
iterable, is not preserved.

Georg

The fatal error is raised if the stack overflows in the process of
handling RuntimeError. In certain cases, the C algorithms won't bail out
if a RuntimeError is raised, and just catch it. If that then causes
another stack overflow, Python gives up. One such case is an overflow
occuring during a comparison operation of a dictionary lookup, IIRC.

Georg, would you have time to submit an improved patch?

Attaching a new patch; this one passes regrtest -R:: of test_io without
apparent reference leaks.

Thanks Georg; please check it in!

Committed r58602.