[Python-Dev] [Python-checkins] Python Regression Test Failures refleak (1)

[Tim Peters]

[Thomas Wouters]
> I did the same narrowing-down last week, and submitted a patch to add
> cycle-GC support to itertools.tee . It really needs it.

I agree.

> Come to think of it, now that I remember how to properly do GC, I think
> the patch cuts some corners, but it solved the problem.

You mean because it didn't supply tp_clear?  That's a funny one.  Some
people take pride in not supplying tp_clear unless it's absolutely
necessary.  For example, despite that tuples can be in cycles, the
tuple type doesn't supply a tp_clear.  This is "because" it's possible
to prove that any cycle involving tuples must involve a non-tuple
gc'ed type too, and that clearing the latter is always sufficient to
break the cycles (which is all tp_clear _needs_ to do:  we just need
that the aggregate of all tp_clear slots that are implemented suffice
to break all possible cycles).

I never saw a point to that cleverness, though.  It makes gc more
obscure, and I'm not sure what it buys.  Maybe the (typically teensy)
bit of code needed to implement a tp_clear slot?  That's all the
(dubious) benefit I can think of.

> Raymond is on it, anyway:
>
>  http://python.org/sf/1444398

You found it, you fix it :-)