[Python-Dev] Nested scopes core dump

[Guido van Rossum]

> >   MH> The dict's resizing, it turns out.
> > 
> > So a hack to make the iteration safe would be to assign and element
> > and then delete it?
> 
> Yes.  This would be gross beyond belief though.  Particularly as the
> normal case is for freevars to be empty.
> 
> >   MH> I note that in PyDict_SetItem, the check to see if the dict
> >   MH> needs resizing occurs *before* it is known whether the key is
> >   MH> already in the dict.  But if this is the problem, how come we
> >   MH> haven't been bitten by this before?
> > 
> > It's probably unusual for a dictionary to be in this state when the
> > compiler decides to update the values.
> 
> What I meant was that there are bits and pieces of code in the Python
> core that blithely pass keys gotten from PyDict_Next into
> PyDict_SetItem.

Where?

> From what I've just learnt, I'd expect this to
> occasionally cause glitches of extreme confusing-ness.  Though on
> investigation, I don't think any of these bits of code are sensitive
> to getting keys out multiple times (which is what happens in this case
> - though you must be able to miss keys too).  Might cause the odd leak
> here and there.

I'd fix the dict implementation, except that that's tricky.

Checking for a dup key in PyDict_SetItem() before calling dictresize()
slows things down.  Checking in insertdict() is wrong because
dictresize() uses that!

Jeremy, is there a way that you could fix your code to work around
this?  Let's talk about this when you get into the office.

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