[IronPython] blocker: objects deleted when they shouldn't be

[William Reade]

Hi all

I'm about to file this on Codeplex, but I think I've tracked down a 
subtle and nasty bug, which is blocking Ironclad with numpy (it fatally 
breaks matrices and masked arrays), and I feel compelled to shout loudly 
about it to ensure it doesn't get lost in the noise of the impending 2.0 
release.

In short, the following code:
-------------------------------------------------------------
from System import GC

class C(object):
    _instance = None
    def __new__(cls):
        if C._instance is None:
            C._instance = object.__new__(cls)
        return C._instance
   
    def __del__(self):
        print 'deleting object with id', id(self)

c = C()
print 'created a C with id', id(c)
d = C()
print 'created another C with id', id(d)

for _ in range(3):
    GC.Collect()
    GC.WaitForPendingFinalizers()

print id(c), id(d)
print c
print d
-------------------------------------------------------------

...produces the following output:
-------------------------------------------------------------
C:\dev\ironclad\build>ipy deltest.py
created a C with id 43
created another C with id 43
deleting object with id 43
43 43
<C object at 0x000000000000002B>
<C object at 0x000000000000002B>
-------------------------------------------------------------

...and I don't think that the 'deleting...' message should be printed at 
that point.

My suspicion is that c's 'magic_finalization_wossname'* slot is being 
overwritten with a new instance after __new__ returns, leaving the 
original 'magic_finalization_wossname' unreferenced; in the fullness of 
time, that gets finalized and calls c.__del__, even though c itself 
still exists.

Cheers
William

* This is probably not the actual member name, but I couldn't find the 
relevant code today, even though I remember reading it in the past... 
hopefully the implementers know what I'm trying to refer to, even if 
nobody else does.