Although PR12004 fixes the problem, something a bit more fundamental may have changed. I have spent time investigating why this is happening and the problem is that in the line:


self.assertEqual(c.attr, 1)

Evil.__eq__ is invoked twice where before it was only invoked once. This does not happen all the time, and it seems to only happen under some combination of hashes in the class dictionary, being the extra __eq__ invocation a check for equality when looking up attr in the c object. When the test succeeds, the extra check is not done (and therefore the attribute is not deleted). I think this happens because some identity is used to check when doing the lookup (comparing the pointers) but I would like to understand exactly how the hash is influencing the different code paths to be sure that nothing important has changed recently.