Don't leak objects in _pickle_27.c load_[short_]binbytes.

[jamadden]

Fixes #36.

Manually verified according to zopefoundation/ZODB#203. Here's the current memory usage of the test case described there:

I'm not sure if/how to do some sort of automated regression test for this...

[Krilivye]

Hi looking at your comment I think it's PyTuple_SET_ITEM who keeps the ref even if it fails !

[Krilivye]

C style review

[azmeuk]

@jamadden As a regression test, what about a simple memory consumption comparison using psutil, like the one in the ZODB bug report?

[jamadden]

I have three concerns about a simple memory consumption comparison using psutil:

1. It's an added binary dependency. That dependency cannot be installed everywhere. For example, it cannot be installed on PyPy on Windows (Appveyor). (We don't currently test that combo, but we probably should.)
2. There are platform differences in both the fields available from psutil (and their meaning) and the Python memory allocators. That makes writing a truly portable test that way difficult. (I'm assuming; it's not like I have [or want to] boot up various virtual machines to try...)
3. It would necessarily be slow. It takes a large number of iterations (approximately 10,000) on my machine to see the memory stabilize, and that's in a process that's only doing the test.

gevent uses reference counting and objgraph to check for leaks, and that code is finicky. But here, objgraph didn't reliably find the dropped tuple, not with show_growth and not with show_most_common_types as in the example documentation. It's not clear to me why. (I probably did something wrong. Or maybe its because tuples containing only strings aren't necessarily tracked by the GC and hence are invisible to objgraph? I'm guessing the cases it could find---very early in the run---had to do with the tuple freelist in some fashion. @mgedmin ?)

Because there are no existing reference counting tests in this codebase, and because of the complexity outlined above, and because Python 2.7 is a "legacy" platform at this point, I'm inclined to ask for a pass on an added regression test. @icemac ?

[mgedmin]

maybe its because tuples containing only strings aren't necessarily tracked by the GC and hence are invisible to objgraph

That is, unfortunately, very true.

[azmeuk]

Because there are no existing reference counting tests in this codebase, and because of the complexity outlined above, and because Python 2.7 is a "legacy" platform at this point, I'm inclined to ask for a pass on an added regression test.

That makes sense.

[icemac]

@jamadden I'm fine with fixing this test without adding a proper test. The verification shown in this PR is enough as the change log links to it.

[jamadden]

@icemac @mgedmin Thanks! I'm going to take that as approval to merge then.

Could we get a PyPI release (or access to do so), please?

[mgedmin]

@jamadden you're now a PyPI owner of zodbpickle.

[jamadden]

@mgedmin Thanks! The new pypi.org (warehouse) sends out notification emails when that happens, so that was nice.