Add Python 3.6 support

[mgedmin]

Closes #24.

[mgedmin]

I'm seeing a new build warning on Python 3.6 (only):

src/zodbpickle/_pickle_33.c: In function ‘Pdata_grow’:
src/zodbpickle/_pickle_33.c:221:23: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
     if (new_allocated > (PY_SSIZE_T_MAX / sizeof(PyObject *)))
                       ^

[jamadden]

I'm seeing a new build warning on Python 3.6 (only)

Interesting. I think that's highly platform dependent, although I don't really understand why it started showing up. new_allocated is defined as Py_ssize_t. The definition of PY_SSIZE_T_MAX is the same in 2.7:

cpython [2.7|…1]$ ack 'define PY_SSIZE_T_MAX'
Include/pyport.h
202:#define PY_SSIZE_T_MAX ((Py_ssize_t)(((size_t)-1)>>1))

as it is in 3.6:

cpython [3.6|…1]$ ack 'define PY_SSIZE_T_MAX'
Include/pyport.h
109:#define PY_SSIZE_T_MAX ((Py_ssize_t)(((size_t)-1)>>1))

AFAICS, Py_ssize_t itself is the same in 2.7 as it is in 3.6, at least on any platforms that have ssize_t defined. If that isn't defined, then it changes: on both versions, the fallback is to make it the same as Py_intptr_t, but the value of that has changed. In 3.6, it's unconditionally intptr_t, but 2.7 supports platforms that don't have that defined, so it could wind up as just int. In any case, those should all be signed.

Now, sizeof is always defined to return an unsigned integer (technically an unsigned integer constant). I forget the rules for C type conversions between unsigned and signed integers but I think promotion is generally to unsigned?

Does this warning go away if you cast the right hand expression to Py_ssize_t?

[jamadden]

FWIW, the 3.6 pickle module does a similar cast. Although it has also been changed to just use size_t instead of Py_ssize_t back in 3.5 because of this very same issue.

So it looks like what may have changed was the addition of more warning flags to the compiler by distutils.