> For the second part, I don't see any undefined behaviour here.  Can you
> explain where the undefined behaviour comes from?  And are you sure that
> this is really just an optimization?  It looks as though it might be
> deliberately there to make sure that the conversion still produces a
> Python 0 even on systems where the NULL pointer *doesn't* give 0 when
> converted to an integer.

If on some platform (uintptr_t)NULL != 0, then some other address can be 
reflected to 0. In shuch case PyLong_FromVoidPtr() returns zero integer for 
both NULL and this non-NULL addresses. Of course this is a hypothetic 
situation.

What about (Py_uintptr_t)p - (Py_uintptr_t)(void *)NULL? Then we should change 
PyLong_AsVoidPtr() too.