Cython mishandles PEP-3135 __class__ cell in methods

[marginalhours]

Hello, I am having trouble with some inconsistent behaviour round the value of __class__ in Cythonized code. This is on Python 3.6.7 with Cython 0.29.1.

Briefly, in uncythonized code __class__ inside a method on a class refers to the class object, whereas once cythonized, it seems to refer to <class 'module'>. This behaviour is a bit unintuitive, and has caused a few issues in a codebase I'm working on, but it's possible I have misunderstood the language semantics!

I have tried to search for reasons / documentation why this is the case, but I couldn't turn up anything. I apologise if there is something really obvious I'm missing here; let me know if so and I'll close the issue at once.

Example:

minimal.py:

class Something:
    def method(self):
        print(__class__)

(in an interpreter, uncythonized):

from minimal import Something
Something().method()
# prints <class 'minimal.Something'>

(after running cythonize minimal.py -3):

from minimal import Something
Something().method()
# prints <class 'module'>   

I believe there is an obvious workaround -- in instance methods, self.__class__ should still have the correct value. But I was curious if this is intended behaviour, and if so, is there something in the docs about it?

[scoder]

I think we handle super() separately but not __class__ (as defined by PEP-3135).

PR welcome, it's probably quite easy. See TransformBuiltinMethods._inject_super() in ParseTreeTransforms.py. The NameNode with name __class__ should be replaced in methods y a ClassCellNode. Tests can go into py3k_super.pyx, although a new test file like pep3135_class_cell.py that also runs uncompiled in Python would also be nice (add a #tag: pure3.0 comment at the top of the file to prevent it from running in Python 2).