Inference for objects with C accelerators and pure-Python equivalents typically returns Uninferable

[pylint-bot]

Originally reported by: BitBucket: ceridwenv, GitHub: ceridwenv

---

For instance, look at lazy_object_proxy.

>>> m = parse('import lazy_object_proxy; lazy_object_proxy.Proxy(lambda: None)')
>>> proxy_call = m.body[1].value
>>> proxy_call
<Call l.1 at 0x7fd5f48f4fd0>
>>> next(proxy_call.infer())
Uninferable

This happens because in lazy_object_proxy's __init__.py, the C accelerator is tried first. This is a typical pattern.

try:
    from .cext import Proxy
    from .cext import identity
except ImportError:
    from .slots import Proxy

Then, the module importing subsystem ends up building an empty module for lazy_object_proxy.cext.

(Pdb) self.astroid_cache['lazy_object_proxy.cext'].print_tree()
Module(
   name='lazy_object_proxy.cext',
   doc=None,
   package=False,
   pure_python=True,
   source_code=b'',
   source_file='<?>',
   body=[])

This happens even on PyPy and Jython, so the presence of any C accelerator will break inference for the pure Python equivalent.

---

• Bitbucket: https://bitbucket.org/logilab/astroid/issue/263

[pylint-bot]

Original comment by Claudiu Popa (BitBucket: PCManticore, GitHub: PCManticore):

---

It might be worth to have a fallback on the pure Python import, if it is available. But since we don't know apriori which import is for a C extension, this could happen whenever a TryExcept is met with two imports on each branch.