Incorrect MRO inferred with multiple inheritance from module-qualified classes

[timmartin]

Steps to reproduce

File other.py:

class A:
    pass

class B:
    pass

Then run:

import astroid

class_node = astroid.extract_node("""
import other

class Child(other.A, other.B): #@
    pass
""")

print([cls.name for cls.name in class_node.mro()])

Current behavior

['Child', 'A', 'object']

Expected behavior

['Child', 'A', 'B', 'object']

python -c "from astroid import __pkginfo__; print(__pkginfo__.version)" output

2.5.0

[timmartin]

I believe this change will fix this, I just need to write a proper PR.

I think the problem is that when we run multiple inferences in _inferred_bases, the context is reused which means the second and subsequent inferences have bad data. Cloning the context before running the inference seems to address this.

[hippo91]

@timmartin thanks for the report. I can reproduce it.
One interesting thing is:

import astroid

class_node = astroid.extract_node("""
from other import A, B

class Child(A, B): #@
    pass
""")

print([cls.name for cls in class_node.mro()])

prints: ['Child', 'A', 'B', 'object']

[timmartin]

Yes, that's correct. I believe this is because in the first case, there is non-trivial inference that must be applied to each expression (other.A, other.B) to determine its type, and sharing the context between the two infer calls causes the second to fail. In the second case, the shared context doesn't affect the second infer call, either because the two expressions are so trivial or because there's no overlap between the names in the two expressions. I'm mostly guessing this from context, I don't yet properly understand how the inference works.