Implicit relative

[MWR27]

Python 2 supports implicit relative imports inside packages. A bare import inside a package can resolve to a sibling module.

Example package:

pkg/
  __init__.py
  foo.py
  bar.py

# pkg/bar.py
import foo

In Python 2, this can resolve to:

pkg.foo

In Python 3, imports are absolute by default, so the same code resolves to:

foo

This means that the code can import a different top-level module or fail with ImportError.

For same environment of

# pkg/bar.py
import foo

Python 2 resolves to pkg.foo.
Python 3 resolves to top-level foo.

I modified import.c to fire warning only if

1. The import uses Python 2 implicit-relative semantics, not explicit-relative.
2. The importing module has package context (i.e. a parent package).
3. Runtime resolution actually selected a sibling submodule.

To ensure above, it checks that the importing module has a parent package, that there are no dots in the import name, and that the resolved name is equivalent to:
parent_package + "." + imported_name

For example, with import foo, there are no dots in foo and

pkg + "." + foo == pkg.foo

Therefore, the import depended on Python 2 implicit relative import behavior, so pygrate2 emits a warning.

[ltratt]

Rather than "..." + ".." f strings would probably be clearer?

[ltratt]

I think we can get to this line with is_from_form uninitialised?

[ltratt]

This looks mostly good, but I have a couple of small comments.

[ltratt]

@MWR27 I didn't notice the changes: if you can ping me when you need a review, that will alert me, thanks.

[ltratt]

Please squash (note: this doesn't mean you necessarily have to squash down to one commit).

[MWR27]

Please squash (note: this doesn't mean you necessarily have to squash down to one commit).

Sorry @ltratt, I've never squashed commits before. I somehow made more commits than there was before. I'll try again.

[ltratt]

Please let me know when you're done.

[MWR27]

@ltratt I'm ready for a review. I added the remaining tests for implicit relative imports and fixed an issue when catching warnings. I plan to fix the other failing test cases not related to imports in a separate pull request.

[ltratt]

I think this is ready: my suggestion (in this case) is that squashing down to one commit is probably the right thing to do.