Hm... I'm not crazy about this approach. It seems to be type-checking the argument expression (e.g. 'x' in d['x']) at the point where it's supposed to just be looking for the __getitem__ function. I feel that maybe these checks ought to be done in check_op_local()?

Is the concern that performing the check in the more-widely-used analyze_member_access() might be too general? I can certainly move the call up to check_op_local(), although I believe the semantics will be equivalent.

Moving the check up to check_op_local() may make it more difficult to recognize direct invocations like point.__getitem__('x') which we may also want to consider, although the current diff doesn't presently recognize that style of invocation.

Supporting point.__getitem__('x') doesn't seem important to me -- I can't find any examples of it in a large internal codebase. For consistency, visit_index_expr_helper seems like the right place for the __getitem__ support. We also special case tuples there, and TypedDict is sufficiently similar.

I've moved __getitem__ recognition to visit_index_expr_helper() as suggested, and added __setitem__ recognition. This PR now covers the entirety of #2486 .

Squished and rebased to the tip of master.

Got all the feedback integrated except for one item:

Add test case for both 'foo' and u'foo' key names in Python 2 mode.

When writing that test case, it crashes mypy when run inside the test runner but not when running mypy normally. Here's the problem: If you run py.test -k TypedDict there is a crash in semanal.py at:

    def normalize_type_alias(self, node: SymbolTableNode,
                             ctx: Context) -> SymbolTableNode:
        if node.fullname in type_aliases:
            # Node refers to an aliased type such as typing.List; normalize.
            node = self.lookup_qualified(type_aliases[node.fullname], ctx)  # <-- HERE
        if node.fullname == 'typing.DefaultDict':  # <-- THERE
            self.add_module_symbol('collections', '__mypy_collections__', False, ctx)
            node = self.lookup_qualified('__mypy_collections__.defaultdict', ctx)
        return node

At the line marked HERE, lookup_qualified is called with 'builtins.dict' which fails, returning a None node. That None node crashes the line marked THERE with a null pointer exception.

How is it that lookup_qualified('__builtins__.dict') could fail? Maybe there's something different in Python 2.7 mode?

Here is the test in question:

[case testCanGetItemOfTypedDictWithValidBytesOrUnicodeLiteralKey]
# flags: --python-version 2.7
(...)
[builtins fixtures/dict.pyi]

• If I run it as-is (in Python 2.7 mode), I get the missing __builtins__.dict problem.
• If I remove the # flags: --python-version 2.7 line then the test succeeds in Python 3.x mode.
• If I remove the # flags: --python-version 2.7 line and the [builtins fixtures/dict.pyi] fixture then I get the missing __builtins__.dict problem again.

The preceding behavior suggests that the [builtins fixtures/dict.pyi] line is being ignored in Python 2.7 mode or perhaps a different dict.pyi fixture is being used. Does this sound plausible?

I will keep investigating...

Ha! Apparently I need to say [builtins_py2 fixtures/dict.pyi] rather than [builtins fixtures/dict.pyi] when in Python 2.x mode.

@JukkaL , I've integrated all the requested feedback. Ready for another look.

Looks good now! Thanks for the updates!