Stub functions should only typecheck for None

[Wilfred]

This passes in mypy 0.521:

def foo(): # type: () -> int
    pass

mypy --py2 --strict demo.py does not show any errors. If a function body does not contain a return statement or a yield statement it should only have a type of None I believe.

[ilevkivskyi]

This is a special case in mypy. If a function have an empty body then it can have arbitrary return type. This is useful for testing, and is necessary for stubs. We have lots of tests that use this, so we can't disable this in non-stub files. Alternatively we can somehow detect whether we are running tests, and display an error otherwise.

[JukkaL]

Mypy should probably eventually enforce this outside stubs. One way would be to make this an option (which doesn't need to be shown in help, perhaps) and turn it on by default, but we'd turn it off for most tests. Detecting stubs is easy, so it won't be a problem. We don't need to actually detect tests if all test runners just explicitly toggle the option.

[ilevkivskyi]

@JukkaL

We don't need to actually detect tests if all test runners just explicitly toggle the option.

Yes, I was thinking about the same idea.

[ilevkivskyi]

See also #2350

[Wilfred]

Aha, I didn't realise that mypy treated stubs specially. My intent was to minimise my example, but I initially encountered this issue on a function that calls exit:

import sys

def fatal_error(message):
    # type: (str) -> dict
    print "Oh no! {}".format(message)
    sys.exit(1)

I was expecting mypy to give a type error here.

[JelleZijlstra]

That's expected: that function basically returns NoReturn, and NoReturn is a subtype of every type.

[gvanrossum]

I propose to change the title of this issue to "Get rid of exception for functions with just pass as body" -- or something better (but I can't think of something better). The current title is too wide and in general incorrect.

[Wilfred]

I've updated the issue title.

[Wilfred]

@JelleZijlstra the NoReturn behaviour was surprising to me. Could I open an issue for a --strict option that disables this behaviour, or is this working as intended?

[ilevkivskyi]

or is this working as intended?

This is intended. It is OK for a function to declare a return type wider than actual (NoReturn in this case).

[Wilfred]

@JelleZijlstra out of interest, what's the benefit of that? When is NoReturn being a subtype of everything useful?

[ilevkivskyi]

@Wilfred NoReturn is the bottom type, i.e. a subtype of every type, so I would put this vice versa, what's the benefit of complicating the type system by not doing that?