Conditional expression type inference behaves unexpectedly

[JukkaL]

The type of the conditional expression 1 if foo else 'x' is object, which is totally unexpected.

To be consistent with the rest of the language, the type should be one of these:

1. Union[int, str] (obvious)
2. Error, unless the type context is a supertype of both int and str. If one of the operands would have type None, the type will always be Optional[...] once mypy supports strict type checking of None values (None would be the only special case).

The rationale for case 2 is that mypy usually tries not to infer non-trivial union types, as these can result in confusing error messages far away from the source of the error. The error would be given if the only common non-union, non-Any supertype of the operands would be object. This is a somewhat arbitrary heuristic, but it would probably work well in most cases.

(These are also sometimes called ternary expressions. -- SEO Guido)

[JukkaL]

This was motivated by an example on the python-ideas mailing list.

[gvanrossum]

What exactly is the context? Assuming the following two examples would work, (2) sounds good:

• Using a type comment:

x = 'no' if a < 0 else sqrt(a)  # type: Union[str, float]

• Using a pre-declared type:

x = Undefined(Union[str, float])
x = 'no' if a < 0 else sqrt(a)

[JukkaL]

Both would work. Examples of type context (yeah, this should be documented better):

• In assignment x = expr, the type context of expr is the type of x (unless this assignment initializes x).
• In assignment x = expr # type: T, the type context of expr is T.
• In call f(expr), the type context of expr is the type of the first argument of f.
• In return expr, the type context of expr is the declared return type of the enclosing function.

[ilevkivskyi]

This is superseded by #3487