Narrow types after 'in' operator.

[ilevkivskyi]

Sometimes it is useful to narrow down a type after an in/not in comparison. For example:

class Option:
    ...
opt: Union[Option, str]
if opt in ('verbose', 'silent'):
    reveal_type(opt)  # we know this is 'str'
else:
    ...  # still a union here

presets: List[Option]
if opt not in presets:
    ...  # still a union
else:
    reveal_type(opt)  # we know this can be only 'Option'

This seems to be especially useful for --strict-optional, for example:

class Node:
    fullname: Optional[str]
n: Node

reverse_aliases: Dict[str, str]
if n.fullname in reverse_aliases:
    reveal_type(n.fullname)  # This should be 'str'

(related #2980)

A similar idea was discussed in #1749, but emitting an error on if 'a' in [1, 2, 3]: ... would require a larger change: emitting errors on unreachable code in binder.

[elazarg]

I think this can also fall out of #4063 (just as a note; of course a dedicated PR is more practical)

[gvanrossum]

Careful. For unconstrained x, after x in (1, 2), x may still be a float. (Or a fraction or decimal?)

…

On Oct 8, 2017 6:15 AM, "Elazar Gershuni" ***@***.***> wrote: I think this can also fall out of #4063 <#4063> (just as a note; of course a dedicated PR is more practical) — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#4071 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACwrMifVTAK8FQdpXnhnADA9IBWzQ_7jks5sqMr8gaJpZM4PxpHz> .

[ilevkivskyi]

@gvanrossum

Careful. For unconstrained x, after x in (1, 2), x may still be a float. (Or a fraction or decimal?)

OK, makes sense, we should not narrow down if we start from Any and should not use promotions (this is actually default in is_overlapping_types). I will update the PR now.

[ilevkivskyi]

@gvanrossum OK, I updated PR #4072
Now it does not narrow type from Any and does not use promotions:

x: Any
if x in [1, 2, 3]:
    reveal_type(x)  # Revealed type is 'Any'
y: float
if y in [1, 2, 3]:
    reveal_type(x)  # Revealed type is 'builtins.float'

It is however not easy to prohibit this:

z: object
if z in [1, 2, 3]:
    reveal_type(z)  # Revealed type is 'builtins.int'

This is still formally speaking unsafe, but I think it is a reasonable price to pay.