inflexibility with diverging functions

[thestinger]

The following code compiles, since the return type is inferred to be Option<()>:

fn main() {
    Some(5).map(|_| ());
}

However, this will not compile because rustc cannot infer a type:

fn main() {
    Some(5).map(|_| fail!());
}

If we used something like Void instead of !, the return type would obviously be Option<Void> here. However, since ! isn't allowed anywhere but a function signature this is a pain.

[thestinger]

#7680 was rejected, but this presents a good reason for changing how we handle this

[Kimundi]

Maybe we could just make ! default to () or Void in case where inference fails?

[thestinger]

In that case it probably makes more sense just to use an uninhabited enum with a lang item called something like Fail with consistent usage everywhere. The return type should really contain an uninhabited enum rather than () or it will still be a pain elsewhere.

[nikomatsakis]

I don't consider this a bug. The type of fail! is correct as is and allows a number of important use cases. The compiler is not wrong to say that there is no constraint on that type. The correct answer is to supply a manual type hint or just not call map -- adding a function like apply would do just as well, or creating a wrapper around fail! that hides the fact that it fails (which is what you are looking to do):

fn fail(s: ~str) { fail!(s); }
my_opt.map(|s| fail(format!(...)); )

[nikomatsakis]

As an example of the kind of code that having fail! have bottom type enables:

let x = match ... {
   case1 => 1,
   case2 => 2,
   case3 => fail!("Impossible!")
};

Another possible fix would be to say that all unconstrained type variables are arbitrarily assigned unit. We did this for a while and reverted it due to bugs I think, and possibly some surprising interactions. I'm inclined not to change the inference rules, personally, partially because I'm not too happy with our inference scheme and I don't want to get locked into defaults we might want to change.

[thestinger]

Okay, I'm convinced.