Don't widen return types of function expressions

[RyanCavanaugh]

This change courtesy @JsonFreeman who is trying it out

(context elided: widening of function expression return types)

The problem with this widening is observable in type argument inference and no implicit any. For type argument inference, we are prone to infer an any, where we should not:

function f<T>(x: T, y: T): T { }
f(1, null); // returns number
f(() => 1, () => null); // returns () => any, but should return () => number

So after we get to parity, I propose we do the following. We do not widen function expressions. A function is simply given a function type. However, a function declaration (and a named function expression) introduces a name whose type is the widened form of the type of the function. Very simple to explain and simple to implement.

I’ve been told that this would be a breaking change, because types that used to be any are now more specific types. But here are some reasons why it would be okay:

• In the places where you actually need the type to be any (because there are no other inference candidates), you would still get any as a result
• In places where there was a better (more specific) type to infer, you’d get the better type.
• With the noImplicitAny flag, you’d get fewer errors because there are actually fewer implicit anys

Questions:

Is a principle of design changes going forward to not switch from 'any' to a more precise type because it can be a breaking change?

Going with 'not a breaking change' here because this is unlikely to break working code, but we need to verify this.

Would this manufacture two types?

In essence, we already have two types: The original and the widened type. So by that measure this is not really a change

Has someone tried it?

Jason willing to try it out and report back

[DanielRosenwasser]

Bumping this since we ran into a situation with a leaked any in our own compiler. Minimal example:

function forEach<T, U>(array: T[], callback: (element: T, index: number) => U): U {
    if (array) {
        for (let i = 0, len = array.length; i < len; i++) {
            let result = callback(array[i], i);
            if (result) {
                return result;
            }
        }
    }
    return undefined;
}


forEach([1, 2, 3, 4], () => { 
    // do stuff
    if (blah) {
        // bam! accidental any
        return undefined;
    }
});

[mhegazy]

We need an implementation proposal for this one.

[JsonFreeman]

What if we just did this:

• Don't widen when you type a function body
• When you widen a type, and it's a function type, you widen the return type
• When you resolveName and it's a function expression or function declaration, you widen its type

[myitcv]

I've arrived here via #7220. Would appreciate some guidance on whether the following is expected behaviour or not (playground):

interface G<K, R> {
    (v: K): R;
}

function F<T>(v: T): T {
    return v;
}

interface A {
    __Type: "A";
}

interface B {
    __Type: "B";
}

let a: A;
let b: B;

a = b; // OK: error as expected, B is not assignable to A
b = a; // OK: error as expected, A is not assignable to B

let x: G<A, B> = F;  // How is this possible?

It's the final assignment I cannot understand... because the compiler does not complain.

[RyanCavanaugh]

During assignment checking, type parameters are replaced with any. So the effective signature of F here is (a: any) => any

[JsonFreeman]

@myitcv See #138

[RyanCavanaugh]

Time to take this back up since it's affecting a lot of people writing reducers

[jdanielson-indeed]

+1 to consider this or adding it as a compiler flag.

[EddyVinck]

I ran into problems with some .map() calls, which seems related to the issue described.

Is this still being worked on?

[folex]

Hi!

Is return type widening a possible cause for the lack of errors in this case?

type T = { foo: number };
type F = () => T

// no type error
const a: F = () => ({
  foo: 1,
  bar: 2,
})

Or is it because of the covariance rule on function subtyping?

[MaksimKlepikov]

Happens for .map()

interface IUser {
  id: number;
}
// No errors
const users: Array<IUser> = [].map<IUser>(() => ({
  id: 1,
  unknownProperty: 'unknownProperty',
}));

// Property 'id' is missing in type '{ unknownProperty: string; }' but required in type 'IUser'.ts(2741)
const users1: Array<IUser> = [].map<IUser>(() => ({
  unknownProperty: 'unknownProperty',
}));
console.log(users);
console.log(users1);

[Akryum]

This issue is also present when writing GraphQL resolvers:

const resolvers: Resolvers = {
  Query: {
    someField: () => ({
      requiredField: 'some value',
      // @ts-expect-error
      optionalFieldWithTyppo: 'some value',
    })
  }
}

As we can see in this example TS is not doing its job of catching the type issue introduced by the typo on the optional field name.

Playground | Other example

AFAIK there is no real workaround (except explicitly putting the return type on the function which defeats the purpose of putting the type in the first place). Is this issue still on the radar of the TypeScript team?