Naked generic type returned from iterator method

[reverofevil]

🔎 Search Terms

generic generator static

🕗 Version & Regression Information

• I was unable to test this on prior versions because there is still no online version of every-ts bisect

⏯ Playground Link

Link

💻 Code

class Base {
    static *[Symbol.iterator]<T>(this: T) {
        yield this;
    }
}
class Child extends Base {}
const test = function* () { yield* Child };
const value = [...test()][0]; // T
console.log(value); // Child

🙁 Actual behavior

The result is typed as T, whatever generic parameter outside of generic code means.

🙂 Expected behavior

T should be instantiated with Child during a call to [Symbol.iterator]

Additional information about the issue

For regular static method calls it works as expected, by instantiating T

class Base {
    static foo<T>(this: T) {
        return this;
    }
}
class Child extends Base {}
const value = Child.foo(); // Child
console.log(value); // Child

On the other hand, this is a nice way to create unique types for opaque type implementation! 🙃

[Andarist]

This is a fun one too:

class Base {
  static *[Symbol.iterator]<T extends [1, 2]>() {
    yield {} as T;
  }
}
class Child extends Base {}
const test = function* () {
  yield* Child;
};
const value = [...test()][0];
//    ^? const value: T extends [1, 2]

Basically, all type parameters leak in this situation.

[reverofevil]

Also it doesn't have to be static. I left it in the example, because otherwise it's harder to see that the behavior is incorrect.

This is all it actually takes:

class Base {
  *[Symbol.iterator]<T>() {
    yield {} as T;
  }
}

const x = [...new Base()][0];
//    ^? const x: T

I tried another Symbol() there, and it works fine.

[Andarist]

The proper fix for this is to thread the obtained [Symbol.iterator] signatures through resolveCall. Just like this PR does things with [Symbol.hasInstance]: #55052

I have this working locally but it needs a lot of cleanup (implementation of iteration protocol has many layers/helper functions) and investigating all of the other ways of iterating through things. The problem is not exclusive to spreads:

class Base {
  *[Symbol.iterator]<T>() {
    yield {} as T;
  }
}

for (const element of new Base()) {
  element;
  // ^? const element: T
}

[reverofevil]

Oof! It actually doesn't do resolveCall stuff at all! Not only this shouldn't compile (foo comes from where?), but if you leave only one signature, it infers correct type.

declare class Base {
  [Symbol.iterator](foo: 1): Generator<3, void, undefined>;
  [Symbol.iterator](foo: 2): Generator<4, void, undefined>;
  [Symbol.iterator](foo: any): Generator<3 | 4, void, undefined>;
}

const x = [...new Base()];
//    ^? const x: never[]

[reverofevil]

@rbuckton Any news on this?