instanceof narrowing should preserve generic types from super to child type

Search Terms

• instanceof generic
• narrow instanceof

Suggestion

The following code has a rather unexpected behavior today:

class Parent<T> {
    x: T;
}
class Child<S> extends Parent<S> {
    y: S;
}

function example(obj: Parent<number>) {
    if (obj instanceof Child) {
        const child = obj;
    }
}

The narrowed type that child gets is Child<any>.

I'm requesting that the type instead gets narrowed to Child<number>, which is what I originally assumed would happen.

My proposal is, in general, to infer the type arguments to the narrowed-to class type whenever they can be determined from the original type.

Precise Behavior

Consider the following code today:

const x: P = ...;
if (x instanceof C) {
    x;
}

Today, if C is a generic class, x gets the narrowed type C<any, any, any, ...>; otherwise it just gets the type C.

My proposal is to consider the following piece of code, and use it to inform

Imagine that C has a no-argument constructor. Then the following piece of code is valid today:

function onlyP(c: P) { ... }

onlyC(new C());

in order to make it valid, the compiler infers type arguments for C that make it into a subtype of P. My proposal is to use the same strategy to infer the type arguments for C in an instanceof narrowing.

Consider the following examples today, and their corresponding instanceof narrowings:

function ex1(x: Parent<string>) { }
ex1(new Child()); // inferred arguments: <string>

function ex2(x: Parent<number | string>) { }
ex2(new Child()); // inferred arguments: <number | string>

function ex3(x: Parent<number> | string) { }
ex3(new Child()); // inferred arguments: <number>

function ex4(x: Parent<number> | Parent<string>) { }
ex4(new Child()); // inferred arguments: Child<number | string>
// Note: the above errors, because it Child<number|string> actually fails to be a subtype of
// Parent<number> | Parent<string>.
// We can either choose to infer Child<any> in this case, or use the (incorrect, but more-precise)
// inference Child<number | string>.

Examples

The original use-case I had in mind was roughly the following:

abstract class Obtainer<T> {
  __phantom: T = null as T;
}

abstract class Fetcher<T> extends Obtainer<T> {
  public abstract fetch(): Promise<T>;
}

abstract class Dependency<T, D> extends Obtainer<T> {
  public abstract dependencies(): Obtainer<D>;
  public abstract compute(got: D): T;
}

async function obtain<T>(obtainer: Obtainer<T>): Promise<T> {
  if (obtainer instanceof Fetcher) {
    // obtainer: Fetcher<T>
    // currently, it's a Fetcher<any>
    return await obtainer.fetch();
  }
  if (obtainer instanceof Dependency) {
    // obtainer: Dependency<T, any>
    // currently, it's a Dependency<any, any>
    const dependencies = obtainer.dependencies();
    return obtainer.compute(await obtain(dependencies));
  }
  throw new Error("not implemented");
}

(note: there's still one extraneous any in the above, since the D parameter cannot be inferred)

Checklist

My suggestion meets these guidelines:

• This wouldn't be a breaking change in existing TypeScript/JavaScript code
  • This is a breaking change. However, I think it could only really break code that was already broken. If really necessary, it could be put behind a new --strictGenericNarrowing flag.
• This wouldn't change the runtime behavior of existing JavaScript code
  • No change to runtime; just inferred generic types.
• This could be implemented without emitting different JS based on the types of the expressions
• [ x This isn't a runtime feature (e.g. library functionality, non-ECMAScript syntax with JavaScript output, etc.)
• This feature would agree with the rest of TypeScript's Design Goals.
  • I think this is an easy win for a more sound type system without any negative impact

Also, I am interested in contributing this change if it's approved!

No se nombran archivos ni pruebas. Empieza reduciendo los ejemplos de Parent/Child, Fetcher y Dependency, y rastreando el narrowing de instanceof de TypeScript junto con la inferencia de tipos genéricos. Se considera terminado cuando los argumentos genéricos estrechados se conservan donde pueden inferirse, y los casos de unión documentados se comportan de forma coherente.