Can't evaluate equality of symbols made by Symbol.for

[falsandtru]

This is also an obstacle of branded types and ghost types. I think resolving this problem is easy because these types are just a new literal type similar to string literal types.

TypeScript Version: 3.7.x-dev.20191228

Search Terms:

Code

const a = Symbol.for('');
const b = Symbol.for('');
const c: typeof b = Symbol.for(''); // error
a === b; // error

Expected behavior:
no error
Actual behavior:
Type 'typeof b' is not assignable to type 'typeof a'.
This condition will always return 'false' since the types 'typeof a' and 'typeof b' have no overlap.
Playground Link: https://www.typescriptlang.org/play/index.html?ts=3.8.0-dev.20191228&ssl=1&ssc=1&pln=4&pc=18#code/MYewdgzgLgBAhjAvDAygTwLYCMQBsB0AZiAE4AUA5BQJQDcAUKJLFkqpjgceVXY+NBjAAXDChoADgFMQhGK2TpseIqUo1aMAPRaYUkiVL0EiU-M069B0kA

Related Issues: #35200

[falsandtru]

Example

// npm package A@v1
namespace A1 {
    // Structural hidden field across versions and packages.
    const tag = Symbol.for('TagName');
    export class A<T> {
        constructor(arg: T) { }
        [tag]: T;
    }
}
// npm package A@v2
namespace A2 {
    // Structural hidden field across versions and packages.
    const tag = Symbol.for('TagName');
    export class A<T> {
        constructor(arg: T) { }
        [tag]: T;
    }
    export declare function f(a: A<'a'>): void;
}
// npm package B
namespace B {
    A2.f(new A1.A('a')); // Should work.
}

// npm package A
namespace A {
    // Structural hidden field across versions and packages.
    const tag = Symbol.for('TagName');
    export class A<T> {
        constructor(arg: T) { }
        [tag]: T;
    }
}
// npm package B
namespace B {
    // Structural hidden field across versions and packages.
    const tag = Symbol.for('TagName');
    export class B<T> {
        constructor(arg: T) { }
        [tag]: T;
    }
}
// npm package C
namespace C {
    declare function f(a: A.A<'a'>): void;
    f(new B.B('a')); // Should work.
}

[ExE-Boss]

Like we have unique symbol, we should probably support:

const foo: global symbol "<name>";

This would allow declaring Node’s built-in symbols as:

declare module "util" {
	namespace inspect {
		// https://github.com/nodejs/node/issues/20821
		const custom: global symbol "nodejs.util.inspect.custom";
	}

	namespace promisify {
		// https://github.com/nodejs/node/issues/31647
		const custom: global symbol "nodejs.util.promisify.custom";
	}
}

[typeofweb]

It's a bug I just stumbled upon and it was really surprising.

[ExE-Boss]

It’s the intended behaviour for unique symbol types, which global symbols actually aren’t, but they’re currently using the same code path.

[typeofweb]

Sounds like it's not expected, more like an unforeseen side effect of adding unique symbols.

[kurtharriger]

I was looking at code for library and noticed usage of Symbol and Symbol.for I was curious what the difference was and found this https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Symbol/for

I then attempted to try it in typescript and erroneously get the error that the conditional will alwasy be false even when the value is actually true.

https://www.typescriptlang.org/play/?ssl=3&ssc=25&pln=1&pc=1#code/MYewdgzgLgBARjAvDAygTwLZxAGwHQBmIATgBQDkcAhseQJQDcAUKJLHAExKqbb5FlKNes1YRcAUzw4QAc1LwkiZJxiMmQA

[cowboyd]

Is the problem that there is no concept of a "symbol literal" in TypeScript?

It would be really nice to have a literal syntax for symbols which would type just like any other primitive literal like 'string', 5, or true.

Syntax is arbitrary, but using@ as a placeholder.

const something = @something; //=> Symbol.for("something")
const somethingElse = @"something.else"; //=> Symbol.for("something.else")

const object: { [something]: string } = { @something: 'hello' };

Seems like this would ease a lot of pain in matching types that use well known symbols as keys. But even in the absence of a literal syntax, what would it take for TS to treat Symbol.for() expressions as literals?

[nicolo-ribaudo]

@cowboyd That would risk big conflicts with decorators, but a while ago I was thinking about the symbol "foo" syntax:

declare function Symbol.for<T extends string>(name: T): symbol T;
declare function Symbol.keyFor<T extends string>(sym: symbol T): T;
declare function Symbol.keyFor(sym: symbol): string | undefined;

let a = Symbol.for("foo"); // type is `symbol "foo"`
let b: symbol "foo" = Symbol.for("bar"); // error
let c = Symbol.for("baz" as string); // symbol string
let d: symbol = Symbol.for("foo"); // ok, symbol T is a subtype of symbol
let d: symbol "foo" = Symbol(); // error

let descA: "foo" = Symbol.keyFor(a); // ok
let descB: "bar" = Symbol.keyFor(a); // err, "foo" is not assignable to "bar"

EDIT: I just noticed that this was already suggested at #35909 (comment) 😬

[cowboyd]

@nicolo-ribaudo you're definitely right, and the @ was just the first thing that popped into my head. And in fact, there isn't any need for a new literal syntax, just a new type that accurately represents the return value of Symbol.for().

I like your suggestions. Also need to consider the difference between Symbol() and Symbol.for(). Maybe something like:

let a = Symbol("foo") // type is `unique symbol "foo"`
let b = Symbol.for("foo") // type is `symbol for "foo"`

Anyhow, as you noted @ExE-Boss already suggested this.... wonder why this thread petered out? Is there something that we can do?

[pjeby]

Throwing in a use case here: I've got a library that exports a symbol for purposes of interop, that other code can declare methods with that symbol, without needing to depend on the library that defines the symbol. But then, in a project that uses both of those libraries, Typescript can't tell that module 2 implements the interface specified by module 1, even though it uses the same Symbol.for().

As it stands, the only way to get this kind of interop is with strings, not symbols, as there's no "global symbol" or "symbol for" syntax in Typescript. Using typeof can't work here because that would force the common dependency, and a shared type definitions file, since merely equivalent types are not sufficient for "unique symbol" to match.

[yuhr]

Excerpting from #60944, as I believe it's worth mentioning here:

1. What do you want to use this for?

Making multiple versions of the same library compatible.

Sometimes a library generates objects with a private property at a unique symbol key and use it as some kind of marker to statically or dynamically determine if an object is really generated or already processed with the library, while allowing users to define arbitrary properties on it as much as possible.

But what if a user has to end up depending on multiple versions of the library somehow? To address this, the library should accept objects from other versions of the same library, so the library might want to start using non-unique symbols for the purpose. This does work at runtime but not at compile time due to the absence of indexed symbol types.

Using a string key instead of a non-unique symbol key invades the string key space of such an object at runtime, and could conflict with an index signature at compile time.