Problems with comparability between string literal types

[DanielRosenwasser]

String literals currently only acquire a string literal type if they are contextually typed by another string literal or a union of string literals. They are then comparable and assertable to any string-like type (i.e. string, another string literal type, and any union with a string literal type inside of it).

By and large, this gives us the behavior you want; however, there are some problems with this approach.

Problems

switch/case

This is valid, which seems bad:

let x: "foo" = "foo";
// ...
switch (x) {
    case "bar":
        console.log("wat");
}

When can x be "bar"? Without bypassing the type system, clearly never.

Equality

This is valid, which seems bad:

let x: "foo";
// ...
if (x === "bar") {
}

This is similar to the switch statement example. When will this condition be true? Again, without fooling the type system, it can't be.

Type assertions

These are both valid, which seems bad:

let x = "hello" as "world";
let y = <"bar" | "baz">"foo";

This is where we allow you to lie to the type system. This is probably way too lenient, so I'd say that this is bad.

Solutions?

Most of the solutions here would require us to tighten our rules regarding allowing these operations between string-like entities. Essentially, we would need to capture some of the same behavior originally proposed in the pull request about a new "comparable" type relationship (#5517).

Widening

Why don't we have every string literal start out with a string literal type and widen to string when needed? We could certainly do that, but the question is when does a string literal type need to be widened?

To quote @JsonFreeman on #5300 (comment), widening as it would stand today would be a problem for type argument inference:

I think the biggest issue with the widening approach is that types are widened after type argument inference, which is an issue we have discussed before. It makes it so that any string literals that get inferred are automatically widened to string, which is a little unfortunate.

Here's the problem Jason was talking about

declare function f<T>(x: T): T;

let x = f("hello");

Here, we would always widen to string after picking a type for T. And even if we didn't, we'd still have other issues:

declare function f<T>(x: T, y: T): T;

let x = f("hello", "world");

When we try to figure out what T should be, we'd have two different types: "hello" and "world. Usually, when trying to infer T given the choice between something like number and string, we'll error. We could change this behavior, and while it might be more desirable to infer "hello" | "world" here, it would certainly be less consistent.

Contextually type case expressions

For the switch/case example, we could contextually type each case clause expression by the type of the switch expression:

let x: "foo" | "bar;
// ...
switch (x) {
    case "baz:": // <- error: type '"baz"' is not compatible with '"foo" | "bar"'
        console.log("wat");
}

But what about the other way around? We can't contextually type in both directions, because that would create a circularity. A case clause expression would try to grab a contextual type from the switch expression, which would try to grab the contextual type from its case clauses... etc.

So we could just make things unidirectional, which still has undesirable results

let x: "foo" | "bar;
// ...
switch ("baz") {
    case x: // <- okay: type '"foo" | "bar"' is compatible with 'string'
        console.log("wat");
}

And admittedly, that code is less likely to be written, but really, what is fundamentally different about the first switch (x) example and the following?

let x: "foo" | "bar";
// ...
if (x === "baz") { // <- okay: type '"foo" | "bar"' is compatible with 'string'
    console.log("wat");
}

Clearly you want an error here too, so it would be kind of weird to specially treat switch statements, but not equality comparisons.

Come up with a some new flow of information, like contextual typing

For equality checks, it'd be nice to have something that can flow both ways. For instance, in this example

let x: "foo" | "bar";
// ...
if (x === "baz") {
    console.log("wat");
}

you want "baz" to grab the type of x, independent of this type relation, and apply the same sort of information towards creating a string literal type as you would given a contextual type. Since x would have the type of "foo" | "bar", which is a union containing string literal types, "baz" would acquire its string literal type for the purpose of this check.

You wouldn't end up catching the following

if ("hello" === "world") {
    // ...
}

because each side would just have type string, and inform the other side with the type string. But you can already do that today, and this is a fairly silly case anyway, so it probably doesn't matter.

The biggest question is how we plan to implement this. It would certainly need a proposal.

Come up with some ad hoc checks

We could come up with some weak checks in these positions to figure out if the user is likely making an error. This would need a proposal.

[RyanCavanaugh]

/cc @Aleksey-Bykov

[JsonFreeman]

For the part where you quoted me, this is not quite what I meant. In the example you gave:

declare function f<T>(x: T, y: T): T;
let x = f("hello", "world");

The widening approach would be a problem here, but not because of the widening itself. Rather, because the string literals are initially typed as string literal types, the type argument inference will fail. Neither "hello" nor "world" is a supertype of the other.

What I meant was in an example like this:

var hello: "hello";
declare function f<T>(param: T): T;
f(hello);

You'd want this to infer "hello" for T, since that's the type you passed. But with widening, it would infer string.

I'm pretty sure widening is not appropriate here. Widening is what you do when some kind of type can arise in an expression context, but you don't want that type to land on any named entity (like a variable). You clearly want to allow variables to have string literal types, so to me widening is automatically ruled out.

One other argument. In the statement var a = b;, a and b should have the same type. If you say that variables can have string literal types, yet string literal types can be widened to string, you break this expectation, like this:

var b: "hello" = "hello";
var a = b; // a would have type string if you widen

So the two variables would not have the same type.

[zpdDG4gta8XKpMCd]

I think that the solution for all 3 problems should be one big "NO". This is a type script after all, don't use literal types if they scare you. "NO" doesn't mean that there is no way around for someone who is looking for troubles:

let x = "i don't feel like" : "making sense today"; // <-- should be a problem
let x = <"making sense today">("i don't feel like" : any); // <-- fine, be this way

So I wish that the right way of doing things was a default, and the wrong way was a harder-to-come-with option.

[falsandtru]

I think,

Nondeterministic intersection type for string literal value

"foo" string literal value has a string&"foo" intersection type. This is normal intersection types. TypeScript uses this intersection types as a nondeterministic type. Probably, almost operations can use existing language functions.

strliteral type

strliteral type is a nondeterministic new type. Set is types of string literal value(string&strliteral) >= strliteral types >= any union types of string literal types("foo"|"bar"). strliteral types infer a minimum string literal type set. It works like a placeholder for string literal types. <strliteral>"foo" means <"foo">"foo".

strliteral is a super set of string literal types. This type can use only with type parameters.

Declarations

let x = "foo"; // `x` is `string&"foo"`
let x: "foo" = "foo"; // `x` is `"foo"`
let x: string = "foo"; // `x` is `string`
let x: "foo" = <string>"foo"; // error
//let x: strliteral&"foo" = "foo"; // `x` is `"foo"`
//let x: string&strliteral = "foo"; // `x` is `string&"foo"`
let x: "foo"|"bar" = "foo"; // `x` is `"foo"|"bar"`
let x: "foo"&"bar" = "foo"; // error
let x: "foo"&"bar"; // `x` is `"foo"&"bar"` but useless
//let x: strliteral = 0; // error
//let x: strliteral = null; // error
//let x: strliteral = undefined; // error

Assertions

let x = "foo"; // `x` is `string&"foo"`
let x = <"foo">"foo"; // `x` is `"foo"`
let x = <string>"foo"; // `x` is `string`
let x = <"foo"><string>"foo"; // error
//let x = <strliteral>"foo"; // `x` is `"foo"`
//let x = <strliteral&"foo">"foo"; // `x` is `"foo"`
//let x = <string&strliteral><string>"foo"; // `x` is `string`
let x = <"bar">"foo"; // error
let x = <"foo"|"bar">"foo"; // `x` is `"foo"|"bar"`
//let x = <strliteral><"foo"|"bar">"foo"; // `x` is `"foo"|"bar"`
//let x = <strliteral>0; // error
//let x = <strliteral>null; // error
//let x = <strliteral>undefined; // error

let x = <string&"bar"><string>"foo"; // ok...

Equality

let x = "foo";
x === "bar"; // ok, string === string
let x: string = "foo";
x === "bar"; // ok, string === string
let x: "foo" = "foo";
x === "bar"; // error, "foo" === "bar"

Inferences

let x = "foo"; // `x` is `string&"foo"`
let x = "foo" + "bar"; // `x` is `string`
let x = "foo" || "bar"; // `x` is `string`
let x = "foo" && "bar"; // `x` is `string`
let x = <"foo">"foo" + "bar"; // error, invalid operation
let x = <"foo">"foo" || "bar"; // error, meaningless code
let x = <"foo">"foo" && "bar"; // error, meaningless code
//let x = <strliteral>"foo" + "bar"; // error, invalid operation
//let x = <strliteral>"foo" || "bar"; // error, meaningless code
//let x = <strliteral>"foo" && "bar"; // error, meaningless code

Inference constraints

declare function f<T>(x: T, y: T): T;

let x = f("hello", "world"); // `T` is `string`
let x = f<string>("hello", "world"); // `T` is `string`
//let x = f<strliteral>("hello", "world"); // `T` is `"hello" | "world"`
let x = f<"hello" | "world">("hello", "world"); // `T` is `"hello" | "world"`
let x = f<"hello">("hello", "world"); // error

declare function f<T extends strliteral>(x: T, y: T): T;
let x = f("hello", "world"); // `T` is `"hello" | "world"`

Type guards

let x = "foo";
switch (x) {
    default: // ok, `x` is `string&"foo"` here
    case "foo": // ok, `x` is `string&"foo"`
    case "bar": // ok, `x` is `string`
    default: // ok, `x` is `string` here
}
switch (<"foo">x) {
//switch (<strliteral>x) {
    case "foo": // ok
    case "bar": // error
    default: // ok, `x` is `string` here, this is a special rule for an exception handling
}
switch (<string>x) {
    case "foo": // ok
    case "bar": // ok
    default: // ok
}
switch (<"foo"|"bar">x) {
    default: // ok, `x` is `"foo"|"bar"` here
    case "foo": // ok, `x` is `"foo"`
    default: // ok, `x` is `"bar"` here
    case "bar": // ok, `x` is `"bar"`
    case "fizzbazz": // error
    default: // ok, `x` is `string` here
}

[DanielRosenwasser]

@JsonFreeman sorry, my mistake. There are two issues there. I'll amend the original content.

[JsonFreeman]

In the example

declare function f<T>(x: T, y: T): T;
let x = f("hello", "world");

You claim that it would be more desirable to infer "hello" | "world". Why is this more desirable? I would think string is the correct inference here. There is no indication that you want string literal types to play any part here, as you want to be able to assign any string to x after it is initialized (I think).

[JsonFreeman]

Why does the type assertion problem happen? I thought in a type assertion, the operand is contextually typed by the asserted type. Doesn't that take care of it?

[DanielRosenwasser]

Why does the type assertion problem happen? I thought in a type assertion, the operand is contextually typed by the asserted type. Doesn't that take care of it?

Here is why something like "foo" as "bar" or "<hello">"world" succeeds.

Prior to string literal types, we would perform to/from assignability checking on all types to see if they are compatible in the three listed locations. However, that couldn't capture the following behavior:

var x: "foo" | boolean;
// ...
var y = x as string; // error: 'string' is not assignable to 'boolean', and vice versa

The comparable relationship was supposed to fix this, but instead we just simply settled for a different fix where a string-like type, or a union containing a string-like type, is compatible with any other string-like type, or union containing a string-like type.

[JsonFreeman]

I see, that makes sense. But I also don't think it is specific to string literal types. In logical terms, what you really want to check is not assignability, but consistency. For all three of the cases you pointed out, whether or not they include string literal types, what matters is whether the types involved have any values in common. In other words, there should be an error if the intersection of the sets corresponding to the types is empty. This is not adequately captured by assignability, regardless of contextual typing or any other typing mechanism. I think this needs to be a new type relation (consistency).

If that is deemed too complex, then I think the only other realistic alternative is to just allow all three cases. Anything else is smoke and mirrors, it's not really addressing the issue at its core. I also do not agree with the notion of any string-like type being assignable to all other string-like types. It is nice for type assertions, but too permissive for other scenarios involving assignability.

[JsonFreeman]

I realize now what is different about string literal types versus object types for example. For object types, consistency is theoretically the goal, but it holds trivially in almost all cases. Pretty much all pairs of objects are consistent if you can take their intersection. So this is not very interesting, and so consistency is not enough to imply that the types are related in some meaningful way.

For string literals on the other hand, because the types map to specific values, consistency is not trivial, and it makes more sense to rely on it for type assertions and equality checks. So I think the answer is a relation that somehow covers consistency, but checks something stronger in the cases where consistency is trivial (like object types).

[DanielRosenwasser]

Spoken offline with @ahejlsberg and @RyanCavanaugh. @ahejlsberg brought up the idea of a specialized "top-level" widening.

Basically, here's the semantics that would be involved.

1. Every string literal gets a string literal type.

2. Widening on a singleton string literal type occurs at every widening location, except when bound to a readonly location. For example:

   // 'c' has type '"hello"'
   const c = "hello";
   
   // 'd' has type 'string'
   let d = c;
   
   // 'e' has type '"hello"'
   let e: "hello" = c;
   
   // 'f' has type 'string'
   let f = e;
   
   // 'g' has type '{ prop: string }'
   const e = {
      prop: c
   };

3. String literal types are not widened when within a union, intersection, or within a property of any other type:

   // 'c' has type '"hello" | "world"'
   const c = randBool() ? "hello" : "world";
   
   // 'd' has type '"hello" | "world"'
   let d = c;
   
   // 'e' has type '{ prop: "hello" | "world" }'
   const e = {
      prop: c
   };

Additionally, another change would be to union multiple string literal return types rather than complaining about no best common type.