Type literal assertion as const to prevent type widening

[cshaa]

Search Terms

String literal type and number literal type in an object literal. Type assertion to make the string literal or number literal also a type literal. Cast string or number to type literal. Make a property definition in object literal behave like constant for the type inference. Define a property as string literal or number literal in an object. Narow the type of a literal to its exact value.

Background

When I declare a constant and assign a string literal or a number literal to it, type inference works differently than when I declare a variable in the same way. While the variable assumes a wide type, the constant settles for a much narrower type literal.

let   a = 'foo'; // type of a is string
const b = 'foo'; // type of b is 'foo'

This behavior comes in handy in many scenarios when exact values are needed. The problem is that the compiler cannot be forced to the "type literal inference mode" outside of the const declaration. For example the only way of defining an object literal with type-literal properties is this:

const o = {
  a: 42 as 42,
  b: 'foo' as 'foo',
  c: 'someReallyLongPropertyValue' as 'someReallyLongPropertyValue',
  d: Symbol('a') as... ahem, what?
};
// type, as well as value, of o is { a: 42, b: 'foo', c: 'some…' }

That is not only incredibly annoying, but also violates the principles of DRY code. In adition it's really annoying.

Suggestion

Since const is already a reserved keyword and can't be used as a name for a type, adding a type assertion expresion as const would cause no harm. This expression would switch the compiler to the "constant inference mode" where it prefers type literals over wide types.

// In objects
const o = {
  a: 42 as const, // type: 42
  b: 'foo' as const, // type: 'foo'
  c: 'someReallyLongPropertyValue' as const
  d: Symbol('a') as const // type: typeof o.d
}


// In generic function
let f = <T>(a: T) => ({ boxed: a });
let x = f(42 as const) // T is 42, typeof x is { boxed: 42 }

Non-simple types

There are currently three competing proposals for the way as const would work with non-simple types.

Shallow literal assertion

The first one would stay true to its name and would treat the type as if it were assigned to a constant.

// Type of a is exactly the same as the type of b, regardless of X
const a = X;
let b = X as const;

However, this would mean that all non-trivial expressions would stay the same as they were without the as const. Therefore I would argue it's less useful than the other two proposals.

Tuple-friendly literal assertion

The second proposal differs in the way it treats array literals. While the shallow assertion treats all array literals as arrays, the tuple-friendly assertion treats them as tuples. Then it recursively propagates deeper untill it stops at a type that is neither a string, number, boolean, symbol, nor Array.

let a = [1, 2] as const; // type: [1, 2]
let b = [ [1, 2], 'foo' ] as const; // type: [ [1, 2], 'foo' ]
let c = [ 1, { a: 1 } ] as const; // type: [ 1, { a: number } ]

// Furthermore new syntax could be invented for arrays of const
// but that is outside the scope of this proposal right now
let d = [ [1,2] as const, [3,4] as const, [5,6] as const ]; // type: Array< [1,2] | [3,4] | [5,6] >
let e = [ [1,2], [3,4], [5,6] ] as const[]; // type: Array< [1,2] | [3,4] | [5,6] >

This is probably the most useful proposal, as it solves both the problem described in Use Cases and the problems described in #11152.

Deep literal assertion

The third proposal would recursively iterate even through object literals. I included it just for sake of completeness, but I don't think it could be any more useful than the tuple-friendly variant.

let c = [ 1, { a: 1, b: [1, 2] } ] as const; // type: [ 1, { a: 1, b: [1, 2] } ]

Use Cases

Say I'm using a library which takes a very long object of type LibraryParams of various parameters and input data. I don't know some of the data right away, I need to compute them in my program, so I'd like to create my object myParams which I would fill and then pass to the library.

Since some of the properties of LibraryParams are optional and I don't want to check for them or assert them every time I use them – I know I've set them, right? – I wouldn't set the type of myParams to LibraryParams. Rather I'd use a narrower type by simply declaring an object literal with the things I need.

However some of the properties need to be selected from a set of exact values and when I add them to myParams, they turn into a string or a number and render my object incompatible with LibraryParams.

There are some ways around it using the existing code, none of which are particularly good. I'll give some examples in the next section.

Examples

Imagine that all of these examples contain much longer programs.

// Type is too wide because of LibraryParams

const myParams: LibraryParams = {
  name: "Foobar",
  favouritePrime: 7,
  sports: [ 'chess' ]
}

if (myFunctions.lovesFootball()) {
  myParams.sports.push('football'); //sports is possibly undefined 🤷
}

Library.doThings(myParams);

// Type is too wide because of literal type widening

const myParams = {
  name: "Foobar",
  favouritePrime: 7,
  sports: [ 'chess' ]
}

if (myFunctions.lovesFootball()) {
  myParams.sports.push('football');
}

Library.doThings(myParams); //number is not assignable to prime 🤷

// Too many type assertions

const myParams = {
  name: "Foobar",
  favouritePrime: 7,
  sports: [ 'chess' ]
}

if (myFunctions.lovesFootball()) {
   // this looks even weirder when you do it for the 10th time 🤮
  (myParams.sports as string[]).push('football');
}

Library.doThings(myParams);

// Too much searching for the correct types

const myParams = {
  name: "Foobar",
  favouritePrime: Library.ParamTypes.Primes.Seven, // 🤮
  sports: [ 'chess' ]
}

if (myFunctions.lovesFootball()) {
  myParams.sports.push('football');
}

Library.doThings(myParams);

// Probably the best solution but definitely not very dry

const myParams = {
  name: "Foobar",
  favouritePrime: 7 as 7 // 😕
  sports: [ 'chess' ]
}

if (myFunctions.lovesFootball()) {
  myParams.sports.push('football');
}

Library.doThings(myParams);

// Top tier 👌😉

const myParams = {
  name: "Foobar",
  favouritePrime: 7 as const // 🤩
  sports: [ 'chess' ]
}

if (myFunctions.lovesFootball()) {
  myParams.sports.push('football');
}

Library.doThings(myParams);

Checklist

My suggestion meets these guidelines:

• This wouldn't be a breaking change in existing TypeScript / JavaScript code
• This wouldn't change the runtime behavior of existing JavaScript code
• This could be implemented without emitting different JS based on the types of the expressions
• This isn't a runtime feature (e.g. new expression-level syntax)

Related

#14745 Bug that allows number literal types to be incremented
#20195 A different approach to the problem of type literals in objects, excluding generic functions
#11152 Use cases for as const in generic functions

[mattmccutchen]

I think this is a duplicate of #10195.

[cshaa]

@mattmccutchen You're right, #10195 proposes a different solution for the same problem. However I think this one could be implemented much more easily as it doesn't have any breaking changes and doesn't colide with something as basic as... parentheses.

[zpdDG4gta8XKpMCd]

"solution" that works today:

const data = new class {
  readonly name = "Foobar";
  readonly favouritePrime = 7;
  readonly sports = [ 'chess' ];
};

[yortus]

I suggested pretty much the same thing in #10195 (comment), but with as unit instead of as const. @DanielRosenwasser responded with some feedback in #10195 (comment).

[connorjclark]

One example of how this comes up in common code (if I am understanding the issue correctly) - React CSS styles.

const style = {
  textAlign: 'center' as 'center'
}

return <div style={style}></div>

[denis-sokolov]

The React CSS styles can be handled with a type annotation on the constant:

const style: React.CSSProperties = {
  textAlign: 'center'
}

return <div style={style}></div>

[gwicksted]

This is somewhat possible with the following notation:

// doesn't quite work right for interface types:
type Exactly<T> = T | never;

// for value types:
// unfortunately you have to echo v into the type param if you want to constrain it to only that value
const Literally = <T>(v: T): Exactly<T> => {
    return v as Readonly<T>;
};

// In objects
const o = {
  a: Literally<42>(42),
  b: Literally<"foo">('foo'),
  c: Literally<'someReallyLongPropertyValue'>('someReallyLongPropertyValue'),
  d: Literally(Symbol('a'))
};

// In generic function
let f = <T>(a: T) => ({ boxed: a });
let x = f(Literally<42>(42));

x.boxed = 43; // compile time error

Example in Playground

[yortus]

@gwicksted your code can be simplified as follows (note no need to repeat all the literals twice):

function literal<T extends string|number|symbol>(v: T): T {
    return v;
};

// In objects
const o = {
  a: literal(42),
  b: literal("foo"),
  c: literal('someReallyLongPropertyValue'),
  d: literal(Symbol('a'))
};


// In generic function
let f = <T>(a: T) => ({ boxed: a });
let x = f(literal(42));

x.boxed = 43;

(Playground link)

But note this only works for string and number literals. Symbols, tuples and booleans are widened (they are widened with your code too).

[gwicksted]

@yortus That's great! It seems the trick is the "extends" and one of the literal types with the notable exceptions you mentioned.

Booleans, Tuples, numbers, and strings all work when given the specific type using as which makes the literal function a shorter option when you don't want to re-type the literal. So other than Symbol, it's possible today to achieve the desired behavior.... just not ideal.

Full example:

function literal<T extends string|number>(v: T): T {
    return v;
};

// In objects
const o = {
  a: 42 as 42, // or literal(42)
  b: "foo" as "foo", // or literal("foo")
  c: literal("someReallyLongPropertyValue"),
  d: Symbol("a"), // no solution yet
  e: false as false,
  f: ["abc", false] as ["abc", false]
};

o.a = 43; // error
o.b = "bar"; // error
o.c = "shorter"; // error
o.d = Symbol("d"); // no error
o.e = true; // error
o.f = ["def", true]; // error


// In generic function
let f = <T>(a: T) => ({ boxed: a });
let x = f(literal(42));

x.boxed = 43; // error

Playground link

IMO it would be great if something as simple as this would prevent widening in all cases:

// PROPOSAL: does not work
function literal<T extends string | number | boolean | symbol>(v: T): T | never {
    return v; // as T | never
}

[cshaa]

@gwicksted @yortus This is a cool walkaround, but it affects the compiled JavaScript. However I guess the interpreter will quickly realize it's an identity function and optimize the heck out of it.

[yortus]

EDIT: figured out tuples of any length.

Here's a slightly more refined literal helper function that works for strings, numbers, booleans (as of the 3.2.0-dev.20180916 nightly build) and tuples:

// Overloaded function - supports strings, numbers, booleans and tuples. Rejects others.
function literal<T extends string | number | boolean>(t: T): T;
function literal<T>(t: Tuple<T>): T;
function literal(t: any) { return t; }
type Tuple<T> = T extends [any?, ...any[]] ? T : never;

let v1 = literal(42);                           // 42
let v2 = literal('foo');                        // 'foo'
let v3 = literal(true);                         // true
let v4 = literal([]);                           // []
let v5 = literal([10, 20]);                     // [number, number]
let v6 = literal([literal(1), false, 'foo']);   // [1, boolean, string]
let v7 = literal([1, 2, 3, literal('foo'),
                  true, literal(true)]);        // [number, number, number, 'foo', boolean, true]
let v8 = literal(Symbol('sym'));                // ERROR

For boolean literals, this depends on #27042 which arrived in the 3.2.0-dev.20180916 nightly.

For tuples, I couldn't find any general way to prevent them widening. But as a simple workaround, we can just overload the literal helper function for as many tuple elements as deemed practical. the Tuple<T> type seems to do the trick, although I'm not sure why this exact form prevents widening but the simpler T extends [...any[]] does not.

[cshaa]

@yortus That's really cool! And when Variadic Generics #5453 land (which is probably still far in the future), this approach could simplify and support even symbols.

[gwicksted]

@yortus nice! It's succinct and hits the 99% with a single function. I think it solves #10195 without compiler changes (just lib). Not sure if it helps with #26841

On that note: should literal(x) remain a copy-and-paste helper until a language feature is added to suppress widening?

It's unfortunate things become verbose with a complex Tuple structure (but those are rare in my experience):

const o: {
  f: ["abc", false, [123, true]] as ["abc", false, [123, true]] // this works
  g: literal([literal("abc"), literal(false), literal([literal(123), literal(true)])])
};

o.f = ["abc", false, [567, false]]; // error on both [567, false]
o.g= ["abc", false, [567, false]]; // (untested, as I do not have dev nightly at the moment)

Can it be written like the following and still work? Then any is never present as an argument type.

type Tuple<T> = T extends [any?, ...any[]] ? T : never;
function literal<T extends string | number | boolean | Tuple<T>>(t: T): T { return t; }