Discuss
  • Map literal

Map literal

# Mohsen Azimi (8 years ago)

I'm using Maps a lot now and I was thinking why there is no "easy" way of declaring them like objects and arrays.

I'm sure I'm not the first one who came up with the idea of having Map literal declaration. There are many ways we can introduce new syntax for declaring Maps via a literal syntax such as:

let map = [window: 'window', document: 'document'];

or

let map = {{window: 'window', document: 'document'}}

and possibly many more.

I searched the discussions but couldn't find a topic on this. Have you discussed this before?

# Alexander Jones (8 years ago)

I agree this is pretty important. Using actual maps really frees up a lot of complexity, but the syntax is cumbersome to say the least.

Whatever the decided syntax, bare words as string keys is a really bad idea IMO. The key syntax should be parsed as an expression, like the values are, and like they are in basically every other language.

Another outstanding issue is that we might want the syntax for Immutable.Map, or WeakMap, or MapTwoPointOh that improves deficiency $x, $y and $z. I'd say introducing a special syntax for Map right now is not ideal.

Rather, we have an opportunity to instead devise a syntax for an abstract map. While we're at it, we might as well do the same for an abstract list. Why should maps have all the fun?

const {List: IList, Map: IMap} = Immutable;
const bar = 43;
const map = IMap#{"foo": 42, bar: 44};  // keys "foo" and 43
const list = IList#[4, 5, 6, 7, Map#{map: "why not?"}];  // 5th element is
a Map with one key, which is the Immutable.Map above
const weakMap = WeakMap#{map: "It's an Immutable", list: "Also Immutable"};
 // WeakMap keys are the objects map and list

It could desugar as, for the sake of example:

Foo#{key: value, ...}
➔
Foo[Symbol.literalOf]([[key, value], ...][Symbol.iterator]())

and

Foo#[value, ...]
➔
Foo[Symbol.literalOf]([value, ...][Symbol.iterator]())

The nice thing about this is it's extensible and future proofs the language a little bit. The actual arrays need not exist if engines choose to implement this more efficiently - the syntax just results in an iterator which yields the elements of the literal. The only difference between the [] and the {} notation ise that the {} notation enforces syntactically valid key-value pairs and are a little less heavy on brackets.

I know literally every proposal ever these days seems to claim the # symbol now, so that's clearly an issue to contend with... :)

Alex

# Jordan Harband (8 years ago)

fwiw, my Object.entries proposal ( ljharb/proposal-object-values-entries ) would allow you to do: new Map(Object.entries({ a: 'b', b: 'c' })).

# Alexander Jones (8 years ago)

Quite verbose, harder to optimize and only supports string keys.

# Tab Atkins Jr. (8 years ago)

On Wed, Oct 28, 2015 at 8:36 AM, Alexander Jones <alex at weej.com> wrote:

I agree this is pretty important. Using actual maps really frees up a lot of complexity, but the syntax is cumbersome to say the least.

Whatever the decided syntax, bare words as string keys is a really bad idea IMO. The key syntax should be parsed as an expression, like the values are, and like they are in basically every other language.

Another outstanding issue is that we might want the syntax for Immutable.Map, or WeakMap, or MapTwoPointOh that improves deficiency $x, $y and $z. I'd say introducing a special syntax for Map right now is not ideal.

Currently, the "extensible literal syntax" for this isn't that bad:

const bar = 43; const map = Immutable.Map([["foo", 42], [bar, 44]]);

It's a little more verbose because the entries have to be surrounded by [], but hey.

# Alexander Jones (8 years ago)

True, but easy to mess up and only be treated to a runtime error. Three nested brackets at the start and end could definitely be better, and this just encourages people to use POJSOs instead. Also not a very uniform interface if you look at how to construct a Map, Set or Immutable.List at present, though admittedly constructor call for the ES6 types would be a partial improvement.

# Viktor Kronvall (8 years ago)
const map = IMap#{"foo": 42, bar: 44};

It could desugar as, for the sake of example:

Foo#{key: value, ...}
➔
Foo[Symbol.literalOf]([[key, value], ...][Symbol.iterator]())

I like this proposal. However, Maps should guarantee insertion order when traversing the keys and values and desugaring it like that does not respect this guarantee or more precisely it will lead to (in my opinion) unexpected order of the keys.

Object.keys({1: 6, bar: 'Hello', 2: 8}); // → [ '1', '2', 'bar' ]

If I'm not mistaken this will be same order for {1: 6, bar: 'Hello', 2: 8}[Symbol.iterator]().

This implies that:

Map#{1: 6, bar: 'Hello', 2: 8};

Will not have entries in the order [[1, 6], ['bar', 'Hello'], [2,8]] but instead [[1,6], [2,8], ['bar','Hello']].

This means that possible future destructuring of a Map will be harder to reason about.

2015-10-28 2:21 GMT+01:00 Alexander Jones <alex at weej.com>:

# Alexander Jones (8 years ago)

Not sure I understand - the desugaring I wrote would absolutely preserve the written ordering because it speaks in terms of an ArrayIterator of key-value pairs. If the map type to which it's applied chooses to forget the ordering then that's fine.

Alex

# Viktor Kronvall (8 years ago)

Hello Alexander,

I see now that I misread your desugaring.

I read:

Map#{1: 6, bar: 'Hello', 2: 8};

as being desugared to:

Map[Symbol.literalOf]({1: 6, bar: 'Hello', 2: 8}[Symbol.iterator]());

But your proposal clearly states that is should be:

Map[Symbol.literalOf]([[1, 6], ['bar', 'Hello'], [2,8]][Symbol.iterator]());

Which would preserve lexical ordering of entries. The fault is completely mine. Sorry.

I like this proposal as it is extensible and not that noisy in syntax. Using the # for this doesn't seem like a bad idea either. People coming from Erlang will be familiar with this as well.

2015-10-28 10:53 GMT+01:00 Alexander Jones <alex at weej.com>:

# Alexander Jones (8 years ago)

Ok, thanks for clarifying. Not only does it preserve order but it also permits non-string keys. You're still missing one detail which is that bar would actually be a variable not a string key.

Another example to clarify that the key part would be an expression:

Map#{foo(42) + 7: "bar"}

I prefer this over the precedent set by object literals which would require that [] are used around a key expression ("computed key") simply due to relieving the syntax noise, which is what this idea is all about. Also, this is how it works in Python and I make no apologies about the similarities ;)

Alex

# Herby Vojčík (8 years ago)

Alexander Jones wrote:

Ok, thanks for clarifying. Not only does it preserve order but it also permits non-string keys. You're still missing one detail which is that bar would actually be a variable not a string key.

Another example to clarify that the key part would be an expression:

Map#{foo(42) + 7: "bar"}

I prefer this over the precedent set by object literals which would require that [] are used around a key expression ("computed key") simply

I, on the other hand, think it should match object literals completely. So your example would be

Map#{[foo(42)+7]: "bar"}

Yes, it's just for consistency and less WTF moment while learning the details.

OTOH, there could be consistent contraproposal of:

Object#{foo(42) + 7: "bar"}
null#{foo(42) + 7: "bar"}
#{foo(42) + 7: "bar"}

where the first is equivalent to {[foo(42)+7]: "bar"}, the second is pure container (Object.create(null)) filled with properties, and the third is the default case, but I don't know which of the previous two - the first is probably less confusing, though the feels more clean.

# Herby Vojčík (8 years ago)

Herby Vojčík wrote:

OTOH, there could be consistent contraproposal of:

Object#{foo(42) + 7: "bar"}
null#{foo(42) + 7: "bar"}
#{foo(42) + 7: "bar"}

where the first is equivalent to {[foo(42)+7]: "bar"}, the second is pure container (Object.create(null)) filled with properties, and the third is the default case, but I don't know which of the previous two - the first is probably less confusing, though the feels more clean. ... though the second feels more ...

# Alexander Jones (8 years ago)

Agree with your concerns about symmetry with object literals, but many of the uses of maps benefit from having non string keys, and in such case generally everything would involve wrapping in []. Trying to use an Array as a key would be quite ugly with the extra squares required

const precachedResults = MyMap#{[[1, 2, 3]]: [1, 4, 9]}
vs
const precachedResults = MyMap#{[1, 2, 3]: [1, 4, 9]}

Perhaps a middle ground could be that if you want to use an expression that would otherwise be a bare word, you enclose in parens. The visual binding of the colon is deceptive anyway, so I tend to do this if the key expression contains a space:

MyMap#{1 + 2 + 3: 6}
vs.
MyMap#{(1 + 2 + 3): 6}

But I think I still prefer that the parsing for the key part is just standard expression evaluation, personally, and the POJSO literal barewords remain the only special case.

Indeed,

Object#{1: "one", Symbol(): "sym"}

Could Object-key-ify the keys, i.e. turn them into strings if not symbols, and Just Work (but a default implementation on the Object prototype is questionable!). That said I'm not sure we should be using Object for this kind of thing. At this point I don't know what a raw #{} should produce... There may be a better use case for it in the future, horrible ASI complexities notwithstanding.

Alex

# Alexander Jones (8 years ago)

Ah, there is actually a good case for keeping barewords in object literals but removing them from map literals, and that's due to objects accessing string properties as bare words, too. This is almost never the case for other map types.

const o = {foo: 42};
o.foo === 42;
o.bar = 43;

const m = Map#{"foo": 42};
m.get("foo") === 42;
m.set("bar", 43);

Would you agree?

# Mohsen Azimi (8 years ago)

When I look at Map#{"foo": 42} I don't see much difference with new Map([['foo', 42]]).

Since you can pass expressions there, it's already possible to do it with current syntax. There is only a bunch of extra brackets([ and ]) that I don't like.

# Michał Wadas (8 years ago)

Difference between any proposed here syntax and current state ( new Map([ [1,2], [2,3] ]); ) is..

3 characters + 2 characters/entry.

2015-10-28 17:22 GMT+01:00 Mohsen Azimi <me at azimi.me>:

# Dave Porter (8 years ago)

I don’t love any of the specific suggestions so far, but saving 3 + 2n keystrokes isn't the point – readability and learnability are. Visually, new Map([[‘foo’, 42]]) is a mess.

# Alexander Jones (8 years ago)

Also I would like to reiterate that errors in the shape of the N-by-2 array are only caught at runtime. That's really not ideal.

# Isiah Meadows (8 years ago)

Why not make it desugar to a direct function call with a single array of pairs? It's so parsed as a regular object, so shorthands can still be used.

Map#{foo: 1, bar: 2, 3: "baz"} Map[Symbol.fromHash]([[foo", 1], ["bar", 2], ["3", "baz]])

Object#{foo: 1, bar: 2, 3: "baz"} Object[Symbol.fromHash]([[foo", 1], ["bar", 2], ["3", "baz]])

Object.null#{foo: 1, bar: 2, 3: "baz"} Object.null[Symbol.fromHash]([[foo", 1], ["bar", 2], ["3", "baz]])

(bar doesn't have [[Construct]]) Object#{foo, bar() {}} Object[Symbol.fromHash]([[foo", foo], ["bar", function () {}]])

And as for implementation, use this:

extend class Map {
  static [Symbol.fromHash](pairs) {
    return new this(pairs);
  }
}

// etc...

function SetKeys(target, pairs) {
  for (const [key, value] of pairs) {
      target[key] = value
    }
    return target
}

extend class Object {
  static [Symbol.fromHash](pairs) {
    return SetKeys({}, pairs)
  }

  static null(pairs) {
    return SetKeys(Object.create(null), pairs)
  }
}

Pretty simple IMHO. A helper decorator could even be made.

# Alexander Jones (8 years ago)

I don't think borrowing object notation is a good idea. What exactly does

const myMap = Map#{
    get foo() { return 100; },
    set foo(v) {}
    constructor() {}
};

mean?

Honestly, a very significant portion of the use cases I have for actual maps don't involve string keys. So to borrow object notation and have to constantly write keys in [] is pretty naff:

const describe = Dict#{
    [1]: "one",
    [[1, 2]]: "array of 1 and 2",
    [null]: "the null value",
}; // please no!

If it makes people feel too weird to have comma separated, colon split key-value pairs within curlies that don't parse like POJSOs, we could have completely non-ambiguous parse with normal parentheses, I think?

const describe = Dict#(
    1: "one",
    [1, 2]: "array of 1 and 2",
    null: "the null value",
);

That might limit confusion while giving a syntactically clean way to define maps. Let's consider that a future mapping type like Dict compares non-primitive keys by abstract value instead of by reference identity. There are tonnes of nice use cases that open up that are taken for granted in other languages and other classes like Immutable.Map - we're not there yet with ES6 built-ins, so perhaps people might not yet appreciate the value of this.

To reiterate a previous point, object property access with a statically defined string key is idiomatically written obj.foo, so it makes sense for symmetry to have foo appear as a bareword in a literal defining obj = {foo: 42}. For most mapping-type classes this symmetry simply does not apply, and frankly neither should it.

Also, I specifically suggested that the consumed value is an ArrayIterator rather than an Array, because I feel having an intermediate Array around is placing too high an importance on the humble Array. If the implementation really wants an Array to work on internally, they can simply call Array.from with little cost. But if they want an Immutable.List they can have that instead without ever seeing an actual Array. (The Symbol.fromHash method is just Symbol.literalOf as I called it - same thing, modulo bikeshed.)

Alex

# Isiah Meadows (8 years ago)

That's something no one here really thought of yet. I don't personally have a lot of investment in this.

(@Alex and yes, much of that was based on your idea. It was a great starting point.)

# Tab Atkins Jr. (8 years ago)

On Thu, Oct 29, 2015 at 6:23 PM, Alexander Jones <alex at weej.com> wrote:

I don't think borrowing object notation is a good idea. What exactly does

const myMap = Map#{
    get foo() { return 100; },
    set foo(v) {}
    constructor() {}
};

mean?

Honestly, a very significant portion of the use cases I have for actual maps don't involve string keys. So to borrow object notation and have to constantly write keys in [] is pretty naff:

const describe = Dict#{
    [1]: "one",
    [[1, 2]]: "array of 1 and 2",
    [null]: "the null value",
}; // please no!

If it makes people feel too weird to have comma separated, colon split key-value pairs within curlies that don't parse like POJSOs, we could have completely non-ambiguous parse with normal parentheses, I think?

const describe = Dict#(
    1: "one",
    [1, 2]: "array of 1 and 2",
    null: "the null value",
);

That might limit confusion while giving a syntactically clean way to define maps. Let's consider that a future mapping type like Dict compares non-primitive keys by abstract value instead of by reference identity. There are tonnes of nice use cases that open up that are taken for granted in other languages and other classes like Immutable.Map - we're not there yet with ES6 built-ins, so perhaps people might not yet appreciate the value of this.

To reiterate a previous point, object property access with a statically defined string key is idiomatically written obj.foo, so it makes sense for symmetry to have foo appear as a bareword in a literal defining obj = {foo: 42}. For most mapping-type classes this symmetry simply does not apply, and frankly neither should it.

Also, I specifically suggested that the consumed value is an ArrayIterator rather than an Array, because I feel having an intermediate Array around is placing too high an importance on the humble Array. If the implementation really wants an Array to work on internally, they can simply call Array.from with little cost. But if they want an Immutable.List they can have that instead without ever seeing an actual Array. (The Symbol.fromHash method is just Symbol.literalOf as I called it - same thing, modulo bikeshed.)

I strongly agree with a lot of the points here, and think they suggest the OP's suggestion was generalized in slightly the wrong way. Producing a Map literal is indeed too specific to justify syntax, but what's suggested is not a special way of calling some constructors, but a literal syntax for 2-value iterators.

We have a literal syntax for 1-value iterators: just use an Array. It's lightweight (2 chars + 1 char per item), and typos in the syntax are caught at compile time. Our existing literal syntax for 2-value iterators (an array of length-2 arrays) fails at both of these: it's heavyweight (4 chars + 4 chars per item), and typos in the syntax are only caught at runtime, when it's actually iterated over.

Having a lightweight, compile-time-checked 2-value iterator literal that desugars to an N×2 Array (or ArrayIterator) fixes all these problems, and makes it easy to write Map literals, Immutable.Dict literals, or anything else. Using the hash-paren syntax suggested above:

let m =  Map(#(1: "one", [1, 2]: "array of 1 and 2", null: "the null value"));

There's no need to invent a new function-calling syntax or add a new well-known symbol to anything. It Just Works™ as long as the function you pass it to expects a 2-value iterator.

(From other languages, there doesn't appear to be any call for N×3 literals or anything higher

# Isiah Meadows (8 years ago)

With that syntax, I'm not even sure it's necessary. It's not much more concise than a list of 2-tuples. Don't quite see the benefit the other than a few characters.

# Alexander Jones (8 years ago)

I think you're overlooking the parse-time shape checking, Isiah, which in the new world order of type inference and checking seems like a necessity to me.

While I fully appreciate that Tab's solution involves the least number of specification additions, I still would rather write this without the extra pair of parens and the new which was omitted, just to cut down on the noise and really reduce the number of reasons people have to use Object instead of a more suitable type. Consider embedding lists as keys in a map which compares keys by value:

let m =  new Dict(#(1: "one", new List(#(1, 2)): "array of 1 and 2", null:
"the null value"));

I think this just starts to look quite gnarly, and the fewer occurrences of new I see in declarative code the better.

Re-consider my offered alternative, (putting #{} and #[] back on the table to visually distinguish abstract list and abstract mapping):

let m =  Dict#{1: "one", List#[1, 2]: "array of 1 and 2", null: "the null
value"};

let platonics = Set#["tetrahedron", "cube", "octahedron", "dodecahedron",
"icosahedron"];

I think it would also make sense to allow ... to include the results of iteration over some other iterable.

let nums = [2, 3, 4];
let setOfNums = Set#[1, 2, 3, ...nums, ...range(20, 50)];
# Francisco Tolmasky (8 years ago)

My one desire (perhaps not completely substantiated) would be that the “list” portion of the syntax exist outside of these expressions as well. That is to say, if:

new Dict#{a: b, c: d }

desugars to:

new Dict([[a,b], [c,d]])

Then I’d like/expect #{a: b, c: d} to also just desugar to [[a,b], [c,d]]. More than anything, I would just expect this to be the case if I ran into it in the wild. As such, perhaps one way to look at this is that what we actually want is a nice tuple syntax. Imagine if we had just a tuple syntax taht a->b desugars to [a,b]. Now, with no further changes, we get this “map

syntax” for free:

new Map([a->b, (4+5)->d, null->7])

That looks pretty good to me, and introduces a new operator that behaves the same everywhere, and has minimal requirements to “new” collection classes. Additionally, I would enjoy returning these as well when referring to key value pairs:

return (a->b) // sugar for return [a,b]
# Isiah Meadows (8 years ago)

I was actually talking about the syntax itself. I'd be fine with less verbose, extensible Map semantics.

# Isiah Meadows (8 years ago)

I like the idea of the #{} syntax working without a required type. But here's my opinions:

  1. It should automatically [[Construct]]. I don't see any other reason why it shouldn't.
  2. I don't like the idea of an -> operator which does that. Also, is a -> b -> c equivalent to [a, b, c] or [a, [b, c]]? I just don't like

it. 3. I can see why you want an untyped version to destructure into a list of 2-tuples, but I fear it would lead into the same pitfalls of the extreme consistency in this. Every ES5 function has its own this, which led people into problems when they forgot about it when using inner functions. Forth is one of the simplest, most consistent languages out there (it's an old concatenative, stack-based language, for those unfamiliar). But it's easy to trip up if your stack has the wrong value, it leaves one too many values on the stack, or you pass one too few values to it. And this fails silently.