Why :: need to be used while defining?

[LongTengDao]

I saw the proposal is const ::toArray=... now.

Why not const toArray=... simply, like old proposal did?

[hax]

First, only binary form need defining first, or import ::{toArray} from ....

The design goal is making extension (virtual) methods/properties behave as close as normal (real) methods/properties. The only difference between them is how to lookup the methods/properties.

• o.foo(...args) Normal methods: lookup method foo in the object o
• o::foo(...args) Binary form: lookup method foo in the defined local extension methods
• o::X:foo(...args) Ternary form: lookup method foo in the object X

When get the method, all invoke the method with the receiver o and arguments args.

Note foo in o.foo not use the local name of foo. So to behave as close as o.foo, o::foo should also use separate lexical namespace, or it will cause some developer experience problem.

One such problem of old proposal (lookup foo from normal namespace, aka. const foo=...; o::foo()) is, it's very common to write let json = obj.json(), but let json = obj::json() doesn't work. This make the refactoring between normal methods and extension methods trouble. There is also shadow problem (json may be shadowed by accidents), See also #4 for similar discussion. Note in many cases, extension methods would use common words like first, last, max, etc. and have very high possibility to be shadowed in local scope.

Another benefit of defining first is it could give error in more early stage.

https://www.smashingmagazine.com/2018/10/taming-this-javascript-bind-operator/ have a good example:

const plus = (x) => this + x;
...
1::plus(1) // expect 2 but got "[object Window]1" or NaN or  possible some other values

With this new proposal, we could throw type error in first place.

Of coz, it seems forcing defining is not as convenience as old proposal, but consider the common usage is importing extension methods from modules, with import ::{first, last} from 'mod' syntax, it have the same ergonomics .

[hax]

Note, I understand it's not common to have separate namespace in JS, I'm still investigating the other options which mitigate the problem.

One possibility is remove const ::x = ... syntax at all, so the only two syntax will be

• import ::{foo, bar} from "module"
• const ::{foo, bar} from extensionObject

which seems more consistent and simpler.

It still in separate namespace but may be more acceptable...

[hax]

Another point, if we really want applying method from a local binding directly, we could have syntax like o::[foo] or o::(foo). Especially o::(foo) is somewhat compatible with the old proposal. But:

1. It's not easy to read code like o::(function square() { return this * this }), I feel most people will eventually write const square = ...; o::(square). From the ergonomic view , there is no much difference with const ::square = ...; o::square.
2. o :: (function () {...}) make it more like pipeline op in fp, but o :: (x => x * x) is wrong, unless we automatically convert arrow functions and use the first param as receiver, but how about function (x) { return x * x }? It seem user expectations will be broken in various manner whichever we choose. This is why that part of old proposal is dead and left to pipeline proposal (though pipeline have their own issues).
3. Use cases like arrayLike::(Array.prototype.map)(x => x * x) should have more ergonomics form, this is one of the reasons of introducing ternary form, so we can write arrayLike::Array:map(x => x * x).
4. As we see, foo in o::(foo) should be a "free method" (an unbound function contains dynamic this outside of classes or object literals) or accessor in property descriptor-like form. It seems both are uncommon in daily codes (especially very rare in exports) and may be confusing to average programmers. (This is also one of the reasons why I'm considering removing const ::foo = ... syntax at all.)

[ljharb]

I’m not sure i see the case for a separate namespace at all. I don’t recall the old proposal being considered to have any problems related to o::json() or similar. Any kind of separate namespace is, to me, a bit of a nonstarter.

[hax]

@ljharb I hope I could express the ideas clearer:

High-level goal: make extension methods behave as close as real methods in syntax, semantic and developer experiences; eliminate the barrier and troubles of using extensions as possible as we can.

Old proposal issues which affect this goal:

1. Refactoring hazard of let json = o.json() <=> let json = o::json() or let size = collection.size <=> let size = collection::size or let max = Math.max(...data) <=> let max = data::max() or many other possible cases.
2. Easy to be shadowed by accidents because extensions tend to use short names and common words. This could be mitigated by linter, but adopting linter and specific rules always have big ecosystem cost, and no-shadow rule is not enabled by default as eslint recommended rules.
3. If use bindings directly, they should be "free methods", and it's uncommon and potentially confusing to average users. This proposal solve this problem by introducing a importing syntax like import ::{last} from 'lodash' which will auto convert the functions to extension methods, avoid such pitfalls and could leverage the ecosystem in the best. Technically such converted last method could still in same namespace, but actually not useful at all, and only cause problems if it wrongly used in other form.

[ljharb]

I'm not clear on how that's a refactoring hazard.

"Adopting a linter" is not really a cost; everyone who's going to use one likely already has one - but shadowing is also a normal part of the language, and it's often desired to shadow.

Why does a function need conversion? function () { return this; } should already be an extension method, as should Array.prototype.slice.

[hax]

The refactoring hazard is programmers are forced to either rename the local variable or rename the extension method name. And naming is hard.

"Adopting a linter" (and adopting special rules) is always a cost. But I don't want to argue this point with you again and again.

but shadowing is also a normal part of the language, and it's often desired to shadow.

I somewhat agree. This is why people often override and disable no-shadow rule when they use some coding style (like airbnb eslint config) which enable no-shadow by default.

The point here is, when someone use let size = 10, it's very impossible they desired to shadow ::size extension method. And vice versa.

Why does a function need conversion

Because most libraries provide util functions like export function last(a) { return a[a.length - 1] } not free methods like export function last() { return this[this.length - 1] }. But people want to use the util functions like extension methods: let last = a::last().

[ljharb]

I don't understand why that's something we'd want to support. Methods used as if they're in a chain should be ones written to be used as such - just like decorators aren't designed to be used with functions that weren't written to be decorators.

A user who wants that can function last() { return utilLast(this); }; the language shouldn't try to handle that (nor assume that such utility functions will always use the first argument).

[hax]

Methods used as if they're in a chain should be ones written to be used as such

Actually I agree that.

The point here is , actually there is no much differences between these two last function, they have same purpose and semantic except one use this receiver, one use first argument. If you look inside lodash, most util functions are match that pattern.

The reason why lodash (and other util libs) is written like that is just because the language do NOT have extension methods. Library authors are forced to provide the functions in util function form.

So if we introduce extensions in language level, it would be better to levage the assets if possible, not force every libraries write the same conversion logic, and publish a new versions , and users have to upgrade to new versions to use the feature.

[ljharb]

I don't think that would be better. I think that libraries like lodash would either provide forms that use this, or, would provide their own adapter method.

[hax]

This question would be better answered by library maintainers. But I at least see some factors :

• require publishing new versions even the conversion is very trivial
• runtime conversion would have perf cost
• do we need publish two package for each form? Note in a big project, there may be someone use extensions someone use utils form. so such policy increase the deploy complexity of application devs.

Note every library need to repeat that. And also there are tons of libraries have no active maintaining.

[ljharb]

As a library maintainer, I would prefer that functions I write to not use this not be suddenly used in ways I did not intend, and to suddenly expose me to a flood of new bug reports as a result.

You don't need two packages, but you could easily have one package with two entry points.

I've used Lodash via "lodash-bound" before and I thought it should be considered when evaluating how 3rd party libraries can adapt to this operator. I used it against the same babel implementation of the :: operator that we've always had. I'm not at a desktop or I might go into greater detail. :)

[hax]

suddenly expose me to a flood of new bug reports

Could you explain what bug reports will occur? I hope we can try to analyze and avoid them.

[ljharb]

Any time someone uses a package in a way the author didn’t intend, there’s a likelihood of new bugs. An automatic adapter is very unlikely to be able to handle all edge cases, including “the second argument is the receiver”, or “it’s sloppy mode but i don’t expect the global object” or “it’s strict mode but i do expect the global object” or “it’s sloppy mode but i do expect a primitive receiver” or the inverse, etc.