Skip to content

Latest commit

 

History

History
317 lines (235 loc) · 17.3 KB

PRIVATE_SYNTAX_FAQ.md

File metadata and controls

317 lines (235 loc) · 17.3 KB
Raw

What's this about #?

# is the new _. Whereas

class A {
  _hidden = 0;
  m() {
    return this._hidden;
  }
}

creates a class with a field which is hidden by convention, but in fact is fully public,

class B {
  #hidden = 0;
  m() {
    return this.#hidden;
  }
}

creates a class with a field which is actually private to the class, with privacy enforced by the language.

We can't just change how _ works for web compatibility reasons.

Why isn't access this.x?

Having a private field named x must not prevent there from being a public field named x, so accessing a private field can't just be a normal lookup.

This is only an issue in JavaScript because of its lack of static types. Statically typed languages use type declarations to distinguish the external-public/internal-private cases without the need of a sigil. But a dynamically typed language doesn't have enough static information to differentiate those cases.

Why does this proposal allow a class to have a private field #x and a public field x?

  1. If private fields conflicted with public fields, it would break encapsulation; see below.

  2. One of the more important properties of private fields is that subclasses don't need to know about them. We'd like to allow a subclass to define a property named x even if its superclass has a private field of that name.

This is something other languages with private fields generally allow. E.g., the following is perfectly legal Java:

class Base {
  private int x = 0;
}

class Derived extends Base {
  public int x = 0;
}

Why not do a runtime check on the type of the receiver to determine whether to access the private or public field named x?

Property access semantics are already complicated, and we don't want to slow down every property access just to add this feature.

It also would allow methods of the class to be tricked into operating on public fields of non-instances as if they were private fields of instances. See this comment for an example.

Why not just always have obj.x refer to a private field inside of a class which declares a private field x?

Class methods often manipulate objects which are not instances of the class. It would be surprising if the code obj.x suddenly stopped referring to public field x of obj, when obj is not expected to be an instance of the class, just because that code happened to occur somewhere within a class which declares a private field named x, possibly deep within said class.

Why not give the this keyword special semantics?

this is already a source of enough confusion in JS; we'd prefer not to make it worse. Also, it's a major refactoring hazard: it would be surprising if this.x had different semantics from const thiz = this; thiz.x.

This also wouldn't allow accessing fields of objects other than this.

Why not only allow accessing fields of this, e.g. by just having a bare x refer to the private field x?

It is a goal of this proposal to allow accessing private fields of other instances of the same class (see below), which requires some syntax. Also, using bare identifiers to refer to properties is not the usual JS way (with the exception of with, which is generally considered to have been a mistake).

Why doesn't this['#x'] access the private field named #x, given that this.#x does?

  1. This would complicate property access semantics.

  2. Dynamic access to private fields is contrary to the notion of 'private'. E.g. this is concerning:

class Dict extends null {
  #data = something_secret;
  add(key, value) {
    this[key] = value;
  }
  get(key) {
    return this[key];
  }
}

(new Dict).get('#data'); // returns something_secret

But doesn't giving this.#x and this['#x'] different semantics break an invariant of current syntax?

Not exactly, but it is a concern. this.#x has never previously been legal syntax, so from one point of view there can be no invariant regarding it.

On the other hand, it might be surprising that they differ, and this is a downside of the current proposal.

Why not have access be this#x, without the dot?

This would work, but would be an ASI hazard if we later introduced a shorthand syntax of #x to mean this.#x.

More generally, the committee has mostly come down on the side of this.# as more clearly implying field access.

Why aren't declarations private x?

This sort of declaration is what other languages use (notably Java), and implies that access would be done with this.x. Assuming that isn't the case (see above), in JavaScript this would silently create or access a public field, rather than throwing an error. This is a major potential source of bugs or invisibly making public fields which were intended to be private.

It also allows a symmetry between declaration and access, just as there is for public fields:

class A {
  pub = 0;
  #priv = 1;
  m() {
    return this.pub + this.#priv;
  }
}

Why does this proposal allow accessing private fields of other instances of the same class? Don't other languages normally forbid that?

It's very useful: see e.g. the equals method in the Point example in the readme. And in fact, other languages allow it for the same reason; e.g. the following is perfectly legal Java:

class Point {
	private int x = 0;
	private int y = 0;
	public boolean equals(Point p) { return this.x == p.x && this.y == p.y; }
}

Why was the sigil # chosen, among all the Unicode code points?

No one came out and said, # is the most beautiful, intuitive thing to indicate private state. Instead, it was more of a process of elimination:

  • @ was the initial favorite, but it was taken by decorators. TC39 considered swapping decorators and private state sigils, but the committee decided to defer to the existing usage of transpiler users.
  • _ would cause compatibility issues with existing JavaScript code, which has allowed _ at the start of an identifier or (public) property name for a long time.
  • Other characters which could be used as infix operators, but not prefix operators, are hypothetically possible, such as %, ^, &, ?, given that our syntax is a bit unique--x.%y is not valid currently, so there would be no ambiguity. However, the shorthand would lead to ASI issues, e.g., the following would look like using the infix operator:
class Foo {
  %x;
  method() {
    calculate().my().value()
    %x.print()
  }
}

Here, the user likely intended to call the print method on this.%x, but instead, the mod operator will be used!

  • Other Unicode code points which are not in ASCII or IDStart could be used, but these might be hard to input for many users; they are not found on common keyboards.

In the end, the only other options are longer sequences of punctuation, which seems suboptimal compared to a single character.

Why not use private.x to refer to a private field of this, and private(that).x to refer to a private field of another object?

There are a few issues with this approach:

  • If declarations are private x;, programmers may be tempted to access fields as this.x rather than private.x; if they do this for both setting and getting the private field (e.g., if there's no initializer), their code will "just work.
  • private(that) looks like an expression on its own, but this leads to problems:
    • If private(that) evaluated to an object that had the private fields in it, it would be difficult to optimize away the additional, separate allocation of that object.
    • if private(that) were a syntax error by itself, and forced to have a property access after it, this would be a surprising break from how functions and pseudo-functions (like import()) work in JavaScript.

Why doesn't this proposal allow some mechanism for reflecting on / accessing private fields from outside the class which declares them (e.g. for testing)? Don't other languages normally allow that?

Doing so would violate encapsulation (see below). That other languages allow it isn't sufficient reason on its own, especially since in some of them (e.g. C++) this is accomplished by modifying memory directly and is not necessarily a goal.

What do you mean by "encapsulation" / "hard private"?

It means that private fields are purely internal: no JS code outside of a class can detect or affect the existence, name, or value of any private field of instances of said class without directly inspecting the class's source, unless the class chooses to reveal them. (This includes subclasses and superclasses.)

This means that reflection methods like getOwnPropertySymbols must not reveal private fields.

This also means that if a class has a private field named x, code outside the class which does obj.x on an instance obj of the class should access the public field x just as it would in the absence of the private field. It must not access the private field or throw an error. Note that this doesn't come up as much in languages like Java, which can be type-checked at compile time and do not allow dynamic access to fields by name except via reflection APIs.

Why is encapsulation a goal of this proposal?

  1. Library authors have found that their users will start to depend on any exposed part of their interface, even undocumented parts. They do not generally consider themselves free to break their user's pages and applications just because those users were depending upon some part of the library's interface which the library author did not intend them to depend upon. As a consequence, they would like to have hard private state to be able to hide implementation details in a more complete way.

  2. While it's already possible to model true encapsulation using either per-instance closures or WeakMaps (see below), both approaches are unergonomic to use with classes and have memory usage semantics which may be surprising. Furthermore, per-instance closures don't allow instances of the same class to access each other's private fields (see above), and with WeakMaps there's a non-obvious risk of exposing private data. WeakMaps are also likely to be slower than private fields could be.

  3. "Hidden" but not encapsulated properties are likewise already possible through the use of Symbols as property names (see below).

This proposal is currently moving forward with hard-private fields, letting decorators or other mechanisms provide classes an opt-in escape hatch. We intend to gather feedback during this process to help determine whether this is the correct semantics.

See this issue for more.

How can you model encapsulation using WeakMaps?

const Person = (function(){
  const privates = new WeakMap;
  let ids = 0;
  return class Person {
    constructor(name) {
      this.name = name;
      privates.set(this, {id: ids++});
    }
    equals(otherPerson) {
      return privates.get(this).id === privates.get(otherPerson).id;
    }
  }
})();
let alice = new Person('Alice');
let bob = new Person('Bob');
alice.equals(bob); // false

There's a potential pitfall with this approach, though. Suppose we want to add a custom callback to construct greetings:

const Person = (function(){
  const privates = new WeakMap;
  let ids = 0;
  return class Person {
    constructor(name, makeGreeting) {
      this.name = name;
      privates.set(this, {id: ids++, makeGreeting});
    }
    equals(otherPerson) {
      return privates.get(this).id === privates.get(otherPerson).id;
    }
    greet(otherPerson) {
      return privates.get(this).makeGreeting(otherPerson.name);
    }
  }
})();
let alice = new Person('Alice', name => `Hello, ${name}!`);
let bob = new Person('Bob', name => `Hi, ${name}.`);
alice.equals(bob); // false
alice.greet(bob); // === 'Hello, Bob!'

At first glance this appears fine, but:

let mallory = new Person('Mallory', function(name) {this.id = 0; return `o/ ${name}`;});
mallory.greet(bob); // === 'o/ Bob'
mallory.equals(alice); // true. Oops!

How can you provide hidden but not encapsulated properties using Symbols?

const Person = (function(){
  const _id = Symbol('id');
  let ids = 0;
  return class Person {
    constructor(name) {
      this.name = name;
      this[_id] = ids++;
    }
    equals(otherPerson) {
      return this[_id] === otherPerson[_id];
    }
  }
})();
let alice = new Person('Alice');
let bob = new Person('Bob');
alice.equals(bob); // false

alice[Object.getOwnPropertySymbols(alice)[0]]; // == 0, which is alice's id.

How does this proposal relate to brand checks?

This proposal only adds private fields to objects during constructors, and it includes errors for accessing private fields that don't exist. Some people have ascribed the term "brand check" to the combination of these properties, but there has been some confusion in the community about exactly which guarantees are provided. To clarify, the only guarantee that you have is that, if a private field declared in a particular class exists on an instance, then that class's constructor has run against that instance.

For example, given the following class definitions:

class F { #f; checkF() { this.#f; } }
class G extends F { #g; checkG() { this.#g; } }

If an object obj doesn't throw in response to F.prototype.checkF.call(obj), we know that obj was created from a new F call.

By contrast, due to mutable prototype chains, we don't have all of that information about an object which doesn't throw from a call to G.prototype.checkG.call(obj)--all we know in this case was that, at some point, new G was called, and the super constructor returned obj. For example, this could be due to a clever case taking advantage of what's called the "super return trick":

let obj = { };
Object.setPrototypeOf(G, class { constructor() { return obj; });
new G;
Object.setPrototypeOf(G, F);
G.prototype.checkG.call(obj);  // doesn't throw
F.prototype.checkF.call(obj);  // throws

This weird phantom object obj has G's "brand" and not F's! This is not considered "broken", it's just that the nature of dynamic prototype chains for classes gives programmers the flexibility to do this sort of thing. Sometimes, it's more useful to have predictability than flexibility.

This pattern can be prevented by preventing mutations to the prototype chain which is dynamically queried when constructing objects (as part of ES2015). For example, you can call Object.preventExtensions(G) to ensure that G's superclass remains F. In this case, an object which doesn't throw in response to checkG will also have a private field #f.

How do private fields interact with Proxy?

Private fields of a Proxy target aren't accessible from the Proxy itself. For example, the following will throw a TypeError:

class K { #x; get x() { return this.x; } }
let k = new Proxy(new K, { });
k.x  // TypeError

This behavior is analogous to internal slots and WeakMaps: the private field in the target is not forwarded through the Proxy. Although this behavior doesn't meet some hopes for Proxy serving as a transparent way to observe operations against objects, the no-forwarding behavior enables the following:

  • Private fields can be used for polyfills matching the behavior of internal slots, without deviation in behavior with respect to Proxy.
  • The Proxy target is encapsulated, without uncontrolled leaking knowledge about the target to the consumer of the Proxy, and without leaking any information about the private field to the Proxy traps.
  • Semantics are analogous to WeakMap, giving programmers a simple mental model (everything is either a property or a WeakMap) rather than a third alternative.

Conversely, Proxy objects can have private fields directly defined on them using the "super return trick", as in the following example:

class Super { constructor(x) { return x; } }
class Trick extends Super {
  #x;
  checkX() { this.#x; }
}

let target = { };
let proxy = new Proxy(target, { });
new Trick(proxy);
Trick.prototype.checkX(proxy);   // No exception thrown
Trick.prototype.checkX(target);  // TypeError

One way to handle the lack of forwarding to Proxy targets is to use a membrane pattern, e.g., as in Salesforce's Observable Membrane.