ECMAScript 4 Netscape Proposal Rationale Operators

Thursday, May 22, 2003

User-overridable operators had originally been part of the ECMAScript 4 proposal but were deferred to a future version of the language in order to keep ECMAScript 4 small. If implemented, operator overriding might behave as described in this section.

Overview

ECMAScript 4 provides a number of operators such as such as +, *, a[b], and -=. These operators are dispatched based on the types of their arguments. Some operators are dispatched only on one operand, in which case the dispatch behaves like property lookup. Other operators are dispatched on two operands simultaneously. Finally, some operators are syntactic sugars for combinations of existing operators.

There are predefined definitions for all ECMAScript operators. In addition, a class may provide additional definitions by overriding the existing behavior. Such overrides apply only when at least one of the dispatched operands is an instance of that class or its descendants. Operator overriding is useful, for example, to implement units without having to add units to the core of the ECMAScript 4 language.

This section presents the expansions of operators in ECMAScript source code, the built-in definitions of those operators, and the means of overriding them.

Invoking Operators

When an operator is invoked in an expression, the corresponding operator is invoked according to the table below. The operator invocations use pseudo-syntax to denote which operator is called; only the expression syntax on the left side of the table can be used to invoke operators.

Operator Dispatch

Unary operators are single-dispatched just like regular methods. in is like a unary operator in that it can only be overridden by its right operand’s class, and it is single-dispatched by its right operand. Like in regular method dispatch, null is considered to be a member of only the types Null and Object.

Binary operator methods are double-dispatched based on the types of both operands. When one of the binary operators is invoked in an expression a  b, all applicable operator ""(x, y) methods are examined to find the most specific one, which is then called. A definition of operator ""(x:X, y:Y) is applicable to a  b if the value of a is a member of X and the value of b is a member of Y, except that, like in regular method dispatch, null is considered to be a member of only the types Null and Object. A definition of operator ""(x:X, y:Y) is most specific if it is applicable and if every applicable definition of operator ""(x:X', y:Y') satisfies X X' and Y Y'. If there is no most specific applicable definition of or if there is more than one most specific applicable definition of , then an error occurs when the expression a  b is evaluated.

One of the operands of an operator expression inside a class C can be a super expression, in which case the set of applicable definitions of the operator is restricted to those defined to take a parameter of type T in that position, where T is any proper ancestor of C. This way a method overriding an operator can invoke its superclass’s definition of that operator.

Expression Grammar Changes

To accommodate super expressions, the grammar rules for expressions would be changed as described below.

Super Expressions

The SuperExpression grammar rule is split into two to make FullSuperExpression into a separate rule:

super, which may only be used inside a class C, can be applied to a subexpression that evaluates to an instance v of C. That subexpression can be either a ParenExpression or omitted, in which case it defaults to this.

As specified in the grammar below, the SuperExpression must be embedded as an operand of one of the following operators:

• The left operand of a . (property lookup), [] (indexing), () (call), new, or new() operator (for the () and new() operators the SuperExpression must be a FullSuperExpression)
• The right operand of an in operator
• The operand of a unary +, -, ~, ++, or -- operator
• Either operand of a binary +, -, *, /, %, <<, >>, >>>, |, ^, &, <, >, <=, >=, ==, !=, ===, or !== operator
• Either operand of a +=, -=, *=, /=, %=, <<=, >>=, >>>=, |=, ^=, or &= operator

super changes the behavior of the operator in which it is embedded by limiting its property search to definitions inherited from class C’s superclass. See property lookup and operator dispatch.

Postfix Expressions

These grammar rules are modified to permit super expressions as operands of () (call), new, new(), and postfix ++ and --. Unmodified rules are not shown below.

Unary Operators

These grammar rules are modified to permit super expressions as operands of unary +, -, and ~ as well as prefix ++ and -- operators.

Multiplicative Operators

These grammar rules are modified to permit super expressions as operands of binary *, /, and % operators.

Additive Operators

These grammar rules are modified to permit super expressions as operands of binary + and - operators.

Bitwise Shift Operators

These grammar rules are modified to permit super expressions as operands of the <<, >>, and >>> operators.

Relational Operators

These grammar rules are modified to permit super expressions as operands of the <, >, <=, and >= operators and the right operand of an in operator.

Equality Operators

These grammar rules are modified to permit super expressions as operands of the ==, !=, ===, and !== operators.

Binary Bitwise Operators

These grammar rules are modified to permit super expressions as operands of the &, ^, and | operators.

Assignment Operators

These grammar rules are modified to permit super expressions as operands of the +=, -=, *=, /=, %=, <<=, >>=, >>>=, |=, ^=, and &= operators.

Built-In Operator Definitions

All operators have built-in definitions. These are summarized in the table below.

Operator Overriding

An operator can be overridden only inside a class C. That class should define a function with the operator attribute (operator is an additional member modifier attribute). The function’s name must be one of the strings in the table below. The name must be a string rather than an identifier; the FunctionName grammar rule is extended to allow strings:

A FunctionName can be a String only for operator overrides.

The table also defines the acceptable signatures that the operator can have. At least one parameter must be specified as taking a value of type C (a type that is a subclass of C is not acceptable). If the operator is ===, then both parameters must take values of type C. A class can define several different signatures of the same binary (but not unary) operator, differing in the choice of parameter types X or Y, which may be omitted if they are Object. The parameters denoted as taking types C, X, or Y in the table below cannot be optional or named parameters. The return type of an operator override can be any type.

The operator methods should return the result of the operator. For the ++ and -- operators, the result is the incremented or decremented value. The (), new, [], delete[], and []= operators take additional argument lists. If desired, these argument lists can include optional, named, or rest arguments. The []= operator must take one more initial required parameter, which is the value stored in the element.

The >, >=, !=, and !== operators cannot be overridden directly because they are syntactic sugars for invocations of the <, <=, ==, and === operators; override the latter operators instead. The !, ||, ^^, &&, and ?: operators cannot be overridden directly, but they are affected by any redefinition of toBoolean. The reason for not overriding ||, &&, and ?: is that they do not always evaluate all of their arguments, and programmers often take advantage of this short-circuiting behavior. ! is not overridable in order to preserve useful identities such as:

• if (!x) A else B if (x) B else A
• !(x && y) !x || !y
• x != y !(x == y)

Definitions of operators cannot specify a namespace. No access restrictions can be imposed on operators.

Example

The following class illustrates overriding a few operators.

class Complex {
  var x:Number, y:Number;

  operator function "+"(a:Complex):Complex {
    return a;
  }

  operator function "-"(a:Complex):Complex {
    return new Complex(x: -a.x, y: -a.y);
  }

  operator function "+"(a:Complex, b:Complex):Complex {
    return new Complex(x: a.x+b.x, y: a.y+b.y);
  }

  operator function "+"(a:Complex, b:Number):Complex {
    return new Complex(x: a.x+b, y: a.y);
  }

  operator function "+"(a:Number, b:Complex):Complex {
    return new Complex(x: a+b.x, y: b.y);
  }

  function toBoolean():Boolean {
    return this.x != 0 || this.y != 0;
  }
}

toBoolean Method

Just like toString, every object has a toBoolean method that can be overridden. This method is called by statements such as if, while, do while, and for and operators such as !, ||, ^^, &&, and ? :. This method should return either true or false.

For-In Operator

An object x’s class can override the meaning of for-in loops such as for (v in x) by overriding the following three methods which are located in the system-defined namespace Iterator:

Rationale

A few other approaches had been considered for overriding the for-in operator. The three-method approach above was chosen for the following reasons:

• Why use an iterator object instead of a function and a callback (i.e. have Iterator::forIn take a parameter f and have it call f for each element in the collection)? The reason is that the iterator model permits future extensions such as iterating through two collections concurrently. This cannot be done using callbacks.
• Why return an object from Iterator::forIn and Iterator::next instead of having Iterator::forIn and Iterator::next return just an iterator and having a separate Iterator::get method that, given an iterator, returns an element? The rationale here is that the separate Iterator::get method would introduce unavoidable concurrency hazards. Since null (or any other value) can be a member of a collection, we can’t use the return value from Iterator::get to detect the end of iteration; instead, we’d have to use a null return value from Iterator::forIn or Iterator::next to indicate the end of iteration. In such a case, however, it would be possible for another thread to delete the last element between the call to Iterator::forIn or Iterator::next and the call to Iterator::get, and Iterator::get would then be stuck. We could use an exception to indicate the end of iteration, but this would put exception handling on the fast path of program execution.
• Why use an object with two named slots rather than a two-element read-only array for the return value? Named slots can have heterogeneous types, while the types of array elements must be the same.
• Why define Iterator::done? Some objects may erect temporary scaffolding to support iteration. This method informs them that they can take down such scaffolding. Since ECMAScript currently lacks finalization, there is no other way for the iterated object to know that the iteration aborted early — the start of another iteration cannot be used as an indication that the previous iteration completed because several iterations can be happening simultaneously on the same object.

Class Reflection Operator

Earlier drafts of this proposal had a .class operator that returned an instance’s class. This operator was defined using the grammar production:

The operator was removed because obtaining an instance’s class breaks abstraction. The client of a factory method m could determine whether m was returning instances of some stated class C or subclasses of C, and m could not evolve to alter its implementation decision of what precise instances to return without potentially breaking clients.

Waldemar Horwat Last modified Thursday, May 22, 2003