ECMAScript 4 Netscape Proposal Rationale Execution Model

Tuesday, September 10, 2002

This page is somewhat out of date.

Introduction

When does a declaration (of a value, function, type, class, method, pragma, etc.) take effect? When are expressions evaluated? The answers to these questions distinguish among major kinds of programming languages. Let’s consider the following function definition in a language with C++ or Java-like syntax:

gadget f(widget x) {
  if ((gizmo)(x) != null)
    return (gizmo)(x);
  return x.owner;
}

In a static language such as Java or C++, all type expressions are evaluated at compile time. Thus, in this example widget and gadget would be evaluated at compile time. If gizmo were a type, then it too would be evaluated at compile time ((gizmo)(x) would become a type cast). Note that we must be able to statically distinguish identifiers used for variables from identifiers used for types so we can decide whether (gizmo)(x) is a one-argument function call (in which case gizmo would be evaluated at run time) or a type cast (in which case gizmo would be evaluated at compile time). In most cases, in a static language a declaration is visible throughout its enclosing scope, although there are exceptions that have been deemed too complicated for a compiler to handle such as the following C++:

typedef int *x;

class foo {
  typedef x *y;
  typedef char *x;
}

Many dynamic languages can construct, evaluate, and manipulate type expressions at run time. Some dynamic languages (such as Common Lisp) distinguish between compile time and run time and provide constructs (eval-when) to evaluate expressions early. The simplest dynamic languages (such as Scheme) process input in a single pass and do not distinguish between compile time and run time. If we evaluated the above function in such a simple language, widget and gadget would be evaluated at the time the function is called.

Challenges

ECMAScript is a scripting language. Many programmers wish to write ECMAScript scripts embedded in web pages that work in a variety of environments. Some of these environments may provide libraries that a script would like to use, while on other environments the script may have to emulate those libraries. Let’s take a look at an example of something one would expect to be able to easily do in a scripting language:

Bob is writing a script for a web page that wants to take advantage of an optional package MacPack that is present on some environments (Macintoshes) but not on others. MacPack provides a class HyperWindoid from which Bob wants to subclass his own class BobWindoid. On other platforms Bob has to define an emulation class BobWindoid' that is implemented differently from BobWindoid — it has a different set of private methods and fields. There also is a class WindoidGuide in Bob’s package; the code and method signatures of classes BobWindoid and BobWindoid' refer to objects of type WindoidGuide, and class WindoidGuide’s code refers to objects of type BobWindoid (or BobWindoid' as appropriate).

Were ECMAScript to use a dynamic execution model (described below), declarations take effect only when executed, and Bob can implement his package as shown below.

class WindoidGuide; // forward declaration

if (onMac()) {
  import "MacPack";

  global class BobWindoid extends HyperWindoid {
    private var x;
    var g:WindoidGuide;

    private function speck() {...};
    public function zoom(a:WindoidGuide, uncle:HyperWindoid = null):WindoidGuide {...};
  }
} else {
  // emulation class BobWindoid'
  global class BobWindoid {
    private var i:Integer, j:Integer;
    var g:WindoidGuide;

    private function advertise(h:WindoidGuide):WindoidGuide {...};
    private function subscribe(h:WindoidGuide):WindoidGuide {...};
    public function zoom(a:WindoidGuide):WindoidGuide {...};
  }
}

class WindoidGuide {
  var currentWindoid:BobWindoid;

  function introduce(arg:BobWindoid):BobWindoid {...};
}

On the other hand, if the language were static (meaning that types are compile-time expressions), Bob would run into problems. How could he declare the two alternatives for the class BobWindoid?

Bob’s first thought was to split his package into three HTML SCRIPT tags (containing BobWindoid, BobWindoid', and WindoidGuide) and turn one of the first two off depending on the platform. Unfortunately this doesn’t work because he gets type errors if he separates the definition of class BobWindoid (or BobWindoid') from the definition of WindoidGuide because these classes mutually refer to each other. Furthermore, Bob would like to share the script among many pages, so he’d like to have the entire script in a single BobUtilities.js file.

Note that this problem would be newly introduced by ECMAScript 4 if it were to evaluate type expressions at compile time. ECMAScript 3 does not suffer from this problem because it does not have a concept of evaluating an expression at compile time, and it is relatively easy to conditionally define a class (which is merely a function) by declaring a single global variable g and conditionally assigning either one or another anonymous function to it.

There exist other alternatives in between the dynamic execution model and the static model that also solve Bob’s problem. One of them is described at the end of this chapter.

Dynamic Execution Model

In a pure dynamic execution model the entire program is processed in one pass. Declarations take effect only when they are executed. A declaration that is never executed is ignored. Scheme follows this model, as did early versions of Visual Basic.

The dynamic execution model considerably simplifies the language and allows an interpreter to treat programs read from a file identically to programs typed in via an interactive console. Also, a dynamic execution model interpreter or just-in-time compiler may start to execute a script even before it has finished downloading all of it.

One of the most significant advantages of the dynamic execution model is that it allows ECMAScript 4 scripts to turn parts of themselves on and off based on dynamically obtained information. For example, a script or library could define additional functions and classes if it runs on an environment that provides a CSS unit arithmetic library while still working on environments that do not.

The dynamic execution model requires identifiers naming functions and variables to be defined before they are used. A use occurs when an identifier is read, written, or called, at which point that identifier is resolved to a variable or a function according to the scoping rules. A reference from within a control statement such as if and while located outside a function is resolved only when execution reaches the reference. References from within the body of a function are resolved only after the function is called; for efficiency, an implementation is allowed to resolve all references within a function or method that does not contain eval at the first time the function is called.

According to these rules, the following program is correct and would print 7:

function f(a:Integer):Integer {
  return a+b;
}

var b:Integer = 4;
print(f(3));

Assuming that variable b is predefined by the host if featurePresent is true, this program would also work:

function f(a:Integer):Integer {
  return a+b;
}

if (!featurePresent) {
  var b:Integer = 4;
}

print(f(3));

On the other hand, the following program would produce an error because f is referenced before it is defined:

print(f(3));

function f(a:Integer):Integer {
  return a*2;
}

Defining mutually recursive functions is not a problem as long as one defines all of them before calling them.

Hybrid Execution Model

ECMAScript 3 does not follow the pure dynamic execution model, and, for reasons of compatibility, ECMAScript 4 strays from that model as well, adopting a hybrid execution model instead. Specifically, ECMAScript 4 inherits the following static execution model aspects from ECMAScript 3:

• Variable declarations of variables at the global scope cause the variables to be created at the time the program is entered rather than at the time the declaractions are evaluated.
• Variable declarations of local variables inside a function cause the variables to be created at the time the function is entered rather than at the time the declaractions are evaluated.
• Unless specified otherwise, declarations of functions at the global scope cause the functions to be created at the time the program is entered rather than at the time the declaractions are evaluated.
• Unless specified otherwise, declarations of functions nested inside a function cause the functions to be created at the time the function is entered rather than at the time the declaractions are evaluated.

In addition to the above, the evaluation of class declarations has special provisions for delayed evaluation to allow mutually-referencing classes.

The second condition above allows the following program to work in ECMAScript 4:

const b:String = "Bee";

function square(a:Integer):Integer {
  b = a;   // Refers to local b defined below, not global b
  return b*a;
  var b:Integer;
}

While allowed, using variables ahead of declaring them, such as in the above example, is considered bad style and may generate a warning.

The third condition above makes the last example from the pure execution model section work:

print(f(3));

function f(a:Integer):Integer {
  return a*2;
}

Again, actually calling a function at the top level before declaring it is considered bad style and may generate a warning. It also will not work with classes.

Compiling The Dynamic Execution Model

Perhaps the easiest way to compile a script under the dynamic execution model is to accumulate function definitions unprocessed and compile them only when they are first called. Many JITs do this anyway because this lets them avoid the overhead of compiling functions that are never called. This process does not impose any more of an overhead than the static model would because under the static model the compiler would need to either scan the source code twice or save all of it unprocessed during the first pass for processing in the second pass.

Compiling a dynamic execution model script off-line also does not present special difficulties as long as eval is restricted to not introduce additional declarations that shadow existing ones (if eval is allowed to do this, it would present problems for any execution model, including the static one). Under the dynamic execution model, once the compiler has reached the end of a scope it can assume that that scope is complete; at that point all identifiers inside that scope can be resolved to the same extent that they would be in the static model.

Conditional Compilation Alternative

Bob’s problem could also be solved by using conditional compilation similar in spirit to C’s preprocessor. If we do this, we have to ask about how expressive the conditional compilation meta-language should be. C’s preprocessor is too weak. In ECMAScript applications we’d often find that we need the full power of ECMAScript so that we can inspect the DOM, the environment, etc. when deciding how to control compilation. Besides, using ECMAScript as the meta-language would reduce the number of languages that a programmer would have to learn.

Here’s one sketch of how this could be done:

• Execution of a script occurs in a series of passes. The last one is called run time. It is preceded by a compile time pass, which is preceded (if necessary) by a compile compile time pass, which is preceded (if necessary) by a compile compile compile time pass, etc. Having anything beyond a compile compile time pass is very rare, but rapid-development tools could generate such code.
• All definitions made in a previous pass (such as compile time) are also visible in a later pass (such as run time). If we define a type to also be a function that casts its argument to it type, this allows an implementation to not worry about whether (x)(y) is a function call of function x or a cast of y to type x.
• TypeExpressions are evaluated at compile time. These appear as types of variables, constants, and fields; argument and result types of functions and methods; and superclasses.
• Declarations are accumulated at compile time. Variables are declared, but variable initializers are not evaluated. Functions and classes are declared and defined at compile time.
• One can lift a Block from run time to compile time by preceding the block with a # symbol. For example, #{var x:int = 3} defines a compile-time constant x and initializes it to 3. One can also lift a var, const, or function declaration directly by preceding it with a # symbol, so #var x:int = 3; would accomplish the same thing.
• Any TypeExpressions in blocks or statements lifted to compile time are evaluated at compile compile time. int in the preceding example is such a TypeExpression.
• Any block or statement that is lifted twice (as in #{#var x:int = 3}) is evaluated at compile compile time, and so forth.
• The compile time pass can selectively exclude source code statements from the run time pass using the construct below. If the first Expression is true, the first Statements list is included in the run time pass. Otherwise, if the second Expression is true, the second Statements list is included in the run time pass, and so on.
  
  # if ( Expression ) Statements [# else if ( Expression ) Statements] ... [# else Statements] # end if
• Unlike in C, the Statements above must have balanced parentheses, braces, brackets, etc. Also, unlike in C, preprocessing is not line-oriented; #’s can appear anywhere on a line.
• The compile compile time pass can use #if to conditionally exclude compile time code, etc.

Note that because variable initializers are not evaluated at compile time, one has to use #var a = int rather than var a = int to define an alias a for a type name int.

This sketch does not address many issues that would have to be resolved, such as how typed variables are handled after they are declared but before they are initialized (this problem doesn’t arise in the dynamic execution model), how the lexical scopes of the run time pass would interact with scoping of the compile time pass, etc.

Comparing the Dynamic Execution Model with Conditional Compilation

Both approaches solve Bob’s problem, but they differ in other areas. In the sequel "conditional compilation" refers to the conditional compilation alternative described above.

• The dynamic execution model is simpler to describe and analyze.
• The dynamic execution model is easier to learn for non-programmers.
• Conditional compilation is more familar to C/C++ programmers.
• Conditional compilation requires either a control meta-language or facilities for lifting execution from run to compile time. These often have unintended consequences (compile-time evaluation does not follow the same scopes as run-time evaluation) and have been a significant source of problems in Common Lisp.
• Conditional compilation can more faithfully emulate ECMAScript 3 semantics because it allows a function to be called at the top level before it is defined.
• Conditional compilation cannot guarantee that variables’ initializers have been run before the variables are referenced. The dynamic execution model guarantees that.
• Conditional compilation does not allow one to use general prototype-based types because prototypes are often calculated at run time. The dynamic execution model allows these.
• Conditional compilation does not support functors and templates without significant additional syntax. The dynamic execution model naturally extends to these.

Compiler Blocks

Another alternative execution model briefly considered but rejected is the idea of allowing compiler blocks. A compiler block has the syntax:

The compile attribute is a hint that the block may be (but does not have to be) evaluated early. The statements inside this block should depend only on each other, on the results of earlier compiler blocks, and on properties of the environment that are designated as being available early. Other than perhaps being evaluated early, compiler blocks respect all of the scope rules and semantics of the enclosing program. Any definitions introduced by a compiler block are saved and reintroduced at normal evaluation time. On the other hand, side effects may or may not be reintroduced at normal evaluation time, so compiler blocks should not rely on side effects.

compile is an attribute, so it may also be applied to individual definitions without enclosing them in a block.

As an example, after defining

compile var x = 2;

function f1() {
  compile {
    var y = 5;
    var x = 1;
    while (y) x *= y--;
  }
  return ++x;
}

function f2() {
  compile {
    var y = x;
  }
  return x+y;
}

the value of global x will still be 2, calling f1() will always return 121, and calling f2() will return 4. If the statement x=5 is then evaluated at the global level, f1() will still return 121 because it uses its own local x. On the other hand, calling f2() may return either 7 or 10 at the implementation’s discretion — 7 if the implementation evaluated the compile block early and saved the value of y or 10 if it didn’t. As this example illustrates, it is poor technique to define variables inside compiler blocks; constants are usually better.

A fully dynamic implementation of ECMAScript 4 may choose to ignore the compile attribute and evaluate all compiler blocks at normal evaluation time. A fully static implementation may require that all user-defined types and attributes be defined inside compiler blocks.

Should const definitions with simple constant expressions such as const four = 2+2 be treated as though they were implicitly compiler definitions (compile const four = 2+2)?

Waldemar Horwat Last modified Tuesday, September 10, 2002