On Sep 25, 2009, at 10:31 AM, Charles Jolley wrote:

But strict mode is not supported in current or downrev IE. So it is imperative that Microsoft fix these old named function expression name-binding bugs before, or at the same time as, script mode along with all of ES5 are supported in a future version of IE.

Wouldn't that violate the principle of strict mode?

Which "that"? IE fixing bad old named function expression bugs violates nothing, not ES5 or ES5 strict, and certainly not ES3 which spec'ed how named function expressions bind their names (not in the variable object! :-P).

Ok, forget IE bugs. It seems you are asking here whether "use strict"; in ES5 engines causes semantic changes that are not backward compatible. The answer is that it does. See Annex C of ES5.

The cases are edgy, e.g., arguments[i] aliasing the i'th formal parameter does not happen in an ES5 implementation under "use strict", but it would when that code is fed to a pre-ES5 engine. And of course, arguemnts.callee and the caller property poison pills, which throw in ES5 strict but don't when code using them is fed into a pre-ES5 engine.


I thought the idea was that writing code in Harmony could basically boil down to ES3 code if I wanted to. But in this case it wouldn't; the same code written for Harmony would appear valid ES3 but would fail.

Right. You have to test to see that "use strict"; code works the way you expect in older browsers. If you steer clear of arguments objects you don't have much to do. There's another runtime semantic change, to eval (it can't create bindings in its dynamic scope), which under ES5 strict will break such examples (at least when I last checked) as extjs.com, which did something like

  function compile(...) {
    eval("var f = function() {" + compile_optimized_query(); + "}");
    return f;
  }

This can be avoided by putting the var f = outside the eval, if you parenthesize the lambda of course -- except Jack Slocum pointed out to me that IE has bugs (perhaps only old IE versions still have these) that bite this alternative! I'm not sure what the latest extjs code does, since it seems hidden by dual licensing stuff I didn't click through.


Sorry if I'm misunderstanding the goal here. I thought the point of "use strict" is the same code could still run on older browsers who don't understand the directive.

The code runs, but certain edge cases don't work the same. In old engines they work as ever, in ES5 engines they may throw due to the poison pills, or they may do something "better" (e.g., eval binds in its own declarative environment).

A strict mode that only prevents programs from reaching runtime would require too much static analysis.

A strict mode that adds runtime errors will require testing in ES5 and pre-ES5 engines.

Given our inability to avoid runtime errors new with strict mode, we managed to reach consensus on a few further runtime semantic changes. If these cause trouble, we may find fewer strict mode uses than we'd like to find. But it's hard to evaluate strict mode adoption for success or failure based on whether these runtime changes bite back -- we can't run the experiment a different way.

/be


-Charles



We talked about this issue yesterday at the end of the TC39 meeting. I think Allen had to leave but I mailed him about it.

/be


Best Fabian


Charles Jolley wrote:

Hi everyone,

First an introduction: I am the lead developer of SproutCore
(http://www.sproutcore.com
), an open source framework for building client-side applications in HTML5. SproutCore is used in several large projects including Apple's MobileMe and iwork.com, OtherInbox, and some other large projects that
are not yet public.

Point is, we write very large applications in JavaScript and HTML5.
I've been following ES5/Harmony closely.  By and large I am very
excited about the features to be added.  One critical feature (imo)
that is missing, however, is "arguments.callee" - at least some way to
identify the current function that is executing.

I spoke with Brenden Eich and Yehuda Katz about this on Friday and
talked to Doug Crockford about it today and they suggested I email
this list to make my case so here it is:

USE CASE

Currently SproutCore implements a class-like system on top of
JavaScript. That is, you can have "classes" with "subclasses" and you can instantiate both. SproutCore is not strictly classical, and JS is
not class-based of course, but I think this is a pattern that many
developers commonly want to implement and use in JS.

The problem comes with implementing methods that "override" methods in superclasses. Take the following example [extend() is the SproutCore
function that creates a new "subclass"]:

ClassA = SC.Object.extend({

foo: function() {
// do something
}

});

ClassB = ClassA.extend({

// NOTE: overrides foo in classA!
foo: function() {
// call ClassA.foo();
// do something else
}

});

--

Now the question is, how can ClassB.foo() call ClassA.foo() in a
generic fashion?

I could force developers to hard code this knowledge  (i.e. when
implementing ClassB.foo() you have to explicitly call ClassA.foo.apply (this)) but this is prone to developer error and also makes the code
not easily transportable from one method to another; violating the
sort of "copy-and-paste" ethos that is part of JavaScript.

I've been told that I could name the functions.  e.g.:

ClassB = ClassA.extend({
foo:  function foo() {
// ..code
}
});

Somehow that should solve my problem, though I can't really work out how. But regardless, asking developers to name each method twice in
the declaration is also error prone and fragile.

--

The way I solve this currently is to implement extend() so that when
it copies an overloaded method, it sets a property ("base") on the
Function to point to the Function it is overloading. In the example
above, for example, this means that ClassB.foo.base === ClassA.foo.

This way I can write a generic call to "super" like so:

ClassB = ClassA.extend({

foo: function() {
arguments.callee.base.apply(this, arguments); // calls super!
// other code
}

});

--

I realize this is not the most elegant looking code, but currently its
the only way I can figure out to implement the "super" pattern in a
generic way in JS.

SOLUTIONS?

With ES5, arguments.callee is gone.

One suggestion Brenden had was to perhaps include a "thisFunction" or some such property that returns the current function instance you are
in.  This seems useful to me for a bunch of other meta-programming
patterns as well but would certainly work here.

Another solution suggested by Douglas would be to provide a way for a
function to get its current "name" AND for ECMAScript to follow a
convention that anonymous functions when declared as part of an object
literal take on the name they are assigned to by default.

Either of these would work for this pattern; there are probably other
solutions as well.

I don't really care what ends up in the final implementation only that
there is some way to generically implement the ultra-common class-
based-with-method-overloading pattern in ECMAScript >= 5 without
jumping through some crazy hoops.

--

So that's my best argument on this.  Any suggestions of alternative
implementations that will work in ECMAScript 5?  If this is a
limitation of the new language, what I can do to agitate for something
to be added?

Thanks,
-Charles


_______________________________________________
es-discuss mailing list
[email protected]
https://mail.mozilla.org/listinfo/es-discuss



--
View this message in context: 
http://www.nabble.com/arguments.callee-in-Harmony-tp25603357p25610255.html
Sent from the Mozilla - ECMAScript 4 discussion mailing list archive at Nabble.com.

_______________________________________________
es-discuss mailing list
[email protected]
https://mail.mozilla.org/listinfo/es-discuss

_______________________________________________
es-discuss mailing list
[email protected]
https://mail.mozilla.org/listinfo/es-discuss


_______________________________________________
es-discuss mailing list
[email protected]
https://mail.mozilla.org/listinfo/es-discuss

Reply via email to