Archive for October, 2011

ES6 – a quick look at Weak Maps

31 Oct

One of my favourite proposals in ECMAScript Harmony is Weak Maps. Simply put, ‘weak maps are unordered tables mapping objects to values’ with references to those objects and values being weak. A weak reference means that the item won’t be protected from garbage collection if no other references are held. For example:

// Using the proposed API (currently implemented in Firefox)
var wm = WeakMap();

(function () {
    var obj = {};
    wm.set(obj, 1);

console.log(wm); //-> "[object WeakMap]"

By now a very obvious use case might have occurred to you, but I’ll get to that in a minute. When the above function finishes execution, obj is no longer reachable, so it can be garbage collected even though our WeakMap object is still reachable. This addition to the language is a useful tool for an age-old concern in JavaScript and the DOM; memory leaks. That aforementioned use case is becoming more obvious now, eh?

Where do Weak Maps fit in with current code?

OK, if you never had to implement this kind of thing before, you might not have thought of jQuery’s data() method. allows us to attach an arbitrary value to a DOM object. Whilst you can do this with expando properties, ordinarily you run the risk of circular references and memory leaks. jQuery tackles this by creating a unique ID for itself (only once) and each element it is attaching data to. It tacks the element’s unique ID, which is really just an incremented counter, onto an expando property whose name is the unique ID. So it looks something like this:

     element[jQuery.expando] = elementId;

Any data you attach to the element is actually stored in an array at index elementId. jQuery is able to avoid memory leaks because there are no circular references, and it can delete all stored data when the page is unloaded (or the element is removed using jQuery’s own methods). However, it does have some limitations. For example — according to the source code — object/embed/applet cannot accept expando properties and, therefore, cannot work with them.

Weak maps come in here because they can do the job much better. They cut out the need for the expando property entirely, along with the requirement of handling JS objects differently to DOM objects. They also expand on jQuery’s ability to allow garbage collection when DOM elements are removed by its own methods, by automatically allowing garbage collection when DOM elements no longer reachable after they’ve been removed by any method. Had WeakMaps been implemented in ECMAScript 3 (if only), jQuery’s workload with data() could have been significantly reduced. Perhaps jQuery wouldn’t have bothered implementing data if we could create weak maps, though.

Another use case to consider is secretly extending objects without the danger of naming collisions. Using someGloballyAvailableObject.someProperty = "some arbitrary value"; as an example, what if someGloballyAvailableObject already has someProperty? We’ve just written a potentially script-breaking piece of code. However, with a weak map we can eliminate this danger entirely:

(function () { 
    var wm = new WeakMap();
    wm.set(someGloballyAvailableObject, {someProperty:"some arbitrary value"});

There’s also a hidden bonus here. I used the word secretly before because no other piece of our script can access this mapped value if we never expose the wm variable.

You might recall (if you read it) my article on creating lookup lists. If you do, you might be thinking that weak maps are a good use for creating lists, but in reality this isn’t the case because weak maps — in the current proposal, at least — will only map a value to an object.

Weak map compatibility shims

A true weak map is technically impossible in ECMAScript 3. We can create a map of objects to values by simply using two arrays:

function WeakMap() {
    this.keys = [/* key01, key02, key03 */];
    //                ↓      ↓      ↓
    this.vals = [/* val01, val02, val03 */];

Unfortunately, there’s 2 very important caveats here. Firstly, these references are not (and cannot be) weak, so they will not be garbage collected as long as the mapping exists. Secondly, for our get(key) method to work, we need to search our keys array. This significantly impacts performance, particularly on large maps.

As for ES5, I was pointed in the direction of this implementation recently. It’s very clever; It eliminates the need for a keys array and uses a custom valueOf function to keep track of the mapped value. I profiled the method in Chrome and it does indeed appear to prevent leaks. However, it does still have its limitations like not working on frozen objects or objects whose valueOf function is not configurable.


Update: HTML5 oninput event plugin for jQuery

24 Oct

About a year ago, I wrote my first special events plugin for jQuery — a cross-browser implementation of the HTML5 oninput event. A while afterwards, though, it became apparent that there were a multitude of issues with browsers that already implemented the event.

  • Opera 10.x doesn’t fire the event for cut, paste, undo & drop operations on <input> elements. They fixed this in Opera 11, though.
  • Safari 4.x and lower don’t support the event for <textarea> elements. The textInput event is supported on them for 3.1+, however.
  • Internet Explorer 9 forgets to fire the event when the backspace key is used to delete text. Other than that majorly horrendous oversight, they almost got it right.
  • Firefox has issues that make feature detection very difficult. In fact, detection is only possible via another Firefox proprietary feature, which I suppose is rather lucky.

With these problems stacking up, my code was stacking up too — how could I detect and work around each of these problems? On top of that, I couldn’t easily make the event work with delegate() or live(). Soon, it became apparent that I was going to have to completely rethink the approach I made in the original plugin.

Feature detection

The original code used feature detection to discover whether browsers supported the event. "oninput" in document.documentElement is enough for this in most browsers, but the aforementioned Firefox issues required an excellent little snippet from Daniel Friesen, which I slightly modified. It was a shame though, as this extra work bulked up the plugin a little more than I would have liked. Safari didn’t like our feature detection either, because even though our snippet told us the browser implemented support for the event, it doesn’t tell us that <textarea> won’t fire it.

I eventually decided that feature detection wasn’t going to work — graceful degradation, on the other hand, might. This involves binding to all known events that fire when an input element is modified, and then setting a flag to tell the event handlers whether or not they need to execute. For events that fire pre-DOM modification, we set a timer with timeout value of 0ms (this will be adjusted to browser’s minimum allowable value). The events that fire post-DOM modification cancel this timer. So, let’s say our browser is Internet Explorer 6, we have some text in our clipboard and we just pressed Ctrl+V in a <textarea> that has our events bound. For the Ctrl key, the onkeydown event sets the timer but when it executes the value hasn’t changed, so the handler does nothing. For the V key, the plugin will handle the events in the following sequence:

Pre-DOM modification
    onkeydown -> timer is set
    onpaste   -> previous timer is cleared, new timer is set

Post-DOM modification - DOM is modified, textarea's value is updated but element hasn't repainted yet
    onpropertychange -> timer is cleared, the handler is triggered

Opera 9, however, the best possible compatibility we can provide is via the onkeydown event, so that sets the timer which executes almost instantly, triggering our special event handler. In Safari 4, the textInput event fires post-DOM modification, so that clears the timer and triggers the handler, giving seamless `oninput` compatibility to the developer.

Late binding

No, not that kind of late binding. In order to mitigate the lack of event bubbling for our Internet Explorer polyfill, I opted to have the plugin not bind the input events to the element right away if it wasn’t an <input> or <textarea>. Instead, it binds several bubbling events — onfocusin, onmouseover and ondragstart — to the element. When one of those events fires, is checked and if it is “INPUT” or “TEXTAREA” then we bind our input events to it. This allows delegate() to work really well in IE and I suspect, but haven’t verified, this is how jQuery is able to offer event delegation for non-propagating events such as onresize.

Problems with the new approach

Undoubtedly, the biggest problem is naming the special event. I can’t think of a way to have a special event named `input` that binds to the native `input` event and completely overrides it because, when we come to trigger the event later, we can’t. In my oh-so finite wisdom, I decided that the best solution I could provide was a differently named special event. So, when using the plugin now, you need to bind to "txtinput". $(selector).input(function () {}); will still work, though.

Another problem was teardown. Suppose we have a DOM structure like body -> div -> input — now, what happens if we bind to the <body> and <div> elements? Actually, this works completely as expected but, what if we unbind from one of those elements? Our teardown function needs to make sure that no other handlers are relying on our bound events before we remove them. The solution makes use of data and, every time an event is bound, increases an internal count to keep track. This counter is decreased for each of those elements in the teardown.

You can see the updated plugin at the original project page.