Implementing a syntax-higlighting JavaScript editor in JavaScript

or

A brutal odyssey to the dark side of the DOM tree

Topic: JavaScript, advanced browser weirdness, cool programming techniques
Audience: Programmers, especially JavaScript programmers
Author: Marijn Haverbeke
Date: May 24th 2007

Note: some of the details given here no longer apply to the current CodeMirror codebase, which has evolved quite a bit in the meantime.

In one of his (very informative) video lectures, Douglas Crockford remarks that writing JavaScript for the web is 'programming in a hostile environment'. I had done my fair share of weird workarounds, and even occasonally gave up an on idea entirely because browsers just wouldn't support it, but before this project I never really realized just how powerless a programmer can be in the face of buggy, incompatible, and poorly designed platforms.

The plan was not ridiculously ambitious. I wanted to 'enhance' a textarea to the point where writing code in it is pleasant. This meant automatic indentation and, if possible at all, syntax highlighting.

In this document I describe the story of implementing this, for your education and amusement. A demonstration of the resulting program, along with the source code, can be found at my website.

Take one: Only indentation

The very first attempt merely added auto-indentation to a textarea element. It would scan backwards through the content of the area, starting from the cursor, until it had enough information to decide how to indent the current line. It took me a while to figure out a decent model for indenting JavaScript code, but in the end this seems to work:

When scanning backwards through code one has to take string values, comments, and regular expressions (which are delimited by slashes) into account, because braces and semicolons and such are not significant when they appear inside them. Single-line ('//') comments turned out to be rather inefficient to check for when doing a backwards scan, since every time you encounter a newline you have to go on to the next newline to determine whether this line ends in a comment or not. Regular expressions are even worse ― without contextual information they are impossible to distinguish from the division operator, and I didn't get them working in this first version.

To find out which line to indent, and to make sure that adding or removing whitespace doesn't cause the cursor to jump in strange ways, it is necessary to determine which text the user has selected. Even though I was working with just a simple textarea at this point, this was already a bit of a headache.

On W3C-standards-respecting browsers, textarea nodes have selectionStart and selectionEnd properties which nicely give you the amount of characters before the start and end of the selection. Great!

Then, there is Internet Explorer. Internet Explorer also has an API for looking at and manipulating selections. It gives you information such as a detailed map of the space the selected lines take up on the screen, in pixels, and of course the text inside the selection. It does, however, not give you much of a clue on where the selection is located in the document.

After some experimentation I managed to work out an elaborate method for getting something similar to the selectionStart and selectionEnd values in other browsers. It worked like this:

That seemed to work, but when resetting the selection after modifying the content of the textarea I ran into another interesting feature of these TextRanges: You can move their endpoints by a given number of characters, which is useful when trying to set a cursor at the Nth character of a textarea, but in this context, newlines are not considered to be characters, so you'll always end up one character too far for every newline you passed. Of course, you can count newlines and compensate for this (though it is still not possible to position the cursor right in front of a newline). Sheesh.

After ragging on Internet Explorer for a while, let us move on and rag on Firefox a bit. It turns out that, in Firefox, getting and setting the text content of a DOM element is unexplainably expensive, especially when there is a lot of text involved. As soon as I tried to use my indentation code to indent itself (some 400 lines), I found myself waiting for over four seconds every time I pressed enter. That seemed a little slow.

designMode it is

The solution was obvious: Since the text inside a textarea can only be manipulated as one single big string, I had to spread it out over multiple nodes. How do you spread editable content over multiple nodes? Right! designMode or contentEditable.

Now I wasn't entirely naive about designMode, I had been looking into writing a non-messy WYSIWYG editor before, and at that time I had concluded two things:

Basically, the good folks at Microsoft designed a really bad interface for putting editable documents in pages, and the other browsers, not wanting to be left behind, more or less copied that. And there isn't much hope for a better way to do this appearing anytime soon. Wise people probably use a Flash movie or (God forbid) a Java applet for these kind of things, though those are not without drawbacks either.

Anyway, seeing how using an editable document would also make syntax highlighting possible, I foolishly went ahead. There is something perversely fascinating about trying to build a complicated system on a lousy, unsuitable platform.

A parser

How does one do decent syntax highlighting? A very simple scanning can tell the difference between strings, comments, keywords, and other code. But this time I wanted to actually be able to recognize regular expressions, so that I didn't have any blatant incorrect behaviour anymore.

That brought me to the idea of doing a serious parse on the code. This would not only make detecting regular expressions much easier, it would also give me detailed information about the code, which can be used to determine proper indentation levels, and to make subtle distinctions in colouring, for example the difference between variable names and property names.

And hey, when we're parsing the whole thing, it would even be possible to make a distinction between local and global variables, and colour them differently. If you've ever programmed JavaScript you can probably imagine how useful this would be ― it is ridiculously easy to accidentally create global instead of local variables. I don't consider myself a JavaScript rookie anymore, but it was (embarrasingly enough) only this week that I realized that my habit of typing for (name in object) ... was creating a global variable name, and that I should be typing for (var name in object) ... instead.

Re-parsing all the code the user has typed in every time he hits a key is obviously not feasible. So how does one combine on-the-fly highlighting with a serious parser? One option would be to split the code into top-level statements (functions, variable definitions, etc.) and parse these separately. This is horribly clunky though, especially considering the fact that modern JavaScripters often put all the code in a file in a single big object or function to prevent namespace pollution.

I have always liked continuation-passing style and generators. So the idea I came up with is this: An interruptable, resumable parser. This is a parser that does not run through a whole document at once, but parses on-demand, a little bit at a time. At any moment you can create a copy of its current state, which can be resumed later. You start parsing at the top of the code, and keep going as long as you like, but throughout the document, for example at every end of line, you store a copy of the current parser state. Later on, when line 106 changes, you grab the interrupted parser that was stored at the end of line 105, and use it to re-parse line 106. It still knows exactly what the context was at that point, which local variables were defined, which unfinished statements were encountered, and so on.

But that, unfortunately, turned out to be not quite as easy as it sounds.

The DOM nodes underfoot

Of course, when working inside an editable frame we don't just have to deal with text. The code will be represented by some kind of DOM tree. My first idea was to set the white-space: pre style for the frame and try to work with mostly text, with the occasional coloured span element. It turned out that support for white-space: pre in browsers, especially in editable frames, is so hopelessly glitchy that this was unworkable.

Next I tried a series of div elements, one per line, with span elements inside them. This seemed to nicely reflect the structure of the code in a shallowly hierarchical way. I soon realized, however, that my code would be much more straightfoward when using no hierarchy whatsoever ― a series of spans, with br tags at the end of every line. This way, the DOM nodes form a flat sequence that corresponds to the sequence of the text ― just extract text from span nodes and substitute newlines for br nodes.

It would be a shame if the editor would fall apart as soon as someone pastes some complicated HTML into it. I wanted it to be able to deal with whatever mess it finds. This means using some kind of HTML-normalizer that takes arbitrary HTML and flattens it into a series of brs and span elements that contain a single text node. Just like the parsing process, it would be best if this did not have to done to the entire buffer every time something changes.

It took some banging my head against my keyboard, but I found a very nice way to model this. It makes heavy use of generators, for which I used MochiKit's iterator framework. Bob Ippolito explains the concepts in this library very well in his blog post about it. (Also notice some of the dismissive comments at the bottom of that post. They say "I don't think I really want to learn this, so I'll make up some silly reason to condemn it.")

The highlighting process consists of the following elements: normalizing the DOM tree, extracting the text from the DOM tree, tokenizing this text, parsing the tokens, and finally adjusting the DOM nodes to reflect the structure of the code.

The first two, I put into a single generator. It scans the DOM tree, fixing anything that is not a simple top-level span or br, and it produces the text content of the nodes (or a newline in case of a br) as its output ― each time it is called, it yields a string. Continuation passing style was a good way to model this process in an iterator, which has to be processed one step at a time. Look at this simplified version:

function traverseDOM(start){
  var cc = function(){scanNode(start, stop);};
  function stop(){
    cc = stop;
    throw StopIteration;
  }
  function yield(value, c){
    cc = c;
    return value;
  }

  function scanNode(node, c){
    if (node.nextSibling)
      var nextc = function(){scanNode(node.nextSibling, c);};
    else
      var nextc = c;

    if (/* node is proper span element */)
      return yield(node.firstChild.nodeValue, nextc);
    else if (/* node is proper br element */)
      return yield("\n", nextc);
    else
      /* flatten node, yield its textual content */;
  }

  return {next: function(){return cc();}};
}

The variable c stands for 'continuation', and cc for 'current continuation' ― that last variable is used to store the function to continue with, when yielding a value to the outside world. Every time control leaves this function, it has to make sure that cc is set to a suitable value, which is what yield and stop take care of.

The object that is returned contains a next method, which is MochiKit's idea of an iterator, and the initial continuation just throws a StopIteration, which is how MochiKit signals that an iterator has reached its end.

The first lines of scanNode extend the continuation with the task of scanning the next node, if there is a next node. The rest of the function decides what kind of value to yield. Note that this is a rather trivial example of this technique, since the process of going through these nodes is basically linear (it was much, much more complex in earlier versions), but still the trick with the continuations makes the code shorter and, for those in the know, clearer than the equivalent 'storing the iterator state in variables' approach.

The next iterator that the input passes through is the tokenizer. Well, actually, there is another iterator in between that isolates the tokenizer from the fact that the DOM traversal yields a bunch of separate strings, and presents them as a single character stream (with a convenient peek operation), but this is not a very interesting one. What the tokenizer returns is a stream of token objects, each of which has a value, its textual content, a type, like "variable", "operator", or just itself, "{" for example, in the case of significant punctuation or special keywords. They also have a style, which is used later by the highlighter to give their span elements a class name (the parser will still adjust this in some cases).

At first I assumed the parser would have to talk back to the tokenizer about the current context, in order to be able to distinguish those accursed regular expressions from divisions, but it seems that regular expressions are only allowed if the previous (non-whitespace, non-comment) token was either an operator, a keyword like new or throw, or a specific kind of punctuation ("[{}(,;:") that indicates a new expression can be started here. This made things considerably easier, since the 'regexp or no regexp' question could stay entirely within the tokenizer.

The next step, then, is the parser. It does not do a very thorough job because, firstly, it has to be fast, and secondly, it should not go to pieces when fed an incorrect program. So only superficial constructs are recognized, keywords that resemble each other in syntax, such as while and if, are treated in precisely the same way, as are try and else ― the parser doesn't mind if an else appears without an if. Stuff that binds variables, var, function, and catch to be precise, is treated with more care, because the parser wants to know about local variables.

Inside the parser, three kinds of context are stored. Firstly, a set of known local variables, which is used to adjust the style of variable tokens. Every time the parser enters a function, a new set of variables is created. If there was already such a set (entering an inner function), a pointer to the old one is stored in the new one. At the end of the function, the current variable set is 'popped' off and the previous one is restored.

The second kind of context is the lexical context, this keeps track of whether we are inside a statement, block, or list. Like the variable context, it also forms a stack of contexts, with each one containing a pointer to the previous ones so that they can be popped off again when they are finished. This information is used for indentation. Every time the parser encounters a newline token, it attaches the current lexical context and a 'copy' of itself (more about that later) to this token.

The third context is a continuation context. This parser does not use straight continuation style, instead it uses a stack of actions that have to be performed. These actions are simple functions, a kind of minilanguage, they act on tokens, and decide what kind of new actions should be pushed onto the stack. Here are some examples:

function expression(type){
  if (type in atomicTypes) cont(maybeoperator);
  else if (type == "function") cont(functiondef);
  else if (type == "(") cont(pushlex("list"), expression, expect(")"), poplex);
  else if (type == "operator") cont(expression);
  else if (type == "[") cont(pushlex("list"), commasep(expression), expect("]"), poplex);
  else if (type == "{") cont(pushlex("list"), commasep(objprop), expect("}"), poplex);
  else if (type == "keyword c") cont(expression);
}

function block(type){
  if (type == "}") cont();
  else pass(statement, block);
}

The function cont (for continue), will push the actions it is given onto the stack (in reverse order, so that the first one will be popped first). Actions such as pushlex and poplex merely adjust the lexical environment, while others, such as expression itself, do actual parsing. pass, as seen in block, is similar to cont, but it does not 'consume' the current token, so the next action will again see this same token. In block, this happens when the function determines that we are not at the end of the block yet, so it pushes the statement function which will interpret the current token as the start of a statement.

These actions are called by a 'driver' function, which filters out the whitespace and comments, so that the parser actions do not have to think about those, and keeps track of some things like the indentation of the current line and the column at which the current token ends, which are stored in the lexical context and used for indentation. After calling an action, if the action called cont, this driver function will return the current token, if pass (or nothing) was called, it will immediately continue with the next action.

This goes to show that it is viable to write a quite elaborate minilanguage in a macro-less language like JavaScript. I don't think it would be possible to do something like this without closures (or similarly powerful abstraction) though, I've certainly never seen anything like it in Java code.

The way a 'copy' of the parser was produced shows a nice usage of closures. Like with the DOM transformer shown above, most of the local state of the parser is held in a closure produced by calling parse(stream). The function copy, which is local to the parser function, produces a new closure, with copies of all the relevant variables:

function copy(){
  var _context = context, _lexical = lexical, _actions = copyArray(actions);

  return function(_tokens){
    context = _context;
    lexical = _lexical;
    actions = copyArray(_actions);
    tokens = _tokens;
    return parser;
  };
}

Where parser is the object that contains the next (driver) function, and a reference to this copy function. When the function that copy produces is called with a token stream as argument, it updates the local variables in the parser closure, and returns the corresponding iterator object.

Moving on, we get to the last stop in this chain of generators, the actual highlighter. You can view this one as taking two streams as input, on the one hand there is the stream of tokens from the parser, and on the other hand there is the DOM tree as left by the DOM transformer. If everything went correctly, these two should be synchronized. The highlighter can look at the current token, see if the span in the DOM tree corresponds to it (has the same text content, and the correct class), and if not it can chop up the DOM nodes to conform to the tokens.

Every time the parser yields a newline token, the highligher encounters a br element in the DOM stream. It takes the copy of the parser and the lexical context from this token and attaches them to the DOM node. This way, a new highlighting process can be started from that node by re-starting the copy of the parser with a new token stream, which reads tokens from the DOM nodes starting at that br element, and the indentation code can use the lexical context information to determine the correct indentation at that point.

Selection woes

All the above can be done using the DOM interface that all major browsers have in common, and which is relatively free of weird bugs and abberrations. However, when the user is typing in new code, this must also be highlighted. For this to happen, the program must know where the cursor currently is, and because it mucks up the DOM tree, it has to restore this cursor position after doing the highlighting.

Re-highlighting always happens per line, because the copy of the parser is stored only at the end of lines. Doing this every time the user presses a key is terribly slow and obnoxious, so what I did was keep a list of 'dirty' nodes, and as soon as the user didn't type anyting for 300 milliseconds the program starts re-highlighting these nodes. If it finds more than ten lines must be re-parsed, it does only ten and waits another 300 milliseconds before it continues, this way the browser never freezes up entirely.

As mentioned earlier, Internet Explorer's selection model is not the most practical one. My attempts to build a wrapper that makes it look like the W3C model all stranded. In the end I came to the conclusion that I only needed two operations:

It turns out that the pixel-based selection model that Internet Explorer uses, which always seemed completely ludricrous to me, is perfect for the first case. Since the DOM transformation (generally) does not change the position of things, storing the pixel offsets of the selection makes it possible to restore that same selection, never mind what happened to the underlying DOM structure.

[Later addition: Note that this, due to the very random design of the TextRange interface, only really works when the whole selection falls within the visible part of the document.]

Doing the same with the W3C selection model is a lot harder. What I ended up with was this:

Now in the second case (getting the top-level node at the cursor) the Internet Explorer cheat does not work. In the W3C model this is rather easy, you have to do some creative parent- and sibling-pointer following to arrive at the correct top-level node, but nothing weird. In Internet Explorer, all we have to go on is the parentElement method on a TextRange, which gives the first element that completely envelops the selection. If the cursor is inside a text node, this is good, that text node tells us where we are. If the cursor is between nodes, for example between two br nodes, you get to top-level node itself back, which is remarkably useless. In cases like this I stoop to a rather ugly hack (which fortunately turned out to be acceptably fast) ― I create a temporary empty span with an ID inside the selection, get a reference to this span by ID, take its previousSibling, and remove it again.

Unfortunately, Opera's selection implementation is buggy, and it will give wildly incorrect Range objects when the cursor is between two nodes. This is a bit of a showstopper, and until I find a workaround for that or it gets fixed, the highlighter doesn't work properly in Opera.

Also, when one presses enter in a designMode document in Firefox or Opera, a br tag is inserted. In Internet Explorer, pressing enter causes some maniacal gnome to come out and start wrapping all the content before and after the cursor in p tags. I suppose there is something to be said for that, in principle, though if you saw the tag soup of fonts and nested paragraphs Internet Explorer generates you would soon enough forget all about principle. Anyway, getting unwanted p tags slowed the highlighter down terribly ― it had to overhaul the whole DOM tree to remove them again, every time the user pressed enter. Fortunately I could fix this by capturing the enter presses and manually inserting a br tag at the cursor.

On the subject of Internet Explorer's tag soup, here is an interesting anecdote: One time, when testing the effect that modifying the content of a selection had, I inspected the DOM tree and found a "/B" element. This was not a closing tag, there are no closing tags in the DOM tree, just elements. The nodeName of this element was actually "/B". That was when I gave up any notions of ever understanding the profound mystery that is Internet Explorer.

Closing thoughts

Well, I despaired at times, but I did end up with a working JavaScript editor. I did not keep track of the amount of time I wasted on this, but I would estimate it to be around fifty hours. Finding workarounds for browser bugs can be a terribly nonlinear process. I just spent half a day working on a weird glitch in Firefox that caused the cursor in the editable frame to be displayed 3/4 line too high when it was at the very end of the document. Then I found out that setting the style.display of the iframe to "block" fixed this (why not?). I'm amazed how often issues that seem hopeless do turn out to be avoidable, even if it takes hours of screwing around and some truly non-obvious ideas.

For a lot of things, JavaScript + DOM elements are a surprisingly powerful platform. Simple interactive documents and forms can be written in browsers with very little effort, generally less than with most 'traditional' platforms (Java, Win32, things like WxWidgets). Libraries like Dojo (and a similar monster I once wrote myself) even make complex, composite widgets workable. However, when applications go sufficiently beyond the things that browsers were designed for, the available APIs do not give enough control, are nonstandard and buggy, and are often poorly designed. Because of this, writing such applications, when it is even possible, is painful process.

And who likes pain? Sure, when finding that crazy workaround, subdueing the damn browser, and getting everything to work, there is a certain macho thrill. But one can't help wondering how much easier things like preventing the user from pasting pictures in his source code would be on another platform. Maybe something like Silverlight or whatever other new browser plugin gizmos people are pushing these days will become the way to solve things like this in the future. But, personally, I would prefer for those browser companies to put some real effort into things like cleaning up and standardising shady things like designMode, fixing their bugs, and getting serious about ECMAScript 4.

Which is probably not realistically going to happen anytime soon.

Some interesting projects similar to this:

If you have any remarks, criticism, or hints related to the above, drop me an e-mail at marijnh@gmail.com. If you say something generally interesting, I'll include your reaction here at the bottom of this page.