JavaScript split Bugs: Fixed!

The String.prototype.split method is very handy, so it's a shame that if you use a regular expression as its delimiter, the results can be so wildly different cross-browser that odds are you've just introduced bugs into your code (unless you know precisely what kind of data you're working with and are able to avoid the issues). Here's one example of other people venting about the problems. Following are the inconsistencies cross-browser when using regexes with split:

  • Internet Explorer excludes almost all empty values from the resulting array (e.g., when two delimiters appear next to each other in the data, or when a delimiter appears at the start or end of the data). This doesn't make any sense to me, since IE does include empty values when using a string as the delimiter.
  • Internet Explorer and Safari do not splice the values of capturing parentheses into the returned array (this functionality can be useful with simple parsers, etc.)
  • Firefox does not splice undefined values into the returned array as the result of non-participating capturing groups.
  • Internet Explorer, Firefox, and Safari have various additional edge-case bugs where they do not follow the split specification (which is actually quite complex).

The situation is so bad that I've simply avoided using regex-based splitting in the past.

That ends now. wink

The following script provides a fast, uniform cross-browser implementation of String.prototype.split, and attempts to precisely follow the relevant spec (ECMA-262 v3 §15.5.4.14, pp.103,104).

I've also created a fairly quick and dirty page where you can test the result of more than 50 usages of JavaScript's split method, and quickly compare your browser's results with the correct implementation. On the test page, the pink lines in the third column highlight incorrect results from the native split method. The rightmost column shows the results of the below script. It's all green in every browser I've tested (IE 5.5 – 7, Firefox 2.0.0.4, Opera 9.21, Safari 3.0.1 beta, and Swift 0.2).

Run the tests in your browser.

Here's the script:

/*!
 * Cross-Browser Split 1.1.1
 * Copyright 2007-2012 Steven Levithan <stevenlevithan.com>
 * Available under the MIT License
 * ECMAScript compliant, uniform cross-browser split method
 */

/**
 * Splits a string into an array of strings using a regex or string separator. Matches of the
 * separator are not included in the result array. However, if `separator` is a regex that contains
 * capturing groups, backreferences are spliced into the result each time `separator` is matched.
 * Fixes browser bugs compared to the native `String.prototype.split` and can be used reliably
 * cross-browser.
 * @param {String} str String to split.
 * @param {RegExp|String} separator Regex or string to use for separating the string.
 * @param {Number} [limit] Maximum number of items to include in the result array.
 * @returns {Array} Array of substrings.
 * @example
 *
 * // Basic use
 * split('a b c d', ' ');
 * // -> ['a', 'b', 'c', 'd']
 *
 * // With limit
 * split('a b c d', ' ', 2);
 * // -> ['a', 'b']
 *
 * // Backreferences in result array
 * split('..word1 word2..', /([a-z]+)(\d+)/i);
 * // -> ['..', 'word', '1', ' ', 'word', '2', '..']
 */
var split;

// Avoid running twice; that would break the `nativeSplit` reference
split = split || function (undef) {

    var nativeSplit = String.prototype.split,
        compliantExecNpcg = /()??/.exec("")[1] === undef, // NPCG: nonparticipating capturing group
        self;

    self = function (str, separator, limit) {
        // If `separator` is not a regex, use `nativeSplit`
        if (Object.prototype.toString.call(separator) !== "[object RegExp]") {
            return nativeSplit.call(str, separator, limit);
        }
        var output = [],
            flags = (separator.ignoreCase ? "i" : "") +
                    (separator.multiline  ? "m" : "") +
                    (separator.extended   ? "x" : "") + // Proposed for ES6
                    (separator.sticky     ? "y" : ""), // Firefox 3+
            lastLastIndex = 0,
            // Make `global` and avoid `lastIndex` issues by working with a copy
            separator = new RegExp(separator.source, flags + "g"),
            separator2, match, lastIndex, lastLength;
        str += ""; // Type-convert
        if (!compliantExecNpcg) {
            // Doesn't need flags gy, but they don't hurt
            separator2 = new RegExp("^" + separator.source + "$(?!\\s)", flags);
        }
        /* Values for `limit`, per the spec:
         * If undefined: 4294967295 // Math.pow(2, 32) - 1
         * If 0, Infinity, or NaN: 0
         * If positive number: limit = Math.floor(limit); if (limit > 4294967295) limit -= 4294967296;
         * If negative number: 4294967296 - Math.floor(Math.abs(limit))
         * If other: Type-convert, then use the above rules
         */
        limit = limit === undef ?
            -1 >>> 0 : // Math.pow(2, 32) - 1
            limit >>> 0; // ToUint32(limit)
        while (match = separator.exec(str)) {
            // `separator.lastIndex` is not reliable cross-browser
            lastIndex = match.index + match[0].length;
            if (lastIndex > lastLastIndex) {
                output.push(str.slice(lastLastIndex, match.index));
                // Fix browsers whose `exec` methods don't consistently return `undefined` for
                // nonparticipating capturing groups
                if (!compliantExecNpcg && match.length > 1) {
                    match[0].replace(separator2, function () {
                        for (var i = 1; i < arguments.length - 2; i++) {
                            if (arguments[i] === undef) {
                                match[i] = undef;
                            }
                        }
                    });
                }
                if (match.length > 1 && match.index < str.length) {
                    Array.prototype.push.apply(output, match.slice(1));
                }
                lastLength = match[0].length;
                lastLastIndex = lastIndex;
                if (output.length >= limit) {
                    break;
                }
            }
            if (separator.lastIndex === match.index) {
                separator.lastIndex++; // Avoid an infinite loop
            }
        }
        if (lastLastIndex === str.length) {
            if (lastLength || !separator.test("")) {
                output.push("");
            }
        } else {
            output.push(str.slice(lastLastIndex));
        }
        return output.length > limit ? output.slice(0, limit) : output;
    };

    // For convenience
    String.prototype.split = function (separator, limit) {
        return self(this, separator, limit);
    };

    return self;

}();

Download it.

Please let me know if you find any problems. Thanks!

Update: This script has become part of my XRegExp library, which includes many other JavaScript regular expression cross-browser compatibility fixes.

Mimicking Lookbehind in JavaScript

Unlike lookaheads, JavaScript doesn't support regex lookbehind syntax. That's unfortunate, but I'm not content with just resigning to that fact. Following are three ways I've come up with to mimic lookbehinds in JavaScript.

For those not familar with the concept of lookbehinds, they are zero-width assertions which, like the more specific \b, ^, and $ metacharacters, don't actually consume anything — they just match a position within text. This can be a very powerful concept. Read this first if you need more details.

Mimicking lookbehind with the replace method and optional capturing groups

This first approach is not much like a real lookbehind, but it might be "good enough" in some simple cases. Here are a few examples:

// Mimic leading, positive lookbehind like replace(/(?<=es)t/g, 'x')
var output = 'testt'.replace(/(es)?t/g, function($0, $1){
	return $1 ? $1 + 'x' : $0;
});
// output: tesxt

// Mimic leading, negative lookbehind like replace(/(?<!es)t/g, 'x')
var output = 'testt'.replace(/(es)?t/g, function($0, $1){
	return $1 ? $0 : 'x';
});
// output: xestx

// Mimic inner, positive lookbehind like replace(/\w(?<=s)t/g, 'x')
var output = 'testt'.replace(/(?:(s)|\w)t/g, function($0, $1){
	return $1 ? 'x' : $0;
});
// output: text

Unfortunately, there are many cases where lookbehinds can't be mimicked using this construct. Here's one example:

// Trying to mimic positive lookbehind, but this doesn't work
var output = 'ttttt'.replace(/(t)?t/g, function($0, $1){
	return $1 ? $1 + 'x' : $0;
});
// output: txtxt
// desired output: txxxx

The problem is that the regexes are relying on actually consuming the characters which should be within zero-width lookbehind assertions, then simply putting back the match unviolated (an effective no-op) if the backreferences contain or don't contain a value. Since the actual matching process here doesn't work anything like real lookbehinds, this only works in a limited number of scenarios. Additionally, it only works with the replace method, since other regex-related methods don't offer a mechanism to dynamically "undo" matches. However, since you can run arbitrary code in the replacement function, it does offer a limited degree of flexibility.

Mimicking lookbehind through reversal

The next approach uses lookaheads to mimic lookbehinds, and relies on manually reversing the data and writing your regex backwards. You'll also need to write the replacement value backwards if using this with the replace method, flip the match index if using this with the search method, etc. If that sounds a bit confusing, it is. I'll show an example in a second, but first we need a way to reverse our test string, since JavaScript doesn't provide this capability natively.

String.prototype.reverse = function () {
	return this.split('').reverse().join('');
};

Now let's try to pull this off:

// Mimicking lookbehind like (?<=es)t
var output = 'testt'.reverse().replace(/t(?=se)/g, 'x').reverse();
// output: tesxt

That actually works quite nicely, and allows mimicking both positive and negative lookbehind. However, writing a more complex regex with all nodes reversed can get a bit confusing, and since lookahead is used to mimic lookbehind, you can't mix what you intend as real lookaheads in the same pattern.

Note that reversing a string and applying regexes with reversed nodes can actually open up entirely new ways to approach a pattern, and in a few cases might make your code faster, even with the overhead of reversing the data. I'll have to save the efficiency discussion for another day, but before moving on to the third lookbehind-mimicking approach, here's one example of a new pattern approach made possible through reversal.

In my last post, I used the following code to add commas every three digits from the right for all numbers which are not preceded by a dot, letter, or underscore:

String.prototype.commafy = function () {
	return this.replace(/(^|[^\w.])(\d{4,})/g, function($0, $1, $2) {
		return $1 + $2.replace(/\d(?=(?:\d\d\d)+(?!\d))/g, '$&,');
	});
}

Here's an alternative implementation:

String.prototype.commafy = function() {
	return this.
		reverse().
		replace(/\d\d\d(?=\d)(?!\d*[a-z._])/gi, '$&,').
		reverse();
};

I'll leave the analysis for your free time.

Finally, we come to the third lookbehind-mimicking approach:

Mimicking lookbehind using a while loop and regexp.lastIndex

This last approach has the following advantages:

  • It's easier to use (no need to reverse your data and regex nodes).
  • It allows lookahead and lookbehind to be used together.
  • It allows you to more easily automate the mimicking process.

However, the trade off is that, in order to avoid interfering with standard regex backtracking, this approach only allows you to use lookbehinds (positive or negative) at the very start and/or end of your regexes. Fortunately, it's quite common to want to use a lookbehind at the start of a regex.

If you're not already familiar with the exec method available for RegExp objects, make sure to read about it at the Mozilla Developer Center before continuing. In particular, look at the examples which use exec within a while loop.

Here's a quick implementation of this approach, in which we'll actually toy with the regex engine's bump-along mechanism to get it to work as we want:

var data = 'ttttt',
	regex = /t/g,
	replacement = 'x',
	match,
	lastLastIndex = 0,
	output = '';

regex.x = {
	gRegex: /t/g,
	startLb: {
		regex: /t$/,
		type: true
	}
};

function lookbehind (data, regex, match) {
	return (
		(regex.x.startLb ? (regex.x.startLb.regex.test(data.substring(0, match.index)) === regex.x.startLb.type) : true) &&
		(regex.x.endLb ? (regex.x.endLb.regex.test(data.substring(0, regex.x.gRegex.lastIndex)) === regex.x.endLb.type) : true)
	);
}

while (match = regex.x.gRegex.exec(data)) {
	/* If the match is preceded/not by start lookbehind, and the end of the match is preceded/not by end lookbehind */
	if (lookbehind(data, regex, match)) {
		/* replacement can be a function */
		output += data.substring(lastLastIndex, match.index) + match[0].replace(regex, replacement);
		if(!regex.global){
			lastLastIndex = regex.gRegex.lastIndex;
			break;
		}
	/* If the inner pattern matched, but the leading or trailing lookbehind failed */
	} else {
		output += match[0].charAt(0);
		/* Set the regex to try again one character after the failed position, rather than at the end of the last match */
		regex.x.gRegex.lastIndex = match.index + 1;
	}
	lastLastIndex = regex.x.gRegex.lastIndex;
}
output += data.substring(lastLastIndex);

// output: txxxx

That's a fair bit of code, but it's quite powerful. It accounts for using both a leading and trailing lookbehind, and allows using a function for the replacement value. Also, this could relatively easily be made into a function which accepts a string for the regex using normal lookbehind syntax (e.g., "(?<=x)x(?<!x)"), then splits it into the various parts in needs before applying it.

Notes:

  • regex.x.gRegex should be an exact copy of regex, with the difference that it must use the g flag whether or not regex does (in order for the exec method to interact with the while loop as we need it to).
  • regex.x.startLb.type and regex.x.endLb.type use true for "positive," and false for "negative."
  • regex.x.startLb.regex and regex.x.endLb.regex are the patterns you want to use for the lookbehinds, but they must contain a trailing $. The dollar sign in this case does not mean end of the data, but rather end of the data segment they will be tested against.

If you're wondering why there hasn't been any discussion of fixed- vs. variable-length lookbehinds, that's because none of these approaches have any such limitations. They support full, variable-length lookbehind, which no regex engines I know of other than .NET and JGsoft (used by products like RegexBuddy) are capable of.

In conclusion, if you take advantage of all of the above approaches, regex lookbehind syntax can be mimicked in JavaScript in the vast majority of cases. Make sure to take advantage of the comment button if you have feedback about any of this stuff.

Update 2012-04: See my followup blog post, JavaScript Regex Lookbehind Redux, where I've posted a collection of short functions that make it much easier to simulate leading lookbehind.

Commafy Numbers

I've never used the few scripts I've seen that add commas to numbers because usually I want to apply the functionality to entire blocks of text. Having to pull out numbers, add commas, then put them back becomes a needlessly complex task without a method which can just do this in one shot. So, here's my attempt at this (if JavaScript regexes supported lookbehind, it could be even shorter):

String.prototype.commafy = function () {
	return this.replace(/(^|[^\w.])(\d{4,})/g, function($0, $1, $2) {
		return $1 + $2.replace(/\d(?=(?:\d\d\d)+(?!\d))/g, "$&,");
	});
}

Number.prototype.commafy = function () {
	return String(this).commafy();
}

Here are a couple examples of how this can be used:

(1000).commafy();
// Output: 1,000

var data = '1\n' +
	'10\n' +
	'100\n' +
	'1000\n' +
	'10000\n' +
	'100000\n' +
	'1000000\n' +
	'12345678901234567890\n' +
	'1000.99\n' +
	'1000.9999\n' +
	'.9999\n' +
	'-1000\n' +
	'$1000\n' +
	'"1000"\n' +
	'1000MHz\n' +
	'Z1000';

data.commafy();
/* Output:
1
10
100
1,000
10,000
100,000
1,000,000
12,345,678,901,234,567,890
1,000.99
1,000.9999
.9999
-1,000
$1,000
"1,000"
1,000MHz
Z1000
*/

Note that it adds commas to numbers followed by non-numeric characters, but avoids adding commas to numbers immediately preceded by a dot (decimal point), letter, or underscore. And as shown, this can be applied to individual numbers or entire blocks of text.

This is a decent example of where regular expressions can help to shorten and simplify code even in places you may not initially think to use them.


Edit: I've included an alternative implementation of the above code in my post Mimicking Lookbehind in JavaScript, which temporarily reverses the string to allow a simplified approach.

JavaScript Date Format

Update: The documentation below has been updated for the new Date Format 1.2. Get it now!

Although JavaScript provides a bunch of methods for getting and setting parts of a date object, it lacks a simple way to format dates and times according to a user-specified mask. There are a few scripts out there which provide this functionality, but I've never seen one that worked well for me… Most are needlessly bulky or slow, tie in unrelated functionality, use complicated mask syntaxes that more or less require you to read the documentation every time you want to use them, or don't account for special cases like escaping mask characters within the generated string.

When choosing which special mask characters to use for my JavaScript date formatter, I looked at PHP's date function and ColdFusion's discrete dateFormat and timeFormat functions. PHP uses a crazy mix of letters (to me at least, since I'm not a PHP programmer) to represent various date entities, and while I'll probably never memorize the full list, it does offer the advantages that you can apply both date and time formatting with one function, and that none of the special characters overlap (unlike ColdFusion where m and mm mean different things depending on whether you're dealing with dates or times). On the other hand, ColdFusion uses very easy to remember special characters for masks.

With my date formatter, I've tried to take the best features from both, and add some sugar of my own. It did end up a lot like the ColdFusion implementation though, since I've primarily used CF's mask syntax.

Before getting into further details, here are some examples of how this script can be used:

var now = new Date();

now.format("m/dd/yy");
// Returns, e.g., 6/09/07

// Can also be used as a standalone function
dateFormat(now, "dddd, mmmm dS, yyyy, h:MM:ss TT");
// Saturday, June 9th, 2007, 5:46:21 PM

// You can use one of several named masks
now.format("isoDateTime");
// 2007-06-09T17:46:21

// ...Or add your own
dateFormat.masks.hammerTime = 'HH:MM! "Can\'t touch this!"';
now.format("hammerTime");
// 17:46! Can't touch this!

// When using the standalone dateFormat function,
// you can also provide the date as a string
dateFormat("Jun 9 2007", "fullDate");
// Saturday, June 9, 2007

// Note that if you don't include the mask argument,
// dateFormat.masks.default is used
now.format();
// Sat Jun 09 2007 17:46:21

// And if you don't include the date argument,
// the current date and time is used
dateFormat();
// Sat Jun 09 2007 17:46:22

// You can also skip the date argument (as long as your mask doesn't
// contain any numbers), in which case the current date/time is used
dateFormat("longTime");
// 5:46:22 PM EST

// And finally, you can convert local time to UTC time. Either pass in
// true as an additional argument (no argument skipping allowed in this case):
dateFormat(now, "longTime", true);
now.format("longTime", true);
// Both lines return, e.g., 10:46:21 PM UTC

// ...Or add the prefix "UTC:" to your mask.
now.format("UTC:h:MM:ss TT Z");
// 10:46:21 PM UTC

Following are the special characters supported. Any differences in meaning from ColdFusion's dateFormat and timeFormat functions are noted.

Mask Description
d Day of the month as digits; no leading zero for single-digit days.
dd Day of the month as digits; leading zero for single-digit days.
ddd Day of the week as a three-letter abbreviation.
dddd Day of the week as its full name.
m Month as digits; no leading zero for single-digit months.
mm Month as digits; leading zero for single-digit months.
mmm Month as a three-letter abbreviation.
mmmm Month as its full name.
yy Year as last two digits; leading zero for years less than 10.
yyyy Year represented by four digits.
h Hours; no leading zero for single-digit hours (12-hour clock).
hh Hours; leading zero for single-digit hours (12-hour clock).
H Hours; no leading zero for single-digit hours (24-hour clock).
HH Hours; leading zero for single-digit hours (24-hour clock).
M Minutes; no leading zero for single-digit minutes.
Uppercase M unlike CF timeFormat's m to avoid conflict with months.
MM Minutes; leading zero for single-digit minutes.
Uppercase MM unlike CF timeFormat's mm to avoid conflict with months.
s Seconds; no leading zero for single-digit seconds.
ss Seconds; leading zero for single-digit seconds.
l or L Milliseconds. l gives 3 digits. L gives 2 digits.
t Lowercase, single-character time marker string: a or p.
No equivalent in CF.
tt Lowercase, two-character time marker string: am or pm.
No equivalent in CF.
T Uppercase, single-character time marker string: A or P.
Uppercase T unlike CF's t to allow for user-specified casing.
TT Uppercase, two-character time marker string: AM or PM.
Uppercase TT unlike CF's tt to allow for user-specified casing.
Z US timezone abbreviation, e.g. EST or MDT. With non-US timezones or in the Opera browser, the GMT/UTC offset is returned, e.g. GMT-0500
No equivalent in CF.
o GMT/UTC timezone offset, e.g. -0500 or +0230.
No equivalent in CF.
S The date's ordinal suffix (st, nd, rd, or th). Works well with d.
No equivalent in CF.
'…' or "…" Literal character sequence. Surrounding quotes are removed.
No equivalent in CF.
UTC: Must be the first four characters of the mask. Converts the date from local time to UTC/GMT/Zulu time before applying the mask. The "UTC:" prefix is removed.
No equivalent in CF.

And here are the named masks provided by default (you can easily change these or add your own):

Name Mask Example
default ddd mmm dd yyyy HH:MM:ss Sat Jun 09 2007 17:46:21
shortDate m/d/yy 6/9/07
mediumDate mmm d, yyyy Jun 9, 2007
longDate mmmm d, yyyy June 9, 2007
fullDate dddd, mmmm d, yyyy Saturday, June 9, 2007
shortTime h:MM TT 5:46 PM
mediumTime h:MM:ss TT 5:46:21 PM
longTime h:MM:ss TT Z 5:46:21 PM EST
isoDate yyyy-mm-dd 2007-06-09
isoTime HH:MM:ss 17:46:21
isoDateTime yyyy-mm-dd'T'HH:MM:ss 2007-06-09T17:46:21
isoUtcDateTime UTC:yyyy-mm-dd'T'HH:MM:ss'Z' 2007-06-09T22:46:21Z

A couple issues:

  • In the unlikely event that there is ambiguity in the meaning of your mask (e.g., m followed by mm, with no separating characters), put a pair of empty quotes between your metasequences. The quotes will be removed automatically.
  • If you need to include literal quotes in your mask, the following rules apply:
    • Unpaired quotes do not need special handling.
    • To include literal quotes inside masks which contain any other quote marks of the same type, you need to enclose them with the alternative quote type (i.e., double quotes for single quotes, and vice versa). E.g., date.format('h "o\'clock, y\'all!"') returns "6 o'clock, y'all". This can get a little hairy, perhaps, but I doubt people will really run into it that often. The previous example can also be written as date.format("h") + "o'clock, y'all!".

Here's the code:

/*
 * Date Format 1.2.3
 * (c) 2007-2009 Steven Levithan <stevenlevithan.com>
 * MIT license
 *
 * Includes enhancements by Scott Trenda <scott.trenda.net>
 * and Kris Kowal <cixar.com/~kris.kowal/>
 *
 * Accepts a date, a mask, or a date and a mask.
 * Returns a formatted version of the given date.
 * The date defaults to the current date/time.
 * The mask defaults to dateFormat.masks.default.
 */

var dateFormat = function () {
	var	token = /d{1,4}|m{1,4}|yy(?:yy)?|([HhMsTt])\1?|[LloSZ]|"[^"]*"|'[^']*'/g,
		timezone = /\b(?:[PMCEA][SDP]T|(?:Pacific|Mountain|Central|Eastern|Atlantic) (?:Standard|Daylight|Prevailing) Time|(?:GMT|UTC)(?:[-+]\d{4})?)\b/g,
		timezoneClip = /[^-+\dA-Z]/g,
		pad = function (val, len) {
			val = String(val);
			len = len || 2;
			while (val.length < len) val = "0" + val;
			return val;
		};

	// Regexes and supporting functions are cached through closure
	return function (date, mask, utc) {
		var dF = dateFormat;

		// You can't provide utc if you skip other args (use the "UTC:" mask prefix)
		if (arguments.length == 1 && Object.prototype.toString.call(date) == "[object String]" && !/\d/.test(date)) {
			mask = date;
			date = undefined;
		}

		// Passing date through Date applies Date.parse, if necessary
		date = date ? new Date(date) : new Date;
		if (isNaN(date)) throw SyntaxError("invalid date");

		mask = String(dF.masks[mask] || mask || dF.masks["default"]);

		// Allow setting the utc argument via the mask
		if (mask.slice(0, 4) == "UTC:") {
			mask = mask.slice(4);
			utc = true;
		}

		var	_ = utc ? "getUTC" : "get",
			d = date[_ + "Date"](),
			D = date[_ + "Day"](),
			m = date[_ + "Month"](),
			y = date[_ + "FullYear"](),
			H = date[_ + "Hours"](),
			M = date[_ + "Minutes"](),
			s = date[_ + "Seconds"](),
			L = date[_ + "Milliseconds"](),
			o = utc ? 0 : date.getTimezoneOffset(),
			flags = {
				d:    d,
				dd:   pad(d),
				ddd:  dF.i18n.dayNames[D],
				dddd: dF.i18n.dayNames[D + 7],
				m:    m + 1,
				mm:   pad(m + 1),
				mmm:  dF.i18n.monthNames[m],
				mmmm: dF.i18n.monthNames[m + 12],
				yy:   String(y).slice(2),
				yyyy: y,
				h:    H % 12 || 12,
				hh:   pad(H % 12 || 12),
				H:    H,
				HH:   pad(H),
				M:    M,
				MM:   pad(M),
				s:    s,
				ss:   pad(s),
				l:    pad(L, 3),
				L:    pad(L > 99 ? Math.round(L / 10) : L),
				t:    H < 12 ? "a"  : "p",
				tt:   H < 12 ? "am" : "pm",
				T:    H < 12 ? "A"  : "P",
				TT:   H < 12 ? "AM" : "PM",
				Z:    utc ? "UTC" : (String(date).match(timezone) || [""]).pop().replace(timezoneClip, ""),
				o:    (o > 0 ? "-" : "+") + pad(Math.floor(Math.abs(o) / 60) * 100 + Math.abs(o) % 60, 4),
				S:    ["th", "st", "nd", "rd"][d % 10 > 3 ? 0 : (d % 100 - d % 10 != 10) * d % 10]
			};

		return mask.replace(token, function ($0) {
			return $0 in flags ? flags[$0] : $0.slice(1, $0.length - 1);
		});
	};
}();

// Some common format strings
dateFormat.masks = {
	"default":      "ddd mmm dd yyyy HH:MM:ss",
	shortDate:      "m/d/yy",
	mediumDate:     "mmm d, yyyy",
	longDate:       "mmmm d, yyyy",
	fullDate:       "dddd, mmmm d, yyyy",
	shortTime:      "h:MM TT",
	mediumTime:     "h:MM:ss TT",
	longTime:       "h:MM:ss TT Z",
	isoDate:        "yyyy-mm-dd",
	isoTime:        "HH:MM:ss",
	isoDateTime:    "yyyy-mm-dd'T'HH:MM:ss",
	isoUtcDateTime: "UTC:yyyy-mm-dd'T'HH:MM:ss'Z'"
};

// Internationalization strings
dateFormat.i18n = {
	dayNames: [
		"Sun", "Mon", "Tue", "Wed", "Thu", "Fri", "Sat",
		"Sunday", "Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday"
	],
	monthNames: [
		"Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec",
		"January", "February", "March", "April", "May", "June", "July", "August", "September", "October", "November", "December"
	]
};

// For convenience...
Date.prototype.format = function (mask, utc) {
	return dateFormat(this, mask, utc);
};

Download it here (1.2 KB when minified and gzipped).

Note that the day and month names can be changed (for internationalization or other purposes) by updating the dateFormat.i18n object.

If you have any suggestions or find any issues, lemme know.


My new blog: Want to learn about aphantasisa and hyperphantasia, the Shen Yun cult, or establishing South Dakota residency to take advantage of nomad friendly laws? All of these and more await at Life After Tech.

Faster JavaScript Trim

Since JavaScript doesn't include a trim method natively, it's included by countless JavaScript libraries – usually as a global function or appended to String.prototype. However, I've never seen an implementation which performs as well as it could, probably because most programmers don't deeply understand or care about regex efficiency issues.

After seeing a particularly bad trim implementation, I decided to do a little research towards finding the most efficient approach. Before getting into the analysis, here are the results:

Method Firefox 2 IE 6
trim1 15ms < 0.5ms
trim2 31ms < 0.5ms
trim3 46ms 31ms
trim4 47ms 46ms
trim5 156ms 1656ms
trim6 172ms 2406ms
trim7 172ms 1640ms
trim8 281ms < 0.5ms
trim9 125ms 78ms
trim10 < 0.5ms < 0.5ms
trim11 < 0.5ms < 0.5ms

Note 1: The comparison is based on trimming the Magna Carta (over 27,600 characters) with a bit of leading and trailing whitespace 20 times on my personal system. However, the data you're trimming can have a major impact on performance, which is detailed below.

Note 2: trim4 and trim6 are the most commonly found in JavaScript libraries today.

Note 3: The aforementioned bad implementation is not included in the comparison, but is shown later.

The analysis

Although there are 11 rows in the table above, they are only the most notable (for various reasons) of about 20 versions I wrote and benchmarked against various types of strings. The following analysis is based on testing in Firefox 2.0.0.4, although I have noted where there are major differences in IE6.

  1. return str.replace(/^\s\s*/, '').replace(/\s\s*$/, '');
    All things considered, this is probably the best all-around approach. Its speed advantage is most notable with long strings — when efficiency matters. The speed is largely due to a number of optimizations internal to JavaScript regex interpreters which the two discrete regexes here trigger. Specifically, the pre-check of required character and start of string anchor optimizations, possibly among others.
  2. return str.replace(/^\s+/, '').replace(/\s+$/, '');
    Very similar to trim1 (above), but a little slower since it doesn't trigger all of the same optimizations.
  3. return str.substring(Math.max(str.search(/\S/), 0), str.search(/\S\s*$/) + 1);
    This is often faster than the following methods, but slower than the above two. Its speed comes from its use of simple, character-index lookups.
  4. return str.replace(/^\s+|\s+$/g, '');
    This commonly thought up approach is easily the most frequently used in JavaScript libraries today. It is generally the fastest implementation of the bunch only when working with short strings which don't include leading or trailing whitespace. This minor advantage is due in part to the initial-character discrimination optimization it triggers. While this is a relatively decent performer, it's slower than the three methods above when working with longer strings, because the top-level alternation prevents a number of optimizations which could otherwise kick in.
  5. str = str.match(/\S+(?:\s+\S+)*/);
    return str ? str[0] : '';

    This is generally the fastest method when working with empty or whitespace-only strings, due to the pre-check of required character optimization it triggers. Note: In IE6, this can be quite slow when working with longer strings.
  6. return str.replace(/^\s*(\S*(\s+\S+)*)\s*$/, '$1');
    This is a relatively common approach, popularized in part by some leading JavaScripters. It's similar in approach (but inferior) to trim8. There's no good reason to use this in JavaScript, especially since it can be very slow in IE6.
  7. return str.replace(/^\s*(\S*(?:\s+\S+)*)\s*$/, '$1');
    The same as trim6, but a bit faster due to the use of a non-capturing group (which doesn't work in IE 5.0 and lower). Again, this can be slow in IE6.
  8. return str.replace(/^\s*((?:[\S\s]*\S)?)\s*$/, '$1');
    This uses a simple, single-pass, greedy approach. In IE6, this is crazy fast! The performance difference indicates that IE has superior optimization for quantification of "any character" tokens.
  9. return str.replace(/^\s*([\S\s]*?)\s*$/, '$1');
    This is generally the fastest with very short strings which contain both non-space characters and edge whitespace. This minor advantage is due to the simple, single-pass, lazy approach it uses. Like trim8, this is significantly faster in IE6 than Firefox 2.

Since I've seen the following additional implementation in one library, I'll include it here as a warning:

return str.replace(/^\s*([\S\s]*)\b\s*$/, '$1');

Although the above is sometimes the fastest method when working with short strings which contain both non-space characters and edge whitespace, it performs very poorly with long strings which contain numerous word boundaries, and it's terrible (!) with long strings comprised of nothing but whitespace, since that triggers an exponentially increasing amount of backtracking. Do not use.

A different endgame

There are two methods in the table at the top of this post which haven't been covered yet. For those, I've used a non-regex and hybrid approach.

After comparing and analyzing all of the above, I wondered how an implementation which used no regular expressions would perform. Here's what I tried:

function trim10 (str) {
	var whitespace = ' \n\r\t\f\x0b\xa0\u2000\u2001\u2002\u2003\u2004\u2005\u2006\u2007\u2008\u2009\u200a\u200b\u2028\u2029\u3000';
	for (var i = 0; i < str.length; i++) {
		if (whitespace.indexOf(str.charAt(i)) === -1) {
			str = str.substring(i);
			break;
		}
	}
	for (i = str.length - 1; i >= 0; i--) {
		if (whitespace.indexOf(str.charAt(i)) === -1) {
			str = str.substring(0, i + 1);
			break;
		}
	}
	return whitespace.indexOf(str.charAt(0)) === -1 ? str : '';
}

How does that perform? Well, with long strings which do not contain excessive leading or trailing whitespace, it blows away the competition (except against trim1/2/8 in IE, which are already insanely fast there).

Does that mean regular expressions are slow in Firefox? No, not at all. The issue here is that although regexes are very well suited for trimming leading whitespace, apart from the .NET library (which offers a somewhat-mysterious "backwards matching" mode), they don't really provide a method to jump to the end of a string without even considering previous characters. However, the non-regex-reliant trim10 function does just that, with the second loop working backwards from the end of the string until it finds a non-whitespace character.

Knowing that, what if we created a hybrid implementation which combined a regex's universal efficiency at trimming leading whitespace with the alternative method's speed at removing trailing characters?

function trim11 (str) {
	str = str.replace(/^\s+/, '');
	for (var i = str.length - 1; i >= 0; i--) {
		if (/\S/.test(str.charAt(i))) {
			str = str.substring(0, i + 1);
			break;
		}
	}
	return str;
}

Although the above is a bit slower than trim10 with some strings, it uses significantly less code and is still lightning fast. Plus, with strings which contain a lot of leading whitespace (which includes strings comprised of nothing but whitespace), it's much faster than trim10.

In conclusion…

Since the differences between the implementations cross-browser and when used with different data are both complex and nuanced (none of them are faster than all the others with any data you can throw at it), here are my general recommendations for a trim method:

  • Use trim1 if you want a general-purpose implementation which is fast cross-browser.
  • Use trim11 if you want to handle long strings exceptionally fast in all browsers.

To test all of the above implementations for yourself, try my very rudimentary benchmarking page. Background processing can cause the results to be severely skewed, so run the test a number of times (regardless of how many iterations you specify) and only consider the fastest results (since averaging the cost of background interference is not very enlightening).

As a final note, although some people like to cache regular expressions (e.g. using global variables) so they can be used repeatedly without recompilation, IMO this does not make much sense for a trim method. All of the above regexes are so simple that they typically take no more than a nanosecond to compile. Additionally, some browsers automatically cache the most recently used regexes, so a typical loop which uses trim and doesn't contain a bunch of other regexes might not encounter recompilation anyway.


Edit (2008-02-04): Shortly after posting this I realized trim10/11 could be better written. Several people have also posted improved versions in the comments. Here's what I use now, which takes the trim11-style hybrid approach:

function trim12 (str) {
	var	str = str.replace(/^\s\s*/, ''),
		ws = /\s/,
		i = str.length;
	while (ws.test(str.charAt(--i)));
	return str.slice(0, i + 1);
}

New library: Are you a JavaScript regex master, or want to be? Then you need my fancy XRegExp library. It adds new regex syntax (including named capture and Unicode properties); s, x, and n flags; powerful regex utils; and it fixes pesky browser inconsistencies. Check it out!