UglifyJS – a JavaScript parser/compressor/beautifier

NEW: UglifyJS2

I started working on UglifyJS’s successor, version 2. It’s almost a full rewrite (except for the parser which is heavily modified, everything else starts from scratch). I’ve detailed my reasons in the README, see the project page.

https://github.com/mishoo/UglifyJS2

Version 1 will continue to be maintained for fixing show-stopper bugs, but no new features should be expected.

UglifyJS — a JavaScript parser/compressor/beautifier

This package implements a general-purpose JavaScript parser/compressor/beautifier toolkit. It is developed on NodeJS, but it should work on any JavaScript platform supporting the CommonJS module system (and if your platform of choice doesn’t support CommonJS, you can easily implement it, or discard the exports.* lines from UglifyJS sources).

The tokenizer/parser generates an abstract syntax tree from JS code. You can then traverse the AST to learn more about the code, or do various manipulations on it. This part is implemented in parse-js.js and it’s a port to JavaScript of the excellent parse-js Common Lisp library from Marijn Haverbeke.

( See cl-uglify-js if you’re looking for the Common Lisp version of UglifyJS. )

The second part of this package, implemented in process.js, inspects and manipulates the AST generated by the parser to provide the following:

• ability to re-generate JavaScript code from the AST. Optionally indented—you can use this if you want to “beautify” a program that has been compressed, so that you can inspect the source. But you can also run our code generator to print out an AST without any whitespace, so you achieve compression as well.
• shorten variable names (usually to single characters). Our mangler will analyze the code and generate proper variable names, depending on scope and usage, and is smart enough to deal with globals defined elsewhere, or with eval() calls or with{} statements. In short, if eval() or with{} are used in some scope, then all variables in that scope and any variables in the parent scopes will remain unmangled, and any references to such variables remain unmangled as well.
• various small optimizations that may lead to faster code but certainly lead to smaller code. Where possible, we do the following:
  • foo[“bar”] ==> foo.bar
  • remove block brackets {}
  • join consecutive var declarations: var a = 10; var b = 20; ==> var a=10,b=20;
  • resolve simple constant expressions: 1 +2 * 3 ==> 7. We only do the replacement if the result occupies less bytes; for example 1/3 would translate to 0.333333333333, so in this case we don’t replace it.
  • consecutive statements in blocks are merged into a sequence; in many cases, this leaves blocks with a single statement, so then we can remove the block brackets.
  • various optimizations for IF statements:
    • if (foo) bar(); else baz(); ==> foo?bar():baz();
    • if (!foo) bar(); else baz(); ==> foo?baz():bar();
    • if (foo) bar(); ==> foo&&bar();
    • if (!foo) bar(); ==> foo||bar();
    • if (foo) return bar(); else return baz(); ==> return foo?bar():baz();
    • if (foo) return bar(); else something(); ==> {if(foo)return bar();something()}
  • remove some unreachable code and warn about it (code that follows a return, throw, break or continue statement, except function/variable declarations).
  • act a limited version of a pre-processor (c.f. the pre-processor of C/C++) to allow you to safely replace selected global symbols with specified values. When combined with the optimisations above this can make UglifyJS operate slightly more like a compilation process, in that when certain symbols are replaced by constant values, entire code blocks may be optimised away as unreachable.

<<Unsafe transformations>>

The following transformations can in theory break code, although they’re probably safe in most practical cases. To enable them you need to pass the --unsafe flag.

Calls involving the global Array constructor

The following transformations occur:

new Array(1, 2, 3, 4)  => [1,2,3,4]
Array(a, b, c)         => [a,b,c]
new Array(5)           => Array(5)
new Array(a)           => Array(a)

These are all safe if the Array name isn’t redefined. JavaScript does allow one to globally redefine Array (and pretty much everything, in fact) but I personally don’t see why would anyone do that.

UglifyJS does handle the case where Array is redefined locally, or even globally but with a function or var declaration. Therefore, in the following cases UglifyJS doesn’t touch calls or instantiations of Array:

// case 1.  globally declared variable
  var Array;
  new Array(1, 2, 3);
  Array(a, b);

  // or (can be declared later)
  new Array(1, 2, 3);
  var Array;

  // or (can be a function)
  new Array(1, 2, 3);
  function Array() { ... }

// case 2.  declared in a function
  (function(){
    a = new Array(1, 2, 3);
    b = Array(5, 6);
    var Array;
  })();

  // or
  (function(Array){
    return Array(5, 6, 7);
  })();

  // or
  (function(){
    return new Array(1, 2, 3, 4);
    function Array() { ... }
  })();

  // etc.

obj.toString() ==> obj+“”

Install (NPM)

UglifyJS is now available through NPM — npm install uglify-js@1 should do the job.

NOTE: The NPM package has been upgraded to UglifyJS2. If you need to install version 1.x you need to add `@1` to the command, as I did above. I strongly suggest you to try to upgrade, though this might not be simple (v2 has a completely different AST structure and API).

Install latest code from GitHub

## clone the repository
mkdir -p /where/you/wanna/put/it
cd /where/you/wanna/put/it
git clone git://github.com/mishoo/UglifyJS.git

## make the module available to Node
mkdir -p ~/.node_libraries/
cd ~/.node_libraries/
ln -s /where/you/wanna/put/it/UglifyJS/uglify-js.js

## and if you want the CLI script too:
mkdir -p ~/bin
cd ~/bin
ln -s /where/you/wanna/put/it/UglifyJS/bin/uglifyjs
  # (then add ~/bin to your $PATH if it's not there already)

Usage

There is a command-line tool that exposes the functionality of this library for your shell-scripting needs:

uglifyjs [ options... ] [ filename ]

filename should be the last argument and should name the file from which to read the JavaScript code. If you don’t specify it, it will read code from STDIN.

Supported options:

• -b or --beautify — output indented code; when passed, additional options control the beautifier:
  • -i N or --indent N — indentation level (number of spaces)
  • -q or --quote-keys — quote keys in literal objects (by default, only keys that cannot be identifier names will be quotes).
• -c or ----consolidate-primitive-values — consolidates null, Boolean, and String values. Known as aliasing in the Closure Compiler. Worsens the data compression ratio of gzip.
• --ascii — pass this argument to encode non-ASCII characters as \uXXXX sequences. By default UglifyJS won’t bother to do it and will output Unicode characters instead. (the output is always encoded in UTF8, but if you pass this option you’ll only get ASCII).
• -nm or --no-mangle — don’t mangle names.
• -nmf or --no-mangle-functions – in case you want to mangle variable names, but not touch function names.
• -ns or --no-squeeze — don’t call ast_squeeze() (which does various optimizations that result in smaller, less readable code).
• -mt or --mangle-toplevel — mangle names in the toplevel scope too (by default we don’t do this).
• --no-seqs — when ast_squeeze() is called (thus, unless you pass --no-squeeze) it will reduce consecutive statements in blocks into a sequence. For example, “a = 10; b = 20; foo();” will be written as “a=10,b=20,foo();”. In various occasions, this allows us to discard the block brackets (since the block becomes a single statement). This is ON by default because it seems safe and saves a few hundred bytes on some libs that I tested it on, but pass --no-seqs to disable it.
• --no-dead-code — by default, UglifyJS will remove code that is obviously unreachable (code that follows a return, throw, break or continue statement and is not a function/variable declaration). Pass this option to disable this optimization.
• -nc or --no-copyright — by default, uglifyjs will keep the initial comment tokens in the generated code (assumed to be copyright information etc.). If you pass this it will discard it.
• -o filename or --output filename — put the result in filename. If this isn’t given, the result goes to standard output (or see next one).
• --overwrite — if the code is read from a file (not from STDIN) and you pass --overwrite then the output will be written in the same file.
• --ast — pass this if you want to get the Abstract Syntax Tree instead of JavaScript as output. Useful for debugging or learning more about the internals.
• -v or --verbose — output some notes on STDERR (for now just how long each operation takes).
• -d SYMBOL[=VALUE] or --define SYMBOL[=VALUE] — will replace all instances of the specified symbol where used as an identifier (except where symbol has properly declared by a var declaration or use as function parameter or similar) with the specified value. This argument may be specified multiple times to define multiple symbols - if no value is specified the symbol will be replaced with the value true, or you can specify a numeric value (such as 1024), a quoted string value (such as =”object”= or =’https://github.com’), or the name of another symbol or keyword (such as =null or document). This allows you, for example, to assign meaningful names to key constant values but discard the symbolic names in the uglified version for brevity/efficiency, or when used wth care, allows UglifyJS to operate as a form of conditional compilation whereby defining appropriate values may, by dint of the constant folding and dead code removal features above, remove entire superfluous code blocks (e.g. completely remove instrumentation or trace code for production use). Where string values are being defined, the handling of quotes are likely to be subject to the specifics of your command shell environment, so you may need to experiment with quoting styles depending on your platform, or you may find the option --define-from-module more suitable for use.
• -define-from-module SOMEMODULE — will load the named module (as per the NodeJS require() function) and iterate all the exported properties of the module defining them as symbol names to be defined (as if by the --define option) per the name of each property (i.e. without the module name prefix) and given the value of the property. This is a much easier way to handle and document groups of symbols to be defined rather than a large number of --define options.
• --unsafe — enable other additional optimizations that are known to be unsafe in some contrived situations, but could still be generally useful. For now only these:
  • foo.toString() ==> foo+””
  • new Array(x,…) ==> [x,…]
  • new Array(x) ==> Array(x)
• --max-line-len (default 32K characters) — add a newline after around 32K characters. I’ve seen both FF and Chrome croak when all the code was on a single line of around 670K. Pass –max-line-len 0 to disable this safety feature.
• --reserved-names — some libraries rely on certain names to be used, as pointed out in issue #92 and #81, so this option allow you to exclude such names from the mangler. For example, to keep names require and $super intact you’d specify –reserved-names “require,$super”.
• --inline-script – when you want to include the output literally in an HTML <script> tag you can use this option to prevent </script from showing up in the output.
• --lift-vars – when you pass this, UglifyJS will apply the following transformations (see the notes in API, ast_lift_variables):
  • put all var declarations at the start of the scope
  • make sure a variable is declared only once
  • discard unused function arguments
  • discard unused inner (named) functions
  • finally, try to merge assignments into that one var declaration, if possible.

API

To use the library from JavaScript, you’d do the following (example for NodeJS):

var jsp = require("uglify-js").parser;
var pro = require("uglify-js").uglify;

var orig_code = "... JS code here";
var ast = jsp.parse(orig_code); // parse code and get the initial AST
ast = pro.ast_mangle(ast); // get a new AST with mangled names
ast = pro.ast_squeeze(ast); // get an AST with compression optimizations
var final_code = pro.gen_code(ast); // compressed code here

The above performs the full compression that is possible right now. As you can see, there are a sequence of steps which you can apply. For example if you want compressed output but for some reason you don’t want to mangle variable names, you would simply skip the line that calls pro.ast_mangle(ast).

Some of these functions take optional arguments. Here’s a description:

• jsp.parse(code, strict_semicolons) – parses JS code and returns an AST. strict_semicolons is optional and defaults to false. If you pass true then the parser will throw an error when it expects a semicolon and it doesn’t find it. For most JS code you don’t want that, but it’s useful if you want to strictly sanitize your code.
• pro.ast_lift_variables(ast) – merge and move var declarations to the scop of the scope; discard unused function arguments or variables; discard unused (named) inner functions. It also tries to merge assignments following the var declaration into it.

  If your code is very hand-optimized concerning var declarations, this lifting variable declarations might actually increase size. For me it helps out. On jQuery it adds 865 bytes (243 after gzip). YMMV. Also note that (since it’s not enabled by default) this operation isn’t yet heavily tested (please report if you find issues!).

  Note that although it might increase the image size (on jQuery it gains 865 bytes, 243 after gzip) it’s technically more correct: in certain situations, dead code removal might drop variable declarations, which would not happen if the variables are lifted in advance.

  Here’s an example of what it does:

function f(a, b, c, d, e) {
    var q;
    var w;
    w = 10;
    q = 20;
    for (var i = 1; i < 10; ++i) {
        var boo = foo(a);
    }
    for (var i = 0; i < 1; ++i) {
        var boo = bar(c);
    }
    function foo(){ ... }
    function bar(){ ... }
    function baz(){ ... }
}

// transforms into ==>

function f(a, b, c) {
    var i, boo, w = 10, q = 20;
    for (i = 1; i < 10; ++i) {
        boo = foo(a);
    }
    for (i = 0; i < 1; ++i) {
        boo = bar(c);
    }
    function foo() { ... }
    function bar() { ... }
}

• pro.ast_mangle(ast, options) – generates a new AST containing mangled (compressed) variable and function names. It supports the following options:
  • toplevel – mangle toplevel names (by default we don’t touch them).
  • except – an array of names to exclude from compression.
  • defines – an object with properties named after symbols to replace (see the --define option for the script) and the values representing the AST replacement value. For example, { defines: { DEBUG: ['name', 'false'], VERSION: ['string', '1.0'] } }
• pro.ast_squeeze(ast, options) – employs further optimizations designed to reduce the size of the code that gen_code would generate from the AST. Returns a new AST. options can be a hash; the supported options are:
  • make_seqs (default true) which will cause consecutive statements in a block to be merged using the “sequence” (comma) operator
  • dead_code (default true) which will remove unreachable code.
• pro.gen_code(ast, options) – generates JS code from the AST. By default it’s minified, but using the options argument you can get nicely formatted output. options is, well, optional :-) and if you pass it it must be an object and supports the following properties (below you can see the default values):
  • beautify: false – pass true if you want indented output
  • indent_start: 0 (only applies when beautify is true) – initial indentation in spaces
  • indent_level: 4 (only applies when beautify is true) – indentation level, in spaces (pass an even number)
  • quote_keys: false – if you pass true it will quote all keys in literal objects
  • space_colon: false (only applies when beautify is true) – wether to put a space before the colon in object literals
  • ascii_only: false – pass true if you want to encode non-ASCII characters as \uXXXX.
  • inline_script: false – pass true to escape occurrences of </script in strings

Beautifier shortcoming – no more comments

The beautifier can be used as a general purpose indentation tool. It’s useful when you want to make a minified file readable. One limitation, though, is that it discards all comments, so you don’t really want to use it to reformat your code, unless you don’t have, or don’t care about, comments.

In fact it’s not the beautifier who discards comments — they are dumped at the parsing stage, when we build the initial AST. Comments don’t really make sense in the AST, and while we could add nodes for them, it would be inconvenient because we’d have to add special rules to ignore them at all the processing stages.

Use as a code pre-processor

The --define option can be used, particularly when combined with the constant folding logic, as a form of pre-processor to enable or remove particular constructions, such as might be used for instrumenting development code, or to produce variations aimed at a specific platform.

The code below illustrates the way this can be done, and how the symbol replacement is performed.

CLAUSE1: if (typeof DEVMODE === 'undefined') {
    DEVMODE = true;
}

CLAUSE2: function init() {
    if (DEVMODE) {
        console.log("init() called");
    }
    ....
    DEVMODE && console.log("init() complete");
}

CLAUSE3: function reportDeviceStatus(device) {
    var DEVMODE = device.mode, DEVNAME = device.name;
    if (DEVMODE === 'open') {
        ....
    }
}

When the above code is normally executed, the undeclared global variable DEVMODE will be assigned the value true (see CLAUSE1) and so the init() function (CLAUSE2) will write messages to the console log when executed, but in CLAUSE3 a locally declared variable will mask access to the DEVMODE global symbol.

If the above code is processed by UglifyJS with an argument of --define DEVMODE=false then UglifyJS will replace DEVMODE with the boolean constant value false within CLAUSE1 and CLAUSE2, but it will leave CLAUSE3 as it stands because there DEVMODE resolves to a validly declared variable.

And more so, the constant-folding features of UglifyJS will recognise that the if condition of CLAUSE1 is thus always false, and so will remove the test and body of CLAUSE1 altogether (including the otherwise slightly problematical statement false = true; which it will have formed by replacing DEVMODE in the body). Similarly, within CLAUSE2 both calls to console.log() will be removed altogether.

In this way you can mimic, to a limited degree, the functionality of the C/C++ pre-processor to enable or completely remove blocks depending on how certain symbols are defined - perhaps using UglifyJS to generate different versions of source aimed at different environments

It is recommmended (but not made mandatory) that symbols designed for this purpose are given names consisting of UPPER_CASE_LETTERS to distinguish them from other (normal) symbols and avoid the sort of clash that CLAUSE3 above illustrates.

Compression – how good is it?

Here are updated statistics. (I also updated my Google Closure and YUI installations).

We’re still a lot better than YUI in terms of compression, though slightly slower. We’re still a lot faster than Closure, and compression after gzip is comparable.

File	UglifyJS	UglifyJS+gzip	Closure	Closure+gzip	YUI	YUI+gzip
jquery-1.6.2.js	91001 (0:01.59)	31896	90678 (0:07.40)	31979	101527 (0:01.82)	34646
paper.js	142023 (0:01.65)	43334	134301 (0:07.42)	42495	173383 (0:01.58)	48785
prototype.js	88544 (0:01.09)	26680	86955 (0:06.97)	26326	92130 (0:00.79)	28624
thelib-full.js (DynarchLIB)	251939 (0:02.55)	72535	249911 (0:09.05)	72696	258869 (0:01.94)	76584

Bugs?

Unfortunately, for the time being there is no automated test suite. But I ran the compressor manually on non-trivial code, and then I tested that the generated code works as expected. A few hundred times.

DynarchLIB was started in times when there was no good JS minifier. Therefore I was quite religious about trying to write short code manually, and as such DL contains a lot of syntactic hacks[1] such as “foo == bar ? a = 10 : b = 20”, though the more readable version would clearly be to use “if/else”.

Since the parser/compressor runs fine on DL and jQuery, I’m quite confident that it’s solid enough for production use. If you can identify any bugs, I’d love to hear about them (use the Google Group or email me directly).

[1] I even reported a few bugs and suggested some fixes in the original parse-js library, and Marijn pushed fixes literally in minutes.

Links

• Twitter: @UglifyJS
• Project at GitHub: http://github.com/mishoo/UglifyJS
• Google Group: http://groups.google.com/group/uglifyjs
• Common Lisp JS parser: http://marijn.haverbeke.nl/parse-js/
• JS-to-Lisp compiler: http://github.com/marijnh/js
• Common Lisp JS uglifier: http://github.com/mishoo/cl-uglify-js

License

UglifyJS is released under the BSD license:

Copyright 2010 (c) Mihai Bazon <mihai.bazon@gmail.com>
Based on parse-js (http://marijn.haverbeke.nl/parse-js/).

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:

    * Redistributions of source code must retain the above
      copyright notice, this list of conditions and the following
      disclaimer.

    * Redistributions in binary form must reproduce the above
      copyright notice, this list of conditions and the following
      disclaimer in the documentation and/or other materials
      provided with the distribution.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER “AS IS” AND ANY
EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY,
OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF
THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
SUCH DAMAGE.