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.
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 orwith{}
statements. In short, ifeval()
orwith{}
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
orcontinue
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.
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.
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.
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).
## 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)
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 callast_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
— whenast_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 areturn
,throw
,break
orcontinue
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 infilename
. 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 valuetrue
, or you can specify a numeric value (such as1024
), a quoted string value (such as =”object”= or =’https://github.com’), or the name of another symbol or keyword (such as =null
ordocument
). 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 NodeJSrequire()
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 namesrequire
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.
- put all
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 tofalse
. If you passtrue
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 movevar
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 thevar
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 thatgen_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) operatordead_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 theoptions
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
– passtrue
if you want indented outputindent_start: 0
(only applies whenbeautify
istrue
) – initial indentation in spacesindent_level: 4
(only applies whenbeautify
istrue
) – indentation level, in spaces (pass an even number)quote_keys: false
– if you passtrue
it will quote all keys in literal objectsspace_colon: false
(only applies whenbeautify
istrue
) – wether to put a space before the colon in object literalsascii_only: false
– passtrue
if you want to encode non-ASCII characters as\uXXXX
.inline_script: false
– passtrue
to escape occurrences of</script
in strings
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.
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.
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 |
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.
- 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
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.