Google Closure ? I'm Not Impressed

We all know that Google is synonym of performances, simplicity, and again performances. A search engine that uses a truncated body for a not valid W3 markup should be the most bytes and performances maniac in the current web era, isn't it?
Well, Google Closure Tools has been a negative surprise, at least this is what I can tell about it after a first quick review.

Closure Compiler

It's since ages I am wondering what kind of tool Big G is using to produce their scripts and finally we got the answer: Closure Compiler ... ooooh yeah!
Packer, YUIC, it does not matter, when Google needs something, it creates something. This is almost intrinsic, as developers, in our DNA: we spot some interesting concept? We rewrite it from the scratch pretending we are doing it better!
This is not the case, or better, something could go terribly wrong!

// ==ClosureCompiler==
// @compilation_level ADVANCED_OPTIMIZATIONS
// @output_file_name default.js
// ==/ClosureCompiler==

(function(){
   "use strict";
   this.myLib = this.myLib || {};
}).call(this);

myLib.test = 123;

The produced output:

(function(){this.a=this.a||{}}).call(this);myLib.test=123;

And 2 warnings:

JSC_USELESS_CODE: Suspicious code. Is there a missing '+' on the previous line? at line 2 character 4
"use strict";
^
JSC_USED_GLOBAL_THIS: dangerous use of the global this object at line 3 character 4
this.myLib = this.myLib || {};
^

Excuse Me?
The sophisticated compiler is able to understand the "use strict" ES5 activation statement and the fact we are passing the global object as this reference in the closure. It does not matter, as showed in the produced code the result will be a broken library, thanks to its missed name, magically transformed into "a".
Advanced Optimization Could Fail to both give us right suggestions and fix or optimize the code, since in that case, as example, this.a won't perform anyhow faster than original this.myLib.
Advanced Optimization parameter could also be dangerous for lazy loaded libraries.
We need to be extremely careful with this option and, as result, rather than a Compiler, we will deal with a "code messer" where hard debug will become automatically the hardest ever.
Read Carefully This Page if you are planning to use this option because under the flag "best ratio" and "removed dead code" we could have massive surprises in the middle of the application.

As summary, SIMPLE_OPTIMIZATIONS as compilation_level directive is so far the recommended one, but at the same time it won't offer that different ratio compared against YUI Compressor or Dean's Packer (NO base62) produced outputs while ADVANCED_OPTIMIZATIONS could be tested for single stand alone files hoping these won't break the global namespace via renamed variables.
In this case a JavaScript closure, the real one, is an absolute must!

Closure Library

This is another part, not strictly related with the Compiler, but apparently able to work with it. The Closure Library is a wide namespace loads of core features and a cross browser User Interface. I have to admit this library is a massive piece of work, but techniques used to make it happen are often hilarious.
First of all, this is the first time I read protected variables called with an underscore at the end, rather than as first char:

// normal code
function Constructor(){
    this._protected = [];
};

// Closure Library Creativity
function Constructor(){
    this.protected_ = [];
};

Why On Earth? Python style a part, where the underscore has a concrete meaning, the technique to use underscore as first character has a reason to exists.

function copyOnlyPublic(o){
    var $o = {}, k;
    for(k in o){
        if(k.charAt() !== "_")
            $o[k] = o[k]
        ;
    };
    return $o;
};


var myC = new Constructor;
var $myC = copyOnlyPublic(myC);

No Way! To make the style "creative" the charAt method with optional 0 as argument needs to become:

if(k.charAt(k.length - 1) !== "_")

Is this what we would expect from the performances king? I don't think so.
Is this faster to read at least for human eyes? Neither!
Gotchas are everywhere in the library ... the most redundant stuff I've ever seen is the array namespace!

goog.array.indexOf = function(arr, obj, opt_fromIndex) {
  if (arr.indexOf) {
    return arr.indexOf(obj, opt_fromIndex);
  }
  if (Array.indexOf) {
    return Array.indexOf(arr, obj, opt_fromIndex);
  }

  var fromIndex = opt_fromIndex == null ?
      0 : (opt_fromIndex < 0 ?
           Math.max(0, arr.length + opt_fromIndex) : opt_fromIndex);
  for (var i = fromIndex; i < arr.length; i++) {
    if (i in arr && arr[i] === obj)
      return i;
  }
  return -1;
};

OMG, I can't believe performances matter only for a missed body tag in the layout ... it cannot be real, can it?
What we have there? 1 to 2 possibly missed checks for each call (IE) and everything just to emulate the native Array.indexOf which is present in basically every browser except truly old or Internet Explorer?

goog.array.indexOf = Array.indexOf || function(arr, obj, opt_fromIndex) {
  var fromIndex = opt_fromIndex == null ?
      0 : (opt_fromIndex < 0 ?
           Math.max(0, arr.length + opt_fromIndex) : opt_fromIndex);
  for (var i = fromIndex; i < arr.length; i++) {
    if (i in arr && arr[i] === obj)
      return i;
  }
  return -1;
};

Array.indexOf is used as fallback if for some unknown reason (and browser...) an Array has not indexOf but Array.indexOf is present ... well, if we can trust that case is there any valid reason to create a performance gap like that for almost every method?
forEach, lastIndexOf, every JavaScript 1.6 or greater emulated method contains redundant checks performed for each call ... where is the performance maniac here?
The feeling is that this library has been created by some Java guy, probably extremely skilled with Java, but definitively not that clever with JavaScript programming style. Google if you need skilled JS developers there are hundreds of us out there. What I mean is that it does not matter if a minifier is able to remove dead code because dead code should not be there at all, isn't it?

// WebReflection Suggestion
goog.array.clone = (function(slice){
try{slice.call(document.childNodes)}catch(e){
    slice = function(){
        var rv = [];
        if(this instanceof Object)
            // suitable for arguments
            // and every other ArrayLike instance
            return rv.slice.call(this)
        ;
        for (var i = 0, len = this.length; i < len; ++i)
            rv[i] = this[i]
        ;
        return rv;
    };
};
return function(arr){
    return slice.call(arr);
};
})(Array.prototype.slice);

Above suggested snippet is based on Features Detection, a technique apparently completely discarded in those files I've read in this library.

Features Detection Cons

• performed runtime, few milliseconds before the library or function is ready to use

Features Detection Pros

• performed once, and never again for the entire session
• best performances, browser independent and browser focused at the same time
• being usually based over most recent standards, features detections could cost a bit more only for deprecated, obsolete, or truly old browsers

Moreover!

If the problem is the wasted millisecond to perform a feature detection, we can always fallback into lazy feature detection.
I agree that web performances are more about download time and round trip, but if Google has V8 as engine monster, do we agree that better JavaScript practices could make even V8 faster?

Lack Of Creativity

Even if this library uses some weird practice, most logical and common techniques to speed up execution and reduce code are often not considered. As example, this is just one method we can find in the crypt file:

goog.crypt.byteArrayToString = function(array) {
  var output = [];
  for (var i = 0; i < array.length; i++) {
    output[i] = String.fromCharCode(array[i]);
  }
  return output.join('');
};

Oh Really? So now a JavaScript developer should create a function which accept an array in order to create another array to populate over a loop and a global method call to perform a join at the end? That's weird, I thought that method could have been written in this way:

// WebReflection Suggestion
goog.crypt.byteArrayToString = function(array) {
  return String.fromCharCode.apply(null, array);
};

... but maybe it's just me that noob that I cannot spot the difference except better performances over less code ... please enlighten me!

Closure Library, The Good Part

Is massive, for basically each file there is a test case so it is robust. For sure as we have seen before, this is not the fastest general purpose library we could find in the net, something I was expecting from Big G, but hey ... we cannot blame the excellent work done so far, can we?

Closure Templates

Well, here I am almost without a word ... I mean, the number 1 search engine able to create monster pages via JavaScript templates? I do hope those files will be performed everywhere but a web page and runtime, 'cause with incoming HTML5 I feel horrified thinking about a new web loads of document.write and nothing else. Where is the semantic? Where is the logic applied before and not during execution? Where are best practices? Looking at examples even V8CGI project is not able to perform that code ... what is that, exactly? And WHY???
The only thing I do like there is the definition file, something simple, clever, easy to parse, exactly the opposite of its implementation via JavaScript: please avoid it!

Closure Conclusion

No this is not another piece of the puzzle, just the end of this post. I've been probably too aggressive and it's only thanks to Google decision that I can write a post like this: Open Source is the key, or your libraries, operating systems, whatever, will be always behind 'cause developers able to help you are not only in your team.
I was expecting excellent ideas, new killer techniques unknown for everybody else, what I have found is "yet another toolkit". I am always up to contribute so if interested put my name in the committers list, I have already lot of stuff to optimize because a Compiler, whatever it does, cannot create a better code, it can simply try to make a bit better and shorter the existent one: never delegate skills to a machine until these will be able to take decisions for us!