This seems misleading. Rollup doesn't ship with an uglifier/minifier, and Webpack ships with UglifyJS. So what you're really comparing is UglifyJS vsthe Closure Compiler, however both webpack _and_ rollup have plugins for utilizing the closure compiler:
I remember back in the day, Google Closure was so advanced they even tried out shuffling code around in the final JS file and measured which way produced the best GZIP compression. They since abandoned that code though, but that's something they did over 5 years ago.
I think it works better for GWT than for Closure because GWT's optimizer permits inlining functions in some circumstances even if they increase code size. This causes repetitious fragments to appear, and re-ordering the code to bring those fragments within GZIP's compression window functions as kind of a late-stage "anti-inliner/common subexpression elimination"
They do (or did) something similar with the back end of GWT (globally sorting method definitions by length). But it's probably easier with GWT output (vs arbitrary js), because it has a well defined structure.
Last time playing with closure, it did quite a good job of dead code elimination.
Is there a specific reason why simple code elimination is not as wide spread as tree shaking, although it seems to be simpler and more effective solution?
Closure fell out of vogue because it's fairly heavy and seriously subsets javascript, which makes it convenient when you're compiling to javascript (from a very different language, not just augmenting/macroing it), but not when you're writing it by hand.
Something interesting is that I believe Google is working on a compiler that converts TypeScript to Closure, which would let you take advantage of the closure compiler while utilizing TS. That would be a very exciting development.
The Javascript version of the Closure-Compiler is
transpiled by GWT from the Java source. For more details
on the differences in behavior see the super sourced
files in the main repo.
That's pretty interesting. Do we have any comparisons on the performance of the outputted code between each? Like do they all perform identically? Are there optimizations one may do that others do not or perhaps even the opposite like a shortcut that cuts down on code but makes it perform worse? Also what about memory performance?
It's been a while since I've used closure. I may look into this to do a comparison on my current open source project and see what difference, if any, it makes.
Unfortunately, recent versions of Closure Compiler don't appear to produce entirely reliable output when run with the advanced optimisations that make it interesting and give the kinds of impressive results observed by the OP here.
Case in point: We just installed the latest version on a new computer yesterday, and it causes some JS code that runs fine in its original form to fail on IE11 after transformation. The original code isn't complicated at all and multiple developers and lint tools think it's correct and not doing anything non-standard, but the transformed version seems to confuse IE. In fairness to Closure Compiler, we also can't immediately see why the transformation it has applied would be invalid, so maybe this is really a bug in IE's JS parser or something along those lines. Unfortunately, the fact remains that the original code works, the slightly different transformed version from a previous version of Closure Compiler also worked, but the transformed code in the new version does not.
Intuitively I would assume that for most codebases dead-code removal would have the biggest impact, especially when importing third-party libraries and only using portions of them.
But from your comment I can see that rollup should be able to handle this too. Yet the data in the linked article seems to imply Google Closure compiler still can optimize the code 33% further.
I find it hard to imagine that inlining functions can make that kind of difference.
Is it the test which contains flawed data? Or is this just an area where it's hard to generalize results due to the varying characteristics of different code-bases? Anyone got an opinion?
I compared the output bundles (ran them back through a beautifier for sanity) and it looks like a lot of the benefit comes from the aggressive renaming.
Compare this section of the rollup output:
return e || (e = new zh({
enableLongStackTrace: dt()
})), e.run(function() {
var r = yh.resolveAndCreate([{
provide: zh,
useValue: e
}], n.injector),
o = t.create(r),
l = o.injector.get(th, null);
if (!l) throw new Error("No ErrorHandler. Is platform module (BrowserModule) included?");
return o.onDestroy(function() {
return _t(n._modules, o)
}), e.onError.subscribe({
next: function(t) {
l.handleError(t)
}
}), bt(l, function() {
return o.injector.get(bh).donePromise.then(function() {
return n._moduleDoBootstrap(o), o
})
})
})
To this from Closure compiler:
var c;
c || (c = new Af({
Lm: yh()
}));
return c.run(function() {
var d = Ne([{
sb: Af,
Be: c
}], a.s),
e = b.create(d),
f = e.s.get(ee, null);
if (!f) throw Error("No ErrorHandler. Is platform module (BrowserModule) included?");
e.ec(function() {
return Hh(a.mi, e)
});
c.hq.subscribe({
next: function(a) {
f.handleError(a)
}
});
return Ch(f, function() {
return e.s.get(Ze).xp.then(function() {
Eh(a, e);
return e
})
})
})
On top of that I did a basic comparison of the number of times the 'function' keyword is present in each bundle:
$ ag --count function eg-closure.js
eg-closure.js:2905
$ ag --count function eg-rollup.js
eg-rollup.js:4421
Whether this is from inlining or 'better' dead-code removal I can't really tell.
There are too many optimizations in Closure Compiler to describe simply what it does. It is a full blown optimizing compiler like C, right down to supporting graph coloring in the naming of locals. It even does type-based optimization by disambiguating property names on prototypes by types (https://github.com/google/closure-compiler/wiki/Type-Based-P...) -- this alone I've measured at 14% code size reduction in my projects, and it can move code at the function level out of the main module into late loaded modules if it detects the code is not used until later.
All of Google's main projects, from the Google Home page, to Docs, Gmail, Maps, etc rely on Closure Compiler. Google Photos makes particularly aggressive use of this splitting functionality (take a look at the network tab in Chrome)
Closure Compiler also pairs well with Google Style Sheets, which is a style sheet compiler like SASS/LESS (but predated them), but it optimizes and prunes CSS.
Closure has it's downsides, the Advanced Mode makes assumptions about the way you write Javascript that can break. However, there are conformance checks you can enable to check for many of these. The compiler has to "see" what you're doing, so if you do something like this:
foo.hello = 42;
function blah() { return "hello"; }
foo[blah()] = 42;
The compiler may not be able to see that 'foo.hello' and 'foo["hello"]' are the same thing, and the former will get renamed to foo.xyz. The price you pay huge savings in code size is being consistent and diligent in your code, and sometimes giving up a little bit of dynamism
Closure compiler understands how the code runs, whereas those other tools only understand how its parsed.
Closure compiler is really awesome but never fully caught on outside of Google because it doesn't let you use some features of JS that aren't optimizable. The main one I can remember is only allowing dot syntax for properties
https://github.com/roman01la/webpack-closure-compiler
https://github.com/camelaissani/rollup-plugin-closure-compil...