Speed up min builds

[zpao]

We're minifiying twice thanks to uglifyify. We need this to remove dead requires. But it also doesn't minify all the way, so we still need uglify as well. I haven't actually looked to see if there's a better way.

[sophiebits]

I believe webpack is smart enough to strip out unneeded requires even before uglification so that might be faster than what we're doing…

[syranide]

@spicyj Is it? I haven't actually checked, but I use require('es5-shim') to apply the shim and it's working in IE8 so I really doubt it's been stripped (unless it's really smart and knows when a module has side-effects).

[sophiebits]

I meant if you do if (false) { require('dev-only-stuff'); } then webpack won't include that module.

[zertosh]

@zpao, I had a similar problem with "dead requires" so I put together unreachable-branch - it's a browserify transform (jstransform visitor under-the-hood) to comment out unreachable branches in if's, ? ternaries and || && logical operations.

[zpao]

Some context... I did end up trying unreachable-branch and it mostly works (except it chokes on if (__DEV__) { /* comment */ code }, which we have at least 1 instance of). Even fixing that I didn't end up with any significant perf wins (and I even combined the multiple jstransforms we have happening at the browserify step).

[tgriesser]

@spicyj is right that webpack is not only smart enough to strip out dead requires, but it strips out dead code paths entirely, so in your build (pre-minified code) instead of:

("production" !== "development" ? invariant(
    children.nodeType !== 1,
    'traverseAllChildren(...): Encountered an invalid child; DOM ' +
    'elements are not valid children of React components.'
) : invariant(children.nodeType !== 1));

you just get the much nicer:

invariant(children.nodeType !== 1,
  'traverseAllChildren(...): Encountered an invalid child; DOM ' +
  'elements are not valid children of React components.'
);

[zertosh]

wow I totally missed your response @zpao. I just pushed an updated version of unreachable-branch-transform that uses recast instead of jstransform, that doesn't have that problem. I did a quick time benchmark between this vs an extra uglify pass, that the time difference was negligible.

Though, I was able to shave off 2 sec by turning off mangle and a few compress options on the first pass of uglifyify - since it's only the dead code removal that matters. For the 2 sec improvement (prob less on faster machines), it wasn't worth PRing the changes to the grunt task that allow passing options to transforms 😞

The real killer is derequire.

[zpao]

I think at this point things are pretty fast. We'll make some more changes to the build process in general.