Consolidate build process with GCC #11483
Conversation
gaearon
changed the title
gaearon
force-pushed
the
fork-gcc-2
branch
3 times, most recently
from
f3b06ea
to
59287b6
Compare
gaearon
changed the title
| const shims = Modules.getShims(bundleType, entry, featureFlags); | ||
| const plugins = [ | ||
| const isProduction = isProductionBundleType(bundleType); | ||
| const isInGlobalScope = bundleType === UMD_DEV || bundleType === UMD_PROD; |
There was a problem hiding this comment.
I’m not sure this is the best name. Maybe isUMDBundle?
There was a problem hiding this comment.
There may be more bundle types where top-level things are declared in global scope. In theory. :-)
| ); | ||
| return plugins; | ||
| }), | ||
| ].filter(Boolean); |
There was a problem hiding this comment.
I’m not sure this is super obvious as to what’s happening?
There was a problem hiding this comment.
I dunno, seems like a common pattern with plugin systems?
Happy to change to something explicit if you prefer.
I only see a ~20 byte difference. If that's what you're referring to, yes, it's expected because we no longer try to compress the license header, and it's consistent for DEV and PROD bundles. |
|
Added a second commit that doesn't change the build output, but breaks out the bundle header/footer logic into a separate file, and restructures it to be more explicit. This makes |
It is easier to understand if we just explicitly type them out.
* Consolidate build process with GCC * Record sizes * Refactor header and footer wrapping It is easier to understand if we just explicitly type them out.
This completely gets rid of Uglify in our build process. GCC provides better DCE for the FB bundles (e.g. the persistence path is now dead code eliminated in www, and mutation path is eliminated in CS). And this unifies our build process with how we already build UMD/CJS.
I forked GCC to add google/closure-compiler#2707. Hopefully this gets merged but if not, it’s a pretty simple change to maintain. Then I recompiled the
google-closure-compiler-jspackage and pointed Yarn lockfile to my fork (easier than forking another dependency layer up from there). I verified with Yarn folks it’s a supported (although a bit unobvious) way to point dependency to a fork (and this is pretty much what custom registry support relies on).GCC doesn't preserve whitespace so I added Prettier after it for FB/RN bundles.
I also took an opportunity to clean up how we apply Rollup plugins in general, and set up a single chain instead of switches. This should make bundles more uniform across the build targets.
Some DEV sizes decreased because I started dropping per-file license headers. FB PROD sizes decreased because of better DCE and less whitespace.
RN DEV sizes increased because I did not add Closure to them whereas in the past they did use Uglify (for DEV). I'm not sure why it was desirable and it seemed better to make them un-Closure'd just like FB_DEV builds are. But I can change that if you want to. Mainly I just want to keep DEV bundles with comments, and Closure can't preserve them.
And that's about it! I hope to test and maybe land this on Thursday after shipping 16.1.