* Rename bundles (prep for webpack optimisation)
This might or might not reduce the overall repo size,
because git can combine similar chunks in the newly generated files
* Optimise webpack build time
Consolidates bundles, since a separation of bundle-framework.js vs. bundle-legacy.js
vs. bundle-lib.js no longer makes sense - they're all loaded upfront anyway,
since we'll be introducing more react-powered logic alongside the "legacy" JavaScript.
By consolidating into fewer bundles, we give the optimisation scripts (UglifyJS)
more options to reduce the overall file size.
The main motivation for a vendor.js is to shorten rebuild times:
Most active development is happening in files required through bundle.js.
This commit drastically reduces the rebuild time for those changes (15s to 4s).
The 'admin' module will be split off from 'framework',
where 'framework' only provides (mostly) frontend-agnostic PHP classes.
For example, HTMLEditorField.php has a TinyMCEConfig.php driver,
but doesn't come with its own JS includes.
Multiple entry points can't result in a single bundle.css with a fixed filename, see
https://github.com/webpack/extract-text-webpack-plugin/issues/179
Until that's resolved, it's easier to keep the 'css' task separate in Webpack,
and have a single entry point for all CSS (bundle.scss).
Also partially reverting "Moved frontend assets into admin/ "module"",
which moved too many files: debug.css and install.css need to remain
as framework (not admin) deps. Split out into a separate `framework-css` Webpack
task in preparation for splitting off the module.
The JavaScript i18n functionality in SilverStripe is used in the CMS as well as form field implementations.
Form fields used to include their own JavaScript for usage outside of CMS. This now requires custom build tooling in a project.
Hence there's no need for an i18n shim (i18nx.js), since the CMS always uses i18n support.
We've removed the ability to directly reference JS and CSS files
for form fields and other SilverStripe features in favour of a common bundle built by Webpack.
The logical next step is to make the framework module free of frontend dependencies,
which should simplify its operation, and avoid another time intensive "npm install" on a module.
There's not a lot of benefit in packaging these separately in terms of initial CMS load size,
so let's simplify the setup. They'll eventually become lazy loaded chunks in a React-based setup
When adding the deps straight into the file (recommended),
the onchange handler in file-upload isn't firing properly when a file is uploaded through
the <input type=file> button - it falls back to default behaviour, which submits the
containing form, failing because the upload is handled by a different URL.
These were originally copied from node_modules via an "npm run thirdparty" task,
in order to have them loadable with oldschool <script> tags.
Since webpack supports CommonJS-style loading, that's no longer required,
we can simply inline those scripts into the bundle.
We need to use imports-loader though, in order to ensure
that "define" is not available in some module scopes,
which triggers AMD behavior that's not compatible with Webpack's loaders.
See http://webpack.github.io/docs/shimming-modules.html
I've had to pin to the exact versions used in the 3.x CMS,
since jquery-upload has introduced an AMD wrapper sometime
between 6.0 and 6.9 (the latest version NPM automatically pulls in).
This AMD wrapper confuses Webpack, since it's trying to resolve the
dependencies contained in it. We could create shims for those instead,
but the easiest way was to simply revert to the versions already used
before the Webpack migration (since the newer versions in node_modules
were never actually copied into thirdparty, they weren't used before).
The bundle is generated by running “webpack” directly - gulp is no
longer needed as an intermediary. The resulting config is a lot shorter,
although more configuration is pushed into lib.js.
Modules are shared between javascript files as global variables.
Although this global state pollution is a bit messy, I don’t think it’s
practically any worse than the previous state, and it highlights the
heavy coupling between the different packages we have in place.
Reducing the width of the coupling between the core javascript and
add-on modules would probably be a better way of dealing with this than
replacing global variables with some other kind of global state.
The web pack execution seems roughly twice as fast - if I clear out my
framework/client/dist/js folder, it takes 13.3s to rebuild. However,
it’s not rebuilding other files inside dist, only the bundle files.
CSS files are now included from javascript and incorporated into
bundle.css by the webpack. Although the style-loader is helpful in some
dev workflows (it allows live reload), it introduces a flash of
unstyled content which makes it inappropriate for production.
Instead ExtractTextPlugin is used to write all the aggregated CSS
into a single bundle.css file. A style-loader-based configuration could
be introduced for dev environments, if we make use of the webpack live
reloader in the future.
Note that the following features have been removed as they don't appear to be
necessary when using Webpack:
- UMD module generation
- thirdparty dist file copying
LeftAndMain.js deps: Without it, ssui.core.js gets loaded too late,
which leads e.g. to buttons being initialised without this added behaviour.
