-
Notifications
You must be signed in to change notification settings - Fork 1.1k
Chunk output for dynamic loading
NOTE: This page started life as a copy of a stackoverflow article by ChadKillingsworth.
Regardless of input type, the compiler will normalize all files and bundle them together. By default, a single output file is produced. However, this may not work for you if your application is really big. In that case, you may want to have the compiler break your code up into multiple chunks that can be loaded separately. You will design your application so that the code you always need gets loaded in an initial chunk, probably from a <script>
tag, then that chunk will load others (which may load still others) as needed in order to support the user's actions. (e.g. The user may request a new display view, which requires you to load the code for showing that view.)
There are two different types of output chunks available specified by the --chunk_output_type
flag.
This is the default option and the compiler will produce standard scripts. This mode is normally paired with the --chunk_wrapper
flag for script isolation and the --rename_prefix_namespace
flag so that symbols can be referenced across chunks.
For this type, you must fully specify your own script loader.
The compiler will output es modules and cross chunk references will utilize the import
and export
statements. Since modules have built in isolation and modern browsers know how to load them, this option is by far the easiest.
Manually building the flags for code splitting is only feasible in very simple applications. In real world applications, the chunk graph quickly becomes very complex. The closure-calculate-chunks utility will both discover the correct set of files for the compiler and automatically split your code into chunks. It will output the correct --js
and --chunk
flags for you to provide to Closure Compiler.
The closure-calculate-chunks utility splits code into chunks where ever it encounters a dynamic import expression (ex import('./my-module.js')
) where the import specifier is a string literal.
There are multiple compiler flags pertaining to chunks.
Each use of the --chunk flag describes an output file and it's dependencies. Each chunk flag follows the following syntax:
--js inputfile.js
--chunk name:num_files:dependency
The resulting output file will be name.js and includes the files specified by the preceding --js flag(s).
The dependency option is what you will be most interested in. It specifies what the parent chunk is. The chunk options must describe a valid dependency tree (you must have a base chunk).
Here's an example:
--js commonfunctions.js
--chunk common:1
--js page1functions.js
--js page1events.js
--chunk page1:2:common
--js page2function.js
--chunk page2:1:common
--js page1addons.js
--chunk page1addons:1:page1
In this case, you are telling the compiler that the page1 and page2 chunks depend on the common chunk and that the page1addons chunk depends on the page1 chunk.
Keep in mind that the compiler can and does move code from one chunk into other chunk output files if it determines that it is only used by that chunk.
None of this requires closure-library or the use of goog.require/provide calls nor does it add any code to your output. If you want the compiler to determine dependencies automatically or to be able to manage those dependencies for you, you'll need to use a module format such as CommonJS, ES2015 modules or goog.require/provide/module calls.
If the compiler flag of --chunk_output_type=ES_MODULES
is used, chunks will be proper ES modules. In this case cross chunk references will use import
and export
statements. These chunks are intended to be referenced in a browser using <script type="module">
. Since modern browsers can natively recognize and handle ES modules, no script/module loader is provided.
If you utilize dynamic import expressions (ex import('./my-chunk.js')
) but also use an output language level lower than ECMASCRIPT_2020, you must specify the --dynamic_import_alias=import_
flag and provide a polyfill for the compiler. You can utilize the fact that browsers shipped dynamic imports well before full ECMASCRIPT_2020 support and provide a wrapping function instead of a polyfill.
<script>function import_(specifier) { return import(specifier); }</script>
By default, the compiler will produce a script type output (not a module) regardless of the types of input files provided.
The chunks are intended to be loaded as scripts, not JS modules. If you load them as JS modules, you may have problems with some global variables not being available when you expect them to be. (See e.g. https://github.com/google/closure-compiler/issues/3752).
Although you can use whatever mechanism you like to trigger loading chunks.
Closure library provides a goog.module.ModuleManager
class and related classes for this purpose.
NOTE: For a long time we used the term "module" to refer to these separately-loadable chunks, but that was too easily confused with goog.module()
and ES modules which came along later. The Closure library namespace and class names reflect this older naming.
Chunks are expected to be global. If you use the --chunk_wrapper
flag to isolate the code, cross chunk references will break. The compiler has the --rename_prefix_namespace=NS
flag to address this. Symbols which are referenced in more than one chunk will be converted to properties of the specified namespace.
Example chunk output wrapper for use with a namespace:
(function(NS){%output%}).call(this, window.NS = window.NS || {})