Fixes#15792.
Previously, budgets would include content for both versions of a differential build. Thus the `initial` budget would count content from the ES5 **and** ES2015 bundles together. This is a very misleading statistic because no user would download both versions. I've updated the budget calculators to take this into account and generate size values for both builds which are then checked independently of each other.
The only calculators I changed are the `InitialCalculator` (for computing `initial` bundle sizes) and `BundleCalculator` (for computing named bundles). Since budgets are handled by Webpack for builds without differential loading, the `initial` bundle will always have those two sizes. The `BundleCalculator` might reference a bundle which does not have differential loading performed (such as a CSS file), so it emits sizes depending on whether or not multiple builds were found for that chunk.
Most of the other calculators don't really need to take differential loading into account. `AnyScriptCalculator` and `AnyCalculator` already apply on a file-by-file basis, so they generate sizes for both build versions already. `AnyComponentStyleCalculator` only applies to CSS which does not have differential builds.
The wierd ones here are `AllCalculator` and `AllScriptCalculator` which reference files with and without differential builds. Conceptually, they should be separated, as a "total" budget specified by an app developer probably wanted it to mean "the total resources a user would have to download", which would only be one differential build at a time. However, I don't see a good way of identifying which assets belong to which differential build. Even if an asset belongs to a chunk with differential builds, we don't know which build takes which assets into account. I decided to leave this for the time being, but it is probably something we should look into separately.
Since budgets take differential loading into account, users might reasonably want to set different budgets for different builds (ie. "initial-es2015 bundle should be capped at 100k, but initial-es5 bundle can go to 150k"). That's more of a feature request, so I also left that out for a future PR.
Refs #15792.
Static files listed in `angular.json` were being accounted in the `initial` bundle budget even when they were deferred asynchronously with `"lazy": true` or `"inject": false`. Webpack belives these files to be `initial`, so this commit corrects that by finding all extra entry points and excluding ones which are explicitly marked by the application developer as asynchronous.
One edge case would be that the main bundle might transitively depend on one of these static files, and thus pull it into the `initial` bundle. However, this is not possible because the files are not present until the end of the build and cannot be depended upon by a Webpack build step. Thus all files listed by the application developer can be safely assumed to truly be loaded asynchronously.
Refs #15792.
This provides access to all the size information necessary because all build steps have already completed. This commit is roughly a no-op because it simply moves the budget checks (for different builds) to be executed post-build.
The lone exception is the AnyComponentStyle budget. Component stylesheet files are not emitted after the build is completed, so there is no size information to work with. Instead, these budgets are checked during a separate plugin (exected for different builds **and** non-differential builds).
This is needed as otherwise under Bazel we hit `Cannot find module '@babel/preset-env` we use `require.resolve` since this is always patched under Bazel.
This is needed as otherwise under Bazel it will fail with `Angular Compiler was detected but it was an instance of the wrong class.` when not using the linker since node_modules will be linked to `external/npm`. Under Bazel `preserveSymlinks` is to be set to true, however with the current behaviour this doesn't effect resolveLoader.
Refs #16193
This detects and filters out error messages about duplicate 3rdpartylicenses.txt. This is a short-term fix intended to get us to 9.0.x release while a follow up effort will work to properly fix this bug. Left a TODO to remove this filter once a fix is ready.
Certain node packages depend on native bindings which cannot be bundled, with this fix we add a way to exclude these from being processed by the bundler.
Closes: #16348
All locale i18n options now support an object form which allows a base HREF to be defined for the locale. Each locale can now optionally define a custom base HREF that will be combined with the base HREF defined for the build configuration. By default if the shorthand form for the locale is used or the field is not present in the longhand form, the locale code will be used as the base HREF. To disable automatic augmentation a base HREF value of an empty string (`""`) can be used. This will prevent anything from being added to the existing base HREF.
For common scenarios, the shorthand form will result in the preferred and recommended outcome of each built locale variant of the application containing a defined base HREF containing the locale code.
Allow version "0.0.0" for integration testing in the angular/angular repository with the generated development @angular/core npm package which is versioned "0.0.0".
This is a pre-req for https://github.com/angular/angular/pull/33927 which runs integration tests against the bazel generated npm packages.
A target of es2015 was previously assumed when using differential loading. This could result in erroneously downleveling an es2016+ output file instead of generating a new es5 output file.
The ES5 polyfills file was erroneously being suffixed with `es2015`. The webpack configuration does not support conditional customization per chunk for the output filenames (`chunkFilename` option schema only supports string values). This change adds an additional small webpack plugin that allows the chunk filenames to be adjusted based on the chunk name. The plugin is only added when differential loading is enabled as this is the only time that a chunk currently requires its filename to be adjusted.
Closes#15915
The node modules directory contains a massive set of directories and files. When watching via polling, that set needs to be queried repeatedly to determine if any files have changed. Changes within node modules are quite rare while using `ng serve` or `ng build --watch`. As a result, polling the node modules directory is rarely useful. This change causes CPU usage to drop from a potential high of ~80% to a more manageable ~5-10%.
TSLint [6.0.0-beta0](https://github.com/palantir/tslint/releases/tag/6.0.0-beta0) was released, and `ng lint` throws with:
TSLint must be version 5.5 or higher.
if we try to use it in a CLI project.
It looks like the current version check allows v5.5+ by checking that both the major and minor versions are > 5. So this fails with 6.0 (but would succeed with 6.5).
This fixes the check to allow using v6.0.
This should be reintroduced in 9.1 or 9.2. The other remaining "hidden" peer dependencies should be included as well (currently `karma` and `node-sass`).
At the moment we are relying on source content to determine if the compilation is under Ivy or VE. However, we do know what compilation we are in from the parsed tsconfig.
With this change we use the `enableIvy` to set the analytics metric
