Summary: @public
Fix the haste resolution algorithm. Changed the map data structure from a list of modules, to a list of modules grouped by platform.
This considerably simplifies the "getModule" method and makes it easy to properly warn about colliding module names.
This also fixes a bug where we used to include `.web` files when we shouldn't (see task).
Reviewed By: @vjeux
Differential Revision: D2482969
Summary: @public
The issue of colliding haste modules have came up many times and have wasted countless engineering hours. This will start warning about it and will also start selecting modules at random so that people don't depend on undefined behavior.
Additionally, this surfaced an issue where with assets we may fatally throw if the directory doesn't exist. This is fixed by checking the existence of the directory before trying to match files in it.
Reviewed By: @jingc
Differential Revision: D2478480
Summary: @public
I've noticed that the logs can be sometimes misleading, as when an Activity ends it doesn't log immediatly. A sync `console.log` would log before it although the Acitivity should've finished before.
Turns out we wait before writing out the logs to the console. I don't see any reason for this. Looking at the `Activity` module it's over-engineered. This diff makes logging sync and simplfies the module.
Reviewed By: @martinbigio
Differential Revision: D2467922
Summary:
We don't currently support platform extensions in asset modules.
This adds supports for it:
```
require('./a.png');
```
Will require 'a.ios.png' if it exists and 'a.png' if it doesn't.
Summary:
There are two fs steps and it wasn't clear why. This now puts the right label:
```
[9:38:25 PM] <START> Building in-memory fs for JavaScript
[9:38:27 PM] <END> Building in-memory fs for JavaScript (2030ms)
[9:38:27 PM] <START> Building in-memory fs for Assets
[9:38:27 PM] <END> Building in-memory fs for Assets (615ms)
```
Summary:
D2319999 introduced a regression where we stopped waiting for the "build haste map" step to finish before we accept any requests. This makes sure that we block on that.
Need to unbreak with this, but will follow up with a test to catch this in the future.
Summary:
Currently the platform selection is controlled by the blacklist. However, since we want to use the same server instance for cross-platform development, we need this to be controlled per request.
One outstanding issue, is that the DependencyGraph class wasn't designed that way and it doesn't have a per-request state. This means that with the current design race conditions is possible. If we got a request for a different platfrom while processing the previous request, we may change the outcome of the previous request.
To fix this a larger refactor is needed. I'll follow up a diff to do that.
Finally, so I don't break the universe like last time, I'll leave it up to the RN guys to update the call sites.
Summary:
Instead of using plain objects and having to convert to and from them we just use the `Module` class across the codebase.
This seems cleaner and can enforce the type as opposed to fuzzy objects.
Summary:
Introduce a Bundler capable of generating the layout of modules for a given entry point. The current algorithm is the most trivial we could come up with: (1)it puts all the sync dependencies into the same bundle and (2) each group of async dependencies with all their dependencies into a separate bundle. For async dependencies we do this recursivelly, meaning that async dependencies could have async dependencies which will end up on separate bundles as well.
The output of of the layout is an array of bundles. Each bundle is just an array for now with the dependencies in the order the requires where processed. Using this information we should be able to generate the actual bundles by using the `/path/to/entry/point.bundle` endpoint. We might change the structure of this json in the future, for instance to account for parent/child bundles relationships.
The next step will be to improve this algorithm to avoid repeating quite a bit dependencies across bundles.
Summary:
The word Package is overloaded, it may mean npm package, or may mean a collection of bundles. Neither is what we mean. We mean `bundle`.
This renames it and modernize some of the Bundler code.
Summary:
The cache is only used for JSTransformer at the moment. We're doing IO and some computation to get each module's name, whether is a haste or node module and it's dependencies. This work happens on startup so by caching this value we shouldbe able to reduce the start up time. Lets promote the Cache to the Packager level to be able to use it by any of the components of the packager. For now, on this diff we'll start using it to cache the mentioned fields.
Also we had to introduce the concept of fields in the cache as manually merging the date we had for each path is not possible as we're using promisses all around. With the new API, each field is a promise.
@amasad and I did some manual testing to measure the impact of this change and looks like it's saves 1 second when building the haste map (which represents 50% of the time). Overall this reduces 1 second of start up time which was currently about 8s on my mac book pro.
Summary:
Teach the resolver about platform-based resolution. The platform extension is inferred from the entry point.
It works for haste modules, as well as node-based resolution.
Summary:
@public
Now that watchman perf issue was fixed we can enable watchman-based fs crawling which is faster than node.
This showed an existing issue with some files missing from the blacklist which I addressed.
Test Plan:
./fbrnios.sh run
click around and scroll all the apps
Summary:
@public
We have a function that detects whether a give file is to be treated as a node_modules. If so it doesn't have access to the haste module map. There is an exception to this rule which is a few modules that are allowed to do that. Currently thats react-native, react-tools, and parse.
The current implementation had a bug where if you had `react-native` (or react-tools etc) in the name before the actual package root then the detection will be off. This fixes the problem by starting from the `lastIndexOf('node_modules')` directory, that way nothing confuses us.
Test Plan:
./runJestTests.sh
export OSS, patch, run e2e test
Summary:
@public
Fixes#773, #1055
The resolver was getting a bit unwieldy because a lot has changed since the initial writing (porting node-haste).
This also splits up a large complex file into the following:
* Makes use of classes: Module, AssetModule, Package, and AssetModule_DEPRECATED (`image!` modules)
* DependencyGraph is lazy for everything that isn't haste modules and packages (need to read ahead of time)
* Lazy makes it fast, easier to reason about, and easier to add new loaders
* Has a centralized filesystem wrapper: fast-fs (ffs)
* ffs is async and lazy for any read operation and sync for directory/file lookup which makes it fast
* we can easily drop in different adapters for ffs to be able to build up the tree: watchman, git ls-files, etc
* use es6 for classes and easier to read promise-based code
Follow up diffs will include:
* Using new types (Module, AssetModule etc) in the rest of the codebase (currently we convert to plain object which is a bit of a hack)
* using watchman to build up the fs
* some caching at the object creation level (we are recreating Modules and Packages many times, we can cache them)
* A plugin system for loaders (e.g. @tadeuzagallo wants to add a native module loader)
Test Plan:
* ./runJestTests.sh react-packager
* ./runJestTests.sh PackagerIntegration
* Export open source and run the e2e test
* reset cache
* ./fbrnios.sh run and click around
