Summary:
This is a revamp of how we decide whether to stop at a refresh boundary or to bubble to the next one.
We used to decide that at module initialization. However, that's both unnecessary overhead on start (for modules you don't plan to edit), and is actually insufficient.
In particular, even if a module only exports components (and thus is a Refresh Boundary), consecutive versions of that module might not be compatible.
For example, any of these changes should trigger reevaluation of parents:
- Adding or removing exports
- Renaming the exported component (which probably means you exported a different one, and we shouldn't preserve state)
- Converting from a class to a function, or back
- Wrapping something in a HOC
The new system handles these cases by comparing the Refresh "families" corresponding to exports, and bubbling the update up if some of them don't match up.
The tests have been rewritten from the webpack-inspired `module.hot` API (which we no longer use directly) to the Refresh API.
Reviewed By: rubennorte
Differential Revision: D16036044
fbshipit-source-id: 18018548d4aaa05877ae6fbaffe40c993c77cdf0
Summary:
Metro symbolication can be expensive in large apps. However, there is no need to symbolicate _runtime stacks from compile errors_. Those are pretty much useless anyway.
This will reduce the workload on Metro workers, and the delays when iterating with Fast Refresh, as the server will be busy much less often.
So I'm special-casing them and not sending the symbolication request anymore.
Reviewed By: rickhanlonii
Differential Revision: D16030087
fbshipit-source-id: 41f83ac01780c0a60cca777014e4ed95c0f3d14b
Summary:
Since we always create `module.hot` objects, the `module.hot` checks were unnecessary. They give a false impression that we're checking for a Hot Reloading mode. However, they're just Flow refinements and always exist in DEV. I made that explicit by throwing early.
Similarly, I removed a `module.hot` check inside `setupReactRefresh`, as it is always truish in DEV.
Finally, I'm adding a new mechanism as an escape hatch. It lets you do:
```
if (__DEV__) {
require.Refresh.forceFullRefresh = true;
}
```
in your entry point and opt into full refreshes on every edit. This sounds similar to "Reload-on-Save". That is because in the next diff, I plan to remove "Reload-on-Save" from user-visible options (but it'll stay for automated workflows).
So this workaround is intended for people who for one reason or another don't want to opt into Hot Reloading as an alternative. We'll need to talk to them and find out why.
Reviewed By: rickhanlonii
Differential Revision: D15958475
fbshipit-source-id: 674187ddf86a4e286dfae28f4182555a8b5d7396
Summary:
This removes React Transform HMR in favor of the new React Refresh implementation. It should only affect the "Enable Hot Reloading" mode. In further diffs I will remove "React Transform HMR" completely.
This is technically a breaking change for Metro so it'll need a version bump.
Reviewed By: rickhanlonii, rubennorte
Differential Revision: D15903585
fbshipit-source-id: 074380b00868cb31857f599a03799d3584c35d87
Summary:
This adds the Fresh Babel plugin and runtime to React Native dependencies. **They're not actually being used or enabled yet**. This is purely additive and just gets the deps setup out of the way for future diffs.
The `react-refresh` source of truth is in the React repo.
Reviewed By: cpojer
Differential Revision: D15828330
fbshipit-source-id: 67ec2dea8c896477ff8b434445f1730e388ea67a
Summary: Developer tools have a lot of dependencies on the bridge, so for now I'm just disabling them in bridgeless mode entirely by guarding everything in `setUpDeveloperTools` with the `RN$Bridgeless` flag. Also consolidating some of the stuff Dan added to `InitializeCore` for hot reloading in here.
Reviewed By: zackargyle
Differential Revision: D15797924
fbshipit-source-id: 360ea81a2844e49f7281eed259fc16a541148ac2
Summary: For bridgeless RN we're not going to use the JSTimers module + Timing native module for timers (e.g. setTimeout, setImmediate, etc.). Instead we're going to install global functions from cpp for each of these, so we can just skip setUpTimers entirely in this case.
Reviewed By: fkgozali
Differential Revision: D15790638
fbshipit-source-id: 1626fe90a27cb8d385cbb700ad932969f708f0cb
Summary:
In bridgeless mode we don't want to set up the batched bridge, which is set up as part of InitializeCore. Instead of deleting InitializeCore completely, let's just skip this step if we're in bridgeless mode, which we'll detect using a global variable set on the runtime from cpp (`RN$Bridgeless`).
This way you still get an error if the bridge is somehow not set up properly when you're not in bridgeless mode (it won't fail silently).
Reviewed By: fkgozali
Differential Revision: D15721940
fbshipit-source-id: 73896e25874dd000f37d1abc9cf6be549ab3434f
Summary:
This is the same diff as D14786123 but with one of the buck targets fixed that only failed on continuous and didn't run during land time.
This moves Map/Set to fb internal. We do not need them in open source any more but we still need this in some apps at FB that use an old version of JSC.
Reviewed By: rickhanlonii
Differential Revision: D15713305
fbshipit-source-id: caec43c76a6255b2af1693c13d8dea31d7d674f5
Summary:
Hot reloading propagates upwards through the inverse dependency tree — from a file you edited, to the files that import it, and so on. However, we can't always reevaluate everything. Many core infra modules can't run twice, and also the more you run, the more the risk of encountering a module with init side effects. So our practical compromise is to stop the propagation when we reach a module whose exports look like React components. We say that such module "accepts" an update. This means that in practice, changes trigger module reevaluation up to the closest component modules from the edited file. (If you edited a component file, it re-executes alone — unless it exports a non-component.)
However, current implementation has a problem. Sometimes there is an inverse dependency path that has no "accepting" modules whatsoever. For example, maybe you're editing some core module, and its inverse dependency tree reaches goes into React Native itself. Or maybe it reaches the entry point with a bunch of side effects that can't be repeated, like registering the app root component.
In the past, such cases would lead to confusing errors like "Expected `FBPrelude.conclude()` to have been called" after hot reload. This was because we kept re-executing modules all the way upwards, even if there is nothing that can accept the update on the path. Eventually we'd reach top-level modules in the import graph that don't like to run twice.
This diff changes the logic so that we *don't attempt* to re-execute the module factories if we know that some inverse dependency path doesn't terminate in a component. In that case we know we simply *can't apply the hot update* because it doesn't stop at a point we can handle, like a React component.
Since the hot update fails in this case, I'm making it fall back to a regular reload. This is similar to how webpack handles a similar situation on the web. This means that hot updates normally don't refresh, but if we can't apply a hot update to this file, we do refresh automatically.
Reviewed By: cpojer
Differential Revision: D15631864
fbshipit-source-id: 52cd1b03739fd760f1b1b1ab8c7276a150cc3c4c
Summary: This moves Map/Set to fb internal. We do not need them in open source any more but we still need this in some apps at FB that use an old version of JSC.
Reviewed By: TheSavior
Differential Revision: D14786123
fbshipit-source-id: 1c49b47d547ad30f2d93c00b44382cf410100b67
Summary: When metro is not running, D15559151 caused infinite exceptions (fetch threw an error if it couldn't connect to localhost:8081) which affected UI. Swallow those errors and everything works well, with or without metro.
Reviewed By: yungsters
Differential Revision: D15588623
fbshipit-source-id: d170ea82478545836a7a22a228196c9778e93ef0
Summary:
The original reason for vendoring the fetch polyfill was to remove the default blob response type but this was reverted.
Here's a little history around the fetch polyfill and the blob issue:
- Original commit introducing the vendored polyfill: #19333, the goal was to fix a memory leak because our blob implementation doesn't release resources automatically. Not an ideal fix but since the issue was pretty severe and the infra for a proper fix was not in place.
- This introduced an issue when downloading images using `fetch` which was fixed by #22063 which re-added the default blob content type. However that re-introduced the original fetch memory leak.
- We have better infra now with jsi and I was able to get blob deallocation working, see #24405
Currently the vendored fetch polyfill is useless since it was changed back to the original version. We can just use the npm version again. I also updated to 3.0 which brings better spec compliance and support for cancellation via `AbortController`, https://github.com/github/fetch/releases/tag/v3.0.0.
## Changelog
[General] [Changed] - Remove vendored fetch polyfill, update to whatwg-fetch@3.0
Pull Request resolved: https://github.com/facebook/react-native/pull/24418
Differential Revision: D14932683
Pulled By: cpojer
fbshipit-source-id: 915e3d25978e8b9d7507ed807e7fba45aa88385a
Summary:
People ask "How do you use `console.log` with React Native?" and there is no good answer. This diff aims to stop people from asking this question.
See https://fb.workplace.com/groups/rn.core/permalink/2372327062999018/ for context.
This logging relies on network requests which can cause logs to show up out-of-order. To reduce the likelihood I queue every log message on the server for a maximum of 200ms. There could be other methods, like using websocket, but that seems more complex than is necessary at least in the beginning.
I considered various throttling strategies because this could be quite chatty and possibly problematic, however I think we can just ship this and iterate based on feedback. On my very underpowered laptop I logged a random number every 10 milliseconds and it didn't cause any issues or slowdown.
Reviewed By: gaearon
Differential Revision: D15559151
fbshipit-source-id: 552001622af0937ae3a37d2bd8c1b96e7ca52020
Summary:
This is another step in moving RN towards standard path-based requires, updating more code to use path-based requires. See the umbrella issue at https://github.com/facebook/react-native/issues/24316 for more detail.
## Changelog
[General] [Changed] - Replace more Haste imports with path-based imports
Pull Request resolved: https://github.com/facebook/react-native/pull/25001
Differential Revision: D15467829
Pulled By: cpojer
fbshipit-source-id: 58c364bb4c1c757689907d5ed0d0f3fac0e22f3f
Summary: This check isn't needed in prod as the bundle is served with the app. In dev mode it's possible to have native and JS out of sync.
Reviewed By: yungsters
Differential Revision: D15268485
fbshipit-source-id: 9aeeb6cf5ca91baa90b85e18c848c3b10d85b0f7
Summary: If some stack frames in a trace fail to symbolicate (or are genuinely unmapped), their `frame` field will be null, and `ExceptionsManager.js` will currently crash. This diff lets it recover gracefully and show whatever information is available.
Reviewed By: dcaspi
Differential Revision: D15296220
fbshipit-source-id: 2b1006b1354295171b25bfc6230c5b3e0c57f67f
Summary:
This PR is related to #24760 and adds the `openURLInBrowser` functionality introduced on react-native-community/cli#383.
[General] [Changed] - Open links from new app in computer's browser.
Pull Request resolved: https://github.com/facebook/react-native/pull/24843
Reviewed By: rickhanlonii
Differential Revision: D15334011
Pulled By: cpojer
fbshipit-source-id: 947ad1b113923989cf706e60851e02a87e1099e8
Summary:
This is the next step in moving RN towards standard path-based requires. All the requires in `Libraries` have been rewritten to use relative requires with a few exceptions, namely, `vendor` and `Renderer/oss` since those need to be changed upstream. This commit uses relative requires instead of `react-native/...` so that if Facebook were to stop syncing out certain folders and therefore remove code from the react-native package, internal code at Facebook would not need to change.
See the umbrella issue at https://github.com/facebook/react-native/issues/24316 for more detail.
[General] [Changed] - Migrate "Libraries" from Haste to standard path-based requires
Pull Request resolved: https://github.com/facebook/react-native/pull/24749
Differential Revision: D15258017
Pulled By: cpojer
fbshipit-source-id: a1f480ea36c05c659b6f37c8f02f6f9216d5a323
Summary:
On iOS, `RCTRedBox` will not update the displayed stack trace if the message string sent with the update differs from the original error message. As JS errors are shown in two stages - before and after symbolication - there was previously a case where the message would differ between the two `updateExceptionMessage` calls, blocking the update and leaving only the unsymbolicated trace visible. This diff fixes that.
Longer term, we should also change `RCTRedBox`'s logic to rely on the JS-provided `exceptionID` instead of the message string - similar to what we do on Android.
Changelog:
[iOS] [Fixed] - Fix redbox JS symbolication when adding JS engine tag to the message
Reviewed By: sahrens
Differential Revision: D15202524
fbshipit-source-id: 237fc090e88b0c609865e0aed842d6a609c1239a
Summary:
This adds https://github.com/mysticatea/abort-controller to polyfill [AbortController](https://developer.mozilla.org/en-US/docs/Web/API/AbortController). This is used to cancel requests when using `fetch`.
This also updates `event-target-shim` to 5.0 to make sure we only have one version of this dependency. This updates required adding a polyfill for `console.assert` which is used by the new version. I made one based on https://github.com/gskinner/console-polyfill/blob/master/console.js#L74.
The polyfill is very small, especially since we already use `event-target-shim` so I think it makes sense to include in core.
Depends on #24418 so that the fetch polyfill supports the `signal` parameter.
Fixes#18115
[General] [Added] - Add support for cancelling fetch requests with AbortController
Pull Request resolved: https://github.com/facebook/react-native/pull/24419
Differential Revision: D14912858
Pulled By: cpojer
fbshipit-source-id: 8a6402910398db51e2f3e3262f07aabdf68fcf72
Summary: Now that React Native ships with a newer version of JSC, we do not need this code to wrap `Object.freeze` any longer.
Reviewed By: mmmulani
Differential Revision: D14779239
fbshipit-source-id: 1a6e1a9c7f4312572bd08ba604fa8c9d6b1927e1
Summary: This is the first diff in an effort to remove Geolocation from React Native. This diff removes the globally injected navigator.geolocation feature and instead requires explicit importing of `Geolocation`. When using Web APIs, people will need to patch `navigator.geolocation` on their own from now on.
Reviewed By: sahrens
Differential Revision: D14692386
fbshipit-source-id: c57b290b49728101250d726d67b1956ff23a9a92
Summary:
I'm changing all callsites to use either global or scoped perf logger explicitly in one diff.
`GlobalPerformanceLogger` is basically a singleton
`scopedPerformanceLogger` is scoped to the React tree by using a React Context
Reviewed By: sahrens
Differential Revision: D14186694
fbshipit-source-id: 062c76eea8fce9d9b531f0eddf153bb79d52f68d
Summary: This is one more step to remove `fbjs` from `react-native-github`. This changes both the internal and external code to use `invariant` from zertosh instead of the copy in fbjs.
Reviewed By: yungsters
Differential Revision: D13195941
fbshipit-source-id: 73564ca1715110e7da9c7ef56dc57374d61377e0
Summary: It seems like `nativePerformanceNow` might not be getting installed in time to be used by InitializeCore on Android. Using PerformanceLogger.currentTimestamp instead, which uses nativeQPLTimestamp (which is what we should be using anyway)
Reviewed By: alexeylang
Differential Revision: D12875187
fbshipit-source-id: 6eda5d2ed7948ba48c63f76b40b2014511c32494
Summary:
Flow is broken for a bunch of people and has been for the last few days. For some reason this isn't causing any builds to fail which we will need to look into. For now, this stops the bleeding.
I *think* this is because of case sensitivity differenes between Mac and where all of our CI machines run.
```
Error ┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈ Libraries/Core/polyfillPromise.js:21:33
Importing from an untyped module makes it any and is not safe! Did you mean to add // flow to the top of Promise?
(untyped-import)
18│ * If you don't need these polyfills, don't use InitializeCore; just directly
19│ * require the modules you need from InitializeCore for setup.
20│ */
21│ polyfillGlobal('Promise', () => require('Promise'));
22│
```
Pull Request resolved: https://github.com/facebook/react-native/pull/22048
Reviewed By: shergin
Differential Revision: D12918418
Pulled By: TheSavior
fbshipit-source-id: 9d9db1b3686eab906244180325c12ffcce92be8c
Summary: Split up InitializeCore into a bunch of modules. The idea here is to make it easier for apps to just get the initialization logic they want and leave behind what they don't; for example, if you don't want the Map/Set polyfills, instead of requiring InitializeCore you can require the modules you want from it.
Reviewed By: yungsters
Differential Revision: D10842564
fbshipit-source-id: 3b12d54fddea8c4ee75886022338c214987a015c
Summary: Instead of having a single point at the top of InitializeCore, let's just create a subspan for it. Initially I just wanted to use this point to track JS start time, but it'll be useful to see how long initializeCore takes, too.
Reviewed By: alexeylang
Differential Revision: D10521595
fbshipit-source-id: 3025c34ffab39b79efc966f0c0eb6f502c91c550
Dan Abramov
Emily Janzer
Christoph Nakazawa
Felix Oghina
Joshua Gross
Rubén Norte
Janic Duplessis
Dan Nguyen
Tom Sanderson
James Ide
Jean Regisser
Wojteg1337
Michał Pierzchała
Mehdi Mulani
Moti Zilberman
Lucas Bento
Brian Zhao
Alexey Lang
Alexander Zhang
George Zahariev
Kevin Gozali
Daniel Sainati
Ben Newman
Matt Hargett
Ignacio Olaciregui
empyrical
