Summary:
`VirtualizedList_EXPERIMENTAL` is a fork of VirtualizedList. It was originally created for the purpose of keyboard/a11y fixes, which required the list to allow rendering discontiguous regions of cells. So, the large part of the refactor is in the data structure used to represent state (now a sparse bitmask), and building out the render function to operate off the bitmask.
This structure allows pre-rendering areas where we know focus must be able to be synchronously be moved to. This is exercised on platforms like Windows and macOS which fire view level focus events. The new implementation otherwise aims to preserve near-exact list behavior as previously.
Apart from the structure change, the refactor has made the code (subjectively) a lot easier to reason about, and there are now stronger internal invariants than before. The bitmask structure also enables new approaches for nested list stability, and implementing OffScreen.
**Why ship it now?**
Having two implementations has multiple times prevented other changes to VirtualizedList, so it is incurring a constant engineer cost.
At this point, we have a lot of reason to be confident in the maturity of the new list implementation (see the test plan below). So, reconciling the implementations not only unlocks the enhancements of the new list, but unblocks future changes to VirtualizedList.
## Test Plan
1. **Unit tests:** A few dozen unit tests were added to VirtualizedList, validating the regions it renders in response to injected input, across its API surface. Both VirtualizedList implementations must pass the same tests, around expected behavior.
2. **RNTester:** Scenarios in RNTester were manually validated, and all worked as before
3. **New invariants:** I added a lot of internal logic checks and bounds checks as invariant violations, to try to produce signal if any part of the refactor went unexpected in the wild.
4. **MSFT Rollout:** I rolled out the changes as a fork of VirtualizedList at MSFT to a couple of high-usage surfaces using Android, iOS, and Windows on Paper. Some invariant violations were surfaced and fixed, but telemetry showed solid system and scenario health with the change.
5. **Meta import:** Fixed all land-blocking test failures when using the new list. These were confined to updating snapshots, and adding additional checks for invalid input.
6. **Panel apps:** Manually validated top-level surfaces of panel apps in debug/release modes. Fixed a couple of invariant violations and novel usages. Profiled new JS structures against an expensive list with Hermes profiler.
7. **Facebook App Rollout:** After some manual validation of top-level surfaces, the change was rolled out to Facebook for Android and iOS as an experiment. New invariant violations were fixed, and the change has sat in production for several weeks at this point, while measuring impact to metrics like error rate, responsiveness, and impressions.
This is the first introduction to all of OSS, and some Meta internal apps. This means there may still be novel usages of VirtualizedList that were not picked up during testing, but I think there has been enough diligence to roll out what has been tested, and forward fix anything unexpected that might still come up.
Changelog:
[General][Changed] - Ship VirtualizedList_EXPERIMENTAL
Reviewed By: javache
Differential Revision: D40259791
fbshipit-source-id: 63eee9381d197a1e38ae663b2158436ff135c0e1
Summary:
Following up on the previous diff to remove the only usage of `listKey` for persistent association, we can remove the need for a manual `listKey`, and instead rely on per-instance association via refs.
A followup change will remove the existing usages.
Changelog:
[General][Removed] - Remove VirtualizedList `listKey` prop
Reviewed By: p-sun
Differential Revision: D39466677
fbshipit-source-id: 6b49f45c987fff9836918ba833fbb16f24414ff8
Summary:
When `getItemLayout` is not supplied, VirtualizedList doesn't know the measurements for its child items ahead of time. It caches most recent cell measurements so that an item which is virtualized away is associated with its correct height. VirtualizedList will not consume space for an item it has not yet rendered/knows measurements for.
This also means an initially created list will only take up the height of its `initialNumToRender()`. This can cause scroll jumping for the scenario of nested VirtualizedLists of the same orientation, since a virtualized away child list will shrink (compared to the spacer), then re-expand, once returned to.
There is logic in the VirtualizedList constructor to load cached frame metrics based on supplied listKey, to enable reloading most recent state. Though locally it seems to be reading the context too early in the lifecycle (compared to other successful checks during componentDidMount, render). It was moved from `componentDidMount` to `constructor` as part of the initial change (https://www.internalfb.com/diff/D6330846 (https://github.com/facebook/react-native/commit/2668dc8e1be7bf93e65ca2e11c87cbeef3310c3e)?dst_version_fbid=131968984136932&transaction_fbid=1578362385612542), and I think has likely been unreliable or dead since.
Restoring state was the only usage of a durable key for nested lists beyond their lifetime. So removing the dead code lets us remove the need for a `listKey`, which has been error-prone. This still leaves the problem present since inception, but there may be alternate solutions like scroll anchoring.
Changelog:
[Internal][Removed] - Remove dead code for restoring state to nested VirtualizedList
Reviewed By: ryancat
Differential Revision: D39466678
fbshipit-source-id: fc2f39802e5cdf9b920974333be64c74211b99a9
Summary:
This change is in preparation of adding a separate `VirtualizedList_EXPERIMENTAL` component.
Both the original, and experimental lists use `VirtualizedListContext`, which itself references back to the VirtualizedList class type. VirtualizedList private methods are currently included in the type system, and are called in other VirtualizedList code (see https://github.com/facebook/react-native/commit/b2f871a6fa9c92dd0712055384b9eca6d828e37d). This prevents Flow from seeing the two classes are compatible if "private" methods change.
My first attempt was to parameterize the context, to allow both `VirtualizedList`, and `VirtualizedList_EXPERIMENTAL` to use the same code without sacrificing type safety or adding further duplication. This added more complexity than it is worth, so I am instead loosening the type on VirtualizedListContext to pass around a more generic handle.
Changelog:
[Internal][Changed] - Allow VirtualizedListContext to support different component type
Reviewed By: javache
Differential Revision: D38017086
fbshipit-source-id: 91e8f6ab2591d3ae9b7f9263711b4a39b78f68e0
Summary:
When a FlatList is nested inside another FlatList, it may be re-rendered whenever the outer FlatList renders. Apply the same optimization we already had in `VirtualizedListContextProvider` to avoid changing the context object if no values have changed.
Changelog: [General][Changed] - Optimized VirtualizedList context when used with nested lists
Reviewed By: genkikondo
Differential Revision: D35905952
fbshipit-source-id: 695253c85db2043d22e208ad94ecc7daa1455055
Summary:
ES Modules implicitly enable strict mode. Adding the "use strict" directive is, therefore, not required.
This diff removes all "use strict" directives from ES modules.
Changelog:
[Internal]
Reviewed By: motiz88
Differential Revision: D26172715
fbshipit-source-id: 57957bcbb672c4c3e62b1db633cf425c1c9d6430
Summary:
Changelog:
[General] [Changed] - Added (DEV-only) `displayName` to some RN contexts to make them more easy to differentiate when debugging.
Reviewed By: lunaleaps
Differential Revision: D24993068
fbshipit-source-id: 4904259eda50444c2f74700a3540ff4fd02ac322
Nick Gerleman
Tim Yung
Nick Gerleman
islandryu
Pieter De Baets
Andres Suarez
Tim Yung
Micha Reiser
Brian Vaughn