This reverts commit
d24c01bd0c.
The original change was reverted because it caused some apps to get
stuck on the splash screen on some phones.
An investigation determined that this was due to a rounding error.
Example: The device reports a physical size of 1008.0 x 2198.0 with a
dpr of 1.912500023841858. Flutter would translate that to a logical size
of 527.0588169589221 x 1149.2810314243163 and use that as the input for
its layout algorithm. Since the constraints here are tight, the layout
algorithm would determine that the resulting logical size of the root
render object must be 527.0588169589221 x 1149.2810314243163.
Translating this back to physical pixels by applying the dpr resulted in
a physical size of 1007.9999999999999 x 2198.0 for the frame. Android
now rejected that frame because it didn't match the expected size of
1008.0 x 2198.0 and since no frame had been rendered would never take
down the splash screen.
Prior to dynamically sized views, this wasn't an issue because we would
hard-code the frame size to whatever the requested size was.
Changes in this PR over the original PR:
* The issue has been fixed now by constraining the calculated physical
size to the input physical constraints which makes sure that we always
end up with a size that is acceptable to the operating system.
* The `ViewConfiguration` was refactored to use the slightly more
convenient `BoxConstraints` over the `ViewConstraints` to represent
constraints. Both essentially represent the same thing, but
`BoxConstraints` are more powerful and we avoid a couple of translations
between the two by translating the` ViewConstraints` from the
`FlutterView` to `BoxConstraints` directly when the `ViewConfiguration`
is created.
All changes over the original PR are contained in the second commit of
this PR.
Fixes b/316813075
Part of https://github.com/flutter/flutter/issues/134501.
Towards https://github.com/flutter/flutter/issues/134501.
This change is based on https://github.com/flutter/engine/pull/48090. It changes the `RenderView` to be dynamically sized based on its content if the `FlutterView` it is configured with allows it (i.e. the `FlutterView` has loose `FlutterView.physicalConstraints`). For that, it uses those `physicalConstraints` as input to the layout algorithm by passing them on to its child (after translating them to logical constraints via the device pixel ratio). The resulting `Size` that the `RenderView` would like to be is then communicated back to the engine by passing it to the `FlutterView.render` call.
Tests will fail until https://github.com/flutter/engine/pull/48090 has rolled into the framework.
Avoid using string replacement on a bogus string to get the relative directory of this test.
https://github.com/flutter/engine/pull/46911 will fix the behavior of Platform.script so it actually has the file being executed, but that will not be the file the test author wrote, it will be `listener.dart.dill` generated by tool code and running in a temp directory for this particular test.
We may want to come up with a more robust way of providing the test for framework test users, but we should do so in a way that will not limit the ability to run multiple tests in a single process if that offers performance benefits.
Fixes https://github.com/flutter/flutter/issues/131627
Originally this code sometimes was returning null and sometimes was failing, when stack frame is in unexpected format.
This PR updates for one of the code paths from failing to returning null.
This change enables Flutter to generate multiple Scenes to be rendered into separate FlutterViews from a single widget tree. Each Scene is described by a separate render tree, which are all associated with the single widget tree.
This PR implements the framework-side mechanisms to describe the content to be rendered into multiple views. Separate engine-side changes are necessary to provide these views to the framework and to draw the framework-generated Scene into them.
## Summary of changes
The details of this change are described in [flutter.dev/go/multiple-views](https://flutter.dev/go/multiple-views). Below is a high-level summary organized by layers.
### Rendering layer changes
* The `RendererBinding` no longer owns a single `renderView`. In fact, it doesn't OWN any `RenderView`s at all anymore. Instead, it offers an API (`addRenderView`/`removeRenderView`) to add and remove `RenderView`s that then will be MANAGED by the binding. The `RenderView` itself is now owned by a higher-level abstraction (e.g. the `RawView` Element of the widgets layer, see below), who is also in charge of adding it to the binding. When added, the binding will interact with the `RenderView` to produce a frame (e.g. by calling `compositeFrame` on it) and to perform hit tests for incoming pointer events. Multiple `RenderView`s can be added to the binding (typically one per `FlutterView`) to produce multiple Scenes.
* Instead of owning a single `pipelineOwner`, the `RendererBinding` now owns the root of the `PipelineOwner` tree (exposed as `rootPipelineOwner` on the binding). Each `PipelineOwner` in that tree (except for the root) typically manages its own render tree typically rooted in one of the `RenderView`s mentioned in the previous bullet. During frame production, the binding will instruct each `PipelineOwner` of that tree to flush layout, paint, semantics etc. A higher-level abstraction (e.g. the widgets layer, see below) is in charge of adding `PipelineOwner`s to this tree.
* Backwards compatibility: The old `renderView` and `pipelineOwner` properties of the `RendererBinding` are retained, but marked as deprecated. Care has been taken to keep their original behavior for the deprecation period, i.e. if you just call `runApp`, the render tree bootstrapped by this call is rooted in the deprecated `RendererBinding.renderView` and managed by the deprecated `RendererBinding.pipelineOwner`.
### Widgets layer changes
* The `WidgetsBinding` no longer attaches the widget tree to an existing render tree. Instead, it bootstraps a stand-alone widget tree that is not backed by a render tree. For this, `RenderObjectToWidgetAdapter` has been replaced by `RootWidget`.
* Multiple render trees can be bootstrapped and attached to the widget tree with the help of the `View` widget, which internally is backed by a `RawView` widget. Configured with a `FlutterView` to render into, the `RawView` creates a new `PipelineOwner` and a new `RenderView` for the new render tree. It adds the new `RenderView` to the `RendererBinding` and its `PipelineOwner` to the pipeline owner tree.
* The `View` widget can only appear in certain well-defined locations in the widget tree since it bootstraps a new render tree and does not insert a `RenderObject` into an ancestor. However, almost all Elements expect that their children insert `RenderObject`s, otherwise they will not function properly. To produce a good error message when the `View` widget is used in an illegal location, the `debugMustInsertRenderObjectIntoSlot` method has been added to Element, where a child can ask whether a given slot must insert a RenderObject into its ancestor or not. In practice, the `View` widget can be used as a child of the `RootWidget`, inside the `view` slot of the `ViewAnchor` (see below) and inside a `ViewCollection` (see below). In those locations, the `View` widget may be wrapped in other non-RenderObjectWidgets (e.g. InheritedWidgets).
* The new `ViewAnchor` can be used to create a side-view inside a parent `View`. The `child` of the `ViewAnchor` widget renders into the parent `View` as usual, but the `view` slot can take on another `View` widget, which has access to all inherited widgets above the `ViewAnchor`. Metaphorically speaking, the view is anchored to the location of the `ViewAnchor` in the widget tree.
* The new `ViewCollection` widget allows for multiple sibling views as it takes a list of `View`s as children. It can be used in all the places that accept a `View` widget.
## Google3
As of July 5, 2023 this change passed a TAP global presubmit (TGP) in google3: tap/OCL:544707016:BASE:545809771:1688597935864:e43dd651
## Note to reviewers
This change is big (sorry). I suggest focusing the initial review on the changes inside of `packages/flutter` first. The majority of the changes describe above are implemented in (listed in suggested review order):
* `rendering/binding.dart`
* `widgets/binding.dart`
* `widgets/view.dart`
* `widgets/framework.dart`
All other changes included in the PR are basically the fallout of what's implemented in those files. Also note that a lot of the lines added in this PR are documentation and tests.
I am also very happy to walk reviewers through the code in person or via video call, if that is helpful.
I appreciate any feedback.
## Feedback to address before submitting ("TODO")
This PR is adding a file of two simple unit tests in order to improve code coverage (covering two simple classes within the annotations.dart file).
The issue this is fixing is the lack of complete test coverage for the annotations.dart file.
## FlutterTimeline
Add a new class `FlutterTimeline` that's a drop-in replacement for `Timeline` from `dart:developer`. In addition to forwarding invocations of `startSync`, `finishSync`, `timeSync`, and `instantSync` to `dart:developer`, provides the following extra methods that make is easy to collect timings for code blocks on a frame-by-frame basis:
* `debugCollect()` - aggregates timings since the last reset, or since the app launched.
* `debugReset()` - forgets all data collected since the previous reset, or since the app launched. This allows clearing data from previous frames so timings can be attributed to the current frame.
* `now` - this was enhanced so that it works on the web by calling `window.performance.now` (in `Timeline` this is a noop in Dart web compilers).
* `collectionEnabled` - a field that controls whether `FlutterTimeline` stores timings in memory. By default this is disabled to avoid unexpected overhead (although the class is designed for minimal and predictable overhead). Specific benchmarks can enable collection to report to Skia Perf.
## Semantics benchmarks
Add `BenchMaterial3Semantics` that benchmarks the cost of semantics when constructing a screen full of Material 3 widgets from nothing. It is expected that semantics will have non-trivial cost in this case, but we should strive to keep it much lower than the rendering cost. This is the case already. This benchmark shows that the cost of semantics is <10%.
Add `BenchMaterial3ScrollSemantics` that benchmarks the cost of scrolling a previously constructed screen full of Material 3 widgets. The expectation should be that semantics will have trivial cost, since we're just shifting some widgets around. As of today, the numbers are not great, with semantics taking >50% of frame time, which is what prompted this PR in the first place. As we optimize this, we want to see this number improve.
... found when looking at leak tracker today.
By the way, shall we add some kind of automatic link linter, which goes through all links and see whether they are alive?
