Some fixes #2
Some fixes #2
Conversation
| unused_layers = unused_layers]() mutable { | ||
| auto task = [this, platform_view_layers = std::move(platform_view_layers), | ||
| callbacks = std::move(callbacks), | ||
| unused_layers = std::move(unused_layers)]() mutable { | ||
| TRACE_EVENT0("flutter", "FlutterPlatformViewsController::SubmitFrame::CATransaction"); |
There was a problem hiding this comment.
We're capturing this anyway and calling methods on it (CompositeWithParams, RemoveUnusedLayers) which modifies variables such as views_to_recomposite_ - holding on to previous value in capture causes crashes when removing views.
We may want to synchronize this (i.e. block raster thread until the block completes) - in case there is something holding the platform thread for longer than expected we can have race condition.
| @@ -783,34 +784,26 @@ static bool ClipRRectContainsPlatformViewBoundingRect(const SkRRect& clip_rrect, | |||
| DisposeViews(); | |||
|
|
|||
| // Composite Platform Views. | |||
| for (auto view_id : views_to_recomposite) { | |||
| CompositeWithParams(view_id, current_composition_params[view_id]); | |||
| for (auto view_id : views_to_recomposite_) { | |||
There was a problem hiding this comment.
DisposeViews modifies views_to_recompose_ so we can't use captured value.
|
|
||
| platform_task_runner_->PostTask(task); | ||
| latch.Wait(); |
There was a problem hiding this comment.
This seems necessary otherwise things quickly crash in release mode. I assume there is some state read on platform thread that is being touched in next preroll. On macOS we can get away with not blocking the raster thread because we have stricter separation of data. Possible could do it here too if the wait causes performance issues.
|
I tried this PR with Wonderous + |
Relands flutter#54917. The change is the same as before, except now the native resources for `SkwasmColorFilter` and `SkwasmImageFilter` classes are no longer GC'd. Instead, we use manually managed native handles (vended and scoped by `withRawColorFilter` and `withRawImageFilter`). The bug in the previous PR was that filter objects were disposed with the paint while the framework continued holding onto them. When GC kicked the finalization registry, it attempted to double-free the filters.
Some fixes I did to get wonderous working.