* Safer checks in ASDisplayNode’s setFrame before assigning bounds and position
* Consolidate checks in one place
* Assert with ASDisplayNodeNonFatal
* Allow position to be negative
* Add inverted property to ASTableNode
* Add inverted property to ASCollectionNode
* Add example
* De-inform data controller and use different CATransform3D
* Add logic to pending state
* Add accessor for inverted property
* [ASDisplayNode] Ensure all subclasses are using base class __instanceLock__ and not re-defining their own.
This also moves the @package definition of the instance variable to +FrameworkPrivate instead of Internal.h,
because Internal.h should ideally not be used outside of the ASDisplayNode file setup. This has greatly reduced
the number of imports of Internal.h.
* [ASDisplayNode] Add ASDisplayNode+FrameworkSubclasses.h to share __instanceLock__ definition.
* Dealloc _pendingViewState if range managed
_ASPendingState objects can add up very quickly when adding
many nodes. This is especially an issue in large collection views
and table views. This needs to be weighed against the cost of
reallocing a _ASPendingState. So in range managed nodes we
delete the pending state, otherwise we just clear it.
* Adlai's comments
Summary:
We observed a deadlock which occurred when two threads were laying out the same set of nodes.
On one thread, layout would occur on a leaf node. It would lock and as part of this layout
process, ASDK walks up the node tree and calls __setNeedsLayout on its supernode until it
reaches the supernode with no supernode. When the supernode gets its call to __setNeedsLayout
it also locks. So leaf node locks and then awaits supernode lock.
On another thread, we're doing a layout pass on the supernode in the above thread. This locks
the supernode and attempts to lock the leaf node. This deadlocks (remember the above thread
is holding onto the leaf lock and awaiting the supernode lock. This thread is holding onto
the supernode lock and awaiting the leaf lock).
This is all exacerbated by the use of recursive locks.
These optimizations are surprisingly impactful. -setNeedsDisplay being called
for every node triggered cancelAsyncDisplay, locking, and memory management overhead
that is completely avoidable because Core Animation triggers first display automatically.
The mutable array copy optimizations reduced this key cost by over 10x, from 52ms to 5ms
on an iPad Air 2 / A8X with a real-world test case.
