It seems there is yet another focus regression in compose ui

Question

It seems there is yet another focus regression in compose ui 1 4 cold sweat ```E java lang IllegalStateException Check failed Full trace in thread at androidx compose ui focus FocusOwnerImpl moveFocus 3ESFkO8 FocusOwnerImpl kt 150 at androidx compose ui platform AndroidComposeView$keyInputModifier$1 invoke ZmokQxo AndroidComposeView android kt 212 ```

vide · Answer

Full trace ```E java lang IllegalStateException Check failed at androidx compose ui focus FocusTargetModifierNode fetchFocusProperties$ui release FocusTargetModifierNode kt 141 at androidx compose ui focus TwoDimensionalFocusSearchKt collectAccessibleChildren TwoDimensionalFocusSearch kt 216 at androidx compose ui focus TwoDimensionalFocusSearchKt findChildCorrespondingToFocusEnter OM vw8 TwoDimensionalFocusSearch kt 121 at androidx compose ui focus TwoDimensionalFocusSearchKt twoDimensionalFocusSearch OM vw8 TwoDimensionalFocusSearch kt 96 at androidx compose ui focus FocusTraversalKt focusSearch sMXa3k8 FocusTraversal kt 105 at androidx compose ui focus FocusOwnerImpl moveFocus 3ESFkO8 FocusOwnerImpl kt 150 at androidx compose ui platform AndroidComposeView$keyInputModifier$1 invoke ZmokQxo AndroidComposeView android kt 212 ```

vide · Answer

I ll try to investigate this

vide · Answer

The specific check that fails is this ``` ExperimentalComposeUiApi internal inline fun DelegatableNode visitAncestors mask Int block Modifier Node gt Unit TODO lmr we might want to add some safety wheels to prevent this from being called while one of the chains is being diffed updated Although that might only be necessary for visiting subtree check node isAttached ```

vide · Answer

I haven t found out a reliable way to reproduce it which makes me think it s probably a race condition somewhere as the node is not attached anymore when running focus search shrug This is a really big problem though for using 1 4 0

vide · Answer

I have 2 stack traces of this and it seems like both involve `findChildCorrespondingToFocusEnter` which might be a lead

jw · Answer

File an issue on the issue tracker

vide · Answer

I definitely will I m just hoping that gathering as much detail as possible would help the devs to narrow down the cause more effectively blob smile

vide · Answer

Also I don t think the description in the issue tracker is editable which might lead to some incoherent text if I notice something just after posting it sweat smile

vide · Answer

Both traces seem to have `twoDimensionalFocusSearch` in the stack only once which to my understanding would indicate the focus search is beginning from the root of the hierarchy I am not yet familiar with the Modifier Node architecture but this would seem to indicate that the root node is for some reason detached from the layout thinking face

jw · Answer

There s no limit to follow up comments though Trust me over communication is never a bad thing when chasing down bugs

vide · Answer

Here we go <https issuetracker google com issues 274897776>

vide · Answer

Update I managed to create a reliable demo of the focus crash and some edge conditions <https issuetracker google com issues 274897776 comment2> It seems to be related to several other bugs I have filed previously Most importantly it seems to be directly connected to the addition removal of nodes in focus modifier trees and the resetting of nodes in lazy layouts I am wondering if anyone has ideas for workarounds I m balancing between having to downgrade back to 1 3 x and backporting other relevant fixes from 1 4 0 as I can t roll to production with this sad panda