I just want to quickly mention that option 3
(https://github.com/openjdk/jfx/pull/185) is no longer a WIP -- all
functionality is working all tests pass.
The PR creates a TreeShowingExpression class which encapsulates the tree
showing logic that was in Node class. This class is then only used by
ProgressIndicatorSkin and PopupWindow. This effectively means that
instead of registering a listener on Window and Scene for all nodes
they're only registered for those classes that need to know the showing
status of a Scene.
--John
On 27/08/2020 02:17, Kevin Rushforth wrote:
Sorry for the badly formatted html. Here it is again.
I see some progress being made on the {Tree}TableView performance issue.
To summarize where I think we are:
There are currently 2 different approaches under review:
1. https://github.com/openjdk/jfx/pull/108 : optimization in javafx.base
to make removing listeners faster
2. https://github.com/openjdk/jfx/pull/125 : optimization in TableView
to reduce the number of add / removes
In addition, the following is a WIP PR that the author thinks could be a
3rd approach:
3. https://github.com/openjdk/jfx/pull/185 : optimization in scene graph
to avoid install treeShowing listeners on Window and Scene for all nodes
Jose has proposed a 4th approach as a comment to PR #108, and as I
understand it, he will propose a PR shortly.
4. Don't clear the list of children in a VirtualFlow when the number of
items changes.
So the first thing that is needed is to evaluate the approaches and
decide which one to pursue.
Options 1 and 3 are more broad in their scope, option #2 is more
targeted (to TableView), but within that area is a larger change. Option
#3 would remove the (internal) treeShowing property as a generally
available concept and only use it for controls like ProgressIndicator
that really need it. Option #4 affects only controls that use
VirtualFlow (ListView, TableVIew, TreeTableView), and seems not to be a
large change (presuming we can verify that no leak is introduced).
I note that these fixes are not mutually exclusive, but I do think we
need to settle on a primary approach and use that to fix this issue. If
there are still performance problems after that fix, we can consider one
(or more) of the others.
Comments?
-- Kevin
