I'd be interested to hear from app developers on this list.
Here are a few quick thoughts I have.
As you note, this is a long-standing problem with layout in FX. You
mention in the performance considerations that for most cases this will
iterate quickly. It would be interesting to know what some of the corner
cases are, so we can see how bad the pathological case will be. I see
that you propose to iterate up to 100 times. Maybe a lower threshold
would be better? We already run layout passes 3 times in many cases. So
also running 3 (or maybe 4 or 5) times seems reasonable, especially when
your testing shows that most of the time it converges in <= 2 passes. So
a smaller threshold than 100 would seem to make sense. If a control
doesn't converge, you might consider logging it (in case an app
developer wants to debug it) perhaps at a "fine" level so it isn't shown
by default.
How do you propose to document this behavioral change -- not so much the
fact that it will (usually) get the right answer now instead of the
wrong one, but more about the multi-pass nature of it, and the fact that
nodes with text will be "preferred". Related to that, how necessary is a
new public API on Node?
-- Kevin
On 3/19/2021 3:02 PM, Michael Strauß wrote:
Trying to use baseline alignment in JavaFX can be pretty hard to get
right. For example: put some shape and a label into a layout
container, and the container's baseline might not reflect what you
would intuitively assume (which is the text baseline). In fact, most
layout containers just take the first child that reports a baseline
other than the special constant BASELINE_OFFSET_SAME_AS_HEIGHT and use
that as the baseline for the entire container.
Since there is no meaningful baseline if the first child is not a text
node, the layout container won't neatly align with other kinds of text
or text composites.
Another issue is that the baseline offset of a resizable node depends
on the height of the node, but the height of the node also depends on
the baseline offset. This circular dependency is currently unaccounted
for in controls such as Labeled (that's the reason for bugs like
JDK-809261).
I've prepared a PR that makes working with baseline alignments easier
and more consistent, and also solves the circular dependency problem
by introducing a multi-pass layout algorithm. The PR includes
before-and-after images that help visualize the problem and the
proposed solution.
Here's the PR: https://github.com/openjdk/jfx/pull/433
