RFR: JDK-8199216: Memory leak and quadratic layout time with nested nodes (hbox) and pseudo-class in style sheet [v2]

[Michael Strauß]

On Thu, 30 Mar 2023 20:34:13 GMT, John Hendrikx <jhendrikx at openjdk.org> wrote:

>> This fix introduces immutable sets of `PseudoClass` almost everywhere, as they are rarely modified.  These are re-used by caching them in a new class `ImmutablePseudoClassSetsCache`.
>> 
>> In order to make this work, `BitSet` had to be cleaned up.  It made assumptions about the collections it is given (which may no longer always be another `BitSet`).  I also added the appropriate null checks to ensure there weren't any other bugs lurking.
>> 
>> Then there was a severe bug in `toArray` in both the subclasses that implement `BitSet`.
>> 
>> The bug in `toArray` was incorrect use of the variable `index` which was used for both advancing the pointer in the array to be generated, as well as for the index to the correct `long` in the `BitSet`.  This must have resulted in other hard to reproduce problems when dealing with `Set<PseudoClass>` or `Set<StyleClass>` if directly or indirectly calling `toArray` (which is for example used by `List.of` and `Set.of`) -- I fixed this bug because I need to call `Set.copyOf` which uses `toArray` -- as the same bug was also present in `StyleClassSet`, I fixed it there as well.
>> 
>> The net result of this change is that there are far fewer `PseudoClassState` objects created; the majority of these are never modified, and the few that are left are where you'd expect to see them modified.
>> 
>> A test with 160 nested HBoxes which were given the hover state shows a 99.99% reduction in `PseudoClassState` instances and a 70% reduction in heap use (220 MB -> 68 MB), see the linked ticket for more details.
>> 
>> Although the test case above was extreme, this change should have positive effects for most applications.
>
> John Hendrikx has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Simplify ImmutablePseudoClassSetsCache and avoid an unnecessary copy

modules/javafx.graphics/src/main/java/com/sun/javafx/css/BitSet.java line 584:

> 582:      * @param obj the object to cast, cannot be {@code null}
> 583:      * @return a type T, or {@code null} if the argument was not of this type
> 584:      * @throws NullPointerException when {@code obj} is {@code null}

Previously, this method always returned an instance of `T`. Now that is not the case, it might also simply return `null` if the argument passed into it is an instance of a different class. I think it makes sense to also return `null` when the argument passed into the method is `null`.

-------------

PR Review Comment: https://git.openjdk.org/jfx/pull/1076#discussion_r1154620769