- The underlying changes are in windowed-list, which the item tree and
all virtualized tables are based on, so if there are bugs they might
show up outside of the collection tree.
- The expensive operation is adding/removing rows, since row offsets
have to be recalculated (this includes collapsing/expanding rows).
- The cost on further drawing while scrolling is constant and shouldn't
affect performance much.
Without the /u (Unicode) flag, [\ud800-\udfff] matches *all* surrogate pair code
points, even when they're in a valid position in the string.
Fixes#3421
Some pages (notably the first DOI.js test case) never get to "complete". I don't
know why! But for practically all translators that don't use defer: true,
"interactive" is ready enough.
Amends 30c70a6ecd.
On some machines, the callback added in that change would never actually run, so
memory usage would grow quickly when switching between tabs. By wrapping in
setTimeout() instead of requestIdleCallback(), we keep the performance gains
from removing once the element is no longer visible (so we can avoid repaints,
rebuilding the layout tree, etc.), but we make sure that the callback gets
called nearly immediately.
#3321
It's now possible to tab through the panes and the buttons, and
left/right-arrow also change the selection, so you can use Left/Right +
Return to move through multiple conflicts.
Fixes#3395
In Z7 on Windows 10 (but not 11 for me), nsIZipReader doesn't properly
close the file after `findEntries()` is called (as discovered by
@abaevbog), so a `remove()` on the downloaded ZIP file during file
syncing triggers an access-denied error. Setting the zip-reader variable
to null and forcing garbage collection seems to fix it. Doing this
everywhere we use nsIZipReader just to be safe.
I found the `forceGC()` in only one test file in fx102, but setting the
reader to null is done more widely, so maybe they just don't try to
delete ZIP files before GC happens and manage to avoid this bug.
Fixes#3369
