The three buttons act on the last color clicked (__color). After a change of reduction or color count, or through ui.dialog.set, that color can be one the table no longer shows, and a click added it to locked, transparent or webShift. The buttons are now enabled only when the table shows that color, the same check Delete Color uses.
Double-clicking one of several Shift-selected width points with the
Width tool now keeps the selection and opens Width Point Edit for all of
them; a double-click on an unselected point still edits that point
alone. ui.widthPointEdit takes {id, indices} for the same dialog, and
refuses an empty, malformed or out-of-range list, or indices given with
index or t. Either half of a discontinuous point brings the other.
Side 1, Side 2 and Total Width show a value only where every selected
point has it, and are blank otherwise. A side width typed in sets that
side of every point and keeps each point's other side (both sides with
the link on); Total Width scales each point's own two sides. With a
discontinuous point among them, Incoming and Outgoing edit its two
halves and both edit the continuous points. OK sets them all with one
stroke.widthProfile.set, one undo step, and nothing when no width
changed; Delete removes them all. With Adjust Adjoining Width Points,
each changed point changes the nearest unselected point before and
after it in proportion to its change since the dialog opened. OK and
Delete are refused when the points moved or changed since the dialog
opened.
Help page: https://helpx.adobe.com/illustrator/using/stroke-object.html
describes Shift-clicking to select several width points and setting
Side 1 and Side 2 for all of them in Width Point Edit. The blank fields,
keeping each point's other side, per-point Total Width and the adjoining
rule for several points are VectorCraft's own.
DocState::proxy_mixed, which decides whether the Fill and Stroke proxies
show "?", walked every node of the document to find the selected
objects, even with one object selected. The Tools panel proxies, the
paint chips and the task bar ask for it on every frame whose revision
changed, which includes each frame of a drag preview, and new art asks
for it after a selection change.
Document::nodes_in_order finds every node with one of the given ids in
tree order. When the id cache holds each id once at a path that still
checks out, it reads the nodes along those paths; otherwise, and for a
selection of an eighth of the document or more, it walks the tree once,
as before. Unlike Document::index_paths it returns every node with a
repeated id, so a damaged file that repeats an id keeps the question
marks it had. proxy_mixed now gathers the selected objects through it.
Tests compare the result with a copy of the old walk for one object,
different fills and strokes, a group, a group with one of its objects,
two layers, a hidden object, Appearance items, picked gradient mesh
points beside another object, ids not in the document and a repeated id
with different paints, with the cache built and dropped, and for 200
random selections. A cost test shows that with one object selected only
that object is read and the tree is not walked. Document tests compare
nodes_in_order with a walk.
The gain shows on large documents. On 50,000 rectangles in 500 groups
on 5 layers (release build on Linux, thread CPU time, median of seven
alternating runs of both versions in one process; an ignored test,
proxy_mixed_timing, times them):
Selection old new
1 object 995 us under 5 us
2 objects, different fills 810 us under 5 us
a group of 100 875 us 5 us
all 500 groups 3,615 us 2,575 us
all 50,000 objects 4,075 us 3,470 us
While Type > Find Font... is open, the dialog ran text.fonts on every
frame, and the missing-font walk on every frame where a folder picker is
available. Each walks all of the document's type and resolves its fonts.
The dialog now keeps both answers for the active document's uid, the
version of its content and the font database's generation, and works
them out again only when one of those changes. A new selection keeps
them. The list, its counts and statuses, and Find in Folder... are the
same as before.
With 2,000 point-type objects of 50 characters each, this took a median
1.53 ms of thread CPU time per frame in a release build on Linux; later
frames of one version now take about 10 ns. With 20 such objects it took
58 us per frame.
`token` on a string copied everything left in the string before scanning
one object, and `search` and `anchorsearch` copied both strings before
comparing them. A program that walks a string by what each call leaves
copied the rest of it on every call. All three now read the strings in
place and release the borrow before the scanned object is checked
against the budget and before the results are pushed. Results, errors
and the operations charged against the budget are the same as before.
Walking a 512 KiB string takes 16 ms instead of 518 ms with `token` and
1.0 ms instead of 27 ms with `search` (release build on Linux under load
from other builds, median of 7 alternating runs); on a 4 KiB string the
difference is under 0.1 ms.
Tests in tests_strings.rs compare each operator with the copying code on
the same operands (stack, error and operations left after every call),
check that the results see later writes to the string, and check that
no string is copied.