A document's composite fonts were lent by writing them into the font
database's shared map, and a second lend of another definition of the same
name was skipped while the first was held. A document drawn in a tiled or
floating window while the front document drew, a document placed in one
that defines the name differently, and an export made while another
document is in front could draw their type in the other document's
definition. Registering the front document's fonts during a render could
also change the definition mid-render.
Lending now applies on the lending thread only: each guard pushes the
document's definitions onto a per-thread stack that lookups read before the
database's own fonts, latest loan first, and removes them when it drops,
also when the thread unwinds. The guard is neither Send nor Sync. Lending no
longer changes the database or its generation; while a thread lends
definitions other than the database's, FontDb::generation gives a number of
its own for them, so kept layouts and render caches made in one document's
definitions are not used for another's.
The test's second and third wheel turns step to whichever families the
process has loaded at that moment. When another test has loaded a test
font without Latin glyphs, such as Alias Jointest Sans from the Glyphs
panel tests, the turn applies it, missing glyph protection keeps "Hi" in
Source Serif 4, and the assert_ne at the second turn fails. The test now
turns the preference off, as the font menu tests in tests/font_menu.rs
do, so each turn changes the text's font whatever family it reaches.
With downsampling on, each piece of an image bent by a Warp effect was
resampled for its own placed size, so a warp wrote one image XObject for
each size a piece was placed at (four for a 4 x 4 warp that stretches
across, eight for a colorized one with its soft masks). The pieces are
now resampled for the widest and the highest size any piece places the
image at, and all paint that one image. Pieces off the page count too,
and another placement of the same image keeps its own size.
Without downsampling, a colorized image and a CMYK image with a grid
mask no longer key their cache by placed size, so each is decoded and
encoded once for all the sizes it is placed at, as plain images are.
Break Text checked every thread of the document for each selected type
object to skip threaded frames, and Remove Breaks searched every thread
for each selected object to find its story, so the work grew with the
number of selected objects times the number of threaded frames.
Both commands now index the document's threads once per command: Break
Text tests membership in that index, and Remove Breaks takes the story
from the indexed thread through threads::story_in, the second half of
story_of. Split Text at Cursor, which takes one object, checks it once
as before. The results are unchanged.
Tests: Break Text on two threaded frames and six point type objects
skips the frames unchanged, breaks the rest in place as one undo step
and goes through the threads once; Remove Breaks reads the selected
story as one text and the other objects on their own, going through
the threads once; Split Text at Cursor refuses a threaded frame.
Offset Path took each selected path in turn, so offsetting many paths, or a
group of many paths, used one CPU. It now reads the selected paths first,
works out their offsets on as many threads as the computer has logical CPUs
(one thread on the web and for a single path), and then makes the new
objects one after another in the selection's order on the editing thread.
The new objects, their ids, their stacking, the selection, the messages and
the undo step are the same as before. When several paths fail, the error is
the one for the first of them in the selection's order, and the paths after
a failure that have not started yet are skipped.
A right-to-left line draws the text before a tab after it, but its width
was measured in the order of its text. A right-aligned line such as
"אב גד<tab>הו זח" with a stop at 150 pt in a 300 pt frame ended 5.59 pt
past the frame's right edge, and right, center and decimal stops aligned
the text before the tab instead of the text drawn after it.
- The line's width is measured in drawing order, each tab up to its stop
from where the glyphs drawn before it end. Right, Auto and centered
lines are aligned by that width. Stops are still measured from the
left edge.
- Right, center and decimal stops align the text drawn after the tab.
- Area type breaks a line with a tab where it fits as drawn, and wraps a
line that fits only in the order of its text. Once the line holds a
tab, it reads on while the least it can be drawn at fits, then ends at
the latest break that fits as drawn: up to 256 breaks are measured
from the latest back, and earlier breaks without a hyphen are
bisected. Hyphenation points are measured as drawn the same way, so
the line hyphenates at a point that fits.
- Only lines of horizontal paragraphs with right-to-left text keep a
list of their break candidates.
- A tab that ends a wrapped right-to-left line is drawn before the line
at its own width instead of pushing the line to the right.
Docs: docs/target-app-parity.md.
Type on a path whose first paragraph is aligned Justify All places its line
from the start bracket, the same as Left, so a threaded story is now measured
with the paragraphs it needs rather than the whole rest. A 1,500-paragraph
story through 40 paths aligned Justify All was laid out over 2,135,202 bytes
while measuring and is now laid out over 20,583. The frames take the same
text as before.
Center, Right, Auto right to left, hanging punctuation, Dynamic Text and
horizontal right-to-left text are still measured with the whole rest: a
centered or right-aligned line starts where the width of all its glyphs puts
it, and negative tracking can make an advance negative.
Tests compare every frame's share with the whole-rest share for Justify All
with spacing, on closed paths, with negative tracking and with punctuation
ending paragraphs.
Add Points and Divide into Dashes read each subpath's selected segments
with a range lookup in the path's ordered segment set instead of
filtering the whole set once per subpath. Add Points at Crossings and
Split at Crossings sort a path's crossing places into a list per
subpath once, in the order they come, instead of filtering the whole
list for every subpath. Places on a subpath that no longer exists are
still left out. The results are unchanged.
Tests cover paths with several subpaths, an empty subpath, an
untouched subpath and selection on the last subpath, for all four
commands, plus the anchor fallback and the grouping of out-of-order
crossing places.
An SVG shown as an image that could not be read (text that isn't XML,
an SVG without a size) still took one of the 256 nested SVG images an
SVG may show, so an image after 256 broken ones was left out. Only
images that decode now count toward that limit; a failed image gives
its place back.
Every nested SVG image read, decoded or not, counts toward new limits
of 1,024 reads and 512 MB of text per import, so a broken image shown
many times through <use> is still read a limited number of times. Past
them, nested SVG images are left out with a warning.
A nested SVG image with text that usvg then refuses no longer warns
that its text isn't imported.
open_menu_with waited once for the samples rendering and then drew its
last frame. A frame starts at most four samples, and none while eight
are rendering, so that last frame could start the samples of the rows
it showed instead of drawing them, and
show_samples_on_white_draws_dark_type_on_white_in_every_brightness
failed with "dark: samples are drawn" when the workers were slow.
wait_for_samples now returns whether any sample was rendering, and the
helper draws and waits until a frame leaves none rendering. Only test
code changes: wait_for_samples is cfg(test) and open_menu_with is a test
helper.
Diagnosis and approach by Taha-Mahmoodi in #7093.
Spot Color Objects listed every spot color linked from a gradient's stops
once any stop was above opacity 0. A spot stop at opacity 0 between other
clear stops, at the end of the ramp after a clear stop, or meeting an
opaque stop only at one offset was listed though it prints no ink, and so
was an opaque stop out of order whose pairs have no width.
A stop now counts when it bounds a stretch of some width that the
gradient sampler mixes and either stop of that stretch is above opacity
0, or when it is the first or last stop and above opacity 0, since those
color the offsets before and after the others. Stops in any order are
judged by the stretches the sampler uses. A freeform gradient keeps the
old rule, since it draws from its points. Type runs follow the same rule.
The document.info params text, docs/mcp.md and the progress line
describe it.
In a PDF without a standard, the warning for a page longer than 14,400 pt
written with a UserUnit above 1 now ends with "turn on Write pages longer
than 200 in at full size in points". The warning already said that apps
ignoring UserUnit show such a page smaller; it now names the Save PDF
option that writes the page at its size in points instead. PDF/A files,
which keep the UserUnit with the option on, and .ai saves, whose options
do not have it, keep the warning as it was.
The longer message has a row in every complete-message catalog. The
params text of document.exportPdf, docs/mcp.md and the evidence line in
docs/target-app-parity.md describe it.
BFlat