After the Ngapi hackfest work, we were assuming that fonts
won't use presentation features to choose specific forms
(eg. conjuncts). As such, we were using auto-joiner behavior
for such features. It proved to be troublesome as many fonts
used presentation forms ('pres') for example to form conjuncts,
which need to be disabled when a ZWJ is inserted.
Two examples:
U+0D2F,U+200D,U+0D4D,U+0D2F with kartika.ttf
U+0995,U+09CD,U+200D,U+09B7 with vrinda.ttf
What we do now is to never do magic to ZWJ during GSUB's main input
match for Indic-style shapers. Note that backtrack/lookahead are still
matched liberally, as is GPOS. This seems to be an acceptable
compromise.
As to the bug that initially started this work, that one needs to
be fixed differently:
Bug 58714 - Kannada u+0cb0 u+200d u+0ccd u+0c95 u+0cbe does not
provide same results as Windows8
https://bugs.freedesktop.org/show_bug.cgi?id=58714
New numbers:
BENGALI: 353689 out of 354188 tests passed. 499 failed (0.140886%)
DEVANAGARI: 707305 out of 707394 tests passed. 89 failed (0.0125814%)
GUJARATI: 366349 out of 366457 tests passed. 108 failed (0.0294714%)
GURMUKHI: 60706 out of 60747 tests passed. 41 failed (0.067493%)
KANNADA: 951030 out of 951913 tests passed. 883 failed (0.0927606%)
KHMER: 299070 out of 299124 tests passed. 54 failed (0.0180527%)
LAO: 53611 out of 53644 tests passed. 33 failed (0.0615167%)
MALAYALAM: 1048102 out of 1048334 tests passed. 232 failed (0.0221304%)
ORIYA: 42320 out of 42329 tests passed. 9 failed (0.021262%)
SINHALA: 271666 out of 271847 tests passed. 181 failed (0.0665816%)
TAMIL: 1091753 out of 1091754 tests passed. 1 failed (9.15957e-05%)
TELUGU: 970555 out of 970573 tests passed. 18 failed (0.00185457%)
TIBETAN: 208469 out of 208469 tests passed. 0 failed (0%)
When a match_func was not set on the matcher_t object (ie. from GPOS),
then the Default_Ignorables (including joiners) were never skipped.
This meant that they were not skipped as they should during GPOS
matching. Fix that.
A few Indic numbers have "regressed": BENGALI and DEVANAGARI went
up from 290 and 58 respectively, but in both cases new results are
superior to Uniscribe, as they apply GPOS when we weren't (and
Uniscribe isn't) before.
BENGALI: 353686 out of 354188 tests passed. 502 failed (0.141733%)
DEVANAGARI: 707305 out of 707394 tests passed. 89 failed (0.0125814%)
GUJARATI: 366262 out of 366457 tests passed. 195 failed (0.0532122%)
GURMUKHI: 60706 out of 60747 tests passed. 41 failed (0.067493%)
KANNADA: 950680 out of 951913 tests passed. 1233 failed (0.129529%)
KHMER: 299074 out of 299124 tests passed. 50 failed (0.0167155%)
LAO: 53611 out of 53644 tests passed. 33 failed (0.0615167%)
MALAYALAM: 1047983 out of 1048334 tests passed. 351 failed (0.0334817%)
ORIYA: 42320 out of 42329 tests passed. 9 failed (0.021262%)
SINHALA: 271539 out of 271847 tests passed. 308 failed (0.113299%)
TAMIL: 1091753 out of 1091754 tests passed. 1 failed (9.15957e-05%)
TELUGU: 970555 out of 970573 tests passed. 18 failed (0.00185457%)
TIBETAN: 208469 out of 208469 tests passed. 0 failed (0%)
Originally we meant to match backtrack/lookahead across syllable
boundaries. But a bug in the code meant that this was NOT done for
backtrack. We "fixed" that in 2c7d0b6b80,
but that broke Myanmar shaping.
We now believe that for Indic-like shapers (which is where syllables are
used), all basic shaping forms should be fully contained within their
syllables, so now we limit backtrack/lookahead matching to the syllable
too. Unbreaks Myanmar.
Not for Arabic, but for Indic-like scripts. ZWJ/ZWNJ have special
meanings in those scripts, so let font lookups take full control.
This undoes the regression caused by automatic-joiners handling
introduced two commits ago.
We only disable automatic joiner handling for the "basic shaping
features" of Indic, Myanmar, and SEAsian shapers. The "presentation
forms" and other features are still applied with automatic-joiner
handling.
This change also changes the test suite failure statistics, such that
a few scripts show more "failures". The most affected is Kannada.
However, upon inspection, we believe that in most, if not all, of the
new failures, we are producing results superior to Uniscribe. Hard to
count those!
Here's an example of what is fixed by the recent joiner-handling
changes:
https://bugs.freedesktop.org/show_bug.cgi?id=58714
New numbers, for future reference:
BENGALI: 353892 out of 354188 tests passed. 296 failed (0.0835714%)
DEVANAGARI: 707336 out of 707394 tests passed. 58 failed (0.00819911%)
GUJARATI: 366262 out of 366457 tests passed. 195 failed (0.0532122%)
GURMUKHI: 60706 out of 60747 tests passed. 41 failed (0.067493%)
KANNADA: 950680 out of 951913 tests passed. 1233 failed (0.129529%)
KHMER: 299074 out of 299124 tests passed. 50 failed (0.0167155%)
LAO: 53611 out of 53644 tests passed. 33 failed (0.0615167%)
MALAYALAM: 1047983 out of 1048334 tests passed. 351 failed (0.0334817%)
ORIYA: 42320 out of 42329 tests passed. 9 failed (0.021262%)
SINHALA: 271539 out of 271847 tests passed. 308 failed (0.113299%)
TAMIL: 1091753 out of 1091754 tests passed. 1 failed (9.15957e-05%)
TELUGU: 970555 out of 970573 tests passed. 18 failed (0.00185457%)
TIBETAN: 208469 out of 208469 tests passed. 0 failed (0%)
When matching lookups, be smart about default-ignorable characters.
In particular:
Do nothing specific about ZWNJ, but for the other default-ignorables:
If the lookup in question uses the ignorable character in a sequence,
then match it as we used to do. However, if the sequence match will
fail because the default-ignorable blocked it, try skipping the
ignorable character and continue.
The most immediate thing it means is that if Lam-Alef forms a ligature,
then Lam-ZWJ-Alef will do to. Finally!
One exception: when matching for GPOS, or for backtrack/lookahead of
GSUB, we ignore ZWNJ too. That's the right thing to do.
It certainly is possible to build fonts that this feature will result
in undesirable glyphs, but it's hard to think of a real-world case
that that would happen.
This *does* break Indic shaping right now, since Indic Unicode has
specific rules for what ZWJ/ZWNJ mean, and skipping ZWJ is breaking
those rules. That will be fixed in upcoming commits.
Before, when matching ligatures, we never skipping over base / liga
glyphs even if that was what the LookupFlags asked for.
Fixed now. We carefully reviewed all instances of this, and tested with
Amiri as well as some Indic scripts, and are confident that this should
NOT break anyone's fonts. It's also how Uniscribe does it, from what
we can tell.
This will eventually allow us to skip marks, as well as (fallback)
attach marks to ligature components of fallback-shaped Arabic.
That would be pretty cool. I kludged GDEF props in, so mark-skipping
works, but the produced ligature id/components will be cleared later
by substitute_start() et al.
Perhaps using a synthetic table for Arabic fallback shaping was a better
idea. The current approach has way too many layering violations...
This reverts commit 24dd4e5674.
Oops. My bad. The change _regressed_ Malayalam test suite, not
improved it. I'll redo it, differentiating between old-spec and
new-spec cases.
The MS Indic specs say "...all classifications are determined ... using
context-free substitutions." However, testing shows that MS's Malayalam
shapers (both old and new), "match" even if there is no zero-context rule.
We follow.
Fixes below-base La (eg. Pa,H,La) with AnjaliNewLipi.ttf (old spec).
Moreover, test suite Malayalam failures are down to 312 from 875! No
change in other scripts.
Current numbers:
BENGALI: 353996 out of 354285 tests passed. 289 failed (0.0815727%)
DEVANAGARI: 707339 out of 707394 tests passed. 55 failed (0.00777502%)
GUJARATI: 366489 out of 366506 tests passed. 17 failed (0.0046384%)
GURMUKHI: 60769 out of 60809 tests passed. 40 failed (0.0657797%)
KANNADA: 951086 out of 951913 tests passed. 827 failed (0.0868777%)
KHMER: 299106 out of 299124 tests passed. 18 failed (0.00601757%)
LAO: 53611 out of 53644 tests passed. 33 failed (0.0615167%)
MALAYALAM: 1047541 out of 1048416 tests passed. 875 failed (0.0834592%)
ORIYA: 42320 out of 42329 tests passed. 9 failed (0.021262%)
SINHALA: 271726 out of 271847 tests passed. 121 failed (0.0445103%)
TAMIL: 1091837 out of 1091837 tests passed. 0 failed (0%)
TELUGU: 970558 out of 970573 tests passed. 15 failed (0.00154548%)
TIBETAN: 208469 out of 208469 tests passed. 0 failed (0%)
This reverts commit 0981068b75.
I was confused. Even if we access coverage[0] unconditionally, we don't
need bound checks since the array machinary already handles that.
We need the font for glyph lookup during GSUB pauses in Indic shaper.
Could perhaps be avoided, but at this point, we don't mean to support
separate substitute()/position() entry points (anymore), so there is
no point in not providing the font to GSUB.
Gives me a good 10% speedup for the Devanagari test case. Less so
for less lookup-intensive tests.
For the Devanagari test case, the false positive rate of the GSUB digest
is 4%.
And use it to speed up the hotspot by checking coverage directly in
the main loop, not 10 functions deep in.
Gives me a solid 20% boost with Indic test suite. Less so for less
lookup-intensive scenarios.
Remove the "fast_path" hack from before.
This does not apply to the context matchings.
This regresses tests right now. And we are not sure whether this is
the right thing to do for GPOS. But we'll figure out.
I couldn't measure significant performance gains out of this; maybe
about 5% (with one million Malayalam strings). Still, not bad.
But reminds me that optimizing this codebase without profiling first
is simply not going to work. Oh well...
I've messed up a lot of stuff recently, different parts of the
shaping process are stumbling on eachother's toes because
manually tracking what's in which buffer var is hard. I'm
going to add some internal API to track those such that mistakes
are discovered as soon as they are introduced.
Behdad Esfahbod