Avalonia

Branch: main

Commit Graph

Author	SHA1	Message	Date
Benedikt Stebner	883a2477a5	[Text] Rework text fallback itemization (#21631 ) * Add GlyphTypeface.CanShapeScript(Script), backed by a lazy read of the GSUB/GPOS ScriptList via the new ScriptListTable parser. Add FontFallbackScriptHints.TryGetComplexShapingTags, which maps 17 complex scripts to the OpenType script tag(s) a font must declare to shape them. Thread a Script shapingScript constraint through FontManager.TryMatchCharacter and FontCollectionBase.TryMatchCharacter as internal overloads, leaving the public signatures untouched. Rework TextCharacters.CreateShapeableRun into a capability preference: a complex script tries shaping-capable fonts first (tier 0), then cmap-only coverage (tier 1); simple scripts keep the single historical tier. Replace the NUL placeholder U+200B (ZERO WIDTH SPACE) with U+2060 (WORD JOINER) so the substitution introduces no line-break opportunity. Suppress previous-run font reuse across a culture change for locale-sensitive scripts, so a zh run's font no longer pins following ja text. Add 3 subset test fonts and 3 new test files (7 new tests) and fix 2 stale doc-comment crefs. * Fix comments * Fix font fallback reusing the wrong cached face from the script bucket FontCollectionBase.TryMatchCharacter caches a resolved fallback family per (script, culture) bucket. Two bugs let that bucket hand back a face that does not match the request: - A positive bucket entry suppressed the Tier D platform lookup for any later codepoint, even one the bucket family cannot cover. A CJK run left the Simplified-only U+534E unmatched (tofu) because the bucket resolved to a Japanese font that lacks it, while another installed font could place it. Tier D now treats only a negative (null) entry as "platform already attempted"; a positive entry no longer blocks the platform, since that family may not cover this codepoint. - A cached covering match of a neighbouring FontCollectionKey was reused for a different requested key. After a Bold run cached a face, a following Normal run reused it and rendered bold (and the symmetric style and stretch cases). Tier A, B and C now keep the exact key: when the cached covering match is a neighbouring key, the exact-key face of the same family is resolved from the platform via TryMatchCharacterFromPlatform, falling back to the neighbour only when the platform has no exact face. Collections without a platform are unchanged. Add regression tests for both: a bucket whose font lacks a later same-script codepoint, and weight and style leaks resolved through Tier B (script cache) and Tier C (the sweep, exercised via a non-locale-sensitive script with no culture).	3 months ago

Author

SHA1

Message

Date

Benedikt Stebner

883a2477a5

[Text] Rework text fallback itemization (#21631 )

* Add GlyphTypeface.CanShapeScript(Script), backed by a lazy read of the GSUB/GPOS ScriptList via the new ScriptListTable parser.
Add FontFallbackScriptHints.TryGetComplexShapingTags, which maps 17 complex scripts to the OpenType script tag(s) a font must declare to shape them.
Thread a Script shapingScript constraint through FontManager.TryMatchCharacter and FontCollectionBase.TryMatchCharacter as internal overloads, leaving the public signatures untouched.
Rework TextCharacters.CreateShapeableRun into a capability preference: a complex script tries shaping-capable fonts first (tier 0), then cmap-only coverage (tier 1); simple scripts keep the single historical tier.
Replace the NUL placeholder U+200B (ZERO WIDTH SPACE) with U+2060 (WORD JOINER) so the substitution introduces no line-break opportunity.
Suppress previous-run font reuse across a culture change for locale-sensitive scripts, so a zh run's font no longer pins following ja text.
Add 3 subset test fonts and 3 new test files (7 new tests) and fix 2 stale doc-comment crefs.

* Fix comments

* Fix font fallback reusing the wrong cached face from the script bucket

FontCollectionBase.TryMatchCharacter caches a resolved fallback family per
(script, culture) bucket. Two bugs let that bucket hand back a face that does
not match the request:

- A positive bucket entry suppressed the Tier D platform lookup for any later
codepoint, even one the bucket family cannot cover. A CJK run left the
Simplified-only U+534E unmatched (tofu) because the bucket resolved to a
Japanese font that lacks it, while another installed font could place it.
Tier D now treats only a negative (null) entry as "platform already
attempted"; a positive entry no longer blocks the platform, since that
family may not cover this codepoint.

- A cached covering match of a neighbouring FontCollectionKey was reused for a
different requested key. After a Bold run cached a face, a following Normal
run reused it and rendered bold (and the symmetric style and stretch cases).
Tier A, B and C now keep the exact key: when the cached covering match is a
neighbouring key, the exact-key face of the same family is resolved from the
platform via TryMatchCharacterFromPlatform, falling back to the neighbour
only when the platform has no exact face. Collections without a platform are
unchanged.

Add regression tests for both: a bucket whose font lacks a later same-script
codepoint, and weight and style leaks resolved through Tier B (script cache)
and Tier C (the sweep, exercised via a non-locale-sensitive script with no
culture).

3 months ago

1 Commits (main)