address review: keep modern-comments precedence for already-owned parts

The shaper lets one call grow to 256 glyphs per input character, as an
embedded font's substitutions decide, and every glyph was then copied out.
Items longer than 1 KiB are now shaped in pieces (cut at a space where one
is near, never before a combining mark or after a joiner), each with the
rest of the paragraph as context and the item's script, and a piece that
expands past 32 glyphs per character falls back to one glyph per character
from the cmap.

crates/fonts/src/shape.rs

@@ -37,56 +37,146 @@ pub struct ShapeParams<'a> {
     pub language: Option<&'a str>,
 }
 
+/// Most bytes of text given to the shaper at once. The shaper lets one call grow its buffers to
+/// 256 glyphs per character (a font's substitutions decide how many), so a long item is shaped in
+/// pieces of this size, each with the rest of the paragraph as context.
+const MAX_SHAPE_CHUNK: usize = 1024;
+/// Most glyphs kept per character of a piece: a piece the font expands further is laid out one
+/// glyph per character from the cmap instead.
+const MAX_GLYPHS_PER_CHAR: usize = 32;
+
 /// Shape `para[range]` with `face`, the rest of `para` as context. Glyphs are in visual order
 /// (left to right) and clusters are byte offsets into `para`. `map` substitutes characters (All
 /// Caps) without moving clusters. An empty or out-of-bounds range shapes nothing.
 pub fn shape_range(face: &FontFace, para: &str, range: Range<usize>, params: &ShapeParams, map: impl Fn(char) -> char) -> Vec<ShapedGlyph> {
-    let (Some(pre), Some(text), Some(post)) = (para.get(..range.start), para.get(range.clone()), para.get(range.end..)) else {
+    let (Some(_), Some(text), Some(_)) = (para.get(..range.start), para.get(range.clone()), para.get(range.end..)) else {
         return Vec::new();
     };
-    let mut out = Vec::with_capacity(text.len());
+    let chunks = chunks(text, MAX_SHAPE_CHUNK);
+    let mut pieces: Vec<Vec<ShapedGlyph>> = Vec::with_capacity(chunks.len());
     let shaped = face.hb().map(|hb| {
         let shaper = face.shaper.shaper(&hb).instance(face.instance.as_ref()).build();
-        let mut buf = UnicodeBuffer::new();
-        buf.set_pre_context(pre);
-        buf.set_post_context(post);
-        for (i, c) in text.char_indices() {
-            buf.add(map(c), (range.start + i) as u32);
-        }
-        buf.guess_segment_properties();
-        if let Some(s) = params.script.and_then(|t| Script::from_iso15924_tag(Tag::new(&t))) {
-            buf.set_script(s);
-        }
-        if let Some(l) = params.language.or_else(|| params.script.and_then(crate::script::default_language)).and_then(|l| l.parse::<Language>().ok())
-        {
-            buf.set_language(l);
-        }
-        buf.set_direction(if params.rtl { Direction::RightToLeft } else { Direction::LeftToRight });
-        let gb = shaper.shape(buf, ShapeOptions::new().features(params.features));
-        for (info, pos) in gb.glyph_infos().iter().zip(gb.glyph_positions()) {
-            out.push(ShapedGlyph {
-                gid: info.glyph_id,
-                cluster: info.cluster as usize,
-                x_advance: pos.x_advance,
-                x_offset: pos.x_offset,
-                y_offset: pos.y_offset,
-            });
+        // Pieces of one item share the script the whole item would be guessed to have.
+        let item_script = (chunks.len() > 1 && params.script.is_none()).then(|| {
+            let mut probe = UnicodeBuffer::new();
+            for c in text.chars() {
+                probe.add(map(c), 0);
+            }
+            probe.guess_segment_properties();
+            probe.script()
+        });
+        for c in &chunks {
+            let r = (range.start + c.start)..(range.start + c.end);
+            let (Some(pre), Some(piece), Some(post)) = (para.get(..r.start), para.get(r.clone()), para.get(r.end..)) else { continue };
+            let mut buf = UnicodeBuffer::new();
+            buf.set_pre_context(pre);
+            buf.set_post_context(post);
+            for (i, c) in piece.char_indices() {
+                buf.add(map(c), (r.start + i) as u32);
+            }
+            buf.guess_segment_properties();
+            if let Some(s) = params.script.and_then(|t| Script::from_iso15924_tag(Tag::new(&t))) {
+                buf.set_script(s);
+            } else if let Some(s) = item_script {
+                buf.set_script(s);
+            }
+            if let Some(l) =
+                params.language.or_else(|| params.script.and_then(crate::script::default_language)).and_then(|l| l.parse::<Language>().ok())
+            {
+                buf.set_language(l);
+            }
+            buf.set_direction(if params.rtl { Direction::RightToLeft } else { Direction::LeftToRight });
+            let gb = shaper.shape(buf, ShapeOptions::new().features(params.features));
+            let budget = piece.chars().count().saturating_mul(MAX_GLYPHS_PER_CHAR);
+            if gb.len() > budget {
+                // Font-controlled expansion past anything real text needs: keep it bounded.
+                pieces.push(cmap_glyphs(face, para, r, params.rtl, &map));
+                continue;
+            }
+            pieces.push(
+                gb.glyph_infos()
+                    .iter()
+                    .zip(gb.glyph_positions())
+                    .map(|(info, pos)| ShapedGlyph {
+                        gid: info.glyph_id,
+                        cluster: info.cluster as usize,
+                        x_advance: pos.x_advance,
+                        x_offset: pos.x_offset,
+                        y_offset: pos.y_offset,
+                    })
+                    .collect(),
+            );
         }
     });
-    if shaped.is_none()
-        && let Some(f) = face.skrifa()
-    {
+    if shaped.is_none() {
         // No shaper for this face: one glyph per character from the cmap, in visual order.
-        let cmap = f.charmap();
-        let gm = f.glyph_metrics(Size::unscaled(), face.location());
-        for (i, c) in text.char_indices() {
-            let g = cmap.map(map(c)).unwrap_or_default();
-            let adv = gm.advance_width(g).unwrap_or(face.upem as f32 * 0.5);
-            out.push(ShapedGlyph { gid: g.to_u32(), cluster: range.start + i, x_advance: adv.round() as i32, x_offset: 0, y_offset: 0 });
-        }
-        if params.rtl {
-            out.reverse();
+        return cmap_glyphs(face, para, range, params.rtl, &map);
+    }
+    // Each piece is in visual order; right to left, the later pieces come first.
+    if params.rtl {
+        pieces.reverse();
+    }
+    let mut out = Vec::with_capacity(pieces.iter().map(Vec::len).sum());
+    for p in pieces {
+        out.extend(p);
+    }
+    out
+}
+
+/// One glyph per character of `para[range]` from the cmap, in visual order.
+fn cmap_glyphs(face: &FontFace, para: &str, range: Range<usize>, rtl: bool, map: &impl Fn(char) -> char) -> Vec<ShapedGlyph> {
+    let (Some(text), Some(f)) = (para.get(range.clone()), face.skrifa()) else { return Vec::new() };
+    let cmap = f.charmap();
+    let gm = f.glyph_metrics(Size::unscaled(), face.location());
+    let mut out = Vec::with_capacity(text.len());
+    for (i, c) in text.char_indices() {
+        let g = cmap.map(map(c)).unwrap_or_default();
+        let adv = gm.advance_width(g).unwrap_or(face.upem as f32 * 0.5);
+        out.push(ShapedGlyph { gid: g.to_u32(), cluster: range.start + i, x_advance: adv.round() as i32, x_offset: 0, y_offset: 0 });
+    }
+    if rtl {
+        out.reverse();
+    }
+    out
+}
+
+/// `text` cut into byte ranges of at most `max` bytes (or one character, if longer), at the last
+/// space of a piece where there is one, else before a character that starts a cluster (not a
+/// combining mark or joiner, not after a joiner). Text that fits is one piece.
+fn chunks(text: &str, max: usize) -> Vec<Range<usize>> {
+    use unicode_bidi::BidiClass::{BN, NSM};
+    let mut out = Vec::new();
+    let mut start = 0;
+    while text.len() - start > max {
+        let window = text.get(start..).unwrap_or_default();
+        let mut space = None;
+        let mut boundary = None;
+        let mut any = None;
+        let mut prev: Option<char> = None;
+        for (i, c) in window.char_indices() {
+            if i > max {
+                break;
+            }
+            if i > 0 {
+                any = Some(i);
+                let joined = prev.is_some_and(|p| matches!(unicode_bidi::bidi_class(p), BN));
+                if !joined && !matches!(unicode_bidi::bidi_class(c), NSM | BN) {
+                    boundary = Some(i);
+                    if prev.is_some_and(char::is_whitespace) {
+                        space = Some(i);
+                    }
+                }
+            }
+            prev = Some(c);
         }
+        // A piece always moves forward by at least one character.
+        let first = window.chars().next().map_or(window.len(), char::len_utf8);
+        let cut = space.filter(|&i| i >= max / 2).or(boundary).or(any).unwrap_or(first).max(first);
+        out.push(start..start + cut);
+        start += cut;
+    }
+    if start < text.len() || out.is_empty() {
+        out.push(start..text.len());
     }
     out
 }
@@ -121,4 +211,37 @@ mod tests {
         // Not a char boundary.
         assert!(shape_range(&face, "سنة", 1..2, &ShapeParams::default(), |c| c).is_empty());
     }
+
+    #[test]
+    fn long_items_are_shaped_in_bounded_pieces() {
+        for text in ["word ".repeat(2_000), "e\u{301}".repeat(2_000), "x".repeat(5_000), "\u{1f469}\u{200d}\u{1f4bb} ".repeat(500)] {
+            let pieces = chunks(&text, MAX_SHAPE_CHUNK);
+            assert_eq!(pieces.first().map(|r| r.start), Some(0));
+            assert_eq!(pieces.last().map(|r| r.end), Some(text.len()));
+            for w in pieces.windows(2) {
+                assert_eq!(w[0].end, w[1].start);
+            }
+            for r in &pieces {
+                assert!(!r.is_empty() && r.len() <= MAX_SHAPE_CHUNK);
+                // Never before a combining mark, never after a joiner.
+                let rest = &text[r.end..];
+                assert!(!rest.starts_with('\u{301}'));
+                assert!(!text[..r.end].ends_with('\u{200d}') || rest.is_empty());
+            }
+        }
+        assert_eq!(chunks("short", MAX_SHAPE_CHUNK), vec![0..5]);
+        assert_eq!(chunks("", MAX_SHAPE_CHUNK), vec![0..0]);
+    }
+
+    #[test]
+    fn long_items_keep_every_cluster_in_visual_order() {
+        let face = FontDb::global().face(DEFAULT_FAMILY, "Regular");
+        let text = "ab cd ".repeat(1_000);
+        let g = shape_range(&face, &text, 0..text.len(), &ShapeParams::default(), |c| c);
+        assert_eq!(g.len(), text.len());
+        assert!(g.windows(2).all(|w| w[0].cluster < w[1].cluster));
+        let rtl = shape_range(&face, &text, 0..text.len(), &ShapeParams { rtl: true, ..Default::default() }, |c| c);
+        assert_eq!(rtl.len(), text.len());
+        assert!(rtl.windows(2).all(|w| w[0].cluster > w[1].cluster));
+    }
 }

crates/pptx/src/read.rs

@@ -663,52 +663,55 @@ fn read_slides(
     let pkg = imp.pkg;
     let fallback_layout = p.masters.first().and_then(|m| m.layouts.first()).map(|l| l.id).unwrap_or_default();
     let mut file_slide_ids: HashMap<u32, SlideId> = HashMap::new();
-    let mut notes_by_target: HashMap<String, TextBody> = HashMap::new();
+    // A notes or comments part belongs to one slide; one shared by many must not be read (and
+    // copied) once per slide that names it.
+    let mut owned_parts: HashSet<String> = HashSet::new();
     for (file_id, path, id) in slide_ids {
         let Some(doc) = docs.get(&opc::norm(path)) else { continue };
         let part = Part::new(pkg, path, doc);
         let layout = part.rels.first_of("slideLayout").and_then(|r| layout_by_part.get(&opc::norm(&r.target)).copied()).unwrap_or(fallback_layout);
         let mut slide = read_slide(imp, &part, doc, *id, layout);
-        if let Some(notes) = slide_notes(imp, &part, &mut notes_by_target) {
+        if let Some(notes) = slide_notes(imp, &part, &mut owned_parts) {
             slide.notes = notes;
         }
-        read_slide_comments(pkg, pres_part, &part, &mut slide);
+        read_slide_comments(pkg, pres_part, &part, &mut slide, &mut owned_parts);
         file_slide_ids.insert(*file_id, *id);
         p.slides.push(Arc::new(slide));
     }
     file_slide_ids
 }
 
-/// The slide's notes text, read once per notes part.
-fn slide_notes(imp: &mut Imp, part: &Part, cache: &mut HashMap<String, TextBody>) -> Option<TextBody> {
+/// The slide's notes text. A notes part is the notes of the first slide that names it.
+fn slide_notes(imp: &mut Imp, part: &Part, owned: &mut HashSet<String>) -> Option<TextBody> {
     let n = part.rels.first_of("notesSlide")?;
-    let key = opc::norm(&n.target);
-    if let Some(cached) = cache.get(&key) {
-        return Some(cached.clone());
+    if !owned.insert(opc::norm(&n.target)) {
+        log::warn!("pptx: notes part {} is named by more than one slide; kept on the first", n.target);
+        return None;
     }
     let pkg = imp.pkg;
     let nd = load(pkg, &n.target)?;
     let np = Part::new(pkg, &n.target, &nd);
-    let notes = read_notes(imp, &np, &nd.root);
-    cache.insert(key, notes.clone());
-    Some(notes)
+    Some(read_notes(imp, &np, &nd.root))
 }
 
 /// Comments. A modern part and the legacy part can describe the same threads (PowerPoint
 /// writes both): the modern one wins, its flattened legacy copy is skipped.
-fn read_slide_comments(pkg: &Package, pres_part: &Part, part: &Part, slide: &mut Slide) {
-    let mut parts: Vec<(Doc, bool)> = vec![];
+/// Each comments part is read once, for the first slide that names it.
+fn read_slide_comments(pkg: &Package, pres_part: &Part, part: &Part, slide: &mut Slide, owned: &mut HashSet<String>) {
+    let mut parts: Vec<(Doc, bool, bool)> = vec![];
     for c in part.rels.all_of("comments") {
+        // An already-owned part is not re-read, but its kind still decides modern precedence.
+        let fresh = owned.insert(opc::norm(&c.target));
         if let Some(cd) = pkg.xml(&c.target) {
             // The content type names the modern part; a package that types it only through a
             // Default entry instead of an Override is still modern when its root is p188.
             let modern = pkg.content_type(&c.target).as_deref() == Some(crate::write::CT_COMMENTS_MODERN) || cd.root.prefix() == "p188";
-            parts.push((cd, modern));
+            parts.push((cd, modern, fresh));
         }
     }
-    let has_modern = parts.iter().any(|(_, m)| *m);
-    for (cd, modern) in parts {
-        if has_modern && !modern {
+    let has_modern = parts.iter().any(|(_, m, _)| *m);
+    for (cd, modern, fresh) in parts {
+        if !fresh || (has_modern && !modern) {
             continue;
         }
         slide.comments.extend(read_comments(pkg, pres_part, &cd.root, modern));

crates/pptx/tests/it/comments.rs

@@ -146,6 +146,33 @@ fn comments_on_several_slides_round_trip() {
     assert_eq!(q.slides[2].comments[0].text, "Slide three");
 }
 
+/// A slide that names a modern part another slide already owns, plus its own legacy part, must
+/// not import the legacy copy: modern precedence holds even when the modern part is not re-read.
+#[test]
+fn owned_modern_part_still_suppresses_later_legacy() {
+    let mut p = deck();
+    let layout = p.slides[0].layout;
+    let s2 = defaults::new_slide(&mut p, layout).expect("slide");
+    p.slides.push(Arc::new(s2));
+    Arc::make_mut(&mut p.slides[1]).comments = vec![Comment { author: "Cara".into(), text: "Slide two".into(), ..Default::default() }];
+    let bytes = deckcraft_pptx::export(&p).expect("export");
+    let rels2 = part(&bytes, "ppt/slides/_rels/slide2.xml.rels").replace("modernComment2.xml", "modernComment1.xml");
+    let mut z = zip::ZipArchive::new(std::io::Cursor::new(&bytes)).expect("zip");
+    let mut out = zip::ZipWriter::new(std::io::Cursor::new(Vec::new()));
+    for i in 0..z.len() {
+        let mut f = z.by_index(i).expect("entry");
+        out.start_file(f.name(), zip::write::SimpleFileOptions::default()).expect("start");
+        if f.name() == "ppt/slides/_rels/slide2.xml.rels" {
+            std::io::Write::write_all(&mut out, rels2.as_bytes()).expect("write");
+        } else {
+            std::io::copy(&mut f, &mut out).expect("copy");
+        }
+    }
+    let q = deckcraft_pptx::import(&out.finish().expect("finish").into_inner()).expect("import");
+    assert_eq!(q.slides[0].comments.len(), 2);
+    assert!(q.slides[1].comments.is_empty(), "{:?}", q.slides[1].comments);
+}
+
 /// A legacy-only package (what older tools read and write) still imports: 1/8 pt positions,
 /// flattened replies.
 #[test]

crates/pptx/tests/it/malformed.rs

@@ -287,3 +287,34 @@ fn unreadable_slide_is_reported() {
     assert_eq!(lost, vec!["ppt/slides/slide2.xml".to_string()]);
     assert_eq!(deckcraft_pptx::import(&broken).expect("import").slides.len(), 1);
 }
+
+/// Many slides naming one notes part: it is read for the first slide only, not copied onto every
+/// slide that names it.
+#[test]
+fn a_shared_notes_part_is_read_once() {
+    use std::sync::Arc;
+    let mut p = defaults::new_presentation(None);
+    let layout = p.slides[0].layout;
+    for _ in 0..3 {
+        let s = defaults::new_slide(&mut p, layout).expect("slide");
+        p.slides.push(Arc::new(s));
+    }
+    for (i, s) in p.slides.iter_mut().enumerate() {
+        Arc::make_mut(s).notes = deckcraft_model::text::TextBody::from_text(&format!("notes {i}"));
+    }
+    let mut es = entries(&deckcraft_pptx::export(&p).expect("export"));
+    for (n, d) in es.iter_mut() {
+        if n.starts_with("ppt/slides/_rels/") {
+            let x = String::from_utf8(d.clone()).expect("utf8");
+            let x = x
+                .replace("notesSlide2.xml", "notesSlide1.xml")
+                .replace("notesSlide3.xml", "notesSlide1.xml")
+                .replace("notesSlide4.xml", "notesSlide1.xml");
+            *d = x.into_bytes();
+        }
+    }
+    let q = deckcraft_pptx::import(&rezip(&es)).expect("import");
+    assert_eq!(q.slides.len(), 4);
+    assert_eq!(q.slides[0].notes.text(), "notes 0");
+    assert!(q.slides[1..].iter().all(|s| s.notes.text().is_empty()), "a shared notes part is not copied");
+}