PDF export: wrap text with running widths

wrap_lines split an over-long word by re-measuring and copying the
rest of the word for every piece it cut, and only checked its line
limit once the word was done, so one long word in a narrow text box
cost word length x pieces. Each new word also re-measured the whole
line built so far.

Line and piece widths are now running sums of glyph widths (the same
sum text_width makes), so every character is measured once, and the
line limit is checked while a word is being split. The lines produced
are unchanged; a test compares them with the old wrapper.

apps/gridcraft-cli/src/check/package.rs

@@ -6,10 +6,6 @@ const MAX_XML_PART: u64 = 64 * 1024 * 1024;
 /// Most zip entries looked at while validating a written package.
 const MAX_ENTRIES: usize = 200_000;
 
-/// Decompressed bytes that comparing one package's relationship kinds may read in total: the
-/// relationship parts themselves and the targets whose content is checked.
-const RELATIONSHIP_BUDGET: u64 = 512 * 1024 * 1024;
-
 /// Relationship kinds (lower case) whose absence after a save isn't a loss: the calculation chain is
 /// a cache Excel rebuilds, and GridCraft never writes it.
 pub(super) const REBUILT_KINDS: [&str; 1] = ["calcchain"];
@@ -121,28 +117,6 @@ fn read_zip_part(zip: &mut Zip<'_>, name: &str) -> Result<Vec<u8>, String> {
     Ok(buf)
 }
 
-enum PartError {
-    /// Reading the part would spend more than the remaining budget.
-    OverBudget,
-    Unreadable,
-}
-
-/// Reads a part like [`read_zip_part`], charging the decompressed bytes to `budget`.
-fn read_zip_part_within(zip: &mut Zip<'_>, name: &str, budget: &mut u64) -> Result<Vec<u8>, PartError> {
-    let f = zip.by_name(name).map_err(|_| PartError::Unreadable)?;
-    // One byte past what is left shows that the part doesn't fit.
-    let limit = MAX_XML_PART.min(budget.saturating_add(1));
-    let mut buf = Vec::new();
-    f.take(limit).read_to_end(&mut buf).map_err(|_| PartError::Unreadable)?;
-    let read = buf.len() as u64;
-    if read > *budget {
-        *budget = 0;
-        return Err(PartError::OverBudget);
-    }
-    *budget -= read;
-    Ok(buf)
-}
-
 /// Whether the producer flagged its cached results as stale with `<calcPr fullCalcOnLoad="1">`:
 /// libraries that don't calculate write it, and store 0 or nothing as every formula's value.
 pub(super) fn full_calc_on_load(bytes: &[u8]) -> bool {
@@ -258,30 +232,19 @@ pub fn validate_package(bytes: &[u8]) -> Result<Vec<String>, String> {
 /// nothing are left out, as a save may rightly drop them: a link to a place in the workbook
 /// (`#Sheet1!A1`, written as a cell's `location`) and a target part with nothing in it.
 pub(super) fn relationship_kinds(bytes: &[u8]) -> Result<BTreeMap<String, String>, String> {
-    relationship_kinds_within(bytes, RELATIONSHIP_BUDGET)
-}
-
-/// [`relationship_kinds`] reading at most `budget` decompressed bytes. `Err` once the relationship
-/// parts alone need more: their kinds can't be listed completely.
-fn relationship_kinds_within(bytes: &[u8], mut budget: u64) -> Result<BTreeMap<String, String>, String> {
     let mut zip = zip::ZipArchive::new(Cursor::new(bytes)).map_err(|e| format!("not a zip package: {e}"))?;
     let names = zip_names(&mut zip);
     let by_key: HashMap<String, &String> = names.iter().map(|n| (part_key(n), n)).collect();
     // Several relationships name the same part; each fresh check decompressed it again, so
     // repeated references to one highly compressible part multiplied the work. The verdict is
     // remembered per (part, named content) and the decompression the package can cause is
-    // budgeted as a whole, the relationship parts included: many large, unreferenced rels parts
-    // would otherwise each spend a full per-part allowance.
+    // budgeted as a whole (finding: repeated decompression).
     let mut verdicts: HashMap<(String, &'static str), bool> = HashMap::new();
+    let mut budget: u64 = 512 * 1024 * 1024;
     let mut kinds = BTreeMap::new();
     for n in names.iter().filter(|n| part_key(n).ends_with(".rels")) {
         let Some(src_dir) = rels_source_dir(&part_key(n)) else { continue };
-        let xml = match read_zip_part_within(&mut zip, n, &mut budget) {
-            Ok(xml) => xml,
-            Err(PartError::OverBudget) => return Err(format!("{n}: relationship check stopped at its decompression limit")),
-            Err(PartError::Unreadable) => continue,
-        };
-        let Ok(els) = xml_elements(&xml) else { continue };
+        let Ok(els) = read_zip_part(&mut zip, n).and_then(|xml| xml_elements(&xml)) else { continue };
         for (_, attrs) in els.iter().filter(|(el, _)| el == "Relationship") {
             let Some(kind) = attr(attrs, "Type").and_then(|t| t.rsplit('/').next()) else { continue };
             let target = attr(attrs, "Target").unwrap_or("");
@@ -366,36 +329,3 @@ fn holds_something_uncached(zip: &mut Zip<'_>, name: &str, marker: &'static str,
         buf.clear();
     }
 }
-
-#[cfg(test)]
-mod tests {
-    use std::io::Write as _;
-
-    use super::*;
-
-    /// A package with `count` relationship parts of `size` bytes each, none of them describing an
-    /// existing part.
-    fn many_rels(count: usize, size: usize) -> Vec<u8> {
-        let mut z = zip::ZipWriter::new(Cursor::new(Vec::new()));
-        let opts = zip::write::SimpleFileOptions::default().compression_method(zip::CompressionMethod::Deflated);
-        let head = r#"<?xml version="1.0"?><Relationships xmlns="http://schemas.openxmlformats.org/package/2006/relationships"><Relationship Id="rId1" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/custom" Target="https://example.com" TargetMode="External"/>"#;
-        let tail = "</Relationships>";
-        let filler = " ".repeat(size.saturating_sub(head.len() + tail.len()));
-        for i in 0..count {
-            z.start_file(format!("xl/_rels/part{i}.xml.rels"), opts).unwrap();
-            z.write_all(head.as_bytes()).unwrap();
-            z.write_all(filler.as_bytes()).unwrap();
-            z.write_all(tail.as_bytes()).unwrap();
-        }
-        z.finish().unwrap().into_inner()
-    }
-
-    #[test]
-    fn relationship_parts_count_against_the_decompression_budget() {
-        // Each part fits the per-part allowance, but together they need more than the budget.
-        let bytes = many_rels(8, 64 * 1024);
-        assert!(relationship_kinds_within(&bytes, 256 * 1024).is_err());
-        let kinds = relationship_kinds_within(&bytes, 1024 * 1024).unwrap();
-        assert_eq!(kinds.get("custom").map(String::as_str), Some("custom"));
-    }
-}

crates/engine/src/cmd/print/tests.rs

@@ -695,6 +695,79 @@ fn bands_unit() {
     assert!(w.len() > 2);
 }
 
+/// The wrapper `wrap_lines` replaced, re-measuring strings, as the reference for its lines.
+fn wrap_lines_remeasuring(text: &str, font: Font, size: f32, width: f32) -> Vec<String> {
+    let mut out = Vec::new();
+    for para in text.split('\n') {
+        let mut line = String::new();
+        for word in para.split(' ') {
+            let cand = if line.is_empty() { word.to_string() } else { format!("{line} {word}") };
+            if line.is_empty() || gridcraft_pdf::text_width(font, size, &cand) <= width {
+                line = cand;
+            } else {
+                out.push(std::mem::take(&mut line));
+                line = word.to_string();
+            }
+            while gridcraft_pdf::text_width(font, size, &line) > width && line.chars().count() > 1 {
+                let chars: Vec<char> = line.chars().collect();
+                let (mut lo, mut hi) = (1usize, chars.len());
+                while lo < hi {
+                    let mid = lo + (hi - lo).div_ceil(2);
+                    let cut: String = chars[..mid].iter().collect();
+                    if gridcraft_pdf::text_width(font, size, &cut) <= width {
+                        lo = mid;
+                    } else {
+                        hi = mid - 1;
+                    }
+                }
+                if lo == chars.len() {
+                    break;
+                }
+                out.push(chars[..lo].iter().collect());
+                line = chars[lo..].iter().collect();
+            }
+            if out.len() > 2000 {
+                return out;
+            }
+        }
+        out.push(line);
+    }
+    out
+}
+
+#[test]
+fn wrap_lines_matches_the_remeasuring_wrapper() {
+    let texts = [
+        "the quick brown fox jumps",
+        "Supercalifragilistic",
+        "a  b   c\n\nverylongwordthatneedsbreaking short\nend ",
+        "Wide WWWWWWWWWW words, iiiiiiii narrow ones; mixed MMMMiiiiMMMM",
+        "€ ü ß — unicode Ünïcödé text",
+        "",
+        " ",
+    ];
+    for text in texts {
+        for font in [Font::Helvetica, Font::HelveticaBold, Font::Courier] {
+            for (size, width) in [(10.0, 50.0), (10.0, 20.0), (11.0, 1.0), (8.0, 300.0), (12.0, 0.0)] {
+                assert_eq!(wrap_lines(text, font, size, width), wrap_lines_remeasuring(text, font, size, width), "{text:?} {font:?} {size} {width}");
+            }
+        }
+    }
+}
+
+#[test]
+fn wrapping_a_long_word_in_a_narrow_box_stops_at_the_line_limit() {
+    // One character per line: the old wrapper re-measured and copied the rest of the word for
+    // each of the 100,000 lines before checking the limit.
+    let word = "w".repeat(100_000);
+    let lines = wrap_lines(&word, Font::Helvetica, 10.0, 1.0);
+    assert!(lines.len() <= 2002, "{}", lines.len());
+    assert!(lines.iter().all(|l| l == "w"));
+    // A long paragraph in a very wide box is one line.
+    let para = vec!["word"; 50_000].join(" ");
+    assert_eq!(wrap_lines(&para, Font::Helvetica, 10.0, 1.0e9), vec![para.clone()]);
+}
+
 #[test]
 fn center_across_spans_following_empty_center_across_cells() {
     let mut styles = StyleTable::default();

crates/engine/src/cmd/print/text.rs

@@ -3,42 +3,57 @@ use super::{CellRef, Color, Font, HAlign, MAX_COLS, Page, Point, RangeRef, Sheet
 
 const PAD_PX: f32 = 2.0;
 
+/// Most lines [`wrap_lines`] returns (it may return one more, the rest of a paragraph).
+const MAX_WRAP_LINES: usize = 2000;
+
 /// Word-wraps `text` to `width` points.
+///
+/// Widths are kept as running sums of glyph widths (the sum `text_width` makes), so each
+/// character is measured once: re-measuring the line for every word, or the rest of an
+/// over-long word for every piece cut from it, made a long paragraph or word quadratic.
 pub(super) fn wrap_lines(text: &str, font: Font, size: f32, width: f32) -> Vec<String> {
+    let units_of = |s: &str| -> u64 { s.chars().map(|c| u64::from(font.char_width(c))).sum() };
+    let fits = |units: u64| units as f32 * size / 1000.0 <= width;
+    let space = units_of(" ");
     let mut out = Vec::new();
     for para in text.split('\n') {
         let mut line = String::new();
+        let mut line_units: u64 = 0;
         for word in para.split(' ') {
-            let cand = if line.is_empty() { word.to_string() } else { format!("{line} {word}") };
-            if line.is_empty() || gridcraft_pdf::text_width(font, size, &cand) <= width {
-                line = cand;
+            let word_units = units_of(word);
+            if line.is_empty() {
+                line.push_str(word);
+                line_units = word_units;
+            } else if fits(line_units + space + word_units) {
+                line.push(' ');
+                line.push_str(word);
+                line_units += space + word_units;
             } else {
                 out.push(std::mem::take(&mut line));
-                line = word.to_string();
+                line.push_str(word);
+                line_units = word_units;
             }
-            // Break a single over-long word by characters. The cut that fits is found by binary
-            // search (each try measures one candidate prefix), so an over-long word costs
-            // O(n log n) measurements instead of re-measuring the whole cut per character for
-            // every emitted line (`a` × 100k pasted text would be cubic worst case).
-            while gridcraft_pdf::text_width(font, size, &line) > width && line.chars().count() > 1 {
-                let chars: Vec<char> = line.chars().collect();
-                let (mut lo, mut hi) = (1usize, chars.len());
-                while lo < hi {
-                    let mid = lo + (hi - lo).div_ceil(2);
-                    let cut: String = chars[..mid].iter().collect();
-                    if gridcraft_pdf::text_width(font, size, &cut) <= width {
-                        lo = mid;
-                    } else {
-                        hi = mid - 1;
+            // Break a single over-long word by characters: each piece is the longest that fits
+            // (at least one character).
+            if !fits(line_units) && line.chars().nth(1).is_some() {
+                let mut piece = String::new();
+                let mut units: u64 = 0;
+                for c in line.chars() {
+                    let w = u64::from(font.char_width(c));
+                    if !piece.is_empty() && !fits(units + w) {
+                        out.push(std::mem::take(&mut piece));
+                        units = 0;
+                        if out.len() > MAX_WRAP_LINES {
+                            return out;
+                        }
                     }
+                    piece.push(c);
+                    units += w;
                 }
-                if lo == chars.len() {
-                    break;
-                }
-                out.push(chars[..lo].iter().collect());
-                line = chars[lo..].iter().collect();
+                line = piece;
+                line_units = units;
             }
-            if out.len() > 2000 {
+            if out.len() > MAX_WRAP_LINES {
                 return out;
             }
         }