xlsx: find a reused picture through a content hash instead of comparing every media part

Out::image_part compared each new picture byte by byte with every
media part written so far, so a workbook of many distinct pictures
sharing a long prefix was exported in quadratic time. Media parts are
now indexed by extension and SHA-256.

crates/numfmt/src/lib.rs

@@ -332,14 +332,6 @@ impl NumberFormat {
             None => Formatted { text: t.to_string(), color: None, fill: None, numeric: false },
         }
     }
-
-    /// [`Self::format_text`]'s text, or `None` once it passes `max_units` UTF-16 code units.
-    fn format_text_within(&self, t: &str, max_units: usize) -> Option<String> {
-        match self.inner.text_section.and_then(|i| self.inner.sections.get(i)) {
-            Some(sec) => render::render_text_within(sec, t, max_units).map(|out| out.text),
-            None => (t.encode_utf16().count() <= max_units).then(|| t.to_string()),
-        }
-    }
 }
 
 /// Formats a cell value for display.
@@ -433,10 +425,6 @@ impl NumberFormat {
 /// The `TEXT()` worksheet function: formats a value with a format code. Numeric text is
 /// converted to a number first; other text and booleans (as `TRUE`/`FALSE`) go through the text
 /// section, or are returned as is.
-/// Longest text a cell holds, in UTF-16 code units (Excel's 32,767 characters, which XLSB also
-/// stores that way), and so the longest `TEXT` result.
-const MAX_CELL_TEXT_CHARS: usize = 32_767;
-
 pub fn text_function(v: &Value, code: &str, sys: DateSystem) -> Result<String, CellError> {
     if code.is_empty() {
         return Ok(String::new());
@@ -452,9 +440,7 @@ pub fn text_function(v: &Value, code: &str, sys: DateSystem) -> Result<String, C
         Value::Bool(b) => return Ok(fmt.format_text(if b { "TRUE" } else { "FALSE" }).text),
         Value::Text(t) => match gridcraft_core::parse::parse_number_text(&t) {
             Some(n) => n,
-            // Every `@` repeats the text: a result past a cell's limit is `#VALUE!`, so nested
-            // TEXT calls can't multiply it further.
-            None => return fmt.format_text_within(&t, MAX_CELL_TEXT_CHARS).ok_or(CellError::Value),
+            None => return Ok(fmt.format_text(&t).text),
         },
         Value::Array(_) => return Err(CellError::Value),
     };

crates/numfmt/src/render.rs

@@ -45,35 +45,6 @@ pub(crate) fn render_text(sec: &Section, text: &str) -> Out {
     out
 }
 
-/// [`render_text`], but `None` once the output passes `max_units` UTF-16 code units (a cell's
-/// limit), checked before each piece is appended, so an oversized result is never built.
-pub(crate) fn render_text_within(sec: &Section, text: &str, max_units: usize) -> Option<Out> {
-    let text_units = text.encode_utf16().count();
-    let mut out = Out::default();
-    let mut units = 0usize;
-    for t in &sec.toks {
-        match t {
-            Tok::At => {
-                units = units.saturating_add(text_units);
-                if units > max_units {
-                    return None;
-                }
-                out.push(text);
-            }
-            Tok::Fill(c) => out.fill(*c),
-            other => {
-                let literal = other.literal_text();
-                units = units.saturating_add(literal.encode_utf16().count());
-                if units > max_units {
-                    return None;
-                }
-                out.push(&literal);
-            }
-        }
-    }
-    Some(out)
-}
-
 /// Renders a number through a General section (literals around the General keyword).
 pub(crate) fn render_general(sec: &Section, v: f64, width: usize) -> Out {
     let mut out = Out::default();

crates/numfmt/src/tests.rs

@@ -396,13 +396,6 @@ fn text_fn() {
     assert_eq!(text_function(&Value::Number(1234.5), "$#,##0.00", S), Ok("$1,234.50".into()));
     assert_eq!(text_function(&Value::from("5"), "0.00", S), Ok("5.00".into()));
     assert_eq!(text_function(&Value::from("abc"), "0.00", S), Ok("abc".into()));
-    // Each `@` repeats the text: up to a cell's 32,767 characters, then #VALUE!.
-    let at = "@".repeat(255);
-    assert_eq!(text_function(&Value::from("x".repeat(128).as_str()), &at, S).map(|s| s.len()), Ok(128 * 255));
-    assert_eq!(text_function(&Value::from("x".repeat(32_000).as_str()), &at, S), Err(CellError::Value));
-    // The limit counts UTF-16 units: 16,384 emoji are 32,768 of them, one too many.
-    assert_eq!(text_function(&Value::from("\u{1F600}".repeat(16_383).as_str()), "@", S).map(|s| s.chars().count()), Ok(16_383));
-    assert_eq!(text_function(&Value::from("\u{1F600}".repeat(16_384).as_str()), "@", S), Err(CellError::Value));
     assert_eq!(text_function(&Value::Error(CellError::NA), "0", S), Err(CellError::NA));
     assert_eq!(text_function(&Value::Number(46302.0), "yyyy", S), Ok("2026".into()));
     assert_eq!(text_function(&Value::Number(0.5), "General", S), Ok("0.5".into()));

crates/xlsx/src/tests/shared_media.rs

@@ -85,3 +85,21 @@ fn different_pictures_with_the_same_length_stay_separate() {
     let back = read_xlsx(&bytes).unwrap().0;
     assert_eq!(back.sheets[0].images.iter().map(|i| i.data.to_vec()).collect::<Vec<_>>(), vec![a, b]);
 }
+
+/// Many distinct pictures that share a long prefix: each is matched against the earlier ones
+/// through a content hash, not by comparing bytes with every one of them.
+#[test]
+fn many_distinct_pictures_with_a_common_prefix_are_each_written_once() {
+    let pictures: Vec<Vec<u8>> = (0..400u32)
+        .map(|i| {
+            let mut v = b"\x89PNG\r\n\x1a\n".to_vec();
+            v.extend(std::iter::repeat_n(7u8, 65_536));
+            v.extend(i.to_le_bytes());
+            v
+        })
+        .collect();
+    let refs: Vec<(u32, &[u8])> = pictures.iter().enumerate().map(|(i, p)| (i as u32 + 1, p.as_slice())).collect();
+    let wb = Workbook { sheets: vec![sheet_with("One", &refs), sheet_with("Two", &refs)], ..Default::default() };
+    let bytes = write_xlsx(&wb).unwrap();
+    assert_eq!(names(&bytes).iter().filter(|n| n.starts_with("xl/media/")).count(), 400);
+}

crates/xlsx/src/write.rs

@@ -6,6 +6,7 @@ use std::fmt::Write as _;
 use gridcraft_core::RangeRef;
 use gridcraft_formula::quote_sheet;
 use gridcraft_model::{CalcMode, CalcSettings, PhoneticString, ShowObjects, Style, Visibility, Workbook, WorkbookView};
+use sha2::Digest as _;
 
 use crate::package::ZipOut;
 use crate::xml::{esc, esc_attr};
@@ -112,6 +113,9 @@ pub struct Out {
     pub dxfs: Vec<Style>,
     pub sst: Sst,
     pub images: u32,
+    /// Media parts written so far, by extension and SHA-256 of their bytes: a picture used again
+    /// is found here instead of being compared byte by byte with every earlier one.
+    media: HashMap<(String, [u8; 32]), String>,
     pub charts: u32,
     /// The workbook's file name, which a PivotChart's link to its PivotTable names in brackets.
     pub book_name: String,
@@ -142,15 +146,14 @@ impl Out {
     /// The `xl/media` file name (`imageN.ext`) holding `data`. A picture used on several sheets,
     /// or twice on one, is written once and every drawing points to that part, as in Excel.
     pub fn image_part(&mut self, ext: &str, data: &[u8]) -> String {
-        let existing = self.parts.iter().find(|(name, bytes)| {
-            name.strip_prefix("xl/media/image").is_some_and(|rest| rest.ends_with(&format!(".{ext}"))) && bytes.as_slice() == data
-        });
-        if let Some((name, _)) = existing {
-            return name.trim_start_matches("xl/media/").to_string();
+        let key = (ext.to_string(), sha2::Sha256::digest(data).into());
+        if let Some(file) = self.media.get(&key) {
+            return file.clone();
         }
         self.images += 1;
         let file = format!("image{}.{ext}", self.images);
         self.part(&format!("xl/media/{file}"), None, data.to_vec());
+        self.media.insert(key, file.clone());
         file
     }
     pub fn default_type(&mut self, ext: &str, ct: &str) {
@@ -249,6 +252,7 @@ pub fn write_xlsx_named(wb: &Workbook, kind: XlsxKind, book_name: Option<&str>)
         dxfs: vec![],
         sst: Sst::default(),
         images: 0,
+        media: HashMap::new(),
         charts: 0,
         book_name: book_name.filter(|n| !n.is_empty() && !n.contains(['[', ']'])).unwrap_or(DEFAULT_BOOK_NAME).to_string(),
         chart_names: vec![],