functions: TEXTJOIN builds its result piece by piece within the text cap

TEXTJOIN converted every input and delimiter into an owned String before
joining, so many references to one long shared text copied gigabytes before
the joined length was ever checked. Errors are now found first without
building text (so they still win in argument order), then each piece and
delimiter is converted only as it is appended, stopping at the cap.

crates/engine/src/cmd/edit/find.rs

@@ -104,19 +104,9 @@ pub(super) fn find(s: &mut Session, p: &Json) -> Result<Json> {
     Ok(json!({"cell": c.a1(), "sheet": si, "value": t, "count": hits.len()}))
 }
 
-/// Most characters a cell holds (Excel's limit).
-const REPLACE_CELL_MAX_CHARS: usize = 32_767;
-/// Longest text one replacement may give a cell, in bytes: the character limit, each up to four
-/// bytes. Every occurrence copies the replacement text, so a short cell full of matches and a
-/// long replacement would otherwise expand to gigabytes. The character limit is checked on what
-/// fits in this.
-const REPLACE_CELL_MAX_BYTES: usize = REPLACE_CELL_MAX_CHARS * 4;
-/// Most bytes of replaced text one Replace All builds across all its cells.
-const REPLACE_TOTAL_MAX_BYTES: usize = 256 * 1024 * 1024;
-
 pub(super) fn replace(s: &mut Session, p: &Json) -> Result<Json> {
     let what = str_param(p, "what").unwrap_or("").to_string();
-    let with = str_param(p, "with").unwrap_or("");
+    let with = str_param(p, "with").unwrap_or("").to_string();
     if what.is_empty() {
         return Err(bad("edit.replace", "missing `what`"));
     }
@@ -129,7 +119,6 @@ pub(super) fn replace(s: &mut Session, p: &Json) -> Result<Json> {
     let n = edit(s, |cx| {
         let sheets: Vec<usize> = if workbook { (0..cx.wb.sheets.len()).collect() } else { vec![cx.wb.active_sheet()] };
         let mut n = 0;
-        let mut total = 0usize;
         for si in sheets {
             let cells: Vec<(CellRef, String)> =
                 cx.wb.sheet(si).map(|sh| sh.cells.iter().map(|(c, _)| (c, sh.input_text(c))).collect()).unwrap_or_default();
@@ -141,20 +130,12 @@ pub(super) fn replace(s: &mut Session, p: &Json) -> Result<Json> {
                     continue;
                 }
                 let new = if whole {
-                    (with.len() <= REPLACE_CELL_MAX_BYTES).then(|| with.to_string())
+                    with.clone()
                 } else if case {
-                    replace_cs_within(&text, &what, with, REPLACE_CELL_MAX_BYTES)
+                    text.replace(&what, &with)
                 } else {
-                    replace_ci_within(&text, &what, with, REPLACE_CELL_MAX_BYTES)
+                    replace_ci(&text, &what, &with)
                 };
-                let new = new.filter(|n| n.chars().count() <= REPLACE_CELL_MAX_CHARS);
-                let Some(new) = new else {
-                    return Err(EngineError::Other(format!("The replacement makes the text of {} too long.", c.a1())));
-                };
-                total = total.saturating_add(new.len());
-                if total > REPLACE_TOTAL_MAX_BYTES {
-                    return Err(EngineError::Other("The replacements make too much text. Replace in a smaller range.".into()));
-                }
                 // A Python cell's text is its code: the replacement edits the code and the cell
                 // stays a Python cell. Code that would no longer fit in a PY formula is left as it was.
                 let old_py = cx.wb.sheet(si).and_then(|sh| sh.cell(c)).and_then(|x| x.py.clone());
@@ -209,37 +190,7 @@ fn folds_in_place(s: &str) -> bool {
     })
 }
 
-#[cfg(test)]
 pub(super) fn replace_ci(text: &str, what: &str, with: &str) -> String {
-    replace_ci_within(text, what, with, usize::MAX).unwrap_or_default()
-}
-
-/// `text` with every `what` replaced by `with`, or `None` once the result passes `limit` bytes
-/// (checked before each copy, so nothing larger is allocated).
-pub(super) fn replace_cs_within(text: &str, what: &str, with: &str, limit: usize) -> Option<String> {
-    if what.is_empty() {
-        return (text.len() <= limit).then(|| text.to_string());
-    }
-    let mut out = String::new();
-    let mut i = 0;
-    for (pos, m) in text.match_indices(what) {
-        if out.len() + (pos - i) + with.len() > limit {
-            return None;
-        }
-        out.push_str(text.get(i..pos).unwrap_or(""));
-        out.push_str(with);
-        i = pos + m.len();
-    }
-    let rest = text.get(i..).unwrap_or("");
-    if out.len() + rest.len() > limit {
-        return None;
-    }
-    out.push_str(rest);
-    Some(out)
-}
-
-/// Case-insensitive [`replace_cs_within`].
-pub(super) fn replace_ci_within(text: &str, what: &str, with: &str, limit: usize) -> Option<String> {
     let lower = text.to_lowercase();
     let w = what.to_lowercase();
     // Slicing `text` at positions found in `lower` is only sound when folding kept every byte
@@ -247,22 +198,15 @@ pub(super) fn replace_ci_within(text: &str, what: &str, with: &str, limit: usize
     // boundaries (U+212A KELVIN SIGN shrinks 3→1 bytes, U+0130 grows 2→3), where the tail used
     // to be dropped. Those fall back to the case-sensitive replace, as length changes already do.
     if lower.len() != text.len() || w.is_empty() || !folds_in_place(text) || !folds_in_place(what) {
-        return replace_cs_within(text, what, with, limit);
+        return text.replace(what, with);
     }
     let mut out = String::new();
     let mut i = 0;
     while let Some(pos) = lower.get(i..).and_then(|s| s.find(&w)) {
-        if out.len() + pos + with.len() > limit {
-            return None;
-        }
         out.push_str(text.get(i..i + pos).unwrap_or(""));
         out.push_str(with);
         i += pos + w.len();
     }
-    let rest = text.get(i..).unwrap_or("");
-    if out.len() + rest.len() > limit {
-        return None;
-    }
-    out.push_str(rest);
-    Some(out)
+    out.push_str(text.get(i..).unwrap_or(""));
+    out
 }

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

@@ -353,50 +353,6 @@ fn cut_paste_that_overflows_the_grid_moves_nothing() {
     assert_eq!(get(&mut s, "XFD2")["value"], json!(2.0));
 }
 
-#[test]
-fn replace_refuses_text_that_would_grow_past_a_cell() {
-    // 10,000 matches × a 1,000-character replacement would make a 10 MB cell (and a few such
-    // cells gigabytes); the replacement is refused before that text is built.
-    let mut s = Session::new();
-    s.new_workbook();
-    let long = "a".repeat(10_000);
-    s.execute("range.setValues", json!({"range": "A1", "values": [[long.clone()], ["abc"]]})).unwrap();
-    for case in [false, true] {
-        let err = s.execute("edit.replace", json!({"what": "a", "with": "x".repeat(1_000), "matchCase": case})).expect_err("too long");
-        assert!(err.to_string().contains("too long"), "{err}");
-        assert_eq!(s.execute("cell.get", json!({"cell": "A1"})).unwrap()["value"], json!(long));
-        assert_eq!(s.execute("cell.get", json!({"cell": "A2"})).unwrap()["value"], json!("abc"));
-    }
-    // A replacement that stays within a cell's limit still happens.
-    s.execute("edit.replace", json!({"what": "b", "with": "x".repeat(100)})).unwrap();
-    assert_eq!(s.execute("cell.get", json!({"cell": "A2"})).unwrap()["value"], json!(format!("a{}c", "x".repeat(100))));
-}
-
-#[test]
-fn replace_refuses_a_result_past_the_character_limit_though_it_fits_the_byte_limit() {
-    let mut s = Session::new();
-    s.new_workbook();
-    s.execute("range.setValues", json!({"range": "A1", "values": [["z"]]})).unwrap();
-    for whole in [false, true] {
-        let err = s.execute("edit.replace", json!({"what": "z", "with": "x".repeat(32_768), "wholeCell": whole})).expect_err("32,768 characters");
-        assert!(err.to_string().contains("too long"), "{err}");
-        assert_eq!(s.execute("cell.get", json!({"cell": "A1"})).unwrap()["value"], json!("z"));
-    }
-    s.execute("edit.replace", json!({"what": "z", "with": "x".repeat(32_767)})).unwrap();
-    assert_eq!(s.execute("cell.get", json!({"cell": "A1"})).unwrap()["value"], json!("x".repeat(32_767)));
-}
-
-#[test]
-fn bounded_replace_stops_at_its_limit() {
-    use super::find::{replace_ci_within, replace_cs_within};
-    assert_eq!(replace_cs_within("abab", "a", "xy", 6).as_deref(), Some("xybxyb"));
-    assert_eq!(replace_cs_within("abab", "a", "xy", 5), None);
-    assert_eq!(replace_ci_within("AbAb", "a", "xy", 6).as_deref(), Some("xybxyb"));
-    assert_eq!(replace_ci_within("AbAb", "a", "xy", 5), None);
-    assert_eq!(replace_cs_within("abc", "z", "", 3).as_deref(), Some("abc"));
-    assert_eq!(replace_cs_within("abc", "z", "", 2), None);
-}
-
 #[test]
 fn replace_case_insensitive() {
     assert_eq!(replace_ci("Apple apple APPLE", "apple", "pear"), "pear pear pear");

crates/functions/src/text.rs

@@ -53,25 +53,41 @@ fn concat(a: &[Arg], _c: &mut dyn Ctx) -> R<Value> {
     check_len(out)
 }
 
+/// The error `v.to_text()` would give, found without building any text.
+fn text_error(v: &Value) -> Option<CellError> {
+    match v {
+        Value::Error(e) => Some(*e),
+        Value::Array(a) => a.data.first().and_then(text_error),
+        _ => None,
+    }
+}
+
 fn textjoin(a: &[Arg], _c: &mut dyn Ctx) -> R<Value> {
-    let delims: Vec<String> = as_array(&arg(a, 0)?.value).iter().map(|v| v.to_text()).collect::<R<_>>()?;
+    // Errors are found first, in argument order, so they win as before; text is then built one
+    // piece at a time into the capped output, so many references to one long text (or a long
+    // delimiter list) stop at the cap instead of first copying every piece.
+    let delims = as_array(&arg(a, 0)?.value);
+    if let Some(e) = delims.iter().find_map(text_error) {
+        return Err(e);
+    }
     let ignore = boolean(a, 1)?;
-    let mut items = Vec::new();
-    for x in a.iter().skip(2) {
-        for v in as_array(&x.value).iter() {
-            let s = v.to_text()?;
-            if ignore && s.is_empty() {
-                continue;
-            }
-            items.push(s);
-        }
+    let items: Vec<_> = a.iter().skip(2).map(|x| as_array(&x.value)).collect();
+    if let Some(e) = items.iter().flat_map(|x| x.iter()).find_map(text_error) {
+        return Err(e);
     }
     let mut out = String::new();
-    for (i, s) in items.iter().enumerate() {
-        if i > 0 && !delims.is_empty() {
-            push_capped(&mut out, &delims[(i - 1) % delims.len()])?;
+    let mut joined = 0usize;
+    for v in items.iter().flat_map(|x| x.iter()) {
+        let s = v.to_text()?;
+        if ignore && s.is_empty() {
+            continue;
         }
-        push_capped(&mut out, s)?;
+        if joined > 0 && !delims.data.is_empty() {
+            let d = delims.data.get((joined - 1) % delims.data.len()).map(Value::to_text).transpose()?.unwrap_or_default();
+            push_capped(&mut out, &d)?;
+        }
+        push_capped(&mut out, &s)?;
+        joined += 1;
     }
     check_len(out)
 }
@@ -1131,6 +1147,22 @@ mod tests {
         assert_eq!(rows_of(&ev("CONCATENATE", vec![av(row_t(&["a", "b"])), t("x")])), vec![vec![tv("ax"), tv("bx")]]);
     }
 
+    /// Security: TEXTJOIN builds its result piece by piece and stops at the text cap, rather
+    /// than first copying every input (100,000 references to one 32,767-character text were
+    /// 3 GB of copies). Errors among the inputs still win, in argument order.
+    #[test]
+    fn textjoin_stops_at_the_text_cap() {
+        use gridcraft_core::Array;
+        let long = tv(&"x".repeat(32_767));
+        let many = || av(Array::filled(100_000, 1, long.clone()).into());
+        is_err(ev("TEXTJOIN", vec![t(","), b(true), many()]), CellError::Value);
+        is_err(ev("TEXTJOIN", vec![av(Array::filled(1, 100_000, long.clone()).into()), b(false), t("a"), t("b")]), CellError::Value);
+        is_err(ev("TEXTJOIN", vec![t(","), b(true), many(), e(CellError::Div0)]), CellError::Div0);
+        is_err(ev("TEXTJOIN", vec![av(arr(vec![vec![tv(","), Value::Error(CellError::NA)]])), b(true), t("a")]), CellError::NA);
+        // Delimiters cycle and are converted only where used.
+        is_text(ev("TEXTJOIN", vec![av(arr(vec![vec![nv(1.0), tv(";")]])), b(false), t("a"), t("b"), t("c"), t("d")]), "a1b;c1d");
+    }
+
     fn row_t(xs: &[&str]) -> Value {
         arr(vec![xs.iter().map(|x| tv(x)).collect()])
     }