functions: stop WORKDAY, PIVOTBY, SORT and the D-functions rescanning their input

- WORKDAY/WORKDAY.INTL counted the holidays each step passed by scanning every
  holiday, once per step; a run of consecutive holidays made that quadratic.
  The working-day holidays are kept sorted and counted by binary search.
- PIVOTBY totals rescanned the whole source (rebuilding keys) once per total
  row and column. The grouping pass now also indexes rows by row key and by
  column key, and totals read from those indexes in the same order.
- SORT copied one key per sort_index entry for every row, so a long
  sort_index with repeated columns multiplied the source. Repeats can't break
  a tie their first use left, so only first uses are kept.
- DSUM and the other database functions tested every record against every
  criteria row. A selection whose records x conditions exceed 5e8 is #NUM!,
  and each criteria label's database column is found once, not per cell.

crates/engine/src/cf.rs

@@ -1,7 +1,7 @@
 //! Conditional formatting evaluation: what extra style, data bar, colour scale or icon a cell
 //! gets.
 
-use std::collections::{HashMap, HashSet};
+use std::collections::HashMap;
 
 use gridcraft_core::{CellRef, RangeRef, Value, compare};
 use gridcraft_model::{CfOperator, CfRule, CfValueKind, Color, CondFormat, PatternType, Sheet, Style, Underline, Workbook, default_icon_thresholds};
@@ -47,8 +47,6 @@ enum StatsNeed {
 fn stats_need(rule: &CfRule) -> Option<StatsNeed> {
     match rule {
         CfRule::Duplicate { .. } | CfRule::Unique { .. } => Some(StatsNeed::Counts),
-        // only two- and three-colour scales paint: no statistics for the others
-        CfRule::ColorScale { stops } if !matches!(stops.len(), 2 | 3) => None,
         CfRule::Top10 { .. } | CfRule::AboveAverage { .. } | CfRule::ColorScale { .. } | CfRule::DataBar { .. } | CfRule::IconSet { .. } => {
             Some(StatsNeed::Numbers)
         }
@@ -72,58 +70,15 @@ thread_local! {
     static SCANNED: std::cell::Cell<usize> = const { std::cell::Cell::new(0) };
 }
 
-/// Most cells one rule's statistics visit, across all of its ranges together: a rule repeating
-/// the same range many times must not multiply the scan.
-const STATS_SCAN_LIMIT: usize = 2_000_000;
-
-/// Steps through the spilled cells inside `r`, row-major: `Some(cell)` for each, and `None` for a
-/// jump past the spills left or right of the columns on a row (one step per row, however many
-/// there are), so unrelated spills between a narrow range's corners cost little.
-fn spill_steps(sh: &Sheet, r: RangeRef) -> impl Iterator<Item = Option<CellRef>> + '_ {
-    let mut cur = Some(r.start);
-    std::iter::from_fn(move || {
-        let at = cur.filter(|at| *at <= r.end)?;
-        let (&c, _) = sh.spill.range(at..=r.end).next()?;
-        if c.col < r.start.col {
-            cur = Some(CellRef::new(c.row, r.start.col));
-            Some(None)
-        } else if c.col > r.end.col {
-            cur = c.row.checked_add(1).map(|row| CellRef::new(row, r.start.col));
-            Some(None)
-        } else {
-            cur = c.col.checked_add(1).map(|col| CellRef::new(c.row, col));
-            Some(Some(c))
-        }
-    })
-}
-
 fn numbers_in(sh: &Sheet, ranges: &[RangeRef], need: StatsNeed) -> (Vec<f64>, HashMap<String, usize>) {
     let mut nums = Vec::new();
     let mut counts = HashMap::new();
-    let mut budget = STATS_SCAN_LIMIT;
-    let mut seen: HashSet<RangeRef> = HashSet::new();
-    let used = sh.used_range();
     for r in ranges {
-        if budget == 0 {
-            break;
-        }
-        let r = match used.and_then(|u| r.intersection(&u)) {
+        let r = match sh.used_range().and_then(|u| r.intersection(&u)) {
             Some(x) => x,
             None => continue,
         };
-        // The same range listed twice counts its cells once.
-        if !seen.insert(r) {
-            continue;
-        }
-        // Only stored and spilled cells hold values: blank cells of a sparse range cost nothing,
-        // so a value far down a huge range is still counted. Every cell looked at is charged.
-        let stored = sh.cells.iter_range(r).map(|(c, _)| (Some(c), true));
-        let spilled = spill_steps(sh, r).map(|c| (c, false));
-        for (c, is_stored) in stored.chain(spilled).take(budget) {
-            budget -= 1;
-            // a skip over spills outside the columns (charged, not read), or a spilled value
-            // under a stored cell, which was read with it
-            let Some(c) = c.filter(|c| is_stored || sh.cells.get(*c).is_none()) else { continue };
+        for c in r.iter().take(2_000_000) {
             #[cfg(test)]
             SCANNED.with(|n| n.set(n.get() + 1));
             let v = sh.value(c);
@@ -323,8 +278,7 @@ impl CfCache {
                     _ => None,
                 },
                 CfRule::ColorScale { stops } => {
-                    // Only two- and three-colour scales paint: evaluate no thresholds for others.
-                    if let Some(n) = v.as_f64().filter(|_| matches!(stops.len(), 2 | 3)) {
+                    if let Some(n) = v.as_f64() {
                         let pts: Vec<(f64, [u8; 3])> = stops
                             .iter()
                             .map(|(k, col)| (threshold(wb, sheet, c, st, k), col.resolve(&wb.theme).unwrap_or([255, 255, 255])))
@@ -606,89 +560,6 @@ mod tests {
         assert_eq!(SCANNED.with(|n| n.get()), 10_000);
     }
 
-    /// A rule listing the same range many times scans it once, and never more than the shared
-    /// limit however many ranges it holds.
-    #[test]
-    fn repeated_ranges_share_one_scan_budget() {
-        let mut s = crate::Session::new();
-        s.new_workbook();
-        s.execute("range.setValues", json!({"range": "A1", "values": [[1], [2], [3]]})).unwrap();
-        s.execute("range.setValues", json!({"range": "XFD1048576", "values": [[4]]})).unwrap();
-        let d = s.doc().unwrap();
-        let sh = d.wb.sheet(0).unwrap();
-        let dense = RangeRef::parse("A1:XFD1048576").unwrap();
-        let small = RangeRef::parse("A1:A3").unwrap();
-        SCANNED.with(|n| n.set(0));
-        let st = RuleStats::build(sh, &vec![small; 1000], StatsNeed::Numbers);
-        assert_eq!(st.sorted, vec![1.0, 2.0, 3.0]);
-        assert_eq!(SCANNED.with(|n| n.get()), 3);
-        let shifted: Vec<RangeRef> = (0..50).map(|i| RangeRef::new(CellRef::new(i, 0), dense.end)).collect();
-        SCANNED.with(|n| n.set(0));
-        RuleStats::build(sh, &shifted, StatsNeed::Numbers);
-        assert!(SCANNED.with(|n| n.get()) <= STATS_SCAN_LIMIT);
-    }
-
-    /// Blank cells of a sparse range are not charged to the scan budget: a value at the end of two
-    /// million-row columns still counts, so duplicates are found.
-    #[test]
-    fn sparse_ranges_keep_their_values_under_the_scan_budget() {
-        let mut s = crate::Session::new();
-        s.new_workbook();
-        s.execute("range.setValues", json!({"range": "A1", "values": [[7]]})).unwrap();
-        s.execute("range.setValues", json!({"range": "B1048576", "values": [[7]]})).unwrap();
-        let d = s.doc().unwrap();
-        let sh = d.wb.sheet(0).unwrap();
-        let ranges = [RangeRef::parse("A1:A1048576").unwrap(), RangeRef::parse("B1:B1048576").unwrap()];
-        SCANNED.with(|n| n.set(0));
-        let st = RuleStats::build(sh, &ranges, StatsNeed::Counts);
-        assert_eq!(st.counts.get("7"), Some(&2));
-        assert_eq!(SCANNED.with(|n| n.get()), 2);
-    }
-
-    /// Spills left and right of a narrow range, between its corners, are skipped a row at a time:
-    /// they do not use the budget that the in-range spills need.
-    #[test]
-    fn spills_outside_a_narrow_range_do_not_use_its_scan_budget() {
-        let mut s = crate::Session::new();
-        s.new_workbook();
-        let d = s.doc().unwrap();
-        let mut sh = d.wb.sheet(0).unwrap().clone();
-        for row in 0..1000 {
-            for col in [0u32, 5, 6, 7] {
-                sh.spill.insert(CellRef::new(row, col), Value::Number(1.0));
-            }
-        }
-        let r = RangeRef::parse("F1:G1000").unwrap();
-        let steps: Vec<_> = spill_steps(&sh, r).collect();
-        assert_eq!(steps.iter().flatten().count(), 2000);
-        // column A is skipped one step a row, never one per cell
-        assert!(steps.len() <= 2000 + 1000, "{}", steps.len());
-        let narrow = RangeRef::parse("C1:C1000").unwrap();
-        assert_eq!(spill_steps(&sh, narrow).flatten().count(), 0);
-    }
-
-    /// A colour scale with an unsupported number of stops paints nothing and evaluates none of them.
-    #[test]
-    fn colour_scales_with_other_stop_counts_paint_nothing() {
-        let mut s = crate::Session::new();
-        s.new_workbook();
-        s.execute("range.setValues", json!({"range": "A1", "values": [[1], [5]]})).unwrap();
-        let mut wb: Workbook = (*s.doc().unwrap().wb).clone();
-        let stop = (CfValueKind::Formula("=1".into()), Color::default());
-        let rule = |n: usize| CondFormat {
-            ranges: vec![RangeRef::parse("A1:A2").unwrap()],
-            rule: CfRule::ColorScale { stops: vec![stop.clone(); n] },
-            priority: 1,
-            stop_if_true: false,
-        };
-        wb.sheet_mut(0).unwrap().cond_formats.push(rule(4));
-        let mut cache = CfCache::new();
-        assert_eq!(cache.look(&wb, 0, CellRef::parse("A1").unwrap()).and_then(|l| l.scale_fill), None);
-        wb.sheet_mut(0).unwrap().cond_formats = vec![rule(2)];
-        let mut cache = CfCache::new();
-        assert!(cache.look(&wb, 0, CellRef::parse("A1").unwrap()).and_then(|l| l.scale_fill).is_some());
-    }
-
     #[test]
     fn cell_value_operator_names() {
         let mut s = crate::Session::new();

crates/functions/src/database.rs

@@ -70,27 +70,42 @@ fn cell_criterion(v: &Value) -> Option<Criterion> {
 /// when the label is unknown, criterion).
 type CriteriaRow = (usize, Vec<(usize, Option<usize>, Criterion)>);
 
+/// Most condition tests one selection may make (records times conditions): every record can be
+/// tested against every criteria row when none matches, so a modest database and criteria table
+/// multiply into billions of tests. Past it the function is `#NUM!`, as lifted lookups are; 5e8
+/// is a few seconds of matching.
+const SELECTION_WORK_BUDGET: usize = 500_000_000;
+
 /// Rows (1-based within the database, excluding the header) of matching records.
-fn matching_records(db: &Array, crit: &Array, ctx: &mut dyn Ctx) -> Vec<usize> {
+fn matching_records(db: &Array, crit: &Array, ctx: &mut dyn Ctx) -> R<Vec<usize>> {
+    // The database column under each criteria label, found once per label rather than per cell.
+    let columns: Vec<Option<usize>> = (0..crit.cols)
+        .map(|c| match crit.get(0, c) {
+            Some(Value::Text(t)) => find_column(db, t),
+            Some(Value::Number(n)) => find_column(db, &gridcraft_core::number_to_text(*n)),
+            _ => None,
+        })
+        .collect();
     // Per criteria row: its index and a list of (criteria column, database column or None when
     // the label is unknown, criterion).
     let mut rows: Vec<CriteriaRow> = Vec::new();
+    let mut tests = 0usize;
     for r in 1..crit.rows {
         let mut conds = Vec::new();
         for c in 0..crit.cols {
             let Some(cell) = crit.get(r, c) else { continue };
             let Some(cr) = cell_criterion(cell) else { continue };
-            let col = match crit.get(0, c) {
-                Some(Value::Text(t)) => find_column(db, t),
-                Some(Value::Number(n)) => find_column(db, &gridcraft_core::number_to_text(*n)),
-                _ => None,
-            };
-            conds.push((c, col, cr));
+            conds.push((c, columns.get(c).copied().flatten(), cr));
         }
+        tests = tests.saturating_add(conds.len().max(1));
         rows.push((r, conds));
     }
+    // Computed criteria (a formula evaluated per record) count like any other test.
+    if db.rows.saturating_sub(1).saturating_mul(tests) > SELECTION_WORK_BUDGET {
+        return Err(CellError::Num);
+    }
     let all_match = rows.is_empty();
-    (1..db.rows)
+    Ok((1..db.rows)
         .filter(|&rec| {
             all_match
                 || rows.iter().any(|(r, conds)| {
@@ -102,7 +117,7 @@ fn matching_records(db: &Array, crit: &Array, ctx: &mut dyn Ctx) -> Vec<usize> {
                     })
                 })
         })
-        .collect()
+        .collect())
 }
 
 struct Selection<'a> {
@@ -123,7 +138,7 @@ fn select<'a>(a: &'a [Arg], ctx: &mut dyn Ctx) -> R<Selection<'a>> {
     let db = as_array(dbv);
     let crit = as_array(critv);
     let field = field_index(&db, a)?;
-    let records = matching_records(&db, &crit, ctx);
+    let records = matching_records(&db, &crit, ctx)?;
     Ok(Selection { db, field, records })
 }
 
@@ -352,6 +367,24 @@ mod tests {
         is_err(d("DSUM", n(9.0)), CellError::Value);
     }
 
+    /// Security: records times criteria conditions past the selection budget is `#NUM!` before
+    /// any record is tested; a nonmatching criteria table used to test every record against
+    /// every criteria row.
+    #[test]
+    fn selection_work_is_bounded() {
+        use gridcraft_core::Array;
+        let column = |label: &str, len: usize, v: Value| {
+            let mut data = vec![tv(label)];
+            data.extend(std::iter::repeat_n(v, len));
+            Value::from(Array::column(data))
+        };
+        let db = column("Tree", 100_000, tv("Apple"));
+        let wide = column("Tree", 5_001, tv("=Plum"));
+        is_err(ev("DCOUNTA", vec![rf(db.clone()), t("Tree"), rf(wide)]), CellError::Num);
+        let narrow = column("Tree", 2, tv("=Apple"));
+        close(ev("DCOUNTA", vec![rf(db), t("Tree"), rf(narrow)]), 100_000.0);
+    }
+
     #[test]
     fn simple_criteria() {
         let c = |cells: Vec<Value>| arr(vec![vec![tv("Tree"), tv("Height")], cells]);

crates/functions/src/date.rs

@@ -416,13 +416,17 @@ fn workday_core(a: &[Arg], c: &mut dyn Ctx, mask: [bool; 7], hol: Option<&Arg>)
     if days == 0 {
         return num_val(start as f64);
     }
+    // The holidays on working days, still sorted (`holidays` sorts them), so each step counts
+    // the ones it passed with two binary searches rather than a scan of every holiday: a run of
+    // consecutive holidays takes one step per holiday.
+    let hs: Vec<i64> = hs.into_iter().filter(|&h| !mask[weekday0(sys, h)]).collect();
     let dir = days.signum();
     let mut cur = start;
     let mut rem = days.abs();
     while rem > 0 {
         let next = advance(sys, cur, rem, dir, &mask, per_week);
         let (lo, hi) = if dir > 0 { (cur + 1, next) } else { (next, cur - 1) };
-        rem = hs.iter().filter(|&&h| h >= lo && h <= hi && !mask[weekday0(sys, h)]).count() as i64;
+        rem = hs.partition_point(|&h| h <= hi).saturating_sub(hs.partition_point(|&h| h < lo)) as i64;
         cur = next;
         if !(0..=MAX_SERIAL as i64).contains(&cur) {
             return Err(CellError::Num);
@@ -755,6 +759,23 @@ mod tests {
         is_err(ev("TIMEVALUE", vec![t("x")]), CellError::Value);
     }
 
+    /// Security: WORKDAY counts the holidays each step passes by binary search, so a long run of
+    /// consecutive holidays is linear work, not one scan of every holiday per holiday.
+    #[test]
+    fn workday_over_many_consecutive_holidays() {
+        let start = d(2000.0, 1.0, 3.0);
+        let hol: Vec<f64> = (1..=200_000).map(|i| start + f64::from(i)).collect();
+        let out = ev("WORKDAY.INTL", vec![n(start), n(1.0), n(11.0), av(col(&hol))]);
+        // Sundays are the only weekend (11): 200,000 days of holidays are skipped, landing on the
+        // first working day after them.
+        let after = start + 200_001.0;
+        let sunday = ev("WEEKDAY", vec![n(after), n(2.0)]) == nv(7.0);
+        close(out, if sunday { after + 1.0 } else { after });
+        // Holidays on a weekend don't push the result out.
+        let fri = d(2026.0, 10.0, 9.0);
+        close(ev("WORKDAY", vec![n(fri), n(1.0), av(col(&[fri + 1.0, fri + 3.0]))]), fri + 4.0);
+    }
+
     #[test]
     fn day_counts() {
         close(ev("DAYS", vec![n(d(2021.0, 3.0, 15.0)), n(d(2021.0, 2.0, 1.0))]), 42.0);

crates/functions/src/dynamic.rs

@@ -178,6 +178,10 @@ fn sort(args: &[Arg], _c: &mut dyn Ctx) -> R<Value> {
         return Err(CellError::Value);
     }
     let orders: Vec<i64> = (0..idx.len()).map(|i| *ord.get(i).or(ord.first()).unwrap_or(&1)).collect();
+    // A column repeated in sort_index can't break a tie its first use left, so only first uses
+    // stay: the keys copied per row are then at most the row itself, however long sort_index is.
+    let mut seen = std::collections::HashSet::new();
+    let (idx, orders): (Vec<i64>, Vec<i64>) = idx.into_iter().zip(orders).filter(|(i, _)| seen.insert(*i)).unzip();
     let mut lines: Vec<(Vec<Value>, Vec<Value>)> =
         rows_of(&a).into_iter().map(|r| (idx.iter().map(|&i| r.get(i as usize - 1).cloned().unwrap_or(Value::Empty)).collect(), r)).collect();
     lines.sort_by(|x, y| cmp_keys(&x.0, &y.0, &orders));
@@ -834,9 +838,16 @@ fn pivotby(args: &[Arg], ctx: &mut dyn Ctx) -> R<Value> {
     let ckeys: Vec<Vec<Value>> = p.rows.iter().map(|&r| key_row(&cf, r)).collect();
     let rk = distinct_keys(&rkeys, if row_sort.unsigned_abs() as usize <= rf.cols { row_sort } else { 0 });
     let ck = distinct_keys(&ckeys, if col_sort.unsigned_abs() as usize <= cf.cols { col_sort } else { 0 });
+    // Source rows of every row key, column key and pair, built in one pass, so a total reads
+    // its rows instead of rescanning the whole source once per total.
     let mut cell: HashMap<(String, String), Vec<usize>> = HashMap::new();
+    let mut by_row: HashMap<String, Vec<usize>> = HashMap::new();
+    let mut by_col: HashMap<String, Vec<usize>> = HashMap::new();
     for ((r, c), &i) in rkeys.iter().zip(ckeys.iter()).zip(p.rows.iter()) {
-        cell.entry((row_key(r), row_key(c))).or_default().push(i);
+        let (r, c) = (row_key(r), row_key(c));
+        by_row.entry(r.clone()).or_default().push(i);
+        by_col.entry(c.clone()).or_default().push(i);
+        cell.entry((r, c)).or_default().push(i);
     }
     let mut agg = |rows: &[usize], c: usize| -> Value {
         if rows.is_empty() {
@@ -846,14 +857,12 @@ fn pivotby(args: &[Arg], ctx: &mut dyn Ctx) -> R<Value> {
         aggregate_with(f, &vs, ctx)
     };
     let rows_for = |r: Option<&Vec<Value>>, c: Option<&Vec<Value>>| -> Vec<usize> {
-        let rkey = r.map(|k| row_key(k));
-        let ckey = c.map(|k| row_key(k));
-        p.rows
-            .iter()
-            .zip(rkeys.iter().zip(ckeys.iter()))
-            .filter(|(_, (a, b))| rkey.as_ref().is_none_or(|k| *k == row_key(a)) && ckey.as_ref().is_none_or(|k| *k == row_key(b)))
-            .map(|(&i, _)| i)
-            .collect()
+        match (r, c) {
+            (Some(r), Some(c)) => cell.get(&(row_key(r), row_key(c))).cloned().unwrap_or_default(),
+            (Some(r), None) => by_row.get(&row_key(r)).cloned().unwrap_or_default(),
+            (None, Some(c)) => by_col.get(&row_key(c)).cloned().unwrap_or_default(),
+            (None, None) => p.rows.clone(),
+        }
     };
     let mut cols: Vec<Option<&Vec<Value>>> = ck.iter().map(Some).collect();
     if col_total > 0 {
@@ -903,10 +912,7 @@ fn pivotby(args: &[Arg], ctx: &mut dyn Ctx) -> R<Value> {
             }
         };
         for c in &cols {
-            let rows = match (r, c) {
-                (Some(rk), Some(ck)) => cell.get(&(row_key(rk), row_key(ck))).cloned().unwrap_or_default(),
-                _ => rows_for(r, *c),
-            };
+            let rows = rows_for(r, *c);
             for vc in 0..vals.cols {
                 line.push(agg(&rows, vc));
             }
@@ -1110,6 +1116,21 @@ mod tests {
         assert_eq!(rows_of(&r), vec![vec![nv(0.0), nv(5.0)], vec![nv(1.0), nv(2.0)], vec![nv(1.0), nv(1.0)]]);
     }
 
+    /// Security: a column repeated in sort_index keys the sort once. The keys copied per row stay
+    /// within the row, however long sort_index is, and a repeat (even with the other order) can't
+    /// change the result, since its first use already ordered the rows.
+    #[test]
+    fn sort_drops_repeated_sort_indices() {
+        let m = || av(arr(vec![vec![nv(1.0), nv(2.0)], vec![nv(1.0), nv(1.0)], vec![nv(0.0), nv(5.0)]]));
+        let want = vec![vec![nv(0.0), nv(5.0)], vec![nv(1.0), nv(2.0)], vec![nv(1.0), nv(1.0)]];
+        let r = ev("SORT", vec![m(), av(row(&[1.0, 1.0, 2.0, 1.0])), av(row(&[1.0, -1.0, -1.0, 1.0]))]);
+        assert_eq!(rows_of(&r), want);
+        let idx = vec![1.0; 1_000_000];
+        let tall = av(gridcraft_core::Array::filled(2_000, 1, nv(1.0)).into());
+        let r = ev("SORT", vec![tall, av(row(&idx))]);
+        assert_eq!(rows_of(&r).len(), 2_000);
+    }
+
     #[test]
     fn sortby() {
         let r = ev("SORTBY", vec![av(col(&[1.0, 2.0, 3.0])), av(col(&[30.0, 10.0, 20.0]))]);
@@ -1281,6 +1302,16 @@ mod tests {
         );
     }
 
+    /// Security: PIVOTBY totals read their source rows from indexes built in the grouping pass,
+    /// rather than rescanning the source once per total, so many distinct keys stay linear.
+    #[test]
+    fn pivotby_totals_with_many_groups() {
+        let keys: Vec<f64> = (0..20_000).map(f64::from).collect();
+        let r = rows_of(&ev("PIVOTBY", vec![av(col(&keys)), av(col(&vec![1.0; 20_000])), av(col(&vec![2.0; 20_000])), t("SUM")]));
+        assert_eq!(r.get(1), Some(&vec![nv(0.0), nv(2.0), nv(2.0)]));
+        assert_eq!(r.last(), Some(&vec![tv("Total"), nv(40_000.0), nv(40_000.0)]));
+    }
+
     #[test]
     fn pivotby_size_check_counts_column_header_rows() {
         // One row, n column fields and n value columns: n header rows + value-name row + body row + total row,