perf(calc): remember SUM/COUNT/AVERAGE of constant-only blocks per pass (#2681)

A block that holds only constants can't change during a pass, so its sum
and count are worked out once and reused by every formula that reads the
same block. A block that is the same rows from the top but longer (a
running total) continues the shorter block's sum, which is the same
additions in the same order as one walk from the start.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

crates/calc/src/eval.rs

@@ -103,6 +103,12 @@ pub trait Host {
     fn visit_range(&mut self, _sheet: usize, _range: RangeRef, _f: &mut dyn FnMut(&Value) -> bool) -> bool {
         false
     }
+    /// `(sum, count)` of the numbers in a block that holds nothing but constants (no formula,
+    /// error or spilled value), added row-major from `0.0` exactly as `visit_range` would hand
+    /// them over. `None`: the block isn't like that, or this host can't tell.
+    fn range_sum(&mut self, _sheet: usize, _range: RangeRef) -> Option<(f64, u64)> {
+        None
+    }
     /// While a formula is calculated on demand inside another one: `stack_position()` where the
     /// outermost one started, which the stack budget is counted from.
     fn stack_top(&self) -> Option<usize> {
@@ -984,6 +990,14 @@ impl<'h> Evaluator<'h> {
     /// Adds one area to `acc`; `Err` is the result the function ends with.
     fn aggregate_area(&mut self, kind: Agg, a: &Area, acc: &mut Acc) -> Result<(), Ev> {
         if !a.range.is_single() {
+            // Adding the block's remembered sum to anything but +0.0 would add in another order.
+            if acc.sum.to_bits() == 0
+                && let Some((sum, count)) = self.host.range_sum(a.sheet, a.range)
+            {
+                acc.sum = sum;
+                acc.count = acc.count.saturating_add(count);
+                return Ok(());
+            }
             let mut error = None;
             let walked = self.host.visit_range(a.sheet, a.range, &mut |v| match acc.add_cell(kind, v) {
                 Ok(()) => true,

crates/calc/src/recalc.rs

@@ -436,6 +436,10 @@ struct PassHost<'a> {
     stack_top: usize,
     /// Whether an evaluation ran past the stack budget since the innermost formula started.
     out_of_stack: bool,
+    /// Sum and count of blocks that hold only constants, for `range_sum`. Keyed by sheet, first
+    /// cell and last column; each value maps a last row to the block's result (`None`: the block
+    /// has a formula, an error value or a spilled value, so every longer one has too).
+    sums: HashMap<(usize, CellRef, u32), std::collections::BTreeMap<u32, Option<(f64, u64)>>>,
 }
 
 impl<'a> PassHost<'a> {
@@ -468,6 +472,7 @@ impl<'a> PassHost<'a> {
             taint: false,
             stack_top: 0,
             out_of_stack: false,
+            sums: HashMap::default(),
         }
     }
 }
@@ -909,6 +914,50 @@ impl Host for PassHost<'_> {
         }
         true
     }
+    fn range_sum(&mut self, sheet: usize, range: RangeRef) -> Option<(f64, u64)> {
+        // Constants can't change during a pass, so a block's result stays valid for it. Spills
+        // found in this pass (and iteration, which tracks what a read depends on) go the long way.
+        if self.iterate || self.spills.keys().any(|a| a.0 == sheet) {
+            return None;
+        }
+        let sh = self.wb.sheet(sheet)?;
+        let (key, last) = ((sheet, range.start, range.end.col), range.end.row);
+        let known = self.sums.entry(key).or_default();
+        if let Some(r) = known.get(&last) {
+            return *r;
+        }
+        // Start from the longest shorter block of the same rows from the top: summing row-major
+        // from there on is the same additions in the same order as one walk from the start.
+        let (from, mut sum, mut count) = match known.range(..last).next_back() {
+            Some((_, None)) => {
+                known.insert(last, None);
+                return None;
+            }
+            Some((&row, &Some((sum, count)))) => (row + 1, sum, count),
+            None => (range.start.row, 0.0, 0),
+        };
+        let rest = RangeRef::new(CellRef::new(from, range.start.col), range.end);
+        let mut plain = sh.spill.range(rest.start..=rest.end).all(|(c, _)| !rest.contains(*c));
+        if plain {
+            for (_, cell) in sh.cells.iter_range(rest) {
+                match &cell.value {
+                    _ if cell.formula.is_some() => plain = false,
+                    Value::Error(_) => plain = false,
+                    Value::Number(n) => {
+                        sum += n;
+                        count = u64::saturating_add(count, 1);
+                    }
+                    _ => {}
+                }
+                if !plain {
+                    break;
+                }
+            }
+        }
+        let result = plain.then_some((sum, count));
+        known.insert(last, result);
+        result
+    }
     fn used_range(&mut self, sheet: usize) -> Option<RangeRef> {
         let wb = self.wb;
         *self.used.entry(sheet).or_insert_with(|| wb.sheet(sheet).and_then(|s| s.used_range()))

crates/calc/src/tests.rs

@@ -2274,3 +2274,116 @@ fn loan_schedule_residue_keeps_excels_sign() {
     }
     assert_eq!(t.get("P7"), Value::Text("".into()));
 }
+
+/// `SUM`, `COUNT` and `AVERAGE` over blocks of constants share one pass-wide answer per block, and a
+/// running total extends the shorter block's. Every result must equal the left-to-right sum of the
+/// cells, bit for bit, with text, blanks, errors, formulas and a nonzero start in the mix.
+#[test]
+fn aggregates_over_ranges_match_a_plain_fold_bit_for_bit() {
+    struct Rng(u64);
+    impl Rng {
+        fn next(&mut self, n: u64) -> u64 {
+            self.0 ^= self.0 << 13;
+            self.0 ^= self.0 >> 7;
+            self.0 ^= self.0 << 17;
+            self.0 % n
+        }
+    }
+    let nums = [0.1, 0.2, 0.3, 1e16, -1e16, 1.0, 3.7, -0.0, 12345.678, 1e-9, 7.0];
+    let rows = 40u32;
+    let mut rng = Rng(0x9E37_79B9_7F4A_7C15);
+    let mut wb = Workbook::new();
+    let fresh = |rng: &mut Rng, plain: bool| match rng.next(if plain { 8 } else { 14 }) {
+        0 => Cell::value(Value::text("t")),
+        1 => Cell::default(),
+        2 => Cell::value(Value::Bool(true)),
+        3 if !plain => Cell::value(Value::Error(CellError::NA)),
+        4 if !plain => Cell::formula(Formula::new("=A1*2")),
+        5 if !plain => Cell::formula(Formula::new("=\"x\"&B2")),
+        _ => Cell::value(Value::Number(nums[rng.next(nums.len() as u64) as usize])),
+    };
+    // Columns A:C mostly numbers; E holds a formula column and some errors only in a few rows.
+    for r in 0..rows {
+        for col in 0..3 {
+            let plain = !(r % 13 == 7 && col == 1);
+            wb.sheet_mut(0).unwrap().cells.set(CellRef::new(r, col), fresh(&mut rng, plain));
+        }
+    }
+    // Formulas in column H and up: (function, optional leading number, blocks).
+    let mut oracle: Vec<(CellRef, &str, Option<f64>, Vec<(u32, u32, u32, u32)>)> = vec![];
+    let funcs = ["SUM", "COUNT", "AVERAGE"];
+    let mut at = 0;
+    let mut add = |wb: &mut Workbook, f: &'static str, lead: Option<f64>, blocks: Vec<(u32, u32, u32, u32)>| {
+        let args: Vec<String> = blocks
+            .iter()
+            .map(|&(r0, c0, r1, c1)| format!("{}{}:{}{}", (b'A' + c0 as u8) as char, r0 + 1, (b'A' + c1 as u8) as char, r1 + 1))
+            .collect();
+        let lead_s = lead.map(|l| format!("{l:e},")).unwrap_or_default();
+        let cell = CellRef::new(at % rows, 7 + at / rows);
+        wb.sheet_mut(0).unwrap().cells.set(cell, Cell::formula(Formula::new(&format!("={f}({lead_s}{})", args.join(",")))));
+        oracle.push((cell, f, lead, blocks));
+        at += 1;
+    };
+    // Running totals over a clean column and over a column with an error further down.
+    for r in 0..rows {
+        add(&mut wb, "SUM", None, vec![(0, 0, r, 0)]);
+        add(&mut wb, "SUM", None, vec![(0, 1, r, 1)]);
+        add(&mut wb, "COUNT", None, vec![(0, 0, r, 2)]);
+    }
+    for _ in 0..120 {
+        let (r0, r1) = (rng.next(rows as u64) as u32, rng.next(rows as u64) as u32);
+        let (c0, c1) = (rng.next(3) as u32, rng.next(3) as u32);
+        let n = 1 + rng.next(2) as usize;
+        let blocks = (0..n).map(|_| (r0.min(r1), c0.min(c1), r0.max(r1), c0.max(c1))).collect();
+        let lead = [None, Some(1e16), Some(0.5), Some(2.0)][rng.next(4) as usize];
+        let f = funcs[rng.next(3) as usize];
+        add(&mut wb, f, lead, blocks);
+    }
+    let check = |wb: &Workbook, when: &str| {
+        let sh = wb.sheet(0).unwrap();
+        for (cell, f, lead, blocks) in &oracle {
+            let (mut sum, mut count, mut error) = (lead.unwrap_or(0.0), u64::from(lead.is_some()), None);
+            for &(r0, c0, r1, c1) in blocks {
+                for r in r0..=r1 {
+                    for col in c0..=c1 {
+                        match sh.value(CellRef::new(r, col)) {
+                            Value::Number(n) => {
+                                sum += n;
+                                count += 1;
+                            }
+                            Value::Error(e) if *f != "COUNT" => error = error.or(Some(e)),
+                            _ => {}
+                        }
+                    }
+                }
+            }
+            let want = match (error, *f) {
+                (Some(e), _) => Value::Error(e),
+                (_, "SUM") => Value::number(sum),
+                (_, "COUNT") => Value::number(count as f64),
+                (_, _) if count == 0 => Value::Error(CellError::Div0),
+                _ => Value::number(sum / count as f64),
+            };
+            let got = sh.value(*cell);
+            let same = match (&got, &want) {
+                (Value::Number(a), Value::Number(b)) => a.to_bits() == b.to_bits(),
+                (a, b) => a == b,
+            };
+            assert!(same, "{when}: {:?} = {:?}, plain fold {:?}", cell, got, want);
+        }
+    };
+    let mut calc = Calc::new();
+    calc.rebuild(&wb);
+    calc.recalc_all(&mut wb);
+    check(&wb, "full recalc");
+    for round in 0..30 {
+        let mut changed = vec![];
+        for _ in 0..1 + rng.next(4) {
+            let at = CellRef::new(rng.next(rows as u64) as u32, rng.next(3) as u32);
+            wb.sheet_mut(0).unwrap().cells.set(at, fresh(&mut rng, round % 2 == 0));
+            changed.push((0, at));
+        }
+        calc.cells_changed(&mut wb, &changed);
+        check(&wb, &format!("edit round {round}"));
+    }
+}