pivot: cached Rank and Running Total, accumulator-width limit, capped subtotal captions and PivotChart series

crates/engine/src/pivot/charts.rs

@@ -57,6 +57,10 @@ fn rows_ref(sheet: &str, origin: CellRef, col: u32, rows: &[u32]) -> Option<Stri
 /// Most category rows a PivotChart plots.
 const MAX_CHART_ROWS: usize = 10_000;
 
+/// Most series a PivotChart gets: the chart renderer reads at most 255 series, and each one owns
+/// reference text as long as its category rows, copied into every linked chart.
+const MAX_CHART_SERIES: usize = 255;
+
 /// A PivotChart's series for a report written at `origin`: categories are the innermost row
 /// items, one series per item column (column items × value fields); totals are left out.
 pub fn chart_series(sheet: &str, origin: CellRef, out: &Output) -> Vec<Series> {
@@ -64,6 +68,7 @@ pub fn chart_series(sheet: &str, origin: CellRef, out: &Output) -> Vec<Series> {
     let categories = out.label_col.and_then(|c| rows_ref(sheet, origin, c, &rows));
     out.chart_cols
         .iter()
+        .take(MAX_CHART_SERIES)
         .filter_map(|(c, name)| {
             Some(Series {
                 name: Some(format!("\"{}\"", name.replace('"', "\"\""))),
@@ -106,3 +111,22 @@ pub fn clear_output(wb: &mut Workbook, sheet: usize, idx: usize) {
         p.last_range = None;
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn a_pivot_chart_gets_at_most_the_series_the_renderer_reads() {
+        let out = Output {
+            grid: vec![],
+            row_kinds: vec![],
+            body_offset: 0,
+            width: 1000,
+            chart_rows: vec![1, 2, 3],
+            chart_cols: (1..1000).map(|c| (c, format!("S{c}"))).collect(),
+            label_col: Some(0),
+        };
+        assert_eq!(chart_series("Sheet1", CellRef::new(0, 0), &out).len(), MAX_CHART_SERIES);
+    }
+}

crates/engine/src/pivot/compute.rs

@@ -1,5 +1,7 @@
 use super::items::{Acc, FieldItems, build_tree, collect_siblings, is_collapsed, item_label};
-use super::output::{AggMap, CellKind, ColKind, ColSpec, Ctx, Grid, Line, LineKind, Output, RowKind, ValSrc, data_cols_exceed_sheet, flatten_cols};
+use super::output::{
+    AggMap, CellKind, ColKind, ColSpec, Ctx, Grid, Line, LineKind, Output, RowKind, ValSrc, caption, data_cols_exceed_sheet, flatten_cols,
+};
 use super::{
     CalcSet, CellError, CellRef, HashMap, HashSet, MAX_AXIS_FIELDS, MAX_COLS, MAX_VALUE_FIELDS, PivotDateGroup, PivotField, PivotFilter, PivotFunc,
     PivotLayout, PivotShowAs, PivotTable, PivotValue, Source, Value, Workbook, read_source,
@@ -92,6 +94,10 @@ struct AggregateData {
     col_leaves: HashSet<Vec<u32>>,
 }
 
+/// Most accumulators (groups times value columns) one report may hold: twice what the 5 million
+/// group limit allowed at one value field.
+const MAX_ACCUMULATORS: usize = 16_000_000;
+
 fn aggregate_record(
     input: &AggregateInput<'_>,
     data: &mut AggregateData,
@@ -150,7 +156,9 @@ fn aggregate_record(
     if !input.cols.is_empty() && data_cols_exceed_sheet(data.col_leaves.len(), input.pt.grand_totals_rows, input.vals.len()) {
         return Err("The PivotTable report will not fit on the sheet.".into());
     }
-    if data.aggs.len() > 5_000_000 {
+    // Each entry holds one accumulator per value field and calculated-field source column, so the
+    // width counts too: a handful of groups over hundreds of fields is large as well.
+    if data.aggs.len() > 5_000_000 || data.aggs.len().saturating_mul(input.vals.len() + input.extra.len()) > MAX_ACCUMULATORS {
         return Err("The PivotTable report is too large.".into());
     }
     Ok(())
@@ -257,8 +265,8 @@ fn write_headers(g: &mut Grid, input: HeaderInput<'_>) -> Result<(), String> {
                     let lv = spec.path.len().saturating_sub(1);
                     let label = spec.path.last().and_then(|it| col_items.get(lv).map(|fi| fi.label(*it))).unwrap_or_default();
                     let text = match vi {
-                        Some(vi) if k > 1 => format!("{label} {}", val_name(*vi)),
-                        _ => format!("{label} Total"),
+                        Some(vi) if k > 1 => caption(&label, &val_name(*vi)),
+                        _ => caption(&label, "Total"),
                     };
                     g.set(first_level + lv, c, Value::text(text), CellKind::ColLabel, 0, None);
                 }
@@ -316,6 +324,7 @@ struct ShowContext<'a> {
     value_index: usize,
     level0: &'a [u32],
     siblings: &'a HashMap<Vec<u32>, Vec<u32>>,
+    cache: &'a ShowCache,
 }
 
 fn number(value: &Option<Value>) -> Option<f64> {
@@ -332,42 +341,65 @@ fn ratio(a: Value, b: Option<Value>) -> Value {
     }
 }
 
+/// What Running Total and Rank need from other cells, computed once per sibling set instead of
+/// once per displayed value (which rescanned the whole set every time).
+type ShowKey = (Vec<u32>, Vec<u32>, usize);
+/// Cumulative values by top-level item, and the total of all of them.
+type RunningTotals = (HashMap<u32, f64>, f64);
+
+#[derive(Default)]
+struct ShowCache {
+    /// Ranked values' sibling numbers, sorted ascending without repeats, by (parent, column, value).
+    ranks: std::cell::RefCell<HashMap<ShowKey, Vec<f64>>>,
+    /// Running totals by top-level item, with the total of all of them, by (rest of path, column, value).
+    running: std::cell::RefCell<HashMap<ShowKey, RunningTotals>>,
+}
+
 fn running_total(raw: Value, lookup: &ValueLookup<'_>, context: &ShowContext<'_>) -> Value {
-    let Some((first, _)) = context.row_path.split_first() else { return raw };
+    let Some((first, rest)) = context.row_path.split_first() else { return raw };
     if !matches!(raw, Value::Number(_)) {
         return raw;
     }
-    let mut total = 0.0;
-    for item in context.level0 {
-        let mut path = context.row_path.to_vec();
-        if let Some(first_item) = path.first_mut() {
-            *first_item = *item;
+    let mut cache = context.cache.running.borrow_mut();
+    let (by_item, all) = cache.entry((rest.to_vec(), context.col_path.to_vec(), context.value_index)).or_insert_with(|| {
+        let mut total = 0.0;
+        let mut by_item = HashMap::new();
+        for item in context.level0 {
+            let mut path = Vec::with_capacity(rest.len() + 1);
+            path.push(*item);
+            path.extend_from_slice(rest);
+            total += number(&lookup.get(&path, context.col_path, context.value_index)).unwrap_or(0.0);
+            by_item.entry(*item).or_insert(total);
         }
-        total += number(&lookup.get(&path, context.col_path, context.value_index)).unwrap_or(0.0);
-        if item == first {
-            break;
-        }
-    }
-    Value::number(total)
+        (by_item, total)
+    });
+    Value::number(by_item.get(first).copied().unwrap_or(*all))
 }
 
 fn ranked_value(raw: Value, lookup: &ValueLookup<'_>, context: &ShowContext<'_>) -> Value {
     let Some((_, parent)) = context.row_path.split_last() else { return Value::Empty };
     let Some(me) = raw.as_f64() else { return raw };
-    let siblings = context.siblings.get(parent).cloned().unwrap_or_default();
+    let mut cache = context.cache.ranks.borrow_mut();
+    let sorted = cache.entry((parent.to_vec(), context.col_path.to_vec(), context.value_index)).or_insert_with(|| {
+        let mut all: Vec<f64> = context
+            .siblings
+            .get(parent)
+            .into_iter()
+            .flatten()
+            .filter_map(|sibling| {
+                let mut path = parent.to_vec();
+                path.push(*sibling);
+                number(&lookup.get(&path, context.col_path, context.value_index))
+            })
+            .collect();
+        all.sort_by(f64::total_cmp);
+        all.dedup();
+        all
+    });
     // Excel's ranks are dense: ties share a rank and the next value takes the next one
     // (5, 4, 3, 3, 4 rank 1, 2, 3, 3, 2), so count the distinct larger values.
-    let mut larger: Vec<f64> = siblings
-        .into_iter()
-        .filter_map(|sibling| {
-            let mut path = parent.to_vec();
-            path.push(sibling);
-            number(&lookup.get(&path, context.col_path, context.value_index)).filter(|value| *value > me)
-        })
-        .collect();
-    larger.sort_by(f64::total_cmp);
-    larger.dedup();
-    Value::number(larger.len() as f64 + 1.0)
+    let larger = sorted.len() - sorted.partition_point(|value| *value <= me);
+    Value::number(larger as f64 + 1.0)
 }
 
 fn show_value(definition: &PivotValue, raw: Value, lookup: &ValueLookup<'_>, context: &ShowContext<'_>) -> Value {
@@ -637,6 +669,7 @@ pub(super) fn compute_from(wb: &Workbook, pt: &PivotTable, src: &Source, mut pro
         })
         .collect();
     let lookup = ValueLookup { wb, source: src, calc: &calc, vals: &vals, extra: &extra, aggs: &aggs, value_count: k };
+    let show_cache = ShowCache::default();
     let mut chart_rows: Vec<u32> = Vec::new();
     let chart_cols: Vec<(u32, String)> = data_cols
         .iter()
@@ -694,7 +727,7 @@ pub(super) fn compute_from(wb: &Workbook, pt: &PivotTable, src: &Source, mut pro
             let Some((_, vdef)) = vals.get(*vi) else { continue };
             let cp = spec.path.as_slice();
             let Some(raw) = lookup.get(rp, cp, *vi) else { continue };
-            let context = ShowContext { row_path: rp, col_path: cp, value_index: *vi, level0: &level0, siblings: &siblings };
+            let context = ShowContext { row_path: rp, col_path: cp, value_index: *vi, level0: &level0, siblings: &siblings, cache: &show_cache };
             let shown = show_value(vdef, raw, &lookup, &context);
             if let Some(pr) = probe.as_deref_mut() {
                 if pr.at == Some((r, l + j)) {

crates/engine/src/pivot/output.rs

@@ -121,6 +121,18 @@ pub(super) enum ColKind {
     Grand,
 }
 
+/// Longest item caption Excel keeps; a subtotal caption repeats its parent's label for every value
+/// field at every level, so a long source label is cut rather than multiplied.
+const MAX_CAPTION_LABEL_CHARS: usize = 255;
+
+/// `label` (cut to [`MAX_CAPTION_LABEL_CHARS`]) and `suffix`, as in "East Total".
+pub(super) fn caption(label: &str, suffix: &str) -> String {
+    let mut out: String = label.chars().take(MAX_CAPTION_LABEL_CHARS).collect();
+    out.push(' ');
+    out.push_str(suffix);
+    out
+}
+
 pub(super) struct ColSpec {
     pub(super) kind: ColKind,
     pub(super) path: Vec<u32>,
@@ -162,7 +174,7 @@ impl Ctx<'_> {
                 leaf: false,
                 path: prefix.to_vec(),
                 values: true,
-                labels: vec![(col, Value::text(format!("{label} {name}")), indent, None)],
+                labels: vec![(col, Value::text(caption(label, name)), indent, None)],
                 vi: Some(vi),
             });
         }
@@ -249,7 +261,7 @@ impl Ctx<'_> {
                         leaf: false,
                         path: prefix.clone(),
                         values: true,
-                        labels: vec![(col, Value::text(format!("{label} Total")), indent, None)],
+                        labels: vec![(col, Value::text(caption(&label, "Total")), indent, None)],
                         vi,
                     });
                 }
@@ -316,7 +328,7 @@ impl Ctx<'_> {
                         leaf: false,
                         path: prefix.clone(),
                         values: true,
-                        labels: vec![(col, Value::text(format!("{label} Total")), 0, None)],
+                        labels: vec![(col, Value::text(caption(&label, "Total")), 0, None)],
                         vi,
                     });
                 }
@@ -438,3 +450,17 @@ pub(super) fn cell_style(lk: &Look, rk: RowKind, cell: &OutCell, last_header: bo
     }
     s
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn captions_cut_a_long_parent_label() {
+        assert_eq!(caption("East", "Total"), "East Total");
+        let long = "é".repeat(100_000);
+        let c = caption(&long, "Sum of Sales");
+        assert_eq!(c.chars().count(), MAX_CAPTION_LABEL_CHARS + " Sum of Sales".len());
+        assert!(c.ends_with(" Sum of Sales"));
+    }
+}