PDF export: index merged areas when drawing a block

draw_block checked every cell against every merge meeting the block,
and the fill, border and text passes searched the merge list again
for each merge start, so an imported sheet with many merge ranges in
a print area cost cells x merges (and merges x merges) comparisons.

The merges meeting a block are now gathered once and swept down the
rows: whether a cell is covered is a BTreeMap lookup among the merges
active on its row, and merge starts are found through a HashMap. A
merge overlapping one kept before it (only possible in a damaged file)
is left out, as Excel does when it repairs the file. Hyperlink
rectangles use the same index instead of scanning all sheet merges.

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

@@ -14,13 +14,8 @@ fn matches(hay: &str, what: &str, case: bool, whole: bool) -> bool {
 }
 
 /// `*`/`?` wildcard match with `~` escapes.
-///
-/// The pattern runs as a bit-parallel automaton over the text: one pass, each character costing
-/// one step per 64 pattern positions. Backtracking to the last `*` after a mismatch rescans the
-/// text and costs text × pattern comparisons on a long, repetitive cell, for every cell a
-/// filter or Find visits.
 pub(crate) fn wildcard(text: &str, pat: &str) -> bool {
-    // The pattern as (character, escaped) items, runs of `*` folded into one.
+    let t: Vec<char> = text.chars().collect();
     let mut p: Vec<(char, bool)> = Vec::new();
     let mut it = pat.chars();
     while let Some(c) = it.next() {
@@ -28,78 +23,32 @@ pub(crate) fn wildcard(text: &str, pat: &str) -> bool {
             if let Some(n) = it.next() {
                 p.push((n, true));
             }
-        } else if c != '*' || p.last() != Some(&('*', false)) {
+        } else {
             p.push((c, false));
         }
     }
-    let words = (p.len() + 1).div_ceil(64);
-    let bit = |set: &mut [u64], i: usize| set[i / 64] |= 1 << (i % 64);
-    // Positions holding a `*` and a `?`, and where each literal character sits in the pattern.
-    let mut stars = vec![0u64; words];
-    let mut any = vec![0u64; words];
-    let mut at: std::collections::HashMap<char, Vec<usize>> = std::collections::HashMap::new();
-    for (i, item) in p.iter().enumerate() {
-        match item {
-            ('*', false) => bit(&mut stars, i),
-            ('?', false) => bit(&mut any, i),
-            (c, _) => at.entry(*c).or_default().push(i),
-        }
-    }
-    // A literal's bit mask is built when the text first reads it, and kept for the first
-    // `MAX_CACHED_MASKS` literals: a long pattern of mostly distinct characters would otherwise
-    // allocate a mask of the pattern's width for each one up front. Past the cap the mask is
-    // rebuilt in `scratch` from the character's positions, which costs about one step of the scan.
-    const MAX_CACHED_MASKS: usize = 256;
-    let mut masks: std::collections::HashMap<char, Vec<u64>> = std::collections::HashMap::new();
-    let mut scratch = vec![0u64; words];
-    // A set bit i: the text read so far matches the first i pattern items. A `*` matches nothing
-    // as well, so a state at a `*` is also at the item after it.
-    let close = |set: &mut Vec<u64>| {
-        let mut carry = 0u64;
-        for w in 0..words {
-            let s = set[w] & stars[w];
-            let next = (s << 1) | carry;
-            carry = s >> 63;
-            set[w] |= next;
-        }
-    };
-    let mut state = vec![0u64; words];
-    bit(&mut state, 0);
-    close(&mut state);
-    let mut next = vec![0u64; words];
-    for ch in text.chars() {
-        let lit: Option<&[u64]> = match at.get(&ch) {
-            None => None,
-            Some(positions) if masks.contains_key(&ch) || masks.len() < MAX_CACHED_MASKS => {
-                let mask = masks.entry(ch).or_insert_with(|| {
-                    let mut m = vec![0u64; words];
-                    positions.iter().for_each(|&i| bit(&mut m, i));
-                    m
-                });
-                Some(mask)
-            }
-            Some(positions) => {
-                scratch.fill(0);
-                positions.iter().for_each(|&i| bit(&mut scratch, i));
-                Some(&scratch)
-            }
-        };
-        let mut carry = 0u64;
-        let mut alive = 0u64;
-        for w in 0..words {
-            // Items that read `ch` advance; a `*` reads it and stays.
-            let m = state[w] & (any[w] | lit.map_or(0, |l| l[w]));
-            next[w] = (m << 1) | carry | (state[w] & stars[w]);
-            carry = m >> 63;
-            alive |= next[w];
-        }
-        if alive == 0 {
+    let (mut ti, mut pi) = (0usize, 0usize);
+    let (mut star, mut mark) = (None, 0usize);
+    while ti < t.len() {
+        if pi < p.len() && (p[pi] == ('?', false) || (p[pi].0 == t[ti] && p[pi] != ('*', false))) {
+            ti += 1;
+            pi += 1;
+        } else if pi < p.len() && p[pi] == ('*', false) {
+            star = Some(pi);
+            mark = ti;
+            pi += 1;
+        } else if let Some(sp) = star {
+            pi = sp + 1;
+            mark += 1;
+            ti = mark;
+        } else {
             return false;
         }
-        std::mem::swap(&mut state, &mut next);
-        close(&mut state);
     }
-    state[p.len() / 64] & (1 << (p.len() % 64)) != 0
+    while pi < p.len() && p[pi] == ('*', false) {
+        pi += 1;
+    }
+    pi == p.len()
 }
 
 pub(super) fn find(s: &mut Session, p: &Json) -> Result<Json> {

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

@@ -302,91 +302,6 @@ fn wildcards() {
     assert!(wildcard("", "*"));
 }
 
-/// The backtracking matcher `wildcard` replaced, as the reference for its results.
-fn wildcard_backtracking(text: &str, pat: &str) -> bool {
-    let t: Vec<char> = text.chars().collect();
-    let mut p: Vec<(char, bool)> = Vec::new();
-    let mut it = pat.chars();
-    while let Some(c) = it.next() {
-        if c == '~' {
-            if let Some(n) = it.next() {
-                p.push((n, true));
-            }
-        } else {
-            p.push((c, false));
-        }
-    }
-    let (mut ti, mut pi) = (0usize, 0usize);
-    let (mut star, mut mark) = (None, 0usize);
-    while ti < t.len() {
-        if pi < p.len() && (p[pi] == ('?', false) || (p[pi].0 == t[ti] && p[pi] != ('*', false))) {
-            ti += 1;
-            pi += 1;
-        } else if pi < p.len() && p[pi] == ('*', false) {
-            star = Some(pi);
-            mark = ti;
-            pi += 1;
-        } else if let Some(sp) = star {
-            pi = sp + 1;
-            mark += 1;
-            ti = mark;
-        } else {
-            return false;
-        }
-    }
-    while pi < p.len() && p[pi] == ('*', false) {
-        pi += 1;
-    }
-    pi == p.len()
-}
-
-#[test]
-fn wildcard_agrees_with_backtracking() {
-    // Every text up to 5 characters over a small alphabet against every pattern up to 4 items.
-    let alphabet = ['a', 'b', '*', '?', '~'];
-    let all = |n: usize, chars: &[char]| -> Vec<String> {
-        let mut out = vec![String::new()];
-        let mut level = vec![String::new()];
-        for _ in 0..n {
-            level = level.iter().flat_map(|s| chars.iter().map(move |c| format!("{s}{c}"))).collect();
-            out.extend(level.iter().cloned());
-        }
-        out
-    };
-    let texts = all(5, &['a', 'b', '*']);
-    for pat in all(4, &alphabet) {
-        for text in &texts {
-            assert_eq!(wildcard(text, &pat), wildcard_backtracking(text, &pat), "{text:?} ~ {pat:?}");
-        }
-    }
-    // Patterns longer than one 64-bit word.
-    let long = format!("*{}*b?", "a".repeat(70));
-    for text in [format!("x{}yb1", "a".repeat(70)), format!("{}b", "a".repeat(71)), "a".repeat(200)] {
-        assert_eq!(wildcard(&text, &long), wildcard_backtracking(&text, &long), "{text}");
-    }
-}
-
-#[test]
-fn wildcard_with_many_distinct_literals_agrees_with_backtracking() {
-    // 600 distinct characters in the pattern: more than the masks kept, so the rest are rebuilt.
-    let chars: Vec<char> = (0x4E00u32..0x4E00 + 600).filter_map(char::from_u32).collect();
-    let word: String = chars.iter().collect();
-    let pat = format!("*{word}?{}", chars[3]);
-    for text in [format!("x{word}yq{}", chars[3]), word.clone(), format!("{word}{word}"), chars.iter().rev().collect::<String>()] {
-        assert_eq!(wildcard(&text, &pat), wildcard_backtracking(&text, &pat), "{} chars", text.chars().count());
-    }
-    assert!(wildcard(&format!("{word}zz{}", chars[3]).replace("zz", "z"), &pat));
-}
-
-#[test]
-fn wildcard_on_long_repetitive_text_is_linear() {
-    // The backtracking matcher compared about text × pattern characters here (10^8).
-    let text = "a".repeat(20_000);
-    let pat = format!("*{}b", "a".repeat(5_000));
-    assert!(!wildcard(&text, &pat));
-    assert!(wildcard(&format!("{text}b"), &pat));
-}
-
 #[test]
 fn tsv() {
     assert_eq!(parse_tsv("a\tb\n1\t2\n"), vec![vec!["a", "b"], vec!["1", "2"]]);

crates/engine/src/cmd/print/draw.rs

@@ -228,6 +228,82 @@ fn draw_border(pg: &mut Page, b: &BorderLine, wb: &Workbook, start: Point, end:
     pg.stroke_line_dashed(Point::new(x0, y0), Point::new(x1, y1), lw, color, &border_dash(b.style, k));
 }
 
+/// Merged areas active at one row of a downward sweep, kept disjoint: by first column, their last
+/// column and last row. Adding, dropping and looking up a merge costs log(active merges).
+#[derive(Default)]
+struct MergeSweep {
+    cols: std::collections::BTreeMap<u32, (u32, u32)>,
+    /// (last row, first column) of the active merges, soonest ending first.
+    ends: std::collections::BinaryHeap<std::cmp::Reverse<(u32, u32)>>,
+}
+
+impl MergeSweep {
+    /// Moves down to `row` (rows never decrease), dropping the merges that end above it.
+    fn advance(&mut self, row: u32) {
+        while let Some(std::cmp::Reverse((end, col))) = self.ends.peek().copied() {
+            if end >= row {
+                break;
+            }
+            self.ends.pop();
+            self.cols.remove(&col);
+        }
+    }
+
+    /// Does an active merge meet columns `c0..=c1`? The intervals are disjoint, so only the last
+    /// one starting at or before `c1` can.
+    fn meets(&self, c0: u32, c1: u32) -> bool {
+        self.cols.range(..=c1).next_back().is_some_and(|(_, (end, _))| *end >= c0)
+    }
+
+    /// Adds merge `m`, at row `row`, unless it overlaps an active one.
+    fn add(&mut self, m: RangeRef, row: u32) -> bool {
+        self.advance(row);
+        if self.meets(m.start.col, m.end.col) {
+            return false;
+        }
+        self.cols.insert(m.start.col, (m.end.col, m.end.row));
+        self.ends.push(std::cmp::Reverse((m.end.row, m.start.col)));
+        true
+    }
+}
+
+/// The merged areas that meet `block`, ordered by the row where they enter it, then by column,
+/// leaving out any that overlaps one kept before it (only a damaged file has overlapping merges;
+/// Excel drops them when it repairs one). One sweep down the rows instead of a scan of every
+/// merge per cell.
+fn block_merges(sh: &Sheet, block: RangeRef) -> Vec<RangeRef> {
+    let top = |m: &RangeRef| m.start.row.max(block.start.row);
+    let mut merges: Vec<RangeRef> = sh.merges.iter().filter(|m| m.intersects(&block)).copied().collect();
+    merges.sort_by_key(|m| (top(m), m.start.col));
+    let mut sweep = MergeSweep::default();
+    merges.retain(|m| sweep.add(*m, top(m)));
+    merges
+}
+
+/// Answers "is this cell inside a merge?" for cells asked in row-major order, from the merges of
+/// [`block_merges`].
+struct Covered<'a> {
+    merges: &'a [RangeRef],
+    top: u32,
+    next: usize,
+    sweep: MergeSweep,
+}
+
+impl<'a> Covered<'a> {
+    fn new(merges: &'a [RangeRef], block: RangeRef) -> Self {
+        Self { merges, top: block.start.row, next: 0, sweep: MergeSweep::default() }
+    }
+
+    fn covers(&mut self, c: CellRef) -> bool {
+        while let Some(m) = self.merges.get(self.next).filter(|m| m.start.row.max(self.top) <= c.row) {
+            self.sweep.add(*m, m.start.row.max(self.top));
+            self.next += 1;
+        }
+        self.sweep.advance(c.row);
+        self.sweep.meets(c.col, c.col)
+    }
+}
+
 /// Draws the cells of a block: fills, gridlines, data bars, borders and text.
 pub(super) fn draw_block(pg: &mut Page, b: &Block, gridlines: bool, cf: &mut Option<crate::cf::CfCache>) {
     let (bx, by) = (b.xs.first().copied().unwrap_or(0.0), b.ys.first().copied().unwrap_or(0.0));
@@ -237,17 +313,18 @@ pub(super) fn draw_block(pg: &mut Page, b: &Block, gridlines: bool, cf: &mut Opt
     }
     let block = b.range();
     let sh = b.sh;
-    let merges: Vec<RangeRef> = sh.merges.iter().filter(|m| m.intersects(&block)).copied().collect();
-    let covered = |c: CellRef| merges.iter().any(|m| m.contains(c));
-    // Cells worth visiting.
+    let merges = block_merges(sh, block);
+    let by_start: std::collections::HashMap<CellRef, RangeRef> = merges.iter().map(|m| (m.start, *m)).collect();
+    // Cells worth visiting, in row-major order (which `Covered` needs).
+    let mut covered = Covered::new(&merges, block);
     let cells: Vec<CellRef> = if block.count() <= DENSE_BLOCK {
-        block.iter().filter(|c| !covered(*c)).collect()
+        block.iter().filter(|c| !covered.covers(*c)).collect()
     } else {
         let mut set: Vec<CellRef> = sh.cells.iter_range(block).map(|(c, _)| c).collect();
         set.extend(sh.spill.range(block.start..=block.end).map(|(c, _)| *c).filter(|c| block.contains(*c)));
         set.sort_unstable();
         set.dedup();
-        set.retain(|c| !covered(*c));
+        set.retain(|c| !covered.covers(*c));
         set
     };
     pg.save();
@@ -277,14 +354,14 @@ pub(super) fn draw_block(pg: &mut Page, b: &Block, gridlines: bool, cf: &mut Opt
         .filter(|(c, l)| {
             l.fill.is_some()
                 || l.cf.is_some()
-                || merges.iter().any(|m| m.start == *c)
+                || by_start.contains_key(c)
                 || !sh.value(*c).is_empty()
                 || l.style.borders != Borders::default()
                 || l.style.num_fmt.as_str() == "checkbox"
         })
         .collect();
     let rect_of = |c: CellRef| -> (f32, f32, f32, f32, bool) {
-        match merges.iter().find(|m| m.start == c) {
+        match by_start.get(&c) {
             Some(m) => {
                 let (x, y, w, h) = b.rect(*m);
                 (x, y, w, h, true)
@@ -322,7 +399,7 @@ pub(super) fn draw_block(pg: &mut Page, b: &Block, gridlines: bool, cf: &mut Opt
             continue;
         }
         let bd = &l.style.borders;
-        let (right, bottom) = match merges.iter().find(|m| m.start == *c).filter(|_| merged) {
+        let (right, bottom) = match by_start.get(c).filter(|_| merged) {
             Some(m) => (
                 b.wb.styles.get(sh.style_id(CellRef::new(m.start.row, m.end.col))).borders.right,
                 b.wb.styles.get(sh.style_id(CellRef::new(m.end.row, m.start.col))).borders.bottom,
@@ -407,11 +484,12 @@ pub(super) struct PrintedBlock {
 pub(super) fn draw_block_links(pg: &mut Page, b: &Block, page: usize, pending: &mut Vec<PendingPdfLink>) {
     let range = b.range();
     let (Some(&bx), Some(&by), Some(&right), Some(&bottom)) = (b.xs.first(), b.ys.first(), b.xs.last(), b.ys.last()) else { return };
+    let by_start: std::collections::HashMap<CellRef, RangeRef> = block_merges(b.sh, range).into_iter().map(|m| (m.start, m)).collect();
     for (cell, link) in b.sh.hyperlinks.range(range.start..=range.end) {
         if !range.contains(*cell) {
             continue;
         }
-        let cell_range = b.sh.merges.iter().find(|m| m.start == *cell).copied().unwrap_or(RangeRef::cell(*cell));
+        let cell_range = by_start.get(cell).copied().unwrap_or(RangeRef::cell(*cell));
         let (x, y, w, h) = b.rect(cell_range);
         let (x0, y0, x1, y1) = (x.max(bx), y.max(by), (x + w).min(right), (y + h).min(bottom));
         if x1 <= x0 || y1 <= y0 {
@@ -503,3 +581,44 @@ pub(super) fn make_block<'a>(
     }
     Block { wb, si, sh, rows, cols, xs, ys, k, scale: lay.scale }
 }
+
+#[cfg(test)]
+mod merge_tests {
+    use super::{CellRef, Covered, RangeRef, Sheet, block_merges};
+
+    fn r(a: &str) -> RangeRef {
+        RangeRef::parse(a).unwrap()
+    }
+
+    #[test]
+    fn block_merges_drop_overlaps_and_cover_like_a_scan() {
+        let mut sh = Sheet::new("S");
+        sh.merges = vec![r("B2:C3"), r("E1:F8"), r("B2:B5"), r("A10:D10"), r("H3:H4"), r("Z1:Z2"), r("C3:D4"), r("A1:A1048576")];
+        let block = r("B1:H12");
+        let merges = block_merges(&sh, block);
+        // B2:B5 and C3:D4 overlap B2:C3, which comes first; Z1:Z2 is outside the block.
+        assert_eq!(merges, vec![r("E1:F8"), r("B2:C3"), r("H3:H4"), r("A10:D10")]);
+        let mut covered = Covered::new(&merges, block);
+        for c in block.iter() {
+            assert_eq!(covered.covers(c), merges.iter().any(|m| m.contains(c)), "{}", c.a1());
+        }
+        // Cells asked sparsely, still in order.
+        let mut covered = Covered::new(&merges, block);
+        for (a, want) in [("C1", false), ("F5", true), ("G5", false), ("D10", true), ("H12", false)] {
+            assert_eq!(covered.covers(CellRef::parse(a).unwrap()), want, "{a}");
+        }
+    }
+
+    #[test]
+    fn many_stacked_merges_are_swept_once() {
+        // 50,000 merges over the same cells: one is kept, and the block is checked against it
+        // without comparing every cell with every merge.
+        let mut sh = Sheet::new("S");
+        sh.merges = (0..50_000u32).map(|i| RangeRef::new(CellRef::new(0, 0), CellRef::new(199, 1 + i % 50))).collect();
+        let block = r("A1:AZ200");
+        let merges = block_merges(&sh, block);
+        assert_eq!(merges, vec![r("A1:B200")]);
+        let mut covered = Covered::new(&merges, block);
+        assert_eq!(block.iter().filter(|c| covered.covers(*c)).count(), 400);
+    }
+}