formualizer-eval 0.8.4

use std::collections::{BTreeMap, HashSet};
use std::ops::ControlFlow;

/// Custom interval tree optimized for spreadsheet cell indexing.
///
/// ## Design decisions:
///
/// 1. **Point intervals are the common case** - Most cells are single points [r,r] or [c,c]
/// 2. **Sparse data** - Even million-row sheets typically have <10K cells
/// 3. **Batch updates** - During shifts, we update many intervals at once
/// 4. **Small value sets** - Each interval maps to a small set of VertexIds
///
/// ## Implementation:
///
/// Uses an augmented BST where each node stores:
/// - Interval [low, high]
/// - Max endpoint in subtree (for efficient pruning)
/// - Value set (HashSet<VertexId>)
///
/// This is simpler than generic interval trees because we optimize for our specific use case.

#[derive(Debug, Clone)]
struct IntervalNode<T: Clone + Eq + std::hash::Hash> {
    high: u32,
    values: HashSet<T>,
}

/// B-Tree based implementation of the interval index.
#[derive(Debug, Clone)]
pub struct IntervalTree<T: Clone + Eq + std::hash::Hash> {
    /// Maps low coordinate to a set of intervals/values starting there.
    /// Internal storage uses IntervalNode, NOT Entry.
    map: BTreeMap<u32, Vec<IntervalNode<T>>>,
    size: usize,
}

impl<T: Clone + Eq + std::hash::Hash> Default for IntervalTree<T> {
    fn default() -> Self {
        Self::new()
    }
}

impl<T: Clone + Eq + std::hash::Hash> IntervalTree<T> {
    pub fn new() -> Self {
        Self {
            map: BTreeMap::new(),
            size: 0,
        }
    }

    pub fn len(&self) -> usize {
        self.size
    }

    pub fn is_empty(&self) -> bool {
        self.size == 0
    }

    /// Get a mutable reference to the values for an exact interval match.
    /// Required by the Entry API.
    pub fn get_mut(&mut self, low: u32, high: u32) -> Option<&mut HashSet<T>> {
        self.map.get_mut(&low).and_then(|nodes| {
            nodes
                .iter_mut()
                .find(|n| n.high == high)
                .map(|n| &mut n.values)
        })
    }

    /// Insert a value for the given interval [low, high]
    pub fn insert(&mut self, low: u32, high: u32, value: T) {
        let entries = self.map.entry(low).or_default();

        if let Some(node) = entries.iter_mut().find(|n| n.high == high) {
            node.values.insert(value);
        } else {
            let mut values = HashSet::new();
            values.insert(value);
            entries.push(IntervalNode { high, values });
            self.size += 1;
        }
    }

    pub fn query(&self, q_low: u32, q_high: u32) -> Vec<(u32, u32, HashSet<T>)> {
        let mut results = Vec::new();
        for (&low, nodes) in self.map.range(..=q_high) {
            for node in nodes {
                if node.high >= q_low {
                    results.push((low, node.high, node.values.clone()));
                }
            }
        }
        results
    }

    /// Returns the values stored at the exact point interval `[point, point]`.
    ///
    /// Unlike [`Self::query`], this performs a direct B-tree lookup and therefore
    /// does not inspect intervals whose lower bound precedes `point`.
    pub(crate) fn point_values(&self, point: u32) -> Option<&HashSet<T>> {
        self.map.get(&point).and_then(|nodes| {
            nodes
                .iter()
                .find(|node| node.high == point)
                .map(|node| &node.values)
        })
    }

    /// Visits point intervals whose coordinate is in the inclusive query range.
    ///
    /// This is a specialized path for indexes that contain only `[x, x]`
    /// intervals. General overlapping-range callers must continue to use
    /// [`Self::query`] or [`Self::visit_query`]. `None` is emitted once for each
    /// point-interval node inspected so callers can account deterministic visits.
    pub(crate) fn visit_point_intervals(
        &self,
        q_low: u32,
        q_high: u32,
        mut visitor: impl FnMut(Option<&T>) -> ControlFlow<()>,
    ) -> ControlFlow<()> {
        if q_low > q_high {
            return ControlFlow::Continue(());
        }
        for (&low, nodes) in self.map.range(q_low..=q_high) {
            for node in nodes {
                if node.high != low {
                    continue;
                }
                visitor(None)?;
                for value in &node.values {
                    visitor(Some(value))?;
                }
            }
        }
        ControlFlow::Continue(())
    }

    /// Counts point-interval nodes and values in an inclusive coordinate range.
    /// This has the same point-only caller contract as [`Self::visit_point_intervals`].
    pub(crate) fn point_interval_stats(&self, q_low: u32, q_high: u32) -> (usize, usize) {
        if q_low > q_high {
            return (0, 0);
        }
        let mut node_count = 0usize;
        let mut value_count = 0usize;
        for (&low, nodes) in self.map.range(q_low..=q_high) {
            for node in nodes {
                if node.high == low {
                    node_count = node_count.saturating_add(1);
                    value_count = value_count.saturating_add(node.values.len());
                }
            }
        }
        (node_count, value_count)
    }

    /// Visits matching values without cloning or materializing the query.
    /// Returning `Break` from the visitor stops traversal immediately.
    pub(crate) fn visit_query(
        &self,
        q_low: u32,
        q_high: u32,
        mut visitor: impl FnMut(Option<&T>) -> ControlFlow<()>,
    ) -> ControlFlow<()> {
        for (_low, nodes) in self.map.range(..=q_high) {
            for node in nodes {
                visitor(None)?;
                if node.high < q_low {
                    continue;
                }
                for value in &node.values {
                    visitor(Some(value))?;
                }
            }
        }
        ControlFlow::Continue(())
    }

    pub(crate) fn estimated_heap_bytes(&self) -> Option<usize> {
        const NODE_ALLOCATION_OVERHEAD: usize = 3 * std::mem::size_of::<usize>();
        let mut bytes = 0usize;
        for nodes in self.map.values() {
            bytes = bytes.checked_add(
                std::mem::size_of::<u32>()
                    .checked_add(std::mem::size_of::<Vec<IntervalNode<T>>>())?
                    .checked_add(NODE_ALLOCATION_OVERHEAD)?,
            )?;
            bytes = bytes.checked_add(
                nodes
                    .capacity()
                    .checked_mul(std::mem::size_of::<IntervalNode<T>>())?,
            )?;
            for node in nodes {
                bytes = bytes.checked_add(node.values.capacity().checked_mul(
                    std::mem::size_of::<T>().checked_add(std::mem::size_of::<usize>())?,
                )?)?;
            }
        }
        Some(bytes)
    }

    pub fn remove(&mut self, low: u32, high: u32, value: &T) -> bool {
        if let Some(nodes) = self.map.get_mut(&low)
            && let Some(node) = nodes.iter_mut().find(|n| n.high == high)
        {
            let removed = node.values.remove(value);

            if removed && node.values.is_empty() {
                nodes.retain(|n| n.high != high);
                self.size -= 1;
                if nodes.is_empty() {
                    self.map.remove(&low);
                }
            }
            return removed;
        }
        false
    }

    pub fn entry(&mut self, low: u32, high: u32) -> BTreeEntry<'_, T> {
        BTreeEntry {
            tree: self,
            low,
            high,
        }
    }

    /// Bulk build optimization for a collection of point intervals [x,x].
    pub fn bulk_build_points(&mut self, mut items: Vec<(u32, HashSet<T>)>) {
        if !self.is_empty() {
            // Fallback: incremental insert to preserve existing nodes
            for (coord, set) in items {
                for val in set {
                    self.insert(coord, coord, val);
                }
            }
            return;
        }

        if items.is_empty() {
            return;
        }

        // 1. Sort by coordinate
        items.sort_by_key(|(k, _)| *k);

        // 2. Process items. BTreeMap handles the balancing (O(log N)).
        for (coord, set) in items {
            let entries = self.map.entry(coord).or_default();

            // Since this is specifically for point intervals, check if [coord, coord] exists
            if let Some(node) = entries.iter_mut().find(|n| n.high == coord) {
                node.values.extend(set);
            } else {
                entries.push(IntervalNode {
                    high: coord,
                    values: set,
                });
                self.size += 1;
            }
        }
    }
}

pub struct BTreeEntry<'a, T: Clone + Eq + std::hash::Hash> {
    tree: &'a mut IntervalTree<T>,
    low: u32,
    high: u32,
}

impl<'a, T: Clone + Eq + std::hash::Hash> BTreeEntry<'a, T> {
    pub fn or_insert_with<F>(self, f: F) -> &'a mut HashSet<T>
    where
        F: FnOnce() -> HashSet<T>,
    {
        if self.tree.get_mut(self.low, self.high).is_none() {
            let values = f();
            let entries = self.tree.map.entry(self.low).or_default();
            entries.push(IntervalNode {
                high: self.high,
                values,
            });
            self.tree.size += 1;
        }
        self.tree.get_mut(self.low, self.high).unwrap()
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_insert_and_query_point_interval() {
        let mut tree = IntervalTree::new();
        tree.insert(5, 5, 100);

        let results = tree.query(5, 5);
        assert_eq!(results.len(), 1);
        assert_eq!(results[0].0, 5);
        assert_eq!(results[0].1, 5);
        assert!(results[0].2.contains(&100));
    }

    #[test]
    fn test_insert_and_query_range() {
        let mut tree = IntervalTree::new();
        tree.insert(10, 20, 1);
        tree.insert(15, 25, 2);
        tree.insert(30, 40, 3);

        // Query overlapping with first two intervals
        let results = tree.query(12, 22);
        assert_eq!(results.len(), 2);

        // Query overlapping with only the third interval
        let results = tree.query(35, 45);
        assert_eq!(results.len(), 1);
        assert!(results[0].2.contains(&3));
    }

    #[test]
    fn point_interval_visit_does_not_scan_coordinate_prefixes() {
        let mut tree = IntervalTree::new();
        for coordinate in 0..10_000 {
            tree.insert(coordinate, coordinate, coordinate);
        }

        let mut nodes = 0;
        let mut values = Vec::new();
        let result = tree.visit_point_intervals(9_999, 9_999, |entry| {
            match entry {
                None => nodes += 1,
                Some(value) => values.push(*value),
            }
            ControlFlow::Continue(())
        });

        assert_eq!(result, ControlFlow::Continue(()));
        assert_eq!(nodes, 1);
        assert_eq!(values, vec![9_999]);
    }

    #[test]
    fn point_interval_path_does_not_change_general_overlap_queries() {
        let mut tree = IntervalTree::new();
        tree.insert(1, 100, "range");
        tree.insert(75, 75, "point");

        let general = tree.query(75, 75);
        assert_eq!(general.len(), 2);
        assert!(general.iter().any(|entry| entry.2.contains("range")));
        assert!(general.iter().any(|entry| entry.2.contains("point")));

        let mut point_values = Vec::new();
        let _ = tree.visit_point_intervals(75, 75, |entry| {
            if let Some(value) = entry {
                point_values.push(*value);
            }
            ControlFlow::Continue(())
        });
        assert_eq!(point_values, vec!["point"]);
    }

    #[test]
    fn test_remove_value() {
        let mut tree = IntervalTree::new();
        tree.insert(5, 5, 100);
        tree.insert(5, 5, 200);

        assert_eq!(tree.query(5, 5).len(), 1);
        assert_eq!(tree.query(5, 5)[0].2.len(), 2);

        tree.remove(5, 5, &100);

        let results = tree.query(5, 5);
        assert_eq!(results.len(), 1);
        assert_eq!(results[0].2.len(), 1);
        assert!(results[0].2.contains(&200));
    }

    #[test]
    fn test_entry_api() {
        let mut tree: IntervalTree<i32> = IntervalTree::new();

        tree.entry(10, 10).or_insert_with(HashSet::new).insert(42);

        tree.entry(10, 10).or_insert_with(HashSet::new).insert(43);

        let results = tree.query(10, 10);
        assert_eq!(results.len(), 1);
        assert_eq!(results[0].2.len(), 2);
        assert!(results[0].2.contains(&42));
        assert!(results[0].2.contains(&43));
    }

    #[test]
    fn test_large_sparse_tree() {
        let mut tree = IntervalTree::new();

        // Simulate sparse spreadsheet
        for i in (0..1_000_000).step_by(10000) {
            tree.insert(i, i, i as i32);
        }

        assert_eq!(tree.len(), 100);

        // Query for high rows
        let results = tree.query(500_000, u32::MAX);
        assert_eq!(results.len(), 50);
    }

    #[test]
    fn test_entry_recursion_bug() {
        let mut tree: IntervalTree<u32> = IntervalTree::new();

        // The bug happens when we insert a value, then use entry()
        // on a coordinate that would be a child of that value.
        let count: u32 = 5000;
        for i in 0..count {
            tree.entry(i, i).or_insert_with(HashSet::new);
        }

        assert_eq!(tree.len(), count as usize);
    }

    #[test]
    fn test_complex_overlaps() {
        let mut tree = IntervalTree::new();
        // Nested intervals
        tree.insert(10, 100, "A");
        tree.insert(20, 50, "B");
        tree.insert(30, 40, "C");

        // Partially overlapping
        tree.insert(5, 15, "D");
        tree.insert(95, 105, "E");

        // Query for the very middle
        let results = tree.query(35, 35);
        assert_eq!(results.len(), 3); // Should hit A, B, and C

        // Query for a range that only hits the "tail" of the large interval and the "head" of the end interval
        let results = tree.query(98, 102);
        assert_eq!(results.len(), 2); // Should hit A and E
    }

    #[test]
    fn test_multiple_values_and_size() {
        let mut tree = IntervalTree::new();

        // Insert same interval twice with different values
        tree.insert(10, 10, "val1");
        tree.insert(10, 10, "val2");
        assert_eq!(tree.len(), 1); // Size should only count unique intervals

        // Insert same value twice
        tree.insert(10, 10, "val1");
        assert_eq!(tree.len(), 1);
        let results = tree.query(10, 10);
        assert_eq!(results[0].2.len(), 2); // HashSet handles the duplicate value "val1"
    }

    #[test]
    fn test_remove_edge_cases() {
        let mut tree = IntervalTree::new();
        tree.insert(10, 20, "A");

        // Try to remove a value that isn't there
        let removed = tree.remove(10, 20, &"B");
        assert!(!removed);
        assert_eq!(tree.query(10, 20)[0].2.len(), 1);

        // Try to remove from an interval that doesn't exist
        let removed = tree.remove(99, 100, &"A");
        assert!(!removed);
    }

    #[test]
    fn test_bulk_build_consistency() {
        let mut incremental_tree = IntervalTree::new();
        let mut bulk_tree = IntervalTree::new();

        let data: Vec<(u32, HashSet<&str>)> = vec![
            (10, vec!["A", "B"].into_iter().collect()),
            (20, vec!["C"].into_iter().collect()),
            (5, vec!["D"].into_iter().collect()),
        ];

        // Build incrementally
        for (coord, values) in &data {
            for val in values {
                incremental_tree.insert(*coord, *coord, *val);
            }
        }

        // Build using bulk
        bulk_tree.bulk_build_points(data.clone());

        // Compare results
        assert_eq!(incremental_tree.len(), bulk_tree.len());
        assert_eq!(incremental_tree.query(0, 100), bulk_tree.query(0, 100));
    }

    #[test]
    fn test_query_stack_safety() {
        let mut tree = IntervalTree::new();
        let count = 10_000;

        // Create a deep right-leaning tree
        for i in 0..count {
            tree.insert(i, i, i);
        }

        // Query the very end of the tree
        // If this causes a SIGABRT, it means query_node() must be made iterative
        let results = tree.query(count - 1, count - 1);
        assert_eq!(results.len(), 1);
    }

    #[test]
    fn test_empty_and_boundaries() {
        let mut tree: IntervalTree<i32> = IntervalTree::new();

        assert!(tree.is_empty());
        assert_eq!(tree.query(0, 100).len(), 0);
        assert!(!tree.remove(0, 0, &1));

        // Test a query that "misses" everything
        tree.insert(50, 60, 1);
        assert_eq!(tree.query(0, 49).len(), 0);
        assert_eq!(tree.query(61, 100).len(), 0);
    }

    #[test]
    fn test_multi_value_interval_size_tracking() {
        let mut tree = IntervalTree::new();
        let iv = (10, 20);

        // 1. Insert two values for the same interval
        // Destructure the tuple into low (iv.0) and high (iv.1)
        tree.insert(iv.0, iv.1, "A");
        tree.insert(iv.0, iv.1, "B");
        assert_eq!(tree.len(), 1, "Should be 1 unique interval");

        // 2. Remove first value - pass as reference &"A"
        assert!(tree.remove(iv.0, iv.1, &"A"));
        assert_eq!(
            tree.len(),
            1,
            "Should still be 1 interval after partial removal"
        );

        // 3. Remove second value - size should now be 0
        assert!(tree.remove(iv.0, iv.1, &"B"));
        assert_eq!(tree.len(), 0, "Should be 0 after last value removed");
    }
}