Override QueryIter::fold to port Query::for_each perf gains to select Iterator combinators

crates/bevy_ecs/src/query/iter.rs

@@ -1,11 +1,11 @@
 use crate::{
-    archetype::{ArchetypeEntity, ArchetypeId, Archetypes},
+    archetype::{Archetype, ArchetypeEntity, ArchetypeId, Archetypes},
     entity::{Entities, Entity},
     prelude::World,
     query::{ArchetypeFilter, DebugCheckedUnwrap, QueryState, WorldQuery},
-    storage::{TableId, TableRow, Tables},
+    storage::{Table, TableId, TableRow, Tables},
 };
-use std::{borrow::Borrow, iter::FusedIterator, marker::PhantomData, mem::MaybeUninit};
+use std::{borrow::Borrow, iter::FusedIterator, marker::PhantomData, mem::MaybeUninit, ops::Range};
 
 use super::ReadOnlyWorldQuery;
 
@@ -39,6 +39,146 @@ impl<'w, 's, Q: WorldQuery, F: ReadOnlyWorldQuery> QueryIter<'w, 's, Q, F> {
             cursor: QueryIterationCursor::init(world, query_state, last_change_tick, change_tick),
         }
     }
+
+    /// Executes the equivalent of [`Iterator::fold`] over a contiguous segment
+    /// from an archetype.
+    ///
+    /// # Safety
+    ///  - all `rows` must be in `[0, tables.entity_count)`.
+    ///  - `table` must match Q and F
+    ///  - Both `Q::IS_DENSE` and `F::IS_DENSE` must be true.
+    #[inline]
+    pub(super) unsafe fn fold_table<B, Func>(
+        &mut self,
+        mut accum: B,
+        func: &mut Func,
+        table: &'w Table,
+        rows: Range<usize>,
+    ) -> B
+    where
+        Func: FnMut(B, Q::Item<'w>) -> B,
+    {
+        Q::set_table(&mut self.cursor.fetch, &self.query_state.fetch_state, table);
+        F::set_table(
+            &mut self.cursor.filter,
+            &self.query_state.filter_state,
+            table,
+        );
+
+        let entities = table.entities();
+        for row in rows {
+            // SAFETY: set_table was called prior.
+            // `current_row` is a table row in range of the current table, because if it was not, then the if above would have been executed.
+            let entity = entities.get_unchecked(row);
+            let row = TableRow::new(row);
+            if !F::filter_fetch(&mut self.cursor.filter, *entity, row) {
+                continue;
+            }
+
+            // SAFETY: set_table was called prior.
+            // `current_row` is a table row in range of the current table, because if it was not, then the if above would have been executed.
+            let item = Q::fetch(&mut self.cursor.fetch, *entity, row);
+
+            accum = func(accum, item);
+        }
+        accum
+    }
+
+    /// Executes the equivalent of [`Iterator::fold`] over a contiguous segment
+    /// from an archetype.
+    ///
+    /// # Safety
+    ///  - all `indicies` must be in `[0, archetype.len())`.
+    ///  - `archetype` must match Q and F
+    ///  - Either `Q::IS_DENSE` or `F::IS_DENSE` must be false.
+    #[inline]
+    pub(super) unsafe fn fold_archetype<B, Func>(
+        &mut self,
+        mut accum: B,
+        func: &mut Func,
+        archetype: &'w Archetype,
+        indices: Range<usize>,
+    ) -> B
+    where
+        Func: FnMut(B, Q::Item<'w>) -> B,
+    {
+        let table = self.tables.get(archetype.table_id()).debug_checked_unwrap();
+        Q::set_archetype(
+            &mut self.cursor.fetch,
+            &self.query_state.fetch_state,
+            archetype,
+            table,
+        );
+        F::set_archetype(
+            &mut self.cursor.filter,
+            &self.query_state.filter_state,
+            archetype,
+            table,
+        );
+
+        let entities = archetype.entities();
+        for index in indices {
+            // SAFETY: set_archetype was called prior.
+            // `current_row` is an archetype index row in range of the current archetype, because if it was not, then the if above would have been executed.
+            let archetype_entity = entities.get_unchecked(index);
+            if !F::filter_fetch(
+                &mut self.cursor.filter,
+                archetype_entity.entity(),
+                archetype_entity.table_row(),
+            ) {
+                continue;
+            }
+
+            // SAFETY: set_archetype was called prior, `current_row` is an archetype index in range of the current archetype
+            // `current_row` is an archetype index row in range of the current archetype, because if it was not, then the if above would have been executed.
+            let item = Q::fetch(
+                &mut self.cursor.fetch,
+                archetype_entity.entity(),
+                archetype_entity.table_row(),
+            );
+
+            accum = func(accum, item);
+        }
+        accum
+    }
+
+    /// Executes the equivalent of [`Iterator::for_each`] over a contiguous segment
+    /// from an table.
+    ///
+    /// # Safety
+    ///  - all `rows` must be in `[0, tables.entity_count)`.
+    ///  - `table` must match Q and F
+    ///  - Both `Q::IS_DENSE` and `F::IS_DENSE` must be true.
+    #[inline]
+    pub(super) unsafe fn for_each_table<Func>(
+        &mut self,
+        func: &mut Func,
+        table: &'w Table,
+        rows: Range<usize>,
+    ) where
+        Func: FnMut(Q::Item<'w>),
+    {
+        self.fold_table((), &mut |_, item| func(item), table, rows);
+    }
+
+    /// Executes the equivalent of [`Iterator::for_each`] over a contiguous segment
+    /// from an archetype.
+    ///
+    /// # Safety
+    ///  - all `indicies` must be in `[0, archetype.len())`.
+    ///  - `archetype` must match Q and F
+    ///  - Either `Q::IS_DENSE` or `F::IS_DENSE` must be false.
+    #[inline]
+    pub(super) unsafe fn for_each_archetype<Func>(
+        &mut self,
+        func: &mut Func,
+        archetype: &'w Archetype,
+        indices: Range<usize>,
+    ) where
+        Func: FnMut(Q::Item<'w>),
+    {
+        self.fold_archetype((), &mut |_, item| func(item), archetype, indices);
+    }
 }
 
 impl<'w, 's, Q: WorldQuery, F: ReadOnlyWorldQuery> Iterator for QueryIter<'w, 's, Q, F> {
@@ -61,6 +201,38 @@ impl<'w, 's, Q: WorldQuery, F: ReadOnlyWorldQuery> Iterator for QueryIter<'w, 's
         let min_size = if archetype_query { max_size } else { 0 };
         (min_size, Some(max_size))
     }
+
+    #[inline]
+    fn fold<B, Func>(mut self, init: B, mut func: Func) -> B
+    where
+        Func: FnMut(B, Self::Item) -> B,
+    {
+        let mut accum = init;
+        // Empty any remaining uniterated values from the current table/archetype
+        while self.cursor.current_row != self.cursor.current_len {
+            let Some(item) = self.next() else { break };
+            accum = func(accum, item);
+        }
+        if Q::IS_DENSE && F::IS_DENSE {
+            for table_id in self.cursor.table_id_iter.clone() {
+                // SAFETY: Matched table IDs are guarenteed to still exist.
+                let table = unsafe { self.tables.get(*table_id).debug_checked_unwrap() };
+                accum =
+                    // SAFETY: The fetched table matches the query
+                    unsafe { self.fold_table(accum, &mut func, table, 0..table.entity_count()) };
+            }
+        } else {
+            for archetype_id in self.cursor.archetype_id_iter.clone() {
+                let archetype =
+                    // SAFETY: Matched archetype IDs are guarenteed to still exist.
+                    unsafe { self.archetypes.get(*archetype_id).debug_checked_unwrap() };
+                accum =
+                    // SAFETY: The fetched archetype and table matches the query
+                    unsafe { self.fold_archetype(accum, &mut func, archetype, 0..archetype.len()) };
+            }
+        }
+        accum
+    }
 }
 
 // This is correct as [`QueryIter`] always returns `None` once exhausted.
@@ -623,9 +795,9 @@ impl<'w, 's, Q: WorldQuery, F: ReadOnlyWorldQuery> QueryIterationCursor<'w, 's,
                 if self.current_row == self.current_len {
                     let archetype_id = self.archetype_id_iter.next()?;
                     let archetype = archetypes.get(*archetype_id).debug_checked_unwrap();
+                    let table = tables.get(archetype.table_id()).debug_checked_unwrap();
                     // SAFETY: `archetype` and `tables` are from the world that `fetch/filter` were created for,
                     // `fetch_state`/`filter_state` are the states that `fetch/filter` were initialized with
-                    let table = tables.get(archetype.table_id()).debug_checked_unwrap();
                     Q::set_archetype(&mut self.fetch, &query_state.fetch_state, archetype, table);
                     F::set_archetype(
                         &mut self.filter,

crates/bevy_ecs/src/query/par_iter.rs

@@ -148,12 +148,13 @@ impl<'w, 's, Q: WorldQuery, F: ReadOnlyWorldQuery> QueryParIter<'w, 's, Q, F> {
     ) {
         let thread_count = ComputeTaskPool::get().thread_num();
         if thread_count <= 1 {
-            self.state.for_each_unchecked_manual(
-                self.world,
-                func,
-                self.world.last_change_tick(),
-                self.world.read_change_tick(),
-            );
+            self.state
+                .iter_unchecked_manual(
+                    self.world,
+                    self.world.last_change_tick(),
+                    self.world.read_change_tick(),
+                )
+                .for_each(func);
         } else {
             // Need a batch size of at least 1.
             let batch_size = self.get_batch_size(thread_count).max(1);