dfir_rs/compiled/pull/half_join_state/
multiset.rs

1use std::collections::VecDeque;
2use std::collections::hash_map::Entry;
3
4use super::HalfJoinState;
5use crate::util::clear::Clear;
6
7type HashMap<K, V> = rustc_hash::FxHashMap<K, V>;
8
9use smallvec::{SmallVec, smallvec};
10#[derive(Debug)]
11pub struct HalfMultisetJoinState<Key, ValBuild, ValProbe> {
12    // Here a smallvec with inline storage of 1 is chosen.
13    // The rationale for this decision is that, I speculate, that joins possibly have a bimodal distribution with regards to how much key contention they have.
14    // That is, I think that there are many joins that have 1 value per key on LHS/RHS, and there are also a large category of joins that have multiple values per key.
15    // For the category of joins that have multiple values per key, it's not clear why they would only have 2, 3, 4, or N specific number of values per key. So there's no good number to set the smallvec storage to.
16    // Instead we can just focus on the first group of joins that have 1 value per key and get benefit there without hurting the other group too much with excessive memory usage.
17    /// Table to probe, vec val contains all matches.
18    table: HashMap<Key, SmallVec<[ValBuild; 1]>>,
19    /// Not-yet emitted matches.
20    current_matches: VecDeque<(Key, ValProbe, ValBuild)>,
21    len: usize,
22}
23impl<Key, ValBuild, ValProbe> Default for HalfMultisetJoinState<Key, ValBuild, ValProbe> {
24    fn default() -> Self {
25        Self {
26            table: HashMap::default(),
27            current_matches: VecDeque::default(),
28            len: 0,
29        }
30    }
31}
32impl<Key, ValBuild, ValProbe> Clear for HalfMultisetJoinState<Key, ValBuild, ValProbe> {
33    fn clear(&mut self) {
34        self.table.clear();
35        self.current_matches.clear();
36        self.len = 0;
37    }
38}
39impl<Key, ValBuild, ValProbe> HalfJoinState<Key, ValBuild, ValProbe>
40    for HalfMultisetJoinState<Key, ValBuild, ValProbe>
41where
42    Key: Clone + Eq + std::hash::Hash,
43    ValBuild: Clone,
44    ValProbe: Clone,
45{
46    fn build(&mut self, k: Key, v: &ValBuild) -> bool {
47        let entry = self.table.entry(k);
48
49        match entry {
50            Entry::Occupied(mut e) => {
51                let vec = e.get_mut();
52
53                vec.push(v.clone());
54                self.len += 1;
55            }
56            Entry::Vacant(e) => {
57                e.insert(smallvec![v.clone()]);
58                self.len += 1;
59            }
60        };
61
62        true
63    }
64
65    fn probe(&mut self, k: &Key, v: &ValProbe) -> Option<(Key, ValProbe, ValBuild)> {
66        // TODO: We currently don't free/shrink the self.current_matches vecdeque to save time.
67        // This mean it will grow to eventually become the largest number of matches in a single probe call.
68        // Maybe we should clear this memory at the beginning of every tick/periodically?
69        let mut iter = self
70            .table
71            .get(k)?
72            .iter()
73            .map(|valbuild| (k.clone(), v.clone(), valbuild.clone()));
74
75        let first = iter.next();
76
77        self.current_matches.extend(iter);
78
79        first
80    }
81
82    fn full_probe(&self, k: &Key) -> std::slice::Iter<'_, ValBuild> {
83        let Some(sv) = self.table.get(k) else {
84            return [].iter();
85        };
86
87        sv.iter()
88    }
89
90    fn pop_match(&mut self) -> Option<(Key, ValProbe, ValBuild)> {
91        self.current_matches.pop_front()
92    }
93
94    fn len(&self) -> usize {
95        self.len
96    }
97    fn iter(&self) -> std::collections::hash_map::Iter<'_, Key, SmallVec<[ValBuild; 1]>> {
98        self.table.iter()
99    }
100}
dfir_rs/compiled/pull/half_join_state/multiset.rs

dfir_rs/compiled/pull/half_join_state/
multiset.rs
