1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150

use crate::fx::{FxHashMap, FxHasher};
use crate::sync::{Lock, LockGuard};
use std::borrow::Borrow;
use std::collections::hash_map::RawEntryMut;
use std::hash::{Hash, Hasher};
use std::mem;

#[derive(Clone, Default)]
#[cfg_attr(parallel_compiler, repr(align(64)))]
struct CacheAligned<T>(T);

#[cfg(parallel_compiler)]
// 32 shards is sufficient to reduce contention on an 8-core Ryzen 7 1700,
// but this should be tested on higher core count CPUs. How the `Sharded` type gets used
// may also affect the ideal number of shards.
const SHARD_BITS: usize = 5;

#[cfg(not(parallel_compiler))]
const SHARD_BITS: usize = 0;

pub const SHARDS: usize = 1 << SHARD_BITS;

/// An array of cache-line aligned inner locked structures with convenience methods.
#[derive(Clone)]
pub struct Sharded<T> {
    shards: [CacheAligned<Lock<T>>; SHARDS],
}

impl<T: Default> Default for Sharded<T> {
    #[inline]
    fn default() -> Self {
        Self::new(T::default)
    }
}

impl<T> Sharded<T> {
    #[inline]
    pub fn new(mut value: impl FnMut() -> T) -> Self {
        Sharded { shards: [(); SHARDS].map(|()| CacheAligned(Lock::new(value()))) }
    }

    /// The shard is selected by hashing `val` with `FxHasher`.
    #[inline]
    pub fn get_shard_by_value<K: Hash + ?Sized>(&self, val: &K) -> &Lock<T> {
        if SHARDS == 1 { &self.shards[0].0 } else { self.get_shard_by_hash(make_hash(val)) }
    }

    #[inline]
    pub fn get_shard_by_hash(&self, hash: u64) -> &Lock<T> {
        &self.shards[get_shard_index_by_hash(hash)].0
    }

    #[inline]
    pub fn get_shard_by_index(&self, i: usize) -> &Lock<T> {
        &self.shards[i].0
    }

    pub fn lock_shards(&self) -> Vec<LockGuard<'_, T>> {
        (0..SHARDS).map(|i| self.shards[i].0.lock()).collect()
    }

    pub fn try_lock_shards(&self) -> Option<Vec<LockGuard<'_, T>>> {
        (0..SHARDS).map(|i| self.shards[i].0.try_lock()).collect()
    }
}

pub type ShardedHashMap<K, V> = Sharded<FxHashMap<K, V>>;

impl<K: Eq, V> ShardedHashMap<K, V> {
    pub fn len(&self) -> usize {
        self.lock_shards().iter().map(|shard| shard.len()).sum()
    }
}

impl<K: Eq + Hash + Copy> ShardedHashMap<K, ()> {
    #[inline]
    pub fn intern_ref<Q: ?Sized>(&self, value: &Q, make: impl FnOnce() -> K) -> K
    where
        K: Borrow<Q>,
        Q: Hash + Eq,
    {
        let hash = make_hash(value);
        let mut shard = self.get_shard_by_hash(hash).lock();
        let entry = shard.raw_entry_mut().from_key_hashed_nocheck(hash, value);

        match entry {
            RawEntryMut::Occupied(e) => *e.key(),
            RawEntryMut::Vacant(e) => {
                let v = make();
                e.insert_hashed_nocheck(hash, v, ());
                v
            }
        }
    }

    #[inline]
    pub fn intern<Q>(&self, value: Q, make: impl FnOnce(Q) -> K) -> K
    where
        K: Borrow<Q>,
        Q: Hash + Eq,
    {
        let hash = make_hash(&value);
        let mut shard = self.get_shard_by_hash(hash).lock();
        let entry = shard.raw_entry_mut().from_key_hashed_nocheck(hash, &value);

        match entry {
            RawEntryMut::Occupied(e) => *e.key(),
            RawEntryMut::Vacant(e) => {
                let v = make(value);
                e.insert_hashed_nocheck(hash, v, ());
                v
            }
        }
    }
}

pub trait IntoPointer {
    /// Returns a pointer which outlives `self`.
    fn into_pointer(&self) -> *const ();
}

impl<K: Eq + Hash + Copy + IntoPointer> ShardedHashMap<K, ()> {
    pub fn contains_pointer_to<T: Hash + IntoPointer>(&self, value: &T) -> bool {
        let hash = make_hash(&value);
        let shard = self.get_shard_by_hash(hash).lock();
        let value = value.into_pointer();
        shard.raw_entry().from_hash(hash, |entry| entry.into_pointer() == value).is_some()
    }
}

#[inline]
pub fn make_hash<K: Hash + ?Sized>(val: &K) -> u64 {
    let mut state = FxHasher::default();
    val.hash(&mut state);
    state.finish()
}

/// Get a shard with a pre-computed hash value. If `get_shard_by_value` is
/// ever used in combination with `get_shard_by_hash` on a single `Sharded`
/// instance, then `hash` must be computed with `FxHasher`. Otherwise,
/// `hash` can be computed with any hasher, so long as that hasher is used
/// consistently for each `Sharded` instance.
#[inline]
pub fn get_shard_index_by_hash(hash: u64) -> usize {
    let hash_len = mem::size_of::<usize>();
    // Ignore the top 7 bits as hashbrown uses these and get the next SHARD_BITS highest bits.
    // hashbrown also uses the lowest bits, so we can't use those
    let bits = (hash >> (hash_len * 8 - 7 - SHARD_BITS)) as usize;
    bits % SHARDS
}