rustc_hash/
lib.rs

1//! A speedy, non-cryptographic hashing algorithm used by `rustc`.
2//!
3//! # Example
4//!
5//! ```rust
6//! # #[cfg(feature = "std")]
7//! # fn main() {
8//! use rustc_hash::FxHashMap;
9//!
10//! let mut map: FxHashMap<u32, u32> = FxHashMap::default();
11//! map.insert(22, 44);
12//! # }
13//! # #[cfg(not(feature = "std"))]
14//! # fn main() { }
15//! ```
16
17#![no_std]
18#![cfg_attr(feature = "nightly", feature(hasher_prefixfree_extras))]
19
20#[cfg(feature = "std")]
21extern crate std;
22
23#[cfg(feature = "rand")]
24extern crate rand;
25
26#[cfg(feature = "rand")]
27mod random_state;
28
29mod seeded_state;
30
31use core::default::Default;
32use core::hash::{BuildHasher, Hasher};
33#[cfg(feature = "std")]
34use std::collections::{HashMap, HashSet};
35
36/// Type alias for a hash map that uses the Fx hashing algorithm.
37#[cfg(feature = "std")]
38pub type FxHashMap<K, V> = HashMap<K, V, FxBuildHasher>;
39
40/// Type alias for a hash set that uses the Fx hashing algorithm.
41#[cfg(feature = "std")]
42pub type FxHashSet<V> = HashSet<V, FxBuildHasher>;
43
44#[cfg(feature = "rand")]
45pub use random_state::{FxHashMapRand, FxHashSetRand, FxRandomState};
46
47pub use seeded_state::FxSeededState;
48#[cfg(feature = "std")]
49pub use seeded_state::{FxHashMapSeed, FxHashSetSeed};
50
51/// A speedy hash algorithm for use within rustc. The hashmap in liballoc
52/// by default uses SipHash which isn't quite as speedy as we want. In the
53/// compiler we're not really worried about DOS attempts, so we use a fast
54/// non-cryptographic hash.
55///
56/// The current implementation is a fast polynomial hash with a single
57/// bit rotation as a finishing step designed by Orson Peters.
58#[derive(Clone)]
59pub struct FxHasher {
60    hash: usize,
61}
62
63// One might view a polynomial hash
64//    m[0] * k    + m[1] * k^2  + m[2] * k^3  + ...
65// as a multilinear hash with keystream k[..]
66//    m[0] * k[0] + m[1] * k[1] + m[2] * k[2] + ...
67// where keystream k just happens to be generated using a multiplicative
68// congruential pseudorandom number generator (MCG). For that reason we chose a
69// constant that was found to be good for a MCG in:
70//     "Computationally Easy, Spectrally Good Multipliers for Congruential
71//     Pseudorandom Number Generators" by Guy Steele and Sebastiano Vigna.
72#[cfg(target_pointer_width = "64")]
73const K: usize = 0xf1357aea2e62a9c5;
74#[cfg(target_pointer_width = "32")]
75const K: usize = 0x93d765dd;
76
77impl FxHasher {
78    /// Creates a `fx` hasher with a given seed.
79    pub const fn with_seed(seed: usize) -> FxHasher {
80        FxHasher { hash: seed }
81    }
82
83    /// Creates a default `fx` hasher.
84    pub const fn default() -> FxHasher {
85        FxHasher { hash: 0 }
86    }
87}
88
89impl Default for FxHasher {
90    #[inline]
91    fn default() -> FxHasher {
92        Self::default()
93    }
94}
95
96impl FxHasher {
97    #[inline]
98    fn add_to_hash(&mut self, i: usize) {
99        self.hash = self.hash.wrapping_add(i).wrapping_mul(K);
100    }
101}
102
103impl Hasher for FxHasher {
104    #[inline]
105    fn write(&mut self, bytes: &[u8]) {
106        // Compress the byte string to a single u64 and add to our hash.
107        self.write_u64(hash_bytes(bytes));
108    }
109
110    #[inline]
111    fn write_u8(&mut self, i: u8) {
112        self.add_to_hash(i as usize);
113    }
114
115    #[inline]
116    fn write_u16(&mut self, i: u16) {
117        self.add_to_hash(i as usize);
118    }
119
120    #[inline]
121    fn write_u32(&mut self, i: u32) {
122        self.add_to_hash(i as usize);
123    }
124
125    #[inline]
126    fn write_u64(&mut self, i: u64) {
127        self.add_to_hash(i as usize);
128        #[cfg(target_pointer_width = "32")]
129        self.add_to_hash((i >> 32) as usize);
130    }
131
132    #[inline]
133    fn write_u128(&mut self, i: u128) {
134        self.add_to_hash(i as usize);
135        #[cfg(target_pointer_width = "32")]
136        self.add_to_hash((i >> 32) as usize);
137        self.add_to_hash((i >> 64) as usize);
138        #[cfg(target_pointer_width = "32")]
139        self.add_to_hash((i >> 96) as usize);
140    }
141
142    #[inline]
143    fn write_usize(&mut self, i: usize) {
144        self.add_to_hash(i);
145    }
146
147    #[cfg(feature = "nightly")]
148    #[inline]
149    fn write_length_prefix(&mut self, _len: usize) {
150        // Most cases will specialize hash_slice to call write(), which encodes
151        // the length already in a more efficient manner than we could here. For
152        // HashDoS-resistance you would still need to include this for the
153        // non-slice collection hashes, but for the purposes of rustc we do not
154        // care and do not wish to pay the performance penalty of mixing in len
155        // for those collections.
156    }
157
158    #[cfg(feature = "nightly")]
159    #[inline]
160    fn write_str(&mut self, s: &str) {
161        // Similarly here, write already encodes the length, so nothing special
162        // is needed.
163        self.write(s.as_bytes())
164    }
165
166    #[inline]
167    fn finish(&self) -> u64 {
168        // Since we used a multiplicative hash our top bits have the most
169        // entropy (with the top bit having the most, decreasing as you go).
170        // As most hash table implementations (including hashbrown) compute
171        // the bucket index from the bottom bits we want to move bits from the
172        // top to the bottom. Ideally we'd rotate left by exactly the hash table
173        // size, but as we don't know this we'll choose 26 bits, giving decent
174        // entropy up until 2^26 table sizes. On 32-bit hosts we'll dial it
175        // back down a bit to 15 bits.
176
177        #[cfg(target_pointer_width = "64")]
178        const ROTATE: u32 = 26;
179        #[cfg(target_pointer_width = "32")]
180        const ROTATE: u32 = 15;
181
182        self.hash.rotate_left(ROTATE) as u64
183
184        // A bit reversal would be even better, except hashbrown also expects
185        // good entropy in the top 7 bits and a bit reverse would fill those
186        // bits with low entropy. More importantly, bit reversals are very slow
187        // on x86-64. A byte reversal is relatively fast, but still has a 2
188        // cycle latency on x86-64 compared to the 1 cycle latency of a rotate.
189        // It also suffers from the hashbrown-top-7-bit-issue.
190    }
191}
192
193// Nothing special, digits of pi.
194const SEED1: u64 = 0x243f6a8885a308d3;
195const SEED2: u64 = 0x13198a2e03707344;
196const PREVENT_TRIVIAL_ZERO_COLLAPSE: u64 = 0xa4093822299f31d0;
197
198#[inline]
199fn multiply_mix(x: u64, y: u64) -> u64 {
200    #[cfg(target_pointer_width = "64")]
201    {
202        // We compute the full u64 x u64 -> u128 product, this is a single mul
203        // instruction on x86-64, one mul plus one mulhi on ARM64.
204        let full = (x as u128) * (y as u128);
205        let lo = full as u64;
206        let hi = (full >> 64) as u64;
207
208        // The middle bits of the full product fluctuate the most with small
209        // changes in the input. This is the top bits of lo and the bottom bits
210        // of hi. We can thus make the entire output fluctuate with small
211        // changes to the input by XOR'ing these two halves.
212        lo ^ hi
213
214        // Unfortunately both 2^64 + 1 and 2^64 - 1 have small prime factors,
215        // otherwise combining with + or - could result in a really strong hash, as:
216        //     x * y = 2^64 * hi + lo = (-1) * hi + lo = lo - hi,   (mod 2^64 + 1)
217        //     x * y = 2^64 * hi + lo =    1 * hi + lo = lo + hi,   (mod 2^64 - 1)
218        // Multiplicative hashing is universal in a field (like mod p).
219    }
220
221    #[cfg(target_pointer_width = "32")]
222    {
223        // u64 x u64 -> u128 product is prohibitively expensive on 32-bit.
224        // Decompose into 32-bit parts.
225        let lx = x as u32;
226        let ly = y as u32;
227        let hx = (x >> 32) as u32;
228        let hy = (y >> 32) as u32;
229
230        // u32 x u32 -> u64 the low bits of one with the high bits of the other.
231        let afull = (lx as u64) * (hy as u64);
232        let bfull = (hx as u64) * (ly as u64);
233
234        // Combine, swapping low/high of one of them so the upper bits of the
235        // product of one combine with the lower bits of the other.
236        afull ^ bfull.rotate_right(32)
237    }
238}
239
240/// A wyhash-inspired non-collision-resistant hash for strings/slices designed
241/// by Orson Peters, with a focus on small strings and small codesize.
242///
243/// The 64-bit version of this hash passes the SMHasher3 test suite on the full
244/// 64-bit output, that is, f(hash_bytes(b) ^ f(seed)) for some good avalanching
245/// permutation f() passed all tests with zero failures. When using the 32-bit
246/// version of multiply_mix this hash has a few non-catastrophic failures where
247/// there are a handful more collisions than an optimal hash would give.
248///
249/// We don't bother avalanching here as we'll feed this hash into a
250/// multiplication after which we take the high bits, which avalanches for us.
251#[inline]
252fn hash_bytes(bytes: &[u8]) -> u64 {
253    let len = bytes.len();
254    let mut s0 = SEED1;
255    let mut s1 = SEED2;
256
257    if len <= 16 {
258        // XOR the input into s0, s1.
259        if len >= 8 {
260            s0 ^= u64::from_le_bytes(bytes[0..8].try_into().unwrap());
261            s1 ^= u64::from_le_bytes(bytes[len - 8..].try_into().unwrap());
262        } else if len >= 4 {
263            s0 ^= u32::from_le_bytes(bytes[0..4].try_into().unwrap()) as u64;
264            s1 ^= u32::from_le_bytes(bytes[len - 4..].try_into().unwrap()) as u64;
265        } else if len > 0 {
266            let lo = bytes[0];
267            let mid = bytes[len / 2];
268            let hi = bytes[len - 1];
269            s0 ^= lo as u64;
270            s1 ^= ((hi as u64) << 8) | mid as u64;
271        }
272    } else {
273        // Handle bulk (can partially overlap with suffix).
274        let mut off = 0;
275        while off < len - 16 {
276            let x = u64::from_le_bytes(bytes[off..off + 8].try_into().unwrap());
277            let y = u64::from_le_bytes(bytes[off + 8..off + 16].try_into().unwrap());
278
279            // Replace s1 with a mix of s0, x, and y, and s0 with s1.
280            // This ensures the compiler can unroll this loop into two
281            // independent streams, one operating on s0, the other on s1.
282            //
283            // Since zeroes are a common input we prevent an immediate trivial
284            // collapse of the hash function by XOR'ing a constant with y.
285            let t = multiply_mix(s0 ^ x, PREVENT_TRIVIAL_ZERO_COLLAPSE ^ y);
286            s0 = s1;
287            s1 = t;
288            off += 16;
289        }
290
291        let suffix = &bytes[len - 16..];
292        s0 ^= u64::from_le_bytes(suffix[0..8].try_into().unwrap());
293        s1 ^= u64::from_le_bytes(suffix[8..16].try_into().unwrap());
294    }
295
296    multiply_mix(s0, s1) ^ (len as u64)
297}
298
299/// An implementation of [`BuildHasher`] that produces [`FxHasher`]s.
300///
301/// ```
302/// use std::hash::BuildHasher;
303/// use rustc_hash::FxBuildHasher;
304/// assert_ne!(FxBuildHasher.hash_one(1), FxBuildHasher.hash_one(2));
305/// ```
306#[derive(Copy, Clone, Default)]
307pub struct FxBuildHasher;
308
309impl BuildHasher for FxBuildHasher {
310    type Hasher = FxHasher;
311    fn build_hasher(&self) -> FxHasher {
312        FxHasher::default()
313    }
314}
315
316#[cfg(test)]
317mod tests {
318    #[cfg(not(any(target_pointer_width = "64", target_pointer_width = "32")))]
319    compile_error!("The test suite only supports 64 bit and 32 bit usize");
320
321    use crate::{FxBuildHasher, FxHasher};
322    use core::hash::{BuildHasher, Hash, Hasher};
323
324    macro_rules! test_hash {
325        (
326            $(
327                hash($value:expr) == $result:expr,
328            )*
329        ) => {
330            $(
331                assert_eq!(FxBuildHasher.hash_one($value), $result);
332            )*
333        };
334    }
335
336    const B32: bool = cfg!(target_pointer_width = "32");
337
338    #[test]
339    fn unsigned() {
340        test_hash! {
341            hash(0_u8) == 0,
342            hash(1_u8) == if B32 { 3001993707 } else { 12157901119326311915 },
343            hash(100_u8) == if B32 { 3844759569 } else { 16751747135202103309 },
344            hash(u8::MAX) == if B32 { 999399879 } else { 1211781028898739645 },
345
346            hash(0_u16) == 0,
347            hash(1_u16) == if B32 { 3001993707 } else { 12157901119326311915 },
348            hash(100_u16) == if B32 { 3844759569 } else { 16751747135202103309 },
349            hash(u16::MAX) == if B32 { 3440503042 } else { 16279819243059860173 },
350
351            hash(0_u32) == 0,
352            hash(1_u32) == if B32 { 3001993707 } else { 12157901119326311915 },
353            hash(100_u32) == if B32 { 3844759569 } else { 16751747135202103309 },
354            hash(u32::MAX) == if B32 { 1293006356 } else { 7729994835221066939 },
355
356            hash(0_u64) == 0,
357            hash(1_u64) == if B32 { 275023839 } else { 12157901119326311915 },
358            hash(100_u64) == if B32 { 1732383522 } else { 16751747135202103309 },
359            hash(u64::MAX) == if B32 { 1017982517 } else { 6288842954450348564 },
360
361            hash(0_u128) == 0,
362            hash(1_u128) == if B32 { 1860738631 } else { 13032756267696824044 },
363            hash(100_u128) == if B32 { 1389515751 } else { 12003541609544029302 },
364            hash(u128::MAX) == if B32 { 2156022013 } else { 11702830760530184999 },
365
366            hash(0_usize) == 0,
367            hash(1_usize) == if B32 { 3001993707 } else { 12157901119326311915 },
368            hash(100_usize) == if B32 { 3844759569 } else { 16751747135202103309 },
369            hash(usize::MAX) == if B32 { 1293006356 } else { 6288842954450348564 },
370        }
371    }
372
373    #[test]
374    fn signed() {
375        test_hash! {
376            hash(i8::MIN) == if B32 { 2000713177 } else { 6684841074112525780 },
377            hash(0_i8) == 0,
378            hash(1_i8) == if B32 { 3001993707 } else { 12157901119326311915 },
379            hash(100_i8) == if B32 { 3844759569 } else { 16751747135202103309 },
380            hash(i8::MAX) == if B32 { 3293686765 } else { 12973684028562874344 },
381
382            hash(i16::MIN) == if B32 { 1073764727 } else { 14218860181193086044 },
383            hash(0_i16) == 0,
384            hash(1_i16) == if B32 { 3001993707 } else { 12157901119326311915 },
385            hash(100_i16) == if B32 { 3844759569 } else { 16751747135202103309 },
386            hash(i16::MAX) == if B32 { 2366738315 } else { 2060959061933882993 },
387
388            hash(i32::MIN) == if B32 { 16384 } else { 9943947977240134995 },
389            hash(0_i32) == 0,
390            hash(1_i32) == if B32 { 3001993707 } else { 12157901119326311915 },
391            hash(100_i32) == if B32 { 3844759569 } else { 16751747135202103309 },
392            hash(i32::MAX) == if B32 { 1293022740 } else { 16232790931690483559 },
393
394            hash(i64::MIN) == if B32 { 16384 } else { 33554432 },
395            hash(0_i64) == 0,
396            hash(1_i64) == if B32 { 275023839 } else { 12157901119326311915 },
397            hash(100_i64) == if B32 { 1732383522 } else { 16751747135202103309 },
398            hash(i64::MAX) == if B32 { 1017998901 } else { 6288842954483902996 },
399
400            hash(i128::MIN) == if B32 { 16384 } else { 33554432 },
401            hash(0_i128) == 0,
402            hash(1_i128) == if B32 { 1860738631 } else { 13032756267696824044 },
403            hash(100_i128) == if B32 { 1389515751 } else { 12003541609544029302 },
404            hash(i128::MAX) == if B32 { 2156005629 } else { 11702830760496630567 },
405
406            hash(isize::MIN) == if B32 { 16384 } else { 33554432 },
407            hash(0_isize) == 0,
408            hash(1_isize) == if B32 { 3001993707 } else { 12157901119326311915 },
409            hash(100_isize) == if B32 { 3844759569 } else { 16751747135202103309 },
410            hash(isize::MAX) == if B32 { 1293022740 } else { 6288842954483902996 },
411        }
412    }
413
414    // Avoid relying on any `Hash` implementations in the standard library.
415    struct HashBytes(&'static [u8]);
416    impl Hash for HashBytes {
417        fn hash<H: core::hash::Hasher>(&self, state: &mut H) {
418            state.write(self.0);
419        }
420    }
421
422    #[test]
423    fn bytes() {
424        test_hash! {
425            hash(HashBytes(&[])) == if B32 { 2673204745 } else { 17606491139363777937 },
426            hash(HashBytes(&[0])) == if B32 { 2948228584 } else { 5448590020104574886 },
427            hash(HashBytes(&[0, 0, 0, 0, 0, 0])) == if B32 { 3223252423 } else { 16766921560080789783 },
428            hash(HashBytes(&[1])) == if B32 { 2943445104 } else { 5922447956811044110 },
429            hash(HashBytes(&[2])) == if B32 { 1055423297 } else { 5229781508510959783 },
430            hash(HashBytes(b"uwu")) == if B32 { 2699662140 } else { 7168164714682931527 },
431            hash(HashBytes(b"These are some bytes for testing rustc_hash.")) == if B32 { 2303640537 } else { 2349210501944688211 },
432        }
433    }
434
435    #[test]
436    fn with_seed_actually_different() {
437        let seeds = [
438            [1, 2],
439            [42, 17],
440            [124436707, 99237],
441            [usize::MIN, usize::MAX],
442        ];
443
444        for [a_seed, b_seed] in seeds {
445            let a = || FxHasher::with_seed(a_seed);
446            let b = || FxHasher::with_seed(b_seed);
447
448            for x in u8::MIN..=u8::MAX {
449                let mut a = a();
450                let mut b = b();
451
452                x.hash(&mut a);
453                x.hash(&mut b);
454
455                assert_ne!(a.finish(), b.finish())
456            }
457        }
458    }
459}