From 702f349f0b7860fe02efa1fea3354b6350161426 Mon Sep 17 00:00:00 2001 From: Orson Peters Date: Wed, 15 May 2024 00:17:25 +0200 Subject: [PATCH 1/6] Replace hash with faster and better finalized hash --- Cargo.toml | 1 + README.md | 21 ++- src/lib.rs | 366 +++++++++++++++++++++++++++++++++++++---------------- 3 files changed, 274 insertions(+), 114 deletions(-) diff --git a/Cargo.toml b/Cargo.toml index 191ef72..4ed7249 100644 --- a/Cargo.toml +++ b/Cargo.toml @@ -12,6 +12,7 @@ edition = "2018" [features] default = ["std"] std = [] +nightly = [] rand = ["dep:rand", "std"] [dependencies] diff --git a/README.md b/README.md index 05991b6..2a1b0cc 100644 --- a/README.md +++ b/README.md @@ -3,13 +3,22 @@ [![crates.io](https://img.shields.io/crates/v/rustc-hash.svg)](https://crates.io/crates/rustc-hash) [![Documentation](https://docs.rs/rustc-hash/badge.svg)](https://docs.rs/rustc-hash) -A speedy, non-cryptographic hashing algorithm used by `rustc` and Firefox. +A speedy, non-cryptographic hashing algorithm used by `rustc`. The [hash map in `std`](https://doc.rust-lang.org/std/collections/struct.HashMap.html) uses SipHash by default, which provides resistance against DOS attacks. -These attacks aren't as much of a concern in the compiler so we prefer to use the quicker, non-cryptographic Fx algorithm. - -The Fx algorithm is a unique one used by Firefox. It is fast because it can hash eight bytes at a time. -Within `rustc`, it consistently outperforms every other tested algorithm (such as FNV). -The collision rate is similar or slightly worse than other low-quality hash functions. +These attacks aren't as much of a concern in the compiler so we prefer to use a +quicker, non-cryptographic hash algorithm. + +The original hash algorithm provided by this crate was one taken from Firefox, +hence the hasher it provides is called FxHasher. This name is kept for backwards +compatibility, but the underlying hash has since been replaced. The current +design for the hasher is a polynomial hash finished with a single bit rotation, +together with a wyhash-inspired compression function for strings/slices, both +designed by Orson Peters. + +Within `rustc`, it consistently outperforms every other tested algorithm or +variation, despite its simplicity. Spending more CPU cycles on a higher quality +hash does not reduce hash collisions enough to be worth the cost on real-world +benchmarks. ## Usage diff --git a/src/lib.rs b/src/lib.rs index b94393b..ce50aa1 100644 --- a/src/lib.rs +++ b/src/lib.rs @@ -1,4 +1,4 @@ -//! A speedy, non-cryptographic hashing algorithm used by `rustc` and Firefox. +//! A speedy, non-cryptographic hashing algorithm used by `rustc`. //! //! # Example //! @@ -15,6 +15,7 @@ //! ``` #![no_std] +#![cfg_attr(feature = "nightly", feature(hasher_prefixfree_extras))] #[cfg(feature = "std")] extern crate std; @@ -32,8 +33,6 @@ use core::default::Default; #[cfg(feature = "std")] use core::hash::BuildHasherDefault; use core::hash::Hasher; -use core::mem::size_of; -use core::ops::BitXor; #[cfg(feature = "std")] use std::collections::{HashMap, HashSet}; @@ -55,33 +54,26 @@ pub use seeded_state::{FxHashMapSeed, FxHashSetSeed, FxSeededState}; /// compiler we're not really worried about DOS attempts, so we use a fast /// non-cryptographic hash. /// -/// This is the same as the algorithm used by Firefox -- which is a homespun -/// one not based on any widely-known algorithm -- though modified to produce -/// 64-bit hash values instead of 32-bit hash values. It consistently -/// out-performs an FNV-based hash within rustc itself -- the collision rate is -/// similar or slightly worse than FNV, but the speed of the hash function -/// itself is much higher because it works on up to 8 bytes at a time. +/// The current implementation is a fast polynomial hash with a single +/// bit rotation as a finishing step designed by Orson Peters. #[derive(Clone)] pub struct FxHasher { hash: usize, } -#[cfg(target_pointer_width = "32")] -const K: usize = 0x9e3779b9; +// One might view a polynomial hash +// m[0] * k + m[1] * k^2 + m[2] * k^3 + ... +// as a multilinear hash with keystream k[..] +// m[0] * k[0] + m[1] * k[1] + m[2] * k[2] + ... +// where keystream k just happens to be generated using a multiplicative +// congrential pseudorandom number generator (MCG). For that reason we chose a +// constant that was found to be good for a MCG in: +// "Computationally Easy, Spectrally Good Multipliers for Congruential +// Pseudorandom Number Generators" by Guy Steele and Sebastiano Vigna. #[cfg(target_pointer_width = "64")] -const K: usize = 0x517cc1b727220a95; - -#[inline] -fn take_first_chunk<'a, const N: usize>(slice: &mut &'a [u8]) -> Option<&'a [u8; N]> { - // TODO: use [T]::split_first_chunk() when stable - if slice.len() < N { - return None; - } - - let (first, rest) = slice.split_at(N); - *slice = rest; - Some(first.try_into().unwrap()) -} +const K: usize = 0xf1357aea2e62a9c5; +#[cfg(target_pointer_width = "32")] +const K: usize = 0x93d765dd; impl FxHasher { /// Creates a `fx` hasher with a given seed. @@ -105,42 +97,200 @@ impl Default for FxHasher { impl FxHasher { #[inline] fn add_to_hash(&mut self, i: usize) { - self.hash = self.hash.rotate_left(5).bitxor(i).wrapping_mul(K); + self.hash = self.hash.wrapping_add(i).wrapping_mul(K); } } impl Hasher for FxHasher { #[inline] - fn write(&mut self, mut bytes: &[u8]) { - // Ensure all bytes will be consumed - const _: () = assert!(size_of::() <= size_of::()); - // Ensure no bytes are discarded by casting to usize - const _: () = assert!(size_of::() <= size_of::()); - // Copy the 1 word sized state to a local variable to ensure it - // is kept in a register. - // See: https://github.com/rust-lang/rustc-hash/pull/34 - let mut state = self.clone(); - while let Some(&usize_bytes) = take_first_chunk(&mut bytes) { - state.add_to_hash(usize::from_ne_bytes(usize_bytes)); - } - if let Some(&u32_bytes) = take_first_chunk(&mut bytes) { - state.add_to_hash(u32::from_ne_bytes(u32_bytes) as usize); - } - if let Some(&u16_bytes) = take_first_chunk(&mut bytes) { - state.add_to_hash(u16::from_ne_bytes(u16_bytes) as usize); - } - if let Some(&[u8_byte]) = take_first_chunk(&mut bytes) { - state.add_to_hash(u8_byte as usize); - } - *self = state; + fn write(&mut self, bytes: &[u8]) { + self.write_u64(hash_bytes(bytes)); + } + + #[inline] + fn write_u8(&mut self, i: u8) { + self.add_to_hash(i as usize); + } + + #[inline] + fn write_u16(&mut self, i: u16) { + self.add_to_hash(i as usize); + } + + #[inline] + fn write_u32(&mut self, i: u32) { + self.add_to_hash(i as usize); + } + + #[inline] + fn write_u64(&mut self, i: u64) { + self.add_to_hash(i as usize); + #[cfg(target_pointer_width = "32")] + self.add_to_hash((i >> 32) as usize); + } + + #[inline] + fn write_u128(&mut self, i: u128) { + self.add_to_hash(i as usize); + #[cfg(target_pointer_width = "32")] + self.add_to_hash((i >> 32) as usize); + self.add_to_hash((i >> 64) as usize); + #[cfg(target_pointer_width = "32")] + self.add_to_hash((i >> 96) as usize); + } + + #[inline] + fn write_usize(&mut self, i: usize) { + self.add_to_hash(i); + } + + #[cfg(feature = "nightly")] + #[inline] + fn write_length_prefix(&mut self, len: usize) { + // Most cases will specialize hash_slice anyway which calls write(), + // which encodes the length already. + } + + #[cfg(feature = "nightly")] + #[inline] + fn write_str(&mut self, s: &str) { + // We don't need anything special here. + self.write(s.as_bytes()) } #[inline] fn finish(&self) -> u64 { - self.hash as u64 + // Since we used a multiplicative hash our top bits have the most + // entropy (with the top bit having the most, decreasing as you go). + // As most hash table implementations (including hashbrown) compute + // the bucket index from the bottom bits we want to move bits from the + // top to the bottom. Ideally we'd rotate left by exactly the hash table + // size, but as we don't know this we'll choose 20 bits, giving decent + // entropy up until 2^20 table sizes. On 32-bit hosts we'll dial it + // back down a bit to 15 bits. + + #[cfg(target_pointer_width = "64")] + const ROTATE: u32 = 20; + #[cfg(target_pointer_width = "32")] + const ROTATE: u32 = 15; + + self.hash.rotate_left(ROTATE) as u64 + + // A bit reversal would be even better, except hashbrown also expects + // good entropy in the top 7 bits and a bit reverse would fill those + // bits with low entropy. More importantly, bit reversals are very slow + // on x86-64. A byte reversal is relatively fast, but still has a 2 + // cycle latency on x86-64 compared to the 1 cycle latency of a rotate. + // It also suffers from the hashbrown-top-7-bit-issue. + } +} + +// Nothing special, digits of pi. +const SEED1: u64 = 0x243f6a8885a308d3; +const SEED2: u64 = 0x13198a2e03707344; +const PREVENT_TRIVIAL_ZERO_COLLAPSE: u64 = 0xa4093822299f31d0; + +#[inline] +fn multiply_mix(x: u64, y: u64) -> u64 { + #[cfg(target_pointer_width = "64")] + { + // We compute the full u64 x u64 -> u128 product, this is a single mul + // instruction on x86-64, one mul plus one mulhi on ARM64. + let full = (x as u128) * (y as u128); + let lo = full as u64; + let hi = (full >> 64) as u64; + + // The middle bits of the full product fluctuate the most with small + // changes in the input. This is the top bits of lo and the bottom bits + // of hi. We can thus make the entire output fluctuate with small + // changes to the input by XOR'ing these two halves. + lo ^ hi + + // Unfortunately both 2^64 + 1 and 2^64 - 1 have small prime factors, + // otherwise combining with + or - could result in a really strong hash, as: + // x * y = 2^64 * hi + lo = (-1) * hi + lo = lo - hi, (mod 2^64 + 1) + // x * y = 2^64 * hi + lo = 1 * hi + lo = lo + hi, (mod 2^64 - 1) + // Multiplicative hashing is universal in a field (like mod p). + } + + #[cfg(target_pointer_width = "32")] + { + // u64 x u64 -> u128 product is prohibitively expensive on 32-bit. + // Decompose into 32-bit parts. + let lx = x as u32; + let ly = y as u32; + let hx = (x >> 32) as u32; + let hy = (y >> 32) as u32; + + // u32 x u32 -> u64 the low bits of one with the high bits of the other. + let afull = (lx as u64) * (hy as u64); + let bfull = (hx as u64) * (ly as u64); + + // Combine, swapping low/high of one of them so the upper bits of the + // product of one combine with the lower bits of the other. + afull ^ bfull.rotate_right(32) } } +/// A wyhash-inspired non-collision-resistant hash for strings/slices designed +/// by Orson Peters, with a focus on small strings and small codesize. +/// +/// The 64-bit version of this hash passes the SMHasher3 test suite on the full +/// 64-bit output, that is, f(hash_bytes(b) ^ f(seed)) for some good avalanching +/// permutation f() passed all tests with zero failures. When using the 32-bit +/// version of multiply_mix this hash has a few non-catastrophic failures where +/// there are a handful more collisions than an optimal hash would give. +/// +/// We don't bother avalanching here as we'll feed this hash into a +/// multiplication after which we take the high bits, which avalanches for us. +#[inline] +fn hash_bytes(bytes: &[u8]) -> u64 { + let len = bytes.len(); + let mut s0 = SEED1; + let mut s1 = SEED2; + + if len <= 16 { + // XOR the input into s0, s1. + if len >= 8 { + s0 ^= u64::from_le_bytes(bytes[0..8].try_into().unwrap()); + s1 ^= u64::from_le_bytes(bytes[len - 8..].try_into().unwrap()); + } else if len >= 4 { + s0 ^= u32::from_le_bytes(bytes[0..4].try_into().unwrap()) as u64; + s1 ^= u32::from_le_bytes(bytes[len - 4..].try_into().unwrap()) as u64; + } else if len > 0 { + let lo = bytes[0]; + let mid = bytes[len / 2]; + let hi = bytes[len - 1]; + s0 ^= lo as u64; + s1 ^= ((hi as u64) << 8) | mid as u64; + } + } else { + // Handle bulk (can partially overlap with suffix). + let mut off = 0; + while off < len - 16 { + let x = u64::from_le_bytes(bytes[off..off + 8].try_into().unwrap()); + let y = u64::from_le_bytes(bytes[off + 8..off + 16].try_into().unwrap()); + + // Replace s1 with a mix of s0, x, and y, and s0 with s1. + // This ensures the compiler can unroll this loop into two + // independent streams, one operating on s0, the other on s1. + // + // Since zeroes are a common input we prevent an immediate trivial + // collapse of the hash function by XOR'ing a constant with y. + let t = multiply_mix(s0 ^ x, PREVENT_TRIVIAL_ZERO_COLLAPSE ^ y); + s0 = s1; + s1 = t; + off += 16; + } + + let suffix = &bytes[len - 16..]; + s0 ^= u64::from_le_bytes(suffix[0..8].try_into().unwrap()); + s1 ^= u64::from_le_bytes(suffix[8..16].try_into().unwrap()); + } + + multiply_mix(s0, s1) ^ (len as u64) +} + #[cfg(test)] mod tests { #[cfg(not(any(target_pointer_width = "64", target_pointer_width = "32")))] @@ -167,75 +317,75 @@ mod tests { fn unsigned() { test_hash! { hash(0_u8) == 0, - hash(1_u8) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(100_u8) == if B32 { 3450571844 } else { 15329034371404145204 }, - hash(u8::MAX) == if B32 { 2571255623 } else { 3117886703346944619 }, + hash(1_u8) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_u8) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(u8::MAX) == if B32 { 999399879 } else { 17600987023830959190 }, hash(0_u16) == 0, - hash(1_u16) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(100_u16) == if B32 { 3450571844 } else { 15329034371404145204 }, - hash(u16::MAX) == if B32 { 3682698823 } else { 8086887590654047595 }, + hash(1_u16) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_u16) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(u16::MAX) == if B32 { 3440503042 } else { 4001367065645062987 }, hash(0_u32) == 0, - hash(1_u32) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(100_u32) == if B32 { 3450571844 } else { 15329034371404145204 }, - hash(u32::MAX) == if B32 { 1640531527 } else { 15394791018899305835 }, + hash(1_u32) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_u32) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(u32::MAX) == if B32 { 1293006356 } else { 17126373362251322066 }, hash(0_u64) == 0, - hash(1_u64) == if B32 { 703266523 } else { 5871781006564002453 }, - hash(100_u64) == if B32 { 2407204753 } else { 15329034371404145204 }, - hash(u64::MAX) == if B32 { 1660667835 } else { 12574963067145549163 }, + hash(1_u64) == if B32 { 275023839 } else { 12583873379513078615 }, + hash(100_u64) == if B32 { 1732383522 } else { 4008740938959785536 }, + hash(u64::MAX) == if B32 { 1017982517 } else { 5862870694197521576 }, hash(0_u128) == 0, - hash(1_u128) == if B32 { 1294492036 } else { 956286968014291186 }, - hash(100_u128) == if B32 { 3411300242 } else { 2770938889503972258 }, - hash(u128::MAX) == if B32 { 3723263291 } else { 15973479568771280466 }, + hash(1_u128) == if B32 { 1860738631 } else { 12885773367358079611 }, + hash(100_u128) == if B32 { 1389515751 } else { 15751995649841559633 }, + hash(u128::MAX) == if B32 { 2156022013 } else { 11423841400550042156 }, hash(0_usize) == 0, - hash(1_usize) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(100_usize) == if B32 { 3450571844 } else { 15329034371404145204 }, - hash(usize::MAX) == if B32 { 1640531527 } else { 12574963067145549163 }, + hash(1_usize) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_usize) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(usize::MAX) == if B32 { 1293006356 } else { 5862870694197521576 }, } } #[test] fn signed() { test_hash! { - hash(i8::MIN) == if B32 { 465362048 } else { 13718205891810249344 }, - hash(0_i8) == 0, - hash(1_i8) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(100_i8) == if B32 { 3450571844 } else { 15329034371404145204 }, - hash(i8::MAX) == if B32 { 2105893575 } else { 7846424885246246891 }, - - hash(i16::MIN) == if B32 { 3168567296 } else { 6979334298609025024 }, - hash(0_i16) == 0, - hash(1_i16) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(100_i16) == if B32 { 3450571844 } else { 15329034371404145204 }, - hash(i16::MAX) == if B32 { 514131527 } else { 1107553292045022571 }, - - hash(i32::MIN) == if B32 { 2147483648 } else { 10633286012731654144 }, - hash(0_i32) == 0, - hash(1_i32) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(100_i32) == if B32 { 3450571844 } else { 15329034371404145204 }, - hash(i32::MAX) == if B32 { 3788015175 } else { 4761505006167651691 }, - - hash(i64::MIN) == if B32 { 2147483648 } else { 9223372036854775808 }, - hash(0_i64) == 0, - hash(1_i64) == if B32 { 703266523 } else { 5871781006564002453 }, - hash(100_i64) == if B32 { 2407204753 } else { 15329034371404145204 }, - hash(i64::MAX) == if B32 { 3808151483 } else { 3351591030290773355 }, - - hash(i128::MIN) == if B32 { 2147483648 } else { 9223372036854775808 }, - hash(0_i128) == 0, - hash(1_i128) == if B32 { 1294492036 } else { 956286968014291186 }, - hash(100_i128) == if B32 { 3411300242 } else { 2770938889503972258 }, - hash(i128::MAX) == if B32 { 1575779643 } else { 6750107531916504658 }, - - hash(isize::MIN) == if B32 { 2147483648 } else { 9223372036854775808 }, - hash(0_isize) == 0, - hash(1_isize) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(100_isize) == if B32 { 3450571844 } else { 15329034371404145204 }, - hash(isize::MAX) == if B32 { 3788015175 } else { 3351591030290773355 }, + hash(i8::MIN) == if B32 { 2000713177 } else { 5869058164817243095 }, + hash(0_i8) == 0, + hash(1_i8) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_i8) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(i8::MAX) == if B32 { 3293686765 } else { 11731928859014764671 }, + + hash(i16::MIN) == if B32 { 1073764727 } else { 8292620222579070801 }, + hash(0_i16) == 0, + hash(1_i16) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_i16) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(i16::MAX) == if B32 { 2366738315 } else { 14155490916776592377 }, + + hash(i32::MIN) == if B32 { 16384 } else { 5631751334026900245 }, + hash(0_i32) == 0, + hash(1_i32) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_i32) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(i32::MAX) == if B32 { 1293022740 } else { 11494622028224421821 }, + + hash(i64::MIN) == if B32 { 16384 } else { 524288 }, + hash(0_i64) == 0, + hash(1_i64) == if B32 { 275023839 } else { 12583873379513078615 }, + hash(100_i64) == if B32 { 1732383522 } else { 4008740938959785536 }, + hash(i64::MAX) == if B32 { 1017998901 } else { 5862870694198045864 }, + + hash(i128::MIN) == if B32 { 16384 } else { 524288 }, + hash(0_i128) == 0, + hash(1_i128) == if B32 { 1860738631 } else { 12885773367358079611 }, + hash(100_i128) == if B32 { 1389515751 } else { 15751995649841559633 }, + hash(i128::MAX) == if B32 { 2156005629 } else { 11423841400549517868 }, + + hash(isize::MIN) == if B32 { 16384 } else { 524288 }, + hash(0_isize) == 0, + hash(1_isize) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_isize) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(isize::MAX) == if B32 { 1293022740 } else { 5862870694198045864 }, } } @@ -250,13 +400,13 @@ mod tests { #[test] fn bytes() { test_hash! { - hash(HashBytes(&[])) == 0, - hash(HashBytes(&[0])) == 0, - hash(HashBytes(&[0, 0, 0, 0, 0, 0])) == 0, - hash(HashBytes(&[1])) == if B32 { 2654435769 } else { 5871781006564002453 }, - hash(HashBytes(&[2])) == if B32 { 1013904242 } else { 11743562013128004906 }, - hash(HashBytes(b"uwu")) == if B32 { 3939043750 } else { 16622306935539548858 }, - hash(HashBytes(b"These are some bytes for testing rustc_hash.")) == if B32 { 2345708736 } else { 12390864548135261390 }, + hash(HashBytes(&[])) == if B32 { 2673204745 } else { 5175017818631658678 }, + hash(HashBytes(&[0])) == if B32 { 2948228584 } else { 11037888512829180254 }, + hash(HashBytes(&[0, 0, 0, 0, 0, 0])) == if B32 { 3223252423 } else { 6891281800865632452 }, + hash(HashBytes(&[1])) == if B32 { 2943445104 } else { 4127763515449136980 }, + hash(HashBytes(&[2])) == if B32 { 1055423297 } else { 11322700005987241762 }, + hash(HashBytes(b"uwu")) == if B32 { 2699662140 } else { 2129615206728903013 }, + hash(HashBytes(b"These are some bytes for testing rustc_hash.")) == if B32 { 2303640537 } else { 5513083560975408889 }, } } From d5b36344139e1271104307686bb2f59c2cc4f46b Mon Sep 17 00:00:00 2001 From: Orson Peters Date: Wed, 15 May 2024 00:50:40 +0200 Subject: [PATCH 2/6] Fix trailing whitespace --- src/lib.rs | 84 +++++++++++++++++++++++++++--------------------------- 1 file changed, 42 insertions(+), 42 deletions(-) diff --git a/src/lib.rs b/src/lib.rs index ce50aa1..c08d4f3 100644 --- a/src/lib.rs +++ b/src/lib.rs @@ -351,41 +351,41 @@ mod tests { #[test] fn signed() { test_hash! { - hash(i8::MIN) == if B32 { 2000713177 } else { 5869058164817243095 }, - hash(0_i8) == 0, - hash(1_i8) == if B32 { 3001993707 } else { 12583873379513078615 }, - hash(100_i8) == if B32 { 3844759569 } else { 4008740938959785536 }, - hash(i8::MAX) == if B32 { 3293686765 } else { 11731928859014764671 }, - - hash(i16::MIN) == if B32 { 1073764727 } else { 8292620222579070801 }, - hash(0_i16) == 0, - hash(1_i16) == if B32 { 3001993707 } else { 12583873379513078615 }, - hash(100_i16) == if B32 { 3844759569 } else { 4008740938959785536 }, - hash(i16::MAX) == if B32 { 2366738315 } else { 14155490916776592377 }, - - hash(i32::MIN) == if B32 { 16384 } else { 5631751334026900245 }, - hash(0_i32) == 0, - hash(1_i32) == if B32 { 3001993707 } else { 12583873379513078615 }, - hash(100_i32) == if B32 { 3844759569 } else { 4008740938959785536 }, - hash(i32::MAX) == if B32 { 1293022740 } else { 11494622028224421821 }, - - hash(i64::MIN) == if B32 { 16384 } else { 524288 }, - hash(0_i64) == 0, - hash(1_i64) == if B32 { 275023839 } else { 12583873379513078615 }, - hash(100_i64) == if B32 { 1732383522 } else { 4008740938959785536 }, - hash(i64::MAX) == if B32 { 1017998901 } else { 5862870694198045864 }, - - hash(i128::MIN) == if B32 { 16384 } else { 524288 }, - hash(0_i128) == 0, - hash(1_i128) == if B32 { 1860738631 } else { 12885773367358079611 }, - hash(100_i128) == if B32 { 1389515751 } else { 15751995649841559633 }, - hash(i128::MAX) == if B32 { 2156005629 } else { 11423841400549517868 }, - - hash(isize::MIN) == if B32 { 16384 } else { 524288 }, - hash(0_isize) == 0, - hash(1_isize) == if B32 { 3001993707 } else { 12583873379513078615 }, - hash(100_isize) == if B32 { 3844759569 } else { 4008740938959785536 }, - hash(isize::MAX) == if B32 { 1293022740 } else { 5862870694198045864 }, + hash(i8::MIN) == if B32 { 2000713177 } else { 5869058164817243095 }, + hash(0_i8) == 0, + hash(1_i8) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_i8) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(i8::MAX) == if B32 { 3293686765 } else { 11731928859014764671 }, + + hash(i16::MIN) == if B32 { 1073764727 } else { 8292620222579070801 }, + hash(0_i16) == 0, + hash(1_i16) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_i16) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(i16::MAX) == if B32 { 2366738315 } else { 14155490916776592377 }, + + hash(i32::MIN) == if B32 { 16384 } else { 5631751334026900245 }, + hash(0_i32) == 0, + hash(1_i32) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_i32) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(i32::MAX) == if B32 { 1293022740 } else { 11494622028224421821 }, + + hash(i64::MIN) == if B32 { 16384 } else { 524288 }, + hash(0_i64) == 0, + hash(1_i64) == if B32 { 275023839 } else { 12583873379513078615 }, + hash(100_i64) == if B32 { 1732383522 } else { 4008740938959785536 }, + hash(i64::MAX) == if B32 { 1017998901 } else { 5862870694198045864 }, + + hash(i128::MIN) == if B32 { 16384 } else { 524288 }, + hash(0_i128) == 0, + hash(1_i128) == if B32 { 1860738631 } else { 12885773367358079611 }, + hash(100_i128) == if B32 { 1389515751 } else { 15751995649841559633 }, + hash(i128::MAX) == if B32 { 2156005629 } else { 11423841400549517868 }, + + hash(isize::MIN) == if B32 { 16384 } else { 524288 }, + hash(0_isize) == 0, + hash(1_isize) == if B32 { 3001993707 } else { 12583873379513078615 }, + hash(100_isize) == if B32 { 3844759569 } else { 4008740938959785536 }, + hash(isize::MAX) == if B32 { 1293022740 } else { 5862870694198045864 }, } } @@ -400,13 +400,13 @@ mod tests { #[test] fn bytes() { test_hash! { - hash(HashBytes(&[])) == if B32 { 2673204745 } else { 5175017818631658678 }, - hash(HashBytes(&[0])) == if B32 { 2948228584 } else { 11037888512829180254 }, - hash(HashBytes(&[0, 0, 0, 0, 0, 0])) == if B32 { 3223252423 } else { 6891281800865632452 }, - hash(HashBytes(&[1])) == if B32 { 2943445104 } else { 4127763515449136980 }, - hash(HashBytes(&[2])) == if B32 { 1055423297 } else { 11322700005987241762 }, - hash(HashBytes(b"uwu")) == if B32 { 2699662140 } else { 2129615206728903013 }, - hash(HashBytes(b"These are some bytes for testing rustc_hash.")) == if B32 { 2303640537 } else { 5513083560975408889 }, + hash(HashBytes(&[])) == if B32 { 2673204745 } else { 5175017818631658678 }, + hash(HashBytes(&[0])) == if B32 { 2948228584 } else { 11037888512829180254 }, + hash(HashBytes(&[0, 0, 0, 0, 0, 0])) == if B32 { 3223252423 } else { 6891281800865632452 }, + hash(HashBytes(&[1])) == if B32 { 2943445104 } else { 4127763515449136980 }, + hash(HashBytes(&[2])) == if B32 { 1055423297 } else { 11322700005987241762 }, + hash(HashBytes(b"uwu")) == if B32 { 2699662140 } else { 2129615206728903013 }, + hash(HashBytes(b"These are some bytes for testing rustc_hash.")) == if B32 { 2303640537 } else { 5513083560975408889 }, } } From 4f6be25b23c194281fcfc8c42ae4473ec9231fb0 Mon Sep 17 00:00:00 2001 From: Orson Peters Date: Sat, 18 May 2024 14:22:32 +0200 Subject: [PATCH 3/6] address review comments --- README.md | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/README.md b/README.md index 2a1b0cc..36b13ca 100644 --- a/README.md +++ b/README.md @@ -5,8 +5,8 @@ A speedy, non-cryptographic hashing algorithm used by `rustc`. The [hash map in `std`](https://doc.rust-lang.org/std/collections/struct.HashMap.html) uses SipHash by default, which provides resistance against DOS attacks. -These attacks aren't as much of a concern in the compiler so we prefer to use a -quicker, non-cryptographic hash algorithm. +These attacks aren't a concern in the compiler so we prefer to use a quicker, +non-cryptographic hash algorithm. The original hash algorithm provided by this crate was one taken from Firefox, hence the hasher it provides is called FxHasher. This name is kept for backwards @@ -15,10 +15,10 @@ design for the hasher is a polynomial hash finished with a single bit rotation, together with a wyhash-inspired compression function for strings/slices, both designed by Orson Peters. -Within `rustc`, it consistently outperforms every other tested algorithm or -variation, despite its simplicity. Spending more CPU cycles on a higher quality -hash does not reduce hash collisions enough to be worth the cost on real-world -benchmarks. +For `rustc` we have tried many different hashing algorithms. Hashing speed is +critical, especially for single integers. Spending more CPU cycles on a higher +quality hash does not reduce hash collisions enough to make the compiler faster +on real-world benchmarks. ## Usage From e07b577e6cb9d0cc42262b87cdcac32ee252c90f Mon Sep 17 00:00:00 2001 From: Orson Peters Date: Sat, 18 May 2024 14:22:47 +0200 Subject: [PATCH 4/6] fix unused function argument warning --- src/lib.rs | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/lib.rs b/src/lib.rs index c08d4f3..b967451 100644 --- a/src/lib.rs +++ b/src/lib.rs @@ -146,7 +146,7 @@ impl Hasher for FxHasher { #[cfg(feature = "nightly")] #[inline] - fn write_length_prefix(&mut self, len: usize) { + fn write_length_prefix(&mut self, _len: usize) { // Most cases will specialize hash_slice anyway which calls write(), // which encodes the length already. } From d3453abd348365257245a4070f8937b07c44d340 Mon Sep 17 00:00:00 2001 From: Orson Peters Date: Sat, 18 May 2024 23:47:07 +0200 Subject: [PATCH 5/6] Address more review comments --- src/lib.rs | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/src/lib.rs b/src/lib.rs index b967451..62fa223 100644 --- a/src/lib.rs +++ b/src/lib.rs @@ -104,6 +104,7 @@ impl FxHasher { impl Hasher for FxHasher { #[inline] fn write(&mut self, bytes: &[u8]) { + // Compress the byte string to a single u64 and add to our hash. self.write_u64(hash_bytes(bytes)); } @@ -147,14 +148,19 @@ impl Hasher for FxHasher { #[cfg(feature = "nightly")] #[inline] fn write_length_prefix(&mut self, _len: usize) { - // Most cases will specialize hash_slice anyway which calls write(), - // which encodes the length already. + // Most cases will specialize hash_slice to call write(), which encodes + // the length already in a more efficient manner than we could here. For + // HashDoS-resistance you would still need to include this for the + // non-slice collection hashes, but for the purposes of rustc we do not + // care and do not wish to pay the performance penalty of mixing in len + // for those collections. } #[cfg(feature = "nightly")] #[inline] fn write_str(&mut self, s: &str) { - // We don't need anything special here. + // Similarly here, write already encodes the length, so nothing special + // is needed. self.write(s.as_bytes()) } From 4cb9c7cfd5349b7f5a9c40d8f9ad40404fdd130d Mon Sep 17 00:00:00 2001 From: Orson Peters Date: Tue, 28 May 2024 11:49:23 +0200 Subject: [PATCH 6/6] Remove mention of firefox from Cargo.toml. --- Cargo.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Cargo.toml b/Cargo.toml index 4ed7249..e6686c4 100644 --- a/Cargo.toml +++ b/Cargo.toml @@ -2,7 +2,7 @@ name = "rustc-hash" version = "1.1.0" authors = ["The Rust Project Developers"] -description = "A speedy, non-cryptographic hashing algorithm used by rustc and Firefox" +description = "A speedy, non-cryptographic hashing algorithm used by rustc" license = "Apache-2.0/MIT" readme = "README.md" keywords = ["hash", "hasher", "fxhash", "rustc"]