/// A trait for describing vector operations used by vectorized searchers.
///
/// The trait is highly constrained to low level vector operations needed.
/// In general, it was invented mostly to be generic over x86's __m128i and
/// __m256i types. At time of writing, it also supports wasm and aarch64
/// 128-bit vector types as well.
///
/// # Safety
///
/// All methods are not safe since they are intended to be implemented using
/// vendor intrinsics, which are also not safe. Callers must ensure that the
/// appropriate target features are enabled in the calling function, and that
/// the current CPU supports them. All implementations should avoid marking the
/// routines with #[target_feature] and instead mark them as #[inline(always)]
/// to ensure they get appropriately inlined. (inline(always) cannot be used
/// with target_feature.)
pub(crate) trait Vector: Copy + core::fmt::Debug {
/// The number of bytes in the vector. That is, this is the size of the
/// vector in memory.
const BYTES: usize;
/// The bits that must be zero in order for a `*const u8` pointer to be
/// correctly aligned to read vector values.
const ALIGN: usize;
/// The type of the value returned by `Vector::movemask`.
///
/// This supports abstracting over the specific representation used in
/// order to accommodate different representations in different ISAs.
type Mask: MoveMask;
/// Create a vector with 8-bit lanes with the given byte repeated into each
/// lane.
unsafe fn splat(byte: u8) -> Self;
/// Read a vector-size number of bytes from the given pointer. The pointer
/// must be aligned to the size of the vector.
///
/// # Safety
///
/// Callers must guarantee that at least `BYTES` bytes are readable from
/// `data` and that `data` is aligned to a `BYTES` boundary.
unsafe fn load_aligned(data: *const u8) -> Self;
/// Read a vector-size number of bytes from the given pointer. The pointer
/// does not need to be aligned.
///
/// # Safety
///
/// Callers must guarantee that at least `BYTES` bytes are readable from
/// `data`.
unsafe fn load_unaligned(data: *const u8) -> Self;
/// _mm_movemask_epi8 or _mm256_movemask_epi8
unsafe fn movemask(self) -> Self::Mask;
/// _mm_cmpeq_epi8 or _mm256_cmpeq_epi8
unsafe fn cmpeq(self, vector2: Self) -> Self;
/// _mm_and_si128 or _mm256_and_si256
unsafe fn and(self, vector2: Self) -> Self;
/// _mm_or or _mm256_or_si256
unsafe fn or(self, vector2: Self) -> Self;
/// Returns true if and only if `Self::movemask` would return a mask that
/// contains at least one non-zero bit.
unsafe fn movemask_will_have_non_zero(self) -> bool {
self.movemask().has_non_zero()
}
}
/// A trait that abstracts over a vector-to-scalar operation called
/// "move mask."
///
/// On x86-64, this is `_mm_movemask_epi8` for SSE2 and `_mm256_movemask_epi8`
/// for AVX2. It takes a vector of `u8` lanes and returns a scalar where the
/// `i`th bit is set if and only if the most significant bit in the `i`th lane
/// of the vector is set. The simd128 ISA for wasm32 also supports this
/// exact same operation natively.
///
/// ... But aarch64 doesn't. So we have to fake it with more instructions and
/// a slightly different representation. We could do extra work to unify the
/// representations, but then would require additional costs in the hot path
/// for `memchr` and `packedpair`. So instead, we abstraction over the specific
/// representation with this trait an ddefine the operations we actually need.
pub(crate) trait MoveMask: Copy + core::fmt::Debug {
/// Return a mask that is all zeros except for the least significant `n`
/// lanes in a corresponding vector.
fn all_zeros_except_least_significant(n: usize) -> Self;
/// Returns true if and only if this mask has a a non-zero bit anywhere.
fn has_non_zero(self) -> bool;
/// Returns the number of bits set to 1 in this mask.
fn count_ones(self) -> usize;
/// Does a bitwise `and` operation between `self` and `other`.
fn and(self, other: Self) -> Self;
/// Does a bitwise `or` operation between `self` and `other`.
fn or(self, other: Self) -> Self;
/// Returns a mask that is equivalent to `self` but with the least
/// significant 1-bit set to 0.
fn clear_least_significant_bit(self) -> Self;
/// Returns the offset of the first non-zero lane this mask represents.
fn first_offset(self) -> usize;
/// Returns the offset of the last non-zero lane this mask represents.
fn last_offset(self) -> usize;
}
/// This is a "sensible" movemask implementation where each bit represents
/// whether the most significant bit is set in each corresponding lane of a
/// vector. This is used on x86-64 and wasm, but such a mask is more expensive
/// to get on aarch64 so we use something a little different.
///
/// We call this "sensible" because this is what we get using native sse/avx
/// movemask instructions. But neon has no such native equivalent.
#[derive(Clone, Copy, Debug)]
pub(crate) struct SensibleMoveMask(u32);
impl SensibleMoveMask {
/// Get the mask in a form suitable for computing offsets.
///
/// Basically, this normalizes to little endian. On big endian, this swaps
/// the bytes.
#[inline(always)]
fn get_for_offset(self) -> u32 {
#[cfg(target_endian = "big")]
{
self.0.swap_bytes()
}
#[cfg(target_endian = "little")]
{
self.0
}
}
}
impl MoveMask for SensibleMoveMask {
#[inline(always)]
fn all_zeros_except_least_significant(n: usize) -> SensibleMoveMask {
debug_assert!(n < 32);
SensibleMoveMask(!((1 << n) - 1))
}
#[inline(always)]
fn has_non_zero(self) -> bool {
self.0 != 0
}
#[inline(always)]
fn count_ones(self) -> usize {
self.0.count_ones() as usize
}
#[inline(always)]
fn and(self, other: SensibleMoveMask) -> SensibleMoveMask {
SensibleMoveMask(self.0 & other.0)
}
#[inline(always)]
fn or(self, other: SensibleMoveMask) -> SensibleMoveMask {
SensibleMoveMask(self.0 | other.0)
}
#[inline(always)]
fn clear_least_significant_bit(self) -> SensibleMoveMask {
SensibleMoveMask(self.0 & (self.0 - 1))
}
#[inline(always)]
fn first_offset(self) -> usize {
// We are dealing with little endian here (and if we aren't, we swap
// the bytes so we are in practice), where the most significant byte
// is at a higher address. That means the least significant bit that
// is set corresponds to the position of our first matching byte.
// That position corresponds to the number of zeros after the least
// significant bit.
self.get_for_offset().trailing_zeros() as usize
}
#[inline(always)]
fn last_offset(self) -> usize {
// We are dealing with little endian here (and if we aren't, we swap
// the bytes so we are in practice), where the most significant byte is
// at a higher address. That means the most significant bit that is set
// corresponds to the position of our last matching byte. The position
// from the end of the mask is therefore the number of leading zeros
// in a 32 bit integer, and the position from the start of the mask is
// therefore 32 - (leading zeros) - 1.
32 - self.get_for_offset().leading_zeros() as usize - 1
}
}
#[cfg(target_arch = "x86_64")]
mod x86sse2 {
use core::arch::x86_64::*;
use super::{SensibleMoveMask, Vector};
impl Vector for __m128i {
const BYTES: usize = 16;
const ALIGN: usize = Self::BYTES - 1;
type Mask = SensibleMoveMask;
#[inline(always)]
unsafe fn splat(byte: u8) -> __m128i {
_mm_set1_epi8(byte as i8)
}
#[inline(always)]
unsafe fn load_aligned(data: *const u8) -> __m128i {
_mm_load_si128(data as *const __m128i)
}
#[inline(always)]
unsafe fn load_unaligned(data: *const u8) -> __m128i {
_mm_loadu_si128(data as *const __m128i)
}
#[inline(always)]
unsafe fn movemask(self) -> SensibleMoveMask {
SensibleMoveMask(_mm_movemask_epi8(self) as u32)
}
#[inline(always)]
unsafe fn cmpeq(self, vector2: Self) -> __m128i {
_mm_cmpeq_epi8(self, vector2)
}
#[inline(always)]
unsafe fn and(self, vector2: Self) -> __m128i {
_mm_and_si128(self, vector2)
}
#[inline(always)]
unsafe fn or(self, vector2: Self) -> __m128i {
_mm_or_si128(self, vector2)
}
}
}
#[cfg(target_arch = "x86_64")]
mod x86avx2 {
use core::arch::x86_64::*;
use super::{SensibleMoveMask, Vector};
impl Vector for __m256i {
const BYTES: usize = 32;
const ALIGN: usize = Self::BYTES - 1;
type Mask = SensibleMoveMask;
#[inline(always)]
unsafe fn splat(byte: u8) -> __m256i {
_mm256_set1_epi8(byte as i8)
}
#[inline(always)]
unsafe fn load_aligned(data: *const u8) -> __m256i {
_mm256_load_si256(data as *const __m256i)
}
#[inline(always)]
unsafe fn load_unaligned(data: *const u8) -> __m256i {
_mm256_loadu_si256(data as *const __m256i)
}
#[inline(always)]
unsafe fn movemask(self) -> SensibleMoveMask {
SensibleMoveMask(_mm256_movemask_epi8(self) as u32)
}
#[inline(always)]
unsafe fn cmpeq(self, vector2: Self) -> __m256i {
_mm256_cmpeq_epi8(self, vector2)
}
#[inline(always)]
unsafe fn and(self, vector2: Self) -> __m256i {
_mm256_and_si256(self, vector2)
}
#[inline(always)]
unsafe fn or(self, vector2: Self) -> __m256i {
_mm256_or_si256(self, vector2)
}
}
}
#[cfg(target_arch = "aarch64")]
mod aarch64neon {
use core::arch::aarch64::*;
use super::{MoveMask, Vector};
impl Vector for uint8x16_t {
const BYTES: usize = 16;
const ALIGN: usize = Self::BYTES - 1;
type Mask = NeonMoveMask;
#[inline(always)]
unsafe fn splat(byte: u8) -> uint8x16_t {
vdupq_n_u8(byte)
}
#[inline(always)]
unsafe fn load_aligned(data: *const u8) -> uint8x16_t {
// I've tried `data.cast::<uint8x16_t>().read()` instead, but
// couldn't observe any benchmark differences.
Self::load_unaligned(data)
}
#[inline(always)]
unsafe fn load_unaligned(data: *const u8) -> uint8x16_t {
vld1q_u8(data)
}
#[inline(always)]
unsafe fn movemask(self) -> NeonMoveMask {
let asu16s = vreinterpretq_u16_u8(self);
let mask = vshrn_n_u16(asu16s, 4);
let asu64 = vreinterpret_u64_u8(mask);
let scalar64 = vget_lane_u64(asu64, 0);
NeonMoveMask(scalar64 & 0x8888888888888888)
}
#[inline(always)]
unsafe fn cmpeq(self, vector2: Self) -> uint8x16_t {
vceqq_u8(self, vector2)
}
#[inline(always)]
unsafe fn and(self, vector2: Self) -> uint8x16_t {
vandq_u8(self, vector2)
}
#[inline(always)]
unsafe fn or(self, vector2: Self) -> uint8x16_t {
vorrq_u8(self, vector2)
}
/// This is the only interesting implementation of this routine.
/// Basically, instead of doing the "shift right narrow" dance, we use
/// adajacent folding max to determine whether there are any non-zero
/// bytes in our mask. If there are, *then* we'll do the "shift right
/// narrow" dance. In benchmarks, this does lead to slightly better
/// throughput, but the win doesn't appear huge.
#[inline(always)]
unsafe fn movemask_will_have_non_zero(self) -> bool {
let low = vreinterpretq_u64_u8(vpmaxq_u8(self, self));
vgetq_lane_u64(low, 0) != 0
}
}
/// Neon doesn't have a `movemask` that works like the one in x86-64, so we
/// wind up using a different method[1]. The different method also produces
/// a mask, but 4 bits are set in the neon case instead of a single bit set
/// in the x86-64 case. We do an extra step to zero out 3 of the 4 bits,
/// but we still wind up with at least 3 zeroes between each set bit. This
/// generally means that we need to do some division by 4 before extracting
/// offsets.
///
/// In fact, the existence of this type is the entire reason that we have
/// the `MoveMask` trait in the first place. This basically lets us keep
/// the different representations of masks without being forced to unify
/// them into a single representation, which could result in extra and
/// unnecessary work.
///
/// [1]:
https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
#[derive(Clone, Copy, Debug)]
pub(crate) struct NeonMoveMask(u64);
impl NeonMoveMask {
/// Get the mask in a form suitable for computing offsets.
///
/// Basically, this normalizes to little endian. On big endian, this
/// swaps the bytes.
#[inline(always)]
fn get_for_offset(self) -> u64 {
#[cfg(target_endian = "big")]
{
self.0.swap_bytes()
}
#[cfg(target_endian = "little")]
{
self.0
}
}
}
impl MoveMask for NeonMoveMask {
#[inline(always)]
fn all_zeros_except_least_significant(n: usize) -> NeonMoveMask {
debug_assert!(n < 16);
NeonMoveMask(!(((1 << n) << 2) - 1))
}
#[inline(always)]
fn has_non_zero(self) -> bool {
self.0 != 0
}
#[inline(always)]
fn count_ones(self) -> usize {
self.0.count_ones() as usize
}
#[inline(always)]
fn and(self, other: NeonMoveMask) -> NeonMoveMask {
NeonMoveMask(self.0 & other.0)
}
#[inline(always)]
fn or(self, other: NeonMoveMask) -> NeonMoveMask {
NeonMoveMask(self.0 | other.0)
}
#[inline(always)]
fn clear_least_significant_bit(self) -> NeonMoveMask {
NeonMoveMask(self.0 & (self.0 - 1))
}
#[inline(always)]
fn first_offset(self) -> usize {
// We are dealing with little endian here (and if we aren't,
// we swap the bytes so we are in practice), where the most
// significant byte is at a higher address. That means the least
// significant bit that is set corresponds to the position of our
// first matching byte. That position corresponds to the number of
// zeros after the least significant bit.
//
// Note that unlike `SensibleMoveMask`, this mask has its bits
// spread out over 64 bits instead of 16 bits (for a 128 bit
// vector). Namely, where as x86-64 will turn
//
// 0x00 0xFF 0x00 0x00 0xFF
//
// into 10010, our neon approach will turn it into
//
// 10000000000010000000
//
// And this happens because neon doesn't have a native `movemask`
// instruction, so we kind of fake it[1]. Thus, we divide the
// number of trailing zeros by 4 to get the "real" offset.
//
// [1]:
https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
(self.get_for_offset().trailing_zeros() >> 2) as usize
}
#[inline(always)]
fn last_offset(self) -> usize {
// See comment in `first_offset` above. This is basically the same,
// but coming from the other direction.
16 - (self.get_for_offset().leading_zeros() >> 2) as usize - 1
}
}
}
#[cfg(all(target_arch = "wasm32", target_feature = "simd128"))]
mod wasm_simd128 {
use core::arch::wasm32::*;
use super::{SensibleMoveMask, Vector};
impl Vector for v128 {
const BYTES: usize = 16;
const ALIGN: usize = Self::BYTES - 1;
type Mask = SensibleMoveMask;
#[inline(always)]
unsafe fn splat(byte: u8) -> v128 {
u8x16_splat(byte)
}
#[inline(always)]
unsafe fn load_aligned(data: *const u8) -> v128 {
*data.cast()
}
#[inline(always)]
unsafe fn load_unaligned(data: *const u8) -> v128 {
v128_load(data.cast())
}
#[inline(always)]
unsafe fn movemask(self) -> SensibleMoveMask {
SensibleMoveMask(u8x16_bitmask(self).into())
}
#[inline(always)]
unsafe fn cmpeq(self, vector2: Self) -> v128 {
u8x16_eq(self, vector2)
}
#[inline(always)]
unsafe fn and(self, vector2: Self) -> v128 {
v128_and(self, vector2)
}
#[inline(always)]
unsafe fn or(self, vector2: Self) -> v128 {
v128_or(self, vector2)
}
}
}
