std/crc64: optimize for x86+SSE4.2

name                           old speed      new speed      delta

wuffs_crc64_ecma_10k/clang14   1.78GB/s ± 0%  5.79GB/s ± 0%  +225.48%  (p=0.008 n=5+5)
wuffs_crc64_ecma_100k/clang14  1.78GB/s ± 0%  6.09GB/s ± 0%  +241.67%  (p=0.008 n=5+5)

wuffs_crc64_ecma_10k/gcc12     2.05GB/s ± 1%  5.79GB/s ± 1%  +182.80%  (p=0.008 n=5+5)
wuffs_crc64_ecma_100k/gcc12    2.01GB/s ± 0%  6.09GB/s ± 0%  +202.64%  (p=0.008 n=5+5)

wuffs_xz_decode_100k/clang14   57.9MB/s ± 1%  60.2MB/s ± 0%    +4.11%  (p=0.008 n=5+5)

wuffs_xz_decode_100k/gcc12     57.0MB/s ± 0%  57.8MB/s ± 1%    +1.39%  (p=0.008 n=5+5)

$ time example-mzcat < linux-6.8.2.tar.xz > /dev/null
Before: user 0m8.122s
After:  user 0m8.115s
This commit is contained in:
Nigel Tao committed 2024-04-02 00:17:43 +11:00
1 parent 5c693c9de8
commit f32bfe95c6
5 files changed
+466 -24

No files matched your search

+1 -1
View File
@@ -648,7 +648,7 @@ func (g *gen) writeBuiltinCPUArchX86(b *buffer, recv *a.Expr, method t.ID, retur
case "make_m128i_single_u32":
fName, tName = "_mm_cvtsi32_si128", "int32_t"
case "make_m128i_single_u64":
fName, tName = "_mm_cvtsi64x_si128", "int64_t"
fName, tName = "_mm_cvtsi64_si128", "int64_t"
case "make_m128i_slice128", "make_m128i_slice_u16lex8":
fName, tName, ptr = "_mm_lddqu_si128", "const __m128i*)(const void*", true
case "make_m128i_zeroes":
+195 -18
View File
@@ -8468,6 +8468,10 @@ struct wuffs_crc64__ecma_hasher__struct {
wuffs_base__vtable null_vtable;
uint64_t f_state;
wuffs_base__empty_struct (*choosy_up)(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x);
} private_impl;
#ifdef __cplusplus
@@ -36626,10 +36630,54 @@ WUFFS_CRC64__ECMA_TABLE[8][256] WUFFS_BASE__POTENTIALLY_UNUSED = {
},
};
static const uint8_t
WUFFS_CRC64__SHUFFLE_707F[16] WUFFS_BASE__POTENTIALLY_UNUSED = {
112u, 113u, 114u, 115u, 116u, 117u, 118u, 119u,
120u, 121u, 122u, 123u, 124u, 125u, 126u, 127u,
};
static const uint8_t
WUFFS_CRC64__SHUFFLE_8F80[16] WUFFS_BASE__POTENTIALLY_UNUSED = {
143u, 142u, 141u, 140u, 139u, 138u, 137u, 136u,
135u, 134u, 133u, 132u, 131u, 130u, 129u, 128u,
};
static const uint8_t
WUFFS_CRC64__ECMA_X86_SSE42_K1K2[16] WUFFS_BASE__POTENTIALLY_UNUSED = {
228u, 58u, 57u, 202u, 151u, 212u, 93u, 224u,
64u, 95u, 135u, 199u, 175u, 149u, 190u, 218u,
};
static const uint8_t
WUFFS_CRC64__ECMA_X86_SSE42_PXMU[16] WUFFS_BASE__POTENTIALLY_UNUSED = {
133u, 30u, 14u, 175u, 43u, 175u, 216u, 146u,
213u, 99u, 41u, 23u, 108u, 70u, 62u, 156u,
};
// ---------------- Private Initializer Prototypes
// ---------------- Private Function Prototypes
WUFFS_BASE__GENERATED_C_CODE
static wuffs_base__empty_struct
wuffs_crc64__ecma_hasher__up(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x);
WUFFS_BASE__GENERATED_C_CODE
static wuffs_base__empty_struct
wuffs_crc64__ecma_hasher__up__choosy_default(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x);
#if defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
WUFFS_BASE__GENERATED_C_CODE
static wuffs_base__empty_struct
wuffs_crc64__ecma_hasher__up_x86_sse42(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x);
#endif // defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
// ---------------- VTables
const wuffs_base__hasher_u64__func_ptrs
@@ -36687,6 +36735,8 @@ wuffs_crc64__ecma_hasher__initialize(
}
}
self->private_impl.choosy_up = &wuffs_crc64__ecma_hasher__up__choosy_default;
self->private_impl.magic = WUFFS_BASE__MAGIC;
self->private_impl.vtable_for__wuffs_base__hasher_u64.vtable_name =
wuffs_base__hasher_u64__vtable_name;
@@ -36770,6 +36820,50 @@ wuffs_crc64__ecma_hasher__update(
return wuffs_base__make_empty_struct();
}
if (self->private_impl.f_state == 0u) {
self->private_impl.choosy_up = (
#if defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
wuffs_base__cpu_arch__have_x86_sse42() ? &wuffs_crc64__ecma_hasher__up_x86_sse42 :
#endif
self->private_impl.choosy_up);
}
wuffs_crc64__ecma_hasher__up(self, a_x);
return wuffs_base__make_empty_struct();
}
// -------- func crc64.ecma_hasher.update_u64
WUFFS_BASE__GENERATED_C_CODE
WUFFS_BASE__MAYBE_STATIC uint64_t
wuffs_crc64__ecma_hasher__update_u64(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x) {
if (!self) {
return 0;
}
if (self->private_impl.magic != WUFFS_BASE__MAGIC) {
return 0;
}
wuffs_crc64__ecma_hasher__update(self, a_x);
return wuffs_crc64__ecma_hasher__checksum_u64(self);
}
// -------- func crc64.ecma_hasher.up
WUFFS_BASE__GENERATED_C_CODE
static wuffs_base__empty_struct
wuffs_crc64__ecma_hasher__up(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x) {
return (*self->private_impl.choosy_up)(self, a_x);
}
WUFFS_BASE__GENERATED_C_CODE
static wuffs_base__empty_struct
wuffs_crc64__ecma_hasher__up__choosy_default(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x) {
uint64_t v_s = 0;
wuffs_base__slice_u8 v_p = {0};
@@ -36810,24 +36904,6 @@ wuffs_crc64__ecma_hasher__update(
return wuffs_base__make_empty_struct();
}
// -------- func crc64.ecma_hasher.update_u64
WUFFS_BASE__GENERATED_C_CODE
WUFFS_BASE__MAYBE_STATIC uint64_t
wuffs_crc64__ecma_hasher__update_u64(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x) {
if (!self) {
return 0;
}
if (self->private_impl.magic != WUFFS_BASE__MAGIC) {
return 0;
}
wuffs_crc64__ecma_hasher__update(self, a_x);
return wuffs_crc64__ecma_hasher__checksum_u64(self);
}
// -------- func crc64.ecma_hasher.checksum_u64
WUFFS_BASE__GENERATED_C_CODE
@@ -36845,6 +36921,107 @@ wuffs_crc64__ecma_hasher__checksum_u64(
return self->private_impl.f_state;
}
// ‼ WUFFS MULTI-FILE SECTION +x86_sse42
// -------- func crc64.ecma_hasher.up_x86_sse42
#if defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
WUFFS_BASE__MAYBE_ATTRIBUTE_TARGET("pclmul,popcnt,sse4.2")
WUFFS_BASE__GENERATED_C_CODE
static wuffs_base__empty_struct
wuffs_crc64__ecma_hasher__up_x86_sse42(
wuffs_crc64__ecma_hasher* self,
wuffs_base__slice_u8 a_x) {
uint64_t v_s = 0;
wuffs_base__slice_u8 v_p = {0};
__m128i v_s0 = {0};
__m128i v_s0_707F = {0};
__m128i v_s0_8F80 = {0};
__m128i v_x0 = {0};
__m128i v_aa = {0};
__m128i v_k1k2 = {0};
__m128i v_t0 = {0};
__m128i v_t1 = {0};
__m128i v_t2 = {0};
__m128i v_u0 = {0};
__m128i v_u1 = {0};
__m128i v_u2 = {0};
__m128i v_v0 = {0};
__m128i v_v1 = {0};
__m128i v_pxmu = {0};
__m128i v_w1 = {0};
__m128i v_w2 = {0};
uint64_t v_tail_index = 0;
v_s = (18446744073709551615u ^ self->private_impl.f_state);
while ((((uint64_t)(a_x.len)) > 0u) && ((15u & ((uint32_t)(0xFFFu & (uintptr_t)(a_x.ptr)))) != 0u)) {
v_s = (WUFFS_CRC64__ECMA_TABLE[0u][((uint8_t)(((uint8_t)(v_s)) ^ a_x.ptr[0u]))] ^ (v_s >> 8u));
a_x = wuffs_base__slice_u8__subslice_i(a_x, 1u);
}
if (((uint64_t)(a_x.len)) < 32u) {
{
wuffs_base__slice_u8 i_slice_p = a_x;
v_p.ptr = i_slice_p.ptr;
v_p.len = 1;
uint8_t* i_end0_p = i_slice_p.ptr + i_slice_p.len;
while (v_p.ptr < i_end0_p) {
v_s = (WUFFS_CRC64__ECMA_TABLE[0u][((uint8_t)(((uint8_t)(v_s)) ^ v_p.ptr[0u]))] ^ (v_s >> 8u));
v_p.ptr += 1;
}
v_p.len = 0;
}
self->private_impl.f_state = (18446744073709551615u ^ v_s);
return wuffs_base__make_empty_struct();
}
v_s0 = _mm_cvtsi64_si128((int64_t)(v_s));
v_s0_707F = _mm_shuffle_epi8(v_s0, _mm_lddqu_si128((const __m128i*)(const void*)(WUFFS_CRC64__SHUFFLE_707F)));
v_s0_8F80 = _mm_shuffle_epi8(v_s0, _mm_lddqu_si128((const __m128i*)(const void*)(WUFFS_CRC64__SHUFFLE_8F80)));
v_x0 = _mm_lddqu_si128((const __m128i*)(const void*)(a_x.ptr + 0u));
a_x = wuffs_base__slice_u8__subslice_i(a_x, 16u);
v_k1k2 = _mm_lddqu_si128((const __m128i*)(const void*)(WUFFS_CRC64__ECMA_X86_SSE42_K1K2));
v_t0 = _mm_xor_si128(v_s0_707F, v_x0);
v_t1 = _mm_clmulepi64_si128(v_t0, v_k1k2, (int32_t)(0u));
v_t2 = _mm_clmulepi64_si128(v_t0, v_k1k2, (int32_t)(17u));
v_aa = _mm_xor_si128(_mm_xor_si128(v_t1, v_t2), v_s0_8F80);
while (((uint64_t)(a_x.len)) >= 32u) {
v_x0 = _mm_lddqu_si128((const __m128i*)(const void*)(a_x.ptr + 0u));
a_x = wuffs_base__slice_u8__subslice_i(a_x, 16u);
v_u0 = _mm_xor_si128(v_aa, v_x0);
v_u1 = _mm_clmulepi64_si128(v_u0, v_k1k2, (int32_t)(0u));
v_u2 = _mm_clmulepi64_si128(v_u0, v_k1k2, (int32_t)(17u));
v_aa = _mm_xor_si128(v_u1, v_u2);
}
if (((uint64_t)(a_x.len)) < 16u) {
return wuffs_base__make_empty_struct();
}
v_x0 = _mm_lddqu_si128((const __m128i*)(const void*)(a_x.ptr + 0u));
a_x = wuffs_base__slice_u8__subslice_i(a_x, 16u);
v_v0 = _mm_xor_si128(v_aa, v_x0);
v_v1 = _mm_clmulepi64_si128(v_v0, v_k1k2, (int32_t)(16u));
v_aa = _mm_xor_si128(v_v1, _mm_srli_si128(v_v0, (int32_t)(8u)));
v_pxmu = _mm_lddqu_si128((const __m128i*)(const void*)(WUFFS_CRC64__ECMA_X86_SSE42_PXMU));
v_w1 = _mm_clmulepi64_si128(v_aa, v_pxmu, (int32_t)(16u));
v_w2 = _mm_clmulepi64_si128(v_w1, v_pxmu, (int32_t)(0u));
v_s = ((uint64_t)(_mm_extract_epi64(_mm_xor_si128(v_aa, _mm_xor_si128(v_w2, _mm_slli_si128(v_w1, (int32_t)(8u)))), (int32_t)(1u))));
v_tail_index = (((uint64_t)(a_x.len)) & 18446744073709551600u);
if (v_tail_index < ((uint64_t)(a_x.len))) {
{
wuffs_base__slice_u8 i_slice_p = wuffs_base__slice_u8__subslice_i(a_x, v_tail_index);
v_p.ptr = i_slice_p.ptr;
v_p.len = 1;
uint8_t* i_end0_p = i_slice_p.ptr + i_slice_p.len;
while (v_p.ptr < i_end0_p) {
v_s = (WUFFS_CRC64__ECMA_TABLE[0u][((uint8_t)(((uint8_t)(v_s)) ^ v_p.ptr[0u]))] ^ (v_s >> 8u));
v_p.ptr += 1;
}
v_p.len = 0;
}
}
self->private_impl.f_state = (18446744073709551615u ^ v_s);
return wuffs_base__make_empty_struct();
}
#endif // defined(WUFFS_BASE__CPU_ARCH__X86_FAMILY)
// ‼ WUFFS MULTI-FILE SECTION -x86_sse42
#endif // !defined(WUFFS_CONFIG__MODULES) || defined(WUFFS_CONFIG__MODULE__CRC64)
#if !defined(WUFFS_CONFIG__MODULES) || defined(WUFFS_CONFIG__MODULE__DEFLATE)
@@ -0,0 +1,110 @@
// Copyright 2024 The Wuffs Authors.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.
//
// SPDX-License-Identifier: Apache-2.0 OR MIT
//go:build ignore
// +build ignore
package main
// print-crc64-x86-sse42-magic-numbers.go prints the std/crc64
// ECMA_X86_SSE42_ETC magic number tables.
//
// It is like print-crc32-x86-sse42-magic-numbers.go but for CRC-64/ECMA, not
// CRC-32/IEEE.
//
// Usage: go run print-crc64-x86-sse42-magic-numbers.go
//
// Output:
// Px' = 0x92D8_AF2B_AF0E_1E85
// k1' = 0xE05D_D497_CA39_3AE4
// k2' = 0xDABE_95AF_C787_5F40
// μ' = 0x9C3E_466C_1729_63D5
import (
"fmt"
"strings"
)
// px is the P(x) polynomial, a bit-reversal (with explicit high bit) of the
// CRC-64/ECMA polynomial sometimes written as 0xC96C_5795_D787_0F42.
// P(x) = 0x1_42F0_E1EB_A9EA_3693
// P(x) = 0b1_01000010_11110000_11100001_11101011_10101001_11101010_00110110_10010011
const px = "10100001011110000111000011110101110101001111010100011011010010011"
var spaces = strings.Repeat(" ", 1024)
func debugf(format string, a ...interface{}) {
if false { // Change false to true to show the long divisions.
fmt.Printf(format, a...)
}
}
func show(name string, value string) {
v := uint64(0)
if strings.Contains(name, "'") {
for i := len(value) - 1; i >= 0; i-- {
v = (v << 1) | uint64(value[i]&1)
}
} else {
for i := 0; i < len(value); i++ {
v = (v << 1) | uint64(value[i]&1)
}
}
fmt.Printf("%s = 0x%04X_%04X_%04X_%04X\n", name,
0xFFFF&(v>>48), 0xFFFF&(v>>32), 0xFFFF&(v>>16), 0xFFFF&(v>>0))
}
func calcKn(name string, power int) {
numerator := "1" + strings.Repeat("0", power)
b := []byte(numerator)
i := 0
debugf(" %s\n", numerator)
for i+len(px) <= len(numerator) {
for j := 0; j < len(px); j++ {
b[i+j] ^= 1 & px[j]
}
debugf(" %s%s\n", spaces[:i], px)
for ; (i < len(b)) && (b[i] == '0'); i++ {
b[i] = ' '
}
debugf(" %s\n", b)
}
show(name, string(b[len(b)-len(px):]))
}
func calcMu(name string) {
numerator := "1" + strings.Repeat("0", 128)
mu := make([]byte, 129)
b := []byte(numerator)
i := 0
debugf(" %s\n", numerator)
for i+len(px) <= len(numerator) {
for j := 0; j < len(px); j++ {
b[i+j] ^= 1 & px[j]
}
debugf(" %s%s\n", spaces[:i], px)
b[i] = ':'
mu[i] = '1'
i++
for ; (i < len(b)) && (b[i] == '0'); i++ {
b[i] = ' '
mu[i] = '0'
}
debugf(" %s\n", b)
}
show(name, string(mu[:65]))
}
func main() {
show("Px'", px)
calcKn("k1'", 128+64)
calcKn("k2'", 128)
calcMu("μ' ")
}
+15 -5
View File
@@ -21,6 +21,21 @@ pub func ecma_hasher.set_quirk!(key: base.u32, value: base.u64) base.status {
}
pub func ecma_hasher.update!(x: roslice base.u8) {
if this.state == 0 {
choose up = [
up_x86_sse42]
}
this.up!(x: args.x)
}
pub func ecma_hasher.update_u64!(x: roslice base.u8) base.u64 {
this.update!(x: args.x)
return this.checksum_u64()
}
pri func ecma_hasher.up!(x: roslice base.u8),
choosy,
{
var s : base.u64
var p : roslice base.u8
@@ -51,11 +66,6 @@ pub func ecma_hasher.update!(x: roslice base.u8) {
this.state = 0xFFFF_FFFF_FFFF_FFFF ^ s
}
pub func ecma_hasher.update_u64!(x: roslice base.u8) base.u64 {
this.update!(x: args.x)
return this.checksum_u64()
}
pub func ecma_hasher.checksum_u64() base.u64 {
return this.state
}
+145
View File
@@ -0,0 +1,145 @@
// Copyright 2024 The Wuffs Authors.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.
//
// SPDX-License-Identifier: Apache-2.0 OR MIT
// --------
// Like std/crc32's x86 SIMD implementation, this is based on Gopal et al.
// "Fast CRC Computation for Generic Polynomials Using PCLMULQDQ Instruction".
pri func ecma_hasher.up_x86_sse42!(x: roslice base.u8),
choose cpu_arch >= x86_sse42,
{
var s : base.u64
var p : roslice base.u8
var util : base.x86_sse42_utility
var s0 : base.x86_m128i
var s0_707F : base.x86_m128i
var s0_8F80 : base.x86_m128i
var x0 : base.x86_m128i
var aa : base.x86_m128i
var k1k2 : base.x86_m128i
var t0 : base.x86_m128i
var t1 : base.x86_m128i
var t2 : base.x86_m128i
var u0 : base.x86_m128i
var u1 : base.x86_m128i
var u2 : base.x86_m128i
var v0 : base.x86_m128i
var v1 : base.x86_m128i
var pxmu : base.x86_m128i
var w1 : base.x86_m128i
var w2 : base.x86_m128i
var tail_index : base.u64
s = 0xFFFF_FFFF_FFFF_FFFF ^ this.state
// Align to a 16-byte boundary.
while (args.x.length() > 0) and ((15 & args.x.uintptr_low_12_bits()) <> 0) {
s = ECMA_TABLE[0][((s & 0xFF) as base.u8) ^ args.x[0]] ^ (s >> 8)
args.x = args.x[1 ..]
} endwhile
// For short inputs, just do a simple loop.
if args.x.length() < 0x20 {
iterate (p = args.x)(length: 1, advance: 1, unroll: 1) {
s = ECMA_TABLE[0][((s & 0xFF) as base.u8) ^ p[0]] ^ (s >> 8)
}
this.state = 0xFFFF_FFFF_FFFF_FFFF ^ s
return nothing
}
// Process N 16-byte chunks (1 on ramp, N-2 main loop, 1 off ramp). The
// algorithm is loosely based on https://github.com/rawrunprotected/crc but
// modified to always start and end on 16-byte alignment boundaries. It
// also doesn't crash on zero-length input.
// On ramp.
s0 = util.make_m128i_single_u64(a: s)
s0_707F = s0._mm_shuffle_epi8(b: util.make_m128i_slice128(a: SHUFFLE_707F[.. 16]))
s0_8F80 = s0._mm_shuffle_epi8(b: util.make_m128i_slice128(a: SHUFFLE_8F80[.. 16]))
x0 = util.make_m128i_slice128(a: args.x[0x00 .. 0x10])
args.x = args.x[0x10 ..]
k1k2 = util.make_m128i_slice128(a: ECMA_X86_SSE42_K1K2[.. 16])
t0 = s0_707F._mm_xor_si128(b: x0)
t1 = t0._mm_clmulepi64_si128(b: k1k2, imm8: 0x00)
t2 = t0._mm_clmulepi64_si128(b: k1k2, imm8: 0x11)
aa = t1._mm_xor_si128(b: t2)._mm_xor_si128(b: s0_8F80)
// Main loop.
while args.x.length() >= 0x20 {
x0 = util.make_m128i_slice128(a: args.x[0x00 .. 0x10])
args.x = args.x[0x10 ..]
u0 = aa._mm_xor_si128(b: x0)
u1 = u0._mm_clmulepi64_si128(b: k1k2, imm8: 0x00)
u2 = u0._mm_clmulepi64_si128(b: k1k2, imm8: 0x11)
aa = u1._mm_xor_si128(b: u2)
} endwhile
// Off ramp.
if args.x.length() < 0x10 {
// Unreachable, but the "if" is needed by the bounds checker.
return nothing
}
x0 = util.make_m128i_slice128(a: args.x[0x00 .. 0x10])
args.x = args.x[0x10 ..]
v0 = aa._mm_xor_si128(b: x0)
v1 = v0._mm_clmulepi64_si128(b: k1k2, imm8: 0x10)
aa = v1._mm_xor_si128(b: v0._mm_srli_si128(imm8: 8))
// Barrett reduction.
pxmu = util.make_m128i_slice128(a: ECMA_X86_SSE42_PXMU[.. 16])
w1 = aa._mm_clmulepi64_si128(b: pxmu, imm8: 0x10)
w2 = w1._mm_clmulepi64_si128(b: pxmu, imm8: 0x00)
s = aa._mm_xor_si128(b:
w2._mm_xor_si128(b:
w1._mm_slli_si128(imm8: 8)))._mm_extract_epi64(imm8: 1)
// Handle the tail of args.x that wasn't a complete 16-byte chunk.
tail_index = args.x.length() & 0xFFFF_FFFF_FFFF_FFF0 // And-not 16.
if tail_index < args.x.length() {
iterate (p = args.x[tail_index ..])(length: 1, advance: 1, unroll: 1) {
s = ECMA_TABLE[0][((s & 0xFF) as base.u8) ^ p[0]] ^ (s >> 8)
}
}
this.state = 0xFFFF_FFFF_FFFF_FFFF ^ s
}
pri const SHUFFLE_707F : roarray[16] base.u8 = [
0x70, 0x71, 0x72, 0x73, 0x74, 0x75, 0x76, 0x77,
0x78, 0x79, 0x7A, 0x7B, 0x7C, 0x7D, 0x7E, 0x7F,
]
pri const SHUFFLE_8F80 : roarray[16] base.u8 = [
0x8F, 0x8E, 0x8D, 0x8C, 0x8B, 0x8A, 0x89, 0x88,
0x87, 0x86, 0x85, 0x84, 0x83, 0x82, 0x81, 0x80,
]
// These constants are reproduced by
// script/print-crc64-x86-sse42-magic-numbers.go
pri const ECMA_X86_SSE42_K1K2 : roarray[16] base.u8 = [
0xE4, 0x3A, 0x39, 0xCA, 0x97, 0xD4, 0x5D, 0xE0, // k1' = 0xE05D_D497_CA39_3AE4
0x40, 0x5F, 0x87, 0xC7, 0xAF, 0x95, 0xBE, 0xDA, // k2' = 0xDABE_95AF_C787_5F40
]
pri const ECMA_X86_SSE42_PXMU : roarray[16] base.u8 = [
0x85, 0x1E, 0x0E, 0xAF, 0x2B, 0xAF, 0xD8, 0x92, // Px' = 0x92D8_AF2B_AF0E_1E85
0xD5, 0x63, 0x29, 0x17, 0x6C, 0x46, 0x3E, 0x9C, // μ' = 0x9C3E_466C_1729_63D5
]