diff options
| author |  Runxi Yu | 2026-02-22 11:38:33 +0800 |
|---|---|---|
| committer | Runxi Yu | 2026-02-22 11:38:33 +0800 |
| commit | 5e26980abad6374917aafad7e6c8120cd2e55d63 (patch) | |
| tree | d09e39258c666be1d513d77cf7ed6f3aec048ee5 /internal | |
| parent | objectstore/chain: Fix lints (diff) | |
| signature | No signature | |
adler32: Import
Diffstat (limited to 'internal')
| -rw-r--r-- | internal/adler32/LICENSE | 30 | ||||
| -rw-r--r-- | internal/adler32/LICENSE.ZLIB | 17 | ||||
| -rw-r--r-- | internal/adler32/README | 1 | ||||
| -rw-r--r-- | internal/adler32/adler32_amd64.go | 93 | ||||
| -rw-r--r-- | internal/adler32/adler32_arm64.go | 73 | ||||
| -rw-r--r-- | internal/adler32/adler32_avx2.go | 6 | ||||
| -rw-r--r-- | internal/adler32/adler32_avx2.s | 263 | ||||
| -rw-r--r-- | internal/adler32/adler32_fallback.go | 19 | ||||
| -rw-r--r-- | internal/adler32/adler32_generic.go | 44 | ||||
| -rw-r--r-- | internal/adler32/adler32_neon.go | 6 | ||||
| -rw-r--r-- | internal/adler32/adler32_neon.s | 208 | ||||
| -rw-r--r-- | internal/adler32/adler32_sse3.go | 6 | ||||
| -rw-r--r-- | internal/adler32/adler32_sse3.s | 214 | ||||
| -rw-r--r-- | internal/adler32/bench_test.go | 22 |
14 files changed, 1002 insertions, 0 deletions
diff --git a/internal/adler32/LICENSE b/internal/adler32/LICENSE new file mode 100644 index 00000000..5cec357a --- /dev/null +++ b/internal/adler32/LICENSE @@ -0,0 +1,30 @@ +Copyright (c) 2024, Michal Hruby +Copyright (c) 2017 The Chromium Authors. All rights reserved. +Copyright (c) 1995-2024 Mark Adler +Copyright (c) 1995-2024 Jean-loup Gailly +Copyright (c) 2022 Adam Stylinski + +BSD 2-Clause License + + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +1. Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +2. Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + diff --git a/internal/adler32/LICENSE.ZLIB b/internal/adler32/LICENSE.ZLIB new file mode 100644 index 00000000..c75c1568 --- /dev/null +++ b/internal/adler32/LICENSE.ZLIB @@ -0,0 +1,17 @@ +Copyright (C) 1995-2024 Jean-loup Gailly and Mark Adler + +This software is provided 'as-is', without any express or implied +warranty. In no event will the authors be held liable for any damages +arising from the use of this software. + +Permission is granted to anyone to use this software for any purpose, +including commercial applications, and to alter it and redistribute it +freely, subject to the following restrictions: + +1. The origin of this software must not be misrepresented; you must not + claim that you wrote the original software. If you use this software + in a product, an acknowledgment in the product documentation would be + appreciated but is not required. +2. Altered source versions must be plainly marked as such, and must not be + misrepresented as being the original software. +3. This notice may not be removed or altered from any source distribution. diff --git a/internal/adler32/README b/internal/adler32/README new file mode 100644 index 00000000..b80acd00 --- /dev/null +++ b/internal/adler32/README @@ -0,0 +1 @@ +This package was mostly copied from github.com/mhr3/adler32-simd. diff --git a/internal/adler32/adler32_amd64.go b/internal/adler32/adler32_amd64.go new file mode 100644 index 00000000..88a854ed --- /dev/null +++ b/internal/adler32/adler32_amd64.go @@ -0,0 +1,93 @@ +//go:build amd64 && !purego + +package adler32 + +import ( + "encoding/binary" + "errors" + "hash" + "hash/adler32" + + "golang.org/x/sys/cpu" +) + +// The size of an Adler-32 checksum in bytes. +const Size = 4 + +var ( + hasSSE3 = cpu.X86.HasSSE3 + hasAVX2 = cpu.X86.HasAVX2 +) + +// digest represents the partial evaluation of a checksum. +// The low 16 bits are s1, the high 16 bits are s2. +type digest uint32 + +func (d *digest) Reset() { *d = 1 } + +// New returns a new hash.Hash32 computing the Adler-32 checksum. +func New() hash.Hash32 { + if !hasSSE3 { + return adler32.New() + } + d := new(digest) + d.Reset() + return d +} + +func (d *digest) MarshalBinary() ([]byte, error) { + b := make([]byte, 0, marshaledSize) + b = append(b, magic...) + b = binary.BigEndian.AppendUint32(b, uint32(*d)) + return b, nil +} + +func (d *digest) UnmarshalBinary(b []byte) error { + if len(b) < len(magic) || string(b[:len(magic)]) != magic { + return errors.New("hash/adler32: invalid hash state identifier") + } + if len(b) != marshaledSize { + return errors.New("hash/adler32: invalid hash state size") + } + *d = digest(binary.BigEndian.Uint32(b[len(magic):])) + return nil +} + +func (d *digest) Size() int { return Size } + +func (d *digest) BlockSize() int { return 4 } + +func (d *digest) Write(data []byte) (nn int, err error) { + if len(data) >= 64 { + var h uint32 + if hasAVX2 { + h = adler32_avx2(uint32(*d), data) + } else { + h = adler32_sse3(uint32(*d), data) + } + *d = digest(h) + } else { + h := update(uint32(*d), data) + *d = digest(h) + } + return len(data), nil +} + +func (d *digest) Sum32() uint32 { return uint32(*d) } + +func (d *digest) Sum(in []byte) []byte { + s := uint32(*d) + return append(in, byte(s>>24), byte(s>>16), byte(s>>8), byte(s)) +} + +// Checksum returns the Adler-32 checksum of data. +func Checksum(data []byte) uint32 { + if !hasSSE3 || len(data) < 64 { + return update(1, data) + } + + if hasAVX2 { + return adler32_avx2(1, data) + } + return adler32_sse3(1, data) +} diff --git a/internal/adler32/adler32_arm64.go b/internal/adler32/adler32_arm64.go new file mode 100644 index 00000000..ddf9cb5e --- /dev/null +++ b/internal/adler32/adler32_arm64.go @@ -0,0 +1,73 @@ +//go:build arm64 && !purego + +package adler32 + +import ( + "encoding/binary" + "errors" + "hash" +) + +// The size of an Adler-32 checksum in bytes. +const Size = 4 + +// digest represents the partial evaluation of a checksum. +// The low 16 bits are s1, the high 16 bits are s2. +type digest uint32 + +func (d *digest) Reset() { *d = 1 } + +// New returns a new hash.Hash32 computing the Adler-32 checksum. +func New() hash.Hash32 { + d := new(digest) + d.Reset() + return d +} + +func (d *digest) MarshalBinary() ([]byte, error) { + b := make([]byte, 0, marshaledSize) + b = append(b, magic...) + b = binary.BigEndian.AppendUint32(b, uint32(*d)) + return b, nil +} + +func (d *digest) UnmarshalBinary(b []byte) error { + if len(b) < len(magic) || string(b[:len(magic)]) != magic { + return errors.New("hash/adler32: invalid hash state identifier") + } + if len(b) != marshaledSize { + return errors.New("hash/adler32: invalid hash state size") + } + *d = digest(binary.BigEndian.Uint32(b[len(magic):])) + return nil +} + +func (d *digest) Size() int { return Size } + +func (d *digest) BlockSize() int { return 4 } + +func (d *digest) Write(data []byte) (nn int, err error) { + if len(data) >= 64 { + h := adler32_neon(uint32(*d), data) + *d = digest(h) + } else { + h := update(uint32(*d), data) + *d = digest(h) + } + return len(data), nil +} + +func (d *digest) Sum32() uint32 { return uint32(*d) } + +func (d *digest) Sum(in []byte) []byte { + s := uint32(*d) + return append(in, byte(s>>24), byte(s>>16), byte(s>>8), byte(s)) +} + +// Checksum returns the Adler-32 checksum of data. +func Checksum(data []byte) uint32 { + if len(data) >= 64 { + return adler32_neon(1, data) + } + return update(1, data) +} diff --git a/internal/adler32/adler32_avx2.go b/internal/adler32/adler32_avx2.go new file mode 100644 index 00000000..042812b8 --- /dev/null +++ b/internal/adler32/adler32_avx2.go @@ -0,0 +1,6 @@ +//go:build !purego && amd64 + +package adler32 + +//go:noescape +func adler32_avx2(in uint32, buf []byte) uint32 diff --git a/internal/adler32/adler32_avx2.s b/internal/adler32/adler32_avx2.s new file mode 100644 index 00000000..60879fcc --- /dev/null +++ b/internal/adler32/adler32_avx2.s @@ -0,0 +1,263 @@ +//go:build !purego && amd64 + +#include "textflag.h" + +DATA LCPI0_0<>+0x00(SB)/8, $0x191a1b1c1d1e1f20 +DATA LCPI0_0<>+0x08(SB)/8, $0x1112131415161718 +DATA LCPI0_0<>+0x10(SB)/8, $0x090a0b0c0d0e0f10 +DATA LCPI0_0<>+0x18(SB)/8, $0x0102030405060708 +GLOBL LCPI0_0<>(SB), (RODATA|NOPTR), $32 + +DATA LCPI0_1<>+0x00(SB)/8, $0x0001000100010001 +DATA LCPI0_1<>+0x08(SB)/8, $0x0001000100010001 +DATA LCPI0_1<>+0x10(SB)/8, $0x0001000100010001 +DATA LCPI0_1<>+0x18(SB)/8, $0x0001000100010001 +GLOBL LCPI0_1<>(SB), (RODATA|NOPTR), $32 + +DATA LCPI0_2<>+0x00(SB)/2, $0x0001 +GLOBL LCPI0_2<>(SB), (RODATA|NOPTR), $2 + +TEXT ·adler32_avx2(SB), NOSPLIT, $0-36 + MOVLQZX in+0(FP), DI + MOVQ buf_base+8(FP), SI + MOVQ buf_len+16(FP), DX + MOVQ buf_cap+24(FP), CX + WORD $0x8548; BYTE $0xf6 // TESTQ SI, SI // test rsi, rsi + JE LBB0_1 // <-- // je .LBB0_1 + WORD $0xf889 // MOVL DI, AX // mov eax, edi + WORD $0x8548; BYTE $0xd2 // TESTQ DX, DX // test rdx, rdx + JE LBB0_2 // <-- // je .LBB0_2 + NOP // (skipped) // push rbp + NOP // (skipped) // mov rbp, rsp + NOP // (skipped) // and rsp, -8 + WORD $0xc189 // MOVL AX, CX // mov ecx, eax + WORD $0xe9c1; BYTE $0x10 // SHRL $0x10, CX // shr ecx, 16 + WORD $0xb70f; BYTE $0xc0 // MOVZX AX, AX // movzx eax, ax + CMPQ DX, $0x20 // <-- // cmp rdx, 32 + JB LBB0_17 // <-- // jb .LBB0_17 + LONG $0x078071bf; BYTE $0x80 // MOVL $-0x7ff87f8f, DI // mov edi, 2147975281 + LONG $0xc0eff9c5 // VPXOR X0, X0, X0 // vpxor xmm0, xmm0, xmm0 + VMOVDQA LCPI0_0<>(SB), Y1 // <-- // vmovdqa ymm1, ymmword ptr [rip + .LCPI0_0] + VPBROADCASTW LCPI0_2<>(SB), Y2 // <-- // vpbroadcastw ymm2, word ptr [rip + .LCPI0_2] + JMP LBB0_6 // <-- // jmp .LBB0_6 + +LBB0_7: + LONG $0xf46ffdc5 // VMOVDQA Y4, Y6 // vmovdqa ymm6, ymm4 + LONG $0xedefd1c5 // VPXOR X5, X5, X5 // vpxor xmm5, xmm5, xmm5 + +LBB0_14: + SUBQ AX, DX // <-- // sub rdx, rax + LONG $0xf572ddc5; BYTE $0x05 // ? // vpslld ymm4, ymm5, 5 + LONG $0xdbfeddc5 // VPADDD Y3, Y4, Y3 // vpaddd ymm3, ymm4, ymm3 + LONG $0x397de3c4; WORD $0x01f4 // VEXTRACTI128 $0x1, Y6, X4 // vextracti128 xmm4, ymm6, 1 + LONG $0xecc6c8c5; BYTE $0x88 // VSHUFPS $-0x78, X4, X6, X5 // vshufps xmm5, xmm6, xmm4, 136 + LONG $0xe470f9c5; BYTE $0x88 // VPSHUFD $-0x78, X4, X4 // vpshufd xmm4, xmm4, 136 + LONG $0xe4fed1c5 // VPADDD X4, X5, X4 // vpaddd xmm4, xmm5, xmm4 + LONG $0xec70f9c5; BYTE $0x55 // VPSHUFD $0x55, X4, X5 // vpshufd xmm5, xmm4, 85 + LONG $0xe4fed1c5 // VPADDD X4, X5, X4 // vpaddd xmm4, xmm5, xmm4 + LONG $0xe07ef9c5 // VMOVD X4, AX // vmovd eax, xmm4 + MOVQ AX, CX // <-- // mov rcx, rax + IMULQ DI, CX // <-- // imul rcx, rdi + SHRQ $0x2f, CX // <-- // shr rcx, 47 + LONG $0xfff1c969; WORD $0x0000 // IMULL $0xfff1, CX, CX // imul ecx, ecx, 65521 + WORD $0xc829 // SUBL CX, AX // sub eax, ecx + LONG $0x397de3c4; WORD $0x01dc // VEXTRACTI128 $0x1, Y3, X4 // vextracti128 xmm4, ymm3, 1 + LONG $0xdbfed9c5 // VPADDD X3, X4, X3 // vpaddd xmm3, xmm4, xmm3 + LONG $0xe370f9c5; BYTE $0xee // VPSHUFD $-0x12, X3, X4 // vpshufd xmm4, xmm3, 238 + LONG $0xdcfee1c5 // VPADDD X4, X3, X3 // vpaddd xmm3, xmm3, xmm4 + LONG $0xe370f9c5; BYTE $0x55 // VPSHUFD $0x55, X3, X4 // vpshufd xmm4, xmm3, 85 + LONG $0xdbfed9c5 // VPADDD X3, X4, X3 // vpaddd xmm3, xmm4, xmm3 + LONG $0xd97ef9c5 // VMOVD X3, CX // vmovd ecx, xmm3 + MOVQ CX, R8 // <-- // mov r8, rcx + IMULQ DI, R8 // <-- // imul r8, rdi + SHRQ $0x2f, R8 // <-- // shr r8, 47 + LONG $0xf1c06945; WORD $0x00ff; BYTE $0x00 // IMULL $0xfff1, R8, R8 // imul r8d, r8d, 65521 + WORD $0x2944; BYTE $0xc1 // SUBL R8, CX // sub ecx, r8d + CMPQ DX, $0x1f // <-- // cmp rdx, 31 + JBE LBB0_15 // <-- // jbe .LBB0_15 + +LBB0_6: + LONG $0xe06ef9c5 // VMOVD AX, X4 // vmovd xmm4, eax + LONG $0xd96ef9c5 // VMOVD CX, X3 // vmovd xmm3, ecx + CMPQ DX, $0x15b0 // <-- // cmp rdx, 5552 + LONG $0x15b0b841; WORD $0x0000 // MOVL $0x15b0, R8 // mov r8d, 5552 + LONG $0xc2420f4c // CMOVB DX, R8 // cmovb r8, rdx + WORD $0x8944; BYTE $0xc0 // MOVL R8, AX // mov eax, r8d + LONG $0x001fe025; BYTE $0x00 // ANDL $0x1fe0, AX // and eax, 8160 + JE LBB0_7 // <-- // je .LBB0_7 + ADDQ $-0x20, R8 // <-- // add r8, -32 + LONG $0xedefd1c5 // VPXOR X5, X5, X5 // vpxor xmm5, xmm5, xmm5 + LONG $0x20c0f641 // TESTL $0x20, R8 // test r8b, 32 + JNE LBB0_9 // <-- // jne .LBB0_9 + LONG $0x2e6ffec5 // VMOVDQU 0(SI), Y5 // vmovdqu ymm5, ymmword ptr [rsi] + ADDQ $0x20, SI // <-- // add rsi, 32 + LEAQ -0x20(AX), CX // <-- // lea rcx, [rax - 32] + LONG $0xf0f6d5c5 // VPSADBW Y0, Y5, Y6 // vpsadbw ymm6, ymm5, ymm0 + LONG $0xf4fecdc5 // VPADDD Y4, Y6, Y6 // vpaddd ymm6, ymm6, ymm4 + LONG $0x0455e2c4; BYTE $0xe9 // VPMADDUBSW Y1, Y5, Y5 // vpmaddubsw ymm5, ymm5, ymm1 + LONG $0xeaf5d5c5 // VPMADDWD Y2, Y5, Y5 // vpmaddwd ymm5, ymm5, ymm2 + LONG $0xdbfed5c5 // VPADDD Y3, Y5, Y3 // vpaddd ymm3, ymm5, ymm3 + LONG $0xec6ffdc5 // VMOVDQA Y4, Y5 // vmovdqa ymm5, ymm4 + LONG $0xe66ffdc5 // VMOVDQA Y6, Y4 // vmovdqa ymm4, ymm6 + CMPQ R8, $0x20 // <-- // cmp r8, 32 + JAE LBB0_12 // <-- // jae .LBB0_12 + JMP LBB0_14 // <-- // jmp .LBB0_14 + +LBB0_9: + MOVQ AX, CX // <-- // mov rcx, rax + CMPQ R8, $0x20 // <-- // cmp r8, 32 + JB LBB0_14 // <-- // jb .LBB0_14 + +LBB0_12: + LONG $0x366ffec5 // VMOVDQU 0(SI), Y6 // vmovdqu ymm6, ymmword ptr [rsi] + LONG $0x7e6ffec5; BYTE $0x20 // VMOVDQU 0x20(SI), Y7 // vmovdqu ymm7, ymmword ptr [rsi + 32] + LONG $0xc0f64dc5 // VPSADBW Y0, Y6, Y8 // vpsadbw ymm8, ymm6, ymm0 + LONG $0xc4fe3dc5 // VPADDD Y4, Y8, Y8 // vpaddd ymm8, ymm8, ymm4 + LONG $0xecfed5c5 // VPADDD Y4, Y5, Y5 // vpaddd ymm5, ymm5, ymm4 + LONG $0x044de2c4; BYTE $0xe1 // VPMADDUBSW Y1, Y6, Y4 // vpmaddubsw ymm4, ymm6, ymm1 + LONG $0xe2f5ddc5 // VPMADDWD Y2, Y4, Y4 // vpmaddwd ymm4, ymm4, ymm2 + LONG $0xdbfeddc5 // VPADDD Y3, Y4, Y3 // vpaddd ymm3, ymm4, ymm3 + ADDQ $0x40, SI // <-- // add rsi, 64 + LONG $0xe0f6c5c5 // VPSADBW Y0, Y7, Y4 // vpsadbw ymm4, ymm7, ymm0 + LONG $0xe4febdc5 // VPADDD Y4, Y8, Y4 // vpaddd ymm4, ymm8, ymm4 + LONG $0xedfebdc5 // VPADDD Y5, Y8, Y5 // vpaddd ymm5, ymm8, ymm5 + LONG $0x0445e2c4; BYTE $0xf1 // VPMADDUBSW Y1, Y7, Y6 // vpmaddubsw ymm6, ymm7, ymm1 + LONG $0xf2f5cdc5 // VPMADDWD Y2, Y6, Y6 // vpmaddwd ymm6, ymm6, ymm2 + LONG $0xdbfecdc5 // VPADDD Y3, Y6, Y3 // vpaddd ymm3, ymm6, ymm3 + ADDQ $-0x40, CX // <-- // add rcx, -64 + JNE LBB0_12 // <-- // jne .LBB0_12 + LONG $0xf46ffdc5 // VMOVDQA Y4, Y6 // vmovdqa ymm6, ymm4 + JMP LBB0_14 // <-- // jmp .LBB0_14 + +LBB0_1: + LONG $0x000001b8; BYTE $0x00 // MOVL $0x1, AX // mov eax, 1 + +LBB0_2: + MOVL AX, ret+32(FP) // <-- + RET // <-- // ret + +LBB0_15: + WORD $0x8548; BYTE $0xd2 // TESTQ DX, DX // test rdx, rdx + JE LBB0_16 // <-- // je .LBB0_16 + +LBB0_17: + CMPQ DX, $0x10 // <-- // cmp rdx, 16 + JB LBB0_20 // <-- // jb .LBB0_20 + WORD $0xb60f; BYTE $0x3e // MOVZX 0(SI), DI // movzx edi, byte ptr [rsi] + WORD $0xf801 // ADDL DI, AX // add eax, edi + WORD $0xc101 // ADDL AX, CX // add ecx, eax + LONG $0x017eb60f // MOVZX 0x1(SI), DI // movzx edi, byte ptr [rsi + 1] + WORD $0xc701 // ADDL AX, DI // add edi, eax + WORD $0xf901 // ADDL DI, CX // add ecx, edi + LONG $0x0246b60f // MOVZX 0x2(SI), AX // movzx eax, byte ptr [rsi + 2] + WORD $0xf801 // ADDL DI, AX // add eax, edi + WORD $0xc101 // ADDL AX, CX // add ecx, eax + LONG $0x037eb60f // MOVZX 0x3(SI), DI // movzx edi, byte ptr [rsi + 3] + WORD $0xc701 // ADDL AX, DI // add edi, eax + WORD $0xf901 // ADDL DI, CX // add ecx, edi + LONG $0x0446b60f // MOVZX 0x4(SI), AX // movzx eax, byte ptr [rsi + 4] + WORD $0xf801 // ADDL DI, AX // add eax, edi + WORD $0xc101 // ADDL AX, CX // add ecx, eax + LONG $0x057eb60f // MOVZX 0x5(SI), DI // movzx edi, byte ptr [rsi + 5] + WORD $0xc701 // ADDL AX, DI // add edi, eax + WORD $0xf901 // ADDL DI, CX // add ecx, edi + LONG $0x0646b60f // MOVZX 0x6(SI), AX // movzx eax, byte ptr [rsi + 6] + WORD $0xf801 // ADDL DI, AX // add eax, edi + WORD $0xc101 // ADDL AX, CX // add ecx, eax + LONG $0x077eb60f // MOVZX 0x7(SI), DI // movzx edi, byte ptr [rsi + 7] + WORD $0xc701 // ADDL AX, DI // add edi, eax + WORD $0xf901 // ADDL DI, CX // add ecx, edi + LONG $0x0846b60f // MOVZX 0x8(SI), AX // movzx eax, byte ptr [rsi + 8] + WORD $0xf801 // ADDL DI, AX // add eax, edi + WORD $0xc101 // ADDL AX, CX // add ecx, eax + LONG $0x097eb60f // MOVZX 0x9(SI), DI // movzx edi, byte ptr [rsi + 9] + WORD $0xc701 // ADDL AX, DI // add edi, eax + WORD $0xf901 // ADDL DI, CX // add ecx, edi + LONG $0x0a46b60f // MOVZX 0xa(SI), AX // movzx eax, byte ptr [rsi + 10] + WORD $0xf801 // ADDL DI, AX // add eax, edi + WORD $0xc101 // ADDL AX, CX // add ecx, eax + LONG $0x0b7eb60f // MOVZX 0xb(SI), DI // movzx edi, byte ptr [rsi + 11] + WORD $0xc701 // ADDL AX, DI // add edi, eax + WORD $0xf901 // ADDL DI, CX // add ecx, edi + LONG $0x0c46b60f // MOVZX 0xc(SI), AX // movzx eax, byte ptr [rsi + 12] + WORD $0xf801 // ADDL DI, AX // add eax, edi + WORD $0xc101 // ADDL AX, CX // add ecx, eax + LONG $0x0d7eb60f // MOVZX 0xd(SI), DI // movzx edi, byte ptr [rsi + 13] + WORD $0xc701 // ADDL AX, DI // add edi, eax + WORD $0xf901 // ADDL DI, CX // add ecx, edi + LONG $0x46b60f44; BYTE $0x0e // MOVZX 0xe(SI), R8 // movzx r8d, byte ptr [rsi + 14] + WORD $0x0141; BYTE $0xf8 // ADDL DI, R8 // add r8d, edi + WORD $0x0144; BYTE $0xc1 // ADDL R8, CX // add ecx, r8d + LONG $0x0f46b60f // MOVZX 0xf(SI), AX // movzx eax, byte ptr [rsi + 15] + WORD $0x0144; BYTE $0xc0 // ADDL R8, AX // add eax, r8d + WORD $0xc101 // ADDL AX, CX // add ecx, eax + ADDQ $-0x10, DX // <-- // add rdx, -16 + JE LBB0_27 // <-- // je .LBB0_27 + ADDQ $0x10, SI // <-- // add rsi, 16 + +LBB0_20: + LEAQ -0x1(DX), DI // <-- // lea rdi, [rdx - 1] + MOVQ DX, R9 // <-- // mov r9, rdx + ANDQ $0x3, R9 // <-- // and r9, 3 + JE LBB0_24 // <-- // je .LBB0_24 + XORL R8, R8 // <-- // xor r8d, r8d + +LBB0_22: + LONG $0x14b60f46; BYTE $0x06 // MOVZX 0(SI)(R8*1), R10 // movzx r10d, byte ptr [rsi + r8] + WORD $0x0144; BYTE $0xd0 // ADDL R10, AX // add eax, r10d + WORD $0xc101 // ADDL AX, CX // add ecx, eax + INCQ R8 // <-- // inc r8 + CMPQ R9, R8 // <-- // cmp r9, r8 + JNE LBB0_22 // <-- // jne .LBB0_22 + ADDQ R8, SI // <-- // add rsi, r8 + SUBQ R8, DX // <-- // sub rdx, r8 + +LBB0_24: + CMPQ DI, $0x3 // <-- // cmp rdi, 3 + JB LBB0_27 // <-- // jb .LBB0_27 + XORL DI, DI // <-- // xor edi, edi + +LBB0_26: + LONG $0x04b60f44; BYTE $0x3e // MOVZX 0(SI)(DI*1), R8 // movzx r8d, byte ptr [rsi + rdi] + WORD $0x0141; BYTE $0xc0 // ADDL AX, R8 // add r8d, eax + WORD $0x0144; BYTE $0xc1 // ADDL R8, CX // add ecx, r8d + LONG $0x3e44b60f; BYTE $0x01 // MOVZX 0x1(SI)(DI*1), AX // movzx eax, byte ptr [rsi + rdi + 1] + WORD $0x0144; BYTE $0xc0 // ADDL R8, AX // add eax, r8d + WORD $0xc101 // ADDL AX, CX // add ecx, eax + LONG $0x44b60f44; WORD $0x023e // MOVZX 0x2(SI)(DI*1), R8 // movzx r8d, byte ptr [rsi + rdi + 2] + WORD $0x0141; BYTE $0xc0 // ADDL AX, R8 // add r8d, eax + WORD $0x0144; BYTE $0xc1 // ADDL R8, CX // add ecx, r8d + LONG $0x3e44b60f; BYTE $0x03 // MOVZX 0x3(SI)(DI*1), AX // movzx eax, byte ptr [rsi + rdi + 3] + WORD $0x0144; BYTE $0xc0 // ADDL R8, AX // add eax, r8d + WORD $0xc101 // ADDL AX, CX // add ecx, eax + ADDQ $0x4, DI // <-- // add rdi, 4 + CMPQ DX, DI // <-- // cmp rdx, rdi + JNE LBB0_26 // <-- // jne .LBB0_26 + +LBB0_27: + LONG $0x000f908d; WORD $0xffff // LEAL -0xfff1(AX), DX // lea edx, [rax - 65521] + CMPL AX, $0xfff1 // <-- // cmp eax, 65521 + WORD $0x420f; BYTE $0xd0 // CMOVB AX, DX // cmovb edx, eax + WORD $0xc889 // MOVL CX, AX // mov eax, ecx + LONG $0x078071be; BYTE $0x80 // MOVL $-0x7ff87f8f, SI // mov esi, 2147975281 + IMULQ AX, SI // <-- // imul rsi, rax + SHRQ $0x2f, SI // <-- // shr rsi, 47 + LONG $0xfff1c669; WORD $0x0000 // IMULL $0xfff1, SI, AX // imul eax, esi, 65521 + WORD $0xc129 // SUBL AX, CX // sub ecx, eax + WORD $0xe1c1; BYTE $0x10 // SHLL $0x10, CX // shl ecx, 16 + WORD $0xd109 // ORL DX, CX // or ecx, edx + WORD $0xc889 // MOVL CX, AX // mov eax, ecx + NOP // (skipped) // mov rsp, rbp + NOP // (skipped) // pop rbp + VZEROUPPER // <-- // vzeroupper + MOVL AX, ret+32(FP) // <-- + RET // <-- // ret + +LBB0_16: + WORD $0xe1c1; BYTE $0x10 // SHLL $0x10, CX // shl ecx, 16 + WORD $0xc809 // ORL CX, AX // or eax, ecx + NOP // (skipped) // mov rsp, rbp + NOP // (skipped) // pop rbp + VZEROUPPER // <-- // vzeroupper + MOVL AX, ret+32(FP) // <-- + RET // <-- // ret diff --git a/internal/adler32/adler32_fallback.go b/internal/adler32/adler32_fallback.go new file mode 100644 index 00000000..c213c3c1 --- /dev/null +++ b/internal/adler32/adler32_fallback.go @@ -0,0 +1,19 @@ +//go:build (!arm64 && !amd64) || purego + +package adler32 + +import ( + "hash" + "hash/adler32" +) + +// The size of an Adler-32 checksum in bytes. +const Size = 4 + +// New returns a new hash.Hash32 computing the Adler-32 checksum. +func New() hash.Hash32 { + return adler32.New() +} + +// Checksum returns the Adler-32 checksum of data. +func Checksum(data []byte) uint32 { return adler32.Checksum(data) } diff --git a/internal/adler32/adler32_generic.go b/internal/adler32/adler32_generic.go new file mode 100644 index 00000000..0908d8f7 --- /dev/null +++ b/internal/adler32/adler32_generic.go @@ -0,0 +1,44 @@ +package adler32 + +const ( + // mod is the largest prime that is less than 65536. + mod = 65521 + // nmax is the largest n such that + // 255 * n * (n+1) / 2 + (n+1) * (mod-1) <= 2^32-1. + // It is mentioned in RFC 1950 (search for "5552"). + nmax = 5552 + + // binary representation compatible with standard library. + magic = "adl\x01" + marshaledSize = len(magic) + 4 +) + +// Add p to the running checksum d. +func update(d uint32, p []byte) uint32 { + s1, s2 := d&0xffff, d>>16 + for len(p) > 0 { + var q []byte + if len(p) > nmax { + p, q = p[:nmax], p[nmax:] + } + for len(p) >= 4 { + s1 += uint32(p[0]) + s2 += s1 + s1 += uint32(p[1]) + s2 += s1 + s1 += uint32(p[2]) + s2 += s1 + s1 += uint32(p[3]) + s2 += s1 + p = p[4:] + } + for _, x := range p { + s1 += uint32(x) + s2 += s1 + } + s1 %= mod + s2 %= mod + p = q + } + return s2<<16 | s1 +} diff --git a/internal/adler32/adler32_neon.go b/internal/adler32/adler32_neon.go new file mode 100644 index 00000000..521b71e0 --- /dev/null +++ b/internal/adler32/adler32_neon.go @@ -0,0 +1,6 @@ +//go:build !purego && arm64 + +package adler32 + +//go:noescape +func adler32_neon(in uint32, buf []byte) uint32 diff --git a/internal/adler32/adler32_neon.s b/internal/adler32/adler32_neon.s new file mode 100644 index 00000000..a60c86ec --- /dev/null +++ b/internal/adler32/adler32_neon.s @@ -0,0 +1,208 @@ +//go:build !purego && arm64 + +#include "textflag.h" + +DATA mult_table<>+0x00(SB)/8, $0x001d001e001f0020 +DATA mult_table<>+0x08(SB)/8, $0x0019001a001b001c +DATA mult_table<>+0x10(SB)/8, $0x0015001600170018 +DATA mult_table<>+0x18(SB)/8, $0x0011001200130014 +DATA mult_table<>+0x20(SB)/8, $0x000d000e000f0010 +DATA mult_table<>+0x28(SB)/8, $0x0009000a000b000c +DATA mult_table<>+0x30(SB)/8, $0x0005000600070008 +DATA mult_table<>+0x38(SB)/8, $0x0001000200030004 +GLOBL mult_table<>(SB), (RODATA|NOPTR), $64 + +TEXT ·adler32_neon(SB), NOSPLIT, $0-36 + MOVW in+0(FP), R0 + MOVD buf_base+8(FP), R1 + MOVD buf_len+16(FP), R2 + MOVD buf_cap+24(FP), R3 + NOP // (skipped) // stp x29, x30, [sp, #-16]! + ANDS $15, R1, R10 // <-- // ands x10, x1, #0xf + ANDW $65535, R0, R8 // <-- // and w8, w0, #0xffff + LSRW $16, R0, R9 // <-- // lsr w9, w0, #16 + NOP // (skipped) // mov x29, sp + BEQ LBB0_4 // <-- // b.eq .LBB0_4 + ADD $1, R1, R11 // <-- // add x11, x1, #1 + MOVD R1, R12 // <-- // mov x12, x1 + +LBB0_2: + WORD $0x3840158d // MOVBU.P 1(R12), R13 // ldrb w13, [x12], #1 + SUB $1, R2, R2 // <-- // sub x2, x2, #1 + TST $15, R11 // <-- // tst x11, #0xf + ADD $1, R11, R11 // <-- // add x11, x11, #1 + ADDW R13, R8, R8 // <-- // add w8, w8, w13 + ADDW R9, R8, R9 // <-- // add w9, w8, w9 + BNE LBB0_2 // <-- // b.ne .LBB0_2 + MOVW $32881, R11 // <-- // mov w11, #32881 + MOVW $65521, R13 // <-- // mov w13, #65521 + MOVKW $(32775<<16), R11 // <-- // movk w11, #32775, lsl #16 + MOVW $4294901775, R12 // <-- // mov w12, #-65521 + MOVW $65520, R14 // <-- // mov w14, #65520 + SUB R10, R1, R10 // <-- // sub x10, x1, x10 + UMULL R11, R9, R11 // <-- // umull x11, w9, w11 + ADDW R12, R8, R12 // <-- // add w12, w8, w12 + CMPW R14, R8 // <-- // cmp w8, w14 + ADD $16, R10, R1 // <-- // add x1, x10, #16 + LSR $47, R11, R11 // <-- // lsr x11, x11, #47 + CSELW HI, R12, R8, R8 // <-- // csel w8, w12, w8, hi + MSUBW R13, R9, R11, R9 // <-- // msub w9, w11, w13, w9 + +LBB0_4: + AND $31, R2, R10 // <-- // and x10, x2, #0x1f + CMP $32, R2 // <-- // cmp x2, #32 + BCC LBB0_9 // <-- // b.lo .LBB0_9 + MOVD $mult_table<>(SB), R11 // <-- // adrp x11, mult_table + ADD $0, R11, R11 // <-- // add x11, x11, :lo12:mult_table + MOVW $32881, R14 // <-- // mov w14, #32881 + MOVW $173, R12 // <-- // mov w12, #173 + MOVD $137438953440, R13 // <-- // mov x13, #137438953440 + MOVKW $(32775<<16), R14 // <-- // movk w14, #32775, lsl #16 + VLD1 (R11), [V0.H8, V1.H8, V2.H8, V3.H8] // <-- // ld1 { v0.8h, v1.8h, v2.8h, v3.8h }, [x11] + LSR $5, R2, R11 // <-- // lsr x11, x2, #5 + MOVW $65521, R15 // <-- // mov w15, #65521 + VEXT $8, V0.B16, V0.B16, V4.B16 // <-- // ext v4.16b, v0.16b, v0.16b, #8 + VEXT $8, V1.B16, V1.B16, V5.B16 // <-- // ext v5.16b, v1.16b, v1.16b, #8 + VEXT $8, V2.B16, V2.B16, V6.B16 // <-- // ext v6.16b, v2.16b, v2.16b, #8 + VEXT $8, V3.B16, V3.B16, V7.B16 // <-- // ext v7.16b, v3.16b, v3.16b, #8 + +LBB0_6: + CMP $173, R11 // <-- // cmp x11, #173 + MOVD R1, R2 // <-- // mov x2, x1 + CSEL LO, R11, R12, R16 // <-- // csel x16, x11, x12, lo + WORD $0x6f00e414 // VMOVI $0, V20.D2 // movi v20.2d, #0000000000000000 + MULW R16, R8, R0 // <-- // mul w0, w8, w16 + ADD R16<<5, R13, R17 // <-- // add x17, x13, x16, lsl #5 + WORD $0x6f00e410 // VMOVI $0, V16.D2 // movi v16.2d, #0000000000000000 + AND $137438953440, R17, R17 // <-- // and x17, x17, #0x1fffffffe0 + WORD $0x6f00e412 // VMOVI $0, V18.D2 // movi v18.2d, #0000000000000000 + WORD $0x6f00e413 // VMOVI $0, V19.D2 // movi v19.2d, #0000000000000000 + WORD $0x6f00e415 // VMOVI $0, V21.D2 // movi v21.2d, #0000000000000000 + VMOV R0, V20.S[3] // <-- // mov v20.s[3], w0 + MOVW R16, R0 // <-- // mov w0, w16 + WORD $0x6f00e411 // VMOVI $0, V17.D2 // movi v17.2d, #0000000000000000 + +LBB0_7: + WORD $0xacc15857 // FLDPQ.P 32(R2), (F23, F22) // ldp q23, q22, [x2], #32 + SUBSW $1, R0, R0 // <-- // subs w0, w0, #1 + VADD V17.S4, V20.S4, V20.S4 // <-- // add v20.4s, v20.4s, v17.4s + WORD $0x2e3712b5 // VUADDW V23.B8, V21.H8, V21.H8 // uaddw v21.8h, v21.8h, v23.8b + WORD $0x6e371273 // VUADDW2 V23.B16, V19.H8, V19.H8 // uaddw2 v19.8h, v19.8h, v23.16b + WORD $0x6e202ad8 // VUADDLP V22.B16, V24.H8 // uaddlp v24.8h, v22.16b + WORD $0x2e361252 // VUADDW V22.B8, V18.H8, V18.H8 // uaddw v18.8h, v18.8h, v22.8b + WORD $0x6e361210 // VUADDW2 V22.B16, V16.H8, V16.H8 // uaddw2 v16.8h, v16.8h, v22.16b + WORD $0x6e206af8 // VUADALP V23.B16, V24.H8 // uadalp v24.8h, v23.16b + WORD $0x6e606b11 // VUADALP V24.H8, V17.S4 // uadalp v17.4s, v24.8h + BNE LBB0_7 // <-- // b.ne .LBB0_7 + VSHL $5, V20.S4, V20.S4 // <-- // shl v20.4s, v20.4s, #5 + ADD R17, R1, R17 // <-- // add x17, x1, x17 + SUBS R16, R11, R11 // <-- // subs x11, x11, x16 + ADD $32, R17, R1 // <-- // add x1, x17, #32 + WORD $0x2e6082b4 // VUMLAL V0.H4, V21.H4, V20.S4 // umlal v20.4s, v21.4h, v0.4h + VEXT $8, V21.B16, V21.B16, V21.B16 // <-- // ext v21.16b, v21.16b, v21.16b, #8 + WORD $0x2e6482b4 // VUMLAL V4.H4, V21.H4, V20.S4 // umlal v20.4s, v21.4h, v4.4h + VEXT $8, V19.B16, V19.B16, V21.B16 // <-- // ext v21.16b, v19.16b, v19.16b, #8 + WORD $0x2e618274 // VUMLAL V1.H4, V19.H4, V20.S4 // umlal v20.4s, v19.4h, v1.4h + VEXT $8, V18.B16, V18.B16, V19.B16 // <-- // ext v19.16b, v18.16b, v18.16b, #8 + WORD $0x2e6582b4 // VUMLAL V5.H4, V21.H4, V20.S4 // umlal v20.4s, v21.4h, v5.4h + WORD $0x2e628254 // VUMLAL V2.H4, V18.H4, V20.S4 // umlal v20.4s, v18.4h, v2.4h + WORD $0x2e668274 // VUMLAL V6.H4, V19.H4, V20.S4 // umlal v20.4s, v19.4h, v6.4h + WORD $0x2e638214 // VUMLAL V3.H4, V16.H4, V20.S4 // umlal v20.4s, v16.4h, v3.4h + VEXT $8, V16.B16, V16.B16, V16.B16 // <-- // ext v16.16b, v16.16b, v16.16b, #8 + WORD $0x2e678214 // VUMLAL V7.H4, V16.H4, V20.S4 // umlal v20.4s, v16.4h, v7.4h + WORD $0x4eb1be30 // VADDP V17.S4, V17.S4, V16.S4 // addp v16.4s, v17.4s, v17.4s + WORD $0x4eb4be91 // VADDP V20.S4, V20.S4, V17.S4 // addp v17.4s, v20.4s, v20.4s + WORD $0x0eb1be10 // VADDP V17.S2, V16.S2, V16.S2 // addp v16.2s, v16.2s, v17.2s + VMOV V16.S[1], R0 // <-- // mov w0, v16.s[1] + FMOVS F16, R2 // <-- // fmov w2, s16 + ADDW R8, R2, R8 // <-- // add w8, w2, w8 + ADDW R9, R0, R9 // <-- // add w9, w0, w9 + UMULL R14, R8, R0 // <-- // umull x0, w8, w14 + UMULL R14, R9, R2 // <-- // umull x2, w9, w14 + LSR $47, R0, R0 // <-- // lsr x0, x0, #47 + LSR $47, R2, R2 // <-- // lsr x2, x2, #47 + MSUBW R15, R8, R0, R8 // <-- // msub w8, w0, w15, w8 + MSUBW R15, R9, R2, R9 // <-- // msub w9, w2, w15, w9 + BNE LBB0_6 // <-- // b.ne .LBB0_6 + +LBB0_9: + CBZ R10, LBB0_15 // <-- // cbz x10, .LBB0_15 + CMP $16, R10 // <-- // cmp x10, #16 + BCC LBB0_13 // <-- // b.lo .LBB0_13 + WORD $0x3940002b // MOVBU (R1), R11 // ldrb w11, [x1] + SUBS $16, R10, R10 // <-- // subs x10, x10, #16 + WORD $0x3940042c // MOVBU 1(R1), R12 // ldrb w12, [x1, #1] + WORD $0x3940082d // MOVBU 2(R1), R13 // ldrb w13, [x1, #2] + ADDW R11, R8, R8 // <-- // add w8, w8, w11 + WORD $0x39400c2b // MOVBU 3(R1), R11 // ldrb w11, [x1, #3] + ADDW R9, R8, R9 // <-- // add w9, w8, w9 + ADDW R12, R8, R8 // <-- // add w8, w8, w12 + WORD $0x3940102c // MOVBU 4(R1), R12 // ldrb w12, [x1, #4] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R13, R8, R8 // <-- // add w8, w8, w13 + WORD $0x3940142d // MOVBU 5(R1), R13 // ldrb w13, [x1, #5] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R11, R8, R8 // <-- // add w8, w8, w11 + WORD $0x3940182b // MOVBU 6(R1), R11 // ldrb w11, [x1, #6] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R12, R8, R8 // <-- // add w8, w8, w12 + WORD $0x39401c2c // MOVBU 7(R1), R12 // ldrb w12, [x1, #7] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R13, R8, R8 // <-- // add w8, w8, w13 + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R11, R8, R8 // <-- // add w8, w8, w11 + WORD $0x3940202b // MOVBU 8(R1), R11 // ldrb w11, [x1, #8] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R12, R8, R8 // <-- // add w8, w8, w12 + WORD $0x3940242c // MOVBU 9(R1), R12 // ldrb w12, [x1, #9] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + WORD $0x3940382d // MOVBU 14(R1), R13 // ldrb w13, [x1, #14] + ADDW R11, R8, R8 // <-- // add w8, w8, w11 + WORD $0x3940282b // MOVBU 10(R1), R11 // ldrb w11, [x1, #10] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R12, R8, R8 // <-- // add w8, w8, w12 + WORD $0x39402c2c // MOVBU 11(R1), R12 // ldrb w12, [x1, #11] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R11, R8, R8 // <-- // add w8, w8, w11 + WORD $0x3940302b // MOVBU 12(R1), R11 // ldrb w11, [x1, #12] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R12, R8, R8 // <-- // add w8, w8, w12 + WORD $0x3940342c // MOVBU 13(R1), R12 // ldrb w12, [x1, #13] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R11, R8, R8 // <-- // add w8, w8, w11 + WORD $0x39403c2b // MOVBU 15(R1), R11 // ldrb w11, [x1, #15] + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R12, R8, R8 // <-- // add w8, w8, w12 + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R13, R8, R8 // <-- // add w8, w8, w13 + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + ADDW R11, R8, R8 // <-- // add w8, w8, w11 + ADDW R8, R9, R9 // <-- // add w9, w9, w8 + BEQ LBB0_14 // <-- // b.eq .LBB0_14 + ADD $16, R1, R1 // <-- // add x1, x1, #16 + +LBB0_13: + WORD $0x3840142b // MOVBU.P 1(R1), R11 // ldrb w11, [x1], #1 + SUBS $1, R10, R10 // <-- // subs x10, x10, #1 + ADDW R11, R8, R8 // <-- // add w8, w8, w11 + ADDW R9, R8, R9 // <-- // add w9, w8, w9 + BNE LBB0_13 // <-- // b.ne .LBB0_13 + +LBB0_14: + MOVW $32881, R10 // <-- // mov w10, #32881 + MOVW $65521, R12 // <-- // mov w12, #65521 + MOVKW $(32775<<16), R10 // <-- // movk w10, #32775, lsl #16 + MOVW $4294901775, R11 // <-- // mov w11, #-65521 + MOVW $65520, R13 // <-- // mov w13, #65520 + ADDW R11, R8, R11 // <-- // add w11, w8, w11 + UMULL R10, R9, R10 // <-- // umull x10, w9, w10 + CMPW R13, R8 // <-- // cmp w8, w13 + CSELW HI, R11, R8, R8 // <-- // csel w8, w11, w8, hi + LSR $47, R10, R10 // <-- // lsr x10, x10, #47 + MSUBW R12, R9, R10, R9 // <-- // msub w9, w10, w12, w9 + +LBB0_15: + ORRW R9<<16, R8, R0 // <-- // orr w0, w8, w9, lsl #16 + NOP // (skipped) // ldp x29, x30, [sp], #16 + MOVW R0, ret+32(FP) // <-- + RET // <-- // ret diff --git a/internal/adler32/adler32_sse3.go b/internal/adler32/adler32_sse3.go new file mode 100644 index 00000000..8e8c8a9b --- /dev/null +++ b/internal/adler32/adler32_sse3.go @@ -0,0 +1,6 @@ +//go:build !purego && amd64 + +package adler32 + +//go:noescape +func adler32_sse3(in uint32, buf []byte) uint32 diff --git a/internal/adler32/adler32_sse3.s b/internal/adler32/adler32_sse3.s new file mode 100644 index 00000000..c652de31 --- /dev/null +++ b/internal/adler32/adler32_sse3.s @@ -0,0 +1,214 @@ +//go:build !purego && amd64 + +#include "textflag.h" + +DATA LCPI0_0<>+0x00(SB)/8, $0x191a1b1c1d1e1f20 +DATA LCPI0_0<>+0x08(SB)/8, $0x1112131415161718 +GLOBL LCPI0_0<>(SB), (RODATA|NOPTR), $16 + +DATA LCPI0_1<>+0x00(SB)/8, $0x0001000100010001 +DATA LCPI0_1<>+0x08(SB)/8, $0x0001000100010001 +GLOBL LCPI0_1<>(SB), (RODATA|NOPTR), $16 + +DATA LCPI0_2<>+0x00(SB)/8, $0x090a0b0c0d0e0f10 +DATA LCPI0_2<>+0x08(SB)/8, $0x0102030405060708 +GLOBL LCPI0_2<>(SB), (RODATA|NOPTR), $16 + +TEXT ·adler32_sse3(SB), NOSPLIT, $0-36 + MOVLQZX in+0(FP), DI + MOVQ buf_base+8(FP), SI + MOVQ buf_len+16(FP), DX + MOVQ buf_cap+24(FP), CX + NOP // (skipped) // push rbp + NOP // (skipped) // mov rbp, rsp + NOP // (skipped) // and rsp, -8 + WORD $0xf889 // MOVL DI, AX // mov eax, edi + LONG $0xc8b70f44 // MOVZX AX, R9 // movzx r9d, ax + WORD $0xe8c1; BYTE $0x10 // SHRL $0x10, AX // shr eax, 16 + WORD $0xd189 // MOVL DX, CX // mov ecx, edx + WORD $0xe183; BYTE $0x1f // ANDL $0x1f, CX // and ecx, 31 + CMPQ DX, $0x20 // <-- // cmp rdx, 32 + JAE LBB0_2 // <-- // jae .LBB0_2 + WORD $0x8944; BYTE $0xcf // MOVL R9, DI // mov edi, r9d + JMP LBB0_6 // <-- // jmp .LBB0_6 + +LBB0_2: + SHRQ $0x5, DX // <-- // shr rdx, 5 + LONG $0xc0ef0f66 // PXOR X0, X0 // pxor xmm0, xmm0 + MOVO LCPI0_0<>(SB), X1 // <-- // movdqa xmm1, xmmword ptr [rip + .LCPI0_0] + MOVO LCPI0_1<>(SB), X2 // <-- // movdqa xmm2, xmmword ptr [rip + .LCPI0_1] + MOVO LCPI0_2<>(SB), X3 // <-- // movdqa xmm3, xmmword ptr [rip + .LCPI0_2] + LONG $0x8071b841; WORD $0x8007 // MOVL $-0x7ff87f8f, R8 // mov r8d, 2147975281 + +LBB0_3: + CMPQ DX, $0xad // <-- // cmp rdx, 173 + LONG $0x00adba41; WORD $0x0000 // MOVL $0xad, R10 // mov r10d, 173 + LONG $0xd2420f4c // CMOVB DX, R10 // cmovb r10, rdx + WORD $0x8944; BYTE $0xcf // MOVL R9, DI // mov edi, r9d + LONG $0xfaaf0f41 // IMULL R10, DI // imul edi, r10d + LONG $0xef6e0f66 // MOVD DI, X5 // movd xmm5, edi + LONG $0xe06e0f66 // MOVD AX, X4 // movd xmm4, eax + WORD $0x8944; BYTE $0xd0 // MOVL R10, AX // mov eax, r10d + LONG $0xf6ef0f66 // PXOR X6, X6 // pxor xmm6, xmm6 + +LBB0_4: + LONG $0x3e6f0ff3 // MOVDQU 0(SI), X7 // movdqu xmm7, xmmword ptr [rsi] + LONG $0x6f0f4466; BYTE $0xc7 // MOVDQA X7, X8 // movdqa xmm8, xmm7 + LONG $0x04380f66; BYTE $0xf9 // PMADDUBSW X1, X7 // pmaddubsw xmm7, xmm1 + LONG $0xfaf50f66 // PMADDWD X2, X7 // pmaddwd xmm7, xmm2 + LONG $0xfcfe0f66 // PADDD X4, X7 // paddd xmm7, xmm4 + LONG $0x666f0ff3; BYTE $0x10 // MOVDQU 0x10(SI), X4 // movdqu xmm4, xmmword ptr [rsi + 16] + LONG $0xeefe0f66 // PADDD X6, X5 // paddd xmm5, xmm6 + LONG $0xf60f4466; BYTE $0xc0 // PSADBW X0, X8 // psadbw xmm8, xmm0 + LONG $0xfe0f4466; BYTE $0xc6 // PADDD X6, X8 // paddd xmm8, xmm6 + LONG $0xf46f0f66 // MOVDQA X4, X6 // movdqa xmm6, xmm4 + LONG $0xf0f60f66 // PSADBW X0, X6 // psadbw xmm6, xmm0 + LONG $0xfe0f4166; BYTE $0xf0 // PADDD X8, X6 // paddd xmm6, xmm8 + LONG $0x04380f66; BYTE $0xe3 // PMADDUBSW X3, X4 // pmaddubsw xmm4, xmm3 + LONG $0xe2f50f66 // PMADDWD X2, X4 // pmaddwd xmm4, xmm2 + LONG $0xe7fe0f66 // PADDD X7, X4 // paddd xmm4, xmm7 + ADDQ $0x20, SI // <-- // add rsi, 32 + WORD $0xc8ff // DECL AX // dec eax + JNE LBB0_4 // <-- // jne .LBB0_4 + LONG $0xf5720f66; BYTE $0x05 // PSLLD $0x5, X5 // pslld xmm5, 5 + LONG $0xe5fe0f66 // PADDD X5, X4 // paddd xmm4, xmm5 + LONG $0xee700f66; BYTE $0xb1 // PSHUFD $0xb1, X6, X5 // pshufd xmm5, xmm6, 177 + LONG $0xeefe0f66 // PADDD X6, X5 // paddd xmm5, xmm6 + LONG $0xf5700f66; BYTE $0xee // PSHUFD $0xee, X5, X6 // pshufd xmm6, xmm5, 238 + LONG $0xf5fe0f66 // PADDD X5, X6 // paddd xmm6, xmm5 + LONG $0xf77e0f66 // MOVD X6, DI // movd edi, xmm6 + WORD $0x0144; BYTE $0xcf // ADDL R9, DI // add edi, r9d + LONG $0xec700f66; BYTE $0xb1 // PSHUFD $0xb1, X4, X5 // pshufd xmm5, xmm4, 177 + LONG $0xecfe0f66 // PADDD X4, X5 // paddd xmm5, xmm4 + LONG $0xe5700f66; BYTE $0xee // PSHUFD $0xee, X5, X4 // pshufd xmm4, xmm5, 238 + LONG $0xe5fe0f66 // PADDD X5, X4 // paddd xmm4, xmm5 + LONG $0xe07e0f66 // MOVD X4, AX // movd eax, xmm4 + MOVQ DI, R9 // <-- // mov r9, rdi + IMULQ R8, R9 // <-- // imul r9, r8 + SHRQ $0x2f, R9 // <-- // shr r9, 47 + LONG $0xf1c96945; WORD $0x00ff; BYTE $0x00 // IMULL $0xfff1, R9, R9 // imul r9d, r9d, 65521 + WORD $0x2944; BYTE $0xcf // SUBL R9, DI // sub edi, r9d + MOVQ AX, R9 // <-- // mov r9, rax + IMULQ R8, R9 // <-- // imul r9, r8 + SHRQ $0x2f, R9 // <-- // shr r9, 47 + LONG $0xf1c96945; WORD $0x00ff; BYTE $0x00 // IMULL $0xfff1, R9, R9 // imul r9d, r9d, 65521 + WORD $0x2944; BYTE $0xc8 // SUBL R9, AX // sub eax, r9d + WORD $0x8941; BYTE $0xf9 // MOVL DI, R9 // mov r9d, edi + SUBQ R10, DX // <-- // sub rdx, r10 + JNE LBB0_3 // <-- // jne .LBB0_3 + +LBB0_6: + WORD $0x8548; BYTE $0xc9 // TESTQ CX, CX // test rcx, rcx + JE LBB0_18 // <-- // je .LBB0_18 + CMPL CX, $0x10 // <-- // cmp ecx, 16 + JB LBB0_10 // <-- // jb .LBB0_10 + WORD $0xb60f; BYTE $0x16 // MOVZX 0(SI), DX // movzx edx, byte ptr [rsi] + WORD $0xd701 // ADDL DX, DI // add edi, edx + WORD $0xf801 // ADDL DI, AX // add eax, edi + LONG $0x0156b60f // MOVZX 0x1(SI), DX // movzx edx, byte ptr [rsi + 1] + WORD $0xfa01 // ADDL DI, DX // add edx, edi + WORD $0xd001 // ADDL DX, AX // add eax, edx + LONG $0x027eb60f // MOVZX 0x2(SI), DI // movzx edi, byte ptr [rsi + 2] + WORD $0xd701 // ADDL DX, DI // add edi, edx + WORD $0xf801 // ADDL DI, AX // add eax, edi + LONG $0x0356b60f // MOVZX 0x3(SI), DX // movzx edx, byte ptr [rsi + 3] + WORD $0xfa01 // ADDL DI, DX // add edx, edi + WORD $0xd001 // ADDL DX, AX // add eax, edx + LONG $0x047eb60f // MOVZX 0x4(SI), DI // movzx edi, byte ptr [rsi + 4] + WORD $0xd701 // ADDL DX, DI // add edi, edx + WORD $0xf801 // ADDL DI, AX // add eax, edi + LONG $0x0556b60f // MOVZX 0x5(SI), DX // movzx edx, byte ptr [rsi + 5] + WORD $0xfa01 // ADDL DI, DX // add edx, edi + WORD $0xd001 // ADDL DX, AX // add eax, edx + LONG $0x067eb60f // MOVZX 0x6(SI), DI // movzx edi, byte ptr [rsi + 6] + WORD $0xd701 // ADDL DX, DI // add edi, edx + WORD $0xf801 // ADDL DI, AX // add eax, edi + LONG $0x0756b60f // MOVZX 0x7(SI), DX // movzx edx, byte ptr [rsi + 7] + WORD $0xfa01 // ADDL DI, DX // add edx, edi + WORD $0xd001 // ADDL DX, AX // add eax, edx + LONG $0x087eb60f // MOVZX 0x8(SI), DI // movzx edi, byte ptr [rsi + 8] + WORD $0xd701 // ADDL DX, DI // add edi, edx + WORD $0xf801 // ADDL DI, AX // add eax, edi + LONG $0x0956b60f // MOVZX 0x9(SI), DX // movzx edx, byte ptr [rsi + 9] + WORD $0xfa01 // ADDL DI, DX // add edx, edi + WORD $0xd001 // ADDL DX, AX // add eax, edx + LONG $0x0a7eb60f // MOVZX 0xa(SI), DI // movzx edi, byte ptr [rsi + 10] + WORD $0xd701 // ADDL DX, DI // add edi, edx + WORD $0xf801 // ADDL DI, AX // add eax, edi + LONG $0x0b56b60f // MOVZX 0xb(SI), DX // movzx edx, byte ptr [rsi + 11] + WORD $0xfa01 // ADDL DI, DX // add edx, edi + WORD $0xd001 // ADDL DX, AX // add eax, edx + LONG $0x0c7eb60f // MOVZX 0xc(SI), DI // movzx edi, byte ptr [rsi + 12] + WORD $0xd701 // ADDL DX, DI // add edi, edx + WORD $0xf801 // ADDL DI, AX // add eax, edi + LONG $0x0d56b60f // MOVZX 0xd(SI), DX // movzx edx, byte ptr [rsi + 13] + WORD $0xfa01 // ADDL DI, DX // add edx, edi + WORD $0xd001 // ADDL DX, AX // add eax, edx + LONG $0x46b60f44; BYTE $0x0e // MOVZX 0xe(SI), R8 // movzx r8d, byte ptr [rsi + 14] + WORD $0x0141; BYTE $0xd0 // ADDL DX, R8 // add r8d, edx + WORD $0x0144; BYTE $0xc0 // ADDL R8, AX // add eax, r8d + LONG $0x0f7eb60f // MOVZX 0xf(SI), DI // movzx edi, byte ptr [rsi + 15] + WORD $0x0144; BYTE $0xc7 // ADDL R8, DI // add edi, r8d + WORD $0xf801 // ADDL DI, AX // add eax, edi + ADDQ $-0x10, CX // <-- // add rcx, -16 + JE LBB0_17 // <-- // je .LBB0_17 + ADDQ $0x10, SI // <-- // add rsi, 16 + +LBB0_10: + LEAQ -0x1(CX), DX // <-- // lea rdx, [rcx - 1] + MOVQ CX, R9 // <-- // mov r9, rcx + ANDQ $0x3, R9 // <-- // and r9, 3 + JE LBB0_14 // <-- // je .LBB0_14 + XORL R8, R8 // <-- // xor r8d, r8d + +LBB0_12: + LONG $0x14b60f46; BYTE $0x06 // MOVZX 0(SI)(R8*1), R10 // movzx r10d, byte ptr [rsi + r8] + WORD $0x0144; BYTE $0xd7 // ADDL R10, DI // add edi, r10d + WORD $0xf801 // ADDL DI, AX // add eax, edi + INCQ R8 // <-- // inc r8 + CMPQ R9, R8 // <-- // cmp r9, r8 + JNE LBB0_12 // <-- // jne .LBB0_12 + ADDQ R8, SI // <-- // add rsi, r8 + SUBQ R8, CX // <-- // sub rcx, r8 + +LBB0_14: + CMPQ DX, $0x3 // <-- // cmp rdx, 3 + JB LBB0_17 // <-- // jb .LBB0_17 + XORL DX, DX // <-- // xor edx, edx + +LBB0_16: + LONG $0x04b60f44; BYTE $0x16 // MOVZX 0(SI)(DX*1), R8 // movzx r8d, byte ptr [rsi + rdx] + WORD $0x0141; BYTE $0xf8 // ADDL DI, R8 // add r8d, edi + WORD $0x0144; BYTE $0xc0 // ADDL R8, AX // add eax, r8d + LONG $0x167cb60f; BYTE $0x01 // MOVZX 0x1(SI)(DX*1), DI // movzx edi, byte ptr [rsi + rdx + 1] + WORD $0x0144; BYTE $0xc7 // ADDL R8, DI // add edi, r8d + WORD $0xf801 // ADDL DI, AX // add eax, edi + LONG $0x44b60f44; WORD $0x0216 // MOVZX 0x2(SI)(DX*1), R8 // movzx r8d, byte ptr [rsi + rdx + 2] + WORD $0x0141; BYTE $0xf8 // ADDL DI, R8 // add r8d, edi + WORD $0x0144; BYTE $0xc0 // ADDL R8, AX // add eax, r8d + LONG $0x167cb60f; BYTE $0x03 // MOVZX 0x3(SI)(DX*1), DI // movzx edi, byte ptr [rsi + rdx + 3] + WORD $0x0144; BYTE $0xc7 // ADDL R8, DI // add edi, r8d + WORD $0xf801 // ADDL DI, AX // add eax, edi + ADDQ $0x4, DX // <-- // add rdx, 4 + CMPQ CX, DX // <-- // cmp rcx, rdx + JNE LBB0_16 // <-- // jne .LBB0_16 + +LBB0_17: + LONG $0x000f8f8d; WORD $0xffff // LEAL -0xfff1(DI), CX // lea ecx, [rdi - 65521] + CMPL DI, $0xfff1 // <-- // cmp edi, 65521 + WORD $0x420f; BYTE $0xcf // CMOVB DI, CX // cmovb ecx, edi + WORD $0xc289 // MOVL AX, DX // mov edx, eax + LONG $0x078071be; BYTE $0x80 // MOVL $-0x7ff87f8f, SI // mov esi, 2147975281 + IMULQ DX, SI // <-- // imul rsi, rdx + SHRQ $0x2f, SI // <-- // shr rsi, 47 + LONG $0xfff1d669; WORD $0x0000 // IMULL $0xfff1, SI, DX // imul edx, esi, 65521 + WORD $0xd029 // SUBL DX, AX // sub eax, edx + WORD $0xcf89 // MOVL CX, DI // mov edi, ecx + +LBB0_18: + WORD $0xe0c1; BYTE $0x10 // SHLL $0x10, AX // shl eax, 16 + WORD $0xf809 // ORL DI, AX // or eax, edi + NOP // (skipped) // mov rsp, rbp + NOP // (skipped) // pop rbp + MOVL AX, ret+32(FP) // <-- + RET // <-- // ret diff --git a/internal/adler32/bench_test.go b/internal/adler32/bench_test.go new file mode 100644 index 00000000..7744b903 --- /dev/null +++ b/internal/adler32/bench_test.go @@ -0,0 +1,22 @@ +package adler32 + +import ( + "testing" +) + +const benchmarkSize = 64 * 1024 + +var data = make([]byte, benchmarkSize) + +func init() { + for i := range benchmarkSize { + data[i] = byte(i % 256) + } +} + +func BenchmarkChecksum(b *testing.B) { + b.ReportAllocs() + for range b.N { + Checksum(data) + } +} |