golang 对象(struct) hash原理

map里面的key的hash是怎么实现的

src/runtime/map.go

type maptype struct {
	typ        _type
	key        *_type
	elem       *_type
	bucket     *_type // internal type representing a hash bucket
	keysize    uint8  // size of key slot
	elemsize   uint8  // size of elem slot
	bucketsize uint16 // size of bucket
	flags      uint32
}

//t *maptype
alg := t.key.alg
hash := alg.hash(key, uintptr(h.hash0))

这里其实就是拿到key对应的类型，然后获取当前key的类型的hash算法。然后调用hash函数。

hash函数的定义在这里：

// typeAlg is also copied/used in reflect/type.go.
// keep them in sync.
type typeAlg struct {
	// function for hashing objects of this type
	// (ptr to object, seed) -> hash
	hash func(unsafe.Pointer, uintptr) uintptr
	// function for comparing objects of this type
	// (ptr to object A, ptr to object B) -> ==?
	equal func(unsafe.Pointer, unsafe.Pointer) bool
}

可以知道，入参是指向key的指针，第二个参数是hash种子。

golang里面的对象的hash原理

src/runtime/alg.go

var algarray = [alg_max]typeAlg{
	alg_NOEQ:     {nil, nil},
	alg_MEM0:     {memhash0, memequal0},
	alg_MEM8:     {memhash8, memequal8},
	alg_MEM16:    {memhash16, memequal16},
	alg_MEM32:    {memhash32, memequal32},
	alg_MEM64:    {memhash64, memequal64},
	alg_MEM128:   {memhash128, memequal128},
	alg_STRING:   {strhash, strequal},
	alg_INTER:    {interhash, interequal},
	alg_NILINTER: {nilinterhash, nilinterequal},
	alg_FLOAT32:  {f32hash, f32equal},
	alg_FLOAT64:  {f64hash, f64equal},
	alg_CPLX64:   {c64hash, c64equal},
	alg_CPLX128:  {c128hash, c128equal},
}

func memhash(p unsafe.Pointer, seed, s uintptr) uintptrstrhash

func strhash(a unsafe.Pointer, h uintptr) uintptr {
	x := (*stringStruct)(a)
	return memhash(x.str, h, uintptr(x.len))
}

memhashsrc/runtime/hash64.go

// p表示需要hash的对象的地址
// seed 是hash 种子，
// s是需要hash的对象的字节数
func memhash(p unsafe.Pointer, seed, s uintptr) uintptr {
	if (GOARCH == "amd64" || GOARCH == "arm64") &&
		GOOS != "nacl" && useAeshash {
		return aeshash(p, seed, s)
	}
	h := uint64(seed + s*hashkey[0])
tail:
	switch {
	case s == 0:
	case s < 4:
		h ^= uint64(*(*byte)(p))
		h ^= uint64(*(*byte)(add(p, s>>1))) << 8
		h ^= uint64(*(*byte)(add(p, s-1))) << 16
		h = rotl_31(h*m1) * m2
	case s <= 8:
		h ^= uint64(readUnaligned32(p))
		h ^= uint64(readUnaligned32(add(p, s-4))) << 32
		h = rotl_31(h*m1) * m2
	case s <= 16:
		h ^= readUnaligned64(p)
		h = rotl_31(h*m1) * m2
		h ^= readUnaligned64(add(p, s-8))
		h = rotl_31(h*m1) * m2
	case s <= 32:
		h ^= readUnaligned64(p)
		h = rotl_31(h*m1) * m2
		h ^= readUnaligned64(add(p, 8))
		h = rotl_31(h*m1) * m2
		h ^= readUnaligned64(add(p, s-16))
		h = rotl_31(h*m1) * m2
		h ^= readUnaligned64(add(p, s-8))
		h = rotl_31(h*m1) * m2
	default:
		v1 := h
		v2 := uint64(seed * hashkey[1])
		v3 := uint64(seed * hashkey[2])
		v4 := uint64(seed * hashkey[3])
		for s >= 32 {
			v1 ^= readUnaligned64(p)
			v1 = rotl_31(v1*m1) * m2
			p = add(p, 8)
			v2 ^= readUnaligned64(p)
			v2 = rotl_31(v2*m2) * m3
			p = add(p, 8)
			v3 ^= readUnaligned64(p)
			v3 = rotl_31(v3*m3) * m4
			p = add(p, 8)
			v4 ^= readUnaligned64(p)
			v4 = rotl_31(v4*m4) * m1
			p = add(p, 8)
			s -= 32
		}
		h = v1 ^ v2 ^ v3 ^ v4
		goto tail
	}

	h ^= h >> 29
	h *= m3
	h ^= h >> 32
	return uintptr(h)
}

aeshash

src/runtime/asm_amd64.sruntime·aeshash

Hash 测试

下面以string作为key，来测试以下三种算法的性能：

1. fnv算法
2. memhash算法
3. aeshash的汇编算法

fnv

测试背景：10000个uuid string;

const LEN = 10000
// the length of element is 36
var keys [LEN][]byte

func init(){
	for i:=0; i<LEN; i++ {
		k := uuid.New().String()
		keys[i] = []byte(k)
	}
}

// 777000ns
func main(){
	start := time.Now().UnixNano()
	for _, k := range keys {
		h := fnv.New64()
		h.Write(k)
		h.Sum64()
	}
	end := time.Now().UnixNano()
	fmt.Println("total time:", end-start, "ns")
}

执行结果：10000次hash大约是727000ns，也就是平均每次hash要花费72ns。

memhash

这个是调用golang runtime里面的内嵌hash代码，这部分是我从runtime里面copy出来的。

代码如下：

package main

const LEN = 10000
// the length of element is 36
var keys [LEN]string

func init(){
	for i:=0; i<LEN; i++ {
		keys[i] = uuid.New().String()
	}
}
// 178000ns
func main()  {
	start := time.Now().UnixNano()
	for _, k := range keys {
		memhash(unsafe.Pointer(&k), 1, 36)
	}
	end := time.Now().UnixNano()
	fmt.Println("total time:", end-start, "ns")
}

// used in hash{32,64}.go to seed the hash function
var hashkey [4]uintptr
const PtrSize = 4 << (^uintptr(0) >> 63)
func init()  {
	for i:=0; i<4; i++ {
		hashkey[i]= uintptr(rand.Int63())
	}
	hashkey[0] |= 1 // make sure these numbers are odd
	hashkey[1] |= 1
	hashkey[2] |= 1
	hashkey[3] |= 1
}
func add(p unsafe.Pointer, x uintptr) unsafe.Pointer {
	return unsafe.Pointer(uintptr(p) + x)
}
func rotl_31(x uint64) uint64 {
	return (x << 31) | (x >> (64 - 31))
}

const (
	// Constants for multiplication: four random odd 64-bit numbers.
	m1 = 16877499708836156737
	m2 = 2820277070424839065
	m3 = 9497967016996688599
	m4 = 15839092249703872147
)

const (
	BigEndian           = false
	DefaultPhysPageSize = 4096
	PCQuantum           = 1
	Int64Align          = 8
	MinFrameSize        = 0
)

func readUnaligned64(p unsafe.Pointer) uint64 {
	q := (*[8]byte)(p)
	if BigEndian {
		return uint64(q[7]) | uint64(q[6])<<8 | uint64(q[5])<<16 | uint64(q[4])<<24 |
			uint64(q[3])<<32 | uint64(q[2])<<40 | uint64(q[1])<<48 | uint64(q[0])<<56
	}
	return uint64(q[0]) | uint64(q[1])<<8 | uint64(q[2])<<16 | uint64(q[3])<<24 | uint64(q[4])<<32 | uint64(q[5])<<40 | uint64(q[6])<<48 | uint64(q[7])<<56
}

// Note: These routines perform the read with an native endianness.
func readUnaligned32(p unsafe.Pointer) uint32 {
	q := (*[4]byte)(p)
	if BigEndian {
		return uint32(q[3]) | uint32(q[2])<<8 | uint32(q[1])<<16 | uint32(q[0])<<24
	}
	return uint32(q[0]) | uint32(q[1])<<8 | uint32(q[2])<<16 | uint32(q[3])<<24
}

func memhash(p unsafe.Pointer, seed, s uintptr) uintptr {
	h := uint64(seed + s*hashkey[0])
tail:
	switch {
	case s == 0:
	case s < 4:
		h ^= uint64(*(*byte)(p))
		h ^= uint64(*(*byte)(add(p, s>>1))) << 8
		h ^= uint64(*(*byte)(add(p, s-1))) << 16
		h = rotl_31(h*m1) * m2
	case s <= 8:
		h ^= uint64(readUnaligned32(p))
		h ^= uint64(readUnaligned32(add(p, s-4))) << 32
		h = rotl_31(h*m1) * m2
	case s <= 16:
		h ^= readUnaligned64(p)
		h = rotl_31(h*m1) * m2
		h ^= readUnaligned64(add(p, s-8))
		h = rotl_31(h*m1) * m2
	case s <= 32:
		h ^= readUnaligned64(p)
		h = rotl_31(h*m1) * m2
		h ^= readUnaligned64(add(p, 8))
		h = rotl_31(h*m1) * m2
		h ^= readUnaligned64(add(p, s-16))
		h = rotl_31(h*m1) * m2
		h ^= readUnaligned64(add(p, s-8))
		h = rotl_31(h*m1) * m2
	default:
		v1 := h
		v2 := uint64(seed * hashkey[1])
		v3 := uint64(seed * hashkey[2])
		v4 := uint64(seed * hashkey[3])
		for s >= 32 {
			v1 ^= readUnaligned64(p)
			v1 = rotl_31(v1*m1) * m2
			p = add(p, 8)
			v2 ^= readUnaligned64(p)
			v2 = rotl_31(v2*m2) * m3
			p = add(p, 8)
			v3 ^= readUnaligned64(p)
			v3 = rotl_31(v3*m3) * m4
			p = add(p, 8)
			v4 ^= readUnaligned64(p)
			v4 = rotl_31(v4*m4) * m1
			p = add(p, 8)
			s -= 32
		}
		h = v1 ^ v2 ^ v3 ^ v4
		goto tail
	}

	h ^= h >> 29
	h *= m3
	h ^= h >> 32
	return uintptr(h)
}

执行结果：10000次hash大约是178000ns，也就是平均每次hash要花费17.8ns。

aeshash汇编

这部分代码也是从runtime asm代码copy出来做测试：
注意，需要把src/runtime/go_tls.h、funcdata.h、textflag.h 这个三个头文件放在当前目录下。

// Copyright 2009 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

#include "go_asm.h"
#include "go_tls.h"
#include "funcdata.h"
#include "textflag.h"

//louyuting
// func aeshashstr(p unsafe.Pointer, h uintptr) uintptr
TEXT ·aeshashstr2(SB),NOSPLIT,$0-24
	MOVQ	p+0(FP), AX	// ptr to string struct
	MOVQ	8(AX), CX	// length of string
	MOVQ	(AX), AX	// string data
	LEAQ	ret+16(FP), DX
	JMP	aeshashbody<>(SB)

// AX: data
// CX: length
// DX: address to put return value
TEXT aeshashbody<>(SB),NOSPLIT,$0-0
	// Fill an SSE register with our seeds.
	MOVQ	h+8(FP), X0			// 64 bits of per-table hash seed
	PINSRW	$4, CX, X0			// 16 bits of length
	PSHUFHW $0, X0, X0			// repeat length 4 times total
	MOVO	X0, X1				// save unscrambled seed
	PXOR	main·aeskeysched(SB), X0	// xor in per-process seed
	AESENC	X0, X0				// scramble seed

	CMPQ	CX, $16
	JB	aes0to15
	JE	aes16
	CMPQ	CX, $32
	JBE	aes17to32
	CMPQ	CX, $64
	JBE	aes33to64
	CMPQ	CX, $128
	JBE	aes65to128
	JMP	aes129plus

aes0to15:
	TESTQ	CX, CX
	JE	aes0

	ADDQ	$16, AX
	TESTW	$0xff0, AX
	JE	endofpage

	// 16 bytes loaded at this address won't cross
	// a page boundary, so we can load it directly.
	MOVOU	-16(AX), X1
	ADDQ	CX, CX
	MOVQ	$masks<>(SB), AX
	PAND	(AX)(CX*8), X1
final1:
	PXOR	X0, X1	// xor data with seed
	AESENC	X1, X1	// scramble combo 3 times
	AESENC	X1, X1
	AESENC	X1, X1
	MOVQ	X1, (DX)
	RET

endofpage:
	// address ends in 1111xxxx. Might be up against
	// a page boundary, so load ending at last byte.
	// Then shift bytes down using pshufb.
	MOVOU	-32(AX)(CX*1), X1
	ADDQ	CX, CX
	MOVQ	$shifts<>(SB), AX
	PSHUFB	(AX)(CX*8), X1
	JMP	final1

aes0:
	// Return scrambled input seed
	AESENC	X0, X0
	MOVQ	X0, (DX)
	RET

aes16:
	MOVOU	(AX), X1
	JMP	final1

aes17to32:
	// make second starting seed
	PXOR	main·aeskeysched+16(SB), X1
	AESENC	X1, X1

	// load data to be hashed
	MOVOU	(AX), X2
	MOVOU	-16(AX)(CX*1), X3

	// xor with seed
	PXOR	X0, X2
	PXOR	X1, X3

	// scramble 3 times
	AESENC	X2, X2
	AESENC	X3, X3
	AESENC	X2, X2
	AESENC	X3, X3
	AESENC	X2, X2
	AESENC	X3, X3

	// combine results
	PXOR	X3, X2
	MOVQ	X2, (DX)
	RET

aes33to64:
	// make 3 more starting seeds
	MOVO	X1, X2
	MOVO	X1, X3
	PXOR	main·aeskeysched+16(SB), X1
	PXOR	main·aeskeysched+32(SB), X2
	PXOR	main·aeskeysched+48(SB), X3
	AESENC	X1, X1
	AESENC	X2, X2
	AESENC	X3, X3

	MOVOU	(AX), X4
	MOVOU	16(AX), X5
	MOVOU	-32(AX)(CX*1), X6
	MOVOU	-16(AX)(CX*1), X7

	PXOR	X0, X4
	PXOR	X1, X5
	PXOR	X2, X6
	PXOR	X3, X7

	AESENC	X4, X4
	AESENC	X5, X5
	AESENC	X6, X6
	AESENC	X7, X7

	AESENC	X4, X4
	AESENC	X5, X5
	AESENC	X6, X6
	AESENC	X7, X7

	AESENC	X4, X4
	AESENC	X5, X5
	AESENC	X6, X6
	AESENC	X7, X7

	PXOR	X6, X4
	PXOR	X7, X5
	PXOR	X5, X4
	MOVQ	X4, (DX)
	RET

aes65to128:
	// make 7 more starting seeds
	MOVO	X1, X2
	MOVO	X1, X3
	MOVO	X1, X4
	MOVO	X1, X5
	MOVO	X1, X6
	MOVO	X1, X7
	PXOR	main·aeskeysched+16(SB), X1
	PXOR	main·aeskeysched+32(SB), X2
	PXOR	main·aeskeysched+48(SB), X3
	PXOR	main·aeskeysched+64(SB), X4
	PXOR	main·aeskeysched+80(SB), X5
	PXOR	main·aeskeysched+96(SB), X6
	PXOR	main·aeskeysched+112(SB), X7
	AESENC	X1, X1
	AESENC	X2, X2
	AESENC	X3, X3
	AESENC	X4, X4
	AESENC	X5, X5
	AESENC	X6, X6
	AESENC	X7, X7

	// load data
	MOVOU	(AX), X8
	MOVOU	16(AX), X9
	MOVOU	32(AX), X10
	MOVOU	48(AX), X11
	MOVOU	-64(AX)(CX*1), X12
	MOVOU	-48(AX)(CX*1), X13
	MOVOU	-32(AX)(CX*1), X14
	MOVOU	-16(AX)(CX*1), X15

	// xor with seed
	PXOR	X0, X8
	PXOR	X1, X9
	PXOR	X2, X10
	PXOR	X3, X11
	PXOR	X4, X12
	PXOR	X5, X13
	PXOR	X6, X14
	PXOR	X7, X15

	// scramble 3 times
	AESENC	X8, X8
	AESENC	X9, X9
	AESENC	X10, X10
	AESENC	X11, X11
	AESENC	X12, X12
	AESENC	X13, X13
	AESENC	X14, X14
	AESENC	X15, X15

	AESENC	X8, X8
	AESENC	X9, X9
	AESENC	X10, X10
	AESENC	X11, X11
	AESENC	X12, X12
	AESENC	X13, X13
	AESENC	X14, X14
	AESENC	X15, X15

	AESENC	X8, X8
	AESENC	X9, X9
	AESENC	X10, X10
	AESENC	X11, X11
	AESENC	X12, X12
	AESENC	X13, X13
	AESENC	X14, X14
	AESENC	X15, X15

	// combine results
	PXOR	X12, X8
	PXOR	X13, X9
	PXOR	X14, X10
	PXOR	X15, X11
	PXOR	X10, X8
	PXOR	X11, X9
	PXOR	X9, X8
	MOVQ	X8, (DX)
	RET

aes129plus:
	// make 7 more starting seeds
	MOVO	X1, X2
	MOVO	X1, X3
	MOVO	X1, X4
	MOVO	X1, X5
	MOVO	X1, X6
	MOVO	X1, X7
	PXOR	main·aeskeysched+16(SB), X1
	PXOR	main·aeskeysched+32(SB), X2
	PXOR	main·aeskeysched+48(SB), X3
	PXOR	main·aeskeysched+64(SB), X4
	PXOR	main·aeskeysched+80(SB), X5
	PXOR	main·aeskeysched+96(SB), X6
	PXOR	main·aeskeysched+112(SB), X7
	AESENC	X1, X1
	AESENC	X2, X2
	AESENC	X3, X3
	AESENC	X4, X4
	AESENC	X5, X5
	AESENC	X6, X6
	AESENC	X7, X7

	// start with last (possibly overlapping) block
	MOVOU	-128(AX)(CX*1), X8
	MOVOU	-112(AX)(CX*1), X9
	MOVOU	-96(AX)(CX*1), X10
	MOVOU	-80(AX)(CX*1), X11
	MOVOU	-64(AX)(CX*1), X12
	MOVOU	-48(AX)(CX*1), X13
	MOVOU	-32(AX)(CX*1), X14
	MOVOU	-16(AX)(CX*1), X15

	// xor in seed
	PXOR	X0, X8
	PXOR	X1, X9
	PXOR	X2, X10
	PXOR	X3, X11
	PXOR	X4, X12
	PXOR	X5, X13
	PXOR	X6, X14
	PXOR	X7, X15

	// compute number of remaining 128-byte blocks
	DECQ	CX
	SHRQ	$7, CX

aesloop:
	// scramble state
	AESENC	X8, X8
	AESENC	X9, X9
	AESENC	X10, X10
	AESENC	X11, X11
	AESENC	X12, X12
	AESENC	X13, X13
	AESENC	X14, X14
	AESENC	X15, X15

	// scramble state, xor in a block
	MOVOU	(AX), X0
	MOVOU	16(AX), X1
	MOVOU	32(AX), X2
	MOVOU	48(AX), X3
	AESENC	X0, X8
	AESENC	X1, X9
	AESENC	X2, X10
	AESENC	X3, X11
	MOVOU	64(AX), X4
	MOVOU	80(AX), X5
	MOVOU	96(AX), X6
	MOVOU	112(AX), X7
	AESENC	X4, X12
	AESENC	X5, X13
	AESENC	X6, X14
	AESENC	X7, X15

	ADDQ	$128, AX
	DECQ	CX
	JNE	aesloop

	// 3 more scrambles to finish
	AESENC	X8, X8
	AESENC	X9, X9
	AESENC	X10, X10
	AESENC	X11, X11
	AESENC	X12, X12
	AESENC	X13, X13
	AESENC	X14, X14
	AESENC	X15, X15
	AESENC	X8, X8
	AESENC	X9, X9
	AESENC	X10, X10
	AESENC	X11, X11
	AESENC	X12, X12
	AESENC	X13, X13
	AESENC	X14, X14
	AESENC	X15, X15
	AESENC	X8, X8
	AESENC	X9, X9
	AESENC	X10, X10
	AESENC	X11, X11
	AESENC	X12, X12
	AESENC	X13, X13
	AESENC	X14, X14
	AESENC	X15, X15

	PXOR	X12, X8
	PXOR	X13, X9
	PXOR	X14, X10
	PXOR	X15, X11
	PXOR	X10, X8
	PXOR	X11, X9
	PXOR	X9, X8
	MOVQ	X8, (DX)
	RET


// simple mask to get rid of data in the high part of the register.
DATA masks<>+0x00(SB)/8, $0x0000000000000000
DATA masks<>+0x08(SB)/8, $0x0000000000000000
DATA masks<>+0x10(SB)/8, $0x00000000000000ff
DATA masks<>+0x18(SB)/8, $0x0000000000000000
DATA masks<>+0x20(SB)/8, $0x000000000000ffff
DATA masks<>+0x28(SB)/8, $0x0000000000000000
DATA masks<>+0x30(SB)/8, $0x0000000000ffffff
DATA masks<>+0x38(SB)/8, $0x0000000000000000
DATA masks<>+0x40(SB)/8, $0x00000000ffffffff
DATA masks<>+0x48(SB)/8, $0x0000000000000000
DATA masks<>+0x50(SB)/8, $0x000000ffffffffff
DATA masks<>+0x58(SB)/8, $0x0000000000000000
DATA masks<>+0x60(SB)/8, $0x0000ffffffffffff
DATA masks<>+0x68(SB)/8, $0x0000000000000000
DATA masks<>+0x70(SB)/8, $0x00ffffffffffffff
DATA masks<>+0x78(SB)/8, $0x0000000000000000
DATA masks<>+0x80(SB)/8, $0xffffffffffffffff
DATA masks<>+0x88(SB)/8, $0x0000000000000000
DATA masks<>+0x90(SB)/8, $0xffffffffffffffff
DATA masks<>+0x98(SB)/8, $0x00000000000000ff
DATA masks<>+0xa0(SB)/8, $0xffffffffffffffff
DATA masks<>+0xa8(SB)/8, $0x000000000000ffff
DATA masks<>+0xb0(SB)/8, $0xffffffffffffffff
DATA masks<>+0xb8(SB)/8, $0x0000000000ffffff
DATA masks<>+0xc0(SB)/8, $0xffffffffffffffff
DATA masks<>+0xc8(SB)/8, $0x00000000ffffffff
DATA masks<>+0xd0(SB)/8, $0xffffffffffffffff
DATA masks<>+0xd8(SB)/8, $0x000000ffffffffff
DATA masks<>+0xe0(SB)/8, $0xffffffffffffffff
DATA masks<>+0xe8(SB)/8, $0x0000ffffffffffff
DATA masks<>+0xf0(SB)/8, $0xffffffffffffffff
DATA masks<>+0xf8(SB)/8, $0x00ffffffffffffff
GLOBL masks<>(SB),RODATA,$256

// func checkASM() bool
TEXT ·checkASM(SB),NOSPLIT,$0-1
	// check that masks<>(SB) and shifts<>(SB) are aligned to 16-byte
	MOVQ	$masks<>(SB), AX
	MOVQ	$shifts<>(SB), BX
	ORQ	BX, AX
	TESTQ	$15, AX
	SETEQ	ret+0(FP)
	RET

// these are arguments to pshufb. They move data down from
// the high bytes of the register to the low bytes of the register.
// index is how many bytes to move.
DATA shifts<>+0x00(SB)/8, $0x0000000000000000
DATA shifts<>+0x08(SB)/8, $0x0000000000000000
DATA shifts<>+0x10(SB)/8, $0xffffffffffffff0f
DATA shifts<>+0x18(SB)/8, $0xffffffffffffffff
DATA shifts<>+0x20(SB)/8, $0xffffffffffff0f0e
DATA shifts<>+0x28(SB)/8, $0xffffffffffffffff
DATA shifts<>+0x30(SB)/8, $0xffffffffff0f0e0d
DATA shifts<>+0x38(SB)/8, $0xffffffffffffffff
DATA shifts<>+0x40(SB)/8, $0xffffffff0f0e0d0c
DATA shifts<>+0x48(SB)/8, $0xffffffffffffffff
DATA shifts<>+0x50(SB)/8, $0xffffff0f0e0d0c0b
DATA shifts<>+0x58(SB)/8, $0xffffffffffffffff
DATA shifts<>+0x60(SB)/8, $0xffff0f0e0d0c0b0a
DATA shifts<>+0x68(SB)/8, $0xffffffffffffffff
DATA shifts<>+0x70(SB)/8, $0xff0f0e0d0c0b0a09
DATA shifts<>+0x78(SB)/8, $0xffffffffffffffff
DATA shifts<>+0x80(SB)/8, $0x0f0e0d0c0b0a0908
DATA shifts<>+0x88(SB)/8, $0xffffffffffffffff
DATA shifts<>+0x90(SB)/8, $0x0e0d0c0b0a090807
DATA shifts<>+0x98(SB)/8, $0xffffffffffffff0f
DATA shifts<>+0xa0(SB)/8, $0x0d0c0b0a09080706
DATA shifts<>+0xa8(SB)/8, $0xffffffffffff0f0e
DATA shifts<>+0xb0(SB)/8, $0x0c0b0a0908070605
DATA shifts<>+0xb8(SB)/8, $0xffffffffff0f0e0d
DATA shifts<>+0xc0(SB)/8, $0x0b0a090807060504
DATA shifts<>+0xc8(SB)/8, $0xffffffff0f0e0d0c
DATA shifts<>+0xd0(SB)/8, $0x0a09080706050403
DATA shifts<>+0xd8(SB)/8, $0xffffff0f0e0d0c0b
DATA shifts<>+0xe0(SB)/8, $0x0908070605040302
DATA shifts<>+0xe8(SB)/8, $0xffff0f0e0d0c0b0a
DATA shifts<>+0xf0(SB)/8, $0x0807060504030201
DATA shifts<>+0xf8(SB)/8, $0xff0f0e0d0c0b0a09
GLOBL shifts<>(SB),RODATA,$256

测试结果来看，10000个记录，大概也是182000ns, 每次大概是18.2ns。

总结：

算法	per hash cost
fnv	72ns
memhash	17.8ns
aeshash 汇编	18.2ns

从测试来看，memhash的hash和aeshash性能差不多，但是比fnv就要好很多了。