.

vendor/github.com/cespare/xxhash/v2/README.md

@@ -0,0 +1,74 @@
+# xxhash
+
+[![Go Reference](https://pkg.go.dev/badge/github.com/cespare/xxhash/v2.svg)](https://pkg.go.dev/github.com/cespare/xxhash/v2)
+[![Test](https://github.com/cespare/xxhash/actions/workflows/test.yml/badge.svg)](https://github.com/cespare/xxhash/actions/workflows/test.yml)
+
+xxhash is a Go implementation of the 64-bit [xxHash] algorithm, XXH64. This is a
+high-quality hashing algorithm that is much faster than anything in the Go
+standard library.
+
+This package provides a straightforward API:
+
+```
+func Sum64(b []byte) uint64
+func Sum64String(s string) uint64
+type Digest struct{ ... }
+    func New() *Digest
+```
+
+The `Digest` type implements hash.Hash64. Its key methods are:
+
+```
+func (*Digest) Write([]byte) (int, error)
+func (*Digest) WriteString(string) (int, error)
+func (*Digest) Sum64() uint64
+```
+
+The package is written with optimized pure Go and also contains even faster
+assembly implementations for amd64 and arm64. If desired, the `purego` build tag
+opts into using the Go code even on those architectures.
+
+[xxHash]: http://cyan4973.github.io/xxHash/
+
+## Compatibility
+
+This package is in a module and the latest code is in version 2 of the module.
+You need a version of Go with at least "minimal module compatibility" to use
+github.com/cespare/xxhash/v2:
+
+* 1.9.7+ for Go 1.9
+* 1.10.3+ for Go 1.10
+* Go 1.11 or later
+
+I recommend using the latest release of Go.
+
+## Benchmarks
+
+Here are some quick benchmarks comparing the pure-Go and assembly
+implementations of Sum64.
+
+| input size | purego    | asm       |
+| ---------- | --------- | --------- |
+| 4 B        |  1.3 GB/s |  1.2 GB/s |
+| 16 B       |  2.9 GB/s |  3.5 GB/s |
+| 100 B      |  6.9 GB/s |  8.1 GB/s |
+| 4 KB       | 11.7 GB/s | 16.7 GB/s |
+| 10 MB      | 12.0 GB/s | 17.3 GB/s |
+
+These numbers were generated on Ubuntu 20.04 with an Intel Xeon Platinum 8252C
+CPU using the following commands under Go 1.19.2:
+
+```
+benchstat <(go test -tags purego -benchtime 500ms -count 15 -bench 'Sum64$')
+benchstat <(go test -benchtime 500ms -count 15 -bench 'Sum64$')
+```
+
+## Projects using this package
+
+- [InfluxDB](https://github.com/influxdata/influxdb)
+- [Prometheus](https://github.com/prometheus/prometheus)
+- [VictoriaMetrics](https://github.com/VictoriaMetrics/VictoriaMetrics)
+- [FreeCache](https://github.com/coocood/freecache)
+- [FastCache](https://github.com/VictoriaMetrics/fastcache)
+- [Ristretto](https://github.com/dgraph-io/ristretto)
+- [Badger](https://github.com/dgraph-io/badger)

vendor/github.com/cespare/xxhash/v2/testall.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+set -eu -o pipefail
+
+# Small convenience script for running the tests with various combinations of
+# arch/tags. This assumes we're running on amd64 and have qemu available.
+
+go test ./...
+go test -tags purego ./...
+GOARCH=arm64 go test
+GOARCH=arm64 go test -tags purego

vendor/github.com/cespare/xxhash/v2/xxhash.go

@@ -0,0 +1,243 @@
+// Package xxhash implements the 64-bit variant of xxHash (XXH64) as described
+// at http://cyan4973.github.io/xxHash/.
+package xxhash
+
+import (
+	"encoding/binary"
+	"errors"
+	"math/bits"
+)
+
+const (
+	prime1 uint64 = 11400714785074694791
+	prime2 uint64 = 14029467366897019727
+	prime3 uint64 = 1609587929392839161
+	prime4 uint64 = 9650029242287828579
+	prime5 uint64 = 2870177450012600261
+)
+
+// Store the primes in an array as well.
+//
+// The consts are used when possible in Go code to avoid MOVs but we need a
+// contiguous array for the assembly code.
+var primes = [...]uint64{prime1, prime2, prime3, prime4, prime5}
+
+// Digest implements hash.Hash64.
+//
+// Note that a zero-valued Digest is not ready to receive writes.
+// Call Reset or create a Digest using New before calling other methods.
+type Digest struct {
+	v1    uint64
+	v2    uint64
+	v3    uint64
+	v4    uint64
+	total uint64
+	mem   [32]byte
+	n     int // how much of mem is used
+}
+
+// New creates a new Digest with a zero seed.
+func New() *Digest {
+	return NewWithSeed(0)
+}
+
+// NewWithSeed creates a new Digest with the given seed.
+func NewWithSeed(seed uint64) *Digest {
+	var d Digest
+	d.ResetWithSeed(seed)
+	return &d
+}
+
+// Reset clears the Digest's state so that it can be reused.
+// It uses a seed value of zero.
+func (d *Digest) Reset() {
+	d.ResetWithSeed(0)
+}
+
+// ResetWithSeed clears the Digest's state so that it can be reused.
+// It uses the given seed to initialize the state.
+func (d *Digest) ResetWithSeed(seed uint64) {
+	d.v1 = seed + prime1 + prime2
+	d.v2 = seed + prime2
+	d.v3 = seed
+	d.v4 = seed - prime1
+	d.total = 0
+	d.n = 0
+}
+
+// Size always returns 8 bytes.
+func (d *Digest) Size() int { return 8 }
+
+// BlockSize always returns 32 bytes.
+func (d *Digest) BlockSize() int { return 32 }
+
+// Write adds more data to d. It always returns len(b), nil.
+func (d *Digest) Write(b []byte) (n int, err error) {
+	n = len(b)
+	d.total += uint64(n)
+
+	memleft := d.mem[d.n&(len(d.mem)-1):]
+
+	if d.n+n < 32 {
+		// This new data doesn't even fill the current block.
+		copy(memleft, b)
+		d.n += n
+		return
+	}
+
+	if d.n > 0 {
+		// Finish off the partial block.
+		c := copy(memleft, b)
+		d.v1 = round(d.v1, u64(d.mem[0:8]))
+		d.v2 = round(d.v2, u64(d.mem[8:16]))
+		d.v3 = round(d.v3, u64(d.mem[16:24]))
+		d.v4 = round(d.v4, u64(d.mem[24:32]))
+		b = b[c:]
+		d.n = 0
+	}
+
+	if len(b) >= 32 {
+		// One or more full blocks left.
+		nw := writeBlocks(d, b)
+		b = b[nw:]
+	}
+
+	// Store any remaining partial block.
+	copy(d.mem[:], b)
+	d.n = len(b)
+
+	return
+}
+
+// Sum appends the current hash to b and returns the resulting slice.
+func (d *Digest) Sum(b []byte) []byte {
+	s := d.Sum64()
+	return append(
+		b,
+		byte(s>>56),
+		byte(s>>48),
+		byte(s>>40),
+		byte(s>>32),
+		byte(s>>24),
+		byte(s>>16),
+		byte(s>>8),
+		byte(s),
+	)
+}
+
+// Sum64 returns the current hash.
+func (d *Digest) Sum64() uint64 {
+	var h uint64
+
+	if d.total >= 32 {
+		v1, v2, v3, v4 := d.v1, d.v2, d.v3, d.v4
+		h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
+		h = mergeRound(h, v1)
+		h = mergeRound(h, v2)
+		h = mergeRound(h, v3)
+		h = mergeRound(h, v4)
+	} else {
+		h = d.v3 + prime5
+	}
+
+	h += d.total
+
+	b := d.mem[:d.n&(len(d.mem)-1)]
+	for ; len(b) >= 8; b = b[8:] {
+		k1 := round(0, u64(b[:8]))
+		h ^= k1
+		h = rol27(h)*prime1 + prime4
+	}
+	if len(b) >= 4 {
+		h ^= uint64(u32(b[:4])) * prime1
+		h = rol23(h)*prime2 + prime3
+		b = b[4:]
+	}
+	for ; len(b) > 0; b = b[1:] {
+		h ^= uint64(b[0]) * prime5
+		h = rol11(h) * prime1
+	}
+
+	h ^= h >> 33
+	h *= prime2
+	h ^= h >> 29
+	h *= prime3
+	h ^= h >> 32
+
+	return h
+}
+
+const (
+	magic         = "xxh\x06"
+	marshaledSize = len(magic) + 8*5 + 32
+)
+
+// MarshalBinary implements the encoding.BinaryMarshaler interface.
+func (d *Digest) MarshalBinary() ([]byte, error) {
+	b := make([]byte, 0, marshaledSize)
+	b = append(b, magic...)
+	b = appendUint64(b, d.v1)
+	b = appendUint64(b, d.v2)
+	b = appendUint64(b, d.v3)
+	b = appendUint64(b, d.v4)
+	b = appendUint64(b, d.total)
+	b = append(b, d.mem[:d.n]...)
+	b = b[:len(b)+len(d.mem)-d.n]
+	return b, nil
+}
+
+// UnmarshalBinary implements the encoding.BinaryUnmarshaler interface.
+func (d *Digest) UnmarshalBinary(b []byte) error {
+	if len(b) < len(magic) || string(b[:len(magic)]) != magic {
+		return errors.New("xxhash: invalid hash state identifier")
+	}
+	if len(b) != marshaledSize {
+		return errors.New("xxhash: invalid hash state size")
+	}
+	b = b[len(magic):]
+	b, d.v1 = consumeUint64(b)
+	b, d.v2 = consumeUint64(b)
+	b, d.v3 = consumeUint64(b)
+	b, d.v4 = consumeUint64(b)
+	b, d.total = consumeUint64(b)
+	copy(d.mem[:], b)
+	d.n = int(d.total % uint64(len(d.mem)))
+	return nil
+}
+
+func appendUint64(b []byte, x uint64) []byte {
+	var a [8]byte
+	binary.LittleEndian.PutUint64(a[:], x)
+	return append(b, a[:]...)
+}
+
+func consumeUint64(b []byte) ([]byte, uint64) {
+	x := u64(b)
+	return b[8:], x
+}
+
+func u64(b []byte) uint64 { return binary.LittleEndian.Uint64(b) }
+func u32(b []byte) uint32 { return binary.LittleEndian.Uint32(b) }
+
+func round(acc, input uint64) uint64 {
+	acc += input * prime2
+	acc = rol31(acc)
+	acc *= prime1
+	return acc
+}
+
+func mergeRound(acc, val uint64) uint64 {
+	val = round(0, val)
+	acc ^= val
+	acc = acc*prime1 + prime4
+	return acc
+}
+
+func rol1(x uint64) uint64  { return bits.RotateLeft64(x, 1) }
+func rol7(x uint64) uint64  { return bits.RotateLeft64(x, 7) }
+func rol11(x uint64) uint64 { return bits.RotateLeft64(x, 11) }
+func rol12(x uint64) uint64 { return bits.RotateLeft64(x, 12) }
+func rol18(x uint64) uint64 { return bits.RotateLeft64(x, 18) }
+func rol23(x uint64) uint64 { return bits.RotateLeft64(x, 23) }
+func rol27(x uint64) uint64 { return bits.RotateLeft64(x, 27) }
+func rol31(x uint64) uint64 { return bits.RotateLeft64(x, 31) }

vendor/github.com/cespare/xxhash/v2/xxhash_amd64.s

@@ -0,0 +1,209 @@
+//go:build !appengine && gc && !purego
+// +build !appengine
+// +build gc
+// +build !purego
+
+#include "textflag.h"
+
+// Registers:
+#define h      AX
+#define d      AX
+#define p      SI // pointer to advance through b
+#define n      DX
+#define end    BX // loop end
+#define v1     R8
+#define v2     R9
+#define v3     R10
+#define v4     R11
+#define x      R12
+#define prime1 R13
+#define prime2 R14
+#define prime4 DI
+
+#define round(acc, x) \
+	IMULQ prime2, x   \
+	ADDQ  x, acc      \
+	ROLQ  $31, acc    \
+	IMULQ prime1, acc
+
+// round0 performs the operation x = round(0, x).
+#define round0(x) \
+	IMULQ prime2, x \
+	ROLQ  $31, x    \
+	IMULQ prime1, x
+
+// mergeRound applies a merge round on the two registers acc and x.
+// It assumes that prime1, prime2, and prime4 have been loaded.
+#define mergeRound(acc, x) \
+	round0(x)         \
+	XORQ  x, acc      \
+	IMULQ prime1, acc \
+	ADDQ  prime4, acc
+
+// blockLoop processes as many 32-byte blocks as possible,
+// updating v1, v2, v3, and v4. It assumes that there is at least one block
+// to process.
+#define blockLoop() \
+loop:  \
+	MOVQ +0(p), x  \
+	round(v1, x)   \
+	MOVQ +8(p), x  \
+	round(v2, x)   \
+	MOVQ +16(p), x \
+	round(v3, x)   \
+	MOVQ +24(p), x \
+	round(v4, x)   \
+	ADDQ $32, p    \
+	CMPQ p, end    \
+	JLE  loop
+
+// func Sum64(b []byte) uint64
+TEXT ·Sum64(SB), NOSPLIT|NOFRAME, $0-32
+	// Load fixed primes.
+	MOVQ ·primes+0(SB), prime1
+	MOVQ ·primes+8(SB), prime2
+	MOVQ ·primes+24(SB), prime4
+
+	// Load slice.
+	MOVQ b_base+0(FP), p
+	MOVQ b_len+8(FP), n
+	LEAQ (p)(n*1), end
+
+	// The first loop limit will be len(b)-32.
+	SUBQ $32, end
+
+	// Check whether we have at least one block.
+	CMPQ n, $32
+	JLT  noBlocks
+
+	// Set up initial state (v1, v2, v3, v4).
+	MOVQ prime1, v1
+	ADDQ prime2, v1
+	MOVQ prime2, v2
+	XORQ v3, v3
+	XORQ v4, v4
+	SUBQ prime1, v4
+
+	blockLoop()
+
+	MOVQ v1, h
+	ROLQ $1, h
+	MOVQ v2, x
+	ROLQ $7, x
+	ADDQ x, h
+	MOVQ v3, x
+	ROLQ $12, x
+	ADDQ x, h
+	MOVQ v4, x
+	ROLQ $18, x
+	ADDQ x, h
+
+	mergeRound(h, v1)
+	mergeRound(h, v2)
+	mergeRound(h, v3)
+	mergeRound(h, v4)
+
+	JMP afterBlocks
+
+noBlocks:
+	MOVQ ·primes+32(SB), h
+
+afterBlocks:
+	ADDQ n, h
+
+	ADDQ $24, end
+	CMPQ p, end
+	JG   try4
+
+loop8:
+	MOVQ  (p), x
+	ADDQ  $8, p
+	round0(x)
+	XORQ  x, h
+	ROLQ  $27, h
+	IMULQ prime1, h
+	ADDQ  prime4, h
+
+	CMPQ p, end
+	JLE  loop8
+
+try4:
+	ADDQ $4, end
+	CMPQ p, end
+	JG   try1
+
+	MOVL  (p), x
+	ADDQ  $4, p
+	IMULQ prime1, x
+	XORQ  x, h
+
+	ROLQ  $23, h
+	IMULQ prime2, h
+	ADDQ  ·primes+16(SB), h
+
+try1:
+	ADDQ $4, end
+	CMPQ p, end
+	JGE  finalize
+
+loop1:
+	MOVBQZX (p), x
+	ADDQ    $1, p
+	IMULQ   ·primes+32(SB), x
+	XORQ    x, h
+	ROLQ    $11, h
+	IMULQ   prime1, h
+
+	CMPQ p, end
+	JL   loop1
+
+finalize:
+	MOVQ  h, x
+	SHRQ  $33, x
+	XORQ  x, h
+	IMULQ prime2, h
+	MOVQ  h, x
+	SHRQ  $29, x
+	XORQ  x, h
+	IMULQ ·primes+16(SB), h
+	MOVQ  h, x
+	SHRQ  $32, x
+	XORQ  x, h
+
+	MOVQ h, ret+24(FP)
+	RET
+
+// func writeBlocks(d *Digest, b []byte) int
+TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40
+	// Load fixed primes needed for round.
+	MOVQ ·primes+0(SB), prime1
+	MOVQ ·primes+8(SB), prime2
+
+	// Load slice.
+	MOVQ b_base+8(FP), p
+	MOVQ b_len+16(FP), n
+	LEAQ (p)(n*1), end
+	SUBQ $32, end
+
+	// Load vN from d.
+	MOVQ s+0(FP), d
+	MOVQ 0(d), v1
+	MOVQ 8(d), v2
+	MOVQ 16(d), v3
+	MOVQ 24(d), v4
+
+	// We don't need to check the loop condition here; this function is
+	// always called with at least one block of data to process.
+	blockLoop()
+
+	// Copy vN back to d.
+	MOVQ v1, 0(d)
+	MOVQ v2, 8(d)
+	MOVQ v3, 16(d)
+	MOVQ v4, 24(d)
+
+	// The number of bytes written is p minus the old base pointer.
+	SUBQ b_base+8(FP), p
+	MOVQ p, ret+32(FP)
+
+	RET

vendor/github.com/cespare/xxhash/v2/xxhash_arm64.s

@@ -0,0 +1,183 @@
+//go:build !appengine && gc && !purego
+// +build !appengine
+// +build gc
+// +build !purego
+
+#include "textflag.h"
+
+// Registers:
+#define digest	R1
+#define h	R2 // return value
+#define p	R3 // input pointer
+#define n	R4 // input length
+#define nblocks	R5 // n / 32
+#define prime1	R7
+#define prime2	R8
+#define prime3	R9
+#define prime4	R10
+#define prime5	R11
+#define v1	R12
+#define v2	R13
+#define v3	R14
+#define v4	R15
+#define x1	R20
+#define x2	R21
+#define x3	R22
+#define x4	R23
+
+#define round(acc, x) \
+	MADD prime2, acc, x, acc \
+	ROR  $64-31, acc         \
+	MUL  prime1, acc
+
+// round0 performs the operation x = round(0, x).
+#define round0(x) \
+	MUL prime2, x \
+	ROR $64-31, x \
+	MUL prime1, x
+
+#define mergeRound(acc, x) \
+	round0(x)                     \
+	EOR  x, acc                   \
+	MADD acc, prime4, prime1, acc
+
+// blockLoop processes as many 32-byte blocks as possible,
+// updating v1, v2, v3, and v4. It assumes that n >= 32.
+#define blockLoop() \
+	LSR     $5, n, nblocks  \
+	PCALIGN $16             \
+	loop:                   \
+	LDP.P   16(p), (x1, x2) \
+	LDP.P   16(p), (x3, x4) \
+	round(v1, x1)           \
+	round(v2, x2)           \
+	round(v3, x3)           \
+	round(v4, x4)           \
+	SUB     $1, nblocks     \
+	CBNZ    nblocks, loop
+
+// func Sum64(b []byte) uint64
+TEXT ·Sum64(SB), NOSPLIT|NOFRAME, $0-32
+	LDP b_base+0(FP), (p, n)
+
+	LDP  ·primes+0(SB), (prime1, prime2)
+	LDP  ·primes+16(SB), (prime3, prime4)
+	MOVD ·primes+32(SB), prime5
+
+	CMP  $32, n
+	CSEL LT, prime5, ZR, h // if n < 32 { h = prime5 } else { h = 0 }
+	BLT  afterLoop
+
+	ADD  prime1, prime2, v1
+	MOVD prime2, v2
+	MOVD $0, v3
+	NEG  prime1, v4
+
+	blockLoop()
+
+	ROR $64-1, v1, x1
+	ROR $64-7, v2, x2
+	ADD x1, x2
+	ROR $64-12, v3, x3
+	ROR $64-18, v4, x4
+	ADD x3, x4
+	ADD x2, x4, h
+
+	mergeRound(h, v1)
+	mergeRound(h, v2)
+	mergeRound(h, v3)
+	mergeRound(h, v4)
+
+afterLoop:
+	ADD n, h
+
+	TBZ   $4, n, try8
+	LDP.P 16(p), (x1, x2)
+
+	round0(x1)
+
+	// NOTE: here and below, sequencing the EOR after the ROR (using a
+	// rotated register) is worth a small but measurable speedup for small
+	// inputs.
+	ROR  $64-27, h
+	EOR  x1 @> 64-27, h, h
+	MADD h, prime4, prime1, h
+
+	round0(x2)
+	ROR  $64-27, h
+	EOR  x2 @> 64-27, h, h
+	MADD h, prime4, prime1, h
+
+try8:
+	TBZ    $3, n, try4
+	MOVD.P 8(p), x1
+
+	round0(x1)
+	ROR  $64-27, h
+	EOR  x1 @> 64-27, h, h
+	MADD h, prime4, prime1, h
+
+try4:
+	TBZ     $2, n, try2
+	MOVWU.P 4(p), x2
+
+	MUL  prime1, x2
+	ROR  $64-23, h
+	EOR  x2 @> 64-23, h, h
+	MADD h, prime3, prime2, h
+
+try2:
+	TBZ     $1, n, try1
+	MOVHU.P 2(p), x3
+	AND     $255, x3, x1
+	LSR     $8, x3, x2
+
+	MUL prime5, x1
+	ROR $64-11, h
+	EOR x1 @> 64-11, h, h
+	MUL prime1, h
+
+	MUL prime5, x2
+	ROR $64-11, h
+	EOR x2 @> 64-11, h, h
+	MUL prime1, h
+
+try1:
+	TBZ   $0, n, finalize
+	MOVBU (p), x4
+
+	MUL prime5, x4
+	ROR $64-11, h
+	EOR x4 @> 64-11, h, h
+	MUL prime1, h
+
+finalize:
+	EOR h >> 33, h
+	MUL prime2, h
+	EOR h >> 29, h
+	MUL prime3, h
+	EOR h >> 32, h
+
+	MOVD h, ret+24(FP)
+	RET
+
+// func writeBlocks(d *Digest, b []byte) int
+TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40
+	LDP ·primes+0(SB), (prime1, prime2)
+
+	// Load state. Assume v[1-4] are stored contiguously.
+	MOVD d+0(FP), digest
+	LDP  0(digest), (v1, v2)
+	LDP  16(digest), (v3, v4)
+
+	LDP b_base+8(FP), (p, n)
+
+	blockLoop()
+
+	// Store updated state.
+	STP (v1, v2), 0(digest)
+	STP (v3, v4), 16(digest)
+
+	BIC  $31, n
+	MOVD n, ret+32(FP)
+	RET

vendor/github.com/cespare/xxhash/v2/xxhash_asm.go

@@ -0,0 +1,15 @@
+//go:build (amd64 || arm64) && !appengine && gc && !purego
+// +build amd64 arm64
+// +build !appengine
+// +build gc
+// +build !purego
+
+package xxhash
+
+// Sum64 computes the 64-bit xxHash digest of b with a zero seed.
+//
+//go:noescape
+func Sum64(b []byte) uint64
+
+//go:noescape
+func writeBlocks(d *Digest, b []byte) int

vendor/github.com/cespare/xxhash/v2/xxhash_other.go

@@ -0,0 +1,76 @@
+//go:build (!amd64 && !arm64) || appengine || !gc || purego
+// +build !amd64,!arm64 appengine !gc purego
+
+package xxhash
+
+// Sum64 computes the 64-bit xxHash digest of b with a zero seed.
+func Sum64(b []byte) uint64 {
+	// A simpler version would be
+	//   d := New()
+	//   d.Write(b)
+	//   return d.Sum64()
+	// but this is faster, particularly for small inputs.
+
+	n := len(b)
+	var h uint64
+
+	if n >= 32 {
+		v1 := primes[0] + prime2
+		v2 := prime2
+		v3 := uint64(0)
+		v4 := -primes[0]
+		for len(b) >= 32 {
+			v1 = round(v1, u64(b[0:8:len(b)]))
+			v2 = round(v2, u64(b[8:16:len(b)]))
+			v3 = round(v3, u64(b[16:24:len(b)]))
+			v4 = round(v4, u64(b[24:32:len(b)]))
+			b = b[32:len(b):len(b)]
+		}
+		h = rol1(v1) + rol7(v2) + rol12(v3) + rol18(v4)
+		h = mergeRound(h, v1)
+		h = mergeRound(h, v2)
+		h = mergeRound(h, v3)
+		h = mergeRound(h, v4)
+	} else {
+		h = prime5
+	}
+
+	h += uint64(n)
+
+	for ; len(b) >= 8; b = b[8:] {
+		k1 := round(0, u64(b[:8]))
+		h ^= k1
+		h = rol27(h)*prime1 + prime4
+	}
+	if len(b) >= 4 {
+		h ^= uint64(u32(b[:4])) * prime1
+		h = rol23(h)*prime2 + prime3
+		b = b[4:]
+	}
+	for ; len(b) > 0; b = b[1:] {
+		h ^= uint64(b[0]) * prime5
+		h = rol11(h) * prime1
+	}
+
+	h ^= h >> 33
+	h *= prime2
+	h ^= h >> 29
+	h *= prime3
+	h ^= h >> 32
+
+	return h
+}
+
+func writeBlocks(d *Digest, b []byte) int {
+	v1, v2, v3, v4 := d.v1, d.v2, d.v3, d.v4
+	n := len(b)
+	for len(b) >= 32 {
+		v1 = round(v1, u64(b[0:8:len(b)]))
+		v2 = round(v2, u64(b[8:16:len(b)]))
+		v3 = round(v3, u64(b[16:24:len(b)]))
+		v4 = round(v4, u64(b[24:32:len(b)]))
+		b = b[32:len(b):len(b)]
+	}
+	d.v1, d.v2, d.v3, d.v4 = v1, v2, v3, v4
+	return n - len(b)
+}

vendor/github.com/cespare/xxhash/v2/xxhash_safe.go

@@ -0,0 +1,16 @@
+//go:build appengine
+// +build appengine
+
+// This file contains the safe implementations of otherwise unsafe-using code.
+
+package xxhash
+
+// Sum64String computes the 64-bit xxHash digest of s with a zero seed.
+func Sum64String(s string) uint64 {
+	return Sum64([]byte(s))
+}
+
+// WriteString adds more data to d. It always returns len(s), nil.
+func (d *Digest) WriteString(s string) (n int, err error) {
+	return d.Write([]byte(s))
+}

vendor/github.com/cespare/xxhash/v2/xxhash_unsafe.go

@@ -0,0 +1,58 @@
+//go:build !appengine
+// +build !appengine
+
+// This file encapsulates usage of unsafe.
+// xxhash_safe.go contains the safe implementations.
+
+package xxhash
+
+import (
+	"unsafe"
+)
+
+// In the future it's possible that compiler optimizations will make these
+// XxxString functions unnecessary by realizing that calls such as
+// Sum64([]byte(s)) don't need to copy s. See https://go.dev/issue/2205.
+// If that happens, even if we keep these functions they can be replaced with
+// the trivial safe code.
+
+// NOTE: The usual way of doing an unsafe string-to-[]byte conversion is:
+//
+//   var b []byte
+//   bh := (*reflect.SliceHeader)(unsafe.Pointer(&b))
+//   bh.Data = (*reflect.StringHeader)(unsafe.Pointer(&s)).Data
+//   bh.Len = len(s)
+//   bh.Cap = len(s)
+//
+// Unfortunately, as of Go 1.15.3 the inliner's cost model assigns a high enough
+// weight to this sequence of expressions that any function that uses it will
+// not be inlined. Instead, the functions below use a different unsafe
+// conversion designed to minimize the inliner weight and allow both to be
+// inlined. There is also a test (TestInlining) which verifies that these are
+// inlined.
+//
+// See https://github.com/golang/go/issues/42739 for discussion.
+
+// Sum64String computes the 64-bit xxHash digest of s with a zero seed.
+// It may be faster than Sum64([]byte(s)) by avoiding a copy.
+func Sum64String(s string) uint64 {
+	b := *(*[]byte)(unsafe.Pointer(&sliceHeader{s, len(s)}))
+	return Sum64(b)
+}
+
+// WriteString adds more data to d. It always returns len(s), nil.
+// It may be faster than Write([]byte(s)) by avoiding a copy.
+func (d *Digest) WriteString(s string) (n int, err error) {
+	d.Write(*(*[]byte)(unsafe.Pointer(&sliceHeader{s, len(s)})))
+	// d.Write always returns len(s), nil.
+	// Ignoring the return output and returning these fixed values buys a
+	// savings of 6 in the inliner's cost model.
+	return len(s), nil
+}
+
+// sliceHeader is similar to reflect.SliceHeader, but it assumes that the layout
+// of the first two words is the same as the layout of a string.
+type sliceHeader struct {
+	s   string
+	cap int
+}