Bug 41048: Remove the kcp-go project

projects/kcp-go/remove-crypt-fec.patch

-From 0b9d0759f979a5d828b747ea51771f307c53d221 Mon Sep 17 00:00:00 2001
-From: David Fifield <david@bamsoftware.com>
-Date: Thu, 9 Apr 2020 11:27:44 -0600
-Subject: [PATCH] Remove crypt and FEC dependencies.
-
----
- crypt.go   | 618 -----------------------------------------------------
- fec.go     | 337 -----------------------------
- removed.go |  29 +++
- 3 files changed, 29 insertions(+), 955 deletions(-)
- delete mode 100644 crypt.go
- delete mode 100644 fec.go
- create mode 100644 removed.go
-
-diff --git a/crypt.go b/crypt.go
-deleted file mode 100644
-index d882852..0000000
---- a/crypt.go
-+++ /dev/null
-@@ -1,618 +0,0 @@
--package kcp
--
--import (
--	"crypto/aes"
--	"crypto/cipher"
--	"crypto/des"
--	"crypto/sha1"
--	"unsafe"
--
--	xor "github.com/templexxx/xorsimd"
--	"github.com/tjfoc/gmsm/sm4"
--
--	"golang.org/x/crypto/blowfish"
--	"golang.org/x/crypto/cast5"
--	"golang.org/x/crypto/pbkdf2"
--	"golang.org/x/crypto/salsa20"
--	"golang.org/x/crypto/tea"
--	"golang.org/x/crypto/twofish"
--	"golang.org/x/crypto/xtea"
--)
--
--var (
--	initialVector = []byte{167, 115, 79, 156, 18, 172, 27, 1, 164, 21, 242, 193, 252, 120, 230, 107}
--	saltxor       = `sH3CIVoF#rWLtJo6`
--)
--
--// BlockCrypt defines encryption/decryption methods for a given byte slice.
--// Notes on implementing: the data to be encrypted contains a builtin
--// nonce at the first 16 bytes
--type BlockCrypt interface {
--	// Encrypt encrypts the whole block in src into dst.
--	// Dst and src may point at the same memory.
--	Encrypt(dst, src []byte)
--
--	// Decrypt decrypts the whole block in src into dst.
--	// Dst and src may point at the same memory.
--	Decrypt(dst, src []byte)
--}
--
--type salsa20BlockCrypt struct {
--	key [32]byte
--}
--
--// NewSalsa20BlockCrypt https://en.wikipedia.org/wiki/Salsa20
--func NewSalsa20BlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(salsa20BlockCrypt)
--	copy(c.key[:], key)
--	return c, nil
--}
--
--func (c *salsa20BlockCrypt) Encrypt(dst, src []byte) {
--	salsa20.XORKeyStream(dst[8:], src[8:], src[:8], &c.key)
--	copy(dst[:8], src[:8])
--}
--func (c *salsa20BlockCrypt) Decrypt(dst, src []byte) {
--	salsa20.XORKeyStream(dst[8:], src[8:], src[:8], &c.key)
--	copy(dst[:8], src[:8])
--}
--
--type sm4BlockCrypt struct {
--	encbuf [sm4.BlockSize]byte // 64bit alignment enc/dec buffer
--	decbuf [2 * sm4.BlockSize]byte
--	block  cipher.Block
--}
--
--// NewSM4BlockCrypt https://github.com/tjfoc/gmsm/tree/master/sm4
--func NewSM4BlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(sm4BlockCrypt)
--	block, err := sm4.NewCipher(key)
--	if err != nil {
--		return nil, err
--	}
--	c.block = block
--	return c, nil
--}
--
--func (c *sm4BlockCrypt) Encrypt(dst, src []byte) { encrypt(c.block, dst, src, c.encbuf[:]) }
--func (c *sm4BlockCrypt) Decrypt(dst, src []byte) { decrypt(c.block, dst, src, c.decbuf[:]) }
--
--type twofishBlockCrypt struct {
--	encbuf [twofish.BlockSize]byte
--	decbuf [2 * twofish.BlockSize]byte
--	block  cipher.Block
--}
--
--// NewTwofishBlockCrypt https://en.wikipedia.org/wiki/Twofish
--func NewTwofishBlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(twofishBlockCrypt)
--	block, err := twofish.NewCipher(key)
--	if err != nil {
--		return nil, err
--	}
--	c.block = block
--	return c, nil
--}
--
--func (c *twofishBlockCrypt) Encrypt(dst, src []byte) { encrypt(c.block, dst, src, c.encbuf[:]) }
--func (c *twofishBlockCrypt) Decrypt(dst, src []byte) { decrypt(c.block, dst, src, c.decbuf[:]) }
--
--type tripleDESBlockCrypt struct {
--	encbuf [des.BlockSize]byte
--	decbuf [2 * des.BlockSize]byte
--	block  cipher.Block
--}
--
--// NewTripleDESBlockCrypt https://en.wikipedia.org/wiki/Triple_DES
--func NewTripleDESBlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(tripleDESBlockCrypt)
--	block, err := des.NewTripleDESCipher(key)
--	if err != nil {
--		return nil, err
--	}
--	c.block = block
--	return c, nil
--}
--
--func (c *tripleDESBlockCrypt) Encrypt(dst, src []byte) { encrypt(c.block, dst, src, c.encbuf[:]) }
--func (c *tripleDESBlockCrypt) Decrypt(dst, src []byte) { decrypt(c.block, dst, src, c.decbuf[:]) }
--
--type cast5BlockCrypt struct {
--	encbuf [cast5.BlockSize]byte
--	decbuf [2 * cast5.BlockSize]byte
--	block  cipher.Block
--}
--
--// NewCast5BlockCrypt https://en.wikipedia.org/wiki/CAST-128
--func NewCast5BlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(cast5BlockCrypt)
--	block, err := cast5.NewCipher(key)
--	if err != nil {
--		return nil, err
--	}
--	c.block = block
--	return c, nil
--}
--
--func (c *cast5BlockCrypt) Encrypt(dst, src []byte) { encrypt(c.block, dst, src, c.encbuf[:]) }
--func (c *cast5BlockCrypt) Decrypt(dst, src []byte) { decrypt(c.block, dst, src, c.decbuf[:]) }
--
--type blowfishBlockCrypt struct {
--	encbuf [blowfish.BlockSize]byte
--	decbuf [2 * blowfish.BlockSize]byte
--	block  cipher.Block
--}
--
--// NewBlowfishBlockCrypt https://en.wikipedia.org/wiki/Blowfish_(cipher)
--func NewBlowfishBlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(blowfishBlockCrypt)
--	block, err := blowfish.NewCipher(key)
--	if err != nil {
--		return nil, err
--	}
--	c.block = block
--	return c, nil
--}
--
--func (c *blowfishBlockCrypt) Encrypt(dst, src []byte) { encrypt(c.block, dst, src, c.encbuf[:]) }
--func (c *blowfishBlockCrypt) Decrypt(dst, src []byte) { decrypt(c.block, dst, src, c.decbuf[:]) }
--
--type aesBlockCrypt struct {
--	encbuf [aes.BlockSize]byte
--	decbuf [2 * aes.BlockSize]byte
--	block  cipher.Block
--}
--
--// NewAESBlockCrypt https://en.wikipedia.org/wiki/Advanced_Encryption_Standard
--func NewAESBlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(aesBlockCrypt)
--	block, err := aes.NewCipher(key)
--	if err != nil {
--		return nil, err
--	}
--	c.block = block
--	return c, nil
--}
--
--func (c *aesBlockCrypt) Encrypt(dst, src []byte) { encrypt(c.block, dst, src, c.encbuf[:]) }
--func (c *aesBlockCrypt) Decrypt(dst, src []byte) { decrypt(c.block, dst, src, c.decbuf[:]) }
--
--type teaBlockCrypt struct {
--	encbuf [tea.BlockSize]byte
--	decbuf [2 * tea.BlockSize]byte
--	block  cipher.Block
--}
--
--// NewTEABlockCrypt https://en.wikipedia.org/wiki/Tiny_Encryption_Algorithm
--func NewTEABlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(teaBlockCrypt)
--	block, err := tea.NewCipherWithRounds(key, 16)
--	if err != nil {
--		return nil, err
--	}
--	c.block = block
--	return c, nil
--}
--
--func (c *teaBlockCrypt) Encrypt(dst, src []byte) { encrypt(c.block, dst, src, c.encbuf[:]) }
--func (c *teaBlockCrypt) Decrypt(dst, src []byte) { decrypt(c.block, dst, src, c.decbuf[:]) }
--
--type xteaBlockCrypt struct {
--	encbuf [xtea.BlockSize]byte
--	decbuf [2 * xtea.BlockSize]byte
--	block  cipher.Block
--}
--
--// NewXTEABlockCrypt https://en.wikipedia.org/wiki/XTEA
--func NewXTEABlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(xteaBlockCrypt)
--	block, err := xtea.NewCipher(key)
--	if err != nil {
--		return nil, err
--	}
--	c.block = block
--	return c, nil
--}
--
--func (c *xteaBlockCrypt) Encrypt(dst, src []byte) { encrypt(c.block, dst, src, c.encbuf[:]) }
--func (c *xteaBlockCrypt) Decrypt(dst, src []byte) { decrypt(c.block, dst, src, c.decbuf[:]) }
--
--type simpleXORBlockCrypt struct {
--	xortbl []byte
--}
--
--// NewSimpleXORBlockCrypt simple xor with key expanding
--func NewSimpleXORBlockCrypt(key []byte) (BlockCrypt, error) {
--	c := new(simpleXORBlockCrypt)
--	c.xortbl = pbkdf2.Key(key, []byte(saltxor), 32, mtuLimit, sha1.New)
--	return c, nil
--}
--
--func (c *simpleXORBlockCrypt) Encrypt(dst, src []byte) { xor.Bytes(dst, src, c.xortbl) }
--func (c *simpleXORBlockCrypt) Decrypt(dst, src []byte) { xor.Bytes(dst, src, c.xortbl) }
--
--type noneBlockCrypt struct{}
--
--// NewNoneBlockCrypt does nothing but copying
--func NewNoneBlockCrypt(key []byte) (BlockCrypt, error) {
--	return new(noneBlockCrypt), nil
--}
--
--func (c *noneBlockCrypt) Encrypt(dst, src []byte) { copy(dst, src) }
--func (c *noneBlockCrypt) Decrypt(dst, src []byte) { copy(dst, src) }
--
--// packet encryption with local CFB mode
--func encrypt(block cipher.Block, dst, src, buf []byte) {
--	switch block.BlockSize() {
--	case 8:
--		encrypt8(block, dst, src, buf)
--	case 16:
--		encrypt16(block, dst, src, buf)
--	default:
--		panic("unsupported cipher block size")
--	}
--}
--
--// optimized encryption for the ciphers which works in 8-bytes
--func encrypt8(block cipher.Block, dst, src, buf []byte) {
--	tbl := buf[:8]
--	block.Encrypt(tbl, initialVector)
--	n := len(src) / 8
--	base := 0
--	repeat := n / 8
--	left := n % 8
--	ptr_tbl := (*uint64)(unsafe.Pointer(&tbl[0]))
--
--	for i := 0; i < repeat; i++ {
--		s := src[base:][0:64]
--		d := dst[base:][0:64]
--		// 1
--		*(*uint64)(unsafe.Pointer(&d[0])) = *(*uint64)(unsafe.Pointer(&s[0])) ^ *ptr_tbl
--		block.Encrypt(tbl, d[0:8])
--		// 2
--		*(*uint64)(unsafe.Pointer(&d[8])) = *(*uint64)(unsafe.Pointer(&s[8])) ^ *ptr_tbl
--		block.Encrypt(tbl, d[8:16])
--		// 3
--		*(*uint64)(unsafe.Pointer(&d[16])) = *(*uint64)(unsafe.Pointer(&s[16])) ^ *ptr_tbl
--		block.Encrypt(tbl, d[16:24])
--		// 4
--		*(*uint64)(unsafe.Pointer(&d[24])) = *(*uint64)(unsafe.Pointer(&s[24])) ^ *ptr_tbl
--		block.Encrypt(tbl, d[24:32])
--		// 5
--		*(*uint64)(unsafe.Pointer(&d[32])) = *(*uint64)(unsafe.Pointer(&s[32])) ^ *ptr_tbl
--		block.Encrypt(tbl, d[32:40])
--		// 6
--		*(*uint64)(unsafe.Pointer(&d[40])) = *(*uint64)(unsafe.Pointer(&s[40])) ^ *ptr_tbl
--		block.Encrypt(tbl, d[40:48])
--		// 7
--		*(*uint64)(unsafe.Pointer(&d[48])) = *(*uint64)(unsafe.Pointer(&s[48])) ^ *ptr_tbl
--		block.Encrypt(tbl, d[48:56])
--		// 8
--		*(*uint64)(unsafe.Pointer(&d[56])) = *(*uint64)(unsafe.Pointer(&s[56])) ^ *ptr_tbl
--		block.Encrypt(tbl, d[56:64])
--		base += 64
--	}
--
--	switch left {
--	case 7:
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *ptr_tbl
--		block.Encrypt(tbl, dst[base:])
--		base += 8
--		fallthrough
--	case 6:
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *ptr_tbl
--		block.Encrypt(tbl, dst[base:])
--		base += 8
--		fallthrough
--	case 5:
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *ptr_tbl
--		block.Encrypt(tbl, dst[base:])
--		base += 8
--		fallthrough
--	case 4:
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *ptr_tbl
--		block.Encrypt(tbl, dst[base:])
--		base += 8
--		fallthrough
--	case 3:
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *ptr_tbl
--		block.Encrypt(tbl, dst[base:])
--		base += 8
--		fallthrough
--	case 2:
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *ptr_tbl
--		block.Encrypt(tbl, dst[base:])
--		base += 8
--		fallthrough
--	case 1:
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *ptr_tbl
--		block.Encrypt(tbl, dst[base:])
--		base += 8
--		fallthrough
--	case 0:
--		xorBytes(dst[base:], src[base:], tbl)
--	}
--}
--
--// optimized encryption for the ciphers which works in 16-bytes
--func encrypt16(block cipher.Block, dst, src, buf []byte) {
--	tbl := buf[:16]
--	block.Encrypt(tbl, initialVector)
--	n := len(src) / 16
--	base := 0
--	repeat := n / 8
--	left := n % 8
--	for i := 0; i < repeat; i++ {
--		s := src[base:][0:128]
--		d := dst[base:][0:128]
--		// 1
--		xor.Bytes16Align(d[0:16], s[0:16], tbl)
--		block.Encrypt(tbl, d[0:16])
--		// 2
--		xor.Bytes16Align(d[16:32], s[16:32], tbl)
--		block.Encrypt(tbl, d[16:32])
--		// 3
--		xor.Bytes16Align(d[32:48], s[32:48], tbl)
--		block.Encrypt(tbl, d[32:48])
--		// 4
--		xor.Bytes16Align(d[48:64], s[48:64], tbl)
--		block.Encrypt(tbl, d[48:64])
--		// 5
--		xor.Bytes16Align(d[64:80], s[64:80], tbl)
--		block.Encrypt(tbl, d[64:80])
--		// 6
--		xor.Bytes16Align(d[80:96], s[80:96], tbl)
--		block.Encrypt(tbl, d[80:96])
--		// 7
--		xor.Bytes16Align(d[96:112], s[96:112], tbl)
--		block.Encrypt(tbl, d[96:112])
--		// 8
--		xor.Bytes16Align(d[112:128], s[112:128], tbl)
--		block.Encrypt(tbl, d[112:128])
--		base += 128
--	}
--
--	switch left {
--	case 7:
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		block.Encrypt(tbl, dst[base:])
--		base += 16
--		fallthrough
--	case 6:
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		block.Encrypt(tbl, dst[base:])
--		base += 16
--		fallthrough
--	case 5:
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		block.Encrypt(tbl, dst[base:])
--		base += 16
--		fallthrough
--	case 4:
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		block.Encrypt(tbl, dst[base:])
--		base += 16
--		fallthrough
--	case 3:
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		block.Encrypt(tbl, dst[base:])
--		base += 16
--		fallthrough
--	case 2:
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		block.Encrypt(tbl, dst[base:])
--		base += 16
--		fallthrough
--	case 1:
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		block.Encrypt(tbl, dst[base:])
--		base += 16
--		fallthrough
--	case 0:
--		xorBytes(dst[base:], src[base:], tbl)
--	}
--}
--
--// decryption
--func decrypt(block cipher.Block, dst, src, buf []byte) {
--	switch block.BlockSize() {
--	case 8:
--		decrypt8(block, dst, src, buf)
--	case 16:
--		decrypt16(block, dst, src, buf)
--	default:
--		panic("unsupported cipher block size")
--	}
--}
--
--// decrypt 8 bytes block, all byte slices are supposed to be 64bit aligned
--func decrypt8(block cipher.Block, dst, src, buf []byte) {
--	tbl := buf[0:8]
--	next := buf[8:16]
--	block.Encrypt(tbl, initialVector)
--	n := len(src) / 8
--	base := 0
--	repeat := n / 8
--	left := n % 8
--	ptr_tbl := (*uint64)(unsafe.Pointer(&tbl[0]))
--	ptr_next := (*uint64)(unsafe.Pointer(&next[0]))
--
--	for i := 0; i < repeat; i++ {
--		s := src[base:][0:64]
--		d := dst[base:][0:64]
--		// 1
--		block.Encrypt(next, s[0:8])
--		*(*uint64)(unsafe.Pointer(&d[0])) = *(*uint64)(unsafe.Pointer(&s[0])) ^ *ptr_tbl
--		// 2
--		block.Encrypt(tbl, s[8:16])
--		*(*uint64)(unsafe.Pointer(&d[8])) = *(*uint64)(unsafe.Pointer(&s[8])) ^ *ptr_next
--		// 3
--		block.Encrypt(next, s[16:24])
--		*(*uint64)(unsafe.Pointer(&d[16])) = *(*uint64)(unsafe.Pointer(&s[16])) ^ *ptr_tbl
--		// 4
--		block.Encrypt(tbl, s[24:32])
--		*(*uint64)(unsafe.Pointer(&d[24])) = *(*uint64)(unsafe.Pointer(&s[24])) ^ *ptr_next
--		// 5
--		block.Encrypt(next, s[32:40])
--		*(*uint64)(unsafe.Pointer(&d[32])) = *(*uint64)(unsafe.Pointer(&s[32])) ^ *ptr_tbl
--		// 6
--		block.Encrypt(tbl, s[40:48])
--		*(*uint64)(unsafe.Pointer(&d[40])) = *(*uint64)(unsafe.Pointer(&s[40])) ^ *ptr_next
--		// 7
--		block.Encrypt(next, s[48:56])
--		*(*uint64)(unsafe.Pointer(&d[48])) = *(*uint64)(unsafe.Pointer(&s[48])) ^ *ptr_tbl
--		// 8
--		block.Encrypt(tbl, s[56:64])
--		*(*uint64)(unsafe.Pointer(&d[56])) = *(*uint64)(unsafe.Pointer(&s[56])) ^ *ptr_next
--		base += 64
--	}
--
--	switch left {
--	case 7:
--		block.Encrypt(next, src[base:])
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *(*uint64)(unsafe.Pointer(&tbl[0]))
--		tbl, next = next, tbl
--		base += 8
--		fallthrough
--	case 6:
--		block.Encrypt(next, src[base:])
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *(*uint64)(unsafe.Pointer(&tbl[0]))
--		tbl, next = next, tbl
--		base += 8
--		fallthrough
--	case 5:
--		block.Encrypt(next, src[base:])
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *(*uint64)(unsafe.Pointer(&tbl[0]))
--		tbl, next = next, tbl
--		base += 8
--		fallthrough
--	case 4:
--		block.Encrypt(next, src[base:])
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *(*uint64)(unsafe.Pointer(&tbl[0]))
--		tbl, next = next, tbl
--		base += 8
--		fallthrough
--	case 3:
--		block.Encrypt(next, src[base:])
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *(*uint64)(unsafe.Pointer(&tbl[0]))
--		tbl, next = next, tbl
--		base += 8
--		fallthrough
--	case 2:
--		block.Encrypt(next, src[base:])
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *(*uint64)(unsafe.Pointer(&tbl[0]))
--		tbl, next = next, tbl
--		base += 8
--		fallthrough
--	case 1:
--		block.Encrypt(next, src[base:])
--		*(*uint64)(unsafe.Pointer(&dst[base])) = *(*uint64)(unsafe.Pointer(&src[base])) ^ *(*uint64)(unsafe.Pointer(&tbl[0]))
--		tbl, next = next, tbl
--		base += 8
--		fallthrough
--	case 0:
--		xorBytes(dst[base:], src[base:], tbl)
--	}
--}
--
--func decrypt16(block cipher.Block, dst, src, buf []byte) {
--	tbl := buf[0:16]
--	next := buf[16:32]
--	block.Encrypt(tbl, initialVector)
--	n := len(src) / 16
--	base := 0
--	repeat := n / 8
--	left := n % 8
--	for i := 0; i < repeat; i++ {
--		s := src[base:][0:128]
--		d := dst[base:][0:128]
--		// 1
--		block.Encrypt(next, s[0:16])
--		xor.Bytes16Align(d[0:16], s[0:16], tbl)
--		// 2
--		block.Encrypt(tbl, s[16:32])
--		xor.Bytes16Align(d[16:32], s[16:32], next)
--		// 3
--		block.Encrypt(next, s[32:48])
--		xor.Bytes16Align(d[32:48], s[32:48], tbl)
--		// 4
--		block.Encrypt(tbl, s[48:64])
--		xor.Bytes16Align(d[48:64], s[48:64], next)
--		// 5
--		block.Encrypt(next, s[64:80])
--		xor.Bytes16Align(d[64:80], s[64:80], tbl)
--		// 6
--		block.Encrypt(tbl, s[80:96])
--		xor.Bytes16Align(d[80:96], s[80:96], next)
--		// 7
--		block.Encrypt(next, s[96:112])
--		xor.Bytes16Align(d[96:112], s[96:112], tbl)
--		// 8
--		block.Encrypt(tbl, s[112:128])
--		xor.Bytes16Align(d[112:128], s[112:128], next)
--		base += 128
--	}
--
--	switch left {
--	case 7:
--		block.Encrypt(next, src[base:])
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		tbl, next = next, tbl
--		base += 16
--		fallthrough
--	case 6:
--		block.Encrypt(next, src[base:])
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		tbl, next = next, tbl
--		base += 16
--		fallthrough
--	case 5:
--		block.Encrypt(next, src[base:])
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		tbl, next = next, tbl
--		base += 16
--		fallthrough
--	case 4:
--		block.Encrypt(next, src[base:])
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		tbl, next = next, tbl
--		base += 16
--		fallthrough
--	case 3:
--		block.Encrypt(next, src[base:])
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		tbl, next = next, tbl
--		base += 16
--		fallthrough
--	case 2:
--		block.Encrypt(next, src[base:])
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		tbl, next = next, tbl
--		base += 16
--		fallthrough
--	case 1:
--		block.Encrypt(next, src[base:])
--		xor.Bytes16Align(dst[base:], src[base:], tbl)
--		tbl, next = next, tbl
--		base += 16
--		fallthrough
--	case 0:
--		xorBytes(dst[base:], src[base:], tbl)
--	}
--}
--
--// per bytes xors
--func xorBytes(dst, a, b []byte) int {
--	n := len(a)
--	if len(b) < n {
--		n = len(b)
--	}
--	if n == 0 {
--		return 0
--	}
--
--	for i := 0; i < n; i++ {
--		dst[i] = a[i] ^ b[i]
--	}
--	return n
--}
-diff --git a/fec.go b/fec.go
-deleted file mode 100644
-index 97cd40b..0000000
---- a/fec.go
-+++ /dev/null
-@@ -1,337 +0,0 @@
--package kcp
--
--import (
--	"encoding/binary"
--	"sync/atomic"
--
--	"github.com/klauspost/reedsolomon"
--)
--
--const (
--	fecHeaderSize      = 6
--	fecHeaderSizePlus2 = fecHeaderSize + 2 // plus 2B data size
--	typeData           = 0xf1
--	typeParity         = 0xf2
--	fecExpire          = 60000
--)
--
--// fecPacket is a decoded FEC packet
--type fecPacket []byte
--
--func (bts fecPacket) seqid() uint32 { return binary.LittleEndian.Uint32(bts) }
--func (bts fecPacket) flag() uint16  { return binary.LittleEndian.Uint16(bts[4:]) }
--func (bts fecPacket) data() []byte  { return bts[6:] }
--
--// fecElement has auxcilliary time field
--type fecElement struct {
--	fecPacket
--	ts uint32
--}
--
--// fecDecoder for decoding incoming packets
--type fecDecoder struct {
--	rxlimit      int // queue size limit
--	dataShards   int
--	parityShards int
--	shardSize    int
--	rx           []fecElement // ordered receive queue
--
--	// caches
--	decodeCache [][]byte
--	flagCache   []bool
--
--	// zeros
--	zeros []byte
--
--	// RS decoder
--	codec reedsolomon.Encoder
--}
--
--func newFECDecoder(rxlimit, dataShards, parityShards int) *fecDecoder {
--	if dataShards <= 0 || parityShards <= 0 {
--		return nil
--	}
--	if rxlimit < dataShards+parityShards {
--		return nil
--	}
--
--	dec := new(fecDecoder)
--	dec.rxlimit = rxlimit
--	dec.dataShards = dataShards
--	dec.parityShards = parityShards
--	dec.shardSize = dataShards + parityShards
--	codec, err := reedsolomon.New(dataShards, parityShards)
--	if err != nil {
--		return nil
--	}
--	dec.codec = codec
--	dec.decodeCache = make([][]byte, dec.shardSize)
--	dec.flagCache = make([]bool, dec.shardSize)
--	dec.zeros = make([]byte, mtuLimit)
--	return dec
--}
--
--// decode a fec packet
--func (dec *fecDecoder) decode(in fecPacket) (recovered [][]byte) {
--	// insertion
--	n := len(dec.rx) - 1
--	insertIdx := 0
--	for i := n; i >= 0; i-- {
--		if in.seqid() == dec.rx[i].seqid() { // de-duplicate
--			return nil
--		} else if _itimediff(in.seqid(), dec.rx[i].seqid()) > 0 { // insertion
--			insertIdx = i + 1
--			break
--		}
--	}
--
--	// make a copy
--	pkt := fecPacket(xmitBuf.Get().([]byte)[:len(in)])
--	copy(pkt, in)
--	elem := fecElement{pkt, currentMs()}
--
--	// insert into ordered rx queue
--	if insertIdx == n+1 {
--		dec.rx = append(dec.rx, elem)
--	} else {
--		dec.rx = append(dec.rx, fecElement{})
--		copy(dec.rx[insertIdx+1:], dec.rx[insertIdx:]) // shift right
--		dec.rx[insertIdx] = elem
--	}
--
--	// shard range for current packet
--	shardBegin := pkt.seqid() - pkt.seqid()%uint32(dec.shardSize)
--	shardEnd := shardBegin + uint32(dec.shardSize) - 1
--
--	// max search range in ordered queue for current shard
--	searchBegin := insertIdx - int(pkt.seqid()%uint32(dec.shardSize))
--	if searchBegin < 0 {
--		searchBegin = 0
--	}
--	searchEnd := searchBegin + dec.shardSize - 1
--	if searchEnd >= len(dec.rx) {
--		searchEnd = len(dec.rx) - 1
--	}
--
--	// re-construct datashards
--	if searchEnd-searchBegin+1 >= dec.dataShards {
--		var numshard, numDataShard, first, maxlen int
--
--		// zero caches
--		shards := dec.decodeCache
--		shardsflag := dec.flagCache
--		for k := range dec.decodeCache {
--			shards[k] = nil
--			shardsflag[k] = false
--		}
--
--		// shard assembly
--		for i := searchBegin; i <= searchEnd; i++ {
--			seqid := dec.rx[i].seqid()
--			if _itimediff(seqid, shardEnd) > 0 {
--				break
--			} else if _itimediff(seqid, shardBegin) >= 0 {
--				shards[seqid%uint32(dec.shardSize)] = dec.rx[i].data()
--				shardsflag[seqid%uint32(dec.shardSize)] = true
--				numshard++
--				if dec.rx[i].flag() == typeData {
--					numDataShard++
--				}
--				if numshard == 1 {
--					first = i
--				}
--				if len(dec.rx[i].data()) > maxlen {
--					maxlen = len(dec.rx[i].data())
--				}
--			}
--		}
--
--		if numDataShard == dec.dataShards {
--			// case 1: no loss on data shards
--			dec.rx = dec.freeRange(first, numshard, dec.rx)
--		} else if numshard >= dec.dataShards {
--			// case 2: loss on data shards, but it's recoverable from parity shards
--			for k := range shards {
--				if shards[k] != nil {
--					dlen := len(shards[k])
--					shards[k] = shards[k][:maxlen]
--					copy(shards[k][dlen:], dec.zeros)
--				} else if k < dec.dataShards {
--					shards[k] = xmitBuf.Get().([]byte)[:0]
--				}
--			}
--			if err := dec.codec.ReconstructData(shards); err == nil {
--				for k := range shards[:dec.dataShards] {
--					if !shardsflag[k] {
--						// recovered data should be recycled
--						recovered = append(recovered, shards[k])
--					}
--				}
--			}
--			dec.rx = dec.freeRange(first, numshard, dec.rx)
--		}
--	}
--
--	// keep rxlimit
--	if len(dec.rx) > dec.rxlimit {
--		if dec.rx[0].flag() == typeData { // track the unrecoverable data
--			atomic.AddUint64(&DefaultSnmp.FECShortShards, 1)
--		}
--		dec.rx = dec.freeRange(0, 1, dec.rx)
--	}
--
--	// timeout policy
--	current := currentMs()
--	numExpired := 0
--	for k := range dec.rx {
--		if _itimediff(current, dec.rx[k].ts) > fecExpire {
--			numExpired++
--			continue
--		}
--		break
--	}
--	if numExpired > 0 {
--		dec.rx = dec.freeRange(0, numExpired, dec.rx)
--	}
--	return
--}
--
--// free a range of fecPacket
--func (dec *fecDecoder) freeRange(first, n int, q []fecElement) []fecElement {
--	for i := first; i < first+n; i++ { // recycle buffer
--		xmitBuf.Put([]byte(q[i].fecPacket))
--	}
--
--	if first == 0 && n < cap(q)/2 {
--		return q[n:]
--	}
--	copy(q[first:], q[first+n:])
--	return q[:len(q)-n]
--}
--
--// release all segments back to xmitBuf
--func (dec *fecDecoder) release() {
--	if n := len(dec.rx); n > 0 {
--		dec.rx = dec.freeRange(0, n, dec.rx)
--	}
--}
--
--type (
--	// fecEncoder for encoding outgoing packets
--	fecEncoder struct {
--		dataShards   int
--		parityShards int
--		shardSize    int
--		paws         uint32 // Protect Against Wrapped Sequence numbers
--		next         uint32 // next seqid
--
--		shardCount int // count the number of datashards collected
--		maxSize    int // track maximum data length in datashard
--
--		headerOffset  int // FEC header offset
--		payloadOffset int // FEC payload offset
--
--		// caches
--		shardCache  [][]byte
--		encodeCache [][]byte
--
--		// zeros
--		zeros []byte
--
--		// RS encoder
--		codec reedsolomon.Encoder
--	}
--)
--
--func newFECEncoder(dataShards, parityShards, offset int) *fecEncoder {
--	if dataShards <= 0 || parityShards <= 0 {
--		return nil
--	}
--	enc := new(fecEncoder)
--	enc.dataShards = dataShards
--	enc.parityShards = parityShards
--	enc.shardSize = dataShards + parityShards
--	enc.paws = 0xffffffff / uint32(enc.shardSize) * uint32(enc.shardSize)
--	enc.headerOffset = offset
--	enc.payloadOffset = enc.headerOffset + fecHeaderSize
--
--	codec, err := reedsolomon.New(dataShards, parityShards)
--	if err != nil {
--		return nil
--	}
--	enc.codec = codec
--
--	// caches
--	enc.encodeCache = make([][]byte, enc.shardSize)
--	enc.shardCache = make([][]byte, enc.shardSize)
--	for k := range enc.shardCache {
--		enc.shardCache[k] = make([]byte, mtuLimit)
--	}
--	enc.zeros = make([]byte, mtuLimit)
--	return enc
--}
--
--// encodes the packet, outputs parity shards if we have collected quorum datashards
--// notice: the contents of 'ps' will be re-written in successive calling
--func (enc *fecEncoder) encode(b []byte) (ps [][]byte) {
--	// The header format:
--	// | FEC SEQID(4B) | FEC TYPE(2B) | SIZE (2B) | PAYLOAD(SIZE-2) |
--	// |<-headerOffset                |<-payloadOffset
--	enc.markData(b[enc.headerOffset:])
--	binary.LittleEndian.PutUint16(b[enc.payloadOffset:], uint16(len(b[enc.payloadOffset:])))
--
--	// copy data from payloadOffset to fec shard cache
--	sz := len(b)
--	enc.shardCache[enc.shardCount] = enc.shardCache[enc.shardCount][:sz]
--	copy(enc.shardCache[enc.shardCount][enc.payloadOffset:], b[enc.payloadOffset:])
--	enc.shardCount++
--
--	// track max datashard length
--	if sz > enc.maxSize {
--		enc.maxSize = sz
--	}
--
--	//  Generation of Reed-Solomon Erasure Code
--	if enc.shardCount == enc.dataShards {
--		// fill '0' into the tail of each datashard
--		for i := 0; i < enc.dataShards; i++ {
--			shard := enc.shardCache[i]
--			slen := len(shard)
--			copy(shard[slen:enc.maxSize], enc.zeros)
--		}
--
--		// construct equal-sized slice with stripped header
--		cache := enc.encodeCache
--		for k := range cache {
--			cache[k] = enc.shardCache[k][enc.payloadOffset:enc.maxSize]
--		}
--
--		// encoding
--		if err := enc.codec.Encode(cache); err == nil {
--			ps = enc.shardCache[enc.dataShards:]
--			for k := range ps {
--				enc.markParity(ps[k][enc.headerOffset:])
--				ps[k] = ps[k][:enc.maxSize]
--			}
--		}
--
--		// counters resetting
--		enc.shardCount = 0
--		enc.maxSize = 0
--	}
--
--	return
--}
--
--func (enc *fecEncoder) markData(data []byte) {
--	binary.LittleEndian.PutUint32(data, enc.next)
--	binary.LittleEndian.PutUint16(data[4:], typeData)
--	enc.next++
--}
--
--func (enc *fecEncoder) markParity(data []byte) {
--	binary.LittleEndian.PutUint32(data, enc.next)
--	binary.LittleEndian.PutUint16(data[4:], typeParity)
--	// sequence wrap will only happen at parity shard
--	enc.next = (enc.next + 1) % enc.paws
--}
-diff --git a/removed.go b/removed.go
-new file mode 100644
-index 0000000..5ecf446
---- /dev/null
-+++ b/removed.go
-@@ -0,0 +1,29 @@
-+package kcp
-+
-+// Dummy implementations for types from crypt.go and fec.go, removed to reduce
-+// dependencies.
-+
-+const (
-+	fecHeaderSize      = 6
-+	fecHeaderSizePlus2 = fecHeaderSize + 2
-+	typeData           = 0xf1
-+	typeParity         = 0xf2
-+)
-+
-+type (
-+	BlockCrypt interface {
-+		Encrypt(_, _ []byte)
-+		Decrypt(_, _ []byte)
-+	}
-+	fecDecoder struct{}
-+	fecEncoder struct{}
-+	fecPacket  []byte
-+)
-+
-+func newFECDecoder(rxlimit, dataShards, parityShards int) *fecDecoder { return nil }
-+func newFECEncoder(dataShards, parityShards, offset int) *fecEncoder  { return nil }
-+
-+func (_ *fecDecoder) decode(in fecPacket) [][]byte { panic("disabled") }
-+func (_ *fecDecoder) release()                     { panic("disabled") }
-+func (_ *fecEncoder) encode(b []byte) [][]byte     { panic("disabled") }
-+func (_ fecPacket) flag() uint16                   { panic("disabled") }
--- 
-2.20.1
-