From 0b8e9e2aa343e5bd5180b27d75658e2fec0058c7 Mon Sep 17 00:00:00 2001
From: Pawel Srokosz <pawel.srokosz@cert.pl>
Date: Wed, 3 Jul 2019 17:22:06 +0200
Subject: [PATCH] Added serpent

---
 docs/crypto.rst                        |    2 +
 malduck/__init__.py                    |    2 +-
 malduck/crypto/components/__init__.py  |    0
 malduck/crypto/components/pyserpent.py | 3003 ++++++++++++++++++++++++
 malduck/crypto/serpent.py              |   13 +
 malduck/short.py                       |    5 +
 tests/test_crypto.py                   |    8 +-
 7 files changed, 3031 insertions(+), 2 deletions(-)
 create mode 100644 malduck/crypto/components/__init__.py
 create mode 100644 malduck/crypto/components/pyserpent.py
 create mode 100644 malduck/crypto/serpent.py

diff --git a/docs/crypto.rst b/docs/crypto.rst
index c118435..9503506 100644
--- a/docs/crypto.rst
+++ b/docs/crypto.rst
@@ -16,3 +16,5 @@ Cryptography
 .. autoclass:: malduck.crypto.rsa.RSA
     :members:
 .. autofunction:: malduck.crypto.xor.xor
+.. autoclass:: malduck.crypto.serpent.Serpent
+    :members:
diff --git a/malduck/__init__.py b/malduck/__init__.py
index d644331..8040db6 100644
--- a/malduck/__init__.py
+++ b/malduck/__init__.py
@@ -20,7 +20,7 @@
 
 from .short import (
     aes, blowfish, des3, rc4, pe, aplib, gzip, procmem, procmempe, cuckoomem, pad, unpad,
-    insn, rsa, verify, base64, rabbit
+    insn, rsa, verify, base64, rabbit, serpent
 )
 
 from .string.bin import (
diff --git a/malduck/crypto/components/__init__.py b/malduck/crypto/components/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/malduck/crypto/components/pyserpent.py b/malduck/crypto/components/pyserpent.py
new file mode 100644
index 0000000..719858b
--- /dev/null
+++ b/malduck/crypto/components/pyserpent.py
@@ -0,0 +1,3003 @@
+# serpent.py - pure Python implementation of the Serpent algorithm.
+# Bjorn Edstrom <be@bjrn.se> 13 december 2007.
+#
+# Found on https://github.com/sysopfb/pyserpent/blob/master/serpent.py
+# Modified for Py2/3 support by psrok1 (03-09-2019)
+#
+# Copyrights
+# ==========
+#
+# This code is a derived from an implementation by Dr Brian Gladman
+# (gladman@seven77.demon.co.uk) which is subject to the following license.
+# This Python implementation is not subject to any other license.
+#
+#/* This is an independent implementation of the encryption algorithm:
+# *
+# * Serpent by Ross Anderson, Eli Biham and Lars Knudsen
+# *
+# * which is a candidate algorithm in the Advanced Encryption Standard
+# * programme of the US National Institute of Standards and Technology
+# *
+# * Copyright in this implementation is held by Dr B R Gladman but I
+# * hereby give permission for its free direct or derivative use subject
+# * to acknowledgment of its origin and compliance with any conditions
+# * that the originators of the algorithm place on its exploitation.
+# *
+# * Dr Brian Gladman (gladman@seven77.demon.co.uk) 14th January 1999
+# */
+#
+# The above copyright notice must not be removed.
+#
+# Information
+# ===========
+#
+# Anyone thinking of using this code should reconsider. It's slow.
+# Try python-mcrypt instead. In case a faster library is not installed
+# on the target system, this code can be used as a portable fallback.
+
+block_size = 16
+key_size = 32
+
+
+class Serpent:
+
+    def __init__(self, key=None):
+        """Serpent."""
+
+        if key:
+            self.set_key(key)
+
+    def set_key(self, key):
+        """Init."""
+
+        key_len = len(key)
+        if key_len % 4:
+            # XXX: add padding?
+            raise KeyError("key not a multiple of 4")
+        if key_len > 32:
+            # XXX: prune?
+            raise KeyError("key_len > 32")
+
+        self.key_context = [0] * 140
+
+        key_word32 = [0] * 32
+        i = 0
+        while key:
+            key_word32[i] = struct.unpack("<L", key[0:4])[0]
+            key = key[4:]
+            i += 1
+
+        set_key(self.key_context, key_word32, key_len)
+        # print(map(hex,self.key_context))
+
+    def decrypt(self, block):
+        """Decrypt blocks."""
+
+        if len(block) % 16:
+            raise ValueError("block size must be a multiple of 16")
+
+        plaintext = b''
+
+        while block:
+            a, b, c, d = struct.unpack("<4L", block[:16])
+            temp = [a, b, c, d]
+            decrypt(self.key_context, temp)
+            plaintext += struct.pack("<4L", *temp)
+            block = block[16:]
+
+        return plaintext
+
+    def encrypt(self, block):
+        """Encrypt blocks."""
+
+        if len(block) % 16:
+            raise ValueError("block size must be a multiple of 16")
+
+        ciphertext = b''
+
+        while block:
+            a, b, c, d = struct.unpack("<4L", block[0:16])
+            temp = [a, b, c, d]
+            encrypt(self.key_context, temp)
+            ciphertext += struct.pack("<4L", *temp)
+            block = block[16:]
+
+        return ciphertext
+
+    def get_name(self):
+        """Return the name of the cipher."""
+
+        return "Serpent"
+
+    def get_block_size(self):
+        """Get cipher block size in bytes."""
+
+        return 16
+
+    def get_key_size(self):
+        """Get cipher key size in bytes."""
+
+        return 32
+
+
+#
+# Private.
+#
+
+import struct
+import sys
+
+WORD_BIGENDIAN = 0
+if sys.byteorder == 'big':
+    WORD_BIGENDIAN = 1
+
+
+def rotr32(x, n):
+    return (x >> n) | ((x << (32 - n)) & 0xFFFFFFFF)
+
+
+def rotl32(x, n):
+    return ((x << n) & 0xFFFFFFFF) | (x >> (32 - n))
+
+
+def byteswap32(x):
+    return ((x & 0xff) << 24) | (((x >> 8) & 0xff) << 16) | \
+           (((x >> 16) & 0xff) << 8) | ((x >> 24) & 0xff)
+
+
+def set_key(l_key, key, key_len):
+    key_len *= 8
+    if key_len > 256:
+        return False
+
+    i = 0
+    lk = (key_len + 31) // 32
+    while i < lk:
+        l_key[i] = key[i]
+        if WORD_BIGENDIAN:
+            l_key[i] = byteswap32(key[i])
+        i += 1
+
+    if key_len < 256:
+        while i < 8:
+            l_key[i] = 0
+            i += 1
+        i = key_len // 32
+        lk = 1 << (key_len % 32)
+        l_key[i] = (l_key[i] & (lk - 1)) | lk
+    for i in range(132):
+        lk = l_key[i] ^ l_key[i + 3] ^ l_key[i + 5] ^ l_key[i + 7] ^ 0x9e3779b9 ^ i
+        l_key[i + 8] = ((lk << 11) & 0xFFFFFFFF) | (lk >> 21)
+
+    key = l_key
+    # serpent_generate.py
+    a = key[4 * 0 + 8]
+    b = key[4 * 0 + 9]
+    c = key[4 * 0 + 10]
+    d = key[4 * 0 + 11]
+    e = 0
+    f = 0
+    g = 0
+    h = 0
+    t1 = 0
+    t2 = 0
+    t3 = 0
+    t4 = 0
+    t5 = 0
+    t6 = 0
+    t7 = 0
+    t8 = 0
+    t9 = 0
+    t10 = 0
+    t11 = 0
+    t12 = 0
+    t13 = 0
+    t14 = 0
+    t15 = 0
+    t16 = 0
+    t1 = a ^ c
+    t2 = d ^ t1
+    t3 = a & t2
+    t4 = d ^ t3
+    t5 = b & t4
+    g = t2 ^ t5
+    t7 = a | g
+    t8 = b | d
+    t11 = a | d
+    t9 = t4 & t7
+    f = t8 ^ t9
+    t12 = b ^ t11
+    t13 = g ^ t9
+    t15 = t3 ^ t8
+    h = t12 ^ t13
+    t16 = c & t15
+    e = t12 ^ t16
+    key[4 * 0 + 8] = e
+    key[4 * 0 + 9] = f
+    key[4 * 0 + 10] = g
+    key[4 * 0 + 11] = h
+    a = key[4 * 1 + 8]
+    b = key[4 * 1 + 9]
+    c = key[4 * 1 + 10]
+    d = key[4 * 1 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ d
+    t3 = c & t1
+    t13 = d | t1
+    e = t2 ^ t3
+    t5 = c ^ t1
+    t6 = c ^ e
+    t7 = b & t6
+    t10 = e | t5
+    h = t5 ^ t7
+    t9 = d | t7
+    t11 = t9 & t10
+    t14 = t2 ^ h
+    g = a ^ t11
+    t15 = g ^ t13
+    f = t14 ^ t15
+    key[4 * 1 + 8] = e
+    key[4 * 1 + 9] = f
+    key[4 * 1 + 10] = g
+    key[4 * 1 + 11] = h
+    a = key[4 * 2 + 8]
+    b = key[4 * 2 + 9]
+    c = key[4 * 2 + 10]
+    d = key[4 * 2 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ t1
+    t3 = a | t2
+    t4 = d | t2
+    t5 = c ^ t3
+    g = d ^ t5
+    t7 = b ^ t4
+    t8 = t2 ^ g
+    t9 = t5 & t7
+    h = t8 ^ t9
+    t11 = t5 ^ t7
+    f = h ^ t11
+    t13 = t8 & t11
+    e = t5 ^ t13
+    key[4 * 2 + 8] = e
+    key[4 * 2 + 9] = f
+    key[4 * 2 + 10] = g
+    key[4 * 2 + 11] = h
+    a = key[4 * 3 + 8]
+    b = key[4 * 3 + 9]
+    c = key[4 * 3 + 10]
+    d = key[4 * 3 + 11]
+    t1 = a ^ d
+    t2 = a & d
+    t3 = c ^ t1
+    t6 = b & t1
+    t4 = b ^ t3
+    t10 = (~t3) % 0x100000000
+    h = t2 ^ t4
+    t7 = a ^ t6
+    t14 = (~t7) % 0x100000000
+    t8 = c | t7
+    t11 = t3 ^ t7
+    g = t4 ^ t8
+    t12 = h & t11
+    f = t10 ^ t12
+    e = t12 ^ t14
+    key[4 * 3 + 8] = e
+    key[4 * 3 + 9] = f
+    key[4 * 3 + 10] = g
+    key[4 * 3 + 11] = h
+    a = key[4 * 4 + 8]
+    b = key[4 * 4 + 9]
+    c = key[4 * 4 + 10]
+    d = key[4 * 4 + 11]
+    t1 = (~c) % 0x100000000
+    t2 = b ^ c
+    t3 = b | t1
+    t4 = d ^ t3
+    t5 = a & t4
+    t7 = a ^ d
+    h = t2 ^ t5
+    t8 = b ^ t5
+    t9 = t2 | t8
+    t11 = d & t3
+    f = t7 ^ t9
+    t12 = t5 ^ f
+    t15 = t1 | t4
+    t13 = h & t12
+    g = t11 ^ t13
+    t16 = t12 ^ g
+    e = t15 ^ t16
+    key[4 * 4 + 8] = e
+    key[4 * 4 + 9] = f
+    key[4 * 4 + 10] = g
+    key[4 * 4 + 11] = h
+    a = key[4 * 5 + 8]
+    b = key[4 * 5 + 9]
+    c = key[4 * 5 + 10]
+    d = key[4 * 5 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ d
+    t3 = b ^ t2
+    t4 = t1 | t2
+    t5 = c ^ t4
+    f = b ^ t5
+    t13 = (~t5) % 0x100000000
+    t7 = t2 | f
+    t8 = d ^ t7
+    t9 = t5 & t8
+    g = t3 ^ t9
+    t11 = t5 ^ t8
+    e = g ^ t11
+    t14 = t3 & t11
+    h = t13 ^ t14
+    key[4 * 5 + 8] = e
+    key[4 * 5 + 9] = f
+    key[4 * 5 + 10] = g
+    key[4 * 5 + 11] = h
+    a = key[4 * 6 + 8]
+    b = key[4 * 6 + 9]
+    c = key[4 * 6 + 10]
+    d = key[4 * 6 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ b
+    t3 = a ^ d
+    t4 = c ^ t1
+    t5 = t2 | t3
+    e = t4 ^ t5
+    t7 = d & e
+    t8 = t2 ^ e
+    t10 = t1 | e
+    f = t7 ^ t8
+    t11 = t2 | t7
+    t12 = t3 ^ t10
+    t14 = b ^ t7
+    g = t11 ^ t12
+    t15 = f & t12
+    h = t14 ^ t15
+    key[4 * 6 + 8] = e
+    key[4 * 6 + 9] = f
+    key[4 * 6 + 10] = g
+    key[4 * 6 + 11] = h
+    a = key[4 * 7 + 8]
+    b = key[4 * 7 + 9]
+    c = key[4 * 7 + 10]
+    d = key[4 * 7 + 11]
+    t1 = a ^ d
+    t2 = d & t1
+    t3 = c ^ t2
+    t4 = b | t3
+    h = t1 ^ t4
+    t6 = (~b) % 0x100000000
+    t7 = t1 | t6
+    e = t3 ^ t7
+    t9 = a & e
+    t10 = t1 ^ t6
+    t11 = t4 & t10
+    g = t9 ^ t11
+    t13 = a ^ t3
+    t14 = t10 & g
+    f = t13 ^ t14
+    key[4 * 7 + 8] = e
+    key[4 * 7 + 9] = f
+    key[4 * 7 + 10] = g
+    key[4 * 7 + 11] = h
+    a = key[4 * 8 + 8]
+    b = key[4 * 8 + 9]
+    c = key[4 * 8 + 10]
+    d = key[4 * 8 + 11]
+    t1 = a ^ c
+    t2 = d ^ t1
+    t3 = a & t2
+    t4 = d ^ t3
+    t5 = b & t4
+    g = t2 ^ t5
+    t7 = a | g
+    t8 = b | d
+    t11 = a | d
+    t9 = t4 & t7
+    f = t8 ^ t9
+    t12 = b ^ t11
+    t13 = g ^ t9
+    t15 = t3 ^ t8
+    h = t12 ^ t13
+    t16 = c & t15
+    e = t12 ^ t16
+    key[4 * 8 + 8] = e
+    key[4 * 8 + 9] = f
+    key[4 * 8 + 10] = g
+    key[4 * 8 + 11] = h
+    a = key[4 * 9 + 8]
+    b = key[4 * 9 + 9]
+    c = key[4 * 9 + 10]
+    d = key[4 * 9 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ d
+    t3 = c & t1
+    t13 = d | t1
+    e = t2 ^ t3
+    t5 = c ^ t1
+    t6 = c ^ e
+    t7 = b & t6
+    t10 = e | t5
+    h = t5 ^ t7
+    t9 = d | t7
+    t11 = t9 & t10
+    t14 = t2 ^ h
+    g = a ^ t11
+    t15 = g ^ t13
+    f = t14 ^ t15
+    key[4 * 9 + 8] = e
+    key[4 * 9 + 9] = f
+    key[4 * 9 + 10] = g
+    key[4 * 9 + 11] = h
+    a = key[4 * 10 + 8]
+    b = key[4 * 10 + 9]
+    c = key[4 * 10 + 10]
+    d = key[4 * 10 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ t1
+    t3 = a | t2
+    t4 = d | t2
+    t5 = c ^ t3
+    g = d ^ t5
+    t7 = b ^ t4
+    t8 = t2 ^ g
+    t9 = t5 & t7
+    h = t8 ^ t9
+    t11 = t5 ^ t7
+    f = h ^ t11
+    t13 = t8 & t11
+    e = t5 ^ t13
+    key[4 * 10 + 8] = e
+    key[4 * 10 + 9] = f
+    key[4 * 10 + 10] = g
+    key[4 * 10 + 11] = h
+    a = key[4 * 11 + 8]
+    b = key[4 * 11 + 9]
+    c = key[4 * 11 + 10]
+    d = key[4 * 11 + 11]
+    t1 = a ^ d
+    t2 = a & d
+    t3 = c ^ t1
+    t6 = b & t1
+    t4 = b ^ t3
+    t10 = (~t3) % 0x100000000
+    h = t2 ^ t4
+    t7 = a ^ t6
+    t14 = (~t7) % 0x100000000
+    t8 = c | t7
+    t11 = t3 ^ t7
+    g = t4 ^ t8
+    t12 = h & t11
+    f = t10 ^ t12
+    e = t12 ^ t14
+    key[4 * 11 + 8] = e
+    key[4 * 11 + 9] = f
+    key[4 * 11 + 10] = g
+    key[4 * 11 + 11] = h
+    a = key[4 * 12 + 8]
+    b = key[4 * 12 + 9]
+    c = key[4 * 12 + 10]
+    d = key[4 * 12 + 11]
+    t1 = (~c) % 0x100000000
+    t2 = b ^ c
+    t3 = b | t1
+    t4 = d ^ t3
+    t5 = a & t4
+    t7 = a ^ d
+    h = t2 ^ t5
+    t8 = b ^ t5
+    t9 = t2 | t8
+    t11 = d & t3
+    f = t7 ^ t9
+    t12 = t5 ^ f
+    t15 = t1 | t4
+    t13 = h & t12
+    g = t11 ^ t13
+    t16 = t12 ^ g
+    e = t15 ^ t16
+    key[4 * 12 + 8] = e
+    key[4 * 12 + 9] = f
+    key[4 * 12 + 10] = g
+    key[4 * 12 + 11] = h
+    a = key[4 * 13 + 8]
+    b = key[4 * 13 + 9]
+    c = key[4 * 13 + 10]
+    d = key[4 * 13 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ d
+    t3 = b ^ t2
+    t4 = t1 | t2
+    t5 = c ^ t4
+    f = b ^ t5
+    t13 = (~t5) % 0x100000000
+    t7 = t2 | f
+    t8 = d ^ t7
+    t9 = t5 & t8
+    g = t3 ^ t9
+    t11 = t5 ^ t8
+    e = g ^ t11
+    t14 = t3 & t11
+    h = t13 ^ t14
+    key[4 * 13 + 8] = e
+    key[4 * 13 + 9] = f
+    key[4 * 13 + 10] = g
+    key[4 * 13 + 11] = h
+    a = key[4 * 14 + 8]
+    b = key[4 * 14 + 9]
+    c = key[4 * 14 + 10]
+    d = key[4 * 14 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ b
+    t3 = a ^ d
+    t4 = c ^ t1
+    t5 = t2 | t3
+    e = t4 ^ t5
+    t7 = d & e
+    t8 = t2 ^ e
+    t10 = t1 | e
+    f = t7 ^ t8
+    t11 = t2 | t7
+    t12 = t3 ^ t10
+    t14 = b ^ t7
+    g = t11 ^ t12
+    t15 = f & t12
+    h = t14 ^ t15
+    key[4 * 14 + 8] = e
+    key[4 * 14 + 9] = f
+    key[4 * 14 + 10] = g
+    key[4 * 14 + 11] = h
+    a = key[4 * 15 + 8]
+    b = key[4 * 15 + 9]
+    c = key[4 * 15 + 10]
+    d = key[4 * 15 + 11]
+    t1 = a ^ d
+    t2 = d & t1
+    t3 = c ^ t2
+    t4 = b | t3
+    h = t1 ^ t4
+    t6 = (~b) % 0x100000000
+    t7 = t1 | t6
+    e = t3 ^ t7
+    t9 = a & e
+    t10 = t1 ^ t6
+    t11 = t4 & t10
+    g = t9 ^ t11
+    t13 = a ^ t3
+    t14 = t10 & g
+    f = t13 ^ t14
+    key[4 * 15 + 8] = e
+    key[4 * 15 + 9] = f
+    key[4 * 15 + 10] = g
+    key[4 * 15 + 11] = h
+    a = key[4 * 16 + 8]
+    b = key[4 * 16 + 9]
+    c = key[4 * 16 + 10]
+    d = key[4 * 16 + 11]
+    t1 = a ^ c
+    t2 = d ^ t1
+    t3 = a & t2
+    t4 = d ^ t3
+    t5 = b & t4
+    g = t2 ^ t5
+    t7 = a | g
+    t8 = b | d
+    t11 = a | d
+    t9 = t4 & t7
+    f = t8 ^ t9
+    t12 = b ^ t11
+    t13 = g ^ t9
+    t15 = t3 ^ t8
+    h = t12 ^ t13
+    t16 = c & t15
+    e = t12 ^ t16
+    key[4 * 16 + 8] = e
+    key[4 * 16 + 9] = f
+    key[4 * 16 + 10] = g
+    key[4 * 16 + 11] = h
+    a = key[4 * 17 + 8]
+    b = key[4 * 17 + 9]
+    c = key[4 * 17 + 10]
+    d = key[4 * 17 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ d
+    t3 = c & t1
+    t13 = d | t1
+    e = t2 ^ t3
+    t5 = c ^ t1
+    t6 = c ^ e
+    t7 = b & t6
+    t10 = e | t5
+    h = t5 ^ t7
+    t9 = d | t7
+    t11 = t9 & t10
+    t14 = t2 ^ h
+    g = a ^ t11
+    t15 = g ^ t13
+    f = t14 ^ t15
+    key[4 * 17 + 8] = e
+    key[4 * 17 + 9] = f
+    key[4 * 17 + 10] = g
+    key[4 * 17 + 11] = h
+    a = key[4 * 18 + 8]
+    b = key[4 * 18 + 9]
+    c = key[4 * 18 + 10]
+    d = key[4 * 18 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ t1
+    t3 = a | t2
+    t4 = d | t2
+    t5 = c ^ t3
+    g = d ^ t5
+    t7 = b ^ t4
+    t8 = t2 ^ g
+    t9 = t5 & t7
+    h = t8 ^ t9
+    t11 = t5 ^ t7
+    f = h ^ t11
+    t13 = t8 & t11
+    e = t5 ^ t13
+    key[4 * 18 + 8] = e
+    key[4 * 18 + 9] = f
+    key[4 * 18 + 10] = g
+    key[4 * 18 + 11] = h
+    a = key[4 * 19 + 8]
+    b = key[4 * 19 + 9]
+    c = key[4 * 19 + 10]
+    d = key[4 * 19 + 11]
+    t1 = a ^ d
+    t2 = a & d
+    t3 = c ^ t1
+    t6 = b & t1
+    t4 = b ^ t3
+    t10 = (~t3) % 0x100000000
+    h = t2 ^ t4
+    t7 = a ^ t6
+    t14 = (~t7) % 0x100000000
+    t8 = c | t7
+    t11 = t3 ^ t7
+    g = t4 ^ t8
+    t12 = h & t11
+    f = t10 ^ t12
+    e = t12 ^ t14
+    key[4 * 19 + 8] = e
+    key[4 * 19 + 9] = f
+    key[4 * 19 + 10] = g
+    key[4 * 19 + 11] = h
+    a = key[4 * 20 + 8]
+    b = key[4 * 20 + 9]
+    c = key[4 * 20 + 10]
+    d = key[4 * 20 + 11]
+    t1 = (~c) % 0x100000000
+    t2 = b ^ c
+    t3 = b | t1
+    t4 = d ^ t3
+    t5 = a & t4
+    t7 = a ^ d
+    h = t2 ^ t5
+    t8 = b ^ t5
+    t9 = t2 | t8
+    t11 = d & t3
+    f = t7 ^ t9
+    t12 = t5 ^ f
+    t15 = t1 | t4
+    t13 = h & t12
+    g = t11 ^ t13
+    t16 = t12 ^ g
+    e = t15 ^ t16
+    key[4 * 20 + 8] = e
+    key[4 * 20 + 9] = f
+    key[4 * 20 + 10] = g
+    key[4 * 20 + 11] = h
+    a = key[4 * 21 + 8]
+    b = key[4 * 21 + 9]
+    c = key[4 * 21 + 10]
+    d = key[4 * 21 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ d
+    t3 = b ^ t2
+    t4 = t1 | t2
+    t5 = c ^ t4
+    f = b ^ t5
+    t13 = (~t5) % 0x100000000
+    t7 = t2 | f
+    t8 = d ^ t7
+    t9 = t5 & t8
+    g = t3 ^ t9
+    t11 = t5 ^ t8
+    e = g ^ t11
+    t14 = t3 & t11
+    h = t13 ^ t14
+    key[4 * 21 + 8] = e
+    key[4 * 21 + 9] = f
+    key[4 * 21 + 10] = g
+    key[4 * 21 + 11] = h
+    a = key[4 * 22 + 8]
+    b = key[4 * 22 + 9]
+    c = key[4 * 22 + 10]
+    d = key[4 * 22 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ b
+    t3 = a ^ d
+    t4 = c ^ t1
+    t5 = t2 | t3
+    e = t4 ^ t5
+    t7 = d & e
+    t8 = t2 ^ e
+    t10 = t1 | e
+    f = t7 ^ t8
+    t11 = t2 | t7
+    t12 = t3 ^ t10
+    t14 = b ^ t7
+    g = t11 ^ t12
+    t15 = f & t12
+    h = t14 ^ t15
+    key[4 * 22 + 8] = e
+    key[4 * 22 + 9] = f
+    key[4 * 22 + 10] = g
+    key[4 * 22 + 11] = h
+    a = key[4 * 23 + 8]
+    b = key[4 * 23 + 9]
+    c = key[4 * 23 + 10]
+    d = key[4 * 23 + 11]
+    t1 = a ^ d
+    t2 = d & t1
+    t3 = c ^ t2
+    t4 = b | t3
+    h = t1 ^ t4
+    t6 = (~b) % 0x100000000
+    t7 = t1 | t6
+    e = t3 ^ t7
+    t9 = a & e
+    t10 = t1 ^ t6
+    t11 = t4 & t10
+    g = t9 ^ t11
+    t13 = a ^ t3
+    t14 = t10 & g
+    f = t13 ^ t14
+    key[4 * 23 + 8] = e
+    key[4 * 23 + 9] = f
+    key[4 * 23 + 10] = g
+    key[4 * 23 + 11] = h
+    a = key[4 * 24 + 8]
+    b = key[4 * 24 + 9]
+    c = key[4 * 24 + 10]
+    d = key[4 * 24 + 11]
+    t1 = a ^ c
+    t2 = d ^ t1
+    t3 = a & t2
+    t4 = d ^ t3
+    t5 = b & t4
+    g = t2 ^ t5
+    t7 = a | g
+    t8 = b | d
+    t11 = a | d
+    t9 = t4 & t7
+    f = t8 ^ t9
+    t12 = b ^ t11
+    t13 = g ^ t9
+    t15 = t3 ^ t8
+    h = t12 ^ t13
+    t16 = c & t15
+    e = t12 ^ t16
+    key[4 * 24 + 8] = e
+    key[4 * 24 + 9] = f
+    key[4 * 24 + 10] = g
+    key[4 * 24 + 11] = h
+    a = key[4 * 25 + 8]
+    b = key[4 * 25 + 9]
+    c = key[4 * 25 + 10]
+    d = key[4 * 25 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ d
+    t3 = c & t1
+    t13 = d | t1
+    e = t2 ^ t3
+    t5 = c ^ t1
+    t6 = c ^ e
+    t7 = b & t6
+    t10 = e | t5
+    h = t5 ^ t7
+    t9 = d | t7
+    t11 = t9 & t10
+    t14 = t2 ^ h
+    g = a ^ t11
+    t15 = g ^ t13
+    f = t14 ^ t15
+    key[4 * 25 + 8] = e
+    key[4 * 25 + 9] = f
+    key[4 * 25 + 10] = g
+    key[4 * 25 + 11] = h
+    a = key[4 * 26 + 8]
+    b = key[4 * 26 + 9]
+    c = key[4 * 26 + 10]
+    d = key[4 * 26 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ t1
+    t3 = a | t2
+    t4 = d | t2
+    t5 = c ^ t3
+    g = d ^ t5
+    t7 = b ^ t4
+    t8 = t2 ^ g
+    t9 = t5 & t7
+    h = t8 ^ t9
+    t11 = t5 ^ t7
+    f = h ^ t11
+    t13 = t8 & t11
+    e = t5 ^ t13
+    key[4 * 26 + 8] = e
+    key[4 * 26 + 9] = f
+    key[4 * 26 + 10] = g
+    key[4 * 26 + 11] = h
+    a = key[4 * 27 + 8]
+    b = key[4 * 27 + 9]
+    c = key[4 * 27 + 10]
+    d = key[4 * 27 + 11]
+    t1 = a ^ d
+    t2 = a & d
+    t3 = c ^ t1
+    t6 = b & t1
+    t4 = b ^ t3
+    t10 = (~t3) % 0x100000000
+    h = t2 ^ t4
+    t7 = a ^ t6
+    t14 = (~t7) % 0x100000000
+    t8 = c | t7
+    t11 = t3 ^ t7
+    g = t4 ^ t8
+    t12 = h & t11
+    f = t10 ^ t12
+    e = t12 ^ t14
+    key[4 * 27 + 8] = e
+    key[4 * 27 + 9] = f
+    key[4 * 27 + 10] = g
+    key[4 * 27 + 11] = h
+    a = key[4 * 28 + 8]
+    b = key[4 * 28 + 9]
+    c = key[4 * 28 + 10]
+    d = key[4 * 28 + 11]
+    t1 = (~c) % 0x100000000
+    t2 = b ^ c
+    t3 = b | t1
+    t4 = d ^ t3
+    t5 = a & t4
+    t7 = a ^ d
+    h = t2 ^ t5
+    t8 = b ^ t5
+    t9 = t2 | t8
+    t11 = d & t3
+    f = t7 ^ t9
+    t12 = t5 ^ f
+    t15 = t1 | t4
+    t13 = h & t12
+    g = t11 ^ t13
+    t16 = t12 ^ g
+    e = t15 ^ t16
+    key[4 * 28 + 8] = e
+    key[4 * 28 + 9] = f
+    key[4 * 28 + 10] = g
+    key[4 * 28 + 11] = h
+    a = key[4 * 29 + 8]
+    b = key[4 * 29 + 9]
+    c = key[4 * 29 + 10]
+    d = key[4 * 29 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ d
+    t3 = b ^ t2
+    t4 = t1 | t2
+    t5 = c ^ t4
+    f = b ^ t5
+    t13 = (~t5) % 0x100000000
+    t7 = t2 | f
+    t8 = d ^ t7
+    t9 = t5 & t8
+    g = t3 ^ t9
+    t11 = t5 ^ t8
+    e = g ^ t11
+    t14 = t3 & t11
+    h = t13 ^ t14
+    key[4 * 29 + 8] = e
+    key[4 * 29 + 9] = f
+    key[4 * 29 + 10] = g
+    key[4 * 29 + 11] = h
+    a = key[4 * 30 + 8]
+    b = key[4 * 30 + 9]
+    c = key[4 * 30 + 10]
+    d = key[4 * 30 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ b
+    t3 = a ^ d
+    t4 = c ^ t1
+    t5 = t2 | t3
+    e = t4 ^ t5
+    t7 = d & e
+    t8 = t2 ^ e
+    t10 = t1 | e
+    f = t7 ^ t8
+    t11 = t2 | t7
+    t12 = t3 ^ t10
+    t14 = b ^ t7
+    g = t11 ^ t12
+    t15 = f & t12
+    h = t14 ^ t15
+    key[4 * 30 + 8] = e
+    key[4 * 30 + 9] = f
+    key[4 * 30 + 10] = g
+    key[4 * 30 + 11] = h
+    a = key[4 * 31 + 8]
+    b = key[4 * 31 + 9]
+    c = key[4 * 31 + 10]
+    d = key[4 * 31 + 11]
+    t1 = a ^ d
+    t2 = d & t1
+    t3 = c ^ t2
+    t4 = b | t3
+    h = t1 ^ t4
+    t6 = (~b) % 0x100000000
+    t7 = t1 | t6
+    e = t3 ^ t7
+    t9 = a & e
+    t10 = t1 ^ t6
+    t11 = t4 & t10
+    g = t9 ^ t11
+    t13 = a ^ t3
+    t14 = t10 & g
+    f = t13 ^ t14
+    key[4 * 31 + 8] = e
+    key[4 * 31 + 9] = f
+    key[4 * 31 + 10] = g
+    key[4 * 31 + 11] = h
+    a = key[4 * 32 + 8]
+    b = key[4 * 32 + 9]
+    c = key[4 * 32 + 10]
+    d = key[4 * 32 + 11]
+    t1 = a ^ c
+    t2 = d ^ t1
+    t3 = a & t2
+    t4 = d ^ t3
+    t5 = b & t4
+    g = t2 ^ t5
+    t7 = a | g
+    t8 = b | d
+    t11 = a | d
+    t9 = t4 & t7
+    f = t8 ^ t9
+    t12 = b ^ t11
+    t13 = g ^ t9
+    t15 = t3 ^ t8
+    h = t12 ^ t13
+    t16 = c & t15
+    e = t12 ^ t16
+    key[4 * 32 + 8] = e
+    key[4 * 32 + 9] = f
+    key[4 * 32 + 10] = g
+    key[4 * 32 + 11] = h
+
+
+def encrypt(key, in_blk):
+    # serpent_generate.py
+    a = in_blk[0]
+    b = in_blk[1]
+    c = in_blk[2]
+    d = in_blk[3]
+    if WORD_BIGENDIAN:
+        a = byteswap32(a)
+        b = byteswap32(b)
+        c = byteswap32(c)
+        d = byteswap32(d)
+    e = 0
+    f = 0
+    g = 0
+    h = 0
+    t1 = 0
+    t2 = 0
+    t3 = 0
+    t4 = 0
+    t5 = 0
+    t6 = 0
+    t7 = 0
+    t8 = 0
+    t9 = 0
+    t10 = 0
+    t11 = 0
+    t12 = 0
+    t13 = 0
+    t14 = 0
+    t15 = 0
+    t16 = 0
+    a ^= key[4 * 0 + 8]
+    b ^= key[4 * 0 + 9]
+    c ^= key[4 * 0 + 10]
+    d ^= key[4 * 0 + 11]
+    t1 = a ^ d
+    t2 = a & d
+    t3 = c ^ t1
+    t6 = b & t1
+    t4 = b ^ t3
+    t10 = (~t3) % 0x100000000
+    h = t2 ^ t4
+    t7 = a ^ t6
+    t14 = (~t7) % 0x100000000
+    t8 = c | t7
+    t11 = t3 ^ t7
+    g = t4 ^ t8
+    t12 = h & t11
+    f = t10 ^ t12
+    e = t12 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 1 + 8]
+    f ^= key[4 * 1 + 9]
+    g ^= key[4 * 1 + 10]
+    h ^= key[4 * 1 + 11]
+    t1 = (~e) % 0x100000000
+    t2 = f ^ t1
+    t3 = e | t2
+    t4 = h | t2
+    t5 = g ^ t3
+    c = h ^ t5
+    t7 = f ^ t4
+    t8 = t2 ^ c
+    t9 = t5 & t7
+    d = t8 ^ t9
+    t11 = t5 ^ t7
+    b = d ^ t11
+    t13 = t8 & t11
+    a = t5 ^ t13
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 2 + 8]
+    b ^= key[4 * 2 + 9]
+    c ^= key[4 * 2 + 10]
+    d ^= key[4 * 2 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ d
+    t3 = c & t1
+    t13 = d | t1
+    e = t2 ^ t3
+    t5 = c ^ t1
+    t6 = c ^ e
+    t7 = b & t6
+    t10 = e | t5
+    h = t5 ^ t7
+    t9 = d | t7
+    t11 = t9 & t10
+    t14 = t2 ^ h
+    g = a ^ t11
+    t15 = g ^ t13
+    f = t14 ^ t15
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 3 + 8]
+    f ^= key[4 * 3 + 9]
+    g ^= key[4 * 3 + 10]
+    h ^= key[4 * 3 + 11]
+    t1 = e ^ g
+    t2 = h ^ t1
+    t3 = e & t2
+    t4 = h ^ t3
+    t5 = f & t4
+    c = t2 ^ t5
+    t7 = e | c
+    t8 = f | h
+    t11 = e | h
+    t9 = t4 & t7
+    b = t8 ^ t9
+    t12 = f ^ t11
+    t13 = c ^ t9
+    t15 = t3 ^ t8
+    d = t12 ^ t13
+    t16 = g & t15
+    a = t12 ^ t16
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 4 + 8]
+    b ^= key[4 * 4 + 9]
+    c ^= key[4 * 4 + 10]
+    d ^= key[4 * 4 + 11]
+    t1 = a ^ d
+    t2 = d & t1
+    t3 = c ^ t2
+    t4 = b | t3
+    h = t1 ^ t4
+    t6 = (~b) % 0x100000000
+    t7 = t1 | t6
+    e = t3 ^ t7
+    t9 = a & e
+    t10 = t1 ^ t6
+    t11 = t4 & t10
+    g = t9 ^ t11
+    t13 = a ^ t3
+    t14 = t10 & g
+    f = t13 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 5 + 8]
+    f ^= key[4 * 5 + 9]
+    g ^= key[4 * 5 + 10]
+    h ^= key[4 * 5 + 11]
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = e ^ h
+    t4 = g ^ t1
+    t5 = t2 | t3
+    a = t4 ^ t5
+    t7 = h & a
+    t8 = t2 ^ a
+    t10 = t1 | a
+    b = t7 ^ t8
+    t11 = t2 | t7
+    t12 = t3 ^ t10
+    t14 = f ^ t7
+    c = t11 ^ t12
+    t15 = b & t12
+    d = t14 ^ t15
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 6 + 8]
+    b ^= key[4 * 6 + 9]
+    c ^= key[4 * 6 + 10]
+    d ^= key[4 * 6 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ d
+    t3 = b ^ t2
+    t4 = t1 | t2
+    t5 = c ^ t4
+    f = b ^ t5
+    t13 = (~t5) % 0x100000000
+    t7 = t2 | f
+    t8 = d ^ t7
+    t9 = t5 & t8
+    g = t3 ^ t9
+    t11 = t5 ^ t8
+    e = g ^ t11
+    t14 = t3 & t11
+    h = t13 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 7 + 8]
+    f ^= key[4 * 7 + 9]
+    g ^= key[4 * 7 + 10]
+    h ^= key[4 * 7 + 11]
+    t1 = (~g) % 0x100000000
+    t2 = f ^ g
+    t3 = f | t1
+    t4 = h ^ t3
+    t5 = e & t4
+    t7 = e ^ h
+    d = t2 ^ t5
+    t8 = f ^ t5
+    t9 = t2 | t8
+    t11 = h & t3
+    b = t7 ^ t9
+    t12 = t5 ^ b
+    t15 = t1 | t4
+    t13 = d & t12
+    c = t11 ^ t13
+    t16 = t12 ^ c
+    a = t15 ^ t16
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 8 + 8]
+    b ^= key[4 * 8 + 9]
+    c ^= key[4 * 8 + 10]
+    d ^= key[4 * 8 + 11]
+    t1 = a ^ d
+    t2 = a & d
+    t3 = c ^ t1
+    t6 = b & t1
+    t4 = b ^ t3
+    t10 = (~t3) % 0x100000000
+    h = t2 ^ t4
+    t7 = a ^ t6
+    t14 = (~t7) % 0x100000000
+    t8 = c | t7
+    t11 = t3 ^ t7
+    g = t4 ^ t8
+    t12 = h & t11
+    f = t10 ^ t12
+    e = t12 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 9 + 8]
+    f ^= key[4 * 9 + 9]
+    g ^= key[4 * 9 + 10]
+    h ^= key[4 * 9 + 11]
+    t1 = (~e) % 0x100000000
+    t2 = f ^ t1
+    t3 = e | t2
+    t4 = h | t2
+    t5 = g ^ t3
+    c = h ^ t5
+    t7 = f ^ t4
+    t8 = t2 ^ c
+    t9 = t5 & t7
+    d = t8 ^ t9
+    t11 = t5 ^ t7
+    b = d ^ t11
+    t13 = t8 & t11
+    a = t5 ^ t13
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 10 + 8]
+    b ^= key[4 * 10 + 9]
+    c ^= key[4 * 10 + 10]
+    d ^= key[4 * 10 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ d
+    t3 = c & t1
+    t13 = d | t1
+    e = t2 ^ t3
+    t5 = c ^ t1
+    t6 = c ^ e
+    t7 = b & t6
+    t10 = e | t5
+    h = t5 ^ t7
+    t9 = d | t7
+    t11 = t9 & t10
+    t14 = t2 ^ h
+    g = a ^ t11
+    t15 = g ^ t13
+    f = t14 ^ t15
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 11 + 8]
+    f ^= key[4 * 11 + 9]
+    g ^= key[4 * 11 + 10]
+    h ^= key[4 * 11 + 11]
+    t1 = e ^ g
+    t2 = h ^ t1
+    t3 = e & t2
+    t4 = h ^ t3
+    t5 = f & t4
+    c = t2 ^ t5
+    t7 = e | c
+    t8 = f | h
+    t11 = e | h
+    t9 = t4 & t7
+    b = t8 ^ t9
+    t12 = f ^ t11
+    t13 = c ^ t9
+    t15 = t3 ^ t8
+    d = t12 ^ t13
+    t16 = g & t15
+    a = t12 ^ t16
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 12 + 8]
+    b ^= key[4 * 12 + 9]
+    c ^= key[4 * 12 + 10]
+    d ^= key[4 * 12 + 11]
+    t1 = a ^ d
+    t2 = d & t1
+    t3 = c ^ t2
+    t4 = b | t3
+    h = t1 ^ t4
+    t6 = (~b) % 0x100000000
+    t7 = t1 | t6
+    e = t3 ^ t7
+    t9 = a & e
+    t10 = t1 ^ t6
+    t11 = t4 & t10
+    g = t9 ^ t11
+    t13 = a ^ t3
+    t14 = t10 & g
+    f = t13 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 13 + 8]
+    f ^= key[4 * 13 + 9]
+    g ^= key[4 * 13 + 10]
+    h ^= key[4 * 13 + 11]
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = e ^ h
+    t4 = g ^ t1
+    t5 = t2 | t3
+    a = t4 ^ t5
+    t7 = h & a
+    t8 = t2 ^ a
+    t10 = t1 | a
+    b = t7 ^ t8
+    t11 = t2 | t7
+    t12 = t3 ^ t10
+    t14 = f ^ t7
+    c = t11 ^ t12
+    t15 = b & t12
+    d = t14 ^ t15
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 14 + 8]
+    b ^= key[4 * 14 + 9]
+    c ^= key[4 * 14 + 10]
+    d ^= key[4 * 14 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ d
+    t3 = b ^ t2
+    t4 = t1 | t2
+    t5 = c ^ t4
+    f = b ^ t5
+    t13 = (~t5) % 0x100000000
+    t7 = t2 | f
+    t8 = d ^ t7
+    t9 = t5 & t8
+    g = t3 ^ t9
+    t11 = t5 ^ t8
+    e = g ^ t11
+    t14 = t3 & t11
+    h = t13 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 15 + 8]
+    f ^= key[4 * 15 + 9]
+    g ^= key[4 * 15 + 10]
+    h ^= key[4 * 15 + 11]
+    t1 = (~g) % 0x100000000
+    t2 = f ^ g
+    t3 = f | t1
+    t4 = h ^ t3
+    t5 = e & t4
+    t7 = e ^ h
+    d = t2 ^ t5
+    t8 = f ^ t5
+    t9 = t2 | t8
+    t11 = h & t3
+    b = t7 ^ t9
+    t12 = t5 ^ b
+    t15 = t1 | t4
+    t13 = d & t12
+    c = t11 ^ t13
+    t16 = t12 ^ c
+    a = t15 ^ t16
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 16 + 8]
+    b ^= key[4 * 16 + 9]
+    c ^= key[4 * 16 + 10]
+    d ^= key[4 * 16 + 11]
+    t1 = a ^ d
+    t2 = a & d
+    t3 = c ^ t1
+    t6 = b & t1
+    t4 = b ^ t3
+    t10 = (~t3) % 0x100000000
+    h = t2 ^ t4
+    t7 = a ^ t6
+    t14 = (~t7) % 0x100000000
+    t8 = c | t7
+    t11 = t3 ^ t7
+    g = t4 ^ t8
+    t12 = h & t11
+    f = t10 ^ t12
+    e = t12 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 17 + 8]
+    f ^= key[4 * 17 + 9]
+    g ^= key[4 * 17 + 10]
+    h ^= key[4 * 17 + 11]
+    t1 = (~e) % 0x100000000
+    t2 = f ^ t1
+    t3 = e | t2
+    t4 = h | t2
+    t5 = g ^ t3
+    c = h ^ t5
+    t7 = f ^ t4
+    t8 = t2 ^ c
+    t9 = t5 & t7
+    d = t8 ^ t9
+    t11 = t5 ^ t7
+    b = d ^ t11
+    t13 = t8 & t11
+    a = t5 ^ t13
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 18 + 8]
+    b ^= key[4 * 18 + 9]
+    c ^= key[4 * 18 + 10]
+    d ^= key[4 * 18 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ d
+    t3 = c & t1
+    t13 = d | t1
+    e = t2 ^ t3
+    t5 = c ^ t1
+    t6 = c ^ e
+    t7 = b & t6
+    t10 = e | t5
+    h = t5 ^ t7
+    t9 = d | t7
+    t11 = t9 & t10
+    t14 = t2 ^ h
+    g = a ^ t11
+    t15 = g ^ t13
+    f = t14 ^ t15
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 19 + 8]
+    f ^= key[4 * 19 + 9]
+    g ^= key[4 * 19 + 10]
+    h ^= key[4 * 19 + 11]
+    t1 = e ^ g
+    t2 = h ^ t1
+    t3 = e & t2
+    t4 = h ^ t3
+    t5 = f & t4
+    c = t2 ^ t5
+    t7 = e | c
+    t8 = f | h
+    t11 = e | h
+    t9 = t4 & t7
+    b = t8 ^ t9
+    t12 = f ^ t11
+    t13 = c ^ t9
+    t15 = t3 ^ t8
+    d = t12 ^ t13
+    t16 = g & t15
+    a = t12 ^ t16
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 20 + 8]
+    b ^= key[4 * 20 + 9]
+    c ^= key[4 * 20 + 10]
+    d ^= key[4 * 20 + 11]
+    t1 = a ^ d
+    t2 = d & t1
+    t3 = c ^ t2
+    t4 = b | t3
+    h = t1 ^ t4
+    t6 = (~b) % 0x100000000
+    t7 = t1 | t6
+    e = t3 ^ t7
+    t9 = a & e
+    t10 = t1 ^ t6
+    t11 = t4 & t10
+    g = t9 ^ t11
+    t13 = a ^ t3
+    t14 = t10 & g
+    f = t13 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 21 + 8]
+    f ^= key[4 * 21 + 9]
+    g ^= key[4 * 21 + 10]
+    h ^= key[4 * 21 + 11]
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = e ^ h
+    t4 = g ^ t1
+    t5 = t2 | t3
+    a = t4 ^ t5
+    t7 = h & a
+    t8 = t2 ^ a
+    t10 = t1 | a
+    b = t7 ^ t8
+    t11 = t2 | t7
+    t12 = t3 ^ t10
+    t14 = f ^ t7
+    c = t11 ^ t12
+    t15 = b & t12
+    d = t14 ^ t15
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 22 + 8]
+    b ^= key[4 * 22 + 9]
+    c ^= key[4 * 22 + 10]
+    d ^= key[4 * 22 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ d
+    t3 = b ^ t2
+    t4 = t1 | t2
+    t5 = c ^ t4
+    f = b ^ t5
+    t13 = (~t5) % 0x100000000
+    t7 = t2 | f
+    t8 = d ^ t7
+    t9 = t5 & t8
+    g = t3 ^ t9
+    t11 = t5 ^ t8
+    e = g ^ t11
+    t14 = t3 & t11
+    h = t13 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 23 + 8]
+    f ^= key[4 * 23 + 9]
+    g ^= key[4 * 23 + 10]
+    h ^= key[4 * 23 + 11]
+    t1 = (~g) % 0x100000000
+    t2 = f ^ g
+    t3 = f | t1
+    t4 = h ^ t3
+    t5 = e & t4
+    t7 = e ^ h
+    d = t2 ^ t5
+    t8 = f ^ t5
+    t9 = t2 | t8
+    t11 = h & t3
+    b = t7 ^ t9
+    t12 = t5 ^ b
+    t15 = t1 | t4
+    t13 = d & t12
+    c = t11 ^ t13
+    t16 = t12 ^ c
+    a = t15 ^ t16
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 24 + 8]
+    b ^= key[4 * 24 + 9]
+    c ^= key[4 * 24 + 10]
+    d ^= key[4 * 24 + 11]
+    t1 = a ^ d
+    t2 = a & d
+    t3 = c ^ t1
+    t6 = b & t1
+    t4 = b ^ t3
+    t10 = (~t3) % 0x100000000
+    h = t2 ^ t4
+    t7 = a ^ t6
+    t14 = (~t7) % 0x100000000
+    t8 = c | t7
+    t11 = t3 ^ t7
+    g = t4 ^ t8
+    t12 = h & t11
+    f = t10 ^ t12
+    e = t12 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 25 + 8]
+    f ^= key[4 * 25 + 9]
+    g ^= key[4 * 25 + 10]
+    h ^= key[4 * 25 + 11]
+    t1 = (~e) % 0x100000000
+    t2 = f ^ t1
+    t3 = e | t2
+    t4 = h | t2
+    t5 = g ^ t3
+    c = h ^ t5
+    t7 = f ^ t4
+    t8 = t2 ^ c
+    t9 = t5 & t7
+    d = t8 ^ t9
+    t11 = t5 ^ t7
+    b = d ^ t11
+    t13 = t8 & t11
+    a = t5 ^ t13
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 26 + 8]
+    b ^= key[4 * 26 + 9]
+    c ^= key[4 * 26 + 10]
+    d ^= key[4 * 26 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = b ^ d
+    t3 = c & t1
+    t13 = d | t1
+    e = t2 ^ t3
+    t5 = c ^ t1
+    t6 = c ^ e
+    t7 = b & t6
+    t10 = e | t5
+    h = t5 ^ t7
+    t9 = d | t7
+    t11 = t9 & t10
+    t14 = t2 ^ h
+    g = a ^ t11
+    t15 = g ^ t13
+    f = t14 ^ t15
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 27 + 8]
+    f ^= key[4 * 27 + 9]
+    g ^= key[4 * 27 + 10]
+    h ^= key[4 * 27 + 11]
+    t1 = e ^ g
+    t2 = h ^ t1
+    t3 = e & t2
+    t4 = h ^ t3
+    t5 = f & t4
+    c = t2 ^ t5
+    t7 = e | c
+    t8 = f | h
+    t11 = e | h
+    t9 = t4 & t7
+    b = t8 ^ t9
+    t12 = f ^ t11
+    t13 = c ^ t9
+    t15 = t3 ^ t8
+    d = t12 ^ t13
+    t16 = g & t15
+    a = t12 ^ t16
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 28 + 8]
+    b ^= key[4 * 28 + 9]
+    c ^= key[4 * 28 + 10]
+    d ^= key[4 * 28 + 11]
+    t1 = a ^ d
+    t2 = d & t1
+    t3 = c ^ t2
+    t4 = b | t3
+    h = t1 ^ t4
+    t6 = (~b) % 0x100000000
+    t7 = t1 | t6
+    e = t3 ^ t7
+    t9 = a & e
+    t10 = t1 ^ t6
+    t11 = t4 & t10
+    g = t9 ^ t11
+    t13 = a ^ t3
+    t14 = t10 & g
+    f = t13 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 29 + 8]
+    f ^= key[4 * 29 + 9]
+    g ^= key[4 * 29 + 10]
+    h ^= key[4 * 29 + 11]
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = e ^ h
+    t4 = g ^ t1
+    t5 = t2 | t3
+    a = t4 ^ t5
+    t7 = h & a
+    t8 = t2 ^ a
+    t10 = t1 | a
+    b = t7 ^ t8
+    t11 = t2 | t7
+    t12 = t3 ^ t10
+    t14 = f ^ t7
+    c = t11 ^ t12
+    t15 = b & t12
+    d = t14 ^ t15
+    a = rotl32(a, 13)
+    c = rotl32(c, 3)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    d = rotl32(d, 7)
+    b = rotl32(b, 1)
+    a ^= b ^ d
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a = rotl32(a, 5)
+    c = rotl32(c, 22)
+    a ^= key[4 * 30 + 8]
+    b ^= key[4 * 30 + 9]
+    c ^= key[4 * 30 + 10]
+    d ^= key[4 * 30 + 11]
+    t1 = (~a) % 0x100000000
+    t2 = a ^ d
+    t3 = b ^ t2
+    t4 = t1 | t2
+    t5 = c ^ t4
+    f = b ^ t5
+    t13 = (~t5) % 0x100000000
+    t7 = t2 | f
+    t8 = d ^ t7
+    t9 = t5 & t8
+    g = t3 ^ t9
+    t11 = t5 ^ t8
+    e = g ^ t11
+    t14 = t3 & t11
+    h = t13 ^ t14
+    e = rotl32(e, 13)
+    g = rotl32(g, 3)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    h = rotl32(h, 7)
+    f = rotl32(f, 1)
+    e ^= f ^ h
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e = rotl32(e, 5)
+    g = rotl32(g, 22)
+    e ^= key[4 * 31 + 8]
+    f ^= key[4 * 31 + 9]
+    g ^= key[4 * 31 + 10]
+    h ^= key[4 * 31 + 11]
+    t1 = (~g) % 0x100000000
+    t2 = f ^ g
+    t3 = f | t1
+    t4 = h ^ t3
+    t5 = e & t4
+    t7 = e ^ h
+    d = t2 ^ t5
+    t8 = f ^ t5
+    t9 = t2 | t8
+    t11 = h & t3
+    b = t7 ^ t9
+    t12 = t5 ^ b
+    t15 = t1 | t4
+    t13 = d & t12
+    c = t11 ^ t13
+    t16 = t12 ^ c
+    a = t15 ^ t16
+    a ^= key[4 * 32 + 8]
+    b ^= key[4 * 32 + 9]
+    c ^= key[4 * 32 + 10]
+    d ^= key[4 * 32 + 11]
+    if WORD_BIGENDIAN:
+        a = byteswap32(a)
+        b = byteswap32(b)
+        c = byteswap32(c)
+        d = byteswap32(d)
+    in_blk[0] = a
+    in_blk[1] = b
+    in_blk[2] = c
+    in_blk[3] = d
+
+
+def decrypt(key, in_blk):
+    # serpent_generate.py
+    a = in_blk[0]
+    b = in_blk[1]
+    c = in_blk[2]
+    d = in_blk[3]
+    if WORD_BIGENDIAN:
+        a = byteswap32(a)
+        b = byteswap32(b)
+        c = byteswap32(c)
+        d = byteswap32(d)
+    e = 0
+    f = 0
+    g = 0
+    h = 0
+    t1 = 0
+    t2 = 0
+    t3 = 0
+    t4 = 0
+    t5 = 0
+    t6 = 0
+    t7 = 0
+    t8 = 0
+    t9 = 0
+    t10 = 0
+    t11 = 0
+    t12 = 0
+    t13 = 0
+    t14 = 0
+    t15 = 0
+    t16 = 0
+    a ^= key[4 * 32 + 8]
+    b ^= key[4 * 32 + 9]
+    c ^= key[4 * 32 + 10]
+    d ^= key[4 * 32 + 11]
+    t1 = a & b
+    t2 = a | b
+    t3 = c | t1
+    t4 = d & t2
+    h = t3 ^ t4
+    t6 = (~d) % 0x100000000
+    t7 = b ^ t4
+    t8 = h ^ t6
+    t11 = c ^ t7
+    t9 = t7 | t8
+    f = a ^ t9
+    t12 = d | f
+    e = t11 ^ t12
+    t14 = a & h
+    t15 = t3 ^ f
+    t16 = e ^ t14
+    g = t15 ^ t16
+    e ^= key[4 * 31 + 8]
+    f ^= key[4 * 31 + 9]
+    g ^= key[4 * 31 + 10]
+    h ^= key[4 * 31 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = g ^ t2
+    t4 = g | t1
+    t5 = h ^ t4
+    t13 = h & t1
+    b = t3 ^ t5
+    t7 = t3 & t5
+    t8 = t2 ^ t7
+    t9 = f | t8
+    d = t5 ^ t9
+    t11 = f | d
+    a = t8 ^ t11
+    t14 = t3 ^ t11
+    c = t13 ^ t14
+    a ^= key[4 * 30 + 8]
+    b ^= key[4 * 30 + 9]
+    c ^= key[4 * 30 + 10]
+    d ^= key[4 * 30 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = (~c) % 0x100000000
+    t2 = b & t1
+    t3 = d ^ t2
+    t4 = a & t3
+    t5 = b ^ t1
+    h = t4 ^ t5
+    t7 = b | h
+    t8 = a & t7
+    f = t3 ^ t8
+    t10 = a | d
+    t11 = t1 ^ t7
+    e = t10 ^ t11
+    t13 = a ^ c
+    t14 = b & t10
+    t15 = t4 | t13
+    g = t14 ^ t15
+    e ^= key[4 * 29 + 8]
+    f ^= key[4 * 29 + 9]
+    g ^= key[4 * 29 + 10]
+    h ^= key[4 * 29 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = g ^ h
+    t2 = g | h
+    t3 = f ^ t2
+    t4 = e & t3
+    b = t1 ^ t4
+    t6 = e ^ h
+    t7 = f | h
+    t8 = t6 & t7
+    d = t3 ^ t8
+    t10 = (~e) % 0x100000000
+    t11 = g ^ d
+    t12 = t10 | t11
+    a = t3 ^ t12
+    t14 = g | t4
+    t15 = t7 ^ t14
+    t16 = d | t10
+    c = t15 ^ t16
+    a ^= key[4 * 28 + 8]
+    b ^= key[4 * 28 + 9]
+    c ^= key[4 * 28 + 10]
+    d ^= key[4 * 28 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = b ^ c
+    t2 = b | c
+    t3 = a ^ c
+    t7 = a ^ d
+    t4 = t2 ^ t3
+    t5 = d | t4
+    t9 = t2 ^ t7
+    e = t1 ^ t5
+    t8 = t1 | t5
+    t11 = a & t4
+    g = t8 ^ t9
+    t12 = e | t9
+    f = t11 ^ t12
+    t14 = a & g
+    t15 = t2 ^ t14
+    t16 = e & t15
+    h = t4 ^ t16
+    e ^= key[4 * 27 + 8]
+    f ^= key[4 * 27 + 9]
+    g ^= key[4 * 27 + 10]
+    h ^= key[4 * 27 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = f ^ h
+    t2 = (~t1) % 0x100000000
+    t3 = e ^ g
+    t4 = g ^ t1
+    t7 = e | t2
+    t5 = f & t4
+    t8 = h ^ t7
+    t11 = (~t4) % 0x100000000
+    a = t3 ^ t5
+    t9 = t3 | t8
+    t14 = h & t11
+    d = t1 ^ t9
+    t12 = a | d
+    b = t11 ^ t12
+    t15 = t3 ^ t12
+    c = t14 ^ t15
+    a ^= key[4 * 26 + 8]
+    b ^= key[4 * 26 + 9]
+    c ^= key[4 * 26 + 10]
+    d ^= key[4 * 26 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = a ^ d
+    t2 = a & b
+    t3 = b ^ c
+    t4 = a ^ t3
+    t5 = b | d
+    t7 = c | t1
+    h = t4 ^ t5
+    t8 = b ^ t7
+    t11 = (~t2) % 0x100000000
+    t9 = t4 & t8
+    f = t1 ^ t9
+    t13 = t9 ^ t11
+    t12 = h & f
+    g = t12 ^ t13
+    t15 = a & d
+    t16 = c ^ t13
+    e = t15 ^ t16
+    e ^= key[4 * 25 + 8]
+    f ^= key[4 * 25 + 9]
+    g ^= key[4 * 25 + 10]
+    h ^= key[4 * 25 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = t1 | t2
+    t4 = h ^ t3
+    t7 = h & t2
+    t5 = g ^ t4
+    t8 = t1 ^ t7
+    c = t2 ^ t5
+    t11 = e & t4
+    t9 = c & t8
+    t14 = t5 ^ t8
+    b = t4 ^ t9
+    t12 = t5 | b
+    d = t11 ^ t12
+    a = d ^ t14
+    a ^= key[4 * 24 + 8]
+    b ^= key[4 * 24 + 9]
+    c ^= key[4 * 24 + 10]
+    d ^= key[4 * 24 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = a & b
+    t2 = a | b
+    t3 = c | t1
+    t4 = d & t2
+    h = t3 ^ t4
+    t6 = (~d) % 0x100000000
+    t7 = b ^ t4
+    t8 = h ^ t6
+    t11 = c ^ t7
+    t9 = t7 | t8
+    f = a ^ t9
+    t12 = d | f
+    e = t11 ^ t12
+    t14 = a & h
+    t15 = t3 ^ f
+    t16 = e ^ t14
+    g = t15 ^ t16
+    e ^= key[4 * 23 + 8]
+    f ^= key[4 * 23 + 9]
+    g ^= key[4 * 23 + 10]
+    h ^= key[4 * 23 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = g ^ t2
+    t4 = g | t1
+    t5 = h ^ t4
+    t13 = h & t1
+    b = t3 ^ t5
+    t7 = t3 & t5
+    t8 = t2 ^ t7
+    t9 = f | t8
+    d = t5 ^ t9
+    t11 = f | d
+    a = t8 ^ t11
+    t14 = t3 ^ t11
+    c = t13 ^ t14
+    a ^= key[4 * 22 + 8]
+    b ^= key[4 * 22 + 9]
+    c ^= key[4 * 22 + 10]
+    d ^= key[4 * 22 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = (~c) % 0x100000000
+    t2 = b & t1
+    t3 = d ^ t2
+    t4 = a & t3
+    t5 = b ^ t1
+    h = t4 ^ t5
+    t7 = b | h
+    t8 = a & t7
+    f = t3 ^ t8
+    t10 = a | d
+    t11 = t1 ^ t7
+    e = t10 ^ t11
+    t13 = a ^ c
+    t14 = b & t10
+    t15 = t4 | t13
+    g = t14 ^ t15
+    e ^= key[4 * 21 + 8]
+    f ^= key[4 * 21 + 9]
+    g ^= key[4 * 21 + 10]
+    h ^= key[4 * 21 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = g ^ h
+    t2 = g | h
+    t3 = f ^ t2
+    t4 = e & t3
+    b = t1 ^ t4
+    t6 = e ^ h
+    t7 = f | h
+    t8 = t6 & t7
+    d = t3 ^ t8
+    t10 = (~e) % 0x100000000
+    t11 = g ^ d
+    t12 = t10 | t11
+    a = t3 ^ t12
+    t14 = g | t4
+    t15 = t7 ^ t14
+    t16 = d | t10
+    c = t15 ^ t16
+    a ^= key[4 * 20 + 8]
+    b ^= key[4 * 20 + 9]
+    c ^= key[4 * 20 + 10]
+    d ^= key[4 * 20 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = b ^ c
+    t2 = b | c
+    t3 = a ^ c
+    t7 = a ^ d
+    t4 = t2 ^ t3
+    t5 = d | t4
+    t9 = t2 ^ t7
+    e = t1 ^ t5
+    t8 = t1 | t5
+    t11 = a & t4
+    g = t8 ^ t9
+    t12 = e | t9
+    f = t11 ^ t12
+    t14 = a & g
+    t15 = t2 ^ t14
+    t16 = e & t15
+    h = t4 ^ t16
+    e ^= key[4 * 19 + 8]
+    f ^= key[4 * 19 + 9]
+    g ^= key[4 * 19 + 10]
+    h ^= key[4 * 19 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = f ^ h
+    t2 = (~t1) % 0x100000000
+    t3 = e ^ g
+    t4 = g ^ t1
+    t7 = e | t2
+    t5 = f & t4
+    t8 = h ^ t7
+    t11 = (~t4) % 0x100000000
+    a = t3 ^ t5
+    t9 = t3 | t8
+    t14 = h & t11
+    d = t1 ^ t9
+    t12 = a | d
+    b = t11 ^ t12
+    t15 = t3 ^ t12
+    c = t14 ^ t15
+    a ^= key[4 * 18 + 8]
+    b ^= key[4 * 18 + 9]
+    c ^= key[4 * 18 + 10]
+    d ^= key[4 * 18 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = a ^ d
+    t2 = a & b
+    t3 = b ^ c
+    t4 = a ^ t3
+    t5 = b | d
+    t7 = c | t1
+    h = t4 ^ t5
+    t8 = b ^ t7
+    t11 = (~t2) % 0x100000000
+    t9 = t4 & t8
+    f = t1 ^ t9
+    t13 = t9 ^ t11
+    t12 = h & f
+    g = t12 ^ t13
+    t15 = a & d
+    t16 = c ^ t13
+    e = t15 ^ t16
+    e ^= key[4 * 17 + 8]
+    f ^= key[4 * 17 + 9]
+    g ^= key[4 * 17 + 10]
+    h ^= key[4 * 17 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = t1 | t2
+    t4 = h ^ t3
+    t7 = h & t2
+    t5 = g ^ t4
+    t8 = t1 ^ t7
+    c = t2 ^ t5
+    t11 = e & t4
+    t9 = c & t8
+    t14 = t5 ^ t8
+    b = t4 ^ t9
+    t12 = t5 | b
+    d = t11 ^ t12
+    a = d ^ t14
+    a ^= key[4 * 16 + 8]
+    b ^= key[4 * 16 + 9]
+    c ^= key[4 * 16 + 10]
+    d ^= key[4 * 16 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = a & b
+    t2 = a | b
+    t3 = c | t1
+    t4 = d & t2
+    h = t3 ^ t4
+    t6 = (~d) % 0x100000000
+    t7 = b ^ t4
+    t8 = h ^ t6
+    t11 = c ^ t7
+    t9 = t7 | t8
+    f = a ^ t9
+    t12 = d | f
+    e = t11 ^ t12
+    t14 = a & h
+    t15 = t3 ^ f
+    t16 = e ^ t14
+    g = t15 ^ t16
+    e ^= key[4 * 15 + 8]
+    f ^= key[4 * 15 + 9]
+    g ^= key[4 * 15 + 10]
+    h ^= key[4 * 15 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = g ^ t2
+    t4 = g | t1
+    t5 = h ^ t4
+    t13 = h & t1
+    b = t3 ^ t5
+    t7 = t3 & t5
+    t8 = t2 ^ t7
+    t9 = f | t8
+    d = t5 ^ t9
+    t11 = f | d
+    a = t8 ^ t11
+    t14 = t3 ^ t11
+    c = t13 ^ t14
+    a ^= key[4 * 14 + 8]
+    b ^= key[4 * 14 + 9]
+    c ^= key[4 * 14 + 10]
+    d ^= key[4 * 14 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = (~c) % 0x100000000
+    t2 = b & t1
+    t3 = d ^ t2
+    t4 = a & t3
+    t5 = b ^ t1
+    h = t4 ^ t5
+    t7 = b | h
+    t8 = a & t7
+    f = t3 ^ t8
+    t10 = a | d
+    t11 = t1 ^ t7
+    e = t10 ^ t11
+    t13 = a ^ c
+    t14 = b & t10
+    t15 = t4 | t13
+    g = t14 ^ t15
+    e ^= key[4 * 13 + 8]
+    f ^= key[4 * 13 + 9]
+    g ^= key[4 * 13 + 10]
+    h ^= key[4 * 13 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = g ^ h
+    t2 = g | h
+    t3 = f ^ t2
+    t4 = e & t3
+    b = t1 ^ t4
+    t6 = e ^ h
+    t7 = f | h
+    t8 = t6 & t7
+    d = t3 ^ t8
+    t10 = (~e) % 0x100000000
+    t11 = g ^ d
+    t12 = t10 | t11
+    a = t3 ^ t12
+    t14 = g | t4
+    t15 = t7 ^ t14
+    t16 = d | t10
+    c = t15 ^ t16
+    a ^= key[4 * 12 + 8]
+    b ^= key[4 * 12 + 9]
+    c ^= key[4 * 12 + 10]
+    d ^= key[4 * 12 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = b ^ c
+    t2 = b | c
+    t3 = a ^ c
+    t7 = a ^ d
+    t4 = t2 ^ t3
+    t5 = d | t4
+    t9 = t2 ^ t7
+    e = t1 ^ t5
+    t8 = t1 | t5
+    t11 = a & t4
+    g = t8 ^ t9
+    t12 = e | t9
+    f = t11 ^ t12
+    t14 = a & g
+    t15 = t2 ^ t14
+    t16 = e & t15
+    h = t4 ^ t16
+    e ^= key[4 * 11 + 8]
+    f ^= key[4 * 11 + 9]
+    g ^= key[4 * 11 + 10]
+    h ^= key[4 * 11 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = f ^ h
+    t2 = (~t1) % 0x100000000
+    t3 = e ^ g
+    t4 = g ^ t1
+    t7 = e | t2
+    t5 = f & t4
+    t8 = h ^ t7
+    t11 = (~t4) % 0x100000000
+    a = t3 ^ t5
+    t9 = t3 | t8
+    t14 = h & t11
+    d = t1 ^ t9
+    t12 = a | d
+    b = t11 ^ t12
+    t15 = t3 ^ t12
+    c = t14 ^ t15
+    a ^= key[4 * 10 + 8]
+    b ^= key[4 * 10 + 9]
+    c ^= key[4 * 10 + 10]
+    d ^= key[4 * 10 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = a ^ d
+    t2 = a & b
+    t3 = b ^ c
+    t4 = a ^ t3
+    t5 = b | d
+    t7 = c | t1
+    h = t4 ^ t5
+    t8 = b ^ t7
+    t11 = (~t2) % 0x100000000
+    t9 = t4 & t8
+    f = t1 ^ t9
+    t13 = t9 ^ t11
+    t12 = h & f
+    g = t12 ^ t13
+    t15 = a & d
+    t16 = c ^ t13
+    e = t15 ^ t16
+    e ^= key[4 * 9 + 8]
+    f ^= key[4 * 9 + 9]
+    g ^= key[4 * 9 + 10]
+    h ^= key[4 * 9 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = t1 | t2
+    t4 = h ^ t3
+    t7 = h & t2
+    t5 = g ^ t4
+    t8 = t1 ^ t7
+    c = t2 ^ t5
+    t11 = e & t4
+    t9 = c & t8
+    t14 = t5 ^ t8
+    b = t4 ^ t9
+    t12 = t5 | b
+    d = t11 ^ t12
+    a = d ^ t14
+    a ^= key[4 * 8 + 8]
+    b ^= key[4 * 8 + 9]
+    c ^= key[4 * 8 + 10]
+    d ^= key[4 * 8 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = a & b
+    t2 = a | b
+    t3 = c | t1
+    t4 = d & t2
+    h = t3 ^ t4
+    t6 = (~d) % 0x100000000
+    t7 = b ^ t4
+    t8 = h ^ t6
+    t11 = c ^ t7
+    t9 = t7 | t8
+    f = a ^ t9
+    t12 = d | f
+    e = t11 ^ t12
+    t14 = a & h
+    t15 = t3 ^ f
+    t16 = e ^ t14
+    g = t15 ^ t16
+    e ^= key[4 * 7 + 8]
+    f ^= key[4 * 7 + 9]
+    g ^= key[4 * 7 + 10]
+    h ^= key[4 * 7 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = g ^ t2
+    t4 = g | t1
+    t5 = h ^ t4
+    t13 = h & t1
+    b = t3 ^ t5
+    t7 = t3 & t5
+    t8 = t2 ^ t7
+    t9 = f | t8
+    d = t5 ^ t9
+    t11 = f | d
+    a = t8 ^ t11
+    t14 = t3 ^ t11
+    c = t13 ^ t14
+    a ^= key[4 * 6 + 8]
+    b ^= key[4 * 6 + 9]
+    c ^= key[4 * 6 + 10]
+    d ^= key[4 * 6 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = (~c) % 0x100000000
+    t2 = b & t1
+    t3 = d ^ t2
+    t4 = a & t3
+    t5 = b ^ t1
+    h = t4 ^ t5
+    t7 = b | h
+    t8 = a & t7
+    f = t3 ^ t8
+    t10 = a | d
+    t11 = t1 ^ t7
+    e = t10 ^ t11
+    t13 = a ^ c
+    t14 = b & t10
+    t15 = t4 | t13
+    g = t14 ^ t15
+    e ^= key[4 * 5 + 8]
+    f ^= key[4 * 5 + 9]
+    g ^= key[4 * 5 + 10]
+    h ^= key[4 * 5 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = g ^ h
+    t2 = g | h
+    t3 = f ^ t2
+    t4 = e & t3
+    b = t1 ^ t4
+    t6 = e ^ h
+    t7 = f | h
+    t8 = t6 & t7
+    d = t3 ^ t8
+    t10 = (~e) % 0x100000000
+    t11 = g ^ d
+    t12 = t10 | t11
+    a = t3 ^ t12
+    t14 = g | t4
+    t15 = t7 ^ t14
+    t16 = d | t10
+    c = t15 ^ t16
+    a ^= key[4 * 4 + 8]
+    b ^= key[4 * 4 + 9]
+    c ^= key[4 * 4 + 10]
+    d ^= key[4 * 4 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = b ^ c
+    t2 = b | c
+    t3 = a ^ c
+    t7 = a ^ d
+    t4 = t2 ^ t3
+    t5 = d | t4
+    t9 = t2 ^ t7
+    e = t1 ^ t5
+    t8 = t1 | t5
+    t11 = a & t4
+    g = t8 ^ t9
+    t12 = e | t9
+    f = t11 ^ t12
+    t14 = a & g
+    t15 = t2 ^ t14
+    t16 = e & t15
+    h = t4 ^ t16
+    e ^= key[4 * 3 + 8]
+    f ^= key[4 * 3 + 9]
+    g ^= key[4 * 3 + 10]
+    h ^= key[4 * 3 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = f ^ h
+    t2 = (~t1) % 0x100000000
+    t3 = e ^ g
+    t4 = g ^ t1
+    t7 = e | t2
+    t5 = f & t4
+    t8 = h ^ t7
+    t11 = (~t4) % 0x100000000
+    a = t3 ^ t5
+    t9 = t3 | t8
+    t14 = h & t11
+    d = t1 ^ t9
+    t12 = a | d
+    b = t11 ^ t12
+    t15 = t3 ^ t12
+    c = t14 ^ t15
+    a ^= key[4 * 2 + 8]
+    b ^= key[4 * 2 + 9]
+    c ^= key[4 * 2 + 10]
+    d ^= key[4 * 2 + 11]
+    c = rotr32(c, 22)
+    a = rotr32(a, 5)
+    c ^= d ^ ((b << 7) & 0xFFFFFFFF)
+    a ^= b ^ d
+    d = rotr32(d, 7)
+    b = rotr32(b, 1)
+    d ^= c ^ ((a << 3) & 0xFFFFFFFF)
+    b ^= a ^ c
+    c = rotr32(c, 3)
+    a = rotr32(a, 13)
+    t1 = a ^ d
+    t2 = a & b
+    t3 = b ^ c
+    t4 = a ^ t3
+    t5 = b | d
+    t7 = c | t1
+    h = t4 ^ t5
+    t8 = b ^ t7
+    t11 = (~t2) % 0x100000000
+    t9 = t4 & t8
+    f = t1 ^ t9
+    t13 = t9 ^ t11
+    t12 = h & f
+    g = t12 ^ t13
+    t15 = a & d
+    t16 = c ^ t13
+    e = t15 ^ t16
+    e ^= key[4 * 1 + 8]
+    f ^= key[4 * 1 + 9]
+    g ^= key[4 * 1 + 10]
+    h ^= key[4 * 1 + 11]
+    g = rotr32(g, 22)
+    e = rotr32(e, 5)
+    g ^= h ^ ((f << 7) & 0xFFFFFFFF)
+    e ^= f ^ h
+    h = rotr32(h, 7)
+    f = rotr32(f, 1)
+    h ^= g ^ ((e << 3) & 0xFFFFFFFF)
+    f ^= e ^ g
+    g = rotr32(g, 3)
+    e = rotr32(e, 13)
+    t1 = (~e) % 0x100000000
+    t2 = e ^ f
+    t3 = t1 | t2
+    t4 = h ^ t3
+    t7 = h & t2
+    t5 = g ^ t4
+    t8 = t1 ^ t7
+    c = t2 ^ t5
+    t11 = e & t4
+    t9 = c & t8
+    t14 = t5 ^ t8
+    b = t4 ^ t9
+    t12 = t5 | b
+    d = t11 ^ t12
+    a = d ^ t14
+    a ^= key[4 * 0 + 8]
+    b ^= key[4 * 0 + 9]
+    c ^= key[4 * 0 + 10]
+    d ^= key[4 * 0 + 11]
+    if WORD_BIGENDIAN:
+        a = byteswap32(a)
+        b = byteswap32(b)
+        c = byteswap32(c)
+        d = byteswap32(d)
+    in_blk[0] = a
+    in_blk[1] = b
+    in_blk[2] = c
+    in_blk[3] = d
+
+
+__testkey = b'\x00\x01\x02\x03\x04\x05\x06\x07\x08\x09\x0a\x0b\x0c\x0d\x0e\x0f'\
+            b'\x10\x11\x12\x13\x14\x15\x16\x17\x18\x19\x1a\x1b\x1c\x1d\x1e\x1f'
+__testdat = b'\x00\x01\x02\x03\x04\x05\x06\x07\x08\t\n\x0b\x0c\r\x0e\x0f'
+assert b'\xde&\x9f\xf83\xe42\xb8[.\x88\xd2p\x1c\xe7\\' == Serpent(__testkey).encrypt(__testdat)
+assert __testdat == Serpent(__testkey).decrypt(b'\xde&\x9f\xf83\xe42\xb8[.\x88\xd2p\x1c\xe7\\')
+
+
+# CBC Encrypt - Jason Reaves
+def serpent_cbc_encrypt(key, data, iv=b'\x00' * 16):
+    out = ""
+    last = iv
+    for i in range(len(data) // 16):
+        temp = data[i * 16:(i + 1) * 16]
+        to_encode = b""
+        for j in range(4):
+            temp1 = struct.unpack_from('<I', temp[j * 4:])[0]
+            temp2 = struct.unpack_from('<I', last[j * 4:])[0]
+            to_encode += struct.pack('<I', ((temp1 ^ temp2) & 0xffffffff))
+        last = Serpent(key).encrypt(to_encode)
+        out += last
+    # print(binascii.hexlify(Serpent(key).encrypt(data)))
+    return out
+
+
+# CBC Decrypt - Jason Reaves
+def serpent_cbc_decrypt(key, data, iv=b'\x00' * 16):
+    out2 = b""
+    last = iv
+    for i in range(len(data) // 16):
+        temp = Serpent(key).decrypt(data[i * 16:(i + 1) * 16])
+        to_decode = b""
+        for j in range(4):
+            temp1 = struct.unpack_from('<I', temp[j * 4:])[0]
+            temp2 = struct.unpack_from('<I', last[j * 4:])[0]
+            to_decode += struct.pack('<I', ((temp1 ^ temp2) & 0xffffffff))
+        out2 += to_decode
+        last = data[i * 16:(i + 1) * 16]
+    return out2
diff --git a/malduck/crypto/serpent.py b/malduck/crypto/serpent.py
new file mode 100644
index 0000000..0ec50eb
--- /dev/null
+++ b/malduck/crypto/serpent.py
@@ -0,0 +1,13 @@
+from .components.pyserpent import serpent_cbc_encrypt, serpent_cbc_decrypt
+
+
+class Serpent(object):
+    def __init__(self, key, iv=None):
+        self.key = key
+        self.iv = iv or b"\x00" * 16
+
+    def decrypt(self, data):
+        return serpent_cbc_decrypt(self.key, data, iv=self.iv)
+
+    def encrypt(self, data):
+        return serpent_cbc_encrypt(self.key, data, iv=self.iv)
diff --git a/malduck/short.py b/malduck/short.py
index 25d6a6c..f978259 100644
--- a/malduck/short.py
+++ b/malduck/short.py
@@ -7,6 +7,7 @@
 from .crypto.aes import AES
 from .crypto.blowfish import Blowfish
 from .crypto.des3 import DES3
+from .crypto.serpent import Serpent
 from .crypto.rabbit import Rabbit
 from .crypto.rc import RC4
 from .crypto.rsa import RSA
@@ -97,6 +98,10 @@ def rabbit(key, iv, data):
     __call__ = rabbit
 
 
+def serpent(key, data, iv=None):
+    return Serpent(key, iv).decrypt(data)
+
+
 blowfish = blowfish_()
 rc4 = rc4_()
 rabbit = rabbit_()
diff --git a/tests/test_crypto.py b/tests/test_crypto.py
index b20c58c..65a1054 100644
--- a/tests/test_crypto.py
+++ b/tests/test_crypto.py
@@ -2,7 +2,7 @@
 # This file is part of Roach - https://github.com/jbremer/roach.
 # See the file 'docs/LICENSE.txt' for copying permission.
 
-from malduck import aes, blowfish, des3, rc4, rsa, xor, base64, unhex, rabbit, p8
+from malduck import aes, blowfish, des3, rc4, rsa, xor, base64, unhex, rabbit, p8, serpent
 
 
 def test_aes():
@@ -217,3 +217,9 @@ def test_rabbit():
     assert rabbit(key1, iv1, b"\x00"*48) == out4
     assert rabbit(key1, iv2, b"\x00"*48) == out5
     assert rabbit(key1, iv3, b"\x00"*48) == out6
+
+
+def test_serpent():
+    payload = unhex("8a516cb035540b5854a18eeccc40299d")
+    key = b"0123456789ABCDEF"
+    assert serpent(key, payload) == b'\xf4\xee\xd1\xec\x04\x01\x00\x00\x00\x04\x00\x00\xc3\xdc\x07\xd4'