serpent.c

Serpent加密算法源代码

/* 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 ([e-mail]) 14th January 1999     */

/* Timing data for Serpent (serpent.c)

Core timing without I/O endian conversion:

128 bit key:
Key Setup:    2402 cycles
Encrypt:       952 cycles =    26.9 mbits/sec
Decrypt:       914 cycles =    28.0 mbits/sec
Mean:          933 cycles =    27.4 mbits/sec

192 bit key:
Key Setup:    2449 cycles
Encrypt:       952 cycles =    26.9 mbits/sec
Decrypt:       914 cycles =    28.0 mbits/sec
Mean:          933 cycles =    27.4 mbits/sec

256 bit key:
Key Setup:    2349 cycles
Encrypt:       952 cycles =    26.9 mbits/sec
Decrypt:       914 cycles =    28.0 mbits/sec
Mean:          933 cycles =    27.4 mbits/sec

Full timing with I/O endian conversion:

128 bit key:
Key Setup:    2415 cycles
Encrypt:       985 cycles =    26.0 mbits/sec
Decrypt:       954 cycles =    26.8 mbits/sec
Mean:          970 cycles =    26.4 mbits/sec

192 bit key:
Key Setup:    2438 cycles
Encrypt:       985 cycles =    26.0 mbits/sec
Decrypt:       954 cycles =    26.8 mbits/sec
Mean:          970 cycles =    26.4 mbits/sec

256 bit key:
Key Setup:    2463 cycles
Encrypt:       985 cycles =    26.0 mbits/sec
Decrypt:       954 cycles =    26.8 mbits/sec
Mean:          970 cycles =    26.4 mbits/sec

*/

#define BLOCK_SWAP

#ifdef CORE_TIME
#  undef    BLOCK_SWAP
#endif

#include "../std_defs.h"

static char *alg_name[] = { "serpent", "serpent.c", "serpent" };

char **cipher_name()
{
    return alg_name;
}

/* Partially optimised Serpent S Box boolean functions derived  */
/* using a recursive descent analyser but without a full search */
/* of all subtrees. This set of S boxes is the result of work   */
/* by Sam Simpson and Brian Gladman using the spare time on a   */
/* cluster of high capacity servers to search for S boxes with  */
/* this customised search engine.                               */
/*                                                              */
/* Copyright:   Dr B. R Gladman ([e-mail])   */
/*              and Sam Simpson ([e-mail])           */ 
/*              17th December 1998                              */
/*                                                              */
/* We hereby give permission for information in this file to be */
/* used freely subject only to acknowledgement of its origin    */

/* 15 terms */

#define sb0(a,b,c,d,e,f,g,h)    \
    t1 = a ^ d;     \
    t2 = a & d;     \
    t3 = c ^ t1;    \
    t6 = b & t1;    \
    t4 = b ^ t3;    \
    t10 = ~t3;      \
    h = t2 ^ t4;    \
    t7 = a ^ t6;    \
    t14 = ~t7;      \
    t8 = c | t7;    \
    t11 = t3 ^ t7;  \
    g = t4 ^ t8;    \
    t12 = h & t11;  \
    f = t10 ^ t12;  \
    e = t12 ^ t14

/* 15 terms */

#define ib0(a,b,c,d,e,f,g,h)    \
    t1 = ~a;        \
    t2 = a ^ b;     \
    t3 = t1 | t2;   \
    t4 = d ^ t3;    \
    t7 = d & t2;    \
    t5 = c ^ t4;    \
    t8 = t1 ^ t7;   \
    g = t2 ^ t5;    \
    t11 = a & t4;   \
    t9 = g & t8;    \
    t14 = t5 ^ t8;  \
    f = t4 ^ t9;    \
    t12 = t5 | f;   \
    h = t11 ^ t12;  \
    e = h ^ t14

/* 14 terms!  */

#define sb1(a,b,c,d,e,f,g,h)    \
    t1 = ~a;        \
    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

/* 17 terms */

#define ib1(a,b,c,d,e,f,g,h)    \
    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;      \
    t9 = t4 & t8;   \
    f = t1 ^ t9;    \
    t13 = t9 ^ t11; \
    t12 = h & f;    \
    g = t12 ^ t13;  \
    t15 = a & d;    \
    t16 = c ^ t13;  \
    e = t15 ^ t16

/* 16 terms */

#define sb2(a,b,c,d,e,f,g,h)    \
    t1 = ~a;        \
    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

/* 16 terms */

#define ib2(a,b,c,d,e,f,g,h)    \
    t1 = b ^ d;     \
    t2 = ~t1;       \
    t3 = a ^ c;     \
    t4 = c ^ t1;    \
    t7 = a | t2;    \
    t5 = b & t4;    \
    t8 = d ^ t7;    \
    t11 = ~t4;      \
    e = t3 ^ t5;    \
    t9 = t3 | t8;   \
    t14 = d & t11;  \
    h = t1 ^ t9;    \
    t12 = e | h;    \
    f = t11 ^ t12;  \
    t15 = t3 ^ t12; \
    g = t14 ^ t15

/* 17 terms */

#define sb3(a,b,c,d,e,f,g,h)    \
    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

/* 16 term solution that performs less well than 17 term one
   in my environment (PPro/PII)                                  

#define sb3(a,b,c,d,e,f,g,h)    \
    t1 = a ^ b;     \
    t2 = a & c;     \
    t3 = a | d;     \
    t4 = c ^ d;     \
    t5 = t1 & t3;   \
    t6 = t2 | t5;   \
    g = t4 ^ t6;    \
    t8 = b ^ t3;    \
    t9 = t6 ^ t8;   \
    t10 = t4 & t9;  \
    e = t1 ^ t10;   \
    t12 = g & e;    \
    f = t9 ^ t12;   \
    t14 = b | d;    \
    t15 = t4 ^ t12; \
    h = t14 ^ t15
*/

/* 17 terms */

#define ib3(a,b,c,d,e,f,g,h)    \
    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

/* 15 terms */

#define sb4(a,b,c,d,e,f,g,h)    \
    t1 = a ^ d;     \
    t2 = d & t1;    \
    t3 = c ^ t2;    \
    t4 = b | t3;    \
    h = t1 ^ t4;    \
    t6 = ~b;        \
    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

/* 17 terms */

#define ib4(a,b,c,d,e,f,g,h)    \
    t1 = c ^ d;     \
    t2 = c | d;     \
    t3 = b ^ t2;    \
    t4 = a & t3;    \
    f = t1 ^ t4;    \
    t6 = a ^ d;     \
    t7 = b | d;     \
    t8 = t6 & t7;   \
    h = t3 ^ t8;    \
    t10 = ~a;       \
    t11 = c ^ h;    \
    t12 = t10 | t11;\
    e = t3 ^ t12;   \
    t14 = c | t4;   \
    t15 = t7 ^ t14; \
    t16 = h | t10;  \
    g = t15 ^ t16

/* 16 terms */

#define sb5(a,b,c,d,e,f,g,h)    \
    t1 = ~a;        \
    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

/* 16 terms */