crypto/rc4/rc4_enc.c

   1 /* crypto/rc4/rc4_enc.c */
   2 /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
   3  * All rights reserved.
   4  *
   5  * This package is an SSL implementation written
   6  * by Eric Young (eay@cryptsoft.com).
   7  * The implementation was written so as to conform with Netscapes SSL.
   8  *
   9  * This library is free for commercial and non-commercial use as long as
  10  * the following conditions are aheared to.  The following conditions
  11  * apply to all code found in this distribution, be it the RC4, RSA,
  12  * lhash, DES, etc., code; not just the SSL code.  The SSL documentation
  13  * included with this distribution is covered by the same copyright terms
  14  * except that the holder is Tim Hudson (tjh@cryptsoft.com).
  15  *
  16  * Copyright remains Eric Young's, and as such any Copyright notices in
  17  * the code are not to be removed.
  18  * If this package is used in a product, Eric Young should be given attribution
  19  * as the author of the parts of the library used.
  20  * This can be in the form of a textual message at program startup or
  21  * in documentation (online or textual) provided with the package.
  22  *
  23  * Redistribution and use in source and binary forms, with or without
  24  * modification, are permitted provided that the following conditions
  25  * are met:
  26  * 1. Redistributions of source code must retain the copyright
  27  *    notice, this list of conditions and the following disclaimer.
  28  * 2. Redistributions in binary form must reproduce the above copyright
  29  *    notice, this list of conditions and the following disclaimer in the
  30  *    documentation and/or other materials provided with the distribution.
  31  * 3. All advertising materials mentioning features or use of this software
  32  *    must display the following acknowledgement:
  33  *    "This product includes cryptographic software written by
  34  *     Eric Young (eay@cryptsoft.com)"
  35  *    The word 'cryptographic' can be left out if the rouines from the library
  36  *    being used are not cryptographic related :-).
  37  * 4. If you include any Windows specific code (or a derivative thereof) from
  38  *    the apps directory (application code) you must include an acknowledgement:
  39  *    "This product includes software written by Tim Hudson (tjh@cryptsoft.com)"
  40  *
  41  * THIS SOFTWARE IS PROVIDED BY ERIC YOUNG ``AS IS'' AND
  42  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  43  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  44  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  45  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  46  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  47  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  48  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  49  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  50  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  51  * SUCH DAMAGE.
  52  *
  53  * The licence and distribution terms for any publically available version or
  54  * derivative of this code cannot be changed.  i.e. this code cannot simply be
  55  * copied and put under another distribution licence
  56  * [including the GNU Public Licence.]
  57  */
  58
  59 #include <openssl/rc4.h>
  60 #include "rc4_locl.h"
  61
  62 /* RC4 as implemented from a posting from
  63  * Newsgroups: sci.crypt
  64  * From: sterndark@netcom.com (David Sterndark)
  65  * Subject: RC4 Algorithm revealed.
  66  * Message-ID: <sternCvKL4B.Hyy@netcom.com>
  67  * Date: Wed, 14 Sep 1994 06:35:31 GMT
  68  */
  69
  70 void RC4(RC4_KEY *key, unsigned long len, const unsigned char *indata,
  71              unsigned char *outdata)
  72         {
  73         register RC4_INT *d;
  74         register RC4_INT x,y,tx,ty;
  75         int i;
  76
  77         x=key->x;
  78         y=key->y;
  79         d=key->data;
  80 #if defined(__ia64) || defined(__ia64__) || defined(_M_IA64)
  81         /* see crypto/rc4/asm/rc4-ia64.S for further details... */
  82         d=(RC4_INT *)(((size_t)(d+255))&~(sizeof(key->data)-1));
  83 #endif
  84
  85 #if defined(RC4_CHUNK)
  86         /*
  87          * The original reason for implementing this(*) was the fact that
  88          * pre-21164a Alpha CPUs don't have byte load/store instructions
  89          * and e.g. a byte store has to be done with 64-bit load, shift,
  90          * and, or and finally 64-bit store. Peaking data and operating
  91          * at natural word size made it possible to reduce amount of
  92          * instructions as well as to perform early read-ahead without
  93          * suffering from RAW (read-after-write) hazard. This resulted
  94          * in ~40%(**) performance improvement on 21064 box with gcc.
  95          * But it's not only Alpha users who win here:-) Thanks to the
  96          * early-n-wide read-ahead this implementation also exhibits
  97          * >40% speed-up on SPARC and 20-30% on 64-bit MIPS (depending
  98          * on sizeof(RC4_INT)).
  99          *
 100          * (*)  "this" means code which recognizes the case when input
 101          *      and output pointers appear to be aligned at natural CPU
 102          *      word boundary
 103          * (**) i.e. according to 'apps/openssl speed rc4' benchmark,
 104          *      crypto/rc4/rc4speed.c exhibits almost 70% speed-up...
 105          *
 106          * Cavets.
 107          *
 108          * - RC4_CHUNK="unsigned long long" should be a #1 choice for
 109          *   UltraSPARC. Unfortunately gcc generates very slow code
 110          *   (2.5-3 times slower than one generated by Sun's WorkShop
 111          *   C) and therefore gcc (at least 2.95 and earlier) should
 112          *   always be told that RC4_CHUNK="unsigned long".
 113          *
 114          *                                      <appro@fy.chalmers.se>
 115          */
 116
 117 # define RC4_STEP       ( \
 118                         x=(x+1) &0xff,  \
 119                         tx=d[x],        \
 120                         y=(tx+y)&0xff,  \
 121                         ty=d[y],        \
 122                         d[y]=tx,        \
 123                         d[x]=ty,        \
 124                         (RC4_CHUNK)d[(tx+ty)&0xff]\
 125                         )
 126
 127         if ( ( ((unsigned long)indata  & (sizeof(RC4_CHUNK)-1)) |
 128                ((unsigned long)outdata & (sizeof(RC4_CHUNK)-1)) ) == 0 )
 129                 {
 130                 RC4_CHUNK ichunk,otp;
 131                 const union { long one; char little; } is_endian = {1};
 132
 133                 /*
 134                  * I reckon we can afford to implement both endian
 135                  * cases and to decide which way to take at run-time
 136                  * because the machine code appears to be very compact
 137                  * and redundant 1-2KB is perfectly tolerable (i.e.
 138                  * in case the compiler fails to eliminate it:-). By
 139                  * suggestion from Terrel Larson <terr@terralogic.net>
 140                  * who also stands for the is_endian union:-)
 141                  *
 142                  * Special notes.
 143                  *
 144                  * - is_endian is declared automatic as doing otherwise
 145                  *   (declaring static) prevents gcc from eliminating
 146                  *   the redundant code;
 147                  * - compilers (those I've tried) don't seem to have
 148                  *   problems eliminating either the operators guarded
 149                  *   by "if (sizeof(RC4_CHUNK)==8)" or the condition
 150                  *   expressions themselves so I've got 'em to replace
 151                  *   corresponding #ifdefs from the previous version;
 152                  * - I chose to let the redundant switch cases when
 153                  *   sizeof(RC4_CHUNK)!=8 be (were also #ifdefed
 154                  *   before);
 155                  * - in case you wonder "&(sizeof(RC4_CHUNK)*8-1)" in
 156                  *   [LB]ESHFT guards against "shift is out of range"
 157                  *   warnings when sizeof(RC4_CHUNK)!=8
 158                  *
 159                  *                      <appro@fy.chalmers.se>
 160                  */
 161                 if (!is_endian.little)
 162                         {       /* BIG-ENDIAN CASE */
 163 # define BESHFT(c)      (((sizeof(RC4_CHUNK)-(c)-1)*8)&(sizeof(RC4_CHUNK)*8-1))
 164                         for (;len&-sizeof(RC4_CHUNK);len-=sizeof(RC4_CHUNK))
 165                                 {
 166                                 ichunk  = *(RC4_CHUNK *)indata;
 167                                 otp  = RC4_STEP<<BESHFT(0);
 168                                 otp |= RC4_STEP<<BESHFT(1);
 169                                 otp |= RC4_STEP<<BESHFT(2);
 170                                 otp |= RC4_STEP<<BESHFT(3);
 171                                 if (sizeof(RC4_CHUNK)==8)
 172                                         {
 173                                         otp |= RC4_STEP<<BESHFT(4);
 174                                         otp |= RC4_STEP<<BESHFT(5);
 175                                         otp |= RC4_STEP<<BESHFT(6);
 176                                         otp |= RC4_STEP<<BESHFT(7);
 177                                         }
 178                                 *(RC4_CHUNK *)outdata = otp^ichunk;
 179                                 indata  += sizeof(RC4_CHUNK);
 180                                 outdata += sizeof(RC4_CHUNK);
 181                                 }
 182                         if (len)
 183                                 {
 184                                 RC4_CHUNK mask=(RC4_CHUNK)-1, ochunk;
 185
 186                                 ichunk = *(RC4_CHUNK *)indata;
 187                                 ochunk = *(RC4_CHUNK *)outdata;
 188                                 otp = 0;
 189                                 i = BESHFT(0);
 190                                 mask <<= (sizeof(RC4_CHUNK)-len)<<3;
 191                                 switch (len&(sizeof(RC4_CHUNK)-1))
 192                                         {
 193                                         case 7: otp  = RC4_STEP<<i, i-=8;
 194                                         case 6: otp |= RC4_STEP<<i, i-=8;
 195                                         case 5: otp |= RC4_STEP<<i, i-=8;
 196                                         case 4: otp |= RC4_STEP<<i, i-=8;
 197                                         case 3: otp |= RC4_STEP<<i, i-=8;
 198                                         case 2: otp |= RC4_STEP<<i, i-=8;
 199                                         case 1: otp |= RC4_STEP<<i, i-=8;
 200                                         case 0: ; /*
 201                                                    * it's never the case,
 202                                                    * but it has to be here
 203                                                    * for ultrix?
 204                                                    */
 205                                         }
 206                                 ochunk &= ~mask;
 207                                 ochunk |= (otp^ichunk) & mask;
 208                                 *(RC4_CHUNK *)outdata = ochunk;
 209                                 }
 210                         key->x=x;
 211                         key->y=y;
 212                         return;
 213                         }
 214                 else
 215                         {       /* LITTLE-ENDIAN CASE */
 216 # define LESHFT(c)      (((c)*8)&(sizeof(RC4_CHUNK)*8-1))
 217                         for (;len&-sizeof(RC4_CHUNK);len-=sizeof(RC4_CHUNK))
 218                                 {
 219                                 ichunk  = *(RC4_CHUNK *)indata;
 220                                 otp  = RC4_STEP;
 221                                 otp |= RC4_STEP<<8;
 222                                 otp |= RC4_STEP<<16;
 223                                 otp |= RC4_STEP<<24;
 224                                 if (sizeof(RC4_CHUNK)==8)
 225                                         {
 226                                         otp |= RC4_STEP<<LESHFT(4);
 227                                         otp |= RC4_STEP<<LESHFT(5);
 228                                         otp |= RC4_STEP<<LESHFT(6);
 229                                         otp |= RC4_STEP<<LESHFT(7);
 230                                         }
 231                                 *(RC4_CHUNK *)outdata = otp^ichunk;
 232                                 indata  += sizeof(RC4_CHUNK);
 233                                 outdata += sizeof(RC4_CHUNK);
 234                                 }
 235                         if (len)
 236                                 {
 237                                 RC4_CHUNK mask=(RC4_CHUNK)-1, ochunk;
 238
 239                                 ichunk = *(RC4_CHUNK *)indata;
 240                                 ochunk = *(RC4_CHUNK *)outdata;
 241                                 otp = 0;
 242                                 i   = 0;
 243                                 mask >>= (sizeof(RC4_CHUNK)-len)<<3;
 244                                 switch (len&(sizeof(RC4_CHUNK)-1))
 245                                         {
 246                                         case 7: otp  = RC4_STEP,    i+=8;
 247                                         case 6: otp |= RC4_STEP<<i, i+=8;
 248                                         case 5: otp |= RC4_STEP<<i, i+=8;
 249                                         case 4: otp |= RC4_STEP<<i, i+=8;
 250                                         case 3: otp |= RC4_STEP<<i, i+=8;
 251                                         case 2: otp |= RC4_STEP<<i, i+=8;
 252                                         case 1: otp |= RC4_STEP<<i, i+=8;
 253                                         case 0: ; /*
 254                                                    * it's never the case,
 255                                                    * but it has to be here
 256                                                    * for ultrix?
 257                                                    */
 258                                         }
 259                                 ochunk &= ~mask;
 260                                 ochunk |= (otp^ichunk) & mask;
 261                                 *(RC4_CHUNK *)outdata = ochunk;
 262                                 }
 263                         key->x=x;
 264                         key->y=y;
 265                         return;
 266                         }
 267                 }
 268 #endif
 269 #define LOOP(in,out) \
 270                 x=((x+1)&0xff); \
 271                 tx=d[x]; \
 272                 y=(tx+y)&0xff; \
 273                 d[x]=ty=d[y]; \
 274                 d[y]=tx; \
 275                 (out) = d[(tx+ty)&0xff]^ (in);
 276
 277 #ifndef RC4_INDEX
 278 #define RC4_LOOP(a,b,i) LOOP(*((a)++),*((b)++))
 279 #else
 280 #define RC4_LOOP(a,b,i) LOOP(a[i],b[i])
 281 #endif
 282
 283         i=(int)(len>>3L);
 284         if (i)
 285                 {
 286                 for (;;)
 287                         {
 288                         RC4_LOOP(indata,outdata,0);
 289                         RC4_LOOP(indata,outdata,1);
 290                         RC4_LOOP(indata,outdata,2);
 291                         RC4_LOOP(indata,outdata,3);
 292                         RC4_LOOP(indata,outdata,4);
 293                         RC4_LOOP(indata,outdata,5);
 294                         RC4_LOOP(indata,outdata,6);
 295                         RC4_LOOP(indata,outdata,7);
 296 #ifdef RC4_INDEX
 297                         indata+=8;
 298                         outdata+=8;
 299 #endif
 300                         if (--i == 0) break;
 301                         }
 302                 }
 303         i=(int)len&0x07;
 304         if (i)
 305                 {
 306                 for (;;)
 307                         {
 308                         RC4_LOOP(indata,outdata,0); if (--i == 0) break;
 309                         RC4_LOOP(indata,outdata,1); if (--i == 0) break;
 310                         RC4_LOOP(indata,outdata,2); if (--i == 0) break;
 311                         RC4_LOOP(indata,outdata,3); if (--i == 0) break;
 312                         RC4_LOOP(indata,outdata,4); if (--i == 0) break;
 313                         RC4_LOOP(indata,outdata,5); if (--i == 0) break;
 314                         RC4_LOOP(indata,outdata,6); if (--i == 0) break;
 315                         }
 316                 }
 317         key->x=x;
 318         key->y=y;
 319         }