crypto/bn/bn_nist.c

   1 /* crypto/bn/bn_nist.c */
   2 /*
   3  * Written by Nils Larsch for the OpenSSL project
   4  */
   5 /* ====================================================================
   6  * Copyright (c) 1998-2005 The OpenSSL Project.  All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  *
  15  * 2. Redistributions in binary form must reproduce the above copyright
  16  *    notice, this list of conditions and the following disclaimer in
  17  *    the documentation and/or other materials provided with the
  18  *    distribution.
  19  *
  20  * 3. All advertising materials mentioning features or use of this
  21  *    software must display the following acknowledgment:
  22  *    "This product includes software developed by the OpenSSL Project
  23  *    for use in the OpenSSL Toolkit. (http://www.openssl.org/)"
  24  *
  25  * 4. The names "OpenSSL Toolkit" and "OpenSSL Project" must not be used to
  26  *    endorse or promote products derived from this software without
  27  *    prior written permission. For written permission, please contact
  28  *    openssl-core@openssl.org.
  29  *
  30  * 5. Products derived from this software may not be called "OpenSSL"
  31  *    nor may "OpenSSL" appear in their names without prior written
  32  *    permission of the OpenSSL Project.
  33  *
  34  * 6. Redistributions of any form whatsoever must retain the following
  35  *    acknowledgment:
  36  *    "This product includes software developed by the OpenSSL Project
  37  *    for use in the OpenSSL Toolkit (http://www.openssl.org/)"
  38  *
  39  * THIS SOFTWARE IS PROVIDED BY THE OpenSSL PROJECT ``AS IS'' AND ANY
  40  * EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  41  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  42  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE OpenSSL PROJECT OR
  43  * ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  44  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  45  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  46  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  47  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  48  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
  49  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
  50  * OF THE POSSIBILITY OF SUCH DAMAGE.
  51  * ====================================================================
  52  *
  53  * This product includes cryptographic software written by Eric Young
  54  * (eay@cryptsoft.com).  This product includes software written by Tim
  55  * Hudson (tjh@cryptsoft.com).
  56  *
  57  */
  58
  59 #include "bn_lcl.h"
  60 #include "cryptlib.h"
  61
  62
  63 #define BN_NIST_192_TOP (192+BN_BITS2-1)/BN_BITS2
  64 #define BN_NIST_224_TOP (224+BN_BITS2-1)/BN_BITS2
  65 #define BN_NIST_256_TOP (256+BN_BITS2-1)/BN_BITS2
  66 #define BN_NIST_384_TOP (384+BN_BITS2-1)/BN_BITS2
  67 #define BN_NIST_521_TOP (521+BN_BITS2-1)/BN_BITS2
  68
  69 /* pre-computed tables are "carry-less" values of modulus*(i+1) */
  70 #if BN_BITS2 == 64
  71 __fips_constseg
  72 static const BN_ULONG _nist_p_192[][BN_NIST_192_TOP] = {
  73         {0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFEULL,0xFFFFFFFFFFFFFFFFULL},
  74         {0xFFFFFFFFFFFFFFFEULL,0xFFFFFFFFFFFFFFFDULL,0xFFFFFFFFFFFFFFFFULL},
  75         {0xFFFFFFFFFFFFFFFDULL,0xFFFFFFFFFFFFFFFCULL,0xFFFFFFFFFFFFFFFFULL}
  76         };
  77 __fips_constseg
  78 static const BN_ULONG _nist_p_192_sqr[] = {
  79         0x0000000000000001ULL,0x0000000000000002ULL,0x0000000000000001ULL,
  80         0xFFFFFFFFFFFFFFFEULL,0xFFFFFFFFFFFFFFFDULL,0xFFFFFFFFFFFFFFFFULL
  81         };
  82 __fips_constseg
  83 static const BN_ULONG _nist_p_224[][BN_NIST_224_TOP] = {
  84         {0x0000000000000001ULL,0xFFFFFFFF00000000ULL,
  85          0xFFFFFFFFFFFFFFFFULL,0x00000000FFFFFFFFULL},
  86         {0x0000000000000002ULL,0xFFFFFFFE00000000ULL,
  87          0xFFFFFFFFFFFFFFFFULL,0x00000001FFFFFFFFULL} /* this one is "carry-full" */
  88         };
  89 __fips_constseg
  90 static const BN_ULONG _nist_p_224_sqr[] = {
  91         0x0000000000000001ULL,0xFFFFFFFE00000000ULL,
  92         0xFFFFFFFFFFFFFFFFULL,0x0000000200000000ULL,
  93         0x0000000000000000ULL,0xFFFFFFFFFFFFFFFEULL,
  94         0xFFFFFFFFFFFFFFFFULL
  95         };
  96 __fips_constseg
  97 static const BN_ULONG _nist_p_256[][BN_NIST_256_TOP] = {
  98         {0xFFFFFFFFFFFFFFFFULL,0x00000000FFFFFFFFULL,
  99          0x0000000000000000ULL,0xFFFFFFFF00000001ULL},
 100         {0xFFFFFFFFFFFFFFFEULL,0x00000001FFFFFFFFULL,
 101          0x0000000000000000ULL,0xFFFFFFFE00000002ULL},
 102         {0xFFFFFFFFFFFFFFFDULL,0x00000002FFFFFFFFULL,
 103          0x0000000000000000ULL,0xFFFFFFFD00000003ULL},
 104         {0xFFFFFFFFFFFFFFFCULL,0x00000003FFFFFFFFULL,
 105          0x0000000000000000ULL,0xFFFFFFFC00000004ULL},
 106         {0xFFFFFFFFFFFFFFFBULL,0x00000004FFFFFFFFULL,
 107          0x0000000000000000ULL,0xFFFFFFFB00000005ULL},
 108         };
 109 __fips_constseg
 110 static const BN_ULONG _nist_p_256_sqr[] = {
 111         0x0000000000000001ULL,0xFFFFFFFE00000000ULL,
 112         0xFFFFFFFFFFFFFFFFULL,0x00000001FFFFFFFEULL,
 113         0x00000001FFFFFFFEULL,0x00000001FFFFFFFEULL,
 114         0xFFFFFFFE00000001ULL,0xFFFFFFFE00000002ULL
 115         };
 116 __fips_constseg
 117 static const BN_ULONG _nist_p_384[][BN_NIST_384_TOP] = {
 118         {0x00000000FFFFFFFFULL,0xFFFFFFFF00000000ULL,0xFFFFFFFFFFFFFFFEULL,
 119          0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL},
 120         {0x00000001FFFFFFFEULL,0xFFFFFFFE00000000ULL,0xFFFFFFFFFFFFFFFDULL,
 121          0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL},
 122         {0x00000002FFFFFFFDULL,0xFFFFFFFD00000000ULL,0xFFFFFFFFFFFFFFFCULL,
 123          0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL},
 124         {0x00000003FFFFFFFCULL,0xFFFFFFFC00000000ULL,0xFFFFFFFFFFFFFFFBULL,
 125          0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL},
 126         {0x00000004FFFFFFFBULL,0xFFFFFFFB00000000ULL,0xFFFFFFFFFFFFFFFAULL,
 127          0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL},
 128         };
 129 __fips_constseg
 130 static const BN_ULONG _nist_p_384_sqr[] = {
 131         0xFFFFFFFE00000001ULL,0x0000000200000000ULL,0xFFFFFFFE00000000ULL,
 132         0x0000000200000000ULL,0x0000000000000001ULL,0x0000000000000000ULL,
 133         0x00000001FFFFFFFEULL,0xFFFFFFFE00000000ULL,0xFFFFFFFFFFFFFFFDULL,
 134         0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL
 135         };
 136 __fips_constseg
 137 static const BN_ULONG _nist_p_521[] =
 138         {0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,
 139         0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,
 140         0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,
 141         0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,
 142         0x00000000000001FFULL};
 143 __fips_constseg
 144 static const BN_ULONG _nist_p_521_sqr[] = {
 145         0x0000000000000001ULL,0x0000000000000000ULL,0x0000000000000000ULL,
 146         0x0000000000000000ULL,0x0000000000000000ULL,0x0000000000000000ULL,
 147         0x0000000000000000ULL,0x0000000000000000ULL,0xFFFFFFFFFFFFFC00ULL,
 148         0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,
 149         0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,0xFFFFFFFFFFFFFFFFULL,
 150         0xFFFFFFFFFFFFFFFFULL,0x000000000003FFFFULL
 151         };
 152 #elif BN_BITS2 == 32
 153 __fips_constseg
 154 static const BN_ULONG _nist_p_192[][BN_NIST_192_TOP] = {
 155         {0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFE,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF},
 156         {0xFFFFFFFE,0xFFFFFFFF,0xFFFFFFFD,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF},
 157         {0xFFFFFFFD,0xFFFFFFFF,0xFFFFFFFC,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF}
 158         };
 159 __fips_constseg
 160 static const BN_ULONG _nist_p_192_sqr[] = {
 161         0x00000001,0x00000000,0x00000002,0x00000000,0x00000001,0x00000000,
 162         0xFFFFFFFE,0xFFFFFFFF,0xFFFFFFFD,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF
 163         };
 164 __fips_constseg
 165 static const BN_ULONG _nist_p_224[][BN_NIST_224_TOP] = {
 166         {0x00000001,0x00000000,0x00000000,0xFFFFFFFF,
 167          0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF},
 168         {0x00000002,0x00000000,0x00000000,0xFFFFFFFE,
 169          0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF}
 170         };
 171 __fips_constseg
 172 static const BN_ULONG _nist_p_224_sqr[] = {
 173         0x00000001,0x00000000,0x00000000,0xFFFFFFFE,
 174         0xFFFFFFFF,0xFFFFFFFF,0x00000000,0x00000002,
 175         0x00000000,0x00000000,0xFFFFFFFE,0xFFFFFFFF,
 176         0xFFFFFFFF,0xFFFFFFFF
 177         };
 178 __fips_constseg
 179 static const BN_ULONG _nist_p_256[][BN_NIST_256_TOP] = {
 180         {0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0x00000000,
 181          0x00000000,0x00000000,0x00000001,0xFFFFFFFF},
 182         {0xFFFFFFFE,0xFFFFFFFF,0xFFFFFFFF,0x00000001,
 183          0x00000000,0x00000000,0x00000002,0xFFFFFFFE},
 184         {0xFFFFFFFD,0xFFFFFFFF,0xFFFFFFFF,0x00000002,
 185          0x00000000,0x00000000,0x00000003,0xFFFFFFFD},
 186         {0xFFFFFFFC,0xFFFFFFFF,0xFFFFFFFF,0x00000003,
 187          0x00000000,0x00000000,0x00000004,0xFFFFFFFC},
 188         {0xFFFFFFFB,0xFFFFFFFF,0xFFFFFFFF,0x00000004,
 189          0x00000000,0x00000000,0x00000005,0xFFFFFFFB},
 190         };
 191 __fips_constseg
 192 static const BN_ULONG _nist_p_256_sqr[] = {
 193         0x00000001,0x00000000,0x00000000,0xFFFFFFFE,
 194         0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFE,0x00000001,
 195         0xFFFFFFFE,0x00000001,0xFFFFFFFE,0x00000001,
 196         0x00000001,0xFFFFFFFE,0x00000002,0xFFFFFFFE
 197         };
 198 __fips_constseg
 199 static const BN_ULONG _nist_p_384[][BN_NIST_384_TOP] = {
 200         {0xFFFFFFFF,0x00000000,0x00000000,0xFFFFFFFF,0xFFFFFFFE,0xFFFFFFFF,
 201          0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF},
 202         {0xFFFFFFFE,0x00000001,0x00000000,0xFFFFFFFE,0xFFFFFFFD,0xFFFFFFFF,
 203          0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF},
 204         {0xFFFFFFFD,0x00000002,0x00000000,0xFFFFFFFD,0xFFFFFFFC,0xFFFFFFFF,
 205          0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF},
 206         {0xFFFFFFFC,0x00000003,0x00000000,0xFFFFFFFC,0xFFFFFFFB,0xFFFFFFFF,
 207          0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF},
 208         {0xFFFFFFFB,0x00000004,0x00000000,0xFFFFFFFB,0xFFFFFFFA,0xFFFFFFFF,
 209          0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF},
 210         };
 211 __fips_constseg
 212 static const BN_ULONG _nist_p_384_sqr[] = {
 213         0x00000001,0xFFFFFFFE,0x00000000,0x00000002,0x00000000,0xFFFFFFFE,
 214         0x00000000,0x00000002,0x00000001,0x00000000,0x00000000,0x00000000,
 215         0xFFFFFFFE,0x00000001,0x00000000,0xFFFFFFFE,0xFFFFFFFD,0xFFFFFFFF,
 216         0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF
 217         };
 218 __fips_constseg
 219 static const BN_ULONG _nist_p_521[] = {0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,
 220         0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,
 221         0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,
 222         0xFFFFFFFF,0x000001FF};
 223 __fips_constseg
 224 static const BN_ULONG _nist_p_521_sqr[] = {
 225         0x00000001,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,
 226         0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,
 227         0x00000000,0x00000000,0x00000000,0x00000000,0xFFFFFC00,0xFFFFFFFF,
 228         0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,
 229         0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,
 230         0xFFFFFFFF,0xFFFFFFFF,0x0003FFFF
 231         };
 232 #else
 233 #error "unsupported BN_BITS2"
 234 #endif
 235
 236
 237 static const BIGNUM _bignum_nist_p_192 =
 238         {
 239         (BN_ULONG *)_nist_p_192[0],
 240         BN_NIST_192_TOP,
 241         BN_NIST_192_TOP,
 242         0,
 243         BN_FLG_STATIC_DATA
 244         };
 245
 246 static const BIGNUM _bignum_nist_p_224 =
 247         {
 248         (BN_ULONG *)_nist_p_224[0],
 249         BN_NIST_224_TOP,
 250         BN_NIST_224_TOP,
 251         0,
 252         BN_FLG_STATIC_DATA
 253         };
 254
 255 static const BIGNUM _bignum_nist_p_256 =
 256         {
 257         (BN_ULONG *)_nist_p_256[0],
 258         BN_NIST_256_TOP,
 259         BN_NIST_256_TOP,
 260         0,
 261         BN_FLG_STATIC_DATA
 262         };
 263
 264 static const BIGNUM _bignum_nist_p_384 =
 265         {
 266         (BN_ULONG *)_nist_p_384[0],
 267         BN_NIST_384_TOP,
 268         BN_NIST_384_TOP,
 269         0,
 270         BN_FLG_STATIC_DATA
 271         };
 272
 273 static const BIGNUM _bignum_nist_p_521 =
 274         {
 275         (BN_ULONG *)_nist_p_521,
 276         BN_NIST_521_TOP,
 277         BN_NIST_521_TOP,
 278         0,
 279         BN_FLG_STATIC_DATA
 280         };
 281
 282
 283 const BIGNUM *BN_get0_nist_prime_192(void)
 284         {
 285         return &_bignum_nist_p_192;
 286         }
 287
 288 const BIGNUM *BN_get0_nist_prime_224(void)
 289         {
 290         return &_bignum_nist_p_224;
 291         }
 292
 293 const BIGNUM *BN_get0_nist_prime_256(void)
 294         {
 295         return &_bignum_nist_p_256;
 296         }
 297
 298 const BIGNUM *BN_get0_nist_prime_384(void)
 299         {
 300         return &_bignum_nist_p_384;
 301         }
 302
 303 const BIGNUM *BN_get0_nist_prime_521(void)
 304         {
 305         return &_bignum_nist_p_521;
 306         }
 307
 308
 309 static void nist_cp_bn_0(BN_ULONG *dst, const BN_ULONG *src, int top, int max)
 310         {
 311         int i;
 312
 313 #ifdef BN_DEBUG
 314         OPENSSL_assert(top <= max);
 315 #endif
 316         for (i = 0; i < top; i++)
 317                 dst[i] = src[i];
 318         for (; i < max; i++)
 319                 dst[i] = 0;
 320         }
 321
 322 static void nist_cp_bn(BN_ULONG *dst, const BN_ULONG *src, int top)
 323         {
 324         int i;
 325
 326         for (i = 0; i < top; i++)
 327                 dst[i] = src[i];
 328         }
 329
 330 #if BN_BITS2 == 64
 331 #define bn_cp_64(to, n, from, m)        (to)[n] = (m>=0)?((from)[m]):0;
 332 #define bn_64_set_0(to, n)              (to)[n] = (BN_ULONG)0;
 333 /*
 334  * two following macros are implemented under assumption that they
 335  * are called in a sequence with *ascending* n, i.e. as they are...
 336  */
 337 #define bn_cp_32_naked(to, n, from, m)  (((n)&1)?(to[(n)/2]|=((m)&1)?(from[(m)/2]&BN_MASK2h):(from[(m)/2]<<32))\
 338                                                 :(to[(n)/2] =((m)&1)?(from[(m)/2]>>32):(from[(m)/2]&BN_MASK2l)))
 339 #define bn_32_set_0(to, n)              (((n)&1)?(to[(n)/2]&=BN_MASK2l):(to[(n)/2]=0));
 340 #define bn_cp_32(to,n,from,m)           ((m)>=0)?bn_cp_32_naked(to,n,from,m):bn_32_set_0(to,n)
 341 # if defined(L_ENDIAN)
 342 #  if defined(__arch64__)
 343 #   define NIST_INT64 long
 344 #  else
 345 #   define NIST_INT64 long long
 346 #  endif
 347 # endif
 348 #else
 349 #define bn_cp_64(to, n, from, m) \
 350         { \
 351         bn_cp_32(to, (n)*2, from, (m)*2); \
 352         bn_cp_32(to, (n)*2+1, from, (m)*2+1); \
 353         }
 354 #define bn_64_set_0(to, n) \
 355         { \
 356         bn_32_set_0(to, (n)*2); \
 357         bn_32_set_0(to, (n)*2+1); \
 358         }
 359 #define bn_cp_32(to, n, from, m)        (to)[n] = (m>=0)?((from)[m]):0;
 360 #define bn_32_set_0(to, n)              (to)[n] = (BN_ULONG)0;
 361 # if defined(_WIN32) && !defined(__GNUC__)
 362 #  define NIST_INT64 __int64
 363 # elif defined(BN_LLONG)
 364 #  define NIST_INT64 long long
 365 # endif
 366 #endif /* BN_BITS2 != 64 */
 367
 368 #define nist_set_192(to, from, a1, a2, a3) \
 369         { \
 370         bn_cp_64(to, 0, from, (a3) - 3) \
 371         bn_cp_64(to, 1, from, (a2) - 3) \
 372         bn_cp_64(to, 2, from, (a1) - 3) \
 373         }
 374
 375 int BN_nist_mod_192(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 376         BN_CTX *ctx)
 377         {
 378         int      top = a->top, i;
 379         int      carry;
 380         register BN_ULONG *r_d, *a_d = a->d;
 381         union   {
 382                 BN_ULONG        bn[BN_NIST_192_TOP];
 383                 unsigned int    ui[BN_NIST_192_TOP*sizeof(BN_ULONG)/sizeof(unsigned int)];
 384                 } buf;
 385         BN_ULONG c_d[BN_NIST_192_TOP],
 386                 *res;
 387         PTR_SIZE_INT mask;
 388         static const BIGNUM _bignum_nist_p_192_sqr = {
 389                 (BN_ULONG *)_nist_p_192_sqr,
 390                 sizeof(_nist_p_192_sqr)/sizeof(_nist_p_192_sqr[0]),
 391                 sizeof(_nist_p_192_sqr)/sizeof(_nist_p_192_sqr[0]),
 392                 0,BN_FLG_STATIC_DATA };
 393
 394         field = &_bignum_nist_p_192; /* just to make sure */
 395
 396         if (BN_is_negative(a) || BN_ucmp(a,&_bignum_nist_p_192_sqr)>=0)
 397                 return BN_nnmod(r, a, field, ctx);
 398
 399         i = BN_ucmp(field, a);
 400         if (i == 0)
 401                 {
 402                 BN_zero(r);
 403                 return 1;
 404                 }
 405         else if (i > 0)
 406                 return (r == a) ? 1 : (BN_copy(r ,a) != NULL);
 407
 408         if (r != a)
 409                 {
 410                 if (!bn_wexpand(r, BN_NIST_192_TOP))
 411                         return 0;
 412                 r_d = r->d;
 413                 nist_cp_bn(r_d, a_d, BN_NIST_192_TOP);
 414                 }
 415         else
 416                 r_d = a_d;
 417
 418         nist_cp_bn_0(buf.bn, a_d + BN_NIST_192_TOP, top - BN_NIST_192_TOP, BN_NIST_192_TOP);
 419
 420 #if defined(NIST_INT64)
 421         {
 422         NIST_INT64              acc;    /* accumulator */
 423         unsigned int            *rp=(unsigned int *)r_d;
 424         const unsigned int      *bp=(const unsigned int *)buf.ui;
 425
 426         acc  = rp[0];   acc += bp[3*2-6];
 427                         acc += bp[5*2-6]; rp[0] = (unsigned int)acc; acc >>= 32;
 428
 429         acc += rp[1];   acc += bp[3*2-5];
 430                         acc += bp[5*2-5]; rp[1] = (unsigned int)acc; acc >>= 32;
 431
 432         acc += rp[2];   acc += bp[3*2-6];
 433                         acc += bp[4*2-6];
 434                         acc += bp[5*2-6]; rp[2] = (unsigned int)acc; acc >>= 32;
 435
 436         acc += rp[3];   acc += bp[3*2-5];
 437                         acc += bp[4*2-5];
 438                         acc += bp[5*2-5]; rp[3] = (unsigned int)acc; acc >>= 32;
 439
 440         acc += rp[4];   acc += bp[4*2-6];
 441                         acc += bp[5*2-6]; rp[4] = (unsigned int)acc; acc >>= 32;
 442
 443         acc += rp[5];   acc += bp[4*2-5];
 444                         acc += bp[5*2-5]; rp[5] = (unsigned int)acc;
 445
 446         carry = (int)(acc>>32);
 447         }
 448 #else
 449         {
 450         BN_ULONG t_d[BN_NIST_192_TOP];
 451
 452         nist_set_192(t_d, buf.bn, 0, 3, 3);
 453         carry = (int)bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP);
 454         nist_set_192(t_d, buf.bn, 4, 4, 0);
 455         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP);
 456         nist_set_192(t_d, buf.bn, 5, 5, 5)
 457         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP);
 458         }
 459 #endif
 460         if (carry > 0)
 461                 carry = (int)bn_sub_words(r_d,r_d,_nist_p_192[carry-1],BN_NIST_192_TOP);
 462         else
 463                 carry = 1;
 464
 465         /*
 466          * we need 'if (carry==0 || result>=modulus) result-=modulus;'
 467          * as comparison implies subtraction, we can write
 468          * 'tmp=result-modulus; if (!carry || !borrow) result=tmp;'
 469          * this is what happens below, but without explicit if:-) a.
 470          */
 471         mask  = 0-(PTR_SIZE_INT)bn_sub_words(c_d,r_d,_nist_p_192[0],BN_NIST_192_TOP);
 472         mask &= 0-(PTR_SIZE_INT)carry;
 473         res   = c_d;
 474         res   = (BN_ULONG *)
 475          (((PTR_SIZE_INT)res&~mask) | ((PTR_SIZE_INT)r_d&mask));
 476         nist_cp_bn(r_d, res, BN_NIST_192_TOP);
 477         r->top = BN_NIST_192_TOP;
 478         bn_correct_top(r);
 479
 480         return 1;
 481         }
 482
 483 typedef BN_ULONG (*bn_addsub_f)(BN_ULONG *,const BN_ULONG *,const BN_ULONG *,int);
 484
 485 #define nist_set_224(to, from, a1, a2, a3, a4, a5, a6, a7) \
 486         { \
 487         bn_cp_32(to, 0, from, (a7) - 7) \
 488         bn_cp_32(to, 1, from, (a6) - 7) \
 489         bn_cp_32(to, 2, from, (a5) - 7) \
 490         bn_cp_32(to, 3, from, (a4) - 7) \
 491         bn_cp_32(to, 4, from, (a3) - 7) \
 492         bn_cp_32(to, 5, from, (a2) - 7) \
 493         bn_cp_32(to, 6, from, (a1) - 7) \
 494         }
 495
 496 int BN_nist_mod_224(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 497         BN_CTX *ctx)
 498         {
 499         int     top = a->top, i;
 500         int     carry;
 501         BN_ULONG *r_d, *a_d = a->d;
 502         union   {
 503                 BN_ULONG        bn[BN_NIST_224_TOP];
 504                 unsigned int    ui[BN_NIST_224_TOP*sizeof(BN_ULONG)/sizeof(unsigned int)];
 505                 } buf;
 506         BN_ULONG c_d[BN_NIST_224_TOP],
 507                 *res;
 508         PTR_SIZE_INT mask;
 509         union { bn_addsub_f f; PTR_SIZE_INT p; } u;
 510         static const BIGNUM _bignum_nist_p_224_sqr = {
 511                 (BN_ULONG *)_nist_p_224_sqr,
 512                 sizeof(_nist_p_224_sqr)/sizeof(_nist_p_224_sqr[0]),
 513                 sizeof(_nist_p_224_sqr)/sizeof(_nist_p_224_sqr[0]),
 514                 0,BN_FLG_STATIC_DATA };
 515
 516
 517         field = &_bignum_nist_p_224; /* just to make sure */
 518
 519         if (BN_is_negative(a) || BN_ucmp(a,&_bignum_nist_p_224_sqr)>=0)
 520                 return BN_nnmod(r, a, field, ctx);
 521
 522         i = BN_ucmp(field, a);
 523         if (i == 0)
 524                 {
 525                 BN_zero(r);
 526                 return 1;
 527                 }
 528         else if (i > 0)
 529                 return (r == a)? 1 : (BN_copy(r ,a) != NULL);
 530
 531         if (r != a)
 532                 {
 533                 if (!bn_wexpand(r, BN_NIST_224_TOP))
 534                         return 0;
 535                 r_d = r->d;
 536                 nist_cp_bn(r_d, a_d, BN_NIST_224_TOP);
 537                 }
 538         else
 539                 r_d = a_d;
 540
 541 #if BN_BITS2==64
 542         /* copy upper 256 bits of 448 bit number ... */
 543         nist_cp_bn_0(c_d, a_d + (BN_NIST_224_TOP-1), top - (BN_NIST_224_TOP-1), BN_NIST_224_TOP);
 544         /* ... and right shift by 32 to obtain upper 224 bits */
 545         nist_set_224(buf.bn, c_d, 14, 13, 12, 11, 10, 9, 8);
 546         /* truncate lower part to 224 bits too */
 547         r_d[BN_NIST_224_TOP-1] &= BN_MASK2l;
 548 #else
 549         nist_cp_bn_0(buf.bn, a_d + BN_NIST_224_TOP, top - BN_NIST_224_TOP, BN_NIST_224_TOP);
 550 #endif
 551
 552 #if defined(NIST_INT64) && BN_BITS2!=64
 553         {
 554         NIST_INT64              acc;    /* accumulator */
 555         unsigned int            *rp=(unsigned int *)r_d;
 556         const unsigned int      *bp=(const unsigned int *)buf.ui;
 557
 558         acc  = rp[0];   acc -= bp[7-7];
 559                         acc -= bp[11-7]; rp[0] = (unsigned int)acc; acc >>= 32;
 560
 561         acc += rp[1];   acc -= bp[8-7];
 562                         acc -= bp[12-7]; rp[1] = (unsigned int)acc; acc >>= 32;
 563
 564         acc += rp[2];   acc -= bp[9-7];
 565                         acc -= bp[13-7]; rp[2] = (unsigned int)acc; acc >>= 32;
 566
 567         acc += rp[3];   acc += bp[7-7];
 568                         acc += bp[11-7];
 569                         acc -= bp[10-7]; rp[3] = (unsigned int)acc; acc>>= 32;
 570
 571         acc += rp[4];   acc += bp[8-7];
 572                         acc += bp[12-7];
 573                         acc -= bp[11-7]; rp[4] = (unsigned int)acc; acc >>= 32;
 574
 575         acc += rp[5];   acc += bp[9-7];
 576                         acc += bp[13-7];
 577                         acc -= bp[12-7]; rp[5] = (unsigned int)acc; acc >>= 32;
 578
 579         acc += rp[6];   acc += bp[10-7];
 580                         acc -= bp[13-7]; rp[6] = (unsigned int)acc;
 581
 582         carry = (int)(acc>>32);
 583 # if BN_BITS2==64
 584         rp[7] = carry;
 585 # endif
 586         }
 587 #else
 588         {
 589         BN_ULONG t_d[BN_NIST_224_TOP];
 590
 591         nist_set_224(t_d, buf.bn, 10, 9, 8, 7, 0, 0, 0);
 592         carry = (int)bn_add_words(r_d, r_d, t_d, BN_NIST_224_TOP);
 593         nist_set_224(t_d, buf.bn, 0, 13, 12, 11, 0, 0, 0);
 594         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_224_TOP);
 595         nist_set_224(t_d, buf.bn, 13, 12, 11, 10, 9, 8, 7);
 596         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_224_TOP);
 597         nist_set_224(t_d, buf.bn, 0, 0, 0, 0, 13, 12, 11);
 598         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_224_TOP);
 599
 600 #if BN_BITS2==64
 601         carry = (int)(r_d[BN_NIST_224_TOP-1]>>32);
 602 #endif
 603         }
 604 #endif
 605         u.f = bn_sub_words;
 606         if (carry > 0)
 607                 {
 608                 carry = (int)bn_sub_words(r_d,r_d,_nist_p_224[carry-1],BN_NIST_224_TOP);
 609 #if BN_BITS2==64
 610                 carry=(int)(~(r_d[BN_NIST_224_TOP-1]>>32))&1;
 611 #endif
 612                 }
 613         else if (carry < 0)
 614                 {
 615                 /* it's a bit more comlicated logic in this case.
 616                  * if bn_add_words yields no carry, then result
 617                  * has to be adjusted by unconditionally *adding*
 618                  * the modulus. but if it does, then result has
 619                  * to be compared to the modulus and conditionally
 620                  * adjusted by *subtracting* the latter. */
 621                 carry = (int)bn_add_words(r_d,r_d,_nist_p_224[-carry-1],BN_NIST_224_TOP);
 622                 mask = 0-(PTR_SIZE_INT)carry;
 623                 u.p = ((PTR_SIZE_INT)bn_sub_words&mask) |
 624                  ((PTR_SIZE_INT)bn_add_words&~mask);
 625                 }
 626         else
 627                 carry = 1;
 628
 629         /* otherwise it's effectively same as in BN_nist_mod_192... */
 630         mask  = 0-(PTR_SIZE_INT)(*u.f)(c_d,r_d,_nist_p_224[0],BN_NIST_224_TOP);
 631         mask &= 0-(PTR_SIZE_INT)carry;
 632         res   = c_d;
 633         res   = (BN_ULONG *)(((PTR_SIZE_INT)res&~mask) |
 634          ((PTR_SIZE_INT)r_d&mask));
 635         nist_cp_bn(r_d, res, BN_NIST_224_TOP);
 636         r->top = BN_NIST_224_TOP;
 637         bn_correct_top(r);
 638
 639         return 1;
 640         }
 641
 642 #define nist_set_256(to, from, a1, a2, a3, a4, a5, a6, a7, a8) \
 643         { \
 644         bn_cp_32(to, 0, from, (a8) - 8) \
 645         bn_cp_32(to, 1, from, (a7) - 8) \
 646         bn_cp_32(to, 2, from, (a6) - 8) \
 647         bn_cp_32(to, 3, from, (a5) - 8) \
 648         bn_cp_32(to, 4, from, (a4) - 8) \
 649         bn_cp_32(to, 5, from, (a3) - 8) \
 650         bn_cp_32(to, 6, from, (a2) - 8) \
 651         bn_cp_32(to, 7, from, (a1) - 8) \
 652         }
 653
 654 int BN_nist_mod_256(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 655         BN_CTX *ctx)
 656         {
 657         int     i, top = a->top;
 658         int     carry = 0;
 659         register BN_ULONG *a_d = a->d, *r_d;
 660         union   {
 661                 BN_ULONG bn[BN_NIST_256_TOP];
 662                 unsigned int ui[BN_NIST_256_TOP*sizeof(BN_ULONG)/sizeof(unsigned int)];
 663                 } buf;
 664         BN_ULONG c_d[BN_NIST_256_TOP],
 665                 *res;
 666         PTR_SIZE_INT mask;
 667         union { bn_addsub_f f; PTR_SIZE_INT p; } u;
 668         static const BIGNUM _bignum_nist_p_256_sqr = {
 669                 (BN_ULONG *)_nist_p_256_sqr,
 670                 sizeof(_nist_p_256_sqr)/sizeof(_nist_p_256_sqr[0]),
 671                 sizeof(_nist_p_256_sqr)/sizeof(_nist_p_256_sqr[0]),
 672                 0,BN_FLG_STATIC_DATA };
 673
 674         field = &_bignum_nist_p_256; /* just to make sure */
 675
 676         if (BN_is_negative(a) || BN_ucmp(a,&_bignum_nist_p_256_sqr)>=0)
 677                 return BN_nnmod(r, a, field, ctx);
 678
 679         i = BN_ucmp(field, a);
 680         if (i == 0)
 681                 {
 682                 BN_zero(r);
 683                 return 1;
 684                 }
 685         else if (i > 0)
 686                 return (r == a)? 1 : (BN_copy(r ,a) != NULL);
 687
 688         if (r != a)
 689                 {
 690                 if (!bn_wexpand(r, BN_NIST_256_TOP))
 691                         return 0;
 692                 r_d = r->d;
 693                 nist_cp_bn(r_d, a_d, BN_NIST_256_TOP);
 694                 }
 695         else
 696                 r_d = a_d;
 697
 698         nist_cp_bn_0(buf.bn, a_d + BN_NIST_256_TOP, top - BN_NIST_256_TOP, BN_NIST_256_TOP);
 699
 700 #if defined(NIST_INT64)
 701         {
 702         NIST_INT64              acc;    /* accumulator */
 703         unsigned int            *rp=(unsigned int *)r_d;
 704         const unsigned int      *bp=(const unsigned int *)buf.ui;
 705
 706         acc = rp[0];    acc += bp[8-8];
 707                         acc += bp[9-8];
 708                         acc -= bp[11-8];
 709                         acc -= bp[12-8];
 710                         acc -= bp[13-8];
 711                         acc -= bp[14-8]; rp[0] = (unsigned int)acc; acc >>= 32;
 712
 713         acc += rp[1];   acc += bp[9-8];
 714                         acc += bp[10-8];
 715                         acc -= bp[12-8];
 716                         acc -= bp[13-8];
 717                         acc -= bp[14-8];
 718                         acc -= bp[15-8]; rp[1] = (unsigned int)acc; acc >>= 32;
 719
 720         acc += rp[2];   acc += bp[10-8];
 721                         acc += bp[11-8];
 722                         acc -= bp[13-8];
 723                         acc -= bp[14-8];
 724                         acc -= bp[15-8]; rp[2] = (unsigned int)acc; acc >>= 32;
 725
 726         acc += rp[3];   acc += bp[11-8];
 727                         acc += bp[11-8];
 728                         acc += bp[12-8];
 729                         acc += bp[12-8];
 730                         acc += bp[13-8];
 731                         acc -= bp[15-8];
 732                         acc -= bp[8-8];
 733                         acc -= bp[9-8];  rp[3] = (unsigned int)acc; acc >>= 32;
 734
 735         acc += rp[4];   acc += bp[12-8];
 736                         acc += bp[12-8];
 737                         acc += bp[13-8];
 738                         acc += bp[13-8];
 739                         acc += bp[14-8];
 740                         acc -= bp[9-8];
 741                         acc -= bp[10-8]; rp[4] = (unsigned int)acc; acc >>= 32;
 742
 743         acc += rp[5];   acc += bp[13-8];
 744                         acc += bp[13-8];
 745                         acc += bp[14-8];
 746                         acc += bp[14-8];
 747                         acc += bp[15-8];
 748                         acc -= bp[10-8];
 749                         acc -= bp[11-8]; rp[5] = (unsigned int)acc; acc >>= 32;
 750
 751         acc += rp[6];   acc += bp[14-8];
 752                         acc += bp[14-8];
 753                         acc += bp[15-8];
 754                         acc += bp[15-8];
 755                         acc += bp[14-8];
 756                         acc += bp[13-8];
 757                         acc -= bp[8-8];
 758                         acc -= bp[9-8];  rp[6] = (unsigned int)acc; acc >>= 32;
 759
 760         acc += rp[7];   acc += bp[15-8];
 761                         acc += bp[15-8];
 762                         acc += bp[15-8];
 763                         acc += bp[8 -8];
 764                         acc -= bp[10-8];
 765                         acc -= bp[11-8];
 766                         acc -= bp[12-8];
 767                         acc -= bp[13-8]; rp[7] = (unsigned int)acc;
 768
 769         carry = (int)(acc>>32);
 770         }
 771 #else
 772         {
 773         BN_ULONG t_d[BN_NIST_256_TOP];
 774
 775         /*S1*/
 776         nist_set_256(t_d, buf.bn, 15, 14, 13, 12, 11, 0, 0, 0);
 777         /*S2*/
 778         nist_set_256(c_d, buf.bn, 0, 15, 14, 13, 12, 0, 0, 0);
 779         carry = (int)bn_add_words(t_d, t_d, c_d, BN_NIST_256_TOP);
 780         /* left shift */
 781                 {
 782                 register BN_ULONG *ap,t,c;
 783                 ap = t_d;
 784                 c=0;
 785                 for (i = BN_NIST_256_TOP; i != 0; --i)
 786                         {
 787                         t= *ap;
 788                         *(ap++)=((t<<1)|c)&BN_MASK2;
 789                         c=(t & BN_TBIT)?1:0;
 790                         }
 791                 carry <<= 1;
 792                 carry  |= c;
 793                 }
 794         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 795         /*S3*/
 796         nist_set_256(t_d, buf.bn, 15, 14, 0, 0, 0, 10, 9, 8);
 797         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 798         /*S4*/
 799         nist_set_256(t_d, buf.bn, 8, 13, 15, 14, 13, 11, 10, 9);
 800         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 801         /*D1*/
 802         nist_set_256(t_d, buf.bn, 10, 8, 0, 0, 0, 13, 12, 11);
 803         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 804         /*D2*/
 805         nist_set_256(t_d, buf.bn, 11, 9, 0, 0, 15, 14, 13, 12);
 806         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 807         /*D3*/
 808         nist_set_256(t_d, buf.bn, 12, 0, 10, 9, 8, 15, 14, 13);
 809         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 810         /*D4*/
 811         nist_set_256(t_d, buf.bn, 13, 0, 11, 10, 9, 0, 15, 14);
 812         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 813
 814         }
 815 #endif
 816         /* see BN_nist_mod_224 for explanation */
 817         u.f = bn_sub_words;
 818         if (carry > 0)
 819                 carry = (int)bn_sub_words(r_d,r_d,_nist_p_256[carry-1],BN_NIST_256_TOP);
 820         else if (carry < 0)
 821                 {
 822                 carry = (int)bn_add_words(r_d,r_d,_nist_p_256[-carry-1],BN_NIST_256_TOP);
 823                 mask = 0-(PTR_SIZE_INT)carry;
 824                 u.p = ((PTR_SIZE_INT)bn_sub_words&mask) |
 825                  ((PTR_SIZE_INT)bn_add_words&~mask);
 826                 }
 827         else
 828                 carry = 1;
 829
 830         mask  = 0-(PTR_SIZE_INT)(*u.f)(c_d,r_d,_nist_p_256[0],BN_NIST_256_TOP);
 831         mask &= 0-(PTR_SIZE_INT)carry;
 832         res   = c_d;
 833         res   = (BN_ULONG *)(((PTR_SIZE_INT)res&~mask) |
 834          ((PTR_SIZE_INT)r_d&mask));
 835         nist_cp_bn(r_d, res, BN_NIST_256_TOP);
 836         r->top = BN_NIST_256_TOP;
 837         bn_correct_top(r);
 838
 839         return 1;
 840         }
 841
 842 #define nist_set_384(to,from,a1,a2,a3,a4,a5,a6,a7,a8,a9,a10,a11,a12) \
 843         { \
 844         bn_cp_32(to, 0, from,  (a12) - 12) \
 845         bn_cp_32(to, 1, from,  (a11) - 12) \
 846         bn_cp_32(to, 2, from,  (a10) - 12) \
 847         bn_cp_32(to, 3, from,  (a9) - 12)  \
 848         bn_cp_32(to, 4, from,  (a8) - 12)  \
 849         bn_cp_32(to, 5, from,  (a7) - 12)  \
 850         bn_cp_32(to, 6, from,  (a6) - 12)  \
 851         bn_cp_32(to, 7, from,  (a5) - 12)  \
 852         bn_cp_32(to, 8, from,  (a4) - 12)  \
 853         bn_cp_32(to, 9, from,  (a3) - 12)  \
 854         bn_cp_32(to, 10, from, (a2) - 12)  \
 855         bn_cp_32(to, 11, from, (a1) - 12)  \
 856         }
 857
 858 int BN_nist_mod_384(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 859         BN_CTX *ctx)
 860         {
 861         int     i, top = a->top;
 862         int     carry = 0;
 863         register BN_ULONG *r_d, *a_d = a->d;
 864         union   {
 865                 BN_ULONG bn[BN_NIST_384_TOP];
 866                 unsigned int ui[BN_NIST_384_TOP*sizeof(BN_ULONG)/sizeof(unsigned int)];
 867                 } buf;
 868         BN_ULONG c_d[BN_NIST_384_TOP],
 869                 *res;
 870         PTR_SIZE_INT mask;
 871         union { bn_addsub_f f; PTR_SIZE_INT p; } u;
 872         static const BIGNUM _bignum_nist_p_384_sqr = {
 873                 (BN_ULONG *)_nist_p_384_sqr,
 874                 sizeof(_nist_p_384_sqr)/sizeof(_nist_p_384_sqr[0]),
 875                 sizeof(_nist_p_384_sqr)/sizeof(_nist_p_384_sqr[0]),
 876                 0,BN_FLG_STATIC_DATA };
 877
 878
 879         field = &_bignum_nist_p_384; /* just to make sure */
 880
 881         if (BN_is_negative(a) || BN_ucmp(a,&_bignum_nist_p_384_sqr)>=0)
 882                 return BN_nnmod(r, a, field, ctx);
 883
 884         i = BN_ucmp(field, a);
 885         if (i == 0)
 886                 {
 887                 BN_zero(r);
 888                 return 1;
 889                 }
 890         else if (i > 0)
 891                 return (r == a)? 1 : (BN_copy(r ,a) != NULL);
 892
 893         if (r != a)
 894                 {
 895                 if (!bn_wexpand(r, BN_NIST_384_TOP))
 896                         return 0;
 897                 r_d = r->d;
 898                 nist_cp_bn(r_d, a_d, BN_NIST_384_TOP);
 899                 }
 900         else
 901                 r_d = a_d;
 902
 903         nist_cp_bn_0(buf.bn, a_d + BN_NIST_384_TOP, top - BN_NIST_384_TOP, BN_NIST_384_TOP);
 904
 905 #if defined(NIST_INT64)
 906         {
 907         NIST_INT64              acc;    /* accumulator */
 908         unsigned int            *rp=(unsigned int *)r_d;
 909         const unsigned int      *bp=(const unsigned int *)buf.ui;
 910
 911         acc = rp[0];    acc += bp[12-12];
 912                         acc += bp[21-12];
 913                         acc += bp[20-12];
 914                         acc -= bp[23-12]; rp[0] = (unsigned int)acc; acc >>= 32;
 915
 916         acc += rp[1];   acc += bp[13-12];
 917                         acc += bp[22-12];
 918                         acc += bp[23-12];
 919                         acc -= bp[12-12];
 920                         acc -= bp[20-12]; rp[1] = (unsigned int)acc; acc >>= 32;
 921
 922         acc += rp[2];   acc += bp[14-12];
 923                         acc += bp[23-12];
 924                         acc -= bp[13-12];
 925                         acc -= bp[21-12]; rp[2] = (unsigned int)acc; acc >>= 32;
 926
 927         acc += rp[3];   acc += bp[15-12];
 928                         acc += bp[12-12];
 929                         acc += bp[20-12];
 930                         acc += bp[21-12];
 931                         acc -= bp[14-12];
 932                         acc -= bp[22-12];
 933                         acc -= bp[23-12]; rp[3] = (unsigned int)acc; acc >>= 32;
 934
 935         acc += rp[4];   acc += bp[21-12];
 936                         acc += bp[21-12];
 937                         acc += bp[16-12];
 938                         acc += bp[13-12];
 939                         acc += bp[12-12];
 940                         acc += bp[20-12];
 941                         acc += bp[22-12];
 942                         acc -= bp[15-12];
 943                         acc -= bp[23-12];
 944                         acc -= bp[23-12]; rp[4] = (unsigned int)acc; acc >>= 32;
 945
 946         acc += rp[5];   acc += bp[22-12];
 947                         acc += bp[22-12];
 948                         acc += bp[17-12];
 949                         acc += bp[14-12];
 950                         acc += bp[13-12];
 951                         acc += bp[21-12];
 952                         acc += bp[23-12];
 953                         acc -= bp[16-12]; rp[5] = (unsigned int)acc; acc >>= 32;
 954
 955         acc += rp[6];   acc += bp[23-12];
 956                         acc += bp[23-12];
 957                         acc += bp[18-12];
 958                         acc += bp[15-12];
 959                         acc += bp[14-12];
 960                         acc += bp[22-12];
 961                         acc -= bp[17-12]; rp[6] = (unsigned int)acc; acc >>= 32;
 962
 963         acc += rp[7];   acc += bp[19-12];
 964                         acc += bp[16-12];
 965                         acc += bp[15-12];
 966                         acc += bp[23-12];
 967                         acc -= bp[18-12]; rp[7] = (unsigned int)acc; acc >>= 32;
 968
 969         acc += rp[8];   acc += bp[20-12];
 970                         acc += bp[17-12];
 971                         acc += bp[16-12];
 972                         acc -= bp[19-12]; rp[8] = (unsigned int)acc; acc >>= 32;
 973
 974         acc += rp[9];   acc += bp[21-12];
 975                         acc += bp[18-12];
 976                         acc += bp[17-12];
 977                         acc -= bp[20-12]; rp[9] = (unsigned int)acc; acc >>= 32;
 978
 979         acc += rp[10];  acc += bp[22-12];
 980                         acc += bp[19-12];
 981                         acc += bp[18-12];
 982                         acc -= bp[21-12]; rp[10] = (unsigned int)acc; acc >>= 32;
 983
 984         acc += rp[11];  acc += bp[23-12];
 985                         acc += bp[20-12];
 986                         acc += bp[19-12];
 987                         acc -= bp[22-12]; rp[11] = (unsigned int)acc;
 988
 989         carry = (int)(acc>>32);
 990         }
 991 #else
 992         {
 993         BN_ULONG t_d[BN_NIST_384_TOP];
 994
 995         /*S1*/
 996         nist_set_256(t_d, buf.bn, 0, 0, 0, 0, 0, 23-4, 22-4, 21-4);
 997                 /* left shift */
 998                 {
 999                 register BN_ULONG *ap,t,c;
1000                 ap = t_d;
1001                 c=0;
1002                 for (i = 3; i != 0; --i)
1003                         {
1004                         t= *ap;
1005                         *(ap++)=((t<<1)|c)&BN_MASK2;
1006                         c=(t & BN_TBIT)?1:0;
1007                         }
1008                 *ap=c;
1009                 }
1010         carry = (int)bn_add_words(r_d+(128/BN_BITS2), r_d+(128/BN_BITS2),
1011                 t_d, BN_NIST_256_TOP);
1012         /*S2 */
1013         carry += (int)bn_add_words(r_d, r_d, buf.bn, BN_NIST_384_TOP);
1014         /*S3*/
1015         nist_set_384(t_d,buf.bn,20,19,18,17,16,15,14,13,12,23,22,21);
1016         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
1017         /*S4*/
1018         nist_set_384(t_d,buf.bn,19,18,17,16,15,14,13,12,20,0,23,0);
1019         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
1020         /*S5*/
1021         nist_set_384(t_d, buf.bn,0,0,0,0,23,22,21,20,0,0,0,0);
1022         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
1023         /*S6*/
1024         nist_set_384(t_d,buf.bn,0,0,0,0,0,0,23,22,21,0,0,20);
1025         carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
1026         /*D1*/
1027         nist_set_384(t_d,buf.bn,22,21,20,19,18,17,16,15,14,13,12,23);
1028         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
1029         /*D2*/
1030         nist_set_384(t_d,buf.bn,0,0,0,0,0,0,0,23,22,21,20,0);
1031         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
1032         /*D3*/
1033         nist_set_384(t_d,buf.bn,0,0,0,0,0,0,0,23,23,0,0,0);
1034         carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
1035
1036         }
1037 #endif
1038         /* see BN_nist_mod_224 for explanation */
1039         u.f = bn_sub_words;
1040         if (carry > 0)
1041                 carry = (int)bn_sub_words(r_d,r_d,_nist_p_384[carry-1],BN_NIST_384_TOP);
1042         else if (carry < 0)
1043                 {
1044                 carry = (int)bn_add_words(r_d,r_d,_nist_p_384[-carry-1],BN_NIST_384_TOP);
1045                 mask = 0-(PTR_SIZE_INT)carry;
1046                 u.p = ((PTR_SIZE_INT)bn_sub_words&mask) |
1047                  ((PTR_SIZE_INT)bn_add_words&~mask);
1048                 }
1049         else
1050                 carry = 1;
1051
1052         mask  = 0-(PTR_SIZE_INT)(*u.f)(c_d,r_d,_nist_p_384[0],BN_NIST_384_TOP);
1053         mask &= 0-(PTR_SIZE_INT)carry;
1054         res   = c_d;
1055         res   = (BN_ULONG *)(((PTR_SIZE_INT)res&~mask) |
1056          ((PTR_SIZE_INT)r_d&mask));
1057         nist_cp_bn(r_d, res, BN_NIST_384_TOP);
1058         r->top = BN_NIST_384_TOP;
1059         bn_correct_top(r);
1060
1061         return 1;
1062         }
1063
1064 #define BN_NIST_521_RSHIFT      (521%BN_BITS2)
1065 #define BN_NIST_521_LSHIFT      (BN_BITS2-BN_NIST_521_RSHIFT)
1066 #define BN_NIST_521_TOP_MASK    ((BN_ULONG)BN_MASK2>>BN_NIST_521_LSHIFT)
1067
1068 int BN_nist_mod_521(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
1069         BN_CTX *ctx)
1070         {
1071         int     top = a->top, i;
1072         BN_ULONG *r_d, *a_d = a->d,
1073                  t_d[BN_NIST_521_TOP],
1074                  val,tmp,*res;
1075         PTR_SIZE_INT mask;
1076         static const BIGNUM _bignum_nist_p_521_sqr = {
1077                 (BN_ULONG *)_nist_p_521_sqr,
1078                 sizeof(_nist_p_521_sqr)/sizeof(_nist_p_521_sqr[0]),
1079                 sizeof(_nist_p_521_sqr)/sizeof(_nist_p_521_sqr[0]),
1080                 0,BN_FLG_STATIC_DATA };
1081
1082         field = &_bignum_nist_p_521; /* just to make sure */
1083
1084         if (BN_is_negative(a) || BN_ucmp(a,&_bignum_nist_p_521_sqr)>=0)
1085                 return BN_nnmod(r, a, field, ctx);
1086
1087         i = BN_ucmp(field, a);
1088         if (i == 0)
1089                 {
1090                 BN_zero(r);
1091                 return 1;
1092                 }
1093         else if (i > 0)
1094                 return (r == a)? 1 : (BN_copy(r ,a) != NULL);
1095
1096         if (r != a)
1097                 {
1098                 if (!bn_wexpand(r,BN_NIST_521_TOP))
1099                         return 0;
1100                 r_d = r->d;
1101                 nist_cp_bn(r_d,a_d, BN_NIST_521_TOP);
1102                 }
1103         else
1104                 r_d = a_d;
1105
1106         /* upper 521 bits, copy ... */
1107         nist_cp_bn_0(t_d,a_d + (BN_NIST_521_TOP-1), top - (BN_NIST_521_TOP-1),BN_NIST_521_TOP);
1108         /* ... and right shift */
1109         for (val=t_d[0],i=0; i<BN_NIST_521_TOP-1; i++)
1110                 {
1111                 tmp = val>>BN_NIST_521_RSHIFT;
1112                 val = t_d[i+1];
1113                 t_d[i] = (tmp | val<<BN_NIST_521_LSHIFT) & BN_MASK2;
1114                 }
1115         t_d[i] = val>>BN_NIST_521_RSHIFT;
1116         /* lower 521 bits */
1117         r_d[i] &= BN_NIST_521_TOP_MASK;
1118
1119         bn_add_words(r_d,r_d,t_d,BN_NIST_521_TOP);
1120         mask = 0-(PTR_SIZE_INT)bn_sub_words(t_d,r_d,_nist_p_521,BN_NIST_521_TOP);
1121         res  = t_d;
1122         res  = (BN_ULONG *)(((PTR_SIZE_INT)res&~mask) |
1123          ((PTR_SIZE_INT)r_d&mask));
1124         nist_cp_bn(r_d,res,BN_NIST_521_TOP);
1125         r->top = BN_NIST_521_TOP;
1126         bn_correct_top(r);
1127
1128         return 1;
1129         }
1130
1131 int (*BN_nist_mod_func(const BIGNUM *p))(BIGNUM *r, const BIGNUM *a, const BIGNUM *field, BN_CTX *ctx)
1132         {
1133         if (BN_ucmp(&_bignum_nist_p_192, p) == 0)
1134                 return BN_nist_mod_192;
1135         if (BN_ucmp(&_bignum_nist_p_224, p) == 0)
1136                 return BN_nist_mod_224;
1137         if (BN_ucmp(&_bignum_nist_p_256, p) == 0)
1138                 return BN_nist_mod_256;
1139         if (BN_ucmp(&_bignum_nist_p_384, p) == 0)
1140                 return BN_nist_mod_384;
1141         if (BN_ucmp(&_bignum_nist_p_521, p) == 0)
1142                 return BN_nist_mod_521;
1143         return 0;
1144         }