crypto/bn/bn_local.h

   1 /*
   2  * Copyright 1995-2019 The OpenSSL Project Authors. All Rights Reserved.
   3  *
   4  * Licensed under the Apache License 2.0 (the "License").  You may not use
   5  * this file except in compliance with the License.  You can obtain a copy
   6  * in the file LICENSE in the source distribution or at
   7  * https://www.openssl.org/source/license.html
   8  */
   9
  10 #ifndef OSSL_CRYPTO_BN_LOCAL_H
  11 # define OSSL_CRYPTO_BN_LOCAL_H
  12
  13 /*
  14  * The EDK2 build doesn't use bn_conf.h; it sets THIRTY_TWO_BIT or
  15  * SIXTY_FOUR_BIT in its own environment since it doesn't re-run our
  16  * Configure script and needs to support both 32-bit and 64-bit.
  17  */
  18 # include <openssl/opensslconf.h>
  19
  20 # if !defined(OPENSSL_SYS_UEFI)
  21 #  include "crypto/bn_conf.h"
  22 # endif
  23
  24 # include "crypto/bn.h"
  25
  26 /*
  27  * These preprocessor symbols control various aspects of the bignum headers
  28  * and library code. They're not defined by any "normal" configuration, as
  29  * they are intended for development and testing purposes. NB: defining all
  30  * three can be useful for debugging application code as well as openssl
  31  * itself. BN_DEBUG - turn on various debugging alterations to the bignum
  32  * code BN_DEBUG_RAND - uses random poisoning of unused words to trip up
  33  * mismanagement of bignum internals. You must also define BN_DEBUG.
  34  */
  35 /* #define BN_DEBUG */
  36 /* #define BN_DEBUG_RAND */
  37
  38 # ifndef OPENSSL_SMALL_FOOTPRINT
  39 #  define BN_MUL_COMBA
  40 #  define BN_SQR_COMBA
  41 #  define BN_RECURSION
  42 # endif
  43
  44 /*
  45  * This next option uses the C libraries (2 word)/(1 word) function. If it is
  46  * not defined, I use my C version (which is slower). The reason for this
  47  * flag is that when the particular C compiler library routine is used, and
  48  * the library is linked with a different compiler, the library is missing.
  49  * This mostly happens when the library is built with gcc and then linked
  50  * using normal cc.  This would be a common occurrence because gcc normally
  51  * produces code that is 2 times faster than system compilers for the big
  52  * number stuff. For machines with only one compiler (or shared libraries),
  53  * this should be on.  Again this in only really a problem on machines using
  54  * "long long's", are 32bit, and are not using my assembler code.
  55  */
  56 # if defined(OPENSSL_SYS_MSDOS) || defined(OPENSSL_SYS_WINDOWS) || \
  57     defined(OPENSSL_SYS_WIN32) || defined(linux)
  58 #  define BN_DIV2W
  59 # endif
  60
  61 /*
  62  * 64-bit processor with LP64 ABI
  63  */
  64 # ifdef SIXTY_FOUR_BIT_LONG
  65 #  define BN_ULLONG       unsigned long long
  66 #  define BN_BITS4        32
  67 #  define BN_MASK2        (0xffffffffffffffffL)
  68 #  define BN_MASK2l       (0xffffffffL)
  69 #  define BN_MASK2h       (0xffffffff00000000L)
  70 #  define BN_MASK2h1      (0xffffffff80000000L)
  71 #  define BN_DEC_CONV     (10000000000000000000UL)
  72 #  define BN_DEC_NUM      19
  73 #  define BN_DEC_FMT1     "%lu"
  74 #  define BN_DEC_FMT2     "%019lu"
  75 # endif
  76
  77 /*
  78  * 64-bit processor other than LP64 ABI
  79  */
  80 # ifdef SIXTY_FOUR_BIT
  81 #  undef BN_LLONG
  82 #  undef BN_ULLONG
  83 #  define BN_BITS4        32
  84 #  define BN_MASK2        (0xffffffffffffffffLL)
  85 #  define BN_MASK2l       (0xffffffffL)
  86 #  define BN_MASK2h       (0xffffffff00000000LL)
  87 #  define BN_MASK2h1      (0xffffffff80000000LL)
  88 #  define BN_DEC_CONV     (10000000000000000000ULL)
  89 #  define BN_DEC_NUM      19
  90 #  define BN_DEC_FMT1     "%llu"
  91 #  define BN_DEC_FMT2     "%019llu"
  92 # endif
  93
  94 # ifdef THIRTY_TWO_BIT
  95 #  ifdef BN_LLONG
  96 #   if defined(_WIN32) && !defined(__GNUC__)
  97 #    define BN_ULLONG     unsigned __int64
  98 #   else
  99 #    define BN_ULLONG     unsigned long long
 100 #   endif
 101 #  endif
 102 #  define BN_BITS4        16
 103 #  define BN_MASK2        (0xffffffffL)
 104 #  define BN_MASK2l       (0xffff)
 105 #  define BN_MASK2h1      (0xffff8000L)
 106 #  define BN_MASK2h       (0xffff0000L)
 107 #  define BN_DEC_CONV     (1000000000L)
 108 #  define BN_DEC_NUM      9
 109 #  define BN_DEC_FMT1     "%u"
 110 #  define BN_DEC_FMT2     "%09u"
 111 # endif
 112
 113
 114 /*-
 115  * Bignum consistency macros
 116  * There is one "API" macro, bn_fix_top(), for stripping leading zeroes from
 117  * bignum data after direct manipulations on the data. There is also an
 118  * "internal" macro, bn_check_top(), for verifying that there are no leading
 119  * zeroes. Unfortunately, some auditing is required due to the fact that
 120  * bn_fix_top() has become an overabused duct-tape because bignum data is
 121  * occasionally passed around in an inconsistent state. So the following
 122  * changes have been made to sort this out;
 123  * - bn_fix_top()s implementation has been moved to bn_correct_top()
 124  * - if BN_DEBUG isn't defined, bn_fix_top() maps to bn_correct_top(), and
 125  *   bn_check_top() is as before.
 126  * - if BN_DEBUG *is* defined;
 127  *   - bn_check_top() tries to pollute unused words even if the bignum 'top' is
 128  *     consistent. (ed: only if BN_DEBUG_RAND is defined)
 129  *   - bn_fix_top() maps to bn_check_top() rather than "fixing" anything.
 130  * The idea is to have debug builds flag up inconsistent bignums when they
 131  * occur. If that occurs in a bn_fix_top(), we examine the code in question; if
 132  * the use of bn_fix_top() was appropriate (ie. it follows directly after code
 133  * that manipulates the bignum) it is converted to bn_correct_top(), and if it
 134  * was not appropriate, we convert it permanently to bn_check_top() and track
 135  * down the cause of the bug. Eventually, no internal code should be using the
 136  * bn_fix_top() macro. External applications and libraries should try this with
 137  * their own code too, both in terms of building against the openssl headers
 138  * with BN_DEBUG defined *and* linking with a version of OpenSSL built with it
 139  * defined. This not only improves external code, it provides more test
 140  * coverage for openssl's own code.
 141  */
 142
 143 # ifdef BN_DEBUG
 144 /*
 145  * The new BN_FLG_FIXED_TOP flag marks vectors that were not treated with
 146  * bn_correct_top, in other words such vectors are permitted to have zeros
 147  * in most significant limbs. Such vectors are used internally to achieve
 148  * execution time invariance for critical operations with private keys.
 149  * It's BN_DEBUG-only flag, because user application is not supposed to
 150  * observe it anyway. Moreover, optimizing compiler would actually remove
 151  * all operations manipulating the bit in question in non-BN_DEBUG build.
 152  */
 153 #  define BN_FLG_FIXED_TOP 0x10000
 154 #  ifdef BN_DEBUG_RAND
 155 #   define bn_pollute(a) \
 156         do { \
 157             const BIGNUM *_bnum1 = (a); \
 158             if (_bnum1->top < _bnum1->dmax) { \
 159                 unsigned char _tmp_char; \
 160                 /* We cast away const without the compiler knowing, any \
 161                  * *genuinely* constant variables that aren't mutable \
 162                  * wouldn't be constructed with top!=dmax. */ \
 163                 BN_ULONG *_not_const; \
 164                 memcpy(&_not_const, &_bnum1->d, sizeof(_not_const)); \
 165                 RAND_bytes(&_tmp_char, 1); /* Debug only - safe to ignore error return */\
 166                 memset(_not_const + _bnum1->top, _tmp_char, \
 167                        sizeof(*_not_const) * (_bnum1->dmax - _bnum1->top)); \
 168             } \
 169         } while(0)
 170 #  else
 171 #   define bn_pollute(a)
 172 #  endif
 173 #  define bn_check_top(a) \
 174         do { \
 175                 const BIGNUM *_bnum2 = (a); \
 176                 if (_bnum2 != NULL) { \
 177                         int _top = _bnum2->top; \
 178                         (void)ossl_assert((_top == 0 && !_bnum2->neg) || \
 179                                   (_top && ((_bnum2->flags & BN_FLG_FIXED_TOP) \
 180                                             || _bnum2->d[_top - 1] != 0))); \
 181                         bn_pollute(_bnum2); \
 182                 } \
 183         } while(0)
 184
 185 #  define bn_fix_top(a)           bn_check_top(a)
 186
 187 #  define bn_check_size(bn, bits) bn_wcheck_size(bn, ((bits+BN_BITS2-1))/BN_BITS2)
 188 #  define bn_wcheck_size(bn, words) \
 189         do { \
 190                 const BIGNUM *_bnum2 = (bn); \
 191                 assert((words) <= (_bnum2)->dmax && \
 192                        (words) >= (_bnum2)->top); \
 193                 /* avoid unused variable warning with NDEBUG */ \
 194                 (void)(_bnum2); \
 195         } while(0)
 196
 197 # else                          /* !BN_DEBUG */
 198
 199 #  define BN_FLG_FIXED_TOP 0
 200 #  define bn_pollute(a)
 201 #  define bn_check_top(a)
 202 #  define bn_fix_top(a)           bn_correct_top(a)
 203 #  define bn_check_size(bn, bits)
 204 #  define bn_wcheck_size(bn, words)
 205
 206 # endif
 207
 208 BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num,
 209                           BN_ULONG w);
 210 BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w);
 211 void bn_sqr_words(BN_ULONG *rp, const BN_ULONG *ap, int num);
 212 BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d);
 213 BN_ULONG bn_add_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
 214                       int num);
 215 BN_ULONG bn_sub_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
 216                       int num);
 217
 218 struct bignum_st {
 219     BN_ULONG *d;                /* Pointer to an array of 'BN_BITS2' bit
 220                                  * chunks. */
 221     int top;                    /* Index of last used d +1. */
 222     /* The next are internal book keeping for bn_expand. */
 223     int dmax;                   /* Size of the d array. */
 224     int neg;                    /* one if the number is negative */
 225     int flags;
 226 };
 227
 228 /* Used for montgomery multiplication */
 229 struct bn_mont_ctx_st {
 230     int ri;                     /* number of bits in R */
 231     BIGNUM RR;                  /* used to convert to montgomery form,
 232                                    possibly zero-padded */
 233     BIGNUM N;                   /* The modulus */
 234     BIGNUM Ni;                  /* R*(1/R mod N) - N*Ni = 1 (Ni is only
 235                                  * stored for bignum algorithm) */
 236     BN_ULONG n0[2];             /* least significant word(s) of Ni; (type
 237                                  * changed with 0.9.9, was "BN_ULONG n0;"
 238                                  * before) */
 239     int flags;
 240 };
 241
 242 /*
 243  * Used for reciprocal division/mod functions It cannot be shared between
 244  * threads
 245  */
 246 struct bn_recp_ctx_st {
 247     BIGNUM N;                   /* the divisor */
 248     BIGNUM Nr;                  /* the reciprocal */
 249     int num_bits;
 250     int shift;
 251     int flags;
 252 };
 253
 254 /* Used for slow "generation" functions. */
 255 struct bn_gencb_st {
 256     unsigned int ver;           /* To handle binary (in)compatibility */
 257     void *arg;                  /* callback-specific data */
 258     union {
 259         /* if (ver==1) - handles old style callbacks */
 260         void (*cb_1) (int, int, void *);
 261         /* if (ver==2) - new callback style */
 262         int (*cb_2) (int, int, BN_GENCB *);
 263     } cb;
 264 };
 265
 266 /*-
 267  * BN_window_bits_for_exponent_size -- macro for sliding window mod_exp functions
 268  *
 269  *
 270  * For window size 'w' (w >= 2) and a random 'b' bits exponent,
 271  * the number of multiplications is a constant plus on average
 272  *
 273  *    2^(w-1) + (b-w)/(w+1);
 274  *
 275  * here  2^(w-1)  is for precomputing the table (we actually need
 276  * entries only for windows that have the lowest bit set), and
 277  * (b-w)/(w+1)  is an approximation for the expected number of
 278  * w-bit windows, not counting the first one.
 279  *
 280  * Thus we should use
 281  *
 282  *    w >= 6  if        b > 671
 283  *     w = 5  if  671 > b > 239
 284  *     w = 4  if  239 > b >  79
 285  *     w = 3  if   79 > b >  23
 286  *    w <= 2  if   23 > b
 287  *
 288  * (with draws in between).  Very small exponents are often selected
 289  * with low Hamming weight, so we use  w = 1  for b <= 23.
 290  */
 291 # define BN_window_bits_for_exponent_size(b) \
 292                 ((b) > 671 ? 6 : \
 293                  (b) > 239 ? 5 : \
 294                  (b) >  79 ? 4 : \
 295                  (b) >  23 ? 3 : 1)
 296
 297 /*
 298  * BN_mod_exp_mont_consttime is based on the assumption that the L1 data cache
 299  * line width of the target processor is at least the following value.
 300  */
 301 # define MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH      ( 64 )
 302 # define MOD_EXP_CTIME_MIN_CACHE_LINE_MASK       (MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH - 1)
 303
 304 /*
 305  * Window sizes optimized for fixed window size modular exponentiation
 306  * algorithm (BN_mod_exp_mont_consttime). To achieve the security goals of
 307  * BN_mode_exp_mont_consttime, the maximum size of the window must not exceed
 308  * log_2(MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH). Window size thresholds are
 309  * defined for cache line sizes of 32 and 64, cache line sizes where
 310  * log_2(32)=5 and log_2(64)=6 respectively. A window size of 7 should only be
 311  * used on processors that have a 128 byte or greater cache line size.
 312  */
 313 # if MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH == 64
 314
 315 #  define BN_window_bits_for_ctime_exponent_size(b) \
 316                 ((b) > 937 ? 6 : \
 317                  (b) > 306 ? 5 : \
 318                  (b) >  89 ? 4 : \
 319                  (b) >  22 ? 3 : 1)
 320 #  define BN_MAX_WINDOW_BITS_FOR_CTIME_EXPONENT_SIZE    (6)
 321
 322 # elif MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH == 32
 323
 324 #  define BN_window_bits_for_ctime_exponent_size(b) \
 325                 ((b) > 306 ? 5 : \
 326                  (b) >  89 ? 4 : \
 327                  (b) >  22 ? 3 : 1)
 328 #  define BN_MAX_WINDOW_BITS_FOR_CTIME_EXPONENT_SIZE    (5)
 329
 330 # endif
 331
 332 /* Pentium pro 16,16,16,32,64 */
 333 /* Alpha       16,16,16,16.64 */
 334 # define BN_MULL_SIZE_NORMAL                     (16)/* 32 */
 335 # define BN_MUL_RECURSIVE_SIZE_NORMAL            (16)/* 32 less than */
 336 # define BN_SQR_RECURSIVE_SIZE_NORMAL            (16)/* 32 */
 337 # define BN_MUL_LOW_RECURSIVE_SIZE_NORMAL        (32)/* 32 */
 338 # define BN_MONT_CTX_SET_SIZE_WORD               (64)/* 32 */
 339
 340 /*
 341  * 2011-02-22 SMS. In various places, a size_t variable or a type cast to
 342  * size_t was used to perform integer-only operations on pointers.  This
 343  * failed on VMS with 64-bit pointers (CC /POINTER_SIZE = 64) because size_t
 344  * is still only 32 bits.  What's needed in these cases is an integer type
 345  * with the same size as a pointer, which size_t is not certain to be. The
 346  * only fix here is VMS-specific.
 347  */
 348 # if defined(OPENSSL_SYS_VMS)
 349 #  if __INITIAL_POINTER_SIZE == 64
 350 #   define PTR_SIZE_INT long long
 351 #  else                         /* __INITIAL_POINTER_SIZE == 64 */
 352 #   define PTR_SIZE_INT int
 353 #  endif                        /* __INITIAL_POINTER_SIZE == 64 [else] */
 354 # elif !defined(PTR_SIZE_INT)   /* defined(OPENSSL_SYS_VMS) */
 355 #  define PTR_SIZE_INT size_t
 356 # endif                         /* defined(OPENSSL_SYS_VMS) [else] */
 357
 358 # if !defined(OPENSSL_NO_ASM) && !defined(OPENSSL_NO_INLINE_ASM) && !defined(PEDANTIC)
 359 /*
 360  * BN_UMULT_HIGH section.
 361  * If the compiler doesn't support 2*N integer type, then you have to
 362  * replace every N*N multiplication with 4 (N/2)*(N/2) accompanied by some
 363  * shifts and additions which unavoidably results in severe performance
 364  * penalties. Of course provided that the hardware is capable of producing
 365  * 2*N result... That's when you normally start considering assembler
 366  * implementation. However! It should be pointed out that some CPUs (e.g.,
 367  * PowerPC, Alpha, and IA-64) provide *separate* instruction calculating
 368  * the upper half of the product placing the result into a general
 369  * purpose register. Now *if* the compiler supports inline assembler,
 370  * then it's not impossible to implement the "bignum" routines (and have
 371  * the compiler optimize 'em) exhibiting "native" performance in C. That's
 372  * what BN_UMULT_HIGH macro is about:-) Note that more recent compilers do
 373  * support 2*64 integer type, which is also used here.
 374  */
 375 #  if defined(__SIZEOF_INT128__) && __SIZEOF_INT128__==16 && \
 376       (defined(SIXTY_FOUR_BIT) || defined(SIXTY_FOUR_BIT_LONG))
 377 #   define BN_UMULT_HIGH(a,b)          (((__uint128_t)(a)*(b))>>64)
 378 #   define BN_UMULT_LOHI(low,high,a,b) ({       \
 379         __uint128_t ret=(__uint128_t)(a)*(b);   \
 380         (high)=ret>>64; (low)=ret;      })
 381 #  elif defined(__alpha) && (defined(SIXTY_FOUR_BIT_LONG) || defined(SIXTY_FOUR_BIT))
 382 #   if defined(__DECC)
 383 #    include <c_asm.h>
 384 #    define BN_UMULT_HIGH(a,b)   (BN_ULONG)asm("umulh %a0,%a1,%v0",(a),(b))
 385 #   elif defined(__GNUC__) && __GNUC__>=2
 386 #    define BN_UMULT_HIGH(a,b)   ({     \
 387         register BN_ULONG ret;          \
 388         asm ("umulh     %1,%2,%0"       \
 389              : "=r"(ret)                \
 390              : "r"(a), "r"(b));         \
 391         ret;                      })
 392 #   endif                       /* compiler */
 393 #  elif defined(_ARCH_PPC64) && defined(SIXTY_FOUR_BIT_LONG)
 394 #   if defined(__GNUC__) && __GNUC__>=2
 395 #    define BN_UMULT_HIGH(a,b)   ({     \
 396         register BN_ULONG ret;          \
 397         asm ("mulhdu    %0,%1,%2"       \
 398              : "=r"(ret)                \
 399              : "r"(a), "r"(b));         \
 400         ret;                      })
 401 #   endif                       /* compiler */
 402 #  elif (defined(__x86_64) || defined(__x86_64__)) && \
 403        (defined(SIXTY_FOUR_BIT_LONG) || defined(SIXTY_FOUR_BIT))
 404 #   if defined(__GNUC__) && __GNUC__>=2
 405 #    define BN_UMULT_HIGH(a,b)   ({     \
 406         register BN_ULONG ret,discard;  \
 407         asm ("mulq      %3"             \
 408              : "=a"(discard),"=d"(ret)  \
 409              : "a"(a), "g"(b)           \
 410              : "cc");                   \
 411         ret;                      })
 412 #    define BN_UMULT_LOHI(low,high,a,b) \
 413         asm ("mulq      %3"             \
 414                 : "=a"(low),"=d"(high)  \
 415                 : "a"(a),"g"(b)         \
 416                 : "cc");
 417 #   endif
 418 #  elif (defined(_M_AMD64) || defined(_M_X64)) && defined(SIXTY_FOUR_BIT)
 419 #   if defined(_MSC_VER) && _MSC_VER>=1400
 420 unsigned __int64 __umulh(unsigned __int64 a, unsigned __int64 b);
 421 unsigned __int64 _umul128(unsigned __int64 a, unsigned __int64 b,
 422                           unsigned __int64 *h);
 423 #    pragma intrinsic(__umulh,_umul128)
 424 #    define BN_UMULT_HIGH(a,b)           __umulh((a),(b))
 425 #    define BN_UMULT_LOHI(low,high,a,b)  ((low)=_umul128((a),(b),&(high)))
 426 #   endif
 427 #  elif defined(__mips) && (defined(SIXTY_FOUR_BIT) || defined(SIXTY_FOUR_BIT_LONG))
 428 #   if defined(__GNUC__) && __GNUC__>=2
 429 #    define BN_UMULT_HIGH(a,b) ({       \
 430         register BN_ULONG ret;          \
 431         asm ("dmultu    %1,%2"          \
 432              : "=h"(ret)                \
 433              : "r"(a), "r"(b) : "l");   \
 434         ret;                    })
 435 #    define BN_UMULT_LOHI(low,high,a,b) \
 436         asm ("dmultu    %2,%3"          \
 437              : "=l"(low),"=h"(high)     \
 438              : "r"(a), "r"(b));
 439 #   endif
 440 #  elif defined(__aarch64__) && defined(SIXTY_FOUR_BIT_LONG)
 441 #   if defined(__GNUC__) && __GNUC__>=2
 442 #    define BN_UMULT_HIGH(a,b)   ({     \
 443         register BN_ULONG ret;          \
 444         asm ("umulh     %0,%1,%2"       \
 445              : "=r"(ret)                \
 446              : "r"(a), "r"(b));         \
 447         ret;                      })
 448 #   endif
 449 #  endif                        /* cpu */
 450 # endif                         /* OPENSSL_NO_ASM */
 451
 452 # ifdef BN_DEBUG_RAND
 453 #  define bn_clear_top2max(a) \
 454         { \
 455         int      ind = (a)->dmax - (a)->top; \
 456         BN_ULONG *ftl = &(a)->d[(a)->top-1]; \
 457         for (; ind != 0; ind--) \
 458                 *(++ftl) = 0x0; \
 459         }
 460 # else
 461 #  define bn_clear_top2max(a)
 462 # endif
 463
 464 # ifdef BN_LLONG
 465 /*******************************************************************
 466  * Using the long long type, has to be twice as wide as BN_ULONG...
 467  */
 468 #  define Lw(t)    (((BN_ULONG)(t))&BN_MASK2)
 469 #  define Hw(t)    (((BN_ULONG)((t)>>BN_BITS2))&BN_MASK2)
 470
 471 #  define mul_add(r,a,w,c) { \
 472         BN_ULLONG t; \
 473         t=(BN_ULLONG)w * (a) + (r) + (c); \
 474         (r)= Lw(t); \
 475         (c)= Hw(t); \
 476         }
 477
 478 #  define mul(r,a,w,c) { \
 479         BN_ULLONG t; \
 480         t=(BN_ULLONG)w * (a) + (c); \
 481         (r)= Lw(t); \
 482         (c)= Hw(t); \
 483         }
 484
 485 #  define sqr(r0,r1,a) { \
 486         BN_ULLONG t; \
 487         t=(BN_ULLONG)(a)*(a); \
 488         (r0)=Lw(t); \
 489         (r1)=Hw(t); \
 490         }
 491
 492 # elif defined(BN_UMULT_LOHI)
 493 #  define mul_add(r,a,w,c) {              \
 494         BN_ULONG high,low,ret,tmp=(a);  \
 495         ret =  (r);                     \
 496         BN_UMULT_LOHI(low,high,w,tmp);  \
 497         ret += (c);                     \
 498         (c) =  (ret<(c))?1:0;           \
 499         (c) += high;                    \
 500         ret += low;                     \
 501         (c) += (ret<low)?1:0;           \
 502         (r) =  ret;                     \
 503         }
 504
 505 #  define mul(r,a,w,c)    {               \
 506         BN_ULONG high,low,ret,ta=(a);   \
 507         BN_UMULT_LOHI(low,high,w,ta);   \
 508         ret =  low + (c);               \
 509         (c) =  high;                    \
 510         (c) += (ret<low)?1:0;           \
 511         (r) =  ret;                     \
 512         }
 513
 514 #  define sqr(r0,r1,a)    {               \
 515         BN_ULONG tmp=(a);               \
 516         BN_UMULT_LOHI(r0,r1,tmp,tmp);   \
 517         }
 518
 519 # elif defined(BN_UMULT_HIGH)
 520 #  define mul_add(r,a,w,c) {              \
 521         BN_ULONG high,low,ret,tmp=(a);  \
 522         ret =  (r);                     \
 523         high=  BN_UMULT_HIGH(w,tmp);    \
 524         ret += (c);                     \
 525         low =  (w) * tmp;               \
 526         (c) =  (ret<(c))?1:0;           \
 527         (c) += high;                    \
 528         ret += low;                     \
 529         (c) += (ret<low)?1:0;           \
 530         (r) =  ret;                     \
 531         }
 532
 533 #  define mul(r,a,w,c)    {               \
 534         BN_ULONG high,low,ret,ta=(a);   \
 535         low =  (w) * ta;                \
 536         high=  BN_UMULT_HIGH(w,ta);     \
 537         ret =  low + (c);               \
 538         (c) =  high;                    \
 539         (c) += (ret<low)?1:0;           \
 540         (r) =  ret;                     \
 541         }
 542
 543 #  define sqr(r0,r1,a)    {               \
 544         BN_ULONG tmp=(a);               \
 545         (r0) = tmp * tmp;               \
 546         (r1) = BN_UMULT_HIGH(tmp,tmp);  \
 547         }
 548
 549 # else
 550 /*************************************************************
 551  * No long long type
 552  */
 553
 554 #  define LBITS(a)        ((a)&BN_MASK2l)
 555 #  define HBITS(a)        (((a)>>BN_BITS4)&BN_MASK2l)
 556 #  define L2HBITS(a)      (((a)<<BN_BITS4)&BN_MASK2)
 557
 558 #  define LLBITS(a)       ((a)&BN_MASKl)
 559 #  define LHBITS(a)       (((a)>>BN_BITS2)&BN_MASKl)
 560 #  define LL2HBITS(a)     ((BN_ULLONG)((a)&BN_MASKl)<<BN_BITS2)
 561
 562 #  define mul64(l,h,bl,bh) \
 563         { \
 564         BN_ULONG m,m1,lt,ht; \
 565  \
 566         lt=l; \
 567         ht=h; \
 568         m =(bh)*(lt); \
 569         lt=(bl)*(lt); \
 570         m1=(bl)*(ht); \
 571         ht =(bh)*(ht); \
 572         m=(m+m1)&BN_MASK2; if (m < m1) ht+=L2HBITS((BN_ULONG)1); \
 573         ht+=HBITS(m); \
 574         m1=L2HBITS(m); \
 575         lt=(lt+m1)&BN_MASK2; if (lt < m1) ht++; \
 576         (l)=lt; \
 577         (h)=ht; \
 578         }
 579
 580 #  define sqr64(lo,ho,in) \
 581         { \
 582         BN_ULONG l,h,m; \
 583  \
 584         h=(in); \
 585         l=LBITS(h); \
 586         h=HBITS(h); \
 587         m =(l)*(h); \
 588         l*=l; \
 589         h*=h; \
 590         h+=(m&BN_MASK2h1)>>(BN_BITS4-1); \
 591         m =(m&BN_MASK2l)<<(BN_BITS4+1); \
 592         l=(l+m)&BN_MASK2; if (l < m) h++; \
 593         (lo)=l; \
 594         (ho)=h; \
 595         }
 596
 597 #  define mul_add(r,a,bl,bh,c) { \
 598         BN_ULONG l,h; \
 599  \
 600         h= (a); \
 601         l=LBITS(h); \
 602         h=HBITS(h); \
 603         mul64(l,h,(bl),(bh)); \
 604  \
 605         /* non-multiply part */ \
 606         l=(l+(c))&BN_MASK2; if (l < (c)) h++; \
 607         (c)=(r); \
 608         l=(l+(c))&BN_MASK2; if (l < (c)) h++; \
 609         (c)=h&BN_MASK2; \
 610         (r)=l; \
 611         }
 612
 613 #  define mul(r,a,bl,bh,c) { \
 614         BN_ULONG l,h; \
 615  \
 616         h= (a); \
 617         l=LBITS(h); \
 618         h=HBITS(h); \
 619         mul64(l,h,(bl),(bh)); \
 620  \
 621         /* non-multiply part */ \
 622         l+=(c); if ((l&BN_MASK2) < (c)) h++; \
 623         (c)=h&BN_MASK2; \
 624         (r)=l&BN_MASK2; \
 625         }
 626 # endif                         /* !BN_LLONG */
 627
 628 void BN_RECP_CTX_init(BN_RECP_CTX *recp);
 629 void BN_MONT_CTX_init(BN_MONT_CTX *ctx);
 630
 631 void bn_init(BIGNUM *a);
 632 void bn_mul_normal(BN_ULONG *r, BN_ULONG *a, int na, BN_ULONG *b, int nb);
 633 void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b);
 634 void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b);
 635 void bn_sqr_normal(BN_ULONG *r, const BN_ULONG *a, int n, BN_ULONG *tmp);
 636 void bn_sqr_comba8(BN_ULONG *r, const BN_ULONG *a);
 637 void bn_sqr_comba4(BN_ULONG *r, const BN_ULONG *a);
 638 int bn_cmp_words(const BN_ULONG *a, const BN_ULONG *b, int n);
 639 int bn_cmp_part_words(const BN_ULONG *a, const BN_ULONG *b, int cl, int dl);
 640 void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
 641                       int dna, int dnb, BN_ULONG *t);
 642 void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b,
 643                            int n, int tna, int tnb, BN_ULONG *t);
 644 void bn_sqr_recursive(BN_ULONG *r, const BN_ULONG *a, int n2, BN_ULONG *t);
 645 void bn_mul_low_normal(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n);
 646 void bn_mul_low_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
 647                           BN_ULONG *t);
 648 BN_ULONG bn_sub_part_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b,
 649                            int cl, int dl);
 650 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
 651                 const BN_ULONG *np, const BN_ULONG *n0, int num);
 652
 653 BIGNUM *int_bn_mod_inverse(BIGNUM *in,
 654                            const BIGNUM *a, const BIGNUM *n, BN_CTX *ctx,
 655                            int *noinv);
 656
 657 static ossl_inline BIGNUM *bn_expand(BIGNUM *a, int bits)
 658 {
 659     if (bits > (INT_MAX - BN_BITS2 + 1))
 660         return NULL;
 661
 662     if (((bits+BN_BITS2-1)/BN_BITS2) <= (a)->dmax)
 663         return a;
 664
 665     return bn_expand2((a),(bits+BN_BITS2-1)/BN_BITS2);
 666 }
 667
 668 #endif