1 /*
2 * Copyright 1995-2023 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 #include "internal/cryptlib.h"
26 #include "internal/numbers.h"
27
28 /*
29 * These preprocessor symbols control various aspects of the bignum headers
30 * and library code. They're not defined by any "normal" configuration, as
31 * they are intended for development and testing purposes. NB: defining
32 * them can be useful for debugging application code as well as openssl
33 * itself. BN_DEBUG - turn on various debugging alterations to the bignum
34 * code BN_RAND_DEBUG - uses random poisoning of unused words to trip up
35 * mismanagement of bignum internals. Enable BN_RAND_DEBUG is known to
36 * break some of the OpenSSL tests.
37 */
38 #if defined(BN_RAND_DEBUG) && !defined(BN_DEBUG)
39 #define BN_DEBUG
40 #endif
41 #if defined(BN_RAND_DEBUG)
42 #include <openssl/rand.h>
43 #endif
44
45 /*
46 * This should limit the stack usage due to alloca to about 4K.
47 * BN_SOFT_LIMIT is a soft limit equivalent to 2*OPENSSL_RSA_MAX_MODULUS_BITS.
48 * Beyond that size bn_mul_mont is no longer used, and the constant time
49 * assembler code is disabled, due to the blatant alloca and bn_mul_mont usage.
50 * Note that bn_mul_mont does an alloca that is hidden away in assembly.
51 * It is not recommended to do computations with numbers exceeding this limit,
52 * since the result will be highly version dependent:
53 * While the current OpenSSL version will use non-optimized, but safe code,
54 * previous versions will use optimized code, that may crash due to unexpected
55 * stack overflow, and future versions may very well turn this into a hard
56 * limit.
57 * Note however, that it is possible to override the size limit using
58 * "./config -DBN_SOFT_LIMIT=<limit>" if necessary, and the O/S specific
59 * stack limit is known and taken into consideration.
60 */
61 #ifndef BN_SOFT_LIMIT
62 #define BN_SOFT_LIMIT (4096 / BN_BYTES)
63 #endif
64
65 #ifndef OPENSSL_SMALL_FOOTPRINT
66 #define BN_MUL_COMBA
67 #define BN_SQR_COMBA
68 #define BN_RECURSION
69 #endif
70
71 /*
72 * This next option uses the C libraries (2 word)/(1 word) function. If it is
73 * not defined, I use my C version (which is slower). The reason for this
74 * flag is that when the particular C compiler library routine is used, and
75 * the library is linked with a different compiler, the library is missing.
76 * This mostly happens when the library is built with gcc and then linked
77 * using normal cc. This would be a common occurrence because gcc normally
78 * produces code that is 2 times faster than system compilers for the big
79 * number stuff. For machines with only one compiler (or shared libraries),
80 * this should be on. Again this in only really a problem on machines using
81 * "long long's", are 32bit, and are not using my assembler code.
82 */
83 #if defined(OPENSSL_SYS_MSDOS) || defined(OPENSSL_SYS_WINDOWS) || defined(OPENSSL_SYS_WIN32) || defined(linux)
84 #define BN_DIV2W
85 #endif
86
87 /*
88 * 64-bit processor with LP64 ABI
89 */
90 #ifdef SIXTY_FOUR_BIT_LONG
91 #define BN_ULLONG unsigned long long
92 #define BN_BITS4 32
93 #define BN_MASK2 (0xffffffffffffffffL)
94 #define BN_MASK2l (0xffffffffL)
95 #define BN_MASK2h (0xffffffff00000000L)
96 #define BN_MASK2h1 (0xffffffff80000000L)
97 #define BN_DEC_CONV (10000000000000000000UL)
98 #define BN_DEC_NUM 19
99 #define BN_DEC_FMT1 "%lu"
100 #define BN_DEC_FMT2 "%019lu"
101 #endif
102
103 /*
104 * 64-bit processor other than LP64 ABI
105 */
106 #ifdef SIXTY_FOUR_BIT
107 #undef BN_LLONG
108 #undef BN_ULLONG
109 #define BN_BITS4 32
110 #define BN_MASK2 (0xffffffffffffffffLL)
111 #define BN_MASK2l (0xffffffffL)
112 #define BN_MASK2h (0xffffffff00000000LL)
113 #define BN_MASK2h1 (0xffffffff80000000LL)
114 #define BN_DEC_CONV (10000000000000000000ULL)
115 #define BN_DEC_NUM 19
116 #define BN_DEC_FMT1 "%llu"
117 #define BN_DEC_FMT2 "%019llu"
118 #endif
119
120 #ifdef THIRTY_TWO_BIT
121 #ifdef BN_LLONG
122 #if defined(_WIN32) && !defined(__GNUC__)
123 #define BN_ULLONG unsigned __int64
124 #else
125 #define BN_ULLONG unsigned long long
126 #endif
127 #endif
128 #define BN_BITS4 16
129 #define BN_MASK2 (0xffffffffL)
130 #define BN_MASK2l (0xffff)
131 #define BN_MASK2h1 (0xffff8000L)
132 #define BN_MASK2h (0xffff0000L)
133 #define BN_DEC_CONV (1000000000L)
134 #define BN_DEC_NUM 9
135 #define BN_DEC_FMT1 "%u"
136 #define BN_DEC_FMT2 "%09u"
137 #endif
138
139 /*-
140 * Bignum consistency macros
141 * There is one "API" macro, bn_fix_top(), for stripping leading zeroes from
142 * bignum data after direct manipulations on the data. There is also an
143 * "internal" macro, bn_check_top(), for verifying that there are no leading
144 * zeroes. Unfortunately, some auditing is required due to the fact that
145 * bn_fix_top() has become an overabused duct-tape because bignum data is
146 * occasionally passed around in an inconsistent state. So the following
147 * changes have been made to sort this out;
148 * - bn_fix_top()s implementation has been moved to bn_correct_top()
149 * - if BN_DEBUG isn't defined, bn_fix_top() maps to bn_correct_top(), and
150 * bn_check_top() is as before.
151 * - if BN_DEBUG *is* defined;
152 * - bn_check_top() tries to pollute unused words even if the bignum 'top' is
153 * consistent. (ed: only if BN_RAND_DEBUG is defined)
154 * - bn_fix_top() maps to bn_check_top() rather than "fixing" anything.
155 * The idea is to have debug builds flag up inconsistent bignums when they
156 * occur. If that occurs in a bn_fix_top(), we examine the code in question; if
157 * the use of bn_fix_top() was appropriate (ie. it follows directly after code
158 * that manipulates the bignum) it is converted to bn_correct_top(), and if it
159 * was not appropriate, we convert it permanently to bn_check_top() and track
160 * down the cause of the bug. Eventually, no internal code should be using the
161 * bn_fix_top() macro. External applications and libraries should try this with
162 * their own code too, both in terms of building against the openssl headers
163 * with BN_DEBUG defined *and* linking with a version of OpenSSL built with it
164 * defined. This not only improves external code, it provides more test
165 * coverage for openssl's own code.
166 */
167
168 #ifdef BN_DEBUG
169 /*
170 * The new BN_FLG_FIXED_TOP flag marks vectors that were not treated with
171 * bn_correct_top, in other words such vectors are permitted to have zeros
172 * in most significant limbs. Such vectors are used internally to achieve
173 * execution time invariance for critical operations with private keys.
174 * It's BN_DEBUG-only flag, because user application is not supposed to
175 * observe it anyway. Moreover, optimizing compiler would actually remove
176 * all operations manipulating the bit in question in non-BN_DEBUG build.
177 */
178 #define BN_FLG_FIXED_TOP 0x10000
179 #ifdef BN_RAND_DEBUG
180 #define bn_pollute(a) \
181 do { \
182 const BIGNUM *_bnum1 = (a); \
183 if (_bnum1->top < _bnum1->dmax) { \
184 unsigned char _tmp_char; \
185 /* We cast away const without the compiler knowing, any \
186 * *genuinely* constant variables that aren't mutable \
187 * wouldn't be constructed with top!=dmax. */ \
188 BN_ULONG *_not_const; \
189 memcpy(&_not_const, &_bnum1->d, sizeof(_not_const)); \
190 (void)RAND_bytes(&_tmp_char, 1); /* Debug only - safe to ignore error return */ \
191 memset(_not_const + _bnum1->top, _tmp_char, \
192 sizeof(*_not_const) * (_bnum1->dmax - _bnum1->top)); \
193 } \
194 } while (0)
195 #else
196 #define bn_pollute(a)
197 #endif
198 #define bn_check_top(a) \
199 do { \
200 const BIGNUM *_bnum2 = (a); \
201 if (_bnum2 != NULL) { \
202 int _top = _bnum2->top; \
203 (void)ossl_assert((_top == 0 && !_bnum2->neg) || (_top && ((_bnum2->flags & BN_FLG_FIXED_TOP) || _bnum2->d[_top - 1] != 0))); \
204 bn_pollute(_bnum2); \
205 } \
206 } while (0)
207
208 #define bn_fix_top(a) bn_check_top(a)
209
210 #define bn_check_size(bn, bits) bn_wcheck_size(bn, ((bits + BN_BITS2 - 1)) / BN_BITS2)
211 #define bn_wcheck_size(bn, words) \
212 do { \
213 const BIGNUM *_bnum2 = (bn); \
214 assert((words) <= (_bnum2)->dmax && (words) >= (_bnum2)->top); \
215 /* avoid unused variable warning with NDEBUG */ \
216 (void)(_bnum2); \
217 } while (0)
218
219 #else /* !BN_DEBUG */
220
221 #define BN_FLG_FIXED_TOP 0
222 #define bn_pollute(a)
223 #define bn_check_top(a)
224 #define bn_fix_top(a) bn_correct_top(a)
225 #define bn_check_size(bn, bits)
226 #define bn_wcheck_size(bn, words)
227
228 #endif
229
230 BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num,
231 BN_ULONG w);
232 BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w);
233 void bn_sqr_words(BN_ULONG *rp, const BN_ULONG *ap, int num);
234 BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d);
235 BN_ULONG bn_add_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
236 int num);
237 BN_ULONG bn_sub_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
238 int num);
239
240 struct bignum_st {
241 BN_ULONG *d; /*
242 * Pointer to an array of 'BN_BITS2' bit
243 * chunks. These chunks are organised in
244 * a least significant chunk first order.
245 */
246 int top; /* Index of last used d +1. */
247 /* The next are internal book keeping for bn_expand. */
248 int dmax; /* Size of the d array. */
249 int neg; /* one if the number is negative */
250 int flags;
251 };
252
253 /* Used for montgomery multiplication */
254 struct bn_mont_ctx_st {
255 int ri; /* number of bits in R */
256 BIGNUM RR; /* used to convert to montgomery form,
257 possibly zero-padded */
258 BIGNUM N; /* The modulus */
259 BIGNUM Ni; /* R*(1/R mod N) - N*Ni = 1 (Ni is only
260 * stored for bignum algorithm) */
261 BN_ULONG n0[2]; /* least significant word(s) of Ni; (type
262 * changed with 0.9.9, was "BN_ULONG n0;"
263 * before) */
264 int flags;
265 };
266
267 /*
268 * Used for reciprocal division/mod functions It cannot be shared between
269 * threads
270 */
271 struct bn_recp_ctx_st {
272 BIGNUM N; /* the divisor */
273 BIGNUM Nr; /* the reciprocal */
274 int num_bits;
275 int shift;
276 int flags;
277 };
278
279 /* Used for slow "generation" functions. */
280 struct bn_gencb_st {
281 unsigned int ver; /* To handle binary (in)compatibility */
282 void *arg; /* callback-specific data */
283 union {
284 /* if (ver==1) - handles old style callbacks */
285 void (*cb_1)(int, int, void *);
286 /* if (ver==2) - new callback style */
287 int (*cb_2)(int, int, BN_GENCB *);
288 } cb;
289 };
290
291 /*-
292 * BN_window_bits_for_exponent_size -- macro for sliding window mod_exp functions
293 *
294 *
295 * For window size 'w' (w >= 2) and a random 'b' bits exponent,
296 * the number of multiplications is a constant plus on average
297 *
298 * 2^(w-1) + (b-w)/(w+1);
299 *
300 * here 2^(w-1) is for precomputing the table (we actually need
301 * entries only for windows that have the lowest bit set), and
302 * (b-w)/(w+1) is an approximation for the expected number of
303 * w-bit windows, not counting the first one.
304 *
305 * Thus we should use
306 *
307 * w >= 6 if b > 671
308 * w = 5 if 671 > b > 239
309 * w = 4 if 239 > b > 79
310 * w = 3 if 79 > b > 23
311 * w <= 2 if 23 > b
312 *
313 * (with draws in between). Very small exponents are often selected
314 * with low Hamming weight, so we use w = 1 for b <= 23.
315 */
316 #define BN_window_bits_for_exponent_size(b) \
317 ((b) > 671 ? 6 : (b) > 239 ? 5 \
318 : (b) > 79 ? 4 \
319 : (b) > 23 ? 3 \
320 : 1)
321
322 /*
323 * BN_mod_exp_mont_consttime is based on the assumption that the L1 data cache
324 * line width of the target processor is at least the following value.
325 */
326 #define MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH (64)
327 #define MOD_EXP_CTIME_MIN_CACHE_LINE_MASK (MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH - 1)
328
329 /*
330 * Window sizes optimized for fixed window size modular exponentiation
331 * algorithm (BN_mod_exp_mont_consttime). To achieve the security goals of
332 * BN_mode_exp_mont_consttime, the maximum size of the window must not exceed
333 * log_2(MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH). Window size thresholds are
334 * defined for cache line sizes of 32 and 64, cache line sizes where
335 * log_2(32)=5 and log_2(64)=6 respectively. A window size of 7 should only be
336 * used on processors that have a 128 byte or greater cache line size.
337 */
338 #if MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH == 64
339
340 #define BN_window_bits_for_ctime_exponent_size(b) \
341 ((b) > 937 ? 6 : (b) > 306 ? 5 \
342 : (b) > 89 ? 4 \
343 : (b) > 22 ? 3 \
344 : 1)
345 #define BN_MAX_WINDOW_BITS_FOR_CTIME_EXPONENT_SIZE (6)
346
347 #elif MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH == 32
348
349 #define BN_window_bits_for_ctime_exponent_size(b) \
350 ((b) > 306 ? 5 : (b) > 89 ? 4 \
351 : (b) > 22 ? 3 \
352 : 1)
353 #define BN_MAX_WINDOW_BITS_FOR_CTIME_EXPONENT_SIZE (5)
354
355 #endif
356
357 /* Pentium pro 16,16,16,32,64 */
358 /* Alpha 16,16,16,16.64 */
359 #define BN_MULL_SIZE_NORMAL (16) /* 32 */
360 #define BN_MUL_RECURSIVE_SIZE_NORMAL (16) /* 32 less than */
361 #define BN_SQR_RECURSIVE_SIZE_NORMAL (16) /* 32 */
362 #define BN_MUL_LOW_RECURSIVE_SIZE_NORMAL (32) /* 32 */
363 #define BN_MONT_CTX_SET_SIZE_WORD (64) /* 32 */
364
365 #if !defined(OPENSSL_NO_ASM) && !defined(OPENSSL_NO_INLINE_ASM) && !defined(PEDANTIC)
366 /*
367 * BN_UMULT_HIGH section.
368 * If the compiler doesn't support 2*N integer type, then you have to
369 * replace every N*N multiplication with 4 (N/2)*(N/2) accompanied by some
370 * shifts and additions which unavoidably results in severe performance
371 * penalties. Of course provided that the hardware is capable of producing
372 * 2*N result... That's when you normally start considering assembler
373 * implementation. However! It should be pointed out that some CPUs (e.g.,
374 * PowerPC, Alpha, and IA-64) provide *separate* instruction calculating
375 * the upper half of the product placing the result into a general
376 * purpose register. Now *if* the compiler supports inline assembler,
377 * then it's not impossible to implement the "bignum" routines (and have
378 * the compiler optimize 'em) exhibiting "native" performance in C. That's
379 * what BN_UMULT_HIGH macro is about:-) Note that more recent compilers do
380 * support 2*64 integer type, which is also used here.
381 */
382 #if defined(__SIZEOF_INT128__) && __SIZEOF_INT128__ == 16 && (defined(SIXTY_FOUR_BIT) || defined(SIXTY_FOUR_BIT_LONG))
383 #define BN_UMULT_HIGH(a, b) (((uint128_t)(a) * (b)) >> 64)
384 #define BN_UMULT_LOHI(low, high, a, b) ({ \
385 uint128_t ret=(uint128_t)(a)*(b); \
386 (high)=ret>>64; (low)=ret; })
387 #elif defined(__alpha) && (defined(SIXTY_FOUR_BIT_LONG) || defined(SIXTY_FOUR_BIT))
388 #if defined(__DECC)
389 #include <c_asm.h>
390 #define BN_UMULT_HIGH(a, b) (BN_ULONG) asm("umulh %a0,%a1,%v0", (a), (b))
391 #elif defined(__GNUC__) && __GNUC__ >= 2
392 #define BN_UMULT_HIGH(a, b) ({ \
393 register BN_ULONG ret; \
394 asm ("umulh %1,%2,%0" \
395 : "=r"(ret) \
396 : "r"(a), "r"(b)); \
397 ret; })
398 #endif /* compiler */
399 #elif defined(_ARCH_PPC64) && defined(SIXTY_FOUR_BIT_LONG)
400 #if defined(__GNUC__) && __GNUC__ >= 2
401 #define BN_UMULT_HIGH(a, b) ({ \
402 register BN_ULONG ret; \
403 asm ("mulhdu %0,%1,%2" \
404 : "=r"(ret) \
405 : "r"(a), "r"(b)); \
406 ret; })
407 #endif /* compiler */
408 #elif (defined(__x86_64) || defined(__x86_64__)) && (defined(SIXTY_FOUR_BIT_LONG) || defined(SIXTY_FOUR_BIT))
409 #if defined(__GNUC__) && __GNUC__ >= 2
410 #define BN_UMULT_HIGH(a, b) ({ \
411 register BN_ULONG ret,discard; \
412 asm ("mulq %3" \
413 : "=a"(discard),"=d"(ret) \
414 : "a"(a), "g"(b) \
415 : "cc"); \
416 ret; })
417 #define BN_UMULT_LOHI(low, high, a, b) \
418 asm("mulq %3" \
419 : "=a"(low), "=d"(high) \
420 : "a"(a), "g"(b) \
421 : "cc");
422 #endif
423 #elif (defined(_M_AMD64) || defined(_M_X64)) && defined(SIXTY_FOUR_BIT)
424 #if defined(_MSC_VER) && _MSC_VER >= 1400
425 unsigned __int64 __umulh(unsigned __int64 a, unsigned __int64 b);
426 unsigned __int64 _umul128(unsigned __int64 a, unsigned __int64 b,
427 unsigned __int64 *h);
428 #pragma intrinsic(__umulh, _umul128)
429 #define BN_UMULT_HIGH(a, b) __umulh((a), (b))
430 #define BN_UMULT_LOHI(low, high, a, b) ((low) = _umul128((a), (b), &(high)))
431 #endif
432 #elif defined(__mips) && (defined(SIXTY_FOUR_BIT) || defined(SIXTY_FOUR_BIT_LONG))
433 #if defined(__GNUC__) && __GNUC__ >= 2
434 #define BN_UMULT_HIGH(a, b) ({ \
435 register BN_ULONG ret; \
436 asm ("dmultu %1,%2" \
437 : "=h"(ret) \
438 : "r"(a), "r"(b) : "l"); \
439 ret; })
440 #define BN_UMULT_LOHI(low, high, a, b) \
441 asm("dmultu %2,%3" \
442 : "=l"(low), "=h"(high) \
443 : "r"(a), "r"(b));
444 #endif
445 #elif defined(__aarch64__) && defined(SIXTY_FOUR_BIT_LONG)
446 #if defined(__GNUC__) && __GNUC__ >= 2
447 #define BN_UMULT_HIGH(a, b) ({ \
448 register BN_ULONG ret; \
449 asm ("umulh %0,%1,%2" \
450 : "=r"(ret) \
451 : "r"(a), "r"(b)); \
452 ret; })
453 #endif
454 #endif /* cpu */
455 #endif /* OPENSSL_NO_ASM */
456
457 #ifdef BN_RAND_DEBUG
458 #define bn_clear_top2max(a) \
459 { \
460 int ind = (a)->dmax - (a)->top; \
461 BN_ULONG *ftl = &(a)->d[(a)->top - 1]; \
462 for (; ind != 0; ind--) \
463 *(++ftl) = 0x0; \
464 }
465 #else
466 #define bn_clear_top2max(a)
467 #endif
468
469 #ifdef BN_LLONG
470 /*******************************************************************
471 * Using the long long type, has to be twice as wide as BN_ULONG...
472 */
473 #define Lw(t) (((BN_ULONG)(t)) & BN_MASK2)
474 #define Hw(t) (((BN_ULONG)((t) >> BN_BITS2)) & BN_MASK2)
475
476 #define mul_add(r, a, w, c) \
477 { \
478 BN_ULLONG t; \
479 t = (BN_ULLONG)w * (a) + (r) + (c); \
480 (r) = Lw(t); \
481 (c) = Hw(t); \
482 }
483
484 #define mul(r, a, w, c) \
485 { \
486 BN_ULLONG t; \
487 t = (BN_ULLONG)w * (a) + (c); \
488 (r) = Lw(t); \
489 (c) = Hw(t); \
490 }
491
492 #define sqr(r0, r1, a) \
493 { \
494 BN_ULLONG t; \
495 t = (BN_ULLONG)(a) * (a); \
496 (r0) = Lw(t); \
497 (r1) = Hw(t); \
498 }
499
500 #elif defined(BN_UMULT_LOHI)
501 #define mul_add(r, a, w, c) \
502 { \
503 BN_ULONG high, low, ret, tmp = (a); \
504 ret = (r); \
505 BN_UMULT_LOHI(low, high, w, tmp); \
506 ret += (c); \
507 (c) = (ret < (c)); \
508 (c) += high; \
509 ret += low; \
510 (c) += (ret < low); \
511 (r) = ret; \
512 }
513
514 #define mul(r, a, w, c) \
515 { \
516 BN_ULONG high, low, ret, ta = (a); \
517 BN_UMULT_LOHI(low, high, w, ta); \
518 ret = low + (c); \
519 (c) = high; \
520 (c) += (ret < low); \
521 (r) = ret; \
522 }
523
524 #define sqr(r0, r1, a) \
525 { \
526 BN_ULONG tmp = (a); \
527 BN_UMULT_LOHI(r0, r1, tmp, tmp); \
528 }
529
530 #elif defined(BN_UMULT_HIGH)
531 #define mul_add(r, a, w, c) \
532 { \
533 BN_ULONG high, low, ret, tmp = (a); \
534 ret = (r); \
535 high = BN_UMULT_HIGH(w, tmp); \
536 ret += (c); \
537 low = (w) * tmp; \
538 (c) = (ret < (c)); \
539 (c) += high; \
540 ret += low; \
541 (c) += (ret < low); \
542 (r) = ret; \
543 }
544
545 #define mul(r, a, w, c) \
546 { \
547 BN_ULONG high, low, ret, ta = (a); \
548 low = (w) * ta; \
549 high = BN_UMULT_HIGH(w, ta); \
550 ret = low + (c); \
551 (c) = high; \
552 (c) += (ret < low); \
553 (r) = ret; \
554 }
555
556 #define sqr(r0, r1, a) \
557 { \
558 BN_ULONG tmp = (a); \
559 (r0) = tmp * tmp; \
560 (r1) = BN_UMULT_HIGH(tmp, tmp); \
561 }
562
563 #else
564 /*************************************************************
565 * No long long type
566 */
567
568 #define LBITS(a) ((a) & BN_MASK2l)
569 #define HBITS(a) (((a) >> BN_BITS4) & BN_MASK2l)
570 #define L2HBITS(a) (((a) << BN_BITS4) & BN_MASK2)
571
572 #define LLBITS(a) ((a) & BN_MASKl)
573 #define LHBITS(a) (((a) >> BN_BITS2) & BN_MASKl)
574 #define LL2HBITS(a) ((BN_ULLONG)((a) & BN_MASKl) << BN_BITS2)
575
576 #define mul64(l, h, bl, bh) \
577 { \
578 BN_ULONG m, m1, lt, ht; \
579 \
580 lt = l; \
581 ht = h; \
582 m = (bh) * (lt); \
583 lt = (bl) * (lt); \
584 m1 = (bl) * (ht); \
585 ht = (bh) * (ht); \
586 m = (m + m1) & BN_MASK2; \
587 ht += L2HBITS((BN_ULONG)(m < m1)); \
588 ht += HBITS(m); \
589 m1 = L2HBITS(m); \
590 lt = (lt + m1) & BN_MASK2; \
591 ht += (lt < m1); \
592 (l) = lt; \
593 (h) = ht; \
594 }
595
596 #define sqr64(lo, ho, in) \
597 { \
598 BN_ULONG l, h, m; \
599 \
600 h = (in); \
601 l = LBITS(h); \
602 h = HBITS(h); \
603 m = (l) * (h); \
604 l *= l; \
605 h *= h; \
606 h += (m & BN_MASK2h1) >> (BN_BITS4 - 1); \
607 m = (m & BN_MASK2l) << (BN_BITS4 + 1); \
608 l = (l + m) & BN_MASK2; \
609 h += (l < m); \
610 (lo) = l; \
611 (ho) = h; \
612 }
613
614 #define mul_add(r, a, bl, bh, c) \
615 { \
616 BN_ULONG l, h; \
617 \
618 h = (a); \
619 l = LBITS(h); \
620 h = HBITS(h); \
621 mul64(l, h, (bl), (bh)); \
622 \
623 /* non-multiply part */ \
624 l = (l + (c)) & BN_MASK2; \
625 h += (l < (c)); \
626 (c) = (r); \
627 l = (l + (c)) & BN_MASK2; \
628 h += (l < (c)); \
629 (c) = h & BN_MASK2; \
630 (r) = l; \
631 }
632
633 #define mul(r, a, bl, bh, c) \
634 { \
635 BN_ULONG l, h; \
636 \
637 h = (a); \
638 l = LBITS(h); \
639 h = HBITS(h); \
640 mul64(l, h, (bl), (bh)); \
641 \
642 /* non-multiply part */ \
643 l += (c); \
644 h += ((l & BN_MASK2) < (c)); \
645 (c) = h & BN_MASK2; \
646 (r) = l & BN_MASK2; \
647 }
648 #endif /* !BN_LLONG */
649
650 void BN_RECP_CTX_init(BN_RECP_CTX *recp);
651 void BN_MONT_CTX_init(BN_MONT_CTX *ctx);
652
653 void bn_init(BIGNUM *a);
654 void bn_mul_normal(BN_ULONG *r, BN_ULONG *a, int na, BN_ULONG *b, int nb);
655 void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b);
656 void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b);
657 void bn_sqr_normal(BN_ULONG *r, const BN_ULONG *a, int n, BN_ULONG *tmp);
658 void bn_sqr_comba8(BN_ULONG *r, const BN_ULONG *a);
659 void bn_sqr_comba4(BN_ULONG *r, const BN_ULONG *a);
660 int bn_cmp_words(const BN_ULONG *a, const BN_ULONG *b, int n);
661 int bn_cmp_part_words(const BN_ULONG *a, const BN_ULONG *b, int cl, int dl);
662 void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
663 int dna, int dnb, BN_ULONG *t);
664 void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b,
665 int n, int tna, int tnb, BN_ULONG *t);
666 void bn_sqr_recursive(BN_ULONG *r, const BN_ULONG *a, int n2, BN_ULONG *t);
667 void bn_mul_low_normal(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n);
668 void bn_mul_low_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
669 BN_ULONG *t);
670 BN_ULONG bn_sub_part_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b,
671 int cl, int dl);
672 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
673 const BN_ULONG *np, const BN_ULONG *n0, int num);
674 void bn_correct_top_consttime(BIGNUM *a);
675 BIGNUM *int_bn_mod_inverse(BIGNUM *in,
676 const BIGNUM *a, const BIGNUM *n, BN_CTX *ctx,
677 int *noinv);
678
bn_expand(BIGNUM * a,int bits)679 static ossl_inline BIGNUM *bn_expand(BIGNUM *a, int bits)
680 {
681 if (bits > (INT_MAX - BN_BITS2 + 1))
682 return NULL;
683
684 if (((bits + BN_BITS2 - 1) / BN_BITS2) <= (a)->dmax)
685 return a;
686
687 return bn_expand2((a), (bits + BN_BITS2 - 1) / BN_BITS2);
688 }
689
690 int ossl_bn_check_prime(const BIGNUM *w, int checks, BN_CTX *ctx,
691 int do_trial_division, BN_GENCB *cb);
692
693 #endif
694