1 /*
2 * Copyright 1995-2021 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 # ifndef OPENSSL_SMALL_FOOTPRINT
46 # define BN_MUL_COMBA
47 # define BN_SQR_COMBA
48 # define BN_RECURSION
49 # endif
50
51 /*
52 * This next option uses the C libraries (2 word)/(1 word) function. If it is
53 * not defined, I use my C version (which is slower). The reason for this
54 * flag is that when the particular C compiler library routine is used, and
55 * the library is linked with a different compiler, the library is missing.
56 * This mostly happens when the library is built with gcc and then linked
57 * using normal cc. This would be a common occurrence because gcc normally
58 * produces code that is 2 times faster than system compilers for the big
59 * number stuff. For machines with only one compiler (or shared libraries),
60 * this should be on. Again this in only really a problem on machines using
61 * "long long's", are 32bit, and are not using my assembler code.
62 */
63 # if defined(OPENSSL_SYS_MSDOS) || defined(OPENSSL_SYS_WINDOWS) || \
64 defined(OPENSSL_SYS_WIN32) || defined(linux)
65 # define BN_DIV2W
66 # endif
67
68 /*
69 * 64-bit processor with LP64 ABI
70 */
71 # ifdef SIXTY_FOUR_BIT_LONG
72 # define BN_ULLONG unsigned long long
73 # define BN_BITS4 32
74 # define BN_MASK2 (0xffffffffffffffffL)
75 # define BN_MASK2l (0xffffffffL)
76 # define BN_MASK2h (0xffffffff00000000L)
77 # define BN_MASK2h1 (0xffffffff80000000L)
78 # define BN_DEC_CONV (10000000000000000000UL)
79 # define BN_DEC_NUM 19
80 # define BN_DEC_FMT1 "%lu"
81 # define BN_DEC_FMT2 "%019lu"
82 # endif
83
84 /*
85 * 64-bit processor other than LP64 ABI
86 */
87 # ifdef SIXTY_FOUR_BIT
88 # undef BN_LLONG
89 # undef BN_ULLONG
90 # define BN_BITS4 32
91 # define BN_MASK2 (0xffffffffffffffffLL)
92 # define BN_MASK2l (0xffffffffL)
93 # define BN_MASK2h (0xffffffff00000000LL)
94 # define BN_MASK2h1 (0xffffffff80000000LL)
95 # define BN_DEC_CONV (10000000000000000000ULL)
96 # define BN_DEC_NUM 19
97 # define BN_DEC_FMT1 "%llu"
98 # define BN_DEC_FMT2 "%019llu"
99 # endif
100
101 # ifdef THIRTY_TWO_BIT
102 # ifdef BN_LLONG
103 # if defined(_WIN32) && !defined(__GNUC__)
104 # define BN_ULLONG unsigned __int64
105 # else
106 # define BN_ULLONG unsigned long long
107 # endif
108 # endif
109 # define BN_BITS4 16
110 # define BN_MASK2 (0xffffffffL)
111 # define BN_MASK2l (0xffff)
112 # define BN_MASK2h1 (0xffff8000L)
113 # define BN_MASK2h (0xffff0000L)
114 # define BN_DEC_CONV (1000000000L)
115 # define BN_DEC_NUM 9
116 # define BN_DEC_FMT1 "%u"
117 # define BN_DEC_FMT2 "%09u"
118 # endif
119
120
121 /*-
122 * Bignum consistency macros
123 * There is one "API" macro, bn_fix_top(), for stripping leading zeroes from
124 * bignum data after direct manipulations on the data. There is also an
125 * "internal" macro, bn_check_top(), for verifying that there are no leading
126 * zeroes. Unfortunately, some auditing is required due to the fact that
127 * bn_fix_top() has become an overabused duct-tape because bignum data is
128 * occasionally passed around in an inconsistent state. So the following
129 * changes have been made to sort this out;
130 * - bn_fix_top()s implementation has been moved to bn_correct_top()
131 * - if BN_DEBUG isn't defined, bn_fix_top() maps to bn_correct_top(), and
132 * bn_check_top() is as before.
133 * - if BN_DEBUG *is* defined;
134 * - bn_check_top() tries to pollute unused words even if the bignum 'top' is
135 * consistent. (ed: only if BN_RAND_DEBUG is defined)
136 * - bn_fix_top() maps to bn_check_top() rather than "fixing" anything.
137 * The idea is to have debug builds flag up inconsistent bignums when they
138 * occur. If that occurs in a bn_fix_top(), we examine the code in question; if
139 * the use of bn_fix_top() was appropriate (ie. it follows directly after code
140 * that manipulates the bignum) it is converted to bn_correct_top(), and if it
141 * was not appropriate, we convert it permanently to bn_check_top() and track
142 * down the cause of the bug. Eventually, no internal code should be using the
143 * bn_fix_top() macro. External applications and libraries should try this with
144 * their own code too, both in terms of building against the openssl headers
145 * with BN_DEBUG defined *and* linking with a version of OpenSSL built with it
146 * defined. This not only improves external code, it provides more test
147 * coverage for openssl's own code.
148 */
149
150 # ifdef BN_DEBUG
151 /*
152 * The new BN_FLG_FIXED_TOP flag marks vectors that were not treated with
153 * bn_correct_top, in other words such vectors are permitted to have zeros
154 * in most significant limbs. Such vectors are used internally to achieve
155 * execution time invariance for critical operations with private keys.
156 * It's BN_DEBUG-only flag, because user application is not supposed to
157 * observe it anyway. Moreover, optimizing compiler would actually remove
158 * all operations manipulating the bit in question in non-BN_DEBUG build.
159 */
160 # define BN_FLG_FIXED_TOP 0x10000
161 # ifdef BN_RAND_DEBUG
162 # define bn_pollute(a) \
163 do { \
164 const BIGNUM *_bnum1 = (a); \
165 if (_bnum1->top < _bnum1->dmax) { \
166 unsigned char _tmp_char; \
167 /* We cast away const without the compiler knowing, any \
168 * *genuinely* constant variables that aren't mutable \
169 * wouldn't be constructed with top!=dmax. */ \
170 BN_ULONG *_not_const; \
171 memcpy(&_not_const, &_bnum1->d, sizeof(_not_const)); \
172 (void)RAND_bytes(&_tmp_char, 1); /* Debug only - safe to ignore error return */\
173 memset(_not_const + _bnum1->top, _tmp_char, \
174 sizeof(*_not_const) * (_bnum1->dmax - _bnum1->top)); \
175 } \
176 } while(0)
177 # else
178 # define bn_pollute(a)
179 # endif
180 # define bn_check_top(a) \
181 do { \
182 const BIGNUM *_bnum2 = (a); \
183 if (_bnum2 != NULL) { \
184 int _top = _bnum2->top; \
185 (void)ossl_assert((_top == 0 && !_bnum2->neg) || \
186 (_top && ((_bnum2->flags & BN_FLG_FIXED_TOP) \
187 || _bnum2->d[_top - 1] != 0))); \
188 bn_pollute(_bnum2); \
189 } \
190 } while(0)
191
192 # define bn_fix_top(a) bn_check_top(a)
193
194 # define bn_check_size(bn, bits) bn_wcheck_size(bn, ((bits+BN_BITS2-1))/BN_BITS2)
195 # define bn_wcheck_size(bn, words) \
196 do { \
197 const BIGNUM *_bnum2 = (bn); \
198 assert((words) <= (_bnum2)->dmax && \
199 (words) >= (_bnum2)->top); \
200 /* avoid unused variable warning with NDEBUG */ \
201 (void)(_bnum2); \
202 } while(0)
203
204 # else /* !BN_DEBUG */
205
206 # define BN_FLG_FIXED_TOP 0
207 # define bn_pollute(a)
208 # define bn_check_top(a)
209 # define bn_fix_top(a) bn_correct_top(a)
210 # define bn_check_size(bn, bits)
211 # define bn_wcheck_size(bn, words)
212
213 # endif
214
215 BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num,
216 BN_ULONG w);
217 BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w);
218 void bn_sqr_words(BN_ULONG *rp, const BN_ULONG *ap, int num);
219 BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d);
220 BN_ULONG bn_add_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
221 int num);
222 BN_ULONG bn_sub_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
223 int num);
224
225 struct bignum_st {
226 BN_ULONG *d; /* Pointer to an array of 'BN_BITS2' bit
227 * chunks. */
228 int top; /* Index of last used d +1. */
229 /* The next are internal book keeping for bn_expand. */
230 int dmax; /* Size of the d array. */
231 int neg; /* one if the number is negative */
232 int flags;
233 };
234
235 /* Used for montgomery multiplication */
236 struct bn_mont_ctx_st {
237 int ri; /* number of bits in R */
238 BIGNUM RR; /* used to convert to montgomery form,
239 possibly zero-padded */
240 BIGNUM N; /* The modulus */
241 BIGNUM Ni; /* R*(1/R mod N) - N*Ni = 1 (Ni is only
242 * stored for bignum algorithm) */
243 BN_ULONG n0[2]; /* least significant word(s) of Ni; (type
244 * changed with 0.9.9, was "BN_ULONG n0;"
245 * before) */
246 int flags;
247 };
248
249 /*
250 * Used for reciprocal division/mod functions It cannot be shared between
251 * threads
252 */
253 struct bn_recp_ctx_st {
254 BIGNUM N; /* the divisor */
255 BIGNUM Nr; /* the reciprocal */
256 int num_bits;
257 int shift;
258 int flags;
259 };
260
261 /* Used for slow "generation" functions. */
262 struct bn_gencb_st {
263 unsigned int ver; /* To handle binary (in)compatibility */
264 void *arg; /* callback-specific data */
265 union {
266 /* if (ver==1) - handles old style callbacks */
267 void (*cb_1) (int, int, void *);
268 /* if (ver==2) - new callback style */
269 int (*cb_2) (int, int, BN_GENCB *);
270 } cb;
271 };
272
273 struct bn_blinding_st {
274 BIGNUM *A;
275 BIGNUM *Ai;
276 BIGNUM *e;
277 BIGNUM *mod; /* just a reference */
278 CRYPTO_THREAD_ID tid;
279 int counter;
280 unsigned long flags;
281 BN_MONT_CTX *m_ctx;
282 int (*bn_mod_exp) (BIGNUM *r, const BIGNUM *a, const BIGNUM *p,
283 const BIGNUM *m, BN_CTX *ctx, BN_MONT_CTX *m_ctx);
284 CRYPTO_RWLOCK *lock;
285 };
286
287 /*-
288 * BN_window_bits_for_exponent_size -- macro for sliding window mod_exp functions
289 *
290 *
291 * For window size 'w' (w >= 2) and a random 'b' bits exponent,
292 * the number of multiplications is a constant plus on average
293 *
294 * 2^(w-1) + (b-w)/(w+1);
295 *
296 * here 2^(w-1) is for precomputing the table (we actually need
297 * entries only for windows that have the lowest bit set), and
298 * (b-w)/(w+1) is an approximation for the expected number of
299 * w-bit windows, not counting the first one.
300 *
301 * Thus we should use
302 *
303 * w >= 6 if b > 671
304 * w = 5 if 671 > b > 239
305 * w = 4 if 239 > b > 79
306 * w = 3 if 79 > b > 23
307 * w <= 2 if 23 > b
308 *
309 * (with draws in between). Very small exponents are often selected
310 * with low Hamming weight, so we use w = 1 for b <= 23.
311 */
312 # define BN_window_bits_for_exponent_size(b) \
313 ((b) > 671 ? 6 : \
314 (b) > 239 ? 5 : \
315 (b) > 79 ? 4 : \
316 (b) > 23 ? 3 : 1)
317
318 /*
319 * BN_mod_exp_mont_consttime is based on the assumption that the L1 data cache
320 * line width of the target processor is at least the following value.
321 */
322 # define MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH ( 64 )
323 # define MOD_EXP_CTIME_MIN_CACHE_LINE_MASK (MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH - 1)
324
325 /*
326 * Window sizes optimized for fixed window size modular exponentiation
327 * algorithm (BN_mod_exp_mont_consttime). To achieve the security goals of
328 * BN_mode_exp_mont_consttime, the maximum size of the window must not exceed
329 * log_2(MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH). Window size thresholds are
330 * defined for cache line sizes of 32 and 64, cache line sizes where
331 * log_2(32)=5 and log_2(64)=6 respectively. A window size of 7 should only be
332 * used on processors that have a 128 byte or greater cache line size.
333 */
334 # if MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH == 64
335
336 # define BN_window_bits_for_ctime_exponent_size(b) \
337 ((b) > 937 ? 6 : \
338 (b) > 306 ? 5 : \
339 (b) > 89 ? 4 : \
340 (b) > 22 ? 3 : 1)
341 # define BN_MAX_WINDOW_BITS_FOR_CTIME_EXPONENT_SIZE (6)
342
343 # elif MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH == 32
344
345 # define BN_window_bits_for_ctime_exponent_size(b) \
346 ((b) > 306 ? 5 : \
347 (b) > 89 ? 4 : \
348 (b) > 22 ? 3 : 1)
349 # define BN_MAX_WINDOW_BITS_FOR_CTIME_EXPONENT_SIZE (5)
350
351 # endif
352
353 /* Pentium pro 16,16,16,32,64 */
354 /* Alpha 16,16,16,16.64 */
355 # define BN_MULL_SIZE_NORMAL (16)/* 32 */
356 # define BN_MUL_RECURSIVE_SIZE_NORMAL (16)/* 32 less than */
357 # define BN_SQR_RECURSIVE_SIZE_NORMAL (16)/* 32 */
358 # define BN_MUL_LOW_RECURSIVE_SIZE_NORMAL (32)/* 32 */
359 # define BN_MONT_CTX_SET_SIZE_WORD (64)/* 32 */
360
361 /*
362 * 2011-02-22 SMS. In various places, a size_t variable or a type cast to
363 * size_t was used to perform integer-only operations on pointers. This
364 * failed on VMS with 64-bit pointers (CC /POINTER_SIZE = 64) because size_t
365 * is still only 32 bits. What's needed in these cases is an integer type
366 * with the same size as a pointer, which size_t is not certain to be. The
367 * only fix here is VMS-specific.
368 */
369 # if defined(OPENSSL_SYS_VMS)
370 # if __INITIAL_POINTER_SIZE == 64
371 # define PTR_SIZE_INT long long
372 # else /* __INITIAL_POINTER_SIZE == 64 */
373 # define PTR_SIZE_INT int
374 # endif /* __INITIAL_POINTER_SIZE == 64 [else] */
375 # elif !defined(PTR_SIZE_INT) /* defined(OPENSSL_SYS_VMS) */
376 # define PTR_SIZE_INT size_t
377 # endif /* defined(OPENSSL_SYS_VMS) [else] */
378
379 # if !defined(OPENSSL_NO_ASM) && !defined(OPENSSL_NO_INLINE_ASM) && !defined(PEDANTIC)
380 /*
381 * BN_UMULT_HIGH section.
382 * If the compiler doesn't support 2*N integer type, then you have to
383 * replace every N*N multiplication with 4 (N/2)*(N/2) accompanied by some
384 * shifts and additions which unavoidably results in severe performance
385 * penalties. Of course provided that the hardware is capable of producing
386 * 2*N result... That's when you normally start considering assembler
387 * implementation. However! It should be pointed out that some CPUs (e.g.,
388 * PowerPC, Alpha, and IA-64) provide *separate* instruction calculating
389 * the upper half of the product placing the result into a general
390 * purpose register. Now *if* the compiler supports inline assembler,
391 * then it's not impossible to implement the "bignum" routines (and have
392 * the compiler optimize 'em) exhibiting "native" performance in C. That's
393 * what BN_UMULT_HIGH macro is about:-) Note that more recent compilers do
394 * support 2*64 integer type, which is also used here.
395 */
396 # if defined(__SIZEOF_INT128__) && __SIZEOF_INT128__==16 && \
397 (defined(SIXTY_FOUR_BIT) || defined(SIXTY_FOUR_BIT_LONG))
398 # define BN_UMULT_HIGH(a,b) (((uint128_t)(a)*(b))>>64)
399 # define BN_UMULT_LOHI(low,high,a,b) ({ \
400 uint128_t ret=(uint128_t)(a)*(b); \
401 (high)=ret>>64; (low)=ret; })
402 # elif defined(__alpha) && (defined(SIXTY_FOUR_BIT_LONG) || defined(SIXTY_FOUR_BIT))
403 # if defined(__DECC)
404 # include <c_asm.h>
405 # define BN_UMULT_HIGH(a,b) (BN_ULONG)asm("umulh %a0,%a1,%v0",(a),(b))
406 # elif defined(__GNUC__) && __GNUC__>=2
407 # define BN_UMULT_HIGH(a,b) ({ \
408 register BN_ULONG ret; \
409 asm ("umulh %1,%2,%0" \
410 : "=r"(ret) \
411 : "r"(a), "r"(b)); \
412 ret; })
413 # endif /* compiler */
414 # elif defined(_ARCH_PPC64) && defined(SIXTY_FOUR_BIT_LONG)
415 # if defined(__GNUC__) && __GNUC__>=2
416 # define BN_UMULT_HIGH(a,b) ({ \
417 register BN_ULONG ret; \
418 asm ("mulhdu %0,%1,%2" \
419 : "=r"(ret) \
420 : "r"(a), "r"(b)); \
421 ret; })
422 # endif /* compiler */
423 # elif (defined(__x86_64) || defined(__x86_64__)) && \
424 (defined(SIXTY_FOUR_BIT_LONG) || defined(SIXTY_FOUR_BIT))
425 # if defined(__GNUC__) && __GNUC__>=2
426 # define BN_UMULT_HIGH(a,b) ({ \
427 register BN_ULONG ret,discard; \
428 asm ("mulq %3" \
429 : "=a"(discard),"=d"(ret) \
430 : "a"(a), "g"(b) \
431 : "cc"); \
432 ret; })
433 # define BN_UMULT_LOHI(low,high,a,b) \
434 asm ("mulq %3" \
435 : "=a"(low),"=d"(high) \
436 : "a"(a),"g"(b) \
437 : "cc");
438 # endif
439 # elif (defined(_M_AMD64) || defined(_M_X64)) && defined(SIXTY_FOUR_BIT)
440 # if defined(_MSC_VER) && _MSC_VER>=1400
441 unsigned __int64 __umulh(unsigned __int64 a, unsigned __int64 b);
442 unsigned __int64 _umul128(unsigned __int64 a, unsigned __int64 b,
443 unsigned __int64 *h);
444 # pragma intrinsic(__umulh,_umul128)
445 # define BN_UMULT_HIGH(a,b) __umulh((a),(b))
446 # define BN_UMULT_LOHI(low,high,a,b) ((low)=_umul128((a),(b),&(high)))
447 # endif
448 # elif defined(__mips) && (defined(SIXTY_FOUR_BIT) || defined(SIXTY_FOUR_BIT_LONG))
449 # if defined(__GNUC__) && __GNUC__>=2
450 # define BN_UMULT_HIGH(a,b) ({ \
451 register BN_ULONG ret; \
452 asm ("dmultu %1,%2" \
453 : "=h"(ret) \
454 : "r"(a), "r"(b) : "l"); \
455 ret; })
456 # define BN_UMULT_LOHI(low,high,a,b) \
457 asm ("dmultu %2,%3" \
458 : "=l"(low),"=h"(high) \
459 : "r"(a), "r"(b));
460 # endif
461 # elif defined(__aarch64__) && defined(SIXTY_FOUR_BIT_LONG)
462 # if defined(__GNUC__) && __GNUC__>=2
463 # define BN_UMULT_HIGH(a,b) ({ \
464 register BN_ULONG ret; \
465 asm ("umulh %0,%1,%2" \
466 : "=r"(ret) \
467 : "r"(a), "r"(b)); \
468 ret; })
469 # endif
470 # endif /* cpu */
471 # endif /* OPENSSL_NO_ASM */
472
473 # ifdef BN_RAND_DEBUG
474 # define bn_clear_top2max(a) \
475 { \
476 int ind = (a)->dmax - (a)->top; \
477 BN_ULONG *ftl = &(a)->d[(a)->top-1]; \
478 for (; ind != 0; ind--) \
479 *(++ftl) = 0x0; \
480 }
481 # else
482 # define bn_clear_top2max(a)
483 # endif
484
485 # ifdef BN_LLONG
486 /*******************************************************************
487 * Using the long long type, has to be twice as wide as BN_ULONG...
488 */
489 # define Lw(t) (((BN_ULONG)(t))&BN_MASK2)
490 # define Hw(t) (((BN_ULONG)((t)>>BN_BITS2))&BN_MASK2)
491
492 # define mul_add(r,a,w,c) { \
493 BN_ULLONG t; \
494 t=(BN_ULLONG)w * (a) + (r) + (c); \
495 (r)= Lw(t); \
496 (c)= Hw(t); \
497 }
498
499 # define mul(r,a,w,c) { \
500 BN_ULLONG t; \
501 t=(BN_ULLONG)w * (a) + (c); \
502 (r)= Lw(t); \
503 (c)= Hw(t); \
504 }
505
506 # define sqr(r0,r1,a) { \
507 BN_ULLONG t; \
508 t=(BN_ULLONG)(a)*(a); \
509 (r0)=Lw(t); \
510 (r1)=Hw(t); \
511 }
512
513 # elif defined(BN_UMULT_LOHI)
514 # define mul_add(r,a,w,c) { \
515 BN_ULONG high,low,ret,tmp=(a); \
516 ret = (r); \
517 BN_UMULT_LOHI(low,high,w,tmp); \
518 ret += (c); \
519 (c) = (ret<(c))?1:0; \
520 (c) += high; \
521 ret += low; \
522 (c) += (ret<low)?1:0; \
523 (r) = ret; \
524 }
525
526 # define mul(r,a,w,c) { \
527 BN_ULONG high,low,ret,ta=(a); \
528 BN_UMULT_LOHI(low,high,w,ta); \
529 ret = low + (c); \
530 (c) = high; \
531 (c) += (ret<low)?1:0; \
532 (r) = ret; \
533 }
534
535 # define sqr(r0,r1,a) { \
536 BN_ULONG tmp=(a); \
537 BN_UMULT_LOHI(r0,r1,tmp,tmp); \
538 }
539
540 # elif defined(BN_UMULT_HIGH)
541 # define mul_add(r,a,w,c) { \
542 BN_ULONG high,low,ret,tmp=(a); \
543 ret = (r); \
544 high= BN_UMULT_HIGH(w,tmp); \
545 ret += (c); \
546 low = (w) * tmp; \
547 (c) = (ret<(c))?1:0; \
548 (c) += high; \
549 ret += low; \
550 (c) += (ret<low)?1:0; \
551 (r) = ret; \
552 }
553
554 # define mul(r,a,w,c) { \
555 BN_ULONG high,low,ret,ta=(a); \
556 low = (w) * ta; \
557 high= BN_UMULT_HIGH(w,ta); \
558 ret = low + (c); \
559 (c) = high; \
560 (c) += (ret<low)?1:0; \
561 (r) = ret; \
562 }
563
564 # define sqr(r0,r1,a) { \
565 BN_ULONG tmp=(a); \
566 (r0) = tmp * tmp; \
567 (r1) = BN_UMULT_HIGH(tmp,tmp); \
568 }
569
570 # else
571 /*************************************************************
572 * No long long type
573 */
574
575 # define LBITS(a) ((a)&BN_MASK2l)
576 # define HBITS(a) (((a)>>BN_BITS4)&BN_MASK2l)
577 # define L2HBITS(a) (((a)<<BN_BITS4)&BN_MASK2)
578
579 # define LLBITS(a) ((a)&BN_MASKl)
580 # define LHBITS(a) (((a)>>BN_BITS2)&BN_MASKl)
581 # define LL2HBITS(a) ((BN_ULLONG)((a)&BN_MASKl)<<BN_BITS2)
582
583 # define mul64(l,h,bl,bh) \
584 { \
585 BN_ULONG m,m1,lt,ht; \
586 \
587 lt=l; \
588 ht=h; \
589 m =(bh)*(lt); \
590 lt=(bl)*(lt); \
591 m1=(bl)*(ht); \
592 ht =(bh)*(ht); \
593 m=(m+m1)&BN_MASK2; if (m < m1) ht+=L2HBITS((BN_ULONG)1); \
594 ht+=HBITS(m); \
595 m1=L2HBITS(m); \
596 lt=(lt+m1)&BN_MASK2; if (lt < m1) ht++; \
597 (l)=lt; \
598 (h)=ht; \
599 }
600
601 # define sqr64(lo,ho,in) \
602 { \
603 BN_ULONG l,h,m; \
604 \
605 h=(in); \
606 l=LBITS(h); \
607 h=HBITS(h); \
608 m =(l)*(h); \
609 l*=l; \
610 h*=h; \
611 h+=(m&BN_MASK2h1)>>(BN_BITS4-1); \
612 m =(m&BN_MASK2l)<<(BN_BITS4+1); \
613 l=(l+m)&BN_MASK2; if (l < m) h++; \
614 (lo)=l; \
615 (ho)=h; \
616 }
617
618 # define mul_add(r,a,bl,bh,c) { \
619 BN_ULONG l,h; \
620 \
621 h= (a); \
622 l=LBITS(h); \
623 h=HBITS(h); \
624 mul64(l,h,(bl),(bh)); \
625 \
626 /* non-multiply part */ \
627 l=(l+(c))&BN_MASK2; if (l < (c)) h++; \
628 (c)=(r); \
629 l=(l+(c))&BN_MASK2; if (l < (c)) h++; \
630 (c)=h&BN_MASK2; \
631 (r)=l; \
632 }
633
634 # define mul(r,a,bl,bh,c) { \
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); if ((l&BN_MASK2) < (c)) h++; \
644 (c)=h&BN_MASK2; \
645 (r)=l&BN_MASK2; \
646 }
647 # endif /* !BN_LLONG */
648
649 void BN_RECP_CTX_init(BN_RECP_CTX *recp);
650 void BN_MONT_CTX_init(BN_MONT_CTX *ctx);
651
652 void bn_init(BIGNUM *a);
653 void bn_mul_normal(BN_ULONG *r, BN_ULONG *a, int na, BN_ULONG *b, int nb);
654 void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b);
655 void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b);
656 void bn_sqr_normal(BN_ULONG *r, const BN_ULONG *a, int n, BN_ULONG *tmp);
657 void bn_sqr_comba8(BN_ULONG *r, const BN_ULONG *a);
658 void bn_sqr_comba4(BN_ULONG *r, const BN_ULONG *a);
659 int bn_cmp_words(const BN_ULONG *a, const BN_ULONG *b, int n);
660 int bn_cmp_part_words(const BN_ULONG *a, const BN_ULONG *b, int cl, int dl);
661 void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
662 int dna, int dnb, BN_ULONG *t);
663 void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b,
664 int n, int tna, int tnb, BN_ULONG *t);
665 void bn_sqr_recursive(BN_ULONG *r, const BN_ULONG *a, int n2, BN_ULONG *t);
666 void bn_mul_low_normal(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n);
667 void bn_mul_low_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
668 BN_ULONG *t);
669 BN_ULONG bn_sub_part_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b,
670 int cl, int dl);
671 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
672 const BN_ULONG *np, const BN_ULONG *n0, int num);
673
674 BIGNUM *int_bn_mod_inverse(BIGNUM *in,
675 const BIGNUM *a, const BIGNUM *n, BN_CTX *ctx,
676 int *noinv);
677
bn_expand(BIGNUM * a,int bits)678 static ossl_inline BIGNUM *bn_expand(BIGNUM *a, int bits)
679 {
680 if (bits > (INT_MAX - BN_BITS2 + 1))
681 return NULL;
682
683 if (((bits+BN_BITS2-1)/BN_BITS2) <= (a)->dmax)
684 return a;
685
686 return bn_expand2((a),(bits+BN_BITS2-1)/BN_BITS2);
687 }
688
689 int ossl_bn_check_prime(const BIGNUM *w, int checks, BN_CTX *ctx,
690 int do_trial_division, BN_GENCB *cb);
691
692 #endif
693