]> git.ipfire.org Git - thirdparty/openssl.git/blob - crypto/bn/bn_asm.c
projects / thirdparty / openssl.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Support for "multiply high" instruction, see BN_UMULT_HIGH comment in
[thirdparty/openssl.git] / crypto / bn / bn_asm.c
1 /* crypto/bn/bn_asm.c */
2 /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
3 * All rights reserved.
4 *
5 * This package is an SSL implementation written
6 * by Eric Young (eay@cryptsoft.com).
7 * The implementation was written so as to conform with Netscapes SSL.
8 *
9 * This library is free for commercial and non-commercial use as long as
10 * the following conditions are aheared to. The following conditions
11 * apply to all code found in this distribution, be it the RC4, RSA,
12 * lhash, DES, etc., code; not just the SSL code. The SSL documentation
13 * included with this distribution is covered by the same copyright terms
14 * except that the holder is Tim Hudson (tjh@cryptsoft.com).
15 *
16 * Copyright remains Eric Young's, and as such any Copyright notices in
17 * the code are not to be removed.
18 * If this package is used in a product, Eric Young should be given attribution
19 * as the author of the parts of the library used.
20 * This can be in the form of a textual message at program startup or
21 * in documentation (online or textual) provided with the package.
22 *
23 * Redistribution and use in source and binary forms, with or without
24 * modification, are permitted provided that the following conditions
25 * are met:
26 * 1. Redistributions of source code must retain the copyright
27 * notice, this list of conditions and the following disclaimer.
28 * 2. Redistributions in binary form must reproduce the above copyright
29 * notice, this list of conditions and the following disclaimer in the
30 * documentation and/or other materials provided with the distribution.
31 * 3. All advertising materials mentioning features or use of this software
32 * must display the following acknowledgement:
33 * "This product includes cryptographic software written by
34 * Eric Young (eay@cryptsoft.com)"
35 * The word 'cryptographic' can be left out if the rouines from the library
36 * being used are not cryptographic related :-).
37 * 4. If you include any Windows specific code (or a derivative thereof) from
38 * the apps directory (application code) you must include an acknowledgement:
39 * "This product includes software written by Tim Hudson (tjh@cryptsoft.com)"
40 *
41 * THIS SOFTWARE IS PROVIDED BY ERIC YOUNG ``AS IS'' AND
42 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
43 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
44 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
45 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
46 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
47 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
48 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
49 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
50 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
51 * SUCH DAMAGE.
52 *
53 * The licence and distribution terms for any publically available version or
54 * derivative of this code cannot be changed. i.e. this code cannot simply be
55 * copied and put under another distribution licence
56 * [including the GNU Public Licence.]
57 */
58
59 #include <stdio.h>
60 #include "cryptlib.h"
61 #include "bn_lcl.h"
62
63 #if defined(BN_LLONG) || defined(BN_UMULT_HIGH)
64
65 BN_ULONG bn_mul_add_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
66 {
67 BN_ULONG c1=0;
68
69 bn_check_num(num);
70 if (num <= 0) return(c1);
71
72 while (num&~3)
73 {
74 mul_add(rp[0],ap[0],w,c1);
75 mul_add(rp[1],ap[1],w,c1);
76 mul_add(rp[2],ap[2],w,c1);
77 mul_add(rp[3],ap[3],w,c1);
78 ap+=4; rp+=4; num-=4;
79 }
80 if (num)
81 {
82 mul_add(rp[0],ap[0],w,c1); if (--num==0) return c1;
83 mul_add(rp[1],ap[1],w,c1); if (--num==0) return c1;
84 mul_add(rp[2],ap[2],w,c1); return c1;
85 }
86
87 return(c1);
88 }
89
90 BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
91 {
92 BN_ULONG c1=0;
93
94 bn_check_num(num);
95 if (num <= 0) return(c1);
96
97 while (num&~3)
98 {
99 mul(rp[0],ap[0],w,c1);
100 mul(rp[1],ap[1],w,c1);
101 mul(rp[2],ap[2],w,c1);
102 mul(rp[3],ap[3],w,c1);
103 ap+=4; rp+=4; num-=4;
104 }
105 if (num)
106 {
107 mul(rp[0],ap[0],w,c1); if (--num == 0) return c1;
108 mul(rp[1],ap[1],w,c1); if (--num == 0) return c1;
109 mul(rp[2],ap[2],w,c1);
110 }
111 return(c1);
112 }
113
114 void bn_sqr_words(BN_ULONG *r, BN_ULONG *a, int n)
115 {
116 bn_check_num(n);
117 if (n <= 0) return;
118 while (n&~3)
119 {
120 sqr(r[0],r[1],a[0]);
121 sqr(r[2],r[3],a[1]);
122 sqr(r[4],r[5],a[2]);
123 sqr(r[6],r[7],a[3]);
124 a+=4; r+=8; n-=4;
125 }
126 if (n)
127 {
128 sqr(r[0],r[1],a[0]); if (--n == 0) return;
129 sqr(r[2],r[3],a[1]); if (--n == 0) return;
130 sqr(r[4],r[5],a[2]);
131 }
132 }
133
134 #else
135
136 BN_ULONG bn_mul_add_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
137 {
138 BN_ULONG c=0;
139 BN_ULONG bl,bh;
140
141 bn_check_num(num);
142 if (num <= 0) return((BN_ULONG)0);
143
144 bl=LBITS(w);
145 bh=HBITS(w);
146
147 for (;;)
148 {
149 mul_add(rp[0],ap[0],bl,bh,c);
150 if (--num == 0) break;
151 mul_add(rp[1],ap[1],bl,bh,c);
152 if (--num == 0) break;
153 mul_add(rp[2],ap[2],bl,bh,c);
154 if (--num == 0) break;
155 mul_add(rp[3],ap[3],bl,bh,c);
156 if (--num == 0) break;
157 ap+=4;
158 rp+=4;
159 }
160 return(c);
161 }
162
163 BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
164 {
165 BN_ULONG carry=0;
166 BN_ULONG bl,bh;
167
168 bn_check_num(num);
169 if (num <= 0) return((BN_ULONG)0);
170
171 bl=LBITS(w);
172 bh=HBITS(w);
173
174 for (;;)
175 {
176 mul(rp[0],ap[0],bl,bh,carry);
177 if (--num == 0) break;
178 mul(rp[1],ap[1],bl,bh,carry);
179 if (--num == 0) break;
180 mul(rp[2],ap[2],bl,bh,carry);
181 if (--num == 0) break;
182 mul(rp[3],ap[3],bl,bh,carry);
183 if (--num == 0) break;
184 ap+=4;
185 rp+=4;
186 }
187 return(carry);
188 }
189
190 void bn_sqr_words(BN_ULONG *r, BN_ULONG *a, int n)
191 {
192 bn_check_num(n);
193 if (n <= 0) return;
194 for (;;)
195 {
196 sqr64(r[0],r[1],a[0]);
197 if (--n == 0) break;
198
199 sqr64(r[2],r[3],a[1]);
200 if (--n == 0) break;
201
202 sqr64(r[4],r[5],a[2]);
203 if (--n == 0) break;
204
205 sqr64(r[6],r[7],a[3]);
206 if (--n == 0) break;
207
208 a+=4;
209 r+=8;
210 }
211 }
212
213 #endif
214
215 #if defined(BN_LLONG) && defined(BN_DIV2W)
216
217 BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d)
218 {
219 return((BN_ULONG)(((((BN_ULLONG)h)<<BN_BITS2)|l)/(BN_ULLONG)d));
220 }
221
222 #else
223
224 /* Divide h-l by d and return the result. */
225 /* I need to test this some more :-( */
226 BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d)
227 {
228 BN_ULONG dh,dl,q,ret=0,th,tl,t;
229 int i,count=2;
230
231 if (d == 0) return(BN_MASK2);
232
233 i=BN_num_bits_word(d);
234 if ((i != BN_BITS2) && (h > (BN_ULONG)1<<i))
235 {
236 #if !defined(NO_STDIO) && !defined(WIN16)
237 fprintf(stderr,"Division would overflow (%d)\n",i);
238 #endif
239 abort();
240 }
241 i=BN_BITS2-i;
242 if (h >= d) h-=d;
243
244 if (i)
245 {
246 d<<=i;
247 h=(h<<i)|(l>>(BN_BITS2-i));
248 l<<=i;
249 }
250 dh=(d&BN_MASK2h)>>BN_BITS4;
251 dl=(d&BN_MASK2l);
252 for (;;)
253 {
254 if ((h>>BN_BITS4) == dh)
255 q=BN_MASK2l;
256 else
257 q=h/dh;
258
259 th=q*dh;
260 tl=dl*q;
261 for (;;)
262 {
263 t=h-th;
264 if ((t&BN_MASK2h) ||
265 ((tl) <= (
266 (t<<BN_BITS4)|
267 ((l&BN_MASK2h)>>BN_BITS4))))
268 break;
269 q--;
270 th-=dh;
271 tl-=dl;
272 }
273 t=(tl>>BN_BITS4);
274 tl=(tl<<BN_BITS4)&BN_MASK2h;
275 th+=t;
276
277 if (l < tl) th++;
278 l-=tl;
279 if (h < th)
280 {
281 h+=d;
282 q--;
283 }
284 h-=th;
285
286 if (--count == 0) break;
287
288 ret=q<<BN_BITS4;
289 h=((h<<BN_BITS4)|(l>>BN_BITS4))&BN_MASK2;
290 l=(l&BN_MASK2l)<<BN_BITS4;
291 }
292 ret|=q;
293 return(ret);
294 }
295 #endif
296
297 #ifdef BN_LLONG
298 BN_ULONG bn_add_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
299 {
300 BN_ULLONG ll=0;
301
302 bn_check_num(n);
303 if (n <= 0) return((BN_ULONG)0);
304
305 for (;;)
306 {
307 ll+=(BN_ULLONG)a[0]+b[0];
308 r[0]=(BN_ULONG)ll&BN_MASK2;
309 ll>>=BN_BITS2;
310 if (--n <= 0) break;
311
312 ll+=(BN_ULLONG)a[1]+b[1];
313 r[1]=(BN_ULONG)ll&BN_MASK2;
314 ll>>=BN_BITS2;
315 if (--n <= 0) break;
316
317 ll+=(BN_ULLONG)a[2]+b[2];
318 r[2]=(BN_ULONG)ll&BN_MASK2;
319 ll>>=BN_BITS2;
320 if (--n <= 0) break;
321
322 ll+=(BN_ULLONG)a[3]+b[3];
323 r[3]=(BN_ULONG)ll&BN_MASK2;
324 ll>>=BN_BITS2;
325 if (--n <= 0) break;
326
327 a+=4;
328 b+=4;
329 r+=4;
330 }
331 return((BN_ULONG)ll);
332 }
333 #else
334 BN_ULONG bn_add_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
335 {
336 BN_ULONG c,l,t;
337
338 bn_check_num(n);
339 if (n <= 0) return((BN_ULONG)0);
340
341 c=0;
342 for (;;)
343 {
344 t=a[0];
345 t=(t+c)&BN_MASK2;
346 c=(t < c);
347 l=(t+b[0])&BN_MASK2;
348 c+=(l < t);
349 r[0]=l;
350 if (--n <= 0) break;
351
352 t=a[1];
353 t=(t+c)&BN_MASK2;
354 c=(t < c);
355 l=(t+b[1])&BN_MASK2;
356 c+=(l < t);
357 r[1]=l;
358 if (--n <= 0) break;
359
360 t=a[2];
361 t=(t+c)&BN_MASK2;
362 c=(t < c);
363 l=(t+b[2])&BN_MASK2;
364 c+=(l < t);
365 r[2]=l;
366 if (--n <= 0) break;
367
368 t=a[3];
369 t=(t+c)&BN_MASK2;
370 c=(t < c);
371 l=(t+b[3])&BN_MASK2;
372 c+=(l < t);
373 r[3]=l;
374 if (--n <= 0) break;
375
376 a+=4;
377 b+=4;
378 r+=4;
379 }
380 return((BN_ULONG)c);
381 }
382 #endif
383
384 BN_ULONG bn_sub_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
385 {
386 BN_ULONG t1,t2;
387 int c=0;
388
389 bn_check_num(n);
390 if (n <= 0) return((BN_ULONG)0);
391
392 for (;;)
393 {
394 t1=a[0]; t2=b[0];
395 r[0]=(t1-t2-c)&BN_MASK2;
396 if (t1 != t2) c=(t1 < t2);
397 if (--n <= 0) break;
398
399 t1=a[1]; t2=b[1];
400 r[1]=(t1-t2-c)&BN_MASK2;
401 if (t1 != t2) c=(t1 < t2);
402 if (--n <= 0) break;
403
404 t1=a[2]; t2=b[2];
405 r[2]=(t1-t2-c)&BN_MASK2;
406 if (t1 != t2) c=(t1 < t2);
407 if (--n <= 0) break;
408
409 t1=a[3]; t2=b[3];
410 r[3]=(t1-t2-c)&BN_MASK2;
411 if (t1 != t2) c=(t1 < t2);
412 if (--n <= 0) break;
413
414 a+=4;
415 b+=4;
416 r+=4;
417 }
418 return(c);
419 }
420
421 #ifdef BN_MUL_COMBA
422
423 #undef bn_mul_comba8
424 #undef bn_mul_comba4
425 #undef bn_sqr_comba8
426 #undef bn_sqr_comba4
427
428 #ifdef BN_LLONG
429 #define mul_add_c(a,b,c0,c1,c2) \
430 t=(BN_ULLONG)a*b; \
431 t1=(BN_ULONG)Lw(t); \
432 t2=(BN_ULONG)Hw(t); \
433 c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \
434 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
435
436 #define mul_add_c2(a,b,c0,c1,c2) \
437 t=(BN_ULLONG)a*b; \
438 tt=(t+t)&BN_MASK; \
439 if (tt < t) c2++; \
440 t1=(BN_ULONG)Lw(tt); \
441 t2=(BN_ULONG)Hw(tt); \
442 c0=(c0+t1)&BN_MASK2; \
443 if ((c0 < t1) && (((++t2)&BN_MASK2) == 0)) c2++; \
444 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
445
446 #define sqr_add_c(a,i,c0,c1,c2) \
447 t=(BN_ULLONG)a[i]*a[i]; \
448 t1=(BN_ULONG)Lw(t); \
449 t2=(BN_ULONG)Hw(t); \
450 c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \
451 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
452
453 #define sqr_add_c2(a,i,j,c0,c1,c2) \
454 mul_add_c2((a)[i],(a)[j],c0,c1,c2)
455
456 #elif defined(BN_UMULT_HIGH)
457
458 #define mul_add_c(a,b,c0,c1,c2) { \
459 BN_ULONG ta=(a),tb=(b); \
460 t1 = ta * tb; \
461 t2 = BN_UMULT_HIGH(ta,tb); \
462 c0 += t1; t2 += (c0<t1)?1:0; \
463 c1 += t2; c2 += (c1<t2)?1:0; \
464 }
465
466 #define mul_add_c2(a,b,c0,c1,c2) { \
467 BN_ULONG ta=(a),tb=(b),t0; \
468 t1 = BN_UMULT_HIGH(ta,tb); \
469 t0 = ta * tb; \
470 t2 = t1+t1; c2 += (t2<t1)?1:0; \
471 t1 = t0+t0; t2 += (t1<t0)?1:0; \
472 c0 += t1; t2 += (c0<t1)?1:0; \
473 c1 += t2; c2 += (c1<t2)?1:0; \
474 }
475
476 #define sqr_add_c(a,i,c0,c1,c2) { \
477 BN_ULONG ta=(a)[i]; \
478 t1 = ta * ta; \
479 t2 = BN_UMULT_HIGH(ta,ta); \
480 c0 += t1; t2 += (c0<t1)?1:0; \
481 c1 += t2; c2 += (c1<t2)?1:0; \
482 }
483
484 #define sqr_add_c2(a,i,j,c0,c1,c2) \
485 mul_add_c2((a)[i],(a)[j],c0,c1,c2)
486
487 #else
488 #define mul_add_c(a,b,c0,c1,c2) \
489 t1=LBITS(a); t2=HBITS(a); \
490 bl=LBITS(b); bh=HBITS(b); \
491 mul64(t1,t2,bl,bh); \
492 c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \
493 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
494
495 #define mul_add_c2(a,b,c0,c1,c2) \
496 t1=LBITS(a); t2=HBITS(a); \
497 bl=LBITS(b); bh=HBITS(b); \
498 mul64(t1,t2,bl,bh); \
499 if (t2 & BN_TBIT) c2++; \
500 t2=(t2+t2)&BN_MASK2; \
501 if (t1 & BN_TBIT) t2++; \
502 t1=(t1+t1)&BN_MASK2; \
503 c0=(c0+t1)&BN_MASK2; \
504 if ((c0 < t1) && (((++t2)&BN_MASK2) == 0)) c2++; \
505 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
506
507 #define sqr_add_c(a,i,c0,c1,c2) \
508 sqr64(t1,t2,(a)[i]); \
509 c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \
510 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
511
512 #define sqr_add_c2(a,i,j,c0,c1,c2) \
513 mul_add_c2((a)[i],(a)[j],c0,c1,c2)
514 #endif
515
516 void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
517 {
518 #ifdef BN_LLONG
519 BN_ULLONG t;
520 #else
521 BN_ULONG bl,bh;
522 #endif
523 BN_ULONG t1,t2;
524 BN_ULONG c1,c2,c3;
525
526 c1=0;
527 c2=0;
528 c3=0;
529 mul_add_c(a[0],b[0],c1,c2,c3);
530 r[0]=c1;
531 c1=0;
532 mul_add_c(a[0],b[1],c2,c3,c1);
533 mul_add_c(a[1],b[0],c2,c3,c1);
534 r[1]=c2;
535 c2=0;
536 mul_add_c(a[2],b[0],c3,c1,c2);
537 mul_add_c(a[1],b[1],c3,c1,c2);
538 mul_add_c(a[0],b[2],c3,c1,c2);
539 r[2]=c3;
540 c3=0;
541 mul_add_c(a[0],b[3],c1,c2,c3);
542 mul_add_c(a[1],b[2],c1,c2,c3);
543 mul_add_c(a[2],b[1],c1,c2,c3);
544 mul_add_c(a[3],b[0],c1,c2,c3);
545 r[3]=c1;
546 c1=0;
547 mul_add_c(a[4],b[0],c2,c3,c1);
548 mul_add_c(a[3],b[1],c2,c3,c1);
549 mul_add_c(a[2],b[2],c2,c3,c1);
550 mul_add_c(a[1],b[3],c2,c3,c1);
551 mul_add_c(a[0],b[4],c2,c3,c1);
552 r[4]=c2;
553 c2=0;
554 mul_add_c(a[0],b[5],c3,c1,c2);
555 mul_add_c(a[1],b[4],c3,c1,c2);
556 mul_add_c(a[2],b[3],c3,c1,c2);
557 mul_add_c(a[3],b[2],c3,c1,c2);
558 mul_add_c(a[4],b[1],c3,c1,c2);
559 mul_add_c(a[5],b[0],c3,c1,c2);
560 r[5]=c3;
561 c3=0;
562 mul_add_c(a[6],b[0],c1,c2,c3);
563 mul_add_c(a[5],b[1],c1,c2,c3);
564 mul_add_c(a[4],b[2],c1,c2,c3);
565 mul_add_c(a[3],b[3],c1,c2,c3);
566 mul_add_c(a[2],b[4],c1,c2,c3);
567 mul_add_c(a[1],b[5],c1,c2,c3);
568 mul_add_c(a[0],b[6],c1,c2,c3);
569 r[6]=c1;
570 c1=0;
571 mul_add_c(a[0],b[7],c2,c3,c1);
572 mul_add_c(a[1],b[6],c2,c3,c1);
573 mul_add_c(a[2],b[5],c2,c3,c1);
574 mul_add_c(a[3],b[4],c2,c3,c1);
575 mul_add_c(a[4],b[3],c2,c3,c1);
576 mul_add_c(a[5],b[2],c2,c3,c1);
577 mul_add_c(a[6],b[1],c2,c3,c1);
578 mul_add_c(a[7],b[0],c2,c3,c1);
579 r[7]=c2;
580 c2=0;
581 mul_add_c(a[7],b[1],c3,c1,c2);
582 mul_add_c(a[6],b[2],c3,c1,c2);
583 mul_add_c(a[5],b[3],c3,c1,c2);
584 mul_add_c(a[4],b[4],c3,c1,c2);
585 mul_add_c(a[3],b[5],c3,c1,c2);
586 mul_add_c(a[2],b[6],c3,c1,c2);
587 mul_add_c(a[1],b[7],c3,c1,c2);
588 r[8]=c3;
589 c3=0;
590 mul_add_c(a[2],b[7],c1,c2,c3);
591 mul_add_c(a[3],b[6],c1,c2,c3);
592 mul_add_c(a[4],b[5],c1,c2,c3);
593 mul_add_c(a[5],b[4],c1,c2,c3);
594 mul_add_c(a[6],b[3],c1,c2,c3);
595 mul_add_c(a[7],b[2],c1,c2,c3);
596 r[9]=c1;
597 c1=0;
598 mul_add_c(a[7],b[3],c2,c3,c1);
599 mul_add_c(a[6],b[4],c2,c3,c1);
600 mul_add_c(a[5],b[5],c2,c3,c1);
601 mul_add_c(a[4],b[6],c2,c3,c1);
602 mul_add_c(a[3],b[7],c2,c3,c1);
603 r[10]=c2;
604 c2=0;
605 mul_add_c(a[4],b[7],c3,c1,c2);
606 mul_add_c(a[5],b[6],c3,c1,c2);
607 mul_add_c(a[6],b[5],c3,c1,c2);
608 mul_add_c(a[7],b[4],c3,c1,c2);
609 r[11]=c3;
610 c3=0;
611 mul_add_c(a[7],b[5],c1,c2,c3);
612 mul_add_c(a[6],b[6],c1,c2,c3);
613 mul_add_c(a[5],b[7],c1,c2,c3);
614 r[12]=c1;
615 c1=0;
616 mul_add_c(a[6],b[7],c2,c3,c1);
617 mul_add_c(a[7],b[6],c2,c3,c1);
618 r[13]=c2;
619 c2=0;
620 mul_add_c(a[7],b[7],c3,c1,c2);
621 r[14]=c3;
622 r[15]=c1;
623 }
624
625 void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
626 {
627 #ifdef BN_LLONG
628 BN_ULLONG t;
629 #else
630 BN_ULONG bl,bh;
631 #endif
632 BN_ULONG t1,t2;
633 BN_ULONG c1,c2,c3;
634
635 c1=0;
636 c2=0;
637 c3=0;
638 mul_add_c(a[0],b[0],c1,c2,c3);
639 r[0]=c1;
640 c1=0;
641 mul_add_c(a[0],b[1],c2,c3,c1);
642 mul_add_c(a[1],b[0],c2,c3,c1);
643 r[1]=c2;
644 c2=0;
645 mul_add_c(a[2],b[0],c3,c1,c2);
646 mul_add_c(a[1],b[1],c3,c1,c2);
647 mul_add_c(a[0],b[2],c3,c1,c2);
648 r[2]=c3;
649 c3=0;
650 mul_add_c(a[0],b[3],c1,c2,c3);
651 mul_add_c(a[1],b[2],c1,c2,c3);
652 mul_add_c(a[2],b[1],c1,c2,c3);
653 mul_add_c(a[3],b[0],c1,c2,c3);
654 r[3]=c1;
655 c1=0;
656 mul_add_c(a[3],b[1],c2,c3,c1);
657 mul_add_c(a[2],b[2],c2,c3,c1);
658 mul_add_c(a[1],b[3],c2,c3,c1);
659 r[4]=c2;
660 c2=0;
661 mul_add_c(a[2],b[3],c3,c1,c2);
662 mul_add_c(a[3],b[2],c3,c1,c2);
663 r[5]=c3;
664 c3=0;
665 mul_add_c(a[3],b[3],c1,c2,c3);
666 r[6]=c1;
667 r[7]=c2;
668 }
669
670 void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a)
671 {
672 #ifdef BN_LLONG
673 BN_ULLONG t,tt;
674 #else
675 BN_ULONG bl,bh;
676 #endif
677 BN_ULONG t1,t2;
678 BN_ULONG c1,c2,c3;
679
680 c1=0;
681 c2=0;
682 c3=0;
683 sqr_add_c(a,0,c1,c2,c3);
684 r[0]=c1;
685 c1=0;
686 sqr_add_c2(a,1,0,c2,c3,c1);
687 r[1]=c2;
688 c2=0;
689 sqr_add_c(a,1,c3,c1,c2);
690 sqr_add_c2(a,2,0,c3,c1,c2);
691 r[2]=c3;
692 c3=0;
693 sqr_add_c2(a,3,0,c1,c2,c3);
694 sqr_add_c2(a,2,1,c1,c2,c3);
695 r[3]=c1;
696 c1=0;
697 sqr_add_c(a,2,c2,c3,c1);
698 sqr_add_c2(a,3,1,c2,c3,c1);
699 sqr_add_c2(a,4,0,c2,c3,c1);
700 r[4]=c2;
701 c2=0;
702 sqr_add_c2(a,5,0,c3,c1,c2);
703 sqr_add_c2(a,4,1,c3,c1,c2);
704 sqr_add_c2(a,3,2,c3,c1,c2);
705 r[5]=c3;
706 c3=0;
707 sqr_add_c(a,3,c1,c2,c3);
708 sqr_add_c2(a,4,2,c1,c2,c3);
709 sqr_add_c2(a,5,1,c1,c2,c3);
710 sqr_add_c2(a,6,0,c1,c2,c3);
711 r[6]=c1;
712 c1=0;
713 sqr_add_c2(a,7,0,c2,c3,c1);
714 sqr_add_c2(a,6,1,c2,c3,c1);
715 sqr_add_c2(a,5,2,c2,c3,c1);
716 sqr_add_c2(a,4,3,c2,c3,c1);
717 r[7]=c2;
718 c2=0;
719 sqr_add_c(a,4,c3,c1,c2);
720 sqr_add_c2(a,5,3,c3,c1,c2);
721 sqr_add_c2(a,6,2,c3,c1,c2);
722 sqr_add_c2(a,7,1,c3,c1,c2);
723 r[8]=c3;
724 c3=0;
725 sqr_add_c2(a,7,2,c1,c2,c3);
726 sqr_add_c2(a,6,3,c1,c2,c3);
727 sqr_add_c2(a,5,4,c1,c2,c3);
728 r[9]=c1;
729 c1=0;
730 sqr_add_c(a,5,c2,c3,c1);
731 sqr_add_c2(a,6,4,c2,c3,c1);
732 sqr_add_c2(a,7,3,c2,c3,c1);
733 r[10]=c2;
734 c2=0;
735 sqr_add_c2(a,7,4,c3,c1,c2);
736 sqr_add_c2(a,6,5,c3,c1,c2);
737 r[11]=c3;
738 c3=0;
739 sqr_add_c(a,6,c1,c2,c3);
740 sqr_add_c2(a,7,5,c1,c2,c3);
741 r[12]=c1;
742 c1=0;
743 sqr_add_c2(a,7,6,c2,c3,c1);
744 r[13]=c2;
745 c2=0;
746 sqr_add_c(a,7,c3,c1,c2);
747 r[14]=c3;
748 r[15]=c1;
749 }
750
751 void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a)
752 {
753 #ifdef BN_LLONG
754 BN_ULLONG t,tt;
755 #else
756 BN_ULONG bl,bh;
757 #endif
758 BN_ULONG t1,t2;
759 BN_ULONG c1,c2,c3;
760
761 c1=0;
762 c2=0;
763 c3=0;
764 sqr_add_c(a,0,c1,c2,c3);
765 r[0]=c1;
766 c1=0;
767 sqr_add_c2(a,1,0,c2,c3,c1);
768 r[1]=c2;
769 c2=0;
770 sqr_add_c(a,1,c3,c1,c2);
771 sqr_add_c2(a,2,0,c3,c1,c2);
772 r[2]=c3;
773 c3=0;
774 sqr_add_c2(a,3,0,c1,c2,c3);
775 sqr_add_c2(a,2,1,c1,c2,c3);
776 r[3]=c1;
777 c1=0;
778 sqr_add_c(a,2,c2,c3,c1);
779 sqr_add_c2(a,3,1,c2,c3,c1);
780 r[4]=c2;
781 c2=0;
782 sqr_add_c2(a,3,2,c3,c1,c2);
783 r[5]=c3;
784 c3=0;
785 sqr_add_c(a,3,c1,c2,c3);
786 r[6]=c1;
787 r[7]=c2;
788 }
789 #else
790
791 /* hmm... is it faster just to do a multiply? */
792 #undef bn_sqr_comba4
793 void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a)
794 {
795 BN_ULONG t[8];
796 bn_sqr_normal(r,a,4,t);
797 }
798
799 #undef bn_sqr_comba8
800 void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a)
801 {
802 BN_ULONG t[16];
803 bn_sqr_normal(r,a,8,t);
804 }
805
806 void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
807 {
808 r[4]=bn_mul_words( &(r[0]),a,4,b[0]);
809 r[5]=bn_mul_add_words(&(r[1]),a,4,b[1]);
810 r[6]=bn_mul_add_words(&(r[2]),a,4,b[2]);
811 r[7]=bn_mul_add_words(&(r[3]),a,4,b[3]);
812 }
813
814 void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
815 {
816 r[ 8]=bn_mul_words( &(r[0]),a,8,b[0]);
817 r[ 9]=bn_mul_add_words(&(r[1]),a,8,b[1]);
818 r[10]=bn_mul_add_words(&(r[2]),a,8,b[2]);
819 r[11]=bn_mul_add_words(&(r[3]),a,8,b[3]);
820 r[12]=bn_mul_add_words(&(r[4]),a,8,b[4]);
821 r[13]=bn_mul_add_words(&(r[5]),a,8,b[5]);
822 r[14]=bn_mul_add_words(&(r[6]),a,8,b[6]);
823 r[15]=bn_mul_add_words(&(r[7]),a,8,b[7]);
824 }
825
826 #endif /* BN_COMBA */
A mirror of the official openssl repository