summaryrefslogtreecommitdiff
path: root/src/lib/libcrypto/bn/bn_asm.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/lib/libcrypto/bn/bn_asm.c')
-rw-r--r--src/lib/libcrypto/bn/bn_asm.c802
1 files changed, 802 insertions, 0 deletions
diff --git a/src/lib/libcrypto/bn/bn_asm.c b/src/lib/libcrypto/bn/bn_asm.c
new file mode 100644
index 0000000000..4d3da16a0c
--- /dev/null
+++ b/src/lib/libcrypto/bn/bn_asm.c
@@ -0,0 +1,802 @@
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#ifdef BN_LLONG
64
65BN_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 for (;;)
73 {
74 mul_add(rp[0],ap[0],w,c1);
75 if (--num == 0) break;
76 mul_add(rp[1],ap[1],w,c1);
77 if (--num == 0) break;
78 mul_add(rp[2],ap[2],w,c1);
79 if (--num == 0) break;
80 mul_add(rp[3],ap[3],w,c1);
81 if (--num == 0) break;
82 ap+=4;
83 rp+=4;
84 }
85
86 return(c1);
87 }
88
89BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
90 {
91 BN_ULONG c1=0;
92
93 bn_check_num(num);
94 if (num <= 0) return(c1);
95
96 /* for (;;) */
97 while (1) /* circumvent egcs-1.1.2 bug */
98 {
99 mul(rp[0],ap[0],w,c1);
100 if (--num == 0) break;
101 mul(rp[1],ap[1],w,c1);
102 if (--num == 0) break;
103 mul(rp[2],ap[2],w,c1);
104 if (--num == 0) break;
105 mul(rp[3],ap[3],w,c1);
106 if (--num == 0) break;
107 ap+=4;
108 rp+=4;
109 }
110 return(c1);
111 }
112
113void bn_sqr_words(BN_ULONG *r, BN_ULONG *a, int n)
114 {
115 bn_check_num(n);
116 if (n <= 0) return;
117 for (;;)
118 {
119 BN_ULLONG t;
120
121 t=(BN_ULLONG)(a[0])*(a[0]);
122 r[0]=Lw(t); r[1]=Hw(t);
123 if (--n == 0) break;
124
125 t=(BN_ULLONG)(a[1])*(a[1]);
126 r[2]=Lw(t); r[3]=Hw(t);
127 if (--n == 0) break;
128
129 t=(BN_ULLONG)(a[2])*(a[2]);
130 r[4]=Lw(t); r[5]=Hw(t);
131 if (--n == 0) break;
132
133 t=(BN_ULLONG)(a[3])*(a[3]);
134 r[6]=Lw(t); r[7]=Hw(t);
135 if (--n == 0) break;
136
137 a+=4;
138 r+=8;
139 }
140 }
141
142#else
143
144BN_ULONG bn_mul_add_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
145 {
146 BN_ULONG c=0;
147 BN_ULONG bl,bh;
148
149 bn_check_num(num);
150 if (num <= 0) return((BN_ULONG)0);
151
152 bl=LBITS(w);
153 bh=HBITS(w);
154
155 for (;;)
156 {
157 mul_add(rp[0],ap[0],bl,bh,c);
158 if (--num == 0) break;
159 mul_add(rp[1],ap[1],bl,bh,c);
160 if (--num == 0) break;
161 mul_add(rp[2],ap[2],bl,bh,c);
162 if (--num == 0) break;
163 mul_add(rp[3],ap[3],bl,bh,c);
164 if (--num == 0) break;
165 ap+=4;
166 rp+=4;
167 }
168 return(c);
169 }
170
171BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
172 {
173 BN_ULONG carry=0;
174 BN_ULONG bl,bh;
175
176 bn_check_num(num);
177 if (num <= 0) return((BN_ULONG)0);
178
179 bl=LBITS(w);
180 bh=HBITS(w);
181
182 for (;;)
183 {
184 mul(rp[0],ap[0],bl,bh,carry);
185 if (--num == 0) break;
186 mul(rp[1],ap[1],bl,bh,carry);
187 if (--num == 0) break;
188 mul(rp[2],ap[2],bl,bh,carry);
189 if (--num == 0) break;
190 mul(rp[3],ap[3],bl,bh,carry);
191 if (--num == 0) break;
192 ap+=4;
193 rp+=4;
194 }
195 return(carry);
196 }
197
198void bn_sqr_words(BN_ULONG *r, BN_ULONG *a, int n)
199 {
200 bn_check_num(n);
201 if (n <= 0) return;
202 for (;;)
203 {
204 sqr64(r[0],r[1],a[0]);
205 if (--n == 0) break;
206
207 sqr64(r[2],r[3],a[1]);
208 if (--n == 0) break;
209
210 sqr64(r[4],r[5],a[2]);
211 if (--n == 0) break;
212
213 sqr64(r[6],r[7],a[3]);
214 if (--n == 0) break;
215
216 a+=4;
217 r+=8;
218 }
219 }
220
221#endif
222
223#if defined(BN_LLONG) && defined(BN_DIV2W)
224
225BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d)
226 {
227 return((BN_ULONG)(((((BN_ULLONG)h)<<BN_BITS2)|l)/(BN_ULLONG)d));
228 }
229
230#else
231
232/* Divide h-l by d and return the result. */
233/* I need to test this some more :-( */
234BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d)
235 {
236 BN_ULONG dh,dl,q,ret=0,th,tl,t;
237 int i,count=2;
238
239 if (d == 0) return(BN_MASK2);
240
241 i=BN_num_bits_word(d);
242 if ((i != BN_BITS2) && (h > (BN_ULONG)1<<i))
243 {
244#if !defined(NO_STDIO) && !defined(WIN16)
245 fprintf(stderr,"Division would overflow (%d)\n",i);
246#endif
247 abort();
248 }
249 i=BN_BITS2-i;
250 if (h >= d) h-=d;
251
252 if (i)
253 {
254 d<<=i;
255 h=(h<<i)|(l>>(BN_BITS2-i));
256 l<<=i;
257 }
258 dh=(d&BN_MASK2h)>>BN_BITS4;
259 dl=(d&BN_MASK2l);
260 for (;;)
261 {
262 if ((h>>BN_BITS4) == dh)
263 q=BN_MASK2l;
264 else
265 q=h/dh;
266
267 th=q*dh;
268 tl=dl*q;
269 for (;;)
270 {
271 t=h-th;
272 if ((t&BN_MASK2h) ||
273 ((tl) <= (
274 (t<<BN_BITS4)|
275 ((l&BN_MASK2h)>>BN_BITS4))))
276 break;
277 q--;
278 th-=dh;
279 tl-=dl;
280 }
281 t=(tl>>BN_BITS4);
282 tl=(tl<<BN_BITS4)&BN_MASK2h;
283 th+=t;
284
285 if (l < tl) th++;
286 l-=tl;
287 if (h < th)
288 {
289 h+=d;
290 q--;
291 }
292 h-=th;
293
294 if (--count == 0) break;
295
296 ret=q<<BN_BITS4;
297 h=((h<<BN_BITS4)|(l>>BN_BITS4))&BN_MASK2;
298 l=(l&BN_MASK2l)<<BN_BITS4;
299 }
300 ret|=q;
301 return(ret);
302 }
303#endif
304
305#ifdef BN_LLONG
306BN_ULONG bn_add_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
307 {
308 BN_ULLONG ll=0;
309
310 bn_check_num(n);
311 if (n <= 0) return((BN_ULONG)0);
312
313 for (;;)
314 {
315 ll+=(BN_ULLONG)a[0]+b[0];
316 r[0]=(BN_ULONG)ll&BN_MASK2;
317 ll>>=BN_BITS2;
318 if (--n <= 0) break;
319
320 ll+=(BN_ULLONG)a[1]+b[1];
321 r[1]=(BN_ULONG)ll&BN_MASK2;
322 ll>>=BN_BITS2;
323 if (--n <= 0) break;
324
325 ll+=(BN_ULLONG)a[2]+b[2];
326 r[2]=(BN_ULONG)ll&BN_MASK2;
327 ll>>=BN_BITS2;
328 if (--n <= 0) break;
329
330 ll+=(BN_ULLONG)a[3]+b[3];
331 r[3]=(BN_ULONG)ll&BN_MASK2;
332 ll>>=BN_BITS2;
333 if (--n <= 0) break;
334
335 a+=4;
336 b+=4;
337 r+=4;
338 }
339 return((BN_ULONG)ll);
340 }
341#else
342BN_ULONG bn_add_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
343 {
344 BN_ULONG c,l,t;
345
346 bn_check_num(n);
347 if (n <= 0) return((BN_ULONG)0);
348
349 c=0;
350 for (;;)
351 {
352 t=a[0];
353 t=(t+c)&BN_MASK2;
354 c=(t < c);
355 l=(t+b[0])&BN_MASK2;
356 c+=(l < t);
357 r[0]=l;
358 if (--n <= 0) break;
359
360 t=a[1];
361 t=(t+c)&BN_MASK2;
362 c=(t < c);
363 l=(t+b[1])&BN_MASK2;
364 c+=(l < t);
365 r[1]=l;
366 if (--n <= 0) break;
367
368 t=a[2];
369 t=(t+c)&BN_MASK2;
370 c=(t < c);
371 l=(t+b[2])&BN_MASK2;
372 c+=(l < t);
373 r[2]=l;
374 if (--n <= 0) break;
375
376 t=a[3];
377 t=(t+c)&BN_MASK2;
378 c=(t < c);
379 l=(t+b[3])&BN_MASK2;
380 c+=(l < t);
381 r[3]=l;
382 if (--n <= 0) break;
383
384 a+=4;
385 b+=4;
386 r+=4;
387 }
388 return((BN_ULONG)c);
389 }
390#endif
391
392BN_ULONG bn_sub_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
393 {
394 BN_ULONG t1,t2;
395 int c=0;
396
397 bn_check_num(n);
398 if (n <= 0) return((BN_ULONG)0);
399
400 for (;;)
401 {
402 t1=a[0]; t2=b[0];
403 r[0]=(t1-t2-c)&BN_MASK2;
404 if (t1 != t2) c=(t1 < t2);
405 if (--n <= 0) break;
406
407 t1=a[1]; t2=b[1];
408 r[1]=(t1-t2-c)&BN_MASK2;
409 if (t1 != t2) c=(t1 < t2);
410 if (--n <= 0) break;
411
412 t1=a[2]; t2=b[2];
413 r[2]=(t1-t2-c)&BN_MASK2;
414 if (t1 != t2) c=(t1 < t2);
415 if (--n <= 0) break;
416
417 t1=a[3]; t2=b[3];
418 r[3]=(t1-t2-c)&BN_MASK2;
419 if (t1 != t2) c=(t1 < t2);
420 if (--n <= 0) break;
421
422 a+=4;
423 b+=4;
424 r+=4;
425 }
426 return(c);
427 }
428
429#ifdef BN_MUL_COMBA
430
431#undef bn_mul_comba8
432#undef bn_mul_comba4
433#undef bn_sqr_comba8
434#undef bn_sqr_comba4
435
436#ifdef BN_LLONG
437#define mul_add_c(a,b,c0,c1,c2) \
438 t=(BN_ULLONG)a*b; \
439 t1=(BN_ULONG)Lw(t); \
440 t2=(BN_ULONG)Hw(t); \
441 c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \
442 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
443
444#define mul_add_c2(a,b,c0,c1,c2) \
445 t=(BN_ULLONG)a*b; \
446 tt=(t+t)&BN_MASK; \
447 if (tt < t) c2++; \
448 t1=(BN_ULONG)Lw(tt); \
449 t2=(BN_ULONG)Hw(tt); \
450 c0=(c0+t1)&BN_MASK2; \
451 if ((c0 < t1) && (((++t2)&BN_MASK2) == 0)) c2++; \
452 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
453
454#define sqr_add_c(a,i,c0,c1,c2) \
455 t=(BN_ULLONG)a[i]*a[i]; \
456 t1=(BN_ULONG)Lw(t); \
457 t2=(BN_ULONG)Hw(t); \
458 c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \
459 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
460
461#define sqr_add_c2(a,i,j,c0,c1,c2) \
462 mul_add_c2((a)[i],(a)[j],c0,c1,c2)
463#else
464#define mul_add_c(a,b,c0,c1,c2) \
465 t1=LBITS(a); t2=HBITS(a); \
466 bl=LBITS(b); bh=HBITS(b); \
467 mul64(t1,t2,bl,bh); \
468 c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \
469 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
470
471#define mul_add_c2(a,b,c0,c1,c2) \
472 t1=LBITS(a); t2=HBITS(a); \
473 bl=LBITS(b); bh=HBITS(b); \
474 mul64(t1,t2,bl,bh); \
475 if (t2 & BN_TBIT) c2++; \
476 t2=(t2+t2)&BN_MASK2; \
477 if (t1 & BN_TBIT) t2++; \
478 t1=(t1+t1)&BN_MASK2; \
479 c0=(c0+t1)&BN_MASK2; \
480 if ((c0 < t1) && (((++t2)&BN_MASK2) == 0)) c2++; \
481 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
482
483#define sqr_add_c(a,i,c0,c1,c2) \
484 sqr64(t1,t2,(a)[i]); \
485 c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \
486 c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++;
487
488#define sqr_add_c2(a,i,j,c0,c1,c2) \
489 mul_add_c2((a)[i],(a)[j],c0,c1,c2)
490#endif
491
492void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
493 {
494#ifdef BN_LLONG
495 BN_ULLONG t;
496#else
497 BN_ULONG bl,bh;
498#endif
499 BN_ULONG t1,t2;
500 BN_ULONG c1,c2,c3;
501
502 c1=0;
503 c2=0;
504 c3=0;
505 mul_add_c(a[0],b[0],c1,c2,c3);
506 r[0]=c1;
507 c1=0;
508 mul_add_c(a[0],b[1],c2,c3,c1);
509 mul_add_c(a[1],b[0],c2,c3,c1);
510 r[1]=c2;
511 c2=0;
512 mul_add_c(a[2],b[0],c3,c1,c2);
513 mul_add_c(a[1],b[1],c3,c1,c2);
514 mul_add_c(a[0],b[2],c3,c1,c2);
515 r[2]=c3;
516 c3=0;
517 mul_add_c(a[0],b[3],c1,c2,c3);
518 mul_add_c(a[1],b[2],c1,c2,c3);
519 mul_add_c(a[2],b[1],c1,c2,c3);
520 mul_add_c(a[3],b[0],c1,c2,c3);
521 r[3]=c1;
522 c1=0;
523 mul_add_c(a[4],b[0],c2,c3,c1);
524 mul_add_c(a[3],b[1],c2,c3,c1);
525 mul_add_c(a[2],b[2],c2,c3,c1);
526 mul_add_c(a[1],b[3],c2,c3,c1);
527 mul_add_c(a[0],b[4],c2,c3,c1);
528 r[4]=c2;
529 c2=0;
530 mul_add_c(a[0],b[5],c3,c1,c2);
531 mul_add_c(a[1],b[4],c3,c1,c2);
532 mul_add_c(a[2],b[3],c3,c1,c2);
533 mul_add_c(a[3],b[2],c3,c1,c2);
534 mul_add_c(a[4],b[1],c3,c1,c2);
535 mul_add_c(a[5],b[0],c3,c1,c2);
536 r[5]=c3;
537 c3=0;
538 mul_add_c(a[6],b[0],c1,c2,c3);
539 mul_add_c(a[5],b[1],c1,c2,c3);
540 mul_add_c(a[4],b[2],c1,c2,c3);
541 mul_add_c(a[3],b[3],c1,c2,c3);
542 mul_add_c(a[2],b[4],c1,c2,c3);
543 mul_add_c(a[1],b[5],c1,c2,c3);
544 mul_add_c(a[0],b[6],c1,c2,c3);
545 r[6]=c1;
546 c1=0;
547 mul_add_c(a[0],b[7],c2,c3,c1);
548 mul_add_c(a[1],b[6],c2,c3,c1);
549 mul_add_c(a[2],b[5],c2,c3,c1);
550 mul_add_c(a[3],b[4],c2,c3,c1);
551 mul_add_c(a[4],b[3],c2,c3,c1);
552 mul_add_c(a[5],b[2],c2,c3,c1);
553 mul_add_c(a[6],b[1],c2,c3,c1);
554 mul_add_c(a[7],b[0],c2,c3,c1);
555 r[7]=c2;
556 c2=0;
557 mul_add_c(a[7],b[1],c3,c1,c2);
558 mul_add_c(a[6],b[2],c3,c1,c2);
559 mul_add_c(a[5],b[3],c3,c1,c2);
560 mul_add_c(a[4],b[4],c3,c1,c2);
561 mul_add_c(a[3],b[5],c3,c1,c2);
562 mul_add_c(a[2],b[6],c3,c1,c2);
563 mul_add_c(a[1],b[7],c3,c1,c2);
564 r[8]=c3;
565 c3=0;
566 mul_add_c(a[2],b[7],c1,c2,c3);
567 mul_add_c(a[3],b[6],c1,c2,c3);
568 mul_add_c(a[4],b[5],c1,c2,c3);
569 mul_add_c(a[5],b[4],c1,c2,c3);
570 mul_add_c(a[6],b[3],c1,c2,c3);
571 mul_add_c(a[7],b[2],c1,c2,c3);
572 r[9]=c1;
573 c1=0;
574 mul_add_c(a[7],b[3],c2,c3,c1);
575 mul_add_c(a[6],b[4],c2,c3,c1);
576 mul_add_c(a[5],b[5],c2,c3,c1);
577 mul_add_c(a[4],b[6],c2,c3,c1);
578 mul_add_c(a[3],b[7],c2,c3,c1);
579 r[10]=c2;
580 c2=0;
581 mul_add_c(a[4],b[7],c3,c1,c2);
582 mul_add_c(a[5],b[6],c3,c1,c2);
583 mul_add_c(a[6],b[5],c3,c1,c2);
584 mul_add_c(a[7],b[4],c3,c1,c2);
585 r[11]=c3;
586 c3=0;
587 mul_add_c(a[7],b[5],c1,c2,c3);
588 mul_add_c(a[6],b[6],c1,c2,c3);
589 mul_add_c(a[5],b[7],c1,c2,c3);
590 r[12]=c1;
591 c1=0;
592 mul_add_c(a[6],b[7],c2,c3,c1);
593 mul_add_c(a[7],b[6],c2,c3,c1);
594 r[13]=c2;
595 c2=0;
596 mul_add_c(a[7],b[7],c3,c1,c2);
597 r[14]=c3;
598 r[15]=c1;
599 }
600
601void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
602 {
603#ifdef BN_LLONG
604 BN_ULLONG t;
605#else
606 BN_ULONG bl,bh;
607#endif
608 BN_ULONG t1,t2;
609 BN_ULONG c1,c2,c3;
610
611 c1=0;
612 c2=0;
613 c3=0;
614 mul_add_c(a[0],b[0],c1,c2,c3);
615 r[0]=c1;
616 c1=0;
617 mul_add_c(a[0],b[1],c2,c3,c1);
618 mul_add_c(a[1],b[0],c2,c3,c1);
619 r[1]=c2;
620 c2=0;
621 mul_add_c(a[2],b[0],c3,c1,c2);
622 mul_add_c(a[1],b[1],c3,c1,c2);
623 mul_add_c(a[0],b[2],c3,c1,c2);
624 r[2]=c3;
625 c3=0;
626 mul_add_c(a[0],b[3],c1,c2,c3);
627 mul_add_c(a[1],b[2],c1,c2,c3);
628 mul_add_c(a[2],b[1],c1,c2,c3);
629 mul_add_c(a[3],b[0],c1,c2,c3);
630 r[3]=c1;
631 c1=0;
632 mul_add_c(a[3],b[1],c2,c3,c1);
633 mul_add_c(a[2],b[2],c2,c3,c1);
634 mul_add_c(a[1],b[3],c2,c3,c1);
635 r[4]=c2;
636 c2=0;
637 mul_add_c(a[2],b[3],c3,c1,c2);
638 mul_add_c(a[3],b[2],c3,c1,c2);
639 r[5]=c3;
640 c3=0;
641 mul_add_c(a[3],b[3],c1,c2,c3);
642 r[6]=c1;
643 r[7]=c2;
644 }
645
646void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a)
647 {
648#ifdef BN_LLONG
649 BN_ULLONG t,tt;
650#else
651 BN_ULONG bl,bh;
652#endif
653 BN_ULONG t1,t2;
654 BN_ULONG c1,c2,c3;
655
656 c1=0;
657 c2=0;
658 c3=0;
659 sqr_add_c(a,0,c1,c2,c3);
660 r[0]=c1;
661 c1=0;
662 sqr_add_c2(a,1,0,c2,c3,c1);
663 r[1]=c2;
664 c2=0;
665 sqr_add_c(a,1,c3,c1,c2);
666 sqr_add_c2(a,2,0,c3,c1,c2);
667 r[2]=c3;
668 c3=0;
669 sqr_add_c2(a,3,0,c1,c2,c3);
670 sqr_add_c2(a,2,1,c1,c2,c3);
671 r[3]=c1;
672 c1=0;
673 sqr_add_c(a,2,c2,c3,c1);
674 sqr_add_c2(a,3,1,c2,c3,c1);
675 sqr_add_c2(a,4,0,c2,c3,c1);
676 r[4]=c2;
677 c2=0;
678 sqr_add_c2(a,5,0,c3,c1,c2);
679 sqr_add_c2(a,4,1,c3,c1,c2);
680 sqr_add_c2(a,3,2,c3,c1,c2);
681 r[5]=c3;
682 c3=0;
683 sqr_add_c(a,3,c1,c2,c3);
684 sqr_add_c2(a,4,2,c1,c2,c3);
685 sqr_add_c2(a,5,1,c1,c2,c3);
686 sqr_add_c2(a,6,0,c1,c2,c3);
687 r[6]=c1;
688 c1=0;
689 sqr_add_c2(a,7,0,c2,c3,c1);
690 sqr_add_c2(a,6,1,c2,c3,c1);
691 sqr_add_c2(a,5,2,c2,c3,c1);
692 sqr_add_c2(a,4,3,c2,c3,c1);
693 r[7]=c2;
694 c2=0;
695 sqr_add_c(a,4,c3,c1,c2);
696 sqr_add_c2(a,5,3,c3,c1,c2);
697 sqr_add_c2(a,6,2,c3,c1,c2);
698 sqr_add_c2(a,7,1,c3,c1,c2);
699 r[8]=c3;
700 c3=0;
701 sqr_add_c2(a,7,2,c1,c2,c3);
702 sqr_add_c2(a,6,3,c1,c2,c3);
703 sqr_add_c2(a,5,4,c1,c2,c3);
704 r[9]=c1;
705 c1=0;
706 sqr_add_c(a,5,c2,c3,c1);
707 sqr_add_c2(a,6,4,c2,c3,c1);
708 sqr_add_c2(a,7,3,c2,c3,c1);
709 r[10]=c2;
710 c2=0;
711 sqr_add_c2(a,7,4,c3,c1,c2);
712 sqr_add_c2(a,6,5,c3,c1,c2);
713 r[11]=c3;
714 c3=0;
715 sqr_add_c(a,6,c1,c2,c3);
716 sqr_add_c2(a,7,5,c1,c2,c3);
717 r[12]=c1;
718 c1=0;
719 sqr_add_c2(a,7,6,c2,c3,c1);
720 r[13]=c2;
721 c2=0;
722 sqr_add_c(a,7,c3,c1,c2);
723 r[14]=c3;
724 r[15]=c1;
725 }
726
727void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a)
728 {
729#ifdef BN_LLONG
730 BN_ULLONG t,tt;
731#else
732 BN_ULONG bl,bh;
733#endif
734 BN_ULONG t1,t2;
735 BN_ULONG c1,c2,c3;
736
737 c1=0;
738 c2=0;
739 c3=0;
740 sqr_add_c(a,0,c1,c2,c3);
741 r[0]=c1;
742 c1=0;
743 sqr_add_c2(a,1,0,c2,c3,c1);
744 r[1]=c2;
745 c2=0;
746 sqr_add_c(a,1,c3,c1,c2);
747 sqr_add_c2(a,2,0,c3,c1,c2);
748 r[2]=c3;
749 c3=0;
750 sqr_add_c2(a,3,0,c1,c2,c3);
751 sqr_add_c2(a,2,1,c1,c2,c3);
752 r[3]=c1;
753 c1=0;
754 sqr_add_c(a,2,c2,c3,c1);
755 sqr_add_c2(a,3,1,c2,c3,c1);
756 r[4]=c2;
757 c2=0;
758 sqr_add_c2(a,3,2,c3,c1,c2);
759 r[5]=c3;
760 c3=0;
761 sqr_add_c(a,3,c1,c2,c3);
762 r[6]=c1;
763 r[7]=c2;
764 }
765#else
766
767/* hmm... is it faster just to do a multiply? */
768#undef bn_sqr_comba4
769void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a)
770 {
771 BN_ULONG t[8];
772 bn_sqr_normal(r,a,4,t);
773 }
774
775#undef bn_sqr_comba8
776void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a)
777 {
778 BN_ULONG t[16];
779 bn_sqr_normal(r,a,8,t);
780 }
781
782void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
783 {
784 r[4]=bn_mul_words( &(r[0]),a,4,b[0]);
785 r[5]=bn_mul_add_words(&(r[1]),a,4,b[1]);
786 r[6]=bn_mul_add_words(&(r[2]),a,4,b[2]);
787 r[7]=bn_mul_add_words(&(r[3]),a,4,b[3]);
788 }
789
790void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
791 {
792 r[ 8]=bn_mul_words( &(r[0]),a,8,b[0]);
793 r[ 9]=bn_mul_add_words(&(r[1]),a,8,b[1]);
794 r[10]=bn_mul_add_words(&(r[2]),a,8,b[2]);
795 r[11]=bn_mul_add_words(&(r[3]),a,8,b[3]);
796 r[12]=bn_mul_add_words(&(r[4]),a,8,b[4]);
797 r[13]=bn_mul_add_words(&(r[5]),a,8,b[5]);
798 r[14]=bn_mul_add_words(&(r[6]),a,8,b[6]);
799 r[15]=bn_mul_add_words(&(r[7]),a,8,b[7]);
800 }
801
802#endif /* BN_COMBA */
generated by cgit v1.2.3-55-g6feb (git 2.39.1) at 2025-12-08 05:45:19 +0000