| 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 |
#ifndef BN_DEBUG |
| 60 |
# undef NDEBUG /* avoid conflicting definitions */ |
| 61 |
# define NDEBUG |
| 62 |
#endif |
| 63 |
|
| 64 |
#include <stdio.h> |
| 65 |
#include <assert.h> |
| 66 |
#include "bn_lcl.h" |
| 67 |
|
| 68 |
#if defined(BN_LLONG) || defined(BN_UMULT_HIGH) |
| 69 |
|
| 70 |
BN_ULONG bn_mul_add_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w) |
| 71 |
{ |
| 72 |
BN_ULONG c1=0; |
| 73 |
|
| 74 |
assert(num >= 0); |
| 75 |
if (num <= 0) return(c1); |
| 76 |
|
| 77 |
while (num&~3) |
| 78 |
{ |
| 79 |
mul_add(rp[0],ap[0],w,c1); |
| 80 |
mul_add(rp[1],ap[1],w,c1); |
| 81 |
mul_add(rp[2],ap[2],w,c1); |
| 82 |
mul_add(rp[3],ap[3],w,c1); |
| 83 |
ap+=4; rp+=4; num-=4; |
| 84 |
} |
| 85 |
if (num) |
| 86 |
{ |
| 87 |
mul_add(rp[0],ap[0],w,c1); if (--num==0) return c1; |
| 88 |
mul_add(rp[1],ap[1],w,c1); if (--num==0) return c1; |
| 89 |
mul_add(rp[2],ap[2],w,c1); return c1; |
| 90 |
} |
| 91 |
|
| 92 |
return(c1); |
| 93 |
} |
| 94 |
|
| 95 |
BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w) |
| 96 |
{ |
| 97 |
BN_ULONG c1=0; |
| 98 |
|
| 99 |
assert(num >= 0); |
| 100 |
if (num <= 0) return(c1); |
| 101 |
|
| 102 |
while (num&~3) |
| 103 |
{ |
| 104 |
mul(rp[0],ap[0],w,c1); |
| 105 |
mul(rp[1],ap[1],w,c1); |
| 106 |
mul(rp[2],ap[2],w,c1); |
| 107 |
mul(rp[3],ap[3],w,c1); |
| 108 |
ap+=4; rp+=4; num-=4; |
| 109 |
} |
| 110 |
if (num) |
| 111 |
{ |
| 112 |
mul(rp[0],ap[0],w,c1); if (--num == 0) return c1; |
| 113 |
mul(rp[1],ap[1],w,c1); if (--num == 0) return c1; |
| 114 |
mul(rp[2],ap[2],w,c1); |
| 115 |
} |
| 116 |
return(c1); |
| 117 |
} |
| 118 |
|
| 119 |
void bn_sqr_words(BN_ULONG *r, BN_ULONG *a, int n) |
| 120 |
{ |
| 121 |
assert(n >= 0); |
| 122 |
if (n <= 0) return; |
| 123 |
while (n&~3) |
| 124 |
{ |
| 125 |
sqr(r[0],r[1],a[0]); |
| 126 |
sqr(r[2],r[3],a[1]); |
| 127 |
sqr(r[4],r[5],a[2]); |
| 128 |
sqr(r[6],r[7],a[3]); |
| 129 |
a+=4; r+=8; n-=4; |
| 130 |
} |
| 131 |
if (n) |
| 132 |
{ |
| 133 |
sqr(r[0],r[1],a[0]); if (--n == 0) return; |
| 134 |
sqr(r[2],r[3],a[1]); if (--n == 0) return; |
| 135 |
sqr(r[4],r[5],a[2]); |
| 136 |
} |
| 137 |
} |
| 138 |
|
| 139 |
#else /* !(defined(BN_LLONG) || defined(BN_UMULT_HIGH)) */ |
| 140 |
|
| 141 |
BN_ULONG bn_mul_add_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w) |
| 142 |
{ |
| 143 |
BN_ULONG c=0; |
| 144 |
BN_ULONG bl,bh; |
| 145 |
|
| 146 |
assert(num >= 0); |
| 147 |
if (num <= 0) return((BN_ULONG)0); |
| 148 |
|
| 149 |
bl=LBITS(w); |
| 150 |
bh=HBITS(w); |
| 151 |
|
| 152 |
for (;;) |
| 153 |
{ |
| 154 |
mul_add(rp[0],ap[0],bl,bh,c); |
| 155 |
if (--num == 0) break; |
| 156 |
mul_add(rp[1],ap[1],bl,bh,c); |
| 157 |
if (--num == 0) break; |
| 158 |
mul_add(rp[2],ap[2],bl,bh,c); |
| 159 |
if (--num == 0) break; |
| 160 |
mul_add(rp[3],ap[3],bl,bh,c); |
| 161 |
if (--num == 0) break; |
| 162 |
ap+=4; |
| 163 |
rp+=4; |
| 164 |
} |
| 165 |
return(c); |
| 166 |
} |
| 167 |
|
| 168 |
BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w) |
| 169 |
{ |
| 170 |
BN_ULONG carry=0; |
| 171 |
BN_ULONG bl,bh; |
| 172 |
|
| 173 |
assert(num >= 0); |
| 174 |
if (num <= 0) return((BN_ULONG)0); |
| 175 |
|
| 176 |
bl=LBITS(w); |
| 177 |
bh=HBITS(w); |
| 178 |
|
| 179 |
for (;;) |
| 180 |
{ |
| 181 |
mul(rp[0],ap[0],bl,bh,carry); |
| 182 |
if (--num == 0) break; |
| 183 |
mul(rp[1],ap[1],bl,bh,carry); |
| 184 |
if (--num == 0) break; |
| 185 |
mul(rp[2],ap[2],bl,bh,carry); |
| 186 |
if (--num == 0) break; |
| 187 |
mul(rp[3],ap[3],bl,bh,carry); |
| 188 |
if (--num == 0) break; |
| 189 |
ap+=4; |
| 190 |
rp+=4; |
| 191 |
} |
| 192 |
return(carry); |
| 193 |
} |
| 194 |
|
| 195 |
void bn_sqr_words(BN_ULONG *r, BN_ULONG *a, int n) |
| 196 |
{ |
| 197 |
assert(n >= 0); |
| 198 |
if (n <= 0) return; |
| 199 |
for (;;) |
| 200 |
{ |
| 201 |
sqr64(r[0],r[1],a[0]); |
| 202 |
if (--n == 0) break; |
| 203 |
|
| 204 |
sqr64(r[2],r[3],a[1]); |
| 205 |
if (--n == 0) break; |
| 206 |
|
| 207 |
sqr64(r[4],r[5],a[2]); |
| 208 |
if (--n == 0) break; |
| 209 |
|
| 210 |
sqr64(r[6],r[7],a[3]); |
| 211 |
if (--n == 0) break; |
| 212 |
|
| 213 |
a+=4; |
| 214 |
r+=8; |
| 215 |
} |
| 216 |
} |
| 217 |
|
| 218 |
#endif /* !(defined(BN_LLONG) || defined(BN_UMULT_HIGH)) */ |
| 219 |
|
| 220 |
#if defined(BN_LLONG) && defined(BN_DIV2W) |
| 221 |
|
| 222 |
BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d) |
| 223 |
{ |
| 224 |
return((BN_ULONG)(((((BN_ULLONG)h)<<BN_BITS2)|l)/(BN_ULLONG)d)); |
| 225 |
} |
| 226 |
|
| 227 |
#else |
| 228 |
|
| 229 |
/* Divide h,l by d and return the result. */ |
| 230 |
/* I need to test this some more :-( */ |
| 231 |
BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d) |
| 232 |
{ |
| 233 |
BN_ULONG dh,dl,q,ret=0,th,tl,t; |
| 234 |
int i,count=2; |
| 235 |
|
| 236 |
if (d == 0) return(BN_MASK2); |
| 237 |
|
| 238 |
i=BN_num_bits_word(d); |
| 239 |
assert((i == BN_BITS2) || (h > (BN_ULONG)1<<i)); |
| 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 /* !defined(BN_LLONG) && defined(BN_DIV2W) */ |
| 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 |
assert(n >= 0); |
| 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 /* !BN_LLONG */ |
| 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 |
assert(n >= 0); |
| 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 /* !BN_LLONG */ |
| 383 |
|
| 384 |
#if 0 |
| 385 |
BN_ULONG bn_sub_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n) |
| 386 |
{ |
| 387 |
BN_ULONG t1,t2; |
| 388 |
int c=0; |
| 389 |
|
| 390 |
assert(n >= 0); |
| 391 |
if (n <= 0) return((BN_ULONG)0); |
| 392 |
|
| 393 |
for (;;) |
| 394 |
{ |
| 395 |
t1=a[0]; t2=b[0]; |
| 396 |
r[0]=(t1-t2-c)&BN_MASK2; |
| 397 |
if (t1 != t2) c=(t1 < t2); |
| 398 |
if (--n <= 0) break; |
| 399 |
|
| 400 |
t1=a[1]; t2=b[1]; |
| 401 |
r[1]=(t1-t2-c)&BN_MASK2; |
| 402 |
if (t1 != t2) c=(t1 < t2); |
| 403 |
if (--n <= 0) break; |
| 404 |
|
| 405 |
t1=a[2]; t2=b[2]; |
| 406 |
r[2]=(t1-t2-c)&BN_MASK2; |
| 407 |
if (t1 != t2) c=(t1 < t2); |
| 408 |
if (--n <= 0) break; |
| 409 |
|
| 410 |
t1=a[3]; t2=b[3]; |
| 411 |
r[3]=(t1-t2-c)&BN_MASK2; |
| 412 |
if (t1 != t2) c=(t1 < t2); |
| 413 |
if (--n <= 0) break; |
| 414 |
|
| 415 |
a+=4; |
| 416 |
b+=4; |
| 417 |
r+=4; |
| 418 |
} |
| 419 |
return(c); |
| 420 |
} |
| 421 |
#endif |
| 422 |
|
| 423 |
#ifdef BN_MUL_COMBA |
| 424 |
|
| 425 |
#undef bn_mul_comba8 |
| 426 |
#undef bn_mul_comba4 |
| 427 |
#undef bn_sqr_comba8 |
| 428 |
#undef bn_sqr_comba4 |
| 429 |
|
| 430 |
/* mul_add_c(a,b,c0,c1,c2) -- c+=a*b for three word number c=(c2,c1,c0) */ |
| 431 |
/* mul_add_c2(a,b,c0,c1,c2) -- c+=2*a*b for three word number c=(c2,c1,c0) */ |
| 432 |
/* sqr_add_c(a,i,c0,c1,c2) -- c+=a[i]^2 for three word number c=(c2,c1,c0) */ |
| 433 |
/* sqr_add_c2(a,i,c0,c1,c2) -- c+=2*a[i]*a[j] for three word number c=(c2,c1,c0) */ |
| 434 |
|
| 435 |
#ifdef BN_LLONG |
| 436 |
#define mul_add_c(a,b,c0,c1,c2) \ |
| 437 |
t=(BN_ULLONG)a*b; \ |
| 438 |
t1=(BN_ULONG)Lw(t); \ |
| 439 |
t2=(BN_ULONG)Hw(t); \ |
| 440 |
c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \ |
| 441 |
c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++; |
| 442 |
|
| 443 |
#define mul_add_c2(a,b,c0,c1,c2) \ |
| 444 |
t=(BN_ULLONG)a*b; \ |
| 445 |
tt=(t+t)&BN_MASK; \ |
| 446 |
if (tt < t) c2++; \ |
| 447 |
t1=(BN_ULONG)Lw(tt); \ |
| 448 |
t2=(BN_ULONG)Hw(tt); \ |
| 449 |
c0=(c0+t1)&BN_MASK2; \ |
| 450 |
if ((c0 < t1) && (((++t2)&BN_MASK2) == 0)) c2++; \ |
| 451 |
c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++; |
| 452 |
|
| 453 |
#define sqr_add_c(a,i,c0,c1,c2) \ |
| 454 |
t=(BN_ULLONG)a[i]*a[i]; \ |
| 455 |
t1=(BN_ULONG)Lw(t); \ |
| 456 |
t2=(BN_ULONG)Hw(t); \ |
| 457 |
c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \ |
| 458 |
c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++; |
| 459 |
|
| 460 |
#define sqr_add_c2(a,i,j,c0,c1,c2) \ |
| 461 |
mul_add_c2((a)[i],(a)[j],c0,c1,c2) |
| 462 |
|
| 463 |
#elif defined(BN_UMULT_HIGH) |
| 464 |
|
| 465 |
#define mul_add_c(a,b,c0,c1,c2) { \ |
| 466 |
BN_ULONG ta=(a),tb=(b); \ |
| 467 |
t1 = ta * tb; \ |
| 468 |
t2 = BN_UMULT_HIGH(ta,tb); \ |
| 469 |
c0 += t1; t2 += (c0<t1)?1:0; \ |
| 470 |
c1 += t2; c2 += (c1<t2)?1:0; \ |
| 471 |
} |
| 472 |
|
| 473 |
#define mul_add_c2(a,b,c0,c1,c2) { \ |
| 474 |
BN_ULONG ta=(a),tb=(b),t0; \ |
| 475 |
t1 = BN_UMULT_HIGH(ta,tb); \ |
| 476 |
t0 = ta * tb; \ |
| 477 |
t2 = t1+t1; c2 += (t2<t1)?1:0; \ |
| 478 |
t1 = t0+t0; t2 += (t1<t0)?1:0; \ |
| 479 |
c0 += t1; t2 += (c0<t1)?1:0; \ |
| 480 |
c1 += t2; c2 += (c1<t2)?1:0; \ |
| 481 |
} |
| 482 |
|
| 483 |
#define sqr_add_c(a,i,c0,c1,c2) { \ |
| 484 |
BN_ULONG ta=(a)[i]; \ |
| 485 |
t1 = ta * ta; \ |
| 486 |
t2 = BN_UMULT_HIGH(ta,ta); \ |
| 487 |
c0 += t1; t2 += (c0<t1)?1:0; \ |
| 488 |
c1 += t2; c2 += (c1<t2)?1:0; \ |
| 489 |
} |
| 490 |
|
| 491 |
#define sqr_add_c2(a,i,j,c0,c1,c2) \ |
| 492 |
mul_add_c2((a)[i],(a)[j],c0,c1,c2) |
| 493 |
|
| 494 |
#else /* !BN_LLONG */ |
| 495 |
#define mul_add_c(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 |
c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \ |
| 500 |
c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++; |
| 501 |
|
| 502 |
#define mul_add_c2(a,b,c0,c1,c2) \ |
| 503 |
t1=LBITS(a); t2=HBITS(a); \ |
| 504 |
bl=LBITS(b); bh=HBITS(b); \ |
| 505 |
mul64(t1,t2,bl,bh); \ |
| 506 |
if (t2 & BN_TBIT) c2++; \ |
| 507 |
t2=(t2+t2)&BN_MASK2; \ |
| 508 |
if (t1 & BN_TBIT) t2++; \ |
| 509 |
t1=(t1+t1)&BN_MASK2; \ |
| 510 |
c0=(c0+t1)&BN_MASK2; \ |
| 511 |
if ((c0 < t1) && (((++t2)&BN_MASK2) == 0)) c2++; \ |
| 512 |
c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++; |
| 513 |
|
| 514 |
#define sqr_add_c(a,i,c0,c1,c2) \ |
| 515 |
sqr64(t1,t2,(a)[i]); \ |
| 516 |
c0=(c0+t1)&BN_MASK2; if ((c0) < t1) t2++; \ |
| 517 |
c1=(c1+t2)&BN_MASK2; if ((c1) < t2) c2++; |
| 518 |
|
| 519 |
#define sqr_add_c2(a,i,j,c0,c1,c2) \ |
| 520 |
mul_add_c2((a)[i],(a)[j],c0,c1,c2) |
| 521 |
#endif /* !BN_LLONG */ |
| 522 |
|
| 523 |
void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) |
| 524 |
{ |
| 525 |
#ifdef BN_LLONG |
| 526 |
BN_ULLONG t; |
| 527 |
#else |
| 528 |
BN_ULONG bl,bh; |
| 529 |
#endif |
| 530 |
BN_ULONG t1,t2; |
| 531 |
BN_ULONG c1,c2,c3; |
| 532 |
|
| 533 |
c1=0; |
| 534 |
c2=0; |
| 535 |
c3=0; |
| 536 |
mul_add_c(a[0],b[0],c1,c2,c3); |
| 537 |
r[0]=c1; |
| 538 |
c1=0; |
| 539 |
mul_add_c(a[0],b[1],c2,c3,c1); |
| 540 |
mul_add_c(a[1],b[0],c2,c3,c1); |
| 541 |
r[1]=c2; |
| 542 |
c2=0; |
| 543 |
mul_add_c(a[2],b[0],c3,c1,c2); |
| 544 |
mul_add_c(a[1],b[1],c3,c1,c2); |
| 545 |
mul_add_c(a[0],b[2],c3,c1,c2); |
| 546 |
r[2]=c3; |
| 547 |
c3=0; |
| 548 |
mul_add_c(a[0],b[3],c1,c2,c3); |
| 549 |
mul_add_c(a[1],b[2],c1,c2,c3); |
| 550 |
mul_add_c(a[2],b[1],c1,c2,c3); |
| 551 |
mul_add_c(a[3],b[0],c1,c2,c3); |
| 552 |
r[3]=c1; |
| 553 |
c1=0; |
| 554 |
mul_add_c(a[4],b[0],c2,c3,c1); |
| 555 |
mul_add_c(a[3],b[1],c2,c3,c1); |
| 556 |
mul_add_c(a[2],b[2],c2,c3,c1); |
| 557 |
mul_add_c(a[1],b[3],c2,c3,c1); |
| 558 |
mul_add_c(a[0],b[4],c2,c3,c1); |
| 559 |
r[4]=c2; |
| 560 |
c2=0; |
| 561 |
mul_add_c(a[0],b[5],c3,c1,c2); |
| 562 |
mul_add_c(a[1],b[4],c3,c1,c2); |
| 563 |
mul_add_c(a[2],b[3],c3,c1,c2); |
| 564 |
mul_add_c(a[3],b[2],c3,c1,c2); |
| 565 |
mul_add_c(a[4],b[1],c3,c1,c2); |
| 566 |
mul_add_c(a[5],b[0],c3,c1,c2); |
| 567 |
r[5]=c3; |
| 568 |
c3=0; |
| 569 |
mul_add_c(a[6],b[0],c1,c2,c3); |
| 570 |
mul_add_c(a[5],b[1],c1,c2,c3); |
| 571 |
mul_add_c(a[4],b[2],c1,c2,c3); |
| 572 |
mul_add_c(a[3],b[3],c1,c2,c3); |
| 573 |
mul_add_c(a[2],b[4],c1,c2,c3); |
| 574 |
mul_add_c(a[1],b[5],c1,c2,c3); |
| 575 |
mul_add_c(a[0],b[6],c1,c2,c3); |
| 576 |
r[6]=c1; |
| 577 |
c1=0; |
| 578 |
mul_add_c(a[0],b[7],c2,c3,c1); |
| 579 |
mul_add_c(a[1],b[6],c2,c3,c1); |
| 580 |
mul_add_c(a[2],b[5],c2,c3,c1); |
| 581 |
mul_add_c(a[3],b[4],c2,c3,c1); |
| 582 |
mul_add_c(a[4],b[3],c2,c3,c1); |
| 583 |
mul_add_c(a[5],b[2],c2,c3,c1); |
| 584 |
mul_add_c(a[6],b[1],c2,c3,c1); |
| 585 |
mul_add_c(a[7],b[0],c2,c3,c1); |
| 586 |
r[7]=c2; |
| 587 |
c2=0; |
| 588 |
mul_add_c(a[7],b[1],c3,c1,c2); |
| 589 |
mul_add_c(a[6],b[2],c3,c1,c2); |
| 590 |
mul_add_c(a[5],b[3],c3,c1,c2); |
| 591 |
mul_add_c(a[4],b[4],c3,c1,c2); |
| 592 |
mul_add_c(a[3],b[5],c3,c1,c2); |
| 593 |
mul_add_c(a[2],b[6],c3,c1,c2); |
| 594 |
mul_add_c(a[1],b[7],c3,c1,c2); |
| 595 |
r[8]=c3; |
| 596 |
c3=0; |
| 597 |
mul_add_c(a[2],b[7],c1,c2,c3); |
| 598 |
mul_add_c(a[3],b[6],c1,c2,c3); |
| 599 |
mul_add_c(a[4],b[5],c1,c2,c3); |
| 600 |
mul_add_c(a[5],b[4],c1,c2,c3); |
| 601 |
mul_add_c(a[6],b[3],c1,c2,c3); |
| 602 |
mul_add_c(a[7],b[2],c1,c2,c3); |
| 603 |
r[9]=c1; |
| 604 |
c1=0; |
| 605 |
mul_add_c(a[7],b[3],c2,c3,c1); |
| 606 |
mul_add_c(a[6],b[4],c2,c3,c1); |
| 607 |
mul_add_c(a[5],b[5],c2,c3,c1); |
| 608 |
mul_add_c(a[4],b[6],c2,c3,c1); |
| 609 |
mul_add_c(a[3],b[7],c2,c3,c1); |
| 610 |
r[10]=c2; |
| 611 |
c2=0; |
| 612 |
mul_add_c(a[4],b[7],c3,c1,c2); |
| 613 |
mul_add_c(a[5],b[6],c3,c1,c2); |
| 614 |
mul_add_c(a[6],b[5],c3,c1,c2); |
| 615 |
mul_add_c(a[7],b[4],c3,c1,c2); |
| 616 |
r[11]=c3; |
| 617 |
c3=0; |
| 618 |
mul_add_c(a[7],b[5],c1,c2,c3); |
| 619 |
mul_add_c(a[6],b[6],c1,c2,c3); |
| 620 |
mul_add_c(a[5],b[7],c1,c2,c3); |
| 621 |
r[12]=c1; |
| 622 |
c1=0; |
| 623 |
mul_add_c(a[6],b[7],c2,c3,c1); |
| 624 |
mul_add_c(a[7],b[6],c2,c3,c1); |
| 625 |
r[13]=c2; |
| 626 |
c2=0; |
| 627 |
mul_add_c(a[7],b[7],c3,c1,c2); |
| 628 |
r[14]=c3; |
| 629 |
r[15]=c1; |
| 630 |
} |
| 631 |
|
| 632 |
void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) |
| 633 |
{ |
| 634 |
#ifdef BN_LLONG |
| 635 |
BN_ULLONG t; |
| 636 |
#else |
| 637 |
BN_ULONG bl,bh; |
| 638 |
#endif |
| 639 |
BN_ULONG t1,t2; |
| 640 |
BN_ULONG c1,c2,c3; |
| 641 |
|
| 642 |
c1=0; |
| 643 |
c2=0; |
| 644 |
c3=0; |
| 645 |
mul_add_c(a[0],b[0],c1,c2,c3); |
| 646 |
r[0]=c1; |
| 647 |
c1=0; |
| 648 |
mul_add_c(a[0],b[1],c2,c3,c1); |
| 649 |
mul_add_c(a[1],b[0],c2,c3,c1); |
| 650 |
r[1]=c2; |
| 651 |
c2=0; |
| 652 |
mul_add_c(a[2],b[0],c3,c1,c2); |
| 653 |
mul_add_c(a[1],b[1],c3,c1,c2); |
| 654 |
mul_add_c(a[0],b[2],c3,c1,c2); |
| 655 |
r[2]=c3; |
| 656 |
c3=0; |
| 657 |
mul_add_c(a[0],b[3],c1,c2,c3); |
| 658 |
mul_add_c(a[1],b[2],c1,c2,c3); |
| 659 |
mul_add_c(a[2],b[1],c1,c2,c3); |
| 660 |
mul_add_c(a[3],b[0],c1,c2,c3); |
| 661 |
r[3]=c1; |
| 662 |
c1=0; |
| 663 |
mul_add_c(a[3],b[1],c2,c3,c1); |
| 664 |
mul_add_c(a[2],b[2],c2,c3,c1); |
| 665 |
mul_add_c(a[1],b[3],c2,c3,c1); |
| 666 |
r[4]=c2; |
| 667 |
c2=0; |
| 668 |
mul_add_c(a[2],b[3],c3,c1,c2); |
| 669 |
mul_add_c(a[3],b[2],c3,c1,c2); |
| 670 |
r[5]=c3; |
| 671 |
c3=0; |
| 672 |
mul_add_c(a[3],b[3],c1,c2,c3); |
| 673 |
r[6]=c1; |
| 674 |
r[7]=c2; |
| 675 |
} |
| 676 |
|
| 677 |
void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a) |
| 678 |
{ |
| 679 |
#ifdef BN_LLONG |
| 680 |
BN_ULLONG t,tt; |
| 681 |
#else |
| 682 |
BN_ULONG bl,bh; |
| 683 |
#endif |
| 684 |
BN_ULONG t1,t2; |
| 685 |
BN_ULONG c1,c2,c3; |
| 686 |
|
| 687 |
c1=0; |
| 688 |
c2=0; |
| 689 |
c3=0; |
| 690 |
sqr_add_c(a,0,c1,c2,c3); |
| 691 |
r[0]=c1; |
| 692 |
c1=0; |
| 693 |
sqr_add_c2(a,1,0,c2,c3,c1); |
| 694 |
r[1]=c2; |
| 695 |
c2=0; |
| 696 |
sqr_add_c(a,1,c3,c1,c2); |
| 697 |
sqr_add_c2(a,2,0,c3,c1,c2); |
| 698 |
r[2]=c3; |
| 699 |
c3=0; |
| 700 |
sqr_add_c2(a,3,0,c1,c2,c3); |
| 701 |
sqr_add_c2(a,2,1,c1,c2,c3); |
| 702 |
r[3]=c1; |
| 703 |
c1=0; |
| 704 |
sqr_add_c(a,2,c2,c3,c1); |
| 705 |
sqr_add_c2(a,3,1,c2,c3,c1); |
| 706 |
sqr_add_c2(a,4,0,c2,c3,c1); |
| 707 |
r[4]=c2; |
| 708 |
c2=0; |
| 709 |
sqr_add_c2(a,5,0,c3,c1,c2); |
| 710 |
sqr_add_c2(a,4,1,c3,c1,c2); |
| 711 |
sqr_add_c2(a,3,2,c3,c1,c2); |
| 712 |
r[5]=c3; |
| 713 |
c3=0; |
| 714 |
sqr_add_c(a,3,c1,c2,c3); |
| 715 |
sqr_add_c2(a,4,2,c1,c2,c3); |
| 716 |
sqr_add_c2(a,5,1,c1,c2,c3); |
| 717 |
sqr_add_c2(a,6,0,c1,c2,c3); |
| 718 |
r[6]=c1; |
| 719 |
c1=0; |
| 720 |
sqr_add_c2(a,7,0,c2,c3,c1); |
| 721 |
sqr_add_c2(a,6,1,c2,c3,c1); |
| 722 |
sqr_add_c2(a,5,2,c2,c3,c1); |
| 723 |
sqr_add_c2(a,4,3,c2,c3,c1); |
| 724 |
r[7]=c2; |
| 725 |
c2=0; |
| 726 |
sqr_add_c(a,4,c3,c1,c2); |
| 727 |
sqr_add_c2(a,5,3,c3,c1,c2); |
| 728 |
sqr_add_c2(a,6,2,c3,c1,c2); |
| 729 |
sqr_add_c2(a,7,1,c3,c1,c2); |
| 730 |
r[8]=c3; |
| 731 |
c3=0; |
| 732 |
sqr_add_c2(a,7,2,c1,c2,c3); |
| 733 |
sqr_add_c2(a,6,3,c1,c2,c3); |
| 734 |
sqr_add_c2(a,5,4,c1,c2,c3); |
| 735 |
r[9]=c1; |
| 736 |
c1=0; |
| 737 |
sqr_add_c(a,5,c2,c3,c1); |
| 738 |
sqr_add_c2(a,6,4,c2,c3,c1); |
| 739 |
sqr_add_c2(a,7,3,c2,c3,c1); |
| 740 |
r[10]=c2; |
| 741 |
c2=0; |
| 742 |
sqr_add_c2(a,7,4,c3,c1,c2); |
| 743 |
sqr_add_c2(a,6,5,c3,c1,c2); |
| 744 |
r[11]=c3; |
| 745 |
c3=0; |
| 746 |
sqr_add_c(a,6,c1,c2,c3); |
| 747 |
sqr_add_c2(a,7,5,c1,c2,c3); |
| 748 |
r[12]=c1; |
| 749 |
c1=0; |
| 750 |
sqr_add_c2(a,7,6,c2,c3,c1); |
| 751 |
r[13]=c2; |
| 752 |
c2=0; |
| 753 |
sqr_add_c(a,7,c3,c1,c2); |
| 754 |
r[14]=c3; |
| 755 |
r[15]=c1; |
| 756 |
} |
| 757 |
|
| 758 |
void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a) |
| 759 |
{ |
| 760 |
#ifdef BN_LLONG |
| 761 |
BN_ULLONG t,tt; |
| 762 |
#else |
| 763 |
BN_ULONG bl,bh; |
| 764 |
#endif |
| 765 |
BN_ULONG t1,t2; |
| 766 |
BN_ULONG c1,c2,c3; |
| 767 |
|
| 768 |
c1=0; |
| 769 |
c2=0; |
| 770 |
c3=0; |
| 771 |
sqr_add_c(a,0,c1,c2,c3); |
| 772 |
r[0]=c1; |
| 773 |
c1=0; |
| 774 |
sqr_add_c2(a,1,0,c2,c3,c1); |
| 775 |
r[1]=c2; |
| 776 |
c2=0; |
| 777 |
sqr_add_c(a,1,c3,c1,c2); |
| 778 |
sqr_add_c2(a,2,0,c3,c1,c2); |
| 779 |
r[2]=c3; |
| 780 |
c3=0; |
| 781 |
sqr_add_c2(a,3,0,c1,c2,c3); |
| 782 |
sqr_add_c2(a,2,1,c1,c2,c3); |
| 783 |
r[3]=c1; |
| 784 |
c1=0; |
| 785 |
sqr_add_c(a,2,c2,c3,c1); |
| 786 |
sqr_add_c2(a,3,1,c2,c3,c1); |
| 787 |
r[4]=c2; |
| 788 |
c2=0; |
| 789 |
sqr_add_c2(a,3,2,c3,c1,c2); |
| 790 |
r[5]=c3; |
| 791 |
c3=0; |
| 792 |
sqr_add_c(a,3,c1,c2,c3); |
| 793 |
r[6]=c1; |
| 794 |
r[7]=c2; |
| 795 |
} |
| 796 |
#else /* !BN_MUL_COMBA */ |
| 797 |
|
| 798 |
/* hmm... is it faster just to do a multiply? */ |
| 799 |
#undef bn_sqr_comba4 |
| 800 |
void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a) |
| 801 |
{ |
| 802 |
BN_ULONG t[8]; |
| 803 |
bn_sqr_normal(r,a,4,t); |
| 804 |
} |
| 805 |
|
| 806 |
#undef bn_sqr_comba8 |
| 807 |
void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a) |
| 808 |
{ |
| 809 |
BN_ULONG t[16]; |
| 810 |
bn_sqr_normal(r,a,8,t); |
| 811 |
} |
| 812 |
|
| 813 |
void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) |
| 814 |
{ |
| 815 |
r[4]=bn_mul_words( &(r[0]),a,4,b[0]); |
| 816 |
r[5]=bn_mul_add_words(&(r[1]),a,4,b[1]); |
| 817 |
r[6]=bn_mul_add_words(&(r[2]),a,4,b[2]); |
| 818 |
r[7]=bn_mul_add_words(&(r[3]),a,4,b[3]); |
| 819 |
} |
| 820 |
|
| 821 |
void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) |
| 822 |
{ |
| 823 |
r[ 8]=bn_mul_words( &(r[0]),a,8,b[0]); |
| 824 |
r[ 9]=bn_mul_add_words(&(r[1]),a,8,b[1]); |
| 825 |
r[10]=bn_mul_add_words(&(r[2]),a,8,b[2]); |
| 826 |
r[11]=bn_mul_add_words(&(r[3]),a,8,b[3]); |
| 827 |
r[12]=bn_mul_add_words(&(r[4]),a,8,b[4]); |
| 828 |
r[13]=bn_mul_add_words(&(r[5]),a,8,b[5]); |
| 829 |
r[14]=bn_mul_add_words(&(r[6]),a,8,b[6]); |
| 830 |
r[15]=bn_mul_add_words(&(r[7]),a,8,b[7]); |
| 831 |
} |
| 832 |
|
| 833 |
#endif /* !BN_MUL_COMBA */ |
Parent Directory
| 
Revision Log