Allow EC_GROUP objects to share precomputation for improved memory
[openssl.git] / crypto / ec / ec_mult.c
1 /* crypto/ec/ec_mult.c */
2 /*
3  * Originally written by Bodo Moeller and Nils Larsch for the OpenSSL project.
4  */
5 /* ====================================================================
6  * Copyright (c) 1998-2003 The OpenSSL Project.  All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer. 
14  *
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in
17  *    the documentation and/or other materials provided with the
18  *    distribution.
19  *
20  * 3. All advertising materials mentioning features or use of this
21  *    software must display the following acknowledgment:
22  *    "This product includes software developed by the OpenSSL Project
23  *    for use in the OpenSSL Toolkit. (http://www.openssl.org/)"
24  *
25  * 4. The names "OpenSSL Toolkit" and "OpenSSL Project" must not be used to
26  *    endorse or promote products derived from this software without
27  *    prior written permission. For written permission, please contact
28  *    openssl-core@openssl.org.
29  *
30  * 5. Products derived from this software may not be called "OpenSSL"
31  *    nor may "OpenSSL" appear in their names without prior written
32  *    permission of the OpenSSL Project.
33  *
34  * 6. Redistributions of any form whatsoever must retain the following
35  *    acknowledgment:
36  *    "This product includes software developed by the OpenSSL Project
37  *    for use in the OpenSSL Toolkit (http://www.openssl.org/)"
38  *
39  * THIS SOFTWARE IS PROVIDED BY THE OpenSSL PROJECT ``AS IS'' AND ANY
40  * EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
41  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
42  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE OpenSSL PROJECT OR
43  * ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
44  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
45  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
46  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
47  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
48  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
49  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
50  * OF THE POSSIBILITY OF SUCH DAMAGE.
51  * ====================================================================
52  *
53  * This product includes cryptographic software written by Eric Young
54  * (eay@cryptsoft.com).  This product includes software written by Tim
55  * Hudson (tjh@cryptsoft.com).
56  *
57  */
58 /* ====================================================================
59  * Copyright 2002 Sun Microsystems, Inc. ALL RIGHTS RESERVED.
60  * Portions of this software developed by SUN MICROSYSTEMS, INC.,
61  * and contributed to the OpenSSL project.
62  */
63
64 #include <string.h>
65
66 #include <openssl/err.h>
67
68 #include "ec_lcl.h"
69
70
71 /*
72  * This file implements the wNAF-based interleaving multi-exponentation method
73  * (<URL:http://www.informatik.tu-darmstadt.de/TI/Mitarbeiter/moeller.html#multiexp>);
74  * for multiplication with precomputation, we use wNAF splitting
75  * (<URL:http://www.informatik.tu-darmstadt.de/TI/Mitarbeiter/moeller.html#fastexp>).
76  */
77
78
79
80
81 /* structure for precomputed multiples of the generator */
82 typedef struct ec_pre_comp_st {
83         const EC_GROUP *group; /* parent EC_GROUP object */
84         size_t blocksize;      /* block size for wNAF splitting */
85         size_t numblocks;      /* max. number of blocks for which we have precomputation */
86         size_t w;              /* window size */
87         EC_POINT **points;     /* array with pre-calculated multiples of generator:
88                                 * 'num' pointers to EC_POINT objects followed by a NULL */
89         size_t num;            /* numblocks * 2^(w-1) */
90         int references;
91 } EC_PRE_COMP;
92  
93 /* functions to manage EC_PRE_COMP within the EC_GROUP extra_data framework */
94 static void *ec_pre_comp_dup(void *);
95 static void ec_pre_comp_free(void *);
96 static void ec_pre_comp_clear_free(void *);
97
98 static EC_PRE_COMP *ec_pre_comp_new(const EC_GROUP *group)
99         {
100         EC_PRE_COMP *ret = NULL;
101
102         if (!group)
103                 return NULL;
104
105         ret = (EC_PRE_COMP *)OPENSSL_malloc(sizeof(EC_PRE_COMP));
106         if (!ret)
107                 return ret;
108         ret->group = group;
109         ret->blocksize = 8; /* default */
110         ret->numblocks = 0;
111         ret->w = 4; /* default */
112         ret->points = NULL;
113         ret->num = 0;
114         ret->references = 1;
115         return ret;
116         }
117
118 static void *ec_pre_comp_dup(void *src_)
119         {
120         EC_PRE_COMP *src = src_;
121
122         /* no need to actually copy, these objects never change! */
123
124         CRYPTO_add(&src->references, 1, CRYPTO_LOCK_EC_PRE_COMP);
125
126         return src_;
127         }
128
129 static void ec_pre_comp_free(void *pre_)
130         {
131         int i;
132         EC_PRE_COMP *pre = pre_;
133
134         if (!pre)
135                 return;
136
137         i = CRYPTO_add(&pre->references, -1, CRYPTO_LOCK_EC_PRE_COMP);
138         if (i > 0)
139                 return;
140
141         if (pre->points)
142                 {
143                 EC_POINT **p;
144
145                 for (p = pre->points; *p != NULL; p++)
146                         EC_POINT_free(*p);
147                 OPENSSL_free(pre->points);
148                 }
149         OPENSSL_free(pre);
150         }
151
152 static void ec_pre_comp_clear_free(void *pre_)
153         {
154         int i;
155         EC_PRE_COMP *pre = pre_;
156
157         if (!pre)
158                 return;
159
160         i = CRYPTO_add(&pre->references, -1, CRYPTO_LOCK_EC_PRE_COMP);
161         if (i > 0)
162                 return;
163
164         if (pre->points)
165                 {
166                 EC_POINT **p;
167
168                 for (p = pre->points; *p != NULL; p++)
169                         EC_POINT_clear_free(*p);
170                 OPENSSL_cleanse(pre->points, sizeof pre->points);
171                 OPENSSL_free(pre->points);
172                 }
173         OPENSSL_cleanse(pre, sizeof pre);
174         OPENSSL_free(pre);
175         }
176
177
178
179
180 /* Determine the modified width-(w+1) Non-Adjacent Form (wNAF) of 'scalar'.
181  * This is an array  r[]  of values that are either zero or odd with an
182  * absolute value less than  2^w  satisfying
183  *     scalar = \sum_j r[j]*2^j
184  * where at most one of any  w+1  consecutive digits is non-zero
185  * with the exception that the most significant digit may be only
186  * w-1 zeros away from that next non-zero digit.
187  */
188 static signed char *compute_wNAF(const BIGNUM *scalar, int w, size_t *ret_len)
189         {
190         int window_val;
191         int ok = 0;
192         signed char *r = NULL;
193         int sign = 1;
194         int bit, next_bit, mask;
195         size_t len = 0, j;
196         
197         if (w <= 0 || w > 7) /* 'signed char' can represent integers with absolute values less than 2^7 */
198                 {
199                 ECerr(EC_F_COMPUTE_WNAF, ERR_R_INTERNAL_ERROR);
200                 goto err;
201                 }
202         bit = 1 << w; /* at most 128 */
203         next_bit = bit << 1; /* at most 256 */
204         mask = next_bit - 1; /* at most 255 */
205
206         if (BN_get_sign(scalar))
207                 {
208                 sign = -1;
209                 }
210
211         len = BN_num_bits(scalar);
212         r = OPENSSL_malloc(len + 1); /* modified wNAF may be one digit longer than binary representation
213                                       * (*ret_len will be set to the actual length, i.e. at most
214                                       * BN_num_bits(scalar) + 1) */
215         if (r == NULL) goto err;
216
217         if (scalar->d == NULL || scalar->top == 0)
218                 {
219                 ECerr(EC_F_COMPUTE_WNAF, ERR_R_INTERNAL_ERROR);
220                 goto err;
221                 }
222         window_val = scalar->d[0] & mask;
223         j = 0;
224         while ((window_val != 0) || (j + w + 1 < len)) /* if j+w+1 >= len, window_val will not increase */
225                 {
226                 int digit = 0;
227
228                 /* 0 <= window_val <= 2^(w+1) */
229
230                 if (window_val & 1)
231                         {
232                         /* 0 < window_val < 2^(w+1) */
233
234                         if (window_val & bit)
235                                 {
236                                 digit = window_val - next_bit; /* -2^w < digit < 0 */
237
238 #if 1 /* modified wNAF */
239                                 if (j + w + 1 >= len)
240                                         {
241                                         /* special case for generating modified wNAFs:
242                                          * no new bits will be added into window_val,
243                                          * so using a positive digit here will decrease
244                                          * the total length of the representation */
245                                         
246                                         digit = window_val & (mask >> 1); /* 0 < digit < 2^w */
247                                         }
248 #endif
249                                 }
250                         else
251                                 {
252                                 digit = window_val; /* 0 < digit < 2^w */
253                                 }
254                         
255                         if (digit <= -bit || digit >= bit || !(digit & 1))
256                                 {
257                                 ECerr(EC_F_COMPUTE_WNAF, ERR_R_INTERNAL_ERROR);
258                                 goto err;
259                                 }
260
261                         window_val -= digit;
262
263                         /* now window_val is 0 or 2^(w+1) in standard wNAF generation;
264                          * for modified window NAFs, it may also be 2^w
265                          */
266                         if (window_val != 0 && window_val != next_bit && window_val != bit)
267                                 {
268                                 ECerr(EC_F_COMPUTE_WNAF, ERR_R_INTERNAL_ERROR);
269                                 goto err;
270                                 }
271                         }
272
273                 r[j++] = sign * digit;
274
275                 window_val >>= 1;
276                 window_val += bit * BN_is_bit_set(scalar, j + w);
277
278                 if (window_val > next_bit)
279                         {
280                         ECerr(EC_F_COMPUTE_WNAF, ERR_R_INTERNAL_ERROR);
281                         goto err;
282                         }
283                 }
284
285         if (j > len + 1)
286                 {
287                 ECerr(EC_F_COMPUTE_WNAF, ERR_R_INTERNAL_ERROR);
288                 goto err;
289                 }
290         len = j;
291         ok = 1;
292
293  err:
294         if (!ok)
295                 {
296                 OPENSSL_free(r);
297                 r = NULL;
298                 }
299         if (ok)
300                 *ret_len = len;
301         return r;
302         }
303
304
305 /* TODO: table should be optimised for the wNAF-based implementation,
306  *       sometimes smaller windows will give better performance
307  *       (thus the boundaries should be increased)
308  */
309 #define EC_window_bits_for_scalar_size(b) \
310                 ((b) >= 2000 ? 6 : \
311                  (b) >=  800 ? 5 : \
312                  (b) >=  300 ? 4 : \
313                  (b) >=   70 ? 3 : \
314                  (b) >=   20 ? 2 : \
315                   1)
316
317 /* Compute
318  *      \sum scalars[i]*points[i],
319  * also including
320  *      scalar*generator
321  * in the addition if scalar != NULL
322  */
323 int ec_wNAF_mul(const EC_GROUP *group, EC_POINT *r, const BIGNUM *scalar,
324         size_t num, const EC_POINT *points[], const BIGNUM *scalars[], BN_CTX *ctx)
325         {
326         BN_CTX *new_ctx = NULL;
327         EC_POINT *generator = NULL;
328         EC_POINT *tmp = NULL;
329         size_t totalnum;
330         size_t blocksize = 0, numblocks = 0; /* for wNAF splitting */
331         size_t pre_points_per_block = 0;
332         size_t i, j;
333         int k;
334         int r_is_inverted = 0;
335         int r_is_at_infinity = 1;
336         size_t *wsize = NULL; /* individual window sizes */
337         signed char **wNAF = NULL; /* individual wNAFs */
338         size_t *wNAF_len = NULL;
339         size_t max_len = 0;
340         size_t num_val;
341         EC_POINT **val = NULL; /* precomputation */
342         EC_POINT **v;
343         EC_POINT ***val_sub = NULL; /* pointers to sub-arrays of 'val' or 'pre_comp->points' */
344         const EC_PRE_COMP *pre_comp = NULL;
345         int num_scalar = 0; /* flag: will be set to 1 if 'scalar' must be treated like other scalars,
346                              * i.e. precomputation is not available */
347         int ret = 0;
348         
349         if (group->meth != r->meth)
350                 {
351                 ECerr(EC_F_EC_WNAF_MUL, EC_R_INCOMPATIBLE_OBJECTS);
352                 return 0;
353                 }
354
355         if ((scalar == NULL) && (num == 0))
356                 {
357                 return EC_POINT_set_to_infinity(group, r);
358                 }
359
360         for (i = 0; i < num; i++)
361                 {
362                 if (group->meth != points[i]->meth)
363                         {
364                         ECerr(EC_F_EC_WNAF_MUL, EC_R_INCOMPATIBLE_OBJECTS);
365                         return 0;
366                         }
367                 }
368
369         if (ctx == NULL)
370                 {
371                 ctx = new_ctx = BN_CTX_new();
372                 if (ctx == NULL)
373                         goto err;
374                 }
375
376         if (scalar != NULL)
377                 {
378                 generator = EC_GROUP_get0_generator(group);
379                 if (generator == NULL)
380                         {
381                         ECerr(EC_F_EC_WNAF_MUL, EC_R_UNDEFINED_GENERATOR);
382                         goto err;
383                         }
384                 
385                 /* look if we can use precomputed multiples of generator */
386
387                 pre_comp = EC_GROUP_get_extra_data(group, ec_pre_comp_dup, ec_pre_comp_free, ec_pre_comp_clear_free);
388
389                 if (pre_comp && pre_comp->numblocks && (EC_POINT_cmp(group, generator, pre_comp->points[0], ctx) == 0))
390                         {
391                         blocksize = pre_comp->blocksize;
392
393                         /* determine maximum number of blocks that wNAF splitting may yield
394                          * (NB: maximum wNAF length is bit length plus one) */
395                         numblocks = (BN_num_bits(scalar) / blocksize) + 1;
396
397                         /* we cannot use more blocks than we have precomputation for */
398                         if (numblocks > pre_comp->numblocks)
399                                 numblocks = pre_comp->numblocks;
400
401                         pre_points_per_block = 1u << (pre_comp->w - 1);
402
403                         /* check that pre_comp looks sane */
404                         if (pre_comp->num != (pre_comp->numblocks * pre_points_per_block))
405                                 {
406                                 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
407                                 goto err;
408                                 }
409                         }
410                 else
411                         {
412                         /* can't use precomputation */
413                         pre_comp = NULL;
414                         numblocks = 1;
415                         num_scalar = 1; /* treat 'scalar' like 'num'-th element of 'scalars' */
416                         }
417                 }
418         
419         totalnum = num + numblocks;
420
421         wsize    = OPENSSL_malloc(totalnum * sizeof wsize[0]);
422         wNAF_len = OPENSSL_malloc(totalnum * sizeof wNAF_len[0]);
423         wNAF     = OPENSSL_malloc((totalnum + 1) * sizeof wNAF[0]); /* includes space for pivot */
424         val_sub  = OPENSSL_malloc(totalnum * sizeof val_sub[0]);
425                  
426         if (!wsize || !wNAF_len || !wNAF || !val_sub)
427                 goto err;
428
429         wNAF[0] = NULL; /* preliminary pivot */
430
431         /* num_val will be the total number of temporarily precomputed points */
432         num_val = 0;
433
434         for (i = 0; i < num + num_scalar; i++)
435                 {
436                 size_t bits;
437
438                 bits = i < num ? BN_num_bits(scalars[i]) : BN_num_bits(scalar);
439                 wsize[i] = EC_window_bits_for_scalar_size(bits);
440                 num_val += 1u << (wsize[i] - 1);
441                 wNAF[i + 1] = NULL; /* make sure we always have a pivot */
442                 wNAF[i] = compute_wNAF((i < num ? scalars[i] : scalar), wsize[i], &wNAF_len[i]);
443                 if (wNAF[i] == NULL)
444                         goto err;
445                 if (wNAF_len[i] > max_len)
446                         max_len = wNAF_len[i];
447                 }
448
449         if (numblocks)
450                 {
451                 /* we go here iff scalar != NULL */
452                 
453                 if (pre_comp == NULL)
454                         {
455                         if (num_scalar != 1)
456                                 {
457                                 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
458                                 goto err;
459                                 }
460                         /* we have already generated a wNAF for 'scalar' */
461                         }
462                 else
463                         {
464                         signed char *tmp_wNAF = NULL;
465                         size_t tmp_len = 0;
466                         
467                         if (num_scalar != 0)
468                                 {
469                                 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
470                                 goto err;
471                                 }
472
473                         /* use the window size for which we have precomputation */
474                         wsize[num] = pre_comp->w;
475                         tmp_wNAF = compute_wNAF(scalar, wsize[num], &tmp_len);
476                         if (!tmp_wNAF)
477                                 goto err;
478
479                         if (tmp_len <= max_len)
480                                 {
481                                 /* One of the other wNAFs is at least as long
482                                  * as the wNAF belonging to the generator,
483                                  * so wNAF splitting will not buy us anything. */
484
485                                 numblocks = 1;
486                                 totalnum = num + 1; /* don't use wNAF splitting */
487                                 wNAF[num] = tmp_wNAF;
488                                 wNAF[num + 1] = NULL;
489                                 wNAF_len[num] = tmp_len;
490                                 if (tmp_len > max_len)
491                                         max_len = tmp_len;
492                                 /* pre_comp->points starts with the points that we need here: */
493                                 val_sub[num] = pre_comp->points;
494                                 }
495                         else
496                                 {
497                                 /* don't include tmp_wNAF directly into wNAF array
498                                  * - use wNAF splitting and include the blocks */
499
500                                 signed char *pp;
501                                 EC_POINT **tmp_points;
502                                 
503                                 if (tmp_len < numblocks * blocksize)
504                                         {
505                                         /* possibly we can do with fewer blocks than estimated */
506                                         numblocks = (tmp_len + blocksize - 1) / blocksize;
507                                         if (numblocks > pre_comp->numblocks)
508                                                 {
509                                                 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
510                                                 goto err;
511                                                 }
512                                         totalnum = num + numblocks;
513                                         }
514                                 
515                                 /* split wNAF in 'numblocks' parts */
516                                 pp = tmp_wNAF;
517                                 tmp_points = pre_comp->points;
518
519                                 for (i = num; i < totalnum; i++)
520                                         {
521                                         if (i < totalnum - 1)
522                                                 {
523                                                 wNAF_len[i] = blocksize;
524                                                 if (tmp_len < blocksize)
525                                                         {
526                                                         ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
527                                                         goto err;
528                                                         }
529                                                 tmp_len -= blocksize;
530                                                 }
531                                         else
532                                                 /* last block gets whatever is left
533                                                  * (this could be more or less than 'blocksize'!) */
534                                                 wNAF_len[i] = tmp_len;
535                                         
536                                         wNAF[i + 1] = NULL;
537                                         wNAF[i] = OPENSSL_malloc(wNAF_len[i]);
538                                         if (wNAF[i] == NULL)
539                                                 {
540                                                 OPENSSL_free(tmp_wNAF);
541                                                 goto err;
542                                                 }
543                                         memcpy(wNAF[i], pp, wNAF_len[i]);
544                                         if (wNAF_len[i] > max_len)
545                                                 max_len = wNAF_len[i];
546
547                                         if (*tmp_points == NULL)
548                                                 {
549                                                 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
550                                                 OPENSSL_free(tmp_wNAF);
551                                                 goto err;
552                                                 }
553                                         val_sub[i] = tmp_points;
554                                         tmp_points += pre_points_per_block;
555                                         pp += blocksize;
556                                         }
557                                 OPENSSL_free(tmp_wNAF);
558                                 }
559                         }
560                 }
561
562         /* All points we precompute now go into a single array 'val'.
563          * 'val_sub[i]' is a pointer to the subarray for the i-th point,
564          * or to a subarray of 'pre_comp->points' if we already have precomputation. */
565         val = OPENSSL_malloc((num_val + 1) * sizeof val[0]);
566         if (val == NULL) goto err;
567         val[num_val] = NULL; /* pivot element */
568
569         /* allocate points for precomputation */
570         v = val;
571         for (i = 0; i < num + num_scalar; i++)
572                 {
573                 val_sub[i] = v;
574                 for (j = 0; j < (1u << (wsize[i] - 1)); j++)
575                         {
576                         *v = EC_POINT_new(group);
577                         if (*v == NULL) goto err;
578                         v++;
579                         }
580                 }
581         if (!(v == val + num_val))
582                 {
583                 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
584                 goto err;
585                 }
586
587         if (!(tmp = EC_POINT_new(group)))
588                 goto err;
589
590         /* prepare precomputed values:
591          *    val_sub[i][0] :=     points[i]
592          *    val_sub[i][1] := 3 * points[i]
593          *    val_sub[i][2] := 5 * points[i]
594          *    ...
595          */
596         for (i = 0; i < num + num_scalar; i++)
597                 {
598                 if (i < num)
599                         {
600                         if (!EC_POINT_copy(val_sub[i][0], points[i])) goto err;
601                         }
602                 else
603                         {
604                         if (!EC_POINT_copy(val_sub[i][0], generator)) goto err;
605                         }
606
607                 if (wsize[i] > 1)
608                         {
609                         if (!EC_POINT_dbl(group, tmp, val_sub[i][0], ctx)) goto err;
610                         for (j = 1; j < (1u << (wsize[i] - 1)); j++)
611                                 {
612                                 if (!EC_POINT_add(group, val_sub[i][j], val_sub[i][j - 1], tmp, ctx)) goto err;
613                                 }
614                         }
615                 }
616
617 #if 1 /* optional; EC_window_bits_for_scalar_size assumes we do this step */
618         if (!EC_POINTs_make_affine(group, num_val, val, ctx))
619                 goto err;
620 #endif
621
622         r_is_at_infinity = 1;
623
624         for (k = max_len - 1; k >= 0; k--)
625                 {
626                 if (!r_is_at_infinity)
627                         {
628                         if (!EC_POINT_dbl(group, r, r, ctx)) goto err;
629                         }
630                 
631                 for (i = 0; i < totalnum; i++)
632                         {
633                         if (wNAF_len[i] > (size_t)k)
634                                 {
635                                 int digit = wNAF[i][k];
636                                 int is_neg;
637
638                                 if (digit) 
639                                         {
640                                         is_neg = digit < 0;
641
642                                         if (is_neg)
643                                                 digit = -digit;
644
645                                         if (is_neg != r_is_inverted)
646                                                 {
647                                                 if (!r_is_at_infinity)
648                                                         {
649                                                         if (!EC_POINT_invert(group, r, ctx)) goto err;
650                                                         }
651                                                 r_is_inverted = !r_is_inverted;
652                                                 }
653
654                                         /* digit > 0 */
655
656                                         if (r_is_at_infinity)
657                                                 {
658                                                 if (!EC_POINT_copy(r, val_sub[i][digit >> 1])) goto err;
659                                                 r_is_at_infinity = 0;
660                                                 }
661                                         else
662                                                 {
663                                                 if (!EC_POINT_add(group, r, r, val_sub[i][digit >> 1], ctx)) goto err;
664                                                 }
665                                         }
666                                 }
667                         }
668                 }
669
670         if (r_is_at_infinity)
671                 {
672                 if (!EC_POINT_set_to_infinity(group, r)) goto err;
673                 }
674         else
675                 {
676                 if (r_is_inverted)
677                         if (!EC_POINT_invert(group, r, ctx)) goto err;
678                 }
679         
680         ret = 1;
681
682  err:
683         if (new_ctx != NULL)
684                 BN_CTX_free(new_ctx);
685         if (tmp != NULL)
686                 EC_POINT_free(tmp);
687         if (wsize != NULL)
688                 OPENSSL_free(wsize);
689         if (wNAF_len != NULL)
690                 OPENSSL_free(wNAF_len);
691         if (wNAF != NULL)
692                 {
693                 signed char **w;
694                 
695                 for (w = wNAF; *w != NULL; w++)
696                         OPENSSL_free(*w);
697                 
698                 OPENSSL_free(wNAF);
699                 }
700         if (val != NULL)
701                 {
702                 for (v = val; *v != NULL; v++)
703                         EC_POINT_clear_free(*v);
704
705                 OPENSSL_free(val);
706                 }
707         if (val_sub != NULL)
708                 {
709                 OPENSSL_free(val_sub);
710                 }
711         return ret;
712         }
713
714
715 /* ec_wNAF_precompute_mult()
716  * creates an EC_PRE_COMP object with preprecomputed multiples of the generator
717  * for use with wNAF splitting as implemented in ec_wNAF_mul().
718  * 
719  * 'pre_comp->points' is an array of multiples of the generator
720  * of the following form:
721  * points[0] =     generator;
722  * points[1] = 3 * generator;
723  * ...
724  * points[2^(w-1)-1] =     (2^(w-1)-1) * generator;
725  * points[2^(w-1)]   =     2^blocksize * generator;
726  * points[2^(w-1)+1] = 3 * 2^blocksize * generator;
727  * ...
728  * points[2^(w-1)*(numblocks-1)-1] = (2^(w-1)) *  2^(blocksize*(numblocks-2)) * generator
729  * points[2^(w-1)*(numblocks-1)]   =              2^(blocksize*(numblocks-1)) * generator
730  * ...
731  * points[2^(w-1)*numblocks-1]     = (2^(w-1)) *  2^(blocksize*(numblocks-1)) * generator
732  * points[2^(w-1)*numblocks]       = NULL
733  */
734 int ec_wNAF_precompute_mult(EC_GROUP *group, BN_CTX *ctx)
735         {
736         const EC_POINT *generator;
737         EC_POINT *tmp_point = NULL, *base = NULL, **var;
738         BN_CTX *new_ctx = NULL;
739         BIGNUM *order;
740         size_t i, bits, w, pre_points_per_block, blocksize, numblocks, num;
741         EC_POINT **points = NULL;
742         EC_PRE_COMP *pre_comp;
743         int ret = 0;
744
745         /* if there is an old EC_PRE_COMP object, throw it away */
746         EC_GROUP_free_extra_data(group, ec_pre_comp_dup, ec_pre_comp_free, ec_pre_comp_clear_free);
747
748         if ((pre_comp = ec_pre_comp_new(group)) == NULL)
749                 return 0;
750
751         generator = EC_GROUP_get0_generator(group);
752         if (generator == NULL)
753                 {
754                 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, EC_R_UNDEFINED_GENERATOR);
755                 goto err;
756                 }
757
758         if (ctx == NULL)
759                 {
760                 ctx = new_ctx = BN_CTX_new();
761                 if (ctx == NULL)
762                         goto err;
763                 }
764         
765         BN_CTX_start(ctx);
766         order = BN_CTX_get(ctx);
767         if (order == NULL) goto err;
768         
769         if (!EC_GROUP_get_order(group, order, ctx)) goto err;           
770         if (BN_is_zero(order))
771                 {
772                 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, EC_R_UNKNOWN_ORDER);
773                 goto err;
774                 }
775
776         bits = BN_num_bits(order);
777         /* The following parameters mean we precompute (approximately)
778          * one point per bit.
779          *
780          * TBD: The combination  8, 4  is perfect for 160 bits; for other
781          * bit lengths, other parameter combinations might provide better
782          * efficiency.
783          */
784         blocksize = 8;
785         w = 4;
786         if (EC_window_bits_for_scalar_size(bits) > w)
787                 {
788                 /* let's not make the window too small ... */
789                 w = EC_window_bits_for_scalar_size(bits);
790                 }
791
792         numblocks = (bits + blocksize - 1) / blocksize; /* max. number of blocks to use for wNAF splitting */
793         
794         pre_points_per_block = 1u << (w - 1);
795         num = pre_points_per_block * numblocks; /* number of points to compute and store */
796
797         points = OPENSSL_malloc(sizeof (EC_POINT*)*(num + 1));
798         if (!points)
799                 {
800                 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, ERR_R_MALLOC_FAILURE);
801                 goto err;
802                 }
803
804         var = points;
805         var[num] = NULL; /* pivot */
806         for (i = 0; i < num; i++)
807                 {
808                 if ((var[i] = EC_POINT_new(group)) == NULL)
809                         {
810                         ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, ERR_R_MALLOC_FAILURE);
811                         goto err;
812                         }
813                 }
814
815         if (!(tmp_point = EC_POINT_new(group)) || !(base = EC_POINT_new(group)))
816                 {
817                 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, ERR_R_MALLOC_FAILURE);
818                 goto err;
819                 }       
820         
821         if (!EC_POINT_copy(base, generator))
822                 goto err;
823         
824         /* do the precomputation */
825         for (i = 0; i < numblocks; i++)
826                 {
827                 size_t j;
828
829                 if (!EC_POINT_dbl(group, tmp_point, base, ctx))
830                         goto err;
831
832                 if (!EC_POINT_copy(*var++, base))
833                         goto err;
834
835                 for (j = 1; j < pre_points_per_block; j++, var++)
836                         {
837                         /* calculate odd multiples of the current base point */
838                         if (!EC_POINT_add(group, *var, tmp_point, *(var - 1), ctx))
839                                 goto err;
840                         }
841
842                 if (i < numblocks - 1)
843                         {
844                         /* get the next base (multiply current one by 2^blocksize) */
845                         size_t k;
846
847                         if (blocksize <= 2)
848                                 {
849                                 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, ERR_R_INTERNAL_ERROR);
850                                 goto err;
851                                 }                               
852
853                         if (!EC_POINT_dbl(group, base, tmp_point, ctx))
854                                 goto err;
855                         for (k = 2; k < blocksize; k++)
856                                 {
857                                 if (!EC_POINT_dbl(group,base,base,ctx))
858                                         goto err;
859                                 }
860                         }
861                 }
862
863         if (!EC_POINTs_make_affine(group, num, points, ctx))
864                 goto err;
865         
866         pre_comp->group = group;
867         pre_comp->blocksize = blocksize;
868         pre_comp->numblocks = numblocks;
869         pre_comp->w = w;
870         pre_comp->points = points;
871         points = NULL;
872         pre_comp->num = num;
873
874         if (!EC_GROUP_set_extra_data(group, pre_comp,
875                 ec_pre_comp_dup, ec_pre_comp_free, ec_pre_comp_clear_free))
876                 goto err;
877         pre_comp = NULL;
878
879         ret = 1;
880  err:
881         BN_CTX_end(ctx);
882         if (new_ctx != NULL)
883                 BN_CTX_free(new_ctx);
884         if (pre_comp)
885                 ec_pre_comp_free(pre_comp);
886         if (points)
887                 {
888                 EC_POINT **p;
889
890                 for (p = points; *p != NULL; p++)
891                         EC_POINT_free(*p);
892                 OPENSSL_free(points);
893                 }
894         if (tmp_point)
895                 EC_POINT_free(tmp_point);
896         if (base)
897                 EC_POINT_free(base);
898         return ret;
899         }
900
901
902 int ec_wNAF_have_precompute_mult(const EC_GROUP *group)
903         {
904         if (EC_GROUP_get_extra_data(group, ec_pre_comp_dup, ec_pre_comp_free, ec_pre_comp_clear_free) != NULL)
905                 return 1;
906         else
907                 return 0;
908         }