2 * Copyright 2001-2020 The OpenSSL Project Authors. All Rights Reserved.
3 * Copyright (c) 2002, Oracle and/or its affiliates. All rights reserved
5 * Licensed under the OpenSSL license (the "License"). You may not use
6 * this file except in compliance with the License. You can obtain a copy
7 * in the file LICENSE in the source distribution or at
8 * https://www.openssl.org/source/license.html
12 #include <openssl/err.h>
14 #include "internal/cryptlib.h"
15 #include "crypto/bn.h"
17 #include "internal/refcount.h"
20 * This file implements the wNAF-based interleaving multi-exponentiation method
22 * http://www.informatik.tu-darmstadt.de/TI/Mitarbeiter/moeller.html#multiexp
23 * You might now find it here:
24 * http://link.springer.com/chapter/10.1007%2F3-540-45537-X_13
25 * http://www.bmoeller.de/pdf/TI-01-08.multiexp.pdf
26 * For multiplication with precomputation, we use wNAF splitting, formerly at:
27 * http://www.informatik.tu-darmstadt.de/TI/Mitarbeiter/moeller.html#fastexp
30 /* structure for precomputed multiples of the generator */
31 struct ec_pre_comp_st {
32 const EC_GROUP *group; /* parent EC_GROUP object */
33 size_t blocksize; /* block size for wNAF splitting */
34 size_t numblocks; /* max. number of blocks for which we have
36 size_t w; /* window size */
37 EC_POINT **points; /* array with pre-calculated multiples of
38 * generator: 'num' pointers to EC_POINT
39 * objects followed by a NULL */
40 size_t num; /* numblocks * 2^(w-1) */
41 CRYPTO_REF_COUNT references;
45 static EC_PRE_COMP *ec_pre_comp_new(const EC_GROUP *group)
47 EC_PRE_COMP *ret = NULL;
52 ret = OPENSSL_zalloc(sizeof(*ret));
54 ECerr(EC_F_EC_PRE_COMP_NEW, ERR_R_MALLOC_FAILURE);
59 ret->blocksize = 8; /* default */
60 ret->w = 4; /* default */
63 ret->lock = CRYPTO_THREAD_lock_new();
64 if (ret->lock == NULL) {
65 ECerr(EC_F_EC_PRE_COMP_NEW, ERR_R_MALLOC_FAILURE);
72 EC_PRE_COMP *EC_ec_pre_comp_dup(EC_PRE_COMP *pre)
76 CRYPTO_UP_REF(&pre->references, &i, pre->lock);
80 void EC_ec_pre_comp_free(EC_PRE_COMP *pre)
87 CRYPTO_DOWN_REF(&pre->references, &i, pre->lock);
88 REF_PRINT_COUNT("EC_ec", pre);
91 REF_ASSERT_ISNT(i < 0);
93 if (pre->points != NULL) {
96 for (pts = pre->points; *pts != NULL; pts++)
98 OPENSSL_free(pre->points);
100 CRYPTO_THREAD_lock_free(pre->lock);
104 #define EC_POINT_BN_set_flags(P, flags) do { \
105 BN_set_flags((P)->X, (flags)); \
106 BN_set_flags((P)->Y, (flags)); \
107 BN_set_flags((P)->Z, (flags)); \
111 * This functions computes a single point multiplication over the EC group,
112 * using, at a high level, a Montgomery ladder with conditional swaps, with
113 * various timing attack defenses.
115 * It performs either a fixed point multiplication
116 * (scalar * generator)
117 * when point is NULL, or a variable point multiplication
119 * when point is not NULL.
121 * `scalar` cannot be NULL and should be in the range [0,n) otherwise all
122 * constant time bets are off (where n is the cardinality of the EC group).
124 * This function expects `group->order` and `group->cardinality` to be well
125 * defined and non-zero: it fails with an error code otherwise.
127 * NB: This says nothing about the constant-timeness of the ladder step
128 * implementation (i.e., the default implementation is based on EC_POINT_add and
129 * EC_POINT_dbl, which of course are not constant time themselves) or the
130 * underlying multiprecision arithmetic.
132 * The product is stored in `r`.
134 * This is an internal function: callers are in charge of ensuring that the
135 * input parameters `group`, `r`, `scalar` and `ctx` are not NULL.
137 * Returns 1 on success, 0 otherwise.
139 int ec_scalar_mul_ladder(const EC_GROUP *group, EC_POINT *r,
140 const BIGNUM *scalar, const EC_POINT *point,
143 int i, cardinality_bits, group_top, kbit, pbit, Z_is_one;
147 BIGNUM *lambda = NULL;
148 BIGNUM *cardinality = NULL;
151 /* early exit if the input point is the point at infinity */
152 if (point != NULL && EC_POINT_is_at_infinity(group, point))
153 return EC_POINT_set_to_infinity(group, r);
155 if (BN_is_zero(group->order)) {
156 ECerr(EC_F_EC_SCALAR_MUL_LADDER, EC_R_UNKNOWN_ORDER);
159 if (BN_is_zero(group->cofactor)) {
160 ECerr(EC_F_EC_SCALAR_MUL_LADDER, EC_R_UNKNOWN_COFACTOR);
166 if (((p = EC_POINT_new(group)) == NULL)
167 || ((s = EC_POINT_new(group)) == NULL)) {
168 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_MALLOC_FAILURE);
173 if (!EC_POINT_copy(p, group->generator)) {
174 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_EC_LIB);
178 if (!EC_POINT_copy(p, point)) {
179 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_EC_LIB);
184 EC_POINT_BN_set_flags(p, BN_FLG_CONSTTIME);
185 EC_POINT_BN_set_flags(r, BN_FLG_CONSTTIME);
186 EC_POINT_BN_set_flags(s, BN_FLG_CONSTTIME);
188 cardinality = BN_CTX_get(ctx);
189 lambda = BN_CTX_get(ctx);
192 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_MALLOC_FAILURE);
196 if (!BN_mul(cardinality, group->order, group->cofactor, ctx)) {
197 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_BN_LIB);
202 * Group cardinalities are often on a word boundary.
203 * So when we pad the scalar, some timing diff might
204 * pop if it needs to be expanded due to carries.
205 * So expand ahead of time.
207 cardinality_bits = BN_num_bits(cardinality);
208 group_top = bn_get_top(cardinality);
209 if ((bn_wexpand(k, group_top + 2) == NULL)
210 || (bn_wexpand(lambda, group_top + 2) == NULL)) {
211 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_BN_LIB);
215 if (!BN_copy(k, scalar)) {
216 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_BN_LIB);
220 BN_set_flags(k, BN_FLG_CONSTTIME);
222 if ((BN_num_bits(k) > cardinality_bits) || (BN_is_negative(k))) {
224 * this is an unusual input, and we don't guarantee
227 if (!BN_nnmod(k, k, cardinality, ctx)) {
228 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_BN_LIB);
233 if (!BN_add(lambda, k, cardinality)) {
234 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_BN_LIB);
237 BN_set_flags(lambda, BN_FLG_CONSTTIME);
238 if (!BN_add(k, lambda, cardinality)) {
239 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_BN_LIB);
243 * lambda := scalar + cardinality
244 * k := scalar + 2*cardinality
246 kbit = BN_is_bit_set(lambda, cardinality_bits);
247 BN_consttime_swap(kbit, k, lambda, group_top + 2);
249 group_top = bn_get_top(group->field);
250 if ((bn_wexpand(s->X, group_top) == NULL)
251 || (bn_wexpand(s->Y, group_top) == NULL)
252 || (bn_wexpand(s->Z, group_top) == NULL)
253 || (bn_wexpand(r->X, group_top) == NULL)
254 || (bn_wexpand(r->Y, group_top) == NULL)
255 || (bn_wexpand(r->Z, group_top) == NULL)
256 || (bn_wexpand(p->X, group_top) == NULL)
257 || (bn_wexpand(p->Y, group_top) == NULL)
258 || (bn_wexpand(p->Z, group_top) == NULL)) {
259 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_BN_LIB);
263 /* ensure input point is in affine coords for ladder step efficiency */
264 if (!p->Z_is_one && !EC_POINT_make_affine(group, p, ctx)) {
265 ECerr(EC_F_EC_SCALAR_MUL_LADDER, ERR_R_EC_LIB);
269 /* Initialize the Montgomery ladder */
270 if (!ec_point_ladder_pre(group, r, s, p, ctx)) {
271 ECerr(EC_F_EC_SCALAR_MUL_LADDER, EC_R_LADDER_PRE_FAILURE);
275 /* top bit is a 1, in a fixed pos */
278 #define EC_POINT_CSWAP(c, a, b, w, t) do { \
279 BN_consttime_swap(c, (a)->X, (b)->X, w); \
280 BN_consttime_swap(c, (a)->Y, (b)->Y, w); \
281 BN_consttime_swap(c, (a)->Z, (b)->Z, w); \
282 t = ((a)->Z_is_one ^ (b)->Z_is_one) & (c); \
283 (a)->Z_is_one ^= (t); \
284 (b)->Z_is_one ^= (t); \
288 * The ladder step, with branches, is
290 * k[i] == 0: S = add(R, S), R = dbl(R)
291 * k[i] == 1: R = add(S, R), S = dbl(S)
293 * Swapping R, S conditionally on k[i] leaves you with state
295 * k[i] == 0: T, U = R, S
296 * k[i] == 1: T, U = S, R
298 * Then perform the ECC ops.
303 * Which leaves you with state
305 * k[i] == 0: U = add(R, S), T = dbl(R)
306 * k[i] == 1: U = add(S, R), T = dbl(S)
308 * Swapping T, U conditionally on k[i] leaves you with state
310 * k[i] == 0: R, S = T, U
311 * k[i] == 1: R, S = U, T
313 * Which leaves you with state
315 * k[i] == 0: S = add(R, S), R = dbl(R)
316 * k[i] == 1: R = add(S, R), S = dbl(S)
318 * So we get the same logic, but instead of a branch it's a
319 * conditional swap, followed by ECC ops, then another conditional swap.
321 * Optimization: The end of iteration i and start of i-1 looks like
328 * CSWAP(k[i-1], R, S)
330 * CSWAP(k[i-1], R, S)
333 * So instead of two contiguous swaps, you can merge the condition
334 * bits and do a single swap.
336 * k[i] k[i-1] Outcome
342 * This is XOR. pbit tracks the previous bit of k.
345 for (i = cardinality_bits - 1; i >= 0; i--) {
346 kbit = BN_is_bit_set(k, i) ^ pbit;
347 EC_POINT_CSWAP(kbit, r, s, group_top, Z_is_one);
349 /* Perform a single step of the Montgomery ladder */
350 if (!ec_point_ladder_step(group, r, s, p, ctx)) {
351 ECerr(EC_F_EC_SCALAR_MUL_LADDER, EC_R_LADDER_STEP_FAILURE);
355 * pbit logic merges this cswap with that of the
360 /* one final cswap to move the right value into r */
361 EC_POINT_CSWAP(pbit, r, s, group_top, Z_is_one);
362 #undef EC_POINT_CSWAP
364 /* Finalize ladder (and recover full point coordinates) */
365 if (!ec_point_ladder_post(group, r, s, p, ctx)) {
366 ECerr(EC_F_EC_SCALAR_MUL_LADDER, EC_R_LADDER_POST_FAILURE);
374 EC_POINT_clear_free(s);
380 #undef EC_POINT_BN_set_flags
383 * TODO: table should be optimised for the wNAF-based implementation,
384 * sometimes smaller windows will give better performance (thus the
385 * boundaries should be increased)
387 #define EC_window_bits_for_scalar_size(b) \
398 * \sum scalars[i]*points[i],
401 * in the addition if scalar != NULL
403 int ec_wNAF_mul(const EC_GROUP *group, EC_POINT *r, const BIGNUM *scalar,
404 size_t num, const EC_POINT *points[], const BIGNUM *scalars[],
407 const EC_POINT *generator = NULL;
408 EC_POINT *tmp = NULL;
410 size_t blocksize = 0, numblocks = 0; /* for wNAF splitting */
411 size_t pre_points_per_block = 0;
414 int r_is_inverted = 0;
415 int r_is_at_infinity = 1;
416 size_t *wsize = NULL; /* individual window sizes */
417 signed char **wNAF = NULL; /* individual wNAFs */
418 size_t *wNAF_len = NULL;
421 EC_POINT **val = NULL; /* precomputation */
423 EC_POINT ***val_sub = NULL; /* pointers to sub-arrays of 'val' or
424 * 'pre_comp->points' */
425 const EC_PRE_COMP *pre_comp = NULL;
426 int num_scalar = 0; /* flag: will be set to 1 if 'scalar' must be
427 * treated like other scalars, i.e.
428 * precomputation is not available */
431 if (!BN_is_zero(group->order) && !BN_is_zero(group->cofactor)) {
433 * Handle the common cases where the scalar is secret, enforcing a
434 * scalar multiplication implementation based on a Montgomery ladder,
435 * with various timing attack defenses.
437 if ((scalar != group->order) && (scalar != NULL) && (num == 0)) {
439 * In this case we want to compute scalar * GeneratorPoint: this
440 * codepath is reached most prominently by (ephemeral) key
441 * generation of EC cryptosystems (i.e. ECDSA keygen and sign setup,
442 * ECDH keygen/first half), where the scalar is always secret. This
443 * is why we ignore if BN_FLG_CONSTTIME is actually set and we
444 * always call the ladder version.
446 return ec_scalar_mul_ladder(group, r, scalar, NULL, ctx);
448 if ((scalar == NULL) && (num == 1) && (scalars[0] != group->order)) {
450 * In this case we want to compute scalar * VariablePoint: this
451 * codepath is reached most prominently by the second half of ECDH,
452 * where the secret scalar is multiplied by the peer's public point.
453 * To protect the secret scalar, we ignore if BN_FLG_CONSTTIME is
454 * actually set and we always call the ladder version.
456 return ec_scalar_mul_ladder(group, r, scalars[0], points[0], ctx);
460 if (scalar != NULL) {
461 generator = EC_GROUP_get0_generator(group);
462 if (generator == NULL) {
463 ECerr(EC_F_EC_WNAF_MUL, EC_R_UNDEFINED_GENERATOR);
467 /* look if we can use precomputed multiples of generator */
469 pre_comp = group->pre_comp.ec;
470 if (pre_comp && pre_comp->numblocks
471 && (EC_POINT_cmp(group, generator, pre_comp->points[0], ctx) ==
473 blocksize = pre_comp->blocksize;
476 * determine maximum number of blocks that wNAF splitting may
477 * yield (NB: maximum wNAF length is bit length plus one)
479 numblocks = (BN_num_bits(scalar) / blocksize) + 1;
482 * we cannot use more blocks than we have precomputation for
484 if (numblocks > pre_comp->numblocks)
485 numblocks = pre_comp->numblocks;
487 pre_points_per_block = (size_t)1 << (pre_comp->w - 1);
489 /* check that pre_comp looks sane */
490 if (pre_comp->num != (pre_comp->numblocks * pre_points_per_block)) {
491 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
495 /* can't use precomputation */
498 num_scalar = 1; /* treat 'scalar' like 'num'-th element of
503 totalnum = num + numblocks;
505 wsize = OPENSSL_malloc(totalnum * sizeof(wsize[0]));
506 wNAF_len = OPENSSL_malloc(totalnum * sizeof(wNAF_len[0]));
507 /* include space for pivot */
508 wNAF = OPENSSL_malloc((totalnum + 1) * sizeof(wNAF[0]));
509 val_sub = OPENSSL_malloc(totalnum * sizeof(val_sub[0]));
511 /* Ensure wNAF is initialised in case we end up going to err */
513 wNAF[0] = NULL; /* preliminary pivot */
515 if (wsize == NULL || wNAF_len == NULL || wNAF == NULL || val_sub == NULL) {
516 ECerr(EC_F_EC_WNAF_MUL, ERR_R_MALLOC_FAILURE);
521 * num_val will be the total number of temporarily precomputed points
525 for (i = 0; i < num + num_scalar; i++) {
528 bits = i < num ? BN_num_bits(scalars[i]) : BN_num_bits(scalar);
529 wsize[i] = EC_window_bits_for_scalar_size(bits);
530 num_val += (size_t)1 << (wsize[i] - 1);
531 wNAF[i + 1] = NULL; /* make sure we always have a pivot */
533 bn_compute_wNAF((i < num ? scalars[i] : scalar), wsize[i],
537 if (wNAF_len[i] > max_len)
538 max_len = wNAF_len[i];
542 /* we go here iff scalar != NULL */
544 if (pre_comp == NULL) {
545 if (num_scalar != 1) {
546 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
549 /* we have already generated a wNAF for 'scalar' */
551 signed char *tmp_wNAF = NULL;
554 if (num_scalar != 0) {
555 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
560 * use the window size for which we have precomputation
562 wsize[num] = pre_comp->w;
563 tmp_wNAF = bn_compute_wNAF(scalar, wsize[num], &tmp_len);
567 if (tmp_len <= max_len) {
569 * One of the other wNAFs is at least as long as the wNAF
570 * belonging to the generator, so wNAF splitting will not buy
575 totalnum = num + 1; /* don't use wNAF splitting */
576 wNAF[num] = tmp_wNAF;
577 wNAF[num + 1] = NULL;
578 wNAF_len[num] = tmp_len;
580 * pre_comp->points starts with the points that we need here:
582 val_sub[num] = pre_comp->points;
585 * don't include tmp_wNAF directly into wNAF array - use wNAF
586 * splitting and include the blocks
590 EC_POINT **tmp_points;
592 if (tmp_len < numblocks * blocksize) {
594 * possibly we can do with fewer blocks than estimated
596 numblocks = (tmp_len + blocksize - 1) / blocksize;
597 if (numblocks > pre_comp->numblocks) {
598 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
599 OPENSSL_free(tmp_wNAF);
602 totalnum = num + numblocks;
605 /* split wNAF in 'numblocks' parts */
607 tmp_points = pre_comp->points;
609 for (i = num; i < totalnum; i++) {
610 if (i < totalnum - 1) {
611 wNAF_len[i] = blocksize;
612 if (tmp_len < blocksize) {
613 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
614 OPENSSL_free(tmp_wNAF);
617 tmp_len -= blocksize;
620 * last block gets whatever is left (this could be
621 * more or less than 'blocksize'!)
623 wNAF_len[i] = tmp_len;
626 wNAF[i] = OPENSSL_malloc(wNAF_len[i]);
627 if (wNAF[i] == NULL) {
628 ECerr(EC_F_EC_WNAF_MUL, ERR_R_MALLOC_FAILURE);
629 OPENSSL_free(tmp_wNAF);
632 memcpy(wNAF[i], pp, wNAF_len[i]);
633 if (wNAF_len[i] > max_len)
634 max_len = wNAF_len[i];
636 if (*tmp_points == NULL) {
637 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
638 OPENSSL_free(tmp_wNAF);
641 val_sub[i] = tmp_points;
642 tmp_points += pre_points_per_block;
645 OPENSSL_free(tmp_wNAF);
651 * All points we precompute now go into a single array 'val'.
652 * 'val_sub[i]' is a pointer to the subarray for the i-th point, or to a
653 * subarray of 'pre_comp->points' if we already have precomputation.
655 val = OPENSSL_malloc((num_val + 1) * sizeof(val[0]));
657 ECerr(EC_F_EC_WNAF_MUL, ERR_R_MALLOC_FAILURE);
660 val[num_val] = NULL; /* pivot element */
662 /* allocate points for precomputation */
664 for (i = 0; i < num + num_scalar; i++) {
666 for (j = 0; j < ((size_t)1 << (wsize[i] - 1)); j++) {
667 *v = EC_POINT_new(group);
673 if (!(v == val + num_val)) {
674 ECerr(EC_F_EC_WNAF_MUL, ERR_R_INTERNAL_ERROR);
678 if ((tmp = EC_POINT_new(group)) == NULL)
682 * prepare precomputed values:
683 * val_sub[i][0] := points[i]
684 * val_sub[i][1] := 3 * points[i]
685 * val_sub[i][2] := 5 * points[i]
688 for (i = 0; i < num + num_scalar; i++) {
690 if (!EC_POINT_copy(val_sub[i][0], points[i]))
693 if (!EC_POINT_copy(val_sub[i][0], generator))
698 if (!EC_POINT_dbl(group, tmp, val_sub[i][0], ctx))
700 for (j = 1; j < ((size_t)1 << (wsize[i] - 1)); j++) {
702 (group, val_sub[i][j], val_sub[i][j - 1], tmp, ctx))
708 if (!EC_POINTs_make_affine(group, num_val, val, ctx))
711 r_is_at_infinity = 1;
713 for (k = max_len - 1; k >= 0; k--) {
714 if (!r_is_at_infinity) {
715 if (!EC_POINT_dbl(group, r, r, ctx))
719 for (i = 0; i < totalnum; i++) {
720 if (wNAF_len[i] > (size_t)k) {
721 int digit = wNAF[i][k];
730 if (is_neg != r_is_inverted) {
731 if (!r_is_at_infinity) {
732 if (!EC_POINT_invert(group, r, ctx))
735 r_is_inverted = !r_is_inverted;
740 if (r_is_at_infinity) {
741 if (!EC_POINT_copy(r, val_sub[i][digit >> 1]))
745 * Apply coordinate blinding for EC_POINT.
747 * The underlying EC_METHOD can optionally implement this function:
748 * ec_point_blind_coordinates() returns 0 in case of errors or 1 on
749 * success or if coordinate blinding is not implemented for this
752 if (!ec_point_blind_coordinates(group, r, ctx)) {
753 ECerr(EC_F_EC_WNAF_MUL, EC_R_POINT_COORDINATES_BLIND_FAILURE);
757 r_is_at_infinity = 0;
760 (group, r, r, val_sub[i][digit >> 1], ctx))
768 if (r_is_at_infinity) {
769 if (!EC_POINT_set_to_infinity(group, r))
773 if (!EC_POINT_invert(group, r, ctx))
782 OPENSSL_free(wNAF_len);
786 for (w = wNAF; *w != NULL; w++)
792 for (v = val; *v != NULL; v++)
793 EC_POINT_clear_free(*v);
797 OPENSSL_free(val_sub);
802 * ec_wNAF_precompute_mult()
803 * creates an EC_PRE_COMP object with preprecomputed multiples of the generator
804 * for use with wNAF splitting as implemented in ec_wNAF_mul().
806 * 'pre_comp->points' is an array of multiples of the generator
807 * of the following form:
808 * points[0] = generator;
809 * points[1] = 3 * generator;
811 * points[2^(w-1)-1] = (2^(w-1)-1) * generator;
812 * points[2^(w-1)] = 2^blocksize * generator;
813 * points[2^(w-1)+1] = 3 * 2^blocksize * generator;
815 * points[2^(w-1)*(numblocks-1)-1] = (2^(w-1)) * 2^(blocksize*(numblocks-2)) * generator
816 * points[2^(w-1)*(numblocks-1)] = 2^(blocksize*(numblocks-1)) * generator
818 * points[2^(w-1)*numblocks-1] = (2^(w-1)) * 2^(blocksize*(numblocks-1)) * generator
819 * points[2^(w-1)*numblocks] = NULL
821 int ec_wNAF_precompute_mult(EC_GROUP *group, BN_CTX *ctx)
823 const EC_POINT *generator;
824 EC_POINT *tmp_point = NULL, *base = NULL, **var;
825 BN_CTX *new_ctx = NULL;
827 size_t i, bits, w, pre_points_per_block, blocksize, numblocks, num;
828 EC_POINT **points = NULL;
829 EC_PRE_COMP *pre_comp;
832 /* if there is an old EC_PRE_COMP object, throw it away */
833 EC_pre_comp_free(group);
834 if ((pre_comp = ec_pre_comp_new(group)) == NULL)
837 generator = EC_GROUP_get0_generator(group);
838 if (generator == NULL) {
839 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, EC_R_UNDEFINED_GENERATOR);
844 ctx = new_ctx = BN_CTX_new();
851 order = EC_GROUP_get0_order(group);
854 if (BN_is_zero(order)) {
855 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, EC_R_UNKNOWN_ORDER);
859 bits = BN_num_bits(order);
861 * The following parameters mean we precompute (approximately) one point
862 * per bit. TBD: The combination 8, 4 is perfect for 160 bits; for other
863 * bit lengths, other parameter combinations might provide better
868 if (EC_window_bits_for_scalar_size(bits) > w) {
869 /* let's not make the window too small ... */
870 w = EC_window_bits_for_scalar_size(bits);
873 numblocks = (bits + blocksize - 1) / blocksize; /* max. number of blocks
877 pre_points_per_block = (size_t)1 << (w - 1);
878 num = pre_points_per_block * numblocks; /* number of points to compute
881 points = OPENSSL_malloc(sizeof(*points) * (num + 1));
882 if (points == NULL) {
883 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, ERR_R_MALLOC_FAILURE);
888 var[num] = NULL; /* pivot */
889 for (i = 0; i < num; i++) {
890 if ((var[i] = EC_POINT_new(group)) == NULL) {
891 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, ERR_R_MALLOC_FAILURE);
896 if ((tmp_point = EC_POINT_new(group)) == NULL
897 || (base = EC_POINT_new(group)) == NULL) {
898 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, ERR_R_MALLOC_FAILURE);
902 if (!EC_POINT_copy(base, generator))
905 /* do the precomputation */
906 for (i = 0; i < numblocks; i++) {
909 if (!EC_POINT_dbl(group, tmp_point, base, ctx))
912 if (!EC_POINT_copy(*var++, base))
915 for (j = 1; j < pre_points_per_block; j++, var++) {
917 * calculate odd multiples of the current base point
919 if (!EC_POINT_add(group, *var, tmp_point, *(var - 1), ctx))
923 if (i < numblocks - 1) {
925 * get the next base (multiply current one by 2^blocksize)
929 if (blocksize <= 2) {
930 ECerr(EC_F_EC_WNAF_PRECOMPUTE_MULT, ERR_R_INTERNAL_ERROR);
934 if (!EC_POINT_dbl(group, base, tmp_point, ctx))
936 for (k = 2; k < blocksize; k++) {
937 if (!EC_POINT_dbl(group, base, base, ctx))
943 if (!EC_POINTs_make_affine(group, num, points, ctx))
946 pre_comp->group = group;
947 pre_comp->blocksize = blocksize;
948 pre_comp->numblocks = numblocks;
950 pre_comp->points = points;
953 SETPRECOMP(group, ec, pre_comp);
959 BN_CTX_free(new_ctx);
960 EC_ec_pre_comp_free(pre_comp);
964 for (p = points; *p != NULL; p++)
966 OPENSSL_free(points);
968 EC_POINT_free(tmp_point);
973 int ec_wNAF_have_precompute_mult(const EC_GROUP *group)
975 return HAVEPRECOMP(group, ec);