]> git.ipfire.org Git - thirdparty/openssl.git/blame - crypto/ec/curve448/curve448.c
More style fixes to Curve448 code based on review feedback
[thirdparty/openssl.git] / crypto / ec / curve448 / curve448.c
CommitLineData
1308e022 1/*
f53c7764 2 * Copyright 2017-2018 The OpenSSL Project Authors. All Rights Reserved.
1308e022 3 * Copyright 2015-2016 Cryptography Research, Inc.
7324473f 4 *
1308e022
MC
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
7324473f 9 *
1308e022 10 * Originally written by Mike Hamburg
7324473f 11 */
b6e388ba 12#include <openssl/crypto.h>
7324473f
MC
13#include "word.h"
14#include "field.h"
15
a2039c87
MC
16#include "point_448.h"
17#include "ed448.h"
ad0a8a5c 18#include "curve448_lcl.h"
7324473f 19
7324473f
MC
20#define COFACTOR 4
21
22/* Comb config: number of combs, n, t, s. */
23#define COMBS_N 5
24#define COMBS_T 5
25#define COMBS_S 18
aeeef83c
MC
26#define C448_WNAF_FIXED_TABLE_BITS 5
27#define C448_WNAF_VAR_TABLE_BITS 3
7324473f 28
7324473f 29static const int EDWARDS_D = -39081;
8d55f844
MC
30static const curve448_scalar_t precomputed_scalarmul_adjustment = {
31 {
32 {
33 SC_LIMB(0xc873d6d54a7bb0cf), SC_LIMB(0xe933d8d723a70aad),
34 SC_LIMB(0xbb124b65129c96fd), SC_LIMB(0x00000008335dc163)
35 }
36 }
205fd638 37};
7324473f 38
7324473f 39#define TWISTED_D ((EDWARDS_D)-1)
7324473f 40
aeeef83c 41#define WBITS C448_WORD_BITS /* NB this may be different from ARCH_WORD_BITS */
7324473f 42
7324473f 43/* Projective Niels coordinates */
205fd638
MC
44typedef struct {
45 gf a, b, c;
46} niels_s, niels_t[1];
47typedef struct {
48 niels_t n;
49 gf z;
35b7c85a 50} VECTOR_ALIGNED pniels_t[1];
7324473f
MC
51
52/* Precomputed base */
205fd638
MC
53struct curve448_precomputed_s {
54 niels_t table[COMBS_N << (COMBS_T - 1)];
55};
7324473f 56
e7772577 57extern const gf curve448_precomputed_base_as_fe[];
88ba7e71 58const curve448_precomputed_s *curve448_precomputed_base =
205fd638 59 (const curve448_precomputed_s *)&curve448_precomputed_base_as_fe;
7324473f 60
8d55f844 61/* Inverse. */
205fd638
MC
62static void gf_invert(gf y, const gf x, int assert_nonzero)
63{
094c071c
MC
64 mask_t ret;
65
7324473f 66 gf t1, t2;
205fd638
MC
67 gf_sqr(t1, x); /* o^2 */
68 ret = gf_isr(t2, t1); /* +-1/sqrt(o^2) = +-1/o */
7324473f 69 (void)ret;
205fd638
MC
70 if (assert_nonzero)
71 assert(ret);
7324473f 72 gf_sqr(t1, t2);
205fd638 73 gf_mul(t2, t1, x); /* not direct to y in case of alias. */
7324473f
MC
74 gf_copy(y, t2);
75}
76
77/** identity = (0,1) */
205fd638
MC
78const curve448_point_t curve448_point_identity =
79 { {{{{0}}}, {{{1}}}, {{{1}}}, {{{0}}}} };
7324473f 80
8d55f844
MC
81static void point_double_internal(curve448_point_t p, const curve448_point_t q,
82 int before_double)
205fd638 83{
7324473f 84 gf a, b, c, d;
8d55f844 85
205fd638
MC
86 gf_sqr(c, q->x);
87 gf_sqr(a, q->y);
88 gf_add_nr(d, c, a); /* 2+e */
89 gf_add_nr(p->t, q->y, q->x); /* 2+e */
90 gf_sqr(b, p->t);
91 gf_subx_nr(b, b, d, 3); /* 4+e */
92 gf_sub_nr(p->t, a, c); /* 3+e */
93 gf_sqr(p->x, q->z);
94 gf_add_nr(p->z, p->x, p->x); /* 2+e */
95 gf_subx_nr(a, p->z, p->t, 4); /* 6+e */
96 if (GF_HEADROOM == 5)
97 gf_weak_reduce(a); /* or 1+e */
98 gf_mul(p->x, a, b);
99 gf_mul(p->z, p->t, a);
100 gf_mul(p->y, p->t, d);
101 if (!before_double)
102 gf_mul(p->t, b, d);
7324473f
MC
103}
104
205fd638
MC
105void curve448_point_double(curve448_point_t p, const curve448_point_t q)
106{
107 point_double_internal(p, q, 0);
7324473f
MC
108}
109
7324473f 110/* Operations on [p]niels */
205fd638
MC
111static ossl_inline void cond_neg_niels(niels_t n, mask_t neg)
112{
7324473f
MC
113 gf_cond_swap(n->a, n->b, neg);
114 gf_cond_neg(n->c, neg);
115}
116
205fd638
MC
117static void pt_to_pniels(pniels_t b, const curve448_point_t a)
118{
119 gf_sub(b->n->a, a->y, a->x);
120 gf_add(b->n->b, a->x, a->y);
121 gf_mulw(b->n->c, a->t, 2 * TWISTED_D);
122 gf_add(b->z, a->z, a->z);
7324473f
MC
123}
124
205fd638
MC
125static void pniels_to_pt(curve448_point_t e, const pniels_t d)
126{
7324473f 127 gf eu;
8d55f844 128
205fd638
MC
129 gf_add(eu, d->n->b, d->n->a);
130 gf_sub(e->y, d->n->b, d->n->a);
131 gf_mul(e->t, e->y, eu);
132 gf_mul(e->x, d->z, e->y);
133 gf_mul(e->y, d->z, eu);
134 gf_sqr(e->z, d->z);
7324473f
MC
135}
136
205fd638
MC
137static void niels_to_pt(curve448_point_t e, const niels_t n)
138{
139 gf_add(e->y, n->b, n->a);
140 gf_sub(e->x, n->b, n->a);
141 gf_mul(e->t, e->y, e->x);
142 gf_copy(e->z, ONE);
7324473f
MC
143}
144
8d55f844
MC
145static void add_niels_to_pt(curve448_point_t d, const niels_t e,
146 int before_double)
205fd638 147{
7324473f 148 gf a, b, c;
8d55f844 149
205fd638
MC
150 gf_sub_nr(b, d->y, d->x); /* 3+e */
151 gf_mul(a, e->a, b);
152 gf_add_nr(b, d->x, d->y); /* 2+e */
153 gf_mul(d->y, e->b, b);
154 gf_mul(d->x, e->c, d->t);
155 gf_add_nr(c, a, d->y); /* 2+e */
156 gf_sub_nr(b, d->y, a); /* 3+e */
157 gf_sub_nr(d->y, d->z, d->x); /* 3+e */
158 gf_add_nr(a, d->x, d->z); /* 2+e */
159 gf_mul(d->z, a, d->y);
160 gf_mul(d->x, d->y, b);
161 gf_mul(d->y, a, c);
162 if (!before_double)
163 gf_mul(d->t, b, c);
7324473f
MC
164}
165
8d55f844
MC
166static void sub_niels_from_pt(curve448_point_t d, const niels_t e,
167 int before_double)
205fd638 168{
7324473f 169 gf a, b, c;
68b20c00 170
205fd638
MC
171 gf_sub_nr(b, d->y, d->x); /* 3+e */
172 gf_mul(a, e->b, b);
173 gf_add_nr(b, d->x, d->y); /* 2+e */
174 gf_mul(d->y, e->a, b);
175 gf_mul(d->x, e->c, d->t);
176 gf_add_nr(c, a, d->y); /* 2+e */
177 gf_sub_nr(b, d->y, a); /* 3+e */
178 gf_add_nr(d->y, d->z, d->x); /* 2+e */
179 gf_sub_nr(a, d->z, d->x); /* 3+e */
180 gf_mul(d->z, a, d->y);
181 gf_mul(d->x, d->y, b);
182 gf_mul(d->y, a, c);
183 if (!before_double)
184 gf_mul(d->t, b, c);
7324473f
MC
185}
186
8d55f844
MC
187static void add_pniels_to_pt(curve448_point_t p, const pniels_t pn,
188 int before_double)
205fd638 189{
7324473f 190 gf L0;
8d55f844 191
205fd638
MC
192 gf_mul(L0, p->z, pn->z);
193 gf_copy(p->z, L0);
194 add_niels_to_pt(p, pn->n, before_double);
7324473f
MC
195}
196
8d55f844
MC
197static void sub_pniels_from_pt(curve448_point_t p, const pniels_t pn,
198 int before_double)
205fd638 199{
7324473f 200 gf L0;
8d55f844 201
205fd638
MC
202 gf_mul(L0, p->z, pn->z);
203 gf_copy(p->z, L0);
204 sub_niels_from_pt(p, pn->n, before_double);
7324473f
MC
205}
206
aeeef83c
MC
207c448_bool_t curve448_point_eq(const curve448_point_t p,
208 const curve448_point_t q)
205fd638 209{
094c071c 210 mask_t succ;
68b20c00 211 gf a, b;
094c071c 212
7324473f 213 /* equality mod 2-torsion compares x/y */
205fd638
MC
214 gf_mul(a, p->y, q->x);
215 gf_mul(b, q->y, p->x);
216 succ = gf_eq(a, b);
a469abf0 217
7324473f
MC
218 return mask_to_bool(succ);
219}
220
aeeef83c 221c448_bool_t curve448_point_valid(const curve448_point_t p)
205fd638 222{
094c071c 223 mask_t out;
205fd638 224 gf a, b, c;
68b20c00 225
205fd638
MC
226 gf_mul(a, p->x, p->y);
227 gf_mul(b, p->z, p->t);
228 out = gf_eq(a, b);
229 gf_sqr(a, p->x);
230 gf_sqr(b, p->y);
231 gf_sub(a, b, a);
232 gf_sqr(b, p->t);
233 gf_mulw(c, b, TWISTED_D);
234 gf_sqr(b, p->z);
235 gf_add(b, b, c);
236 out &= gf_eq(a, b);
237 out &= ~gf_eq(p->z, ZERO);
7324473f
MC
238 return mask_to_bool(out);
239}
240
52a9587c 241static ossl_inline void constant_time_lookup_niels(niels_s * RESTRICT ni,
8d55f844
MC
242 const niels_t * table,
243 int nelts, int idx)
205fd638 244{
7324473f
MC
245 constant_time_lookup(ni, table, sizeof(niels_s), nelts, idx);
246}
247
205fd638
MC
248void curve448_precomputed_scalarmul(curve448_point_t out,
249 const curve448_precomputed_s * table,
250 const curve448_scalar_t scalar)
251{
68b20c00 252 unsigned int i, j, k;
7324473f 253 const unsigned int n = COMBS_N, t = COMBS_T, s = COMBS_S;
094c071c 254 niels_t ni;
88ba7e71 255 curve448_scalar_t scalar1x;
68b20c00 256
e7772577 257 curve448_scalar_add(scalar1x, scalar, precomputed_scalarmul_adjustment);
205fd638
MC
258 curve448_scalar_halve(scalar1x, scalar1x);
259
68b20c00
MC
260 for (i = s; i > 0; i--) {
261 if (i != s)
205fd638
MC
262 point_double_internal(out, out, 0);
263
264 for (j = 0; j < n; j++) {
7324473f 265 int tab = 0;
094c071c 266 mask_t invert;
205fd638
MC
267
268 for (k = 0; k < t; k++) {
68b20c00
MC
269 unsigned int bit = (i - 1) + s * (k + j * t);
270
db90b274 271 if (bit < C448_SCALAR_BITS) {
205fd638
MC
272 tab |=
273 (scalar1x->limb[bit / WBITS] >> (bit % WBITS) & 1) << k;
7324473f
MC
274 }
275 }
205fd638
MC
276
277 invert = (tab >> (t - 1)) - 1;
7324473f 278 tab ^= invert;
205fd638 279 tab &= (1 << (t - 1)) - 1;
7324473f 280
205fd638
MC
281 constant_time_lookup_niels(ni, &table->table[j << (t - 1)],
282 1 << (t - 1), tab);
7324473f
MC
283
284 cond_neg_niels(ni, invert);
68b20c00
MC
285 if ((i != s) || j != 0) {
286 add_niels_to_pt(out, ni, j == n - 1 && i != 1);
7324473f
MC
287 } else {
288 niels_to_pt(out, ni);
289 }
290 }
291 }
205fd638
MC
292
293 OPENSSL_cleanse(ni, sizeof(ni));
294 OPENSSL_cleanse(scalar1x, sizeof(scalar1x));
7324473f
MC
295}
296
8d55f844 297void curve448_point_mul_by_ratio_and_encode_like_eddsa(
db90b274 298 uint8_t enc[EDDSA_448_PUBLIC_BYTES],
8d55f844 299 const curve448_point_t p)
205fd638 300{
7324473f 301 gf x, y, z, t;
88ba7e71 302 curve448_point_t q;
68b20c00
MC
303
304 /* The point is now on the twisted curve. Move it to untwisted. */
205fd638 305 curve448_point_copy(q, p);
a469abf0 306
7324473f
MC
307 {
308 /* 4-isogeny: 2xy/(y^+x^2), (y^2-x^2)/(2z^2-y^2+x^2) */
309 gf u;
8d55f844 310
205fd638
MC
311 gf_sqr(x, q->x);
312 gf_sqr(t, q->y);
313 gf_add(u, x, t);
314 gf_add(z, q->y, q->x);
315 gf_sqr(y, z);
316 gf_sub(y, y, u);
317 gf_sub(z, t, x);
318 gf_sqr(x, q->z);
319 gf_add(t, x, x);
320 gf_sub(t, t, z);
321 gf_mul(x, t, y);
322 gf_mul(y, z, u);
323 gf_mul(z, u, t);
324 OPENSSL_cleanse(u, sizeof(u));
7324473f 325 }
a469abf0 326
7324473f 327 /* Affinize */
205fd638
MC
328 gf_invert(z, z, 1);
329 gf_mul(t, x, z);
330 gf_mul(x, y, z);
331
7324473f 332 /* Encode */
db90b274 333 enc[EDDSA_448_PRIVATE_BYTES - 1] = 0;
7324473f 334 gf_serialize(enc, x, 1);
db90b274 335 enc[EDDSA_448_PRIVATE_BYTES - 1] |= 0x80 & gf_lobit(t);
7324473f 336
205fd638
MC
337 OPENSSL_cleanse(x, sizeof(x));
338 OPENSSL_cleanse(y, sizeof(y));
339 OPENSSL_cleanse(z, sizeof(z));
340 OPENSSL_cleanse(t, sizeof(t));
e7772577 341 curve448_point_destroy(q);
7324473f
MC
342}
343
aeeef83c 344c448_error_t curve448_point_decode_like_eddsa_and_mul_by_ratio(
8d55f844 345 curve448_point_t p,
db90b274 346 const uint8_t enc[EDDSA_448_PUBLIC_BYTES])
205fd638 347{
db90b274 348 uint8_t enc2[EDDSA_448_PUBLIC_BYTES];
094c071c
MC
349 mask_t low;
350 mask_t succ;
351
205fd638
MC
352 memcpy(enc2, enc, sizeof(enc2));
353
db90b274
MC
354 low = ~word_is_zero(enc2[EDDSA_448_PRIVATE_BYTES - 1] & 0x80);
355 enc2[EDDSA_448_PRIVATE_BYTES - 1] &= ~0x80;
7324473f 356
094c071c 357 succ = gf_deserialize(p->y, enc2, 1, 0);
db90b274 358 succ &= word_is_zero(enc2[EDDSA_448_PRIVATE_BYTES - 1]);
7324473f 359
205fd638
MC
360 gf_sqr(p->x, p->y);
361 gf_sub(p->z, ONE, p->x); /* num = 1-y^2 */
362 gf_mulw(p->t, p->x, EDWARDS_D); /* dy^2 */
363 gf_sub(p->t, ONE, p->t); /* denom = 1-dy^2 or 1-d + dy^2 */
364
365 gf_mul(p->x, p->z, p->t);
366 succ &= gf_isr(p->t, p->x); /* 1/sqrt(num * denom) */
367
368 gf_mul(p->x, p->t, p->z); /* sqrt(num / denom) */
369 gf_cond_neg(p->x, gf_lobit(p->x) ^ low);
370 gf_copy(p->z, ONE);
371
7324473f 372 {
7324473f 373 gf a, b, c, d;
68b20c00
MC
374
375 /* 4-isogeny 2xy/(y^2-ax^2), (y^2+ax^2)/(2-y^2-ax^2) */
205fd638
MC
376 gf_sqr(c, p->x);
377 gf_sqr(a, p->y);
378 gf_add(d, c, a);
379 gf_add(p->t, p->y, p->x);
380 gf_sqr(b, p->t);
381 gf_sub(b, b, d);
382 gf_sub(p->t, a, c);
383 gf_sqr(p->x, p->z);
384 gf_add(p->z, p->x, p->x);
385 gf_sub(a, p->z, d);
386 gf_mul(p->x, a, b);
387 gf_mul(p->z, p->t, a);
388 gf_mul(p->y, p->t, d);
389 gf_mul(p->t, b, d);
390 OPENSSL_cleanse(a, sizeof(a));
391 OPENSSL_cleanse(b, sizeof(b));
392 OPENSSL_cleanse(c, sizeof(c));
393 OPENSSL_cleanse(d, sizeof(d));
7324473f 394 }
205fd638
MC
395
396 OPENSSL_cleanse(enc2, sizeof(enc2));
e7772577 397 assert(curve448_point_valid(p) || ~succ);
205fd638 398
aeeef83c 399 return c448_succeed_if(mask_to_bool(succ));
7324473f
MC
400}
401
db90b274
MC
402c448_error_t x448_int(uint8_t out[X_PUBLIC_BYTES],
403 const uint8_t base[X_PUBLIC_BYTES],
404 const uint8_t scalar[X_PRIVATE_BYTES])
205fd638 405{
7324473f 406 gf x1, x2, z2, x3, z3, t1, t2;
094c071c
MC
407 int t;
408 mask_t swap = 0;
409 mask_t nz;
410
205fd638
MC
411 ignore_result(gf_deserialize(x1, base, 1, 0));
412 gf_copy(x2, ONE);
413 gf_copy(z2, ZERO);
414 gf_copy(x3, x1);
415 gf_copy(z3, ONE);
416
417 for (t = X_PRIVATE_BITS - 1; t >= 0; t--) {
418 uint8_t sb = scalar[t / 8];
094c071c 419 mask_t k_t;
205fd638 420
7324473f 421 /* Scalar conditioning */
205fd638
MC
422 if (t / 8 == 0)
423 sb &= -(uint8_t)COFACTOR;
424 else if (t == X_PRIVATE_BITS - 1)
425 sb = -1;
426
427 k_t = (sb >> (t % 8)) & 1;
52a9587c 428 k_t = 0 - k_t; /* set to all 0s or all 1s */
205fd638 429
7324473f 430 swap ^= k_t;
205fd638
MC
431 gf_cond_swap(x2, x3, swap);
432 gf_cond_swap(z2, z3, swap);
7324473f 433 swap = k_t;
205fd638
MC
434
435 gf_add_nr(t1, x2, z2); /* A = x2 + z2 *//* 2+e */
436 gf_sub_nr(t2, x2, z2); /* B = x2 - z2 *//* 3+e */
437 gf_sub_nr(z2, x3, z3); /* D = x3 - z3 *//* 3+e */
438 gf_mul(x2, t1, z2); /* DA */
439 gf_add_nr(z2, z3, x3); /* C = x3 + z3 *//* 2+e */
440 gf_mul(x3, t2, z2); /* CB */
441 gf_sub_nr(z3, x2, x3); /* DA-CB *//* 3+e */
442 gf_sqr(z2, z3); /* (DA-CB)^2 */
443 gf_mul(z3, x1, z2); /* z3 = x1(DA-CB)^2 */
444 gf_add_nr(z2, x2, x3); /* (DA+CB) *//* 2+e */
445 gf_sqr(x3, z2); /* x3 = (DA+CB)^2 */
446
447 gf_sqr(z2, t1); /* AA = A^2 */
448 gf_sqr(t1, t2); /* BB = B^2 */
449 gf_mul(x2, z2, t1); /* x2 = AA*BB */
450 gf_sub_nr(t2, z2, t1); /* E = AA-BB *//* 3+e */
451
452 gf_mulw(t1, t2, -EDWARDS_D); /* E*-d = a24*E */
453 gf_add_nr(t1, t1, z2); /* AA + a24*E *//* 2+e */
454 gf_mul(z2, t2, t1); /* z2 = E(AA+a24*E) */
7324473f 455 }
205fd638 456
7324473f 457 /* Finish */
205fd638
MC
458 gf_cond_swap(x2, x3, swap);
459 gf_cond_swap(z2, z3, swap);
460 gf_invert(z2, z2, 0);
461 gf_mul(x1, x2, z2);
462 gf_serialize(out, x1, 1);
463 nz = ~gf_eq(x1, ZERO);
464
465 OPENSSL_cleanse(x1, sizeof(x1));
466 OPENSSL_cleanse(x2, sizeof(x2));
467 OPENSSL_cleanse(z2, sizeof(z2));
468 OPENSSL_cleanse(x3, sizeof(x3));
469 OPENSSL_cleanse(z3, sizeof(z3));
470 OPENSSL_cleanse(t1, sizeof(t1));
471 OPENSSL_cleanse(t2, sizeof(t2));
472
aeeef83c 473 return c448_succeed_if(mask_to_bool(nz));
7324473f
MC
474}
475
205fd638
MC
476void curve448_point_mul_by_ratio_and_encode_like_x448(uint8_t
477 out[X_PUBLIC_BYTES],
478 const curve448_point_t p)
479{
88ba7e71 480 curve448_point_t q;
68b20c00 481
205fd638
MC
482 curve448_point_copy(q, p);
483 gf_invert(q->t, q->x, 0); /* 1/x */
484 gf_mul(q->z, q->t, q->y); /* y/x */
485 gf_sqr(q->y, q->z); /* (y/x)^2 */
486 gf_serialize(out, q->y, 1);
e7772577 487 curve448_point_destroy(q);
7324473f
MC
488}
489
db90b274
MC
490void x448_derive_public_key(uint8_t out[X_PUBLIC_BYTES],
491 const uint8_t scalar[X_PRIVATE_BYTES])
205fd638 492{
7324473f
MC
493 /* Scalar conditioning */
494 uint8_t scalar2[X_PRIVATE_BYTES];
094c071c
MC
495 curve448_scalar_t the_scalar;
496 curve448_point_t p;
497 unsigned int i;
498
205fd638 499 memcpy(scalar2, scalar, sizeof(scalar2));
7324473f 500 scalar2[0] &= -(uint8_t)COFACTOR;
205fd638 501
52a9587c 502 scalar2[X_PRIVATE_BYTES - 1] &= ~((0u - 1u) << ((X_PRIVATE_BITS + 7) % 8));
205fd638
MC
503 scalar2[X_PRIVATE_BYTES - 1] |= 1 << ((X_PRIVATE_BITS + 7) % 8);
504
505 curve448_scalar_decode_long(the_scalar, scalar2, sizeof(scalar2));
506
7324473f 507 /* Compensate for the encoding ratio */
db90b274 508 for (i = 1; i < X448_ENCODE_RATIO; i <<= 1) {
205fd638 509 curve448_scalar_halve(the_scalar, the_scalar);
7324473f 510 }
205fd638
MC
511 curve448_precomputed_scalarmul(p, curve448_precomputed_base, the_scalar);
512 curve448_point_mul_by_ratio_and_encode_like_x448(out, p);
e7772577 513 curve448_point_destroy(p);
7324473f
MC
514}
515
8d55f844 516/* Control for variable-time scalar multiply algorithms. */
7324473f 517struct smvt_control {
205fd638 518 int power, addend;
7324473f
MC
519};
520
52a9587c
MC
521#if defined(__GNUC__) || defined(__clang__)
522# define NUMTRAILINGZEROS __builtin_ctz
523#else
524# define NUMTRAILINGZEROS numtrailingzeros
525static uint32_t numtrailingzeros(uint32_t i)
526{
68b20c00 527 uint32_t tmp;
52a9587c
MC
528 uint32_t num = 31;
529
530 if (i == 0)
531 return 32;
532
533 tmp = i << 16;
534 if (tmp != 0) {
535 i = tmp;
536 num -= 16;
537 }
538 tmp = i << 8;
539 if (tmp != 0) {
540 i = tmp;
541 num -= 8;
542 }
543 tmp = i << 4;
544 if (tmp != 0) {
545 i = tmp;
546 num -= 4;
547 }
548 tmp = i << 2;
549 if (tmp != 0) {
550 i = tmp;
551 num -= 2;
552 }
68b20c00
MC
553 tmp = i << 1;
554 if (tmp != 0)
52a9587c
MC
555 num--;
556
557 return num;
558}
559#endif
560
8d55f844
MC
561static int recode_wnaf(struct smvt_control *control,
562 /* [nbits/(table_bits + 1) + 3] */
563 const curve448_scalar_t scalar,
564 unsigned int table_bits)
205fd638 565{
db90b274 566 unsigned int table_size = C448_SCALAR_BITS / (table_bits + 1) + 3;
7324473f 567 int position = table_size - 1; /* at the end */
094c071c 568 uint64_t current = scalar->limb[0] & 0xFFFF;
205fd638 569 uint32_t mask = (1 << (table_bits + 1)) - 1;
094c071c
MC
570 unsigned int w;
571 const unsigned int B_OVER_16 = sizeof(scalar->limb[0]) / 2;
572 unsigned int n, i;
573
7324473f
MC
574 /* place the end marker */
575 control[position].power = -1;
576 control[position].addend = 0;
577 position--;
578
205fd638
MC
579 /*
580 * PERF: Could negate scalar if it's large. But then would need more cases
581 * in the actual code that uses it, all for an expected reduction of like
582 * 1/5 op. Probably not worth it.
7324473f 583 */
7324473f 584
db90b274
MC
585 for (w = 1; w < (C448_SCALAR_BITS - 1) / 16 + 3; w++) {
586 if (w < (C448_SCALAR_BITS - 1) / 16 + 1) {
7324473f 587 /* Refill the 16 high bits of current */
8d55f844
MC
588 current += (uint32_t)((scalar->limb[w / B_OVER_16]
589 >> (16 * (w % B_OVER_16))) << 16);
7324473f 590 }
205fd638 591
7324473f 592 while (current & 0xFFFF) {
52a9587c 593 uint32_t pos = NUMTRAILINGZEROS((uint32_t)current);
8d55f844 594 uint32_t odd = (uint32_t)current >> pos;
7324473f 595 int32_t delta = odd & mask;
094c071c
MC
596
597 assert(position >= 0);
68b20c00 598 if (odd & (1 << (table_bits + 1)))
205fd638 599 delta -= (1 << (table_bits + 1));
7324473f 600 current -= delta << pos;
205fd638 601 control[position].power = pos + 16 * (w - 1);
7324473f
MC
602 control[position].addend = delta;
603 position--;
604 }
605 current >>= 16;
606 }
205fd638
MC
607 assert(current == 0);
608
7324473f 609 position++;
094c071c 610 n = table_size - position;
68b20c00 611 for (i = 0; i < n; i++)
205fd638 612 control[i] = control[i + position];
68b20c00 613
205fd638 614 return n - 1;
7324473f
MC
615}
616
8d55f844
MC
617static void prepare_wnaf_table(pniels_t * output,
618 const curve448_point_t working,
619 unsigned int tbits)
205fd638 620{
88ba7e71 621 curve448_point_t tmp;
7324473f 622 int i;
094c071c
MC
623 pniels_t twop;
624
7324473f
MC
625 pt_to_pniels(output[0], working);
626
205fd638
MC
627 if (tbits == 0)
628 return;
7324473f 629
205fd638 630 curve448_point_double(tmp, working);
7324473f
MC
631 pt_to_pniels(twop, tmp);
632
205fd638 633 add_pniels_to_pt(tmp, output[0], 0);
7324473f
MC
634 pt_to_pniels(output[1], tmp);
635
205fd638
MC
636 for (i = 2; i < 1 << tbits; i++) {
637 add_pniels_to_pt(tmp, twop, 0);
7324473f
MC
638 pt_to_pniels(output[i], tmp);
639 }
205fd638 640
e7772577 641 curve448_point_destroy(tmp);
205fd638 642 OPENSSL_cleanse(twop, sizeof(twop));
7324473f
MC
643}
644
e7772577 645extern const gf curve448_precomputed_wnaf_as_fe[];
205fd638
MC
646static const niels_t *curve448_wnaf_base =
647 (const niels_t *)curve448_precomputed_wnaf_as_fe;
648
649void curve448_base_double_scalarmul_non_secret(curve448_point_t combo,
650 const curve448_scalar_t scalar1,
651 const curve448_point_t base2,
652 const curve448_scalar_t scalar2)
653{
68b20c00
MC
654 const int table_bits_var = C448_WNAF_VAR_TABLE_BITS;
655 const int table_bits_pre = C448_WNAF_FIXED_TABLE_BITS;
db90b274 656 struct smvt_control control_var[C448_SCALAR_BITS /
aeeef83c 657 (C448_WNAF_VAR_TABLE_BITS + 1) + 3];
db90b274 658 struct smvt_control control_pre[C448_SCALAR_BITS /
aeeef83c 659 (C448_WNAF_FIXED_TABLE_BITS + 1) + 3];
7324473f
MC
660 int ncb_pre = recode_wnaf(control_pre, scalar1, table_bits_pre);
661 int ncb_var = recode_wnaf(control_var, scalar2, table_bits_var);
aeeef83c 662 pniels_t precmp_var[1 << C448_WNAF_VAR_TABLE_BITS];
205fd638 663 int contp = 0, contv = 0, i;
094c071c 664
7324473f 665 prepare_wnaf_table(precmp_var, base2, table_bits_var);
094c071c 666 i = control_var[0].power;
7324473f
MC
667
668 if (i < 0) {
e7772577 669 curve448_point_copy(combo, curve448_point_identity);
7324473f
MC
670 return;
671 } else if (i > control_pre[0].power) {
672 pniels_to_pt(combo, precmp_var[control_var[0].addend >> 1]);
673 contv++;
205fd638 674 } else if (i == control_pre[0].power && i >= 0) {
7324473f 675 pniels_to_pt(combo, precmp_var[control_var[0].addend >> 1]);
205fd638
MC
676 add_niels_to_pt(combo, curve448_wnaf_base[control_pre[0].addend >> 1],
677 i);
678 contv++;
679 contp++;
7324473f
MC
680 } else {
681 i = control_pre[0].power;
e7772577 682 niels_to_pt(combo, curve448_wnaf_base[control_pre[0].addend >> 1]);
7324473f
MC
683 contp++;
684 }
205fd638 685
7324473f 686 for (i--; i >= 0; i--) {
68b20c00
MC
687 int cv = (i == control_var[contv].power);
688 int cp = (i == control_pre[contp].power);
689
205fd638 690 point_double_internal(combo, combo, i && !(cv || cp));
7324473f
MC
691
692 if (cv) {
693 assert(control_var[contv].addend);
694
68b20c00 695 if (control_var[contv].addend > 0)
205fd638 696 add_pniels_to_pt(combo,
8d55f844
MC
697 precmp_var[control_var[contv].addend >> 1],
698 i && !cp);
68b20c00 699 else
205fd638 700 sub_pniels_from_pt(combo,
8d55f844
MC
701 precmp_var[(-control_var[contv].addend)
702 >> 1], i && !cp);
7324473f
MC
703 contv++;
704 }
705
706 if (cp) {
707 assert(control_pre[contp].addend);
708
68b20c00 709 if (control_pre[contp].addend > 0)
205fd638 710 add_niels_to_pt(combo,
8d55f844
MC
711 curve448_wnaf_base[control_pre[contp].addend
712 >> 1], i);
68b20c00 713 else
205fd638
MC
714 sub_niels_from_pt(combo,
715 curve448_wnaf_base[(-control_pre
716 [contp].addend) >> 1], i);
7324473f
MC
717 contp++;
718 }
719 }
7324473f 720
205fd638
MC
721 /* This function is non-secret, but whatever this is cheap. */
722 OPENSSL_cleanse(control_var, sizeof(control_var));
723 OPENSSL_cleanse(control_pre, sizeof(control_pre));
724 OPENSSL_cleanse(precmp_var, sizeof(precmp_var));
725
726 assert(contv == ncb_var);
727 (void)ncb_var;
728 assert(contp == ncb_pre);
729 (void)ncb_pre;
7324473f
MC
730}
731
205fd638
MC
732void curve448_point_destroy(curve448_point_t point)
733{
88ba7e71 734 OPENSSL_cleanse(point, sizeof(curve448_point_t));
7324473f
MC
735}
736
ad0a8a5c
MC
737int X448(uint8_t out_shared_key[56], const uint8_t private_key[56],
738 const uint8_t peer_public_value[56])
739{
db90b274
MC
740 return x448_int(out_shared_key, peer_public_value, private_key)
741 == C448_SUCCESS;
ad0a8a5c
MC
742}
743
744void X448_public_from_private(uint8_t out_public_value[56],
745 const uint8_t private_key[56])
746{
db90b274 747 x448_derive_public_key(out_public_value, private_key);
ad0a8a5c 748}