xref: /freebsd/contrib/bearssl/src/ec/ec_p256_m64.c (revision 6be3386466ab79a84b48429ae66244f21526d3df)
1 /*
2  * Copyright (c) 2018 Thomas Pornin <pornin@bolet.org>
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining
5  * a copy of this software and associated documentation files (the
6  * "Software"), to deal in the Software without restriction, including
7  * without limitation the rights to use, copy, modify, merge, publish,
8  * distribute, sublicense, and/or sell copies of the Software, and to
9  * permit persons to whom the Software is furnished to do so, subject to
10  * the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be
13  * included in all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #include "inner.h"
26 
27 #if BR_INT128 || BR_UMUL128
28 
29 #if BR_UMUL128
30 #include <intrin.h>
31 #endif
32 
33 static const unsigned char P256_G[] = {
34 	0x04, 0x6B, 0x17, 0xD1, 0xF2, 0xE1, 0x2C, 0x42, 0x47, 0xF8,
35 	0xBC, 0xE6, 0xE5, 0x63, 0xA4, 0x40, 0xF2, 0x77, 0x03, 0x7D,
36 	0x81, 0x2D, 0xEB, 0x33, 0xA0, 0xF4, 0xA1, 0x39, 0x45, 0xD8,
37 	0x98, 0xC2, 0x96, 0x4F, 0xE3, 0x42, 0xE2, 0xFE, 0x1A, 0x7F,
38 	0x9B, 0x8E, 0xE7, 0xEB, 0x4A, 0x7C, 0x0F, 0x9E, 0x16, 0x2B,
39 	0xCE, 0x33, 0x57, 0x6B, 0x31, 0x5E, 0xCE, 0xCB, 0xB6, 0x40,
40 	0x68, 0x37, 0xBF, 0x51, 0xF5
41 };
42 
43 static const unsigned char P256_N[] = {
44 	0xFF, 0xFF, 0xFF, 0xFF, 0x00, 0x00, 0x00, 0x00, 0xFF, 0xFF,
45 	0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xBC, 0xE6, 0xFA, 0xAD,
46 	0xA7, 0x17, 0x9E, 0x84, 0xF3, 0xB9, 0xCA, 0xC2, 0xFC, 0x63,
47 	0x25, 0x51
48 };
49 
50 static const unsigned char *
51 api_generator(int curve, size_t *len)
52 {
53 	(void)curve;
54 	*len = sizeof P256_G;
55 	return P256_G;
56 }
57 
58 static const unsigned char *
59 api_order(int curve, size_t *len)
60 {
61 	(void)curve;
62 	*len = sizeof P256_N;
63 	return P256_N;
64 }
65 
66 static size_t
67 api_xoff(int curve, size_t *len)
68 {
69 	(void)curve;
70 	*len = 32;
71 	return 1;
72 }
73 
74 /*
75  * A field element is encoded as four 64-bit integers, in basis 2^64.
76  * Values may reach up to 2^256-1. Montgomery multiplication is used.
77  */
78 
79 /* R = 2^256 mod p */
80 static const uint64_t F256_R[] = {
81 	0x0000000000000001, 0xFFFFFFFF00000000,
82 	0xFFFFFFFFFFFFFFFF, 0x00000000FFFFFFFE
83 };
84 
85 /* Curve equation is y^2 = x^3 - 3*x + B. This constant is B*R mod p
86    (Montgomery representation of B). */
87 static const uint64_t P256_B_MONTY[] = {
88 	0xD89CDF6229C4BDDF, 0xACF005CD78843090,
89 	0xE5A220ABF7212ED6, 0xDC30061D04874834
90 };
91 
92 /*
93  * Addition in the field.
94  */
95 static inline void
96 f256_add(uint64_t *d, const uint64_t *a, const uint64_t *b)
97 {
98 #if BR_INT128
99 	unsigned __int128 w;
100 	uint64_t t;
101 
102 	w = (unsigned __int128)a[0] + b[0];
103 	d[0] = (uint64_t)w;
104 	w = (unsigned __int128)a[1] + b[1] + (w >> 64);
105 	d[1] = (uint64_t)w;
106 	w = (unsigned __int128)a[2] + b[2] + (w >> 64);
107 	d[2] = (uint64_t)w;
108 	w = (unsigned __int128)a[3] + b[3] + (w >> 64);
109 	d[3] = (uint64_t)w;
110 	t = (uint64_t)(w >> 64);
111 
112 	/*
113 	 * 2^256 = 2^224 - 2^192 - 2^96 + 1 in the field.
114 	 */
115 	w = (unsigned __int128)d[0] + t;
116 	d[0] = (uint64_t)w;
117 	w = (unsigned __int128)d[1] + (w >> 64) - (t << 32);
118 	d[1] = (uint64_t)w;
119 	/* Here, carry "w >> 64" can only be 0 or -1 */
120 	w = (unsigned __int128)d[2] - ((w >> 64) & 1);
121 	d[2] = (uint64_t)w;
122 	/* Again, carry is 0 or -1 */
123 	d[3] += (uint64_t)(w >> 64) + (t << 32) - t;
124 
125 #elif BR_UMUL128
126 
127 	unsigned char cc;
128 	uint64_t t;
129 
130 	cc = _addcarry_u64(0, a[0], b[0], &d[0]);
131 	cc = _addcarry_u64(cc, a[1], b[1], &d[1]);
132 	cc = _addcarry_u64(cc, a[2], b[2], &d[2]);
133 	cc = _addcarry_u64(cc, a[3], b[3], &d[3]);
134 
135 	/*
136 	 * If there is a carry, then we want to subtract p, which we
137 	 * do by adding 2^256 - p.
138 	 */
139 	t = cc;
140 	cc = _addcarry_u64(cc, d[0], 0, &d[0]);
141 	cc = _addcarry_u64(cc, d[1], -(t << 32), &d[1]);
142 	cc = _addcarry_u64(cc, d[2], -t, &d[2]);
143 	(void)_addcarry_u64(cc, d[3], (t << 32) - (t << 1), &d[3]);
144 
145 #endif
146 }
147 
148 /*
149  * Subtraction in the field.
150  */
151 static inline void
152 f256_sub(uint64_t *d, const uint64_t *a, const uint64_t *b)
153 {
154 #if BR_INT128
155 
156 	unsigned __int128 w;
157 	uint64_t t;
158 
159 	w = (unsigned __int128)a[0] - b[0];
160 	d[0] = (uint64_t)w;
161 	w = (unsigned __int128)a[1] - b[1] - ((w >> 64) & 1);
162 	d[1] = (uint64_t)w;
163 	w = (unsigned __int128)a[2] - b[2] - ((w >> 64) & 1);
164 	d[2] = (uint64_t)w;
165 	w = (unsigned __int128)a[3] - b[3] - ((w >> 64) & 1);
166 	d[3] = (uint64_t)w;
167 	t = (uint64_t)(w >> 64) & 1;
168 
169 	/*
170 	 * p = 2^256 - 2^224 + 2^192 + 2^96 - 1.
171 	 */
172 	w = (unsigned __int128)d[0] - t;
173 	d[0] = (uint64_t)w;
174 	w = (unsigned __int128)d[1] + (t << 32) - ((w >> 64) & 1);
175 	d[1] = (uint64_t)w;
176 	/* Here, carry "w >> 64" can only be 0 or +1 */
177 	w = (unsigned __int128)d[2] + (w >> 64);
178 	d[2] = (uint64_t)w;
179 	/* Again, carry is 0 or +1 */
180 	d[3] += (uint64_t)(w >> 64) - (t << 32) + t;
181 
182 #elif BR_UMUL128
183 
184 	unsigned char cc;
185 	uint64_t t;
186 
187 	cc = _subborrow_u64(0, a[0], b[0], &d[0]);
188 	cc = _subborrow_u64(cc, a[1], b[1], &d[1]);
189 	cc = _subborrow_u64(cc, a[2], b[2], &d[2]);
190 	cc = _subborrow_u64(cc, a[3], b[3], &d[3]);
191 
192 	/*
193 	 * If there is a carry, then we need to add p.
194 	 */
195 	t = cc;
196 	cc = _addcarry_u64(0, d[0], -t, &d[0]);
197 	cc = _addcarry_u64(cc, d[1], (-t) >> 32, &d[1]);
198 	cc = _addcarry_u64(cc, d[2], 0, &d[2]);
199 	(void)_addcarry_u64(cc, d[3], t - (t << 32), &d[3]);
200 
201 #endif
202 }
203 
204 /*
205  * Montgomery multiplication in the field.
206  */
207 static void
208 f256_montymul(uint64_t *d, const uint64_t *a, const uint64_t *b)
209 {
210 #if BR_INT128
211 
212 	uint64_t x, f, t0, t1, t2, t3, t4;
213 	unsigned __int128 z, ff;
214 	int i;
215 
216 	/*
217 	 * When computing d <- d + a[u]*b, we also add f*p such
218 	 * that d + a[u]*b + f*p is a multiple of 2^64. Since
219 	 * p = -1 mod 2^64, we can compute f = d[0] + a[u]*b[0] mod 2^64.
220 	 */
221 
222 	/*
223 	 * Step 1: t <- (a[0]*b + f*p) / 2^64
224 	 * We have f = a[0]*b[0] mod 2^64. Since p = -1 mod 2^64, this
225 	 * ensures that (a[0]*b + f*p) is a multiple of 2^64.
226 	 *
227 	 * We also have: f*p = f*2^256 - f*2^224 + f*2^192 + f*2^96 - f.
228 	 */
229 	x = a[0];
230 	z = (unsigned __int128)b[0] * x;
231 	f = (uint64_t)z;
232 	z = (unsigned __int128)b[1] * x + (z >> 64) + (uint64_t)(f << 32);
233 	t0 = (uint64_t)z;
234 	z = (unsigned __int128)b[2] * x + (z >> 64) + (uint64_t)(f >> 32);
235 	t1 = (uint64_t)z;
236 	z = (unsigned __int128)b[3] * x + (z >> 64) + f;
237 	t2 = (uint64_t)z;
238 	t3 = (uint64_t)(z >> 64);
239 	ff = ((unsigned __int128)f << 64) - ((unsigned __int128)f << 32);
240 	z = (unsigned __int128)t2 + (uint64_t)ff;
241 	t2 = (uint64_t)z;
242 	z = (unsigned __int128)t3 + (z >> 64) + (ff >> 64);
243 	t3 = (uint64_t)z;
244 	t4 = (uint64_t)(z >> 64);
245 
246 	/*
247 	 * Steps 2 to 4: t <- (t + a[i]*b + f*p) / 2^64
248 	 */
249 	for (i = 1; i < 4; i ++) {
250 		x = a[i];
251 
252 		/* t <- (t + x*b - f) / 2^64 */
253 		z = (unsigned __int128)b[0] * x + t0;
254 		f = (uint64_t)z;
255 		z = (unsigned __int128)b[1] * x + t1 + (z >> 64);
256 		t0 = (uint64_t)z;
257 		z = (unsigned __int128)b[2] * x + t2 + (z >> 64);
258 		t1 = (uint64_t)z;
259 		z = (unsigned __int128)b[3] * x + t3 + (z >> 64);
260 		t2 = (uint64_t)z;
261 		z = t4 + (z >> 64);
262 		t3 = (uint64_t)z;
263 		t4 = (uint64_t)(z >> 64);
264 
265 		/* t <- t + f*2^32, carry in the upper half of z */
266 		z = (unsigned __int128)t0 + (uint64_t)(f << 32);
267 		t0 = (uint64_t)z;
268 		z = (z >> 64) + (unsigned __int128)t1 + (uint64_t)(f >> 32);
269 		t1 = (uint64_t)z;
270 
271 		/* t <- t + f*2^192 - f*2^160 + f*2^128 */
272 		ff = ((unsigned __int128)f << 64)
273 			- ((unsigned __int128)f << 32) + f;
274 		z = (z >> 64) + (unsigned __int128)t2 + (uint64_t)ff;
275 		t2 = (uint64_t)z;
276 		z = (unsigned __int128)t3 + (z >> 64) + (ff >> 64);
277 		t3 = (uint64_t)z;
278 		t4 += (uint64_t)(z >> 64);
279 	}
280 
281 	/*
282 	 * At that point, we have computed t = (a*b + F*p) / 2^256, where
283 	 * F is a 256-bit integer whose limbs are the "f" coefficients
284 	 * in the steps above. We have:
285 	 *   a <= 2^256-1
286 	 *   b <= 2^256-1
287 	 *   F <= 2^256-1
288 	 * Hence:
289 	 *   a*b + F*p <= (2^256-1)*(2^256-1) + p*(2^256-1)
290 	 *   a*b + F*p <= 2^256*(2^256 - 2 + p) + 1 - p
291 	 * Therefore:
292 	 *   t < 2^256 + p - 2
293 	 * Since p < 2^256, it follows that:
294 	 *   t4 can be only 0 or 1
295 	 *   t - p < 2^256
296 	 * We can therefore subtract p from t, conditionally on t4, to
297 	 * get a nonnegative result that fits on 256 bits.
298 	 */
299 	z = (unsigned __int128)t0 + t4;
300 	t0 = (uint64_t)z;
301 	z = (unsigned __int128)t1 - (t4 << 32) + (z >> 64);
302 	t1 = (uint64_t)z;
303 	z = (unsigned __int128)t2 - (z >> 127);
304 	t2 = (uint64_t)z;
305 	t3 = t3 - (uint64_t)(z >> 127) - t4 + (t4 << 32);
306 
307 	d[0] = t0;
308 	d[1] = t1;
309 	d[2] = t2;
310 	d[3] = t3;
311 
312 #elif BR_UMUL128
313 
314 	uint64_t x, f, t0, t1, t2, t3, t4;
315 	uint64_t zl, zh, ffl, ffh;
316 	unsigned char k, m;
317 	int i;
318 
319 	/*
320 	 * When computing d <- d + a[u]*b, we also add f*p such
321 	 * that d + a[u]*b + f*p is a multiple of 2^64. Since
322 	 * p = -1 mod 2^64, we can compute f = d[0] + a[u]*b[0] mod 2^64.
323 	 */
324 
325 	/*
326 	 * Step 1: t <- (a[0]*b + f*p) / 2^64
327 	 * We have f = a[0]*b[0] mod 2^64. Since p = -1 mod 2^64, this
328 	 * ensures that (a[0]*b + f*p) is a multiple of 2^64.
329 	 *
330 	 * We also have: f*p = f*2^256 - f*2^224 + f*2^192 + f*2^96 - f.
331 	 */
332 	x = a[0];
333 
334 	zl = _umul128(b[0], x, &zh);
335 	f = zl;
336 	t0 = zh;
337 
338 	zl = _umul128(b[1], x, &zh);
339 	k = _addcarry_u64(0, zl, t0, &zl);
340 	(void)_addcarry_u64(k, zh, 0, &zh);
341 	k = _addcarry_u64(0, zl, f << 32, &zl);
342 	(void)_addcarry_u64(k, zh, 0, &zh);
343 	t0 = zl;
344 	t1 = zh;
345 
346 	zl = _umul128(b[2], x, &zh);
347 	k = _addcarry_u64(0, zl, t1, &zl);
348 	(void)_addcarry_u64(k, zh, 0, &zh);
349 	k = _addcarry_u64(0, zl, f >> 32, &zl);
350 	(void)_addcarry_u64(k, zh, 0, &zh);
351 	t1 = zl;
352 	t2 = zh;
353 
354 	zl = _umul128(b[3], x, &zh);
355 	k = _addcarry_u64(0, zl, t2, &zl);
356 	(void)_addcarry_u64(k, zh, 0, &zh);
357 	k = _addcarry_u64(0, zl, f, &zl);
358 	(void)_addcarry_u64(k, zh, 0, &zh);
359 	t2 = zl;
360 	t3 = zh;
361 
362 	t4 = _addcarry_u64(0, t3, f, &t3);
363 	k = _subborrow_u64(0, t2, f << 32, &t2);
364 	k = _subborrow_u64(k, t3, f >> 32, &t3);
365 	(void)_subborrow_u64(k, t4, 0, &t4);
366 
367 	/*
368 	 * Steps 2 to 4: t <- (t + a[i]*b + f*p) / 2^64
369 	 */
370 	for (i = 1; i < 4; i ++) {
371 		x = a[i];
372 		/* f = t0 + x * b[0]; -- computed below */
373 
374 		/* t <- (t + x*b - f) / 2^64 */
375 		zl = _umul128(b[0], x, &zh);
376 		k = _addcarry_u64(0, zl, t0, &f);
377 		(void)_addcarry_u64(k, zh, 0, &t0);
378 
379 		zl = _umul128(b[1], x, &zh);
380 		k = _addcarry_u64(0, zl, t0, &zl);
381 		(void)_addcarry_u64(k, zh, 0, &zh);
382 		k = _addcarry_u64(0, zl, t1, &t0);
383 		(void)_addcarry_u64(k, zh, 0, &t1);
384 
385 		zl = _umul128(b[2], x, &zh);
386 		k = _addcarry_u64(0, zl, t1, &zl);
387 		(void)_addcarry_u64(k, zh, 0, &zh);
388 		k = _addcarry_u64(0, zl, t2, &t1);
389 		(void)_addcarry_u64(k, zh, 0, &t2);
390 
391 		zl = _umul128(b[3], x, &zh);
392 		k = _addcarry_u64(0, zl, t2, &zl);
393 		(void)_addcarry_u64(k, zh, 0, &zh);
394 		k = _addcarry_u64(0, zl, t3, &t2);
395 		(void)_addcarry_u64(k, zh, 0, &t3);
396 
397 		t4 = _addcarry_u64(0, t3, t4, &t3);
398 
399 		/* t <- t + f*2^32, carry in k */
400 		k = _addcarry_u64(0, t0, f << 32, &t0);
401 		k = _addcarry_u64(k, t1, f >> 32, &t1);
402 
403 		/* t <- t + f*2^192 - f*2^160 + f*2^128 */
404 		m = _subborrow_u64(0, f, f << 32, &ffl);
405 		(void)_subborrow_u64(m, f, f >> 32, &ffh);
406 		k = _addcarry_u64(k, t2, ffl, &t2);
407 		k = _addcarry_u64(k, t3, ffh, &t3);
408 		(void)_addcarry_u64(k, t4, 0, &t4);
409 	}
410 
411 	/*
412 	 * At that point, we have computed t = (a*b + F*p) / 2^256, where
413 	 * F is a 256-bit integer whose limbs are the "f" coefficients
414 	 * in the steps above. We have:
415 	 *   a <= 2^256-1
416 	 *   b <= 2^256-1
417 	 *   F <= 2^256-1
418 	 * Hence:
419 	 *   a*b + F*p <= (2^256-1)*(2^256-1) + p*(2^256-1)
420 	 *   a*b + F*p <= 2^256*(2^256 - 2 + p) + 1 - p
421 	 * Therefore:
422 	 *   t < 2^256 + p - 2
423 	 * Since p < 2^256, it follows that:
424 	 *   t4 can be only 0 or 1
425 	 *   t - p < 2^256
426 	 * We can therefore subtract p from t, conditionally on t4, to
427 	 * get a nonnegative result that fits on 256 bits.
428 	 */
429 	k = _addcarry_u64(0, t0, t4, &t0);
430 	k = _addcarry_u64(k, t1, -(t4 << 32), &t1);
431 	k = _addcarry_u64(k, t2, -t4, &t2);
432 	(void)_addcarry_u64(k, t3, (t4 << 32) - (t4 << 1), &t3);
433 
434 	d[0] = t0;
435 	d[1] = t1;
436 	d[2] = t2;
437 	d[3] = t3;
438 
439 #endif
440 }
441 
442 /*
443  * Montgomery squaring in the field; currently a basic wrapper around
444  * multiplication (inline, should be optimized away).
445  * TODO: see if some extra speed can be gained here.
446  */
447 static inline void
448 f256_montysquare(uint64_t *d, const uint64_t *a)
449 {
450 	f256_montymul(d, a, a);
451 }
452 
453 /*
454  * Convert to Montgomery representation.
455  */
456 static void
457 f256_tomonty(uint64_t *d, const uint64_t *a)
458 {
459 	/*
460 	 * R2 = 2^512 mod p.
461 	 * If R = 2^256 mod p, then R2 = R^2 mod p; and the Montgomery
462 	 * multiplication of a by R2 is: a*R2/R = a*R mod p, i.e. the
463 	 * conversion to Montgomery representation.
464 	 */
465 	static const uint64_t R2[] = {
466 		0x0000000000000003,
467 		0xFFFFFFFBFFFFFFFF,
468 		0xFFFFFFFFFFFFFFFE,
469 		0x00000004FFFFFFFD
470 	};
471 
472 	f256_montymul(d, a, R2);
473 }
474 
475 /*
476  * Convert from Montgomery representation.
477  */
478 static void
479 f256_frommonty(uint64_t *d, const uint64_t *a)
480 {
481 	/*
482 	 * Montgomery multiplication by 1 is division by 2^256 modulo p.
483 	 */
484 	static const uint64_t one[] = { 1, 0, 0, 0 };
485 
486 	f256_montymul(d, a, one);
487 }
488 
489 /*
490  * Inversion in the field. If the source value is 0 modulo p, then this
491  * returns 0 or p. This function uses Montgomery representation.
492  */
493 static void
494 f256_invert(uint64_t *d, const uint64_t *a)
495 {
496 	/*
497 	 * We compute a^(p-2) mod p. The exponent pattern (from high to
498 	 * low) is:
499 	 *  - 32 bits of value 1
500 	 *  - 31 bits of value 0
501 	 *  - 1 bit of value 1
502 	 *  - 96 bits of value 0
503 	 *  - 94 bits of value 1
504 	 *  - 1 bit of value 0
505 	 *  - 1 bit of value 1
506 	 * To speed up the square-and-multiply algorithm, we precompute
507 	 * a^(2^31-1).
508 	 */
509 
510 	uint64_t r[4], t[4];
511 	int i;
512 
513 	memcpy(t, a, sizeof t);
514 	for (i = 0; i < 30; i ++) {
515 		f256_montysquare(t, t);
516 		f256_montymul(t, t, a);
517 	}
518 
519 	memcpy(r, t, sizeof t);
520 	for (i = 224; i >= 0; i --) {
521 		f256_montysquare(r, r);
522 		switch (i) {
523 		case 0:
524 		case 2:
525 		case 192:
526 		case 224:
527 			f256_montymul(r, r, a);
528 			break;
529 		case 3:
530 		case 34:
531 		case 65:
532 			f256_montymul(r, r, t);
533 			break;
534 		}
535 	}
536 	memcpy(d, r, sizeof r);
537 }
538 
539 /*
540  * Finalize reduction.
541  * Input value fits on 256 bits. This function subtracts p if and only
542  * if the input is greater than or equal to p.
543  */
544 static inline void
545 f256_final_reduce(uint64_t *a)
546 {
547 #if BR_INT128
548 
549 	uint64_t t0, t1, t2, t3, cc;
550 	unsigned __int128 z;
551 
552 	/*
553 	 * We add 2^224 - 2^192 - 2^96 + 1 to a. If there is no carry,
554 	 * then a < p; otherwise, the addition result we computed is
555 	 * the value we must return.
556 	 */
557 	z = (unsigned __int128)a[0] + 1;
558 	t0 = (uint64_t)z;
559 	z = (unsigned __int128)a[1] + (z >> 64) - ((uint64_t)1 << 32);
560 	t1 = (uint64_t)z;
561 	z = (unsigned __int128)a[2] - (z >> 127);
562 	t2 = (uint64_t)z;
563 	z = (unsigned __int128)a[3] - (z >> 127) + 0xFFFFFFFF;
564 	t3 = (uint64_t)z;
565 	cc = -(uint64_t)(z >> 64);
566 
567 	a[0] ^= cc & (a[0] ^ t0);
568 	a[1] ^= cc & (a[1] ^ t1);
569 	a[2] ^= cc & (a[2] ^ t2);
570 	a[3] ^= cc & (a[3] ^ t3);
571 
572 #elif BR_UMUL128
573 
574 	uint64_t t0, t1, t2, t3, m;
575 	unsigned char k;
576 
577 	k = _addcarry_u64(0, a[0], (uint64_t)1, &t0);
578 	k = _addcarry_u64(k, a[1], -((uint64_t)1 << 32), &t1);
579 	k = _addcarry_u64(k, a[2], -(uint64_t)1, &t2);
580 	k = _addcarry_u64(k, a[3], ((uint64_t)1 << 32) - 2, &t3);
581 	m = -(uint64_t)k;
582 
583 	a[0] ^= m & (a[0] ^ t0);
584 	a[1] ^= m & (a[1] ^ t1);
585 	a[2] ^= m & (a[2] ^ t2);
586 	a[3] ^= m & (a[3] ^ t3);
587 
588 #endif
589 }
590 
591 /*
592  * Points in affine and Jacobian coordinates.
593  *
594  *  - In affine coordinates, the point-at-infinity cannot be encoded.
595  *  - Jacobian coordinates (X,Y,Z) correspond to affine (X/Z^2,Y/Z^3);
596  *    if Z = 0 then this is the point-at-infinity.
597  */
598 typedef struct {
599 	uint64_t x[4];
600 	uint64_t y[4];
601 } p256_affine;
602 
603 typedef struct {
604 	uint64_t x[4];
605 	uint64_t y[4];
606 	uint64_t z[4];
607 } p256_jacobian;
608 
609 /*
610  * Decode a point. The returned point is in Jacobian coordinates, but
611  * with z = 1. If the encoding is invalid, or encodes a point which is
612  * not on the curve, or encodes the point at infinity, then this function
613  * returns 0. Otherwise, 1 is returned.
614  *
615  * The buffer is assumed to have length exactly 65 bytes.
616  */
617 static uint32_t
618 point_decode(p256_jacobian *P, const unsigned char *buf)
619 {
620 	uint64_t x[4], y[4], t[4], x3[4], tt;
621 	uint32_t r;
622 
623 	/*
624 	 * Header byte shall be 0x04.
625 	 */
626 	r = EQ(buf[0], 0x04);
627 
628 	/*
629 	 * Decode X and Y coordinates, and convert them into
630 	 * Montgomery representation.
631 	 */
632 	x[3] = br_dec64be(buf +  1);
633 	x[2] = br_dec64be(buf +  9);
634 	x[1] = br_dec64be(buf + 17);
635 	x[0] = br_dec64be(buf + 25);
636 	y[3] = br_dec64be(buf + 33);
637 	y[2] = br_dec64be(buf + 41);
638 	y[1] = br_dec64be(buf + 49);
639 	y[0] = br_dec64be(buf + 57);
640 	f256_tomonty(x, x);
641 	f256_tomonty(y, y);
642 
643 	/*
644 	 * Verify y^2 = x^3 + A*x + B. In curve P-256, A = -3.
645 	 * Note that the Montgomery representation of 0 is 0. We must
646 	 * take care to apply the final reduction to make sure we have
647 	 * 0 and not p.
648 	 */
649 	f256_montysquare(t, y);
650 	f256_montysquare(x3, x);
651 	f256_montymul(x3, x3, x);
652 	f256_sub(t, t, x3);
653 	f256_add(t, t, x);
654 	f256_add(t, t, x);
655 	f256_add(t, t, x);
656 	f256_sub(t, t, P256_B_MONTY);
657 	f256_final_reduce(t);
658 	tt = t[0] | t[1] | t[2] | t[3];
659 	r &= EQ((uint32_t)(tt | (tt >> 32)), 0);
660 
661 	/*
662 	 * Return the point in Jacobian coordinates (and Montgomery
663 	 * representation).
664 	 */
665 	memcpy(P->x, x, sizeof x);
666 	memcpy(P->y, y, sizeof y);
667 	memcpy(P->z, F256_R, sizeof F256_R);
668 	return r;
669 }
670 
671 /*
672  * Final conversion for a point:
673  *  - The point is converted back to affine coordinates.
674  *  - Final reduction is performed.
675  *  - The point is encoded into the provided buffer.
676  *
677  * If the point is the point-at-infinity, all operations are performed,
678  * but the buffer contents are indeterminate, and 0 is returned. Otherwise,
679  * the encoded point is written in the buffer, and 1 is returned.
680  */
681 static uint32_t
682 point_encode(unsigned char *buf, const p256_jacobian *P)
683 {
684 	uint64_t t1[4], t2[4], z;
685 
686 	/* Set t1 = 1/z^2 and t2 = 1/z^3. */
687 	f256_invert(t2, P->z);
688 	f256_montysquare(t1, t2);
689 	f256_montymul(t2, t2, t1);
690 
691 	/* Compute affine coordinates x (in t1) and y (in t2). */
692 	f256_montymul(t1, P->x, t1);
693 	f256_montymul(t2, P->y, t2);
694 
695 	/* Convert back from Montgomery representation, and finalize
696 	   reductions. */
697 	f256_frommonty(t1, t1);
698 	f256_frommonty(t2, t2);
699 	f256_final_reduce(t1);
700 	f256_final_reduce(t2);
701 
702 	/* Encode. */
703 	buf[0] = 0x04;
704 	br_enc64be(buf +  1, t1[3]);
705 	br_enc64be(buf +  9, t1[2]);
706 	br_enc64be(buf + 17, t1[1]);
707 	br_enc64be(buf + 25, t1[0]);
708 	br_enc64be(buf + 33, t2[3]);
709 	br_enc64be(buf + 41, t2[2]);
710 	br_enc64be(buf + 49, t2[1]);
711 	br_enc64be(buf + 57, t2[0]);
712 
713 	/* Return success if and only if P->z != 0. */
714 	z = P->z[0] | P->z[1] | P->z[2] | P->z[3];
715 	return NEQ((uint32_t)(z | z >> 32), 0);
716 }
717 
718 /*
719  * Point doubling in Jacobian coordinates: point P is doubled.
720  * Note: if the source point is the point-at-infinity, then the result is
721  * still the point-at-infinity, which is correct. Moreover, if the three
722  * coordinates were zero, then they still are zero in the returned value.
723  *
724  * (Note: this is true even without the final reduction: if the three
725  * coordinates are encoded as four words of value zero each, then the
726  * result will also have all-zero coordinate encodings, not the alternate
727  * encoding as the integer p.)
728  */
729 static void
730 p256_double(p256_jacobian *P)
731 {
732 	/*
733 	 * Doubling formulas are:
734 	 *
735 	 *   s = 4*x*y^2
736 	 *   m = 3*(x + z^2)*(x - z^2)
737 	 *   x' = m^2 - 2*s
738 	 *   y' = m*(s - x') - 8*y^4
739 	 *   z' = 2*y*z
740 	 *
741 	 * These formulas work for all points, including points of order 2
742 	 * and points at infinity:
743 	 *   - If y = 0 then z' = 0. But there is no such point in P-256
744 	 *     anyway.
745 	 *   - If z = 0 then z' = 0.
746 	 */
747 	uint64_t t1[4], t2[4], t3[4], t4[4];
748 
749 	/*
750 	 * Compute z^2 in t1.
751 	 */
752 	f256_montysquare(t1, P->z);
753 
754 	/*
755 	 * Compute x-z^2 in t2 and x+z^2 in t1.
756 	 */
757 	f256_add(t2, P->x, t1);
758 	f256_sub(t1, P->x, t1);
759 
760 	/*
761 	 * Compute 3*(x+z^2)*(x-z^2) in t1.
762 	 */
763 	f256_montymul(t3, t1, t2);
764 	f256_add(t1, t3, t3);
765 	f256_add(t1, t3, t1);
766 
767 	/*
768 	 * Compute 4*x*y^2 (in t2) and 2*y^2 (in t3).
769 	 */
770 	f256_montysquare(t3, P->y);
771 	f256_add(t3, t3, t3);
772 	f256_montymul(t2, P->x, t3);
773 	f256_add(t2, t2, t2);
774 
775 	/*
776 	 * Compute x' = m^2 - 2*s.
777 	 */
778 	f256_montysquare(P->x, t1);
779 	f256_sub(P->x, P->x, t2);
780 	f256_sub(P->x, P->x, t2);
781 
782 	/*
783 	 * Compute z' = 2*y*z.
784 	 */
785 	f256_montymul(t4, P->y, P->z);
786 	f256_add(P->z, t4, t4);
787 
788 	/*
789 	 * Compute y' = m*(s - x') - 8*y^4. Note that we already have
790 	 * 2*y^2 in t3.
791 	 */
792 	f256_sub(t2, t2, P->x);
793 	f256_montymul(P->y, t1, t2);
794 	f256_montysquare(t4, t3);
795 	f256_add(t4, t4, t4);
796 	f256_sub(P->y, P->y, t4);
797 }
798 
799 /*
800  * Point addition (Jacobian coordinates): P1 is replaced with P1+P2.
801  * This function computes the wrong result in the following cases:
802  *
803  *   - If P1 == 0 but P2 != 0
804  *   - If P1 != 0 but P2 == 0
805  *   - If P1 == P2
806  *
807  * In all three cases, P1 is set to the point at infinity.
808  *
809  * Returned value is 0 if one of the following occurs:
810  *
811  *   - P1 and P2 have the same Y coordinate.
812  *   - P1 == 0 and P2 == 0.
813  *   - The Y coordinate of one of the points is 0 and the other point is
814  *     the point at infinity.
815  *
816  * The third case cannot actually happen with valid points, since a point
817  * with Y == 0 is a point of order 2, and there is no point of order 2 on
818  * curve P-256.
819  *
820  * Therefore, assuming that P1 != 0 and P2 != 0 on input, then the caller
821  * can apply the following:
822  *
823  *   - If the result is not the point at infinity, then it is correct.
824  *   - Otherwise, if the returned value is 1, then this is a case of
825  *     P1+P2 == 0, so the result is indeed the point at infinity.
826  *   - Otherwise, P1 == P2, so a "double" operation should have been
827  *     performed.
828  *
829  * Note that you can get a returned value of 0 with a correct result,
830  * e.g. if P1 and P2 have the same Y coordinate, but distinct X coordinates.
831  */
832 static uint32_t
833 p256_add(p256_jacobian *P1, const p256_jacobian *P2)
834 {
835 	/*
836 	 * Addtions formulas are:
837 	 *
838 	 *   u1 = x1 * z2^2
839 	 *   u2 = x2 * z1^2
840 	 *   s1 = y1 * z2^3
841 	 *   s2 = y2 * z1^3
842 	 *   h = u2 - u1
843 	 *   r = s2 - s1
844 	 *   x3 = r^2 - h^3 - 2 * u1 * h^2
845 	 *   y3 = r * (u1 * h^2 - x3) - s1 * h^3
846 	 *   z3 = h * z1 * z2
847 	 */
848 	uint64_t t1[4], t2[4], t3[4], t4[4], t5[4], t6[4], t7[4], tt;
849 	uint32_t ret;
850 
851 	/*
852 	 * Compute u1 = x1*z2^2 (in t1) and s1 = y1*z2^3 (in t3).
853 	 */
854 	f256_montysquare(t3, P2->z);
855 	f256_montymul(t1, P1->x, t3);
856 	f256_montymul(t4, P2->z, t3);
857 	f256_montymul(t3, P1->y, t4);
858 
859 	/*
860 	 * Compute u2 = x2*z1^2 (in t2) and s2 = y2*z1^3 (in t4).
861 	 */
862 	f256_montysquare(t4, P1->z);
863 	f256_montymul(t2, P2->x, t4);
864 	f256_montymul(t5, P1->z, t4);
865 	f256_montymul(t4, P2->y, t5);
866 
867 	/*
868 	 * Compute h = h2 - u1 (in t2) and r = s2 - s1 (in t4).
869 	 * We need to test whether r is zero, so we will do some extra
870 	 * reduce.
871 	 */
872 	f256_sub(t2, t2, t1);
873 	f256_sub(t4, t4, t3);
874 	f256_final_reduce(t4);
875 	tt = t4[0] | t4[1] | t4[2] | t4[3];
876 	ret = (uint32_t)(tt | (tt >> 32));
877 	ret = (ret | -ret) >> 31;
878 
879 	/*
880 	 * Compute u1*h^2 (in t6) and h^3 (in t5);
881 	 */
882 	f256_montysquare(t7, t2);
883 	f256_montymul(t6, t1, t7);
884 	f256_montymul(t5, t7, t2);
885 
886 	/*
887 	 * Compute x3 = r^2 - h^3 - 2*u1*h^2.
888 	 */
889 	f256_montysquare(P1->x, t4);
890 	f256_sub(P1->x, P1->x, t5);
891 	f256_sub(P1->x, P1->x, t6);
892 	f256_sub(P1->x, P1->x, t6);
893 
894 	/*
895 	 * Compute y3 = r*(u1*h^2 - x3) - s1*h^3.
896 	 */
897 	f256_sub(t6, t6, P1->x);
898 	f256_montymul(P1->y, t4, t6);
899 	f256_montymul(t1, t5, t3);
900 	f256_sub(P1->y, P1->y, t1);
901 
902 	/*
903 	 * Compute z3 = h*z1*z2.
904 	 */
905 	f256_montymul(t1, P1->z, P2->z);
906 	f256_montymul(P1->z, t1, t2);
907 
908 	return ret;
909 }
910 
911 /*
912  * Point addition (mixed coordinates): P1 is replaced with P1+P2.
913  * This is a specialised function for the case when P2 is a non-zero point
914  * in affine coordinates.
915  *
916  * This function computes the wrong result in the following cases:
917  *
918  *   - If P1 == 0
919  *   - If P1 == P2
920  *
921  * In both cases, P1 is set to the point at infinity.
922  *
923  * Returned value is 0 if one of the following occurs:
924  *
925  *   - P1 and P2 have the same Y (affine) coordinate.
926  *   - The Y coordinate of P2 is 0 and P1 is the point at infinity.
927  *
928  * The second case cannot actually happen with valid points, since a point
929  * with Y == 0 is a point of order 2, and there is no point of order 2 on
930  * curve P-256.
931  *
932  * Therefore, assuming that P1 != 0 on input, then the caller
933  * can apply the following:
934  *
935  *   - If the result is not the point at infinity, then it is correct.
936  *   - Otherwise, if the returned value is 1, then this is a case of
937  *     P1+P2 == 0, so the result is indeed the point at infinity.
938  *   - Otherwise, P1 == P2, so a "double" operation should have been
939  *     performed.
940  *
941  * Again, a value of 0 may be returned in some cases where the addition
942  * result is correct.
943  */
944 static uint32_t
945 p256_add_mixed(p256_jacobian *P1, const p256_affine *P2)
946 {
947 	/*
948 	 * Addtions formulas are:
949 	 *
950 	 *   u1 = x1
951 	 *   u2 = x2 * z1^2
952 	 *   s1 = y1
953 	 *   s2 = y2 * z1^3
954 	 *   h = u2 - u1
955 	 *   r = s2 - s1
956 	 *   x3 = r^2 - h^3 - 2 * u1 * h^2
957 	 *   y3 = r * (u1 * h^2 - x3) - s1 * h^3
958 	 *   z3 = h * z1
959 	 */
960 	uint64_t t1[4], t2[4], t3[4], t4[4], t5[4], t6[4], t7[4], tt;
961 	uint32_t ret;
962 
963 	/*
964 	 * Compute u1 = x1 (in t1) and s1 = y1 (in t3).
965 	 */
966 	memcpy(t1, P1->x, sizeof t1);
967 	memcpy(t3, P1->y, sizeof t3);
968 
969 	/*
970 	 * Compute u2 = x2*z1^2 (in t2) and s2 = y2*z1^3 (in t4).
971 	 */
972 	f256_montysquare(t4, P1->z);
973 	f256_montymul(t2, P2->x, t4);
974 	f256_montymul(t5, P1->z, t4);
975 	f256_montymul(t4, P2->y, t5);
976 
977 	/*
978 	 * Compute h = h2 - u1 (in t2) and r = s2 - s1 (in t4).
979 	 * We need to test whether r is zero, so we will do some extra
980 	 * reduce.
981 	 */
982 	f256_sub(t2, t2, t1);
983 	f256_sub(t4, t4, t3);
984 	f256_final_reduce(t4);
985 	tt = t4[0] | t4[1] | t4[2] | t4[3];
986 	ret = (uint32_t)(tt | (tt >> 32));
987 	ret = (ret | -ret) >> 31;
988 
989 	/*
990 	 * Compute u1*h^2 (in t6) and h^3 (in t5);
991 	 */
992 	f256_montysquare(t7, t2);
993 	f256_montymul(t6, t1, t7);
994 	f256_montymul(t5, t7, t2);
995 
996 	/*
997 	 * Compute x3 = r^2 - h^3 - 2*u1*h^2.
998 	 */
999 	f256_montysquare(P1->x, t4);
1000 	f256_sub(P1->x, P1->x, t5);
1001 	f256_sub(P1->x, P1->x, t6);
1002 	f256_sub(P1->x, P1->x, t6);
1003 
1004 	/*
1005 	 * Compute y3 = r*(u1*h^2 - x3) - s1*h^3.
1006 	 */
1007 	f256_sub(t6, t6, P1->x);
1008 	f256_montymul(P1->y, t4, t6);
1009 	f256_montymul(t1, t5, t3);
1010 	f256_sub(P1->y, P1->y, t1);
1011 
1012 	/*
1013 	 * Compute z3 = h*z1*z2.
1014 	 */
1015 	f256_montymul(P1->z, P1->z, t2);
1016 
1017 	return ret;
1018 }
1019 
1020 #if 0
1021 /* unused */
1022 /*
1023  * Point addition (mixed coordinates, complete): P1 is replaced with P1+P2.
1024  * This is a specialised function for the case when P2 is a non-zero point
1025  * in affine coordinates.
1026  *
1027  * This function returns the correct result in all cases.
1028  */
1029 static uint32_t
1030 p256_add_complete_mixed(p256_jacobian *P1, const p256_affine *P2)
1031 {
1032 	/*
1033 	 * Addtions formulas, in the general case, are:
1034 	 *
1035 	 *   u1 = x1
1036 	 *   u2 = x2 * z1^2
1037 	 *   s1 = y1
1038 	 *   s2 = y2 * z1^3
1039 	 *   h = u2 - u1
1040 	 *   r = s2 - s1
1041 	 *   x3 = r^2 - h^3 - 2 * u1 * h^2
1042 	 *   y3 = r * (u1 * h^2 - x3) - s1 * h^3
1043 	 *   z3 = h * z1
1044 	 *
1045 	 * These formulas mishandle the two following cases:
1046 	 *
1047 	 *  - If P1 is the point-at-infinity (z1 = 0), then z3 is
1048 	 *    incorrectly set to 0.
1049 	 *
1050 	 *  - If P1 = P2, then u1 = u2 and s1 = s2, and x3, y3 and z3
1051 	 *    are all set to 0.
1052 	 *
1053 	 * However, if P1 + P2 = 0, then u1 = u2 but s1 != s2, and then
1054 	 * we correctly get z3 = 0 (the point-at-infinity).
1055 	 *
1056 	 * To fix the case P1 = 0, we perform at the end a copy of P2
1057 	 * over P1, conditional to z1 = 0.
1058 	 *
1059 	 * For P1 = P2: in that case, both h and r are set to 0, and
1060 	 * we get x3, y3 and z3 equal to 0. We can test for that
1061 	 * occurrence to make a mask which will be all-one if P1 = P2,
1062 	 * or all-zero otherwise; then we can compute the double of P2
1063 	 * and add it, combined with the mask, to (x3,y3,z3).
1064 	 *
1065 	 * Using the doubling formulas in p256_double() on (x2,y2),
1066 	 * simplifying since P2 is affine (i.e. z2 = 1, implicitly),
1067 	 * we get:
1068 	 *   s = 4*x2*y2^2
1069 	 *   m = 3*(x2 + 1)*(x2 - 1)
1070 	 *   x' = m^2 - 2*s
1071 	 *   y' = m*(s - x') - 8*y2^4
1072 	 *   z' = 2*y2
1073 	 * which requires only 6 multiplications. Added to the 11
1074 	 * multiplications of the normal mixed addition in Jacobian
1075 	 * coordinates, we get a cost of 17 multiplications in total.
1076 	 */
1077 	uint64_t t1[4], t2[4], t3[4], t4[4], t5[4], t6[4], t7[4], tt, zz;
1078 	int i;
1079 
1080 	/*
1081 	 * Set zz to -1 if P1 is the point at infinity, 0 otherwise.
1082 	 */
1083 	zz = P1->z[0] | P1->z[1] | P1->z[2] | P1->z[3];
1084 	zz = ((zz | -zz) >> 63) - (uint64_t)1;
1085 
1086 	/*
1087 	 * Compute u1 = x1 (in t1) and s1 = y1 (in t3).
1088 	 */
1089 	memcpy(t1, P1->x, sizeof t1);
1090 	memcpy(t3, P1->y, sizeof t3);
1091 
1092 	/*
1093 	 * Compute u2 = x2*z1^2 (in t2) and s2 = y2*z1^3 (in t4).
1094 	 */
1095 	f256_montysquare(t4, P1->z);
1096 	f256_montymul(t2, P2->x, t4);
1097 	f256_montymul(t5, P1->z, t4);
1098 	f256_montymul(t4, P2->y, t5);
1099 
1100 	/*
1101 	 * Compute h = h2 - u1 (in t2) and r = s2 - s1 (in t4).
1102 	 * reduce.
1103 	 */
1104 	f256_sub(t2, t2, t1);
1105 	f256_sub(t4, t4, t3);
1106 
1107 	/*
1108 	 * If both h = 0 and r = 0, then P1 = P2, and we want to set
1109 	 * the mask tt to -1; otherwise, the mask will be 0.
1110 	 */
1111 	f256_final_reduce(t2);
1112 	f256_final_reduce(t4);
1113 	tt = t2[0] | t2[1] | t2[2] | t2[3] | t4[0] | t4[1] | t4[2] | t4[3];
1114 	tt = ((tt | -tt) >> 63) - (uint64_t)1;
1115 
1116 	/*
1117 	 * Compute u1*h^2 (in t6) and h^3 (in t5);
1118 	 */
1119 	f256_montysquare(t7, t2);
1120 	f256_montymul(t6, t1, t7);
1121 	f256_montymul(t5, t7, t2);
1122 
1123 	/*
1124 	 * Compute x3 = r^2 - h^3 - 2*u1*h^2.
1125 	 */
1126 	f256_montysquare(P1->x, t4);
1127 	f256_sub(P1->x, P1->x, t5);
1128 	f256_sub(P1->x, P1->x, t6);
1129 	f256_sub(P1->x, P1->x, t6);
1130 
1131 	/*
1132 	 * Compute y3 = r*(u1*h^2 - x3) - s1*h^3.
1133 	 */
1134 	f256_sub(t6, t6, P1->x);
1135 	f256_montymul(P1->y, t4, t6);
1136 	f256_montymul(t1, t5, t3);
1137 	f256_sub(P1->y, P1->y, t1);
1138 
1139 	/*
1140 	 * Compute z3 = h*z1.
1141 	 */
1142 	f256_montymul(P1->z, P1->z, t2);
1143 
1144 	/*
1145 	 * The "double" result, in case P1 = P2.
1146 	 */
1147 
1148 	/*
1149 	 * Compute z' = 2*y2 (in t1).
1150 	 */
1151 	f256_add(t1, P2->y, P2->y);
1152 
1153 	/*
1154 	 * Compute 2*(y2^2) (in t2) and s = 4*x2*(y2^2) (in t3).
1155 	 */
1156 	f256_montysquare(t2, P2->y);
1157 	f256_add(t2, t2, t2);
1158 	f256_add(t3, t2, t2);
1159 	f256_montymul(t3, P2->x, t3);
1160 
1161 	/*
1162 	 * Compute m = 3*(x2^2 - 1) (in t4).
1163 	 */
1164 	f256_montysquare(t4, P2->x);
1165 	f256_sub(t4, t4, F256_R);
1166 	f256_add(t5, t4, t4);
1167 	f256_add(t4, t4, t5);
1168 
1169 	/*
1170 	 * Compute x' = m^2 - 2*s (in t5).
1171 	 */
1172 	f256_montysquare(t5, t4);
1173 	f256_sub(t5, t3);
1174 	f256_sub(t5, t3);
1175 
1176 	/*
1177 	 * Compute y' = m*(s - x') - 8*y2^4 (in t6).
1178 	 */
1179 	f256_sub(t6, t3, t5);
1180 	f256_montymul(t6, t6, t4);
1181 	f256_montysquare(t7, t2);
1182 	f256_sub(t6, t6, t7);
1183 	f256_sub(t6, t6, t7);
1184 
1185 	/*
1186 	 * We now have the alternate (doubling) coordinates in (t5,t6,t1).
1187 	 * We combine them with (x3,y3,z3).
1188 	 */
1189 	for (i = 0; i < 4; i ++) {
1190 		P1->x[i] |= tt & t5[i];
1191 		P1->y[i] |= tt & t6[i];
1192 		P1->z[i] |= tt & t1[i];
1193 	}
1194 
1195 	/*
1196 	 * If P1 = 0, then we get z3 = 0 (which is invalid); if z1 is 0,
1197 	 * then we want to replace the result with a copy of P2. The
1198 	 * test on z1 was done at the start, in the zz mask.
1199 	 */
1200 	for (i = 0; i < 4; i ++) {
1201 		P1->x[i] ^= zz & (P1->x[i] ^ P2->x[i]);
1202 		P1->y[i] ^= zz & (P1->y[i] ^ P2->y[i]);
1203 		P1->z[i] ^= zz & (P1->z[i] ^ F256_R[i]);
1204 	}
1205 }
1206 #endif
1207 
1208 /*
1209  * Inner function for computing a point multiplication. A window is
1210  * provided, with points 1*P to 15*P in affine coordinates.
1211  *
1212  * Assumptions:
1213  *  - All provided points are valid points on the curve.
1214  *  - Multiplier is non-zero, and smaller than the curve order.
1215  *  - Everything is in Montgomery representation.
1216  */
1217 static void
1218 point_mul_inner(p256_jacobian *R, const p256_affine *W,
1219 	const unsigned char *k, size_t klen)
1220 {
1221 	p256_jacobian Q;
1222 	uint32_t qz;
1223 
1224 	memset(&Q, 0, sizeof Q);
1225 	qz = 1;
1226 	while (klen -- > 0) {
1227 		int i;
1228 		unsigned bk;
1229 
1230 		bk = *k ++;
1231 		for (i = 0; i < 2; i ++) {
1232 			uint32_t bits;
1233 			uint32_t bnz;
1234 			p256_affine T;
1235 			p256_jacobian U;
1236 			uint32_t n;
1237 			int j;
1238 			uint64_t m;
1239 
1240 			p256_double(&Q);
1241 			p256_double(&Q);
1242 			p256_double(&Q);
1243 			p256_double(&Q);
1244 			bits = (bk >> 4) & 0x0F;
1245 			bnz = NEQ(bits, 0);
1246 
1247 			/*
1248 			 * Lookup point in window. If the bits are 0,
1249 			 * we get something invalid, which is not a
1250 			 * problem because we will use it only if the
1251 			 * bits are non-zero.
1252 			 */
1253 			memset(&T, 0, sizeof T);
1254 			for (n = 0; n < 15; n ++) {
1255 				m = -(uint64_t)EQ(bits, n + 1);
1256 				T.x[0] |= m & W[n].x[0];
1257 				T.x[1] |= m & W[n].x[1];
1258 				T.x[2] |= m & W[n].x[2];
1259 				T.x[3] |= m & W[n].x[3];
1260 				T.y[0] |= m & W[n].y[0];
1261 				T.y[1] |= m & W[n].y[1];
1262 				T.y[2] |= m & W[n].y[2];
1263 				T.y[3] |= m & W[n].y[3];
1264 			}
1265 
1266 			U = Q;
1267 			p256_add_mixed(&U, &T);
1268 
1269 			/*
1270 			 * If qz is still 1, then Q was all-zeros, and this
1271 			 * is conserved through p256_double().
1272 			 */
1273 			m = -(uint64_t)(bnz & qz);
1274 			for (j = 0; j < 4; j ++) {
1275 				Q.x[j] |= m & T.x[j];
1276 				Q.y[j] |= m & T.y[j];
1277 				Q.z[j] |= m & F256_R[j];
1278 			}
1279 			CCOPY(bnz & ~qz, &Q, &U, sizeof Q);
1280 			qz &= ~bnz;
1281 			bk <<= 4;
1282 		}
1283 	}
1284 	*R = Q;
1285 }
1286 
1287 /*
1288  * Convert a window from Jacobian to affine coordinates. A single
1289  * field inversion is used. This function works for windows up to
1290  * 32 elements.
1291  *
1292  * The destination array (aff[]) and the source array (jac[]) may
1293  * overlap, provided that the start of aff[] is not after the start of
1294  * jac[]. Even if the arrays do _not_ overlap, the source array is
1295  * modified.
1296  */
1297 static void
1298 window_to_affine(p256_affine *aff, p256_jacobian *jac, int num)
1299 {
1300 	/*
1301 	 * Convert the window points to affine coordinates. We use the
1302 	 * following trick to mutualize the inversion computation: if
1303 	 * we have z1, z2, z3, and z4, and want to inverse all of them,
1304 	 * we compute u = 1/(z1*z2*z3*z4), and then we have:
1305 	 *   1/z1 = u*z2*z3*z4
1306 	 *   1/z2 = u*z1*z3*z4
1307 	 *   1/z3 = u*z1*z2*z4
1308 	 *   1/z4 = u*z1*z2*z3
1309 	 *
1310 	 * The partial products are computed recursively:
1311 	 *
1312 	 *  - on input (z_1,z_2), return (z_2,z_1) and z_1*z_2
1313 	 *  - on input (z_1,z_2,... z_n):
1314 	 *       recurse on (z_1,z_2,... z_(n/2)) -> r1 and m1
1315 	 *       recurse on (z_(n/2+1),z_(n/2+2)... z_n) -> r2 and m2
1316 	 *       multiply elements of r1 by m2 -> s1
1317 	 *       multiply elements of r2 by m1 -> s2
1318 	 *       return r1||r2 and m1*m2
1319 	 *
1320 	 * In the example below, we suppose that we have 14 elements.
1321 	 * Let z1, z2,... zE be the 14 values to invert (index noted in
1322 	 * hexadecimal, starting at 1).
1323 	 *
1324 	 *  - Depth 1:
1325 	 *      swap(z1, z2); z12 = z1*z2
1326 	 *      swap(z3, z4); z34 = z3*z4
1327 	 *      swap(z5, z6); z56 = z5*z6
1328 	 *      swap(z7, z8); z78 = z7*z8
1329 	 *      swap(z9, zA); z9A = z9*zA
1330 	 *      swap(zB, zC); zBC = zB*zC
1331 	 *      swap(zD, zE); zDE = zD*zE
1332 	 *
1333 	 *  - Depth 2:
1334 	 *      z1 <- z1*z34, z2 <- z2*z34, z3 <- z3*z12, z4 <- z4*z12
1335 	 *      z1234 = z12*z34
1336 	 *      z5 <- z5*z78, z6 <- z6*z78, z7 <- z7*z56, z8 <- z8*z56
1337 	 *      z5678 = z56*z78
1338 	 *      z9 <- z9*zBC, zA <- zA*zBC, zB <- zB*z9A, zC <- zC*z9A
1339 	 *      z9ABC = z9A*zBC
1340 	 *
1341 	 *  - Depth 3:
1342 	 *      z1 <- z1*z5678, z2 <- z2*z5678, z3 <- z3*z5678, z4 <- z4*z5678
1343 	 *      z5 <- z5*z1234, z6 <- z6*z1234, z7 <- z7*z1234, z8 <- z8*z1234
1344 	 *      z12345678 = z1234*z5678
1345 	 *      z9 <- z9*zDE, zA <- zA*zDE, zB <- zB*zDE, zC <- zC*zDE
1346 	 *      zD <- zD*z9ABC, zE*z9ABC
1347 	 *      z9ABCDE = z9ABC*zDE
1348 	 *
1349 	 *  - Depth 4:
1350 	 *      multiply z1..z8 by z9ABCDE
1351 	 *      multiply z9..zE by z12345678
1352 	 *      final z = z12345678*z9ABCDE
1353 	 */
1354 
1355 	uint64_t z[16][4];
1356 	int i, k, s;
1357 #define zt   (z[15])
1358 #define zu   (z[14])
1359 #define zv   (z[13])
1360 
1361 	/*
1362 	 * First recursion step (pairwise swapping and multiplication).
1363 	 * If there is an odd number of elements, then we "invent" an
1364 	 * extra one with coordinate Z = 1 (in Montgomery representation).
1365 	 */
1366 	for (i = 0; (i + 1) < num; i += 2) {
1367 		memcpy(zt, jac[i].z, sizeof zt);
1368 		memcpy(jac[i].z, jac[i + 1].z, sizeof zt);
1369 		memcpy(jac[i + 1].z, zt, sizeof zt);
1370 		f256_montymul(z[i >> 1], jac[i].z, jac[i + 1].z);
1371 	}
1372 	if ((num & 1) != 0) {
1373 		memcpy(z[num >> 1], jac[num - 1].z, sizeof zt);
1374 		memcpy(jac[num - 1].z, F256_R, sizeof F256_R);
1375 	}
1376 
1377 	/*
1378 	 * Perform further recursion steps. At the entry of each step,
1379 	 * the process has been done for groups of 's' points. The
1380 	 * integer k is the log2 of s.
1381 	 */
1382 	for (k = 1, s = 2; s < num; k ++, s <<= 1) {
1383 		int n;
1384 
1385 		for (i = 0; i < num; i ++) {
1386 			f256_montymul(jac[i].z, jac[i].z, z[(i >> k) ^ 1]);
1387 		}
1388 		n = (num + s - 1) >> k;
1389 		for (i = 0; i < (n >> 1); i ++) {
1390 			f256_montymul(z[i], z[i << 1], z[(i << 1) + 1]);
1391 		}
1392 		if ((n & 1) != 0) {
1393 			memmove(z[n >> 1], z[n], sizeof zt);
1394 		}
1395 	}
1396 
1397 	/*
1398 	 * Invert the final result, and convert all points.
1399 	 */
1400 	f256_invert(zt, z[0]);
1401 	for (i = 0; i < num; i ++) {
1402 		f256_montymul(zv, jac[i].z, zt);
1403 		f256_montysquare(zu, zv);
1404 		f256_montymul(zv, zv, zu);
1405 		f256_montymul(aff[i].x, jac[i].x, zu);
1406 		f256_montymul(aff[i].y, jac[i].y, zv);
1407 	}
1408 }
1409 
1410 /*
1411  * Multiply the provided point by an integer.
1412  * Assumptions:
1413  *  - Source point is a valid curve point.
1414  *  - Source point is not the point-at-infinity.
1415  *  - Integer is not 0, and is lower than the curve order.
1416  * If these conditions are not met, then the result is indeterminate
1417  * (but the process is still constant-time).
1418  */
1419 static void
1420 p256_mul(p256_jacobian *P, const unsigned char *k, size_t klen)
1421 {
1422 	union {
1423 		p256_affine aff[15];
1424 		p256_jacobian jac[15];
1425 	} window;
1426 	int i;
1427 
1428 	/*
1429 	 * Compute window, in Jacobian coordinates.
1430 	 */
1431 	window.jac[0] = *P;
1432 	for (i = 2; i < 16; i ++) {
1433 		window.jac[i - 1] = window.jac[(i >> 1) - 1];
1434 		if ((i & 1) == 0) {
1435 			p256_double(&window.jac[i - 1]);
1436 		} else {
1437 			p256_add(&window.jac[i - 1], &window.jac[i >> 1]);
1438 		}
1439 	}
1440 
1441 	/*
1442 	 * Convert the window points to affine coordinates. Point
1443 	 * window[0] is the source point, already in affine coordinates.
1444 	 */
1445 	window_to_affine(window.aff, window.jac, 15);
1446 
1447 	/*
1448 	 * Perform point multiplication.
1449 	 */
1450 	point_mul_inner(P, window.aff, k, klen);
1451 }
1452 
1453 /*
1454  * Precomputed window for the conventional generator: P256_Gwin[n]
1455  * contains (n+1)*G (affine coordinates, in Montgomery representation).
1456  */
1457 static const p256_affine P256_Gwin[] = {
1458 	{
1459 		{ 0x79E730D418A9143C, 0x75BA95FC5FEDB601,
1460 		  0x79FB732B77622510, 0x18905F76A53755C6 },
1461 		{ 0xDDF25357CE95560A, 0x8B4AB8E4BA19E45C,
1462 		  0xD2E88688DD21F325, 0x8571FF1825885D85 }
1463 	},
1464 	{
1465 		{ 0x850046D410DDD64D, 0xAA6AE3C1A433827D,
1466 		  0x732205038D1490D9, 0xF6BB32E43DCF3A3B },
1467 		{ 0x2F3648D361BEE1A5, 0x152CD7CBEB236FF8,
1468 		  0x19A8FB0E92042DBE, 0x78C577510A5B8A3B }
1469 	},
1470 	{
1471 		{ 0xFFAC3F904EEBC127, 0xB027F84A087D81FB,
1472 		  0x66AD77DD87CBBC98, 0x26936A3FB6FF747E },
1473 		{ 0xB04C5C1FC983A7EB, 0x583E47AD0861FE1A,
1474 		  0x788208311A2EE98E, 0xD5F06A29E587CC07 }
1475 	},
1476 	{
1477 		{ 0x74B0B50D46918DCC, 0x4650A6EDC623C173,
1478 		  0x0CDAACACE8100AF2, 0x577362F541B0176B },
1479 		{ 0x2D96F24CE4CBABA6, 0x17628471FAD6F447,
1480 		  0x6B6C36DEE5DDD22E, 0x84B14C394C5AB863 }
1481 	},
1482 	{
1483 		{ 0xBE1B8AAEC45C61F5, 0x90EC649A94B9537D,
1484 		  0x941CB5AAD076C20C, 0xC9079605890523C8 },
1485 		{ 0xEB309B4AE7BA4F10, 0x73C568EFE5EB882B,
1486 		  0x3540A9877E7A1F68, 0x73A076BB2DD1E916 }
1487 	},
1488 	{
1489 		{ 0x403947373E77664A, 0x55AE744F346CEE3E,
1490 		  0xD50A961A5B17A3AD, 0x13074B5954213673 },
1491 		{ 0x93D36220D377E44B, 0x299C2B53ADFF14B5,
1492 		  0xF424D44CEF639F11, 0xA4C9916D4A07F75F }
1493 	},
1494 	{
1495 		{ 0x0746354EA0173B4F, 0x2BD20213D23C00F7,
1496 		  0xF43EAAB50C23BB08, 0x13BA5119C3123E03 },
1497 		{ 0x2847D0303F5B9D4D, 0x6742F2F25DA67BDD,
1498 		  0xEF933BDC77C94195, 0xEAEDD9156E240867 }
1499 	},
1500 	{
1501 		{ 0x27F14CD19499A78F, 0x462AB5C56F9B3455,
1502 		  0x8F90F02AF02CFC6B, 0xB763891EB265230D },
1503 		{ 0xF59DA3A9532D4977, 0x21E3327DCF9EBA15,
1504 		  0x123C7B84BE60BBF0, 0x56EC12F27706DF76 }
1505 	},
1506 	{
1507 		{ 0x75C96E8F264E20E8, 0xABE6BFED59A7A841,
1508 		  0x2CC09C0444C8EB00, 0xE05B3080F0C4E16B },
1509 		{ 0x1EB7777AA45F3314, 0x56AF7BEDCE5D45E3,
1510 		  0x2B6E019A88B12F1A, 0x086659CDFD835F9B }
1511 	},
1512 	{
1513 		{ 0x2C18DBD19DC21EC8, 0x98F9868A0FCF8139,
1514 		  0x737D2CD648250B49, 0xCC61C94724B3428F },
1515 		{ 0x0C2B407880DD9E76, 0xC43A8991383FBE08,
1516 		  0x5F7D2D65779BE5D2, 0x78719A54EB3B4AB5 }
1517 	},
1518 	{
1519 		{ 0xEA7D260A6245E404, 0x9DE407956E7FDFE0,
1520 		  0x1FF3A4158DAC1AB5, 0x3E7090F1649C9073 },
1521 		{ 0x1A7685612B944E88, 0x250F939EE57F61C8,
1522 		  0x0C0DAA891EAD643D, 0x68930023E125B88E }
1523 	},
1524 	{
1525 		{ 0x04B71AA7D2697768, 0xABDEDEF5CA345A33,
1526 		  0x2409D29DEE37385E, 0x4EE1DF77CB83E156 },
1527 		{ 0x0CAC12D91CBB5B43, 0x170ED2F6CA895637,
1528 		  0x28228CFA8ADE6D66, 0x7FF57C9553238ACA }
1529 	},
1530 	{
1531 		{ 0xCCC425634B2ED709, 0x0E356769856FD30D,
1532 		  0xBCBCD43F559E9811, 0x738477AC5395B759 },
1533 		{ 0x35752B90C00EE17F, 0x68748390742ED2E3,
1534 		  0x7CD06422BD1F5BC1, 0xFBC08769C9E7B797 }
1535 	},
1536 	{
1537 		{ 0xA242A35BB0CF664A, 0x126E48F77F9707E3,
1538 		  0x1717BF54C6832660, 0xFAAE7332FD12C72E },
1539 		{ 0x27B52DB7995D586B, 0xBE29569E832237C2,
1540 		  0xE8E4193E2A65E7DB, 0x152706DC2EAA1BBB }
1541 	},
1542 	{
1543 		{ 0x72BCD8B7BC60055B, 0x03CC23EE56E27E4B,
1544 		  0xEE337424E4819370, 0xE2AA0E430AD3DA09 },
1545 		{ 0x40B8524F6383C45D, 0xD766355442A41B25,
1546 		  0x64EFA6DE778A4797, 0x2042170A7079ADF4 }
1547 	}
1548 };
1549 
1550 /*
1551  * Multiply the conventional generator of the curve by the provided
1552  * integer. Return is written in *P.
1553  *
1554  * Assumptions:
1555  *  - Integer is not 0, and is lower than the curve order.
1556  * If this conditions is not met, then the result is indeterminate
1557  * (but the process is still constant-time).
1558  */
1559 static void
1560 p256_mulgen(p256_jacobian *P, const unsigned char *k, size_t klen)
1561 {
1562 	point_mul_inner(P, P256_Gwin, k, klen);
1563 }
1564 
1565 /*
1566  * Return 1 if all of the following hold:
1567  *  - klen <= 32
1568  *  - k != 0
1569  *  - k is lower than the curve order
1570  * Otherwise, return 0.
1571  *
1572  * Constant-time behaviour: only klen may be observable.
1573  */
1574 static uint32_t
1575 check_scalar(const unsigned char *k, size_t klen)
1576 {
1577 	uint32_t z;
1578 	int32_t c;
1579 	size_t u;
1580 
1581 	if (klen > 32) {
1582 		return 0;
1583 	}
1584 	z = 0;
1585 	for (u = 0; u < klen; u ++) {
1586 		z |= k[u];
1587 	}
1588 	if (klen == 32) {
1589 		c = 0;
1590 		for (u = 0; u < klen; u ++) {
1591 			c |= -(int32_t)EQ0(c) & CMP(k[u], P256_N[u]);
1592 		}
1593 	} else {
1594 		c = -1;
1595 	}
1596 	return NEQ(z, 0) & LT0(c);
1597 }
1598 
1599 static uint32_t
1600 api_mul(unsigned char *G, size_t Glen,
1601 	const unsigned char *k, size_t klen, int curve)
1602 {
1603 	uint32_t r;
1604 	p256_jacobian P;
1605 
1606 	(void)curve;
1607 	if (Glen != 65) {
1608 		return 0;
1609 	}
1610 	r = check_scalar(k, klen);
1611 	r &= point_decode(&P, G);
1612 	p256_mul(&P, k, klen);
1613 	r &= point_encode(G, &P);
1614 	return r;
1615 }
1616 
1617 static size_t
1618 api_mulgen(unsigned char *R,
1619 	const unsigned char *k, size_t klen, int curve)
1620 {
1621 	p256_jacobian P;
1622 
1623 	(void)curve;
1624 	p256_mulgen(&P, k, klen);
1625 	point_encode(R, &P);
1626 	return 65;
1627 }
1628 
1629 static uint32_t
1630 api_muladd(unsigned char *A, const unsigned char *B, size_t len,
1631 	const unsigned char *x, size_t xlen,
1632 	const unsigned char *y, size_t ylen, int curve)
1633 {
1634 	/*
1635 	 * We might want to use Shamir's trick here: make a composite
1636 	 * window of u*P+v*Q points, to merge the two doubling-ladders
1637 	 * into one. This, however, has some complications:
1638 	 *
1639 	 *  - During the computation, we may hit the point-at-infinity.
1640 	 *    Thus, we would need p256_add_complete_mixed() (complete
1641 	 *    formulas for point addition), with a higher cost (17 muls
1642 	 *    instead of 11).
1643 	 *
1644 	 *  - A 4-bit window would be too large, since it would involve
1645 	 *    16*16-1 = 255 points. For the same window size as in the
1646 	 *    p256_mul() case, we would need to reduce the window size
1647 	 *    to 2 bits, and thus perform twice as many non-doubling
1648 	 *    point additions.
1649 	 *
1650 	 *  - The window may itself contain the point-at-infinity, and
1651 	 *    thus cannot be in all generality be made of affine points.
1652 	 *    Instead, we would need to make it a window of points in
1653 	 *    Jacobian coordinates. Even p256_add_complete_mixed() would
1654 	 *    be inappropriate.
1655 	 *
1656 	 * For these reasons, the code below performs two separate
1657 	 * point multiplications, then computes the final point addition
1658 	 * (which is both a "normal" addition, and a doubling, to handle
1659 	 * all cases).
1660 	 */
1661 
1662 	p256_jacobian P, Q;
1663 	uint32_t r, t, s;
1664 	uint64_t z;
1665 
1666 	(void)curve;
1667 	if (len != 65) {
1668 		return 0;
1669 	}
1670 	r = point_decode(&P, A);
1671 	p256_mul(&P, x, xlen);
1672 	if (B == NULL) {
1673 		p256_mulgen(&Q, y, ylen);
1674 	} else {
1675 		r &= point_decode(&Q, B);
1676 		p256_mul(&Q, y, ylen);
1677 	}
1678 
1679 	/*
1680 	 * The final addition may fail in case both points are equal.
1681 	 */
1682 	t = p256_add(&P, &Q);
1683 	f256_final_reduce(P.z);
1684 	z = P.z[0] | P.z[1] | P.z[2] | P.z[3];
1685 	s = EQ((uint32_t)(z | (z >> 32)), 0);
1686 	p256_double(&Q);
1687 
1688 	/*
1689 	 * If s is 1 then either P+Q = 0 (t = 1) or P = Q (t = 0). So we
1690 	 * have the following:
1691 	 *
1692 	 *   s = 0, t = 0   return P (normal addition)
1693 	 *   s = 0, t = 1   return P (normal addition)
1694 	 *   s = 1, t = 0   return Q (a 'double' case)
1695 	 *   s = 1, t = 1   report an error (P+Q = 0)
1696 	 */
1697 	CCOPY(s & ~t, &P, &Q, sizeof Q);
1698 	point_encode(A, &P);
1699 	r &= ~(s & t);
1700 	return r;
1701 }
1702 
1703 /* see bearssl_ec.h */
1704 const br_ec_impl br_ec_p256_m64 = {
1705 	(uint32_t)0x00800000,
1706 	&api_generator,
1707 	&api_order,
1708 	&api_xoff,
1709 	&api_mul,
1710 	&api_mulgen,
1711 	&api_muladd
1712 };
1713 
1714 /* see bearssl_ec.h */
1715 const br_ec_impl *
1716 br_ec_p256_m64_get(void)
1717 {
1718 	return &br_ec_p256_m64;
1719 }
1720 
1721 #else
1722 
1723 /* see bearssl_ec.h */
1724 const br_ec_impl *
1725 br_ec_p256_m64_get(void)
1726 {
1727 	return 0;
1728 }
1729 
1730 #endif
1731