xref: /aosp_15_r20/external/boringssl/src/crypto/curve25519/curve25519.c (revision 8fb009dc861624b67b6cdb62ea21f0f22d0c584b)
1 /* Copyright (c) 2020, Google Inc.
2  *
3  * Permission to use, copy, modify, and/or distribute this software for any
4  * purpose with or without fee is hereby granted, provided that the above
5  * copyright notice and this permission notice appear in all copies.
6  *
7  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
8  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
9  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY
10  * SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
11  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION
12  * OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN
13  * CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. */
14 
15 // Some of this code is taken from the ref10 version of Ed25519 in SUPERCOP
16 // 20141124 (http://bench.cr.yp.to/supercop.html). That code is released as
17 // public domain. Other parts have been replaced to call into code generated by
18 // Fiat (https://github.com/mit-plv/fiat-crypto) in //third_party/fiat.
19 //
20 // The field functions are shared by Ed25519 and X25519 where possible.
21 
22 #include <assert.h>
23 #include <string.h>
24 
25 #include <openssl/mem.h>
26 #include <openssl/rand.h>
27 #include <openssl/sha.h>
28 
29 #include "internal.h"
30 #include "../internal.h"
31 
32 // Various pre-computed constants.
33 #include "./curve25519_tables.h"
34 
35 #if defined(BORINGSSL_HAS_UINT128)
36 #include "../../third_party/fiat/curve25519_64.h"
37 #elif defined(OPENSSL_64_BIT)
38 #include "../../third_party/fiat/curve25519_64_msvc.h"
39 #else
40 #include "../../third_party/fiat/curve25519_32.h"
41 #endif
42 
43 
44 // Low-level intrinsic operations
45 
load_3(const uint8_t * in)46 static uint64_t load_3(const uint8_t *in) {
47   uint64_t result;
48   result = (uint64_t)in[0];
49   result |= ((uint64_t)in[1]) << 8;
50   result |= ((uint64_t)in[2]) << 16;
51   return result;
52 }
53 
load_4(const uint8_t * in)54 static uint64_t load_4(const uint8_t *in) {
55   uint64_t result;
56   result = (uint64_t)in[0];
57   result |= ((uint64_t)in[1]) << 8;
58   result |= ((uint64_t)in[2]) << 16;
59   result |= ((uint64_t)in[3]) << 24;
60   return result;
61 }
62 
63 
64 // Field operations.
65 
66 #if defined(OPENSSL_64_BIT)
67 
68 typedef uint64_t fe_limb_t;
69 #define FE_NUM_LIMBS 5
70 
71 // assert_fe asserts that |f| satisfies bounds:
72 //
73 //  [[0x0 ~> 0x8cccccccccccc],
74 //   [0x0 ~> 0x8cccccccccccc],
75 //   [0x0 ~> 0x8cccccccccccc],
76 //   [0x0 ~> 0x8cccccccccccc],
77 //   [0x0 ~> 0x8cccccccccccc]]
78 //
79 // See comments in curve25519_64.h for which functions use these bounds for
80 // inputs or outputs.
81 #define assert_fe(f)                                                    \
82   do {                                                                  \
83     for (unsigned _assert_fe_i = 0; _assert_fe_i < 5; _assert_fe_i++) { \
84       declassify_assert(f[_assert_fe_i] <= UINT64_C(0x8cccccccccccc));  \
85     }                                                                   \
86   } while (0)
87 
88 // assert_fe_loose asserts that |f| satisfies bounds:
89 //
90 //  [[0x0 ~> 0x1a666666666664],
91 //   [0x0 ~> 0x1a666666666664],
92 //   [0x0 ~> 0x1a666666666664],
93 //   [0x0 ~> 0x1a666666666664],
94 //   [0x0 ~> 0x1a666666666664]]
95 //
96 // See comments in curve25519_64.h for which functions use these bounds for
97 // inputs or outputs.
98 #define assert_fe_loose(f)                                              \
99   do {                                                                  \
100     for (unsigned _assert_fe_i = 0; _assert_fe_i < 5; _assert_fe_i++) { \
101       declassify_assert(f[_assert_fe_i] <= UINT64_C(0x1a666666666664)); \
102     }                                                                   \
103   } while (0)
104 
105 #else
106 
107 typedef uint32_t fe_limb_t;
108 #define FE_NUM_LIMBS 10
109 
110 // assert_fe asserts that |f| satisfies bounds:
111 //
112 //  [[0x0 ~> 0x4666666], [0x0 ~> 0x2333333],
113 //   [0x0 ~> 0x4666666], [0x0 ~> 0x2333333],
114 //   [0x0 ~> 0x4666666], [0x0 ~> 0x2333333],
115 //   [0x0 ~> 0x4666666], [0x0 ~> 0x2333333],
116 //   [0x0 ~> 0x4666666], [0x0 ~> 0x2333333]]
117 //
118 // See comments in curve25519_32.h for which functions use these bounds for
119 // inputs or outputs.
120 #define assert_fe(f)                                                     \
121   do {                                                                   \
122     for (unsigned _assert_fe_i = 0; _assert_fe_i < 10; _assert_fe_i++) { \
123       declassify_assert(f[_assert_fe_i] <=                               \
124                         ((_assert_fe_i & 1) ? 0x2333333u : 0x4666666u)); \
125     }                                                                    \
126   } while (0)
127 
128 // assert_fe_loose asserts that |f| satisfies bounds:
129 //
130 //  [[0x0 ~> 0xd333332], [0x0 ~> 0x6999999],
131 //   [0x0 ~> 0xd333332], [0x0 ~> 0x6999999],
132 //   [0x0 ~> 0xd333332], [0x0 ~> 0x6999999],
133 //   [0x0 ~> 0xd333332], [0x0 ~> 0x6999999],
134 //   [0x0 ~> 0xd333332], [0x0 ~> 0x6999999]]
135 //
136 // See comments in curve25519_32.h for which functions use these bounds for
137 // inputs or outputs.
138 #define assert_fe_loose(f)                                               \
139   do {                                                                   \
140     for (unsigned _assert_fe_i = 0; _assert_fe_i < 10; _assert_fe_i++) { \
141       declassify_assert(f[_assert_fe_i] <=                               \
142                         ((_assert_fe_i & 1) ? 0x6999999u : 0xd333332u)); \
143     }                                                                    \
144   } while (0)
145 
146 #endif  // OPENSSL_64_BIT
147 
148 static_assert(sizeof(fe) == sizeof(fe_limb_t) * FE_NUM_LIMBS,
149               "fe_limb_t[FE_NUM_LIMBS] is inconsistent with fe");
150 
fe_frombytes_strict(fe * h,const uint8_t s[32])151 static void fe_frombytes_strict(fe *h, const uint8_t s[32]) {
152   // |fiat_25519_from_bytes| requires the top-most bit be clear.
153   declassify_assert((s[31] & 0x80) == 0);
154   fiat_25519_from_bytes(h->v, s);
155   assert_fe(h->v);
156 }
157 
fe_frombytes(fe * h,const uint8_t s[32])158 static void fe_frombytes(fe *h, const uint8_t s[32]) {
159   uint8_t s_copy[32];
160   OPENSSL_memcpy(s_copy, s, 32);
161   s_copy[31] &= 0x7f;
162   fe_frombytes_strict(h, s_copy);
163 }
164 
fe_tobytes(uint8_t s[32],const fe * f)165 static void fe_tobytes(uint8_t s[32], const fe *f) {
166   assert_fe(f->v);
167   fiat_25519_to_bytes(s, f->v);
168 }
169 
170 // h = 0
fe_0(fe * h)171 static void fe_0(fe *h) {
172   OPENSSL_memset(h, 0, sizeof(fe));
173 }
174 
fe_loose_0(fe_loose * h)175 static void fe_loose_0(fe_loose *h) {
176   OPENSSL_memset(h, 0, sizeof(fe_loose));
177 }
178 
179 // h = 1
fe_1(fe * h)180 static void fe_1(fe *h) {
181   OPENSSL_memset(h, 0, sizeof(fe));
182   h->v[0] = 1;
183 }
184 
fe_loose_1(fe_loose * h)185 static void fe_loose_1(fe_loose *h) {
186   OPENSSL_memset(h, 0, sizeof(fe_loose));
187   h->v[0] = 1;
188 }
189 
190 // h = f + g
191 // Can overlap h with f or g.
fe_add(fe_loose * h,const fe * f,const fe * g)192 static void fe_add(fe_loose *h, const fe *f, const fe *g) {
193   assert_fe(f->v);
194   assert_fe(g->v);
195   fiat_25519_add(h->v, f->v, g->v);
196   assert_fe_loose(h->v);
197 }
198 
199 // h = f - g
200 // Can overlap h with f or g.
fe_sub(fe_loose * h,const fe * f,const fe * g)201 static void fe_sub(fe_loose *h, const fe *f, const fe *g) {
202   assert_fe(f->v);
203   assert_fe(g->v);
204   fiat_25519_sub(h->v, f->v, g->v);
205   assert_fe_loose(h->v);
206 }
207 
fe_carry(fe * h,const fe_loose * f)208 static void fe_carry(fe *h, const fe_loose* f) {
209   assert_fe_loose(f->v);
210   fiat_25519_carry(h->v, f->v);
211   assert_fe(h->v);
212 }
213 
fe_mul_impl(fe_limb_t out[FE_NUM_LIMBS],const fe_limb_t in1[FE_NUM_LIMBS],const fe_limb_t in2[FE_NUM_LIMBS])214 static void fe_mul_impl(fe_limb_t out[FE_NUM_LIMBS],
215                         const fe_limb_t in1[FE_NUM_LIMBS],
216                         const fe_limb_t in2[FE_NUM_LIMBS]) {
217   assert_fe_loose(in1);
218   assert_fe_loose(in2);
219   fiat_25519_carry_mul(out, in1, in2);
220   assert_fe(out);
221 }
222 
fe_mul_ltt(fe_loose * h,const fe * f,const fe * g)223 static void fe_mul_ltt(fe_loose *h, const fe *f, const fe *g) {
224   fe_mul_impl(h->v, f->v, g->v);
225 }
226 
fe_mul_llt(fe_loose * h,const fe_loose * f,const fe * g)227 static void fe_mul_llt(fe_loose *h, const fe_loose *f, const fe *g) {
228   fe_mul_impl(h->v, f->v, g->v);
229 }
230 
fe_mul_ttt(fe * h,const fe * f,const fe * g)231 static void fe_mul_ttt(fe *h, const fe *f, const fe *g) {
232   fe_mul_impl(h->v, f->v, g->v);
233 }
234 
fe_mul_tlt(fe * h,const fe_loose * f,const fe * g)235 static void fe_mul_tlt(fe *h, const fe_loose *f, const fe *g) {
236   fe_mul_impl(h->v, f->v, g->v);
237 }
238 
fe_mul_ttl(fe * h,const fe * f,const fe_loose * g)239 static void fe_mul_ttl(fe *h, const fe *f, const fe_loose *g) {
240   fe_mul_impl(h->v, f->v, g->v);
241 }
242 
fe_mul_tll(fe * h,const fe_loose * f,const fe_loose * g)243 static void fe_mul_tll(fe *h, const fe_loose *f, const fe_loose *g) {
244   fe_mul_impl(h->v, f->v, g->v);
245 }
246 
fe_sq_tl(fe * h,const fe_loose * f)247 static void fe_sq_tl(fe *h, const fe_loose *f) {
248   assert_fe_loose(f->v);
249   fiat_25519_carry_square(h->v, f->v);
250   assert_fe(h->v);
251 }
252 
fe_sq_tt(fe * h,const fe * f)253 static void fe_sq_tt(fe *h, const fe *f) {
254   assert_fe_loose(f->v);
255   fiat_25519_carry_square(h->v, f->v);
256   assert_fe(h->v);
257 }
258 
259 // Replace (f,g) with (g,f) if b == 1;
260 // replace (f,g) with (f,g) if b == 0.
261 //
262 // Preconditions: b in {0,1}.
fe_cswap(fe * f,fe * g,fe_limb_t b)263 static void fe_cswap(fe *f, fe *g, fe_limb_t b) {
264   b = 0-b;
265   for (unsigned i = 0; i < FE_NUM_LIMBS; i++) {
266     fe_limb_t x = f->v[i] ^ g->v[i];
267     x &= b;
268     f->v[i] ^= x;
269     g->v[i] ^= x;
270   }
271 }
272 
fe_mul121666(fe * h,const fe_loose * f)273 static void fe_mul121666(fe *h, const fe_loose *f) {
274   assert_fe_loose(f->v);
275   fiat_25519_carry_scmul_121666(h->v, f->v);
276   assert_fe(h->v);
277 }
278 
279 // h = -f
fe_neg(fe_loose * h,const fe * f)280 static void fe_neg(fe_loose *h, const fe *f) {
281   assert_fe(f->v);
282   fiat_25519_opp(h->v, f->v);
283   assert_fe_loose(h->v);
284 }
285 
286 // Replace (f,g) with (g,g) if b == 1;
287 // replace (f,g) with (f,g) if b == 0.
288 //
289 // Preconditions: b in {0,1}.
fe_cmov(fe_loose * f,const fe_loose * g,fe_limb_t b)290 static void fe_cmov(fe_loose *f, const fe_loose *g, fe_limb_t b) {
291   // Silence an unused function warning. |fiat_25519_selectznz| isn't quite the
292   // calling convention the rest of this code wants, so implement it by hand.
293   //
294   // TODO(davidben): Switch to fiat's calling convention, or ask fiat to emit a
295   // different one.
296   (void)fiat_25519_selectznz;
297 
298   b = 0-b;
299   for (unsigned i = 0; i < FE_NUM_LIMBS; i++) {
300     fe_limb_t x = f->v[i] ^ g->v[i];
301     x &= b;
302     f->v[i] ^= x;
303   }
304 }
305 
306 // h = f
fe_copy(fe * h,const fe * f)307 static void fe_copy(fe *h, const fe *f) {
308   OPENSSL_memmove(h, f, sizeof(fe));
309 }
310 
fe_copy_lt(fe_loose * h,const fe * f)311 static void fe_copy_lt(fe_loose *h, const fe *f) {
312   static_assert(sizeof(fe_loose) == sizeof(fe), "fe and fe_loose mismatch");
313   OPENSSL_memmove(h, f, sizeof(fe));
314 }
315 
fe_loose_invert(fe * out,const fe_loose * z)316 static void fe_loose_invert(fe *out, const fe_loose *z) {
317   fe t0;
318   fe t1;
319   fe t2;
320   fe t3;
321   int i;
322 
323   fe_sq_tl(&t0, z);
324   fe_sq_tt(&t1, &t0);
325   for (i = 1; i < 2; ++i) {
326     fe_sq_tt(&t1, &t1);
327   }
328   fe_mul_tlt(&t1, z, &t1);
329   fe_mul_ttt(&t0, &t0, &t1);
330   fe_sq_tt(&t2, &t0);
331   fe_mul_ttt(&t1, &t1, &t2);
332   fe_sq_tt(&t2, &t1);
333   for (i = 1; i < 5; ++i) {
334     fe_sq_tt(&t2, &t2);
335   }
336   fe_mul_ttt(&t1, &t2, &t1);
337   fe_sq_tt(&t2, &t1);
338   for (i = 1; i < 10; ++i) {
339     fe_sq_tt(&t2, &t2);
340   }
341   fe_mul_ttt(&t2, &t2, &t1);
342   fe_sq_tt(&t3, &t2);
343   for (i = 1; i < 20; ++i) {
344     fe_sq_tt(&t3, &t3);
345   }
346   fe_mul_ttt(&t2, &t3, &t2);
347   fe_sq_tt(&t2, &t2);
348   for (i = 1; i < 10; ++i) {
349     fe_sq_tt(&t2, &t2);
350   }
351   fe_mul_ttt(&t1, &t2, &t1);
352   fe_sq_tt(&t2, &t1);
353   for (i = 1; i < 50; ++i) {
354     fe_sq_tt(&t2, &t2);
355   }
356   fe_mul_ttt(&t2, &t2, &t1);
357   fe_sq_tt(&t3, &t2);
358   for (i = 1; i < 100; ++i) {
359     fe_sq_tt(&t3, &t3);
360   }
361   fe_mul_ttt(&t2, &t3, &t2);
362   fe_sq_tt(&t2, &t2);
363   for (i = 1; i < 50; ++i) {
364     fe_sq_tt(&t2, &t2);
365   }
366   fe_mul_ttt(&t1, &t2, &t1);
367   fe_sq_tt(&t1, &t1);
368   for (i = 1; i < 5; ++i) {
369     fe_sq_tt(&t1, &t1);
370   }
371   fe_mul_ttt(out, &t1, &t0);
372 }
373 
fe_invert(fe * out,const fe * z)374 static void fe_invert(fe *out, const fe *z) {
375   fe_loose l;
376   fe_copy_lt(&l, z);
377   fe_loose_invert(out, &l);
378 }
379 
380 // return 0 if f == 0
381 // return 1 if f != 0
fe_isnonzero(const fe_loose * f)382 static int fe_isnonzero(const fe_loose *f) {
383   fe tight;
384   fe_carry(&tight, f);
385   uint8_t s[32];
386   fe_tobytes(s, &tight);
387 
388   static const uint8_t zero[32] = {0};
389   return CRYPTO_memcmp(s, zero, sizeof(zero)) != 0;
390 }
391 
392 // return 1 if f is in {1,3,5,...,q-2}
393 // return 0 if f is in {0,2,4,...,q-1}
fe_isnegative(const fe * f)394 static int fe_isnegative(const fe *f) {
395   uint8_t s[32];
396   fe_tobytes(s, f);
397   return s[0] & 1;
398 }
399 
fe_sq2_tt(fe * h,const fe * f)400 static void fe_sq2_tt(fe *h, const fe *f) {
401   // h = f^2
402   fe_sq_tt(h, f);
403 
404   // h = h + h
405   fe_loose tmp;
406   fe_add(&tmp, h, h);
407   fe_carry(h, &tmp);
408 }
409 
fe_pow22523(fe * out,const fe * z)410 static void fe_pow22523(fe *out, const fe *z) {
411   fe t0;
412   fe t1;
413   fe t2;
414   int i;
415 
416   fe_sq_tt(&t0, z);
417   fe_sq_tt(&t1, &t0);
418   for (i = 1; i < 2; ++i) {
419     fe_sq_tt(&t1, &t1);
420   }
421   fe_mul_ttt(&t1, z, &t1);
422   fe_mul_ttt(&t0, &t0, &t1);
423   fe_sq_tt(&t0, &t0);
424   fe_mul_ttt(&t0, &t1, &t0);
425   fe_sq_tt(&t1, &t0);
426   for (i = 1; i < 5; ++i) {
427     fe_sq_tt(&t1, &t1);
428   }
429   fe_mul_ttt(&t0, &t1, &t0);
430   fe_sq_tt(&t1, &t0);
431   for (i = 1; i < 10; ++i) {
432     fe_sq_tt(&t1, &t1);
433   }
434   fe_mul_ttt(&t1, &t1, &t0);
435   fe_sq_tt(&t2, &t1);
436   for (i = 1; i < 20; ++i) {
437     fe_sq_tt(&t2, &t2);
438   }
439   fe_mul_ttt(&t1, &t2, &t1);
440   fe_sq_tt(&t1, &t1);
441   for (i = 1; i < 10; ++i) {
442     fe_sq_tt(&t1, &t1);
443   }
444   fe_mul_ttt(&t0, &t1, &t0);
445   fe_sq_tt(&t1, &t0);
446   for (i = 1; i < 50; ++i) {
447     fe_sq_tt(&t1, &t1);
448   }
449   fe_mul_ttt(&t1, &t1, &t0);
450   fe_sq_tt(&t2, &t1);
451   for (i = 1; i < 100; ++i) {
452     fe_sq_tt(&t2, &t2);
453   }
454   fe_mul_ttt(&t1, &t2, &t1);
455   fe_sq_tt(&t1, &t1);
456   for (i = 1; i < 50; ++i) {
457     fe_sq_tt(&t1, &t1);
458   }
459   fe_mul_ttt(&t0, &t1, &t0);
460   fe_sq_tt(&t0, &t0);
461   for (i = 1; i < 2; ++i) {
462     fe_sq_tt(&t0, &t0);
463   }
464   fe_mul_ttt(out, &t0, z);
465 }
466 
467 
468 // Group operations.
469 
x25519_ge_tobytes(uint8_t s[32],const ge_p2 * h)470 void x25519_ge_tobytes(uint8_t s[32], const ge_p2 *h) {
471   fe recip;
472   fe x;
473   fe y;
474 
475   fe_invert(&recip, &h->Z);
476   fe_mul_ttt(&x, &h->X, &recip);
477   fe_mul_ttt(&y, &h->Y, &recip);
478   fe_tobytes(s, &y);
479   s[31] ^= fe_isnegative(&x) << 7;
480 }
481 
ge_p3_tobytes(uint8_t s[32],const ge_p3 * h)482 static void ge_p3_tobytes(uint8_t s[32], const ge_p3 *h) {
483   fe recip;
484   fe x;
485   fe y;
486 
487   fe_invert(&recip, &h->Z);
488   fe_mul_ttt(&x, &h->X, &recip);
489   fe_mul_ttt(&y, &h->Y, &recip);
490   fe_tobytes(s, &y);
491   s[31] ^= fe_isnegative(&x) << 7;
492 }
493 
x25519_ge_frombytes_vartime(ge_p3 * h,const uint8_t s[32])494 int x25519_ge_frombytes_vartime(ge_p3 *h, const uint8_t s[32]) {
495   fe u;
496   fe_loose v;
497   fe w;
498   fe vxx;
499   fe_loose check;
500 
501   fe_frombytes(&h->Y, s);
502   fe_1(&h->Z);
503   fe_sq_tt(&w, &h->Y);
504   fe_mul_ttt(&vxx, &w, &d);
505   fe_sub(&v, &w, &h->Z);  // u = y^2-1
506   fe_carry(&u, &v);
507   fe_add(&v, &vxx, &h->Z);  // v = dy^2+1
508 
509   fe_mul_ttl(&w, &u, &v);  // w = u*v
510   fe_pow22523(&h->X, &w);  // x = w^((q-5)/8)
511   fe_mul_ttt(&h->X, &h->X, &u);  // x = u*w^((q-5)/8)
512 
513   fe_sq_tt(&vxx, &h->X);
514   fe_mul_ttl(&vxx, &vxx, &v);
515   fe_sub(&check, &vxx, &u);
516   if (fe_isnonzero(&check)) {
517     fe_add(&check, &vxx, &u);
518     if (fe_isnonzero(&check)) {
519       return 0;
520     }
521     fe_mul_ttt(&h->X, &h->X, &sqrtm1);
522   }
523 
524   if (fe_isnegative(&h->X) != (s[31] >> 7)) {
525     fe_loose t;
526     fe_neg(&t, &h->X);
527     fe_carry(&h->X, &t);
528   }
529 
530   fe_mul_ttt(&h->T, &h->X, &h->Y);
531   return 1;
532 }
533 
ge_p2_0(ge_p2 * h)534 static void ge_p2_0(ge_p2 *h) {
535   fe_0(&h->X);
536   fe_1(&h->Y);
537   fe_1(&h->Z);
538 }
539 
ge_p3_0(ge_p3 * h)540 static void ge_p3_0(ge_p3 *h) {
541   fe_0(&h->X);
542   fe_1(&h->Y);
543   fe_1(&h->Z);
544   fe_0(&h->T);
545 }
546 
ge_cached_0(ge_cached * h)547 static void ge_cached_0(ge_cached *h) {
548   fe_loose_1(&h->YplusX);
549   fe_loose_1(&h->YminusX);
550   fe_loose_1(&h->Z);
551   fe_loose_0(&h->T2d);
552 }
553 
ge_precomp_0(ge_precomp * h)554 static void ge_precomp_0(ge_precomp *h) {
555   fe_loose_1(&h->yplusx);
556   fe_loose_1(&h->yminusx);
557   fe_loose_0(&h->xy2d);
558 }
559 
560 // r = p
ge_p3_to_p2(ge_p2 * r,const ge_p3 * p)561 static void ge_p3_to_p2(ge_p2 *r, const ge_p3 *p) {
562   fe_copy(&r->X, &p->X);
563   fe_copy(&r->Y, &p->Y);
564   fe_copy(&r->Z, &p->Z);
565 }
566 
567 // r = p
x25519_ge_p3_to_cached(ge_cached * r,const ge_p3 * p)568 void x25519_ge_p3_to_cached(ge_cached *r, const ge_p3 *p) {
569   fe_add(&r->YplusX, &p->Y, &p->X);
570   fe_sub(&r->YminusX, &p->Y, &p->X);
571   fe_copy_lt(&r->Z, &p->Z);
572   fe_mul_ltt(&r->T2d, &p->T, &d2);
573 }
574 
575 // r = p
x25519_ge_p1p1_to_p2(ge_p2 * r,const ge_p1p1 * p)576 void x25519_ge_p1p1_to_p2(ge_p2 *r, const ge_p1p1 *p) {
577   fe_mul_tll(&r->X, &p->X, &p->T);
578   fe_mul_tll(&r->Y, &p->Y, &p->Z);
579   fe_mul_tll(&r->Z, &p->Z, &p->T);
580 }
581 
582 // r = p
x25519_ge_p1p1_to_p3(ge_p3 * r,const ge_p1p1 * p)583 void x25519_ge_p1p1_to_p3(ge_p3 *r, const ge_p1p1 *p) {
584   fe_mul_tll(&r->X, &p->X, &p->T);
585   fe_mul_tll(&r->Y, &p->Y, &p->Z);
586   fe_mul_tll(&r->Z, &p->Z, &p->T);
587   fe_mul_tll(&r->T, &p->X, &p->Y);
588 }
589 
590 // r = p
ge_p1p1_to_cached(ge_cached * r,const ge_p1p1 * p)591 static void ge_p1p1_to_cached(ge_cached *r, const ge_p1p1 *p) {
592   ge_p3 t;
593   x25519_ge_p1p1_to_p3(&t, p);
594   x25519_ge_p3_to_cached(r, &t);
595 }
596 
597 // r = 2 * p
ge_p2_dbl(ge_p1p1 * r,const ge_p2 * p)598 static void ge_p2_dbl(ge_p1p1 *r, const ge_p2 *p) {
599   fe trX, trZ, trT;
600   fe t0;
601 
602   fe_sq_tt(&trX, &p->X);
603   fe_sq_tt(&trZ, &p->Y);
604   fe_sq2_tt(&trT, &p->Z);
605   fe_add(&r->Y, &p->X, &p->Y);
606   fe_sq_tl(&t0, &r->Y);
607 
608   fe_add(&r->Y, &trZ, &trX);
609   fe_sub(&r->Z, &trZ, &trX);
610   fe_carry(&trZ, &r->Y);
611   fe_sub(&r->X, &t0, &trZ);
612   fe_carry(&trZ, &r->Z);
613   fe_sub(&r->T, &trT, &trZ);
614 }
615 
616 // r = 2 * p
ge_p3_dbl(ge_p1p1 * r,const ge_p3 * p)617 static void ge_p3_dbl(ge_p1p1 *r, const ge_p3 *p) {
618   ge_p2 q;
619   ge_p3_to_p2(&q, p);
620   ge_p2_dbl(r, &q);
621 }
622 
623 // r = p + q
ge_madd(ge_p1p1 * r,const ge_p3 * p,const ge_precomp * q)624 static void ge_madd(ge_p1p1 *r, const ge_p3 *p, const ge_precomp *q) {
625   fe trY, trZ, trT;
626 
627   fe_add(&r->X, &p->Y, &p->X);
628   fe_sub(&r->Y, &p->Y, &p->X);
629   fe_mul_tll(&trZ, &r->X, &q->yplusx);
630   fe_mul_tll(&trY, &r->Y, &q->yminusx);
631   fe_mul_tlt(&trT, &q->xy2d, &p->T);
632   fe_add(&r->T, &p->Z, &p->Z);
633   fe_sub(&r->X, &trZ, &trY);
634   fe_add(&r->Y, &trZ, &trY);
635   fe_carry(&trZ, &r->T);
636   fe_add(&r->Z, &trZ, &trT);
637   fe_sub(&r->T, &trZ, &trT);
638 }
639 
640 // r = p - q
ge_msub(ge_p1p1 * r,const ge_p3 * p,const ge_precomp * q)641 static void ge_msub(ge_p1p1 *r, const ge_p3 *p, const ge_precomp *q) {
642   fe trY, trZ, trT;
643 
644   fe_add(&r->X, &p->Y, &p->X);
645   fe_sub(&r->Y, &p->Y, &p->X);
646   fe_mul_tll(&trZ, &r->X, &q->yminusx);
647   fe_mul_tll(&trY, &r->Y, &q->yplusx);
648   fe_mul_tlt(&trT, &q->xy2d, &p->T);
649   fe_add(&r->T, &p->Z, &p->Z);
650   fe_sub(&r->X, &trZ, &trY);
651   fe_add(&r->Y, &trZ, &trY);
652   fe_carry(&trZ, &r->T);
653   fe_sub(&r->Z, &trZ, &trT);
654   fe_add(&r->T, &trZ, &trT);
655 }
656 
657 // r = p + q
x25519_ge_add(ge_p1p1 * r,const ge_p3 * p,const ge_cached * q)658 void x25519_ge_add(ge_p1p1 *r, const ge_p3 *p, const ge_cached *q) {
659   fe trX, trY, trZ, trT;
660 
661   fe_add(&r->X, &p->Y, &p->X);
662   fe_sub(&r->Y, &p->Y, &p->X);
663   fe_mul_tll(&trZ, &r->X, &q->YplusX);
664   fe_mul_tll(&trY, &r->Y, &q->YminusX);
665   fe_mul_tlt(&trT, &q->T2d, &p->T);
666   fe_mul_ttl(&trX, &p->Z, &q->Z);
667   fe_add(&r->T, &trX, &trX);
668   fe_sub(&r->X, &trZ, &trY);
669   fe_add(&r->Y, &trZ, &trY);
670   fe_carry(&trZ, &r->T);
671   fe_add(&r->Z, &trZ, &trT);
672   fe_sub(&r->T, &trZ, &trT);
673 }
674 
675 // r = p - q
x25519_ge_sub(ge_p1p1 * r,const ge_p3 * p,const ge_cached * q)676 void x25519_ge_sub(ge_p1p1 *r, const ge_p3 *p, const ge_cached *q) {
677   fe trX, trY, trZ, trT;
678 
679   fe_add(&r->X, &p->Y, &p->X);
680   fe_sub(&r->Y, &p->Y, &p->X);
681   fe_mul_tll(&trZ, &r->X, &q->YminusX);
682   fe_mul_tll(&trY, &r->Y, &q->YplusX);
683   fe_mul_tlt(&trT, &q->T2d, &p->T);
684   fe_mul_ttl(&trX, &p->Z, &q->Z);
685   fe_add(&r->T, &trX, &trX);
686   fe_sub(&r->X, &trZ, &trY);
687   fe_add(&r->Y, &trZ, &trY);
688   fe_carry(&trZ, &r->T);
689   fe_sub(&r->Z, &trZ, &trT);
690   fe_add(&r->T, &trZ, &trT);
691 }
692 
cmov(ge_precomp * t,const ge_precomp * u,uint8_t b)693 static void cmov(ge_precomp *t, const ge_precomp *u, uint8_t b) {
694   fe_cmov(&t->yplusx, &u->yplusx, b);
695   fe_cmov(&t->yminusx, &u->yminusx, b);
696   fe_cmov(&t->xy2d, &u->xy2d, b);
697 }
698 
x25519_ge_scalarmult_small_precomp(ge_p3 * h,const uint8_t a[32],const uint8_t precomp_table[15* 2* 32])699 void x25519_ge_scalarmult_small_precomp(
700     ge_p3 *h, const uint8_t a[32], const uint8_t precomp_table[15 * 2 * 32]) {
701   // precomp_table is first expanded into matching |ge_precomp|
702   // elements.
703   ge_precomp multiples[15];
704 
705   unsigned i;
706   for (i = 0; i < 15; i++) {
707     // The precomputed table is assumed to already clear the top bit, so
708     // |fe_frombytes_strict| may be used directly.
709     const uint8_t *bytes = &precomp_table[i*(2 * 32)];
710     fe x, y;
711     fe_frombytes_strict(&x, bytes);
712     fe_frombytes_strict(&y, bytes + 32);
713 
714     ge_precomp *out = &multiples[i];
715     fe_add(&out->yplusx, &y, &x);
716     fe_sub(&out->yminusx, &y, &x);
717     fe_mul_ltt(&out->xy2d, &x, &y);
718     fe_mul_llt(&out->xy2d, &out->xy2d, &d2);
719   }
720 
721   // See the comment above |k25519SmallPrecomp| about the structure of the
722   // precomputed elements. This loop does 64 additions and 64 doublings to
723   // calculate the result.
724   ge_p3_0(h);
725 
726   for (i = 63; i < 64; i--) {
727     unsigned j;
728     signed char index = 0;
729 
730     for (j = 0; j < 4; j++) {
731       const uint8_t bit = 1 & (a[(8 * j) + (i / 8)] >> (i & 7));
732       index |= (bit << j);
733     }
734 
735     ge_precomp e;
736     ge_precomp_0(&e);
737 
738     for (j = 1; j < 16; j++) {
739       cmov(&e, &multiples[j-1], 1&constant_time_eq_w(index, j));
740     }
741 
742     ge_cached cached;
743     ge_p1p1 r;
744     x25519_ge_p3_to_cached(&cached, h);
745     x25519_ge_add(&r, h, &cached);
746     x25519_ge_p1p1_to_p3(h, &r);
747 
748     ge_madd(&r, h, &e);
749     x25519_ge_p1p1_to_p3(h, &r);
750   }
751 }
752 
753 #if defined(OPENSSL_SMALL)
754 
x25519_ge_scalarmult_base(ge_p3 * h,const uint8_t a[32])755 void x25519_ge_scalarmult_base(ge_p3 *h, const uint8_t a[32]) {
756   x25519_ge_scalarmult_small_precomp(h, a, k25519SmallPrecomp);
757 }
758 
759 #else
760 
table_select(ge_precomp * t,const int pos,const signed char b)761 static void table_select(ge_precomp *t, const int pos, const signed char b) {
762   uint8_t bnegative = constant_time_msb_w(b);
763   uint8_t babs = b - ((bnegative & b) << 1);
764 
765   uint8_t t_bytes[3][32] = {
766       {constant_time_is_zero_w(b) & 1}, {constant_time_is_zero_w(b) & 1}, {0}};
767 #if defined(__clang__) // materialize for vectorization, 6% speedup
768   __asm__("" : "+m" (t_bytes) : /*no inputs*/);
769 #endif
770   static_assert(sizeof(t_bytes) == sizeof(k25519Precomp[pos][0]), "");
771   for (int i = 0; i < 8; i++) {
772     constant_time_conditional_memxor(t_bytes, k25519Precomp[pos][i],
773                                      sizeof(t_bytes),
774                                      constant_time_eq_w(babs, 1 + i));
775   }
776 
777   fe yplusx, yminusx, xy2d;
778   fe_frombytes_strict(&yplusx, t_bytes[0]);
779   fe_frombytes_strict(&yminusx, t_bytes[1]);
780   fe_frombytes_strict(&xy2d, t_bytes[2]);
781 
782   fe_copy_lt(&t->yplusx, &yplusx);
783   fe_copy_lt(&t->yminusx, &yminusx);
784   fe_copy_lt(&t->xy2d, &xy2d);
785 
786   ge_precomp minust;
787   fe_copy_lt(&minust.yplusx, &yminusx);
788   fe_copy_lt(&minust.yminusx, &yplusx);
789   fe_neg(&minust.xy2d, &xy2d);
790   cmov(t, &minust, bnegative>>7);
791 }
792 
793 // h = a * B
794 // where a = a[0]+256*a[1]+...+256^31 a[31]
795 // B is the Ed25519 base point (x,4/5) with x positive.
796 //
797 // Preconditions:
798 //   a[31] <= 127
x25519_ge_scalarmult_base(ge_p3 * h,const uint8_t a[32])799 void x25519_ge_scalarmult_base(ge_p3 *h, const uint8_t a[32]) {
800 #if defined(BORINGSSL_FE25519_ADX)
801   if (CRYPTO_is_BMI1_capable() && CRYPTO_is_BMI2_capable() &&
802       CRYPTO_is_ADX_capable()) {
803     uint8_t t[4][32];
804     x25519_ge_scalarmult_base_adx(t, a);
805     fiat_25519_from_bytes(h->X.v, t[0]);
806     fiat_25519_from_bytes(h->Y.v, t[1]);
807     fiat_25519_from_bytes(h->Z.v, t[2]);
808     fiat_25519_from_bytes(h->T.v, t[3]);
809     return;
810   }
811 #endif
812   signed char e[64];
813   signed char carry;
814   ge_p1p1 r;
815   ge_p2 s;
816   ge_precomp t;
817   int i;
818 
819   for (i = 0; i < 32; ++i) {
820     e[2 * i + 0] = (a[i] >> 0) & 15;
821     e[2 * i + 1] = (a[i] >> 4) & 15;
822   }
823   // each e[i] is between 0 and 15
824   // e[63] is between 0 and 7
825 
826   carry = 0;
827   for (i = 0; i < 63; ++i) {
828     e[i] += carry;
829     carry = e[i] + 8;
830     carry >>= 4;
831     e[i] -= carry << 4;
832   }
833   e[63] += carry;
834   // each e[i] is between -8 and 8
835 
836   ge_p3_0(h);
837   for (i = 1; i < 64; i += 2) {
838     table_select(&t, i / 2, e[i]);
839     ge_madd(&r, h, &t);
840     x25519_ge_p1p1_to_p3(h, &r);
841   }
842 
843   ge_p3_dbl(&r, h);
844   x25519_ge_p1p1_to_p2(&s, &r);
845   ge_p2_dbl(&r, &s);
846   x25519_ge_p1p1_to_p2(&s, &r);
847   ge_p2_dbl(&r, &s);
848   x25519_ge_p1p1_to_p2(&s, &r);
849   ge_p2_dbl(&r, &s);
850   x25519_ge_p1p1_to_p3(h, &r);
851 
852   for (i = 0; i < 64; i += 2) {
853     table_select(&t, i / 2, e[i]);
854     ge_madd(&r, h, &t);
855     x25519_ge_p1p1_to_p3(h, &r);
856   }
857 }
858 
859 #endif
860 
cmov_cached(ge_cached * t,ge_cached * u,uint8_t b)861 static void cmov_cached(ge_cached *t, ge_cached *u, uint8_t b) {
862   fe_cmov(&t->YplusX, &u->YplusX, b);
863   fe_cmov(&t->YminusX, &u->YminusX, b);
864   fe_cmov(&t->Z, &u->Z, b);
865   fe_cmov(&t->T2d, &u->T2d, b);
866 }
867 
868 // r = scalar * A.
869 // where a = a[0]+256*a[1]+...+256^31 a[31].
x25519_ge_scalarmult(ge_p2 * r,const uint8_t * scalar,const ge_p3 * A)870 void x25519_ge_scalarmult(ge_p2 *r, const uint8_t *scalar, const ge_p3 *A) {
871   ge_p2 Ai_p2[8];
872   ge_cached Ai[16];
873   ge_p1p1 t;
874 
875   ge_cached_0(&Ai[0]);
876   x25519_ge_p3_to_cached(&Ai[1], A);
877   ge_p3_to_p2(&Ai_p2[1], A);
878 
879   unsigned i;
880   for (i = 2; i < 16; i += 2) {
881     ge_p2_dbl(&t, &Ai_p2[i / 2]);
882     ge_p1p1_to_cached(&Ai[i], &t);
883     if (i < 8) {
884       x25519_ge_p1p1_to_p2(&Ai_p2[i], &t);
885     }
886     x25519_ge_add(&t, A, &Ai[i]);
887     ge_p1p1_to_cached(&Ai[i + 1], &t);
888     if (i < 7) {
889       x25519_ge_p1p1_to_p2(&Ai_p2[i + 1], &t);
890     }
891   }
892 
893   ge_p2_0(r);
894   ge_p3 u;
895 
896   for (i = 0; i < 256; i += 4) {
897     ge_p2_dbl(&t, r);
898     x25519_ge_p1p1_to_p2(r, &t);
899     ge_p2_dbl(&t, r);
900     x25519_ge_p1p1_to_p2(r, &t);
901     ge_p2_dbl(&t, r);
902     x25519_ge_p1p1_to_p2(r, &t);
903     ge_p2_dbl(&t, r);
904     x25519_ge_p1p1_to_p3(&u, &t);
905 
906     uint8_t index = scalar[31 - i/8];
907     index >>= 4 - (i & 4);
908     index &= 0xf;
909 
910     unsigned j;
911     ge_cached selected;
912     ge_cached_0(&selected);
913     for (j = 0; j < 16; j++) {
914       cmov_cached(&selected, &Ai[j], 1&constant_time_eq_w(index, j));
915     }
916 
917     x25519_ge_add(&t, &u, &selected);
918     x25519_ge_p1p1_to_p2(r, &t);
919   }
920 }
921 
slide(signed char * r,const uint8_t * a)922 static void slide(signed char *r, const uint8_t *a) {
923   int i;
924   int b;
925   int k;
926 
927   for (i = 0; i < 256; ++i) {
928     r[i] = 1 & (a[i >> 3] >> (i & 7));
929   }
930 
931   for (i = 0; i < 256; ++i) {
932     if (r[i]) {
933       for (b = 1; b <= 6 && i + b < 256; ++b) {
934         if (r[i + b]) {
935           if (r[i] + (r[i + b] << b) <= 15) {
936             r[i] += r[i + b] << b;
937             r[i + b] = 0;
938           } else if (r[i] - (r[i + b] << b) >= -15) {
939             r[i] -= r[i + b] << b;
940             for (k = i + b; k < 256; ++k) {
941               if (!r[k]) {
942                 r[k] = 1;
943                 break;
944               }
945               r[k] = 0;
946             }
947           } else {
948             break;
949           }
950         }
951       }
952     }
953   }
954 }
955 
956 // r = a * A + b * B
957 // where a = a[0]+256*a[1]+...+256^31 a[31].
958 // and b = b[0]+256*b[1]+...+256^31 b[31].
959 // B is the Ed25519 base point (x,4/5) with x positive.
ge_double_scalarmult_vartime(ge_p2 * r,const uint8_t * a,const ge_p3 * A,const uint8_t * b)960 static void ge_double_scalarmult_vartime(ge_p2 *r, const uint8_t *a,
961                                          const ge_p3 *A, const uint8_t *b) {
962   signed char aslide[256];
963   signed char bslide[256];
964   ge_cached Ai[8];  // A,3A,5A,7A,9A,11A,13A,15A
965   ge_p1p1 t;
966   ge_p3 u;
967   ge_p3 A2;
968   int i;
969 
970   slide(aslide, a);
971   slide(bslide, b);
972 
973   x25519_ge_p3_to_cached(&Ai[0], A);
974   ge_p3_dbl(&t, A);
975   x25519_ge_p1p1_to_p3(&A2, &t);
976   x25519_ge_add(&t, &A2, &Ai[0]);
977   x25519_ge_p1p1_to_p3(&u, &t);
978   x25519_ge_p3_to_cached(&Ai[1], &u);
979   x25519_ge_add(&t, &A2, &Ai[1]);
980   x25519_ge_p1p1_to_p3(&u, &t);
981   x25519_ge_p3_to_cached(&Ai[2], &u);
982   x25519_ge_add(&t, &A2, &Ai[2]);
983   x25519_ge_p1p1_to_p3(&u, &t);
984   x25519_ge_p3_to_cached(&Ai[3], &u);
985   x25519_ge_add(&t, &A2, &Ai[3]);
986   x25519_ge_p1p1_to_p3(&u, &t);
987   x25519_ge_p3_to_cached(&Ai[4], &u);
988   x25519_ge_add(&t, &A2, &Ai[4]);
989   x25519_ge_p1p1_to_p3(&u, &t);
990   x25519_ge_p3_to_cached(&Ai[5], &u);
991   x25519_ge_add(&t, &A2, &Ai[5]);
992   x25519_ge_p1p1_to_p3(&u, &t);
993   x25519_ge_p3_to_cached(&Ai[6], &u);
994   x25519_ge_add(&t, &A2, &Ai[6]);
995   x25519_ge_p1p1_to_p3(&u, &t);
996   x25519_ge_p3_to_cached(&Ai[7], &u);
997 
998   ge_p2_0(r);
999 
1000   for (i = 255; i >= 0; --i) {
1001     if (aslide[i] || bslide[i]) {
1002       break;
1003     }
1004   }
1005 
1006   for (; i >= 0; --i) {
1007     ge_p2_dbl(&t, r);
1008 
1009     if (aslide[i] > 0) {
1010       x25519_ge_p1p1_to_p3(&u, &t);
1011       x25519_ge_add(&t, &u, &Ai[aslide[i] / 2]);
1012     } else if (aslide[i] < 0) {
1013       x25519_ge_p1p1_to_p3(&u, &t);
1014       x25519_ge_sub(&t, &u, &Ai[(-aslide[i]) / 2]);
1015     }
1016 
1017     if (bslide[i] > 0) {
1018       x25519_ge_p1p1_to_p3(&u, &t);
1019       ge_madd(&t, &u, &Bi[bslide[i] / 2]);
1020     } else if (bslide[i] < 0) {
1021       x25519_ge_p1p1_to_p3(&u, &t);
1022       ge_msub(&t, &u, &Bi[(-bslide[i]) / 2]);
1023     }
1024 
1025     x25519_ge_p1p1_to_p2(r, &t);
1026   }
1027 }
1028 
1029 // int64_lshift21 returns |a << 21| but is defined when shifting bits into the
1030 // sign bit. This works around a language flaw in C.
int64_lshift21(int64_t a)1031 static inline int64_t int64_lshift21(int64_t a) {
1032   return (int64_t)((uint64_t)a << 21);
1033 }
1034 
1035 // The set of scalars is \Z/l
1036 // where l = 2^252 + 27742317777372353535851937790883648493.
1037 
1038 // Input:
1039 //   s[0]+256*s[1]+...+256^63*s[63] = s
1040 //
1041 // Output:
1042 //   s[0]+256*s[1]+...+256^31*s[31] = s mod l
1043 //   where l = 2^252 + 27742317777372353535851937790883648493.
1044 //   Overwrites s in place.
x25519_sc_reduce(uint8_t s[64])1045 void x25519_sc_reduce(uint8_t s[64]) {
1046   int64_t s0 = 2097151 & load_3(s);
1047   int64_t s1 = 2097151 & (load_4(s + 2) >> 5);
1048   int64_t s2 = 2097151 & (load_3(s + 5) >> 2);
1049   int64_t s3 = 2097151 & (load_4(s + 7) >> 7);
1050   int64_t s4 = 2097151 & (load_4(s + 10) >> 4);
1051   int64_t s5 = 2097151 & (load_3(s + 13) >> 1);
1052   int64_t s6 = 2097151 & (load_4(s + 15) >> 6);
1053   int64_t s7 = 2097151 & (load_3(s + 18) >> 3);
1054   int64_t s8 = 2097151 & load_3(s + 21);
1055   int64_t s9 = 2097151 & (load_4(s + 23) >> 5);
1056   int64_t s10 = 2097151 & (load_3(s + 26) >> 2);
1057   int64_t s11 = 2097151 & (load_4(s + 28) >> 7);
1058   int64_t s12 = 2097151 & (load_4(s + 31) >> 4);
1059   int64_t s13 = 2097151 & (load_3(s + 34) >> 1);
1060   int64_t s14 = 2097151 & (load_4(s + 36) >> 6);
1061   int64_t s15 = 2097151 & (load_3(s + 39) >> 3);
1062   int64_t s16 = 2097151 & load_3(s + 42);
1063   int64_t s17 = 2097151 & (load_4(s + 44) >> 5);
1064   int64_t s18 = 2097151 & (load_3(s + 47) >> 2);
1065   int64_t s19 = 2097151 & (load_4(s + 49) >> 7);
1066   int64_t s20 = 2097151 & (load_4(s + 52) >> 4);
1067   int64_t s21 = 2097151 & (load_3(s + 55) >> 1);
1068   int64_t s22 = 2097151 & (load_4(s + 57) >> 6);
1069   int64_t s23 = (load_4(s + 60) >> 3);
1070   int64_t carry0;
1071   int64_t carry1;
1072   int64_t carry2;
1073   int64_t carry3;
1074   int64_t carry4;
1075   int64_t carry5;
1076   int64_t carry6;
1077   int64_t carry7;
1078   int64_t carry8;
1079   int64_t carry9;
1080   int64_t carry10;
1081   int64_t carry11;
1082   int64_t carry12;
1083   int64_t carry13;
1084   int64_t carry14;
1085   int64_t carry15;
1086   int64_t carry16;
1087 
1088   s11 += s23 * 666643;
1089   s12 += s23 * 470296;
1090   s13 += s23 * 654183;
1091   s14 -= s23 * 997805;
1092   s15 += s23 * 136657;
1093   s16 -= s23 * 683901;
1094   s23 = 0;
1095 
1096   s10 += s22 * 666643;
1097   s11 += s22 * 470296;
1098   s12 += s22 * 654183;
1099   s13 -= s22 * 997805;
1100   s14 += s22 * 136657;
1101   s15 -= s22 * 683901;
1102   s22 = 0;
1103 
1104   s9 += s21 * 666643;
1105   s10 += s21 * 470296;
1106   s11 += s21 * 654183;
1107   s12 -= s21 * 997805;
1108   s13 += s21 * 136657;
1109   s14 -= s21 * 683901;
1110   s21 = 0;
1111 
1112   s8 += s20 * 666643;
1113   s9 += s20 * 470296;
1114   s10 += s20 * 654183;
1115   s11 -= s20 * 997805;
1116   s12 += s20 * 136657;
1117   s13 -= s20 * 683901;
1118   s20 = 0;
1119 
1120   s7 += s19 * 666643;
1121   s8 += s19 * 470296;
1122   s9 += s19 * 654183;
1123   s10 -= s19 * 997805;
1124   s11 += s19 * 136657;
1125   s12 -= s19 * 683901;
1126   s19 = 0;
1127 
1128   s6 += s18 * 666643;
1129   s7 += s18 * 470296;
1130   s8 += s18 * 654183;
1131   s9 -= s18 * 997805;
1132   s10 += s18 * 136657;
1133   s11 -= s18 * 683901;
1134   s18 = 0;
1135 
1136   carry6 = (s6 + (1 << 20)) >> 21;
1137   s7 += carry6;
1138   s6 -= int64_lshift21(carry6);
1139   carry8 = (s8 + (1 << 20)) >> 21;
1140   s9 += carry8;
1141   s8 -= int64_lshift21(carry8);
1142   carry10 = (s10 + (1 << 20)) >> 21;
1143   s11 += carry10;
1144   s10 -= int64_lshift21(carry10);
1145   carry12 = (s12 + (1 << 20)) >> 21;
1146   s13 += carry12;
1147   s12 -= int64_lshift21(carry12);
1148   carry14 = (s14 + (1 << 20)) >> 21;
1149   s15 += carry14;
1150   s14 -= int64_lshift21(carry14);
1151   carry16 = (s16 + (1 << 20)) >> 21;
1152   s17 += carry16;
1153   s16 -= int64_lshift21(carry16);
1154 
1155   carry7 = (s7 + (1 << 20)) >> 21;
1156   s8 += carry7;
1157   s7 -= int64_lshift21(carry7);
1158   carry9 = (s9 + (1 << 20)) >> 21;
1159   s10 += carry9;
1160   s9 -= int64_lshift21(carry9);
1161   carry11 = (s11 + (1 << 20)) >> 21;
1162   s12 += carry11;
1163   s11 -= int64_lshift21(carry11);
1164   carry13 = (s13 + (1 << 20)) >> 21;
1165   s14 += carry13;
1166   s13 -= int64_lshift21(carry13);
1167   carry15 = (s15 + (1 << 20)) >> 21;
1168   s16 += carry15;
1169   s15 -= int64_lshift21(carry15);
1170 
1171   s5 += s17 * 666643;
1172   s6 += s17 * 470296;
1173   s7 += s17 * 654183;
1174   s8 -= s17 * 997805;
1175   s9 += s17 * 136657;
1176   s10 -= s17 * 683901;
1177   s17 = 0;
1178 
1179   s4 += s16 * 666643;
1180   s5 += s16 * 470296;
1181   s6 += s16 * 654183;
1182   s7 -= s16 * 997805;
1183   s8 += s16 * 136657;
1184   s9 -= s16 * 683901;
1185   s16 = 0;
1186 
1187   s3 += s15 * 666643;
1188   s4 += s15 * 470296;
1189   s5 += s15 * 654183;
1190   s6 -= s15 * 997805;
1191   s7 += s15 * 136657;
1192   s8 -= s15 * 683901;
1193   s15 = 0;
1194 
1195   s2 += s14 * 666643;
1196   s3 += s14 * 470296;
1197   s4 += s14 * 654183;
1198   s5 -= s14 * 997805;
1199   s6 += s14 * 136657;
1200   s7 -= s14 * 683901;
1201   s14 = 0;
1202 
1203   s1 += s13 * 666643;
1204   s2 += s13 * 470296;
1205   s3 += s13 * 654183;
1206   s4 -= s13 * 997805;
1207   s5 += s13 * 136657;
1208   s6 -= s13 * 683901;
1209   s13 = 0;
1210 
1211   s0 += s12 * 666643;
1212   s1 += s12 * 470296;
1213   s2 += s12 * 654183;
1214   s3 -= s12 * 997805;
1215   s4 += s12 * 136657;
1216   s5 -= s12 * 683901;
1217   s12 = 0;
1218 
1219   carry0 = (s0 + (1 << 20)) >> 21;
1220   s1 += carry0;
1221   s0 -= int64_lshift21(carry0);
1222   carry2 = (s2 + (1 << 20)) >> 21;
1223   s3 += carry2;
1224   s2 -= int64_lshift21(carry2);
1225   carry4 = (s4 + (1 << 20)) >> 21;
1226   s5 += carry4;
1227   s4 -= int64_lshift21(carry4);
1228   carry6 = (s6 + (1 << 20)) >> 21;
1229   s7 += carry6;
1230   s6 -= int64_lshift21(carry6);
1231   carry8 = (s8 + (1 << 20)) >> 21;
1232   s9 += carry8;
1233   s8 -= int64_lshift21(carry8);
1234   carry10 = (s10 + (1 << 20)) >> 21;
1235   s11 += carry10;
1236   s10 -= int64_lshift21(carry10);
1237 
1238   carry1 = (s1 + (1 << 20)) >> 21;
1239   s2 += carry1;
1240   s1 -= int64_lshift21(carry1);
1241   carry3 = (s3 + (1 << 20)) >> 21;
1242   s4 += carry3;
1243   s3 -= int64_lshift21(carry3);
1244   carry5 = (s5 + (1 << 20)) >> 21;
1245   s6 += carry5;
1246   s5 -= int64_lshift21(carry5);
1247   carry7 = (s7 + (1 << 20)) >> 21;
1248   s8 += carry7;
1249   s7 -= int64_lshift21(carry7);
1250   carry9 = (s9 + (1 << 20)) >> 21;
1251   s10 += carry9;
1252   s9 -= int64_lshift21(carry9);
1253   carry11 = (s11 + (1 << 20)) >> 21;
1254   s12 += carry11;
1255   s11 -= int64_lshift21(carry11);
1256 
1257   s0 += s12 * 666643;
1258   s1 += s12 * 470296;
1259   s2 += s12 * 654183;
1260   s3 -= s12 * 997805;
1261   s4 += s12 * 136657;
1262   s5 -= s12 * 683901;
1263   s12 = 0;
1264 
1265   carry0 = s0 >> 21;
1266   s1 += carry0;
1267   s0 -= int64_lshift21(carry0);
1268   carry1 = s1 >> 21;
1269   s2 += carry1;
1270   s1 -= int64_lshift21(carry1);
1271   carry2 = s2 >> 21;
1272   s3 += carry2;
1273   s2 -= int64_lshift21(carry2);
1274   carry3 = s3 >> 21;
1275   s4 += carry3;
1276   s3 -= int64_lshift21(carry3);
1277   carry4 = s4 >> 21;
1278   s5 += carry4;
1279   s4 -= int64_lshift21(carry4);
1280   carry5 = s5 >> 21;
1281   s6 += carry5;
1282   s5 -= int64_lshift21(carry5);
1283   carry6 = s6 >> 21;
1284   s7 += carry6;
1285   s6 -= int64_lshift21(carry6);
1286   carry7 = s7 >> 21;
1287   s8 += carry7;
1288   s7 -= int64_lshift21(carry7);
1289   carry8 = s8 >> 21;
1290   s9 += carry8;
1291   s8 -= int64_lshift21(carry8);
1292   carry9 = s9 >> 21;
1293   s10 += carry9;
1294   s9 -= int64_lshift21(carry9);
1295   carry10 = s10 >> 21;
1296   s11 += carry10;
1297   s10 -= int64_lshift21(carry10);
1298   carry11 = s11 >> 21;
1299   s12 += carry11;
1300   s11 -= int64_lshift21(carry11);
1301 
1302   s0 += s12 * 666643;
1303   s1 += s12 * 470296;
1304   s2 += s12 * 654183;
1305   s3 -= s12 * 997805;
1306   s4 += s12 * 136657;
1307   s5 -= s12 * 683901;
1308   s12 = 0;
1309 
1310   carry0 = s0 >> 21;
1311   s1 += carry0;
1312   s0 -= int64_lshift21(carry0);
1313   carry1 = s1 >> 21;
1314   s2 += carry1;
1315   s1 -= int64_lshift21(carry1);
1316   carry2 = s2 >> 21;
1317   s3 += carry2;
1318   s2 -= int64_lshift21(carry2);
1319   carry3 = s3 >> 21;
1320   s4 += carry3;
1321   s3 -= int64_lshift21(carry3);
1322   carry4 = s4 >> 21;
1323   s5 += carry4;
1324   s4 -= int64_lshift21(carry4);
1325   carry5 = s5 >> 21;
1326   s6 += carry5;
1327   s5 -= int64_lshift21(carry5);
1328   carry6 = s6 >> 21;
1329   s7 += carry6;
1330   s6 -= int64_lshift21(carry6);
1331   carry7 = s7 >> 21;
1332   s8 += carry7;
1333   s7 -= int64_lshift21(carry7);
1334   carry8 = s8 >> 21;
1335   s9 += carry8;
1336   s8 -= int64_lshift21(carry8);
1337   carry9 = s9 >> 21;
1338   s10 += carry9;
1339   s9 -= int64_lshift21(carry9);
1340   carry10 = s10 >> 21;
1341   s11 += carry10;
1342   s10 -= int64_lshift21(carry10);
1343 
1344   s[0] = s0 >> 0;
1345   s[1] = s0 >> 8;
1346   s[2] = (s0 >> 16) | (s1 << 5);
1347   s[3] = s1 >> 3;
1348   s[4] = s1 >> 11;
1349   s[5] = (s1 >> 19) | (s2 << 2);
1350   s[6] = s2 >> 6;
1351   s[7] = (s2 >> 14) | (s3 << 7);
1352   s[8] = s3 >> 1;
1353   s[9] = s3 >> 9;
1354   s[10] = (s3 >> 17) | (s4 << 4);
1355   s[11] = s4 >> 4;
1356   s[12] = s4 >> 12;
1357   s[13] = (s4 >> 20) | (s5 << 1);
1358   s[14] = s5 >> 7;
1359   s[15] = (s5 >> 15) | (s6 << 6);
1360   s[16] = s6 >> 2;
1361   s[17] = s6 >> 10;
1362   s[18] = (s6 >> 18) | (s7 << 3);
1363   s[19] = s7 >> 5;
1364   s[20] = s7 >> 13;
1365   s[21] = s8 >> 0;
1366   s[22] = s8 >> 8;
1367   s[23] = (s8 >> 16) | (s9 << 5);
1368   s[24] = s9 >> 3;
1369   s[25] = s9 >> 11;
1370   s[26] = (s9 >> 19) | (s10 << 2);
1371   s[27] = s10 >> 6;
1372   s[28] = (s10 >> 14) | (s11 << 7);
1373   s[29] = s11 >> 1;
1374   s[30] = s11 >> 9;
1375   s[31] = s11 >> 17;
1376 }
1377 
1378 // Input:
1379 //   a[0]+256*a[1]+...+256^31*a[31] = a
1380 //   b[0]+256*b[1]+...+256^31*b[31] = b
1381 //   c[0]+256*c[1]+...+256^31*c[31] = c
1382 //
1383 // Output:
1384 //   s[0]+256*s[1]+...+256^31*s[31] = (ab+c) mod l
1385 //   where l = 2^252 + 27742317777372353535851937790883648493.
sc_muladd(uint8_t * s,const uint8_t * a,const uint8_t * b,const uint8_t * c)1386 static void sc_muladd(uint8_t *s, const uint8_t *a, const uint8_t *b,
1387                       const uint8_t *c) {
1388   int64_t a0 = 2097151 & load_3(a);
1389   int64_t a1 = 2097151 & (load_4(a + 2) >> 5);
1390   int64_t a2 = 2097151 & (load_3(a + 5) >> 2);
1391   int64_t a3 = 2097151 & (load_4(a + 7) >> 7);
1392   int64_t a4 = 2097151 & (load_4(a + 10) >> 4);
1393   int64_t a5 = 2097151 & (load_3(a + 13) >> 1);
1394   int64_t a6 = 2097151 & (load_4(a + 15) >> 6);
1395   int64_t a7 = 2097151 & (load_3(a + 18) >> 3);
1396   int64_t a8 = 2097151 & load_3(a + 21);
1397   int64_t a9 = 2097151 & (load_4(a + 23) >> 5);
1398   int64_t a10 = 2097151 & (load_3(a + 26) >> 2);
1399   int64_t a11 = (load_4(a + 28) >> 7);
1400   int64_t b0 = 2097151 & load_3(b);
1401   int64_t b1 = 2097151 & (load_4(b + 2) >> 5);
1402   int64_t b2 = 2097151 & (load_3(b + 5) >> 2);
1403   int64_t b3 = 2097151 & (load_4(b + 7) >> 7);
1404   int64_t b4 = 2097151 & (load_4(b + 10) >> 4);
1405   int64_t b5 = 2097151 & (load_3(b + 13) >> 1);
1406   int64_t b6 = 2097151 & (load_4(b + 15) >> 6);
1407   int64_t b7 = 2097151 & (load_3(b + 18) >> 3);
1408   int64_t b8 = 2097151 & load_3(b + 21);
1409   int64_t b9 = 2097151 & (load_4(b + 23) >> 5);
1410   int64_t b10 = 2097151 & (load_3(b + 26) >> 2);
1411   int64_t b11 = (load_4(b + 28) >> 7);
1412   int64_t c0 = 2097151 & load_3(c);
1413   int64_t c1 = 2097151 & (load_4(c + 2) >> 5);
1414   int64_t c2 = 2097151 & (load_3(c + 5) >> 2);
1415   int64_t c3 = 2097151 & (load_4(c + 7) >> 7);
1416   int64_t c4 = 2097151 & (load_4(c + 10) >> 4);
1417   int64_t c5 = 2097151 & (load_3(c + 13) >> 1);
1418   int64_t c6 = 2097151 & (load_4(c + 15) >> 6);
1419   int64_t c7 = 2097151 & (load_3(c + 18) >> 3);
1420   int64_t c8 = 2097151 & load_3(c + 21);
1421   int64_t c9 = 2097151 & (load_4(c + 23) >> 5);
1422   int64_t c10 = 2097151 & (load_3(c + 26) >> 2);
1423   int64_t c11 = (load_4(c + 28) >> 7);
1424   int64_t s0;
1425   int64_t s1;
1426   int64_t s2;
1427   int64_t s3;
1428   int64_t s4;
1429   int64_t s5;
1430   int64_t s6;
1431   int64_t s7;
1432   int64_t s8;
1433   int64_t s9;
1434   int64_t s10;
1435   int64_t s11;
1436   int64_t s12;
1437   int64_t s13;
1438   int64_t s14;
1439   int64_t s15;
1440   int64_t s16;
1441   int64_t s17;
1442   int64_t s18;
1443   int64_t s19;
1444   int64_t s20;
1445   int64_t s21;
1446   int64_t s22;
1447   int64_t s23;
1448   int64_t carry0;
1449   int64_t carry1;
1450   int64_t carry2;
1451   int64_t carry3;
1452   int64_t carry4;
1453   int64_t carry5;
1454   int64_t carry6;
1455   int64_t carry7;
1456   int64_t carry8;
1457   int64_t carry9;
1458   int64_t carry10;
1459   int64_t carry11;
1460   int64_t carry12;
1461   int64_t carry13;
1462   int64_t carry14;
1463   int64_t carry15;
1464   int64_t carry16;
1465   int64_t carry17;
1466   int64_t carry18;
1467   int64_t carry19;
1468   int64_t carry20;
1469   int64_t carry21;
1470   int64_t carry22;
1471 
1472   s0 = c0 + a0 * b0;
1473   s1 = c1 + a0 * b1 + a1 * b0;
1474   s2 = c2 + a0 * b2 + a1 * b1 + a2 * b0;
1475   s3 = c3 + a0 * b3 + a1 * b2 + a2 * b1 + a3 * b0;
1476   s4 = c4 + a0 * b4 + a1 * b3 + a2 * b2 + a3 * b1 + a4 * b0;
1477   s5 = c5 + a0 * b5 + a1 * b4 + a2 * b3 + a3 * b2 + a4 * b1 + a5 * b0;
1478   s6 = c6 + a0 * b6 + a1 * b5 + a2 * b4 + a3 * b3 + a4 * b2 + a5 * b1 + a6 * b0;
1479   s7 = c7 + a0 * b7 + a1 * b6 + a2 * b5 + a3 * b4 + a4 * b3 + a5 * b2 +
1480        a6 * b1 + a7 * b0;
1481   s8 = c8 + a0 * b8 + a1 * b7 + a2 * b6 + a3 * b5 + a4 * b4 + a5 * b3 +
1482        a6 * b2 + a7 * b1 + a8 * b0;
1483   s9 = c9 + a0 * b9 + a1 * b8 + a2 * b7 + a3 * b6 + a4 * b5 + a5 * b4 +
1484        a6 * b3 + a7 * b2 + a8 * b1 + a9 * b0;
1485   s10 = c10 + a0 * b10 + a1 * b9 + a2 * b8 + a3 * b7 + a4 * b6 + a5 * b5 +
1486         a6 * b4 + a7 * b3 + a8 * b2 + a9 * b1 + a10 * b0;
1487   s11 = c11 + a0 * b11 + a1 * b10 + a2 * b9 + a3 * b8 + a4 * b7 + a5 * b6 +
1488         a6 * b5 + a7 * b4 + a8 * b3 + a9 * b2 + a10 * b1 + a11 * b0;
1489   s12 = a1 * b11 + a2 * b10 + a3 * b9 + a4 * b8 + a5 * b7 + a6 * b6 + a7 * b5 +
1490         a8 * b4 + a9 * b3 + a10 * b2 + a11 * b1;
1491   s13 = a2 * b11 + a3 * b10 + a4 * b9 + a5 * b8 + a6 * b7 + a7 * b6 + a8 * b5 +
1492         a9 * b4 + a10 * b3 + a11 * b2;
1493   s14 = a3 * b11 + a4 * b10 + a5 * b9 + a6 * b8 + a7 * b7 + a8 * b6 + a9 * b5 +
1494         a10 * b4 + a11 * b3;
1495   s15 = a4 * b11 + a5 * b10 + a6 * b9 + a7 * b8 + a8 * b7 + a9 * b6 + a10 * b5 +
1496         a11 * b4;
1497   s16 = a5 * b11 + a6 * b10 + a7 * b9 + a8 * b8 + a9 * b7 + a10 * b6 + a11 * b5;
1498   s17 = a6 * b11 + a7 * b10 + a8 * b9 + a9 * b8 + a10 * b7 + a11 * b6;
1499   s18 = a7 * b11 + a8 * b10 + a9 * b9 + a10 * b8 + a11 * b7;
1500   s19 = a8 * b11 + a9 * b10 + a10 * b9 + a11 * b8;
1501   s20 = a9 * b11 + a10 * b10 + a11 * b9;
1502   s21 = a10 * b11 + a11 * b10;
1503   s22 = a11 * b11;
1504   s23 = 0;
1505 
1506   carry0 = (s0 + (1 << 20)) >> 21;
1507   s1 += carry0;
1508   s0 -= int64_lshift21(carry0);
1509   carry2 = (s2 + (1 << 20)) >> 21;
1510   s3 += carry2;
1511   s2 -= int64_lshift21(carry2);
1512   carry4 = (s4 + (1 << 20)) >> 21;
1513   s5 += carry4;
1514   s4 -= int64_lshift21(carry4);
1515   carry6 = (s6 + (1 << 20)) >> 21;
1516   s7 += carry6;
1517   s6 -= int64_lshift21(carry6);
1518   carry8 = (s8 + (1 << 20)) >> 21;
1519   s9 += carry8;
1520   s8 -= int64_lshift21(carry8);
1521   carry10 = (s10 + (1 << 20)) >> 21;
1522   s11 += carry10;
1523   s10 -= int64_lshift21(carry10);
1524   carry12 = (s12 + (1 << 20)) >> 21;
1525   s13 += carry12;
1526   s12 -= int64_lshift21(carry12);
1527   carry14 = (s14 + (1 << 20)) >> 21;
1528   s15 += carry14;
1529   s14 -= int64_lshift21(carry14);
1530   carry16 = (s16 + (1 << 20)) >> 21;
1531   s17 += carry16;
1532   s16 -= int64_lshift21(carry16);
1533   carry18 = (s18 + (1 << 20)) >> 21;
1534   s19 += carry18;
1535   s18 -= int64_lshift21(carry18);
1536   carry20 = (s20 + (1 << 20)) >> 21;
1537   s21 += carry20;
1538   s20 -= int64_lshift21(carry20);
1539   carry22 = (s22 + (1 << 20)) >> 21;
1540   s23 += carry22;
1541   s22 -= int64_lshift21(carry22);
1542 
1543   carry1 = (s1 + (1 << 20)) >> 21;
1544   s2 += carry1;
1545   s1 -= int64_lshift21(carry1);
1546   carry3 = (s3 + (1 << 20)) >> 21;
1547   s4 += carry3;
1548   s3 -= int64_lshift21(carry3);
1549   carry5 = (s5 + (1 << 20)) >> 21;
1550   s6 += carry5;
1551   s5 -= int64_lshift21(carry5);
1552   carry7 = (s7 + (1 << 20)) >> 21;
1553   s8 += carry7;
1554   s7 -= int64_lshift21(carry7);
1555   carry9 = (s9 + (1 << 20)) >> 21;
1556   s10 += carry9;
1557   s9 -= int64_lshift21(carry9);
1558   carry11 = (s11 + (1 << 20)) >> 21;
1559   s12 += carry11;
1560   s11 -= int64_lshift21(carry11);
1561   carry13 = (s13 + (1 << 20)) >> 21;
1562   s14 += carry13;
1563   s13 -= int64_lshift21(carry13);
1564   carry15 = (s15 + (1 << 20)) >> 21;
1565   s16 += carry15;
1566   s15 -= int64_lshift21(carry15);
1567   carry17 = (s17 + (1 << 20)) >> 21;
1568   s18 += carry17;
1569   s17 -= int64_lshift21(carry17);
1570   carry19 = (s19 + (1 << 20)) >> 21;
1571   s20 += carry19;
1572   s19 -= int64_lshift21(carry19);
1573   carry21 = (s21 + (1 << 20)) >> 21;
1574   s22 += carry21;
1575   s21 -= int64_lshift21(carry21);
1576 
1577   s11 += s23 * 666643;
1578   s12 += s23 * 470296;
1579   s13 += s23 * 654183;
1580   s14 -= s23 * 997805;
1581   s15 += s23 * 136657;
1582   s16 -= s23 * 683901;
1583   s23 = 0;
1584 
1585   s10 += s22 * 666643;
1586   s11 += s22 * 470296;
1587   s12 += s22 * 654183;
1588   s13 -= s22 * 997805;
1589   s14 += s22 * 136657;
1590   s15 -= s22 * 683901;
1591   s22 = 0;
1592 
1593   s9 += s21 * 666643;
1594   s10 += s21 * 470296;
1595   s11 += s21 * 654183;
1596   s12 -= s21 * 997805;
1597   s13 += s21 * 136657;
1598   s14 -= s21 * 683901;
1599   s21 = 0;
1600 
1601   s8 += s20 * 666643;
1602   s9 += s20 * 470296;
1603   s10 += s20 * 654183;
1604   s11 -= s20 * 997805;
1605   s12 += s20 * 136657;
1606   s13 -= s20 * 683901;
1607   s20 = 0;
1608 
1609   s7 += s19 * 666643;
1610   s8 += s19 * 470296;
1611   s9 += s19 * 654183;
1612   s10 -= s19 * 997805;
1613   s11 += s19 * 136657;
1614   s12 -= s19 * 683901;
1615   s19 = 0;
1616 
1617   s6 += s18 * 666643;
1618   s7 += s18 * 470296;
1619   s8 += s18 * 654183;
1620   s9 -= s18 * 997805;
1621   s10 += s18 * 136657;
1622   s11 -= s18 * 683901;
1623   s18 = 0;
1624 
1625   carry6 = (s6 + (1 << 20)) >> 21;
1626   s7 += carry6;
1627   s6 -= int64_lshift21(carry6);
1628   carry8 = (s8 + (1 << 20)) >> 21;
1629   s9 += carry8;
1630   s8 -= int64_lshift21(carry8);
1631   carry10 = (s10 + (1 << 20)) >> 21;
1632   s11 += carry10;
1633   s10 -= int64_lshift21(carry10);
1634   carry12 = (s12 + (1 << 20)) >> 21;
1635   s13 += carry12;
1636   s12 -= int64_lshift21(carry12);
1637   carry14 = (s14 + (1 << 20)) >> 21;
1638   s15 += carry14;
1639   s14 -= int64_lshift21(carry14);
1640   carry16 = (s16 + (1 << 20)) >> 21;
1641   s17 += carry16;
1642   s16 -= int64_lshift21(carry16);
1643 
1644   carry7 = (s7 + (1 << 20)) >> 21;
1645   s8 += carry7;
1646   s7 -= int64_lshift21(carry7);
1647   carry9 = (s9 + (1 << 20)) >> 21;
1648   s10 += carry9;
1649   s9 -= int64_lshift21(carry9);
1650   carry11 = (s11 + (1 << 20)) >> 21;
1651   s12 += carry11;
1652   s11 -= int64_lshift21(carry11);
1653   carry13 = (s13 + (1 << 20)) >> 21;
1654   s14 += carry13;
1655   s13 -= int64_lshift21(carry13);
1656   carry15 = (s15 + (1 << 20)) >> 21;
1657   s16 += carry15;
1658   s15 -= int64_lshift21(carry15);
1659 
1660   s5 += s17 * 666643;
1661   s6 += s17 * 470296;
1662   s7 += s17 * 654183;
1663   s8 -= s17 * 997805;
1664   s9 += s17 * 136657;
1665   s10 -= s17 * 683901;
1666   s17 = 0;
1667 
1668   s4 += s16 * 666643;
1669   s5 += s16 * 470296;
1670   s6 += s16 * 654183;
1671   s7 -= s16 * 997805;
1672   s8 += s16 * 136657;
1673   s9 -= s16 * 683901;
1674   s16 = 0;
1675 
1676   s3 += s15 * 666643;
1677   s4 += s15 * 470296;
1678   s5 += s15 * 654183;
1679   s6 -= s15 * 997805;
1680   s7 += s15 * 136657;
1681   s8 -= s15 * 683901;
1682   s15 = 0;
1683 
1684   s2 += s14 * 666643;
1685   s3 += s14 * 470296;
1686   s4 += s14 * 654183;
1687   s5 -= s14 * 997805;
1688   s6 += s14 * 136657;
1689   s7 -= s14 * 683901;
1690   s14 = 0;
1691 
1692   s1 += s13 * 666643;
1693   s2 += s13 * 470296;
1694   s3 += s13 * 654183;
1695   s4 -= s13 * 997805;
1696   s5 += s13 * 136657;
1697   s6 -= s13 * 683901;
1698   s13 = 0;
1699 
1700   s0 += s12 * 666643;
1701   s1 += s12 * 470296;
1702   s2 += s12 * 654183;
1703   s3 -= s12 * 997805;
1704   s4 += s12 * 136657;
1705   s5 -= s12 * 683901;
1706   s12 = 0;
1707 
1708   carry0 = (s0 + (1 << 20)) >> 21;
1709   s1 += carry0;
1710   s0 -= int64_lshift21(carry0);
1711   carry2 = (s2 + (1 << 20)) >> 21;
1712   s3 += carry2;
1713   s2 -= int64_lshift21(carry2);
1714   carry4 = (s4 + (1 << 20)) >> 21;
1715   s5 += carry4;
1716   s4 -= int64_lshift21(carry4);
1717   carry6 = (s6 + (1 << 20)) >> 21;
1718   s7 += carry6;
1719   s6 -= int64_lshift21(carry6);
1720   carry8 = (s8 + (1 << 20)) >> 21;
1721   s9 += carry8;
1722   s8 -= int64_lshift21(carry8);
1723   carry10 = (s10 + (1 << 20)) >> 21;
1724   s11 += carry10;
1725   s10 -= int64_lshift21(carry10);
1726 
1727   carry1 = (s1 + (1 << 20)) >> 21;
1728   s2 += carry1;
1729   s1 -= int64_lshift21(carry1);
1730   carry3 = (s3 + (1 << 20)) >> 21;
1731   s4 += carry3;
1732   s3 -= int64_lshift21(carry3);
1733   carry5 = (s5 + (1 << 20)) >> 21;
1734   s6 += carry5;
1735   s5 -= int64_lshift21(carry5);
1736   carry7 = (s7 + (1 << 20)) >> 21;
1737   s8 += carry7;
1738   s7 -= int64_lshift21(carry7);
1739   carry9 = (s9 + (1 << 20)) >> 21;
1740   s10 += carry9;
1741   s9 -= int64_lshift21(carry9);
1742   carry11 = (s11 + (1 << 20)) >> 21;
1743   s12 += carry11;
1744   s11 -= int64_lshift21(carry11);
1745 
1746   s0 += s12 * 666643;
1747   s1 += s12 * 470296;
1748   s2 += s12 * 654183;
1749   s3 -= s12 * 997805;
1750   s4 += s12 * 136657;
1751   s5 -= s12 * 683901;
1752   s12 = 0;
1753 
1754   carry0 = s0 >> 21;
1755   s1 += carry0;
1756   s0 -= int64_lshift21(carry0);
1757   carry1 = s1 >> 21;
1758   s2 += carry1;
1759   s1 -= int64_lshift21(carry1);
1760   carry2 = s2 >> 21;
1761   s3 += carry2;
1762   s2 -= int64_lshift21(carry2);
1763   carry3 = s3 >> 21;
1764   s4 += carry3;
1765   s3 -= int64_lshift21(carry3);
1766   carry4 = s4 >> 21;
1767   s5 += carry4;
1768   s4 -= int64_lshift21(carry4);
1769   carry5 = s5 >> 21;
1770   s6 += carry5;
1771   s5 -= int64_lshift21(carry5);
1772   carry6 = s6 >> 21;
1773   s7 += carry6;
1774   s6 -= int64_lshift21(carry6);
1775   carry7 = s7 >> 21;
1776   s8 += carry7;
1777   s7 -= int64_lshift21(carry7);
1778   carry8 = s8 >> 21;
1779   s9 += carry8;
1780   s8 -= int64_lshift21(carry8);
1781   carry9 = s9 >> 21;
1782   s10 += carry9;
1783   s9 -= int64_lshift21(carry9);
1784   carry10 = s10 >> 21;
1785   s11 += carry10;
1786   s10 -= int64_lshift21(carry10);
1787   carry11 = s11 >> 21;
1788   s12 += carry11;
1789   s11 -= int64_lshift21(carry11);
1790 
1791   s0 += s12 * 666643;
1792   s1 += s12 * 470296;
1793   s2 += s12 * 654183;
1794   s3 -= s12 * 997805;
1795   s4 += s12 * 136657;
1796   s5 -= s12 * 683901;
1797   s12 = 0;
1798 
1799   carry0 = s0 >> 21;
1800   s1 += carry0;
1801   s0 -= int64_lshift21(carry0);
1802   carry1 = s1 >> 21;
1803   s2 += carry1;
1804   s1 -= int64_lshift21(carry1);
1805   carry2 = s2 >> 21;
1806   s3 += carry2;
1807   s2 -= int64_lshift21(carry2);
1808   carry3 = s3 >> 21;
1809   s4 += carry3;
1810   s3 -= int64_lshift21(carry3);
1811   carry4 = s4 >> 21;
1812   s5 += carry4;
1813   s4 -= int64_lshift21(carry4);
1814   carry5 = s5 >> 21;
1815   s6 += carry5;
1816   s5 -= int64_lshift21(carry5);
1817   carry6 = s6 >> 21;
1818   s7 += carry6;
1819   s6 -= int64_lshift21(carry6);
1820   carry7 = s7 >> 21;
1821   s8 += carry7;
1822   s7 -= int64_lshift21(carry7);
1823   carry8 = s8 >> 21;
1824   s9 += carry8;
1825   s8 -= int64_lshift21(carry8);
1826   carry9 = s9 >> 21;
1827   s10 += carry9;
1828   s9 -= int64_lshift21(carry9);
1829   carry10 = s10 >> 21;
1830   s11 += carry10;
1831   s10 -= int64_lshift21(carry10);
1832 
1833   s[0] = s0 >> 0;
1834   s[1] = s0 >> 8;
1835   s[2] = (s0 >> 16) | (s1 << 5);
1836   s[3] = s1 >> 3;
1837   s[4] = s1 >> 11;
1838   s[5] = (s1 >> 19) | (s2 << 2);
1839   s[6] = s2 >> 6;
1840   s[7] = (s2 >> 14) | (s3 << 7);
1841   s[8] = s3 >> 1;
1842   s[9] = s3 >> 9;
1843   s[10] = (s3 >> 17) | (s4 << 4);
1844   s[11] = s4 >> 4;
1845   s[12] = s4 >> 12;
1846   s[13] = (s4 >> 20) | (s5 << 1);
1847   s[14] = s5 >> 7;
1848   s[15] = (s5 >> 15) | (s6 << 6);
1849   s[16] = s6 >> 2;
1850   s[17] = s6 >> 10;
1851   s[18] = (s6 >> 18) | (s7 << 3);
1852   s[19] = s7 >> 5;
1853   s[20] = s7 >> 13;
1854   s[21] = s8 >> 0;
1855   s[22] = s8 >> 8;
1856   s[23] = (s8 >> 16) | (s9 << 5);
1857   s[24] = s9 >> 3;
1858   s[25] = s9 >> 11;
1859   s[26] = (s9 >> 19) | (s10 << 2);
1860   s[27] = s10 >> 6;
1861   s[28] = (s10 >> 14) | (s11 << 7);
1862   s[29] = s11 >> 1;
1863   s[30] = s11 >> 9;
1864   s[31] = s11 >> 17;
1865 }
1866 
ED25519_keypair(uint8_t out_public_key[32],uint8_t out_private_key[64])1867 void ED25519_keypair(uint8_t out_public_key[32], uint8_t out_private_key[64]) {
1868   uint8_t seed[32];
1869   RAND_bytes(seed, 32);
1870   ED25519_keypair_from_seed(out_public_key, out_private_key, seed);
1871 }
1872 
ED25519_sign(uint8_t out_sig[64],const uint8_t * message,size_t message_len,const uint8_t private_key[64])1873 int ED25519_sign(uint8_t out_sig[64], const uint8_t *message,
1874                  size_t message_len, const uint8_t private_key[64]) {
1875   // NOTE: The documentation on this function says that it returns zero on
1876   // allocation failure. While that can't happen with the current
1877   // implementation, we want to reserve the ability to allocate in this
1878   // implementation in the future.
1879 
1880   uint8_t az[SHA512_DIGEST_LENGTH];
1881   SHA512(private_key, 32, az);
1882 
1883   az[0] &= 248;
1884   az[31] &= 63;
1885   az[31] |= 64;
1886 
1887   SHA512_CTX hash_ctx;
1888   SHA512_Init(&hash_ctx);
1889   SHA512_Update(&hash_ctx, az + 32, 32);
1890   SHA512_Update(&hash_ctx, message, message_len);
1891   uint8_t nonce[SHA512_DIGEST_LENGTH];
1892   SHA512_Final(nonce, &hash_ctx);
1893 
1894   x25519_sc_reduce(nonce);
1895   ge_p3 R;
1896   x25519_ge_scalarmult_base(&R, nonce);
1897   ge_p3_tobytes(out_sig, &R);
1898 
1899   SHA512_Init(&hash_ctx);
1900   SHA512_Update(&hash_ctx, out_sig, 32);
1901   SHA512_Update(&hash_ctx, private_key + 32, 32);
1902   SHA512_Update(&hash_ctx, message, message_len);
1903   uint8_t hram[SHA512_DIGEST_LENGTH];
1904   SHA512_Final(hram, &hash_ctx);
1905 
1906   x25519_sc_reduce(hram);
1907   sc_muladd(out_sig + 32, hram, az, nonce);
1908 
1909   // The signature is computed from the private key, but is public.
1910   CONSTTIME_DECLASSIFY(out_sig, 64);
1911   return 1;
1912 }
1913 
ED25519_verify(const uint8_t * message,size_t message_len,const uint8_t signature[64],const uint8_t public_key[32])1914 int ED25519_verify(const uint8_t *message, size_t message_len,
1915                    const uint8_t signature[64], const uint8_t public_key[32]) {
1916   ge_p3 A;
1917   if ((signature[63] & 224) != 0 ||
1918       !x25519_ge_frombytes_vartime(&A, public_key)) {
1919     return 0;
1920   }
1921 
1922   fe_loose t;
1923   fe_neg(&t, &A.X);
1924   fe_carry(&A.X, &t);
1925   fe_neg(&t, &A.T);
1926   fe_carry(&A.T, &t);
1927 
1928   uint8_t pkcopy[32];
1929   OPENSSL_memcpy(pkcopy, public_key, 32);
1930   uint8_t rcopy[32];
1931   OPENSSL_memcpy(rcopy, signature, 32);
1932   uint8_t scopy[32];
1933   OPENSSL_memcpy(scopy, signature + 32, 32);
1934 
1935   // https://tools.ietf.org/html/rfc8032#section-5.1.7 requires that s be in
1936   // the range [0, order) in order to prevent signature malleability.
1937 
1938   // kOrder is the order of Curve25519 in little-endian form.
1939   static const uint64_t kOrder[4] = {
1940     UINT64_C(0x5812631a5cf5d3ed),
1941     UINT64_C(0x14def9dea2f79cd6),
1942     0,
1943     UINT64_C(0x1000000000000000),
1944   };
1945   for (size_t i = 3;; i--) {
1946     uint64_t word = CRYPTO_load_u64_le(scopy + i * 8);
1947     if (word > kOrder[i]) {
1948       return 0;
1949     } else if (word < kOrder[i]) {
1950       break;
1951     } else if (i == 0) {
1952       return 0;
1953     }
1954   }
1955 
1956   SHA512_CTX hash_ctx;
1957   SHA512_Init(&hash_ctx);
1958   SHA512_Update(&hash_ctx, signature, 32);
1959   SHA512_Update(&hash_ctx, public_key, 32);
1960   SHA512_Update(&hash_ctx, message, message_len);
1961   uint8_t h[SHA512_DIGEST_LENGTH];
1962   SHA512_Final(h, &hash_ctx);
1963 
1964   x25519_sc_reduce(h);
1965 
1966   ge_p2 R;
1967   ge_double_scalarmult_vartime(&R, h, &A, scopy);
1968 
1969   uint8_t rcheck[32];
1970   x25519_ge_tobytes(rcheck, &R);
1971 
1972   return CRYPTO_memcmp(rcheck, rcopy, sizeof(rcheck)) == 0;
1973 }
1974 
ED25519_keypair_from_seed(uint8_t out_public_key[32],uint8_t out_private_key[64],const uint8_t seed[32])1975 void ED25519_keypair_from_seed(uint8_t out_public_key[32],
1976                                uint8_t out_private_key[64],
1977                                const uint8_t seed[32]) {
1978   uint8_t az[SHA512_DIGEST_LENGTH];
1979   SHA512(seed, 32, az);
1980 
1981   az[0] &= 248;
1982   az[31] &= 127;
1983   az[31] |= 64;
1984 
1985   ge_p3 A;
1986   x25519_ge_scalarmult_base(&A, az);
1987   ge_p3_tobytes(out_public_key, &A);
1988   // The public key is derived from the private key, but it is public.
1989   CONSTTIME_DECLASSIFY(out_public_key, 32);
1990 
1991   OPENSSL_memcpy(out_private_key, seed, 32);
1992   OPENSSL_memcpy(out_private_key + 32, out_public_key, 32);
1993 }
1994 
1995 
x25519_scalar_mult_generic(uint8_t out[32],const uint8_t scalar[32],const uint8_t point[32])1996 static void x25519_scalar_mult_generic(uint8_t out[32],
1997                                        const uint8_t scalar[32],
1998                                        const uint8_t point[32]) {
1999   fe x1, x2, z2, x3, z3, tmp0, tmp1;
2000   fe_loose x2l, z2l, x3l, tmp0l, tmp1l;
2001 
2002   uint8_t e[32];
2003   OPENSSL_memcpy(e, scalar, 32);
2004   e[0] &= 248;
2005   e[31] &= 127;
2006   e[31] |= 64;
2007 
2008   // The following implementation was transcribed to Coq and proven to
2009   // correspond to unary scalar multiplication in affine coordinates given that
2010   // x1 != 0 is the x coordinate of some point on the curve. It was also checked
2011   // in Coq that doing a ladderstep with x1 = x3 = 0 gives z2' = z3' = 0, and z2
2012   // = z3 = 0 gives z2' = z3' = 0. The statement was quantified over the
2013   // underlying field, so it applies to Curve25519 itself and the quadratic
2014   // twist of Curve25519. It was not proven in Coq that prime-field arithmetic
2015   // correctly simulates extension-field arithmetic on prime-field values.
2016   // The decoding of the byte array representation of e was not considered.
2017   // Specification of Montgomery curves in affine coordinates:
2018   // <https://github.com/mit-plv/fiat-crypto/blob/2456d821825521f7e03e65882cc3521795b0320f/src/Spec/MontgomeryCurve.v#L27>
2019   // Proof that these form a group that is isomorphic to a Weierstrass curve:
2020   // <https://github.com/mit-plv/fiat-crypto/blob/2456d821825521f7e03e65882cc3521795b0320f/src/Curves/Montgomery/AffineProofs.v#L35>
2021   // Coq transcription and correctness proof of the loop (where scalarbits=255):
2022   // <https://github.com/mit-plv/fiat-crypto/blob/2456d821825521f7e03e65882cc3521795b0320f/src/Curves/Montgomery/XZ.v#L118>
2023   // <https://github.com/mit-plv/fiat-crypto/blob/2456d821825521f7e03e65882cc3521795b0320f/src/Curves/Montgomery/XZProofs.v#L278>
2024   // preconditions: 0 <= e < 2^255 (not necessarily e < order), fe_invert(0) = 0
2025   fe_frombytes(&x1, point);
2026   fe_1(&x2);
2027   fe_0(&z2);
2028   fe_copy(&x3, &x1);
2029   fe_1(&z3);
2030 
2031   unsigned swap = 0;
2032   int pos;
2033   for (pos = 254; pos >= 0; --pos) {
2034     // loop invariant as of right before the test, for the case where x1 != 0:
2035     //   pos >= -1; if z2 = 0 then x2 is nonzero; if z3 = 0 then x3 is nonzero
2036     //   let r := e >> (pos+1) in the following equalities of projective points:
2037     //   to_xz (r*P)     === if swap then (x3, z3) else (x2, z2)
2038     //   to_xz ((r+1)*P) === if swap then (x2, z2) else (x3, z3)
2039     //   x1 is the nonzero x coordinate of the nonzero point (r*P-(r+1)*P)
2040     unsigned b = 1 & (e[pos / 8] >> (pos & 7));
2041     swap ^= b;
2042     fe_cswap(&x2, &x3, swap);
2043     fe_cswap(&z2, &z3, swap);
2044     swap = b;
2045     // Coq transcription of ladderstep formula (called from transcribed loop):
2046     // <https://github.com/mit-plv/fiat-crypto/blob/2456d821825521f7e03e65882cc3521795b0320f/src/Curves/Montgomery/XZ.v#L89>
2047     // <https://github.com/mit-plv/fiat-crypto/blob/2456d821825521f7e03e65882cc3521795b0320f/src/Curves/Montgomery/XZProofs.v#L131>
2048     // x1 != 0 <https://github.com/mit-plv/fiat-crypto/blob/2456d821825521f7e03e65882cc3521795b0320f/src/Curves/Montgomery/XZProofs.v#L217>
2049     // x1  = 0 <https://github.com/mit-plv/fiat-crypto/blob/2456d821825521f7e03e65882cc3521795b0320f/src/Curves/Montgomery/XZProofs.v#L147>
2050     fe_sub(&tmp0l, &x3, &z3);
2051     fe_sub(&tmp1l, &x2, &z2);
2052     fe_add(&x2l, &x2, &z2);
2053     fe_add(&z2l, &x3, &z3);
2054     fe_mul_tll(&z3, &tmp0l, &x2l);
2055     fe_mul_tll(&z2, &z2l, &tmp1l);
2056     fe_sq_tl(&tmp0, &tmp1l);
2057     fe_sq_tl(&tmp1, &x2l);
2058     fe_add(&x3l, &z3, &z2);
2059     fe_sub(&z2l, &z3, &z2);
2060     fe_mul_ttt(&x2, &tmp1, &tmp0);
2061     fe_sub(&tmp1l, &tmp1, &tmp0);
2062     fe_sq_tl(&z2, &z2l);
2063     fe_mul121666(&z3, &tmp1l);
2064     fe_sq_tl(&x3, &x3l);
2065     fe_add(&tmp0l, &tmp0, &z3);
2066     fe_mul_ttt(&z3, &x1, &z2);
2067     fe_mul_tll(&z2, &tmp1l, &tmp0l);
2068   }
2069   // here pos=-1, so r=e, so to_xz (e*P) === if swap then (x3, z3) else (x2, z2)
2070   fe_cswap(&x2, &x3, swap);
2071   fe_cswap(&z2, &z3, swap);
2072 
2073   fe_invert(&z2, &z2);
2074   fe_mul_ttt(&x2, &x2, &z2);
2075   fe_tobytes(out, &x2);
2076 }
2077 
x25519_scalar_mult(uint8_t out[32],const uint8_t scalar[32],const uint8_t point[32])2078 static void x25519_scalar_mult(uint8_t out[32], const uint8_t scalar[32],
2079                                const uint8_t point[32]) {
2080 #if defined(BORINGSSL_X25519_NEON)
2081   if (CRYPTO_is_NEON_capable()) {
2082     x25519_NEON(out, scalar, point);
2083     return;
2084   }
2085 #elif defined(BORINGSSL_FE25519_ADX)
2086   if (CRYPTO_is_BMI1_capable() && CRYPTO_is_BMI2_capable() &&
2087       CRYPTO_is_ADX_capable()) {
2088     x25519_scalar_mult_adx(out, scalar, point);
2089     return;
2090   }
2091 #endif
2092 
2093   x25519_scalar_mult_generic(out, scalar, point);
2094 }
2095 
X25519_keypair(uint8_t out_public_value[32],uint8_t out_private_key[32])2096 void X25519_keypair(uint8_t out_public_value[32], uint8_t out_private_key[32]) {
2097   RAND_bytes(out_private_key, 32);
2098 
2099   // All X25519 implementations should decode scalars correctly (see
2100   // https://tools.ietf.org/html/rfc7748#section-5). However, if an
2101   // implementation doesn't then it might interoperate with random keys a
2102   // fraction of the time because they'll, randomly, happen to be correctly
2103   // formed.
2104   //
2105   // Thus we do the opposite of the masking here to make sure that our private
2106   // keys are never correctly masked and so, hopefully, any incorrect
2107   // implementations are deterministically broken.
2108   //
2109   // This does not affect security because, although we're throwing away
2110   // entropy, a valid implementation of scalarmult should throw away the exact
2111   // same bits anyway.
2112   out_private_key[0] |= ~248;
2113   out_private_key[31] &= ~64;
2114   out_private_key[31] |= ~127;
2115 
2116   X25519_public_from_private(out_public_value, out_private_key);
2117 }
2118 
X25519(uint8_t out_shared_key[32],const uint8_t private_key[32],const uint8_t peer_public_value[32])2119 int X25519(uint8_t out_shared_key[32], const uint8_t private_key[32],
2120            const uint8_t peer_public_value[32]) {
2121   static const uint8_t kZeros[32] = {0};
2122   x25519_scalar_mult(out_shared_key, private_key, peer_public_value);
2123   // The all-zero output results when the input is a point of small order.
2124   return constant_time_declassify_int(
2125              CRYPTO_memcmp(kZeros, out_shared_key, 32)) != 0;
2126 }
2127 
X25519_public_from_private(uint8_t out_public_value[32],const uint8_t private_key[32])2128 void X25519_public_from_private(uint8_t out_public_value[32],
2129                                 const uint8_t private_key[32]) {
2130 #if defined(BORINGSSL_X25519_NEON)
2131   if (CRYPTO_is_NEON_capable()) {
2132     static const uint8_t kMongomeryBasePoint[32] = {9};
2133     x25519_NEON(out_public_value, private_key, kMongomeryBasePoint);
2134     return;
2135   }
2136 #endif
2137 
2138   uint8_t e[32];
2139   OPENSSL_memcpy(e, private_key, 32);
2140   e[0] &= 248;
2141   e[31] &= 127;
2142   e[31] |= 64;
2143 
2144   ge_p3 A;
2145   x25519_ge_scalarmult_base(&A, e);
2146 
2147   // We only need the u-coordinate of the curve25519 point. The map is
2148   // u=(y+1)/(1-y). Since y=Y/Z, this gives u=(Z+Y)/(Z-Y).
2149   fe_loose zplusy, zminusy;
2150   fe zminusy_inv;
2151   fe_add(&zplusy, &A.Z, &A.Y);
2152   fe_sub(&zminusy, &A.Z, &A.Y);
2153   fe_loose_invert(&zminusy_inv, &zminusy);
2154   fe_mul_tlt(&zminusy_inv, &zplusy, &zminusy_inv);
2155   fe_tobytes(out_public_value, &zminusy_inv);
2156   CONSTTIME_DECLASSIFY(out_public_value, 32);
2157 }
2158