crypto/poly1305/poly1305.c

   1 /*
   2  * Copyright 2015-2016 The OpenSSL Project Authors. All Rights Reserved.
   3  *
   4  * Licensed under the OpenSSL license (the "License").  You may not use
   5  * this file except in compliance with the License.  You can obtain a copy
   6  * in the file LICENSE in the source distribution or at
   7  * https://www.openssl.org/source/license.html
   8  */
   9
  10 #include <stdlib.h>
  11 #include <string.h>
  12
  13 #include "internal/poly1305.h"
  14
  15 typedef void (*poly1305_blocks_f) (void *ctx, const unsigned char *inp,
  16                                    size_t len, unsigned int padbit);
  17 typedef void (*poly1305_emit_f) (void *ctx, unsigned char mac[16],
  18                                  const unsigned int nonce[4]);
  19
  20 struct poly1305_context {
  21     double opaque[24];  /* large enough to hold internal state, declared
  22                          * 'double' to ensure at least 64-bit invariant
  23                          * alignment across all platforms and
  24                          * configurations */
  25     unsigned int nonce[4];
  26     unsigned char data[POLY1305_BLOCK_SIZE];
  27     size_t num;
  28     struct {
  29         poly1305_blocks_f blocks;
  30         poly1305_emit_f emit;
  31     } func;
  32 };
  33
  34 size_t Poly1305_ctx_size ()
  35 {
  36     return sizeof(struct poly1305_context);
  37 }
  38
  39 /* pick 32-bit unsigned integer in little endian order */
  40 static unsigned int U8TOU32(const unsigned char *p)
  41 {
  42     return (((unsigned int)(p[0] & 0xff)) |
  43             ((unsigned int)(p[1] & 0xff) << 8) |
  44             ((unsigned int)(p[2] & 0xff) << 16) |
  45             ((unsigned int)(p[3] & 0xff) << 24));
  46 }
  47
  48 /*
  49  * Implementations can be classified by amount of significant bits in
  50  * words making up the multi-precision value, or in other words radix
  51  * or base of numerical representation, e.g. base 2^64, base 2^32,
  52  * base 2^26. Complementary characteristic is how wide is the result of
  53  * multiplication of pair of digits, e.g. it would take 128 bits to
  54  * accommodate multiplication result in base 2^64 case. These are used
  55  * interchangeably. To describe implementation that is. But interface
  56  * is designed to isolate this so that low-level primitives implemented
  57  * in assembly can be self-contained/self-coherent.
  58  */
  59 #ifndef POLY1305_ASM
  60 /*
  61  * Even though there is __int128 reference implementation targeting
  62  * 64-bit platforms provided below, it's not obvious that it's optimal
  63  * choice for every one of them. Depending on instruction set overall
  64  * amount of instructions can be comparable to one in __int64
  65  * implementation. Amount of multiplication instructions would be lower,
  66  * but not necessarily overall. And in out-of-order execution context,
  67  * it is the latter that can be crucial...
  68  *
  69  * On related note. Poly1305 author, D. J. Bernstein, discusses and
  70  * provides floating-point implementations of the algorithm in question.
  71  * It made a lot of sense by the time of introduction, because most
  72  * then-modern processors didn't have pipelined integer multiplier.
  73  * [Not to mention that some had non-constant timing for integer
  74  * multiplications.] Floating-point instructions on the other hand could
  75  * be issued every cycle, which allowed to achieve better performance.
  76  * Nowadays, with SIMD and/or out-or-order execution, shared or
  77  * even emulated FPU, it's more complicated, and floating-point
  78  * implementation is not necessarily optimal choice in every situation,
  79  * rather contrary...
  80  *
  81  *                                              <appro@openssl.org>
  82  */
  83
  84 typedef unsigned int u32;
  85
  86 /*
  87  * poly1305_blocks processes a multiple of POLY1305_BLOCK_SIZE blocks
  88  * of |inp| no longer than |len|. Behaviour for |len| not divisible by
  89  * block size is unspecified in general case, even though in reference
  90  * implementation the trailing chunk is simply ignored. Per algorithm
  91  * specification, every input block, complete or last partial, is to be
  92  * padded with a bit past most significant byte. The latter kind is then
  93  * padded with zeros till block size. This last partial block padding
  94  * is caller(*)'s responsibility, and because of this the last partial
  95  * block is always processed with separate call with |len| set to
  96  * POLY1305_BLOCK_SIZE and |padbit| to 0. In all other cases |padbit|
  97  * should be set to 1 to perform implicit padding with 128th bit.
  98  * poly1305_blocks does not actually check for this constraint though,
  99  * it's caller(*)'s responsibility to comply.
 100  *
 101  * (*)  In the context "caller" is not application code, but higher
 102  *      level Poly1305_* from this very module, so that quirks are
 103  *      handled locally.
 104  */
 105 static void
 106 poly1305_blocks(void *ctx, const unsigned char *inp, size_t len, u32 padbit);
 107
 108 /*
 109  * Type-agnostic "rip-off" from constant_time_locl.h
 110  */
 111 # define CONSTANT_TIME_CARRY(a,b) ( \
 112          (a ^ ((a ^ b) | ((a - b) ^ b))) >> (sizeof(a) * 8 - 1) \
 113          )
 114
 115 # if !defined(PEDANTIC) && \
 116      (defined(__SIZEOF_INT128__) && __SIZEOF_INT128__==16) && \
 117      (defined(__SIZEOF_LONG__) && __SIZEOF_LONG__==8)
 118
 119 typedef unsigned long u64;
 120 typedef unsigned __int128 u128;
 121
 122 typedef struct {
 123     u64 h[3];
 124     u64 r[2];
 125 } poly1305_internal;
 126
 127 /* pick 32-bit unsigned integer in little endian order */
 128 static u64 U8TOU64(const unsigned char *p)
 129 {
 130     return (((u64)(p[0] & 0xff)) |
 131             ((u64)(p[1] & 0xff) << 8) |
 132             ((u64)(p[2] & 0xff) << 16) |
 133             ((u64)(p[3] & 0xff) << 24) |
 134             ((u64)(p[4] & 0xff) << 32) |
 135             ((u64)(p[5] & 0xff) << 40) |
 136             ((u64)(p[6] & 0xff) << 48) |
 137             ((u64)(p[7] & 0xff) << 56));
 138 }
 139
 140 /* store a 32-bit unsigned integer in little endian */
 141 static void U64TO8(unsigned char *p, u64 v)
 142 {
 143     p[0] = (unsigned char)((v) & 0xff);
 144     p[1] = (unsigned char)((v >> 8) & 0xff);
 145     p[2] = (unsigned char)((v >> 16) & 0xff);
 146     p[3] = (unsigned char)((v >> 24) & 0xff);
 147     p[4] = (unsigned char)((v >> 32) & 0xff);
 148     p[5] = (unsigned char)((v >> 40) & 0xff);
 149     p[6] = (unsigned char)((v >> 48) & 0xff);
 150     p[7] = (unsigned char)((v >> 56) & 0xff);
 151 }
 152
 153 static void poly1305_init(void *ctx, const unsigned char key[16])
 154 {
 155     poly1305_internal *st = (poly1305_internal *) ctx;
 156
 157     /* h = 0 */
 158     st->h[0] = 0;
 159     st->h[1] = 0;
 160     st->h[2] = 0;
 161
 162     /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
 163     st->r[0] = U8TOU64(&key[0]) & 0x0ffffffc0fffffff;
 164     st->r[1] = U8TOU64(&key[8]) & 0x0ffffffc0ffffffc;
 165 }
 166
 167 static void
 168 poly1305_blocks(void *ctx, const unsigned char *inp, size_t len, u32 padbit)
 169 {
 170     poly1305_internal *st = (poly1305_internal *)ctx;
 171     u64 r0, r1;
 172     u64 s1;
 173     u64 h0, h1, h2, c;
 174     u128 d0, d1;
 175
 176     r0 = st->r[0];
 177     r1 = st->r[1];
 178
 179     s1 = r1 + (r1 >> 2);
 180
 181     h0 = st->h[0];
 182     h1 = st->h[1];
 183     h2 = st->h[2];
 184
 185     while (len >= POLY1305_BLOCK_SIZE) {
 186         /* h += m[i] */
 187         h0 = (u64)(d0 = (u128)h0 + U8TOU64(inp + 0));
 188         h1 = (u64)(d1 = (u128)h1 + (d0 >> 64) + U8TOU64(inp + 8));
 189         /*
 190          * padbit can be zero only when original len was
 191          * POLY1306_BLOCK_SIZE, but we don't check
 192          */
 193         h2 += (u64)(d1 >> 64) + padbit;
 194
 195         /* h *= r "%" p, where "%" stands for "partial remainder" */
 196         d0 = ((u128)h0 * r0) +
 197              ((u128)h1 * s1);
 198         d1 = ((u128)h0 * r1) +
 199              ((u128)h1 * r0) +
 200              (h2 * s1);
 201         h2 = (h2 * r0);
 202
 203         /* last reduction step: */
 204         /* a) h2:h0 = h2<<128 + d1<<64 + d0 */
 205         h0 = (u64)d0;
 206         h1 = (u64)(d1 += d0 >> 64);
 207         h2 += (u64)(d1 >> 64);
 208         /* b) (h2:h0 += (h2:h0>>130) * 5) %= 2^130 */
 209         c = (h2 >> 2) + (h2 & ~3UL);
 210         h2 &= 3;
 211         h0 += c;
 212         h1 += (c = CONSTANT_TIME_CARRY(h0,c));
 213         h2 += CONSTANT_TIME_CARRY(h1,c);
 214         /*
 215          * Occasional overflows to 3rd bit of h2 are taken care of
 216          * "naturally". If after this point we end up at the top of
 217          * this loop, then the overflow bit will be accounted for
 218          * in next iteration. If we end up in poly1305_emit, then
 219          * comparison to modulus below will still count as "carry
 220          * into 131st bit", so that properly reduced value will be
 221          * picked in conditional move.
 222          */
 223
 224         inp += POLY1305_BLOCK_SIZE;
 225         len -= POLY1305_BLOCK_SIZE;
 226     }
 227
 228     st->h[0] = h0;
 229     st->h[1] = h1;
 230     st->h[2] = h2;
 231 }
 232
 233 static void poly1305_emit(void *ctx, unsigned char mac[16],
 234                           const u32 nonce[4])
 235 {
 236     poly1305_internal *st = (poly1305_internal *) ctx;
 237     u64 h0, h1, h2;
 238     u64 g0, g1, g2;
 239     u128 t;
 240     u64 mask;
 241
 242     h0 = st->h[0];
 243     h1 = st->h[1];
 244     h2 = st->h[2];
 245
 246     /* compare to modulus by computing h + -p */
 247     g0 = (u64)(t = (u128)h0 + 5);
 248     g1 = (u64)(t = (u128)h1 + (t >> 64));
 249     g2 = h2 + (u64)(t >> 64);
 250
 251     /* if there was carry into 131st bit, h1:h0 = g1:g0 */
 252     mask = 0 - (g2 >> 2);
 253     g0 &= mask;
 254     g1 &= mask;
 255     mask = ~mask;
 256     h0 = (h0 & mask) | g0;
 257     h1 = (h1 & mask) | g1;
 258
 259     /* mac = (h + nonce) % (2^128) */
 260     h0 = (u64)(t = (u128)h0 + nonce[0] + ((u64)nonce[1]<<32));
 261     h1 = (u64)(t = (u128)h1 + nonce[2] + ((u64)nonce[3]<<32) + (t >> 64));
 262
 263     U64TO8(mac + 0, h0);
 264     U64TO8(mac + 8, h1);
 265 }
 266
 267 # else
 268
 269 #  if defined(_WIN32) && !defined(__MINGW32__)
 270 typedef unsigned __int64 u64;
 271 #  elif defined(__arch64__)
 272 typedef unsigned long u64;
 273 #  else
 274 typedef unsigned long long u64;
 275 #  endif
 276
 277 typedef struct {
 278     u32 h[5];
 279     u32 r[4];
 280 } poly1305_internal;
 281
 282 /* store a 32-bit unsigned integer in little endian */
 283 static void U32TO8(unsigned char *p, unsigned int v)
 284 {
 285     p[0] = (unsigned char)((v) & 0xff);
 286     p[1] = (unsigned char)((v >> 8) & 0xff);
 287     p[2] = (unsigned char)((v >> 16) & 0xff);
 288     p[3] = (unsigned char)((v >> 24) & 0xff);
 289 }
 290
 291 static void poly1305_init(void *ctx, const unsigned char key[16])
 292 {
 293     poly1305_internal *st = (poly1305_internal *) ctx;
 294
 295     /* h = 0 */
 296     st->h[0] = 0;
 297     st->h[1] = 0;
 298     st->h[2] = 0;
 299     st->h[3] = 0;
 300     st->h[4] = 0;
 301
 302     /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
 303     st->r[0] = U8TOU32(&key[0]) & 0x0fffffff;
 304     st->r[1] = U8TOU32(&key[4]) & 0x0ffffffc;
 305     st->r[2] = U8TOU32(&key[8]) & 0x0ffffffc;
 306     st->r[3] = U8TOU32(&key[12]) & 0x0ffffffc;
 307 }
 308
 309 static void
 310 poly1305_blocks(void *ctx, const unsigned char *inp, size_t len, u32 padbit)
 311 {
 312     poly1305_internal *st = (poly1305_internal *)ctx;
 313     u32 r0, r1, r2, r3;
 314     u32 s1, s2, s3;
 315     u32 h0, h1, h2, h3, h4, c;
 316     u64 d0, d1, d2, d3;
 317
 318     r0 = st->r[0];
 319     r1 = st->r[1];
 320     r2 = st->r[2];
 321     r3 = st->r[3];
 322
 323     s1 = r1 + (r1 >> 2);
 324     s2 = r2 + (r2 >> 2);
 325     s3 = r3 + (r3 >> 2);
 326
 327     h0 = st->h[0];
 328     h1 = st->h[1];
 329     h2 = st->h[2];
 330     h3 = st->h[3];
 331     h4 = st->h[4];
 332
 333     while (len >= POLY1305_BLOCK_SIZE) {
 334         /* h += m[i] */
 335         h0 = (u32)(d0 = (u64)h0 + U8TOU32(inp + 0));
 336         h1 = (u32)(d1 = (u64)h1 + (d0 >> 32) + U8TOU32(inp + 4));
 337         h2 = (u32)(d2 = (u64)h2 + (d1 >> 32) + U8TOU32(inp + 8));
 338         h3 = (u32)(d3 = (u64)h3 + (d2 >> 32) + U8TOU32(inp + 12));
 339         h4 += (u32)(d3 >> 32) + padbit;
 340
 341         /* h *= r "%" p, where "%" stands for "partial remainder" */
 342         d0 = ((u64)h0 * r0) +
 343              ((u64)h1 * s3) +
 344              ((u64)h2 * s2) +
 345              ((u64)h3 * s1);
 346         d1 = ((u64)h0 * r1) +
 347              ((u64)h1 * r0) +
 348              ((u64)h2 * s3) +
 349              ((u64)h3 * s2) +
 350              (h4 * s1);
 351         d2 = ((u64)h0 * r2) +
 352              ((u64)h1 * r1) +
 353              ((u64)h2 * r0) +
 354              ((u64)h3 * s3) +
 355              (h4 * s2);
 356         d3 = ((u64)h0 * r3) +
 357              ((u64)h1 * r2) +
 358              ((u64)h2 * r1) +
 359              ((u64)h3 * r0) +
 360              (h4 * s3);
 361         h4 = (h4 * r0);
 362
 363         /* last reduction step: */
 364         /* a) h4:h0 = h4<<128 + d3<<96 + d2<<64 + d1<<32 + d0 */
 365         h0 = (u32)d0;
 366         h1 = (u32)(d1 += d0 >> 32);
 367         h2 = (u32)(d2 += d1 >> 32);
 368         h3 = (u32)(d3 += d2 >> 32);
 369         h4 += (u32)(d3 >> 32);
 370         /* b) (h4:h0 += (h4:h0>>130) * 5) %= 2^130 */
 371         c = (h4 >> 2) + (h4 & ~3U);
 372         h4 &= 3;
 373         h0 += c;
 374         h1 += (c = CONSTANT_TIME_CARRY(h0,c));
 375         h2 += (c = CONSTANT_TIME_CARRY(h1,c));
 376         h3 += (c = CONSTANT_TIME_CARRY(h2,c));
 377         h4 += CONSTANT_TIME_CARRY(h3,c);
 378         /*
 379          * Occasional overflows to 3rd bit of h4 are taken care of
 380          * "naturally". If after this point we end up at the top of
 381          * this loop, then the overflow bit will be accounted for
 382          * in next iteration. If we end up in poly1305_emit, then
 383          * comparison to modulus below will still count as "carry
 384          * into 131st bit", so that properly reduced value will be
 385          * picked in conditional move.
 386          */
 387
 388         inp += POLY1305_BLOCK_SIZE;
 389         len -= POLY1305_BLOCK_SIZE;
 390     }
 391
 392     st->h[0] = h0;
 393     st->h[1] = h1;
 394     st->h[2] = h2;
 395     st->h[3] = h3;
 396     st->h[4] = h4;
 397 }
 398
 399 static void poly1305_emit(void *ctx, unsigned char mac[16],
 400                           const u32 nonce[4])
 401 {
 402     poly1305_internal *st = (poly1305_internal *) ctx;
 403     u32 h0, h1, h2, h3, h4;
 404     u32 g0, g1, g2, g3, g4;
 405     u64 t;
 406     u32 mask;
 407
 408     h0 = st->h[0];
 409     h1 = st->h[1];
 410     h2 = st->h[2];
 411     h3 = st->h[3];
 412     h4 = st->h[4];
 413
 414     /* compare to modulus by computing h + -p */
 415     g0 = (u32)(t = (u64)h0 + 5);
 416     g1 = (u32)(t = (u64)h1 + (t >> 32));
 417     g2 = (u32)(t = (u64)h2 + (t >> 32));
 418     g3 = (u32)(t = (u64)h3 + (t >> 32));
 419     g4 = h4 + (u32)(t >> 32);
 420
 421     /* if there was carry into 131st bit, h3:h0 = g3:g0 */
 422     mask = 0 - (g4 >> 2);
 423     g0 &= mask;
 424     g1 &= mask;
 425     g2 &= mask;
 426     g3 &= mask;
 427     mask = ~mask;
 428     h0 = (h0 & mask) | g0;
 429     h1 = (h1 & mask) | g1;
 430     h2 = (h2 & mask) | g2;
 431     h3 = (h3 & mask) | g3;
 432
 433     /* mac = (h + nonce) % (2^128) */
 434     h0 = (u32)(t = (u64)h0 + nonce[0]);
 435     h1 = (u32)(t = (u64)h1 + (t >> 32) + nonce[1]);
 436     h2 = (u32)(t = (u64)h2 + (t >> 32) + nonce[2]);
 437     h3 = (u32)(t = (u64)h3 + (t >> 32) + nonce[3]);
 438
 439     U32TO8(mac + 0, h0);
 440     U32TO8(mac + 4, h1);
 441     U32TO8(mac + 8, h2);
 442     U32TO8(mac + 12, h3);
 443 }
 444 # endif
 445 #else
 446 int poly1305_init(void *ctx, const unsigned char key[16], void *func);
 447 void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
 448                      unsigned int padbit);
 449 void poly1305_emit(void *ctx, unsigned char mac[16],
 450                    const unsigned int nonce[4]);
 451 #endif
 452
 453 void Poly1305_Init(POLY1305 *ctx, const unsigned char key[32])
 454 {
 455     ctx->nonce[0] = U8TOU32(&key[16]);
 456     ctx->nonce[1] = U8TOU32(&key[20]);
 457     ctx->nonce[2] = U8TOU32(&key[24]);
 458     ctx->nonce[3] = U8TOU32(&key[28]);
 459
 460 #ifndef POLY1305_ASM
 461     poly1305_init(ctx->opaque, key);
 462 #else
 463     /*
 464      * Unlike reference poly1305_init assembly counterpart is expected
 465      * to return a value: non-zero if it initializes ctx->func, and zero
 466      * otherwise. Latter is to simplify assembly in cases when there no
 467      * multiple code paths to switch between.
 468      */
 469     if (!poly1305_init(ctx->opaque, key, &ctx->func)) {
 470         ctx->func.blocks = poly1305_blocks;
 471         ctx->func.emit = poly1305_emit;
 472     }
 473 #endif
 474
 475     ctx->num = 0;
 476
 477 }
 478
 479 #ifdef POLY1305_ASM
 480 /*
 481  * This "eclipses" poly1305_blocks and poly1305_emit, but it's
 482  * conscious choice imposed by -Wshadow compiler warnings.
 483  */
 484 # define poly1305_blocks (*poly1305_blocks_p)
 485 # define poly1305_emit   (*poly1305_emit_p)
 486 #endif
 487
 488 void Poly1305_Update(POLY1305 *ctx, const unsigned char *inp, size_t len)
 489 {
 490 #ifdef POLY1305_ASM
 491     /*
 492      * As documented, poly1305_blocks is never called with input
 493      * longer than single block and padbit argument set to 0. This
 494      * property is fluently used in assembly modules to optimize
 495      * padbit handling on loop boundary.
 496      */
 497     poly1305_blocks_f poly1305_blocks_p = ctx->func.blocks;
 498 #endif
 499     size_t rem, num;
 500
 501     if ((num = ctx->num)) {
 502         rem = POLY1305_BLOCK_SIZE - num;
 503         if (len >= rem) {
 504             memcpy(ctx->data + num, inp, rem);
 505             poly1305_blocks(ctx->opaque, ctx->data, POLY1305_BLOCK_SIZE, 1);
 506             inp += rem;
 507             len -= rem;
 508         } else {
 509             /* Still not enough data to process a block. */
 510             memcpy(ctx->data + num, inp, len);
 511             ctx->num = num + len;
 512             return;
 513         }
 514     }
 515
 516     rem = len % POLY1305_BLOCK_SIZE;
 517     len -= rem;
 518
 519     if (len >= POLY1305_BLOCK_SIZE) {
 520         poly1305_blocks(ctx->opaque, inp, len, 1);
 521         inp += len;
 522     }
 523
 524     if (rem)
 525         memcpy(ctx->data, inp, rem);
 526
 527     ctx->num = rem;
 528 }
 529
 530 void Poly1305_Final(POLY1305 *ctx, unsigned char mac[16])
 531 {
 532 #ifdef POLY1305_ASM
 533     poly1305_blocks_f poly1305_blocks_p = ctx->func.blocks;
 534     poly1305_emit_f poly1305_emit_p = ctx->func.emit;
 535 #endif
 536     size_t num;
 537
 538     if ((num = ctx->num)) {
 539         ctx->data[num++] = 1;   /* pad bit */
 540         while (num < POLY1305_BLOCK_SIZE)
 541             ctx->data[num++] = 0;
 542         poly1305_blocks(ctx->opaque, ctx->data, POLY1305_BLOCK_SIZE, 0);
 543     }
 544
 545     poly1305_emit(ctx->opaque, mac, ctx->nonce);
 546
 547     /* zero out the state */
 548     memset(ctx, 0, sizeof(*ctx));
 549 }
 550
 551 #ifdef SELFTEST
 552 #include <stdio.h>
 553
 554 struct poly1305_test {
 555     const char *inputhex;
 556     const char *keyhex;
 557     const char *outhex;
 558 };
 559
 560 static const struct poly1305_test poly1305_tests[] = {
 561     /*
 562      * RFC7539
 563      */
 564     {
 565      "43727970746f6772617068696320466f72756d2052657365617263682047726f"
 566      "7570",
 567      "85d6be7857556d337f4452fe42d506a8""0103808afb0db2fd4abff6af4149f51b",
 568      "a8061dc1305136c6c22b8baf0c0127a9"
 569     },
 570     /*
 571      * test vectors from "The Poly1305-AES message-authentication code"
 572      */
 573     {
 574      "f3f6",
 575      "851fc40c3467ac0be05cc20404f3f700""580b3b0f9447bb1e69d095b5928b6dbc",
 576      "f4c633c3044fc145f84f335cb81953de"
 577     },
 578     {
 579      "",
 580      "a0f3080000f46400d0c7e9076c834403""dd3fab2251f11ac759f0887129cc2ee7",
 581      "dd3fab2251f11ac759f0887129cc2ee7"
 582     },
 583     {
 584      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136",
 585      "48443d0bb0d21109c89a100b5ce2c208""83149c69b561dd88298a1798b10716ef",
 586      "0ee1c16bb73f0f4fd19881753c01cdbe"
 587     },
 588     {
 589      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 590      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9",
 591      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 592      "5154ad0d2cb26e01274fc51148491f1b"
 593     },
 594     /*
 595      * self-generated vectors exercise "significant" lengths, such that
 596      * are handled by different code paths
 597      */
 598     {
 599      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 600      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af",
 601      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 602      "812059a5da198637cac7c4a631bee466"
 603     },
 604     {
 605      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 606      "990c62e48b8018b2c3e4a0fa3134cb67",
 607      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 608      "5b88d7f6228b11e2e28579a5c0c1f761"
 609     },
 610     {
 611      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 612      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 613      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136",
 614      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 615      "bbb613b2b6d753ba07395b916aaece15"
 616     },
 617     {
 618      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 619      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 620      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 621      "663cea190ffb83d89593f3f476b6bc24",
 622      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 623      "c794d7057d1778c4bbee0a39b3d97342"
 624     },
 625     {
 626      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 627      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 628      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 629      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136",
 630      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 631      "ffbcb9b371423152d7fca5ad042fbaa9"
 632     },
 633     {
 634      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 635      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 636      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 637      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 638      "812059a5da198637cac7c4a631bee466",
 639      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 640      "069ed6b8ef0f207b3e243bb1019fe632"
 641     },
 642     {
 643      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 644      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 645      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 646      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 647      "812059a5da198637cac7c4a631bee4665b88d7f6228b11e2e28579a5c0c1f761",
 648      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 649      "cca339d9a45fa2368c2c68b3a4179133"
 650     },
 651     {
 652      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 653      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 654      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 655      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 656      "812059a5da198637cac7c4a631bee4665b88d7f6228b11e2e28579a5c0c1f761"
 657      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 658      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 659      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 660      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136",
 661      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 662      "53f6e828a2f0fe0ee815bf0bd5841a34"
 663     },
 664     {
 665      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 666      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 667      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 668      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 669      "812059a5da198637cac7c4a631bee4665b88d7f6228b11e2e28579a5c0c1f761"
 670      "ab0812724a7f1e342742cbed374d94d136c6b8795d45b3819830f2c04491faf0"
 671      "990c62e48b8018b2c3e4a0fa3134cb67fa83e158c994d961c4cb21095c1bf9af"
 672      "48443d0bb0d21109c89a100b5ce2c20883149c69b561dd88298a1798b10716ef"
 673      "663cea190ffb83d89593f3f476b6bc24d7e679107ea26adb8caf6652d0656136"
 674      "812059a5da198637cac7c4a631bee4665b88d7f6228b11e2e28579a5c0c1f761",
 675      "12976a08c4426d0ce8a82407c4f48207""80f8c20aa71202d1e29179cbcb555a57",
 676      "b846d44e9bbd53cedffbfbb6b7fa4933"
 677     },
 678     /*
 679      * 4th power of the key spills to 131th bit in SIMD key setup
 680      */
 681     {
 682      "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
 683      "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
 684      "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
 685      "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
 686      "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
 687      "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
 688      "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
 689      "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff",
 690      "ad628107e8351d0f2c231a05dc4a4106""00000000000000000000000000000000",
 691      "07145a4c02fe5fa32036de68fabe9066"
 692     },
 693     {
 694     /*
 695      * poly1305_ieee754.c failed this in final stage
 696      */
 697      "842364e156336c0998b933a6237726180d9e3fdcbde4cd5d17080fc3beb49614"
 698      "d7122c037463ff104d73f19c12704628d417c4c54a3fe30d3c3d7714382d43b0"
 699      "382a50a5dee54be844b076e8df88201a1cd43b90eb21643fa96f39b518aa8340"
 700      "c942ff3c31baf7c9bdbf0f31ae3fa096bf8c63030609829fe72e179824890bc8"
 701      "e08c315c1cce2a83144dbbff09f74e3efc770b54d0984a8f19b14719e6363564"
 702      "1d6b1eedf63efbf080e1783d32445412114c20de0b837a0dfa33d6b82825fff4"
 703      "4c9a70ea54ce47f07df698e6b03323b53079364a5fc3e9dd034392bdde86dccd"
 704      "da94321c5e44060489336cb65bf3989c36f7282c2f5d2b882c171e74",
 705      "95d5c005503e510d8cd0aa072c4a4d06""6eabc52d11653df47fbf63ab198bcc26",
 706      "f248312e578d9d58f8b7bb4d19105431"
 707     },
 708     /*
 709      * AVX2 in poly1305-x86.pl failed this with 176+32 split
 710      */
 711     {
 712     "248ac31085b6c2adaaa38259a0d7192c5c35d1bb4ef39ad94c38d1c82479e2dd"
 713     "2159a077024b0589bc8a20101b506f0a1ad0bbab76e83a83f1b94be6beae74e8"
 714     "74cab692c5963a75436b776121ec9f62399a3e66b2d22707dae81933b6277f3c"
 715     "8516bcbe26dbbd86f373103d7cf4cad1888c952118fbfbd0d7b4bedc4ae4936a"
 716     "ff91157e7aa47c54442ea78d6ac251d324a0fbe49d89cc3521b66d16e9c66a37"
 717     "09894e4eb0a4eedc4ae19468e66b81f2"
 718     "71351b1d921ea551047abcc6b87a901fde7db79fa1818c11336dbc07244a40eb",
 719     "000102030405060708090a0b0c0d0e0f""00000000000000000000000000000000",
 720     "bc939bc5281480fa99c6d68c258ec42f"
 721     },
 722     /*
 723      * test vectors from Google
 724      */
 725     {
 726      "",
 727      "c8afaac331ee372cd6082de134943b17""4710130e9f6fea8d72293850a667d86c",
 728      "4710130e9f6fea8d72293850a667d86c",
 729     },
 730     {
 731      "48656c6c6f20776f726c6421",
 732      "746869732069732033322d6279746520""6b657920666f7220506f6c7931333035",
 733      "a6f745008f81c916a20dcc74eef2b2f0"
 734     },
 735     {
 736      "0000000000000000000000000000000000000000000000000000000000000000",
 737      "746869732069732033322d6279746520""6b657920666f7220506f6c7931333035",
 738      "49ec78090e481ec6c26b33b91ccc0307"
 739     },
 740     {
 741      "89dab80b7717c1db5db437860a3f70218e93e1b8f461fb677f16f35f6f87e2a9"
 742      "1c99bc3a47ace47640cc95c345be5ecca5a3523c35cc01893af0b64a62033427"
 743      "0372ec12482d1b1e363561698a578b359803495bb4e2ef1930b17a5190b580f1"
 744      "41300df30adbeca28f6427a8bc1a999fd51c554a017d095d8c3e3127daf9f595",
 745      "2d773be37adb1e4d683bf0075e79c4ee""037918535a7f99ccb7040fb5f5f43aea",
 746      "c85d15ed44c378d6b00e23064c7bcd51"
 747     },
 748     {
 749      "000000000000000b1703030200000000"
 750      "06db1f1f368d696a810a349c0c714c9a5e7850c2407d721acded95e018d7a852"
 751      "66a6e1289cdb4aeb18da5ac8a2b0026d24a59ad485227f3eaedbb2e7e35e1c66"
 752      "cd60f9abf716dcc9ac42682dd7dab287a7024c4eefc321cc0574e16793e37cec"
 753      "03c5bda42b54c114a80b57af26416c7be742005e20855c73e21dc8e2edc9d435"
 754      "cb6f6059280011c270b71570051c1c9b3052126620bc1e2730fa066c7a509d53"
 755      "c60e5ae1b40aa6e39e49669228c90eecb4a50db32a50bc49e90b4f4b359a1dfd"
 756      "11749cd3867fcf2fb7bb6cd4738f6a4ad6f7ca5058f7618845af9f020f6c3b96"
 757      "7b8f4cd4a91e2813b507ae66f2d35c18284f7292186062e10fd5510d18775351"
 758      "ef334e7634ab4743f5b68f49adcab384d3fd75f7390f4006ef2a295c8c7a076a"
 759      "d54546cd25d2107fbe1436c840924aaebe5b370893cd63d1325b8616fc481088"
 760      "6bc152c53221b6df373119393255ee72bcaa880174f1717f9184fa91646f17a2"
 761      "4ac55d16bfddca9581a92eda479201f0edbf633600d6066d1ab36d5d2415d713"
 762      "51bbcd608a25108d25641992c1f26c531cf9f90203bc4cc19f5927d834b0a471"
 763      "16d3884bbb164b8ec883d1ac832e56b3918a98601a08d171881541d594db399c"
 764      "6ae6151221745aec814c45b0b05b565436fd6f137aa10a0c0b643761dbd6f9a9"
 765      "dcb99b1a6e690854ce0769cde39761d82fcdec15f0d92d7d8e94ade8eb83fbe0",
 766      "99e5822dd4173c995e3dae0ddefb9774""3fde3b080134b39f76e9bf8d0e88d546",
 767      "2637408fe13086ea73f971e3425e2820"
 768     },
 769     /*
 770      * test vectors from Hanno Böck
 771      */
 772     {
 773      "cccccccccccccccccccccccccccccccccccccccccccccccccc80cccccccccccc"
 774      "cccccccccccccccccccccccccccccccccccccccccccccccccccccccccecccccc"
 775      "ccccccccccccccccccccccccccccccc5cccccccccccccccccccccccccccccccc"
 776      "cccccccccce3cccccccccccccccccccccccccccccccccccccccccccccccccccc"
 777      "ccccccccaccccccccccccccccccccce6cccccccccc000000afcccccccccccccc"
 778      "ccccfffffff50000000000000000000000000000000000000000000000000000"
 779      "00ffffffe7000000000000000000000000000000000000000000000000000000"
 780      "0000000000000000000000000000000000000000000000000000719205a8521d"
 781      "fc",
 782      "7f1b0264000000000000000000000000""0000000000000000cccccccccccccccc",
 783      "8559b876eceed66eb37798c0457baff9"
 784     },
 785     {
 786      "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa0000000000"
 787      "00000000800264",
 788      "e0001600000000000000000000000000""0000aaaaaaaaaaaaaaaaaaaaaaaaaaaa",
 789      "00bd1258978e205444c9aaaa82006fed"
 790     },
 791     {
 792      "02fc",
 793      "0c0c0c0c0c0c0c0c0c0c0c0c0c0c0c0c""0c0c0c0c0c0c0c0c0c0c0c0c0c0c0c0c",
 794      "06120c0c0c0c0c0c0c0c0c0c0c0c0c0c"
 795     },
 796     {
 797      "7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b"
 798      "7b7b7b7b7b7b7a7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b"
 799      "7b7b5c7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b"
 800      "7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b6e7b007b7b7b7b7b7b7b7b7b"
 801      "7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7a7b7b7b7b7b7b7b7b7b7b7b7b"
 802      "7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b5c7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b"
 803      "7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b7b"
 804      "7b6e7b001300000000b300000000000000000000000000000000000000000000"
 805      "f20000000000000000000000000000000000002000efff000900000000000000"
 806      "0000000000100000000009000000640000000000000000000000001300000000"
 807      "b300000000000000000000000000000000000000000000f20000000000000000"
 808      "000000000000000000002000efff00090000000000000000007a000010000000"
 809      "000900000064000000000000000000000000000000000000000000000000fc",
 810      "00ff0000000000000000000000000000""00000000001e00000000000000007b7b",
 811      "33205bbf9e9f8f7212ab9e2ab9b7e4a5"
 812     },
 813     {
 814      "7777777777777777777777777777777777777777777777777777777777777777"
 815      "7777777777777777777777777777777777777777777777777777777777777777"
 816      "777777777777777777777777ffffffe9e9acacacacacacacacacacac0000acac"
 817      "ec0100acacac2caca2acacacacacacacacacacac64f2",
 818      "0000007f0000007f0100002000000000""0000cf77777777777777777777777777",
 819      "02ee7c8c546ddeb1a467e4c3981158b9"
 820     },
 821     /*
 822      * test vectors from Andrew Moon
 823      */
 824     {   /* nacl */
 825      "8e993b9f48681273c29650ba32fc76ce48332ea7164d96a4476fb8c531a1186a"
 826      "c0dfc17c98dce87b4da7f011ec48c97271d2c20f9b928fe2270d6fb863d51738"
 827      "b48eeee314a7cc8ab932164548e526ae90224368517acfeabd6bb3732bc0e9da"
 828      "99832b61ca01b6de56244a9e88d5f9b37973f622a43d14a6599b1f654cb45a74"
 829      "e355a5",
 830      "eea6a7251c1e72916d11c2cb214d3c25""2539121d8e234e652d651fa4c8cff880",
 831      "f3ffc7703f9400e52a7dfb4b3d3305d9"
 832     },
 833     {   /* wrap 2^130-5 */
 834      "ffffffffffffffffffffffffffffffff",
 835      "02000000000000000000000000000000""00000000000000000000000000000000",
 836      "03000000000000000000000000000000"
 837     },
 838     {   /* wrap 2^128 */
 839      "02000000000000000000000000000000",
 840      "02000000000000000000000000000000""ffffffffffffffffffffffffffffffff",
 841      "03000000000000000000000000000000"
 842     },
 843     {   /* limb carry */
 844      "fffffffffffffffffffffffffffffffff0ffffffffffffffffffffffffffffff"
 845      "11000000000000000000000000000000",
 846      "01000000000000000000000000000000""00000000000000000000000000000000",
 847      "05000000000000000000000000000000"
 848     },
 849     {   /* 2^130-5 */
 850      "fffffffffffffffffffffffffffffffffbfefefefefefefefefefefefefefefe"
 851      "01010101010101010101010101010101",
 852      "01000000000000000000000000000000""00000000000000000000000000000000",
 853      "00000000000000000000000000000000"
 854     },
 855     {   /* 2^130-6 */
 856      "fdffffffffffffffffffffffffffffff",
 857      "02000000000000000000000000000000""00000000000000000000000000000000",
 858      "faffffffffffffffffffffffffffffff"
 859     },
 860     {   /* 5*H+L reduction intermediate */
 861      "e33594d7505e43b900000000000000003394d7505e4379cd0100000000000000"
 862      "0000000000000000000000000000000001000000000000000000000000000000",
 863      "01000000000000000400000000000000""00000000000000000000000000000000",
 864      "14000000000000005500000000000000"
 865     },
 866     {   /* 5*H+L reduction final */
 867      "e33594d7505e43b900000000000000003394d7505e4379cd0100000000000000"
 868      "00000000000000000000000000000000",
 869      "01000000000000000400000000000000""00000000000000000000000000000000",
 870      "13000000000000000000000000000000"
 871     }
 872 };
 873
 874 static unsigned char hex_digit(char h)
 875 {
 876     int i = OPENSSL_hexchar2int(h);
 877
 878     if (i < 0)
 879         abort();
 880     return i;
 881 }
 882
 883 static void hex_decode(unsigned char *out, const char *hex)
 884 {
 885     size_t j = 0;
 886
 887     while (*hex != 0) {
 888         unsigned char v = hex_digit(*hex++);
 889         v <<= 4;
 890         v |= hex_digit(*hex++);
 891         out[j++] = v;
 892     }
 893 }
 894
 895 static void hexdump(unsigned char *a, size_t len)
 896 {
 897     size_t i;
 898
 899     for (i = 0; i < len; i++)
 900         printf("%02x", a[i]);
 901 }
 902
 903 int main()
 904 {
 905     static const unsigned num_tests =
 906         sizeof(poly1305_tests) / sizeof(struct poly1305_test);
 907     unsigned i;
 908     unsigned char key[32], out[16], expected[16];
 909     POLY1305 poly1305;
 910
 911     for (i = 0; i < num_tests; i++) {
 912         const struct poly1305_test *test = &poly1305_tests[i];
 913         unsigned char *in;
 914         size_t inlen = strlen(test->inputhex);
 915
 916         if (strlen(test->keyhex) != sizeof(key) * 2 ||
 917             strlen(test->outhex) != sizeof(out) * 2 || (inlen & 1) == 1)
 918             return 1;
 919
 920         inlen /= 2;
 921
 922         hex_decode(key, test->keyhex);
 923         hex_decode(expected, test->outhex);
 924
 925         in = malloc(inlen);
 926
 927         hex_decode(in, test->inputhex);
 928
 929         Poly1305_Init(&poly1305, key);
 930         Poly1305_Update(&poly1305, in, inlen);
 931         Poly1305_Final(&poly1305, out);
 932
 933         if (memcmp(out, expected, sizeof(expected)) != 0) {
 934             printf("Poly1305 test #%d failed.\n", i);
 935             printf("got:      ");
 936             hexdump(out, sizeof(out));
 937             printf("\nexpected: ");
 938             hexdump(expected, sizeof(expected));
 939             printf("\n");
 940             return 1;
 941         }
 942
 943         if (inlen > 16) {
 944             Poly1305_Init(&poly1305, key);
 945             Poly1305_Update(&poly1305, in, 1);
 946             Poly1305_Update(&poly1305, in+1, inlen-1);
 947             Poly1305_Final(&poly1305, out);
 948
 949             if (memcmp(out, expected, sizeof(expected)) != 0) {
 950                 printf("Poly1305 test #%d/1+(N-1) failed.\n", i);
 951                 printf("got:      ");
 952                 hexdump(out, sizeof(out));
 953                 printf("\nexpected: ");
 954                 hexdump(expected, sizeof(expected));
 955                 printf("\n");
 956                 return 1;
 957             }
 958         }
 959
 960         if (inlen > 32) {
 961             size_t half = inlen / 2;
 962
 963             Poly1305_Init(&poly1305, key);
 964             Poly1305_Update(&poly1305, in, half);
 965             Poly1305_Update(&poly1305, in+half, inlen-half);
 966             Poly1305_Final(&poly1305, out);
 967
 968             if (memcmp(out, expected, sizeof(expected)) != 0) {
 969                 printf("Poly1305 test #%d/2 failed.\n", i);
 970                 printf("got:      ");
 971                 hexdump(out, sizeof(out));
 972                 printf("\nexpected: ");
 973                 hexdump(expected, sizeof(expected));
 974                 printf("\n");
 975                 return 1;
 976             }
 977
 978             for (half = 16; half < inlen; half += 16) {
 979                 Poly1305_Init(&poly1305, key);
 980                 Poly1305_Update(&poly1305, in, half);
 981                 Poly1305_Update(&poly1305, in+half, inlen-half);
 982                 Poly1305_Final(&poly1305, out);
 983
 984                 if (memcmp(out, expected, sizeof(expected)) != 0) {
 985                     printf("Poly1305 test #%d/%d+%d failed.\n",
 986                                            i, half, inlen-half);
 987                     printf("got:      ");
 988                     hexdump(out, sizeof(out));
 989                     printf("\nexpected: ");
 990                     hexdump(expected, sizeof(expected));
 991                     printf("\n");
 992                     return 1;
 993                 }
 994             }
 995         }
 996
 997         free(in);
 998     }
 999
1000     printf("PASS\n");
1001
1002 # ifdef OPENSSL_CPUID_OBJ
1003     {
1004         unsigned char buf[8192];
1005         unsigned long long stopwatch;
1006         unsigned long long OPENSSL_rdtsc();
1007
1008         memset (buf,0x55,sizeof(buf));
1009         memset (key,0xAA,sizeof(key));
1010
1011         Poly1305_Init(&poly1305, key);
1012
1013         for (i=0;i<100000;i++)
1014             Poly1305_Update(&poly1305,buf,sizeof(buf));
1015
1016         stopwatch = OPENSSL_rdtsc();
1017         for (i=0;i<10000;i++)
1018             Poly1305_Update(&poly1305,buf,sizeof(buf));
1019         stopwatch = OPENSSL_rdtsc() - stopwatch;
1020
1021         printf("%g\n",stopwatch/(double)(i*sizeof(buf)));
1022
1023         stopwatch = OPENSSL_rdtsc();
1024         for (i=0;i<10000;i++) {
1025             Poly1305_Init(&poly1305, key);
1026             Poly1305_Update(&poly1305,buf,16);
1027             Poly1305_Final(&poly1305,buf);
1028         }
1029         stopwatch = OPENSSL_rdtsc() - stopwatch;
1030
1031         printf("%g\n",stopwatch/(double)(i));
1032     }
1033 # endif
1034     return 0;
1035 }
1036 #endif