git.zerfleddert.de Git - proxmark3-svn/blob - client/hardnested/hardnested_bitarray_core.c

   1 //-----------------------------------------------------------------------------
   2 // Copyright (C) 2016, 2017 by piwi
   3 //
   4 // This code is licensed to you under the terms of the GNU GPL, version 2 or,
   5 // at your option, any later version. See the LICENSE.txt file for the text of
   6 // the license.ch b
   7 //-----------------------------------------------------------------------------
   8 // Implements a card only attack based on crypto text (encrypted nonces
   9 // received during a nested authentication) only. Unlike other card only
  10 // attacks this doesn't rely on implementation errors but only on the
  11 // inherent weaknesses of the crypto1 cypher. Described in
  12 //   Carlo Meijer, Roel Verdult, "Ciphertext-only Cryptanalysis on Hardened
  13 //   Mifare Classic Cards" in Proceedings of the 22nd ACM SIGSAC Conference on
  14 //   Computer and Communications Security, 2015
  15 //-----------------------------------------------------------------------------
  16 // some helper functions which can benefit from SIMD instructions or other special instructions
  17 //
  18
  19 #include "hardnested_bitarray_core.h"
  20
  21 #include <stdint.h>
  22 #include <stdio.h>
  23 #include <stdlib.h>
  24 #include <malloc.h>
  25
  26 // this needs to be compiled several times for each instruction set.
  27 // For each instruction set, define a dedicated function name:
  28 #if defined (__AVX512F__)
  29 #define MALLOC_BITARRAY malloc_bitarray_AVX512
  30 #define FREE_BITARRAY free_bitarray_AVX512
  31 #define BITCOUNT bitcount_AVX512
  32 #define COUNT_STATES count_states_AVX512
  33 #define BITARRAY_AND bitarray_AND_AVX512
  34 #define BITARRAY_LOW20_AND bitarray_low20_AND_AVX512
  35 #define COUNT_BITARRAY_AND count_bitarray_AND_AVX512
  36 #define COUNT_BITARRAY_LOW20_AND count_bitarray_low20_AND_AVX512
  37 #define BITARRAY_AND4 bitarray_AND4_AVX512
  38 #define BITARRAY_OR bitarray_OR_AVX512
  39 #define COUNT_BITARRAY_AND2 count_bitarray_AND2_AVX512
  40 #define COUNT_BITARRAY_AND3 count_bitarray_AND3_AVX512
  41 #define COUNT_BITARRAY_AND4 count_bitarray_AND4_AVX512
  42 #elif defined (__AVX2__)
  43 #define MALLOC_BITARRAY malloc_bitarray_AVX2
  44 #define FREE_BITARRAY free_bitarray_AVX2
  45 #define BITCOUNT bitcount_AVX2
  46 #define COUNT_STATES count_states_AVX2
  47 #define BITARRAY_AND bitarray_AND_AVX2
  48 #define BITARRAY_LOW20_AND bitarray_low20_AND_AVX2
  49 #define COUNT_BITARRAY_AND count_bitarray_AND_AVX2
  50 #define COUNT_BITARRAY_LOW20_AND count_bitarray_low20_AND_AVX2
  51 #define BITARRAY_AND4 bitarray_AND4_AVX2
  52 #define BITARRAY_OR bitarray_OR_AVX2
  53 #define COUNT_BITARRAY_AND2 count_bitarray_AND2_AVX2
  54 #define COUNT_BITARRAY_AND3 count_bitarray_AND3_AVX2
  55 #define COUNT_BITARRAY_AND4 count_bitarray_AND4_AVX2
  56 #elif defined (__AVX__)
  57 #define MALLOC_BITARRAY malloc_bitarray_AVX
  58 #define FREE_BITARRAY free_bitarray_AVX
  59 #define BITCOUNT bitcount_AVX
  60 #define COUNT_STATES count_states_AVX
  61 #define BITARRAY_AND bitarray_AND_AVX
  62 #define BITARRAY_LOW20_AND bitarray_low20_AND_AVX
  63 #define COUNT_BITARRAY_AND count_bitarray_AND_AVX
  64 #define COUNT_BITARRAY_LOW20_AND count_bitarray_low20_AND_AVX
  65 #define BITARRAY_AND4 bitarray_AND4_AVX
  66 #define BITARRAY_OR bitarray_OR_AVX
  67 #define COUNT_BITARRAY_AND2 count_bitarray_AND2_AVX
  68 #define COUNT_BITARRAY_AND3 count_bitarray_AND3_AVX
  69 #define COUNT_BITARRAY_AND4 count_bitarray_AND4_AVX
  70 #elif defined (__SSE2__)
  71 #define MALLOC_BITARRAY malloc_bitarray_SSE2
  72 #define FREE_BITARRAY free_bitarray_SSE2
  73 #define BITCOUNT bitcount_SSE2
  74 #define COUNT_STATES count_states_SSE2
  75 #define BITARRAY_AND bitarray_AND_SSE2
  76 #define BITARRAY_LOW20_AND bitarray_low20_AND_SSE2
  77 #define COUNT_BITARRAY_AND count_bitarray_AND_SSE2
  78 #define COUNT_BITARRAY_LOW20_AND count_bitarray_low20_AND_SSE2
  79 #define BITARRAY_AND4 bitarray_AND4_SSE2
  80 #define BITARRAY_OR bitarray_OR_SSE2
  81 #define COUNT_BITARRAY_AND2 count_bitarray_AND2_SSE2
  82 #define COUNT_BITARRAY_AND3 count_bitarray_AND3_SSE2
  83 #define COUNT_BITARRAY_AND4 count_bitarray_AND4_SSE2
  84 #elif defined (__MMX__)
  85 #define MALLOC_BITARRAY malloc_bitarray_MMX
  86 #define FREE_BITARRAY free_bitarray_MMX
  87 #define BITCOUNT bitcount_MMX
  88 #define COUNT_STATES count_states_MMX
  89 #define BITARRAY_AND bitarray_AND_MMX
  90 #define BITARRAY_LOW20_AND bitarray_low20_AND_MMX
  91 #define COUNT_BITARRAY_AND count_bitarray_AND_MMX
  92 #define COUNT_BITARRAY_LOW20_AND count_bitarray_low20_AND_MMX
  93 #define BITARRAY_AND4 bitarray_AND4_MMX
  94 #define BITARRAY_OR bitarray_OR_MMX
  95 #define COUNT_BITARRAY_AND2 count_bitarray_AND2_MMX
  96 #define COUNT_BITARRAY_AND3 count_bitarray_AND3_MMX
  97 #define COUNT_BITARRAY_AND4 count_bitarray_AND4_MMX
  98 #else
  99 #define MALLOC_BITARRAY malloc_bitarray_NOSIMD
 100 #define FREE_BITARRAY free_bitarray_NOSIMD
 101 #define BITCOUNT bitcount_NOSIMD
 102 #define COUNT_STATES count_states_NOSIMD
 103 #define BITARRAY_AND bitarray_AND_NOSIMD
 104 #define BITARRAY_LOW20_AND bitarray_low20_AND_NOSIMD
 105 #define COUNT_BITARRAY_AND count_bitarray_AND_NOSIMD
 106 #define COUNT_BITARRAY_LOW20_AND count_bitarray_low20_AND_NOSIMD
 107 #define BITARRAY_AND4 bitarray_AND4_NOSIMD
 108 #define BITARRAY_OR bitarray_OR_NOSIMD
 109 #define COUNT_BITARRAY_AND2 count_bitarray_AND2_NOSIMD
 110 #define COUNT_BITARRAY_AND3 count_bitarray_AND3_NOSIMD
 111 #define COUNT_BITARRAY_AND4 count_bitarray_AND4_NOSIMD
 112 #endif
 113
 114
 115 // typedefs and declaration of functions:
 116 typedef uint32_t* malloc_bitarray_t(uint32_t);
 117 malloc_bitarray_t malloc_bitarray_AVX512, malloc_bitarray_AVX2, malloc_bitarray_AVX, malloc_bitarray_SSE2, malloc_bitarray_MMX, malloc_bitarray_NOSIMD, malloc_bitarray_dispatch;
 118 typedef void free_bitarray_t(uint32_t*);
 119 free_bitarray_t free_bitarray_AVX512, free_bitarray_AVX2, free_bitarray_AVX, free_bitarray_SSE2, free_bitarray_MMX, free_bitarray_NOSIMD, free_bitarray_dispatch;
 120 typedef uint32_t bitcount_t(uint32_t);
 121 bitcount_t bitcount_AVX512, bitcount_AVX2, bitcount_AVX, bitcount_SSE2, bitcount_MMX, bitcount_NOSIMD, bitcount_dispatch;
 122 typedef uint32_t count_states_t(uint32_t*);
 123 count_states_t count_states_AVX512, count_states_AVX2, count_states_AVX, count_states_SSE2, count_states_MMX, count_states_NOSIMD, count_states_dispatch;
 124 typedef void bitarray_AND_t(uint32_t[], uint32_t[]);
 125 bitarray_AND_t bitarray_AND_AVX512, bitarray_AND_AVX2, bitarray_AND_AVX, bitarray_AND_SSE2, bitarray_AND_MMX, bitarray_AND_NOSIMD, bitarray_AND_dispatch;
 126 typedef void bitarray_low20_AND_t(uint32_t*, uint32_t*);
 127 bitarray_low20_AND_t bitarray_low20_AND_AVX512, bitarray_low20_AND_AVX2, bitarray_low20_AND_AVX, bitarray_low20_AND_SSE2, bitarray_low20_AND_MMX, bitarray_low20_AND_NOSIMD, bitarray_low20_AND_dispatch;
 128 typedef uint32_t count_bitarray_AND_t(uint32_t*, uint32_t*);
 129 count_bitarray_AND_t count_bitarray_AND_AVX512, count_bitarray_AND_AVX2, count_bitarray_AND_AVX, count_bitarray_AND_SSE2, count_bitarray_AND_MMX, count_bitarray_AND_NOSIMD, count_bitarray_AND_dispatch;
 130 typedef uint32_t count_bitarray_low20_AND_t(uint32_t*, uint32_t*);
 131 count_bitarray_low20_AND_t count_bitarray_low20_AND_AVX512, count_bitarray_low20_AND_AVX2, count_bitarray_low20_AND_AVX, count_bitarray_low20_AND_SSE2, count_bitarray_low20_AND_MMX, count_bitarray_low20_AND_NOSIMD, count_bitarray_low20_AND_dispatch;
 132 typedef void bitarray_AND4_t(uint32_t*, uint32_t*, uint32_t*, uint32_t*);
 133 bitarray_AND4_t bitarray_AND4_AVX512, bitarray_AND4_AVX2, bitarray_AND4_AVX, bitarray_AND4_SSE2, bitarray_AND4_MMX, bitarray_AND4_NOSIMD, bitarray_AND4_dispatch;
 134 typedef void bitarray_OR_t(uint32_t[], uint32_t[]);
 135 bitarray_OR_t bitarray_OR_AVX512, bitarray_OR_AVX2, bitarray_OR_AVX, bitarray_OR_SSE2, bitarray_OR_MMX, bitarray_OR_NOSIMD, bitarray_OR_dispatch;
 136 typedef uint32_t count_bitarray_AND2_t(uint32_t*, uint32_t*);
 137 count_bitarray_AND2_t count_bitarray_AND2_AVX512, count_bitarray_AND2_AVX2, count_bitarray_AND2_AVX, count_bitarray_AND2_SSE2, count_bitarray_AND2_MMX, count_bitarray_AND2_NOSIMD, count_bitarray_AND2_dispatch;
 138 typedef uint32_t count_bitarray_AND3_t(uint32_t*, uint32_t*, uint32_t*);
 139 count_bitarray_AND3_t count_bitarray_AND3_AVX512, count_bitarray_AND3_AVX2, count_bitarray_AND3_AVX, count_bitarray_AND3_SSE2, count_bitarray_AND3_MMX, count_bitarray_AND3_NOSIMD, count_bitarray_AND3_dispatch;
 140 typedef uint32_t count_bitarray_AND4_t(uint32_t*, uint32_t*, uint32_t*, uint32_t*);
 141 count_bitarray_AND4_t count_bitarray_AND4_AVX512, count_bitarray_AND4_AVX2, count_bitarray_AND4_AVX, count_bitarray_AND4_SSE2, count_bitarray_AND4_MMX, count_bitarray_AND4_NOSIMD, count_bitarray_AND4_dispatch;
 142
 143
 144 inline uint32_t *MALLOC_BITARRAY(uint32_t x)
 145 {
 146 #ifdef _WIN32
 147         return __builtin_assume_aligned(_aligned_malloc((x), __BIGGEST_ALIGNMENT__), __BIGGEST_ALIGNMENT__);
 148 #else
 149         return __builtin_assume_aligned(memalign(__BIGGEST_ALIGNMENT__, (x)), __BIGGEST_ALIGNMENT__);
 150 #endif
 151 }
 152
 153
 154 inline void FREE_BITARRAY(uint32_t *x)
 155 {
 156 #ifdef _WIN32
 157         _aligned_free(x);
 158 #else
 159         free(x);
 160 #endif
 161 }
 162
 163
 164 inline uint32_t BITCOUNT(uint32_t a)
 165 {
 166         return __builtin_popcountl(a);
 167 }
 168
 169
 170 inline uint32_t COUNT_STATES(uint32_t *A)
 171 {
 172         uint32_t count = 0;
 173         for (uint32_t i = 0; i < (1<<19); i++) {
 174                 count += BITCOUNT(A[i]);
 175         }
 176         return count;
 177 }
 178
 179
 180 inline void BITARRAY_AND(uint32_t *restrict A, uint32_t *restrict B)
 181 {
 182         A = __builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 183         B = __builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 184         for (uint32_t i = 0; i < (1<<19); i++) {
 185                 A[i] &= B[i];
 186         }
 187 }
 188
 189
 190 inline void BITARRAY_LOW20_AND(uint32_t *restrict A, uint32_t *restrict B)
 191 {
 192         uint16_t *a = (uint16_t *)__builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 193         uint16_t *b = (uint16_t *)__builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 194
 195         for (uint32_t i = 0; i < (1<<20); i++) {
 196                 if (!b[i]) {
 197                         a[i] = 0;
 198                 }
 199         }
 200 }
 201
 202
 203 inline uint32_t COUNT_BITARRAY_AND(uint32_t *restrict A, uint32_t *restrict B)
 204 {
 205         A = __builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 206         B = __builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 207         uint32_t count = 0;
 208         for (uint32_t i = 0; i < (1<<19); i++) {
 209                 A[i] &= B[i];
 210                 count += BITCOUNT(A[i]);
 211         }
 212         return count;
 213 }
 214
 215
 216 inline uint32_t COUNT_BITARRAY_LOW20_AND(uint32_t *restrict A, uint32_t *restrict B)
 217 {
 218         uint16_t *a = (uint16_t *)__builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 219         uint16_t *b = (uint16_t *)__builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 220         uint32_t count = 0;
 221
 222         for (uint32_t i = 0; i < (1<<20); i++) {
 223                 if (!b[i]) {
 224                         a[i] = 0;
 225                 }
 226                 count += BITCOUNT(a[i]);
 227         }
 228         return count;
 229 }
 230
 231
 232 inline void BITARRAY_AND4(uint32_t *restrict A, uint32_t *restrict B, uint32_t *restrict C, uint32_t *restrict D)
 233 {
 234         A = __builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 235         B = __builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 236         C = __builtin_assume_aligned(C, __BIGGEST_ALIGNMENT__);
 237         D = __builtin_assume_aligned(D, __BIGGEST_ALIGNMENT__);
 238         for (uint32_t i = 0; i < (1<<19); i++) {
 239                 A[i] = B[i] & C[i] & D[i];
 240         }
 241 }
 242
 243
 244 inline void BITARRAY_OR(uint32_t *restrict A, uint32_t *restrict B)
 245 {
 246         A = __builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 247         B = __builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 248         for (uint32_t i = 0; i < (1<<19); i++) {
 249                 A[i] |= B[i];
 250         }
 251 }
 252
 253
 254 inline uint32_t COUNT_BITARRAY_AND2(uint32_t *restrict A, uint32_t *restrict B)
 255 {
 256         A = __builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 257         B = __builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 258         uint32_t count = 0;
 259         for (uint32_t i = 0; i < (1<<19); i++) {
 260                 count += BITCOUNT(A[i] & B[i]);
 261         }
 262         return count;
 263 }
 264
 265
 266 inline uint32_t COUNT_BITARRAY_AND3(uint32_t *restrict A, uint32_t *restrict B, uint32_t *restrict C)
 267 {
 268         A = __builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 269         B = __builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 270         C = __builtin_assume_aligned(C, __BIGGEST_ALIGNMENT__);
 271         uint32_t count = 0;
 272         for (uint32_t i = 0; i < (1<<19); i++) {
 273                 count += BITCOUNT(A[i] & B[i] & C[i]);
 274         }
 275         return count;
 276 }
 277
 278
 279 inline uint32_t COUNT_BITARRAY_AND4(uint32_t *restrict A, uint32_t *restrict B, uint32_t *restrict C, uint32_t *restrict D)
 280 {
 281         A = __builtin_assume_aligned(A, __BIGGEST_ALIGNMENT__);
 282         B = __builtin_assume_aligned(B, __BIGGEST_ALIGNMENT__);
 283         C = __builtin_assume_aligned(C, __BIGGEST_ALIGNMENT__);
 284         D = __builtin_assume_aligned(D, __BIGGEST_ALIGNMENT__);
 285         uint32_t count = 0;
 286         for (uint32_t i = 0; i < (1<<19); i++) {
 287                 count += BITCOUNT(A[i] & B[i] & C[i] & D[i]);
 288         }
 289         return count;
 290 }
 291
 292
 293 #ifndef __MMX__
 294
 295 // pointers to functions:
 296 malloc_bitarray_t *malloc_bitarray_function_p = &malloc_bitarray_dispatch;
 297 free_bitarray_t *free_bitarray_function_p = &free_bitarray_dispatch;
 298 bitcount_t *bitcount_function_p = &bitcount_dispatch;
 299 count_states_t *count_states_function_p = &count_states_dispatch;
 300 bitarray_AND_t *bitarray_AND_function_p = &bitarray_AND_dispatch;
 301 bitarray_low20_AND_t *bitarray_low20_AND_function_p = &bitarray_low20_AND_dispatch;
 302 count_bitarray_AND_t *count_bitarray_AND_function_p = &count_bitarray_AND_dispatch;
 303 count_bitarray_low20_AND_t *count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_dispatch;
 304 bitarray_AND4_t *bitarray_AND4_function_p = &bitarray_AND4_dispatch;
 305 bitarray_OR_t *bitarray_OR_function_p = &bitarray_OR_dispatch;
 306 count_bitarray_AND2_t *count_bitarray_AND2_function_p = &count_bitarray_AND2_dispatch;
 307 count_bitarray_AND3_t *count_bitarray_AND3_function_p = &count_bitarray_AND3_dispatch;
 308 count_bitarray_AND4_t *count_bitarray_AND4_function_p = &count_bitarray_AND4_dispatch;
 309
 310 // determine the available instruction set at runtime and call the correct function
 311 uint32_t *malloc_bitarray_dispatch(uint32_t x) {
 312 #if defined (__i386__) || defined (__x86_64__)
 313         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 314         if (__builtin_cpu_supports("avx512f")) malloc_bitarray_function_p = &malloc_bitarray_AVX512;
 315         else if (__builtin_cpu_supports("avx2")) malloc_bitarray_function_p = &malloc_bitarray_AVX2;
 316         #else
 317         if (__builtin_cpu_supports("avx2")) malloc_bitarray_function_p = &malloc_bitarray_AVX2;
 318         #endif
 319         else if (__builtin_cpu_supports("avx")) malloc_bitarray_function_p = &malloc_bitarray_AVX;
 320         else if (__builtin_cpu_supports("sse2")) malloc_bitarray_function_p = &malloc_bitarray_SSE2;
 321         else if (__builtin_cpu_supports("mmx")) malloc_bitarray_function_p = &malloc_bitarray_MMX;
 322     else
 323 #endif
 324                 malloc_bitarray_function_p = &malloc_bitarray_NOSIMD;
 325
 326     // call the most optimized function for this CPU
 327     return (*malloc_bitarray_function_p)(x);
 328 }
 329
 330 void free_bitarray_dispatch(uint32_t *x) {
 331 #if defined (__i386__) || defined (__x86_64__)
 332         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 333         if (__builtin_cpu_supports("avx512f")) free_bitarray_function_p = &free_bitarray_AVX512;
 334         else if (__builtin_cpu_supports("avx2")) free_bitarray_function_p = &free_bitarray_AVX2;
 335         #else
 336         if (__builtin_cpu_supports("avx2")) free_bitarray_function_p = &free_bitarray_AVX2;
 337         #endif
 338         else if (__builtin_cpu_supports("avx")) free_bitarray_function_p = &free_bitarray_AVX;
 339         else if (__builtin_cpu_supports("sse2")) free_bitarray_function_p = &free_bitarray_SSE2;
 340         else if (__builtin_cpu_supports("mmx")) free_bitarray_function_p = &free_bitarray_MMX;
 341     else
 342 #endif
 343                 free_bitarray_function_p = &free_bitarray_NOSIMD;
 344
 345     // call the most optimized function for this CPU
 346     (*free_bitarray_function_p)(x);
 347 }
 348
 349 uint32_t bitcount_dispatch(uint32_t a) {
 350 #if defined (__i386__) || defined (__x86_64__)
 351         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 352         if (__builtin_cpu_supports("avx512f")) bitcount_function_p = &bitcount_AVX512;
 353         else if (__builtin_cpu_supports("avx2")) bitcount_function_p = &bitcount_AVX2;
 354         #else
 355         if (__builtin_cpu_supports("avx2")) bitcount_function_p = &bitcount_AVX2;
 356         #endif
 357         else if (__builtin_cpu_supports("avx")) bitcount_function_p = &bitcount_AVX;
 358         else if (__builtin_cpu_supports("sse2")) bitcount_function_p = &bitcount_SSE2;
 359         else if (__builtin_cpu_supports("mmx")) bitcount_function_p = &bitcount_MMX;
 360     else
 361 #endif
 362                 bitcount_function_p = &bitcount_NOSIMD;
 363
 364     // call the most optimized function for this CPU
 365     return (*bitcount_function_p)(a);
 366 }
 367
 368 uint32_t count_states_dispatch(uint32_t *bitarray) {
 369 #if defined (__i386__) || defined (__x86_64__)
 370         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 371         if (__builtin_cpu_supports("avx512f")) count_states_function_p = &count_states_AVX512;
 372         else if (__builtin_cpu_supports("avx2")) count_states_function_p = &count_states_AVX2;
 373         #else
 374         if (__builtin_cpu_supports("avx2")) count_states_function_p = &count_states_AVX2;
 375         #endif
 376         else if (__builtin_cpu_supports("avx")) count_states_function_p = &count_states_AVX;
 377         else if (__builtin_cpu_supports("sse2")) count_states_function_p = &count_states_SSE2;
 378         else if (__builtin_cpu_supports("mmx")) count_states_function_p = &count_states_MMX;
 379     else
 380 #endif
 381                 count_states_function_p = &count_states_NOSIMD;
 382
 383     // call the most optimized function for this CPU
 384     return (*count_states_function_p)(bitarray);
 385 }
 386
 387 void bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
 388 #if defined (__i386__) || defined (__x86_64__)
 389         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 390         if (__builtin_cpu_supports("avx512f")) bitarray_AND_function_p = &bitarray_AND_AVX512;
 391         else if (__builtin_cpu_supports("avx2")) bitarray_AND_function_p = &bitarray_AND_AVX2;
 392         #else
 393         if (__builtin_cpu_supports("avx2")) bitarray_AND_function_p = &bitarray_AND_AVX2;
 394         #endif
 395         else if (__builtin_cpu_supports("avx")) bitarray_AND_function_p = &bitarray_AND_AVX;
 396         else if (__builtin_cpu_supports("sse2")) bitarray_AND_function_p = &bitarray_AND_SSE2;
 397         else if (__builtin_cpu_supports("mmx")) bitarray_AND_function_p = &bitarray_AND_MMX;
 398     else
 399 #endif
 400                 bitarray_AND_function_p = &bitarray_AND_NOSIMD;
 401
 402     // call the most optimized function for this CPU
 403     (*bitarray_AND_function_p)(A,B);
 404 }
 405
 406 void bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
 407 #if defined (__i386__) || defined (__x86_64__)
 408         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 409         if (__builtin_cpu_supports("avx512f")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX512;
 410         else if (__builtin_cpu_supports("avx2")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX2;
 411         #else
 412         if (__builtin_cpu_supports("avx2")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX2;
 413         #endif
 414         else if (__builtin_cpu_supports("avx")) bitarray_low20_AND_function_p = &bitarray_low20_AND_AVX;
 415         else if (__builtin_cpu_supports("sse2")) bitarray_low20_AND_function_p = &bitarray_low20_AND_SSE2;
 416         else if (__builtin_cpu_supports("mmx")) bitarray_low20_AND_function_p = &bitarray_low20_AND_MMX;
 417     else
 418 #endif
 419                 bitarray_low20_AND_function_p = &bitarray_low20_AND_NOSIMD;
 420
 421     // call the most optimized function for this CPU
 422     (*bitarray_low20_AND_function_p)(A, B);
 423 }
 424
 425 uint32_t count_bitarray_AND_dispatch(uint32_t *A, uint32_t *B) {
 426 #if defined (__i386__) || defined (__x86_64__)
 427         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 428         if (__builtin_cpu_supports("avx512f")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX512;
 429         else if (__builtin_cpu_supports("avx2")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX2;
 430         #else
 431         if (__builtin_cpu_supports("avx2")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX2;
 432         #endif
 433         else if (__builtin_cpu_supports("avx")) count_bitarray_AND_function_p = &count_bitarray_AND_AVX;
 434         else if (__builtin_cpu_supports("sse2")) count_bitarray_AND_function_p = &count_bitarray_AND_SSE2;
 435         else if (__builtin_cpu_supports("mmx")) count_bitarray_AND_function_p = &count_bitarray_AND_MMX;
 436     else
 437 #endif
 438                 count_bitarray_AND_function_p = &count_bitarray_AND_NOSIMD;
 439
 440     // call the most optimized function for this CPU
 441     return (*count_bitarray_AND_function_p)(A, B);
 442 }
 443
 444 uint32_t count_bitarray_low20_AND_dispatch(uint32_t *A, uint32_t *B) {
 445 #if defined (__i386__) || defined (__x86_64__)
 446         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 447         if (__builtin_cpu_supports("avx512f")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX512;
 448         else if (__builtin_cpu_supports("avx2")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX2;
 449         #else
 450         if (__builtin_cpu_supports("avx2")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX2;
 451         #endif
 452         else if (__builtin_cpu_supports("avx")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_AVX;
 453         else if (__builtin_cpu_supports("sse2")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_SSE2;
 454         else if (__builtin_cpu_supports("mmx")) count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_MMX;
 455     else
 456 #endif
 457                 count_bitarray_low20_AND_function_p = &count_bitarray_low20_AND_NOSIMD;
 458
 459     // call the most optimized function for this CPU
 460     return (*count_bitarray_low20_AND_function_p)(A, B);
 461 }
 462
 463 void bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
 464 #if defined (__i386__) || defined (__x86_64__)
 465         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 466         if (__builtin_cpu_supports("avx512f")) bitarray_AND4_function_p = &bitarray_AND4_AVX512;
 467         else if (__builtin_cpu_supports("avx2")) bitarray_AND4_function_p = &bitarray_AND4_AVX2;
 468         #else
 469         if (__builtin_cpu_supports("avx2")) bitarray_AND4_function_p = &bitarray_AND4_AVX2;
 470         #endif
 471         else if (__builtin_cpu_supports("avx")) bitarray_AND4_function_p = &bitarray_AND4_AVX;
 472         else if (__builtin_cpu_supports("sse2")) bitarray_AND4_function_p = &bitarray_AND4_SSE2;
 473         else if (__builtin_cpu_supports("mmx")) bitarray_AND4_function_p = &bitarray_AND4_MMX;
 474     else
 475 #endif
 476                 bitarray_AND4_function_p = &bitarray_AND4_NOSIMD;
 477
 478     // call the most optimized function for this CPU
 479     (*bitarray_AND4_function_p)(A, B, C, D);
 480 }
 481
 482 void bitarray_OR_dispatch(uint32_t *A, uint32_t *B) {
 483 #if defined (__i386__) || defined (__x86_64__)
 484         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 485         if (__builtin_cpu_supports("avx512f")) bitarray_OR_function_p = &bitarray_OR_AVX512;
 486         else if (__builtin_cpu_supports("avx2")) bitarray_OR_function_p = &bitarray_OR_AVX2;
 487         #else
 488         if (__builtin_cpu_supports("avx2")) bitarray_OR_function_p = &bitarray_OR_AVX2;
 489         #endif
 490         else if (__builtin_cpu_supports("avx")) bitarray_OR_function_p = &bitarray_OR_AVX;
 491         else if (__builtin_cpu_supports("sse2")) bitarray_OR_function_p = &bitarray_OR_SSE2;
 492         else if (__builtin_cpu_supports("mmx")) bitarray_OR_function_p = &bitarray_OR_MMX;
 493     else
 494 #endif
 495                 bitarray_OR_function_p = &bitarray_OR_NOSIMD;
 496
 497     // call the most optimized function for this CPU
 498     (*bitarray_OR_function_p)(A,B);
 499 }
 500
 501 uint32_t count_bitarray_AND2_dispatch(uint32_t *A, uint32_t *B) {
 502 #if defined (__i386__) || defined (__x86_64__)
 503         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 504         if (__builtin_cpu_supports("avx512f")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX512;
 505         else if (__builtin_cpu_supports("avx2")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX2;
 506         #else
 507         if (__builtin_cpu_supports("avx2")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX2;
 508         #endif
 509         else if (__builtin_cpu_supports("avx")) count_bitarray_AND2_function_p = &count_bitarray_AND2_AVX;
 510         else if (__builtin_cpu_supports("sse2")) count_bitarray_AND2_function_p = &count_bitarray_AND2_SSE2;
 511         else if (__builtin_cpu_supports("mmx")) count_bitarray_AND2_function_p = &count_bitarray_AND2_MMX;
 512     else
 513 #endif
 514                 count_bitarray_AND2_function_p = &count_bitarray_AND2_NOSIMD;
 515
 516     // call the most optimized function for this CPU
 517     return (*count_bitarray_AND2_function_p)(A, B);
 518 }
 519
 520 uint32_t count_bitarray_AND3_dispatch(uint32_t *A, uint32_t *B, uint32_t *C) {
 521 #if defined (__i386__) || defined (__x86_64__)
 522         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 523         if (__builtin_cpu_supports("avx512f")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX512;
 524         else if (__builtin_cpu_supports("avx2")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX2;
 525         #else
 526         if (__builtin_cpu_supports("avx2")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX2;
 527         #endif
 528         else if (__builtin_cpu_supports("avx")) count_bitarray_AND3_function_p = &count_bitarray_AND3_AVX;
 529         else if (__builtin_cpu_supports("sse2")) count_bitarray_AND3_function_p = &count_bitarray_AND3_SSE2;
 530         else if (__builtin_cpu_supports("mmx")) count_bitarray_AND3_function_p = &count_bitarray_AND3_MMX;
 531     else
 532 #endif
 533                 count_bitarray_AND3_function_p = &count_bitarray_AND3_NOSIMD;
 534
 535     // call the most optimized function for this CPU
 536     return (*count_bitarray_AND3_function_p)(A, B, C);
 537 }
 538
 539 uint32_t count_bitarray_AND4_dispatch(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
 540 #if defined (__i386__) || defined (__x86_64__)
 541         #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
 542         if (__builtin_cpu_supports("avx512f")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX512;
 543         else if (__builtin_cpu_supports("avx2")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX2;
 544         #else
 545         if (__builtin_cpu_supports("avx2")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX2;
 546         #endif
 547         else if (__builtin_cpu_supports("avx")) count_bitarray_AND4_function_p = &count_bitarray_AND4_AVX;
 548         else if (__builtin_cpu_supports("sse2")) count_bitarray_AND4_function_p = &count_bitarray_AND4_SSE2;
 549         else if (__builtin_cpu_supports("mmx")) count_bitarray_AND4_function_p = &count_bitarray_AND4_MMX;
 550     else
 551 #endif
 552                 count_bitarray_AND4_function_p = &count_bitarray_AND4_NOSIMD;
 553
 554     // call the most optimized function for this CPU
 555     return (*count_bitarray_AND4_function_p)(A, B, C, D);
 556 }
 557
 558
 559 ///////////////////////////////////////////////77
 560 // Entries to dispatched function calls
 561
 562 uint32_t *malloc_bitarray(uint32_t x) {
 563     return (*malloc_bitarray_function_p)(x);
 564 }
 565
 566 void free_bitarray(uint32_t *x) {
 567     (*free_bitarray_function_p)(x);
 568 }
 569
 570 uint32_t bitcount(uint32_t a) {
 571     return (*bitcount_function_p)(a);
 572 }
 573
 574 uint32_t count_states(uint32_t *bitarray) {
 575     return (*count_states_function_p)(bitarray);
 576 }
 577
 578 void bitarray_AND(uint32_t *A, uint32_t *B) {
 579     (*bitarray_AND_function_p)(A, B);
 580 }
 581
 582 void bitarray_low20_AND(uint32_t *A, uint32_t *B) {
 583     (*bitarray_low20_AND_function_p)(A, B);
 584 }
 585
 586 uint32_t count_bitarray_AND(uint32_t *A, uint32_t *B) {
 587     return (*count_bitarray_AND_function_p)(A, B);
 588 }
 589
 590 uint32_t count_bitarray_low20_AND(uint32_t *A, uint32_t *B) {
 591     return (*count_bitarray_low20_AND_function_p)(A, B);
 592 }
 593
 594 void bitarray_AND4(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
 595     (*bitarray_AND4_function_p)(A, B, C, D);
 596 }
 597
 598 void bitarray_OR(uint32_t *A, uint32_t *B) {
 599     (*bitarray_OR_function_p)(A, B);
 600 }
 601
 602 uint32_t count_bitarray_AND2(uint32_t *A, uint32_t *B) {
 603     return (*count_bitarray_AND2_function_p)(A, B);
 604 }
 605
 606 uint32_t count_bitarray_AND3(uint32_t *A, uint32_t *B, uint32_t *C) {
 607     return (*count_bitarray_AND3_function_p)(A, B, C);
 608 }
 609
 610 uint32_t count_bitarray_AND4(uint32_t *A, uint32_t *B, uint32_t *C, uint32_t *D) {
 611     return (*count_bitarray_AND4_function_p)(A, B, C, D);
 612 }
 613
 614 #endif
 615