master
h 6,773 lines 247 KB
Raw
1 /*
2 * xxHash - Extremely Fast Hash algorithm
3 * Header File
4 * Copyright (C) 2012-2023 Yann Collet
5 *
6 * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions are
10 * met:
11 *
12 * * Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * * Redistributions in binary form must reproduce the above
15 * copyright notice, this list of conditions and the following disclaimer
16 * in the documentation and/or other materials provided with the
17 * distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
22 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
23 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
24 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
25 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
29 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30 *
31 * You can contact the author at:
32 * - xxHash homepage: https://www.xxhash.com
33 * - xxHash source repository: https://github.com/Cyan4973/xxHash
34 */
35
36 /*!
37 * @mainpage xxHash
38 *
39 * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
40 * limits.
41 *
42 * It is proposed in four flavors, in three families:
43 * 1. @ref XXH32_family
44 * - Classic 32-bit hash function. Simple, compact, and runs on almost all
45 * 32-bit and 64-bit systems.
46 * 2. @ref XXH64_family
47 * - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
48 * 64-bit systems (but _not_ 32-bit systems).
49 * 3. @ref XXH3_family
50 * - Modern 64-bit and 128-bit hash function family which features improved
51 * strength and performance across the board, especially on smaller data.
52 * It benefits greatly from SIMD and 64-bit without requiring it.
53 *
54 * Benchmarks
55 * ---
56 * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
57 * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
58 *
59 * | Hash Name | ISA ext | Width | Large Data Speed | Small Data Velocity |
60 * | -------------------- | ------- | ----: | ---------------: | ------------------: |
61 * | XXH3_64bits() | @b AVX2 | 64 | 59.4 GB/s | 133.1 |
62 * | MeowHash | AES-NI | 128 | 58.2 GB/s | 52.5 |
63 * | XXH3_128bits() | @b AVX2 | 128 | 57.9 GB/s | 118.1 |
64 * | CLHash | PCLMUL | 64 | 37.1 GB/s | 58.1 |
65 * | XXH3_64bits() | @b SSE2 | 64 | 31.5 GB/s | 133.1 |
66 * | XXH3_128bits() | @b SSE2 | 128 | 29.6 GB/s | 118.1 |
67 * | RAM sequential read | | N/A | 28.0 GB/s | N/A |
68 * | ahash | AES-NI | 64 | 22.5 GB/s | 107.2 |
69 * | City64 | | 64 | 22.0 GB/s | 76.6 |
70 * | T1ha2 | | 64 | 22.0 GB/s | 99.0 |
71 * | City128 | | 128 | 21.7 GB/s | 57.7 |
72 * | FarmHash | AES-NI | 64 | 21.3 GB/s | 71.9 |
73 * | XXH64() | | 64 | 19.4 GB/s | 71.0 |
74 * | SpookyHash | | 64 | 19.3 GB/s | 53.2 |
75 * | Mum | | 64 | 18.0 GB/s | 67.0 |
76 * | CRC32C | SSE4.2 | 32 | 13.0 GB/s | 57.9 |
77 * | XXH32() | | 32 | 9.7 GB/s | 71.9 |
78 * | City32 | | 32 | 9.1 GB/s | 66.0 |
79 * | Blake3* | @b AVX2 | 256 | 4.4 GB/s | 8.1 |
80 * | Murmur3 | | 32 | 3.9 GB/s | 56.1 |
81 * | SipHash* | | 64 | 3.0 GB/s | 43.2 |
82 * | Blake3* | @b SSE2 | 256 | 2.4 GB/s | 8.1 |
83 * | HighwayHash | | 64 | 1.4 GB/s | 6.0 |
84 * | FNV64 | | 64 | 1.2 GB/s | 62.7 |
85 * | Blake2* | | 256 | 1.1 GB/s | 5.1 |
86 * | SHA1* | | 160 | 0.8 GB/s | 5.6 |
87 * | MD5* | | 128 | 0.6 GB/s | 7.8 |
88 * @note
89 * - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
90 * even though it is mandatory on x64.
91 * - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
92 * by modern standards.
93 * - Small data velocity is a rough average of algorithm's efficiency for small
94 * data. For more accurate information, see the wiki.
95 * - More benchmarks and strength tests are found on the wiki:
96 * https://github.com/Cyan4973/xxHash/wiki
97 *
98 * Usage
99 * ------
100 * All xxHash variants use a similar API. Changing the algorithm is a trivial
101 * substitution.
102 *
103 * @pre
104 * For functions which take an input and length parameter, the following
105 * requirements are assumed:
106 * - The range from [`input`, `input + length`) is valid, readable memory.
107 * - The only exception is if the `length` is `0`, `input` may be `NULL`.
108 * - For C++, the objects must have the *TriviallyCopyable* property, as the
109 * functions access bytes directly as if it was an array of `unsigned char`.
110 *
111 * @anchor single_shot_example
112 * **Single Shot**
113 *
114 * These functions are stateless functions which hash a contiguous block of memory,
115 * immediately returning the result. They are the easiest and usually the fastest
116 * option.
117 *
118 * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
119 *
120 * @code{.c}
121 * #include <string.h>
122 * #include "xxhash.h"
123 *
124 * // Example for a function which hashes a null terminated string with XXH32().
125 * XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
126 * {
127 * // NULL pointers are only valid if the length is zero
128 * size_t length = (string == NULL) ? 0 : strlen(string);
129 * return XXH32(string, length, seed);
130 * }
131 * @endcode
132 *
133 * @anchor streaming_example
134 * **Streaming**
135 *
136 * These groups of functions allow incremental hashing of unknown size, even
137 * more than what would fit in a size_t.
138 *
139 * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
140 *
141 * @code{.c}
142 * #include <stdio.h>
143 * #include <assert.h>
144 * #include "xxhash.h"
145 * // Example for a function which hashes a FILE incrementally with XXH3_64bits().
146 * XXH64_hash_t hashFile(FILE* f)
147 * {
148 * // Allocate a state struct. Do not just use malloc() or new.
149 * XXH3_state_t* state = XXH3_createState();
150 * assert(state != NULL && "Out of memory!");
151 * // Reset the state to start a new hashing session.
152 * XXH3_64bits_reset(state);
153 * char buffer[4096];
154 * size_t count;
155 * // Read the file in chunks
156 * while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
157 * // Run update() as many times as necessary to process the data
158 * XXH3_64bits_update(state, buffer, count);
159 * }
160 * // Retrieve the finalized hash. This will not change the state.
161 * XXH64_hash_t result = XXH3_64bits_digest(state);
162 * // Free the state. Do not use free().
163 * XXH3_freeState(state);
164 * return result;
165 * }
166 * @endcode
167 *
168 * @file xxhash.h
169 * xxHash prototypes and implementation
170 */
171
172 #if defined (__cplusplus)
173 extern "C" {
174 #endif
175
176 /* ****************************
177 * INLINE mode
178 ******************************/
179 /*!
180 * @defgroup public Public API
181 * Contains details on the public xxHash functions.
182 * @{
183 */
184 #ifdef XXH_DOXYGEN
185 /*!
186 * @brief Gives access to internal state declaration, required for static allocation.
187 *
188 * Incompatible with dynamic linking, due to risks of ABI changes.
189 *
190 * Usage:
191 * @code{.c}
192 * #define XXH_STATIC_LINKING_ONLY
193 * #include "xxhash.h"
194 * @endcode
195 */
196 # define XXH_STATIC_LINKING_ONLY
197 /* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */
198
199 /*!
200 * @brief Gives access to internal definitions.
201 *
202 * Usage:
203 * @code{.c}
204 * #define XXH_STATIC_LINKING_ONLY
205 * #define XXH_IMPLEMENTATION
206 * #include "xxhash.h"
207 * @endcode
208 */
209 # define XXH_IMPLEMENTATION
210 /* Do not undef XXH_IMPLEMENTATION for Doxygen */
211
212 /*!
213 * @brief Exposes the implementation and marks all functions as `inline`.
214 *
215 * Use these build macros to inline xxhash into the target unit.
216 * Inlining improves performance on small inputs, especially when the length is
217 * expressed as a compile-time constant:
218 *
219 * https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
220 *
221 * It also keeps xxHash symbols private to the unit, so they are not exported.
222 *
223 * Usage:
224 * @code{.c}
225 * #define XXH_INLINE_ALL
226 * #include "xxhash.h"
227 * @endcode
228 * Do not compile and link xxhash.o as a separate object, as it is not useful.
229 */
230 # define XXH_INLINE_ALL
231 # undef XXH_INLINE_ALL
232 /*!
233 * @brief Exposes the implementation without marking functions as inline.
234 */
235 # define XXH_PRIVATE_API
236 # undef XXH_PRIVATE_API
237 /*!
238 * @brief Emulate a namespace by transparently prefixing all symbols.
239 *
240 * If you want to include _and expose_ xxHash functions from within your own
241 * library, but also want to avoid symbol collisions with other libraries which
242 * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
243 * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
244 * (therefore, avoid empty or numeric values).
245 *
246 * Note that no change is required within the calling program as long as it
247 * includes `xxhash.h`: Regular symbol names will be automatically translated
248 * by this header.
249 */
250 # define XXH_NAMESPACE /* YOUR NAME HERE */
251 # undef XXH_NAMESPACE
252 #endif
253
254 #if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
255 && !defined(XXH_INLINE_ALL_31684351384)
256 /* this section should be traversed only once */
257 # define XXH_INLINE_ALL_31684351384
258 /* give access to the advanced API, required to compile implementations */
259 # undef XXH_STATIC_LINKING_ONLY /* avoid macro redef */
260 # define XXH_STATIC_LINKING_ONLY
261 /* make all functions private */
262 # undef XXH_PUBLIC_API
263 # if defined(__GNUC__)
264 # define XXH_PUBLIC_API static __inline __attribute__((unused))
265 # elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
266 # define XXH_PUBLIC_API static inline
267 # elif defined(_MSC_VER)
268 # define XXH_PUBLIC_API static __inline
269 # else
270 /* note: this version may generate warnings for unused static functions */
271 # define XXH_PUBLIC_API static
272 # endif
273
274 /*
275 * This part deals with the special case where a unit wants to inline xxHash,
276 * but "xxhash.h" has previously been included without XXH_INLINE_ALL,
277 * such as part of some previously included *.h header file.
278 * Without further action, the new include would just be ignored,
279 * and functions would effectively _not_ be inlined (silent failure).
280 * The following macros solve this situation by prefixing all inlined names,
281 * avoiding naming collision with previous inclusions.
282 */
283 /* Before that, we unconditionally #undef all symbols,
284 * in case they were already defined with XXH_NAMESPACE.
285 * They will then be redefined for XXH_INLINE_ALL
286 */
287 # undef XXH_versionNumber
288 /* XXH32 */
289 # undef XXH32
290 # undef XXH32_createState
291 # undef XXH32_freeState
292 # undef XXH32_reset
293 # undef XXH32_update
294 # undef XXH32_digest
295 # undef XXH32_copyState
296 # undef XXH32_canonicalFromHash
297 # undef XXH32_hashFromCanonical
298 /* XXH64 */
299 # undef XXH64
300 # undef XXH64_createState
301 # undef XXH64_freeState
302 # undef XXH64_reset
303 # undef XXH64_update
304 # undef XXH64_digest
305 # undef XXH64_copyState
306 # undef XXH64_canonicalFromHash
307 # undef XXH64_hashFromCanonical
308 /* XXH3_64bits */
309 # undef XXH3_64bits
310 # undef XXH3_64bits_withSecret
311 # undef XXH3_64bits_withSeed
312 # undef XXH3_64bits_withSecretandSeed
313 # undef XXH3_createState
314 # undef XXH3_freeState
315 # undef XXH3_copyState
316 # undef XXH3_64bits_reset
317 # undef XXH3_64bits_reset_withSeed
318 # undef XXH3_64bits_reset_withSecret
319 # undef XXH3_64bits_update
320 # undef XXH3_64bits_digest
321 # undef XXH3_generateSecret
322 /* XXH3_128bits */
323 # undef XXH128
324 # undef XXH3_128bits
325 # undef XXH3_128bits_withSeed
326 # undef XXH3_128bits_withSecret
327 # undef XXH3_128bits_reset
328 # undef XXH3_128bits_reset_withSeed
329 # undef XXH3_128bits_reset_withSecret
330 # undef XXH3_128bits_reset_withSecretandSeed
331 # undef XXH3_128bits_update
332 # undef XXH3_128bits_digest
333 # undef XXH128_isEqual
334 # undef XXH128_cmp
335 # undef XXH128_canonicalFromHash
336 # undef XXH128_hashFromCanonical
337 /* Finally, free the namespace itself */
338 # undef XXH_NAMESPACE
339
340 /* employ the namespace for XXH_INLINE_ALL */
341 # define XXH_NAMESPACE XXH_INLINE_
342 /*
343 * Some identifiers (enums, type names) are not symbols,
344 * but they must nonetheless be renamed to avoid redeclaration.
345 * Alternative solution: do not redeclare them.
346 * However, this requires some #ifdefs, and has a more dispersed impact.
347 * Meanwhile, renaming can be achieved in a single place.
348 */
349 # define XXH_IPREF(Id) XXH_NAMESPACE ## Id
350 # define XXH_OK XXH_IPREF(XXH_OK)
351 # define XXH_ERROR XXH_IPREF(XXH_ERROR)
352 # define XXH_errorcode XXH_IPREF(XXH_errorcode)
353 # define XXH32_canonical_t XXH_IPREF(XXH32_canonical_t)
354 # define XXH64_canonical_t XXH_IPREF(XXH64_canonical_t)
355 # define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t)
356 # define XXH32_state_s XXH_IPREF(XXH32_state_s)
357 # define XXH32_state_t XXH_IPREF(XXH32_state_t)
358 # define XXH64_state_s XXH_IPREF(XXH64_state_s)
359 # define XXH64_state_t XXH_IPREF(XXH64_state_t)
360 # define XXH3_state_s XXH_IPREF(XXH3_state_s)
361 # define XXH3_state_t XXH_IPREF(XXH3_state_t)
362 # define XXH128_hash_t XXH_IPREF(XXH128_hash_t)
363 /* Ensure the header is parsed again, even if it was previously included */
364 # undef XXHASH_H_5627135585666179
365 # undef XXHASH_H_STATIC_13879238742
366 #endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */
367
368 /* ****************************************************************
369 * Stable API
370 *****************************************************************/
371 #ifndef XXHASH_H_5627135585666179
372 #define XXHASH_H_5627135585666179 1
373
374 /*! @brief Marks a global symbol. */
375 #if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
376 # if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
377 # ifdef XXH_EXPORT
378 # define XXH_PUBLIC_API __declspec(dllexport)
379 # elif XXH_IMPORT
380 # define XXH_PUBLIC_API __declspec(dllimport)
381 # endif
382 # else
383 # define XXH_PUBLIC_API /* do nothing */
384 # endif
385 #endif
386
387 #ifdef XXH_NAMESPACE
388 # define XXH_CAT(A,B) A##B
389 # define XXH_NAME2(A,B) XXH_CAT(A,B)
390 # define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
391 /* XXH32 */
392 # define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
393 # define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
394 # define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
395 # define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
396 # define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
397 # define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
398 # define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
399 # define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
400 # define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
401 /* XXH64 */
402 # define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
403 # define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
404 # define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
405 # define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
406 # define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
407 # define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
408 # define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
409 # define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
410 # define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
411 /* XXH3_64bits */
412 # define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
413 # define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
414 # define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
415 # define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed)
416 # define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
417 # define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
418 # define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
419 # define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
420 # define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
421 # define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
422 # define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed)
423 # define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
424 # define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
425 # define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
426 # define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed)
427 /* XXH3_128bits */
428 # define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
429 # define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
430 # define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
431 # define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
432 # define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed)
433 # define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
434 # define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
435 # define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
436 # define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed)
437 # define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
438 # define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
439 # define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
440 # define XXH128_cmp XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
441 # define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
442 # define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
443 #endif
444
445
446 /* *************************************
447 * Compiler specifics
448 ***************************************/
449
450 /* specific declaration modes for Windows */
451 #if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
452 # if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
453 # ifdef XXH_EXPORT
454 # define XXH_PUBLIC_API __declspec(dllexport)
455 # elif XXH_IMPORT
456 # define XXH_PUBLIC_API __declspec(dllimport)
457 # endif
458 # else
459 # define XXH_PUBLIC_API /* do nothing */
460 # endif
461 #endif
462
463 #if defined (__GNUC__)
464 # define XXH_CONSTF __attribute__((const))
465 # define XXH_PUREF __attribute__((pure))
466 # define XXH_MALLOCF __attribute__((malloc))
467 #else
468 # define XXH_CONSTF /* disable */
469 # define XXH_PUREF
470 # define XXH_MALLOCF
471 #endif
472
473 /* *************************************
474 * Version
475 ***************************************/
476 #define XXH_VERSION_MAJOR 0
477 #define XXH_VERSION_MINOR 8
478 #define XXH_VERSION_RELEASE 2
479 /*! @brief Version number, encoded as two digits each */
480 #define XXH_VERSION_NUMBER (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
481
482 /*!
483 * @brief Obtains the xxHash version.
484 *
485 * This is mostly useful when xxHash is compiled as a shared library,
486 * since the returned value comes from the library, as opposed to header file.
487 *
488 * @return @ref XXH_VERSION_NUMBER of the invoked library.
489 */
490 XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
491
492
493 /* ****************************
494 * Common basic types
495 ******************************/
496 #include <stddef.h> /* size_t */
497 /*!
498 * @brief Exit code for the streaming API.
499 */
500 typedef enum {
501 XXH_OK = 0, /*!< OK */
502 XXH_ERROR /*!< Error */
503 } XXH_errorcode;
504
505
506 /*-**********************************************************************
507 * 32-bit hash
508 ************************************************************************/
509 #if defined(XXH_DOXYGEN) /* Don't show <stdint.h> include */
510 /*!
511 * @brief An unsigned 32-bit integer.
512 *
513 * Not necessarily defined to `uint32_t` but functionally equivalent.
514 */
515 typedef uint32_t XXH32_hash_t;
516
517 #elif !defined (__VMS) \
518 && (defined (__cplusplus) \
519 || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
520 # include <stdint.h>
521 typedef uint32_t XXH32_hash_t;
522
523 #else
524 # include <limits.h>
525 # if UINT_MAX == 0xFFFFFFFFUL
526 typedef unsigned int XXH32_hash_t;
527 # elif ULONG_MAX == 0xFFFFFFFFUL
528 typedef unsigned long XXH32_hash_t;
529 # else
530 # error "unsupported platform: need a 32-bit type"
531 # endif
532 #endif
533
534 /*!
535 * @}
536 *
537 * @defgroup XXH32_family XXH32 family
538 * @ingroup public
539 * Contains functions used in the classic 32-bit xxHash algorithm.
540 *
541 * @note
542 * XXH32 is useful for older platforms, with no or poor 64-bit performance.
543 * Note that the @ref XXH3_family provides competitive speed for both 32-bit
544 * and 64-bit systems, and offers true 64/128 bit hash results.
545 *
546 * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
547 * @see @ref XXH32_impl for implementation details
548 * @{
549 */
550
551 /*!
552 * @brief Calculates the 32-bit hash of @p input using xxHash32.
553 *
554 * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s
555 *
556 * See @ref single_shot_example "Single Shot Example" for an example.
557 *
558 * @param input The block of data to be hashed, at least @p length bytes in size.
559 * @param length The length of @p input, in bytes.
560 * @param seed The 32-bit seed to alter the hash's output predictably.
561 *
562 * @pre
563 * The memory between @p input and @p input + @p length must be valid,
564 * readable, contiguous memory. However, if @p length is `0`, @p input may be
565 * `NULL`. In C++, this also must be *TriviallyCopyable*.
566 *
567 * @return The calculated 32-bit hash value.
568 *
569 * @see
570 * XXH64(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
571 * Direct equivalents for the other variants of xxHash.
572 * @see
573 * XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version.
574 */
575 XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
576
577 #ifndef XXH_NO_STREAM
578 /*!
579 * Streaming functions generate the xxHash value from an incremental input.
580 * This method is slower than single-call functions, due to state management.
581 * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
582 *
583 * An XXH state must first be allocated using `XXH*_createState()`.
584 *
585 * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
586 *
587 * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
588 *
589 * The function returns an error code, with 0 meaning OK, and any other value
590 * meaning there is an error.
591 *
592 * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
593 * This function returns the nn-bits hash as an int or long long.
594 *
595 * It's still possible to continue inserting input into the hash state after a
596 * digest, and generate new hash values later on by invoking `XXH*_digest()`.
597 *
598 * When done, release the state using `XXH*_freeState()`.
599 *
600 * @see streaming_example at the top of @ref xxhash.h for an example.
601 */
602
603 /*!
604 * @typedef struct XXH32_state_s XXH32_state_t
605 * @brief The opaque state struct for the XXH32 streaming API.
606 *
607 * @see XXH32_state_s for details.
608 */
609 typedef struct XXH32_state_s XXH32_state_t;
610
611 /*!
612 * @brief Allocates an @ref XXH32_state_t.
613 *
614 * Must be freed with XXH32_freeState().
615 * @return An allocated XXH32_state_t on success, `NULL` on failure.
616 */
617 XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
618 /*!
619 * @brief Frees an @ref XXH32_state_t.
620 *
621 * Must be allocated with XXH32_createState().
622 * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
623 * @return XXH_OK.
624 */
625 XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr);
626 /*!
627 * @brief Copies one @ref XXH32_state_t to another.
628 *
629 * @param dst_state The state to copy to.
630 * @param src_state The state to copy from.
631 * @pre
632 * @p dst_state and @p src_state must not be `NULL` and must not overlap.
633 */
634 XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
635
636 /*!
637 * @brief Resets an @ref XXH32_state_t to begin a new hash.
638 *
639 * This function resets and seeds a state. Call it before @ref XXH32_update().
640 *
641 * @param statePtr The state struct to reset.
642 * @param seed The 32-bit seed to alter the hash result predictably.
643 *
644 * @pre
645 * @p statePtr must not be `NULL`.
646 *
647 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
648 */
649 XXH_PUBLIC_API XXH_errorcode XXH32_reset (XXH32_state_t* statePtr, XXH32_hash_t seed);
650
651 /*!
652 * @brief Consumes a block of @p input to an @ref XXH32_state_t.
653 *
654 * Call this to incrementally consume blocks of data.
655 *
656 * @param statePtr The state struct to update.
657 * @param input The block of data to be hashed, at least @p length bytes in size.
658 * @param length The length of @p input, in bytes.
659 *
660 * @pre
661 * @p statePtr must not be `NULL`.
662 * @pre
663 * The memory between @p input and @p input + @p length must be valid,
664 * readable, contiguous memory. However, if @p length is `0`, @p input may be
665 * `NULL`. In C++, this also must be *TriviallyCopyable*.
666 *
667 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
668 */
669 XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
670
671 /*!
672 * @brief Returns the calculated hash value from an @ref XXH32_state_t.
673 *
674 * @note
675 * Calling XXH32_digest() will not affect @p statePtr, so you can update,
676 * digest, and update again.
677 *
678 * @param statePtr The state struct to calculate the hash from.
679 *
680 * @pre
681 * @p statePtr must not be `NULL`.
682 *
683 * @return The calculated xxHash32 value from that state.
684 */
685 XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
686 #endif /* !XXH_NO_STREAM */
687
688 /******* Canonical representation *******/
689
690 /*
691 * The default return values from XXH functions are unsigned 32 and 64 bit
692 * integers.
693 * This the simplest and fastest format for further post-processing.
694 *
695 * However, this leaves open the question of what is the order on the byte level,
696 * since little and big endian conventions will store the same number differently.
697 *
698 * The canonical representation settles this issue by mandating big-endian
699 * convention, the same convention as human-readable numbers (large digits first).
700 *
701 * When writing hash values to storage, sending them over a network, or printing
702 * them, it's highly recommended to use the canonical representation to ensure
703 * portability across a wider range of systems, present and future.
704 *
705 * The following functions allow transformation of hash values to and from
706 * canonical format.
707 */
708
709 /*!
710 * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
711 */
712 typedef struct {
713 unsigned char digest[4]; /*!< Hash bytes, big endian */
714 } XXH32_canonical_t;
715
716 /*!
717 * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
718 *
719 * @param dst The @ref XXH32_canonical_t pointer to be stored to.
720 * @param hash The @ref XXH32_hash_t to be converted.
721 *
722 * @pre
723 * @p dst must not be `NULL`.
724 */
725 XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
726
727 /*!
728 * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t.
729 *
730 * @param src The @ref XXH32_canonical_t to convert.
731 *
732 * @pre
733 * @p src must not be `NULL`.
734 *
735 * @return The converted hash.
736 */
737 XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
738
739
740 /*! @cond Doxygen ignores this part */
741 #ifdef __has_attribute
742 # define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
743 #else
744 # define XXH_HAS_ATTRIBUTE(x) 0
745 #endif
746 /*! @endcond */
747
748 /*! @cond Doxygen ignores this part */
749 /*
750 * C23 __STDC_VERSION__ number hasn't been specified yet. For now
751 * leave as `201711L` (C17 + 1).
752 * TODO: Update to correct value when its been specified.
753 */
754 #define XXH_C23_VN 201711L
755 /*! @endcond */
756
757 /*! @cond Doxygen ignores this part */
758 /* C-language Attributes are added in C23. */
759 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
760 # define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
761 #else
762 # define XXH_HAS_C_ATTRIBUTE(x) 0
763 #endif
764 /*! @endcond */
765
766 /*! @cond Doxygen ignores this part */
767 #if defined(__cplusplus) && defined(__has_cpp_attribute)
768 # define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
769 #else
770 # define XXH_HAS_CPP_ATTRIBUTE(x) 0
771 #endif
772 /*! @endcond */
773
774 /*! @cond Doxygen ignores this part */
775 /*
776 * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
777 * introduced in CPP17 and C23.
778 * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
779 * C23 : https://en.cppreference.com/w/c/language/attributes/fallthrough
780 */
781 #if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough)
782 # define XXH_FALLTHROUGH [[fallthrough]]
783 #elif XXH_HAS_ATTRIBUTE(__fallthrough__)
784 # define XXH_FALLTHROUGH __attribute__ ((__fallthrough__))
785 #else
786 # define XXH_FALLTHROUGH /* fallthrough */
787 #endif
788 /*! @endcond */
789
790 /*! @cond Doxygen ignores this part */
791 /*
792 * Define XXH_NOESCAPE for annotated pointers in public API.
793 * https://clang.llvm.org/docs/AttributeReference.html#noescape
794 * As of writing this, only supported by clang.
795 */
796 #if XXH_HAS_ATTRIBUTE(noescape)
797 # define XXH_NOESCAPE __attribute__((noescape))
798 #else
799 # define XXH_NOESCAPE
800 #endif
801 /*! @endcond */
802
803
804 /*!
805 * @}
806 * @ingroup public
807 * @{
808 */
809
810 #ifndef XXH_NO_LONG_LONG
811 /*-**********************************************************************
812 * 64-bit hash
813 ************************************************************************/
814 #if defined(XXH_DOXYGEN) /* don't include <stdint.h> */
815 /*!
816 * @brief An unsigned 64-bit integer.
817 *
818 * Not necessarily defined to `uint64_t` but functionally equivalent.
819 */
820 typedef uint64_t XXH64_hash_t;
821 #elif !defined (__VMS) \
822 && (defined (__cplusplus) \
823 || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
824 # include <stdint.h>
825 typedef uint64_t XXH64_hash_t;
826 #else
827 # include <limits.h>
828 # if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL
829 /* LP64 ABI says uint64_t is unsigned long */
830 typedef unsigned long XXH64_hash_t;
831 # else
832 /* the following type must have a width of 64-bit */
833 typedef unsigned long long XXH64_hash_t;
834 # endif
835 #endif
836
837 /*!
838 * @}
839 *
840 * @defgroup XXH64_family XXH64 family
841 * @ingroup public
842 * @{
843 * Contains functions used in the classic 64-bit xxHash algorithm.
844 *
845 * @note
846 * XXH3 provides competitive speed for both 32-bit and 64-bit systems,
847 * and offers true 64/128 bit hash results.
848 * It provides better speed for systems with vector processing capabilities.
849 */
850
851 /*!
852 * @brief Calculates the 64-bit hash of @p input using xxHash64.
853 *
854 * This function usually runs faster on 64-bit systems, but slower on 32-bit
855 * systems (see benchmark).
856 *
857 * @param input The block of data to be hashed, at least @p length bytes in size.
858 * @param length The length of @p input, in bytes.
859 * @param seed The 64-bit seed to alter the hash's output predictably.
860 *
861 * @pre
862 * The memory between @p input and @p input + @p length must be valid,
863 * readable, contiguous memory. However, if @p length is `0`, @p input may be
864 * `NULL`. In C++, this also must be *TriviallyCopyable*.
865 *
866 * @return The calculated 64-bit hash.
867 *
868 * @see
869 * XXH32(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
870 * Direct equivalents for the other variants of xxHash.
871 * @see
872 * XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
873 */
874 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
875
876 /******* Streaming *******/
877 #ifndef XXH_NO_STREAM
878 /*!
879 * @brief The opaque state struct for the XXH64 streaming API.
880 *
881 * @see XXH64_state_s for details.
882 */
883 typedef struct XXH64_state_s XXH64_state_t; /* incomplete type */
884
885 /*!
886 * @brief Allocates an @ref XXH64_state_t.
887 *
888 * Must be freed with XXH64_freeState().
889 * @return An allocated XXH64_state_t on success, `NULL` on failure.
890 */
891 XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
892
893 /*!
894 * @brief Frees an @ref XXH64_state_t.
895 *
896 * Must be allocated with XXH64_createState().
897 * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState().
898 * @return XXH_OK.
899 */
900 XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr);
901
902 /*!
903 * @brief Copies one @ref XXH64_state_t to another.
904 *
905 * @param dst_state The state to copy to.
906 * @param src_state The state to copy from.
907 * @pre
908 * @p dst_state and @p src_state must not be `NULL` and must not overlap.
909 */
910 XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
911
912 /*!
913 * @brief Resets an @ref XXH64_state_t to begin a new hash.
914 *
915 * This function resets and seeds a state. Call it before @ref XXH64_update().
916 *
917 * @param statePtr The state struct to reset.
918 * @param seed The 64-bit seed to alter the hash result predictably.
919 *
920 * @pre
921 * @p statePtr must not be `NULL`.
922 *
923 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
924 */
925 XXH_PUBLIC_API XXH_errorcode XXH64_reset (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
926
927 /*!
928 * @brief Consumes a block of @p input to an @ref XXH64_state_t.
929 *
930 * Call this to incrementally consume blocks of data.
931 *
932 * @param statePtr The state struct to update.
933 * @param input The block of data to be hashed, at least @p length bytes in size.
934 * @param length The length of @p input, in bytes.
935 *
936 * @pre
937 * @p statePtr must not be `NULL`.
938 * @pre
939 * The memory between @p input and @p input + @p length must be valid,
940 * readable, contiguous memory. However, if @p length is `0`, @p input may be
941 * `NULL`. In C++, this also must be *TriviallyCopyable*.
942 *
943 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
944 */
945 XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
946
947 /*!
948 * @brief Returns the calculated hash value from an @ref XXH64_state_t.
949 *
950 * @note
951 * Calling XXH64_digest() will not affect @p statePtr, so you can update,
952 * digest, and update again.
953 *
954 * @param statePtr The state struct to calculate the hash from.
955 *
956 * @pre
957 * @p statePtr must not be `NULL`.
958 *
959 * @return The calculated xxHash64 value from that state.
960 */
961 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
962 #endif /* !XXH_NO_STREAM */
963 /******* Canonical representation *******/
964
965 /*!
966 * @brief Canonical (big endian) representation of @ref XXH64_hash_t.
967 */
968 typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
969
970 /*!
971 * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t.
972 *
973 * @param dst The @ref XXH64_canonical_t pointer to be stored to.
974 * @param hash The @ref XXH64_hash_t to be converted.
975 *
976 * @pre
977 * @p dst must not be `NULL`.
978 */
979 XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
980
981 /*!
982 * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t.
983 *
984 * @param src The @ref XXH64_canonical_t to convert.
985 *
986 * @pre
987 * @p src must not be `NULL`.
988 *
989 * @return The converted hash.
990 */
991 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
992
993 #ifndef XXH_NO_XXH3
994
995 /*!
996 * @}
997 * ************************************************************************
998 * @defgroup XXH3_family XXH3 family
999 * @ingroup public
1000 * @{
1001 *
1002 * XXH3 is a more recent hash algorithm featuring:
1003 * - Improved speed for both small and large inputs
1004 * - True 64-bit and 128-bit outputs
1005 * - SIMD acceleration
1006 * - Improved 32-bit viability
1007 *
1008 * Speed analysis methodology is explained here:
1009 *
1010 * https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
1011 *
1012 * Compared to XXH64, expect XXH3 to run approximately
1013 * ~2x faster on large inputs and >3x faster on small ones,
1014 * exact differences vary depending on platform.
1015 *
1016 * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
1017 * but does not require it.
1018 * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
1019 * at competitive speeds, even without vector support. Further details are
1020 * explained in the implementation.
1021 *
1022 * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD
1023 * implementations for many common platforms:
1024 * - AVX512
1025 * - AVX2
1026 * - SSE2
1027 * - ARM NEON
1028 * - WebAssembly SIMD128
1029 * - POWER8 VSX
1030 * - s390x ZVector
1031 * This can be controlled via the @ref XXH_VECTOR macro, but it automatically
1032 * selects the best version according to predefined macros. For the x86 family, an
1033 * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c.
1034 *
1035 * XXH3 implementation is portable:
1036 * it has a generic C90 formulation that can be compiled on any platform,
1037 * all implementations generate exactly the same hash value on all platforms.
1038 * Starting from v0.8.0, it's also labelled "stable", meaning that
1039 * any future version will also generate the same hash value.
1040 *
1041 * XXH3 offers 2 variants, _64bits and _128bits.
1042 *
1043 * When only 64 bits are needed, prefer invoking the _64bits variant, as it
1044 * reduces the amount of mixing, resulting in faster speed on small inputs.
1045 * It's also generally simpler to manipulate a scalar return type than a struct.
1046 *
1047 * The API supports one-shot hashing, streaming mode, and custom secrets.
1048 */
1049 /*-**********************************************************************
1050 * XXH3 64-bit variant
1051 ************************************************************************/
1052
1053 /*!
1054 * @brief 64-bit unseeded variant of XXH3.
1055 *
1056 * This is equivalent to @ref XXH3_64bits_withSeed() with a seed of 0, however
1057 * it may have slightly better performance due to constant propagation of the
1058 * defaults.
1059 *
1060 * @see
1061 * XXH32(), XXH64(), XXH3_128bits(): equivalent for the other xxHash algorithms
1062 * @see
1063 * XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
1064 * @see
1065 * XXH3_64bits_reset(), XXH3_64bits_update(), XXH3_64bits_digest(): Streaming version.
1066 */
1067 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
1068
1069 /*!
1070 * @brief 64-bit seeded variant of XXH3
1071 *
1072 * This variant generates a custom secret on the fly based on default secret
1073 * altered using the `seed` value.
1074 *
1075 * While this operation is decently fast, note that it's not completely free.
1076 *
1077 * @note
1078 * seed == 0 produces the same results as @ref XXH3_64bits().
1079 *
1080 * @param input The data to hash
1081 * @param length The length
1082 * @param seed The 64-bit seed to alter the state.
1083 */
1084 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
1085
1086 /*!
1087 * The bare minimum size for a custom secret.
1088 *
1089 * @see
1090 * XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(),
1091 * XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
1092 */
1093 #define XXH3_SECRET_SIZE_MIN 136
1094
1095 /*!
1096 * @brief 64-bit variant of XXH3 with a custom "secret".
1097 *
1098 * It's possible to provide any blob of bytes as a "secret" to generate the hash.
1099 * This makes it more difficult for an external actor to prepare an intentional collision.
1100 * The main condition is that secretSize *must* be large enough (>= XXH3_SECRET_SIZE_MIN).
1101 * However, the quality of the secret impacts the dispersion of the hash algorithm.
1102 * Therefore, the secret _must_ look like a bunch of random bytes.
1103 * Avoid "trivial" or structured data such as repeated sequences or a text document.
1104 * Whenever in doubt about the "randomness" of the blob of bytes,
1105 * consider employing "XXH3_generateSecret()" instead (see below).
1106 * It will generate a proper high entropy secret derived from the blob of bytes.
1107 * Another advantage of using XXH3_generateSecret() is that
1108 * it guarantees that all bits within the initial blob of bytes
1109 * will impact every bit of the output.
1110 * This is not necessarily the case when using the blob of bytes directly
1111 * because, when hashing _small_ inputs, only a portion of the secret is employed.
1112 */
1113 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
1114
1115
1116 /******* Streaming *******/
1117 #ifndef XXH_NO_STREAM
1118 /*
1119 * Streaming requires state maintenance.
1120 * This operation costs memory and CPU.
1121 * As a consequence, streaming is slower than one-shot hashing.
1122 * For better performance, prefer one-shot functions whenever applicable.
1123 */
1124
1125 /*!
1126 * @brief The state struct for the XXH3 streaming API.
1127 *
1128 * @see XXH3_state_s for details.
1129 */
1130 typedef struct XXH3_state_s XXH3_state_t;
1131 XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
1132 XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
1133
1134 /*!
1135 * @brief Copies one @ref XXH3_state_t to another.
1136 *
1137 * @param dst_state The state to copy to.
1138 * @param src_state The state to copy from.
1139 * @pre
1140 * @p dst_state and @p src_state must not be `NULL` and must not overlap.
1141 */
1142 XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
1143
1144 /*!
1145 * @brief Resets an @ref XXH3_state_t to begin a new hash.
1146 *
1147 * This function resets `statePtr` and generate a secret with default parameters. Call it before @ref XXH3_64bits_update().
1148 * Digest will be equivalent to `XXH3_64bits()`.
1149 *
1150 * @param statePtr The state struct to reset.
1151 *
1152 * @pre
1153 * @p statePtr must not be `NULL`.
1154 *
1155 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
1156 *
1157 */
1158 XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
1159
1160 /*!
1161 * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
1162 *
1163 * This function resets `statePtr` and generate a secret from `seed`. Call it before @ref XXH3_64bits_update().
1164 * Digest will be equivalent to `XXH3_64bits_withSeed()`.
1165 *
1166 * @param statePtr The state struct to reset.
1167 * @param seed The 64-bit seed to alter the state.
1168 *
1169 * @pre
1170 * @p statePtr must not be `NULL`.
1171 *
1172 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
1173 *
1174 */
1175 XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
1176
1177 /*!
1178 * XXH3_64bits_reset_withSecret():
1179 * `secret` is referenced, it _must outlive_ the hash streaming session.
1180 * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`,
1181 * and the quality of produced hash values depends on secret's entropy
1182 * (secret's content should look like a bunch of random bytes).
1183 * When in doubt about the randomness of a candidate `secret`,
1184 * consider employing `XXH3_generateSecret()` instead (see below).
1185 */
1186 XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
1187
1188 /*!
1189 * @brief Consumes a block of @p input to an @ref XXH3_state_t.
1190 *
1191 * Call this to incrementally consume blocks of data.
1192 *
1193 * @param statePtr The state struct to update.
1194 * @param input The block of data to be hashed, at least @p length bytes in size.
1195 * @param length The length of @p input, in bytes.
1196 *
1197 * @pre
1198 * @p statePtr must not be `NULL`.
1199 * @pre
1200 * The memory between @p input and @p input + @p length must be valid,
1201 * readable, contiguous memory. However, if @p length is `0`, @p input may be
1202 * `NULL`. In C++, this also must be *TriviallyCopyable*.
1203 *
1204 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
1205 */
1206 XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
1207
1208 /*!
1209 * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t.
1210 *
1211 * @note
1212 * Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update,
1213 * digest, and update again.
1214 *
1215 * @param statePtr The state struct to calculate the hash from.
1216 *
1217 * @pre
1218 * @p statePtr must not be `NULL`.
1219 *
1220 * @return The calculated XXH3 64-bit hash value from that state.
1221 */
1222 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
1223 #endif /* !XXH_NO_STREAM */
1224
1225 /* note : canonical representation of XXH3 is the same as XXH64
1226 * since they both produce XXH64_hash_t values */
1227
1228
1229 /*-**********************************************************************
1230 * XXH3 128-bit variant
1231 ************************************************************************/
1232
1233 /*!
1234 * @brief The return value from 128-bit hashes.
1235 *
1236 * Stored in little endian order, although the fields themselves are in native
1237 * endianness.
1238 */
1239 typedef struct {
1240 XXH64_hash_t low64; /*!< `value & 0xFFFFFFFFFFFFFFFF` */
1241 XXH64_hash_t high64; /*!< `value >> 64` */
1242 } XXH128_hash_t;
1243
1244 /*!
1245 * @brief Unseeded 128-bit variant of XXH3
1246 *
1247 * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
1248 * for shorter inputs.
1249 *
1250 * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of 0, however
1251 * it may have slightly better performance due to constant propagation of the
1252 * defaults.
1253 *
1254 * @see
1255 * XXH32(), XXH64(), XXH3_64bits(): equivalent for the other xxHash algorithms
1256 * @see
1257 * XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
1258 * @see
1259 * XXH3_128bits_reset(), XXH3_128bits_update(), XXH3_128bits_digest(): Streaming version.
1260 */
1261 XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
1262 /*! @brief Seeded 128-bit variant of XXH3. @see XXH3_64bits_withSeed(). */
1263 XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
1264 /*! @brief Custom secret 128-bit variant of XXH3. @see XXH3_64bits_withSecret(). */
1265 XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
1266
1267 /******* Streaming *******/
1268 #ifndef XXH_NO_STREAM
1269 /*
1270 * Streaming requires state maintenance.
1271 * This operation costs memory and CPU.
1272 * As a consequence, streaming is slower than one-shot hashing.
1273 * For better performance, prefer one-shot functions whenever applicable.
1274 *
1275 * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits().
1276 * Use already declared XXH3_createState() and XXH3_freeState().
1277 *
1278 * All reset and streaming functions have same meaning as their 64-bit counterpart.
1279 */
1280
1281 /*!
1282 * @brief Resets an @ref XXH3_state_t to begin a new hash.
1283 *
1284 * This function resets `statePtr` and generate a secret with default parameters. Call it before @ref XXH3_128bits_update().
1285 * Digest will be equivalent to `XXH3_128bits()`.
1286 *
1287 * @param statePtr The state struct to reset.
1288 *
1289 * @pre
1290 * @p statePtr must not be `NULL`.
1291 *
1292 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
1293 *
1294 */
1295 XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
1296
1297 /*!
1298 * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
1299 *
1300 * This function resets `statePtr` and generate a secret from `seed`. Call it before @ref XXH3_128bits_update().
1301 * Digest will be equivalent to `XXH3_128bits_withSeed()`.
1302 *
1303 * @param statePtr The state struct to reset.
1304 * @param seed The 64-bit seed to alter the state.
1305 *
1306 * @pre
1307 * @p statePtr must not be `NULL`.
1308 *
1309 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
1310 *
1311 */
1312 XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
1313 /*! @brief Custom secret 128-bit variant of XXH3. @see XXH_64bits_reset_withSecret(). */
1314 XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
1315
1316 /*!
1317 * @brief Consumes a block of @p input to an @ref XXH3_state_t.
1318 *
1319 * Call this to incrementally consume blocks of data.
1320 *
1321 * @param statePtr The state struct to update.
1322 * @param input The block of data to be hashed, at least @p length bytes in size.
1323 * @param length The length of @p input, in bytes.
1324 *
1325 * @pre
1326 * @p statePtr must not be `NULL`.
1327 * @pre
1328 * The memory between @p input and @p input + @p length must be valid,
1329 * readable, contiguous memory. However, if @p length is `0`, @p input may be
1330 * `NULL`. In C++, this also must be *TriviallyCopyable*.
1331 *
1332 * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
1333 */
1334 XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
1335
1336 /*!
1337 * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t.
1338 *
1339 * @note
1340 * Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update,
1341 * digest, and update again.
1342 *
1343 * @param statePtr The state struct to calculate the hash from.
1344 *
1345 * @pre
1346 * @p statePtr must not be `NULL`.
1347 *
1348 * @return The calculated XXH3 128-bit hash value from that state.
1349 */
1350 XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
1351 #endif /* !XXH_NO_STREAM */
1352
1353 /* Following helper functions make it possible to compare XXH128_hast_t values.
1354 * Since XXH128_hash_t is a structure, this capability is not offered by the language.
1355 * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
1356
1357 /*!
1358 * XXH128_isEqual():
1359 * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
1360 */
1361 XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
1362
1363 /*!
1364 * @brief Compares two @ref XXH128_hash_t
1365 * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
1366 *
1367 * @return: >0 if *h128_1 > *h128_2
1368 * =0 if *h128_1 == *h128_2
1369 * <0 if *h128_1 < *h128_2
1370 */
1371 XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
1372
1373
1374 /******* Canonical representation *******/
1375 typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
1376
1377
1378 /*!
1379 * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t.
1380 *
1381 * @param dst The @ref XXH128_canonical_t pointer to be stored to.
1382 * @param hash The @ref XXH128_hash_t to be converted.
1383 *
1384 * @pre
1385 * @p dst must not be `NULL`.
1386 */
1387 XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
1388
1389 /*!
1390 * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t.
1391 *
1392 * @param src The @ref XXH128_canonical_t to convert.
1393 *
1394 * @pre
1395 * @p src must not be `NULL`.
1396 *
1397 * @return The converted hash.
1398 */
1399 XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
1400
1401
1402 #endif /* !XXH_NO_XXH3 */
1403 #endif /* XXH_NO_LONG_LONG */
1404
1405 /*!
1406 * @}
1407 */
1408 #endif /* XXHASH_H_5627135585666179 */
1409
1410
1411
1412 #if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
1413 #define XXHASH_H_STATIC_13879238742
1414 /* ****************************************************************************
1415 * This section contains declarations which are not guaranteed to remain stable.
1416 * They may change in future versions, becoming incompatible with a different
1417 * version of the library.
1418 * These declarations should only be used with static linking.
1419 * Never use them in association with dynamic linking!
1420 ***************************************************************************** */
1421
1422 /*
1423 * These definitions are only present to allow static allocation
1424 * of XXH states, on stack or in a struct, for example.
1425 * Never **ever** access their members directly.
1426 */
1427
1428 /*!
1429 * @internal
1430 * @brief Structure for XXH32 streaming API.
1431 *
1432 * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
1433 * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
1434 * an opaque type. This allows fields to safely be changed.
1435 *
1436 * Typedef'd to @ref XXH32_state_t.
1437 * Do not access the members of this struct directly.
1438 * @see XXH64_state_s, XXH3_state_s
1439 */
1440 struct XXH32_state_s {
1441 XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
1442 XXH32_hash_t large_len; /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
1443 XXH32_hash_t v[4]; /*!< Accumulator lanes */
1444 XXH32_hash_t mem32[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
1445 XXH32_hash_t memsize; /*!< Amount of data in @ref mem32 */
1446 XXH32_hash_t reserved; /*!< Reserved field. Do not read nor write to it. */
1447 }; /* typedef'd to XXH32_state_t */
1448
1449
1450 #ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */
1451
1452 /*!
1453 * @internal
1454 * @brief Structure for XXH64 streaming API.
1455 *
1456 * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
1457 * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
1458 * an opaque type. This allows fields to safely be changed.
1459 *
1460 * Typedef'd to @ref XXH64_state_t.
1461 * Do not access the members of this struct directly.
1462 * @see XXH32_state_s, XXH3_state_s
1463 */
1464 struct XXH64_state_s {
1465 XXH64_hash_t total_len; /*!< Total length hashed. This is always 64-bit. */
1466 XXH64_hash_t v[4]; /*!< Accumulator lanes */
1467 XXH64_hash_t mem64[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
1468 XXH32_hash_t memsize; /*!< Amount of data in @ref mem64 */
1469 XXH32_hash_t reserved32; /*!< Reserved field, needed for padding anyways*/
1470 XXH64_hash_t reserved64; /*!< Reserved field. Do not read or write to it. */
1471 }; /* typedef'd to XXH64_state_t */
1472
1473 #ifndef XXH_NO_XXH3
1474
1475 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
1476 # include <stdalign.h>
1477 # define XXH_ALIGN(n) alignas(n)
1478 #elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */
1479 /* In C++ alignas() is a keyword */
1480 # define XXH_ALIGN(n) alignas(n)
1481 #elif defined(__GNUC__)
1482 # define XXH_ALIGN(n) __attribute__ ((aligned(n)))
1483 #elif defined(_MSC_VER)
1484 # define XXH_ALIGN(n) __declspec(align(n))
1485 #else
1486 # define XXH_ALIGN(n) /* disabled */
1487 #endif
1488
1489 /* Old GCC versions only accept the attribute after the type in structures. */
1490 #if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) /* C11+ */ \
1491 && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \
1492 && defined(__GNUC__)
1493 # define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
1494 #else
1495 # define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
1496 #endif
1497
1498 /*!
1499 * @brief The size of the internal XXH3 buffer.
1500 *
1501 * This is the optimal update size for incremental hashing.
1502 *
1503 * @see XXH3_64b_update(), XXH3_128b_update().
1504 */
1505 #define XXH3_INTERNALBUFFER_SIZE 256
1506
1507 /*!
1508 * @internal
1509 * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
1510 *
1511 * This is the size used in @ref XXH3_kSecret and the seeded functions.
1512 *
1513 * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
1514 */
1515 #define XXH3_SECRET_DEFAULT_SIZE 192
1516
1517 /*!
1518 * @internal
1519 * @brief Structure for XXH3 streaming API.
1520 *
1521 * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
1522 * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined.
1523 * Otherwise it is an opaque type.
1524 * Never use this definition in combination with dynamic library.
1525 * This allows fields to safely be changed in the future.
1526 *
1527 * @note ** This structure has a strict alignment requirement of 64 bytes!! **
1528 * Do not allocate this with `malloc()` or `new`,
1529 * it will not be sufficiently aligned.
1530 * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation.
1531 *
1532 * Typedef'd to @ref XXH3_state_t.
1533 * Do never access the members of this struct directly.
1534 *
1535 * @see XXH3_INITSTATE() for stack initialization.
1536 * @see XXH3_createState(), XXH3_freeState().
1537 * @see XXH32_state_s, XXH64_state_s
1538 */
1539 struct XXH3_state_s {
1540 XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
1541 /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
1542 XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
1543 /*!< Used to store a custom secret generated from a seed. */
1544 XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
1545 /*!< The internal buffer. @see XXH32_state_s::mem32 */
1546 XXH32_hash_t bufferedSize;
1547 /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
1548 XXH32_hash_t useSeed;
1549 /*!< Reserved field. Needed for padding on 64-bit. */
1550 size_t nbStripesSoFar;
1551 /*!< Number or stripes processed. */
1552 XXH64_hash_t totalLen;
1553 /*!< Total length hashed. 64-bit even on 32-bit targets. */
1554 size_t nbStripesPerBlock;
1555 /*!< Number of stripes per block. */
1556 size_t secretLimit;
1557 /*!< Size of @ref customSecret or @ref extSecret */
1558 XXH64_hash_t seed;
1559 /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
1560 XXH64_hash_t reserved64;
1561 /*!< Reserved field. */
1562 const unsigned char* extSecret;
1563 /*!< Reference to an external secret for the _withSecret variants, NULL
1564 * for other variants. */
1565 /* note: there may be some padding at the end due to alignment on 64 bytes */
1566 }; /* typedef'd to XXH3_state_t */
1567
1568 #undef XXH_ALIGN_MEMBER
1569
1570 /*!
1571 * @brief Initializes a stack-allocated `XXH3_state_s`.
1572 *
1573 * When the @ref XXH3_state_t structure is merely emplaced on stack,
1574 * it should be initialized with XXH3_INITSTATE() or a memset()
1575 * in case its first reset uses XXH3_NNbits_reset_withSeed().
1576 * This init can be omitted if the first reset uses default or _withSecret mode.
1577 * This operation isn't necessary when the state is created with XXH3_createState().
1578 * Note that this doesn't prepare the state for a streaming operation,
1579 * it's still necessary to use XXH3_NNbits_reset*() afterwards.
1580 */
1581 #define XXH3_INITSTATE(XXH3_state_ptr) \
1582 do { \
1583 XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
1584 tmp_xxh3_state_ptr->seed = 0; \
1585 tmp_xxh3_state_ptr->extSecret = NULL; \
1586 } while(0)
1587
1588
1589 /*!
1590 * simple alias to pre-selected XXH3_128bits variant
1591 */
1592 XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
1593
1594
1595 /* === Experimental API === */
1596 /* Symbols defined below must be considered tied to a specific library version. */
1597
1598 /*!
1599 * XXH3_generateSecret():
1600 *
1601 * Derive a high-entropy secret from any user-defined content, named customSeed.
1602 * The generated secret can be used in combination with `*_withSecret()` functions.
1603 * The `_withSecret()` variants are useful to provide a higher level of protection
1604 * than 64-bit seed, as it becomes much more difficult for an external actor to
1605 * guess how to impact the calculation logic.
1606 *
1607 * The function accepts as input a custom seed of any length and any content,
1608 * and derives from it a high-entropy secret of length @p secretSize into an
1609 * already allocated buffer @p secretBuffer.
1610 *
1611 * The generated secret can then be used with any `*_withSecret()` variant.
1612 * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(),
1613 * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret()
1614 * are part of this list. They all accept a `secret` parameter
1615 * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN)
1616 * _and_ feature very high entropy (consist of random-looking bytes).
1617 * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
1618 * be employed to ensure proper quality.
1619 *
1620 * @p customSeed can be anything. It can have any size, even small ones,
1621 * and its content can be anything, even "poor entropy" sources such as a bunch
1622 * of zeroes. The resulting `secret` will nonetheless provide all required qualities.
1623 *
1624 * @pre
1625 * - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN
1626 * - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
1627 *
1628 * Example code:
1629 * @code{.c}
1630 * #include <stdio.h>
1631 * #include <stdlib.h>
1632 * #include <string.h>
1633 * #define XXH_STATIC_LINKING_ONLY // expose unstable API
1634 * #include "xxhash.h"
1635 * // Hashes argv[2] using the entropy from argv[1].
1636 * int main(int argc, char* argv[])
1637 * {
1638 * char secret[XXH3_SECRET_SIZE_MIN];
1639 * if (argv != 3) { return 1; }
1640 * XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1]));
1641 * XXH64_hash_t h = XXH3_64bits_withSecret(
1642 * argv[2], strlen(argv[2]),
1643 * secret, sizeof(secret)
1644 * );
1645 * printf("%016llx\n", (unsigned long long) h);
1646 * }
1647 * @endcode
1648 */
1649 XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize);
1650
1651 /*!
1652 * @brief Generate the same secret as the _withSeed() variants.
1653 *
1654 * The generated secret can be used in combination with
1655 *`*_withSecret()` and `_withSecretandSeed()` variants.
1656 *
1657 * Example C++ `std::string` hash class:
1658 * @code{.cpp}
1659 * #include <string>
1660 * #define XXH_STATIC_LINKING_ONLY // expose unstable API
1661 * #include "xxhash.h"
1662 * // Slow, seeds each time
1663 * class HashSlow {
1664 * XXH64_hash_t seed;
1665 * public:
1666 * HashSlow(XXH64_hash_t s) : seed{s} {}
1667 * size_t operator()(const std::string& x) const {
1668 * return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)};
1669 * }
1670 * };
1671 * // Fast, caches the seeded secret for future uses.
1672 * class HashFast {
1673 * unsigned char secret[XXH3_SECRET_SIZE_MIN];
1674 * public:
1675 * HashFast(XXH64_hash_t s) {
1676 * XXH3_generateSecret_fromSeed(secret, seed);
1677 * }
1678 * size_t operator()(const std::string& x) const {
1679 * return size_t{
1680 * XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret))
1681 * };
1682 * }
1683 * };
1684 * @endcode
1685 * @param secretBuffer A writable buffer of @ref XXH3_SECRET_SIZE_MIN bytes
1686 * @param seed The seed to seed the state.
1687 */
1688 XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
1689
1690 /*!
1691 * These variants generate hash values using either
1692 * @p seed for "short" keys (< XXH3_MIDSIZE_MAX = 240 bytes)
1693 * or @p secret for "large" keys (>= XXH3_MIDSIZE_MAX).
1694 *
1695 * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
1696 * `_withSeed()` has to generate the secret on the fly for "large" keys.
1697 * It's fast, but can be perceptible for "not so large" keys (< 1 KB).
1698 * `_withSecret()` has to generate the masks on the fly for "small" keys,
1699 * which requires more instructions than _withSeed() variants.
1700 * Therefore, _withSecretandSeed variant combines the best of both worlds.
1701 *
1702 * When @p secret has been generated by XXH3_generateSecret_fromSeed(),
1703 * this variant produces *exactly* the same results as `_withSeed()` variant,
1704 * hence offering only a pure speed benefit on "large" input,
1705 * by skipping the need to regenerate the secret for every large input.
1706 *
1707 * Another usage scenario is to hash the secret to a 64-bit hash value,
1708 * for example with XXH3_64bits(), which then becomes the seed,
1709 * and then employ both the seed and the secret in _withSecretandSeed().
1710 * On top of speed, an added benefit is that each bit in the secret
1711 * has a 50% chance to swap each bit in the output, via its impact to the seed.
1712 *
1713 * This is not guaranteed when using the secret directly in "small data" scenarios,
1714 * because only portions of the secret are employed for small data.
1715 */
1716 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
1717 XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len,
1718 XXH_NOESCAPE const void* secret, size_t secretSize,
1719 XXH64_hash_t seed);
1720 /*! @copydoc XXH3_64bits_withSecretandSeed() */
1721 XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
1722 XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
1723 XXH_NOESCAPE const void* secret, size_t secretSize,
1724 XXH64_hash_t seed64);
1725 #ifndef XXH_NO_STREAM
1726 /*! @copydoc XXH3_64bits_withSecretandSeed() */
1727 XXH_PUBLIC_API XXH_errorcode
1728 XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
1729 XXH_NOESCAPE const void* secret, size_t secretSize,
1730 XXH64_hash_t seed64);
1731 /*! @copydoc XXH3_64bits_withSecretandSeed() */
1732 XXH_PUBLIC_API XXH_errorcode
1733 XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
1734 XXH_NOESCAPE const void* secret, size_t secretSize,
1735 XXH64_hash_t seed64);
1736 #endif /* !XXH_NO_STREAM */
1737
1738 #endif /* !XXH_NO_XXH3 */
1739 #endif /* XXH_NO_LONG_LONG */
1740 #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
1741 # define XXH_IMPLEMENTATION
1742 #endif
1743
1744 #endif /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
1745
1746
1747 /* ======================================================================== */
1748 /* ======================================================================== */
1749 /* ======================================================================== */
1750
1751
1752 /*-**********************************************************************
1753 * xxHash implementation
1754 *-**********************************************************************
1755 * xxHash's implementation used to be hosted inside xxhash.c.
1756 *
1757 * However, inlining requires implementation to be visible to the compiler,
1758 * hence be included alongside the header.
1759 * Previously, implementation was hosted inside xxhash.c,
1760 * which was then #included when inlining was activated.
1761 * This construction created issues with a few build and install systems,
1762 * as it required xxhash.c to be stored in /include directory.
1763 *
1764 * xxHash implementation is now directly integrated within xxhash.h.
1765 * As a consequence, xxhash.c is no longer needed in /include.
1766 *
1767 * xxhash.c is still available and is still useful.
1768 * In a "normal" setup, when xxhash is not inlined,
1769 * xxhash.h only exposes the prototypes and public symbols,
1770 * while xxhash.c can be built into an object file xxhash.o
1771 * which can then be linked into the final binary.
1772 ************************************************************************/
1773
1774 #if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
1775 || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
1776 # define XXH_IMPLEM_13a8737387
1777
1778 /* *************************************
1779 * Tuning parameters
1780 ***************************************/
1781
1782 /*!
1783 * @defgroup tuning Tuning parameters
1784 * @{
1785 *
1786 * Various macros to control xxHash's behavior.
1787 */
1788 #ifdef XXH_DOXYGEN
1789 /*!
1790 * @brief Define this to disable 64-bit code.
1791 *
1792 * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
1793 */
1794 # define XXH_NO_LONG_LONG
1795 # undef XXH_NO_LONG_LONG /* don't actually */
1796 /*!
1797 * @brief Controls how unaligned memory is accessed.
1798 *
1799 * By default, access to unaligned memory is controlled by `memcpy()`, which is
1800 * safe and portable.
1801 *
1802 * Unfortunately, on some target/compiler combinations, the generated assembly
1803 * is sub-optimal.
1804 *
1805 * The below switch allow selection of a different access method
1806 * in the search for improved performance.
1807 *
1808 * @par Possible options:
1809 *
1810 * - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy`
1811 * @par
1812 * Use `memcpy()`. Safe and portable. Note that most modern compilers will
1813 * eliminate the function call and treat it as an unaligned access.
1814 *
1815 * - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))`
1816 * @par
1817 * Depends on compiler extensions and is therefore not portable.
1818 * This method is safe _if_ your compiler supports it,
1819 * and *generally* as fast or faster than `memcpy`.
1820 *
1821 * - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast
1822 * @par
1823 * Casts directly and dereferences. This method doesn't depend on the
1824 * compiler, but it violates the C standard as it directly dereferences an
1825 * unaligned pointer. It can generate buggy code on targets which do not
1826 * support unaligned memory accesses, but in some circumstances, it's the
1827 * only known way to get the most performance.
1828 *
1829 * - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift
1830 * @par
1831 * Also portable. This can generate the best code on old compilers which don't
1832 * inline small `memcpy()` calls, and it might also be faster on big-endian
1833 * systems which lack a native byteswap instruction. However, some compilers
1834 * will emit literal byteshifts even if the target supports unaligned access.
1835 *
1836 *
1837 * @warning
1838 * Methods 1 and 2 rely on implementation-defined behavior. Use these with
1839 * care, as what works on one compiler/platform/optimization level may cause
1840 * another to read garbage data or even crash.
1841 *
1842 * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
1843 *
1844 * Prefer these methods in priority order (0 > 3 > 1 > 2)
1845 */
1846 # define XXH_FORCE_MEMORY_ACCESS 0
1847
1848 /*!
1849 * @def XXH_SIZE_OPT
1850 * @brief Controls how much xxHash optimizes for size.
1851 *
1852 * xxHash, when compiled, tends to result in a rather large binary size. This
1853 * is mostly due to heavy usage to forced inlining and constant folding of the
1854 * @ref XXH3_family to increase performance.
1855 *
1856 * However, some developers prefer size over speed. This option can
1857 * significantly reduce the size of the generated code. When using the `-Os`
1858 * or `-Oz` options on GCC or Clang, this is defined to 1 by default,
1859 * otherwise it is defined to 0.
1860 *
1861 * Most of these size optimizations can be controlled manually.
1862 *
1863 * This is a number from 0-2.
1864 * - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed
1865 * comes first.
1866 * - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more
1867 * conservative and disables hacks that increase code size. It implies the
1868 * options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0,
1869 * and @ref XXH3_NEON_LANES == 8 if they are not already defined.
1870 * - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible.
1871 * Performance may cry. For example, the single shot functions just use the
1872 * streaming API.
1873 */
1874 # define XXH_SIZE_OPT 0
1875
1876 /*!
1877 * @def XXH_FORCE_ALIGN_CHECK
1878 * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
1879 * and XXH64() only).
1880 *
1881 * This is an important performance trick for architectures without decent
1882 * unaligned memory access performance.
1883 *
1884 * It checks for input alignment, and when conditions are met, uses a "fast
1885 * path" employing direct 32-bit/64-bit reads, resulting in _dramatically
1886 * faster_ read speed.
1887 *
1888 * The check costs one initial branch per hash, which is generally negligible,
1889 * but not zero.
1890 *
1891 * Moreover, it's not useful to generate an additional code path if memory
1892 * access uses the same instruction for both aligned and unaligned
1893 * addresses (e.g. x86 and aarch64).
1894 *
1895 * In these cases, the alignment check can be removed by setting this macro to 0.
1896 * Then the code will always use unaligned memory access.
1897 * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips
1898 * which are platforms known to offer good unaligned memory accesses performance.
1899 *
1900 * It is also disabled by default when @ref XXH_SIZE_OPT >= 1.
1901 *
1902 * This option does not affect XXH3 (only XXH32 and XXH64).
1903 */
1904 # define XXH_FORCE_ALIGN_CHECK 0
1905
1906 /*!
1907 * @def XXH_NO_INLINE_HINTS
1908 * @brief When non-zero, sets all functions to `static`.
1909 *
1910 * By default, xxHash tries to force the compiler to inline almost all internal
1911 * functions.
1912 *
1913 * This can usually improve performance due to reduced jumping and improved
1914 * constant folding, but significantly increases the size of the binary which
1915 * might not be favorable.
1916 *
1917 * Additionally, sometimes the forced inlining can be detrimental to performance,
1918 * depending on the architecture.
1919 *
1920 * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
1921 * compiler full control on whether to inline or not.
1922 *
1923 * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if
1924 * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
1925 */
1926 # define XXH_NO_INLINE_HINTS 0
1927
1928 /*!
1929 * @def XXH3_INLINE_SECRET
1930 * @brief Determines whether to inline the XXH3 withSecret code.
1931 *
1932 * When the secret size is known, the compiler can improve the performance
1933 * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret().
1934 *
1935 * However, if the secret size is not known, it doesn't have any benefit. This
1936 * happens when xxHash is compiled into a global symbol. Therefore, if
1937 * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0.
1938 *
1939 * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers
1940 * that are *sometimes* force inline on -Og, and it is impossible to automatically
1941 * detect this optimization level.
1942 */
1943 # define XXH3_INLINE_SECRET 0
1944
1945 /*!
1946 * @def XXH32_ENDJMP
1947 * @brief Whether to use a jump for `XXH32_finalize`.
1948 *
1949 * For performance, `XXH32_finalize` uses multiple branches in the finalizer.
1950 * This is generally preferable for performance,
1951 * but depending on exact architecture, a jmp may be preferable.
1952 *
1953 * This setting is only possibly making a difference for very small inputs.
1954 */
1955 # define XXH32_ENDJMP 0
1956
1957 /*!
1958 * @internal
1959 * @brief Redefines old internal names.
1960 *
1961 * For compatibility with code that uses xxHash's internals before the names
1962 * were changed to improve namespacing. There is no other reason to use this.
1963 */
1964 # define XXH_OLD_NAMES
1965 # undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
1966
1967 /*!
1968 * @def XXH_NO_STREAM
1969 * @brief Disables the streaming API.
1970 *
1971 * When xxHash is not inlined and the streaming functions are not used, disabling
1972 * the streaming functions can improve code size significantly, especially with
1973 * the @ref XXH3_family which tends to make constant folded copies of itself.
1974 */
1975 # define XXH_NO_STREAM
1976 # undef XXH_NO_STREAM /* don't actually */
1977 #endif /* XXH_DOXYGEN */
1978 /*!
1979 * @}
1980 */
1981
1982 #ifndef XXH_FORCE_MEMORY_ACCESS /* can be defined externally, on command line for example */
1983 /* prefer __packed__ structures (method 1) for GCC
1984 * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
1985 * which for some reason does unaligned loads. */
1986 # if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED))
1987 # define XXH_FORCE_MEMORY_ACCESS 1
1988 # endif
1989 #endif
1990
1991 #ifndef XXH_SIZE_OPT
1992 /* default to 1 for -Os or -Oz */
1993 # if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__)
1994 # define XXH_SIZE_OPT 1
1995 # else
1996 # define XXH_SIZE_OPT 0
1997 # endif
1998 #endif
1999
2000 #ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */
2001 /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */
2002 # if XXH_SIZE_OPT >= 1 || \
2003 defined(__i386) || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
2004 || defined(_M_IX86) || defined(_M_X64) || defined(_M_ARM64) || defined(_M_ARM) /* visual */
2005 # define XXH_FORCE_ALIGN_CHECK 0
2006 # else
2007 # define XXH_FORCE_ALIGN_CHECK 1
2008 # endif
2009 #endif
2010
2011 #ifndef XXH_NO_INLINE_HINTS
2012 # if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__) /* -O0, -fno-inline */
2013 # define XXH_NO_INLINE_HINTS 1
2014 # else
2015 # define XXH_NO_INLINE_HINTS 0
2016 # endif
2017 #endif
2018
2019 #ifndef XXH3_INLINE_SECRET
2020 # if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \
2021 || !defined(XXH_INLINE_ALL)
2022 # define XXH3_INLINE_SECRET 0
2023 # else
2024 # define XXH3_INLINE_SECRET 1
2025 # endif
2026 #endif
2027
2028 #ifndef XXH32_ENDJMP
2029 /* generally preferable for performance */
2030 # define XXH32_ENDJMP 0
2031 #endif
2032
2033 /*!
2034 * @defgroup impl Implementation
2035 * @{
2036 */
2037
2038
2039 /* *************************************
2040 * Includes & Memory related functions
2041 ***************************************/
2042 #if defined(XXH_NO_STREAM)
2043 /* nothing */
2044 #elif defined(XXH_NO_STDLIB)
2045
2046 /* When requesting to disable any mention of stdlib,
2047 * the library loses the ability to invoked malloc / free.
2048 * In practice, it means that functions like `XXH*_createState()`
2049 * will always fail, and return NULL.
2050 * This flag is useful in situations where
2051 * xxhash.h is integrated into some kernel, embedded or limited environment
2052 * without access to dynamic allocation.
2053 */
2054
2055 static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
2056 static void XXH_free(void* p) { (void)p; }
2057
2058 #else
2059
2060 /*
2061 * Modify the local functions below should you wish to use
2062 * different memory routines for malloc() and free()
2063 */
2064 #include <stdlib.h>
2065
2066 /*!
2067 * @internal
2068 * @brief Modify this function to use a different routine than malloc().
2069 */
2070 static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
2071
2072 /*!
2073 * @internal
2074 * @brief Modify this function to use a different routine than free().
2075 */
2076 static void XXH_free(void* p) { free(p); }
2077
2078 #endif /* XXH_NO_STDLIB */
2079
2080 #include <string.h>
2081
2082 /*!
2083 * @internal
2084 * @brief Modify this function to use a different routine than memcpy().
2085 */
2086 static void* XXH_memcpy(void* dest, const void* src, size_t size)
2087 {
2088 return memcpy(dest,src,size);
2089 }
2090
2091 #include <limits.h> /* ULLONG_MAX */
2092
2093
2094 /* *************************************
2095 * Compiler Specific Options
2096 ***************************************/
2097 #ifdef _MSC_VER /* Visual Studio warning fix */
2098 # pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
2099 #endif
2100
2101 #if XXH_NO_INLINE_HINTS /* disable inlining hints */
2102 # if defined(__GNUC__) || defined(__clang__)
2103 # define XXH_FORCE_INLINE static __attribute__((unused))
2104 # else
2105 # define XXH_FORCE_INLINE static
2106 # endif
2107 # define XXH_NO_INLINE static
2108 /* enable inlining hints */
2109 #elif defined(__GNUC__) || defined(__clang__)
2110 # define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused))
2111 # define XXH_NO_INLINE static __attribute__((noinline))
2112 #elif defined(_MSC_VER) /* Visual Studio */
2113 # define XXH_FORCE_INLINE static __forceinline
2114 # define XXH_NO_INLINE static __declspec(noinline)
2115 #elif defined (__cplusplus) \
2116 || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) /* C99 */
2117 # define XXH_FORCE_INLINE static inline
2118 # define XXH_NO_INLINE static
2119 #else
2120 # define XXH_FORCE_INLINE static
2121 # define XXH_NO_INLINE static
2122 #endif
2123
2124 #if XXH3_INLINE_SECRET
2125 # define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
2126 #else
2127 # define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
2128 #endif
2129
2130
2131 /* *************************************
2132 * Debug
2133 ***************************************/
2134 /*!
2135 * @ingroup tuning
2136 * @def XXH_DEBUGLEVEL
2137 * @brief Sets the debugging level.
2138 *
2139 * XXH_DEBUGLEVEL is expected to be defined externally, typically via the
2140 * compiler's command line options. The value must be a number.
2141 */
2142 #ifndef XXH_DEBUGLEVEL
2143 # ifdef DEBUGLEVEL /* backwards compat */
2144 # define XXH_DEBUGLEVEL DEBUGLEVEL
2145 # else
2146 # define XXH_DEBUGLEVEL 0
2147 # endif
2148 #endif
2149
2150 #if (XXH_DEBUGLEVEL>=1)
2151 # include <assert.h> /* note: can still be disabled with NDEBUG */
2152 # define XXH_ASSERT(c) assert(c)
2153 #else
2154 # if defined(__INTEL_COMPILER)
2155 # define XXH_ASSERT(c) XXH_ASSUME((unsigned char) (c))
2156 # else
2157 # define XXH_ASSERT(c) XXH_ASSUME(c)
2158 # endif
2159 #endif
2160
2161 /* note: use after variable declarations */
2162 #ifndef XXH_STATIC_ASSERT
2163 # if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11 */
2164 # define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
2165 # elif defined(__cplusplus) && (__cplusplus >= 201103L) /* C++11 */
2166 # define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
2167 # else
2168 # define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0)
2169 # endif
2170 # define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c)
2171 #endif
2172
2173 /*!
2174 * @internal
2175 * @def XXH_COMPILER_GUARD(var)
2176 * @brief Used to prevent unwanted optimizations for @p var.
2177 *
2178 * It uses an empty GCC inline assembly statement with a register constraint
2179 * which forces @p var into a general purpose register (eg eax, ebx, ecx
2180 * on x86) and marks it as modified.
2181 *
2182 * This is used in a few places to avoid unwanted autovectorization (e.g.
2183 * XXH32_round()). All vectorization we want is explicit via intrinsics,
2184 * and _usually_ isn't wanted elsewhere.
2185 *
2186 * We also use it to prevent unwanted constant folding for AArch64 in
2187 * XXH3_initCustomSecret_scalar().
2188 */
2189 #if defined(__GNUC__) || defined(__clang__)
2190 # define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
2191 #else
2192 # define XXH_COMPILER_GUARD(var) ((void)0)
2193 #endif
2194
2195 /* Specifically for NEON vectors which use the "w" constraint, on
2196 * Clang. */
2197 #if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__)
2198 # define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var))
2199 #else
2200 # define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0)
2201 #endif
2202
2203 /* *************************************
2204 * Basic Types
2205 ***************************************/
2206 #if !defined (__VMS) \
2207 && (defined (__cplusplus) \
2208 || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
2209 # include <stdint.h>
2210 typedef uint8_t xxh_u8;
2211 #else
2212 typedef unsigned char xxh_u8;
2213 #endif
2214 typedef XXH32_hash_t xxh_u32;
2215
2216 #ifdef XXH_OLD_NAMES
2217 # warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly"
2218 # define BYTE xxh_u8
2219 # define U8 xxh_u8
2220 # define U32 xxh_u32
2221 #endif
2222
2223 /* *** Memory access *** */
2224
2225 /*!
2226 * @internal
2227 * @fn xxh_u32 XXH_read32(const void* ptr)
2228 * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness.
2229 *
2230 * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
2231 *
2232 * @param ptr The pointer to read from.
2233 * @return The 32-bit native endian integer from the bytes at @p ptr.
2234 */
2235
2236 /*!
2237 * @internal
2238 * @fn xxh_u32 XXH_readLE32(const void* ptr)
2239 * @brief Reads an unaligned 32-bit little endian integer from @p ptr.
2240 *
2241 * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
2242 *
2243 * @param ptr The pointer to read from.
2244 * @return The 32-bit little endian integer from the bytes at @p ptr.
2245 */
2246
2247 /*!
2248 * @internal
2249 * @fn xxh_u32 XXH_readBE32(const void* ptr)
2250 * @brief Reads an unaligned 32-bit big endian integer from @p ptr.
2251 *
2252 * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
2253 *
2254 * @param ptr The pointer to read from.
2255 * @return The 32-bit big endian integer from the bytes at @p ptr.
2256 */
2257
2258 /*!
2259 * @internal
2260 * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align)
2261 * @brief Like @ref XXH_readLE32(), but has an option for aligned reads.
2262 *
2263 * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
2264 * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is
2265 * always @ref XXH_alignment::XXH_unaligned.
2266 *
2267 * @param ptr The pointer to read from.
2268 * @param align Whether @p ptr is aligned.
2269 * @pre
2270 * If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte
2271 * aligned.
2272 * @return The 32-bit little endian integer from the bytes at @p ptr.
2273 */
2274
2275 #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
2276 /*
2277 * Manual byteshift. Best for old compilers which don't inline memcpy.
2278 * We actually directly use XXH_readLE32 and XXH_readBE32.
2279 */
2280 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
2281
2282 /*
2283 * Force direct memory access. Only works on CPU which support unaligned memory
2284 * access in hardware.
2285 */
2286 static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
2287
2288 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
2289
2290 /*
2291 * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
2292 * documentation claimed that it only increased the alignment, but actually it
2293 * can decrease it on gcc, clang, and icc:
2294 * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
2295 * https://gcc.godbolt.org/z/xYez1j67Y.
2296 */
2297 #ifdef XXH_OLD_NAMES
2298 typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
2299 #endif
2300 static xxh_u32 XXH_read32(const void* ptr)
2301 {
2302 typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32;
2303 return *((const xxh_unalign32*)ptr);
2304 }
2305
2306 #else
2307
2308 /*
2309 * Portable and safe solution. Generally efficient.
2310 * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
2311 */
2312 static xxh_u32 XXH_read32(const void* memPtr)
2313 {
2314 xxh_u32 val;
2315 XXH_memcpy(&val, memPtr, sizeof(val));
2316 return val;
2317 }
2318
2319 #endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
2320
2321
2322 /* *** Endianness *** */
2323
2324 /*!
2325 * @ingroup tuning
2326 * @def XXH_CPU_LITTLE_ENDIAN
2327 * @brief Whether the target is little endian.
2328 *
2329 * Defined to 1 if the target is little endian, or 0 if it is big endian.
2330 * It can be defined externally, for example on the compiler command line.
2331 *
2332 * If it is not defined,
2333 * a runtime check (which is usually constant folded) is used instead.
2334 *
2335 * @note
2336 * This is not necessarily defined to an integer constant.
2337 *
2338 * @see XXH_isLittleEndian() for the runtime check.
2339 */
2340 #ifndef XXH_CPU_LITTLE_ENDIAN
2341 /*
2342 * Try to detect endianness automatically, to avoid the nonstandard behavior
2343 * in `XXH_isLittleEndian()`
2344 */
2345 # if defined(_WIN32) /* Windows is always little endian */ \
2346 || defined(__LITTLE_ENDIAN__) \
2347 || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
2348 # define XXH_CPU_LITTLE_ENDIAN 1
2349 # elif defined(__BIG_ENDIAN__) \
2350 || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2351 # define XXH_CPU_LITTLE_ENDIAN 0
2352 # else
2353 /*!
2354 * @internal
2355 * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN.
2356 *
2357 * Most compilers will constant fold this.
2358 */
2359 static int XXH_isLittleEndian(void)
2360 {
2361 /*
2362 * Portable and well-defined behavior.
2363 * Don't use static: it is detrimental to performance.
2364 */
2365 const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
2366 return one.c[0];
2367 }
2368 # define XXH_CPU_LITTLE_ENDIAN XXH_isLittleEndian()
2369 # endif
2370 #endif
2371
2372
2373
2374
2375 /* ****************************************
2376 * Compiler-specific Functions and Macros
2377 ******************************************/
2378 #define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
2379
2380 #ifdef __has_builtin
2381 # define XXH_HAS_BUILTIN(x) __has_builtin(x)
2382 #else
2383 # define XXH_HAS_BUILTIN(x) 0
2384 #endif
2385
2386
2387
2388 /*
2389 * C23 and future versions have standard "unreachable()".
2390 * Once it has been implemented reliably we can add it as an
2391 * additional case:
2392 *
2393 * ```
2394 * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN)
2395 * # include <stddef.h>
2396 * # ifdef unreachable
2397 * # define XXH_UNREACHABLE() unreachable()
2398 * # endif
2399 * #endif
2400 * ```
2401 *
2402 * Note C++23 also has std::unreachable() which can be detected
2403 * as follows:
2404 * ```
2405 * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L)
2406 * # include <utility>
2407 * # define XXH_UNREACHABLE() std::unreachable()
2408 * #endif
2409 * ```
2410 * NB: `__cpp_lib_unreachable` is defined in the `<version>` header.
2411 * We don't use that as including `<utility>` in `extern "C"` blocks
2412 * doesn't work on GCC12
2413 */
2414
2415 #if XXH_HAS_BUILTIN(__builtin_unreachable)
2416 # define XXH_UNREACHABLE() __builtin_unreachable()
2417
2418 #elif defined(_MSC_VER)
2419 # define XXH_UNREACHABLE() __assume(0)
2420
2421 #else
2422 # define XXH_UNREACHABLE()
2423 #endif
2424
2425 #if XXH_HAS_BUILTIN(__builtin_assume)
2426 # define XXH_ASSUME(c) __builtin_assume(c)
2427 #else
2428 # define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
2429 #endif
2430
2431 /*!
2432 * @internal
2433 * @def XXH_rotl32(x,r)
2434 * @brief 32-bit rotate left.
2435 *
2436 * @param x The 32-bit integer to be rotated.
2437 * @param r The number of bits to rotate.
2438 * @pre
2439 * @p r > 0 && @p r < 32
2440 * @note
2441 * @p x and @p r may be evaluated multiple times.
2442 * @return The rotated result.
2443 */
2444 #if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \
2445 && XXH_HAS_BUILTIN(__builtin_rotateleft64)
2446 # define XXH_rotl32 __builtin_rotateleft32
2447 # define XXH_rotl64 __builtin_rotateleft64
2448 /* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
2449 #elif defined(_MSC_VER)
2450 # define XXH_rotl32(x,r) _rotl(x,r)
2451 # define XXH_rotl64(x,r) _rotl64(x,r)
2452 #else
2453 # define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
2454 # define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
2455 #endif
2456
2457 /*!
2458 * @internal
2459 * @fn xxh_u32 XXH_swap32(xxh_u32 x)
2460 * @brief A 32-bit byteswap.
2461 *
2462 * @param x The 32-bit integer to byteswap.
2463 * @return @p x, byteswapped.
2464 */
2465 #if defined(_MSC_VER) /* Visual Studio */
2466 # define XXH_swap32 _byteswap_ulong
2467 #elif XXH_GCC_VERSION >= 403
2468 # define XXH_swap32 __builtin_bswap32
2469 #else
2470 static xxh_u32 XXH_swap32 (xxh_u32 x)
2471 {
2472 return ((x << 24) & 0xff000000 ) |
2473 ((x << 8) & 0x00ff0000 ) |
2474 ((x >> 8) & 0x0000ff00 ) |
2475 ((x >> 24) & 0x000000ff );
2476 }
2477 #endif
2478
2479
2480 /* ***************************
2481 * Memory reads
2482 *****************************/
2483
2484 /*!
2485 * @internal
2486 * @brief Enum to indicate whether a pointer is aligned.
2487 */
2488 typedef enum {
2489 XXH_aligned, /*!< Aligned */
2490 XXH_unaligned /*!< Possibly unaligned */
2491 } XXH_alignment;
2492
2493 /*
2494 * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
2495 *
2496 * This is ideal for older compilers which don't inline memcpy.
2497 */
2498 #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
2499
2500 XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
2501 {
2502 const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
2503 return bytePtr[0]
2504 | ((xxh_u32)bytePtr[1] << 8)
2505 | ((xxh_u32)bytePtr[2] << 16)
2506 | ((xxh_u32)bytePtr[3] << 24);
2507 }
2508
2509 XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
2510 {
2511 const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
2512 return bytePtr[3]
2513 | ((xxh_u32)bytePtr[2] << 8)
2514 | ((xxh_u32)bytePtr[1] << 16)
2515 | ((xxh_u32)bytePtr[0] << 24);
2516 }
2517
2518 #else
2519 XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
2520 {
2521 return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
2522 }
2523
2524 static xxh_u32 XXH_readBE32(const void* ptr)
2525 {
2526 return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
2527 }
2528 #endif
2529
2530 XXH_FORCE_INLINE xxh_u32
2531 XXH_readLE32_align(const void* ptr, XXH_alignment align)
2532 {
2533 if (align==XXH_unaligned) {
2534 return XXH_readLE32(ptr);
2535 } else {
2536 return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
2537 }
2538 }
2539
2540
2541 /* *************************************
2542 * Misc
2543 ***************************************/
2544 /*! @ingroup public */
2545 XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
2546
2547
2548 /* *******************************************************************
2549 * 32-bit hash functions
2550 *********************************************************************/
2551 /*!
2552 * @}
2553 * @defgroup XXH32_impl XXH32 implementation
2554 * @ingroup impl
2555 *
2556 * Details on the XXH32 implementation.
2557 * @{
2558 */
2559 /* #define instead of static const, to be used as initializers */
2560 #define XXH_PRIME32_1 0x9E3779B1U /*!< 0b10011110001101110111100110110001 */
2561 #define XXH_PRIME32_2 0x85EBCA77U /*!< 0b10000101111010111100101001110111 */
2562 #define XXH_PRIME32_3 0xC2B2AE3DU /*!< 0b11000010101100101010111000111101 */
2563 #define XXH_PRIME32_4 0x27D4EB2FU /*!< 0b00100111110101001110101100101111 */
2564 #define XXH_PRIME32_5 0x165667B1U /*!< 0b00010110010101100110011110110001 */
2565
2566 #ifdef XXH_OLD_NAMES
2567 # define PRIME32_1 XXH_PRIME32_1
2568 # define PRIME32_2 XXH_PRIME32_2
2569 # define PRIME32_3 XXH_PRIME32_3
2570 # define PRIME32_4 XXH_PRIME32_4
2571 # define PRIME32_5 XXH_PRIME32_5
2572 #endif
2573
2574 /*!
2575 * @internal
2576 * @brief Normal stripe processing routine.
2577 *
2578 * This shuffles the bits so that any bit from @p input impacts several bits in
2579 * @p acc.
2580 *
2581 * @param acc The accumulator lane.
2582 * @param input The stripe of input to mix.
2583 * @return The mixed accumulator lane.
2584 */
2585 static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
2586 {
2587 acc += input * XXH_PRIME32_2;
2588 acc = XXH_rotl32(acc, 13);
2589 acc *= XXH_PRIME32_1;
2590 #if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
2591 /*
2592 * UGLY HACK:
2593 * A compiler fence is the only thing that prevents GCC and Clang from
2594 * autovectorizing the XXH32 loop (pragmas and attributes don't work for some
2595 * reason) without globally disabling SSE4.1.
2596 *
2597 * The reason we want to avoid vectorization is because despite working on
2598 * 4 integers at a time, there are multiple factors slowing XXH32 down on
2599 * SSE4:
2600 * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
2601 * newer chips!) making it slightly slower to multiply four integers at
2602 * once compared to four integers independently. Even when pmulld was
2603 * fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
2604 * just to multiply unless doing a long operation.
2605 *
2606 * - Four instructions are required to rotate,
2607 * movqda tmp, v // not required with VEX encoding
2608 * pslld tmp, 13 // tmp <<= 13
2609 * psrld v, 19 // x >>= 19
2610 * por v, tmp // x |= tmp
2611 * compared to one for scalar:
2612 * roll v, 13 // reliably fast across the board
2613 * shldl v, v, 13 // Sandy Bridge and later prefer this for some reason
2614 *
2615 * - Instruction level parallelism is actually more beneficial here because
2616 * the SIMD actually serializes this operation: While v1 is rotating, v2
2617 * can load data, while v3 can multiply. SSE forces them to operate
2618 * together.
2619 *
2620 * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
2621 * the loop. NEON is only faster on the A53, and with the newer cores, it is less
2622 * than half the speed.
2623 *
2624 * Additionally, this is used on WASM SIMD128 because it JITs to the same
2625 * SIMD instructions and has the same issue.
2626 */
2627 XXH_COMPILER_GUARD(acc);
2628 #endif
2629 return acc;
2630 }
2631
2632 /*!
2633 * @internal
2634 * @brief Mixes all bits to finalize the hash.
2635 *
2636 * The final mix ensures that all input bits have a chance to impact any bit in
2637 * the output digest, resulting in an unbiased distribution.
2638 *
2639 * @param hash The hash to avalanche.
2640 * @return The avalanched hash.
2641 */
2642 static xxh_u32 XXH32_avalanche(xxh_u32 hash)
2643 {
2644 hash ^= hash >> 15;
2645 hash *= XXH_PRIME32_2;
2646 hash ^= hash >> 13;
2647 hash *= XXH_PRIME32_3;
2648 hash ^= hash >> 16;
2649 return hash;
2650 }
2651
2652 #define XXH_get32bits(p) XXH_readLE32_align(p, align)
2653
2654 /*!
2655 * @internal
2656 * @brief Processes the last 0-15 bytes of @p ptr.
2657 *
2658 * There may be up to 15 bytes remaining to consume from the input.
2659 * This final stage will digest them to ensure that all input bytes are present
2660 * in the final mix.
2661 *
2662 * @param hash The hash to finalize.
2663 * @param ptr The pointer to the remaining input.
2664 * @param len The remaining length, modulo 16.
2665 * @param align Whether @p ptr is aligned.
2666 * @return The finalized hash.
2667 * @see XXH64_finalize().
2668 */
2669 static XXH_PUREF xxh_u32
2670 XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
2671 {
2672 #define XXH_PROCESS1 do { \
2673 hash += (*ptr++) * XXH_PRIME32_5; \
2674 hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1; \
2675 } while (0)
2676
2677 #define XXH_PROCESS4 do { \
2678 hash += XXH_get32bits(ptr) * XXH_PRIME32_3; \
2679 ptr += 4; \
2680 hash = XXH_rotl32(hash, 17) * XXH_PRIME32_4; \
2681 } while (0)
2682
2683 if (ptr==NULL) XXH_ASSERT(len == 0);
2684
2685 /* Compact rerolled version; generally faster */
2686 if (!XXH32_ENDJMP) {
2687 len &= 15;
2688 while (len >= 4) {
2689 XXH_PROCESS4;
2690 len -= 4;
2691 }
2692 while (len > 0) {
2693 XXH_PROCESS1;
2694 --len;
2695 }
2696 return XXH32_avalanche(hash);
2697 } else {
2698 switch(len&15) /* or switch(bEnd - p) */ {
2699 case 12: XXH_PROCESS4;
2700 XXH_FALLTHROUGH; /* fallthrough */
2701 case 8: XXH_PROCESS4;
2702 XXH_FALLTHROUGH; /* fallthrough */
2703 case 4: XXH_PROCESS4;
2704 return XXH32_avalanche(hash);
2705
2706 case 13: XXH_PROCESS4;
2707 XXH_FALLTHROUGH; /* fallthrough */
2708 case 9: XXH_PROCESS4;
2709 XXH_FALLTHROUGH; /* fallthrough */
2710 case 5: XXH_PROCESS4;
2711 XXH_PROCESS1;
2712 return XXH32_avalanche(hash);
2713
2714 case 14: XXH_PROCESS4;
2715 XXH_FALLTHROUGH; /* fallthrough */
2716 case 10: XXH_PROCESS4;
2717 XXH_FALLTHROUGH; /* fallthrough */
2718 case 6: XXH_PROCESS4;
2719 XXH_PROCESS1;
2720 XXH_PROCESS1;
2721 return XXH32_avalanche(hash);
2722
2723 case 15: XXH_PROCESS4;
2724 XXH_FALLTHROUGH; /* fallthrough */
2725 case 11: XXH_PROCESS4;
2726 XXH_FALLTHROUGH; /* fallthrough */
2727 case 7: XXH_PROCESS4;
2728 XXH_FALLTHROUGH; /* fallthrough */
2729 case 3: XXH_PROCESS1;
2730 XXH_FALLTHROUGH; /* fallthrough */
2731 case 2: XXH_PROCESS1;
2732 XXH_FALLTHROUGH; /* fallthrough */
2733 case 1: XXH_PROCESS1;
2734 XXH_FALLTHROUGH; /* fallthrough */
2735 case 0: return XXH32_avalanche(hash);
2736 }
2737 XXH_ASSERT(0);
2738 return hash; /* reaching this point is deemed impossible */
2739 }
2740 }
2741
2742 #ifdef XXH_OLD_NAMES
2743 # define PROCESS1 XXH_PROCESS1
2744 # define PROCESS4 XXH_PROCESS4
2745 #else
2746 # undef XXH_PROCESS1
2747 # undef XXH_PROCESS4
2748 #endif
2749
2750 /*!
2751 * @internal
2752 * @brief The implementation for @ref XXH32().
2753 *
2754 * @param input , len , seed Directly passed from @ref XXH32().
2755 * @param align Whether @p input is aligned.
2756 * @return The calculated hash.
2757 */
2758 XXH_FORCE_INLINE XXH_PUREF xxh_u32
2759 XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
2760 {
2761 xxh_u32 h32;
2762
2763 if (input==NULL) XXH_ASSERT(len == 0);
2764
2765 if (len>=16) {
2766 const xxh_u8* const bEnd = input + len;
2767 const xxh_u8* const limit = bEnd - 15;
2768 xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
2769 xxh_u32 v2 = seed + XXH_PRIME32_2;
2770 xxh_u32 v3 = seed + 0;
2771 xxh_u32 v4 = seed - XXH_PRIME32_1;
2772
2773 do {
2774 v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
2775 v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
2776 v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
2777 v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
2778 } while (input < limit);
2779
2780 h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7)
2781 + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
2782 } else {
2783 h32 = seed + XXH_PRIME32_5;
2784 }
2785
2786 h32 += (xxh_u32)len;
2787
2788 return XXH32_finalize(h32, input, len&15, align);
2789 }
2790
2791 /*! @ingroup XXH32_family */
2792 XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
2793 {
2794 #if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
2795 /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
2796 XXH32_state_t state;
2797 XXH32_reset(&state, seed);
2798 XXH32_update(&state, (const xxh_u8*)input, len);
2799 return XXH32_digest(&state);
2800 #else
2801 if (XXH_FORCE_ALIGN_CHECK) {
2802 if ((((size_t)input) & 3) == 0) { /* Input is 4-bytes aligned, leverage the speed benefit */
2803 return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
2804 } }
2805
2806 return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
2807 #endif
2808 }
2809
2810
2811
2812 /******* Hash streaming *******/
2813 #ifndef XXH_NO_STREAM
2814 /*! @ingroup XXH32_family */
2815 XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
2816 {
2817 return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
2818 }
2819 /*! @ingroup XXH32_family */
2820 XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
2821 {
2822 XXH_free(statePtr);
2823 return XXH_OK;
2824 }
2825
2826 /*! @ingroup XXH32_family */
2827 XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
2828 {
2829 XXH_memcpy(dstState, srcState, sizeof(*dstState));
2830 }
2831
2832 /*! @ingroup XXH32_family */
2833 XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
2834 {
2835 XXH_ASSERT(statePtr != NULL);
2836 memset(statePtr, 0, sizeof(*statePtr));
2837 statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
2838 statePtr->v[1] = seed + XXH_PRIME32_2;
2839 statePtr->v[2] = seed + 0;
2840 statePtr->v[3] = seed - XXH_PRIME32_1;
2841 return XXH_OK;
2842 }
2843
2844
2845 /*! @ingroup XXH32_family */
2846 XXH_PUBLIC_API XXH_errorcode
2847 XXH32_update(XXH32_state_t* state, const void* input, size_t len)
2848 {
2849 if (input==NULL) {
2850 XXH_ASSERT(len == 0);
2851 return XXH_OK;
2852 }
2853
2854 { const xxh_u8* p = (const xxh_u8*)input;
2855 const xxh_u8* const bEnd = p + len;
2856
2857 state->total_len_32 += (XXH32_hash_t)len;
2858 state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
2859
2860 if (state->memsize + len < 16) { /* fill in tmp buffer */
2861 XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
2862 state->memsize += (XXH32_hash_t)len;
2863 return XXH_OK;
2864 }
2865
2866 if (state->memsize) { /* some data left from previous update */
2867 XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
2868 { const xxh_u32* p32 = state->mem32;
2869 state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++;
2870 state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++;
2871 state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++;
2872 state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32));
2873 }
2874 p += 16-state->memsize;
2875 state->memsize = 0;
2876 }
2877
2878 if (p <= bEnd-16) {
2879 const xxh_u8* const limit = bEnd - 16;
2880
2881 do {
2882 state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4;
2883 state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4;
2884 state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4;
2885 state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4;
2886 } while (p<=limit);
2887
2888 }
2889
2890 if (p < bEnd) {
2891 XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
2892 state->memsize = (unsigned)(bEnd-p);
2893 }
2894 }
2895
2896 return XXH_OK;
2897 }
2898
2899
2900 /*! @ingroup XXH32_family */
2901 XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
2902 {
2903 xxh_u32 h32;
2904
2905 if (state->large_len) {
2906 h32 = XXH_rotl32(state->v[0], 1)
2907 + XXH_rotl32(state->v[1], 7)
2908 + XXH_rotl32(state->v[2], 12)
2909 + XXH_rotl32(state->v[3], 18);
2910 } else {
2911 h32 = state->v[2] /* == seed */ + XXH_PRIME32_5;
2912 }
2913
2914 h32 += state->total_len_32;
2915
2916 return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
2917 }
2918 #endif /* !XXH_NO_STREAM */
2919
2920 /******* Canonical representation *******/
2921
2922 /*!
2923 * @ingroup XXH32_family
2924 * The default return values from XXH functions are unsigned 32 and 64 bit
2925 * integers.
2926 *
2927 * The canonical representation uses big endian convention, the same convention
2928 * as human-readable numbers (large digits first).
2929 *
2930 * This way, hash values can be written into a file or buffer, remaining
2931 * comparable across different systems.
2932 *
2933 * The following functions allow transformation of hash values to and from their
2934 * canonical format.
2935 */
2936 XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
2937 {
2938 XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
2939 if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
2940 XXH_memcpy(dst, &hash, sizeof(*dst));
2941 }
2942 /*! @ingroup XXH32_family */
2943 XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
2944 {
2945 return XXH_readBE32(src);
2946 }
2947
2948
2949 #ifndef XXH_NO_LONG_LONG
2950
2951 /* *******************************************************************
2952 * 64-bit hash functions
2953 *********************************************************************/
2954 /*!
2955 * @}
2956 * @ingroup impl
2957 * @{
2958 */
2959 /******* Memory access *******/
2960
2961 typedef XXH64_hash_t xxh_u64;
2962
2963 #ifdef XXH_OLD_NAMES
2964 # define U64 xxh_u64
2965 #endif
2966
2967 #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
2968 /*
2969 * Manual byteshift. Best for old compilers which don't inline memcpy.
2970 * We actually directly use XXH_readLE64 and XXH_readBE64.
2971 */
2972 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
2973
2974 /* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
2975 static xxh_u64 XXH_read64(const void* memPtr)
2976 {
2977 return *(const xxh_u64*) memPtr;
2978 }
2979
2980 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
2981
2982 /*
2983 * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
2984 * documentation claimed that it only increased the alignment, but actually it
2985 * can decrease it on gcc, clang, and icc:
2986 * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
2987 * https://gcc.godbolt.org/z/xYez1j67Y.
2988 */
2989 #ifdef XXH_OLD_NAMES
2990 typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
2991 #endif
2992 static xxh_u64 XXH_read64(const void* ptr)
2993 {
2994 typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64;
2995 return *((const xxh_unalign64*)ptr);
2996 }
2997
2998 #else
2999
3000 /*
3001 * Portable and safe solution. Generally efficient.
3002 * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
3003 */
3004 static xxh_u64 XXH_read64(const void* memPtr)
3005 {
3006 xxh_u64 val;
3007 XXH_memcpy(&val, memPtr, sizeof(val));
3008 return val;
3009 }
3010
3011 #endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
3012
3013 #if defined(_MSC_VER) /* Visual Studio */
3014 # define XXH_swap64 _byteswap_uint64
3015 #elif XXH_GCC_VERSION >= 403
3016 # define XXH_swap64 __builtin_bswap64
3017 #else
3018 static xxh_u64 XXH_swap64(xxh_u64 x)
3019 {
3020 return ((x << 56) & 0xff00000000000000ULL) |
3021 ((x << 40) & 0x00ff000000000000ULL) |
3022 ((x << 24) & 0x0000ff0000000000ULL) |
3023 ((x << 8) & 0x000000ff00000000ULL) |
3024 ((x >> 8) & 0x00000000ff000000ULL) |
3025 ((x >> 24) & 0x0000000000ff0000ULL) |
3026 ((x >> 40) & 0x000000000000ff00ULL) |
3027 ((x >> 56) & 0x00000000000000ffULL);
3028 }
3029 #endif
3030
3031
3032 /* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
3033 #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
3034
3035 XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
3036 {
3037 const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
3038 return bytePtr[0]
3039 | ((xxh_u64)bytePtr[1] << 8)
3040 | ((xxh_u64)bytePtr[2] << 16)
3041 | ((xxh_u64)bytePtr[3] << 24)
3042 | ((xxh_u64)bytePtr[4] << 32)
3043 | ((xxh_u64)bytePtr[5] << 40)
3044 | ((xxh_u64)bytePtr[6] << 48)
3045 | ((xxh_u64)bytePtr[7] << 56);
3046 }
3047
3048 XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
3049 {
3050 const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
3051 return bytePtr[7]
3052 | ((xxh_u64)bytePtr[6] << 8)
3053 | ((xxh_u64)bytePtr[5] << 16)
3054 | ((xxh_u64)bytePtr[4] << 24)
3055 | ((xxh_u64)bytePtr[3] << 32)
3056 | ((xxh_u64)bytePtr[2] << 40)
3057 | ((xxh_u64)bytePtr[1] << 48)
3058 | ((xxh_u64)bytePtr[0] << 56);
3059 }
3060
3061 #else
3062 XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
3063 {
3064 return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
3065 }
3066
3067 static xxh_u64 XXH_readBE64(const void* ptr)
3068 {
3069 return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
3070 }
3071 #endif
3072
3073 XXH_FORCE_INLINE xxh_u64
3074 XXH_readLE64_align(const void* ptr, XXH_alignment align)
3075 {
3076 if (align==XXH_unaligned)
3077 return XXH_readLE64(ptr);
3078 else
3079 return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
3080 }
3081
3082
3083 /******* xxh64 *******/
3084 /*!
3085 * @}
3086 * @defgroup XXH64_impl XXH64 implementation
3087 * @ingroup impl
3088 *
3089 * Details on the XXH64 implementation.
3090 * @{
3091 */
3092 /* #define rather that static const, to be used as initializers */
3093 #define XXH_PRIME64_1 0x9E3779B185EBCA87ULL /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */
3094 #define XXH_PRIME64_2 0xC2B2AE3D27D4EB4FULL /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */
3095 #define XXH_PRIME64_3 0x165667B19E3779F9ULL /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */
3096 #define XXH_PRIME64_4 0x85EBCA77C2B2AE63ULL /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */
3097 #define XXH_PRIME64_5 0x27D4EB2F165667C5ULL /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */
3098
3099 #ifdef XXH_OLD_NAMES
3100 # define PRIME64_1 XXH_PRIME64_1
3101 # define PRIME64_2 XXH_PRIME64_2
3102 # define PRIME64_3 XXH_PRIME64_3
3103 # define PRIME64_4 XXH_PRIME64_4
3104 # define PRIME64_5 XXH_PRIME64_5
3105 #endif
3106
3107 /*! @copydoc XXH32_round */
3108 static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
3109 {
3110 acc += input * XXH_PRIME64_2;
3111 acc = XXH_rotl64(acc, 31);
3112 acc *= XXH_PRIME64_1;
3113 return acc;
3114 }
3115
3116 static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
3117 {
3118 val = XXH64_round(0, val);
3119 acc ^= val;
3120 acc = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
3121 return acc;
3122 }
3123
3124 /*! @copydoc XXH32_avalanche */
3125 static xxh_u64 XXH64_avalanche(xxh_u64 hash)
3126 {
3127 hash ^= hash >> 33;
3128 hash *= XXH_PRIME64_2;
3129 hash ^= hash >> 29;
3130 hash *= XXH_PRIME64_3;
3131 hash ^= hash >> 32;
3132 return hash;
3133 }
3134
3135
3136 #define XXH_get64bits(p) XXH_readLE64_align(p, align)
3137
3138 /*!
3139 * @internal
3140 * @brief Processes the last 0-31 bytes of @p ptr.
3141 *
3142 * There may be up to 31 bytes remaining to consume from the input.
3143 * This final stage will digest them to ensure that all input bytes are present
3144 * in the final mix.
3145 *
3146 * @param hash The hash to finalize.
3147 * @param ptr The pointer to the remaining input.
3148 * @param len The remaining length, modulo 32.
3149 * @param align Whether @p ptr is aligned.
3150 * @return The finalized hash
3151 * @see XXH32_finalize().
3152 */
3153 static XXH_PUREF xxh_u64
3154 XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
3155 {
3156 if (ptr==NULL) XXH_ASSERT(len == 0);
3157 len &= 31;
3158 while (len >= 8) {
3159 xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
3160 ptr += 8;
3161 hash ^= k1;
3162 hash = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
3163 len -= 8;
3164 }
3165 if (len >= 4) {
3166 hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
3167 ptr += 4;
3168 hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
3169 len -= 4;
3170 }
3171 while (len > 0) {
3172 hash ^= (*ptr++) * XXH_PRIME64_5;
3173 hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
3174 --len;
3175 }
3176 return XXH64_avalanche(hash);
3177 }
3178
3179 #ifdef XXH_OLD_NAMES
3180 # define PROCESS1_64 XXH_PROCESS1_64
3181 # define PROCESS4_64 XXH_PROCESS4_64
3182 # define PROCESS8_64 XXH_PROCESS8_64
3183 #else
3184 # undef XXH_PROCESS1_64
3185 # undef XXH_PROCESS4_64
3186 # undef XXH_PROCESS8_64
3187 #endif
3188
3189 /*!
3190 * @internal
3191 * @brief The implementation for @ref XXH64().
3192 *
3193 * @param input , len , seed Directly passed from @ref XXH64().
3194 * @param align Whether @p input is aligned.
3195 * @return The calculated hash.
3196 */
3197 XXH_FORCE_INLINE XXH_PUREF xxh_u64
3198 XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
3199 {
3200 xxh_u64 h64;
3201 if (input==NULL) XXH_ASSERT(len == 0);
3202
3203 if (len>=32) {
3204 const xxh_u8* const bEnd = input + len;
3205 const xxh_u8* const limit = bEnd - 31;
3206 xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
3207 xxh_u64 v2 = seed + XXH_PRIME64_2;
3208 xxh_u64 v3 = seed + 0;
3209 xxh_u64 v4 = seed - XXH_PRIME64_1;
3210
3211 do {
3212 v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
3213 v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
3214 v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
3215 v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
3216 } while (input<limit);
3217
3218 h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
3219 h64 = XXH64_mergeRound(h64, v1);
3220 h64 = XXH64_mergeRound(h64, v2);
3221 h64 = XXH64_mergeRound(h64, v3);
3222 h64 = XXH64_mergeRound(h64, v4);
3223
3224 } else {
3225 h64 = seed + XXH_PRIME64_5;
3226 }
3227
3228 h64 += (xxh_u64) len;
3229
3230 return XXH64_finalize(h64, input, len, align);
3231 }
3232
3233
3234 /*! @ingroup XXH64_family */
3235 XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
3236 {
3237 #if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
3238 /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
3239 XXH64_state_t state;
3240 XXH64_reset(&state, seed);
3241 XXH64_update(&state, (const xxh_u8*)input, len);
3242 return XXH64_digest(&state);
3243 #else
3244 if (XXH_FORCE_ALIGN_CHECK) {
3245 if ((((size_t)input) & 7)==0) { /* Input is aligned, let's leverage the speed advantage */
3246 return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
3247 } }
3248
3249 return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
3250
3251 #endif
3252 }
3253
3254 /******* Hash Streaming *******/
3255 #ifndef XXH_NO_STREAM
3256 /*! @ingroup XXH64_family*/
3257 XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
3258 {
3259 return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
3260 }
3261 /*! @ingroup XXH64_family */
3262 XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
3263 {
3264 XXH_free(statePtr);
3265 return XXH_OK;
3266 }
3267
3268 /*! @ingroup XXH64_family */
3269 XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
3270 {
3271 XXH_memcpy(dstState, srcState, sizeof(*dstState));
3272 }
3273
3274 /*! @ingroup XXH64_family */
3275 XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
3276 {
3277 XXH_ASSERT(statePtr != NULL);
3278 memset(statePtr, 0, sizeof(*statePtr));
3279 statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
3280 statePtr->v[1] = seed + XXH_PRIME64_2;
3281 statePtr->v[2] = seed + 0;
3282 statePtr->v[3] = seed - XXH_PRIME64_1;
3283 return XXH_OK;
3284 }
3285
3286 /*! @ingroup XXH64_family */
3287 XXH_PUBLIC_API XXH_errorcode
3288 XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
3289 {
3290 if (input==NULL) {
3291 XXH_ASSERT(len == 0);
3292 return XXH_OK;
3293 }
3294
3295 { const xxh_u8* p = (const xxh_u8*)input;
3296 const xxh_u8* const bEnd = p + len;
3297
3298 state->total_len += len;
3299
3300 if (state->memsize + len < 32) { /* fill in tmp buffer */
3301 XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
3302 state->memsize += (xxh_u32)len;
3303 return XXH_OK;
3304 }
3305
3306 if (state->memsize) { /* tmp buffer is full */
3307 XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
3308 state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0));
3309 state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1));
3310 state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2));
3311 state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3));
3312 p += 32 - state->memsize;
3313 state->memsize = 0;
3314 }
3315
3316 if (p+32 <= bEnd) {
3317 const xxh_u8* const limit = bEnd - 32;
3318
3319 do {
3320 state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8;
3321 state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8;
3322 state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8;
3323 state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8;
3324 } while (p<=limit);
3325
3326 }
3327
3328 if (p < bEnd) {
3329 XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
3330 state->memsize = (unsigned)(bEnd-p);
3331 }
3332 }
3333
3334 return XXH_OK;
3335 }
3336
3337
3338 /*! @ingroup XXH64_family */
3339 XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
3340 {
3341 xxh_u64 h64;
3342
3343 if (state->total_len >= 32) {
3344 h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18);
3345 h64 = XXH64_mergeRound(h64, state->v[0]);
3346 h64 = XXH64_mergeRound(h64, state->v[1]);
3347 h64 = XXH64_mergeRound(h64, state->v[2]);
3348 h64 = XXH64_mergeRound(h64, state->v[3]);
3349 } else {
3350 h64 = state->v[2] /*seed*/ + XXH_PRIME64_5;
3351 }
3352
3353 h64 += (xxh_u64) state->total_len;
3354
3355 return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
3356 }
3357 #endif /* !XXH_NO_STREAM */
3358
3359 /******* Canonical representation *******/
3360
3361 /*! @ingroup XXH64_family */
3362 XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash)
3363 {
3364 XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
3365 if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
3366 XXH_memcpy(dst, &hash, sizeof(*dst));
3367 }
3368
3369 /*! @ingroup XXH64_family */
3370 XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
3371 {
3372 return XXH_readBE64(src);
3373 }
3374
3375 #ifndef XXH_NO_XXH3
3376
3377 /* *********************************************************************
3378 * XXH3
3379 * New generation hash designed for speed on small keys and vectorization
3380 ************************************************************************ */
3381 /*!
3382 * @}
3383 * @defgroup XXH3_impl XXH3 implementation
3384 * @ingroup impl
3385 * @{
3386 */
3387
3388 /* === Compiler specifics === */
3389
3390 #if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
3391 # define XXH_RESTRICT /* disable */
3392 #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */
3393 # define XXH_RESTRICT restrict
3394 #elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \
3395 || (defined (__clang__)) \
3396 || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \
3397 || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300))
3398 /*
3399 * There are a LOT more compilers that recognize __restrict but this
3400 * covers the major ones.
3401 */
3402 # define XXH_RESTRICT __restrict
3403 #else
3404 # define XXH_RESTRICT /* disable */
3405 #endif
3406
3407 #if (defined(__GNUC__) && (__GNUC__ >= 3)) \
3408 || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
3409 || defined(__clang__)
3410 # define XXH_likely(x) __builtin_expect(x, 1)
3411 # define XXH_unlikely(x) __builtin_expect(x, 0)
3412 #else
3413 # define XXH_likely(x) (x)
3414 # define XXH_unlikely(x) (x)
3415 #endif
3416
3417 #ifndef XXH_HAS_INCLUDE
3418 # ifdef __has_include
3419 # define XXH_HAS_INCLUDE(x) __has_include(x)
3420 # else
3421 # define XXH_HAS_INCLUDE(x) 0
3422 # endif
3423 #endif
3424
3425 #if defined(__GNUC__) || defined(__clang__)
3426 # if defined(__ARM_FEATURE_SVE)
3427 # include <arm_sve.h>
3428 # endif
3429 # if defined(__ARM_NEON__) || defined(__ARM_NEON) \
3430 || (defined(_M_ARM) && _M_ARM >= 7) \
3431 || defined(_M_ARM64) || defined(_M_ARM64EC) \
3432 || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* WASM SIMD128 via SIMDe */
3433 # define inline __inline__ /* circumvent a clang bug */
3434 # include <arm_neon.h>
3435 # undef inline
3436 # elif defined(__AVX2__)
3437 # include <immintrin.h>
3438 # elif defined(__SSE2__)
3439 # include <emmintrin.h>
3440 # endif
3441 #endif
3442
3443 #if defined(_MSC_VER)
3444 # include <intrin.h>
3445 #endif
3446
3447 /*
3448 * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
3449 * remaining a true 64-bit/128-bit hash function.
3450 *
3451 * This is done by prioritizing a subset of 64-bit operations that can be
3452 * emulated without too many steps on the average 32-bit machine.
3453 *
3454 * For example, these two lines seem similar, and run equally fast on 64-bit:
3455 *
3456 * xxh_u64 x;
3457 * x ^= (x >> 47); // good
3458 * x ^= (x >> 13); // bad
3459 *
3460 * However, to a 32-bit machine, there is a major difference.
3461 *
3462 * x ^= (x >> 47) looks like this:
3463 *
3464 * x.lo ^= (x.hi >> (47 - 32));
3465 *
3466 * while x ^= (x >> 13) looks like this:
3467 *
3468 * // note: funnel shifts are not usually cheap.
3469 * x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
3470 * x.hi ^= (x.hi >> 13);
3471 *
3472 * The first one is significantly faster than the second, simply because the
3473 * shift is larger than 32. This means:
3474 * - All the bits we need are in the upper 32 bits, so we can ignore the lower
3475 * 32 bits in the shift.
3476 * - The shift result will always fit in the lower 32 bits, and therefore,
3477 * we can ignore the upper 32 bits in the xor.
3478 *
3479 * Thanks to this optimization, XXH3 only requires these features to be efficient:
3480 *
3481 * - Usable unaligned access
3482 * - A 32-bit or 64-bit ALU
3483 * - If 32-bit, a decent ADC instruction
3484 * - A 32 or 64-bit multiply with a 64-bit result
3485 * - For the 128-bit variant, a decent byteswap helps short inputs.
3486 *
3487 * The first two are already required by XXH32, and almost all 32-bit and 64-bit
3488 * platforms which can run XXH32 can run XXH3 efficiently.
3489 *
3490 * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
3491 * notable exception.
3492 *
3493 * First of all, Thumb-1 lacks support for the UMULL instruction which
3494 * performs the important long multiply. This means numerous __aeabi_lmul
3495 * calls.
3496 *
3497 * Second of all, the 8 functional registers are just not enough.
3498 * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
3499 * Lo registers, and this shuffling results in thousands more MOVs than A32.
3500 *
3501 * A32 and T32 don't have this limitation. They can access all 14 registers,
3502 * do a 32->64 multiply with UMULL, and the flexible operand allowing free
3503 * shifts is helpful, too.
3504 *
3505 * Therefore, we do a quick sanity check.
3506 *
3507 * If compiling Thumb-1 for a target which supports ARM instructions, we will
3508 * emit a warning, as it is not a "sane" platform to compile for.
3509 *
3510 * Usually, if this happens, it is because of an accident and you probably need
3511 * to specify -march, as you likely meant to compile for a newer architecture.
3512 *
3513 * Credit: large sections of the vectorial and asm source code paths
3514 * have been contributed by @easyaspi314
3515 */
3516 #if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
3517 # warning "XXH3 is highly inefficient without ARM or Thumb-2."
3518 #endif
3519
3520 /* ==========================================
3521 * Vectorization detection
3522 * ========================================== */
3523
3524 #ifdef XXH_DOXYGEN
3525 /*!
3526 * @ingroup tuning
3527 * @brief Overrides the vectorization implementation chosen for XXH3.
3528 *
3529 * Can be defined to 0 to disable SIMD or any of the values mentioned in
3530 * @ref XXH_VECTOR_TYPE.
3531 *
3532 * If this is not defined, it uses predefined macros to determine the best
3533 * implementation.
3534 */
3535 # define XXH_VECTOR XXH_SCALAR
3536 /*!
3537 * @ingroup tuning
3538 * @brief Possible values for @ref XXH_VECTOR.
3539 *
3540 * Note that these are actually implemented as macros.
3541 *
3542 * If this is not defined, it is detected automatically.
3543 * internal macro XXH_X86DISPATCH overrides this.
3544 */
3545 enum XXH_VECTOR_TYPE /* fake enum */ {
3546 XXH_SCALAR = 0, /*!< Portable scalar version */
3547 XXH_SSE2 = 1, /*!<
3548 * SSE2 for Pentium 4, Opteron, all x86_64.
3549 *
3550 * @note SSE2 is also guaranteed on Windows 10, macOS, and
3551 * Android x86.
3552 */
3553 XXH_AVX2 = 2, /*!< AVX2 for Haswell and Bulldozer */
3554 XXH_AVX512 = 3, /*!< AVX512 for Skylake and Icelake */
3555 XXH_NEON = 4, /*!<
3556 * NEON for most ARMv7-A, all AArch64, and WASM SIMD128
3557 * via the SIMDeverywhere polyfill provided with the
3558 * Emscripten SDK.
3559 */
3560 XXH_VSX = 5, /*!< VSX and ZVector for POWER8/z13 (64-bit) */
3561 XXH_SVE = 6, /*!< SVE for some ARMv8-A and ARMv9-A */
3562 };
3563 /*!
3564 * @ingroup tuning
3565 * @brief Selects the minimum alignment for XXH3's accumulators.
3566 *
3567 * When using SIMD, this should match the alignment required for said vector
3568 * type, so, for example, 32 for AVX2.
3569 *
3570 * Default: Auto detected.
3571 */
3572 # define XXH_ACC_ALIGN 8
3573 #endif
3574
3575 /* Actual definition */
3576 #ifndef XXH_DOXYGEN
3577 # define XXH_SCALAR 0
3578 # define XXH_SSE2 1
3579 # define XXH_AVX2 2
3580 # define XXH_AVX512 3
3581 # define XXH_NEON 4
3582 # define XXH_VSX 5
3583 # define XXH_SVE 6
3584 #endif
3585
3586 #ifndef XXH_VECTOR /* can be defined on command line */
3587 # if defined(__ARM_FEATURE_SVE)
3588 # define XXH_VECTOR XXH_SVE
3589 # elif ( \
3590 defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
3591 || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
3592 || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* wasm simd128 via SIMDe */ \
3593 ) && ( \
3594 defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
3595 || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
3596 )
3597 # define XXH_VECTOR XXH_NEON
3598 # elif defined(__AVX512F__)
3599 # define XXH_VECTOR XXH_AVX512
3600 # elif defined(__AVX2__)
3601 # define XXH_VECTOR XXH_AVX2
3602 # elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
3603 # define XXH_VECTOR XXH_SSE2
3604 # elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
3605 || (defined(__s390x__) && defined(__VEC__)) \
3606 && defined(__GNUC__) /* TODO: IBM XL */
3607 # define XXH_VECTOR XXH_VSX
3608 # else
3609 # define XXH_VECTOR XXH_SCALAR
3610 # endif
3611 #endif
3612
3613 /* __ARM_FEATURE_SVE is only supported by GCC & Clang. */
3614 #if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE)
3615 # ifdef _MSC_VER
3616 # pragma warning(once : 4606)
3617 # else
3618 # warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead."
3619 # endif
3620 # undef XXH_VECTOR
3621 # define XXH_VECTOR XXH_SCALAR
3622 #endif
3623
3624 /*
3625 * Controls the alignment of the accumulator,
3626 * for compatibility with aligned vector loads, which are usually faster.
3627 */
3628 #ifndef XXH_ACC_ALIGN
3629 # if defined(XXH_X86DISPATCH)
3630 # define XXH_ACC_ALIGN 64 /* for compatibility with avx512 */
3631 # elif XXH_VECTOR == XXH_SCALAR /* scalar */
3632 # define XXH_ACC_ALIGN 8
3633 # elif XXH_VECTOR == XXH_SSE2 /* sse2 */
3634 # define XXH_ACC_ALIGN 16
3635 # elif XXH_VECTOR == XXH_AVX2 /* avx2 */
3636 # define XXH_ACC_ALIGN 32
3637 # elif XXH_VECTOR == XXH_NEON /* neon */
3638 # define XXH_ACC_ALIGN 16
3639 # elif XXH_VECTOR == XXH_VSX /* vsx */
3640 # define XXH_ACC_ALIGN 16
3641 # elif XXH_VECTOR == XXH_AVX512 /* avx512 */
3642 # define XXH_ACC_ALIGN 64
3643 # elif XXH_VECTOR == XXH_SVE /* sve */
3644 # define XXH_ACC_ALIGN 64
3645 # endif
3646 #endif
3647
3648 #if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
3649 || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
3650 # define XXH_SEC_ALIGN XXH_ACC_ALIGN
3651 #elif XXH_VECTOR == XXH_SVE
3652 # define XXH_SEC_ALIGN XXH_ACC_ALIGN
3653 #else
3654 # define XXH_SEC_ALIGN 8
3655 #endif
3656
3657 #if defined(__GNUC__) || defined(__clang__)
3658 # define XXH_ALIASING __attribute__((may_alias))
3659 #else
3660 # define XXH_ALIASING /* nothing */
3661 #endif
3662
3663 /*
3664 * UGLY HACK:
3665 * GCC usually generates the best code with -O3 for xxHash.
3666 *
3667 * However, when targeting AVX2, it is overzealous in its unrolling resulting
3668 * in code roughly 3/4 the speed of Clang.
3669 *
3670 * There are other issues, such as GCC splitting _mm256_loadu_si256 into
3671 * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
3672 * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
3673 *
3674 * That is why when compiling the AVX2 version, it is recommended to use either
3675 * -O2 -mavx2 -march=haswell
3676 * or
3677 * -O2 -mavx2 -mno-avx256-split-unaligned-load
3678 * for decent performance, or to use Clang instead.
3679 *
3680 * Fortunately, we can control the first one with a pragma that forces GCC into
3681 * -O2, but the other one we can't control without "failed to inline always
3682 * inline function due to target mismatch" warnings.
3683 */
3684 #if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
3685 && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
3686 && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
3687 # pragma GCC push_options
3688 # pragma GCC optimize("-O2")
3689 #endif
3690
3691 #if XXH_VECTOR == XXH_NEON
3692
3693 /*
3694 * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3
3695 * optimizes out the entire hashLong loop because of the aliasing violation.
3696 *
3697 * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
3698 * so the only option is to mark it as aliasing.
3699 */
3700 typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
3701
3702 /*!
3703 * @internal
3704 * @brief `vld1q_u64` but faster and alignment-safe.
3705 *
3706 * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only
3707 * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86).
3708 *
3709 * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it
3710 * prohibits load-store optimizations. Therefore, a direct dereference is used.
3711 *
3712 * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe
3713 * unaligned load.
3714 */
3715 #if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
3716 XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */
3717 {
3718 return *(xxh_aliasing_uint64x2_t const *)ptr;
3719 }
3720 #else
3721 XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
3722 {
3723 return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
3724 }
3725 #endif
3726
3727 /*!
3728 * @internal
3729 * @brief `vmlal_u32` on low and high halves of a vector.
3730 *
3731 * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with
3732 * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32`
3733 * with `vmlal_u32`.
3734 */
3735 #if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
3736 XXH_FORCE_INLINE uint64x2_t
3737 XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
3738 {
3739 /* Inline assembly is the only way */
3740 __asm__("umlal %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs));
3741 return acc;
3742 }
3743 XXH_FORCE_INLINE uint64x2_t
3744 XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
3745 {
3746 /* This intrinsic works as expected */
3747 return vmlal_high_u32(acc, lhs, rhs);
3748 }
3749 #else
3750 /* Portable intrinsic versions */
3751 XXH_FORCE_INLINE uint64x2_t
3752 XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
3753 {
3754 return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
3755 }
3756 /*! @copydoc XXH_vmlal_low_u32
3757 * Assume the compiler converts this to vmlal_high_u32 on aarch64 */
3758 XXH_FORCE_INLINE uint64x2_t
3759 XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
3760 {
3761 return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
3762 }
3763 #endif
3764
3765 /*!
3766 * @ingroup tuning
3767 * @brief Controls the NEON to scalar ratio for XXH3
3768 *
3769 * This can be set to 2, 4, 6, or 8.
3770 *
3771 * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
3772 *
3773 * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those
3774 * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU
3775 * bandwidth.
3776 *
3777 * This is even more noticeable on the more advanced cores like the Cortex-A76 which
3778 * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
3779 *
3780 * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes
3781 * and 2 scalar lanes, which is chosen by default.
3782 *
3783 * This does not apply to Apple processors or 32-bit processors, which run better with
3784 * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes.
3785 *
3786 * This change benefits CPUs with large micro-op buffers without negatively affecting
3787 * most other CPUs:
3788 *
3789 * | Chipset | Dispatch type | NEON only | 6:2 hybrid | Diff. |
3790 * |:----------------------|:--------------------|----------:|-----------:|------:|
3791 * | Snapdragon 730 (A76) | 2 NEON/8 micro-ops | 8.8 GB/s | 10.1 GB/s | ~16% |
3792 * | Snapdragon 835 (A73) | 2 NEON/3 micro-ops | 5.1 GB/s | 5.3 GB/s | ~5% |
3793 * | Marvell PXA1928 (A53) | In-order dual-issue | 1.9 GB/s | 1.9 GB/s | 0% |
3794 * | Apple M1 | 4 NEON/8 micro-ops | 37.3 GB/s | 36.1 GB/s | ~-3% |
3795 *
3796 * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
3797 *
3798 * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning
3799 * it effectively becomes worse 4.
3800 *
3801 * @see XXH3_accumulate_512_neon()
3802 */
3803 # ifndef XXH3_NEON_LANES
3804 # if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
3805 && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
3806 # define XXH3_NEON_LANES 6
3807 # else
3808 # define XXH3_NEON_LANES XXH_ACC_NB
3809 # endif
3810 # endif
3811 #endif /* XXH_VECTOR == XXH_NEON */
3812
3813 /*
3814 * VSX and Z Vector helpers.
3815 *
3816 * This is very messy, and any pull requests to clean this up are welcome.
3817 *
3818 * There are a lot of problems with supporting VSX and s390x, due to
3819 * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
3820 */
3821 #if XXH_VECTOR == XXH_VSX
3822 /* Annoyingly, these headers _may_ define three macros: `bool`, `vector`,
3823 * and `pixel`. This is a problem for obvious reasons.
3824 *
3825 * These keywords are unnecessary; the spec literally says they are
3826 * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd
3827 * after including the header.
3828 *
3829 * We use pragma push_macro/pop_macro to keep the namespace clean. */
3830 # pragma push_macro("bool")
3831 # pragma push_macro("vector")
3832 # pragma push_macro("pixel")
3833 /* silence potential macro redefined warnings */
3834 # undef bool
3835 # undef vector
3836 # undef pixel
3837
3838 # if defined(__s390x__)
3839 # include <s390intrin.h>
3840 # else
3841 # include <altivec.h>
3842 # endif
3843
3844 /* Restore the original macro values, if applicable. */
3845 # pragma pop_macro("pixel")
3846 # pragma pop_macro("vector")
3847 # pragma pop_macro("bool")
3848
3849 typedef __vector unsigned long long xxh_u64x2;
3850 typedef __vector unsigned char xxh_u8x16;
3851 typedef __vector unsigned xxh_u32x4;
3852
3853 /*
3854 * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue.
3855 */
3856 typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING;
3857
3858 # ifndef XXH_VSX_BE
3859 # if defined(__BIG_ENDIAN__) \
3860 || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
3861 # define XXH_VSX_BE 1
3862 # elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
3863 # warning "-maltivec=be is not recommended. Please use native endianness."
3864 # define XXH_VSX_BE 1
3865 # else
3866 # define XXH_VSX_BE 0
3867 # endif
3868 # endif /* !defined(XXH_VSX_BE) */
3869
3870 # if XXH_VSX_BE
3871 # if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
3872 # define XXH_vec_revb vec_revb
3873 # else
3874 /*!
3875 * A polyfill for POWER9's vec_revb().
3876 */
3877 XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
3878 {
3879 xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
3880 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
3881 return vec_perm(val, val, vByteSwap);
3882 }
3883 # endif
3884 # endif /* XXH_VSX_BE */
3885
3886 /*!
3887 * Performs an unaligned vector load and byte swaps it on big endian.
3888 */
3889 XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
3890 {
3891 xxh_u64x2 ret;
3892 XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2));
3893 # if XXH_VSX_BE
3894 ret = XXH_vec_revb(ret);
3895 # endif
3896 return ret;
3897 }
3898
3899 /*
3900 * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
3901 *
3902 * These intrinsics weren't added until GCC 8, despite existing for a while,
3903 * and they are endian dependent. Also, their meaning swap depending on version.
3904 * */
3905 # if defined(__s390x__)
3906 /* s390x is always big endian, no issue on this platform */
3907 # define XXH_vec_mulo vec_mulo
3908 # define XXH_vec_mule vec_mule
3909 # elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__)
3910 /* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
3911 /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */
3912 # define XXH_vec_mulo __builtin_altivec_vmulouw
3913 # define XXH_vec_mule __builtin_altivec_vmuleuw
3914 # else
3915 /* gcc needs inline assembly */
3916 /* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
3917 XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
3918 {
3919 xxh_u64x2 result;
3920 __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
3921 return result;
3922 }
3923 XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
3924 {
3925 xxh_u64x2 result;
3926 __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
3927 return result;
3928 }
3929 # endif /* XXH_vec_mulo, XXH_vec_mule */
3930 #endif /* XXH_VECTOR == XXH_VSX */
3931
3932 #if XXH_VECTOR == XXH_SVE
3933 #define ACCRND(acc, offset) \
3934 do { \
3935 svuint64_t input_vec = svld1_u64(mask, xinput + offset); \
3936 svuint64_t secret_vec = svld1_u64(mask, xsecret + offset); \
3937 svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec); \
3938 svuint64_t swapped = svtbl_u64(input_vec, kSwap); \
3939 svuint64_t mixed_lo = svextw_u64_x(mask, mixed); \
3940 svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32); \
3941 svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
3942 acc = svadd_u64_x(mask, acc, mul); \
3943 } while (0)
3944 #endif /* XXH_VECTOR == XXH_SVE */
3945
3946 /* prefetch
3947 * can be disabled, by declaring XXH_NO_PREFETCH build macro */
3948 #if defined(XXH_NO_PREFETCH)
3949 # define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */
3950 #else
3951 # if XXH_SIZE_OPT >= 1
3952 # define XXH_PREFETCH(ptr) (void)(ptr)
3953 # elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) /* _mm_prefetch() not defined outside of x86/x64 */
3954 # include <mmintrin.h> /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
3955 # define XXH_PREFETCH(ptr) _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
3956 # elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
3957 # define XXH_PREFETCH(ptr) __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
3958 # else
3959 # define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */
3960 # endif
3961 #endif /* XXH_NO_PREFETCH */
3962
3963
3964 /* ==========================================
3965 * XXH3 default settings
3966 * ========================================== */
3967
3968 #define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */
3969
3970 #if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
3971 # error "default keyset is not large enough"
3972 #endif
3973
3974 /*! Pseudorandom secret taken directly from FARSH. */
3975 XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
3976 0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
3977 0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
3978 0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
3979 0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
3980 0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
3981 0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
3982 0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
3983 0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
3984 0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
3985 0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
3986 0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
3987 0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
3988 };
3989
3990 static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL; /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */
3991 static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL; /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */
3992
3993 #ifdef XXH_OLD_NAMES
3994 # define kSecret XXH3_kSecret
3995 #endif
3996
3997 #ifdef XXH_DOXYGEN
3998 /*!
3999 * @brief Calculates a 32-bit to 64-bit long multiply.
4000 *
4001 * Implemented as a macro.
4002 *
4003 * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't
4004 * need to (but it shouldn't need to anyways, it is about 7 instructions to do
4005 * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we
4006 * use that instead of the normal method.
4007 *
4008 * If you are compiling for platforms like Thumb-1 and don't have a better option,
4009 * you may also want to write your own long multiply routine here.
4010 *
4011 * @param x, y Numbers to be multiplied
4012 * @return 64-bit product of the low 32 bits of @p x and @p y.
4013 */
4014 XXH_FORCE_INLINE xxh_u64
4015 XXH_mult32to64(xxh_u64 x, xxh_u64 y)
4016 {
4017 return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
4018 }
4019 #elif defined(_MSC_VER) && defined(_M_IX86)
4020 # define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
4021 #else
4022 /*
4023 * Downcast + upcast is usually better than masking on older compilers like
4024 * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
4025 *
4026 * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
4027 * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
4028 */
4029 # define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
4030 #endif
4031
4032 /*!
4033 * @brief Calculates a 64->128-bit long multiply.
4034 *
4035 * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
4036 * version.
4037 *
4038 * @param lhs , rhs The 64-bit integers to be multiplied
4039 * @return The 128-bit result represented in an @ref XXH128_hash_t.
4040 */
4041 static XXH128_hash_t
4042 XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
4043 {
4044 /*
4045 * GCC/Clang __uint128_t method.
4046 *
4047 * On most 64-bit targets, GCC and Clang define a __uint128_t type.
4048 * This is usually the best way as it usually uses a native long 64-bit
4049 * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
4050 *
4051 * Usually.
4052 *
4053 * Despite being a 32-bit platform, Clang (and emscripten) define this type
4054 * despite not having the arithmetic for it. This results in a laggy
4055 * compiler builtin call which calculates a full 128-bit multiply.
4056 * In that case it is best to use the portable one.
4057 * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
4058 */
4059 #if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
4060 && defined(__SIZEOF_INT128__) \
4061 || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
4062
4063 __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
4064 XXH128_hash_t r128;
4065 r128.low64 = (xxh_u64)(product);
4066 r128.high64 = (xxh_u64)(product >> 64);
4067 return r128;
4068
4069 /*
4070 * MSVC for x64's _umul128 method.
4071 *
4072 * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
4073 *
4074 * This compiles to single operand MUL on x64.
4075 */
4076 #elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)
4077
4078 #ifndef _MSC_VER
4079 # pragma intrinsic(_umul128)
4080 #endif
4081 xxh_u64 product_high;
4082 xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
4083 XXH128_hash_t r128;
4084 r128.low64 = product_low;
4085 r128.high64 = product_high;
4086 return r128;
4087
4088 /*
4089 * MSVC for ARM64's __umulh method.
4090 *
4091 * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
4092 */
4093 #elif defined(_M_ARM64) || defined(_M_ARM64EC)
4094
4095 #ifndef _MSC_VER
4096 # pragma intrinsic(__umulh)
4097 #endif
4098 XXH128_hash_t r128;
4099 r128.low64 = lhs * rhs;
4100 r128.high64 = __umulh(lhs, rhs);
4101 return r128;
4102
4103 #else
4104 /*
4105 * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
4106 *
4107 * This is a fast and simple grade school multiply, which is shown below
4108 * with base 10 arithmetic instead of base 0x100000000.
4109 *
4110 * 9 3 // D2 lhs = 93
4111 * x 7 5 // D2 rhs = 75
4112 * ----------
4113 * 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
4114 * 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
4115 * 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
4116 * + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
4117 * ---------
4118 * 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
4119 * + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
4120 * ---------
4121 * 6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
4122 *
4123 * The reasons for adding the products like this are:
4124 * 1. It avoids manual carry tracking. Just like how
4125 * (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
4126 * This avoids a lot of complexity.
4127 *
4128 * 2. It hints for, and on Clang, compiles to, the powerful UMAAL
4129 * instruction available in ARM's Digital Signal Processing extension
4130 * in 32-bit ARMv6 and later, which is shown below:
4131 *
4132 * void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
4133 * {
4134 * xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
4135 * *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
4136 * *RdHi = (xxh_u32)(product >> 32);
4137 * }
4138 *
4139 * This instruction was designed for efficient long multiplication, and
4140 * allows this to be calculated in only 4 instructions at speeds
4141 * comparable to some 64-bit ALUs.
4142 *
4143 * 3. It isn't terrible on other platforms. Usually this will be a couple
4144 * of 32-bit ADD/ADCs.
4145 */
4146
4147 /* First calculate all of the cross products. */
4148 xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
4149 xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF);
4150 xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
4151 xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32);
4152
4153 /* Now add the products together. These will never overflow. */
4154 xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
4155 xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi;
4156 xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
4157
4158 XXH128_hash_t r128;
4159 r128.low64 = lower;
4160 r128.high64 = upper;
4161 return r128;
4162 #endif
4163 }
4164
4165 /*!
4166 * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it.
4167 *
4168 * The reason for the separate function is to prevent passing too many structs
4169 * around by value. This will hopefully inline the multiply, but we don't force it.
4170 *
4171 * @param lhs , rhs The 64-bit integers to multiply
4172 * @return The low 64 bits of the product XOR'd by the high 64 bits.
4173 * @see XXH_mult64to128()
4174 */
4175 static xxh_u64
4176 XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
4177 {
4178 XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
4179 return product.low64 ^ product.high64;
4180 }
4181
4182 /*! Seems to produce slightly better code on GCC for some reason. */
4183 XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
4184 {
4185 XXH_ASSERT(0 <= shift && shift < 64);
4186 return v64 ^ (v64 >> shift);
4187 }
4188
4189 /*
4190 * This is a fast avalanche stage,
4191 * suitable when input bits are already partially mixed
4192 */
4193 static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
4194 {
4195 h64 = XXH_xorshift64(h64, 37);
4196 h64 *= PRIME_MX1;
4197 h64 = XXH_xorshift64(h64, 32);
4198 return h64;
4199 }
4200
4201 /*
4202 * This is a stronger avalanche,
4203 * inspired by Pelle Evensen's rrmxmx
4204 * preferable when input has not been previously mixed
4205 */
4206 static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
4207 {
4208 /* this mix is inspired by Pelle Evensen's rrmxmx */
4209 h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
4210 h64 *= PRIME_MX2;
4211 h64 ^= (h64 >> 35) + len ;
4212 h64 *= PRIME_MX2;
4213 return XXH_xorshift64(h64, 28);
4214 }
4215
4216
4217 /* ==========================================
4218 * Short keys
4219 * ==========================================
4220 * One of the shortcomings of XXH32 and XXH64 was that their performance was
4221 * sub-optimal on short lengths. It used an iterative algorithm which strongly
4222 * favored lengths that were a multiple of 4 or 8.
4223 *
4224 * Instead of iterating over individual inputs, we use a set of single shot
4225 * functions which piece together a range of lengths and operate in constant time.
4226 *
4227 * Additionally, the number of multiplies has been significantly reduced. This
4228 * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
4229 *
4230 * Depending on the platform, this may or may not be faster than XXH32, but it
4231 * is almost guaranteed to be faster than XXH64.
4232 */
4233
4234 /*
4235 * At very short lengths, there isn't enough input to fully hide secrets, or use
4236 * the entire secret.
4237 *
4238 * There is also only a limited amount of mixing we can do before significantly
4239 * impacting performance.
4240 *
4241 * Therefore, we use different sections of the secret and always mix two secret
4242 * samples with an XOR. This should have no effect on performance on the
4243 * seedless or withSeed variants because everything _should_ be constant folded
4244 * by modern compilers.
4245 *
4246 * The XOR mixing hides individual parts of the secret and increases entropy.
4247 *
4248 * This adds an extra layer of strength for custom secrets.
4249 */
4250 XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
4251 XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
4252 {
4253 XXH_ASSERT(input != NULL);
4254 XXH_ASSERT(1 <= len && len <= 3);
4255 XXH_ASSERT(secret != NULL);
4256 /*
4257 * len = 1: combined = { input[0], 0x01, input[0], input[0] }
4258 * len = 2: combined = { input[1], 0x02, input[0], input[1] }
4259 * len = 3: combined = { input[2], 0x03, input[0], input[1] }
4260 */
4261 { xxh_u8 const c1 = input[0];
4262 xxh_u8 const c2 = input[len >> 1];
4263 xxh_u8 const c3 = input[len - 1];
4264 xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24)
4265 | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
4266 xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
4267 xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
4268 return XXH64_avalanche(keyed);
4269 }
4270 }
4271
4272 XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
4273 XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
4274 {
4275 XXH_ASSERT(input != NULL);
4276 XXH_ASSERT(secret != NULL);
4277 XXH_ASSERT(4 <= len && len <= 8);
4278 seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
4279 { xxh_u32 const input1 = XXH_readLE32(input);
4280 xxh_u32 const input2 = XXH_readLE32(input + len - 4);
4281 xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
4282 xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
4283 xxh_u64 const keyed = input64 ^ bitflip;
4284 return XXH3_rrmxmx(keyed, len);
4285 }
4286 }
4287
4288 XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
4289 XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
4290 {
4291 XXH_ASSERT(input != NULL);
4292 XXH_ASSERT(secret != NULL);
4293 XXH_ASSERT(9 <= len && len <= 16);
4294 { xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
4295 xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
4296 xxh_u64 const input_lo = XXH_readLE64(input) ^ bitflip1;
4297 xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
4298 xxh_u64 const acc = len
4299 + XXH_swap64(input_lo) + input_hi
4300 + XXH3_mul128_fold64(input_lo, input_hi);
4301 return XXH3_avalanche(acc);
4302 }
4303 }
4304
4305 XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
4306 XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
4307 {
4308 XXH_ASSERT(len <= 16);
4309 { if (XXH_likely(len > 8)) return XXH3_len_9to16_64b(input, len, secret, seed);
4310 if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
4311 if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
4312 return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
4313 }
4314 }
4315
4316 /*
4317 * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
4318 * multiplication by zero, affecting hashes of lengths 17 to 240.
4319 *
4320 * However, they are very unlikely.
4321 *
4322 * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
4323 * unseeded non-cryptographic hashes, it does not attempt to defend itself
4324 * against specially crafted inputs, only random inputs.
4325 *
4326 * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
4327 * cancelling out the secret is taken an arbitrary number of times (addressed
4328 * in XXH3_accumulate_512), this collision is very unlikely with random inputs
4329 * and/or proper seeding:
4330 *
4331 * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
4332 * function that is only called up to 16 times per hash with up to 240 bytes of
4333 * input.
4334 *
4335 * This is not too bad for a non-cryptographic hash function, especially with
4336 * only 64 bit outputs.
4337 *
4338 * The 128-bit variant (which trades some speed for strength) is NOT affected
4339 * by this, although it is always a good idea to use a proper seed if you care
4340 * about strength.
4341 */
4342 XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
4343 const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
4344 {
4345 #if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
4346 && defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \
4347 && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */
4348 /*
4349 * UGLY HACK:
4350 * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
4351 * slower code.
4352 *
4353 * By forcing seed64 into a register, we disrupt the cost model and
4354 * cause it to scalarize. See `XXH32_round()`
4355 *
4356 * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
4357 * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
4358 * GCC 9.2, despite both emitting scalar code.
4359 *
4360 * GCC generates much better scalar code than Clang for the rest of XXH3,
4361 * which is why finding a more optimal codepath is an interest.
4362 */
4363 XXH_COMPILER_GUARD(seed64);
4364 #endif
4365 { xxh_u64 const input_lo = XXH_readLE64(input);
4366 xxh_u64 const input_hi = XXH_readLE64(input+8);
4367 return XXH3_mul128_fold64(
4368 input_lo ^ (XXH_readLE64(secret) + seed64),
4369 input_hi ^ (XXH_readLE64(secret+8) - seed64)
4370 );
4371 }
4372 }
4373
4374 /* For mid range keys, XXH3 uses a Mum-hash variant. */
4375 XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
4376 XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
4377 const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
4378 XXH64_hash_t seed)
4379 {
4380 XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
4381 XXH_ASSERT(16 < len && len <= 128);
4382
4383 { xxh_u64 acc = len * XXH_PRIME64_1;
4384 #if XXH_SIZE_OPT >= 1
4385 /* Smaller and cleaner, but slightly slower. */
4386 unsigned int i = (unsigned int)(len - 1) / 32;
4387 do {
4388 acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
4389 acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
4390 } while (i-- != 0);
4391 #else
4392 if (len > 32) {
4393 if (len > 64) {
4394 if (len > 96) {
4395 acc += XXH3_mix16B(input+48, secret+96, seed);
4396 acc += XXH3_mix16B(input+len-64, secret+112, seed);
4397 }
4398 acc += XXH3_mix16B(input+32, secret+64, seed);
4399 acc += XXH3_mix16B(input+len-48, secret+80, seed);
4400 }
4401 acc += XXH3_mix16B(input+16, secret+32, seed);
4402 acc += XXH3_mix16B(input+len-32, secret+48, seed);
4403 }
4404 acc += XXH3_mix16B(input+0, secret+0, seed);
4405 acc += XXH3_mix16B(input+len-16, secret+16, seed);
4406 #endif
4407 return XXH3_avalanche(acc);
4408 }
4409 }
4410
4411 #define XXH3_MIDSIZE_MAX 240
4412
4413 XXH_NO_INLINE XXH_PUREF XXH64_hash_t
4414 XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
4415 const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
4416 XXH64_hash_t seed)
4417 {
4418 XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
4419 XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
4420
4421 #define XXH3_MIDSIZE_STARTOFFSET 3
4422 #define XXH3_MIDSIZE_LASTOFFSET 17
4423
4424 { xxh_u64 acc = len * XXH_PRIME64_1;
4425 xxh_u64 acc_end;
4426 unsigned int const nbRounds = (unsigned int)len / 16;
4427 unsigned int i;
4428 XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
4429 for (i=0; i<8; i++) {
4430 acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
4431 }
4432 /* last bytes */
4433 acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
4434 XXH_ASSERT(nbRounds >= 8);
4435 acc = XXH3_avalanche(acc);
4436 #if defined(__clang__) /* Clang */ \
4437 && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
4438 && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */
4439 /*
4440 * UGLY HACK:
4441 * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
4442 * In everywhere else, it uses scalar code.
4443 *
4444 * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
4445 * would still be slower than UMAAL (see XXH_mult64to128).
4446 *
4447 * Unfortunately, Clang doesn't handle the long multiplies properly and
4448 * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
4449 * scalarized into an ugly mess of VMOV.32 instructions.
4450 *
4451 * This mess is difficult to avoid without turning autovectorization
4452 * off completely, but they are usually relatively minor and/or not
4453 * worth it to fix.
4454 *
4455 * This loop is the easiest to fix, as unlike XXH32, this pragma
4456 * _actually works_ because it is a loop vectorization instead of an
4457 * SLP vectorization.
4458 */
4459 #pragma clang loop vectorize(disable)
4460 #endif
4461 for (i=8 ; i < nbRounds; i++) {
4462 /*
4463 * Prevents clang for unrolling the acc loop and interleaving with this one.
4464 */
4465 XXH_COMPILER_GUARD(acc);
4466 acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
4467 }
4468 return XXH3_avalanche(acc + acc_end);
4469 }
4470 }
4471
4472
4473 /* ======= Long Keys ======= */
4474
4475 #define XXH_STRIPE_LEN 64
4476 #define XXH_SECRET_CONSUME_RATE 8 /* nb of secret bytes consumed at each accumulation */
4477 #define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))
4478
4479 #ifdef XXH_OLD_NAMES
4480 # define STRIPE_LEN XXH_STRIPE_LEN
4481 # define ACC_NB XXH_ACC_NB
4482 #endif
4483
4484 #ifndef XXH_PREFETCH_DIST
4485 # ifdef __clang__
4486 # define XXH_PREFETCH_DIST 320
4487 # else
4488 # if (XXH_VECTOR == XXH_AVX512)
4489 # define XXH_PREFETCH_DIST 512
4490 # else
4491 # define XXH_PREFETCH_DIST 384
4492 # endif
4493 # endif /* __clang__ */
4494 #endif /* XXH_PREFETCH_DIST */
4495
4496 /*
4497 * These macros are to generate an XXH3_accumulate() function.
4498 * The two arguments select the name suffix and target attribute.
4499 *
4500 * The name of this symbol is XXH3_accumulate_<name>() and it calls
4501 * XXH3_accumulate_512_<name>().
4502 *
4503 * It may be useful to hand implement this function if the compiler fails to
4504 * optimize the inline function.
4505 */
4506 #define XXH3_ACCUMULATE_TEMPLATE(name) \
4507 void \
4508 XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc, \
4509 const xxh_u8* XXH_RESTRICT input, \
4510 const xxh_u8* XXH_RESTRICT secret, \
4511 size_t nbStripes) \
4512 { \
4513 size_t n; \
4514 for (n = 0; n < nbStripes; n++ ) { \
4515 const xxh_u8* const in = input + n*XXH_STRIPE_LEN; \
4516 XXH_PREFETCH(in + XXH_PREFETCH_DIST); \
4517 XXH3_accumulate_512_##name( \
4518 acc, \
4519 in, \
4520 secret + n*XXH_SECRET_CONSUME_RATE); \
4521 } \
4522 }
4523
4524
4525 XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
4526 {
4527 if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
4528 XXH_memcpy(dst, &v64, sizeof(v64));
4529 }
4530
4531 /* Several intrinsic functions below are supposed to accept __int64 as argument,
4532 * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ .
4533 * However, several environments do not define __int64 type,
4534 * requiring a workaround.
4535 */
4536 #if !defined (__VMS) \
4537 && (defined (__cplusplus) \
4538 || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
4539 typedef int64_t xxh_i64;
4540 #else
4541 /* the following type must have a width of 64-bit */
4542 typedef long long xxh_i64;
4543 #endif
4544
4545
4546 /*
4547 * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
4548 *
4549 * It is a hardened version of UMAC, based off of FARSH's implementation.
4550 *
4551 * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
4552 * implementations, and it is ridiculously fast.
4553 *
4554 * We harden it by mixing the original input to the accumulators as well as the product.
4555 *
4556 * This means that in the (relatively likely) case of a multiply by zero, the
4557 * original input is preserved.
4558 *
4559 * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
4560 * cross-pollination, as otherwise the upper and lower halves would be
4561 * essentially independent.
4562 *
4563 * This doesn't matter on 64-bit hashes since they all get merged together in
4564 * the end, so we skip the extra step.
4565 *
4566 * Both XXH3_64bits and XXH3_128bits use this subroutine.
4567 */
4568
4569 #if (XXH_VECTOR == XXH_AVX512) \
4570 || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0)
4571
4572 #ifndef XXH_TARGET_AVX512
4573 # define XXH_TARGET_AVX512 /* disable attribute target */
4574 #endif
4575
4576 XXH_FORCE_INLINE XXH_TARGET_AVX512 void
4577 XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
4578 const void* XXH_RESTRICT input,
4579 const void* XXH_RESTRICT secret)
4580 {
4581 __m512i* const xacc = (__m512i *) acc;
4582 XXH_ASSERT((((size_t)acc) & 63) == 0);
4583 XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
4584
4585 {
4586 /* data_vec = input[0]; */
4587 __m512i const data_vec = _mm512_loadu_si512 (input);
4588 /* key_vec = secret[0]; */
4589 __m512i const key_vec = _mm512_loadu_si512 (secret);
4590 /* data_key = data_vec ^ key_vec; */
4591 __m512i const data_key = _mm512_xor_si512 (data_vec, key_vec);
4592 /* data_key_lo = data_key >> 32; */
4593 __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32);
4594 /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
4595 __m512i const product = _mm512_mul_epu32 (data_key, data_key_lo);
4596 /* xacc[0] += swap(data_vec); */
4597 __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2));
4598 __m512i const sum = _mm512_add_epi64(*xacc, data_swap);
4599 /* xacc[0] += product; */
4600 *xacc = _mm512_add_epi64(product, sum);
4601 }
4602 }
4603 XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)
4604
4605 /*
4606 * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
4607 *
4608 * Multiplication isn't perfect, as explained by Google in HighwayHash:
4609 *
4610 * // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
4611 * // varying degrees. In descending order of goodness, bytes
4612 * // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
4613 * // As expected, the upper and lower bytes are much worse.
4614 *
4615 * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
4616 *
4617 * Since our algorithm uses a pseudorandom secret to add some variance into the
4618 * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
4619 *
4620 * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
4621 * extraction.
4622 *
4623 * Both XXH3_64bits and XXH3_128bits use this subroutine.
4624 */
4625
4626 XXH_FORCE_INLINE XXH_TARGET_AVX512 void
4627 XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
4628 {
4629 XXH_ASSERT((((size_t)acc) & 63) == 0);
4630 XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
4631 { __m512i* const xacc = (__m512i*) acc;
4632 const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
4633
4634 /* xacc[0] ^= (xacc[0] >> 47) */
4635 __m512i const acc_vec = *xacc;
4636 __m512i const shifted = _mm512_srli_epi64 (acc_vec, 47);
4637 /* xacc[0] ^= secret; */
4638 __m512i const key_vec = _mm512_loadu_si512 (secret);
4639 __m512i const data_key = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */);
4640
4641 /* xacc[0] *= XXH_PRIME32_1; */
4642 __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32);
4643 __m512i const prod_lo = _mm512_mul_epu32 (data_key, prime32);
4644 __m512i const prod_hi = _mm512_mul_epu32 (data_key_hi, prime32);
4645 *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
4646 }
4647 }
4648
4649 XXH_FORCE_INLINE XXH_TARGET_AVX512 void
4650 XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
4651 {
4652 XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0);
4653 XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64);
4654 XXH_ASSERT(((size_t)customSecret & 63) == 0);
4655 (void)(&XXH_writeLE64);
4656 { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
4657 __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64);
4658 __m512i const seed = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos);
4659
4660 const __m512i* const src = (const __m512i*) ((const void*) XXH3_kSecret);
4661 __m512i* const dest = ( __m512i*) customSecret;
4662 int i;
4663 XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
4664 XXH_ASSERT(((size_t)dest & 63) == 0);
4665 for (i=0; i < nbRounds; ++i) {
4666 dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
4667 } }
4668 }
4669
4670 #endif
4671
4672 #if (XXH_VECTOR == XXH_AVX2) \
4673 || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0)
4674
4675 #ifndef XXH_TARGET_AVX2
4676 # define XXH_TARGET_AVX2 /* disable attribute target */
4677 #endif
4678
4679 XXH_FORCE_INLINE XXH_TARGET_AVX2 void
4680 XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
4681 const void* XXH_RESTRICT input,
4682 const void* XXH_RESTRICT secret)
4683 {
4684 XXH_ASSERT((((size_t)acc) & 31) == 0);
4685 { __m256i* const xacc = (__m256i *) acc;
4686 /* Unaligned. This is mainly for pointer arithmetic, and because
4687 * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
4688 const __m256i* const xinput = (const __m256i *) input;
4689 /* Unaligned. This is mainly for pointer arithmetic, and because
4690 * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
4691 const __m256i* const xsecret = (const __m256i *) secret;
4692
4693 size_t i;
4694 for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
4695 /* data_vec = xinput[i]; */
4696 __m256i const data_vec = _mm256_loadu_si256 (xinput+i);
4697 /* key_vec = xsecret[i]; */
4698 __m256i const key_vec = _mm256_loadu_si256 (xsecret+i);
4699 /* data_key = data_vec ^ key_vec; */
4700 __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec);
4701 /* data_key_lo = data_key >> 32; */
4702 __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32);
4703 /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
4704 __m256i const product = _mm256_mul_epu32 (data_key, data_key_lo);
4705 /* xacc[i] += swap(data_vec); */
4706 __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
4707 __m256i const sum = _mm256_add_epi64(xacc[i], data_swap);
4708 /* xacc[i] += product; */
4709 xacc[i] = _mm256_add_epi64(product, sum);
4710 } }
4711 }
4712 XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)
4713
4714 XXH_FORCE_INLINE XXH_TARGET_AVX2 void
4715 XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
4716 {
4717 XXH_ASSERT((((size_t)acc) & 31) == 0);
4718 { __m256i* const xacc = (__m256i*) acc;
4719 /* Unaligned. This is mainly for pointer arithmetic, and because
4720 * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
4721 const __m256i* const xsecret = (const __m256i *) secret;
4722 const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1);
4723
4724 size_t i;
4725 for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
4726 /* xacc[i] ^= (xacc[i] >> 47) */
4727 __m256i const acc_vec = xacc[i];
4728 __m256i const shifted = _mm256_srli_epi64 (acc_vec, 47);
4729 __m256i const data_vec = _mm256_xor_si256 (acc_vec, shifted);
4730 /* xacc[i] ^= xsecret; */
4731 __m256i const key_vec = _mm256_loadu_si256 (xsecret+i);
4732 __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec);
4733
4734 /* xacc[i] *= XXH_PRIME32_1; */
4735 __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32);
4736 __m256i const prod_lo = _mm256_mul_epu32 (data_key, prime32);
4737 __m256i const prod_hi = _mm256_mul_epu32 (data_key_hi, prime32);
4738 xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
4739 }
4740 }
4741 }
4742
4743 XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
4744 {
4745 XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0);
4746 XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6);
4747 XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64);
4748 (void)(&XXH_writeLE64);
4749 XXH_PREFETCH(customSecret);
4750 { __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);
4751
4752 const __m256i* const src = (const __m256i*) ((const void*) XXH3_kSecret);
4753 __m256i* dest = ( __m256i*) customSecret;
4754
4755 # if defined(__GNUC__) || defined(__clang__)
4756 /*
4757 * On GCC & Clang, marking 'dest' as modified will cause the compiler:
4758 * - do not extract the secret from sse registers in the internal loop
4759 * - use less common registers, and avoid pushing these reg into stack
4760 */
4761 XXH_COMPILER_GUARD(dest);
4762 # endif
4763 XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
4764 XXH_ASSERT(((size_t)dest & 31) == 0);
4765
4766 /* GCC -O2 need unroll loop manually */
4767 dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed);
4768 dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed);
4769 dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed);
4770 dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed);
4771 dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed);
4772 dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed);
4773 }
4774 }
4775
4776 #endif
4777
4778 /* x86dispatch always generates SSE2 */
4779 #if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)
4780
4781 #ifndef XXH_TARGET_SSE2
4782 # define XXH_TARGET_SSE2 /* disable attribute target */
4783 #endif
4784
4785 XXH_FORCE_INLINE XXH_TARGET_SSE2 void
4786 XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
4787 const void* XXH_RESTRICT input,
4788 const void* XXH_RESTRICT secret)
4789 {
4790 /* SSE2 is just a half-scale version of the AVX2 version. */
4791 XXH_ASSERT((((size_t)acc) & 15) == 0);
4792 { __m128i* const xacc = (__m128i *) acc;
4793 /* Unaligned. This is mainly for pointer arithmetic, and because
4794 * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
4795 const __m128i* const xinput = (const __m128i *) input;
4796 /* Unaligned. This is mainly for pointer arithmetic, and because
4797 * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
4798 const __m128i* const xsecret = (const __m128i *) secret;
4799
4800 size_t i;
4801 for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
4802 /* data_vec = xinput[i]; */
4803 __m128i const data_vec = _mm_loadu_si128 (xinput+i);
4804 /* key_vec = xsecret[i]; */
4805 __m128i const key_vec = _mm_loadu_si128 (xsecret+i);
4806 /* data_key = data_vec ^ key_vec; */
4807 __m128i const data_key = _mm_xor_si128 (data_vec, key_vec);
4808 /* data_key_lo = data_key >> 32; */
4809 __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
4810 /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
4811 __m128i const product = _mm_mul_epu32 (data_key, data_key_lo);
4812 /* xacc[i] += swap(data_vec); */
4813 __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
4814 __m128i const sum = _mm_add_epi64(xacc[i], data_swap);
4815 /* xacc[i] += product; */
4816 xacc[i] = _mm_add_epi64(product, sum);
4817 } }
4818 }
4819 XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)
4820
4821 XXH_FORCE_INLINE XXH_TARGET_SSE2 void
4822 XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
4823 {
4824 XXH_ASSERT((((size_t)acc) & 15) == 0);
4825 { __m128i* const xacc = (__m128i*) acc;
4826 /* Unaligned. This is mainly for pointer arithmetic, and because
4827 * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
4828 const __m128i* const xsecret = (const __m128i *) secret;
4829 const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1);
4830
4831 size_t i;
4832 for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
4833 /* xacc[i] ^= (xacc[i] >> 47) */
4834 __m128i const acc_vec = xacc[i];
4835 __m128i const shifted = _mm_srli_epi64 (acc_vec, 47);
4836 __m128i const data_vec = _mm_xor_si128 (acc_vec, shifted);
4837 /* xacc[i] ^= xsecret[i]; */
4838 __m128i const key_vec = _mm_loadu_si128 (xsecret+i);
4839 __m128i const data_key = _mm_xor_si128 (data_vec, key_vec);
4840
4841 /* xacc[i] *= XXH_PRIME32_1; */
4842 __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
4843 __m128i const prod_lo = _mm_mul_epu32 (data_key, prime32);
4844 __m128i const prod_hi = _mm_mul_epu32 (data_key_hi, prime32);
4845 xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
4846 }
4847 }
4848 }
4849
4850 XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
4851 {
4852 XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
4853 (void)(&XXH_writeLE64);
4854 { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);
4855
4856 # if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
4857 /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */
4858 XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
4859 __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
4860 # else
4861 __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64);
4862 # endif
4863 int i;
4864
4865 const void* const src16 = XXH3_kSecret;
4866 __m128i* dst16 = (__m128i*) customSecret;
4867 # if defined(__GNUC__) || defined(__clang__)
4868 /*
4869 * On GCC & Clang, marking 'dest' as modified will cause the compiler:
4870 * - do not extract the secret from sse registers in the internal loop
4871 * - use less common registers, and avoid pushing these reg into stack
4872 */
4873 XXH_COMPILER_GUARD(dst16);
4874 # endif
4875 XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */
4876 XXH_ASSERT(((size_t)dst16 & 15) == 0);
4877
4878 for (i=0; i < nbRounds; ++i) {
4879 dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
4880 } }
4881 }
4882
4883 #endif
4884
4885 #if (XXH_VECTOR == XXH_NEON)
4886
4887 /* forward declarations for the scalar routines */
4888 XXH_FORCE_INLINE void
4889 XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input,
4890 void const* XXH_RESTRICT secret, size_t lane);
4891
4892 XXH_FORCE_INLINE void
4893 XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
4894 void const* XXH_RESTRICT secret, size_t lane);
4895
4896 /*!
4897 * @internal
4898 * @brief The bulk processing loop for NEON and WASM SIMD128.
4899 *
4900 * The NEON code path is actually partially scalar when running on AArch64. This
4901 * is to optimize the pipelining and can have up to 15% speedup depending on the
4902 * CPU, and it also mitigates some GCC codegen issues.
4903 *
4904 * @see XXH3_NEON_LANES for configuring this and details about this optimization.
4905 *
4906 * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
4907 * integers instead of the other platforms which mask full 64-bit vectors,
4908 * so the setup is more complicated than just shifting right.
4909 *
4910 * Additionally, there is an optimization for 4 lanes at once noted below.
4911 *
4912 * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
4913 * there needs to be *three* versions of the accumulate operation used
4914 * for the remaining 2 lanes.
4915 *
4916 * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap
4917 * nearly perfectly.
4918 */
4919
4920 XXH_FORCE_INLINE void
4921 XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
4922 const void* XXH_RESTRICT input,
4923 const void* XXH_RESTRICT secret)
4924 {
4925 XXH_ASSERT((((size_t)acc) & 15) == 0);
4926 XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
4927 { /* GCC for darwin arm64 does not like aliasing here */
4928 xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
4929 /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
4930 uint8_t const* xinput = (const uint8_t *) input;
4931 uint8_t const* xsecret = (const uint8_t *) secret;
4932
4933 size_t i;
4934 #ifdef __wasm_simd128__
4935 /*
4936 * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret
4937 * is constant propagated, which results in it converting it to this
4938 * inside the loop:
4939 *
4940 * a = v128.load(XXH3_kSecret + 0 + $secret_offset, offset = 0)
4941 * b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0)
4942 * ...
4943 *
4944 * This requires a full 32-bit address immediate (and therefore a 6 byte
4945 * instruction) as well as an add for each offset.
4946 *
4947 * Putting an asm guard prevents it from folding (at the cost of losing
4948 * the alignment hint), and uses the free offset in `v128.load` instead
4949 * of adding secret_offset each time which overall reduces code size by
4950 * about a kilobyte and improves performance.
4951 */
4952 XXH_COMPILER_GUARD(xsecret);
4953 #endif
4954 /* Scalar lanes use the normal scalarRound routine */
4955 for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
4956 XXH3_scalarRound(acc, input, secret, i);
4957 }
4958 i = 0;
4959 /* 4 NEON lanes at a time. */
4960 for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
4961 /* data_vec = xinput[i]; */
4962 uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput + (i * 16));
4963 uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput + ((i+1) * 16));
4964 /* key_vec = xsecret[i]; */
4965 uint64x2_t key_vec_1 = XXH_vld1q_u64(xsecret + (i * 16));
4966 uint64x2_t key_vec_2 = XXH_vld1q_u64(xsecret + ((i+1) * 16));
4967 /* data_swap = swap(data_vec) */
4968 uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
4969 uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1);
4970 /* data_key = data_vec ^ key_vec; */
4971 uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
4972 uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
4973
4974 /*
4975 * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
4976 * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
4977 * get one vector with the low 32 bits of each lane, and one vector
4978 * with the high 32 bits of each lane.
4979 *
4980 * The intrinsic returns a double vector because the original ARMv7-a
4981 * instruction modified both arguments in place. AArch64 and SIMD128 emit
4982 * two instructions from this intrinsic.
4983 *
4984 * [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
4985 * [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
4986 */
4987 uint32x4x2_t unzipped = vuzpq_u32(
4988 vreinterpretq_u32_u64(data_key_1),
4989 vreinterpretq_u32_u64(data_key_2)
4990 );
4991 /* data_key_lo = data_key & 0xFFFFFFFF */
4992 uint32x4_t data_key_lo = unzipped.val[0];
4993 /* data_key_hi = data_key >> 32 */
4994 uint32x4_t data_key_hi = unzipped.val[1];
4995 /*
4996 * Then, we can split the vectors horizontally and multiply which, as for most
4997 * widening intrinsics, have a variant that works on both high half vectors
4998 * for free on AArch64. A similar instruction is available on SIMD128.
4999 *
5000 * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
Showing first 5,000 of 6,773 lines. View raw