/* * xxHash - Extremely Fast Hash algorithm * Header File * Copyright (C) 2012-2021 Yann Collet * * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php) * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are * met: * * * Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * * Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following disclaimer * in the documentation and/or other materials provided with the * distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * You can contact the author at: * - xxHash homepage: https://www.xxhash.com * - xxHash source repository: https://github.com/Cyan4973/xxHash
*/
/*! * @mainpage xxHash * * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed * limits. * * It is proposed in four flavors, in three families: * 1. @ref XXH32_family * - Classic 32-bit hash function. Simple, compact, and runs on almost all * 32-bit and 64-bit systems. * 2. @ref XXH64_family * - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most * 64-bit systems (but _not_ 32-bit systems). * 3. @ref XXH3_family * - Modern 64-bit and 128-bit hash function family which features improved * strength and performance across the board, especially on smaller data. * It benefits greatly from SIMD and 64-bit without requiring it. * * Benchmarks * --- * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04. * The open source benchmark program is compiled with clang v10.0 using -O3 flag. * * | Hash Name | ISA ext | Width | Large Data Speed | Small Data Velocity | * | -------------------- | ------- | ----: | ---------------: | ------------------: | * | XXH3_64bits() | @b AVX2 | 64 | 59.4 GB/s | 133.1 | * | MeowHash | AES-NI | 128 | 58.2 GB/s | 52.5 | * | XXH3_128bits() | @b AVX2 | 128 | 57.9 GB/s | 118.1 | * | CLHash | PCLMUL | 64 | 37.1 GB/s | 58.1 | * | XXH3_64bits() | @b SSE2 | 64 | 31.5 GB/s | 133.1 | * | XXH3_128bits() | @b SSE2 | 128 | 29.6 GB/s | 118.1 | * | RAM sequential read | | N/A | 28.0 GB/s | N/A | * | ahash | AES-NI | 64 | 22.5 GB/s | 107.2 | * | City64 | | 64 | 22.0 GB/s | 76.6 | * | T1ha2 | | 64 | 22.0 GB/s | 99.0 | * | City128 | | 128 | 21.7 GB/s | 57.7 | * | FarmHash | AES-NI | 64 | 21.3 GB/s | 71.9 | * | XXH64() | | 64 | 19.4 GB/s | 71.0 | * | SpookyHash | | 64 | 19.3 GB/s | 53.2 | * | Mum | | 64 | 18.0 GB/s | 67.0 | * | CRC32C | SSE4.2 | 32 | 13.0 GB/s | 57.9 | * | XXH32() | | 32 | 9.7 GB/s | 71.9 | * | City32 | | 32 | 9.1 GB/s | 66.0 | * | Blake3* | @b AVX2 | 256 | 4.4 GB/s | 8.1 | * | Murmur3 | | 32 | 3.9 GB/s | 56.1 | * | SipHash* | | 64 | 3.0 GB/s | 43.2 | * | Blake3* | @b SSE2 | 256 | 2.4 GB/s | 8.1 | * | HighwayHash | | 64 | 1.4 GB/s | 6.0 | * | FNV64 | | 64 | 1.2 GB/s | 62.7 | * | Blake2* | | 256 | 1.1 GB/s | 5.1 | * | SHA1* | | 160 | 0.8 GB/s | 5.6 | * | MD5* | | 128 | 0.6 GB/s | 7.8 | * @note * - Hashes which require a specific ISA extension are noted. SSE2 is also noted, * even though it is mandatory on x64. * - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic * by modern standards. * - Small data velocity is a rough average of algorithm's efficiency for small * data. For more accurate information, see the wiki. * - More benchmarks and strength tests are found on the wiki: * https://github.com/Cyan4973/xxHash/wiki * * Usage * ------ * All xxHash variants use a similar API. Changing the algorithm is a trivial * substitution. * * @pre * For functions which take an input and length parameter, the following * requirements are assumed: * - The range from [`input`, `input + length`) is valid, readable memory. * - The only exception is if the `length` is `0`, `input` may be `NULL`. * - For C++, the objects must have the *TriviallyCopyable* property, as the * functions access bytes directly as if it was an array of `unsigned char`. * * @anchor single_shot_example * **Single Shot** * * These functions are stateless functions which hash a contiguous block of memory, * immediately returning the result. They are the easiest and usually the fastest * option. * * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits() * * @code{.c} * #include <string.h> * #include "xxhash.h" * * // Example for a function which hashes a null terminated string with XXH32(). * XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed) * { * // NULL pointers are only valid if the length is zero * size_t length = (string == NULL) ? 0 : strlen(string); * return XXH32(string, length, seed); * } * @endcode * * @anchor streaming_example * **Streaming** * * These groups of functions allow incremental hashing of unknown size, even * more than what would fit in a size_t. * * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset() * * @code{.c} * #include <stdio.h> * #include <assert.h> * #include "xxhash.h" * // Example for a function which hashes a FILE incrementally with XXH3_64bits(). * XXH64_hash_t hashFile(FILE* f) * { * // Allocate a state struct. Do not just use malloc() or new. * XXH3_state_t* state = XXH3_createState(); * assert(state != NULL && "Out of memory!"); * // Reset the state to start a new hashing session. * XXH3_64bits_reset(state); * char buffer[4096]; * size_t count; * // Read the file in chunks * while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) { * // Run update() as many times as necessary to process the data * XXH3_64bits_update(state, buffer, count); * } * // Retrieve the finalized hash. This will not change the state. * XXH64_hash_t result = XXH3_64bits_digest(state); * // Free the state. Do not use free(). * XXH3_freeState(state); * return result; * } * @endcode * * @file xxhash.h * xxHash prototypes and implementation
*/
#ifdefined (__cplusplus) extern"C" { #endif
/* **************************** * INLINE mode
******************************/ /*! * @defgroup public Public API * Contains details on the public xxHash functions. * @{
*/ #ifdef XXH_DOXYGEN /*! * @brief Gives access to internal state declaration, required for static allocation. * * Incompatible with dynamic linking, due to risks of ABI changes. * * Usage: * @code{.c} * #define XXH_STATIC_LINKING_ONLY * #include "xxhash.h" * @endcode
*/ # define XXH_STATIC_LINKING_ONLY /* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */
/*! * @brief Gives access to internal definitions. * * Usage: * @code{.c} * #define XXH_STATIC_LINKING_ONLY * #define XXH_IMPLEMENTATION * #include "xxhash.h" * @endcode
*/ # define XXH_IMPLEMENTATION /* Do not undef XXH_IMPLEMENTATION for Doxygen */
/*! * @brief Exposes the implementation and marks all functions as `inline`. * * Use these build macros to inline xxhash into the target unit. * Inlining improves performance on small inputs, especially when the length is * expressed as a compile-time constant: * * https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html * * It also keeps xxHash symbols private to the unit, so they are not exported. * * Usage: * @code{.c} * #define XXH_INLINE_ALL * #include "xxhash.h" * @endcode * Do not compile and link xxhash.o as a separate object, as it is not useful.
*/ # define XXH_INLINE_ALL # undef XXH_INLINE_ALL /*! * @brief Exposes the implementation without marking functions as inline.
*/ # define XXH_PRIVATE_API # undef XXH_PRIVATE_API /*! * @brief Emulate a namespace by transparently prefixing all symbols. * * If you want to include _and expose_ xxHash functions from within your own * library, but also want to avoid symbol collisions with other libraries which * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE * (therefore, avoid empty or numeric values). * * Note that no change is required within the calling program as long as it * includes `xxhash.h`: Regular symbol names will be automatically translated * by this header.
*/ # define XXH_NAMESPACE /* YOUR NAME HERE */ # undef XXH_NAMESPACE #endif
#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
&& !defined(XXH_INLINE_ALL_31684351384) /* this section should be traversed only once */ # define XXH_INLINE_ALL_31684351384 /* give access to the advanced API, required to compile implementations */ # undef XXH_STATIC_LINKING_ONLY /* avoid macro redef */ # define XXH_STATIC_LINKING_ONLY /* make all functions private */ # undef XXH_PUBLIC_API # ifdefined(__GNUC__) # define XXH_PUBLIC_API static __inline __attribute__((unused)) # elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) # define XXH_PUBLIC_API staticinline # elif defined(_MSC_VER) # define XXH_PUBLIC_API static __inline # else /* note: this version may generate warnings for unused static functions */ # define XXH_PUBLIC_API static # endif
/* * This part deals with the special case where a unit wants to inline xxHash, * but "xxhash.h" has previously been included without XXH_INLINE_ALL, * such as part of some previously included *.h header file. * Without further action, the new include would just be ignored, * and functions would effectively _not_ be inlined (silent failure). * The following macros solve this situation by prefixing all inlined names, * avoiding naming collision with previous inclusions.
*/ /* Before that, we unconditionally #undef all symbols, * in case they were already defined with XXH_NAMESPACE. * They will then be redefined for XXH_INLINE_ALL
*/ # undef XXH_versionNumber /* XXH32 */ # undef XXH32 # undef XXH32_createState # undef XXH32_freeState # undef XXH32_reset # undef XXH32_update # undef XXH32_digest # undef XXH32_copyState # undef XXH32_canonicalFromHash # undef XXH32_hashFromCanonical /* XXH64 */ # undef XXH64 # undef XXH64_createState # undef XXH64_freeState # undef XXH64_reset # undef XXH64_update # undef XXH64_digest # undef XXH64_copyState # undef XXH64_canonicalFromHash # undef XXH64_hashFromCanonical /* XXH3_64bits */ # undef XXH3_64bits # undef XXH3_64bits_withSecret # undef XXH3_64bits_withSeed # undef XXH3_64bits_withSecretandSeed # undef XXH3_createState # undef XXH3_freeState # undef XXH3_copyState # undef XXH3_64bits_reset # undef XXH3_64bits_reset_withSeed # undef XXH3_64bits_reset_withSecret # undef XXH3_64bits_update # undef XXH3_64bits_digest # undef XXH3_generateSecret /* XXH3_128bits */ # undef XXH128 # undef XXH3_128bits # undef XXH3_128bits_withSeed # undef XXH3_128bits_withSecret # undef XXH3_128bits_reset # undef XXH3_128bits_reset_withSeed # undef XXH3_128bits_reset_withSecret # undef XXH3_128bits_reset_withSecretandSeed # undef XXH3_128bits_update # undef XXH3_128bits_digest # undef XXH128_isEqual # undef XXH128_cmp # undef XXH128_canonicalFromHash # undef XXH128_hashFromCanonical /* Finally, free the namespace itself */ # undef XXH_NAMESPACE
/* employ the namespace for XXH_INLINE_ALL */ # define XXH_NAMESPACE XXH_INLINE_ /* * Some identifiers (enums, type names) are not symbols, * but they must nonetheless be renamed to avoid redeclaration. * Alternative solution: do not redeclare them. * However, this requires some #ifdefs, and has a more dispersed impact. * Meanwhile, renaming can be achieved in a single place.
*/ # define XXH_IPREF(Id) XXH_NAMESPACE ## Id # define XXH_OK XXH_IPREF(XXH_OK) # define XXH_ERROR XXH_IPREF(XXH_ERROR) # define XXH_errorcode XXH_IPREF(XXH_errorcode) # define XXH32_canonical_t XXH_IPREF(XXH32_canonical_t) # define XXH64_canonical_t XXH_IPREF(XXH64_canonical_t) # define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t) # define XXH32_state_s XXH_IPREF(XXH32_state_s) # define XXH32_state_t XXH_IPREF(XXH32_state_t) # define XXH64_state_s XXH_IPREF(XXH64_state_s) # define XXH64_state_t XXH_IPREF(XXH64_state_t) # define XXH3_state_s XXH_IPREF(XXH3_state_s) # define XXH3_state_t XXH_IPREF(XXH3_state_t) # define XXH128_hash_t XXH_IPREF(XXH128_hash_t) /* Ensure the header is parsed again, even if it was previously included */ # undef XXHASH_H_5627135585666179 # undef XXHASH_H_STATIC_13879238742 #endif/* XXH_INLINE_ALL || XXH_PRIVATE_API */
/* ************************************* * Version
***************************************/ #define XXH_VERSION_MAJOR 0 #define XXH_VERSION_MINOR 8 #define XXH_VERSION_RELEASE 2 /*! @brief Version number, encoded as two digits each */ #define XXH_VERSION_NUMBER (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
/*! * @brief Obtains the xxHash version. * * This is mostly useful when xxHash is compiled as a shared library, * since the returned value comes from the library, as opposed to header file. * * @return @ref XXH_VERSION_NUMBER of the invoked library.
*/
XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
/* **************************** * Common basic types
******************************/ #include <stddef.h> /* size_t */ /*! * @brief Exit code for the streaming API.
*/ typedefenum {
XXH_OK = 0, /*!< OK */
XXH_ERROR /*!< Error */
} XXH_errorcode;
/*-********************************************************************** * 32-bit hash
************************************************************************/ #ifdefined(XXH_DOXYGEN) /* Don't show <stdint.h> include */ /*! * @brief An unsigned 32-bit integer. * * Not necessarily defined to `uint32_t` but functionally equivalent.
*/ typedef uint32_t XXH32_hash_t;
#else # include <limits.h> # if UINT_MAX == 0xFFFFFFFFUL typedefunsignedint XXH32_hash_t; # elif ULONG_MAX == 0xFFFFFFFFUL typedefunsignedlong XXH32_hash_t; # else # error "unsupported platform: need a 32-bit type" # endif #endif
/*! * @} * * @defgroup XXH32_family XXH32 family * @ingroup public * Contains functions used in the classic 32-bit xxHash algorithm. * * @note * XXH32 is useful for older platforms, with no or poor 64-bit performance. * Note that the @ref XXH3_family provides competitive speed for both 32-bit * and 64-bit systems, and offers true 64/128 bit hash results. * * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families * @see @ref XXH32_impl for implementation details * @{
*/
/*! * @brief Calculates the 32-bit hash of @p input using xxHash32. * * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s * * See @ref single_shot_example "Single Shot Example" for an example. * * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * @param seed The 32-bit seed to alter the hash's output predictably. * * @pre * The memory between @p input and @p input + @p length must be valid, * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * * @return The calculated 32-bit hash value. * * @see * XXH64(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128(): * Direct equivalents for the other variants of xxHash. * @see * XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version.
*/
XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (constvoid* input, size_t length, XXH32_hash_t seed);
#ifndef XXH_NO_STREAM /*! * Streaming functions generate the xxHash value from an incremental input. * This method is slower than single-call functions, due to state management. * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized. * * An XXH state must first be allocated using `XXH*_createState()`. * * Start a new hash by initializing the state with a seed using `XXH*_reset()`. * * Then, feed the hash state by calling `XXH*_update()` as many times as necessary. * * The function returns an error code, with 0 meaning OK, and any other value * meaning there is an error. * * Finally, a hash value can be produced anytime, by using `XXH*_digest()`. * This function returns the nn-bits hash as an int or long long. * * It's still possible to continue inserting input into the hash state after a * digest, and generate new hash values later on by invoking `XXH*_digest()`. * * When done, release the state using `XXH*_freeState()`. * * @see streaming_example at the top of @ref xxhash.h for an example.
*/
/*! * @typedef struct XXH32_state_s XXH32_state_t * @brief The opaque state struct for the XXH32 streaming API. * * @see XXH32_state_s for details.
*/ typedefstruct XXH32_state_s XXH32_state_t;
/*! * @brief Allocates an @ref XXH32_state_t. * * Must be freed with XXH32_freeState(). * @return An allocated XXH32_state_t on success, `NULL` on failure.
*/
XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void); /*! * @brief Frees an @ref XXH32_state_t. * * Must be allocated with XXH32_createState(). * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState(). * @return XXH_OK.
*/
XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr); /*! * @brief Copies one @ref XXH32_state_t to another. * * @param dst_state The state to copy to. * @param src_state The state to copy from. * @pre * @p dst_state and @p src_state must not be `NULL` and must not overlap.
*/
XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
/*! * @brief Resets an @ref XXH32_state_t to begin a new hash. * * This function resets and seeds a state. Call it before @ref XXH32_update(). * * @param statePtr The state struct to reset. * @param seed The 32-bit seed to alter the hash result predictably. * * @pre * @p statePtr must not be `NULL`. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
*/
XXH_PUBLIC_API XXH_errorcode XXH32_reset (XXH32_state_t* statePtr, XXH32_hash_t seed);
/*! * @brief Consumes a block of @p input to an @ref XXH32_state_t. * * Call this to incrementally consume blocks of data. * * @param statePtr The state struct to update. * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * * @pre * @p statePtr must not be `NULL`. * @pre * The memory between @p input and @p input + @p length must be valid, * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
*/
XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, constvoid* input, size_t length);
/*! * @brief Returns the calculated hash value from an @ref XXH32_state_t. * * @note * Calling XXH32_digest() will not affect @p statePtr, so you can update, * digest, and update again. * * @param statePtr The state struct to calculate the hash from. * * @pre * @p statePtr must not be `NULL`. * * @return The calculated xxHash32 value from that state.
*/
XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr); #endif/* !XXH_NO_STREAM */
/******* Canonical representation *******/
/* * The default return values from XXH functions are unsigned 32 and 64 bit * integers. * This the simplest and fastest format for further post-processing. * * However, this leaves open the question of what is the order on the byte level, * since little and big endian conventions will store the same number differently. * * The canonical representation settles this issue by mandating big-endian * convention, the same convention as human-readable numbers (large digits first). * * When writing hash values to storage, sending them over a network, or printing * them, it's highly recommended to use the canonical representation to ensure * portability across a wider range of systems, present and future. * * The following functions allow transformation of hash values to and from * canonical format.
*/
/*! * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t. * * @param dst The @ref XXH32_canonical_t pointer to be stored to. * @param hash The @ref XXH32_hash_t to be converted. * * @pre * @p dst must not be `NULL`.
*/
XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
/*! * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t. * * @param src The @ref XXH32_canonical_t to convert. * * @pre * @p src must not be `NULL`. * * @return The converted hash.
*/
XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
/*! @cond Doxygen ignores this part */ /* * C23 __STDC_VERSION__ number hasn't been specified yet. For now * leave as `201711L` (C17 + 1). * TODO: Update to correct value when its been specified.
*/ #define XXH_C23_VN 201711L /*! @endcond */
/*! @cond Doxygen ignores this part */ /* C-language Attributes are added in C23. */ #ifdefined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute) # define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x) #else # define XXH_HAS_C_ATTRIBUTE(x) 0 #endif /*! @endcond */
/*! @cond Doxygen ignores this part */ /* * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute * introduced in CPP17 and C23. * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough * C23 : https://en.cppreference.com/w/c/language/attributes/fallthrough
*/ #if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough) # define XXH_FALLTHROUGH [[fallthrough]] #elif XXH_HAS_ATTRIBUTE(__fallthrough__) # define XXH_FALLTHROUGH __attribute__ ((__fallthrough__)) #else # define XXH_FALLTHROUGH /* fallthrough */ #endif /*! @endcond */
/*! @cond Doxygen ignores this part */ /* * Define XXH_NOESCAPE for annotated pointers in public API. * https://clang.llvm.org/docs/AttributeReference.html#noescape * As of writing this, only supported by clang.
*/ #if XXH_HAS_ATTRIBUTE(noescape) # define XXH_NOESCAPE __attribute__((noescape)) #else # define XXH_NOESCAPE #endif /*! @endcond */
/*! * @} * @ingroup public * @{
*/
#ifndef XXH_NO_LONG_LONG /*-********************************************************************** * 64-bit hash
************************************************************************/ #ifdefined(XXH_DOXYGEN) /* don't include <stdint.h> */ /*! * @brief An unsigned 64-bit integer. * * Not necessarily defined to `uint64_t` but functionally equivalent.
*/ typedef uint64_t XXH64_hash_t; #elif !defined (__VMS) \
&& (defined (__cplusplus) \
|| (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) # include <stdint.h> typedef uint64_t XXH64_hash_t; #else # include <limits.h> # ifdefined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL /* LP64 ABI says uint64_t is unsigned long */ typedefunsignedlong XXH64_hash_t; # else /* the following type must have a width of 64-bit */ typedefunsignedlonglong XXH64_hash_t; # endif #endif
/*! * @} * * @defgroup XXH64_family XXH64 family * @ingroup public * @{ * Contains functions used in the classic 64-bit xxHash algorithm. * * @note * XXH3 provides competitive speed for both 32-bit and 64-bit systems, * and offers true 64/128 bit hash results. * It provides better speed for systems with vector processing capabilities.
*/
/*! * @brief Calculates the 64-bit hash of @p input using xxHash64. * * This function usually runs faster on 64-bit systems, but slower on 32-bit * systems (see benchmark). * * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * @param seed The 64-bit seed to alter the hash's output predictably. * * @pre * The memory between @p input and @p input + @p length must be valid, * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * * @return The calculated 64-bit hash. * * @see * XXH32(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128(): * Direct equivalents for the other variants of xxHash. * @see * XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
*/
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE constvoid* input, size_t length, XXH64_hash_t seed);
/******* Streaming *******/ #ifndef XXH_NO_STREAM /*! * @brief The opaque state struct for the XXH64 streaming API. * * @see XXH64_state_s for details.
*/ typedefstruct XXH64_state_s XXH64_state_t; /* incomplete type */
/*! * @brief Allocates an @ref XXH64_state_t. * * Must be freed with XXH64_freeState(). * @return An allocated XXH64_state_t on success, `NULL` on failure.
*/
XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
/*! * @brief Frees an @ref XXH64_state_t. * * Must be allocated with XXH64_createState(). * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState(). * @return XXH_OK.
*/
XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr);
/*! * @brief Copies one @ref XXH64_state_t to another. * * @param dst_state The state to copy to. * @param src_state The state to copy from. * @pre * @p dst_state and @p src_state must not be `NULL` and must not overlap.
*/
XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
/*! * @brief Resets an @ref XXH64_state_t to begin a new hash. * * This function resets and seeds a state. Call it before @ref XXH64_update(). * * @param statePtr The state struct to reset. * @param seed The 64-bit seed to alter the hash result predictably. * * @pre * @p statePtr must not be `NULL`. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
*/
XXH_PUBLIC_API XXH_errorcode XXH64_reset (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
/*! * @brief Consumes a block of @p input to an @ref XXH64_state_t. * * Call this to incrementally consume blocks of data. * * @param statePtr The state struct to update. * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * * @pre * @p statePtr must not be `NULL`. * @pre * The memory between @p input and @p input + @p length must be valid, * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
*/
XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE constvoid* input, size_t length);
/*! * @brief Returns the calculated hash value from an @ref XXH64_state_t. * * @note * Calling XXH64_digest() will not affect @p statePtr, so you can update, * digest, and update again. * * @param statePtr The state struct to calculate the hash from. * * @pre * @p statePtr must not be `NULL`. * * @return The calculated xxHash64 value from that state.
*/
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr); #endif/* !XXH_NO_STREAM */ /******* Canonical representation *******/
/*! * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t. * * @param dst The @ref XXH64_canonical_t pointer to be stored to. * @param hash The @ref XXH64_hash_t to be converted. * * @pre * @p dst must not be `NULL`.
*/
XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
/*! * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t. * * @param src The @ref XXH64_canonical_t to convert. * * @pre * @p src must not be `NULL`. * * @return The converted hash.
*/
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
#ifndef XXH_NO_XXH3
/*! * @} * ************************************************************************ * @defgroup XXH3_family XXH3 family * @ingroup public * @{ * * XXH3 is a more recent hash algorithm featuring: * - Improved speed for both small and large inputs * - True 64-bit and 128-bit outputs * - SIMD acceleration * - Improved 32-bit viability * * Speed analysis methodology is explained here: * * https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html * * Compared to XXH64, expect XXH3 to run approximately * ~2x faster on large inputs and >3x faster on small ones, * exact differences vary depending on platform. * * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic, * but does not require it. * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3 * at competitive speeds, even without vector support. Further details are * explained in the implementation. * * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD * implementations for many common platforms: * - AVX512 * - AVX2 * - SSE2 * - ARM NEON * - WebAssembly SIMD128 * - POWER8 VSX * - s390x ZVector * This can be controlled via the @ref XXH_VECTOR macro, but it automatically * selects the best version according to predefined macros. For the x86 family, an * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c. * * XXH3 implementation is portable: * it has a generic C90 formulation that can be compiled on any platform, * all implementations generate exactly the same hash value on all platforms. * Starting from v0.8.0, it's also labelled "stable", meaning that * any future version will also generate the same hash value. * * XXH3 offers 2 variants, _64bits and _128bits. * * When only 64 bits are needed, prefer invoking the _64bits variant, as it * reduces the amount of mixing, resulting in faster speed on small inputs. * It's also generally simpler to manipulate a scalar return type than a struct. * * The API supports one-shot hashing, streaming mode, and custom secrets.
*/ /*-********************************************************************** * XXH3 64-bit variant
************************************************************************/
/*! * @brief 64-bit unseeded variant of XXH3. * * This is equivalent to @ref XXH3_64bits_withSeed() with a seed of 0, however * it may have slightly better performance due to constant propagation of the * defaults. * * @see * XXH32(), XXH64(), XXH3_128bits(): equivalent for the other xxHash algorithms * @see * XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants * @see * XXH3_64bits_reset(), XXH3_64bits_update(), XXH3_64bits_digest(): Streaming version.
*/
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE constvoid* input, size_t length);
/*! * @brief 64-bit seeded variant of XXH3 * * This variant generates a custom secret on the fly based on default secret * altered using the `seed` value. * * While this operation is decently fast, note that it's not completely free. * * @note * seed == 0 produces the same results as @ref XXH3_64bits(). * * @param input The data to hash * @param length The length * @param seed The 64-bit seed to alter the state.
*/
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE constvoid* input, size_t length, XXH64_hash_t seed);
/*! * The bare minimum size for a custom secret. * * @see * XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(), * XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
*/ #define XXH3_SECRET_SIZE_MIN 136
/*! * @brief 64-bit variant of XXH3 with a custom "secret". * * It's possible to provide any blob of bytes as a "secret" to generate the hash. * This makes it more difficult for an external actor to prepare an intentional collision. * The main condition is that secretSize *must* be large enough (>= XXH3_SECRET_SIZE_MIN). * However, the quality of the secret impacts the dispersion of the hash algorithm. * Therefore, the secret _must_ look like a bunch of random bytes. * Avoid "trivial" or structured data such as repeated sequences or a text document. * Whenever in doubt about the "randomness" of the blob of bytes, * consider employing "XXH3_generateSecret()" instead (see below). * It will generate a proper high entropy secret derived from the blob of bytes. * Another advantage of using XXH3_generateSecret() is that * it guarantees that all bits within the initial blob of bytes * will impact every bit of the output. * This is not necessarily the case when using the blob of bytes directly * because, when hashing _small_ inputs, only a portion of the secret is employed.
*/
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE constvoid* data, size_t len, XXH_NOESCAPE constvoid* secret, size_t secretSize);
/******* Streaming *******/ #ifndef XXH_NO_STREAM /* * Streaming requires state maintenance. * This operation costs memory and CPU. * As a consequence, streaming is slower than one-shot hashing. * For better performance, prefer one-shot functions whenever applicable.
*/
/*! * @brief The state struct for the XXH3 streaming API. * * @see XXH3_state_s for details.
*/ typedefstruct XXH3_state_s XXH3_state_t;
XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
/*! * @brief Copies one @ref XXH3_state_t to another. * * @param dst_state The state to copy to. * @param src_state The state to copy from. * @pre * @p dst_state and @p src_state must not be `NULL` and must not overlap.
*/
XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
/*! * @brief Resets an @ref XXH3_state_t to begin a new hash. * * This function resets `statePtr` and generate a secret with default parameters. Call it before @ref XXH3_64bits_update(). * Digest will be equivalent to `XXH3_64bits()`. * * @param statePtr The state struct to reset. * * @pre * @p statePtr must not be `NULL`. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure. *
*/
XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
/*! * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash. * * This function resets `statePtr` and generate a secret from `seed`. Call it before @ref XXH3_64bits_update(). * Digest will be equivalent to `XXH3_64bits_withSeed()`. * * @param statePtr The state struct to reset. * @param seed The 64-bit seed to alter the state. * * @pre * @p statePtr must not be `NULL`. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure. *
*/
XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
/*! * XXH3_64bits_reset_withSecret(): * `secret` is referenced, it _must outlive_ the hash streaming session. * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`, * and the quality of produced hash values depends on secret's entropy * (secret's content should look like a bunch of random bytes). * When in doubt about the randomness of a candidate `secret`, * consider employing `XXH3_generateSecret()` instead (see below).
*/
XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE constvoid* secret, size_t secretSize);
/*! * @brief Consumes a block of @p input to an @ref XXH3_state_t. * * Call this to incrementally consume blocks of data. * * @param statePtr The state struct to update. * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * * @pre * @p statePtr must not be `NULL`. * @pre * The memory between @p input and @p input + @p length must be valid, * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
*/
XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE constvoid* input, size_t length);
/*! * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t. * * @note * Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update, * digest, and update again. * * @param statePtr The state struct to calculate the hash from. * * @pre * @p statePtr must not be `NULL`. * * @return The calculated XXH3 64-bit hash value from that state.
*/
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr); #endif/* !XXH_NO_STREAM */
/* note : canonical representation of XXH3 is the same as XXH64
* since they both produce XXH64_hash_t values */
/*! * @brief The return value from 128-bit hashes. * * Stored in little endian order, although the fields themselves are in native * endianness.
*/ typedefstruct {
XXH64_hash_t low64; /*!< `value & 0xFFFFFFFFFFFFFFFF` */
XXH64_hash_t high64; /*!< `value >> 64` */
} XXH128_hash_t;
/*! * @brief Unseeded 128-bit variant of XXH3 * * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead * for shorter inputs. * * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of 0, however * it may have slightly better performance due to constant propagation of the * defaults. * * @see * XXH32(), XXH64(), XXH3_64bits(): equivalent for the other xxHash algorithms * @see * XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants * @see * XXH3_128bits_reset(), XXH3_128bits_update(), XXH3_128bits_digest(): Streaming version.
*/
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE constvoid* data, size_t len); /*! @brief Seeded 128-bit variant of XXH3. @see XXH3_64bits_withSeed(). */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE constvoid* data, size_t len, XXH64_hash_t seed); /*! @brief Custom secret 128-bit variant of XXH3. @see XXH3_64bits_withSecret(). */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE constvoid* data, size_t len, XXH_NOESCAPE constvoid* secret, size_t secretSize);
/******* Streaming *******/ #ifndef XXH_NO_STREAM /* * Streaming requires state maintenance. * This operation costs memory and CPU. * As a consequence, streaming is slower than one-shot hashing. * For better performance, prefer one-shot functions whenever applicable. * * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits(). * Use already declared XXH3_createState() and XXH3_freeState(). * * All reset and streaming functions have same meaning as their 64-bit counterpart.
*/
/*! * @brief Resets an @ref XXH3_state_t to begin a new hash. * * This function resets `statePtr` and generate a secret with default parameters. Call it before @ref XXH3_128bits_update(). * Digest will be equivalent to `XXH3_128bits()`. * * @param statePtr The state struct to reset. * * @pre * @p statePtr must not be `NULL`. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure. *
*/
XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
/*! * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash. * * This function resets `statePtr` and generate a secret from `seed`. Call it before @ref XXH3_128bits_update(). * Digest will be equivalent to `XXH3_128bits_withSeed()`. * * @param statePtr The state struct to reset. * @param seed The 64-bit seed to alter the state. * * @pre * @p statePtr must not be `NULL`. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure. *
*/
XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed); /*! @brief Custom secret 128-bit variant of XXH3. @see XXH_64bits_reset_withSecret(). */
XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE constvoid* secret, size_t secretSize);
/*! * @brief Consumes a block of @p input to an @ref XXH3_state_t. * * Call this to incrementally consume blocks of data. * * @param statePtr The state struct to update. * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * * @pre * @p statePtr must not be `NULL`. * @pre * The memory between @p input and @p input + @p length must be valid, * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
*/
XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE constvoid* input, size_t length);
/*! * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t. * * @note * Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update, * digest, and update again. * * @param statePtr The state struct to calculate the hash from. * * @pre * @p statePtr must not be `NULL`. * * @return The calculated XXH3 128-bit hash value from that state.
*/
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr); #endif/* !XXH_NO_STREAM */
/* Following helper functions make it possible to compare XXH128_hast_t values. * Since XXH128_hash_t is a structure, this capability is not offered by the language.
* Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
/*! * XXH128_isEqual(): * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
*/
XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
/*! * @brief Compares two @ref XXH128_hash_t * This comparator is compatible with stdlib's `qsort()`/`bsearch()`. * * @return: >0 if *h128_1 > *h128_2 * =0 if *h128_1 == *h128_2 * <0 if *h128_1 < *h128_2
*/
XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE constvoid* h128_1, XXH_NOESCAPE constvoid* h128_2);
/*! * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t. * * @param dst The @ref XXH128_canonical_t pointer to be stored to. * @param hash The @ref XXH128_hash_t to be converted. * * @pre * @p dst must not be `NULL`.
*/
XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
/*! * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t. * * @param src The @ref XXH128_canonical_t to convert. * * @pre * @p src must not be `NULL`. * * @return The converted hash.
*/
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
#ifdefined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) #define XXHASH_H_STATIC_13879238742 /* **************************************************************************** * This section contains declarations which are not guaranteed to remain stable. * They may change in future versions, becoming incompatible with a different * version of the library. * These declarations should only be used with static linking. * Never use them in association with dynamic linking!
***************************************************************************** */
/* * These definitions are only present to allow static allocation * of XXH states, on stack or in a struct, for example. * Never **ever** access their members directly.
*/
/*! * @internal * @brief Structure for XXH32 streaming API. * * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY, * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is * an opaque type. This allows fields to safely be changed. * * Typedef'd to @ref XXH32_state_t. * Do not access the members of this struct directly. * @see XXH64_state_s, XXH3_state_s
*/ struct XXH32_state_s {
XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
XXH32_hash_t large_len; /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
XXH32_hash_t v[4]; /*!< Accumulator lanes */
XXH32_hash_t mem32[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
XXH32_hash_t memsize; /*!< Amount of data in @ref mem32 */
XXH32_hash_t reserved; /*!< Reserved field. Do not read nor write to it. */
}; /* typedef'd to XXH32_state_t */
#ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */
/*! * @internal * @brief Structure for XXH64 streaming API. * * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY, * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is * an opaque type. This allows fields to safely be changed. * * Typedef'd to @ref XXH64_state_t. * Do not access the members of this struct directly. * @see XXH32_state_s, XXH3_state_s
*/ struct XXH64_state_s {
XXH64_hash_t total_len; /*!< Total length hashed. This is always 64-bit. */
XXH64_hash_t v[4]; /*!< Accumulator lanes */
XXH64_hash_t mem64[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
XXH32_hash_t memsize; /*!< Amount of data in @ref mem64 */
XXH32_hash_t reserved32; /*!< Reserved field, needed for padding anyways*/
XXH64_hash_t reserved64; /*!< Reserved field. Do not read or write to it. */
}; /* typedef'd to XXH64_state_t */
/* Old GCC versions only accept the attribute after the type in structures. */ #if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) /* C11+ */ \
&& ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \
&& defined(__GNUC__) # define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align) #else # define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type #endif
/*! * @brief The size of the internal XXH3 buffer. * * This is the optimal update size for incremental hashing. * * @see XXH3_64b_update(), XXH3_128b_update().
*/ #define XXH3_INTERNALBUFFER_SIZE 256
/*! * @internal * @brief Default size of the secret buffer (and @ref XXH3_kSecret). * * This is the size used in @ref XXH3_kSecret and the seeded functions. * * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
*/ #define XXH3_SECRET_DEFAULT_SIZE 192
/*! * @internal * @brief Structure for XXH3 streaming API. * * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY, * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. * Otherwise it is an opaque type. * Never use this definition in combination with dynamic library. * This allows fields to safely be changed in the future. * * @note ** This structure has a strict alignment requirement of 64 bytes!! ** * Do not allocate this with `malloc()` or `new`, * it will not be sufficiently aligned. * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation. * * Typedef'd to @ref XXH3_state_t. * Do never access the members of this struct directly. * * @see XXH3_INITSTATE() for stack initialization. * @see XXH3_createState(), XXH3_freeState(). * @see XXH32_state_s, XXH64_state_s
*/ struct XXH3_state_s {
XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]); /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
XXH_ALIGN_MEMBER(64, unsignedchar customSecret[XXH3_SECRET_DEFAULT_SIZE]); /*!< Used to store a custom secret generated from a seed. */
XXH_ALIGN_MEMBER(64, unsignedchar buffer[XXH3_INTERNALBUFFER_SIZE]); /*!< The internal buffer. @see XXH32_state_s::mem32 */
XXH32_hash_t bufferedSize; /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
XXH32_hash_t useSeed; /*!< Reserved field. Needed for padding on 64-bit. */
size_t nbStripesSoFar; /*!< Number or stripes processed. */
XXH64_hash_t totalLen; /*!< Total length hashed. 64-bit even on 32-bit targets. */
size_t nbStripesPerBlock; /*!< Number of stripes per block. */
size_t secretLimit; /*!< Size of @ref customSecret or @ref extSecret */
XXH64_hash_t seed; /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
XXH64_hash_t reserved64; /*!< Reserved field. */ constunsignedchar* extSecret; /*!< Reference to an external secret for the _withSecret variants, NULL
* for other variants. */ /* note: there may be some padding at the end due to alignment on 64 bytes */
}; /* typedef'd to XXH3_state_t */
#undef XXH_ALIGN_MEMBER
/*! * @brief Initializes a stack-allocated `XXH3_state_s`. * * When the @ref XXH3_state_t structure is merely emplaced on stack, * it should be initialized with XXH3_INITSTATE() or a memset() * in case its first reset uses XXH3_NNbits_reset_withSeed(). * This init can be omitted if the first reset uses default or _withSecret mode. * This operation isn't necessary when the state is created with XXH3_createState(). * Note that this doesn't prepare the state for a streaming operation, * it's still necessary to use XXH3_NNbits_reset*() afterwards.
*/ #define XXH3_INITSTATE(XXH3_state_ptr) \ do { \
XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
tmp_xxh3_state_ptr->seed = 0; \
tmp_xxh3_state_ptr->extSecret = NULL; \
} while(0)
/*! * simple alias to pre-selected XXH3_128bits variant
*/
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE constvoid* data, size_t len, XXH64_hash_t seed);
/* === Experimental API === */ /* Symbols defined below must be considered tied to a specific library version. */
/*! * XXH3_generateSecret(): * * Derive a high-entropy secret from any user-defined content, named customSeed. * The generated secret can be used in combination with `*_withSecret()` functions. * The `_withSecret()` variants are useful to provide a higher level of protection * than 64-bit seed, as it becomes much more difficult for an external actor to * guess how to impact the calculation logic. * * The function accepts as input a custom seed of any length and any content, * and derives from it a high-entropy secret of length @p secretSize into an * already allocated buffer @p secretBuffer. * * The generated secret can then be used with any `*_withSecret()` variant. * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(), * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret() * are part of this list. They all accept a `secret` parameter * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN) * _and_ feature very high entropy (consist of random-looking bytes). * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can * be employed to ensure proper quality. * * @p customSeed can be anything. It can have any size, even small ones, * and its content can be anything, even "poor entropy" sources such as a bunch * of zeroes. The resulting `secret` will nonetheless provide all required qualities. * * @pre * - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN * - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior. * * Example code: * @code{.c} * #include <stdio.h> * #include <stdlib.h> * #include <string.h> * #define XXH_STATIC_LINKING_ONLY // expose unstable API * #include "xxhash.h" * // Hashes argv[2] using the entropy from argv[1]. * int main(int argc, char* argv[]) * { * char secret[XXH3_SECRET_SIZE_MIN]; * if (argv != 3) { return 1; } * XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1])); * XXH64_hash_t h = XXH3_64bits_withSecret( * argv[2], strlen(argv[2]), * secret, sizeof(secret) * ); * printf("%016llx\n", (unsigned long long) h); * } * @endcode
*/
XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE constvoid* customSeed, size_t customSeedSize);
/*! * @brief Generate the same secret as the _withSeed() variants. * * The generated secret can be used in combination with *`*_withSecret()` and `_withSecretandSeed()` variants. * * Example C++ `std::string` hash class: * @code{.cpp} * #include <string> * #define XXH_STATIC_LINKING_ONLY // expose unstable API * #include "xxhash.h" * // Slow, seeds each time * class HashSlow { * XXH64_hash_t seed; * public: * HashSlow(XXH64_hash_t s) : seed{s} {} * size_t operator()(const std::string& x) const { * return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)}; * } * }; * // Fast, caches the seeded secret for future uses. * class HashFast { * unsigned char secret[XXH3_SECRET_SIZE_MIN]; * public: * HashFast(XXH64_hash_t s) { * XXH3_generateSecret_fromSeed(secret, seed); * } * size_t operator()(const std::string& x) const { * return size_t{ * XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret)) * }; * } * }; * @endcode * @param secretBuffer A writable buffer of @ref XXH3_SECRET_SIZE_MIN bytes * @param seed The seed to seed the state.
*/
XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
/*! * These variants generate hash values using either * @p seed for "short" keys (< XXH3_MIDSIZE_MAX = 240 bytes) * or @p secret for "large" keys (>= XXH3_MIDSIZE_MAX). * * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`. * `_withSeed()` has to generate the secret on the fly for "large" keys. * It's fast, but can be perceptible for "not so large" keys (< 1 KB). * `_withSecret()` has to generate the masks on the fly for "small" keys, * which requires more instructions than _withSeed() variants. * Therefore, _withSecretandSeed variant combines the best of both worlds. * * When @p secret has been generated by XXH3_generateSecret_fromSeed(), * this variant produces *exactly* the same results as `_withSeed()` variant, * hence offering only a pure speed benefit on "large" input, * by skipping the need to regenerate the secret for every large input. * * Another usage scenario is to hash the secret to a 64-bit hash value, * for example with XXH3_64bits(), which then becomes the seed, * and then employ both the seed and the secret in _withSecretandSeed(). * On top of speed, an added benefit is that each bit in the secret * has a 50% chance to swap each bit in the output, via its impact to the seed. * * This is not guaranteed when using the secret directly in "small data" scenarios, * because only portions of the secret are employed for small data.
*/
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
XXH3_64bits_withSecretandSeed(XXH_NOESCAPE constvoid* data, size_t len,
XXH_NOESCAPE constvoid* secret, size_t secretSize,
XXH64_hash_t seed); /*! @copydoc XXH3_64bits_withSecretandSeed() */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
XXH3_128bits_withSecretandSeed(XXH_NOESCAPE constvoid* input, size_t length,
XXH_NOESCAPE constvoid* secret, size_t secretSize,
XXH64_hash_t seed64); #ifndef XXH_NO_STREAM /*! @copydoc XXH3_64bits_withSecretandSeed() */
XXH_PUBLIC_API XXH_errorcode
XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
XXH_NOESCAPE constvoid* secret, size_t secretSize,
XXH64_hash_t seed64); /*! @copydoc XXH3_64bits_withSecretandSeed() */
XXH_PUBLIC_API XXH_errorcode
XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
XXH_NOESCAPE constvoid* secret, size_t secretSize,
XXH64_hash_t seed64); #endif/* !XXH_NO_STREAM */
/*-********************************************************************** * xxHash implementation *-********************************************************************** * xxHash's implementation used to be hosted inside xxhash.c. * * However, inlining requires implementation to be visible to the compiler, * hence be included alongside the header. * Previously, implementation was hosted inside xxhash.c, * which was then #included when inlining was activated. * This construction created issues with a few build and install systems, * as it required xxhash.c to be stored in /include directory. * * xxHash implementation is now directly integrated within xxhash.h. * As a consequence, xxhash.c is no longer needed in /include. * * xxhash.c is still available and is still useful. * In a "normal" setup, when xxhash is not inlined, * xxhash.h only exposes the prototypes and public symbols, * while xxhash.c can be built into an object file xxhash.o * which can then be linked into the final binary.
************************************************************************/
/*! * @defgroup tuning Tuning parameters * @{ * * Various macros to control xxHash's behavior.
*/ #ifdef XXH_DOXYGEN /*! * @brief Define this to disable 64-bit code. * * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
*/ # define XXH_NO_LONG_LONG # undef XXH_NO_LONG_LONG /* don't actually */ /*! * @brief Controls how unaligned memory is accessed. * * By default, access to unaligned memory is controlled by `memcpy()`, which is * safe and portable. * * Unfortunately, on some target/compiler combinations, the generated assembly * is sub-optimal. * * The below switch allow selection of a different access method * in the search for improved performance. * * @par Possible options: * * - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy` * @par * Use `memcpy()`. Safe and portable. Note that most modern compilers will * eliminate the function call and treat it as an unaligned access. * * - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))` * @par * Depends on compiler extensions and is therefore not portable. * This method is safe _if_ your compiler supports it, * and *generally* as fast or faster than `memcpy`. * * - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast * @par * Casts directly and dereferences. This method doesn't depend on the * compiler, but it violates the C standard as it directly dereferences an * unaligned pointer. It can generate buggy code on targets which do not * support unaligned memory accesses, but in some circumstances, it's the * only known way to get the most performance. * * - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift * @par * Also portable. This can generate the best code on old compilers which don't * inline small `memcpy()` calls, and it might also be faster on big-endian * systems which lack a native byteswap instruction. However, some compilers * will emit literal byteshifts even if the target supports unaligned access. * * * @warning * Methods 1 and 2 rely on implementation-defined behavior. Use these with * care, as what works on one compiler/platform/optimization level may cause * another to read garbage data or even crash. * * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.htmlfor details. * * Prefer these methods in priority order (0 > 3 > 1 > 2)
*/ # define XXH_FORCE_MEMORY_ACCESS 0
/*! * @def XXH_SIZE_OPT * @brief Controls how much xxHash optimizes for size. * * xxHash, when compiled, tends to result in a rather large binary size. This * is mostly due to heavy usage to forced inlining and constant folding of the * @ref XXH3_family to increase performance. * * However, some developers prefer size over speed. This option can * significantly reduce the size of the generated code. When using the `-Os` * or `-Oz` options on GCC or Clang, this is defined to 1 by default, * otherwise it is defined to 0. * * Most of these size optimizations can be controlled manually. * * This is a number from 0-2. * - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed * comes first. * - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more * conservative and disables hacks that increase code size. It implies the * options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0, * and @ref XXH3_NEON_LANES == 8 if they are not already defined. * - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible. * Performance may cry. For example, the single shot functions just use the * streaming API.
*/ # define XXH_SIZE_OPT 0
/*! * @def XXH_FORCE_ALIGN_CHECK * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32() * and XXH64() only). * * This is an important performance trick for architectures without decent * unaligned memory access performance. * * It checks for input alignment, and when conditions are met, uses a "fast * path" employing direct 32-bit/64-bit reads, resulting in _dramatically * faster_ read speed. * * The check costs one initial branch per hash, which is generally negligible, * but not zero. * * Moreover, it's not useful to generate an additional code path if memory * access uses the same instruction for both aligned and unaligned * addresses (e.g. x86 and aarch64). * * In these cases, the alignment check can be removed by setting this macro to 0. * Then the code will always use unaligned memory access. * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips * which are platforms known to offer good unaligned memory accesses performance. * * It is also disabled by default when @ref XXH_SIZE_OPT >= 1. * * This option does not affect XXH3 (only XXH32 and XXH64).
*/ # define XXH_FORCE_ALIGN_CHECK 0
/*! * @def XXH_NO_INLINE_HINTS * @brief When non-zero, sets all functions to `static`. * * By default, xxHash tries to force the compiler to inline almost all internal * functions. * * This can usually improve performance due to reduced jumping and improved * constant folding, but significantly increases the size of the binary which * might not be favorable. * * Additionally, sometimes the forced inlining can be detrimental to performance, * depending on the architecture. * * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the * compiler full control on whether to inline or not. * * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
*/ # define XXH_NO_INLINE_HINTS 0
/*! * @def XXH3_INLINE_SECRET * @brief Determines whether to inline the XXH3 withSecret code. * * When the secret size is known, the compiler can improve the performance * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret(). * * However, if the secret size is not known, it doesn't have any benefit. This * happens when xxHash is compiled into a global symbol. Therefore, if * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0. * * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers * that are *sometimes* force inline on -Og, and it is impossible to automatically * detect this optimization level.
*/ # define XXH3_INLINE_SECRET 0
/*! * @def XXH32_ENDJMP * @brief Whether to use a jump for `XXH32_finalize`. * * For performance, `XXH32_finalize` uses multiple branches in the finalizer. * This is generally preferable for performance, * but depending on exact architecture, a jmp may be preferable. * * This setting is only possibly making a difference for very small inputs.
*/ # define XXH32_ENDJMP 0
/*! * @internal * @brief Redefines old internal names. * * For compatibility with code that uses xxHash's internals before the names * were changed to improve namespacing. There is no other reason to use this.
*/ # define XXH_OLD_NAMES # undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
/*! * @def XXH_NO_STREAM * @brief Disables the streaming API. * * When xxHash is not inlined and the streaming functions are not used, disabling * the streaming functions can improve code size significantly, especially with * the @ref XXH3_family which tends to make constant folded copies of itself.
*/ # define XXH_NO_STREAM # undef XXH_NO_STREAM /* don't actually */ #endif/* XXH_DOXYGEN */ /*! * @}
*/
#ifndef XXH_FORCE_MEMORY_ACCESS /* can be defined externally, on command line for example */ /* prefer __packed__ structures (method 1) for GCC * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
* which for some reason does unaligned loads. */ # ifdefined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED)) # define XXH_FORCE_MEMORY_ACCESS 1 # endif #endif
#ifndef XXH_SIZE_OPT /* default to 1 for -Os or -Oz */ # if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__) # define XXH_SIZE_OPT 1 # else # define XXH_SIZE_OPT 0 # endif #endif
#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */ /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */ # if XXH_SIZE_OPT >= 1 || \ defined(__i386) || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
|| defined(_M_IX86) || defined(_M_X64) || defined(_M_ARM64) || defined(_M_ARM) /* visual */ # define XXH_FORCE_ALIGN_CHECK 0 # else # define XXH_FORCE_ALIGN_CHECK 1 # endif #endif
#ifndef XXH32_ENDJMP /* generally preferable for performance */ # define XXH32_ENDJMP 0 #endif
/*! * @defgroup impl Implementation * @{
*/
/* ************************************* * Includes & Memory related functions
***************************************/ #ifdefined(XXH_NO_STREAM) /* nothing */ #elifdefined(XXH_NO_STDLIB)
/* When requesting to disable any mention of stdlib, * the library loses the ability to invoked malloc / free. * In practice, it means that functions like `XXH*_createState()` * will always fail, and return NULL. * This flag is useful in situations where * xxhash.h is integrated into some kernel, embedded or limited environment * without access to dynamic allocation.
*/
/* * Modify the local functions below should you wish to use * different memory routines for malloc() and free()
*/ #include <stdlib.h>
/*! * @internal * @brief Modify this function to use a different routine than malloc().
*/ static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
/*! * @internal * @brief Modify this function to use a different routine than free().
*/ staticvoid XXH_free(void* p) { free(p); }
#endif/* XXH_NO_STDLIB */
#include <string.h>
/*! * @internal * @brief Modify this function to use a different routine than memcpy().
*/ staticvoid* XXH_memcpy(void* dest, constvoid* src, size_t size)
{ return memcpy(dest,src,size);
}
#include <limits.h> /* ULLONG_MAX */
/* ************************************* * Compiler Specific Options
***************************************/ #ifdef _MSC_VER /* Visual Studio warning fix */ # pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */ #endif
/* ************************************* * Debug
***************************************/ /*! * @ingroup tuning * @def XXH_DEBUGLEVEL * @brief Sets the debugging level. * * XXH_DEBUGLEVEL is expected to be defined externally, typically via the * compiler's command line options. The value must be a number.
*/ #ifndef XXH_DEBUGLEVEL # ifdef DEBUGLEVEL /* backwards compat */ # define XXH_DEBUGLEVEL DEBUGLEVEL # else # define XXH_DEBUGLEVEL 0 # endif #endif
#if (XXH_DEBUGLEVEL>=1) # include <assert.h> /* note: can still be disabled with NDEBUG */ # define XXH_ASSERT(c) assert(c) #else # ifdefined(__INTEL_COMPILER) # define XXH_ASSERT(c) XXH_ASSUME((unsignedchar) (c)) # else # define XXH_ASSERT(c) XXH_ASSUME(c) # endif #endif
/*! * @internal * @def XXH_COMPILER_GUARD(var) * @brief Used to prevent unwanted optimizations for @p var. * * It uses an empty GCC inline assembly statement with a register constraint * which forces @p var into a general purpose register (eg eax, ebx, ecx * on x86) and marks it as modified. * * This is used in a few places to avoid unwanted autovectorization (e.g. * XXH32_round()). All vectorization we want is explicit via intrinsics, * and _usually_ isn't wanted elsewhere. * * We also use it to prevent unwanted constant folding for AArch64 in * XXH3_initCustomSecret_scalar().
*/ #ifdefined(__GNUC__) || defined(__clang__) # define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var)) #else # define XXH_COMPILER_GUARD(var) ((void)0) #endif
/* Specifically for NEON vectors which use the "w" constraint, on
* Clang. */ #ifdefined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__) # define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var)) #else # define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0) #endif
#ifdef XXH_OLD_NAMES # 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" # define BYTE xxh_u8 # define U8 xxh_u8 # define U32 xxh_u32 #endif
/* *** Memory access *** */
/*! * @internal * @fn xxh_u32 XXH_read32(const void* ptr) * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness. * * Affected by @ref XXH_FORCE_MEMORY_ACCESS. * * @param ptr The pointer to read from. * @return The 32-bit native endian integer from the bytes at @p ptr.
*/
/*! * @internal * @fn xxh_u32 XXH_readLE32(const void* ptr) * @brief Reads an unaligned 32-bit little endian integer from @p ptr. * * Affected by @ref XXH_FORCE_MEMORY_ACCESS. * * @param ptr The pointer to read from. * @return The 32-bit little endian integer from the bytes at @p ptr.
*/
/*! * @internal * @fn xxh_u32 XXH_readBE32(const void* ptr) * @brief Reads an unaligned 32-bit big endian integer from @p ptr. * * Affected by @ref XXH_FORCE_MEMORY_ACCESS. * * @param ptr The pointer to read from. * @return The 32-bit big endian integer from the bytes at @p ptr.
*/
/*! * @internal * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align) * @brief Like @ref XXH_readLE32(), but has an option for aligned reads. * * Affected by @ref XXH_FORCE_MEMORY_ACCESS. * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is * always @ref XXH_alignment::XXH_unaligned. * * @param ptr The pointer to read from. * @param align Whether @p ptr is aligned. * @pre * If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte * aligned. * @return The 32-bit little endian integer from the bytes at @p ptr.
*/
#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) /* * Manual byteshift. Best for old compilers which don't inline memcpy. * We actually directly use XXH_readLE32 and XXH_readBE32.
*/ #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
/* * Force direct memory access. Only works on CPU which support unaligned memory * access in hardware.
*/ static xxh_u32 XXH_read32(constvoid* memPtr) { return *(const xxh_u32*) memPtr; }
/* * __attribute__((aligned(1))) is supported by gcc and clang. Originally the * documentation claimed that it only increased the alignment, but actually it * can decrease it on gcc, clang, and icc: * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502, * https://gcc.godbolt.org/z/xYez1j67Y.
*/ #ifdef XXH_OLD_NAMES typedefunion { xxh_u32 u32; } __attribute__((packed)) unalign; #endif static xxh_u32 XXH_read32(constvoid* ptr)
{ typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32; return *((const xxh_unalign32*)ptr);
}
/*! * @ingroup tuning * @def XXH_CPU_LITTLE_ENDIAN * @brief Whether the target is little endian. * * Defined to 1 if the target is little endian, or 0 if it is big endian. * It can be defined externally, for example on the compiler command line. * * If it is not defined, * a runtime check (which is usually constant folded) is used instead. * * @note * This is not necessarily defined to an integer constant. * * @see XXH_isLittleEndian() for the runtime check.
*/ #ifndef XXH_CPU_LITTLE_ENDIAN /* * Try to detect endianness automatically, to avoid the nonstandard behavior * in `XXH_isLittleEndian()`
*/ # ifdefined(_WIN32) /* Windows is always little endian */ \
|| defined(__LITTLE_ENDIAN__) \
|| (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) # define XXH_CPU_LITTLE_ENDIAN 1 # elif defined(__BIG_ENDIAN__) \
|| (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) # define XXH_CPU_LITTLE_ENDIAN 0 # else /*! * @internal * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN. * * Most compilers will constant fold this.
*/ staticint XXH_isLittleEndian(void)
{ /* * Portable and well-defined behavior. * Don't use static: it is detrimental to performance.
*/ constunion { xxh_u32 u; xxh_u8 c[4]; } one = { 1 }; return one.c[0];
} # define XXH_CPU_LITTLE_ENDIAN XXH_isLittleEndian() # endif #endif
/* * C23 and future versions have standard "unreachable()". * Once it has been implemented reliably we can add it as an * additional case: * * ``` * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) * # include <stddef.h> * # ifdef unreachable * # define XXH_UNREACHABLE() unreachable() * # endif * #endif * ``` * * Note C++23 also has std::unreachable() which can be detected * as follows: * ``` * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L) * # include <utility> * # define XXH_UNREACHABLE() std::unreachable() * #endif * ``` * NB: `__cpp_lib_unreachable` is defined in the `<version>` header. * We don't use that as including `<utility>` in `extern "C"` blocks * doesn't work on GCC12
*/
/*! * @internal * @brief Enum to indicate whether a pointer is aligned.
*/ typedefenum {
XXH_aligned, /*!< Aligned */
XXH_unaligned /*!< Possibly unaligned */
} XXH_alignment;
/* * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. * * This is ideal for older compilers which don't inline memcpy.
*/ #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
/*! * @internal * @brief Normal stripe processing routine. * * This shuffles the bits so that any bit from @p input impacts several bits in * @p acc. * * @param acc The accumulator lane. * @param input The stripe of input to mix. * @return The mixed accumulator lane.
*/ static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
{
acc += input * XXH_PRIME32_2;
acc = XXH_rotl32(acc, 13);
acc *= XXH_PRIME32_1; #if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE) /* * UGLY HACK: * A compiler fence is the only thing that prevents GCC and Clang from * autovectorizing the XXH32 loop (pragmas and attributes don't work for some * reason) without globally disabling SSE4.1. * * The reason we want to avoid vectorization is because despite working on * 4 integers at a time, there are multiple factors slowing XXH32 down on * SSE4: * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on * newer chips!) making it slightly slower to multiply four integers at * once compared to four integers independently. Even when pmulld was * fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE * just to multiply unless doing a long operation. * * - Four instructions are required to rotate, * movqda tmp, v // not required with VEX encoding * pslld tmp, 13 // tmp <<= 13 * psrld v, 19 // x >>= 19 * por v, tmp // x |= tmp * compared to one for scalar: * roll v, 13 // reliably fast across the board * shldl v, v, 13 // Sandy Bridge and later prefer this for some reason * * - Instruction level parallelism is actually more beneficial here because * the SIMD actually serializes this operation: While v1 is rotating, v2 * can load data, while v3 can multiply. SSE forces them to operate * together. * * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing * the loop. NEON is only faster on the A53, and with the newer cores, it is less * than half the speed. * * Additionally, this is used on WASM SIMD128 because it JITs to the same * SIMD instructions and has the same issue.
*/
XXH_COMPILER_GUARD(acc); #endif return acc;
}
/*! * @internal * @brief Mixes all bits to finalize the hash. * * The final mix ensures that all input bits have a chance to impact any bit in * the output digest, resulting in an unbiased distribution. * * @param hash The hash to avalanche. * @return The avalanched hash.
*/ static xxh_u32 XXH32_avalanche(xxh_u32 hash)
{
hash ^= hash >> 15;
hash *= XXH_PRIME32_2;
hash ^= hash >> 13;
hash *= XXH_PRIME32_3;
hash ^= hash >> 16; return hash;
}
/*! * @internal * @brief Processes the last 0-15 bytes of @p ptr. * * There may be up to 15 bytes remaining to consume from the input. * This final stage will digest them to ensure that all input bytes are present * in the final mix. * * @param hash The hash to finalize. * @param ptr The pointer to the remaining input. * @param len The remaining length, modulo 16. * @param align Whether @p ptr is aligned. * @return The finalized hash. * @see XXH64_finalize().
*/ static XXH_PUREF xxh_u32
XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
{ #define XXH_PROCESS1 do { \
hash += (*ptr++) * XXH_PRIME32_5; \
hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1; \
} while (0)
/*! * @ingroup XXH32_family * The default return values from XXH functions are unsigned 32 and 64 bit * integers. * * The canonical representation uses big endian convention, the same convention * as human-readable numbers (large digits first). * * This way, hash values can be written into a file or buffer, remaining * comparable across different systems. * * The following functions allow transformation of hash values to and from their * canonical format.
*/
XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
{
XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t)); if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
XXH_memcpy(dst, &hash, sizeof(*dst));
} /*! @ingroup XXH32_family */
XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
{ return XXH_readBE32(src);
}
#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) /* * Manual byteshift. Best for old compilers which don't inline memcpy. * We actually directly use XXH_readLE64 and XXH_readBE64.
*/ #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */ static xxh_u64 XXH_read64(constvoid* memPtr)
{ return *(const xxh_u64*) memPtr;
}
/* * __attribute__((aligned(1))) is supported by gcc and clang. Originally the * documentation claimed that it only increased the alignment, but actually it * can decrease it on gcc, clang, and icc: * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502, * https://gcc.godbolt.org/z/xYez1j67Y.
*/ #ifdef XXH_OLD_NAMES typedefunion { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64; #endif static xxh_u64 XXH_read64(constvoid* ptr)
{ typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64; return *((const xxh_unalign64*)ptr);
}
/*! * @internal * @brief Processes the last 0-31 bytes of @p ptr. * * There may be up to 31 bytes remaining to consume from the input. * This final stage will digest them to ensure that all input bytes are present * in the final mix. * * @param hash The hash to finalize. * @param ptr The pointer to the remaining input. * @param len The remaining length, modulo 32. * @param align Whether @p ptr is aligned. * @return The finalized hash * @see XXH32_finalize().
*/ static XXH_PUREF xxh_u64
XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
{ if (ptr==NULL) XXH_ASSERT(len == 0);
len &= 31; while (len >= 8) {
xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
ptr += 8;
hash ^= k1;
hash = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
len -= 8;
} if (len >= 4) {
hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
ptr += 4;
hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
len -= 4;
} while (len > 0) {
hash ^= (*ptr++) * XXH_PRIME64_5;
hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
--len;
} return XXH64_avalanche(hash);
}
/* * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while * remaining a true 64-bit/128-bit hash function. * * This is done by prioritizing a subset of 64-bit operations that can be * emulated without too many steps on the average 32-bit machine. * * For example, these two lines seem similar, and run equally fast on 64-bit: * * xxh_u64 x; * x ^= (x >> 47); // good * x ^= (x >> 13); // bad * * However, to a 32-bit machine, there is a major difference. * * x ^= (x >> 47) looks like this: * * x.lo ^= (x.hi >> (47 - 32)); * * while x ^= (x >> 13) looks like this: * * // note: funnel shifts are not usually cheap. * x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13)); * x.hi ^= (x.hi >> 13); * * The first one is significantly faster than the second, simply because the * shift is larger than 32. This means: * - All the bits we need are in the upper 32 bits, so we can ignore the lower * 32 bits in the shift. * - The shift result will always fit in the lower 32 bits, and therefore, * we can ignore the upper 32 bits in the xor. * * Thanks to this optimization, XXH3 only requires these features to be efficient: * * - Usable unaligned access * - A 32-bit or 64-bit ALU * - If 32-bit, a decent ADC instruction * - A 32 or 64-bit multiply with a 64-bit result * - For the 128-bit variant, a decent byteswap helps short inputs. * * The first two are already required by XXH32, and almost all 32-bit and 64-bit * platforms which can run XXH32 can run XXH3 efficiently. * * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one * notable exception. * * First of all, Thumb-1 lacks support for the UMULL instruction which * performs the important long multiply. This means numerous __aeabi_lmul * calls. * * Second of all, the 8 functional registers are just not enough. * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need * Lo registers, and this shuffling results in thousands more MOVs than A32. * * A32 and T32 don't have this limitation. They can access all 14 registers, * do a 32->64 multiply with UMULL, and the flexible operand allowing free * shifts is helpful, too. * * Therefore, we do a quick sanity check. * * If compiling Thumb-1 for a target which supports ARM instructions, we will * emit a warning, as it is not a "sane" platform to compile for. * * Usually, if this happens, it is because of an accident and you probably need * to specify -march, as you likely meant to compile for a newer architecture. * * Credit: large sections of the vectorial and asm source code paths * have been contributed by @easyaspi314
*/ #ifdefined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM) # warning "XXH3 is highly inefficient without ARM or Thumb-2." #endif
#ifdef XXH_DOXYGEN /*! * @ingroup tuning * @brief Overrides the vectorization implementation chosen for XXH3. * * Can be defined to 0 to disable SIMD or any of the values mentioned in * @ref XXH_VECTOR_TYPE. * * If this is not defined, it uses predefined macros to determine the best * implementation.
*/ # define XXH_VECTOR XXH_SCALAR /*! * @ingroup tuning * @brief Possible values for @ref XXH_VECTOR. * * Note that these are actually implemented as macros. * * If this is not defined, it is detected automatically. * internal macro XXH_X86DISPATCH overrides this.
*/ enum XXH_VECTOR_TYPE /* fake enum */ {
XXH_SCALAR = 0, /*!< Portable scalar version */
XXH_SSE2 = 1, /*!< * SSE2 for Pentium 4, Opteron, all x86_64. * * @note SSE2 is also guaranteed on Windows 10, macOS, and * Android x86.
*/
XXH_AVX2 = 2, /*!< AVX2 for Haswell and Bulldozer */
XXH_AVX512 = 3, /*!< AVX512 for Skylake and Icelake */
XXH_NEON = 4, /*!< * NEON for most ARMv7-A, all AArch64, and WASM SIMD128 * via the SIMDeverywhere polyfill provided with the * Emscripten SDK.
*/
XXH_VSX = 5, /*!< VSX and ZVector for POWER8/z13 (64-bit) */
XXH_SVE = 6, /*!< SVE for some ARMv8-A and ARMv9-A */
}; /*! * @ingroup tuning * @brief Selects the minimum alignment for XXH3's accumulators. * * When using SIMD, this should match the alignment required for said vector * type, so, for example, 32 for AVX2. * * Default: Auto detected.
*/ # define XXH_ACC_ALIGN 8 #endif
/* * UGLY HACK: * GCC usually generates the best code with -O3 for xxHash. * * However, when targeting AVX2, it is overzealous in its unrolling resulting * in code roughly 3/4 the speed of Clang. * * There are other issues, such as GCC splitting _mm256_loadu_si256 into * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which * only applies to Sandy and Ivy Bridge... which don't even support AVX2. * * That is why when compiling the AVX2 version, it is recommended to use either * -O2 -mavx2 -march=haswell * or * -O2 -mavx2 -mno-avx256-split-unaligned-load * for decent performance, or to use Clang instead. * * Fortunately, we can control the first one with a pragma that forces GCC into * -O2, but the other one we can't control without "failed to inline always * inline function due to target mismatch" warnings.
*/ #if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
&& defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
&& defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */ # pragma GCC push_options # pragma GCC optimize("-O2") #endif
#if XXH_VECTOR == XXH_NEON
/* * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3 * optimizes out the entire hashLong loop because of the aliasing violation. * * However, GCC is also inefficient at load-store optimization with vld1q/vst1q, * so the only option is to mark it as aliasing.
*/ typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
/*! * @internal * @brief `vld1q_u64` but faster and alignment-safe. * * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86). * * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it * prohibits load-store optimizations. Therefore, a direct dereference is used. * * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe * unaligned load.
*/ #ifdefined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(voidconst* ptr) /* silence -Wcast-align */
{ return *(xxh_aliasing_uint64x2_t const *)ptr;
} #else
XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(voidconst* ptr)
{ return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
} #endif
/*! * @internal * @brief `vmlal_u32` on low and high halves of a vector. * * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32` * with `vmlal_u32`.
*/ #ifdefined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
XXH_FORCE_INLINE uint64x2_t
XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
{ /* Inline assembly is the only way */
__asm__("umlal %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs)); return acc;
}
XXH_FORCE_INLINE uint64x2_t
XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
{ /* This intrinsic works as expected */ return vmlal_high_u32(acc, lhs, rhs);
} #else /* Portable intrinsic versions */
XXH_FORCE_INLINE uint64x2_t
XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
{ return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
} /*! @copydoc XXH_vmlal_low_u32
* Assume the compiler converts this to vmlal_high_u32 on aarch64 */
XXH_FORCE_INLINE uint64x2_t
XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
{ return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
} #endif
/*! * @ingroup tuning * @brief Controls the NEON to scalar ratio for XXH3 * * This can be set to 2, 4, 6, or 8. * * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used. * * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU * bandwidth. * * This is even more noticeable on the more advanced cores like the Cortex-A76 which * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once. * * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes * and 2 scalar lanes, which is chosen by default. * * This does not apply to Apple processors or 32-bit processors, which run better with * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes. * * This change benefits CPUs with large micro-op buffers without negatively affecting * most other CPUs: * * | Chipset | Dispatch type | NEON only | 6:2 hybrid | Diff. | * |:----------------------|:--------------------|----------:|-----------:|------:| * | Snapdragon 730 (A76) | 2 NEON/8 micro-ops | 8.8 GB/s | 10.1 GB/s | ~16% | * | Snapdragon 835 (A73) | 2 NEON/3 micro-ops | 5.1 GB/s | 5.3 GB/s | ~5% | * | Marvell PXA1928 (A53) | In-order dual-issue | 1.9 GB/s | 1.9 GB/s | 0% | * | Apple M1 | 4 NEON/8 micro-ops | 37.3 GB/s | 36.1 GB/s | ~-3% | * * It also seems to fix some bad codegen on GCC, making it almost as fast as clang. * * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning * it effectively becomes worse 4. * * @see XXH3_accumulate_512_neon()
*/ # ifndef XXH3_NEON_LANES # if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
&& !defined(__APPLE__) && XXH_SIZE_OPT <= 0 # define XXH3_NEON_LANES 6 # else # define XXH3_NEON_LANES XXH_ACC_NB # endif # endif #endif/* XXH_VECTOR == XXH_NEON */
/* * VSX and Z Vector helpers. * * This is very messy, and any pull requests to clean this up are welcome. * * There are a lot of problems with supporting VSX and s390x, due to * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
*/ #if XXH_VECTOR == XXH_VSX /* Annoyingly, these headers _may_ define three macros: `bool`, `vector`, * and `pixel`. This is a problem for obvious reasons. * * These keywords are unnecessary; the spec literally says they are * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd * after including the header. *
* We use pragma push_macro/pop_macro to keep the namespace clean. */ # pragma push_macro("bool") # pragma push_macro("vector") # pragma push_macro("pixel") /* silence potential macro redefined warnings */ # undef bool # undef vector # undef pixel
# ifdefined(__s390x__) # include <s390intrin.h> # else # include <altivec.h> # endif
/* Restore the original macro values, if applicable. */ # pragma pop_macro("pixel") # pragma pop_macro("vector") # pragma pop_macro("bool")
/*! * Performs an unaligned vector load and byte swaps it on big endian.
*/
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(constvoid *ptr)
{
xxh_u64x2 ret;
XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2)); # if XXH_VSX_BE
ret = XXH_vec_revb(ret); # endif return ret;
}
/* * vec_mulo and vec_mule are very problematic intrinsics on PowerPC * * These intrinsics weren't added until GCC 8, despite existing for a while, * and they are endian dependent. Also, their meaning swap depending on version.
* */ # ifdefined(__s390x__) /* s390x is always big endian, no issue on this platform */ # define XXH_vec_mulo vec_mulo # define XXH_vec_mule vec_mule # elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__) /* Clang has a better way to control this, we can just use the builtin which doesn't swap. */ /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */ # define XXH_vec_mulo __builtin_altivec_vmulouw # define XXH_vec_mule __builtin_altivec_vmuleuw # else /* gcc needs inline assembly */ /* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
{
xxh_u64x2 result;
__asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); return result;
}
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
{
xxh_u64x2 result;
__asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); return result;
} # endif /* XXH_vec_mulo, XXH_vec_mule */ #endif/* XXH_VECTOR == XXH_VSX */
#ifdef XXH_DOXYGEN /*! * @brief Calculates a 32-bit to 64-bit long multiply. * * Implemented as a macro. * * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't * need to (but it shouldn't need to anyways, it is about 7 instructions to do * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we * use that instead of the normal method. * * If you are compiling for platforms like Thumb-1 and don't have a better option, * you may also want to write your own long multiply routine here. * * @param x, y Numbers to be multiplied * @return 64-bit product of the low 32 bits of @p x and @p y.
*/
XXH_FORCE_INLINE xxh_u64
XXH_mult32to64(xxh_u64 x, xxh_u64 y)
{ return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
} #elifdefined(_MSC_VER) && defined(_M_IX86) # define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y)) #else /* * Downcast + upcast is usually better than masking on older compilers like * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers. * * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
*/ # define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y)) #endif
/*! * @brief Calculates a 64->128-bit long multiply. * * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar * version. * * @param lhs , rhs The 64-bit integers to be multiplied * @return The 128-bit result represented in an @ref XXH128_hash_t.
*/ static XXH128_hash_t
XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
{ /* * GCC/Clang __uint128_t method. * * On most 64-bit targets, GCC and Clang define a __uint128_t type. * This is usually the best way as it usually uses a native long 64-bit * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64. * * Usually. * * Despite being a 32-bit platform, Clang (and emscripten) define this type * despite not having the arithmetic for it. This results in a laggy * compiler builtin call which calculates a full 128-bit multiply. * In that case it is best to use the portable one. * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
*/ #if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
&& defined(__SIZEOF_INT128__) \
|| (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
/* * MSVC for ARM64's __umulh method. * * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
*/ #elifdefined(_M_ARM64) || defined(_M_ARM64EC)
/*! * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it. * * The reason for the separate function is to prevent passing too many structs * around by value. This will hopefully inline the multiply, but we don't force it. * * @param lhs , rhs The 64-bit integers to multiply * @return The low 64 bits of the product XOR'd by the high 64 bits. * @see XXH_mult64to128()
*/ static xxh_u64
XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
{
XXH128_hash_t product = XXH_mult64to128(lhs, rhs); return product.low64 ^ product.high64;
}
/*! Seems to produce slightly better code on GCC for some reason. */
XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
{
XXH_ASSERT(0 <= shift && shift < 64); return v64 ^ (v64 >> shift);
}
/* * This is a fast avalanche stage, * suitable when input bits are already partially mixed
*/ static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
{
h64 = XXH_xorshift64(h64, 37);
h64 *= PRIME_MX1;
h64 = XXH_xorshift64(h64, 32); return h64;
}
/* * This is a stronger avalanche, * inspired by Pelle Evensen's rrmxmx * preferable when input has not been previously mixed
*/ static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
{ /* this mix is inspired by Pelle Evensen's rrmxmx */
h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
h64 *= PRIME_MX2;
h64 ^= (h64 >> 35) + len ;
h64 *= PRIME_MX2; return XXH_xorshift64(h64, 28);
}
/* ========================================== * Short keys * ========================================== * One of the shortcomings of XXH32 and XXH64 was that their performance was * sub-optimal on short lengths. It used an iterative algorithm which strongly * favored lengths that were a multiple of 4 or 8. * * Instead of iterating over individual inputs, we use a set of single shot * functions which piece together a range of lengths and operate in constant time. * * Additionally, the number of multiplies has been significantly reduced. This * reduces latency, especially when emulating 64-bit multiplies on 32-bit. * * Depending on the platform, this may or may not be faster than XXH32, but it * is almost guaranteed to be faster than XXH64.
*/
/* * At very short lengths, there isn't enough input to fully hide secrets, or use * the entire secret. * * There is also only a limited amount of mixing we can do before significantly * impacting performance. * * Therefore, we use different sections of the secret and always mix two secret * samples with an XOR. This should have no effect on performance on the * seedless or withSeed variants because everything _should_ be constant folded * by modern compilers. * * The XOR mixing hides individual parts of the secret and increases entropy. * * This adds an extra layer of strength for custom secrets.
*/
XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
XXH_ASSERT(input != NULL);
XXH_ASSERT(1 <= len && len <= 3);
XXH_ASSERT(secret != NULL); /* * len = 1: combined = { input[0], 0x01, input[0], input[0] } * len = 2: combined = { input[1], 0x02, input[0], input[1] } * len = 3: combined = { input[2], 0x03, input[0], input[1] }
*/
{ xxh_u8 const c1 = input[0];
xxh_u8 const c2 = input[len >> 1];
xxh_u8 const c3 = input[len - 1];
xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24)
| ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
xxh_u64 const keyed = (xxh_u64)combined ^ bitflip; return XXH64_avalanche(keyed);
}
}
/* * DISCLAIMER: There are known *seed-dependent* multicollisions here due to * multiplication by zero, affecting hashes of lengths 17 to 240. * * However, they are very unlikely. * * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all * unseeded non-cryptographic hashes, it does not attempt to defend itself * against specially crafted inputs, only random inputs. * * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes * cancelling out the secret is taken an arbitrary number of times (addressed * in XXH3_accumulate_512), this collision is very unlikely with random inputs * and/or proper seeding: * * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a * function that is only called up to 16 times per hash with up to 240 bytes of * input. * * This is not too bad for a non-cryptographic hash function, especially with * only 64 bit outputs. * * The 128-bit variant (which trades some speed for strength) is NOT affected * by this, although it is always a good idea to use a proper seed if you care * about strength.
*/
XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input, const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
{ #ifdefined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
&& defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \
&& !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */ /* * UGLY HACK: * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in * slower code. * * By forcing seed64 into a register, we disrupt the cost model and * cause it to scalarize. See `XXH32_round()` * * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600, * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on * GCC 9.2, despite both emitting scalar code. * * GCC generates much better scalar code than Clang for the rest of XXH3, * which is why finding a more optimal codepath is an interest.
*/
XXH_COMPILER_GUARD(seed64); #endif
{ xxh_u64 const input_lo = XXH_readLE64(input);
xxh_u64 const input_hi = XXH_readLE64(input+8); return XXH3_mul128_fold64(
input_lo ^ (XXH_readLE64(secret) + seed64),
input_hi ^ (XXH_readLE64(secret+8) - seed64)
);
}
}
/* For mid range keys, XXH3 uses a Mum-hash variant. */
XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len, const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
XXH64_hash_t seed)
{
XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
XXH_ASSERT(16 < len && len <= 128);
{ xxh_u64 acc = len * XXH_PRIME64_1;
xxh_u64 acc_end; unsignedintconst nbRounds = (unsignedint)len / 16; unsignedint i;
XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); for (i=0; i<8; i++) {
acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
} /* last bytes */
acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
XXH_ASSERT(nbRounds >= 8);
acc = XXH3_avalanche(acc); #ifdefined(__clang__) /* Clang */ \
&& (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
&& !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ /* * UGLY HACK: * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86. * In everywhere else, it uses scalar code. * * For 64->128-bit multiplies, even if the NEON was 100% optimal, it * would still be slower than UMAAL (see XXH_mult64to128). * * Unfortunately, Clang doesn't handle the long multiplies properly and * converts them to the nonexistent "vmulq_u64" intrinsic, which is then * scalarized into an ugly mess of VMOV.32 instructions. * * This mess is difficult to avoid without turning autovectorization * off completely, but they are usually relatively minor and/or not * worth it to fix. * * This loop is the easiest to fix, as unlike XXH32, this pragma * _actually works_ because it is a loop vectorization instead of an * SLP vectorization.
*/ #pragma clang loop vectorize(disable) #endif for (i=8 ; i < nbRounds; i++) { /* * Prevents clang for unrolling the acc loop and interleaving with this one.
*/
XXH_COMPILER_GUARD(acc);
acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
} return XXH3_avalanche(acc + acc_end);
}
}
/* ======= Long Keys ======= */
#define XXH_STRIPE_LEN 64 #define XXH_SECRET_CONSUME_RATE 8 /* nb of secret bytes consumed at each accumulation */ #define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))
/* * These macros are to generate an XXH3_accumulate() function. * The two arguments select the name suffix and target attribute. * * The name of this symbol is XXH3_accumulate_<name>() and it calls * XXH3_accumulate_512_<name>(). * * It may be useful to hand implement this function if the compiler fails to * optimize the inline function.
*/ #define XXH3_ACCUMULATE_TEMPLATE(name) \ void \
XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc, \ const xxh_u8* XXH_RESTRICT input, \ const xxh_u8* XXH_RESTRICT secret, \
size_t nbStripes) \
{ \
size_t n; \ for (n = 0; n < nbStripes; n++ ) { \ const xxh_u8* const in = input + n*XXH_STRIPE_LEN; \
XXH_PREFETCH(in + XXH_PREFETCH_DIST); \
XXH3_accumulate_512_##name( \
acc, \
in, \
secret + n*XXH_SECRET_CONSUME_RATE); \
} \
}
/* Several intrinsic functions below are supposed to accept __int64 as argument, * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ . * However, several environments do not define __int64 type, * requiring a workaround.
*/ #if !defined (__VMS) \
&& (defined (__cplusplus) \
|| (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) typedef int64_t xxh_i64; #else /* the following type must have a width of 64-bit */ typedeflonglong xxh_i64; #endif
/* * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized. * * It is a hardened version of UMAC, based off of FARSH's implementation. * * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD * implementations, and it is ridiculously fast. * * We harden it by mixing the original input to the accumulators as well as the product. * * This means that in the (relatively likely) case of a multiply by zero, the * original input is preserved. * * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve * cross-pollination, as otherwise the upper and lower halves would be * essentially independent. * * This doesn't matter on 64-bit hashes since they all get merged together in * the end, so we skip the extra step. * * Both XXH3_64bits and XXH3_128bits use this subroutine.
*/
/* * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing. * * Multiplication isn't perfect, as explained by Google in HighwayHash: * * // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to * // varying degrees. In descending order of goodness, bytes * // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32. * // As expected, the upper and lower bytes are much worse. * * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291 * * Since our algorithm uses a pseudorandom secret to add some variance into the * mix, we don't need to (or want to) mix as often or as much as HighwayHash does. * * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid * extraction. * * Both XXH3_64bits and XXH3_128bits use this subroutine.
*/
# ifdefined(__GNUC__) || defined(__clang__) /* * On GCC & Clang, marking 'dest' as modified will cause the compiler: * - do not extract the secret from sse registers in the internal loop * - use less common registers, and avoid pushing these reg into stack
*/
XXH_COMPILER_GUARD(dest); # endif
XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
XXH_ASSERT(((size_t)dest & 31) == 0);
XXH_FORCE_INLINE XXH_TARGET_SSE2 void
XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc, constvoid* XXH_RESTRICT input, constvoid* XXH_RESTRICT secret)
{ /* SSE2 is just a half-scale version of the AVX2 version. */
XXH_ASSERT((((size_t)acc) & 15) == 0);
{ __m128i* const xacc = (__m128i *) acc; /* Unaligned. This is mainly for pointer arithmetic, and because
* _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ const __m128i* const xinput = (const __m128i *) input; /* Unaligned. This is mainly for pointer arithmetic, and because
* _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ const __m128i* const xsecret = (const __m128i *) secret;
# ifdefined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900 /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */
XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
__m128i const seed = _mm_load_si128((__m128i const*)seed64x2); # else
__m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64); # endif int i;
constvoid* const src16 = XXH3_kSecret;
__m128i* dst16 = (__m128i*) customSecret; # ifdefined(__GNUC__) || defined(__clang__) /* * On GCC & Clang, marking 'dest' as modified will cause the compiler: * - do not extract the secret from sse registers in the internal loop * - use less common registers, and avoid pushing these reg into stack
*/
XXH_COMPILER_GUARD(dst16); # endif
XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */
XXH_ASSERT(((size_t)dst16 & 15) == 0);
for (i=0; i < nbRounds; ++i) {
dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
} }
}
/*! * @internal * @brief The bulk processing loop for NEON and WASM SIMD128. * * The NEON code path is actually partially scalar when running on AArch64. This * is to optimize the pipelining and can have up to 15% speedup depending on the * CPU, and it also mitigates some GCC codegen issues. * * @see XXH3_NEON_LANES for configuring this and details about this optimization. * * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit * integers instead of the other platforms which mask full 64-bit vectors, * so the setup is more complicated than just shifting right. * * Additionally, there is an optimization for 4 lanes at once noted below. * * Since, as stated, the most optimal amount of lanes for Cortexes is 6, * there needs to be *three* versions of the accumulate operation used * for the remaining 2 lanes. * * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap * nearly perfectly.
*/
XXH_FORCE_INLINE void
XXH3_accumulate_512_neon( void* XXH_RESTRICT acc, constvoid* XXH_RESTRICT input, constvoid* XXH_RESTRICT secret)
{
XXH_ASSERT((((size_t)acc) & 15) == 0);
XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
{ /* GCC for darwin arm64 does not like aliasing here */
xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc; /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
uint8_t const* xinput = (const uint8_t *) input;
uint8_t const* xsecret = (const uint8_t *) secret;
size_t i; #ifdef __wasm_simd128__ /* * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret * is constant propagated, which results in it converting it to this * inside the loop: * * a = v128.load(XXH3_kSecret + 0 + $secret_offset, offset = 0) * b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0) * ... * * This requires a full 32-bit address immediate (and therefore a 6 byte * instruction) as well as an add for each offset. * * Putting an asm guard prevents it from folding (at the cost of losing * the alignment hint), and uses the free offset in `v128.load` instead * of adding secret_offset each time which overall reduces code size by * about a kilobyte and improves performance.
*/
XXH_COMPILER_GUARD(xsecret); #endif /* Scalar lanes use the normal scalarRound routine */ for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
XXH3_scalarRound(acc, input, secret, i);
}
i = 0; /* 4 NEON lanes at a time. */ for (; i+1 < XXH3_NEON_LANES / 2; i+=2) { /* data_vec = xinput[i]; */
uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput + (i * 16));
uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput + ((i+1) * 16)); /* key_vec = xsecret[i]; */
uint64x2_t key_vec_1 = XXH_vld1q_u64(xsecret + (i * 16));
uint64x2_t key_vec_2 = XXH_vld1q_u64(xsecret + ((i+1) * 16)); /* data_swap = swap(data_vec) */
uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1); /* data_key = data_vec ^ key_vec; */
uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
/* * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to * get one vector with the low 32 bits of each lane, and one vector * with the high 32 bits of each lane. * * The intrinsic returns a double vector because the original ARMv7-a * instruction modified both arguments in place. AArch64 and SIMD128 emit * two instructions from this intrinsic. * * [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ] * [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
*/
uint32x4x2_t unzipped = vuzpq_u32(
vreinterpretq_u32_u64(data_key_1),
vreinterpretq_u32_u64(data_key_2)
); /* data_key_lo = data_key & 0xFFFFFFFF */
uint32x4_t data_key_lo = unzipped.val[0]; /* data_key_hi = data_key >> 32 */
uint32x4_t data_key_hi = unzipped.val[1]; /* * Then, we can split the vectors horizontally and multiply which, as for most * widening intrinsics, have a variant that works on both high half vectors * for free on AArch64. A similar instruction is available on SIMD128. * * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
*/
uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi); /* * Clang reorders * a += b * c; // umlal swap.2d, dkl.2s, dkh.2s * c += a; // add acc.2d, acc.2d, swap.2d * to * c += a; // add acc.2d, acc.2d, swap.2d * c += b * c; // umlal acc.2d, dkl.2s, dkh.2s * * While it would make sense in theory since the addition is faster, * for reasons likely related to umlal being limited to certain NEON * pipelines, this is worse. A compiler guard fixes this.
*/
XXH_COMPILER_GUARD_CLANG_NEON(sum_1);
XXH_COMPILER_GUARD_CLANG_NEON(sum_2); /* xacc[i] = acc_vec + sum; */
xacc[i] = vaddq_u64(xacc[i], sum_1);
xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
} /* Operate on the remaining NEON lanes 2 at a time. */ for (; i < XXH3_NEON_LANES / 2; i++) { /* data_vec = xinput[i]; */
uint64x2_t data_vec = XXH_vld1q_u64(xinput + (i * 16)); /* key_vec = xsecret[i]; */
uint64x2_t key_vec = XXH_vld1q_u64(xsecret + (i * 16)); /* acc_vec_2 = swap(data_vec) */
uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1); /* data_key = data_vec ^ key_vec; */
uint64x2_t data_key = veorq_u64(data_vec, key_vec); /* For two lanes, just use VMOVN and VSHRN. */ /* data_key_lo = data_key & 0xFFFFFFFF; */
uint32x2_t data_key_lo = vmovn_u64(data_key); /* data_key_hi = data_key >> 32; */
uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32); /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi); /* Same Clang workaround as before */
XXH_COMPILER_GUARD_CLANG_NEON(sum); /* xacc[i] = acc_vec + sum; */
xacc[i] = vaddq_u64 (xacc[i], sum);
}
}
}
XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
/* AArch64 uses both scalar and neon at the same time */ for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
XXH3_scalarScrambleRound(acc, secret, i);
} for (i=0; i < XXH3_NEON_LANES / 2; i++) { /* xacc[i] ^= (xacc[i] >> 47); */
uint64x2_t acc_vec = xacc[i];
uint64x2_t shifted = vshrq_n_u64(acc_vec, 47);
uint64x2_t data_vec = veorq_u64(acc_vec, shifted);
/* xacc[i] ^= xsecret[i]; */
uint64x2_t key_vec = XXH_vld1q_u64(xsecret + (i * 16));
uint64x2_t data_key = veorq_u64(data_vec, key_vec); /* xacc[i] *= XXH_PRIME32_1 */ #ifdef __wasm_simd128__ /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */
xacc[i] = data_key * XXH_PRIME32_1; #else /* * Expanded version with portable NEON intrinsics * * lo(x) * lo(y) + (hi(x) * lo(y) << 32) * * prod_hi = hi(data_key) * lo(prime) << 32 * * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits * and avoid the shift.
*/
uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi); /* Extract low bits for vmlal_u32 */
uint32x2_t data_key_lo = vmovn_u64(data_key); /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */
xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo); #endif
}
}
} #endif
#ifdefined(__aarch64__) && (defined(__GNUC__) || defined(__clang__)) /* * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they * emit an excess mask and a full 64-bit multiply-add (MADD X-form). * * While this might not seem like much, as AArch64 is a 64-bit architecture, only * big Cortex designs have a full 64-bit multiplier. * * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit * multiplies expand to 2-3 multiplies in microcode. This has a major penalty * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline. * * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does * not have this penalty and does the mask automatically.
*/
XXH_FORCE_INLINE xxh_u64
XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
{
xxh_u64 ret; /* note: %x = 64-bit register, %w = 32-bit register */
__asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc)); return ret;
} #else
XXH_FORCE_INLINE xxh_u64
XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
{ return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc;
} #endif
/*! * @internal * @brief Scalar round for @ref XXH3_accumulate_512_scalar(). * * This is extracted to its own function because the NEON path uses a combination * of NEON and scalar.
*/
XXH_FORCE_INLINE void
XXH3_scalarRound(void* XXH_RESTRICT acc, voidconst* XXH_RESTRICT input, voidconst* XXH_RESTRICT secret,
size_t lane)
{
xxh_u64* xacc = (xxh_u64*) acc;
xxh_u8 const* xinput = (xxh_u8 const*) input;
xxh_u8 const* xsecret = (xxh_u8 const*) secret;
XXH_ASSERT(lane < XXH_ACC_NB);
XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
{
xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8);
xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8);
xacc[lane ^ 1] += data_val; /* swap adjacent lanes */
xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]);
}
}
/*! * @internal * @brief Processes a 64 byte block of data using the scalar path.
*/
XXH_FORCE_INLINE void
XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc, constvoid* XXH_RESTRICT input, constvoid* XXH_RESTRICT secret)
{
size_t i; /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */ #ifdefined(__GNUC__) && !defined(__clang__) \
&& (defined(__arm__) || defined(__thumb2__)) \
&& defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
&& XXH_SIZE_OPT <= 0 # pragma GCC unroll 8 #endif for (i=0; i < XXH_ACC_NB; i++) {
XXH3_scalarRound(acc, input, secret, i);
}
}
XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
/*! * @internal * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar(). * * This is extracted to its own function because the NEON path uses a combination * of NEON and scalar.
*/
XXH_FORCE_INLINE void
XXH3_scalarScrambleRound(void* XXH_RESTRICT acc, voidconst* XXH_RESTRICT secret,
size_t lane)
{
xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */ const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */
XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
XXH_ASSERT(lane < XXH_ACC_NB);
{
xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8);
xxh_u64 acc64 = xacc[lane];
acc64 = XXH_xorshift64(acc64, 47);
acc64 ^= key64;
acc64 *= XXH_PRIME32_1;
xacc[lane] = acc64;
}
}
/*! * @internal * @brief Scrambles the accumulators after a large chunk has been read
*/
XXH_FORCE_INLINE void
XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, constvoid* XXH_RESTRICT secret)
{
size_t i; for (i=0; i < XXH_ACC_NB; i++) {
XXH3_scalarScrambleRound(acc, secret, i);
}
}
XXH_FORCE_INLINE void
XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
{ /* * We need a separate pointer for the hack below, * which requires a non-const pointer. * Any decent compiler will optimize this out otherwise.
*/ const xxh_u8* kSecretPtr = XXH3_kSecret;
XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
#ifdefined(__GNUC__) && defined(__aarch64__) /* * UGLY HACK: * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are * placed sequentially, in order, at the top of the unrolled loop. * * While MOVK is great for generating constants (2 cycles for a 64-bit * constant compared to 4 cycles for LDR), it fights for bandwidth with * the arithmetic instructions. * * I L S * MOVK * MOVK * MOVK * MOVK * ADD * SUB STR * STR * By forcing loads from memory (as the asm line causes the compiler to assume * that XXH3_kSecretPtr has been changed), the pipelines are used more * efficiently: * I L S * LDR * ADD LDR * SUB STR * STR * * See XXH3_NEON_LANES for details on the pipsline. * * XXH3_64bits_withSeed, len == 256, Snapdragon 835 * without hack: 2654.4 MB/s * with hack: 3202.9 MB/s
*/
XXH_COMPILER_GUARD(kSecretPtr); #endif
{ intconst nbRounds = XXH_SECRET_DEFAULT_SIZE / 16; int i; for (i=0; i < nbRounds; i++) { /* * The asm hack causes the compiler to assume that kSecretPtr aliases with * customSecret, and on aarch64, this prevented LDP from merging two * loads together for free. Putting the loads together before the stores * properly generates LDP.
*/
xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i) + seed64;
xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
XXH_writeLE64((xxh_u8*)customSecret + 16*i, lo);
XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
} }
}
/* last stripe */
{ const xxh_u8* const p = input + len - XXH_STRIPE_LEN; #define XXH_SECRET_LASTACC_START 7 /* not aligned on 8, last secret is different from acc & ;scrambler */
XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
} }
}
/* converge into final hash */
XXH_STATIC_ASSERT(sizeof(acc) == 64); /* do not align on 8, so that the secret is different from the accumulator */ #define XXH_SECRET_MERGEACCS_START 11
XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1);
}
/* * It's important for performance to transmit secret's size (when it's static) * so that the compiler can properly optimize the vectorized loop. * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set. * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE * breaks -Og, this is XXH_NO_INLINE.
*/
XXH3_WITH_SECRET_INLINE XXH64_hash_t
XXH3_hashLong_64b_withSecret(constvoid* XXH_RESTRICT input, size_t len,
XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
{
(void)seed64; return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
}
/* * It's preferable for performance that XXH3_hashLong is not inlined, * as it results in a smaller function for small data, easier to the instruction cache. * Note that inside this no_inline function, we do inline the internal loop, * and provide a statically defined secret size to allow optimization of vector loop.
*/
XXH_NO_INLINE XXH_PUREF XXH64_hash_t
XXH3_hashLong_64b_default(constvoid* XXH_RESTRICT input, size_t len,
XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
{
(void)seed64; (void)secret; (void)secretLen; return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
}
/* * XXH3_hashLong_64b_withSeed(): * Generate a custom key based on alteration of default XXH3_kSecret with the seed, * and then use this key for long mode hashing. * * This operation is decently fast but nonetheless costs a little bit of time. * Try to avoid it whenever possible (typically when seed==0). * * It's important for performance that XXH3_hashLong is not inlined. Not sure * why (uop cache maybe?), but the difference is large and easily measurable.
*/
XXH_FORCE_INLINE XXH64_hash_t
XXH3_hashLong_64b_withSeed_internal(constvoid* input, size_t len,
XXH64_hash_t seed,
XXH3_f_accumulate f_acc,
XXH3_f_scrambleAcc f_scramble,
XXH3_f_initCustomSecret f_initSec)
{ #if XXH_SIZE_OPT <= 0 if (seed == 0) return XXH3_hashLong_64b_internal(input, len,
XXH3_kSecret, sizeof(XXH3_kSecret),
f_acc, f_scramble); #endif
{ XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
f_initSec(secret, seed); return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
f_acc, f_scramble);
}
}
/* * It's important for performance that XXH3_hashLong is not inlined.
*/
XXH_NO_INLINE XXH64_hash_t
XXH3_hashLong_64b_withSeed(constvoid* XXH_RESTRICT input, size_t len,
XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
{
(void)secret; (void)secretLen; return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
}
XXH_FORCE_INLINE XXH64_hash_t
XXH3_64bits_internal(constvoid* XXH_RESTRICT input, size_t len,
XXH64_hash_t seed64, constvoid* XXH_RESTRICT secret, size_t secretLen,
XXH3_hashLong64_f f_hashLong)
{
XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN); /* * If an action is to be taken if `secretLen` condition is not respected, * it should be done here. * For now, it's a contract pre-condition. * Adding a check and a branch here would cost performance at every hash. * Also, note that function signature doesn't offer room to return an error.
*/ if (len <= 16) return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64); if (len <= 128) return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen);
}
/* === XXH3 streaming === */ #ifndef XXH_NO_STREAM /* * Malloc's a pointer that is always aligned to align. * * This must be freed with `XXH_alignedFree()`. * * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2 * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON. * * This underalignment previously caused a rather obvious crash which went * completely unnoticed due to XXH3_createState() not actually being tested. * Credit to RedSpah for noticing this bug. * * The alignment is done manually: Functions like posix_memalign or _mm_malloc * are avoided: To maintain portability, we would have to write a fallback * like this anyways, and besides, testing for the existence of library * functions without relying on external build tools is impossible. * * The method is simple: Overallocate, manually align, and store the offset * to the original behind the returned pointer. * * Align must be a power of 2 and 8 <= align <= 128.
*/ static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align)
{
XXH_ASSERT(align <= 128 && align >= 8); /* range check */
XXH_ASSERT((align & (align-1)) == 0); /* power of 2 */
XXH_ASSERT(s != 0 && s < (s + align)); /* empty/overflow */
{ /* Overallocate to make room for manual realignment and an offset byte */
xxh_u8* base = (xxh_u8*)XXH_malloc(s + align); if (base != NULL) { /* * Get the offset needed to align this pointer. * * Even if the returned pointer is aligned, there will always be * at least one byte to store the offset to the original pointer.
*/
size_t offset = align - ((size_t)base & (align - 1)); /* base % align */ /* Add the offset for the now-aligned pointer */
xxh_u8* ptr = base + offset;
XXH_ASSERT((size_t)ptr % align == 0);
/* Store the offset immediately before the returned pointer. */
ptr[-1] = (xxh_u8)offset; return ptr;
} return NULL;
}
} /* * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
*/ staticvoid XXH_alignedFree(void* p)
{ if (p != NULL) {
xxh_u8* ptr = (xxh_u8*)p; /* Get the offset byte we added in XXH_malloc. */
xxh_u8 offset = ptr[-1]; /* Free the original malloc'd pointer */
xxh_u8* base = ptr - offset;
XXH_free(base);
}
} /*! @ingroup XXH3_family */ /*! * @brief Allocate an @ref XXH3_state_t. * * Must be freed with XXH3_freeState(). * @return An allocated XXH3_state_t on success, `NULL` on failure.
*/
XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
{
XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64); if (state==NULL) return NULL;
XXH3_INITSTATE(state); return state;
}
/*! @ingroup XXH3_family */ /*! * @brief Frees an @ref XXH3_state_t. * * Must be allocated with XXH3_createState(). * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState(). * @return XXH_OK.
*/
XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
{
XXH_alignedFree(statePtr); return XXH_OK;
}
/*! * @internal * @brief Processes a large input for XXH3_update() and XXH3_digest_long(). * * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block. * * @param acc Pointer to the 8 accumulator lanes * @param nbStripesSoFarPtr In/out pointer to the number of leftover stripes in the block* * @param nbStripesPerBlock Number of stripes in a block * @param input Input pointer * @param nbStripes Number of stripes to process * @param secret Secret pointer * @param secretLimit Offset of the last block in @p secret * @param f_acc Pointer to an XXH3_accumulate implementation * @param f_scramble Pointer to an XXH3_scrambleAcc implementation * @return Pointer past the end of @p input after processing
*/
XXH_FORCE_INLINE const xxh_u8 *
XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock, const xxh_u8* XXH_RESTRICT input, size_t nbStripes, const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
XXH3_f_accumulate f_acc,
XXH3_f_scrambleAcc f_scramble)
{ const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE; /* Process full blocks */ if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) { /* Process the initial partial block... */
size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr;
do { /* Accumulate and scramble */
f_acc(acc, input, initialSecret, nbStripesThisIter);
f_scramble(acc, secret + secretLimit);
input += nbStripesThisIter * XXH_STRIPE_LEN;
nbStripes -= nbStripesThisIter; /* Then continue the loop with the full block size */
nbStripesThisIter = nbStripesPerBlock;
initialSecret = secret;
} while (nbStripes >= nbStripesPerBlock);
*nbStripesSoFarPtr = 0;
} /* Process a partial block */ if (nbStripes > 0) {
f_acc(acc, input, initialSecret, nbStripes);
input += nbStripes * XXH_STRIPE_LEN;
*nbStripesSoFarPtr += nbStripes;
} /* Return end pointer */ return input;
}
#ifndef XXH3_STREAM_USE_STACK # if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */ # define XXH3_STREAM_USE_STACK 1 # endif #endif /* * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
*/
XXH_FORCE_INLINE XXH_errorcode
XXH3_update(XXH3_state_t* XXH_RESTRICT const state, const xxh_u8* XXH_RESTRICT input, size_t len,
XXH3_f_accumulate f_acc,
XXH3_f_scrambleAcc f_scramble)
{ if (input==NULL) {
XXH_ASSERT(len == 0); return XXH_OK;
}
XXH_ASSERT(state != NULL);
{ const xxh_u8* const bEnd = input + len; constunsignedchar* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret; #ifdefined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1 /* For some reason, gcc and MSVC seem to suffer greatly * when operating accumulators directly into state. * Operating into stack space seems to enable proper optimization.
* clang, on the other hand, doesn't seem to need this trick */
XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
XXH_memcpy(acc, state->acc, sizeof(acc)); #else
xxh_u64* XXH_RESTRICT const acc = state->acc; #endif
state->totalLen += len;
XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
/* small input : just fill in tmp buffer */ if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
XXH_memcpy(state->buffer + state->bufferedSize, input, len);
state->bufferedSize += (XXH32_hash_t)len; return XXH_OK;
}
/* total input is now > XXH3_INTERNALBUFFER_SIZE */ #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN)
XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0); /* clean multiple */
/* ========================================== * XXH3 128 bits (a.k.a XXH128) * ========================================== * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant, * even without counting the significantly larger output size. * * For example, extra steps are taken to avoid the seed-dependent collisions * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B). * * This strength naturally comes at the cost of some speed, especially on short * lengths. Note that longer hashes are about as fast as the 64-bit version * due to it using only a slight modification of the 64-bit loop. * * XXH128 is also more oriented towards 64-bit machines. It is still extremely * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
*/
/* Shift len to the left to ensure it is even, this avoids even multiplies. */
XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2));
XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
XXH_ASSERT(input != NULL);
XXH_ASSERT(secret != NULL);
XXH_ASSERT(9 <= len && len <= 16);
{ xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
xxh_u64 const input_lo = XXH_readLE64(input);
xxh_u64 input_hi = XXH_readLE64(input + len - 8);
XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1); /* * Put len in the middle of m128 to ensure that the length gets mixed to * both the low and high bits in the 128x64 multiply below.
*/
m128.low64 += (xxh_u64)(len - 1) << 54;
input_hi ^= bitfliph; /* * Add the high 32 bits of input_hi to the high 32 bits of m128, then * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to * the high 64 bits of m128. * * The best approach to this operation is different on 32-bit and 64-bit.
*/ if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */ /* * 32-bit optimized version, which is more readable. * * On 32-bit, it removes an ADC and delays a dependency between the two * halves of m128.high64, but it generates an extra mask on 64-bit.
*/
m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2);
} else { /* * 64-bit optimized (albeit more confusing) version. * * Uses some properties of addition and multiplication to remove the mask: * * Let: * a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF) * b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000) * c = XXH_PRIME32_2 * * a + (b * c) * Inverse Property: x + y - x == y * a + (b * (1 + c - 1)) * Distributive Property: x * (y + z) == (x * y) + (x * z) * a + (b * 1) + (b * (c - 1)) * Identity Property: x * 1 == x * a + b + (b * (c - 1)) * * Substitute a, b, and c: * input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1)) * * Since input_hi.hi + input_hi.lo == input_hi, we get this: * input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
*/
m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1);
} /* m128 ^= XXH_swap64(m128 >> 64); */
m128.low64 ^= XXH_swap64(m128.high64);
{ XXH128_hash_t acc; unsigned i;
acc.low64 = len * XXH_PRIME64_1;
acc.high64 = 0; /* * We set as `i` as offset + 32. We do this so that unchanged * `len` can be used as upper bound. This reaches a sweet spot * where both x86 and aarch64 get simple agen and good codegen * for the loop.
*/ for (i = 32; i < 160; i += 32) {
acc = XXH128_mix32B(acc,
input + i - 32,
input + i - 16,
secret + i - 32,
seed);
}
acc.low64 = XXH3_avalanche(acc.low64);
acc.high64 = XXH3_avalanche(acc.high64); /* * NB: `i <= len` will duplicate the last 32-bytes if * len % 32 was zero. This is an unfortunate necessity to keep * the hash result stable.
*/ for (i=160; i <= len; i += 32) {
acc = XXH128_mix32B(acc,
input + i - 32,
input + i - 16,
secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
seed);
} /* last bytes */
acc = XXH128_mix32B(acc,
input + len - 16,
input + len - 32,
secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
(XXH64_hash_t)0 - seed);
/* * It's important for performance that XXH3_hashLong() is not inlined.
*/
XXH_NO_INLINE XXH_PUREF XXH128_hash_t
XXH3_hashLong_128b_default(constvoid* XXH_RESTRICT input, size_t len,
XXH64_hash_t seed64, constvoid* XXH_RESTRICT secret, size_t secretLen)
{
(void)seed64; (void)secret; (void)secretLen; return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
XXH3_accumulate, XXH3_scrambleAcc);
}
/* * It's important for performance to pass @p secretLen (when it's static) * to the compiler, so that it can properly optimize the vectorized loop. * * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE * breaks -Og, this is XXH_NO_INLINE.
*/
XXH3_WITH_SECRET_INLINE XXH128_hash_t
XXH3_hashLong_128b_withSecret(constvoid* XXH_RESTRICT input, size_t len,
XXH64_hash_t seed64, constvoid* XXH_RESTRICT secret, size_t secretLen)
{
(void)seed64; return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
XXH3_accumulate, XXH3_scrambleAcc);
}
XXH_FORCE_INLINE XXH128_hash_t
XXH3_128bits_internal(constvoid* input, size_t len,
XXH64_hash_t seed64, constvoid* XXH_RESTRICT secret, size_t secretLen,
XXH3_hashLong128_f f_hl128)
{
XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN); /* * If an action is to be taken if `secret` conditions are not respected, * it should be done here. * For now, it's a contract pre-condition. * Adding a check and a branch here would cost performance at every hash.
*/ if (len <= 16) return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64); if (len <= 128) return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); return f_hl128(input, len, seed64, secret, secretLen);
}
/* === XXH3 128-bit streaming === */ #ifndef XXH_NO_STREAM /* * All initialization and update functions are identical to 64-bit streaming variant. * The only difference is the finalization routine.
*/
/* return : 1 is equal, 0 if different */ /*! @ingroup XXH3_family */
XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
{ /* note : XXH128_hash_t is compact, it has no padding byte */ return !(memcmp(&h1, &h2, sizeof(h1)));
}
/* This prototype is compatible with stdlib's qsort(). * @return : >0 if *h128_1 > *h128_2 * <0 if *h128_1 < *h128_2
* =0 if *h128_1 == *h128_2 */ /*! @ingroup XXH3_family */
XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE constvoid* h128_1, XXH_NOESCAPE constvoid* h128_2)
{
XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2; intconst hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64); /* note : bets that, in most cases, hash values are different */ if (hcmp) return hcmp; return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
}
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.
