@nxtedition/rocksdb 8.1.3 → 8.1.5

package/deps/rocksdb/rocksdb/util/xxhash.h

@@ -20,7 +20,7 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
  * Header File
- * Copyright (C) 2012-2020 Yann Collet
+ * Copyright (C) 2012-2021 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
@@ -51,49 +51,142 @@
  *   - 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
  */
-/* TODO: update */
-/* Notice extracted from xxHash homepage:
-
-xxHash is an extremely fast hash algorithm, running at RAM speed limits.
-It also successfully passes all tests from the SMHasher suite.
-
-Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo @3GHz)
-
-Name            Speed       Q.Score   Author
-xxHash          5.4 GB/s     10
-CrapWow         3.2 GB/s      2       Andrew
-MurmurHash 3a   2.7 GB/s     10       Austin Appleby
-SpookyHash      2.0 GB/s     10       Bob Jenkins
-SBox            1.4 GB/s      9       Bret Mulvey
-Lookup3         1.2 GB/s      9       Bob Jenkins
-SuperFastHash   1.2 GB/s      1       Paul Hsieh
-CityHash64      1.05 GB/s    10       Pike & Alakuijala
-FNV             0.55 GB/s     5       Fowler, Noll, Vo
-CRC32           0.43 GB/s     9
-MD5-32          0.33 GB/s    10       Ronald L. Rivest
-SHA1-32         0.28 GB/s    10
-
-Q.Score is a measure of quality of the hash function.
-It depends on successfully passing SMHasher test set.
-10 is a perfect score.
-
-Note: SMHasher's CRC32 implementation is not the fastest one.
-Other speed-oriented implementations can be faster,
-especially in combination with PCLMUL instruction:
-https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html?showComment=1552696407071#c3490092340461170735
-
-A 64-bit version, named XXH64, is available since r35.
-It offers much better speed, but for 64-bit applications only.
-Name     Speed on 64 bits    Speed on 32 bits
-XXH64       13.8 GB/s            1.9 GB/s
-XXH32        6.8 GB/s            6.0 GB/s
-*/
 
 #if defined (__cplusplus)
 extern "C" {
@@ -103,21 +196,53 @@ extern "C" {
  *  INLINE mode
  ******************************/
 /*!
- * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+#ifdef XXH_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
+ *  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 */
@@ -140,29 +265,80 @@ extern "C" {
 
    /*
     * 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.
+    * 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.
     */
-#  ifdef XXH_NAMESPACE
-#    error "XXH_INLINE_ALL with XXH_NAMESPACE is not supported"
-     /*
-      * Note: Alternative: #undef all symbols (it's a pretty large list).
-      * Without #error: it compiles, but functions are actually not inlined.
-      */
-#  endif
+   /* 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
-    * still be renamed to avoid redeclaration.
+    * 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 is a more dispersed action.
-    * Meanwhile, renaming can be achieved in a single block
+    * 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_INLINE_ ## Id
+#  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)
@@ -181,21 +357,13 @@ extern "C" {
 #  undef XXHASH_H_STATIC_13879238742
 #endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */
 
-
-
 /* ****************************************************************
  *  Stable API
  *****************************************************************/
 #ifndef XXHASH_H_5627135585666179
 #define XXHASH_H_5627135585666179 1
 
-
-/*!
- * @defgroup public Public API
- * Contains details on the public xxHash functions.
- * @{
- */
-/* specific declaration modes for Windows */
+/*! @brief Marks a global symbol. */
 #if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
 #  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
 #    ifdef XXH_EXPORT
@@ -208,24 +376,6 @@ extern "C" {
 #  endif
 #endif
 
-#ifdef XXH_DOXYGEN
-/*!
- * @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 XXH_NAMESPACE to automatically prefix
- * any public symbol from xxhash library with the value of 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
-
 #ifdef XXH_NAMESPACE
 #  define XXH_CAT(A,B) A##B
 #  define XXH_NAME2(A,B) XXH_CAT(A,B)
@@ -254,23 +404,28 @@ extern "C" {
 #  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
 #  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
 #  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+#  define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed)
 #  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
 #  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
 #  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
 #  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
 #  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
 #  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
+#  define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed)
 #  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
 #  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
 #  define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
+#  define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed)
 /* XXH3_128bits */
 #  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
 #  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
 #  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
 #  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+#  define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed)
 #  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
 #  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
 #  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
+#  define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed)
 #  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
 #  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
 #  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
@@ -280,30 +435,64 @@ extern "C" {
 #endif
 
 
+/* *************************************
+*  Compiler specifics
+***************************************/
+
+/* specific declaration modes for Windows */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
+#endif
+
+#if defined (__GNUC__)
+# define XXH_CONSTF  __attribute__((const))
+# define XXH_PUREF   __attribute__((pure))
+# define XXH_MALLOCF __attribute__((malloc))
+#else
+# define XXH_CONSTF  /* disable */
+# define XXH_PUREF
+# define XXH_MALLOCF
+#endif
+
 /* *************************************
 *  Version
 ***************************************/
 #define XXH_VERSION_MAJOR    0
 #define XXH_VERSION_MINOR    8
 #define XXH_VERSION_RELEASE  1
+/*! @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 only useful when xxHash is compiled as a shared library, as it is
- * independent of the version defined in the header.
+ * 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 `XXH_VERSION_NUMBER` as of when the libray was compiled.
+ * @return @ref XXH_VERSION_NUMBER of the invoked library.
  */
-XXH_PUBLIC_API unsigned XXH_versionNumber (void);
+XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
 
 
 /* ****************************
-*  Definitions
+*  Common basic types
 ******************************/
 #include <stddef.h>   /* size_t */
-typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
+/*!
+ * @brief Exit code for the streaming API.
+ */
+typedef enum {
+    XXH_OK = 0, /*!< OK */
+    XXH_ERROR   /*!< Error */
+} XXH_errorcode;
 
 
 /*-**********************************************************************
@@ -316,39 +505,38 @@ typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
  * Not necessarily defined to `uint32_t` but functionally equivalent.
  */
 typedef uint32_t XXH32_hash_t;
+
 #elif !defined (__VMS) \
   && (defined (__cplusplus) \
   || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
 #   include <stdint.h>
     typedef uint32_t XXH32_hash_t;
+
 #else
 #   include <limits.h>
 #   if UINT_MAX == 0xFFFFFFFFUL
       typedef unsigned int XXH32_hash_t;
+#   elif ULONG_MAX == 0xFFFFFFFFUL
+      typedef unsigned long XXH32_hash_t;
 #   else
-#     if ULONG_MAX == 0xFFFFFFFFUL
-        typedef unsigned long XXH32_hash_t;
-#     else
-#       error "unsupported platform: need a 32-bit type"
-#     endif
+#     error "unsupported platform: need a 32-bit type"
 #   endif
 #endif
 
 /*!
  * @}
  *
- * @defgroup xxh32_family XXH32 family
+ * @defgroup XXH32_family XXH32 family
  * @ingroup public
  * Contains functions used in the classic 32-bit xxHash algorithm.
  *
  * @note
- *   XXH32 is considered rather weak by today's standards.
- *   The @ref xxh3_family provides competitive speed for both 32-bit and 64-bit
- *   systems, and offers true 64/128 bit hash results. It provides a superior
- *   level of dispersion, and greatly reduces the risks of collisions.
+ *   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
+ * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
+ * @see @ref XXH32_impl for implementation details
  * @{
  */
 
@@ -357,6 +545,8 @@ typedef uint32_t XXH32_hash_t;
  *
  * 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.
@@ -374,8 +564,9 @@ typedef uint32_t XXH32_hash_t;
  * @see
  *    XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version.
  */
-XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* 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.
@@ -398,32 +589,7 @@ XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_
  *
  * When done, release the state using `XXH*_freeState()`.
  *
- * Example code for incrementally hashing a file:
- * @code{.c}
- *    #include <stdio.h>
- *    #include <xxhash.h>
- *    #define BUFFER_SIZE 256
- *
- *    // Note: XXH64 and XXH3 use the same interface.
- *    XXH32_hash_t
- *    hashFile(FILE* stream)
- *    {
- *        XXH32_state_t* state;
- *        unsigned char buf[BUFFER_SIZE];
- *        size_t amt;
- *        XXH32_hash_t hash;
- *
- *        state = XXH32_createState();       // Create a state
- *        assert(state != NULL);             // Error check here
- *        XXH32_reset(state, 0xbaad5eed);    // Reset state with our seed
- *        while ((amt = fread(buf, 1, sizeof(buf), stream)) != 0) {
- *            XXH32_update(state, buf, amt); // Hash the file in chunks
- *        }
- *        hash = XXH32_digest(state);        // Finalize the hash
- *        XXH32_freeState(state);            // Clean up
- *        return hash;
- *    }
- * @endcode
+ * @see streaming_example at the top of @ref xxhash.h for an example.
  */
 
 /*!
@@ -440,7 +606,7 @@ typedef struct XXH32_state_s XXH32_state_t;
  * Must be freed with XXH32_freeState().
  * @return An allocated XXH32_state_t on success, `NULL` on failure.
  */
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
+XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
 /*!
  * @brief Frees an @ref XXH32_state_t.
  *
@@ -508,7 +674,8 @@ XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void*
  *
  * @return The calculated xxHash32 value from that state.
  */
-XXH_PUBLIC_API XXH32_hash_t  XXH32_digest (const XXH32_state_t* statePtr);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
 /*******   Canonical representation   *******/
 
@@ -559,7 +726,52 @@ XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t
  *
  * @return The converted hash.
  */
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+
+
+#ifdef __has_attribute
+# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
+#else
+# define XXH_HAS_ATTRIBUTE(x) 0
+#endif
+
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
+# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define XXH_HAS_C_ATTRIBUTE(x) 0
+#endif
+
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define XXH_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+
+/*
+ * 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
+
+/*
+ * 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
 
 
 /*!
@@ -598,18 +810,17 @@ typedef uint64_t XXH64_hash_t;
 /*!
  * @}
  *
- * @defgroup xxh64_family XXH64 family
+ * @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 a superior level of
- *   dispersion, and greatly reduces the risks of collisions.
+ *   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.
  *
@@ -633,32 +844,35 @@ typedef uint64_t XXH64_hash_t;
  * @see
  *    XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH64(const void* input, size_t length, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* 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.
  */
 typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
+XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
 XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dst_state, const XXH64_state_t* src_state);
-
-XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH64_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH64_digest (const XXH64_state_t* statePtr);
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
 
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 /*******   Canonical representation   *******/
 typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src);
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t 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
+ * @defgroup XXH3_family XXH3 family
  * @ingroup public
  * @{
  *
@@ -678,12 +892,14 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
  *
  * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
  * but does not require it.
- * Any 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.
+ * 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.
  *
  * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8,
- * ZVector and scalar targets. This can be controlled via the XXH_VECTOR macro.
+ * ZVector and scalar targets. This can be controlled via the @ref XXH_VECTOR
+ * macro. For the x86 family, an automatic 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,
@@ -699,24 +915,42 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
  *
  * The API supports one-shot hashing, streaming mode, and custom secrets.
  */
-
 /*-**********************************************************************
 *  XXH3 64-bit variant
 ************************************************************************/
 
-/* XXH3_64bits():
- * default 64-bit variant, using default secret and default seed of 0.
- * It's the fastest variant. */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* data, size_t len);
+/*!
+ * @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 const void* input, size_t length);
 
-/*
- * XXH3_64bits_withSeed():
- * This variant generates a custom secret on the fly
- * based on default secret altered using the `seed` value.
+/*!
+ * @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 XXH3_64bits().
+ *
+ * @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 XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
 
 /*!
  * The bare minimum size for a custom secret.
@@ -727,23 +961,29 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, X
  */
 #define XXH3_SECRET_SIZE_MIN 136
 
-/*
- * XXH3_64bits_withSecret():
+/*!
+ * @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 produced hash values depends on secret's entropy.
- * Technically, the secret must look like a bunch of random bytes.
+ * 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 unsure about the "randomness" of the blob of bytes,
- * consider relabelling it as a "custom seed" instead,
- * and employ "XXH3_generateSecret()" (see below)
- * to generate a high entropy secret derived from the custom seed.
+ * 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 XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
 
 
 /*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*
  * Streaming requires state maintenance.
  * This operation costs memory and CPU.
@@ -757,23 +997,23 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len,
  * @see XXH3_state_s for details.
  */
 typedef struct XXH3_state_s XXH3_state_t;
-XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
 XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
-XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state);
+XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
 
 /*
  * XXH3_64bits_reset():
  * Initialize with default parameters.
  * digest will be equivalent to `XXH3_64bits()`.
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
 /*
  * XXH3_64bits_reset_withSeed():
  * Generate a custom secret from `seed`, and store it into `statePtr`.
  * digest will be equivalent to `XXH3_64bits_withSeed()`.
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-/*
+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`,
@@ -782,10 +1022,11 @@ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr,
  * 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(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
 
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_digest (const XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+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 */
@@ -806,11 +1047,31 @@ typedef struct {
     XXH64_hash_t high64;  /*!< `value >> 64` */
 } XXH128_hash_t;
 
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+/*!
+ * @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 const void* 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 const void* 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 const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
 
 /*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*
  * Streaming requires state maintenance.
  * This operation costs memory and CPU.
@@ -823,12 +1084,13 @@ XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t le
  * All reset and streaming functions have same meaning as their 64-bit counterpart.
  */
 
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
 
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+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.
@@ -838,26 +1100,26 @@ XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr);
  * XXH128_isEqual():
  * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
  */
-XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
 
 /*!
- * XXH128_cmp():
- *
+ * @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
+ * @return: >0 if *h128_1  > *h128_2
+ *          =0 if *h128_1 == *h128_2
+ *          <0 if *h128_1  < *h128_2
  */
-XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2);
+XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
 
 
 /*******   Canonical representation   *******/
 typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
-XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash);
-XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src);
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
 
 
+#endif  /* !XXH_NO_XXH3 */
 #endif  /* XXH_NO_LONG_LONG */
 
 /*!
@@ -898,13 +1160,10 @@ XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t*
 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 v1;           /*!< First accumulator lane */
-   XXH32_hash_t v2;           /*!< Second accumulator lane */
-   XXH32_hash_t v3;           /*!< Third accumulator lane */
-   XXH32_hash_t v4;           /*!< Fourth accumulator lane */
+   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 or write to it, it may be removed. */
+   XXH32_hash_t reserved;     /*!< Reserved field. Do not read nor write to it. */
 };   /* typedef'd to XXH32_state_t */
 
 
@@ -924,19 +1183,21 @@ struct XXH32_state_s {
  */
 struct XXH64_state_s {
    XXH64_hash_t total_len;    /*!< Total length hashed. This is always 64-bit. */
-   XXH64_hash_t v1;           /*!< First accumulator lane */
-   XXH64_hash_t v2;           /*!< Second accumulator lane */
-   XXH64_hash_t v3;           /*!< Third accumulator lane */
-   XXH64_hash_t v4;           /*!< Fourth accumulator lane */
+   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, it may be removed. */
+   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it. */
 };   /* typedef'd to XXH64_state_t */
 
-#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)   /* C11+ */
+#ifndef XXH_NO_XXH3
+
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
 #  include <stdalign.h>
 #  define XXH_ALIGN(n)      alignas(n)
+#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */
+/* In C++ alignas() is a keyword */
+#  define XXH_ALIGN(n)      alignas(n)
 #elif defined(__GNUC__)
 #  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
 #elif defined(_MSC_VER)
@@ -947,6 +1208,7 @@ struct XXH64_state_s {
 
 /* 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
@@ -976,16 +1238,18 @@ struct XXH64_state_s {
  * @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. This allows fields to safely be changed.
+ * @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.
+ * @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 not access the members of this struct directly.
+ * Do never access the members of this struct directly.
  *
  * @see XXH3_INITSTATE() for stack initialization.
  * @see XXH3_createState(), XXH3_freeState().
@@ -993,14 +1257,14 @@ struct XXH64_state_s {
  */
 struct XXH3_state_s {
    XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
-       /*!< The 8 accumulators. Similar to `vN` in @ref XXH32_state_s::v1 and @ref XXH64_state_s */
+       /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
    XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
        /*!< Used to store a custom secret generated from a seed. */
    XXH_ALIGN_MEMBER(64, unsigned char 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 reserved32;
+   XXH32_hash_t useSeed;
        /*!< Reserved field. Needed for padding on 64-bit. */
    size_t nbStripesSoFar;
        /*!< Number or stripes processed. */
@@ -1036,45 +1300,156 @@ struct XXH3_state_s {
 #define XXH3_INITSTATE(XXH3_state_ptr)   { (XXH3_state_ptr)->seed = 0; }
 
 
+/*!
+ * simple alias to pre-selected XXH3_128bits variant
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* 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 `_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 XXH3_SECRET_DEFAULT_SIZE
- * into an already allocated buffer secretBuffer.
- * The generated secret is _always_ XXH_SECRET_DEFAULT_SIZE bytes long.
+ * 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.
- * Functions `XXH3_128bits_withSecret()`, `XXH3_64bits_withSecret()`,
- * `XXH3_128bits_reset_withSecret()` and `XXH3_64bits_reset_withSecret()`
+ * 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 very long for implementation reasons (>= XXH3_SECRET_SIZE_MIN)
+ * 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
- * this function can be used to generate a secret of proper quality.
+ * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
+ * be employed to ensure proper quality.
  *
- * customSeed can be anything. It can have any size, even small ones,
- * and its content can be anything, even stupidly "low entropy" source such as a bunch of zeroes.
- * The resulting `secret` will nonetheless provide all expected qualities.
+ * @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.
  *
- * Supplying NULL as the customSeed copies the default secret into `secretBuffer`.
- * When 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 void XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSeedSize);
-
+XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize);
 
-/* simple short-cut to pre-selected XXH3_128bits variant */
-XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed);
+/*!
+ * @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 const void* data, size_t len,
+                              XXH_NOESCAPE const void* 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 const void* input, size_t length,
+                               XXH_NOESCAPE const void* 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 const void* 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 const void* secret, size_t secretSize,
+                                     XXH64_hash_t seed64);
+#endif /* !XXH_NO_STREAM */
 
+#endif  /* !XXH_NO_XXH3 */
 #endif  /* XXH_NO_LONG_LONG */
 #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
 #  define XXH_IMPLEMENTATION
@@ -1128,7 +1503,7 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
 /*!
  * @brief Define this to disable 64-bit code.
  *
- * Useful if only using the @ref xxh32_family and you have a strict C90 compiler.
+ * 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 */
@@ -1151,7 +1526,7 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
  *     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__((packed))`
+ *  - `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,
@@ -1178,22 +1553,40 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
  *   care, as what works on one compiler/platform/optimization level may cause
  *   another to read garbage data or even crash.
  *
- * See https://stackoverflow.com/a/32095106/646947 for details.
+ * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
  *
  * Prefer these methods in priority order (0 > 3 > 1 > 2)
  */
 #  define XXH_FORCE_MEMORY_ACCESS 0
+
 /*!
- * @def XXH_ACCEPT_NULL_INPUT_POINTER
- * @brief Whether to add explicit `NULL` checks.
+ * @def XXH_SIZE_OPT
+ * @brief Controls how much xxHash optimizes for size.
  *
- * If the input pointer is `NULL` and the length is non-zero, xxHash's default
- * behavior is to dereference it, triggering a segfault.
+ * 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.
  *
- * When this macro is enabled, xxHash actively checks the input for a null pointer.
- * If it is, the result for null input pointers is the same as a zero-length input.
+ * 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_ACCEPT_NULL_INPUT_POINTER 0
+#  define XXH_SIZE_OPT 0
+
 /*!
  * @def XXH_FORCE_ALIGN_CHECK
  * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
@@ -1215,9 +1608,11 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
  *
  * 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,
+ * 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
@@ -1239,24 +1634,22 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
  * 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), optimizing for size (-Os, -Oz), or using
- * -fno-inline with GCC or Clang, this will automatically be defined.
+ * 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 XXH_REROLL
- * @brief Whether to reroll `XXH32_finalize` and `XXH64_finalize`.
+ * @def XXH32_ENDJMP
+ * @brief Whether to use a jump for `XXH32_finalize`.
  *
- * For performance, `XXH32_finalize` and `XXH64_finalize` use an unrolled loop
- * in the form of a switch statement.
+ * 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 is not always desirable, as it generates larger code, and depending on
- * the architecture, may even be slower
- *
- * This is automatically defined with `-Os`/`-Oz` on GCC and Clang.
+ * This setting is only possibly making a difference for very small inputs.
  */
-#  define XXH_REROLL 0
+#  define XXH32_ENDJMP 0
 
 /*!
  * @internal
@@ -1267,27 +1660,46 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
  */
 #  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 on armv7 and armv8 */
-#  if !defined(__clang__) && ( \
-    (defined(__INTEL_COMPILER) && !defined(_WIN32)) || \
-    (defined(__GNUC__) && (defined(__ARM_ARCH) && __ARM_ARCH >= 7)) )
+   /* 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. */
+#  if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED))
 #    define XXH_FORCE_MEMORY_ACCESS 1
 #  endif
 #endif
 
-#ifndef XXH_ACCEPT_NULL_INPUT_POINTER   /* can be defined externally */
-#  define XXH_ACCEPT_NULL_INPUT_POINTER 0
+#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 */
-#  if defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) \
-   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64) /* visual */
+   /* 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) \
+   || defined(__loongarch64) /* visual */
 #    define XXH_FORCE_ALIGN_CHECK 0
 #  else
 #    define XXH_FORCE_ALIGN_CHECK 1
@@ -1295,20 +1707,16 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
 #endif
 
 #ifndef XXH_NO_INLINE_HINTS
-#  if defined(__OPTIMIZE_SIZE__) /* -Os, -Oz */ \
-   || defined(__NO_INLINE__)     /* -O0, -fno-inline */
+#  if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__)  /* -O0, -fno-inline */
 #    define XXH_NO_INLINE_HINTS 1
 #  else
 #    define XXH_NO_INLINE_HINTS 0
 #  endif
 #endif
 
-#ifndef XXH_REROLL
-#  if defined(__OPTIMIZE_SIZE__)
-#    define XXH_REROLL 1
-#  else
-#    define XXH_REROLL 0
-#  endif
+#ifndef XXH32_ENDJMP
+/* generally preferable for performance */
+#  define XXH32_ENDJMP 0
 #endif
 
 /*!
@@ -1320,6 +1728,24 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
 /* *************************************
 *  Includes & Memory related functions
 ***************************************/
+#if defined(XXH_NO_STREAM)
+/* nothing */
+#elif defined(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.
+ */
+
+static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
+static void XXH_free(void* p) { (void)p; }
+
+#else
+
 /*
  * Modify the local functions below should you wish to use
  * different memory routines for malloc() and free()
@@ -1330,7 +1756,7 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t s
  * @internal
  * @brief Modify this function to use a different routine than malloc().
  */
-static void* XXH_malloc(size_t s) { return malloc(s); }
+static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
 
 /*!
  * @internal
@@ -1338,6 +1764,8 @@ static void* XXH_malloc(size_t s) { return malloc(s); }
  */
 static void XXH_free(void* p) { free(p); }
 
+#endif  /* XXH_NO_STDLIB */
+
 #include <string.h>
 
 /*!
@@ -1360,19 +1788,19 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size)
 #endif
 
 #if XXH_NO_INLINE_HINTS  /* disable inlining hints */
-#  if defined(__GNUC__)
+#  if defined(__GNUC__) || defined(__clang__)
 #    define XXH_FORCE_INLINE static __attribute__((unused))
 #  else
 #    define XXH_FORCE_INLINE static
 #  endif
 #  define XXH_NO_INLINE static
 /* enable inlining hints */
+#elif defined(__GNUC__) || defined(__clang__)
+#  define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused))
+#  define XXH_NO_INLINE static __attribute__((noinline))
 #elif defined(_MSC_VER)  /* Visual Studio */
 #  define XXH_FORCE_INLINE static __forceinline
 #  define XXH_NO_INLINE static __declspec(noinline)
-#elif defined(__GNUC__)
-#  define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused))
-#  define XXH_NO_INLINE static __attribute__((noinline))
 #elif defined (__cplusplus) \
   || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* C99 */
 #  define XXH_FORCE_INLINE static inline
@@ -1407,11 +1835,20 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size)
 #  include <assert.h>   /* note: can still be disabled with NDEBUG */
 #  define XXH_ASSERT(c)   assert(c)
 #else
-#  define XXH_ASSERT(c)   ((void)0)
+#  define XXH_ASSERT(c)   XXH_ASSUME(c)
 #endif
 
 /* note: use after variable declarations */
-#define XXH_STATIC_ASSERT(c)  do { enum { XXH_sa = 1/(int)(!!(c)) }; } while (0)
+#ifndef XXH_STATIC_ASSERT
+#  if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)    /* C11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
+#  elif defined(__cplusplus) && (__cplusplus >= 201103L)            /* C++11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
+#  else
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0)
+#  endif
+#  define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c)
+#endif
 
 /*!
  * @internal
@@ -1429,12 +1866,18 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size)
  * We also use it to prevent unwanted constant folding for AArch64 in
  * XXH3_initCustomSecret_scalar().
  */
-#ifdef __GNUC__
+#if defined(__GNUC__) || defined(__clang__)
 #  define XXH_COMPILER_GUARD(var) __asm__ __volatile__("" : "+r" (var))
 #else
 #  define XXH_COMPILER_GUARD(var) ((void)0)
 #endif
 
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_COMPILER_GUARD_W(var) __asm__ __volatile__("" : "+w" (var))
+#else
+#  define XXH_COMPILER_GUARD_W(var) ((void)0)
+#endif
+
 /* *************************************
 *  Basic Types
 ***************************************/
@@ -1522,30 +1965,31 @@ static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr;
 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
 
 /*
- * __pack instructions are safer but compiler specific, hence potentially
- * problematic for some compilers.
- *
- * Currently only defined for GCC and ICC.
+ * __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
 typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
 #endif
 static xxh_u32 XXH_read32(const void* ptr)
 {
-    typedef union { xxh_u32 u32; } __attribute__((packed)) xxh_unalign;
-    return ((const xxh_unalign*)ptr)->u32;
+    typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32;
+    return *((const xxh_unalign32*)ptr);
 }
 
 #else
 
 /*
  * Portable and safe solution. Generally efficient.
- * see: https://stackoverflow.com/a/32095106/646947
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
  */
 static xxh_u32 XXH_read32(const void* memPtr)
 {
     xxh_u32 val;
-    memcpy(&val, memPtr, sizeof(val));
+    XXH_memcpy(&val, memPtr, sizeof(val));
     return val;
 }
 
@@ -1553,6 +1997,7 @@ static xxh_u32 XXH_read32(const void* memPtr)
 
 
 /* ***   Endianness   *** */
+
 /*!
  * @ingroup tuning
  * @def XXH_CPU_LITTLE_ENDIAN
@@ -1561,8 +2006,8 @@ static xxh_u32 XXH_read32(const void* memPtr)
  * 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.
+ * 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.
@@ -1615,6 +2060,29 @@ static int XXH_isLittleEndian(void)
 #  define XXH_HAS_BUILTIN(x) 0
 #endif
 
+
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L)
+/* C23 and future versions have standard "unreachable()" */
+#  include <stddef.h>
+#  define XXH_UNREACHABLE() unreachable()
+
+#elif defined(__cplusplus) && (__cplusplus > 202002L)
+/* C++23 and future versions have std::unreachable() */
+#  include <utility> /* std::unreachable() */
+#  define XXH_UNREACHABLE() std::unreachable()
+
+#elif XXH_HAS_BUILTIN(__builtin_unreachable)
+#  define XXH_UNREACHABLE() __builtin_unreachable()
+
+#elif defined(_MSC_VER)
+#  define XXH_UNREACHABLE() __assume(0)
+
+#else
+#  define XXH_UNREACHABLE()
+#endif
+
+#define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
+
 /*!
  * @internal
  * @def XXH_rotl32(x,r)
@@ -1737,8 +2205,10 @@ XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
 *********************************************************************/
 /*!
  * @}
- * @defgroup xxh32_impl XXH32 implementation
+ * @defgroup XXH32_impl XXH32 implementation
  * @ingroup impl
+ *
+ * Details on the XXH32 implementation.
  * @{
  */
  /* #define instead of static const, to be used as initializers */
@@ -1818,17 +2288,17 @@ static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
  * 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 h32 The hash to avalanche.
+ * @param hash The hash to avalanche.
  * @return The avalanched hash.
  */
-static xxh_u32 XXH32_avalanche(xxh_u32 h32)
+static xxh_u32 XXH32_avalanche(xxh_u32 hash)
 {
-    h32 ^= h32 >> 15;
-    h32 *= XXH_PRIME32_2;
-    h32 ^= h32 >> 13;
-    h32 *= XXH_PRIME32_3;
-    h32 ^= h32 >> 16;
-    return(h32);
+    hash ^= hash >> 15;
+    hash *= XXH_PRIME32_2;
+    hash ^= hash >> 13;
+    hash *= XXH_PRIME32_3;
+    hash ^= hash >> 16;
+    return hash;
 }
 
 #define XXH_get32bits(p) XXH_readLE32_align(p, align)
@@ -1841,28 +2311,31 @@ static xxh_u32 XXH32_avalanche(xxh_u32 h32)
  * This final stage will digest them to ensure that all input bytes are present
  * in the final mix.
  *
- * @param h32 The hash to finalize.
+ * @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_u32
-XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
+static XXH_PUREF xxh_u32
+XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
 {
-#define XXH_PROCESS1 do {                           \
-    h32 += (*ptr++) * XXH_PRIME32_5;                \
-    h32 = XXH_rotl32(h32, 11) * XXH_PRIME32_1;      \
+#define XXH_PROCESS1 do {                             \
+    hash += (*ptr++) * XXH_PRIME32_5;                 \
+    hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1;      \
 } while (0)
 
-#define XXH_PROCESS4 do {                           \
-    h32 += XXH_get32bits(ptr) * XXH_PRIME32_3;      \
-    ptr += 4;                                   \
-    h32  = XXH_rotl32(h32, 17) * XXH_PRIME32_4;     \
+#define XXH_PROCESS4 do {                             \
+    hash += XXH_get32bits(ptr) * XXH_PRIME32_3;       \
+    ptr += 4;                                         \
+    hash  = XXH_rotl32(hash, 17) * XXH_PRIME32_4;     \
 } while (0)
 
-    /* Compact rerolled version */
-    if (XXH_REROLL) {
+    if (ptr==NULL) XXH_ASSERT(len == 0);
+
+    /* Compact rerolled version; generally faster */
+    if (!XXH32_ENDJMP) {
         len &= 15;
         while (len >= 4) {
             XXH_PROCESS4;
@@ -1872,49 +2345,49 @@ XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
             XXH_PROCESS1;
             --len;
         }
-        return XXH32_avalanche(h32);
+        return XXH32_avalanche(hash);
     } else {
          switch(len&15) /* or switch(bEnd - p) */ {
            case 12:      XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 8:       XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 4:       XXH_PROCESS4;
-                         return XXH32_avalanche(h32);
+                         return XXH32_avalanche(hash);
 
            case 13:      XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 9:       XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 5:       XXH_PROCESS4;
                          XXH_PROCESS1;
-                         return XXH32_avalanche(h32);
+                         return XXH32_avalanche(hash);
 
            case 14:      XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 10:      XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 6:       XXH_PROCESS4;
                          XXH_PROCESS1;
                          XXH_PROCESS1;
-                         return XXH32_avalanche(h32);
+                         return XXH32_avalanche(hash);
 
            case 15:      XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 11:      XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 7:       XXH_PROCESS4;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 3:       XXH_PROCESS1;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 2:       XXH_PROCESS1;
-                         FALLTHROUGH_INTENDED;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 1:       XXH_PROCESS1;
-                         FALLTHROUGH_INTENDED;
-           case 0:       return XXH32_avalanche(h32);
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 0:       return XXH32_avalanche(hash);
         }
         XXH_ASSERT(0);
-        return h32;   /* reaching this point is deemed impossible */
+        return hash;   /* reaching this point is deemed impossible */
     }
 }
 
@@ -1930,24 +2403,19 @@ XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
  * @internal
  * @brief The implementation for @ref XXH32().
  *
- * @param input, len, seed Directly passed from @ref XXH32().
+ * @param input , len , seed Directly passed from @ref XXH32().
  * @param align Whether @p input is aligned.
  * @return The calculated hash.
  */
-XXH_FORCE_INLINE xxh_u32
+XXH_FORCE_INLINE XXH_PUREF xxh_u32
 XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
 {
-    const xxh_u8* bEnd = input ? input + len : NULL;
     xxh_u32 h32;
 
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-    if (input==NULL) {
-        len=0;
-        bEnd=input=(const xxh_u8*)(size_t)16;
-    }
-#endif
+    if (input==NULL) XXH_ASSERT(len == 0);
 
     if (len>=16) {
+        const xxh_u8* const bEnd = input + len;
         const xxh_u8* const limit = bEnd - 15;
         xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
         xxh_u32 v2 = seed + XXH_PRIME32_2;
@@ -1972,10 +2440,10 @@ XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment
     return XXH32_finalize(h32, input, len&15, align);
 }
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
 {
-#if 0
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
     /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
     XXH32_state_t state;
     XXH32_reset(&state, seed);
@@ -1994,51 +2462,46 @@ XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t s
 
 
 /*******   Hash streaming   *******/
-/*!
- * @ingroup xxh32_family
- */
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
 {
     return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
 }
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
 {
     XXH_free(statePtr);
     return XXH_OK;
 }
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
 {
-    memcpy(dstState, srcState, sizeof(*dstState));
+    XXH_memcpy(dstState, srcState, sizeof(*dstState));
 }
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
 {
-    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state));
-    state.v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
-    state.v2 = seed + XXH_PRIME32_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - XXH_PRIME32_1;
-    /* do not write into reserved, planned to be removed in a future version */
-    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved));
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+    statePtr->v[1] = seed + XXH_PRIME32_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME32_1;
     return XXH_OK;
 }
 
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH_errorcode
 XXH32_update(XXH32_state_t* state, const void* input, size_t len)
 {
-    if (input==NULL)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
         return XXH_OK;
-#else
-        return XXH_ERROR;
-#endif
+    }
 
     {   const xxh_u8* p = (const xxh_u8*)input;
         const xxh_u8* const bEnd = p + len;
@@ -2055,35 +2518,25 @@ XXH32_update(XXH32_state_t* state, const void* input, size_t len)
         if (state->memsize) {   /* some data left from previous update */
             XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
             {   const xxh_u32* p32 = state->mem32;
-                state->v1 = XXH32_round(state->v1, XXH_readLE32(p32)); p32++;
-                state->v2 = XXH32_round(state->v2, XXH_readLE32(p32)); p32++;
-                state->v3 = XXH32_round(state->v3, XXH_readLE32(p32)); p32++;
-                state->v4 = XXH32_round(state->v4, XXH_readLE32(p32));
+                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++;
+                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++;
+                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++;
+                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32));
             }
             p += 16-state->memsize;
             state->memsize = 0;
         }
 
-        /* uintptr_t casts avoid UB or compiler warning on out-of-bounds
-         * pointer arithmetic */
-        if ((uintptr_t)p <= (uintptr_t)bEnd - 16) {
-            const uintptr_t limit = (uintptr_t)bEnd - 16;
-            xxh_u32 v1 = state->v1;
-            xxh_u32 v2 = state->v2;
-            xxh_u32 v3 = state->v3;
-            xxh_u32 v4 = state->v4;
+        if (p <= bEnd-16) {
+            const xxh_u8* const limit = bEnd - 16;
 
             do {
-                v1 = XXH32_round(v1, XXH_readLE32(p)); p+=4;
-                v2 = XXH32_round(v2, XXH_readLE32(p)); p+=4;
-                v3 = XXH32_round(v3, XXH_readLE32(p)); p+=4;
-                v4 = XXH32_round(v4, XXH_readLE32(p)); p+=4;
-            } while ((uintptr_t)p<=limit);
-
-            state->v1 = v1;
-            state->v2 = v2;
-            state->v3 = v3;
-            state->v4 = v4;
+                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4;
+                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4;
+                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4;
+                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4;
+            } while (p<=limit);
+
         }
 
         if (p < bEnd) {
@@ -2096,30 +2549,30 @@ XXH32_update(XXH32_state_t* state, const void* input, size_t len)
 }
 
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
 {
     xxh_u32 h32;
 
     if (state->large_len) {
-        h32 = XXH_rotl32(state->v1, 1)
-            + XXH_rotl32(state->v2, 7)
-            + XXH_rotl32(state->v3, 12)
-            + XXH_rotl32(state->v4, 18);
+        h32 = XXH_rotl32(state->v[0], 1)
+            + XXH_rotl32(state->v[1], 7)
+            + XXH_rotl32(state->v[2], 12)
+            + XXH_rotl32(state->v[3], 18);
     } else {
-        h32 = state->v3 /* == seed */ + XXH_PRIME32_5;
+        h32 = state->v[2] /* == seed */ + XXH_PRIME32_5;
     }
 
     h32 += state->total_len_32;
 
     return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
 }
-
+#endif /* !XXH_NO_STREAM */
 
 /*******   Canonical representation   *******/
 
 /*!
- * @ingroup xxh32_family
+ * @ingroup XXH32_family
  * The default return values from XXH functions are unsigned 32 and 64 bit
  * integers.
  *
@@ -2136,9 +2589,9 @@ XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t
 {
     XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
     if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
-    memcpy(dst, &hash, sizeof(*dst));
+    XXH_memcpy(dst, &hash, sizeof(*dst));
 }
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
 {
     return XXH_readBE32(src);
@@ -2179,30 +2632,31 @@ static xxh_u64 XXH_read64(const void* memPtr)
 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
 
 /*
- * __pack instructions are safer, but compiler specific, hence potentially
- * problematic for some compilers.
- *
- * Currently only defined for GCC and ICC.
+ * __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
 typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
 #endif
 static xxh_u64 XXH_read64(const void* ptr)
 {
-    typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) xxh_unalign64;
-    return ((const xxh_unalign64*)ptr)->u64;
+    typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64;
+    return *((const xxh_unalign64*)ptr);
 }
 
 #else
 
 /*
  * Portable and safe solution. Generally efficient.
- * see: https://stackoverflow.com/a/32095106/646947
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
  */
 static xxh_u64 XXH_read64(const void* memPtr)
 {
     xxh_u64 val;
-    memcpy(&val, memPtr, sizeof(val));
+    XXH_memcpy(&val, memPtr, sizeof(val));
     return val;
 }
 
@@ -2281,8 +2735,10 @@ XXH_readLE64_align(const void* ptr, XXH_alignment align)
 /*******   xxh64   *******/
 /*!
  * @}
- * @defgroup xxh64_impl XXH64 implementation
+ * @defgroup XXH64_impl XXH64 implementation
  * @ingroup impl
+ *
+ * Details on the XXH64 implementation.
  * @{
  */
 /* #define rather that static const, to be used as initializers */
@@ -2300,6 +2756,7 @@ XXH_readLE64_align(const void* ptr, XXH_alignment align)
 #  define PRIME64_5 XXH_PRIME64_5
 #endif
 
+/*! @copydoc XXH32_round */
 static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
 {
     acc += input * XXH_PRIME64_2;
@@ -2316,42 +2773,59 @@ static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
     return acc;
 }
 
-static xxh_u64 XXH64_avalanche(xxh_u64 h64)
+/*! @copydoc XXH32_avalanche */
+static xxh_u64 XXH64_avalanche(xxh_u64 hash)
 {
-    h64 ^= h64 >> 33;
-    h64 *= XXH_PRIME64_2;
-    h64 ^= h64 >> 29;
-    h64 *= XXH_PRIME64_3;
-    h64 ^= h64 >> 32;
-    return h64;
+    hash ^= hash >> 33;
+    hash *= XXH_PRIME64_2;
+    hash ^= hash >> 29;
+    hash *= XXH_PRIME64_3;
+    hash ^= hash >> 32;
+    return hash;
 }
 
 
 #define XXH_get64bits(p) XXH_readLE64_align(p, align)
 
-static xxh_u64
-XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align)
-{
-    len &= 31;
-    while (len >= 8) {
-        xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
-        ptr += 8;
-        h64 ^= k1;
-        h64  = XXH_rotl64(h64,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+/*!
+ * @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) {
-        h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+        hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
         ptr += 4;
-        h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+        hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
         len -= 4;
     }
     while (len > 0) {
-        h64 ^= (*ptr++) * XXH_PRIME64_5;
-        h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;
+        hash ^= (*ptr++) * XXH_PRIME64_5;
+        hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
         --len;
     }
-    return  XXH64_avalanche(h64);
+    return  XXH64_avalanche(hash);
 }
 
 #ifdef XXH_OLD_NAMES
@@ -2364,21 +2838,23 @@ XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align)
 #  undef XXH_PROCESS8_64
 #endif
 
-XXH_FORCE_INLINE xxh_u64
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH64().
+ *
+ * @param input , len , seed Directly passed from @ref XXH64().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u64
 XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
 {
-    const xxh_u8* bEnd = input ? input + len : NULL;
     xxh_u64 h64;
-
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-    if (input==NULL) {
-        len=0;
-        bEnd=input=(const xxh_u8*)(size_t)32;
-    }
-#endif
+    if (input==NULL) XXH_ASSERT(len == 0);
 
     if (len>=32) {
-        const xxh_u8* const limit = bEnd - 32;
+        const xxh_u8* const bEnd = input + len;
+        const xxh_u8* const limit = bEnd - 31;
         xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
         xxh_u64 v2 = seed + XXH_PRIME64_2;
         xxh_u64 v3 = seed + 0;
@@ -2389,7 +2865,7 @@ XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment
             v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
             v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
             v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
-        } while (input<=limit);
+        } while (input<limit);
 
         h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
         h64 = XXH64_mergeRound(h64, v1);
@@ -2407,10 +2883,10 @@ XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment
 }
 
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t seed)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
-#if 0
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
     /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
     XXH64_state_t state;
     XXH64_reset(&state, seed);
@@ -2428,49 +2904,45 @@ XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t s
 }
 
 /*******   Hash Streaming   *******/
-
-/*! @ingroup xxh64_family*/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH64_family*/
 XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
 {
     return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
 }
-/*! @ingroup xxh64_family */
+/*! @ingroup XXH64_family */
 XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
 {
     XXH_free(statePtr);
     return XXH_OK;
 }
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t* srcState)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
 {
-    memcpy(dstState, srcState, sizeof(*dstState));
+    XXH_memcpy(dstState, srcState, sizeof(*dstState));
 }
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
 {
-    XXH64_state_t state;   /* use a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state));
-    state.v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
-    state.v2 = seed + XXH_PRIME64_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - XXH_PRIME64_1;
-     /* do not write into reserved64, might be removed in a future version */
-    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved64));
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+    statePtr->v[1] = seed + XXH_PRIME64_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME64_1;
     return XXH_OK;
 }
 
-/*! @ingroup xxh64_family */
+/*! @ingroup XXH64_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH64_update (XXH64_state_t* state, const void* input, size_t len)
+XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
-    if (input==NULL)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
         return XXH_OK;
-#else
-        return XXH_ERROR;
-#endif
+    }
 
     {   const xxh_u8* p = (const xxh_u8*)input;
         const xxh_u8* const bEnd = p + len;
@@ -2485,34 +2957,24 @@ XXH64_update (XXH64_state_t* state, const void* input, size_t len)
 
         if (state->memsize) {   /* tmp buffer is full */
             XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
-            state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0));
-            state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1));
-            state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2));
-            state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3));
+            state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0));
+            state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1));
+            state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2));
+            state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3));
             p += 32 - state->memsize;
             state->memsize = 0;
         }
 
-        /* uintptr_t casts avoid UB or compiler warning on out-of-bounds
-         * pointer arithmetic */
-        if ((uintptr_t)p + 32 <= (uintptr_t)bEnd) {
-            const uintptr_t limit = (uintptr_t)bEnd - 32;
-            xxh_u64 v1 = state->v1;
-            xxh_u64 v2 = state->v2;
-            xxh_u64 v3 = state->v3;
-            xxh_u64 v4 = state->v4;
+        if (p+32 <= bEnd) {
+            const xxh_u8* const limit = bEnd - 32;
 
             do {
-                v1 = XXH64_round(v1, XXH_readLE64(p)); p+=8;
-                v2 = XXH64_round(v2, XXH_readLE64(p)); p+=8;
-                v3 = XXH64_round(v3, XXH_readLE64(p)); p+=8;
-                v4 = XXH64_round(v4, XXH_readLE64(p)); p+=8;
-            } while ((uintptr_t)p<=limit);
-
-            state->v1 = v1;
-            state->v2 = v2;
-            state->v3 = v3;
-            state->v4 = v4;
+                state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8;
+                state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8;
+                state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8;
+                state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8;
+            } while (p<=limit);
+
         }
 
         if (p < bEnd) {
@@ -2525,44 +2987,39 @@ XXH64_update (XXH64_state_t* state, const void* input, size_t len)
 }
 
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API XXH64_hash_t XXH64_digest(const XXH64_state_t* state)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
 {
     xxh_u64 h64;
 
     if (state->total_len >= 32) {
-        xxh_u64 const v1 = state->v1;
-        xxh_u64 const v2 = state->v2;
-        xxh_u64 const v3 = state->v3;
-        xxh_u64 const v4 = state->v4;
-
-        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
-        h64 = XXH64_mergeRound(h64, v1);
-        h64 = XXH64_mergeRound(h64, v2);
-        h64 = XXH64_mergeRound(h64, v3);
-        h64 = XXH64_mergeRound(h64, v4);
+        h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18);
+        h64 = XXH64_mergeRound(h64, state->v[0]);
+        h64 = XXH64_mergeRound(h64, state->v[1]);
+        h64 = XXH64_mergeRound(h64, state->v[2]);
+        h64 = XXH64_mergeRound(h64, state->v[3]);
     } else {
-        h64  = state->v3 /*seed*/ + XXH_PRIME64_5;
+        h64  = state->v[2] /*seed*/ + XXH_PRIME64_5;
     }
 
     h64 += (xxh_u64) state->total_len;
 
     return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
 }
-
+#endif /* !XXH_NO_STREAM */
 
 /******* Canonical representation   *******/
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash)
 {
     XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
     if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
-    memcpy(dst, &hash, sizeof(*dst));
+    XXH_memcpy(dst, &hash, sizeof(*dst));
 }
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
 {
     return XXH_readBE64(src);
 }
@@ -2575,7 +3032,7 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
 ************************************************************************ */
 /*!
  * @}
- * @defgroup xxh3_impl XXH3 implementation
+ * @defgroup XXH3_impl XXH3 implementation
  * @ingroup impl
  * @{
  */
@@ -2601,17 +3058,23 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
 #    define XXH_unlikely(x) (x)
 #endif
 
-#if defined(__GNUC__)
-#  if defined(__AVX2__)
-#    include <immintrin.h>
-#  elif defined(__SSE2__)
-#    include <emmintrin.h>
-#  elif defined(__ARM_NEON__) || defined(__ARM_NEON)
+#if defined(__GNUC__) || defined(__clang__)
+#  if defined(__ARM_FEATURE_SVE)
+#    include <arm_sve.h>
+#  elif defined(__ARM_NEON__) || defined(__ARM_NEON) \
+   || defined(__aarch64__)  || defined(_M_ARM) \
+   || defined(_M_ARM64)     || defined(_M_ARM64EC)
 #    define inline __inline__  /* circumvent a clang bug */
 #    include <arm_neon.h>
 #    undef inline
+#  elif defined(__AVX2__)
+#    include <immintrin.h>
+#  elif defined(__SSE2__)
+#    include <emmintrin.h>
 #  endif
-#elif defined(_MSC_VER)
+#endif
+
+#if defined(_MSC_VER)
 #  include <intrin.h>
 #endif
 
@@ -2725,12 +3188,13 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
     XXH_AVX512 = 3,  /*!< AVX512 for Skylake and Icelake */
     XXH_NEON   = 4,  /*!< NEON for most ARMv7-A and all AArch64 */
     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 reqired for said vector
+ * When using SIMD, this should match the alignment required for said vector
  * type, so, for example, 32 for AVX2.
  *
  * Default: Auto detected.
@@ -2746,20 +3210,26 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
 #  define XXH_AVX512 3
 #  define XXH_NEON   4
 #  define XXH_VSX    5
+#  define XXH_SVE    6
 #endif
 
 #ifndef XXH_VECTOR    /* can be defined on command line */
-#  if defined(__AVX512F__)
+#  if defined(__ARM_FEATURE_SVE)
+#    define XXH_VECTOR XXH_SVE
+#  elif ( \
+        defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
+     || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
+   ) && ( \
+        defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
+    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
+   )
+#    define XXH_VECTOR XXH_NEON
+#  elif defined(__AVX512F__)
 #    define XXH_VECTOR XXH_AVX512
 #  elif defined(__AVX2__)
 #    define XXH_VECTOR XXH_AVX2
 #  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
 #    define XXH_VECTOR XXH_SSE2
-#  elif defined(__GNUC__) /* msvc support maybe later */ \
-  && (defined(__ARM_NEON__) || defined(__ARM_NEON)) \
-  && (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \
-    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__))
-#    define XXH_VECTOR XXH_NEON
 #  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
      || (defined(__s390x__) && defined(__VEC__)) \
      && defined(__GNUC__) /* TODO: IBM XL */
@@ -2769,6 +3239,17 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
 #  endif
 #endif
 
+/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */
+#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE)
+#  ifdef _MSC_VER
+#    pragma warning(once : 4606)
+#  else
+#    warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead."
+#  endif
+#  undef XXH_VECTOR
+#  define XXH_VECTOR XXH_SCALAR
+#endif
+
 /*
  * Controls the alignment of the accumulator,
  * for compatibility with aligned vector loads, which are usually faster.
@@ -2788,12 +3269,16 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
 #     define XXH_ACC_ALIGN 16
 #  elif XXH_VECTOR == XXH_AVX512  /* avx512 */
 #     define XXH_ACC_ALIGN 64
+#  elif XXH_VECTOR == XXH_SVE   /* sve */
+#     define XXH_ACC_ALIGN 64
 #  endif
 #endif
 
 #if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
     || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
 #  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#elif XXH_VECTOR == XXH_SVE
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
 #else
 #  define XXH_SEC_ALIGN 8
 #endif
@@ -2821,7 +3306,7 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
  */
 #if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
   && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
 #  pragma GCC push_options
 #  pragma GCC optimize("-O2")
 #endif
@@ -2909,8 +3394,8 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
  * }
  */
 # if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \
-   && defined(__GNUC__) \
-   && !defined(__aarch64__) && !defined(__arm64__)
+   && (defined(__GNUC__) || defined(__clang__)) \
+   && (defined(__arm__) || defined(__thumb__) || defined(_M_ARM))
 #  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                              \
     do {                                                                                    \
       /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = upper D half */ \
@@ -2927,6 +3412,78 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
       (outHi) = vshrn_n_u64  ((in), 32);                                                  \
     } while (0)
 # endif
+
+/*!
+ * @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.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */
+{
+    return *(uint64x2_t const*)ptr;
+}
+#else
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
+{
+    return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
+}
+#endif
+/*!
+ * @ingroup tuning
+ * @brief Controls the NEON to scalar ratio for XXH3
+ *
+ * On AArch64 when not optimizing for size, XXH3 will run 6 lanes using NEON and
+ * 2 lanes on scalar by default (except on Apple platforms, as Apple CPUs benefit
+ * from only using NEON).
+ *
+ * This can be set to 2, 4, 6, or 8. ARMv7 will default to all 8 NEON lanes, as the
+ * emulated 64-bit arithmetic is too slow.
+ *
+ * Modern ARM CPUs are _very_ sensitive to how their pipelines are used.
+ *
+ * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but it can't
+ * have more than 2 NEON (F0/F1) micro-ops. 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 A76 which
+ * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
+ *
+ * Therefore, @ref XXH3_NEON_LANES lanes will be processed using NEON, and the
+ * remaining lanes will use scalar instructions. This improves the bandwidth
+ * and also gives the integer pipelines something to do besides twiddling loop
+ * counters and pointers.
+ *
+ * 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.
+ *
+ * @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 */
 
 /*
@@ -2938,23 +3495,33 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
  * 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
+
 #  if defined(__s390x__)
 #    include <s390intrin.h>
 #  else
-/* gcc's altivec.h can have the unwanted consequence to unconditionally
- * #define bool, vector, and pixel keywords,
- * with bad consequences for programs already using these keywords for other purposes.
- * The paragraph defining these macros is skipped when __APPLE_ALTIVEC__ is defined.
- * __APPLE_ALTIVEC__ is _generally_ defined automatically by the compiler,
- * but it seems that, in some cases, it isn't.
- * Force the build macro to be defined, so that keywords are not altered.
- */
-#    if defined(__GNUC__) && !defined(__APPLE_ALTIVEC__)
-#      define __APPLE_ALTIVEC__
-#    endif
 #    include <altivec.h>
 #  endif
 
+/* Restore the original macro values, if applicable. */
+#  pragma pop_macro("pixel")
+#  pragma pop_macro("vector")
+#  pragma pop_macro("bool")
+
 typedef __vector unsigned long long xxh_u64x2;
 typedef __vector unsigned char xxh_u8x16;
 typedef __vector unsigned xxh_u32x4;
@@ -2993,7 +3560,7 @@ XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
 XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
 {
     xxh_u64x2 ret;
-    memcpy(&ret, ptr, sizeof(xxh_u64x2));
+    XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2));
 # if XXH_VSX_BE
     ret = XXH_vec_revb(ret);
 # endif
@@ -3010,8 +3577,9 @@ XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
  /* 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)
+# 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
@@ -3032,13 +3600,29 @@ XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
 # endif /* XXH_vec_mulo, XXH_vec_mule */
 #endif /* XXH_VECTOR == XXH_VSX */
 
+#if XXH_VECTOR == XXH_SVE
+#define ACCRND(acc, offset) \
+do { \
+    svuint64_t input_vec = svld1_u64(mask, xinput + offset);         \
+    svuint64_t secret_vec = svld1_u64(mask, xsecret + offset);       \
+    svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec);     \
+    svuint64_t swapped = svtbl_u64(input_vec, kSwap);                \
+    svuint64_t mixed_lo = svextw_u64_x(mask, mixed);                 \
+    svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32);            \
+    svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
+    acc = svadd_u64_x(mask, acc, mul);                               \
+} while (0)
+#endif /* XXH_VECTOR == XXH_SVE */
+
 
 /* prefetch
  * can be disabled, by declaring XXH_NO_PREFETCH build macro */
 #if defined(XXH_NO_PREFETCH)
 #  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
 #else
-#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
+#  if XXH_SIZE_OPT >= 1
+#    define XXH_PREFETCH(ptr) (void)(ptr)
+#  elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
 #    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
 #    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
 #  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
@@ -3103,7 +3687,6 @@ XXH_mult32to64(xxh_u64 x, xxh_u64 y)
    return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
 }
 #elif defined(_MSC_VER) && defined(_M_IX86)
-#    include <intrin.h>
 #    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
 #else
 /*
@@ -3122,7 +3705,7 @@ XXH_mult32to64(xxh_u64 x, xxh_u64 y)
  * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
  * version.
  *
- * @param lhs, rhs The 64-bit integers to be multiplied
+ * @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
@@ -3143,7 +3726,7 @@ XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
      * In that case it is best to use the portable one.
      * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
      */
-#if defined(__GNUC__) && !defined(__wasm__) \
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
     && defined(__SIZEOF_INT128__) \
     || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
 
@@ -3160,7 +3743,7 @@ XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
      *
      * This compiles to single operand MUL on x64.
      */
-#elif defined(_M_X64) || defined(_M_IA64)
+#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)
 
 #ifndef _MSC_VER
 #   pragma intrinsic(_umul128)
@@ -3172,6 +3755,21 @@ XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
     r128.high64 = product_high;
     return r128;
 
+    /*
+     * MSVC for ARM64's __umulh method.
+     *
+     * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
+     */
+#elif defined(_M_ARM64) || defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(__umulh)
+#endif
+    XXH128_hash_t r128;
+    r128.low64  = lhs * rhs;
+    r128.high64 = __umulh(lhs, rhs);
+    return r128;
+
 #else
     /*
      * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
@@ -3240,7 +3838,7 @@ XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
  * 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
+ * @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()
  */
@@ -3252,7 +3850,7 @@ XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
 }
 
 /*! Seems to produce slightly better code on GCC for some reason. */
-XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
+XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
 {
     XXH_ASSERT(0 <= shift && shift < 64);
     return v64 ^ (v64 >> shift);
@@ -3319,7 +3917,7 @@ static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
  *
  * This adds an extra layer of strength for custom secrets.
  */
-XXH_FORCE_INLINE XXH64_hash_t
+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);
@@ -3341,7 +3939,7 @@ XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_h
     }
 }
 
-XXH_FORCE_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -3357,7 +3955,7 @@ XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_h
     }
 }
 
-XXH_FORCE_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -3374,7 +3972,7 @@ XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_
     }
 }
 
-XXH_FORCE_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(len <= 16);
@@ -3444,7 +4042,7 @@ XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
 }
 
 /* For mid range keys, XXH3 uses a Mum-hash variant. */
-XXH_FORCE_INLINE XXH64_hash_t
+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)
@@ -3452,29 +4050,39 @@ XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
     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 = len * XXH_PRIME64_1, acc_end;
+#if XXH_SIZE_OPT >= 1
+        /* Smaller and cleaner, but slightly slower. */
+        unsigned int i = (unsigned int)(len - 1) / 32;
+        do {
+            acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
+            acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
+        } while (i-- != 0);
+        acc_end = 0;
+#else
+        acc += XXH3_mix16B(input+0, secret+0, seed);
+        acc_end = XXH3_mix16B(input+len-16, secret+16, seed);
         if (len > 32) {
+            acc += XXH3_mix16B(input+16, secret+32, seed);
+            acc_end += XXH3_mix16B(input+len-32, secret+48, seed);
             if (len > 64) {
+                acc += XXH3_mix16B(input+32, secret+64, seed);
+                acc_end += XXH3_mix16B(input+len-48, secret+80, seed);
+
                 if (len > 96) {
                     acc += XXH3_mix16B(input+48, secret+96, seed);
-                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
+                    acc_end += XXH3_mix16B(input+len-64, secret+112, seed);
                 }
-                acc += XXH3_mix16B(input+32, secret+64, seed);
-                acc += XXH3_mix16B(input+len-48, secret+80, seed);
             }
-            acc += XXH3_mix16B(input+16, secret+32, seed);
-            acc += XXH3_mix16B(input+len-32, secret+48, seed);
         }
-        acc += XXH3_mix16B(input+0, secret+0, seed);
-        acc += XXH3_mix16B(input+len-16, secret+16, seed);
-
-        return XXH3_avalanche(acc);
+#endif
+        return XXH3_avalanche(acc + acc_end);
     }
 }
 
 #define XXH3_MIDSIZE_MAX 240
 
-XXH_NO_INLINE XXH64_hash_t
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                       XXH64_hash_t seed)
@@ -3486,13 +4094,17 @@ XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
     #define XXH3_MIDSIZE_LASTOFFSET  17
 
     {   xxh_u64 acc = len * XXH_PRIME64_1;
-        int const nbRounds = (int)len / 16;
-        int i;
+        xxh_u64 acc_end;
+        unsigned int const nbRounds = (unsigned int)len / 16;
+        unsigned int 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);
         }
-        acc = XXH3_avalanche(acc);
+        /* 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);
 #if defined(__clang__)                                /* Clang */ \
     && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
     && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
@@ -3519,11 +4131,13 @@ XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
         #pragma clang loop vectorize(disable)
 #endif
         for (i=8 ; i < nbRounds; i++) {
-            acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+            /*
+             * 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);
         }
-        /* last bytes */
-        acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
-        return XXH3_avalanche(acc);
+        return XXH3_avalanche(acc + acc_end);
     }
 }
 
@@ -3539,10 +4153,51 @@ XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
 #  define ACC_NB XXH_ACC_NB
 #endif
 
+#ifndef XXH_PREFETCH_DIST
+#  ifdef __clang__
+#    define XXH_PREFETCH_DIST 320
+#  else
+#    if (XXH_VECTOR == XXH_AVX512)
+#      define XXH_PREFETCH_DIST 512
+#    else
+#      define XXH_PREFETCH_DIST 384
+#    endif
+#  endif  /* __clang__ */
+#endif  /* XXH_PREFETCH_DIST */
+
+/*
+ * 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);       \
+    }                                                       \
+}
+
+
 XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
 {
     if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
-    memcpy(dst, &v64, sizeof(v64));
+    XXH_memcpy(dst, &v64, sizeof(v64));
 }
 
 /* Several intrinsic functions below are supposed to accept __int64 as argument,
@@ -3559,6 +4214,7 @@ XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
     typedef long long xxh_i64;
 #endif
 
+
 /*
  * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
  *
@@ -3594,7 +4250,7 @@ XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
                      const void* XXH_RESTRICT input,
                      const void* XXH_RESTRICT secret)
 {
-    XXH_ALIGN(64) __m512i* const xacc = (__m512i *) acc;
+    __m512i* const xacc = (__m512i *) acc;
     XXH_ASSERT((((size_t)acc) & 63) == 0);
     XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
 
@@ -3606,7 +4262,7 @@ XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
         /* data_key    = data_vec ^ key_vec; */
         __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
         /* data_key_lo = data_key >> 32; */
-        __m512i const data_key_lo = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32);
         /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
         __m512i const product     = _mm512_mul_epu32     (data_key, data_key_lo);
         /* xacc[0] += swap(data_vec); */
@@ -3616,6 +4272,7 @@ XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
         *xacc = _mm512_add_epi64(product, sum);
     }
 }
+XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)
 
 /*
  * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
@@ -3643,19 +4300,18 @@ XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
     XXH_ASSERT((((size_t)acc) & 63) == 0);
     XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
-    {   XXH_ALIGN(64) __m512i* const xacc = (__m512i*) acc;
+    {   __m512i* const xacc = (__m512i*) acc;
         const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
 
         /* xacc[0] ^= (xacc[0] >> 47) */
         __m512i const acc_vec     = *xacc;
         __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
-        __m512i const data_vec    = _mm512_xor_si512     (acc_vec, shifted);
         /* xacc[0] ^= secret; */
         __m512i const key_vec     = _mm512_loadu_si512   (secret);
-        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+        __m512i const data_key    = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */);
 
         /* xacc[0] *= XXH_PRIME32_1; */
-        __m512i const data_key_hi = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32);
         __m512i const prod_lo     = _mm512_mul_epu32     (data_key, prime32);
         __m512i const prod_hi     = _mm512_mul_epu32     (data_key_hi, prime32);
         *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
@@ -3670,20 +4326,16 @@ XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
     XXH_ASSERT(((size_t)customSecret & 63) == 0);
     (void)(&XXH_writeLE64);
     {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
-        __m512i const seed = _mm512_mask_set1_epi64(_mm512_set1_epi64((xxh_i64)seed64), 0xAA, (xxh_i64)(0U - seed64));
+        __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64);
+        __m512i const seed     = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos);
 
-        XXH_ALIGN(64) const __m512i* const src  = (const __m512i*) XXH3_kSecret;
-        XXH_ALIGN(64)       __m512i* const dest = (      __m512i*) customSecret;
+        const __m512i* const src  = (const __m512i*) ((const void*) XXH3_kSecret);
+              __m512i* const dest = (      __m512i*) customSecret;
         int i;
+        XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 63) == 0);
         for (i=0; i < nbRounds; ++i) {
-            /* GCC has a bug, _mm512_stream_load_si512 accepts 'void*', not 'void const*',
-             * this will warn "discards 'const' qualifier". */
-            union {
-                XXH_ALIGN(64) const __m512i* cp;
-                XXH_ALIGN(64) void* p;
-            } remote_const_void;
-            remote_const_void.cp = src + i;
-            dest[i] = _mm512_add_epi64(_mm512_stream_load_si512(remote_const_void.p), seed);
+            dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
     }   }
 }
 
@@ -3702,7 +4354,7 @@ XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT secret)
 {
     XXH_ASSERT((((size_t)acc) & 31) == 0);
-    {   XXH_ALIGN(32) __m256i* const xacc    =       (__m256i *) acc;
+    {   __m256i* const xacc    =       (__m256i *) acc;
         /* Unaligned. This is mainly for pointer arithmetic, and because
          * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
         const         __m256i* const xinput  = (const __m256i *) input;
@@ -3719,7 +4371,7 @@ XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
             /* data_key    = data_vec ^ key_vec; */
             __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
             /* data_key_lo = data_key >> 32; */
-            __m256i const data_key_lo = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32);
             /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
             __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
             /* xacc[i] += swap(data_vec); */
@@ -3729,12 +4381,13 @@ XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
             xacc[i] = _mm256_add_epi64(product, sum);
     }   }
 }
+XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)
 
 XXH_FORCE_INLINE XXH_TARGET_AVX2 void
 XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
     XXH_ASSERT((((size_t)acc) & 31) == 0);
-    {   XXH_ALIGN(32) __m256i* const xacc = (__m256i*) acc;
+    {   __m256i* const xacc = (__m256i*) acc;
         /* Unaligned. This is mainly for pointer arithmetic, and because
          * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
         const         __m256i* const xsecret = (const __m256i *) secret;
@@ -3751,7 +4404,7 @@ XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
             __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
 
             /* xacc[i] *= XXH_PRIME32_1; */
-            __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32);
             __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
             __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
             xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
@@ -3768,8 +4421,8 @@ XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTR
     XXH_PREFETCH(customSecret);
     {   __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);
 
-        XXH_ALIGN(64) const __m256i* const src  = (const __m256i*) XXH3_kSecret;
-        XXH_ALIGN(64)       __m256i*       dest = (      __m256i*) customSecret;
+        const __m256i* const src  = (const __m256i*) ((const void*) XXH3_kSecret);
+              __m256i*       dest = (      __m256i*) customSecret;
 
 #       if defined(__GNUC__) || defined(__clang__)
         /*
@@ -3779,14 +4432,16 @@ XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTR
          */
         XXH_COMPILER_GUARD(dest);
 #       endif
+        XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 31) == 0);
 
         /* GCC -O2 need unroll loop manually */
-        dest[0] = _mm256_add_epi64(_mm256_stream_load_si256(src+0), seed);
-        dest[1] = _mm256_add_epi64(_mm256_stream_load_si256(src+1), seed);
-        dest[2] = _mm256_add_epi64(_mm256_stream_load_si256(src+2), seed);
-        dest[3] = _mm256_add_epi64(_mm256_stream_load_si256(src+3), seed);
-        dest[4] = _mm256_add_epi64(_mm256_stream_load_si256(src+4), seed);
-        dest[5] = _mm256_add_epi64(_mm256_stream_load_si256(src+5), seed);
+        dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed);
+        dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed);
+        dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed);
+        dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed);
+        dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed);
+        dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed);
     }
 }
 
@@ -3806,7 +4461,7 @@ XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
 {
     /* SSE2 is just a half-scale version of the AVX2 version. */
     XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {   XXH_ALIGN(16) __m128i* const xacc    =       (__m128i *) acc;
+    {   __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;
@@ -3833,12 +4488,13 @@ XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
             xacc[i] = _mm_add_epi64(product, sum);
     }   }
 }
+XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)
 
 XXH_FORCE_INLINE XXH_TARGET_SSE2 void
 XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
     XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {   XXH_ALIGN(16) __m128i* const xacc = (__m128i*) acc;
+    {   __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 xsecret = (const __m128i *) secret;
@@ -3870,7 +4526,7 @@ XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTR
     {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);
 
 #       if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
-        // MSVC 32bit mode does not support _mm_set_epi64x before 2015
+        /* 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
@@ -3878,19 +4534,21 @@ XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTR
 #       endif
         int i;
 
-        XXH_ALIGN(64)        const float* const src  = (float const*) XXH3_kSecret;
-        XXH_ALIGN(XXH_SEC_ALIGN) __m128i*       dest = (__m128i*) customSecret;
+        const void* const src16 = XXH3_kSecret;
+        __m128i* dst16 = (__m128i*) customSecret;
 #       if defined(__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);
+        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) {
-            dest[i] = _mm_add_epi64(_mm_castps_si128(_mm_load_ps(src+i*4)), seed);
+            dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
     }   }
 }
 
@@ -3898,42 +4556,112 @@ XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTR
 
 #if (XXH_VECTOR == XXH_NEON)
 
+/* forward declarations for the scalar routines */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret, size_t lane);
+
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret, size_t lane);
+
+/*!
+ * @internal
+ * @brief The bulk processing loop for NEON.
+ *
+ * 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.
+ */
 XXH_FORCE_INLINE void
 XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT input,
                     const void* 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);
     {
-        XXH_ALIGN(16) uint64x2_t* const xacc = (uint64x2_t *) acc;
+        uint64x2_t* const xacc = (uint64x2_t *) acc;
         /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
         uint8_t const* const xinput = (const uint8_t *) input;
         uint8_t const* const xsecret  = (const uint8_t *) secret;
 
         size_t i;
-        for (i=0; i < XXH_STRIPE_LEN / sizeof(uint64x2_t); i++) {
+        /* AArch64 uses both scalar and neon at the same time */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarRound(acc, input, secret, i);
+        }
+        i = 0;
+        for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
+            uint64x2_t acc_vec1 = xacc[i];
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec1 = XXH_vld1q_u64(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec1  = XXH_vld1q_u64(xsecret + (i * 16));
+            /* acc_vec_2 = swap(data_vec) */
+            uint64x2_t acc_vec_21 = vextq_u64(data_vec1, data_vec1, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key1 = veorq_u64(data_vec1, key_vec1);
+
+            uint64x2_t acc_vec2 = xacc[i+1];
             /* data_vec = xinput[i]; */
-            uint8x16_t data_vec    = vld1q_u8(xinput  + (i * 16));
+            uint64x2_t data_vec2 = XXH_vld1q_u64(xinput  + ((i+1) * 16));
             /* key_vec  = xsecret[i];  */
-            uint8x16_t key_vec     = vld1q_u8(xsecret + (i * 16));
+            uint64x2_t key_vec2  = XXH_vld1q_u64(xsecret + ((i+1) * 16));
+            /* acc_vec_2 = swap(data_vec) */
+            uint64x2_t acc_vec_22 = vextq_u64(data_vec2, data_vec2, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key2 = veorq_u64(data_vec2, key_vec2);
+
+            /* data_key_lo = {(data_key1 & 0xFFFFFFFF), (data_key2 & 0xFFFFFFFF)};
+             * data_key_hi = {(data_key1 >> 32), (data_key2 >> 32)};
+             */
+            uint32x4x2_t zipped = vuzpq_u32(vreinterpretq_u32_u64(data_key1), vreinterpretq_u32_u64(data_key2));
+            uint32x4_t data_key_lo = zipped.val[0];
+            uint32x4_t data_key_hi = zipped.val[1];
+
+            /* acc_vec_2 += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
+            acc_vec_21 = vmlal_u32 (acc_vec_21, vget_low_u32(data_key_lo), vget_low_u32(data_key_hi));
+            XXH_COMPILER_GUARD_W(acc_vec_21);
+            /* xacc[i] += acc_vec_2; */
+            acc_vec1 = vaddq_u64 (acc_vec1, acc_vec_21);
+            xacc[i] = acc_vec1;
+            /* acc_vec_2 += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
+            acc_vec_22 = vmlal_u32 (acc_vec_22, vget_high_u32(data_key_lo), vget_high_u32(data_key_hi));
+            XXH_COMPILER_GUARD_W(acc_vec_22);
+            /* xacc[i] += acc_vec_2; */
+            acc_vec2 = vaddq_u64 (acc_vec2, acc_vec_22);
+            xacc[i+1] = acc_vec2;
+        }
+        for (; i < XXH3_NEON_LANES / 2; i++) {
+            uint64x2_t acc_vec = xacc[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));
             uint64x2_t data_key;
             uint32x2_t data_key_lo, data_key_hi;
-            /* xacc[i] += swap(data_vec); */
-            uint64x2_t const data64  = vreinterpretq_u64_u8(data_vec);
-            uint64x2_t const swapped = vextq_u64(data64, data64, 1);
-            xacc[i] = vaddq_u64 (xacc[i], swapped);
+            /* acc_vec_2 = swap(data_vec) */
+            uint64x2_t acc_vec_2 = vextq_u64(data_vec, data_vec, 1);
             /* data_key = data_vec ^ key_vec; */
-            data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec));
+            data_key = veorq_u64(data_vec, key_vec);
             /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF);
              * data_key_hi = (uint32x2_t) (data_key >> 32);
              * data_key = UNDEFINED; */
             XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
-            /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
-            xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi);
-
+            /* acc_vec_2 += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
+            acc_vec_2 = vmlal_u32 (acc_vec_2, data_key_lo, data_key_hi);
+            XXH_COMPILER_GUARD_W(acc_vec_2);
+            /* xacc[i] += acc_vec_2; */
+            acc_vec = vaddq_u64 (acc_vec, acc_vec_2);
+            xacc[i] = acc_vec;
         }
+
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
 
 XXH_FORCE_INLINE void
 XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
@@ -3945,15 +4673,19 @@ XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
         uint32x2_t prime       = vdup_n_u32 (XXH_PRIME32_1);
 
         size_t i;
-        for (i=0; i < XXH_STRIPE_LEN/sizeof(uint64x2_t); i++) {
+        /* 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);
+            uint64x2_t shifted  = vshrq_n_u64   (acc_vec, 47);
+            uint64x2_t data_vec = veorq_u64     (acc_vec, shifted);
 
             /* xacc[i] ^= xsecret[i]; */
-            uint8x16_t key_vec  = vld1q_u8(xsecret + (i * 16));
-            uint64x2_t data_key = veorq_u64(data_vec, vreinterpretq_u64_u8(key_vec));
+            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 */
             uint32x2_t data_key_lo, data_key_hi;
@@ -3981,11 +4713,12 @@ XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
                  */
                 uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime);
                 /* xacc[i] = prod_hi << 32; */
-                xacc[i] = vshlq_n_u64(prod_hi, 32);
+                prod_hi = vshlq_n_u64(prod_hi, 32);
                 /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */
-                xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime);
+                xacc[i] = vmlal_u32(prod_hi, data_key_lo, prime);
             }
-    }   }
+        }
+    }
 }
 
 #endif
@@ -3997,7 +4730,8 @@ XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT input,
                     const void* XXH_RESTRICT secret)
 {
-          xxh_u64x2* const xacc     =       (xxh_u64x2*) acc;    /* presumed aligned */
+    /* presumed aligned */
+    unsigned int* const xacc = (unsigned int*) acc;
     xxh_u64x2 const* const xinput   = (xxh_u64x2 const*) input;   /* no alignment restriction */
     xxh_u64x2 const* const xsecret  = (xxh_u64x2 const*) secret;    /* no alignment restriction */
     xxh_u64x2 const v32 = { 32, 32 };
@@ -4012,16 +4746,21 @@ XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
         xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32);
         /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */
         xxh_u64x2 const product  = XXH_vec_mulo((xxh_u32x4)data_key, shuffled);
-        xacc[i] += product;
+        /* acc_vec = xacc[i]; */
+        xxh_u64x2 acc_vec        = (xxh_u64x2)vec_xl(0, xacc + 4 * i);
+        acc_vec += product;
 
         /* swap high and low halves */
 #ifdef __s390x__
-        xacc[i] += vec_permi(data_vec, data_vec, 2);
+        acc_vec += vec_permi(data_vec, data_vec, 2);
 #else
-        xacc[i] += vec_xxpermdi(data_vec, data_vec, 2);
+        acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
 #endif
+        /* xacc[i] = acc_vec; */
+        vec_xst((xxh_u32x4)acc_vec, 0, xacc + 4 * i);
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(vsx)
 
 XXH_FORCE_INLINE void
 XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
@@ -4055,40 +4794,202 @@ XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 
 #endif
 
+#if (XXH_VECTOR == XXH_SVE)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_sve( void* XXH_RESTRICT acc,
+                   const void* XXH_RESTRICT input,
+                   const void* XXH_RESTRICT secret)
+{
+    uint64_t *xacc = (uint64_t *)acc;
+    const uint64_t *xinput = (const uint64_t *)(const void *)input;
+    const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+    svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+    uint64_t element_count = svcntd();
+    if (element_count >= 8) {
+        svbool_t mask = svptrue_pat_b64(SV_VL8);
+        svuint64_t vacc = svld1_u64(mask, xacc);
+        ACCRND(vacc, 0);
+        svst1_u64(mask, xacc, vacc);
+    } else if (element_count == 2) {   /* sve128 */
+        svbool_t mask = svptrue_pat_b64(SV_VL2);
+        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+        svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+        svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+        svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+        ACCRND(acc0, 0);
+        ACCRND(acc1, 2);
+        ACCRND(acc2, 4);
+        ACCRND(acc3, 6);
+        svst1_u64(mask, xacc + 0, acc0);
+        svst1_u64(mask, xacc + 2, acc1);
+        svst1_u64(mask, xacc + 4, acc2);
+        svst1_u64(mask, xacc + 6, acc3);
+    } else {
+        svbool_t mask = svptrue_pat_b64(SV_VL4);
+        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+        svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+        ACCRND(acc0, 0);
+        ACCRND(acc1, 4);
+        svst1_u64(mask, xacc + 0, acc0);
+        svst1_u64(mask, xacc + 4, acc1);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc,
+               const xxh_u8* XXH_RESTRICT input,
+               const xxh_u8* XXH_RESTRICT secret,
+               size_t nbStripes)
+{
+    if (nbStripes != 0) {
+        uint64_t *xacc = (uint64_t *)acc;
+        const uint64_t *xinput = (const uint64_t *)(const void *)input;
+        const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+        svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+        uint64_t element_count = svcntd();
+        if (element_count >= 8) {
+            svbool_t mask = svptrue_pat_b64(SV_VL8);
+            svuint64_t vacc = svld1_u64(mask, xacc + 0);
+            do {
+                /* svprfd(svbool_t, void *, enum svfprop); */
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(vacc, 0);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, vacc);
+        } else if (element_count == 2) { /* sve128 */
+            svbool_t mask = svptrue_pat_b64(SV_VL2);
+            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+            svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+            svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+            svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+            do {
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(acc0, 0);
+                ACCRND(acc1, 2);
+                ACCRND(acc2, 4);
+                ACCRND(acc3, 6);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, acc0);
+           svst1_u64(mask, xacc + 2, acc1);
+           svst1_u64(mask, xacc + 4, acc2);
+           svst1_u64(mask, xacc + 6, acc3);
+        } else {
+            svbool_t mask = svptrue_pat_b64(SV_VL4);
+            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+            svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+            do {
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(acc0, 0);
+                ACCRND(acc1, 4);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, acc0);
+           svst1_u64(mask, xacc + 4, acc1);
+       }
+    }
+}
+
+#endif
+
 /* scalar variants - universal */
 
+/*!
+ * @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,
+                 void const* XXH_RESTRICT input,
+                 void const* 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(data_key & 0xFFFFFFFF, data_key >> 32);
+    }
+}
+
+/*!
+ * @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,
                      const void* XXH_RESTRICT input,
                      const void* XXH_RESTRICT secret)
 {
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */
-    const xxh_u8* const xinput  = (const xxh_u8*) input;  /* no alignment restriction */
-    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
     size_t i;
-    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+    /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
+#if defined(__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++) {
-        xxh_u64 const data_val = XXH_readLE64(xinput + 8*i);
-        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i*8);
-        xacc[i ^ 1] += data_val; /* swap adjacent lanes */
-        xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32);
+        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_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret,
+                         size_t lane)
 {
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
+    xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
     const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
-    size_t i;
     XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
-    for (i=0; i < XXH_ACC_NB; i++) {
-        xxh_u64 const key64 = XXH_readLE64(xsecret + 8*i);
-        xxh_u64 acc64 = xacc[i];
+    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[i] = acc64;
+        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, const void* XXH_RESTRICT secret)
+{
+    size_t i;
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarScrambleRound(acc, secret, i);
     }
 }
 
@@ -4110,8 +5011,9 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
      * 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), long MOVK chains stall the
-     * integer pipelines:
+     * constant compared to 4 cycles for LDR), it fights for bandwidth with
+     * the arithmetic instructions.
+     *
      *   I   L   S
      * MOVK
      * MOVK
@@ -4128,6 +5030,9 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
      *  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
@@ -4157,7 +5062,7 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
 }
 
 
-typedef void (*XXH3_f_accumulate_512)(void* XXH_RESTRICT, const void*, const void*);
+typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t);
 typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
 typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
 
@@ -4165,82 +5070,63 @@ typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
 #if (XXH_VECTOR == XXH_AVX512)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_avx512
+#define XXH3_accumulate     XXH3_accumulate_avx512
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_avx512
 #define XXH3_initCustomSecret XXH3_initCustomSecret_avx512
 
 #elif (XXH_VECTOR == XXH_AVX2)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_avx2
+#define XXH3_accumulate     XXH3_accumulate_avx2
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_avx2
 #define XXH3_initCustomSecret XXH3_initCustomSecret_avx2
 
 #elif (XXH_VECTOR == XXH_SSE2)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_sse2
+#define XXH3_accumulate     XXH3_accumulate_sse2
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_sse2
 #define XXH3_initCustomSecret XXH3_initCustomSecret_sse2
 
 #elif (XXH_VECTOR == XXH_NEON)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_neon
+#define XXH3_accumulate     XXH3_accumulate_neon
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_neon
 #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
 #elif (XXH_VECTOR == XXH_VSX)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_vsx
+#define XXH3_accumulate     XXH3_accumulate_vsx
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_vsx
 #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
+#elif (XXH_VECTOR == XXH_SVE)
+#define XXH3_accumulate_512 XXH3_accumulate_512_sve
+#define XXH3_accumulate     XXH3_accumulate_sve
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
 #else /* scalar */
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_accumulate     XXH3_accumulate_scalar
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
 #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
 #endif
 
-
-
-#ifndef XXH_PREFETCH_DIST
-#  ifdef __clang__
-#    define XXH_PREFETCH_DIST 320
-#  else
-#    if (XXH_VECTOR == XXH_AVX512)
-#      define XXH_PREFETCH_DIST 512
-#    else
-#      define XXH_PREFETCH_DIST 384
-#    endif
-#  endif  /* __clang__ */
-#endif  /* XXH_PREFETCH_DIST */
-
-/*
- * XXH3_accumulate()
- * Loops over XXH3_accumulate_512().
- * Assumption: nbStripes will not overflow the secret size
- */
-XXH_FORCE_INLINE void
-XXH3_accumulate(     xxh_u64* XXH_RESTRICT acc,
-                const xxh_u8* XXH_RESTRICT input,
-                const xxh_u8* XXH_RESTRICT secret,
-                      size_t nbStripes,
-                      XXH3_f_accumulate_512 f_acc512)
-{
-    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);
-        f_acc512(acc,
-                 in,
-                 secret + n*XXH_SECRET_CONSUME_RATE);
-    }
-}
+#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */
+#  undef XXH3_initCustomSecret
+#  define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+#endif
 
 XXH_FORCE_INLINE void
 XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
                       const xxh_u8* XXH_RESTRICT input, size_t len,
                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                            XXH3_f_accumulate_512 f_acc512,
+                            XXH3_f_accumulate f_acc,
                             XXH3_f_scrambleAcc f_scramble)
 {
     size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
@@ -4252,7 +5138,7 @@ XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
     XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
 
     for (n = 0; n < nb_blocks; n++) {
-        XXH3_accumulate(acc, input + n*block_len, secret, nbStripesPerBlock, f_acc512);
+        f_acc(acc, input + n*block_len, secret, nbStripesPerBlock);
         f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
     }
 
@@ -4260,12 +5146,12 @@ XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
     XXH_ASSERT(len > XXH_STRIPE_LEN);
     {   size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
         XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
-        XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, f_acc512);
+        f_acc(acc, input + nb_blocks*block_len, secret, nbStripes);
 
         /* 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 */
-            f_acc512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+            XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
     }   }
 }
 
@@ -4310,12 +5196,12 @@ XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secre
 XXH_FORCE_INLINE XXH64_hash_t
 XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
                            const void* XXH_RESTRICT secret, size_t secretSize,
-                           XXH3_f_accumulate_512 f_acc512,
+                           XXH3_f_accumulate f_acc,
                            XXH3_f_scrambleAcc f_scramble)
 {
     XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
 
-    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc512, f_scramble);
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc, f_scramble);
 
     /* converge into final hash */
     XXH_STATIC_ASSERT(sizeof(acc) == 64);
@@ -4326,29 +5212,30 @@ XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
 }
 
 /*
- * It's important for performance that XXH3_hashLong is not inlined.
+ * 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.
  */
-XXH_NO_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH64_hash_t
 XXH3_hashLong_64b_withSecret(const void* 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_512, XXH3_scrambleAcc);
+    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 /*
- * It's important for performance that XXH3_hashLong is not inlined.
- * Since the function is not inlined, the compiler may not be able to understand that,
- * in some scenarios, its `secret` argument is actually a compile time constant.
- * This variant enforces that the compiler can detect that,
- * and uses this opportunity to streamline the generated code for better performance.
+ * 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 XXH64_hash_t
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
 XXH3_hashLong_64b_default(const void* 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_512, XXH3_scrambleAcc);
+    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 /*
@@ -4365,18 +5252,20 @@ XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
 XXH_FORCE_INLINE XXH64_hash_t
 XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
                                     XXH64_hash_t seed,
-                                    XXH3_f_accumulate_512 f_acc512,
+                                    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_acc512, f_scramble);
+                                          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_acc512, f_scramble);
+                                          f_acc, f_scramble);
     }
 }
 
@@ -4384,12 +5273,12 @@ XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
  * It's important for performance that XXH3_hashLong is not inlined.
  */
 XXH_NO_INLINE XXH64_hash_t
-XXH3_hashLong_64b_withSeed(const void* input, size_t len,
-                           XXH64_hash_t seed, const xxh_u8* secret, size_t secretLen)
+XXH3_hashLong_64b_withSeed(const void* 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_512, XXH3_scrambleAcc, XXH3_initCustomSecret);
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
 }
 
 
@@ -4421,29 +5310,37 @@ XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
 
 /* ===   Public entry point   === */
 
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
 {
-    return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+    return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
-    return XXH3_64bits_internal(input, len, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+    return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
 {
-    return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+    return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
 }
 
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (length <= XXH3_MIDSIZE_MAX)
+        return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize);
+}
 
-/* ===   XXH3 streaming   === */
 
+/* ===   XXH3 streaming   === */
+#ifndef XXH_NO_STREAM
 /*
  * Malloc's a pointer that is always aligned to align.
  *
@@ -4467,7 +5364,7 @@ XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
  *
  * Align must be a power of 2 and 8 <= align <= 128.
  */
-static void* XXH_alignedMalloc(size_t s, size_t align)
+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 */
@@ -4509,7 +5406,7 @@ static void XXH_alignedFree(void* p)
         XXH_free(base);
     }
 }
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
 {
     XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
@@ -4518,24 +5415,24 @@ XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
     return state;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
 {
     XXH_alignedFree(statePtr);
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API void
-XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state)
+XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state)
 {
-    memcpy(dst_state, src_state, sizeof(*dst_state));
+    XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
 }
 
 static void
 XXH3_reset_internal(XXH3_state_t* statePtr,
-                           XXH64_hash_t seed,
-                           const void* secret, size_t secretSize)
+                    XXH64_hash_t seed,
+                    const void* secret, size_t secretSize)
 {
     size_t const initStart = offsetof(XXH3_state_t, bufferedSize);
     size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart;
@@ -4552,24 +5449,25 @@ XXH3_reset_internal(XXH3_state_t* statePtr,
     statePtr->acc[6] = XXH_PRIME64_5;
     statePtr->acc[7] = XXH_PRIME32_1;
     statePtr->seed = seed;
+    statePtr->useSeed = (seed != 0);
     statePtr->extSecret = (const unsigned char*)secret;
     XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
     statePtr->secretLimit = secretSize - XXH_STRIPE_LEN;
     statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset(XXH3_state_t* statePtr)
+XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
 {
     if (statePtr == NULL) return XXH_ERROR;
     XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
+XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
     if (statePtr == NULL) return XXH_ERROR;
     XXH3_reset_internal(statePtr, 0, secret, secretSize);
@@ -4578,17 +5476,30 @@ XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
+XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
 {
     if (statePtr == NULL) return XXH_ERROR;
     if (seed==0) return XXH3_64bits_reset(statePtr);
-    if (seed != statePtr->seed) XXH3_initCustomSecret(statePtr->customSecret, seed);
+    if ((seed != statePtr->seed) || (statePtr->extSecret != NULL))
+        XXH3_initCustomSecret(statePtr->customSecret, seed);
     XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
     return XXH_OK;
 }
 
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, seed64, secret, secretSize);
+    statePtr->useSeed = 1; /* always, even if seed64==0 */
+    return XXH_OK;
+}
+
 /* Note : when XXH3_consumeStripes() is invoked,
  * there must be a guarantee that at least one more byte must be consumed from input
  * so that the function can blindly consume all stripes using the "normal" secret segment */
@@ -4597,7 +5508,7 @@ 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_512 f_acc512,
+                    XXH3_f_accumulate f_acc,
                     XXH3_f_scrambleAcc f_scramble)
 {
     XXH_ASSERT(nbStripes <= nbStripesPerBlock);  /* can handle max 1 scramble per invocation */
@@ -4606,45 +5517,58 @@ XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
         /* need a scrambling operation */
         size_t const nbStripesToEndofBlock = nbStripesPerBlock - *nbStripesSoFarPtr;
         size_t const nbStripesAfterBlock = nbStripes - nbStripesToEndofBlock;
-        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripesToEndofBlock, f_acc512);
+        f_acc(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripesToEndofBlock);
         f_scramble(acc, secret + secretLimit);
-        XXH3_accumulate(acc, input + nbStripesToEndofBlock * XXH_STRIPE_LEN, secret, nbStripesAfterBlock, f_acc512);
+        f_acc(acc, input + nbStripesToEndofBlock * XXH_STRIPE_LEN, secret, nbStripesAfterBlock);
         *nbStripesSoFarPtr = nbStripesAfterBlock;
     } else {
-        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, f_acc512);
+        f_acc(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes);
         *nbStripesSoFarPtr += nbStripes;
     }
 }
 
+#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* state,
-            const xxh_u8* input, size_t len,
-            XXH3_f_accumulate_512 f_acc512,
+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)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
         return XXH_OK;
-#else
-        return XXH_ERROR;
-#endif
+    }
 
+    XXH_ASSERT(state != NULL);
     {   const xxh_u8* const bEnd = input + len;
         const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
-
+#if defined(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]; 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);
 
-        if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) {  /* fill in tmp buffer */
+        /* small input : just fill in tmp buffer */
+        if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) {
             XXH_memcpy(state->buffer + state->bufferedSize, input, len);
             state->bufferedSize += (XXH32_hash_t)len;
             return XXH_OK;
         }
-        /* total input is now > XXH3_INTERNALBUFFER_SIZE */
 
+        /* 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 */
 
@@ -4656,45 +5580,82 @@ XXH3_update(XXH3_state_t* state,
             size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
             XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
             input += loadSize;
-            XXH3_consumeStripes(state->acc,
+            XXH3_consumeStripes(acc,
                                &state->nbStripesSoFar, state->nbStripesPerBlock,
                                 state->buffer, XXH3_INTERNALBUFFER_STRIPES,
                                 secret, state->secretLimit,
-                                f_acc512, f_scramble);
+                                f_acc, f_scramble);
             state->bufferedSize = 0;
         }
         XXH_ASSERT(input < bEnd);
 
-        /* Consume input by a multiple of internal buffer size */
-        if (input+XXH3_INTERNALBUFFER_SIZE < bEnd) {
-            const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE;
-            do {
-                XXH3_consumeStripes(state->acc,
-                                   &state->nbStripesSoFar, state->nbStripesPerBlock,
-                                    input, XXH3_INTERNALBUFFER_STRIPES,
-                                    secret, state->secretLimit,
-                                    f_acc512, f_scramble);
-                input += XXH3_INTERNALBUFFER_SIZE;
-            } while (input<limit);
-            /* for last partial stripe */
-            memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+        /* large input to consume : ingest per full block */
+        if ((size_t)(bEnd - input) > state->nbStripesPerBlock * XXH_STRIPE_LEN) {
+            size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
+            XXH_ASSERT(state->nbStripesPerBlock >= state->nbStripesSoFar);
+            /* join to current block's end */
+            {   size_t const nbStripesToEnd = state->nbStripesPerBlock - state->nbStripesSoFar;
+                XXH_ASSERT(nbStripesToEnd <= nbStripes);
+                f_acc(acc, input, secret + state->nbStripesSoFar * XXH_SECRET_CONSUME_RATE, nbStripesToEnd);
+                f_scramble(acc, secret + state->secretLimit);
+                state->nbStripesSoFar = 0;
+                input += nbStripesToEnd * XXH_STRIPE_LEN;
+                nbStripes -= nbStripesToEnd;
+            }
+            /* consume per entire blocks */
+            while(nbStripes >= state->nbStripesPerBlock) {
+                f_acc(acc, input, secret, state->nbStripesPerBlock);
+                f_scramble(acc, secret + state->secretLimit);
+                input += state->nbStripesPerBlock * XXH_STRIPE_LEN;
+                nbStripes -= state->nbStripesPerBlock;
+            }
+            /* consume last partial block */
+            f_acc(acc, input, secret, nbStripes);
+            input += nbStripes * XXH_STRIPE_LEN;
+            XXH_ASSERT(input < bEnd);  /* at least some bytes left */
+            state->nbStripesSoFar = nbStripes;
+            /* buffer predecessor of last partial stripe */
+            XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+            XXH_ASSERT(bEnd - input <= XXH_STRIPE_LEN);
+        } else {
+            /* content to consume <= block size */
+            /* Consume input by a multiple of internal buffer size */
+            if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
+                const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE;
+                do {
+                    XXH3_consumeStripes(acc,
+                                       &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                        input, XXH3_INTERNALBUFFER_STRIPES,
+                                        secret, state->secretLimit,
+                                        f_acc, f_scramble);
+                    input += XXH3_INTERNALBUFFER_SIZE;
+                } while (input<limit);
+                /* buffer predecessor of last partial stripe */
+                XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+            }
         }
-        XXH_ASSERT(input < bEnd);
 
         /* Some remaining input (always) : buffer it */
+        XXH_ASSERT(input < bEnd);
+        XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE);
+        XXH_ASSERT(state->bufferedSize == 0);
         XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
         state->bufferedSize = (XXH32_hash_t)(bEnd-input);
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* save stack accumulators into state */
+        memcpy(state->acc, acc, sizeof(acc));
+#endif
     }
 
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len)
+XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
     return XXH3_update(state, (const xxh_u8*)input, len,
-                       XXH3_accumulate_512, XXH3_scrambleAcc);
+                       XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 
@@ -4707,7 +5668,7 @@ XXH3_digest_long (XXH64_hash_t* acc,
      * Digest on a local copy. This way, the state remains unaltered, and it can
      * continue ingesting more input afterwards.
      */
-    memcpy(acc, state->acc, sizeof(state->acc));
+    XXH_memcpy(acc, state->acc, sizeof(state->acc));
     if (state->bufferedSize >= XXH_STRIPE_LEN) {
         size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
         size_t nbStripesSoFar = state->nbStripesSoFar;
@@ -4715,7 +5676,7 @@ XXH3_digest_long (XXH64_hash_t* acc,
                            &nbStripesSoFar, state->nbStripesPerBlock,
                             state->buffer, nbStripes,
                             secret, state->secretLimit,
-                            XXH3_accumulate_512, XXH3_scrambleAcc);
+                            XXH3_accumulate, XXH3_scrambleAcc);
         /* last stripe */
         XXH3_accumulate_512(acc,
                             state->buffer + state->bufferedSize - XXH_STRIPE_LEN,
@@ -4724,16 +5685,16 @@ XXH3_digest_long (XXH64_hash_t* acc,
         xxh_u8 lastStripe[XXH_STRIPE_LEN];
         size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
         XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
-        memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
-        memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
+        XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
+        XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
         XXH3_accumulate_512(acc,
                             lastStripe,
                             secret + state->secretLimit - XXH_SECRET_LASTACC_START);
     }
 }
 
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
 {
     const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
     if (state->totalLen > XXH3_MIDSIZE_MAX) {
@@ -4744,57 +5705,12 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
                               (xxh_u64)state->totalLen * XXH_PRIME64_1);
     }
     /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
-    if (state->seed)
+    if (state->useSeed)
         return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
     return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
                                   secret, state->secretLimit + XXH_STRIPE_LEN);
 }
-
-
-#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
-
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API void
-XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSeedSize)
-{
-    XXH_ASSERT(secretBuffer != NULL);
-    if (customSeedSize == 0) {
-        memcpy(secretBuffer, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
-        return;
-    }
-    XXH_ASSERT(customSeed != NULL);
-
-    {   size_t const segmentSize = sizeof(XXH128_hash_t);
-        size_t const nbSegments = XXH_SECRET_DEFAULT_SIZE / segmentSize;
-        XXH128_canonical_t scrambler;
-        XXH64_hash_t seeds[12];
-        size_t segnb;
-        XXH_ASSERT(nbSegments == 12);
-        XXH_ASSERT(segmentSize * nbSegments == XXH_SECRET_DEFAULT_SIZE); /* exact multiple */
-        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
-
-        /*
-        * Copy customSeed to seeds[], truncating or repeating as necessary.
-        */
-        {   size_t toFill = XXH_MIN(customSeedSize, sizeof(seeds));
-            size_t filled = toFill;
-            memcpy(seeds, customSeed, toFill);
-            while (filled < sizeof(seeds)) {
-                toFill = XXH_MIN(filled, sizeof(seeds) - filled);
-                memcpy((char*)seeds + filled, seeds, toFill);
-                filled += toFill;
-        }   }
-
-        /* generate secret */
-        memcpy(secretBuffer, &scrambler, sizeof(scrambler));
-        for (segnb=1; segnb < nbSegments; segnb++) {
-            size_t const segmentStart = segnb * segmentSize;
-            XXH128_canonical_t segment;
-            XXH128_canonicalFromHash(&segment,
-                XXH128(&scrambler, sizeof(scrambler), XXH_readLE64(seeds + segnb) + segnb) );
-            memcpy((char*)secretBuffer + segmentStart, &segment, sizeof(segment));
-    }   }
-}
+#endif /* !XXH_NO_STREAM */
 
 
 /* ==========================================
@@ -4814,7 +5730,7 @@ XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSee
  * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
  */
 
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     /* A doubled version of 1to3_64b with different constants. */
@@ -4843,7 +5759,7 @@ XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_
     }
 }
 
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -4870,7 +5786,7 @@ XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_
     }
 }
 
-XXH_FORCE_INLINE XXH128_hash_t
+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);
@@ -4945,7 +5861,7 @@ XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64
 /*
  * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
  */
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(len <= 16);
@@ -4976,7 +5892,7 @@ XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
 }
 
 
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                       XXH64_hash_t seed)
@@ -4987,6 +5903,16 @@ XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
     {   XXH128_hash_t acc;
         acc.low64 = len * XXH_PRIME64_1;
         acc.high64 = 0;
+
+#if XXH_SIZE_OPT >= 1
+        {
+            /* Smaller, but slightly slower. */
+            unsigned int i = (unsigned int)(len - 1) / 32;
+            do {
+                acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed);
+            } while (i-- != 0);
+        }
+#else
         if (len > 32) {
             if (len > 64) {
                 if (len > 96) {
@@ -4997,6 +5923,7 @@ XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
             acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
         }
         acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
+#endif
         {   XXH128_hash_t h128;
             h128.low64  = acc.low64 + acc.high64;
             h128.high64 = (acc.low64    * XXH_PRIME64_1)
@@ -5009,7 +5936,7 @@ XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
     }
 }
 
-XXH_NO_INLINE XXH128_hash_t
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
                        const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                        XXH64_hash_t seed)
@@ -5018,25 +5945,34 @@ XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
     XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
 
     {   XXH128_hash_t acc;
-        int const nbRounds = (int)len / 32;
-        int i;
+        unsigned i;
         acc.low64 = len * XXH_PRIME64_1;
         acc.high64 = 0;
-        for (i=0; i<4; i++) {
+        /*
+         *  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  + (32 * i),
-                                input  + (32 * i) + 16,
-                                secret + (32 * i),
+                                input  + i - 32,
+                                input  + i - 16,
+                                secret + i - 32,
                                 seed);
         }
         acc.low64 = XXH3_avalanche(acc.low64);
         acc.high64 = XXH3_avalanche(acc.high64);
-        XXH_ASSERT(nbRounds >= 4);
-        for (i=4 ; i < nbRounds; i++) {
+        /*
+         * 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 + (32 * i),
-                                input + (32 * i) + 16,
-                                secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)),
+                                input + i - 32,
+                                input + i - 16,
+                                secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
                                 seed);
         }
         /* last bytes */
@@ -5044,7 +5980,7 @@ XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
                             input + len - 16,
                             input + len - 32,
                             secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
-                            0ULL - seed);
+                            (XXH64_hash_t)0 - seed);
 
         {   XXH128_hash_t h128;
             h128.low64  = acc.low64 + acc.high64;
@@ -5061,12 +5997,12 @@ XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
 XXH_FORCE_INLINE XXH128_hash_t
 XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
                             const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                            XXH3_f_accumulate_512 f_acc512,
+                            XXH3_f_accumulate f_acc,
                             XXH3_f_scrambleAcc f_scramble)
 {
     XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
 
-    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc512, f_scramble);
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble);
 
     /* converge into final hash */
     XXH_STATIC_ASSERT(sizeof(acc) == 64);
@@ -5084,46 +6020,47 @@ XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
 }
 
 /*
- * It's important for performance that XXH3_hashLong is not inlined.
+ * It's important for performance that XXH3_hashLong() is not inlined.
  */
-XXH_NO_INLINE XXH128_hash_t
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
 XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
                            XXH64_hash_t seed64,
                            const void* 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_512, XXH3_scrambleAcc);
+                                       XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 /*
- * It's important for performance that XXH3_hashLong is not inlined.
+ * 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.
  */
-XXH_NO_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH128_hash_t
 XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len,
                               XXH64_hash_t seed64,
                               const void* XXH_RESTRICT secret, size_t secretLen)
 {
     (void)seed64;
     return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
-                                       XXH3_accumulate_512, XXH3_scrambleAcc);
+                                       XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 XXH_FORCE_INLINE XXH128_hash_t
 XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len,
                                 XXH64_hash_t seed64,
-                                XXH3_f_accumulate_512 f_acc512,
+                                XXH3_f_accumulate f_acc,
                                 XXH3_f_scrambleAcc f_scramble,
                                 XXH3_f_initCustomSecret f_initSec)
 {
     if (seed64 == 0)
         return XXH3_hashLong_128b_internal(input, len,
                                            XXH3_kSecret, sizeof(XXH3_kSecret),
-                                           f_acc512, f_scramble);
+                                           f_acc, f_scramble);
     {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
         f_initSec(secret, seed64);
         return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret),
-                                           f_acc512, f_scramble);
+                                           f_acc, f_scramble);
     }
 }
 
@@ -5136,7 +6073,7 @@ XXH3_hashLong_128b_withSeed(const void* input, size_t len,
 {
     (void)secret; (void)secretLen;
     return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
-                XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret);
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
 }
 
 typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
@@ -5166,88 +6103,94 @@ XXH3_128bits_internal(const void* input, size_t len,
 
 /* ===   Public XXH128 API   === */
 
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len)
 {
     return XXH3_128bits_internal(input, len, 0,
                                  XXH3_kSecret, sizeof(XXH3_kSecret),
                                  XXH3_hashLong_128b_default);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
     return XXH3_128bits_internal(input, len, 0,
                                  (const xxh_u8*)secret, secretSize,
                                  XXH3_hashLong_128b_withSecret);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
     return XXH3_128bits_internal(input, len, seed,
                                  XXH3_kSecret, sizeof(XXH3_kSecret),
                                  XXH3_hashLong_128b_withSeed);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize);
+}
+
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH128(const void* input, size_t len, XXH64_hash_t seed)
+XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
     return XXH3_128bits_withSeed(input, len, seed);
 }
 
 
 /* ===   XXH3 128-bit streaming   === */
-
+#ifndef XXH_NO_STREAM
 /*
- * All the functions are actually the same as for 64-bit streaming variant.
+ * All initialization and update functions are identical to 64-bit streaming variant.
  * The only difference is the finalization routine.
  */
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset(XXH3_state_t* statePtr)
+XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
 {
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
-    return XXH_OK;
+    return XXH3_64bits_reset(statePtr);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
+XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_reset_internal(statePtr, 0, secret, secretSize);
-    if (secret == NULL) return XXH_ERROR;
-    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
-    return XXH_OK;
+    return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
+XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
 {
-    if (statePtr == NULL) return XXH_ERROR;
-    if (seed==0) return XXH3_128bits_reset(statePtr);
-    if (seed != statePtr->seed) XXH3_initCustomSecret(statePtr->customSecret, seed);
-    XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
-    return XXH_OK;
+    return XXH3_64bits_reset_withSeed(statePtr, seed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_update(XXH3_state_t* state, const void* input, size_t len)
+XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
     return XXH3_update(state, (const xxh_u8*)input, len,
-                       XXH3_accumulate_512, XXH3_scrambleAcc);
+                       XXH3_accumulate, XXH3_scrambleAcc);
 }
 
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
 {
     const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
     if (state->totalLen > XXH3_MIDSIZE_MAX) {
@@ -5271,13 +6214,13 @@ XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state)
     return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
                                    secret, state->secretLimit + XXH_STRIPE_LEN);
 }
-
+#endif /* !XXH_NO_STREAM */
 /* 128-bit utility functions */
 
 #include <string.h>   /* memcmp, memcpy */
 
 /* return : 1 is equal, 0 if different */
-/*! @ingroup xxh3_family */
+/*! @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 */
@@ -5285,11 +6228,11 @@ XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
 }
 
 /* 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(const void* h128_1, const void* h128_2)
+ * @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 const void* h128_1, XXH_NOESCAPE const void* h128_2)
 {
     XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
     XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
@@ -5301,22 +6244,22 @@ XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2)
 
 
 /*======   Canonical representation   ======*/
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API void
-XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash)
+XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash)
 {
     XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
     if (XXH_CPU_LITTLE_ENDIAN) {
         hash.high64 = XXH_swap64(hash.high64);
         hash.low64  = XXH_swap64(hash.low64);
     }
-    memcpy(dst, &hash.high64, sizeof(hash.high64));
-    memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
+    XXH_memcpy(dst, &hash.high64, sizeof(hash.high64));
+    XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH128_hashFromCanonical(const XXH128_canonical_t* src)
+XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src)
 {
     XXH128_hash_t h;
     h.high64 = XXH_readBE64(src);
@@ -5324,10 +6267,81 @@ XXH128_hashFromCanonical(const XXH128_canonical_t* src)
     return h;
 }
 
+
+
+/* ==========================================
+ * Secret generators
+ * ==========================================
+ */
+#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
+
+XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
+{
+    XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 );
+    XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize)
+{
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(secretBuffer != NULL);
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+#else
+    /* production mode, assert() are disabled */
+    if (secretBuffer == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+#endif
+
+    if (customSeedSize == 0) {
+        customSeed = XXH3_kSecret;
+        customSeedSize = XXH_SECRET_DEFAULT_SIZE;
+    }
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(customSeed != NULL);
+#else
+    if (customSeed == NULL) return XXH_ERROR;
+#endif
+
+    /* Fill secretBuffer with a copy of customSeed - repeat as needed */
+    {   size_t pos = 0;
+        while (pos < secretSize) {
+            size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize);
+            memcpy((char*)secretBuffer + pos, customSeed, toCopy);
+            pos += toCopy;
+    }   }
+
+    {   size_t const nbSeg16 = secretSize / 16;
+        size_t n;
+        XXH128_canonical_t scrambler;
+        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
+        for (n=0; n<nbSeg16; n++) {
+            XXH128_hash_t const h128 = XXH128(&scrambler, sizeof(scrambler), n);
+            XXH3_combine16((char*)secretBuffer + n*16, h128);
+        }
+        /* last segment */
+        XXH3_combine16((char*)secretBuffer + secretSize - 16, XXH128_hashFromCanonical(&scrambler));
+    }
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed)
+{
+    XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+    XXH3_initCustomSecret(secret, seed);
+    XXH_ASSERT(secretBuffer != NULL);
+    memcpy(secretBuffer, secret, XXH_SECRET_DEFAULT_SIZE);
+}
+
+
+
 /* Pop our optimization override from above */
 #if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
   && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
 #  pragma GCC pop_options
 #endif