digest-xxhash 0.2.5 → 0.2.6

data/ext/digest/xxhash/xxhash.h

@@ -716,8 +716,15 @@ XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canoni
 # define XXH_HAS_ATTRIBUTE(x) 0
 #endif
 
+/*
+ * C23 __STDC_VERSION__ number hasn't been specified yet. For now
+ * leave as `201711L` (C17 + 1).
+ * TODO: Update to correct value when its been specified.
+ */
+#define XXH_C23_VN 201711L
+
 /* C-language Attributes are added in C23. */
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
 # define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
 #else
 # define XXH_HAS_C_ATTRIBUTE(x) 0
@@ -743,6 +750,18 @@ XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canoni
 # 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
+
+
 /*!
  * @}
  * @ingroup public
@@ -813,7 +832,7 @@ typedef uint64_t XXH64_hash_t;
  * @see
  *    XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
  */
-XXH_PUBLIC_API XXH_PUREF 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
@@ -825,16 +844,16 @@ XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(const void* input, size_t length, XX
 typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
 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 void XXH64_copyState(XXH_NOESCAPE 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 XXH_PUREF XXH64_hash_t XXH64_digest (const XXH64_state_t* statePtr);
+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 XXH_PUREF 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
 
@@ -872,7 +891,7 @@ XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canoni
  *
  * XXH3 implementation is portable:
  * it has a generic C90 formulation that can be compiled on any platform,
- * all implementations generage exactly the same hash value on all platforms.
+ * all implementations generate exactly the same hash value on all platforms.
  * Starting from v0.8.0, it's also labelled "stable", meaning that
  * any future version will also generate the same hash value.
  *
@@ -902,7 +921,7 @@ XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canoni
  * @see
  *    XXH3_64bits_reset(), XXH3_64bits_update(), XXH3_64bits_digest(): Streaming version.
  */
-XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(const void* input, size_t length);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
 
 /*!
  * @brief 64-bit seeded variant of XXH3
@@ -919,7 +938,7 @@ XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(const void* input, size_t leng
  * @param length The length
  * @param seed The 64-bit seed to alter the state.
  */
-XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(const void* input, size_t length, 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.
@@ -948,7 +967,7 @@ XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(const void* input, si
  * This is not necessarily the case when using the blob of bytes directly
  * because, when hashing _small_ inputs, only a portion of the secret is employed.
  */
-XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(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   *******/
@@ -968,20 +987,20 @@ XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(const void* data, s
 typedef struct XXH3_state_s XXH3_state_t;
 XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
 XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
-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.
@@ -991,10 +1010,10 @@ 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 XXH_PUREF 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
@@ -1033,11 +1052,11 @@ typedef struct {
  * @see
  *    XXH3_128bits_reset(), XXH3_128bits_update(), XXH3_128bits_digest(): Streaming version.
  */
-XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(const void* data, size_t len);
+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(const void* data, size_t len, XXH64_hash_t seed);
+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(const void* data, size_t len, const void* secret, size_t secretSize);
+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
@@ -1053,12 +1072,12 @@ XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(const void* data,
  * 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 XXH_PUREF 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.
@@ -1079,13 +1098,13 @@ XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
  *          =0 if *h128_1 == *h128_2
  *          <0 if *h128_1  < *h128_2
  */
-XXH_PUBLIC_API XXH_PUREF 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 XXH_PUREF 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 */
@@ -1266,13 +1285,18 @@ struct XXH3_state_s {
  * Note that this doesn't prepare the state for a streaming operation,
  * it's still necessary to use XXH3_NNbits_reset*() afterwards.
  */
-#define XXH3_INITSTATE(XXH3_state_ptr)   { (XXH3_state_ptr)->seed = 0; }
+#define XXH3_INITSTATE(XXH3_state_ptr)                       \
+    do {                                                     \
+        XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
+        tmp_xxh3_state_ptr->seed = 0;                        \
+        tmp_xxh3_state_ptr->extSecret = NULL;                \
+    } while(0)
 
 
 /*!
  * simple alias to pre-selected XXH3_128bits variant
  */
-XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
 
 
 /* ===   Experimental API   === */
@@ -1329,7 +1353,7 @@ XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(const void* data, size_t len, XXH6
  *    }
  * @endcode
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(void* secretBuffer, size_t secretSize, 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);
 
 /*!
  * @brief Generate the same secret as the _withSeed() variants.
@@ -1368,7 +1392,7 @@ XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(void* secretBuffer, size_t secr
  * @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(void* secretBuffer, XXH64_hash_t seed);
+XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
 
 /*!
  * These variants generate hash values using either
@@ -1397,24 +1421,24 @@ XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_
  * because only portions of the secret are employed for small data.
  */
 XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
-XXH3_64bits_withSecretandSeed(const void* data, size_t len,
-                              const void* secret, size_t secretSize,
+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(const void* input, size_t length,
-                               const void* secret, size_t secretSize,
+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(XXH3_state_t* statePtr,
-                                    const void* secret, size_t secretSize,
+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(XXH3_state_t* statePtr,
-                                     const void* secret, size_t secretSize,
+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 */
 
@@ -1522,7 +1546,7 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  *   care, as what works on one compiler/platform/optimization level may cause
  *   another to read garbage data or even crash.
  *
- * See http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html 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)
  */
@@ -1608,6 +1632,23 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  */
 #  define XXH_NO_INLINE_HINTS 0
 
+/*!
+ * @def XXH3_INLINE_SECRET
+ * @brief Determines whether to inline the XXH3 withSecret code.
+ *
+ * When the secret size is known, the compiler can improve the performance
+ * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret().
+ *
+ * However, if the secret size is not known, it doesn't have any benefit. This
+ * happens when xxHash is compiled into a global symbol. Therefore, if
+ * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0.
+ *
+ * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers
+ * that are *sometimes* force inline on -Og, and it is impossible to automatically
+ * detect this optimization level.
+ */
+#  define XXH3_INLINE_SECRET 0
+
 /*!
  * @def XXH32_ENDJMP
  * @brief Whether to use a jump for `XXH32_finalize`.
@@ -1682,6 +1723,15 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
 #  endif
 #endif
 
+#ifndef XXH3_INLINE_SECRET
+#  if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \
+     || !defined(XXH_INLINE_ALL)
+#    define XXH3_INLINE_SECRET 0
+#  else
+#    define XXH3_INLINE_SECRET 1
+#  endif
+#endif
+
 #ifndef XXH32_ENDJMP
 /* generally preferable for performance */
 #  define XXH32_ENDJMP 0
@@ -1778,6 +1828,11 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size)
 #  define XXH_NO_INLINE static
 #endif
 
+#if XXH3_INLINE_SECRET
+#  define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
+#else
+#  define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
+#endif
 
 
 /* *************************************
@@ -1803,7 +1858,7 @@ 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 */
@@ -1835,11 +1890,17 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size)
  * XXH3_initCustomSecret_scalar().
  */
 #if defined(__GNUC__) || defined(__clang__)
-#  define XXH_COMPILER_GUARD(var) __asm__ __volatile__("" : "+r" (var))
+#  define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
 #else
 #  define XXH_COMPILER_GUARD(var) ((void)0)
 #endif
 
+#if defined(__clang__)
+#  define XXH_COMPILER_GUARD_W(var) __asm__("" : "+w" (var))
+#else
+#  define XXH_COMPILER_GUARD_W(var) ((void)0)
+#endif
+
 /* *************************************
 *  Basic Types
 ***************************************/
@@ -1946,7 +2007,7 @@ static xxh_u32 XXH_read32(const void* ptr)
 
 /*
  * Portable and safe solution. Generally efficient.
- * see: http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
  */
 static xxh_u32 XXH_read32(const void* memPtr)
 {
@@ -2022,6 +2083,51 @@ static int XXH_isLittleEndian(void)
 #  define XXH_HAS_BUILTIN(x) 0
 #endif
 
+
+
+/*
+ * C23 and future versions have standard "unreachable()".
+ * Once it has been implemented reliably we can add it as an
+ * additional case:
+ *
+ * ```
+ * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN)
+ * #  include <stddef.h>
+ * #  ifdef unreachable
+ * #    define XXH_UNREACHABLE() unreachable()
+ * #  endif
+ * #endif
+ * ```
+ *
+ * Note C++23 also has std::unreachable() which can be detected
+ * as follows:
+ * ```
+ * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L)
+ * #  include <utility>
+ * #  define XXH_UNREACHABLE() std::unreachable()
+ * #endif
+ * ```
+ * NB: `__cpp_lib_unreachable` is defined in the `<version>` header.
+ * We don't use that as including `<utility>` in `extern "C"` blocks
+ * doesn't work on GCC12
+ */
+
+#if XXH_HAS_BUILTIN(__builtin_unreachable)
+#  define XXH_UNREACHABLE() __builtin_unreachable()
+
+#elif defined(_MSC_VER)
+#  define XXH_UNREACHABLE() __assume(0)
+
+#else
+#  define XXH_UNREACHABLE()
+#endif
+
+#if XXH_HAS_BUILTIN(__builtin_assume)
+#  define XXH_ASSUME(c) __builtin_assume(c)
+#else
+#  define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
+#endif
+
 /*!
  * @internal
  * @def XXH_rotl32(x,r)
@@ -2211,9 +2317,9 @@ static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
      *   can load data, while v3 can multiply. SSE forces them to operate
      *   together.
      *
-     * This is also enabled on AArch64, as Clang autovectorizes it incorrectly
-     * and it is pointless writing a NEON implementation that is basically the
-     * same speed as scalar for XXH32.
+     * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
+     * the loop. NEON is only faster on the A53, and with the newer cores, it is less
+     * than half the speed.
      */
     XXH_COMPILER_GUARD(acc);
 #endif
@@ -2288,41 +2394,41 @@ XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
     } else {
          switch(len&15) /* or switch(bEnd - p) */ {
            case 12:      XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 8:       XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 4:       XXH_PROCESS4;
                          return XXH32_avalanche(hash);
 
            case 13:      XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 9:       XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 5:       XXH_PROCESS4;
                          XXH_PROCESS1;
                          return XXH32_avalanche(hash);
 
            case 14:      XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 10:      XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 6:       XXH_PROCESS4;
                          XXH_PROCESS1;
                          XXH_PROCESS1;
                          return XXH32_avalanche(hash);
 
            case 15:      XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 11:      XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 7:       XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 3:       XXH_PROCESS1;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 2:       XXH_PROCESS1;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 1:       XXH_PROCESS1;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 0:       return XXH32_avalanche(hash);
         }
         XXH_ASSERT(0);
@@ -2590,7 +2696,7 @@ static xxh_u64 XXH_read64(const void* ptr)
 
 /*
  * Portable and safe solution. Generally efficient.
- * see: http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
  */
 static xxh_u64 XXH_read64(const void* memPtr)
 {
@@ -2823,7 +2929,7 @@ 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)
+XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
 #if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
     /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
@@ -2857,13 +2963,13 @@ XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
 }
 
 /*! @ingroup XXH64_family */
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t* srcState)
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
 {
     XXH_memcpy(dstState, srcState, sizeof(*dstState));
 }
 
 /*! @ingroup XXH64_family */
-XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed)
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
 {
     XXH_ASSERT(statePtr != NULL);
     memset(statePtr, 0, sizeof(*statePtr));
@@ -2876,7 +2982,7 @@ XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t s
 
 /*! @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) {
         XXH_ASSERT(len == 0);
@@ -2927,7 +3033,7 @@ 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)
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
 {
     xxh_u64 h64;
 
@@ -2950,7 +3056,7 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_digest(const XXH64_state_t* state)
 /******* Canonical representation   *******/
 
 /*! @ingroup XXH64_family */
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash)
+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);
@@ -2958,7 +3064,7 @@ XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t
 }
 
 /*! @ingroup XXH64_family */
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
 {
     return XXH_readBE64(src);
 }
@@ -2979,11 +3085,19 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
 /* ===   Compiler specifics   === */
 
 #if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
-#  define XXH_RESTRICT /* disable */
+#  define XXH_RESTRICT   /* disable */
 #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
 #  define XXH_RESTRICT   restrict
+#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \
+   || (defined (__clang__)) \
+   || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \
+   || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300))
+/*
+ * There are a LOT more compilers that recognize __restrict but this
+ * covers the major ones.
+ */
+#  define XXH_RESTRICT   __restrict
 #else
-/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */
 #  define XXH_RESTRICT   /* disable */
 #endif
 
@@ -2998,9 +3112,12 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
 #endif
 
 #if defined(__GNUC__) || defined(__clang__)
+#  if defined(__ARM_FEATURE_SVE)
+#    include <arm_sve.h>
+#  endif
 #  if defined(__ARM_NEON__) || defined(__ARM_NEON) \
-   || defined(__aarch64__)  || defined(_M_ARM) \
-   || defined(_M_ARM64)     || defined(_M_ARM64EC)
+   || (defined(_M_ARM) && _M_ARM >= 7) \
+   || defined(_M_ARM64) || defined(_M_ARM64EC)
 #    define inline __inline__  /* circumvent a clang bug */
 #    include <arm_neon.h>
 #    undef inline
@@ -3125,12 +3242,13 @@ 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.
@@ -3146,10 +3264,13 @@ 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 ( \
+#  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 */ \
    ) && ( \
@@ -3172,6 +3293,17 @@ 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.
@@ -3191,16 +3323,26 @@ 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
 
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_ALIASING __attribute__((may_alias))
+#else
+#  define XXH_ALIASING /* nothing */
+#endif
+
 /*
  * UGLY HACK:
  * GCC usually generates the best code with -O3 for xxHash.
@@ -3229,107 +3371,16 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
 #  pragma GCC optimize("-O2")
 #endif
 
-
 #if XXH_VECTOR == XXH_NEON
+
 /*
- * NEON's setup for vmlal_u32 is a little more complicated than it is on
- * SSE2, AVX2, and VSX.
- *
- * While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an upcast.
- *
- * To do the same operation, the 128-bit 'Q' register needs to be split into
- * two 64-bit 'D' registers, performing this operation::
- *
- *   [                a                 |                 b                ]
- *            |              '---------. .--------'                |
- *            |                         x                          |
- *            |              .---------' '--------.                |
- *   [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[    a >> 32     |     b >> 32    ]
- *
- * Due to significant changes in aarch64, the fastest method for aarch64 is
- * completely different than the fastest method for ARMv7-A.
- *
- * ARMv7-A treats D registers as unions overlaying Q registers, so modifying
- * D11 will modify the high half of Q5. This is similar to how modifying AH
- * will only affect bits 8-15 of AX on x86.
- *
- * VZIP takes two registers, and puts even lanes in one register and odd lanes
- * in the other.
- *
- * On ARMv7-A, this strangely modifies both parameters in place instead of
- * taking the usual 3-operand form.
- *
- * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the
- * lower and upper halves of the Q register to end up with the high and low
- * halves where we want - all in one instruction.
- *
- *   vzip.32   d10, d11       @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] }
- *
- * Unfortunately we need inline assembly for this: Instructions modifying two
- * registers at once is not possible in GCC or Clang's IR, and they have to
- * create a copy.
+ * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3
+ * optimizes out the entire hashLong loop because of the aliasing violation.
  *
- * aarch64 requires a different approach.
- *
- * In order to make it easier to write a decent compiler for aarch64, many
- * quirks were removed, such as conditional execution.
- *
- * NEON was also affected by this.
- *
- * aarch64 cannot access the high bits of a Q-form register, and writes to a
- * D-form register zero the high bits, similar to how writes to W-form scalar
- * registers (or DWORD registers on x86_64) work.
- *
- * The formerly free vget_high intrinsics now require a vext (with a few
- * exceptions)
- *
- * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent
- * of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one
- * operand.
- *
- * The equivalent of the VZIP.32 on the lower and upper halves would be this
- * mess:
- *
- *   ext     v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] }
- *   zip1    v1.2s, v0.2s, v2.2s     // v1 = { v0[0], v2[0] }
- *   zip2    v0.2s, v0.2s, v1.2s     // v0 = { v0[1], v2[1] }
- *
- * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN):
- *
- *   shrn    v1.2s, v0.2d, #32  // v1 = (uint32x2_t)(v0 >> 32);
- *   xtn     v0.2s, v0.2d       // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF);
- *
- * This is available on ARMv7-A, but is less efficient than a single VZIP.32.
+ * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
+ * so the only option is to mark it as aliasing.
  */
-
-/*!
- * Function-like macro:
- * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t &outHi)
- * {
- *     outLo = (uint32x2_t)(in & 0xFFFFFFFF);
- *     outHi = (uint32x2_t)(in >> 32);
- *     in = UNDEFINED;
- * }
- */
-# if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \
-   && (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 */ \
-      /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 */     \
-      /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 */ \
-      __asm__("vzip.32  %e0, %f0" : "+w" (in));                                             \
-      (outLo) = vget_low_u32 (vreinterpretq_u32_u64(in));                                   \
-      (outHi) = vget_high_u32(vreinterpretq_u32_u64(in));                                   \
-   } while (0)
-# else
-#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                            \
-    do {                                                                                  \
-      (outLo) = vmovn_u64    (in);                                                        \
-      (outHi) = vshrn_n_u64  ((in), 32);                                                  \
-    } while (0)
-# endif
+typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
 
 /*!
  * @internal
@@ -3347,7 +3398,7 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
 #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;
+    return *(xxh_aliasing_uint64x2_t const *)ptr;
 }
 #else
 XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
@@ -3355,38 +3406,75 @@ XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
     return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
 }
 #endif
+
+/*!
+ * @internal
+ * @brief `vmlal_u32` on low and high halves of a vector.
+ *
+ * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with
+ * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32`
+ * with `vmlal_u32`.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* Inline assembly is the only way */
+    __asm__("umlal   %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs));
+    return acc;
+}
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* This intrinsic works as expected */
+    return vmlal_high_u32(acc, lhs, rhs);
+}
+#else
+/* Portable intrinsic versions */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
+}
+/*! @copydoc XXH_vmlal_low_u32
+ * Assume the compiler converts this to vmlal_high_u32 on aarch64 */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
+}
+#endif
+
 /*!
  * @ingroup tuning
  * @brief Controls the NEON to scalar ratio for XXH3
  *
- * On AArch64 when not optimizing for size, XXH3 will run 6 lanes using NEON and
- * 2 lanes on scalar by default.
+ * This can be set to 2, 4, 6, or 8.
  *
- * 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.
+ * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
  *
- * 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 only 2 of those
+ * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU
+ * bandwidth.
  *
- * 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 noticable on the more advanced cores like the A76 which
+ * This is even more noticeable on the more advanced cores like the Cortex-A76 which
  * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
  *
- * Therefore, @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.
+ * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes
+ * and 2 scalar lanes, which is chosen by default.
+ *
+ * This does not apply to Apple processors or 32-bit processors, which run better with
+ * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes.
  *
  * This change benefits CPUs with large micro-op buffers without negatively affecting
- * other CPUs:
+ * 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.
  *
@@ -3394,7 +3482,7 @@ XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
  */
 # ifndef XXH3_NEON_LANES
 #  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
-   && XXH_SIZE_OPT <= 0
+   && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
 #   define XXH3_NEON_LANES 6
 #  else
 #   define XXH3_NEON_LANES XXH_ACC_NB
@@ -3442,6 +3530,11 @@ typedef __vector unsigned long long xxh_u64x2;
 typedef __vector unsigned char xxh_u8x16;
 typedef __vector unsigned xxh_u32x4;
 
+/*
+ * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue.
+ */
+typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING;
+
 # ifndef XXH_VSX_BE
 #  if defined(__BIG_ENDIAN__) \
   || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
@@ -3516,6 +3609,20 @@ 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 */
@@ -3952,31 +4059,33 @@ 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. */
-        size_t i = (len - 1) / 32;
+        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);
 #endif
-        return XXH3_avalanche(acc);
+        return XXH3_avalanche(acc + acc_end);
     }
 }
 
@@ -3994,13 +4103,17 @@ 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 */
@@ -4027,11 +4140,13 @@ 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);
     }
 }
 
@@ -4047,6 +4162,47 @@ 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);
@@ -4115,7 +4271,7 @@ 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); */
@@ -4125,6 +4281,7 @@ 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.
@@ -4158,13 +4315,12 @@ XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
         /* 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));
@@ -4179,7 +4335,8 @@ 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);
 
         const __m512i* const src  = (const __m512i*) ((const void*) XXH3_kSecret);
               __m512i* const dest = (      __m512i*) customSecret;
@@ -4187,14 +4344,7 @@ XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
         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 {
-                const __m512i* cp;
-                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);
     }   }
 }
 
@@ -4230,7 +4380,7 @@ 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); */
@@ -4240,6 +4390,7 @@ 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)
@@ -4262,7 +4413,7 @@ 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));
@@ -4294,12 +4445,12 @@ XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTR
         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);
     }
 }
 
@@ -4346,6 +4497,7 @@ 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)
@@ -4431,6 +4583,16 @@ XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
  * CPU, and it also mitigates some GCC codegen issues.
  *
  * @see XXH3_NEON_LANES for configuring this and details about this optimization.
+ *
+ * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
+ * integers instead of the other platforms which mask full 64-bit vectors,
+ * so the setup is more complicated than just shifting right.
+ *
+ * Additionally, there is an optimization for 4 lanes at once noted below.
+ *
+ * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
+ * there needs to be *three* versions of the accumulate operation used
+ * for the remaining 2 lanes.
  */
 XXH_FORCE_INLINE void
 XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
@@ -4439,49 +4601,113 @@ XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
 {
     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);
-    {
-        uint64x2_t* const xacc = (uint64x2_t *) acc;
+    {   /* GCC for darwin arm64 does not like aliasing here */
+        xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
         /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
         uint8_t const* const xinput = (const uint8_t *) input;
         uint8_t const* const xsecret  = (const uint8_t *) secret;
 
         size_t i;
-        /* AArch64 uses both scalar and neon at the same time */
+        /* Scalar lanes use the normal scalarRound routine */
         for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
             XXH3_scalarRound(acc, input, secret, i);
         }
-        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
-            uint64x2_t acc_vec = xacc[i];
+        i = 0;
+        /* 4 NEON lanes at a time. */
+        for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput  + (i * 16));
+            uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput  + ((i+1) * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec_1  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t key_vec_2  = XXH_vld1q_u64(xsecret + ((i+1) * 16));
+            /* data_swap = swap(data_vec) */
+            uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
+            uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
+            uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
+
+            /*
+             * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
+             * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
+             * get one vector with the low 32 bits of each lane, and one vector
+             * with the high 32 bits of each lane.
+             *
+             * This compiles to two instructions on AArch64 and has a paired vector
+             * result, which is an artifact from ARMv7a's version which modified both
+             * vectors in place.
+             *
+             *  [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
+             *  [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
+             */
+            uint32x4x2_t unzipped = vuzpq_u32(
+                vreinterpretq_u32_u64(data_key_1),
+                vreinterpretq_u32_u64(data_key_2)
+            );
+            /* data_key_lo = data_key & 0xFFFFFFFF */
+            uint32x4_t data_key_lo = unzipped.val[0];
+            /* data_key_hi = data_key >> 32 */
+            uint32x4_t data_key_hi = unzipped.val[1];
+            /*
+             * Then, we can split the vectors horizontally and multiply which, as for most
+             * widening intrinsics, have a variant that works on both high half vectors
+             * for free on AArch64.
+             *
+             * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
+             */
+            uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
+            uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
+            /*
+             * Clang reorders
+             *    a += b * c;     // umlal   swap.2d, dkl.2s, dkh.2s
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             * to
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             *    c += b * c;     // umlal   acc.2d, dkl.2s, dkh.2s
+             *
+             * While it would make sense in theory since the addition is faster,
+             * for reasons likely related to umlal being limited to certain NEON
+             * pipelines, this is worse. A compiler guard fixes this.
+             */
+            XXH_COMPILER_GUARD_W(sum_1);
+            XXH_COMPILER_GUARD_W(sum_2);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i]   = vaddq_u64(xacc[i], sum_1);
+            xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
+        }
+        /* Operate on the remaining NEON lanes 2 at a time. */
+        for (; i < XXH3_NEON_LANES / 2; i++) {
             /* data_vec = xinput[i]; */
             uint64x2_t data_vec = XXH_vld1q_u64(xinput  + (i * 16));
             /* key_vec  = xsecret[i];  */
             uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
-            uint64x2_t data_key;
-            uint32x2_t data_key_lo, data_key_hi;
             /* acc_vec_2 = swap(data_vec) */
-            uint64x2_t acc_vec_2 = vextq_u64(data_vec, data_vec, 1);
+            uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1);
             /* data_key = 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);
-            /* 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);
-            /* xacc[i] += acc_vec_2; */
-            acc_vec = vaddq_u64 (acc_vec, acc_vec_2);
-            xacc[i] = acc_vec;
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* For two lanes, just use VMOVN and VSHRN. */
+            /* data_key_lo = data_key & 0xFFFFFFFF; */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* data_key_hi = data_key >> 32; */
+            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+            /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
+            uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi);
+            /* Same Clang workaround as before */
+            XXH_COMPILER_GUARD_W(sum);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i] = vaddq_u64 (xacc[i], sum);
         }
-
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
 
 XXH_FORCE_INLINE void
 XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
     XXH_ASSERT((((size_t)acc) & 15) == 0);
 
-    {   uint64x2_t* xacc       = (uint64x2_t*) acc;
+    {   xxh_aliasing_uint64x2_t* xacc       = (xxh_aliasing_uint64x2_t*) acc;
         uint8_t const* xsecret = (uint8_t const*) secret;
         uint32x2_t prime       = vdup_n_u32 (XXH_PRIME32_1);
 
@@ -4493,47 +4719,42 @@ XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
         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]; */
-            uint64x2_t key_vec  = XXH_vld1q_u64 (xsecret + (i * 16));
-            uint64x2_t data_key = veorq_u64     (data_vec, 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;
-            /* data_key_lo = (uint32x2_t) (xacc[i] & 0xFFFFFFFF);
-             * data_key_hi = (uint32x2_t) (xacc[i] >> 32);
-             * xacc[i] = UNDEFINED; */
-            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
-            {   /*
-                 * prod_hi = (data_key >> 32) * XXH_PRIME32_1;
-                 *
-                 * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will
-                 * incorrectly "optimize" this:
-                 *   tmp     = vmul_u32(vmovn_u64(a), vmovn_u64(b));
-                 *   shifted = vshll_n_u32(tmp, 32);
-                 * to this:
-                 *   tmp     = "vmulq_u64"(a, b); // no such thing!
-                 *   shifted = vshlq_n_u64(tmp, 32);
-                 *
-                 * However, unlike SSE, Clang lacks a 64-bit multiply routine
-                 * for NEON, and it scalarizes two 64-bit multiplies instead.
-                 *
-                 * vmull_u32 has the same timing as vmul_u32, and it avoids
-                 * this bug completely.
-                 * See https://bugs.llvm.org/show_bug.cgi?id=39967
-                 */
-                uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime);
-                /* xacc[i] = prod_hi << 32; */
-                prod_hi = vshlq_n_u64(prod_hi, 32);
-                /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */
-                xacc[i] = vmlal_u32(prod_hi, data_key_lo, prime);
-            }
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+            /*
+             * prod_hi = (data_key >> 32) * XXH_PRIME32_1;
+             *
+             * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will
+             * incorrectly "optimize" this:
+             *   tmp     = vmul_u32(vmovn_u64(a), vmovn_u64(b));
+             *   shifted = vshll_n_u32(tmp, 32);
+             * to this:
+             *   tmp     = "vmulq_u64"(a, b); // no such thing!
+             *   shifted = vshlq_n_u64(tmp, 32);
+             *
+             * However, unlike SSE, Clang lacks a 64-bit multiply routine
+             * for NEON, and it scalarizes two 64-bit multiplies instead.
+             *
+             * vmull_u32 has the same timing as vmul_u32, and it avoids
+             * this bug completely.
+             * See https://bugs.llvm.org/show_bug.cgi?id=39967
+             */
+            uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime);
+            /* xacc[i] = prod_hi << 32; */
+            prod_hi = vshlq_n_u64(prod_hi, 32);
+            /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */
+            xacc[i] = vmlal_u32(prod_hi, data_key_lo, prime);
         }
     }
 }
-
 #endif
 
 #if (XXH_VECTOR == XXH_VSX)
@@ -4544,23 +4765,23 @@ XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT secret)
 {
     /* 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_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+    xxh_u8 const* const xinput   = (xxh_u8 const*) input;   /* no alignment restriction */
+    xxh_u8 const* const xsecret  = (xxh_u8 const*) secret;    /* no alignment restriction */
     xxh_u64x2 const v32 = { 32, 32 };
     size_t i;
     for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
         /* data_vec = xinput[i]; */
-        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + i);
+        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + 16*i);
         /* key_vec = xsecret[i]; */
-        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + i);
+        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
         xxh_u64x2 const data_key = data_vec ^ key_vec;
         /* shuffled = (data_key << 32) | (data_key >> 32); */
         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);
         /* acc_vec = xacc[i]; */
-        xxh_u64x2 acc_vec        = (xxh_u64x2)vec_xl(0, xacc + 4 * i);
+        xxh_u64x2 acc_vec        = xacc[i];
         acc_vec += product;
 
         /* swap high and low halves */
@@ -4569,18 +4790,18 @@ XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
 #else
         acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
 #endif
-        /* xacc[i] = acc_vec; */
-        vec_xst((xxh_u32x4)acc_vec, 0, xacc + 4 * i);
+        xacc[i] = acc_vec;
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(vsx)
 
 XXH_FORCE_INLINE void
 XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
     XXH_ASSERT((((size_t)acc) & 15) == 0);
 
-    {         xxh_u64x2* const xacc    =       (xxh_u64x2*) acc;
-        const xxh_u64x2* const xsecret = (const xxh_u64x2*) secret;
+    {   xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+        const xxh_u8* const xsecret = (const xxh_u8*) secret;
         /* constants */
         xxh_u64x2 const v32  = { 32, 32 };
         xxh_u64x2 const v47 = { 47, 47 };
@@ -4592,7 +4813,7 @@ XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
             xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
 
             /* xacc[i] ^= xsecret[i]; */
-            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + i);
+            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
             xxh_u64x2 const data_key = data_vec ^ key_vec;
 
             /* xacc[i] *= XXH_PRIME32_1 */
@@ -4606,8 +4827,148 @@ 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 */
 
+#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__))
+/*
+ * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they
+ * emit an excess mask and a full 64-bit multiply-add (MADD X-form).
+ *
+ * While this might not seem like much, as AArch64 is a 64-bit architecture, only
+ * big Cortex designs have a full 64-bit multiplier.
+ *
+ * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit
+ * multiplies expand to 2-3 multiplies in microcode. This has a major penalty
+ * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline.
+ *
+ * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does
+ * not have this penalty and does the mask automatically.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
+{
+    xxh_u64 ret;
+    /* note: %x = 64-bit register, %w = 32-bit register */
+    __asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc));
+    return ret;
+}
+#else
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
+{
+    return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc;
+}
+#endif
+
 /*!
  * @internal
  * @brief Scalar round for @ref XXH3_accumulate_512_scalar().
@@ -4630,7 +4991,7 @@ XXH3_scalarRound(void* XXH_RESTRICT acc,
         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);
+        xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]);
     }
 }
 
@@ -4655,6 +5016,7 @@ XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
         XXH3_scalarRound(acc, input, secret, i);
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
 
 /*!
  * @internal
@@ -4706,10 +5068,10 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
     const xxh_u8* kSecretPtr = XXH3_kSecret;
     XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
 
-#if defined(__clang__) && defined(__aarch64__)
+#if defined(__GNUC__) && defined(__aarch64__)
     /*
      * UGLY HACK:
-     * Clang generates a bunch of MOV/MOVK pairs for aarch64, and they are
+     * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are
      * placed sequentially, in order, at the top of the unrolled loop.
      *
      * While MOVK is great for generating constants (2 cycles for a 64-bit
@@ -4724,7 +5086,7 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
      * ADD
      * SUB      STR
      *          STR
-     * By forcing loads from memory (as the asm line causes Clang to assume
+     * By forcing loads from memory (as the asm line causes the compiler to assume
      * that XXH3_kSecretPtr has been changed), the pipelines are used more
      * efficiently:
      *   I   L   S
@@ -4741,17 +5103,11 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
      */
     XXH_COMPILER_GUARD(kSecretPtr);
 #endif
-    /*
-     * Note: in debug mode, this overrides the asm optimization
-     * and Clang will emit MOVK chains again.
-     */
-    XXH_ASSERT(kSecretPtr == XXH3_kSecret);
-
     {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
         int i;
         for (i=0; i < nbRounds; i++) {
             /*
-             * The asm hack causes Clang to assume that kSecretPtr aliases with
+             * The asm hack causes the compiler to assume that kSecretPtr aliases with
              * customSecret, and on aarch64, this prevented LDP from merging two
              * loads together for free. Putting the loads together before the stores
              * properly generates LDP.
@@ -4764,7 +5120,7 @@ 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);
 
@@ -4772,36 +5128,48 @@ 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
 
@@ -4812,45 +5180,11 @@ typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
 #  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);
-    }
-}
-
 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;
@@ -4862,7 +5196,7 @@ 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);
     }
 
@@ -4870,12 +5204,12 @@ 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);
     }   }
 }
 
@@ -4920,12 +5254,12 @@ 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);
@@ -4939,13 +5273,15 @@ XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
  * It's important for performance to transmit secret's size (when it's static)
  * so that the compiler can properly optimize the vectorized loop.
  * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set.
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
  */
-XXH_FORCE_INLINE XXH64_hash_t
+XXH3_WITH_SECRET_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);
 }
 
 /*
@@ -4959,7 +5295,7 @@ 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);
 }
 
 /*
@@ -4976,7 +5312,7 @@ 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)
 {
@@ -4984,12 +5320,12 @@ XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
     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);
     }
 }
 
@@ -4997,12 +5333,12 @@ 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);
 }
 
 
@@ -5035,27 +5371,27 @@ 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 length)
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
 {
     return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
 }
 
 /*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSecret(const void* input, size_t length, 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, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
 }
 
 /*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSeed(const void* input, size_t length, XXH64_hash_t seed)
+XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
 {
     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(const void* input, size_t length, const void* secret, size_t secretSize, XXH64_hash_t seed)
+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);
@@ -5148,7 +5484,7 @@ XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
 
 /*! @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)
 {
     XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
 }
@@ -5182,7 +5518,7 @@ XXH3_reset_internal(XXH3_state_t* statePtr,
 
 /*! @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);
@@ -5191,7 +5527,7 @@ XXH3_64bits_reset(XXH3_state_t* statePtr)
 
 /*! @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);
@@ -5202,7 +5538,7 @@ XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t
 
 /*! @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);
@@ -5214,7 +5550,7 @@ XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
 
 /*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSecretandSeed(XXH3_state_t* statePtr, const void* secret, size_t secretSize, XXH64_hash_t seed64)
+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;
@@ -5224,31 +5560,57 @@ XXH3_64bits_reset_withSecretandSeed(XXH3_state_t* statePtr, const void* secret,
     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 */
-XXH_FORCE_INLINE void
+/*!
+ * @internal
+ * @brief Processes a large input for XXH3_update() and XXH3_digest_long().
+ *
+ * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block.
+ *
+ * @param acc                Pointer to the 8 accumulator lanes
+ * @param nbStripesSoFarPtr  In/out pointer to the number of leftover stripes in the block*
+ * @param nbStripesPerBlock  Number of stripes in a block
+ * @param input              Input pointer
+ * @param nbStripes          Number of stripes to process
+ * @param secret             Secret pointer
+ * @param secretLimit        Offset of the last block in @p secret
+ * @param f_acc              Pointer to an XXH3_accumulate implementation
+ * @param f_scramble         Pointer to an XXH3_scrambleAcc implementation
+ * @return                   Pointer past the end of @p input after processing
+ */
+XXH_FORCE_INLINE const xxh_u8 *
 XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
                     size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
                     const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
                     const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
-                    XXH3_f_accumulate_512 f_acc512,
+                    XXH3_f_accumulate f_acc,
                     XXH3_f_scrambleAcc f_scramble)
 {
-    XXH_ASSERT(nbStripes <= nbStripesPerBlock);  /* can handle max 1 scramble per invocation */
-    XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock);
-    if (nbStripesPerBlock - *nbStripesSoFarPtr <= nbStripes) {
-        /* 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_scramble(acc, secret + secretLimit);
-        XXH3_accumulate(acc, input + nbStripesToEndofBlock * XXH_STRIPE_LEN, secret, nbStripesAfterBlock, f_acc512);
-        *nbStripesSoFarPtr = nbStripesAfterBlock;
-    } else {
-        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, f_acc512);
+    const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE;
+    /* Process full blocks */
+    if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) {
+        /* Process the initial partial block... */
+        size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr;
+
+        do {
+            /* Accumulate and scramble */
+            f_acc(acc, input, initialSecret, nbStripesThisIter);
+            f_scramble(acc, secret + secretLimit);
+            input += nbStripesThisIter * XXH_STRIPE_LEN;
+            nbStripes -= nbStripesThisIter;
+            /* Then continue the loop with the full block size */
+            nbStripesThisIter = nbStripesPerBlock;
+            initialSecret = secret;
+        } while (nbStripes >= nbStripesPerBlock);
+        *nbStripesSoFarPtr = 0;
+    }
+    /* Process a partial block */
+    if (nbStripes > 0) {
+        f_acc(acc, input, initialSecret, nbStripes);
+        input += nbStripes * XXH_STRIPE_LEN;
         *nbStripesSoFarPtr += nbStripes;
     }
+    /* Return end pointer */
+    return input;
 }
 
 #ifndef XXH3_STREAM_USE_STACK
@@ -5262,7 +5624,7 @@ XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
 XXH_FORCE_INLINE XXH_errorcode
 XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
             const xxh_u8* XXH_RESTRICT input, size_t len,
-            XXH3_f_accumulate_512 f_acc512,
+            XXH3_f_accumulate f_acc,
             XXH3_f_scrambleAcc f_scramble)
 {
     if (input==NULL) {
@@ -5278,7 +5640,8 @@ XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
          * 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));
+        XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
+        XXH_memcpy(acc, state->acc, sizeof(acc));
 #else
         xxh_u64* XXH_RESTRICT const acc = state->acc;
 #endif
@@ -5286,7 +5649,7 @@ XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
         XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
 
         /* small input : just fill in tmp buffer */
-        if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) {
+        if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
             XXH_memcpy(state->buffer + state->bufferedSize, input, len);
             state->bufferedSize += (XXH32_hash_t)len;
             return XXH_OK;
@@ -5308,57 +5671,20 @@ XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
                                &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);
-
-        /* large input to consume : ingest per full block */
-        if ((size_t)(bEnd - input) > state->nbStripesPerBlock * XXH_STRIPE_LEN) {
+        if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
             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);
-                XXH3_accumulate(acc, input, secret + state->nbStripesSoFar * XXH_SECRET_CONSUME_RATE, nbStripesToEnd, f_acc512);
-                f_scramble(acc, secret + state->secretLimit);
-                state->nbStripesSoFar = 0;
-                input += nbStripesToEnd * XXH_STRIPE_LEN;
-                nbStripes -= nbStripesToEnd;
-            }
-            /* consume per entire blocks */
-            while(nbStripes >= state->nbStripesPerBlock) {
-                XXH3_accumulate(acc, input, secret, state->nbStripesPerBlock, f_acc512);
-                f_scramble(acc, secret + state->secretLimit);
-                input += state->nbStripesPerBlock * XXH_STRIPE_LEN;
-                nbStripes -= state->nbStripesPerBlock;
-            }
-            /* consume last partial block */
-            XXH3_accumulate(acc, input, secret, nbStripes, f_acc512);
-            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,
+            input = XXH3_consumeStripes(acc,
                                        &state->nbStripesSoFar, state->nbStripesPerBlock,
-                                        input, XXH3_INTERNALBUFFER_STRIPES,
-                                        secret, state->secretLimit,
-                                        f_acc512, 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);
-            }
-        }
+                                       input, nbStripes,
+                                       secret, state->secretLimit,
+                                       f_acc, f_scramble);
+            XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
 
+        }
         /* Some remaining input (always) : buffer it */
         XXH_ASSERT(input < bEnd);
         XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE);
@@ -5367,7 +5693,7 @@ XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
         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));
+        XXH_memcpy(state->acc, acc, sizeof(acc));
 #endif
     }
 
@@ -5376,10 +5702,10 @@ XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
 
 /*! @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);
 }
 
 
@@ -5388,37 +5714,40 @@ XXH3_digest_long (XXH64_hash_t* acc,
                   const XXH3_state_t* state,
                   const unsigned char* secret)
 {
+    xxh_u8 lastStripe[XXH_STRIPE_LEN];
+    const xxh_u8* lastStripePtr;
+
     /*
      * Digest on a local copy. This way, the state remains unaltered, and it can
      * continue ingesting more input afterwards.
      */
     XXH_memcpy(acc, state->acc, sizeof(state->acc));
     if (state->bufferedSize >= XXH_STRIPE_LEN) {
+        /* Consume remaining stripes then point to remaining data in buffer */
         size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
         size_t nbStripesSoFar = state->nbStripesSoFar;
         XXH3_consumeStripes(acc,
                            &nbStripesSoFar, state->nbStripesPerBlock,
                             state->buffer, nbStripes,
                             secret, state->secretLimit,
-                            XXH3_accumulate_512, XXH3_scrambleAcc);
-        /* last stripe */
-        XXH3_accumulate_512(acc,
-                            state->buffer + state->bufferedSize - XXH_STRIPE_LEN,
-                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+                            XXH3_accumulate, XXH3_scrambleAcc);
+        lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN;
     } else {  /* bufferedSize < XXH_STRIPE_LEN */
-        xxh_u8 lastStripe[XXH_STRIPE_LEN];
+        /* Copy to temp buffer */
         size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
         XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
         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);
+        lastStripePtr = lastStripe;
     }
+    /* Last stripe */
+    XXH3_accumulate_512(acc,
+                        lastStripePtr,
+                        secret + state->secretLimit - XXH_SECRET_LASTACC_START);
 }
 
 /*! @ingroup XXH3_family */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
+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) {
@@ -5631,7 +5960,7 @@ XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
 #if XXH_SIZE_OPT >= 1
         {
             /* Smaller, but slightly slower. */
-            size_t i = (len - 1) / 32;
+            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);
@@ -5669,25 +5998,34 @@ 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 */
@@ -5695,7 +6033,7 @@ 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;
@@ -5712,12 +6050,12 @@ 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);
@@ -5744,38 +6082,41 @@ XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
 {
     (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 to pass @p secretLen (when it's static)
  * to the compiler, so that it can properly optimize the vectorized loop.
+ *
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
  */
-XXH_FORCE_INLINE XXH128_hash_t
+XXH3_WITH_SECRET_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);
     }
 }
 
@@ -5788,7 +6129,7 @@ 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,
@@ -5819,7 +6160,7 @@ 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)
+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),
@@ -5828,7 +6169,7 @@ 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_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,
@@ -5837,7 +6178,7 @@ XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_
 
 /*! @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),
@@ -5846,7 +6187,7 @@ XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
 
 /*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSecretandSeed(const void* input, size_t len, const void* secret, size_t secretSize, XXH64_hash_t seed)
+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);
@@ -5855,7 +6196,7 @@ XXH3_128bits_withSecretandSeed(const void* input, size_t len, const void* secret
 
 /*! @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);
 }
@@ -5870,42 +6211,41 @@ XXH128(const void* input, size_t len, XXH64_hash_t seed)
 
 /*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset(XXH3_state_t* statePtr)
+XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
 {
     return XXH3_64bits_reset(statePtr);
 }
 
 /*! @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)
 {
     return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
 }
 
 /*! @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)
 {
     return XXH3_64bits_reset_withSeed(statePtr, seed);
 }
 
 /*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr, const void* secret, size_t secretSize, XXH64_hash_t seed)
+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 */
 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);
+    return XXH3_64bits_update(state, input, len);
 }
 
 /*! @ingroup XXH3_family */
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state)
+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) {
@@ -5947,7 +6287,7 @@ XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
  *           <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)
+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;
@@ -5961,7 +6301,7 @@ XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2)
 /*======   Canonical representation   ======*/
 /*! @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) {
@@ -5974,7 +6314,7 @@ XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash)
 
 /*! @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);
@@ -5998,7 +6338,7 @@ XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
 
 /*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_generateSecret(void* secretBuffer, size_t secretSize, const void* customSeed, size_t customSeedSize)
+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);
@@ -6043,7 +6383,7 @@ XXH3_generateSecret(void* secretBuffer, size_t secretSize, const void* customSee
 
 /*! @ingroup XXH3_family */
 XXH_PUBLIC_API void
-XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_t seed)
+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);
@@ -6071,5 +6411,5 @@ XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_t seed)
 
 
 #if defined (__cplusplus)
-}
+} /* extern "C" */
 #endif