zstd: Update to upstream version 1.5.6

Release notes:
- https://github.com/facebook/zstd/releases/tag/v1.5.6

18 files changed, 2646 insertions, 1131 deletions

diff --git a/thirdparty/zstd/common/allocations.h b/thirdparty/zstd/common/allocations.h
index a3153c4bac..5e89955010 100644
--- a/thirdparty/zstd/common/allocations.h
+++ b/thirdparty/zstd/common/allocations.h
@@ -14,7 +14,7 @@
 #define ZSTD_DEPS_NEED_MALLOC
 #include "zstd_deps.h"   /* ZSTD_malloc, ZSTD_calloc, ZSTD_free, ZSTD_memset */
 
-#include "mem.h" /* MEM_STATIC */
+#include "compiler.h" /* MEM_STATIC */
 #define ZSTD_STATIC_LINKING_ONLY
 #include "../zstd.h" /* ZSTD_customMem */
 
diff --git a/thirdparty/zstd/common/bitstream.h b/thirdparty/zstd/common/bitstream.h
index 72b0b3df22..676044989c 100644
--- a/thirdparty/zstd/common/bitstream.h
+++ b/thirdparty/zstd/common/bitstream.h
@@ -90,19 +90,20 @@ MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC);
 /*-********************************************
 *  bitStream decoding API (read backward)
 **********************************************/
+typedef size_t BitContainerType;
 typedef struct {
-    size_t   bitContainer;
+    BitContainerType bitContainer;
     unsigned bitsConsumed;
     const char* ptr;
     const char* start;
     const char* limitPtr;
 } BIT_DStream_t;
 
-typedef enum { BIT_DStream_unfinished = 0,
-               BIT_DStream_endOfBuffer = 1,
-               BIT_DStream_completed = 2,
-               BIT_DStream_overflow = 3 } BIT_DStream_status;  /* result of BIT_reloadDStream() */
-               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+typedef enum { BIT_DStream_unfinished = 0,  /* fully refilled */
+               BIT_DStream_endOfBuffer = 1, /* still some bits left in bitstream */
+               BIT_DStream_completed = 2,   /* bitstream entirely consumed, bit-exact */
+               BIT_DStream_overflow = 3     /* user requested more bits than present in bitstream */
+    } BIT_DStream_status;  /* result of BIT_reloadDStream() */
 
 MEM_STATIC size_t   BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
 MEM_STATIC size_t   BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
@@ -112,7 +113,7 @@ MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* bitD);
 
 /* Start by invoking BIT_initDStream().
 *  A chunk of the bitStream is then stored into a local register.
-*  Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t).
+*  Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (BitContainerType).
 *  You can then retrieve bitFields stored into the local register, **in reverse order**.
 *  Local register is explicitly reloaded from memory by the BIT_reloadDStream() method.
 *  A reload guarantee a minimum of ((8*sizeof(bitD->bitContainer))-7) bits when its result is BIT_DStream_unfinished.
@@ -162,7 +163,7 @@ MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC,
     return 0;
 }
 
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
+FORCE_INLINE_TEMPLATE size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
 {
 #if defined(STATIC_BMI2) && STATIC_BMI2 == 1 && !defined(ZSTD_NO_INTRINSICS)
     return  _bzhi_u64(bitContainer, nbBits);
@@ -267,22 +268,22 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
         bitD->bitContainer = *(const BYTE*)(bitD->start);
         switch(srcSize)
         {
-        case 7: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
+        case 7: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
                 ZSTD_FALLTHROUGH;
 
-        case 6: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
+        case 6: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
                 ZSTD_FALLTHROUGH;
 
-        case 5: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
+        case 5: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
                 ZSTD_FALLTHROUGH;
 
-        case 4: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[3]) << 24;
+        case 4: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[3]) << 24;
                 ZSTD_FALLTHROUGH;
 
-        case 3: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[2]) << 16;
+        case 3: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[2]) << 16;
                 ZSTD_FALLTHROUGH;
 
-        case 2: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[1]) <<  8;
+        case 2: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[1]) <<  8;
                 ZSTD_FALLTHROUGH;
 
         default: break;
@@ -297,12 +298,12 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
     return srcSize;
 }
 
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getUpperBits(size_t bitContainer, U32 const start)
+FORCE_INLINE_TEMPLATE size_t BIT_getUpperBits(BitContainerType bitContainer, U32 const start)
 {
     return bitContainer >> start;
 }
 
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getMiddleBits(size_t bitContainer, U32 const start, U32 const nbBits)
+FORCE_INLINE_TEMPLATE size_t BIT_getMiddleBits(BitContainerType bitContainer, U32 const start, U32 const nbBits)
 {
     U32 const regMask = sizeof(bitContainer)*8 - 1;
     /* if start > regMask, bitstream is corrupted, and result is undefined */
@@ -325,7 +326,7 @@ MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getMiddleBits(size_t bitContainer, U32 c
  *  On 32-bits, maxNbBits==24.
  *  On 64-bits, maxNbBits==56.
  * @return : value extracted */
-MEM_STATIC  FORCE_INLINE_ATTR size_t BIT_lookBits(const BIT_DStream_t*  bitD, U32 nbBits)
+FORCE_INLINE_TEMPLATE size_t BIT_lookBits(const BIT_DStream_t*  bitD, U32 nbBits)
 {
     /* arbitrate between double-shift and shift+mask */
 #if 1
@@ -348,7 +349,7 @@ MEM_STATIC size_t BIT_lookBitsFast(const BIT_DStream_t* bitD, U32 nbBits)
     return (bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> (((regMask+1)-nbBits) & regMask);
 }
 
-MEM_STATIC FORCE_INLINE_ATTR void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+FORCE_INLINE_TEMPLATE void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
 {
     bitD->bitsConsumed += nbBits;
 }
@@ -357,7 +358,7 @@ MEM_STATIC FORCE_INLINE_ATTR void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
  *  Read (consume) next n bits from local register and update.
  *  Pay attention to not read more than nbBits contained into local register.
  * @return : extracted value. */
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
+FORCE_INLINE_TEMPLATE size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
 {
     size_t const value = BIT_lookBits(bitD, nbBits);
     BIT_skipBits(bitD, nbBits);
@@ -374,6 +375,21 @@ MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits)
     return value;
 }
 
+/*! BIT_reloadDStream_internal() :
+ *  Simple variant of BIT_reloadDStream(), with two conditions:
+ *  1. bitstream is valid : bitsConsumed <= sizeof(bitD->bitContainer)*8
+ *  2. look window is valid after shifted down : bitD->ptr >= bitD->start
+ */
+MEM_STATIC BIT_DStream_status BIT_reloadDStream_internal(BIT_DStream_t* bitD)
+{
+    assert(bitD->bitsConsumed <= sizeof(bitD->bitContainer)*8);
+    bitD->ptr -= bitD->bitsConsumed >> 3;
+    assert(bitD->ptr >= bitD->start);
+    bitD->bitsConsumed &= 7;
+    bitD->bitContainer = MEM_readLEST(bitD->ptr);
+    return BIT_DStream_unfinished;
+}
+
 /*! BIT_reloadDStreamFast() :
  *  Similar to BIT_reloadDStream(), but with two differences:
  *  1. bitsConsumed <= sizeof(bitD->bitContainer)*8 must hold!
@@ -384,31 +400,35 @@ MEM_STATIC BIT_DStream_status BIT_reloadDStreamFast(BIT_DStream_t* bitD)
 {
     if (UNLIKELY(bitD->ptr < bitD->limitPtr))
         return BIT_DStream_overflow;
-    assert(bitD->bitsConsumed <= sizeof(bitD->bitContainer)*8);
-    bitD->ptr -= bitD->bitsConsumed >> 3;
-    bitD->bitsConsumed &= 7;
-    bitD->bitContainer = MEM_readLEST(bitD->ptr);
-    return BIT_DStream_unfinished;
+    return BIT_reloadDStream_internal(bitD);
 }
 
 /*! BIT_reloadDStream() :
  *  Refill `bitD` from buffer previously set in BIT_initDStream() .
- *  This function is safe, it guarantees it will not read beyond src buffer.
+ *  This function is safe, it guarantees it will not never beyond src buffer.
  * @return : status of `BIT_DStream_t` internal register.
  *           when status == BIT_DStream_unfinished, internal register is filled with at least 25 or 57 bits */
-MEM_STATIC FORCE_INLINE_ATTR BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
+FORCE_INLINE_TEMPLATE BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
 {
-    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* overflow detected, like end of stream */
+    /* note : once in overflow mode, a bitstream remains in this mode until it's reset */
+    if (UNLIKELY(bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))) {
+        static const BitContainerType zeroFilled = 0;
+        bitD->ptr = (const char*)&zeroFilled; /* aliasing is allowed for char */
+        /* overflow detected, erroneous scenario or end of stream: no update */
         return BIT_DStream_overflow;
+    }
+
+    assert(bitD->ptr >= bitD->start);
 
     if (bitD->ptr >= bitD->limitPtr) {
-        return BIT_reloadDStreamFast(bitD);
+        return BIT_reloadDStream_internal(bitD);
     }
     if (bitD->ptr == bitD->start) {
+        /* reached end of bitStream => no update */
         if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BIT_DStream_endOfBuffer;
         return BIT_DStream_completed;
     }
-    /* start < ptr < limitPtr */
+    /* start < ptr < limitPtr => cautious update */
     {   U32 nbBytes = bitD->bitsConsumed >> 3;
         BIT_DStream_status result = BIT_DStream_unfinished;
         if (bitD->ptr - nbBytes < bitD->start) {
diff --git a/thirdparty/zstd/common/compiler.h b/thirdparty/zstd/common/compiler.h
index 73f8d01998..31880ecbe1 100644
--- a/thirdparty/zstd/common/compiler.h
+++ b/thirdparty/zstd/common/compiler.h
@@ -11,6 +11,8 @@
 #ifndef ZSTD_COMPILER_H
 #define ZSTD_COMPILER_H
 
+#include <stddef.h>
+
 #include "portability_macros.h"
 
 /*-*******************************************************
@@ -51,12 +53,19 @@
 #  define WIN_CDECL
 #endif
 
+/* UNUSED_ATTR tells the compiler it is okay if the function is unused. */
+#if defined(__GNUC__)
+#  define UNUSED_ATTR __attribute__((unused))
+#else
+#  define UNUSED_ATTR
+#endif
+
 /**
  * FORCE_INLINE_TEMPLATE is used to define C "templates", which take constant
  * parameters. They must be inlined for the compiler to eliminate the constant
  * branches.
  */
-#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR
+#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR UNUSED_ATTR
 /**
  * HINT_INLINE is used to help the compiler generate better code. It is *not*
  * used for "templates", so it can be tweaked based on the compilers
@@ -71,14 +80,28 @@
 #if !defined(__clang__) && defined(__GNUC__) && __GNUC__ >= 4 && __GNUC_MINOR__ >= 8 && __GNUC__ < 5
 #  define HINT_INLINE static INLINE_KEYWORD
 #else
-#  define HINT_INLINE static INLINE_KEYWORD FORCE_INLINE_ATTR
+#  define HINT_INLINE FORCE_INLINE_TEMPLATE
 #endif
 
-/* UNUSED_ATTR tells the compiler it is okay if the function is unused. */
+/* "soft" inline :
+ * The compiler is free to select if it's a good idea to inline or not.
+ * The main objective is to silence compiler warnings
+ * when a defined function in included but not used.
+ *
+ * Note : this macro is prefixed `MEM_` because it used to be provided by `mem.h` unit.
+ * Updating the prefix is probably preferable, but requires a fairly large codemod,
+ * since this name is used everywhere.
+ */
+#ifndef MEM_STATIC  /* already defined in Linux Kernel mem.h */
 #if defined(__GNUC__)
-#  define UNUSED_ATTR __attribute__((unused))
+#  define MEM_STATIC static __inline UNUSED_ATTR
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#  define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+#  define MEM_STATIC static __inline
 #else
-#  define UNUSED_ATTR
+#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
 #endif
 
 /* force no inlining */
@@ -109,10 +132,10 @@
 /* prefetch
  * can be disabled, by declaring NO_PREFETCH build macro */
 #if defined(NO_PREFETCH)
-#  define PREFETCH_L1(ptr)  (void)(ptr)  /* disabled */
-#  define PREFETCH_L2(ptr)  (void)(ptr)  /* disabled */
+#  define PREFETCH_L1(ptr)  do { (void)(ptr); } while (0)  /* disabled */
+#  define PREFETCH_L2(ptr)  do { (void)(ptr); } while (0)  /* disabled */
 #else
-#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86))  /* _mm_prefetch() is not defined outside of x86/x64 */
+#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86)) && !defined(_M_ARM64EC)  /* _mm_prefetch() is not defined outside of x86/x64 */
 #    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
 #    define PREFETCH_L1(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
 #    define PREFETCH_L2(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T1)
@@ -120,24 +143,25 @@
 #    define PREFETCH_L1(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
 #    define PREFETCH_L2(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 2 /* locality */)
 #  elif defined(__aarch64__)
-#    define PREFETCH_L1(ptr)  __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr)))
-#    define PREFETCH_L2(ptr)  __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr)))
+#    define PREFETCH_L1(ptr)  do { __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr))); } while (0)
+#    define PREFETCH_L2(ptr)  do { __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr))); } while (0)
 #  else
-#    define PREFETCH_L1(ptr) (void)(ptr)  /* disabled */
-#    define PREFETCH_L2(ptr) (void)(ptr)  /* disabled */
+#    define PREFETCH_L1(ptr) do { (void)(ptr); } while (0)  /* disabled */
+#    define PREFETCH_L2(ptr) do { (void)(ptr); } while (0)  /* disabled */
 #  endif
 #endif  /* NO_PREFETCH */
 
 #define CACHELINE_SIZE 64
 
-#define PREFETCH_AREA(p, s)  {            \
-    const char* const _ptr = (const char*)(p);  \
-    size_t const _size = (size_t)(s);     \
-    size_t _pos;                          \
-    for (_pos=0; _pos<_size; _pos+=CACHELINE_SIZE) {  \
-        PREFETCH_L2(_ptr + _pos);         \
-    }                                     \
-}
+#define PREFETCH_AREA(p, s)                              \
+    do {                                                 \
+        const char* const _ptr = (const char*)(p);       \
+        size_t const _size = (size_t)(s);                \
+        size_t _pos;                                     \
+        for (_pos=0; _pos<_size; _pos+=CACHELINE_SIZE) { \
+            PREFETCH_L2(_ptr + _pos);                    \
+        }                                                \
+    } while (0)
 
 /* vectorization
  * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax,
@@ -166,9 +190,9 @@
 #endif
 
 #if __has_builtin(__builtin_unreachable) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 5)))
-#  define ZSTD_UNREACHABLE { assert(0), __builtin_unreachable(); }
+#  define ZSTD_UNREACHABLE do { assert(0), __builtin_unreachable(); } while (0)
 #else
-#  define ZSTD_UNREACHABLE { assert(0); }
+#  define ZSTD_UNREACHABLE do { assert(0); } while (0)
 #endif
 
 /* disable warnings */
@@ -281,6 +305,74 @@
 *  Sanitizer
 *****************************************************************/
 
+/**
+ * Zstd relies on pointer overflow in its decompressor.
+ * We add this attribute to functions that rely on pointer overflow.
+ */
+#ifndef ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+#  if __has_attribute(no_sanitize)
+#    if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 8
+       /* gcc < 8 only has signed-integer-overlow which triggers on pointer overflow */
+#      define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR __attribute__((no_sanitize("signed-integer-overflow")))
+#    else
+       /* older versions of clang [3.7, 5.0) will warn that pointer-overflow is ignored. */
+#      define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR __attribute__((no_sanitize("pointer-overflow")))
+#    endif
+#  else
+#    define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+#  endif
+#endif
+
+/**
+ * Helper function to perform a wrapped pointer difference without trigging
+ * UBSAN.
+ *
+ * @returns lhs - rhs with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+ptrdiff_t ZSTD_wrappedPtrDiff(unsigned char const* lhs, unsigned char const* rhs)
+{
+    return lhs - rhs;
+}
+
+/**
+ * Helper function to perform a wrapped pointer add without triggering UBSAN.
+ *
+ * @return ptr + add with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+unsigned char const* ZSTD_wrappedPtrAdd(unsigned char const* ptr, ptrdiff_t add)
+{
+    return ptr + add;
+}
+
+/**
+ * Helper function to perform a wrapped pointer subtraction without triggering
+ * UBSAN.
+ *
+ * @return ptr - sub with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+unsigned char const* ZSTD_wrappedPtrSub(unsigned char const* ptr, ptrdiff_t sub)
+{
+    return ptr - sub;
+}
+
+/**
+ * Helper function to add to a pointer that works around C's undefined behavior
+ * of adding 0 to NULL.
+ *
+ * @returns `ptr + add` except it defines `NULL + 0 == NULL`.
+ */
+MEM_STATIC
+unsigned char* ZSTD_maybeNullPtrAdd(unsigned char* ptr, ptrdiff_t add)
+{
+    return add > 0 ? ptr + add : ptr;
+}
+
 /* Issue #3240 reports an ASAN failure on an llvm-mingw build. Out of an
  * abundance of caution, disable our custom poisoning on mingw. */
 #ifdef __MINGW32__
diff --git a/thirdparty/zstd/common/cpu.h b/thirdparty/zstd/common/cpu.h
index 8bc34a36da..0e684d9ad8 100644
--- a/thirdparty/zstd/common/cpu.h
+++ b/thirdparty/zstd/common/cpu.h
@@ -35,6 +35,7 @@ MEM_STATIC ZSTD_cpuid_t ZSTD_cpuid(void) {
     U32 f7b = 0;
     U32 f7c = 0;
 #if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
+#if !defined(__clang__)
     int reg[4];
     __cpuid((int*)reg, 0);
     {
@@ -50,6 +51,41 @@ MEM_STATIC ZSTD_cpuid_t ZSTD_cpuid(void) {
             f7c = (U32)reg[2];
         }
     }
+#else
+    /* Clang compiler has a bug (fixed in https://reviews.llvm.org/D101338) in
+     * which the `__cpuid` intrinsic does not save and restore `rbx` as it needs
+     * to due to being a reserved register. So in that case, do the `cpuid`
+     * ourselves. Clang supports inline assembly anyway.
+     */
+    U32 n;
+    __asm__(
+        "pushq %%rbx\n\t"
+        "cpuid\n\t"
+        "popq %%rbx\n\t"
+        : "=a"(n)
+        : "a"(0)
+        : "rcx", "rdx");
+    if (n >= 1) {
+      U32 f1a;
+      __asm__(
+          "pushq %%rbx\n\t"
+          "cpuid\n\t"
+          "popq %%rbx\n\t"
+          : "=a"(f1a), "=c"(f1c), "=d"(f1d)
+          : "a"(1)
+          :);
+    }
+    if (n >= 7) {
+      __asm__(
+          "pushq %%rbx\n\t"
+          "cpuid\n\t"
+          "movq %%rbx, %%rax\n\t"
+          "popq %%rbx"
+          : "=a"(f7b), "=c"(f7c)
+          : "a"(7), "c"(0)
+          : "rdx");
+    }
+#endif
 #elif defined(__i386__) && defined(__PIC__) && !defined(__clang__) && defined(__GNUC__)
     /* The following block like the normal cpuid branch below, but gcc
      * reserves ebx for use of its pic register so we must specially
diff --git a/thirdparty/zstd/common/debug.c b/thirdparty/zstd/common/debug.c
index ebf7bfccfa..9d0b7d229c 100644
--- a/thirdparty/zstd/common/debug.c
+++ b/thirdparty/zstd/common/debug.c
@@ -21,4 +21,10 @@
 
 #include "debug.h"
 
+#if !defined(ZSTD_LINUX_KERNEL) || (DEBUGLEVEL>=2)
+/* We only use this when DEBUGLEVEL>=2, but we get -Werror=pedantic errors if a
+ * translation unit is empty. So remove this from Linux kernel builds, but
+ * otherwise just leave it in.
+ */
 int g_debuglevel = DEBUGLEVEL;
+#endif
diff --git a/thirdparty/zstd/common/debug.h b/thirdparty/zstd/common/debug.h
index 0e9817ea6d..a16b69e574 100644
--- a/thirdparty/zstd/common/debug.h
+++ b/thirdparty/zstd/common/debug.h
@@ -85,18 +85,27 @@ extern int g_debuglevel; /* the variable is only declared,
                             It's useful when enabling very verbose levels
                             on selective conditions (such as position in src) */
 
-#  define RAWLOG(l, ...) {                                       \
-                if (l<=g_debuglevel) {                           \
-                    ZSTD_DEBUG_PRINT(__VA_ARGS__);               \
-            }   }
-#  define DEBUGLOG(l, ...) {                                     \
-                if (l<=g_debuglevel) {                           \
-                    ZSTD_DEBUG_PRINT(__FILE__ ": " __VA_ARGS__); \
-                    ZSTD_DEBUG_PRINT(" \n");                     \
-            }   }
+#  define RAWLOG(l, ...)                   \
+    do {                                   \
+        if (l<=g_debuglevel) {             \
+            ZSTD_DEBUG_PRINT(__VA_ARGS__); \
+        }                                  \
+    } while (0)
+
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define LINE_AS_STRING TOSTRING(__LINE__)
+
+#  define DEBUGLOG(l, ...)                               \
+    do {                                                 \
+        if (l<=g_debuglevel) {                           \
+            ZSTD_DEBUG_PRINT(__FILE__ ":" LINE_AS_STRING ": " __VA_ARGS__); \
+            ZSTD_DEBUG_PRINT(" \n");                     \
+        }                                                \
+    } while (0)
 #else
-#  define RAWLOG(l, ...)      {}    /* disabled */
-#  define DEBUGLOG(l, ...)    {}    /* disabled */
+#  define RAWLOG(l, ...)   do { } while (0)    /* disabled */
+#  define DEBUGLOG(l, ...) do { } while (0)    /* disabled */
 #endif
 
 
diff --git a/thirdparty/zstd/common/error_private.h b/thirdparty/zstd/common/error_private.h
index 325daad404..0156010c74 100644
--- a/thirdparty/zstd/common/error_private.h
+++ b/thirdparty/zstd/common/error_private.h
@@ -60,8 +60,13 @@ ERR_STATIC unsigned ERR_isError(size_t code) { return (code > ERROR(maxCode)); }
 ERR_STATIC ERR_enum ERR_getErrorCode(size_t code) { if (!ERR_isError(code)) return (ERR_enum)0; return (ERR_enum) (0-code); }
 
 /* check and forward error code */
-#define CHECK_V_F(e, f) size_t const e = f; if (ERR_isError(e)) return e
-#define CHECK_F(f)   { CHECK_V_F(_var_err__, f); }
+#define CHECK_V_F(e, f)     \
+    size_t const e = f;     \
+    do {                    \
+        if (ERR_isError(e)) \
+            return e;       \
+    } while (0)
+#define CHECK_F(f)   do { CHECK_V_F(_var_err__, f); } while (0)
 
 
 /*-****************************************
@@ -95,10 +100,12 @@ void _force_has_format_string(const char *format, ...) {
  * We want to force this function invocation to be syntactically correct, but
  * we don't want to force runtime evaluation of its arguments.
  */
-#define _FORCE_HAS_FORMAT_STRING(...) \
-  if (0) { \
-    _force_has_format_string(__VA_ARGS__); \
-  }
+#define _FORCE_HAS_FORMAT_STRING(...)              \
+    do {                                           \
+        if (0) {                                   \
+            _force_has_format_string(__VA_ARGS__); \
+        }                                          \
+    } while (0)
 
 #define ERR_QUOTE(str) #str
 
@@ -109,48 +116,50 @@ void _force_has_format_string(const char *format, ...) {
  * In order to do that (particularly, printing the conditional that failed),
  * this can't just wrap RETURN_ERROR().
  */
-#define RETURN_ERROR_IF(cond, err, ...) \
-  if (cond) { \
-    RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
-           __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  }
+#define RETURN_ERROR_IF(cond, err, ...)                                        \
+    do {                                                                       \
+        if (cond) {                                                            \
+            RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s",          \
+                  __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
+            _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                             \
+            RAWLOG(3, ": " __VA_ARGS__);                                       \
+            RAWLOG(3, "\n");                                                   \
+            return ERROR(err);                                                 \
+        }                                                                      \
+    } while (0)
 
 /**
  * Unconditionally return the specified error.
  *
  * In debug modes, prints additional information.
  */
-#define RETURN_ERROR(err, ...) \
-  do { \
-    RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
-           __FILE__, __LINE__, ERR_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  } while(0);
+#define RETURN_ERROR(err, ...)                                               \
+    do {                                                                     \
+        RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
+              __FILE__, __LINE__, ERR_QUOTE(ERROR(err)));                    \
+        _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                               \
+        RAWLOG(3, ": " __VA_ARGS__);                                         \
+        RAWLOG(3, "\n");                                                     \
+        return ERROR(err);                                                   \
+    } while(0)
 
 /**
  * If the provided expression evaluates to an error code, returns that error code.
  *
  * In debug modes, prints additional information.
  */
-#define FORWARD_IF_ERROR(err, ...) \
-  do { \
-    size_t const err_code = (err); \
-    if (ERR_isError(err_code)) { \
-      RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
-             __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
-      _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-      RAWLOG(3, ": " __VA_ARGS__); \
-      RAWLOG(3, "\n"); \
-      return err_code; \
-    } \
-  } while(0);
+#define FORWARD_IF_ERROR(err, ...)                                                 \
+    do {                                                                           \
+        size_t const err_code = (err);                                             \
+        if (ERR_isError(err_code)) {                                               \
+            RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s",                 \
+                  __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
+            _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                                 \
+            RAWLOG(3, ": " __VA_ARGS__);                                           \
+            RAWLOG(3, "\n");                                                       \
+            return err_code;                                                       \
+        }                                                                          \
+    } while(0)
 
 #if defined (__cplusplus)
 }
diff --git a/thirdparty/zstd/common/fse.h b/thirdparty/zstd/common/fse.h
index 02a1f0bc53..2ae128e60d 100644
--- a/thirdparty/zstd/common/fse.h
+++ b/thirdparty/zstd/common/fse.h
@@ -229,6 +229,7 @@ If there is an error, the function will return an error code, which can be teste
 
 #endif  /* FSE_H */
 
+
 #if defined(FSE_STATIC_LINKING_ONLY) && !defined(FSE_H_FSE_STATIC_LINKING_ONLY)
 #define FSE_H_FSE_STATIC_LINKING_ONLY
 
@@ -464,13 +465,13 @@ MEM_STATIC void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* statePtr, un
     FSE_symbolCompressionTransform const symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
     const U16* const stateTable = (const U16*)(statePtr->stateTable);
     U32 const nbBitsOut  = (U32)((statePtr->value + symbolTT.deltaNbBits) >> 16);
-    BIT_addBits(bitC, statePtr->value, nbBitsOut);
+    BIT_addBits(bitC,  (size_t)statePtr->value, nbBitsOut);
     statePtr->value = stateTable[ (statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
 }
 
 MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePtr)
 {
-    BIT_addBits(bitC, statePtr->value, statePtr->stateLog);
+    BIT_addBits(bitC, (size_t)statePtr->value, statePtr->stateLog);
     BIT_flushBits(bitC);
 }
 
diff --git a/thirdparty/zstd/common/fse_decompress.c b/thirdparty/zstd/common/fse_decompress.c
index 1e1c9f92d6..0dcc4640d0 100644
--- a/thirdparty/zstd/common/fse_decompress.c
+++ b/thirdparty/zstd/common/fse_decompress.c
@@ -22,8 +22,7 @@
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
 #include "error_private.h"
-#define ZSTD_DEPS_NEED_MALLOC
-#include "zstd_deps.h"
+#include "zstd_deps.h"  /* ZSTD_memcpy */
 #include "bits.h"       /* ZSTD_highbit32 */
 
 
@@ -84,7 +83,7 @@ static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCo
                     symbolNext[s] = 1;
                 } else {
                     if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
-                    symbolNext[s] = normalizedCounter[s];
+                    symbolNext[s] = (U16)normalizedCounter[s];
         }   }   }
         ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
     }
@@ -99,8 +98,7 @@ static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCo
          * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
          * our buffer to handle the over-write.
          */
-        {
-            U64 const add = 0x0101010101010101ull;
+        {   U64 const add = 0x0101010101010101ull;
             size_t pos = 0;
             U64 sv = 0;
             U32 s;
@@ -111,9 +109,8 @@ static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCo
                 for (i = 8; i < n; i += 8) {
                     MEM_write64(spread + pos + i, sv);
                 }
-                pos += n;
-            }
-        }
+                pos += (size_t)n;
+        }   }
         /* Now we spread those positions across the table.
          * The benefit of doing it in two stages is that we avoid the
          * variable size inner loop, which caused lots of branch misses.
@@ -232,12 +229,12 @@ FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
             break;
     }   }
 
-    return op-ostart;
+    assert(op >= ostart);
+    return (size_t)(op-ostart);
 }
 
 typedef struct {
     short ncount[FSE_MAX_SYMBOL_VALUE + 1];
-    FSE_DTable dtable[1]; /* Dynamically sized */
 } FSE_DecompressWksp;
 
 
@@ -252,13 +249,18 @@ FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
     unsigned tableLog;
     unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
     FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
+    size_t const dtablePos = sizeof(FSE_DecompressWksp) / sizeof(FSE_DTable);
+    FSE_DTable* const dtable = (FSE_DTable*)workSpace + dtablePos;
 
-    DEBUG_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
+    FSE_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
     if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);
 
+    /* correct offset to dtable depends on this property */
+    FSE_STATIC_ASSERT(sizeof(FSE_DecompressWksp) % sizeof(FSE_DTable) == 0);
+
     /* normal FSE decoding mode */
-    {
-        size_t const NCountLength = FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
+    {   size_t const NCountLength =
+            FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
         if (FSE_isError(NCountLength)) return NCountLength;
         if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
         assert(NCountLength <= cSrcSize);
@@ -271,16 +273,16 @@ FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
     workSpace = (BYTE*)workSpace + sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
     wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
 
-    CHECK_F( FSE_buildDTable_internal(wksp->dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
+    CHECK_F( FSE_buildDTable_internal(dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
 
     {
-        const void* ptr = wksp->dtable;
+        const void* ptr = dtable;
         const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
         const U32 fastMode = DTableH->fastMode;
 
         /* select fast mode (static) */
-        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 1);
-        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 0);
+        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 1);
+        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 0);
     }
 }
 
diff --git a/thirdparty/zstd/common/huf.h b/thirdparty/zstd/common/huf.h
index 73d1ee5654..99bf85d6f4 100644
--- a/thirdparty/zstd/common/huf.h
+++ b/thirdparty/zstd/common/huf.h
@@ -197,9 +197,22 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void
 
 /** HUF_getNbBitsFromCTable() :
  *  Read nbBits from CTable symbolTable, for symbol `symbolValue` presumed <= HUF_SYMBOLVALUE_MAX
- *  Note 1 : is not inlined, as HUF_CElt definition is private */
+ *  Note 1 : If symbolValue > HUF_readCTableHeader(symbolTable).maxSymbolValue, returns 0
+ *  Note 2 : is not inlined, as HUF_CElt definition is private
+ */
 U32 HUF_getNbBitsFromCTable(const HUF_CElt* symbolTable, U32 symbolValue);
 
+typedef struct {
+    BYTE tableLog;
+    BYTE maxSymbolValue;
+    BYTE unused[sizeof(size_t) - 2];
+} HUF_CTableHeader;
+
+/** HUF_readCTableHeader() :
+ *  @returns The header from the CTable specifying the tableLog and the maxSymbolValue.
+ */
+HUF_CTableHeader HUF_readCTableHeader(HUF_CElt const* ctable);
+
 /*
  * HUF_decompress() does the following:
  * 1. select the decompression algorithm (X1, X2) based on pre-computed heuristics
diff --git a/thirdparty/zstd/common/mem.h b/thirdparty/zstd/common/mem.h
index 98dd47a047..096f4be519 100644
--- a/thirdparty/zstd/common/mem.h
+++ b/thirdparty/zstd/common/mem.h
@@ -31,15 +31,6 @@ extern "C" {
 #   include <stdlib.h>  /* _byteswap_ulong */
 #   include <intrin.h>  /* _byteswap_* */
 #endif
-#if defined(__GNUC__)
-#  define MEM_STATIC static __inline __attribute__((unused))
-#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
-#  define MEM_STATIC static inline
-#elif defined(_MSC_VER)
-#  define MEM_STATIC static __inline
-#else
-#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
-#endif
 
 /*-**************************************************************
 *  Basic Types
diff --git a/thirdparty/zstd/common/pool.c b/thirdparty/zstd/common/pool.c
index d5ca5a7808..3adcefc9a5 100644
--- a/thirdparty/zstd/common/pool.c
+++ b/thirdparty/zstd/common/pool.c
@@ -223,7 +223,7 @@ static int POOL_resize_internal(POOL_ctx* ctx, size_t numThreads)
     {   ZSTD_pthread_t* const threadPool = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), ctx->customMem);
         if (!threadPool) return 1;
         /* replace existing thread pool */
-        ZSTD_memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(*threadPool));
+        ZSTD_memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(ZSTD_pthread_t));
         ZSTD_customFree(ctx->threads, ctx->customMem);
         ctx->threads = threadPool;
         /* Initialize additional threads */
diff --git a/thirdparty/zstd/common/pool.h b/thirdparty/zstd/common/pool.h
index eb22ff509f..cca4de73a8 100644
--- a/thirdparty/zstd/common/pool.h
+++ b/thirdparty/zstd/common/pool.h
@@ -47,7 +47,7 @@ void POOL_joinJobs(POOL_ctx* ctx);
 /*! POOL_resize() :
  *  Expands or shrinks pool's number of threads.
  *  This is more efficient than releasing + creating a new context,
- *  since it tries to preserve and re-use existing threads.
+ *  since it tries to preserve and reuse existing threads.
  * `numThreads` must be at least 1.
  * @return : 0 when resize was successful,
  *           !0 (typically 1) if there is an error.
diff --git a/thirdparty/zstd/common/portability_macros.h b/thirdparty/zstd/common/portability_macros.h
index 8fd6ea82d1..e50314a78e 100644
--- a/thirdparty/zstd/common/portability_macros.h
+++ b/thirdparty/zstd/common/portability_macros.h
@@ -68,6 +68,8 @@
 /* Mark the internal assembly functions as hidden  */
 #ifdef __ELF__
 # define ZSTD_HIDE_ASM_FUNCTION(func) .hidden func
+#elif defined(__APPLE__)
+# define ZSTD_HIDE_ASM_FUNCTION(func) .private_extern func
 #else
 # define ZSTD_HIDE_ASM_FUNCTION(func)
 #endif
diff --git a/thirdparty/zstd/common/threading.c b/thirdparty/zstd/common/threading.c
index ca155b9b9d..25bb8b9810 100644
--- a/thirdparty/zstd/common/threading.c
+++ b/thirdparty/zstd/common/threading.c
@@ -73,10 +73,12 @@ int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
     ZSTD_thread_params_t thread_param;
     (void)unused;
 
+    if (thread==NULL) return -1;
+    *thread = NULL;
+
     thread_param.start_routine = start_routine;
     thread_param.arg = arg;
     thread_param.initialized = 0;
-    *thread = NULL;
 
     /* Setup thread initialization synchronization */
     if(ZSTD_pthread_cond_init(&thread_param.initialized_cond, NULL)) {
@@ -91,7 +93,7 @@ int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
 
     /* Spawn thread */
     *thread = (HANDLE)_beginthreadex(NULL, 0, worker, &thread_param, 0, NULL);
-    if (!thread) {
+    if (*thread==NULL) {
         ZSTD_pthread_mutex_destroy(&thread_param.initialized_mutex);
         ZSTD_pthread_cond_destroy(&thread_param.initialized_cond);
         return errno;
@@ -137,6 +139,7 @@ int ZSTD_pthread_join(ZSTD_pthread_t thread)
 
 int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t const* attr)
 {
+    assert(mutex != NULL);
     *mutex = (pthread_mutex_t*)ZSTD_malloc(sizeof(pthread_mutex_t));
     if (!*mutex)
         return 1;
@@ -145,6 +148,7 @@ int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t con
 
 int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex)
 {
+    assert(mutex != NULL);
     if (!*mutex)
         return 0;
     {
@@ -156,6 +160,7 @@ int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex)
 
 int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const* attr)
 {
+    assert(cond != NULL);
     *cond = (pthread_cond_t*)ZSTD_malloc(sizeof(pthread_cond_t));
     if (!*cond)
         return 1;
@@ -164,6 +169,7 @@ int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const*
 
 int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond)
 {
+    assert(cond != NULL);
     if (!*cond)
         return 0;
     {
diff --git a/thirdparty/zstd/common/xxhash.c b/thirdparty/zstd/common/xxhash.c
index fd237c9062..052cd52282 100644
--- a/thirdparty/zstd/common/xxhash.c
+++ b/thirdparty/zstd/common/xxhash.c
@@ -1,24 +1,18 @@
 /*
- *  xxHash - Fast Hash algorithm
- *  Copyright (c) Meta Platforms, Inc. and affiliates.
- *
- *  You can contact the author at :
- *  - xxHash homepage: https://cyan4973.github.io/xxHash/
- *  - xxHash source repository : https://github.com/Cyan4973/xxHash
+ * xxHash - Extremely Fast Hash algorithm
+ * Copyright (c) Yann Collet - Meta Platforms, Inc
  *
  * This source code is licensed under both the BSD-style license (found in the
  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
  * in the COPYING file in the root directory of this source tree).
  * You may select, at your option, one of the above-listed licenses.
-*/
-
-
+ */
 
 /*
  * xxhash.c instantiates functions defined in xxhash.h
  */
 
-#define XXH_STATIC_LINKING_ONLY   /* access advanced declarations */
-#define XXH_IMPLEMENTATION   /* access definitions */
+#define XXH_STATIC_LINKING_ONLY /* access advanced declarations */
+#define XXH_IMPLEMENTATION      /* access definitions */
 
 #include "xxhash.h"
diff --git a/thirdparty/zstd/common/xxhash.h b/thirdparty/zstd/common/xxhash.h
index b8b73290bb..e59e44267c 100644
--- a/thirdparty/zstd/common/xxhash.h
+++ b/thirdparty/zstd/common/xxhash.h
@@ -1,17 +1,15 @@
 /*
- *  xxHash - Fast Hash algorithm
- *  Copyright (c) Meta Platforms, Inc. and affiliates.
- *
- *  You can contact the author at :
- *  - xxHash homepage: https://cyan4973.github.io/xxHash/
- *  - xxHash source repository : https://github.com/Cyan4973/xxHash
+ * xxHash - Extremely Fast Hash algorithm
+ * Header File
+ * Copyright (c) Yann Collet - Meta Platforms, Inc
  *
  * This source code is licensed under both the BSD-style license (found in the
  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
  * in the COPYING file in the root directory of this source tree).
  * You may select, at your option, one of the above-listed licenses.
-*/
+ */
 
+/* Local adaptations for Zstandard */
 
 #ifndef XXH_NO_XXH3
 # define XXH_NO_XXH3
@@ -24,46 +22,210 @@
 /*!
  * @mainpage xxHash
  *
+ * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
+ * limits.
+ *
+ * It is proposed in four flavors, in three families:
+ * 1. @ref XXH32_family
+ *   - Classic 32-bit hash function. Simple, compact, and runs on almost all
+ *     32-bit and 64-bit systems.
+ * 2. @ref XXH64_family
+ *   - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
+ *     64-bit systems (but _not_ 32-bit systems).
+ * 3. @ref XXH3_family
+ *   - Modern 64-bit and 128-bit hash function family which features improved
+ *     strength and performance across the board, especially on smaller data.
+ *     It benefits greatly from SIMD and 64-bit without requiring it.
+ *
+ * Benchmarks
+ * ---
+ * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
+ * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
+ *
+ * | Hash Name            | ISA ext | Width | Large Data Speed | Small Data Velocity |
+ * | -------------------- | ------- | ----: | ---------------: | ------------------: |
+ * | XXH3_64bits()        | @b AVX2 |    64 |        59.4 GB/s |               133.1 |
+ * | MeowHash             | AES-NI  |   128 |        58.2 GB/s |                52.5 |
+ * | XXH3_128bits()       | @b AVX2 |   128 |        57.9 GB/s |               118.1 |
+ * | CLHash               | PCLMUL  |    64 |        37.1 GB/s |                58.1 |
+ * | XXH3_64bits()        | @b SSE2 |    64 |        31.5 GB/s |               133.1 |
+ * | XXH3_128bits()       | @b SSE2 |   128 |        29.6 GB/s |               118.1 |
+ * | RAM sequential read  |         |   N/A |        28.0 GB/s |                 N/A |
+ * | ahash                | AES-NI  |    64 |        22.5 GB/s |               107.2 |
+ * | City64               |         |    64 |        22.0 GB/s |                76.6 |
+ * | T1ha2                |         |    64 |        22.0 GB/s |                99.0 |
+ * | City128              |         |   128 |        21.7 GB/s |                57.7 |
+ * | FarmHash             | AES-NI  |    64 |        21.3 GB/s |                71.9 |
+ * | XXH64()              |         |    64 |        19.4 GB/s |                71.0 |
+ * | SpookyHash           |         |    64 |        19.3 GB/s |                53.2 |
+ * | Mum                  |         |    64 |        18.0 GB/s |                67.0 |
+ * | CRC32C               | SSE4.2  |    32 |        13.0 GB/s |                57.9 |
+ * | XXH32()              |         |    32 |         9.7 GB/s |                71.9 |
+ * | City32               |         |    32 |         9.1 GB/s |                66.0 |
+ * | Blake3*              | @b AVX2 |   256 |         4.4 GB/s |                 8.1 |
+ * | Murmur3              |         |    32 |         3.9 GB/s |                56.1 |
+ * | SipHash*             |         |    64 |         3.0 GB/s |                43.2 |
+ * | Blake3*              | @b SSE2 |   256 |         2.4 GB/s |                 8.1 |
+ * | HighwayHash          |         |    64 |         1.4 GB/s |                 6.0 |
+ * | FNV64                |         |    64 |         1.2 GB/s |                62.7 |
+ * | Blake2*              |         |   256 |         1.1 GB/s |                 5.1 |
+ * | SHA1*                |         |   160 |         0.8 GB/s |                 5.6 |
+ * | MD5*                 |         |   128 |         0.6 GB/s |                 7.8 |
+ * @note
+ *   - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
+ *     even though it is mandatory on x64.
+ *   - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
+ *     by modern standards.
+ *   - Small data velocity is a rough average of algorithm's efficiency for small
+ *     data. For more accurate information, see the wiki.
+ *   - More benchmarks and strength tests are found on the wiki:
+ *         https://github.com/Cyan4973/xxHash/wiki
+ *
+ * Usage
+ * ------
+ * All xxHash variants use a similar API. Changing the algorithm is a trivial
+ * substitution.
+ *
+ * @pre
+ *    For functions which take an input and length parameter, the following
+ *    requirements are assumed:
+ *    - The range from [`input`, `input + length`) is valid, readable memory.
+ *      - The only exception is if the `length` is `0`, `input` may be `NULL`.
+ *    - For C++, the objects must have the *TriviallyCopyable* property, as the
+ *      functions access bytes directly as if it was an array of `unsigned char`.
+ *
+ * @anchor single_shot_example
+ * **Single Shot**
+ *
+ * These functions are stateless functions which hash a contiguous block of memory,
+ * immediately returning the result. They are the easiest and usually the fastest
+ * option.
+ *
+ * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
+ *
+ * @code{.c}
+ *   #include <string.h>
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which hashes a null terminated string with XXH32().
+ *   XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
+ *   {
+ *       // NULL pointers are only valid if the length is zero
+ *       size_t length = (string == NULL) ? 0 : strlen(string);
+ *       return XXH32(string, length, seed);
+ *   }
+ * @endcode
+ *
+ *
+ * @anchor streaming_example
+ * **Streaming**
+ *
+ * These groups of functions allow incremental hashing of unknown size, even
+ * more than what would fit in a size_t.
+ *
+ * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
+ *
+ * @code{.c}
+ *   #include <stdio.h>
+ *   #include <assert.h>
+ *   #include "xxhash.h"
+ *   // Example for a function which hashes a FILE incrementally with XXH3_64bits().
+ *   XXH64_hash_t hashFile(FILE* f)
+ *   {
+ *       // Allocate a state struct. Do not just use malloc() or new.
+ *       XXH3_state_t* state = XXH3_createState();
+ *       assert(state != NULL && "Out of memory!");
+ *       // Reset the state to start a new hashing session.
+ *       XXH3_64bits_reset(state);
+ *       char buffer[4096];
+ *       size_t count;
+ *       // Read the file in chunks
+ *       while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
+ *           // Run update() as many times as necessary to process the data
+ *           XXH3_64bits_update(state, buffer, count);
+ *       }
+ *       // Retrieve the finalized hash. This will not change the state.
+ *       XXH64_hash_t result = XXH3_64bits_digest(state);
+ *       // Free the state. Do not use free().
+ *       XXH3_freeState(state);
+ *       return result;
+ *   }
+ * @endcode
+ *
+ * Streaming functions generate the xxHash value from an incremental input.
+ * This method is slower than single-call functions, due to state management.
+ * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+ *
+ * An XXH state must first be allocated using `XXH*_createState()`.
+ *
+ * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ *
+ * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ *
+ * The function returns an error code, with 0 meaning OK, and any other value
+ * meaning there is an error.
+ *
+ * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
+ * This function returns the nn-bits hash as an int or long long.
+ *
+ * It's still possible to continue inserting input into the hash state after a
+ * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ *
+ * When done, release the state using `XXH*_freeState()`.
+ *
+ *
+ * @anchor canonical_representation_example
+ * **Canonical Representation**
+ *
+ * The default return values from XXH functions are unsigned 32, 64 and 128 bit
+ * integers.
+ * This the simplest and fastest format for further post-processing.
+ *
+ * However, this leaves open the question of what is the order on the byte level,
+ * since little and big endian conventions will store the same number differently.
+ *
+ * The canonical representation settles this issue by mandating big-endian
+ * convention, the same convention as human-readable numbers (large digits first).
+ *
+ * When writing hash values to storage, sending them over a network, or printing
+ * them, it's highly recommended to use the canonical representation to ensure
+ * portability across a wider range of systems, present and future.
+ *
+ * The following functions allow transformation of hash values to and from
+ * canonical format.
+ *
+ * XXH32_canonicalFromHash(), XXH32_hashFromCanonical(),
+ * XXH64_canonicalFromHash(), XXH64_hashFromCanonical(),
+ * XXH128_canonicalFromHash(), XXH128_hashFromCanonical(),
+ *
+ * @code{.c}
+ *   #include <stdio.h>
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which prints XXH32_hash_t in human readable format
+ *   void printXxh32(XXH32_hash_t hash)
+ *   {
+ *       XXH32_canonical_t cano;
+ *       XXH32_canonicalFromHash(&cano, hash);
+ *       size_t i;
+ *       for(i = 0; i < sizeof(cano.digest); ++i) {
+ *           printf("%02x", cano.digest[i]);
+ *       }
+ *       printf("\n");
+ *   }
+ *
+ *   // Example for a function which converts XXH32_canonical_t to XXH32_hash_t
+ *   XXH32_hash_t convertCanonicalToXxh32(XXH32_canonical_t cano)
+ *   {
+ *       XXH32_hash_t hash = XXH32_hashFromCanonical(&cano);
+ *       return hash;
+ *   }
+ * @endcode
+ *
+ *
  * @file xxhash.h
  * xxHash prototypes and implementation
  */
-/* TODO: update */
-/* Notice extracted from xxHash homepage:
-
-xxHash is an extremely fast hash algorithm, running at RAM speed limits.
-It also successfully passes all tests from the SMHasher suite.
-
-Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo @3GHz)
-
-Name            Speed       Q.Score   Author
-xxHash          5.4 GB/s     10
-CrapWow         3.2 GB/s      2       Andrew
-MurmurHash 3a   2.7 GB/s     10       Austin Appleby
-SpookyHash      2.0 GB/s     10       Bob Jenkins
-SBox            1.4 GB/s      9       Bret Mulvey
-Lookup3         1.2 GB/s      9       Bob Jenkins
-SuperFastHash   1.2 GB/s      1       Paul Hsieh
-CityHash64      1.05 GB/s    10       Pike & Alakuijala
-FNV             0.55 GB/s     5       Fowler, Noll, Vo
-CRC32           0.43 GB/s     9
-MD5-32          0.33 GB/s    10       Ronald L. Rivest
-SHA1-32         0.28 GB/s    10
-
-Q.Score is a measure of quality of the hash function.
-It depends on successfully passing SMHasher test set.
-10 is a perfect score.
-
-Note: SMHasher's CRC32 implementation is not the fastest one.
-Other speed-oriented implementations can be faster,
-especially in combination with PCLMUL instruction:
-https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html?showComment=1552696407071#c3490092340461170735
-
-A 64-bit version, named XXH64, is available since r35.
-It offers much better speed, but for 64-bit applications only.
-Name     Speed on 64 bits    Speed on 32 bits
-XXH64       13.8 GB/s            1.9 GB/s
-XXH32        6.8 GB/s            6.0 GB/s
-*/
 
 #if defined (__cplusplus)
 extern "C" {
@@ -73,21 +235,80 @@ extern "C" {
  *  INLINE mode
  ******************************/
 /*!
- * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Gives access to internal state declaration, required for static allocation.
+ *
+ * Incompatible with dynamic linking, due to risks of ABI changes.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_STATIC_LINKING_ONLY
+/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */
+
+/*!
+ * @brief Gives access to internal definitions.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #define XXH_IMPLEMENTATION
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_IMPLEMENTATION
+/* Do not undef XXH_IMPLEMENTATION for Doxygen */
+
+/*!
+ * @brief Exposes the implementation and marks all functions as `inline`.
+ *
  * Use these build macros to inline xxhash into the target unit.
  * Inlining improves performance on small inputs, especially when the length is
  * expressed as a compile-time constant:
  *
- *      https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
+ *  https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
  *
  * It also keeps xxHash symbols private to the unit, so they are not exported.
  *
  * Usage:
+ * @code{.c}
  *     #define XXH_INLINE_ALL
  *     #include "xxhash.h"
- *
+ * @endcode
  * Do not compile and link xxhash.o as a separate object, as it is not useful.
  */
+#  define XXH_INLINE_ALL
+#  undef XXH_INLINE_ALL
+/*!
+ * @brief Exposes the implementation without marking functions as inline.
+ */
+#  define XXH_PRIVATE_API
+#  undef XXH_PRIVATE_API
+/*!
+ * @brief Emulate a namespace by transparently prefixing all symbols.
+ *
+ * If you want to include _and expose_ xxHash functions from within your own
+ * library, but also want to avoid symbol collisions with other libraries which
+ * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
+ * (therefore, avoid empty or numeric values).
+ *
+ * Note that no change is required within the calling program as long as it
+ * includes `xxhash.h`: Regular symbol names will be automatically translated
+ * by this header.
+ */
+#  define XXH_NAMESPACE /* YOUR NAME HERE */
+#  undef XXH_NAMESPACE
+#endif
+
 #if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
     && !defined(XXH_INLINE_ALL_31684351384)
    /* this section should be traversed only once */
@@ -202,21 +423,13 @@ extern "C" {
 #  undef XXHASH_H_STATIC_13879238742
 #endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */
 
-
-
 /* ****************************************************************
  *  Stable API
  *****************************************************************/
 #ifndef XXHASH_H_5627135585666179
 #define XXHASH_H_5627135585666179 1
 
-
-/*!
- * @defgroup public Public API
- * Contains details on the public xxHash functions.
- * @{
- */
-/* specific declaration modes for Windows */
+/*! @brief Marks a global symbol. */
 #if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
 #  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
 #    ifdef XXH_EXPORT
@@ -229,24 +442,6 @@ extern "C" {
 #  endif
 #endif
 
-#ifdef XXH_DOXYGEN
-/*!
- * @brief Emulate a namespace by transparently prefixing all symbols.
- *
- * If you want to include _and expose_ xxHash functions from within your own
- * library, but also want to avoid symbol collisions with other libraries which
- * may also include xxHash, you can use XXH_NAMESPACE to automatically prefix
- * any public symbol from xxhash library with the value of XXH_NAMESPACE
- * (therefore, avoid empty or numeric values).
- *
- * Note that no change is required within the calling program as long as it
- * includes `xxhash.h`: Regular symbol names will be automatically translated
- * by this header.
- */
-#  define XXH_NAMESPACE /* YOUR NAME HERE */
-#  undef XXH_NAMESPACE
-#endif
-
 #ifdef XXH_NAMESPACE
 #  define XXH_CAT(A,B) A##B
 #  define XXH_NAME2(A,B) XXH_CAT(A,B)
@@ -307,11 +502,39 @@ extern "C" {
 
 
 /* *************************************
+*  Compiler specifics
+***************************************/
+
+/* specific declaration modes for Windows */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
+#endif
+
+#if defined (__GNUC__)
+# define XXH_CONSTF  __attribute__((const))
+# define XXH_PUREF   __attribute__((pure))
+# define XXH_MALLOCF __attribute__((malloc))
+#else
+# define XXH_CONSTF  /* disable */
+# define XXH_PUREF
+# define XXH_MALLOCF
+#endif
+
+/* *************************************
 *  Version
 ***************************************/
 #define XXH_VERSION_MAJOR    0
 #define XXH_VERSION_MINOR    8
-#define XXH_VERSION_RELEASE  1
+#define XXH_VERSION_RELEASE  2
+/*! @brief Version number, encoded as two digits each */
 #define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
 
 /*!
@@ -320,16 +543,22 @@ extern "C" {
  * This is mostly useful when xxHash is compiled as a shared library,
  * since the returned value comes from the library, as opposed to header file.
  *
- * @return `XXH_VERSION_NUMBER` of the invoked library.
+ * @return @ref XXH_VERSION_NUMBER of the invoked library.
  */
-XXH_PUBLIC_API unsigned XXH_versionNumber (void);
+XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
 
 
 /* ****************************
 *  Common basic types
 ******************************/
 #include <stddef.h>   /* size_t */
-typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
+/*!
+ * @brief Exit code for the streaming API.
+ */
+typedef enum {
+    XXH_OK = 0, /*!< OK */
+    XXH_ERROR   /*!< Error */
+} XXH_errorcode;
 
 
 /*-**********************************************************************
@@ -346,44 +575,44 @@ typedef uint32_t XXH32_hash_t;
 #elif !defined (__VMS) \
   && (defined (__cplusplus) \
   || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-#   include <stdint.h>
+#   ifdef _AIX
+#     include <inttypes.h>
+#   else
+#     include <stdint.h>
+#   endif
     typedef uint32_t XXH32_hash_t;
 
 #else
 #   include <limits.h>
 #   if UINT_MAX == 0xFFFFFFFFUL
       typedef unsigned int XXH32_hash_t;
+#   elif ULONG_MAX == 0xFFFFFFFFUL
+      typedef unsigned long XXH32_hash_t;
 #   else
-#     if ULONG_MAX == 0xFFFFFFFFUL
-        typedef unsigned long XXH32_hash_t;
-#     else
-#       error "unsupported platform: need a 32-bit type"
-#     endif
+#     error "unsupported platform: need a 32-bit type"
 #   endif
 #endif
 
 /*!
  * @}
  *
- * @defgroup xxh32_family XXH32 family
+ * @defgroup XXH32_family XXH32 family
  * @ingroup public
  * Contains functions used in the classic 32-bit xxHash algorithm.
  *
  * @note
  *   XXH32 is useful for older platforms, with no or poor 64-bit performance.
- *   Note that @ref xxh3_family provides competitive speed
- *   for both 32-bit and 64-bit systems, and offers true 64/128 bit hash results.
+ *   Note that the @ref XXH3_family provides competitive speed for both 32-bit
+ *   and 64-bit systems, and offers true 64/128 bit hash results.
  *
- * @see @ref xxh64_family, @ref xxh3_family : Other xxHash families
- * @see @ref xxh32_impl for implementation details
+ * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
+ * @see @ref XXH32_impl for implementation details
  * @{
  */
 
 /*!
  * @brief Calculates the 32-bit hash of @p input using xxHash32.
  *
- * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s
- *
  * @param input The block of data to be hashed, at least @p length bytes in size.
  * @param length The length of @p input, in bytes.
  * @param seed The 32-bit seed to alter the hash's output predictably.
@@ -393,66 +622,13 @@ typedef uint32_t XXH32_hash_t;
  *   readable, contiguous memory. However, if @p length is `0`, @p input may be
  *   `NULL`. In C++, this also must be *TriviallyCopyable*.
  *
- * @return The calculated 32-bit hash value.
- *
- * @see
- *    XXH64(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
- *    Direct equivalents for the other variants of xxHash.
- * @see
- *    XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version.
- */
-XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
-
-/*!
- * Streaming functions generate the xxHash value from an incremental input.
- * This method is slower than single-call functions, due to state management.
- * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
- *
- * An XXH state must first be allocated using `XXH*_createState()`.
- *
- * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
- *
- * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
- *
- * The function returns an error code, with 0 meaning OK, and any other value
- * meaning there is an error.
- *
- * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
- * This function returns the nn-bits hash as an int or long long.
- *
- * It's still possible to continue inserting input into the hash state after a
- * digest, and generate new hash values later on by invoking `XXH*_digest()`.
- *
- * When done, release the state using `XXH*_freeState()`.
- *
- * Example code for incrementally hashing a file:
- * @code{.c}
- *    #include <stdio.h>
- *    #include <xxhash.h>
- *    #define BUFFER_SIZE 256
+ * @return The calculated 32-bit xxHash32 value.
  *
- *    // Note: XXH64 and XXH3 use the same interface.
- *    XXH32_hash_t
- *    hashFile(FILE* stream)
- *    {
- *        XXH32_state_t* state;
- *        unsigned char buf[BUFFER_SIZE];
- *        size_t amt;
- *        XXH32_hash_t hash;
- *
- *        state = XXH32_createState();       // Create a state
- *        assert(state != NULL);             // Error check here
- *        XXH32_reset(state, 0xbaad5eed);    // Reset state with our seed
- *        while ((amt = fread(buf, 1, sizeof(buf), stream)) != 0) {
- *            XXH32_update(state, buf, amt); // Hash the file in chunks
- *        }
- *        hash = XXH32_digest(state);        // Finalize the hash
- *        XXH32_freeState(state);            // Clean up
- *        return hash;
- *    }
- * @endcode
+ * @see @ref single_shot_example "Single Shot Example" for an example.
  */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
 
+#ifndef XXH_NO_STREAM
 /*!
  * @typedef struct XXH32_state_s XXH32_state_t
  * @brief The opaque state struct for the XXH32 streaming API.
@@ -464,16 +640,21 @@ typedef struct XXH32_state_s XXH32_state_t;
 /*!
  * @brief Allocates an @ref XXH32_state_t.
  *
- * Must be freed with XXH32_freeState().
- * @return An allocated XXH32_state_t on success, `NULL` on failure.
+ * @return An allocated pointer of @ref XXH32_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH32_freeState().
  */
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
+XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
 /*!
  * @brief Frees an @ref XXH32_state_t.
  *
- * Must be allocated with XXH32_createState().
  * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
- * @return XXH_OK.
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH32_createState().
+ *
  */
 XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
 /*!
@@ -489,23 +670,22 @@ XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_
 /*!
  * @brief Resets an @ref XXH32_state_t to begin a new hash.
  *
- * This function resets and seeds a state. Call it before @ref XXH32_update().
- *
  * @param statePtr The state struct to reset.
  * @param seed The 32-bit seed to alter the hash result predictably.
  *
  * @pre
  *   @p statePtr must not be `NULL`.
  *
- * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH32_update().
  */
 XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t seed);
 
 /*!
  * @brief Consumes a block of @p input to an @ref XXH32_state_t.
  *
- * Call this to incrementally consume blocks of data.
- *
  * @param statePtr The state struct to update.
  * @param input The block of data to be hashed, at least @p length bytes in size.
  * @param length The length of @p input, in bytes.
@@ -517,47 +697,32 @@ XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t
  *   readable, contiguous memory. However, if @p length is `0`, @p input may be
  *   `NULL`. In C++, this also must be *TriviallyCopyable*.
  *
- * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
  */
 XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
 
 /*!
  * @brief Returns the calculated hash value from an @ref XXH32_state_t.
  *
- * @note
- *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
- *   digest, and update again.
- *
  * @param statePtr The state struct to calculate the hash from.
  *
  * @pre
  *  @p statePtr must not be `NULL`.
  *
- * @return The calculated xxHash32 value from that state.
+ * @return The calculated 32-bit xxHash32 value from that state.
+ *
+ * @note
+ *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
  */
-XXH_PUBLIC_API XXH32_hash_t  XXH32_digest (const XXH32_state_t* statePtr);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
 /*******   Canonical representation   *******/
 
-/*
- * The default return values from XXH functions are unsigned 32 and 64 bit
- * integers.
- * This the simplest and fastest format for further post-processing.
- *
- * However, this leaves open the question of what is the order on the byte level,
- * since little and big endian conventions will store the same number differently.
- *
- * The canonical representation settles this issue by mandating big-endian
- * convention, the same convention as human-readable numbers (large digits first).
- *
- * When writing hash values to storage, sending them over a network, or printing
- * them, it's highly recommended to use the canonical representation to ensure
- * portability across a wider range of systems, present and future.
- *
- * The following functions allow transformation of hash values to and from
- * canonical format.
- */
-
 /*!
  * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
  */
@@ -568,11 +733,13 @@ typedef struct {
 /*!
  * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
  *
- * @param dst The @ref XXH32_canonical_t pointer to be stored to.
+ * @param dst  The @ref XXH32_canonical_t pointer to be stored to.
  * @param hash The @ref XXH32_hash_t to be converted.
  *
  * @pre
  *   @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
  */
 XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
 
@@ -585,44 +752,75 @@ XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t
  *   @p src must not be `NULL`.
  *
  * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
  */
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
 
 
+/*! @cond Doxygen ignores this part */
 #ifdef __has_attribute
 # define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
 #else
 # define XXH_HAS_ATTRIBUTE(x) 0
 #endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * C23 __STDC_VERSION__ number hasn't been specified yet. For now
+ * leave as `201711L` (C17 + 1).
+ * TODO: Update to correct value when its been specified.
+ */
+#define XXH_C23_VN 201711L
+/*! @endcond */
 
+/*! @cond Doxygen ignores this part */
 /* C-language Attributes are added in C23. */
-#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
 #endif
+/*! @endcond */
 
+/*! @cond Doxygen ignores this part */
 #if defined(__cplusplus) && defined(__has_cpp_attribute)
 # define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
 #else
 # define XXH_HAS_CPP_ATTRIBUTE(x) 0
 #endif
+/*! @endcond */
 
+/*! @cond Doxygen ignores this part */
 /*
-Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
-introduced in CPP17 and C23.
-CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
-C23   : https://en.cppreference.com/w/c/language/attributes/fallthrough
-*/
-#if XXH_HAS_C_ATTRIBUTE(x)
-# define XXH_FALLTHROUGH [[fallthrough]]
-#elif XXH_HAS_CPP_ATTRIBUTE(x)
+ * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
+ * introduced in CPP17 and C23.
+ * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * C23   : https://en.cppreference.com/w/c/language/attributes/fallthrough
+ */
+#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough)
 # define XXH_FALLTHROUGH [[fallthrough]]
 #elif XXH_HAS_ATTRIBUTE(__fallthrough__)
-# define XXH_FALLTHROUGH __attribute__ ((fallthrough))
+# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__))
 #else
-# define XXH_FALLTHROUGH
+# define XXH_FALLTHROUGH /* fallthrough */
 #endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_NOESCAPE for annotated pointers in public API.
+ * https://clang.llvm.org/docs/AttributeReference.html#noescape
+ * As of writing this, only supported by clang.
+ */
+#if XXH_HAS_ATTRIBUTE(noescape)
+# define XXH_NOESCAPE __attribute__((noescape))
+#else
+# define XXH_NOESCAPE
+#endif
+/*! @endcond */
+
 
 /*!
  * @}
@@ -644,7 +842,11 @@ typedef uint64_t XXH64_hash_t;
 #elif !defined (__VMS) \
   && (defined (__cplusplus) \
   || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-#  include <stdint.h>
+#   ifdef _AIX
+#     include <inttypes.h>
+#   else
+#     include <stdint.h>
+#   endif
    typedef uint64_t XXH64_hash_t;
 #else
 #  include <limits.h>
@@ -660,7 +862,7 @@ typedef uint64_t XXH64_hash_t;
 /*!
  * @}
  *
- * @defgroup xxh64_family XXH64 family
+ * @defgroup XXH64_family XXH64 family
  * @ingroup public
  * @{
  * Contains functions used in the classic 64-bit xxHash algorithm.
@@ -671,13 +873,9 @@ typedef uint64_t XXH64_hash_t;
  *   It provides better speed for systems with vector processing capabilities.
  */
 
-
 /*!
  * @brief Calculates the 64-bit hash of @p input using xxHash64.
  *
- * This function usually runs faster on 64-bit systems, but slower on 32-bit
- * systems (see benchmark).
- *
  * @param input The block of data to be hashed, at least @p length bytes in size.
  * @param length The length of @p input, in bytes.
  * @param seed The 64-bit seed to alter the hash's output predictably.
@@ -687,41 +885,145 @@ typedef uint64_t XXH64_hash_t;
  *   readable, contiguous memory. However, if @p length is `0`, @p input may be
  *   `NULL`. In C++, this also must be *TriviallyCopyable*.
  *
- * @return The calculated 64-bit hash.
+ * @return The calculated 64-bit xxHash64 value.
  *
- * @see
- *    XXH32(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
- *    Direct equivalents for the other variants of xxHash.
- * @see
- *    XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
+ * @see @ref single_shot_example "Single Shot Example" for an example.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH64(const void* input, size_t length, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
 
 /*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*!
  * @brief The opaque state struct for the XXH64 streaming API.
  *
  * @see XXH64_state_s for details.
  */
 typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
+
+/*!
+ * @brief Allocates an @ref XXH64_state_t.
+ *
+ * @return An allocated pointer of @ref XXH64_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH64_freeState().
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
+
+/*!
+ * @brief Frees an @ref XXH64_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH64_createState().
+ */
 XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dst_state, const XXH64_state_t* src_state);
 
-XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH64_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH64_digest (const XXH64_state_t* statePtr);
+/*!
+ * @brief Copies one @ref XXH64_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH64_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH64_update().
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
 
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated 64-bit xxHash64 value from that state.
+ *
+ * @note
+ *   Calling XXH64_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 /*******   Canonical representation   *******/
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH64_hash_t.
+ */
 typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src);
+
+/*!
+ * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t.
+ *
+ * @param dst The @ref XXH64_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH64_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t.
+ *
+ * @param src The @ref XXH64_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
 
 #ifndef XXH_NO_XXH3
+
 /*!
  * @}
  * ************************************************************************
- * @defgroup xxh3_family XXH3 family
+ * @defgroup XXH3_family XXH3 family
  * @ingroup public
  * @{
  *
@@ -741,16 +1043,26 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
  *
  * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
  * but does not require it.
- * Any 32-bit and 64-bit targets that can run XXH32 smoothly
- * can run XXH3 at competitive speeds, even without vector support.
- * Further details are explained in the implementation.
- *
- * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8,
- * ZVector and scalar targets. This can be controlled via the XXH_VECTOR macro.
+ * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
+ * at competitive speeds, even without vector support. Further details are
+ * explained in the implementation.
+ *
+ * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD
+ * implementations for many common platforms:
+ *   - AVX512
+ *   - AVX2
+ *   - SSE2
+ *   - ARM NEON
+ *   - WebAssembly SIMD128
+ *   - POWER8 VSX
+ *   - s390x ZVector
+ * This can be controlled via the @ref XXH_VECTOR macro, but it automatically
+ * selects the best version according to predefined macros. For the x86 family, an
+ * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c.
  *
  * XXH3 implementation is portable:
  * it has a generic C90 formulation that can be compiled on any platform,
- * all implementations 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.
  *
@@ -762,24 +1074,59 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
  *
  * The API supports one-shot hashing, streaming mode, and custom secrets.
  */
-
 /*-**********************************************************************
 *  XXH3 64-bit variant
 ************************************************************************/
 
-/* XXH3_64bits():
- * default 64-bit variant, using default secret and default seed of 0.
- * It's the fastest variant. */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* data, size_t len);
+/*!
+ * @brief Calculates 64-bit unseeded variant of XXH3 hash of @p input.
+ *
+ * @param input  The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ *   This is equivalent to @ref XXH3_64bits_withSeed() with a seed of `0`, however
+ *   it may have slightly better performance due to constant propagation of the
+ *   defaults.
+ *
+ * @see
+ *    XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
 
-/*
- * XXH3_64bits_withSeed():
- * This variant generates a custom secret on the fly
- * based on default secret altered using the `seed` value.
+/*!
+ * @brief Calculates 64-bit seeded variant of XXH3 hash of @p input.
+ *
+ * @param input  The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed   The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
  * While this operation is decently fast, note that it's not completely free.
- * Note: seed==0 produces the same results as XXH3_64bits().
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
 
 /*!
  * The bare minimum size for a custom secret.
@@ -790,27 +1137,43 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, X
  */
 #define XXH3_SECRET_SIZE_MIN 136
 
-/*
- * XXH3_64bits_withSecret():
+/*!
+ * @brief Calculates 64-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @pre
+ *   The memory between @p data and @p data + @p len must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p data may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
  * It's possible to provide any blob of bytes as a "secret" to generate the hash.
  * This makes it more difficult for an external actor to prepare an intentional collision.
- * The main condition is that secretSize *must* be large enough (>= XXH3_SECRET_SIZE_MIN).
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
  * However, the quality of the secret impacts the dispersion of the hash algorithm.
  * Therefore, the secret _must_ look like a bunch of random bytes.
  * Avoid "trivial" or structured data such as repeated sequences or a text document.
  * Whenever in doubt about the "randomness" of the blob of bytes,
- * consider employing "XXH3_generateSecret()" instead (see below).
+ * consider employing @ref XXH3_generateSecret() instead (see below).
  * It will generate a proper high entropy secret derived from the blob of bytes.
  * Another advantage of using XXH3_generateSecret() is that
  * it guarantees that all bits within the initial blob of bytes
  * will impact every bit of the output.
  * This is not necessarily the case when using the blob of bytes directly
  * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
 
 
 /*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*
  * Streaming requires state maintenance.
  * This operation costs memory and CPU.
@@ -819,40 +1182,124 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len,
  */
 
 /*!
- * @brief The state struct for the XXH3 streaming API.
+ * @brief The opaque state struct for the XXH3 streaming API.
  *
  * @see XXH3_state_s for details.
  */
 typedef struct XXH3_state_s XXH3_state_t;
-XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
 XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
-XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state);
 
-/*
- * XXH3_64bits_reset():
- * Initialize with default parameters.
- * digest will be equivalent to `XXH3_64bits()`.
+/*!
+ * @brief Copies one @ref XXH3_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(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 void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret with default parameters.
+ *   - Call this function before @ref XXH3_64bits_update().
+ *   - Digest will be equivalent to `XXH3_64bits()`.
+ *
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-/*
- * XXH3_64bits_reset_withSecret():
- * `secret` is referenced, it _must outlive_ the hash streaming session.
- * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`,
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret from `seed`.
+ *   - Call this function before @ref XXH3_64bits_update().
+ *   - Digest will be equivalent to `XXH3_64bits_withSeed()`.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   `secret` is referenced, it _must outlive_ the hash streaming session.
+ *
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
  * and the quality of produced hash values depends on secret's entropy
  * (secret's content should look like a bunch of random bytes).
  * When in doubt about the randomness of a candidate `secret`,
  * consider employing `XXH3_generateSecret()` instead (see below).
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
 
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_digest (const XXH3_state_t* statePtr);
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 64-bit hash value from that state.
+ *
+ * @note
+ *   Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t  XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
 /* note : canonical representation of XXH3 is the same as XXH64
  * since they both produce XXH64_hash_t values */
@@ -873,11 +1320,76 @@ typedef struct {
     XXH64_hash_t high64;  /*!< `value >> 64` */
 } XXH128_hash_t;
 
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+/*!
+ * @brief Calculates 128-bit unseeded variant of XXH3 of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len  The length of @p data, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
+ * for shorter inputs.
+ *
+ * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of `0`, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
+/*! @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len  The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see XXH3_128bits(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+/*!
+ * @brief Calculates 128-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
 
 /*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*
  * Streaming requires state maintenance.
  * This operation costs memory and CPU.
@@ -890,39 +1402,163 @@ XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t le
  * All reset and streaming functions have same meaning as their 64-bit counterpart.
  */
 
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret with default parameters.
+ *   - Call it before @ref XXH3_128bits_update().
+ *   - Digest will be equivalent to `XXH3_128bits()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret from `seed`.
+ *   - Call it before @ref XXH3_128bits_update().
+ *   - Digest will be equivalent to `XXH3_128bits_withSeed()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   The state struct to reset.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ */
+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_errorcode XXH3_128bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr);
+/*!
+ * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 128-bit hash value from that state.
+ *
+ * @note
+ *   Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
 /* Following helper functions make it possible to compare XXH128_hast_t values.
  * Since XXH128_hash_t is a structure, this capability is not offered by the language.
  * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
 
 /*!
- * XXH128_isEqual():
- * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
+ * @brief Check equality of two XXH128_hash_t values
+ *
+ * @param h1 The 128-bit hash value.
+ * @param h2 Another 128-bit hash value.
+ *
+ * @return `1` if `h1` and `h2` are equal.
+ * @return `0` if they are not.
  */
-XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
 
 /*!
- * XXH128_cmp():
+ * @brief Compares two @ref XXH128_hash_t
  *
  * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
  *
- * return: >0 if *h128_1  > *h128_2
- *         =0 if *h128_1 == *h128_2
- *         <0 if *h128_1  < *h128_2
+ * @param h128_1 Left-hand side value
+ * @param h128_2 Right-hand side value
+ *
+ * @return >0 if @p h128_1  > @p h128_2
+ * @return =0 if @p h128_1 == @p h128_2
+ * @return <0 if @p h128_1  < @p h128_2
  */
-XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2);
+XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
 
 
 /*******   Canonical representation   *******/
 typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
-XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash);
-XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src);
+
+
+/*!
+ * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t.
+ *
+ * @param dst  The @ref XXH128_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH128_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t.
+ *
+ * @param src The @ref XXH128_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
 
 
 #endif  /* !XXH_NO_XXH3 */
@@ -996,7 +1632,6 @@ struct XXH64_state_s {
    XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it. */
 };   /* typedef'd to XXH64_state_t */
 
-
 #ifndef XXH_NO_XXH3
 
 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
@@ -1032,6 +1667,7 @@ struct XXH64_state_s {
 #define XXH3_INTERNALBUFFER_SIZE 256
 
 /*!
+ * @internal
  * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
  *
  * This is the size used in @ref XXH3_kSecret and the seeded functions.
@@ -1064,7 +1700,7 @@ struct XXH64_state_s {
  */
 struct XXH3_state_s {
    XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
-       /*!< The 8 accumulators. Similar to `vN` in @ref XXH32_state_s::v1 and @ref XXH64_state_s */
+       /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
    XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
        /*!< Used to store a custom secret generated from a seed. */
    XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
@@ -1104,69 +1740,148 @@ 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)
 
 
-/* XXH128() :
- * simple alias to pre-selected XXH3_128bits variant
+/*!
+ * @brief Calculates the 128-bit hash of @p data using XXH3.
+ *
+ * @param data The block of data to be hashed, at least @p len bytes in size.
+ * @param len  The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p data and @p data + @p len must be valid,
+ *   readable, contiguous memory. However, if @p len is `0`, @p data may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 128-bit XXH3 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
  */
-XXH_PUBLIC_API 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   === */
 /* Symbols defined below must be considered tied to a specific library version. */
 
-/*
- * XXH3_generateSecret():
+/*!
+ * @brief Derive a high-entropy secret from any user-defined content, named customSeed.
+ *
+ * @param secretBuffer    A writable buffer for derived high-entropy secret data.
+ * @param secretSize      Size of secretBuffer, in bytes.  Must be >= XXH3_SECRET_DEFAULT_SIZE.
+ * @param customSeed      A user-defined content.
+ * @param customSeedSize  Size of customSeed, in bytes.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
  *
- * Derive a high-entropy secret from any user-defined content, named customSeed.
  * The generated secret can be used in combination with `*_withSecret()` functions.
- * The `_withSecret()` variants are useful to provide a higher level of protection than 64-bit seed,
- * as it becomes much more difficult for an external actor to guess how to impact the calculation logic.
+ * The `_withSecret()` variants are useful to provide a higher level of protection
+ * than 64-bit seed, as it becomes much more difficult for an external actor to
+ * guess how to impact the calculation logic.
  *
  * The function accepts as input a custom seed of any length and any content,
- * and derives from it a high-entropy secret of length @secretSize
- * into an already allocated buffer @secretBuffer.
- * @secretSize must be >= XXH3_SECRET_SIZE_MIN
+ * and derives from it a high-entropy secret of length @p secretSize into an
+ * already allocated buffer @p secretBuffer.
  *
  * The generated secret can then be used with any `*_withSecret()` variant.
- * Functions `XXH3_128bits_withSecret()`, `XXH3_64bits_withSecret()`,
- * `XXH3_128bits_reset_withSecret()` and `XXH3_64bits_reset_withSecret()`
+ * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(),
+ * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret()
  * are part of this list. They all accept a `secret` parameter
- * which must be large enough for implementation reasons (>= XXH3_SECRET_SIZE_MIN)
+ * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN)
  * _and_ feature very high entropy (consist of random-looking bytes).
- * These conditions can be a high bar to meet, so
- * XXH3_generateSecret() can be employed to ensure proper quality.
+ * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
+ * be employed to ensure proper quality.
+ *
+ * @p customSeed can be anything. It can have any size, even small ones,
+ * and its content can be anything, even "poor entropy" sources such as a bunch
+ * of zeroes. The resulting `secret` will nonetheless provide all required qualities.
  *
- * customSeed can be anything. It can have any size, even small ones,
- * and its content can be anything, even "poor entropy" sources such as a bunch of zeroes.
- * The resulting `secret` will nonetheless provide all required qualities.
+ * @pre
+ *   - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN
+ *   - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
  *
- * When customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ * Example code:
+ * @code{.c}
+ *    #include <stdio.h>
+ *    #include <stdlib.h>
+ *    #include <string.h>
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Hashes argv[2] using the entropy from argv[1].
+ *    int main(int argc, char* argv[])
+ *    {
+ *        char secret[XXH3_SECRET_SIZE_MIN];
+ *        if (argv != 3) { return 1; }
+ *        XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1]));
+ *        XXH64_hash_t h = XXH3_64bits_withSecret(
+ *             argv[2], strlen(argv[2]),
+ *             secret, sizeof(secret)
+ *        );
+ *        printf("%016llx\n", (unsigned long long) h);
+ *    }
+ * @endcode
  */
-XXH_PUBLIC_API 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);
 
-
-/*
- * XXH3_generateSecret_fromSeed():
- *
- * Generate the same secret as the _withSeed() variants.
+/*!
+ * @brief Generate the same secret as the _withSeed() variants.
  *
- * The resulting secret has a length of XXH3_SECRET_DEFAULT_SIZE (necessarily).
- * @secretBuffer must be already allocated, of size at least XXH3_SECRET_DEFAULT_SIZE bytes.
+ * @param secretBuffer A writable buffer of @ref XXH3_SECRET_SIZE_MIN bytes
+ * @param seed         The 64-bit seed to alter the hash result predictably.
  *
  * The generated secret can be used in combination with
  *`*_withSecret()` and `_withSecretandSeed()` variants.
- * This generator is notably useful in combination with `_withSecretandSeed()`,
- * as a way to emulate a faster `_withSeed()` variant.
+ *
+ * Example C++ `std::string` hash class:
+ * @code{.cpp}
+ *    #include <string>
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Slow, seeds each time
+ *    class HashSlow {
+ *        XXH64_hash_t seed;
+ *    public:
+ *        HashSlow(XXH64_hash_t s) : seed{s} {}
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)};
+ *        }
+ *    };
+ *    // Fast, caches the seeded secret for future uses.
+ *    class HashFast {
+ *        unsigned char secret[XXH3_SECRET_SIZE_MIN];
+ *    public:
+ *        HashFast(XXH64_hash_t s) {
+ *            XXH3_generateSecret_fromSeed(secret, seed);
+ *        }
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{
+ *                XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret))
+ *            };
+ *        }
+ *    };
+ * @endcode
  */
-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);
 
-/*
- * *_withSecretandSeed() :
+/*!
+ * @brief Calculates 64/128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed       The 64-bit seed to alter the hash result predictably.
+ *
  * These variants generate hash values using either
- * @seed for "short" keys (< XXH3_MIDSIZE_MAX = 240 bytes)
- * or @secret for "large" keys (>= XXH3_MIDSIZE_MAX).
+ * @p seed for "short" keys (< @ref XXH3_MIDSIZE_MAX = 240 bytes)
+ * or @p secret for "large" keys (>= @ref XXH3_MIDSIZE_MAX).
  *
  * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
  * `_withSeed()` has to generate the secret on the fly for "large" keys.
@@ -1175,7 +1890,7 @@ XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_
  * which requires more instructions than _withSeed() variants.
  * Therefore, _withSecretandSeed variant combines the best of both worlds.
  *
- * When @secret has been generated by XXH3_generateSecret_fromSeed(),
+ * When @p secret has been generated by XXH3_generateSecret_fromSeed(),
  * this variant produces *exactly* the same results as `_withSeed()` variant,
  * hence offering only a pure speed benefit on "large" input,
  * by skipping the need to regenerate the secret for every large input.
@@ -1184,33 +1899,71 @@ XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_
  * for example with XXH3_64bits(), which then becomes the seed,
  * and then employ both the seed and the secret in _withSecretandSeed().
  * On top of speed, an added benefit is that each bit in the secret
- * has a 50% chance to swap each bit in the output,
- * via its impact to the seed.
+ * has a 50% chance to swap each bit in the output, via its impact to the seed.
+ *
  * This is not guaranteed when using the secret directly in "small data" scenarios,
  * because only portions of the secret are employed for small data.
  */
-XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSecretandSeed(const void* data, size_t len,
-                              const void* secret, size_t secretSize,
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len,
+                              XXH_NOESCAPE const void* secret, size_t secretSize,
                               XXH64_hash_t seed);
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSecretandSeed(const void* data, size_t len,
-                               const void* secret, size_t secretSize,
+/*!
+ * @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param input      The block of data to be hashed, at least @p len bytes in size.
+ * @param length     The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed()
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
+                               XXH_NOESCAPE const void* secret, size_t secretSize,
                                XXH64_hash_t seed64);
-
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see 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);
-
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see 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 */
 
-
-#endif  /* XXH_NO_XXH3 */
+#endif  /* !XXH_NO_XXH3 */
 #endif  /* XXH_NO_LONG_LONG */
 #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
 #  define XXH_IMPLEMENTATION
@@ -1264,7 +2017,7 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
 /*!
  * @brief Define this to disable 64-bit code.
  *
- * Useful if only using the @ref xxh32_family and you have a strict C90 compiler.
+ * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
  */
 #  define XXH_NO_LONG_LONG
 #  undef XXH_NO_LONG_LONG /* don't actually */
@@ -1287,7 +2040,7 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  *     Use `memcpy()`. Safe and portable. Note that most modern compilers will
  *     eliminate the function call and treat it as an unaligned access.
  *
- *  - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((packed))`
+ *  - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))`
  *   @par
  *     Depends on compiler extensions and is therefore not portable.
  *     This method is safe _if_ your compiler supports it,
@@ -1307,7 +2060,7 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  *     inline small `memcpy()` calls, and it might also be faster on big-endian
  *     systems which lack a native byteswap instruction. However, some compilers
  *     will emit literal byteshifts even if the target supports unaligned access.
- *  .
+ *
  *
  * @warning
  *   Methods 1 and 2 rely on implementation-defined behavior. Use these with
@@ -1321,6 +2074,34 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
 #  define XXH_FORCE_MEMORY_ACCESS 0
 
 /*!
+ * @def XXH_SIZE_OPT
+ * @brief Controls how much xxHash optimizes for size.
+ *
+ * xxHash, when compiled, tends to result in a rather large binary size. This
+ * is mostly due to heavy usage to forced inlining and constant folding of the
+ * @ref XXH3_family to increase performance.
+ *
+ * However, some developers prefer size over speed. This option can
+ * significantly reduce the size of the generated code. When using the `-Os`
+ * or `-Oz` options on GCC or Clang, this is defined to 1 by default,
+ * otherwise it is defined to 0.
+ *
+ * Most of these size optimizations can be controlled manually.
+ *
+ * This is a number from 0-2.
+ *  - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed
+ *    comes first.
+ *  - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more
+ *    conservative and disables hacks that increase code size. It implies the
+ *    options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0,
+ *    and @ref XXH3_NEON_LANES == 8 if they are not already defined.
+ *  - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible.
+ *    Performance may cry. For example, the single shot functions just use the
+ *    streaming API.
+ */
+#  define XXH_SIZE_OPT 0
+
+/*!
  * @def XXH_FORCE_ALIGN_CHECK
  * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
  * and XXH64() only).
@@ -1341,9 +2122,11 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  *
  * In these cases, the alignment check can be removed by setting this macro to 0.
  * Then the code will always use unaligned memory access.
- * Align check is automatically disabled on x86, x64 & arm64,
+ * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips
  * which are platforms known to offer good unaligned memory accesses performance.
  *
+ * It is also disabled by default when @ref XXH_SIZE_OPT >= 1.
+ *
  * This option does not affect XXH3 (only XXH32 and XXH64).
  */
 #  define XXH_FORCE_ALIGN_CHECK 0
@@ -1365,12 +2148,29 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
  * compiler full control on whether to inline or not.
  *
- * When not optimizing (-O0), optimizing for size (-Os, -Oz), or using
- * -fno-inline with GCC or Clang, this will automatically be defined.
+ * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if
+ * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
  */
 #  define XXH_NO_INLINE_HINTS 0
 
 /*!
+ * @def 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`.
  *
@@ -1391,34 +2191,45 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  */
 #  define XXH_OLD_NAMES
 #  undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
+
+/*!
+ * @def XXH_NO_STREAM
+ * @brief Disables the streaming API.
+ *
+ * When xxHash is not inlined and the streaming functions are not used, disabling
+ * the streaming functions can improve code size significantly, especially with
+ * the @ref XXH3_family which tends to make constant folded copies of itself.
+ */
+#  define XXH_NO_STREAM
+#  undef XXH_NO_STREAM /* don't actually */
 #endif /* XXH_DOXYGEN */
 /*!
  * @}
  */
 
 #ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-   /* prefer __packed__ structures (method 1) for gcc on armv7+ and mips */
-#  if !defined(__clang__) && \
-( \
-    (defined(__INTEL_COMPILER) && !defined(_WIN32)) || \
-    ( \
-        defined(__GNUC__) && ( \
-            (defined(__ARM_ARCH) && __ARM_ARCH >= 7) || \
-            ( \
-                defined(__mips__) && \
-                (__mips <= 5 || __mips_isa_rev < 6) && \
-                (!defined(__mips16) || defined(__mips_mips16e2)) \
-            ) \
-        ) \
-    ) \
-)
+   /* prefer __packed__ structures (method 1) for GCC
+    * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
+    * which for some reason does unaligned loads. */
+#  if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED))
 #    define XXH_FORCE_MEMORY_ACCESS 1
 #  endif
 #endif
 
+#ifndef XXH_SIZE_OPT
+   /* default to 1 for -Os or -Oz */
+#  if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__)
+#    define XXH_SIZE_OPT 1
+#  else
+#    define XXH_SIZE_OPT 0
+#  endif
+#endif
+
 #ifndef XXH_FORCE_ALIGN_CHECK  /* can be defined externally */
-#  if defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) \
-   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64) /* visual */
+   /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */
+#  if XXH_SIZE_OPT >= 1 || \
+      defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
+   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64)    || defined(_M_ARM) /* visual */
 #    define XXH_FORCE_ALIGN_CHECK 0
 #  else
 #    define XXH_FORCE_ALIGN_CHECK 1
@@ -1426,14 +2237,22 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
 #endif
 
 #ifndef XXH_NO_INLINE_HINTS
-#  if defined(__OPTIMIZE_SIZE__) /* -Os, -Oz */ \
-   || defined(__NO_INLINE__)     /* -O0, -fno-inline */
+#  if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__)  /* -O0, -fno-inline */
 #    define XXH_NO_INLINE_HINTS 1
 #  else
 #    define XXH_NO_INLINE_HINTS 0
 #  endif
 #endif
 
+#ifndef 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
@@ -1448,13 +2267,56 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
 /* *************************************
 *  Includes & Memory related functions
 ***************************************/
-/* Modify the local functions below should you wish to use some other memory routines */
-/* for ZSTD_malloc(), ZSTD_free() */
-#define ZSTD_DEPS_NEED_MALLOC
-#include "zstd_deps.h"  /* size_t, ZSTD_malloc, ZSTD_free, ZSTD_memcpy */
-static void* XXH_malloc(size_t s) { return ZSTD_malloc(s); }
-static void  XXH_free  (void* p)  { ZSTD_free(p); }
-static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_memcpy(dest,src,size); }
+#if defined(XXH_NO_STREAM)
+/* nothing */
+#elif defined(XXH_NO_STDLIB)
+
+/* When requesting to disable any mention of stdlib,
+ * the library loses the ability to invoked malloc / free.
+ * In practice, it means that functions like `XXH*_createState()`
+ * will always fail, and return NULL.
+ * This flag is useful in situations where
+ * xxhash.h is integrated into some kernel, embedded or limited environment
+ * without access to dynamic allocation.
+ */
+
+static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
+static void XXH_free(void* p) { (void)p; }
+
+#else
+
+/*
+ * Modify the local functions below should you wish to use
+ * different memory routines for malloc() and free()
+ */
+#include <stdlib.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than malloc().
+ */
+static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than free().
+ */
+static void XXH_free(void* p) { free(p); }
+
+#endif  /* XXH_NO_STDLIB */
+
+#include <string.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than memcpy().
+ */
+static void* XXH_memcpy(void* dest, const void* src, size_t size)
+{
+    return memcpy(dest,src,size);
+}
+
+#include <limits.h>   /* ULLONG_MAX */
 
 
 /* *************************************
@@ -1487,6 +2349,11 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_
 #  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
 
 
 /* *************************************
@@ -1512,14 +2379,17 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_
 #  include <assert.h>   /* note: can still be disabled with NDEBUG */
 #  define XXH_ASSERT(c)   assert(c)
 #else
-#  define XXH_ASSERT(c)   ((void)0)
+#  if defined(__INTEL_COMPILER)
+#    define XXH_ASSERT(c)   XXH_ASSUME((unsigned char) (c))
+#  else
+#    define XXH_ASSERT(c)   XXH_ASSUME(c)
+#  endif
 #endif
 
 /* note: use after variable declarations */
 #ifndef XXH_STATIC_ASSERT
 #  if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)    /* C11 */
-#    include <assert.h>
-#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
 #  elif defined(__cplusplus) && (__cplusplus >= 201103L)            /* C++11 */
 #    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
 #  else
@@ -1534,7 +2404,7 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_
  * @brief Used to prevent unwanted optimizations for @p var.
  *
  * It uses an empty GCC inline assembly statement with a register constraint
- * which forces @p var into a general purpose register (e.g. eax, ebx, ecx
+ * which forces @p var into a general purpose register (eg eax, ebx, ecx
  * on x86) and marks it as modified.
  *
  * This is used in a few places to avoid unwanted autovectorization (e.g.
@@ -1545,18 +2415,30 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_
  * 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
 
+/* Specifically for NEON vectors which use the "w" constraint, on
+ * Clang. */
+#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__)
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var))
+#else
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0)
+#endif
+
 /* *************************************
 *  Basic Types
 ***************************************/
 #if !defined (__VMS) \
  && (defined (__cplusplus) \
  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-# include <stdint.h>
+# ifdef _AIX
+#   include <inttypes.h>
+# else
+#   include <stdint.h>
+# endif
   typedef uint8_t xxh_u8;
 #else
   typedef unsigned char xxh_u8;
@@ -1564,6 +2446,7 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_
 typedef XXH32_hash_t xxh_u32;
 
 #ifdef XXH_OLD_NAMES
+#  warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly"
 #  define BYTE xxh_u8
 #  define U8   xxh_u8
 #  define U32  xxh_u32
@@ -1637,18 +2520,19 @@ static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr;
 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
 
 /*
- * __pack instructions are safer but compiler specific, hence potentially
- * problematic for some compilers.
- *
- * Currently only defined for GCC and ICC.
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
  */
 #ifdef XXH_OLD_NAMES
 typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
 #endif
 static xxh_u32 XXH_read32(const void* ptr)
 {
-    typedef union { xxh_u32 u32; } __attribute__((packed)) xxh_unalign;
-    return ((const xxh_unalign*)ptr)->u32;
+    typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32;
+    return *((const xxh_unalign32*)ptr);
 }
 
 #else
@@ -1731,6 +2615,51 @@ 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)
@@ -1853,8 +2782,10 @@ XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
 *********************************************************************/
 /*!
  * @}
- * @defgroup xxh32_impl XXH32 implementation
+ * @defgroup XXH32_impl XXH32 implementation
  * @ingroup impl
+ *
+ * Details on the XXH32 implementation.
  * @{
  */
  /* #define instead of static const, to be used as initializers */
@@ -1888,7 +2819,7 @@ static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
     acc += input * XXH_PRIME32_2;
     acc  = XXH_rotl32(acc, 13);
     acc *= XXH_PRIME32_1;
-#if (defined(__SSE4_1__) || defined(__aarch64__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
     /*
      * UGLY HACK:
      * A compiler fence is the only thing that prevents GCC and Clang from
@@ -1918,9 +2849,12 @@ 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.
+     *
+     * Additionally, this is used on WASM SIMD128 because it JITs to the same
+     * SIMD instructions and has the same issue.
      */
     XXH_COMPILER_GUARD(acc);
 #endif
@@ -1934,17 +2868,17 @@ static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
  * The final mix ensures that all input bits have a chance to impact any bit in
  * the output digest, resulting in an unbiased distribution.
  *
- * @param h32 The hash to avalanche.
+ * @param hash The hash to avalanche.
  * @return The avalanched hash.
  */
-static xxh_u32 XXH32_avalanche(xxh_u32 h32)
+static xxh_u32 XXH32_avalanche(xxh_u32 hash)
 {
-    h32 ^= h32 >> 15;
-    h32 *= XXH_PRIME32_2;
-    h32 ^= h32 >> 13;
-    h32 *= XXH_PRIME32_3;
-    h32 ^= h32 >> 16;
-    return(h32);
+    hash ^= hash >> 15;
+    hash *= XXH_PRIME32_2;
+    hash ^= hash >> 13;
+    hash *= XXH_PRIME32_3;
+    hash ^= hash >> 16;
+    return hash;
 }
 
 #define XXH_get32bits(p) XXH_readLE32_align(p, align)
@@ -1957,24 +2891,25 @@ static xxh_u32 XXH32_avalanche(xxh_u32 h32)
  * This final stage will digest them to ensure that all input bytes are present
  * in the final mix.
  *
- * @param h32 The hash to finalize.
+ * @param hash The hash to finalize.
  * @param ptr The pointer to the remaining input.
  * @param len The remaining length, modulo 16.
  * @param align Whether @p ptr is aligned.
  * @return The finalized hash.
+ * @see XXH64_finalize().
  */
-static xxh_u32
-XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
+static XXH_PUREF xxh_u32
+XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
 {
-#define XXH_PROCESS1 do {                           \
-    h32 += (*ptr++) * XXH_PRIME32_5;                \
-    h32 = XXH_rotl32(h32, 11) * XXH_PRIME32_1;      \
+#define XXH_PROCESS1 do {                             \
+    hash += (*ptr++) * XXH_PRIME32_5;                 \
+    hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1;      \
 } while (0)
 
-#define XXH_PROCESS4 do {                           \
-    h32 += XXH_get32bits(ptr) * XXH_PRIME32_3;      \
-    ptr += 4;                                   \
-    h32  = XXH_rotl32(h32, 17) * XXH_PRIME32_4;     \
+#define XXH_PROCESS4 do {                             \
+    hash += XXH_get32bits(ptr) * XXH_PRIME32_3;       \
+    ptr += 4;                                         \
+    hash  = XXH_rotl32(hash, 17) * XXH_PRIME32_4;     \
 } while (0)
 
     if (ptr==NULL) XXH_ASSERT(len == 0);
@@ -1990,49 +2925,49 @@ XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
             XXH_PROCESS1;
             --len;
         }
-        return XXH32_avalanche(h32);
+        return XXH32_avalanche(hash);
     } else {
          switch(len&15) /* or switch(bEnd - p) */ {
            case 12:      XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 8:       XXH_PROCESS4;
-                         XXH_FALLTHROUGH;
+                         XXH_FALLTHROUGH;  /* fallthrough */
            case 4:       XXH_PROCESS4;
-                         return XXH32_avalanche(h32);
+                         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(h32);
+                         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(h32);
+                         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;
-           case 0:       return XXH32_avalanche(h32);
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 0:       return XXH32_avalanche(hash);
         }
         XXH_ASSERT(0);
-        return h32;   /* reaching this point is deemed impossible */
+        return hash;   /* reaching this point is deemed impossible */
     }
 }
 
@@ -2052,7 +2987,7 @@ XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
  * @param align Whether @p input is aligned.
  * @return The calculated hash.
  */
-XXH_FORCE_INLINE xxh_u32
+XXH_FORCE_INLINE XXH_PUREF xxh_u32
 XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
 {
     xxh_u32 h32;
@@ -2085,10 +3020,10 @@ XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment
     return XXH32_finalize(h32, input, len&15, align);
 }
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
 {
-#if 0
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
     /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
     XXH32_state_t state;
     XXH32_reset(&state, seed);
@@ -2107,27 +3042,26 @@ XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t s
 
 
 /*******   Hash streaming   *******/
-/*!
- * @ingroup xxh32_family
- */
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
 {
     return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
 }
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
 {
     XXH_free(statePtr);
     return XXH_OK;
 }
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
 {
     XXH_memcpy(dstState, srcState, sizeof(*dstState));
 }
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
 {
     XXH_ASSERT(statePtr != NULL);
@@ -2140,7 +3074,7 @@ XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t s
 }
 
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH_errorcode
 XXH32_update(XXH32_state_t* state, const void* input, size_t len)
 {
@@ -2195,7 +3129,7 @@ XXH32_update(XXH32_state_t* state, const void* input, size_t len)
 }
 
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
 {
     xxh_u32 h32;
@@ -2213,31 +3147,18 @@ XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
 
     return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
 }
-
+#endif /* !XXH_NO_STREAM */
 
 /*******   Canonical representation   *******/
 
-/*!
- * @ingroup xxh32_family
- * The default return values from XXH functions are unsigned 32 and 64 bit
- * integers.
- *
- * The canonical representation uses big endian convention, the same convention
- * as human-readable numbers (large digits first).
- *
- * This way, hash values can be written into a file or buffer, remaining
- * comparable across different systems.
- *
- * The following functions allow transformation of hash values to and from their
- * canonical format.
- */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
 {
-    /* XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t)); */
+    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
     if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
     XXH_memcpy(dst, &hash, sizeof(*dst));
 }
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
 {
     return XXH_readBE32(src);
@@ -2278,18 +3199,19 @@ static xxh_u64 XXH_read64(const void* memPtr)
 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
 
 /*
- * __pack instructions are safer, but compiler specific, hence potentially
- * problematic for some compilers.
- *
- * Currently only defined for GCC and ICC.
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
  */
 #ifdef XXH_OLD_NAMES
 typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
 #endif
 static xxh_u64 XXH_read64(const void* ptr)
 {
-    typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) xxh_unalign64;
-    return ((const xxh_unalign64*)ptr)->u64;
+    typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64;
+    return *((const xxh_unalign64*)ptr);
 }
 
 #else
@@ -2380,8 +3302,10 @@ XXH_readLE64_align(const void* ptr, XXH_alignment align)
 /*******   xxh64   *******/
 /*!
  * @}
- * @defgroup xxh64_impl XXH64 implementation
+ * @defgroup XXH64_impl XXH64 implementation
  * @ingroup impl
+ *
+ * Details on the XXH64 implementation.
  * @{
  */
 /* #define rather that static const, to be used as initializers */
@@ -2399,11 +3323,29 @@ XXH_readLE64_align(const void* ptr, XXH_alignment align)
 #  define PRIME64_5 XXH_PRIME64_5
 #endif
 
+/*! @copydoc XXH32_round */
 static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
 {
     acc += input * XXH_PRIME64_2;
     acc  = XXH_rotl64(acc, 31);
     acc *= XXH_PRIME64_1;
+#if (defined(__AVX512F__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * DISABLE AUTOVECTORIZATION:
+     * A compiler fence is used to prevent GCC and Clang from
+     * autovectorizing the XXH64 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling AVX512.
+     *
+     * Autovectorization of XXH64 tends to be detrimental,
+     * though the exact outcome may change depending on exact cpu and compiler version.
+     * For information, it has been reported as detrimental for Skylake-X,
+     * but possibly beneficial for Zen4.
+     *
+     * The default is to disable auto-vectorization,
+     * but you can select to enable it instead using `XXH_ENABLE_AUTOVECTORIZE` build variable.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
     return acc;
 }
 
@@ -2415,43 +3357,59 @@ static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
     return acc;
 }
 
-static xxh_u64 XXH64_avalanche(xxh_u64 h64)
+/*! @copydoc XXH32_avalanche */
+static xxh_u64 XXH64_avalanche(xxh_u64 hash)
 {
-    h64 ^= h64 >> 33;
-    h64 *= XXH_PRIME64_2;
-    h64 ^= h64 >> 29;
-    h64 *= XXH_PRIME64_3;
-    h64 ^= h64 >> 32;
-    return h64;
+    hash ^= hash >> 33;
+    hash *= XXH_PRIME64_2;
+    hash ^= hash >> 29;
+    hash *= XXH_PRIME64_3;
+    hash ^= hash >> 32;
+    return hash;
 }
 
 
 #define XXH_get64bits(p) XXH_readLE64_align(p, align)
 
-static xxh_u64
-XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align)
+/*!
+ * @internal
+ * @brief Processes the last 0-31 bytes of @p ptr.
+ *
+ * There may be up to 31 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 32.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash
+ * @see XXH32_finalize().
+ */
+static XXH_PUREF xxh_u64
+XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
 {
     if (ptr==NULL) XXH_ASSERT(len == 0);
     len &= 31;
     while (len >= 8) {
         xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
         ptr += 8;
-        h64 ^= k1;
-        h64  = XXH_rotl64(h64,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+        hash ^= k1;
+        hash  = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
         len -= 8;
     }
     if (len >= 4) {
-        h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+        hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
         ptr += 4;
-        h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+        hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
         len -= 4;
     }
     while (len > 0) {
-        h64 ^= (*ptr++) * XXH_PRIME64_5;
-        h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;
+        hash ^= (*ptr++) * XXH_PRIME64_5;
+        hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
         --len;
     }
-    return  XXH64_avalanche(h64);
+    return  XXH64_avalanche(hash);
 }
 
 #ifdef XXH_OLD_NAMES
@@ -2464,7 +3422,15 @@ XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align)
 #  undef XXH_PROCESS8_64
 #endif
 
-XXH_FORCE_INLINE xxh_u64
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH64().
+ *
+ * @param input , len , seed Directly passed from @ref XXH64().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u64
 XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
 {
     xxh_u64 h64;
@@ -2501,10 +3467,10 @@ XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment
 }
 
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t seed)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
-#if 0
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
     /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
     XXH64_state_t state;
     XXH64_reset(&state, seed);
@@ -2522,27 +3488,27 @@ XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t s
 }
 
 /*******   Hash Streaming   *******/
-
-/*! @ingroup xxh64_family*/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH64_family*/
 XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
 {
     return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
 }
-/*! @ingroup xxh64_family */
+/*! @ingroup XXH64_family */
 XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
 {
     XXH_free(statePtr);
     return XXH_OK;
 }
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t* srcState)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
 {
     XXH_memcpy(dstState, srcState, sizeof(*dstState));
 }
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
 {
     XXH_ASSERT(statePtr != NULL);
     memset(statePtr, 0, sizeof(*statePtr));
@@ -2553,9 +3519,9 @@ XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t s
     return XXH_OK;
 }
 
-/*! @ingroup xxh64_family */
+/*! @ingroup XXH64_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH64_update (XXH64_state_t* state, const void* input, size_t len)
+XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
     if (input==NULL) {
         XXH_ASSERT(len == 0);
@@ -2605,8 +3571,8 @@ XXH64_update (XXH64_state_t* state, const void* input, size_t len)
 }
 
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API XXH64_hash_t XXH64_digest(const XXH64_state_t* state)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
 {
     xxh_u64 h64;
 
@@ -2624,20 +3590,20 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_digest(const XXH64_state_t* state)
 
     return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
 }
-
+#endif /* !XXH_NO_STREAM */
 
 /******* Canonical representation   *******/
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash)
 {
-    /* XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t)); */
+    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
     if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
     XXH_memcpy(dst, &hash, sizeof(*dst));
 }
 
-/*! @ingroup xxh64_family */
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
 {
     return XXH_readBE64(src);
 }
@@ -2650,7 +3616,7 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
 ************************************************************************ */
 /*!
  * @}
- * @defgroup xxh3_impl XXH3 implementation
+ * @defgroup XXH3_impl XXH3 implementation
  * @ingroup impl
  * @{
  */
@@ -2658,11 +3624,19 @@ 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
 
@@ -2676,10 +3650,26 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
 #    define XXH_unlikely(x) (x)
 #endif
 
+#ifndef XXH_HAS_INCLUDE
+#  ifdef __has_include
+/*
+ * Not defined as XXH_HAS_INCLUDE(x) (function-like) because
+ * this causes segfaults in Apple Clang 4.2 (on Mac OS X 10.7 Lion)
+ */
+#    define XXH_HAS_INCLUDE __has_include
+#  else
+#    define XXH_HAS_INCLUDE(x) 0
+#  endif
+#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) \
+   || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* WASM SIMD128 via SIMDe */
 #    define inline __inline__  /* circumvent a clang bug */
 #    include <arm_neon.h>
 #    undef inline
@@ -2790,7 +3780,7 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
  * Note that these are actually implemented as macros.
  *
  * If this is not defined, it is detected automatically.
- * @ref XXH_X86DISPATCH overrides this.
+ * internal macro XXH_X86DISPATCH overrides this.
  */
 enum XXH_VECTOR_TYPE /* fake enum */ {
     XXH_SCALAR = 0,  /*!< Portable scalar version */
@@ -2802,8 +3792,13 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
                       */
     XXH_AVX2   = 2,  /*!< AVX2 for Haswell and Bulldozer */
     XXH_AVX512 = 3,  /*!< AVX512 for Skylake and Icelake */
-    XXH_NEON   = 4,  /*!< NEON for most ARMv7-A and all AArch64 */
+    XXH_NEON   = 4,  /*!<
+                       * NEON for most ARMv7-A, all AArch64, and WASM SIMD128
+                       * via the SIMDeverywhere polyfill provided with the
+                       * Emscripten SDK.
+                       */
     XXH_VSX    = 5,  /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+    XXH_SVE    = 6,  /*!< SVE for some ARMv8-A and ARMv9-A */
 };
 /*!
  * @ingroup tuning
@@ -2825,12 +3820,16 @@ 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 */ \
+     || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* wasm simd128 via SIMDe */ \
    ) && ( \
         defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
@@ -2851,6 +3850,17 @@ 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.
@@ -2870,16 +3880,26 @@ 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.
@@ -2903,153 +3923,126 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
  */
 #if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
   && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
 #  pragma GCC push_options
 #  pragma GCC optimize("-O2")
 #endif
 
-
 #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.
+ * 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.
  *
- * 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.
- *
- * 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] }
+ * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
+ * so the only option is to mark it as aliasing.
+ */
+typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
+
+/*!
+ * @internal
+ * @brief `vld1q_u64` but faster and alignment-safe.
  *
- * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN):
+ * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only
+ * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86).
  *
- *   shrn    v1.2s, v0.2d, #32  // v1 = (uint32x2_t)(v0 >> 32);
- *   xtn     v0.2s, v0.2d       // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF);
+ * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it
+ * prohibits load-store optimizations. Therefore, a direct dereference is used.
  *
- * This is available on ARMv7-A, but is less efficient than a single VZIP.32.
+ * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe
+ * unaligned load.
  */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */
+{
+    return *(xxh_aliasing_uint64x2_t const *)ptr;
+}
+#else
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
+{
+    return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
+}
+#endif
 
 /*!
- * 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;
- * }
+ * @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(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
+#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 noticeable 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.
  *
+ * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning
+ * it effectively becomes worse 4.
+ *
  * @see XXH3_accumulate_512_neon()
  */
 # ifndef XXH3_NEON_LANES
 #  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
-   && !defined(__OPTIMIZE_SIZE__)
+   && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
 #   define XXH3_NEON_LANES 6
 #  else
 #   define XXH3_NEON_LANES XXH_ACC_NB
@@ -3066,27 +4059,42 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
  * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
  */
 #if XXH_VECTOR == XXH_VSX
+/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`,
+ * and `pixel`. This is a problem for obvious reasons.
+ *
+ * These keywords are unnecessary; the spec literally says they are
+ * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd
+ * after including the header.
+ *
+ * We use pragma push_macro/pop_macro to keep the namespace clean. */
+#  pragma push_macro("bool")
+#  pragma push_macro("vector")
+#  pragma push_macro("pixel")
+/* silence potential macro redefined warnings */
+#  undef bool
+#  undef vector
+#  undef pixel
+
 #  if defined(__s390x__)
 #    include <s390intrin.h>
 #  else
-/* gcc's altivec.h can have the unwanted consequence to unconditionally
- * #define bool, vector, and pixel keywords,
- * with bad consequences for programs already using these keywords for other purposes.
- * The paragraph defining these macros is skipped when __APPLE_ALTIVEC__ is defined.
- * __APPLE_ALTIVEC__ is _generally_ defined automatically by the compiler,
- * but it seems that, in some cases, it isn't.
- * Force the build macro to be defined, so that keywords are not altered.
- */
-#    if defined(__GNUC__) && !defined(__APPLE_ALTIVEC__)
-#      define __APPLE_ALTIVEC__
-#    endif
 #    include <altivec.h>
 #  endif
 
+/* Restore the original macro values, if applicable. */
+#  pragma pop_macro("pixel")
+#  pragma pop_macro("vector")
+#  pragma pop_macro("bool")
+
 typedef __vector unsigned long long xxh_u64x2;
 typedef __vector unsigned char xxh_u8x16;
 typedef __vector unsigned xxh_u32x4;
 
+/*
+ * 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__)
@@ -3138,8 +4146,9 @@ XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
  /* s390x is always big endian, no issue on this platform */
 #  define XXH_vec_mulo vec_mulo
 #  define XXH_vec_mule vec_mule
-# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw)
+# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__)
 /* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
+ /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */
 #  define XXH_vec_mulo __builtin_altivec_vmulouw
 #  define XXH_vec_mule __builtin_altivec_vmuleuw
 # else
@@ -3160,13 +4169,28 @@ XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
 # endif /* XXH_vec_mulo, XXH_vec_mule */
 #endif /* XXH_VECTOR == XXH_VSX */
 
+#if XXH_VECTOR == XXH_SVE
+#define ACCRND(acc, offset) \
+do { \
+    svuint64_t input_vec = svld1_u64(mask, xinput + offset);         \
+    svuint64_t secret_vec = svld1_u64(mask, xsecret + offset);       \
+    svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec);     \
+    svuint64_t swapped = svtbl_u64(input_vec, kSwap);                \
+    svuint64_t mixed_lo = svextw_u64_x(mask, mixed);                 \
+    svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32);            \
+    svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
+    acc = svadd_u64_x(mask, acc, mul);                               \
+} while (0)
+#endif /* XXH_VECTOR == XXH_SVE */
 
 /* prefetch
  * can be disabled, by declaring XXH_NO_PREFETCH build macro */
 #if defined(XXH_NO_PREFETCH)
 #  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
 #else
-#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
+#  if XXH_SIZE_OPT >= 1
+#    define XXH_PREFETCH(ptr) (void)(ptr)
+#  elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
 #    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
 #    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
 #  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
@@ -3203,6 +4227,8 @@ XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
     0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
 };
 
+static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL;  /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */
+static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL;  /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */
 
 #ifdef XXH_OLD_NAMES
 #  define kSecret XXH3_kSecret
@@ -3394,7 +4420,7 @@ XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
 }
 
 /*! Seems to produce slightly better code on GCC for some reason. */
-XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
+XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
 {
     XXH_ASSERT(0 <= shift && shift < 64);
     return v64 ^ (v64 >> shift);
@@ -3407,7 +4433,7 @@ XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
 static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
 {
     h64 = XXH_xorshift64(h64, 37);
-    h64 *= 0x165667919E3779F9ULL;
+    h64 *= PRIME_MX1;
     h64 = XXH_xorshift64(h64, 32);
     return h64;
 }
@@ -3421,9 +4447,9 @@ static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
 {
     /* this mix is inspired by Pelle Evensen's rrmxmx */
     h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
-    h64 *= 0x9FB21C651E98DF25ULL;
+    h64 *= PRIME_MX2;
     h64 ^= (h64 >> 35) + len ;
-    h64 *= 0x9FB21C651E98DF25ULL;
+    h64 *= PRIME_MX2;
     return XXH_xorshift64(h64, 28);
 }
 
@@ -3461,7 +4487,7 @@ static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
  *
  * This adds an extra layer of strength for custom secrets.
  */
-XXH_FORCE_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -3483,7 +4509,7 @@ XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_h
     }
 }
 
-XXH_FORCE_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -3499,7 +4525,7 @@ XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_h
     }
 }
 
-XXH_FORCE_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -3516,7 +4542,7 @@ XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_
     }
 }
 
-XXH_FORCE_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(len <= 16);
@@ -3586,7 +4612,7 @@ XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
 }
 
 /* For mid range keys, XXH3 uses a Mum-hash variant. */
-XXH_FORCE_INLINE XXH64_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                      XXH64_hash_t seed)
@@ -3595,6 +4621,14 @@ XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
     XXH_ASSERT(16 < len && len <= 128);
 
     {   xxh_u64 acc = len * XXH_PRIME64_1;
+#if XXH_SIZE_OPT >= 1
+        /* Smaller and cleaner, but slightly slower. */
+        unsigned int i = (unsigned int)(len - 1) / 32;
+        do {
+            acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
+            acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
+        } while (i-- != 0);
+#else
         if (len > 32) {
             if (len > 64) {
                 if (len > 96) {
@@ -3609,14 +4643,17 @@ XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
         }
         acc += XXH3_mix16B(input+0, secret+0, seed);
         acc += XXH3_mix16B(input+len-16, secret+16, seed);
-
+#endif
         return XXH3_avalanche(acc);
     }
 }
 
+/*!
+ * @brief Maximum size of "short" key in bytes.
+ */
 #define XXH3_MIDSIZE_MAX 240
 
-XXH_NO_INLINE XXH64_hash_t
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                       XXH64_hash_t seed)
@@ -3628,13 +4665,17 @@ 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 */
@@ -3661,11 +4702,13 @@ 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);
     }
 }
 
@@ -3681,6 +4724,47 @@ 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);
@@ -3749,7 +4833,7 @@ 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); */
@@ -3759,6 +4843,7 @@ 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.
@@ -3792,13 +4877,12 @@ 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));
@@ -3813,7 +4897,8 @@ 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;
@@ -3821,14 +4906,7 @@ 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);
     }   }
 }
 
@@ -3864,7 +4942,7 @@ 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); */
@@ -3874,6 +4952,7 @@ 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)
@@ -3896,7 +4975,7 @@ 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));
@@ -3928,12 +5007,12 @@ 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);
     }
 }
 
@@ -3980,6 +5059,7 @@ 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)
@@ -4058,14 +5138,28 @@ XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
 
 /*!
  * @internal
- * @brief The bulk processing loop for NEON.
+ * @brief The bulk processing loop for NEON and WASM SIMD128.
  *
  * The NEON code path is actually partially scalar when running on AArch64. This
  * is to optimize the pipelining and can have up to 15% speedup depending on the
  * CPU, and it also mitigates some GCC codegen issues.
  *
  * @see XXH3_NEON_LANES for configuring this and details about this optimization.
+ *
+ * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
+ * integers instead of the other platforms which mask full 64-bit vectors,
+ * so the setup is more complicated than just shifting right.
+ *
+ * Additionally, there is an optimization for 4 lanes at once noted below.
+ *
+ * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
+ * there needs to be *three* versions of the accumulate operation used
+ * for the remaining 2 lanes.
+ *
+ * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap
+ * nearly perfectly.
  */
+
 XXH_FORCE_INLINE void
 XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT input,
@@ -4073,101 +5167,182 @@ 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;
+        uint8_t const* xinput = (const uint8_t *) input;
+        uint8_t const* xsecret  = (const uint8_t *) secret;
 
         size_t i;
-        /* NEON for the first few lanes (these loops are normally interleaved) */
-        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
+#ifdef __wasm_simd128__
+        /*
+         * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret
+         * is constant propagated, which results in it converting it to this
+         * inside the loop:
+         *
+         *    a = v128.load(XXH3_kSecret +  0 + $secret_offset, offset = 0)
+         *    b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0)
+         *    ...
+         *
+         * This requires a full 32-bit address immediate (and therefore a 6 byte
+         * instruction) as well as an add for each offset.
+         *
+         * Putting an asm guard prevents it from folding (at the cost of losing
+         * the alignment hint), and uses the free offset in `v128.load` instead
+         * of adding secret_offset each time which overall reduces code size by
+         * about a kilobyte and improves performance.
+         */
+        XXH_COMPILER_GUARD(xsecret);
+#endif
+        /* Scalar lanes use the normal scalarRound routine */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarRound(acc, input, secret, i);
+        }
+        i = 0;
+        /* 4 NEON lanes at a time. */
+        for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
             /* data_vec = xinput[i]; */
-            uint8x16_t data_vec    = vld1q_u8(xinput  + (i * 16));
+            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];  */
-            uint8x16_t key_vec     = vld1q_u8(xsecret + (i * 16));
-            uint64x2_t data_key;
-            uint32x2_t data_key_lo, data_key_hi;
-            /* xacc[i] += swap(data_vec); */
-            uint64x2_t const data64  = vreinterpretq_u64_u8(data_vec);
-            uint64x2_t const swapped = vextq_u64(data64, data64, 1);
-            xacc[i] = vaddq_u64 (xacc[i], swapped);
+            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; */
-            data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec));
-            /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF);
-             * data_key_hi = (uint32x2_t) (data_key >> 32);
-             * data_key = UNDEFINED; */
-            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
-            /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
-            xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi);
+            uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
+            uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
 
+            /*
+             * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
+             * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
+             * get one vector with the low 32 bits of each lane, and one vector
+             * with the high 32 bits of each lane.
+             *
+             * The intrinsic returns a double vector because the original ARMv7-a
+             * instruction modified both arguments in place. AArch64 and SIMD128 emit
+             * two instructions from this intrinsic.
+             *
+             *  [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
+             *  [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
+             */
+            uint32x4x2_t unzipped = vuzpq_u32(
+                vreinterpretq_u32_u64(data_key_1),
+                vreinterpretq_u32_u64(data_key_2)
+            );
+            /* data_key_lo = data_key & 0xFFFFFFFF */
+            uint32x4_t data_key_lo = unzipped.val[0];
+            /* data_key_hi = data_key >> 32 */
+            uint32x4_t data_key_hi = unzipped.val[1];
+            /*
+             * Then, we can split the vectors horizontally and multiply which, as for most
+             * widening intrinsics, have a variant that works on both high half vectors
+             * for free on AArch64. A similar instruction is available on SIMD128.
+             *
+             * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
+             */
+            uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
+            uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
+            /*
+             * Clang reorders
+             *    a += b * c;     // umlal   swap.2d, dkl.2s, dkh.2s
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             * to
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             *    c += b * c;     // umlal   acc.2d, dkl.2s, dkh.2s
+             *
+             * While it would make sense in theory since the addition is faster,
+             * for reasons likely related to umlal being limited to certain NEON
+             * pipelines, this is worse. A compiler guard fixes this.
+             */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_1);
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_2);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i]   = vaddq_u64(xacc[i], sum_1);
+            xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
         }
-        /* Scalar for the remainder. This may be a zero iteration loop. */
-        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
-            XXH3_scalarRound(acc, input, secret, i);
+        /* Operate on the remaining NEON lanes 2 at a time. */
+        for (; i < XXH3_NEON_LANES / 2; i++) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec = XXH_vld1q_u64(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            /* acc_vec_2 = swap(data_vec) */
+            uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* For two lanes, just use VMOVN and VSHRN. */
+            /* data_key_lo = data_key & 0xFFFFFFFF; */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* data_key_hi = data_key >> 32; */
+            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+            /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
+            uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi);
+            /* Same Clang workaround as before */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i] = vaddq_u64 (xacc[i], sum);
         }
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
 
 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);
 
         size_t i;
-        /* NEON for the first few lanes (these loops are normally interleaved) */
+        /* WASM uses operator overloads and doesn't need these. */
+#ifndef __wasm_simd128__
+        /* { prime32_1, prime32_1 } */
+        uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1);
+        /* { 0, prime32_1, 0, prime32_1 } */
+        uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32));
+#endif
+
+        /* AArch64 uses both scalar and neon at the same time */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarScrambleRound(acc, secret, i);
+        }
         for (i=0; i < XXH3_NEON_LANES / 2; i++) {
             /* xacc[i] ^= (xacc[i] >> 47); */
             uint64x2_t acc_vec  = xacc[i];
-            uint64x2_t shifted  = vshrq_n_u64 (acc_vec, 47);
-            uint64x2_t data_vec = veorq_u64   (acc_vec, shifted);
+            uint64x2_t shifted  = vshrq_n_u64(acc_vec, 47);
+            uint64x2_t data_vec = veorq_u64(acc_vec, shifted);
 
             /* xacc[i] ^= xsecret[i]; */
-            uint8x16_t key_vec  = vld1q_u8    (xsecret + (i * 16));
-            uint64x2_t data_key = veorq_u64   (data_vec, vreinterpretq_u64_u8(key_vec));
-
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
             /* xacc[i] *= XXH_PRIME32_1 */
-            uint32x2_t data_key_lo, data_key_hi;
-            /* 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; */
-                xacc[i] = vshlq_n_u64(prod_hi, 32);
-                /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */
-                xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime);
-            }
-        }
-        /* Scalar for the remainder. This may be a zero iteration loop. */
-        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
-            XXH3_scalarScrambleRound(acc, secret, i);
+#ifdef __wasm_simd128__
+            /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */
+            xacc[i] = data_key * XXH_PRIME32_1;
+#else
+            /*
+             * Expanded version with portable NEON intrinsics
+             *
+             *    lo(x) * lo(y) + (hi(x) * lo(y) << 32)
+             *
+             * prod_hi = hi(data_key) * lo(prime) << 32
+             *
+             * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector
+             * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits
+             * and avoid the shift.
+             */
+            uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi);
+            /* Extract low bits for vmlal_u32  */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */
+            xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo);
+#endif
         }
     }
 }
-
 #endif
 
 #if (XXH_VECTOR == XXH_VSX)
@@ -4178,23 +5353,23 @@ 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 */
@@ -4203,18 +5378,18 @@ 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 };
@@ -4226,7 +5401,7 @@ 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 */
@@ -4240,8 +5415,148 @@ 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().
@@ -4264,7 +5579,7 @@ 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]);
     }
 }
 
@@ -4278,10 +5593,18 @@ XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
                      const void* XXH_RESTRICT secret)
 {
     size_t i;
+    /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
+#if defined(__GNUC__) && !defined(__clang__) \
+  && (defined(__arm__) || defined(__thumb2__)) \
+  && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
+  && XXH_SIZE_OPT <= 0
+#  pragma GCC unroll 8
+#endif
     for (i=0; i < XXH_ACC_NB; i++) {
         XXH3_scalarRound(acc, input, secret, i);
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
 
 /*!
  * @internal
@@ -4333,10 +5656,10 @@ 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
@@ -4351,7 +5674,7 @@ 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
@@ -4368,17 +5691,11 @@ 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.
@@ -4391,7 +5708,7 @@ 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);
 
@@ -4399,82 +5716,63 @@ typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
 #if (XXH_VECTOR == XXH_AVX512)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_avx512
+#define XXH3_accumulate     XXH3_accumulate_avx512
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_avx512
 #define XXH3_initCustomSecret XXH3_initCustomSecret_avx512
 
 #elif (XXH_VECTOR == XXH_AVX2)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_avx2
+#define XXH3_accumulate     XXH3_accumulate_avx2
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_avx2
 #define XXH3_initCustomSecret XXH3_initCustomSecret_avx2
 
 #elif (XXH_VECTOR == XXH_SSE2)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_sse2
+#define XXH3_accumulate     XXH3_accumulate_sse2
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_sse2
 #define XXH3_initCustomSecret XXH3_initCustomSecret_sse2
 
 #elif (XXH_VECTOR == XXH_NEON)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_neon
+#define XXH3_accumulate     XXH3_accumulate_neon
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_neon
 #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
 #elif (XXH_VECTOR == XXH_VSX)
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_vsx
+#define XXH3_accumulate     XXH3_accumulate_vsx
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_vsx
 #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
+#elif (XXH_VECTOR == XXH_SVE)
+#define XXH3_accumulate_512 XXH3_accumulate_512_sve
+#define XXH3_accumulate     XXH3_accumulate_sve
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
 #else /* scalar */
 
 #define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_accumulate     XXH3_accumulate_scalar
 #define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
 #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
 #endif
 
-
-
-#ifndef XXH_PREFETCH_DIST
-#  ifdef __clang__
-#    define XXH_PREFETCH_DIST 320
-#  else
-#    if (XXH_VECTOR == XXH_AVX512)
-#      define XXH_PREFETCH_DIST 512
-#    else
-#      define XXH_PREFETCH_DIST 384
-#    endif
-#  endif  /* __clang__ */
-#endif  /* XXH_PREFETCH_DIST */
-
-/*
- * XXH3_accumulate()
- * Loops over XXH3_accumulate_512().
- * Assumption: nbStripes will not overflow the secret size
- */
-XXH_FORCE_INLINE void
-XXH3_accumulate(     xxh_u64* XXH_RESTRICT acc,
-                const xxh_u8* XXH_RESTRICT input,
-                const xxh_u8* XXH_RESTRICT secret,
-                      size_t nbStripes,
-                      XXH3_f_accumulate_512 f_acc512)
-{
-    size_t n;
-    for (n = 0; n < nbStripes; n++ ) {
-        const xxh_u8* const in = input + n*XXH_STRIPE_LEN;
-        XXH_PREFETCH(in + XXH_PREFETCH_DIST);
-        f_acc512(acc,
-                 in,
-                 secret + n*XXH_SECRET_CONSUME_RATE);
-    }
-}
+#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */
+#  undef XXH3_initCustomSecret
+#  define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+#endif
 
 XXH_FORCE_INLINE void
 XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
                       const xxh_u8* XXH_RESTRICT input, size_t len,
                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                            XXH3_f_accumulate_512 f_acc512,
+                            XXH3_f_accumulate f_acc,
                             XXH3_f_scrambleAcc f_scramble)
 {
     size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
@@ -4486,7 +5784,7 @@ 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);
     }
 
@@ -4494,12 +5792,12 @@ 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);
     }   }
 }
 
@@ -4544,12 +5842,12 @@ 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);
@@ -4563,13 +5861,15 @@ 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);
 }
 
 /*
@@ -4578,12 +5878,12 @@ XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
  * Note that inside this no_inline function, we do inline the internal loop,
  * and provide a statically defined secret size to allow optimization of vector loop.
  */
-XXH_NO_INLINE XXH64_hash_t
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
 XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
                           XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
 {
     (void)seed64; (void)secret; (void)secretLen;
-    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate_512, XXH3_scrambleAcc);
+    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 /*
@@ -4600,18 +5900,20 @@ XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
 XXH_FORCE_INLINE XXH64_hash_t
 XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
                                     XXH64_hash_t seed,
-                                    XXH3_f_accumulate_512 f_acc512,
+                                    XXH3_f_accumulate f_acc,
                                     XXH3_f_scrambleAcc f_scramble,
                                     XXH3_f_initCustomSecret f_initSec)
 {
+#if XXH_SIZE_OPT <= 0
     if (seed == 0)
         return XXH3_hashLong_64b_internal(input, len,
                                           XXH3_kSecret, sizeof(XXH3_kSecret),
-                                          f_acc512, f_scramble);
+                                          f_acc, f_scramble);
+#endif
     {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
         f_initSec(secret, seed);
         return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
-                                          f_acc512, f_scramble);
+                                          f_acc, f_scramble);
     }
 }
 
@@ -4619,12 +5921,12 @@ 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);
 }
 
 
@@ -4656,37 +5958,37 @@ XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
 
 /* ===   Public entry point   === */
 
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
 {
-    return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+    return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
-    return XXH3_64bits_internal(input, len, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+    return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
 {
-    return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+    return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
 }
 
 XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSecretandSeed(const void* input, size_t len, 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 (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
-    return XXH3_hashLong_64b_withSecret(input, len, seed, (const xxh_u8*)secret, secretSize);
+    if (length <= XXH3_MIDSIZE_MAX)
+        return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize);
 }
 
 
 /* ===   XXH3 streaming   === */
-
+#ifndef XXH_NO_STREAM
 /*
  * Malloc's a pointer that is always aligned to align.
  *
@@ -4710,7 +6012,7 @@ XXH3_64bits_withSecretandSeed(const void* input, size_t len, const void* secret,
  *
  * Align must be a power of 2 and 8 <= align <= 128.
  */
-static void* XXH_alignedMalloc(size_t s, size_t align)
+static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align)
 {
     XXH_ASSERT(align <= 128 && align >= 8); /* range check */
     XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
@@ -4752,7 +6054,15 @@ static void XXH_alignedFree(void* p)
         XXH_free(base);
     }
 }
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Allocate an @ref XXH3_state_t.
+ *
+ * @return An allocated pointer of @ref XXH3_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH3_freeState().
+ */
 XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
 {
     XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
@@ -4761,16 +6071,25 @@ XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
     return state;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Frees an @ref XXH3_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note Must be allocated with XXH3_createState().
+ */
 XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
 {
     XXH_alignedFree(statePtr);
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API void
-XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state)
+XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state)
 {
     XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
 }
@@ -4802,18 +6121,18 @@ XXH3_reset_internal(XXH3_state_t* statePtr,
     statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset(XXH3_state_t* statePtr)
+XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
 {
     if (statePtr == NULL) return XXH_ERROR;
     XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
+XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
     if (statePtr == NULL) return XXH_ERROR;
     XXH3_reset_internal(statePtr, 0, secret, secretSize);
@@ -4822,9 +6141,9 @@ XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
+XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
 {
     if (statePtr == NULL) return XXH_ERROR;
     if (seed==0) return XXH3_64bits_reset(statePtr);
@@ -4834,9 +6153,9 @@ XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @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;
@@ -4846,35 +6165,61 @@ 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
-# ifndef __clang__ /* clang doesn't need additional stack space */
+# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */
 #   define XXH3_STREAM_USE_STACK 1
 # endif
 #endif
@@ -4884,7 +6229,7 @@ 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) {
@@ -4900,7 +6245,8 @@ 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
@@ -4908,7 +6254,7 @@ 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;
@@ -4930,57 +6276,20 @@ 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);
@@ -4989,19 +6298,19 @@ 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
     }
 
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len)
+XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
     return XXH3_update(state, (const xxh_u8*)input, len,
-                       XXH3_accumulate_512, XXH3_scrambleAcc);
+                       XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 
@@ -5010,37 +6319,40 @@ 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)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
 {
     const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
     if (state->totalLen > XXH3_MIDSIZE_MAX) {
@@ -5056,7 +6368,7 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
     return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
                                   secret, state->secretLimit + XXH_STRIPE_LEN);
 }
-
+#endif /* !XXH_NO_STREAM */
 
 
 /* ==========================================
@@ -5076,7 +6388,7 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
  * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
  */
 
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     /* A doubled version of 1to3_64b with different constants. */
@@ -5105,7 +6417,7 @@ XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_
     }
 }
 
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -5125,14 +6437,14 @@ XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_
         m128.low64  ^= (m128.high64 >> 3);
 
         m128.low64   = XXH_xorshift64(m128.low64, 35);
-        m128.low64  *= 0x9FB21C651E98DF25ULL;
+        m128.low64  *= PRIME_MX2;
         m128.low64   = XXH_xorshift64(m128.low64, 28);
         m128.high64  = XXH3_avalanche(m128.high64);
         return m128;
     }
 }
 
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -5207,7 +6519,7 @@ XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64
 /*
  * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
  */
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(len <= 16);
@@ -5238,7 +6550,7 @@ XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
 }
 
 
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                       XXH64_hash_t seed)
@@ -5249,6 +6561,16 @@ XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
     {   XXH128_hash_t acc;
         acc.low64 = len * XXH_PRIME64_1;
         acc.high64 = 0;
+
+#if XXH_SIZE_OPT >= 1
+        {
+            /* Smaller, but slightly slower. */
+            unsigned int i = (unsigned int)(len - 1) / 32;
+            do {
+                acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed);
+            } while (i-- != 0);
+        }
+#else
         if (len > 32) {
             if (len > 64) {
                 if (len > 96) {
@@ -5259,6 +6581,7 @@ XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
             acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
         }
         acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
+#endif
         {   XXH128_hash_t h128;
             h128.low64  = acc.low64 + acc.high64;
             h128.high64 = (acc.low64    * XXH_PRIME64_1)
@@ -5271,7 +6594,7 @@ XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
     }
 }
 
-XXH_NO_INLINE XXH128_hash_t
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
                        const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                        XXH64_hash_t seed)
@@ -5280,25 +6603,34 @@ 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 */
@@ -5306,7 +6638,7 @@ 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;
@@ -5323,12 +6655,12 @@ 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);
@@ -5346,47 +6678,50 @@ XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
 }
 
 /*
- * It's important for performance that XXH3_hashLong is not inlined.
+ * It's important for performance that XXH3_hashLong() is not inlined.
  */
-XXH_NO_INLINE XXH128_hash_t
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
 XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
                            XXH64_hash_t seed64,
                            const void* XXH_RESTRICT secret, size_t secretLen)
 {
     (void)seed64; (void)secret; (void)secretLen;
     return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
-                                       XXH3_accumulate_512, XXH3_scrambleAcc);
+                                       XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 /*
- * It's important for performance to pass @secretLen (when it's static)
+ * 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);
     }
 }
 
@@ -5399,7 +6734,7 @@ 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,
@@ -5429,94 +6764,93 @@ XXH3_128bits_internal(const void* input, size_t len,
 
 /* ===   Public XXH128 API   === */
 
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len)
 {
     return XXH3_128bits_internal(input, len, 0,
                                  XXH3_kSecret, sizeof(XXH3_kSecret),
                                  XXH3_hashLong_128b_default);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
     return XXH3_128bits_internal(input, len, 0,
                                  (const xxh_u8*)secret, secretSize,
                                  XXH3_hashLong_128b_withSecret);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
     return XXH3_128bits_internal(input, len, seed,
                                  XXH3_kSecret, sizeof(XXH3_kSecret),
                                  XXH3_hashLong_128b_withSeed);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSecretandSeed(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);
     return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH128(const void* input, size_t len, XXH64_hash_t seed)
+XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
     return XXH3_128bits_withSeed(input, len, seed);
 }
 
 
 /* ===   XXH3 128-bit streaming   === */
-
+#ifndef XXH_NO_STREAM
 /*
  * All initialization and update functions are identical to 64-bit streaming variant.
  * The only difference is the finalization routine.
  */
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset(XXH3_state_t* statePtr)
+XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
 {
     return XXH3_64bits_reset(statePtr);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
+XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
     return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
+XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
 {
     return XXH3_64bits_reset_withSeed(statePtr, seed);
 }
 
-/*! @ingroup xxh3_family */
+/*! @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 */
+/*! @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)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
 {
     const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
     if (state->totalLen > XXH3_MIDSIZE_MAX) {
@@ -5540,13 +6874,13 @@ XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state)
     return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
                                    secret, state->secretLimit + XXH_STRIPE_LEN);
 }
-
+#endif /* !XXH_NO_STREAM */
 /* 128-bit utility functions */
 
 #include <string.h>   /* memcmp, memcpy */
 
 /* return : 1 is equal, 0 if different */
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
 {
     /* note : XXH128_hash_t is compact, it has no padding byte */
@@ -5554,11 +6888,11 @@ XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
 }
 
 /* This prototype is compatible with stdlib's qsort().
- * return : >0 if *h128_1  > *h128_2
- *          <0 if *h128_1  < *h128_2
- *          =0 if *h128_1 == *h128_2  */
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2)
+ * @return : >0 if *h128_1  > *h128_2
+ *           <0 if *h128_1  < *h128_2
+ *           =0 if *h128_1 == *h128_2  */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2)
 {
     XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
     XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
@@ -5570,9 +6904,9 @@ XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2)
 
 
 /*======   Canonical representation   ======*/
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API void
-XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash)
+XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash)
 {
     XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
     if (XXH_CPU_LITTLE_ENDIAN) {
@@ -5583,9 +6917,9 @@ XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash)
     XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH128_hash_t
-XXH128_hashFromCanonical(const XXH128_canonical_t* src)
+XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src)
 {
     XXH128_hash_t h;
     h.high64 = XXH_readBE64(src);
@@ -5607,9 +6941,9 @@ XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
     XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
 }
 
-/*! @ingroup xxh3_family */
+/*! @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);
@@ -5652,9 +6986,9 @@ XXH3_generateSecret(void* secretBuffer, size_t secretSize, const void* customSee
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @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);
@@ -5667,7 +7001,7 @@ XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_t seed)
 /* Pop our optimization override from above */
 #if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
   && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
 #  pragma GCC pop_options
 #endif
 
@@ -5682,5 +7016,5 @@ XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_t seed)
 
 
 #if defined (__cplusplus)
-}
+} /* extern "C" */
 #endif
diff --git a/thirdparty/zstd/common/zstd_internal.h b/thirdparty/zstd/common/zstd_internal.h
index 1f942f27bf..ecb9cfba87 100644
--- a/thirdparty/zstd/common/zstd_internal.h
+++ b/thirdparty/zstd/common/zstd_internal.h
@@ -178,7 +178,7 @@ static void ZSTD_copy8(void* dst, const void* src) {
     ZSTD_memcpy(dst, src, 8);
 #endif
 }
-#define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+#define COPY8(d,s) do { ZSTD_copy8(d,s); d+=8; s+=8; } while (0)
 
 /* Need to use memmove here since the literal buffer can now be located within
    the dst buffer. In circumstances where the op "catches up" to where the
@@ -198,7 +198,7 @@ static void ZSTD_copy16(void* dst, const void* src) {
     ZSTD_memcpy(dst, copy16_buf, 16);
 #endif
 }
-#define COPY16(d,s) { ZSTD_copy16(d,s); d+=16; s+=16; }
+#define COPY16(d,s) do { ZSTD_copy16(d,s); d+=16; s+=16; } while (0)
 
 #define WILDCOPY_OVERLENGTH 32
 #define WILDCOPY_VECLEN 16
@@ -227,7 +227,7 @@ void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e
     if (ovtype == ZSTD_overlap_src_before_dst && diff < WILDCOPY_VECLEN) {
         /* Handle short offset copies. */
         do {
-            COPY8(op, ip)
+            COPY8(op, ip);
         } while (op < oend);
     } else {
         assert(diff >= WILDCOPY_VECLEN || diff <= -WILDCOPY_VECLEN);
@@ -366,13 +366,13 @@ typedef struct {
 
 /*! ZSTD_getcBlockSize() :
  *  Provides the size of compressed block from block header `src` */
-/* Used by: decompress, fullbench (does not get its definition from here) */
+/*  Used by: decompress, fullbench */
 size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
                           blockProperties_t* bpPtr);
 
 /*! ZSTD_decodeSeqHeaders() :
  *  decode sequence header from src */
-/* Used by: decompress, fullbench (does not get its definition from here) */
+/*  Used by: zstd_decompress_block, fullbench */
 size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                        const void* src, size_t srcSize);