resolved most of the compiler warnings/errors under gcc

also update xxhash and tinyformat

7 files changed, 2115 insertions, 1139 deletions

diff --git a/source/extern/.gitattributes b/source/extern/.gitattributes
new file mode 100644
index 0000000..c5fb0d0
--- /dev/null
+++ b/source/extern/.gitattributes
@@ -0,0 +1 @@
+** linguist-vendored=true
diff --git a/source/extern/tinyformat.h b/source/extern/tinyformat.h
index 6aa00a7..b588bf9 100644
--- a/source/extern/tinyformat.h
+++ b/source/extern/tinyformat.h
@@ -33,6 +33,7 @@
 //
 // * Type safety and extensibility for user defined types.
 // * C99 printf() compatibility, to the extent possible using std::ostream
+// * POSIX extension for positional arguments
 // * Simplicity and minimalism.  A single header file to include and distribute
 //   with your projects.
 // * Augment rather than replace the standard stream formatting mechanism
@@ -42,7 +43,7 @@
 // Main interface example usage
 // ----------------------------
 //
-// To print a date to std::cout:
+// To print a date to std::cout for American usage:
 //
 //   std::string weekday = "Wednesday";
 //   const char* month = "July";
@@ -52,6 +53,14 @@
 //
 //   tfm::printf("%s, %s %d, %.2d:%.2d\n", weekday, month, day, hour, min);
 //
+// POSIX extension for positional arguments is available.
+// The ability to rearrange formatting arguments is an important feature
+// for localization because the word order may vary in different languages.
+//
+// Previous example for German usage. Arguments are reordered:
+//
+//   tfm::printf("%1$s, %3$d. %2$s, %4$d:%5$.2d\n", weekday, month, day, hour, min);
+//
 // The strange types here emphasize the type safety of the interface; it is
 // possible to print a std::string using the "%s" conversion, and a
 // size_t using the "%d" conversion.  A similar result could be achieved
@@ -133,11 +142,17 @@ namespace tfm = tinyformat;
 //------------------------------------------------------------------------------
 // Implementation details.
 #include <algorithm>
-#include <cassert>
 #include <iostream>
 #include <sstream>
+#include <cmath>
+
+#ifndef TINYFORMAT_ASSERT
+#   include <cassert>
+#   define TINYFORMAT_ASSERT(cond) assert(cond)
+#endif
 
 #ifndef TINYFORMAT_ERROR
+#   include <cassert>
 #   define TINYFORMAT_ERROR(reason) assert(0 && reason)
 #endif
 
@@ -147,23 +162,14 @@ namespace tfm = tinyformat;
 #   endif
 #endif
 
-#ifdef TINYFORMAT_USE_VARIADIC_TEMPLATES
-#   include <array>
-#   if defined(_MSC_VER) && _MSC_VER <= 1800 // VS2013
-#       define TINYFORMAT_BRACED_INIT_WORKAROUND(x) (x)
-#   else
-#       define TINYFORMAT_BRACED_INIT_WORKAROUND(x) {x}
-#   endif
-#endif
-
 #if defined(__GLIBCXX__) && __GLIBCXX__ < 20080201
-//  std::showpos is broken on old libstdc++ as provided with OSX.  See
+//  std::showpos is broken on old libstdc++ as provided with macOS.  See
 //  http://gcc.gnu.org/ml/libstdc++/2007-11/msg00075.html
 #   define TINYFORMAT_OLD_LIBSTDCPLUSPLUS_WORKAROUND
 #endif
 
 #ifdef __APPLE__
-// Workaround OSX linker warning: xcode uses different default symbol
+// Workaround macOS linker warning: Xcode uses different default symbol
 // visibilities for static libs vs executables (see issue #25)
 #   define TINYFORMAT_HIDDEN __attribute__((visibility("hidden")))
 #else
@@ -219,7 +225,7 @@ template<int n> struct is_wchar<wchar_t[n]> {};
 template<typename T, typename fmtT, bool convertible = is_convertible<T, fmtT>::value>
 struct formatValueAsType
 {
-    static void invoke(std::ostream& /*out*/, const T& /*value*/) { assert(0); }
+    static void invoke(std::ostream& /*out*/, const T& /*value*/) { TINYFORMAT_ASSERT(0); }
 };
 // Specialized version for types that can actually be converted to fmtT, as
 // indicated by the "convertible" template parameter.
@@ -241,8 +247,7 @@ struct formatZeroIntegerWorkaround<T,true>
 {
     static bool invoke(std::ostream& out, const T& value)
     {
-        if (static_cast<int>(value) == 0 && out.flags() & std::ios::showpos)
-        {
+        if (static_cast<int>(value) == 0 && out.flags() & std::ios::showpos) {
             out << "+0";
             return true;
         }
@@ -277,13 +282,13 @@ inline void formatTruncated(std::ostream& out, const T& value, int ntrunc)
     std::ostringstream tmp;
     tmp << value;
     std::string result = tmp.str();
-    out.write(result.c_str(), std::min(ntrunc, static_cast<int>(result.size())));
+    out.write(result.c_str(), (std::min)(ntrunc, static_cast<int>(result.size())));
 }
 #define TINYFORMAT_DEFINE_FORMAT_TRUNCATED_CSTR(type)       \
 inline void formatTruncated(std::ostream& out, type* value, int ntrunc) \
 {                                                           \
     std::streamsize len = 0;                                \
-    while(len < ntrunc && value[len] != 0)                  \
+    while (len < ntrunc && value[len] != 0)                 \
         ++len;                                              \
     out.write(value, len);                                  \
 }
@@ -293,6 +298,21 @@ TINYFORMAT_DEFINE_FORMAT_TRUNCATED_CSTR(const char)
 TINYFORMAT_DEFINE_FORMAT_TRUNCATED_CSTR(char)
 #undef TINYFORMAT_DEFINE_FORMAT_TRUNCATED_CSTR
 
+template<typename T>
+void spaceFillIfNotFinite(std::ostream& out, const T& value) { }
+// TODO: type_traits would clearly be better here. Should consider moving all
+// these workarounds into a big pre-C++11 section.
+#define TINYFORMAT_SETFILL_NOT_FINITE_FLOATING(type)             \
+inline void spaceFillIfNotFinite(std::ostream& out, type value)  \
+{                                                                \
+    if (out.fill() == '0' && !std::isfinite(value))              \
+        out.fill(' ');                                           \
+}
+TINYFORMAT_SETFILL_NOT_FINITE_FLOATING(float)
+TINYFORMAT_SETFILL_NOT_FINITE_FLOATING(double)
+TINYFORMAT_SETFILL_NOT_FINITE_FLOATING(long double)
+#undef TINYFORMAT_SETFILL_NOT_FINITE_FLOATING
+
 } // namespace detail
 
 
@@ -327,17 +347,17 @@ inline void formatValue(std::ostream& out, const char* /*fmtBegin*/,
     // void* respectively and format that instead of the value itself.  For the
     // %p conversion it's important to avoid dereferencing the pointer, which
     // could otherwise lead to a crash when printing a dangling (const char*).
-    bool canConvertToChar = detail::is_convertible<T,char>::value;
-    bool canConvertToVoidPtr = detail::is_convertible<T, const void*>::value;
-    if(canConvertToChar && *(fmtEnd-1) == 'c')
+    const bool canConvertToChar = detail::is_convertible<T,char>::value;
+    const bool canConvertToVoidPtr = detail::is_convertible<T, const void*>::value;
+    detail::spaceFillIfNotFinite(out, value);
+    if (canConvertToChar && *(fmtEnd-1) == 'c')
         detail::formatValueAsType<T, char>::invoke(out, value);
-    else if(canConvertToVoidPtr && *(fmtEnd-1) == 'p')
+    else if (canConvertToVoidPtr && *(fmtEnd-1) == 'p')
         detail::formatValueAsType<T, const void*>::invoke(out, value);
 #ifdef TINYFORMAT_OLD_LIBSTDCPLUSPLUS_WORKAROUND
-    else if(detail::formatZeroIntegerWorkaround<T>::invoke(out, value)) /**/;
+    else if (detail::formatZeroIntegerWorkaround<T>::invoke(out, value)) /**/;
 #endif
-    else if(ntrunc >= 0)
-    {
+    else if (ntrunc >= 0) {
         // Take care not to overread C strings in truncating conversions like
         // "%.4s" where at most 4 characters may be read.
         detail::formatTruncated(out, value, ntrunc);
@@ -352,8 +372,7 @@ inline void formatValue(std::ostream& out, const char* /*fmtBegin*/,
 inline void formatValue(std::ostream& out, const char* /*fmtBegin*/,  \
                         const char* fmtEnd, int /**/, charType value) \
 {                                                                     \
-    switch(*(fmtEnd-1))                                               \
-    {                                                                 \
+    switch (*(fmtEnd-1)) {                                            \
         case 'u': case 'd': case 'i': case 'o': case 'X': case 'x':   \
             out << static_cast<int>(value); break;                    \
         default:                                                      \
@@ -378,11 +397,11 @@ TINYFORMAT_DEFINE_FORMATVALUE_CHAR(unsigned char)
 #define TINYFORMAT_PASSARGS_TAIL(n) TINYFORMAT_PASSARGS_TAIL_ ## n
 
 // To keep it as transparent as possible, the macros below have been generated
-// using python via the excellent cog.py code generation script.  This avoids
+// using python via the excellent cog code generation script.  This avoids
 // the need for a bunch of complex (but more general) preprocessor tricks as
 // used in boost.preprocessor.
 //
-// To rerun the code generation in place, use `cog.py -r tinyformat.h`
+// To rerun the code generation in place, use `cog -r tinyformat.h`
 // (see http://nedbatchelder.com/code/cog).  Alternatively you can just create
 // extra versions by hand.
 
@@ -491,16 +510,22 @@ namespace detail {
 
 // Type-opaque holder for an argument to format(), with associated actions on
 // the type held as explicit function pointers.  This allows FormatArg's for
-// each argument to be allocated as a homogenous array inside FormatList
+// each argument to be allocated as a homogeneous array inside FormatList
 // whereas a naive implementation based on inheritance does not.
 class FormatArg
 {
     public:
-        FormatArg() {}
+        FormatArg()
+            : m_value(NULL),
+            m_formatImpl(NULL),
+            m_toIntImpl(NULL)
+        { }
 
         template<typename T>
         FormatArg(const T& value)
-            : m_value(static_cast<const void*>(&value)),
+            // C-style cast here allows us to also remove volatile; we put it
+            // back in the *Impl functions before dereferencing to avoid UB.
+            : m_value((const void*)(&value)),
             m_formatImpl(&formatImpl<T>),
             m_toIntImpl(&toIntImpl<T>)
         { }
@@ -508,11 +533,15 @@ class FormatArg
         void format(std::ostream& out, const char* fmtBegin,
                     const char* fmtEnd, int ntrunc) const
         {
+            TINYFORMAT_ASSERT(m_value);
+            TINYFORMAT_ASSERT(m_formatImpl);
             m_formatImpl(out, fmtBegin, fmtEnd, ntrunc, m_value);
         }
 
         int toInt() const
         {
+            TINYFORMAT_ASSERT(m_value);
+            TINYFORMAT_ASSERT(m_toIntImpl);
             return m_toIntImpl(m_value);
         }
 
@@ -542,34 +571,68 @@ class FormatArg
 inline int parseIntAndAdvance(const char*& c)
 {
     int i = 0;
-    for(;*c >= '0' && *c <= '9'; ++c)
+    for (;*c >= '0' && *c <= '9'; ++c)
         i = 10*i + (*c - '0');
     return i;
 }
 
-// Print literal part of format string and return next format spec
-// position.
+// Parse width or precision `n` from format string pointer `c`, and advance it
+// to the next character. If an indirection is requested with `*`, the argument
+// is read from `args[argIndex]` and `argIndex` is incremented (or read
+// from `args[n]` in positional mode). Returns true if one or more
+// characters were read.
+inline bool parseWidthOrPrecision(int& n, const char*& c, bool positionalMode,
+                                  const detail::FormatArg* args,
+                                  int& argIndex, int numArgs)
+{
+    if (*c >= '0' && *c <= '9') {
+        n = parseIntAndAdvance(c);
+    }
+    else if (*c == '*') {
+        ++c;
+        n = 0;
+        if (positionalMode) {
+            int pos = parseIntAndAdvance(c) - 1;
+            if (*c != '$')
+                TINYFORMAT_ERROR("tinyformat: Non-positional argument used after a positional one");
+            if (pos >= 0 && pos < numArgs)
+                n = args[pos].toInt();
+            else
+                TINYFORMAT_ERROR("tinyformat: Positional argument out of range");
+            ++c;
+        }
+        else {
+            if (argIndex < numArgs)
+                n = args[argIndex++].toInt();
+            else
+                TINYFORMAT_ERROR("tinyformat: Not enough arguments to read variable width or precision");
+        }
+    }
+    else {
+        return false;
+    }
+    return true;
+}
+
+// Print literal part of format string and return next format spec position.
 //
-// Skips over any occurrences of '%%', printing a literal '%' to the
-// output.  The position of the first % character of the next
-// nontrivial format spec is returned, or the end of string.
+// Skips over any occurrences of '%%', printing a literal '%' to the output.
+// The position of the first % character of the next nontrivial format spec is
+// returned, or the end of string.
 inline const char* printFormatStringLiteral(std::ostream& out, const char* fmt)
 {
     const char* c = fmt;
-    for(;; ++c)
-    {
-        switch(*c)
-        {
-            case '\0':
-                out.write(fmt, static_cast<std::streamsize>(c - fmt));
+    for (;; ++c) {
+        if (*c == '\0') {
+            out.write(fmt, c - fmt);
+            return c;
+        }
+        else if (*c == '%') {
+            out.write(fmt, c - fmt);
+            if (*(c+1) != '%')
                 return c;
-            case '%':
-                out.write(fmt, static_cast<std::streamsize>(c - fmt));
-                if(*(c+1) != '%')
-                    return c;
-                // for "%%", tack trailing % onto next literal section.
-                fmt = ++c;
-                break;
+            // for "%%", tack trailing % onto next literal section.
+            fmt = ++c;
         }
     }
 }
@@ -578,23 +641,43 @@ inline const char* printFormatStringLiteral(std::ostream& out, const char* fmt)
 // Parse a format string and set the stream state accordingly.
 //
 // The format mini-language recognized here is meant to be the one from C99,
-// with the form "%[flags][width][.precision][length]type".
+// with the form "%[flags][width][.precision][length]type" with POSIX
+// positional arguments extension.
+//
+// POSIX positional arguments extension:
+// Conversions can be applied to the nth argument after the format in
+// the argument list, rather than to the next unused argument. In this case,
+// the conversion specifier character % (see below) is replaced by the sequence
+// "%n$", where n is a decimal integer in the range [1,{NL_ARGMAX}],
+// giving the position of the argument in the argument list. This feature
+// provides for the definition of format strings that select arguments
+// in an order appropriate to specific languages.
+//
+// The format can contain either numbered argument conversion specifications
+// (that is, "%n$" and "*m$"), or unnumbered argument conversion specifications
+// (that is, % and * ), but not both. The only exception to this is that %%
+// can be mixed with the "%n$" form. The results of mixing numbered and
+// unnumbered argument specifications in a format string are undefined.
+// When numbered argument specifications are used, specifying the Nth argument
+// requires that all the leading arguments, from the first to the (N-1)th,
+// are specified in the format string.
+//
+// In format strings containing the "%n$" form of conversion specification,
+// numbered arguments in the argument list can be referenced from the format
+// string as many times as required.
 //
 // Formatting options which can't be natively represented using the ostream
 // state are returned in spacePadPositive (for space padded positive numbers)
 // and ntrunc (for truncating conversions).  argIndex is incremented if
-// necessary to pull out variable width and precision .  The function returns a
+// necessary to pull out variable width and precision.  The function returns a
 // pointer to the character after the end of the current format spec.
-inline const char* streamStateFromFormat(std::ostream& out, bool& spacePadPositive,
+inline const char* streamStateFromFormat(std::ostream& out, bool& positionalMode,
+                                         bool& spacePadPositive,
                                          int& ntrunc, const char* fmtStart,
-                                         const detail::FormatArg* formatters,
-                                         int& argIndex, int numFormatters)
+                                         const detail::FormatArg* args,
+                                         int& argIndex, int numArgs)
 {
-    if(*fmtStart != '%')
-    {
-        TINYFORMAT_ERROR("tinyformat: Not enough conversion specifiers in format string");
-        return fmtStart;
-    }
+    TINYFORMAT_ASSERT(*fmtStart == '%');
     // Reset stream state to defaults.
     out.width(0);
     out.precision(6);
@@ -607,98 +690,113 @@ inline const char* streamStateFromFormat(std::ostream& out, bool& spacePadPositi
     bool widthSet = false;
     int widthExtra = 0;
     const char* c = fmtStart + 1;
-    // 1) Parse flags
-    for(;; ++c)
-    {
-        switch(*c)
-        {
-            case '#':
-                out.setf(std::ios::showpoint | std::ios::showbase);
-                continue;
-            case '0':
-                // overridden by left alignment ('-' flag)
-                if(!(out.flags() & std::ios::left))
-                {
-                    // Use internal padding so that numeric values are
-                    // formatted correctly, eg -00010 rather than 000-10
-                    out.fill('0');
-                    out.setf(std::ios::internal, std::ios::adjustfield);
-                }
-                continue;
-            case '-':
-                out.fill(' ');
-                out.setf(std::ios::left, std::ios::adjustfield);
-                continue;
-            case ' ':
-                // overridden by show positive sign, '+' flag.
-                if(!(out.flags() & std::ios::showpos))
-                    spacePadPositive = true;
-                continue;
-            case '+':
-                out.setf(std::ios::showpos);
-                spacePadPositive = false;
-                widthExtra = 1;
-                continue;
+
+    // 1) Parse an argument index (if followed by '$') or a width possibly
+    // preceded with '0' flag.
+    if (*c >= '0' && *c <= '9') {
+        const char tmpc = *c;
+        int value = parseIntAndAdvance(c);
+        if (*c == '$') {
+            // value is an argument index
+            if (value > 0 && value <= numArgs)
+                argIndex = value - 1;
+            else
+                TINYFORMAT_ERROR("tinyformat: Positional argument out of range");
+            ++c;
+            positionalMode = true;
+        }
+        else if (positionalMode) {
+            TINYFORMAT_ERROR("tinyformat: Non-positional argument used after a positional one");
+        }
+        else {
+            if (tmpc == '0') {
+                // Use internal padding so that numeric values are
+                // formatted correctly, eg -00010 rather than 000-10
+                out.fill('0');
+                out.setf(std::ios::internal, std::ios::adjustfield);
+            }
+            if (value != 0) {
+                // Nonzero value means that we parsed width.
+                widthSet = true;
+                out.width(value);
+            }
         }
-        break;
     }
-    // 2) Parse width
-    if(*c >= '0' && *c <= '9')
-    {
-        widthSet = true;
-        out.width(parseIntAndAdvance(c));
+    else if (positionalMode) {
+        TINYFORMAT_ERROR("tinyformat: Non-positional argument used after a positional one");
     }
-    if(*c == '*')
-    {
-        widthSet = true;
+    // 2) Parse flags and width if we did not do it in previous step.
+    if (!widthSet) {
+        // Parse flags
+        for (;; ++c) {
+            switch (*c) {
+                case '#':
+                    out.setf(std::ios::showpoint | std::ios::showbase);
+                    continue;
+                case '0':
+                    // overridden by left alignment ('-' flag)
+                    if (!(out.flags() & std::ios::left)) {
+                        // Use internal padding so that numeric values are
+                        // formatted correctly, eg -00010 rather than 000-10
+                        out.fill('0');
+                        out.setf(std::ios::internal, std::ios::adjustfield);
+                    }
+                    continue;
+                case '-':
+                    out.fill(' ');
+                    out.setf(std::ios::left, std::ios::adjustfield);
+                    continue;
+                case ' ':
+                    // overridden by show positive sign, '+' flag.
+                    if (!(out.flags() & std::ios::showpos))
+                        spacePadPositive = true;
+                    continue;
+                case '+':
+                    out.setf(std::ios::showpos);
+                    spacePadPositive = false;
+                    widthExtra = 1;
+                    continue;
+                default:
+                    break;
+            }
+            break;
+        }
+        // Parse width
         int width = 0;
-        if(argIndex < numFormatters)
-            width = formatters[argIndex++].toInt();
-        else
-            TINYFORMAT_ERROR("tinyformat: Not enough arguments to read variable width");
-        if(width < 0)
-        {
-            // negative widths correspond to '-' flag set
-            out.fill(' ');
-            out.setf(std::ios::left, std::ios::adjustfield);
-            width = -width;
+        widthSet = parseWidthOrPrecision(width, c, positionalMode,
+                                         args, argIndex, numArgs);
+        if (widthSet) {
+            if (width < 0) {
+                // negative widths correspond to '-' flag set
+                out.fill(' ');
+                out.setf(std::ios::left, std::ios::adjustfield);
+                width = -width;
+            }
+            out.width(width);
         }
-        out.width(width);
-        ++c;
     }
     // 3) Parse precision
-    if(*c == '.')
-    {
+    if (*c == '.') {
         ++c;
         int precision = 0;
-        if(*c == '*')
-        {
-            ++c;
-            if(argIndex < numFormatters)
-                precision = formatters[argIndex++].toInt();
-            else
-                TINYFORMAT_ERROR("tinyformat: Not enough arguments to read variable precision");
-        }
-        else
-        {
-            if(*c >= '0' && *c <= '9')
-                precision = parseIntAndAdvance(c);
-            else if(*c == '-') // negative precisions ignored, treated as zero.
-                parseIntAndAdvance(++c);
-        }
-        out.precision(precision);
-        precisionSet = true;
+        parseWidthOrPrecision(precision, c, positionalMode,
+                              args, argIndex, numArgs);
+        // Presence of `.` indicates precision set, unless the inferred value
+        // was negative in which case the default is used.
+        precisionSet = precision >= 0;
+        if (precisionSet)
+            out.precision(precision);
     }
     // 4) Ignore any C99 length modifier
-    while(*c == 'l' || *c == 'h' || *c == 'L' ||
-          *c == 'j' || *c == 'z' || *c == 't')
+    while (*c == 'l' || *c == 'h' || *c == 'L' ||
+           *c == 'j' || *c == 'z' || *c == 't') {
         ++c;
+    }
     // 5) We're up to the conversion specifier character.
     // Set stream flags based on conversion specifier (thanks to the
     // boost::format class for forging the way here).
     bool intConversion = false;
-    switch(*c)
-    {
+    switch (*c) {
         case 'u': case 'd': case 'i':
             out.setf(std::ios::dec, std::ios::basefield);
             intConversion = true;
@@ -709,39 +807,51 @@ inline const char* streamStateFromFormat(std::ostream& out, bool& spacePadPositi
             break;
         case 'X':
             out.setf(std::ios::uppercase);
+            // Falls through
         case 'x': case 'p':
             out.setf(std::ios::hex, std::ios::basefield);
             intConversion = true;
             break;
         case 'E':
             out.setf(std::ios::uppercase);
+            // Falls through
         case 'e':
             out.setf(std::ios::scientific, std::ios::floatfield);
             out.setf(std::ios::dec, std::ios::basefield);
             break;
         case 'F':
             out.setf(std::ios::uppercase);
+            // Falls through
         case 'f':
             out.setf(std::ios::fixed, std::ios::floatfield);
             break;
+        case 'A':
+            out.setf(std::ios::uppercase);
+            // Falls through
+        case 'a':
+#           ifdef _MSC_VER
+            // Workaround https://developercommunity.visualstudio.com/content/problem/520472/hexfloat-stream-output-does-not-ignore-precision-a.html
+            // by always setting maximum precision on MSVC to avoid precision
+            // loss for doubles.
+            out.precision(13);
+#           endif
+            out.setf(std::ios::fixed | std::ios::scientific, std::ios::floatfield);
+            break;
         case 'G':
             out.setf(std::ios::uppercase);
+            // Falls through
         case 'g':
             out.setf(std::ios::dec, std::ios::basefield);
             // As in boost::format, let stream decide float format.
             out.flags(out.flags() & ~std::ios::floatfield);
             break;
-        case 'a': case 'A':
-            TINYFORMAT_ERROR("tinyformat: the %a and %A conversion specs "
-                             "are not supported");
-            break;
         case 'c':
             // Handled as special case inside formatValue()
             break;
         case 's':
-            if(precisionSet)
+            if (precisionSet)
                 ntrunc = static_cast<int>(out.precision());
-            // Make %s print booleans as "true" and "false"
+            // Make %s print Booleans as "true" and "false"
             out.setf(std::ios::boolalpha);
             break;
         case 'n':
@@ -752,9 +862,10 @@ inline const char* streamStateFromFormat(std::ostream& out, bool& spacePadPositi
             TINYFORMAT_ERROR("tinyformat: Conversion spec incorrectly "
                              "terminated by end of string");
             return c;
+        default:
+            break;
     }
-    if(intConversion && precisionSet && !widthSet)
-    {
+    if (intConversion && precisionSet && !widthSet) {
         // "precision" for integers gives the minimum number of digits (to be
         // padded with zeros on the left).  This isn't really supported by the
         // iostreams, but we can approximately simulate it with the width if
@@ -769,8 +880,8 @@ inline const char* streamStateFromFormat(std::ostream& out, bool& spacePadPositi
 
 //------------------------------------------------------------------------------
 inline void formatImpl(std::ostream& out, const char* fmt,
-                       const detail::FormatArg* formatters,
-                       int numFormatters)
+                       const detail::FormatArg* args,
+                       int numArgs)
 {
     // Saved stream state
     std::streamsize origWidth = out.width();
@@ -778,26 +889,34 @@ inline void formatImpl(std::ostream& out, const char* fmt,
     std::ios::fmtflags origFlags = out.flags();
     char origFill = out.fill();
 
-    for (int argIndex = 0; argIndex < numFormatters; ++argIndex)
-    {
-        // Parse the format string
+    // "Positional mode" means all format specs should be of the form "%n$..."
+    // with `n` an integer. We detect this in `streamStateFromFormat`.
+    bool positionalMode = false;
+    int argIndex = 0;
+    while (true) {
         fmt = printFormatStringLiteral(out, fmt);
+        if (*fmt == '\0') {
+            if (!positionalMode && argIndex < numArgs) {
+                TINYFORMAT_ERROR("tinyformat: Not enough conversion specifiers in format string");
+            }
+            break;
+        }
         bool spacePadPositive = false;
         int ntrunc = -1;
-        const char* fmtEnd = streamStateFromFormat(out, spacePadPositive, ntrunc, fmt,
-                                                   formatters, argIndex, numFormatters);
-        if (argIndex >= numFormatters)
-        {
-            // Check args remain after reading any variable width/precision
-            TINYFORMAT_ERROR("tinyformat: Not enough format arguments");
+        const char* fmtEnd = streamStateFromFormat(out, positionalMode, spacePadPositive, ntrunc, fmt,
+                                                   args, argIndex, numArgs);
+        // NB: argIndex may be incremented by reading variable width/precision
+        // in `streamStateFromFormat`, so do the bounds check here.
+        if (argIndex >= numArgs) {
+            TINYFORMAT_ERROR("tinyformat: Too many conversion specifiers in format string");
             return;
         }
-        const FormatArg& arg = formatters[argIndex];
+        const FormatArg& arg = args[argIndex];
         // Format the arg into the stream.
-        if(!spacePadPositive)
+        if (!spacePadPositive) {
             arg.format(out, fmt, fmtEnd, ntrunc);
-        else
-        {
+        }
+        else {
             // The following is a special case with no direct correspondence
             // between stream formatting and the printf() behaviour.  Simulate
             // it crudely by formatting into a temporary string stream and
@@ -807,18 +926,17 @@ inline void formatImpl(std::ostream& out, const char* fmt,
             tmpStream.setf(std::ios::showpos);
             arg.format(tmpStream, fmt, fmtEnd, ntrunc);
             std::string result = tmpStream.str(); // allocates... yuck.
-            for(size_t i = 0, iend = result.size(); i < iend; ++i)
-                if(result[i] == '+') result[i] = ' ';
+            for (size_t i = 0, iend = result.size(); i < iend; ++i) {
+                if (result[i] == '+')
+                    result[i] = ' ';
+            }
             out << result;
         }
+        if (!positionalMode)
+            ++argIndex;
         fmt = fmtEnd;
     }
 
-    // Print remaining part of format string.
-    fmt = printFormatStringLiteral(out, fmt);
-    if(*fmt != '\0')
-        TINYFORMAT_ERROR("tinyformat: Too many conversion specifiers in format string");
-
     // Restore stream state
     out.width(origWidth);
     out.precision(origPrecision);
@@ -838,17 +956,14 @@ inline void formatImpl(std::ostream& out, const char* fmt,
 class FormatList
 {
     public:
+        FormatList(detail::FormatArg* args, int N)
+            : m_args(args), m_N(N) { }
+
         friend void vformat(std::ostream& out, const char* fmt,
                             const FormatList& list);
 
-    protected:
-        void setFormatters(detail::FormatArg* formatters, int N) {
-            m_formatters = formatters;
-            m_N = N;
-        }
-
     private:
-        const detail::FormatArg* m_formatters;
+        const detail::FormatArg* m_args;
         int m_N;
 };
 
@@ -859,53 +974,48 @@ typedef const FormatList& FormatListRef;
 namespace detail {
 
 // Format list subclass with fixed storage to avoid dynamic allocation
-template<int N>
+template<std::size_t N>
 class FormatListN : public FormatList
 {
     public:
 #ifdef TINYFORMAT_USE_VARIADIC_TEMPLATES
         template<typename... Args>
         FormatListN(const Args&... args)
-            : m_formatterStore TINYFORMAT_BRACED_INIT_WORKAROUND({ FormatArg(args)... })
-        {
-            static_assert(sizeof...(args) == N, "Number of args must be N");
-            setFormatters(&m_formatterStore[0], N);
-        }
+            : FormatList(&m_formatterStore[0], N),
+            m_formatterStore { FormatArg(args)... }
+        { static_assert(sizeof...(args) == N, "Number of args must be N"); }
 #else // C++98 version
         void init(int) {}
-#       define TINYFORMAT_MAKE_FORMATLIST_CONSTRUCTOR(n)       \
-                                                               \
-        template<TINYFORMAT_ARGTYPES(n)>                       \
-        FormatListN(TINYFORMAT_VARARGS(n))                     \
-        {                                                      \
-          setFormatters(&m_formatterStore[0], n);             \
-          assert(n == N);                                      \
-          init(0, TINYFORMAT_PASSARGS(n));                     \
-        }                                                      \
-                                                               \
-        template<TINYFORMAT_ARGTYPES(n)>                       \
-        void init(int i, TINYFORMAT_VARARGS(n))                \
-        {                                                      \
-            m_formatterStore[i] = FormatArg(v1);               \
-            init(i+1 TINYFORMAT_PASSARGS_TAIL(n));             \
+#       define TINYFORMAT_MAKE_FORMATLIST_CONSTRUCTOR(n)                \
+                                                                        \
+        template<TINYFORMAT_ARGTYPES(n)>                                \
+        FormatListN(TINYFORMAT_VARARGS(n))                              \
+            : FormatList(&m_formatterStore[0], n)                       \
+        { TINYFORMAT_ASSERT(n == N); init(0, TINYFORMAT_PASSARGS(n)); } \
+                                                                        \
+        template<TINYFORMAT_ARGTYPES(n)>                                \
+        void init(int i, TINYFORMAT_VARARGS(n))                         \
+        {                                                               \
+            m_formatterStore[i] = FormatArg(v1);                        \
+            init(i+1 TINYFORMAT_PASSARGS_TAIL(n));                      \
         }
 
         TINYFORMAT_FOREACH_ARGNUM(TINYFORMAT_MAKE_FORMATLIST_CONSTRUCTOR)
 #       undef TINYFORMAT_MAKE_FORMATLIST_CONSTRUCTOR
 #endif
+        FormatListN(const FormatListN& other)
+            : FormatList(&m_formatterStore[0], N)
+        { std::copy(&other.m_formatterStore[0], &other.m_formatterStore[N],
+                    &m_formatterStore[0]); }
 
     private:
-#ifdef TINYFORMAT_USE_VARIADIC_TEMPLATES
-        std::array<FormatArg, N> m_formatterStore;
-#else // C++98 version
         FormatArg m_formatterStore[N];
-#endif
 };
 
 // Special 0-arg version - MSVC says zero-sized C array in struct is nonstandard
 template<> class FormatListN<0> : public FormatList
 {
-    public: FormatListN() { setFormatters(0, 0); }
+    public: FormatListN() : FormatList(0, 0) {}
 };
 
 } // namespace detail
@@ -951,7 +1061,7 @@ TINYFORMAT_FOREACH_ARGNUM(TINYFORMAT_MAKE_MAKEFORMATLIST)
 /// list of format arguments is held in a single function argument.
 inline void vformat(std::ostream& out, const char* fmt, FormatListRef list)
 {
-    detail::formatImpl(out, fmt, list.m_formatters, list.m_N);
+    detail::formatImpl(out, fmt, list.m_args, list.m_N);
 }
 
 
diff --git a/source/extern/xxh3.h b/source/extern/xxh3.h
index f7dc195..7e3ce68 100644
--- a/source/extern/xxh3.h
+++ b/source/extern/xxh3.h
@@ -1,7 +1,7 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
  * Development source file for `xxh3`
- * Copyright (C) 2019-2020 Yann Collet
+ * Copyright (C) 2019-2021 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
diff --git a/source/extern/xxh_x86dispatch.c b/source/extern/xxh_x86dispatch.c
index 399bad9..871b499 100644
--- a/source/extern/xxh_x86dispatch.c
+++ b/source/extern/xxh_x86dispatch.c
@@ -1,6 +1,6 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
- * Copyright (C) 2020 Yann Collet
+ * Copyright (C) 2020-2021 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
@@ -36,7 +36,7 @@
 /*!
  * @file xxh_x86dispatch.c
  *
- * Automatic dispatcher code for the @ref xxh3_family on x86-based targets.
+ * Automatic dispatcher code for the @ref XXH3_family on x86-based targets.
  *
  * Optional add-on.
  *
@@ -60,29 +60,28 @@ extern "C" {
  * @def XXH_X86DISPATCH_ALLOW_AVX
  * @brief Disables the AVX sanity check.
  *
- * Don't compile xxh_x86dispatch.c with options like `-mavx*`, `-march=native`,
- * or `/arch:AVX*`. It is intended to be compiled for the minimum target, and
+ * xxh_x86dispatch.c is intended to be compiled for the minimum target, and
  * it selectively enables SSE2, AVX2, and AVX512 when it is needed.
  *
- * Using this option _globally_ allows this feature, and therefore makes it
+ * Compiling with options like `-mavx*`, `-march=native`, or `/arch:AVX*`
+ * _globally_ will always enable this feature, and therefore makes it
  * undefined behavior to execute on any CPU without said feature.
  *
  * Even if the source code isn't directly using AVX intrinsics in a function,
  * the compiler can still generate AVX code from autovectorization and by
  * "upgrading" SSE2 intrinsics to use the VEX prefixes (a.k.a. AVX128).
  *
- * Use the same flags that you use to compile the rest of the program; this
- * file will safely generate SSE2, AVX2, and AVX512 without these flags.
- *
- * Define XXH_X86DISPATCH_ALLOW_AVX to ignore this check, and feel free to open
- * an issue if there is a target in the future where AVX is a default feature.
+ * Define XXH_X86DISPATCH_ALLOW_AVX to ignore this check,
+ * thus accepting that the produced binary will not work correctly
+ * on any CPU with less features than the ones stated at compilation time.
  */
 #ifdef XXH_DOXYGEN
 #  define XXH_X86DISPATCH_ALLOW_AVX
 #endif
 
 #if defined(__AVX__) && !defined(XXH_X86DISPATCH_ALLOW_AVX)
-#  error "Do not compile xxh_x86dispatch.c with AVX enabled! See the comment above."
+#  error "Error: if xxh_x86dispatch.c is compiled with AVX enabled, the resulting binary will crash on sse2-only cpus !! " \
+         "If you nonetheless want to do that, please enable the XXH_X86DISPATCH_ALLOW_AVX build variable"
 #endif
 
 #ifdef __has_include
@@ -207,6 +206,23 @@ extern "C" {
 #define XXH_X86DISPATCH
 #include "xxhash.h"
 
+#ifndef XXH_HAS_ATTRIBUTE
+#  ifdef __has_attribute
+#    define XXH_HAS_ATTRIBUTE(...) __has_attribute(__VA_ARGS__)
+#  else
+#    define XXH_HAS_ATTRIBUTE(...) 0
+#  endif
+#endif
+
+#if XXH_HAS_ATTRIBUTE(constructor)
+#  define XXH_CONSTRUCTOR __attribute__((constructor))
+#  define XXH_DISPATCH_MAYBE_NULL 0
+#else
+#  define XXH_CONSTRUCTOR
+#  define XXH_DISPATCH_MAYBE_NULL 1
+#endif
+
+
 /*
  * Support both AT&T and Intel dialects
  *
@@ -316,7 +332,7 @@ static xxh_u64 XXH_xgetbv(void)
  *
  * Runs various CPUID/XGETBV tests to try and determine the best implementation.
  *
- * @ret The best @ref XXH_VECTOR implementation.
+ * @return The best @ref XXH_VECTOR implementation.
  * @see XXH_VECTOR_TYPES
  */
 static int XXH_featureTest(void)
@@ -455,22 +471,23 @@ static int XXH_featureTest(void)
 /* ===   XXH3, default variants   === */                                      \
                                                                               \
 XXH_NO_INLINE target XXH64_hash_t                                             \
-XXHL64_default_##suffix(const void* XXH_RESTRICT input, size_t len)           \
+XXHL64_default_##suffix(XXH_NOESCAPE const void* XXH_RESTRICT input,          \
+                        size_t len)                                           \
 {                                                                             \
     return XXH3_hashLong_64b_internal(                                        \
                input, len, XXH3_kSecret, sizeof(XXH3_kSecret),                \
-               XXH3_accumulate_512_##suffix, XXH3_scrambleAcc_##suffix        \
+               XXH3_accumulate_##suffix, XXH3_scrambleAcc_##suffix            \
     );                                                                        \
 }                                                                             \
                                                                               \
 /* ===   XXH3, Seeded variants   === */                                       \
                                                                               \
 XXH_NO_INLINE target XXH64_hash_t                                             \
-XXHL64_seed_##suffix(const void* XXH_RESTRICT input, size_t len,              \
+XXHL64_seed_##suffix(XXH_NOESCAPE const void* XXH_RESTRICT input, size_t len, \
                      XXH64_hash_t seed)                                       \
 {                                                                             \
     return XXH3_hashLong_64b_withSeed_internal(                               \
-                    input, len, seed, XXH3_accumulate_512_##suffix,           \
+                    input, len, seed, XXH3_accumulate_##suffix,               \
                     XXH3_scrambleAcc_##suffix, XXH3_initCustomSecret_##suffix \
     );                                                                        \
 }                                                                             \
@@ -478,54 +495,59 @@ XXHL64_seed_##suffix(const void* XXH_RESTRICT input, size_t len,              \
 /* ===   XXH3, Secret variants   === */                                       \
                                                                               \
 XXH_NO_INLINE target XXH64_hash_t                                             \
-XXHL64_secret_##suffix(const void* XXH_RESTRICT input, size_t len,            \
-                       const void* secret, size_t secretLen)                  \
+XXHL64_secret_##suffix(XXH_NOESCAPE const void* XXH_RESTRICT input,           \
+                       size_t len, XXH_NOESCAPE const void* secret,           \
+                       size_t secretLen)                                      \
 {                                                                             \
     return XXH3_hashLong_64b_internal(                                        \
                     input, len, secret, secretLen,                            \
-                    XXH3_accumulate_512_##suffix, XXH3_scrambleAcc_##suffix   \
+                    XXH3_accumulate_##suffix, XXH3_scrambleAcc_##suffix       \
     );                                                                        \
 }                                                                             \
                                                                               \
 /* ===   XXH3 update variants   === */                                        \
                                                                               \
 XXH_NO_INLINE target XXH_errorcode                                            \
-XXH3_update_##suffix(XXH3_state_t* state, const void* input, size_t len)      \
+XXH3_update_##suffix(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_##suffix, XXH3_scrambleAcc_##suffix); \
+                    XXH3_accumulate_##suffix, XXH3_scrambleAcc_##suffix);     \
 }                                                                             \
                                                                               \
 /* ===   XXH128 default variants   === */                                     \
                                                                               \
 XXH_NO_INLINE target XXH128_hash_t                                            \
-XXHL128_default_##suffix(const void* XXH_RESTRICT input, size_t len)          \
+XXHL128_default_##suffix(XXH_NOESCAPE  const void* XXH_RESTRICT input,        \
+                         size_t len)                                          \
 {                                                                             \
     return XXH3_hashLong_128b_internal(                                       \
                     input, len, XXH3_kSecret, sizeof(XXH3_kSecret),           \
-                    XXH3_accumulate_512_##suffix, XXH3_scrambleAcc_##suffix   \
+                    XXH3_accumulate_##suffix, XXH3_scrambleAcc_##suffix       \
     );                                                                        \
 }                                                                             \
                                                                               \
 /* ===   XXH128 Secret variants   === */                                      \
                                                                               \
 XXH_NO_INLINE target XXH128_hash_t                                            \
-XXHL128_secret_##suffix(const void* XXH_RESTRICT input, size_t len,           \
-                        const void* XXH_RESTRICT secret, size_t secretLen)    \
+XXHL128_secret_##suffix(XXH_NOESCAPE const void* XXH_RESTRICT input,          \
+                        size_t len,                                           \
+                        XXH_NOESCAPE const void* XXH_RESTRICT secret,         \
+                        size_t secretLen)                                     \
 {                                                                             \
     return XXH3_hashLong_128b_internal(                                       \
                     input, len, (const xxh_u8*)secret, secretLen,             \
-                    XXH3_accumulate_512_##suffix, XXH3_scrambleAcc_##suffix); \
+                    XXH3_accumulate_##suffix, XXH3_scrambleAcc_##suffix);     \
 }                                                                             \
                                                                               \
 /* ===   XXH128 Seeded variants   === */                                      \
                                                                               \
 XXH_NO_INLINE target XXH128_hash_t                                            \
-XXHL128_seed_##suffix(const void* XXH_RESTRICT input, size_t len,             \
+XXHL128_seed_##suffix(XXH_NOESCAPE const void* XXH_RESTRICT input, size_t len,\
                       XXH64_hash_t seed)                                      \
 {                                                                             \
     return XXH3_hashLong_128b_withSeed_internal(input, len, seed,             \
-                    XXH3_accumulate_512_##suffix, XXH3_scrambleAcc_##suffix,  \
+                    XXH3_accumulate_##suffix, XXH3_scrambleAcc_##suffix,      \
                     XXH3_initCustomSecret_##suffix);                          \
 }
 
@@ -545,13 +567,13 @@ XXH_DEFINE_DISPATCH_FUNCS(avx512, XXH_TARGET_AVX512)
 
 /* ====    Dispatchers    ==== */
 
-typedef XXH64_hash_t (*XXH3_dispatchx86_hashLong64_default)(const void* XXH_RESTRICT, size_t);
+typedef XXH64_hash_t (*XXH3_dispatchx86_hashLong64_default)(XXH_NOESCAPE const void* XXH_RESTRICT, size_t);
 
-typedef XXH64_hash_t (*XXH3_dispatchx86_hashLong64_withSeed)(const void* XXH_RESTRICT, size_t, XXH64_hash_t);
+typedef XXH64_hash_t (*XXH3_dispatchx86_hashLong64_withSeed)(XXH_NOESCAPE const void* XXH_RESTRICT, size_t, XXH64_hash_t);
 
-typedef XXH64_hash_t (*XXH3_dispatchx86_hashLong64_withSecret)(const void* XXH_RESTRICT, size_t, const void* XXH_RESTRICT, size_t);
+typedef XXH64_hash_t (*XXH3_dispatchx86_hashLong64_withSecret)(XXH_NOESCAPE const void* XXH_RESTRICT, size_t, XXH_NOESCAPE const void* XXH_RESTRICT, size_t);
 
-typedef XXH_errorcode (*XXH3_dispatchx86_update)(XXH3_state_t*, const void*, size_t);
+typedef XXH_errorcode (*XXH3_dispatchx86_update)(XXH_NOESCAPE XXH3_state_t*, XXH_NOESCAPE const void*, size_t);
 
 typedef struct {
     XXH3_dispatchx86_hashLong64_default    hashLong64_default;
@@ -593,11 +615,11 @@ static const XXH_dispatchFunctions_s XXH_kDispatch[XXH_NB_DISPATCHES] = {
 static XXH_dispatchFunctions_s XXH_g_dispatch = { NULL, NULL, NULL, NULL };
 
 
-typedef XXH128_hash_t (*XXH3_dispatchx86_hashLong128_default)(const void* XXH_RESTRICT, size_t);
+typedef XXH128_hash_t (*XXH3_dispatchx86_hashLong128_default)(XXH_NOESCAPE const void* XXH_RESTRICT, size_t);
 
-typedef XXH128_hash_t (*XXH3_dispatchx86_hashLong128_withSeed)(const void* XXH_RESTRICT, size_t, XXH64_hash_t);
+typedef XXH128_hash_t (*XXH3_dispatchx86_hashLong128_withSeed)(XXH_NOESCAPE const void* XXH_RESTRICT, size_t, XXH64_hash_t);
 
-typedef XXH128_hash_t (*XXH3_dispatchx86_hashLong128_withSecret)(const void* XXH_RESTRICT, size_t, const void* XXH_RESTRICT, size_t);
+typedef XXH128_hash_t (*XXH3_dispatchx86_hashLong128_withSecret)(XXH_NOESCAPE const void* XXH_RESTRICT, size_t, const void* XXH_RESTRICT, size_t);
 
 typedef struct {
     XXH3_dispatchx86_hashLong128_default    hashLong128_default;
@@ -642,7 +664,7 @@ static XXH_dispatch128Functions_s XXH_g_dispatch128 = { NULL, NULL, NULL, NULL }
  * @internal
  * @brief Runs a CPUID check and sets the correct dispatch tables.
  */
-static void XXH_setDispatch(void)
+static XXH_CONSTRUCTOR void XXH_setDispatch(void)
 {
     int vecID = XXH_featureTest();
     XXH_STATIC_ASSERT(XXH_AVX512 == XXH_NB_DISPATCHES-1);
@@ -668,11 +690,12 @@ XXH3_hashLong_64b_defaultSecret_selection(const void* input, size_t len,
                                           XXH64_hash_t seed64, const xxh_u8* secret, size_t secretLen)
 {
     (void)seed64; (void)secret; (void)secretLen;
-    if (XXH_g_dispatch.hashLong64_default == NULL) XXH_setDispatch();
+    if (XXH_DISPATCH_MAYBE_NULL && XXH_g_dispatch.hashLong64_default == NULL)
+        XXH_setDispatch();
     return XXH_g_dispatch.hashLong64_default(input, len);
 }
 
-XXH64_hash_t XXH3_64bits_dispatch(const void* input, size_t len)
+XXH64_hash_t XXH3_64bits_dispatch(XXH_NOESCAPE const void* input, size_t len)
 {
     return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_defaultSecret_selection);
 }
@@ -682,11 +705,12 @@ XXH3_hashLong_64b_withSeed_selection(const void* input, size_t len,
                                      XXH64_hash_t seed64, const xxh_u8* secret, size_t secretLen)
 {
     (void)secret; (void)secretLen;
-    if (XXH_g_dispatch.hashLong64_seed == NULL) XXH_setDispatch();
+    if (XXH_DISPATCH_MAYBE_NULL && XXH_g_dispatch.hashLong64_seed == NULL)
+        XXH_setDispatch();
     return XXH_g_dispatch.hashLong64_seed(input, len, seed64);
 }
 
-XXH64_hash_t XXH3_64bits_withSeed_dispatch(const void* input, size_t len, XXH64_hash_t seed)
+XXH64_hash_t XXH3_64bits_withSeed_dispatch(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
     return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed_selection);
 }
@@ -696,19 +720,22 @@ XXH3_hashLong_64b_withSecret_selection(const void* input, size_t len,
                                        XXH64_hash_t seed64, const xxh_u8* secret, size_t secretLen)
 {
     (void)seed64;
-    if (XXH_g_dispatch.hashLong64_secret == NULL) XXH_setDispatch();
+    if (XXH_DISPATCH_MAYBE_NULL && XXH_g_dispatch.hashLong64_secret == NULL)
+        XXH_setDispatch();
     return XXH_g_dispatch.hashLong64_secret(input, len, secret, secretLen);
 }
 
-XXH64_hash_t XXH3_64bits_withSecret_dispatch(const void* input, size_t len, const void* secret, size_t secretLen)
+XXH64_hash_t XXH3_64bits_withSecret_dispatch(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretLen)
 {
     return XXH3_64bits_internal(input, len, 0, secret, secretLen, XXH3_hashLong_64b_withSecret_selection);
 }
 
 XXH_errorcode
-XXH3_64bits_update_dispatch(XXH3_state_t* state, const void* input, size_t len)
+XXH3_64bits_update_dispatch(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
-    if (XXH_g_dispatch.update == NULL) XXH_setDispatch();
+    if (XXH_DISPATCH_MAYBE_NULL && XXH_g_dispatch.update == NULL)
+        XXH_setDispatch();
+
     return XXH_g_dispatch.update(state, (const xxh_u8*)input, len);
 }
 
@@ -720,25 +747,27 @@ XXH3_hashLong_128b_defaultSecret_selection(const void* input, size_t len,
                                            XXH64_hash_t seed64, const void* secret, size_t secretLen)
 {
     (void)seed64; (void)secret; (void)secretLen;
-    if (XXH_g_dispatch128.hashLong128_default == NULL) XXH_setDispatch();
+    if (XXH_DISPATCH_MAYBE_NULL && XXH_g_dispatch128.hashLong128_default == NULL)
+        XXH_setDispatch();
     return XXH_g_dispatch128.hashLong128_default(input, len);
 }
 
-XXH128_hash_t XXH3_128bits_dispatch(const void* input, size_t len)
+XXH128_hash_t XXH3_128bits_dispatch(XXH_NOESCAPE const void* input, size_t len)
 {
     return XXH3_128bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_128b_defaultSecret_selection);
 }
 
 static XXH128_hash_t
 XXH3_hashLong_128b_withSeed_selection(const void* input, size_t len,
-                                     XXH64_hash_t seed64, const void* secret, size_t secretLen)
+                                      XXH64_hash_t seed64, const void* secret, size_t secretLen)
 {
     (void)secret; (void)secretLen;
-    if (XXH_g_dispatch128.hashLong128_seed == NULL) XXH_setDispatch();
+    if (XXH_DISPATCH_MAYBE_NULL && XXH_g_dispatch128.hashLong128_seed == NULL)
+        XXH_setDispatch();
     return XXH_g_dispatch128.hashLong128_seed(input, len, seed64);
 }
 
-XXH128_hash_t XXH3_128bits_withSeed_dispatch(const void* input, size_t len, XXH64_hash_t seed)
+XXH128_hash_t XXH3_128bits_withSeed_dispatch(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_selection);
 }
@@ -748,19 +777,21 @@ XXH3_hashLong_128b_withSecret_selection(const void* input, size_t len,
                                         XXH64_hash_t seed64, const void* secret, size_t secretLen)
 {
     (void)seed64;
-    if (XXH_g_dispatch128.hashLong128_secret == NULL) XXH_setDispatch();
+    if (XXH_DISPATCH_MAYBE_NULL && XXH_g_dispatch128.hashLong128_secret == NULL)
+        XXH_setDispatch();
     return XXH_g_dispatch128.hashLong128_secret(input, len, secret, secretLen);
 }
 
-XXH128_hash_t XXH3_128bits_withSecret_dispatch(const void* input, size_t len, const void* secret, size_t secretLen)
+XXH128_hash_t XXH3_128bits_withSecret_dispatch(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretLen)
 {
     return XXH3_128bits_internal(input, len, 0, secret, secretLen, XXH3_hashLong_128b_withSecret_selection);
 }
 
 XXH_errorcode
-XXH3_128bits_update_dispatch(XXH3_state_t* state, const void* input, size_t len)
+XXH3_128bits_update_dispatch(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
-    if (XXH_g_dispatch128.update == NULL) XXH_setDispatch();
+    if (XXH_DISPATCH_MAYBE_NULL && XXH_g_dispatch128.update == NULL)
+        XXH_setDispatch();
     return XXH_g_dispatch128.update(state, (const xxh_u8*)input, len);
 }
 
diff --git a/source/extern/xxh_x86dispatch.h b/source/extern/xxh_x86dispatch.h
index 6bc17bc..b87cea9 100644
--- a/source/extern/xxh_x86dispatch.h
+++ b/source/extern/xxh_x86dispatch.h
@@ -1,6 +1,6 @@
 /*
  * xxHash - XXH3 Dispatcher for x86-based targets
- * Copyright (C) 2020 Yann Collet
+ * Copyright (C) 2020-2021 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
@@ -41,15 +41,15 @@
 extern "C" {
 #endif
 
-XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_dispatch(const void* input, size_t len);
-XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_withSeed_dispatch(const void* input, size_t len, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_withSecret_dispatch(const void* input, size_t len, const void* secret, size_t secretLen);
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update_dispatch(XXH3_state_t* state, const void* input, size_t len);
+XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_dispatch(XXH_NOESCAPE const void* input, size_t len);
+XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_withSeed_dispatch(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_withSecret_dispatch(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretLen);
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update_dispatch(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len);
 
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_dispatch(const void* input, size_t len);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed_dispatch(const void* input, size_t len, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret_dispatch(const void* input, size_t len, const void* secret, size_t secretLen);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update_dispatch(XXH3_state_t* state, const void* input, size_t len);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_dispatch(XXH_NOESCAPE const void* input, size_t len);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed_dispatch(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret_dispatch(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretLen);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update_dispatch(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len);
 
 #if defined (__cplusplus)
 }
@@ -71,7 +71,6 @@ XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update_dispatch(XXH3_state_t* state, c
 
 # undef  XXH128
 # define XXH128 XXH3_128bits_withSeed_dispatch
-# define XXH3_128bits XXH3_128bits_dispatch
 # undef  XXH3_128bits
 # define XXH3_128bits XXH3_128bits_dispatch
 # undef  XXH3_128bits_withSeed
diff --git a/source/extern/xxhash.c b/source/extern/xxhash.c
index 0fae88c..083b039 100644
--- a/source/extern/xxhash.c
+++ b/source/extern/xxhash.c
@@ -1,6 +1,6 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
- * Copyright (C) 2012-2020 Yann Collet
+ * Copyright (C) 2012-2021 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
diff --git a/source/extern/xxhash.h b/source/extern/xxhash.h
index 08ab794..a0758d9 100644
--- a/source/extern/xxhash.h
+++ b/source/extern/xxhash.h
@@ -1,7 +1,7 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
  * Header File
- * Copyright (C) 2012-2020 Yann Collet
+ * Copyright (C) 2012-2021 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
@@ -32,49 +32,142 @@
  *   - xxHash homepage: https://www.xxhash.com
  *   - xxHash source repository: https://github.com/Cyan4973/xxHash
  */
+
 /*!
  * @mainpage xxHash
  *
+ * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
+ * limits.
+ *
+ * It is proposed in four flavors, in three families:
+ * 1. @ref XXH32_family
+ *   - Classic 32-bit hash function. Simple, compact, and runs on almost all
+ *     32-bit and 64-bit systems.
+ * 2. @ref XXH64_family
+ *   - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
+ *     64-bit systems (but _not_ 32-bit systems).
+ * 3. @ref XXH3_family
+ *   - Modern 64-bit and 128-bit hash function family which features improved
+ *     strength and performance across the board, especially on smaller data.
+ *     It benefits greatly from SIMD and 64-bit without requiring it.
+ *
+ * Benchmarks
+ * ---
+ * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
+ * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
+ *
+ * | Hash Name            | ISA ext | Width | Large Data Speed | Small Data Velocity |
+ * | -------------------- | ------- | ----: | ---------------: | ------------------: |
+ * | XXH3_64bits()        | @b AVX2 |    64 |        59.4 GB/s |               133.1 |
+ * | MeowHash             | AES-NI  |   128 |        58.2 GB/s |                52.5 |
+ * | XXH3_128bits()       | @b AVX2 |   128 |        57.9 GB/s |               118.1 |
+ * | CLHash               | PCLMUL  |    64 |        37.1 GB/s |                58.1 |
+ * | XXH3_64bits()        | @b SSE2 |    64 |        31.5 GB/s |               133.1 |
+ * | XXH3_128bits()       | @b SSE2 |   128 |        29.6 GB/s |               118.1 |
+ * | RAM sequential read  |         |   N/A |        28.0 GB/s |                 N/A |
+ * | ahash                | AES-NI  |    64 |        22.5 GB/s |               107.2 |
+ * | City64               |         |    64 |        22.0 GB/s |                76.6 |
+ * | T1ha2                |         |    64 |        22.0 GB/s |                99.0 |
+ * | City128              |         |   128 |        21.7 GB/s |                57.7 |
+ * | FarmHash             | AES-NI  |    64 |        21.3 GB/s |                71.9 |
+ * | XXH64()              |         |    64 |        19.4 GB/s |                71.0 |
+ * | SpookyHash           |         |    64 |        19.3 GB/s |                53.2 |
+ * | Mum                  |         |    64 |        18.0 GB/s |                67.0 |
+ * | CRC32C               | SSE4.2  |    32 |        13.0 GB/s |                57.9 |
+ * | XXH32()              |         |    32 |         9.7 GB/s |                71.9 |
+ * | City32               |         |    32 |         9.1 GB/s |                66.0 |
+ * | Blake3*              | @b AVX2 |   256 |         4.4 GB/s |                 8.1 |
+ * | Murmur3              |         |    32 |         3.9 GB/s |                56.1 |
+ * | SipHash*             |         |    64 |         3.0 GB/s |                43.2 |
+ * | Blake3*              | @b SSE2 |   256 |         2.4 GB/s |                 8.1 |
+ * | HighwayHash          |         |    64 |         1.4 GB/s |                 6.0 |
+ * | FNV64                |         |    64 |         1.2 GB/s |                62.7 |
+ * | Blake2*              |         |   256 |         1.1 GB/s |                 5.1 |
+ * | SHA1*                |         |   160 |         0.8 GB/s |                 5.6 |
+ * | MD5*                 |         |   128 |         0.6 GB/s |                 7.8 |
+ * @note
+ *   - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
+ *     even though it is mandatory on x64.
+ *   - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
+ *     by modern standards.
+ *   - Small data velocity is a rough average of algorithm's efficiency for small
+ *     data. For more accurate information, see the wiki.
+ *   - More benchmarks and strength tests are found on the wiki:
+ *         https://github.com/Cyan4973/xxHash/wiki
+ *
+ * Usage
+ * ------
+ * All xxHash variants use a similar API. Changing the algorithm is a trivial
+ * substitution.
+ *
+ * @pre
+ *    For functions which take an input and length parameter, the following
+ *    requirements are assumed:
+ *    - The range from [`input`, `input + length`) is valid, readable memory.
+ *      - The only exception is if the `length` is `0`, `input` may be `NULL`.
+ *    - For C++, the objects must have the *TriviallyCopyable* property, as the
+ *      functions access bytes directly as if it was an array of `unsigned char`.
+ *
+ * @anchor single_shot_example
+ * **Single Shot**
+ *
+ * These functions are stateless functions which hash a contiguous block of memory,
+ * immediately returning the result. They are the easiest and usually the fastest
+ * option.
+ *
+ * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
+ *
+ * @code{.c}
+ *   #include <string.h>
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which hashes a null terminated string with XXH32().
+ *   XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
+ *   {
+ *       // NULL pointers are only valid if the length is zero
+ *       size_t length = (string == NULL) ? 0 : strlen(string);
+ *       return XXH32(string, length, seed);
+ *   }
+ * @endcode
+ *
+ * @anchor streaming_example
+ * **Streaming**
+ *
+ * These groups of functions allow incremental hashing of unknown size, even
+ * more than what would fit in a size_t.
+ *
+ * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
+ *
+ * @code{.c}
+ *   #include <stdio.h>
+ *   #include <assert.h>
+ *   #include "xxhash.h"
+ *   // Example for a function which hashes a FILE incrementally with XXH3_64bits().
+ *   XXH64_hash_t hashFile(FILE* f)
+ *   {
+ *       // Allocate a state struct. Do not just use malloc() or new.
+ *       XXH3_state_t* state = XXH3_createState();
+ *       assert(state != NULL && "Out of memory!");
+ *       // Reset the state to start a new hashing session.
+ *       XXH3_64bits_reset(state);
+ *       char buffer[4096];
+ *       size_t count;
+ *       // Read the file in chunks
+ *       while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
+ *           // Run update() as many times as necessary to process the data
+ *           XXH3_64bits_update(state, buffer, count);
+ *       }
+ *       // Retrieve the finalized hash. This will not change the state.
+ *       XXH64_hash_t result = XXH3_64bits_digest(state);
+ *       // Free the state. Do not use free().
+ *       XXH3_freeState(state);
+ *       return result;
+ *   }
+ * @endcode
+ *
  * @file xxhash.h
  * xxHash prototypes and implementation
  */
-/* TODO: update */
-/* Notice extracted from xxHash homepage:
-
-xxHash is an extremely fast hash algorithm, running at RAM speed limits.
-It also successfully passes all tests from the SMHasher suite.
-
-Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo @3GHz)
-
-Name            Speed       Q.Score   Author
-xxHash          5.4 GB/s     10
-CrapWow         3.2 GB/s      2       Andrew
-MurmurHash 3a   2.7 GB/s     10       Austin Appleby
-SpookyHash      2.0 GB/s     10       Bob Jenkins
-SBox            1.4 GB/s      9       Bret Mulvey
-Lookup3         1.2 GB/s      9       Bob Jenkins
-SuperFastHash   1.2 GB/s      1       Paul Hsieh
-CityHash64      1.05 GB/s    10       Pike & Alakuijala
-FNV             0.55 GB/s     5       Fowler, Noll, Vo
-CRC32           0.43 GB/s     9
-MD5-32          0.33 GB/s    10       Ronald L. Rivest
-SHA1-32         0.28 GB/s    10
-
-Q.Score is a measure of quality of the hash function.
-It depends on successfully passing SMHasher test set.
-10 is a perfect score.
-
-Note: SMHasher's CRC32 implementation is not the fastest one.
-Other speed-oriented implementations can be faster,
-especially in combination with PCLMUL instruction:
-https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html?showComment=1552696407071#c3490092340461170735
-
-A 64-bit version, named XXH64, is available since r35.
-It offers much better speed, but for 64-bit applications only.
-Name     Speed on 64 bits    Speed on 32 bits
-XXH64       13.8 GB/s            1.9 GB/s
-XXH32        6.8 GB/s            6.0 GB/s
-*/
 
 #if defined (__cplusplus)
 extern "C" {
@@ -84,21 +177,53 @@ extern "C" {
  *  INLINE mode
  ******************************/
 /*!
- * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Exposes the implementation and marks all functions as `inline`.
+ *
  * Use these build macros to inline xxhash into the target unit.
  * Inlining improves performance on small inputs, especially when the length is
  * expressed as a compile-time constant:
  *
- *      https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
+ *  https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
  *
  * It also keeps xxHash symbols private to the unit, so they are not exported.
  *
  * Usage:
+ * @code{.c}
  *     #define XXH_INLINE_ALL
  *     #include "xxhash.h"
- *
+ * @endcode
  * Do not compile and link xxhash.o as a separate object, as it is not useful.
  */
+#  define XXH_INLINE_ALL
+#  undef XXH_INLINE_ALL
+/*!
+ * @brief Exposes the implementation without marking functions as inline.
+ */
+#  define XXH_PRIVATE_API
+#  undef XXH_PRIVATE_API
+/*!
+ * @brief Emulate a namespace by transparently prefixing all symbols.
+ *
+ * If you want to include _and expose_ xxHash functions from within your own
+ * library, but also want to avoid symbol collisions with other libraries which
+ * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
+ * (therefore, avoid empty or numeric values).
+ *
+ * Note that no change is required within the calling program as long as it
+ * includes `xxhash.h`: Regular symbol names will be automatically translated
+ * by this header.
+ */
+#  define XXH_NAMESPACE /* YOUR NAME HERE */
+#  undef XXH_NAMESPACE
+#endif
+
 #if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
     && !defined(XXH_INLINE_ALL_31684351384)
    /* this section should be traversed only once */
@@ -213,21 +338,13 @@ 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
@@ -240,24 +357,6 @@ 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)
@@ -318,11 +417,39 @@ 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
+/*! @brief Version number, encoded as two digits each */
 #define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
 
 /*!
@@ -331,16 +458,22 @@ 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;
 
 
 /*-**********************************************************************
@@ -364,29 +497,27 @@ typedef uint32_t XXH32_hash_t;
 #   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
  * @{
  */
 
@@ -395,6 +526,8 @@ typedef uint32_t XXH32_hash_t;
  *
  * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s
  *
+ * See @ref single_shot_example "Single Shot Example" for an example.
+ *
  * @param input The block of data to be hashed, at least @p length bytes in size.
  * @param length The length of @p input, in bytes.
  * @param seed The 32-bit seed to alter the hash's output predictably.
@@ -412,8 +545,9 @@ typedef uint32_t XXH32_hash_t;
  * @see
  *    XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version.
  */
-XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
 
+#ifndef XXH_NO_STREAM
 /*!
  * Streaming functions generate the xxHash value from an incremental input.
  * This method is slower than single-call functions, due to state management.
@@ -436,32 +570,7 @@ XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_
  *
  * When done, release the state using `XXH*_freeState()`.
  *
- * Example code for incrementally hashing a file:
- * @code{.c}
- *    #include <stdio.h>
- *    #include <xxhash.h>
- *    #define BUFFER_SIZE 256
- *
- *    // Note: XXH64 and XXH3 use the same interface.
- *    XXH32_hash_t
- *    hashFile(FILE* stream)
- *    {
- *        XXH32_state_t* state;
- *        unsigned char buf[BUFFER_SIZE];
- *        size_t amt;
- *        XXH32_hash_t hash;
- *
- *        state = XXH32_createState();       // Create a state
- *        assert(state != NULL);             // Error check here
- *        XXH32_reset(state, 0xbaad5eed);    // Reset state with our seed
- *        while ((amt = fread(buf, 1, sizeof(buf), stream)) != 0) {
- *            XXH32_update(state, buf, amt); // Hash the file in chunks
- *        }
- *        hash = XXH32_digest(state);        // Finalize the hash
- *        XXH32_freeState(state);            // Clean up
- *        return hash;
- *    }
- * @endcode
+ * @see streaming_example at the top of @ref xxhash.h for an example.
  */
 
 /*!
@@ -478,7 +587,7 @@ typedef struct XXH32_state_s XXH32_state_t;
  * Must be freed with XXH32_freeState().
  * @return An allocated XXH32_state_t on success, `NULL` on failure.
  */
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
+XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
 /*!
  * @brief Frees an @ref XXH32_state_t.
  *
@@ -546,7 +655,8 @@ XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void*
  *
  * @return The calculated xxHash32 value from that state.
  */
-XXH_PUBLIC_API XXH32_hash_t  XXH32_digest (const XXH32_state_t* statePtr);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
 /*******   Canonical representation   *******/
 
@@ -597,7 +707,7 @@ XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t
  *
  * @return The converted hash.
  */
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
 
 
 #ifdef __has_attribute
@@ -606,8 +716,15 @@ XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src
 # define XXH_HAS_ATTRIBUTE(x) 0
 #endif
 
+/*
+ * C23 __STDC_VERSION__ number hasn't been specified yet. For now
+ * leave as `201711L` (C17 + 1).
+ * TODO: Update to correct value when its been specified.
+ */
+#define XXH_C23_VN 201711L
+
 /* C-language Attributes are added in C23. */
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
 # define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
 #else
 # define XXH_HAS_C_ATTRIBUTE(x) 0
@@ -620,21 +737,31 @@ XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src
 #endif
 
 /*
-Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
-introduced in CPP17 and C23.
-CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
-C23   : https://en.cppreference.com/w/c/language/attributes/fallthrough
-*/
-#if XXH_HAS_C_ATTRIBUTE(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 /* fallthrough */
+#endif
+
+/*
+ * Define XXH_NOESCAPE for annotated pointers in public API.
+ * https://clang.llvm.org/docs/AttributeReference.html#noescape
+ * As of writing this, only supported by clang.
+ */
+#if XXH_HAS_ATTRIBUTE(noescape)
+# define XXH_NOESCAPE __attribute__((noescape))
 #else
-# define XXH_FALLTHROUGH
+# define XXH_NOESCAPE
 #endif
 
+
 /*!
  * @}
  * @ingroup public
@@ -671,7 +798,7 @@ 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.
@@ -682,7 +809,6 @@ 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.
  *
@@ -706,32 +832,35 @@ typedef uint64_t XXH64_hash_t;
  * @see
  *    XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH64(const void* input, size_t length, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
 
 /*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*!
  * @brief The opaque state struct for the XXH64 streaming API.
  *
  * @see XXH64_state_s for details.
  */
 typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
+XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
 XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dst_state, const XXH64_state_t* src_state);
-
-XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH64_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH64_digest (const XXH64_state_t* statePtr);
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
 
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 /*******   Canonical representation   *******/
 typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src);
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
+
+#ifndef XXH_NO_XXH3
 
 /*!
  * @}
  * ************************************************************************
- * @defgroup xxh3_family XXH3 family
+ * @defgroup XXH3_family XXH3 family
  * @ingroup public
  * @{
  *
@@ -751,16 +880,18 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
  *
  * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
  * but does not require it.
- * Any 32-bit and 64-bit targets that can run XXH32 smoothly
- * can run XXH3 at competitive speeds, even without vector support.
- * Further details are explained in the implementation.
+ * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
+ * at competitive speeds, even without vector support. Further details are
+ * explained in the implementation.
  *
  * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8,
- * ZVector and scalar targets. This can be controlled via the XXH_VECTOR macro.
+ * ZVector and scalar targets. This can be controlled via the @ref XXH_VECTOR
+ * macro. For the x86 family, an automatic dispatcher is included separately
+ * in @ref xxh_x86dispatch.c.
  *
  * XXH3 implementation is portable:
  * it has a generic C90 formulation that can be compiled on any platform,
- * 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.
  *
@@ -772,24 +903,42 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
  *
  * The API supports one-shot hashing, streaming mode, and custom secrets.
  */
-
 /*-**********************************************************************
 *  XXH3 64-bit variant
 ************************************************************************/
 
-/* XXH3_64bits():
- * default 64-bit variant, using default secret and default seed of 0.
- * It's the fastest variant. */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* data, size_t len);
+/*!
+ * @brief 64-bit unseeded variant of XXH3.
+ *
+ * This is equivalent to @ref XXH3_64bits_withSeed() with a seed of 0, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see
+ *    XXH32(), XXH64(), XXH3_128bits(): equivalent for the other xxHash algorithms
+ * @see
+ *    XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
+ * @see
+ *    XXH3_64bits_reset(), XXH3_64bits_update(), XXH3_64bits_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
 
-/*
- * XXH3_64bits_withSeed():
- * This variant generates a custom secret on the fly
- * based on default secret altered using the `seed` value.
+/*!
+ * @brief 64-bit seeded variant of XXH3
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the `seed` value.
+ *
  * While this operation is decently fast, note that it's not completely free.
- * Note: seed==0 produces the same results as XXH3_64bits().
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * @param input The data to hash
+ * @param length The length
+ * @param seed The 64-bit seed to alter the state.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
 
 /*!
  * The bare minimum size for a custom secret.
@@ -800,8 +949,9 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, X
  */
 #define XXH3_SECRET_SIZE_MIN 136
 
-/*
- * XXH3_64bits_withSecret():
+/*!
+ * @brief 64-bit variant of XXH3 with a custom "secret".
+ *
  * It's possible to provide any blob of bytes as a "secret" to generate the hash.
  * This makes it more difficult for an external actor to prepare an intentional collision.
  * The main condition is that secretSize *must* be large enough (>= XXH3_SECRET_SIZE_MIN).
@@ -817,10 +967,11 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, X
  * This is not necessarily the case when using the blob of bytes directly
  * because, when hashing _small_ inputs, only a portion of the secret is employed.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
 
 
 /*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*
  * Streaming requires state maintenance.
  * This operation costs memory and CPU.
@@ -834,23 +985,23 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len,
  * @see XXH3_state_s for details.
  */
 typedef struct XXH3_state_s XXH3_state_t;
-XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
 XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
-XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state);
+XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
 
 /*
  * XXH3_64bits_reset():
  * Initialize with default parameters.
  * digest will be equivalent to `XXH3_64bits()`.
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
 /*
  * XXH3_64bits_reset_withSeed():
  * Generate a custom secret from `seed`, and store it into `statePtr`.
  * digest will be equivalent to `XXH3_64bits_withSeed()`.
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-/*
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*!
  * XXH3_64bits_reset_withSecret():
  * `secret` is referenced, it _must outlive_ the hash streaming session.
  * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`,
@@ -859,10 +1010,11 @@ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr,
  * When in doubt about the randomness of a candidate `secret`,
  * consider employing `XXH3_generateSecret()` instead (see below).
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
 
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_digest (const XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t  XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
 /* note : canonical representation of XXH3 is the same as XXH64
  * since they both produce XXH64_hash_t values */
@@ -883,11 +1035,31 @@ typedef struct {
     XXH64_hash_t high64;  /*!< `value >> 64` */
 } XXH128_hash_t;
 
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+/*!
+ * @brief Unseeded 128-bit variant of XXH3
+ *
+ * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
+ * for shorter inputs.
+ *
+ * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of 0, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see
+ *    XXH32(), XXH64(), XXH3_64bits(): equivalent for the other xxHash algorithms
+ * @see
+ *    XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
+ * @see
+ *    XXH3_128bits_reset(), XXH3_128bits_update(), XXH3_128bits_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
+/*! @brief Seeded 128-bit variant of XXH3. @see XXH3_64bits_withSeed(). */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+/*! @brief Custom secret 128-bit variant of XXH3. @see XXH3_64bits_withSecret(). */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
 
 /*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*
  * Streaming requires state maintenance.
  * This operation costs memory and CPU.
@@ -900,12 +1072,13 @@ XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t le
  * All reset and streaming functions have same meaning as their 64-bit counterpart.
  */
 
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
 
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
 /* Following helper functions make it possible to compare XXH128_hast_t values.
  * Since XXH128_hash_t is a structure, this capability is not offered by the language.
@@ -915,26 +1088,26 @@ XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr);
  * XXH128_isEqual():
  * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
  */
-XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
 
 /*!
- * XXH128_cmp():
- *
+ * @brief Compares two @ref XXH128_hash_t
  * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
  *
- * return: >0 if *h128_1  > *h128_2
- *         =0 if *h128_1 == *h128_2
- *         <0 if *h128_1  < *h128_2
+ * @return: >0 if *h128_1  > *h128_2
+ *          =0 if *h128_1 == *h128_2
+ *          <0 if *h128_1  < *h128_2
  */
-XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2);
+XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
 
 
 /*******   Canonical representation   *******/
 typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
-XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash);
-XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src);
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
 
 
+#endif  /* !XXH_NO_XXH3 */
 #endif  /* XXH_NO_LONG_LONG */
 
 /*!
@@ -978,7 +1151,7 @@ struct XXH32_state_s {
    XXH32_hash_t v[4];         /*!< Accumulator lanes */
    XXH32_hash_t mem32[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
    XXH32_hash_t memsize;      /*!< Amount of data in @ref mem32 */
-   XXH32_hash_t reserved;     /*!< Reserved field. Do not read or write to it, it may be removed. */
+   XXH32_hash_t reserved;     /*!< Reserved field. Do not read nor write to it. */
 };   /* typedef'd to XXH32_state_t */
 
 
@@ -1002,9 +1175,11 @@ struct XXH64_state_s {
    XXH64_hash_t mem64[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
    XXH32_hash_t memsize;      /*!< Amount of data in @ref mem64 */
    XXH32_hash_t reserved32;   /*!< Reserved field, needed for padding anyways*/
-   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it, it may be removed. */
+   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it. */
 };   /* typedef'd to XXH64_state_t */
 
+#ifndef XXH_NO_XXH3
+
 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
 #  include <stdalign.h>
 #  define XXH_ALIGN(n)      alignas(n)
@@ -1070,7 +1245,7 @@ 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]);
@@ -1110,69 +1285,119 @@ 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
  */
-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():
  *
  * 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.
- *
- * 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.
+/*!
+ * @brief Generate the same secret as the _withSeed() variants.
  *
  * 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
+ * @param secretBuffer A writable buffer of @ref XXH3_SECRET_SIZE_MIN bytes
+ * @param seed The seed to seed the state.
  */
-XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_t seed);
+XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
 
-/*
- * *_withSecretandSeed() :
+/*!
  * 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 (< XXH3_MIDSIZE_MAX = 240 bytes)
+ * or @p secret for "large" keys (>= XXH3_MIDSIZE_MAX).
  *
  * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
  * `_withSeed()` has to generate the secret on the fly for "large" keys.
@@ -1181,7 +1406,7 @@ 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.
@@ -1190,32 +1415,34 @@ 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,
+/*! @copydoc XXH3_64bits_withSecretandSeed() */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
+                               XXH_NOESCAPE const void* secret, size_t secretSize,
                                XXH64_hash_t seed64);
-
+#ifndef XXH_NO_STREAM
+/*! @copydoc XXH3_64bits_withSecretandSeed() */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
-                                    const void* secret, size_t secretSize,
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                    XXH_NOESCAPE const void* secret, size_t secretSize,
                                     XXH64_hash_t seed64);
-
+/*! @copydoc XXH3_64bits_withSecretandSeed() */
 XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
-                                     const void* secret, size_t secretSize,
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                     XXH_NOESCAPE const void* secret, size_t secretSize,
                                      XXH64_hash_t seed64);
+#endif /* !XXH_NO_STREAM */
 
-
+#endif  /* !XXH_NO_XXH3 */
 #endif  /* XXH_NO_LONG_LONG */
 #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
 #  define XXH_IMPLEMENTATION
@@ -1269,7 +1496,7 @@ 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 */
@@ -1292,7 +1519,7 @@ 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,
@@ -1319,13 +1546,41 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  *   care, as what works on one compiler/platform/optimization level may cause
  *   another to read garbage data or even crash.
  *
- * See http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
+ * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
  *
  * Prefer these methods in priority order (0 > 3 > 1 > 2)
  */
 #  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).
@@ -1346,9 +1601,11 @@ 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
@@ -1370,12 +1627,29 @@ 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`.
  *
@@ -1396,34 +1670,45 @@ 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
@@ -1431,14 +1716,22 @@ 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
@@ -1453,6 +1746,24 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
 /* *************************************
 *  Includes & Memory related functions
 ***************************************/
+#if defined(XXH_NO_STREAM)
+/* nothing */
+#elif defined(XXH_NO_STDLIB)
+
+/* When requesting to disable any mention of stdlib,
+ * the library loses the ability to invoked malloc / free.
+ * In practice, it means that functions like `XXH*_createState()`
+ * will always fail, and return NULL.
+ * This flag is useful in situations where
+ * xxhash.h is integrated into some kernel, embedded or limited environment
+ * without access to dynamic allocation.
+ */
+
+static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
+static void XXH_free(void* p) { (void)p; }
+
+#else
+
 /*
  * Modify the local functions below should you wish to use
  * different memory routines for malloc() and free()
@@ -1463,7 +1774,7 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  * @internal
  * @brief Modify this function to use a different routine than malloc().
  */
-static void* XXH_malloc(size_t s) { return malloc(s); }
+static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
 
 /*!
  * @internal
@@ -1471,6 +1782,8 @@ static void* XXH_malloc(size_t s) { return malloc(s); }
  */
 static void XXH_free(void* p) { free(p); }
 
+#endif  /* XXH_NO_STDLIB */
+
 #include <string.h>
 
 /*!
@@ -1515,6 +1828,11 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size)
 #  define XXH_NO_INLINE static
 #endif
 
+#if XXH3_INLINE_SECRET
+#  define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
+#else
+#  define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
+#endif
 
 
 /* *************************************
@@ -1540,14 +1858,13 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size)
 #  include <assert.h>   /* note: can still be disabled with NDEBUG */
 #  define XXH_ASSERT(c)   assert(c)
 #else
-#  define XXH_ASSERT(c)   ((void)0)
+#  define XXH_ASSERT(c)   XXH_ASSUME(c)
 #endif
 
 /* note: use after variable declarations */
 #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
@@ -1573,11 +1890,17 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size)
  * XXH3_initCustomSecret_scalar().
  */
 #if defined(__GNUC__) || defined(__clang__)
-#  define XXH_COMPILER_GUARD(var) __asm__ __volatile__("" : "+r" (var))
+#  define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
 #else
 #  define XXH_COMPILER_GUARD(var) ((void)0)
 #endif
 
+#if defined(__clang__)
+#  define XXH_COMPILER_GUARD_W(var) __asm__("" : "+w" (var))
+#else
+#  define XXH_COMPILER_GUARD_W(var) ((void)0)
+#endif
+
 /* *************************************
 *  Basic Types
 ***************************************/
@@ -1665,25 +1988,26 @@ static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr;
 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
 
 /*
- * __pack instructions are safer but compiler specific, hence potentially
- * problematic for some compilers.
- *
- * Currently only defined for GCC and ICC.
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
  */
 #ifdef XXH_OLD_NAMES
 typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
 #endif
 static xxh_u32 XXH_read32(const void* ptr)
 {
-    typedef union { xxh_u32 u32; } __attribute__((packed)) xxh_unalign;
-    return ((const xxh_unalign*)ptr)->u32;
+    typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32;
+    return *((const xxh_unalign32*)ptr);
 }
 
 #else
 
 /*
  * Portable and safe solution. Generally efficient.
- * see: http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
  */
 static xxh_u32 XXH_read32(const void* memPtr)
 {
@@ -1759,6 +2083,51 @@ static int XXH_isLittleEndian(void)
 #  define XXH_HAS_BUILTIN(x) 0
 #endif
 
+
+
+/*
+ * C23 and future versions have standard "unreachable()".
+ * Once it has been implemented reliably we can add it as an
+ * additional case:
+ *
+ * ```
+ * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN)
+ * #  include <stddef.h>
+ * #  ifdef unreachable
+ * #    define XXH_UNREACHABLE() unreachable()
+ * #  endif
+ * #endif
+ * ```
+ *
+ * Note C++23 also has std::unreachable() which can be detected
+ * as follows:
+ * ```
+ * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L)
+ * #  include <utility>
+ * #  define XXH_UNREACHABLE() std::unreachable()
+ * #endif
+ * ```
+ * NB: `__cpp_lib_unreachable` is defined in the `<version>` header.
+ * We don't use that as including `<utility>` in `extern "C"` blocks
+ * doesn't work on GCC12
+ */
+
+#if XXH_HAS_BUILTIN(__builtin_unreachable)
+#  define XXH_UNREACHABLE() __builtin_unreachable()
+
+#elif defined(_MSC_VER)
+#  define XXH_UNREACHABLE() __assume(0)
+
+#else
+#  define XXH_UNREACHABLE()
+#endif
+
+#if XXH_HAS_BUILTIN(__builtin_assume)
+#  define XXH_ASSUME(c) __builtin_assume(c)
+#else
+#  define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
+#endif
+
 /*!
  * @internal
  * @def XXH_rotl32(x,r)
@@ -1881,8 +2250,10 @@ 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 */
@@ -1946,9 +2317,9 @@ static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
      *   can load data, while v3 can multiply. SSE forces them to operate
      *   together.
      *
-     * This is also enabled on AArch64, as Clang autovectorizes it incorrectly
-     * and it is pointless writing a NEON implementation that is basically the
-     * same speed as scalar for XXH32.
+     * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
+     * the loop. NEON is only faster on the A53, and with the newer cores, it is less
+     * than half the speed.
      */
     XXH_COMPILER_GUARD(acc);
 #endif
@@ -1962,17 +2333,17 @@ 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)
@@ -1985,24 +2356,25 @@ 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);
@@ -2018,49 +2390,49 @@ 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 */
     }
 }
 
@@ -2080,7 +2452,7 @@ 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;
@@ -2113,10 +2485,10 @@ 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);
@@ -2135,42 +2507,39 @@ 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)
 {
-    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state));
-    state.v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
-    state.v[1] = seed + XXH_PRIME32_2;
-    state.v[2] = seed + 0;
-    state.v[3] = seed - XXH_PRIME32_1;
-    /* do not write into reserved, planned to be removed in a future version */
-    XXH_memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved));
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+    statePtr->v[1] = seed + XXH_PRIME32_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME32_1;
     return XXH_OK;
 }
 
 
-/*! @ingroup xxh32_family */
+/*! @ingroup XXH32_family */
 XXH_PUBLIC_API XXH_errorcode
 XXH32_update(XXH32_state_t* state, const void* input, size_t len)
 {
@@ -2225,7 +2594,7 @@ 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;
@@ -2243,12 +2612,12 @@ 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
+ * @ingroup XXH32_family
  * The default return values from XXH functions are unsigned 32 and 64 bit
  * integers.
  *
@@ -2267,7 +2636,7 @@ XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, 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);
@@ -2308,25 +2677,26 @@ static xxh_u64 XXH_read64(const void* memPtr)
 #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
 
 /*
- * __pack instructions are safer, but compiler specific, hence potentially
- * problematic for some compilers.
- *
- * Currently only defined for GCC and ICC.
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
  */
 #ifdef XXH_OLD_NAMES
 typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
 #endif
 static xxh_u64 XXH_read64(const void* ptr)
 {
-    typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) xxh_unalign64;
-    return ((const xxh_unalign64*)ptr)->u64;
+    typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64;
+    return *((const xxh_unalign64*)ptr);
 }
 
 #else
 
 /*
  * Portable and safe solution. Generally efficient.
- * see: http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
  */
 static xxh_u64 XXH_read64(const void* memPtr)
 {
@@ -2410,8 +2780,10 @@ 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 */
@@ -2429,6 +2801,7 @@ XXH_readLE64_align(const void* ptr, XXH_alignment align)
 #  define PRIME64_5 XXH_PRIME64_5
 #endif
 
+/*! @copydoc XXH32_round */
 static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
 {
     acc += input * XXH_PRIME64_2;
@@ -2445,43 +2818,59 @@ 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
@@ -2494,7 +2883,15 @@ 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;
@@ -2531,10 +2928,10 @@ 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);
@@ -2552,42 +2949,40 @@ 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)
 {
-    XXH64_state_t state;   /* use a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state));
-    state.v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
-    state.v[1] = seed + XXH_PRIME64_2;
-    state.v[2] = seed + 0;
-    state.v[3] = seed - XXH_PRIME64_1;
-     /* do not write into reserved64, might be removed in a future version */
-    XXH_memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved64));
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+    statePtr->v[1] = seed + XXH_PRIME64_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME64_1;
     return XXH_OK;
 }
 
-/*! @ingroup xxh64_family */
+/*! @ingroup XXH64_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH64_update (XXH64_state_t* state, const void* input, size_t len)
+XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
     if (input==NULL) {
         XXH_ASSERT(len == 0);
@@ -2637,8 +3032,8 @@ 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;
 
@@ -2656,20 +3051,20 @@ 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));
     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);
 }
@@ -2682,7 +3077,7 @@ 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
  * @{
  */
@@ -2690,11 +3085,19 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
 /* ===   Compiler specifics   === */
 
 #if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
-#  define XXH_RESTRICT /* disable */
+#  define XXH_RESTRICT   /* disable */
 #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
 #  define XXH_RESTRICT   restrict
+#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \
+   || (defined (__clang__)) \
+   || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \
+   || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300))
+/*
+ * There are a LOT more compilers that recognize __restrict but this
+ * covers the major ones.
+ */
+#  define XXH_RESTRICT   __restrict
 #else
-/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */
 #  define XXH_RESTRICT   /* disable */
 #endif
 
@@ -2708,17 +3111,24 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src
 #    define XXH_unlikely(x) (x)
 #endif
 
-#if defined(__GNUC__)
-#  if defined(__AVX2__)
-#    include <immintrin.h>
-#  elif defined(__SSE2__)
-#    include <emmintrin.h>
-#  elif defined(__ARM_NEON__) || defined(__ARM_NEON)
+#if defined(__GNUC__) || defined(__clang__)
+#  if defined(__ARM_FEATURE_SVE)
+#    include <arm_sve.h>
+#  endif
+#  if defined(__ARM_NEON__) || defined(__ARM_NEON) \
+   || (defined(_M_ARM) && _M_ARM >= 7) \
+   || defined(_M_ARM64) || defined(_M_ARM64EC)
 #    define inline __inline__  /* circumvent a clang bug */
 #    include <arm_neon.h>
 #    undef inline
+#  elif defined(__AVX2__)
+#    include <immintrin.h>
+#  elif defined(__SSE2__)
+#    include <emmintrin.h>
 #  endif
-#elif defined(_MSC_VER)
+#endif
+
+#if defined(_MSC_VER)
 #  include <intrin.h>
 #endif
 
@@ -2832,12 +3242,13 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
     XXH_AVX512 = 3,  /*!< AVX512 for Skylake and Icelake */
     XXH_NEON   = 4,  /*!< NEON for most ARMv7-A and all AArch64 */
     XXH_VSX    = 5,  /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+    XXH_SVE    = 6,  /*!< SVE for some ARMv8-A and ARMv9-A */
 };
 /*!
  * @ingroup tuning
  * @brief Selects the minimum alignment for XXH3's accumulators.
  *
- * When using SIMD, this should match the alignment reqired for said vector
+ * When using SIMD, this should match the alignment required for said vector
  * type, so, for example, 32 for AVX2.
  *
  * Default: Auto detected.
@@ -2853,23 +3264,26 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
 #  define XXH_AVX512 3
 #  define XXH_NEON   4
 #  define XXH_VSX    5
+#  define XXH_SVE    6
 #endif
 
 #ifndef XXH_VECTOR    /* can be defined on command line */
-#  if defined(__AVX512F__)
-#    define XXH_VECTOR XXH_AVX512
-#  elif defined(__AVX2__)
-#    define XXH_VECTOR XXH_AVX2
-#  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
-#    define XXH_VECTOR XXH_SSE2
+#  if defined(__ARM_FEATURE_SVE)
+#    define XXH_VECTOR XXH_SVE
 #  elif ( \
         defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
-     || defined(_M_ARM64) || defined(_M_ARM_ARMV7VE) /* msvc */ \
+     || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
    ) && ( \
         defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
    )
 #    define XXH_VECTOR XXH_NEON
+#  elif defined(__AVX512F__)
+#    define XXH_VECTOR XXH_AVX512
+#  elif defined(__AVX2__)
+#    define XXH_VECTOR XXH_AVX2
+#  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
+#    define XXH_VECTOR XXH_SSE2
 #  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
      || (defined(__s390x__) && defined(__VEC__)) \
      && defined(__GNUC__) /* TODO: IBM XL */
@@ -2879,6 +3293,17 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
 #  endif
 #endif
 
+/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */
+#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE)
+#  ifdef _MSC_VER
+#    pragma warning(once : 4606)
+#  else
+#    warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead."
+#  endif
+#  undef XXH_VECTOR
+#  define XXH_VECTOR XXH_SCALAR
+#endif
+
 /*
  * Controls the alignment of the accumulator,
  * for compatibility with aligned vector loads, which are usually faster.
@@ -2898,16 +3323,26 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
 #     define XXH_ACC_ALIGN 16
 #  elif XXH_VECTOR == XXH_AVX512  /* avx512 */
 #     define XXH_ACC_ALIGN 64
+#  elif XXH_VECTOR == XXH_SVE   /* sve */
+#     define XXH_ACC_ALIGN 64
 #  endif
 #endif
 
 #if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
     || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
 #  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#elif XXH_VECTOR == XXH_SVE
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
 #else
 #  define XXH_SEC_ALIGN 8
 #endif
 
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_ALIASING __attribute__((may_alias))
+#else
+#  define XXH_ALIASING /* nothing */
+#endif
+
 /*
  * UGLY HACK:
  * GCC usually generates the best code with -O3 for xxHash.
@@ -2931,111 +3366,127 @@ 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.
- *
- * 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.
+ * 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.
  *
- * On ARMv7-A, this strangely modifies both parameters in place instead of
- * taking the usual 3-operand form.
+ * 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.
  *
- * 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.
+ * 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).
  *
- *   vzip.32   d10, d11       @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] }
+ * 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.
  *
- * 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.
+ * 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
+
+/*!
+ * @internal
+ * @brief `vmlal_u32` on low and high halves of a vector.
  *
- * aarch64 requires a different approach.
+ * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with
+ * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32`
+ * with `vmlal_u32`.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* Inline assembly is the only way */
+    __asm__("umlal   %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs));
+    return acc;
+}
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* This intrinsic works as expected */
+    return vmlal_high_u32(acc, lhs, rhs);
+}
+#else
+/* Portable intrinsic versions */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
+}
+/*! @copydoc XXH_vmlal_low_u32
+ * Assume the compiler converts this to vmlal_high_u32 on aarch64 */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
+}
+#endif
+
+/*!
+ * @ingroup tuning
+ * @brief Controls the NEON to scalar ratio for XXH3
  *
- * In order to make it easier to write a decent compiler for aarch64, many
- * quirks were removed, such as conditional execution.
+ * This can be set to 2, 4, 6, or 8.
  *
- * NEON was also affected by this.
+ * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
  *
- * 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.
+ * 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.
  *
- * The formerly free vget_high intrinsics now require a vext (with a few
- * exceptions)
+ * 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.
  *
- * 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.
+ * 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.
  *
- * The equivalent of the VZIP.32 on the lower and upper halves would be this
- * mess:
+ * 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.
  *
- *   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] }
+ * This change benefits CPUs with large micro-op buffers without negatively affecting
+ * most other CPUs:
  *
- * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN):
+ *  | 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% |
  *
- *   shrn    v1.2s, v0.2d, #32  // v1 = (uint32x2_t)(v0 >> 32);
- *   xtn     v0.2s, v0.2d       // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF);
+ * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
  *
- * This is available on ARMv7-A, but is less efficient than a single VZIP.32.
+ * @see XXH3_accumulate_512_neon()
  */
-
-/*!
- * Function-like macro:
- * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t &outHi)
- * {
- *     outLo = (uint32x2_t)(in & 0xFFFFFFFF);
- *     outHi = (uint32x2_t)(in >> 32);
- *     in = UNDEFINED;
- * }
- */
-# if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \
-   && defined(__GNUC__) \
-   && !defined(__aarch64__) && !defined(__arm64__) && !defined(_M_ARM64)
-#  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)
+# ifndef XXH3_NEON_LANES
+#  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
+   && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
+#   define XXH3_NEON_LANES 6
+#  else
+#   define XXH3_NEON_LANES XXH_ACC_NB
+#  endif
 # endif
 #endif  /* XXH_VECTOR == XXH_NEON */
 
@@ -3048,27 +3499,42 @@ 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__)
@@ -3120,8 +3586,9 @@ 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
@@ -3142,13 +3609,29 @@ XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
 # endif /* XXH_vec_mulo, XXH_vec_mule */
 #endif /* XXH_VECTOR == XXH_VSX */
 
+#if XXH_VECTOR == XXH_SVE
+#define ACCRND(acc, offset) \
+do { \
+    svuint64_t input_vec = svld1_u64(mask, xinput + offset);         \
+    svuint64_t secret_vec = svld1_u64(mask, xsecret + offset);       \
+    svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec);     \
+    svuint64_t swapped = svtbl_u64(input_vec, kSwap);                \
+    svuint64_t mixed_lo = svextw_u64_x(mask, mixed);                 \
+    svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32);            \
+    svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
+    acc = svadd_u64_x(mask, acc, mul);                               \
+} while (0)
+#endif /* XXH_VECTOR == XXH_SVE */
+
 
 /* prefetch
  * can be disabled, by declaring XXH_NO_PREFETCH build macro */
 #if defined(XXH_NO_PREFETCH)
 #  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
 #else
-#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
+#  if XXH_SIZE_OPT >= 1
+#    define XXH_PREFETCH(ptr) (void)(ptr)
+#  elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
 #    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
 #    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
 #  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
@@ -3213,7 +3696,6 @@ XXH_mult32to64(xxh_u64 x, xxh_u64 y)
    return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
 }
 #elif defined(_MSC_VER) && defined(_M_IX86)
-#    include <intrin.h>
 #    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
 #else
 /*
@@ -3253,7 +3735,7 @@ XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
      * In that case it is best to use the portable one.
      * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
      */
-#if defined(__GNUC__) && !defined(__wasm__) \
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
     && defined(__SIZEOF_INT128__) \
     || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
 
@@ -3270,7 +3752,7 @@ XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
      *
      * This compiles to single operand MUL on x64.
      */
-#elif defined(_M_X64) || defined(_M_IA64)
+#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)
 
 #ifndef _MSC_VER
 #   pragma intrinsic(_umul128)
@@ -3287,7 +3769,7 @@ XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
      *
      * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
      */
-#elif defined(_M_ARM64)
+#elif defined(_M_ARM64) || defined(_M_ARM64EC)
 
 #ifndef _MSC_VER
 #   pragma intrinsic(__umulh)
@@ -3377,7 +3859,7 @@ 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);
@@ -3444,7 +3926,7 @@ 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);
@@ -3466,7 +3948,7 @@ 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);
@@ -3482,7 +3964,7 @@ 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);
@@ -3499,7 +3981,7 @@ 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);
@@ -3569,7 +4051,7 @@ 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)
@@ -3577,29 +4059,39 @@ XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
     XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
     XXH_ASSERT(16 < len && len <= 128);
 
-    {   xxh_u64 acc = len * XXH_PRIME64_1;
+    {   xxh_u64 acc = len * XXH_PRIME64_1, acc_end;
+#if XXH_SIZE_OPT >= 1
+        /* Smaller and cleaner, but slightly slower. */
+        unsigned int i = (unsigned int)(len - 1) / 32;
+        do {
+            acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
+            acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
+        } while (i-- != 0);
+        acc_end = 0;
+#else
+        acc += XXH3_mix16B(input+0, secret+0, seed);
+        acc_end = XXH3_mix16B(input+len-16, secret+16, seed);
         if (len > 32) {
+            acc += XXH3_mix16B(input+16, secret+32, seed);
+            acc_end += XXH3_mix16B(input+len-32, secret+48, seed);
             if (len > 64) {
+                acc += XXH3_mix16B(input+32, secret+64, seed);
+                acc_end += XXH3_mix16B(input+len-48, secret+80, seed);
+
                 if (len > 96) {
                     acc += XXH3_mix16B(input+48, secret+96, seed);
-                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
+                    acc_end += XXH3_mix16B(input+len-64, secret+112, seed);
                 }
-                acc += XXH3_mix16B(input+32, secret+64, seed);
-                acc += XXH3_mix16B(input+len-48, secret+80, seed);
             }
-            acc += XXH3_mix16B(input+16, secret+32, seed);
-            acc += XXH3_mix16B(input+len-32, secret+48, seed);
         }
-        acc += XXH3_mix16B(input+0, secret+0, seed);
-        acc += XXH3_mix16B(input+len-16, secret+16, seed);
-
-        return XXH3_avalanche(acc);
+#endif
+        return XXH3_avalanche(acc + acc_end);
     }
 }
 
 #define XXH3_MIDSIZE_MAX 240
 
-XXH_NO_INLINE XXH64_hash_t
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
 XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                       XXH64_hash_t seed)
@@ -3611,13 +4103,17 @@ XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
     #define XXH3_MIDSIZE_LASTOFFSET  17
 
     {   xxh_u64 acc = len * XXH_PRIME64_1;
-        int const nbRounds = (int)len / 16;
-        int i;
+        xxh_u64 acc_end;
+        unsigned int const nbRounds = (unsigned int)len / 16;
+        unsigned int i;
+        XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
         for (i=0; i<8; i++) {
             acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
         }
-        acc = XXH3_avalanche(acc);
+        /* last bytes */
+        acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
         XXH_ASSERT(nbRounds >= 8);
+        acc = XXH3_avalanche(acc);
 #if defined(__clang__)                                /* Clang */ \
     && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
     && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
@@ -3644,11 +4140,13 @@ XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
         #pragma clang loop vectorize(disable)
 #endif
         for (i=8 ; i < nbRounds; i++) {
-            acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+            /*
+             * Prevents clang for unrolling the acc loop and interleaving with this one.
+             */
+            XXH_COMPILER_GUARD(acc);
+            acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
         }
-        /* last bytes */
-        acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
-        return XXH3_avalanche(acc);
+        return XXH3_avalanche(acc + acc_end);
     }
 }
 
@@ -3664,6 +4162,47 @@ XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
 #  define ACC_NB XXH_ACC_NB
 #endif
 
+#ifndef XXH_PREFETCH_DIST
+#  ifdef __clang__
+#    define XXH_PREFETCH_DIST 320
+#  else
+#    if (XXH_VECTOR == XXH_AVX512)
+#      define XXH_PREFETCH_DIST 512
+#    else
+#      define XXH_PREFETCH_DIST 384
+#    endif
+#  endif  /* __clang__ */
+#endif  /* XXH_PREFETCH_DIST */
+
+/*
+ * These macros are to generate an XXH3_accumulate() function.
+ * The two arguments select the name suffix and target attribute.
+ *
+ * The name of this symbol is XXH3_accumulate_<name>() and it calls
+ * XXH3_accumulate_512_<name>().
+ *
+ * It may be useful to hand implement this function if the compiler fails to
+ * optimize the inline function.
+ */
+#define XXH3_ACCUMULATE_TEMPLATE(name)                      \
+void                                                        \
+XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc,           \
+                       const xxh_u8* XXH_RESTRICT input,    \
+                       const xxh_u8* XXH_RESTRICT secret,   \
+                       size_t nbStripes)                    \
+{                                                           \
+    size_t n;                                               \
+    for (n = 0; n < nbStripes; n++ ) {                      \
+        const xxh_u8* const in = input + n*XXH_STRIPE_LEN;  \
+        XXH_PREFETCH(in + XXH_PREFETCH_DIST);               \
+        XXH3_accumulate_512_##name(                         \
+                 acc,                                       \
+                 in,                                        \
+                 secret + n*XXH_SECRET_CONSUME_RATE);       \
+    }                                                       \
+}
+
+
 XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
 {
     if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
@@ -3684,6 +4223,7 @@ XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
     typedef long long xxh_i64;
 #endif
 
+
 /*
  * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
  *
@@ -3731,7 +4271,7 @@ XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
         /* data_key    = data_vec ^ key_vec; */
         __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
         /* data_key_lo = data_key >> 32; */
-        __m512i const data_key_lo = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32);
         /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
         __m512i const product     = _mm512_mul_epu32     (data_key, data_key_lo);
         /* xacc[0] += swap(data_vec); */
@@ -3741,6 +4281,7 @@ XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
         *xacc = _mm512_add_epi64(product, sum);
     }
 }
+XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)
 
 /*
  * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
@@ -3774,13 +4315,12 @@ XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
         /* xacc[0] ^= (xacc[0] >> 47) */
         __m512i const acc_vec     = *xacc;
         __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
-        __m512i const data_vec    = _mm512_xor_si512     (acc_vec, shifted);
         /* xacc[0] ^= secret; */
         __m512i const key_vec     = _mm512_loadu_si512   (secret);
-        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+        __m512i const data_key    = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */);
 
         /* xacc[0] *= XXH_PRIME32_1; */
-        __m512i const data_key_hi = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32);
         __m512i const prod_lo     = _mm512_mul_epu32     (data_key, prime32);
         __m512i const prod_hi     = _mm512_mul_epu32     (data_key_hi, prime32);
         *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
@@ -3795,7 +4335,8 @@ XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
     XXH_ASSERT(((size_t)customSecret & 63) == 0);
     (void)(&XXH_writeLE64);
     {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
-        __m512i const seed = _mm512_mask_set1_epi64(_mm512_set1_epi64((xxh_i64)seed64), 0xAA, (xxh_i64)(0U - seed64));
+        __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64);
+        __m512i const seed     = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos);
 
         const __m512i* const src  = (const __m512i*) ((const void*) XXH3_kSecret);
               __m512i* const dest = (      __m512i*) customSecret;
@@ -3803,14 +4344,7 @@ XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
         XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
         XXH_ASSERT(((size_t)dest & 63) == 0);
         for (i=0; i < nbRounds; ++i) {
-            /* GCC has a bug, _mm512_stream_load_si512 accepts 'void*', not 'void const*',
-             * this will warn "discards 'const' qualifier". */
-            union {
-                const __m512i* cp;
-                void* p;
-            } remote_const_void;
-            remote_const_void.cp = src + i;
-            dest[i] = _mm512_add_epi64(_mm512_stream_load_si512(remote_const_void.p), seed);
+            dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
     }   }
 }
 
@@ -3846,7 +4380,7 @@ XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
             /* data_key    = data_vec ^ key_vec; */
             __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
             /* data_key_lo = data_key >> 32; */
-            __m256i const data_key_lo = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32);
             /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
             __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
             /* xacc[i] += swap(data_vec); */
@@ -3856,6 +4390,7 @@ XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
             xacc[i] = _mm256_add_epi64(product, sum);
     }   }
 }
+XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)
 
 XXH_FORCE_INLINE XXH_TARGET_AVX2 void
 XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
@@ -3878,7 +4413,7 @@ XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
             __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
 
             /* xacc[i] *= XXH_PRIME32_1; */
-            __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32);
             __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
             __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
             xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
@@ -3910,12 +4445,12 @@ XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTR
         XXH_ASSERT(((size_t)dest & 31) == 0);
 
         /* GCC -O2 need unroll loop manually */
-        dest[0] = _mm256_add_epi64(_mm256_stream_load_si256(src+0), seed);
-        dest[1] = _mm256_add_epi64(_mm256_stream_load_si256(src+1), seed);
-        dest[2] = _mm256_add_epi64(_mm256_stream_load_si256(src+2), seed);
-        dest[3] = _mm256_add_epi64(_mm256_stream_load_si256(src+3), seed);
-        dest[4] = _mm256_add_epi64(_mm256_stream_load_si256(src+4), seed);
-        dest[5] = _mm256_add_epi64(_mm256_stream_load_si256(src+5), seed);
+        dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed);
+        dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed);
+        dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed);
+        dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed);
+        dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed);
+        dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed);
     }
 }
 
@@ -3962,6 +4497,7 @@ XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
             xacc[i] = _mm_add_epi64(product, sum);
     }   }
 }
+XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)
 
 XXH_FORCE_INLINE XXH_TARGET_SSE2 void
 XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
@@ -4029,96 +4565,196 @@ XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTR
 
 #if (XXH_VECTOR == XXH_NEON)
 
+/* forward declarations for the scalar routines */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret, size_t lane);
+
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret, size_t lane);
+
+/*!
+ * @internal
+ * @brief The bulk processing loop for NEON.
+ *
+ * The NEON code path is actually partially scalar when running on AArch64. This
+ * is to optimize the pipelining and can have up to 15% speedup depending on the
+ * CPU, and it also mitigates some GCC codegen issues.
+ *
+ * @see XXH3_NEON_LANES for configuring this and details about this optimization.
+ *
+ * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
+ * integers instead of the other platforms which mask full 64-bit vectors,
+ * so the setup is more complicated than just shifting right.
+ *
+ * Additionally, there is an optimization for 4 lanes at once noted below.
+ *
+ * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
+ * there needs to be *three* versions of the accumulate operation used
+ * for the remaining 2 lanes.
+ */
 XXH_FORCE_INLINE void
 XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT input,
                     const void* XXH_RESTRICT secret)
 {
     XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {
-        uint64x2_t* const xacc = (uint64x2_t *) acc;
+    XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
+    {   /* GCC for darwin arm64 does not like aliasing here */
+        xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
         /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
         uint8_t const* const xinput = (const uint8_t *) input;
         uint8_t const* const xsecret  = (const uint8_t *) secret;
 
         size_t i;
-        for (i=0; i < XXH_STRIPE_LEN / sizeof(uint64x2_t); i++) {
+        /* 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.
+             *
+             * This compiles to two instructions on AArch64 and has a paired vector
+             * result, which is an artifact from ARMv7a's version which modified both
+             * vectors in place.
+             *
+             *  [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
+             *  [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
+             */
+            uint32x4x2_t unzipped = vuzpq_u32(
+                vreinterpretq_u32_u64(data_key_1),
+                vreinterpretq_u32_u64(data_key_2)
+            );
+            /* data_key_lo = data_key & 0xFFFFFFFF */
+            uint32x4_t data_key_lo = unzipped.val[0];
+            /* data_key_hi = data_key >> 32 */
+            uint32x4_t data_key_hi = unzipped.val[1];
+            /*
+             * Then, we can split the vectors horizontally and multiply which, as for most
+             * widening intrinsics, have a variant that works on both high half vectors
+             * for free on AArch64.
+             *
+             * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
+             */
+            uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
+            uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
+            /*
+             * Clang reorders
+             *    a += b * c;     // umlal   swap.2d, dkl.2s, dkh.2s
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             * to
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             *    c += b * c;     // umlal   acc.2d, dkl.2s, dkh.2s
+             *
+             * While it would make sense in theory since the addition is faster,
+             * for reasons likely related to umlal being limited to certain NEON
+             * pipelines, this is worse. A compiler guard fixes this.
+             */
+            XXH_COMPILER_GUARD_W(sum_1);
+            XXH_COMPILER_GUARD_W(sum_2);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i]   = vaddq_u64(xacc[i], sum_1);
+            xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
+        }
+        /* Operate on the remaining NEON lanes 2 at a time. */
+        for (; i < XXH3_NEON_LANES / 2; i++) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec = XXH_vld1q_u64(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            /* 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_W(sum);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i] = vaddq_u64 (xacc[i], sum);
         }
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
 
 XXH_FORCE_INLINE void
 XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
     XXH_ASSERT((((size_t)acc) & 15) == 0);
 
-    {   uint64x2_t* xacc       = (uint64x2_t*) acc;
+    {   xxh_aliasing_uint64x2_t* xacc       = (xxh_aliasing_uint64x2_t*) acc;
         uint8_t const* xsecret = (uint8_t const*) secret;
         uint32x2_t prime       = vdup_n_u32 (XXH_PRIME32_1);
 
         size_t i;
-        for (i=0; i < XXH_STRIPE_LEN/sizeof(uint64x2_t); i++) {
+        /* AArch64 uses both scalar and neon at the same time */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarScrambleRound(acc, secret, i);
+        }
+        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
             /* xacc[i] ^= (xacc[i] >> 47); */
             uint64x2_t acc_vec  = xacc[i];
-            uint64x2_t shifted  = vshrq_n_u64 (acc_vec, 47);
-            uint64x2_t data_vec = veorq_u64   (acc_vec, shifted);
+            uint64x2_t shifted  = vshrq_n_u64(acc_vec, 47);
+            uint64x2_t data_vec = veorq_u64(acc_vec, shifted);
 
             /* xacc[i] ^= xsecret[i]; */
-            uint8x16_t key_vec  = vld1q_u8    (xsecret + (i * 16));
-            uint64x2_t data_key = veorq_u64   (data_vec, vreinterpretq_u64_u8(key_vec));
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
 
             /* xacc[i] *= XXH_PRIME32_1 */
-            uint32x2_t data_key_lo, data_key_hi;
-            /* 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);
-            }
-    }   }
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+            /*
+             * prod_hi = (data_key >> 32) * XXH_PRIME32_1;
+             *
+             * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will
+             * incorrectly "optimize" this:
+             *   tmp     = vmul_u32(vmovn_u64(a), vmovn_u64(b));
+             *   shifted = vshll_n_u32(tmp, 32);
+             * to this:
+             *   tmp     = "vmulq_u64"(a, b); // no such thing!
+             *   shifted = vshlq_n_u64(tmp, 32);
+             *
+             * However, unlike SSE, Clang lacks a 64-bit multiply routine
+             * for NEON, and it scalarizes two 64-bit multiplies instead.
+             *
+             * vmull_u32 has the same timing as vmul_u32, and it avoids
+             * this bug completely.
+             * See https://bugs.llvm.org/show_bug.cgi?id=39967
+             */
+            uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime);
+            /* xacc[i] = prod_hi << 32; */
+            prod_hi = vshlq_n_u64(prod_hi, 32);
+            /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */
+            xacc[i] = vmlal_u32(prod_hi, data_key_lo, prime);
+        }
+    }
 }
-
 #endif
 
 #if (XXH_VECTOR == XXH_VSX)
@@ -4129,23 +4765,23 @@ XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT secret)
 {
     /* presumed aligned */
-    unsigned long long* const xacc = (unsigned long long*) 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        = vec_xl(0, xacc + 2 * i);
+        xxh_u64x2 acc_vec        = xacc[i];
         acc_vec += product;
 
         /* swap high and low halves */
@@ -4154,18 +4790,18 @@ XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
 #else
         acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
 #endif
-        /* xacc[i] = acc_vec; */
-        vec_xst(acc_vec, 0, xacc + 2 * 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 };
@@ -4177,7 +4813,7 @@ XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
             xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
 
             /* xacc[i] ^= xsecret[i]; */
-            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + i);
+            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
             xxh_u64x2 const data_key = data_vec ^ key_vec;
 
             /* xacc[i] *= XXH_PRIME32_1 */
@@ -4191,40 +4827,233 @@ 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().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc,
+                 void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret,
+                 size_t lane)
+{
+    xxh_u64* xacc = (xxh_u64*) acc;
+    xxh_u8 const* xinput  = (xxh_u8 const*) input;
+    xxh_u8 const* xsecret = (xxh_u8 const*) secret;
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+    {
+        xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8);
+        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8);
+        xacc[lane ^ 1] += data_val; /* swap adjacent lanes */
+        xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]);
+    }
+}
+
+/*!
+ * @internal
+ * @brief Processes a 64 byte block of data using the scalar path.
+ */
 XXH_FORCE_INLINE void
 XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
                      const void* XXH_RESTRICT input,
                      const void* XXH_RESTRICT secret)
 {
-    xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */
-    const xxh_u8* const xinput  = (const xxh_u8*) input;  /* no alignment restriction */
-    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
     size_t i;
-    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+    /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
+#if defined(__GNUC__) && !defined(__clang__) \
+  && (defined(__arm__) || defined(__thumb2__)) \
+  && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
+  && XXH_SIZE_OPT <= 0
+#  pragma GCC unroll 8
+#endif
     for (i=0; i < XXH_ACC_NB; i++) {
-        xxh_u64 const data_val = XXH_readLE64(xinput + 8*i);
-        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i*8);
-        xacc[i ^ 1] += data_val; /* swap adjacent lanes */
-        xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32);
+        XXH3_scalarRound(acc, input, secret, i);
     }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
 
+/*!
+ * @internal
+ * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
 XXH_FORCE_INLINE void
-XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret,
+                         size_t lane)
 {
     xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
     const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
-    size_t i;
     XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
-    for (i=0; i < XXH_ACC_NB; i++) {
-        xxh_u64 const key64 = XXH_readLE64(xsecret + 8*i);
-        xxh_u64 acc64 = xacc[i];
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    {
+        xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8);
+        xxh_u64 acc64 = xacc[lane];
         acc64 = XXH_xorshift64(acc64, 47);
         acc64 ^= key64;
         acc64 *= XXH_PRIME32_1;
-        xacc[i] = acc64;
+        xacc[lane] = acc64;
+    }
+}
+
+/*!
+ * @internal
+ * @brief Scrambles the accumulators after a large chunk has been read
+ */
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    size_t i;
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarScrambleRound(acc, secret, i);
     }
 }
 
@@ -4239,15 +5068,16 @@ 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
-     * constant compared to 4 cycles for LDR), long MOVK chains stall the
-     * integer pipelines:
+     * constant compared to 4 cycles for LDR), it fights for bandwidth with
+     * the arithmetic instructions.
+     *
      *   I   L   S
      * MOVK
      * MOVK
@@ -4256,7 +5086,7 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
      * ADD
      * SUB      STR
      *          STR
-     * By forcing loads from memory (as the asm line causes Clang to assume
+     * By forcing loads from memory (as the asm line causes the compiler to assume
      * that XXH3_kSecretPtr has been changed), the pipelines are used more
      * efficiently:
      *   I   L   S
@@ -4264,23 +5094,20 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
      *  ADD LDR
      *  SUB     STR
      *          STR
+     *
+     * See XXH3_NEON_LANES for details on the pipsline.
+     *
      * XXH3_64bits_withSeed, len == 256, Snapdragon 835
      *   without hack: 2654.4 MB/s
      *   with hack:    3202.9 MB/s
      */
     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.
@@ -4293,7 +5120,7 @@ XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
 }
 
 
-typedef void (*XXH3_f_accumulate_512)(void* XXH_RESTRICT, const void*, const void*);
+typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t);
 typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
 typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
 
@@ -4301,82 +5128,63 @@ 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;
@@ -4388,7 +5196,7 @@ XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
     XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
 
     for (n = 0; n < nb_blocks; n++) {
-        XXH3_accumulate(acc, input + n*block_len, secret, nbStripesPerBlock, f_acc512);
+        f_acc(acc, input + n*block_len, secret, nbStripesPerBlock);
         f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
     }
 
@@ -4396,12 +5204,12 @@ XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
     XXH_ASSERT(len > XXH_STRIPE_LEN);
     {   size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
         XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
-        XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, f_acc512);
+        f_acc(acc, input + nb_blocks*block_len, secret, nbStripes);
 
         /* last stripe */
         {   const xxh_u8* const p = input + len - XXH_STRIPE_LEN;
 #define XXH_SECRET_LASTACC_START 7  /* not aligned on 8, last secret is different from acc & scrambler */
-            f_acc512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+            XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
     }   }
 }
 
@@ -4446,12 +5254,12 @@ XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secre
 XXH_FORCE_INLINE XXH64_hash_t
 XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
                            const void* XXH_RESTRICT secret, size_t secretSize,
-                           XXH3_f_accumulate_512 f_acc512,
+                           XXH3_f_accumulate f_acc,
                            XXH3_f_scrambleAcc f_scramble)
 {
     XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
 
-    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc512, f_scramble);
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc, f_scramble);
 
     /* converge into final hash */
     XXH_STATIC_ASSERT(sizeof(acc) == 64);
@@ -4465,13 +5273,15 @@ XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
  * It's important for performance to transmit secret's size (when it's static)
  * so that the compiler can properly optimize the vectorized loop.
  * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set.
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
  */
-XXH_FORCE_INLINE XXH64_hash_t
+XXH3_WITH_SECRET_INLINE XXH64_hash_t
 XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
                              XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
 {
     (void)seed64;
-    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate_512, XXH3_scrambleAcc);
+    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
 }
 
 /*
@@ -4480,12 +5290,12 @@ 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);
 }
 
 /*
@@ -4502,18 +5312,20 @@ 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);
     }
 }
 
@@ -4521,12 +5333,12 @@ XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
  * It's important for performance that XXH3_hashLong is not inlined.
  */
 XXH_NO_INLINE XXH64_hash_t
-XXH3_hashLong_64b_withSeed(const void* input, size_t len,
-                           XXH64_hash_t seed, const xxh_u8* secret, size_t secretLen)
+XXH3_hashLong_64b_withSeed(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
 {
     (void)secret; (void)secretLen;
     return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
-                XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret);
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
 }
 
 
@@ -4558,37 +5370,37 @@ 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.
  *
@@ -4612,7 +5424,7 @@ 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 */
@@ -4654,7 +5466,7 @@ static void XXH_alignedFree(void* p)
         XXH_free(base);
     }
 }
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
 {
     XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
@@ -4663,16 +5475,16 @@ XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
     return state;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
 {
     XXH_alignedFree(statePtr);
     return XXH_OK;
 }
 
-/*! @ingroup xxh3_family */
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API void
-XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state)
+XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state)
 {
     XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
 }
@@ -4704,18 +5516,18 @@ 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);
@@ -4724,9 +5536,9 @@ 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);
@@ -4736,9 +5548,9 @@ 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;
@@ -4748,35 +5560,61 @@ 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
@@ -4786,7 +5624,7 @@ XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
 XXH_FORCE_INLINE XXH_errorcode
 XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
             const xxh_u8* XXH_RESTRICT input, size_t len,
-            XXH3_f_accumulate_512 f_acc512,
+            XXH3_f_accumulate f_acc,
             XXH3_f_scrambleAcc f_scramble)
 {
     if (input==NULL) {
@@ -4802,7 +5640,8 @@ XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
          * when operating accumulators directly into state.
          * Operating into stack space seems to enable proper optimization.
          * clang, on the other hand, doesn't seem to need this trick */
-        XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8]; memcpy(acc, state->acc, sizeof(acc));
+        XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
+        XXH_memcpy(acc, state->acc, sizeof(acc));
 #else
         xxh_u64* XXH_RESTRICT const acc = state->acc;
 #endif
@@ -4810,7 +5649,7 @@ XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
         XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
 
         /* small input : just fill in tmp buffer */
-        if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) {
+        if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
             XXH_memcpy(state->buffer + state->bufferedSize, input, len);
             state->bufferedSize += (XXH32_hash_t)len;
             return XXH_OK;
@@ -4832,57 +5671,20 @@ XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
                                &state->nbStripesSoFar, state->nbStripesPerBlock,
                                 state->buffer, XXH3_INTERNALBUFFER_STRIPES,
                                 secret, state->secretLimit,
-                                f_acc512, f_scramble);
+                                f_acc, f_scramble);
             state->bufferedSize = 0;
         }
         XXH_ASSERT(input < bEnd);
-
-        /* large input to consume : ingest per full block */
-        if ((size_t)(bEnd - input) > state->nbStripesPerBlock * XXH_STRIPE_LEN) {
+        if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
             size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
-            XXH_ASSERT(state->nbStripesPerBlock >= state->nbStripesSoFar);
-            /* join to current block's end */
-            {   size_t const nbStripesToEnd = state->nbStripesPerBlock - state->nbStripesSoFar;
-                XXH_ASSERT(nbStripes <= 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);
@@ -4891,19 +5693,19 @@ 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);
 }
 
 
@@ -4912,37 +5714,40 @@ XXH3_digest_long (XXH64_hash_t* acc,
                   const XXH3_state_t* state,
                   const unsigned char* secret)
 {
+    xxh_u8 lastStripe[XXH_STRIPE_LEN];
+    const xxh_u8* lastStripePtr;
+
     /*
      * Digest on a local copy. This way, the state remains unaltered, and it can
      * continue ingesting more input afterwards.
      */
     XXH_memcpy(acc, state->acc, sizeof(state->acc));
     if (state->bufferedSize >= XXH_STRIPE_LEN) {
+        /* Consume remaining stripes then point to remaining data in buffer */
         size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
         size_t nbStripesSoFar = state->nbStripesSoFar;
         XXH3_consumeStripes(acc,
                            &nbStripesSoFar, state->nbStripesPerBlock,
                             state->buffer, nbStripes,
                             secret, state->secretLimit,
-                            XXH3_accumulate_512, XXH3_scrambleAcc);
-        /* last stripe */
-        XXH3_accumulate_512(acc,
-                            state->buffer + state->bufferedSize - XXH_STRIPE_LEN,
-                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+                            XXH3_accumulate, XXH3_scrambleAcc);
+        lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN;
     } else {  /* bufferedSize < XXH_STRIPE_LEN */
-        xxh_u8 lastStripe[XXH_STRIPE_LEN];
+        /* Copy to temp buffer */
         size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
         XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
         XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
         XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
-        XXH3_accumulate_512(acc,
-                            lastStripe,
-                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+        lastStripePtr = lastStripe;
     }
+    /* Last stripe */
+    XXH3_accumulate_512(acc,
+                        lastStripePtr,
+                        secret + state->secretLimit - XXH_SECRET_LASTACC_START);
 }
 
-/*! @ingroup xxh3_family */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
+/*! @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) {
@@ -4958,7 +5763,7 @@ 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 */
 
 
 /* ==========================================
@@ -4978,7 +5783,7 @@ 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. */
@@ -5007,7 +5812,7 @@ 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);
@@ -5034,7 +5839,7 @@ XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_
     }
 }
 
-XXH_FORCE_INLINE XXH128_hash_t
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
 XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
     XXH_ASSERT(input != NULL);
@@ -5109,7 +5914,7 @@ 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);
@@ -5140,7 +5945,7 @@ 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)
@@ -5151,6 +5956,16 @@ 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) {
@@ -5161,6 +5976,7 @@ 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)
@@ -5173,7 +5989,7 @@ 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)
@@ -5182,25 +5998,34 @@ XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
     XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
 
     {   XXH128_hash_t acc;
-        int const nbRounds = (int)len / 32;
-        int i;
+        unsigned i;
         acc.low64 = len * XXH_PRIME64_1;
         acc.high64 = 0;
-        for (i=0; i<4; i++) {
+        /*
+         *  We set as `i` as offset + 32. We do this so that unchanged
+         * `len` can be used as upper bound. This reaches a sweet spot
+         * where both x86 and aarch64 get simple agen and good codegen
+         * for the loop.
+         */
+        for (i = 32; i < 160; i += 32) {
             acc = XXH128_mix32B(acc,
-                                input  + (32 * i),
-                                input  + (32 * i) + 16,
-                                secret + (32 * i),
+                                input  + i - 32,
+                                input  + i - 16,
+                                secret + i - 32,
                                 seed);
         }
         acc.low64 = XXH3_avalanche(acc.low64);
         acc.high64 = XXH3_avalanche(acc.high64);
-        XXH_ASSERT(nbRounds >= 4);
-        for (i=4 ; i < nbRounds; i++) {
+        /*
+         * NB: `i <= len` will duplicate the last 32-bytes if
+         * len % 32 was zero. This is an unfortunate necessity to keep
+         * the hash result stable.
+         */
+        for (i=160; i <= len; i += 32) {
             acc = XXH128_mix32B(acc,
-                                input + (32 * i),
-                                input + (32 * i) + 16,
-                                secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)),
+                                input + i - 32,
+                                input + i - 16,
+                                secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
                                 seed);
         }
         /* last bytes */
@@ -5208,7 +6033,7 @@ XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
                             input + len - 16,
                             input + len - 32,
                             secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
-                            0ULL - seed);
+                            (XXH64_hash_t)0 - seed);
 
         {   XXH128_hash_t h128;
             h128.low64  = acc.low64 + acc.high64;
@@ -5225,12 +6050,12 @@ XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
 XXH_FORCE_INLINE XXH128_hash_t
 XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
                             const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                            XXH3_f_accumulate_512 f_acc512,
+                            XXH3_f_accumulate f_acc,
                             XXH3_f_scrambleAcc f_scramble)
 {
     XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
 
-    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc512, f_scramble);
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble);
 
     /* converge into final hash */
     XXH_STATIC_ASSERT(sizeof(acc) == 64);
@@ -5248,47 +6073,50 @@ 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);
     }
 }
 
@@ -5301,7 +6129,7 @@ XXH3_hashLong_128b_withSeed(const void* input, size_t len,
 {
     (void)secret; (void)secretLen;
     return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
-                XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret);
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
 }
 
 typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
@@ -5331,94 +6159,93 @@ 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) {
@@ -5442,13 +6269,13 @@ 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 */
@@ -5456,11 +6283,11 @@ 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;
@@ -5472,9 +6299,9 @@ 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) {
@@ -5485,9 +6312,9 @@ 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);
@@ -5503,26 +6330,34 @@ XXH128_hashFromCanonical(const XXH128_canonical_t* src)
  */
 #define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
 
-static void XXH3_combine16(void* dst, XXH128_hash_t h128)
+XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
 {
     XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 );
     XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
 }
 
-/*! @ingroup xxh3_family */
+/*! @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);
-    if (secretBuffer == NULL) return XXH_ERROR;
     XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+#else
+    /* production mode, assert() are disabled */
+    if (secretBuffer == NULL) return XXH_ERROR;
     if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+#endif
+
     if (customSeedSize == 0) {
         customSeed = XXH3_kSecret;
         customSeedSize = XXH_SECRET_DEFAULT_SIZE;
     }
+#if (XXH_DEBUGLEVEL >= 1)
     XXH_ASSERT(customSeed != NULL);
+#else
     if (customSeed == NULL) return XXH_ERROR;
+#endif
 
     /* Fill secretBuffer with a copy of customSeed - repeat as needed */
     {   size_t pos = 0;
@@ -5546,9 +6381,9 @@ 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);
@@ -5561,7 +6396,7 @@ 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
 
@@ -5576,5 +6411,5 @@ XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_t seed)
 
 
 #if defined (__cplusplus)
-}
+} /* extern "C" */
 #endif