diff --git a/.github/workflows/build-bsd.yml b/.github/workflows/build-bsd.yml
new file mode 100644
index 0000000000..add3f4ad58
--- /dev/null
+++ b/.github/workflows/build-bsd.yml
@@ -0,0 +1,107 @@
+name: BSD
+
+on:
+  push:
+    branches:
+      - master
+      - ci/*
+  pull_request:
+    branches:
+      - master
+
+env:
+  MELONDS_GIT_BRANCH: ${{ github.ref }}
+  MELONDS_GIT_HASH: ${{ github.sha }}
+  MELONDS_BUILD_PROVIDER: GitHub Actions
+
+jobs:
+  build:
+    strategy:
+      matrix:
+        os: [freebsd, netbsd, openbsd]
+        # The emulated aarch64 VMs are quite slow, so disable aarch64
+        # runs for now.
+        arch: [x86_64] #, aarch64]
+        include:
+          - os: freebsd
+            os_name: FreeBSD
+            # CMake's autorcc for Qt tries to lock files, so we need to
+            # start lockd for NFS, otherwise the build will fail.
+            prepare_vm: service lockd onestart
+            install_deps: >
+              pkg install -y cmake kf6-extra-cmake-modules pkgconf ninja
+              qt6-base qt6-multimedia qt6-svg sdl2 libarchive zstd enet
+              faad2
+
+          - os: netbsd
+            os_name: NetBSD
+            # The NetBSD VM image does not come with X sets by default,
+            # so download and unpack the sets we need manually.
+            # See https://github.com/vmactions/netbsd-vm/issues/2
+            # Also, for some reason pkg_add is not in $PATH, so we have
+            # to call it using its absolute location.
+            prepare_vm: |
+              /usr/sbin/pkg_add pkgin curl
+              mach=$(uname -m | sed "s/evbarm/evbarm-$(uname -p)/")
+              for set in xbase xcomp; do
+                curl -sL ${PKG_PATH%%/pkgsrc*}/NetBSD/NetBSD-$(uname -r
+                  )/$mach/binary/sets/$set.tar.xz | tar xJpf - -C /
+              done
+            install_deps: >
+              pkgin -y install cmake extra-cmake-modules pkgconf
+              ninja-build qt6-qtbase qt6-qtmultimedia qt6-qtsvg SDL2
+              libarchive zstd enet faad2
+
+          - os: openbsd
+            os_name: OpenBSD
+            install_deps: >
+              pkg_add -v cmake kf6-extra-cmake-modules pkgconf ninja
+              qt6-qtbase qt6-qtmultimedia qt6-qtsvg sdl2 libarchive zstd
+              enet faad
+
+
+    name: ${{ matrix.os_name }} / ${{ matrix.arch }}
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        shell: ${{ matrix.os }} {0}
+
+    steps:
+    - uses: actions/checkout@v4
+      name: Check out sources
+
+    - name: Start ${{ matrix.os_name }} VM
+      uses: jenseng/dynamic-uses@v1
+      with:
+        uses: vmactions/${{ matrix.os }}-vm@v1
+        # As we are using multiple steps, we need to use NFS to sync the
+        # workspace between host and VM, as with the default, rsync, the
+        # workspace is only synced before and after the setup VM step.
+        with: >
+          {
+            "arch": ${{ matrix.arch }},
+            "envs": "MELONDS_GIT_BRANCH MELONDS_GIT_HASH
+                     MELONDS_BUILD_PROVIDER",
+            "prepare": ${{ toJSON(matrix.prepare_vm) }},
+            "sync": "nfs",
+            "usesh": "true",
+          }
+
+    - name: Install dependencies
+      run: ${{ matrix.install_deps }}
+
+    - name: Configure
+      run: |
+        cd $GITHUB_WORKSPACE
+        cmake -B build -G Ninja -DMELONDS_EMBED_BUILD_INFO=ON
+
+    - name: Build
+      run: |
+        cd $GITHUB_WORKSPACE
+        cmake --build build
+
+    - name: Upload binary
+      uses: actions/upload-artifact@v4
+      with:
+        name: melonDS-${{ matrix.os }}-${{ matrix.arch }}
+        path: build/melonDS
diff --git a/.github/workflows/build-macos.yml b/.github/workflows/build-macos.yml
index cf9c7f45cd..7d5e94a4c5 100644
--- a/.github/workflows/build-macos.yml
+++ b/.github/workflows/build-macos.yml
@@ -10,7 +10,6 @@ on:
       - master
 
 env:
-  VCPKG_COMMIT: 2ad004460f5db4d3b66f62f5799ff66c265c4b5d
   MELONDS_GIT_BRANCH: ${{ github.ref }}
   MELONDS_GIT_HASH: ${{ github.sha }}
   MELONDS_BUILD_PROVIDER: GitHub Actions
@@ -24,6 +23,10 @@ jobs:
 
     name: ${{ matrix.arch }}
     runs-on: macos-14
+
+    env:
+      VCPKG_BINARY_SOURCES: clear;files,${{ github.workspace }}/vcpkg_cache,readwrite
+
     steps:
       - name: Check out sources
         uses: actions/checkout@v3
@@ -32,10 +35,16 @@ jobs:
           brew install autoconf automake autoconf-archive libtool python-setuptools
       - name: Set up CMake
         uses: lukka/get-cmake@latest
+      - name: Restore vcpkg cache
+        uses: actions/cache/restore@v4
+        with:
+          path: ${{ github.workspace }}/vcpkg_cache
+          key: vcpkg-${{ matrix.arch }}-osx-${{ hashFiles('vcpkg.json') }}
+          restore-keys: vcpkg-${{ matrix.arch }}-osx-
       - name: Set up vcpkg
         uses: lukka/run-vcpkg@v11
         with:
-          vcpkgGitCommitId: ${{ env.VCPKG_COMMIT }}
+          vcpkgJsonGlob: vcpkg.json
       - name: Build
         uses: lukka/run-cmake@v10
         with:
@@ -53,6 +62,11 @@ jobs:
           name: macOS-${{ matrix.arch }}
           path: macOS-${{ matrix.arch }}.zip
           retention-days: 1
+      - name: Save vcpkg cache
+        uses: actions/cache/save@v4
+        with:
+          path: ${{ github.workspace }}/vcpkg_cache
+          key: vcpkg-${{ matrix.arch }}-osx-${{ hashFiles('vcpkg.json') }}
 
   universal-binary:
     name: Universal binary
diff --git a/.github/workflows/build-windows.yml b/.github/workflows/build-windows.yml
index f582c37996..c8597619ed 100644
--- a/.github/workflows/build-windows.yml
+++ b/.github/workflows/build-windows.yml
@@ -10,7 +10,6 @@ on:
       - master
 
 env:
-  VCPKG_COMMIT: 2ad004460f5db4d3b66f62f5799ff66c265c4b5d
   MELONDS_GIT_BRANCH: ${{ github.ref }}
   MELONDS_GIT_HASH: ${{ github.sha }}
   MELONDS_BUILD_PROVIDER: GitHub Actions
@@ -18,10 +17,12 @@ env:
 
 jobs:
   build:
-    runs-on: windows-latest
+    runs-on: windows-2025
     defaults:
       run:
         shell: msys2 {0}
+    env:
+      VCPKG_BINARY_SOURCES: clear;files,${{ github.workspace }}/vcpkg_cache,readwrite
     steps:
     - name: Check out sources
       uses: actions/checkout@v3
@@ -31,10 +32,16 @@ jobs:
         msystem: ucrt64
         update: true
         pacboy: gcc:p cmake:p ninja:p make:p
+    - name: Restore vcpkg cache
+      uses: actions/cache/restore@v4
+      with:
+        path: ${{ github.workspace }}/vcpkg_cache
+        key: vcpkg-x86_64-mingw-${{ hashFiles('vcpkg.json') }}
+        restore-keys: vcpkg-x86_64-mingw-
     - name: Set up vcpkg
       uses: lukka/run-vcpkg@v11
       with:
-        vcpkgGitCommitId: ${{ env.VCPKG_COMMIT }}
+        vcpkgJsonGlob: vcpkg.json
     - name: Configure
       run: cmake --preset=release-mingw-x86_64 -DMELONDS_EMBED_BUILD_INFO=ON
     - name: Build
@@ -43,3 +50,8 @@ jobs:
       with:
         name: melonDS-windows-x86_64
         path: .\build\release-mingw-x86_64\melonDS.exe
+    - name: Save vcpkg cache
+      uses: actions/cache/save@v4
+      with:
+        path: ${{ github.workspace }}/vcpkg_cache
+        key: vcpkg-x86_64-mingw-${{ hashFiles('vcpkg.json') }}
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 531ca19f15..c74f6b93a7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -17,7 +17,7 @@ if (USE_VCPKG)
 endif()
 
 project(melonDS
-    VERSION 1.0
+    VERSION 1.1
     DESCRIPTION "DS emulator, sorta"
     HOMEPAGE_URL "https://melonds.kuribo64.net"
     LANGUAGES C CXX)
@@ -63,7 +63,7 @@ detect_architecture("__arm__" ARM)
 detect_architecture("__aarch64__" ARM64)
 
 cmake_dependent_option(ENABLE_JIT "Enable JIT recompiler" ON
-    "ARCHITECTURE STREQUAL x86_64 OR ARCHITECTURE STREQUAL ARM64" OFF)
+    "ARCHITECTURE STREQUAL x86_64 OR ARCHITECTURE STREQUAL ARM64;NOT ARCHITECTURE STREQUAL x86_64 OR NOT APPLE" OFF)
 cmake_dependent_option(ENABLE_JIT_PROFILING "Enable JIT profiling with VTune" OFF "ENABLE_JIT" OFF)
 option(ENABLE_OGLRENDERER "Enable OpenGL renderer" ON)
 
@@ -85,10 +85,15 @@ endif()
 
 if (WIN32)
     option(BUILD_STATIC "Statically link dependencies" OFF)
-endif()
-
-if (BUILD_STATIC AND WIN32)
-    set(CMAKE_FIND_LIBRARY_SUFFIXES .a)
+    add_definitions(-DWIN32_LEAN_AND_MEAN -DNOMINMAX)
+    if (NOT MINGW)
+        add_definitions(-D__WIN32__ -D_CRT_SECURE_NO_WARNINGS -D_USE_MATH_DEFINES)
+        if (BUILD_STATIC)
+            set(CMAKE_MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
+        endif()
+    elseif (BUILD_STATIC AND MINGW)
+        set(CMAKE_FIND_LIBRARY_SUFFIXES .a)
+    endif()
 endif()
 
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
diff --git a/README.md b/README.md
index e08ca2b80f..db919b498b 100644
--- a/README.md
+++ b/README.md
@@ -2,7 +2,7 @@
 <h2 align="center"><b>melonDS</b></h2>
 <p align="center">
 <a href="http://melonds.kuribo64.net/" alt="melonDS website"><img src="https://img.shields.io/badge/website-melonds.kuribo64.net-%2331352e.svg"></a>
-<a href="http://melonds.kuribo64.net/downloads.php" alt="Release: 1.0"><img src="https://img.shields.io/badge/release-1.0-%235c913b.svg"></a>
+<a href="http://melonds.kuribo64.net/downloads.php" alt="Release: 1.1"><img src="https://img.shields.io/badge/release-1.1-%235c913b.svg"></a>
 <a href="https://www.gnu.org/licenses/gpl-3.0" alt="License: GPLv3"><img src="https://img.shields.io/badge/License-GPL%20v3-%23ff554d.svg"></a>
 <a href="https://kiwiirc.com/client/irc.badnik.net/?nick=IRC-Source_?#melonds" alt="IRC channel: #melonds"><img src="https://img.shields.io/badge/IRC%20chat-%23melonds-%23dd2e44.svg"></a>
 <a href="https://discord.gg/pAMAtExcqV" alt="Discord"><img src="https://img.shields.io/badge/Discord-Kuribo64-7289da?logo=discord&logoColor=white"></a>
@@ -10,6 +10,7 @@
 <a href="https://github.com/melonDS-emu/melonDS/actions/workflows/build-windows.yml?query=event%3Apush"><img src="https://github.com/melonDS-emu/melonDS/actions/workflows/build-windows.yml/badge.svg" /></a>
 <a href="https://github.com/melonDS-emu/melonDS/actions/workflows/build-ubuntu.yml?query=event%3Apush"><img src="https://github.com/melonDS-emu/melonDS/actions/workflows/build-ubuntu.yml/badge.svg" /></a>
 <a href="https://github.com/melonDS-emu/melonDS/actions/workflows/build-macos.yml?query=event%3Apush"><img src="https://github.com/melonDS-emu/melonDS/actions/workflows/build-macos.yml/badge.svg" /></a>
+<a href="https://github.com/melonDS-emu/melonDS/actions/workflows/build-bsd.yml?query=event%3Apush"><img src="https://github.com/melonDS-emu/melonDS/actions/workflows/build-bsd.yml/badge.svg" /></a>
 </p>
 DS emulator, sorta
 
diff --git a/cmake/ConfigureVcpkg.cmake b/cmake/ConfigureVcpkg.cmake
index 2f8e9ad305..b02587c5e1 100644
--- a/cmake/ConfigureVcpkg.cmake
+++ b/cmake/ConfigureVcpkg.cmake
@@ -9,17 +9,23 @@ if (VCPKG_ROOT STREQUAL "${_DEFAULT_VCPKG_ROOT}")
     endif()
     FetchContent_Declare(vcpkg
         GIT_REPOSITORY "https://github.com/Microsoft/vcpkg.git"
-        GIT_TAG 2ad004460f5db4d3b66f62f5799ff66c265c4b5d
+        GIT_TAG HEAD
         EXCLUDE_FROM_ALL
         SOURCE_DIR "${CMAKE_SOURCE_DIR}/vcpkg")
     FetchContent_MakeAvailable(vcpkg)
 endif()
 
-set(VCPKG_OVERLAY_TRIPLETS "${CMAKE_SOURCE_DIR}/cmake/overlay-triplets")
-set(VCPKG_OVERLAY_PORTS "${CMAKE_SOURCE_DIR}/cmake/overlay-ports/faad2")
-
 option(USE_RECOMMENDED_TRIPLETS "Use the recommended triplets that are used for official builds" ON)
 
+# Required for Windows non-MinGW debug builds, optional for all other platforms
+if (WIN32)
+    if (DEFINED ENV{MINGW_PREFIX})
+        option(ENABLE_DEBUG_DEPS "Enable debug builds of vcpkg dependencies" OFF)
+    else()
+        option(ENABLE_DEBUG_DEPS "Enable debug builds of vcpkg dependencies" ON)
+    endif()
+endif()
+
 # Duplicated here because it needs to be set before project()
 option(USE_QT6 "Use Qt 6 instead of Qt 5" ON)
 
@@ -62,9 +68,30 @@ if (USE_RECOMMENDED_TRIPLETS)
             set(CMAKE_OSX_DEPLOYMENT_TARGET 10.15)
         endif()
     elseif(WIN32)
-        # TODO Windows arm64 if possible
         set(_CAN_TARGET_AS_HOST ON)
-        set(_WANTED_TRIPLET x64-mingw-static-release)
+        if (ENABLE_DEBUG_DEPS)
+            set(_RELEASE_SUFFIX "")
+        else()
+            set(_RELEASE_SUFFIX "-release")
+        endif()
+
+        if (DEFINED ENV{MINGW_PREFIX})
+            set(_WIN32_ENV mingw)
+        else()
+            set(_WIN32_ENV windows)
+        endif()
+
+        if ("$ENV{PROCESSOR_ARCHITECTURE}" STREQUAL "AMD64")
+            set(_WIN32_ARCH x64)
+        elseif ("$ENV{PROCESSOR_ARCHITECTURE}" STREQUAL "ARM64")
+            set(_WIN32_ARCH arm64)
+        elseif ("$ENV{PROCESSOR_ARCHITECTURE}" STREQUAL "x86")
+            set(_WIN32_ARCH x86)
+        else()
+            message(FATAL_ERROR "Unknown processor architecture '$ENV{PROCESSOR_ARCHITECTURE}'. Please disable USE_RECOMMENDED_TRIPLETS and set your vcpkg settings manually.")
+        endif()
+
+        set(_WANTED_TRIPLET ${_WIN32_ARCH}-${_WIN32_ENV}-static${_RELEASE_SUFFIX})
     elseif(CMAKE_HOST_SYSTEM_NAME STREQUAL Linux)
         # Can't really detect cross compiling here.
         set(_CAN_TARGET_AS_HOST ON)
diff --git a/flake.lock b/flake.lock
index 85be38faac..725c6d355e 100644
--- a/flake.lock
+++ b/flake.lock
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1739020877,
-        "narHash": "sha256-mIvECo/NNdJJ/bXjNqIh8yeoSjVLAuDuTUzAo7dzs8Y=",
+        "lastModified": 1761672384,
+        "narHash": "sha256-o9KF3DJL7g7iYMZq9SWgfS1BFlNbsm6xplRjVlOCkXI=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "a79cfe0ebd24952b580b1cf08cd906354996d547",
+        "rev": "08dacfca559e1d7da38f3cf05f1f45ee9bfd213c",
         "type": "github"
       },
       "original": {
diff --git a/flake.nix b/flake.nix
index 8dedde8f0d..efa9be50c3 100644
--- a/flake.nix
+++ b/flake.nix
@@ -40,6 +40,7 @@
           libGL
           libslirp
           enet
+          faad2
         ]) ++ optionals (!isDarwin) (with pkgs; [
           kdePackages.extra-cmake-modules
           qt6.qtwayland
@@ -79,6 +80,12 @@
           packages = with pkgs; [
             qt6.qttools
           ];
+
+          # https://discourse.nixos.org/t/how-to-create-a-qt-devshell/58529/5
+          shellHook = ''
+            export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$APPEND_LIBRARY_PATH"
+            eval $(echo "''${qtWrapperArgs[@]}"|perl -n -e '$_ =~ s/--prefix (\w+) : ([\w-.\/]+)/export ''${1}="''${2}:\''${''${1}}";/g;print')
+          '';
         };
 
         # Shell for building static melonDS release builds with vcpkg
diff --git a/melonDLDI/melonDLDI.s b/melonDLDI/melonDLDI.s
index 399ad032ee..eb98a67118 100755
--- a/melonDLDI/melonDLDI.s
+++ b/melonDLDI/melonDLDI.s
@@ -7,7 +7,7 @@ _start:
 .word 0xBF8DA5ED
 .string " Chishm"
 .byte 1
-.byte 9 @ size
+.byte 10 @ size
 .byte 0
 .byte 0
 .string "melonDS DLDI driver"
diff --git a/res/melon.rc.in b/res/melon.rc.in
index c96d0f9e2c..ac2bb84b57 100644
--- a/res/melon.rc.in
+++ b/res/melon.rc.in
@@ -31,4 +31,6 @@ FILETYPE VFT_APP
 	}
 }
 
+#if (0${MINGW})
 1 24 "xp.manifest"
+#endif
diff --git a/src/ARCodeFile.cpp b/src/ARCodeFile.cpp
index deb8bc403f..df6d26e4ce 100644
--- a/src/ARCodeFile.cpp
+++ b/src/ARCodeFile.cpp
@@ -16,9 +16,10 @@
     with melonDS. If not, see http://www.gnu.org/licenses/.
 */
 
-#include <stdio.h>
-#include <string.h>
+#include <cstdio>
+#include <cstring>
 #include "ARCodeFile.h"
+#include "ARDatabaseDAT.h"
 #include "Platform.h"
 
 namespace melonDS
@@ -44,10 +45,21 @@ std::vector<ARCode> ARCodeFile::GetCodes() const noexcept
 
     std::vector<ARCode> codes;
 
-    for (const ARCodeCat& cat : Categories)
+    for (auto& item : RootCat.Children)
     {
-        for (const ARCode& code : cat.Codes)
+        if (std::holds_alternative<ARCodeCat>(item))
         {
+            auto& cat = std::get<ARCodeCat>(item);
+
+            for (auto& childitem : cat.Children)
+            {
+                auto& code = std::get<ARCode>(childitem);
+                codes.push_back(code);
+            }
+        }
+        else
+        {
+            auto& code = std::get<ARCode>(item);
             codes.push_back(code);
         }
     }
@@ -60,13 +72,17 @@ bool ARCodeFile::Load()
     FileHandle* f = OpenFile(Filename, FileMode::ReadText);
     if (!f) return true;
 
-    Categories.clear();
+    RootCat.Parent = nullptr;
+    RootCat.OnlyOneCodeEnabled = false;
+    RootCat.Children.clear();
 
     bool isincat = false;
-    ARCodeCat curcat;
+    ARCodeCat* curcat = &RootCat;
 
     bool isincode = false;
-    ARCode curcode;
+    ARCode* curcode = nullptr;
+
+    int lastentry = 0;
 
     char linebuf[1024];
     while (!IsEndOfFile(f))
@@ -83,27 +99,54 @@ bool ARCodeFile::Load()
         if (start[0]=='#' || start[0]=='\r' || start[0]=='\n' || start[0]=='\0')
             continue;
 
-        if (!strncasecmp(start, "CAT", 3))
+        if (!strncasecmp(start, "ROOT", 4))
+        {
+            isincode = false;
+            isincat = true;
+
+            curcat = &RootCat;
+            lastentry = 0;
+        }
+        else if (!strncasecmp(start, "CAT", 3))
         {
             char catname[128];
-            int ret = sscanf(start, "CAT %127[^\r\n]", catname);
+            int ret, retchk;
+            int onlyone;
+            if (start[3] == ' ' && (start[4] == '0' || start[4] == '1') && start[5] == ' ')
+            {
+                retchk = 2;
+                ret = sscanf(start, "CAT %d %127[^\r\n]", &onlyone, catname);
+            }
+            else
+            {
+                // backwards compatibility
+                onlyone = 0;
+                retchk = 1;
+                ret = sscanf(start, "CAT %127[^\r\n]", catname);
+            }
             catname[127] = '\0';
 
-            if (ret < 1)
+            if (ret < retchk)
             {
                 Log(LogLevel::Error, "AR: malformed CAT line: %s\n", start);
                 CloseFile(f);
                 return false;
             }
 
-            if (isincode) curcat.Codes.push_back(curcode);
             isincode = false;
-
-            if (isincat) Categories.push_back(curcat);
             isincat = true;
 
-            curcat.Name = catname;
-            curcat.Codes.clear();
+            ARCodeCat cat = {
+                .Parent = &RootCat,
+                .Name = catname,
+                .Description = "",
+                .OnlyOneCodeEnabled = onlyone!=0,
+                .Children = {}
+            };
+            RootCat.Children.emplace_back(cat);
+            curcat = &std::get<ARCodeCat>(RootCat.Children.back());
+
+            lastentry = 1;
         }
         else if (!strncasecmp(start, "CODE", 4))
         {
@@ -126,12 +169,39 @@ bool ARCodeFile::Load()
                 return false;
             }
 
-            if (isincode) curcat.Codes.push_back(curcode);
             isincode = true;
 
-            curcode.Name = codename;
-            curcode.Enabled = enable!=0;
-            curcode.Code.clear();
+            ARCode code = {
+                .Parent = curcat,
+                .Name = codename,
+                .Description = "",
+                .Enabled = enable!=0,
+                .Code = {}
+            };
+            curcat->Children.emplace_back(code);
+            curcode = &std::get<ARCode>(curcat->Children.back());
+
+            lastentry = 2;
+        }
+        else if (!strncasecmp(start, "DESC", 4))
+        {
+            char desc[256];
+            int ret = sscanf(start, "DESC %255[^\r\n]", desc);
+            desc[255] = '\0';
+
+            if (ret < 1)
+                continue;
+
+            if (lastentry == 2)
+                curcode->Description = desc;
+            else if (lastentry == 1)
+                curcat->Description = desc;
+            else
+            {
+                Log(LogLevel::Error, "AR: encountered DESC line not part of anything\n");
+                CloseFile(f);
+                return false;
+            }
         }
         else
         {
@@ -152,13 +222,12 @@ bool ARCodeFile::Load()
                 return false;
             }
 
-            curcode.Code.push_back(c0);
-            curcode.Code.push_back(c1);
+            curcode->Code.push_back(c0);
+            curcode->Code.push_back(c1);
         }
     }
 
-    if (isincode) curcat.Codes.push_back(curcode);
-    if (isincat) Categories.push_back(curcat);
+    FinalizeList();
 
     CloseFile(f);
     return true;
@@ -169,22 +238,51 @@ bool ARCodeFile::Save()
     FileHandle* f = Platform::OpenFile(Filename, FileMode::WriteText);
     if (!f) return false;
 
-    for (ARCodeCatList::iterator it = Categories.begin(); it != Categories.end(); it++)
+    bool isincat = true;
+
+    for (auto& item : RootCat.Children)
     {
-        ARCodeCat& cat = *it;
+        if (std::holds_alternative<ARCodeCat>(item))
+        {
+            auto& cat = std::get<ARCodeCat>(item);
 
-        if (it != Categories.begin()) FileWriteFormatted(f, "\n");
-        FileWriteFormatted(f, "CAT %s\n\n", cat.Name.c_str());
+            FileWriteFormatted(f, "CAT %d %s\n", cat.OnlyOneCodeEnabled, cat.Name.c_str());
+            if (!cat.Description.empty())
+                FileWriteFormatted(f, "DESC %s\n", cat.Description.c_str());
+            FileWriteFormatted(f, "\n");
+
+            isincat = true;
+
+            for (auto& childitem : cat.Children)
+            {
+                auto& code = std::get<ARCode>(childitem);
 
-        for (ARCodeList::iterator jt = cat.Codes.begin(); jt != cat.Codes.end(); jt++)
+                FileWriteFormatted(f, "CODE %d %s\n", code.Enabled, code.Name.c_str());
+                if (!code.Description.empty())
+                    FileWriteFormatted(f, "DESC %s\n", code.Description.c_str());
+
+                for (size_t i = 0; i < code.Code.size(); i+=2)
+                    FileWriteFormatted(f, "%08X %08X\n", code.Code[i], code.Code[i + 1]);
+
+                FileWriteFormatted(f, "\n");
+            }
+        }
+        else
         {
-            ARCode& code = *jt;
+            auto& code = std::get<ARCode>(item);
+
+            if (isincat)
+            {
+                isincat = false;
+                FileWriteFormatted(f, "ROOT\n\n");
+            }
+
             FileWriteFormatted(f, "CODE %d %s\n", code.Enabled, code.Name.c_str());
+            if (!code.Description.empty())
+                FileWriteFormatted(f, "DESC %s\n", code.Description.c_str());
 
             for (size_t i = 0; i < code.Code.size(); i+=2)
-            {
                 FileWriteFormatted(f, "%08X %08X\n", code.Code[i], code.Code[i + 1]);
-            }
 
             FileWriteFormatted(f, "\n");
         }
@@ -194,4 +292,109 @@ bool ARCodeFile::Save()
     return true;
 }
 
+void ARCodeFile::Import(ARDatabaseEntry& dbentry, ARCodeEnableMap& enablemap, bool clear)
+{
+    bool hasenablemap = !enablemap.empty();
+
+    if (clear)
+        RootCat.Children.clear();
+
+    for (auto& item : dbentry.RootCat.Children)
+    {
+        if (std::holds_alternative<ARCodeCat>(item))
+        {
+            auto& cat = std::get<ARCodeCat>(item);
+
+            bool shouldimport = false;
+            if (hasenablemap)
+            {
+                for (auto& childitem: cat.Children)
+                {
+                    auto& code = std::get<ARCode>(childitem);
+                    if (enablemap[&code])
+                    {
+                        shouldimport = true;
+                        break;
+                    }
+                }
+            }
+            else
+                shouldimport = true;
+
+            if (!shouldimport)
+                continue;
+
+            ARCodeCat newcat = {
+                .Parent = &RootCat,
+                .Name = cat.Name,
+                .Description = cat.Description,
+                .OnlyOneCodeEnabled = cat.OnlyOneCodeEnabled,
+                .Children = {}
+            };
+            RootCat.Children.emplace_back(newcat);
+            ARCodeCat* parentptr = &std::get<ARCodeCat>(RootCat.Children.back());
+
+            for (auto& childitem : cat.Children)
+            {
+                auto& code = std::get<ARCode>(childitem);
+                if (hasenablemap && (!enablemap[&code]))
+                    continue;
+
+                ARCode newcode = {
+                    .Parent = parentptr,
+                    .Name = code.Name,
+                    .Description = code.Description,
+                    .Enabled = code.Enabled,
+                    .Code = code.Code
+                };
+                parentptr->Children.emplace_back(newcode);
+            }
+        }
+        else
+        {
+            auto& code = std::get<ARCode>(item);
+            if (hasenablemap && (!enablemap[&code]))
+                continue;
+
+            ARCode newcode = {
+                .Parent = &RootCat,
+                .Name = code.Name,
+                .Description = code.Description,
+                .Enabled = code.Enabled,
+                .Code = code.Code
+            };
+            RootCat.Children.emplace_back(newcode);
+        }
+    }
+
+    FinalizeList();
+}
+
+void ARCodeFile::FinalizeList()
+{
+    for (auto& item : RootCat.Children)
+    {
+        if (!std::holds_alternative<ARCodeCat>(item))
+            continue;
+
+        auto& cat = std::get<ARCodeCat>(item);
+        if (!cat.OnlyOneCodeEnabled)
+            continue;
+
+        // for categories that only allow one code to be enabled:
+        // make sure we don't have multiple ones enabled
+
+        bool foundone = false;
+        for (auto& childitem : cat.Children)
+        {
+            auto& code = std::get<ARCode>(childitem);
+            if (!code.Enabled) continue;
+            if (foundone)
+                code.Enabled = false;
+            else
+                foundone = true;
+        }
+    }
+}
+
 }
\ No newline at end of file
diff --git a/src/ARCodeFile.h b/src/ARCodeFile.h
index 823871da70..200259e8d5 100644
--- a/src/ARCodeFile.h
+++ b/src/ARCodeFile.h
@@ -22,26 +22,38 @@
 #include <string>
 #include <list>
 #include <vector>
+#include <unordered_map>
+#include <variant>
 #include "types.h"
 
 namespace melonDS
 {
+
+struct ARCodeCat;
+
 struct ARCode
 {
+    ARCodeCat* Parent;
     std::string Name;
+    std::string Description;
     bool Enabled;
     std::vector<u32> Code;
 };
 
-typedef std::list<ARCode> ARCodeList;
+typedef std::variant<ARCode, ARCodeCat> ARCodeItem;
+typedef std::list<ARCodeItem> ARCodeItemList;
 
 struct ARCodeCat
 {
+    ARCodeCat* Parent;
     std::string Name;
-    ARCodeList Codes;
+    std::string Description;
+    bool OnlyOneCodeEnabled;
+    ARCodeItemList Children;
 };
 
-typedef std::list<ARCodeCat> ARCodeCatList;
+struct ARDatabaseEntry;
+typedef std::unordered_map<ARCode*, bool> ARCodeEnableMap;
 
 
 class ARCodeFile
@@ -57,10 +69,14 @@ class ARCodeFile
     bool Load();
     bool Save();
 
-    ARCodeCatList Categories {};
+    void Import(ARDatabaseEntry& dbentry, ARCodeEnableMap& enablemap, bool clear);
+
+    ARCodeCat RootCat {};
 
 private:
     std::string Filename;
+
+    void FinalizeList();
 };
 
 }
diff --git a/src/ARDatabaseDAT.cpp b/src/ARDatabaseDAT.cpp
new file mode 100644
index 0000000000..7246480ceb
--- /dev/null
+++ b/src/ARDatabaseDAT.cpp
@@ -0,0 +1,374 @@
+/*
+    Copyright 2016-2025 melonDS team
+
+    This file is part of melonDS.
+
+    melonDS is free software: you can redistribute it and/or modify it under
+    the terms of the GNU General Public License as published by the Free
+    Software Foundation, either version 3 of the License, or (at your option)
+    any later version.
+
+    melonDS is distributed in the hope that it will be useful, but WITHOUT ANY
+    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+    FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with melonDS. If not, see http://www.gnu.org/licenses/.
+*/
+
+#include <cstring>
+#include "ARDatabaseDAT.h"
+#include "Platform.h"
+
+/* FILE FORMAT for usrcheat.dat
+
+ header:
+ 00: ID string "R4 CheatCode" (12 bytes)
+ 0C: version? must be 0x100
+ 10: database description
+ 4C: ??? (gets set by r4cce.exe, but zero in usrcheat.dat)
+ 50: database active flag
+ 100: entry list
+
+ entry:
+ 00: game code
+ 04: checksum of ROM header (logical NOT of CRC32 over first 512 bytes)
+ 08: offset to cheat data (absolute)
+ 0C: zero
+
+ the entry list is terminated with an all-zero entry
+
+ cheat data: header then items
+
+ header:
+ * game name (null-term string)
+ * padding (to align entry below to 4-byte boundary)
+ * flags (word): bit0-27 = total number of items, bit28-31 = game active (0=inactive, F=active)
+ * master codes (8 words) - use unclear
+
+ item:
+ * flags (word): bit0-23 = data length, bit24 = flag, bit28 = type
+ * item name (null-term string)
+ * item description (null-term string)
+ * padding (to align data below to 4-byte boundary)
+ * item data
+
+ meaning of item type/flag:
+ * type=0: cheat, bit0-23 = total data length in words, bit24 = cheat active
+ * type=1: category, bit0-23 = number of cheats inside, bit24 = only one cheat may be active in this category
+
+ empty description still takes up one extra byte (for the null terminator)
+
+ for a category: item data is a list of cheat items
+
+ for a cheat: item data is as follows:
+ * number of code words (word)
+ * code
+*/
+
+namespace melonDS
+{
+using namespace Platform;
+
+// TODO: more user-friendly error reporting
+
+std::string ReadNTString(Platform::FileHandle* f);
+void AlignFilePos(Platform::FileHandle* f);
+
+
+ARDatabaseDAT::ARDatabaseDAT(const std::string& filename)
+{
+    Filename = filename;
+
+    if (!LoadEntries())
+        Error = true;
+}
+
+bool ARDatabaseDAT::FindGameCode(u32 gamecode)
+{
+    auto it = EntryList.find(gamecode);
+    if (it == EntryList.end())
+        return false;
+    if ((*it).second.empty())
+        return false;
+    return true;
+}
+
+ARDatabaseEntryList ARDatabaseDAT::GetEntriesByGameCode(u32 gamecode)
+{
+    ARDatabaseEntryList ret;
+    auto it = EntryList.find(gamecode);
+    if (it == EntryList.end())
+        return ret;
+
+    for (auto& info : (*it).second)
+    {
+        ARDatabaseEntry entry;
+        LoadCheatCodes(info, entry);
+        ret.push_back(entry);
+    }
+
+    return ret;
+}
+
+bool ARDatabaseDAT::LoadEntries()
+{
+    FileHandle* f = OpenFile(Filename, FileMode::Read);
+    if (!f) return false;
+
+    u64 filelen = FileLength(f);
+    if (filelen > 0xFFFFFFFFULL)
+    {
+        CloseFile(f);
+        return false;
+    }
+
+    char header[16];
+    FileRead(header, 16, 1, f);
+    //if (strncmp(header, "R4 CheatCode", 12) != 0)
+    if (memcmp(header, "R4 CheatCode\x00\x01\x00\x00", 16) != 0)
+    {
+        CloseFile(f);
+        return false;
+    }
+
+    char name[0x3D] = {0};
+    FileRead(name, 0x3C, 1, f);
+    DBName = name;
+
+    FileSeek(f, 0x100, FileSeekOrigin::Start);
+    while (!IsEndOfFile(f))
+    {
+        u32 entrydata[4];
+        FileRead(entrydata, 16, 1, f);
+
+        // a zero entry marks the end of the entry list
+        if (entrydata[0] == 0)
+            break;
+
+        if ((entrydata[2] < 0x100) || (entrydata[2] >= filelen))
+        {
+            Log(LogLevel::Error, "AR: malformed database file (invalid offset %08X)\n", entrydata[2]);
+            CloseFile(f);
+            return false;
+        }
+
+        EntryInfo entry;
+        entry.GameCode = entrydata[0];
+        entry.Checksum = entrydata[1];
+        entry.Offset = entrydata[2];
+
+        EntryList[entry.GameCode].push_back(entry);
+    }
+
+    CloseFile(f);
+    return true;
+}
+
+bool ARDatabaseDAT::LoadCheatCodes(EntryInfo& info, ARDatabaseEntry& entry)
+{
+    FileHandle* f = OpenFile(Filename, FileMode::Read);
+    if (!f) return false;
+
+    u64 filelen = FileLength(f);
+    if (filelen > 0xFFFFFFFFULL)
+    {
+        CloseFile(f);
+        return false;
+    }
+
+    if ((info.Offset < 0x100) || (info.Offset >= filelen))
+    {
+        CloseFile(f);
+        return false;
+    }
+
+    entry.GameCode = info.GameCode;
+    entry.Checksum = info.Checksum;
+
+    FileSeek(f, info.Offset, FileSeekOrigin::Start);
+
+    entry.Name = ReadNTString(f);
+    AlignFilePos(f);
+
+    u32 flags[9] = {0};
+    FileRead(flags, 4*9, 1, f);
+
+    entry.RootCat.Parent = nullptr;
+    entry.RootCat.OnlyOneCodeEnabled = false;
+    entry.RootCat.Children.clear();
+
+    ARCodeCat* curcat = &entry.RootCat;
+    int catlen = 0;
+
+    u32 numentries = flags[0] & 0xFFFFFF;
+    for (u32 i = 0; i < numentries; i++)
+    {
+        if (IsEndOfFile((f)))
+            break;
+
+        u32 itemflags = 0;
+        FileRead(&itemflags, 4, 1, f);
+
+        u32 totallen = itemflags & 0xFFFFFF;
+
+        std::string itemname = ReadNTString(f);
+        std::string itemdesc = ReadNTString(f);
+        AlignFilePos(f);
+
+        if (itemflags & (1<<28))
+        {
+            // this item is a category
+
+            if ((totallen >= 0x10000) || (totallen == 0))
+            {
+                Log(LogLevel::Error, "AR: unreasonable category length %08X\n",
+                    totallen);
+                Log(LogLevel::Error, "game=%s, offset=%08X, cat=%s\n",
+                    entry.Name.c_str(), info.Offset, itemname.c_str());
+
+                CloseFile(f);
+                return false;
+            }
+
+            ARCodeCat cat = {
+                    .Parent = &entry.RootCat,
+                    .Name = itemname,
+                    .Description = itemdesc,
+                    .OnlyOneCodeEnabled = !!(itemflags & (1<<24)),
+                    .Children = {}
+            };
+            entry.RootCat.Children.emplace_back(cat);
+            curcat = &std::get<ARCodeCat>(entry.RootCat.Children.back());
+
+            catlen = totallen;
+        }
+        else
+        {
+            // this item is a code
+
+            u32 codelen = 0;
+            FileRead(&codelen, 4, 1, f);
+
+            u32 chklen = itemname.length() + 1 + itemdesc.length() + 1;
+            chklen = ((chklen + 3) >> 2) + 1 + codelen;
+            if (chklen != totallen)
+            {
+                Log(LogLevel::Error, "AR: malformed code entry, codelen=%08X, totallen=%08X (expected %08X)\n",
+                    codelen, totallen, chklen);
+                Log(LogLevel::Error, "game=%s, offset=%08X, cheat=%s\n",
+                    entry.Name.c_str(), info.Offset, itemname.c_str());
+
+                CloseFile(f);
+                return false;
+            }
+
+            if ((codelen >= 0x100000) || (codelen & 1))
+            {
+                Log(LogLevel::Error, "AR: unreasonable code length %08X\n",
+                    codelen);
+                Log(LogLevel::Error, "game=%s, offset=%08X, cheat=%s\n",
+                    entry.Name.c_str(), info.Offset, itemname.c_str());
+
+                CloseFile(f);
+                return false;
+            }
+
+            if (catlen == 0)
+            {
+                curcat = &entry.RootCat;
+            }
+
+            ARCode code;
+            code.Name = itemname;
+            code.Description = itemdesc;
+            code.Enabled = !!(itemflags & (1<<24));
+
+            u32* rawcode = new u32[codelen];
+            FileRead(rawcode, codelen*4, 1, f);
+
+            for (u32 j = 0; j < codelen; j+=2)
+            {
+                code.Code.push_back(rawcode[j]);
+                code.Code.push_back(rawcode[j+1]);
+            }
+
+            delete[] rawcode;
+
+            curcat->Children.emplace_back(code);
+
+            if (catlen >= 0)
+                catlen--;
+        }
+    }
+
+    for (auto& item : entry.RootCat.Children)
+    {
+        if (!std::holds_alternative<ARCodeCat>(item))
+            continue;
+
+        auto& cat = std::get<ARCodeCat>(item);
+        if (!cat.OnlyOneCodeEnabled)
+            continue;
+
+        // for categories that only allow one code to be enabled:
+        // make sure we don't have multiple ones enabled
+
+        bool foundone = false;
+        for (auto& childitem : cat.Children)
+        {
+            auto& code = std::get<ARCode>(childitem);
+            if (!code.Enabled) continue;
+            if (foundone)
+                code.Enabled = false;
+            else
+                foundone = true;
+        }
+    }
+
+    CloseFile(f);
+    return true;
+}
+
+
+std::string ReadNTString(Platform::FileHandle* f)
+{
+    char tmp[256];
+    std::string ret;
+    int readlen = 0;
+    u64 startpos = FilePosition(f);
+
+    // TODO might break with UTF8 and shit
+
+    while (!IsEndOfFile(f))
+    {
+        // read 256 bytes of data, see where the string actually ends
+        memset(tmp, 0, 256);
+        u64 nread = FileRead(tmp, 1, 256, f);
+
+        bool done = false;
+        for (int i = 0; i < nread; i++)
+        {
+            readlen++;
+            if (!tmp[i]) { done = true; break; }
+            ret += tmp[i];
+        }
+
+        if (done)
+            break;
+    }
+
+    // correct the file position to point right after the end of this string
+    FileSeek(f, startpos + readlen, FileSeekOrigin::Start);
+
+    return ret;
+}
+
+void AlignFilePos(Platform::FileHandle* f)
+{
+    u64 pos = FilePosition(f);
+    if (pos & 3)
+        FileSeek(f, (pos + 3) & (~3), FileSeekOrigin::Start);
+}
+
+}
diff --git a/src/ARDatabaseDAT.h b/src/ARDatabaseDAT.h
new file mode 100644
index 0000000000..acf0fb862f
--- /dev/null
+++ b/src/ARDatabaseDAT.h
@@ -0,0 +1,74 @@
+/*
+    Copyright 2016-2025 melonDS team
+
+    This file is part of melonDS.
+
+    melonDS is free software: you can redistribute it and/or modify it under
+    the terms of the GNU General Public License as published by the Free
+    Software Foundation, either version 3 of the License, or (at your option)
+    any later version.
+
+    melonDS is distributed in the hope that it will be useful, but WITHOUT ANY
+    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+    FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with melonDS. If not, see http://www.gnu.org/licenses/.
+*/
+
+#ifndef ARDATABASEDAT_H
+#define ARDATABASEDAT_H
+
+#include <string>
+#include <list>
+#include <vector>
+#include <unordered_map>
+#include "types.h"
+#include "ARCodeFile.h"
+
+namespace melonDS
+{
+
+struct ARDatabaseEntry
+{
+    u32 GameCode;
+    u32 Checksum;
+    std::string Name;
+    ARCodeCat RootCat;
+};
+
+typedef std::vector<ARDatabaseEntry> ARDatabaseEntryList;
+
+
+class ARDatabaseDAT
+{
+public:
+    ARDatabaseDAT(const std::string& filename);
+    ~ARDatabaseDAT() noexcept = default;
+
+    bool Error = false;
+
+    std::string GetDBName() const { return DBName; }
+    bool FindGameCode(u32 gamecode);
+    ARDatabaseEntryList GetEntriesByGameCode(u32 gamecode);
+
+private:
+    std::string Filename;
+    std::string DBName;
+
+    struct EntryInfo
+    {
+        u32 GameCode;
+        u32 Checksum;
+        u32 Offset;
+    };
+
+    // list of entries per gamecode
+    std::unordered_map<u32, std::vector<EntryInfo>> EntryList;
+
+    bool LoadEntries();
+    bool LoadCheatCodes(EntryInfo& info, ARDatabaseEntry& entry);
+};
+
+}
+#endif // ARDATABASEDAT_H
diff --git a/src/AREngine.cpp b/src/AREngine.cpp
index 8a5b088117..2a5233cbc3 100644
--- a/src/AREngine.cpp
+++ b/src/AREngine.cpp
@@ -285,8 +285,39 @@ void AREngine::RunCheat(const ARCode& arcode)
             offset = b;
             break;
 
-        case 0xD4: // datareg += b
-            datareg += b;
+        case 0xD4: // data op
+            switch (a & 0xFF)
+            {
+                case 0x00: datareg += b; break;
+                case 0x01: datareg |= b; break;
+                case 0x02: datareg &= b; break;
+                case 0x03: datareg ^= b; break;
+
+                case 0x04:
+                    b &= 0xFF;
+                    if (b > 31) datareg = 0;
+                    else        datareg <<= b;
+                    break;
+                case 0x05:
+                    b &= 0xFF;
+                    if (b > 31) datareg = 0;
+                    else        datareg >>= b;
+                    break;
+                case 0x06:
+                    datareg = ROR(datareg, b & 0x1F);
+                    break;
+                case 0x07:
+                    b &= 0xFF;
+                    if (b > 31) datareg = ((s32)datareg) >> 31;
+                    else        datareg = ((s32)datareg) >> b;
+                    break;
+
+                case 0x08: datareg *= b; break;
+
+                default:
+                    Log(LogLevel::Warn, "!! bad AR D4 opcode %08X %08X\n", a, b);
+                    break;
+            }
             break;
 
         case 0xD5: // datareg = b
diff --git a/src/ARMInterpreter_Branch.cpp b/src/ARMInterpreter_Branch.cpp
index df24f20964..0c8b6676b0 100644
--- a/src/ARMInterpreter_Branch.cpp
+++ b/src/ARMInterpreter_Branch.cpp
@@ -106,11 +106,22 @@ void T_BL_LONG_2(ARM* cpu)
 {
     s32 offset = (cpu->CurInstr & 0x7FF) << 1;
     u32 pc = cpu->R[14] + offset;
-    cpu->R[14] = (cpu->R[15] - 2) | 1;
 
-    if ((cpu->Num==1) || (cpu->CurInstr & (1<<12)))
+    if ((cpu->Num==1) || (cpu->CurInstr & (1<<12))) // BL
+    {
         pc |= 1;
+    }
+    else // BLX
+    {
+        if (cpu->CurInstr & 1) // lsb of immediate is set, implying halfword offset; this raises undefined.
+            return T_UNK(cpu);
 
+        // instruction always switches to arm mode
+        // interworking bit should be cleared.
+        pc &= ~1;
+    }
+
+    cpu->R[14] = (cpu->R[15] - 2) | 1;
     cpu->JumpTo(pc);
 }
 
diff --git a/src/ARMJIT.cpp b/src/ARMJIT.cpp
index 39abdd59bb..a9489cac0d 100644
--- a/src/ARMJIT.cpp
+++ b/src/ARMJIT.cpp
@@ -1161,6 +1161,8 @@ void ARMJIT::JitEnableWrite() noexcept
     #if defined(__APPLE__) && defined(__aarch64__)
         if (__builtin_available(macOS 11.0, *))
             pthread_jit_write_protect_np(false);
+    #elif defined(__NetBSD__)
+        mprotect(JITCompiler.CodeMemBase, ARMJIT_Global::CodeMemorySliceSize, PROT_READ | PROT_WRITE);
     #endif
 }
 
@@ -1169,6 +1171,8 @@ void ARMJIT::JitEnableExecute() noexcept
     #if defined(__APPLE__) && defined(__aarch64__)
         if (__builtin_available(macOS 11.0, *))
             pthread_jit_write_protect_np(true);
+    #elif defined(__NetBSD__)
+        mprotect(JITCompiler.CodeMemBase, ARMJIT_Global::CodeMemorySliceSize, PROT_READ | PROT_EXEC);
     #endif
 }
 
diff --git a/src/ARMJIT_Global.cpp b/src/ARMJIT_Global.cpp
index c9432cbf39..e4aa42f9b3 100644
--- a/src/ARMJIT_Global.cpp
+++ b/src/ARMJIT_Global.cpp
@@ -25,7 +25,7 @@ std::mutex globalMutex;
 #define APPLE_AARCH64
 #endif
 
-#ifndef APPLE_AARCH64
+#if !defined(APPLE_AARCH64) && !defined(__NetBSD__) && !defined(__OpenBSD__)
 static constexpr size_t NumCodeMemSlices = 4;
 static constexpr size_t CodeMemoryAlignedSize = NumCodeMemSlices * CodeMemorySliceSize;
 
@@ -46,7 +46,7 @@ void* AllocateCodeMem()
 {
     std::lock_guard guard(globalMutex);
 
-#ifndef APPLE_AARCH64
+#if !defined(APPLE_AARCH64) && !defined(__NetBSD__) && !defined(__OpenBSD__)
     if (AvailableCodeMemSlices)
     {
         int slice = __builtin_ctz(AvailableCodeMemSlices);
@@ -61,6 +61,8 @@ void* AllocateCodeMem()
     return VirtualAlloc(nullptr, CodeMemorySliceSize, MEM_RESERVE|MEM_COMMIT, PAGE_EXECUTE_READWRITE);
 #elif defined(APPLE_AARCH64)
     return mmap(NULL, CodeMemorySliceSize, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS | MAP_JIT,-1, 0);
+#elif defined(__NetBSD__)
+    return mmap(nullptr, CodeMemorySliceSize, PROT_MPROTECT(PROT_READ | PROT_WRITE | PROT_EXEC), MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 #else
     //printf("mmaping...\n");
     return mmap(nullptr, CodeMemorySliceSize, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
@@ -71,7 +73,7 @@ void FreeCodeMem(void* codeMem)
 {
     std::lock_guard guard(globalMutex);
 
-#ifndef APPLE_AARCH64
+#if !defined(APPLE_AARCH64) && !defined(__NetBSD__) && !defined(__OpenBSD__)
     for (int i = 0; i < NumCodeMemSlices; i++)
     {
         if (codeMem == &GetAlignedCodeMemoryStart()[CodeMemorySliceSize * i])
@@ -100,7 +102,7 @@ void Init()
         #ifdef _WIN32
             DWORD dummy;
             VirtualProtect(GetAlignedCodeMemoryStart(), CodeMemoryAlignedSize, PAGE_EXECUTE_READWRITE, &dummy);
-        #elif defined(APPLE_AARCH64)
+        #elif defined(APPLE_AARCH64) || defined(__NetBSD__) || defined(__OpenBSD__)
             // Apple aarch64 always uses dynamic allocation
         #else
             mprotect(GetAlignedCodeMemoryStart(), CodeMemoryAlignedSize, PROT_EXEC | PROT_READ | PROT_WRITE);
diff --git a/src/ARMJIT_Memory.cpp b/src/ARMJIT_Memory.cpp
index cc971df014..9c509f2e1e 100644
--- a/src/ARMJIT_Memory.cpp
+++ b/src/ARMJIT_Memory.cpp
@@ -83,6 +83,8 @@
         #define CONTEXT_PC uc_mcontext.mc_rip
     #elif defined(__NetBSD__)
         #define CONTEXT_PC uc_mcontext.__gregs[_REG_RIP]
+    #elif defined(__OpenBSD__)
+        #define CONTEXT_PC sc_rip
     #endif
 #elif defined(__aarch64__)
     #if defined(_WIN32)
@@ -95,6 +97,8 @@
         #define CONTEXT_PC uc_mcontext.mc_gpregs.gp_elr
     #elif defined(__NetBSD__)
         #define CONTEXT_PC uc_mcontext.__gregs[_REG_PC]
+    #elif defined(__OpenBSD__)
+        #define CONTEXT_PC sc_elr
     #endif
 #endif
 
@@ -746,7 +750,7 @@ bool ARMJIT_Memory::IsFastMemSupported()
 #ifdef WATERBOX
     return false;
 #endif
-#ifdef __APPLE__
+#if defined(__APPLE__) || defined(__OpenBSD__)
     return false;
 #else
     static bool initialised = false;
@@ -837,7 +841,11 @@ bool ARMJIT_Memory::FaultHandler(FaultDescription& faultDesc, melonDS::NDS& nds)
             rewriteToSlowPath = !nds.JIT.Memory.MapAtAddress(faultDesc.EmulatedFaultAddr);
 
         if (rewriteToSlowPath)
+        {
+            nds.JIT.JitEnableWrite();
             faultDesc.FaultPC = nds.JIT.JITCompiler.RewriteMemAccess(faultDesc.FaultPC);
+            nds.JIT.JitEnableExecute();
+        }
 
         return true;
     }
diff --git a/src/ARMJIT_x64/ARMJIT_Compiler.cpp b/src/ARMJIT_x64/ARMJIT_Compiler.cpp
index b52344307b..976b151d49 100644
--- a/src/ARMJIT_x64/ARMJIT_Compiler.cpp
+++ b/src/ARMJIT_x64/ARMJIT_Compiler.cpp
@@ -228,6 +228,8 @@ Compiler::Compiler(melonDS::NDS& nds) : XEmitter(), NDS(nds)
     ARMJIT_Global::Init();
 
     CodeMemBase = static_cast<u8*>(ARMJIT_Global::AllocateCodeMem());
+    nds.JIT.JitEnableWrite();
+
     CodeMemSize = ARMJIT_Global::CodeMemorySliceSize;
 
     ResetStart = CodeMemBase;
diff --git a/src/ARMJIT_x64/ARMJIT_Linkage.S b/src/ARMJIT_x64/ARMJIT_Linkage.S
index 838e0515ce..96968d6a10 100644
--- a/src/ARMJIT_x64/ARMJIT_Linkage.S
+++ b/src/ARMJIT_x64/ARMJIT_Linkage.S
@@ -20,6 +20,11 @@
 
 #include "ARMJIT_Offsets.h"
 
+// Fixes a stupid bug on clang-cl builds (defines _WIN64 but not _MSC_VER (this one is correct) or WIN64 (this one is MinGW weirdness))
+#if !defined(WIN64) && defined(_WIN64)
+#define WIN64 _WIN64
+#endif
+
 .text
 
 #define RCPU rbp
@@ -105,7 +110,7 @@ ARM_Ret:
 
     ret
 
-#if !defined(__APPLE__) && !defined(WIN64)
+#if !defined(__APPLE__) && !defined(WIN64) && !defined(_MSC_VER)
 .section .note.GNU-stack,"",@progbits
 #endif
 
diff --git a/src/Args.h b/src/Args.h
index cc1fc86b4a..7a54175672 100644
--- a/src/Args.h
+++ b/src/Args.h
@@ -106,6 +106,7 @@ struct NDSArgs
 
     AudioBitDepth BitDepth = AudioBitDepth::Auto;
     AudioInterpolation Interpolation = AudioInterpolation::None;
+    double OutputSampleRate = 48000.f;
 
     /// How the GDB stub should be handled.
     /// Defaults to disabled.
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 15daa2e1fc..e85dd42958 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -4,6 +4,7 @@ include(FixInterfaceIncludes)
 
 add_library(core STATIC
     ARCodeFile.cpp
+    ARDatabaseDAT.cpp
     AREngine.cpp
     ARM.cpp
     ARM_InstrTable.h
@@ -72,7 +73,9 @@ add_library(core STATIC
 
     sha1/sha1.c
     tiny-AES-c/aes.c
-    xxhash/xxhash.c)
+    xxhash/xxhash.c
+
+    blip-buf/blip_buf.c)
 
 if (ENABLE_GDBSTUB)
     message(NOTICE "Enabling GDB stub")
@@ -115,6 +118,10 @@ if (ENABLE_JIT)
         target_link_libraries(core PRIVATE onecore)
     endif()
 
+    if (CMAKE_SYSTEM_NAME STREQUAL "OpenBSD")
+        target_link_options(core PUBLIC -Wl,-z,wxneeded)
+    endif()
+
     if (ARCHITECTURE STREQUAL x86_64)
         target_sources(core PRIVATE
             dolphin/x64ABI.cpp
@@ -178,6 +185,10 @@ if (NOT MSVC)
     # (which is fundamentally non-portable, so this is fine)
 endif()
 
+# Fix xxhash compile errors when using -Og
+# See https://github.com/Cyan4973/xxHash/blob/136cc1f8fe4d5ea62a7c16c8424d4fa5158f6d68/README.md?plain=1#L159-L163
+target_compile_definitions(core PRIVATE $<$<CONFIG:DEBUG>:XXH_NO_INLINE_HINTS=1>)
+
 find_library(m MATH_LIBRARY)
 
 if (MATH_LIBRARY)
@@ -195,7 +206,11 @@ endif()
 
 if (WIN32)
     target_link_libraries(core PRIVATE ole32 comctl32 wsock32 ws2_32)
-    target_compile_definitions(core PUBLIC WIN32_LEAN_AND_MEAN NOMINMAX)
+
+    if (NOT MINGW)
+        find_path(DIRENT_INCLUDE_DIRS "dirent.h")
+        target_include_directories(core PUBLIC ${DIRENT_INCLUDE_DIRS})
+    endif()
 elseif(NOT APPLE AND NOT HAIKU)
     check_library_exists(rt shm_open "" NEED_LIBRT)
     if (NEED_LIBRT)
diff --git a/src/DSi.cpp b/src/DSi.cpp
index 1a32932090..dafbbe927b 100644
--- a/src/DSi.cpp
+++ b/src/DSi.cpp
@@ -421,6 +421,8 @@ void DSi::SetupDirectBoot()
     u32 cartid = NDSCartSlot.GetCart()->ID();
     DSi_TSC* tsc = (DSi_TSC*)SPI.GetTSC();
 
+    // TODO properly setup SCFG_EXT
+
     // TODO: add controls for forcing DS or DSi mode?
     if (!(header.UnitCode & 0x02))
         dsmode = true;
@@ -448,6 +450,9 @@ void DSi::SetupDirectBoot()
 
         tsc->SetMode(0x00);
         Set_SCFG_Clock9(0x0000);
+
+        SCFG_EXT[0] &= ~0xC000;
+        ApplyNewRAMSize(0);
     }
     else
     {
@@ -599,8 +604,6 @@ void DSi::SetupDirectBoot()
         ARM9Write8(0x02FFFDFB, 0x01);
     }
 
-    // TODO: for DS-mode ROMs, switch RAM size here
-
     u32 arm9start = 0;
 
     // load the ARM9 secure area
@@ -1134,12 +1137,6 @@ void DSi::MapNWRAM_B(u32 num, u8 val)
     u8 oldval = (MBK[0][mbkn] >> mbks) & 0xFF;
     if (oldval == val) return;
 
-    // TODO REMOVE ME
-    if ((val&3)==2 || (val&3)==3)
-    {
-        //debug(0);
-    }
-
     JIT.Memory.RemapNWRAM(1);
 
     MBK[0][mbkn] &= ~(0xFF << mbks);
@@ -1317,6 +1314,19 @@ void DSi::ApplyNewRAMSize(u32 size)
         Log(LogLevel::Debug, "RAM: 16MB\n");
         break;
     }
+
+    // mirror the RAM size setting to the ARM7 register
+    SCFG_EXT[1] &= ~0xC000;
+    SCFG_EXT[1] |= (size << 14);
+}
+
+void DSi::CheckDSiLoaderHack()
+{
+    if (!(SCFG_EXT[0] & SCFG_DSiLoaderHack))
+        return;
+
+    SCFG_EXT[0] &= ~SCFG_DSiLoaderHack;
+    ApplyNewRAMSize((SCFG_EXT[0] >> 14) & 0x3);
 }
 
 
@@ -2591,30 +2601,51 @@ void DSi::ARM9IOWrite32(u32 addr, u32 val)
 
             SCFG_EXT[0] &= ~0x8007F19F;
             SCFG_EXT[0] |= (val & 0x8007F19F);
-            SCFG_EXT[1] &= ~0x0000F080;
-            SCFG_EXT[1] |= (val & 0x0000F080);
+            SCFG_EXT[1] &= ~0x00003080;
+            SCFG_EXT[1] |= (val & 0x00003080);
             Log(LogLevel::Debug, "SCFG_EXT = %08X / %08X (val9 %08X)\n", SCFG_EXT[0], SCFG_EXT[1], val);
-            /*switch ((SCFG_EXT[0] >> 14) & 0x3)
+            
+            if (newram != oldram)
             {
-            case 0:
-            case 1:
-                NDS::MainRAMMask = 0x3FFFFF;
-                printf("RAM: 4MB\n");
-                //baziderp=true;
-                break;
-            case 2:
-            case 3: // TODO: debug console w/ 32MB?
-                NDS::MainRAMMask = 0xFFFFFF;
-                printf("RAM: 16MB\n");
-                break;
-            }*/
-            // HAX!!
-            // a change to the RAM size setting is supposed to apply immediately (it does so on hardware)
-            // however, doing so will cause DS-mode app startup to break, because the change happens while the ARM7
-            // is still busy clearing/relocating shit
-            //if (newram != oldram)
-            //    NDS::ScheduleEvent(NDS::Event_DSi_RAMSizeChange, false, 512*512*512, ApplyNewRAMSize, newram);
-            Log(LogLevel::Debug, "from %08X, ARM7 %08X, %08X\n", NDS::GetPC(0), NDS::GetPC(1), ARM7.R[1]);
+                bool isDSiLoader = (ARM9.R[15] == 0x023FEED0) && (IPCSync9 == 0x0505);
+                if (isDSiLoader)
+                {
+                    /*
+                     -- DSI LOADER HACK --
+
+                     When loading a title, the DSi loader does the following: (after other steps)
+                     * ARM7 sends IPCSYNC=5 and waits for ARM9
+                     * ARM9 sends IPCSYNC=5
+                     * ARM7 clears/moves a bunch of memory regions, then sends IPCSYNC=0
+                     * ARM9 clears DTCM (clear/move lists are empty, so that's it)
+                     * ARM9 changes RAM size if required (when loading a DS title)
+                     * ARM9 waits for ARM7, then sends IPCSYNC=0
+                     * the title is booted
+
+                     When loading a DS game, the loader crashes, because the ARM9 tries to change the RAM size
+                     while the ARM7 is still clearing/moving RAM, causing it to overwrite the ARM9 code.
+
+                     The ARM9 has caches off. The ARM7 has priority over main RAM.
+                     On hardware, the loader works due to main RAM contention: the ARM7-side memory clearing/moving
+                     slows down the ARM9 a lot, because the two are actively competing for main RAM access.
+                     Thus, the ARM9 is only able to change the RAM size after the ARM7 has finished its work.
+                     If EXMEMCNT is changed to give priority to the ARM9, the loader will crash.
+
+                     Main RAM contention would be incredibly difficult to emulate without a cycle-accurate emulator
+                     (which melonDS is not).
+                     A possible workaround might be to count main RAM accesses over a time slice and calculate a
+                     penalty to apply to either ARM9 or ARM7. This would be gross and have potential to cause new
+                     timing issues unless it is calibrated precisely.
+
+                     In the meantime, this hack will take care of the DSi loader's shoddy programming, hopefully
+                     with no side effects.
+                     */
+
+                    SCFG_EXT[0] |= SCFG_DSiLoaderHack;
+                }
+                else
+                    ApplyNewRAMSize(newram);
+            }
 
             UpdateVRAMTimings();
         }
@@ -3025,6 +3056,12 @@ void DSi::ARM7IOWrite16(u32 addr, u16 val)
     assert(ConsoleType == 1);
     switch (addr)
     {
+        case 0x04000180:
+            // DSi loader hack hook
+            if ((val & 0x0F00) == 0x0000)
+                CheckDSiLoaderHack();
+            return NDS::ARM7IOWrite16(addr, val);
+
         case 0x04000218: NDS::IE2 = (val & 0x7FF7); NDS::UpdateIRQ(1); return;
         case 0x0400021C: NDS::IF2 &= ~(val & 0x7FF7); NDS::UpdateIRQ(1); return;
 
@@ -3139,6 +3176,12 @@ void DSi::ARM7IOWrite32(u32 addr, u32 val)
     assert(ConsoleType == 1);
     switch (addr)
     {
+    case 0x04000180:
+        // DSi loader hack hook
+        if ((val & 0x0F00) == 0x0000)
+            CheckDSiLoaderHack();
+        return NDS::ARM7IOWrite32(addr, val);
+
     case 0x04000218: NDS::IE2 = (val & 0x7FF7); NDS::UpdateIRQ(1); return;
     case 0x0400021C: NDS::IF2 &= ~(val & 0x7FF7); NDS::UpdateIRQ(1); return;
 
diff --git a/src/DSi.h b/src/DSi.h
index 250d5d3c6f..416453cd4a 100644
--- a/src/DSi.h
+++ b/src/DSi.h
@@ -42,6 +42,11 @@ namespace DSi_NAND
     class NANDImage;
 }
 
+enum
+{
+    SCFG_DSiLoaderHack = (1<<30),
+};
+
 class DSi final : public NDS
 {
 protected:
@@ -188,6 +193,7 @@ class DSi final : public NDS
     void Set_SCFG_MC(u32 val);
     void DecryptModcryptArea(u32 offset, u32 size, const u8* iv);
     void ApplyNewRAMSize(u32 size);
+    void CheckDSiLoaderHack();
 };
 
 }
diff --git a/src/DSi_Camera.cpp b/src/DSi_Camera.cpp
index 55c9c14427..e66bea3eac 100644
--- a/src/DSi_Camera.cpp
+++ b/src/DSi_Camera.cpp
@@ -319,7 +319,7 @@ u16 DSi_CamModule::Read16(u32 addr)
 {
     switch (addr)
     {
-    case 0x04004200: return ModuleCnt;
+    case 0x04004200: return ModuleCnt | (1<<7);
     case 0x04004202: return Cnt | (Transferring ? (1<<15) : 0);
     }
 
@@ -366,21 +366,23 @@ void DSi_CamModule::Write16(u32 addr, u16 val)
     {
     case 0x04004200:
         {
+            if (IsTransferring()) return;
+
             u16 oldcnt = ModuleCnt;
             ModuleCnt = val;
 
-            if ((ModuleCnt & (1<<1)) && !(oldcnt & (1<<1)))
-            {
-                // reset shit to zero
-                // CHECKME
+            // bit 1 doesn't seem to actually do anything?
+            // bit 5 controls camera reset
 
-                Cnt = 0;
-                Transferring = false;
+            if ((!(ModuleCnt & (1<<5))) && (oldcnt & (1<<5)))
+            {
+                Camera0->Reset();
+                Camera1->Reset();
             }
-
-            if ((ModuleCnt & (1<<5)) && !(oldcnt & (1<<5)))
+            else if ((ModuleCnt & (1<<5)) && (!(oldcnt & (1<<5))))
             {
-                // TODO: reset I2C??
+                Camera0->ReleaseReset();
+                Camera1->ReleaseReset();
             }
         }
         return;
@@ -483,6 +485,8 @@ void DSi_Camera::Reset()
 {
     Platform::Camera_Stop(Num, DSi.UserData);
 
+    ResetHeld = true;
+
     DataPos = 0;
     RegAddr = 0;
     RegData = 0;
@@ -505,6 +509,11 @@ void DSi_Camera::Reset()
     memset(FrameBuffer, 0, (640*480/2)*sizeof(u32));
 }
 
+void DSi_Camera::ReleaseReset()
+{
+    ResetHeld = false;
+}
+
 void DSi_Camera::Stop()
 {
     Platform::Camera_Stop(Num, DSi.UserData);
@@ -625,6 +634,8 @@ void DSi_Camera::Acquire()
 
 u8 DSi_Camera::Read(bool last)
 {
+    if (ResetHeld) return 0xFF;
+
     u8 ret;
 
     if (DataPos & 0x1)
@@ -646,6 +657,8 @@ u8 DSi_Camera::Read(bool last)
 
 void DSi_Camera::Write(u8 val, bool last)
 {
+    if (ResetHeld) return;
+
     if (DataPos < 2)
     {
         if (DataPos == 0)
diff --git a/src/DSi_Camera.h b/src/DSi_Camera.h
index 4e93e0a8ec..01c51218ff 100644
--- a/src/DSi_Camera.h
+++ b/src/DSi_Camera.h
@@ -37,6 +37,7 @@ class DSi_Camera : public DSi_I2CDevice
     void DoSavestate(Savestate* file) override;
 
     void Reset() override;
+    void ReleaseReset();
     void Stop();
     bool IsActivated() const;
 
@@ -55,6 +56,8 @@ class DSi_Camera : public DSi_I2CDevice
     u32 Num;
 
 private:
+    bool ResetHeld;
+
     u32 DataPos;
     u32 RegAddr;
     u16 RegData;
diff --git a/src/DSi_SD.cpp b/src/DSi_SD.cpp
index 08f4172725..aab460b776 100644
--- a/src/DSi_SD.cpp
+++ b/src/DSi_SD.cpp
@@ -287,8 +287,10 @@ void DSi_SDHost::SetCardIRQ()
     u16 oldflags = CardIRQStatus & ~CardIRQMask;
     DSi_SDDevice* dev = Ports[PortSelect & 0x1].get();
 
-    if (dev->IRQ) CardIRQStatus |=  (1<<0);
-    else          CardIRQStatus &= ~(1<<0);
+    if (dev && dev->IRQ)
+        CardIRQStatus |=  (1<<0);
+    else
+        CardIRQStatus &= ~(1<<0);
 
     u16 newflags = CardIRQStatus & ~CardIRQMask;
 
diff --git a/src/FATStorage.cpp b/src/FATStorage.cpp
index 7a870d55a3..c93e90b99a 100644
--- a/src/FATStorage.cpp
+++ b/src/FATStorage.cpp
@@ -192,16 +192,14 @@ u64 FATStorage::GetSectorCount() const
 ff_disk_read_cb FATStorage::FF_ReadStorage() const noexcept
 {
     return [this](BYTE* buf, LBA_t sector, UINT num) {
-        u32 res = ReadSectorsInternal(File, FileSize, sector, num, buf);
-        return res / 0x200;
+        return ReadSectorsInternal(File, FileSize, sector, num, buf);
     };
 }
 
 ff_disk_write_cb FATStorage::FF_WriteStorage() const noexcept
 {
     return [this](const BYTE* buf, LBA_t sector, UINT num) {
-        u32 res = WriteSectorsInternal(File, FileSize, sector, num, buf);
-        return res / 0x200;
+        return WriteSectorsInternal(File, FileSize, sector, num, buf);
     };
 }
 
@@ -1164,4 +1162,4 @@ bool FATStorage::Save()
     return true;
 }
 
-}
+}
\ No newline at end of file
diff --git a/src/FIFO.h b/src/FIFO.h
index 159752f80c..62ca883eb5 100644
--- a/src/FIFO.h
+++ b/src/FIFO.h
@@ -225,7 +225,7 @@ class RingBuffer
             u32 part1 = Size - WritePos;
             memcpy(&Buffer[WritePos], data, part1);
             if (len > part1)
-                memcpy(Buffer, &((u8*)data)[part1], len - part1);
+                memcpy(Buffer, &((const u8*)data)[part1], len - part1);
             WritePos = len - part1;
         }
         else
diff --git a/src/GPU3D_Compute.cpp b/src/GPU3D_Compute.cpp
index 3e89f02a55..eef2bf0f23 100644
--- a/src/GPU3D_Compute.cpp
+++ b/src/GPU3D_Compute.cpp
@@ -19,6 +19,9 @@
 #include "GPU3D_Compute.h"
 
 #include <assert.h>
+#include <algorithm>
+
+#include "Utils.h"
 
 #include "OpenGLSupport.h"
 
@@ -50,6 +53,14 @@ bool ComputeRenderer::CompileShader(GLuint& shader, const std::string& source, c
     shaderSource += std::to_string(ScreenHeight);
     shaderSource += "\n#define MaxWorkTiles ";
     shaderSource += std::to_string(MaxWorkTiles);
+    shaderSource += "\n#define TileSize ";
+    shaderSource += std::to_string(TileSize);
+    shaderSource += "\nconst int CoarseTileCountY = ";
+    shaderSource += std::to_string(CoarseTileCountY) + ";";
+    shaderSource += "\n#define CoarseTileArea ";
+    shaderSource += std::to_string(CoarseTileArea);
+    shaderSource += "\n#define ClearCoarseBinMaskLocalSize ";
+    shaderSource += std::to_string(ClearCoarseBinMaskLocalSize);
 
     shaderSource += ComputeRendererShaders::Common;
     shaderSource += source;
@@ -297,6 +308,8 @@ void ComputeRenderer::Reset(GPU& gpu)
 
 void ComputeRenderer::SetRenderSettings(int scale, bool highResolutionCoordinates)
 {
+    u8 TileScale;
+
     CurGLCompositor.SetScaleFactor(scale);
 
     if (ScaleFactor != -1)
@@ -310,6 +323,22 @@ void ComputeRenderer::SetRenderSettings(int scale, bool highResolutionCoordinate
     ScreenWidth = 256 * ScaleFactor;
     ScreenHeight = 192 * ScaleFactor;
 
+    //Starting at 4.5x we want to double TileSize every time scale doubles
+    TileScale = 2 * ScaleFactor / 9;
+    TileScale = GetMSBit(TileScale);
+    TileScale <<= 1;
+    TileScale += TileScale == 0;
+
+    std::printf("Scale: %d\n", ScaleFactor);
+    std::printf("TileScale: %d\n", TileScale);
+    
+    TileSize = std::min(8 * TileScale, 32);
+    CoarseTileCountY = TileSize < 32 ? 4 : 6;
+    ClearCoarseBinMaskLocalSize = TileSize < 32 ? 64 : 48;
+    CoarseTileArea = CoarseTileCountX * CoarseTileCountY;
+    CoarseTileW = CoarseTileCountX * TileSize;
+    CoarseTileH = CoarseTileCountY * TileSize;
+
     TilesPerLine = ScreenWidth/TileSize;
     TileLines = ScreenHeight/TileSize;
 
@@ -918,7 +947,7 @@ void ComputeRenderer::RenderFrame(GPU& gpu)
     glBindBufferBase(GL_UNIFORM_BUFFER, 0, MetaUniformMemory);
 
     glUseProgram(ShaderClearCoarseBinMask);
-    glDispatchCompute(TilesPerLine*TileLines/32, 1, 1);
+    glDispatchCompute(TilesPerLine*TileLines/ClearCoarseBinMaskLocalSize, 1, 1);
 
     bool wbuffer = false;
     if (numYSpans > 0)
@@ -932,23 +961,23 @@ void ComputeRenderer::RenderFrame(GPU& gpu)
         glBindImageTexture(0, YSpanIndicesTexture, 0, GL_FALSE, 0, GL_READ_ONLY, GL_RGBA16UI);
         glUseProgram(ShaderInterpXSpans[wbuffer]);
         glDispatchCompute((numSetupIndices + 31) / 32, 1, 1);
-        glMemoryBarrier(GL_SHADER_STORAGE_BUFFER);
+        glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT | GL_COMMAND_BARRIER_BIT);
 
         // bin polygons
         glUseProgram(ShaderBinCombined);
         glDispatchCompute(((gpu.GPU3D.RenderNumPolygons + 31) / 32), ScreenWidth/CoarseTileW, ScreenHeight/CoarseTileH);
-        glMemoryBarrier(GL_SHADER_STORAGE_BUFFER);
+        glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT | GL_COMMAND_BARRIER_BIT);
 
         // calculate list offsets
         glUseProgram(ShaderCalculateWorkListOffset);
         glDispatchCompute((numVariants + 31) / 32, 1, 1);
-        glMemoryBarrier(GL_SHADER_STORAGE_BUFFER);
+        glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT | GL_COMMAND_BARRIER_BIT);
 
         // sort shader work
         glUseProgram(ShaderSortWork);
         glBindBuffer(GL_DISPATCH_INDIRECT_BUFFER, BinResultMemory);
         glDispatchComputeIndirect(offsetof(BinResultHeader, SortWorkWorkCount));
-        glMemoryBarrier(GL_SHADER_STORAGE_BUFFER);
+        glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT | GL_COMMAND_BARRIER_BIT);
 
         glActiveTexture(GL_TEXTURE0);
 
@@ -1134,4 +1163,4 @@ void ComputeRenderer::Stop(const GPU &gpu)
     CurGLCompositor.Stop(gpu);
 }
 
-}
\ No newline at end of file
+}
diff --git a/src/GPU3D_Compute.h b/src/GPU3D_Compute.h
index 0429f3aef7..06d60311f1 100644
--- a/src/GPU3D_Compute.h
+++ b/src/GPU3D_Compute.h
@@ -163,11 +163,13 @@ class ComputeRenderer : public Renderer3D
         float TextureLayer;
     };
 
-    static constexpr int TileSize = 8;
+    int TileSize;
     static constexpr int CoarseTileCountX = 8;
-    static constexpr int CoarseTileCountY = 4;
-    static constexpr int CoarseTileW = CoarseTileCountX * TileSize;
-    static constexpr int CoarseTileH = CoarseTileCountY * TileSize;
+    int CoarseTileCountY;
+    int CoarseTileArea;
+    int CoarseTileW;
+    int CoarseTileH;
+    int ClearCoarseBinMaskLocalSize;
 
     static constexpr int BinStride = 2048/32;
     static constexpr int CoarseBinStride = BinStride/32;
diff --git a/src/GPU3D_Compute_shaders.h b/src/GPU3D_Compute_shaders.h
index 8b8f6c5a00..dab28a6ea4 100644
--- a/src/GPU3D_Compute_shaders.h
+++ b/src/GPU3D_Compute_shaders.h
@@ -339,9 +339,7 @@ const uint ResultAttrStart = ResultDepthStart+ScreenWidth*ScreenHeight*2;
 
 const char* Common = R"(
 
-#define TileSize 8
 const int CoarseTileCountX = 8;
-const int CoarseTileCountY = 4;
 const int CoarseTileW = (CoarseTileCountX * TileSize);
 const int CoarseTileH = (CoarseTileCountY * TileSize);
 
@@ -848,7 +846,7 @@ void main()
 
 const std::string ClearCoarseBinMask =
     BinningBuffer + R"(
-layout (local_size_x = 32) in;
+layout (local_size_x = ClearCoarseBinMaskLocalSize) in;
 
 void main()
 {
@@ -864,7 +862,7 @@ const std::string BinCombined =
     XSpanSetupBuffer +
     WorkDescBuffer + R"(
 
-layout (local_size_x = 32) in;
+layout (local_size_x = CoarseTileArea) in;
 
 bool BinPolygon(Polygon polygon, ivec2 topLeft, ivec2 botRight)
 {
diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp
index 93be00537e..e2d7f617fc 100644
--- a/src/GPU3D_Soft.cpp
+++ b/src/GPU3D_Soft.cpp
@@ -585,7 +585,7 @@ void SoftRenderer::SetupPolygonLeftEdge(SoftRenderer::RendererPolygon* rp, s32 y
 
     rp->XL = rp->SlopeL.Setup(polygon->Vertices[rp->CurVL]->FinalPosition[0], polygon->Vertices[rp->NextVL]->FinalPosition[0],
                               polygon->Vertices[rp->CurVL]->FinalPosition[1], polygon->Vertices[rp->NextVL]->FinalPosition[1],
-                              polygon->FinalW[rp->CurVL], polygon->FinalW[rp->NextVL], y);
+                              polygon->FinalW[rp->CurVL], polygon->FinalW[rp->NextVL], y, polygon->WBuffer);
 }
 
 void SoftRenderer::SetupPolygonRightEdge(SoftRenderer::RendererPolygon* rp, s32 y) const
@@ -612,7 +612,7 @@ void SoftRenderer::SetupPolygonRightEdge(SoftRenderer::RendererPolygon* rp, s32
 
     rp->XR = rp->SlopeR.Setup(polygon->Vertices[rp->CurVR]->FinalPosition[0], polygon->Vertices[rp->NextVR]->FinalPosition[0],
                               polygon->Vertices[rp->CurVR]->FinalPosition[1], polygon->Vertices[rp->NextVR]->FinalPosition[1],
-                              polygon->FinalW[rp->CurVR], polygon->FinalW[rp->NextVR], y);
+                              polygon->FinalW[rp->CurVR], polygon->FinalW[rp->NextVR], y, polygon->WBuffer);
 }
 
 void SoftRenderer::SetupPolygon(SoftRenderer::RendererPolygon* rp, Polygon* polygon) const
@@ -658,8 +658,8 @@ void SoftRenderer::SetupPolygon(SoftRenderer::RendererPolygon* rp, Polygon* poly
         rp->CurVL = vtop; rp->NextVL = vtop;
         rp->CurVR = vbot; rp->NextVR = vbot;
 
-        rp->XL = rp->SlopeL.SetupDummy(polygon->Vertices[rp->CurVL]->FinalPosition[0]);
-        rp->XR = rp->SlopeR.SetupDummy(polygon->Vertices[rp->CurVR]->FinalPosition[0]);
+        rp->XL = rp->SlopeL.SetupDummy(polygon->Vertices[rp->CurVL]->FinalPosition[0], polygon->WBuffer);
+        rp->XR = rp->SlopeR.SetupDummy(polygon->Vertices[rp->CurVR]->FinalPosition[0], polygon->WBuffer);
     }
     else
     {
@@ -718,8 +718,8 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
     s32 wl = rp->SlopeL.Interp.Interpolate(polygon->FinalW[rp->CurVL], polygon->FinalW[rp->NextVL]);
     s32 wr = rp->SlopeR.Interp.Interpolate(polygon->FinalW[rp->CurVR], polygon->FinalW[rp->NextVR]);
 
-    s32 zl = rp->SlopeL.Interp.InterpolateZ(polygon->FinalZ[rp->CurVL], polygon->FinalZ[rp->NextVL], polygon->WBuffer);
-    s32 zr = rp->SlopeR.Interp.InterpolateZ(polygon->FinalZ[rp->CurVR], polygon->FinalZ[rp->NextVR], polygon->WBuffer);
+    s32 zl = rp->SlopeL.Interp.InterpolateZ(polygon->FinalZ[rp->CurVL], polygon->FinalZ[rp->NextVL]);
+    s32 zr = rp->SlopeR.Interp.InterpolateZ(polygon->FinalZ[rp->CurVR], polygon->FinalZ[rp->NextVR]);
 
     // right vertical edges are pushed 1px to the left as long as either:
     // the left edge slope is not 0, or the span is not 0 pixels wide, and it is not at the leftmost pixel of the screen
@@ -804,7 +804,7 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
     int edge;
 
     s32 x = xstart;
-    Interpolator<0> interpX(xstart, xend+1, wl, wr);
+    Interpolator<0> interpX(xstart, xend+1, wl, wr, polygon->WBuffer);
 
     if (x < 0) x = 0;
     s32 xlimit;
@@ -826,7 +826,7 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
 
         interpX.SetX(x);
 
-        s32 z = interpX.InterpolateZ(zl, zr, polygon->WBuffer);
+        s32 z = interpX.InterpolateZ(zl, zr);
         u32 dstattr = AttrBuffer[pixeladdr];
 
         if (!fnDepthTest(DepthBuffer[pixeladdr], z, dstattr))
@@ -852,7 +852,7 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
 
         interpX.SetX(x);
 
-        s32 z = interpX.InterpolateZ(zl, zr, polygon->WBuffer);
+        s32 z = interpX.InterpolateZ(zl, zr);
         u32 dstattr = AttrBuffer[pixeladdr];
 
         if (!fnDepthTest(DepthBuffer[pixeladdr], z, dstattr))
@@ -878,7 +878,7 @@ void SoftRenderer::RenderShadowMaskScanline(const GPU3D& gpu3d, RendererPolygon*
 
         interpX.SetX(x);
 
-        s32 z = interpX.InterpolateZ(zl, zr, polygon->WBuffer);
+        s32 z = interpX.InterpolateZ(zl, zr);
         u32 dstattr = AttrBuffer[pixeladdr];
 
         if (!fnDepthTest(DepthBuffer[pixeladdr], z, dstattr))
@@ -943,8 +943,8 @@ void SoftRenderer::RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s3
     s32 wl = rp->SlopeL.Interp.Interpolate(polygon->FinalW[rp->CurVL], polygon->FinalW[rp->NextVL]);
     s32 wr = rp->SlopeR.Interp.Interpolate(polygon->FinalW[rp->CurVR], polygon->FinalW[rp->NextVR]);
 
-    s32 zl = rp->SlopeL.Interp.InterpolateZ(polygon->FinalZ[rp->CurVL], polygon->FinalZ[rp->NextVL], polygon->WBuffer);
-    s32 zr = rp->SlopeR.Interp.InterpolateZ(polygon->FinalZ[rp->CurVR], polygon->FinalZ[rp->NextVR], polygon->WBuffer);
+    s32 zl = rp->SlopeL.Interp.InterpolateZ(polygon->FinalZ[rp->CurVL], polygon->FinalZ[rp->NextVL]);
+    s32 zr = rp->SlopeR.Interp.InterpolateZ(polygon->FinalZ[rp->CurVR], polygon->FinalZ[rp->NextVR]);
 
     // right vertical edges are pushed 1px to the left as long as either:
     // the left edge slope is not 0, or the span is not 0 pixels wide, and it is not at the leftmost pixel of the screen
@@ -1054,7 +1054,7 @@ void SoftRenderer::RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s3
     int edge;
 
     s32 x = xstart;
-    Interpolator<0> interpX(xstart, xend+1, wl, wr);
+    Interpolator<0> interpX(xstart, xend+1, wl, wr, polygon->WBuffer);
 
     if (x < 0) x = 0;
     s32 xlimit;
@@ -1093,7 +1093,7 @@ void SoftRenderer::RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s3
 
         interpX.SetX(x);
 
-        s32 z = interpX.InterpolateZ(zl, zr, polygon->WBuffer);
+        s32 z = interpX.InterpolateZ(zl, zr);
 
         // if depth test against the topmost pixel fails, test
         // against the pixel underneath
@@ -1189,7 +1189,7 @@ void SoftRenderer::RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s3
 
         interpX.SetX(x);
 
-        s32 z = interpX.InterpolateZ(zl, zr, polygon->WBuffer);
+        s32 z = interpX.InterpolateZ(zl, zr);
 
         // if depth test against the topmost pixel fails, test
         // against the pixel underneath
@@ -1281,7 +1281,7 @@ void SoftRenderer::RenderPolygonScanline(const GPU& gpu, RendererPolygon* rp, s3
 
         interpX.SetX(x);
 
-        s32 z = interpX.InterpolateZ(zl, zr, polygon->WBuffer);
+        s32 z = interpX.InterpolateZ(zl, zr);
 
         // if depth test against the topmost pixel fails, test
         // against the pixel underneath
diff --git a/src/GPU3D_Soft.h b/src/GPU3D_Soft.h
index 886df84832..7520f2c3d6 100644
--- a/src/GPU3D_Soft.h
+++ b/src/GPU3D_Soft.h
@@ -69,16 +69,17 @@ class SoftRenderer : public Renderer3D
     {
     public:
         constexpr Interpolator() {}
-        constexpr Interpolator(s32 x0, s32 x1, s32 w0, s32 w1)
+        constexpr Interpolator(s32 x0, s32 x1, s32 w0, s32 w1, bool wbuffer)
         {
-            Setup(x0, x1, w0, w1);
+            Setup(x0, x1, w0, w1, wbuffer);
         }
 
-        constexpr void Setup(s32 x0, s32 x1, s32 w0, s32 w1)
+        constexpr void Setup(s32 x0, s32 x1, s32 w0, s32 w1, bool wbuffer)
         {
             this->x0 = x0;
             this->x1 = x1;
             this->xdiff = x1 - x0;
+            this->wbuffer = wbuffer;
 
             // calculate reciprocal for Z interpolation
             // TODO eventually: use a faster reciprocal function?
@@ -123,7 +124,7 @@ class SoftRenderer : public Renderer3D
         {
             x -= x0;
             this->x = x;
-            if (xdiff != 0 && !linear)
+            if ((xdiff != 0) && ((!linear) || wbuffer))
             {
                 u32 num = (x * w0n) << shift;
                 u32 den = (x * w0d) + ((xdiff-x) * w1d);
@@ -157,7 +158,7 @@ class SoftRenderer : public Renderer3D
             }
         }
 
-        constexpr s32 InterpolateZ(s32 z0, s32 z1, bool wbuffer) const
+        constexpr s32 InterpolateZ(s32 z0, s32 z1) const
         {
             if (xdiff == 0 || z0 == z1) return z0;
 
@@ -212,6 +213,7 @@ class SoftRenderer : public Renderer3D
 
         int shift;
         bool linear;
+        bool wbuffer;
 
         s32 xrecip_z;
         s32 w0n, w0d, w1d;
@@ -226,7 +228,7 @@ class SoftRenderer : public Renderer3D
     public:
         constexpr Slope() {}
 
-        constexpr s32 SetupDummy(s32 x0)
+        constexpr s32 SetupDummy(s32 x0, bool wbuffer)
         {
             dx = 0;
 
@@ -237,7 +239,7 @@ class SoftRenderer : public Renderer3D
             Increment = 0;
             XMajor = false;
 
-            Interp.Setup(0, 0, 0, 0);
+            Interp.Setup(0, 0, 0, 0, wbuffer);
             Interp.SetX(0);
 
             xcov_incr = 0;
@@ -245,7 +247,7 @@ class SoftRenderer : public Renderer3D
             return x0;
         }
 
-        constexpr s32 Setup(s32 x0, s32 x1, s32 y0, s32 y1, s32 w0, s32 w1, s32 y)
+        constexpr s32 Setup(s32 x0, s32 x1, s32 y0, s32 y1, s32 w0, s32 w1, s32 y, bool wbuffer)
         {
             this->x0 = x0;
             this->y = y;
@@ -311,7 +313,7 @@ class SoftRenderer : public Renderer3D
             s32 x = XVal();
 
             int interpoffset = (Increment >= 0x40000) && (side ^ Negative);
-            Interp.Setup(y0-interpoffset, y1-interpoffset, w0, w1);
+            Interp.Setup(y0-interpoffset, y1-interpoffset, w0, w1, wbuffer);
             Interp.SetX(y);
 
             // used for calculating AA coverage
diff --git a/src/Mic.cpp b/src/Mic.cpp
index 31aea02209..a777fad4ed 100644
--- a/src/Mic.cpp
+++ b/src/Mic.cpp
@@ -204,13 +204,13 @@ void Mic::Advance(u32 cycles)
     }
 
     // NDS mic input has no explicit start/stop control
-    // so this counter will turn it off if the TSC AUX input doesn't get sampled for one video frame
+    // so this counter will turn it off if the TSC AUX input doesn't get sampled for two video frames
     // other sources also get similar counters, in case we are dealing with stupid code that keeps
     // turning the mic on and off (hi DSi-mode libnds)
     for (int i = 0; i < 3; i++)
     {
         StopCount[i] += cycles;
-        if (StopCount[i] >= 560190)
+        if (StopCount[i] >= 560190*2)
             DoStop((MicSource)i);
     }
 }
diff --git a/src/NDS.cpp b/src/NDS.cpp
index d32ff08bbf..982f385ff4 100644
--- a/src/NDS.cpp
+++ b/src/NDS.cpp
@@ -95,7 +95,7 @@ NDS::NDS(NDSArgs&& args, int type, void* userdata) noexcept :
     ARM7BIOSNative(CRC32(ARM7BIOS.data(), ARM7BIOS.size()) == ARM7BIOSCRC32),
     ARM9BIOSNative(CRC32(ARM9BIOS.data(), ARM9BIOS.size()) == ARM9BIOSCRC32),
     JIT(*this, args.JIT),
-    SPU(*this, args.BitDepth, args.Interpolation),
+    SPU(*this, args.BitDepth, args.Interpolation, args.OutputSampleRate),
     Mic(*this),
     GPU(*this, std::move(args.Renderer3D)),
     SPI(*this, std::move(args.Firmware)),
@@ -221,7 +221,7 @@ void NDS::SetJITArgs(std::optional<JITArgs> args) noexcept
     }
     else if (args.has_value() != EnableJIT)
     { // Else if we want to turn the JIT off, and it wasn't already off...
-        JIT.ResetBlockCache();
+        JIT.Reset();
     }
 
     EnableJIT = args.has_value();
@@ -895,7 +895,7 @@ void NDS::RunSystemSleep(u64 timestamp)
     for (int i = 0; i < Event_MAX; i++)
     {
         if (!mask) break;
-        if (i == Event_SPU || i == Event_RTC)
+        if (i == Event_RTC)
         {
             if (mask & 0x1)
             {
@@ -905,14 +905,8 @@ void NDS::RunSystemSleep(u64 timestamp)
                 {
                     SchedListMask &= ~(1<<i);
 
-                    u32 param;
-                    if (i == Event_SPU)
-                        param = 1;
-                    else
-                        param = evt.Param;
-
                     EventFunc func = evt.Funcs[evt.FuncID];
-                    func(evt.That, param);
+                    func(evt.That, evt.Param);
                 }
             }
         }
@@ -1066,6 +1060,9 @@ u32 NDS::RunFrame()
         break;
     }
 
+    // Signal to the SPU that it should read out the buffered audio from blip-buf to a buffer for the frontend to consume
+    SPU.EndFrame();
+
     // In the context of TASes, frame count is traditionally the primary measure of emulated time,
     // so it needs to be tracked even if NDS is powered off.
     NumFrames++;
diff --git a/src/NDS.h b/src/NDS.h
index 1b7054b792..e67f01df14 100644
--- a/src/NDS.h
+++ b/src/NDS.h
@@ -520,7 +520,7 @@ class NDS
     void SetGdbArgs(std::optional<GDBArgs> args) noexcept {}
 #endif
 
-private:
+protected:
     void InitTimings();
     u32 SchedListMask;
     u64 SysTimestamp;
diff --git a/src/Platform.h b/src/Platform.h
index 92aee471a0..c83a007d30 100644
--- a/src/Platform.h
+++ b/src/Platform.h
@@ -213,6 +213,9 @@ bool IsEndOfFile(FileHandle* file);
 /// @see fgets
 bool FileReadLine(char* str, int count, FileHandle* file);
 
+/// @see ftell
+u64 FilePosition(FileHandle* file);
+
 /// @see fseek
 bool FileSeek(FileHandle* file, s64 offset, FileSeekOrigin origin);
 
diff --git a/src/SPU.cpp b/src/SPU.cpp
index d510452a0f..3c4aca77ad 100644
--- a/src/SPU.cpp
+++ b/src/SPU.cpp
@@ -26,6 +26,10 @@
 #include "DSi_I2S.h"
 #include "SPU.h"
 
+#include "blip-buf/blip_buf.h"
+
+#define INTERNAL_SAMPLE_RATE 16756991.f
+
 namespace melonDS
 {
 using Platform::Log;
@@ -177,7 +181,7 @@ const std::array<s16, 0x200> InterpSNESGauss = {
     0x513, 0x514, 0x514, 0x515, 0x516, 0x516, 0x517, 0x517, 0x517, 0x518, 0x518, 0x518, 0x518, 0x518, 0x519, 0x519
 };
 
-SPU::SPU(melonDS::NDS& nds, AudioBitDepth bitdepth, AudioInterpolation interpolation) :
+SPU::SPU(melonDS::NDS& nds, AudioBitDepth bitdepth, AudioInterpolation interpolation, double outputSampleRate) :
     NDS(nds),
     Channels {
         SPUChannel(0, nds, interpolation),
@@ -201,34 +205,38 @@ SPU::SPU(melonDS::NDS& nds, AudioBitDepth bitdepth, AudioInterpolation interpola
         SPUCaptureUnit(0, nds),
         SPUCaptureUnit(1, nds),
     },
-    Degrade10Bit(bitdepth == AudioBitDepth::_10Bit || (nds.ConsoleType == 1 && bitdepth == AudioBitDepth::Auto))
+    AudioLock(Platform::Mutex_Create()),
+    Degrade10Bit(bitdepth == AudioBitDepth::_10Bit || (nds.ConsoleType == 1 && bitdepth == AudioBitDepth::Auto)),
+    OutputSampleRate(outputSampleRate),
+    OutputBuffer(nullptr)
 {
-    BlipL = blip_new(4096);
-    BlipR = blip_new(4096);
-    blip_set_rates(BlipL, 16756991, 44100);
-    blip_set_rates(BlipR, 16756991, 44100);
-
     NDS.RegisterEventFuncs(Event_SPU, this, {MakeEventThunk(SPU, Mix)});
 
     ApplyBias = true;
+    Degrade10Bit = false;
+
+    BlipLeft = blip_new(8192);
+    BlipRight = blip_new(8192);
+
+    OutputBufferReadPos = 0;
+    OutputBufferWritePos = 0;
 
     SetSampleRate(AudioSampleRate::_32KHz);
 }
 
 SPU::~SPU()
 {
-    blip_delete(BlipL);
-    blip_delete(BlipR);
+    Platform::Mutex_Free(AudioLock);
+    AudioLock = nullptr;
+    blip_delete(BlipLeft);
+    blip_delete(BlipRight);
+
     NDS.UnregisterEventFuncs(Event_SPU);
 }
 
 void SPU::Reset()
 {
-    blip_clear(BlipL);
-    blip_clear(BlipR);
-    BlipAccumulate = 0;
-    LeftLatch = 0;
-    RightLatch = 0;
+    InitOutput();
 
     Cnt = 0;
     MasterVolume = 0;
@@ -246,8 +254,16 @@ void SPU::Reset()
 
 void SPU::Stop()
 {
-    blip_clear(BlipL);
-    blip_clear(BlipR);
+    Platform::Mutex_Lock(AudioLock);
+    memset(OutputBuffer, 0, 2*OutputBufferSize*2);
+
+    blip_clear(BlipLeft);
+    blip_clear(BlipRight);
+    BlipTimer = 0;
+
+    OutputBufferReadPos = 0;
+    OutputBufferWritePos = 0;
+    Platform::Mutex_Unlock(AudioLock);
 }
 
 void SPU::DoSavestate(Savestate* file)
@@ -258,8 +274,6 @@ void SPU::DoSavestate(Savestate* file)
     file->Var8(&MasterVolume);
     file->Var16(&Bias);
 
-    file->Var8(&OutputSamplePos);
-    file->Var8(&OutputSampleInc);
     file->VarArray(OutputLastSamples, sizeof(OutputLastSamples));
 
     file->Var32(&MixInterval);
@@ -285,15 +299,12 @@ void SPU::SetSampleRate(AudioSampleRate rate)
     if (rate == AudioSampleRate::_47KHz)
     {
         MixInterval = 704;
-        OutputSampleInc = 16;
     }
     else
     {
         MixInterval = 1024;
-        OutputSampleInc = 11;
     }
 
-    OutputSamplePos = 0;
     memset(OutputLastSamples, 0, sizeof(OutputLastSamples));
 }
 
@@ -989,36 +1000,180 @@ void SPU::Mix(u32 spucycles)
         output[1] &= 0xFFC0;
     }
 
-    BlipAccumulate += spucycles;
-    // prevent blip from gaining too many samples
-    if (BlipAccumulate > 512 * 8192) BlipAccumulate = 512 * 8192;
+    BlipTimer += spucycles;
+    if (BlipTimer >= 8191 * 512)
+        BlipTimer = 8191 * 512;
+
+    if (output[0] != OutputLastSamples[0])
+        blip_add_delta(BlipLeft, BlipTimer, (int) output[0] - OutputLastSamples[0]);
+    if (output[1] != OutputLastSamples[1])
+        blip_add_delta(BlipRight, BlipTimer, (int) output[1] - OutputLastSamples[1]);
 
-    if (LeftLatch != output[0])
+    OutputLastSamples[0] = output[0];
+    OutputLastSamples[1] = output[1];
+
+    NDS.ScheduleEvent(Event_SPU, true, MixInterval, 0, MixInterval >> 1);
+}
+
+void SPU::EndFrame()
+{
+    blip_end_frame(BlipLeft, BlipTimer);
+    blip_end_frame(BlipRight, BlipTimer);
+    BlipTimer = 0;
+
+    int avail = blip_samples_avail(BlipLeft);
+    s16 temp[avail * 2];
+    blip_read_samples(BlipLeft, temp, avail, true);
+    blip_read_samples(BlipRight, temp + 1, avail, true);
+
+    Platform::Mutex_Lock(AudioLock);
+    for (int i = 0; i < avail * 2; i += 2)
     {
-        s32 diff = output[0] - LeftLatch;
-        LeftLatch = output[0];
-        blip_add_delta(BlipL, BlipAccumulate, diff);
+        OutputBuffer[OutputBufferWritePos++] = temp[i];
+        OutputBuffer[OutputBufferWritePos++] = temp[i+1];
+
+        OutputBufferWritePos &= ((2*OutputBufferSize)-1);
+
+        if (OutputBufferWritePos == OutputBufferReadPos)
+        {
+            // advance the read position too, to avoid losing the entire FIFO
+            OutputBufferReadPos += 2;
+            OutputBufferReadPos &= ((2*OutputBufferSize)-1);
+        }
     }
+    Platform::Mutex_Unlock(AudioLock);
+}
+
+void SPU::TrimOutput()
+{
+    Platform::Mutex_Lock(AudioLock);
+    const int halflimit = (OutputBufferSize / 2);
 
-    if (RightLatch != output[1])
+    int readpos = OutputBufferWritePos - (halflimit*2);
+    if (readpos < 0) readpos += (OutputBufferSize*2);
+
+    OutputBufferReadPos = readpos;
+    Platform::Mutex_Unlock(AudioLock);
+}
+
+void SPU::DrainOutput()
+{
+    Platform::Mutex_Lock(AudioLock);
+    OutputBufferReadPos = 0;
+    OutputBufferWritePos = 0;
+    Platform::Mutex_Unlock(AudioLock);
+}
+
+void SPU::InitOutput()
+{
+    Platform::Mutex_Lock(AudioLock);
+
+    blip_set_rates(BlipLeft, INTERNAL_SAMPLE_RATE * OutputSkew, OutputSampleRate);
+    blip_set_rates(BlipRight, INTERNAL_SAMPLE_RATE * OutputSkew, OutputSampleRate);
+
+    u32 needSamples = (u32) ceil(INTERNAL_SAMPLE_RATE / 60 / INTERNAL_SAMPLE_RATE * OutputSampleRate);
+    u32 newBufferSize = 512;
+    while (newBufferSize < needSamples)
+        newBufferSize <<= 1;
+    newBufferSize <<= 1;
+
+    if (newBufferSize != OutputBufferSize)
     {
-        s32 diff = output[1] - RightLatch;
-        RightLatch = output[1];
-        blip_add_delta(BlipR, BlipAccumulate, diff);
+        if (OutputBuffer != nullptr)
+            free(OutputBuffer);
+        OutputBuffer = (s16*) malloc(2 * newBufferSize * 2);
+        OutputBufferSize = newBufferSize;
     }
 
-    NDS.ScheduleEvent(Event_SPU, true, MixInterval, 0, MixInterval >> 1);
+    memset(OutputBuffer, 0, 2*OutputBufferSize*2);
+    OutputBufferReadPos = 0;
+    OutputBufferWritePos = 0;
+    Platform::Mutex_Unlock(AudioLock);
+}
+
+int SPU::GetOutputSize() const
+{
+    Platform::Mutex_Lock(AudioLock);
+
+    int ret;
+    if (OutputBufferWritePos >= OutputBufferReadPos)
+        ret = OutputBufferWritePos - OutputBufferReadPos;
+    else
+        ret = (OutputBufferSize*2) - OutputBufferReadPos + OutputBufferWritePos;
+
+    ret >>= 1;
+
+    Platform::Mutex_Unlock(AudioLock);
+    return ret;
+}
+
+void SPU::Sync(bool wait)
+{
+    // this function is currently not used anywhere
+    // depending on the usage context the thread safety measures could be made
+    // a lot faster
+
+    // sync to audio output in case the core is running too fast
+    // * wait=true: wait until enough audio data has been played
+    // * wait=false: merely skip some audio data to avoid a FIFO overflow
+
+    const int halflimit = (OutputBufferSize / 2);
+
+    if (wait)
+    {
+        // TODO: less CPU-intensive wait?
+        while (GetOutputSize() > halflimit);
+    }
+    else if (GetOutputSize() > halflimit)
+    {
+        Platform::Mutex_Lock(AudioLock);
+
+        int readpos = OutputBufferWritePos - (halflimit*2);
+        if (readpos < 0) readpos += (OutputBufferSize*2);
+
+        OutputBufferReadPos = readpos;
+
+        Platform::Mutex_Unlock(AudioLock);
+    }
+}
+
+int SPU::ReadOutput(s16* data, int samples)
+{
+    Platform::Mutex_Lock(AudioLock);
+    if (OutputBufferReadPos == OutputBufferWritePos)
+    {
+        Platform::Mutex_Unlock(AudioLock);
+        return 0;
+    }
+
+    for (int i = 0; i < samples; i++)
+    {
+        *data++ = OutputBuffer[OutputBufferReadPos++];
+        *data++ = OutputBuffer[OutputBufferReadPos++];
+        OutputBufferReadPos &= ((2*OutputBufferSize)-1);
+
+        if (OutputBufferWritePos == OutputBufferReadPos)
+        {
+            Platform::Mutex_Unlock(AudioLock);
+            return i+1;
+        }
+    }
+
+    Platform::Mutex_Unlock(AudioLock);
+    return samples;
+}
+
+void SPU::SetOutputSampleRate(double rate)
+{
+    OutputSampleRate = rate;
+    InitOutput();
 }
 
-u32 SPU::ReadOutput(s16* data, u32 maxSamples)
+void SPU::SetOutputSkew(double skew)
 {
-    blip_end_frame(BlipL, BlipAccumulate);
-    blip_end_frame(BlipR, BlipAccumulate);
-    BlipAccumulate = 0;
-    u32 samples = blip_samples_avail(BlipL);
-    blip_read_samples(BlipL, data + 0, maxSamples, true);
-    blip_read_samples(BlipR, data + 1, maxSamples, true);
-    return std::min(samples, maxSamples);
+    blip_set_rates(BlipLeft, INTERNAL_SAMPLE_RATE * skew, OutputSampleRate);
+    blip_set_rates(BlipRight, INTERNAL_SAMPLE_RATE * skew, OutputSampleRate);
+    OutputSkew = skew;
 }
 
 
diff --git a/src/SPU.h b/src/SPU.h
index 0e6b1d472c..ef7f895bca 100644
--- a/src/SPU.h
+++ b/src/SPU.h
@@ -22,10 +22,11 @@
 #include "Savestate.h"
 #include "Platform.h"
 
-#include "blip_buf/blip_buf.h"
+struct blip_t;
 
 namespace melonDS
 {
+
 class NDS;
 class SPU;
 
@@ -224,7 +225,7 @@ class SPUCaptureUnit
 class SPU
 {
 public:
-    explicit SPU(melonDS::NDS& nds, AudioBitDepth bitdepth, AudioInterpolation interpolation);
+    explicit SPU(melonDS::NDS& nds, AudioBitDepth bitdepth, AudioInterpolation interpolation, double outputSampleRate);
     ~SPU();
     void Reset();
     void DoSavestate(Savestate* file);
@@ -244,8 +245,16 @@ class SPU
     void SetApplyBias(bool enable);
 
     void Mix(u32 spucycles);
+    void EndFrame();
 
-    u32 ReadOutput(s16* data, u32 maxSamples);
+    void TrimOutput();
+    void DrainOutput();
+    void InitOutput();
+    int GetOutputSize() const;
+    void Sync(bool wait);
+    int ReadOutput(s16* data, int samples);
+    void SetOutputSampleRate(double rate);
+    void SetOutputSkew(double skew);
 
     u8 Read8(u32 addr);
     u16 Read16(u32 addr);
@@ -255,21 +264,23 @@ class SPU
     void Write32(u32 addr, u32 val);
 
 private:
+    u32 OutputBufferSize = 0;
+    double OutputSampleRate;
+    double OutputSkew = 1.0;
     melonDS::NDS& NDS;
 
-    // sample pos/inc are 4-bit fractional
-    // 32KHz sample rate is 11/16 of 47KHz
-    // so they don't need to be very precise
-    u8 OutputSamplePos;
-    u8 OutputSampleInc;
+    blip_t* BlipLeft;
+    blip_t* BlipRight;
+    int BlipTimer = 0;
+
+    s16* OutputBuffer;
+    u32 OutputBufferWritePos = 0;
+    u32 OutputBufferReadPos = 0;
     s16 OutputLastSamples[2];
 
     u32 MixInterval;
 
-    blip_t* BlipL;
-    blip_t* BlipR;
-    u32 BlipAccumulate = 0;
-    s32 LeftLatch = 0, RightLatch = 0;
+    Platform::Mutex* AudioLock;
 
     u16 Cnt = 0;
     u8 MasterVolume = 0;
diff --git a/src/Utils.h b/src/Utils.h
index 41e14c86d8..b497cbc9b0 100644
--- a/src/Utils.h
+++ b/src/Utils.h
@@ -38,6 +38,20 @@ std::pair<std::unique_ptr<u8[]>, u32> PadToPowerOf2(const u8* data, u32 len) noe
 
 std::unique_ptr<u8[]> CopyToUnique(const u8* data, u32 len) noexcept;
 
+template <typename T>
+T GetMSBit(T val)
+{
+    val |= (val >>  1);
+    val |= (val >>  2);
+    val |= (val >>  4);
+
+    if constexpr(sizeof(val) > 1) val |= (val >>  8);
+    if constexpr(sizeof(val) > 2) val |= (val >> 16);
+    if constexpr(sizeof(val) > 4) val |= (val >> 32);
+    
+    return val - (val >> 1);
+}
+
 }
 
 #endif // MELONDS_UTILS_H
diff --git a/src/blip-buf/blip_buf.c b/src/blip-buf/blip_buf.c
new file mode 100644
index 0000000000..929b8fc669
--- /dev/null
+++ b/src/blip-buf/blip_buf.c
@@ -0,0 +1,344 @@
+/* blip_buf $vers. http://www.slack.net/~ant/ */
+
+#include "blip_buf.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <string.h>
+#include <stdlib.h>
+
+/* Library Copyright (C) 2003-2009 Shay Green. This library is free software;
+you can redistribute it and/or modify it under the terms of the GNU Lesser
+General Public License as published by the Free Software Foundation; either
+version 2.1 of the License, or (at your option) any later version. This
+library is distributed in the hope that it will be useful, but WITHOUT ANY
+WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
+A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for more
+details. You should have received a copy of the GNU Lesser General Public
+License along with this module; if not, write to the Free Software Foundation,
+Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */
+
+#if defined (BLARGG_TEST) && BLARGG_TEST
+	#include "blargg_test.h"
+#endif
+
+/* Equivalent to ULONG_MAX >= 0xFFFFFFFF00000000.
+Avoids constants that don't fit in 32 bits. */
+#if ULONG_MAX/0xFFFFFFFF > 0xFFFFFFFF
+	typedef unsigned long fixed_t;
+	enum { pre_shift = 32 };
+
+#elif defined(ULLONG_MAX)
+	typedef unsigned long long fixed_t;
+	enum { pre_shift = 32 };
+
+#else
+	typedef unsigned fixed_t;
+	enum { pre_shift = 0 };
+
+#endif
+
+enum { time_bits = pre_shift + 20 };
+
+static fixed_t const time_unit = (fixed_t) 1 << time_bits;
+
+enum { bass_shift  = 9 }; /* affects high-pass filter breakpoint frequency */
+enum { end_frame_extra = 2 }; /* allows deltas slightly after frame length */
+
+enum { half_width  = 8 };
+enum { buf_extra   = half_width*2 + end_frame_extra };
+enum { phase_bits  = 5 };
+enum { phase_count = 1 << phase_bits };
+enum { delta_bits  = 15 };
+enum { delta_unit  = 1 << delta_bits };
+enum { frac_bits = time_bits - pre_shift };
+
+/* We could eliminate avail and encode whole samples in offset, but that would
+limit the total buffered samples to blip_max_frame. That could only be
+increased by decreasing time_bits, which would reduce resample ratio accuracy.
+*/
+
+/** Sample buffer that resamples to output rate and accumulates samples
+until they're read out */
+struct blip_t
+{
+	fixed_t factor;
+	fixed_t offset;
+	int avail;
+	int size;
+	int integrator;
+};
+
+typedef int buf_t;
+
+/* probably not totally portable */
+#define SAMPLES( buf ) ((buf_t*) ((buf) + 1))
+
+/* Arithmetic (sign-preserving) right shift */
+#define ARITH_SHIFT( n, shift ) \
+	((n) >> (shift))
+
+enum { max_sample = +32767 };
+enum { min_sample = -32768 };
+
+#define CLAMP( n ) \
+	{\
+		if ( (short) n != n )\
+			n = ARITH_SHIFT( n, 16 ) ^ max_sample;\
+	}
+
+static void check_assumptions( void )
+{
+	int n;
+	
+	#if INT_MAX < 0x7FFFFFFF || UINT_MAX < 0xFFFFFFFF
+		#error "int must be at least 32 bits"
+	#endif
+	
+	assert( (-3 >> 1) == -2 ); /* right shift must preserve sign */
+	
+	n = max_sample * 2;
+	CLAMP( n );
+	assert( n == max_sample );
+	
+	n = min_sample * 2;
+	CLAMP( n );
+	assert( n == min_sample );
+	
+	assert( blip_max_ratio <= time_unit );
+	assert( blip_max_frame <= (fixed_t) -1 >> time_bits );
+}
+
+blip_t* blip_new( int size )
+{
+	blip_t* m;
+	assert( size >= 0 );
+	
+	m = (blip_t*) malloc( sizeof *m + (size + buf_extra) * sizeof (buf_t) );
+	if ( m )
+	{
+		m->factor = time_unit / blip_max_ratio;
+		m->size   = size;
+		blip_clear( m );
+		check_assumptions();
+	}
+	return m;
+}
+
+void blip_delete( blip_t* m )
+{
+	if ( m != NULL )
+	{
+		/* Clear fields in case user tries to use after freeing */
+		memset( m, 0, sizeof *m );
+		free( m );
+	}
+}
+
+void blip_set_rates( blip_t* m, double clock_rate, double sample_rate )
+{
+	double factor = time_unit * sample_rate / clock_rate;
+	m->factor = (fixed_t) factor;
+	
+	/* Fails if clock_rate exceeds maximum, relative to sample_rate */
+	assert( 0 <= factor - m->factor && factor - m->factor < 1 );
+	
+	/* Avoid requiring math.h. Equivalent to
+	m->factor = (int) ceil( factor ) */
+	if ( m->factor < factor )
+		m->factor++;
+	
+	/* At this point, factor is most likely rounded up, but could still
+	have been rounded down in the floating-point calculation. */
+}
+
+void blip_clear( blip_t* m )
+{
+	/* We could set offset to 0, factor/2, or factor-1. 0 is suitable if
+	factor is rounded up. factor-1 is suitable if factor is rounded down.
+	Since we don't know rounding direction, factor/2 accommodates either,
+	with the slight loss of showing an error in half the time. Since for
+	a 64-bit factor this is years, the halving isn't a problem. */
+	
+	m->offset     = m->factor / 2;
+	m->avail      = 0;
+	m->integrator = 0;
+	memset( SAMPLES( m ), 0, (m->size + buf_extra) * sizeof (buf_t) );
+}
+
+int blip_clocks_needed( const blip_t* m, int samples )
+{
+	fixed_t needed;
+	
+	/* Fails if buffer can't hold that many more samples */
+	assert( samples >= 0 && m->avail + samples <= m->size );
+	
+	needed = (fixed_t) samples * time_unit;
+	if ( needed < m->offset )
+		return 0;
+	
+	return (needed - m->offset + m->factor - 1) / m->factor;
+}
+
+void blip_end_frame( blip_t* m, unsigned t )
+{
+	fixed_t off = t * m->factor + m->offset;
+	m->avail += off >> time_bits;
+	m->offset = off & (time_unit - 1);
+	
+	/* Fails if buffer size was exceeded */
+	assert( m->avail <= m->size );
+}
+
+int blip_samples_avail( const blip_t* m )
+{
+	return m->avail;
+}
+
+static void remove_samples( blip_t* m, int count )
+{
+	buf_t* buf = SAMPLES( m );
+	int remain = m->avail + buf_extra - count;
+	m->avail -= count;
+	
+	memmove( &buf [0], &buf [count], remain * sizeof buf [0] );
+	memset( &buf [remain], 0, count * sizeof buf [0] );
+}
+
+int blip_read_samples( blip_t* m, short out [], int count, int stereo )
+{
+	assert( count >= 0 );
+	
+	if ( count > m->avail )
+		count = m->avail;
+	
+	if ( count )
+	{
+		int const step = stereo ? 2 : 1;
+		buf_t const* in  = SAMPLES( m );
+		buf_t const* end = in + count;
+		int sum = m->integrator;
+		do
+		{
+			/* Eliminate fraction */
+			int s = ARITH_SHIFT( sum, delta_bits );
+			
+			sum += *in++;
+			
+			CLAMP( s );
+			
+			*out = s;
+			out += step;
+			
+			/* High-pass filter */
+			sum -= s << (delta_bits - bass_shift);
+		}
+		while ( in != end );
+		m->integrator = sum;
+		
+		remove_samples( m, count );
+	}
+	
+	return count;
+}
+
+/* Things that didn't help performance on x86:
+	__attribute__((aligned(128)))
+	#define short int
+	restrict
+*/
+
+/* Sinc_Generator( 0.9, 0.55, 4.5 ) */
+static short const bl_step [phase_count + 1] [half_width] =
+{
+{   43, -115,  350, -488, 1136, -914, 5861,21022},
+{   44, -118,  348, -473, 1076, -799, 5274,21001},
+{   45, -121,  344, -454, 1011, -677, 4706,20936},
+{   46, -122,  336, -431,  942, -549, 4156,20829},
+{   47, -123,  327, -404,  868, -418, 3629,20679},
+{   47, -122,  316, -375,  792, -285, 3124,20488},
+{   47, -120,  303, -344,  714, -151, 2644,20256},
+{   46, -117,  289, -310,  634,  -17, 2188,19985},
+{   46, -114,  273, -275,  553,  117, 1758,19675},
+{   44, -108,  255, -237,  471,  247, 1356,19327},
+{   43, -103,  237, -199,  390,  373,  981,18944},
+{   42,  -98,  218, -160,  310,  495,  633,18527},
+{   40,  -91,  198, -121,  231,  611,  314,18078},
+{   38,  -84,  178,  -81,  153,  722,   22,17599},
+{   36,  -76,  157,  -43,   80,  824, -241,17092},
+{   34,  -68,  135,   -3,    8,  919, -476,16558},
+{   32,  -61,  115,   34,  -60, 1006, -683,16001},
+{   29,  -52,   94,   70, -123, 1083, -862,15422},
+{   27,  -44,   73,  106, -184, 1152,-1015,14824},
+{   25,  -36,   53,  139, -239, 1211,-1142,14210},
+{   22,  -27,   34,  170, -290, 1261,-1244,13582},
+{   20,  -20,   16,  199, -335, 1301,-1322,12942},
+{   18,  -12,   -3,  226, -375, 1331,-1376,12293},
+{   15,   -4,  -19,  250, -410, 1351,-1408,11638},
+{   13,    3,  -35,  272, -439, 1361,-1419,10979},
+{   11,    9,  -49,  292, -464, 1362,-1410,10319},
+{    9,   16,  -63,  309, -483, 1354,-1383, 9660},
+{    7,   22,  -75,  322, -496, 1337,-1339, 9005},
+{    6,   26,  -85,  333, -504, 1312,-1280, 8355},
+{    4,   31,  -94,  341, -507, 1278,-1205, 7713},
+{    3,   35, -102,  347, -506, 1238,-1119, 7082},
+{    1,   40, -110,  350, -499, 1190,-1021, 6464},
+{    0,   43, -115,  350, -488, 1136, -914, 5861}
+};
+
+/* Shifting by pre_shift allows calculation using unsigned int rather than
+possibly-wider fixed_t. On 32-bit platforms, this is likely more efficient.
+And by having pre_shift 32, a 32-bit platform can easily do the shift by
+simply ignoring the low half. */
+
+void blip_add_delta( blip_t* m, unsigned time, int delta )
+{
+	unsigned fixed = (unsigned) ((time * m->factor + m->offset) >> pre_shift);
+	buf_t* out = SAMPLES( m ) + m->avail + (fixed >> frac_bits);
+	
+	int const phase_shift = frac_bits - phase_bits;
+	int phase = fixed >> phase_shift & (phase_count - 1);
+	short const* in  = bl_step [phase];
+	short const* rev = bl_step [phase_count - phase];
+	
+	int interp = fixed >> (phase_shift - delta_bits) & (delta_unit - 1);
+	int delta2 = (delta * interp) >> delta_bits;
+	delta -= delta2;
+	
+	/* Fails if buffer size was exceeded */
+	assert( out <= &SAMPLES( m ) [m->size + end_frame_extra] );
+	
+	out [0] += in[0]*delta + in[half_width+0]*delta2;
+	out [1] += in[1]*delta + in[half_width+1]*delta2;
+	out [2] += in[2]*delta + in[half_width+2]*delta2;
+	out [3] += in[3]*delta + in[half_width+3]*delta2;
+	out [4] += in[4]*delta + in[half_width+4]*delta2;
+	out [5] += in[5]*delta + in[half_width+5]*delta2;
+	out [6] += in[6]*delta + in[half_width+6]*delta2;
+	out [7] += in[7]*delta + in[half_width+7]*delta2;
+	
+	in = rev;
+	out [ 8] += in[7]*delta + in[7-half_width]*delta2;
+	out [ 9] += in[6]*delta + in[6-half_width]*delta2;
+	out [10] += in[5]*delta + in[5-half_width]*delta2;
+	out [11] += in[4]*delta + in[4-half_width]*delta2;
+	out [12] += in[3]*delta + in[3-half_width]*delta2;
+	out [13] += in[2]*delta + in[2-half_width]*delta2;
+	out [14] += in[1]*delta + in[1-half_width]*delta2;
+	out [15] += in[0]*delta + in[0-half_width]*delta2;
+}
+
+void blip_add_delta_fast( blip_t* m, unsigned time, int delta )
+{
+	unsigned fixed = (unsigned) ((time * m->factor + m->offset) >> pre_shift);
+	buf_t* out = SAMPLES( m ) + m->avail + (fixed >> frac_bits);
+	
+	int interp = fixed >> (frac_bits - delta_bits) & (delta_unit - 1);
+	int delta2 = delta * interp;
+	
+	/* Fails if buffer size was exceeded */
+	assert( out <= &SAMPLES( m ) [m->size + end_frame_extra] );
+	
+	out [7] += delta * delta_unit - delta2;
+	out [8] += delta2;
+}
diff --git a/src/blip-buf/blip_buf.h b/src/blip-buf/blip_buf.h
new file mode 100644
index 0000000000..0853595efe
--- /dev/null
+++ b/src/blip-buf/blip_buf.h
@@ -0,0 +1,72 @@
+/** \file
+Sample buffer that resamples from input clock rate to output sample rate */
+
+/* blip_buf $vers */
+#ifndef BLIP_BUF_H 
+#define BLIP_BUF_H
+
+#ifdef __cplusplus
+	extern "C" {
+#endif
+
+/** First parameter of most functions is blip_t*, or const blip_t* if nothing
+is changed. */
+typedef struct blip_t blip_t;
+
+/** Creates new buffer that can hold at most sample_count samples. Sets rates
+so that there are blip_max_ratio clocks per sample. Returns pointer to new
+buffer, or NULL if insufficient memory. */
+blip_t* blip_new( int sample_count );
+
+/** Sets approximate input clock rate and output sample rate. For every
+clock_rate input clocks, approximately sample_rate samples are generated. */
+void blip_set_rates( blip_t*, double clock_rate, double sample_rate );
+
+enum { /** Maximum clock_rate/sample_rate ratio. For a given sample_rate,
+clock_rate must not be greater than sample_rate*blip_max_ratio. */
+blip_max_ratio = 1 << 20 };
+
+/** Clears entire buffer. Afterwards, blip_samples_avail() == 0. */
+void blip_clear( blip_t* );
+
+/** Adds positive/negative delta into buffer at specified clock time. */
+void blip_add_delta( blip_t*, unsigned int clock_time, int delta );
+
+/** Same as blip_add_delta(), but uses faster, lower-quality synthesis. */
+void blip_add_delta_fast( blip_t*, unsigned int clock_time, int delta );
+
+/** Length of time frame, in clocks, needed to make sample_count additional
+samples available. */
+int blip_clocks_needed( const blip_t*, int sample_count );
+
+enum { /** Maximum number of samples that can be generated from one time frame. */
+blip_max_frame = 4000 };
+
+/** Makes input clocks before clock_duration available for reading as output
+samples. Also begins new time frame at clock_duration, so that clock time 0 in
+the new time frame specifies the same clock as clock_duration in the old time
+frame specified. Deltas can have been added slightly past clock_duration (up to
+however many clocks there are in two output samples). */
+void blip_end_frame( blip_t*, unsigned int clock_duration );
+
+/** Number of buffered samples available for reading. */
+int blip_samples_avail( const blip_t* );
+
+/** Reads and removes at most 'count' samples and writes them to 'out'. If
+'stereo' is true, writes output to every other element of 'out', allowing easy
+interleaving of two buffers into a stereo sample stream. Outputs 16-bit signed
+samples. Returns number of samples actually read.  */
+int blip_read_samples( blip_t*, short out [], int count, int stereo );
+
+/** Frees buffer. No effect if NULL is passed. */
+void blip_delete( blip_t* );
+
+
+/* Deprecated */
+typedef blip_t blip_buffer_t;
+
+#ifdef __cplusplus
+	}
+#endif
+
+#endif
diff --git a/src/blip-buf/license.txt b/src/blip-buf/license.txt
new file mode 100644
index 0000000000..5faba9d48c
--- /dev/null
+++ b/src/blip-buf/license.txt
@@ -0,0 +1,504 @@
+		  GNU LESSER GENERAL PUBLIC LICENSE
+		       Version 2.1, February 1999
+
+ Copyright (C) 1991, 1999 Free Software Foundation, Inc.
+ 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+[This is the first released version of the Lesser GPL.  It also counts
+ as the successor of the GNU Library Public License, version 2, hence
+ the version number 2.1.]
+
+			    Preamble
+
+  The licenses for most software are designed to take away your
+freedom to share and change it.  By contrast, the GNU General Public
+Licenses are intended to guarantee your freedom to share and change
+free software--to make sure the software is free for all its users.
+
+  This license, the Lesser General Public License, applies to some
+specially designated software packages--typically libraries--of the
+Free Software Foundation and other authors who decide to use it.  You
+can use it too, but we suggest you first think carefully about whether
+this license or the ordinary General Public License is the better
+strategy to use in any particular case, based on the explanations below.
+
+  When we speak of free software, we are referring to freedom of use,
+not price.  Our General Public Licenses are designed to make sure that
+you have the freedom to distribute copies of free software (and charge
+for this service if you wish); that you receive source code or can get
+it if you want it; that you can change the software and use pieces of
+it in new free programs; and that you are informed that you can do
+these things.
+
+  To protect your rights, we need to make restrictions that forbid
+distributors to deny you these rights or to ask you to surrender these
+rights.  These restrictions translate to certain responsibilities for
+you if you distribute copies of the library or if you modify it.
+
+  For example, if you distribute copies of the library, whether gratis
+or for a fee, you must give the recipients all the rights that we gave
+you.  You must make sure that they, too, receive or can get the source
+code.  If you link other code with the library, you must provide
+complete object files to the recipients, so that they can relink them
+with the library after making changes to the library and recompiling
+it.  And you must show them these terms so they know their rights.
+
+  We protect your rights with a two-step method: (1) we copyright the
+library, and (2) we offer you this license, which gives you legal
+permission to copy, distribute and/or modify the library.
+
+  To protect each distributor, we want to make it very clear that
+there is no warranty for the free library.  Also, if the library is
+modified by someone else and passed on, the recipients should know
+that what they have is not the original version, so that the original
+author's reputation will not be affected by problems that might be
+introduced by others.
+
+  Finally, software patents pose a constant threat to the existence of
+any free program.  We wish to make sure that a company cannot
+effectively restrict the users of a free program by obtaining a
+restrictive license from a patent holder.  Therefore, we insist that
+any patent license obtained for a version of the library must be
+consistent with the full freedom of use specified in this license.
+
+  Most GNU software, including some libraries, is covered by the
+ordinary GNU General Public License.  This license, the GNU Lesser
+General Public License, applies to certain designated libraries, and
+is quite different from the ordinary General Public License.  We use
+this license for certain libraries in order to permit linking those
+libraries into non-free programs.
+
+  When a program is linked with a library, whether statically or using
+a shared library, the combination of the two is legally speaking a
+combined work, a derivative of the original library.  The ordinary
+General Public License therefore permits such linking only if the
+entire combination fits its criteria of freedom.  The Lesser General
+Public License permits more lax criteria for linking other code with
+the library.
+
+  We call this license the "Lesser" General Public License because it
+does Less to protect the user's freedom than the ordinary General
+Public License.  It also provides other free software developers Less
+of an advantage over competing non-free programs.  These disadvantages
+are the reason we use the ordinary General Public License for many
+libraries.  However, the Lesser license provides advantages in certain
+special circumstances.
+
+  For example, on rare occasions, there may be a special need to
+encourage the widest possible use of a certain library, so that it becomes
+a de-facto standard.  To achieve this, non-free programs must be
+allowed to use the library.  A more frequent case is that a free
+library does the same job as widely used non-free libraries.  In this
+case, there is little to gain by limiting the free library to free
+software only, so we use the Lesser General Public License.
+
+  In other cases, permission to use a particular library in non-free
+programs enables a greater number of people to use a large body of
+free software.  For example, permission to use the GNU C Library in
+non-free programs enables many more people to use the whole GNU
+operating system, as well as its variant, the GNU/Linux operating
+system.
+
+  Although the Lesser General Public License is Less protective of the
+users' freedom, it does ensure that the user of a program that is
+linked with the Library has the freedom and the wherewithal to run
+that program using a modified version of the Library.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.  Pay close attention to the difference between a
+"work based on the library" and a "work that uses the library".  The
+former contains code derived from the library, whereas the latter must
+be combined with the library in order to run.
+
+		  GNU LESSER GENERAL PUBLIC LICENSE
+   TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
+
+  0. This License Agreement applies to any software library or other
+program which contains a notice placed by the copyright holder or
+other authorized party saying it may be distributed under the terms of
+this Lesser General Public License (also called "this License").
+Each licensee is addressed as "you".
+
+  A "library" means a collection of software functions and/or data
+prepared so as to be conveniently linked with application programs
+(which use some of those functions and data) to form executables.
+
+  The "Library", below, refers to any such software library or work
+which has been distributed under these terms.  A "work based on the
+Library" means either the Library or any derivative work under
+copyright law: that is to say, a work containing the Library or a
+portion of it, either verbatim or with modifications and/or translated
+straightforwardly into another language.  (Hereinafter, translation is
+included without limitation in the term "modification".)
+
+  "Source code" for a work means the preferred form of the work for
+making modifications to it.  For a library, complete source code means
+all the source code for all modules it contains, plus any associated
+interface definition files, plus the scripts used to control compilation
+and installation of the library.
+
+  Activities other than copying, distribution and modification are not
+covered by this License; they are outside its scope.  The act of
+running a program using the Library is not restricted, and output from
+such a program is covered only if its contents constitute a work based
+on the Library (independent of the use of the Library in a tool for
+writing it).  Whether that is true depends on what the Library does
+and what the program that uses the Library does.
+  
+  1. You may copy and distribute verbatim copies of the Library's
+complete source code as you receive it, in any medium, provided that
+you conspicuously and appropriately publish on each copy an
+appropriate copyright notice and disclaimer of warranty; keep intact
+all the notices that refer to this License and to the absence of any
+warranty; and distribute a copy of this License along with the
+Library.
+
+  You may charge a fee for the physical act of transferring a copy,
+and you may at your option offer warranty protection in exchange for a
+fee.
+
+  2. You may modify your copy or copies of the Library or any portion
+of it, thus forming a work based on the Library, and copy and
+distribute such modifications or work under the terms of Section 1
+above, provided that you also meet all of these conditions:
+
+    a) The modified work must itself be a software library.
+
+    b) You must cause the files modified to carry prominent notices
+    stating that you changed the files and the date of any change.
+
+    c) You must cause the whole of the work to be licensed at no
+    charge to all third parties under the terms of this License.
+
+    d) If a facility in the modified Library refers to a function or a
+    table of data to be supplied by an application program that uses
+    the facility, other than as an argument passed when the facility
+    is invoked, then you must make a good faith effort to ensure that,
+    in the event an application does not supply such function or
+    table, the facility still operates, and performs whatever part of
+    its purpose remains meaningful.
+
+    (For example, a function in a library to compute square roots has
+    a purpose that is entirely well-defined independent of the
+    application.  Therefore, Subsection 2d requires that any
+    application-supplied function or table used by this function must
+    be optional: if the application does not supply it, the square
+    root function must still compute square roots.)
+
+These requirements apply to the modified work as a whole.  If
+identifiable sections of that work are not derived from the Library,
+and can be reasonably considered independent and separate works in
+themselves, then this License, and its terms, do not apply to those
+sections when you distribute them as separate works.  But when you
+distribute the same sections as part of a whole which is a work based
+on the Library, the distribution of the whole must be on the terms of
+this License, whose permissions for other licensees extend to the
+entire whole, and thus to each and every part regardless of who wrote
+it.
+
+Thus, it is not the intent of this section to claim rights or contest
+your rights to work written entirely by you; rather, the intent is to
+exercise the right to control the distribution of derivative or
+collective works based on the Library.
+
+In addition, mere aggregation of another work not based on the Library
+with the Library (or with a work based on the Library) on a volume of
+a storage or distribution medium does not bring the other work under
+the scope of this License.
+
+  3. You may opt to apply the terms of the ordinary GNU General Public
+License instead of this License to a given copy of the Library.  To do
+this, you must alter all the notices that refer to this License, so
+that they refer to the ordinary GNU General Public License, version 2,
+instead of to this License.  (If a newer version than version 2 of the
+ordinary GNU General Public License has appeared, then you can specify
+that version instead if you wish.)  Do not make any other change in
+these notices.
+
+  Once this change is made in a given copy, it is irreversible for
+that copy, so the ordinary GNU General Public License applies to all
+subsequent copies and derivative works made from that copy.
+
+  This option is useful when you wish to copy part of the code of
+the Library into a program that is not a library.
+
+  4. You may copy and distribute the Library (or a portion or
+derivative of it, under Section 2) in object code or executable form
+under the terms of Sections 1 and 2 above provided that you accompany
+it with the complete corresponding machine-readable source code, which
+must be distributed under the terms of Sections 1 and 2 above on a
+medium customarily used for software interchange.
+
+  If distribution of object code is made by offering access to copy
+from a designated place, then offering equivalent access to copy the
+source code from the same place satisfies the requirement to
+distribute the source code, even though third parties are not
+compelled to copy the source along with the object code.
+
+  5. A program that contains no derivative of any portion of the
+Library, but is designed to work with the Library by being compiled or
+linked with it, is called a "work that uses the Library".  Such a
+work, in isolation, is not a derivative work of the Library, and
+therefore falls outside the scope of this License.
+
+  However, linking a "work that uses the Library" with the Library
+creates an executable that is a derivative of the Library (because it
+contains portions of the Library), rather than a "work that uses the
+library".  The executable is therefore covered by this License.
+Section 6 states terms for distribution of such executables.
+
+  When a "work that uses the Library" uses material from a header file
+that is part of the Library, the object code for the work may be a
+derivative work of the Library even though the source code is not.
+Whether this is true is especially significant if the work can be
+linked without the Library, or if the work is itself a library.  The
+threshold for this to be true is not precisely defined by law.
+
+  If such an object file uses only numerical parameters, data
+structure layouts and accessors, and small macros and small inline
+functions (ten lines or less in length), then the use of the object
+file is unrestricted, regardless of whether it is legally a derivative
+work.  (Executables containing this object code plus portions of the
+Library will still fall under Section 6.)
+
+  Otherwise, if the work is a derivative of the Library, you may
+distribute the object code for the work under the terms of Section 6.
+Any executables containing that work also fall under Section 6,
+whether or not they are linked directly with the Library itself.
+
+  6. As an exception to the Sections above, you may also combine or
+link a "work that uses the Library" with the Library to produce a
+work containing portions of the Library, and distribute that work
+under terms of your choice, provided that the terms permit
+modification of the work for the customer's own use and reverse
+engineering for debugging such modifications.
+
+  You must give prominent notice with each copy of the work that the
+Library is used in it and that the Library and its use are covered by
+this License.  You must supply a copy of this License.  If the work
+during execution displays copyright notices, you must include the
+copyright notice for the Library among them, as well as a reference
+directing the user to the copy of this License.  Also, you must do one
+of these things:
+
+    a) Accompany the work with the complete corresponding
+    machine-readable source code for the Library including whatever
+    changes were used in the work (which must be distributed under
+    Sections 1 and 2 above); and, if the work is an executable linked
+    with the Library, with the complete machine-readable "work that
+    uses the Library", as object code and/or source code, so that the
+    user can modify the Library and then relink to produce a modified
+    executable containing the modified Library.  (It is understood
+    that the user who changes the contents of definitions files in the
+    Library will not necessarily be able to recompile the application
+    to use the modified definitions.)
+
+    b) Use a suitable shared library mechanism for linking with the
+    Library.  A suitable mechanism is one that (1) uses at run time a
+    copy of the library already present on the user's computer system,
+    rather than copying library functions into the executable, and (2)
+    will operate properly with a modified version of the library, if
+    the user installs one, as long as the modified version is
+    interface-compatible with the version that the work was made with.
+
+    c) Accompany the work with a written offer, valid for at
+    least three years, to give the same user the materials
+    specified in Subsection 6a, above, for a charge no more
+    than the cost of performing this distribution.
+
+    d) If distribution of the work is made by offering access to copy
+    from a designated place, offer equivalent access to copy the above
+    specified materials from the same place.
+
+    e) Verify that the user has already received a copy of these
+    materials or that you have already sent this user a copy.
+
+  For an executable, the required form of the "work that uses the
+Library" must include any data and utility programs needed for
+reproducing the executable from it.  However, as a special exception,
+the materials to be distributed need not include anything that is
+normally distributed (in either source or binary form) with the major
+components (compiler, kernel, and so on) of the operating system on
+which the executable runs, unless that component itself accompanies
+the executable.
+
+  It may happen that this requirement contradicts the license
+restrictions of other proprietary libraries that do not normally
+accompany the operating system.  Such a contradiction means you cannot
+use both them and the Library together in an executable that you
+distribute.
+
+  7. You may place library facilities that are a work based on the
+Library side-by-side in a single library together with other library
+facilities not covered by this License, and distribute such a combined
+library, provided that the separate distribution of the work based on
+the Library and of the other library facilities is otherwise
+permitted, and provided that you do these two things:
+
+    a) Accompany the combined library with a copy of the same work
+    based on the Library, uncombined with any other library
+    facilities.  This must be distributed under the terms of the
+    Sections above.
+
+    b) Give prominent notice with the combined library of the fact
+    that part of it is a work based on the Library, and explaining
+    where to find the accompanying uncombined form of the same work.
+
+  8. You may not copy, modify, sublicense, link with, or distribute
+the Library except as expressly provided under this License.  Any
+attempt otherwise to copy, modify, sublicense, link with, or
+distribute the Library is void, and will automatically terminate your
+rights under this License.  However, parties who have received copies,
+or rights, from you under this License will not have their licenses
+terminated so long as such parties remain in full compliance.
+
+  9. You are not required to accept this License, since you have not
+signed it.  However, nothing else grants you permission to modify or
+distribute the Library or its derivative works.  These actions are
+prohibited by law if you do not accept this License.  Therefore, by
+modifying or distributing the Library (or any work based on the
+Library), you indicate your acceptance of this License to do so, and
+all its terms and conditions for copying, distributing or modifying
+the Library or works based on it.
+
+  10. Each time you redistribute the Library (or any work based on the
+Library), the recipient automatically receives a license from the
+original licensor to copy, distribute, link with or modify the Library
+subject to these terms and conditions.  You may not impose any further
+restrictions on the recipients' exercise of the rights granted herein.
+You are not responsible for enforcing compliance by third parties with
+this License.
+
+  11. If, as a consequence of a court judgment or allegation of patent
+infringement or for any other reason (not limited to patent issues),
+conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot
+distribute so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you
+may not distribute the Library at all.  For example, if a patent
+license would not permit royalty-free redistribution of the Library by
+all those who receive copies directly or indirectly through you, then
+the only way you could satisfy both it and this License would be to
+refrain entirely from distribution of the Library.
+
+If any portion of this section is held invalid or unenforceable under any
+particular circumstance, the balance of the section is intended to apply,
+and the section as a whole is intended to apply in other circumstances.
+
+It is not the purpose of this section to induce you to infringe any
+patents or other property right claims or to contest validity of any
+such claims; this section has the sole purpose of protecting the
+integrity of the free software distribution system which is
+implemented by public license practices.  Many people have made
+generous contributions to the wide range of software distributed
+through that system in reliance on consistent application of that
+system; it is up to the author/donor to decide if he or she is willing
+to distribute software through any other system and a licensee cannot
+impose that choice.
+
+This section is intended to make thoroughly clear what is believed to
+be a consequence of the rest of this License.
+
+  12. If the distribution and/or use of the Library is restricted in
+certain countries either by patents or by copyrighted interfaces, the
+original copyright holder who places the Library under this License may add
+an explicit geographical distribution limitation excluding those countries,
+so that distribution is permitted only in or among countries not thus
+excluded.  In such case, this License incorporates the limitation as if
+written in the body of this License.
+
+  13. The Free Software Foundation may publish revised and/or new
+versions of the Lesser General Public License from time to time.
+Such new versions will be similar in spirit to the present version,
+but may differ in detail to address new problems or concerns.
+
+Each version is given a distinguishing version number.  If the Library
+specifies a version number of this License which applies to it and
+"any later version", you have the option of following the terms and
+conditions either of that version or of any later version published by
+the Free Software Foundation.  If the Library does not specify a
+license version number, you may choose any version ever published by
+the Free Software Foundation.
+
+  14. If you wish to incorporate parts of the Library into other free
+programs whose distribution conditions are incompatible with these,
+write to the author to ask for permission.  For software which is
+copyrighted by the Free Software Foundation, write to the Free
+Software Foundation; we sometimes make exceptions for this.  Our
+decision will be guided by the two goals of preserving the free status
+of all derivatives of our free software and of promoting the sharing
+and reuse of software generally.
+
+			    NO WARRANTY
+
+  15. BECAUSE THE LIBRARY IS LICENSED FREE OF CHARGE, THERE IS NO
+WARRANTY FOR THE LIBRARY, TO THE EXTENT PERMITTED BY APPLICABLE LAW.
+EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR
+OTHER PARTIES PROVIDE THE LIBRARY "AS IS" WITHOUT WARRANTY OF ANY
+KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE
+LIBRARY IS WITH YOU.  SHOULD THE LIBRARY PROVE DEFECTIVE, YOU ASSUME
+THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN
+WRITING WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY
+AND/OR REDISTRIBUTE THE LIBRARY AS PERMITTED ABOVE, BE LIABLE TO YOU
+FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR
+CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE
+LIBRARY (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING
+RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A
+FAILURE OF THE LIBRARY TO OPERATE WITH ANY OTHER SOFTWARE), EVEN IF
+SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
+DAMAGES.
+
+		     END OF TERMS AND CONDITIONS
+
+           How to Apply These Terms to Your New Libraries
+
+  If you develop a new library, and you want it to be of the greatest
+possible use to the public, we recommend making it free software that
+everyone can redistribute and change.  You can do so by permitting
+redistribution under these terms (or, alternatively, under the terms of the
+ordinary General Public License).
+
+  To apply these terms, attach the following notices to the library.  It is
+safest to attach them to the start of each source file to most effectively
+convey the exclusion of warranty; and each file should have at least the
+"copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the library's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This library is free software; you can redistribute it and/or
+    modify it under the terms of the GNU Lesser General Public
+    License as published by the Free Software Foundation; either
+    version 2.1 of the License, or (at your option) any later version.
+
+    This library is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+    Lesser General Public License for more details.
+
+    You should have received a copy of the GNU Lesser General Public
+    License along with this library; if not, write to the Free Software
+    Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
+
+Also add information on how to contact you by electronic and paper mail.
+
+You should also get your employer (if you work as a programmer) or your
+school, if any, to sign a "copyright disclaimer" for the library, if
+necessary.  Here is a sample; alter the names:
+
+  Yoyodyne, Inc., hereby disclaims all copyright interest in the
+  library `Frob' (a library for tweaking knobs) written by James Random Hacker.
+
+  <signature of Ty Coon>, 1 April 1990
+  Ty Coon, President of Vice
+
+That's all there is to it!
+
+
diff --git a/src/blip-buf/readme.txt b/src/blip-buf/readme.txt
new file mode 100644
index 0000000000..6a4aa1c3e8
--- /dev/null
+++ b/src/blip-buf/readme.txt
@@ -0,0 +1,62 @@
+blip_buf $vers: Band-Limited Audio Buffer
+-----------------------------------------
+Blip_buf is a small waveform synthesis library meant for use in classic
+video game sound chip emulation. It greatly simplifies sound chip
+emulation code by handling all the details of resampling. The emulator
+merely sets the input clock rate and output sample rate, adds waveforms
+by specifying the clock times where their amplitude changes, then reads
+the resulting output samples. For a more full-features synthesis
+library, get Blip_Buffer.
+
+* Simple C interface and usage.
+* Several code examples, including simple sound chip emulator.
+* Uses fast, high-quality band-limited resampling algorithm (BLEP).
+* Output is low-pass and high-pass filtered and clamped to 16-bit range.
+* Supports mono, stereo, and multi-channel synthesis.
+
+Author  : Shay Green <gblargg@gmail.com>
+Website : http://www.slack.net/~ant/
+License : GNU Lesser General Public License (LGPL)
+Language: C
+
+
+Getting Started
+---------------
+Build the demos by typing "make" at the command-line. If that doesn't
+work, manually build a program from demo.c, blip_buf.c, and
+wave_writer.c. Run "demo", which should record a square wave sweep to
+"out.wav". The others demo advanced topics, and their source code can be
+used as a starting point for experimentation (after making any changes,
+just type "make" to do any necessary recompilation).
+
+To build the SDL demo, type "make demo_sdl" at the command-line, or
+manually build a program from demo_sdl.c, blip_buf.c, and the SDL
+multimedia library.
+
+See blip_buf.h for reference and blip_buf.txt for documentation.
+
+
+Files
+-----
+readme.txt            Essential information
+blip_buf.txt          Library documentation
+license.txt           GNU Lesser General Public License
+
+makefile              Builds demos (type "make" at command-line)
+
+demo.c                Generates square wave sweep
+demo_stereo.c         Generates stereo sound using two blip buffers
+demo_fixed.c          Works in fixed-point time rather than clocks
+demo_sdl.c            Plays sound live using SDL multimedia library
+demo_chip.c           Emulates sound hardware and plays back log.txt
+demo_log.txt          Log of sound hardware writes for demo_chip.c
+wave_writer.h         Simple wave sound file writer used by demos
+wave_writer.c
+
+blip_buf.h            Library header and reference (Doxygen-enabled)
+blip_buf.c            source code
+
+tests/                Unit tests (build with "make test")
+
+-- 
+Shay Green <gblargg@gmail.com>
diff --git a/src/debug/GdbCmds.cpp b/src/debug/GdbCmds.cpp
index c2b1f9c48d..1730efcf11 100644
--- a/src/debug/GdbCmds.cpp
+++ b/src/debug/GdbCmds.cpp
@@ -886,14 +886,14 @@ ExecResult GdbStub::Handle_v_MustReplyEmpty(GdbStub* stub, const u8* cmd, ssize_
 
 ExecResult GdbStub::Handle_v_Cont(GdbStub* stub, const u8* cmd, ssize_t len)
 {
-	if (len < 1)
+	if (len < 2)
 	{
 		printf("insufficient length");
 		stub->RespStr("E01");
 		return ExecResult::Ok;
 	}
 
-	switch (cmd[0])
+	switch (cmd[1])
 	{
 	case 'c':
 		stub->RespStr("OK");
@@ -905,7 +905,7 @@ ExecResult GdbStub::Handle_v_Cont(GdbStub* stub, const u8* cmd, ssize_t len)
 		stub->RespStr("OK");
 		return ExecResult::MustBreak;
 	default:
-		printf("invalid continue %c %s\n", cmd[0], cmd);
+		printf("invalid continue %c %s\n", cmd[1], cmd);
 		stub->RespStr("E01");
 		return ExecResult::Ok;
 	}
diff --git a/src/debug/GdbProto.cpp b/src/debug/GdbProto.cpp
index 33f11dd90f..ece2aa5db0 100644
--- a/src/debug/GdbProto.cpp
+++ b/src/debug/GdbProto.cpp
@@ -8,13 +8,13 @@
 #include <assert.h>
 #include <stdlib.h>
 #include <stdio.h>
-#include <unistd.h>
 #include <string.h>
 #include <sys/types.h>
 #include <errno.h>
 #include <algorithm>
 
 #ifndef _WIN32
+#include <unistd.h>
 #include <poll.h>
 #include <sys/select.h>
 #include <sys/socket.h>
diff --git a/src/debug/GdbStub.cpp b/src/debug/GdbStub.cpp
index 3bc54b1d68..b1dcea4daa 100644
--- a/src/debug/GdbStub.cpp
+++ b/src/debug/GdbStub.cpp
@@ -8,7 +8,6 @@
 #include <stdarg.h>
 #include <stdlib.h>
 #include <stdio.h>
-#include <unistd.h>
 #include <string.h>
 #include <sys/types.h>
 
@@ -20,6 +19,10 @@
 #include <arpa/inet.h>
 #include <netinet/in.h>
 #include <netinet/tcp.h>
+#include <unistd.h>
+// actually do the define for unix platforms rather than windows so that
+// it's clearer what the code does
+#define closesocket(x) close(x)
 #endif
 
 
@@ -139,11 +142,7 @@ bool GdbStub::Init(int port)
 err:
 	if (SockFd != 0)
 	{
-#ifdef _WIN32
 		closesocket(SockFd);
-#else
-		close(SockFd);
-#endif
 		SockFd = 0;
 	}
 
@@ -153,13 +152,13 @@ bool GdbStub::Init(int port)
 void GdbStub::Close()
 {
 	Disconnect();
-	if (SockFd > 0) close(SockFd);
+	if (SockFd > 0) closesocket(SockFd);
 	SockFd = 0;
 }
 
 void GdbStub::Disconnect()
 {
-	if (ConnFd > 0) close(ConnFd);
+	if (ConnFd > 0) closesocket(ConnFd);
 	ConnFd = 0;
 }
 
@@ -318,7 +317,7 @@ StubState GdbStub::Poll(bool wait)
 		if (WaitAckBlocking(&a, 1000) < 0)
 		{
 			Log(LogLevel::Error, "[GDB] inital handshake: didn't receive inital ack!\n");
-			close(ConnFd);
+			closesocket(ConnFd);
 			ConnFd = 0;
 			return StubState::Disconnect;
 		}
@@ -404,7 +403,7 @@ StubState GdbStub::Poll(bool wait)
 	case_gdbp_eof:
 	case ReadResult::Eof:
 		Log(LogLevel::Info, "[GDB] EOF!\n");
-		close(ConnFd);
+		closesocket(ConnFd);
 		ConnFd = 0;
 		return StubState::Disconnect;
 	case ReadResult::CksumErr:
diff --git a/src/dolphin/x64CPUDetect.cpp b/src/dolphin/x64CPUDetect.cpp
index 2339048646..3b4b3b528c 100644
--- a/src/dolphin/x64CPUDetect.cpp
+++ b/src/dolphin/x64CPUDetect.cpp
@@ -8,7 +8,9 @@
 #include "CPUDetect.h"
 #include "../types.h"
 
-#ifndef _MSVC_VER
+using namespace melonDS;
+
+#ifndef _MSC_VER
 
 #ifdef __FreeBSD__
 #include <unistd.h>
@@ -17,8 +19,6 @@
 #include <sys/types.h>
 #endif
 
-using namespace melonDS;
-
 static inline void __cpuidex(int info[4], int function_id, int subfunction_id)
 {
 #ifdef __FreeBSD__
@@ -40,7 +40,7 @@ static inline void __cpuid(int info[4], int function_id)
 
 #endif  // ifndef _WIN32
 
-#ifdef _MSVC_VER
+#ifdef _MSC_VER
 
 static u64 xgetbv(u32 index)
 {
diff --git a/src/frontend/qt_sdl/CMakeLists.txt b/src/frontend/qt_sdl/CMakeLists.txt
index 1afa856f1f..db2d413ea7 100644
--- a/src/frontend/qt_sdl/CMakeLists.txt
+++ b/src/frontend/qt_sdl/CMakeLists.txt
@@ -11,6 +11,7 @@ set(SOURCES_QT_SDL
     EmuInstanceAudio.cpp
     EmuInstanceInput.cpp
     EmuThread.cpp
+    CheatImportDialog.cpp
     CheatsDialog.cpp
     Config.cpp
     DateTimeDialog.cpp
@@ -109,6 +110,13 @@ if (WIN32)
     if (MINGW AND CMAKE_HOST_SYSTEM_NAME STREQUAL "Linux")
         set_property(TARGET melonDS PROPERTY AUTORCC_OPTIONS "--no-zstd")
     endif()
+
+    if (BUILD_STATIC)
+        find_package(SDL2 CONFIG COMPONENTS SDL2main)
+        if (TARGET SDL2::SDL2main)
+            target_link_libraries(melonDS PRIVATE SDL2::SDL2main)
+        endif()
+    endif()
 elseif (APPLE)
     if (NOT USE_QT6)
         find_library(COCOA_LIB Cocoa)
@@ -205,10 +213,19 @@ if (WIN32)
         string(TOLOWER "${ARCHITECTURE}" WIN32_ARCHITECTURE)
     endif()
 
-    configure_file("${CMAKE_SOURCE_DIR}/res/xp.manifest.in" "${CMAKE_BINARY_DIR}/res/xp.manifest")
 
     target_link_libraries(melonDS PRIVATE ws2_32 iphlpapi)
-    set_target_properties(melonDS PROPERTIES LINK_FLAGS_DEBUG "-mconsole")
+    set_target_properties(melonDS PROPERTIES WIN32_EXECUTABLE $<IF:$<CONFIG:Debug>,OFF,ON>)
+
+    if (MINGW)
+        configure_file("${CMAKE_SOURCE_DIR}/res/xp.manifest.in" "${CMAKE_BINARY_DIR}/res/xp.manifest")
+    else()
+        # Work around faad2 unconditionally linking libm even on Windows
+        if (USE_VCPKG)
+            add_library(m STATIC)
+            target_link_libraries(melonDS PUBLIC m)
+        endif()
+    endif()
 endif()
 
 if (APPLE)
diff --git a/src/frontend/qt_sdl/CheatImportDialog.cpp b/src/frontend/qt_sdl/CheatImportDialog.cpp
new file mode 100644
index 0000000000..4ff71b855b
--- /dev/null
+++ b/src/frontend/qt_sdl/CheatImportDialog.cpp
@@ -0,0 +1,366 @@
+/*
+    Copyright 2016-2025 melonDS team
+
+    This file is part of melonDS.
+
+    melonDS is free software: you can redistribute it and/or modify it under
+    the terms of the GNU General Public License as published by the Free
+    Software Foundation, either version 3 of the License, or (at your option)
+    any later version.
+
+    melonDS is distributed in the hope that it will be useful, but WITHOUT ANY
+    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+    FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with melonDS. If not, see http://www.gnu.org/licenses/.
+*/
+
+#include <QtGlobal>
+#include <QFileDialog>
+#include <QMessageBox>
+#include <QStandardItemModel>
+
+#include "types.h"
+#include "Platform.h"
+#include "Config.h"
+#include "EmuInstance.h"
+#include "ARDatabaseDAT.h"
+
+#include "CheatImportDialog.h"
+#include "ui_CheatImportDialog.h"
+
+using namespace melonDS;
+using Platform::Log;
+using Platform::LogLevel;
+
+
+CheatImportDialog::CheatImportDialog(QWidget *parent, ARDatabaseDAT* db, u32 gamecode, u32 checksum)
+: QDialog(parent), ui(new Ui::CheatImportDialog), database(db), gameCode(gamecode), gameChecksum(checksum)
+{
+    ui->setupUi(this);
+    setAttribute(Qt::WA_DeleteOnClose);
+
+    dbEntriesByGameCode = database->GetEntriesByGameCode(gameCode);
+    assert(!dbEntriesByGameCode.empty());
+
+    hasChecksumMatches = false;
+    for (auto& entry : dbEntriesByGameCode)
+    {
+        if (entry.Checksum == gameChecksum)
+        {
+            hasChecksumMatches = true;
+            break;
+        }
+    }
+
+    if (hasChecksumMatches)
+    {
+        ui->chkShowAllMatches->setEnabled(true);
+        ui->chkShowAllMatches->setCheckState(Qt::Unchecked);
+    }
+    else
+    {
+        ui->chkShowAllMatches->setEnabled(false);
+        ui->chkShowAllMatches->setCheckState(Qt::Checked);
+    }
+
+    ui->chkRemoveOld->setCheckState(Qt::Checked);
+
+    auto dbname = database->GetDBName();
+    ui->lblDatabaseName->setText(QString::fromStdString(dbname));
+
+    ui->gbCodeInfo->hide();
+    ui->vlRightPanel->addStretch();
+
+    QImage blank(64, 64, QImage::Format_Alpha8);
+    blank.fill(0);
+    blankIcon = QIcon(QPixmap::fromImage(blank.copy()));
+
+    matchIcon = style()->standardIcon(QStyle::SP_DialogApplyButton);
+
+    updatingImportChk = true;
+    auto treemodel = new QStandardItemModel(ui->tvCheatList);
+    ui->tvCheatList->setModel(treemodel);
+    connect(treemodel, &QStandardItemModel::itemChanged, this, &CheatImportDialog::onCheatEntryModified);
+    connect(ui->tvCheatList->selectionModel(), &QItemSelectionModel::selectionChanged, this, &CheatImportDialog::onCheatSelectionChanged);
+
+    populateEntryList();
+    updatingImportChk = false;
+}
+
+CheatImportDialog::~CheatImportDialog()
+{
+    delete ui;
+}
+
+ARDatabaseEntry& CheatImportDialog::getImportCheats()
+{
+    QVariant data = ui->cbEntryList->currentData();
+    auto entry = data.value<ARDatabaseEntry*>();
+    return *entry;
+}
+
+ARCodeEnableMap& CheatImportDialog::getImportEnableMap()
+{
+    return importEnableMap;
+}
+
+bool CheatImportDialog::getRemoveOldCodes()
+{
+    return (ui->chkRemoveOld->checkState() == Qt::Checked);
+}
+
+void CheatImportDialog::on_chkShowAllMatches_clicked(bool checked)
+{
+    populateEntryList();
+}
+
+void CheatImportDialog::on_cbEntryList_currentIndexChanged(int idx)
+{
+    populateEntryInfo();
+}
+
+void CheatImportDialog::onCheatSelectionChanged(const QItemSelection &sel, const QItemSelection &desel)
+{
+    populateCheatInfo();
+}
+
+void CheatImportDialog::onCheatEntryModified(QStandardItem* item)
+{
+    int itemtype = item->data(Qt::UserRole).toInt();
+    if (itemtype == 2)
+    {
+        // sync up the enable map
+        auto code = item->data(Qt::UserRole+1).value<ARCode*>();
+        importEnableMap[code] = (item->checkState() == Qt::Checked);
+    }
+
+    if (updatingImportChk) return;
+    updatingImportChk = true;
+
+    if (item->hasChildren())
+    {
+        // if this item is a category, we need to check/uncheck all the children
+
+        Qt::CheckState chk = item->checkState();
+        if (chk == Qt::PartiallyChecked)
+        {
+            updatingImportChk = false;
+            return;
+        }
+
+        int rows = item->rowCount();
+        for (int i = 0; i < rows; i++)
+            item->child(i)->setCheckState(chk);
+    }
+    else
+    {
+        // if this item is a code, we need to update the check on the parent
+
+        QStandardItem* parent = item->parent();
+        QStandardItem* root = ((QStandardItemModel*)ui->tvCheatList->model())->invisibleRootItem();
+        if ((!parent) || (parent == root))
+        {
+            updatingImportChk = false;
+            return;
+        }
+
+        bool allyes = true;
+        bool allno = true;
+        int rows = parent->rowCount();
+        for (int i = 0; i < rows; i++)
+        {
+            Qt::CheckState chk = parent->child(i)->checkState();
+            if (chk != Qt::Checked) allyes = false;
+            if (chk != Qt::Unchecked) allno = false;
+        }
+
+        if (allyes)     parent->setCheckState(Qt::Checked);
+        else if (allno) parent->setCheckState(Qt::Unchecked);
+        else            parent->setCheckState(Qt::PartiallyChecked);
+    }
+
+    updatingImportChk = false;
+}
+
+void CheatImportDialog::populateEntryList()
+{
+    ui->cbEntryList->clear();
+
+    bool showall = ui->chkShowAllMatches->isChecked();
+    int i = 0, sel = -1;
+    for (auto& entry : dbEntriesByGameCode)
+    {
+        bool crcmatch = (entry.Checksum == gameChecksum);
+        if ((!showall) && (!crcmatch))
+            continue;
+
+        QIcon& icon = crcmatch ? matchIcon : blankIcon;
+        QString name = QString::fromStdString(entry.Name);
+        ui->cbEntryList->addItem(icon, name, QVariant::fromValue(&entry));
+
+        if (crcmatch && (sel == -1))
+            sel = i;
+        i++;
+    }
+
+    if (sel != -1)
+        ui->cbEntryList->setCurrentIndex(sel);
+
+    populateEntryInfo();
+}
+
+void CheatImportDialog::populateEntryInfo()
+{
+    int idx = ui->cbEntryList->currentIndex();
+    if ((idx < 0) || (idx >= ui->cbEntryList->count()))
+    {
+        // nothing selected
+
+        ui->gbEntryInfo->hide();
+        return;
+    }
+
+    ui->gbEntryInfo->show();
+
+    QVariant data = ui->cbEntryList->currentData();
+    auto entry = data.value<ARDatabaseEntry*>();
+
+    ui->lblEntryName->setText(QString::fromStdString(entry->Name));
+
+    char gamecode[5] = {0};
+    memcpy(gamecode, &entry->GameCode, 4);
+    ui->lblEntryGameCode->setText(gamecode);
+
+    auto chksum = QString::asprintf("0x%08X (ROM: 0x%08X)", entry->Checksum, gameChecksum);
+    ui->lblEntryChecksum->setText(chksum);
+
+    populateCheatList();
+}
+
+void CheatImportDialog::populateCheatListCat(QStandardItem* parentitem, ARCodeCat& parentcat)
+{
+    for (auto& item : parentcat.Children)
+    {
+        if (std::holds_alternative<ARCodeCat>(item))
+        {
+            auto& cat = std::get<ARCodeCat>(item);
+
+            QString catname = QString::fromStdString(cat.Name);
+
+            auto catitem = new QStandardItem(catname);
+            parentitem->appendRow(catitem);
+
+            catitem->setData(1, Qt::UserRole);
+            catitem->setData(QVariant::fromValue(&cat), Qt::UserRole+1);
+            catitem->setCheckable(true);
+            catitem->setCheckState(Qt::Checked);
+
+            populateCheatListCat(catitem, cat);
+        }
+        else
+        {
+            auto& code = std::get<ARCode>(item);
+
+            QString codename = QString::fromStdString(code.Name);
+
+            auto codeitem = new QStandardItem(codename);
+            parentitem->appendRow(codeitem);
+
+            codeitem->setData(2, Qt::UserRole);
+            codeitem->setData(QVariant::fromValue(&code), Qt::UserRole+1);
+            codeitem->setCheckable(true);
+            codeitem->setCheckState(Qt::Checked);
+
+            importEnableMap[&code] = true;
+        }
+    }
+}
+
+void CheatImportDialog::populateCheatList()
+{
+    updatingImportChk = true;
+
+    auto treemodel = (QStandardItemModel*)ui->tvCheatList->model();
+    treemodel->clear();
+    importEnableMap.clear();
+
+    int idx = ui->cbEntryList->currentIndex();
+    if ((idx < 0) || (idx >= ui->cbEntryList->count()))
+    {
+        updatingImportChk = false;
+        populateCheatInfo();
+        return;
+    }
+
+    QStandardItem* rootitem = treemodel->invisibleRootItem();
+
+    QVariant data = ui->cbEntryList->currentData();
+    auto entry = data.value<ARDatabaseEntry*>();
+    populateCheatListCat(rootitem, entry->RootCat);
+
+    updatingImportChk = false;
+    populateCheatInfo();
+}
+
+void CheatImportDialog::populateCheatInfo()
+{
+    auto selmodel = ui->tvCheatList->selectionModel();
+    if (!selmodel->hasSelection())
+    {
+        ui->gbCodeInfo->hide();
+        return;
+    }
+
+    auto index = selmodel->selectedIndexes()[0];
+    int itemtype = index.data(Qt::UserRole).toInt();
+    if (itemtype == 1)
+    {
+        // category
+
+        ui->gbCodeInfo->show();
+        ui->gbCodeInfo->setTitle("Selected category");
+
+        auto cat = index.data(Qt::UserRole+1).value<ARCodeCat*>();
+        QString catname = QString::fromStdString(cat->Name);
+        QString catdesc = QString::fromStdString(cat->Description);
+        QString codeenable = cat->OnlyOneCodeEnabled ? "Only one" : "Multiple";
+
+        ui->lblCodeName->setText(catname);
+        ui->lblCodeDesc->setText(catdesc);
+        ui->lblCodeStatusLabel->setText("Code enable:");
+        ui->lblCodeStatus->setText(codeenable);
+
+        ui->lblCodeCodeLabel->hide();
+        ui->txtCodeCode->hide();
+    }
+    else if (itemtype == 2)
+    {
+        // cheat
+
+        ui->gbCodeInfo->show();
+        ui->gbCodeInfo->setTitle("Selected code");
+
+        auto code = index.data(Qt::UserRole+1).value<ARCode*>();
+        QString codename = QString::fromStdString(code->Name);
+        QString codedesc = QString::fromStdString(code->Description);
+        QString codeenable = code->Enabled ? "Enabled" : "Disabled";
+
+        ui->lblCodeName->setText(codename);
+        ui->lblCodeDesc->setText(codedesc);
+        ui->lblCodeStatusLabel->setText("Status:");
+        ui->lblCodeStatus->setText(codeenable);
+
+        QString codecode = "";
+        for (int i = 0; i < code->Code.size(); i+=2)
+        {
+            if (i > 0) codecode += "\n";
+            codecode += QString::asprintf("%08X %08X", code->Code[i], code->Code[i+1]);
+        }
+
+        ui->lblCodeCodeLabel->show();
+        ui->txtCodeCode->show();
+        ui->txtCodeCode->setPlainText(codecode);
+    }
+}
diff --git a/src/frontend/qt_sdl/CheatImportDialog.h b/src/frontend/qt_sdl/CheatImportDialog.h
new file mode 100644
index 0000000000..79698754e7
--- /dev/null
+++ b/src/frontend/qt_sdl/CheatImportDialog.h
@@ -0,0 +1,73 @@
+/*
+    Copyright 2016-2025 melonDS team
+
+    This file is part of melonDS.
+
+    melonDS is free software: you can redistribute it and/or modify it under
+    the terms of the GNU General Public License as published by the Free
+    Software Foundation, either version 3 of the License, or (at your option)
+    any later version.
+
+    melonDS is distributed in the hope that it will be useful, but WITHOUT ANY
+    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+    FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with melonDS. If not, see http://www.gnu.org/licenses/.
+*/
+
+#ifndef CHEATIMPORTDIALOG_H
+#define CHEATIMPORTDIALOG_H
+
+#include <QDialog>
+#include <QAbstractListModel>
+#include <QStandardItem>
+#include <QItemSelection>
+
+#include "ARCodeFile.h"
+#include "ARDatabaseDAT.h"
+
+namespace Ui { class CheatImportDialog; }
+class CheatImportDialog;
+
+class CheatImportDialog : public QDialog
+{
+    Q_OBJECT
+
+public:
+    explicit CheatImportDialog(QWidget* parent, melonDS::ARDatabaseDAT* db, melonDS::u32 gamecode, melonDS::u32 checksum);
+    ~CheatImportDialog();
+
+    melonDS::ARDatabaseEntry& getImportCheats();
+    melonDS::ARCodeEnableMap& getImportEnableMap();
+    bool getRemoveOldCodes();
+
+private slots:
+    void on_chkShowAllMatches_clicked(bool checked);
+    void on_cbEntryList_currentIndexChanged(int idx);
+    void onCheatSelectionChanged(const QItemSelection& sel, const QItemSelection& desel);
+    void onCheatEntryModified(QStandardItem* item);
+
+private:
+    Ui::CheatImportDialog* ui;
+    QIcon blankIcon, matchIcon;
+
+    melonDS::ARDatabaseDAT* database;
+    melonDS::u32 gameCode;
+    melonDS::u32 gameChecksum;
+
+    melonDS::ARDatabaseEntryList dbEntriesByGameCode;
+    bool hasChecksumMatches;
+
+    bool updatingImportChk;
+
+    melonDS::ARCodeEnableMap importEnableMap;
+
+    void populateEntryList();
+    void populateEntryInfo();
+    void populateCheatListCat(QStandardItem* parentitem, melonDS::ARCodeCat& parentcat);
+    void populateCheatList();
+    void populateCheatInfo();
+};
+
+#endif // CHEATIMPORTDIALOG_H
diff --git a/src/frontend/qt_sdl/CheatImportDialog.ui b/src/frontend/qt_sdl/CheatImportDialog.ui
new file mode 100644
index 0000000000..23093d757f
--- /dev/null
+++ b/src/frontend/qt_sdl/CheatImportDialog.ui
@@ -0,0 +1,385 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<ui version="4.0">
+ <class>CheatImportDialog</class>
+ <widget class="QDialog" name="CheatImportDialog">
+  <property name="geometry">
+   <rect>
+    <x>0</x>
+    <y>0</y>
+    <width>712</width>
+    <height>555</height>
+   </rect>
+  </property>
+  <property name="windowTitle">
+   <string>Import cheat codes - melonDS</string>
+  </property>
+  <layout class="QVBoxLayout" name="verticalLayout_3">
+   <item>
+    <widget class="QSplitter" name="splitter">
+     <property name="orientation">
+      <enum>Qt::Orientation::Horizontal</enum>
+     </property>
+     <property name="handleWidth">
+      <number>2</number>
+     </property>
+     <property name="childrenCollapsible">
+      <bool>false</bool>
+     </property>
+     <widget class="QWidget" name="layoutWidget">
+      <layout class="QVBoxLayout" name="verticalLayout">
+       <property name="rightMargin">
+        <number>4</number>
+       </property>
+       <item>
+        <widget class="QCheckBox" name="chkShowAllMatches">
+         <property name="text">
+          <string>Show all game code matches</string>
+         </property>
+        </widget>
+       </item>
+       <item>
+        <widget class="QComboBox" name="cbEntryList"/>
+       </item>
+       <item>
+        <widget class="QCheckBox" name="chkRemoveOld">
+         <property name="text">
+          <string>Clear cheat file before import</string>
+         </property>
+        </widget>
+       </item>
+       <item>
+        <widget class="QLabel" name="label_8">
+         <property name="text">
+          <string>Check the cheat codes to import in the list below.</string>
+         </property>
+        </widget>
+       </item>
+       <item>
+        <widget class="QTreeView" name="tvCheatList">
+         <property name="sizePolicy">
+          <sizepolicy hsizetype="Expanding" vsizetype="Expanding">
+           <horstretch>0</horstretch>
+           <verstretch>0</verstretch>
+          </sizepolicy>
+         </property>
+         <property name="editTriggers">
+          <set>QAbstractItemView::EditTrigger::NoEditTriggers</set>
+         </property>
+         <property name="headerHidden">
+          <bool>true</bool>
+         </property>
+        </widget>
+       </item>
+      </layout>
+     </widget>
+     <widget class="QWidget" name="layoutWidget2">
+      <layout class="QVBoxLayout" name="vlRightPanel" stretch="0,0,1">
+       <property name="leftMargin">
+        <number>4</number>
+       </property>
+       <item>
+        <widget class="QGroupBox" name="gbDatabaseInfo">
+         <property name="title">
+          <string>Database</string>
+         </property>
+         <layout class="QVBoxLayout" name="verticalLayout_4">
+          <item>
+           <widget class="QLabel" name="lblDatabaseName">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Expanding" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>super duper cheat database by epic haxor team</string>
+            </property>
+            <property name="textFormat">
+             <enum>Qt::TextFormat::PlainText</enum>
+            </property>
+            <property name="wordWrap">
+             <bool>true</bool>
+            </property>
+           </widget>
+          </item>
+         </layout>
+        </widget>
+       </item>
+       <item>
+        <widget class="QGroupBox" name="gbEntryInfo">
+         <property name="sizePolicy">
+          <sizepolicy hsizetype="Preferred" vsizetype="Preferred">
+           <horstretch>0</horstretch>
+           <verstretch>0</verstretch>
+          </sizepolicy>
+         </property>
+         <property name="title">
+          <string>Selected game</string>
+         </property>
+         <layout class="QFormLayout" name="formLayout">
+          <item row="0" column="0">
+           <widget class="QLabel" name="label_2">
+            <property name="text">
+             <string>Name:</string>
+            </property>
+           </widget>
+          </item>
+          <item row="0" column="1">
+           <widget class="QLabel" name="lblEntryName">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Expanding" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>Super Mario Odyssey DS Edition</string>
+            </property>
+            <property name="textFormat">
+             <enum>Qt::TextFormat::PlainText</enum>
+            </property>
+            <property name="wordWrap">
+             <bool>true</bool>
+            </property>
+           </widget>
+          </item>
+          <item row="1" column="0">
+           <widget class="QLabel" name="label_3">
+            <property name="text">
+             <string>Game code:</string>
+            </property>
+           </widget>
+          </item>
+          <item row="1" column="1">
+           <widget class="QLabel" name="lblEntryGameCode">
+            <property name="text">
+             <string>FART</string>
+            </property>
+            <property name="textFormat">
+             <enum>Qt::TextFormat::PlainText</enum>
+            </property>
+           </widget>
+          </item>
+          <item row="2" column="0">
+           <widget class="QLabel" name="label_5">
+            <property name="text">
+             <string>Checksum:</string>
+            </property>
+           </widget>
+          </item>
+          <item row="2" column="1">
+           <widget class="QLabel" name="lblEntryChecksum">
+            <property name="text">
+             <string>0xDEADBEEF</string>
+            </property>
+            <property name="textFormat">
+             <enum>Qt::TextFormat::PlainText</enum>
+            </property>
+           </widget>
+          </item>
+         </layout>
+        </widget>
+       </item>
+       <item>
+        <widget class="QGroupBox" name="gbCodeInfo">
+         <property name="sizePolicy">
+          <sizepolicy hsizetype="Preferred" vsizetype="Preferred">
+           <horstretch>0</horstretch>
+           <verstretch>0</verstretch>
+          </sizepolicy>
+         </property>
+         <property name="title">
+          <string>Selected code</string>
+         </property>
+         <layout class="QFormLayout" name="formLayout_3">
+          <item row="0" column="0">
+           <widget class="QLabel" name="label">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Preferred" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>Name:</string>
+            </property>
+            <property name="alignment">
+             <set>Qt::AlignmentFlag::AlignLeading|Qt::AlignmentFlag::AlignLeft|Qt::AlignmentFlag::AlignTop</set>
+            </property>
+           </widget>
+          </item>
+          <item row="0" column="1">
+           <widget class="QLabel" name="lblCodeName">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Expanding" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>super duper code</string>
+            </property>
+            <property name="textFormat">
+             <enum>Qt::TextFormat::PlainText</enum>
+            </property>
+            <property name="wordWrap">
+             <bool>true</bool>
+            </property>
+           </widget>
+          </item>
+          <item row="1" column="0">
+           <widget class="QLabel" name="label_6">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Preferred" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>Description:</string>
+            </property>
+            <property name="alignment">
+             <set>Qt::AlignmentFlag::AlignLeading|Qt::AlignmentFlag::AlignLeft|Qt::AlignmentFlag::AlignTop</set>
+            </property>
+           </widget>
+          </item>
+          <item row="1" column="1">
+           <widget class="QLabel" name="lblCodeDesc">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Expanding" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>this code is absolutely epic</string>
+            </property>
+            <property name="textFormat">
+             <enum>Qt::TextFormat::PlainText</enum>
+            </property>
+            <property name="wordWrap">
+             <bool>true</bool>
+            </property>
+           </widget>
+          </item>
+          <item row="2" column="0">
+           <widget class="QLabel" name="lblCodeStatusLabel">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Preferred" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>Status:</string>
+            </property>
+           </widget>
+          </item>
+          <item row="2" column="1">
+           <widget class="QLabel" name="lblCodeStatus">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Preferred" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>enabled</string>
+            </property>
+            <property name="textFormat">
+             <enum>Qt::TextFormat::PlainText</enum>
+            </property>
+           </widget>
+          </item>
+          <item row="3" column="0">
+           <widget class="QLabel" name="lblCodeCodeLabel">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Preferred" vsizetype="Preferred">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="text">
+             <string>Code:</string>
+            </property>
+            <property name="alignment">
+             <set>Qt::AlignmentFlag::AlignLeading|Qt::AlignmentFlag::AlignLeft|Qt::AlignmentFlag::AlignTop</set>
+            </property>
+           </widget>
+          </item>
+          <item row="3" column="1">
+           <widget class="QPlainTextEdit" name="txtCodeCode">
+            <property name="sizePolicy">
+             <sizepolicy hsizetype="Expanding" vsizetype="Expanding">
+              <horstretch>0</horstretch>
+              <verstretch>0</verstretch>
+             </sizepolicy>
+            </property>
+            <property name="font">
+             <font>
+              <family>Noto Sans Mono</family>
+             </font>
+            </property>
+            <property name="sizeAdjustPolicy">
+             <enum>QAbstractScrollArea::SizeAdjustPolicy::AdjustToContents</enum>
+            </property>
+            <property name="readOnly">
+             <bool>true</bool>
+            </property>
+           </widget>
+          </item>
+         </layout>
+        </widget>
+       </item>
+      </layout>
+     </widget>
+    </widget>
+   </item>
+   <item>
+    <widget class="QDialogButtonBox" name="buttonBox">
+     <property name="orientation">
+      <enum>Qt::Orientation::Horizontal</enum>
+     </property>
+     <property name="standardButtons">
+      <set>QDialogButtonBox::StandardButton::Cancel|QDialogButtonBox::StandardButton::Ok</set>
+     </property>
+    </widget>
+   </item>
+  </layout>
+ </widget>
+ <resources/>
+ <connections>
+  <connection>
+   <sender>buttonBox</sender>
+   <signal>accepted()</signal>
+   <receiver>CheatImportDialog</receiver>
+   <slot>accept()</slot>
+   <hints>
+    <hint type="sourcelabel">
+     <x>248</x>
+     <y>254</y>
+    </hint>
+    <hint type="destinationlabel">
+     <x>157</x>
+     <y>274</y>
+    </hint>
+   </hints>
+  </connection>
+  <connection>
+   <sender>buttonBox</sender>
+   <signal>rejected()</signal>
+   <receiver>CheatImportDialog</receiver>
+   <slot>reject()</slot>
+   <hints>
+    <hint type="sourcelabel">
+     <x>316</x>
+     <y>260</y>
+    </hint>
+    <hint type="destinationlabel">
+     <x>286</x>
+     <y>274</y>
+    </hint>
+   </hints>
+  </connection>
+ </connections>
+</ui>
diff --git a/src/frontend/qt_sdl/CheatsDialog.cpp b/src/frontend/qt_sdl/CheatsDialog.cpp
index 75279a2afd..ec906e7a73 100644
--- a/src/frontend/qt_sdl/CheatsDialog.cpp
+++ b/src/frontend/qt_sdl/CheatsDialog.cpp
@@ -16,18 +16,19 @@
     with melonDS. If not, see http://www.gnu.org/licenses/.
 */
 
-#include <stdio.h>
 #include <QtGlobal>
 #include <QFileDialog>
 #include <QMessageBox>
+#include <QMimeData>
 
 #include "types.h"
 #include "Platform.h"
-#include "Config.h"
 #include "EmuInstance.h"
+#include "ARDatabaseDAT.h"
 
 #include "CheatsDialog.h"
 #include "ui_CheatsDialog.h"
+#include "CheatImportDialog.h"
 
 using namespace melonDS;
 using Platform::Log;
@@ -41,71 +42,43 @@ CheatsDialog::CheatsDialog(QWidget* parent) : QDialog(parent), ui(new Ui::Cheats
     ui->setupUi(this);
     setAttribute(Qt::WA_DeleteOnClose);
 
+    importDB = nullptr;
+    importDlg = nullptr;
+    updatingEnableChk = false;
+
+    ui->splitter->setStretchFactor(0, 2);
+    ui->splitter->setStretchFactor(1, 3);
+
     emuInstance = ((MainWindow*)parent)->getEmuInstance();
 
+    auto rom = emuInstance->getNDS()->NDSCartSlot.GetCart();
+    gameCode = rom->GetHeader().GameCodeAsU32();
+    gameChecksum = ~CRC32(rom->GetROM(), 0x200, 0);
+
     codeFile = emuInstance->getCheatFile();
 
-    QStandardItemModel* model = new QStandardItemModel();
+    auto* model = new CheatListModel(ui->tvCodeList);
     ui->tvCodeList->setModel(model);
     connect(model, &QStandardItemModel::itemChanged, this, &CheatsDialog::onCheatEntryModified);
     connect(ui->tvCodeList->selectionModel(), &QItemSelectionModel::selectionChanged, this, &CheatsDialog::onCheatSelectionChanged);
 
-    {
-        QStandardItem* root = model->invisibleRootItem();
-
-        for (ARCodeCatList::iterator i = codeFile->Categories.begin(); i != codeFile->Categories.end(); i++)
-        {
-            ARCodeCat& cat = *i;
-
-            QStandardItem* catitem = new QStandardItem(QString::fromStdString(cat.Name));
-            catitem->setEditable(true);
-            catitem->setData(QVariant::fromValue(i));
-            root->appendRow(catitem);
+    populateCheatList();
 
-            for (ARCodeList::iterator j = cat.Codes.begin(); j != cat.Codes.end(); j++)
-            {
-                ARCode& code = *j;
-
-                QStandardItem* codeitem = new QStandardItem(QString::fromStdString(code.Name));
-                codeitem->setEditable(true);
-                codeitem->setCheckable(true);
-                codeitem->setCheckState(code.Enabled ? Qt::Checked : Qt::Unchecked);
-                codeitem->setData(QVariant::fromValue(j));
-                catitem->appendRow(codeitem);
-            }
-        }
-    }
-
-    ui->txtCode->setPlaceholderText("");
     codeChecker = new ARCodeChecker(ui->txtCode->document());
-
-    ui->btnNewARCode->setEnabled(false);
-    ui->btnDeleteCode->setEnabled(false);
-    ui->txtCode->setEnabled(false);
 }
 
 CheatsDialog::~CheatsDialog()
 {
-    QAbstractItemModel* model = ui->tvCodeList->model();
-    ui->tvCodeList->setModel(nullptr);
-    delete model;
-
     delete codeChecker;
 
     delete ui;
 }
 
-void CheatsDialog::on_CheatsDialog_accepted()
+void CheatsDialog::done(int r)
 {
     codeFile->Save();
 
-    closeDlg();
-}
-
-void CheatsDialog::on_CheatsDialog_rejected()
-{
-    codeFile->Load();
-
+    QDialog::done(r);
     closeDlg();
 }
 
@@ -113,70 +86,72 @@ void CheatsDialog::on_btnNewCat_clicked()
 {
     QStandardItem* root = ((QStandardItemModel*)ui->tvCodeList->model())->invisibleRootItem();
 
+    ARCodeCat& parentcat = codeFile->RootCat;
+
     ARCodeCat cat;
-    cat.Codes.clear();
+    cat.Parent = &parentcat;
     cat.Name = "(new category)";
+    cat.OnlyOneCodeEnabled = false;
 
-    codeFile->Categories.push_back(cat);
-    ARCodeCatList::iterator id = codeFile->Categories.end(); id--;
+    parentcat.Children.emplace_back(cat);
+    auto catptr = &std::get<ARCodeCat>(parentcat.Children.back());
 
-    QStandardItem* catitem = new QStandardItem(QString::fromStdString(cat.Name));
-    catitem->setEditable(true);
-    catitem->setData(QVariant::fromValue(id));
+    auto catitem = new QStandardItem(QString::fromStdString(cat.Name));
+    catitem->setDropEnabled(true);
+    catitem->setData(1, Qt::UserRole);
+    catitem->setData(QVariant::fromValue(catptr), Qt::UserRole+1);
     root->appendRow(catitem);
 
     ui->tvCodeList->selectionModel()->select(catitem->index(), QItemSelectionModel::ClearAndSelect);
-    ui->tvCodeList->edit(catitem->index());
+    ui->tvCodeList->scrollTo(catitem->index());
+
+    ui->btnEditCode->click();
 }
 
 void CheatsDialog::on_btnNewARCode_clicked()
 {
-    QModelIndexList indices = ui->tvCodeList->selectionModel()->selectedIndexes();
-    if (indices.isEmpty())
-    {
-        // ????
-        return;
-    }
-
-    QStandardItemModel* model = (QStandardItemModel*)ui->tvCodeList->model();
-    QStandardItem* item = model->itemFromIndex(indices.first());
-    QStandardItem* parentitem;
+    auto model = (QStandardItemModel*)ui->tvCodeList->model();
+    auto selmodel = ui->tvCodeList->selectionModel();
+    QStandardItem* parentitem = nullptr;
+    ARCodeCat* cat;
 
-    QVariant data = item->data();
-    if (data.canConvert<ARCodeCatList::iterator>())
-    {
-        parentitem = item;
-    }
-    else if (data.canConvert<ARCodeList::iterator>())
-    {
-        parentitem = item->parent();
-    }
-    else
+    if (selmodel->hasSelection())
     {
-        Log(LogLevel::Warn, "what?? :(\n");
-        return;
+        QModelIndexList indices = selmodel->selectedIndexes();
+        QStandardItem *item = model->itemFromIndex(indices.first());
+
+        int itemtype = item->data(Qt::UserRole).toInt();
+        if (itemtype == 1)
+            parentitem = item;
+        else if (itemtype == 2)
+            parentitem = item->parent();
+        else
+            assert(false);
     }
 
-    ARCodeCatList::iterator it_cat = parentitem->data().value<ARCodeCatList::iterator>();
-    ARCodeCat& cat = *it_cat;
+    if (!parentitem) parentitem = model->invisibleRootItem();
+    cat = parentitem->data(Qt::UserRole+1).value<ARCodeCat*>();
 
     ARCode code;
+    code.Parent = cat;
     code.Name = "(new AR code)";
-    code.Enabled = true;
-    code.Code.clear();
+    code.Enabled = false;
 
-    cat.Codes.push_back(code);
-    ARCodeList::iterator id = cat.Codes.end(); id--;
+    cat->Children.emplace_back(code);
+    auto codeptr = &std::get<ARCode>(cat->Children.back());
 
-    QStandardItem* codeitem = new QStandardItem(QString::fromStdString(code.Name));
-    codeitem->setEditable(true);
+    auto codeitem = new QStandardItem(QString::fromStdString(code.Name));
     codeitem->setCheckable(true);
     codeitem->setCheckState(code.Enabled ? Qt::Checked : Qt::Unchecked);
-    codeitem->setData(QVariant::fromValue(id));
+    codeitem->setDropEnabled(false);
+    codeitem->setData(2, Qt::UserRole);
+    codeitem->setData(QVariant::fromValue(codeptr), Qt::UserRole+1);
     parentitem->appendRow(codeitem);
 
     ui->tvCodeList->selectionModel()->select(codeitem->index(), QItemSelectionModel::ClearAndSelect);
-    ui->tvCodeList->edit(codeitem->index());
+    ui->tvCodeList->scrollTo(codeitem->index());
+
+    ui->btnEditCode->click();
 }
 
 void CheatsDialog::on_btnDeleteCode_clicked()
@@ -195,127 +170,381 @@ void CheatsDialog::on_btnDeleteCode_clicked()
                                                             QMessageBox::No);
     if (res != QMessageBox::Yes) return;
 
-    QStandardItemModel* model = (QStandardItemModel*)ui->tvCodeList->model();
+    auto model = (QStandardItemModel*)ui->tvCodeList->model();
     QStandardItem* item = model->itemFromIndex(indices.first());
 
-    QVariant data = item->data();
-    if (data.canConvert<ARCodeCatList::iterator>())
+    QStandardItem* itemparent = item->parent();
+    if (!itemparent) itemparent = model->invisibleRootItem();
+
+    int itemtype = item->data(Qt::UserRole).toInt();
+    if (itemtype == 1)
     {
-        ARCodeCatList::iterator it_cat = data.value<ARCodeCatList::iterator>();
+        auto cat = item->data(Qt::UserRole+1).value<ARCodeCat*>();
 
-        (*it_cat).Codes.clear();
-        codeFile->Categories.erase(it_cat);
+        auto parentcat = cat->Parent;
+        for (auto it = parentcat->Children.begin(); it != parentcat->Children.end(); it++)
+        {
+            auto& childitem = *it;
+            if (!std::holds_alternative<ARCodeCat>(childitem))
+                continue;
+            if (&std::get<ARCodeCat>(childitem) != cat)
+                continue;
 
-        model->invisibleRootItem()->removeRow(item->row());
+            parentcat->Children.erase(it);
+            break;
+        }
     }
-    else if (data.canConvert<ARCodeList::iterator>())
+    else if (itemtype == 2)
     {
-        ARCodeList::iterator it_code = data.value<ARCodeList::iterator>();
-        ARCodeCatList::iterator it_cat = item->parent()->data().value<ARCodeCatList::iterator>();
+        auto code = item->data(Qt::UserRole+1).value<ARCode*>();
 
-        (*it_cat).Codes.erase(it_code);
+        auto parentcat = code->Parent;
+        for (auto it = parentcat->Children.begin(); it != parentcat->Children.end(); it++)
+        {
+            auto& childitem = *it;
+            if (!std::holds_alternative<ARCode>(childitem))
+                continue;
+            if (&std::get<ARCode>(childitem) != code)
+                continue;
 
-        item->parent()->removeRow(item->row());
+            parentcat->Children.erase(it);
+            break;
+        }
     }
+    else
+        assert(false);
+
+    itemparent->removeRow(item->row());
 }
 
-void CheatsDialog::onCheatSelectionChanged(const QItemSelection& sel, const QItemSelection& desel)
+void CheatsDialog::on_btnImportCheats_clicked()
 {
-    QModelIndexList indices = sel.indexes();
-    if (indices.isEmpty())
+    QString file = QFileDialog::getOpenFileName(this,
+                                                "Select cheat database...",
+                                                emuDirectory,
+                                                "R4 cheat database (*.dat);;Any file (*.*)");
+
+    if (file.isEmpty()) return;
+
+    importDB = new ARDatabaseDAT(file.toStdString());
+    if (importDB->Error)
     {
-        ui->btnNewARCode->setEnabled(false);
-        ui->btnDeleteCode->setEnabled(false);
-        ui->txtCode->setEnabled(false);
-        ui->txtCode->setPlaceholderText("");
-        ui->txtCode->clear();
+        QMessageBox::critical(this, "melonDS",
+                              "Failed to open this cheat database file.");
+        delete importDB;
+        return;
     }
-    else
+
+    if (!importDB->FindGameCode(gameCode))
     {
-        QStandardItem* item = ((QStandardItemModel*)ui->tvCodeList->model())->itemFromIndex(indices.first());
+        QMessageBox::critical(this, "melonDS",
+                              "No cheat codes were found in this database for the current game.");
+        delete importDB;
+        return;
+    }
 
-        QVariant data = item->data();
-        if (data.canConvert<ARCodeCatList::iterator>())
-        {
-            ui->btnDeleteCode->setEnabled(true);
-            ui->txtCode->setEnabled(false);
-            ui->txtCode->setPlaceholderText("");
-            ui->txtCode->clear();
-        }
-        else if (data.canConvert<ARCodeList::iterator>())
-        {
-            ARCode& code = *(data.value<ARCodeList::iterator>());
+    importDlg = new CheatImportDialog(this, importDB, gameCode, gameChecksum);
+    importDlg->open();
+    connect(importDlg, &CheatImportDialog::finished, this, &CheatsDialog::onImportCheatsFinished);
 
-            ui->btnDeleteCode->setEnabled(true);
-            ui->txtCode->setEnabled(true);
-            ui->txtCode->setPlaceholderText("(enter AR code here)");
+    importDlg->show();
+}
 
-            QString codestr = "";
-            for (size_t i = 0; i < code.Code.size(); i += 2)
-            {
-                u32 c0 = code.Code[i+0];
-                u32 c1 = code.Code[i+1];
-                //codestr += QString("%1 %2\n").arg(c0, 8, 16, '0').arg(c1, 8, 16, '0').toUpper();
-                codestr += QString::asprintf("%08X %08X\n", c0, c1);
-            }
-            ui->txtCode->setPlainText(codestr);
-        }
+void CheatsDialog::onImportCheatsFinished(int res)
+{
+    if (!importDlg) return;
+
+    if (res == QDialog::Accepted)
+    {
+        auto& cheats = importDlg->getImportCheats();
+        auto& enablemap = importDlg->getImportEnableMap();
+        auto removeold = importDlg->getRemoveOldCodes();
 
-        ui->btnNewARCode->setEnabled(true);
+        codeFile->Import(cheats, enablemap, removeold);
+
+        populateCheatList();
     }
+
+    delete importDB;
+    importDB = nullptr;
+    importDlg = nullptr;
+}
+
+void CheatsDialog::onCheatSelectionChanged(const QItemSelection& sel, const QItemSelection& desel)
+{
+    populateCheatInfo();
 }
 
 void CheatsDialog::onCheatEntryModified(QStandardItem* item)
 {
-    QVariant data = item->data();
-    if (data.canConvert<ARCodeCatList::iterator>())
+    if (updatingEnableChk) return;
+
+    auto selmodel = ui->tvCodeList->selectionModel();
+    int itemtype = item->data(Qt::UserRole).toInt();
+    if (itemtype == 2)
     {
-        ARCodeCat& cat = *(data.value<ARCodeCatList::iterator>());
+        auto code = item->data(Qt::UserRole+1).value<ARCode*>();
 
-        if (item->text().isEmpty())
+        code->Enabled = (item->checkState() == Qt::Checked);
+
+        if (code->Enabled && code->Parent->OnlyOneCodeEnabled)
         {
-            QString oldname = QString::fromStdString(cat.Name);
-            item->setText(oldname.isEmpty() ? "(blank category name?)" : oldname);
+            // if needed, disable other codes
+
+            updatingEnableChk = true;
+
+            auto parent = item->parent();
+            for (int i = 0; i < parent->rowCount(); i++)
+            {
+                auto child = parent->child(i);
+                if (child == item) continue;
+
+                int childtype = child->data(Qt::UserRole).toInt();
+                if (childtype == 2)
+                {
+                    auto childcode = child->data(Qt::UserRole+1).value<ARCode*>();
+                    childcode->Enabled = false;
+
+                    child->setCheckState(Qt::Unchecked);
+
+                    if (selmodel->hasSelection())
+                    {
+                        auto index = selmodel->selectedIndexes()[0];
+                        if (index == child->index())
+                        {
+                            ui->chkItemOption->setChecked(childcode->Enabled);
+                        }
+                    }
+                }
+            }
+
+            updatingEnableChk = false;
         }
-        else
+
+        // if this item is selected, reflect the checkbox change
+        if (selmodel->hasSelection())
         {
-            cat.Name = item->text().toStdString();
+            auto index = selmodel->selectedIndexes()[0];
+            if (index == item->index())
+            {
+                ui->chkItemOption->setChecked(code->Enabled);
+            }
         }
     }
-    else if (data.canConvert<ARCodeList::iterator>())
+}
+
+void CheatsDialog::on_btnEditCode_clicked()
+{
+    ui->btnNewCat->setEnabled(false);
+    ui->btnNewARCode->setEnabled(false);
+    ui->btnImportCheats->setEnabled(false);
+    ui->tvCodeList->setEnabled(false);
+
+    ui->btnEditCode->hide();
+    ui->btnDeleteCode->setEnabled(false);
+    ui->btnSaveCode->show();
+    ui->btnCancelEdit->show();
+
+    ui->txtItemName->setReadOnly(false);
+    ui->txtItemDesc->setReadOnly(false);
+    ui->chkItemOption->setEnabled(true);
+    ui->txtCode->setReadOnly(false);
+}
+
+void CheatsDialog::on_btnSaveCode_clicked()
+{
+    auto selmodel = ui->tvCodeList->selectionModel();
+    assert(selmodel->hasSelection());
+    auto index = selmodel->selectedIndexes()[0];
+
+    if (ui->txtItemName->text().trimmed().isEmpty())
+    {
+        QMessageBox::critical(this, "melonDS", "Error: no name entered.");
+        return;
+    }
+
+    int itemtype = index.data(Qt::UserRole).toInt();
+    if (itemtype == 1)
     {
-        ARCode& code = *(data.value<ARCodeList::iterator>());
+        auto cat = index.data(Qt::UserRole+1).value<ARCodeCat*>();
 
-        if (item->text().isEmpty())
-        {
-            QString oldname = QString::fromStdString(code.Name);
-            item->setText(oldname.isEmpty() ? "(blank code name?)" : oldname);
-        }
-        else
+        cat->Name = ui->txtItemName->text().toStdString();
+        cat->Description = ui->txtItemDesc->text().toStdString();
+        cat->OnlyOneCodeEnabled = ui->chkItemOption->isChecked();
+
+        auto tvmodel = (QStandardItemModel*)ui->tvCodeList->model();
+        auto tvitem = tvmodel->itemFromIndex(index);
+        tvitem->setText(ui->txtItemName->text());
+    }
+    else if (itemtype == 2)
+    {
+        auto code = index.data(Qt::UserRole+1).value<ARCode*>();
+
+        auto codeconv = convertCodeInput();
+        if (codeconv.empty())
         {
-            code.Name = item->text().toStdString();
+            QMessageBox::critical(this, "melonDS", "Error: the code entered is empty or invalid.");
+            return;
         }
 
-        code.Enabled = (item->checkState() == Qt::Checked);
+        code->Name = ui->txtItemName->text().toStdString();
+        code->Description = ui->txtItemDesc->text().toStdString();
+        code->Enabled = ui->chkItemOption->isChecked();
+        code->Code = codeconv;
+
+        //updatingEnableChk = true;
+
+        auto tvmodel = (QStandardItemModel*)ui->tvCodeList->model();
+        auto tvitem = tvmodel->itemFromIndex(index);
+        tvitem->setText(ui->txtItemName->text());
+        tvitem->setCheckState(code->Enabled ? Qt::Checked : Qt::Unchecked);
+
+        //updatingEnableChk = false;
     }
+    else
+        assert(false);
+
+    ui->btnNewCat->setEnabled(true);
+    ui->btnNewARCode->setEnabled(true);
+    ui->btnImportCheats->setEnabled(true);
+    ui->tvCodeList->setEnabled(true);
+
+    ui->btnEditCode->show();
+    ui->btnDeleteCode->setEnabled(true);
+    ui->btnSaveCode->hide();
+    ui->btnCancelEdit->hide();
+
+    populateCheatInfo();
 }
 
-void CheatsDialog::on_txtCode_textChanged()
+void CheatsDialog::on_btnCancelEdit_clicked()
 {
-    QModelIndexList indices = ui->tvCodeList->selectionModel()->selectedIndexes();
-    if (indices.isEmpty())
-        return;
+    ui->btnNewCat->setEnabled(true);
+    ui->btnNewARCode->setEnabled(true);
+    ui->btnImportCheats->setEnabled(true);
+    ui->tvCodeList->setEnabled(true);
+
+    ui->btnEditCode->show();
+    ui->btnDeleteCode->setEnabled(true);
+    ui->btnSaveCode->hide();
+    ui->btnCancelEdit->hide();
+
+    populateCheatInfo();
+}
+
+void CheatsDialog::populateCheatList()
+{
+    auto model = (CheatListModel*)ui->tvCodeList->model();
+    model->clear();
+
+    QStandardItem* root = model->invisibleRootItem();
+    root->setData(0, Qt::UserRole);
+    root->setData(QVariant::fromValue(&codeFile->RootCat), Qt::UserRole+1);
+    model->populateCheatListCat(root, codeFile->RootCat);
+
+    populateCheatInfo();
+}
+
+void CheatsDialog::populateCheatInfo()
+{
+    auto selmodel = ui->tvCodeList->selectionModel();
+    if (!selmodel->hasSelection())
+    {
+        ui->splitter->widget(1)->setEnabled(false);
+
+        ui->btnDeleteCode->show();
+        ui->btnDeleteCode->setEnabled(false);
+        ui->btnEditCode->show();
+        ui->btnEditCode->setEnabled(false);
+        ui->btnSaveCode->hide();
+        ui->btnCancelEdit->hide();
+
+        ui->txtItemName->clear();
+        ui->txtItemName->setReadOnly(true);
+        ui->txtItemDesc->clear();
+        ui->txtItemDesc->setReadOnly(true);
+
+        ui->chkItemOption->setText("Enabled");
+        ui->chkItemOption->setChecked(false);
+        ui->chkItemOption->setEnabled(false);
+
+        ui->lblCode->show();
+        ui->txtCode->show();
+        ui->txtCode->setPlainText("");
+        ui->txtCode->setReadOnly(true);
 
-    QStandardItem* item = ((QStandardItemModel*)ui->tvCodeList->model())->itemFromIndex(indices.first());
-    QVariant data = item->data();
-    if (!data.canConvert<ARCodeList::iterator>())
         return;
+    }
+
+    ui->splitter->widget(1)->setEnabled(true);
 
-    bool error = false;
+    auto index = selmodel->selectedIndexes()[0];
+    int itemtype = index.data(Qt::UserRole).toInt();
+    if (itemtype == 1)
+    {
+        auto cat = index.data(Qt::UserRole+1).value<ARCodeCat*>();
+
+        ui->btnDeleteCode->show();
+        ui->btnDeleteCode->setEnabled(true);
+        ui->btnEditCode->show();
+        ui->btnEditCode->setEnabled(true);
+        ui->btnSaveCode->hide();
+        ui->btnCancelEdit->hide();
+
+        ui->txtItemName->setText(QString::fromStdString(cat->Name));
+        ui->txtItemName->setReadOnly(true);
+        ui->txtItemDesc->setText(QString::fromStdString(cat->Description));
+        ui->txtItemDesc->setReadOnly(true);
+
+        ui->chkItemOption->setText("Only allow one code enabled");
+        ui->chkItemOption->setChecked(cat->OnlyOneCodeEnabled);
+        ui->chkItemOption->setEnabled(false);
+
+        ui->lblCode->hide();
+        ui->txtCode->hide();
+    }
+    else if (itemtype == 2)
+    {
+        auto code = index.data(Qt::UserRole+1).value<ARCode*>();
+
+        ui->btnDeleteCode->show();
+        ui->btnDeleteCode->setEnabled(true);
+        ui->btnEditCode->show();
+        ui->btnEditCode->setEnabled(true);
+        ui->btnSaveCode->hide();
+        ui->btnCancelEdit->hide();
+
+        ui->txtItemName->setText(QString::fromStdString(code->Name));
+        ui->txtItemName->setReadOnly(true);
+        ui->txtItemDesc->setText(QString::fromStdString(code->Description));
+        ui->txtItemDesc->setReadOnly(true);
+
+        ui->chkItemOption->setText("Enabled");
+        ui->chkItemOption->setChecked(code->Enabled);
+        ui->chkItemOption->setEnabled(false);
+
+        ui->lblCode->show();
+        ui->txtCode->show();
+
+        QString codestr = "";
+        for (size_t i = 0; i < code->Code.size(); i += 2)
+        {
+            u32 c0 = code->Code[i+0];
+            u32 c1 = code->Code[i+1];
+            //codestr += QString("%1 %2\n").arg(c0, 8, 16, '0').arg(c1, 8, 16, '0').toUpper();
+            codestr += QString::asprintf("%08X %08X\n", c0, c1);
+        }
+        ui->txtCode->setPlainText(codestr);
+        ui->txtCode->setReadOnly(true);
+    }
+}
+
+std::vector<u32> CheatsDialog::convertCodeInput()
+{
     std::vector<u32> codeout;
-    codeout.reserve(64);
 
-    QString text = ui->txtCode->document()->toPlainText();
+    QString text = ui->txtCode->toPlainText();
+    text = text.replace((QChar)'o', (QChar)'0', Qt::CaseInsensitive);
 #if QT_VERSION >= QT_VERSION_CHECK(5, 14, 0)
     QStringList lines = text.split('\n', Qt::SkipEmptyParts);
 #else
@@ -329,15 +558,13 @@ void CheatsDialog::on_txtCode_textChanged()
 
         if (line.length() > 17)
         {
-            error = true;
-            break;
+            return {};
         }
 
         QStringList numbers = line.split(' ');
         if (numbers.length() != 2)
         {
-            error = true;
-            break;
+            return {};
         }
 
         QStringList::iterator jt = numbers.begin();
@@ -352,32 +579,281 @@ void CheatsDialog::on_txtCode_textChanged()
 
         if (!c0good || !c1good)
         {
-            error = true;
-            break;
+            return {};
         }
 
         codeout.push_back(c0);
         codeout.push_back(c1);
     }
 
-    ui->btnNewCat->setEnabled(!error);
-    ui->btnNewARCode->setEnabled(!error);
-    ui->btnDeleteCode->setEnabled(!error);
-    ui->tvCodeList->setEnabled(!error);
-    ui->buttonBox->button(QDialogButtonBox::Ok)->setEnabled(!error);
+    return codeout;
+}
+
+
+void CheatListModel::populateCheatListCat(QStandardItem* parentitem, ARCodeCat& parentcat)
+{
+    for (auto& item : parentcat.Children)
+    {
+        if (std::holds_alternative<ARCodeCat>(item))
+        {
+            auto& cat = std::get<ARCodeCat>(item);
+
+            auto catitem = new QStandardItem(QString::fromStdString(cat.Name));
+            catitem->setDropEnabled(true);
+            catitem->setData(1, Qt::UserRole);
+            catitem->setData(QVariant::fromValue(&cat), Qt::UserRole+1);
+            parentitem->appendRow(catitem);
+
+            populateCheatListCat(catitem, cat);
+        }
+        else
+        {
+            auto& code = std::get<ARCode>(item);
+
+            auto codeitem = new QStandardItem(QString::fromStdString(code.Name));
+            codeitem->setCheckable(true);
+            codeitem->setCheckState(code.Enabled ? Qt::Checked : Qt::Unchecked);
+            codeitem->setDropEnabled(false);
+            codeitem->setData(2, Qt::UserRole);
+            codeitem->setData(QVariant::fromValue(&code), Qt::UserRole+1);
+            parentitem->appendRow(codeitem);
+        }
+    }
+}
+
+QMimeData* CheatListModel::mimeData(const QModelIndexList& indexes) const
+{
+    QMimeData* mime = new QMimeData();
+    QByteArray array;
+
+    auto& index = indexes[0];
+    int itemtype = index.data(Qt::UserRole).toInt();
+    quintptr itemptr;
+
+    auto parentlist = (QTreeView*)this->parent();
+    unsigned char flags = (unsigned char)itemtype;
+    if (parentlist->isExpanded(index))
+        flags |= 0x10;
+
+    array.append((char)flags);
+    if (itemtype == 1)
+    {
+        auto cat = index.data(Qt::UserRole+1).value<ARCodeCat*>();
+        itemptr = (quintptr)cat;
+
+    }
+    else if (itemtype == 2)
+    {
+        auto code = index.data(Qt::UserRole+1).value<ARCode*>();
+        itemptr = (quintptr)code;
+    }
+
+    for (int i = 0; i < sizeof(quintptr); i++)
+    {
+        array.append((char)(itemptr & 0xFF));
+        itemptr >>= 8;
+    }
+
+    mime->setData("application/x-qabstractitemmodeldatalist", array);
+    return mime;
+}
+
+bool CheatListModel::canDropMimeData(const QMimeData* mime, Qt::DropAction action, int row, int column, const QModelIndex& parent) const
+{
+    if (action != Qt::MoveAction)
+        return false;
+
+    QByteArray data = mime->data("application/x-qabstractitemmodeldatalist");
+    if (data.length() != (1 + sizeof(quintptr)))
+        return false;
+
+    int flags = (unsigned char)data[0];
+    int itemtype = flags & 0xF;
+
+    QStandardItem* parentitem = itemFromIndex(parent);
+    if (!parentitem) parentitem = invisibleRootItem();
+
+    if (itemtype == 1)
+    {
+        // if we are trying to drop a category into another category: reject this move
+        if (parentitem != invisibleRootItem())
+            return false;
+    }
+
+    return true;
+}
+
+bool CheatListModel::dropMimeData(const QMimeData* mime, Qt::DropAction action, int row, int column, const QModelIndex& parent)
+{
+    if (action != Qt::MoveAction)
+        return false;
+
+    QByteArray data = mime->data("application/x-qabstractitemmodeldatalist");
+    if (data.length() != (1 + sizeof(quintptr)))
+        return false;
+
+    int flags = (unsigned char)data[0];
+    int itemtype = flags & 0xF;
+
+    quintptr itemptr = 0;
+    for (int i = sizeof(quintptr) - 1; i >= 0; i--)
+    {
+        itemptr <<= 8;
+        itemptr |= (unsigned char)data[1 + i];
+    }
+
+    QStandardItem* parentitem = itemFromIndex(parent);
+    if (!parentitem) parentitem = invisibleRootItem();
+    if (row < 0) row = parentitem->rowCount();
+
+    auto newparent = parentitem->data(Qt::UserRole+1).value<ARCodeCat*>();
+    QStandardItem* newitem;
+
+    if (itemtype == 1)
+    {
+        if (parentitem != invisibleRootItem())
+            return false;
+
+        ARCodeCat* cat = (ARCodeCat*)itemptr;
+        auto oldparent = cat->Parent;
+
+        // retrieve iterator to insert new category at
+        auto newit = newparent->Children.begin();
+        for (int i = 0; newit != newparent->Children.end(); i++)
+        {
+            if (i == row) break;
+            newit++;
+        }
+
+        // create new category
+        ARCodeCat newcat = {
+            .Parent = newparent,
+            .Name = cat->Name,
+            .Description = cat->Description,
+            .OnlyOneCodeEnabled = cat->OnlyOneCodeEnabled,
+            .Children = cat->Children
+        };
+
+        // delete category from old parent
+        for (auto it = oldparent->Children.begin(); it != oldparent->Children.end(); it++)
+        {
+            auto& childitem = *it;
+            if (!std::holds_alternative<ARCodeCat>(childitem))
+                continue;
+            if (&std::get<ARCodeCat>(childitem) != cat)
+                continue;
+
+            oldparent->Children.erase(it);
+            break;
+        }
+
+        // add it to new parent
+        newit = newparent->Children.emplace(newit, newcat);
+        ARCodeCat* newcatptr = &std::get<ARCodeCat>(*newit);
+
+        // reparent the children
+        for (auto& childitem : newcatptr->Children)
+        {
+            if (std::holds_alternative<ARCodeCat>(childitem))
+            {
+                auto& thiscat = std::get<ARCodeCat>(childitem);
+                thiscat.Parent = newcatptr;
+            }
+            else
+            {
+                auto& thiscode = std::get<ARCode>(childitem);
+                thiscode.Parent = newcatptr;
+            }
+        }
+
+        // create and add new item
+        auto catitem = new QStandardItem(QString::fromStdString(newcatptr->Name));
+        catitem->setDropEnabled(true);
+        catitem->setData(1, Qt::UserRole);
+        catitem->setData(QVariant::fromValue(newcatptr), Qt::UserRole+1);
+
+        parentitem->insertRow(row, catitem);
+        newitem = catitem;
+
+        populateCheatListCat(catitem, *newcatptr);
+    }
+    else if (itemtype == 2)
+    {
+        ARCode* code = (ARCode*)itemptr;
+        auto oldparent = code->Parent;
+
+        // retrieve iterator to insert new code at
+        auto newit = newparent->Children.begin();
+        for (int i = 0; newit != newparent->Children.end(); i++)
+        {
+            if (i == row) break;
+            newit++;
+        }
+
+        // create new code
+        ARCode newcode = {
+            .Parent = newparent,
+            .Name = code->Name,
+            .Description = code->Description,
+            .Enabled = code->Enabled,
+            .Code = code->Code
+        };
+
+        // delete code from old parent
+        for (auto it = oldparent->Children.begin(); it != oldparent->Children.end(); it++)
+        {
+            auto& childitem = *it;
+            if (!std::holds_alternative<ARCode>(childitem))
+                continue;
+            if (&std::get<ARCode>(childitem) != code)
+                continue;
+
+            oldparent->Children.erase(it);
+            break;
+        }
+
+        // add it to new parent
+        newit = newparent->Children.emplace(newit, newcode);
+        ARCode* newcodeptr = &std::get<ARCode>(*newit);
 
-    if (error) return;
+        // create and add new item
+        auto codeitem = new QStandardItem(QString::fromStdString(newcodeptr->Name));
+        codeitem->setCheckable(true);
+        codeitem->setCheckState(newcodeptr->Enabled ? Qt::Checked : Qt::Unchecked);
+        codeitem->setDropEnabled(false);
+        codeitem->setData(2, Qt::UserRole);
+        codeitem->setData(QVariant::fromValue(newcodeptr), Qt::UserRole+1);
 
-    ARCode& code = *(data.value<ARCodeList::iterator>());
-    code.Code = codeout;
+        parentitem->insertRow(row, codeitem);
+        newitem = codeitem;
+    }
+    else
+        return false;
+
+    auto parentlist = (QTreeView*)this->parent();
+    auto newindex = newitem->index();
+    if (flags & 0x10)
+        parentlist->expand(newindex);
+
+    // quality of life: select the item after it's been moved
+    // we have to add a small delay here
+    // if we do this immediately, it will cause the dragdrop operation to fail
+    QTimer::singleShot(1, [=]()
+    {
+        parentlist->selectionModel()->select(newindex, QItemSelectionModel::ClearAndSelect);
+        parentlist->scrollTo(newindex);
+    });
+
+    return true;
 }
 
+
 void ARCodeChecker::highlightBlock(const QString& text)
 {
     QTextCharFormat errformat; errformat.setForeground(Qt::red);
 
     {
-        QRegularExpression expr("^\\s*[0-9A-Fa-f]{1,8} [0-9A-Fa-f]{1,8}\\s*$");
+        QRegularExpression expr("^\\s*[0-9A-Fa-foO]{1,8} [0-9A-Fa-foO]{1,8}\\s*$");
         QRegularExpressionMatchIterator it = expr.globalMatch(text);
         if (!it.hasNext())
         {
diff --git a/src/frontend/qt_sdl/CheatsDialog.h b/src/frontend/qt_sdl/CheatsDialog.h
index 9f666b118e..6c0902495d 100644
--- a/src/frontend/qt_sdl/CheatsDialog.h
+++ b/src/frontend/qt_sdl/CheatsDialog.h
@@ -27,14 +27,34 @@
 
 #include "ARCodeFile.h"
 
-Q_DECLARE_METATYPE(melonDS::ARCodeList::iterator)
-Q_DECLARE_METATYPE(melonDS::ARCodeCatList::iterator)
-
 namespace Ui { class CheatsDialog; }
 class CheatsDialog;
+class CheatImportDialog;
 
 class EmuInstance;
 
+namespace melonDS
+{
+    class ARDatabaseDAT;
+}
+
+class CheatListModel : public QStandardItemModel
+{
+    Q_OBJECT
+
+public:
+    CheatListModel(QObject* parent = nullptr) : QStandardItemModel(parent) {}
+
+    void populateCheatListCat(QStandardItem* parentitem, melonDS::ARCodeCat& parentcat);
+
+    Qt::DropActions supportedDragActions() const override { return Qt::MoveAction; }
+    Qt::DropActions supportedDropActions() const override { return Qt::MoveAction; }
+
+    QMimeData* mimeData(const QModelIndexList& indexes) const override;
+    bool canDropMimeData(const QMimeData* mime, Qt::DropAction action, int row, int column, const QModelIndex& parent) const override;
+    bool dropMimeData(const QMimeData* mime, Qt::DropAction action, int row, int column, const QModelIndex& parent) override;
+};
+
 class ARCodeChecker : public QSyntaxHighlighter
 {
     Q_OBJECT
@@ -74,25 +94,40 @@ class CheatsDialog : public QDialog
     }
 
 private slots:
-    void on_CheatsDialog_accepted();
-    void on_CheatsDialog_rejected();
+    void done(int r) override;
 
     void on_btnNewCat_clicked();
     void on_btnNewARCode_clicked();
     void on_btnDeleteCode_clicked();
+    void on_btnImportCheats_clicked();
+    void onImportCheatsFinished(int res);
 
     void onCheatSelectionChanged(const QItemSelection& sel, const QItemSelection& desel);
     void onCheatEntryModified(QStandardItem* item);
 
-    void on_txtCode_textChanged();
+    void on_btnEditCode_clicked();
+    void on_btnSaveCode_clicked();
+    void on_btnCancelEdit_clicked();
 
 private:
     Ui::CheatsDialog* ui;
 
     EmuInstance* emuInstance;
 
+    melonDS::u32 gameCode;
+    melonDS::u32 gameChecksum;
+
     melonDS::ARCodeFile* codeFile;
     ARCodeChecker* codeChecker;
+
+    melonDS::ARDatabaseDAT* importDB;
+    CheatImportDialog* importDlg;
+
+    bool updatingEnableChk;
+
+    void populateCheatList();
+    void populateCheatInfo();
+    std::vector<melonDS::u32> convertCodeInput();
 };
 
 #endif // CHEATSDIALOG_H
diff --git a/src/frontend/qt_sdl/CheatsDialog.ui b/src/frontend/qt_sdl/CheatsDialog.ui
index d0bc586a37..5445b283f5 100644
--- a/src/frontend/qt_sdl/CheatsDialog.ui
+++ b/src/frontend/qt_sdl/CheatsDialog.ui
@@ -6,101 +6,252 @@
    <rect>
     <x>0</x>
     <y>0</y>
-    <width>609</width>
-    <height>417</height>
+    <width>711</width>
+    <height>511</height>
    </rect>
   </property>
   <property name="windowTitle">
    <string>Cheat code editor - melonDS</string>
   </property>
-  <layout class="QVBoxLayout" name="verticalLayout_2">
+  <layout class="QVBoxLayout" name="verticalLayout_5">
    <item>
-    <layout class="QHBoxLayout" name="horizontalLayout">
-     <property name="sizeConstraint">
-      <enum>QLayout::SetDefaultConstraint</enum>
+    <widget class="QSplitter" name="splitter">
+     <property name="orientation">
+      <enum>Qt::Orientation::Horizontal</enum>
      </property>
-     <item>
-      <widget class="QPushButton" name="btnNewCat">
-       <property name="text">
-        <string>New category</string>
-       </property>
-      </widget>
-     </item>
-     <item>
-      <widget class="QPushButton" name="btnNewARCode">
-       <property name="text">
-        <string>New AR code</string>
-       </property>
-      </widget>
-     </item>
-     <item>
-      <widget class="QPushButton" name="btnDeleteCode">
-       <property name="text">
-        <string>Delete</string>
-       </property>
-      </widget>
-     </item>
-     <item>
-      <spacer name="horizontalSpacer">
-       <property name="orientation">
-        <enum>Qt::Horizontal</enum>
-       </property>
-       <property name="sizeType">
-        <enum>QSizePolicy::Expanding</enum>
-       </property>
-       <property name="sizeHint" stdset="0">
-        <size>
-         <width>40</width>
-         <height>20</height>
-        </size>
-       </property>
-      </spacer>
-     </item>
-    </layout>
-   </item>
-   <item>
-    <layout class="QHBoxLayout" name="horizontalLayout_2">
-     <item>
-      <widget class="QTreeView" name="tvCodeList">
-       <property name="sizePolicy">
-        <sizepolicy hsizetype="Expanding" vsizetype="Expanding">
-         <horstretch>2</horstretch>
-         <verstretch>0</verstretch>
-        </sizepolicy>
+     <property name="handleWidth">
+      <number>2</number>
+     </property>
+     <widget class="QWidget" name="verticalLayoutWidget">
+      <layout class="QVBoxLayout" name="verticalLayout_2">
+       <property name="topMargin">
+        <number>0</number>
        </property>
-       <property name="minimumSize">
-        <size>
-         <width>200</width>
-         <height>0</height>
-        </size>
+       <property name="rightMargin">
+        <number>4</number>
        </property>
-       <property name="headerHidden">
-        <bool>true</bool>
+       <property name="bottomMargin">
+        <number>0</number>
        </property>
-      </widget>
-     </item>
-     <item>
-      <widget class="QPlainTextEdit" name="txtCode">
-       <property name="sizePolicy">
-        <sizepolicy hsizetype="Expanding" vsizetype="Expanding">
-         <horstretch>3</horstretch>
-         <verstretch>0</verstretch>
-        </sizepolicy>
+       <item>
+        <layout class="QHBoxLayout" name="horizontalLayout">
+         <property name="sizeConstraint">
+          <enum>QLayout::SizeConstraint::SetDefaultConstraint</enum>
+         </property>
+         <item>
+          <widget class="QPushButton" name="btnNewCat">
+           <property name="text">
+            <string>New category</string>
+           </property>
+          </widget>
+         </item>
+         <item>
+          <widget class="QPushButton" name="btnNewARCode">
+           <property name="text">
+            <string>New AR code</string>
+           </property>
+          </widget>
+         </item>
+         <item>
+          <widget class="QPushButton" name="btnImportCheats">
+           <property name="text">
+            <string>Import...</string>
+           </property>
+          </widget>
+         </item>
+         <item>
+          <spacer name="horizontalSpacer">
+           <property name="orientation">
+            <enum>Qt::Orientation::Horizontal</enum>
+           </property>
+           <property name="sizeType">
+            <enum>QSizePolicy::Policy::Expanding</enum>
+           </property>
+           <property name="sizeHint" stdset="0">
+            <size>
+             <width>40</width>
+             <height>20</height>
+            </size>
+           </property>
+          </spacer>
+         </item>
+        </layout>
+       </item>
+       <item>
+        <widget class="QTreeView" name="tvCodeList">
+         <property name="sizePolicy">
+          <sizepolicy hsizetype="Expanding" vsizetype="Expanding">
+           <horstretch>0</horstretch>
+           <verstretch>0</verstretch>
+          </sizepolicy>
+         </property>
+         <property name="minimumSize">
+          <size>
+           <width>200</width>
+           <height>0</height>
+          </size>
+         </property>
+         <property name="editTriggers">
+          <set>QAbstractItemView::EditTrigger::NoEditTriggers</set>
+         </property>
+         <property name="dragEnabled">
+          <bool>true</bool>
+         </property>
+         <property name="dragDropMode">
+          <enum>QAbstractItemView::DragDropMode::InternalMove</enum>
+         </property>
+         <property name="defaultDropAction">
+          <enum>Qt::DropAction::IgnoreAction</enum>
+         </property>
+         <property name="headerHidden">
+          <bool>true</bool>
+         </property>
+        </widget>
+       </item>
+      </layout>
+     </widget>
+     <widget class="QWidget" name="layoutWidget">
+      <layout class="QVBoxLayout" name="verticalLayout" stretch="0,0">
+       <property name="leftMargin">
+        <number>4</number>
        </property>
-       <property name="minimumSize">
-        <size>
-         <width>200</width>
-         <height>0</height>
-        </size>
+       <property name="rightMargin">
+        <number>0</number>
        </property>
-      </widget>
-     </item>
-    </layout>
+       <item>
+        <layout class="QHBoxLayout" name="horizontalLayout_2">
+         <item>
+          <widget class="QPushButton" name="btnEditCode">
+           <property name="text">
+            <string>Edit</string>
+           </property>
+          </widget>
+         </item>
+         <item>
+          <widget class="QPushButton" name="btnSaveCode">
+           <property name="text">
+            <string>Save</string>
+           </property>
+          </widget>
+         </item>
+         <item>
+          <widget class="QPushButton" name="btnCancelEdit">
+           <property name="text">
+            <string>Cancel edit</string>
+           </property>
+          </widget>
+         </item>
+         <item>
+          <spacer name="horizontalSpacer_2">
+           <property name="orientation">
+            <enum>Qt::Orientation::Horizontal</enum>
+           </property>
+           <property name="sizeHint" stdset="0">
+            <size>
+             <width>40</width>
+             <height>20</height>
+            </size>
+           </property>
+          </spacer>
+         </item>
+         <item>
+          <widget class="QPushButton" name="btnDeleteCode">
+           <property name="text">
+            <string>Delete</string>
+           </property>
+          </widget>
+         </item>
+        </layout>
+       </item>
+       <item>
+        <layout class="QFormLayout" name="formLayout">
+         <item row="0" column="0">
+          <widget class="QLabel" name="label">
+           <property name="text">
+            <string>Name:</string>
+           </property>
+          </widget>
+         </item>
+         <item row="0" column="1">
+          <widget class="QLineEdit" name="txtItemName"/>
+         </item>
+         <item row="1" column="0">
+          <widget class="QLabel" name="label_2">
+           <property name="text">
+            <string>Description:</string>
+           </property>
+          </widget>
+         </item>
+         <item row="1" column="1">
+          <widget class="QLineEdit" name="txtItemDesc"/>
+         </item>
+         <item row="2" column="0">
+          <widget class="QLabel" name="label_3">
+           <property name="text">
+            <string/>
+           </property>
+          </widget>
+         </item>
+         <item row="2" column="1">
+          <widget class="QCheckBox" name="chkItemOption">
+           <property name="text">
+            <string>CheckBox</string>
+           </property>
+          </widget>
+         </item>
+         <item row="3" column="0">
+          <widget class="QLabel" name="lblCode">
+           <property name="sizePolicy">
+            <sizepolicy hsizetype="Preferred" vsizetype="Preferred">
+             <horstretch>0</horstretch>
+             <verstretch>0</verstretch>
+            </sizepolicy>
+           </property>
+           <property name="text">
+            <string>Code:</string>
+           </property>
+           <property name="alignment">
+            <set>Qt::AlignmentFlag::AlignLeading|Qt::AlignmentFlag::AlignLeft|Qt::AlignmentFlag::AlignTop</set>
+           </property>
+          </widget>
+         </item>
+         <item row="3" column="1">
+          <widget class="QPlainTextEdit" name="txtCode">
+           <property name="sizePolicy">
+            <sizepolicy hsizetype="Expanding" vsizetype="Expanding">
+             <horstretch>0</horstretch>
+             <verstretch>0</verstretch>
+            </sizepolicy>
+           </property>
+           <property name="minimumSize">
+            <size>
+             <width>200</width>
+             <height>0</height>
+            </size>
+           </property>
+           <property name="font">
+            <font>
+             <family>Noto Sans Mono</family>
+            </font>
+           </property>
+          </widget>
+         </item>
+        </layout>
+       </item>
+      </layout>
+     </widget>
+    </widget>
    </item>
    <item>
     <widget class="QDialogButtonBox" name="buttonBox">
+     <property name="sizePolicy">
+      <sizepolicy hsizetype="Expanding" vsizetype="Fixed">
+       <horstretch>0</horstretch>
+       <verstretch>0</verstretch>
+      </sizepolicy>
+     </property>
      <property name="standardButtons">
-      <set>QDialogButtonBox::Cancel|QDialogButtonBox::Ok</set>
+      <set>QDialogButtonBox::StandardButton::Close</set>
      </property>
     </widget>
    </item>
@@ -115,12 +266,12 @@
    <slot>accept()</slot>
    <hints>
     <hint type="sourcelabel">
-     <x>304</x>
-     <y>396</y>
+     <x>649</x>
+     <y>23</y>
     </hint>
     <hint type="destinationlabel">
-     <x>304</x>
-     <y>208</y>
+     <x>349</x>
+     <y>230</y>
     </hint>
    </hints>
   </connection>
@@ -131,12 +282,12 @@
    <slot>reject()</slot>
    <hints>
     <hint type="sourcelabel">
-     <x>304</x>
-     <y>396</y>
+     <x>649</x>
+     <y>23</y>
     </hint>
     <hint type="destinationlabel">
-     <x>304</x>
-     <y>208</y>
+     <x>349</x>
+     <y>230</y>
     </hint>
    </hints>
   </connection>
diff --git a/src/frontend/qt_sdl/EmuInstance.cpp b/src/frontend/qt_sdl/EmuInstance.cpp
index 51dfaa6d16..d6d662fa93 100755
--- a/src/frontend/qt_sdl/EmuInstance.cpp
+++ b/src/frontend/qt_sdl/EmuInstance.cpp
@@ -99,6 +99,7 @@ EmuInstance::EmuInstance(int inst) : deleting(false),
         targetFPS = 60.0;
     }
     else targetFPS = val;
+    curFPS = targetFPS;
 
     val = globalCfg.GetDouble("FastForwardFPS");
     if (val == 0.0)
@@ -1318,6 +1319,7 @@ bool EmuInstance::updateConsole() noexcept
             jitargs,
             static_cast<AudioBitDepth>(globalCfg.GetInt("Audio.BitDepth")),
             static_cast<AudioInterpolation>(globalCfg.GetInt("Audio.Interpolation")),
+            (double) audioFreq,
             gdbargs,
     };
     NDSArgs* args = &ndsargs;
diff --git a/src/frontend/qt_sdl/EmuInstance.h b/src/frontend/qt_sdl/EmuInstance.h
index c8e756ea6e..f143496573 100755
--- a/src/frontend/qt_sdl/EmuInstance.h
+++ b/src/frontend/qt_sdl/EmuInstance.h
@@ -232,7 +232,6 @@ class EmuInstance
     void setupMicInputData();
 
     int audioGetNumSamplesOut(int outlen);
-    void audioResample(melonDS::s16* inbuf, int inlen, melonDS::s16* outbuf, int outlen, int volume);
     static void audioCallback(void* data, Uint8* stream, int len);
 
     int micGetNumSamplesIn(int inlen);
diff --git a/src/frontend/qt_sdl/EmuInstanceAudio.cpp b/src/frontend/qt_sdl/EmuInstanceAudio.cpp
index 0ada9388b1..40656de47f 100644
--- a/src/frontend/qt_sdl/EmuInstanceAudio.cpp
+++ b/src/frontend/qt_sdl/EmuInstanceAudio.cpp
@@ -26,6 +26,7 @@
 
 using namespace melonDS;
 
+#define INTERNAL_FRAME_RATE 59.8260982880808f
 
 // --- AUDIO OUTPUT -----------------------------------------------------------
 
@@ -40,7 +41,8 @@ void EmuInstance::audioInit()
     audioSyncLock = SDL_CreateMutex();
 
     audioFreq = 48000; // TODO: make both of these configurable?
-    audioBufSize = 1024;
+    audioBufSize = 512;
+
     SDL_AudioSpec whatIwant, whatIget;
     memset(&whatIwant, 0, sizeof(SDL_AudioSpec));
     whatIwant.freq = audioFreq;
@@ -49,7 +51,7 @@ void EmuInstance::audioInit()
     whatIwant.samples = audioBufSize;
     whatIwant.callback = audioCallback;
     whatIwant.userdata = this;
-    audioDevice = SDL_OpenAudioDevice(NULL, 0, &whatIwant, &whatIget, SDL_AUDIO_ALLOW_FREQUENCY_CHANGE | SDL_AUDIO_ALLOW_SAMPLES_CHANGE);
+    audioDevice = SDL_OpenAudioDevice(NULL, 0, &whatIwant, &whatIget, SDL_AUDIO_ALLOW_FREQUENCY_CHANGE);
     if (!audioDevice)
     {
         Platform::Log(Platform::LogLevel::Error, "Audio init failed: %s\n", SDL_GetError());
@@ -126,7 +128,7 @@ void EmuInstance::audioSync()
     if (audioDevice)
     {
         SDL_LockMutex(audioSyncLock);
-        while (nds->SPU.GetOutputSize() > audioBufSize)
+        while (nds->SPU.GetOutputSize() >= audioFreq / INTERNAL_FRAME_RATE)
         {
             int ret = SDL_CondWaitTimeout(audioSyncCond, audioSyncLock, 500);
             if (ret == SDL_MUTEX_TIMEDOUT) break;
@@ -137,8 +139,7 @@ void EmuInstance::audioSync()
 
 int EmuInstance::audioGetNumSamplesOut(int outlen)
 {
-    //float f_len_in = (outlen * 32823.6328125 * (curFPS/60.0)) / (float)audioFreq;
-    float f_len_in = (outlen * 47743.4659091 * (curFPS/60.0)) / (float)audioFreq;
+    float f_len_in = outlen * (curFPS/targetFPS);
     f_len_in += audioSampleFrac;
     int len_in = (int)floor(f_len_in);
     audioSampleFrac = f_len_in - len_in;
@@ -146,48 +147,20 @@ int EmuInstance::audioGetNumSamplesOut(int outlen)
     return len_in;
 }
 
-void EmuInstance::audioResample(s16* inbuf, int inlen, s16* outbuf, int outlen, int volume)
-{
-    float res_incr = inlen / (float)outlen;
-    float res_timer = -0.5;
-    int res_pos = 0;
-
-    for (int i = 0; i < outlen; i++)
-    {
-        s16 l1 = inbuf[res_pos * 2];
-        s16 l2 = inbuf[res_pos * 2 + 2];
-        s16 r1 = inbuf[res_pos * 2 + 1];
-        s16 r2 = inbuf[res_pos * 2 + 3];
-
-        float l = (float)l1 + ((l2 - l1) * res_timer);
-        float r = (float)r1 + ((r2 - r1) * res_timer);
-
-        outbuf[i*2  ] = (s16)(((s32)round(l) * volume) >> 8);
-        outbuf[i*2+1] = (s16)(((s32)round(r) * volume) >> 8);
-
-        res_timer += res_incr;
-        while (res_timer >= 1.0)
-        {
-            res_timer -= 1.0;
-            res_pos++;
-        }
-    }
-}
-
 void EmuInstance::audioCallback(void* data, Uint8* stream, int len)
 {
     EmuInstance* inst = (EmuInstance*)data;
     len /= (sizeof(s16) * 2);
 
-    // resample incoming audio to match the output sample rate
+    double skew = std::clamp(inst->targetFPS / INTERNAL_FRAME_RATE, 0.995, 1.005);
+    inst->nds->SPU.SetOutputSkew(skew);
 
     int len_in = inst->audioGetNumSamplesOut(len);
     if (len_in > inst->audioBufSize) len_in = inst->audioBufSize;
     s16 buf_in[inst->audioBufSize*2];
-    int num_in;
 
     SDL_LockMutex(inst->audioSyncLock);
-    num_in = inst->nds->SPU.ReadOutput(buf_in, len_in);
+    int num_in = inst->nds->SPU.ReadOutput((s16*) stream, len_in);
     SDL_CondSignal(inst->audioSyncCond);
     SDL_UnlockMutex(inst->audioSyncLock);
 
@@ -197,18 +170,21 @@ void EmuInstance::audioCallback(void* data, Uint8* stream, int len)
         return;
     }
 
+    if (inst->audioVolume < 256)
+    {
+        s16* samples = (s16*) stream;
+        for (int i = 0; i < num_in * 2; i++)
+            samples[i] = ((s32) samples[i] * inst->audioVolume) >> 8;
+    }
+
     int margin = 6;
     if (num_in < len_in-margin)
     {
         int last = num_in-1;
 
         for (int i = num_in; i < len_in-margin; i++)
-            ((u32*)buf_in)[i] = ((u32*)buf_in)[last];
-
-        num_in = len_in-margin;
+            ((u32*)stream)[i] = ((u32*)stream)[last];
     }
-
-    inst->audioResample(buf_in, num_in, (s16*)stream, len, inst->audioVolume);
 }
 
 
@@ -249,7 +225,7 @@ void EmuInstance::micOpen()
     {
         mic = micDeviceName.c_str();
     }
-    micDevice = SDL_OpenAudioDevice(mic, 1, &whatIwant, &whatIget, SDL_AUDIO_ALLOW_FREQUENCY_CHANGE | SDL_AUDIO_ALLOW_SAMPLES_CHANGE);
+    micDevice = SDL_OpenAudioDevice(mic, 1, &whatIwant, &whatIget, SDL_AUDIO_ALLOW_FREQUENCY_CHANGE);
     if (!micDevice)
     {
         Platform::Log(Platform::LogLevel::Error, "Mic init failed: %s\n", SDL_GetError());
diff --git a/src/frontend/qt_sdl/Platform.cpp b/src/frontend/qt_sdl/Platform.cpp
index c9e2f07661..6ad06e5957 100644
--- a/src/frontend/qt_sdl/Platform.cpp
+++ b/src/frontend/qt_sdl/Platform.cpp
@@ -42,8 +42,11 @@
 #include "SPI_Firmware.h"
 
 #ifdef __WIN32__
+#include <io.h>
+#define fdopen _fdopen
 #define fseek _fseeki64
 #define ftell _ftelli64
+#define dup _dup
 #endif // __WIN32__
 
 extern CameraManager* camManager[2];
@@ -136,20 +139,19 @@ FileHandle* OpenFile(const std::string& path, FileMode mode)
     std::string modeString = GetModeString(mode, QFile::exists(qpath));
     QIODevice::OpenMode qmode = GetQMode(mode);
     QFile qfile{qpath};
-    qfile.open(qmode);
-    FILE* file = fdopen(dup(qfile.handle()), modeString.c_str());
-    qfile.close();
-
-    if (file)
+    if (qfile.open(qmode))
     {
-        Log(LogLevel::Debug, "Opened \"%s\" with FileMode 0x%x (effective mode \"%s\")\n", path.c_str(), mode, modeString.c_str());
-        return reinterpret_cast<FileHandle *>(file);
-    }
-    else
-    {
-        Log(LogLevel::Warn, "Failed to open \"%s\" with FileMode 0x%x (effective mode \"%s\")\n", path.c_str(), mode, modeString.c_str());
-        return nullptr;
+        FILE* file = fdopen(dup(qfile.handle()), modeString.c_str());
+        qfile.close();
+
+        if (file)
+        {
+            Log(LogLevel::Debug, "Opened \"%s\" with FileMode 0x%x (effective mode \"%s\")\n", path.c_str(), mode, modeString.c_str());
+            return reinterpret_cast<FileHandle *>(file);
+        }
     }
+    Log(LogLevel::Warn, "Failed to open \"%s\" with FileMode 0x%x (effective mode \"%s\")\n", path.c_str(), mode, modeString.c_str());
+    return nullptr;
 }
 
 std::string GetLocalFilePath(const std::string& filename)
@@ -263,6 +265,11 @@ void FileRewind(FileHandle* file)
     rewind(reinterpret_cast<FILE *>(file));
 }
 
+u64 FilePosition(FileHandle* file)
+{
+    return ftell(reinterpret_cast<FILE *>(file));
+}
+
 u64 FileRead(void* data, u64 size, u64 count, FileHandle* file)
 {
     return fread(data, size, count, reinterpret_cast<FILE *>(file));
diff --git a/src/frontend/qt_sdl/SaveManager.h b/src/frontend/qt_sdl/SaveManager.h
index f9c8611072..6706c39f6a 100644
--- a/src/frontend/qt_sdl/SaveManager.h
+++ b/src/frontend/qt_sdl/SaveManager.h
@@ -20,7 +20,9 @@
 #define SAVEMANAGER_H
 
 #include <string>
+#ifndef _MSC_VER
 #include <unistd.h>
+#endif
 #include <time.h>
 #include <atomic>
 #include <memory>
diff --git a/src/frontend/qt_sdl/Screen.cpp b/src/frontend/qt_sdl/Screen.cpp
index 17fd4b8644..6d58fff981 100755
--- a/src/frontend/qt_sdl/Screen.cpp
+++ b/src/frontend/qt_sdl/Screen.cpp
@@ -23,11 +23,7 @@
 
 #include <QPaintEvent>
 #include <QPainter>
-#ifndef _WIN32
-#ifndef APPLE
-#include <qpa/qplatformnativeinterface.h>
-#endif
-#endif
+
 #include <QDateTime>
 
 #include "OpenGLSupport.h"
@@ -50,6 +46,14 @@
 
 using namespace melonDS;
 
+#if !defined(_WIN32) && !defined(APPLE)
+#if QT_VERSION >= QT_VERSION_CHECK(6, 5, 0)
+using namespace QNativeInterface;
+#else
+#include <qpa/qplatformnativeinterface.h>
+#endif
+#endif
+
 
 const u32 kOSDMargin = 6;
 const int kLogoWidth = 192;
@@ -1285,8 +1289,25 @@ std::optional<WindowInfo> ScreenPanelGL::getWindowInfo()
     wi.type = WindowInfo::Type::MacOS;
     wi.window_handle = reinterpret_cast<void*>(winId());
     #else
-    QPlatformNativeInterface* pni = QGuiApplication::platformNativeInterface();
     const QString platform_name = QGuiApplication::platformName();
+
+    #if QT_VERSION >= QT_VERSION_CHECK(6, 5, 0)
+    if (platform_name == QStringLiteral("xcb"))
+    {
+        wi.type = WindowInfo::Type::X11;
+        const QX11Application* x11 = qApp->nativeInterface<QX11Application>();
+        wi.display_connection = x11->display();
+        wi.window_handle = reinterpret_cast<void*>(winId());
+    }
+    else if (platform_name == QStringLiteral("wayland"))
+    {
+        wi.type = WindowInfo::Type::Wayland;
+        const QWaylandApplication* wl = qApp->nativeInterface<QWaylandApplication>();
+        wi.display_connection = wl->display();
+        wi.window_handle = reinterpret_cast<void*>(winId());
+    }
+    #else
+    QPlatformNativeInterface* pni = QGuiApplication::platformNativeInterface();
     if (platform_name == QStringLiteral("xcb"))
     {
         wi.type = WindowInfo::Type::X11;
@@ -1303,9 +1324,9 @@ std::optional<WindowInfo> ScreenPanelGL::getWindowInfo()
         wi.display_connection = pni->nativeResourceForWindow("display", handle);
         wi.window_handle = pni->nativeResourceForWindow("surface", handle);
     }
+    #endif
     else
     {
-        //qCritical() << "Unknown PNI platform " << platform_name;
         Platform::Log(Platform::LogLevel::Error, "Unknown PNI platform %s\n", platform_name.toStdString().c_str());
         return std::nullopt;
     }
diff --git a/src/frontend/qt_sdl/TitleManagerDialog.cpp b/src/frontend/qt_sdl/TitleManagerDialog.cpp
index a06f578194..f65ca1d419 100644
--- a/src/frontend/qt_sdl/TitleManagerDialog.cpp
+++ b/src/frontend/qt_sdl/TitleManagerDialog.cpp
@@ -119,7 +119,7 @@ void TitleManagerDialog::createTitleItem(u32 category, u32 titleid)
 
     // TODO: make it possible to select other languages?
     QString title = QString::fromUtf16(banner.EnglishTitle, 128);
-    title = title.left(title.indexOf('\0'));
+    title = title.left(title.indexOf(QChar('\0')));
     title.replace("\n", " · ");
 
     char gamecode[5];
diff --git a/src/frontend/qt_sdl/Window.cpp b/src/frontend/qt_sdl/Window.cpp
index b5a9b69064..9b3ab95c2d 100755
--- a/src/frontend/qt_sdl/Window.cpp
+++ b/src/frontend/qt_sdl/Window.cpp
@@ -47,9 +47,6 @@
 #include <unistd.h>
 #include <sys/socket.h>
 #include <signal.h>
-#ifndef APPLE
-#include <qpa/qplatformnativeinterface.h>
-#endif
 #endif
 
 #include "main.h"
diff --git a/src/frontend/qt_sdl/main.cpp b/src/frontend/qt_sdl/main.cpp
index 651426638f..f5bb3c9549 100644
--- a/src/frontend/qt_sdl/main.cpp
+++ b/src/frontend/qt_sdl/main.cpp
@@ -42,9 +42,6 @@
 #include <unistd.h>
 #include <sys/socket.h>
 #include <signal.h>
-#ifndef APPLE
-#include <qpa/qplatformnativeinterface.h>
-#endif
 #endif
 
 #include <SDL2/SDL.h>
@@ -349,6 +346,10 @@ int main(int argc, char** argv)
         }
     }
 
+    // fix for Wayland OpenGL glitches
+    QGuiApplication::setAttribute(Qt::AA_NativeWindows, false);
+    QGuiApplication::setAttribute(Qt::AA_DontCreateNativeWidgetSiblings, true);
+
     // default MP interface type is local MP
     // this will be changed if a LAN or netplay session is initiated
     setMPInterface(MPInterface_Local);
@@ -401,24 +402,3 @@ int main(int argc, char** argv)
     SDL_Quit();
     return ret;
 }
-
-#ifdef __WIN32__
-
-#include <windows.h>
-
-int CALLBACK WinMain(HINSTANCE hinst, HINSTANCE hprev, LPSTR cmdline, int cmdshow)
-{
-    if (AttachConsole(ATTACH_PARENT_PROCESS) && GetStdHandle(STD_OUTPUT_HANDLE))
-    {
-        freopen("CONOUT$", "w", stdout);
-        freopen("CONOUT$", "w", stderr);
-    }
-
-    int ret = main(__argc, __argv);
-
-    printf("\n\n>");
-
-    return ret;
-}
-
-#endif
diff --git a/src/melonDLDI.h b/src/melonDLDI.h
index 271e07137d..ee77b78a4c 100644
--- a/src/melonDLDI.h
+++ b/src/melonDLDI.h
@@ -25,7 +25,7 @@ namespace melonDS
 {
 const u8 melonDLDI[] =
 {
-    0xed, 0xa5, 0x8d, 0xbf, 0x20, 0x43, 0x68, 0x69, 0x73, 0x68, 0x6d, 0x00, 0x01, 0x09, 0x00, 0x00,
+    0xed, 0xa5, 0x8d, 0xbf, 0x20, 0x43, 0x68, 0x69, 0x73, 0x68, 0x6d, 0x00, 0x01, 0x0a, 0x00, 0x00,
     0x6d, 0x65, 0x6c, 0x6f, 0x6e, 0x44, 0x53, 0x20, 0x44, 0x4c, 0x44, 0x49, 0x20, 0x64, 0x72, 0x69,
     0x76, 0x65, 0x72, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
     0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
diff --git a/src/net/Net_Slirp.cpp b/src/net/Net_Slirp.cpp
index 3937c6bc16..c36ba84d4f 100644
--- a/src/net/Net_Slirp.cpp
+++ b/src/net/Net_Slirp.cpp
@@ -50,6 +50,9 @@ const u8 kServerMAC[6] = {0x00, 0xAB, 0x33, 0x28, 0x99, 0x44};
 #ifdef __WIN32__
 
 #define poll WSAPoll
+#ifndef CLOCK_MONOTONIC
+#define CLOCK_MONOTONIC 1
+#endif
 
 // https://stackoverflow.com/questions/5404277/porting-clock-gettime-to-windows
 
diff --git a/src/net/libslirp/glib/glib.h b/src/net/libslirp/glib/glib.h
index 3563f7f5e0..f48993e825 100644
--- a/src/net/libslirp/glib/glib.h
+++ b/src/net/libslirp/glib/glib.h
@@ -11,6 +11,11 @@
 
 #if defined(WIN32) || defined(_WIN32) || defined(_MSC_VER)
 #define G_OS_WIN32 1
+
+#include <io.h>
+#define open _open
+#define close _close
+typedef ptrdiff_t ssize_t;
 #endif
 
 #define G_LITTLE_ENDIAN 0
diff --git a/src/types.h b/src/types.h
index 461c84f224..75b24f09ae 100644
--- a/src/types.h
+++ b/src/types.h
@@ -35,5 +35,12 @@ typedef int64_t     s64;
 
 template<class T, std::size_t A, std::size_t B>
 using array2d = std::array<std::array<T, B>, A>;
+
+#ifdef _MSC_VER
+#define strcasecmp _stricmp
+#define strncasecmp _strnicmp
+
+typedef ptrdiff_t ssize_t;
+#endif
 }
 #endif // TYPES_H
diff --git a/src/xxhash/xxh3.h b/src/xxhash/xxh3.h
index 5d5faf8774..7e3ce68e0c 100644
--- a/src/xxhash/xxh3.h
+++ b/src/xxhash/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)
  *
@@ -34,2357 +34,22 @@
  */
 
 /*
- * Note: This file is separated for development purposes.
- * It will be integrated into `xxhash.h` when development stage is completed.
+ * Note: This file used to host the source code of XXH3_* variants.
+ * during the development period.
+ * The source code is now properly integrated within xxhash.h.
  *
- * Credit: most of the work on vectorial and asm variants comes from @easyaspi314
- */
-
-#ifndef XXH3_H_1397135465
-#define XXH3_H_1397135465
-
-/* ===   Dependencies   === */
-#ifndef XXHASH_H_5627135585666179
-/* special: when including `xxh3.h` directly, turn on XXH_INLINE_ALL */
-#  undef XXH_INLINE_ALL   /* avoid redefinition */
-#  define XXH_INLINE_ALL
-#endif
-#include "xxhash.h"
-
-
-/* ===   Compiler specifics   === */
-
-#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
-#  define XXH_RESTRICT   restrict
-#else
-/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */
-#  define XXH_RESTRICT   /* disable */
-#endif
-
-#if (defined(__GNUC__) && (__GNUC__ >= 3))  \
-  || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
-  || defined(__clang__)
-#    define XXH_likely(x) __builtin_expect(x, 1)
-#    define XXH_unlikely(x) __builtin_expect(x, 0)
-#else
-#    define XXH_likely(x) (x)
-#    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)
-#    define inline __inline__  /* clang bug */
-#    include <arm_neon.h>
-#    undef inline
-#  endif
-#elif defined(_MSC_VER)
-#  include <intrin.h>
-#endif
-
-/*
- * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
- * remaining a true 64-bit/128-bit hash function.
- *
- * This is done by prioritizing a subset of 64-bit operations that can be
- * emulated without too many steps on the average 32-bit machine.
- *
- * For example, these two lines seem similar, and run equally fast on 64-bit:
- *
- *   xxh_u64 x;
- *   x ^= (x >> 47); // good
- *   x ^= (x >> 13); // bad
- *
- * However, to a 32-bit machine, there is a major difference.
- *
- * x ^= (x >> 47) looks like this:
- *
- *   x.lo ^= (x.hi >> (47 - 32));
- *
- * while x ^= (x >> 13) looks like this:
- *
- *   // note: funnel shifts are not usually cheap.
- *   x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
- *   x.hi ^= (x.hi >> 13);
- *
- * The first one is significantly faster than the second, simply because the
- * shift is larger than 32. This means:
- *  - All the bits we need are in the upper 32 bits, so we can ignore the lower
- *    32 bits in the shift.
- *  - The shift result will always fit in the lower 32 bits, and therefore,
- *    we can ignore the upper 32 bits in the xor.
- *
- * Thanks to this optimization, XXH3 only requires these features to be efficient:
- *
- *  - Usable unaligned access
- *  - A 32-bit or 64-bit ALU
- *      - If 32-bit, a decent ADC instruction
- *  - A 32 or 64-bit multiply with a 64-bit result
- *  - For the 128-bit variant, a decent byteswap helps short inputs.
- *
- * The first two are already required by XXH32, and almost all 32-bit and 64-bit
- * platforms which can run XXH32 can run XXH3 efficiently.
- *
- * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
- * notable exception.
- *
- * First of all, Thumb-1 lacks support for the UMULL instruction which
- * performs the important long multiply. This means numerous __aeabi_lmul
- * calls.
- *
- * Second of all, the 8 functional registers are just not enough.
- * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
- * Lo registers, and this shuffling results in thousands more MOVs than A32.
- *
- * A32 and T32 don't have this limitation. They can access all 14 registers,
- * do a 32->64 multiply with UMULL, and the flexible operand allowing free
- * shifts is helpful, too.
- *
- * Therefore, we do a quick sanity check.
- *
- * If compiling Thumb-1 for a target which supports ARM instructions, we will
- * emit a warning, as it is not a "sane" platform to compile for.
- *
- * Usually, if this happens, it is because of an accident and you probably need
- * to specify -march, as you likely meant to compile for a newer architecture.
- */
-#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
-#   warning "XXH3 is highly inefficient without ARM or Thumb-2."
-#endif
-
-/* ==========================================
- * Vectorization detection
- * ========================================== */
-#define XXH_SCALAR 0 /* Portable scalar version */
-#define XXH_SSE2   1 /* SSE2 for Pentium 4 and all x86_64 */
-#define XXH_AVX2   2 /* AVX2 for Haswell and Bulldozer */
-#define XXH_NEON   3 /* NEON for most ARMv7-A and all AArch64 */
-#define XXH_VSX    4 /* VSX and ZVector for POWER8/z13 */
-#define XXH_AVX512 5 /* AVX512 for Skylake and Icelake */
-
-#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
-#  elif defined(__GNUC__) /* msvc support maybe later */ \
-  && (defined(__ARM_NEON__) || defined(__ARM_NEON)) \
-  && (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \
-    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__))
-#    define XXH_VECTOR XXH_NEON
-#  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
-     || (defined(__s390x__) && defined(__VEC__)) \
-     && defined(__GNUC__) /* TODO: IBM XL */
-#    define XXH_VECTOR XXH_VSX
-#  else
-#    define XXH_VECTOR XXH_SCALAR
-#  endif
-#endif
-
-/*
- * Controls the alignment of the accumulator.
- * This is for compatibility with aligned vector loads, which are usually faster.
- */
-#ifndef XXH_ACC_ALIGN
-#  if XXH_VECTOR == XXH_SCALAR  /* scalar */
-#     define XXH_ACC_ALIGN 8
-#  elif XXH_VECTOR == XXH_SSE2  /* sse2 */
-#     define XXH_ACC_ALIGN 16
-#  elif XXH_VECTOR == XXH_AVX2  /* avx2 */
-#     define XXH_ACC_ALIGN 32
-#  elif XXH_VECTOR == XXH_NEON  /* neon */
-#     define XXH_ACC_ALIGN 16
-#  elif XXH_VECTOR == XXH_VSX   /* vsx */
-#     define XXH_ACC_ALIGN 16
-#  elif XXH_VECTOR == XXH_AVX512 /* avx512 */
-#     define XXH_ACC_ALIGN 64
-#  endif
-#endif
-
-/*
- * UGLY HACK:
- * GCC usually generates the best code with -O3 for xxHash.
- *
- * However, when targeting AVX2, it is overzealous in its unrolling resulting
- * in code roughly 3/4 the speed of Clang.
- *
- * There are other issues, such as GCC splitting _mm256_loadu_si256 into
- * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
- * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
- *
- * That is why when compiling the AVX2 version, it is recommended to use either
- *   -O2 -mavx2 -march=haswell
- * or
- *   -O2 -mavx2 -mno-avx256-split-unaligned-load
- * for decent performance, or to use Clang instead.
- *
- * Fortunately, we can control the first one with a pragma that forces GCC into
- * -O2, but the other one we can't control without "failed to inline always
- * inline function due to target mismatch" warnings.
- */
-#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
-  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* 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.
- *
- * On ARMv7-A, this strangely modifies both parameters in place instead of
- * taking the usual 3-operand form.
- *
- * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the
- * lower and upper halves of the Q register to end up with the high and low
- * halves where we want - all in one instruction.
- *
- *   vzip.32   d10, d11       @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] }
- *
- * Unfortunately we need inline assembly for this: Instructions modifying two
- * registers at once is not possible in GCC or Clang's IR, and they have to
- * create a copy.
- *
- * aarch64 requires a different approach.
- *
- * In order to make it easier to write a decent compiler for aarch64, many
- * quirks were removed, such as conditional execution.
- *
- * NEON was also affected by this.
- *
- * aarch64 cannot access the high bits of a Q-form register, and writes to a
- * D-form register zero the high bits, similar to how writes to W-form scalar
- * registers (or DWORD registers on x86_64) work.
- *
- * The formerly free vget_high intrinsics now require a vext (with a few
- * exceptions)
- *
- * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent
- * of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one
- * operand.
- *
- * The equivalent of the VZIP.32 on the lower and upper halves would be this
- * mess:
- *
- *   ext     v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] }
- *   zip1    v1.2s, v0.2s, v2.2s     // v1 = { v0[0], v2[0] }
- *   zip2    v0.2s, v0.2s, v1.2s     // v0 = { v0[1], v2[1] }
- *
- * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN):
- *
- *   shrn    v1.2s, v0.2d, #32  // v1 = (uint32x2_t)(v0 >> 32);
- *   xtn     v0.2s, v0.2d       // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF);
- *
- * This is available on ARMv7-A, but is less efficient than a single VZIP.32.
- */
-
-/*
- * 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__)
-#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                              \
-    do {                                                                                    \
-      /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = upper D half */ \
-      /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 */     \
-      /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 */ \
-      __asm__("vzip.32  %e0, %f0" : "+w" (in));                                             \
-      (outLo) = vget_low_u32 (vreinterpretq_u32_u64(in));                                   \
-      (outHi) = vget_high_u32(vreinterpretq_u32_u64(in));                                   \
-   } while (0)
-# else
-#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                            \
-    do {                                                                                  \
-      (outLo) = vmovn_u64    (in);                                                        \
-      (outHi) = vshrn_n_u64  ((in), 32);                                                  \
-    } while (0)
-# endif
-#endif  /* XXH_VECTOR == XXH_NEON */
-
-/*
- * VSX and Z Vector helpers.
- *
- * This is very messy, and any pull requests to clean this up are welcome.
- *
- * There are a lot of problems with supporting VSX and s390x, due to
- * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
- */
-#if XXH_VECTOR == XXH_VSX
-#  if defined(__s390x__)
-#    include <s390intrin.h>
-#  else
-#    include <altivec.h>
-#  endif
-
-#  undef vector /* Undo the pollution */
-
-typedef __vector unsigned long long xxh_u64x2;
-typedef __vector unsigned char xxh_u8x16;
-typedef __vector unsigned xxh_u32x4;
-
-# ifndef XXH_VSX_BE
-#  if defined(__BIG_ENDIAN__) \
-  || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
-#    define XXH_VSX_BE 1
-#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
-#    warning "-maltivec=be is not recommended. Please use native endianness."
-#    define XXH_VSX_BE 1
-#  else
-#    define XXH_VSX_BE 0
-#  endif
-# endif /* !defined(XXH_VSX_BE) */
-
-# if XXH_VSX_BE
-/* A wrapper for POWER9's vec_revb. */
-#  if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
-#    define XXH_vec_revb vec_revb
-#  else
-XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
-{
-    xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
-                                  0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
-    return vec_perm(val, val, vByteSwap);
-}
-#  endif
-# endif /* XXH_VSX_BE */
-
-/*
- * Performs an unaligned load and byte swaps it on big endian.
- */
-XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
-{
-    xxh_u64x2 ret;
-    memcpy(&ret, ptr, sizeof(xxh_u64x2));
-# if XXH_VSX_BE
-    ret = XXH_vec_revb(ret);
-# endif
-    return ret;
-}
-
-/*
- * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
- *
- * These intrinsics weren't added until GCC 8, despite existing for a while,
- * and they are endian dependent. Also, their meaning swap depending on version.
- * */
-# if defined(__s390x__)
- /* s390x is always big endian, no issue on this platform */
-#  define XXH_vec_mulo vec_mulo
-#  define XXH_vec_mule vec_mule
-# elif defined(__clang__) && __has_builtin(__builtin_altivec_vmuleuw)
-/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
-#  define XXH_vec_mulo __builtin_altivec_vmulouw
-#  define XXH_vec_mule __builtin_altivec_vmuleuw
-# else
-/* gcc needs inline assembly */
-/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
-XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
-{
-    xxh_u64x2 result;
-    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
-    return result;
-}
-XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
-{
-    xxh_u64x2 result;
-    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
-    return result;
-}
-# endif /* XXH_vec_mulo, XXH_vec_mule */
-#endif /* XXH_VECTOR == XXH_VSX */
-
-
-/* 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_I86))  /* _mm_prefetch() is not defined outside of x86/x64 */
-#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
-#    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
-#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
-#    define XXH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
-#  else
-#    define XXH_PREFETCH(ptr) (void)(ptr)  /* disabled */
-#  endif
-#endif  /* XXH_NO_PREFETCH */
-
-
-/* ==========================================
- * XXH3 default settings
- * ========================================== */
-
-#define XXH_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
-
-#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
-#  error "default keyset is not large enough"
-#endif
-
-/* Pseudorandom secret taken directly from FARSH */
-XXH_ALIGN(64) static const xxh_u8 kSecret[XXH_SECRET_DEFAULT_SIZE] = {
-    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
-    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
-    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
-    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
-    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
-    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
-    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
-    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
-
-    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
-    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
-    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
-    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
-};
-
-/*
- * Calculates a 32-bit to 64-bit long multiply.
- *
- * Wraps __emulu on MSVC x86 because it tends to call __allmul when it doesn't
- * need to (but it shouldn't need to anyways, it is about 7 instructions to do
- * a 64x64 multiply...). Since we know that this will _always_ emit MULL, we
- * use that instead of the normal method.
- *
- * If you are compiling for platforms like Thumb-1 and don't have a better option,
- * you may also want to write your own long multiply routine here.
- *
- * XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y)
- * {
- *    return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
- * }
- */
-#if defined(_MSC_VER) && defined(_M_IX86)
-#    include <intrin.h>
-#    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
-#else
-/*
- * Downcast + upcast is usually better than masking on older compilers like
- * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
- *
- * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
- * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
- */
-#    define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
-#endif
-
-/*
- * Calculates a 64->128-bit long multiply.
- *
- * Uses __uint128_t and _umul128 if available, otherwise uses a scalar version.
- */
-static XXH128_hash_t
-XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
-{
-    /*
-     * GCC/Clang __uint128_t method.
-     *
-     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
-     * This is usually the best way as it usually uses a native long 64-bit
-     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
-     *
-     * Usually.
-     *
-     * Despite being a 32-bit platform, Clang (and emscripten) define this type
-     * despite not having the arithmetic for it. This results in a laggy
-     * compiler builtin call which calculates a full 128-bit multiply.
-     * In that case it is best to use the portable one.
-     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
-     */
-#if defined(__GNUC__) && !defined(__wasm__) \
-    && defined(__SIZEOF_INT128__) \
-    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
-
-    __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
-    XXH128_hash_t r128;
-    r128.low64  = (xxh_u64)(product);
-    r128.high64 = (xxh_u64)(product >> 64);
-    return r128;
-
-    /*
-     * MSVC for x64's _umul128 method.
-     *
-     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
-     *
-     * This compiles to single operand MUL on x64.
-     */
-#elif defined(_M_X64) || defined(_M_IA64)
-
-#ifndef _MSC_VER
-#   pragma intrinsic(_umul128)
-#endif
-    xxh_u64 product_high;
-    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
-    XXH128_hash_t r128;
-    r128.low64  = product_low;
-    r128.high64 = product_high;
-    return r128;
-
-#else
-    /*
-     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
-     *
-     * This is a fast and simple grade school multiply, which is shown below
-     * with base 10 arithmetic instead of base 0x100000000.
-     *
-     *           9 3 // D2 lhs = 93
-     *         x 7 5 // D2 rhs = 75
-     *     ----------
-     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
-     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
-     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
-     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
-     *     ---------
-     *         2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
-     *     + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
-     *     ---------
-     *       6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
-     *
-     * The reasons for adding the products like this are:
-     *  1. It avoids manual carry tracking. Just like how
-     *     (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
-     *     This avoids a lot of complexity.
-     *
-     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
-     *     instruction available in ARM's Digital Signal Processing extension
-     *     in 32-bit ARMv6 and later, which is shown below:
-     *
-     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
-     *         {
-     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
-     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
-     *             *RdHi = (xxh_u32)(product >> 32);
-     *         }
-     *
-     *     This instruction was designed for efficient long multiplication, and
-     *     allows this to be calculated in only 4 instructions at speeds
-     *     comparable to some 64-bit ALUs.
-     *
-     *  3. It isn't terrible on other platforms. Usually this will be a couple
-     *     of 32-bit ADD/ADCs.
-     */
-
-    /* First calculate all of the cross products. */
-    xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
-    xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
-    xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
-    xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32,        rhs >> 32);
-
-    /* Now add the products together. These will never overflow. */
-    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
-    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
-    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
-
-    XXH128_hash_t r128;
-    r128.low64  = lower;
-    r128.high64 = upper;
-    return r128;
-#endif
-}
-
-/*
- * Does a 64-bit to 128-bit multiply, then XOR folds it.
- *
- * The reason for the separate function is to prevent passing too many structs
- * around by value. This will hopefully inline the multiply, but we don't force it.
- */
-static xxh_u64
-XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
-{
-    XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
-    return product.low64 ^ product.high64;
-}
-
-/* Seems to produce slightly better code on GCC for some reason. */
-XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
-{
-    XXH_ASSERT(0 <= shift && shift < 64);
-    return v64 ^ (v64 >> shift);
-}
-
-/*
- * We don't need to (or want to) mix as much as XXH64.
- *
- * Short hashes are more evenly distributed, so it isn't necessary.
- */
-static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
-{
-    h64 = XXH_xorshift64(h64, 37);
-    h64 *= 0x165667919E3779F9ULL;
-    h64 = XXH_xorshift64(h64, 32);
-    return h64;
-}
-
-
-/* ==========================================
- * Short keys
- * ==========================================
- * One of the shortcomings of XXH32 and XXH64 was that their performance was
- * sub-optimal on short lengths. It used an iterative algorithm which strongly
- * favored lengths that were a multiple of 4 or 8.
- *
- * Instead of iterating over individual inputs, we use a set of single shot
- * functions which piece together a range of lengths and operate in constant time.
- *
- * Additionally, the number of multiplies has been significantly reduced. This
- * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
- *
- * Depending on the platform, this may or may not be faster than XXH32, but it
- * is almost guaranteed to be faster than XXH64.
- */
-
-/*
- * At very short lengths, there isn't enough input to fully hide secrets, or use
- * the entire secret.
- *
- * There is also only a limited amount of mixing we can do before significantly
- * impacting performance.
- *
- * Therefore, we use different sections of the secret and always mix two secret
- * samples with an XOR. This should have no effect on performance on the
- * seedless or withSeed variants because everything _should_ be constant folded
- * by modern compilers.
- *
- * The XOR mixing hides individual parts of the secret and increases entropy.
- *
- * This adds an extra layer of strength for custom secrets.
- */
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(1 <= len && len <= 3);
-    XXH_ASSERT(secret != NULL);
-    /*
-     * len = 1: combined = { input[0], 0x01, input[0], input[0] }
-     * len = 2: combined = { input[1], 0x02, input[0], input[1] }
-     * len = 3: combined = { input[2], 0x03, input[0], input[1] }
-     */
-    {   xxh_u8 const c1 = input[0];
-        xxh_u8 const c2 = input[len >> 1];
-        xxh_u8 const c3 = input[len - 1];
-        xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2  << 24)
-                               | ((xxh_u32)c3 <<  0) | ((xxh_u32)len << 8);
-        xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
-        xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
-        xxh_u64 const mixed = keyed * PRIME64_1;
-        return XXH3_avalanche(mixed);
-    }
-}
-
-XXH_FORCE_INLINE 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);
-    XXH_ASSERT(secret != NULL);
-    XXH_ASSERT(4 <= len && len < 8);
-    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
-    {   xxh_u32 const input1 = XXH_readLE32(input);
-        xxh_u32 const input2 = XXH_readLE32(input + len - 4);
-        xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
-        xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
-        xxh_u64 x = input64 ^ bitflip;
-        /* this mix is inspired by Pelle Evensen's rrmxmx */
-        x ^= XXH_rotl64(x, 49) ^ XXH_rotl64(x, 24);
-        x *= 0x9FB21C651E98DF25ULL;
-        x ^= (x >> 35) + len ;
-        x *= 0x9FB21C651E98DF25ULL;
-        return XXH_xorshift64(x, 28);
-    }
-}
-
-XXH_FORCE_INLINE 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);
-    XXH_ASSERT(secret != NULL);
-    XXH_ASSERT(8 <= len && len <= 16);
-    {   xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
-        xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
-        xxh_u64 const input_lo = XXH_readLE64(input)           ^ bitflip1;
-        xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
-        xxh_u64 const acc = len
-                          + XXH_swap64(input_lo) + input_hi
-                          + XXH3_mul128_fold64(input_lo, input_hi);
-        return XXH3_avalanche(acc);
-    }
-}
-
-XXH_FORCE_INLINE 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);
-    {   if (XXH_likely(len >  8)) return XXH3_len_9to16_64b(input, len, secret, seed);
-        if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
-        if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
-        return XXH3_avalanche((PRIME64_1 + seed) ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
-    }
-}
-
-/*
- * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
- * multiplication by zero, affecting hashes of lengths 17 to 240.
- *
- * However, they are very unlikely.
- *
- * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
- * unseeded non-cryptographic hashes, it does not attempt to defend itself
- * against specially crafted inputs, only random inputs.
- *
- * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
- * cancelling out the secret is taken an arbitrary number of times (addressed
- * in XXH3_accumulate_512), this collision is very unlikely with random inputs
- * and/or proper seeding:
- *
- * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
- * function that is only called up to 16 times per hash with up to 240 bytes of
- * input.
- *
- * This is not too bad for a non-cryptographic hash function, especially with
- * only 64 bit outputs.
- *
- * The 128-bit variant (which trades some speed for strength) is NOT affected
- * by this, although it is always a good idea to use a proper seed if you care
- * about strength.
- */
-XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
-                                     const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
-{
-#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__i386__) && defined(__SSE2__)  /* x86 + SSE2 */ \
-  && !defined(XXH_ENABLE_AUTOVECTORIZE)      /* Define to disable like XXH32 hack */
-    /*
-     * UGLY HACK:
-     * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
-     * slower code.
-     *
-     * By forcing seed64 into a register, we disrupt the cost model and
-     * cause it to scalarize. See `XXH32_round()`
-     *
-     * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
-     * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
-     * GCC 9.2, despite both emitting scalar code.
-     *
-     * GCC generates much better scalar code than Clang for the rest of XXH3,
-     * which is why finding a more optimal codepath is an interest.
-     */
-    __asm__ ("" : "+r" (seed64));
-#endif
-    {   xxh_u64 const input_lo = XXH_readLE64(input);
-        xxh_u64 const input_hi = XXH_readLE64(input+8);
-        return XXH3_mul128_fold64(
-            input_lo ^ (XXH_readLE64(secret)   + seed64),
-            input_hi ^ (XXH_readLE64(secret+8) - seed64)
-        );
-    }
-}
-
-/* For mid range keys, XXH3 uses a Mum-hash variant. */
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
-                     const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                     XXH64_hash_t seed)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
-    XXH_ASSERT(16 < len && len <= 128);
-
-    {   xxh_u64 acc = len * PRIME64_1;
-        if (len > 32) {
-            if (len > 64) {
-                if (len > 96) {
-                    acc += XXH3_mix16B(input+48, secret+96, seed);
-                    acc += 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);
-    }
-}
-
-#define XXH3_MIDSIZE_MAX 240
-
-XXH_NO_INLINE 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)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
-    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
-
-    #define XXH3_MIDSIZE_STARTOFFSET 3
-    #define XXH3_MIDSIZE_LASTOFFSET  17
-
-    {   xxh_u64 acc = len * PRIME64_1;
-        int const nbRounds = (int)len / 16;
-        int i;
-        for (i=0; i<8; i++) {
-            acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
-        }
-        acc = XXH3_avalanche(acc);
-        XXH_ASSERT(nbRounds >= 8);
-#if defined(__clang__)                                /* Clang */ \
-    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
-    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
-        /*
-         * UGLY HACK:
-         * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
-         * In everywhere else, it uses scalar code.
-         *
-         * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
-         * would still be slower than UMAAL (see XXH_mult64to128).
-         *
-         * Unfortunately, Clang doesn't handle the long multiplies properly and
-         * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
-         * scalarized into an ugly mess of VMOV.32 instructions.
-         *
-         * This mess is difficult to avoid without turning autovectorization
-         * off completely, but they are usually relatively minor and/or not
-         * worth it to fix.
-         *
-         * This loop is the easiest to fix, as unlike XXH32, this pragma
-         * _actually works_ because it is a loop vectorization instead of an
-         * SLP vectorization.
-         */
-        #pragma clang loop vectorize(disable)
-#endif
-        for (i=8 ; i < nbRounds; i++) {
-            acc += 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);
-    }
-}
-
-
-/* ===    Long Keys    === */
-
-#define STRIPE_LEN 64
-#define XXH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
-#define ACC_NB (STRIPE_LEN / sizeof(xxh_u64))
-
-typedef enum { XXH3_acc_64bits, XXH3_acc_128bits } XXH3_accWidth_e;
-
-/*
- * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
- *
- * It is a hardened version of UMAC, based off of FARSH's implementation.
- *
- * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
- * implementations, and it is ridiculously fast.
- *
- * We harden it by mixing the original input to the accumulators as well as the product.
- *
- * This means that in the (relatively likely) case of a multiply by zero, the
- * original input is preserved.
- *
- * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
- * cross-pollination, as otherwise the upper and lower halves would be
- * essentially independent.
- *
- * This doesn't matter on 64-bit hashes since they all get merged together in
- * the end, so we skip the extra step.
- *
- * Both XXH3_64bits and XXH3_128bits use this subroutine.
- */
-XXH_FORCE_INLINE void
-XXH3_accumulate_512(      void* XXH_RESTRICT acc,
-                    const void* XXH_RESTRICT input,
-                    const void* XXH_RESTRICT secret,
-                    XXH3_accWidth_e accWidth)
-{
-#if (XXH_VECTOR == XXH_AVX512)
-
-    XXH_ASSERT((((size_t)acc) & 63) == 0);
-    XXH_STATIC_ASSERT(STRIPE_LEN == sizeof(__m512i));
-    {   XXH_ALIGN(64) __m512i* const xacc    =       (__m512i *) acc;
-
-        /* data_vec    = input[0]; */
-        __m512i const data_vec    = _mm512_loadu_si512   (input);
-        /* key_vec     = secret[0]; */
-        __m512i const key_vec     = _mm512_loadu_si512   (secret);
-        /* 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_SHUFFLE(0, 3, 0, 1));
-        /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
-        __m512i const product     = _mm512_mul_epu32     (data_key, data_key_lo);
-        if (accWidth == XXH3_acc_128bits) {
-            /* xacc[0] += swap(data_vec); */
-            __m512i const data_swap = _mm512_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
-            __m512i const sum       = _mm512_add_epi64(*xacc, data_swap);
-            /* xacc[0] += product; */
-            *xacc = _mm512_add_epi64(product, sum);
-        } else {  /* XXH3_acc_64bits */
-            /* xacc[0] += data_vec; */
-            __m512i const sum = _mm512_add_epi64(*xacc, data_vec);
-            /* xacc[0] += product; */
-            *xacc = _mm512_add_epi64(product, sum);
-        }
-    }
-
-#elif (XXH_VECTOR == XXH_AVX2)
-
-    XXH_ASSERT((((size_t)acc) & 31) == 0);
-    {   XXH_ALIGN(32) __m256i* const xacc    =       (__m256i *) acc;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
-        const         __m256i* const xinput  = (const __m256i *) input;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
-        const         __m256i* const xsecret = (const __m256i *) secret;
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) {
-            /* data_vec    = xinput[i]; */
-            __m256i const data_vec    = _mm256_loadu_si256    (xinput+i);
-            /* key_vec     = xsecret[i]; */
-            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
-            /* 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));
-            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
-            __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
-            if (accWidth == XXH3_acc_128bits) {
-                /* xacc[i] += swap(data_vec); */
-                __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
-                __m256i const sum       = _mm256_add_epi64(xacc[i], data_swap);
-                /* xacc[i] += product; */
-                xacc[i] = _mm256_add_epi64(product, sum);
-            } else {  /* XXH3_acc_64bits */
-                /* xacc[i] += data_vec; */
-                __m256i const sum = _mm256_add_epi64(xacc[i], data_vec);
-                /* xacc[i] += product; */
-                xacc[i] = _mm256_add_epi64(product, sum);
-            }
-    }   }
-
-#elif (XXH_VECTOR == XXH_SSE2)
-
-    /* SSE2 is just a half-scale version of the AVX2 version. */
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {   XXH_ALIGN(16) __m128i* const xacc    =       (__m128i *) acc;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
-        const         __m128i* const xinput  = (const __m128i *) input;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
-        const         __m128i* const xsecret = (const __m128i *) secret;
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) {
-            /* data_vec    = xinput[i]; */
-            __m128i const data_vec    = _mm_loadu_si128   (xinput+i);
-            /* key_vec     = xsecret[i]; */
-            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
-            /* data_key    = data_vec ^ key_vec; */
-            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
-            /* data_key_lo = data_key >> 32; */
-            __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
-            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
-            __m128i const product     = _mm_mul_epu32     (data_key, data_key_lo);
-            if (accWidth == XXH3_acc_128bits) {
-                /* xacc[i] += swap(data_vec); */
-                __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
-                __m128i const sum       = _mm_add_epi64(xacc[i], data_swap);
-                /* xacc[i] += product; */
-                xacc[i] = _mm_add_epi64(product, sum);
-            } else {  /* XXH3_acc_64bits */
-                /* xacc[i] += data_vec; */
-                __m128i const sum = _mm_add_epi64(xacc[i], data_vec);
-                /* xacc[i] += product; */
-                xacc[i] = _mm_add_epi64(product, sum);
-            }
-    }   }
-
-#elif (XXH_VECTOR == XXH_NEON)
-
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {
-        XXH_ALIGN(16) uint64x2_t* const xacc = (uint64x2_t *) acc;
-        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
-        uint8_t const* const xinput = (const uint8_t *) input;
-        uint8_t const* const xsecret  = (const uint8_t *) secret;
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN / sizeof(uint64x2_t); i++) {
-            /* data_vec = xinput[i]; */
-            uint8x16_t data_vec    = vld1q_u8(xinput  + (i * 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;
-            if (accWidth == XXH3_acc_64bits) {
-                /* xacc[i] += data_vec; */
-                xacc[i] = vaddq_u64 (xacc[i], vreinterpretq_u64_u8(data_vec));
-            } else {  /* XXH3_acc_128bits */
-                /* 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);
-            }
-            /* 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);
-
-        }
-    }
-
-#elif (XXH_VECTOR == XXH_VSX)
-          xxh_u64x2* const xacc     =       (xxh_u64x2*) acc;    /* presumed aligned */
-    xxh_u64x2 const* const xinput   = (xxh_u64x2 const*) input;   /* no alignment restriction */
-    xxh_u64x2 const* const xsecret  = (xxh_u64x2 const*) secret;    /* no alignment restriction */
-    xxh_u64x2 const v32 = { 32, 32 };
-    size_t i;
-    for (i = 0; i < STRIPE_LEN / sizeof(xxh_u64x2); i++) {
-        /* data_vec = xinput[i]; */
-        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + i);
-        /* key_vec = xsecret[i]; */
-        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 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);
-        xacc[i] += product;
-
-        if (accWidth == XXH3_acc_64bits) {
-            xacc[i] += data_vec;
-        } else {  /* XXH3_acc_128bits */
-            /* swap high and low halves */
-#ifdef __s390x__
-            xxh_u64x2 const data_swapped = vec_permi(data_vec, data_vec, 2);
-#else
-            xxh_u64x2 const data_swapped = vec_xxpermdi(data_vec, data_vec, 2);
-#endif
-            xacc[i] += data_swapped;
-        }
-    }
-
-#else   /* scalar variant of Accumulator - universal */
-
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */
-    const xxh_u8* const xinput  = (const xxh_u8*) input;  /* no alignment restriction */
-    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
-    size_t i;
-    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
-    for (i=0; i < 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);
-
-        if (accWidth == XXH3_acc_64bits) {
-            xacc[i] += data_val;
-        } else {
-            xacc[i ^ 1] += data_val; /* swap adjacent lanes */
-        }
-        xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32);
-    }
-#endif
-}
-
-/*
- * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
- *
- * Multiplication isn't perfect, as explained by Google in HighwayHash:
- *
- *  // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
- *  // varying degrees. In descending order of goodness, bytes
- *  // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
- *  // As expected, the upper and lower bytes are much worse.
- *
- * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
- *
- * Since our algorithm uses a pseudorandom secret to add some variance into the
- * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
- *
- * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
- * extraction.
- *
- * Both XXH3_64bits and XXH3_128bits use this subroutine.
- */
-XXH_FORCE_INLINE void
-XXH3_scrambleAcc(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
-{
-#if (XXH_VECTOR == XXH_AVX512)
-
-    XXH_ASSERT((((size_t)acc) & 63) == 0);
-    XXH_STATIC_ASSERT(STRIPE_LEN == sizeof(__m512i));
-    {   XXH_ALIGN(64) __m512i* const xacc = (__m512i*) acc;
-        const __m512i prime32 = _mm512_set1_epi32((int)PRIME32_1);
-
-        /* xacc[0] ^= (xacc[0] >> 47) */
-        __m512i const acc_vec     = *xacc;
-        __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
-        __m512i const data_vec    = _mm512_xor_si512     (acc_vec, shifted);
-        /* xacc[0] ^= secret; */
-        __m512i const key_vec     = _mm512_loadu_si512   (secret);
-        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
-
-        /* xacc[0] *= PRIME32_1; */
-        __m512i const data_key_hi = _mm512_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
-        __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));
-    }
-
-#elif (XXH_VECTOR == XXH_AVX2)
-
-    XXH_ASSERT((((size_t)acc) & 31) == 0);
-    {   XXH_ALIGN(32) __m256i* const xacc = (__m256i*) acc;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
-        const         __m256i* const xsecret = (const __m256i *) secret;
-        const __m256i prime32 = _mm256_set1_epi32((int)PRIME32_1);
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) {
-            /* xacc[i] ^= (xacc[i] >> 47) */
-            __m256i const acc_vec     = xacc[i];
-            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
-            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
-            /* xacc[i] ^= xsecret; */
-            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
-            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
-
-            /* xacc[i] *= PRIME32_1; */
-            __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
-            __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));
-        }
-    }
-
-#elif (XXH_VECTOR == XXH_SSE2)
-
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {   XXH_ALIGN(16) __m128i* const xacc = (__m128i*) acc;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
-        const         __m128i* const xsecret = (const __m128i *) secret;
-        const __m128i prime32 = _mm_set1_epi32((int)PRIME32_1);
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) {
-            /* xacc[i] ^= (xacc[i] >> 47) */
-            __m128i const acc_vec     = xacc[i];
-            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
-            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
-            /* xacc[i] ^= xsecret[i]; */
-            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
-            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
-
-            /* xacc[i] *= PRIME32_1; */
-            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
-            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
-            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
-            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
-        }
-    }
-
-#elif (XXH_VECTOR == XXH_NEON)
-
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-
-    {   uint64x2_t* xacc       = (uint64x2_t*) acc;
-        uint8_t const* xsecret = (uint8_t const*) secret;
-        uint32x2_t prime       = vdup_n_u32 (PRIME32_1);
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(uint64x2_t); i++) {
-            /* xacc[i] ^= (xacc[i] >> 47); */
-            uint64x2_t acc_vec  = xacc[i];
-            uint64x2_t shifted  = vshrq_n_u64 (acc_vec, 47);
-            uint64x2_t data_vec = veorq_u64   (acc_vec, shifted);
-
-            /* xacc[i] ^= xsecret[i]; */
-            uint8x16_t key_vec  = vld1q_u8(xsecret + (i * 16));
-            uint64x2_t data_key = veorq_u64(data_vec, vreinterpretq_u64_u8(key_vec));
-
-            /* xacc[i] *= 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) * 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) * PRIME32_1; */
-                xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime);
-            }
-    }   }
-
-#elif (XXH_VECTOR == XXH_VSX)
-
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-
-    {         xxh_u64x2* const xacc    =       (xxh_u64x2*) acc;
-        const xxh_u64x2* const xsecret = (const xxh_u64x2*) secret;
-        /* constants */
-        xxh_u64x2 const v32  = { 32, 32 };
-        xxh_u64x2 const v47 = { 47, 47 };
-        xxh_u32x4 const prime = { PRIME32_1, PRIME32_1, PRIME32_1, PRIME32_1 };
-        size_t i;
-        for (i = 0; i < STRIPE_LEN / sizeof(xxh_u64x2); i++) {
-            /* xacc[i] ^= (xacc[i] >> 47); */
-            xxh_u64x2 const acc_vec  = xacc[i];
-            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 data_key = data_vec ^ key_vec;
-
-            /* xacc[i] *= PRIME32_1 */
-            /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF);  */
-            xxh_u64x2 const prod_even  = XXH_vec_mule((xxh_u32x4)data_key, prime);
-            /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32);  */
-            xxh_u64x2 const prod_odd  = XXH_vec_mulo((xxh_u32x4)data_key, prime);
-            xacc[i] = prod_odd + (prod_even << v32);
-    }   }
-
-#else   /* scalar variant of Scrambler - universal */
-
-    XXH_ALIGN(XXH_ACC_ALIGN) 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 < ACC_NB; i++) {
-        xxh_u64 const key64 = XXH_readLE64(xsecret + 8*i);
-        xxh_u64 acc64 = xacc[i];
-        acc64 = XXH_xorshift64(acc64, 47);
-        acc64 ^= key64;
-        acc64 *= PRIME32_1;
-        xacc[i] = acc64;
-    }
-
-#endif
-}
-
-#define XXH_PREFETCH_DIST 384
-
-#ifdef __clang__ // for clang
-#  define XXH_PREFETCH_DIST_AVX512_64  320
-#  define XXH_PREFETCH_DIST_AVX512_128 320
-#else // for gcc
-#  define XXH_PREFETCH_DIST_AVX512_64  640
-#  define XXH_PREFETCH_DIST_AVX512_128 512
-#endif
-
-/*
- * 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_accWidth_e accWidth)
-{
-    size_t n;
-    for (n = 0; n < nbStripes; n++ ) {
-        const xxh_u8* const in = input + n*STRIPE_LEN;
-#if (XXH_VECTOR == XXH_AVX512)
-        if (accWidth == XXH3_acc_64bits) XXH_PREFETCH(in + XXH_PREFETCH_DIST_AVX512_64);
-        else                             XXH_PREFETCH(in + XXH_PREFETCH_DIST_AVX512_128);
-#else
-        XXH_PREFETCH(in + XXH_PREFETCH_DIST);
-#endif
-        XXH3_accumulate_512(acc,
-                            in,
-                            secret + n*XXH_SECRET_CONSUME_RATE,
-                            accWidth);
-    }
-}
-
-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_accWidth_e accWidth)
-{
-    size_t const nb_rounds = (secretSize - STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
-    size_t const block_len = STRIPE_LEN * nb_rounds;
-    size_t const nb_blocks = len / block_len;
-
-    size_t n;
-
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
-
-    for (n = 0; n < nb_blocks; n++) {
-        XXH3_accumulate(acc, input + n*block_len, secret, nb_rounds, accWidth);
-        XXH3_scrambleAcc(acc, secret + secretSize - STRIPE_LEN);
-    }
-
-    /* last partial block */
-    XXH_ASSERT(len > STRIPE_LEN);
-    {   size_t const nbStripes = (len - (block_len * nb_blocks)) / STRIPE_LEN;
-        XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
-        XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, accWidth);
-
-        /* last stripe */
-        if (len & (STRIPE_LEN - 1)) {
-            const xxh_u8* const p = input + len - STRIPE_LEN;
-            /* Do not align on 8, so that the secret is different from the scrambler */
-#define XXH_SECRET_LASTACC_START 7
-            XXH3_accumulate_512(acc, p, secret + secretSize - STRIPE_LEN - XXH_SECRET_LASTACC_START, accWidth);
-    }   }
-}
-
-XXH_FORCE_INLINE xxh_u64
-XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
-{
-    return XXH3_mul128_fold64(
-               acc[0] ^ XXH_readLE64(secret),
-               acc[1] ^ XXH_readLE64(secret+8) );
-}
-
-static XXH64_hash_t
-XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
-{
-    xxh_u64 result64 = start;
-    size_t i = 0;
-
-    for (i = 0; i < 4; i++) {
-        result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
-#if defined(__clang__)                                /* Clang */ \
-    && (defined(__arm__) || defined(__thumb__))       /* ARMv7 */ \
-    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */  \
-    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
-        /*
-         * UGLY HACK:
-         * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
-         * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
-         * XXH3_64bits, len == 256, Snapdragon 835:
-         *   without hack: 2063.7 MB/s
-         *   with hack:    2560.7 MB/s
-         */
-        __asm__("" : "+r" (result64));
-#endif
-    }
-
-    return XXH3_avalanche(result64);
-}
-
-#define XXH3_INIT_ACC { PRIME32_3, PRIME64_1, PRIME64_2, PRIME64_3, \
-                        PRIME64_4, PRIME32_2, PRIME64_5, PRIME32_1 }
-
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_hashLong_64b_internal(const xxh_u8* XXH_RESTRICT input, size_t len,
-                           const xxh_u8* XXH_RESTRICT secret, size_t secretSize)
-{
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXH3_INIT_ACC;
-
-    XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXH3_acc_64bits);
-
-    /* converge into final hash */
-    XXH_STATIC_ASSERT(sizeof(acc) == 64);
-    /* do not align on 8, so that the secret is different from the accumulator */
-#define XXH_SECRET_MERGEACCS_START 11
-    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
-    return XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * PRIME64_1);
-}
-
-XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
-{
-    if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
-    memcpy(dst, &v64, sizeof(v64));
-}
-
-/* XXH3_initCustomSecret() :
- * destination `customSecret` is presumed allocated and same size as `kSecret`.
- */
-XXH_FORCE_INLINE void XXH3_initCustomSecret(xxh_u8* XXH_RESTRICT customSecret, xxh_u64 seed64)
-{
-    int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
-    int i;
-    /*
-     * We need a separate pointer for the hack below.
-     * Any decent compiler will optimize this out otherwise.
-     */
-    const xxh_u8 *kSecretPtr = kSecret;
-
-    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
-
-#if defined(__clang__) && defined(__aarch64__)
-    /*
-     * UGLY HACK:
-     * Clang generates 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:
-     *   I   L   S
-     * MOVK
-     * MOVK
-     * MOVK
-     * MOVK
-     * ADD
-     * SUB      STR
-     *          STR
-     * By forcing loads from memory (as the asm line causes Clang to assume
-     * that kSecretPtr has been changed), the pipelines are used more efficiently:
-     *   I   L   S
-     *      LDR
-     *  ADD LDR
-     *  SUB     STR
-     *          STR
-     * XXH3_64bits_withSeed, len == 256, Snapdragon 835
-     *   without hack: 2654.4 MB/s
-     *   with hack:    3202.9 MB/s
-     */
-    __asm__("" : "+r" (kSecretPtr));
-#endif
-    /*
-     * Note: in debug mode, this overrides the asm optimization
-     * and Clang will emit MOVK chains again.
-     */
-    XXH_ASSERT(kSecretPtr == kSecret);
-
-    for (i=0; i < nbRounds; i++) {
-        /*
-         * The asm hack causes Clang to assume that kSecretPtr aliases with
-         * customSecret, and on aarch64, this prevented LDP from merging two
-         * loads together for free. Putting the loads together before the stores
-         * properly generates LDP.
-         */
-        xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
-        xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
-        XXH_writeLE64(customSecret + 16*i,     lo);
-        XXH_writeLE64(customSecret + 16*i + 8, hi);
-    }
-}
-
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH64_hash_t
-XXH3_hashLong_64b_defaultSecret(const xxh_u8* XXH_RESTRICT input, size_t len)
-{
-    return XXH3_hashLong_64b_internal(input, len, kSecret, sizeof(kSecret));
-}
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH64_hash_t
-XXH3_hashLong_64b_withSecret(const xxh_u8* XXH_RESTRICT input, size_t len,
-                             const xxh_u8* XXH_RESTRICT secret, size_t secretSize)
-{
-    return XXH3_hashLong_64b_internal(input, len, secret, secretSize);
-}
-
-/*
- * XXH3_hashLong_64b_withSeed():
- * Generate a custom key based on alteration of default kSecret with the seed,
- * and then use this key for long mode hashing.
- *
- * This operation is decently fast but nonetheless costs a little bit of time.
- * Try to avoid it whenever possible (typically when seed==0).
- *
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH64_hash_t
-XXH3_hashLong_64b_withSeed(const xxh_u8* input, size_t len, XXH64_hash_t seed)
-{
-    XXH_ALIGN(8) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
-    if (seed==0) return XXH3_hashLong_64b_defaultSecret(input, len);
-    XXH3_initCustomSecret(secret, seed);
-    return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret));
-}
-
-/* ===   Public entry point   === */
-
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len)
-{
-    if (len <= 16)
-        return XXH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, 0);
-    if (len <= 128)
-        return XXH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
-    if (len <= XXH3_MIDSIZE_MAX)
-         return XXH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
-    return XXH3_hashLong_64b_defaultSecret((const xxh_u8*)input, len);
-}
-
-XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
-    /*
-     * If an action is to be taken if `secret` conditions are not respected,
-     * it should be done here.
-     * For now, it's a contract pre-condition.
-     * Adding a check and a branch here would cost performance at every hash.
-     */
-    if (len <= 16)
-        return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0);
-    if (len <= 128)
-        return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
-    if (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
-    return XXH3_hashLong_64b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize);
-}
-
-XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
-{
-    if (len <= 16)
-        return XXH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, seed);
-    if (len <= 128)
-        return XXH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
-    if (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
-    return XXH3_hashLong_64b_withSeed((const xxh_u8*)input, len, seed);
-}
-
-/* ===   XXH3 streaming   === */
-
-
-/*
- * Malloc's a pointer that is always aligned to align.
- *
- * This must be freed with `XXH_alignedFree()`.
- *
- * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
- * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
- * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
+ * xxh3.h is no longer useful,
+ * but it is still provided for compatibility with source code
+ * which used to include it directly.
  *
- * This underalignment previously caused a rather obvious crash which went
- * completely unnoticed due to XXH3_createState() not actually being tested.
- * Credit to RedSpah for noticing this bug.
+ * Programs are now highly discouraged to include xxh3.h.
+ * Include `xxhash.h` instead, which is the officially supported interface.
  *
- * The alignment is done manually: Functions like posix_memalign or _mm_malloc
- * are avoided: To maintain portability, we would have to write a fallback
- * like this anyways, and besides, testing for the existence of library
- * functions without relying on external build tools is impossible.
- *
- * The method is simple: Overallocate, manually align, and store the offset
- * to the original behind the returned pointer.
- *
- * Align must be a power of 2 and 8 <= align <= 128.
- */
-static void* XXH_alignedMalloc(size_t s, size_t align)
-{
-    XXH_ASSERT(align <= 128 && align >= 8); /* range check */
-    XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
-    XXH_ASSERT(s != 0 && s < (s + align));  /* empty/overflow */
-    {   /* Overallocate to make room for manual realignment and an offset byte */
-        xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
-        if (base != NULL) {
-            /*
-             * Get the offset needed to align this pointer.
-             *
-             * Even if the returned pointer is aligned, there will always be
-             * at least one byte to store the offset to the original pointer.
-             */
-            size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
-            /* Add the offset for the now-aligned pointer */
-            xxh_u8* ptr = base + offset;
-
-            XXH_ASSERT((size_t)ptr % align == 0);
-
-            /* Store the offset immediately before the returned pointer. */
-            ptr[-1] = (xxh_u8)offset;
-            return ptr;
-        }
-        return NULL;
-    }
-}
-/*
- * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
- * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
+ * In the future, xxh3.h will start to generate warnings, then errors,
+ * then it will be removed from source package and from include directory.
  */
-static void XXH_alignedFree(void* p)
-{
-    if (p != NULL) {
-        xxh_u8* ptr = (xxh_u8*)p;
-        /* Get the offset byte we added in XXH_malloc. */
-        xxh_u8 offset = ptr[-1];
-        /* Free the original malloc'd pointer */
-        xxh_u8* base = ptr - offset;
-        XXH_free(base);
-    }
-}
-XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
-{
-    return (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
-}
-
-XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
-{
-    XXH_alignedFree(statePtr);
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API void
-XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state)
-{
-    memcpy(dst_state, src_state, sizeof(*dst_state));
-}
-
-static void
-XXH3_64bits_reset_internal(XXH3_state_t* statePtr,
-                           XXH64_hash_t seed,
-                           const xxh_u8* secret, size_t secretSize)
-{
-    XXH_ASSERT(statePtr != NULL);
-    memset(statePtr, 0, sizeof(*statePtr));
-    statePtr->acc[0] = PRIME32_3;
-    statePtr->acc[1] = PRIME64_1;
-    statePtr->acc[2] = PRIME64_2;
-    statePtr->acc[3] = PRIME64_3;
-    statePtr->acc[4] = PRIME64_4;
-    statePtr->acc[5] = PRIME32_2;
-    statePtr->acc[6] = PRIME64_5;
-    statePtr->acc[7] = PRIME32_1;
-    statePtr->seed = seed;
-    XXH_ASSERT(secret != NULL);
-    statePtr->secret = secret;
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
-    statePtr->secretLimit = (XXH32_hash_t)(secretSize - STRIPE_LEN);
-    statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset(XXH3_state_t* statePtr)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_64bits_reset_internal(statePtr, 0, kSecret, XXH_SECRET_DEFAULT_SIZE);
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_64bits_reset_internal(statePtr, 0, (const xxh_u8*)secret, secretSize);
-    if (secret == NULL) return XXH_ERROR;
-    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_64bits_reset_internal(statePtr, seed, kSecret, XXH_SECRET_DEFAULT_SIZE);
-    XXH3_initCustomSecret(statePtr->customSecret, seed);
-    statePtr->secret = statePtr->customSecret;
-    return XXH_OK;
-}
-
-XXH_FORCE_INLINE void
-XXH3_consumeStripes( xxh_u64* acc,
-                    XXH32_hash_t* nbStripesSoFarPtr, XXH32_hash_t nbStripesPerBlock,
-                    const xxh_u8* input, size_t totalStripes,
-                    const xxh_u8* secret, size_t secretLimit,
-                    XXH3_accWidth_e accWidth)
-{
-    XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock);
-    if (nbStripesPerBlock - *nbStripesSoFarPtr <= totalStripes) {
-        /* need a scrambling operation */
-        size_t const nbStripes = nbStripesPerBlock - *nbStripesSoFarPtr;
-        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, accWidth);
-        XXH3_scrambleAcc(acc, secret + secretLimit);
-        XXH3_accumulate(acc, input + nbStripes * STRIPE_LEN, secret, totalStripes - nbStripes, accWidth);
-        *nbStripesSoFarPtr = (XXH32_hash_t)(totalStripes - nbStripes);
-    } else {
-        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, totalStripes, accWidth);
-        *nbStripesSoFarPtr += (XXH32_hash_t)totalStripes;
-    }
-}
-
-/*
- * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
- */
-XXH_FORCE_INLINE XXH_errorcode
-XXH3_update(XXH3_state_t* state, const xxh_u8* input, size_t len, XXH3_accWidth_e accWidth)
-{
-    if (input==NULL)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-        return XXH_OK;
-#else
-        return XXH_ERROR;
-#endif
-
-    {   const xxh_u8* const bEnd = input + len;
-
-        state->totalLen += len;
-
-        if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) {  /* fill in tmp buffer */
-            XXH_memcpy(state->buffer + state->bufferedSize, input, len);
-            state->bufferedSize += (XXH32_hash_t)len;
-            return XXH_OK;
-        }
-        /* input is now > XXH3_INTERNALBUFFER_SIZE */
-
-        #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / STRIPE_LEN)
-        XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % STRIPE_LEN == 0);   /* clean multiple */
-
-        /*
-         * There is some input left inside the internal buffer.
-         * Fill it, then consume it.
-         */
-        if (state->bufferedSize) {
-            size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
-            XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
-            input += loadSize;
-            XXH3_consumeStripes(state->acc,
-                               &state->nbStripesSoFar, state->nbStripesPerBlock,
-                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
-                                state->secret, state->secretLimit,
-                                accWidth);
-            state->bufferedSize = 0;
-        }
-
-        /* Consume input by full buffer quantities */
-        if (input+XXH3_INTERNALBUFFER_SIZE <= bEnd) {
-            const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE;
-            do {
-                XXH3_consumeStripes(state->acc,
-                                   &state->nbStripesSoFar, state->nbStripesPerBlock,
-                                    input, XXH3_INTERNALBUFFER_STRIPES,
-                                    state->secret, state->secretLimit,
-                                    accWidth);
-                input += XXH3_INTERNALBUFFER_SIZE;
-            } while (input<=limit);
-        }
-
-        if (input < bEnd) { /* Some remaining input: buffer it */
-            XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
-            state->bufferedSize = (XXH32_hash_t)(bEnd-input);
-        }
-    }
 
-    return XXH_OK;
-}
+/* Simulate the same impact as including the old xxh3.h source file */
 
-XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len)
-{
-    return XXH3_update(state, (const xxh_u8*)input, len, XXH3_acc_64bits);
-}
-
-
-XXH_FORCE_INLINE void
-XXH3_digest_long (XXH64_hash_t* acc, const XXH3_state_t* state, XXH3_accWidth_e accWidth)
-{
-    /*
-     * Digest on a local copy. This way, the state remains unaltered, and it can
-     * continue ingesting more input afterwards.
-     */
-    memcpy(acc, state->acc, sizeof(state->acc));
-    if (state->bufferedSize >= STRIPE_LEN) {
-        size_t const totalNbStripes = state->bufferedSize / STRIPE_LEN;
-        XXH32_hash_t nbStripesSoFar = state->nbStripesSoFar;
-        XXH3_consumeStripes(acc,
-                           &nbStripesSoFar, state->nbStripesPerBlock,
-                            state->buffer, totalNbStripes,
-                            state->secret, state->secretLimit,
-                            accWidth);
-        if (state->bufferedSize % STRIPE_LEN) {  /* one last partial stripe */
-            XXH3_accumulate_512(acc,
-                                state->buffer + state->bufferedSize - STRIPE_LEN,
-                                state->secret + state->secretLimit - XXH_SECRET_LASTACC_START,
-                                accWidth);
-        }
-    } else {  /* bufferedSize < STRIPE_LEN */
-        if (state->bufferedSize) { /* one last stripe */
-            xxh_u8 lastStripe[STRIPE_LEN];
-            size_t const catchupSize = STRIPE_LEN - state->bufferedSize;
-            memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
-            memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
-            XXH3_accumulate_512(acc,
-                                lastStripe,
-                                state->secret + state->secretLimit - XXH_SECRET_LASTACC_START,
-                                accWidth);
-    }   }
-}
-
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
-{
-    if (state->totalLen > XXH3_MIDSIZE_MAX) {
-        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[ACC_NB];
-        XXH3_digest_long(acc, state, XXH3_acc_64bits);
-        return XXH3_mergeAccs(acc,
-                              state->secret + XXH_SECRET_MERGEACCS_START,
-                              (xxh_u64)state->totalLen * PRIME64_1);
-    }
-    /* len <= XXH3_MIDSIZE_MAX: short code */
-    if (state->seed)
-        return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
-    return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
-                                  state->secret, state->secretLimit + STRIPE_LEN);
-}
-
-/* ==========================================
- * XXH3 128 bits (a.k.a XXH128)
- * ==========================================
- * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
- * even without counting the significantly larger output size.
- *
- * For example, extra steps are taken to avoid the seed-dependent collisions
- * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
- *
- * This strength naturally comes at the cost of some speed, especially on short
- * lengths. Note that longer hashes are about as fast as the 64-bit version
- * due to it using only a slight modification of the 64-bit loop.
- *
- * XXH128 is also more oriented towards 64-bit machines. It is still extremely
- * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
- */
-
-XXH_FORCE_INLINE 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. */
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(1 <= len && len <= 3);
-    XXH_ASSERT(secret != NULL);
-    /*
-     * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
-     * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
-     * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
-     */
-    {   xxh_u8 const c1 = input[0];
-        xxh_u8 const c2 = input[len >> 1];
-        xxh_u8 const c3 = input[len - 1];
-        xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
-                                | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
-        xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
-        xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
-        xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
-        xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
-        xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
-        xxh_u64 const mixedl = keyed_lo * PRIME64_1;
-        xxh_u64 const mixedh = keyed_hi * PRIME64_5;
-        XXH128_hash_t h128;
-        h128.low64  = XXH3_avalanche(mixedl);
-        h128.high64 = XXH3_avalanche(mixedh);
-        return h128;
-    }
-}
-
-XXH_FORCE_INLINE 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);
-    XXH_ASSERT(secret != NULL);
-    XXH_ASSERT(4 <= len && len <= 8);
-    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
-    {   xxh_u32 const input_lo = XXH_readLE32(input);
-        xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
-        xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
-        xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
-        xxh_u64 const keyed = input_64 ^ bitflip;
-
-        /* Shift len to the left to ensure it is even, this avoids even multiplies. */
-        XXH128_hash_t m128 = XXH_mult64to128(keyed, PRIME64_1 + (len << 2));
-
-        m128.high64 += (m128.low64 << 1);
-        m128.low64  ^= (m128.high64 >> 3);
-
-        m128.low64   = XXH_xorshift64(m128.low64, 35);
-        m128.low64  *= 0x9FB21C651E98DF25ULL;
-        m128.low64   = XXH_xorshift64(m128.low64, 28);
-        m128.high64  = XXH3_avalanche(m128.high64);
-        return m128;
-    }
-}
-
-XXH_FORCE_INLINE XXH128_hash_t
-XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(secret != NULL);
-    XXH_ASSERT(9 <= len && len <= 16);
-    {   xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
-        xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
-        xxh_u64 const input_lo = XXH_readLE64(input);
-        xxh_u64       input_hi = XXH_readLE64(input + len - 8);
-        XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, PRIME64_1);
-        /*
-         * Put len in the middle of m128 to ensure that the length gets mixed to
-         * both the low and high bits in the 128x64 multiply below.
-         */
-        m128.low64 += (xxh_u64)(len - 1) << 54;
-        input_hi   ^= bitfliph;
-        /*
-         * Add the high 32 bits of input_hi to the high 32 bits of m128, then
-         * add the long product of the low 32 bits of input_hi and PRIME32_2 to
-         * the high 64 bits of m128.
-         *
-         * The best approach to this operation is different on 32-bit and 64-bit.
-         */
-        if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
-            /*
-             * 32-bit optimized version, which is more readable.
-             *
-             * On 32-bit, it removes an ADC and delays a dependency between the two
-             * halves of m128.high64, but it generates an extra mask on 64-bit.
-             */
-            m128.high64 += (input_hi & 0xFFFFFFFF00000000) + XXH_mult32to64((xxh_u32)input_hi, PRIME32_2);
-        } else {
-            /*
-             * 64-bit optimized (albeit more confusing) version.
-             *
-             * Uses some properties of addition and multiplication to remove the mask:
-             *
-             * Let:
-             *    a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
-             *    b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
-             *    c = PRIME32_2
-             *
-             *    a + (b * c)
-             * Inverse Property: x + y - x == y
-             *    a + (b * (1 + c - 1))
-             * Distributive Property: x * (y + z) == (x * y) + (x * z)
-             *    a + (b * 1) + (b * (c - 1))
-             * Identity Property: x * 1 == x
-             *    a + b + (b * (c - 1))
-             *
-             * Substitute a, b, and c:
-             *    input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (PRIME32_2 - 1))
-             *
-             * Since input_hi.hi + input_hi.lo == input_hi, we get this:
-             *    input_hi + ((xxh_u64)input_hi.lo * (PRIME32_2 - 1))
-             */
-            m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, PRIME32_2 - 1);
-        }
-        /* m128 ^= XXH_swap64(m128 >> 64); */
-        m128.low64  ^= XXH_swap64(m128.high64);
-
-        {   /* 128x64 multiply: h128 = m128 * PRIME64_2; */
-            XXH128_hash_t h128 = XXH_mult64to128(m128.low64, PRIME64_2);
-            h128.high64 += m128.high64 * PRIME64_2;
-
-            h128.low64   = XXH3_avalanche(h128.low64);
-            h128.high64  = XXH3_avalanche(h128.high64);
-            return h128;
-    }   }
-}
-
-/*
- * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
- */
-XXH_FORCE_INLINE 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);
-    {   if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
-        if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
-        if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
-        {   XXH128_hash_t h128;
-            xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
-            xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
-            h128.low64 = XXH3_avalanche((PRIME64_1 + seed) ^ bitflipl);
-            h128.high64 = XXH3_avalanche((PRIME64_2 - seed) ^ bitfliph);
-            return h128;
-    }   }
-}
-
-/*
- * A bit slower than XXH3_mix16B, but handles multiply by zero better.
- */
-XXH_FORCE_INLINE XXH128_hash_t
-XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
-              const xxh_u8* secret, XXH64_hash_t seed)
-{
-    acc.low64  += XXH3_mix16B (input_1, secret+0, seed);
-    acc.low64  ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
-    acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
-    acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
-    return acc;
-}
-
-
-XXH_FORCE_INLINE 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)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
-    XXH_ASSERT(16 < len && len <= 128);
-
-    {   XXH128_hash_t acc;
-        acc.low64 = len * PRIME64_1;
-        acc.high64 = 0;
-        if (len > 32) {
-            if (len > 64) {
-                if (len > 96) {
-                    acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
-                }
-                acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
-            }
-            acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
-        }
-        acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
-        {   XXH128_hash_t h128;
-            h128.low64  = acc.low64 + acc.high64;
-            h128.high64 = (acc.low64    * PRIME64_1)
-                        + (acc.high64   * PRIME64_4)
-                        + ((len - seed) * PRIME64_2);
-            h128.low64  = XXH3_avalanche(h128.low64);
-            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
-            return h128;
-        }
-    }
-}
-
-XXH_NO_INLINE 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)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
-    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
-
-    {   XXH128_hash_t acc;
-        int const nbRounds = (int)len / 32;
-        int i;
-        acc.low64 = len * PRIME64_1;
-        acc.high64 = 0;
-        for (i=0; i<4; i++) {
-            acc = XXH128_mix32B(acc,
-                                input  + (32 * i),
-                                input  + (32 * i) + 16,
-                                secret + (32 * i),
-                                seed);
-        }
-        acc.low64 = XXH3_avalanche(acc.low64);
-        acc.high64 = XXH3_avalanche(acc.high64);
-        XXH_ASSERT(nbRounds >= 4);
-        for (i=4 ; i < nbRounds; i++) {
-            acc = XXH128_mix32B(acc,
-                                input + (32 * i),
-                                input + (32 * i) + 16,
-                                secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)),
-                                seed);
-        }
-        /* last bytes */
-        acc = XXH128_mix32B(acc,
-                            input + len - 16,
-                            input + len - 32,
-                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
-                            0ULL - seed);
-
-        {   XXH128_hash_t h128;
-            h128.low64  = acc.low64 + acc.high64;
-            h128.high64 = (acc.low64    * PRIME64_1)
-                        + (acc.high64   * PRIME64_4)
-                        + ((len - seed) * PRIME64_2);
-            h128.low64  = XXH3_avalanche(h128.low64);
-            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
-            return h128;
-        }
-    }
-}
-
-XXH_FORCE_INLINE XXH128_hash_t
-XXH3_hashLong_128b_internal(const xxh_u8* XXH_RESTRICT input, size_t len,
-                            const xxh_u8* XXH_RESTRICT secret, size_t secretSize)
-{
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXH3_INIT_ACC;
-
-    XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXH3_acc_128bits);
-
-    /* converge into final hash */
-    XXH_STATIC_ASSERT(sizeof(acc) == 64);
-    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
-    {   XXH128_hash_t h128;
-        h128.low64  = XXH3_mergeAccs(acc,
-                                     secret + XXH_SECRET_MERGEACCS_START,
-                                     (xxh_u64)len * PRIME64_1);
-        h128.high64 = XXH3_mergeAccs(acc,
-                                     secret + secretSize
-                                            - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
-                                     ~((xxh_u64)len * PRIME64_2));
-        return h128;
-    }
-}
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH128_hash_t
-XXH3_hashLong_128b_defaultSecret(const xxh_u8* input, size_t len)
-{
-    return XXH3_hashLong_128b_internal(input, len, kSecret, sizeof(kSecret));
-}
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH128_hash_t
-XXH3_hashLong_128b_withSecret(const xxh_u8* input, size_t len,
-                              const xxh_u8* secret, size_t secretSize)
-{
-    return XXH3_hashLong_128b_internal(input, len, secret, secretSize);
-}
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH128_hash_t
-XXH3_hashLong_128b_withSeed(const xxh_u8* input, size_t len, XXH64_hash_t seed)
-{
-    XXH_ALIGN(8) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
-    if (seed == 0) return XXH3_hashLong_128b_defaultSecret(input, len);
-    XXH3_initCustomSecret(secret, seed);
-    return XXH3_hashLong_128b_internal(input, len, secret, sizeof(secret));
-}
-
-
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len)
-{
-    if (len <= 16)
-        return XXH3_len_0to16_128b((const xxh_u8*)input, len, kSecret, 0);
-    if (len <= 128)
-        return XXH3_len_17to128_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
-    if (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_len_129to240_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
-    return XXH3_hashLong_128b_defaultSecret((const xxh_u8*)input, len);
-}
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
-    /*
-     * If an action is to be taken if `secret` conditions are not respected,
-     * it should be done here.
-     * For now, it's a contract pre-condition.
-     * Adding a check and a branch here would cost performance at every hash.
-     */
-    if (len <= 16)
-        return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0);
-    if (len <= 128)
-        return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
-    if (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
-    return XXH3_hashLong_128b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize);
-}
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
-{
-    if (len <= 16)
-        return XXH3_len_0to16_128b((const xxh_u8*)input, len, kSecret, seed);
-    if (len <= 128)
-         return XXH3_len_17to128_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
-    if (len <= XXH3_MIDSIZE_MAX)
-         return XXH3_len_129to240_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
-    return XXH3_hashLong_128b_withSeed((const xxh_u8*)input, len, seed);
-}
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH128(const void* input, size_t len, XXH64_hash_t seed)
-{
-    return XXH3_128bits_withSeed(input, len, seed);
-}
-
-
-/* ===   XXH3 128-bit streaming   === */
-
-/*
- * All the functions are actually the same as for 64-bit streaming variant.
- * The only difference is the finalizatiom routine.
- */
-
-static void
-XXH3_128bits_reset_internal(XXH3_state_t* statePtr,
-                            XXH64_hash_t seed,
-                            const xxh_u8* secret, size_t secretSize)
-{
-    XXH3_64bits_reset_internal(statePtr, seed, secret, secretSize);
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset(XXH3_state_t* statePtr)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_128bits_reset_internal(statePtr, 0, kSecret, XXH_SECRET_DEFAULT_SIZE);
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_128bits_reset_internal(statePtr, 0, (const xxh_u8*)secret, secretSize);
-    if (secret == NULL) return XXH_ERROR;
-    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_128bits_reset_internal(statePtr, seed, kSecret, XXH_SECRET_DEFAULT_SIZE);
-    XXH3_initCustomSecret(statePtr->customSecret, seed);
-    statePtr->secret = statePtr->customSecret;
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_update(XXH3_state_t* state, const void* input, size_t len)
-{
-    return XXH3_update(state, (const xxh_u8*)input, len, XXH3_acc_128bits);
-}
-
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state)
-{
-    if (state->totalLen > XXH3_MIDSIZE_MAX) {
-        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[ACC_NB];
-        XXH3_digest_long(acc, state, XXH3_acc_128bits);
-        XXH_ASSERT(state->secretLimit + STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
-        {   XXH128_hash_t h128;
-            h128.low64  = XXH3_mergeAccs(acc,
-                                         state->secret + XXH_SECRET_MERGEACCS_START,
-                                         (xxh_u64)state->totalLen * PRIME64_1);
-            h128.high64 = XXH3_mergeAccs(acc,
-                                         state->secret + state->secretLimit + STRIPE_LEN
-                                                       - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
-                                         ~((xxh_u64)state->totalLen * PRIME64_2));
-            return h128;
-        }
-    }
-    /* len <= XXH3_MIDSIZE_MAX : short code */
-    if (state->seed)
-        return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
-    return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
-                                   state->secret, state->secretLimit + STRIPE_LEN);
-}
-
-/* 128-bit utility functions */
-
-#include <string.h>   /* memcmp, memcpy */
-
-/* return : 1 is equal, 0 if different */
-XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
-{
-    /* note : XXH128_hash_t is compact, it has no padding byte */
-    return !(memcmp(&h1, &h2, sizeof(h1)));
-}
-
-/* This prototype is compatible with stdlib's qsort().
- * return : >0 if *h128_1  > *h128_2
- *          <0 if *h128_1  < *h128_2
- *          =0 if *h128_1 == *h128_2  */
-XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, 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;
-    int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
-    /* note : bets that, in most cases, hash values are different */
-    if (hcmp) return hcmp;
-    return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
-}
-
-
-/*======   Canonical representation   ======*/
-XXH_PUBLIC_API void
-XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash)
-{
-    XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
-    if (XXH_CPU_LITTLE_ENDIAN) {
-        hash.high64 = XXH_swap64(hash.high64);
-        hash.low64  = XXH_swap64(hash.low64);
-    }
-    memcpy(dst, &hash.high64, sizeof(hash.high64));
-    memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
-}
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH128_hashFromCanonical(const XXH128_canonical_t* src)
-{
-    XXH128_hash_t h;
-    h.high64 = XXH_readBE64(src);
-    h.low64  = XXH_readBE64(src->digest + 8);
-    return h;
-}
-
-/* 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 */
-#  pragma GCC pop_options
-#endif
-
-#endif  /* XXH3_H_1397135465 */
+#define XXH_INLINE_ALL
+#include "xxhash.h"
diff --git a/src/xxhash/xxhash.c b/src/xxhash/xxhash.c
index 0fae88c5d6..e60cc37f13 100644
--- a/src/xxhash/xxhash.c
+++ b/src/xxhash/xxhash.c
@@ -1,6 +1,6 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
- * Copyright (C) 2012-2020 Yann Collet
+ * Copyright (C) 2012-2023 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
@@ -32,12 +32,11 @@
  *   - xxHash source repository: https://github.com/Cyan4973/xxHash
  */
 
-
 /*
  * xxhash.c instantiates functions defined in xxhash.h
  */
 
-#define XXH_STATIC_LINKING_ONLY   /* access advanced declarations */
-#define XXH_IMPLEMENTATION   /* access definitions */
+#define XXH_STATIC_LINKING_ONLY /* access advanced declarations */
+#define XXH_IMPLEMENTATION      /* access definitions */
 
 #include "xxhash.h"
diff --git a/src/xxhash/xxhash.h b/src/xxhash/xxhash.h
index 67a5887f39..78fc2e8dbf 100644
--- a/src/xxhash/xxhash.h
+++ b/src/xxhash/xxhash.h
@@ -1,7 +1,7 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
  * Header File
- * Copyright (C) 2012-2020 Yann Collet
+ * Copyright (C) 2012-2023 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
@@ -33,43 +33,213 @@
  *   - xxHash source repository: https://github.com/Cyan4973/xxHash
  */
 
-/* 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
-MumurHash 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
-*/
+/*!
+ * @mainpage xxHash
+ *
+ * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
+ * limits.
+ *
+ * It is proposed in four flavors, in three families:
+ * 1. @ref XXH32_family
+ *   - Classic 32-bit hash function. Simple, compact, and runs on almost all
+ *     32-bit and 64-bit systems.
+ * 2. @ref XXH64_family
+ *   - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
+ *     64-bit systems (but _not_ 32-bit systems).
+ * 3. @ref XXH3_family
+ *   - Modern 64-bit and 128-bit hash function family which features improved
+ *     strength and performance across the board, especially on smaller data.
+ *     It benefits greatly from SIMD and 64-bit without requiring it.
+ *
+ * Benchmarks
+ * ---
+ * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
+ * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
+ *
+ * | Hash Name            | ISA ext | Width | Large Data Speed | Small Data Velocity |
+ * | -------------------- | ------- | ----: | ---------------: | ------------------: |
+ * | XXH3_64bits()        | @b AVX2 |    64 |        59.4 GB/s |               133.1 |
+ * | MeowHash             | AES-NI  |   128 |        58.2 GB/s |                52.5 |
+ * | XXH3_128bits()       | @b AVX2 |   128 |        57.9 GB/s |               118.1 |
+ * | CLHash               | PCLMUL  |    64 |        37.1 GB/s |                58.1 |
+ * | XXH3_64bits()        | @b SSE2 |    64 |        31.5 GB/s |               133.1 |
+ * | XXH3_128bits()       | @b SSE2 |   128 |        29.6 GB/s |               118.1 |
+ * | RAM sequential read  |         |   N/A |        28.0 GB/s |                 N/A |
+ * | ahash                | AES-NI  |    64 |        22.5 GB/s |               107.2 |
+ * | City64               |         |    64 |        22.0 GB/s |                76.6 |
+ * | T1ha2                |         |    64 |        22.0 GB/s |                99.0 |
+ * | City128              |         |   128 |        21.7 GB/s |                57.7 |
+ * | FarmHash             | AES-NI  |    64 |        21.3 GB/s |                71.9 |
+ * | XXH64()              |         |    64 |        19.4 GB/s |                71.0 |
+ * | SpookyHash           |         |    64 |        19.3 GB/s |                53.2 |
+ * | Mum                  |         |    64 |        18.0 GB/s |                67.0 |
+ * | CRC32C               | SSE4.2  |    32 |        13.0 GB/s |                57.9 |
+ * | XXH32()              |         |    32 |         9.7 GB/s |                71.9 |
+ * | City32               |         |    32 |         9.1 GB/s |                66.0 |
+ * | Blake3*              | @b AVX2 |   256 |         4.4 GB/s |                 8.1 |
+ * | Murmur3              |         |    32 |         3.9 GB/s |                56.1 |
+ * | SipHash*             |         |    64 |         3.0 GB/s |                43.2 |
+ * | Blake3*              | @b SSE2 |   256 |         2.4 GB/s |                 8.1 |
+ * | HighwayHash          |         |    64 |         1.4 GB/s |                 6.0 |
+ * | FNV64                |         |    64 |         1.2 GB/s |                62.7 |
+ * | Blake2*              |         |   256 |         1.1 GB/s |                 5.1 |
+ * | SHA1*                |         |   160 |         0.8 GB/s |                 5.6 |
+ * | MD5*                 |         |   128 |         0.6 GB/s |                 7.8 |
+ * @note
+ *   - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
+ *     even though it is mandatory on x64.
+ *   - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
+ *     by modern standards.
+ *   - Small data velocity is a rough average of algorithm's efficiency for small
+ *     data. For more accurate information, see the wiki.
+ *   - More benchmarks and strength tests are found on the wiki:
+ *         https://github.com/Cyan4973/xxHash/wiki
+ *
+ * Usage
+ * ------
+ * All xxHash variants use a similar API. Changing the algorithm is a trivial
+ * substitution.
+ *
+ * @pre
+ *    For functions which take an input and length parameter, the following
+ *    requirements are assumed:
+ *    - The range from [`input`, `input + length`) is valid, readable memory.
+ *      - The only exception is if the `length` is `0`, `input` may be `NULL`.
+ *    - For C++, the objects must have the *TriviallyCopyable* property, as the
+ *      functions access bytes directly as if it was an array of `unsigned char`.
+ *
+ * @anchor single_shot_example
+ * **Single Shot**
+ *
+ * These functions are stateless functions which hash a contiguous block of memory,
+ * immediately returning the result. They are the easiest and usually the fastest
+ * option.
+ *
+ * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
+ *
+ * @code{.c}
+ *   #include <string.h>
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which hashes a null terminated string with XXH32().
+ *   XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
+ *   {
+ *       // NULL pointers are only valid if the length is zero
+ *       size_t length = (string == NULL) ? 0 : strlen(string);
+ *       return XXH32(string, length, seed);
+ *   }
+ * @endcode
+ *
+ *
+ * @anchor streaming_example
+ * **Streaming**
+ *
+ * These groups of functions allow incremental hashing of unknown size, even
+ * more than what would fit in a size_t.
+ *
+ * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
+ *
+ * @code{.c}
+ *   #include <stdio.h>
+ *   #include <assert.h>
+ *   #include "xxhash.h"
+ *   // Example for a function which hashes a FILE incrementally with XXH3_64bits().
+ *   XXH64_hash_t hashFile(FILE* f)
+ *   {
+ *       // Allocate a state struct. Do not just use malloc() or new.
+ *       XXH3_state_t* state = XXH3_createState();
+ *       assert(state != NULL && "Out of memory!");
+ *       // Reset the state to start a new hashing session.
+ *       XXH3_64bits_reset(state);
+ *       char buffer[4096];
+ *       size_t count;
+ *       // Read the file in chunks
+ *       while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
+ *           // Run update() as many times as necessary to process the data
+ *           XXH3_64bits_update(state, buffer, count);
+ *       }
+ *       // Retrieve the finalized hash. This will not change the state.
+ *       XXH64_hash_t result = XXH3_64bits_digest(state);
+ *       // Free the state. Do not use free().
+ *       XXH3_freeState(state);
+ *       return result;
+ *   }
+ * @endcode
+ *
+ * Streaming functions generate the xxHash value from an incremental input.
+ * This method is slower than single-call functions, due to state management.
+ * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+ *
+ * An XXH state must first be allocated using `XXH*_createState()`.
+ *
+ * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ *
+ * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ *
+ * The function returns an error code, with 0 meaning OK, and any other value
+ * meaning there is an error.
+ *
+ * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
+ * This function returns the nn-bits hash as an int or long long.
+ *
+ * It's still possible to continue inserting input into the hash state after a
+ * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ *
+ * When done, release the state using `XXH*_freeState()`.
+ *
+ *
+ * @anchor canonical_representation_example
+ * **Canonical Representation**
+ *
+ * The default return values from XXH functions are unsigned 32, 64 and 128 bit
+ * integers.
+ * This the simplest and fastest format for further post-processing.
+ *
+ * However, this leaves open the question of what is the order on the byte level,
+ * since little and big endian conventions will store the same number differently.
+ *
+ * The canonical representation settles this issue by mandating big-endian
+ * convention, the same convention as human-readable numbers (large digits first).
+ *
+ * When writing hash values to storage, sending them over a network, or printing
+ * them, it's highly recommended to use the canonical representation to ensure
+ * portability across a wider range of systems, present and future.
+ *
+ * The following functions allow transformation of hash values to and from
+ * canonical format.
+ *
+ * XXH32_canonicalFromHash(), XXH32_hashFromCanonical(),
+ * XXH64_canonicalFromHash(), XXH64_hashFromCanonical(),
+ * XXH128_canonicalFromHash(), XXH128_hashFromCanonical(),
+ *
+ * @code{.c}
+ *   #include <stdio.h>
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which prints XXH32_hash_t in human readable format
+ *   void printXxh32(XXH32_hash_t hash)
+ *   {
+ *       XXH32_canonical_t cano;
+ *       XXH32_canonicalFromHash(&cano, hash);
+ *       size_t i;
+ *       for(i = 0; i < sizeof(cano.digest); ++i) {
+ *           printf("%02x", cano.digest[i]);
+ *       }
+ *       printf("\n");
+ *   }
+ *
+ *   // Example for a function which converts XXH32_canonical_t to XXH32_hash_t
+ *   XXH32_hash_t convertCanonicalToXxh32(XXH32_canonical_t cano)
+ *   {
+ *       XXH32_hash_t hash = XXH32_hashFromCanonical(&cano);
+ *       return hash;
+ *   }
+ * @endcode
+ *
+ *
+ * @file xxhash.h
+ * xxHash prototypes and implementation
+ */
 
 #if defined (__cplusplus)
 extern "C" {
@@ -79,21 +249,80 @@ extern "C" {
  *  INLINE mode
  ******************************/
 /*!
- * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Gives access to internal state declaration, required for static allocation.
+ *
+ * Incompatible with dynamic linking, due to risks of ABI changes.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_STATIC_LINKING_ONLY
+/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */
+
+/*!
+ * @brief Gives access to internal definitions.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #define XXH_IMPLEMENTATION
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_IMPLEMENTATION
+/* Do not undef XXH_IMPLEMENTATION for Doxygen */
+
+/*!
+ * @brief Exposes the implementation and marks all functions as `inline`.
+ *
  * Use these build macros to inline xxhash into the target unit.
  * Inlining improves performance on small inputs, especially when the length is
  * expressed as a compile-time constant:
  *
- *      https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
+ *  https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
  *
  * It also keeps xxHash symbols private to the unit, so they are not exported.
  *
  * Usage:
+ * @code{.c}
  *     #define XXH_INLINE_ALL
  *     #include "xxhash.h"
- *
+ * @endcode
  * Do not compile and link xxhash.o as a separate object, as it is not useful.
  */
+#  define XXH_INLINE_ALL
+#  undef XXH_INLINE_ALL
+/*!
+ * @brief Exposes the implementation without marking functions as inline.
+ */
+#  define XXH_PRIVATE_API
+#  undef XXH_PRIVATE_API
+/*!
+ * @brief Emulate a namespace by transparently prefixing all symbols.
+ *
+ * If you want to include _and expose_ xxHash functions from within your own
+ * library, but also want to avoid symbol collisions with other libraries which
+ * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
+ * (therefore, avoid empty or numeric values).
+ *
+ * Note that no change is required within the calling program as long as it
+ * includes `xxhash.h`: Regular symbol names will be automatically translated
+ * by this header.
+ */
+#  define XXH_NAMESPACE /* YOUR NAME HERE */
+#  undef XXH_NAMESPACE
+#endif
+
 #if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
     && !defined(XXH_INLINE_ALL_31684351384)
    /* this section should be traversed only once */
@@ -104,7 +333,7 @@ extern "C" {
    /* make all functions private */
 #  undef XXH_PUBLIC_API
 #  if defined(__GNUC__)
-#    define XXH_PUBLIC_API static __inline __attribute__((unused))
+#    define XXH_PUBLIC_API static __inline __attribute__((__unused__))
 #  elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
 #    define XXH_PUBLIC_API static inline
 #  elif defined(_MSC_VER)
@@ -116,29 +345,80 @@ extern "C" {
 
    /*
     * This part deals with the special case where a unit wants to inline xxHash,
-    * but "xxhash.h" has previously been included without XXH_INLINE_ALL, such
-    * as part of some previously included *.h header file.
+    * but "xxhash.h" has previously been included without XXH_INLINE_ALL,
+    * such as part of some previously included *.h header file.
     * Without further action, the new include would just be ignored,
     * and functions would effectively _not_ be inlined (silent failure).
     * The following macros solve this situation by prefixing all inlined names,
     * avoiding naming collision with previous inclusions.
     */
-#  ifdef XXH_NAMESPACE
-#    error "XXH_INLINE_ALL with XXH_NAMESPACE is not supported"
-     /*
-      * Note: Alternative: #undef all symbols (it's a pretty large list).
-      * Without #error: it compiles, but functions are actually not inlined.
-      */
-#  endif
+   /* Before that, we unconditionally #undef all symbols,
+    * in case they were already defined with XXH_NAMESPACE.
+    * They will then be redefined for XXH_INLINE_ALL
+    */
+#  undef XXH_versionNumber
+    /* XXH32 */
+#  undef XXH32
+#  undef XXH32_createState
+#  undef XXH32_freeState
+#  undef XXH32_reset
+#  undef XXH32_update
+#  undef XXH32_digest
+#  undef XXH32_copyState
+#  undef XXH32_canonicalFromHash
+#  undef XXH32_hashFromCanonical
+    /* XXH64 */
+#  undef XXH64
+#  undef XXH64_createState
+#  undef XXH64_freeState
+#  undef XXH64_reset
+#  undef XXH64_update
+#  undef XXH64_digest
+#  undef XXH64_copyState
+#  undef XXH64_canonicalFromHash
+#  undef XXH64_hashFromCanonical
+    /* XXH3_64bits */
+#  undef XXH3_64bits
+#  undef XXH3_64bits_withSecret
+#  undef XXH3_64bits_withSeed
+#  undef XXH3_64bits_withSecretandSeed
+#  undef XXH3_createState
+#  undef XXH3_freeState
+#  undef XXH3_copyState
+#  undef XXH3_64bits_reset
+#  undef XXH3_64bits_reset_withSeed
+#  undef XXH3_64bits_reset_withSecret
+#  undef XXH3_64bits_update
+#  undef XXH3_64bits_digest
+#  undef XXH3_generateSecret
+    /* XXH3_128bits */
+#  undef XXH128
+#  undef XXH3_128bits
+#  undef XXH3_128bits_withSeed
+#  undef XXH3_128bits_withSecret
+#  undef XXH3_128bits_reset
+#  undef XXH3_128bits_reset_withSeed
+#  undef XXH3_128bits_reset_withSecret
+#  undef XXH3_128bits_reset_withSecretandSeed
+#  undef XXH3_128bits_update
+#  undef XXH3_128bits_digest
+#  undef XXH128_isEqual
+#  undef XXH128_cmp
+#  undef XXH128_canonicalFromHash
+#  undef XXH128_hashFromCanonical
+    /* Finally, free the namespace itself */
+#  undef XXH_NAMESPACE
+
+    /* employ the namespace for XXH_INLINE_ALL */
 #  define XXH_NAMESPACE XXH_INLINE_
    /*
-    * Some identifiers (enums, type names) are not symbols, but they must
-    * still be renamed to avoid redeclaration.
+    * Some identifiers (enums, type names) are not symbols,
+    * but they must nonetheless be renamed to avoid redeclaration.
     * Alternative solution: do not redeclare them.
-    * However, this requires some #ifdefs, and is a more dispersed action.
-    * Meanwhile, renaming can be achieved in a single block
+    * However, this requires some #ifdefs, and has a more dispersed impact.
+    * Meanwhile, renaming can be achieved in a single place.
     */
-#  define XXH_IPREF(Id)   XXH_INLINE_ ## Id
+#  define XXH_IPREF(Id)   XXH_NAMESPACE ## Id
 #  define XXH_OK XXH_IPREF(XXH_OK)
 #  define XXH_ERROR XXH_IPREF(XXH_ERROR)
 #  define XXH_errorcode XXH_IPREF(XXH_errorcode)
@@ -157,17 +437,15 @@ 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
 
-/* 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))
+#  if defined(_WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
 #    ifdef XXH_EXPORT
 #      define XXH_PUBLIC_API __declspec(dllexport)
 #    elif XXH_IMPORT
@@ -178,23 +456,11 @@ extern "C" {
 #  endif
 #endif
 
-/*!
- * XXH_NAMESPACE, aka Namespace Emulation:
- *
- * 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.
- */
 #ifdef XXH_NAMESPACE
 #  define XXH_CAT(A,B) A##B
 #  define XXH_NAME2(A,B) XXH_CAT(A,B)
 #  define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
+/* XXH32 */
 #  define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
 #  define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
 #  define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
@@ -204,6 +470,7 @@ extern "C" {
 #  define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
 #  define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
 #  define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
+/* XXH64 */
 #  define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
 #  define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
 #  define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
@@ -213,1753 +480,6759 @@ extern "C" {
 #  define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
 #  define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
 #  define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
+/* XXH3_64bits */
+#  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
+#  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
+#  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+#  define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed)
+#  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
+#  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
+#  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
+#  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
+#  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
+#  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
+#  define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed)
+#  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
+#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
+#  define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
+#  define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed)
+/* XXH3_128bits */
+#  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
+#  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
+#  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
+#  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+#  define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed)
+#  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
+#  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
+#  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
+#  define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed)
+#  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
+#  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
+#  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
+#  define XXH128_cmp     XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
+#  define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
+#  define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
+#endif
+
+
+/* *************************************
+*  Compiler specifics
+***************************************/
+
+/* specific declaration modes for Windows */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(_WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
 #endif
 
+#if defined (__GNUC__)
+# define XXH_CONSTF  __attribute__((__const__))
+# define XXH_PUREF   __attribute__((__pure__))
+# define XXH_MALLOCF __attribute__((__malloc__))
+#else
+# define XXH_CONSTF  /* disable */
+# define XXH_PUREF
+# define XXH_MALLOCF
+#endif
 
 /* *************************************
 *  Version
 ***************************************/
 #define XXH_VERSION_MAJOR    0
-#define XXH_VERSION_MINOR    7
-#define XXH_VERSION_RELEASE  4
+#define XXH_VERSION_MINOR    8
+#define XXH_VERSION_RELEASE  3
+/*! @brief Version number, encoded as two digits each */
 #define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
-XXH_PUBLIC_API unsigned XXH_versionNumber (void);
+
+/*!
+ * @brief Obtains the xxHash version.
+ *
+ * This is mostly useful when xxHash is compiled as a shared library,
+ * since the returned value comes from the library, as opposed to header file.
+ *
+ * @return @ref XXH_VERSION_NUMBER of the invoked library.
+ */
+XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
 
 
 /* ****************************
-*  Definitions
+*  Common basic types
 ******************************/
 #include <stddef.h>   /* size_t */
-typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
+/*!
+ * @brief Exit code for the streaming API.
+ */
+typedef enum {
+    XXH_OK = 0, /*!< OK */
+    XXH_ERROR   /*!< Error */
+} XXH_errorcode;
 
 
 /*-**********************************************************************
 *  32-bit hash
 ************************************************************************/
-#if !defined (__VMS) \
+#if defined(XXH_DOXYGEN) /* Don't show <stdint.h> include */
+/*!
+ * @brief An unsigned 32-bit integer.
+ *
+ * Not necessarily defined to `uint32_t` but functionally equivalent.
+ */
+typedef uint32_t XXH32_hash_t;
+
+#elif !defined (__VMS) \
   && (defined (__cplusplus) \
   || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-#   include <stdint.h>
+#   ifdef _AIX
+#     include <inttypes.h>
+#   else
+#     include <stdint.h>
+#   endif
     typedef uint32_t XXH32_hash_t;
+
 #else
 #   include <limits.h>
 #   if UINT_MAX == 0xFFFFFFFFUL
       typedef unsigned int XXH32_hash_t;
+#   elif ULONG_MAX == 0xFFFFFFFFUL
+      typedef unsigned long XXH32_hash_t;
 #   else
-#     if ULONG_MAX == 0xFFFFFFFFUL
-        typedef unsigned long XXH32_hash_t;
-#     else
-#       error "unsupported platform: need a 32-bit type"
-#     endif
+#     error "unsupported platform: need a 32-bit type"
 #   endif
 #endif
 
 /*!
- * XXH32():
- *  Calculate the 32-bit hash of sequence "length" bytes stored at memory address "input".
- *  The memory between input & input+length must be valid (allocated and read-accessible).
- *  "seed" can be used to alter the result predictably.
- *  Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s
+ * @}
+ *
+ * @defgroup XXH32_family XXH32 family
+ * @ingroup public
+ * Contains functions used in the classic 32-bit xxHash algorithm.
  *
- * Note: XXH3 provides competitive speed for both 32-bit and 64-bit systems,
- * and offers true 64/128 bit hash results. It provides a superior level of
- * dispersion, and greatly reduces the risks of collisions.
+ * @note
+ *   XXH32 is useful for older platforms, with no or poor 64-bit performance.
+ *   Note that the @ref XXH3_family provides competitive speed for both 32-bit
+ *   and 64-bit systems, and offers true 64/128 bit hash results.
+ *
+ * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
+ * @see @ref XXH32_impl for implementation details
+ * @{
  */
-XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
 
-/*******   Streaming   *******/
+/*!
+ * @brief Calculates the 32-bit hash of @p input using xxHash32.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 32-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 32-bit xxHash32 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
 
-/*
- * Streaming functions generate the xxHash value from an incrememtal input.
- * This method is slower than single-call functions, due to state management.
- * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+#ifndef XXH_NO_STREAM
+/*!
+ * @typedef struct XXH32_state_s XXH32_state_t
+ * @brief The opaque state struct for the XXH32 streaming API.
  *
- * An XXH state must first be allocated using `XXH*_createState()`.
+ * @see XXH32_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
+ */
+typedef struct XXH32_state_s XXH32_state_t;
+
+/*!
+ * @brief Allocates an @ref XXH32_state_t.
  *
- * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ * @return An allocated pointer of @ref XXH32_state_t on success.
+ * @return `NULL` on failure.
  *
- * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ * @note Must be freed with XXH32_freeState().
  *
- * The function returns an error code, with 0 meaning OK, and any other value
- * meaning there is an error.
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
+/*!
+ * @brief Frees an @ref XXH32_state_t.
  *
- * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
- * This function returns the nn-bits hash as an int or long long.
+ * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
  *
- * It's still possible to continue inserting input into the hash state after a
- * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH32_createState().
+ *
+ * @see @ref streaming_example "Streaming Example"
  *
- * When done, release the state using `XXH*_freeState()`.
  */
-
-typedef struct XXH32_state_s XXH32_state_t;   /* incomplete type */
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
 XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
+/*!
+ * @brief Copies one @ref XXH32_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
 XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
 
+/*!
+ * @brief Resets an @ref XXH32_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 32-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH32_update().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
 XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH32_hash_t  XXH32_digest (const XXH32_state_t* statePtr);
 
-/*******   Canonical representation   *******/
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH32_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
 
-/*
- * The default return values from XXH functions are unsigned 32 and 64 bit
- * integers.
- * This the simplest and fastest format for further post-processing.
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH32_state_t.
  *
- * However, this leaves open the question of what is the order on the byte level,
- * since little and big endian conventions will store the same number differently.
+ * @param statePtr The state struct to calculate the hash from.
  *
- * The canonical representation settles this issue by mandating big-endian
- * convention, the same convention as human-readable numbers (large digits first).
+ * @pre
+ *  @p statePtr must not be `NULL`.
  *
- * When writing hash values to storage, sending them over a network, or printing
- * them, it's highly recommended to use the canonical representation to ensure
- * portability across a wider range of systems, present and future.
+ * @return The calculated 32-bit xxHash32 value from that state.
  *
- * The following functions allow transformation of hash values to and from
- * canonical format.
+ * @note
+ *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
  */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
-typedef struct { unsigned char digest[4]; } XXH32_canonical_t;
-XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+/*******   Canonical representation   *******/
 
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
+ */
+typedef struct {
+    unsigned char digest[4]; /*!< Hash bytes, big endian */
+} XXH32_canonical_t;
 
-#ifndef XXH_NO_LONG_LONG
-/*-**********************************************************************
-*  64-bit hash
-************************************************************************/
-#if !defined (__VMS) \
-  && (defined (__cplusplus) \
-  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-#   include <stdint.h>
-    typedef uint64_t XXH64_hash_t;
-#else
-    /* the following type must have a width of 64-bit */
-    typedef unsigned long long XXH64_hash_t;
-#endif
+/*!
+ * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
+ *
+ * @param dst  The @ref XXH32_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH32_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
 
 /*!
- * XXH64():
- * Returns the 64-bit hash of sequence of length @length stored at memory
- * address @input.
- * @seed can be used to alter the result predictably.
+ * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t.
+ *
+ * @param src The @ref XXH32_canonical_t to convert.
  *
- * This function usually runs faster on 64-bit systems, but slower on 32-bit
- * systems (see benchmark).
+ * @pre
+ *   @p src must not be `NULL`.
  *
- * Note: XXH3 provides competitive speed for both 32-bit and 64-bit systems,
- * and offers true 64/128 bit hash results. It provides a superior level of
- * dispersion, and greatly reduces the risks of collisions.
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
  */
-XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t length, XXH64_hash_t seed);
-
-/*******   Streaming   *******/
-typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
-XXH_PUBLIC_API 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);
-
-/*******   Canonical representation   *******/
-typedef struct { unsigned char digest[8]; } 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 XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
 
 
-#endif  /* XXH_NO_LONG_LONG */
+/*! @cond Doxygen ignores this part */
+#ifdef __has_attribute
+# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
+#else
+# define XXH_HAS_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
 
-#endif /* XXHASH_H_5627135585666179 */
+/*! @cond Doxygen ignores this part */
+/*
+ * C23 __STDC_VERSION__ number hasn't been specified yet. For now
+ * leave as `201711L` (C17 + 1).
+ * TODO: Update to correct value when its been specified.
+ */
+#define XXH_C23_VN 201711L
+/*! @endcond */
 
+/*! @cond Doxygen ignores this part */
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
+# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define XXH_HAS_C_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
 
+/*! @cond Doxygen ignores this part */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define XXH_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
 
-#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
-#define XXHASH_H_STATIC_13879238742
-/* ****************************************************************************
- * This section contains declarations which are not guaranteed to remain stable.
- * They may change in future versions, becoming incompatible with a different
- * version of the library.
- * These declarations should only be used with static linking.
- * Never use them in association with dynamic linking!
- ***************************************************************************** */
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
+ * introduced in CPP17 and C23.
+ * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * C23   : https://en.cppreference.com/w/c/language/attributes/fallthrough
+ */
+#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough)
+# define XXH_FALLTHROUGH [[fallthrough]]
+#elif XXH_HAS_ATTRIBUTE(__fallthrough__)
+# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__))
+#else
+# define XXH_FALLTHROUGH /* fallthrough */
+#endif
+/*! @endcond */
 
+/*! @cond Doxygen ignores this part */
 /*
- * These definitions are only present to allow static allocation of an XXH
- * state, for example, on the stack or in a struct.
- * Never **ever** access members directly.
+ * Define XXH_NOESCAPE for annotated pointers in public API.
+ * https://clang.llvm.org/docs/AttributeReference.html#noescape
+ * As of writing this, only supported by clang.
  */
+#if XXH_HAS_ATTRIBUTE(noescape)
+# define XXH_NOESCAPE __attribute__((__noescape__))
+#else
+# define XXH_NOESCAPE
+#endif
+/*! @endcond */
 
-struct XXH32_state_s {
-   XXH32_hash_t total_len_32;
-   XXH32_hash_t large_len;
-   XXH32_hash_t v1;
-   XXH32_hash_t v2;
-   XXH32_hash_t v3;
-   XXH32_hash_t v4;
-   XXH32_hash_t mem32[4];
-   XXH32_hash_t memsize;
-   XXH32_hash_t reserved;   /* never read nor write, might be removed in a future version */
-};   /* typedef'd to XXH32_state_t */
 
+/*!
+ * @}
+ * @ingroup public
+ * @{
+ */
 
-#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+#ifndef XXH_NO_LONG_LONG
+/*-**********************************************************************
+*  64-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* don't include <stdint.h> */
+/*!
+ * @brief An unsigned 64-bit integer.
+ *
+ * Not necessarily defined to `uint64_t` but functionally equivalent.
+ */
+typedef uint64_t XXH64_hash_t;
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   ifdef _AIX
+#     include <inttypes.h>
+#   else
+#     include <stdint.h>
+#   endif
+   typedef uint64_t XXH64_hash_t;
+#else
+#  include <limits.h>
+#  if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL
+     /* LP64 ABI says uint64_t is unsigned long */
+     typedef unsigned long XXH64_hash_t;
+#  else
+     /* the following type must have a width of 64-bit */
+     typedef unsigned long long XXH64_hash_t;
+#  endif
+#endif
 
-struct XXH64_state_s {
-   XXH64_hash_t total_len;
-   XXH64_hash_t v1;
-   XXH64_hash_t v2;
-   XXH64_hash_t v3;
-   XXH64_hash_t v4;
-   XXH64_hash_t mem64[4];
-   XXH32_hash_t memsize;
-   XXH32_hash_t reserved32;  /* required for padding anyway */
-   XXH64_hash_t reserved64;  /* never read nor write, might be removed in a future version */
-};   /* typedef'd to XXH64_state_t */
+/*!
+ * @}
+ *
+ * @defgroup XXH64_family XXH64 family
+ * @ingroup public
+ * @{
+ * Contains functions used in the classic 64-bit xxHash algorithm.
+ *
+ * @note
+ *   XXH3 provides competitive speed for both 32-bit and 64-bit systems,
+ *   and offers true 64/128 bit hash results.
+ *   It provides better speed for systems with vector processing capabilities.
+ */
 
+/*!
+ * @brief Calculates the 64-bit hash of @p input using xxHash64.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit xxHash64 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
 
-/*-**********************************************************************
-*  XXH3
-*  New experimental hash
-************************************************************************/
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief The opaque state struct for the XXH64 streaming API.
+ *
+ * @see XXH64_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
+ */
+typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
 
-/* ************************************************************************
- * XXH3 is a new hash algorithm featuring:
- *  - Improved speed for both small and large inputs
- *  - True 64-bit and 128-bit outputs
- *  - SIMD acceleration
- *  - Improved 32-bit viability
+/*!
+ * @brief Allocates an @ref XXH64_state_t.
  *
- * Speed analysis methodology is explained here:
+ * @return An allocated pointer of @ref XXH64_state_t on success.
+ * @return `NULL` on failure.
  *
- *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ * @note Must be freed with XXH64_freeState().
  *
- * In general, expect XXH3 to run about ~2x faster on large inputs and >3x
- * faster on small ones compared to XXH64, though exact differences depend on
- * the platform.
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
+
+/*!
+ * @brief Frees an @ref XXH64_state_t.
  *
- * The algorithm is portable: Like XXH32 and XXH64, it generates the same hash
- * on all platforms.
+ * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState().
  *
- * It benefits greatly from SIMD and 64-bit arithmetic, but does not require it.
+ * @return @ref XXH_OK.
  *
- * Almost all 32-bit and 64-bit targets that can run XXH32 smoothly can run
- * XXH3 at competitive speeds, even if XXH64 runs slowly. Further details are
- * explained in the implementation.
+ * @note @p statePtr must be allocated with XXH64_createState().
  *
- * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8,
- * ZVector and scalar targets. This can be controlled with the XXH_VECTOR macro.
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH64_state_t to another.
  *
- * XXH3 offers 2 variants, _64bits and _128bits.
- * When only 64 bits are needed, prefer calling the _64bits variant, as it
- * reduces the amount of mixing, resulting in faster speed on small inputs.
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH64_state_t to begin a new hash.
  *
- * It's also generally simpler to manipulate a scalar return type than a struct.
+ * @param statePtr The state struct to reset.
+ * @param seed The 64-bit seed to alter the hash result predictably.
  *
- * The 128-bit version adds additional strength, but it is slightly slower.
+ * @pre
+ *   @p statePtr must not be `NULL`.
  *
- * The XXH3 algorithm is still in development.
- * The results it produces may still change in future versions.
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
  *
- * Results produced by v0.7.x are not comparable with results from v0.7.y.
- * However, the API is completely stable, and it can safely be used for
- * ephemeral data (local sessions).
+ * @note This function resets and seeds a state. Call it before @ref XXH64_update().
  *
- * Avoid storing values in long-term storage until the algorithm is finalized.
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH64_state_t.
  *
- * Since v0.7.3, XXH3 has reached "release candidate" status, meaning that, if
- * everything remains fine, its current format will be "frozen" and become the
- * final one.
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
  *
- * After which, return values of XXH3 and XXH128 will no longer change in
- * future versions.
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
  *
- * XXH3's return values will be officially finalized upon reaching v0.8.0.
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
  *
- * The API supports one-shot hashing, streaming mode, and custom secrets.
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
  */
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
 
-#ifdef XXH_NAMESPACE
-#  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
-#  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
-#  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated 64-bit xxHash64 value from that state.
+ *
+ * @note
+ *   Calling XXH64_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+/*******   Canonical representation   *******/
 
-#  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
-#  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
-#  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH64_hash_t.
+ */
+typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
 
-#  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
-#  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
-#  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
-#  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
-#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
-#endif
+/*!
+ * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t.
+ *
+ * @param dst The @ref XXH64_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH64_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t.
+ *
+ * @param src The @ref XXH64_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
 
-/* 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);
+#ifndef XXH_NO_XXH3
 
-/*
- * XXH3_64bits_withSecret():
- * 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 secret *must* be large enough (>= XXH3_SECRET_SIZE_MIN).
- * It should consist of random bytes.
- * Avoid trivial sequences, such as repeating sequences and especially '\0',
- * as this can cancel out itself.
- * Failure to respect these conditions will result in a poor quality hash.
+/*!
+ * @}
+ * ************************************************************************
+ * @defgroup XXH3_family XXH3 family
+ * @ingroup public
+ * @{
+ *
+ * XXH3 is a more recent hash algorithm featuring:
+ *  - Improved speed for both small and large inputs
+ *  - True 64-bit and 128-bit outputs
+ *  - SIMD acceleration
+ *  - Improved 32-bit viability
+ *
+ * Speed analysis methodology is explained here:
+ *
+ *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ *
+ * Compared to XXH64, expect XXH3 to run approximately
+ * ~2x faster on large inputs and >3x faster on small ones,
+ * exact differences vary depending on platform.
+ *
+ * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
+ * but does not require it.
+ * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
+ * at competitive speeds, even without vector support. Further details are
+ * explained in the implementation.
+ *
+ * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD
+ * implementations for many common platforms:
+ *   - AVX512
+ *   - AVX2
+ *   - SSE2
+ *   - ARM NEON
+ *   - WebAssembly SIMD128
+ *   - POWER8 VSX
+ *   - s390x ZVector
+ * This can be controlled via the @ref XXH_VECTOR macro, but it automatically
+ * selects the best version according to predefined macros. For the x86 family, an
+ * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c.
+ *
+ * XXH3 implementation is portable:
+ * it has a generic C90 formulation that can be compiled on any platform,
+ * all implementations generate exactly the same hash value on all platforms.
+ * Starting from v0.8.0, it's also labelled "stable", meaning that
+ * any future version will also generate the same hash value.
+ *
+ * XXH3 offers 2 variants, _64bits and _128bits.
+ *
+ * When only 64 bits are needed, prefer invoking the _64bits variant, as it
+ * reduces the amount of mixing, resulting in faster speed on small inputs.
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
  */
-#define XXH3_SECRET_SIZE_MIN 136
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
 
-/*
- * XXH3_64bits_withSeed():
- * This variant generates a custom secret on the fly based on the 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().
+/*!
+ * @ingroup tuning
+ * @brief Possible values for @ref XXH_VECTOR.
+ *
+ * Unless set explicitly, determined automatically.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+#  define XXH_SCALAR 0 /*!< Portable scalar version */
+#  define XXH_SSE2   1 /*!< SSE2 for Pentium 4, Opteron, all x86_64. */
+#  define XXH_AVX2   2 /*!< AVX2 for Haswell and Bulldozer */
+#  define XXH_AVX512 3 /*!< AVX512 for Skylake and Icelake */
+#  define XXH_NEON   4 /*!< NEON for most ARMv7-A, all AArch64, and WASM SIMD128 */
+#  define XXH_VSX    5 /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+#  define XXH_SVE    6 /*!< SVE for some ARMv8-A and ARMv9-A */
+#  define XXH_LSX    7 /*!< LSX (128-bit SIMD) for LoongArch64 */
 
 
-/* streaming 64-bit */
+/*-**********************************************************************
+*  XXH3 64-bit variant
+************************************************************************/
 
-#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)   /* C11+ */
-#  include <stdalign.h>
-#  define XXH_ALIGN(n)      alignas(n)
-#elif defined(__GNUC__)
-#  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
-#elif defined(_MSC_VER)
-#  define XXH_ALIGN(n)      __declspec(align(n))
-#else
-#  define XXH_ALIGN(n)   /* disabled */
-#endif
+/*!
+ * @brief Calculates 64-bit unseeded variant of XXH3 hash of @p input.
+ *
+ * @param input  The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ *   This is equivalent to @ref XXH3_64bits_withSeed() with a seed of `0`, however
+ *   it may have slightly better performance due to constant propagation of the
+ *   defaults.
+ *
+ * @see
+ *    XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
 
-/* Old GCC versions only accept the attribute after the type in structures. */
-#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L))   /* C11+ */ \
-    && defined(__GNUC__)
-#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
-#else
-#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
-#endif
+/*!
+ * @brief Calculates 64-bit seeded variant of XXH3 hash of @p input.
+ *
+ * @param input  The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed   The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
 
-typedef struct XXH3_state_s XXH3_state_t;
+/*!
+ * The bare minimum size for a custom secret.
+ *
+ * @see
+ *  XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(),
+ *  XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
+ */
+#define XXH3_SECRET_SIZE_MIN 136
 
-#define XXH3_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
-#define XXH3_INTERNALBUFFER_SIZE 256
-struct XXH3_state_s {
-   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
-  /* used to store a custom secret generated from the seed. Makes state larger.
-   * Design might change */
-   XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
-   XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
-   XXH32_hash_t bufferedSize;
-   XXH32_hash_t nbStripesPerBlock;
-   XXH32_hash_t nbStripesSoFar;
-   XXH32_hash_t secretLimit;
-   XXH32_hash_t reserved32;
-   XXH32_hash_t reserved32_2;
-   XXH64_hash_t totalLen;
-   XXH64_hash_t seed;
-   XXH64_hash_t reserved64;
-   /* note: there is some padding after due to alignment on 64 bytes */
-   const unsigned char* secret;
-};   /* typedef'd to XXH3_state_t */
+/*!
+ * @brief Calculates 64-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @pre
+ *   The memory between @p data and @p data + @p len must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p data may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
 
-#undef XXH_ALIGN_MEMBER
 
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
 /*
  * Streaming requires state maintenance.
  * This operation costs memory and CPU.
  * As a consequence, streaming is slower than one-shot hashing.
- * For better performance, prefer one-shot functions whenever possible.
+ * For better performance, prefer one-shot functions whenever applicable.
  */
-XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void);
-XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
-XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state);
 
-
-/*
- * XXH3_64bits_reset():
- * Initialize with the default parameters.
- * The result will be equivalent to `XXH3_64bits()`.
- */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr);
-/*
- * XXH3_64bits_reset_withSeed():
- * Generate a custom secret from `seed`, and store it into `statePtr`.
- * digest will be equivalent to `XXH3_64bits_withSeed()`.
+/*!
+ * @brief The opaque state struct for the XXH3 streaming API.
+ *
+ * @see XXH3_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-/*
- * XXH3_64bits_reset_withSecret():
- * `secret` is referenced, and must outlive the hash streaming session, so
- * be careful when using stack arrays.
- * `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`.
+typedef struct XXH3_state_s XXH3_state_t;
+XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH3_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
 
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_digest (const XXH3_state_t* statePtr);
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret with default parameters.
+ *   - Call this function before @ref XXH3_64bits_update().
+ *   - Digest will be equivalent to `XXH3_64bits()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
 
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret from `seed`.
+ *   - Call this function before @ref XXH3_64bits_update().
+ *   - Digest will be equivalent to `XXH3_64bits_withSeed()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
 
-/* 128-bit */
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   `secret` is referenced, it _must outlive_ the hash streaming session.
+ *
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
 
-#ifdef XXH_NAMESPACE
-#  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
-#  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
-#  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
-#  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
 
-#  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
-#  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
-#  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
-#  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
-#  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
+/*!
+ * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 64-bit hash value from that state.
+ *
+ * @note
+ *   Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t  XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
-#  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
-#  define XXH128_cmp     XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
-#  define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
-#  define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
-#endif
+/* note : canonical representation of XXH3 is the same as XXH64
+ * since they both produce XXH64_hash_t values */
+
+
+/*-**********************************************************************
+*  XXH3 128-bit variant
+************************************************************************/
 
+/*!
+ * @brief The return value from 128-bit hashes.
+ *
+ * Stored in little endian order, although the fields themselves are in native
+ * endianness.
+ */
 typedef struct {
-    XXH64_hash_t low64;
-    XXH64_hash_t high64;
+    XXH64_hash_t low64;   /*!< `value & 0xFFFFFFFFFFFFFFFF` */
+    XXH64_hash_t high64;  /*!< `value >> 64` */
 } XXH128_hash_t;
 
-XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed);
-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);  /* == XXH128() */
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+/*!
+ * @brief Calculates 128-bit unseeded variant of XXH3 of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len  The length of @p data, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
+ * for shorter inputs.
+ *
+ * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of `0`, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
+/*! @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len  The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see XXH3_128bits(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+/*!
+ * @brief Calculates 128-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ *
+ * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits().
+ * Use already declared XXH3_createState() and XXH3_freeState().
+ *
+ * All reset and streaming functions have same meaning as their 64-bit counterpart.
+ */
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret with default parameters.
+ *   - Call it before @ref XXH3_128bits_update().
+ *   - Digest will be equivalent to `XXH3_128bits()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
 
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret from `seed`.
+ *   - Call it before @ref XXH3_128bits_update().
+ *   - Digest will be equivalent to `XXH3_128bits_withSeed()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   The state struct to reset.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+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);
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
 
+/*!
+ * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 128-bit hash value from that state.
+ *
+ * @note
+ *   Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
-/* Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
+/* Following helper functions make it possible to compare XXH128_hast_t values.
+ * Since XXH128_hash_t is a structure, this capability is not offered by the language.
+ * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
 
 /*!
- * XXH128_isEqual():
- * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
+ * @brief Check equality of two XXH128_hash_t values
+ *
+ * @param h1 The 128-bit hash value.
+ * @param h2 Another 128-bit hash value.
+ *
+ * @return `1` if `h1` and `h2` are equal.
+ * @return `0` if they are not.
  */
-XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
 
 /*!
- * XXH128_cmp():
+ * @brief Compares two @ref XXH128_hash_t
  *
  * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
  *
- * return: >0 if *h128_1  > *h128_2
- *         <0 if *h128_1  < *h128_2
- *         =0 if *h128_1 == *h128_2
+ * @param h128_1 Left-hand side value
+ * @param h128_2 Right-hand side value
+ *
+ * @return >0 if @p h128_1  > @p h128_2
+ * @return =0 if @p h128_1 == @p h128_2
+ * @return <0 if @p h128_1  < @p h128_2
  */
-XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2);
+XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
 
 
 /*******   Canonical representation   *******/
-typedef struct { unsigned char digest[16]; } 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);
+typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
+
+
+/*!
+ * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t.
+ *
+ * @param dst  The @ref XXH128_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH128_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t.
+ *
+ * @param src The @ref XXH128_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
 
 
+#endif  /* !XXH_NO_XXH3 */
 #endif  /* XXH_NO_LONG_LONG */
 
-#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
-#  define XXH_IMPLEMENTATION
+/*!
+ * @}
+ */
+#endif /* XXHASH_H_5627135585666179 */
+
+
+
+#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
+#define XXHASH_H_STATIC_13879238742
+/* ****************************************************************************
+ * This section contains declarations which are not guaranteed to remain stable.
+ * They may change in future versions, becoming incompatible with a different
+ * version of the library.
+ * These declarations should only be used with static linking.
+ * Never use them in association with dynamic linking!
+ ***************************************************************************** */
+
+/*
+ * These definitions are only present to allow static allocation
+ * of XXH states, on stack or in a struct, for example.
+ * Never **ever** access their members directly.
+ */
+
+/*!
+ * @internal
+ * @brief Structure for XXH32 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH32_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH64_state_s, XXH3_state_s
+ */
+struct XXH32_state_s {
+   XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
+   XXH32_hash_t large_len;    /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
+   XXH32_hash_t acc[4];       /*!< Accumulator lanes */
+   unsigned char buffer[16];  /*!< Internal buffer for partial reads. */
+   XXH32_hash_t bufferedSize; /*!< Amount of data in @ref buffer */
+   XXH32_hash_t reserved;     /*!< Reserved field. Do not read nor write to it. */
+};   /* typedef'd to XXH32_state_t */
+
+
+#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+
+/*!
+ * @internal
+ * @brief Structure for XXH64 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH64_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH32_state_s, XXH3_state_s
+ */
+struct XXH64_state_s {
+   XXH64_hash_t total_len;    /*!< Total length hashed. This is always 64-bit. */
+   XXH64_hash_t acc[4];       /*!< Accumulator lanes */
+   unsigned char buffer[32];  /*!< Internal buffer for partial reads.. */
+   XXH32_hash_t bufferedSize; /*!< Amount of data in @ref buffer */
+   XXH32_hash_t reserved32;   /*!< Reserved field, needed for padding anyways*/
+   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it. */
+};   /* typedef'd to XXH64_state_t */
+
+#ifndef XXH_NO_XXH3
+
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
+#  define XXH_ALIGN(n)      _Alignas(n)
+#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */
+/* In C++ alignas() is a keyword */
+#  define XXH_ALIGN(n)      alignas(n)
+#elif defined(__GNUC__)
+#  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
+#elif defined(_MSC_VER)
+#  define XXH_ALIGN(n)      __declspec(align(n))
+#else
+#  define XXH_ALIGN(n)   /* disabled */
 #endif
 
-#endif  /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
+/* Old GCC versions only accept the attribute after the type in structures. */
+#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L))   /* C11+ */ \
+    && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \
+    && defined(__GNUC__)
+#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
+#else
+#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
+#endif
+
+/*!
+ * @brief The size of the internal XXH3 buffer.
+ *
+ * This is the optimal update size for incremental hashing.
+ *
+ * @see XXH3_64b_update(), XXH3_128b_update().
+ */
+#define XXH3_INTERNALBUFFER_SIZE 256
 
+/*!
+ * @internal
+ * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
+ *
+ * This is the size used in @ref XXH3_kSecret and the seeded functions.
+ *
+ * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
+ */
+#define XXH3_SECRET_DEFAULT_SIZE 192
 
-/* ======================================================================== */
-/* ======================================================================== */
-/* ======================================================================== */
+/*!
+ * @internal
+ * @brief Structure for XXH3 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined.
+ * Otherwise it is an opaque type.
+ * Never use this definition in combination with dynamic library.
+ * This allows fields to safely be changed in the future.
+ *
+ * @note ** This structure has a strict alignment requirement of 64 bytes!! **
+ * Do not allocate this with `malloc()` or `new`,
+ * it will not be sufficiently aligned.
+ * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation.
+ *
+ * Typedef'd to @ref XXH3_state_t.
+ * Do never access the members of this struct directly.
+ *
+ * @see XXH3_INITSTATE() for stack initialization.
+ * @see XXH3_createState(), XXH3_freeState().
+ * @see XXH32_state_s, XXH64_state_s
+ */
+struct XXH3_state_s {
+   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
+       /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
+   XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
+       /*!< Used to store a custom secret generated from a seed. */
+   XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
+       /*!< The internal buffer. @see XXH32_state_s::mem32 */
+   XXH32_hash_t bufferedSize;
+       /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
+   XXH32_hash_t useSeed;
+       /*!< Reserved field. Needed for padding on 64-bit. */
+   size_t nbStripesSoFar;
+       /*!< Number or stripes processed. */
+   XXH64_hash_t totalLen;
+       /*!< Total length hashed. 64-bit even on 32-bit targets. */
+   size_t nbStripesPerBlock;
+       /*!< Number of stripes per block. */
+   size_t secretLimit;
+       /*!< Size of @ref customSecret or @ref extSecret */
+   XXH64_hash_t seed;
+       /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
+   XXH64_hash_t reserved64;
+       /*!< Reserved field. */
+   const unsigned char* extSecret;
+       /*!< Reference to an external secret for the _withSecret variants, NULL
+        *   for other variants. */
+   /* note: there may be some padding at the end due to alignment on 64 bytes */
+}; /* typedef'd to XXH3_state_t */
 
+#undef XXH_ALIGN_MEMBER
 
-/*-**********************************************************************
- * xxHash implementation
- *-**********************************************************************
- * xxHash's implementation used to be found in xxhash.c.
+/*!
+ * @brief Initializes a stack-allocated `XXH3_state_s`.
+ *
+ * When the @ref XXH3_state_t structure is merely emplaced on stack,
+ * it should be initialized with XXH3_INITSTATE() or a memset()
+ * in case its first reset uses XXH3_NNbits_reset_withSeed().
+ * This init can be omitted if the first reset uses default or _withSecret mode.
+ * This operation isn't necessary when the state is created with XXH3_createState().
+ * Note that this doesn't prepare the state for a streaming operation,
+ * it's still necessary to use XXH3_NNbits_reset*() afterwards.
+ */
+#define XXH3_INITSTATE(XXH3_state_ptr)                       \
+    do {                                                     \
+        XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
+        tmp_xxh3_state_ptr->seed = 0;                        \
+        tmp_xxh3_state_ptr->extSecret = NULL;                \
+    } while(0)
+
+
+/*!
+ * @brief Calculates the 128-bit hash of @p data using XXH3.
  *
- * However, code inlining requires the implementation to be visible to the
- * compiler, usually within the header.
+ * @param data The block of data to be hashed, at least @p len bytes in size.
+ * @param len  The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
  *
- * As a workaround, xxhash.c used to be included within xxhash.h. This caused
- * some issues with some build systems, especially ones which treat .c files
- * as source files.
+ * @pre
+ *   The memory between @p data and @p data + @p len must be valid,
+ *   readable, contiguous memory. However, if @p len is `0`, @p data may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
  *
- * Therefore, the implementation is now directly integrated within xxhash.h.
- * Another small advantage is that xxhash.c is no longer needed in /include.
- ************************************************************************/
+ * @return The calculated 128-bit XXH3 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
 
-#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
-   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
-#  define XXH_IMPLEM_13a8737387
 
-/* *************************************
-*  Tuning parameters
-***************************************/
+/* ===   Experimental API   === */
+/* Symbols defined below must be considered tied to a specific library version. */
+
 /*!
- * XXH_FORCE_MEMORY_ACCESS:
- * By default, access to unaligned memory is controlled by `memcpy()`, which is
- * safe and portable.
+ * @brief Derive a high-entropy secret from any user-defined content, named customSeed.
  *
- * Unfortunately, on some target/compiler combinations, the generated assembly
- * is sub-optimal.
+ * @param secretBuffer    A writable buffer for derived high-entropy secret data.
+ * @param secretSize      Size of secretBuffer, in bytes.  Must be >= XXH3_SECRET_SIZE_MIN.
+ * @param customSeed      A user-defined content.
+ * @param customSeedSize  Size of customSeed, in bytes.
  *
- * The below switch allow to select a different access method for improved
- * performance.
- * Method 0 (default):
- *     Use `memcpy()`. Safe and portable.
- * Method 1:
- *     `__attribute__((packed))` statement. It depends on compiler extensions
- *     and is therefore not portable.
- *     This method is safe if your compiler supports it, and *generally* as
- *     fast or faster than `memcpy`.
- * Method 2:
- *     Direct access via cast. This method doesn't depend on the compiler but
- *     violates the C standard.
- *     It can generate buggy code on targets which do not support unaligned
- *     memory accesses.
- *     But in some circumstances, it's the only known way to get the most
- *     performance (ie GCC + ARMv6)
- * Method 3:
- *     Byteshift. This can generate the best code on old compilers which don't
- *     inline small `memcpy()` calls, and it might also be faster on big-endian
- *     systems which lack a native byteswap instruction.
- * See https://stackoverflow.com/a/32095106/646947 for details.
- * Prefer these methods in priority order (0 > 1 > 2 > 3)
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * The generated secret can be used in combination with `*_withSecret()` functions.
+ * The `_withSecret()` variants are useful to provide a higher level of protection
+ * than 64-bit seed, as it becomes much more difficult for an external actor to
+ * guess how to impact the calculation logic.
+ *
+ * The function accepts as input a custom seed of any length and any content,
+ * and derives from it a high-entropy secret of length @p secretSize into an
+ * already allocated buffer @p secretBuffer.
+ *
+ * The generated secret can then be used with any `*_withSecret()` variant.
+ * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(),
+ * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret()
+ * are part of this list. They all accept a `secret` parameter
+ * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN)
+ * _and_ feature very high entropy (consist of random-looking bytes).
+ * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
+ * be employed to ensure proper quality.
+ *
+ * @p customSeed can be anything. It can have any size, even small ones,
+ * and its content can be anything, even "poor entropy" sources such as a bunch
+ * of zeroes. The resulting `secret` will nonetheless provide all required qualities.
+ *
+ * @pre
+ *   - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN
+ *   - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ *
+ * Example code:
+ * @code{.c}
+ *    #include <stdio.h>
+ *    #include <stdlib.h>
+ *    #include <string.h>
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Hashes argv[2] using the entropy from argv[1].
+ *    int main(int argc, char* argv[])
+ *    {
+ *        char secret[XXH3_SECRET_SIZE_MIN];
+ *        if (argv != 3) { return 1; }
+ *        XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1]));
+ *        XXH64_hash_t h = XXH3_64bits_withSecret(
+ *             argv[2], strlen(argv[2]),
+ *             secret, sizeof(secret)
+ *        );
+ *        printf("%016llx\n", (unsigned long long) h);
+ *    }
+ * @endcode
  */
-#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if !defined(__clang__) && defined(__GNUC__) && defined(__ARM_FEATURE_UNALIGNED) && defined(__ARM_ARCH) && (__ARM_ARCH == 6)
-#    define XXH_FORCE_MEMORY_ACCESS 2
-#  elif !defined(__clang__) && ((defined(__INTEL_COMPILER) && !defined(_WIN32)) || \
-  (defined(__GNUC__) && (defined(__ARM_ARCH) && __ARM_ARCH >= 7)))
-#    define XXH_FORCE_MEMORY_ACCESS 1
-#  endif
+XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize);
+
+/*!
+ * @brief Generate the same secret as the _withSeed() variants.
+ *
+ * @param secretBuffer A writable buffer of @ref XXH3_SECRET_DEFAULT_SIZE bytes
+ * @param seed         The 64-bit seed to alter the hash result predictably.
+ *
+ * The generated secret can be used in combination with
+ *`*_withSecret()` and `_withSecretandSeed()` variants.
+ *
+ * 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_DEFAULT_SIZE];
+ *    public:
+ *        HashFast(XXH64_hash_t s) {
+ *            XXH3_generateSecret_fromSeed(secret, seed);
+ *        }
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{
+ *                XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret))
+ *            };
+ *        }
+ *    };
+ * @endcode
+ */
+XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
+
+/*!
+ * @brief Maximum size of "short" key in bytes.
+ */
+#define XXH3_MIDSIZE_MAX 240
+
+/*!
+ * @brief Calculates 64/128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed       The 64-bit seed to alter the hash result predictably.
+ *
+ * These variants generate hash values using either:
+ * - @p seed for "short" keys (< @ref XXH3_MIDSIZE_MAX = 240 bytes)
+ * - @p secret for "large" keys (>= @ref XXH3_MIDSIZE_MAX).
+ *
+ * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
+ * `_withSeed()` has to generate the secret on the fly for "large" keys.
+ * It's fast, but can be perceptible for "not so large" keys (< 1 KB).
+ * `_withSecret()` has to generate the masks on the fly for "small" keys,
+ * which requires more instructions than _withSeed() variants.
+ * Therefore, _withSecretandSeed variant combines the best of both worlds.
+ *
+ * When @p secret has been generated by XXH3_generateSecret_fromSeed(),
+ * this variant produces *exactly* the same results as `_withSeed()` variant,
+ * hence offering only a pure speed benefit on "large" input,
+ * by skipping the need to regenerate the secret for every large input.
+ *
+ * Another usage scenario is to hash the secret to a 64-bit hash value,
+ * for example with XXH3_64bits(), which then becomes the seed,
+ * and then employ both the seed and the secret in _withSecretandSeed().
+ * On top of speed, an added benefit is that each bit in the secret
+ * has a 50% chance to swap each bit in the output, via its impact to the seed.
+ *
+ * This is not guaranteed when using the secret directly in "small data" scenarios,
+ * because only portions of the secret are employed for small data.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len,
+                              XXH_NOESCAPE const void* secret, size_t secretSize,
+                              XXH64_hash_t seed);
+
+/*!
+ * @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data       The memory segment to be hashed, at least @p len bytes in size.
+ * @param length     The length of @p data, in bytes.
+ * @param secret     The secret used to alter hash result predictably.
+ * @param secretSize The length of @p secret, in bytes (must be >= XXH3_SECRET_SIZE_MIN)
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(): contract is the same.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
+                               XXH_NOESCAPE const void* secret, size_t secretSize,
+                               XXH64_hash_t seed64);
+
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(). Contract is identical.
+ */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                    XXH_NOESCAPE const void* secret, size_t secretSize,
+                                    XXH64_hash_t seed64);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(). Contract is identical.
+ *
+ * Note: there was a bug in an earlier version of this function (<= v0.8.2)
+ * that would make it generate an incorrect hash value
+ * when @p seed == 0 and @p length < XXH3_MIDSIZE_MAX
+ * and @p secret is different from XXH3_generateSecret_fromSeed().
+ * As stated in the contract, the correct hash result must be
+ * the same as XXH3_128bits_withSeed() when @p length <= XXH3_MIDSIZE_MAX.
+ * Results generated by this older version are wrong, hence not comparable.
+ */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                     XXH_NOESCAPE const void* secret, size_t secretSize,
+                                     XXH64_hash_t seed64);
+
+#endif /* !XXH_NO_STREAM */
+
+#endif  /* !XXH_NO_XXH3 */
+#endif  /* XXH_NO_LONG_LONG */
+#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
+#  define XXH_IMPLEMENTATION
+#endif
+
+#endif  /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
+
+
+/* ======================================================================== */
+/* ======================================================================== */
+/* ======================================================================== */
+
+
+/*-**********************************************************************
+ * xxHash implementation
+ *-**********************************************************************
+ * xxHash's implementation used to be hosted inside xxhash.c.
+ *
+ * However, inlining requires implementation to be visible to the compiler,
+ * hence be included alongside the header.
+ * Previously, implementation was hosted inside xxhash.c,
+ * which was then #included when inlining was activated.
+ * This construction created issues with a few build and install systems,
+ * as it required xxhash.c to be stored in /include directory.
+ *
+ * xxHash implementation is now directly integrated within xxhash.h.
+ * As a consequence, xxhash.c is no longer needed in /include.
+ *
+ * xxhash.c is still available and is still useful.
+ * In a "normal" setup, when xxhash is not inlined,
+ * xxhash.h only exposes the prototypes and public symbols,
+ * while xxhash.c can be built into an object file xxhash.o
+ * which can then be linked into the final binary.
+ ************************************************************************/
+
+#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
+   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
+#  define XXH_IMPLEM_13a8737387
+
+/* *************************************
+*  Tuning parameters
+***************************************/
+
+/*!
+ * @defgroup tuning Tuning parameters
+ * @{
+ *
+ * Various macros to control xxHash's behavior.
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Define this to disable 64-bit code.
+ *
+ * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
+ */
+#  define XXH_NO_LONG_LONG
+#  undef XXH_NO_LONG_LONG /* don't actually */
+/*!
+ * @brief Controls how unaligned memory is accessed.
+ *
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is
+ * safe and portable.
+ *
+ * Unfortunately, on some target/compiler combinations, the generated assembly
+ * is sub-optimal.
+ *
+ * The below switch allow selection of a different access method
+ * in the search for improved performance.
+ *
+ * @par Possible options:
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy`
+ *   @par
+ *     Use `memcpy()`. Safe and portable. Note that most modern compilers will
+ *     eliminate the function call and treat it as an unaligned access.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))`
+ *   @par
+ *     Depends on compiler extensions and is therefore not portable.
+ *     This method is safe _if_ your compiler supports it,
+ *     and *generally* as fast or faster than `memcpy`.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast
+ *  @par
+ *     Casts directly and dereferences. This method doesn't depend on the
+ *     compiler, but it violates the C standard as it directly dereferences an
+ *     unaligned pointer. It can generate buggy code on targets which do not
+ *     support unaligned memory accesses, but in some circumstances, it's the
+ *     only known way to get the most performance.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift
+ *  @par
+ *     Also portable. This can generate the best code on old compilers which don't
+ *     inline small `memcpy()` calls, and it might also be faster on big-endian
+ *     systems which lack a native byteswap instruction. However, some compilers
+ *     will emit literal byteshifts even if the target supports unaligned access.
+ *
+ *
+ * @warning
+ *   Methods 1 and 2 rely on implementation-defined behavior. Use these with
+ *   care, as what works on one compiler/platform/optimization level may cause
+ *   another to read garbage data or even crash.
+ *
+ * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.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).
+ *
+ * This is an important performance trick for architectures without decent
+ * unaligned memory access performance.
+ *
+ * It checks for input alignment, and when conditions are met, uses a "fast
+ * path" employing direct 32-bit/64-bit reads, resulting in _dramatically
+ * faster_ read speed.
+ *
+ * The check costs one initial branch per hash, which is generally negligible,
+ * but not zero.
+ *
+ * Moreover, it's not useful to generate an additional code path if memory
+ * access uses the same instruction for both aligned and unaligned
+ * addresses (e.g. x86 and aarch64).
+ *
+ * In these cases, the alignment check can be removed by setting this macro to 0.
+ * Then the code will always use unaligned memory access.
+ * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips
+ * which are platforms known to offer good unaligned memory accesses performance.
+ *
+ * It is also disabled by default when @ref XXH_SIZE_OPT >= 1.
+ *
+ * This option does not affect XXH3 (only XXH32 and XXH64).
+ */
+#  define XXH_FORCE_ALIGN_CHECK 0
+
+/*!
+ * @def XXH_NO_INLINE_HINTS
+ * @brief When non-zero, sets all functions to `static`.
+ *
+ * By default, xxHash tries to force the compiler to inline almost all internal
+ * functions.
+ *
+ * This can usually improve performance due to reduced jumping and improved
+ * constant folding, but significantly increases the size of the binary which
+ * might not be favorable.
+ *
+ * Additionally, sometimes the forced inlining can be detrimental to performance,
+ * depending on the architecture.
+ *
+ * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
+ * compiler full control on whether to inline or not.
+ *
+ * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if
+ * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
+ */
+#  define XXH_NO_INLINE_HINTS 0
+
+/*!
+ * @def XXH3_INLINE_SECRET
+ * @brief Determines whether to inline the XXH3 withSecret code.
+ *
+ * When the secret size is known, the compiler can improve the performance
+ * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret().
+ *
+ * However, if the secret size is not known, it doesn't have any benefit. This
+ * happens when xxHash is compiled into a global symbol. Therefore, if
+ * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0.
+ *
+ * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers
+ * that are *sometimes* force inline on -Og, and it is impossible to automatically
+ * detect this optimization level.
+ */
+#  define XXH3_INLINE_SECRET 0
+
+/*!
+ * @def XXH32_ENDJMP
+ * @brief Whether to use a jump for `XXH32_finalize`.
+ *
+ * For performance, `XXH32_finalize` uses multiple branches in the finalizer.
+ * This is generally preferable for performance,
+ * but depending on exact architecture, a jmp may be preferable.
+ *
+ * This setting is only possibly making a difference for very small inputs.
+ */
+#  define XXH32_ENDJMP 0
+
+/*!
+ * @internal
+ * @brief Redefines old internal names.
+ *
+ * For compatibility with code that uses xxHash's internals before the names
+ * were changed to improve namespacing. There is no other reason to use this.
+ */
+#  define XXH_OLD_NAMES
+#  undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
+
+/*!
+ * @def XXH_NO_STREAM
+ * @brief Disables the streaming API.
+ *
+ * When xxHash is not inlined and the streaming functions are not used, disabling
+ * the streaming functions can improve code size significantly, especially with
+ * the @ref XXH3_family which tends to make constant folded copies of itself.
+ */
+#  define XXH_NO_STREAM
+#  undef XXH_NO_STREAM /* don't actually */
+#endif /* XXH_DOXYGEN */
+/*!
+ * @}
+ */
+
+#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+   /* prefer __packed__ structures (method 1) for GCC
+    * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
+    * which for some reason does unaligned loads. */
+#  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 */
+   /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */
+#  if XXH_SIZE_OPT >= 1 || \
+      defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
+   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64)    || defined(_M_ARM) /* visual */
+#    define XXH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+#ifndef XXH_NO_INLINE_HINTS
+#  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
+#endif
+
+/*!
+ * @defgroup impl Implementation
+ * @{
+ */
+
+
+/* *************************************
+*  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()
+ */
+#include <stdlib.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than malloc().
+ */
+static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than free().
+ */
+static void XXH_free(void* p) { free(p); }
+
+#endif  /* XXH_NO_STDLIB */
+
+#include <string.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than memcpy().
+ */
+static void* XXH_memcpy(void* dest, const void* src, size_t size)
+{
+    return memcpy(dest,src,size);
+}
+
+#include <limits.h>   /* ULLONG_MAX */
+
+
+/* *************************************
+*  Compiler Specific Options
+***************************************/
+#ifdef _MSC_VER /* Visual Studio warning fix */
+#  pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+#endif
+
+#if XXH_NO_INLINE_HINTS  /* disable inlining hints */
+#  if defined(__GNUC__) || defined(__clang__)
+#    define XXH_FORCE_INLINE static __attribute__((__unused__))
+#  else
+#    define XXH_FORCE_INLINE static
+#  endif
+#  define XXH_NO_INLINE static
+/* enable inlining hints */
+#elif defined(__GNUC__) || defined(__clang__)
+#  define XXH_FORCE_INLINE static __inline__ __attribute__((__always_inline__, __unused__))
+#  define XXH_NO_INLINE static __attribute__((__noinline__))
+#elif defined(_MSC_VER)  /* Visual Studio */
+#  define XXH_FORCE_INLINE static __forceinline
+#  define XXH_NO_INLINE static __declspec(noinline)
+#elif defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* C99 */
+#  define XXH_FORCE_INLINE static inline
+#  define XXH_NO_INLINE static
+#else
+#  define XXH_FORCE_INLINE static
+#  define XXH_NO_INLINE static
+#endif
+
+#if defined(XXH_INLINE_ALL)
+#  define XXH_STATIC XXH_FORCE_INLINE
+#else
+#  define XXH_STATIC static
+#endif
+
+#if XXH3_INLINE_SECRET
+#  define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
+#else
+#  define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
+#endif
+
+#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
+#  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
+#  define XXH_RESTRICT   /* disable */
+#endif
+
+/* *************************************
+*  Debug
+***************************************/
+/*!
+ * @ingroup tuning
+ * @def XXH_DEBUGLEVEL
+ * @brief Sets the debugging level.
+ *
+ * XXH_DEBUGLEVEL is expected to be defined externally, typically via the
+ * compiler's command line options. The value must be a number.
+ */
+#ifndef XXH_DEBUGLEVEL
+#  ifdef DEBUGLEVEL /* backwards compat */
+#    define XXH_DEBUGLEVEL DEBUGLEVEL
+#  else
+#    define XXH_DEBUGLEVEL 0
+#  endif
+#endif
+
+#if (XXH_DEBUGLEVEL>=1)
+#  include <assert.h>   /* note: can still be disabled with NDEBUG */
+#  define XXH_ASSERT(c)   assert(c)
+#else
+#  if defined(__INTEL_COMPILER)
+#    define XXH_ASSERT(c)   XXH_ASSUME((unsigned char) (c))
+#  else
+#    define XXH_ASSERT(c)   XXH_ASSUME(c)
+#  endif
+#endif
+
+/* note: use after variable declarations */
+#ifndef XXH_STATIC_ASSERT
+#  if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)    /* C11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
+#  elif defined(__cplusplus) && (__cplusplus >= 201103L)            /* C++11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
+#  else
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0)
+#  endif
+#  define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c)
+#endif
+
+/*!
+ * @internal
+ * @def XXH_COMPILER_GUARD(var)
+ * @brief Used to prevent unwanted optimizations for @p var.
+ *
+ * It uses an empty GCC inline assembly statement with a register constraint
+ * which forces @p var into a general purpose register (eg eax, ebx, ecx
+ * on x86) and marks it as modified.
+ *
+ * This is used in a few places to avoid unwanted autovectorization (e.g.
+ * XXH32_round()). All vectorization we want is explicit via intrinsics,
+ * and _usually_ isn't wanted elsewhere.
+ *
+ * We also use it to prevent unwanted constant folding for AArch64 in
+ * XXH3_initCustomSecret_scalar().
+ */
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
+#else
+#  define XXH_COMPILER_GUARD(var) ((void)0)
+#endif
+
+/* Specifically for NEON vectors which use the "w" constraint, on
+ * Clang. */
+#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__)
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var))
+#else
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0)
+#endif
+
+/* *************************************
+*  Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   ifdef _AIX
+#     include <inttypes.h>
+#   else
+#     include <stdint.h>
+#   endif
+    typedef uint8_t xxh_u8;
+#else
+    typedef unsigned char xxh_u8;
+#endif
+typedef XXH32_hash_t xxh_u32;
+
+#ifdef XXH_OLD_NAMES
+#  warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly"
+#  define BYTE xxh_u8
+#  define U8   xxh_u8
+#  define U32  xxh_u32
+#endif
+
+/* ***   Memory access   *** */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_read32(const void* ptr)
+ * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit native endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit little endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readBE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit big endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit big endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align)
+ * @brief Like @ref XXH_readLE32(), but has an option for aligned reads.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is
+ * always @ref XXH_alignment::XXH_unaligned.
+ *
+ * @param ptr The pointer to read from.
+ * @param align Whether @p ptr is aligned.
+ * @pre
+ *   If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte
+ *   aligned.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE32 and XXH_readBE32.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/*
+ * Force direct memory access. Only works on CPU which support unaligned memory
+ * access in hardware.
+ */
+static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __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 __attribute__((__aligned__(1))) xxh_u32 xxh_unalign32;
+    return *((const xxh_unalign32*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u32 XXH_read32(const void* memPtr)
+{
+    xxh_u32 val;
+    XXH_memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ***   Endianness   *** */
+
+/*!
+ * @ingroup tuning
+ * @def XXH_CPU_LITTLE_ENDIAN
+ * @brief Whether the target is little endian.
+ *
+ * Defined to 1 if the target is little endian, or 0 if it is big endian.
+ * It can be defined externally, for example on the compiler command line.
+ *
+ * If it is not defined,
+ * a runtime check (which is usually constant folded) is used instead.
+ *
+ * @note
+ *   This is not necessarily defined to an integer constant.
+ *
+ * @see XXH_isLittleEndian() for the runtime check.
+ */
+#ifndef XXH_CPU_LITTLE_ENDIAN
+/*
+ * Try to detect endianness automatically, to avoid the nonstandard behavior
+ * in `XXH_isLittleEndian()`
+ */
+#  if defined(_WIN32) /* Windows is always little endian */ \
+     || defined(__LITTLE_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 1
+#  elif defined(__BIG_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 0
+#  else
+/*!
+ * @internal
+ * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN.
+ *
+ * Most compilers will constant fold this.
+ */
+static int XXH_isLittleEndian(void)
+{
+    /*
+     * Portable and well-defined behavior.
+     * Don't use static: it is detrimental to performance.
+     */
+    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
+    return one.c[0];
+}
+#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
+#  endif
+#endif
+
+
+
+
+/* ****************************************
+*  Compiler-specific Functions and Macros
+******************************************/
+#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+#ifdef __has_builtin
+#  define XXH_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  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)
+ * @brief 32-bit rotate left.
+ *
+ * @param x The 32-bit integer to be rotated.
+ * @param r The number of bits to rotate.
+ * @pre
+ *   @p r > 0 && @p r < 32
+ * @note
+ *   @p x and @p r may be evaluated multiple times.
+ * @return The rotated result.
+ */
+#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \
+                               && XXH_HAS_BUILTIN(__builtin_rotateleft64)
+#  define XXH_rotl32 __builtin_rotateleft32
+#  define XXH_rotl64 __builtin_rotateleft64
+#elif XXH_HAS_BUILTIN(__builtin_stdc_rotate_left)
+#  define XXH_rotl32 __builtin_stdc_rotate_left
+#  define XXH_rotl64 __builtin_stdc_rotate_left
+/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
+#elif defined(_MSC_VER)
+#  define XXH_rotl32(x,r) _rotl(x,r)
+#  define XXH_rotl64(x,r) _rotl64(x,r)
+#else
+#  define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
+#  define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
+#endif
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_swap32(xxh_u32 x)
+ * @brief A 32-bit byteswap.
+ *
+ * @param x The 32-bit integer to byteswap.
+ * @return @p x, byteswapped.
+ */
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap32 _byteswap_ulong
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap32 __builtin_bswap32
+#else
+static xxh_u32 XXH_swap32 (xxh_u32 x)
+{
+    return  ((x << 24) & 0xff000000 ) |
+            ((x <<  8) & 0x00ff0000 ) |
+            ((x >>  8) & 0x0000ff00 ) |
+            ((x >> 24) & 0x000000ff );
+}
+#endif
+
+
+/* ***************************
+*  Memory reads
+*****************************/
+
+/*!
+ * @internal
+ * @brief Enum to indicate whether a pointer is aligned.
+ */
+typedef enum {
+    XXH_aligned,  /*!< Aligned */
+    XXH_unaligned /*!< Possibly unaligned */
+} XXH_alignment;
+
+/*
+ * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
+ *
+ * This is ideal for older compilers which don't inline memcpy.
+ */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u32)bytePtr[1] << 8)
+         | ((xxh_u32)bytePtr[2] << 16)
+         | ((xxh_u32)bytePtr[3] << 24);
+}
+
+XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[3]
+         | ((xxh_u32)bytePtr[2] << 8)
+         | ((xxh_u32)bytePtr[1] << 16)
+         | ((xxh_u32)bytePtr[0] << 24);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+}
+
+static xxh_u32 XXH_readBE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u32
+XXH_readLE32_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned) {
+        return XXH_readLE32(ptr);
+    } else {
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
+    }
+}
+
+
+/* *************************************
+*  Misc
+***************************************/
+/*! @ingroup public */
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* *******************************************************************
+*  32-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @defgroup XXH32_impl XXH32 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH32 implementation.
+ * @{
+ */
+ /* #define instead of static const, to be used as initializers */
+#define XXH_PRIME32_1  0x9E3779B1U  /*!< 0b10011110001101110111100110110001 */
+#define XXH_PRIME32_2  0x85EBCA77U  /*!< 0b10000101111010111100101001110111 */
+#define XXH_PRIME32_3  0xC2B2AE3DU  /*!< 0b11000010101100101010111000111101 */
+#define XXH_PRIME32_4  0x27D4EB2FU  /*!< 0b00100111110101001110101100101111 */
+#define XXH_PRIME32_5  0x165667B1U  /*!< 0b00010110010101100110011110110001 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME32_1 XXH_PRIME32_1
+#  define PRIME32_2 XXH_PRIME32_2
+#  define PRIME32_3 XXH_PRIME32_3
+#  define PRIME32_4 XXH_PRIME32_4
+#  define PRIME32_5 XXH_PRIME32_5
+#endif
+
+/*!
+ * @internal
+ * @brief Normal stripe processing routine.
+ *
+ * This shuffles the bits so that any bit from @p input impacts several bits in
+ * @p acc.
+ *
+ * @param acc The accumulator lane.
+ * @param input The stripe of input to mix.
+ * @return The mixed accumulator lane.
+ */
+static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+{
+    acc += input * XXH_PRIME32_2;
+    acc  = XXH_rotl32(acc, 13);
+    acc *= XXH_PRIME32_1;
+#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * UGLY HACK:
+     * A compiler fence is used to prevent GCC and Clang from
+     * autovectorizing the XXH32 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling SSE4.1.
+     *
+     * The reason we want to avoid vectorization is because despite working on
+     * 4 integers at a time, there are multiple factors slowing XXH32 down on
+     * SSE4:
+     * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
+     *   newer chips!) making it slightly slower to multiply four integers at
+     *   once compared to four integers independently. Even when pmulld was
+     *   fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
+     *   just to multiply unless doing a long operation.
+     *
+     * - Four instructions are required to rotate,
+     *      movqda tmp,  v // not required with VEX encoding
+     *      pslld  tmp, 13 // tmp <<= 13
+     *      psrld  v,   19 // x >>= 19
+     *      por    v,  tmp // x |= tmp
+     *   compared to one for scalar:
+     *      roll   v, 13    // reliably fast across the board
+     *      shldl  v, v, 13 // Sandy Bridge and later prefer this for some reason
+     *
+     * - Instruction level parallelism is actually more beneficial here because
+     *   the SIMD actually serializes this operation: While v1 is rotating, v2
+     *   can load data, while v3 can multiply. SSE forces them to operate
+     *   together.
+     *
+     * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
+     * the loop. NEON is only faster on the A53, and with the newer cores, it is less
+     * than half the speed.
+     *
+     * Additionally, this is used on WASM SIMD128 because it JITs to the same
+     * SIMD instructions and has the same issue.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
+    return acc;
+}
+
+/*!
+ * @internal
+ * @brief Mixes all bits to finalize the hash.
+ *
+ * The final mix ensures that all input bits have a chance to impact any bit in
+ * the output digest, resulting in an unbiased distribution.
+ *
+ * @param hash The hash to avalanche.
+ * @return The avalanched hash.
+ */
+static xxh_u32 XXH32_avalanche(xxh_u32 hash)
+{
+    hash ^= hash >> 15;
+    hash *= XXH_PRIME32_2;
+    hash ^= hash >> 13;
+    hash *= XXH_PRIME32_3;
+    hash ^= hash >> 16;
+    return hash;
+}
+
+#define XXH_get32bits(p) XXH_readLE32_align(p, align)
+
+/*!
+ * @internal
+ * @brief Sets up the initial accumulator state for XXH32().
+ */
+XXH_FORCE_INLINE void
+XXH32_initAccs(xxh_u32 *acc, xxh_u32 seed)
+{
+    XXH_ASSERT(acc != NULL);
+    acc[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+    acc[1] = seed + XXH_PRIME32_2;
+    acc[2] = seed + 0;
+    acc[3] = seed - XXH_PRIME32_1;
+}
+
+/*!
+ * @internal
+ * @brief Consumes a block of data for XXH32().
+ *
+ * @return the end input pointer.
+ */
+XXH_FORCE_INLINE const xxh_u8 *
+XXH32_consumeLong(
+    xxh_u32 *XXH_RESTRICT acc,
+    xxh_u8 const *XXH_RESTRICT input,
+    size_t len,
+    XXH_alignment align
+)
+{
+    const xxh_u8* const bEnd = input + len;
+    const xxh_u8* const limit = bEnd - 15;
+    XXH_ASSERT(acc != NULL);
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(len >= 16);
+    do {
+        acc[0] = XXH32_round(acc[0], XXH_get32bits(input)); input += 4;
+        acc[1] = XXH32_round(acc[1], XXH_get32bits(input)); input += 4;
+        acc[2] = XXH32_round(acc[2], XXH_get32bits(input)); input += 4;
+        acc[3] = XXH32_round(acc[3], XXH_get32bits(input)); input += 4;
+    } while (input < limit);
+
+    return input;
+}
+
+/*!
+ * @internal
+ * @brief Merges the accumulator lanes together for XXH32()
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u32
+XXH32_mergeAccs(const xxh_u32 *acc)
+{
+    XXH_ASSERT(acc != NULL);
+    return XXH_rotl32(acc[0], 1)  + XXH_rotl32(acc[1], 7)
+         + XXH_rotl32(acc[2], 12) + XXH_rotl32(acc[3], 18);
+}
+
+/*!
+ * @internal
+ * @brief Processes the last 0-15 bytes of @p ptr.
+ *
+ * There may be up to 15 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 16.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash.
+ * @see XXH64_finalize().
+ */
+static XXH_PUREF xxh_u32
+XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+#define XXH_PROCESS1 do {                             \
+    hash += (*ptr++) * XXH_PRIME32_5;                 \
+    hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1;      \
+} while (0)
+
+#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);
+
+    /* Compact rerolled version; generally faster */
+    if (!XXH32_ENDJMP) {
+        len &= 15;
+        while (len >= 4) {
+            XXH_PROCESS4;
+            len -= 4;
+        }
+        while (len > 0) {
+            XXH_PROCESS1;
+            --len;
+        }
+        return XXH32_avalanche(hash);
+    } else {
+         switch(len&15) /* or switch(bEnd - p) */ {
+           case 12:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 8:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 4:       XXH_PROCESS4;
+                         return XXH32_avalanche(hash);
+
+           case 13:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 9:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 5:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(hash);
+
+           case 14:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 10:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 6:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(hash);
+
+           case 15:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 11:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 7:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 3:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 2:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 1:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 0:       return XXH32_avalanche(hash);
+        }
+        XXH_ASSERT(0);
+        return hash;   /* reaching this point is deemed impossible */
+    }
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1 XXH_PROCESS1
+#  define PROCESS4 XXH_PROCESS4
+#else
+#  undef XXH_PROCESS1
+#  undef XXH_PROCESS4
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH32().
+ *
+ * @param input , len , seed Directly passed from @ref XXH32().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u32
+XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+{
+    xxh_u32 h32;
+
+    if (input==NULL) XXH_ASSERT(len == 0);
+
+    if (len>=16) {
+        xxh_u32 acc[4];
+        XXH32_initAccs(acc, seed);
+
+        input = XXH32_consumeLong(acc, input, len, align);
+
+        h32 = XXH32_mergeAccs(acc);
+    } else {
+        h32  = seed + XXH_PRIME32_5;
+    }
+
+    h32 += (xxh_u32)len;
+
+    return XXH32_finalize(h32, input, len&15, align);
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
+{
+#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);
+    XXH32_update(&state, (const xxh_u8*)input, len);
+    return XXH32_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
+            return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+#endif
+}
+
+
+
+/*******   Hash streaming   *******/
+#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 */
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @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 */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
+{
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    XXH32_initAccs(statePtr->acc, seed);
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH32_update(XXH32_state_t* state, const void* input, size_t len)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    state->total_len_32 += (XXH32_hash_t)len;
+    state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
+
+    XXH_ASSERT(state->bufferedSize < sizeof(state->buffer));
+    if (len < sizeof(state->buffer) - state->bufferedSize)  {   /* fill in tmp buffer */
+        XXH_memcpy(state->buffer + state->bufferedSize, input, len);
+        state->bufferedSize += (XXH32_hash_t)len;
+        return XXH_OK;
+    }
+
+    {   const xxh_u8* xinput = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = xinput + len;
+
+        if (state->bufferedSize) {   /* non-empty buffer: complete first */
+            XXH_memcpy(state->buffer + state->bufferedSize, xinput, sizeof(state->buffer) - state->bufferedSize);
+            xinput += sizeof(state->buffer) - state->bufferedSize;
+            /* then process one round */
+            (void)XXH32_consumeLong(state->acc, state->buffer, sizeof(state->buffer), XXH_aligned);
+            state->bufferedSize = 0;
+        }
+
+        XXH_ASSERT(xinput <= bEnd);
+        if ((size_t)(bEnd - xinput) >= sizeof(state->buffer)) {
+            /* Process the remaining data */
+            xinput = XXH32_consumeLong(state->acc, xinput, (size_t)(bEnd - xinput), XXH_unaligned);
+        }
+
+        if (xinput < bEnd) {
+            /* Copy the leftover to the tmp buffer */
+            XXH_memcpy(state->buffer, xinput, (size_t)(bEnd-xinput));
+            state->bufferedSize = (unsigned)(bEnd-xinput);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
+{
+    xxh_u32 h32;
+
+    if (state->large_len) {
+        h32 = XXH32_mergeAccs(state->acc);
+    } else {
+        h32 = state->acc[2] /* == seed */ + XXH_PRIME32_5;
+    }
+
+    h32 += state->total_len_32;
+
+    return XXH32_finalize(h32, state->buffer, state->bufferedSize, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/*******   Canonical representation   *******/
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
+    XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+{
+    return XXH_readBE32(src);
+}
+
+
+#ifndef XXH_NO_LONG_LONG
+
+/* *******************************************************************
+*  64-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @ingroup impl
+ * @{
+ */
+/*******   Memory access   *******/
+
+typedef XXH64_hash_t xxh_u64;
+
+#ifdef XXH_OLD_NAMES
+#  define U64 xxh_u64
+#endif
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE64 and XXH_readBE64.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    return *(const xxh_u64*) memPtr;
+}
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __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 __attribute__((__aligned__(1))) xxh_u64 xxh_unalign64;
+    return *((const xxh_unalign64*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    xxh_u64 val;
+    XXH_memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap64 _byteswap_uint64
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap64 __builtin_bswap64
+#else
+static xxh_u64 XXH_swap64(xxh_u64 x)
+{
+    return  ((x << 56) & 0xff00000000000000ULL) |
+            ((x << 40) & 0x00ff000000000000ULL) |
+            ((x << 24) & 0x0000ff0000000000ULL) |
+            ((x << 8)  & 0x000000ff00000000ULL) |
+            ((x >> 8)  & 0x00000000ff000000ULL) |
+            ((x >> 24) & 0x0000000000ff0000ULL) |
+            ((x >> 40) & 0x000000000000ff00ULL) |
+            ((x >> 56) & 0x00000000000000ffULL);
+}
+#endif
+
+
+/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u64)bytePtr[1] << 8)
+         | ((xxh_u64)bytePtr[2] << 16)
+         | ((xxh_u64)bytePtr[3] << 24)
+         | ((xxh_u64)bytePtr[4] << 32)
+         | ((xxh_u64)bytePtr[5] << 40)
+         | ((xxh_u64)bytePtr[6] << 48)
+         | ((xxh_u64)bytePtr[7] << 56);
+}
+
+XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[7]
+         | ((xxh_u64)bytePtr[6] << 8)
+         | ((xxh_u64)bytePtr[5] << 16)
+         | ((xxh_u64)bytePtr[4] << 24)
+         | ((xxh_u64)bytePtr[3] << 32)
+         | ((xxh_u64)bytePtr[2] << 40)
+         | ((xxh_u64)bytePtr[1] << 48)
+         | ((xxh_u64)bytePtr[0] << 56);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+}
+
+static xxh_u64 XXH_readBE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u64
+XXH_readLE64_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned)
+        return XXH_readLE64(ptr);
+    else
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
+}
+
+
+/*******   xxh64   *******/
+/*!
+ * @}
+ * @defgroup XXH64_impl XXH64 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH64 implementation.
+ * @{
+ */
+/* #define rather that static const, to be used as initializers */
+#define XXH_PRIME64_1  0x9E3779B185EBCA87ULL  /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */
+#define XXH_PRIME64_2  0xC2B2AE3D27D4EB4FULL  /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */
+#define XXH_PRIME64_3  0x165667B19E3779F9ULL  /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */
+#define XXH_PRIME64_4  0x85EBCA77C2B2AE63ULL  /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */
+#define XXH_PRIME64_5  0x27D4EB2F165667C5ULL  /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME64_1 XXH_PRIME64_1
+#  define PRIME64_2 XXH_PRIME64_2
+#  define PRIME64_3 XXH_PRIME64_3
+#  define PRIME64_4 XXH_PRIME64_4
+#  define PRIME64_5 XXH_PRIME64_5
+#endif
+
+/*! @copydoc XXH32_round */
+static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+{
+    acc += input * XXH_PRIME64_2;
+    acc  = XXH_rotl64(acc, 31);
+    acc *= XXH_PRIME64_1;
+#if (defined(__AVX512F__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * DISABLE AUTOVECTORIZATION:
+     * A compiler fence is used to prevent GCC and Clang from
+     * autovectorizing the XXH64 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling AVX512.
+     *
+     * Autovectorization of XXH64 tends to be detrimental,
+     * though the exact outcome may change depending on exact cpu and compiler version.
+     * For information, it has been reported as detrimental for Skylake-X,
+     * but possibly beneficial for Zen4.
+     *
+     * The default is to disable auto-vectorization,
+     * but you can select to enable it instead using `XXH_ENABLE_AUTOVECTORIZE` build variable.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
+    return acc;
+}
+
+static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+{
+    val  = XXH64_round(0, val);
+    acc ^= val;
+    acc  = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
+    return acc;
+}
+
+/*! @copydoc XXH32_avalanche */
+static xxh_u64 XXH64_avalanche(xxh_u64 hash)
+{
+    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)
+
+/*!
+ * @internal
+ * @brief Sets up the initial accumulator state for XXH64().
+ */
+XXH_FORCE_INLINE void
+XXH64_initAccs(xxh_u64 *acc, xxh_u64 seed)
+{
+    XXH_ASSERT(acc != NULL);
+    acc[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+    acc[1] = seed + XXH_PRIME64_2;
+    acc[2] = seed + 0;
+    acc[3] = seed - XXH_PRIME64_1;
+}
+
+/*!
+ * @internal
+ * @brief Consumes a block of data for XXH64().
+ *
+ * @return the end input pointer.
+ */
+XXH_FORCE_INLINE const xxh_u8 *
+XXH64_consumeLong(
+    xxh_u64 *XXH_RESTRICT acc,
+    xxh_u8 const *XXH_RESTRICT input,
+    size_t len,
+    XXH_alignment align
+)
+{
+    const xxh_u8* const bEnd = input + len;
+    const xxh_u8* const limit = bEnd - 31;
+    XXH_ASSERT(acc != NULL);
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(len >= 32);
+    do {
+        /* reroll on 32-bit */
+        if (sizeof(void *) < sizeof(xxh_u64)) {
+            size_t i;
+            for (i = 0; i < 4; i++) {
+                acc[i] = XXH64_round(acc[i], XXH_get64bits(input));
+                input += 8;
+            }
+        } else {
+            acc[0] = XXH64_round(acc[0], XXH_get64bits(input)); input += 8;
+            acc[1] = XXH64_round(acc[1], XXH_get64bits(input)); input += 8;
+            acc[2] = XXH64_round(acc[2], XXH_get64bits(input)); input += 8;
+            acc[3] = XXH64_round(acc[3], XXH_get64bits(input)); input += 8;
+        }
+    } while (input < limit);
+
+    return input;
+}
+
+/*!
+ * @internal
+ * @brief Merges the accumulator lanes together for XXH64()
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u64
+XXH64_mergeAccs(const xxh_u64 *acc)
+{
+    XXH_ASSERT(acc != NULL);
+    {
+        xxh_u64 h64 = XXH_rotl64(acc[0], 1) + XXH_rotl64(acc[1], 7)
+                    + XXH_rotl64(acc[2], 12) + XXH_rotl64(acc[3], 18);
+        /* reroll on 32-bit */
+        if (sizeof(void *) < sizeof(xxh_u64)) {
+            size_t i;
+            for (i = 0; i < 4; i++) {
+                h64 = XXH64_mergeRound(h64, acc[i]);
+            }
+        } else {
+            h64 = XXH64_mergeRound(h64, acc[0]);
+            h64 = XXH64_mergeRound(h64, acc[1]);
+            h64 = XXH64_mergeRound(h64, acc[2]);
+            h64 = XXH64_mergeRound(h64, acc[3]);
+        }
+        return h64;
+    }
+}
+
+/*!
+ * @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().
+ */
+XXH_STATIC XXH_PUREF xxh_u64
+XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+    if (ptr==NULL) XXH_ASSERT(len == 0);
+    len &= 31;
+    while (len >= 8) {
+        xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
+        ptr += 8;
+        hash ^= k1;
+        hash  = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+        len -= 8;
+    }
+    if (len >= 4) {
+        hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+        ptr += 4;
+        hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+        len -= 4;
+    }
+    while (len > 0) {
+        hash ^= (*ptr++) * XXH_PRIME64_5;
+        hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
+        --len;
+    }
+    return  XXH64_avalanche(hash);
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1_64 XXH_PROCESS1_64
+#  define PROCESS4_64 XXH_PROCESS4_64
+#  define PROCESS8_64 XXH_PROCESS8_64
+#else
+#  undef XXH_PROCESS1_64
+#  undef XXH_PROCESS4_64
+#  undef XXH_PROCESS8_64
+#endif
+
+/*!
+ * @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;
+    if (input==NULL) XXH_ASSERT(len == 0);
+
+    if (len>=32) {  /* Process a large block of data */
+        xxh_u64 acc[4];
+        XXH64_initAccs(acc, seed);
+
+        input = XXH64_consumeLong(acc, input, len, align);
+
+        h64 = XXH64_mergeAccs(acc);
+    } else {
+        h64  = seed + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) len;
+
+    return XXH64_finalize(h64, input, len, align);
+}
+
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH64_state_t state;
+    XXH64_reset(&state, seed);
+    XXH64_update(&state, (const xxh_u8*)input, len);
+    return XXH64_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
+            return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+
+#endif
+}
+
+/*******   Hash Streaming   *******/
+#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 */
+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(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(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
+{
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    XXH64_initAccs(statePtr->acc, seed);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    state->total_len += len;
+
+    XXH_ASSERT(state->bufferedSize <= sizeof(state->buffer));
+    if (len < sizeof(state->buffer) - state->bufferedSize)  {   /* fill in tmp buffer */
+        XXH_memcpy(state->buffer + state->bufferedSize, input, len);
+        state->bufferedSize += (XXH32_hash_t)len;
+        return XXH_OK;
+    }
+
+    {   const xxh_u8* xinput = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = xinput + len;
+
+        if (state->bufferedSize) {   /* non-empty buffer => complete first */
+            XXH_memcpy(state->buffer + state->bufferedSize, xinput, sizeof(state->buffer) - state->bufferedSize);
+            xinput += sizeof(state->buffer) - state->bufferedSize;
+            /* and process one round */
+            (void)XXH64_consumeLong(state->acc, state->buffer, sizeof(state->buffer), XXH_aligned);
+            state->bufferedSize = 0;
+        }
+
+        XXH_ASSERT(xinput <= bEnd);
+        if ((size_t)(bEnd - xinput) >= sizeof(state->buffer)) {
+            /* Process the remaining data */
+            xinput = XXH64_consumeLong(state->acc, xinput, (size_t)(bEnd - xinput), XXH_unaligned);
+        }
+
+        if (xinput < bEnd) {
+            /* Copy the leftover to the tmp buffer */
+            XXH_memcpy(state->buffer, xinput, (size_t)(bEnd-xinput));
+            state->bufferedSize = (unsigned)(bEnd-xinput);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
+{
+    xxh_u64 h64;
+
+    if (state->total_len >= 32) {
+        h64 = XXH64_mergeAccs(state->acc);
+    } else {
+        h64  = state->acc[2] /*seed*/ + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) state->total_len;
+
+    return XXH64_finalize(h64, state->buffer, (size_t)state->total_len, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/******* Canonical representation   *******/
+
+/*! @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(XXH_NOESCAPE const XXH64_canonical_t* src)
+{
+    return XXH_readBE64(src);
+}
+
+#ifndef XXH_NO_XXH3
+
+/* *********************************************************************
+*  XXH3
+*  New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+/*!
+ * @}
+ * @defgroup XXH3_impl XXH3 implementation
+ * @ingroup impl
+ * @{
+ */
+
+/* ===   Compiler specifics   === */
+
+
+#if (defined(__GNUC__) && (__GNUC__ >= 3))  \
+  || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
+  || defined(__clang__)
+#    define XXH_likely(x) __builtin_expect(x, 1)
+#    define XXH_unlikely(x) __builtin_expect(x, 0)
+#else
+#    define XXH_likely(x) (x)
+#    define XXH_unlikely(x) (x)
+#endif
+
+#ifndef XXH_HAS_INCLUDE
+#  ifdef __has_include
+/*
+ * Not defined as XXH_HAS_INCLUDE(x) (function-like) because
+ * this causes segfaults in Apple Clang 4.2 (on Mac OS X 10.7 Lion)
+ */
+#    define XXH_HAS_INCLUDE __has_include
+#  else
+#    define XXH_HAS_INCLUDE(x) 0
+#  endif
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#  if defined(__ARM_FEATURE_SVE)
+#    include <arm_sve.h>
+#  endif
+#  if defined(__ARM_NEON__) || defined(__ARM_NEON) \
+   || (defined(_M_ARM) && _M_ARM >= 7) \
+   || defined(_M_ARM64) || defined(_M_ARM64EC) \
+   || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* WASM SIMD128 via SIMDe */
+#    define inline __inline__  /* circumvent a clang bug */
+#    include <arm_neon.h>
+#    undef inline
+#  elif defined(__AVX2__)
+#    include <immintrin.h>
+#  elif defined(__SSE2__)
+#    include <emmintrin.h>
+#  elif defined(__loongarch_sx)
+#    include <lsxintrin.h>
+#  endif
+#endif
+
+#if defined(_MSC_VER)
+#  include <intrin.h>
+#endif
+
+/*
+ * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
+ * remaining a true 64-bit/128-bit hash function.
+ *
+ * This is done by prioritizing a subset of 64-bit operations that can be
+ * emulated without too many steps on the average 32-bit machine.
+ *
+ * For example, these two lines seem similar, and run equally fast on 64-bit:
+ *
+ *   xxh_u64 x;
+ *   x ^= (x >> 47); // good
+ *   x ^= (x >> 13); // bad
+ *
+ * However, to a 32-bit machine, there is a major difference.
+ *
+ * x ^= (x >> 47) looks like this:
+ *
+ *   x.lo ^= (x.hi >> (47 - 32));
+ *
+ * while x ^= (x >> 13) looks like this:
+ *
+ *   // note: funnel shifts are not usually cheap.
+ *   x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
+ *   x.hi ^= (x.hi >> 13);
+ *
+ * The first one is significantly faster than the second, simply because the
+ * shift is larger than 32. This means:
+ *  - All the bits we need are in the upper 32 bits, so we can ignore the lower
+ *    32 bits in the shift.
+ *  - The shift result will always fit in the lower 32 bits, and therefore,
+ *    we can ignore the upper 32 bits in the xor.
+ *
+ * Thanks to this optimization, XXH3 only requires these features to be efficient:
+ *
+ *  - Usable unaligned access
+ *  - A 32-bit or 64-bit ALU
+ *      - If 32-bit, a decent ADC instruction
+ *  - A 32 or 64-bit multiply with a 64-bit result
+ *  - For the 128-bit variant, a decent byteswap helps short inputs.
+ *
+ * The first two are already required by XXH32, and almost all 32-bit and 64-bit
+ * platforms which can run XXH32 can run XXH3 efficiently.
+ *
+ * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
+ * notable exception.
+ *
+ * First of all, Thumb-1 lacks support for the UMULL instruction which
+ * performs the important long multiply. This means numerous __aeabi_lmul
+ * calls.
+ *
+ * Second of all, the 8 functional registers are just not enough.
+ * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
+ * Lo registers, and this shuffling results in thousands more MOVs than A32.
+ *
+ * A32 and T32 don't have this limitation. They can access all 14 registers,
+ * do a 32->64 multiply with UMULL, and the flexible operand allowing free
+ * shifts is helpful, too.
+ *
+ * Therefore, we do a quick sanity check.
+ *
+ * If compiling Thumb-1 for a target which supports ARM instructions, we will
+ * emit a warning, as it is not a "sane" platform to compile for.
+ *
+ * Usually, if this happens, it is because of an accident and you probably need
+ * to specify -march, as you likely meant to compile for a newer architecture.
+ *
+ * Credit: large sections of the vectorial and asm source code paths
+ *         have been contributed by @easyaspi314
+ */
+#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
+#   warning "XXH3 is highly inefficient without ARM or Thumb-2."
+#endif
+
+/* ==========================================
+ * Vectorization detection
+ * ========================================== */
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @ingroup tuning
+ * @brief Overrides the vectorization implementation chosen for XXH3.
+ *
+ * Can be defined to 0 to disable SIMD or any of the values mentioned in
+ * @ref XXH_VECTOR_TYPE.
+ *
+ * If this is not defined, it uses predefined macros to determine the best
+ * implementation.
+ */
+#  define XXH_VECTOR XXH_SCALAR
+/*!
+ * @ingroup tuning
+ * @brief Selects the minimum alignment for XXH3's accumulators.
+ *
+ * When using SIMD, this should match the alignment required for said vector
+ * type, so, for example, 32 for AVX2.
+ *
+ * Default: Auto detected.
+ */
+#  define XXH_ACC_ALIGN 8
+#endif
+
+/* Actual definition */
+#ifndef XXH_DOXYGEN
+#endif
+
+#ifndef XXH_VECTOR    /* can be defined on command line */
+#  if defined(__ARM_FEATURE_SVE)
+#    define XXH_VECTOR XXH_SVE
+#  elif ( \
+        defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
+     || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
+     || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* wasm simd128 via SIMDe */ \
+   ) && ( \
+        defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
+    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
+   )
+#    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 */
+#    define XXH_VECTOR XXH_VSX
+#  elif defined(__loongarch_sx)
+#    define XXH_VECTOR XXH_LSX
+#  else
+#    define XXH_VECTOR XXH_SCALAR
+#  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.
+ */
+#ifndef XXH_ACC_ALIGN
+#  if defined(XXH_X86DISPATCH)
+#     define XXH_ACC_ALIGN 64  /* for compatibility with avx512 */
+#  elif XXH_VECTOR == XXH_SCALAR  /* scalar */
+#     define XXH_ACC_ALIGN 8
+#  elif XXH_VECTOR == XXH_SSE2  /* sse2 */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX2  /* avx2 */
+#     define XXH_ACC_ALIGN 32
+#  elif XXH_VECTOR == XXH_NEON  /* neon */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_VSX   /* vsx */
+#     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
+#  elif XXH_VECTOR == XXH_LSX   /* lsx */
+#     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.
+ *
+ * However, when targeting AVX2, it is overzealous in its unrolling resulting
+ * in code roughly 3/4 the speed of Clang.
+ *
+ * There are other issues, such as GCC splitting _mm256_loadu_si256 into
+ * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
+ * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
+ *
+ * That is why when compiling the AVX2 version, it is recommended to use either
+ *   -O2 -mavx2 -march=haswell
+ * or
+ *   -O2 -mavx2 -mno-avx256-split-unaligned-load
+ * for decent performance, or to use Clang instead.
+ *
+ * Fortunately, we can control the first one with a pragma that forces GCC into
+ * -O2, but the other one we can't control without "failed to inline always
+ * inline function due to target mismatch" warnings.
+ */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+#  pragma GCC push_options
+#  pragma GCC optimize("-O2")
+#endif
+
+#if XXH_VECTOR == XXH_NEON
+
+/*
+ * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3
+ * optimizes out the entire hashLong loop because of the aliasing violation.
+ *
+ * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
+ * so the only option is to mark it as aliasing.
+ */
+typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
+
+/*!
+ * @internal
+ * @brief `vld1q_u64` but faster and alignment-safe.
+ *
+ * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only
+ * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86).
+ *
+ * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it
+ * prohibits load-store optimizations. Therefore, a direct dereference is used.
+ *
+ * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe
+ * unaligned load.
+ */
+#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.
+ *
+ * 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
+ *
+ * This can be set to 2, 4, 6, or 8.
+ *
+ * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
+ *
+ * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those
+ * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU
+ * bandwidth.
+ *
+ * This is even more noticeable on the more advanced cores like the Cortex-A76 which
+ * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
+ *
+ * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes
+ * and 2 scalar lanes, which is chosen by default.
+ *
+ * This does not apply to Apple processors or 32-bit processors, which run better with
+ * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes.
+ *
+ * This change benefits CPUs with large micro-op buffers without negatively affecting
+ * most other CPUs:
+ *
+ *  | Chipset               | Dispatch type       | NEON only | 6:2 hybrid | Diff. |
+ *  |:----------------------|:--------------------|----------:|-----------:|------:|
+ *  | Snapdragon 730 (A76)  | 2 NEON/8 micro-ops  |  8.8 GB/s |  10.1 GB/s |  ~16% |
+ *  | Snapdragon 835 (A73)  | 2 NEON/3 micro-ops  |  5.1 GB/s |   5.3 GB/s |   ~5% |
+ *  | Marvell PXA1928 (A53) | In-order dual-issue |  1.9 GB/s |   1.9 GB/s |    0% |
+ *  | Apple M1              | 4 NEON/8 micro-ops  | 37.3 GB/s |  36.1 GB/s |  ~-3% |
+ *
+ * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
+ *
+ * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning
+ * it effectively becomes worse 4.
+ *
+ * @see XXH3_accumulate_512_neon()
+ */
+# ifndef XXH3_NEON_LANES
+#  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
+   && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
+#   define XXH3_NEON_LANES 6
+#  else
+#   define XXH3_NEON_LANES XXH_ACC_NB
+#  endif
+# endif
+#endif  /* XXH_VECTOR == XXH_NEON */
+
+/*
+ * VSX and Z Vector helpers.
+ *
+ * This is very messy, and any pull requests to clean this up are welcome.
+ *
+ * There are a lot of problems with supporting VSX and s390x, due to
+ * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
+ */
+#if XXH_VECTOR == XXH_VSX
+/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`,
+ * and `pixel`. This is a problem for obvious reasons.
+ *
+ * These keywords are unnecessary; the spec literally says they are
+ * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd
+ * after including the header.
+ *
+ * We use pragma push_macro/pop_macro to keep the namespace clean. */
+#  pragma push_macro("bool")
+#  pragma push_macro("vector")
+#  pragma push_macro("pixel")
+/* silence potential macro redefined warnings */
+#  undef bool
+#  undef vector
+#  undef pixel
+
+#  if defined(__s390x__)
+#    include <s390intrin.h>
+#  else
+#    include <altivec.h>
+#  endif
+
+/* Restore the original macro values, if applicable. */
+#  pragma pop_macro("pixel")
+#  pragma pop_macro("vector")
+#  pragma pop_macro("bool")
+
+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__)
+#    define XXH_VSX_BE 1
+#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
+#    warning "-maltivec=be is not recommended. Please use native endianness."
+#    define XXH_VSX_BE 1
+#  else
+#    define XXH_VSX_BE 0
+#  endif
+# endif /* !defined(XXH_VSX_BE) */
+
+# if XXH_VSX_BE
+#  if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
+#    define XXH_vec_revb vec_revb
+#  else
+/*!
+ * A polyfill for POWER9's vec_revb().
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
+{
+    xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
+                                  0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+    return vec_perm(val, val, vByteSwap);
+}
+#  endif
+# endif /* XXH_VSX_BE */
+
+/*!
+ * Performs an unaligned vector load and byte swaps it on big endian.
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
+{
+    xxh_u64x2 ret;
+    XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2));
+# if XXH_VSX_BE
+    ret = XXH_vec_revb(ret);
+# endif
+    return ret;
+}
+
+/*
+ * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
+ *
+ * These intrinsics weren't added until GCC 8, despite existing for a while,
+ * and they are endian dependent. Also, their meaning swap depending on version.
+ * */
+# if defined(__s390x__)
+ /* s390x is always big endian, no issue on this platform */
+#  define XXH_vec_mulo vec_mulo
+#  define XXH_vec_mule vec_mule
+# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__)
+/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
+ /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */
+#  define XXH_vec_mulo __builtin_altivec_vmulouw
+#  define XXH_vec_mule __builtin_altivec_vmuleuw
+# else
+/* gcc needs inline assembly */
+/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+# endif /* XXH_vec_mulo, XXH_vec_mule */
+#endif /* XXH_VECTOR == XXH_VSX */
+
+#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 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) ) )
+#    define XXH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
+#  else
+#    define XXH_PREFETCH(ptr) (void)(ptr)  /* disabled */
+#  endif
+#endif  /* XXH_NO_PREFETCH */
+
+
+/* ==========================================
+ * XXH3 default settings
+ * ========================================== */
+
+#define XXH_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
+
+#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
+#  error "default keyset is not large enough"
+#endif
+
+/*! Pseudorandom secret taken directly from FARSH. */
+XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
+    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
+    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
+    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
+    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
+    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
+    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
+    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
+    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
+    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
+    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
+    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
+    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
+};
+
+static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL;  /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */
+static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL;  /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */
+
+#ifdef XXH_OLD_NAMES
+#  define kSecret XXH3_kSecret
+#endif
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Calculates a 32-bit to 64-bit long multiply.
+ *
+ * Implemented as a macro.
+ *
+ * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't
+ * need to (but it shouldn't need to anyways, it is about 7 instructions to do
+ * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we
+ * use that instead of the normal method.
+ *
+ * If you are compiling for platforms like Thumb-1 and don't have a better option,
+ * you may also want to write your own long multiply routine here.
+ *
+ * @param x, y Numbers to be multiplied
+ * @return 64-bit product of the low 32 bits of @p x and @p y.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64(xxh_u64 x, xxh_u64 y)
+{
+   return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
+}
+#elif defined(_MSC_VER) && defined(_M_IX86)
+#    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
+#else
+/*
+ * Downcast + upcast is usually better than masking on older compilers like
+ * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
+ *
+ * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
+ * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
+ */
+#    define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
+#endif
+
+/*!
+ * @brief Calculates a 64->128-bit long multiply.
+ *
+ * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
+ * version.
+ *
+ * @param lhs , rhs The 64-bit integers to be multiplied
+ * @return The 128-bit result represented in an @ref XXH128_hash_t.
+ */
+static XXH128_hash_t
+XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
+{
+    /*
+     * GCC/Clang __uint128_t method.
+     *
+     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
+     * This is usually the best way as it usually uses a native long 64-bit
+     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
+     *
+     * Usually.
+     *
+     * Despite being a 32-bit platform, Clang (and emscripten) define this type
+     * despite not having the arithmetic for it. This results in a laggy
+     * compiler builtin call which calculates a full 128-bit multiply.
+     * In that case it is best to use the portable one.
+     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
+     */
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
+    && defined(__SIZEOF_INT128__) \
+    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
+
+    __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
+    XXH128_hash_t r128;
+    r128.low64  = (xxh_u64)(product);
+    r128.high64 = (xxh_u64)(product >> 64);
+    return r128;
+
+    /*
+     * MSVC for x64's _umul128 method.
+     *
+     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
+     *
+     * This compiles to single operand MUL on x64.
+     */
+#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(_umul128)
+#endif
+    xxh_u64 product_high;
+    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
+    XXH128_hash_t r128;
+    r128.low64  = product_low;
+    r128.high64 = product_high;
+    return r128;
+
+    /*
+     * MSVC for ARM64's __umulh method.
+     *
+     * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
+     */
+#elif defined(_M_ARM64) || defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(__umulh)
+#endif
+    XXH128_hash_t r128;
+    r128.low64  = lhs * rhs;
+    r128.high64 = __umulh(lhs, rhs);
+    return r128;
+
+#else
+    /*
+     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
+     *
+     * This is a fast and simple grade school multiply, which is shown below
+     * with base 10 arithmetic instead of base 0x100000000.
+     *
+     *           9 3 // D2 lhs = 93
+     *         x 7 5 // D2 rhs = 75
+     *     ----------
+     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
+     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
+     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
+     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
+     *     ---------
+     *         2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
+     *     + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
+     *     ---------
+     *       6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
+     *
+     * The reasons for adding the products like this are:
+     *  1. It avoids manual carry tracking. Just like how
+     *     (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
+     *     This avoids a lot of complexity.
+     *
+     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
+     *     instruction available in ARM's Digital Signal Processing extension
+     *     in 32-bit ARMv6 and later, which is shown below:
+     *
+     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
+     *         {
+     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
+     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
+     *             *RdHi = (xxh_u32)(product >> 32);
+     *         }
+     *
+     *     This instruction was designed for efficient long multiplication, and
+     *     allows this to be calculated in only 4 instructions at speeds
+     *     comparable to some 64-bit ALUs.
+     *
+     *  3. It isn't terrible on other platforms. Usually this will be a couple
+     *     of 32-bit ADD/ADCs.
+     */
+
+    /* First calculate all of the cross products. */
+    xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
+    xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
+    xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
+    xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32,        rhs >> 32);
+
+    /* Now add the products together. These will never overflow. */
+    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
+    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
+    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
+
+    XXH128_hash_t r128;
+    r128.low64  = lower;
+    r128.high64 = upper;
+    return r128;
+#endif
+}
+
+/*!
+ * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it.
+ *
+ * The reason for the separate function is to prevent passing too many structs
+ * around by value. This will hopefully inline the multiply, but we don't force it.
+ *
+ * @param lhs , rhs The 64-bit integers to multiply
+ * @return The low 64 bits of the product XOR'd by the high 64 bits.
+ * @see XXH_mult64to128()
+ */
+static xxh_u64
+XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
+{
+    XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
+    return product.low64 ^ product.high64;
+}
+
+/*! Seems to produce slightly better code on GCC for some reason. */
+XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
+{
+    XXH_ASSERT(0 <= shift && shift < 64);
+    return v64 ^ (v64 >> shift);
+}
+
+/*
+ * This is a fast avalanche stage,
+ * suitable when input bits are already partially mixed
+ */
+static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
+{
+    h64 = XXH_xorshift64(h64, 37);
+    h64 *= PRIME_MX1;
+    h64 = XXH_xorshift64(h64, 32);
+    return h64;
+}
+
+/*
+ * This is a stronger avalanche,
+ * inspired by Pelle Evensen's rrmxmx
+ * preferable when input has not been previously mixed
+ */
+static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
+{
+    /* this mix is inspired by Pelle Evensen's rrmxmx */
+    h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
+    h64 *= PRIME_MX2;
+    h64 ^= (h64 >> 35) + len ;
+    h64 *= PRIME_MX2;
+    return XXH_xorshift64(h64, 28);
+}
+
+
+/* ==========================================
+ * Short keys
+ * ==========================================
+ * One of the shortcomings of XXH32 and XXH64 was that their performance was
+ * sub-optimal on short lengths. It used an iterative algorithm which strongly
+ * favored lengths that were a multiple of 4 or 8.
+ *
+ * Instead of iterating over individual inputs, we use a set of single shot
+ * functions which piece together a range of lengths and operate in constant time.
+ *
+ * Additionally, the number of multiplies has been significantly reduced. This
+ * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
+ *
+ * Depending on the platform, this may or may not be faster than XXH32, but it
+ * is almost guaranteed to be faster than XXH64.
+ */
+
+/*
+ * At very short lengths, there isn't enough input to fully hide secrets, or use
+ * the entire secret.
+ *
+ * There is also only a limited amount of mixing we can do before significantly
+ * impacting performance.
+ *
+ * Therefore, we use different sections of the secret and always mix two secret
+ * samples with an XOR. This should have no effect on performance on the
+ * seedless or withSeed variants because everything _should_ be constant folded
+ * by modern compilers.
+ *
+ * The XOR mixing hides individual parts of the secret and increases entropy.
+ *
+ * This adds an extra layer of strength for custom secrets.
+ */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combined = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combined = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combined = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8  const c1 = input[0];
+        xxh_u8  const c2 = input[len >> 1];
+        xxh_u8  const c3 = input[len - 1];
+        xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2  << 24)
+                               | ((xxh_u32)c3 <<  0) | ((xxh_u32)len << 8);
+        xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
+        return XXH64_avalanche(keyed);
+    }
+}
+
+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);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input1 = XXH_readLE32(input);
+        xxh_u32 const input2 = XXH_readLE32(input + len - 4);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
+        xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
+        xxh_u64 const keyed = input64 ^ bitflip;
+        return XXH3_rrmxmx(keyed, len);
+    }
+}
+
+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);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
+        xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
+        xxh_u64 const input_lo = XXH_readLE64(input)           ^ bitflip1;
+        xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
+        xxh_u64 const acc = len
+                          + XXH_swap64(input_lo) + input_hi
+                          + XXH3_mul128_fold64(input_lo, input_hi);
+        return XXH3_avalanche(acc);
+    }
+}
+
+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);
+    {   if (XXH_likely(len >  8)) return XXH3_len_9to16_64b(input, len, secret, seed);
+        if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
+        return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
+    }
+}
+
+/*
+ * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
+ * multiplication by zero, affecting hashes of lengths 17 to 240.
+ *
+ * However, they are very unlikely.
+ *
+ * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
+ * unseeded non-cryptographic hashes, it does not attempt to defend itself
+ * against specially crafted inputs, only random inputs.
+ *
+ * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
+ * cancelling out the secret is taken an arbitrary number of times (addressed
+ * in XXH3_accumulate_512), this collision is very unlikely with random inputs
+ * and/or proper seeding:
+ *
+ * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
+ * function that is only called up to 16 times per hash with up to 240 bytes of
+ * input.
+ *
+ * This is not too bad for a non-cryptographic hash function, especially with
+ * only 64 bit outputs.
+ *
+ * The 128-bit variant (which trades some speed for strength) is NOT affected
+ * by this, although it is always a good idea to use a proper seed if you care
+ * about strength.
+ */
+XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
+                                     const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
+{
+#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__i386__) && defined(__SSE2__)  /* x86 + SSE2 */ \
+  && !defined(XXH_ENABLE_AUTOVECTORIZE)      /* Define to disable like XXH32 hack */
+    /*
+     * UGLY HACK:
+     * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
+     * slower code.
+     *
+     * By forcing seed64 into a register, we disrupt the cost model and
+     * cause it to scalarize. See `XXH32_round()`
+     *
+     * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
+     * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
+     * GCC 9.2, despite both emitting scalar code.
+     *
+     * GCC generates much better scalar code than Clang for the rest of XXH3,
+     * which is why finding a more optimal codepath is an interest.
+     */
+    XXH_COMPILER_GUARD(seed64);
+#endif
+    {   xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64 const input_hi = XXH_readLE64(input+8);
+        return XXH3_mul128_fold64(
+            input_lo ^ (XXH_readLE64(secret)   + seed64),
+            input_hi ^ (XXH_readLE64(secret+8) - seed64)
+        );
+    }
+}
+
+/* For mid range keys, XXH3 uses a Mum-hash variant. */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                     const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                     XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+#if XXH_SIZE_OPT >= 1
+        /* Smaller and cleaner, but slightly slower. */
+        unsigned int i = (unsigned int)(len - 1) / 32;
+        do {
+            acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
+            acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
+        } while (i-- != 0);
+#else
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc += XXH3_mix16B(input+48, secret+96, seed);
+                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
+                }
+                acc += XXH3_mix16B(input+32, secret+64, seed);
+                acc += XXH3_mix16B(input+len-48, secret+80, seed);
+            }
+            acc += XXH3_mix16B(input+16, secret+32, seed);
+            acc += XXH3_mix16B(input+len-32, secret+48, seed);
+        }
+        acc += XXH3_mix16B(input+0, secret+0, seed);
+        acc += XXH3_mix16B(input+len-16, secret+16, seed);
+#endif
+        return XXH3_avalanche(acc);
+    }
+}
+
+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)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    #define XXH3_MIDSIZE_STARTOFFSET 3
+    #define XXH3_MIDSIZE_LASTOFFSET  17
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+        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);
+        }
+        /* 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 */
+        /*
+         * UGLY HACK:
+         * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
+         * In everywhere else, it uses scalar code.
+         *
+         * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
+         * would still be slower than UMAAL (see XXH_mult64to128).
+         *
+         * Unfortunately, Clang doesn't handle the long multiplies properly and
+         * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
+         * scalarized into an ugly mess of VMOV.32 instructions.
+         *
+         * This mess is difficult to avoid without turning autovectorization
+         * off completely, but they are usually relatively minor and/or not
+         * worth it to fix.
+         *
+         * This loop is the easiest to fix, as unlike XXH32, this pragma
+         * _actually works_ because it is a loop vectorization instead of an
+         * SLP vectorization.
+         */
+        #pragma clang loop vectorize(disable)
+#endif
+        for (i=8 ; i < nbRounds; i++) {
+            /*
+             * Prevents clang for unrolling the acc loop and interleaving with this one.
+             */
+            XXH_COMPILER_GUARD(acc);
+            acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+        }
+        return XXH3_avalanche(acc + acc_end);
+    }
+}
+
+
+/* =======     Long Keys     ======= */
+
+#define XXH_STRIPE_LEN 64
+#define XXH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
+#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))
+
+#ifdef XXH_OLD_NAMES
+#  define STRIPE_LEN XXH_STRIPE_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);
+    XXH_memcpy(dst, &v64, sizeof(v64));
+}
+
+/* Several intrinsic functions below are supposed to accept __int64 as argument,
+ * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ .
+ * However, several environments do not define __int64 type,
+ * requiring a workaround.
+ */
+#if !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+    typedef int64_t xxh_i64;
+#else
+    /* the following type must have a width of 64-bit */
+    typedef long long xxh_i64;
+#endif
+
+
+/*
+ * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
+ *
+ * It is a hardened version of UMAC, based off of FARSH's implementation.
+ *
+ * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
+ * implementations, and it is ridiculously fast.
+ *
+ * We harden it by mixing the original input to the accumulators as well as the product.
+ *
+ * This means that in the (relatively likely) case of a multiply by zero, the
+ * original input is preserved.
+ *
+ * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
+ * cross-pollination, as otherwise the upper and lower halves would be
+ * essentially independent.
+ *
+ * This doesn't matter on 64-bit hashes since they all get merged together in
+ * the end, so we skip the extra step.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+#if (XXH_VECTOR == XXH_AVX512) \
+     || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0)
+
+#ifndef XXH_TARGET_AVX512
+# define XXH_TARGET_AVX512  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    __m512i* const xacc = (__m512i *) acc;
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+
+    {
+        /* data_vec    = input[0]; */
+        __m512i const data_vec    = _mm512_loadu_si512   (input);
+        /* key_vec     = secret[0]; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        /* 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_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); */
+        __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2));
+        __m512i const sum       = _mm512_add_epi64(*xacc, data_swap);
+        /* xacc[0] += product; */
+        *xacc = _mm512_add_epi64(product, sum);
+    }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)
+
+/*
+ * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
+ *
+ * Multiplication isn't perfect, as explained by Google in HighwayHash:
+ *
+ *  // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
+ *  // varying degrees. In descending order of goodness, bytes
+ *  // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
+ *  // As expected, the upper and lower bytes are much worse.
+ *
+ * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
+ *
+ * Since our algorithm uses a pseudorandom secret to add some variance into the
+ * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
+ *
+ * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
+ * extraction.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+    {   __m512i* const xacc = (__m512i*) acc;
+        const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
+
+        /* xacc[0] ^= (xacc[0] >> 47) */
+        __m512i const acc_vec     = *xacc;
+        __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
+        /* xacc[0] ^= secret; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        __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_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));
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64);
+    XXH_ASSERT(((size_t)customSecret & 63) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
+        __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;
+        int i;
+        XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 63) == 0);
+        for (i=0; i < nbRounds; ++i) {
+            dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_AVX2) \
+    || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0)
+
+#ifndef XXH_TARGET_AVX2
+# define XXH_TARGET_AVX2  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   __m256i* const xacc    =       (__m256i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
+        const         __m256i* const xinput  = (const __m256i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m256i const data_vec    = _mm256_loadu_si256    (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            /* 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_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); */
+            __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
+            __m256i const sum       = _mm256_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm256_add_epi64(product, sum);
+    }   }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   __m256i* const xacc = (__m256i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+        const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m256i const acc_vec     = xacc[i];
+            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
+            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_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));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0);
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64);
+    (void)(&XXH_writeLE64);
+    XXH_PREFETCH(customSecret);
+    {   __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);
+
+        const __m256i* const src  = (const __m256i*) ((const void*) XXH3_kSecret);
+              __m256i*       dest = (      __m256i*) customSecret;
+
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dest);
+#       endif
+        XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 31) == 0);
+
+        /* GCC -O2 need unroll loop manually */
+        dest[0] = _mm256_add_epi64(_mm256_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);
+    }
+}
+
 #endif
 
-/*!
- *XXH_ACCEPT_NULL_INPUT_POINTER:
- * If the input pointer is NULL, xxHash's default behavior is to dereference it,
- * triggering a segfault.
- * When this macro is enabled, xxHash actively checks the input for a null pointer.
- * If it is, the result for null input pointers is the same as a zero-length input.
- */
-#ifndef XXH_ACCEPT_NULL_INPUT_POINTER   /* can be defined externally */
-#  define XXH_ACCEPT_NULL_INPUT_POINTER 0
+/* x86dispatch always generates SSE2 */
+#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)
+
+#ifndef XXH_TARGET_SSE2
+# define XXH_TARGET_SSE2  /* disable attribute target */
 #endif
 
-/*!
- * XXH_FORCE_ALIGN_CHECK:
- * This is a minor performance trick, only useful with lots of very small keys.
- * It means: check for aligned/unaligned input.
- * The check costs one initial branch per hash;
- * Set it to 0 when the input is guaranteed to be aligned or when alignment
- * doesn't matter for performance.
- *
- * This option does not affect XXH3.
- */
-#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */
-#  if defined(__i386) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64)
-#    define XXH_FORCE_ALIGN_CHECK 0
-#  else
-#    define XXH_FORCE_ALIGN_CHECK 1
-#  endif
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    /* SSE2 is just a half-scale version of the AVX2 version. */
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   __m128i* const xacc    =       (__m128i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xinput  = (const __m128i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m128i const data_vec    = _mm_loadu_si128   (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m128i const product     = _mm_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
+            __m128i const sum       = _mm_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm_add_epi64(product, sum);
+    }   }
+}
+XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   __m128i* const xacc = (__m128i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+        const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m128i const acc_vec     = xacc[i];
+            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
+            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
+            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);
+
+#       if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
+        /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */
+        XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
+        __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
+#       else
+        __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64);
+#       endif
+        int i;
+
+        const void* const src16 = XXH3_kSecret;
+        __m128i* dst16 = (__m128i*) customSecret;
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dst16);
+#       endif
+        XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dst16 & 15) == 0);
+
+        for (i=0; i < nbRounds; ++i) {
+            dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
+    }   }
+}
+
 #endif
 
+#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);
+
 /*!
- * XXH_NO_INLINE_HINTS:
+ * @internal
+ * @brief The bulk processing loop for NEON and WASM SIMD128.
  *
- * By default, xxHash tries to force the compiler to inline almost all internal
- * functions.
+ * 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.
  *
- * This can usually improve performance due to reduced jumping and improved
- * constant folding, but significantly increases the size of the binary which
- * might not be favorable.
+ * @see XXH3_NEON_LANES for configuring this and details about this optimization.
  *
- * Additionally, sometimes the forced inlining can be detrimental to performance,
- * depending on the architecture.
+ * 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.
  *
- * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
- * compiler full control on whether to inline or not.
+ * Additionally, there is an optimization for 4 lanes at once noted below.
  *
- * When not optimizing (-O0), optimizing for size (-Os, -Oz), or using
- * -fno-inline with GCC or Clang, this will automatically be defined.
- */
-#ifndef XXH_NO_INLINE_HINTS
-#  if defined(__OPTIMIZE_SIZE__) /* -Os, -Oz */ \
-   || defined(__NO_INLINE__)     /* -O0, -fno-inline */
-#    define XXH_NO_INLINE_HINTS 1
-#  else
-#    define XXH_NO_INLINE_HINTS 0
-#  endif
-#endif
-
-/*!
- * XXH_REROLL:
- * Whether to reroll XXH32_finalize, and XXH64_finalize,
- * instead of using an unrolled jump table/if statement loop.
+ * 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.
  *
- * This is automatically defined on -Os/-Oz on GCC and Clang.
- */
-#ifndef XXH_REROLL
-#  if defined(__OPTIMIZE_SIZE__)
-#    define XXH_REROLL 1
-#  else
-#    define XXH_REROLL 0
-#  endif
-#endif
-
-
-/* *************************************
-*  Includes & Memory related functions
-***************************************/
-/*!
- * Modify the local functions below should you wish to use some other memory
- * routines for malloc() and free()
+ * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap
+ * nearly perfectly.
  */
-#include <stdlib.h>
-
-static void* XXH_malloc(size_t s) { return malloc(s); }
-static void XXH_free(void* p) { free(p); }
 
-/*! and for memcpy() */
-#include <string.h>
-static void* XXH_memcpy(void* dest, const void* src, size_t size)
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
 {
-    return memcpy(dest,src,size);
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
+    {   /* GCC for darwin arm64 does not like aliasing here */
+        xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
+        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
+        uint8_t const* xinput = (const uint8_t *) input;
+        uint8_t const* xsecret  = (const uint8_t *) secret;
+
+        size_t i;
+#ifdef __wasm_simd128__
+        /*
+         * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret
+         * is constant propagated, which results in it converting it to this
+         * inside the loop:
+         *
+         *    a = v128.load(XXH3_kSecret +  0 + $secret_offset, offset = 0)
+         *    b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0)
+         *    ...
+         *
+         * This requires a full 32-bit address immediate (and therefore a 6 byte
+         * instruction) as well as an add for each offset.
+         *
+         * Putting an asm guard prevents it from folding (at the cost of losing
+         * the alignment hint), and uses the free offset in `v128.load` instead
+         * of adding secret_offset each time which overall reduces code size by
+         * about a kilobyte and improves performance.
+         */
+        XXH_COMPILER_GUARD(xsecret);
+#endif
+        /* Scalar lanes use the normal scalarRound routine */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarRound(acc, input, secret, i);
+        }
+        i = 0;
+        /* 4 NEON lanes at a time. */
+        for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput  + (i * 16));
+            uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput  + ((i+1) * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec_1  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t key_vec_2  = XXH_vld1q_u64(xsecret + ((i+1) * 16));
+            /* data_swap = swap(data_vec) */
+            uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
+            uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
+            uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
+
+            /*
+             * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
+             * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
+             * get one vector with the low 32 bits of each lane, and one vector
+             * with the high 32 bits of each lane.
+             *
+             * The intrinsic returns a double vector because the original ARMv7-a
+             * instruction modified both arguments in place. AArch64 and SIMD128 emit
+             * two instructions from this intrinsic.
+             *
+             *  [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
+             *  [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
+             */
+            uint32x4x2_t unzipped = vuzpq_u32(
+                vreinterpretq_u32_u64(data_key_1),
+                vreinterpretq_u32_u64(data_key_2)
+            );
+            /* data_key_lo = data_key & 0xFFFFFFFF */
+            uint32x4_t data_key_lo = unzipped.val[0];
+            /* data_key_hi = data_key >> 32 */
+            uint32x4_t data_key_hi = unzipped.val[1];
+            /*
+             * Then, we can split the vectors horizontally and multiply which, as for most
+             * widening intrinsics, have a variant that works on both high half vectors
+             * for free on AArch64. A similar instruction is available on SIMD128.
+             *
+             * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
+             */
+            uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
+            uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
+            /*
+             * Clang reorders
+             *    a += b * c;     // umlal   swap.2d, dkl.2s, dkh.2s
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             * to
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             *    c += b * c;     // umlal   acc.2d, dkl.2s, dkh.2s
+             *
+             * While it would make sense in theory since the addition is faster,
+             * for reasons likely related to umlal being limited to certain NEON
+             * pipelines, this is worse. A compiler guard fixes this.
+             */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_1);
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_2);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i]   = vaddq_u64(xacc[i], sum_1);
+            xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
+        }
+        /* Operate on the remaining NEON lanes 2 at a time. */
+        for (; i < XXH3_NEON_LANES / 2; i++) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec = XXH_vld1q_u64(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            /* acc_vec_2 = swap(data_vec) */
+            uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* For two lanes, just use VMOVN and VSHRN. */
+            /* data_key_lo = data_key & 0xFFFFFFFF; */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* data_key_hi = data_key >> 32; */
+            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+            /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
+            uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi);
+            /* Same Clang workaround as before */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i] = vaddq_u64 (xacc[i], sum);
+        }
+    }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
 
-#include <limits.h>   /* ULLONG_MAX */
-
-
-/* *************************************
-*  Compiler Specific Options
-***************************************/
-#ifdef _MSC_VER /* Visual Studio warning fix */
-#  pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   xxh_aliasing_uint64x2_t* xacc       = (xxh_aliasing_uint64x2_t*) acc;
+        uint8_t const* xsecret = (uint8_t const*) secret;
+
+        size_t i;
+        /* WASM uses operator overloads and doesn't need these. */
+#ifndef __wasm_simd128__
+        /* { prime32_1, prime32_1 } */
+        uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1);
+        /* { 0, prime32_1, 0, prime32_1 } */
+        uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32));
 #endif
 
-#if XXH_NO_INLINE_HINTS /* disable inlining hints */
-#  define XXH_FORCE_INLINE static
-#  define XXH_NO_INLINE static
-#elif defined(_MSC_VER)    /* Visual Studio */
-#  define XXH_FORCE_INLINE static __forceinline
-#  define XXH_NO_INLINE static __declspec(noinline)
+        /* AArch64 uses both scalar and neon at the same time */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarScrambleRound(acc, secret, i);
+        }
+        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            uint64x2_t acc_vec  = xacc[i];
+            uint64x2_t shifted  = vshrq_n_u64(acc_vec, 47);
+            uint64x2_t data_vec = veorq_u64(acc_vec, shifted);
+
+            /* xacc[i] ^= xsecret[i]; */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* xacc[i] *= XXH_PRIME32_1 */
+#ifdef __wasm_simd128__
+            /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */
+            xacc[i] = data_key * XXH_PRIME32_1;
 #else
-#  if defined (__cplusplus) \
-    || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
-#    ifdef __GNUC__
-#      define XXH_FORCE_INLINE static inline __attribute__((always_inline))
-#      define XXH_NO_INLINE static __attribute__((noinline))
-#    else
-#      define XXH_FORCE_INLINE static inline
-#      define XXH_NO_INLINE static
-#    endif
-#  else
-#    define XXH_FORCE_INLINE static
-#    define XXH_NO_INLINE static
-#  endif /* __STDC_VERSION__ */
+            /*
+             * Expanded version with portable NEON intrinsics
+             *
+             *    lo(x) * lo(y) + (hi(x) * lo(y) << 32)
+             *
+             * prod_hi = hi(data_key) * lo(prime) << 32
+             *
+             * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector
+             * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits
+             * and avoid the shift.
+             */
+            uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi);
+            /* Extract low bits for vmlal_u32  */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */
+            xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo);
+#endif
+        }
+    }
+}
 #endif
 
+#if (XXH_VECTOR == XXH_VSX)
 
-
-/* *************************************
-*  Debug
-***************************************/
-/*
- * DEBUGLEVEL is expected to be defined externally, typically via the compiler's
- * command line options. The value must be a number.
- */
-#ifndef DEBUGLEVEL
-#  define DEBUGLEVEL 0
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    /* presumed aligned */
+    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 + 16*i);
+        /* key_vec = 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        = xacc[i];
+        acc_vec += product;
+
+        /* swap high and low halves */
+#ifdef __s390x__
+        acc_vec += vec_permi(data_vec, data_vec, 2);
+#else
+        acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
 #endif
+        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_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 };
+        xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 };
+        size_t i;
+        for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            xxh_u64x2 const acc_vec  = xacc[i];
+            xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
+
+            /* xacc[i] ^= 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 */
+            /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF);  */
+            xxh_u64x2 const prod_even  = XXH_vec_mule((xxh_u32x4)data_key, prime);
+            /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32);  */
+            xxh_u64x2 const prod_odd  = XXH_vec_mulo((xxh_u32x4)data_key, prime);
+            xacc[i] = prod_odd + (prod_even << v32);
+    }   }
+}
 
-#if (DEBUGLEVEL>=1)
-#  include <assert.h>   /* note: can still be disabled with NDEBUG */
-#  define XXH_ASSERT(c)   assert(c)
-#else
-#  define XXH_ASSERT(c)   ((void)0)
 #endif
 
-/* note: use after variable declarations */
-#define XXH_STATIC_ASSERT(c)  do { enum { XXH_sa = 1/(int)(!!(c)) }; } while (0)
+#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);
+       }
+    }
+}
 
-/* *************************************
-*  Basic Types
-***************************************/
-#if !defined (__VMS) \
- && (defined (__cplusplus) \
- || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-# include <stdint.h>
-  typedef uint8_t  xxh_u8;
-#else
-  typedef unsigned char      xxh_u8;
 #endif
-typedef XXH32_hash_t xxh_u32;
 
+#if (XXH_VECTOR == XXH_LSX)
+#define _LSX_SHUFFLE(z, y, x, w) (((z) << 6) | ((y) << 4) | ((x) << 2) | (w))
 
-/* ***   Memory access   *** */
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_lsx( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {
+        __m128i* const xacc    =       (__m128i *) acc;
+        const __m128i* const xinput  = (const __m128i *) input;
+        const __m128i* const xsecret = (const __m128i *) secret;
+
+        for (size_t i = 0; i < XXH_STRIPE_LEN / sizeof(__m128i); i++) {
+            /* data_vec = xinput[i]; */
+            __m128i const data_vec = __lsx_vld(xinput + i, 0);
+            /* key_vec = xsecret[i]; */
+            __m128i const key_vec = __lsx_vld(xsecret + i, 0);
+            /* data_key = data_vec ^ key_vec; */
+            __m128i const data_key = __lsx_vxor_v(data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m128i const data_key_lo = __lsx_vsrli_d(data_key, 32);
+            // __m128i const data_key_lo = __lsx_vsrli_d(data_key, 32);
+            /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m128i const product = __lsx_vmulwev_d_wu(data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m128i const data_swap = __lsx_vshuf4i_w(data_vec, _LSX_SHUFFLE(1, 0, 3, 2));
+            __m128i const sum = __lsx_vadd_d(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = __lsx_vadd_d(product, sum);
+        }
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(lsx)
 
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
-/*
- * Manual byteshift. Best for old compilers which don't inline memcpy.
- * We actually directly use XXH_readLE32 and XXH_readBE32.
- */
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_lsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {
+        __m128i* const xacc = (__m128i*) acc;
+        const __m128i* const xsecret = (const __m128i *) secret;
+        const __m128i prime32 = __lsx_vreplgr2vr_w((int)XXH_PRIME32_1);
+
+        for (size_t i = 0; i < XXH_STRIPE_LEN / sizeof(__m128i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m128i const acc_vec = xacc[i];
+            __m128i const shifted = __lsx_vsrli_d(acc_vec, 47);
+            __m128i const data_vec = __lsx_vxor_v(acc_vec, shifted);
+            /* xacc[i] ^= xsecret[i]; */
+            __m128i const key_vec = __lsx_vld(xsecret + i, 0);
+            __m128i const data_key = __lsx_vxor_v(data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m128i const data_key_hi = __lsx_vsrli_d(data_key, 32);
+            __m128i const prod_lo = __lsx_vmulwev_d_wu(data_key, prime32);
+            __m128i const prod_hi = __lsx_vmulwev_d_wu(data_key_hi, prime32);
+            xacc[i] = __lsx_vadd_d(prod_lo, __lsx_vslli_d(prod_hi, 32));
+        }
+    }
+}
 
-/*
- * Force direct memory access. Only works on CPU which support unaligned memory
- * access in hardware.
- */
-static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+#endif
 
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+/* scalar variants - universal */
 
+#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__))
 /*
- * __pack instructions are safer but compiler specific, hence potentially
- * problematic for some compilers.
+ * 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).
  *
- * Currently only defined for GCC and ICC.
+ * 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.
  */
-typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
-static xxh_u32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-
+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
 
-/*
- * Portable and safe solution. Generally efficient.
- * see: https://stackoverflow.com/a/32095106/646947
+/*!
+ * @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.
  */
-static xxh_u32 XXH_read32(const void* memPtr)
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc,
+                 void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret,
+                 size_t lane)
 {
-    xxh_u32 val;
-    memcpy(&val, memPtr, sizeof(val));
-    return val;
+    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]);
+    }
 }
 
-#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
-
-
-/* ***   Endianess   *** */
-typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess;
+/*!
+ * @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)
+{
+    size_t i;
+    /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
+#if defined(__GNUC__) && !defined(__clang__) \
+  && (defined(__arm__) || defined(__thumb2__)) \
+  && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
+  && XXH_SIZE_OPT <= 0
+#  pragma GCC unroll 8
+#endif
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarRound(acc, input, secret, i);
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
 
 /*!
- * XXH_CPU_LITTLE_ENDIAN:
- * Defined to 1 if the target is little endian, or 0 if it is big endian.
- * It can be defined externally, for example on the compiler command line.
+ * @internal
+ * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar().
  *
- * If it is not defined, a runtime check (which is usually constant folded)
- * is used instead.
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
  */
-#ifndef XXH_CPU_LITTLE_ENDIAN
-/*
- * Try to detect endianness automatically, to avoid the nonstandard behavior
- * in `XXH_isLittleEndian()`
+XXH_FORCE_INLINE void
+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 */
+    XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    {
+        xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8);
+        xxh_u64 acc64 = xacc[lane];
+        acc64 = XXH_xorshift64(acc64, 47);
+        acc64 ^= key64;
+        acc64 *= XXH_PRIME32_1;
+        xacc[lane] = acc64;
+    }
+}
+
+/*!
+ * @internal
+ * @brief Scrambles the accumulators after a large chunk has been read
  */
-#  if defined(_WIN32) /* Windows is always little endian */ \
-     || defined(__LITTLE_ENDIAN__) \
-     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
-#    define XXH_CPU_LITTLE_ENDIAN 1
-#  elif defined(__BIG_ENDIAN__) \
-     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
-#    define XXH_CPU_LITTLE_ENDIAN 0
-#  else
-static int XXH_isLittleEndian(void)
+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);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
 {
     /*
-     * Nonstandard, but well-defined behavior in practice.
-     * Don't use static: it is detrimental to performance.
+     * We need a separate pointer for the hack below,
+     * which requires a non-const pointer.
+     * Any decent compiler will optimize this out otherwise.
      */
-    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
-    return one.c[0];
-}
-#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
-#  endif
+    const xxh_u8* kSecretPtr = XXH3_kSecret;
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+
+#if defined(__GNUC__) && defined(__aarch64__)
+    /*
+     * UGLY HACK:
+     * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are
+     * placed sequentially, in order, at the top of the unrolled loop.
+     *
+     * While MOVK is great for generating constants (2 cycles for a 64-bit
+     * constant compared to 4 cycles for LDR), it fights for bandwidth with
+     * the arithmetic instructions.
+     *
+     *   I   L   S
+     * MOVK
+     * MOVK
+     * MOVK
+     * MOVK
+     * ADD
+     * SUB      STR
+     *          STR
+     * By forcing loads from memory (as the asm line causes the compiler to assume
+     * that XXH3_kSecretPtr has been changed), the pipelines are used more
+     * efficiently:
+     *   I   L   S
+     *      LDR
+     *  ADD LDR
+     *  SUB     STR
+     *          STR
+     *
+     * See XXH3_NEON_LANES for details on the pipsline.
+     *
+     * XXH3_64bits_withSeed, len == 256, Snapdragon 835
+     *   without hack: 2654.4 MB/s
+     *   with hack:    3202.9 MB/s
+     */
+    XXH_COMPILER_GUARD(kSecretPtr);
 #endif
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
+        int i;
+        for (i=0; i < nbRounds; i++) {
+            /*
+             * The asm hack causes the compiler to assume that kSecretPtr aliases with
+             * customSecret, and on aarch64, this prevented LDP from merging two
+             * loads together for free. Putting the loads together before the stores
+             * properly generates LDP.
+             */
+            xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
+            xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i,     lo);
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
+    }   }
+}
 
 
+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);
 
 
-/* ****************************************
-*  Compiler-specific Functions and Macros
-******************************************/
-#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+#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
+
+#elif (XXH_VECTOR == XXH_LSX)
+#define XXH3_accumulate_512 XXH3_accumulate_512_lsx
+#define XXH3_accumulate     XXH3_accumulate_lsx
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_lsx
+#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
 
-#ifndef __has_builtin
-#  define __has_builtin(x) 0
 #endif
 
-#if !defined(NO_CLANG_BUILTIN) && __has_builtin(__builtin_rotateleft32) \
-                               && __has_builtin(__builtin_rotateleft64)
-#  define XXH_rotl32 __builtin_rotateleft32
-#  define XXH_rotl64 __builtin_rotateleft64
-/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
-#elif defined(_MSC_VER)
-#  define XXH_rotl32(x,r) _rotl(x,r)
-#  define XXH_rotl64(x,r) _rotl64(x,r)
-#else
-#  define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
-#  define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
+#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */
+#  undef XXH3_initCustomSecret
+#  define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 #endif
 
-#if defined(_MSC_VER)     /* Visual Studio */
-#  define XXH_swap32 _byteswap_ulong
-#elif XXH_GCC_VERSION >= 403
-#  define XXH_swap32 __builtin_bswap32
-#else
-static xxh_u32 XXH_swap32 (xxh_u32 x)
+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 f_acc,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
+    size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock;
+    size_t const nb_blocks = (len - 1) / block_len;
+
+    size_t n;
+
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+
+    for (n = 0; n < nb_blocks; n++) {
+        f_acc(acc, input + n*block_len, secret, nbStripesPerBlock);
+        f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
+    }
+
+    /* last partial block */
+    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));
+        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 */
+            XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+    }   }
+}
+
+XXH_FORCE_INLINE xxh_u64
+XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
 {
-    return  ((x << 24) & 0xff000000 ) |
-            ((x <<  8) & 0x00ff0000 ) |
-            ((x >>  8) & 0x0000ff00 ) |
-            ((x >> 24) & 0x000000ff );
+    return XXH3_mul128_fold64(
+               acc[0] ^ XXH_readLE64(secret),
+               acc[1] ^ XXH_readLE64(secret+8) );
 }
-#endif
 
+static XXH_PUREF XXH64_hash_t
+XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
+{
+    xxh_u64 result64 = start;
+    size_t i = 0;
+
+    for (i = 0; i < 4; i++) {
+        result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__arm__) || defined(__thumb__))       /* ARMv7 */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */  \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
+         * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
+         * XXH3_64bits, len == 256, Snapdragon 835:
+         *   without hack: 2063.7 MB/s
+         *   with hack:    2560.7 MB/s
+         */
+        XXH_COMPILER_GUARD(result64);
+#endif
+    }
 
-/* ***************************
-*  Memory reads
-*****************************/
-typedef enum { XXH_aligned, XXH_unaligned } XXH_alignment;
+    return XXH3_avalanche(result64);
+}
 
-/*
- * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
- *
- * This is ideal for older compilers which don't inline memcpy.
- */
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/* do not align on 8, so that the secret is different from the accumulator */
+#define XXH_SECRET_MERGEACCS_START 11
 
-XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
+static XXH_PUREF XXH64_hash_t
+XXH3_finalizeLong_64b(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 len)
 {
-    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
-    return bytePtr[0]
-         | ((xxh_u32)bytePtr[1] << 8)
-         | ((xxh_u32)bytePtr[2] << 16)
-         | ((xxh_u32)bytePtr[3] << 24);
+    return XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, len * XXH_PRIME64_1);
 }
 
-XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
+#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \
+                        XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 }
+
+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 f_acc,
+                           XXH3_f_scrambleAcc f_scramble)
 {
-    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
-    return bytePtr[3]
-         | ((xxh_u32)bytePtr[2] << 8)
-         | ((xxh_u32)bytePtr[1] << 16)
-         | ((xxh_u32)bytePtr[0] << 24);
+    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_acc, f_scramble);
+
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    return XXH3_finalizeLong_64b(acc, (const xxh_u8*)secret, (xxh_u64)len);
 }
 
-#else
-XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
+/*
+ * It's important for performance to transmit secret's size (when it's static)
+ * so that the compiler can properly optimize the vectorized loop.
+ * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set.
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                             XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
 {
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+    (void)seed64;
+    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
 }
 
-static xxh_u32 XXH_readBE32(const void* ptr)
+/*
+ * It's preferable for performance that XXH3_hashLong is not inlined,
+ * as it results in a smaller function for small data, easier to the instruction cache.
+ * Note that inside this no_inline function, we do inline the internal loop,
+ * and provide a statically defined secret size to allow optimization of vector loop.
+ */
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
+                          XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
 {
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
 }
-#endif
 
-XXH_FORCE_INLINE xxh_u32
-XXH_readLE32_align(const void* ptr, XXH_alignment align)
+/*
+ * XXH3_hashLong_64b_withSeed():
+ * Generate a custom key based on alteration of default XXH3_kSecret with the seed,
+ * and then use this key for long mode hashing.
+ *
+ * This operation is decently fast but nonetheless costs a little bit of time.
+ * Try to avoid it whenever possible (typically when seed==0).
+ *
+ * It's important for performance that XXH3_hashLong is not inlined. Not sure
+ * why (uop cache maybe?), but the difference is large and easily measurable.
+ */
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
+                                    XXH64_hash_t seed,
+                                    XXH3_f_accumulate f_acc,
+                                    XXH3_f_scrambleAcc f_scramble,
+                                    XXH3_f_initCustomSecret f_initSec)
 {
-    if (align==XXH_unaligned) {
-        return XXH_readLE32(ptr);
-    } else {
-        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
+#if XXH_SIZE_OPT <= 0
+    if (seed == 0)
+        return XXH3_hashLong_64b_internal(input, len,
+                                          XXH3_kSecret, sizeof(XXH3_kSecret),
+                                          f_acc, f_scramble);
+#endif
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed);
+        return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
+                                          f_acc, f_scramble);
     }
 }
 
-
-/* *************************************
-*  Misc
-***************************************/
-XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+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, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
 
 
-/* *******************************************************************
-*  32-bit hash functions
-*********************************************************************/
-static const xxh_u32 PRIME32_1 = 0x9E3779B1U;   /* 0b10011110001101110111100110110001 */
-static const xxh_u32 PRIME32_2 = 0x85EBCA77U;   /* 0b10000101111010111100101001110111 */
-static const xxh_u32 PRIME32_3 = 0xC2B2AE3DU;   /* 0b11000010101100101010111000111101 */
-static const xxh_u32 PRIME32_4 = 0x27D4EB2FU;   /* 0b00100111110101001110101100101111 */
-static const xxh_u32 PRIME32_5 = 0x165667B1U;   /* 0b00010110010101100110011110110001 */
+typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t,
+                                          XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t);
 
-static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
+                     XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                     XXH3_hashLong64_f f_hashLong)
 {
-    acc += input * PRIME32_2;
-    acc  = XXH_rotl32(acc, 13);
-    acc *= PRIME32_1;
-#if defined(__GNUC__) && defined(__SSE4_1__) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
     /*
-     * UGLY HACK:
-     * This inline assembly hack forces acc into a normal register. This is the
-     * only thing that prevents GCC and Clang from autovectorizing the XXH32
-     * loop (pragmas and attributes don't work for some resason) without globally
-     * disabling SSE4.1.
-     *
-     * The reason we want to avoid vectorization is because despite working on
-     * 4 integers at a time, there are multiple factors slowing XXH32 down on
-     * SSE4:
-     * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
-     *   newer chips!) making it slightly slower to multiply four integers at
-     *   once compared to four integers independently. Even when pmulld was
-     *   fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
-     *   just to multiply unless doing a long operation.
-     *
-     * - Four instructions are required to rotate,
-     *      movqda tmp,  v // not required with VEX encoding
-     *      pslld  tmp, 13 // tmp <<= 13
-     *      psrld  v,   19 // x >>= 19
-     *      por    v,  tmp // x |= tmp
-     *   compared to one for scalar:
-     *      roll   v, 13    // reliably fast across the board
-     *      shldl  v, v, 13 // Sandy Bridge and later prefer this for some reason
-     *
-     * - Instruction level parallelism is actually more beneficial here because
-     *   the SIMD actually serializes this operation: While v1 is rotating, v2
-     *   can load data, while v3 can multiply. SSE forces them to operate
-     *   together.
-     *
-     * How this hack works:
-     * __asm__(""       // Declare an assembly block but don't declare any instructions
-     *          :       // However, as an Input/Output Operand,
-     *          "+r"    // constrain a read/write operand (+) as a general purpose register (r).
-     *          (acc)   // and set acc as the operand
-     * );
-     *
-     * Because of the 'r', the compiler has promised that seed will be in a
-     * general purpose register and the '+' says that it will be 'read/write',
-     * so it has to assume it has changed. It is like volatile without all the
-     * loads and stores.
-     *
-     * Since the argument has to be in a normal register (not an SSE register),
-     * each time XXH32_round is called, it is impossible to vectorize.
+     * If an action is to be taken if `secretLen` condition is not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     * Also, note that function signature doesn't offer room to return an error.
      */
-    __asm__("" : "+r" (acc));
-#endif
-    return acc;
+    if (len <= 16)
+        return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen);
 }
 
-/* mix all bits */
-static xxh_u32 XXH32_avalanche(xxh_u32 h32)
+
+/* ===   Public entry point   === */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
 {
-    h32 ^= h32 >> 15;
-    h32 *= PRIME32_2;
-    h32 ^= h32 >> 13;
-    h32 *= PRIME32_3;
-    h32 ^= h32 >> 16;
-    return(h32);
+    return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
 }
 
-#define XXH_get32bits(p) XXH_readLE32_align(p, align)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+}
 
-static xxh_u32
-XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
 {
-#define PROCESS1 do {                           \
-    h32 += (*ptr++) * PRIME32_5;                \
-    h32 = XXH_rotl32(h32, 11) * PRIME32_1;      \
-} while (0)
+    return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+}
 
-#define PROCESS4 do {                           \
-    h32 += XXH_get32bits(ptr) * PRIME32_3;      \
-    ptr += 4;                                   \
-    h32  = XXH_rotl32(h32, 17) * PRIME32_4;     \
-} while (0)
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (length <= XXH3_MIDSIZE_MAX)
+        return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize);
+}
 
-    /* Compact rerolled version */
-    if (XXH_REROLL) {
-        len &= 15;
-        while (len >= 4) {
-            PROCESS4;
-            len -= 4;
-        }
-        while (len > 0) {
-            PROCESS1;
-            --len;
-        }
-        return XXH32_avalanche(h32);
-    } else {
-         switch(len&15) /* or switch(bEnd - p) */ {
-           case 12:      PROCESS4;
-                         /* fallthrough */
-           case 8:       PROCESS4;
-                         /* fallthrough */
-           case 4:       PROCESS4;
-                         return XXH32_avalanche(h32);
-
-           case 13:      PROCESS4;
-                         /* fallthrough */
-           case 9:       PROCESS4;
-                         /* fallthrough */
-           case 5:       PROCESS4;
-                         PROCESS1;
-                         return XXH32_avalanche(h32);
-
-           case 14:      PROCESS4;
-                         /* fallthrough */
-           case 10:      PROCESS4;
-                         /* fallthrough */
-           case 6:       PROCESS4;
-                         PROCESS1;
-                         PROCESS1;
-                         return XXH32_avalanche(h32);
-
-           case 15:      PROCESS4;
-                         /* fallthrough */
-           case 11:      PROCESS4;
-                         /* fallthrough */
-           case 7:       PROCESS4;
-                         /* fallthrough */
-           case 3:       PROCESS1;
-                         /* fallthrough */
-           case 2:       PROCESS1;
-                         /* fallthrough */
-           case 1:       PROCESS1;
-                         /* fallthrough */
-           case 0:       return XXH32_avalanche(h32);
+
+/* ===   XXH3 streaming   === */
+#ifndef XXH_NO_STREAM
+/*
+ * Malloc's a pointer that is always aligned to @align.
+ *
+ * This must be freed with `XXH_alignedFree()`.
+ *
+ * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
+ * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
+ * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
+ *
+ * This underalignment previously caused a rather obvious crash which went
+ * completely unnoticed due to XXH3_createState() not actually being tested.
+ * Credit to RedSpah for noticing this bug.
+ *
+ * The alignment is done manually: Functions like posix_memalign or _mm_malloc
+ * are avoided: To maintain portability, we would have to write a fallback
+ * like this anyways, and besides, testing for the existence of library
+ * functions without relying on external build tools is impossible.
+ *
+ * The method is simple: Overallocate, manually align, and store the offset
+ * to the original behind the returned pointer.
+ *
+ * Align must be a power of 2 and 8 <= align <= 128.
+ */
+static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align)
+{
+    XXH_ASSERT(align <= 128 && align >= 8); /* range check */
+    XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
+    XXH_ASSERT(s != 0 && s < (s + align));  /* empty/overflow */
+    {   /* Overallocate to make room for manual realignment and an offset byte */
+        xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
+        if (base != NULL) {
+            /*
+             * Get the offset needed to align this pointer.
+             *
+             * Even if the returned pointer is aligned, there will always be
+             * at least one byte to store the offset to the original pointer.
+             */
+            size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
+            /* Add the offset for the now-aligned pointer */
+            xxh_u8* ptr = base + offset;
+
+            XXH_ASSERT((size_t)ptr % align == 0);
+
+            /* Store the offset immediately before the returned pointer. */
+            ptr[-1] = (xxh_u8)offset;
+            return ptr;
         }
-        XXH_ASSERT(0);
-        return h32;   /* reaching this point is deemed impossible */
+        return NULL;
     }
 }
-
-XXH_FORCE_INLINE xxh_u32
-XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+/*
+ * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
+ * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
+ */
+static void XXH_alignedFree(void* p)
 {
-    const xxh_u8* bEnd = input + len;
-    xxh_u32 h32;
-
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-    if (input==NULL) {
-        len=0;
-        bEnd=input=(const xxh_u8*)(size_t)16;
-    }
-#endif
-
-    if (len>=16) {
-        const xxh_u8* const limit = bEnd - 15;
-        xxh_u32 v1 = seed + PRIME32_1 + PRIME32_2;
-        xxh_u32 v2 = seed + PRIME32_2;
-        xxh_u32 v3 = seed + 0;
-        xxh_u32 v4 = seed - PRIME32_1;
-
-        do {
-            v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
-            v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
-            v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
-            v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
-        } while (input < limit);
-
-        h32 = XXH_rotl32(v1, 1)  + XXH_rotl32(v2, 7)
-            + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
-    } else {
-        h32  = seed + PRIME32_5;
+    if (p != NULL) {
+        xxh_u8* ptr = (xxh_u8*)p;
+        /* Get the offset byte we added in XXH_malloc. */
+        xxh_u8 offset = ptr[-1];
+        /* Free the original malloc'd pointer */
+        xxh_u8* base = ptr - offset;
+        XXH_free(base);
     }
-
-    h32 += (xxh_u32)len;
-
-    return XXH32_finalize(h32, input, len&15, align);
 }
-
-
-XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Allocate an @ref XXH3_state_t.
+ *
+ * @return An allocated pointer of @ref XXH3_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH3_freeState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
 {
-#if 0
-    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
-    XXH32_state_t state;
-    XXH32_reset(&state, seed);
-    XXH32_update(&state, (const xxh_u8*)input, len);
-    return XXH32_digest(&state);
-
-#else
-
-    if (XXH_FORCE_ALIGN_CHECK) {
-        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
-            return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
-    }   }
-
-    return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
-#endif
+    XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
+    if (state==NULL) return NULL;
+    XXH3_INITSTATE(state);
+    return state;
 }
 
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Frees an @ref XXH3_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note Must be allocated with XXH3_createState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
+{
+    XXH_alignedFree(statePtr);
+    return XXH_OK;
+}
 
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+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));
+}
 
-/*******   Hash streaming   *******/
+static void
+XXH3_reset_internal(XXH3_state_t* statePtr,
+                    XXH64_hash_t seed,
+                    const void* secret, size_t secretSize)
+{
+    size_t const initStart = offsetof(XXH3_state_t, bufferedSize);
+    size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart;
+    XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart);
+    XXH_ASSERT(statePtr != NULL);
+    /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */
+    memset((char*)statePtr + initStart, 0, initLength);
+    statePtr->acc[0] = XXH_PRIME32_3;
+    statePtr->acc[1] = XXH_PRIME64_1;
+    statePtr->acc[2] = XXH_PRIME64_2;
+    statePtr->acc[3] = XXH_PRIME64_3;
+    statePtr->acc[4] = XXH_PRIME64_4;
+    statePtr->acc[5] = XXH_PRIME32_2;
+    statePtr->acc[6] = XXH_PRIME64_5;
+    statePtr->acc[7] = XXH_PRIME32_1;
+    statePtr->seed = seed;
+    statePtr->useSeed = (seed != 0);
+    statePtr->extSecret = (const unsigned char*)secret;
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+    statePtr->secretLimit = secretSize - XXH_STRIPE_LEN;
+    statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
+}
 
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
 {
-    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
 }
-XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
-    XXH_free(statePtr);
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, secret, secretSize);
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
     return XXH_OK;
 }
 
-XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
 {
-    memcpy(dstState, srcState, sizeof(*dstState));
+    if (statePtr == NULL) return XXH_ERROR;
+    if (seed==0) return XXH3_64bits_reset(statePtr);
+    if ((seed != statePtr->seed) || (statePtr->extSecret != NULL))
+        XXH3_initCustomSecret(statePtr->customSecret, seed);
+    XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
 }
 
-XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64)
 {
-    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state));
-    state.v1 = seed + PRIME32_1 + PRIME32_2;
-    state.v2 = seed + PRIME32_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - PRIME32_1;
-    /* do not write into reserved, planned to be removed in a future version */
-    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved));
+    if (statePtr == NULL) return XXH_ERROR;
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, seed64, secret, secretSize);
+    statePtr->useSeed = 1; /* always, even if seed64==0 */
     return XXH_OK;
 }
 
+/*!
+ * @internal
+ * @brief Processes a large input for XXH3_update() and XXH3_digest_long().
+ *
+ * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block.
+ *
+ * @param acc                Pointer to the 8 accumulator lanes
+ * @param nbStripesSoFarPtr  In/out pointer to the number of leftover stripes in the block*
+ * @param nbStripesPerBlock  Number of stripes in a block
+ * @param input              Input pointer
+ * @param nbStripes          Number of stripes to process
+ * @param secret             Secret pointer
+ * @param secretLimit        Offset of the last block in @p secret
+ * @param f_acc              Pointer to an XXH3_accumulate implementation
+ * @param f_scramble         Pointer to an XXH3_scrambleAcc implementation
+ * @return                   Pointer past the end of @p input after processing
+ */
+XXH_FORCE_INLINE const xxh_u8 *
+XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
+                    size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
+                    const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
+                    const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
+                    XXH3_f_accumulate f_acc,
+                    XXH3_f_scrambleAcc f_scramble)
+{
+    const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE;
+    /* Process full blocks */
+    if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) {
+        /* Process the initial partial block... */
+        size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr;
 
-XXH_PUBLIC_API XXH_errorcode
-XXH32_update(XXH32_state_t* state, const void* input, size_t len)
+        do {
+            /* Accumulate and scramble */
+            f_acc(acc, input, initialSecret, nbStripesThisIter);
+            f_scramble(acc, secret + secretLimit);
+            input += nbStripesThisIter * XXH_STRIPE_LEN;
+            nbStripes -= nbStripesThisIter;
+            /* Then continue the loop with the full block size */
+            nbStripesThisIter = nbStripesPerBlock;
+            initialSecret = secret;
+        } while (nbStripes >= nbStripesPerBlock);
+        *nbStripesSoFarPtr = 0;
+    }
+    /* Process a partial block */
+    if (nbStripes > 0) {
+        f_acc(acc, input, initialSecret, nbStripes);
+        input += nbStripes * XXH_STRIPE_LEN;
+        *nbStripesSoFarPtr += nbStripes;
+    }
+    /* Return end pointer */
+    return input;
+}
+
+#ifndef XXH3_STREAM_USE_STACK
+# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */
+#   define XXH3_STREAM_USE_STACK 1
+# endif
+#endif
+/*
+ * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
+ */
+XXH_FORCE_INLINE XXH_errorcode
+XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
+            const xxh_u8* XXH_RESTRICT input, size_t len,
+            XXH3_f_accumulate f_acc,
+            XXH3_f_scrambleAcc f_scramble)
 {
-    if (input==NULL)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
         return XXH_OK;
+    }
+
+    XXH_ASSERT(state != NULL);
+    {   const xxh_u8* const bEnd = input + len;
+        const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* For some reason, gcc and MSVC seem to suffer greatly
+         * when operating accumulators directly into state.
+         * Operating into stack space seems to enable proper optimization.
+         * clang, on the other hand, doesn't seem to need this trick */
+        XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
+        XXH_memcpy(acc, state->acc, sizeof(acc));
 #else
-        return XXH_ERROR;
+        xxh_u64* XXH_RESTRICT const acc = state->acc;
 #endif
+        state->totalLen += len;
+        XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
 
-    {   const xxh_u8* p = (const xxh_u8*)input;
-        const xxh_u8* const bEnd = p + len;
-
-        state->total_len_32 += (XXH32_hash_t)len;
-        state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
-
-        if (state->memsize + len < 16)  {   /* fill in tmp buffer */
-            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
-            state->memsize += (XXH32_hash_t)len;
+        /* small input : just fill in tmp buffer */
+        if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
+            XXH_memcpy(state->buffer + state->bufferedSize, input, len);
+            state->bufferedSize += (XXH32_hash_t)len;
             return XXH_OK;
         }
 
-        if (state->memsize) {   /* some data left from previous update */
-            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
-            {   const xxh_u32* p32 = state->mem32;
-                state->v1 = XXH32_round(state->v1, XXH_readLE32(p32)); p32++;
-                state->v2 = XXH32_round(state->v2, XXH_readLE32(p32)); p32++;
-                state->v3 = XXH32_round(state->v3, XXH_readLE32(p32)); p32++;
-                state->v4 = XXH32_round(state->v4, XXH_readLE32(p32));
-            }
-            p += 16-state->memsize;
-            state->memsize = 0;
-        }
-
-        if (p <= bEnd-16) {
-            const xxh_u8* const limit = bEnd - 16;
-            xxh_u32 v1 = state->v1;
-            xxh_u32 v2 = state->v2;
-            xxh_u32 v3 = state->v3;
-            xxh_u32 v4 = state->v4;
-
-            do {
-                v1 = XXH32_round(v1, XXH_readLE32(p)); p+=4;
-                v2 = XXH32_round(v2, XXH_readLE32(p)); p+=4;
-                v3 = XXH32_round(v3, XXH_readLE32(p)); p+=4;
-                v4 = XXH32_round(v4, XXH_readLE32(p)); p+=4;
-            } while (p<=limit);
-
-            state->v1 = v1;
-            state->v2 = v2;
-            state->v3 = v3;
-            state->v4 = v4;
+        /* total input is now > XXH3_INTERNALBUFFER_SIZE */
+        #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN)
+        XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0);   /* clean multiple */
+
+        /*
+         * Internal buffer is partially filled (always, except at beginning)
+         * Complete it, then consume it.
+         */
+        if (state->bufferedSize) {
+            size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
+            XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
+            input += loadSize;
+            XXH3_consumeStripes(acc,
+                               &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
+                                secret, state->secretLimit,
+                                f_acc, f_scramble);
+            state->bufferedSize = 0;
         }
+        XXH_ASSERT(input < bEnd);
+        if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
+            size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
+            input = XXH3_consumeStripes(acc,
+                                       &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                       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);
 
-        if (p < bEnd) {
-            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
-            state->memsize = (unsigned)(bEnd-p);
         }
+        /* Some remaining input (always) : buffer it */
+        XXH_ASSERT(input < bEnd);
+        XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE);
+        XXH_ASSERT(state->bufferedSize == 0);
+        XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
+        state->bufferedSize = (XXH32_hash_t)(bEnd-input);
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* save stack accumulators into state */
+        XXH_memcpy(state->acc, acc, sizeof(acc));
+#endif
     }
 
     return XXH_OK;
 }
 
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+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, XXH3_scrambleAcc);
+}
+
 
-XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* state)
+XXH_FORCE_INLINE void
+XXH3_digest_long (XXH64_hash_t* acc,
+                  const XXH3_state_t* state,
+                  const unsigned char* secret)
 {
-    xxh_u32 h32;
+    xxh_u8 lastStripe[XXH_STRIPE_LEN];
+    const xxh_u8* lastStripePtr;
 
-    if (state->large_len) {
-        h32 = XXH_rotl32(state->v1, 1)
-            + XXH_rotl32(state->v2, 7)
-            + XXH_rotl32(state->v3, 12)
-            + XXH_rotl32(state->v4, 18);
-    } else {
-        h32 = state->v3 /* == seed */ + PRIME32_5;
+    /*
+     * 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, XXH3_scrambleAcc);
+        lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN;
+    } else {  /* bufferedSize < 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);
+        lastStripePtr = lastStripe;
     }
+    /* Last stripe */
+    XXH3_accumulate_512(acc,
+                        lastStripePtr,
+                        secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+}
 
-    h32 += state->total_len_32;
+/*! @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) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        return XXH3_finalizeLong_64b(acc, secret, (xxh_u64)state->totalLen);
+    }
+    /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
+    if (state->useSeed)
+        return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                  secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+#endif /* !XXH_NO_STREAM */
+
+
+/* ==========================================
+ * XXH3 128 bits (a.k.a XXH128)
+ * ==========================================
+ * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
+ * even without counting the significantly larger output size.
+ *
+ * For example, extra steps are taken to avoid the seed-dependent collisions
+ * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
+ *
+ * This strength naturally comes at the cost of some speed, especially on short
+ * lengths. Note that longer hashes are about as fast as the 64-bit version
+ * due to it using only a slight modification of the 64-bit loop.
+ *
+ * XXH128 is also more oriented towards 64-bit machines. It is still extremely
+ * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
+ */
 
-    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+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. */
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8 const c1 = input[0];
+        xxh_u8 const c2 = input[len >> 1];
+        xxh_u8 const c3 = input[len - 1];
+        xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
+                                | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
+        xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
+        xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
+        xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
+        xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
+        XXH128_hash_t h128;
+        h128.low64  = XXH64_avalanche(keyed_lo);
+        h128.high64 = XXH64_avalanche(keyed_hi);
+        return h128;
+    }
+}
+
+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);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input_lo = XXH_readLE32(input);
+        xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
+        xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
+        xxh_u64 const keyed = input_64 ^ bitflip;
+
+        /* Shift len to the left to ensure it is even, this avoids even multiplies. */
+        XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2));
+
+        m128.high64 += (m128.low64 << 1);
+        m128.low64  ^= (m128.high64 >> 3);
+
+        m128.low64   = XXH_xorshift64(m128.low64, 35);
+        m128.low64  *= PRIME_MX2;
+        m128.low64   = XXH_xorshift64(m128.low64, 28);
+        m128.high64  = XXH3_avalanche(m128.high64);
+        return m128;
+    }
 }
 
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
+        xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
+        xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64       input_hi = XXH_readLE64(input + len - 8);
+        XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1);
+        /*
+         * Put len in the middle of m128 to ensure that the length gets mixed to
+         * both the low and high bits in the 128x64 multiply below.
+         */
+        m128.low64 += (xxh_u64)(len - 1) << 54;
+        input_hi   ^= bitfliph;
+        /*
+         * Add the high 32 bits of input_hi to the high 32 bits of m128, then
+         * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to
+         * the high 64 bits of m128.
+         *
+         * The best approach to this operation is different on 32-bit and 64-bit.
+         */
+        if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
+            /*
+             * 32-bit optimized version, which is more readable.
+             *
+             * On 32-bit, it removes an ADC and delays a dependency between the two
+             * halves of m128.high64, but it generates an extra mask on 64-bit.
+             */
+            m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2);
+        } else {
+            /*
+             * 64-bit optimized (albeit more confusing) version.
+             *
+             * Uses some properties of addition and multiplication to remove the mask:
+             *
+             * Let:
+             *    a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
+             *    b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
+             *    c = XXH_PRIME32_2
+             *
+             *    a + (b * c)
+             * Inverse Property: x + y - x == y
+             *    a + (b * (1 + c - 1))
+             * Distributive Property: x * (y + z) == (x * y) + (x * z)
+             *    a + (b * 1) + (b * (c - 1))
+             * Identity Property: x * 1 == x
+             *    a + b + (b * (c - 1))
+             *
+             * Substitute a, b, and c:
+             *    input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             *
+             * Since input_hi.hi + input_hi.lo == input_hi, we get this:
+             *    input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             */
+            m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1);
+        }
+        /* m128 ^= XXH_swap64(m128 >> 64); */
+        m128.low64  ^= XXH_swap64(m128.high64);
 
-/*******   Canonical representation   *******/
+        {   /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */
+            XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2);
+            h128.high64 += m128.high64 * XXH_PRIME64_2;
+
+            h128.low64   = XXH3_avalanche(h128.low64);
+            h128.high64  = XXH3_avalanche(h128.high64);
+            return h128;
+    }   }
+}
 
 /*
- * The default return values from XXH functions are unsigned 32 and 64 bit
- * integers.
- *
- * The canonical representation uses big endian convention, the same convention
- * as human-readable numbers (large digits first).
- *
- * This way, hash values can be written into a file or buffer, remaining
- * comparable across different systems.
- *
- * The following functions allow transformation of hash values to and from their
- * canonical format.
+ * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
  */
-XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+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_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
-    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
-    memcpy(dst, &hash, sizeof(*dst));
+    XXH_ASSERT(len <= 16);
+    {   if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
+        if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
+        {   XXH128_hash_t h128;
+            xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
+            xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
+            h128.low64 = XXH64_avalanche(seed ^ bitflipl);
+            h128.high64 = XXH64_avalanche( seed ^ bitfliph);
+            return h128;
+    }   }
 }
 
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+/*
+ * A bit slower than XXH3_mix16B, but handles multiply by zero better.
+ */
+XXH_FORCE_INLINE XXH128_hash_t
+XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
+              const xxh_u8* secret, XXH64_hash_t seed)
 {
-    return XXH_readBE32(src);
+    acc.low64  += XXH3_mix16B (input_1, secret+0, seed);
+    acc.low64  ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
+    acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
+    acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
+    return acc;
 }
 
 
-#ifndef XXH_NO_LONG_LONG
+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)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
 
-/* *******************************************************************
-*  64-bit hash functions
-*********************************************************************/
+    {   XXH128_hash_t acc;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
 
-/*******   Memory access   *******/
+#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) {
+                    acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
+                }
+                acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
+            }
+            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)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
+    }
+}
 
-typedef XXH64_hash_t xxh_u64;
+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)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    {   XXH128_hash_t acc;
+        unsigned i;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+        /*
+         *  We set as `i` as offset + 32. We do this so that unchanged
+         * `len` can be used as upper bound. This reaches a sweet spot
+         * where both x86 and aarch64 get simple agen and good codegen
+         * for the loop.
+         */
+        for (i = 32; i < 160; i += 32) {
+            acc = XXH128_mix32B(acc,
+                                input  + i - 32,
+                                input  + i - 16,
+                                secret + i - 32,
+                                seed);
+        }
+        acc.low64 = XXH3_avalanche(acc.low64);
+        acc.high64 = XXH3_avalanche(acc.high64);
+        /*
+         * NB: `i <= len` will duplicate the last 32-bytes if
+         * len % 32 was zero. This is an unfortunate necessity to keep
+         * the hash result stable.
+         */
+        for (i=160; i <= len; i += 32) {
+            acc = XXH128_mix32B(acc,
+                                input + i - 32,
+                                input + i - 16,
+                                secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
+                                seed);
+        }
+        /* last bytes */
+        acc = XXH128_mix32B(acc,
+                            input + len - 16,
+                            input + len - 32,
+                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
+                            (XXH64_hash_t)0 - seed);
+
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
+    }
+}
 
+static XXH_PUREF XXH128_hash_t
+XXH3_finalizeLong_128b(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, size_t secretSize, xxh_u64 len)
+{
+    XXH128_hash_t h128;
+    h128.low64 = XXH3_finalizeLong_64b(acc, secret, len);
+    h128.high64 = XXH3_mergeAccs(acc, secret + secretSize
+                                             - XXH_STRIPE_LEN - XXH_SECRET_MERGEACCS_START,
+                                             ~(len * XXH_PRIME64_2));
+    return h128;
+}
 
-/*!
- * XXH_REROLL_XXH64:
- * Whether to reroll the XXH64_finalize() loop.
- *
- * Just like XXH32, we can unroll the XXH64_finalize() loop. This can be a
- * performance gain on 64-bit hosts, as only one jump is required.
- *
- * However, on 32-bit hosts, because arithmetic needs to be done with two 32-bit
- * registers, and 64-bit arithmetic needs to be simulated, it isn't beneficial
- * to unroll. The code becomes ridiculously large (the largest function in the
- * binary on i386!), and rerolling it saves anywhere from 3kB to 20kB. It is
- * also slightly faster because it fits into cache better and is more likely
- * to be inlined by the compiler.
- *
- * If XXH_REROLL is defined, this is ignored and the loop is always rerolled.
- */
-#ifndef XXH_REROLL_XXH64
-#  if (defined(__ILP32__) || defined(_ILP32)) /* ILP32 is often defined on 32-bit GCC family */ \
-   || !(defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) /* x86-64 */ \
-     || defined(_M_ARM64) || defined(__aarch64__) || defined(__arm64__) /* aarch64 */ \
-     || defined(__PPC64__) || defined(__PPC64LE__) || defined(__ppc64__) || defined(__powerpc64__) /* ppc64 */ \
-     || defined(__mips64__) || defined(__mips64)) /* mips64 */ \
-   || (!defined(SIZE_MAX) || SIZE_MAX < ULLONG_MAX) /* check limits */
-#    define XXH_REROLL_XXH64 1
-#  else
-#    define XXH_REROLL_XXH64 0
-#  endif
-#endif /* !defined(XXH_REROLL_XXH64) */
+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 f_acc,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
 
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
-/*
- * Manual byteshift. Best for old compilers which don't inline memcpy.
- * We actually directly use XXH_readLE64 and XXH_readBE64.
- */
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble);
 
-/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
-static xxh_u64 XXH_read64(const void* memPtr) { return *(const xxh_u64*) memPtr; }
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    return XXH3_finalizeLong_128b(acc, secret, secretSize, (xxh_u64)len);
+}
 
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+/*
+ * It's important for performance that XXH3_hashLong() is not inlined.
+ */
+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, XXH3_scrambleAcc);
+}
 
 /*
- * __pack instructions are safer, but compiler specific, hence potentially
- * problematic for some compilers.
+ * 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.
  *
- * Currently only defined for GCC and ICC.
+ * 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.
  */
-typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
-static xxh_u64 XXH_read64(const void* ptr) { return ((const unalign64*)ptr)->u64; }
+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, XXH3_scrambleAcc);
+}
 
-#else
+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 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_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_acc, f_scramble);
+    }
+}
 
 /*
- * Portable and safe solution. Generally efficient.
- * see: https://stackoverflow.com/a/32095106/646947
+ * It's important for performance that XXH3_hashLong is not inlined.
  */
-static xxh_u64 XXH_read64(const void* memPtr)
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed(const void* input, size_t len,
+                            XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen)
 {
-    xxh_u64 val;
-    memcpy(&val, memPtr, sizeof(val));
-    return val;
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
 }
 
-#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
+                                            XXH64_hash_t, const void* XXH_RESTRICT, size_t);
 
-#if defined(_MSC_VER)     /* Visual Studio */
-#  define XXH_swap64 _byteswap_uint64
-#elif XXH_GCC_VERSION >= 403
-#  define XXH_swap64 __builtin_bswap64
-#else
-static xxh_u64 XXH_swap64 (xxh_u64 x)
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_128bits_internal(const void* input, size_t len,
+                      XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                      XXH3_hashLong128_f f_hl128)
 {
-    return  ((x << 56) & 0xff00000000000000ULL) |
-            ((x << 40) & 0x00ff000000000000ULL) |
-            ((x << 24) & 0x0000ff0000000000ULL) |
-            ((x << 8)  & 0x000000ff00000000ULL) |
-            ((x >> 8)  & 0x00000000ff000000ULL) |
-            ((x >> 24) & 0x0000000000ff0000ULL) |
-            ((x >> 40) & 0x000000000000ff00ULL) |
-            ((x >> 56) & 0x00000000000000ffULL);
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secret` conditions are not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hl128(input, len, seed64, secret, secretLen);
 }
-#endif
 
 
-/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/* ===   Public XXH128 API   === */
 
-XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len)
 {
-    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
-    return bytePtr[0]
-         | ((xxh_u64)bytePtr[1] << 8)
-         | ((xxh_u64)bytePtr[2] << 16)
-         | ((xxh_u64)bytePtr[3] << 24)
-         | ((xxh_u64)bytePtr[4] << 32)
-         | ((xxh_u64)bytePtr[5] << 40)
-         | ((xxh_u64)bytePtr[6] << 48)
-         | ((xxh_u64)bytePtr[7] << 56);
+    return XXH3_128bits_internal(input, len, 0,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_default);
 }
 
-XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
-    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
-    return bytePtr[7]
-         | ((xxh_u64)bytePtr[6] << 8)
-         | ((xxh_u64)bytePtr[5] << 16)
-         | ((xxh_u64)bytePtr[4] << 24)
-         | ((xxh_u64)bytePtr[3] << 32)
-         | ((xxh_u64)bytePtr[2] << 40)
-         | ((xxh_u64)bytePtr[1] << 48)
-         | ((xxh_u64)bytePtr[0] << 56);
+    return XXH3_128bits_internal(input, len, 0,
+                                 (const xxh_u8*)secret, secretSize,
+                                 XXH3_hashLong_128b_withSecret);
 }
 
-#else
-XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+    return XXH3_128bits_internal(input, len, seed,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_withSeed);
 }
 
-static xxh_u64 XXH_readBE64(const void* ptr)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
 {
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+    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);
 }
-#endif
 
-XXH_FORCE_INLINE xxh_u64
-XXH_readLE64_align(const void* ptr, XXH_alignment align)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
-    if (align==XXH_unaligned)
-        return XXH_readLE64(ptr);
-    else
-        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
+    return XXH3_128bits_withSeed(input, len, seed);
 }
 
 
-/*******   xxh64   *******/
-
-static const xxh_u64 PRIME64_1 = 0x9E3779B185EBCA87ULL;   /* 0b1001111000110111011110011011000110000101111010111100101010000111 */
-static const xxh_u64 PRIME64_2 = 0xC2B2AE3D27D4EB4FULL;   /* 0b1100001010110010101011100011110100100111110101001110101101001111 */
-static const xxh_u64 PRIME64_3 = 0x165667B19E3779F9ULL;   /* 0b0001011001010110011001111011000110011110001101110111100111111001 */
-static const xxh_u64 PRIME64_4 = 0x85EBCA77C2B2AE63ULL;   /* 0b1000010111101011110010100111011111000010101100101010111001100011 */
-static const xxh_u64 PRIME64_5 = 0x27D4EB2F165667C5ULL;   /* 0b0010011111010100111010110010111100010110010101100110011111000101 */
+/* ===   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.
+ */
 
-static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
 {
-    acc += input * PRIME64_2;
-    acc  = XXH_rotl64(acc, 31);
-    acc *= PRIME64_1;
-    return acc;
+    return XXH3_64bits_reset(statePtr);
 }
 
-static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
-    val  = XXH64_round(0, val);
-    acc ^= val;
-    acc  = acc * PRIME64_1 + PRIME64_4;
-    return acc;
+    return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
 }
 
-static xxh_u64 XXH64_avalanche(xxh_u64 h64)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
 {
-    h64 ^= h64 >> 33;
-    h64 *= PRIME64_2;
-    h64 ^= h64 >> 29;
-    h64 *= PRIME64_3;
-    h64 ^= h64 >> 32;
-    return h64;
+    return XXH3_64bits_reset_withSeed(statePtr, seed);
 }
 
-
-#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)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
 {
-#define PROCESS1_64 do {                                   \
-    h64 ^= (*ptr++) * PRIME64_5;                           \
-    h64 = XXH_rotl64(h64, 11) * PRIME64_1;                 \
-} while (0)
-
-#define PROCESS4_64 do {                                   \
-    h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * PRIME64_1;      \
-    ptr += 4;                                              \
-    h64 = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3;     \
-} while (0)
-
-#define PROCESS8_64 do {                                   \
-    xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr)); \
-    ptr += 8;                                              \
-    h64 ^= k1;                                             \
-    h64  = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4;     \
-} while (0)
-
-    /* Rerolled version for 32-bit targets is faster and much smaller. */
-    if (XXH_REROLL || XXH_REROLL_XXH64) {
-        len &= 31;
-        while (len >= 8) {
-            PROCESS8_64;
-            len -= 8;
-        }
-        if (len >= 4) {
-            PROCESS4_64;
-            len -= 4;
-        }
-        while (len > 0) {
-            PROCESS1_64;
-            --len;
-        }
-         return  XXH64_avalanche(h64);
-    } else {
-        switch(len & 31) {
-           case 24: PROCESS8_64;
-                         /* fallthrough */
-           case 16: PROCESS8_64;
-                         /* fallthrough */
-           case  8: PROCESS8_64;
-                    return XXH64_avalanche(h64);
-
-           case 28: PROCESS8_64;
-                         /* fallthrough */
-           case 20: PROCESS8_64;
-                         /* fallthrough */
-           case 12: PROCESS8_64;
-                         /* fallthrough */
-           case  4: PROCESS4_64;
-                    return XXH64_avalanche(h64);
-
-           case 25: PROCESS8_64;
-                         /* fallthrough */
-           case 17: PROCESS8_64;
-                         /* fallthrough */
-           case  9: PROCESS8_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 29: PROCESS8_64;
-                         /* fallthrough */
-           case 21: PROCESS8_64;
-                         /* fallthrough */
-           case 13: PROCESS8_64;
-                         /* fallthrough */
-           case  5: PROCESS4_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 26: PROCESS8_64;
-                         /* fallthrough */
-           case 18: PROCESS8_64;
-                         /* fallthrough */
-           case 10: PROCESS8_64;
-                    PROCESS1_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 30: PROCESS8_64;
-                         /* fallthrough */
-           case 22: PROCESS8_64;
-                         /* fallthrough */
-           case 14: PROCESS8_64;
-                         /* fallthrough */
-           case  6: PROCESS4_64;
-                    PROCESS1_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 27: PROCESS8_64;
-                         /* fallthrough */
-           case 19: PROCESS8_64;
-                         /* fallthrough */
-           case 11: PROCESS8_64;
-                    PROCESS1_64;
-                    PROCESS1_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 31: PROCESS8_64;
-                         /* fallthrough */
-           case 23: PROCESS8_64;
-                         /* fallthrough */
-           case 15: PROCESS8_64;
-                         /* fallthrough */
-           case  7: PROCESS4_64;
-                         /* fallthrough */
-           case  3: PROCESS1_64;
-                         /* fallthrough */
-           case  2: PROCESS1_64;
-                         /* fallthrough */
-           case  1: PROCESS1_64;
-                         /* fallthrough */
-           case  0: return XXH64_avalanche(h64);
-        }
-    }
-    /* impossible to reach */
-    XXH_ASSERT(0);
-    return 0;  /* unreachable, but some compilers complain without it */
+    return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed);
 }
 
-XXH_FORCE_INLINE xxh_u64
-XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
-    const xxh_u8* bEnd = input + len;
-    xxh_u64 h64;
-
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-    if (input==NULL) {
-        len=0;
-        bEnd=input=(const xxh_u8*)(size_t)32;
-    }
-#endif
-
-    if (len>=32) {
-        const xxh_u8* const limit = bEnd - 32;
-        xxh_u64 v1 = seed + PRIME64_1 + PRIME64_2;
-        xxh_u64 v2 = seed + PRIME64_2;
-        xxh_u64 v3 = seed + 0;
-        xxh_u64 v4 = seed - PRIME64_1;
-
-        do {
-            v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
-            v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
-            v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
-            v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
-        } while (input<=limit);
-
-        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
-        h64 = XXH64_mergeRound(h64, v1);
-        h64 = XXH64_mergeRound(h64, v2);
-        h64 = XXH64_mergeRound(h64, v3);
-        h64 = XXH64_mergeRound(h64, v4);
+    return XXH3_64bits_update(state, input, len);
+}
 
-    } else {
-        h64  = seed + PRIME64_5;
+/*! @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) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+        return XXH3_finalizeLong_128b(acc, secret, state->secretLimit + XXH_STRIPE_LEN,  (xxh_u64)state->totalLen);
     }
-
-    h64 += (xxh_u64) len;
-
-    return XXH64_finalize(h64, input, len, align);
+    /* len <= XXH3_MIDSIZE_MAX : short code */
+    if (state->useSeed)
+        return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    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 */
 
-XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t seed)
+/* return : 1 is equal, 0 if different */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
 {
-#if 0
-    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
-    XXH64_state_t state;
-    XXH64_reset(&state, seed);
-    XXH64_update(&state, (const xxh_u8*)input, len);
-    return XXH64_digest(&state);
-
-#else
-
-    if (XXH_FORCE_ALIGN_CHECK) {
-        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
-            return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
-    }   }
-
-    return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
-
-#endif
+    /* note : XXH128_hash_t is compact, it has no padding byte */
+    return !(memcmp(&h1, &h2, sizeof(h1)));
 }
 
-/*******   Hash Streaming   *******/
-
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+/* This prototype is compatible with stdlib's qsort().
+ * @return : >0 if *h128_1  > *h128_2
+ *           <0 if *h128_1  < *h128_2
+ *           =0 if *h128_1 == *h128_2  */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2)
 {
-    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+    XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
+    XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
+    int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
+    /* note : bets that, in most cases, hash values are different */
+    if (hcmp) return hcmp;
+    return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
 }
-XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+
+
+/*======   Canonical representation   ======*/
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash)
 {
-    XXH_free(statePtr);
-    return XXH_OK;
+    XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) {
+        hash.high64 = XXH_swap64(hash.high64);
+        hash.low64  = XXH_swap64(hash.low64);
+    }
+    XXH_memcpy(dst, &hash.high64, sizeof(hash.high64));
+    XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
 }
 
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t* srcState)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src)
 {
-    memcpy(dstState, srcState, sizeof(*dstState));
+    XXH128_hash_t h;
+    h.high64 = XXH_readBE64(src);
+    h.low64  = XXH_readBE64(src->digest + 8);
+    return h;
 }
 
-XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed)
+
+
+/* ==========================================
+ * Secret generators
+ * ==========================================
+ */
+#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
+
+XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
 {
-    XXH64_state_t state;   /* use a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state));
-    state.v1 = seed + PRIME64_1 + PRIME64_2;
-    state.v2 = seed + PRIME64_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - PRIME64_1;
-     /* do not write into reserved64, might be removed in a future version */
-    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved64));
-    return XXH_OK;
+    XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 );
+    XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
 }
 
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH64_update (XXH64_state_t* state, const void* input, size_t len)
+XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize)
 {
-    if (input==NULL)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-        return XXH_OK;
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(secretBuffer != NULL);
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
 #else
-        return XXH_ERROR;
+    /* production mode, assert() are disabled */
+    if (secretBuffer == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
 #endif
 
-    {   const xxh_u8* p = (const xxh_u8*)input;
-        const xxh_u8* const bEnd = p + len;
-
-        state->total_len += len;
-
-        if (state->memsize + len < 32) {  /* fill in tmp buffer */
-            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
-            state->memsize += (xxh_u32)len;
-            return XXH_OK;
-        }
-
-        if (state->memsize) {   /* tmp buffer is full */
-            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
-            state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0));
-            state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1));
-            state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2));
-            state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3));
-            p += 32-state->memsize;
-            state->memsize = 0;
-        }
-
-        if (p+32 <= bEnd) {
-            const xxh_u8* const limit = bEnd - 32;
-            xxh_u64 v1 = state->v1;
-            xxh_u64 v2 = state->v2;
-            xxh_u64 v3 = state->v3;
-            xxh_u64 v4 = state->v4;
+    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
 
-            do {
-                v1 = XXH64_round(v1, XXH_readLE64(p)); p+=8;
-                v2 = XXH64_round(v2, XXH_readLE64(p)); p+=8;
-                v3 = XXH64_round(v3, XXH_readLE64(p)); p+=8;
-                v4 = XXH64_round(v4, XXH_readLE64(p)); p+=8;
-            } while (p<=limit);
-
-            state->v1 = v1;
-            state->v2 = v2;
-            state->v3 = v3;
-            state->v4 = v4;
-        }
+    /* Fill secretBuffer with a copy of customSeed - repeat as needed */
+    {   size_t pos = 0;
+        while (pos < secretSize) {
+            size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize);
+            memcpy((char*)secretBuffer + pos, customSeed, toCopy);
+            pos += toCopy;
+    }   }
 
-        if (p < bEnd) {
-            XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
-            state->memsize = (unsigned)(bEnd-p);
+    {   size_t const nbSeg16 = secretSize / 16;
+        size_t n;
+        XXH128_canonical_t scrambler;
+        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
+        for (n=0; n<nbSeg16; n++) {
+            XXH128_hash_t const h128 = XXH128(&scrambler, sizeof(scrambler), n);
+            XXH3_combine16((char*)secretBuffer + n*16, h128);
         }
+        /* last segment */
+        XXH3_combine16((char*)secretBuffer + secretSize - 16, XXH128_hashFromCanonical(&scrambler));
     }
-
     return XXH_OK;
 }
 
-
-XXH_PUBLIC_API XXH64_hash_t XXH64_digest (const XXH64_state_t* state)
-{
-    xxh_u64 h64;
-
-    if (state->total_len >= 32) {
-        xxh_u64 const v1 = state->v1;
-        xxh_u64 const v2 = state->v2;
-        xxh_u64 const v3 = state->v3;
-        xxh_u64 const v4 = state->v4;
-
-        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
-        h64 = XXH64_mergeRound(h64, v1);
-        h64 = XXH64_mergeRound(h64, v2);
-        h64 = XXH64_mergeRound(h64, v3);
-        h64 = XXH64_mergeRound(h64, v4);
-    } else {
-        h64  = state->v3 /*seed*/ + PRIME64_5;
-    }
-
-    h64 += (xxh_u64) state->total_len;
-
-    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
-}
-
-
-/******* Canonical representation   *******/
-
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash)
-{
-    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
-    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
-    memcpy(dst, &hash, sizeof(*dst));
-}
-
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed)
 {
-    return XXH_readBE64(src);
+    XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+    XXH3_initCustomSecret(secret, seed);
+    XXH_ASSERT(secretBuffer != NULL);
+    memcpy(secretBuffer, secret, XXH_SECRET_DEFAULT_SIZE);
 }
 
 
 
-/* *********************************************************************
-*  XXH3
-*  New generation hash designed for speed on small keys and vectorization
-************************************************************************ */
-
-#include "xxh3.h"
-
+/* Pop our optimization override from above */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+#  pragma GCC pop_options
+#endif
 
 #endif  /* XXH_NO_LONG_LONG */
 
+#endif  /* XXH_NO_XXH3 */
 
+/*!
+ * @}
+ */
 #endif  /* XXH_IMPLEMENTATION */
 
 
 #if defined (__cplusplus)
-}
+} /* extern "C" */
 #endif
diff --git a/vcpkg.json b/vcpkg.json
index 68968b4702..f356de5ca4 100644
--- a/vcpkg.json
+++ b/vcpkg.json
@@ -1,4 +1,14 @@
 {
+	"vcpkg-configuration": {
+		"default-registry": {
+			"kind": "git",
+			"repository": "https://github.com/Microsoft/vcpkg",
+			"baseline": "dd3097e305afa53f7b4312371f62058d2e665320",
+			"reference": "dd3097e305afa53f7b4312371f62058d2e665320"
+		},
+		"overlay-ports": [ "./cmake/overlay-ports" ],
+		"overlay-triplets": [ "./cmake/overlay-triplets" ]
+	},
 	"default-features": ["qt6"],
 	"dependencies": [
 		"sdl2",
@@ -18,7 +28,9 @@
 			"name": "libslirp",
 			"platform": "linux"
 		},
-		"faad2"
+		"faad2",
+		"pkgconf",
+		{ "name": "dirent", "platform": "windows" }
 	],
 	"features": {
 		"qt6": {