Adding upstream version 86.0.1.upstream/86.0.1upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

11 files changed, 12840 insertions, 0 deletions

diff --git a/mfbt/lz4/LICENSE b/mfbt/lz4/LICENSE
new file mode 100644
index 0000000000..74c2cdd7d5
--- /dev/null
+++ b/mfbt/lz4/LICENSE
@@ -0,0 +1,24 @@
+LZ4 Library
+Copyright (c) 2011-2016, Yann Collet
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this
+  list of conditions and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright notice, this
+  list of conditions and the following disclaimer in the documentation and/or
+  other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/mfbt/lz4/README.md b/mfbt/lz4/README.md
new file mode 100644
index 0000000000..e2af868ff4
--- /dev/null
+++ b/mfbt/lz4/README.md
@@ -0,0 +1,137 @@
+LZ4 - Library Files
+================================
+
+The `/lib` directory contains many files, but depending on project's objectives,
+not all of them are necessary.
+
+#### Minimal LZ4 build
+
+The minimum required is **`lz4.c`** and **`lz4.h`**,
+which provides the fast compression and decompression algorithms.
+They generate and decode data using the [LZ4 block format].
+
+
+#### High Compression variant
+
+For more compression ratio at the cost of compression speed,
+the High Compression variant called **lz4hc** is available.
+Add files **`lz4hc.c`** and **`lz4hc.h`**.
+This variant also compresses data using the [LZ4 block format],
+and depends on regular `lib/lz4.*` source files.
+
+
+#### Frame support, for interoperability
+
+In order to produce compressed data compatible with `lz4` command line utility,
+it's necessary to use the [official interoperable frame format].
+This format is generated and decoded automatically by the **lz4frame** library.
+Its public API is described in `lib/lz4frame.h`.
+In order to work properly, lz4frame needs all other modules present in `/lib`,
+including, lz4 and lz4hc, and also **xxhash**.
+So it's necessary to include all `*.c` and `*.h` files present in `/lib`.
+
+
+#### Advanced / Experimental API
+
+Definitions which are not guaranteed to remain stable in future versions,
+are protected behind macros, such as `LZ4_STATIC_LINKING_ONLY`.
+As the name strongly implies, these definitions should only be invoked
+in the context of static linking ***only***.
+Otherwise, dependent application may fail on API or ABI break in the future.
+The associated symbols are also not exposed by the dynamic library by default.
+Should they be nonetheless needed, it's possible to force their publication
+by using build macros `LZ4_PUBLISH_STATIC_FUNCTIONS`
+and `LZ4F_PUBLISH_STATIC_FUNCTIONS`.
+
+
+#### Build macros
+
+The following build macro can be selected to adjust source code behavior at compilation time :
+
+- `LZ4_FAST_DEC_LOOP` : this triggers a speed optimized decompression loop, more powerful on modern cpus.
+  This loop works great on `x86`, `x64` and `aarch64` cpus, and is automatically enabled for them.
+  It's also possible to enable or disable it manually, by passing `LZ4_FAST_DEC_LOOP=1` or `0` to the preprocessor.
+  For example, with `gcc` : `-DLZ4_FAST_DEC_LOOP=1`,
+  and with `make` : `CPPFLAGS+=-DLZ4_FAST_DEC_LOOP=1 make lz4`.
+
+- `LZ4_DISTANCE_MAX` : control the maximum offset that the compressor will allow.
+  Set to 65535 by default, which is the maximum value supported by lz4 format.
+  Reducing maximum distance will reduce opportunities for LZ4 to find matches,
+  hence will produce a worse compression ratio.
+  However, a smaller max distance can allow compatibility with specific decoders using limited memory budget.
+  This build macro only influences the compressed output of the compressor.
+
+- `LZ4_DISABLE_DEPRECATE_WARNINGS` : invoking a deprecated function will make the compiler generate a warning.
+  This is meant to invite users to update their source code.
+  Should this be a problem, it's generally possible to make the compiler ignore these warnings,
+  for example with `-Wno-deprecated-declarations` on `gcc`,
+  or `_CRT_SECURE_NO_WARNINGS` for Visual Studio.
+  This build macro offers another project-specific method
+  by defining `LZ4_DISABLE_DEPRECATE_WARNINGS` before including the LZ4 header files.
+
+- `LZ4_USER_MEMORY_FUNCTIONS` : replace calls to <stdlib>'s `malloc`, `calloc` and `free`
+  by user-defined functions, which must be called `LZ4_malloc()`, `LZ4_calloc()` and `LZ4_free()`.
+  User functions must be available at link time.
+
+- `LZ4_FORCE_SW_BITCOUNT` : by default, the compression algorithm tries to determine lengths
+  by using bitcount instructions, generally implemented as fast single instructions in many cpus.
+  In case the target cpus doesn't support it, or compiler intrinsic doesn't work, or feature bad performance,
+  it's possible to use an optimized software path instead.
+  This is achieved by setting this build macros .
+  In most cases, it's not expected to be necessary,
+  but it can be legitimately considered for less common platforms.
+
+- `LZ4_ALIGN_TEST` : alignment test ensures that the memory area
+  passed as argument to become a compression state is suitably aligned.
+  This test can be disabled if it proves flaky, by setting this value to 0.
+
+
+#### Amalgamation
+
+lz4 source code can be amalgamated into a single file.
+One can combine all source code into `lz4_all.c` by using following command:
+```
+cat lz4.c lz4hc.c lz4frame.c > lz4_all.c
+```
+(`cat` file order is important) then compile `lz4_all.c`.
+All `*.h` files present in `/lib` remain necessary to compile `lz4_all.c`.
+
+
+#### Windows : using MinGW+MSYS to create DLL
+
+DLL can be created using MinGW+MSYS with the `make liblz4` command.
+This command creates `dll\liblz4.dll` and the import library `dll\liblz4.lib`.
+To override the `dlltool` command  when cross-compiling on Linux, just set the `DLLTOOL` variable. Example of cross compilation on Linux with mingw-w64 64 bits:
+```
+make BUILD_STATIC=no CC=x86_64-w64-mingw32-gcc DLLTOOL=x86_64-w64-mingw32-dlltool OS=Windows_NT
+```
+The import library is only required with Visual C++.
+The header files `lz4.h`, `lz4hc.h`, `lz4frame.h` and the dynamic library
+`dll\liblz4.dll` are required to compile a project using gcc/MinGW.
+The dynamic library has to be added to linking options.
+It means that if a project that uses LZ4 consists of a single `test-dll.c`
+file it should be linked with `dll\liblz4.dll`. For example:
+```
+    $(CC) $(CFLAGS) -Iinclude/ test-dll.c -o test-dll dll\liblz4.dll
+```
+The compiled executable will require LZ4 DLL which is available at `dll\liblz4.dll`.
+
+
+#### Miscellaneous
+
+Other files present in the directory are not source code. They are :
+
+ - `LICENSE` : contains the BSD license text
+ - `Makefile` : `make` script to compile and install lz4 library (static and dynamic)
+ - `liblz4.pc.in` : for `pkg-config` (used in `make install`)
+ - `README.md` : this file
+
+[official interoperable frame format]: ../doc/lz4_Frame_format.md
+[LZ4 block format]: ../doc/lz4_Block_format.md
+
+
+#### License
+
+All source material within __lib__ directory are BSD 2-Clause licensed.
+See [LICENSE](LICENSE) for details.
+The license is also reminded at the top of each source file.
diff --git a/mfbt/lz4/lz4.c b/mfbt/lz4/lz4.c
new file mode 100644
index 0000000000..9f5e9bfa08
--- /dev/null
+++ b/mfbt/lz4/lz4.c
@@ -0,0 +1,2495 @@
+/*
+   LZ4 - Fast LZ compression algorithm
+   Copyright (C) 2011-present, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+    - LZ4 homepage : http://www.lz4.org
+    - LZ4 source repository : https://github.com/lz4/lz4
+*/
+
+/*-************************************
+*  Tuning parameters
+**************************************/
+/*
+ * LZ4_HEAPMODE :
+ * Select how default compression functions will allocate memory for their hash table,
+ * in memory stack (0:default, fastest), or in memory heap (1:requires malloc()).
+ */
+#ifndef LZ4_HEAPMODE
+#  define LZ4_HEAPMODE 0
+#endif
+
+/*
+ * LZ4_ACCELERATION_DEFAULT :
+ * Select "acceleration" for LZ4_compress_fast() when parameter value <= 0
+ */
+#define LZ4_ACCELERATION_DEFAULT 1
+/*
+ * LZ4_ACCELERATION_MAX :
+ * Any "acceleration" value higher than this threshold
+ * get treated as LZ4_ACCELERATION_MAX instead (fix #876)
+ */
+#define LZ4_ACCELERATION_MAX 65537
+
+
+/*-************************************
+*  CPU Feature Detection
+**************************************/
+/* LZ4_FORCE_MEMORY_ACCESS
+ * 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 to select different access method for improved performance.
+ * Method 0 (default) : use `memcpy()`. Safe and portable.
+ * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
+ *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ *            It can generate buggy code on targets which assembly generation depends on alignment.
+ *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
+ * See https://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
+ * Prefer these methods in priority order (0 > 1 > 2)
+ */
+#ifndef LZ4_FORCE_MEMORY_ACCESS   /* can be defined externally */
+#  if defined(__GNUC__) && \
+  ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) \
+  || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
+#    define LZ4_FORCE_MEMORY_ACCESS 2
+#  elif (defined(__INTEL_COMPILER) && !defined(_WIN32)) || defined(__GNUC__)
+#    define LZ4_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+/*
+ * LZ4_FORCE_SW_BITCOUNT
+ * Define this parameter if your target system or compiler does not support hardware bit count
+ */
+#if defined(_MSC_VER) && defined(_WIN32_WCE)   /* Visual Studio for WinCE doesn't support Hardware bit count */
+#  undef  LZ4_FORCE_SW_BITCOUNT  /* avoid double def */
+#  define LZ4_FORCE_SW_BITCOUNT
+#endif
+
+
+
+/*-************************************
+*  Dependency
+**************************************/
+/*
+ * LZ4_SRC_INCLUDED:
+ * Amalgamation flag, whether lz4.c is included
+ */
+#ifndef LZ4_SRC_INCLUDED
+#  define LZ4_SRC_INCLUDED 1
+#endif
+
+#ifndef LZ4_STATIC_LINKING_ONLY
+#define LZ4_STATIC_LINKING_ONLY
+#endif
+
+#ifndef LZ4_DISABLE_DEPRECATE_WARNINGS
+#define LZ4_DISABLE_DEPRECATE_WARNINGS /* due to LZ4_decompress_safe_withPrefix64k */
+#endif
+
+#define LZ4_STATIC_LINKING_ONLY  /* LZ4_DISTANCE_MAX */
+#include "lz4.h"
+/* see also "memory routines" below */
+
+
+/*-************************************
+*  Compiler Options
+**************************************/
+#if defined(_MSC_VER) && (_MSC_VER >= 1400)  /* Visual Studio 2005+ */
+#  include <intrin.h>               /* only present in VS2005+ */
+#  pragma warning(disable : 4127)   /* disable: C4127: conditional expression is constant */
+#endif  /* _MSC_VER */
+
+#ifndef LZ4_FORCE_INLINE
+#  ifdef _MSC_VER    /* Visual Studio */
+#    define LZ4_FORCE_INLINE static __forceinline
+#  else
+#    if defined (__cplusplus) || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
+#      ifdef __GNUC__
+#        define LZ4_FORCE_INLINE static inline __attribute__((always_inline))
+#      else
+#        define LZ4_FORCE_INLINE static inline
+#      endif
+#    else
+#      define LZ4_FORCE_INLINE static
+#    endif /* __STDC_VERSION__ */
+#  endif  /* _MSC_VER */
+#endif /* LZ4_FORCE_INLINE */
+
+/* LZ4_FORCE_O2 and LZ4_FORCE_INLINE
+ * gcc on ppc64le generates an unrolled SIMDized loop for LZ4_wildCopy8,
+ * together with a simple 8-byte copy loop as a fall-back path.
+ * However, this optimization hurts the decompression speed by >30%,
+ * because the execution does not go to the optimized loop
+ * for typical compressible data, and all of the preamble checks
+ * before going to the fall-back path become useless overhead.
+ * This optimization happens only with the -O3 flag, and -O2 generates
+ * a simple 8-byte copy loop.
+ * With gcc on ppc64le, all of the LZ4_decompress_* and LZ4_wildCopy8
+ * functions are annotated with __attribute__((optimize("O2"))),
+ * and also LZ4_wildCopy8 is forcibly inlined, so that the O2 attribute
+ * of LZ4_wildCopy8 does not affect the compression speed.
+ */
+#if defined(__PPC64__) && defined(__LITTLE_ENDIAN__) && defined(__GNUC__) && !defined(__clang__)
+#  define LZ4_FORCE_O2  __attribute__((optimize("O2")))
+#  undef LZ4_FORCE_INLINE
+#  define LZ4_FORCE_INLINE  static __inline __attribute__((optimize("O2"),always_inline))
+#else
+#  define LZ4_FORCE_O2
+#endif
+
+#if (defined(__GNUC__) && (__GNUC__ >= 3)) || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) || defined(__clang__)
+#  define expect(expr,value)    (__builtin_expect ((expr),(value)) )
+#else
+#  define expect(expr,value)    (expr)
+#endif
+
+#ifndef likely
+#define likely(expr)     expect((expr) != 0, 1)
+#endif
+#ifndef unlikely
+#define unlikely(expr)   expect((expr) != 0, 0)
+#endif
+
+/* Should the alignment test prove unreliable, for some reason,
+ * it can be disabled by setting LZ4_ALIGN_TEST to 0 */
+#ifndef LZ4_ALIGN_TEST  /* can be externally provided */
+# define LZ4_ALIGN_TEST 1
+#endif
+
+
+/*-************************************
+*  Memory routines
+**************************************/
+#ifdef LZ4_USER_MEMORY_FUNCTIONS
+/* memory management functions can be customized by user project.
+ * Below functions must exist somewhere in the Project
+ * and be available at link time */
+void* LZ4_malloc(size_t s);
+void* LZ4_calloc(size_t n, size_t s);
+void  LZ4_free(void* p);
+# define ALLOC(s)          LZ4_malloc(s)
+# define ALLOC_AND_ZERO(s) LZ4_calloc(1,s)
+# define FREEMEM(p)        LZ4_free(p)
+#else
+# include <stdlib.h>   /* malloc, calloc, free */
+# define ALLOC(s)          malloc(s)
+# define ALLOC_AND_ZERO(s) calloc(1,s)
+# define FREEMEM(p)        free(p)
+#endif
+
+#include <string.h>   /* memset, memcpy */
+#define MEM_INIT(p,v,s)   memset((p),(v),(s))
+
+
+/*-************************************
+*  Common Constants
+**************************************/
+#define MINMATCH 4
+
+#define WILDCOPYLENGTH 8
+#define LASTLITERALS   5   /* see ../doc/lz4_Block_format.md#parsing-restrictions */
+#define MFLIMIT       12   /* see ../doc/lz4_Block_format.md#parsing-restrictions */
+#define MATCH_SAFEGUARD_DISTANCE  ((2*WILDCOPYLENGTH) - MINMATCH)   /* ensure it's possible to write 2 x wildcopyLength without overflowing output buffer */
+#define FASTLOOP_SAFE_DISTANCE 64
+static const int LZ4_minLength = (MFLIMIT+1);
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define LZ4_DISTANCE_ABSOLUTE_MAX 65535
+#if (LZ4_DISTANCE_MAX > LZ4_DISTANCE_ABSOLUTE_MAX)   /* max supported by LZ4 format */
+#  error "LZ4_DISTANCE_MAX is too big : must be <= 65535"
+#endif
+
+#define ML_BITS  4
+#define ML_MASK  ((1U<<ML_BITS)-1)
+#define RUN_BITS (8-ML_BITS)
+#define RUN_MASK ((1U<<RUN_BITS)-1)
+
+
+/*-************************************
+*  Error detection
+**************************************/
+#if defined(LZ4_DEBUG) && (LZ4_DEBUG>=1)
+#  include <assert.h>
+#else
+#  ifndef assert
+#    define assert(condition) ((void)0)
+#  endif
+#endif
+
+#define LZ4_STATIC_ASSERT(c)   { enum { LZ4_static_assert = 1/(int)(!!(c)) }; }   /* use after variable declarations */
+
+#if defined(LZ4_DEBUG) && (LZ4_DEBUG>=2)
+#  include <stdio.h>
+   static int g_debuglog_enable = 1;
+#  define DEBUGLOG(l, ...) {                          \
+        if ((g_debuglog_enable) && (l<=LZ4_DEBUG)) {  \
+            fprintf(stderr, __FILE__ ": ");           \
+            fprintf(stderr, __VA_ARGS__);             \
+            fprintf(stderr, " \n");                   \
+    }   }
+#else
+#  define DEBUGLOG(l, ...) {}    /* disabled */
+#endif
+
+static int LZ4_isAligned(const void* ptr, size_t alignment)
+{
+    return ((size_t)ptr & (alignment -1)) == 0;
+}
+
+
+/*-************************************
+*  Types
+**************************************/
+#include <limits.h>
+#if defined(__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# include <stdint.h>
+  typedef  uint8_t BYTE;
+  typedef uint16_t U16;
+  typedef uint32_t U32;
+  typedef  int32_t S32;
+  typedef uint64_t U64;
+  typedef uintptr_t uptrval;
+#else
+# if UINT_MAX != 4294967295UL
+#   error "LZ4 code (when not C++ or C99) assumes that sizeof(int) == 4"
+# endif
+  typedef unsigned char       BYTE;
+  typedef unsigned short      U16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+  typedef size_t              uptrval;   /* generally true, except OpenVMS-64 */
+#endif
+
+#if defined(__x86_64__)
+  typedef U64    reg_t;   /* 64-bits in x32 mode */
+#else
+  typedef size_t reg_t;   /* 32-bits in x32 mode */
+#endif
+
+typedef enum {
+    notLimited = 0,
+    limitedOutput = 1,
+    fillOutput = 2
+} limitedOutput_directive;
+
+
+/*-************************************
+*  Reading and writing into memory
+**************************************/
+
+/**
+ * LZ4 relies on memcpy with a constant size being inlined. In freestanding
+ * environments, the compiler can't assume the implementation of memcpy() is
+ * standard compliant, so it can't apply its specialized memcpy() inlining
+ * logic. When possible, use __builtin_memcpy() to tell the compiler to analyze
+ * memcpy() as if it were standard compliant, so it can inline it in freestanding
+ * environments. This is needed when decompressing the Linux Kernel, for example.
+ */
+#if defined(__GNUC__) && (__GNUC__ >= 4)
+#define LZ4_memcpy(dst, src, size) __builtin_memcpy(dst, src, size)
+#else
+#define LZ4_memcpy(dst, src, size) memcpy(dst, src, size)
+#endif
+
+static unsigned LZ4_isLittleEndian(void)
+{
+    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental */
+    return one.c[0];
+}
+
+
+#if defined(LZ4_FORCE_MEMORY_ACCESS) && (LZ4_FORCE_MEMORY_ACCESS==2)
+/* lie to the compiler about data alignment; use with caution */
+
+static U16 LZ4_read16(const void* memPtr) { return *(const U16*) memPtr; }
+static U32 LZ4_read32(const void* memPtr) { return *(const U32*) memPtr; }
+static reg_t LZ4_read_ARCH(const void* memPtr) { return *(const reg_t*) memPtr; }
+
+static void LZ4_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+static void LZ4_write32(void* memPtr, U32 value) { *(U32*)memPtr = value; }
+
+#elif defined(LZ4_FORCE_MEMORY_ACCESS) && (LZ4_FORCE_MEMORY_ACCESS==1)
+
+/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
+/* currently only defined for gcc and icc */
+typedef union { U16 u16; U32 u32; reg_t uArch; } __attribute__((packed)) unalign;
+
+static U16 LZ4_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
+static U32 LZ4_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+static reg_t LZ4_read_ARCH(const void* ptr) { return ((const unalign*)ptr)->uArch; }
+
+static void LZ4_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
+static void LZ4_write32(void* memPtr, U32 value) { ((unalign*)memPtr)->u32 = value; }
+
+#else  /* safe and portable access using memcpy() */
+
+static U16 LZ4_read16(const void* memPtr)
+{
+    U16 val; LZ4_memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+static U32 LZ4_read32(const void* memPtr)
+{
+    U32 val; LZ4_memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+static reg_t LZ4_read_ARCH(const void* memPtr)
+{
+    reg_t val; LZ4_memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+static void LZ4_write16(void* memPtr, U16 value)
+{
+    LZ4_memcpy(memPtr, &value, sizeof(value));
+}
+
+static void LZ4_write32(void* memPtr, U32 value)
+{
+    LZ4_memcpy(memPtr, &value, sizeof(value));
+}
+
+#endif /* LZ4_FORCE_MEMORY_ACCESS */
+
+
+static U16 LZ4_readLE16(const void* memPtr)
+{
+    if (LZ4_isLittleEndian()) {
+        return LZ4_read16(memPtr);
+    } else {
+        const BYTE* p = (const BYTE*)memPtr;
+        return (U16)((U16)p[0] + (p[1]<<8));
+    }
+}
+
+static void LZ4_writeLE16(void* memPtr, U16 value)
+{
+    if (LZ4_isLittleEndian()) {
+        LZ4_write16(memPtr, value);
+    } else {
+        BYTE* p = (BYTE*)memPtr;
+        p[0] = (BYTE) value;
+        p[1] = (BYTE)(value>>8);
+    }
+}
+
+/* customized variant of memcpy, which can overwrite up to 8 bytes beyond dstEnd */
+LZ4_FORCE_INLINE
+void LZ4_wildCopy8(void* dstPtr, const void* srcPtr, void* dstEnd)
+{
+    BYTE* d = (BYTE*)dstPtr;
+    const BYTE* s = (const BYTE*)srcPtr;
+    BYTE* const e = (BYTE*)dstEnd;
+
+    do { LZ4_memcpy(d,s,8); d+=8; s+=8; } while (d<e);
+}
+
+static const unsigned inc32table[8] = {0, 1, 2,  1,  0,  4, 4, 4};
+static const int      dec64table[8] = {0, 0, 0, -1, -4,  1, 2, 3};
+
+
+#ifndef LZ4_FAST_DEC_LOOP
+#  if defined __i386__ || defined _M_IX86 || defined __x86_64__ || defined _M_X64
+#    define LZ4_FAST_DEC_LOOP 1
+#  elif defined(__aarch64__) && !defined(__clang__)
+     /* On aarch64, we disable this optimization for clang because on certain
+      * mobile chipsets, performance is reduced with clang. For information
+      * refer to https://github.com/lz4/lz4/pull/707 */
+#    define LZ4_FAST_DEC_LOOP 1
+#  else
+#    define LZ4_FAST_DEC_LOOP 0
+#  endif
+#endif
+
+#if LZ4_FAST_DEC_LOOP
+
+LZ4_FORCE_INLINE void
+LZ4_memcpy_using_offset_base(BYTE* dstPtr, const BYTE* srcPtr, BYTE* dstEnd, const size_t offset)
+{
+    assert(srcPtr + offset == dstPtr);
+    if (offset < 8) {
+        LZ4_write32(dstPtr, 0);   /* silence an msan warning when offset==0 */
+        dstPtr[0] = srcPtr[0];
+        dstPtr[1] = srcPtr[1];
+        dstPtr[2] = srcPtr[2];
+        dstPtr[3] = srcPtr[3];
+        srcPtr += inc32table[offset];
+        LZ4_memcpy(dstPtr+4, srcPtr, 4);
+        srcPtr -= dec64table[offset];
+        dstPtr += 8;
+    } else {
+        LZ4_memcpy(dstPtr, srcPtr, 8);
+        dstPtr += 8;
+        srcPtr += 8;
+    }
+
+    LZ4_wildCopy8(dstPtr, srcPtr, dstEnd);
+}
+
+/* customized variant of memcpy, which can overwrite up to 32 bytes beyond dstEnd
+ * this version copies two times 16 bytes (instead of one time 32 bytes)
+ * because it must be compatible with offsets >= 16. */
+LZ4_FORCE_INLINE void
+LZ4_wildCopy32(void* dstPtr, const void* srcPtr, void* dstEnd)
+{
+    BYTE* d = (BYTE*)dstPtr;
+    const BYTE* s = (const BYTE*)srcPtr;
+    BYTE* const e = (BYTE*)dstEnd;
+
+    do { LZ4_memcpy(d,s,16); LZ4_memcpy(d+16,s+16,16); d+=32; s+=32; } while (d<e);
+}
+
+/* LZ4_memcpy_using_offset()  presumes :
+ * - dstEnd >= dstPtr + MINMATCH
+ * - there is at least 8 bytes available to write after dstEnd */
+LZ4_FORCE_INLINE void
+LZ4_memcpy_using_offset(BYTE* dstPtr, const BYTE* srcPtr, BYTE* dstEnd, const size_t offset)
+{
+    BYTE v[8];
+
+    assert(dstEnd >= dstPtr + MINMATCH);
+
+    switch(offset) {
+    case 1:
+        MEM_INIT(v, *srcPtr, 8);
+        break;
+    case 2:
+        LZ4_memcpy(v, srcPtr, 2);
+        LZ4_memcpy(&v[2], srcPtr, 2);
+        LZ4_memcpy(&v[4], v, 4);
+        break;
+    case 4:
+        LZ4_memcpy(v, srcPtr, 4);
+        LZ4_memcpy(&v[4], srcPtr, 4);
+        break;
+    default:
+        LZ4_memcpy_using_offset_base(dstPtr, srcPtr, dstEnd, offset);
+        return;
+    }
+
+    LZ4_memcpy(dstPtr, v, 8);
+    dstPtr += 8;
+    while (dstPtr < dstEnd) {
+        LZ4_memcpy(dstPtr, v, 8);
+        dstPtr += 8;
+    }
+}
+#endif
+
+
+/*-************************************
+*  Common functions
+**************************************/
+static unsigned LZ4_NbCommonBytes (reg_t val)
+{
+    assert(val != 0);
+    if (LZ4_isLittleEndian()) {
+        if (sizeof(val) == 8) {
+#       if defined(_MSC_VER) && (_MSC_VER >= 1800) && defined(_M_AMD64) && !defined(LZ4_FORCE_SW_BITCOUNT)
+            /* x64 CPUS without BMI support interpret `TZCNT` as `REP BSF` */
+            return (unsigned)_tzcnt_u64(val) >> 3;
+#       elif defined(_MSC_VER) && defined(_WIN64) && !defined(LZ4_FORCE_SW_BITCOUNT)
+            unsigned long r = 0;
+            _BitScanForward64(&r, (U64)val);
+            return (unsigned)r >> 3;
+#       elif (defined(__clang__) || (defined(__GNUC__) && ((__GNUC__ > 3) || \
+                            ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))) && \
+                                        !defined(LZ4_FORCE_SW_BITCOUNT)
+            return (unsigned)__builtin_ctzll((U64)val) >> 3;
+#       else
+            const U64 m = 0x0101010101010101ULL;
+            val ^= val - 1;
+            return (unsigned)(((U64)((val & (m - 1)) * m)) >> 56);
+#       endif
+        } else /* 32 bits */ {
+#       if defined(_MSC_VER) && (_MSC_VER >= 1400) && !defined(LZ4_FORCE_SW_BITCOUNT)
+            unsigned long r;
+            _BitScanForward(&r, (U32)val);
+            return (unsigned)r >> 3;
+#       elif (defined(__clang__) || (defined(__GNUC__) && ((__GNUC__ > 3) || \
+                            ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))) && \
+                        !defined(__TINYC__) && !defined(LZ4_FORCE_SW_BITCOUNT)
+            return (unsigned)__builtin_ctz((U32)val) >> 3;
+#       else
+            const U32 m = 0x01010101;
+            return (unsigned)((((val - 1) ^ val) & (m - 1)) * m) >> 24;
+#       endif
+        }
+    } else   /* Big Endian CPU */ {
+        if (sizeof(val)==8) {
+#       if (defined(__clang__) || (defined(__GNUC__) && ((__GNUC__ > 3) || \
+                            ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))) && \
+                        !defined(__TINYC__) && !defined(LZ4_FORCE_SW_BITCOUNT)
+            return (unsigned)__builtin_clzll((U64)val) >> 3;
+#       else
+#if 1
+            /* this method is probably faster,
+             * but adds a 128 bytes lookup table */
+            static const unsigned char ctz7_tab[128] = {
+                7, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
+                4, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
+                5, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
+                4, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
+                6, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
+                4, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
+                5, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
+                4, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0,
+            };
+            U64 const mask = 0x0101010101010101ULL;
+            U64 const t = (((val >> 8) - mask) | val) & mask;
+            return ctz7_tab[(t * 0x0080402010080402ULL) >> 57];
+#else
+            /* this method doesn't consume memory space like the previous one,
+             * but it contains several branches,
+             * that may end up slowing execution */
+            static const U32 by32 = sizeof(val)*4;  /* 32 on 64 bits (goal), 16 on 32 bits.
+            Just to avoid some static analyzer complaining about shift by 32 on 32-bits target.
+            Note that this code path is never triggered in 32-bits mode. */
+            unsigned r;
+            if (!(val>>by32)) { r=4; } else { r=0; val>>=by32; }
+            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
+            r += (!val);
+            return r;
+#endif
+#       endif
+        } else /* 32 bits */ {
+#       if (defined(__clang__) || (defined(__GNUC__) && ((__GNUC__ > 3) || \
+                            ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))) && \
+                                        !defined(LZ4_FORCE_SW_BITCOUNT)
+            return (unsigned)__builtin_clz((U32)val) >> 3;
+#       else
+            val >>= 8;
+            val = ((((val + 0x00FFFF00) | 0x00FFFFFF) + val) |
+              (val + 0x00FF0000)) >> 24;
+            return (unsigned)val ^ 3;
+#       endif
+        }
+    }
+}
+
+
+#define STEPSIZE sizeof(reg_t)
+LZ4_FORCE_INLINE
+unsigned LZ4_count(const BYTE* pIn, const BYTE* pMatch, const BYTE* pInLimit)
+{
+    const BYTE* const pStart = pIn;
+
+    if (likely(pIn < pInLimit-(STEPSIZE-1))) {
+        reg_t const diff = LZ4_read_ARCH(pMatch) ^ LZ4_read_ARCH(pIn);
+        if (!diff) {
+            pIn+=STEPSIZE; pMatch+=STEPSIZE;
+        } else {
+            return LZ4_NbCommonBytes(diff);
+    }   }
+
+    while (likely(pIn < pInLimit-(STEPSIZE-1))) {
+        reg_t const diff = LZ4_read_ARCH(pMatch) ^ LZ4_read_ARCH(pIn);
+        if (!diff) { pIn+=STEPSIZE; pMatch+=STEPSIZE; continue; }
+        pIn += LZ4_NbCommonBytes(diff);
+        return (unsigned)(pIn - pStart);
+    }
+
+    if ((STEPSIZE==8) && (pIn<(pInLimit-3)) && (LZ4_read32(pMatch) == LZ4_read32(pIn))) { pIn+=4; pMatch+=4; }
+    if ((pIn<(pInLimit-1)) && (LZ4_read16(pMatch) == LZ4_read16(pIn))) { pIn+=2; pMatch+=2; }
+    if ((pIn<pInLimit) && (*pMatch == *pIn)) pIn++;
+    return (unsigned)(pIn - pStart);
+}
+
+
+#ifndef LZ4_COMMONDEFS_ONLY
+/*-************************************
+*  Local Constants
+**************************************/
+static const int LZ4_64Klimit = ((64 KB) + (MFLIMIT-1));
+static const U32 LZ4_skipTrigger = 6;  /* Increase this value ==> compression run slower on incompressible data */
+
+
+/*-************************************
+*  Local Structures and types
+**************************************/
+typedef enum { clearedTable = 0, byPtr, byU32, byU16 } tableType_t;
+
+/**
+ * This enum distinguishes several different modes of accessing previous
+ * content in the stream.
+ *
+ * - noDict        : There is no preceding content.
+ * - withPrefix64k : Table entries up to ctx->dictSize before the current blob
+ *                   blob being compressed are valid and refer to the preceding
+ *                   content (of length ctx->dictSize), which is available
+ *                   contiguously preceding in memory the content currently
+ *                   being compressed.
+ * - usingExtDict  : Like withPrefix64k, but the preceding content is somewhere
+ *                   else in memory, starting at ctx->dictionary with length
+ *                   ctx->dictSize.
+ * - usingDictCtx  : Like usingExtDict, but everything concerning the preceding
+ *                   content is in a separate context, pointed to by
+ *                   ctx->dictCtx. ctx->dictionary, ctx->dictSize, and table
+ *                   entries in the current context that refer to positions
+ *                   preceding the beginning of the current compression are
+ *                   ignored. Instead, ctx->dictCtx->dictionary and ctx->dictCtx
+ *                   ->dictSize describe the location and size of the preceding
+ *                   content, and matches are found by looking in the ctx
+ *                   ->dictCtx->hashTable.
+ */
+typedef enum { noDict = 0, withPrefix64k, usingExtDict, usingDictCtx } dict_directive;
+typedef enum { noDictIssue = 0, dictSmall } dictIssue_directive;
+
+
+/*-************************************
+*  Local Utils
+**************************************/
+int LZ4_versionNumber (void) { return LZ4_VERSION_NUMBER; }
+const char* LZ4_versionString(void) { return LZ4_VERSION_STRING; }
+int LZ4_compressBound(int isize)  { return LZ4_COMPRESSBOUND(isize); }
+int LZ4_sizeofState(void) { return LZ4_STREAMSIZE; }
+
+
+/*-************************************
+*  Internal Definitions used in Tests
+**************************************/
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+int LZ4_compress_forceExtDict (LZ4_stream_t* LZ4_dict, const char* source, char* dest, int srcSize);
+
+int LZ4_decompress_safe_forceExtDict(const char* source, char* dest,
+                                     int compressedSize, int maxOutputSize,
+                                     const void* dictStart, size_t dictSize);
+
+#if defined (__cplusplus)
+}
+#endif
+
+/*-******************************
+*  Compression functions
+********************************/
+LZ4_FORCE_INLINE U32 LZ4_hash4(U32 sequence, tableType_t const tableType)
+{
+    if (tableType == byU16)
+        return ((sequence * 2654435761U) >> ((MINMATCH*8)-(LZ4_HASHLOG+1)));
+    else
+        return ((sequence * 2654435761U) >> ((MINMATCH*8)-LZ4_HASHLOG));
+}
+
+LZ4_FORCE_INLINE U32 LZ4_hash5(U64 sequence, tableType_t const tableType)
+{
+    const U32 hashLog = (tableType == byU16) ? LZ4_HASHLOG+1 : LZ4_HASHLOG;
+    if (LZ4_isLittleEndian()) {
+        const U64 prime5bytes = 889523592379ULL;
+        return (U32)(((sequence << 24) * prime5bytes) >> (64 - hashLog));
+    } else {
+        const U64 prime8bytes = 11400714785074694791ULL;
+        return (U32)(((sequence >> 24) * prime8bytes) >> (64 - hashLog));
+    }
+}
+
+LZ4_FORCE_INLINE U32 LZ4_hashPosition(const void* const p, tableType_t const tableType)
+{
+    if ((sizeof(reg_t)==8) && (tableType != byU16)) return LZ4_hash5(LZ4_read_ARCH(p), tableType);
+    return LZ4_hash4(LZ4_read32(p), tableType);
+}
+
+LZ4_FORCE_INLINE void LZ4_clearHash(U32 h, void* tableBase, tableType_t const tableType)
+{
+    switch (tableType)
+    {
+    default: /* fallthrough */
+    case clearedTable: { /* illegal! */ assert(0); return; }
+    case byPtr: { const BYTE** hashTable = (const BYTE**)tableBase; hashTable[h] = NULL; return; }
+    case byU32: { U32* hashTable = (U32*) tableBase; hashTable[h] = 0; return; }
+    case byU16: { U16* hashTable = (U16*) tableBase; hashTable[h] = 0; return; }
+    }
+}
+
+LZ4_FORCE_INLINE void LZ4_putIndexOnHash(U32 idx, U32 h, void* tableBase, tableType_t const tableType)
+{
+    switch (tableType)
+    {
+    default: /* fallthrough */
+    case clearedTable: /* fallthrough */
+    case byPtr: { /* illegal! */ assert(0); return; }
+    case byU32: { U32* hashTable = (U32*) tableBase; hashTable[h] = idx; return; }
+    case byU16: { U16* hashTable = (U16*) tableBase; assert(idx < 65536); hashTable[h] = (U16)idx; return; }
+    }
+}
+
+LZ4_FORCE_INLINE void LZ4_putPositionOnHash(const BYTE* p, U32 h,
+                                  void* tableBase, tableType_t const tableType,
+                            const BYTE* srcBase)
+{
+    switch (tableType)
+    {
+    case clearedTable: { /* illegal! */ assert(0); return; }
+    case byPtr: { const BYTE** hashTable = (const BYTE**)tableBase; hashTable[h] = p; return; }
+    case byU32: { U32* hashTable = (U32*) tableBase; hashTable[h] = (U32)(p-srcBase); return; }
+    case byU16: { U16* hashTable = (U16*) tableBase; hashTable[h] = (U16)(p-srcBase); return; }
+    }
+}
+
+LZ4_FORCE_INLINE void LZ4_putPosition(const BYTE* p, void* tableBase, tableType_t tableType, const BYTE* srcBase)
+{
+    U32 const h = LZ4_hashPosition(p, tableType);
+    LZ4_putPositionOnHash(p, h, tableBase, tableType, srcBase);
+}
+
+/* LZ4_getIndexOnHash() :
+ * Index of match position registered in hash table.
+ * hash position must be calculated by using base+index, or dictBase+index.
+ * Assumption 1 : only valid if tableType == byU32 or byU16.
+ * Assumption 2 : h is presumed valid (within limits of hash table)
+ */
+LZ4_FORCE_INLINE U32 LZ4_getIndexOnHash(U32 h, const void* tableBase, tableType_t tableType)
+{
+    LZ4_STATIC_ASSERT(LZ4_MEMORY_USAGE > 2);
+    if (tableType == byU32) {
+        const U32* const hashTable = (const U32*) tableBase;
+        assert(h < (1U << (LZ4_MEMORY_USAGE-2)));
+        return hashTable[h];
+    }
+    if (tableType == byU16) {
+        const U16* const hashTable = (const U16*) tableBase;
+        assert(h < (1U << (LZ4_MEMORY_USAGE-1)));
+        return hashTable[h];
+    }
+    assert(0); return 0;  /* forbidden case */
+}
+
+static const BYTE* LZ4_getPositionOnHash(U32 h, const void* tableBase, tableType_t tableType, const BYTE* srcBase)
+{
+    if (tableType == byPtr) { const BYTE* const* hashTable = (const BYTE* const*) tableBase; return hashTable[h]; }
+    if (tableType == byU32) { const U32* const hashTable = (const U32*) tableBase; return hashTable[h] + srcBase; }
+    { const U16* const hashTable = (const U16*) tableBase; return hashTable[h] + srcBase; }   /* default, to ensure a return */
+}
+
+LZ4_FORCE_INLINE const BYTE*
+LZ4_getPosition(const BYTE* p,
+                const void* tableBase, tableType_t tableType,
+                const BYTE* srcBase)
+{
+    U32 const h = LZ4_hashPosition(p, tableType);
+    return LZ4_getPositionOnHash(h, tableBase, tableType, srcBase);
+}
+
+LZ4_FORCE_INLINE void
+LZ4_prepareTable(LZ4_stream_t_internal* const cctx,
+           const int inputSize,
+           const tableType_t tableType) {
+    /* If the table hasn't been used, it's guaranteed to be zeroed out, and is
+     * therefore safe to use no matter what mode we're in. Otherwise, we figure
+     * out if it's safe to leave as is or whether it needs to be reset.
+     */
+    if ((tableType_t)cctx->tableType != clearedTable) {
+        assert(inputSize >= 0);
+        if ((tableType_t)cctx->tableType != tableType
+          || ((tableType == byU16) && cctx->currentOffset + (unsigned)inputSize >= 0xFFFFU)
+          || ((tableType == byU32) && cctx->currentOffset > 1 GB)
+          || tableType == byPtr
+          || inputSize >= 4 KB)
+        {
+            DEBUGLOG(4, "LZ4_prepareTable: Resetting table in %p", cctx);
+            MEM_INIT(cctx->hashTable, 0, LZ4_HASHTABLESIZE);
+            cctx->currentOffset = 0;
+            cctx->tableType = (U32)clearedTable;
+        } else {
+            DEBUGLOG(4, "LZ4_prepareTable: Re-use hash table (no reset)");
+        }
+    }
+
+    /* Adding a gap, so all previous entries are > LZ4_DISTANCE_MAX back, is faster
+     * than compressing without a gap. However, compressing with
+     * currentOffset == 0 is faster still, so we preserve that case.
+     */
+    if (cctx->currentOffset != 0 && tableType == byU32) {
+        DEBUGLOG(5, "LZ4_prepareTable: adding 64KB to currentOffset");
+        cctx->currentOffset += 64 KB;
+    }
+
+    /* Finally, clear history */
+    cctx->dictCtx = NULL;
+    cctx->dictionary = NULL;
+    cctx->dictSize = 0;
+}
+
+/** LZ4_compress_generic() :
+ *  inlined, to ensure branches are decided at compilation time.
+ *  Presumed already validated at this stage:
+ *  - source != NULL
+ *  - inputSize > 0
+ */
+LZ4_FORCE_INLINE int LZ4_compress_generic_validated(
+                 LZ4_stream_t_internal* const cctx,
+                 const char* const source,
+                 char* const dest,
+                 const int inputSize,
+                 int *inputConsumed, /* only written when outputDirective == fillOutput */
+                 const int maxOutputSize,
+                 const limitedOutput_directive outputDirective,
+                 const tableType_t tableType,
+                 const dict_directive dictDirective,
+                 const dictIssue_directive dictIssue,
+                 const int acceleration)
+{
+    int result;
+    const BYTE* ip = (const BYTE*) source;
+
+    U32 const startIndex = cctx->currentOffset;
+    const BYTE* base = (const BYTE*) source - startIndex;
+    const BYTE* lowLimit;
+
+    const LZ4_stream_t_internal* dictCtx = (const LZ4_stream_t_internal*) cctx->dictCtx;
+    const BYTE* const dictionary =
+        dictDirective == usingDictCtx ? dictCtx->dictionary : cctx->dictionary;
+    const U32 dictSize =
+        dictDirective == usingDictCtx ? dictCtx->dictSize : cctx->dictSize;
+    const U32 dictDelta = (dictDirective == usingDictCtx) ? startIndex - dictCtx->currentOffset : 0;   /* make indexes in dictCtx comparable with index in current context */
+
+    int const maybe_extMem = (dictDirective == usingExtDict) || (dictDirective == usingDictCtx);
+    U32 const prefixIdxLimit = startIndex - dictSize;   /* used when dictDirective == dictSmall */
+    const BYTE* const dictEnd = dictionary ? dictionary + dictSize : dictionary;
+    const BYTE* anchor = (const BYTE*) source;
+    const BYTE* const iend = ip + inputSize;
+    const BYTE* const mflimitPlusOne = iend - MFLIMIT + 1;
+    const BYTE* const matchlimit = iend - LASTLITERALS;
+
+    /* the dictCtx currentOffset is indexed on the start of the dictionary,
+     * while a dictionary in the current context precedes the currentOffset */
+    const BYTE* dictBase = !dictionary ? NULL : (dictDirective == usingDictCtx) ?
+                            dictionary + dictSize - dictCtx->currentOffset :
+                            dictionary + dictSize - startIndex;
+
+    BYTE* op = (BYTE*) dest;
+    BYTE* const olimit = op + maxOutputSize;
+
+    U32 offset = 0;
+    U32 forwardH;
+
+    DEBUGLOG(5, "LZ4_compress_generic_validated: srcSize=%i, tableType=%u", inputSize, tableType);
+    assert(ip != NULL);
+    /* If init conditions are not met, we don't have to mark stream
+     * as having dirty context, since no action was taken yet */
+    if (outputDirective == fillOutput && maxOutputSize < 1) { return 0; } /* Impossible to store anything */
+    if ((tableType == byU16) && (inputSize>=LZ4_64Klimit)) { return 0; }  /* Size too large (not within 64K limit) */
+    if (tableType==byPtr) assert(dictDirective==noDict);      /* only supported use case with byPtr */
+    assert(acceleration >= 1);
+
+    lowLimit = (const BYTE*)source - (dictDirective == withPrefix64k ? dictSize : 0);
+
+    /* Update context state */
+    if (dictDirective == usingDictCtx) {
+        /* Subsequent linked blocks can't use the dictionary. */
+        /* Instead, they use the block we just compressed. */
+        cctx->dictCtx = NULL;
+        cctx->dictSize = (U32)inputSize;
+    } else {
+        cctx->dictSize += (U32)inputSize;
+    }
+    cctx->currentOffset += (U32)inputSize;
+    cctx->tableType = (U32)tableType;
+
+    if (inputSize<LZ4_minLength) goto _last_literals;        /* Input too small, no compression (all literals) */
+
+    /* First Byte */
+    LZ4_putPosition(ip, cctx->hashTable, tableType, base);
+    ip++; forwardH = LZ4_hashPosition(ip, tableType);
+
+    /* Main Loop */
+    for ( ; ; ) {
+        const BYTE* match;
+        BYTE* token;
+        const BYTE* filledIp;
+
+        /* Find a match */
+        if (tableType == byPtr) {
+            const BYTE* forwardIp = ip;
+            int step = 1;
+            int searchMatchNb = acceleration << LZ4_skipTrigger;
+            do {
+                U32 const h = forwardH;
+                ip = forwardIp;
+                forwardIp += step;
+                step = (searchMatchNb++ >> LZ4_skipTrigger);
+
+                if (unlikely(forwardIp > mflimitPlusOne)) goto _last_literals;
+                assert(ip < mflimitPlusOne);
+
+                match = LZ4_getPositionOnHash(h, cctx->hashTable, tableType, base);
+                forwardH = LZ4_hashPosition(forwardIp, tableType);
+                LZ4_putPositionOnHash(ip, h, cctx->hashTable, tableType, base);
+
+            } while ( (match+LZ4_DISTANCE_MAX < ip)
+                   || (LZ4_read32(match) != LZ4_read32(ip)) );
+
+        } else {   /* byU32, byU16 */
+
+            const BYTE* forwardIp = ip;
+            int step = 1;
+            int searchMatchNb = acceleration << LZ4_skipTrigger;
+            do {
+                U32 const h = forwardH;
+                U32 const current = (U32)(forwardIp - base);
+                U32 matchIndex = LZ4_getIndexOnHash(h, cctx->hashTable, tableType);
+                assert(matchIndex <= current);
+                assert(forwardIp - base < (ptrdiff_t)(2 GB - 1));
+                ip = forwardIp;
+                forwardIp += step;
+                step = (searchMatchNb++ >> LZ4_skipTrigger);
+
+                if (unlikely(forwardIp > mflimitPlusOne)) goto _last_literals;
+                assert(ip < mflimitPlusOne);
+
+                if (dictDirective == usingDictCtx) {
+                    if (matchIndex < startIndex) {
+                        /* there was no match, try the dictionary */
+                        assert(tableType == byU32);
+                        matchIndex = LZ4_getIndexOnHash(h, dictCtx->hashTable, byU32);
+                        match = dictBase + matchIndex;
+                        matchIndex += dictDelta;   /* make dictCtx index comparable with current context */
+                        lowLimit = dictionary;
+                    } else {
+                        match = base + matchIndex;
+                        lowLimit = (const BYTE*)source;
+                    }
+                } else if (dictDirective==usingExtDict) {
+                    if (matchIndex < startIndex) {
+                        DEBUGLOG(7, "extDict candidate: matchIndex=%5u  <  startIndex=%5u", matchIndex, startIndex);
+                        assert(startIndex - matchIndex >= MINMATCH);
+                        match = dictBase + matchIndex;
+                        lowLimit = dictionary;
+                    } else {
+                        match = base + matchIndex;
+                        lowLimit = (const BYTE*)source;
+                    }
+                } else {   /* single continuous memory segment */
+                    match = base + matchIndex;
+                }
+                forwardH = LZ4_hashPosition(forwardIp, tableType);
+                LZ4_putIndexOnHash(current, h, cctx->hashTable, tableType);
+
+                DEBUGLOG(7, "candidate at pos=%u  (offset=%u \n", matchIndex, current - matchIndex);
+                if ((dictIssue == dictSmall) && (matchIndex < prefixIdxLimit)) { continue; }    /* match outside of valid area */
+                assert(matchIndex < current);
+                if ( ((tableType != byU16) || (LZ4_DISTANCE_MAX < LZ4_DISTANCE_ABSOLUTE_MAX))
+                  && (matchIndex+LZ4_DISTANCE_MAX < current)) {
+                    continue;
+                } /* too far */
+                assert((current - matchIndex) <= LZ4_DISTANCE_MAX);  /* match now expected within distance */
+
+                if (LZ4_read32(match) == LZ4_read32(ip)) {
+                    if (maybe_extMem) offset = current - matchIndex;
+                    break;   /* match found */
+                }
+
+            } while(1);
+        }
+
+        /* Catch up */
+        filledIp = ip;
+        while (((ip>anchor) & (match > lowLimit)) && (unlikely(ip[-1]==match[-1]))) { ip--; match--; }
+
+        /* Encode Literals */
+        {   unsigned const litLength = (unsigned)(ip - anchor);
+            token = op++;
+            if ((outputDirective == limitedOutput) &&  /* Check output buffer overflow */
+                (unlikely(op + litLength + (2 + 1 + LASTLITERALS) + (litLength/255) > olimit)) ) {
+                return 0;   /* cannot compress within `dst` budget. Stored indexes in hash table are nonetheless fine */
+            }
+            if ((outputDirective == fillOutput) &&
+                (unlikely(op + (litLength+240)/255 /* litlen */ + litLength /* literals */ + 2 /* offset */ + 1 /* token */ + MFLIMIT - MINMATCH /* min last literals so last match is <= end - MFLIMIT */ > olimit))) {
+                op--;
+                goto _last_literals;
+            }
+            if (litLength >= RUN_MASK) {
+                int len = (int)(litLength - RUN_MASK);
+                *token = (RUN_MASK<<ML_BITS);
+                for(; len >= 255 ; len-=255) *op++ = 255;
+                *op++ = (BYTE)len;
+            }
+            else *token = (BYTE)(litLength<<ML_BITS);
+
+            /* Copy Literals */
+            LZ4_wildCopy8(op, anchor, op+litLength);
+            op+=litLength;
+            DEBUGLOG(6, "seq.start:%i, literals=%u, match.start:%i",
+                        (int)(anchor-(const BYTE*)source), litLength, (int)(ip-(const BYTE*)source));
+        }
+
+_next_match:
+        /* at this stage, the following variables must be correctly set :
+         * - ip : at start of LZ operation
+         * - match : at start of previous pattern occurence; can be within current prefix, or within extDict
+         * - offset : if maybe_ext_memSegment==1 (constant)
+         * - lowLimit : must be == dictionary to mean "match is within extDict"; must be == source otherwise
+         * - token and *token : position to write 4-bits for match length; higher 4-bits for literal length supposed already written
+         */
+
+        if ((outputDirective == fillOutput) &&
+            (op + 2 /* offset */ + 1 /* token */ + MFLIMIT - MINMATCH /* min last literals so last match is <= end - MFLIMIT */ > olimit)) {
+            /* the match was too close to the end, rewind and go to last literals */
+            op = token;
+            goto _last_literals;
+        }
+
+        /* Encode Offset */
+        if (maybe_extMem) {   /* static test */
+            DEBUGLOG(6, "             with offset=%u  (ext if > %i)", offset, (int)(ip - (const BYTE*)source));
+            assert(offset <= LZ4_DISTANCE_MAX && offset > 0);
+            LZ4_writeLE16(op, (U16)offset); op+=2;
+        } else  {
+            DEBUGLOG(6, "             with offset=%u  (same segment)", (U32)(ip - match));
+            assert(ip-match <= LZ4_DISTANCE_MAX);
+            LZ4_writeLE16(op, (U16)(ip - match)); op+=2;
+        }
+
+        /* Encode MatchLength */
+        {   unsigned matchCode;
+
+            if ( (dictDirective==usingExtDict || dictDirective==usingDictCtx)
+              && (lowLimit==dictionary) /* match within extDict */ ) {
+                const BYTE* limit = ip + (dictEnd-match);
+                assert(dictEnd > match);
+                if (limit > matchlimit) limit = matchlimit;
+                matchCode = LZ4_count(ip+MINMATCH, match+MINMATCH, limit);
+                ip += (size_t)matchCode + MINMATCH;
+                if (ip==limit) {
+                    unsigned const more = LZ4_count(limit, (const BYTE*)source, matchlimit);
+                    matchCode += more;
+                    ip += more;
+                }
+                DEBUGLOG(6, "             with matchLength=%u starting in extDict", matchCode+MINMATCH);
+            } else {
+                matchCode = LZ4_count(ip+MINMATCH, match+MINMATCH, matchlimit);
+                ip += (size_t)matchCode + MINMATCH;
+                DEBUGLOG(6, "             with matchLength=%u", matchCode+MINMATCH);
+            }
+
+            if ((outputDirective) &&    /* Check output buffer overflow */
+                (unlikely(op + (1 + LASTLITERALS) + (matchCode+240)/255 > olimit)) ) {
+                if (outputDirective == fillOutput) {
+                    /* Match description too long : reduce it */
+                    U32 newMatchCode = 15 /* in token */ - 1 /* to avoid needing a zero byte */ + ((U32)(olimit - op) - 1 - LASTLITERALS) * 255;
+                    ip -= matchCode - newMatchCode;
+                    assert(newMatchCode < matchCode);
+                    matchCode = newMatchCode;
+                    if (unlikely(ip <= filledIp)) {
+                        /* We have already filled up to filledIp so if ip ends up less than filledIp
+                         * we have positions in the hash table beyond the current position. This is
+                         * a problem if we reuse the hash table. So we have to remove these positions
+                         * from the hash table.
+                         */
+                        const BYTE* ptr;
+                        DEBUGLOG(5, "Clearing %u positions", (U32)(filledIp - ip));
+                        for (ptr = ip; ptr <= filledIp; ++ptr) {
+                            U32 const h = LZ4_hashPosition(ptr, tableType);
+                            LZ4_clearHash(h, cctx->hashTable, tableType);
+                        }
+                    }
+                } else {
+                    assert(outputDirective == limitedOutput);
+                    return 0;   /* cannot compress within `dst` budget. Stored indexes in hash table are nonetheless fine */
+                }
+            }
+            if (matchCode >= ML_MASK) {
+                *token += ML_MASK;
+                matchCode -= ML_MASK;
+                LZ4_write32(op, 0xFFFFFFFF);
+                while (matchCode >= 4*255) {
+                    op+=4;
+                    LZ4_write32(op, 0xFFFFFFFF);
+                    matchCode -= 4*255;
+                }
+                op += matchCode / 255;
+                *op++ = (BYTE)(matchCode % 255);
+            } else
+                *token += (BYTE)(matchCode);
+        }
+        /* Ensure we have enough space for the last literals. */
+        assert(!(outputDirective == fillOutput && op + 1 + LASTLITERALS > olimit));
+
+        anchor = ip;
+
+        /* Test end of chunk */
+        if (ip >= mflimitPlusOne) break;
+
+        /* Fill table */
+        LZ4_putPosition(ip-2, cctx->hashTable, tableType, base);
+
+        /* Test next position */
+        if (tableType == byPtr) {
+
+            match = LZ4_getPosition(ip, cctx->hashTable, tableType, base);
+            LZ4_putPosition(ip, cctx->hashTable, tableType, base);
+            if ( (match+LZ4_DISTANCE_MAX >= ip)
+              && (LZ4_read32(match) == LZ4_read32(ip)) )
+            { token=op++; *token=0; goto _next_match; }
+
+        } else {   /* byU32, byU16 */
+
+            U32 const h = LZ4_hashPosition(ip, tableType);
+            U32 const current = (U32)(ip-base);
+            U32 matchIndex = LZ4_getIndexOnHash(h, cctx->hashTable, tableType);
+            assert(matchIndex < current);
+            if (dictDirective == usingDictCtx) {
+                if (matchIndex < startIndex) {
+                    /* there was no match, try the dictionary */
+                    matchIndex = LZ4_getIndexOnHash(h, dictCtx->hashTable, byU32);
+                    match = dictBase + matchIndex;
+                    lowLimit = dictionary;   /* required for match length counter */
+                    matchIndex += dictDelta;
+                } else {
+                    match = base + matchIndex;
+                    lowLimit = (const BYTE*)source;  /* required for match length counter */
+                }
+            } else if (dictDirective==usingExtDict) {
+                if (matchIndex < startIndex) {
+                    match = dictBase + matchIndex;
+                    lowLimit = dictionary;   /* required for match length counter */
+                } else {
+                    match = base + matchIndex;
+                    lowLimit = (const BYTE*)source;   /* required for match length counter */
+                }
+            } else {   /* single memory segment */
+                match = base + matchIndex;
+            }
+            LZ4_putIndexOnHash(current, h, cctx->hashTable, tableType);
+            assert(matchIndex < current);
+            if ( ((dictIssue==dictSmall) ? (matchIndex >= prefixIdxLimit) : 1)
+              && (((tableType==byU16) && (LZ4_DISTANCE_MAX == LZ4_DISTANCE_ABSOLUTE_MAX)) ? 1 : (matchIndex+LZ4_DISTANCE_MAX >= current))
+              && (LZ4_read32(match) == LZ4_read32(ip)) ) {
+                token=op++;
+                *token=0;
+                if (maybe_extMem) offset = current - matchIndex;
+                DEBUGLOG(6, "seq.start:%i, literals=%u, match.start:%i",
+                            (int)(anchor-(const BYTE*)source), 0, (int)(ip-(const BYTE*)source));
+                goto _next_match;
+            }
+        }
+
+        /* Prepare next loop */
+        forwardH = LZ4_hashPosition(++ip, tableType);
+
+    }
+
+_last_literals:
+    /* Encode Last Literals */
+    {   size_t lastRun = (size_t)(iend - anchor);
+        if ( (outputDirective) &&  /* Check output buffer overflow */
+            (op + lastRun + 1 + ((lastRun+255-RUN_MASK)/255) > olimit)) {
+            if (outputDirective == fillOutput) {
+                /* adapt lastRun to fill 'dst' */
+                assert(olimit >= op);
+                lastRun  = (size_t)(olimit-op) - 1/*token*/;
+                lastRun -= (lastRun + 256 - RUN_MASK) / 256;  /*additional length tokens*/
+            } else {
+                assert(outputDirective == limitedOutput);
+                return 0;   /* cannot compress within `dst` budget. Stored indexes in hash table are nonetheless fine */
+            }
+        }
+        DEBUGLOG(6, "Final literal run : %i literals", (int)lastRun);
+        if (lastRun >= RUN_MASK) {
+            size_t accumulator = lastRun - RUN_MASK;
+            *op++ = RUN_MASK << ML_BITS;
+            for(; accumulator >= 255 ; accumulator-=255) *op++ = 255;
+            *op++ = (BYTE) accumulator;
+        } else {
+            *op++ = (BYTE)(lastRun<<ML_BITS);
+        }
+        LZ4_memcpy(op, anchor, lastRun);
+        ip = anchor + lastRun;
+        op += lastRun;
+    }
+
+    if (outputDirective == fillOutput) {
+        *inputConsumed = (int) (((const char*)ip)-source);
+    }
+    result = (int)(((char*)op) - dest);
+    assert(result > 0);
+    DEBUGLOG(5, "LZ4_compress_generic: compressed %i bytes into %i bytes", inputSize, result);
+    return result;
+}
+
+/** LZ4_compress_generic() :
+ *  inlined, to ensure branches are decided at compilation time;
+ *  takes care of src == (NULL, 0)
+ *  and forward the rest to LZ4_compress_generic_validated */
+LZ4_FORCE_INLINE int LZ4_compress_generic(
+                 LZ4_stream_t_internal* const cctx,
+                 const char* const src,
+                 char* const dst,
+                 const int srcSize,
+                 int *inputConsumed, /* only written when outputDirective == fillOutput */
+                 const int dstCapacity,
+                 const limitedOutput_directive outputDirective,
+                 const tableType_t tableType,
+                 const dict_directive dictDirective,
+                 const dictIssue_directive dictIssue,
+                 const int acceleration)
+{
+    DEBUGLOG(5, "LZ4_compress_generic: srcSize=%i, dstCapacity=%i",
+                srcSize, dstCapacity);
+
+    if ((U32)srcSize > (U32)LZ4_MAX_INPUT_SIZE) { return 0; }  /* Unsupported srcSize, too large (or negative) */
+    if (srcSize == 0) {   /* src == NULL supported if srcSize == 0 */
+        if (outputDirective != notLimited && dstCapacity <= 0) return 0;  /* no output, can't write anything */
+        DEBUGLOG(5, "Generating an empty block");
+        assert(outputDirective == notLimited || dstCapacity >= 1);
+        assert(dst != NULL);
+        dst[0] = 0;
+        if (outputDirective == fillOutput) {
+            assert (inputConsumed != NULL);
+            *inputConsumed = 0;
+        }
+        return 1;
+    }
+    assert(src != NULL);
+
+    return LZ4_compress_generic_validated(cctx, src, dst, srcSize,
+                inputConsumed, /* only written into if outputDirective == fillOutput */
+                dstCapacity, outputDirective,
+                tableType, dictDirective, dictIssue, acceleration);
+}
+
+
+int LZ4_compress_fast_extState(void* state, const char* source, char* dest, int inputSize, int maxOutputSize, int acceleration)
+{
+    LZ4_stream_t_internal* const ctx = & LZ4_initStream(state, sizeof(LZ4_stream_t)) -> internal_donotuse;
+    assert(ctx != NULL);
+    if (acceleration < 1) acceleration = LZ4_ACCELERATION_DEFAULT;
+    if (acceleration > LZ4_ACCELERATION_MAX) acceleration = LZ4_ACCELERATION_MAX;
+    if (maxOutputSize >= LZ4_compressBound(inputSize)) {
+        if (inputSize < LZ4_64Klimit) {
+            return LZ4_compress_generic(ctx, source, dest, inputSize, NULL, 0, notLimited, byU16, noDict, noDictIssue, acceleration);
+        } else {
+            const tableType_t tableType = ((sizeof(void*)==4) && ((uptrval)source > LZ4_DISTANCE_MAX)) ? byPtr : byU32;
+            return LZ4_compress_generic(ctx, source, dest, inputSize, NULL, 0, notLimited, tableType, noDict, noDictIssue, acceleration);
+        }
+    } else {
+        if (inputSize < LZ4_64Klimit) {
+            return LZ4_compress_generic(ctx, source, dest, inputSize, NULL, maxOutputSize, limitedOutput, byU16, noDict, noDictIssue, acceleration);
+        } else {
+            const tableType_t tableType = ((sizeof(void*)==4) && ((uptrval)source > LZ4_DISTANCE_MAX)) ? byPtr : byU32;
+            return LZ4_compress_generic(ctx, source, dest, inputSize, NULL, maxOutputSize, limitedOutput, tableType, noDict, noDictIssue, acceleration);
+        }
+    }
+}
+
+/**
+ * LZ4_compress_fast_extState_fastReset() :
+ * A variant of LZ4_compress_fast_extState().
+ *
+ * Using this variant avoids an expensive initialization step. It is only safe
+ * to call if the state buffer is known to be correctly initialized already
+ * (see comment in lz4.h on LZ4_resetStream_fast() for a definition of
+ * "correctly initialized").
+ */
+int LZ4_compress_fast_extState_fastReset(void* state, const char* src, char* dst, int srcSize, int dstCapacity, int acceleration)
+{
+    LZ4_stream_t_internal* ctx = &((LZ4_stream_t*)state)->internal_donotuse;
+    if (acceleration < 1) acceleration = LZ4_ACCELERATION_DEFAULT;
+    if (acceleration > LZ4_ACCELERATION_MAX) acceleration = LZ4_ACCELERATION_MAX;
+
+    if (dstCapacity >= LZ4_compressBound(srcSize)) {
+        if (srcSize < LZ4_64Klimit) {
+            const tableType_t tableType = byU16;
+            LZ4_prepareTable(ctx, srcSize, tableType);
+            if (ctx->currentOffset) {
+                return LZ4_compress_generic(ctx, src, dst, srcSize, NULL, 0, notLimited, tableType, noDict, dictSmall, acceleration);
+            } else {
+                return LZ4_compress_generic(ctx, src, dst, srcSize, NULL, 0, notLimited, tableType, noDict, noDictIssue, acceleration);
+            }
+        } else {
+            const tableType_t tableType = ((sizeof(void*)==4) && ((uptrval)src > LZ4_DISTANCE_MAX)) ? byPtr : byU32;
+            LZ4_prepareTable(ctx, srcSize, tableType);
+            return LZ4_compress_generic(ctx, src, dst, srcSize, NULL, 0, notLimited, tableType, noDict, noDictIssue, acceleration);
+        }
+    } else {
+        if (srcSize < LZ4_64Klimit) {
+            const tableType_t tableType = byU16;
+            LZ4_prepareTable(ctx, srcSize, tableType);
+            if (ctx->currentOffset) {
+                return LZ4_compress_generic(ctx, src, dst, srcSize, NULL, dstCapacity, limitedOutput, tableType, noDict, dictSmall, acceleration);
+            } else {
+                return LZ4_compress_generic(ctx, src, dst, srcSize, NULL, dstCapacity, limitedOutput, tableType, noDict, noDictIssue, acceleration);
+            }
+        } else {
+            const tableType_t tableType = ((sizeof(void*)==4) && ((uptrval)src > LZ4_DISTANCE_MAX)) ? byPtr : byU32;
+            LZ4_prepareTable(ctx, srcSize, tableType);
+            return LZ4_compress_generic(ctx, src, dst, srcSize, NULL, dstCapacity, limitedOutput, tableType, noDict, noDictIssue, acceleration);
+        }
+    }
+}
+
+
+int LZ4_compress_fast(const char* source, char* dest, int inputSize, int maxOutputSize, int acceleration)
+{
+    int result;
+#if (LZ4_HEAPMODE)
+    LZ4_stream_t* ctxPtr = ALLOC(sizeof(LZ4_stream_t));   /* malloc-calloc always properly aligned */
+    if (ctxPtr == NULL) return 0;
+#else
+    LZ4_stream_t ctx;
+    LZ4_stream_t* const ctxPtr = &ctx;
+#endif
+    result = LZ4_compress_fast_extState(ctxPtr, source, dest, inputSize, maxOutputSize, acceleration);
+
+#if (LZ4_HEAPMODE)
+    FREEMEM(ctxPtr);
+#endif
+    return result;
+}
+
+
+int LZ4_compress_default(const char* src, char* dst, int srcSize, int maxOutputSize)
+{
+    return LZ4_compress_fast(src, dst, srcSize, maxOutputSize, 1);
+}
+
+
+/* Note!: This function leaves the stream in an unclean/broken state!
+ * It is not safe to subsequently use the same state with a _fastReset() or
+ * _continue() call without resetting it. */
+static int LZ4_compress_destSize_extState (LZ4_stream_t* state, const char* src, char* dst, int* srcSizePtr, int targetDstSize)
+{
+    void* const s = LZ4_initStream(state, sizeof (*state));
+    assert(s != NULL); (void)s;
+
+    if (targetDstSize >= LZ4_compressBound(*srcSizePtr)) {  /* compression success is guaranteed */
+        return LZ4_compress_fast_extState(state, src, dst, *srcSizePtr, targetDstSize, 1);
+    } else {
+        if (*srcSizePtr < LZ4_64Klimit) {
+            return LZ4_compress_generic(&state->internal_donotuse, src, dst, *srcSizePtr, srcSizePtr, targetDstSize, fillOutput, byU16, noDict, noDictIssue, 1);
+        } else {
+            tableType_t const addrMode = ((sizeof(void*)==4) && ((uptrval)src > LZ4_DISTANCE_MAX)) ? byPtr : byU32;
+            return LZ4_compress_generic(&state->internal_donotuse, src, dst, *srcSizePtr, srcSizePtr, targetDstSize, fillOutput, addrMode, noDict, noDictIssue, 1);
+    }   }
+}
+
+
+int LZ4_compress_destSize(const char* src, char* dst, int* srcSizePtr, int targetDstSize)
+{
+#if (LZ4_HEAPMODE)
+    LZ4_stream_t* ctx = (LZ4_stream_t*)ALLOC(sizeof(LZ4_stream_t));   /* malloc-calloc always properly aligned */
+    if (ctx == NULL) return 0;
+#else
+    LZ4_stream_t ctxBody;
+    LZ4_stream_t* ctx = &ctxBody;
+#endif
+
+    int result = LZ4_compress_destSize_extState(ctx, src, dst, srcSizePtr, targetDstSize);
+
+#if (LZ4_HEAPMODE)
+    FREEMEM(ctx);
+#endif
+    return result;
+}
+
+
+
+/*-******************************
+*  Streaming functions
+********************************/
+
+LZ4_stream_t* LZ4_createStream(void)
+{
+    LZ4_stream_t* const lz4s = (LZ4_stream_t*)ALLOC(sizeof(LZ4_stream_t));
+    LZ4_STATIC_ASSERT(LZ4_STREAMSIZE >= sizeof(LZ4_stream_t_internal));    /* A compilation error here means LZ4_STREAMSIZE is not large enough */
+    DEBUGLOG(4, "LZ4_createStream %p", lz4s);
+    if (lz4s == NULL) return NULL;
+    LZ4_initStream(lz4s, sizeof(*lz4s));
+    return lz4s;
+}
+
+static size_t LZ4_stream_t_alignment(void)
+{
+#if LZ4_ALIGN_TEST
+    typedef struct { char c; LZ4_stream_t t; } t_a;
+    return sizeof(t_a) - sizeof(LZ4_stream_t);
+#else
+    return 1;  /* effectively disabled */
+#endif
+}
+
+LZ4_stream_t* LZ4_initStream (void* buffer, size_t size)
+{
+    DEBUGLOG(5, "LZ4_initStream");
+    if (buffer == NULL) { return NULL; }
+    if (size < sizeof(LZ4_stream_t)) { return NULL; }
+    if (!LZ4_isAligned(buffer, LZ4_stream_t_alignment())) return NULL;
+    MEM_INIT(buffer, 0, sizeof(LZ4_stream_t_internal));
+    return (LZ4_stream_t*)buffer;
+}
+
+/* resetStream is now deprecated,
+ * prefer initStream() which is more general */
+void LZ4_resetStream (LZ4_stream_t* LZ4_stream)
+{
+    DEBUGLOG(5, "LZ4_resetStream (ctx:%p)", LZ4_stream);
+    MEM_INIT(LZ4_stream, 0, sizeof(LZ4_stream_t_internal));
+}
+
+void LZ4_resetStream_fast(LZ4_stream_t* ctx) {
+    LZ4_prepareTable(&(ctx->internal_donotuse), 0, byU32);
+}
+
+int LZ4_freeStream (LZ4_stream_t* LZ4_stream)
+{
+    if (!LZ4_stream) return 0;   /* support free on NULL */
+    DEBUGLOG(5, "LZ4_freeStream %p", LZ4_stream);
+    FREEMEM(LZ4_stream);
+    return (0);
+}
+
+
+#define HASH_UNIT sizeof(reg_t)
+int LZ4_loadDict (LZ4_stream_t* LZ4_dict, const char* dictionary, int dictSize)
+{
+    LZ4_stream_t_internal* dict = &LZ4_dict->internal_donotuse;
+    const tableType_t tableType = byU32;
+    const BYTE* p = (const BYTE*)dictionary;
+    const BYTE* const dictEnd = p + dictSize;
+    const BYTE* base;
+
+    DEBUGLOG(4, "LZ4_loadDict (%i bytes from %p into %p)", dictSize, dictionary, LZ4_dict);
+
+    /* It's necessary to reset the context,
+     * and not just continue it with prepareTable()
+     * to avoid any risk of generating overflowing matchIndex
+     * when compressing using this dictionary */
+    LZ4_resetStream(LZ4_dict);
+
+    /* We always increment the offset by 64 KB, since, if the dict is longer,
+     * we truncate it to the last 64k, and if it's shorter, we still want to
+     * advance by a whole window length so we can provide the guarantee that
+     * there are only valid offsets in the window, which allows an optimization
+     * in LZ4_compress_fast_continue() where it uses noDictIssue even when the
+     * dictionary isn't a full 64k. */
+    dict->currentOffset += 64 KB;
+
+    if (dictSize < (int)HASH_UNIT) {
+        return 0;
+    }
+
+    if ((dictEnd - p) > 64 KB) p = dictEnd - 64 KB;
+    base = dictEnd - dict->currentOffset;
+    dict->dictionary = p;
+    dict->dictSize = (U32)(dictEnd - p);
+    dict->tableType = (U32)tableType;
+
+    while (p <= dictEnd-HASH_UNIT) {
+        LZ4_putPosition(p, dict->hashTable, tableType, base);
+        p+=3;
+    }
+
+    return (int)dict->dictSize;
+}
+
+void LZ4_attach_dictionary(LZ4_stream_t* workingStream, const LZ4_stream_t* dictionaryStream) {
+    const LZ4_stream_t_internal* dictCtx = dictionaryStream == NULL ? NULL :
+        &(dictionaryStream->internal_donotuse);
+
+    DEBUGLOG(4, "LZ4_attach_dictionary (%p, %p, size %u)",
+             workingStream, dictionaryStream,
+             dictCtx != NULL ? dictCtx->dictSize : 0);
+
+    if (dictCtx != NULL) {
+        /* If the current offset is zero, we will never look in the
+         * external dictionary context, since there is no value a table
+         * entry can take that indicate a miss. In that case, we need
+         * to bump the offset to something non-zero.
+         */
+        if (workingStream->internal_donotuse.currentOffset == 0) {
+            workingStream->internal_donotuse.currentOffset = 64 KB;
+        }
+
+        /* Don't actually attach an empty dictionary.
+         */
+        if (dictCtx->dictSize == 0) {
+            dictCtx = NULL;
+        }
+    }
+    workingStream->internal_donotuse.dictCtx = dictCtx;
+}
+
+
+static void LZ4_renormDictT(LZ4_stream_t_internal* LZ4_dict, int nextSize)
+{
+    assert(nextSize >= 0);
+    if (LZ4_dict->currentOffset + (unsigned)nextSize > 0x80000000) {   /* potential ptrdiff_t overflow (32-bits mode) */
+        /* rescale hash table */
+        U32 const delta = LZ4_dict->currentOffset - 64 KB;
+        const BYTE* dictEnd = LZ4_dict->dictionary + LZ4_dict->dictSize;
+        int i;
+        DEBUGLOG(4, "LZ4_renormDictT");
+        for (i=0; i<LZ4_HASH_SIZE_U32; i++) {
+            if (LZ4_dict->hashTable[i] < delta) LZ4_dict->hashTable[i]=0;
+            else LZ4_dict->hashTable[i] -= delta;
+        }
+        LZ4_dict->currentOffset = 64 KB;
+        if (LZ4_dict->dictSize > 64 KB) LZ4_dict->dictSize = 64 KB;
+        LZ4_dict->dictionary = dictEnd - LZ4_dict->dictSize;
+    }
+}
+
+
+int LZ4_compress_fast_continue (LZ4_stream_t* LZ4_stream,
+                                const char* source, char* dest,
+                                int inputSize, int maxOutputSize,
+                                int acceleration)
+{
+    const tableType_t tableType = byU32;
+    LZ4_stream_t_internal* streamPtr = &LZ4_stream->internal_donotuse;
+    const BYTE* dictEnd = streamPtr->dictionary + streamPtr->dictSize;
+
+    DEBUGLOG(5, "LZ4_compress_fast_continue (inputSize=%i)", inputSize);
+
+    LZ4_renormDictT(streamPtr, inputSize);   /* avoid index overflow */
+    if (acceleration < 1) acceleration = LZ4_ACCELERATION_DEFAULT;
+    if (acceleration > LZ4_ACCELERATION_MAX) acceleration = LZ4_ACCELERATION_MAX;
+
+    /* invalidate tiny dictionaries */
+    if ( (streamPtr->dictSize-1 < 4-1)   /* intentional underflow */
+      && (dictEnd != (const BYTE*)source) ) {
+        DEBUGLOG(5, "LZ4_compress_fast_continue: dictSize(%u) at addr:%p is too small", streamPtr->dictSize, streamPtr->dictionary);
+        streamPtr->dictSize = 0;
+        streamPtr->dictionary = (const BYTE*)source;
+        dictEnd = (const BYTE*)source;
+    }
+
+    /* Check overlapping input/dictionary space */
+    {   const BYTE* sourceEnd = (const BYTE*) source + inputSize;
+        if ((sourceEnd > streamPtr->dictionary) && (sourceEnd < dictEnd)) {
+            streamPtr->dictSize = (U32)(dictEnd - sourceEnd);
+            if (streamPtr->dictSize > 64 KB) streamPtr->dictSize = 64 KB;
+            if (streamPtr->dictSize < 4) streamPtr->dictSize = 0;
+            streamPtr->dictionary = dictEnd - streamPtr->dictSize;
+        }
+    }
+
+    /* prefix mode : source data follows dictionary */
+    if (dictEnd == (const BYTE*)source) {
+        if ((streamPtr->dictSize < 64 KB) && (streamPtr->dictSize < streamPtr->currentOffset))
+            return LZ4_compress_generic(streamPtr, source, dest, inputSize, NULL, maxOutputSize, limitedOutput, tableType, withPrefix64k, dictSmall, acceleration);
+        else
+            return LZ4_compress_generic(streamPtr, source, dest, inputSize, NULL, maxOutputSize, limitedOutput, tableType, withPrefix64k, noDictIssue, acceleration);
+    }
+
+    /* external dictionary mode */
+    {   int result;
+        if (streamPtr->dictCtx) {
+            /* We depend here on the fact that dictCtx'es (produced by
+             * LZ4_loadDict) guarantee that their tables contain no references
+             * to offsets between dictCtx->currentOffset - 64 KB and
+             * dictCtx->currentOffset - dictCtx->dictSize. This makes it safe
+             * to use noDictIssue even when the dict isn't a full 64 KB.
+             */
+            if (inputSize > 4 KB) {
+                /* For compressing large blobs, it is faster to pay the setup
+                 * cost to copy the dictionary's tables into the active context,
+                 * so that the compression loop is only looking into one table.
+                 */
+                LZ4_memcpy(streamPtr, streamPtr->dictCtx, sizeof(*streamPtr));
+                result = LZ4_compress_generic(streamPtr, source, dest, inputSize, NULL, maxOutputSize, limitedOutput, tableType, usingExtDict, noDictIssue, acceleration);
+            } else {
+                result = LZ4_compress_generic(streamPtr, source, dest, inputSize, NULL, maxOutputSize, limitedOutput, tableType, usingDictCtx, noDictIssue, acceleration);
+            }
+        } else {
+            if ((streamPtr->dictSize < 64 KB) && (streamPtr->dictSize < streamPtr->currentOffset)) {
+                result = LZ4_compress_generic(streamPtr, source, dest, inputSize, NULL, maxOutputSize, limitedOutput, tableType, usingExtDict, dictSmall, acceleration);
+            } else {
+                result = LZ4_compress_generic(streamPtr, source, dest, inputSize, NULL, maxOutputSize, limitedOutput, tableType, usingExtDict, noDictIssue, acceleration);
+            }
+        }
+        streamPtr->dictionary = (const BYTE*)source;
+        streamPtr->dictSize = (U32)inputSize;
+        return result;
+    }
+}
+
+
+/* Hidden debug function, to force-test external dictionary mode */
+int LZ4_compress_forceExtDict (LZ4_stream_t* LZ4_dict, const char* source, char* dest, int srcSize)
+{
+    LZ4_stream_t_internal* streamPtr = &LZ4_dict->internal_donotuse;
+    int result;
+
+    LZ4_renormDictT(streamPtr, srcSize);
+
+    if ((streamPtr->dictSize < 64 KB) && (streamPtr->dictSize < streamPtr->currentOffset)) {
+        result = LZ4_compress_generic(streamPtr, source, dest, srcSize, NULL, 0, notLimited, byU32, usingExtDict, dictSmall, 1);
+    } else {
+        result = LZ4_compress_generic(streamPtr, source, dest, srcSize, NULL, 0, notLimited, byU32, usingExtDict, noDictIssue, 1);
+    }
+
+    streamPtr->dictionary = (const BYTE*)source;
+    streamPtr->dictSize = (U32)srcSize;
+
+    return result;
+}
+
+
+/*! LZ4_saveDict() :
+ *  If previously compressed data block is not guaranteed to remain available at its memory location,
+ *  save it into a safer place (char* safeBuffer).
+ *  Note : you don't need to call LZ4_loadDict() afterwards,
+ *         dictionary is immediately usable, you can therefore call LZ4_compress_fast_continue().
+ *  Return : saved dictionary size in bytes (necessarily <= dictSize), or 0 if error.
+ */
+int LZ4_saveDict (LZ4_stream_t* LZ4_dict, char* safeBuffer, int dictSize)
+{
+    LZ4_stream_t_internal* const dict = &LZ4_dict->internal_donotuse;
+    const BYTE* const previousDictEnd = dict->dictionary + dict->dictSize;
+
+    if ((U32)dictSize > 64 KB) { dictSize = 64 KB; } /* useless to define a dictionary > 64 KB */
+    if ((U32)dictSize > dict->dictSize) { dictSize = (int)dict->dictSize; }
+
+    if (safeBuffer == NULL) assert(dictSize == 0);
+    if (dictSize > 0)
+        memmove(safeBuffer, previousDictEnd - dictSize, dictSize);
+
+    dict->dictionary = (const BYTE*)safeBuffer;
+    dict->dictSize = (U32)dictSize;
+
+    return dictSize;
+}
+
+
+
+/*-*******************************
+ *  Decompression functions
+ ********************************/
+
+typedef enum { endOnOutputSize = 0, endOnInputSize = 1 } endCondition_directive;
+typedef enum { decode_full_block = 0, partial_decode = 1 } earlyEnd_directive;
+
+#undef MIN
+#define MIN(a,b)    ( (a) < (b) ? (a) : (b) )
+
+/* Read the variable-length literal or match length.
+ *
+ * ip - pointer to use as input.
+ * lencheck - end ip.  Return an error if ip advances >= lencheck.
+ * loop_check - check ip >= lencheck in body of loop.  Returns loop_error if so.
+ * initial_check - check ip >= lencheck before start of loop.  Returns initial_error if so.
+ * error (output) - error code.  Should be set to 0 before call.
+ */
+typedef enum { loop_error = -2, initial_error = -1, ok = 0 } variable_length_error;
+LZ4_FORCE_INLINE unsigned
+read_variable_length(const BYTE**ip, const BYTE* lencheck,
+                     int loop_check, int initial_check,
+                     variable_length_error* error)
+{
+    U32 length = 0;
+    U32 s;
+    if (initial_check && unlikely((*ip) >= lencheck)) {    /* overflow detection */
+        *error = initial_error;
+        return length;
+    }
+    do {
+        s = **ip;
+        (*ip)++;
+        length += s;
+        if (loop_check && unlikely((*ip) >= lencheck)) {    /* overflow detection */
+            *error = loop_error;
+            return length;
+        }
+    } while (s==255);
+
+    return length;
+}
+
+/*! LZ4_decompress_generic() :
+ *  This generic decompression function covers all use cases.
+ *  It shall be instantiated several times, using different sets of directives.
+ *  Note that it is important for performance that this function really get inlined,
+ *  in order to remove useless branches during compilation optimization.
+ */
+LZ4_FORCE_INLINE int
+LZ4_decompress_generic(
+                 const char* const src,
+                 char* const dst,
+                 int srcSize,
+                 int outputSize,         /* If endOnInput==endOnInputSize, this value is `dstCapacity` */
+
+                 endCondition_directive endOnInput,   /* endOnOutputSize, endOnInputSize */
+                 earlyEnd_directive partialDecoding,  /* full, partial */
+                 dict_directive dict,                 /* noDict, withPrefix64k, usingExtDict */
+                 const BYTE* const lowPrefix,  /* always <= dst, == dst when no prefix */
+                 const BYTE* const dictStart,  /* only if dict==usingExtDict */
+                 const size_t dictSize         /* note : = 0 if noDict */
+                 )
+{
+    if (src == NULL) { return -1; }
+
+    {   const BYTE* ip = (const BYTE*) src;
+        const BYTE* const iend = ip + srcSize;
+
+        BYTE* op = (BYTE*) dst;
+        BYTE* const oend = op + outputSize;
+        BYTE* cpy;
+
+        const BYTE* const dictEnd = (dictStart == NULL) ? NULL : dictStart + dictSize;
+
+        const int safeDecode = (endOnInput==endOnInputSize);
+        const int checkOffset = ((safeDecode) && (dictSize < (int)(64 KB)));
+
+
+        /* Set up the "end" pointers for the shortcut. */
+        const BYTE* const shortiend = iend - (endOnInput ? 14 : 8) /*maxLL*/ - 2 /*offset*/;
+        const BYTE* const shortoend = oend - (endOnInput ? 14 : 8) /*maxLL*/ - 18 /*maxML*/;
+
+        const BYTE* match;
+        size_t offset;
+        unsigned token;
+        size_t length;
+
+
+        DEBUGLOG(5, "LZ4_decompress_generic (srcSize:%i, dstSize:%i)", srcSize, outputSize);
+
+        /* Special cases */
+        assert(lowPrefix <= op);
+        if ((endOnInput) && (unlikely(outputSize==0))) {
+            /* Empty output buffer */
+            if (partialDecoding) return 0;
+            return ((srcSize==1) && (*ip==0)) ? 0 : -1;
+        }
+        if ((!endOnInput) && (unlikely(outputSize==0))) { return (*ip==0 ? 1 : -1); }
+        if ((endOnInput) && unlikely(srcSize==0)) { return -1; }
+
+	/* Currently the fast loop shows a regression on qualcomm arm chips. */
+#if LZ4_FAST_DEC_LOOP
+        if ((oend - op) < FASTLOOP_SAFE_DISTANCE) {
+            DEBUGLOG(6, "skip fast decode loop");
+            goto safe_decode;
+        }
+
+        /* Fast loop : decode sequences as long as output < iend-FASTLOOP_SAFE_DISTANCE */
+        while (1) {
+            /* Main fastloop assertion: We can always wildcopy FASTLOOP_SAFE_DISTANCE */
+            assert(oend - op >= FASTLOOP_SAFE_DISTANCE);
+            if (endOnInput) { assert(ip < iend); }
+            token = *ip++;
+            length = token >> ML_BITS;  /* literal length */
+
+            assert(!endOnInput || ip <= iend); /* ip < iend before the increment */
+
+            /* decode literal length */
+            if (length == RUN_MASK) {
+                variable_length_error error = ok;
+                length += read_variable_length(&ip, iend-RUN_MASK, (int)endOnInput, (int)endOnInput, &error);
+                if (error == initial_error) { goto _output_error; }
+                if ((safeDecode) && unlikely((uptrval)(op)+length<(uptrval)(op))) { goto _output_error; } /* overflow detection */
+                if ((safeDecode) && unlikely((uptrval)(ip)+length<(uptrval)(ip))) { goto _output_error; } /* overflow detection */
+
+                /* copy literals */
+                cpy = op+length;
+                LZ4_STATIC_ASSERT(MFLIMIT >= WILDCOPYLENGTH);
+                if (endOnInput) {  /* LZ4_decompress_safe() */
+                    if ((cpy>oend-32) || (ip+length>iend-32)) { goto safe_literal_copy; }
+                    LZ4_wildCopy32(op, ip, cpy);
+                } else {   /* LZ4_decompress_fast() */
+                    if (cpy>oend-8) { goto safe_literal_copy; }
+                    LZ4_wildCopy8(op, ip, cpy); /* LZ4_decompress_fast() cannot copy more than 8 bytes at a time :
+                                                 * it doesn't know input length, and only relies on end-of-block properties */
+                }
+                ip += length; op = cpy;
+            } else {
+                cpy = op+length;
+                if (endOnInput) {  /* LZ4_decompress_safe() */
+                    DEBUGLOG(7, "copy %u bytes in a 16-bytes stripe", (unsigned)length);
+                    /* We don't need to check oend, since we check it once for each loop below */
+                    if (ip > iend-(16 + 1/*max lit + offset + nextToken*/)) { goto safe_literal_copy; }
+                    /* Literals can only be 14, but hope compilers optimize if we copy by a register size */
+                    LZ4_memcpy(op, ip, 16);
+                } else {  /* LZ4_decompress_fast() */
+                    /* LZ4_decompress_fast() cannot copy more than 8 bytes at a time :
+                     * it doesn't know input length, and relies on end-of-block properties */
+                    LZ4_memcpy(op, ip, 8);
+                    if (length > 8) { LZ4_memcpy(op+8, ip+8, 8); }
+                }
+                ip += length; op = cpy;
+            }
+
+            /* get offset */
+            offset = LZ4_readLE16(ip); ip+=2;
+            match = op - offset;
+            assert(match <= op);
+
+            /* get matchlength */
+            length = token & ML_MASK;
+
+            if (length == ML_MASK) {
+                variable_length_error error = ok;
+                if ((checkOffset) && (unlikely(match + dictSize < lowPrefix))) { goto _output_error; } /* Error : offset outside buffers */
+                length += read_variable_length(&ip, iend - LASTLITERALS + 1, (int)endOnInput, 0, &error);
+                if (error != ok) { goto _output_error; }
+                if ((safeDecode) && unlikely((uptrval)(op)+length<(uptrval)op)) { goto _output_error; } /* overflow detection */
+                length += MINMATCH;
+                if (op + length >= oend - FASTLOOP_SAFE_DISTANCE) {
+                    goto safe_match_copy;
+                }
+            } else {
+                length += MINMATCH;
+                if (op + length >= oend - FASTLOOP_SAFE_DISTANCE) {
+                    goto safe_match_copy;
+                }
+
+                /* Fastpath check: Avoids a branch in LZ4_wildCopy32 if true */
+                if ((dict == withPrefix64k) || (match >= lowPrefix)) {
+                    if (offset >= 8) {
+                        assert(match >= lowPrefix);
+                        assert(match <= op);
+                        assert(op + 18 <= oend);
+
+                        LZ4_memcpy(op, match, 8);
+                        LZ4_memcpy(op+8, match+8, 8);
+                        LZ4_memcpy(op+16, match+16, 2);
+                        op += length;
+                        continue;
+            }   }   }
+
+            if (checkOffset && (unlikely(match + dictSize < lowPrefix))) { goto _output_error; } /* Error : offset outside buffers */
+            /* match starting within external dictionary */
+            if ((dict==usingExtDict) && (match < lowPrefix)) {
+                if (unlikely(op+length > oend-LASTLITERALS)) {
+                    if (partialDecoding) {
+                        DEBUGLOG(7, "partialDecoding: dictionary match, close to dstEnd");
+                        length = MIN(length, (size_t)(oend-op));
+                    } else {
+                        goto _output_error;  /* end-of-block condition violated */
+                }   }
+
+                if (length <= (size_t)(lowPrefix-match)) {
+                    /* match fits entirely within external dictionary : just copy */
+                    memmove(op, dictEnd - (lowPrefix-match), length);
+                    op += length;
+                } else {
+                    /* match stretches into both external dictionary and current block */
+                    size_t const copySize = (size_t)(lowPrefix - match);
+                    size_t const restSize = length - copySize;
+                    LZ4_memcpy(op, dictEnd - copySize, copySize);
+                    op += copySize;
+                    if (restSize > (size_t)(op - lowPrefix)) {  /* overlap copy */
+                        BYTE* const endOfMatch = op + restSize;
+                        const BYTE* copyFrom = lowPrefix;
+                        while (op < endOfMatch) { *op++ = *copyFrom++; }
+                    } else {
+                        LZ4_memcpy(op, lowPrefix, restSize);
+                        op += restSize;
+                }   }
+                continue;
+            }
+
+            /* copy match within block */
+            cpy = op + length;
+
+            assert((op <= oend) && (oend-op >= 32));
+            if (unlikely(offset<16)) {
+                LZ4_memcpy_using_offset(op, match, cpy, offset);
+            } else {
+                LZ4_wildCopy32(op, match, cpy);
+            }
+
+            op = cpy;   /* wildcopy correction */
+        }
+    safe_decode:
+#endif
+
+        /* Main Loop : decode remaining sequences where output < FASTLOOP_SAFE_DISTANCE */
+        while (1) {
+            token = *ip++;
+            length = token >> ML_BITS;  /* literal length */
+
+            assert(!endOnInput || ip <= iend); /* ip < iend before the increment */
+
+            /* A two-stage shortcut for the most common case:
+             * 1) If the literal length is 0..14, and there is enough space,
+             * enter the shortcut and copy 16 bytes on behalf of the literals
+             * (in the fast mode, only 8 bytes can be safely copied this way).
+             * 2) Further if the match length is 4..18, copy 18 bytes in a similar
+             * manner; but we ensure that there's enough space in the output for
+             * those 18 bytes earlier, upon entering the shortcut (in other words,
+             * there is a combined check for both stages).
+             */
+            if ( (endOnInput ? length != RUN_MASK : length <= 8)
+                /* strictly "less than" on input, to re-enter the loop with at least one byte */
+              && likely((endOnInput ? ip < shortiend : 1) & (op <= shortoend)) ) {
+                /* Copy the literals */
+                LZ4_memcpy(op, ip, endOnInput ? 16 : 8);
+                op += length; ip += length;
+
+                /* The second stage: prepare for match copying, decode full info.
+                 * If it doesn't work out, the info won't be wasted. */
+                length = token & ML_MASK; /* match length */
+                offset = LZ4_readLE16(ip); ip += 2;
+                match = op - offset;
+                assert(match <= op); /* check overflow */
+
+                /* Do not deal with overlapping matches. */
+                if ( (length != ML_MASK)
+                  && (offset >= 8)
+                  && (dict==withPrefix64k || match >= lowPrefix) ) {
+                    /* Copy the match. */
+                    LZ4_memcpy(op + 0, match + 0, 8);
+                    LZ4_memcpy(op + 8, match + 8, 8);
+                    LZ4_memcpy(op +16, match +16, 2);
+                    op += length + MINMATCH;
+                    /* Both stages worked, load the next token. */
+                    continue;
+                }
+
+                /* The second stage didn't work out, but the info is ready.
+                 * Propel it right to the point of match copying. */
+                goto _copy_match;
+            }
+
+            /* decode literal length */
+            if (length == RUN_MASK) {
+                variable_length_error error = ok;
+                length += read_variable_length(&ip, iend-RUN_MASK, (int)endOnInput, (int)endOnInput, &error);
+                if (error == initial_error) { goto _output_error; }
+                if ((safeDecode) && unlikely((uptrval)(op)+length<(uptrval)(op))) { goto _output_error; } /* overflow detection */
+                if ((safeDecode) && unlikely((uptrval)(ip)+length<(uptrval)(ip))) { goto _output_error; } /* overflow detection */
+            }
+
+            /* copy literals */
+            cpy = op+length;
+#if LZ4_FAST_DEC_LOOP
+        safe_literal_copy:
+#endif
+            LZ4_STATIC_ASSERT(MFLIMIT >= WILDCOPYLENGTH);
+            if ( ((endOnInput) && ((cpy>oend-MFLIMIT) || (ip+length>iend-(2+1+LASTLITERALS))) )
+              || ((!endOnInput) && (cpy>oend-WILDCOPYLENGTH)) )
+            {
+                /* We've either hit the input parsing restriction or the output parsing restriction.
+                 * In the normal scenario, decoding a full block, it must be the last sequence,
+                 * otherwise it's an error (invalid input or dimensions).
+                 * In partialDecoding scenario, it's necessary to ensure there is no buffer overflow.
+                 */
+                if (partialDecoding) {
+                    /* Since we are partial decoding we may be in this block because of the output parsing
+                     * restriction, which is not valid since the output buffer is allowed to be undersized.
+                     */
+                    assert(endOnInput);
+                    DEBUGLOG(7, "partialDecoding: copying literals, close to input or output end")
+                    DEBUGLOG(7, "partialDecoding: literal length = %u", (unsigned)length);
+                    DEBUGLOG(7, "partialDecoding: remaining space in dstBuffer : %i", (int)(oend - op));
+                    DEBUGLOG(7, "partialDecoding: remaining space in srcBuffer : %i", (int)(iend - ip));
+                    /* Finishing in the middle of a literals segment,
+                     * due to lack of input.
+                     */
+                    if (ip+length > iend) {
+                        length = (size_t)(iend-ip);
+                        cpy = op + length;
+                    }
+                    /* Finishing in the middle of a literals segment,
+                     * due to lack of output space.
+                     */
+                    if (cpy > oend) {
+                        cpy = oend;
+                        assert(op<=oend);
+                        length = (size_t)(oend-op);
+                    }
+                } else {
+                    /* We must be on the last sequence because of the parsing limitations so check
+                     * that we exactly regenerate the original size (must be exact when !endOnInput).
+                     */
+                    if ((!endOnInput) && (cpy != oend)) { goto _output_error; }
+                     /* We must be on the last sequence (or invalid) because of the parsing limitations
+                      * so check that we exactly consume the input and don't overrun the output buffer.
+                      */
+                    if ((endOnInput) && ((ip+length != iend) || (cpy > oend))) {
+                        DEBUGLOG(6, "should have been last run of literals")
+                        DEBUGLOG(6, "ip(%p) + length(%i) = %p != iend (%p)", ip, (int)length, ip+length, iend);
+                        DEBUGLOG(6, "or cpy(%p) > oend(%p)", cpy, oend);
+                        goto _output_error;
+                    }
+                }
+                memmove(op, ip, length);  /* supports overlapping memory regions; only matters for in-place decompression scenarios */
+                ip += length;
+                op += length;
+                /* Necessarily EOF when !partialDecoding.
+                 * When partialDecoding, it is EOF if we've either
+                 * filled the output buffer or
+                 * can't proceed with reading an offset for following match.
+                 */
+                if (!partialDecoding || (cpy == oend) || (ip >= (iend-2))) {
+                    break;
+                }
+            } else {
+                LZ4_wildCopy8(op, ip, cpy);   /* may overwrite up to WILDCOPYLENGTH beyond cpy */
+                ip += length; op = cpy;
+            }
+
+            /* get offset */
+            offset = LZ4_readLE16(ip); ip+=2;
+            match = op - offset;
+
+            /* get matchlength */
+            length = token & ML_MASK;
+
+    _copy_match:
+            if (length == ML_MASK) {
+              variable_length_error error = ok;
+              length += read_variable_length(&ip, iend - LASTLITERALS + 1, (int)endOnInput, 0, &error);
+              if (error != ok) goto _output_error;
+                if ((safeDecode) && unlikely((uptrval)(op)+length<(uptrval)op)) goto _output_error;   /* overflow detection */
+            }
+            length += MINMATCH;
+
+#if LZ4_FAST_DEC_LOOP
+        safe_match_copy:
+#endif
+            if ((checkOffset) && (unlikely(match + dictSize < lowPrefix))) goto _output_error;   /* Error : offset outside buffers */
+            /* match starting within external dictionary */
+            if ((dict==usingExtDict) && (match < lowPrefix)) {
+                if (unlikely(op+length > oend-LASTLITERALS)) {
+                    if (partialDecoding) length = MIN(length, (size_t)(oend-op));
+                    else goto _output_error;   /* doesn't respect parsing restriction */
+                }
+
+                if (length <= (size_t)(lowPrefix-match)) {
+                    /* match fits entirely within external dictionary : just copy */
+                    memmove(op, dictEnd - (lowPrefix-match), length);
+                    op += length;
+                } else {
+                    /* match stretches into both external dictionary and current block */
+                    size_t const copySize = (size_t)(lowPrefix - match);
+                    size_t const restSize = length - copySize;
+                    LZ4_memcpy(op, dictEnd - copySize, copySize);
+                    op += copySize;
+                    if (restSize > (size_t)(op - lowPrefix)) {  /* overlap copy */
+                        BYTE* const endOfMatch = op + restSize;
+                        const BYTE* copyFrom = lowPrefix;
+                        while (op < endOfMatch) *op++ = *copyFrom++;
+                    } else {
+                        LZ4_memcpy(op, lowPrefix, restSize);
+                        op += restSize;
+                }   }
+                continue;
+            }
+            assert(match >= lowPrefix);
+
+            /* copy match within block */
+            cpy = op + length;
+
+            /* partialDecoding : may end anywhere within the block */
+            assert(op<=oend);
+            if (partialDecoding && (cpy > oend-MATCH_SAFEGUARD_DISTANCE)) {
+                size_t const mlen = MIN(length, (size_t)(oend-op));
+                const BYTE* const matchEnd = match + mlen;
+                BYTE* const copyEnd = op + mlen;
+                if (matchEnd > op) {   /* overlap copy */
+                    while (op < copyEnd) { *op++ = *match++; }
+                } else {
+                    LZ4_memcpy(op, match, mlen);
+                }
+                op = copyEnd;
+                if (op == oend) { break; }
+                continue;
+            }
+
+            if (unlikely(offset<8)) {
+                LZ4_write32(op, 0);   /* silence msan warning when offset==0 */
+                op[0] = match[0];
+                op[1] = match[1];
+                op[2] = match[2];
+                op[3] = match[3];
+                match += inc32table[offset];
+                LZ4_memcpy(op+4, match, 4);
+                match -= dec64table[offset];
+            } else {
+                LZ4_memcpy(op, match, 8);
+                match += 8;
+            }
+            op += 8;
+
+            if (unlikely(cpy > oend-MATCH_SAFEGUARD_DISTANCE)) {
+                BYTE* const oCopyLimit = oend - (WILDCOPYLENGTH-1);
+                if (cpy > oend-LASTLITERALS) { goto _output_error; } /* Error : last LASTLITERALS bytes must be literals (uncompressed) */
+                if (op < oCopyLimit) {
+                    LZ4_wildCopy8(op, match, oCopyLimit);
+                    match += oCopyLimit - op;
+                    op = oCopyLimit;
+                }
+                while (op < cpy) { *op++ = *match++; }
+            } else {
+                LZ4_memcpy(op, match, 8);
+                if (length > 16)  { LZ4_wildCopy8(op+8, match+8, cpy); }
+            }
+            op = cpy;   /* wildcopy correction */
+        }
+
+        /* end of decoding */
+        if (endOnInput) {
+            DEBUGLOG(5, "decoded %i bytes", (int) (((char*)op)-dst));
+           return (int) (((char*)op)-dst);     /* Nb of output bytes decoded */
+       } else {
+           return (int) (((const char*)ip)-src);   /* Nb of input bytes read */
+       }
+
+        /* Overflow error detected */
+    _output_error:
+        return (int) (-(((const char*)ip)-src))-1;
+    }
+}
+
+
+/*===== Instantiate the API decoding functions. =====*/
+
+LZ4_FORCE_O2
+int LZ4_decompress_safe(const char* source, char* dest, int compressedSize, int maxDecompressedSize)
+{
+    return LZ4_decompress_generic(source, dest, compressedSize, maxDecompressedSize,
+                                  endOnInputSize, decode_full_block, noDict,
+                                  (BYTE*)dest, NULL, 0);
+}
+
+LZ4_FORCE_O2
+int LZ4_decompress_safe_partial(const char* src, char* dst, int compressedSize, int targetOutputSize, int dstCapacity)
+{
+    dstCapacity = MIN(targetOutputSize, dstCapacity);
+    return LZ4_decompress_generic(src, dst, compressedSize, dstCapacity,
+                                  endOnInputSize, partial_decode,
+                                  noDict, (BYTE*)dst, NULL, 0);
+}
+
+LZ4_FORCE_O2
+int LZ4_decompress_fast(const char* source, char* dest, int originalSize)
+{
+    return LZ4_decompress_generic(source, dest, 0, originalSize,
+                                  endOnOutputSize, decode_full_block, withPrefix64k,
+                                  (BYTE*)dest - 64 KB, NULL, 0);
+}
+
+/*===== Instantiate a few more decoding cases, used more than once. =====*/
+
+LZ4_FORCE_O2 /* Exported, an obsolete API function. */
+int LZ4_decompress_safe_withPrefix64k(const char* source, char* dest, int compressedSize, int maxOutputSize)
+{
+    return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize,
+                                  endOnInputSize, decode_full_block, withPrefix64k,
+                                  (BYTE*)dest - 64 KB, NULL, 0);
+}
+
+/* Another obsolete API function, paired with the previous one. */
+int LZ4_decompress_fast_withPrefix64k(const char* source, char* dest, int originalSize)
+{
+    /* LZ4_decompress_fast doesn't validate match offsets,
+     * and thus serves well with any prefixed dictionary. */
+    return LZ4_decompress_fast(source, dest, originalSize);
+}
+
+LZ4_FORCE_O2
+static int LZ4_decompress_safe_withSmallPrefix(const char* source, char* dest, int compressedSize, int maxOutputSize,
+                                               size_t prefixSize)
+{
+    return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize,
+                                  endOnInputSize, decode_full_block, noDict,
+                                  (BYTE*)dest-prefixSize, NULL, 0);
+}
+
+LZ4_FORCE_O2
+int LZ4_decompress_safe_forceExtDict(const char* source, char* dest,
+                                     int compressedSize, int maxOutputSize,
+                                     const void* dictStart, size_t dictSize)
+{
+    return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize,
+                                  endOnInputSize, decode_full_block, usingExtDict,
+                                  (BYTE*)dest, (const BYTE*)dictStart, dictSize);
+}
+
+LZ4_FORCE_O2
+static int LZ4_decompress_fast_extDict(const char* source, char* dest, int originalSize,
+                                       const void* dictStart, size_t dictSize)
+{
+    return LZ4_decompress_generic(source, dest, 0, originalSize,
+                                  endOnOutputSize, decode_full_block, usingExtDict,
+                                  (BYTE*)dest, (const BYTE*)dictStart, dictSize);
+}
+
+/* The "double dictionary" mode, for use with e.g. ring buffers: the first part
+ * of the dictionary is passed as prefix, and the second via dictStart + dictSize.
+ * These routines are used only once, in LZ4_decompress_*_continue().
+ */
+LZ4_FORCE_INLINE
+int LZ4_decompress_safe_doubleDict(const char* source, char* dest, int compressedSize, int maxOutputSize,
+                                   size_t prefixSize, const void* dictStart, size_t dictSize)
+{
+    return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize,
+                                  endOnInputSize, decode_full_block, usingExtDict,
+                                  (BYTE*)dest-prefixSize, (const BYTE*)dictStart, dictSize);
+}
+
+LZ4_FORCE_INLINE
+int LZ4_decompress_fast_doubleDict(const char* source, char* dest, int originalSize,
+                                   size_t prefixSize, const void* dictStart, size_t dictSize)
+{
+    return LZ4_decompress_generic(source, dest, 0, originalSize,
+                                  endOnOutputSize, decode_full_block, usingExtDict,
+                                  (BYTE*)dest-prefixSize, (const BYTE*)dictStart, dictSize);
+}
+
+/*===== streaming decompression functions =====*/
+
+LZ4_streamDecode_t* LZ4_createStreamDecode(void)
+{
+    LZ4_streamDecode_t* lz4s = (LZ4_streamDecode_t*) ALLOC_AND_ZERO(sizeof(LZ4_streamDecode_t));
+    LZ4_STATIC_ASSERT(LZ4_STREAMDECODESIZE >= sizeof(LZ4_streamDecode_t_internal));    /* A compilation error here means LZ4_STREAMDECODESIZE is not large enough */
+    return lz4s;
+}
+
+int LZ4_freeStreamDecode (LZ4_streamDecode_t* LZ4_stream)
+{
+    if (LZ4_stream == NULL) { return 0; }  /* support free on NULL */
+    FREEMEM(LZ4_stream);
+    return 0;
+}
+
+/*! LZ4_setStreamDecode() :
+ *  Use this function to instruct where to find the dictionary.
+ *  This function is not necessary if previous data is still available where it was decoded.
+ *  Loading a size of 0 is allowed (same effect as no dictionary).
+ * @return : 1 if OK, 0 if error
+ */
+int LZ4_setStreamDecode (LZ4_streamDecode_t* LZ4_streamDecode, const char* dictionary, int dictSize)
+{
+    LZ4_streamDecode_t_internal* lz4sd = &LZ4_streamDecode->internal_donotuse;
+    lz4sd->prefixSize = (size_t) dictSize;
+    lz4sd->prefixEnd = (const BYTE*) dictionary + dictSize;
+    lz4sd->externalDict = NULL;
+    lz4sd->extDictSize  = 0;
+    return 1;
+}
+
+/*! LZ4_decoderRingBufferSize() :
+ *  when setting a ring buffer for streaming decompression (optional scenario),
+ *  provides the minimum size of this ring buffer
+ *  to be compatible with any source respecting maxBlockSize condition.
+ *  Note : in a ring buffer scenario,
+ *  blocks are presumed decompressed next to each other.
+ *  When not enough space remains for next block (remainingSize < maxBlockSize),
+ *  decoding resumes from beginning of ring buffer.
+ * @return : minimum ring buffer size,
+ *           or 0 if there is an error (invalid maxBlockSize).
+ */
+int LZ4_decoderRingBufferSize(int maxBlockSize)
+{
+    if (maxBlockSize < 0) return 0;
+    if (maxBlockSize > LZ4_MAX_INPUT_SIZE) return 0;
+    if (maxBlockSize < 16) maxBlockSize = 16;
+    return LZ4_DECODER_RING_BUFFER_SIZE(maxBlockSize);
+}
+
+/*
+*_continue() :
+    These decoding functions allow decompression of multiple blocks in "streaming" mode.
+    Previously decoded blocks must still be available at the memory position where they were decoded.
+    If it's not possible, save the relevant part of decoded data into a safe buffer,
+    and indicate where it stands using LZ4_setStreamDecode()
+*/
+LZ4_FORCE_O2
+int LZ4_decompress_safe_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* source, char* dest, int compressedSize, int maxOutputSize)
+{
+    LZ4_streamDecode_t_internal* lz4sd = &LZ4_streamDecode->internal_donotuse;
+    int result;
+
+    if (lz4sd->prefixSize == 0) {
+        /* The first call, no dictionary yet. */
+        assert(lz4sd->extDictSize == 0);
+        result = LZ4_decompress_safe(source, dest, compressedSize, maxOutputSize);
+        if (result <= 0) return result;
+        lz4sd->prefixSize = (size_t)result;
+        lz4sd->prefixEnd = (BYTE*)dest + result;
+    } else if (lz4sd->prefixEnd == (BYTE*)dest) {
+        /* They're rolling the current segment. */
+        if (lz4sd->prefixSize >= 64 KB - 1)
+            result = LZ4_decompress_safe_withPrefix64k(source, dest, compressedSize, maxOutputSize);
+        else if (lz4sd->extDictSize == 0)
+            result = LZ4_decompress_safe_withSmallPrefix(source, dest, compressedSize, maxOutputSize,
+                                                         lz4sd->prefixSize);
+        else
+            result = LZ4_decompress_safe_doubleDict(source, dest, compressedSize, maxOutputSize,
+                                                    lz4sd->prefixSize, lz4sd->externalDict, lz4sd->extDictSize);
+        if (result <= 0) return result;
+        lz4sd->prefixSize += (size_t)result;
+        lz4sd->prefixEnd  += result;
+    } else {
+        /* The buffer wraps around, or they're switching to another buffer. */
+        lz4sd->extDictSize = lz4sd->prefixSize;
+        lz4sd->externalDict = lz4sd->prefixEnd - lz4sd->extDictSize;
+        result = LZ4_decompress_safe_forceExtDict(source, dest, compressedSize, maxOutputSize,
+                                                  lz4sd->externalDict, lz4sd->extDictSize);
+        if (result <= 0) return result;
+        lz4sd->prefixSize = (size_t)result;
+        lz4sd->prefixEnd  = (BYTE*)dest + result;
+    }
+
+    return result;
+}
+
+LZ4_FORCE_O2
+int LZ4_decompress_fast_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* source, char* dest, int originalSize)
+{
+    LZ4_streamDecode_t_internal* lz4sd = &LZ4_streamDecode->internal_donotuse;
+    int result;
+    assert(originalSize >= 0);
+
+    if (lz4sd->prefixSize == 0) {
+        assert(lz4sd->extDictSize == 0);
+        result = LZ4_decompress_fast(source, dest, originalSize);
+        if (result <= 0) return result;
+        lz4sd->prefixSize = (size_t)originalSize;
+        lz4sd->prefixEnd = (BYTE*)dest + originalSize;
+    } else if (lz4sd->prefixEnd == (BYTE*)dest) {
+        if (lz4sd->prefixSize >= 64 KB - 1 || lz4sd->extDictSize == 0)
+            result = LZ4_decompress_fast(source, dest, originalSize);
+        else
+            result = LZ4_decompress_fast_doubleDict(source, dest, originalSize,
+                                                    lz4sd->prefixSize, lz4sd->externalDict, lz4sd->extDictSize);
+        if (result <= 0) return result;
+        lz4sd->prefixSize += (size_t)originalSize;
+        lz4sd->prefixEnd  += originalSize;
+    } else {
+        lz4sd->extDictSize = lz4sd->prefixSize;
+        lz4sd->externalDict = lz4sd->prefixEnd - lz4sd->extDictSize;
+        result = LZ4_decompress_fast_extDict(source, dest, originalSize,
+                                             lz4sd->externalDict, lz4sd->extDictSize);
+        if (result <= 0) return result;
+        lz4sd->prefixSize = (size_t)originalSize;
+        lz4sd->prefixEnd  = (BYTE*)dest + originalSize;
+    }
+
+    return result;
+}
+
+
+/*
+Advanced decoding functions :
+*_usingDict() :
+    These decoding functions work the same as "_continue" ones,
+    the dictionary must be explicitly provided within parameters
+*/
+
+int LZ4_decompress_safe_usingDict(const char* source, char* dest, int compressedSize, int maxOutputSize, const char* dictStart, int dictSize)
+{
+    if (dictSize==0)
+        return LZ4_decompress_safe(source, dest, compressedSize, maxOutputSize);
+    if (dictStart+dictSize == dest) {
+        if (dictSize >= 64 KB - 1) {
+            return LZ4_decompress_safe_withPrefix64k(source, dest, compressedSize, maxOutputSize);
+        }
+        assert(dictSize >= 0);
+        return LZ4_decompress_safe_withSmallPrefix(source, dest, compressedSize, maxOutputSize, (size_t)dictSize);
+    }
+    assert(dictSize >= 0);
+    return LZ4_decompress_safe_forceExtDict(source, dest, compressedSize, maxOutputSize, dictStart, (size_t)dictSize);
+}
+
+int LZ4_decompress_fast_usingDict(const char* source, char* dest, int originalSize, const char* dictStart, int dictSize)
+{
+    if (dictSize==0 || dictStart+dictSize == dest)
+        return LZ4_decompress_fast(source, dest, originalSize);
+    assert(dictSize >= 0);
+    return LZ4_decompress_fast_extDict(source, dest, originalSize, dictStart, (size_t)dictSize);
+}
+
+
+/*=*************************************************
+*  Obsolete Functions
+***************************************************/
+/* obsolete compression functions */
+int LZ4_compress_limitedOutput(const char* source, char* dest, int inputSize, int maxOutputSize)
+{
+    return LZ4_compress_default(source, dest, inputSize, maxOutputSize);
+}
+int LZ4_compress(const char* src, char* dest, int srcSize)
+{
+    return LZ4_compress_default(src, dest, srcSize, LZ4_compressBound(srcSize));
+}
+int LZ4_compress_limitedOutput_withState (void* state, const char* src, char* dst, int srcSize, int dstSize)
+{
+    return LZ4_compress_fast_extState(state, src, dst, srcSize, dstSize, 1);
+}
+int LZ4_compress_withState (void* state, const char* src, char* dst, int srcSize)
+{
+    return LZ4_compress_fast_extState(state, src, dst, srcSize, LZ4_compressBound(srcSize), 1);
+}
+int LZ4_compress_limitedOutput_continue (LZ4_stream_t* LZ4_stream, const char* src, char* dst, int srcSize, int dstCapacity)
+{
+    return LZ4_compress_fast_continue(LZ4_stream, src, dst, srcSize, dstCapacity, 1);
+}
+int LZ4_compress_continue (LZ4_stream_t* LZ4_stream, const char* source, char* dest, int inputSize)
+{
+    return LZ4_compress_fast_continue(LZ4_stream, source, dest, inputSize, LZ4_compressBound(inputSize), 1);
+}
+
+/*
+These decompression functions are deprecated and should no longer be used.
+They are only provided here for compatibility with older user programs.
+- LZ4_uncompress is totally equivalent to LZ4_decompress_fast
+- LZ4_uncompress_unknownOutputSize is totally equivalent to LZ4_decompress_safe
+*/
+int LZ4_uncompress (const char* source, char* dest, int outputSize)
+{
+    return LZ4_decompress_fast(source, dest, outputSize);
+}
+int LZ4_uncompress_unknownOutputSize (const char* source, char* dest, int isize, int maxOutputSize)
+{
+    return LZ4_decompress_safe(source, dest, isize, maxOutputSize);
+}
+
+/* Obsolete Streaming functions */
+
+int LZ4_sizeofStreamState(void) { return LZ4_STREAMSIZE; }
+
+int LZ4_resetStreamState(void* state, char* inputBuffer)
+{
+    (void)inputBuffer;
+    LZ4_resetStream((LZ4_stream_t*)state);
+    return 0;
+}
+
+void* LZ4_create (char* inputBuffer)
+{
+    (void)inputBuffer;
+    return LZ4_createStream();
+}
+
+char* LZ4_slideInputBuffer (void* state)
+{
+    /* avoid const char * -> char * conversion warning */
+    return (char *)(uptrval)((LZ4_stream_t*)state)->internal_donotuse.dictionary;
+}
+
+#endif   /* LZ4_COMMONDEFS_ONLY */
diff --git a/mfbt/lz4/lz4.h b/mfbt/lz4/lz4.h
new file mode 100644
index 0000000000..7ab1e483a9
--- /dev/null
+++ b/mfbt/lz4/lz4.h
@@ -0,0 +1,774 @@
+/*
+ *  LZ4 - Fast LZ compression algorithm
+ *  Header File
+ *  Copyright (C) 2011-present, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+    - LZ4 homepage : http://www.lz4.org
+    - LZ4 source repository : https://github.com/lz4/lz4
+*/
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef LZ4_H_2983827168210
+#define LZ4_H_2983827168210
+
+/* --- Dependency --- */
+#include <stddef.h>   /* size_t */
+
+
+/**
+  Introduction
+
+  LZ4 is lossless compression algorithm, providing compression speed >500 MB/s per core,
+  scalable with multi-cores CPU. It features an extremely fast decoder, with speed in
+  multiple GB/s per core, typically reaching RAM speed limits on multi-core systems.
+
+  The LZ4 compression library provides in-memory compression and decompression functions.
+  It gives full buffer control to user.
+  Compression can be done in:
+    - a single step (described as Simple Functions)
+    - a single step, reusing a context (described in Advanced Functions)
+    - unbounded multiple steps (described as Streaming compression)
+
+  lz4.h generates and decodes LZ4-compressed blocks (doc/lz4_Block_format.md).
+  Decompressing such a compressed block requires additional metadata.
+  Exact metadata depends on exact decompression function.
+  For the typical case of LZ4_decompress_safe(),
+  metadata includes block's compressed size, and maximum bound of decompressed size.
+  Each application is free to encode and pass such metadata in whichever way it wants.
+
+  lz4.h only handle blocks, it can not generate Frames.
+
+  Blocks are different from Frames (doc/lz4_Frame_format.md).
+  Frames bundle both blocks and metadata in a specified manner.
+  Embedding metadata is required for compressed data to be self-contained and portable.
+  Frame format is delivered through a companion API, declared in lz4frame.h.
+  The `lz4` CLI can only manage frames.
+*/
+
+/*^***************************************************************
+*  Export parameters
+*****************************************************************/
+/*
+*  LZ4_DLL_EXPORT :
+*  Enable exporting of functions when building a Windows DLL
+*  LZ4LIB_VISIBILITY :
+*  Control library symbols visibility.
+*/
+#ifndef LZ4LIB_VISIBILITY
+#  if defined(__GNUC__) && (__GNUC__ >= 4)
+#    define LZ4LIB_VISIBILITY __attribute__ ((visibility ("default")))
+#  else
+#    define LZ4LIB_VISIBILITY
+#  endif
+#endif
+#if defined(LZ4_DLL_EXPORT) && (LZ4_DLL_EXPORT==1)
+#  define LZ4LIB_API __declspec(dllexport) LZ4LIB_VISIBILITY
+#elif defined(LZ4_DLL_IMPORT) && (LZ4_DLL_IMPORT==1)
+#  define LZ4LIB_API __declspec(dllimport) LZ4LIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#else
+#  define LZ4LIB_API LZ4LIB_VISIBILITY
+#endif
+
+/*------   Version   ------*/
+#define LZ4_VERSION_MAJOR    1    /* for breaking interface changes  */
+#define LZ4_VERSION_MINOR    9    /* for new (non-breaking) interface capabilities */
+#define LZ4_VERSION_RELEASE  3    /* for tweaks, bug-fixes, or development */
+
+#define LZ4_VERSION_NUMBER (LZ4_VERSION_MAJOR *100*100 + LZ4_VERSION_MINOR *100 + LZ4_VERSION_RELEASE)
+
+#define LZ4_LIB_VERSION LZ4_VERSION_MAJOR.LZ4_VERSION_MINOR.LZ4_VERSION_RELEASE
+#define LZ4_QUOTE(str) #str
+#define LZ4_EXPAND_AND_QUOTE(str) LZ4_QUOTE(str)
+#define LZ4_VERSION_STRING LZ4_EXPAND_AND_QUOTE(LZ4_LIB_VERSION)
+
+LZ4LIB_API int LZ4_versionNumber (void);  /**< library version number; useful to check dll version */
+LZ4LIB_API const char* LZ4_versionString (void);   /**< library version string; useful to check dll version */
+
+
+/*-************************************
+*  Tuning parameter
+**************************************/
+/*!
+ * LZ4_MEMORY_USAGE :
+ * Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+ * Increasing memory usage improves compression ratio.
+ * Reduced memory usage may improve speed, thanks to better cache locality.
+ * Default value is 14, for 16KB, which nicely fits into Intel x86 L1 cache
+ */
+#ifndef LZ4_MEMORY_USAGE
+# define LZ4_MEMORY_USAGE 14
+#endif
+
+
+/*-************************************
+*  Simple Functions
+**************************************/
+/*! LZ4_compress_default() :
+ *  Compresses 'srcSize' bytes from buffer 'src'
+ *  into already allocated 'dst' buffer of size 'dstCapacity'.
+ *  Compression is guaranteed to succeed if 'dstCapacity' >= LZ4_compressBound(srcSize).
+ *  It also runs faster, so it's a recommended setting.
+ *  If the function cannot compress 'src' into a more limited 'dst' budget,
+ *  compression stops *immediately*, and the function result is zero.
+ *  In which case, 'dst' content is undefined (invalid).
+ *      srcSize : max supported value is LZ4_MAX_INPUT_SIZE.
+ *      dstCapacity : size of buffer 'dst' (which must be already allocated)
+ *     @return  : the number of bytes written into buffer 'dst' (necessarily <= dstCapacity)
+ *                or 0 if compression fails
+ * Note : This function is protected against buffer overflow scenarios (never writes outside 'dst' buffer, nor read outside 'source' buffer).
+ */
+LZ4LIB_API int LZ4_compress_default(const char* src, char* dst, int srcSize, int dstCapacity);
+
+/*! LZ4_decompress_safe() :
+ *  compressedSize : is the exact complete size of the compressed block.
+ *  dstCapacity : is the size of destination buffer (which must be already allocated), presumed an upper bound of decompressed size.
+ * @return : the number of bytes decompressed into destination buffer (necessarily <= dstCapacity)
+ *           If destination buffer is not large enough, decoding will stop and output an error code (negative value).
+ *           If the source stream is detected malformed, the function will stop decoding and return a negative result.
+ * Note 1 : This function is protected against malicious data packets :
+ *          it will never writes outside 'dst' buffer, nor read outside 'source' buffer,
+ *          even if the compressed block is maliciously modified to order the decoder to do these actions.
+ *          In such case, the decoder stops immediately, and considers the compressed block malformed.
+ * Note 2 : compressedSize and dstCapacity must be provided to the function, the compressed block does not contain them.
+ *          The implementation is free to send / store / derive this information in whichever way is most beneficial.
+ *          If there is a need for a different format which bundles together both compressed data and its metadata, consider looking at lz4frame.h instead.
+ */
+LZ4LIB_API int LZ4_decompress_safe (const char* src, char* dst, int compressedSize, int dstCapacity);
+
+
+/*-************************************
+*  Advanced Functions
+**************************************/
+#define LZ4_MAX_INPUT_SIZE        0x7E000000   /* 2 113 929 216 bytes */
+#define LZ4_COMPRESSBOUND(isize)  ((unsigned)(isize) > (unsigned)LZ4_MAX_INPUT_SIZE ? 0 : (isize) + ((isize)/255) + 16)
+
+/*! LZ4_compressBound() :
+    Provides the maximum size that LZ4 compression may output in a "worst case" scenario (input data not compressible)
+    This function is primarily useful for memory allocation purposes (destination buffer size).
+    Macro LZ4_COMPRESSBOUND() is also provided for compilation-time evaluation (stack memory allocation for example).
+    Note that LZ4_compress_default() compresses faster when dstCapacity is >= LZ4_compressBound(srcSize)
+        inputSize  : max supported value is LZ4_MAX_INPUT_SIZE
+        return : maximum output size in a "worst case" scenario
+              or 0, if input size is incorrect (too large or negative)
+*/
+LZ4LIB_API int LZ4_compressBound(int inputSize);
+
+/*! LZ4_compress_fast() :
+    Same as LZ4_compress_default(), but allows selection of "acceleration" factor.
+    The larger the acceleration value, the faster the algorithm, but also the lesser the compression.
+    It's a trade-off. It can be fine tuned, with each successive value providing roughly +~3% to speed.
+    An acceleration value of "1" is the same as regular LZ4_compress_default()
+    Values <= 0 will be replaced by LZ4_ACCELERATION_DEFAULT (currently == 1, see lz4.c).
+    Values > LZ4_ACCELERATION_MAX will be replaced by LZ4_ACCELERATION_MAX (currently == 65537, see lz4.c).
+*/
+LZ4LIB_API int LZ4_compress_fast (const char* src, char* dst, int srcSize, int dstCapacity, int acceleration);
+
+
+/*! LZ4_compress_fast_extState() :
+ *  Same as LZ4_compress_fast(), using an externally allocated memory space for its state.
+ *  Use LZ4_sizeofState() to know how much memory must be allocated,
+ *  and allocate it on 8-bytes boundaries (using `malloc()` typically).
+ *  Then, provide this buffer as `void* state` to compression function.
+ */
+LZ4LIB_API int LZ4_sizeofState(void);
+LZ4LIB_API int LZ4_compress_fast_extState (void* state, const char* src, char* dst, int srcSize, int dstCapacity, int acceleration);
+
+
+/*! LZ4_compress_destSize() :
+ *  Reverse the logic : compresses as much data as possible from 'src' buffer
+ *  into already allocated buffer 'dst', of size >= 'targetDestSize'.
+ *  This function either compresses the entire 'src' content into 'dst' if it's large enough,
+ *  or fill 'dst' buffer completely with as much data as possible from 'src'.
+ *  note: acceleration parameter is fixed to "default".
+ *
+ * *srcSizePtr : will be modified to indicate how many bytes where read from 'src' to fill 'dst'.
+ *               New value is necessarily <= input value.
+ * @return : Nb bytes written into 'dst' (necessarily <= targetDestSize)
+ *           or 0 if compression fails.
+ *
+ * Note : from v1.8.2 to v1.9.1, this function had a bug (fixed un v1.9.2+):
+ *        the produced compressed content could, in specific circumstances,
+ *        require to be decompressed into a destination buffer larger
+ *        by at least 1 byte than the content to decompress.
+ *        If an application uses `LZ4_compress_destSize()`,
+ *        it's highly recommended to update liblz4 to v1.9.2 or better.
+ *        If this can't be done or ensured,
+ *        the receiving decompression function should provide
+ *        a dstCapacity which is > decompressedSize, by at least 1 byte.
+ *        See https://github.com/lz4/lz4/issues/859 for details
+ */
+LZ4LIB_API int LZ4_compress_destSize (const char* src, char* dst, int* srcSizePtr, int targetDstSize);
+
+
+/*! LZ4_decompress_safe_partial() :
+ *  Decompress an LZ4 compressed block, of size 'srcSize' at position 'src',
+ *  into destination buffer 'dst' of size 'dstCapacity'.
+ *  Up to 'targetOutputSize' bytes will be decoded.
+ *  The function stops decoding on reaching this objective.
+ *  This can be useful to boost performance
+ *  whenever only the beginning of a block is required.
+ *
+ * @return : the number of bytes decoded in `dst` (necessarily <= targetOutputSize)
+ *           If source stream is detected malformed, function returns a negative result.
+ *
+ *  Note 1 : @return can be < targetOutputSize, if compressed block contains less data.
+ *
+ *  Note 2 : targetOutputSize must be <= dstCapacity
+ *
+ *  Note 3 : this function effectively stops decoding on reaching targetOutputSize,
+ *           so dstCapacity is kind of redundant.
+ *           This is because in older versions of this function,
+ *           decoding operation would still write complete sequences.
+ *           Therefore, there was no guarantee that it would stop writing at exactly targetOutputSize,
+ *           it could write more bytes, though only up to dstCapacity.
+ *           Some "margin" used to be required for this operation to work properly.
+ *           Thankfully, this is no longer necessary.
+ *           The function nonetheless keeps the same signature, in an effort to preserve API compatibility.
+ *
+ *  Note 4 : If srcSize is the exact size of the block,
+ *           then targetOutputSize can be any value,
+ *           including larger than the block's decompressed size.
+ *           The function will, at most, generate block's decompressed size.
+ *
+ *  Note 5 : If srcSize is _larger_ than block's compressed size,
+ *           then targetOutputSize **MUST** be <= block's decompressed size.
+ *           Otherwise, *silent corruption will occur*.
+ */
+LZ4LIB_API int LZ4_decompress_safe_partial (const char* src, char* dst, int srcSize, int targetOutputSize, int dstCapacity);
+
+
+/*-*********************************************
+*  Streaming Compression Functions
+***********************************************/
+typedef union LZ4_stream_u LZ4_stream_t;  /* incomplete type (defined later) */
+
+LZ4LIB_API LZ4_stream_t* LZ4_createStream(void);
+LZ4LIB_API int           LZ4_freeStream (LZ4_stream_t* streamPtr);
+
+/*! LZ4_resetStream_fast() : v1.9.0+
+ *  Use this to prepare an LZ4_stream_t for a new chain of dependent blocks
+ *  (e.g., LZ4_compress_fast_continue()).
+ *
+ *  An LZ4_stream_t must be initialized once before usage.
+ *  This is automatically done when created by LZ4_createStream().
+ *  However, should the LZ4_stream_t be simply declared on stack (for example),
+ *  it's necessary to initialize it first, using LZ4_initStream().
+ *
+ *  After init, start any new stream with LZ4_resetStream_fast().
+ *  A same LZ4_stream_t can be re-used multiple times consecutively
+ *  and compress multiple streams,
+ *  provided that it starts each new stream with LZ4_resetStream_fast().
+ *
+ *  LZ4_resetStream_fast() is much faster than LZ4_initStream(),
+ *  but is not compatible with memory regions containing garbage data.
+ *
+ *  Note: it's only useful to call LZ4_resetStream_fast()
+ *        in the context of streaming compression.
+ *        The *extState* functions perform their own resets.
+ *        Invoking LZ4_resetStream_fast() before is redundant, and even counterproductive.
+ */
+LZ4LIB_API void LZ4_resetStream_fast (LZ4_stream_t* streamPtr);
+
+/*! LZ4_loadDict() :
+ *  Use this function to reference a static dictionary into LZ4_stream_t.
+ *  The dictionary must remain available during compression.
+ *  LZ4_loadDict() triggers a reset, so any previous data will be forgotten.
+ *  The same dictionary will have to be loaded on decompression side for successful decoding.
+ *  Dictionary are useful for better compression of small data (KB range).
+ *  While LZ4 accept any input as dictionary,
+ *  results are generally better when using Zstandard's Dictionary Builder.
+ *  Loading a size of 0 is allowed, and is the same as reset.
+ * @return : loaded dictionary size, in bytes (necessarily <= 64 KB)
+ */
+LZ4LIB_API int LZ4_loadDict (LZ4_stream_t* streamPtr, const char* dictionary, int dictSize);
+
+/*! LZ4_compress_fast_continue() :
+ *  Compress 'src' content using data from previously compressed blocks, for better compression ratio.
+ * 'dst' buffer must be already allocated.
+ *  If dstCapacity >= LZ4_compressBound(srcSize), compression is guaranteed to succeed, and runs faster.
+ *
+ * @return : size of compressed block
+ *           or 0 if there is an error (typically, cannot fit into 'dst').
+ *
+ *  Note 1 : Each invocation to LZ4_compress_fast_continue() generates a new block.
+ *           Each block has precise boundaries.
+ *           Each block must be decompressed separately, calling LZ4_decompress_*() with relevant metadata.
+ *           It's not possible to append blocks together and expect a single invocation of LZ4_decompress_*() to decompress them together.
+ *
+ *  Note 2 : The previous 64KB of source data is __assumed__ to remain present, unmodified, at same address in memory !
+ *
+ *  Note 3 : When input is structured as a double-buffer, each buffer can have any size, including < 64 KB.
+ *           Make sure that buffers are separated, by at least one byte.
+ *           This construction ensures that each block only depends on previous block.
+ *
+ *  Note 4 : If input buffer is a ring-buffer, it can have any size, including < 64 KB.
+ *
+ *  Note 5 : After an error, the stream status is undefined (invalid), it can only be reset or freed.
+ */
+LZ4LIB_API int LZ4_compress_fast_continue (LZ4_stream_t* streamPtr, const char* src, char* dst, int srcSize, int dstCapacity, int acceleration);
+
+/*! LZ4_saveDict() :
+ *  If last 64KB data cannot be guaranteed to remain available at its current memory location,
+ *  save it into a safer place (char* safeBuffer).
+ *  This is schematically equivalent to a memcpy() followed by LZ4_loadDict(),
+ *  but is much faster, because LZ4_saveDict() doesn't need to rebuild tables.
+ * @return : saved dictionary size in bytes (necessarily <= maxDictSize), or 0 if error.
+ */
+LZ4LIB_API int LZ4_saveDict (LZ4_stream_t* streamPtr, char* safeBuffer, int maxDictSize);
+
+
+/*-**********************************************
+*  Streaming Decompression Functions
+*  Bufferless synchronous API
+************************************************/
+typedef union LZ4_streamDecode_u LZ4_streamDecode_t;   /* tracking context */
+
+/*! LZ4_createStreamDecode() and LZ4_freeStreamDecode() :
+ *  creation / destruction of streaming decompression tracking context.
+ *  A tracking context can be re-used multiple times.
+ */
+LZ4LIB_API LZ4_streamDecode_t* LZ4_createStreamDecode(void);
+LZ4LIB_API int                 LZ4_freeStreamDecode (LZ4_streamDecode_t* LZ4_stream);
+
+/*! LZ4_setStreamDecode() :
+ *  An LZ4_streamDecode_t context can be allocated once and re-used multiple times.
+ *  Use this function to start decompression of a new stream of blocks.
+ *  A dictionary can optionally be set. Use NULL or size 0 for a reset order.
+ *  Dictionary is presumed stable : it must remain accessible and unmodified during next decompression.
+ * @return : 1 if OK, 0 if error
+ */
+LZ4LIB_API int LZ4_setStreamDecode (LZ4_streamDecode_t* LZ4_streamDecode, const char* dictionary, int dictSize);
+
+/*! LZ4_decoderRingBufferSize() : v1.8.2+
+ *  Note : in a ring buffer scenario (optional),
+ *  blocks are presumed decompressed next to each other
+ *  up to the moment there is not enough remaining space for next block (remainingSize < maxBlockSize),
+ *  at which stage it resumes from beginning of ring buffer.
+ *  When setting such a ring buffer for streaming decompression,
+ *  provides the minimum size of this ring buffer
+ *  to be compatible with any source respecting maxBlockSize condition.
+ * @return : minimum ring buffer size,
+ *           or 0 if there is an error (invalid maxBlockSize).
+ */
+LZ4LIB_API int LZ4_decoderRingBufferSize(int maxBlockSize);
+#define LZ4_DECODER_RING_BUFFER_SIZE(maxBlockSize) (65536 + 14 + (maxBlockSize))  /* for static allocation; maxBlockSize presumed valid */
+
+/*! LZ4_decompress_*_continue() :
+ *  These decoding functions allow decompression of consecutive blocks in "streaming" mode.
+ *  A block is an unsplittable entity, it must be presented entirely to a decompression function.
+ *  Decompression functions only accepts one block at a time.
+ *  The last 64KB of previously decoded data *must* remain available and unmodified at the memory position where they were decoded.
+ *  If less than 64KB of data has been decoded, all the data must be present.
+ *
+ *  Special : if decompression side sets a ring buffer, it must respect one of the following conditions :
+ *  - Decompression buffer size is _at least_ LZ4_decoderRingBufferSize(maxBlockSize).
+ *    maxBlockSize is the maximum size of any single block. It can have any value > 16 bytes.
+ *    In which case, encoding and decoding buffers do not need to be synchronized.
+ *    Actually, data can be produced by any source compliant with LZ4 format specification, and respecting maxBlockSize.
+ *  - Synchronized mode :
+ *    Decompression buffer size is _exactly_ the same as compression buffer size,
+ *    and follows exactly same update rule (block boundaries at same positions),
+ *    and decoding function is provided with exact decompressed size of each block (exception for last block of the stream),
+ *    _then_ decoding & encoding ring buffer can have any size, including small ones ( < 64 KB).
+ *  - Decompression buffer is larger than encoding buffer, by a minimum of maxBlockSize more bytes.
+ *    In which case, encoding and decoding buffers do not need to be synchronized,
+ *    and encoding ring buffer can have any size, including small ones ( < 64 KB).
+ *
+ *  Whenever these conditions are not possible,
+ *  save the last 64KB of decoded data into a safe buffer where it can't be modified during decompression,
+ *  then indicate where this data is saved using LZ4_setStreamDecode(), before decompressing next block.
+*/
+LZ4LIB_API int LZ4_decompress_safe_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* src, char* dst, int srcSize, int dstCapacity);
+
+
+/*! LZ4_decompress_*_usingDict() :
+ *  These decoding functions work the same as
+ *  a combination of LZ4_setStreamDecode() followed by LZ4_decompress_*_continue()
+ *  They are stand-alone, and don't need an LZ4_streamDecode_t structure.
+ *  Dictionary is presumed stable : it must remain accessible and unmodified during decompression.
+ *  Performance tip : Decompression speed can be substantially increased
+ *                    when dst == dictStart + dictSize.
+ */
+LZ4LIB_API int LZ4_decompress_safe_usingDict (const char* src, char* dst, int srcSize, int dstCapcity, const char* dictStart, int dictSize);
+
+#endif /* LZ4_H_2983827168210 */
+
+
+/*^*************************************
+ * !!!!!!   STATIC LINKING ONLY   !!!!!!
+ ***************************************/
+
+/*-****************************************************************************
+ * Experimental section
+ *
+ * Symbols declared in this section must be considered unstable. Their
+ * signatures or semantics may change, or they may be removed altogether in the
+ * future. They are therefore only safe to depend on when the caller is
+ * statically linked against the library.
+ *
+ * To protect against unsafe usage, not only are the declarations guarded,
+ * the definitions are hidden by default
+ * when building LZ4 as a shared/dynamic library.
+ *
+ * In order to access these declarations,
+ * define LZ4_STATIC_LINKING_ONLY in your application
+ * before including LZ4's headers.
+ *
+ * In order to make their implementations accessible dynamically, you must
+ * define LZ4_PUBLISH_STATIC_FUNCTIONS when building the LZ4 library.
+ ******************************************************************************/
+
+#ifdef LZ4_STATIC_LINKING_ONLY
+
+#ifndef LZ4_STATIC_3504398509
+#define LZ4_STATIC_3504398509
+
+#ifdef LZ4_PUBLISH_STATIC_FUNCTIONS
+#define LZ4LIB_STATIC_API LZ4LIB_API
+#else
+#define LZ4LIB_STATIC_API
+#endif
+
+
+/*! LZ4_compress_fast_extState_fastReset() :
+ *  A variant of LZ4_compress_fast_extState().
+ *
+ *  Using this variant avoids an expensive initialization step.
+ *  It is only safe to call if the state buffer is known to be correctly initialized already
+ *  (see above comment on LZ4_resetStream_fast() for a definition of "correctly initialized").
+ *  From a high level, the difference is that
+ *  this function initializes the provided state with a call to something like LZ4_resetStream_fast()
+ *  while LZ4_compress_fast_extState() starts with a call to LZ4_resetStream().
+ */
+LZ4LIB_STATIC_API int LZ4_compress_fast_extState_fastReset (void* state, const char* src, char* dst, int srcSize, int dstCapacity, int acceleration);
+
+/*! LZ4_attach_dictionary() :
+ *  This is an experimental API that allows
+ *  efficient use of a static dictionary many times.
+ *
+ *  Rather than re-loading the dictionary buffer into a working context before
+ *  each compression, or copying a pre-loaded dictionary's LZ4_stream_t into a
+ *  working LZ4_stream_t, this function introduces a no-copy setup mechanism,
+ *  in which the working stream references the dictionary stream in-place.
+ *
+ *  Several assumptions are made about the state of the dictionary stream.
+ *  Currently, only streams which have been prepared by LZ4_loadDict() should
+ *  be expected to work.
+ *
+ *  Alternatively, the provided dictionaryStream may be NULL,
+ *  in which case any existing dictionary stream is unset.
+ *
+ *  If a dictionary is provided, it replaces any pre-existing stream history.
+ *  The dictionary contents are the only history that can be referenced and
+ *  logically immediately precede the data compressed in the first subsequent
+ *  compression call.
+ *
+ *  The dictionary will only remain attached to the working stream through the
+ *  first compression call, at the end of which it is cleared. The dictionary
+ *  stream (and source buffer) must remain in-place / accessible / unchanged
+ *  through the completion of the first compression call on the stream.
+ */
+LZ4LIB_STATIC_API void LZ4_attach_dictionary(LZ4_stream_t* workingStream, const LZ4_stream_t* dictionaryStream);
+
+
+/*! In-place compression and decompression
+ *
+ * It's possible to have input and output sharing the same buffer,
+ * for highly contrained memory environments.
+ * In both cases, it requires input to lay at the end of the buffer,
+ * and decompression to start at beginning of the buffer.
+ * Buffer size must feature some margin, hence be larger than final size.
+ *
+ * |<------------------------buffer--------------------------------->|
+ *                             |<-----------compressed data--------->|
+ * |<-----------decompressed size------------------>|
+ *                                                  |<----margin---->|
+ *
+ * This technique is more useful for decompression,
+ * since decompressed size is typically larger,
+ * and margin is short.
+ *
+ * In-place decompression will work inside any buffer
+ * which size is >= LZ4_DECOMPRESS_INPLACE_BUFFER_SIZE(decompressedSize).
+ * This presumes that decompressedSize > compressedSize.
+ * Otherwise, it means compression actually expanded data,
+ * and it would be more efficient to store such data with a flag indicating it's not compressed.
+ * This can happen when data is not compressible (already compressed, or encrypted).
+ *
+ * For in-place compression, margin is larger, as it must be able to cope with both
+ * history preservation, requiring input data to remain unmodified up to LZ4_DISTANCE_MAX,
+ * and data expansion, which can happen when input is not compressible.
+ * As a consequence, buffer size requirements are much higher,
+ * and memory savings offered by in-place compression are more limited.
+ *
+ * There are ways to limit this cost for compression :
+ * - Reduce history size, by modifying LZ4_DISTANCE_MAX.
+ *   Note that it is a compile-time constant, so all compressions will apply this limit.
+ *   Lower values will reduce compression ratio, except when input_size < LZ4_DISTANCE_MAX,
+ *   so it's a reasonable trick when inputs are known to be small.
+ * - Require the compressor to deliver a "maximum compressed size".
+ *   This is the `dstCapacity` parameter in `LZ4_compress*()`.
+ *   When this size is < LZ4_COMPRESSBOUND(inputSize), then compression can fail,
+ *   in which case, the return code will be 0 (zero).
+ *   The caller must be ready for these cases to happen,
+ *   and typically design a backup scheme to send data uncompressed.
+ * The combination of both techniques can significantly reduce
+ * the amount of margin required for in-place compression.
+ *
+ * In-place compression can work in any buffer
+ * which size is >= (maxCompressedSize)
+ * with maxCompressedSize == LZ4_COMPRESSBOUND(srcSize) for guaranteed compression success.
+ * LZ4_COMPRESS_INPLACE_BUFFER_SIZE() depends on both maxCompressedSize and LZ4_DISTANCE_MAX,
+ * so it's possible to reduce memory requirements by playing with them.
+ */
+
+#define LZ4_DECOMPRESS_INPLACE_MARGIN(compressedSize)          (((compressedSize) >> 8) + 32)
+#define LZ4_DECOMPRESS_INPLACE_BUFFER_SIZE(decompressedSize)   ((decompressedSize) + LZ4_DECOMPRESS_INPLACE_MARGIN(decompressedSize))  /**< note: presumes that compressedSize < decompressedSize. note2: margin is overestimated a bit, since it could use compressedSize instead */
+
+#ifndef LZ4_DISTANCE_MAX   /* history window size; can be user-defined at compile time */
+#  define LZ4_DISTANCE_MAX 65535   /* set to maximum value by default */
+#endif
+
+#define LZ4_COMPRESS_INPLACE_MARGIN                           (LZ4_DISTANCE_MAX + 32)   /* LZ4_DISTANCE_MAX can be safely replaced by srcSize when it's smaller */
+#define LZ4_COMPRESS_INPLACE_BUFFER_SIZE(maxCompressedSize)   ((maxCompressedSize) + LZ4_COMPRESS_INPLACE_MARGIN)  /**< maxCompressedSize is generally LZ4_COMPRESSBOUND(inputSize), but can be set to any lower value, with the risk that compression can fail (return code 0(zero)) */
+
+#endif   /* LZ4_STATIC_3504398509 */
+#endif   /* LZ4_STATIC_LINKING_ONLY */
+
+
+
+#ifndef LZ4_H_98237428734687
+#define LZ4_H_98237428734687
+
+/*-************************************************************
+ *  Private Definitions
+ **************************************************************
+ * Do not use these definitions directly.
+ * They are only exposed to allow static allocation of `LZ4_stream_t` and `LZ4_streamDecode_t`.
+ * Accessing members will expose user code to API and/or ABI break in future versions of the library.
+ **************************************************************/
+#define LZ4_HASHLOG   (LZ4_MEMORY_USAGE-2)
+#define LZ4_HASHTABLESIZE (1 << LZ4_MEMORY_USAGE)
+#define LZ4_HASH_SIZE_U32 (1 << LZ4_HASHLOG)       /* required as macro for static allocation */
+
+#if defined(__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# include <stdint.h>
+  typedef  int8_t  LZ4_i8;
+  typedef uint8_t  LZ4_byte;
+  typedef uint16_t LZ4_u16;
+  typedef uint32_t LZ4_u32;
+#else
+  typedef   signed char  LZ4_i8;
+  typedef unsigned char  LZ4_byte;
+  typedef unsigned short LZ4_u16;
+  typedef unsigned int   LZ4_u32;
+#endif
+
+typedef struct LZ4_stream_t_internal LZ4_stream_t_internal;
+struct LZ4_stream_t_internal {
+    LZ4_u32 hashTable[LZ4_HASH_SIZE_U32];
+    LZ4_u32 currentOffset;
+    LZ4_u32 tableType;
+    const LZ4_byte* dictionary;
+    const LZ4_stream_t_internal* dictCtx;
+    LZ4_u32 dictSize;
+};
+
+typedef struct {
+    const LZ4_byte* externalDict;
+    size_t extDictSize;
+    const LZ4_byte* prefixEnd;
+    size_t prefixSize;
+} LZ4_streamDecode_t_internal;
+
+
+/*! LZ4_stream_t :
+ *  Do not use below internal definitions directly !
+ *  Declare or allocate an LZ4_stream_t instead.
+ *  LZ4_stream_t can also be created using LZ4_createStream(), which is recommended.
+ *  The structure definition can be convenient for static allocation
+ *  (on stack, or as part of larger structure).
+ *  Init this structure with LZ4_initStream() before first use.
+ *  note : only use this definition in association with static linking !
+ *  this definition is not API/ABI safe, and may change in future versions.
+ */
+#define LZ4_STREAMSIZE       16416  /* static size, for inter-version compatibility */
+#define LZ4_STREAMSIZE_VOIDP (LZ4_STREAMSIZE / sizeof(void*))
+union LZ4_stream_u {
+    void* table[LZ4_STREAMSIZE_VOIDP];
+    LZ4_stream_t_internal internal_donotuse;
+}; /* previously typedef'd to LZ4_stream_t */
+
+
+/*! LZ4_initStream() : v1.9.0+
+ *  An LZ4_stream_t structure must be initialized at least once.
+ *  This is automatically done when invoking LZ4_createStream(),
+ *  but it's not when the structure is simply declared on stack (for example).
+ *
+ *  Use LZ4_initStream() to properly initialize a newly declared LZ4_stream_t.
+ *  It can also initialize any arbitrary buffer of sufficient size,
+ *  and will @return a pointer of proper type upon initialization.
+ *
+ *  Note : initialization fails if size and alignment conditions are not respected.
+ *         In which case, the function will @return NULL.
+ *  Note2: An LZ4_stream_t structure guarantees correct alignment and size.
+ *  Note3: Before v1.9.0, use LZ4_resetStream() instead
+ */
+LZ4LIB_API LZ4_stream_t* LZ4_initStream (void* buffer, size_t size);
+
+
+/*! LZ4_streamDecode_t :
+ *  information structure to track an LZ4 stream during decompression.
+ *  init this structure  using LZ4_setStreamDecode() before first use.
+ *  note : only use in association with static linking !
+ *         this definition is not API/ABI safe,
+ *         and may change in a future version !
+ */
+#define LZ4_STREAMDECODESIZE_U64 (4 + ((sizeof(void*)==16) ? 2 : 0) /*AS-400*/ )
+#define LZ4_STREAMDECODESIZE     (LZ4_STREAMDECODESIZE_U64 * sizeof(unsigned long long))
+union LZ4_streamDecode_u {
+    unsigned long long table[LZ4_STREAMDECODESIZE_U64];
+    LZ4_streamDecode_t_internal internal_donotuse;
+} ;   /* previously typedef'd to LZ4_streamDecode_t */
+
+
+
+/*-************************************
+*  Obsolete Functions
+**************************************/
+
+/*! Deprecation warnings
+ *
+ *  Deprecated functions make the compiler generate a warning when invoked.
+ *  This is meant to invite users to update their source code.
+ *  Should deprecation warnings be a problem, it is generally possible to disable them,
+ *  typically with -Wno-deprecated-declarations for gcc
+ *  or _CRT_SECURE_NO_WARNINGS in Visual.
+ *
+ *  Another method is to define LZ4_DISABLE_DEPRECATE_WARNINGS
+ *  before including the header file.
+ */
+#ifdef LZ4_DISABLE_DEPRECATE_WARNINGS
+#  define LZ4_DEPRECATED(message)   /* disable deprecation warnings */
+#else
+#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
+#    define LZ4_DEPRECATED(message) [[deprecated(message)]]
+#  elif defined(_MSC_VER)
+#    define LZ4_DEPRECATED(message) __declspec(deprecated(message))
+#  elif defined(__clang__) || (defined(__GNUC__) && (__GNUC__ * 10 + __GNUC_MINOR__ >= 45))
+#    define LZ4_DEPRECATED(message) __attribute__((deprecated(message)))
+#  elif defined(__GNUC__) && (__GNUC__ * 10 + __GNUC_MINOR__ >= 31)
+#    define LZ4_DEPRECATED(message) __attribute__((deprecated))
+#  else
+#    pragma message("WARNING: LZ4_DEPRECATED needs custom implementation for this compiler")
+#    define LZ4_DEPRECATED(message)   /* disabled */
+#  endif
+#endif /* LZ4_DISABLE_DEPRECATE_WARNINGS */
+
+/*! Obsolete compression functions (since v1.7.3) */
+LZ4_DEPRECATED("use LZ4_compress_default() instead")       LZ4LIB_API int LZ4_compress               (const char* src, char* dest, int srcSize);
+LZ4_DEPRECATED("use LZ4_compress_default() instead")       LZ4LIB_API int LZ4_compress_limitedOutput (const char* src, char* dest, int srcSize, int maxOutputSize);
+LZ4_DEPRECATED("use LZ4_compress_fast_extState() instead") LZ4LIB_API int LZ4_compress_withState               (void* state, const char* source, char* dest, int inputSize);
+LZ4_DEPRECATED("use LZ4_compress_fast_extState() instead") LZ4LIB_API int LZ4_compress_limitedOutput_withState (void* state, const char* source, char* dest, int inputSize, int maxOutputSize);
+LZ4_DEPRECATED("use LZ4_compress_fast_continue() instead") LZ4LIB_API int LZ4_compress_continue                (LZ4_stream_t* LZ4_streamPtr, const char* source, char* dest, int inputSize);
+LZ4_DEPRECATED("use LZ4_compress_fast_continue() instead") LZ4LIB_API int LZ4_compress_limitedOutput_continue  (LZ4_stream_t* LZ4_streamPtr, const char* source, char* dest, int inputSize, int maxOutputSize);
+
+/*! Obsolete decompression functions (since v1.8.0) */
+LZ4_DEPRECATED("use LZ4_decompress_fast() instead") LZ4LIB_API int LZ4_uncompress (const char* source, char* dest, int outputSize);
+LZ4_DEPRECATED("use LZ4_decompress_safe() instead") LZ4LIB_API int LZ4_uncompress_unknownOutputSize (const char* source, char* dest, int isize, int maxOutputSize);
+
+/* Obsolete streaming functions (since v1.7.0)
+ * degraded functionality; do not use!
+ *
+ * In order to perform streaming compression, these functions depended on data
+ * that is no longer tracked in the state. They have been preserved as well as
+ * possible: using them will still produce a correct output. However, they don't
+ * actually retain any history between compression calls. The compression ratio
+ * achieved will therefore be no better than compressing each chunk
+ * independently.
+ */
+LZ4_DEPRECATED("Use LZ4_createStream() instead") LZ4LIB_API void* LZ4_create (char* inputBuffer);
+LZ4_DEPRECATED("Use LZ4_createStream() instead") LZ4LIB_API int   LZ4_sizeofStreamState(void);
+LZ4_DEPRECATED("Use LZ4_resetStream() instead")  LZ4LIB_API int   LZ4_resetStreamState(void* state, char* inputBuffer);
+LZ4_DEPRECATED("Use LZ4_saveDict() instead")     LZ4LIB_API char* LZ4_slideInputBuffer (void* state);
+
+/*! Obsolete streaming decoding functions (since v1.7.0) */
+LZ4_DEPRECATED("use LZ4_decompress_safe_usingDict() instead") LZ4LIB_API int LZ4_decompress_safe_withPrefix64k (const char* src, char* dst, int compressedSize, int maxDstSize);
+LZ4_DEPRECATED("use LZ4_decompress_fast_usingDict() instead") LZ4LIB_API int LZ4_decompress_fast_withPrefix64k (const char* src, char* dst, int originalSize);
+
+/*! Obsolete LZ4_decompress_fast variants (since v1.9.0) :
+ *  These functions used to be faster than LZ4_decompress_safe(),
+ *  but this is no longer the case. They are now slower.
+ *  This is because LZ4_decompress_fast() doesn't know the input size,
+ *  and therefore must progress more cautiously into the input buffer to not read beyond the end of block.
+ *  On top of that `LZ4_decompress_fast()` is not protected vs malformed or malicious inputs, making it a security liability.
+ *  As a consequence, LZ4_decompress_fast() is strongly discouraged, and deprecated.
+ *
+ *  The last remaining LZ4_decompress_fast() specificity is that
+ *  it can decompress a block without knowing its compressed size.
+ *  Such functionality can be achieved in a more secure manner
+ *  by employing LZ4_decompress_safe_partial().
+ *
+ *  Parameters:
+ *  originalSize : is the uncompressed size to regenerate.
+ *                 `dst` must be already allocated, its size must be >= 'originalSize' bytes.
+ * @return : number of bytes read from source buffer (== compressed size).
+ *           The function expects to finish at block's end exactly.
+ *           If the source stream is detected malformed, the function stops decoding and returns a negative result.
+ *  note : LZ4_decompress_fast*() requires originalSize. Thanks to this information, it never writes past the output buffer.
+ *         However, since it doesn't know its 'src' size, it may read an unknown amount of input, past input buffer bounds.
+ *         Also, since match offsets are not validated, match reads from 'src' may underflow too.
+ *         These issues never happen if input (compressed) data is correct.
+ *         But they may happen if input data is invalid (error or intentional tampering).
+ *         As a consequence, use these functions in trusted environments with trusted data **only**.
+ */
+LZ4_DEPRECATED("This function is deprecated and unsafe. Consider using LZ4_decompress_safe() instead")
+LZ4LIB_API int LZ4_decompress_fast (const char* src, char* dst, int originalSize);
+LZ4_DEPRECATED("This function is deprecated and unsafe. Consider using LZ4_decompress_safe_continue() instead")
+LZ4LIB_API int LZ4_decompress_fast_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* src, char* dst, int originalSize);
+LZ4_DEPRECATED("This function is deprecated and unsafe. Consider using LZ4_decompress_safe_usingDict() instead")
+LZ4LIB_API int LZ4_decompress_fast_usingDict (const char* src, char* dst, int originalSize, const char* dictStart, int dictSize);
+
+/*! LZ4_resetStream() :
+ *  An LZ4_stream_t structure must be initialized at least once.
+ *  This is done with LZ4_initStream(), or LZ4_resetStream().
+ *  Consider switching to LZ4_initStream(),
+ *  invoking LZ4_resetStream() will trigger deprecation warnings in the future.
+ */
+LZ4LIB_API void LZ4_resetStream (LZ4_stream_t* streamPtr);
+
+
+#endif /* LZ4_H_98237428734687 */
+
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/mfbt/lz4/lz4frame.c b/mfbt/lz4/lz4frame.c
new file mode 100644
index 0000000000..ec02c92f72
--- /dev/null
+++ b/mfbt/lz4/lz4frame.c
@@ -0,0 +1,1899 @@
+/*
+ * LZ4 auto-framing library
+ * Copyright (C) 2011-2016, Yann Collet.
+ *
+ * BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ * - Redistributions of source code must retain the above copyright
+ *   notice, this list of conditions and the following disclaimer.
+ * - Redistributions in binary form must reproduce the above
+ *   copyright notice, this list of conditions and the following disclaimer
+ *   in the documentation and/or other materials provided with the
+ *   distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * You can contact the author at :
+ * - LZ4 homepage : http://www.lz4.org
+ * - LZ4 source repository : https://github.com/lz4/lz4
+ */
+
+/* LZ4F is a stand-alone API to create LZ4-compressed Frames
+ * in full conformance with specification v1.6.1 .
+ * This library rely upon memory management capabilities (malloc, free)
+ * provided either by <stdlib.h>,
+ * or redirected towards another library of user's choice
+ * (see Memory Routines below).
+ */
+
+
+/*-************************************
+*  Compiler Options
+**************************************/
+#ifdef _MSC_VER    /* Visual Studio */
+#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
+#endif
+
+
+/*-************************************
+*  Tuning parameters
+**************************************/
+/*
+ * LZ4F_HEAPMODE :
+ * Select how default compression functions will allocate memory for their hash table,
+ * in memory stack (0:default, fastest), or in memory heap (1:requires malloc()).
+ */
+#ifndef LZ4F_HEAPMODE
+#  define LZ4F_HEAPMODE 0
+#endif
+
+
+/*-************************************
+*  Memory routines
+**************************************/
+/*
+ * User may redirect invocations of
+ * malloc(), calloc() and free()
+ * towards another library or solution of their choice
+ * by modifying below section.
+ */
+#ifndef LZ4_SRC_INCLUDED   /* avoid redefinition when sources are coalesced */
+#  include <stdlib.h>   /* malloc, calloc, free */
+#  define ALLOC(s)          malloc(s)
+#  define ALLOC_AND_ZERO(s) calloc(1,(s))
+#  define FREEMEM(p)        free(p)
+#endif
+
+#include <string.h>   /* memset, memcpy, memmove */
+#ifndef LZ4_SRC_INCLUDED  /* avoid redefinition when sources are coalesced */
+#  define MEM_INIT(p,v,s)   memset((p),(v),(s))
+#endif
+
+
+/*-************************************
+*  Library declarations
+**************************************/
+#define LZ4F_STATIC_LINKING_ONLY
+#include "lz4frame.h"
+#define LZ4_STATIC_LINKING_ONLY
+#include "lz4.h"
+#define LZ4_HC_STATIC_LINKING_ONLY
+#include "lz4hc.h"
+#define XXH_STATIC_LINKING_ONLY
+#include "xxhash.h"
+
+
+/*-************************************
+*  Debug
+**************************************/
+#if defined(LZ4_DEBUG) && (LZ4_DEBUG>=1)
+#  include <assert.h>
+#else
+#  ifndef assert
+#    define assert(condition) ((void)0)
+#  endif
+#endif
+
+#define LZ4F_STATIC_ASSERT(c)    { enum { LZ4F_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
+
+#if defined(LZ4_DEBUG) && (LZ4_DEBUG>=2) && !defined(DEBUGLOG)
+#  include <stdio.h>
+static int g_debuglog_enable = 1;
+#  define DEBUGLOG(l, ...) {                                  \
+                if ((g_debuglog_enable) && (l<=LZ4_DEBUG)) {  \
+                    fprintf(stderr, __FILE__ ": ");           \
+                    fprintf(stderr, __VA_ARGS__);             \
+                    fprintf(stderr, " \n");                   \
+            }   }
+#else
+#  define DEBUGLOG(l, ...)      {}    /* disabled */
+#endif
+
+
+/*-************************************
+*  Basic Types
+**************************************/
+#if !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef  uint8_t BYTE;
+  typedef uint16_t U16;
+  typedef uint32_t U32;
+  typedef  int32_t S32;
+  typedef uint64_t U64;
+#else
+  typedef unsigned char       BYTE;
+  typedef unsigned short      U16;
+  typedef unsigned int        U32;
+  typedef   signed int        S32;
+  typedef unsigned long long  U64;
+#endif
+
+
+/* unoptimized version; solves endianess & alignment issues */
+static U32 LZ4F_readLE32 (const void* src)
+{
+    const BYTE* const srcPtr = (const BYTE*)src;
+    U32 value32 = srcPtr[0];
+    value32 += ((U32)srcPtr[1])<< 8;
+    value32 += ((U32)srcPtr[2])<<16;
+    value32 += ((U32)srcPtr[3])<<24;
+    return value32;
+}
+
+static void LZ4F_writeLE32 (void* dst, U32 value32)
+{
+    BYTE* const dstPtr = (BYTE*)dst;
+    dstPtr[0] = (BYTE)value32;
+    dstPtr[1] = (BYTE)(value32 >> 8);
+    dstPtr[2] = (BYTE)(value32 >> 16);
+    dstPtr[3] = (BYTE)(value32 >> 24);
+}
+
+static U64 LZ4F_readLE64 (const void* src)
+{
+    const BYTE* const srcPtr = (const BYTE*)src;
+    U64 value64 = srcPtr[0];
+    value64 += ((U64)srcPtr[1]<<8);
+    value64 += ((U64)srcPtr[2]<<16);
+    value64 += ((U64)srcPtr[3]<<24);
+    value64 += ((U64)srcPtr[4]<<32);
+    value64 += ((U64)srcPtr[5]<<40);
+    value64 += ((U64)srcPtr[6]<<48);
+    value64 += ((U64)srcPtr[7]<<56);
+    return value64;
+}
+
+static void LZ4F_writeLE64 (void* dst, U64 value64)
+{
+    BYTE* const dstPtr = (BYTE*)dst;
+    dstPtr[0] = (BYTE)value64;
+    dstPtr[1] = (BYTE)(value64 >> 8);
+    dstPtr[2] = (BYTE)(value64 >> 16);
+    dstPtr[3] = (BYTE)(value64 >> 24);
+    dstPtr[4] = (BYTE)(value64 >> 32);
+    dstPtr[5] = (BYTE)(value64 >> 40);
+    dstPtr[6] = (BYTE)(value64 >> 48);
+    dstPtr[7] = (BYTE)(value64 >> 56);
+}
+
+
+/*-************************************
+*  Constants
+**************************************/
+#ifndef LZ4_SRC_INCLUDED   /* avoid double definition */
+#  define KB *(1<<10)
+#  define MB *(1<<20)
+#  define GB *(1<<30)
+#endif
+
+#define _1BIT  0x01
+#define _2BITS 0x03
+#define _3BITS 0x07
+#define _4BITS 0x0F
+#define _8BITS 0xFF
+
+#define LZ4F_MAGIC_SKIPPABLE_START 0x184D2A50U
+#define LZ4F_MAGICNUMBER 0x184D2204U
+#define LZ4F_BLOCKUNCOMPRESSED_FLAG 0x80000000U
+#define LZ4F_BLOCKSIZEID_DEFAULT LZ4F_max64KB
+
+static const size_t minFHSize = LZ4F_HEADER_SIZE_MIN;   /*  7 */
+static const size_t maxFHSize = LZ4F_HEADER_SIZE_MAX;   /* 19 */
+static const size_t BHSize = LZ4F_BLOCK_HEADER_SIZE;  /* block header : size, and compress flag */
+static const size_t BFSize = LZ4F_BLOCK_CHECKSUM_SIZE;  /* block footer : checksum (optional) */
+
+
+/*-************************************
+*  Structures and local types
+**************************************/
+typedef struct LZ4F_cctx_s
+{
+    LZ4F_preferences_t prefs;
+    U32    version;
+    U32    cStage;
+    const LZ4F_CDict* cdict;
+    size_t maxBlockSize;
+    size_t maxBufferSize;
+    BYTE*  tmpBuff;
+    BYTE*  tmpIn;
+    size_t tmpInSize;
+    U64    totalInSize;
+    XXH32_state_t xxh;
+    void*  lz4CtxPtr;
+    U16    lz4CtxAlloc; /* sized for: 0 = none, 1 = lz4 ctx, 2 = lz4hc ctx */
+    U16    lz4CtxState; /* in use as: 0 = none, 1 = lz4 ctx, 2 = lz4hc ctx */
+} LZ4F_cctx_t;
+
+
+/*-************************************
+*  Error management
+**************************************/
+#define LZ4F_GENERATE_STRING(STRING) #STRING,
+static const char* LZ4F_errorStrings[] = { LZ4F_LIST_ERRORS(LZ4F_GENERATE_STRING) };
+
+
+unsigned LZ4F_isError(LZ4F_errorCode_t code)
+{
+    return (code > (LZ4F_errorCode_t)(-LZ4F_ERROR_maxCode));
+}
+
+const char* LZ4F_getErrorName(LZ4F_errorCode_t code)
+{
+    static const char* codeError = "Unspecified error code";
+    if (LZ4F_isError(code)) return LZ4F_errorStrings[-(int)(code)];
+    return codeError;
+}
+
+LZ4F_errorCodes LZ4F_getErrorCode(size_t functionResult)
+{
+    if (!LZ4F_isError(functionResult)) return LZ4F_OK_NoError;
+    return (LZ4F_errorCodes)(-(ptrdiff_t)functionResult);
+}
+
+static LZ4F_errorCode_t err0r(LZ4F_errorCodes code)
+{
+    /* A compilation error here means sizeof(ptrdiff_t) is not large enough */
+    LZ4F_STATIC_ASSERT(sizeof(ptrdiff_t) >= sizeof(size_t));
+    return (LZ4F_errorCode_t)-(ptrdiff_t)code;
+}
+
+unsigned LZ4F_getVersion(void) { return LZ4F_VERSION; }
+
+int LZ4F_compressionLevel_max(void) { return LZ4HC_CLEVEL_MAX; }
+
+size_t LZ4F_getBlockSize(unsigned blockSizeID)
+{
+    static const size_t blockSizes[4] = { 64 KB, 256 KB, 1 MB, 4 MB };
+
+    if (blockSizeID == 0) blockSizeID = LZ4F_BLOCKSIZEID_DEFAULT;
+    if (blockSizeID < LZ4F_max64KB || blockSizeID > LZ4F_max4MB)
+        return err0r(LZ4F_ERROR_maxBlockSize_invalid);
+    blockSizeID -= LZ4F_max64KB;
+    return blockSizes[blockSizeID];
+}
+
+/*-************************************
+*  Private functions
+**************************************/
+#define MIN(a,b)   ( (a) < (b) ? (a) : (b) )
+
+static BYTE LZ4F_headerChecksum (const void* header, size_t length)
+{
+    U32 const xxh = XXH32(header, length, 0);
+    return (BYTE)(xxh >> 8);
+}
+
+
+/*-************************************
+*  Simple-pass compression functions
+**************************************/
+static LZ4F_blockSizeID_t LZ4F_optimalBSID(const LZ4F_blockSizeID_t requestedBSID,
+                                           const size_t srcSize)
+{
+    LZ4F_blockSizeID_t proposedBSID = LZ4F_max64KB;
+    size_t maxBlockSize = 64 KB;
+    while (requestedBSID > proposedBSID) {
+        if (srcSize <= maxBlockSize)
+            return proposedBSID;
+        proposedBSID = (LZ4F_blockSizeID_t)((int)proposedBSID + 1);
+        maxBlockSize <<= 2;
+    }
+    return requestedBSID;
+}
+
+/*! LZ4F_compressBound_internal() :
+ *  Provides dstCapacity given a srcSize to guarantee operation success in worst case situations.
+ *  prefsPtr is optional : if NULL is provided, preferences will be set to cover worst case scenario.
+ * @return is always the same for a srcSize and prefsPtr, so it can be relied upon to size reusable buffers.
+ *  When srcSize==0, LZ4F_compressBound() provides an upper bound for LZ4F_flush() and LZ4F_compressEnd() operations.
+ */
+static size_t LZ4F_compressBound_internal(size_t srcSize,
+                                    const LZ4F_preferences_t* preferencesPtr,
+                                          size_t alreadyBuffered)
+{
+    LZ4F_preferences_t prefsNull = LZ4F_INIT_PREFERENCES;
+    prefsNull.frameInfo.contentChecksumFlag = LZ4F_contentChecksumEnabled;   /* worst case */
+    prefsNull.frameInfo.blockChecksumFlag = LZ4F_blockChecksumEnabled;   /* worst case */
+    {   const LZ4F_preferences_t* const prefsPtr = (preferencesPtr==NULL) ? &prefsNull : preferencesPtr;
+        U32 const flush = prefsPtr->autoFlush | (srcSize==0);
+        LZ4F_blockSizeID_t const blockID = prefsPtr->frameInfo.blockSizeID;
+        size_t const blockSize = LZ4F_getBlockSize(blockID);
+        size_t const maxBuffered = blockSize - 1;
+        size_t const bufferedSize = MIN(alreadyBuffered, maxBuffered);
+        size_t const maxSrcSize = srcSize + bufferedSize;
+        unsigned const nbFullBlocks = (unsigned)(maxSrcSize / blockSize);
+        size_t const partialBlockSize = maxSrcSize & (blockSize-1);
+        size_t const lastBlockSize = flush ? partialBlockSize : 0;
+        unsigned const nbBlocks = nbFullBlocks + (lastBlockSize>0);
+
+        size_t const blockCRCSize = BFSize * prefsPtr->frameInfo.blockChecksumFlag;
+        size_t const frameEnd = BHSize + (prefsPtr->frameInfo.contentChecksumFlag*BFSize);
+
+        return ((BHSize + blockCRCSize) * nbBlocks) +
+               (blockSize * nbFullBlocks) + lastBlockSize + frameEnd;
+    }
+}
+
+size_t LZ4F_compressFrameBound(size_t srcSize, const LZ4F_preferences_t* preferencesPtr)
+{
+    LZ4F_preferences_t prefs;
+    size_t const headerSize = maxFHSize;      /* max header size, including optional fields */
+
+    if (preferencesPtr!=NULL) prefs = *preferencesPtr;
+    else MEM_INIT(&prefs, 0, sizeof(prefs));
+    prefs.autoFlush = 1;
+
+    return headerSize + LZ4F_compressBound_internal(srcSize, &prefs, 0);;
+}
+
+
+/*! LZ4F_compressFrame_usingCDict() :
+ *  Compress srcBuffer using a dictionary, in a single step.
+ *  cdict can be NULL, in which case, no dictionary is used.
+ *  dstBuffer MUST be >= LZ4F_compressFrameBound(srcSize, preferencesPtr).
+ *  The LZ4F_preferences_t structure is optional : you may provide NULL as argument,
+ *  however, it's the only way to provide a dictID, so it's not recommended.
+ * @return : number of bytes written into dstBuffer,
+ *           or an error code if it fails (can be tested using LZ4F_isError())
+ */
+size_t LZ4F_compressFrame_usingCDict(LZ4F_cctx* cctx,
+                                     void* dstBuffer, size_t dstCapacity,
+                               const void* srcBuffer, size_t srcSize,
+                               const LZ4F_CDict* cdict,
+                               const LZ4F_preferences_t* preferencesPtr)
+{
+    LZ4F_preferences_t prefs;
+    LZ4F_compressOptions_t options;
+    BYTE* const dstStart = (BYTE*) dstBuffer;
+    BYTE* dstPtr = dstStart;
+    BYTE* const dstEnd = dstStart + dstCapacity;
+
+    if (preferencesPtr!=NULL)
+        prefs = *preferencesPtr;
+    else
+        MEM_INIT(&prefs, 0, sizeof(prefs));
+    if (prefs.frameInfo.contentSize != 0)
+        prefs.frameInfo.contentSize = (U64)srcSize;   /* auto-correct content size if selected (!=0) */
+
+    prefs.frameInfo.blockSizeID = LZ4F_optimalBSID(prefs.frameInfo.blockSizeID, srcSize);
+    prefs.autoFlush = 1;
+    if (srcSize <= LZ4F_getBlockSize(prefs.frameInfo.blockSizeID))
+        prefs.frameInfo.blockMode = LZ4F_blockIndependent;   /* only one block => no need for inter-block link */
+
+    MEM_INIT(&options, 0, sizeof(options));
+    options.stableSrc = 1;
+
+    if (dstCapacity < LZ4F_compressFrameBound(srcSize, &prefs))  /* condition to guarantee success */
+        return err0r(LZ4F_ERROR_dstMaxSize_tooSmall);
+
+    { size_t const headerSize = LZ4F_compressBegin_usingCDict(cctx, dstBuffer, dstCapacity, cdict, &prefs);  /* write header */
+      if (LZ4F_isError(headerSize)) return headerSize;
+      dstPtr += headerSize;   /* header size */ }
+
+    assert(dstEnd >= dstPtr);
+    { size_t const cSize = LZ4F_compressUpdate(cctx, dstPtr, (size_t)(dstEnd-dstPtr), srcBuffer, srcSize, &options);
+      if (LZ4F_isError(cSize)) return cSize;
+      dstPtr += cSize; }
+
+    assert(dstEnd >= dstPtr);
+    { size_t const tailSize = LZ4F_compressEnd(cctx, dstPtr, (size_t)(dstEnd-dstPtr), &options);   /* flush last block, and generate suffix */
+      if (LZ4F_isError(tailSize)) return tailSize;
+      dstPtr += tailSize; }
+
+    assert(dstEnd >= dstStart);
+    return (size_t)(dstPtr - dstStart);
+}
+
+
+/*! LZ4F_compressFrame() :
+ *  Compress an entire srcBuffer into a valid LZ4 frame, in a single step.
+ *  dstBuffer MUST be >= LZ4F_compressFrameBound(srcSize, preferencesPtr).
+ *  The LZ4F_preferences_t structure is optional : you can provide NULL as argument. All preferences will be set to default.
+ * @return : number of bytes written into dstBuffer.
+ *           or an error code if it fails (can be tested using LZ4F_isError())
+ */
+size_t LZ4F_compressFrame(void* dstBuffer, size_t dstCapacity,
+                    const void* srcBuffer, size_t srcSize,
+                    const LZ4F_preferences_t* preferencesPtr)
+{
+    size_t result;
+#if (LZ4F_HEAPMODE)
+    LZ4F_cctx_t *cctxPtr;
+    result = LZ4F_createCompressionContext(&cctxPtr, LZ4F_VERSION);
+    if (LZ4F_isError(result)) return result;
+#else
+    LZ4F_cctx_t cctx;
+    LZ4_stream_t lz4ctx;
+    LZ4F_cctx_t *cctxPtr = &cctx;
+
+    DEBUGLOG(4, "LZ4F_compressFrame");
+    MEM_INIT(&cctx, 0, sizeof(cctx));
+    cctx.version = LZ4F_VERSION;
+    cctx.maxBufferSize = 5 MB;   /* mess with real buffer size to prevent dynamic allocation; works only because autoflush==1 & stableSrc==1 */
+    if (preferencesPtr == NULL ||
+        preferencesPtr->compressionLevel < LZ4HC_CLEVEL_MIN)
+    {
+        LZ4_initStream(&lz4ctx, sizeof(lz4ctx));
+        cctxPtr->lz4CtxPtr = &lz4ctx;
+        cctxPtr->lz4CtxAlloc = 1;
+        cctxPtr->lz4CtxState = 1;
+    }
+#endif
+
+    result = LZ4F_compressFrame_usingCDict(cctxPtr, dstBuffer, dstCapacity,
+                                           srcBuffer, srcSize,
+                                           NULL, preferencesPtr);
+
+#if (LZ4F_HEAPMODE)
+    LZ4F_freeCompressionContext(cctxPtr);
+#else
+    if (preferencesPtr != NULL &&
+        preferencesPtr->compressionLevel >= LZ4HC_CLEVEL_MIN)
+    {
+        FREEMEM(cctxPtr->lz4CtxPtr);
+    }
+#endif
+    return result;
+}
+
+
+/*-***************************************************
+*   Dictionary compression
+*****************************************************/
+
+struct LZ4F_CDict_s {
+    void* dictContent;
+    LZ4_stream_t* fastCtx;
+    LZ4_streamHC_t* HCCtx;
+}; /* typedef'd to LZ4F_CDict within lz4frame_static.h */
+
+/*! LZ4F_createCDict() :
+ *  When compressing multiple messages / blocks with the same dictionary, it's recommended to load it just once.
+ *  LZ4F_createCDict() will create a digested dictionary, ready to start future compression operations without startup delay.
+ *  LZ4F_CDict can be created once and shared by multiple threads concurrently, since its usage is read-only.
+ * `dictBuffer` can be released after LZ4F_CDict creation, since its content is copied within CDict
+ * @return : digested dictionary for compression, or NULL if failed */
+LZ4F_CDict* LZ4F_createCDict(const void* dictBuffer, size_t dictSize)
+{
+    const char* dictStart = (const char*)dictBuffer;
+    LZ4F_CDict* cdict = (LZ4F_CDict*) ALLOC(sizeof(*cdict));
+    DEBUGLOG(4, "LZ4F_createCDict");
+    if (!cdict) return NULL;
+    if (dictSize > 64 KB) {
+        dictStart += dictSize - 64 KB;
+        dictSize = 64 KB;
+    }
+    cdict->dictContent = ALLOC(dictSize);
+    cdict->fastCtx = LZ4_createStream();
+    cdict->HCCtx = LZ4_createStreamHC();
+    if (!cdict->dictContent || !cdict->fastCtx || !cdict->HCCtx) {
+        LZ4F_freeCDict(cdict);
+        return NULL;
+    }
+    memcpy(cdict->dictContent, dictStart, dictSize);
+    LZ4_loadDict (cdict->fastCtx, (const char*)cdict->dictContent, (int)dictSize);
+    LZ4_setCompressionLevel(cdict->HCCtx, LZ4HC_CLEVEL_DEFAULT);
+    LZ4_loadDictHC(cdict->HCCtx, (const char*)cdict->dictContent, (int)dictSize);
+    return cdict;
+}
+
+void LZ4F_freeCDict(LZ4F_CDict* cdict)
+{
+    if (cdict==NULL) return;  /* support free on NULL */
+    FREEMEM(cdict->dictContent);
+    LZ4_freeStream(cdict->fastCtx);
+    LZ4_freeStreamHC(cdict->HCCtx);
+    FREEMEM(cdict);
+}
+
+
+/*-*********************************
+*  Advanced compression functions
+***********************************/
+
+/*! LZ4F_createCompressionContext() :
+ *  The first thing to do is to create a compressionContext object, which will be used in all compression operations.
+ *  This is achieved using LZ4F_createCompressionContext(), which takes as argument a version and an LZ4F_preferences_t structure.
+ *  The version provided MUST be LZ4F_VERSION. It is intended to track potential incompatible differences between different binaries.
+ *  The function will provide a pointer to an allocated LZ4F_compressionContext_t object.
+ *  If the result LZ4F_errorCode_t is not OK_NoError, there was an error during context creation.
+ *  Object can release its memory using LZ4F_freeCompressionContext();
+ */
+LZ4F_errorCode_t LZ4F_createCompressionContext(LZ4F_cctx** LZ4F_compressionContextPtr, unsigned version)
+{
+    LZ4F_cctx_t* const cctxPtr = (LZ4F_cctx_t*)ALLOC_AND_ZERO(sizeof(LZ4F_cctx_t));
+    if (cctxPtr==NULL) return err0r(LZ4F_ERROR_allocation_failed);
+
+    cctxPtr->version = version;
+    cctxPtr->cStage = 0;   /* Next stage : init stream */
+
+    *LZ4F_compressionContextPtr = cctxPtr;
+
+    return LZ4F_OK_NoError;
+}
+
+
+LZ4F_errorCode_t LZ4F_freeCompressionContext(LZ4F_cctx* cctxPtr)
+{
+    if (cctxPtr != NULL) {  /* support free on NULL */
+       FREEMEM(cctxPtr->lz4CtxPtr);  /* note: LZ4_streamHC_t and LZ4_stream_t are simple POD types */
+       FREEMEM(cctxPtr->tmpBuff);
+       FREEMEM(cctxPtr);
+    }
+
+    return LZ4F_OK_NoError;
+}
+
+
+/**
+ * This function prepares the internal LZ4(HC) stream for a new compression,
+ * resetting the context and attaching the dictionary, if there is one.
+ *
+ * It needs to be called at the beginning of each independent compression
+ * stream (i.e., at the beginning of a frame in blockLinked mode, or at the
+ * beginning of each block in blockIndependent mode).
+ */
+static void LZ4F_initStream(void* ctx,
+                            const LZ4F_CDict* cdict,
+                            int level,
+                            LZ4F_blockMode_t blockMode) {
+    if (level < LZ4HC_CLEVEL_MIN) {
+        if (cdict != NULL || blockMode == LZ4F_blockLinked) {
+            /* In these cases, we will call LZ4_compress_fast_continue(),
+             * which needs an already reset context. Otherwise, we'll call a
+             * one-shot API. The non-continued APIs internally perform their own
+             * resets at the beginning of their calls, where they know what
+             * tableType they need the context to be in. So in that case this
+             * would be misguided / wasted work. */
+            LZ4_resetStream_fast((LZ4_stream_t*)ctx);
+        }
+        LZ4_attach_dictionary((LZ4_stream_t *)ctx, cdict ? cdict->fastCtx : NULL);
+    } else {
+        LZ4_resetStreamHC_fast((LZ4_streamHC_t*)ctx, level);
+        LZ4_attach_HC_dictionary((LZ4_streamHC_t *)ctx, cdict ? cdict->HCCtx : NULL);
+    }
+}
+
+
+/*! LZ4F_compressBegin_usingCDict() :
+ *  init streaming compression and writes frame header into dstBuffer.
+ *  dstBuffer must be >= LZ4F_HEADER_SIZE_MAX bytes.
+ * @return : number of bytes written into dstBuffer for the header
+ *           or an error code (can be tested using LZ4F_isError())
+ */
+size_t LZ4F_compressBegin_usingCDict(LZ4F_cctx* cctxPtr,
+                          void* dstBuffer, size_t dstCapacity,
+                          const LZ4F_CDict* cdict,
+                          const LZ4F_preferences_t* preferencesPtr)
+{
+    LZ4F_preferences_t prefNull;
+    BYTE* const dstStart = (BYTE*)dstBuffer;
+    BYTE* dstPtr = dstStart;
+    BYTE* headerStart;
+
+    if (dstCapacity < maxFHSize) return err0r(LZ4F_ERROR_dstMaxSize_tooSmall);
+    MEM_INIT(&prefNull, 0, sizeof(prefNull));
+    if (preferencesPtr == NULL) preferencesPtr = &prefNull;
+    cctxPtr->prefs = *preferencesPtr;
+
+    /* Ctx Management */
+    {   U16 const ctxTypeID = (cctxPtr->prefs.compressionLevel < LZ4HC_CLEVEL_MIN) ? 1 : 2;
+        if (cctxPtr->lz4CtxAlloc < ctxTypeID) {
+            FREEMEM(cctxPtr->lz4CtxPtr);
+            if (cctxPtr->prefs.compressionLevel < LZ4HC_CLEVEL_MIN) {
+                cctxPtr->lz4CtxPtr = LZ4_createStream();
+            } else {
+                cctxPtr->lz4CtxPtr = LZ4_createStreamHC();
+            }
+            if (cctxPtr->lz4CtxPtr == NULL)
+                return err0r(LZ4F_ERROR_allocation_failed);
+            cctxPtr->lz4CtxAlloc = ctxTypeID;
+            cctxPtr->lz4CtxState = ctxTypeID;
+        } else if (cctxPtr->lz4CtxState != ctxTypeID) {
+            /* otherwise, a sufficient buffer is allocated, but we need to
+             * reset it to the correct context type */
+            if (cctxPtr->prefs.compressionLevel < LZ4HC_CLEVEL_MIN) {
+                LZ4_initStream((LZ4_stream_t *) cctxPtr->lz4CtxPtr, sizeof (LZ4_stream_t));
+            } else {
+                LZ4_initStreamHC((LZ4_streamHC_t *) cctxPtr->lz4CtxPtr, sizeof(LZ4_streamHC_t));
+                LZ4_setCompressionLevel((LZ4_streamHC_t *) cctxPtr->lz4CtxPtr, cctxPtr->prefs.compressionLevel);
+            }
+            cctxPtr->lz4CtxState = ctxTypeID;
+        }
+    }
+
+    /* Buffer Management */
+    if (cctxPtr->prefs.frameInfo.blockSizeID == 0)
+        cctxPtr->prefs.frameInfo.blockSizeID = LZ4F_BLOCKSIZEID_DEFAULT;
+    cctxPtr->maxBlockSize = LZ4F_getBlockSize(cctxPtr->prefs.frameInfo.blockSizeID);
+
+    {   size_t const requiredBuffSize = preferencesPtr->autoFlush ?
+                ((cctxPtr->prefs.frameInfo.blockMode == LZ4F_blockLinked) ? 64 KB : 0) :  /* only needs past data up to window size */
+                cctxPtr->maxBlockSize + ((cctxPtr->prefs.frameInfo.blockMode == LZ4F_blockLinked) ? 128 KB : 0);
+
+        if (cctxPtr->maxBufferSize < requiredBuffSize) {
+            cctxPtr->maxBufferSize = 0;
+            FREEMEM(cctxPtr->tmpBuff);
+            cctxPtr->tmpBuff = (BYTE*)ALLOC_AND_ZERO(requiredBuffSize);
+            if (cctxPtr->tmpBuff == NULL) return err0r(LZ4F_ERROR_allocation_failed);
+            cctxPtr->maxBufferSize = requiredBuffSize;
+    }   }
+    cctxPtr->tmpIn = cctxPtr->tmpBuff;
+    cctxPtr->tmpInSize = 0;
+    (void)XXH32_reset(&(cctxPtr->xxh), 0);
+
+    /* context init */
+    cctxPtr->cdict = cdict;
+    if (cctxPtr->prefs.frameInfo.blockMode == LZ4F_blockLinked) {
+        /* frame init only for blockLinked : blockIndependent will be init at each block */
+        LZ4F_initStream(cctxPtr->lz4CtxPtr, cdict, cctxPtr->prefs.compressionLevel, LZ4F_blockLinked);
+    }
+    if (preferencesPtr->compressionLevel >= LZ4HC_CLEVEL_MIN) {
+        LZ4_favorDecompressionSpeed((LZ4_streamHC_t*)cctxPtr->lz4CtxPtr, (int)preferencesPtr->favorDecSpeed);
+    }
+
+    /* Magic Number */
+    LZ4F_writeLE32(dstPtr, LZ4F_MAGICNUMBER);
+    dstPtr += 4;
+    headerStart = dstPtr;
+
+    /* FLG Byte */
+    *dstPtr++ = (BYTE)(((1 & _2BITS) << 6)    /* Version('01') */
+        + ((cctxPtr->prefs.frameInfo.blockMode & _1BIT ) << 5)
+        + ((cctxPtr->prefs.frameInfo.blockChecksumFlag & _1BIT ) << 4)
+        + ((unsigned)(cctxPtr->prefs.frameInfo.contentSize > 0) << 3)
+        + ((cctxPtr->prefs.frameInfo.contentChecksumFlag & _1BIT ) << 2)
+        +  (cctxPtr->prefs.frameInfo.dictID > 0) );
+    /* BD Byte */
+    *dstPtr++ = (BYTE)((cctxPtr->prefs.frameInfo.blockSizeID & _3BITS) << 4);
+    /* Optional Frame content size field */
+    if (cctxPtr->prefs.frameInfo.contentSize) {
+        LZ4F_writeLE64(dstPtr, cctxPtr->prefs.frameInfo.contentSize);
+        dstPtr += 8;
+        cctxPtr->totalInSize = 0;
+    }
+    /* Optional dictionary ID field */
+    if (cctxPtr->prefs.frameInfo.dictID) {
+        LZ4F_writeLE32(dstPtr, cctxPtr->prefs.frameInfo.dictID);
+        dstPtr += 4;
+    }
+    /* Header CRC Byte */
+    *dstPtr = LZ4F_headerChecksum(headerStart, (size_t)(dstPtr - headerStart));
+    dstPtr++;
+
+    cctxPtr->cStage = 1;   /* header written, now request input data block */
+    return (size_t)(dstPtr - dstStart);
+}
+
+
+/*! LZ4F_compressBegin() :
+ *  init streaming compression and writes frame header into dstBuffer.
+ *  dstBuffer must be >= LZ4F_HEADER_SIZE_MAX bytes.
+ *  preferencesPtr can be NULL, in which case default parameters are selected.
+ * @return : number of bytes written into dstBuffer for the header
+ *        or an error code (can be tested using LZ4F_isError())
+ */
+size_t LZ4F_compressBegin(LZ4F_cctx* cctxPtr,
+                          void* dstBuffer, size_t dstCapacity,
+                          const LZ4F_preferences_t* preferencesPtr)
+{
+    return LZ4F_compressBegin_usingCDict(cctxPtr, dstBuffer, dstCapacity,
+                                         NULL, preferencesPtr);
+}
+
+
+/*  LZ4F_compressBound() :
+ * @return minimum capacity of dstBuffer for a given srcSize to handle worst case scenario.
+ *  LZ4F_preferences_t structure is optional : if NULL, preferences will be set to cover worst case scenario.
+ *  This function cannot fail.
+ */
+size_t LZ4F_compressBound(size_t srcSize, const LZ4F_preferences_t* preferencesPtr)
+{
+    if (preferencesPtr && preferencesPtr->autoFlush) {
+        return LZ4F_compressBound_internal(srcSize, preferencesPtr, 0);
+    }
+    return LZ4F_compressBound_internal(srcSize, preferencesPtr, (size_t)-1);
+}
+
+
+typedef int (*compressFunc_t)(void* ctx, const char* src, char* dst, int srcSize, int dstSize, int level, const LZ4F_CDict* cdict);
+
+
+/*! LZ4F_makeBlock():
+ *  compress a single block, add header and optional checksum.
+ *  assumption : dst buffer capacity is >= BHSize + srcSize + crcSize
+ */
+static size_t LZ4F_makeBlock(void* dst,
+                       const void* src, size_t srcSize,
+                             compressFunc_t compress, void* lz4ctx, int level,
+                       const LZ4F_CDict* cdict,
+                             LZ4F_blockChecksum_t crcFlag)
+{
+    BYTE* const cSizePtr = (BYTE*)dst;
+    U32 cSize = (U32)compress(lz4ctx, (const char*)src, (char*)(cSizePtr+BHSize),
+                                      (int)(srcSize), (int)(srcSize-1),
+                                      level, cdict);
+    if (cSize == 0) {  /* compression failed */
+        DEBUGLOG(5, "LZ4F_makeBlock: compression failed, creating a raw block (size %u)", (U32)srcSize);
+        cSize = (U32)srcSize;
+        LZ4F_writeLE32(cSizePtr, cSize | LZ4F_BLOCKUNCOMPRESSED_FLAG);
+        memcpy(cSizePtr+BHSize, src, srcSize);
+    } else {
+        LZ4F_writeLE32(cSizePtr, cSize);
+    }
+    if (crcFlag) {
+        U32 const crc32 = XXH32(cSizePtr+BHSize, cSize, 0);  /* checksum of compressed data */
+        LZ4F_writeLE32(cSizePtr+BHSize+cSize, crc32);
+    }
+    return BHSize + cSize + ((U32)crcFlag)*BFSize;
+}
+
+
+static int LZ4F_compressBlock(void* ctx, const char* src, char* dst, int srcSize, int dstCapacity, int level, const LZ4F_CDict* cdict)
+{
+    int const acceleration = (level < 0) ? -level + 1 : 1;
+    LZ4F_initStream(ctx, cdict, level, LZ4F_blockIndependent);
+    if (cdict) {
+        return LZ4_compress_fast_continue((LZ4_stream_t*)ctx, src, dst, srcSize, dstCapacity, acceleration);
+    } else {
+        return LZ4_compress_fast_extState_fastReset(ctx, src, dst, srcSize, dstCapacity, acceleration);
+    }
+}
+
+static int LZ4F_compressBlock_continue(void* ctx, const char* src, char* dst, int srcSize, int dstCapacity, int level, const LZ4F_CDict* cdict)
+{
+    int const acceleration = (level < 0) ? -level + 1 : 1;
+    (void)cdict; /* init once at beginning of frame */
+    return LZ4_compress_fast_continue((LZ4_stream_t*)ctx, src, dst, srcSize, dstCapacity, acceleration);
+}
+
+static int LZ4F_compressBlockHC(void* ctx, const char* src, char* dst, int srcSize, int dstCapacity, int level, const LZ4F_CDict* cdict)
+{
+    LZ4F_initStream(ctx, cdict, level, LZ4F_blockIndependent);
+    if (cdict) {
+        return LZ4_compress_HC_continue((LZ4_streamHC_t*)ctx, src, dst, srcSize, dstCapacity);
+    }
+    return LZ4_compress_HC_extStateHC_fastReset(ctx, src, dst, srcSize, dstCapacity, level);
+}
+
+static int LZ4F_compressBlockHC_continue(void* ctx, const char* src, char* dst, int srcSize, int dstCapacity, int level, const LZ4F_CDict* cdict)
+{
+    (void)level; (void)cdict; /* init once at beginning of frame */
+    return LZ4_compress_HC_continue((LZ4_streamHC_t*)ctx, src, dst, srcSize, dstCapacity);
+}
+
+static compressFunc_t LZ4F_selectCompression(LZ4F_blockMode_t blockMode, int level)
+{
+    if (level < LZ4HC_CLEVEL_MIN) {
+        if (blockMode == LZ4F_blockIndependent) return LZ4F_compressBlock;
+        return LZ4F_compressBlock_continue;
+    }
+    if (blockMode == LZ4F_blockIndependent) return LZ4F_compressBlockHC;
+    return LZ4F_compressBlockHC_continue;
+}
+
+static int LZ4F_localSaveDict(LZ4F_cctx_t* cctxPtr)
+{
+    if (cctxPtr->prefs.compressionLevel < LZ4HC_CLEVEL_MIN)
+        return LZ4_saveDict ((LZ4_stream_t*)(cctxPtr->lz4CtxPtr), (char*)(cctxPtr->tmpBuff), 64 KB);
+    return LZ4_saveDictHC ((LZ4_streamHC_t*)(cctxPtr->lz4CtxPtr), (char*)(cctxPtr->tmpBuff), 64 KB);
+}
+
+typedef enum { notDone, fromTmpBuffer, fromSrcBuffer } LZ4F_lastBlockStatus;
+
+/*! LZ4F_compressUpdate() :
+ *  LZ4F_compressUpdate() can be called repetitively to compress as much data as necessary.
+ *  dstBuffer MUST be >= LZ4F_compressBound(srcSize, preferencesPtr).
+ *  LZ4F_compressOptions_t structure is optional : you can provide NULL as argument.
+ * @return : the number of bytes written into dstBuffer. It can be zero, meaning input data was just buffered.
+ *           or an error code if it fails (which can be tested using LZ4F_isError())
+ */
+size_t LZ4F_compressUpdate(LZ4F_cctx* cctxPtr,
+                           void* dstBuffer, size_t dstCapacity,
+                     const void* srcBuffer, size_t srcSize,
+                     const LZ4F_compressOptions_t* compressOptionsPtr)
+{
+    LZ4F_compressOptions_t cOptionsNull;
+    size_t const blockSize = cctxPtr->maxBlockSize;
+    const BYTE* srcPtr = (const BYTE*)srcBuffer;
+    const BYTE* const srcEnd = srcPtr + srcSize;
+    BYTE* const dstStart = (BYTE*)dstBuffer;
+    BYTE* dstPtr = dstStart;
+    LZ4F_lastBlockStatus lastBlockCompressed = notDone;
+    compressFunc_t const compress = LZ4F_selectCompression(cctxPtr->prefs.frameInfo.blockMode, cctxPtr->prefs.compressionLevel);
+
+    DEBUGLOG(4, "LZ4F_compressUpdate (srcSize=%zu)", srcSize);
+
+    if (cctxPtr->cStage != 1) return err0r(LZ4F_ERROR_GENERIC);
+    if (dstCapacity < LZ4F_compressBound_internal(srcSize, &(cctxPtr->prefs), cctxPtr->tmpInSize))
+        return err0r(LZ4F_ERROR_dstMaxSize_tooSmall);
+    MEM_INIT(&cOptionsNull, 0, sizeof(cOptionsNull));
+    if (compressOptionsPtr == NULL) compressOptionsPtr = &cOptionsNull;
+
+    /* complete tmp buffer */
+    if (cctxPtr->tmpInSize > 0) {   /* some data already within tmp buffer */
+        size_t const sizeToCopy = blockSize - cctxPtr->tmpInSize;
+        if (sizeToCopy > srcSize) {
+            /* add src to tmpIn buffer */
+            memcpy(cctxPtr->tmpIn + cctxPtr->tmpInSize, srcBuffer, srcSize);
+            srcPtr = srcEnd;
+            cctxPtr->tmpInSize += srcSize;
+            /* still needs some CRC */
+        } else {
+            /* complete tmpIn block and then compress it */
+            lastBlockCompressed = fromTmpBuffer;
+            memcpy(cctxPtr->tmpIn + cctxPtr->tmpInSize, srcBuffer, sizeToCopy);
+            srcPtr += sizeToCopy;
+
+            dstPtr += LZ4F_makeBlock(dstPtr,
+                                     cctxPtr->tmpIn, blockSize,
+                                     compress, cctxPtr->lz4CtxPtr, cctxPtr->prefs.compressionLevel,
+                                     cctxPtr->cdict,
+                                     cctxPtr->prefs.frameInfo.blockChecksumFlag);
+
+            if (cctxPtr->prefs.frameInfo.blockMode==LZ4F_blockLinked) cctxPtr->tmpIn += blockSize;
+            cctxPtr->tmpInSize = 0;
+        }
+    }
+
+    while ((size_t)(srcEnd - srcPtr) >= blockSize) {
+        /* compress full blocks */
+        lastBlockCompressed = fromSrcBuffer;
+        dstPtr += LZ4F_makeBlock(dstPtr,
+                                 srcPtr, blockSize,
+                                 compress, cctxPtr->lz4CtxPtr, cctxPtr->prefs.compressionLevel,
+                                 cctxPtr->cdict,
+                                 cctxPtr->prefs.frameInfo.blockChecksumFlag);
+        srcPtr += blockSize;
+    }
+
+    if ((cctxPtr->prefs.autoFlush) && (srcPtr < srcEnd)) {
+        /* compress remaining input < blockSize */
+        lastBlockCompressed = fromSrcBuffer;
+        dstPtr += LZ4F_makeBlock(dstPtr,
+                                 srcPtr, (size_t)(srcEnd - srcPtr),
+                                 compress, cctxPtr->lz4CtxPtr, cctxPtr->prefs.compressionLevel,
+                                 cctxPtr->cdict,
+                                 cctxPtr->prefs.frameInfo.blockChecksumFlag);
+        srcPtr  = srcEnd;
+    }
+
+    /* preserve dictionary if necessary */
+    if ((cctxPtr->prefs.frameInfo.blockMode==LZ4F_blockLinked) && (lastBlockCompressed==fromSrcBuffer)) {
+        if (compressOptionsPtr->stableSrc) {
+            cctxPtr->tmpIn = cctxPtr->tmpBuff;
+        } else {
+            int const realDictSize = LZ4F_localSaveDict(cctxPtr);
+            if (realDictSize==0) return err0r(LZ4F_ERROR_GENERIC);
+            cctxPtr->tmpIn = cctxPtr->tmpBuff + realDictSize;
+        }
+    }
+
+    /* keep tmpIn within limits */
+    if ((cctxPtr->tmpIn + blockSize) > (cctxPtr->tmpBuff + cctxPtr->maxBufferSize)   /* necessarily LZ4F_blockLinked && lastBlockCompressed==fromTmpBuffer */
+        && !(cctxPtr->prefs.autoFlush))
+    {
+        int const realDictSize = LZ4F_localSaveDict(cctxPtr);
+        cctxPtr->tmpIn = cctxPtr->tmpBuff + realDictSize;
+    }
+
+    /* some input data left, necessarily < blockSize */
+    if (srcPtr < srcEnd) {
+        /* fill tmp buffer */
+        size_t const sizeToCopy = (size_t)(srcEnd - srcPtr);
+        memcpy(cctxPtr->tmpIn, srcPtr, sizeToCopy);
+        cctxPtr->tmpInSize = sizeToCopy;
+    }
+
+    if (cctxPtr->prefs.frameInfo.contentChecksumFlag == LZ4F_contentChecksumEnabled)
+        (void)XXH32_update(&(cctxPtr->xxh), srcBuffer, srcSize);
+
+    cctxPtr->totalInSize += srcSize;
+    return (size_t)(dstPtr - dstStart);
+}
+
+
+/*! LZ4F_flush() :
+ *  When compressed data must be sent immediately, without waiting for a block to be filled,
+ *  invoke LZ4_flush(), which will immediately compress any remaining data stored within LZ4F_cctx.
+ *  The result of the function is the number of bytes written into dstBuffer.
+ *  It can be zero, this means there was no data left within LZ4F_cctx.
+ *  The function outputs an error code if it fails (can be tested using LZ4F_isError())
+ *  LZ4F_compressOptions_t* is optional. NULL is a valid argument.
+ */
+size_t LZ4F_flush(LZ4F_cctx* cctxPtr,
+                  void* dstBuffer, size_t dstCapacity,
+            const LZ4F_compressOptions_t* compressOptionsPtr)
+{
+    BYTE* const dstStart = (BYTE*)dstBuffer;
+    BYTE* dstPtr = dstStart;
+    compressFunc_t compress;
+
+    if (cctxPtr->tmpInSize == 0) return 0;   /* nothing to flush */
+    if (cctxPtr->cStage != 1) return err0r(LZ4F_ERROR_GENERIC);
+    if (dstCapacity < (cctxPtr->tmpInSize + BHSize + BFSize))
+        return err0r(LZ4F_ERROR_dstMaxSize_tooSmall);
+    (void)compressOptionsPtr;   /* not yet useful */
+
+    /* select compression function */
+    compress = LZ4F_selectCompression(cctxPtr->prefs.frameInfo.blockMode, cctxPtr->prefs.compressionLevel);
+
+    /* compress tmp buffer */
+    dstPtr += LZ4F_makeBlock(dstPtr,
+                             cctxPtr->tmpIn, cctxPtr->tmpInSize,
+                             compress, cctxPtr->lz4CtxPtr, cctxPtr->prefs.compressionLevel,
+                             cctxPtr->cdict,
+                             cctxPtr->prefs.frameInfo.blockChecksumFlag);
+    assert(((void)"flush overflows dstBuffer!", (size_t)(dstPtr - dstStart) <= dstCapacity));
+
+    if (cctxPtr->prefs.frameInfo.blockMode == LZ4F_blockLinked)
+        cctxPtr->tmpIn += cctxPtr->tmpInSize;
+    cctxPtr->tmpInSize = 0;
+
+    /* keep tmpIn within limits */
+    if ((cctxPtr->tmpIn + cctxPtr->maxBlockSize) > (cctxPtr->tmpBuff + cctxPtr->maxBufferSize)) {  /* necessarily LZ4F_blockLinked */
+        int const realDictSize = LZ4F_localSaveDict(cctxPtr);
+        cctxPtr->tmpIn = cctxPtr->tmpBuff + realDictSize;
+    }
+
+    return (size_t)(dstPtr - dstStart);
+}
+
+
+/*! LZ4F_compressEnd() :
+ *  When you want to properly finish the compressed frame, just call LZ4F_compressEnd().
+ *  It will flush whatever data remained within compressionContext (like LZ4_flush())
+ *  but also properly finalize the frame, with an endMark and an (optional) checksum.
+ *  LZ4F_compressOptions_t structure is optional : you can provide NULL as argument.
+ * @return: the number of bytes written into dstBuffer (necessarily >= 4 (endMark size))
+ *       or an error code if it fails (can be tested using LZ4F_isError())
+ *  The context can then be used again to compress a new frame, starting with LZ4F_compressBegin().
+ */
+size_t LZ4F_compressEnd(LZ4F_cctx* cctxPtr,
+                        void* dstBuffer, size_t dstCapacity,
+                  const LZ4F_compressOptions_t* compressOptionsPtr)
+{
+    BYTE* const dstStart = (BYTE*)dstBuffer;
+    BYTE* dstPtr = dstStart;
+
+    size_t const flushSize = LZ4F_flush(cctxPtr, dstBuffer, dstCapacity, compressOptionsPtr);
+    DEBUGLOG(5,"LZ4F_compressEnd: dstCapacity=%u", (unsigned)dstCapacity);
+    if (LZ4F_isError(flushSize)) return flushSize;
+    dstPtr += flushSize;
+
+    assert(flushSize <= dstCapacity);
+    dstCapacity -= flushSize;
+
+    if (dstCapacity < 4) return err0r(LZ4F_ERROR_dstMaxSize_tooSmall);
+    LZ4F_writeLE32(dstPtr, 0);
+    dstPtr += 4;   /* endMark */
+
+    if (cctxPtr->prefs.frameInfo.contentChecksumFlag == LZ4F_contentChecksumEnabled) {
+        U32 const xxh = XXH32_digest(&(cctxPtr->xxh));
+        if (dstCapacity < 8) return err0r(LZ4F_ERROR_dstMaxSize_tooSmall);
+        DEBUGLOG(5,"Writing 32-bit content checksum");
+        LZ4F_writeLE32(dstPtr, xxh);
+        dstPtr+=4;   /* content Checksum */
+    }
+
+    cctxPtr->cStage = 0;   /* state is now re-usable (with identical preferences) */
+    cctxPtr->maxBufferSize = 0;  /* reuse HC context */
+
+    if (cctxPtr->prefs.frameInfo.contentSize) {
+        if (cctxPtr->prefs.frameInfo.contentSize != cctxPtr->totalInSize)
+            return err0r(LZ4F_ERROR_frameSize_wrong);
+    }
+
+    return (size_t)(dstPtr - dstStart);
+}
+
+
+/*-***************************************************
+*   Frame Decompression
+*****************************************************/
+
+typedef enum {
+    dstage_getFrameHeader=0, dstage_storeFrameHeader,
+    dstage_init,
+    dstage_getBlockHeader, dstage_storeBlockHeader,
+    dstage_copyDirect, dstage_getBlockChecksum,
+    dstage_getCBlock, dstage_storeCBlock,
+    dstage_flushOut,
+    dstage_getSuffix, dstage_storeSuffix,
+    dstage_getSFrameSize, dstage_storeSFrameSize,
+    dstage_skipSkippable
+} dStage_t;
+
+struct LZ4F_dctx_s {
+    LZ4F_frameInfo_t frameInfo;
+    U32    version;
+    dStage_t dStage;
+    U64    frameRemainingSize;
+    size_t maxBlockSize;
+    size_t maxBufferSize;
+    BYTE*  tmpIn;
+    size_t tmpInSize;
+    size_t tmpInTarget;
+    BYTE*  tmpOutBuffer;
+    const BYTE* dict;
+    size_t dictSize;
+    BYTE*  tmpOut;
+    size_t tmpOutSize;
+    size_t tmpOutStart;
+    XXH32_state_t xxh;
+    XXH32_state_t blockChecksum;
+    BYTE   header[LZ4F_HEADER_SIZE_MAX];
+};  /* typedef'd to LZ4F_dctx in lz4frame.h */
+
+
+/*! LZ4F_createDecompressionContext() :
+ *  Create a decompressionContext object, which will track all decompression operations.
+ *  Provides a pointer to a fully allocated and initialized LZ4F_decompressionContext object.
+ *  Object can later be released using LZ4F_freeDecompressionContext().
+ * @return : if != 0, there was an error during context creation.
+ */
+LZ4F_errorCode_t LZ4F_createDecompressionContext(LZ4F_dctx** LZ4F_decompressionContextPtr, unsigned versionNumber)
+{
+    LZ4F_dctx* const dctx = (LZ4F_dctx*)ALLOC_AND_ZERO(sizeof(LZ4F_dctx));
+    if (dctx == NULL) {  /* failed allocation */
+        *LZ4F_decompressionContextPtr = NULL;
+        return err0r(LZ4F_ERROR_allocation_failed);
+    }
+
+    dctx->version = versionNumber;
+    *LZ4F_decompressionContextPtr = dctx;
+    return LZ4F_OK_NoError;
+}
+
+LZ4F_errorCode_t LZ4F_freeDecompressionContext(LZ4F_dctx* dctx)
+{
+    LZ4F_errorCode_t result = LZ4F_OK_NoError;
+    if (dctx != NULL) {   /* can accept NULL input, like free() */
+      result = (LZ4F_errorCode_t)dctx->dStage;
+      FREEMEM(dctx->tmpIn);
+      FREEMEM(dctx->tmpOutBuffer);
+      FREEMEM(dctx);
+    }
+    return result;
+}
+
+
+/*==---   Streaming Decompression operations   ---==*/
+
+void LZ4F_resetDecompressionContext(LZ4F_dctx* dctx)
+{
+    dctx->dStage = dstage_getFrameHeader;
+    dctx->dict = NULL;
+    dctx->dictSize = 0;
+}
+
+
+/*! LZ4F_decodeHeader() :
+ *  input   : `src` points at the **beginning of the frame**
+ *  output  : set internal values of dctx, such as
+ *            dctx->frameInfo and dctx->dStage.
+ *            Also allocates internal buffers.
+ *  @return : nb Bytes read from src (necessarily <= srcSize)
+ *            or an error code (testable with LZ4F_isError())
+ */
+static size_t LZ4F_decodeHeader(LZ4F_dctx* dctx, const void* src, size_t srcSize)
+{
+    unsigned blockMode, blockChecksumFlag, contentSizeFlag, contentChecksumFlag, dictIDFlag, blockSizeID;
+    size_t frameHeaderSize;
+    const BYTE* srcPtr = (const BYTE*)src;
+
+    DEBUGLOG(5, "LZ4F_decodeHeader");
+    /* need to decode header to get frameInfo */
+    if (srcSize < minFHSize) return err0r(LZ4F_ERROR_frameHeader_incomplete);   /* minimal frame header size */
+    MEM_INIT(&(dctx->frameInfo), 0, sizeof(dctx->frameInfo));
+
+    /* special case : skippable frames */
+    if ((LZ4F_readLE32(srcPtr) & 0xFFFFFFF0U) == LZ4F_MAGIC_SKIPPABLE_START) {
+        dctx->frameInfo.frameType = LZ4F_skippableFrame;
+        if (src == (void*)(dctx->header)) {
+            dctx->tmpInSize = srcSize;
+            dctx->tmpInTarget = 8;
+            dctx->dStage = dstage_storeSFrameSize;
+            return srcSize;
+        } else {
+            dctx->dStage = dstage_getSFrameSize;
+            return 4;
+        }
+    }
+
+    /* control magic number */
+#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
+    if (LZ4F_readLE32(srcPtr) != LZ4F_MAGICNUMBER) {
+        DEBUGLOG(4, "frame header error : unknown magic number");
+        return err0r(LZ4F_ERROR_frameType_unknown);
+    }
+#endif
+    dctx->frameInfo.frameType = LZ4F_frame;
+
+    /* Flags */
+    {   U32 const FLG = srcPtr[4];
+        U32 const version = (FLG>>6) & _2BITS;
+        blockChecksumFlag = (FLG>>4) & _1BIT;
+        blockMode = (FLG>>5) & _1BIT;
+        contentSizeFlag = (FLG>>3) & _1BIT;
+        contentChecksumFlag = (FLG>>2) & _1BIT;
+        dictIDFlag = FLG & _1BIT;
+        /* validate */
+        if (((FLG>>1)&_1BIT) != 0) return err0r(LZ4F_ERROR_reservedFlag_set); /* Reserved bit */
+        if (version != 1) return err0r(LZ4F_ERROR_headerVersion_wrong);        /* Version Number, only supported value */
+    }
+
+    /* Frame Header Size */
+    frameHeaderSize = minFHSize + (contentSizeFlag?8:0) + (dictIDFlag?4:0);
+
+    if (srcSize < frameHeaderSize) {
+        /* not enough input to fully decode frame header */
+        if (srcPtr != dctx->header)
+            memcpy(dctx->header, srcPtr, srcSize);
+        dctx->tmpInSize = srcSize;
+        dctx->tmpInTarget = frameHeaderSize;
+        dctx->dStage = dstage_storeFrameHeader;
+        return srcSize;
+    }
+
+    {   U32 const BD = srcPtr[5];
+        blockSizeID = (BD>>4) & _3BITS;
+        /* validate */
+        if (((BD>>7)&_1BIT) != 0) return err0r(LZ4F_ERROR_reservedFlag_set);   /* Reserved bit */
+        if (blockSizeID < 4) return err0r(LZ4F_ERROR_maxBlockSize_invalid);    /* 4-7 only supported values for the time being */
+        if (((BD>>0)&_4BITS) != 0) return err0r(LZ4F_ERROR_reservedFlag_set);  /* Reserved bits */
+    }
+
+    /* check header */
+    assert(frameHeaderSize > 5);
+#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
+    {   BYTE const HC = LZ4F_headerChecksum(srcPtr+4, frameHeaderSize-5);
+        if (HC != srcPtr[frameHeaderSize-1])
+            return err0r(LZ4F_ERROR_headerChecksum_invalid);
+    }
+#endif
+
+    /* save */
+    dctx->frameInfo.blockMode = (LZ4F_blockMode_t)blockMode;
+    dctx->frameInfo.blockChecksumFlag = (LZ4F_blockChecksum_t)blockChecksumFlag;
+    dctx->frameInfo.contentChecksumFlag = (LZ4F_contentChecksum_t)contentChecksumFlag;
+    dctx->frameInfo.blockSizeID = (LZ4F_blockSizeID_t)blockSizeID;
+    dctx->maxBlockSize = LZ4F_getBlockSize(blockSizeID);
+    if (contentSizeFlag)
+        dctx->frameRemainingSize =
+            dctx->frameInfo.contentSize = LZ4F_readLE64(srcPtr+6);
+    if (dictIDFlag)
+        dctx->frameInfo.dictID = LZ4F_readLE32(srcPtr + frameHeaderSize - 5);
+
+    dctx->dStage = dstage_init;
+
+    return frameHeaderSize;
+}
+
+
+/*! LZ4F_headerSize() :
+ * @return : size of frame header
+ *           or an error code, which can be tested using LZ4F_isError()
+ */
+size_t LZ4F_headerSize(const void* src, size_t srcSize)
+{
+    if (src == NULL) return err0r(LZ4F_ERROR_srcPtr_wrong);
+
+    /* minimal srcSize to determine header size */
+    if (srcSize < LZ4F_MIN_SIZE_TO_KNOW_HEADER_LENGTH)
+        return err0r(LZ4F_ERROR_frameHeader_incomplete);
+
+    /* special case : skippable frames */
+    if ((LZ4F_readLE32(src) & 0xFFFFFFF0U) == LZ4F_MAGIC_SKIPPABLE_START)
+        return 8;
+
+    /* control magic number */
+#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
+    if (LZ4F_readLE32(src) != LZ4F_MAGICNUMBER)
+        return err0r(LZ4F_ERROR_frameType_unknown);
+#endif
+
+    /* Frame Header Size */
+    {   BYTE const FLG = ((const BYTE*)src)[4];
+        U32 const contentSizeFlag = (FLG>>3) & _1BIT;
+        U32 const dictIDFlag = FLG & _1BIT;
+        return minFHSize + (contentSizeFlag?8:0) + (dictIDFlag?4:0);
+    }
+}
+
+/*! LZ4F_getFrameInfo() :
+ *  This function extracts frame parameters (max blockSize, frame checksum, etc.).
+ *  Usage is optional. Objective is to provide relevant information for allocation purposes.
+ *  This function works in 2 situations :
+ *   - At the beginning of a new frame, in which case it will decode this information from `srcBuffer`, and start the decoding process.
+ *     Amount of input data provided must be large enough to successfully decode the frame header.
+ *     A header size is variable, but is guaranteed to be <= LZ4F_HEADER_SIZE_MAX bytes. It's possible to provide more input data than this minimum.
+ *   - After decoding has been started. In which case, no input is read, frame parameters are extracted from dctx.
+ *  The number of bytes consumed from srcBuffer will be updated within *srcSizePtr (necessarily <= original value).
+ *  Decompression must resume from (srcBuffer + *srcSizePtr).
+ * @return : an hint about how many srcSize bytes LZ4F_decompress() expects for next call,
+ *           or an error code which can be tested using LZ4F_isError()
+ *  note 1 : in case of error, dctx is not modified. Decoding operations can resume from where they stopped.
+ *  note 2 : frame parameters are *copied into* an already allocated LZ4F_frameInfo_t structure.
+ */
+LZ4F_errorCode_t LZ4F_getFrameInfo(LZ4F_dctx* dctx,
+                                   LZ4F_frameInfo_t* frameInfoPtr,
+                             const void* srcBuffer, size_t* srcSizePtr)
+{
+    LZ4F_STATIC_ASSERT(dstage_getFrameHeader < dstage_storeFrameHeader);
+    if (dctx->dStage > dstage_storeFrameHeader) {
+        /* frameInfo already decoded */
+        size_t o=0, i=0;
+        *srcSizePtr = 0;
+        *frameInfoPtr = dctx->frameInfo;
+        /* returns : recommended nb of bytes for LZ4F_decompress() */
+        return LZ4F_decompress(dctx, NULL, &o, NULL, &i, NULL);
+    } else {
+        if (dctx->dStage == dstage_storeFrameHeader) {
+            /* frame decoding already started, in the middle of header => automatic fail */
+            *srcSizePtr = 0;
+            return err0r(LZ4F_ERROR_frameDecoding_alreadyStarted);
+        } else {
+            size_t const hSize = LZ4F_headerSize(srcBuffer, *srcSizePtr);
+            if (LZ4F_isError(hSize)) { *srcSizePtr=0; return hSize; }
+            if (*srcSizePtr < hSize) {
+                *srcSizePtr=0;
+                return err0r(LZ4F_ERROR_frameHeader_incomplete);
+            }
+
+            {   size_t decodeResult = LZ4F_decodeHeader(dctx, srcBuffer, hSize);
+                if (LZ4F_isError(decodeResult)) {
+                    *srcSizePtr = 0;
+                } else {
+                    *srcSizePtr = decodeResult;
+                    decodeResult = BHSize;   /* block header size */
+                }
+                *frameInfoPtr = dctx->frameInfo;
+                return decodeResult;
+    }   }   }
+}
+
+
+/* LZ4F_updateDict() :
+ * only used for LZ4F_blockLinked mode
+ * Condition : dstPtr != NULL
+ */
+static void LZ4F_updateDict(LZ4F_dctx* dctx,
+                      const BYTE* dstPtr, size_t dstSize, const BYTE* dstBufferStart,
+                      unsigned withinTmp)
+{
+    assert(dstPtr != NULL);
+    if (dctx->dictSize==0) {
+        dctx->dict = (const BYTE*)dstPtr;   /* priority to prefix mode */
+    }
+    assert(dctx->dict != NULL);
+
+    if (dctx->dict + dctx->dictSize == dstPtr) {  /* prefix mode, everything within dstBuffer */
+        dctx->dictSize += dstSize;
+        return;
+    }
+
+    assert(dstPtr >= dstBufferStart);
+    if ((size_t)(dstPtr - dstBufferStart) + dstSize >= 64 KB) {  /* history in dstBuffer becomes large enough to become dictionary */
+        dctx->dict = (const BYTE*)dstBufferStart;
+        dctx->dictSize = (size_t)(dstPtr - dstBufferStart) + dstSize;
+        return;
+    }
+
+    assert(dstSize < 64 KB);   /* if dstSize >= 64 KB, dictionary would be set into dstBuffer directly */
+
+    /* dstBuffer does not contain whole useful history (64 KB), so it must be saved within tmpOutBuffer */
+    assert(dctx->tmpOutBuffer != NULL);
+
+    if (withinTmp && (dctx->dict == dctx->tmpOutBuffer)) {   /* continue history within tmpOutBuffer */
+        /* withinTmp expectation : content of [dstPtr,dstSize] is same as [dict+dictSize,dstSize], so we just extend it */
+        assert(dctx->dict + dctx->dictSize == dctx->tmpOut + dctx->tmpOutStart);
+        dctx->dictSize += dstSize;
+        return;
+    }
+
+    if (withinTmp) { /* copy relevant dict portion in front of tmpOut within tmpOutBuffer */
+        size_t const preserveSize = (size_t)(dctx->tmpOut - dctx->tmpOutBuffer);
+        size_t copySize = 64 KB - dctx->tmpOutSize;
+        const BYTE* const oldDictEnd = dctx->dict + dctx->dictSize - dctx->tmpOutStart;
+        if (dctx->tmpOutSize > 64 KB) copySize = 0;
+        if (copySize > preserveSize) copySize = preserveSize;
+
+        memcpy(dctx->tmpOutBuffer + preserveSize - copySize, oldDictEnd - copySize, copySize);
+
+        dctx->dict = dctx->tmpOutBuffer;
+        dctx->dictSize = preserveSize + dctx->tmpOutStart + dstSize;
+        return;
+    }
+
+    if (dctx->dict == dctx->tmpOutBuffer) {    /* copy dst into tmp to complete dict */
+        if (dctx->dictSize + dstSize > dctx->maxBufferSize) {  /* tmp buffer not large enough */
+            size_t const preserveSize = 64 KB - dstSize;
+            memcpy(dctx->tmpOutBuffer, dctx->dict + dctx->dictSize - preserveSize, preserveSize);
+            dctx->dictSize = preserveSize;
+        }
+        memcpy(dctx->tmpOutBuffer + dctx->dictSize, dstPtr, dstSize);
+        dctx->dictSize += dstSize;
+        return;
+    }
+
+    /* join dict & dest into tmp */
+    {   size_t preserveSize = 64 KB - dstSize;
+        if (preserveSize > dctx->dictSize) preserveSize = dctx->dictSize;
+        memcpy(dctx->tmpOutBuffer, dctx->dict + dctx->dictSize - preserveSize, preserveSize);
+        memcpy(dctx->tmpOutBuffer + preserveSize, dstPtr, dstSize);
+        dctx->dict = dctx->tmpOutBuffer;
+        dctx->dictSize = preserveSize + dstSize;
+    }
+}
+
+
+
+/*! LZ4F_decompress() :
+ *  Call this function repetitively to regenerate compressed data in srcBuffer.
+ *  The function will attempt to decode up to *srcSizePtr bytes from srcBuffer
+ *  into dstBuffer of capacity *dstSizePtr.
+ *
+ *  The number of bytes regenerated into dstBuffer will be provided within *dstSizePtr (necessarily <= original value).
+ *
+ *  The number of bytes effectively read from srcBuffer will be provided within *srcSizePtr (necessarily <= original value).
+ *  If number of bytes read is < number of bytes provided, then decompression operation is not complete.
+ *  Remaining data will have to be presented again in a subsequent invocation.
+ *
+ *  The function result is an hint of the better srcSize to use for next call to LZ4F_decompress.
+ *  Schematically, it's the size of the current (or remaining) compressed block + header of next block.
+ *  Respecting the hint provides a small boost to performance, since it allows less buffer shuffling.
+ *  Note that this is just a hint, and it's always possible to any srcSize value.
+ *  When a frame is fully decoded, @return will be 0.
+ *  If decompression failed, @return is an error code which can be tested using LZ4F_isError().
+ */
+size_t LZ4F_decompress(LZ4F_dctx* dctx,
+                       void* dstBuffer, size_t* dstSizePtr,
+                       const void* srcBuffer, size_t* srcSizePtr,
+                       const LZ4F_decompressOptions_t* decompressOptionsPtr)
+{
+    LZ4F_decompressOptions_t optionsNull;
+    const BYTE* const srcStart = (const BYTE*)srcBuffer;
+    const BYTE* const srcEnd = srcStart + *srcSizePtr;
+    const BYTE* srcPtr = srcStart;
+    BYTE* const dstStart = (BYTE*)dstBuffer;
+    BYTE* const dstEnd = dstStart ? dstStart + *dstSizePtr : NULL;
+    BYTE* dstPtr = dstStart;
+    const BYTE* selectedIn = NULL;
+    unsigned doAnotherStage = 1;
+    size_t nextSrcSizeHint = 1;
+
+
+    DEBUGLOG(5, "LZ4F_decompress : %p,%u => %p,%u",
+            srcBuffer, (unsigned)*srcSizePtr, dstBuffer, (unsigned)*dstSizePtr);
+    if (dstBuffer == NULL) assert(*dstSizePtr == 0);
+    MEM_INIT(&optionsNull, 0, sizeof(optionsNull));
+    if (decompressOptionsPtr==NULL) decompressOptionsPtr = &optionsNull;
+    *srcSizePtr = 0;
+    *dstSizePtr = 0;
+    assert(dctx != NULL);
+
+    /* behaves as a state machine */
+
+    while (doAnotherStage) {
+
+        switch(dctx->dStage)
+        {
+
+        case dstage_getFrameHeader:
+            DEBUGLOG(6, "dstage_getFrameHeader");
+            if ((size_t)(srcEnd-srcPtr) >= maxFHSize) {  /* enough to decode - shortcut */
+                size_t const hSize = LZ4F_decodeHeader(dctx, srcPtr, (size_t)(srcEnd-srcPtr));  /* will update dStage appropriately */
+                if (LZ4F_isError(hSize)) return hSize;
+                srcPtr += hSize;
+                break;
+            }
+            dctx->tmpInSize = 0;
+            if (srcEnd-srcPtr == 0) return minFHSize;   /* 0-size input */
+            dctx->tmpInTarget = minFHSize;   /* minimum size to decode header */
+            dctx->dStage = dstage_storeFrameHeader;
+            /* fall-through */
+
+        case dstage_storeFrameHeader:
+            DEBUGLOG(6, "dstage_storeFrameHeader");
+            {   size_t const sizeToCopy = MIN(dctx->tmpInTarget - dctx->tmpInSize, (size_t)(srcEnd - srcPtr));
+                memcpy(dctx->header + dctx->tmpInSize, srcPtr, sizeToCopy);
+                dctx->tmpInSize += sizeToCopy;
+                srcPtr += sizeToCopy;
+            }
+            if (dctx->tmpInSize < dctx->tmpInTarget) {
+                nextSrcSizeHint = (dctx->tmpInTarget - dctx->tmpInSize) + BHSize;   /* rest of header + nextBlockHeader */
+                doAnotherStage = 0;   /* not enough src data, ask for some more */
+                break;
+            }
+            {   size_t const hSize = LZ4F_decodeHeader(dctx, dctx->header, dctx->tmpInTarget);  /* will update dStage appropriately */
+                if (LZ4F_isError(hSize)) return hSize;
+            }
+            break;
+
+        case dstage_init:
+            DEBUGLOG(6, "dstage_init");
+            if (dctx->frameInfo.contentChecksumFlag) (void)XXH32_reset(&(dctx->xxh), 0);
+            /* internal buffers allocation */
+            {   size_t const bufferNeeded = dctx->maxBlockSize
+                    + ((dctx->frameInfo.blockMode==LZ4F_blockLinked) ? 128 KB : 0);
+                if (bufferNeeded > dctx->maxBufferSize) {   /* tmp buffers too small */
+                    dctx->maxBufferSize = 0;   /* ensure allocation will be re-attempted on next entry*/
+                    FREEMEM(dctx->tmpIn);
+                    dctx->tmpIn = (BYTE*)ALLOC(dctx->maxBlockSize + BFSize /* block checksum */);
+                    if (dctx->tmpIn == NULL)
+                        return err0r(LZ4F_ERROR_allocation_failed);
+                    FREEMEM(dctx->tmpOutBuffer);
+                    dctx->tmpOutBuffer= (BYTE*)ALLOC(bufferNeeded);
+                    if (dctx->tmpOutBuffer== NULL)
+                        return err0r(LZ4F_ERROR_allocation_failed);
+                    dctx->maxBufferSize = bufferNeeded;
+            }   }
+            dctx->tmpInSize = 0;
+            dctx->tmpInTarget = 0;
+            dctx->tmpOut = dctx->tmpOutBuffer;
+            dctx->tmpOutStart = 0;
+            dctx->tmpOutSize = 0;
+
+            dctx->dStage = dstage_getBlockHeader;
+            /* fall-through */
+
+        case dstage_getBlockHeader:
+            if ((size_t)(srcEnd - srcPtr) >= BHSize) {
+                selectedIn = srcPtr;
+                srcPtr += BHSize;
+            } else {
+                /* not enough input to read cBlockSize field */
+                dctx->tmpInSize = 0;
+                dctx->dStage = dstage_storeBlockHeader;
+            }
+
+            if (dctx->dStage == dstage_storeBlockHeader)   /* can be skipped */
+        case dstage_storeBlockHeader:
+            {   size_t const remainingInput = (size_t)(srcEnd - srcPtr);
+                size_t const wantedData = BHSize - dctx->tmpInSize;
+                size_t const sizeToCopy = MIN(wantedData, remainingInput);
+                memcpy(dctx->tmpIn + dctx->tmpInSize, srcPtr, sizeToCopy);
+                srcPtr += sizeToCopy;
+                dctx->tmpInSize += sizeToCopy;
+
+                if (dctx->tmpInSize < BHSize) {   /* not enough input for cBlockSize */
+                    nextSrcSizeHint = BHSize - dctx->tmpInSize;
+                    doAnotherStage  = 0;
+                    break;
+                }
+                selectedIn = dctx->tmpIn;
+            }   /* if (dctx->dStage == dstage_storeBlockHeader) */
+
+        /* decode block header */
+            {   U32 const blockHeader = LZ4F_readLE32(selectedIn);
+                size_t const nextCBlockSize = blockHeader & 0x7FFFFFFFU;
+                size_t const crcSize = dctx->frameInfo.blockChecksumFlag * BFSize;
+                if (blockHeader==0) {  /* frameEnd signal, no more block */
+                    DEBUGLOG(5, "end of frame");
+                    dctx->dStage = dstage_getSuffix;
+                    break;
+                }
+                if (nextCBlockSize > dctx->maxBlockSize) {
+                    return err0r(LZ4F_ERROR_maxBlockSize_invalid);
+                }
+                if (blockHeader & LZ4F_BLOCKUNCOMPRESSED_FLAG) {
+                    /* next block is uncompressed */
+                    dctx->tmpInTarget = nextCBlockSize;
+                    DEBUGLOG(5, "next block is uncompressed (size %u)", (U32)nextCBlockSize);
+                    if (dctx->frameInfo.blockChecksumFlag) {
+                        (void)XXH32_reset(&dctx->blockChecksum, 0);
+                    }
+                    dctx->dStage = dstage_copyDirect;
+                    break;
+                }
+                /* next block is a compressed block */
+                dctx->tmpInTarget = nextCBlockSize + crcSize;
+                dctx->dStage = dstage_getCBlock;
+                if (dstPtr==dstEnd || srcPtr==srcEnd) {
+                    nextSrcSizeHint = BHSize + nextCBlockSize + crcSize;
+                    doAnotherStage = 0;
+                }
+                break;
+            }
+
+        case dstage_copyDirect:   /* uncompressed block */
+            DEBUGLOG(6, "dstage_copyDirect");
+            {   size_t sizeToCopy;
+                if (dstPtr == NULL) {
+                    sizeToCopy = 0;
+                } else {
+                    size_t const minBuffSize = MIN((size_t)(srcEnd-srcPtr), (size_t)(dstEnd-dstPtr));
+                    sizeToCopy = MIN(dctx->tmpInTarget, minBuffSize);
+                    memcpy(dstPtr, srcPtr, sizeToCopy);
+                    if (dctx->frameInfo.blockChecksumFlag) {
+                        (void)XXH32_update(&dctx->blockChecksum, srcPtr, sizeToCopy);
+                    }
+                    if (dctx->frameInfo.contentChecksumFlag)
+                        (void)XXH32_update(&dctx->xxh, srcPtr, sizeToCopy);
+                    if (dctx->frameInfo.contentSize)
+                        dctx->frameRemainingSize -= sizeToCopy;
+
+                    /* history management (linked blocks only)*/
+                    if (dctx->frameInfo.blockMode == LZ4F_blockLinked) {
+                        LZ4F_updateDict(dctx, dstPtr, sizeToCopy, dstStart, 0);
+                }   }
+
+                srcPtr += sizeToCopy;
+                dstPtr += sizeToCopy;
+                if (sizeToCopy == dctx->tmpInTarget) {   /* all done */
+                    if (dctx->frameInfo.blockChecksumFlag) {
+                        dctx->tmpInSize = 0;
+                        dctx->dStage = dstage_getBlockChecksum;
+                    } else
+                        dctx->dStage = dstage_getBlockHeader;  /* new block */
+                    break;
+                }
+                dctx->tmpInTarget -= sizeToCopy;  /* need to copy more */
+            }
+            nextSrcSizeHint = dctx->tmpInTarget +
+                            +(dctx->frameInfo.blockChecksumFlag ? BFSize : 0)
+                            + BHSize /* next header size */;
+            doAnotherStage = 0;
+            break;
+
+        /* check block checksum for recently transferred uncompressed block */
+        case dstage_getBlockChecksum:
+            DEBUGLOG(6, "dstage_getBlockChecksum");
+            {   const void* crcSrc;
+                if ((srcEnd-srcPtr >= 4) && (dctx->tmpInSize==0)) {
+                    crcSrc = srcPtr;
+                    srcPtr += 4;
+                } else {
+                    size_t const stillToCopy = 4 - dctx->tmpInSize;
+                    size_t const sizeToCopy = MIN(stillToCopy, (size_t)(srcEnd-srcPtr));
+                    memcpy(dctx->header + dctx->tmpInSize, srcPtr, sizeToCopy);
+                    dctx->tmpInSize += sizeToCopy;
+                    srcPtr += sizeToCopy;
+                    if (dctx->tmpInSize < 4) {  /* all input consumed */
+                        doAnotherStage = 0;
+                        break;
+                    }
+                    crcSrc = dctx->header;
+                }
+                {   U32 const readCRC = LZ4F_readLE32(crcSrc);
+                    U32 const calcCRC = XXH32_digest(&dctx->blockChecksum);
+#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
+                    DEBUGLOG(6, "compare block checksum");
+                    if (readCRC != calcCRC) {
+                        DEBUGLOG(4, "incorrect block checksum: %08X != %08X",
+                                readCRC, calcCRC);
+                        return err0r(LZ4F_ERROR_blockChecksum_invalid);
+                    }
+#else
+                    (void)readCRC;
+                    (void)calcCRC;
+#endif
+            }   }
+            dctx->dStage = dstage_getBlockHeader;  /* new block */
+            break;
+
+        case dstage_getCBlock:
+            DEBUGLOG(6, "dstage_getCBlock");
+            if ((size_t)(srcEnd-srcPtr) < dctx->tmpInTarget) {
+                dctx->tmpInSize = 0;
+                dctx->dStage = dstage_storeCBlock;
+                break;
+            }
+            /* input large enough to read full block directly */
+            selectedIn = srcPtr;
+            srcPtr += dctx->tmpInTarget;
+
+            if (0)  /* always jump over next block */
+        case dstage_storeCBlock:
+            {   size_t const wantedData = dctx->tmpInTarget - dctx->tmpInSize;
+                size_t const inputLeft = (size_t)(srcEnd-srcPtr);
+                size_t const sizeToCopy = MIN(wantedData, inputLeft);
+                memcpy(dctx->tmpIn + dctx->tmpInSize, srcPtr, sizeToCopy);
+                dctx->tmpInSize += sizeToCopy;
+                srcPtr += sizeToCopy;
+                if (dctx->tmpInSize < dctx->tmpInTarget) { /* need more input */
+                    nextSrcSizeHint = (dctx->tmpInTarget - dctx->tmpInSize)
+                                    + (dctx->frameInfo.blockChecksumFlag ? BFSize : 0)
+                                    + BHSize /* next header size */;
+                    doAnotherStage = 0;
+                    break;
+                }
+                selectedIn = dctx->tmpIn;
+            }
+
+            /* At this stage, input is large enough to decode a block */
+            if (dctx->frameInfo.blockChecksumFlag) {
+                dctx->tmpInTarget -= 4;
+                assert(selectedIn != NULL);  /* selectedIn is defined at this stage (either srcPtr, or dctx->tmpIn) */
+                {   U32 const readBlockCrc = LZ4F_readLE32(selectedIn + dctx->tmpInTarget);
+                    U32 const calcBlockCrc = XXH32(selectedIn, dctx->tmpInTarget, 0);
+#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
+                    if (readBlockCrc != calcBlockCrc)
+                        return err0r(LZ4F_ERROR_blockChecksum_invalid);
+#else
+                    (void)readBlockCrc;
+                    (void)calcBlockCrc;
+#endif
+            }   }
+
+            if ((size_t)(dstEnd-dstPtr) >= dctx->maxBlockSize) {
+                const char* dict = (const char*)dctx->dict;
+                size_t dictSize = dctx->dictSize;
+                int decodedSize;
+                assert(dstPtr != NULL);
+                if (dict && dictSize > 1 GB) {
+                    /* the dictSize param is an int, avoid truncation / sign issues */
+                    dict += dictSize - 64 KB;
+                    dictSize = 64 KB;
+                }
+                /* enough capacity in `dst` to decompress directly there */
+                decodedSize = LZ4_decompress_safe_usingDict(
+                        (const char*)selectedIn, (char*)dstPtr,
+                        (int)dctx->tmpInTarget, (int)dctx->maxBlockSize,
+                        dict, (int)dictSize);
+                if (decodedSize < 0) return err0r(LZ4F_ERROR_GENERIC);   /* decompression failed */
+                if (dctx->frameInfo.contentChecksumFlag)
+                    XXH32_update(&(dctx->xxh), dstPtr, (size_t)decodedSize);
+                if (dctx->frameInfo.contentSize)
+                    dctx->frameRemainingSize -= (size_t)decodedSize;
+
+                /* dictionary management */
+                if (dctx->frameInfo.blockMode==LZ4F_blockLinked) {
+                    LZ4F_updateDict(dctx, dstPtr, (size_t)decodedSize, dstStart, 0);
+                }
+
+                dstPtr += decodedSize;
+                dctx->dStage = dstage_getBlockHeader;
+                break;
+            }
+
+            /* not enough place into dst : decode into tmpOut */
+            /* ensure enough place for tmpOut */
+            if (dctx->frameInfo.blockMode == LZ4F_blockLinked) {
+                if (dctx->dict == dctx->tmpOutBuffer) {
+                    if (dctx->dictSize > 128 KB) {
+                        memcpy(dctx->tmpOutBuffer, dctx->dict + dctx->dictSize - 64 KB, 64 KB);
+                        dctx->dictSize = 64 KB;
+                    }
+                    dctx->tmpOut = dctx->tmpOutBuffer + dctx->dictSize;
+                } else {  /* dict not within tmp */
+                    size_t const reservedDictSpace = MIN(dctx->dictSize, 64 KB);
+                    dctx->tmpOut = dctx->tmpOutBuffer + reservedDictSpace;
+            }   }
+
+            /* Decode block */
+            {   const char* dict = (const char*)dctx->dict;
+                size_t dictSize = dctx->dictSize;
+                int decodedSize;
+                if (dict && dictSize > 1 GB) {
+                    /* the dictSize param is an int, avoid truncation / sign issues */
+                    dict += dictSize - 64 KB;
+                    dictSize = 64 KB;
+                }
+                decodedSize = LZ4_decompress_safe_usingDict(
+                        (const char*)selectedIn, (char*)dctx->tmpOut,
+                        (int)dctx->tmpInTarget, (int)dctx->maxBlockSize,
+                        dict, (int)dictSize);
+                if (decodedSize < 0)  /* decompression failed */
+                    return err0r(LZ4F_ERROR_decompressionFailed);
+                if (dctx->frameInfo.contentChecksumFlag)
+                    XXH32_update(&(dctx->xxh), dctx->tmpOut, (size_t)decodedSize);
+                if (dctx->frameInfo.contentSize)
+                    dctx->frameRemainingSize -= (size_t)decodedSize;
+                dctx->tmpOutSize = (size_t)decodedSize;
+                dctx->tmpOutStart = 0;
+                dctx->dStage = dstage_flushOut;
+            }
+            /* fall-through */
+
+        case dstage_flushOut:  /* flush decoded data from tmpOut to dstBuffer */
+            DEBUGLOG(6, "dstage_flushOut");
+            if (dstPtr != NULL) {
+                size_t const sizeToCopy = MIN(dctx->tmpOutSize - dctx->tmpOutStart, (size_t)(dstEnd-dstPtr));
+                memcpy(dstPtr, dctx->tmpOut + dctx->tmpOutStart, sizeToCopy);
+
+                /* dictionary management */
+                if (dctx->frameInfo.blockMode == LZ4F_blockLinked)
+                    LZ4F_updateDict(dctx, dstPtr, sizeToCopy, dstStart, 1 /*withinTmp*/);
+
+                dctx->tmpOutStart += sizeToCopy;
+                dstPtr += sizeToCopy;
+            }
+            if (dctx->tmpOutStart == dctx->tmpOutSize) { /* all flushed */
+                dctx->dStage = dstage_getBlockHeader;  /* get next block */
+                break;
+            }
+            /* could not flush everything : stop there, just request a block header */
+            doAnotherStage = 0;
+            nextSrcSizeHint = BHSize;
+            break;
+
+        case dstage_getSuffix:
+            if (dctx->frameRemainingSize)
+                return err0r(LZ4F_ERROR_frameSize_wrong);   /* incorrect frame size decoded */
+            if (!dctx->frameInfo.contentChecksumFlag) {  /* no checksum, frame is completed */
+                nextSrcSizeHint = 0;
+                LZ4F_resetDecompressionContext(dctx);
+                doAnotherStage = 0;
+                break;
+            }
+            if ((srcEnd - srcPtr) < 4) {  /* not enough size for entire CRC */
+                dctx->tmpInSize = 0;
+                dctx->dStage = dstage_storeSuffix;
+            } else {
+                selectedIn = srcPtr;
+                srcPtr += 4;
+            }
+
+            if (dctx->dStage == dstage_storeSuffix)   /* can be skipped */
+        case dstage_storeSuffix:
+            {   size_t const remainingInput = (size_t)(srcEnd - srcPtr);
+                size_t const wantedData = 4 - dctx->tmpInSize;
+                size_t const sizeToCopy = MIN(wantedData, remainingInput);
+                memcpy(dctx->tmpIn + dctx->tmpInSize, srcPtr, sizeToCopy);
+                srcPtr += sizeToCopy;
+                dctx->tmpInSize += sizeToCopy;
+                if (dctx->tmpInSize < 4) { /* not enough input to read complete suffix */
+                    nextSrcSizeHint = 4 - dctx->tmpInSize;
+                    doAnotherStage=0;
+                    break;
+                }
+                selectedIn = dctx->tmpIn;
+            }   /* if (dctx->dStage == dstage_storeSuffix) */
+
+        /* case dstage_checkSuffix: */   /* no direct entry, avoid initialization risks */
+            {   U32 const readCRC = LZ4F_readLE32(selectedIn);
+                U32 const resultCRC = XXH32_digest(&(dctx->xxh));
+#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
+                if (readCRC != resultCRC)
+                    return err0r(LZ4F_ERROR_contentChecksum_invalid);
+#else
+                (void)readCRC;
+                (void)resultCRC;
+#endif
+                nextSrcSizeHint = 0;
+                LZ4F_resetDecompressionContext(dctx);
+                doAnotherStage = 0;
+                break;
+            }
+
+        case dstage_getSFrameSize:
+            if ((srcEnd - srcPtr) >= 4) {
+                selectedIn = srcPtr;
+                srcPtr += 4;
+            } else {
+                /* not enough input to read cBlockSize field */
+                dctx->tmpInSize = 4;
+                dctx->tmpInTarget = 8;
+                dctx->dStage = dstage_storeSFrameSize;
+            }
+
+            if (dctx->dStage == dstage_storeSFrameSize)
+        case dstage_storeSFrameSize:
+            {   size_t const sizeToCopy = MIN(dctx->tmpInTarget - dctx->tmpInSize,
+                                             (size_t)(srcEnd - srcPtr) );
+                memcpy(dctx->header + dctx->tmpInSize, srcPtr, sizeToCopy);
+                srcPtr += sizeToCopy;
+                dctx->tmpInSize += sizeToCopy;
+                if (dctx->tmpInSize < dctx->tmpInTarget) {
+                    /* not enough input to get full sBlockSize; wait for more */
+                    nextSrcSizeHint = dctx->tmpInTarget - dctx->tmpInSize;
+                    doAnotherStage = 0;
+                    break;
+                }
+                selectedIn = dctx->header + 4;
+            }   /* if (dctx->dStage == dstage_storeSFrameSize) */
+
+        /* case dstage_decodeSFrameSize: */   /* no direct entry */
+            {   size_t const SFrameSize = LZ4F_readLE32(selectedIn);
+                dctx->frameInfo.contentSize = SFrameSize;
+                dctx->tmpInTarget = SFrameSize;
+                dctx->dStage = dstage_skipSkippable;
+                break;
+            }
+
+        case dstage_skipSkippable:
+            {   size_t const skipSize = MIN(dctx->tmpInTarget, (size_t)(srcEnd-srcPtr));
+                srcPtr += skipSize;
+                dctx->tmpInTarget -= skipSize;
+                doAnotherStage = 0;
+                nextSrcSizeHint = dctx->tmpInTarget;
+                if (nextSrcSizeHint) break;  /* still more to skip */
+                /* frame fully skipped : prepare context for a new frame */
+                LZ4F_resetDecompressionContext(dctx);
+                break;
+            }
+        }   /* switch (dctx->dStage) */
+    }   /* while (doAnotherStage) */
+
+    /* preserve history within tmp whenever necessary */
+    LZ4F_STATIC_ASSERT((unsigned)dstage_init == 2);
+    if ( (dctx->frameInfo.blockMode==LZ4F_blockLinked)  /* next block will use up to 64KB from previous ones */
+      && (dctx->dict != dctx->tmpOutBuffer)             /* dictionary is not already within tmp */
+      && (dctx->dict != NULL)                           /* dictionary exists */
+      && (!decompressOptionsPtr->stableDst)             /* cannot rely on dst data to remain there for next call */
+      && ((unsigned)(dctx->dStage)-2 < (unsigned)(dstage_getSuffix)-2) )  /* valid stages : [init ... getSuffix[ */
+    {
+        if (dctx->dStage == dstage_flushOut) {
+            size_t const preserveSize = (size_t)(dctx->tmpOut - dctx->tmpOutBuffer);
+            size_t copySize = 64 KB - dctx->tmpOutSize;
+            const BYTE* oldDictEnd = dctx->dict + dctx->dictSize - dctx->tmpOutStart;
+            if (dctx->tmpOutSize > 64 KB) copySize = 0;
+            if (copySize > preserveSize) copySize = preserveSize;
+            assert(dctx->tmpOutBuffer != NULL);
+
+            memcpy(dctx->tmpOutBuffer + preserveSize - copySize, oldDictEnd - copySize, copySize);
+
+            dctx->dict = dctx->tmpOutBuffer;
+            dctx->dictSize = preserveSize + dctx->tmpOutStart;
+        } else {
+            const BYTE* const oldDictEnd = dctx->dict + dctx->dictSize;
+            size_t const newDictSize = MIN(dctx->dictSize, 64 KB);
+
+            memcpy(dctx->tmpOutBuffer, oldDictEnd - newDictSize, newDictSize);
+
+            dctx->dict = dctx->tmpOutBuffer;
+            dctx->dictSize = newDictSize;
+            dctx->tmpOut = dctx->tmpOutBuffer + newDictSize;
+        }
+    }
+
+    *srcSizePtr = (size_t)(srcPtr - srcStart);
+    *dstSizePtr = (size_t)(dstPtr - dstStart);
+    return nextSrcSizeHint;
+}
+
+/*! LZ4F_decompress_usingDict() :
+ *  Same as LZ4F_decompress(), using a predefined dictionary.
+ *  Dictionary is used "in place", without any preprocessing.
+ *  It must remain accessible throughout the entire frame decoding.
+ */
+size_t LZ4F_decompress_usingDict(LZ4F_dctx* dctx,
+                       void* dstBuffer, size_t* dstSizePtr,
+                       const void* srcBuffer, size_t* srcSizePtr,
+                       const void* dict, size_t dictSize,
+                       const LZ4F_decompressOptions_t* decompressOptionsPtr)
+{
+    if (dctx->dStage <= dstage_init) {
+        dctx->dict = (const BYTE*)dict;
+        dctx->dictSize = dictSize;
+    }
+    return LZ4F_decompress(dctx, dstBuffer, dstSizePtr,
+                           srcBuffer, srcSizePtr,
+                           decompressOptionsPtr);
+}
diff --git a/mfbt/lz4/lz4frame.h b/mfbt/lz4/lz4frame.h
new file mode 100644
index 0000000000..4573317ef2
--- /dev/null
+++ b/mfbt/lz4/lz4frame.h
@@ -0,0 +1,623 @@
+/*
+   LZ4 auto-framing library
+   Header File
+   Copyright (C) 2011-2017, Yann Collet.
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - LZ4 source repository : https://github.com/lz4/lz4
+   - LZ4 public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+/* LZ4F is a stand-alone API able to create and decode LZ4 frames
+ * conformant with specification v1.6.1 in doc/lz4_Frame_format.md .
+ * Generated frames are compatible with `lz4` CLI.
+ *
+ * LZ4F also offers streaming capabilities.
+ *
+ * lz4.h is not required when using lz4frame.h,
+ * except to extract common constant such as LZ4_VERSION_NUMBER.
+ * */
+
+#ifndef LZ4F_H_09782039843
+#define LZ4F_H_09782039843
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* ---   Dependency   --- */
+#include <stddef.h>   /* size_t */
+
+
+/**
+  Introduction
+
+  lz4frame.h implements LZ4 frame specification (doc/lz4_Frame_format.md).
+  lz4frame.h provides frame compression functions that take care
+  of encoding standard metadata alongside LZ4-compressed blocks.
+*/
+
+/*-***************************************************************
+ *  Compiler specifics
+ *****************************************************************/
+/*  LZ4_DLL_EXPORT :
+ *  Enable exporting of functions when building a Windows DLL
+ *  LZ4FLIB_VISIBILITY :
+ *  Control library symbols visibility.
+ */
+#ifndef LZ4FLIB_VISIBILITY
+#  if defined(__GNUC__) && (__GNUC__ >= 4)
+#    define LZ4FLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#  else
+#    define LZ4FLIB_VISIBILITY
+#  endif
+#endif
+#if defined(LZ4_DLL_EXPORT) && (LZ4_DLL_EXPORT==1)
+#  define LZ4FLIB_API __declspec(dllexport) LZ4FLIB_VISIBILITY
+#elif defined(LZ4_DLL_IMPORT) && (LZ4_DLL_IMPORT==1)
+#  define LZ4FLIB_API __declspec(dllimport) LZ4FLIB_VISIBILITY
+#else
+#  define LZ4FLIB_API LZ4FLIB_VISIBILITY
+#endif
+
+#ifdef LZ4F_DISABLE_DEPRECATE_WARNINGS
+#  define LZ4F_DEPRECATE(x) x
+#else
+#  if defined(_MSC_VER)
+#    define LZ4F_DEPRECATE(x) x   /* __declspec(deprecated) x - only works with C++ */
+#  elif defined(__clang__) || (defined(__GNUC__) && (__GNUC__ >= 6))
+#    define LZ4F_DEPRECATE(x) x __attribute__((deprecated))
+#  else
+#    define LZ4F_DEPRECATE(x) x   /* no deprecation warning for this compiler */
+#  endif
+#endif
+
+
+/*-************************************
+ *  Error management
+ **************************************/
+typedef size_t LZ4F_errorCode_t;
+
+LZ4FLIB_API unsigned    LZ4F_isError(LZ4F_errorCode_t code);   /**< tells when a function result is an error code */
+LZ4FLIB_API const char* LZ4F_getErrorName(LZ4F_errorCode_t code);   /**< return error code string; for debugging */
+
+
+/*-************************************
+ *  Frame compression types
+ ************************************* */
+/* #define LZ4F_ENABLE_OBSOLETE_ENUMS   // uncomment to enable obsolete enums */
+#ifdef LZ4F_ENABLE_OBSOLETE_ENUMS
+#  define LZ4F_OBSOLETE_ENUM(x) , LZ4F_DEPRECATE(x) = LZ4F_##x
+#else
+#  define LZ4F_OBSOLETE_ENUM(x)
+#endif
+
+/* The larger the block size, the (slightly) better the compression ratio,
+ * though there are diminishing returns.
+ * Larger blocks also increase memory usage on both compression and decompression sides.
+ */
+typedef enum {
+    LZ4F_default=0,
+    LZ4F_max64KB=4,
+    LZ4F_max256KB=5,
+    LZ4F_max1MB=6,
+    LZ4F_max4MB=7
+    LZ4F_OBSOLETE_ENUM(max64KB)
+    LZ4F_OBSOLETE_ENUM(max256KB)
+    LZ4F_OBSOLETE_ENUM(max1MB)
+    LZ4F_OBSOLETE_ENUM(max4MB)
+} LZ4F_blockSizeID_t;
+
+/* Linked blocks sharply reduce inefficiencies when using small blocks,
+ * they compress better.
+ * However, some LZ4 decoders are only compatible with independent blocks */
+typedef enum {
+    LZ4F_blockLinked=0,
+    LZ4F_blockIndependent
+    LZ4F_OBSOLETE_ENUM(blockLinked)
+    LZ4F_OBSOLETE_ENUM(blockIndependent)
+} LZ4F_blockMode_t;
+
+typedef enum {
+    LZ4F_noContentChecksum=0,
+    LZ4F_contentChecksumEnabled
+    LZ4F_OBSOLETE_ENUM(noContentChecksum)
+    LZ4F_OBSOLETE_ENUM(contentChecksumEnabled)
+} LZ4F_contentChecksum_t;
+
+typedef enum {
+    LZ4F_noBlockChecksum=0,
+    LZ4F_blockChecksumEnabled
+} LZ4F_blockChecksum_t;
+
+typedef enum {
+    LZ4F_frame=0,
+    LZ4F_skippableFrame
+    LZ4F_OBSOLETE_ENUM(skippableFrame)
+} LZ4F_frameType_t;
+
+#ifdef LZ4F_ENABLE_OBSOLETE_ENUMS
+typedef LZ4F_blockSizeID_t blockSizeID_t;
+typedef LZ4F_blockMode_t blockMode_t;
+typedef LZ4F_frameType_t frameType_t;
+typedef LZ4F_contentChecksum_t contentChecksum_t;
+#endif
+
+/*! LZ4F_frameInfo_t :
+ *  makes it possible to set or read frame parameters.
+ *  Structure must be first init to 0, using memset() or LZ4F_INIT_FRAMEINFO,
+ *  setting all parameters to default.
+ *  It's then possible to update selectively some parameters */
+typedef struct {
+  LZ4F_blockSizeID_t     blockSizeID;         /* max64KB, max256KB, max1MB, max4MB; 0 == default */
+  LZ4F_blockMode_t       blockMode;           /* LZ4F_blockLinked, LZ4F_blockIndependent; 0 == default */
+  LZ4F_contentChecksum_t contentChecksumFlag; /* 1: frame terminated with 32-bit checksum of decompressed data; 0: disabled (default) */
+  LZ4F_frameType_t       frameType;           /* read-only field : LZ4F_frame or LZ4F_skippableFrame */
+  unsigned long long     contentSize;         /* Size of uncompressed content ; 0 == unknown */
+  unsigned               dictID;              /* Dictionary ID, sent by compressor to help decoder select correct dictionary; 0 == no dictID provided */
+  LZ4F_blockChecksum_t   blockChecksumFlag;   /* 1: each block followed by a checksum of block's compressed data; 0: disabled (default) */
+} LZ4F_frameInfo_t;
+
+#define LZ4F_INIT_FRAMEINFO   { LZ4F_default, LZ4F_blockLinked, LZ4F_noContentChecksum, LZ4F_frame, 0ULL, 0U, LZ4F_noBlockChecksum }    /* v1.8.3+ */
+
+/*! LZ4F_preferences_t :
+ *  makes it possible to supply advanced compression instructions to streaming interface.
+ *  Structure must be first init to 0, using memset() or LZ4F_INIT_PREFERENCES,
+ *  setting all parameters to default.
+ *  All reserved fields must be set to zero. */
+typedef struct {
+  LZ4F_frameInfo_t frameInfo;
+  int      compressionLevel;    /* 0: default (fast mode); values > LZ4HC_CLEVEL_MAX count as LZ4HC_CLEVEL_MAX; values < 0 trigger "fast acceleration" */
+  unsigned autoFlush;           /* 1: always flush; reduces usage of internal buffers */
+  unsigned favorDecSpeed;       /* 1: parser favors decompression speed vs compression ratio. Only works for high compression modes (>= LZ4HC_CLEVEL_OPT_MIN) */  /* v1.8.2+ */
+  unsigned reserved[3];         /* must be zero for forward compatibility */
+} LZ4F_preferences_t;
+
+#define LZ4F_INIT_PREFERENCES   { LZ4F_INIT_FRAMEINFO, 0, 0u, 0u, { 0u, 0u, 0u } }    /* v1.8.3+ */
+
+
+/*-*********************************
+*  Simple compression function
+***********************************/
+
+LZ4FLIB_API int LZ4F_compressionLevel_max(void);   /* v1.8.0+ */
+
+/*! LZ4F_compressFrameBound() :
+ *  Returns the maximum possible compressed size with LZ4F_compressFrame() given srcSize and preferences.
+ * `preferencesPtr` is optional. It can be replaced by NULL, in which case, the function will assume default preferences.
+ *  Note : this result is only usable with LZ4F_compressFrame().
+ *         It may also be used with LZ4F_compressUpdate() _if no flush() operation_ is performed.
+ */
+LZ4FLIB_API size_t LZ4F_compressFrameBound(size_t srcSize, const LZ4F_preferences_t* preferencesPtr);
+
+/*! LZ4F_compressFrame() :
+ *  Compress an entire srcBuffer into a valid LZ4 frame.
+ *  dstCapacity MUST be >= LZ4F_compressFrameBound(srcSize, preferencesPtr).
+ *  The LZ4F_preferences_t structure is optional : you can provide NULL as argument. All preferences will be set to default.
+ * @return : number of bytes written into dstBuffer.
+ *           or an error code if it fails (can be tested using LZ4F_isError())
+ */
+LZ4FLIB_API size_t LZ4F_compressFrame(void* dstBuffer, size_t dstCapacity,
+                                const void* srcBuffer, size_t srcSize,
+                                const LZ4F_preferences_t* preferencesPtr);
+
+
+/*-***********************************
+*  Advanced compression functions
+*************************************/
+typedef struct LZ4F_cctx_s LZ4F_cctx;   /* incomplete type */
+typedef LZ4F_cctx* LZ4F_compressionContext_t;   /* for compatibility with previous API version */
+
+typedef struct {
+  unsigned stableSrc;    /* 1 == src content will remain present on future calls to LZ4F_compress(); skip copying src content within tmp buffer */
+  unsigned reserved[3];
+} LZ4F_compressOptions_t;
+
+/*---   Resource Management   ---*/
+
+#define LZ4F_VERSION 100    /* This number can be used to check for an incompatible API breaking change */
+LZ4FLIB_API unsigned LZ4F_getVersion(void);
+
+/*! LZ4F_createCompressionContext() :
+ * The first thing to do is to create a compressionContext object, which will be used in all compression operations.
+ * This is achieved using LZ4F_createCompressionContext(), which takes as argument a version.
+ * The version provided MUST be LZ4F_VERSION. It is intended to track potential version mismatch, notably when using DLL.
+ * The function will provide a pointer to a fully allocated LZ4F_cctx object.
+ * If @return != zero, there was an error during context creation.
+ * Object can release its memory using LZ4F_freeCompressionContext();
+ */
+LZ4FLIB_API LZ4F_errorCode_t LZ4F_createCompressionContext(LZ4F_cctx** cctxPtr, unsigned version);
+LZ4FLIB_API LZ4F_errorCode_t LZ4F_freeCompressionContext(LZ4F_cctx* cctx);
+
+
+/*----    Compression    ----*/
+
+#define LZ4F_HEADER_SIZE_MIN  7   /* LZ4 Frame header size can vary, depending on selected paramaters */
+#define LZ4F_HEADER_SIZE_MAX 19
+
+/* Size in bytes of a block header in little-endian format. Highest bit indicates if block data is uncompressed */
+#define LZ4F_BLOCK_HEADER_SIZE 4
+
+/* Size in bytes of a block checksum footer in little-endian format. */
+#define LZ4F_BLOCK_CHECKSUM_SIZE 4
+
+/* Size in bytes of the content checksum. */
+#define LZ4F_CONTENT_CHECKSUM_SIZE 4
+
+/*! LZ4F_compressBegin() :
+ *  will write the frame header into dstBuffer.
+ *  dstCapacity must be >= LZ4F_HEADER_SIZE_MAX bytes.
+ * `prefsPtr` is optional : you can provide NULL as argument, all preferences will then be set to default.
+ * @return : number of bytes written into dstBuffer for the header
+ *           or an error code (which can be tested using LZ4F_isError())
+ */
+LZ4FLIB_API size_t LZ4F_compressBegin(LZ4F_cctx* cctx,
+                                      void* dstBuffer, size_t dstCapacity,
+                                      const LZ4F_preferences_t* prefsPtr);
+
+/*! LZ4F_compressBound() :
+ *  Provides minimum dstCapacity required to guarantee success of
+ *  LZ4F_compressUpdate(), given a srcSize and preferences, for a worst case scenario.
+ *  When srcSize==0, LZ4F_compressBound() provides an upper bound for LZ4F_flush() and LZ4F_compressEnd() instead.
+ *  Note that the result is only valid for a single invocation of LZ4F_compressUpdate().
+ *  When invoking LZ4F_compressUpdate() multiple times,
+ *  if the output buffer is gradually filled up instead of emptied and re-used from its start,
+ *  one must check if there is enough remaining capacity before each invocation, using LZ4F_compressBound().
+ * @return is always the same for a srcSize and prefsPtr.
+ *  prefsPtr is optional : when NULL is provided, preferences will be set to cover worst case scenario.
+ *  tech details :
+ * @return if automatic flushing is not enabled, includes the possibility that internal buffer might already be filled by up to (blockSize-1) bytes.
+ *  It also includes frame footer (ending + checksum), since it might be generated by LZ4F_compressEnd().
+ * @return doesn't include frame header, as it was already generated by LZ4F_compressBegin().
+ */
+LZ4FLIB_API size_t LZ4F_compressBound(size_t srcSize, const LZ4F_preferences_t* prefsPtr);
+
+/*! LZ4F_compressUpdate() :
+ *  LZ4F_compressUpdate() can be called repetitively to compress as much data as necessary.
+ *  Important rule: dstCapacity MUST be large enough to ensure operation success even in worst case situations.
+ *  This value is provided by LZ4F_compressBound().
+ *  If this condition is not respected, LZ4F_compress() will fail (result is an errorCode).
+ *  LZ4F_compressUpdate() doesn't guarantee error recovery.
+ *  When an error occurs, compression context must be freed or resized.
+ * `cOptPtr` is optional : NULL can be provided, in which case all options are set to default.
+ * @return : number of bytes written into `dstBuffer` (it can be zero, meaning input data was just buffered).
+ *           or an error code if it fails (which can be tested using LZ4F_isError())
+ */
+LZ4FLIB_API size_t LZ4F_compressUpdate(LZ4F_cctx* cctx,
+                                       void* dstBuffer, size_t dstCapacity,
+                                 const void* srcBuffer, size_t srcSize,
+                                 const LZ4F_compressOptions_t* cOptPtr);
+
+/*! LZ4F_flush() :
+ *  When data must be generated and sent immediately, without waiting for a block to be completely filled,
+ *  it's possible to call LZ4_flush(). It will immediately compress any data buffered within cctx.
+ * `dstCapacity` must be large enough to ensure the operation will be successful.
+ * `cOptPtr` is optional : it's possible to provide NULL, all options will be set to default.
+ * @return : nb of bytes written into dstBuffer (can be zero, when there is no data stored within cctx)
+ *           or an error code if it fails (which can be tested using LZ4F_isError())
+ *  Note : LZ4F_flush() is guaranteed to be successful when dstCapacity >= LZ4F_compressBound(0, prefsPtr).
+ */
+LZ4FLIB_API size_t LZ4F_flush(LZ4F_cctx* cctx,
+                              void* dstBuffer, size_t dstCapacity,
+                        const LZ4F_compressOptions_t* cOptPtr);
+
+/*! LZ4F_compressEnd() :
+ *  To properly finish an LZ4 frame, invoke LZ4F_compressEnd().
+ *  It will flush whatever data remained within `cctx` (like LZ4_flush())
+ *  and properly finalize the frame, with an endMark and a checksum.
+ * `cOptPtr` is optional : NULL can be provided, in which case all options will be set to default.
+ * @return : nb of bytes written into dstBuffer, necessarily >= 4 (endMark),
+ *           or an error code if it fails (which can be tested using LZ4F_isError())
+ *  Note : LZ4F_compressEnd() is guaranteed to be successful when dstCapacity >= LZ4F_compressBound(0, prefsPtr).
+ *  A successful call to LZ4F_compressEnd() makes `cctx` available again for another compression task.
+ */
+LZ4FLIB_API size_t LZ4F_compressEnd(LZ4F_cctx* cctx,
+                                    void* dstBuffer, size_t dstCapacity,
+                              const LZ4F_compressOptions_t* cOptPtr);
+
+
+/*-*********************************
+*  Decompression functions
+***********************************/
+typedef struct LZ4F_dctx_s LZ4F_dctx;   /* incomplete type */
+typedef LZ4F_dctx* LZ4F_decompressionContext_t;   /* compatibility with previous API versions */
+
+typedef struct {
+  unsigned stableDst;    /* pledges that last 64KB decompressed data will remain available unmodified. This optimization skips storage operations in tmp buffers. */
+  unsigned reserved[3];  /* must be set to zero for forward compatibility */
+} LZ4F_decompressOptions_t;
+
+
+/* Resource management */
+
+/*! LZ4F_createDecompressionContext() :
+ *  Create an LZ4F_dctx object, to track all decompression operations.
+ *  The version provided MUST be LZ4F_VERSION.
+ *  The function provides a pointer to an allocated and initialized LZ4F_dctx object.
+ *  The result is an errorCode, which can be tested using LZ4F_isError().
+ *  dctx memory can be released using LZ4F_freeDecompressionContext();
+ *  Result of LZ4F_freeDecompressionContext() indicates current state of decompressionContext when being released.
+ *  That is, it should be == 0 if decompression has been completed fully and correctly.
+ */
+LZ4FLIB_API LZ4F_errorCode_t LZ4F_createDecompressionContext(LZ4F_dctx** dctxPtr, unsigned version);
+LZ4FLIB_API LZ4F_errorCode_t LZ4F_freeDecompressionContext(LZ4F_dctx* dctx);
+
+
+/*-***********************************
+*  Streaming decompression functions
+*************************************/
+
+#define LZ4F_MIN_SIZE_TO_KNOW_HEADER_LENGTH 5
+
+/*! LZ4F_headerSize() : v1.9.0+
+ *  Provide the header size of a frame starting at `src`.
+ * `srcSize` must be >= LZ4F_MIN_SIZE_TO_KNOW_HEADER_LENGTH,
+ *  which is enough to decode the header length.
+ * @return : size of frame header
+ *           or an error code, which can be tested using LZ4F_isError()
+ *  note : Frame header size is variable, but is guaranteed to be
+ *         >= LZ4F_HEADER_SIZE_MIN bytes, and <= LZ4F_HEADER_SIZE_MAX bytes.
+ */
+LZ4FLIB_API size_t LZ4F_headerSize(const void* src, size_t srcSize);
+
+/*! LZ4F_getFrameInfo() :
+ *  This function extracts frame parameters (max blockSize, dictID, etc.).
+ *  Its usage is optional: user can call LZ4F_decompress() directly.
+ *
+ *  Extracted information will fill an existing LZ4F_frameInfo_t structure.
+ *  This can be useful for allocation and dictionary identification purposes.
+ *
+ *  LZ4F_getFrameInfo() can work in the following situations :
+ *
+ *  1) At the beginning of a new frame, before any invocation of LZ4F_decompress().
+ *     It will decode header from `srcBuffer`,
+ *     consuming the header and starting the decoding process.
+ *
+ *     Input size must be large enough to contain the full frame header.
+ *     Frame header size can be known beforehand by LZ4F_headerSize().
+ *     Frame header size is variable, but is guaranteed to be >= LZ4F_HEADER_SIZE_MIN bytes,
+ *     and not more than <= LZ4F_HEADER_SIZE_MAX bytes.
+ *     Hence, blindly providing LZ4F_HEADER_SIZE_MAX bytes or more will always work.
+ *     It's allowed to provide more input data than the header size,
+ *     LZ4F_getFrameInfo() will only consume the header.
+ *
+ *     If input size is not large enough,
+ *     aka if it's smaller than header size,
+ *     function will fail and return an error code.
+ *
+ *  2) After decoding has been started,
+ *     it's possible to invoke LZ4F_getFrameInfo() anytime
+ *     to extract already decoded frame parameters stored within dctx.
+ *
+ *     Note that, if decoding has barely started,
+ *     and not yet read enough information to decode the header,
+ *     LZ4F_getFrameInfo() will fail.
+ *
+ *  The number of bytes consumed from srcBuffer will be updated in *srcSizePtr (necessarily <= original value).
+ *  LZ4F_getFrameInfo() only consumes bytes when decoding has not yet started,
+ *  and when decoding the header has been successful.
+ *  Decompression must then resume from (srcBuffer + *srcSizePtr).
+ *
+ * @return : a hint about how many srcSize bytes LZ4F_decompress() expects for next call,
+ *           or an error code which can be tested using LZ4F_isError().
+ *  note 1 : in case of error, dctx is not modified. Decoding operation can resume from beginning safely.
+ *  note 2 : frame parameters are *copied into* an already allocated LZ4F_frameInfo_t structure.
+ */
+LZ4FLIB_API size_t LZ4F_getFrameInfo(LZ4F_dctx* dctx,
+                                     LZ4F_frameInfo_t* frameInfoPtr,
+                                     const void* srcBuffer, size_t* srcSizePtr);
+
+/*! LZ4F_decompress() :
+ *  Call this function repetitively to regenerate data compressed in `srcBuffer`.
+ *
+ *  The function requires a valid dctx state.
+ *  It will read up to *srcSizePtr bytes from srcBuffer,
+ *  and decompress data into dstBuffer, of capacity *dstSizePtr.
+ *
+ *  The nb of bytes consumed from srcBuffer will be written into *srcSizePtr (necessarily <= original value).
+ *  The nb of bytes decompressed into dstBuffer will be written into *dstSizePtr (necessarily <= original value).
+ *
+ *  The function does not necessarily read all input bytes, so always check value in *srcSizePtr.
+ *  Unconsumed source data must be presented again in subsequent invocations.
+ *
+ * `dstBuffer` can freely change between each consecutive function invocation.
+ * `dstBuffer` content will be overwritten.
+ *
+ * @return : an hint of how many `srcSize` bytes LZ4F_decompress() expects for next call.
+ *  Schematically, it's the size of the current (or remaining) compressed block + header of next block.
+ *  Respecting the hint provides some small speed benefit, because it skips intermediate buffers.
+ *  This is just a hint though, it's always possible to provide any srcSize.
+ *
+ *  When a frame is fully decoded, @return will be 0 (no more data expected).
+ *  When provided with more bytes than necessary to decode a frame,
+ *  LZ4F_decompress() will stop reading exactly at end of current frame, and @return 0.
+ *
+ *  If decompression failed, @return is an error code, which can be tested using LZ4F_isError().
+ *  After a decompression error, the `dctx` context is not resumable.
+ *  Use LZ4F_resetDecompressionContext() to return to clean state.
+ *
+ *  After a frame is fully decoded, dctx can be used again to decompress another frame.
+ */
+LZ4FLIB_API size_t LZ4F_decompress(LZ4F_dctx* dctx,
+                                   void* dstBuffer, size_t* dstSizePtr,
+                                   const void* srcBuffer, size_t* srcSizePtr,
+                                   const LZ4F_decompressOptions_t* dOptPtr);
+
+
+/*! LZ4F_resetDecompressionContext() : added in v1.8.0
+ *  In case of an error, the context is left in "undefined" state.
+ *  In which case, it's necessary to reset it, before re-using it.
+ *  This method can also be used to abruptly stop any unfinished decompression,
+ *  and start a new one using same context resources. */
+LZ4FLIB_API void LZ4F_resetDecompressionContext(LZ4F_dctx* dctx);   /* always successful */
+
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* LZ4F_H_09782039843 */
+
+#if defined(LZ4F_STATIC_LINKING_ONLY) && !defined(LZ4F_H_STATIC_09782039843)
+#define LZ4F_H_STATIC_09782039843
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* These declarations are not stable and may change in the future.
+ * They are therefore only safe to depend on
+ * when the caller is statically linked against the library.
+ * To access their declarations, define LZ4F_STATIC_LINKING_ONLY.
+ *
+ * By default, these symbols aren't published into shared/dynamic libraries.
+ * You can override this behavior and force them to be published
+ * by defining LZ4F_PUBLISH_STATIC_FUNCTIONS.
+ * Use at your own risk.
+ */
+#ifdef LZ4F_PUBLISH_STATIC_FUNCTIONS
+# define LZ4FLIB_STATIC_API LZ4FLIB_API
+#else
+# define LZ4FLIB_STATIC_API
+#endif
+
+
+/* ---   Error List   --- */
+#define LZ4F_LIST_ERRORS(ITEM) \
+        ITEM(OK_NoError) \
+        ITEM(ERROR_GENERIC) \
+        ITEM(ERROR_maxBlockSize_invalid) \
+        ITEM(ERROR_blockMode_invalid) \
+        ITEM(ERROR_contentChecksumFlag_invalid) \
+        ITEM(ERROR_compressionLevel_invalid) \
+        ITEM(ERROR_headerVersion_wrong) \
+        ITEM(ERROR_blockChecksum_invalid) \
+        ITEM(ERROR_reservedFlag_set) \
+        ITEM(ERROR_allocation_failed) \
+        ITEM(ERROR_srcSize_tooLarge) \
+        ITEM(ERROR_dstMaxSize_tooSmall) \
+        ITEM(ERROR_frameHeader_incomplete) \
+        ITEM(ERROR_frameType_unknown) \
+        ITEM(ERROR_frameSize_wrong) \
+        ITEM(ERROR_srcPtr_wrong) \
+        ITEM(ERROR_decompressionFailed) \
+        ITEM(ERROR_headerChecksum_invalid) \
+        ITEM(ERROR_contentChecksum_invalid) \
+        ITEM(ERROR_frameDecoding_alreadyStarted) \
+        ITEM(ERROR_maxCode)
+
+#define LZ4F_GENERATE_ENUM(ENUM) LZ4F_##ENUM,
+
+/* enum list is exposed, to handle specific errors */
+typedef enum { LZ4F_LIST_ERRORS(LZ4F_GENERATE_ENUM)
+              _LZ4F_dummy_error_enum_for_c89_never_used } LZ4F_errorCodes;
+
+LZ4FLIB_STATIC_API LZ4F_errorCodes LZ4F_getErrorCode(size_t functionResult);
+
+LZ4FLIB_STATIC_API size_t LZ4F_getBlockSize(unsigned);
+
+/**********************************
+ *  Bulk processing dictionary API
+ *********************************/
+
+/* A Dictionary is useful for the compression of small messages (KB range).
+ * It dramatically improves compression efficiency.
+ *
+ * LZ4 can ingest any input as dictionary, though only the last 64 KB are useful.
+ * Best results are generally achieved by using Zstandard's Dictionary Builder
+ * to generate a high-quality dictionary from a set of samples.
+ *
+ * Loading a dictionary has a cost, since it involves construction of tables.
+ * The Bulk processing dictionary API makes it possible to share this cost
+ * over an arbitrary number of compression jobs, even concurrently,
+ * markedly improving compression latency for these cases.
+ *
+ * The same dictionary will have to be used on the decompression side
+ * for decoding to be successful.
+ * To help identify the correct dictionary at decoding stage,
+ * the frame header allows optional embedding of a dictID field.
+ */
+typedef struct LZ4F_CDict_s LZ4F_CDict;
+
+/*! LZ4_createCDict() :
+ *  When compressing multiple messages / blocks using the same dictionary, it's recommended to load it just once.
+ *  LZ4_createCDict() will create a digested dictionary, ready to start future compression operations without startup delay.
+ *  LZ4_CDict can be created once and shared by multiple threads concurrently, since its usage is read-only.
+ * `dictBuffer` can be released after LZ4_CDict creation, since its content is copied within CDict */
+LZ4FLIB_STATIC_API LZ4F_CDict* LZ4F_createCDict(const void* dictBuffer, size_t dictSize);
+LZ4FLIB_STATIC_API void        LZ4F_freeCDict(LZ4F_CDict* CDict);
+
+
+/*! LZ4_compressFrame_usingCDict() :
+ *  Compress an entire srcBuffer into a valid LZ4 frame using a digested Dictionary.
+ *  cctx must point to a context created by LZ4F_createCompressionContext().
+ *  If cdict==NULL, compress without a dictionary.
+ *  dstBuffer MUST be >= LZ4F_compressFrameBound(srcSize, preferencesPtr).
+ *  If this condition is not respected, function will fail (@return an errorCode).
+ *  The LZ4F_preferences_t structure is optional : you may provide NULL as argument,
+ *  but it's not recommended, as it's the only way to provide dictID in the frame header.
+ * @return : number of bytes written into dstBuffer.
+ *           or an error code if it fails (can be tested using LZ4F_isError()) */
+LZ4FLIB_STATIC_API size_t LZ4F_compressFrame_usingCDict(
+    LZ4F_cctx* cctx,
+    void* dst, size_t dstCapacity,
+    const void* src, size_t srcSize,
+    const LZ4F_CDict* cdict,
+    const LZ4F_preferences_t* preferencesPtr);
+
+
+/*! LZ4F_compressBegin_usingCDict() :
+ *  Inits streaming dictionary compression, and writes the frame header into dstBuffer.
+ *  dstCapacity must be >= LZ4F_HEADER_SIZE_MAX bytes.
+ * `prefsPtr` is optional : you may provide NULL as argument,
+ *  however, it's the only way to provide dictID in the frame header.
+ * @return : number of bytes written into dstBuffer for the header,
+ *           or an error code (which can be tested using LZ4F_isError()) */
+LZ4FLIB_STATIC_API size_t LZ4F_compressBegin_usingCDict(
+    LZ4F_cctx* cctx,
+    void* dstBuffer, size_t dstCapacity,
+    const LZ4F_CDict* cdict,
+    const LZ4F_preferences_t* prefsPtr);
+
+
+/*! LZ4F_decompress_usingDict() :
+ *  Same as LZ4F_decompress(), using a predefined dictionary.
+ *  Dictionary is used "in place", without any preprocessing.
+ *  It must remain accessible throughout the entire frame decoding. */
+LZ4FLIB_STATIC_API size_t LZ4F_decompress_usingDict(
+    LZ4F_dctx* dctxPtr,
+    void* dstBuffer, size_t* dstSizePtr,
+    const void* srcBuffer, size_t* srcSizePtr,
+    const void* dict, size_t dictSize,
+    const LZ4F_decompressOptions_t* decompressOptionsPtr);
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif  /* defined(LZ4F_STATIC_LINKING_ONLY) && !defined(LZ4F_H_STATIC_09782039843) */
diff --git a/mfbt/lz4/lz4frame_static.h b/mfbt/lz4/lz4frame_static.h
new file mode 100644
index 0000000000..925a2c5c33
--- /dev/null
+++ b/mfbt/lz4/lz4frame_static.h
@@ -0,0 +1,47 @@
+/*
+   LZ4 auto-framing library
+   Header File for static linking only
+   Copyright (C) 2011-2016, Yann Collet.
+
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - LZ4 source repository : https://github.com/lz4/lz4
+   - LZ4 public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+
+#ifndef LZ4FRAME_STATIC_H_0398209384
+#define LZ4FRAME_STATIC_H_0398209384
+
+/* The declarations that formerly were made here have been merged into
+ * lz4frame.h, protected by the LZ4F_STATIC_LINKING_ONLY macro. Going forward,
+ * it is recommended to simply include that header directly.
+ */
+
+#define LZ4F_STATIC_LINKING_ONLY
+#include "lz4frame.h"
+
+#endif /* LZ4FRAME_STATIC_H_0398209384 */
diff --git a/mfbt/lz4/lz4hc.c b/mfbt/lz4/lz4hc.c
new file mode 100644
index 0000000000..a556d47920
--- /dev/null
+++ b/mfbt/lz4/lz4hc.c
@@ -0,0 +1,1619 @@
+/*
+    LZ4 HC - High Compression Mode of LZ4
+    Copyright (C) 2011-2017, Yann Collet.
+
+    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are
+    met:
+
+    * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+    copyright notice, this list of conditions and the following disclaimer
+    in the documentation and/or other materials provided with the
+    distribution.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+    OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+    You can contact the author at :
+       - LZ4 source repository : https://github.com/lz4/lz4
+       - LZ4 public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+/* note : lz4hc is not an independent module, it requires lz4.h/lz4.c for proper compilation */
+
+
+/* *************************************
+*  Tuning Parameter
+***************************************/
+
+/*! HEAPMODE :
+ *  Select how default compression function will allocate workplace memory,
+ *  in stack (0:fastest), or in heap (1:requires malloc()).
+ *  Since workplace is rather large, heap mode is recommended.
+ */
+#ifndef LZ4HC_HEAPMODE
+#  define LZ4HC_HEAPMODE 1
+#endif
+
+
+/*===    Dependency    ===*/
+#define LZ4_HC_STATIC_LINKING_ONLY
+#include "lz4hc.h"
+
+
+/*===   Common definitions   ===*/
+#if defined(__GNUC__)
+#  pragma GCC diagnostic ignored "-Wunused-function"
+#endif
+#if defined (__clang__)
+#  pragma clang diagnostic ignored "-Wunused-function"
+#endif
+
+#define LZ4_COMMONDEFS_ONLY
+#ifndef LZ4_SRC_INCLUDED
+#include "lz4.c"   /* LZ4_count, constants, mem */
+#endif
+
+
+/*===   Enums   ===*/
+typedef enum { noDictCtx, usingDictCtxHc } dictCtx_directive;
+
+
+/*===   Constants   ===*/
+#define OPTIMAL_ML (int)((ML_MASK-1)+MINMATCH)
+#define LZ4_OPT_NUM   (1<<12)
+
+
+/*===   Macros   ===*/
+#define MIN(a,b)   ( (a) < (b) ? (a) : (b) )
+#define MAX(a,b)   ( (a) > (b) ? (a) : (b) )
+#define HASH_FUNCTION(i)         (((i) * 2654435761U) >> ((MINMATCH*8)-LZ4HC_HASH_LOG))
+#define DELTANEXTMAXD(p)         chainTable[(p) & LZ4HC_MAXD_MASK]    /* flexible, LZ4HC_MAXD dependent */
+#define DELTANEXTU16(table, pos) table[(U16)(pos)]   /* faster */
+/* Make fields passed to, and updated by LZ4HC_encodeSequence explicit */
+#define UPDATABLE(ip, op, anchor) &ip, &op, &anchor
+
+static U32 LZ4HC_hashPtr(const void* ptr) { return HASH_FUNCTION(LZ4_read32(ptr)); }
+
+
+/**************************************
+*  HC Compression
+**************************************/
+static void LZ4HC_clearTables (LZ4HC_CCtx_internal* hc4)
+{
+    MEM_INIT(hc4->hashTable, 0, sizeof(hc4->hashTable));
+    MEM_INIT(hc4->chainTable, 0xFF, sizeof(hc4->chainTable));
+}
+
+static void LZ4HC_init_internal (LZ4HC_CCtx_internal* hc4, const BYTE* start)
+{
+    uptrval startingOffset = (uptrval)(hc4->end - hc4->base);
+    if (startingOffset > 1 GB) {
+        LZ4HC_clearTables(hc4);
+        startingOffset = 0;
+    }
+    startingOffset += 64 KB;
+    hc4->nextToUpdate = (U32) startingOffset;
+    hc4->base = start - startingOffset;
+    hc4->end = start;
+    hc4->dictBase = start - startingOffset;
+    hc4->dictLimit = (U32) startingOffset;
+    hc4->lowLimit = (U32) startingOffset;
+}
+
+
+/* Update chains up to ip (excluded) */
+LZ4_FORCE_INLINE void LZ4HC_Insert (LZ4HC_CCtx_internal* hc4, const BYTE* ip)
+{
+    U16* const chainTable = hc4->chainTable;
+    U32* const hashTable  = hc4->hashTable;
+    const BYTE* const base = hc4->base;
+    U32 const target = (U32)(ip - base);
+    U32 idx = hc4->nextToUpdate;
+
+    while (idx < target) {
+        U32 const h = LZ4HC_hashPtr(base+idx);
+        size_t delta = idx - hashTable[h];
+        if (delta>LZ4_DISTANCE_MAX) delta = LZ4_DISTANCE_MAX;
+        DELTANEXTU16(chainTable, idx) = (U16)delta;
+        hashTable[h] = idx;
+        idx++;
+    }
+
+    hc4->nextToUpdate = target;
+}
+
+/** LZ4HC_countBack() :
+ * @return : negative value, nb of common bytes before ip/match */
+LZ4_FORCE_INLINE
+int LZ4HC_countBack(const BYTE* const ip, const BYTE* const match,
+                    const BYTE* const iMin, const BYTE* const mMin)
+{
+    int back = 0;
+    int const min = (int)MAX(iMin - ip, mMin - match);
+    assert(min <= 0);
+    assert(ip >= iMin); assert((size_t)(ip-iMin) < (1U<<31));
+    assert(match >= mMin); assert((size_t)(match - mMin) < (1U<<31));
+    while ( (back > min)
+         && (ip[back-1] == match[back-1]) )
+            back--;
+    return back;
+}
+
+#if defined(_MSC_VER)
+#  define LZ4HC_rotl32(x,r) _rotl(x,r)
+#else
+#  define LZ4HC_rotl32(x,r) ((x << r) | (x >> (32 - r)))
+#endif
+
+
+static U32 LZ4HC_rotatePattern(size_t const rotate, U32 const pattern)
+{
+    size_t const bitsToRotate = (rotate & (sizeof(pattern) - 1)) << 3;
+    if (bitsToRotate == 0) return pattern;
+    return LZ4HC_rotl32(pattern, (int)bitsToRotate);
+}
+
+/* LZ4HC_countPattern() :
+ * pattern32 must be a sample of repetitive pattern of length 1, 2 or 4 (but not 3!) */
+static unsigned
+LZ4HC_countPattern(const BYTE* ip, const BYTE* const iEnd, U32 const pattern32)
+{
+    const BYTE* const iStart = ip;
+    reg_t const pattern = (sizeof(pattern)==8) ?
+        (reg_t)pattern32 + (((reg_t)pattern32) << (sizeof(pattern)*4)) : pattern32;
+
+    while (likely(ip < iEnd-(sizeof(pattern)-1))) {
+        reg_t const diff = LZ4_read_ARCH(ip) ^ pattern;
+        if (!diff) { ip+=sizeof(pattern); continue; }
+        ip += LZ4_NbCommonBytes(diff);
+        return (unsigned)(ip - iStart);
+    }
+
+    if (LZ4_isLittleEndian()) {
+        reg_t patternByte = pattern;
+        while ((ip<iEnd) && (*ip == (BYTE)patternByte)) {
+            ip++; patternByte >>= 8;
+        }
+    } else {  /* big endian */
+        U32 bitOffset = (sizeof(pattern)*8) - 8;
+        while (ip < iEnd) {
+            BYTE const byte = (BYTE)(pattern >> bitOffset);
+            if (*ip != byte) break;
+            ip ++; bitOffset -= 8;
+        }
+    }
+
+    return (unsigned)(ip - iStart);
+}
+
+/* LZ4HC_reverseCountPattern() :
+ * pattern must be a sample of repetitive pattern of length 1, 2 or 4 (but not 3!)
+ * read using natural platform endianess */
+static unsigned
+LZ4HC_reverseCountPattern(const BYTE* ip, const BYTE* const iLow, U32 pattern)
+{
+    const BYTE* const iStart = ip;
+
+    while (likely(ip >= iLow+4)) {
+        if (LZ4_read32(ip-4) != pattern) break;
+        ip -= 4;
+    }
+    {   const BYTE* bytePtr = (const BYTE*)(&pattern) + 3; /* works for any endianess */
+        while (likely(ip>iLow)) {
+            if (ip[-1] != *bytePtr) break;
+            ip--; bytePtr--;
+    }   }
+    return (unsigned)(iStart - ip);
+}
+
+/* LZ4HC_protectDictEnd() :
+ * Checks if the match is in the last 3 bytes of the dictionary, so reading the
+ * 4 byte MINMATCH would overflow.
+ * @returns true if the match index is okay.
+ */
+static int LZ4HC_protectDictEnd(U32 const dictLimit, U32 const matchIndex)
+{
+    return ((U32)((dictLimit - 1) - matchIndex) >= 3);
+}
+
+typedef enum { rep_untested, rep_not, rep_confirmed } repeat_state_e;
+typedef enum { favorCompressionRatio=0, favorDecompressionSpeed } HCfavor_e;
+
+LZ4_FORCE_INLINE int
+LZ4HC_InsertAndGetWiderMatch (
+    LZ4HC_CCtx_internal* hc4,
+    const BYTE* const ip,
+    const BYTE* const iLowLimit,
+    const BYTE* const iHighLimit,
+    int longest,
+    const BYTE** matchpos,
+    const BYTE** startpos,
+    const int maxNbAttempts,
+    const int patternAnalysis,
+    const int chainSwap,
+    const dictCtx_directive dict,
+    const HCfavor_e favorDecSpeed)
+{
+    U16* const chainTable = hc4->chainTable;
+    U32* const HashTable = hc4->hashTable;
+    const LZ4HC_CCtx_internal * const dictCtx = hc4->dictCtx;
+    const BYTE* const base = hc4->base;
+    const U32 dictLimit = hc4->dictLimit;
+    const BYTE* const lowPrefixPtr = base + dictLimit;
+    const U32 ipIndex = (U32)(ip - base);
+    const U32 lowestMatchIndex = (hc4->lowLimit + (LZ4_DISTANCE_MAX + 1) > ipIndex) ? hc4->lowLimit : ipIndex - LZ4_DISTANCE_MAX;
+    const BYTE* const dictBase = hc4->dictBase;
+    int const lookBackLength = (int)(ip-iLowLimit);
+    int nbAttempts = maxNbAttempts;
+    U32 matchChainPos = 0;
+    U32 const pattern = LZ4_read32(ip);
+    U32 matchIndex;
+    repeat_state_e repeat = rep_untested;
+    size_t srcPatternLength = 0;
+
+    DEBUGLOG(7, "LZ4HC_InsertAndGetWiderMatch");
+    /* First Match */
+    LZ4HC_Insert(hc4, ip);
+    matchIndex = HashTable[LZ4HC_hashPtr(ip)];
+    DEBUGLOG(7, "First match at index %u / %u (lowestMatchIndex)",
+                matchIndex, lowestMatchIndex);
+
+    while ((matchIndex>=lowestMatchIndex) && (nbAttempts>0)) {
+        int matchLength=0;
+        nbAttempts--;
+        assert(matchIndex < ipIndex);
+        if (favorDecSpeed && (ipIndex - matchIndex < 8)) {
+            /* do nothing */
+        } else if (matchIndex >= dictLimit) {   /* within current Prefix */
+            const BYTE* const matchPtr = base + matchIndex;
+            assert(matchPtr >= lowPrefixPtr);
+            assert(matchPtr < ip);
+            assert(longest >= 1);
+            if (LZ4_read16(iLowLimit + longest - 1) == LZ4_read16(matchPtr - lookBackLength + longest - 1)) {
+                if (LZ4_read32(matchPtr) == pattern) {
+                    int const back = lookBackLength ? LZ4HC_countBack(ip, matchPtr, iLowLimit, lowPrefixPtr) : 0;
+                    matchLength = MINMATCH + (int)LZ4_count(ip+MINMATCH, matchPtr+MINMATCH, iHighLimit);
+                    matchLength -= back;
+                    if (matchLength > longest) {
+                        longest = matchLength;
+                        *matchpos = matchPtr + back;
+                        *startpos = ip + back;
+            }   }   }
+        } else {   /* lowestMatchIndex <= matchIndex < dictLimit */
+            const BYTE* const matchPtr = dictBase + matchIndex;
+            if (LZ4_read32(matchPtr) == pattern) {
+                const BYTE* const dictStart = dictBase + hc4->lowLimit;
+                int back = 0;
+                const BYTE* vLimit = ip + (dictLimit - matchIndex);
+                if (vLimit > iHighLimit) vLimit = iHighLimit;
+                matchLength = (int)LZ4_count(ip+MINMATCH, matchPtr+MINMATCH, vLimit) + MINMATCH;
+                if ((ip+matchLength == vLimit) && (vLimit < iHighLimit))
+                    matchLength += LZ4_count(ip+matchLength, lowPrefixPtr, iHighLimit);
+                back = lookBackLength ? LZ4HC_countBack(ip, matchPtr, iLowLimit, dictStart) : 0;
+                matchLength -= back;
+                if (matchLength > longest) {
+                    longest = matchLength;
+                    *matchpos = base + matchIndex + back;   /* virtual pos, relative to ip, to retrieve offset */
+                    *startpos = ip + back;
+        }   }   }
+
+        if (chainSwap && matchLength==longest) {    /* better match => select a better chain */
+            assert(lookBackLength==0);   /* search forward only */
+            if (matchIndex + (U32)longest <= ipIndex) {
+                int const kTrigger = 4;
+                U32 distanceToNextMatch = 1;
+                int const end = longest - MINMATCH + 1;
+                int step = 1;
+                int accel = 1 << kTrigger;
+                int pos;
+                for (pos = 0; pos < end; pos += step) {
+                    U32 const candidateDist = DELTANEXTU16(chainTable, matchIndex + (U32)pos);
+                    step = (accel++ >> kTrigger);
+                    if (candidateDist > distanceToNextMatch) {
+                        distanceToNextMatch = candidateDist;
+                        matchChainPos = (U32)pos;
+                        accel = 1 << kTrigger;
+                    }
+                }
+                if (distanceToNextMatch > 1) {
+                    if (distanceToNextMatch > matchIndex) break;   /* avoid overflow */
+                    matchIndex -= distanceToNextMatch;
+                    continue;
+        }   }   }
+
+        {   U32 const distNextMatch = DELTANEXTU16(chainTable, matchIndex);
+            if (patternAnalysis && distNextMatch==1 && matchChainPos==0) {
+                U32 const matchCandidateIdx = matchIndex-1;
+                /* may be a repeated pattern */
+                if (repeat == rep_untested) {
+                    if ( ((pattern & 0xFFFF) == (pattern >> 16))
+                      &  ((pattern & 0xFF)   == (pattern >> 24)) ) {
+                        repeat = rep_confirmed;
+                        srcPatternLength = LZ4HC_countPattern(ip+sizeof(pattern), iHighLimit, pattern) + sizeof(pattern);
+                    } else {
+                        repeat = rep_not;
+                }   }
+                if ( (repeat == rep_confirmed) && (matchCandidateIdx >= lowestMatchIndex)
+                  && LZ4HC_protectDictEnd(dictLimit, matchCandidateIdx) ) {
+                    const int extDict = matchCandidateIdx < dictLimit;
+                    const BYTE* const matchPtr = (extDict ? dictBase : base) + matchCandidateIdx;
+                    if (LZ4_read32(matchPtr) == pattern) {  /* good candidate */
+                        const BYTE* const dictStart = dictBase + hc4->lowLimit;
+                        const BYTE* const iLimit = extDict ? dictBase + dictLimit : iHighLimit;
+                        size_t forwardPatternLength = LZ4HC_countPattern(matchPtr+sizeof(pattern), iLimit, pattern) + sizeof(pattern);
+                        if (extDict && matchPtr + forwardPatternLength == iLimit) {
+                            U32 const rotatedPattern = LZ4HC_rotatePattern(forwardPatternLength, pattern);
+                            forwardPatternLength += LZ4HC_countPattern(lowPrefixPtr, iHighLimit, rotatedPattern);
+                        }
+                        {   const BYTE* const lowestMatchPtr = extDict ? dictStart : lowPrefixPtr;
+                            size_t backLength = LZ4HC_reverseCountPattern(matchPtr, lowestMatchPtr, pattern);
+                            size_t currentSegmentLength;
+                            if (!extDict && matchPtr - backLength == lowPrefixPtr && hc4->lowLimit < dictLimit) {
+                                U32 const rotatedPattern = LZ4HC_rotatePattern((U32)(-(int)backLength), pattern);
+                                backLength += LZ4HC_reverseCountPattern(dictBase + dictLimit, dictStart, rotatedPattern);
+                            }
+                            /* Limit backLength not go further than lowestMatchIndex */
+                            backLength = matchCandidateIdx - MAX(matchCandidateIdx - (U32)backLength, lowestMatchIndex);
+                            assert(matchCandidateIdx - backLength >= lowestMatchIndex);
+                            currentSegmentLength = backLength + forwardPatternLength;
+                            /* Adjust to end of pattern if the source pattern fits, otherwise the beginning of the pattern */
+                            if ( (currentSegmentLength >= srcPatternLength)   /* current pattern segment large enough to contain full srcPatternLength */
+                              && (forwardPatternLength <= srcPatternLength) ) { /* haven't reached this position yet */
+                                U32 const newMatchIndex = matchCandidateIdx + (U32)forwardPatternLength - (U32)srcPatternLength;  /* best position, full pattern, might be followed by more match */
+                                if (LZ4HC_protectDictEnd(dictLimit, newMatchIndex))
+                                    matchIndex = newMatchIndex;
+                                else {
+                                    /* Can only happen if started in the prefix */
+                                    assert(newMatchIndex >= dictLimit - 3 && newMatchIndex < dictLimit && !extDict);
+                                    matchIndex = dictLimit;
+                                }
+                            } else {
+                                U32 const newMatchIndex = matchCandidateIdx - (U32)backLength;   /* farthest position in current segment, will find a match of length currentSegmentLength + maybe some back */
+                                if (!LZ4HC_protectDictEnd(dictLimit, newMatchIndex)) {
+                                    assert(newMatchIndex >= dictLimit - 3 && newMatchIndex < dictLimit && !extDict);
+                                    matchIndex = dictLimit;
+                                } else {
+                                    matchIndex = newMatchIndex;
+                                    if (lookBackLength==0) {  /* no back possible */
+                                        size_t const maxML = MIN(currentSegmentLength, srcPatternLength);
+                                        if ((size_t)longest < maxML) {
+                                            assert(base + matchIndex != ip);
+                                            if ((size_t)(ip - base) - matchIndex > LZ4_DISTANCE_MAX) break;
+                                            assert(maxML < 2 GB);
+                                            longest = (int)maxML;
+                                            *matchpos = base + matchIndex;   /* virtual pos, relative to ip, to retrieve offset */
+                                            *startpos = ip;
+                                        }
+                                        {   U32 const distToNextPattern = DELTANEXTU16(chainTable, matchIndex);
+                                            if (distToNextPattern > matchIndex) break;  /* avoid overflow */
+                                            matchIndex -= distToNextPattern;
+                        }   }   }   }   }
+                        continue;
+                }   }
+        }   }   /* PA optimization */
+
+        /* follow current chain */
+        matchIndex -= DELTANEXTU16(chainTable, matchIndex + matchChainPos);
+
+    }  /* while ((matchIndex>=lowestMatchIndex) && (nbAttempts)) */
+
+    if ( dict == usingDictCtxHc
+      && nbAttempts > 0
+      && ipIndex - lowestMatchIndex < LZ4_DISTANCE_MAX) {
+        size_t const dictEndOffset = (size_t)(dictCtx->end - dictCtx->base);
+        U32 dictMatchIndex = dictCtx->hashTable[LZ4HC_hashPtr(ip)];
+        assert(dictEndOffset <= 1 GB);
+        matchIndex = dictMatchIndex + lowestMatchIndex - (U32)dictEndOffset;
+        while (ipIndex - matchIndex <= LZ4_DISTANCE_MAX && nbAttempts--) {
+            const BYTE* const matchPtr = dictCtx->base + dictMatchIndex;
+
+            if (LZ4_read32(matchPtr) == pattern) {
+                int mlt;
+                int back = 0;
+                const BYTE* vLimit = ip + (dictEndOffset - dictMatchIndex);
+                if (vLimit > iHighLimit) vLimit = iHighLimit;
+                mlt = (int)LZ4_count(ip+MINMATCH, matchPtr+MINMATCH, vLimit) + MINMATCH;
+                back = lookBackLength ? LZ4HC_countBack(ip, matchPtr, iLowLimit, dictCtx->base + dictCtx->dictLimit) : 0;
+                mlt -= back;
+                if (mlt > longest) {
+                    longest = mlt;
+                    *matchpos = base + matchIndex + back;
+                    *startpos = ip + back;
+            }   }
+
+            {   U32 const nextOffset = DELTANEXTU16(dictCtx->chainTable, dictMatchIndex);
+                dictMatchIndex -= nextOffset;
+                matchIndex -= nextOffset;
+    }   }   }
+
+    return longest;
+}
+
+LZ4_FORCE_INLINE
+int LZ4HC_InsertAndFindBestMatch(LZ4HC_CCtx_internal* const hc4,   /* Index table will be updated */
+                                 const BYTE* const ip, const BYTE* const iLimit,
+                                 const BYTE** matchpos,
+                                 const int maxNbAttempts,
+                                 const int patternAnalysis,
+                                 const dictCtx_directive dict)
+{
+    const BYTE* uselessPtr = ip;
+    /* note : LZ4HC_InsertAndGetWiderMatch() is able to modify the starting position of a match (*startpos),
+     * but this won't be the case here, as we define iLowLimit==ip,
+     * so LZ4HC_InsertAndGetWiderMatch() won't be allowed to search past ip */
+    return LZ4HC_InsertAndGetWiderMatch(hc4, ip, ip, iLimit, MINMATCH-1, matchpos, &uselessPtr, maxNbAttempts, patternAnalysis, 0 /*chainSwap*/, dict, favorCompressionRatio);
+}
+
+/* LZ4HC_encodeSequence() :
+ * @return : 0 if ok,
+ *           1 if buffer issue detected */
+LZ4_FORCE_INLINE int LZ4HC_encodeSequence (
+    const BYTE** _ip,
+    BYTE** _op,
+    const BYTE** _anchor,
+    int matchLength,
+    const BYTE* const match,
+    limitedOutput_directive limit,
+    BYTE* oend)
+{
+#define ip      (*_ip)
+#define op      (*_op)
+#define anchor  (*_anchor)
+
+    size_t length;
+    BYTE* const token = op++;
+
+#if defined(LZ4_DEBUG) && (LZ4_DEBUG >= 6)
+    static const BYTE* start = NULL;
+    static U32 totalCost = 0;
+    U32 const pos = (start==NULL) ? 0 : (U32)(anchor - start);
+    U32 const ll = (U32)(ip - anchor);
+    U32 const llAdd = (ll>=15) ? ((ll-15) / 255) + 1 : 0;
+    U32 const mlAdd = (matchLength>=19) ? ((matchLength-19) / 255) + 1 : 0;
+    U32 const cost = 1 + llAdd + ll + 2 + mlAdd;
+    if (start==NULL) start = anchor;  /* only works for single segment */
+    /* g_debuglog_enable = (pos >= 2228) & (pos <= 2262); */
+    DEBUGLOG(6, "pos:%7u -- literals:%4u, match:%4i, offset:%5u, cost:%4u + %5u",
+                pos,
+                (U32)(ip - anchor), matchLength, (U32)(ip-match),
+                cost, totalCost);
+    totalCost += cost;
+#endif
+
+    /* Encode Literal length */
+    length = (size_t)(ip - anchor);
+    LZ4_STATIC_ASSERT(notLimited == 0);
+    /* Check output limit */
+    if (limit && ((op + (length / 255) + length + (2 + 1 + LASTLITERALS)) > oend)) {
+        DEBUGLOG(6, "Not enough room to write %i literals (%i bytes remaining)",
+                (int)length, (int)(oend - op));
+        return 1;
+    }
+    if (length >= RUN_MASK) {
+        size_t len = length - RUN_MASK;
+        *token = (RUN_MASK << ML_BITS);
+        for(; len >= 255 ; len -= 255) *op++ = 255;
+        *op++ = (BYTE)len;
+    } else {
+        *token = (BYTE)(length << ML_BITS);
+    }
+
+    /* Copy Literals */
+    LZ4_wildCopy8(op, anchor, op + length);
+    op += length;
+
+    /* Encode Offset */
+    assert( (ip - match) <= LZ4_DISTANCE_MAX );   /* note : consider providing offset as a value, rather than as a pointer difference */
+    LZ4_writeLE16(op, (U16)(ip - match)); op += 2;
+
+    /* Encode MatchLength */
+    assert(matchLength >= MINMATCH);
+    length = (size_t)matchLength - MINMATCH;
+    if (limit && (op + (length / 255) + (1 + LASTLITERALS) > oend)) {
+        DEBUGLOG(6, "Not enough room to write match length");
+        return 1;   /* Check output limit */
+    }
+    if (length >= ML_MASK) {
+        *token += ML_MASK;
+        length -= ML_MASK;
+        for(; length >= 510 ; length -= 510) { *op++ = 255; *op++ = 255; }
+        if (length >= 255) { length -= 255; *op++ = 255; }
+        *op++ = (BYTE)length;
+    } else {
+        *token += (BYTE)(length);
+    }
+
+    /* Prepare next loop */
+    ip += matchLength;
+    anchor = ip;
+
+    return 0;
+}
+#undef ip
+#undef op
+#undef anchor
+
+LZ4_FORCE_INLINE int LZ4HC_compress_hashChain (
+    LZ4HC_CCtx_internal* const ctx,
+    const char* const source,
+    char* const dest,
+    int* srcSizePtr,
+    int const maxOutputSize,
+    int maxNbAttempts,
+    const limitedOutput_directive limit,
+    const dictCtx_directive dict
+    )
+{
+    const int inputSize = *srcSizePtr;
+    const int patternAnalysis = (maxNbAttempts > 128);   /* levels 9+ */
+
+    const BYTE* ip = (const BYTE*) source;
+    const BYTE* anchor = ip;
+    const BYTE* const iend = ip + inputSize;
+    const BYTE* const mflimit = iend - MFLIMIT;
+    const BYTE* const matchlimit = (iend - LASTLITERALS);
+
+    BYTE* optr = (BYTE*) dest;
+    BYTE* op = (BYTE*) dest;
+    BYTE* oend = op + maxOutputSize;
+
+    int   ml0, ml, ml2, ml3;
+    const BYTE* start0;
+    const BYTE* ref0;
+    const BYTE* ref = NULL;
+    const BYTE* start2 = NULL;
+    const BYTE* ref2 = NULL;
+    const BYTE* start3 = NULL;
+    const BYTE* ref3 = NULL;
+
+    /* init */
+    *srcSizePtr = 0;
+    if (limit == fillOutput) oend -= LASTLITERALS;                  /* Hack for support LZ4 format restriction */
+    if (inputSize < LZ4_minLength) goto _last_literals;             /* Input too small, no compression (all literals) */
+
+    /* Main Loop */
+    while (ip <= mflimit) {
+        ml = LZ4HC_InsertAndFindBestMatch(ctx, ip, matchlimit, &ref, maxNbAttempts, patternAnalysis, dict);
+        if (ml<MINMATCH) { ip++; continue; }
+
+        /* saved, in case we would skip too much */
+        start0 = ip; ref0 = ref; ml0 = ml;
+
+_Search2:
+        if (ip+ml <= mflimit) {
+            ml2 = LZ4HC_InsertAndGetWiderMatch(ctx,
+                            ip + ml - 2, ip + 0, matchlimit, ml, &ref2, &start2,
+                            maxNbAttempts, patternAnalysis, 0, dict, favorCompressionRatio);
+        } else {
+            ml2 = ml;
+        }
+
+        if (ml2 == ml) { /* No better match => encode ML1 */
+            optr = op;
+            if (LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), ml, ref, limit, oend)) goto _dest_overflow;
+            continue;
+        }
+
+        if (start0 < ip) {   /* first match was skipped at least once */
+            if (start2 < ip + ml0) {  /* squeezing ML1 between ML0(original ML1) and ML2 */
+                ip = start0; ref = ref0; ml = ml0;  /* restore initial ML1 */
+        }   }
+
+        /* Here, start0==ip */
+        if ((start2 - ip) < 3) {  /* First Match too small : removed */
+            ml = ml2;
+            ip = start2;
+            ref =ref2;
+            goto _Search2;
+        }
+
+_Search3:
+        /* At this stage, we have :
+        *  ml2 > ml1, and
+        *  ip1+3 <= ip2 (usually < ip1+ml1) */
+        if ((start2 - ip) < OPTIMAL_ML) {
+            int correction;
+            int new_ml = ml;
+            if (new_ml > OPTIMAL_ML) new_ml = OPTIMAL_ML;
+            if (ip+new_ml > start2 + ml2 - MINMATCH) new_ml = (int)(start2 - ip) + ml2 - MINMATCH;
+            correction = new_ml - (int)(start2 - ip);
+            if (correction > 0) {
+                start2 += correction;
+                ref2 += correction;
+                ml2 -= correction;
+            }
+        }
+        /* Now, we have start2 = ip+new_ml, with new_ml = min(ml, OPTIMAL_ML=18) */
+
+        if (start2 + ml2 <= mflimit) {
+            ml3 = LZ4HC_InsertAndGetWiderMatch(ctx,
+                            start2 + ml2 - 3, start2, matchlimit, ml2, &ref3, &start3,
+                            maxNbAttempts, patternAnalysis, 0, dict, favorCompressionRatio);
+        } else {
+            ml3 = ml2;
+        }
+
+        if (ml3 == ml2) {  /* No better match => encode ML1 and ML2 */
+            /* ip & ref are known; Now for ml */
+            if (start2 < ip+ml)  ml = (int)(start2 - ip);
+            /* Now, encode 2 sequences */
+            optr = op;
+            if (LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), ml, ref, limit, oend)) goto _dest_overflow;
+            ip = start2;
+            optr = op;
+            if (LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), ml2, ref2, limit, oend)) {
+                ml  = ml2;
+                ref = ref2;
+                goto _dest_overflow;
+            }
+            continue;
+        }
+
+        if (start3 < ip+ml+3) {  /* Not enough space for match 2 : remove it */
+            if (start3 >= (ip+ml)) {  /* can write Seq1 immediately ==> Seq2 is removed, so Seq3 becomes Seq1 */
+                if (start2 < ip+ml) {
+                    int correction = (int)(ip+ml - start2);
+                    start2 += correction;
+                    ref2 += correction;
+                    ml2 -= correction;
+                    if (ml2 < MINMATCH) {
+                        start2 = start3;
+                        ref2 = ref3;
+                        ml2 = ml3;
+                    }
+                }
+
+                optr = op;
+                if (LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), ml, ref, limit, oend)) goto _dest_overflow;
+                ip  = start3;
+                ref = ref3;
+                ml  = ml3;
+
+                start0 = start2;
+                ref0 = ref2;
+                ml0 = ml2;
+                goto _Search2;
+            }
+
+            start2 = start3;
+            ref2 = ref3;
+            ml2 = ml3;
+            goto _Search3;
+        }
+
+        /*
+        * OK, now we have 3 ascending matches;
+        * let's write the first one ML1.
+        * ip & ref are known; Now decide ml.
+        */
+        if (start2 < ip+ml) {
+            if ((start2 - ip) < OPTIMAL_ML) {
+                int correction;
+                if (ml > OPTIMAL_ML) ml = OPTIMAL_ML;
+                if (ip + ml > start2 + ml2 - MINMATCH) ml = (int)(start2 - ip) + ml2 - MINMATCH;
+                correction = ml - (int)(start2 - ip);
+                if (correction > 0) {
+                    start2 += correction;
+                    ref2 += correction;
+                    ml2 -= correction;
+                }
+            } else {
+                ml = (int)(start2 - ip);
+            }
+        }
+        optr = op;
+        if (LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), ml, ref, limit, oend)) goto _dest_overflow;
+
+        /* ML2 becomes ML1 */
+        ip = start2; ref = ref2; ml = ml2;
+
+        /* ML3 becomes ML2 */
+        start2 = start3; ref2 = ref3; ml2 = ml3;
+
+        /* let's find a new ML3 */
+        goto _Search3;
+    }
+
+_last_literals:
+    /* Encode Last Literals */
+    {   size_t lastRunSize = (size_t)(iend - anchor);  /* literals */
+        size_t llAdd = (lastRunSize + 255 - RUN_MASK) / 255;
+        size_t const totalSize = 1 + llAdd + lastRunSize;
+        if (limit == fillOutput) oend += LASTLITERALS;  /* restore correct value */
+        if (limit && (op + totalSize > oend)) {
+            if (limit == limitedOutput) return 0;
+            /* adapt lastRunSize to fill 'dest' */
+            lastRunSize  = (size_t)(oend - op) - 1 /*token*/;
+            llAdd = (lastRunSize + 256 - RUN_MASK) / 256;
+            lastRunSize -= llAdd;
+        }
+        DEBUGLOG(6, "Final literal run : %i literals", (int)lastRunSize);
+        ip = anchor + lastRunSize;  /* can be != iend if limit==fillOutput */
+
+        if (lastRunSize >= RUN_MASK) {
+            size_t accumulator = lastRunSize - RUN_MASK;
+            *op++ = (RUN_MASK << ML_BITS);
+            for(; accumulator >= 255 ; accumulator -= 255) *op++ = 255;
+            *op++ = (BYTE) accumulator;
+        } else {
+            *op++ = (BYTE)(lastRunSize << ML_BITS);
+        }
+        memcpy(op, anchor, lastRunSize);
+        op += lastRunSize;
+    }
+
+    /* End */
+    *srcSizePtr = (int) (((const char*)ip) - source);
+    return (int) (((char*)op)-dest);
+
+_dest_overflow:
+    if (limit == fillOutput) {
+        /* Assumption : ip, anchor, ml and ref must be set correctly */
+        size_t const ll = (size_t)(ip - anchor);
+        size_t const ll_addbytes = (ll + 240) / 255;
+        size_t const ll_totalCost = 1 + ll_addbytes + ll;
+        BYTE* const maxLitPos = oend - 3; /* 2 for offset, 1 for token */
+        DEBUGLOG(6, "Last sequence overflowing");
+        op = optr;  /* restore correct out pointer */
+        if (op + ll_totalCost <= maxLitPos) {
+            /* ll validated; now adjust match length */
+            size_t const bytesLeftForMl = (size_t)(maxLitPos - (op+ll_totalCost));
+            size_t const maxMlSize = MINMATCH + (ML_MASK-1) + (bytesLeftForMl * 255);
+            assert(maxMlSize < INT_MAX); assert(ml >= 0);
+            if ((size_t)ml > maxMlSize) ml = (int)maxMlSize;
+            if ((oend + LASTLITERALS) - (op + ll_totalCost + 2) - 1 + ml >= MFLIMIT) {
+                LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), ml, ref, notLimited, oend);
+        }   }
+        goto _last_literals;
+    }
+    /* compression failed */
+    return 0;
+}
+
+
+static int LZ4HC_compress_optimal( LZ4HC_CCtx_internal* ctx,
+    const char* const source, char* dst,
+    int* srcSizePtr, int dstCapacity,
+    int const nbSearches, size_t sufficient_len,
+    const limitedOutput_directive limit, int const fullUpdate,
+    const dictCtx_directive dict,
+    const HCfavor_e favorDecSpeed);
+
+
+LZ4_FORCE_INLINE int LZ4HC_compress_generic_internal (
+    LZ4HC_CCtx_internal* const ctx,
+    const char* const src,
+    char* const dst,
+    int* const srcSizePtr,
+    int const dstCapacity,
+    int cLevel,
+    const limitedOutput_directive limit,
+    const dictCtx_directive dict
+    )
+{
+    typedef enum { lz4hc, lz4opt } lz4hc_strat_e;
+    typedef struct {
+        lz4hc_strat_e strat;
+        int nbSearches;
+        U32 targetLength;
+    } cParams_t;
+    static const cParams_t clTable[LZ4HC_CLEVEL_MAX+1] = {
+        { lz4hc,     2, 16 },  /* 0, unused */
+        { lz4hc,     2, 16 },  /* 1, unused */
+        { lz4hc,     2, 16 },  /* 2, unused */
+        { lz4hc,     4, 16 },  /* 3 */
+        { lz4hc,     8, 16 },  /* 4 */
+        { lz4hc,    16, 16 },  /* 5 */
+        { lz4hc,    32, 16 },  /* 6 */
+        { lz4hc,    64, 16 },  /* 7 */
+        { lz4hc,   128, 16 },  /* 8 */
+        { lz4hc,   256, 16 },  /* 9 */
+        { lz4opt,   96, 64 },  /*10==LZ4HC_CLEVEL_OPT_MIN*/
+        { lz4opt,  512,128 },  /*11 */
+        { lz4opt,16384,LZ4_OPT_NUM },  /* 12==LZ4HC_CLEVEL_MAX */
+    };
+
+    DEBUGLOG(4, "LZ4HC_compress_generic(ctx=%p, src=%p, srcSize=%d, limit=%d)",
+                ctx, src, *srcSizePtr, limit);
+
+    if (limit == fillOutput && dstCapacity < 1) return 0;   /* Impossible to store anything */
+    if ((U32)*srcSizePtr > (U32)LZ4_MAX_INPUT_SIZE) return 0;    /* Unsupported input size (too large or negative) */
+
+    ctx->end += *srcSizePtr;
+    if (cLevel < 1) cLevel = LZ4HC_CLEVEL_DEFAULT;   /* note : convention is different from lz4frame, maybe something to review */
+    cLevel = MIN(LZ4HC_CLEVEL_MAX, cLevel);
+    {   cParams_t const cParam = clTable[cLevel];
+        HCfavor_e const favor = ctx->favorDecSpeed ? favorDecompressionSpeed : favorCompressionRatio;
+        int result;
+
+        if (cParam.strat == lz4hc) {
+            result = LZ4HC_compress_hashChain(ctx,
+                                src, dst, srcSizePtr, dstCapacity,
+                                cParam.nbSearches, limit, dict);
+        } else {
+            assert(cParam.strat == lz4opt);
+            result = LZ4HC_compress_optimal(ctx,
+                                src, dst, srcSizePtr, dstCapacity,
+                                cParam.nbSearches, cParam.targetLength, limit,
+                                cLevel == LZ4HC_CLEVEL_MAX,   /* ultra mode */
+                                dict, favor);
+        }
+        if (result <= 0) ctx->dirty = 1;
+        return result;
+    }
+}
+
+static void LZ4HC_setExternalDict(LZ4HC_CCtx_internal* ctxPtr, const BYTE* newBlock);
+
+static int
+LZ4HC_compress_generic_noDictCtx (
+        LZ4HC_CCtx_internal* const ctx,
+        const char* const src,
+        char* const dst,
+        int* const srcSizePtr,
+        int const dstCapacity,
+        int cLevel,
+        limitedOutput_directive limit
+        )
+{
+    assert(ctx->dictCtx == NULL);
+    return LZ4HC_compress_generic_internal(ctx, src, dst, srcSizePtr, dstCapacity, cLevel, limit, noDictCtx);
+}
+
+static int
+LZ4HC_compress_generic_dictCtx (
+        LZ4HC_CCtx_internal* const ctx,
+        const char* const src,
+        char* const dst,
+        int* const srcSizePtr,
+        int const dstCapacity,
+        int cLevel,
+        limitedOutput_directive limit
+        )
+{
+    const size_t position = (size_t)(ctx->end - ctx->base) - ctx->lowLimit;
+    assert(ctx->dictCtx != NULL);
+    if (position >= 64 KB) {
+        ctx->dictCtx = NULL;
+        return LZ4HC_compress_generic_noDictCtx(ctx, src, dst, srcSizePtr, dstCapacity, cLevel, limit);
+    } else if (position == 0 && *srcSizePtr > 4 KB) {
+        memcpy(ctx, ctx->dictCtx, sizeof(LZ4HC_CCtx_internal));
+        LZ4HC_setExternalDict(ctx, (const BYTE *)src);
+        ctx->compressionLevel = (short)cLevel;
+        return LZ4HC_compress_generic_noDictCtx(ctx, src, dst, srcSizePtr, dstCapacity, cLevel, limit);
+    } else {
+        return LZ4HC_compress_generic_internal(ctx, src, dst, srcSizePtr, dstCapacity, cLevel, limit, usingDictCtxHc);
+    }
+}
+
+static int
+LZ4HC_compress_generic (
+        LZ4HC_CCtx_internal* const ctx,
+        const char* const src,
+        char* const dst,
+        int* const srcSizePtr,
+        int const dstCapacity,
+        int cLevel,
+        limitedOutput_directive limit
+        )
+{
+    if (ctx->dictCtx == NULL) {
+        return LZ4HC_compress_generic_noDictCtx(ctx, src, dst, srcSizePtr, dstCapacity, cLevel, limit);
+    } else {
+        return LZ4HC_compress_generic_dictCtx(ctx, src, dst, srcSizePtr, dstCapacity, cLevel, limit);
+    }
+}
+
+
+int LZ4_sizeofStateHC(void) { return (int)sizeof(LZ4_streamHC_t); }
+
+static size_t LZ4_streamHC_t_alignment(void)
+{
+#if LZ4_ALIGN_TEST
+    typedef struct { char c; LZ4_streamHC_t t; } t_a;
+    return sizeof(t_a) - sizeof(LZ4_streamHC_t);
+#else
+    return 1;  /* effectively disabled */
+#endif
+}
+
+/* state is presumed correctly initialized,
+ * in which case its size and alignment have already been validate */
+int LZ4_compress_HC_extStateHC_fastReset (void* state, const char* src, char* dst, int srcSize, int dstCapacity, int compressionLevel)
+{
+    LZ4HC_CCtx_internal* const ctx = &((LZ4_streamHC_t*)state)->internal_donotuse;
+    if (!LZ4_isAligned(state, LZ4_streamHC_t_alignment())) return 0;
+    LZ4_resetStreamHC_fast((LZ4_streamHC_t*)state, compressionLevel);
+    LZ4HC_init_internal (ctx, (const BYTE*)src);
+    if (dstCapacity < LZ4_compressBound(srcSize))
+        return LZ4HC_compress_generic (ctx, src, dst, &srcSize, dstCapacity, compressionLevel, limitedOutput);
+    else
+        return LZ4HC_compress_generic (ctx, src, dst, &srcSize, dstCapacity, compressionLevel, notLimited);
+}
+
+int LZ4_compress_HC_extStateHC (void* state, const char* src, char* dst, int srcSize, int dstCapacity, int compressionLevel)
+{
+    LZ4_streamHC_t* const ctx = LZ4_initStreamHC(state, sizeof(*ctx));
+    if (ctx==NULL) return 0;   /* init failure */
+    return LZ4_compress_HC_extStateHC_fastReset(state, src, dst, srcSize, dstCapacity, compressionLevel);
+}
+
+int LZ4_compress_HC(const char* src, char* dst, int srcSize, int dstCapacity, int compressionLevel)
+{
+#if defined(LZ4HC_HEAPMODE) && LZ4HC_HEAPMODE==1
+    LZ4_streamHC_t* const statePtr = (LZ4_streamHC_t*)ALLOC(sizeof(LZ4_streamHC_t));
+#else
+    LZ4_streamHC_t state;
+    LZ4_streamHC_t* const statePtr = &state;
+#endif
+    int const cSize = LZ4_compress_HC_extStateHC(statePtr, src, dst, srcSize, dstCapacity, compressionLevel);
+#if defined(LZ4HC_HEAPMODE) && LZ4HC_HEAPMODE==1
+    FREEMEM(statePtr);
+#endif
+    return cSize;
+}
+
+/* state is presumed sized correctly (>= sizeof(LZ4_streamHC_t)) */
+int LZ4_compress_HC_destSize(void* state, const char* source, char* dest, int* sourceSizePtr, int targetDestSize, int cLevel)
+{
+    LZ4_streamHC_t* const ctx = LZ4_initStreamHC(state, sizeof(*ctx));
+    if (ctx==NULL) return 0;   /* init failure */
+    LZ4HC_init_internal(&ctx->internal_donotuse, (const BYTE*) source);
+    LZ4_setCompressionLevel(ctx, cLevel);
+    return LZ4HC_compress_generic(&ctx->internal_donotuse, source, dest, sourceSizePtr, targetDestSize, cLevel, fillOutput);
+}
+
+
+
+/**************************************
+*  Streaming Functions
+**************************************/
+/* allocation */
+LZ4_streamHC_t* LZ4_createStreamHC(void)
+{
+    LZ4_streamHC_t* const state =
+        (LZ4_streamHC_t*)ALLOC_AND_ZERO(sizeof(LZ4_streamHC_t));
+    if (state == NULL) return NULL;
+    LZ4_setCompressionLevel(state, LZ4HC_CLEVEL_DEFAULT);
+    return state;
+}
+
+int LZ4_freeStreamHC (LZ4_streamHC_t* LZ4_streamHCPtr)
+{
+    DEBUGLOG(4, "LZ4_freeStreamHC(%p)", LZ4_streamHCPtr);
+    if (!LZ4_streamHCPtr) return 0;  /* support free on NULL */
+    FREEMEM(LZ4_streamHCPtr);
+    return 0;
+}
+
+
+LZ4_streamHC_t* LZ4_initStreamHC (void* buffer, size_t size)
+{
+    LZ4_streamHC_t* const LZ4_streamHCPtr = (LZ4_streamHC_t*)buffer;
+    /* if compilation fails here, LZ4_STREAMHCSIZE must be increased */
+    LZ4_STATIC_ASSERT(sizeof(LZ4HC_CCtx_internal) <= LZ4_STREAMHCSIZE);
+    DEBUGLOG(4, "LZ4_initStreamHC(%p, %u)", buffer, (unsigned)size);
+    /* check conditions */
+    if (buffer == NULL) return NULL;
+    if (size < sizeof(LZ4_streamHC_t)) return NULL;
+    if (!LZ4_isAligned(buffer, LZ4_streamHC_t_alignment())) return NULL;
+    /* init */
+    { LZ4HC_CCtx_internal* const hcstate = &(LZ4_streamHCPtr->internal_donotuse);
+      MEM_INIT(hcstate, 0, sizeof(*hcstate)); }
+    LZ4_setCompressionLevel(LZ4_streamHCPtr, LZ4HC_CLEVEL_DEFAULT);
+    return LZ4_streamHCPtr;
+}
+
+/* just a stub */
+void LZ4_resetStreamHC (LZ4_streamHC_t* LZ4_streamHCPtr, int compressionLevel)
+{
+    LZ4_initStreamHC(LZ4_streamHCPtr, sizeof(*LZ4_streamHCPtr));
+    LZ4_setCompressionLevel(LZ4_streamHCPtr, compressionLevel);
+}
+
+void LZ4_resetStreamHC_fast (LZ4_streamHC_t* LZ4_streamHCPtr, int compressionLevel)
+{
+    DEBUGLOG(4, "LZ4_resetStreamHC_fast(%p, %d)", LZ4_streamHCPtr, compressionLevel);
+    if (LZ4_streamHCPtr->internal_donotuse.dirty) {
+        LZ4_initStreamHC(LZ4_streamHCPtr, sizeof(*LZ4_streamHCPtr));
+    } else {
+        /* preserve end - base : can trigger clearTable's threshold */
+        if (LZ4_streamHCPtr->internal_donotuse.end != NULL) {
+            LZ4_streamHCPtr->internal_donotuse.end -= (uptrval)LZ4_streamHCPtr->internal_donotuse.base;
+        } else {
+            assert(LZ4_streamHCPtr->internal_donotuse.base == NULL);
+        }
+        LZ4_streamHCPtr->internal_donotuse.base = NULL;
+        LZ4_streamHCPtr->internal_donotuse.dictCtx = NULL;
+    }
+    LZ4_setCompressionLevel(LZ4_streamHCPtr, compressionLevel);
+}
+
+void LZ4_setCompressionLevel(LZ4_streamHC_t* LZ4_streamHCPtr, int compressionLevel)
+{
+    DEBUGLOG(5, "LZ4_setCompressionLevel(%p, %d)", LZ4_streamHCPtr, compressionLevel);
+    if (compressionLevel < 1) compressionLevel = LZ4HC_CLEVEL_DEFAULT;
+    if (compressionLevel > LZ4HC_CLEVEL_MAX) compressionLevel = LZ4HC_CLEVEL_MAX;
+    LZ4_streamHCPtr->internal_donotuse.compressionLevel = (short)compressionLevel;
+}
+
+void LZ4_favorDecompressionSpeed(LZ4_streamHC_t* LZ4_streamHCPtr, int favor)
+{
+    LZ4_streamHCPtr->internal_donotuse.favorDecSpeed = (favor!=0);
+}
+
+/* LZ4_loadDictHC() :
+ * LZ4_streamHCPtr is presumed properly initialized */
+int LZ4_loadDictHC (LZ4_streamHC_t* LZ4_streamHCPtr,
+              const char* dictionary, int dictSize)
+{
+    LZ4HC_CCtx_internal* const ctxPtr = &LZ4_streamHCPtr->internal_donotuse;
+    DEBUGLOG(4, "LZ4_loadDictHC(ctx:%p, dict:%p, dictSize:%d)", LZ4_streamHCPtr, dictionary, dictSize);
+    assert(LZ4_streamHCPtr != NULL);
+    if (dictSize > 64 KB) {
+        dictionary += (size_t)dictSize - 64 KB;
+        dictSize = 64 KB;
+    }
+    /* need a full initialization, there are bad side-effects when using resetFast() */
+    {   int const cLevel = ctxPtr->compressionLevel;
+        LZ4_initStreamHC(LZ4_streamHCPtr, sizeof(*LZ4_streamHCPtr));
+        LZ4_setCompressionLevel(LZ4_streamHCPtr, cLevel);
+    }
+    LZ4HC_init_internal (ctxPtr, (const BYTE*)dictionary);
+    ctxPtr->end = (const BYTE*)dictionary + dictSize;
+    if (dictSize >= 4) LZ4HC_Insert (ctxPtr, ctxPtr->end-3);
+    return dictSize;
+}
+
+void LZ4_attach_HC_dictionary(LZ4_streamHC_t *working_stream, const LZ4_streamHC_t *dictionary_stream) {
+    working_stream->internal_donotuse.dictCtx = dictionary_stream != NULL ? &(dictionary_stream->internal_donotuse) : NULL;
+}
+
+/* compression */
+
+static void LZ4HC_setExternalDict(LZ4HC_CCtx_internal* ctxPtr, const BYTE* newBlock)
+{
+    DEBUGLOG(4, "LZ4HC_setExternalDict(%p, %p)", ctxPtr, newBlock);
+    if (ctxPtr->end >= ctxPtr->base + ctxPtr->dictLimit + 4)
+        LZ4HC_Insert (ctxPtr, ctxPtr->end-3);   /* Referencing remaining dictionary content */
+
+    /* Only one memory segment for extDict, so any previous extDict is lost at this stage */
+    ctxPtr->lowLimit  = ctxPtr->dictLimit;
+    ctxPtr->dictLimit = (U32)(ctxPtr->end - ctxPtr->base);
+    ctxPtr->dictBase  = ctxPtr->base;
+    ctxPtr->base = newBlock - ctxPtr->dictLimit;
+    ctxPtr->end  = newBlock;
+    ctxPtr->nextToUpdate = ctxPtr->dictLimit;   /* match referencing will resume from there */
+
+    /* cannot reference an extDict and a dictCtx at the same time */
+    ctxPtr->dictCtx = NULL;
+}
+
+static int
+LZ4_compressHC_continue_generic (LZ4_streamHC_t* LZ4_streamHCPtr,
+                                 const char* src, char* dst,
+                                 int* srcSizePtr, int dstCapacity,
+                                 limitedOutput_directive limit)
+{
+    LZ4HC_CCtx_internal* const ctxPtr = &LZ4_streamHCPtr->internal_donotuse;
+    DEBUGLOG(5, "LZ4_compressHC_continue_generic(ctx=%p, src=%p, srcSize=%d, limit=%d)",
+                LZ4_streamHCPtr, src, *srcSizePtr, limit);
+    assert(ctxPtr != NULL);
+    /* auto-init if forgotten */
+    if (ctxPtr->base == NULL) LZ4HC_init_internal (ctxPtr, (const BYTE*) src);
+
+    /* Check overflow */
+    if ((size_t)(ctxPtr->end - ctxPtr->base) > 2 GB) {
+        size_t dictSize = (size_t)(ctxPtr->end - ctxPtr->base) - ctxPtr->dictLimit;
+        if (dictSize > 64 KB) dictSize = 64 KB;
+        LZ4_loadDictHC(LZ4_streamHCPtr, (const char*)(ctxPtr->end) - dictSize, (int)dictSize);
+    }
+
+    /* Check if blocks follow each other */
+    if ((const BYTE*)src != ctxPtr->end)
+        LZ4HC_setExternalDict(ctxPtr, (const BYTE*)src);
+
+    /* Check overlapping input/dictionary space */
+    {   const BYTE* sourceEnd = (const BYTE*) src + *srcSizePtr;
+        const BYTE* const dictBegin = ctxPtr->dictBase + ctxPtr->lowLimit;
+        const BYTE* const dictEnd   = ctxPtr->dictBase + ctxPtr->dictLimit;
+        if ((sourceEnd > dictBegin) && ((const BYTE*)src < dictEnd)) {
+            if (sourceEnd > dictEnd) sourceEnd = dictEnd;
+            ctxPtr->lowLimit = (U32)(sourceEnd - ctxPtr->dictBase);
+            if (ctxPtr->dictLimit - ctxPtr->lowLimit < 4) ctxPtr->lowLimit = ctxPtr->dictLimit;
+    }   }
+
+    return LZ4HC_compress_generic (ctxPtr, src, dst, srcSizePtr, dstCapacity, ctxPtr->compressionLevel, limit);
+}
+
+int LZ4_compress_HC_continue (LZ4_streamHC_t* LZ4_streamHCPtr, const char* src, char* dst, int srcSize, int dstCapacity)
+{
+    if (dstCapacity < LZ4_compressBound(srcSize))
+        return LZ4_compressHC_continue_generic (LZ4_streamHCPtr, src, dst, &srcSize, dstCapacity, limitedOutput);
+    else
+        return LZ4_compressHC_continue_generic (LZ4_streamHCPtr, src, dst, &srcSize, dstCapacity, notLimited);
+}
+
+int LZ4_compress_HC_continue_destSize (LZ4_streamHC_t* LZ4_streamHCPtr, const char* src, char* dst, int* srcSizePtr, int targetDestSize)
+{
+    return LZ4_compressHC_continue_generic(LZ4_streamHCPtr, src, dst, srcSizePtr, targetDestSize, fillOutput);
+}
+
+
+
+/* LZ4_saveDictHC :
+ * save history content
+ * into a user-provided buffer
+ * which is then used to continue compression
+ */
+int LZ4_saveDictHC (LZ4_streamHC_t* LZ4_streamHCPtr, char* safeBuffer, int dictSize)
+{
+    LZ4HC_CCtx_internal* const streamPtr = &LZ4_streamHCPtr->internal_donotuse;
+    int const prefixSize = (int)(streamPtr->end - (streamPtr->base + streamPtr->dictLimit));
+    DEBUGLOG(5, "LZ4_saveDictHC(%p, %p, %d)", LZ4_streamHCPtr, safeBuffer, dictSize);
+    assert(prefixSize >= 0);
+    if (dictSize > 64 KB) dictSize = 64 KB;
+    if (dictSize < 4) dictSize = 0;
+    if (dictSize > prefixSize) dictSize = prefixSize;
+    if (safeBuffer == NULL) assert(dictSize == 0);
+    if (dictSize > 0)
+        memmove(safeBuffer, streamPtr->end - dictSize, dictSize);
+    {   U32 const endIndex = (U32)(streamPtr->end - streamPtr->base);
+        streamPtr->end = (const BYTE*)safeBuffer + dictSize;
+        streamPtr->base = streamPtr->end - endIndex;
+        streamPtr->dictLimit = endIndex - (U32)dictSize;
+        streamPtr->lowLimit = endIndex - (U32)dictSize;
+        if (streamPtr->nextToUpdate < streamPtr->dictLimit)
+            streamPtr->nextToUpdate = streamPtr->dictLimit;
+    }
+    return dictSize;
+}
+
+
+/***************************************************
+*  Deprecated Functions
+***************************************************/
+
+/* These functions currently generate deprecation warnings */
+
+/* Wrappers for deprecated compression functions */
+int LZ4_compressHC(const char* src, char* dst, int srcSize) { return LZ4_compress_HC (src, dst, srcSize, LZ4_compressBound(srcSize), 0); }
+int LZ4_compressHC_limitedOutput(const char* src, char* dst, int srcSize, int maxDstSize) { return LZ4_compress_HC(src, dst, srcSize, maxDstSize, 0); }
+int LZ4_compressHC2(const char* src, char* dst, int srcSize, int cLevel) { return LZ4_compress_HC (src, dst, srcSize, LZ4_compressBound(srcSize), cLevel); }
+int LZ4_compressHC2_limitedOutput(const char* src, char* dst, int srcSize, int maxDstSize, int cLevel) { return LZ4_compress_HC(src, dst, srcSize, maxDstSize, cLevel); }
+int LZ4_compressHC_withStateHC (void* state, const char* src, char* dst, int srcSize) { return LZ4_compress_HC_extStateHC (state, src, dst, srcSize, LZ4_compressBound(srcSize), 0); }
+int LZ4_compressHC_limitedOutput_withStateHC (void* state, const char* src, char* dst, int srcSize, int maxDstSize) { return LZ4_compress_HC_extStateHC (state, src, dst, srcSize, maxDstSize, 0); }
+int LZ4_compressHC2_withStateHC (void* state, const char* src, char* dst, int srcSize, int cLevel) { return LZ4_compress_HC_extStateHC(state, src, dst, srcSize, LZ4_compressBound(srcSize), cLevel); }
+int LZ4_compressHC2_limitedOutput_withStateHC (void* state, const char* src, char* dst, int srcSize, int maxDstSize, int cLevel) { return LZ4_compress_HC_extStateHC(state, src, dst, srcSize, maxDstSize, cLevel); }
+int LZ4_compressHC_continue (LZ4_streamHC_t* ctx, const char* src, char* dst, int srcSize) { return LZ4_compress_HC_continue (ctx, src, dst, srcSize, LZ4_compressBound(srcSize)); }
+int LZ4_compressHC_limitedOutput_continue (LZ4_streamHC_t* ctx, const char* src, char* dst, int srcSize, int maxDstSize) { return LZ4_compress_HC_continue (ctx, src, dst, srcSize, maxDstSize); }
+
+
+/* Deprecated streaming functions */
+int LZ4_sizeofStreamStateHC(void) { return LZ4_STREAMHCSIZE; }
+
+/* state is presumed correctly sized, aka >= sizeof(LZ4_streamHC_t)
+ * @return : 0 on success, !=0 if error */
+int LZ4_resetStreamStateHC(void* state, char* inputBuffer)
+{
+    LZ4_streamHC_t* const hc4 = LZ4_initStreamHC(state, sizeof(*hc4));
+    if (hc4 == NULL) return 1;   /* init failed */
+    LZ4HC_init_internal (&hc4->internal_donotuse, (const BYTE*)inputBuffer);
+    return 0;
+}
+
+void* LZ4_createHC (const char* inputBuffer)
+{
+    LZ4_streamHC_t* const hc4 = LZ4_createStreamHC();
+    if (hc4 == NULL) return NULL;   /* not enough memory */
+    LZ4HC_init_internal (&hc4->internal_donotuse, (const BYTE*)inputBuffer);
+    return hc4;
+}
+
+int LZ4_freeHC (void* LZ4HC_Data)
+{
+    if (!LZ4HC_Data) return 0;  /* support free on NULL */
+    FREEMEM(LZ4HC_Data);
+    return 0;
+}
+
+int LZ4_compressHC2_continue (void* LZ4HC_Data, const char* src, char* dst, int srcSize, int cLevel)
+{
+    return LZ4HC_compress_generic (&((LZ4_streamHC_t*)LZ4HC_Data)->internal_donotuse, src, dst, &srcSize, 0, cLevel, notLimited);
+}
+
+int LZ4_compressHC2_limitedOutput_continue (void* LZ4HC_Data, const char* src, char* dst, int srcSize, int dstCapacity, int cLevel)
+{
+    return LZ4HC_compress_generic (&((LZ4_streamHC_t*)LZ4HC_Data)->internal_donotuse, src, dst, &srcSize, dstCapacity, cLevel, limitedOutput);
+}
+
+char* LZ4_slideInputBufferHC(void* LZ4HC_Data)
+{
+    LZ4_streamHC_t *ctx = (LZ4_streamHC_t*)LZ4HC_Data;
+    const BYTE *bufferStart = ctx->internal_donotuse.base + ctx->internal_donotuse.lowLimit;
+    LZ4_resetStreamHC_fast(ctx, ctx->internal_donotuse.compressionLevel);
+    /* avoid const char * -> char * conversion warning :( */
+    return (char *)(uptrval)bufferStart;
+}
+
+
+/* ================================================
+ *  LZ4 Optimal parser (levels [LZ4HC_CLEVEL_OPT_MIN - LZ4HC_CLEVEL_MAX])
+ * ===============================================*/
+typedef struct {
+    int price;
+    int off;
+    int mlen;
+    int litlen;
+} LZ4HC_optimal_t;
+
+/* price in bytes */
+LZ4_FORCE_INLINE int LZ4HC_literalsPrice(int const litlen)
+{
+    int price = litlen;
+    assert(litlen >= 0);
+    if (litlen >= (int)RUN_MASK)
+        price += 1 + ((litlen-(int)RUN_MASK) / 255);
+    return price;
+}
+
+
+/* requires mlen >= MINMATCH */
+LZ4_FORCE_INLINE int LZ4HC_sequencePrice(int litlen, int mlen)
+{
+    int price = 1 + 2 ; /* token + 16-bit offset */
+    assert(litlen >= 0);
+    assert(mlen >= MINMATCH);
+
+    price += LZ4HC_literalsPrice(litlen);
+
+    if (mlen >= (int)(ML_MASK+MINMATCH))
+        price += 1 + ((mlen-(int)(ML_MASK+MINMATCH)) / 255);
+
+    return price;
+}
+
+
+typedef struct {
+    int off;
+    int len;
+} LZ4HC_match_t;
+
+LZ4_FORCE_INLINE LZ4HC_match_t
+LZ4HC_FindLongerMatch(LZ4HC_CCtx_internal* const ctx,
+                      const BYTE* ip, const BYTE* const iHighLimit,
+                      int minLen, int nbSearches,
+                      const dictCtx_directive dict,
+                      const HCfavor_e favorDecSpeed)
+{
+    LZ4HC_match_t match = { 0 , 0 };
+    const BYTE* matchPtr = NULL;
+    /* note : LZ4HC_InsertAndGetWiderMatch() is able to modify the starting position of a match (*startpos),
+     * but this won't be the case here, as we define iLowLimit==ip,
+     * so LZ4HC_InsertAndGetWiderMatch() won't be allowed to search past ip */
+    int matchLength = LZ4HC_InsertAndGetWiderMatch(ctx, ip, ip, iHighLimit, minLen, &matchPtr, &ip, nbSearches, 1 /*patternAnalysis*/, 1 /*chainSwap*/, dict, favorDecSpeed);
+    if (matchLength <= minLen) return match;
+    if (favorDecSpeed) {
+        if ((matchLength>18) & (matchLength<=36)) matchLength=18;   /* favor shortcut */
+    }
+    match.len = matchLength;
+    match.off = (int)(ip-matchPtr);
+    return match;
+}
+
+
+static int LZ4HC_compress_optimal ( LZ4HC_CCtx_internal* ctx,
+                                    const char* const source,
+                                    char* dst,
+                                    int* srcSizePtr,
+                                    int dstCapacity,
+                                    int const nbSearches,
+                                    size_t sufficient_len,
+                                    const limitedOutput_directive limit,
+                                    int const fullUpdate,
+                                    const dictCtx_directive dict,
+                                    const HCfavor_e favorDecSpeed)
+{
+    int retval = 0;
+#define TRAILING_LITERALS 3
+#ifdef LZ4HC_HEAPMODE
+    LZ4HC_optimal_t* const opt = (LZ4HC_optimal_t*)ALLOC(sizeof(LZ4HC_optimal_t) * (LZ4_OPT_NUM + TRAILING_LITERALS));
+#else
+    LZ4HC_optimal_t opt[LZ4_OPT_NUM + TRAILING_LITERALS];   /* ~64 KB, which is a bit large for stack... */
+#endif
+
+    const BYTE* ip = (const BYTE*) source;
+    const BYTE* anchor = ip;
+    const BYTE* const iend = ip + *srcSizePtr;
+    const BYTE* const mflimit = iend - MFLIMIT;
+    const BYTE* const matchlimit = iend - LASTLITERALS;
+    BYTE* op = (BYTE*) dst;
+    BYTE* opSaved = (BYTE*) dst;
+    BYTE* oend = op + dstCapacity;
+    int ovml = MINMATCH;  /* overflow - last sequence */
+    const BYTE* ovref = NULL;
+
+    /* init */
+#ifdef LZ4HC_HEAPMODE
+    if (opt == NULL) goto _return_label;
+#endif
+    DEBUGLOG(5, "LZ4HC_compress_optimal(dst=%p, dstCapa=%u)", dst, (unsigned)dstCapacity);
+    *srcSizePtr = 0;
+    if (limit == fillOutput) oend -= LASTLITERALS;   /* Hack for support LZ4 format restriction */
+    if (sufficient_len >= LZ4_OPT_NUM) sufficient_len = LZ4_OPT_NUM-1;
+
+    /* Main Loop */
+    while (ip <= mflimit) {
+         int const llen = (int)(ip - anchor);
+         int best_mlen, best_off;
+         int cur, last_match_pos = 0;
+
+         LZ4HC_match_t const firstMatch = LZ4HC_FindLongerMatch(ctx, ip, matchlimit, MINMATCH-1, nbSearches, dict, favorDecSpeed);
+         if (firstMatch.len==0) { ip++; continue; }
+
+         if ((size_t)firstMatch.len > sufficient_len) {
+             /* good enough solution : immediate encoding */
+             int const firstML = firstMatch.len;
+             const BYTE* const matchPos = ip - firstMatch.off;
+             opSaved = op;
+             if ( LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), firstML, matchPos, limit, oend) ) {  /* updates ip, op and anchor */
+                 ovml = firstML;
+                 ovref = matchPos;
+                 goto _dest_overflow;
+             }
+             continue;
+         }
+
+         /* set prices for first positions (literals) */
+         {   int rPos;
+             for (rPos = 0 ; rPos < MINMATCH ; rPos++) {
+                 int const cost = LZ4HC_literalsPrice(llen + rPos);
+                 opt[rPos].mlen = 1;
+                 opt[rPos].off = 0;
+                 opt[rPos].litlen = llen + rPos;
+                 opt[rPos].price = cost;
+                 DEBUGLOG(7, "rPos:%3i => price:%3i (litlen=%i) -- initial setup",
+                             rPos, cost, opt[rPos].litlen);
+         }   }
+         /* set prices using initial match */
+         {   int mlen = MINMATCH;
+             int const matchML = firstMatch.len;   /* necessarily < sufficient_len < LZ4_OPT_NUM */
+             int const offset = firstMatch.off;
+             assert(matchML < LZ4_OPT_NUM);
+             for ( ; mlen <= matchML ; mlen++) {
+                 int const cost = LZ4HC_sequencePrice(llen, mlen);
+                 opt[mlen].mlen = mlen;
+                 opt[mlen].off = offset;
+                 opt[mlen].litlen = llen;
+                 opt[mlen].price = cost;
+                 DEBUGLOG(7, "rPos:%3i => price:%3i (matchlen=%i) -- initial setup",
+                             mlen, cost, mlen);
+         }   }
+         last_match_pos = firstMatch.len;
+         {   int addLit;
+             for (addLit = 1; addLit <= TRAILING_LITERALS; addLit ++) {
+                 opt[last_match_pos+addLit].mlen = 1; /* literal */
+                 opt[last_match_pos+addLit].off = 0;
+                 opt[last_match_pos+addLit].litlen = addLit;
+                 opt[last_match_pos+addLit].price = opt[last_match_pos].price + LZ4HC_literalsPrice(addLit);
+                 DEBUGLOG(7, "rPos:%3i => price:%3i (litlen=%i) -- initial setup",
+                             last_match_pos+addLit, opt[last_match_pos+addLit].price, addLit);
+         }   }
+
+         /* check further positions */
+         for (cur = 1; cur < last_match_pos; cur++) {
+             const BYTE* const curPtr = ip + cur;
+             LZ4HC_match_t newMatch;
+
+             if (curPtr > mflimit) break;
+             DEBUGLOG(7, "rPos:%u[%u] vs [%u]%u",
+                     cur, opt[cur].price, opt[cur+1].price, cur+1);
+             if (fullUpdate) {
+                 /* not useful to search here if next position has same (or lower) cost */
+                 if ( (opt[cur+1].price <= opt[cur].price)
+                   /* in some cases, next position has same cost, but cost rises sharply after, so a small match would still be beneficial */
+                   && (opt[cur+MINMATCH].price < opt[cur].price + 3/*min seq price*/) )
+                     continue;
+             } else {
+                 /* not useful to search here if next position has same (or lower) cost */
+                 if (opt[cur+1].price <= opt[cur].price) continue;
+             }
+
+             DEBUGLOG(7, "search at rPos:%u", cur);
+             if (fullUpdate)
+                 newMatch = LZ4HC_FindLongerMatch(ctx, curPtr, matchlimit, MINMATCH-1, nbSearches, dict, favorDecSpeed);
+             else
+                 /* only test matches of minimum length; slightly faster, but misses a few bytes */
+                 newMatch = LZ4HC_FindLongerMatch(ctx, curPtr, matchlimit, last_match_pos - cur, nbSearches, dict, favorDecSpeed);
+             if (!newMatch.len) continue;
+
+             if ( ((size_t)newMatch.len > sufficient_len)
+               || (newMatch.len + cur >= LZ4_OPT_NUM) ) {
+                 /* immediate encoding */
+                 best_mlen = newMatch.len;
+                 best_off = newMatch.off;
+                 last_match_pos = cur + 1;
+                 goto encode;
+             }
+
+             /* before match : set price with literals at beginning */
+             {   int const baseLitlen = opt[cur].litlen;
+                 int litlen;
+                 for (litlen = 1; litlen < MINMATCH; litlen++) {
+                     int const price = opt[cur].price - LZ4HC_literalsPrice(baseLitlen) + LZ4HC_literalsPrice(baseLitlen+litlen);
+                     int const pos = cur + litlen;
+                     if (price < opt[pos].price) {
+                         opt[pos].mlen = 1; /* literal */
+                         opt[pos].off = 0;
+                         opt[pos].litlen = baseLitlen+litlen;
+                         opt[pos].price = price;
+                         DEBUGLOG(7, "rPos:%3i => price:%3i (litlen=%i)",
+                                     pos, price, opt[pos].litlen);
+             }   }   }
+
+             /* set prices using match at position = cur */
+             {   int const matchML = newMatch.len;
+                 int ml = MINMATCH;
+
+                 assert(cur + newMatch.len < LZ4_OPT_NUM);
+                 for ( ; ml <= matchML ; ml++) {
+                     int const pos = cur + ml;
+                     int const offset = newMatch.off;
+                     int price;
+                     int ll;
+                     DEBUGLOG(7, "testing price rPos %i (last_match_pos=%i)",
+                                 pos, last_match_pos);
+                     if (opt[cur].mlen == 1) {
+                         ll = opt[cur].litlen;
+                         price = ((cur > ll) ? opt[cur - ll].price : 0)
+                               + LZ4HC_sequencePrice(ll, ml);
+                     } else {
+                         ll = 0;
+                         price = opt[cur].price + LZ4HC_sequencePrice(0, ml);
+                     }
+
+                    assert((U32)favorDecSpeed <= 1);
+                     if (pos > last_match_pos+TRAILING_LITERALS
+                      || price <= opt[pos].price - (int)favorDecSpeed) {
+                         DEBUGLOG(7, "rPos:%3i => price:%3i (matchlen=%i)",
+                                     pos, price, ml);
+                         assert(pos < LZ4_OPT_NUM);
+                         if ( (ml == matchML)  /* last pos of last match */
+                           && (last_match_pos < pos) )
+                             last_match_pos = pos;
+                         opt[pos].mlen = ml;
+                         opt[pos].off = offset;
+                         opt[pos].litlen = ll;
+                         opt[pos].price = price;
+             }   }   }
+             /* complete following positions with literals */
+             {   int addLit;
+                 for (addLit = 1; addLit <= TRAILING_LITERALS; addLit ++) {
+                     opt[last_match_pos+addLit].mlen = 1; /* literal */
+                     opt[last_match_pos+addLit].off = 0;
+                     opt[last_match_pos+addLit].litlen = addLit;
+                     opt[last_match_pos+addLit].price = opt[last_match_pos].price + LZ4HC_literalsPrice(addLit);
+                     DEBUGLOG(7, "rPos:%3i => price:%3i (litlen=%i)", last_match_pos+addLit, opt[last_match_pos+addLit].price, addLit);
+             }   }
+         }  /* for (cur = 1; cur <= last_match_pos; cur++) */
+
+         assert(last_match_pos < LZ4_OPT_NUM + TRAILING_LITERALS);
+         best_mlen = opt[last_match_pos].mlen;
+         best_off = opt[last_match_pos].off;
+         cur = last_match_pos - best_mlen;
+
+encode: /* cur, last_match_pos, best_mlen, best_off must be set */
+         assert(cur < LZ4_OPT_NUM);
+         assert(last_match_pos >= 1);  /* == 1 when only one candidate */
+         DEBUGLOG(6, "reverse traversal, looking for shortest path (last_match_pos=%i)", last_match_pos);
+         {   int candidate_pos = cur;
+             int selected_matchLength = best_mlen;
+             int selected_offset = best_off;
+             while (1) {  /* from end to beginning */
+                 int const next_matchLength = opt[candidate_pos].mlen;  /* can be 1, means literal */
+                 int const next_offset = opt[candidate_pos].off;
+                 DEBUGLOG(7, "pos %i: sequence length %i", candidate_pos, selected_matchLength);
+                 opt[candidate_pos].mlen = selected_matchLength;
+                 opt[candidate_pos].off = selected_offset;
+                 selected_matchLength = next_matchLength;
+                 selected_offset = next_offset;
+                 if (next_matchLength > candidate_pos) break; /* last match elected, first match to encode */
+                 assert(next_matchLength > 0);  /* can be 1, means literal */
+                 candidate_pos -= next_matchLength;
+         }   }
+
+         /* encode all recorded sequences in order */
+         {   int rPos = 0;  /* relative position (to ip) */
+             while (rPos < last_match_pos) {
+                 int const ml = opt[rPos].mlen;
+                 int const offset = opt[rPos].off;
+                 if (ml == 1) { ip++; rPos++; continue; }  /* literal; note: can end up with several literals, in which case, skip them */
+                 rPos += ml;
+                 assert(ml >= MINMATCH);
+                 assert((offset >= 1) && (offset <= LZ4_DISTANCE_MAX));
+                 opSaved = op;
+                 if ( LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), ml, ip - offset, limit, oend) ) {  /* updates ip, op and anchor */
+                     ovml = ml;
+                     ovref = ip - offset;
+                     goto _dest_overflow;
+         }   }   }
+     }  /* while (ip <= mflimit) */
+
+_last_literals:
+     /* Encode Last Literals */
+     {   size_t lastRunSize = (size_t)(iend - anchor);  /* literals */
+         size_t llAdd = (lastRunSize + 255 - RUN_MASK) / 255;
+         size_t const totalSize = 1 + llAdd + lastRunSize;
+         if (limit == fillOutput) oend += LASTLITERALS;  /* restore correct value */
+         if (limit && (op + totalSize > oend)) {
+             if (limit == limitedOutput) { /* Check output limit */
+                retval = 0;
+                goto _return_label;
+             }
+             /* adapt lastRunSize to fill 'dst' */
+             lastRunSize  = (size_t)(oend - op) - 1 /*token*/;
+             llAdd = (lastRunSize + 256 - RUN_MASK) / 256;
+             lastRunSize -= llAdd;
+         }
+         DEBUGLOG(6, "Final literal run : %i literals", (int)lastRunSize);
+         ip = anchor + lastRunSize; /* can be != iend if limit==fillOutput */
+
+         if (lastRunSize >= RUN_MASK) {
+             size_t accumulator = lastRunSize - RUN_MASK;
+             *op++ = (RUN_MASK << ML_BITS);
+             for(; accumulator >= 255 ; accumulator -= 255) *op++ = 255;
+             *op++ = (BYTE) accumulator;
+         } else {
+             *op++ = (BYTE)(lastRunSize << ML_BITS);
+         }
+         memcpy(op, anchor, lastRunSize);
+         op += lastRunSize;
+     }
+
+     /* End */
+     *srcSizePtr = (int) (((const char*)ip) - source);
+     retval = (int) ((char*)op-dst);
+     goto _return_label;
+
+_dest_overflow:
+if (limit == fillOutput) {
+     /* Assumption : ip, anchor, ovml and ovref must be set correctly */
+     size_t const ll = (size_t)(ip - anchor);
+     size_t const ll_addbytes = (ll + 240) / 255;
+     size_t const ll_totalCost = 1 + ll_addbytes + ll;
+     BYTE* const maxLitPos = oend - 3; /* 2 for offset, 1 for token */
+     DEBUGLOG(6, "Last sequence overflowing (only %i bytes remaining)", (int)(oend-1-opSaved));
+     op = opSaved;  /* restore correct out pointer */
+     if (op + ll_totalCost <= maxLitPos) {
+         /* ll validated; now adjust match length */
+         size_t const bytesLeftForMl = (size_t)(maxLitPos - (op+ll_totalCost));
+         size_t const maxMlSize = MINMATCH + (ML_MASK-1) + (bytesLeftForMl * 255);
+         assert(maxMlSize < INT_MAX); assert(ovml >= 0);
+         if ((size_t)ovml > maxMlSize) ovml = (int)maxMlSize;
+         if ((oend + LASTLITERALS) - (op + ll_totalCost + 2) - 1 + ovml >= MFLIMIT) {
+             DEBUGLOG(6, "Space to end : %i + ml (%i)", (int)((oend + LASTLITERALS) - (op + ll_totalCost + 2) - 1), ovml);
+             DEBUGLOG(6, "Before : ip = %p, anchor = %p", ip, anchor);
+             LZ4HC_encodeSequence(UPDATABLE(ip, op, anchor), ovml, ovref, notLimited, oend);
+             DEBUGLOG(6, "After : ip = %p, anchor = %p", ip, anchor);
+     }   }
+     goto _last_literals;
+}
+_return_label:
+#ifdef LZ4HC_HEAPMODE
+     FREEMEM(opt);
+#endif
+     return retval;
+}
diff --git a/mfbt/lz4/lz4hc.h b/mfbt/lz4/lz4hc.h
new file mode 100644
index 0000000000..3d441fb6fa
--- /dev/null
+++ b/mfbt/lz4/lz4hc.h
@@ -0,0 +1,413 @@
+/*
+   LZ4 HC - High Compression Mode of LZ4
+   Header File
+   Copyright (C) 2011-2017, Yann Collet.
+   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+       * Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+       * Redistributions in binary form must reproduce the above
+   copyright notice, this list of conditions and the following disclaimer
+   in the documentation and/or other materials provided with the
+   distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+   You can contact the author at :
+   - LZ4 source repository : https://github.com/lz4/lz4
+   - LZ4 public forum : https://groups.google.com/forum/#!forum/lz4c
+*/
+#ifndef LZ4_HC_H_19834876238432
+#define LZ4_HC_H_19834876238432
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* --- Dependency --- */
+/* note : lz4hc requires lz4.h/lz4.c for compilation */
+#include "lz4.h"   /* stddef, LZ4LIB_API, LZ4_DEPRECATED */
+
+
+/* --- Useful constants --- */
+#define LZ4HC_CLEVEL_MIN         3
+#define LZ4HC_CLEVEL_DEFAULT     9
+#define LZ4HC_CLEVEL_OPT_MIN    10
+#define LZ4HC_CLEVEL_MAX        12
+
+
+/*-************************************
+ *  Block Compression
+ **************************************/
+/*! LZ4_compress_HC() :
+ *  Compress data from `src` into `dst`, using the powerful but slower "HC" algorithm.
+ * `dst` must be already allocated.
+ *  Compression is guaranteed to succeed if `dstCapacity >= LZ4_compressBound(srcSize)` (see "lz4.h")
+ *  Max supported `srcSize` value is LZ4_MAX_INPUT_SIZE (see "lz4.h")
+ * `compressionLevel` : any value between 1 and LZ4HC_CLEVEL_MAX will work.
+ *                      Values > LZ4HC_CLEVEL_MAX behave the same as LZ4HC_CLEVEL_MAX.
+ * @return : the number of bytes written into 'dst'
+ *           or 0 if compression fails.
+ */
+LZ4LIB_API int LZ4_compress_HC (const char* src, char* dst, int srcSize, int dstCapacity, int compressionLevel);
+
+
+/* Note :
+ *   Decompression functions are provided within "lz4.h" (BSD license)
+ */
+
+
+/*! LZ4_compress_HC_extStateHC() :
+ *  Same as LZ4_compress_HC(), but using an externally allocated memory segment for `state`.
+ * `state` size is provided by LZ4_sizeofStateHC().
+ *  Memory segment must be aligned on 8-bytes boundaries (which a normal malloc() should do properly).
+ */
+LZ4LIB_API int LZ4_sizeofStateHC(void);
+LZ4LIB_API int LZ4_compress_HC_extStateHC(void* stateHC, const char* src, char* dst, int srcSize, int maxDstSize, int compressionLevel);
+
+
+/*! LZ4_compress_HC_destSize() : v1.9.0+
+ *  Will compress as much data as possible from `src`
+ *  to fit into `targetDstSize` budget.
+ *  Result is provided in 2 parts :
+ * @return : the number of bytes written into 'dst' (necessarily <= targetDstSize)
+ *           or 0 if compression fails.
+ * `srcSizePtr` : on success, *srcSizePtr is updated to indicate how much bytes were read from `src`
+ */
+LZ4LIB_API int LZ4_compress_HC_destSize(void* stateHC,
+                                  const char* src, char* dst,
+                                        int* srcSizePtr, int targetDstSize,
+                                        int compressionLevel);
+
+
+/*-************************************
+ *  Streaming Compression
+ *  Bufferless synchronous API
+ **************************************/
+ typedef union LZ4_streamHC_u LZ4_streamHC_t;   /* incomplete type (defined later) */
+
+/*! LZ4_createStreamHC() and LZ4_freeStreamHC() :
+ *  These functions create and release memory for LZ4 HC streaming state.
+ *  Newly created states are automatically initialized.
+ *  A same state can be used multiple times consecutively,
+ *  starting with LZ4_resetStreamHC_fast() to start a new stream of blocks.
+ */
+LZ4LIB_API LZ4_streamHC_t* LZ4_createStreamHC(void);
+LZ4LIB_API int             LZ4_freeStreamHC (LZ4_streamHC_t* streamHCPtr);
+
+/*
+  These functions compress data in successive blocks of any size,
+  using previous blocks as dictionary, to improve compression ratio.
+  One key assumption is that previous blocks (up to 64 KB) remain read-accessible while compressing next blocks.
+  There is an exception for ring buffers, which can be smaller than 64 KB.
+  Ring-buffer scenario is automatically detected and handled within LZ4_compress_HC_continue().
+
+  Before starting compression, state must be allocated and properly initialized.
+  LZ4_createStreamHC() does both, though compression level is set to LZ4HC_CLEVEL_DEFAULT.
+
+  Selecting the compression level can be done with LZ4_resetStreamHC_fast() (starts a new stream)
+  or LZ4_setCompressionLevel() (anytime, between blocks in the same stream) (experimental).
+  LZ4_resetStreamHC_fast() only works on states which have been properly initialized at least once,
+  which is automatically the case when state is created using LZ4_createStreamHC().
+
+  After reset, a first "fictional block" can be designated as initial dictionary,
+  using LZ4_loadDictHC() (Optional).
+
+  Invoke LZ4_compress_HC_continue() to compress each successive block.
+  The number of blocks is unlimited.
+  Previous input blocks, including initial dictionary when present,
+  must remain accessible and unmodified during compression.
+
+  It's allowed to update compression level anytime between blocks,
+  using LZ4_setCompressionLevel() (experimental).
+
+  'dst' buffer should be sized to handle worst case scenarios
+  (see LZ4_compressBound(), it ensures compression success).
+  In case of failure, the API does not guarantee recovery,
+  so the state _must_ be reset.
+  To ensure compression success
+  whenever `dst` buffer size cannot be made >= LZ4_compressBound(),
+  consider using LZ4_compress_HC_continue_destSize().
+
+  Whenever previous input blocks can't be preserved unmodified in-place during compression of next blocks,
+  it's possible to copy the last blocks into a more stable memory space, using LZ4_saveDictHC().
+  Return value of LZ4_saveDictHC() is the size of dictionary effectively saved into 'safeBuffer' (<= 64 KB)
+
+  After completing a streaming compression,
+  it's possible to start a new stream of blocks, using the same LZ4_streamHC_t state,
+  just by resetting it, using LZ4_resetStreamHC_fast().
+*/
+
+LZ4LIB_API void LZ4_resetStreamHC_fast(LZ4_streamHC_t* streamHCPtr, int compressionLevel);   /* v1.9.0+ */
+LZ4LIB_API int  LZ4_loadDictHC (LZ4_streamHC_t* streamHCPtr, const char* dictionary, int dictSize);
+
+LZ4LIB_API int LZ4_compress_HC_continue (LZ4_streamHC_t* streamHCPtr,
+                                   const char* src, char* dst,
+                                         int srcSize, int maxDstSize);
+
+/*! LZ4_compress_HC_continue_destSize() : v1.9.0+
+ *  Similar to LZ4_compress_HC_continue(),
+ *  but will read as much data as possible from `src`
+ *  to fit into `targetDstSize` budget.
+ *  Result is provided into 2 parts :
+ * @return : the number of bytes written into 'dst' (necessarily <= targetDstSize)
+ *           or 0 if compression fails.
+ * `srcSizePtr` : on success, *srcSizePtr will be updated to indicate how much bytes were read from `src`.
+ *           Note that this function may not consume the entire input.
+ */
+LZ4LIB_API int LZ4_compress_HC_continue_destSize(LZ4_streamHC_t* LZ4_streamHCPtr,
+                                           const char* src, char* dst,
+                                                 int* srcSizePtr, int targetDstSize);
+
+LZ4LIB_API int LZ4_saveDictHC (LZ4_streamHC_t* streamHCPtr, char* safeBuffer, int maxDictSize);
+
+
+
+/*^**********************************************
+ * !!!!!!   STATIC LINKING ONLY   !!!!!!
+ ***********************************************/
+
+/*-******************************************************************
+ * PRIVATE DEFINITIONS :
+ * Do not use these definitions directly.
+ * They are merely exposed to allow static allocation of `LZ4_streamHC_t`.
+ * Declare an `LZ4_streamHC_t` directly, rather than any type below.
+ * Even then, only do so in the context of static linking, as definitions may change between versions.
+ ********************************************************************/
+
+#define LZ4HC_DICTIONARY_LOGSIZE 16
+#define LZ4HC_MAXD (1<<LZ4HC_DICTIONARY_LOGSIZE)
+#define LZ4HC_MAXD_MASK (LZ4HC_MAXD - 1)
+
+#define LZ4HC_HASH_LOG 15
+#define LZ4HC_HASHTABLESIZE (1 << LZ4HC_HASH_LOG)
+#define LZ4HC_HASH_MASK (LZ4HC_HASHTABLESIZE - 1)
+
+
+typedef struct LZ4HC_CCtx_internal LZ4HC_CCtx_internal;
+struct LZ4HC_CCtx_internal
+{
+    LZ4_u32   hashTable[LZ4HC_HASHTABLESIZE];
+    LZ4_u16   chainTable[LZ4HC_MAXD];
+    const LZ4_byte* end;       /* next block here to continue on current prefix */
+    const LZ4_byte* base;      /* All index relative to this position */
+    const LZ4_byte* dictBase;  /* alternate base for extDict */
+    LZ4_u32   dictLimit;       /* below that point, need extDict */
+    LZ4_u32   lowLimit;        /* below that point, no more dict */
+    LZ4_u32   nextToUpdate;    /* index from which to continue dictionary update */
+    short     compressionLevel;
+    LZ4_i8    favorDecSpeed;   /* favor decompression speed if this flag set,
+                                  otherwise, favor compression ratio */
+    LZ4_i8    dirty;           /* stream has to be fully reset if this flag is set */
+    const LZ4HC_CCtx_internal* dictCtx;
+};
+
+
+/* Do not use these definitions directly !
+ * Declare or allocate an LZ4_streamHC_t instead.
+ */
+#define LZ4_STREAMHCSIZE       262200  /* static size, for inter-version compatibility */
+#define LZ4_STREAMHCSIZE_VOIDP (LZ4_STREAMHCSIZE / sizeof(void*))
+union LZ4_streamHC_u {
+    void* table[LZ4_STREAMHCSIZE_VOIDP];
+    LZ4HC_CCtx_internal internal_donotuse;
+}; /* previously typedef'd to LZ4_streamHC_t */
+
+/* LZ4_streamHC_t :
+ * This structure allows static allocation of LZ4 HC streaming state.
+ * This can be used to allocate statically, on state, or as part of a larger structure.
+ *
+ * Such state **must** be initialized using LZ4_initStreamHC() before first use.
+ *
+ * Note that invoking LZ4_initStreamHC() is not required when
+ * the state was created using LZ4_createStreamHC() (which is recommended).
+ * Using the normal builder, a newly created state is automatically initialized.
+ *
+ * Static allocation shall only be used in combination with static linking.
+ */
+
+/* LZ4_initStreamHC() : v1.9.0+
+ * Required before first use of a statically allocated LZ4_streamHC_t.
+ * Before v1.9.0 : use LZ4_resetStreamHC() instead
+ */
+LZ4LIB_API LZ4_streamHC_t* LZ4_initStreamHC (void* buffer, size_t size);
+
+
+/*-************************************
+*  Deprecated Functions
+**************************************/
+/* see lz4.h LZ4_DISABLE_DEPRECATE_WARNINGS to turn off deprecation warnings */
+
+/* deprecated compression functions */
+LZ4_DEPRECATED("use LZ4_compress_HC() instead") LZ4LIB_API int LZ4_compressHC               (const char* source, char* dest, int inputSize);
+LZ4_DEPRECATED("use LZ4_compress_HC() instead") LZ4LIB_API int LZ4_compressHC_limitedOutput (const char* source, char* dest, int inputSize, int maxOutputSize);
+LZ4_DEPRECATED("use LZ4_compress_HC() instead") LZ4LIB_API int LZ4_compressHC2              (const char* source, char* dest, int inputSize, int compressionLevel);
+LZ4_DEPRECATED("use LZ4_compress_HC() instead") LZ4LIB_API int LZ4_compressHC2_limitedOutput(const char* source, char* dest, int inputSize, int maxOutputSize, int compressionLevel);
+LZ4_DEPRECATED("use LZ4_compress_HC_extStateHC() instead") LZ4LIB_API int LZ4_compressHC_withStateHC               (void* state, const char* source, char* dest, int inputSize);
+LZ4_DEPRECATED("use LZ4_compress_HC_extStateHC() instead") LZ4LIB_API int LZ4_compressHC_limitedOutput_withStateHC (void* state, const char* source, char* dest, int inputSize, int maxOutputSize);
+LZ4_DEPRECATED("use LZ4_compress_HC_extStateHC() instead") LZ4LIB_API int LZ4_compressHC2_withStateHC              (void* state, const char* source, char* dest, int inputSize, int compressionLevel);
+LZ4_DEPRECATED("use LZ4_compress_HC_extStateHC() instead") LZ4LIB_API int LZ4_compressHC2_limitedOutput_withStateHC(void* state, const char* source, char* dest, int inputSize, int maxOutputSize, int compressionLevel);
+LZ4_DEPRECATED("use LZ4_compress_HC_continue() instead") LZ4LIB_API int LZ4_compressHC_continue               (LZ4_streamHC_t* LZ4_streamHCPtr, const char* source, char* dest, int inputSize);
+LZ4_DEPRECATED("use LZ4_compress_HC_continue() instead") LZ4LIB_API int LZ4_compressHC_limitedOutput_continue (LZ4_streamHC_t* LZ4_streamHCPtr, const char* source, char* dest, int inputSize, int maxOutputSize);
+
+/* Obsolete streaming functions; degraded functionality; do not use!
+ *
+ * In order to perform streaming compression, these functions depended on data
+ * that is no longer tracked in the state. They have been preserved as well as
+ * possible: using them will still produce a correct output. However, use of
+ * LZ4_slideInputBufferHC() will truncate the history of the stream, rather
+ * than preserve a window-sized chunk of history.
+ */
+LZ4_DEPRECATED("use LZ4_createStreamHC() instead") LZ4LIB_API void* LZ4_createHC (const char* inputBuffer);
+LZ4_DEPRECATED("use LZ4_saveDictHC() instead") LZ4LIB_API     char* LZ4_slideInputBufferHC (void* LZ4HC_Data);
+LZ4_DEPRECATED("use LZ4_freeStreamHC() instead") LZ4LIB_API   int   LZ4_freeHC (void* LZ4HC_Data);
+LZ4_DEPRECATED("use LZ4_compress_HC_continue() instead") LZ4LIB_API int LZ4_compressHC2_continue               (void* LZ4HC_Data, const char* source, char* dest, int inputSize, int compressionLevel);
+LZ4_DEPRECATED("use LZ4_compress_HC_continue() instead") LZ4LIB_API int LZ4_compressHC2_limitedOutput_continue (void* LZ4HC_Data, const char* source, char* dest, int inputSize, int maxOutputSize, int compressionLevel);
+LZ4_DEPRECATED("use LZ4_createStreamHC() instead") LZ4LIB_API int   LZ4_sizeofStreamStateHC(void);
+LZ4_DEPRECATED("use LZ4_initStreamHC() instead") LZ4LIB_API  int   LZ4_resetStreamStateHC(void* state, char* inputBuffer);
+
+
+/* LZ4_resetStreamHC() is now replaced by LZ4_initStreamHC().
+ * The intention is to emphasize the difference with LZ4_resetStreamHC_fast(),
+ * which is now the recommended function to start a new stream of blocks,
+ * but cannot be used to initialize a memory segment containing arbitrary garbage data.
+ *
+ * It is recommended to switch to LZ4_initStreamHC().
+ * LZ4_resetStreamHC() will generate deprecation warnings in a future version.
+ */
+LZ4LIB_API void LZ4_resetStreamHC (LZ4_streamHC_t* streamHCPtr, int compressionLevel);
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* LZ4_HC_H_19834876238432 */
+
+
+/*-**************************************************
+ * !!!!!     STATIC LINKING ONLY     !!!!!
+ * Following definitions are considered experimental.
+ * They should not be linked from DLL,
+ * as there is no guarantee of API stability yet.
+ * Prototypes will be promoted to "stable" status
+ * after successfull usage in real-life scenarios.
+ ***************************************************/
+#ifdef LZ4_HC_STATIC_LINKING_ONLY   /* protection macro */
+#ifndef LZ4_HC_SLO_098092834
+#define LZ4_HC_SLO_098092834
+
+#define LZ4_STATIC_LINKING_ONLY   /* LZ4LIB_STATIC_API */
+#include "lz4.h"
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*! LZ4_setCompressionLevel() : v1.8.0+ (experimental)
+ *  It's possible to change compression level
+ *  between successive invocations of LZ4_compress_HC_continue*()
+ *  for dynamic adaptation.
+ */
+LZ4LIB_STATIC_API void LZ4_setCompressionLevel(
+    LZ4_streamHC_t* LZ4_streamHCPtr, int compressionLevel);
+
+/*! LZ4_favorDecompressionSpeed() : v1.8.2+ (experimental)
+ *  Opt. Parser will favor decompression speed over compression ratio.
+ *  Only applicable to levels >= LZ4HC_CLEVEL_OPT_MIN.
+ */
+LZ4LIB_STATIC_API void LZ4_favorDecompressionSpeed(
+    LZ4_streamHC_t* LZ4_streamHCPtr, int favor);
+
+/*! LZ4_resetStreamHC_fast() : v1.9.0+
+ *  When an LZ4_streamHC_t is known to be in a internally coherent state,
+ *  it can often be prepared for a new compression with almost no work, only
+ *  sometimes falling back to the full, expensive reset that is always required
+ *  when the stream is in an indeterminate state (i.e., the reset performed by
+ *  LZ4_resetStreamHC()).
+ *
+ *  LZ4_streamHCs are guaranteed to be in a valid state when:
+ *  - returned from LZ4_createStreamHC()
+ *  - reset by LZ4_resetStreamHC()
+ *  - memset(stream, 0, sizeof(LZ4_streamHC_t))
+ *  - the stream was in a valid state and was reset by LZ4_resetStreamHC_fast()
+ *  - the stream was in a valid state and was then used in any compression call
+ *    that returned success
+ *  - the stream was in an indeterminate state and was used in a compression
+ *    call that fully reset the state (LZ4_compress_HC_extStateHC()) and that
+ *    returned success
+ *
+ *  Note:
+ *  A stream that was last used in a compression call that returned an error
+ *  may be passed to this function. However, it will be fully reset, which will
+ *  clear any existing history and settings from the context.
+ */
+LZ4LIB_STATIC_API void LZ4_resetStreamHC_fast(
+    LZ4_streamHC_t* LZ4_streamHCPtr, int compressionLevel);
+
+/*! LZ4_compress_HC_extStateHC_fastReset() :
+ *  A variant of LZ4_compress_HC_extStateHC().
+ *
+ *  Using this variant avoids an expensive initialization step. It is only safe
+ *  to call if the state buffer is known to be correctly initialized already
+ *  (see above comment on LZ4_resetStreamHC_fast() for a definition of
+ *  "correctly initialized"). From a high level, the difference is that this
+ *  function initializes the provided state with a call to
+ *  LZ4_resetStreamHC_fast() while LZ4_compress_HC_extStateHC() starts with a
+ *  call to LZ4_resetStreamHC().
+ */
+LZ4LIB_STATIC_API int LZ4_compress_HC_extStateHC_fastReset (
+    void* state,
+    const char* src, char* dst,
+    int srcSize, int dstCapacity,
+    int compressionLevel);
+
+/*! LZ4_attach_HC_dictionary() :
+ *  This is an experimental API that allows for the efficient use of a
+ *  static dictionary many times.
+ *
+ *  Rather than re-loading the dictionary buffer into a working context before
+ *  each compression, or copying a pre-loaded dictionary's LZ4_streamHC_t into a
+ *  working LZ4_streamHC_t, this function introduces a no-copy setup mechanism,
+ *  in which the working stream references the dictionary stream in-place.
+ *
+ *  Several assumptions are made about the state of the dictionary stream.
+ *  Currently, only streams which have been prepared by LZ4_loadDictHC() should
+ *  be expected to work.
+ *
+ *  Alternatively, the provided dictionary stream pointer may be NULL, in which
+ *  case any existing dictionary stream is unset.
+ *
+ *  A dictionary should only be attached to a stream without any history (i.e.,
+ *  a stream that has just been reset).
+ *
+ *  The dictionary will remain attached to the working stream only for the
+ *  current stream session. Calls to LZ4_resetStreamHC(_fast) will remove the
+ *  dictionary context association from the working stream. The dictionary
+ *  stream (and source buffer) must remain in-place / accessible / unchanged
+ *  through the lifetime of the stream session.
+ */
+LZ4LIB_STATIC_API void LZ4_attach_HC_dictionary(
+          LZ4_streamHC_t *working_stream,
+    const LZ4_streamHC_t *dictionary_stream);
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif   /* LZ4_HC_SLO_098092834 */
+#endif   /* LZ4_HC_STATIC_LINKING_ONLY */
diff --git a/mfbt/lz4/xxhash.c b/mfbt/lz4/xxhash.c
new file mode 100644
index 0000000000..0fae88c5d6
--- /dev/null
+++ b/mfbt/lz4/xxhash.c
@@ -0,0 +1,43 @@
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Copyright (C) 2012-2020 Yann Collet
+ *
+ * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *    * Redistributions of source code must retain the above copyright
+ *      notice, this list of conditions and the following disclaimer.
+ *    * Redistributions in binary form must reproduce the above
+ *      copyright notice, this list of conditions and the following disclaimer
+ *      in the documentation and/or other materials provided with the
+ *      distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * You can contact the author at:
+ *   - xxHash homepage: https://www.xxhash.com
+ *   - 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 */
+
+#include "xxhash.h"
diff --git a/mfbt/lz4/xxhash.h b/mfbt/lz4/xxhash.h
new file mode 100644
index 0000000000..2d56d23c5d
--- /dev/null
+++ b/mfbt/lz4/xxhash.h
@@ -0,0 +1,4766 @@
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Header File
+ * Copyright (C) 2012-2020 Yann Collet
+ *
+ * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *    * Redistributions of source code must retain the above copyright
+ *      notice, this list of conditions and the following disclaimer.
+ *    * Redistributions in binary form must reproduce the above
+ *      copyright notice, this list of conditions and the following disclaimer
+ *      in the documentation and/or other materials provided with the
+ *      distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * You can contact the author at:
+ *   - xxHash homepage: https://www.xxhash.com
+ *   - 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
+*/
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* ****************************
+ *  INLINE mode
+ ******************************/
+/*!
+ * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * 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
+ *
+ * It also keeps xxHash symbols private to the unit, so they are not exported.
+ *
+ * Usage:
+ *     #define XXH_INLINE_ALL
+ *     #include "xxhash.h"
+ *
+ * Do not compile and link xxhash.o as a separate object, as it is not useful.
+ */
+#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
+    && !defined(XXH_INLINE_ALL_31684351384)
+   /* this section should be traversed only once */
+#  define XXH_INLINE_ALL_31684351384
+   /* give access to the advanced API, required to compile implementations */
+#  undef XXH_STATIC_LINKING_ONLY   /* avoid macro redef */
+#  define XXH_STATIC_LINKING_ONLY
+   /* make all functions private */
+#  undef XXH_PUBLIC_API
+#  if defined(__GNUC__)
+#    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)
+#    define XXH_PUBLIC_API static __inline
+#  else
+     /* note: this version may generate warnings for unused static functions */
+#    define XXH_PUBLIC_API static
+#  endif
+
+   /*
+    * 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.
+    * 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
+#  define XXH_NAMESPACE XXH_INLINE_
+   /*
+    * Some identifiers (enums, type names) are not symbols, but they must
+    * still 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
+    */
+#  define XXH_IPREF(Id)   XXH_INLINE_ ## Id
+#  define XXH_OK XXH_IPREF(XXH_OK)
+#  define XXH_ERROR XXH_IPREF(XXH_ERROR)
+#  define XXH_errorcode XXH_IPREF(XXH_errorcode)
+#  define XXH32_canonical_t  XXH_IPREF(XXH32_canonical_t)
+#  define XXH64_canonical_t  XXH_IPREF(XXH64_canonical_t)
+#  define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t)
+#  define XXH32_state_s XXH_IPREF(XXH32_state_s)
+#  define XXH32_state_t XXH_IPREF(XXH32_state_t)
+#  define XXH64_state_s XXH_IPREF(XXH64_state_s)
+#  define XXH64_state_t XXH_IPREF(XXH64_state_t)
+#  define XXH3_state_s  XXH_IPREF(XXH3_state_s)
+#  define XXH3_state_t  XXH_IPREF(XXH3_state_t)
+#  define XXH128_hash_t XXH_IPREF(XXH128_hash_t)
+   /* Ensure the header is parsed again, even if it was previously included */
+#  undef XXHASH_H_5627135585666179
+#  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 */
+#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
+
+/*!
+ * 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)
+#  define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
+#  define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
+#  define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
+#  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)
+#  define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
+#  define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
+#  define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
+#  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_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_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)
+/* 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_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)
+#  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
+
+
+/* *************************************
+*  Version
+***************************************/
+#define XXH_VERSION_MAJOR    0
+#define XXH_VERSION_MINOR    8
+#define XXH_VERSION_RELEASE  0
+#define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
+XXH_PUBLIC_API unsigned XXH_versionNumber (void);
+
+
+/* ****************************
+*  Definitions
+******************************/
+#include <stddef.h>   /* size_t */
+typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
+
+
+/*-**********************************************************************
+*  32-bit hash
+************************************************************************/
+#if !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   include <stdint.h>
+    typedef uint32_t XXH32_hash_t;
+#else
+#   include <limits.h>
+#   if UINT_MAX == 0xFFFFFFFFUL
+      typedef unsigned int XXH32_hash_t;
+#   else
+#     if ULONG_MAX == 0xFFFFFFFFUL
+        typedef unsigned long XXH32_hash_t;
+#     else
+#       error "unsupported platform: need a 32-bit type"
+#     endif
+#   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
+ *
+ * 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.
+ */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+
+/*******   Streaming   *******/
+
+/*
+ * 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.
+ *
+ * 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()`.
+ */
+
+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);
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
+
+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   *******/
+
+/*
+ * The default return values from XXH functions are unsigned 32 and 64 bit
+ * integers.
+ * This the simplest and fastest format for further post-processing.
+ *
+ * However, this leaves open the question of what is the order on the byte level,
+ * since little and big endian conventions will store the same number differently.
+ *
+ * The canonical representation settles this issue by mandating big-endian
+ * convention, the same convention as human-readable numbers (large digits first).
+ *
+ * When writing hash values to storage, sending them over a network, or printing
+ * them, it's highly recommended to use the canonical representation to ensure
+ * portability across a wider range of systems, present and future.
+ *
+ * The following functions allow transformation of hash values to and from
+ * canonical format.
+ */
+
+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);
+
+
+#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
+
+/*!
+ * 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.
+ *
+ * This function usually runs faster on 64-bit systems, but slower on 32-bit
+ * systems (see benchmark).
+ *
+ * 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.
+ */
+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[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src);
+
+
+/*-**********************************************************************
+*  XXH3 64-bit variant
+************************************************************************/
+
+/* ************************************************************************
+ * 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
+ *
+ * Speed analysis methodology is explained here:
+ *
+ *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ *
+ * 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.
+ *
+ * The algorithm is portable: Like XXH32 and XXH64, it generates the same hash
+ * on all platforms.
+ *
+ * It benefits greatly from SIMD and 64-bit arithmetic, but does not require it.
+ *
+ * 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.
+ *
+ * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8,
+ * ZVector and scalar targets. This can be controlled with the XXH_VECTOR macro.
+ *
+ * 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.
+ *
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The 128-bit version adds additional strength, but it is slightly slower.
+ *
+ * The XXH3 algorithm is still in development.
+ * The results it produces may still change in future versions.
+ *
+ * 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).
+ *
+ * Avoid storing values in long-term storage until the algorithm is finalized.
+ * XXH3's return values will be officially finalized upon reaching v0.8.0.
+ *
+ * After which, return values of XXH3 and XXH128 will no longer change in
+ * future versions.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
+ */
+
+/* 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);
+
+/*
+ * XXH3_64bits_withSeed():
+ * This variant generates a custom secret on the fly
+ * based on default secret altered using the `seed` value.
+ * While this operation is decently fast, note that it's not completely free.
+ * Note: seed==0 produces the same results as XXH3_64bits().
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+
+/*
+ * 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 main condition is that secretSize *must* be large enough (>= XXH3_SECRET_SIZE_MIN).
+ * However, the quality of produced hash values depends on secret's entropy.
+ * Technically, the secret must look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever unsure about the "randomness" of the blob of bytes,
+ * consider relabelling it as a "custom seed" instead,
+ * and employ "XXH3_generateSecret()" (see below)
+ * to generate a high entropy secret derived from the custom seed.
+ */
+#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);
+
+
+/*******   Streaming   *******/
+/*
+ * 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.
+ */
+typedef struct XXH3_state_s XXH3_state_t;
+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 default parameters.
+ * digest 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()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*
+ * XXH3_64bits_reset_withSecret():
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_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);
+
+/* note : canonical representation of XXH3 is the same as XXH64
+ * since they both produce XXH64_hash_t values */
+
+
+/*-**********************************************************************
+*  XXH3 128-bit variant
+************************************************************************/
+
+typedef struct {
+ XXH64_hash_t low64;
+ XXH64_hash_t high64;
+} XXH128_hash_t;
+
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+
+/*******   Streaming   *******/
+/*
+ * 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.
+ */
+
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_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);
+
+/* 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.
+ */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+
+/*!
+ * XXH128_cmp():
+ *
+ * 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
+ */
+XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2);
+
+
+/*******   Canonical representation   *******/
+typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash);
+XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src);
+
+
+#endif  /* XXH_NO_LONG_LONG */
+
+#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.
+ */
+
+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 */
+
+
+#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+
+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 */
+
+#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
+
+/* 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
+
+#define XXH3_INTERNALBUFFER_SIZE 256
+#define XXH3_SECRET_DEFAULT_SIZE 192
+struct XXH3_state_s {
+   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
+   /* used to store a custom secret generated from a seed */
+   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 reserved32;
+   size_t nbStripesSoFar;
+   XXH64_hash_t totalLen;
+   size_t nbStripesPerBlock;
+   size_t secretLimit;
+   XXH64_hash_t seed;
+   XXH64_hash_t reserved64;
+   const unsigned char* extSecret;  /* reference to external secret;
+                                     * if == NULL, use .customSecret instead */
+   /* 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
+
+/* When the 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)   { (XXH3_state_ptr)->seed = 0; }
+
+
+/* ===   Experimental API   === */
+/* Symbols defined below must be considered tied to a specific library version. */
+
+/*
+ * XXH3_generateSecret():
+ *
+ * Derive a high-entropy secret from any user-defined content, named customSeed.
+ * The generated secret can be used in combination with `*_withSecret()` functions.
+ * The `_withSecret()` variants are useful to provide a higher level of protection than 64-bit seed,
+ * as it becomes much more difficult for an external actor to guess how to impact the calculation logic.
+ *
+ * The function accepts as input a custom seed of any length and any content,
+ * and derives from it a high-entropy secret of length XXH3_SECRET_DEFAULT_SIZE
+ * into an already allocated buffer secretBuffer.
+ * The generated secret is _always_ XXH_SECRET_DEFAULT_SIZE bytes long.
+ *
+ * The generated secret can then be used with any `*_withSecret()` variant.
+ * Functions `XXH3_128bits_withSecret()`, `XXH3_64bits_withSecret()`,
+ * `XXH3_128bits_reset_withSecret()` and `XXH3_64bits_reset_withSecret()`
+ * are part of this list. They all accept a `secret` parameter
+ * which must be very long for implementation reasons (>= XXH3_SECRET_SIZE_MIN)
+ * _and_ feature very high entropy (consist of random-looking bytes).
+ * These conditions can be a high bar to meet, so
+ * this function can be used to generate a secret of proper quality.
+ *
+ * customSeed can be anything. It can have any size, even small ones,
+ * and its content can be anything, even stupidly "low entropy" source such as a bunch of zeroes.
+ * The resulting `secret` will nonetheless provide all expected qualities.
+ *
+ * Supplying NULL as the customSeed copies the default secret into `secretBuffer`.
+ * When customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ */
+XXH_PUBLIC_API void XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSeedSize);
+
+
+/* simple short-cut to pre-selected XXH3_128bits variant */
+XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed);
+
+
+#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
+***************************************/
+/*!
+ * XXH_FORCE_MEMORY_ACCESS:
+ * 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.
+ * Method 0 (default):
+ *     Use `memcpy()`. Safe and portable. Default.
+ * 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 (example: 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)
+ */
+#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
+#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
+#endif
+
+/*!
+ * XXH_FORCE_ALIGN_CHECK:
+ * 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 read,
+ * 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 binary for an additional code path
+ * if memory access uses same instruction for both aligned and unaligned adresses.
+ *
+ * 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,
+ * which are platforms known to offer good unaligned memory accesses performance.
+ *
+ * This option does not affect XXH3 (only XXH32 and XXH64).
+ */
+#ifndef XXH_FORCE_ALIGN_CHECK  /* can be defined externally */
+#  if defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) \
+   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64) /* visual */
+#    define XXH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+/*!
+ * XXH_NO_INLINE_HINTS:
+ *
+ * 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), 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.
+ *
+ * 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
+ * different memory routines for malloc() and free()
+ */
+#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)
+{
+    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__)
+#    define XXH_FORCE_INLINE static __attribute__((unused))
+#  else
+#    define XXH_FORCE_INLINE static
+#  endif
+#  define XXH_NO_INLINE static
+/* enable inlining hints */
+#elif defined(_MSC_VER)  /* Visual Studio */
+#  define XXH_FORCE_INLINE static __forceinline
+#  define XXH_NO_INLINE static __declspec(noinline)
+#elif defined(__GNUC__)
+#  define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused))
+#  define XXH_NO_INLINE static __attribute__((noinline))
+#elif defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* C99 */
+#  define XXH_FORCE_INLINE static inline
+#  define XXH_NO_INLINE static
+#else
+#  define XXH_FORCE_INLINE static
+#  define XXH_NO_INLINE static
+#endif
+
+
+
+/* *************************************
+*  Debug
+***************************************/
+/*
+ * 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
+#  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)
+
+
+/* *************************************
+*  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;
+
+#ifdef XXH_OLD_NAMES
+#  define BYTE xxh_u8
+#  define U8   xxh_u8
+#  define U32  xxh_u32
+#endif
+
+/* ***   Memory access   *** */
+
+#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))
+
+/*
+ * __pack instructions are safer but compiler specific, hence potentially
+ * problematic for some compilers.
+ *
+ * Currently only defined for GCC and ICC.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
+#endif
+static xxh_u32 XXH_read32(const void* ptr)
+{
+    typedef union { xxh_u32 u32; } __attribute__((packed)) xxh_unalign;
+    return ((const xxh_unalign*)ptr)->u32;
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://stackoverflow.com/a/32095106/646947
+ */
+static xxh_u32 XXH_read32(const void* memPtr)
+{
+    xxh_u32 val;
+    memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ***   Endianess   *** */
+typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess;
+
+/*!
+ * 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.
+ *
+ * If it is not defined, a runtime check (which is usually constant folded)
+ * is used instead.
+ */
+#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
+/*
+ * runtime test, presumed to simplify to a constant by compiler
+ */
+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
+
+#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
+/* 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
+
+#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
+*****************************/
+typedef enum { XXH_aligned, XXH_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
+***************************************/
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* *******************************************************************
+*  32-bit hash functions
+*********************************************************************/
+static const xxh_u32 XXH_PRIME32_1 = 0x9E3779B1U;   /* 0b10011110001101110111100110110001 */
+static const xxh_u32 XXH_PRIME32_2 = 0x85EBCA77U;   /* 0b10000101111010111100101001110111 */
+static const xxh_u32 XXH_PRIME32_3 = 0xC2B2AE3DU;   /* 0b11000010101100101010111000111101 */
+static const xxh_u32 XXH_PRIME32_4 = 0x27D4EB2FU;   /* 0b00100111110101001110101100101111 */
+static const xxh_u32 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
+
+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(__GNUC__) && defined(__SSE4_1__) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * 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.
+     */
+    __asm__("" : "+r" (acc));
+#endif
+    return acc;
+}
+
+/* mix all bits */
+static xxh_u32 XXH32_avalanche(xxh_u32 h32)
+{
+    h32 ^= h32 >> 15;
+    h32 *= XXH_PRIME32_2;
+    h32 ^= h32 >> 13;
+    h32 *= XXH_PRIME32_3;
+    h32 ^= h32 >> 16;
+    return(h32);
+}
+
+#define XXH_get32bits(p) XXH_readLE32_align(p, align)
+
+static xxh_u32
+XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+#define XXH_PROCESS1 do {                           \
+    h32 += (*ptr++) * XXH_PRIME32_5;                \
+    h32 = XXH_rotl32(h32, 11) * XXH_PRIME32_1;      \
+} while (0)
+
+#define XXH_PROCESS4 do {                           \
+    h32 += XXH_get32bits(ptr) * XXH_PRIME32_3;      \
+    ptr += 4;                                   \
+    h32  = XXH_rotl32(h32, 17) * XXH_PRIME32_4;     \
+} while (0)
+
+    /* Compact rerolled version */
+    if (XXH_REROLL) {
+        len &= 15;
+        while (len >= 4) {
+            XXH_PROCESS4;
+            len -= 4;
+        }
+        while (len > 0) {
+            XXH_PROCESS1;
+            --len;
+        }
+        return XXH32_avalanche(h32);
+    } else {
+         switch(len&15) /* or switch(bEnd - p) */ {
+           case 12:      XXH_PROCESS4;
+                         /* fallthrough */
+           case 8:       XXH_PROCESS4;
+                         /* fallthrough */
+           case 4:       XXH_PROCESS4;
+                         return XXH32_avalanche(h32);
+
+           case 13:      XXH_PROCESS4;
+                         /* fallthrough */
+           case 9:       XXH_PROCESS4;
+                         /* fallthrough */
+           case 5:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(h32);
+
+           case 14:      XXH_PROCESS4;
+                         /* fallthrough */
+           case 10:      XXH_PROCESS4;
+                         /* fallthrough */
+           case 6:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(h32);
+
+           case 15:      XXH_PROCESS4;
+                         /* fallthrough */
+           case 11:      XXH_PROCESS4;
+                         /* fallthrough */
+           case 7:       XXH_PROCESS4;
+                         /* fallthrough */
+           case 3:       XXH_PROCESS1;
+                         /* fallthrough */
+           case 2:       XXH_PROCESS1;
+                         /* fallthrough */
+           case 1:       XXH_PROCESS1;
+                         /* fallthrough */
+           case 0:       return XXH32_avalanche(h32);
+        }
+        XXH_ASSERT(0);
+        return h32;   /* 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
+
+XXH_FORCE_INLINE xxh_u32
+XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+{
+    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 + XXH_PRIME32_1 + XXH_PRIME32_2;
+        xxh_u32 v2 = seed + XXH_PRIME32_2;
+        xxh_u32 v3 = seed + 0;
+        xxh_u32 v4 = seed - XXH_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 + XXH_PRIME32_5;
+    }
+
+    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)
+{
+#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
+}
+
+
+
+/*******   Hash streaming   *******/
+
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
+{
+    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
+}
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
+{
+    memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
+{
+    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
+    memset(&state, 0, sizeof(state));
+    state.v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+    state.v2 = seed + XXH_PRIME32_2;
+    state.v3 = seed + 0;
+    state.v4 = seed - XXH_PRIME32_1;
+    /* do not write into reserved, planned to be removed in a future version */
+    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved));
+    return XXH_OK;
+}
+
+
+XXH_PUBLIC_API XXH_errorcode
+XXH32_update(XXH32_state_t* state, const void* input, size_t len)
+{
+    if (input==NULL)
+#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+        return XXH_OK;
+#else
+        return XXH_ERROR;
+#endif
+
+    {   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;
+            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;
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* state)
+{
+    xxh_u32 h32;
+
+    if (state->large_len) {
+        h32 = XXH_rotl32(state->v1, 1)
+            + XXH_rotl32(state->v2, 7)
+            + XXH_rotl32(state->v3, 12)
+            + XXH_rotl32(state->v4, 18);
+    } else {
+        h32 = state->v3 /* == seed */ + XXH_PRIME32_5;
+    }
+
+    h32 += state->total_len_32;
+
+    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+}
+
+
+/*******   Canonical representation   *******/
+
+/*
+ * 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.
+ */
+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);
+    memcpy(dst, &hash, sizeof(*dst));
+}
+
+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
+*********************************************************************/
+
+/*******   Memory access   *******/
+
+typedef XXH64_hash_t xxh_u64;
+
+#ifdef XXH_OLD_NAMES
+#  define U64 xxh_u64
+#endif
+
+/*!
+ * 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) */
+
+#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))
+
+/*
+ * __pack instructions are safer, but compiler specific, hence potentially
+ * problematic for some compilers.
+ *
+ * Currently only defined for GCC and ICC.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
+#endif
+static xxh_u64 XXH_read64(const void* ptr)
+{
+    typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) xxh_unalign64;
+    return ((const xxh_unalign64*)ptr)->u64;
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://stackoverflow.com/a/32095106/646947
+ */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    xxh_u64 val;
+    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   *******/
+
+static const xxh_u64 XXH_PRIME64_1 = 0x9E3779B185EBCA87ULL;   /* 0b1001111000110111011110011011000110000101111010111100101010000111 */
+static const xxh_u64 XXH_PRIME64_2 = 0xC2B2AE3D27D4EB4FULL;   /* 0b1100001010110010101011100011110100100111110101001110101101001111 */
+static const xxh_u64 XXH_PRIME64_3 = 0x165667B19E3779F9ULL;   /* 0b0001011001010110011001111011000110011110001101110111100111111001 */
+static const xxh_u64 XXH_PRIME64_4 = 0x85EBCA77C2B2AE63ULL;   /* 0b1000010111101011110010100111011111000010101100101010111001100011 */
+static const xxh_u64 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
+
+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;
+    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;
+}
+
+static xxh_u64 XXH64_avalanche(xxh_u64 h64)
+{
+    h64 ^= h64 >> 33;
+    h64 *= XXH_PRIME64_2;
+    h64 ^= h64 >> 29;
+    h64 *= XXH_PRIME64_3;
+    h64 ^= h64 >> 32;
+    return h64;
+}
+
+
+#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)
+{
+#define XXH_PROCESS1_64 do {                                   \
+    h64 ^= (*ptr++) * XXH_PRIME64_5;                           \
+    h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;                 \
+} while (0)
+
+#define XXH_PROCESS4_64 do {                                   \
+    h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;      \
+    ptr += 4;                                              \
+    h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;     \
+} while (0)
+
+#define XXH_PROCESS8_64 do {                                   \
+    xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr)); \
+    ptr += 8;                                              \
+    h64 ^= k1;                                             \
+    h64  = XXH_rotl64(h64,27) * XXH_PRIME64_1 + XXH_PRIME64_4;     \
+} while (0)
+
+    /* Rerolled version for 32-bit targets is faster and much smaller. */
+    if (XXH_REROLL || XXH_REROLL_XXH64) {
+        len &= 31;
+        while (len >= 8) {
+            XXH_PROCESS8_64;
+            len -= 8;
+        }
+        if (len >= 4) {
+            XXH_PROCESS4_64;
+            len -= 4;
+        }
+        while (len > 0) {
+            XXH_PROCESS1_64;
+            --len;
+        }
+         return  XXH64_avalanche(h64);
+    } else {
+        switch(len & 31) {
+           case 24: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 16: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case  8: XXH_PROCESS8_64;
+                    return XXH64_avalanche(h64);
+
+           case 28: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 20: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 12: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case  4: XXH_PROCESS4_64;
+                    return XXH64_avalanche(h64);
+
+           case 25: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 17: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case  9: XXH_PROCESS8_64;
+                    XXH_PROCESS1_64;
+                    return XXH64_avalanche(h64);
+
+           case 29: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 21: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 13: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case  5: XXH_PROCESS4_64;
+                    XXH_PROCESS1_64;
+                    return XXH64_avalanche(h64);
+
+           case 26: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 18: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 10: XXH_PROCESS8_64;
+                    XXH_PROCESS1_64;
+                    XXH_PROCESS1_64;
+                    return XXH64_avalanche(h64);
+
+           case 30: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 22: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 14: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case  6: XXH_PROCESS4_64;
+                    XXH_PROCESS1_64;
+                    XXH_PROCESS1_64;
+                    return XXH64_avalanche(h64);
+
+           case 27: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 19: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 11: XXH_PROCESS8_64;
+                    XXH_PROCESS1_64;
+                    XXH_PROCESS1_64;
+                    XXH_PROCESS1_64;
+                    return XXH64_avalanche(h64);
+
+           case 31: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 23: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case 15: XXH_PROCESS8_64;
+                         /* fallthrough */
+           case  7: XXH_PROCESS4_64;
+                         /* fallthrough */
+           case  3: XXH_PROCESS1_64;
+                         /* fallthrough */
+           case  2: XXH_PROCESS1_64;
+                         /* fallthrough */
+           case  1: XXH_PROCESS1_64;
+                         /* fallthrough */
+           case  0: return XXH64_avalanche(h64);
+        }
+    }
+    /* impossible to reach */
+    XXH_ASSERT(0);
+    return 0;  /* unreachable, but some compilers complain without it */
+}
+
+#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
+
+XXH_FORCE_INLINE xxh_u64
+XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
+{
+    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 + XXH_PRIME64_1 + XXH_PRIME64_2;
+        xxh_u64 v2 = seed + XXH_PRIME64_2;
+        xxh_u64 v3 = seed + 0;
+        xxh_u64 v4 = seed - XXH_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);
+
+    } else {
+        h64  = seed + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) len;
+
+    return XXH64_finalize(h64, input, len, align);
+}
+
+
+XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t seed)
+{
+#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
+}
+
+/*******   Hash Streaming   *******/
+
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+{
+    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+}
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t* srcState)
+{
+    memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed)
+{
+    XXH64_state_t state;   /* use a local state to memcpy() in order to avoid strict-aliasing warnings */
+    memset(&state, 0, sizeof(state));
+    state.v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+    state.v2 = seed + XXH_PRIME64_2;
+    state.v3 = seed + 0;
+    state.v4 = seed - XXH_PRIME64_1;
+     /* do not write into reserved64, might be removed in a future version */
+    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved64));
+    return XXH_OK;
+}
+
+XXH_PUBLIC_API XXH_errorcode
+XXH64_update (XXH64_state_t* state, const void* input, size_t len)
+{
+    if (input==NULL)
+#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
+        return XXH_OK;
+#else
+        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;
+
+            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;
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    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*/ + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) state->total_len;
+
+    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
+}
+
+
+/******* 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)
+{
+    return XXH_readBE64(src);
+}
+
+
+
+/* *********************************************************************
+*  XXH3
+*  New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+
+/* ===   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__  /* circumvent a 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.
+ *
+ * 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
+ * ========================================== */
+#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_AVX512 3  /* AVX512 for Skylake and Icelake */
+#define XXH_NEON   4  /* NEON for most ARMv7-A and all AArch64 */
+#define XXH_VSX    5  /* VSX and ZVector for POWER8/z13 */
+
+#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,
+ * 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
+#  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
+#else
+#  define XXH_SEC_ALIGN 8
+#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
+/* gcc's altivec.h can have the unwanted consequence to unconditionally
+ * #define bool, vector, and pixel keywords,
+ * with bad consequences for programs already using these keywords for other purposes.
+ * The paragraph defining these macros is skipped when __APPLE_ALTIVEC__ is defined.
+ * __APPLE_ALTIVEC__ is _generally_ defined automatically by the compiler,
+ * but it seems that, in some cases, it isn't.
+ * Force the build macro to be defined, so that keywords are not altered.
+ */
+#    if defined(__GNUC__) && !defined(__APPLE_ALTIVEC__)
+#      define __APPLE_ALTIVEC__
+#    endif
+#    include <altivec.h>
+#  endif
+
+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__) && XXH_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 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,
+};
+
+
+#ifdef XXH_OLD_NAMES
+#  define kSecret XXH3_kSecret
+#endif
+
+/*
+ * 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);
+}
+
+/*
+ * 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 *= 0x165667919E3779F9ULL;
+    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 *= 0x9FB21C651E98DF25ULL;
+    h64 ^= (h64 >> 35) + len ;
+    h64 *= 0x9FB21C651E98DF25ULL;
+    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 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 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 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 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.
+     */
+    __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 * XXH_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 * XXH_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 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
+
+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));
+}
+
+/* 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_X86DISPATCH)
+
+#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)
+{
+    XXH_ALIGN(64) __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_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_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);
+        /* 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);
+    }
+}
+
+/*
+ * 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));
+    {   XXH_ALIGN(64) __m512i* const xacc = (__m512i*) acc;
+        const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
+
+        /* xacc[0] ^= (xacc[0] >> 47) */
+        __m512i const acc_vec     = *xacc;
+        __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
+        __m512i const data_vec    = _mm512_xor_si512     (acc_vec, shifted);
+        /* xacc[0] ^= secret; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+
+        /* xacc[0] *= XXH_PRIME32_1; */
+        __m512i const data_key_hi = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        __m512i const 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 = _mm512_mask_set1_epi64(_mm512_set1_epi64((xxh_i64)seed64), 0xAA, -(xxh_i64)seed64);
+
+        XXH_ALIGN(64) const __m512i* const src  = (const __m512i*) XXH3_kSecret;
+        XXH_ALIGN(64)       __m512i* const dest = (      __m512i*) customSecret;
+        int i;
+        for (i=0; i < nbRounds; ++i) {
+            /* GCC has a bug, _mm512_stream_load_si512 accepts 'void*', not 'void const*',
+             * this will warn "discards ‘const’ qualifier". */
+            union {
+                XXH_ALIGN(64) const __m512i* cp;
+                XXH_ALIGN(64) void* p;
+            } remote_const_void;
+            remote_const_void.cp = src + i;
+            dest[i] = _mm512_add_epi64(_mm512_stream_load_si512(remote_const_void.p), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_AVX2) || defined(XXH_X86DISPATCH)
+
+#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);
+    {   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 < 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_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);
+            /* 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 void
+XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   XXH_ALIGN(32) __m256i* const xacc = (__m256i*) acc;
+        /* 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_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));
+        }
+    }
+}
+
+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)seed64, (xxh_i64)seed64, -(xxh_i64)seed64, (xxh_i64)seed64);
+
+        XXH_ALIGN(64) const __m256i* const src  = (const __m256i*) XXH3_kSecret;
+        XXH_ALIGN(64)       __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
+         * The asm hack causes Clang to assume that XXH3_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.
+         */
+        __asm__("" : "+r" (dest));
+#       endif
+
+        /* GCC -O2 need unroll loop manually */
+        dest[0] = _mm256_add_epi64(_mm256_stream_load_si256(src+0), seed);
+        dest[1] = _mm256_add_epi64(_mm256_stream_load_si256(src+1), seed);
+        dest[2] = _mm256_add_epi64(_mm256_stream_load_si256(src+2), seed);
+        dest[3] = _mm256_add_epi64(_mm256_stream_load_si256(src+3), seed);
+        dest[4] = _mm256_add_epi64(_mm256_stream_load_si256(src+4), seed);
+        dest[5] = _mm256_add_epi64(_mm256_stream_load_si256(src+5), seed);
+    }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)
+
+#ifndef XXH_TARGET_SSE2
+# define XXH_TARGET_SSE2  /* disable attribute target */
+#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);
+    {   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 < 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 void
+XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   XXH_ALIGN(16) __m128i* const xacc = (__m128i*) acc;
+        /* 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)seed64 };
+        __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
+#       else
+        __m128i const seed = _mm_set_epi64x(-(xxh_i64)seed64, (xxh_i64)seed64);
+#       endif
+        int i;
+
+        XXH_ALIGN(64)        const float* const src  = (float const*) XXH3_kSecret;
+        XXH_ALIGN(XXH_SEC_ALIGN) __m128i*       dest = (__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
+         */
+        __asm__("" : "+r" (dest));
+#       endif
+
+        for (i=0; i < nbRounds; ++i) {
+            dest[i] = _mm_add_epi64(_mm_castps_si128(_mm_load_ps(src+i*4)), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_NEON)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {
+        XXH_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 < XXH_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;
+            /* 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);
+
+        }
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   uint64x2_t* xacc       = (uint64x2_t*) acc;
+        uint8_t const* xsecret = (uint8_t const*) secret;
+        uint32x2_t prime       = vdup_n_u32 (XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(uint64x2_t); i++) {
+            /* 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] *= XXH_PRIME32_1 */
+            uint32x2_t data_key_lo, data_key_hi;
+            /* data_key_lo = (uint32x2_t) (xacc[i] & 0xFFFFFFFF);
+             * data_key_hi = (uint32x2_t) (xacc[i] >> 32);
+             * xacc[i] = UNDEFINED; */
+            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
+            {   /*
+                 * prod_hi = (data_key >> 32) * XXH_PRIME32_1;
+                 *
+                 * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will
+                 * incorrectly "optimize" this:
+                 *   tmp     = vmul_u32(vmovn_u64(a), vmovn_u64(b));
+                 *   shifted = vshll_n_u32(tmp, 32);
+                 * to this:
+                 *   tmp     = "vmulq_u64"(a, b); // no such thing!
+                 *   shifted = vshlq_n_u64(tmp, 32);
+                 *
+                 * However, unlike SSE, Clang lacks a 64-bit multiply routine
+                 * for NEON, and it scalarizes two 64-bit multiplies instead.
+                 *
+                 * vmull_u32 has the same timing as vmul_u32, and it avoids
+                 * this bug completely.
+                 * See https://bugs.llvm.org/show_bug.cgi?id=39967
+                 */
+                uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime);
+                /* xacc[i] = prod_hi << 32; */
+                xacc[i] = vshlq_n_u64(prod_hi, 32);
+                /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */
+                xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime);
+            }
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_VSX)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+          xxh_u64x2* const xacc     =       (xxh_u64x2*) acc;    /* presumed aligned */
+    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 < XXH_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;
+
+        /* swap high and low halves */
+#ifdef __s390x__
+        xacc[i] += vec_permi(data_vec, data_vec, 2);
+#else
+        xacc[i] += vec_xxpermdi(data_vec, data_vec, 2);
+#endif
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {         xxh_u64x2* const xacc    =       (xxh_u64x2*) acc;
+        const xxh_u64x2* const xsecret = (const xxh_u64x2*) secret;
+        /* 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 + 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);
+    }   }
+}
+
+#endif
+
+/* scalar variants - universal */
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */
+    const xxh_u8* const xinput  = (const xxh_u8*) input;  /* no alignment restriction */
+    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
+    size_t i;
+    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+    for (i=0; i < XXH_ACC_NB; i++) {
+        xxh_u64 const data_val = XXH_readLE64(xinput + 8*i);
+        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i*8);
+        xacc[i ^ 1] += data_val; /* swap adjacent lanes */
+        xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    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 < XXH_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 *= XXH_PRIME32_1;
+        xacc[i] = acc64;
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    /*
+     * We need a separate pointer for the hack below,
+     * which requires a non-const pointer.
+     * Any decent compiler will optimize this out otherwise.
+     */
+    const xxh_u8* kSecretPtr = XXH3_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 XXH3_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 == XXH3_kSecret);
+
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
+        int i;
+        for (i=0; i < nbRounds; i++) {
+            /*
+             * The asm hack causes Clang to assume that kSecretPtr aliases with
+             * 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_512)(void* XXH_RESTRICT, const void*, const void*);
+typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
+typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
+
+
+#if (XXH_VECTOR == XXH_AVX512)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx512
+#define XXH3_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_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_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_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_scrambleAcc    XXH3_scrambleAcc_vsx
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#else /* scalar */
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#endif
+
+
+
+#ifndef XXH_PREFETCH_DIST
+#  ifdef __clang__
+#    define XXH_PREFETCH_DIST 320
+#  else
+#    if (XXH_VECTOR == XXH_AVX512)
+#      define XXH_PREFETCH_DIST 512
+#    else
+#      define XXH_PREFETCH_DIST 384
+#    endif
+#  endif  /* __clang__ */
+#endif  /* XXH_PREFETCH_DIST */
+
+/*
+ * XXH3_accumulate()
+ * Loops over XXH3_accumulate_512().
+ * Assumption: nbStripes will not overflow the secret size
+ */
+XXH_FORCE_INLINE void
+XXH3_accumulate(     xxh_u64* XXH_RESTRICT acc,
+                const xxh_u8* XXH_RESTRICT input,
+                const xxh_u8* XXH_RESTRICT secret,
+                      size_t nbStripes,
+                      XXH3_f_accumulate_512 f_acc512)
+{
+    size_t n;
+    for (n = 0; n < nbStripes; n++ ) {
+        const xxh_u8* const in = input + n*XXH_STRIPE_LEN;
+        XXH_PREFETCH(in + XXH_PREFETCH_DIST);
+        f_acc512(acc,
+                 in,
+                 secret + n*XXH_SECRET_CONSUME_RATE);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
+                      const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate_512 f_acc512,
+                            XXH3_f_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++) {
+        XXH3_accumulate(acc, input + n*block_len, secret, nbStripesPerBlock, f_acc512);
+        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));
+        XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, f_acc512);
+
+        /* last stripe */
+        {   const xxh_u8* const p = input + len - XXH_STRIPE_LEN;
+#define XXH_SECRET_LASTACC_START 7  /* not aligned on 8, last secret is different from acc & scrambler */
+            f_acc512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+    }   }
+}
+
+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 { 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_512 f_acc512,
+                           XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc512, f_scramble);
+
+    /* 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, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1);
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                             XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ * Since the function is not inlined, the compiler may not be able to understand that,
+ * in some scenarios, its `secret` argument is actually a compile time constant.
+ * This variant enforces that the compiler can detect that,
+ * and uses this opportunity to streamline the generated code for better performance.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
+                          XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+/*
+ * 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_512 f_acc512,
+                                    XXH3_f_scrambleAcc f_scramble,
+                                    XXH3_f_initCustomSecret f_initSec)
+{
+    if (seed == 0)
+        return XXH3_hashLong_64b_internal(input, len,
+                                          XXH3_kSecret, sizeof(XXH3_kSecret),
+                                          f_acc512, f_scramble);
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed);
+        return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
+                                          f_acc512, f_scramble);
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed(const void* input, size_t len,
+                           XXH64_hash_t seed, const xxh_u8* secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
+                XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+
+typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t,
+                                          XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t);
+
+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)
+{
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * 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.
+     */
+    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);
+}
+
+
+/* ===   Public entry point   === */
+
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len)
+{
+    return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+}
+
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_internal(input, len, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+}
+
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+}
+
+
+/* ===   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.
+ *
+ * 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 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.
+ */
+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)
+{
+    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;
+}
+
+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 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->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 XXH_errorcode
+XXH3_64bits_reset(XXH3_state_t* statePtr)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_64bits_reset_internal(statePtr, 0, XXH3_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, 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;
+    if (seed==0) return XXH3_64bits_reset(statePtr);
+    if (seed != statePtr->seed) XXH3_initCustomSecret(statePtr->customSecret, seed);
+    XXH3_64bits_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/* Note : when XXH3_consumeStripes() is invoked,
+ * there must be a guarantee that at least one more byte must be consumed from input
+ * so that the function can blindly consume all stripes using the "normal" secret segment */
+XXH_FORCE_INLINE void
+XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
+                    size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
+                    const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
+                    const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
+                    XXH3_f_accumulate_512 f_acc512,
+                    XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ASSERT(nbStripes <= nbStripesPerBlock);  /* can handle max 1 scramble per invocation */
+    XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock);
+    if (nbStripesPerBlock - *nbStripesSoFarPtr <= nbStripes) {
+        /* need a scrambling operation */
+        size_t const nbStripesToEndofBlock = nbStripesPerBlock - *nbStripesSoFarPtr;
+        size_t const nbStripesAfterBlock = nbStripes - nbStripesToEndofBlock;
+        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripesToEndofBlock, f_acc512);
+        f_scramble(acc, secret + secretLimit);
+        XXH3_accumulate(acc, input + nbStripesToEndofBlock * XXH_STRIPE_LEN, secret, nbStripesAfterBlock, f_acc512);
+        *nbStripesSoFarPtr = nbStripesAfterBlock;
+    } else {
+        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, f_acc512);
+        *nbStripesSoFarPtr += nbStripes;
+    }
+}
+
+/*
+ * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
+ */
+XXH_FORCE_INLINE XXH_errorcode
+XXH3_update(XXH3_state_t* state,
+            const xxh_u8* input, size_t len,
+            XXH3_f_accumulate_512 f_acc512,
+            XXH3_f_scrambleAcc f_scramble)
+{
+    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;
+        const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+
+        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;
+        }
+        /* 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(state->acc,
+                               &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
+                                secret, state->secretLimit,
+                                f_acc512, f_scramble);
+            state->bufferedSize = 0;
+        }
+        XXH_ASSERT(input < bEnd);
+
+        /* Consume input by a multiple of internal buffer size */
+        if (input+XXH3_INTERNALBUFFER_SIZE < bEnd) {
+            const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE;
+            do {
+                XXH3_consumeStripes(state->acc,
+                                   &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                    input, XXH3_INTERNALBUFFER_STRIPES,
+                                    secret, state->secretLimit,
+                                    f_acc512, f_scramble);
+                input += XXH3_INTERNALBUFFER_SIZE;
+            } while (input<limit);
+            /* for last partial stripe */
+            memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+        }
+        XXH_ASSERT(input < bEnd);
+
+        /* Some remaining input (always) : buffer it */
+        XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
+        state->bufferedSize = (XXH32_hash_t)(bEnd-input);
+    }
+
+    return XXH_OK;
+}
+
+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_accumulate_512, XXH3_scrambleAcc);
+}
+
+
+XXH_FORCE_INLINE void
+XXH3_digest_long (XXH64_hash_t* acc,
+                  const XXH3_state_t* state,
+                  const unsigned char* secret)
+{
+    /*
+     * 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 >= XXH_STRIPE_LEN) {
+        size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
+        size_t nbStripesSoFar = state->nbStripesSoFar;
+        XXH3_consumeStripes(acc,
+                           &nbStripesSoFar, state->nbStripesPerBlock,
+                            state->buffer, nbStripes,
+                            secret, state->secretLimit,
+                            XXH3_accumulate_512, XXH3_scrambleAcc);
+        /* last stripe */
+        XXH3_accumulate_512(acc,
+                            state->buffer + state->bufferedSize - XXH_STRIPE_LEN,
+                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+    } else {  /* bufferedSize < XXH_STRIPE_LEN */
+        xxh_u8 lastStripe[XXH_STRIPE_LEN];
+        size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
+        XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
+        memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
+        memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
+        XXH3_accumulate_512(acc,
+                            lastStripe,
+                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+    }
+}
+
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (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_mergeAccs(acc,
+                              secret + XXH_SECRET_MERGEACCS_START,
+                              (xxh_u64)state->totalLen * XXH_PRIME64_1);
+    }
+    /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
+    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),
+                                  secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+
+
+#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
+
+XXH_PUBLIC_API void
+XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSeedSize)
+{
+    XXH_ASSERT(secretBuffer != NULL);
+    if (customSeedSize == 0) {
+        memcpy(secretBuffer, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+        return;
+    }
+    XXH_ASSERT(customSeed != NULL);
+
+    {   size_t const segmentSize = sizeof(XXH128_hash_t);
+        size_t const nbSegments = XXH_SECRET_DEFAULT_SIZE / segmentSize;
+        XXH128_canonical_t scrambler;
+        XXH64_hash_t seeds[12];
+        size_t segnb;
+        XXH_ASSERT(nbSegments == 12);
+        XXH_ASSERT(segmentSize * nbSegments == XXH_SECRET_DEFAULT_SIZE); /* exact multiple */
+        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
+
+        /*
+        * Copy customSeed to seeds[], truncating or repeating as necessary.
+        */
+        {   size_t toFill = XXH_MIN(customSeedSize, sizeof(seeds));
+            size_t filled = toFill;
+            memcpy(seeds, customSeed, toFill);
+            while (filled < sizeof(seeds)) {
+                toFill = XXH_MIN(filled, sizeof(seeds) - filled);
+                memcpy((char*)seeds + filled, seeds, toFill);
+                filled += toFill;
+        }   }
+
+        /* generate secret */
+        memcpy(secretBuffer, &scrambler, sizeof(scrambler));
+        for (segnb=1; segnb < nbSegments; segnb++) {
+            size_t const segmentStart = segnb * segmentSize;
+            XXH128_canonical_t segment;
+            XXH128_canonicalFromHash(&segment,
+                XXH128(&scrambler, sizeof(scrambler), XXH_readLE64(seeds + segnb) + segnb) );
+            memcpy((char*)secretBuffer + segmentStart, &segment, sizeof(segment));
+    }   }
+}
+
+
+/* ==========================================
+ * 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;
+        XXH128_hash_t h128;
+        h128.low64  = XXH64_avalanche(keyed_lo);
+        h128.high64 = XXH64_avalanche(keyed_hi);
+        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, XXH_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, 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);
+
+        {   /* 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;
+    }   }
+}
+
+/*
+ * 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 = XXH64_avalanche(seed ^ bitflipl);
+            h128.high64 = XXH64_avalanche( 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 * XXH_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    * 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;
+        }
+    }
+}
+
+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 * XXH_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    * 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;
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
+                            const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate_512 f_acc512,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc512, f_scramble);
+
+    /* 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 * XXH_PRIME64_1);
+        h128.high64 = XXH3_mergeAccs(acc,
+                                     secret + secretSize
+                                            - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                     ~((xxh_u64)len * XXH_PRIME64_2));
+        return h128;
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed64,
+                           const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
+                                       XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                              XXH64_hash_t seed64,
+                              const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
+                                       XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len,
+                                XXH64_hash_t seed64,
+                                XXH3_f_accumulate_512 f_acc512,
+                                XXH3_f_scrambleAcc f_scramble,
+                                XXH3_f_initCustomSecret f_initSec)
+{
+    if (seed64 == 0)
+        return XXH3_hashLong_128b_internal(input, len,
+                                           XXH3_kSecret, sizeof(XXH3_kSecret),
+                                           f_acc512, f_scramble);
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed64);
+        return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret),
+                                           f_acc512, f_scramble);
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+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)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
+                XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
+                                            XXH64_hash_t, const void* XXH_RESTRICT, size_t);
+
+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)
+{
+    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);
+}
+
+
+/* ===   Public XXH128 API   === */
+
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_default);
+}
+
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 (const xxh_u8*)secret, secretSize,
+                                 XXH3_hashLong_128b_withSecret);
+}
+
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_internal(input, len, seed,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_withSeed);
+}
+
+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 void* 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, XXH3_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, 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;
+    if (seed==0) return XXH3_128bits_reset(statePtr);
+    if (seed != statePtr->seed) XXH3_initCustomSecret(statePtr->customSecret, seed);
+    XXH3_128bits_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+    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_accumulate_512, XXH3_scrambleAcc);
+}
+
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (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);
+        {   XXH128_hash_t h128;
+            h128.low64  = XXH3_mergeAccs(acc,
+                                         secret + XXH_SECRET_MERGEACCS_START,
+                                         (xxh_u64)state->totalLen * XXH_PRIME64_1);
+            h128.high64 = XXH3_mergeAccs(acc,
+                                         secret + state->secretLimit + XXH_STRIPE_LEN
+                                                - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                         ~((xxh_u64)state->totalLen * XXH_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),
+                                   secret, state->secretLimit + XXH_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  /* XXH_NO_LONG_LONG */
+
+
+#endif  /* XXH_IMPLEMENTATION */
+
+
+#if defined (__cplusplus)
+}
+#endif