Merge tag 'zstd-linus-v6.15-rc1' of https://github.com/terrelln/linux

Pull zstd updates from Nick Terrell:
"Update zstd to the latest upstream release v1.5.7.

The two major motivations for updating Zstandard are to keep the code
up to date, and to expose API's needed by Intel for the QAT
compression accelerator.

Imported cleanly from the upstream tag v1.5.7-kernel, which is signed
by upstream's signing key EF8FE99528B52FFD"

Link: https://github.com/facebook/zstd/releases/tag/v1.5.7
Link: https://github.com/facebook/zstd/releases/tag/v1.5.7-kernel
Link: https://keyserver.ubuntu.com/pks/lookup?search=EF8FE99528B52FFD&fingerprint=on&op=index

* tag 'zstd-linus-v6.15-rc1' of https://github.com/terrelln/linux:
zstd: Import upstream v1.5.7

include/linux/zstd.h

···
 /* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 zstd_parameters zstd_get_params(int level,
 	unsigned long long estimated_src_size);
 
-
 /**
  * zstd_get_cparams() - returns zstd_compression_parameters for selected level
  * @level:              The compression level
···
 zstd_compression_parameters zstd_get_cparams(int level,
 	unsigned long long estimated_src_size, size_t dict_size);
 
-/* ======   Single-pass Compression   ====== */
-
 typedef ZSTD_CCtx zstd_cctx;
+typedef ZSTD_cParameter zstd_cparameter;
+
+/**
+ * zstd_cctx_set_param() - sets a compression parameter
+ * @cctx:         The context. Must have been initialized with zstd_init_cctx().
+ * @param:        The parameter to set.
+ * @value:        The value to set the parameter to.
+ *
+ * Return:        Zero or an error, which can be checked using zstd_is_error().
+ */
+size_t zstd_cctx_set_param(zstd_cctx *cctx, zstd_cparameter param, int value);
+
+/* ======   Single-pass Compression   ====== */
 
 /**
  * zstd_cctx_workspace_bound() - max memory needed to initialize a zstd_cctx
···
  *              zstd_init_cctx().
  */
 size_t zstd_cctx_workspace_bound(const zstd_compression_parameters *parameters);
+
+/**
+ * zstd_cctx_workspace_bound_with_ext_seq_prod() - max memory needed to
+ * initialize a zstd_cctx when using the block-level external sequence
+ * producer API.
+ * @parameters: The compression parameters to be used.
+ *
+ * If multiple compression parameters might be used, the caller must call
+ * this function for each set of parameters and use the maximum size.
+ *
+ * Return:      A lower bound on the size of the workspace that is passed to
+ *              zstd_init_cctx().
+ */
+size_t zstd_cctx_workspace_bound_with_ext_seq_prod(const zstd_compression_parameters *parameters);
 
 /**
  * zstd_init_cctx() - initialize a zstd compression context
···
 size_t zstd_cstream_workspace_bound(const zstd_compression_parameters *cparams);
 
 /**
+ * zstd_cstream_workspace_bound_with_ext_seq_prod() - memory needed to initialize
+ * a zstd_cstream when using the block-level external sequence producer API.
+ * @cparams: The compression parameters to be used for compression.
+ *
+ * Return:   A lower bound on the size of the workspace that is passed to
+ *           zstd_init_cstream().
+ */
+size_t zstd_cstream_workspace_bound_with_ext_seq_prod(const zstd_compression_parameters *cparams);
+
+/**
  * zstd_init_cstream() - initialize a zstd streaming compression context
  * @parameters        The zstd parameters to use for compression.
  * @pledged_src_size: If params.fParams.contentSizeFlag == 1 then the caller
···
 size_t zstd_find_frame_compressed_size(const void *src, size_t src_size);
 
 /**
+ * zstd_register_sequence_producer() - exposes the zstd library function
+ * ZSTD_registerSequenceProducer(). This is used for the block-level external
+ * sequence producer API. See upstream zstd.h for detailed documentation.
+ */
+typedef ZSTD_sequenceProducer_F zstd_sequence_producer_f;
+void zstd_register_sequence_producer(
+  zstd_cctx *cctx,
+  void* sequence_producer_state,
+  zstd_sequence_producer_f sequence_producer
+);
+
+/**
  * struct zstd_frame_params - zstd frame parameters stored in the frame header
  * @frameContentSize: The frame content size, or ZSTD_CONTENTSIZE_UNKNOWN if not
  *                    present.
···
  *
  * See zstd_lib.h.
  */
-typedef ZSTD_frameHeader zstd_frame_header;
+typedef ZSTD_FrameHeader zstd_frame_header;
 
 /**
  * zstd_get_frame_header() - extracts parameters from a zstd or skippable frame
···
  */
 size_t zstd_get_frame_header(zstd_frame_header *params, const void *src,
 	size_t src_size);
+
+/**
+ * struct zstd_sequence - a sequence of literals or a match
+ *
+ * @offset: The offset of the match
+ * @litLength: The literal length of the sequence
+ * @matchLength: The match length of the sequence
+ * @rep: Represents which repeat offset is used
+ */
+typedef ZSTD_Sequence zstd_sequence;
+
+/**
+ * zstd_compress_sequences_and_literals() - compress an array of zstd_sequence and literals
+ *
+ * @cctx: The zstd compression context.
+ * @dst: The buffer to compress the data into.
+ * @dst_capacity: The size of the destination buffer.
+ * @in_seqs: The array of zstd_sequence to compress.
+ * @in_seqs_size: The number of sequences in in_seqs.
+ * @literals: The literals associated to the sequences to be compressed.
+ * @lit_size: The size of the literals in the literals buffer.
+ * @lit_capacity: The size of the literals buffer.
+ * @decompressed_size: The size of the input data
+ *
+ * Return: The compressed size or an error, which can be checked using
+ * 	   zstd_is_error().
+ */
+size_t zstd_compress_sequences_and_literals(zstd_cctx *cctx, void* dst, size_t dst_capacity,
+					    const zstd_sequence *in_seqs, size_t in_seqs_size,
+					    const void* literals, size_t lit_size, size_t lit_capacity,
+					    size_t decompressed_size);
 
 #endif  /* LINUX_ZSTD_H */

include/linux/zstd_errors.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #define ZSTD_ERRORS_H_398273423
 
 
-/*===== dependency =====*/
-#include <linux/types.h>   /* size_t */
-
-
 /* =====   ZSTDERRORLIB_API : control library symbols visibility   ===== */
-#define ZSTDERRORLIB_VISIBILITY 
-#define ZSTDERRORLIB_API ZSTDERRORLIB_VISIBILITY
+#define ZSTDERRORLIB_VISIBLE 
+
+#ifndef ZSTDERRORLIB_HIDDEN
+#  if (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZSTDERRORLIB_HIDDEN __attribute__ ((visibility ("hidden")))
+#  else
+#    define ZSTDERRORLIB_HIDDEN
+#  endif
+#endif
+
+#define ZSTDERRORLIB_API ZSTDERRORLIB_VISIBLE
 
 /*-*********************************************
  *  Error codes list
···
   ZSTD_error_frameParameter_windowTooLarge = 16,
   ZSTD_error_corruption_detected = 20,
   ZSTD_error_checksum_wrong      = 22,
+  ZSTD_error_literals_headerWrong = 24,
   ZSTD_error_dictionary_corrupted      = 30,
   ZSTD_error_dictionary_wrong          = 32,
   ZSTD_error_dictionaryCreation_failed = 34,
   ZSTD_error_parameter_unsupported   = 40,
+  ZSTD_error_parameter_combination_unsupported = 41,
   ZSTD_error_parameter_outOfBound    = 42,
   ZSTD_error_tableLog_tooLarge       = 44,
   ZSTD_error_maxSymbolValue_tooLarge = 46,
   ZSTD_error_maxSymbolValue_tooSmall = 48,
+  ZSTD_error_cannotProduce_uncompressedBlock = 49,
+  ZSTD_error_stabilityCondition_notRespected = 50,
   ZSTD_error_stage_wrong       = 60,
   ZSTD_error_init_missing      = 62,
   ZSTD_error_memory_allocation = 64,
···
   ZSTD_error_dstSize_tooSmall = 70,
   ZSTD_error_srcSize_wrong    = 72,
   ZSTD_error_dstBuffer_null   = 74,
+  ZSTD_error_noForwardProgress_destFull = 80,
+  ZSTD_error_noForwardProgress_inputEmpty = 82,
   /* following error codes are __NOT STABLE__, they can be removed or changed in future versions */
   ZSTD_error_frameIndex_tooLarge = 100,
   ZSTD_error_seekableIO          = 102,
   ZSTD_error_dstBuffer_wrong     = 104,
   ZSTD_error_srcBuffer_wrong     = 105,
+  ZSTD_error_sequenceProducer_failed = 106,
+  ZSTD_error_externalSequences_invalid = 107,
   ZSTD_error_maxCode = 120  /* never EVER use this value directly, it can change in future versions! Use ZSTD_isError() instead */
 } ZSTD_ErrorCode;
 
-/*! ZSTD_getErrorCode() :
-    convert a `size_t` function result into a `ZSTD_ErrorCode` enum type,
-    which can be used to compare with enum list published above */
-ZSTDERRORLIB_API ZSTD_ErrorCode ZSTD_getErrorCode(size_t functionResult);
 ZSTDERRORLIB_API const char* ZSTD_getErrorString(ZSTD_ErrorCode code);   /*< Same as ZSTD_getErrorName, but using a `ZSTD_ErrorCode` enum argument */
 
 

include/linux/zstd_lib.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #ifndef ZSTD_H_235446
 #define ZSTD_H_235446
 
-/* ======   Dependency   ======*/
-#include <linux/limits.h>   /* INT_MAX */
+
+/* ======   Dependencies   ======*/
 #include <linux/types.h>   /* size_t */
+
+#include <linux/zstd_errors.h> /* list of errors */
+#if !defined(ZSTD_H_ZSTD_STATIC_LINKING_ONLY)
+#include <linux/limits.h>   /* INT_MAX */
+#endif /* ZSTD_STATIC_LINKING_ONLY */
 
 
 /* =====   ZSTDLIB_API : control library symbols visibility   ===== */
-#ifndef ZSTDLIB_VISIBLE
+#define ZSTDLIB_VISIBLE 
+
+#ifndef ZSTDLIB_HIDDEN
 #  if (__GNUC__ >= 4) && !defined(__MINGW32__)
-#    define ZSTDLIB_VISIBLE __attribute__ ((visibility ("default")))
 #    define ZSTDLIB_HIDDEN __attribute__ ((visibility ("hidden")))
 #  else
-#    define ZSTDLIB_VISIBLE
 #    define ZSTDLIB_HIDDEN
 #  endif
 #endif
+
 #define ZSTDLIB_API ZSTDLIB_VISIBLE
+
+/* Deprecation warnings :
+ * Should these 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.
+ * Otherwise, it's also possible to define ZSTD_DISABLE_DEPRECATE_WARNINGS.
+ */
+#ifdef ZSTD_DISABLE_DEPRECATE_WARNINGS
+#  define ZSTD_DEPRECATED(message) /* disable deprecation warnings */
+#else
+#  if (defined(GNUC) && (GNUC > 4 || (GNUC == 4 && GNUC_MINOR >= 5))) || defined(__clang__) || defined(__IAR_SYSTEMS_ICC__)
+#    define ZSTD_DEPRECATED(message) __attribute__((deprecated(message)))
+#  elif (__GNUC__ >= 3)
+#    define ZSTD_DEPRECATED(message) __attribute__((deprecated))
+#  else
+#    pragma message("WARNING: You need to implement ZSTD_DEPRECATED for this compiler")
+#    define ZSTD_DEPRECATED(message)
+#  endif
+#endif /* ZSTD_DISABLE_DEPRECATE_WARNINGS */
 
 
 /* *****************************************************************************
···
 /*------   Version   ------*/
 #define ZSTD_VERSION_MAJOR    1
 #define ZSTD_VERSION_MINOR    5
-#define ZSTD_VERSION_RELEASE  2
+#define ZSTD_VERSION_RELEASE  7
 #define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
 
 /*! ZSTD_versionNumber() :
···
 
 
 /* *************************************
-*  Simple API
+*  Simple Core API
 ***************************************/
 /*! ZSTD_compress() :
  *  Compresses `src` content as a single zstd compressed frame into already allocated `dst`.
- *  Hint : compression runs faster if `dstCapacity` >=  `ZSTD_compressBound(srcSize)`.
+ *  NOTE: Providing `dstCapacity >= ZSTD_compressBound(srcSize)` guarantees that zstd will have
+ *        enough space to successfully compress the data.
  *  @return : compressed size written into `dst` (<= `dstCapacity),
  *            or an error code if it fails (which can be tested using ZSTD_isError()). */
 ZSTDLIB_API size_t ZSTD_compress( void* dst, size_t dstCapacity,
···
                                   int compressionLevel);
 
 /*! ZSTD_decompress() :
- *  `compressedSize` : must be the _exact_ size of some number of compressed and/or skippable frames.
- *  `dstCapacity` is an upper bound of originalSize to regenerate.
- *  If user cannot imply a maximum upper bound, it's better to use streaming mode to decompress data.
- *  @return : the number of bytes decompressed into `dst` (<= `dstCapacity`),
- *            or an errorCode if it fails (which can be tested using ZSTD_isError()). */
+ * `compressedSize` : must be the _exact_ size of some number of compressed and/or skippable frames.
+ *  Multiple compressed frames can be decompressed at once with this method.
+ *  The result will be the concatenation of all decompressed frames, back to back.
+ * `dstCapacity` is an upper bound of originalSize to regenerate.
+ *  First frame's decompressed size can be extracted using ZSTD_getFrameContentSize().
+ *  If maximum upper bound isn't known, prefer using streaming mode to decompress data.
+ * @return : the number of bytes decompressed into `dst` (<= `dstCapacity`),
+ *           or an errorCode if it fails (which can be tested using ZSTD_isError()). */
 ZSTDLIB_API size_t ZSTD_decompress( void* dst, size_t dstCapacity,
                               const void* src, size_t compressedSize);
 
+
+/*======  Decompression helper functions  ======*/
+
 /*! ZSTD_getFrameContentSize() : requires v1.3.0+
- *  `src` should point to the start of a ZSTD encoded frame.
- *  `srcSize` must be at least as large as the frame header.
- *            hint : any size >= `ZSTD_frameHeaderSize_max` is large enough.
- *  @return : - decompressed size of `src` frame content, if known
- *            - ZSTD_CONTENTSIZE_UNKNOWN if the size cannot be determined
- *            - ZSTD_CONTENTSIZE_ERROR if an error occurred (e.g. invalid magic number, srcSize too small)
- *   note 1 : a 0 return value means the frame is valid but "empty".
- *   note 2 : decompressed size is an optional field, it may not be present, typically in streaming mode.
- *            When `return==ZSTD_CONTENTSIZE_UNKNOWN`, data to decompress could be any size.
- *            In which case, it's necessary to use streaming mode to decompress data.
- *            Optionally, application can rely on some implicit limit,
- *            as ZSTD_decompress() only needs an upper bound of decompressed size.
- *            (For example, data could be necessarily cut into blocks <= 16 KB).
- *   note 3 : decompressed size is always present when compression is completed using single-pass functions,
- *            such as ZSTD_compress(), ZSTD_compressCCtx() ZSTD_compress_usingDict() or ZSTD_compress_usingCDict().
- *   note 4 : decompressed size can be very large (64-bits value),
- *            potentially larger than what local system can handle as a single memory segment.
- *            In which case, it's necessary to use streaming mode to decompress data.
- *   note 5 : If source is untrusted, decompressed size could be wrong or intentionally modified.
- *            Always ensure return value fits within application's authorized limits.
- *            Each application can set its own limits.
- *   note 6 : This function replaces ZSTD_getDecompressedSize() */
+ * `src` should point to the start of a ZSTD encoded frame.
+ * `srcSize` must be at least as large as the frame header.
+ *           hint : any size >= `ZSTD_frameHeaderSize_max` is large enough.
+ * @return : - decompressed size of `src` frame content, if known
+ *           - ZSTD_CONTENTSIZE_UNKNOWN if the size cannot be determined
+ *           - ZSTD_CONTENTSIZE_ERROR if an error occurred (e.g. invalid magic number, srcSize too small)
+ *  note 1 : a 0 return value means the frame is valid but "empty".
+ *           When invoking this method on a skippable frame, it will return 0.
+ *  note 2 : decompressed size is an optional field, it may not be present (typically in streaming mode).
+ *           When `return==ZSTD_CONTENTSIZE_UNKNOWN`, data to decompress could be any size.
+ *           In which case, it's necessary to use streaming mode to decompress data.
+ *           Optionally, application can rely on some implicit limit,
+ *           as ZSTD_decompress() only needs an upper bound of decompressed size.
+ *           (For example, data could be necessarily cut into blocks <= 16 KB).
+ *  note 3 : decompressed size is always present when compression is completed using single-pass functions,
+ *           such as ZSTD_compress(), ZSTD_compressCCtx() ZSTD_compress_usingDict() or ZSTD_compress_usingCDict().
+ *  note 4 : decompressed size can be very large (64-bits value),
+ *           potentially larger than what local system can handle as a single memory segment.
+ *           In which case, it's necessary to use streaming mode to decompress data.
+ *  note 5 : If source is untrusted, decompressed size could be wrong or intentionally modified.
+ *           Always ensure return value fits within application's authorized limits.
+ *           Each application can set its own limits.
+ *  note 6 : This function replaces ZSTD_getDecompressedSize() */
 #define ZSTD_CONTENTSIZE_UNKNOWN (0ULL - 1)
 #define ZSTD_CONTENTSIZE_ERROR   (0ULL - 2)
 ZSTDLIB_API unsigned long long ZSTD_getFrameContentSize(const void *src, size_t srcSize);
 
-/*! ZSTD_getDecompressedSize() :
- *  NOTE: This function is now obsolete, in favor of ZSTD_getFrameContentSize().
+/*! ZSTD_getDecompressedSize() (obsolete):
+ *  This function is now obsolete, in favor of ZSTD_getFrameContentSize().
  *  Both functions work the same way, but ZSTD_getDecompressedSize() blends
  *  "empty", "unknown" and "error" results to the same return value (0),
  *  while ZSTD_getFrameContentSize() gives them separate return values.
  * @return : decompressed size of `src` frame content _if known and not empty_, 0 otherwise. */
+ZSTD_DEPRECATED("Replaced by ZSTD_getFrameContentSize")
 ZSTDLIB_API unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize);
 
 /*! ZSTD_findFrameCompressedSize() : Requires v1.4.0+
···
  * `srcSize` must be >= first frame size
  * @return : the compressed size of the first frame starting at `src`,
  *           suitable to pass as `srcSize` to `ZSTD_decompress` or similar,
- *        or an error code if input is invalid */
+ *           or an error code if input is invalid
+ *  Note 1: this method is called _find*() because it's not enough to read the header,
+ *          it may have to scan through the frame's content, to reach its end.
+ *  Note 2: this method also works with Skippable Frames. In which case,
+ *          it returns the size of the complete skippable frame,
+ *          which is always equal to its content size + 8 bytes for headers. */
 ZSTDLIB_API size_t ZSTD_findFrameCompressedSize(const void* src, size_t srcSize);
 
 
-/*======  Helper functions  ======*/
-#define ZSTD_COMPRESSBOUND(srcSize)   ((srcSize) + ((srcSize)>>8) + (((srcSize) < (128<<10)) ? (((128<<10) - (srcSize)) >> 11) /* margin, from 64 to 0 */ : 0))  /* this formula ensures that bound(A) + bound(B) <= bound(A+B) as long as A and B >= 128 KB */
-ZSTDLIB_API size_t      ZSTD_compressBound(size_t srcSize); /*!< maximum compressed size in worst case single-pass scenario */
-ZSTDLIB_API unsigned    ZSTD_isError(size_t code);          /*!< tells if a `size_t` function result is an error code */
-ZSTDLIB_API const char* ZSTD_getErrorName(size_t code);     /*!< provides readable string from an error code */
-ZSTDLIB_API int         ZSTD_minCLevel(void);               /*!< minimum negative compression level allowed, requires v1.4.0+ */
-ZSTDLIB_API int         ZSTD_maxCLevel(void);               /*!< maximum compression level available */
-ZSTDLIB_API int         ZSTD_defaultCLevel(void);           /*!< default compression level, specified by ZSTD_CLEVEL_DEFAULT, requires v1.5.0+ */
+/*======  Compression helper functions  ======*/
+
+/*! ZSTD_compressBound() :
+ * maximum compressed size in worst case single-pass scenario.
+ * When invoking `ZSTD_compress()`, or any other one-pass compression function,
+ * it's recommended to provide @dstCapacity >= ZSTD_compressBound(srcSize)
+ * as it eliminates one potential failure scenario,
+ * aka not enough room in dst buffer to write the compressed frame.
+ * Note : ZSTD_compressBound() itself can fail, if @srcSize >= ZSTD_MAX_INPUT_SIZE .
+ *        In which case, ZSTD_compressBound() will return an error code
+ *        which can be tested using ZSTD_isError().
+ *
+ * ZSTD_COMPRESSBOUND() :
+ * same as ZSTD_compressBound(), but as a macro.
+ * It can be used to produce constants, which can be useful for static allocation,
+ * for example to size a static array on stack.
+ * Will produce constant value 0 if srcSize is too large.
+ */
+#define ZSTD_MAX_INPUT_SIZE ((sizeof(size_t)==8) ? 0xFF00FF00FF00FF00ULL : 0xFF00FF00U)
+#define ZSTD_COMPRESSBOUND(srcSize)   (((size_t)(srcSize) >= ZSTD_MAX_INPUT_SIZE) ? 0 : (srcSize) + ((srcSize)>>8) + (((srcSize) < (128<<10)) ? (((128<<10) - (srcSize)) >> 11) /* margin, from 64 to 0 */ : 0))  /* this formula ensures that bound(A) + bound(B) <= bound(A+B) as long as A and B >= 128 KB */
+ZSTDLIB_API size_t ZSTD_compressBound(size_t srcSize); /*!< maximum compressed size in worst case single-pass scenario */
+
+
+/*======  Error helper functions  ======*/
+/* ZSTD_isError() :
+ * Most ZSTD_* functions returning a size_t value can be tested for error,
+ * using ZSTD_isError().
+ * @return 1 if error, 0 otherwise
+ */
+ZSTDLIB_API unsigned     ZSTD_isError(size_t result);      /*!< tells if a `size_t` function result is an error code */
+ZSTDLIB_API ZSTD_ErrorCode ZSTD_getErrorCode(size_t functionResult); /* convert a result into an error code, which can be compared to error enum list */
+ZSTDLIB_API const char*  ZSTD_getErrorName(size_t result); /*!< provides readable string from a function result */
+ZSTDLIB_API int          ZSTD_minCLevel(void);             /*!< minimum negative compression level allowed, requires v1.4.0+ */
+ZSTDLIB_API int          ZSTD_maxCLevel(void);             /*!< maximum compression level available */
+ZSTDLIB_API int          ZSTD_defaultCLevel(void);         /*!< default compression level, specified by ZSTD_CLEVEL_DEFAULT, requires v1.5.0+ */
 
 
 /* *************************************
···
 ***************************************/
 /*= Compression context
  *  When compressing many times,
- *  it is recommended to allocate a context just once,
- *  and re-use it for each successive compression operation.
- *  This will make workload friendlier for system's memory.
+ *  it is recommended to allocate a compression context just once,
+ *  and reuse it for each successive compression operation.
+ *  This will make the workload easier for system's memory.
  *  Note : re-using context is just a speed / resource optimization.
  *         It doesn't change the compression ratio, which remains identical.
- *  Note 2 : In multi-threaded environments,
- *         use one different context per thread for parallel execution.
+ *  Note 2: For parallel execution in multi-threaded environments,
+ *         use one different context per thread .
  */
 typedef struct ZSTD_CCtx_s ZSTD_CCtx;
 ZSTDLIB_API ZSTD_CCtx* ZSTD_createCCtx(void);
-ZSTDLIB_API size_t     ZSTD_freeCCtx(ZSTD_CCtx* cctx);  /* accept NULL pointer */
+ZSTDLIB_API size_t     ZSTD_freeCCtx(ZSTD_CCtx* cctx);  /* compatible with NULL pointer */
 
 /*! ZSTD_compressCCtx() :
  *  Same as ZSTD_compress(), using an explicit ZSTD_CCtx.
- *  Important : in order to behave similarly to `ZSTD_compress()`,
- *  this function compresses at requested compression level,
- *  __ignoring any other parameter__ .
+ *  Important : in order to mirror `ZSTD_compress()` behavior,
+ *  this function compresses at the requested compression level,
+ *  __ignoring any other advanced parameter__ .
  *  If any advanced parameter was set using the advanced API,
- *  they will all be reset. Only `compressionLevel` remains.
+ *  they will all be reset. Only @compressionLevel remains.
  */
 ZSTDLIB_API size_t ZSTD_compressCCtx(ZSTD_CCtx* cctx,
                                      void* dst, size_t dstCapacity,
···
 /*= Decompression context
  *  When decompressing many times,
  *  it is recommended to allocate a context only once,
- *  and re-use it for each successive compression operation.
+ *  and reuse it for each successive compression operation.
  *  This will make workload friendlier for system's memory.
  *  Use one context per thread for parallel execution. */
 typedef struct ZSTD_DCtx_s ZSTD_DCtx;
···
 /*! ZSTD_decompressDCtx() :
  *  Same as ZSTD_decompress(),
  *  requires an allocated ZSTD_DCtx.
- *  Compatible with sticky parameters.
+ *  Compatible with sticky parameters (see below).
  */
 ZSTDLIB_API size_t ZSTD_decompressDCtx(ZSTD_DCtx* dctx,
                                        void* dst, size_t dstCapacity,
···
  *   using ZSTD_CCtx_set*() functions.
  *   Pushed parameters are sticky : they are valid for next compressed frame, and any subsequent frame.
  *   "sticky" parameters are applicable to `ZSTD_compress2()` and `ZSTD_compressStream*()` !
- *   __They do not apply to "simple" one-shot variants such as ZSTD_compressCCtx()__ .
+ *   __They do not apply to one-shot variants such as ZSTD_compressCCtx()__ .
  *
  *   It's possible to reset all parameters to "default" using ZSTD_CCtx_reset().
  *
  *   This API supersedes all other "advanced" API entry points in the experimental section.
- *   In the future, we expect to remove from experimental API entry points which are redundant with this API.
+ *   In the future, we expect to remove API entry points from experimental which are redundant with this API.
  */
 
 
···
                               * The higher the value of selected strategy, the more complex it is,
                               * resulting in stronger and slower compression.
                               * Special: value 0 means "use default strategy". */
+
+    ZSTD_c_targetCBlockSize=130, /* v1.5.6+
+                                  * Attempts to fit compressed block size into approximately targetCBlockSize.
+                                  * Bound by ZSTD_TARGETCBLOCKSIZE_MIN and ZSTD_TARGETCBLOCKSIZE_MAX.
+                                  * Note that it's not a guarantee, just a convergence target (default:0).
+                                  * No target when targetCBlockSize == 0.
+                                  * This is helpful in low bandwidth streaming environments to improve end-to-end latency,
+                                  * when a client can make use of partial documents (a prominent example being Chrome).
+                                  * Note: this parameter is stable since v1.5.6.
+                                  * It was present as an experimental parameter in earlier versions,
+                                  * but it's not recommended using it with earlier library versions
+                                  * due to massive performance regressions.
+                                  */
     /* LDM mode parameters */
     ZSTD_c_enableLongDistanceMatching=160, /* Enable long distance matching.
                                      * This parameter is designed to improve compression ratio
···
      * ZSTD_c_forceMaxWindow
      * ZSTD_c_forceAttachDict
      * ZSTD_c_literalCompressionMode
-     * ZSTD_c_targetCBlockSize
      * ZSTD_c_srcSizeHint
      * ZSTD_c_enableDedicatedDictSearch
      * ZSTD_c_stableInBuffer
      * ZSTD_c_stableOutBuffer
      * ZSTD_c_blockDelimiters
      * ZSTD_c_validateSequences
-     * ZSTD_c_useBlockSplitter
+     * ZSTD_c_blockSplitterLevel
+     * ZSTD_c_splitAfterSequences
      * ZSTD_c_useRowMatchFinder
+     * ZSTD_c_prefetchCDictTables
+     * ZSTD_c_enableSeqProducerFallback
+     * ZSTD_c_maxBlockSize
      * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them.
      * note : never ever use experimentalParam? names directly;
      *        also, the enums values themselves are unstable and can still change.
···
      ZSTD_c_experimentalParam3=1000,
      ZSTD_c_experimentalParam4=1001,
      ZSTD_c_experimentalParam5=1002,
-     ZSTD_c_experimentalParam6=1003,
+     /* was ZSTD_c_experimentalParam6=1003; is now ZSTD_c_targetCBlockSize */
      ZSTD_c_experimentalParam7=1004,
      ZSTD_c_experimentalParam8=1005,
      ZSTD_c_experimentalParam9=1006,
···
      ZSTD_c_experimentalParam12=1009,
      ZSTD_c_experimentalParam13=1010,
      ZSTD_c_experimentalParam14=1011,
-     ZSTD_c_experimentalParam15=1012
+     ZSTD_c_experimentalParam15=1012,
+     ZSTD_c_experimentalParam16=1013,
+     ZSTD_c_experimentalParam17=1014,
+     ZSTD_c_experimentalParam18=1015,
+     ZSTD_c_experimentalParam19=1016,
+     ZSTD_c_experimentalParam20=1017
 } ZSTD_cParameter;
 
 typedef struct {
···
  *                  They will be used to compress next frame.
  *                  Resetting session never fails.
  *  - The parameters : changes all parameters back to "default".
- *                  This removes any reference to any dictionary too.
+ *                  This also removes any reference to any dictionary or external sequence producer.
  *                  Parameters can only be changed between 2 sessions (i.e. no compression is currently ongoing)
  *                  otherwise the reset fails, and function returns an error value (which can be tested using ZSTD_isError())
  *  - Both : similar to resetting the session, followed by resetting parameters.
···
 
 /*! ZSTD_compress2() :
  *  Behave the same as ZSTD_compressCCtx(), but compression parameters are set using the advanced API.
+ *  (note that this entry point doesn't even expose a compression level parameter).
  *  ZSTD_compress2() always starts a new frame.
  *  Should cctx hold data from a previously unfinished frame, everything about it is forgotten.
  *  - Compression parameters are pushed into CCtx before starting compression, using ZSTD_CCtx_set*()
  *  - The function is always blocking, returns when compression is completed.
- *  Hint : compression runs faster if `dstCapacity` >=  `ZSTD_compressBound(srcSize)`.
+ *  NOTE: Providing `dstCapacity >= ZSTD_compressBound(srcSize)` guarantees that zstd will have
+ *        enough space to successfully compress the data, though it is possible it fails for other reasons.
  * @return : compressed size written into `dst` (<= `dstCapacity),
  *           or an error code if it fails (which can be tested using ZSTD_isError()).
  */
···
      * ZSTD_d_stableOutBuffer
      * ZSTD_d_forceIgnoreChecksum
      * ZSTD_d_refMultipleDDicts
+     * ZSTD_d_disableHuffmanAssembly
+     * ZSTD_d_maxBlockSize
      * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them.
      * note : never ever use experimentalParam? names directly
      */
      ZSTD_d_experimentalParam1=1000,
      ZSTD_d_experimentalParam2=1001,
      ZSTD_d_experimentalParam3=1002,
-     ZSTD_d_experimentalParam4=1003
+     ZSTD_d_experimentalParam4=1003,
+     ZSTD_d_experimentalParam5=1004,
+     ZSTD_d_experimentalParam6=1005
 
 } ZSTD_dParameter;
 
···
 *  A ZSTD_CStream object is required to track streaming operation.
 *  Use ZSTD_createCStream() and ZSTD_freeCStream() to create/release resources.
 *  ZSTD_CStream objects can be reused multiple times on consecutive compression operations.
-*  It is recommended to re-use ZSTD_CStream since it will play nicer with system's memory, by re-using already allocated memory.
+*  It is recommended to reuse ZSTD_CStream since it will play nicer with system's memory, by re-using already allocated memory.
 *
 *  For parallel execution, use one separate ZSTD_CStream per thread.
 *
 *  note : since v1.3.0, ZSTD_CStream and ZSTD_CCtx are the same thing.
 *
 *  Parameters are sticky : when starting a new compression on the same context,
-*  it will re-use the same sticky parameters as previous compression session.
+*  it will reuse the same sticky parameters as previous compression session.
 *  When in doubt, it's recommended to fully initialize the context before usage.
 *  Use ZSTD_CCtx_reset() to reset the context and ZSTD_CCtx_setParameter(),
 *  ZSTD_CCtx_setPledgedSrcSize(), or ZSTD_CCtx_loadDictionary() and friends to
···
  *            only ZSTD_e_end or ZSTD_e_flush operations are allowed.
  *            Before starting a new compression job, or changing compression parameters,
  *            it is required to fully flush internal buffers.
+ *  - note: if an operation ends with an error, it may leave @cctx in an undefined state.
+ *          Therefore, it's UB to invoke ZSTD_compressStream2() of ZSTD_compressStream() on such a state.
+ *          In order to be re-employed after an error, a state must be reset,
+ *          which can be done explicitly (ZSTD_CCtx_reset()),
+ *          or is sometimes implied by methods starting a new compression job (ZSTD_initCStream(), ZSTD_compressCCtx())
  */
 ZSTDLIB_API size_t ZSTD_compressStream2( ZSTD_CCtx* cctx,
                                          ZSTD_outBuffer* output,
···
  * This following is a legacy streaming API, available since v1.0+ .
  * It can be replaced by ZSTD_CCtx_reset() and ZSTD_compressStream2().
  * It is redundant, but remains fully supported.
- * Streaming in combination with advanced parameters and dictionary compression
- * can only be used through the new API.
  ******************************************************************************/
 
 /*!
···
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_refCDict(zcs, NULL); // clear the dictionary (if any)
  *     ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel);
+ *
+ * Note that ZSTD_initCStream() clears any previously set dictionary. Use the new API
+ * to compress with a dictionary.
  */
 ZSTDLIB_API size_t ZSTD_initCStream(ZSTD_CStream* zcs, int compressionLevel);
 /*!
···
 *
 *  A ZSTD_DStream object is required to track streaming operations.
 *  Use ZSTD_createDStream() and ZSTD_freeDStream() to create/release resources.
-*  ZSTD_DStream objects can be re-used multiple times.
+*  ZSTD_DStream objects can be re-employed multiple times.
 *
 *  Use ZSTD_initDStream() to start a new decompression operation.
 * @return : recommended first input size
···
 *  The function will update both `pos` fields.
 *  If `input.pos < input.size`, some input has not been consumed.
 *  It's up to the caller to present again remaining data.
+*
 *  The function tries to flush all data decoded immediately, respecting output buffer size.
 *  If `output.pos < output.size`, decoder has flushed everything it could.
-*  But if `output.pos == output.size`, there might be some data left within internal buffers.,
+*
+*  However, when `output.pos == output.size`, it's more difficult to know.
+*  If @return > 0, the frame is not complete, meaning
+*  either there is still some data left to flush within internal buffers,
+*  or there is more input to read to complete the frame (or both).
 *  In which case, call ZSTD_decompressStream() again to flush whatever remains in the buffer.
 *  Note : with no additional input provided, amount of data flushed is necessarily <= ZSTD_BLOCKSIZE_MAX.
 * @return : 0 when a frame is completely decoded and fully flushed,
 *        or an error code, which can be tested using ZSTD_isError(),
 *        or any other value > 0, which means there is still some decoding or flushing to do to complete current frame :
 *                                the return value is a suggested next input size (just a hint for better latency)
-*                                that will never request more than the remaining frame size.
+*                                that will never request more than the remaining content of the compressed frame.
 * *******************************************************************************/
 
 typedef ZSTD_DCtx ZSTD_DStream;  /*< DCtx and DStream are now effectively same object (>= v1.3.0) */
···
 
 /*===== Streaming decompression functions =====*/
 
-/* This function is redundant with the advanced API and equivalent to:
+/*! ZSTD_initDStream() :
+ * Initialize/reset DStream state for new decompression operation.
+ * Call before new decompression operation using same DStream.
  *
+ * Note : This function is redundant with the advanced API and equivalent to:
  *     ZSTD_DCtx_reset(zds, ZSTD_reset_session_only);
  *     ZSTD_DCtx_refDDict(zds, NULL);
  */
 ZSTDLIB_API size_t ZSTD_initDStream(ZSTD_DStream* zds);
 
+/*! ZSTD_decompressStream() :
+ * Streaming decompression function.
+ * Call repetitively to consume full input updating it as necessary.
+ * Function will update both input and output `pos` fields exposing current state via these fields:
+ * - `input.pos < input.size`, some input remaining and caller should provide remaining input
+ *   on the next call.
+ * - `output.pos < output.size`, decoder flushed internal output buffer.
+ * - `output.pos == output.size`, unflushed data potentially present in the internal buffers,
+ *   check ZSTD_decompressStream() @return value,
+ *   if > 0, invoke it again to flush remaining data to output.
+ * Note : with no additional input, amount of data flushed <= ZSTD_BLOCKSIZE_MAX.
+ *
+ * @return : 0 when a frame is completely decoded and fully flushed,
+ *           or an error code, which can be tested using ZSTD_isError(),
+ *           or any other value > 0, which means there is some decoding or flushing to do to complete current frame.
+ *
+ * Note: when an operation returns with an error code, the @zds state may be left in undefined state.
+ *       It's UB to invoke `ZSTD_decompressStream()` on such a state.
+ *       In order to re-use such a state, it must be first reset,
+ *       which can be done explicitly (`ZSTD_DCtx_reset()`),
+ *       or is implied for operations starting some new decompression job (`ZSTD_initDStream`, `ZSTD_decompressDCtx()`, `ZSTD_decompress_usingDict()`)
+ */
 ZSTDLIB_API size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inBuffer* input);
 
 ZSTDLIB_API size_t ZSTD_DStreamInSize(void);    /*!< recommended size for input buffer */
···
  *  If @return == 0, the dictID could not be decoded.
  *  This could for one of the following reasons :
  *  - The frame does not require a dictionary to be decoded (most common case).
- *  - The frame was built with dictID intentionally removed. Whatever dictionary is necessary is a hidden information.
+ *  - The frame was built with dictID intentionally removed. Whatever dictionary is necessary is a hidden piece of information.
  *    Note : this use case also happens when using a non-conformant dictionary.
  *  - `srcSize` is too small, and as a result, the frame header could not be decoded (only possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`).
  *  - This is not a Zstandard frame.
···
  * Advanced dictionary and prefix API (Requires v1.4.0+)
  *
  * This API allows dictionaries to be used with ZSTD_compress2(),
- * ZSTD_compressStream2(), and ZSTD_decompressDCtx(). Dictionaries are sticky, and
- * only reset with the context is reset with ZSTD_reset_parameters or
- * ZSTD_reset_session_and_parameters. Prefixes are single-use.
+ * ZSTD_compressStream2(), and ZSTD_decompressDCtx().
+ * Dictionaries are sticky, they remain valid when same context is reused,
+ * they only reset when the context is reset
+ * with ZSTD_reset_parameters or ZSTD_reset_session_and_parameters.
+ * In contrast, Prefixes are single-use.
  ******************************************************************************/
 
 
···
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  *  Special: Loading a NULL (or 0-size) dictionary invalidates previous dictionary,
  *           meaning "return to no-dictionary mode".
- *  Note 1 : Dictionary is sticky, it will be used for all future compressed frames.
- *           To return to "no-dictionary" situation, load a NULL dictionary (or reset parameters).
+ *  Note 1 : Dictionary is sticky, it will be used for all future compressed frames,
+ *           until parameters are reset, a new dictionary is loaded, or the dictionary
+ *           is explicitly invalidated by loading a NULL dictionary.
  *  Note 2 : Loading a dictionary involves building tables.
  *           It's also a CPU consuming operation, with non-negligible impact on latency.
  *           Tables are dependent on compression parameters, and for this reason,
···
  *           Use experimental ZSTD_CCtx_loadDictionary_byReference() to reference content instead.
  *           In such a case, dictionary buffer must outlive its users.
  *  Note 4 : Use ZSTD_CCtx_loadDictionary_advanced()
- *           to precisely select how dictionary content must be interpreted. */
+ *           to precisely select how dictionary content must be interpreted.
+ *  Note 5 : This method does not benefit from LDM (long distance mode).
+ *           If you want to employ LDM on some large dictionary content,
+ *           prefer employing ZSTD_CCtx_refPrefix() described below.
+ */
 ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
 
 /*! ZSTD_CCtx_refCDict() : Requires v1.4.0+
- *  Reference a prepared dictionary, to be used for all next compressed frames.
+ *  Reference a prepared dictionary, to be used for all future compressed frames.
  *  Note that compression parameters are enforced from within CDict,
  *  and supersede any compression parameter previously set within CCtx.
  *  The parameters ignored are labelled as "superseded-by-cdict" in the ZSTD_cParameter enum docs.
···
  *  Decompression will need same prefix to properly regenerate data.
  *  Compressing with a prefix is similar in outcome as performing a diff and compressing it,
  *  but performs much faster, especially during decompression (compression speed is tunable with compression level).
+ *  This method is compatible with LDM (long distance mode).
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  *  Special: Adding any prefix (including NULL) invalidates any previous prefix or dictionary
  *  Note 1 : Prefix buffer is referenced. It **must** outlive compression.
···
                                  const void* prefix, size_t prefixSize);
 
 /*! ZSTD_DCtx_loadDictionary() : Requires v1.4.0+
- *  Create an internal DDict from dict buffer,
- *  to be used to decompress next frames.
- *  The dictionary remains valid for all future frames, until explicitly invalidated.
+ *  Create an internal DDict from dict buffer, to be used to decompress all future frames.
+ *  The dictionary remains valid for all future frames, until explicitly invalidated, or
+ *  a new dictionary is loaded.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  *  Special : Adding a NULL (or 0-size) dictionary invalidates any previous dictionary,
  *            meaning "return to no-dictionary mode".
···
  *  The memory for the table is allocated on the first call to refDDict, and can be
  *  freed with ZSTD_freeDCtx().
  *
+ *  If called with ZSTD_d_refMultipleDDicts disabled (the default), only one dictionary
+ *  will be managed, and referencing a dictionary effectively "discards" any previous one.
+ *
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
- *  Note 1 : Currently, only one dictionary can be managed.
- *           Referencing a new dictionary effectively "discards" any previous one.
  *  Special: referencing a NULL DDict means "return to no-dictionary mode".
  *  Note 2 : DDict is just referenced, its lifetime must outlive its usage from DCtx.
  */
···
 ZSTDLIB_API size_t ZSTD_sizeof_CDict(const ZSTD_CDict* cdict);
 ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
 
+
 #endif  /* ZSTD_H_235446 */
 
 
···
 #if !defined(ZSTD_H_ZSTD_STATIC_LINKING_ONLY)
 #define ZSTD_H_ZSTD_STATIC_LINKING_ONLY
 
+
 /* This can be overridden externally to hide static symbols. */
 #ifndef ZSTDLIB_STATIC_API
 #define ZSTDLIB_STATIC_API ZSTDLIB_VISIBLE
 #endif
-
-/* Deprecation warnings :
- * Should these 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.
- * Otherwise, it's also possible to define ZSTD_DISABLE_DEPRECATE_WARNINGS.
- */
-#ifdef ZSTD_DISABLE_DEPRECATE_WARNINGS
-#  define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API  /* disable deprecation warnings */
-#else
-#  if (defined(GNUC) && (GNUC > 4 || (GNUC == 4 && GNUC_MINOR >= 5))) || defined(__clang__)
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API __attribute__((deprecated(message)))
-#  elif (__GNUC__ >= 3)
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API __attribute__((deprecated))
-#  else
-#    pragma message("WARNING: You need to implement ZSTD_DEPRECATED for this compiler")
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API
-#  endif
-#endif /* ZSTD_DISABLE_DEPRECATE_WARNINGS */
 
 /* **************************************************************************************
  *   experimental API (static linking only)
···
 #define ZSTD_TARGETLENGTH_MIN     0   /* note : comparing this constant to an unsigned results in a tautological test */
 #define ZSTD_STRATEGY_MIN        ZSTD_fast
 #define ZSTD_STRATEGY_MAX        ZSTD_btultra2
+#define ZSTD_BLOCKSIZE_MAX_MIN (1 << 10) /* The minimum valid max blocksize. Maximum blocksizes smaller than this make compressBound() inaccurate. */
 
 
 #define ZSTD_OVERLAPLOG_MIN       0
···
 #define ZSTD_LDM_HASHRATELOG_MAX (ZSTD_WINDOWLOG_MAX - ZSTD_HASHLOG_MIN)
 
 /* Advanced parameter bounds */
-#define ZSTD_TARGETCBLOCKSIZE_MIN   64
+#define ZSTD_TARGETCBLOCKSIZE_MIN   1340 /* suitable to fit into an ethernet / wifi / 4G transport frame */
 #define ZSTD_TARGETCBLOCKSIZE_MAX   ZSTD_BLOCKSIZE_MAX
 #define ZSTD_SRCSIZEHINT_MIN        0
 #define ZSTD_SRCSIZEHINT_MAX        INT_MAX
···
                                *
                                * Note: This field is optional. ZSTD_generateSequences() will calculate the value of
                                * 'rep', but repeat offsets do not necessarily need to be calculated from an external
-                               * sequence provider's perspective. For example, ZSTD_compressSequences() does not
+                               * sequence provider perspective. For example, ZSTD_compressSequences() does not
                                * use this 'rep' field at all (as of now).
                                */
 } ZSTD_Sequence;
···
 } ZSTD_literalCompressionMode_e;
 
 typedef enum {
-  /* Note: This enum controls features which are conditionally beneficial. Zstd typically will make a final
-   * decision on whether or not to enable the feature (ZSTD_ps_auto), but setting the switch to ZSTD_ps_enable
-   * or ZSTD_ps_disable allow for a force enable/disable the feature.
+  /* Note: This enum controls features which are conditionally beneficial.
+   * Zstd can take a decision on whether or not to enable the feature (ZSTD_ps_auto),
+   * but setting the switch to ZSTD_ps_enable or ZSTD_ps_disable force enable/disable the feature.
    */
   ZSTD_ps_auto = 0,         /* Let the library automatically determine whether the feature shall be enabled */
   ZSTD_ps_enable = 1,       /* Force-enable the feature */
   ZSTD_ps_disable = 2       /* Do not use the feature */
-} ZSTD_paramSwitch_e;
+} ZSTD_ParamSwitch_e;
+#define ZSTD_paramSwitch_e ZSTD_ParamSwitch_e  /* old name */
 
 /* *************************************
-*  Frame size functions
+*  Frame header and size functions
 ***************************************/
 
 /*! ZSTD_findDecompressedSize() :
···
 ZSTDLIB_STATIC_API unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize);
 
 /*! ZSTD_frameHeaderSize() :
- *  srcSize must be >= ZSTD_FRAMEHEADERSIZE_PREFIX.
+ *  srcSize must be large enough, aka >= ZSTD_FRAMEHEADERSIZE_PREFIX.
  * @return : size of the Frame Header,
  *           or an error code (if srcSize is too small) */
 ZSTDLIB_STATIC_API size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize);
 
+typedef enum { ZSTD_frame, ZSTD_skippableFrame } ZSTD_FrameType_e;
+#define ZSTD_frameType_e ZSTD_FrameType_e /* old name */
+typedef struct {
+    unsigned long long frameContentSize; /* if == ZSTD_CONTENTSIZE_UNKNOWN, it means this field is not available. 0 means "empty" */
+    unsigned long long windowSize;       /* can be very large, up to <= frameContentSize */
+    unsigned blockSizeMax;
+    ZSTD_FrameType_e frameType;          /* if == ZSTD_skippableFrame, frameContentSize is the size of skippable content */
+    unsigned headerSize;
+    unsigned dictID;                     /* for ZSTD_skippableFrame, contains the skippable magic variant [0-15] */
+    unsigned checksumFlag;
+    unsigned _reserved1;
+    unsigned _reserved2;
+} ZSTD_FrameHeader;
+#define ZSTD_frameHeader ZSTD_FrameHeader /* old name */
+
+/*! ZSTD_getFrameHeader() :
+ *  decode Frame Header into `zfhPtr`, or requires larger `srcSize`.
+ * @return : 0 => header is complete, `zfhPtr` is correctly filled,
+ *          >0 => `srcSize` is too small, @return value is the wanted `srcSize` amount, `zfhPtr` is not filled,
+ *           or an error code, which can be tested using ZSTD_isError() */
+ZSTDLIB_STATIC_API size_t ZSTD_getFrameHeader(ZSTD_FrameHeader* zfhPtr, const void* src, size_t srcSize);
+/*! ZSTD_getFrameHeader_advanced() :
+ *  same as ZSTD_getFrameHeader(),
+ *  with added capability to select a format (like ZSTD_f_zstd1_magicless) */
+ZSTDLIB_STATIC_API size_t ZSTD_getFrameHeader_advanced(ZSTD_FrameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format);
+
+/*! ZSTD_decompressionMargin() :
+ * Zstd supports in-place decompression, where the input and output buffers overlap.
+ * In this case, the output buffer must be at least (Margin + Output_Size) bytes large,
+ * and the input buffer must be at the end of the output buffer.
+ *
+ *  _______________________ Output Buffer ________________________
+ * |                                                              |
+ * |                                        ____ Input Buffer ____|
+ * |                                       |                      |
+ * v                                       v                      v
+ * |---------------------------------------|-----------|----------|
+ * ^                                                   ^          ^
+ * |___________________ Output_Size ___________________|_ Margin _|
+ *
+ * NOTE: See also ZSTD_DECOMPRESSION_MARGIN().
+ * NOTE: This applies only to single-pass decompression through ZSTD_decompress() or
+ * ZSTD_decompressDCtx().
+ * NOTE: This function supports multi-frame input.
+ *
+ * @param src The compressed frame(s)
+ * @param srcSize The size of the compressed frame(s)
+ * @returns The decompression margin or an error that can be checked with ZSTD_isError().
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_decompressionMargin(const void* src, size_t srcSize);
+
+/*! ZSTD_DECOMPRESS_MARGIN() :
+ * Similar to ZSTD_decompressionMargin(), but instead of computing the margin from
+ * the compressed frame, compute it from the original size and the blockSizeLog.
+ * See ZSTD_decompressionMargin() for details.
+ *
+ * WARNING: This macro does not support multi-frame input, the input must be a single
+ * zstd frame. If you need that support use the function, or implement it yourself.
+ *
+ * @param originalSize The original uncompressed size of the data.
+ * @param blockSize    The block size == MIN(windowSize, ZSTD_BLOCKSIZE_MAX).
+ *                     Unless you explicitly set the windowLog smaller than
+ *                     ZSTD_BLOCKSIZELOG_MAX you can just use ZSTD_BLOCKSIZE_MAX.
+ */
+#define ZSTD_DECOMPRESSION_MARGIN(originalSize, blockSize) ((size_t)(                                              \
+        ZSTD_FRAMEHEADERSIZE_MAX                                                              /* Frame header */ + \
+        4                                                                                         /* checksum */ + \
+        ((originalSize) == 0 ? 0 : 3 * (((originalSize) + (blockSize) - 1) / blockSize)) /* 3 bytes per block */ + \
+        (blockSize)                                                                    /* One block of margin */   \
+    ))
+
 typedef enum {
-  ZSTD_sf_noBlockDelimiters = 0,         /* Representation of ZSTD_Sequence has no block delimiters, sequences only */
-  ZSTD_sf_explicitBlockDelimiters = 1    /* Representation of ZSTD_Sequence contains explicit block delimiters */
-} ZSTD_sequenceFormat_e;
+  ZSTD_sf_noBlockDelimiters = 0,         /* ZSTD_Sequence[] has no block delimiters, just sequences */
+  ZSTD_sf_explicitBlockDelimiters = 1    /* ZSTD_Sequence[] contains explicit block delimiters */
+} ZSTD_SequenceFormat_e;
+#define ZSTD_sequenceFormat_e ZSTD_SequenceFormat_e /* old name */
+
+/*! ZSTD_sequenceBound() :
+ * `srcSize` : size of the input buffer
+ *  @return : upper-bound for the number of sequences that can be generated
+ *            from a buffer of srcSize bytes
+ *
+ *  note : returns number of sequences - to get bytes, multiply by sizeof(ZSTD_Sequence).
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_sequenceBound(size_t srcSize);
 
 /*! ZSTD_generateSequences() :
- * Generate sequences using ZSTD_compress2, given a source buffer.
+ * WARNING: This function is meant for debugging and informational purposes ONLY!
+ * Its implementation is flawed, and it will be deleted in a future version.
+ * It is not guaranteed to succeed, as there are several cases where it will give
+ * up and fail. You should NOT use this function in production code.
+ *
+ * This function is deprecated, and will be removed in a future version.
+ *
+ * Generate sequences using ZSTD_compress2(), given a source buffer.
+ *
+ * @param zc The compression context to be used for ZSTD_compress2(). Set any
+ *           compression parameters you need on this context.
+ * @param outSeqs The output sequences buffer of size @p outSeqsSize
+ * @param outSeqsCapacity The size of the output sequences buffer.
+ *                    ZSTD_sequenceBound(srcSize) is an upper bound on the number
+ *                    of sequences that can be generated.
+ * @param src The source buffer to generate sequences from of size @p srcSize.
+ * @param srcSize The size of the source buffer.
  *
  * Each block will end with a dummy sequence
  * with offset == 0, matchLength == 0, and litLength == length of last literals.
  * litLength may be == 0, and if so, then the sequence of (of: 0 ml: 0 ll: 0)
  * simply acts as a block delimiter.
  *
- * zc can be used to insert custom compression params.
- * This function invokes ZSTD_compress2
- *
- * The output of this function can be fed into ZSTD_compressSequences() with CCtx
- * setting of ZSTD_c_blockDelimiters as ZSTD_sf_explicitBlockDelimiters
- * @return : number of sequences generated
+ * @returns The number of sequences generated, necessarily less than
+ *          ZSTD_sequenceBound(srcSize), or an error code that can be checked
+ *          with ZSTD_isError().
  */
-
-ZSTDLIB_STATIC_API size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
-                                          size_t outSeqsSize, const void* src, size_t srcSize);
+ZSTD_DEPRECATED("For debugging only, will be replaced by ZSTD_extractSequences()")
+ZSTDLIB_STATIC_API size_t
+ZSTD_generateSequences(ZSTD_CCtx* zc,
+                       ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+                       const void* src, size_t srcSize);
 
 /*! ZSTD_mergeBlockDelimiters() :
  * Given an array of ZSTD_Sequence, remove all sequences that represent block delimiters/last literals
···
 ZSTDLIB_STATIC_API size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, size_t seqsSize);
 
 /*! ZSTD_compressSequences() :
- * Compress an array of ZSTD_Sequence, generated from the original source buffer, into dst.
- * If a dictionary is included, then the cctx should reference the dict. (see: ZSTD_CCtx_refCDict(), ZSTD_CCtx_loadDictionary(), etc.)
+ * Compress an array of ZSTD_Sequence, associated with @src buffer, into dst.
+ * @src contains the entire input (not just the literals).
+ * If @srcSize > sum(sequence.length), the remaining bytes are considered all literals
+ * If a dictionary is included, then the cctx should reference the dict (see: ZSTD_CCtx_refCDict(), ZSTD_CCtx_loadDictionary(), etc.).
  * The entire source is compressed into a single frame.
  *
  * The compression behavior changes based on cctx params. In particular:
···
  *    the block size derived from the cctx, and sequences may be split. This is the default setting.
  *
  *    If ZSTD_c_blockDelimiters == ZSTD_sf_explicitBlockDelimiters, the array of ZSTD_Sequence is expected to contain
- *    block delimiters (defined in ZSTD_Sequence). Behavior is undefined if no block delimiters are provided.
+ *    valid block delimiters (defined in ZSTD_Sequence). Behavior is undefined if no block delimiters are provided.
  *
- *    If ZSTD_c_validateSequences == 0, this function will blindly accept the sequences provided. Invalid sequences cause undefined
- *    behavior. If ZSTD_c_validateSequences == 1, then if sequence is invalid (see doc/zstd_compression_format.md for
- *    specifics regarding offset/matchlength requirements) then the function will bail out and return an error.
+ *    When ZSTD_c_blockDelimiters == ZSTD_sf_explicitBlockDelimiters, it's possible to decide generating repcodes
+ *    using the advanced parameter ZSTD_c_repcodeResolution. Repcodes will improve compression ratio, though the benefit
+ *    can vary greatly depending on Sequences. On the other hand, repcode resolution is an expensive operation.
+ *    By default, it's disabled at low (<10) compression levels, and enabled above the threshold (>=10).
+ *    ZSTD_c_repcodeResolution makes it possible to directly manage this processing in either direction.
+ *
+ *    If ZSTD_c_validateSequences == 0, this function blindly accepts the Sequences provided. Invalid Sequences cause undefined
+ *    behavior. If ZSTD_c_validateSequences == 1, then the function will detect invalid Sequences (see doc/zstd_compression_format.md for
+ *    specifics regarding offset/matchlength requirements) and then bail out and return an error.
  *
  *    In addition to the two adjustable experimental params, there are other important cctx params.
  *    - ZSTD_c_minMatch MUST be set as less than or equal to the smallest match generated by the match finder. It has a minimum value of ZSTD_MINMATCH_MIN.
···
  *    - ZSTD_c_windowLog affects offset validation: this function will return an error at higher debug levels if a provided offset
  *      is larger than what the spec allows for a given window log and dictionary (if present). See: doc/zstd_compression_format.md
  *
- * Note: Repcodes are, as of now, always re-calculated within this function, so ZSTD_Sequence::rep is unused.
- * Note 2: Once we integrate ability to ingest repcodes, the explicit block delims mode must respect those repcodes exactly,
- *         and cannot emit an RLE block that disagrees with the repcode history
- * @return : final compressed size or a ZSTD error.
+ * Note: Repcodes are, as of now, always re-calculated within this function, ZSTD_Sequence.rep is effectively unused.
+ * Dev Note: Once ability to ingest repcodes become available, the explicit block delims mode must respect those repcodes exactly,
+ *         and cannot emit an RLE block that disagrees with the repcode history.
+ * @return : final compressed size, or a ZSTD error code.
  */
-ZSTDLIB_STATIC_API size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* dst, size_t dstSize,
-                                  const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
-                                  const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t
+ZSTD_compressSequences(ZSTD_CCtx* cctx,
+                       void* dst, size_t dstCapacity,
+                 const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
+                 const void* src, size_t srcSize);
+
+
+/*! ZSTD_compressSequencesAndLiterals() :
+ * This is a variant of ZSTD_compressSequences() which,
+ * instead of receiving (src,srcSize) as input parameter, receives (literals,litSize),
+ * aka all the literals, already extracted and laid out into a single continuous buffer.
+ * This can be useful if the process generating the sequences also happens to generate the buffer of literals,
+ * thus skipping an extraction + caching stage.
+ * It's a speed optimization, useful when the right conditions are met,
+ * but it also features the following limitations:
+ * - Only supports explicit delimiter mode
+ * - Currently does not support Sequences validation (so input Sequences are trusted)
+ * - Not compatible with frame checksum, which must be disabled
+ * - If any block is incompressible, will fail and return an error
+ * - @litSize must be == sum of all @.litLength fields in @inSeqs. Any discrepancy will generate an error.
+ * - @litBufCapacity is the size of the underlying buffer into which literals are written, starting at address @literals.
+ *   @litBufCapacity must be at least 8 bytes larger than @litSize.
+ * - @decompressedSize must be correct, and correspond to the sum of all Sequences. Any discrepancy will generate an error.
+ * @return : final compressed size, or a ZSTD error code.
+ */
+ZSTDLIB_STATIC_API size_t
+ZSTD_compressSequencesAndLiterals(ZSTD_CCtx* cctx,
+                                  void* dst, size_t dstCapacity,
+                            const ZSTD_Sequence* inSeqs, size_t nbSequences,
+                            const void* literals, size_t litSize, size_t litBufCapacity,
+                            size_t decompressedSize);
 
 
 /*! ZSTD_writeSkippableFrame() :
···
  *
  * Skippable frames begin with a 4-byte magic number. There are 16 possible choices of magic number,
  * ranging from ZSTD_MAGIC_SKIPPABLE_START to ZSTD_MAGIC_SKIPPABLE_START+15.
- * As such, the parameter magicVariant controls the exact skippable frame magic number variant used, so
- * the magic number used will be ZSTD_MAGIC_SKIPPABLE_START + magicVariant.
+ * As such, the parameter magicVariant controls the exact skippable frame magic number variant used,
+ * so the magic number used will be ZSTD_MAGIC_SKIPPABLE_START + magicVariant.
  *
  * Returns an error if destination buffer is not large enough, if the source size is not representable
  * with a 4-byte unsigned int, or if the parameter magicVariant is greater than 15 (and therefore invalid).
···
  * @return : number of bytes written or a ZSTD error.
  */
 ZSTDLIB_STATIC_API size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity,
-                                            const void* src, size_t srcSize, unsigned magicVariant);
+                                             const void* src, size_t srcSize,
+                                                   unsigned magicVariant);
 
 /*! ZSTD_readSkippableFrame() :
- * Retrieves a zstd skippable frame containing data given by src, and writes it to dst buffer.
+ * Retrieves the content of a zstd skippable frame starting at @src, and writes it to @dst buffer.
  *
- * The parameter magicVariant will receive the magicVariant that was supplied when the frame was written,
- * i.e. magicNumber - ZSTD_MAGIC_SKIPPABLE_START.  This can be NULL if the caller is not interested
- * in the magicVariant.
+ * The parameter @magicVariant will receive the magicVariant that was supplied when the frame was written,
+ * i.e. magicNumber - ZSTD_MAGIC_SKIPPABLE_START.
+ * This can be NULL if the caller is not interested in the magicVariant.
  *
  * Returns an error if destination buffer is not large enough, or if the frame is not skippable.
  *
  * @return : number of bytes written or a ZSTD error.
  */
-ZSTDLIB_API size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity, unsigned* magicVariant,
-                                            const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity,
+                                                  unsigned* magicVariant,
+                                                  const void* src, size_t srcSize);
 
 /*! ZSTD_isSkippableFrame() :
  *  Tells if the content of `buffer` starts with a valid Frame Identifier for a skippable frame.
  */
-ZSTDLIB_API unsigned ZSTD_isSkippableFrame(const void* buffer, size_t size);
+ZSTDLIB_STATIC_API unsigned ZSTD_isSkippableFrame(const void* buffer, size_t size);
 
 
 
···
 /*! ZSTD_estimate*() :
  *  These functions make it possible to estimate memory usage
  *  of a future {D,C}Ctx, before its creation.
+ *  This is useful in combination with ZSTD_initStatic(),
+ *  which makes it possible to employ a static buffer for ZSTD_CCtx* state.
  *
  *  ZSTD_estimateCCtxSize() will provide a memory budget large enough
- *  for any compression level up to selected one.
- *  Note : Unlike ZSTD_estimateCStreamSize*(), this estimate
- *         does not include space for a window buffer.
- *         Therefore, the estimation is only guaranteed for single-shot compressions, not streaming.
+ *  to compress data of any size using one-shot compression ZSTD_compressCCtx() or ZSTD_compress2()
+ *  associated with any compression level up to max specified one.
  *  The estimate will assume the input may be arbitrarily large,
  *  which is the worst case.
  *
+ *  Note that the size estimation is specific for one-shot compression,
+ *  it is not valid for streaming (see ZSTD_estimateCStreamSize*())
+ *  nor other potential ways of using a ZSTD_CCtx* state.
+ *
  *  When srcSize can be bound by a known and rather "small" value,
- *  this fact can be used to provide a tighter estimation
- *  because the CCtx compression context will need less memory.
- *  This tighter estimation can be provided by more advanced functions
+ *  this knowledge can be used to provide a tighter budget estimation
+ *  because the ZSTD_CCtx* state will need less memory for small inputs.
+ *  This tighter estimation can be provided by employing more advanced functions
  *  ZSTD_estimateCCtxSize_usingCParams(), which can be used in tandem with ZSTD_getCParams(),
  *  and ZSTD_estimateCCtxSize_usingCCtxParams(), which can be used in tandem with ZSTD_CCtxParams_setParameter().
  *  Both can be used to estimate memory using custom compression parameters and arbitrary srcSize limits.
  *
- *  Note 2 : only single-threaded compression is supported.
+ *  Note : only single-threaded compression is supported.
  *  ZSTD_estimateCCtxSize_usingCCtxParams() will return an error code if ZSTD_c_nbWorkers is >= 1.
  */
-ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize(int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize(int maxCompressionLevel);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateDCtxSize(void);
 
 /*! ZSTD_estimateCStreamSize() :
- *  ZSTD_estimateCStreamSize() will provide a budget large enough for any compression level up to selected one.
- *  It will also consider src size to be arbitrarily "large", which is worst case.
+ *  ZSTD_estimateCStreamSize() will provide a memory budget large enough for streaming compression
+ *  using any compression level up to the max specified one.
+ *  It will also consider src size to be arbitrarily "large", which is a worst case scenario.
  *  If srcSize is known to always be small, ZSTD_estimateCStreamSize_usingCParams() can provide a tighter estimation.
  *  ZSTD_estimateCStreamSize_usingCParams() can be used in tandem with ZSTD_getCParams() to create cParams from compressionLevel.
  *  ZSTD_estimateCStreamSize_usingCCtxParams() can be used in tandem with ZSTD_CCtxParams_setParameter(). Only single-threaded compression is supported. This function will return an error code if ZSTD_c_nbWorkers is >= 1.
  *  Note : CStream size estimation is only correct for single-threaded compression.
- *  ZSTD_DStream memory budget depends on window Size.
+ *  ZSTD_estimateCStreamSize_usingCCtxParams() will return an error code if ZSTD_c_nbWorkers is >= 1.
+ *  Note 2 : ZSTD_estimateCStreamSize* functions are not compatible with the Block-Level Sequence Producer API at this time.
+ *  Size estimates assume that no external sequence producer is registered.
+ *
+ *  ZSTD_DStream memory budget depends on frame's window Size.
  *  This information can be passed manually, using ZSTD_estimateDStreamSize,
  *  or deducted from a valid frame Header, using ZSTD_estimateDStreamSize_fromFrame();
+ *  Any frame requesting a window size larger than max specified one will be rejected.
  *  Note : if streaming is init with function ZSTD_init?Stream_usingDict(),
  *         an internal ?Dict will be created, which additional size is not estimated here.
- *         In this case, get total size by adding ZSTD_estimate?DictSize */
-ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize(int compressionLevel);
+ *         In this case, get total size by adding ZSTD_estimate?DictSize
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize(int maxCompressionLevel);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params);
-ZSTDLIB_STATIC_API size_t ZSTD_estimateDStreamSize(size_t windowSize);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDStreamSize(size_t maxWindowSize);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize);
 
 /*! ZSTD_estimate?DictSize() :
···
 typedef struct { ZSTD_allocFunction customAlloc; ZSTD_freeFunction customFree; void* opaque; } ZSTD_customMem;
 static
 __attribute__((__unused__))
+
+#if defined(__clang__) && __clang_major__ >= 5
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wzero-as-null-pointer-constant"
+#endif
 ZSTD_customMem const ZSTD_defaultCMem = { NULL, NULL, NULL };  /*< this constant defers to stdlib's functions */
+#if defined(__clang__) && __clang_major__ >= 5
+#pragma clang diagnostic pop
+#endif
 
 ZSTDLIB_STATIC_API ZSTD_CCtx*    ZSTD_createCCtx_advanced(ZSTD_customMem customMem);
 ZSTDLIB_STATIC_API ZSTD_CStream* ZSTD_createCStream_advanced(ZSTD_customMem customMem);
···
  *  This function never fails (wide contract) */
 ZSTDLIB_STATIC_API ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize);
 
+/*! ZSTD_CCtx_setCParams() :
+ *  Set all parameters provided within @p cparams into the working @p cctx.
+ *  Note : if modifying parameters during compression (MT mode only),
+ *         note that changes to the .windowLog parameter will be ignored.
+ * @return 0 on success, or an error code (can be checked with ZSTD_isError()).
+ *         On failure, no parameters are updated.
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_setCParams(ZSTD_CCtx* cctx, ZSTD_compressionParameters cparams);
+
+/*! ZSTD_CCtx_setFParams() :
+ *  Set all parameters provided within @p fparams into the working @p cctx.
+ * @return 0 on success, or an error code (can be checked with ZSTD_isError()).
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_setFParams(ZSTD_CCtx* cctx, ZSTD_frameParameters fparams);
+
+/*! ZSTD_CCtx_setParams() :
+ *  Set all parameters provided within @p params into the working @p cctx.
+ * @return 0 on success, or an error code (can be checked with ZSTD_isError()).
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_setParams(ZSTD_CCtx* cctx, ZSTD_parameters params);
+
 /*! ZSTD_compress_advanced() :
  *  Note : this function is now DEPRECATED.
  *         It can be replaced by ZSTD_compress2(), in combination with ZSTD_CCtx_setParameter() and other parameter setters.
  *  This prototype will generate compilation warnings. */
 ZSTD_DEPRECATED("use ZSTD_compress2")
+ZSTDLIB_STATIC_API
 size_t ZSTD_compress_advanced(ZSTD_CCtx* cctx,
-                                          void* dst, size_t dstCapacity,
-                                    const void* src, size_t srcSize,
-                                    const void* dict,size_t dictSize,
-                                          ZSTD_parameters params);
+                              void* dst, size_t dstCapacity,
+                        const void* src, size_t srcSize,
+                        const void* dict,size_t dictSize,
+                              ZSTD_parameters params);
 
 /*! ZSTD_compress_usingCDict_advanced() :
  *  Note : this function is now DEPRECATED.
  *         It can be replaced by ZSTD_compress2(), in combination with ZSTD_CCtx_loadDictionary() and other parameter setters.
  *  This prototype will generate compilation warnings. */
 ZSTD_DEPRECATED("use ZSTD_compress2 with ZSTD_CCtx_loadDictionary")
+ZSTDLIB_STATIC_API
 size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
                                               void* dst, size_t dstCapacity,
                                         const void* src, size_t srcSize,
···
  * See the comments on that enum for an explanation of the feature. */
 #define ZSTD_c_forceAttachDict ZSTD_c_experimentalParam4
 
-/* Controlled with ZSTD_paramSwitch_e enum.
+/* Controlled with ZSTD_ParamSwitch_e enum.
  * Default is ZSTD_ps_auto.
  * Set to ZSTD_ps_disable to never compress literals.
  * Set to ZSTD_ps_enable to always compress literals. (Note: uncompressed literals
···
  * negative compression levels do not use literal compression.
  */
 #define ZSTD_c_literalCompressionMode ZSTD_c_experimentalParam5
-
-/* Tries to fit compressed block size to be around targetCBlockSize.
- * No target when targetCBlockSize == 0.
- * There is no guarantee on compressed block size (default:0) */
-#define ZSTD_c_targetCBlockSize ZSTD_c_experimentalParam6
 
 /* User's best guess of source size.
  * Hint is not valid when srcSizeHint == 0.
···
  * Experimental parameter.
  * Default is 0 == disabled. Set to 1 to enable.
  *
- * Tells the compressor that the ZSTD_inBuffer will ALWAYS be the same
- * between calls, except for the modifications that zstd makes to pos (the
- * caller must not modify pos). This is checked by the compressor, and
- * compression will fail if it ever changes. This means the only flush
- * mode that makes sense is ZSTD_e_end, so zstd will error if ZSTD_e_end
- * is not used. The data in the ZSTD_inBuffer in the range [src, src + pos)
- * MUST not be modified during compression or you will get data corruption.
+ * Tells the compressor that input data presented with ZSTD_inBuffer
+ * will ALWAYS be the same between calls.
+ * Technically, the @src pointer must never be changed,
+ * and the @pos field can only be updated by zstd.
+ * However, it's possible to increase the @size field,
+ * allowing scenarios where more data can be appended after compressions starts.
+ * These conditions are checked by the compressor,
+ * and compression will fail if they are not respected.
+ * Also, data in the ZSTD_inBuffer within the range [src, src + pos)
+ * MUST not be modified during compression or it will result in data corruption.
  *
  * When this flag is enabled zstd won't allocate an input window buffer,
  * because the user guarantees it can reference the ZSTD_inBuffer until
···
  * large enough to fit a block (see ZSTD_c_stableOutBuffer). This will also
  * avoid the memcpy() from the input buffer to the input window buffer.
  *
- * NOTE: ZSTD_compressStream2() will error if ZSTD_e_end is not used.
- * That means this flag cannot be used with ZSTD_compressStream().
- *
  * NOTE: So long as the ZSTD_inBuffer always points to valid memory, using
  * this flag is ALWAYS memory safe, and will never access out-of-bounds
- * memory. However, compression WILL fail if you violate the preconditions.
+ * memory. However, compression WILL fail if conditions are not respected.
  *
- * WARNING: The data in the ZSTD_inBuffer in the range [dst, dst + pos) MUST
- * not be modified during compression or you will get data corruption. This
- * is because zstd needs to reference data in the ZSTD_inBuffer to find
+ * WARNING: The data in the ZSTD_inBuffer in the range [src, src + pos) MUST
+ * not be modified during compression or it will result in data corruption.
+ * This is because zstd needs to reference data in the ZSTD_inBuffer to find
  * matches. Normally zstd maintains its own window buffer for this purpose,
- * but passing this flag tells zstd to use the user provided buffer.
+ * but passing this flag tells zstd to rely on user provided buffer instead.
  */
 #define ZSTD_c_stableInBuffer ZSTD_c_experimentalParam9
 
···
 /* ZSTD_c_validateSequences
  * Default is 0 == disabled. Set to 1 to enable sequence validation.
  *
- * For use with sequence compression API: ZSTD_compressSequences().
- * Designates whether or not we validate sequences provided to ZSTD_compressSequences()
+ * For use with sequence compression API: ZSTD_compressSequences*().
+ * Designates whether or not provided sequences are validated within ZSTD_compressSequences*()
  * during function execution.
  *
- * Without validation, providing a sequence that does not conform to the zstd spec will cause
- * undefined behavior, and may produce a corrupted block.
+ * When Sequence validation is disabled (default), Sequences are compressed as-is,
+ * so they must correct, otherwise it would result in a corruption error.
  *
- * With validation enabled, a if sequence is invalid (see doc/zstd_compression_format.md for
+ * Sequence validation adds some protection, by ensuring that all values respect boundary conditions.
+ * If a Sequence is detected invalid (see doc/zstd_compression_format.md for
  * specifics regarding offset/matchlength requirements) then the function will bail out and
  * return an error.
- *
  */
 #define ZSTD_c_validateSequences ZSTD_c_experimentalParam12
 
-/* ZSTD_c_useBlockSplitter
- * Controlled with ZSTD_paramSwitch_e enum.
+/* ZSTD_c_blockSplitterLevel
+ * note: this parameter only influences the first splitter stage,
+ *       which is active before producing the sequences.
+ *       ZSTD_c_splitAfterSequences controls the next splitter stage,
+ *       which is active after sequence production.
+ *       Note that both can be combined.
+ * Allowed values are between 0 and ZSTD_BLOCKSPLITTER_LEVEL_MAX included.
+ * 0 means "auto", which will select a value depending on current ZSTD_c_strategy.
+ * 1 means no splitting.
+ * Then, values from 2 to 6 are sorted in increasing cpu load order.
+ *
+ * Note that currently the first block is never split,
+ * to ensure expansion guarantees in presence of incompressible data.
+ */
+#define ZSTD_BLOCKSPLITTER_LEVEL_MAX 6
+#define ZSTD_c_blockSplitterLevel ZSTD_c_experimentalParam20
+
+/* ZSTD_c_splitAfterSequences
+ * This is a stronger splitter algorithm,
+ * based on actual sequences previously produced by the selected parser.
+ * It's also slower, and as a consequence, mostly used for high compression levels.
+ * While the post-splitter does overlap with the pre-splitter,
+ * both can nonetheless be combined,
+ * notably with ZSTD_c_blockSplitterLevel at ZSTD_BLOCKSPLITTER_LEVEL_MAX,
+ * resulting in higher compression ratio than just one of them.
+ *
  * Default is ZSTD_ps_auto.
  * Set to ZSTD_ps_disable to never use block splitter.
  * Set to ZSTD_ps_enable to always use block splitter.
···
  * By default, in ZSTD_ps_auto, the library will decide at runtime whether to use
  * block splitting based on the compression parameters.
  */
-#define ZSTD_c_useBlockSplitter ZSTD_c_experimentalParam13
+#define ZSTD_c_splitAfterSequences ZSTD_c_experimentalParam13
 
 /* ZSTD_c_useRowMatchFinder
- * Controlled with ZSTD_paramSwitch_e enum.
+ * Controlled with ZSTD_ParamSwitch_e enum.
  * Default is ZSTD_ps_auto.
  * Set to ZSTD_ps_disable to never use row-based matchfinder.
  * Set to ZSTD_ps_enable to force usage of row-based matchfinder.
···
  * cost to memcpy() the data.
  */
 #define ZSTD_c_deterministicRefPrefix ZSTD_c_experimentalParam15
+
+/* ZSTD_c_prefetchCDictTables
+ * Controlled with ZSTD_ParamSwitch_e enum. Default is ZSTD_ps_auto.
+ *
+ * In some situations, zstd uses CDict tables in-place rather than copying them
+ * into the working context. (See docs on ZSTD_dictAttachPref_e above for details).
+ * In such situations, compression speed is seriously impacted when CDict tables are
+ * "cold" (outside CPU cache). This parameter instructs zstd to prefetch CDict tables
+ * when they are used in-place.
+ *
+ * For sufficiently small inputs, the cost of the prefetch will outweigh the benefit.
+ * For sufficiently large inputs, zstd will by default memcpy() CDict tables
+ * into the working context, so there is no need to prefetch. This parameter is
+ * targeted at a middle range of input sizes, where a prefetch is cheap enough to be
+ * useful but memcpy() is too expensive. The exact range of input sizes where this
+ * makes sense is best determined by careful experimentation.
+ *
+ * Note: for this parameter, ZSTD_ps_auto is currently equivalent to ZSTD_ps_disable,
+ * but in the future zstd may conditionally enable this feature via an auto-detection
+ * heuristic for cold CDicts.
+ * Use ZSTD_ps_disable to opt out of prefetching under any circumstances.
+ */
+#define ZSTD_c_prefetchCDictTables ZSTD_c_experimentalParam16
+
+/* ZSTD_c_enableSeqProducerFallback
+ * Allowed values are 0 (disable) and 1 (enable). The default setting is 0.
+ *
+ * Controls whether zstd will fall back to an internal sequence producer if an
+ * external sequence producer is registered and returns an error code. This fallback
+ * is block-by-block: the internal sequence producer will only be called for blocks
+ * where the external sequence producer returns an error code. Fallback parsing will
+ * follow any other cParam settings, such as compression level, the same as in a
+ * normal (fully-internal) compression operation.
+ *
+ * The user is strongly encouraged to read the full Block-Level Sequence Producer API
+ * documentation (below) before setting this parameter. */
+#define ZSTD_c_enableSeqProducerFallback ZSTD_c_experimentalParam17
+
+/* ZSTD_c_maxBlockSize
+ * Allowed values are between 1KB and ZSTD_BLOCKSIZE_MAX (128KB).
+ * The default is ZSTD_BLOCKSIZE_MAX, and setting to 0 will set to the default.
+ *
+ * This parameter can be used to set an upper bound on the blocksize
+ * that overrides the default ZSTD_BLOCKSIZE_MAX. It cannot be used to set upper
+ * bounds greater than ZSTD_BLOCKSIZE_MAX or bounds lower than 1KB (will make
+ * compressBound() inaccurate). Only currently meant to be used for testing.
+ */
+#define ZSTD_c_maxBlockSize ZSTD_c_experimentalParam18
+
+/* ZSTD_c_repcodeResolution
+ * This parameter only has an effect if ZSTD_c_blockDelimiters is
+ * set to ZSTD_sf_explicitBlockDelimiters (may change in the future).
+ *
+ * This parameter affects how zstd parses external sequences,
+ * provided via the ZSTD_compressSequences*() API
+ * or from an external block-level sequence producer.
+ *
+ * If set to ZSTD_ps_enable, the library will check for repeated offsets within
+ * external sequences, even if those repcodes are not explicitly indicated in
+ * the "rep" field. Note that this is the only way to exploit repcode matches
+ * while using compressSequences*() or an external sequence producer, since zstd
+ * currently ignores the "rep" field of external sequences.
+ *
+ * If set to ZSTD_ps_disable, the library will not exploit repeated offsets in
+ * external sequences, regardless of whether the "rep" field has been set. This
+ * reduces sequence compression overhead by about 25% while sacrificing some
+ * compression ratio.
+ *
+ * The default value is ZSTD_ps_auto, for which the library will enable/disable
+ * based on compression level (currently: level<10 disables, level>=10 enables).
+ */
+#define ZSTD_c_repcodeResolution ZSTD_c_experimentalParam19
+#define ZSTD_c_searchForExternalRepcodes ZSTD_c_experimentalParam19 /* older name */
+
 
 /*! ZSTD_CCtx_getParameter() :
  *  Get the requested compression parameter value, selected by enum ZSTD_cParameter,
···
  * in the range [dst, dst + pos) MUST not be modified during decompression
  * or you will get data corruption.
  *
- * When this flags is enabled zstd won't allocate an output buffer, because
+ * When this flag is enabled zstd won't allocate an output buffer, because
  * it can write directly to the ZSTD_outBuffer, but it will still allocate
  * an input buffer large enough to fit any compressed block. This will also
  * avoid the memcpy() from the internal output buffer to the ZSTD_outBuffer.
···
  */
 #define ZSTD_d_refMultipleDDicts ZSTD_d_experimentalParam4
 
+/* ZSTD_d_disableHuffmanAssembly
+ * Set to 1 to disable the Huffman assembly implementation.
+ * The default value is 0, which allows zstd to use the Huffman assembly
+ * implementation if available.
+ *
+ * This parameter can be used to disable Huffman assembly at runtime.
+ * If you want to disable it at compile time you can define the macro
+ * ZSTD_DISABLE_ASM.
+ */
+#define ZSTD_d_disableHuffmanAssembly ZSTD_d_experimentalParam5
+
+/* ZSTD_d_maxBlockSize
+ * Allowed values are between 1KB and ZSTD_BLOCKSIZE_MAX (128KB).
+ * The default is ZSTD_BLOCKSIZE_MAX, and setting to 0 will set to the default.
+ *
+ * Forces the decompressor to reject blocks whose content size is
+ * larger than the configured maxBlockSize. When maxBlockSize is
+ * larger than the windowSize, the windowSize is used instead.
+ * This saves memory on the decoder when you know all blocks are small.
+ *
+ * This option is typically used in conjunction with ZSTD_c_maxBlockSize.
+ *
+ * WARNING: This causes the decoder to reject otherwise valid frames
+ * that have block sizes larger than the configured maxBlockSize.
+ */
+#define ZSTD_d_maxBlockSize ZSTD_d_experimentalParam6
+
 
 /*! ZSTD_DCtx_setFormat() :
  *  This function is REDUNDANT. Prefer ZSTD_DCtx_setParameter().
···
  *  such ZSTD_f_zstd1_magicless for example.
  * @return : 0, or an error code (which can be tested using ZSTD_isError()). */
 ZSTD_DEPRECATED("use ZSTD_DCtx_setParameter() instead")
+ZSTDLIB_STATIC_API
 size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
 
 /*! ZSTD_decompressStream_simpleArgs() :
···
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
                          int compressionLevel,
                          unsigned long long pledgedSrcSize);
···
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
                      const void* dict, size_t dictSize,
                            int compressionLevel);
 
 /*! ZSTD_initCStream_advanced() :
- * This function is DEPRECATED, and is approximately equivalent to:
+ * This function is DEPRECATED, and is equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
- *     // Pseudocode: Set each zstd parameter and leave the rest as-is.
- *     for ((param, value) : params) {
- *         ZSTD_CCtx_setParameter(zcs, param, value);
- *     }
+ *     ZSTD_CCtx_setParams(zcs, params);
  *     ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize);
  *     ZSTD_CCtx_loadDictionary(zcs, dict, dictSize);
  *
···
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
                     const void* dict, size_t dictSize,
                           ZSTD_parameters params,
···
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset and ZSTD_CCtx_refCDict, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict);
 
 /*! ZSTD_initCStream_usingCDict_advanced() :
- *   This function is DEPRECATED, and is approximately equivalent to:
+ *   This function is DEPRECATED, and is equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
- *     // Pseudocode: Set each zstd frame parameter and leave the rest as-is.
- *     for ((fParam, value) : fParams) {
- *         ZSTD_CCtx_setParameter(zcs, fParam, value);
- *     }
+ *     ZSTD_CCtx_setFParams(zcs, fParams);
  *     ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize);
  *     ZSTD_CCtx_refCDict(zcs, cdict);
  *
···
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset and ZSTD_CCtx_refCDict, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
                                const ZSTD_CDict* cdict,
                                      ZSTD_frameParameters fParams,
···
  *       explicitly specified.
  *
  *  start a new frame, using same parameters from previous frame.
- *  This is typically useful to skip dictionary loading stage, since it will re-use it in-place.
+ *  This is typically useful to skip dictionary loading stage, since it will reuse it in-place.
  *  Note that zcs must be init at least once before using ZSTD_resetCStream().
  *  If pledgedSrcSize is not known at reset time, use macro ZSTD_CONTENTSIZE_UNKNOWN.
  *  If pledgedSrcSize > 0, its value must be correct, as it will be written in header, and controlled at the end.
···
  *  This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pledgedSrcSize);
 
 
···
  *     ZSTD_DCtx_loadDictionary(zds, dict, dictSize);
  *
  * note: no dictionary will be used if dict == NULL or dictSize < 8
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
  */
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset + ZSTD_DCtx_loadDictionary, see zstd.h for detailed instructions")
 ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize);
 
 /*!
···
  *     ZSTD_DCtx_refDDict(zds, ddict);
  *
  * note : ddict is referenced, it must outlive decompression session
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
  */
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset + ZSTD_DCtx_refDDict, see zstd.h for detailed instructions")
 ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const ZSTD_DDict* ddict);
 
 /*!
···
  *
  *     ZSTD_DCtx_reset(zds, ZSTD_reset_session_only);
  *
- * re-use decompression parameters from previous init; saves dictionary loading
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * reuse decompression parameters from previous init; saves dictionary loading
  */
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset, see zstd.h for detailed instructions")
 ZSTDLIB_STATIC_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
 
 
+/* ********************* BLOCK-LEVEL SEQUENCE PRODUCER API *********************
+ *
+ * *** OVERVIEW ***
+ * The Block-Level Sequence Producer API allows users to provide their own custom
+ * sequence producer which libzstd invokes to process each block. The produced list
+ * of sequences (literals and matches) is then post-processed by libzstd to produce
+ * valid compressed blocks.
+ *
+ * This block-level offload API is a more granular complement of the existing
+ * frame-level offload API compressSequences() (introduced in v1.5.1). It offers
+ * an easier migration story for applications already integrated with libzstd: the
+ * user application continues to invoke the same compression functions
+ * ZSTD_compress2() or ZSTD_compressStream2() as usual, and transparently benefits
+ * from the specific advantages of the external sequence producer. For example,
+ * the sequence producer could be tuned to take advantage of known characteristics
+ * of the input, to offer better speed / ratio, or could leverage hardware
+ * acceleration not available within libzstd itself.
+ *
+ * See contrib/externalSequenceProducer for an example program employing the
+ * Block-Level Sequence Producer API.
+ *
+ * *** USAGE ***
+ * The user is responsible for implementing a function of type
+ * ZSTD_sequenceProducer_F. For each block, zstd will pass the following
+ * arguments to the user-provided function:
+ *
+ *   - sequenceProducerState: a pointer to a user-managed state for the sequence
+ *     producer.
+ *
+ *   - outSeqs, outSeqsCapacity: an output buffer for the sequence producer.
+ *     outSeqsCapacity is guaranteed >= ZSTD_sequenceBound(srcSize). The memory
+ *     backing outSeqs is managed by the CCtx.
+ *
+ *   - src, srcSize: an input buffer for the sequence producer to parse.
+ *     srcSize is guaranteed to be <= ZSTD_BLOCKSIZE_MAX.
+ *
+ *   - dict, dictSize: a history buffer, which may be empty, which the sequence
+ *     producer may reference as it parses the src buffer. Currently, zstd will
+ *     always pass dictSize == 0 into external sequence producers, but this will
+ *     change in the future.
+ *
+ *   - compressionLevel: a signed integer representing the zstd compression level
+ *     set by the user for the current operation. The sequence producer may choose
+ *     to use this information to change its compression strategy and speed/ratio
+ *     tradeoff. Note: the compression level does not reflect zstd parameters set
+ *     through the advanced API.
+ *
+ *   - windowSize: a size_t representing the maximum allowed offset for external
+ *     sequences. Note that sequence offsets are sometimes allowed to exceed the
+ *     windowSize if a dictionary is present, see doc/zstd_compression_format.md
+ *     for details.
+ *
+ * The user-provided function shall return a size_t representing the number of
+ * sequences written to outSeqs. This return value will be treated as an error
+ * code if it is greater than outSeqsCapacity. The return value must be non-zero
+ * if srcSize is non-zero. The ZSTD_SEQUENCE_PRODUCER_ERROR macro is provided
+ * for convenience, but any value greater than outSeqsCapacity will be treated as
+ * an error code.
+ *
+ * If the user-provided function does not return an error code, the sequences
+ * written to outSeqs must be a valid parse of the src buffer. Data corruption may
+ * occur if the parse is not valid. A parse is defined to be valid if the
+ * following conditions hold:
+ *   - The sum of matchLengths and literalLengths must equal srcSize.
+ *   - All sequences in the parse, except for the final sequence, must have
+ *     matchLength >= ZSTD_MINMATCH_MIN. The final sequence must have
+ *     matchLength >= ZSTD_MINMATCH_MIN or matchLength == 0.
+ *   - All offsets must respect the windowSize parameter as specified in
+ *     doc/zstd_compression_format.md.
+ *   - If the final sequence has matchLength == 0, it must also have offset == 0.
+ *
+ * zstd will only validate these conditions (and fail compression if they do not
+ * hold) if the ZSTD_c_validateSequences cParam is enabled. Note that sequence
+ * validation has a performance cost.
+ *
+ * If the user-provided function returns an error, zstd will either fall back
+ * to an internal sequence producer or fail the compression operation. The user can
+ * choose between the two behaviors by setting the ZSTD_c_enableSeqProducerFallback
+ * cParam. Fallback compression will follow any other cParam settings, such as
+ * compression level, the same as in a normal compression operation.
+ *
+ * The user shall instruct zstd to use a particular ZSTD_sequenceProducer_F
+ * function by calling
+ *         ZSTD_registerSequenceProducer(cctx,
+ *                                       sequenceProducerState,
+ *                                       sequenceProducer)
+ * This setting will persist until the next parameter reset of the CCtx.
+ *
+ * The sequenceProducerState must be initialized by the user before calling
+ * ZSTD_registerSequenceProducer(). The user is responsible for destroying the
+ * sequenceProducerState.
+ *
+ * *** LIMITATIONS ***
+ * This API is compatible with all zstd compression APIs which respect advanced parameters.
+ * However, there are three limitations:
+ *
+ * First, the ZSTD_c_enableLongDistanceMatching cParam is not currently supported.
+ * COMPRESSION WILL FAIL if it is enabled and the user tries to compress with a block-level
+ * external sequence producer.
+ *   - Note that ZSTD_c_enableLongDistanceMatching is auto-enabled by default in some
+ *     cases (see its documentation for details). Users must explicitly set
+ *     ZSTD_c_enableLongDistanceMatching to ZSTD_ps_disable in such cases if an external
+ *     sequence producer is registered.
+ *   - As of this writing, ZSTD_c_enableLongDistanceMatching is disabled by default
+ *     whenever ZSTD_c_windowLog < 128MB, but that's subject to change. Users should
+ *     check the docs on ZSTD_c_enableLongDistanceMatching whenever the Block-Level Sequence
+ *     Producer API is used in conjunction with advanced settings (like ZSTD_c_windowLog).
+ *
+ * Second, history buffers are not currently supported. Concretely, zstd will always pass
+ * dictSize == 0 to the external sequence producer (for now). This has two implications:
+ *   - Dictionaries are not currently supported. Compression will *not* fail if the user
+ *     references a dictionary, but the dictionary won't have any effect.
+ *   - Stream history is not currently supported. All advanced compression APIs, including
+ *     streaming APIs, work with external sequence producers, but each block is treated as
+ *     an independent chunk without history from previous blocks.
+ *
+ * Third, multi-threading within a single compression is not currently supported. In other words,
+ * COMPRESSION WILL FAIL if ZSTD_c_nbWorkers > 0 and an external sequence producer is registered.
+ * Multi-threading across compressions is fine: simply create one CCtx per thread.
+ *
+ * Long-term, we plan to overcome all three limitations. There is no technical blocker to
+ * overcoming them. It is purely a question of engineering effort.
+ */
+
+#define ZSTD_SEQUENCE_PRODUCER_ERROR ((size_t)(-1))
+
+typedef size_t (*ZSTD_sequenceProducer_F) (
+  void* sequenceProducerState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+);
+
+/*! ZSTD_registerSequenceProducer() :
+ * Instruct zstd to use a block-level external sequence producer function.
+ *
+ * The sequenceProducerState must be initialized by the caller, and the caller is
+ * responsible for managing its lifetime. This parameter is sticky across
+ * compressions. It will remain set until the user explicitly resets compression
+ * parameters.
+ *
+ * Sequence producer registration is considered to be an "advanced parameter",
+ * part of the "advanced API". This means it will only have an effect on compression
+ * APIs which respect advanced parameters, such as compress2() and compressStream2().
+ * Older compression APIs such as compressCCtx(), which predate the introduction of
+ * "advanced parameters", will ignore any external sequence producer setting.
+ *
+ * The sequence producer can be "cleared" by registering a NULL function pointer. This
+ * removes all limitations described above in the "LIMITATIONS" section of the API docs.
+ *
+ * The user is strongly encouraged to read the full API documentation (above) before
+ * calling this function. */
+ZSTDLIB_STATIC_API void
+ZSTD_registerSequenceProducer(
+  ZSTD_CCtx* cctx,
+  void* sequenceProducerState,
+  ZSTD_sequenceProducer_F sequenceProducer
+);
+
+/*! ZSTD_CCtxParams_registerSequenceProducer() :
+ * Same as ZSTD_registerSequenceProducer(), but operates on ZSTD_CCtx_params.
+ * This is used for accurate size estimation with ZSTD_estimateCCtxSize_usingCCtxParams(),
+ * which is needed when creating a ZSTD_CCtx with ZSTD_initStaticCCtx().
+ *
+ * If you are using the external sequence producer API in a scenario where ZSTD_initStaticCCtx()
+ * is required, then this function is for you. Otherwise, you probably don't need it.
+ *
+ * See tests/zstreamtest.c for example usage. */
+ZSTDLIB_STATIC_API void
+ZSTD_CCtxParams_registerSequenceProducer(
+  ZSTD_CCtx_params* params,
+  void* sequenceProducerState,
+  ZSTD_sequenceProducer_F sequenceProducer
+);
+
+
 /* *******************************************************************
-*  Buffer-less and synchronous inner streaming functions
+*  Buffer-less and synchronous inner streaming functions (DEPRECATED)
 *
-*  This is an advanced API, giving full control over buffer management, for users which need direct control over memory.
-*  But it's also a complex one, with several restrictions, documented below.
-*  Prefer normal streaming API for an easier experience.
+*  This API is deprecated, and will be removed in a future version.
+*  It allows streaming (de)compression with user allocated buffers.
+*  However, it is hard to use, and not as well tested as the rest of
+*  our API.
+*
+*  Please use the normal streaming API instead: ZSTD_compressStream2,
+*  and ZSTD_decompressStream.
+*  If there is functionality that you need, but it doesn't provide,
+*  please open an issue on our GitHub.
 ********************************************************************* */
 
 /*
···
 
   A ZSTD_CCtx object is required to track streaming operations.
   Use ZSTD_createCCtx() / ZSTD_freeCCtx() to manage resource.
-  ZSTD_CCtx object can be re-used multiple times within successive compression operations.
+  ZSTD_CCtx object can be reused multiple times within successive compression operations.
 
   Start by initializing a context.
   Use ZSTD_compressBegin(), or ZSTD_compressBegin_usingDict() for dictionary compression.
-  It's also possible to duplicate a reference context which has already been initialized, using ZSTD_copyCCtx()
 
   Then, consume your input using ZSTD_compressContinue().
   There are some important considerations to keep in mind when using this advanced function :
···
   It's possible to use srcSize==0, in which case, it will write a final empty block to end the frame.
   Without last block mark, frames are considered unfinished (hence corrupted) by compliant decoders.
 
-  `ZSTD_CCtx` object can be re-used (ZSTD_compressBegin()) to compress again.
+  `ZSTD_CCtx` object can be reused (ZSTD_compressBegin()) to compress again.
 */
 
 /*=====   Buffer-less streaming compression functions  =====*/
+ZSTD_DEPRECATED("The buffer-less API is deprecated in favor of the normal streaming API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel);
+ZSTD_DEPRECATED("The buffer-less API is deprecated in favor of the normal streaming API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
+ZSTD_DEPRECATED("The buffer-less API is deprecated in favor of the normal streaming API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict); /*< note: fails if cdict==NULL */
-ZSTDLIB_STATIC_API size_t ZSTD_copyCCtx(ZSTD_CCtx* cctx, const ZSTD_CCtx* preparedCCtx, unsigned long long pledgedSrcSize); /*<  note: if pledgedSrcSize is not known, use ZSTD_CONTENTSIZE_UNKNOWN */
 
+ZSTD_DEPRECATED("This function will likely be removed in a future release. It is misleading and has very limited utility.")
+ZSTDLIB_STATIC_API
+size_t ZSTD_copyCCtx(ZSTD_CCtx* cctx, const ZSTD_CCtx* preparedCCtx, unsigned long long pledgedSrcSize); /*<  note: if pledgedSrcSize is not known, use ZSTD_CONTENTSIZE_UNKNOWN */
+
+ZSTD_DEPRECATED("The buffer-less API is deprecated in favor of the normal streaming API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_compressContinue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTD_DEPRECATED("The buffer-less API is deprecated in favor of the normal streaming API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_compressEnd(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
 /* The ZSTD_compressBegin_advanced() and ZSTD_compressBegin_usingCDict_advanced() are now DEPRECATED and will generate a compiler warning */
 ZSTD_DEPRECATED("use advanced API to access custom parameters")
+ZSTDLIB_STATIC_API
 size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_parameters params, unsigned long long pledgedSrcSize); /*< pledgedSrcSize : If srcSize is not known at init time, use ZSTD_CONTENTSIZE_UNKNOWN */
 ZSTD_DEPRECATED("use advanced API to access custom parameters")
+ZSTDLIB_STATIC_API
 size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize);   /* compression parameters are already set within cdict. pledgedSrcSize must be correct. If srcSize is not known, use macro ZSTD_CONTENTSIZE_UNKNOWN */
 /*
   Buffer-less streaming decompression (synchronous mode)
 
   A ZSTD_DCtx object is required to track streaming operations.
   Use ZSTD_createDCtx() / ZSTD_freeDCtx() to manage it.
-  A ZSTD_DCtx object can be re-used multiple times.
+  A ZSTD_DCtx object can be reused multiple times.
 
   First typical operation is to retrieve frame parameters, using ZSTD_getFrameHeader().
   Frame header is extracted from the beginning of compressed frame, so providing only the frame's beginning is enough.
   Data fragment must be large enough to ensure successful decoding.
  `ZSTD_frameHeaderSize_max` bytes is guaranteed to always be large enough.
-  @result : 0 : successful decoding, the `ZSTD_frameHeader` structure is correctly filled.
-           >0 : `srcSize` is too small, please provide at least @result bytes on next attempt.
+  result  : 0 : successful decoding, the `ZSTD_frameHeader` structure is correctly filled.
+           >0 : `srcSize` is too small, please provide at least result bytes on next attempt.
            errorCode, which can be tested using ZSTD_isError().
 
-  It fills a ZSTD_frameHeader structure with important information to correctly decode the frame,
+  It fills a ZSTD_FrameHeader structure with important information to correctly decode the frame,
   such as the dictionary ID, content size, or maximum back-reference distance (`windowSize`).
   Note that these values could be wrong, either because of data corruption, or because a 3rd party deliberately spoofs false information.
   As a consequence, check that values remain within valid application range.
···
 
   The most memory efficient way is to use a round buffer of sufficient size.
   Sufficient size is determined by invoking ZSTD_decodingBufferSize_min(),
-  which can @return an error code if required value is too large for current system (in 32-bits mode).
+  which can return an error code if required value is too large for current system (in 32-bits mode).
   In a round buffer methodology, ZSTD_decompressContinue() decompresses each block next to previous one,
   up to the moment there is not enough room left in the buffer to guarantee decoding another full block,
   which maximum size is provided in `ZSTD_frameHeader` structure, field `blockSizeMax`.
···
   ZSTD_nextSrcSizeToDecompress() tells how many bytes to provide as 'srcSize' to ZSTD_decompressContinue().
   ZSTD_decompressContinue() requires this _exact_ amount of bytes, or it will fail.
 
- @result of ZSTD_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity).
+  result of ZSTD_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity).
   It can be zero : it just means ZSTD_decompressContinue() has decoded some metadata item.
   It can also be an error code, which can be tested with ZSTD_isError().
 
···
 */
 
 /*=====   Buffer-less streaming decompression functions  =====*/
-typedef enum { ZSTD_frame, ZSTD_skippableFrame } ZSTD_frameType_e;
-typedef struct {
-    unsigned long long frameContentSize; /* if == ZSTD_CONTENTSIZE_UNKNOWN, it means this field is not available. 0 means "empty" */
-    unsigned long long windowSize;       /* can be very large, up to <= frameContentSize */
-    unsigned blockSizeMax;
-    ZSTD_frameType_e frameType;          /* if == ZSTD_skippableFrame, frameContentSize is the size of skippable content */
-    unsigned headerSize;
-    unsigned dictID;
-    unsigned checksumFlag;
-} ZSTD_frameHeader;
 
-/*! ZSTD_getFrameHeader() :
- *  decode Frame Header, or requires larger `srcSize`.
- * @return : 0, `zfhPtr` is correctly filled,
- *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
- *           or an error code, which can be tested using ZSTD_isError() */
-ZSTDLIB_STATIC_API size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize);   /*< doesn't consume input */
-/*! ZSTD_getFrameHeader_advanced() :
- *  same as ZSTD_getFrameHeader(),
- *  with added capability to select a format (like ZSTD_f_zstd1_magicless) */
-ZSTDLIB_STATIC_API size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format);
 ZSTDLIB_STATIC_API size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize);  /*< when frame content size is not known, pass in frameContentSize == ZSTD_CONTENTSIZE_UNKNOWN */
 
 ZSTDLIB_STATIC_API size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx);
···
 ZSTDLIB_STATIC_API size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
 /* misc */
+ZSTD_DEPRECATED("This function will likely be removed in the next minor release. It is misleading and has very limited utility.")
 ZSTDLIB_STATIC_API void   ZSTD_copyDCtx(ZSTD_DCtx* dctx, const ZSTD_DCtx* preparedDCtx);
 typedef enum { ZSTDnit_frameHeader, ZSTDnit_blockHeader, ZSTDnit_block, ZSTDnit_lastBlock, ZSTDnit_checksum, ZSTDnit_skippableFrame } ZSTD_nextInputType_e;
 ZSTDLIB_STATIC_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
···
 
 
 
-/* ============================ */
-/*       Block level API       */
-/* ============================ */
+/* ========================================= */
+/*       Block level API (DEPRECATED)       */
+/* ========================================= */
 
 /*!
+
+    This API is deprecated in favor of the regular compression API.
+    You can get the frame header down to 2 bytes by setting:
+      - ZSTD_c_format = ZSTD_f_zstd1_magicless
+      - ZSTD_c_contentSizeFlag = 0
+      - ZSTD_c_checksumFlag = 0
+      - ZSTD_c_dictIDFlag = 0
+
+    This API is not as well tested as our normal API, so we recommend not using it.
+    We will be removing it in a future version. If the normal API doesn't provide
+    the functionality you need, please open a GitHub issue.
+
     Block functions produce and decode raw zstd blocks, without frame metadata.
     Frame metadata cost is typically ~12 bytes, which can be non-negligible for very small blocks (< 100 bytes).
     But users will have to take in charge needed metadata to regenerate data, such as compressed and content sizes.
···
     - It is necessary to init context before starting
       + compression : any ZSTD_compressBegin*() variant, including with dictionary
       + decompression : any ZSTD_decompressBegin*() variant, including with dictionary
-      + copyCCtx() and copyDCtx() can be used too
     - Block size is limited, it must be <= ZSTD_getBlockSize() <= ZSTD_BLOCKSIZE_MAX == 128 KB
       + If input is larger than a block size, it's necessary to split input data into multiple blocks
       + For inputs larger than a single block, consider using regular ZSTD_compress() instead.
···
 */
 
 /*=====   Raw zstd block functions  =====*/
+ZSTD_DEPRECATED("The block API is deprecated in favor of the normal compression API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_getBlockSize   (const ZSTD_CCtx* cctx);
+ZSTD_DEPRECATED("The block API is deprecated in favor of the normal compression API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_compressBlock  (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTD_DEPRECATED("The block API is deprecated in favor of the normal compression API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTD_DEPRECATED("The block API is deprecated in favor of the normal compression API. See docs.")
 ZSTDLIB_STATIC_API size_t ZSTD_insertBlock    (ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize);  /*< insert uncompressed block into `dctx` history. Useful for multi-blocks decompression. */
 
 
 #endif   /* ZSTD_H_ZSTD_STATIC_LINKING_ONLY */
-

lib/zstd/Makefile

···
 # SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 # ################################################################
-# Copyright (c) Facebook, Inc.
+# Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
 #
 # This source code is licensed under both the BSD-style license (found in the
···
 		compress/zstd_lazy.o \
 		compress/zstd_ldm.o \
 		compress/zstd_opt.o \
+		compress/zstd_preSplit.o \
 
 zstd_decompress-y := \
 		zstd_decompress_module.o \

lib/zstd/common/allocations.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* This file provides custom allocation primitives
+ */
+
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h"   /* ZSTD_malloc, ZSTD_calloc, ZSTD_free, ZSTD_memset */
+
+#include "compiler.h" /* MEM_STATIC */
+#define ZSTD_STATIC_LINKING_ONLY
+#include <linux/zstd.h> /* ZSTD_customMem */
+
+#ifndef ZSTD_ALLOCATIONS_H
+#define ZSTD_ALLOCATIONS_H
+
+/* custom memory allocation functions */
+
+MEM_STATIC void* ZSTD_customMalloc(size_t size, ZSTD_customMem customMem)
+{
+    if (customMem.customAlloc)
+        return customMem.customAlloc(customMem.opaque, size);
+    return ZSTD_malloc(size);
+}
+
+MEM_STATIC void* ZSTD_customCalloc(size_t size, ZSTD_customMem customMem)
+{
+    if (customMem.customAlloc) {
+        /* calloc implemented as malloc+memset;
+         * not as efficient as calloc, but next best guess for custom malloc */
+        void* const ptr = customMem.customAlloc(customMem.opaque, size);
+        ZSTD_memset(ptr, 0, size);
+        return ptr;
+    }
+    return ZSTD_calloc(1, size);
+}
+
+MEM_STATIC void ZSTD_customFree(void* ptr, ZSTD_customMem customMem)
+{
+    if (ptr!=NULL) {
+        if (customMem.customFree)
+            customMem.customFree(customMem.opaque, ptr);
+        else
+            ZSTD_free(ptr);
+    }
+}
+
+#endif /* ZSTD_ALLOCATIONS_H */

lib/zstd/common/bits.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_BITS_H
+#define ZSTD_BITS_H
+
+#include "mem.h"
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros32_fallback(U32 val)
+{
+    assert(val != 0);
+    {
+        static const U32 DeBruijnBytePos[32] = {0, 1, 28, 2, 29, 14, 24, 3,
+                                                30, 22, 20, 15, 25, 17, 4, 8,
+                                                31, 27, 13, 23, 21, 19, 16, 7,
+                                                26, 12, 18, 6, 11, 5, 10, 9};
+        return DeBruijnBytePos[((U32) ((val & -(S32) val) * 0x077CB531U)) >> 27];
+    }
+}
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros32(U32 val)
+{
+    assert(val != 0);
+#if (__GNUC__ >= 4)
+    return (unsigned)__builtin_ctz(val);
+#else
+    return ZSTD_countTrailingZeros32_fallback(val);
+#endif
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros32_fallback(U32 val)
+{
+    assert(val != 0);
+    {
+        static const U32 DeBruijnClz[32] = {0, 9, 1, 10, 13, 21, 2, 29,
+                                            11, 14, 16, 18, 22, 25, 3, 30,
+                                            8, 12, 20, 28, 15, 17, 24, 7,
+                                            19, 27, 23, 6, 26, 5, 4, 31};
+        val |= val >> 1;
+        val |= val >> 2;
+        val |= val >> 4;
+        val |= val >> 8;
+        val |= val >> 16;
+        return 31 - DeBruijnClz[(val * 0x07C4ACDDU) >> 27];
+    }
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros32(U32 val)
+{
+    assert(val != 0);
+#if (__GNUC__ >= 4)
+    return (unsigned)__builtin_clz(val);
+#else
+    return ZSTD_countLeadingZeros32_fallback(val);
+#endif
+}
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros64(U64 val)
+{
+    assert(val != 0);
+#if (__GNUC__ >= 4) && defined(__LP64__)
+    return (unsigned)__builtin_ctzll(val);
+#else
+    {
+        U32 mostSignificantWord = (U32)(val >> 32);
+        U32 leastSignificantWord = (U32)val;
+        if (leastSignificantWord == 0) {
+            return 32 + ZSTD_countTrailingZeros32(mostSignificantWord);
+        } else {
+            return ZSTD_countTrailingZeros32(leastSignificantWord);
+        }
+    }
+#endif
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros64(U64 val)
+{
+    assert(val != 0);
+#if (__GNUC__ >= 4)
+    return (unsigned)(__builtin_clzll(val));
+#else
+    {
+        U32 mostSignificantWord = (U32)(val >> 32);
+        U32 leastSignificantWord = (U32)val;
+        if (mostSignificantWord == 0) {
+            return 32 + ZSTD_countLeadingZeros32(leastSignificantWord);
+        } else {
+            return ZSTD_countLeadingZeros32(mostSignificantWord);
+        }
+    }
+#endif
+}
+
+MEM_STATIC unsigned ZSTD_NbCommonBytes(size_t val)
+{
+    if (MEM_isLittleEndian()) {
+        if (MEM_64bits()) {
+            return ZSTD_countTrailingZeros64((U64)val) >> 3;
+        } else {
+            return ZSTD_countTrailingZeros32((U32)val) >> 3;
+        }
+    } else {  /* Big Endian CPU */
+        if (MEM_64bits()) {
+            return ZSTD_countLeadingZeros64((U64)val) >> 3;
+        } else {
+            return ZSTD_countLeadingZeros32((U32)val) >> 3;
+        }
+    }
+}
+
+MEM_STATIC unsigned ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
+{
+    assert(val != 0);
+    return 31 - ZSTD_countLeadingZeros32(val);
+}
+
+/* ZSTD_rotateRight_*():
+ * Rotates a bitfield to the right by "count" bits.
+ * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
+ */
+MEM_STATIC
+U64 ZSTD_rotateRight_U64(U64 const value, U32 count) {
+    assert(count < 64);
+    count &= 0x3F; /* for fickle pattern recognition */
+    return (value >> count) | (U64)(value << ((0U - count) & 0x3F));
+}
+
+MEM_STATIC
+U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
+    assert(count < 32);
+    count &= 0x1F; /* for fickle pattern recognition */
+    return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
+}
+
+MEM_STATIC
+U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
+    assert(count < 16);
+    count &= 0x0F; /* for fickle pattern recognition */
+    return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
+}
+
+#endif /* ZSTD_BITS_H */

lib/zstd/common/bitstream.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /* ******************************************************************
  * bitstream
  * Part of FSE library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 #include "compiler.h"       /* UNLIKELY() */
 #include "debug.h"          /* assert(), DEBUGLOG(), RAWLOG() */
 #include "error_private.h"  /* error codes and messages */
-
+#include "bits.h"           /* ZSTD_highbit32 */
 
 /*=========================================
 *  Target specific
···
 /*-******************************************
 *  bitStream encoding API (write forward)
 ********************************************/
+typedef size_t BitContainerType;
 /* bitStream can mix input from multiple sources.
  * A critical property of these streams is that they encode and decode in **reverse** direction.
  * So the first bit sequence you add will be the last to be read, like a LIFO stack.
  */
 typedef struct {
-    size_t bitContainer;
+    BitContainerType bitContainer;
     unsigned bitPos;
     char*  startPtr;
     char*  ptr;
···
 } BIT_CStream_t;
 
 MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC, void* dstBuffer, size_t dstCapacity);
-MEM_STATIC void   BIT_addBits(BIT_CStream_t* bitC, size_t value, unsigned nbBits);
+MEM_STATIC void   BIT_addBits(BIT_CStream_t* bitC, BitContainerType value, unsigned nbBits);
 MEM_STATIC void   BIT_flushBits(BIT_CStream_t* bitC);
 MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC);
 
···
 *  `dstCapacity` must be >= sizeof(bitD->bitContainer), otherwise @return will be an error code.
 *
 *  bits are first added to a local register.
-*  Local register is size_t, hence 64-bits on 64-bits systems, or 32-bits on 32-bits systems.
+*  Local register is BitContainerType, 64-bits on 64-bits systems, or 32-bits on 32-bits systems.
 *  Writing data into memory is an explicit operation, performed by the flushBits function.
 *  Hence keep track how many bits are potentially stored into local register to avoid register overflow.
 *  After a flushBits, a maximum of 7 bits might still be stored into local register.
···
 *  bitStream decoding API (read backward)
 **********************************************/
 typedef struct {
-    size_t   bitContainer;
+    BitContainerType bitContainer;
     unsigned bitsConsumed;
     const char* ptr;
     const char* start;
     const char* limitPtr;
 } BIT_DStream_t;
 
-typedef enum { BIT_DStream_unfinished = 0,
-               BIT_DStream_endOfBuffer = 1,
-               BIT_DStream_completed = 2,
-               BIT_DStream_overflow = 3 } BIT_DStream_status;  /* result of BIT_reloadDStream() */
-               /* 1,2,4,8 would be better for bitmap combinations, but slows down performance a bit ... :( */
+typedef enum { BIT_DStream_unfinished = 0,  /* fully refilled */
+               BIT_DStream_endOfBuffer = 1, /* still some bits left in bitstream */
+               BIT_DStream_completed = 2,   /* bitstream entirely consumed, bit-exact */
+               BIT_DStream_overflow = 3     /* user requested more bits than present in bitstream */
+    } BIT_DStream_status;  /* result of BIT_reloadDStream() */
 
 MEM_STATIC size_t   BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
-MEM_STATIC size_t   BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BitContainerType BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
 MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD);
 MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* bitD);
 
 
 /* Start by invoking BIT_initDStream().
 *  A chunk of the bitStream is then stored into a local register.
-*  Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t).
+*  Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (BitContainerType).
 *  You can then retrieve bitFields stored into the local register, **in reverse order**.
 *  Local register is explicitly reloaded from memory by the BIT_reloadDStream() method.
 *  A reload guarantee a minimum of ((8*sizeof(bitD->bitContainer))-7) bits when its result is BIT_DStream_unfinished.
···
 /*-****************************************
 *  unsafe API
 ******************************************/
-MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC, size_t value, unsigned nbBits);
+MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC, BitContainerType value, unsigned nbBits);
 /* faster, but works only if value is "clean", meaning all high bits above nbBits are 0 */
 
 MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC);
···
 
 MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
 /* faster, but works only if nbBits >= 1 */
-
-
-
-/*-**************************************************************
-*  Internal functions
-****************************************************************/
-MEM_STATIC unsigned BIT_highbit32 (U32 val)
-{
-    assert(val != 0);
-    {
-#   if (__GNUC__ >= 3)   /* Use GCC Intrinsic */
-        return __builtin_clz (val) ^ 31;
-#   else   /* Software version */
-        static const unsigned DeBruijnClz[32] = { 0,  9,  1, 10, 13, 21,  2, 29,
-                                                 11, 14, 16, 18, 22, 25,  3, 30,
-                                                  8, 12, 20, 28, 15, 17, 24,  7,
-                                                 19, 27, 23,  6, 26,  5,  4, 31 };
-        U32 v = val;
-        v |= v >> 1;
-        v |= v >> 2;
-        v |= v >> 4;
-        v |= v >> 8;
-        v |= v >> 16;
-        return DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
-#   endif
-    }
-}
 
 /*=====    Local Constants   =====*/
 static const unsigned BIT_mask[] = {
···
     return 0;
 }
 
+FORCE_INLINE_TEMPLATE BitContainerType BIT_getLowerBits(BitContainerType bitContainer, U32 const nbBits)
+{
+    assert(nbBits < BIT_MASK_SIZE);
+    return bitContainer & BIT_mask[nbBits];
+}
+
 /*! BIT_addBits() :
  *  can add up to 31 bits into `bitC`.
  *  Note : does not check for register overflow ! */
 MEM_STATIC void BIT_addBits(BIT_CStream_t* bitC,
-                            size_t value, unsigned nbBits)
+                            BitContainerType value, unsigned nbBits)
 {
     DEBUG_STATIC_ASSERT(BIT_MASK_SIZE == 32);
     assert(nbBits < BIT_MASK_SIZE);
     assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
-    bitC->bitContainer |= (value & BIT_mask[nbBits]) << bitC->bitPos;
+    bitC->bitContainer |= BIT_getLowerBits(value, nbBits) << bitC->bitPos;
     bitC->bitPos += nbBits;
 }
 
···
  *  works only if `value` is _clean_,
  *  meaning all high bits above nbBits are 0 */
 MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC,
-                                size_t value, unsigned nbBits)
+                                BitContainerType value, unsigned nbBits)
 {
     assert((value>>nbBits) == 0);
     assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
···
     BIT_addBitsFast(bitC, 1, 1);   /* endMark */
     BIT_flushBits(bitC);
     if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
-    return (bitC->ptr - bitC->startPtr) + (bitC->bitPos > 0);
+    return (size_t)(bitC->ptr - bitC->startPtr) + (bitC->bitPos > 0);
 }
 
 
···
         bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(bitD->bitContainer);
         bitD->bitContainer = MEM_readLEST(bitD->ptr);
         { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
-          bitD->bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;  /* ensures bitsConsumed is always set */
+          bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;  /* ensures bitsConsumed is always set */
           if (lastByte == 0) return ERROR(GENERIC); /* endMark not present */ }
     } else {
         bitD->ptr   = bitD->start;
         bitD->bitContainer = *(const BYTE*)(bitD->start);
         switch(srcSize)
         {
-        case 7: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
+        case 7: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
                 ZSTD_FALLTHROUGH;
 
-        case 6: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
+        case 6: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
                 ZSTD_FALLTHROUGH;
 
-        case 5: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
+        case 5: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
                 ZSTD_FALLTHROUGH;
 
-        case 4: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[3]) << 24;
+        case 4: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[3]) << 24;
                 ZSTD_FALLTHROUGH;
 
-        case 3: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[2]) << 16;
+        case 3: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[2]) << 16;
                 ZSTD_FALLTHROUGH;
 
-        case 2: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[1]) <<  8;
+        case 2: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[1]) <<  8;
                 ZSTD_FALLTHROUGH;
 
         default: break;
         }
         {   BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
-            bitD->bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;
+            bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
             if (lastByte == 0) return ERROR(corruption_detected);  /* endMark not present */
         }
         bitD->bitsConsumed += (U32)(sizeof(bitD->bitContainer) - srcSize)*8;
···
     return srcSize;
 }
 
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getUpperBits(size_t bitContainer, U32 const start)
+FORCE_INLINE_TEMPLATE BitContainerType BIT_getUpperBits(BitContainerType bitContainer, U32 const start)
 {
     return bitContainer >> start;
 }
 
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getMiddleBits(size_t bitContainer, U32 const start, U32 const nbBits)
+FORCE_INLINE_TEMPLATE BitContainerType BIT_getMiddleBits(BitContainerType bitContainer, U32 const start, U32 const nbBits)
 {
     U32 const regMask = sizeof(bitContainer)*8 - 1;
     /* if start > regMask, bitstream is corrupted, and result is undefined */
···
      * such cpus old (pre-Haswell, 2013) and their performance is not of that
      * importance.
      */
-#if defined(__x86_64__) || defined(_M_X86)
+#if defined(__x86_64__) || defined(_M_X64)
     return (bitContainer >> (start & regMask)) & ((((U64)1) << nbBits) - 1);
 #else
     return (bitContainer >> (start & regMask)) & BIT_mask[nbBits];
 #endif
-}
-
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
-{
-    assert(nbBits < BIT_MASK_SIZE);
-    return bitContainer & BIT_mask[nbBits];
 }
 
 /*! BIT_lookBits() :
···
  *  On 32-bits, maxNbBits==24.
  *  On 64-bits, maxNbBits==56.
  * @return : value extracted */
-MEM_STATIC  FORCE_INLINE_ATTR size_t BIT_lookBits(const BIT_DStream_t*  bitD, U32 nbBits)
+FORCE_INLINE_TEMPLATE BitContainerType BIT_lookBits(const BIT_DStream_t*  bitD, U32 nbBits)
 {
     /* arbitrate between double-shift and shift+mask */
 #if 1
···
 
 /*! BIT_lookBitsFast() :
  *  unsafe version; only works if nbBits >= 1 */
-MEM_STATIC size_t BIT_lookBitsFast(const BIT_DStream_t* bitD, U32 nbBits)
+MEM_STATIC BitContainerType BIT_lookBitsFast(const BIT_DStream_t* bitD, U32 nbBits)
 {
     U32 const regMask = sizeof(bitD->bitContainer)*8 - 1;
     assert(nbBits >= 1);
     return (bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> (((regMask+1)-nbBits) & regMask);
 }
 
-MEM_STATIC FORCE_INLINE_ATTR void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+FORCE_INLINE_TEMPLATE void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
 {
     bitD->bitsConsumed += nbBits;
 }
···
  *  Read (consume) next n bits from local register and update.
  *  Pay attention to not read more than nbBits contained into local register.
  * @return : extracted value. */
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
+FORCE_INLINE_TEMPLATE BitContainerType BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
 {
-    size_t const value = BIT_lookBits(bitD, nbBits);
+    BitContainerType const value = BIT_lookBits(bitD, nbBits);
     BIT_skipBits(bitD, nbBits);
     return value;
 }
 
 /*! BIT_readBitsFast() :
- *  unsafe version; only works only if nbBits >= 1 */
-MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits)
+ *  unsafe version; only works if nbBits >= 1 */
+MEM_STATIC BitContainerType BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits)
 {
-    size_t const value = BIT_lookBitsFast(bitD, nbBits);
+    BitContainerType const value = BIT_lookBitsFast(bitD, nbBits);
     assert(nbBits >= 1);
     BIT_skipBits(bitD, nbBits);
     return value;
+}
+
+/*! BIT_reloadDStream_internal() :
+ *  Simple variant of BIT_reloadDStream(), with two conditions:
+ *  1. bitstream is valid : bitsConsumed <= sizeof(bitD->bitContainer)*8
+ *  2. look window is valid after shifted down : bitD->ptr >= bitD->start
+ */
+MEM_STATIC BIT_DStream_status BIT_reloadDStream_internal(BIT_DStream_t* bitD)
+{
+    assert(bitD->bitsConsumed <= sizeof(bitD->bitContainer)*8);
+    bitD->ptr -= bitD->bitsConsumed >> 3;
+    assert(bitD->ptr >= bitD->start);
+    bitD->bitsConsumed &= 7;
+    bitD->bitContainer = MEM_readLEST(bitD->ptr);
+    return BIT_DStream_unfinished;
 }
 
 /*! BIT_reloadDStreamFast() :
···
 {
     if (UNLIKELY(bitD->ptr < bitD->limitPtr))
         return BIT_DStream_overflow;
-    assert(bitD->bitsConsumed <= sizeof(bitD->bitContainer)*8);
-    bitD->ptr -= bitD->bitsConsumed >> 3;
-    bitD->bitsConsumed &= 7;
-    bitD->bitContainer = MEM_readLEST(bitD->ptr);
-    return BIT_DStream_unfinished;
+    return BIT_reloadDStream_internal(bitD);
 }
 
 /*! BIT_reloadDStream() :
  *  Refill `bitD` from buffer previously set in BIT_initDStream() .
- *  This function is safe, it guarantees it will not read beyond src buffer.
+ *  This function is safe, it guarantees it will not never beyond src buffer.
  * @return : status of `BIT_DStream_t` internal register.
  *           when status == BIT_DStream_unfinished, internal register is filled with at least 25 or 57 bits */
-MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
+FORCE_INLINE_TEMPLATE BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
 {
-    if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* overflow detected, like end of stream */
+    /* note : once in overflow mode, a bitstream remains in this mode until it's reset */
+    if (UNLIKELY(bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))) {
+        static const BitContainerType zeroFilled = 0;
+        bitD->ptr = (const char*)&zeroFilled; /* aliasing is allowed for char */
+        /* overflow detected, erroneous scenario or end of stream: no update */
         return BIT_DStream_overflow;
+    }
+
+    assert(bitD->ptr >= bitD->start);
 
     if (bitD->ptr >= bitD->limitPtr) {
-        return BIT_reloadDStreamFast(bitD);
+        return BIT_reloadDStream_internal(bitD);
     }
     if (bitD->ptr == bitD->start) {
+        /* reached end of bitStream => no update */
         if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BIT_DStream_endOfBuffer;
         return BIT_DStream_completed;
     }
-    /* start < ptr < limitPtr */
+    /* start < ptr < limitPtr => cautious update */
     {   U32 nbBytes = bitD->bitsConsumed >> 3;
         BIT_DStream_status result = BIT_DStream_unfinished;
         if (bitD->ptr - nbBytes < bitD->start) {
···
 {
     return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
 }
-
 
 #endif /* BITSTREAM_H_MODULE */

lib/zstd/common/compiler.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 
 #ifndef ZSTD_COMPILER_H
 #define ZSTD_COMPILER_H
+
+#include <linux/types.h>
 
 #include "portability_macros.h"
 
···
 */
 #define WIN_CDECL
 
+/* UNUSED_ATTR tells the compiler it is okay if the function is unused. */
+#define UNUSED_ATTR __attribute__((unused))
+
 /*
  * FORCE_INLINE_TEMPLATE is used to define C "templates", which take constant
  * parameters. They must be inlined for the compiler to eliminate the constant
  * branches.
  */
-#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR
+#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR UNUSED_ATTR
 /*
  * HINT_INLINE is used to help the compiler generate better code. It is *not*
  * used for "templates", so it can be tweaked based on the compilers
···
 #if !defined(__clang__) && defined(__GNUC__) && __GNUC__ >= 4 && __GNUC_MINOR__ >= 8 && __GNUC__ < 5
 #  define HINT_INLINE static INLINE_KEYWORD
 #else
-#  define HINT_INLINE static INLINE_KEYWORD FORCE_INLINE_ATTR
+#  define HINT_INLINE FORCE_INLINE_TEMPLATE
 #endif
 
-/* UNUSED_ATTR tells the compiler it is okay if the function is unused. */
-#define UNUSED_ATTR __attribute__((unused))
+/* "soft" inline :
+ * The compiler is free to select if it's a good idea to inline or not.
+ * The main objective is to silence compiler warnings
+ * when a defined function in included but not used.
+ *
+ * Note : this macro is prefixed `MEM_` because it used to be provided by `mem.h` unit.
+ * Updating the prefix is probably preferable, but requires a fairly large codemod,
+ * since this name is used everywhere.
+ */
+#ifndef MEM_STATIC  /* already defined in Linux Kernel mem.h */
+#define MEM_STATIC static __inline UNUSED_ATTR
+#endif
 
 /* force no inlining */
 #define FORCE_NOINLINE static __attribute__((__noinline__))
···
 #  define PREFETCH_L1(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
 #  define PREFETCH_L2(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 2 /* locality */)
 #elif defined(__aarch64__)
-#  define PREFETCH_L1(ptr)  __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr)))
-#  define PREFETCH_L2(ptr)  __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr)))
+#  define PREFETCH_L1(ptr)  do { __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr))); } while (0)
+#  define PREFETCH_L2(ptr)  do { __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr))); } while (0)
 #else
-#  define PREFETCH_L1(ptr) (void)(ptr)  /* disabled */
-#  define PREFETCH_L2(ptr) (void)(ptr)  /* disabled */
+#  define PREFETCH_L1(ptr) do { (void)(ptr); } while (0)  /* disabled */
+#  define PREFETCH_L2(ptr) do { (void)(ptr); } while (0)  /* disabled */
 #endif  /* NO_PREFETCH */
 
 #define CACHELINE_SIZE 64
 
-#define PREFETCH_AREA(p, s)  {            \
-    const char* const _ptr = (const char*)(p);  \
-    size_t const _size = (size_t)(s);     \
-    size_t _pos;                          \
-    for (_pos=0; _pos<_size; _pos+=CACHELINE_SIZE) {  \
-        PREFETCH_L2(_ptr + _pos);         \
-    }                                     \
-}
+#define PREFETCH_AREA(p, s)                              \
+    do {                                                 \
+        const char* const _ptr = (const char*)(p);       \
+        size_t const _size = (size_t)(s);                \
+        size_t _pos;                                     \
+        for (_pos=0; _pos<_size; _pos+=CACHELINE_SIZE) { \
+            PREFETCH_L2(_ptr + _pos);                    \
+        }                                                \
+    } while (0)
 
 /* vectorization
  * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax,
···
 #define UNLIKELY(x) (__builtin_expect((x), 0))
 
 #if __has_builtin(__builtin_unreachable) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 5)))
-#  define ZSTD_UNREACHABLE { assert(0), __builtin_unreachable(); }
+#  define ZSTD_UNREACHABLE do { assert(0), __builtin_unreachable(); } while (0)
 #else
-#  define ZSTD_UNREACHABLE { assert(0); }
+#  define ZSTD_UNREACHABLE do { assert(0); } while (0)
 #endif
 
 /* disable warnings */
-
-/*Like DYNAMIC_BMI2 but for compile time determination of BMI2 support*/
-
 
 /* compile time determination of SIMD support */
 
···
 #define ZSTD_FALLTHROUGH fallthrough
 
 /*-**************************************************************
-*  Alignment check
+*  Alignment
 *****************************************************************/
+
+/* @return 1 if @u is a 2^n value, 0 otherwise
+ * useful to check a value is valid for alignment restrictions */
+MEM_STATIC int ZSTD_isPower2(size_t u) {
+    return (u & (u-1)) == 0;
+}
 
 /* this test was initially positioned in mem.h,
  * but this file is removed (or replaced) for linux kernel
···
 
 #endif /* ZSTD_ALIGNOF */
 
+#ifndef ZSTD_ALIGNED
+/* C90-compatible alignment macro (GCC/Clang). Adjust for other compilers if needed. */
+#define ZSTD_ALIGNED(a) __attribute__((aligned(a)))
+#endif /* ZSTD_ALIGNED */
+
+
 /*-**************************************************************
 *  Sanitizer
 *****************************************************************/
+
+/*
+ * Zstd relies on pointer overflow in its decompressor.
+ * We add this attribute to functions that rely on pointer overflow.
+ */
+#ifndef ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+#  if __has_attribute(no_sanitize)
+#    if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 8
+       /* gcc < 8 only has signed-integer-overlow which triggers on pointer overflow */
+#      define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR __attribute__((no_sanitize("signed-integer-overflow")))
+#    else
+       /* older versions of clang [3.7, 5.0) will warn that pointer-overflow is ignored. */
+#      define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR __attribute__((no_sanitize("pointer-overflow")))
+#    endif
+#  else
+#    define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+#  endif
+#endif
+
+/*
+ * Helper function to perform a wrapped pointer difference without triggering
+ * UBSAN.
+ *
+ * @returns lhs - rhs with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+ptrdiff_t ZSTD_wrappedPtrDiff(unsigned char const* lhs, unsigned char const* rhs)
+{
+    return lhs - rhs;
+}
+
+/*
+ * Helper function to perform a wrapped pointer add without triggering UBSAN.
+ *
+ * @return ptr + add with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+unsigned char const* ZSTD_wrappedPtrAdd(unsigned char const* ptr, ptrdiff_t add)
+{
+    return ptr + add;
+}
+
+/*
+ * Helper function to perform a wrapped pointer subtraction without triggering
+ * UBSAN.
+ *
+ * @return ptr - sub with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+unsigned char const* ZSTD_wrappedPtrSub(unsigned char const* ptr, ptrdiff_t sub)
+{
+    return ptr - sub;
+}
+
+/*
+ * Helper function to add to a pointer that works around C's undefined behavior
+ * of adding 0 to NULL.
+ *
+ * @returns `ptr + add` except it defines `NULL + 0 == NULL`.
+ */
+MEM_STATIC
+unsigned char* ZSTD_maybeNullPtrAdd(unsigned char* ptr, ptrdiff_t add)
+{
+    return add > 0 ? ptr + add : ptr;
+}
+
+/* Issue #3240 reports an ASAN failure on an llvm-mingw build. Out of an
+ * abundance of caution, disable our custom poisoning on mingw. */
+#ifdef __MINGW32__
+#ifndef ZSTD_ASAN_DONT_POISON_WORKSPACE
+#define ZSTD_ASAN_DONT_POISON_WORKSPACE 1
+#endif
+#ifndef ZSTD_MSAN_DONT_POISON_WORKSPACE
+#define ZSTD_MSAN_DONT_POISON_WORKSPACE 1
+#endif
+#endif
 
 
 

lib/zstd/common/cpu.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

lib/zstd/common/debug.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /* ******************************************************************
  * debug
  * Part of FSE library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 
 #include "debug.h"
 
+#if (DEBUGLEVEL>=2)
+/* We only use this when DEBUGLEVEL>=2, but we get -Werror=pedantic errors if a
+ * translation unit is empty. So remove this from Linux kernel builds, but
+ * otherwise just leave it in.
+ */
 int g_debuglevel = DEBUGLEVEL;
+#endif

lib/zstd/common/debug.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /* ******************************************************************
  * debug
  * Part of FSE library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 
 #ifndef DEBUG_H_12987983217
 #define DEBUG_H_12987983217
-
 
 
 /* static assert is triggered at compile time, leaving no runtime artefact.
···
                             It's useful when enabling very verbose levels
                             on selective conditions (such as position in src) */
 
-#  define RAWLOG(l, ...) {                                       \
-                if (l<=g_debuglevel) {                           \
-                    ZSTD_DEBUG_PRINT(__VA_ARGS__);               \
-            }   }
-#  define DEBUGLOG(l, ...) {                                     \
-                if (l<=g_debuglevel) {                           \
-                    ZSTD_DEBUG_PRINT(__FILE__ ": " __VA_ARGS__); \
-                    ZSTD_DEBUG_PRINT(" \n");                     \
-            }   }
+#  define RAWLOG(l, ...)                   \
+    do {                                   \
+        if (l<=g_debuglevel) {             \
+            ZSTD_DEBUG_PRINT(__VA_ARGS__); \
+        }                                  \
+    } while (0)
+
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define LINE_AS_STRING TOSTRING(__LINE__)
+
+#  define DEBUGLOG(l, ...)                               \
+    do {                                                 \
+        if (l<=g_debuglevel) {                           \
+            ZSTD_DEBUG_PRINT(__FILE__ ":" LINE_AS_STRING ": " __VA_ARGS__); \
+            ZSTD_DEBUG_PRINT(" \n");                     \
+        }                                                \
+    } while (0)
 #else
-#  define RAWLOG(l, ...)      {}    /* disabled */
-#  define DEBUGLOG(l, ...)    {}    /* disabled */
+#  define RAWLOG(l, ...)   do { } while (0)    /* disabled */
+#  define DEBUGLOG(l, ...) do { } while (0)    /* disabled */
 #endif
-
-
 
 #endif /* DEBUG_H_12987983217 */

lib/zstd/common/entropy_common.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /* ******************************************************************
  * Common functions of New Generation Entropy library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 #include "error_private.h"       /* ERR_*, ERROR */
 #define FSE_STATIC_LINKING_ONLY  /* FSE_MIN_TABLELOG */
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY  /* HUF_TABLELOG_ABSOLUTEMAX */
 #include "huf.h"
+#include "bits.h"                /* ZSDT_highbit32, ZSTD_countTrailingZeros32 */
 
 
 /*===   Version   ===*/
···
 /*-**************************************************************
 *  FSE NCount encoding-decoding
 ****************************************************************/
-static U32 FSE_ctz(U32 val)
-{
-    assert(val != 0);
-    {
-#   if (__GNUC__ >= 3)   /* GCC Intrinsic */
-        return __builtin_ctz(val);
-#   else   /* Software version */
-        U32 count = 0;
-        while ((val & 1) == 0) {
-            val >>= 1;
-            ++count;
-        }
-        return count;
-#   endif
-    }
-}
-
 FORCE_INLINE_TEMPLATE
 size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
                            const void* headerBuffer, size_t hbSize)
···
              * repeat.
              * Avoid UB by setting the high bit to 1.
              */
-            int repeats = FSE_ctz(~bitStream | 0x80000000) >> 1;
+            int repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
             while (repeats >= 12) {
                 charnum += 3 * 12;
                 if (LIKELY(ip <= iend-7)) {
···
                     ip = iend - 4;
                 }
                 bitStream = MEM_readLE32(ip) >> bitCount;
-                repeats = FSE_ctz(~bitStream | 0x80000000) >> 1;
+                repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
             }
             charnum += 3 * repeats;
             bitStream >>= 2 * repeats;
···
                  * know that threshold > 1.
                  */
                 if (remaining <= 1) break;
-                nbBits = BIT_highbit32(remaining) + 1;
+                nbBits = ZSTD_highbit32(remaining) + 1;
                 threshold = 1 << (nbBits - 1);
             }
             if (charnum >= maxSV1) break;
···
                      const void* src, size_t srcSize)
 {
     U32 wksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
-    return HUF_readStats_wksp(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, wksp, sizeof(wksp), /* bmi2 */ 0);
+    return HUF_readStats_wksp(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, wksp, sizeof(wksp), /* flags */ 0);
 }
 
 FORCE_INLINE_TEMPLATE size_t
···
     if (weightTotal == 0) return ERROR(corruption_detected);
 
     /* get last non-null symbol weight (implied, total must be 2^n) */
-    {   U32 const tableLog = BIT_highbit32(weightTotal) + 1;
+    {   U32 const tableLog = ZSTD_highbit32(weightTotal) + 1;
         if (tableLog > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
         *tableLogPtr = tableLog;
         /* determine last weight */
         {   U32 const total = 1 << tableLog;
             U32 const rest = total - weightTotal;
-            U32 const verif = 1 << BIT_highbit32(rest);
-            U32 const lastWeight = BIT_highbit32(rest) + 1;
+            U32 const verif = 1 << ZSTD_highbit32(rest);
+            U32 const lastWeight = ZSTD_highbit32(rest) + 1;
             if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
             huffWeight[oSize] = (BYTE)lastWeight;
             rankStats[lastWeight]++;
···
                      U32* nbSymbolsPtr, U32* tableLogPtr,
                      const void* src, size_t srcSize,
                      void* workSpace, size_t wkspSize,
-                     int bmi2)
+                     int flags)
 {
 #if DYNAMIC_BMI2
-    if (bmi2) {
+    if (flags & HUF_flags_bmi2) {
         return HUF_readStats_body_bmi2(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
     }
 #endif
-    (void)bmi2;
+    (void)flags;
     return HUF_readStats_body_default(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
 }

lib/zstd/common/error_private.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
     case PREFIX(version_unsupported): return "Version not supported";
     case PREFIX(frameParameter_unsupported): return "Unsupported frame parameter";
     case PREFIX(frameParameter_windowTooLarge): return "Frame requires too much memory for decoding";
-    case PREFIX(corruption_detected): return "Corrupted block detected";
+    case PREFIX(corruption_detected): return "Data corruption detected";
     case PREFIX(checksum_wrong): return "Restored data doesn't match checksum";
+    case PREFIX(literals_headerWrong): return "Header of Literals' block doesn't respect format specification";
     case PREFIX(parameter_unsupported): return "Unsupported parameter";
+    case PREFIX(parameter_combination_unsupported): return "Unsupported combination of parameters";
     case PREFIX(parameter_outOfBound): return "Parameter is out of bound";
     case PREFIX(init_missing): return "Context should be init first";
     case PREFIX(memory_allocation): return "Allocation error : not enough memory";
···
     case PREFIX(tableLog_tooLarge): return "tableLog requires too much memory : unsupported";
     case PREFIX(maxSymbolValue_tooLarge): return "Unsupported max Symbol Value : too large";
     case PREFIX(maxSymbolValue_tooSmall): return "Specified maxSymbolValue is too small";
+    case PREFIX(cannotProduce_uncompressedBlock): return "This mode cannot generate an uncompressed block";
+    case PREFIX(stabilityCondition_notRespected): return "pledged buffer stability condition is not respected";
     case PREFIX(dictionary_corrupted): return "Dictionary is corrupted";
     case PREFIX(dictionary_wrong): return "Dictionary mismatch";
     case PREFIX(dictionaryCreation_failed): return "Cannot create Dictionary from provided samples";
     case PREFIX(dstSize_tooSmall): return "Destination buffer is too small";
     case PREFIX(srcSize_wrong): return "Src size is incorrect";
     case PREFIX(dstBuffer_null): return "Operation on NULL destination buffer";
+    case PREFIX(noForwardProgress_destFull): return "Operation made no progress over multiple calls, due to output buffer being full";
+    case PREFIX(noForwardProgress_inputEmpty): return "Operation made no progress over multiple calls, due to input being empty";
         /* following error codes are not stable and may be removed or changed in a future version */
     case PREFIX(frameIndex_tooLarge): return "Frame index is too large";
     case PREFIX(seekableIO): return "An I/O error occurred when reading/seeking";
     case PREFIX(dstBuffer_wrong): return "Destination buffer is wrong";
     case PREFIX(srcBuffer_wrong): return "Source buffer is wrong";
+    case PREFIX(sequenceProducer_failed): return "Block-level external sequence producer returned an error code";
+    case PREFIX(externalSequences_invalid): return "External sequences are not valid";
     case PREFIX(maxCode):
     default: return notErrorCode;
     }

lib/zstd/common/error_private.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #ifndef ERROR_H_MODULE
 #define ERROR_H_MODULE
 
-
-
 /* ****************************************
 *  Dependencies
 ******************************************/
···
 #include "compiler.h"
 #include "debug.h"
 #include "zstd_deps.h"       /* size_t */
-
 
 /* ****************************************
 *  Compiler-specific
···
 ERR_STATIC ERR_enum ERR_getErrorCode(size_t code) { if (!ERR_isError(code)) return (ERR_enum)0; return (ERR_enum) (0-code); }
 
 /* check and forward error code */
-#define CHECK_V_F(e, f) size_t const e = f; if (ERR_isError(e)) return e
-#define CHECK_F(f)   { CHECK_V_F(_var_err__, f); }
+#define CHECK_V_F(e, f)     \
+    size_t const e = f;     \
+    do {                    \
+        if (ERR_isError(e)) \
+            return e;       \
+    } while (0)
+#define CHECK_F(f)   do { CHECK_V_F(_var_err__, f); } while (0)
 
 
 /*-****************************************
···
  * We want to force this function invocation to be syntactically correct, but
  * we don't want to force runtime evaluation of its arguments.
  */
-#define _FORCE_HAS_FORMAT_STRING(...) \
-  if (0) { \
-    _force_has_format_string(__VA_ARGS__); \
-  }
+#define _FORCE_HAS_FORMAT_STRING(...)              \
+    do {                                           \
+        if (0) {                                   \
+            _force_has_format_string(__VA_ARGS__); \
+        }                                          \
+    } while (0)
 
 #define ERR_QUOTE(str) #str
 
···
  * In order to do that (particularly, printing the conditional that failed),
  * this can't just wrap RETURN_ERROR().
  */
-#define RETURN_ERROR_IF(cond, err, ...) \
-  if (cond) { \
-    RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
-           __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  }
+#define RETURN_ERROR_IF(cond, err, ...)                                        \
+    do {                                                                       \
+        if (cond) {                                                            \
+            RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s",          \
+                  __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
+            _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                             \
+            RAWLOG(3, ": " __VA_ARGS__);                                       \
+            RAWLOG(3, "\n");                                                   \
+            return ERROR(err);                                                 \
+        }                                                                      \
+    } while (0)
 
 /*
  * Unconditionally return the specified error.
  *
  * In debug modes, prints additional information.
  */
-#define RETURN_ERROR(err, ...) \
-  do { \
-    RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
-           __FILE__, __LINE__, ERR_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  } while(0);
+#define RETURN_ERROR(err, ...)                                               \
+    do {                                                                     \
+        RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
+              __FILE__, __LINE__, ERR_QUOTE(ERROR(err)));                    \
+        _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                               \
+        RAWLOG(3, ": " __VA_ARGS__);                                         \
+        RAWLOG(3, "\n");                                                     \
+        return ERROR(err);                                                   \
+    } while(0)
 
 /*
  * If the provided expression evaluates to an error code, returns that error code.
  *
  * In debug modes, prints additional information.
  */
-#define FORWARD_IF_ERROR(err, ...) \
-  do { \
-    size_t const err_code = (err); \
-    if (ERR_isError(err_code)) { \
-      RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
-             __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
-      _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-      RAWLOG(3, ": " __VA_ARGS__); \
-      RAWLOG(3, "\n"); \
-      return err_code; \
-    } \
-  } while(0);
-
+#define FORWARD_IF_ERROR(err, ...)                                                 \
+    do {                                                                           \
+        size_t const err_code = (err);                                             \
+        if (ERR_isError(err_code)) {                                               \
+            RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s",                 \
+                  __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
+            _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                                 \
+            RAWLOG(3, ": " __VA_ARGS__);                                           \
+            RAWLOG(3, "\n");                                                       \
+            return err_code;                                                       \
+        }                                                                          \
+    } while(0)
 
 #endif /* ERROR_H_MODULE */

lib/zstd/common/fse.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /* ******************************************************************
  * FSE : Finite State Entropy codec
  * Public Prototypes declaration
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
  * in the COPYING file in the root directory of this source tree).
  * You may select, at your option, one of the above-listed licenses.
 ****************************************************************** */
-
-
 #ifndef FSE_H
 #define FSE_H
 
···
 *  Dependencies
 ******************************************/
 #include "zstd_deps.h"    /* size_t, ptrdiff_t */
-
 
 /*-*****************************************
 *  FSE_PUBLIC_API : control library symbols visibility
···
 FSE_PUBLIC_API unsigned FSE_versionNumber(void);   /*< library version number; to be used when checking dll version */
 
 
-/*-****************************************
-*  FSE simple functions
-******************************************/
-/*! FSE_compress() :
-    Compress content of buffer 'src', of size 'srcSize', into destination buffer 'dst'.
-    'dst' buffer must be already allocated. Compression runs faster is dstCapacity >= FSE_compressBound(srcSize).
-    @return : size of compressed data (<= dstCapacity).
-    Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
-                     if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression instead.
-                     if FSE_isError(return), compression failed (more details using FSE_getErrorName())
-*/
-FSE_PUBLIC_API size_t FSE_compress(void* dst, size_t dstCapacity,
-                             const void* src, size_t srcSize);
-
-/*! FSE_decompress():
-    Decompress FSE data from buffer 'cSrc', of size 'cSrcSize',
-    into already allocated destination buffer 'dst', of size 'dstCapacity'.
-    @return : size of regenerated data (<= maxDstSize),
-              or an error code, which can be tested using FSE_isError() .
-
-    ** Important ** : FSE_decompress() does not decompress non-compressible nor RLE data !!!
-    Why ? : making this distinction requires a header.
-    Header management is intentionally delegated to the user layer, which can better manage special cases.
-*/
-FSE_PUBLIC_API size_t FSE_decompress(void* dst,  size_t dstCapacity,
-                               const void* cSrc, size_t cSrcSize);
-
-
 /*-*****************************************
 *  Tool functions
 ******************************************/
···
 /* Error Management */
 FSE_PUBLIC_API unsigned    FSE_isError(size_t code);        /* tells if a return value is an error code */
 FSE_PUBLIC_API const char* FSE_getErrorName(size_t code);   /* provides error code string (useful for debugging) */
-
-
-/*-*****************************************
-*  FSE advanced functions
-******************************************/
-/*! FSE_compress2() :
-    Same as FSE_compress(), but allows the selection of 'maxSymbolValue' and 'tableLog'
-    Both parameters can be defined as '0' to mean : use default value
-    @return : size of compressed data
-    Special values : if return == 0, srcData is not compressible => Nothing is stored within cSrc !!!
-                     if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression.
-                     if FSE_isError(return), it's an error code.
-*/
-FSE_PUBLIC_API size_t FSE_compress2 (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
 
 
 /*-*****************************************
···
 /*! Constructor and Destructor of FSE_CTable.
     Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */
 typedef unsigned FSE_CTable;   /* don't allocate that. It's only meant to be more restrictive than void* */
-FSE_PUBLIC_API FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog);
-FSE_PUBLIC_API void        FSE_freeCTable (FSE_CTable* ct);
 
 /*! FSE_buildCTable():
     Builds `ct`, which must be already allocated, using FSE_createCTable().
···
                            unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
                            const void* rBuffer, size_t rBuffSize, int bmi2);
 
-/*! Constructor and Destructor of FSE_DTable.
-    Note that its size depends on 'tableLog' */
 typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
-FSE_PUBLIC_API FSE_DTable* FSE_createDTable(unsigned tableLog);
-FSE_PUBLIC_API void        FSE_freeDTable(FSE_DTable* dt);
-
-/*! FSE_buildDTable():
-    Builds 'dt', which must be already allocated, using FSE_createDTable().
-    return : 0, or an errorCode, which can be tested using FSE_isError() */
-FSE_PUBLIC_API size_t FSE_buildDTable (FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
-
-/*! FSE_decompress_usingDTable():
-    Decompress compressed source `cSrc` of size `cSrcSize` using `dt`
-    into `dst` which must be already allocated.
-    @return : size of regenerated data (necessarily <= `dstCapacity`),
-              or an errorCode, which can be tested using FSE_isError() */
-FSE_PUBLIC_API size_t FSE_decompress_usingDTable(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, const FSE_DTable* dt);
 
 /*!
 Tutorial :
···
 
 #endif  /* FSE_H */
 
+
 #if !defined(FSE_H_FSE_STATIC_LINKING_ONLY)
 #define FSE_H_FSE_STATIC_LINKING_ONLY
-
-/* *** Dependency *** */
 #include "bitstream.h"
-
 
 /* *****************************************
 *  Static allocation
···
 unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);
 /*< same as FSE_optimalTableLog(), which used `minus==2` */
 
-/* FSE_compress_wksp() :
- * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
- * FSE_COMPRESS_WKSP_SIZE_U32() provides the minimum size required for `workSpace` as a table of FSE_CTable.
- */
-#define FSE_COMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue)   ( FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) + ((maxTableLog > 12) ? (1 << (maxTableLog - 2)) : 1024) )
-size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
-
-size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits);
-/*< build a fake FSE_CTable, designed for a flat distribution, where each symbol uses nbBits */
-
 size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);
 /*< build a fake FSE_CTable, designed to compress always the same symbolValue */
 
···
 FSE_PUBLIC_API size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
 /*< Same as FSE_buildDTable(), using an externally allocated `workspace` produced with `FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxSymbolValue)` */
 
-size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
-/*< build a fake FSE_DTable, designed to read a flat distribution where each symbol uses nbBits */
-
-size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
-/*< build a fake FSE_DTable, designed to always generate the same symbolValue */
-
-#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) + (FSE_MAX_SYMBOL_VALUE + 1) / 2 + 1)
+#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + 1 + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) + (FSE_MAX_SYMBOL_VALUE + 1) / 2 + 1)
 #define FSE_DECOMPRESS_WKSP_SIZE(maxTableLog, maxSymbolValue) (FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(unsigned))
-size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize);
-/*< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)` */
-
 size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2);
-/*< Same as FSE_decompress_wksp() but with dynamic BMI2 support. Pass 1 if your CPU supports BMI2 or 0 if it doesn't. */
+/*< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)`.
+ * Set bmi2 to 1 if your CPU supports BMI2 or 0 if it doesn't */
 
 typedef enum {
    FSE_repeat_none,  /*< Cannot use the previous table */
···
     FSE_symbolCompressionTransform const symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
     const U16* const stateTable = (const U16*)(statePtr->stateTable);
     U32 const nbBitsOut  = (U32)((statePtr->value + symbolTT.deltaNbBits) >> 16);
-    BIT_addBits(bitC, statePtr->value, nbBitsOut);
+    BIT_addBits(bitC, (BitContainerType)statePtr->value, nbBitsOut);
     statePtr->value = stateTable[ (statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
 }
 
 MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePtr)
 {
-    BIT_addBits(bitC, statePtr->value, statePtr->stateLog);
+    BIT_addBits(bitC, (BitContainerType)statePtr->value, statePtr->stateLog);
     BIT_flushBits(bitC);
 }
 
 
 /* FSE_getMaxNbBits() :
  * Approximate maximum cost of a symbol, in bits.
- * Fractional get rounded up (i.e : a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
+ * Fractional get rounded up (i.e. a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
  * note 1 : assume symbolValue is valid (<= maxSymbolValue)
  * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
 MEM_STATIC U32 FSE_getMaxNbBits(const void* symbolTTPtr, U32 symbolValue)
···
 
 #define FSE_TABLESTEP(tableSize) (((tableSize)>>1) + ((tableSize)>>3) + 3)
 
-
 #endif /* FSE_STATIC_LINKING_ONLY */
-
-

lib/zstd/common/fse_decompress.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /* ******************************************************************
  * FSE : Finite State Entropy decoder
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
 #include "error_private.h"
-#define ZSTD_DEPS_NEED_MALLOC
-#include "zstd_deps.h"
+#include "zstd_deps.h"  /* ZSTD_memcpy */
+#include "bits.h"       /* ZSTD_highbit32 */
 
 
 /* **************************************************************
···
 #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
 #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
 
-
-/* Function templates */
-FSE_DTable* FSE_createDTable (unsigned tableLog)
-{
-    if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
-    return (FSE_DTable*)ZSTD_malloc( FSE_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
-}
-
-void FSE_freeDTable (FSE_DTable* dt)
-{
-    ZSTD_free(dt);
-}
-
 static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
 {
     void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
···
                     symbolNext[s] = 1;
                 } else {
                     if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
-                    symbolNext[s] = normalizedCounter[s];
+                    symbolNext[s] = (U16)normalizedCounter[s];
         }   }   }
         ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
     }
···
          * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
          * our buffer to handle the over-write.
          */
-        {
-            U64 const add = 0x0101010101010101ull;
+        {   U64 const add = 0x0101010101010101ull;
             size_t pos = 0;
             U64 sv = 0;
             U32 s;
···
                 for (i = 8; i < n; i += 8) {
                     MEM_write64(spread + pos + i, sv);
                 }
-                pos += n;
-            }
-        }
+                pos += (size_t)n;
+        }   }
         /* Now we spread those positions across the table.
-         * The benefit of doing it in two stages is that we avoid the the
+         * The benefit of doing it in two stages is that we avoid the
          * variable size inner loop, which caused lots of branch misses.
          * Now we can run through all the positions without any branch misses.
-         * We unroll the loop twice, since that is what emperically worked best.
+         * We unroll the loop twice, since that is what empirically worked best.
          */
         {
             size_t position = 0;
···
         for (u=0; u<tableSize; u++) {
             FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
             U32 const nextState = symbolNext[symbol]++;
-            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
+            tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
             tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
     }   }
 
···
 /*-*******************************************************
 *  Decompression (Byte symbols)
 *********************************************************/
-size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
-{
-    void* ptr = dt;
-    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
-    void* dPtr = dt + 1;
-    FSE_decode_t* const cell = (FSE_decode_t*)dPtr;
-
-    DTableH->tableLog = 0;
-    DTableH->fastMode = 0;
-
-    cell->newState = 0;
-    cell->symbol = symbolValue;
-    cell->nbBits = 0;
-
-    return 0;
-}
-
-
-size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
-{
-    void* ptr = dt;
-    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
-    void* dPtr = dt + 1;
-    FSE_decode_t* const dinfo = (FSE_decode_t*)dPtr;
-    const unsigned tableSize = 1 << nbBits;
-    const unsigned tableMask = tableSize - 1;
-    const unsigned maxSV1 = tableMask+1;
-    unsigned s;
-
-    /* Sanity checks */
-    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
-
-    /* Build Decoding Table */
-    DTableH->tableLog = (U16)nbBits;
-    DTableH->fastMode = 1;
-    for (s=0; s<maxSV1; s++) {
-        dinfo[s].newState = 0;
-        dinfo[s].symbol = (BYTE)s;
-        dinfo[s].nbBits = (BYTE)nbBits;
-    }
-
-    return 0;
-}
 
 FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
           void* dst, size_t maxDstSize,
···
 
     FSE_initDState(&state1, &bitD, dt);
     FSE_initDState(&state2, &bitD, dt);
+
+    RETURN_ERROR_IF(BIT_reloadDStream(&bitD)==BIT_DStream_overflow, corruption_detected, "");
 
 #define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
 
···
             break;
     }   }
 
-    return op-ostart;
-}
-
-
-size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
-                            const void* cSrc, size_t cSrcSize,
-                            const FSE_DTable* dt)
-{
-    const void* ptr = dt;
-    const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
-    const U32 fastMode = DTableH->fastMode;
-
-    /* select fast mode (static) */
-    if (fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
-    return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
-}
-
-
-size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
-{
-    return FSE_decompress_wksp_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, /* bmi2 */ 0);
+    assert(op >= ostart);
+    return (size_t)(op-ostart);
 }
 
 typedef struct {
     short ncount[FSE_MAX_SYMBOL_VALUE + 1];
-    FSE_DTable dtable[]; /* Dynamically sized */
 } FSE_DecompressWksp;
 
 
···
     unsigned tableLog;
     unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
     FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
+    size_t const dtablePos = sizeof(FSE_DecompressWksp) / sizeof(FSE_DTable);
+    FSE_DTable* const dtable = (FSE_DTable*)workSpace + dtablePos;
 
-    DEBUG_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
+    FSE_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
     if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);
 
+    /* correct offset to dtable depends on this property */
+    FSE_STATIC_ASSERT(sizeof(FSE_DecompressWksp) % sizeof(FSE_DTable) == 0);
+
     /* normal FSE decoding mode */
-    {
-        size_t const NCountLength = FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
+    {   size_t const NCountLength =
+            FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
         if (FSE_isError(NCountLength)) return NCountLength;
         if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
         assert(NCountLength <= cSrcSize);
···
     }
 
     if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
-    workSpace = wksp->dtable + FSE_DTABLE_SIZE_U32(tableLog);
+    assert(sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog) <= wkspSize);
+    workSpace = (BYTE*)workSpace + sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
     wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
 
-    CHECK_F( FSE_buildDTable_internal(wksp->dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
+    CHECK_F( FSE_buildDTable_internal(dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
 
     {
-        const void* ptr = wksp->dtable;
+        const void* ptr = dtable;
         const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
         const U32 fastMode = DTableH->fastMode;
 
         /* select fast mode (static) */
-        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 1);
-        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 0);
+        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 1);
+        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 0);
     }
 }
 
···
     (void)bmi2;
     return FSE_decompress_wksp_body_default(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
 }
-
-
-typedef FSE_DTable DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
-
-
 
 #endif   /* FSE_COMMONDEFS_ONLY */

lib/zstd/common/huf.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /* ******************************************************************
  * huff0 huffman codec,
  * part of Finite State Entropy library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
  * You may select, at your option, one of the above-listed licenses.
 ****************************************************************** */
 
-
 #ifndef HUF_H_298734234
 #define HUF_H_298734234
 
 /* *** Dependencies *** */
 #include "zstd_deps.h"    /* size_t */
-
-
-/* *** library symbols visibility *** */
-/* Note : when linking with -fvisibility=hidden on gcc, or by default on Visual,
- *        HUF symbols remain "private" (internal symbols for library only).
- *        Set macro FSE_DLL_EXPORT to 1 if you want HUF symbols visible on DLL interface */
-#if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)
-#  define HUF_PUBLIC_API __attribute__ ((visibility ("default")))
-#elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1)   /* Visual expected */
-#  define HUF_PUBLIC_API __declspec(dllexport)
-#elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)
-#  define HUF_PUBLIC_API __declspec(dllimport)  /* not required, just to generate faster code (saves a function pointer load from IAT and an indirect jump) */
-#else
-#  define HUF_PUBLIC_API
-#endif
-
-
-/* ========================== */
-/* ***  simple functions  *** */
-/* ========================== */
-
-/* HUF_compress() :
- *  Compress content from buffer 'src', of size 'srcSize', into buffer 'dst'.
- * 'dst' buffer must be already allocated.
- *  Compression runs faster if `dstCapacity` >= HUF_compressBound(srcSize).
- * `srcSize` must be <= `HUF_BLOCKSIZE_MAX` == 128 KB.
- * @return : size of compressed data (<= `dstCapacity`).
- *  Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
- *                   if HUF_isError(return), compression failed (more details using HUF_getErrorName())
- */
-HUF_PUBLIC_API size_t HUF_compress(void* dst, size_t dstCapacity,
-                             const void* src, size_t srcSize);
-
-/* HUF_decompress() :
- *  Decompress HUF data from buffer 'cSrc', of size 'cSrcSize',
- *  into already allocated buffer 'dst', of minimum size 'dstSize'.
- * `originalSize` : **must** be the ***exact*** size of original (uncompressed) data.
- *  Note : in contrast with FSE, HUF_decompress can regenerate
- *         RLE (cSrcSize==1) and uncompressed (cSrcSize==dstSize) data,
- *         because it knows size to regenerate (originalSize).
- * @return : size of regenerated data (== originalSize),
- *           or an error code, which can be tested using HUF_isError()
- */
-HUF_PUBLIC_API size_t HUF_decompress(void* dst,  size_t originalSize,
-                               const void* cSrc, size_t cSrcSize);
-
-
-/* ***   Tool functions *** */
-#define HUF_BLOCKSIZE_MAX (128 * 1024)                  /*< maximum input size for a single block compressed with HUF_compress */
-HUF_PUBLIC_API size_t HUF_compressBound(size_t size);   /*< maximum compressed size (worst case) */
-
-/* Error Management */
-HUF_PUBLIC_API unsigned    HUF_isError(size_t code);       /*< tells if a return value is an error code */
-HUF_PUBLIC_API const char* HUF_getErrorName(size_t code);  /*< provides error code string (useful for debugging) */
-
-
-/* ***   Advanced function   *** */
-
-/* HUF_compress2() :
- *  Same as HUF_compress(), but offers control over `maxSymbolValue` and `tableLog`.
- * `maxSymbolValue` must be <= HUF_SYMBOLVALUE_MAX .
- * `tableLog` must be `<= HUF_TABLELOG_MAX` . */
-HUF_PUBLIC_API size_t HUF_compress2 (void* dst, size_t dstCapacity,
-                               const void* src, size_t srcSize,
-                               unsigned maxSymbolValue, unsigned tableLog);
-
-/* HUF_compress4X_wksp() :
- *  Same as HUF_compress2(), but uses externally allocated `workSpace`.
- * `workspace` must be at least as large as HUF_WORKSPACE_SIZE */
-#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
-#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))
-HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
-                                     const void* src, size_t srcSize,
-                                     unsigned maxSymbolValue, unsigned tableLog,
-                                     void* workSpace, size_t wkspSize);
-
-#endif   /* HUF_H_298734234 */
-
-/* ******************************************************************
- *  WARNING !!
- *  The following section contains advanced and experimental definitions
- *  which shall never be used in the context of a dynamic library,
- *  because they are not guaranteed to remain stable in the future.
- *  Only consider them in association with static linking.
- * *****************************************************************/
-#if !defined(HUF_H_HUF_STATIC_LINKING_ONLY)
-#define HUF_H_HUF_STATIC_LINKING_ONLY
-
-/* *** Dependencies *** */
-#include "mem.h"   /* U32 */
+#include "mem.h"          /* U32 */
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
 
+/* ***   Tool functions *** */
+#define HUF_BLOCKSIZE_MAX (128 * 1024)   /*< maximum input size for a single block compressed with HUF_compress */
+size_t HUF_compressBound(size_t size);   /*< maximum compressed size (worst case) */
+
+/* Error Management */
+unsigned    HUF_isError(size_t code);       /*< tells if a return value is an error code */
+const char* HUF_getErrorName(size_t code);  /*< provides error code string (useful for debugging) */
+
+
+#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
+#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))
 
 /* *** Constants *** */
 #define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_TABLELOG_ABSOLUTEMAX */
···
 /* ****************************************
 *  Advanced decompression functions
 ******************************************/
-size_t HUF_decompress4X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /*< single-symbol decoder */
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /*< double-symbols decoder */
-#endif
 
-size_t HUF_decompress4X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /*< decodes RLE and uncompressed */
-size_t HUF_decompress4X_hufOnly(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize); /*< considers RLE and uncompressed as errors */
-size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize); /*< considers RLE and uncompressed as errors */
-size_t HUF_decompress4X1_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /*< single-symbol decoder */
-size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /*< single-symbol decoder */
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /*< double-symbols decoder */
-size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /*< double-symbols decoder */
-#endif
+/*
+ * Huffman flags bitset.
+ * For all flags, 0 is the default value.
+ */
+typedef enum {
+    /*
+     * If compiled with DYNAMIC_BMI2: Set flag only if the CPU supports BMI2 at runtime.
+     * Otherwise: Ignored.
+     */
+    HUF_flags_bmi2 = (1 << 0),
+    /*
+     * If set: Test possible table depths to find the one that produces the smallest header + encoded size.
+     * If unset: Use heuristic to find the table depth.
+     */
+    HUF_flags_optimalDepth = (1 << 1),
+    /*
+     * If set: If the previous table can encode the input, always reuse the previous table.
+     * If unset: If the previous table can encode the input, reuse the previous table if it results in a smaller output.
+     */
+    HUF_flags_preferRepeat = (1 << 2),
+    /*
+     * If set: Sample the input and check if the sample is uncompressible, if it is then don't attempt to compress.
+     * If unset: Always histogram the entire input.
+     */
+    HUF_flags_suspectUncompressible = (1 << 3),
+    /*
+     * If set: Don't use assembly implementations
+     * If unset: Allow using assembly implementations
+     */
+    HUF_flags_disableAsm = (1 << 4),
+    /*
+     * If set: Don't use the fast decoding loop, always use the fallback decoding loop.
+     * If unset: Use the fast decoding loop when possible.
+     */
+    HUF_flags_disableFast = (1 << 5)
+} HUF_flags_e;
 
 
 /* ****************************************
  *  HUF detailed API
  * ****************************************/
+#define HUF_OPTIMAL_DEPTH_THRESHOLD ZSTD_btultra
 
 /*! HUF_compress() does the following:
  *  1. count symbol occurrence from source[] into table count[] using FSE_count() (exposed within "fse.h")
···
  *  For example, it's possible to compress several blocks using the same 'CTable',
  *  or to save and regenerate 'CTable' using external methods.
  */
-unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);
-size_t HUF_buildCTable (HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits);   /* @return : maxNbBits; CTable and count can overlap. In which case, CTable will overwrite count content */
-size_t HUF_writeCTable (void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog);
+unsigned HUF_minTableLog(unsigned symbolCardinality);
+unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue);
+unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, void* workSpace,
+ size_t wkspSize, HUF_CElt* table, const unsigned* count, int flags); /* table is used as scratch space for building and testing tables, not a return value */
 size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog, void* workspace, size_t workspaceSize);
-size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable);
-size_t HUF_compress4X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2);
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
 size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
 int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
 
···
    HUF_repeat_check, /*< Can use the previous table but it must be checked. Note : The previous table must have been constructed by HUF_compress{1, 4}X_repeat */
    HUF_repeat_valid  /*< Can use the previous table and it is assumed to be valid */
  } HUF_repeat;
+
 /* HUF_compress4X_repeat() :
  *  Same as HUF_compress4X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
  *  If it uses hufTable it does not modify hufTable or repeat.
···
                        const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned tableLog,
                        void* workSpace, size_t wkspSize,    /*< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
-                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2, unsigned suspectUncompressible);
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int flags);
 
 /* HUF_buildCTable_wksp() :
  *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
  * `workSpace` must be aligned on 4-bytes boundaries, and its size must be >= HUF_CTABLE_WORKSPACE_SIZE.
  */
-#define HUF_CTABLE_WORKSPACE_SIZE_U32 (2*HUF_SYMBOLVALUE_MAX +1 +1)
+#define HUF_CTABLE_WORKSPACE_SIZE_U32 ((4 * (HUF_SYMBOLVALUE_MAX + 1)) + 192)
 #define HUF_CTABLE_WORKSPACE_SIZE (HUF_CTABLE_WORKSPACE_SIZE_U32 * sizeof(unsigned))
 size_t HUF_buildCTable_wksp (HUF_CElt* tree,
                        const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
···
                           U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
                           const void* src, size_t srcSize,
                           void* workspace, size_t wkspSize,
-                          int bmi2);
+                          int flags);
 
 /* HUF_readCTable() :
  *  Loading a CTable saved with HUF_writeCTable() */
···
 
 /* HUF_getNbBitsFromCTable() :
  *  Read nbBits from CTable symbolTable, for symbol `symbolValue` presumed <= HUF_SYMBOLVALUE_MAX
- *  Note 1 : is not inlined, as HUF_CElt definition is private */
+ *  Note 1 : If symbolValue > HUF_readCTableHeader(symbolTable).maxSymbolValue, returns 0
+ *  Note 2 : is not inlined, as HUF_CElt definition is private
+ */
 U32 HUF_getNbBitsFromCTable(const HUF_CElt* symbolTable, U32 symbolValue);
+
+typedef struct {
+    BYTE tableLog;
+    BYTE maxSymbolValue;
+    BYTE unused[sizeof(size_t) - 2];
+} HUF_CTableHeader;
+
+/* HUF_readCTableHeader() :
+ *  @returns The header from the CTable specifying the tableLog and the maxSymbolValue.
+ */
+HUF_CTableHeader HUF_readCTableHeader(HUF_CElt const* ctable);
 
 /*
  * HUF_decompress() does the following:
···
 #define HUF_DECOMPRESS_WORKSPACE_SIZE ((2 << 10) + (1 << 9))
 #define HUF_DECOMPRESS_WORKSPACE_SIZE_U32 (HUF_DECOMPRESS_WORKSPACE_SIZE / sizeof(U32))
 
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_readDTableX1 (HUF_DTable* DTable, const void* src, size_t srcSize);
-size_t HUF_readDTableX1_wksp (HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize);
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_readDTableX2 (HUF_DTable* DTable, const void* src, size_t srcSize);
-size_t HUF_readDTableX2_wksp (HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize);
-#endif
-
-size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress4X1_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#endif
-
 
 /* ====================== */
 /* single stream variants */
 /* ====================== */
 
-size_t HUF_compress1X (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
-size_t HUF_compress1X_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);  /*< `workSpace` must be a table of at least HUF_WORKSPACE_SIZE_U64 U64 */
-size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable);
-size_t HUF_compress1X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2);
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
 /* HUF_compress1X_repeat() :
  *  Same as HUF_compress1X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
  *  If it uses hufTable it does not modify hufTable or repeat.
···
                        const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned tableLog,
                        void* workSpace, size_t wkspSize,   /*< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
-                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2, unsigned suspectUncompressible);
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int flags);
 
-size_t HUF_decompress1X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbol decoder */
-#endif
-
-size_t HUF_decompress1X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
-size_t HUF_decompress1X_DCtx_wksp (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /*< single-symbol decoder */
-size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /*< single-symbol decoder */
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /*< double-symbols decoder */
-size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /*< double-symbols decoder */
-#endif
-
-size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);   /*< automatic selection of sing or double symbol decoder, based on DTable */
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);   /*< double-symbols decoder */
 #endif
 
 /* BMI2 variants.
  * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
  */
-size_t HUF_decompress1X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
-size_t HUF_decompress4X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
-size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_readDTableX1_wksp_bmi2(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
 #ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_readDTableX2_wksp_bmi2(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_readDTableX2_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
 
-#endif /* HUF_STATIC_LINKING_ONLY */
-
+#endif   /* HUF_H_298734234 */

lib/zstd/common/mem.h

···
 /* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 /*-****************************************
 *  Compiler specifics
 ******************************************/
+#undef MEM_STATIC /* may be already defined from common/compiler.h */
 #define MEM_STATIC static inline
 
 /*-**************************************************************

lib/zstd/common/portability_macros.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #define ZSTD_PORTABILITY_MACROS_H
 
 /*
- * This header file contains macro defintions to support portability.
+ * This header file contains macro definitions to support portability.
  * This header is shared between C and ASM code, so it MUST only
  * contain macro definitions. It MUST not contain any C code.
  *
···
 /* Mark the internal assembly functions as hidden  */
 #ifdef __ELF__
 # define ZSTD_HIDE_ASM_FUNCTION(func) .hidden func
+#elif defined(__APPLE__)
+# define ZSTD_HIDE_ASM_FUNCTION(func) .private_extern func
 #else
 # define ZSTD_HIDE_ASM_FUNCTION(func)
 #endif
+
+/* Compile time determination of BMI2 support */
+
 
 /* Enable runtime BMI2 dispatch based on the CPU.
  * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
  */
 #ifndef DYNAMIC_BMI2
-  #if ((defined(__clang__) && __has_attribute(__target__)) \
+#  if ((defined(__clang__) && __has_attribute(__target__)) \
       || (defined(__GNUC__) \
           && (__GNUC__ >= 11))) \
-      && (defined(__x86_64__) || defined(_M_X64)) \
+      && (defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)) \
       && !defined(__BMI2__)
-  #  define DYNAMIC_BMI2 1
-  #else
-  #  define DYNAMIC_BMI2 0
-  #endif
+#    define DYNAMIC_BMI2 1
+#  else
+#    define DYNAMIC_BMI2 0
+#  endif
 #endif
 
 /*
- * Only enable assembly for GNUC comptabile compilers,
+ * Only enable assembly for GNU C compatible compilers,
  * because other platforms may not support GAS assembly syntax.
  *
- * Only enable assembly for Linux / MacOS, other platforms may
+ * Only enable assembly for Linux / MacOS / Win32, other platforms may
  * work, but they haven't been tested. This could likely be
  * extended to BSD systems.
  *
···
  *   - BMI2 is supported at compile time
  */
 #define ZSTD_ENABLE_ASM_X86_64_BMI2 0
+
+/*
+ * For x86 ELF targets, add .note.gnu.property section for Intel CET in
+ * assembly sources when CET is enabled.
+ *
+ * Additionally, any function that may be called indirectly must begin
+ * with ZSTD_CET_ENDBRANCH.
+ */
+#if defined(__ELF__) && (defined(__x86_64__) || defined(__i386__)) \
+    && defined(__has_include)
+# if __has_include(<cet.h>)
+#  include <cet.h>
+#  define ZSTD_CET_ENDBRANCH _CET_ENDBR
+# endif
+#endif
+
+#ifndef ZSTD_CET_ENDBRANCH
+# define ZSTD_CET_ENDBRANCH
+#endif
 
 #endif /* ZSTD_PORTABILITY_MACROS_H */

lib/zstd/common/zstd_common.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 *  Dependencies
 ***************************************/
 #define ZSTD_DEPS_NEED_MALLOC
-#include "zstd_deps.h"   /* ZSTD_malloc, ZSTD_calloc, ZSTD_free, ZSTD_memset */
 #include "error_private.h"
 #include "zstd_internal.h"
 
···
 /*! ZSTD_getErrorString() :
  *  provides error code string from enum */
 const char* ZSTD_getErrorString(ZSTD_ErrorCode code) { return ERR_getErrorString(code); }
-
-
-
-/*=**************************************************************
-*  Custom allocator
-****************************************************************/
-void* ZSTD_customMalloc(size_t size, ZSTD_customMem customMem)
-{
-    if (customMem.customAlloc)
-        return customMem.customAlloc(customMem.opaque, size);
-    return ZSTD_malloc(size);
-}
-
-void* ZSTD_customCalloc(size_t size, ZSTD_customMem customMem)
-{
-    if (customMem.customAlloc) {
-        /* calloc implemented as malloc+memset;
-         * not as efficient as calloc, but next best guess for custom malloc */
-        void* const ptr = customMem.customAlloc(customMem.opaque, size);
-        ZSTD_memset(ptr, 0, size);
-        return ptr;
-    }
-    return ZSTD_calloc(1, size);
-}
-
-void ZSTD_customFree(void* ptr, ZSTD_customMem customMem)
-{
-    if (ptr!=NULL) {
-        if (customMem.customFree)
-            customMem.customFree(customMem.opaque, ptr);
-        else
-            ZSTD_free(ptr);
-    }
-}

lib/zstd/common/zstd_deps.h

···
 /* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 
 #endif /* ZSTD_DEPS_IO */
 #endif /* ZSTD_DEPS_NEED_IO */
+
+/*
+ * Only requested when MSAN is enabled.
+ * Need:
+ * intptr_t
+ */
+#ifdef ZSTD_DEPS_NEED_STDINT
+#ifndef ZSTD_DEPS_STDINT
+#define ZSTD_DEPS_STDINT
+
+/* intptr_t already provided by ZSTD_DEPS_COMMON */
+
+#endif /* ZSTD_DEPS_STDINT */
+#endif /* ZSTD_DEPS_NEED_STDINT */

lib/zstd/common/zstd_internal.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #include <linux/zstd.h>
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
 #include <linux/xxhash.h>                /* XXH_reset, update, digest */
 #define ZSTD_TRACE 0
-
 
 /* ---- static assert (debug) --- */
 #define ZSTD_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)
···
 #define ZSTD_FRAMECHECKSUMSIZE 4
 
 #define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
-#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
+#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */)   /* for a non-null block */
+#define MIN_LITERALS_FOR_4_STREAMS 6
 
-#define HufLog 12
-typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingType_e;
+typedef enum { set_basic, set_rle, set_compressed, set_repeat } SymbolEncodingType_e;
 
 #define LONGNBSEQ 0x7F00
 
 #define MINMATCH 3
 
 #define Litbits  8
+#define LitHufLog 11
 #define MaxLit ((1<<Litbits) - 1)
 #define MaxML   52
 #define MaxLL   35
···
 #define LLFSELog    9
 #define OffFSELog   8
 #define MaxFSELog  MAX(MAX(MLFSELog, LLFSELog), OffFSELog)
+#define MaxMLBits 16
+#define MaxLLBits 16
 
 #define ZSTD_MAX_HUF_HEADER_SIZE 128 /* header + <= 127 byte tree description */
 /* Each table cannot take more than #symbols * FSELog bits */
···
     ZSTD_memcpy(dst, src, 8);
 #endif
 }
-#define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+#define COPY8(d,s) do { ZSTD_copy8(d,s); d+=8; s+=8; } while (0)
 
 /* Need to use memmove here since the literal buffer can now be located within
    the dst buffer. In circumstances where the op "catches up" to where the
···
     ZSTD_memcpy(dst, copy16_buf, 16);
 #endif
 }
-#define COPY16(d,s) { ZSTD_copy16(d,s); d+=16; s+=16; }
+#define COPY16(d,s) do { ZSTD_copy16(d,s); d+=16; s+=16; } while (0)
 
 #define WILDCOPY_OVERLENGTH 32
 #define WILDCOPY_VECLEN 16
···
     if (ovtype == ZSTD_overlap_src_before_dst && diff < WILDCOPY_VECLEN) {
         /* Handle short offset copies. */
         do {
-            COPY8(op, ip)
+            COPY8(op, ip);
         } while (op < oend);
     } else {
         assert(diff >= WILDCOPY_VECLEN || diff <= -WILDCOPY_VECLEN);
···
          * one COPY16() in the first call. Then, do two calls per loop since
          * at that point it is more likely to have a high trip count.
          */
-#ifdef __aarch64__
-        do {
-            COPY16(op, ip);
-        }
-        while (op < oend);
-#else
         ZSTD_copy16(op, ip);
         if (16 >= length) return;
         op += 16;
···
             COPY16(op, ip);
         }
         while (op < oend);
-#endif
     }
 }
 
···
 /*-*******************************************
 *  Private declarations
 *********************************************/
-typedef struct seqDef_s {
-    U32 offBase;   /* offBase == Offset + ZSTD_REP_NUM, or repcode 1,2,3 */
-    U16 litLength;
-    U16 mlBase;    /* mlBase == matchLength - MINMATCH */
-} seqDef;
-
-/* Controls whether seqStore has a single "long" litLength or matchLength. See seqStore_t. */
-typedef enum {
-    ZSTD_llt_none = 0,             /* no longLengthType */
-    ZSTD_llt_literalLength = 1,    /* represents a long literal */
-    ZSTD_llt_matchLength = 2       /* represents a long match */
-} ZSTD_longLengthType_e;
-
-typedef struct {
-    seqDef* sequencesStart;
-    seqDef* sequences;      /* ptr to end of sequences */
-    BYTE* litStart;
-    BYTE* lit;              /* ptr to end of literals */
-    BYTE* llCode;
-    BYTE* mlCode;
-    BYTE* ofCode;
-    size_t maxNbSeq;
-    size_t maxNbLit;
-
-    /* longLengthPos and longLengthType to allow us to represent either a single litLength or matchLength
-     * in the seqStore that has a value larger than U16 (if it exists). To do so, we increment
-     * the existing value of the litLength or matchLength by 0x10000.
-     */
-    ZSTD_longLengthType_e   longLengthType;
-    U32                     longLengthPos;  /* Index of the sequence to apply long length modification to */
-} seqStore_t;
-
-typedef struct {
-    U32 litLength;
-    U32 matchLength;
-} ZSTD_sequenceLength;
-
-/*
- * Returns the ZSTD_sequenceLength for the given sequences. It handles the decoding of long sequences
- * indicated by longLengthPos and longLengthType, and adds MINMATCH back to matchLength.
- */
-MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore, seqDef const* seq)
-{
-    ZSTD_sequenceLength seqLen;
-    seqLen.litLength = seq->litLength;
-    seqLen.matchLength = seq->mlBase + MINMATCH;
-    if (seqStore->longLengthPos == (U32)(seq - seqStore->sequencesStart)) {
-        if (seqStore->longLengthType == ZSTD_llt_literalLength) {
-            seqLen.litLength += 0xFFFF;
-        }
-        if (seqStore->longLengthType == ZSTD_llt_matchLength) {
-            seqLen.matchLength += 0xFFFF;
-        }
-    }
-    return seqLen;
-}
 
 /*
  * Contains the compressed frame size and an upper-bound for the decompressed frame size.
···
  *          `decompressedBound != ZSTD_CONTENTSIZE_ERROR`
  */
 typedef struct {
+    size_t nbBlocks;
     size_t compressedSize;
     unsigned long long decompressedBound;
 } ZSTD_frameSizeInfo;   /* decompress & legacy */
-
-const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx);   /* compress & dictBuilder */
-void ZSTD_seqToCodes(const seqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
-
-/* custom memory allocation functions */
-void* ZSTD_customMalloc(size_t size, ZSTD_customMem customMem);
-void* ZSTD_customCalloc(size_t size, ZSTD_customMem customMem);
-void ZSTD_customFree(void* ptr, ZSTD_customMem customMem);
-
-
-MEM_STATIC U32 ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
-{
-    assert(val != 0);
-    {
-#   if (__GNUC__ >= 3)   /* GCC Intrinsic */
-        return __builtin_clz (val) ^ 31;
-#   else   /* Software version */
-        static const U32 DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
-        U32 v = val;
-        v |= v >> 1;
-        v |= v >> 2;
-        v |= v >> 4;
-        v |= v >> 8;
-        v |= v >> 16;
-        return DeBruijnClz[(v * 0x07C4ACDDU) >> 27];
-#   endif
-    }
-}
-
-/*
- * Counts the number of trailing zeros of a `size_t`.
- * Most compilers should support CTZ as a builtin. A backup
- * implementation is provided if the builtin isn't supported, but
- * it may not be terribly efficient.
- */
-MEM_STATIC unsigned ZSTD_countTrailingZeros(size_t val)
-{
-    if (MEM_64bits()) {
-#       if (__GNUC__ >= 4)
-            return __builtin_ctzll((U64)val);
-#       else
-            static const int DeBruijnBytePos[64] = {  0,  1,  2,  7,  3, 13,  8, 19,
-                                                      4, 25, 14, 28,  9, 34, 20, 56,
-                                                      5, 17, 26, 54, 15, 41, 29, 43,
-                                                      10, 31, 38, 35, 21, 45, 49, 57,
-                                                      63,  6, 12, 18, 24, 27, 33, 55,
-                                                      16, 53, 40, 42, 30, 37, 44, 48,
-                                                      62, 11, 23, 32, 52, 39, 36, 47,
-                                                      61, 22, 51, 46, 60, 50, 59, 58 };
-            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
-#       endif
-    } else { /* 32 bits */
-#       if (__GNUC__ >= 3)
-            return __builtin_ctz((U32)val);
-#       else
-            static const int DeBruijnBytePos[32] = {  0,  1, 28,  2, 29, 14, 24,  3,
-                                                     30, 22, 20, 15, 25, 17,  4,  8,
-                                                     31, 27, 13, 23, 21, 19, 16,  7,
-                                                     26, 12, 18,  6, 11,  5, 10,  9 };
-            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
-#       endif
-    }
-}
-
 
 /* ZSTD_invalidateRepCodes() :
  * ensures next compression will not use repcodes from previous block.
···
 
 /*! ZSTD_getcBlockSize() :
  *  Provides the size of compressed block from block header `src` */
-/* Used by: decompress, fullbench (does not get its definition from here) */
+/*  Used by: decompress, fullbench */
 size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
                           blockProperties_t* bpPtr);
 
 /*! ZSTD_decodeSeqHeaders() :
  *  decode sequence header from src */
-/* Used by: decompress, fullbench (does not get its definition from here) */
+/*  Used by: zstd_decompress_block, fullbench */
 size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                        const void* src, size_t srcSize);
 
···
     ZSTD_cpuid_t cpuid = ZSTD_cpuid();
     return ZSTD_cpuid_bmi1(cpuid) && ZSTD_cpuid_bmi2(cpuid);
 }
-
 
 #endif   /* ZSTD_CCOMMON_H_MODULE */

lib/zstd/compress/clevels.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

lib/zstd/compress/fse_compress.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /* ******************************************************************
  * FSE : Finite State Entropy encoder
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 #include "../common/error_private.h"
 #define ZSTD_DEPS_NEED_MALLOC
 #define ZSTD_DEPS_NEED_MATH64
-#include "../common/zstd_deps.h"  /* ZSTD_malloc, ZSTD_free, ZSTD_memcpy, ZSTD_memset */
+#include "../common/zstd_deps.h"  /* ZSTD_memset */
+#include "../common/bits.h" /* ZSTD_highbit32 */
 
 
 /* **************************************************************
···
     assert(tableLog < 16);   /* required for threshold strategy to work */
 
     /* For explanations on how to distribute symbol values over the table :
-     * http://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
+     * https://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
 
      #ifdef __clang_analyzer__
      ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
···
                 break;
             default :
                 assert(normalizedCounter[s] > 1);
-                {   U32 const maxBitsOut = tableLog - BIT_highbit32 ((U32)normalizedCounter[s]-1);
+                {   U32 const maxBitsOut = tableLog - ZSTD_highbit32 ((U32)normalizedCounter[s]-1);
                     U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut;
                     symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
                     symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]);
···
     size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog
                                    + 4 /* bitCount initialized at 4 */
                                    + 2 /* first two symbols may use one additional bit each */) / 8)
-                                    + 1 /* round up to whole nb bytes */
-                                    + 2 /* additional two bytes for bitstream flush */;
+                                   + 1 /* round up to whole nb bytes */
+                                   + 2 /* additional two bytes for bitstream flush */;
     return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;  /* maxSymbolValue==0 ? use default */
 }
 
···
     /* Init */
     remaining = tableSize+1;   /* +1 for extra accuracy */
     threshold = tableSize;
-    nbBits = tableLog+1;
+    nbBits = (int)tableLog+1;
 
     while ((symbol < alphabetSize) && (remaining>1)) {  /* stops at 1 */
         if (previousIs0) {
···
             }
             while (symbol >= start+3) {
                 start+=3;
-                bitStream += 3 << bitCount;
+                bitStream += 3U << bitCount;
                 bitCount += 2;
             }
             bitStream += (symbol-start) << bitCount;
···
             count++;   /* +1 for extra accuracy */
             if (count>=threshold)
                 count += max;   /* [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ */
-            bitStream += count << bitCount;
+            bitStream += (U32)count << bitCount;
             bitCount  += nbBits;
             bitCount  -= (count<max);
             previousIs0  = (count==1);
···
     out[1] = (BYTE)(bitStream>>8);
     out+= (bitCount+7) /8;
 
-    return (out-ostart);
+    assert(out >= ostart);
+    return (size_t)(out-ostart);
 }
 
 
···
 *  FSE Compression Code
 ****************************************************************/
 
-FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog)
-{
-    size_t size;
-    if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
-    size = FSE_CTABLE_SIZE_U32 (tableLog, maxSymbolValue) * sizeof(U32);
-    return (FSE_CTable*)ZSTD_malloc(size);
-}
-
-void FSE_freeCTable (FSE_CTable* ct) { ZSTD_free(ct); }
-
 /* provides the minimum logSize to safely represent a distribution */
 static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
 {
-    U32 minBitsSrc = BIT_highbit32((U32)(srcSize)) + 1;
-    U32 minBitsSymbols = BIT_highbit32(maxSymbolValue) + 2;
+    U32 minBitsSrc = ZSTD_highbit32((U32)(srcSize)) + 1;
+    U32 minBitsSymbols = ZSTD_highbit32(maxSymbolValue) + 2;
     U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
     assert(srcSize > 1); /* Not supported, RLE should be used instead */
     return minBits;
···
 
 unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus)
 {
-    U32 maxBitsSrc = BIT_highbit32((U32)(srcSize - 1)) - minus;
+    U32 maxBitsSrc = ZSTD_highbit32((U32)(srcSize - 1)) - minus;
     U32 tableLog = maxTableLog;
     U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue);
     assert(srcSize > 1); /* Not supported, RLE should be used instead */
···
     return tableLog;
 }
 
-
-/* fake FSE_CTable, for raw (uncompressed) input */
-size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits)
-{
-    const unsigned tableSize = 1 << nbBits;
-    const unsigned tableMask = tableSize - 1;
-    const unsigned maxSymbolValue = tableMask;
-    void* const ptr = ct;
-    U16* const tableU16 = ( (U16*) ptr) + 2;
-    void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableSize>>1);   /* assumption : tableLog >= 1 */
-    FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
-    unsigned s;
-
-    /* Sanity checks */
-    if (nbBits < 1) return ERROR(GENERIC);             /* min size */
-
-    /* header */
-    tableU16[-2] = (U16) nbBits;
-    tableU16[-1] = (U16) maxSymbolValue;
-
-    /* Build table */
-    for (s=0; s<tableSize; s++)
-        tableU16[s] = (U16)(tableSize + s);
-
-    /* Build Symbol Transformation Table */
-    {   const U32 deltaNbBits = (nbBits << 16) - (1 << nbBits);
-        for (s=0; s<=maxSymbolValue; s++) {
-            symbolTT[s].deltaNbBits = deltaNbBits;
-            symbolTT[s].deltaFindState = s-1;
-    }   }
-
-    return 0;
-}
-
 /* fake FSE_CTable, for rle input (always same symbol) */
 size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue)
 {
···
 
 
 size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }
-
 
 #endif   /* FSE_COMMONDEFS_ONLY */

lib/zstd/compress/hist.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 /*-**************************************************************
  *  Histogram functions
  ****************************************************************/
+void HIST_add(unsigned* count, const void* src, size_t srcSize)
+{
+    const BYTE* ip = (const BYTE*)src;
+    const BYTE* const end = ip + srcSize;
+
+    while (ip<end) {
+        count[*ip++]++;
+    }
+}
+
 unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr,
                            const void* src, size_t srcSize)
 {

lib/zstd/compress/hist.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
  */
 unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr,
                            const void* src, size_t srcSize);
+
+/*! HIST_add() :
+ *  Lowest level: just add nb of occurrences of characters from @src into @count.
+ *  @count is not reset. @count array is presumed large enough (i.e. 1 KB).
+ @  This function does not need any additional stack memory.
+ */
+void HIST_add(unsigned* count, const void* src, size_t srcSize);

lib/zstd/compress/huf_compress.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /* ******************************************************************
  * Huffman encoder, part of New Generation Entropy library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 #include "hist.h"
 #define FSE_STATIC_LINKING_ONLY   /* FSE_optimalTableLog_internal */
 #include "../common/fse.h"        /* header compression */
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"
 #include "../common/error_private.h"
+#include "../common/bits.h"       /* ZSTD_highbit32 */
 
 
 /* **************************************************************
···
 
 
 /* **************************************************************
-*  Utils
+*  Required declarations
 ****************************************************************/
-unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
+typedef struct nodeElt_s {
+    U32 count;
+    U16 parent;
+    BYTE byte;
+    BYTE nbBits;
+} nodeElt;
+
+
+/* **************************************************************
+*  Debug Traces
+****************************************************************/
+
+#if DEBUGLEVEL >= 2
+
+static size_t showU32(const U32* arr, size_t size)
 {
-    return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", arr[u]); (void)arr;
+    }
+    RAWLOG(6, " \n");
+    return size;
 }
+
+static size_t HUF_getNbBits(HUF_CElt elt);
+
+static size_t showCTableBits(const HUF_CElt* ctable, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %zu", HUF_getNbBits(ctable[u])); (void)ctable;
+    }
+    RAWLOG(6, " \n");
+    return size;
+
+}
+
+static size_t showHNodeSymbols(const nodeElt* hnode, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", hnode[u].byte); (void)hnode;
+    }
+    RAWLOG(6, " \n");
+    return size;
+}
+
+static size_t showHNodeBits(const nodeElt* hnode, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", hnode[u].nbBits); (void)hnode;
+    }
+    RAWLOG(6, " \n");
+    return size;
+}
+
+#endif
 
 
 /* *******************************************************
···
     S16 norm[HUF_TABLELOG_MAX+1];
 } HUF_CompressWeightsWksp;
 
-static size_t HUF_compressWeights(void* dst, size_t dstSize, const void* weightTable, size_t wtSize, void* workspace, size_t workspaceSize)
+static size_t
+HUF_compressWeights(void* dst, size_t dstSize,
+              const void* weightTable, size_t wtSize,
+                    void* workspace, size_t workspaceSize)
 {
     BYTE* const ostart = (BYTE*) dst;
     BYTE* op = ostart;
···
 
 static size_t HUF_getValue(HUF_CElt elt)
 {
-    return elt & ~0xFF;
+    return elt & ~(size_t)0xFF;
 }
 
 static size_t HUF_getValueFast(HUF_CElt elt)
···
     }
 }
 
+HUF_CTableHeader HUF_readCTableHeader(HUF_CElt const* ctable)
+{
+    HUF_CTableHeader header;
+    ZSTD_memcpy(&header, ctable, sizeof(header));
+    return header;
+}
+
+static void HUF_writeCTableHeader(HUF_CElt* ctable, U32 tableLog, U32 maxSymbolValue)
+{
+    HUF_CTableHeader header;
+    HUF_STATIC_ASSERT(sizeof(ctable[0]) == sizeof(header));
+    ZSTD_memset(&header, 0, sizeof(header));
+    assert(tableLog < 256);
+    header.tableLog = (BYTE)tableLog;
+    assert(maxSymbolValue < 256);
+    header.maxSymbolValue = (BYTE)maxSymbolValue;
+    ZSTD_memcpy(ctable, &header, sizeof(header));
+}
+
 typedef struct {
     HUF_CompressWeightsWksp wksp;
     BYTE bitsToWeight[HUF_TABLELOG_MAX + 1];   /* precomputed conversion table */
···
     BYTE* op = (BYTE*)dst;
     U32 n;
     HUF_WriteCTableWksp* wksp = (HUF_WriteCTableWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32));
+
+    HUF_STATIC_ASSERT(HUF_CTABLE_WORKSPACE_SIZE >= sizeof(HUF_WriteCTableWksp));
+
+    assert(HUF_readCTableHeader(CTable).maxSymbolValue == maxSymbolValue);
+    assert(HUF_readCTableHeader(CTable).tableLog == huffLog);
 
     /* check conditions */
     if (workspaceSize < sizeof(HUF_WriteCTableWksp)) return ERROR(GENERIC);
···
     return ((maxSymbolValue+1)/2) + 1;
 }
 
-/*! HUF_writeCTable() :
-    `CTable` : Huffman tree to save, using huf representation.
-    @return : size of saved CTable */
-size_t HUF_writeCTable (void* dst, size_t maxDstSize,
-                        const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog)
-{
-    HUF_WriteCTableWksp wksp;
-    return HUF_writeCTable_wksp(dst, maxDstSize, CTable, maxSymbolValue, huffLog, &wksp, sizeof(wksp));
-}
-
 
 size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned* hasZeroWeights)
 {
···
     if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
     if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall);
 
-    CTable[0] = tableLog;
+    *maxSymbolValuePtr = nbSymbols - 1;
+
+    HUF_writeCTableHeader(CTable, tableLog, *maxSymbolValuePtr);
 
     /* Prepare base value per rank */
     {   U32 n, nextRankStart = 0;
···
         { U32 n; for (n=0; n<nbSymbols; n++) HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++); }
     }
 
-    *maxSymbolValuePtr = nbSymbols - 1;
     return readSize;
 }
 
 U32 HUF_getNbBitsFromCTable(HUF_CElt const* CTable, U32 symbolValue)
 {
-    const HUF_CElt* ct = CTable + 1;
+    const HUF_CElt* const ct = CTable + 1;
     assert(symbolValue <= HUF_SYMBOLVALUE_MAX);
+    if (symbolValue > HUF_readCTableHeader(CTable).maxSymbolValue)
+        return 0;
     return (U32)HUF_getNbBits(ct[symbolValue]);
 }
 
 
-typedef struct nodeElt_s {
-    U32 count;
-    U16 parent;
-    BYTE byte;
-    BYTE nbBits;
-} nodeElt;
-
 /*
  * HUF_setMaxHeight():
- * Enforces maxNbBits on the Huffman tree described in huffNode.
+ * Try to enforce @targetNbBits on the Huffman tree described in @huffNode.
  *
- * It sets all nodes with nbBits > maxNbBits to be maxNbBits. Then it adjusts
- * the tree to so that it is a valid canonical Huffman tree.
+ * It attempts to convert all nodes with nbBits > @targetNbBits
+ * to employ @targetNbBits instead. Then it adjusts the tree
+ * so that it remains a valid canonical Huffman tree.
  *
  * @pre               The sum of the ranks of each symbol == 2^largestBits,
  *                    where largestBits == huffNode[lastNonNull].nbBits.
  * @post              The sum of the ranks of each symbol == 2^largestBits,
- *                    where largestBits is the return value <= maxNbBits.
+ *                    where largestBits is the return value (expected <= targetNbBits).
  *
- * @param huffNode    The Huffman tree modified in place to enforce maxNbBits.
+ * @param huffNode    The Huffman tree modified in place to enforce targetNbBits.
+ *                    It's presumed sorted, from most frequent to rarest symbol.
  * @param lastNonNull The symbol with the lowest count in the Huffman tree.
- * @param maxNbBits   The maximum allowed number of bits, which the Huffman tree
+ * @param targetNbBits  The allowed number of bits, which the Huffman tree
  *                    may not respect. After this function the Huffman tree will
- *                    respect maxNbBits.
- * @return            The maximum number of bits of the Huffman tree after adjustment,
- *                    necessarily no more than maxNbBits.
+ *                    respect targetNbBits.
+ * @return            The maximum number of bits of the Huffman tree after adjustment.
  */
-static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
+static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 targetNbBits)
 {
     const U32 largestBits = huffNode[lastNonNull].nbBits;
-    /* early exit : no elt > maxNbBits, so the tree is already valid. */
-    if (largestBits <= maxNbBits) return largestBits;
+    /* early exit : no elt > targetNbBits, so the tree is already valid. */
+    if (largestBits <= targetNbBits) return largestBits;
+
+    DEBUGLOG(5, "HUF_setMaxHeight (targetNbBits = %u)", targetNbBits);
 
     /* there are several too large elements (at least >= 2) */
     {   int totalCost = 0;
-        const U32 baseCost = 1 << (largestBits - maxNbBits);
+        const U32 baseCost = 1 << (largestBits - targetNbBits);
         int n = (int)lastNonNull;
 
-        /* Adjust any ranks > maxNbBits to maxNbBits.
+        /* Adjust any ranks > targetNbBits to targetNbBits.
          * Compute totalCost, which is how far the sum of the ranks is
          * we are over 2^largestBits after adjust the offending ranks.
          */
-        while (huffNode[n].nbBits > maxNbBits) {
+        while (huffNode[n].nbBits > targetNbBits) {
             totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits));
-            huffNode[n].nbBits = (BYTE)maxNbBits;
+            huffNode[n].nbBits = (BYTE)targetNbBits;
             n--;
         }
-        /* n stops at huffNode[n].nbBits <= maxNbBits */
-        assert(huffNode[n].nbBits <= maxNbBits);
-        /* n end at index of smallest symbol using < maxNbBits */
-        while (huffNode[n].nbBits == maxNbBits) --n;
+        /* n stops at huffNode[n].nbBits <= targetNbBits */
+        assert(huffNode[n].nbBits <= targetNbBits);
+        /* n end at index of smallest symbol using < targetNbBits */
+        while (huffNode[n].nbBits == targetNbBits) --n;
 
-        /* renorm totalCost from 2^largestBits to 2^maxNbBits
+        /* renorm totalCost from 2^largestBits to 2^targetNbBits
          * note : totalCost is necessarily a multiple of baseCost */
-        assert((totalCost & (baseCost - 1)) == 0);
-        totalCost >>= (largestBits - maxNbBits);
+        assert(((U32)totalCost & (baseCost - 1)) == 0);
+        totalCost >>= (largestBits - targetNbBits);
         assert(totalCost > 0);
 
         /* repay normalized cost */
···
 
             /* Get pos of last (smallest = lowest cum. count) symbol per rank */
             ZSTD_memset(rankLast, 0xF0, sizeof(rankLast));
-            {   U32 currentNbBits = maxNbBits;
+            {   U32 currentNbBits = targetNbBits;
                 int pos;
                 for (pos=n ; pos >= 0; pos--) {
                     if (huffNode[pos].nbBits >= currentNbBits) continue;
-                    currentNbBits = huffNode[pos].nbBits;   /* < maxNbBits */
-                    rankLast[maxNbBits-currentNbBits] = (U32)pos;
+                    currentNbBits = huffNode[pos].nbBits;   /* < targetNbBits */
+                    rankLast[targetNbBits-currentNbBits] = (U32)pos;
             }   }
 
             while (totalCost > 0) {
                 /* Try to reduce the next power of 2 above totalCost because we
                  * gain back half the rank.
                  */
-                U32 nBitsToDecrease = BIT_highbit32((U32)totalCost) + 1;
+                U32 nBitsToDecrease = ZSTD_highbit32((U32)totalCost) + 1;
                 for ( ; nBitsToDecrease > 1; nBitsToDecrease--) {
                     U32 const highPos = rankLast[nBitsToDecrease];
                     U32 const lowPos = rankLast[nBitsToDecrease-1];
···
                     rankLast[nBitsToDecrease] = noSymbol;
                 else {
                     rankLast[nBitsToDecrease]--;
-                    if (huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease)
+                    if (huffNode[rankLast[nBitsToDecrease]].nbBits != targetNbBits-nBitsToDecrease)
                         rankLast[nBitsToDecrease] = noSymbol;   /* this rank is now empty */
                 }
             }   /* while (totalCost > 0) */
···
              * TODO.
              */
             while (totalCost < 0) {  /* Sometimes, cost correction overshoot */
-                /* special case : no rank 1 symbol (using maxNbBits-1);
-                 * let's create one from largest rank 0 (using maxNbBits).
+                /* special case : no rank 1 symbol (using targetNbBits-1);
+                 * let's create one from largest rank 0 (using targetNbBits).
                  */
                 if (rankLast[1] == noSymbol) {
-                    while (huffNode[n].nbBits == maxNbBits) n--;
+                    while (huffNode[n].nbBits == targetNbBits) n--;
                     huffNode[n+1].nbBits--;
                     assert(n >= 0);
                     rankLast[1] = (U32)(n+1);
···
         }   /* repay normalized cost */
     }   /* there are several too large elements (at least >= 2) */
 
-    return maxNbBits;
+    return targetNbBits;
 }
 
 typedef struct {
···
     U16 curr;
 } rankPos;
 
-typedef nodeElt huffNodeTable[HUF_CTABLE_WORKSPACE_SIZE_U32];
+typedef nodeElt huffNodeTable[2 * (HUF_SYMBOLVALUE_MAX + 1)];
 
 /* Number of buckets available for HUF_sort() */
 #define RANK_POSITION_TABLE_SIZE 192
···
  * Let buckets 166 to 192 represent all remaining counts up to RANK_POSITION_MAX_COUNT_LOG using log2 bucketing.
  */
 #define RANK_POSITION_MAX_COUNT_LOG 32
-#define RANK_POSITION_LOG_BUCKETS_BEGIN (RANK_POSITION_TABLE_SIZE - 1) - RANK_POSITION_MAX_COUNT_LOG - 1 /* == 158 */
-#define RANK_POSITION_DISTINCT_COUNT_CUTOFF RANK_POSITION_LOG_BUCKETS_BEGIN + BIT_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN) /* == 166 */
+#define RANK_POSITION_LOG_BUCKETS_BEGIN ((RANK_POSITION_TABLE_SIZE - 1) - RANK_POSITION_MAX_COUNT_LOG - 1 /* == 158 */)
+#define RANK_POSITION_DISTINCT_COUNT_CUTOFF (RANK_POSITION_LOG_BUCKETS_BEGIN + ZSTD_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN) /* == 166 */)
 
 /* Return the appropriate bucket index for a given count. See definition of
  * RANK_POSITION_DISTINCT_COUNT_CUTOFF for explanation of bucketing strategy.
···
 static U32 HUF_getIndex(U32 const count) {
     return (count < RANK_POSITION_DISTINCT_COUNT_CUTOFF)
         ? count
-        : BIT_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN;
+        : ZSTD_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN;
 }
 
 /* Helper swap function for HUF_quickSortPartition() */
···
 
     /* Sort each bucket. */
     for (n = RANK_POSITION_DISTINCT_COUNT_CUTOFF; n < RANK_POSITION_TABLE_SIZE - 1; ++n) {
-        U32 const bucketSize = rankPosition[n].curr-rankPosition[n].base;
+        int const bucketSize = rankPosition[n].curr - rankPosition[n].base;
         U32 const bucketStartIdx = rankPosition[n].base;
         if (bucketSize > 1) {
             assert(bucketStartIdx < maxSymbolValue1);
···
 
     assert(HUF_isSorted(huffNode, maxSymbolValue1));
 }
+
 
 /* HUF_buildCTable_wksp() :
  *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
···
     int lowS, lowN;
     int nodeNb = STARTNODE;
     int n, nodeRoot;
+    DEBUGLOG(5, "HUF_buildTree (alphabet size = %u)", maxSymbolValue + 1);
     /* init for parents */
     nonNullRank = (int)maxSymbolValue;
     while(huffNode[nonNullRank].count == 0) nonNullRank--;
···
         huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
     for (n=0; n<=nonNullRank; n++)
         huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
+
+    DEBUGLOG(6, "Initial distribution of bits completed (%zu sorted symbols)", showHNodeBits(huffNode, maxSymbolValue+1));
 
     return nonNullRank;
 }
···
         HUF_setNbBits(ct + huffNode[n].byte, huffNode[n].nbBits);   /* push nbBits per symbol, symbol order */
     for (n=0; n<alphabetSize; n++)
         HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++);   /* assign value within rank, symbol order */
-    CTable[0] = maxNbBits;
+
+    HUF_writeCTableHeader(CTable, maxNbBits, maxSymbolValue);
 }
 
-size_t HUF_buildCTable_wksp (HUF_CElt* CTable, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize)
+size_t
+HUF_buildCTable_wksp(HUF_CElt* CTable, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
+                     void* workSpace, size_t wkspSize)
 {
-    HUF_buildCTable_wksp_tables* const wksp_tables = (HUF_buildCTable_wksp_tables*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(U32));
+    HUF_buildCTable_wksp_tables* const wksp_tables =
+        (HUF_buildCTable_wksp_tables*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(U32));
     nodeElt* const huffNode0 = wksp_tables->huffNodeTbl;
     nodeElt* const huffNode = huffNode0+1;
     int nonNullRank;
 
+    HUF_STATIC_ASSERT(HUF_CTABLE_WORKSPACE_SIZE == sizeof(HUF_buildCTable_wksp_tables));
+
+    DEBUGLOG(5, "HUF_buildCTable_wksp (alphabet size = %u)", maxSymbolValue+1);
+
     /* safety checks */
     if (wkspSize < sizeof(HUF_buildCTable_wksp_tables))
-      return ERROR(workSpace_tooSmall);
+        return ERROR(workSpace_tooSmall);
     if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT;
     if (maxSymbolValue > HUF_SYMBOLVALUE_MAX)
-      return ERROR(maxSymbolValue_tooLarge);
+        return ERROR(maxSymbolValue_tooLarge);
     ZSTD_memset(huffNode0, 0, sizeof(huffNodeTable));
 
     /* sort, decreasing order */
     HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition);
+    DEBUGLOG(6, "sorted symbols completed (%zu symbols)", showHNodeSymbols(huffNode, maxSymbolValue+1));
 
     /* build tree */
     nonNullRank = HUF_buildTree(huffNode, maxSymbolValue);
 
-    /* enforce maxTableLog */
+    /* determine and enforce maxTableLog */
     maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits);
     if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC);   /* check fit into table */
 
···
 }
 
 int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) {
-  HUF_CElt const* ct = CTable + 1;
-  int bad = 0;
-  int s;
-  for (s = 0; s <= (int)maxSymbolValue; ++s) {
-    bad |= (count[s] != 0) & (HUF_getNbBits(ct[s]) == 0);
-  }
-  return !bad;
+    HUF_CTableHeader header = HUF_readCTableHeader(CTable);
+    HUF_CElt const* ct = CTable + 1;
+    int bad = 0;
+    int s;
+
+    assert(header.tableLog <= HUF_TABLELOG_ABSOLUTEMAX);
+
+    if (header.maxSymbolValue < maxSymbolValue)
+        return 0;
+
+    for (s = 0; s <= (int)maxSymbolValue; ++s) {
+        bad |= (count[s] != 0) & (HUF_getNbBits(ct[s]) == 0);
+    }
+    return !bad;
 }
 
 size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); }
···
 #if DEBUGLEVEL >= 1
     {
         size_t const nbBits = HUF_getNbBits(elt);
-        size_t const dirtyBits = nbBits == 0 ? 0 : BIT_highbit32((U32)nbBits) + 1;
+        size_t const dirtyBits = nbBits == 0 ? 0 : ZSTD_highbit32((U32)nbBits) + 1;
         (void)dirtyBits;
         /* Middle bits are 0. */
         assert(((elt >> dirtyBits) << (dirtyBits + nbBits)) == 0);
···
     {
         size_t const nbBits = bitC->bitPos[0] & 0xFF;
         if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
-        return (bitC->ptr - bitC->startPtr) + (nbBits > 0);
+        return (size_t)(bitC->ptr - bitC->startPtr) + (nbBits > 0);
     }
 }
 
···
                                    const void* src, size_t srcSize,
                                    const HUF_CElt* CTable)
 {
-    U32 const tableLog = (U32)CTable[0];
+    U32 const tableLog = HUF_readCTableHeader(CTable).tableLog;
     HUF_CElt const* ct = CTable + 1;
     const BYTE* ip = (const BYTE*) src;
     BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstSize;
-    BYTE* op = ostart;
     HUF_CStream_t bitC;
 
     /* init */
     if (dstSize < 8) return 0;   /* not enough space to compress */
-    { size_t const initErr = HUF_initCStream(&bitC, op, (size_t)(oend-op));
+    { BYTE* op = ostart;
+      size_t const initErr = HUF_initCStream(&bitC, op, (size_t)(oend-op));
       if (HUF_isError(initErr)) return 0; }
 
     if (dstSize < HUF_tightCompressBound(srcSize, (size_t)tableLog) || tableLog > 11)
···
 static size_t
 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, const int bmi2)
+                              const HUF_CElt* CTable, const int flags)
 {
-    if (bmi2) {
+    if (flags & HUF_flags_bmi2) {
         return HUF_compress1X_usingCTable_internal_bmi2(dst, dstSize, src, srcSize, CTable);
     }
     return HUF_compress1X_usingCTable_internal_default(dst, dstSize, src, srcSize, CTable);
···
 static size_t
 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, const int bmi2)
+                              const HUF_CElt* CTable, const int flags)
 {
-    (void)bmi2;
+    (void)flags;
     return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
 }
 
 #endif
 
-size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags)
 {
-    return HUF_compress1X_usingCTable_bmi2(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
-}
-
-size_t HUF_compress1X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2)
-{
-    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, bmi2);
+    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, flags);
 }
 
 static size_t
 HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, int bmi2)
+                              const HUF_CElt* CTable, int flags)
 {
     size_t const segmentSize = (srcSize+3)/4;   /* first 3 segments */
     const BYTE* ip = (const BYTE*) src;
···
     op += 6;   /* jumpTable */
 
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
         if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart, (U16)cSize);
         op += cSize;
···
 
     ip += segmentSize;
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
         if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart+2, (U16)cSize);
         op += cSize;
···
 
     ip += segmentSize;
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
         if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart+4, (U16)cSize);
         op += cSize;
···
     ip += segmentSize;
     assert(op <= oend);
     assert(ip <= iend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, bmi2) );
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, flags) );
         if (cSize == 0 || cSize > 65535) return 0;
         op += cSize;
     }
···
     return (size_t)(op-ostart);
 }
 
-size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags)
 {
-    return HUF_compress4X_usingCTable_bmi2(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
-}
-
-size_t HUF_compress4X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2)
-{
-    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, bmi2);
+    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, flags);
 }
 
 typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e;
···
 static size_t HUF_compressCTable_internal(
                 BYTE* const ostart, BYTE* op, BYTE* const oend,
                 const void* src, size_t srcSize,
-                HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int bmi2)
+                HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int flags)
 {
     size_t const cSize = (nbStreams==HUF_singleStream) ?
-                         HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2) :
-                         HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2);
+                         HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, flags) :
+                         HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, flags);
     if (HUF_isError(cSize)) { return cSize; }
     if (cSize==0) { return 0; }   /* uncompressible */
     op += cSize;
···
 #define SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE 4096
 #define SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO 10  /* Must be >= 2 */
 
+unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue)
+{
+    unsigned cardinality = 0;
+    unsigned i;
+
+    for (i = 0; i < maxSymbolValue + 1; i++) {
+        if (count[i] != 0) cardinality += 1;
+    }
+
+    return cardinality;
+}
+
+unsigned HUF_minTableLog(unsigned symbolCardinality)
+{
+    U32 minBitsSymbols = ZSTD_highbit32(symbolCardinality) + 1;
+    return minBitsSymbols;
+}
+
+unsigned HUF_optimalTableLog(
+            unsigned maxTableLog,
+            size_t srcSize,
+            unsigned maxSymbolValue,
+            void* workSpace, size_t wkspSize,
+            HUF_CElt* table,
+      const unsigned* count,
+            int flags)
+{
+    assert(srcSize > 1); /* Not supported, RLE should be used instead */
+    assert(wkspSize >= sizeof(HUF_buildCTable_wksp_tables));
+
+    if (!(flags & HUF_flags_optimalDepth)) {
+        /* cheap evaluation, based on FSE */
+        return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
+    }
+
+    {   BYTE* dst = (BYTE*)workSpace + sizeof(HUF_WriteCTableWksp);
+        size_t dstSize = wkspSize - sizeof(HUF_WriteCTableWksp);
+        size_t hSize, newSize;
+        const unsigned symbolCardinality = HUF_cardinality(count, maxSymbolValue);
+        const unsigned minTableLog = HUF_minTableLog(symbolCardinality);
+        size_t optSize = ((size_t) ~0) - 1;
+        unsigned optLog = maxTableLog, optLogGuess;
+
+        DEBUGLOG(6, "HUF_optimalTableLog: probing huf depth (srcSize=%zu)", srcSize);
+
+        /* Search until size increases */
+        for (optLogGuess = minTableLog; optLogGuess <= maxTableLog; optLogGuess++) {
+            DEBUGLOG(7, "checking for huffLog=%u", optLogGuess);
+
+            {   size_t maxBits = HUF_buildCTable_wksp(table, count, maxSymbolValue, optLogGuess, workSpace, wkspSize);
+                if (ERR_isError(maxBits)) continue;
+
+                if (maxBits < optLogGuess && optLogGuess > minTableLog) break;
+
+                hSize = HUF_writeCTable_wksp(dst, dstSize, table, maxSymbolValue, (U32)maxBits, workSpace, wkspSize);
+            }
+
+            if (ERR_isError(hSize)) continue;
+
+            newSize = HUF_estimateCompressedSize(table, count, maxSymbolValue) + hSize;
+
+            if (newSize > optSize + 1) {
+                break;
+            }
+
+            if (newSize < optSize) {
+                optSize = newSize;
+                optLog = optLogGuess;
+            }
+        }
+        assert(optLog <= HUF_TABLELOG_MAX);
+        return optLog;
+    }
+}
+
 /* HUF_compress_internal() :
  * `workSpace_align4` must be aligned on 4-bytes boundaries,
  * and occupies the same space as a table of HUF_WORKSPACE_SIZE_U64 unsigned */
···
                        unsigned maxSymbolValue, unsigned huffLog,
                        HUF_nbStreams_e nbStreams,
                        void* workSpace, size_t wkspSize,
-                       HUF_CElt* oldHufTable, HUF_repeat* repeat, int preferRepeat,
-                 const int bmi2, unsigned suspectUncompressible)
+                       HUF_CElt* oldHufTable, HUF_repeat* repeat, int flags)
 {
     HUF_compress_tables_t* const table = (HUF_compress_tables_t*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(size_t));
     BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstSize;
     BYTE* op = ostart;
 
+    DEBUGLOG(5, "HUF_compress_internal (srcSize=%zu)", srcSize);
     HUF_STATIC_ASSERT(sizeof(*table) + HUF_WORKSPACE_MAX_ALIGNMENT <= HUF_WORKSPACE_SIZE);
 
     /* checks & inits */
···
     if (!huffLog) huffLog = HUF_TABLELOG_DEFAULT;
 
     /* Heuristic : If old table is valid, use it for small inputs */
-    if (preferRepeat && repeat && *repeat == HUF_repeat_valid) {
+    if ((flags & HUF_flags_preferRepeat) && repeat && *repeat == HUF_repeat_valid) {
         return HUF_compressCTable_internal(ostart, op, oend,
                                            src, srcSize,
-                                           nbStreams, oldHufTable, bmi2);
+                                           nbStreams, oldHufTable, flags);
     }
 
     /* If uncompressible data is suspected, do a smaller sampling first */
     DEBUG_STATIC_ASSERT(SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO >= 2);
-    if (suspectUncompressible && srcSize >= (SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE * SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO)) {
+    if ((flags & HUF_flags_suspectUncompressible) && srcSize >= (SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE * SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO)) {
         size_t largestTotal = 0;
+        DEBUGLOG(5, "input suspected incompressible : sampling to check");
         {   unsigned maxSymbolValueBegin = maxSymbolValue;
             CHECK_V_F(largestBegin, HIST_count_simple (table->count, &maxSymbolValueBegin, (const BYTE*)src, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
             largestTotal += largestBegin;
···
         if (largest == srcSize) { *ostart = ((const BYTE*)src)[0]; return 1; }   /* single symbol, rle */
         if (largest <= (srcSize >> 7)+4) return 0;   /* heuristic : probably not compressible enough */
     }
+    DEBUGLOG(6, "histogram detail completed (%zu symbols)", showU32(table->count, maxSymbolValue+1));
 
     /* Check validity of previous table */
     if ( repeat
···
         *repeat = HUF_repeat_none;
     }
     /* Heuristic : use existing table for small inputs */
-    if (preferRepeat && repeat && *repeat != HUF_repeat_none) {
+    if ((flags & HUF_flags_preferRepeat) && repeat && *repeat != HUF_repeat_none) {
         return HUF_compressCTable_internal(ostart, op, oend,
                                            src, srcSize,
-                                           nbStreams, oldHufTable, bmi2);
+                                           nbStreams, oldHufTable, flags);
     }
 
     /* Build Huffman Tree */
-    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
+    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue, &table->wksps, sizeof(table->wksps), table->CTable, table->count, flags);
     {   size_t const maxBits = HUF_buildCTable_wksp(table->CTable, table->count,
                                             maxSymbolValue, huffLog,
                                             &table->wksps.buildCTable_wksp, sizeof(table->wksps.buildCTable_wksp));
         CHECK_F(maxBits);
         huffLog = (U32)maxBits;
-    }
-    /* Zero unused symbols in CTable, so we can check it for validity */
-    {
-        size_t const ctableSize = HUF_CTABLE_SIZE_ST(maxSymbolValue);
-        size_t const unusedSize = sizeof(table->CTable) - ctableSize * sizeof(HUF_CElt);
-        ZSTD_memset(table->CTable + ctableSize, 0, unusedSize);
+        DEBUGLOG(6, "bit distribution completed (%zu symbols)", showCTableBits(table->CTable + 1, maxSymbolValue+1));
     }
 
     /* Write table description header */
···
             if (oldSize <= hSize + newSize || hSize + 12 >= srcSize) {
                 return HUF_compressCTable_internal(ostart, op, oend,
                                                    src, srcSize,
-                                                   nbStreams, oldHufTable, bmi2);
+                                                   nbStreams, oldHufTable, flags);
         }   }
 
         /* Use the new huffman table */
···
     }
     return HUF_compressCTable_internal(ostart, op, oend,
                                        src, srcSize,
-                                       nbStreams, table->CTable, bmi2);
-}
-
-
-size_t HUF_compress1X_wksp (void* dst, size_t dstSize,
-                      const void* src, size_t srcSize,
-                      unsigned maxSymbolValue, unsigned huffLog,
-                      void* workSpace, size_t wkspSize)
-{
-    return HUF_compress_internal(dst, dstSize, src, srcSize,
-                                 maxSymbolValue, huffLog, HUF_singleStream,
-                                 workSpace, wkspSize,
-                                 NULL, NULL, 0, 0 /*bmi2*/, 0);
+                                       nbStreams, table->CTable, flags);
 }
 
 size_t HUF_compress1X_repeat (void* dst, size_t dstSize,
                       const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned huffLog,
                       void* workSpace, size_t wkspSize,
-                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat,
-                      int bmi2, unsigned suspectUncompressible)
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int flags)
 {
+    DEBUGLOG(5, "HUF_compress1X_repeat (srcSize = %zu)", srcSize);
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_singleStream,
                                  workSpace, wkspSize, hufTable,
-                                 repeat, preferRepeat, bmi2, suspectUncompressible);
-}
-
-/* HUF_compress4X_repeat():
- * compress input using 4 streams.
- * provide workspace to generate compression tables */
-size_t HUF_compress4X_wksp (void* dst, size_t dstSize,
-                      const void* src, size_t srcSize,
-                      unsigned maxSymbolValue, unsigned huffLog,
-                      void* workSpace, size_t wkspSize)
-{
-    return HUF_compress_internal(dst, dstSize, src, srcSize,
-                                 maxSymbolValue, huffLog, HUF_fourStreams,
-                                 workSpace, wkspSize,
-                                 NULL, NULL, 0, 0 /*bmi2*/, 0);
+                                 repeat, flags);
 }
 
 /* HUF_compress4X_repeat():
  * compress input using 4 streams.
  * consider skipping quickly
- * re-use an existing huffman compression table */
+ * reuse an existing huffman compression table */
 size_t HUF_compress4X_repeat (void* dst, size_t dstSize,
                       const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned huffLog,
                       void* workSpace, size_t wkspSize,
-                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2, unsigned suspectUncompressible)
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int flags)
 {
+    DEBUGLOG(5, "HUF_compress4X_repeat (srcSize = %zu)", srcSize);
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_fourStreams,
                                  workSpace, wkspSize,
-                                 hufTable, repeat, preferRepeat, bmi2, suspectUncompressible);
+                                 hufTable, repeat, flags);
 }
-

lib/zstd/compress/zstd_compress.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 /*-*************************************
 *  Dependencies
 ***************************************/
+#include "../common/allocations.h"  /* ZSTD_customMalloc, ZSTD_customCalloc, ZSTD_customFree */
 #include "../common/zstd_deps.h"  /* INT_MAX, ZSTD_memset, ZSTD_memcpy */
 #include "../common/mem.h"
+#include "../common/error_private.h"
 #include "hist.h"           /* HIST_countFast_wksp */
 #define FSE_STATIC_LINKING_ONLY   /* FSE_encodeSymbol */
 #include "../common/fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"
 #include "zstd_compress_internal.h"
 #include "zstd_compress_sequences.h"
···
 #include "zstd_opt.h"
 #include "zstd_ldm.h"
 #include "zstd_compress_superblock.h"
+#include  "../common/bits.h"      /* ZSTD_highbit32, ZSTD_rotateRight_U64 */
 
 /* ***************************************************************
 *  Tuning parameters
···
  * in log format, aka 17 => 1 << 17 == 128Ki positions.
  * This structure is only used in zstd_opt.
  * Since allocation is centralized for all strategies, it has to be known here.
- * The actual (selected) size of the hash table is then stored in ZSTD_matchState_t.hashLog3,
+ * The actual (selected) size of the hash table is then stored in ZSTD_MatchState_t.hashLog3,
  * so that zstd_opt.c doesn't need to know about this constant.
  */
 #ifndef ZSTD_HASHLOG3_MAX
···
 *  Helper functions
 ***************************************/
 /* ZSTD_compressBound()
- * Note that the result from this function is only compatible with the "normal"
- * full-block strategy.
- * When there are a lot of small blocks due to frequent flush in streaming mode
- * the overhead of headers can make the compressed data to be larger than the
- * return value of ZSTD_compressBound().
+ * Note that the result from this function is only valid for
+ * the one-pass compression functions.
+ * When employing the streaming mode,
+ * if flushes are frequently altering the size of blocks,
+ * the overhead from block headers can make the compressed data larger
+ * than the return value of ZSTD_compressBound().
  */
 size_t ZSTD_compressBound(size_t srcSize) {
-    return ZSTD_COMPRESSBOUND(srcSize);
+    size_t const r = ZSTD_COMPRESSBOUND(srcSize);
+    if (r==0) return ERROR(srcSize_wrong);
+    return r;
 }
 
 
···
     ZSTD_dictContentType_e dictContentType; /* The dictContentType the CDict was created with */
     U32* entropyWorkspace; /* entropy workspace of HUF_WORKSPACE_SIZE bytes */
     ZSTD_cwksp workspace;
-    ZSTD_matchState_t matchState;
+    ZSTD_MatchState_t matchState;
     ZSTD_compressedBlockState_t cBlockState;
     ZSTD_customMem customMem;
     U32 dictID;
     int compressionLevel; /* 0 indicates that advanced API was used to select CDict params */
-    ZSTD_paramSwitch_e useRowMatchFinder; /* Indicates whether the CDict was created with params that would use
+    ZSTD_ParamSwitch_e useRowMatchFinder; /* Indicates whether the CDict was created with params that would use
                                            * row-based matchfinder. Unless the cdict is reloaded, we will use
                                            * the same greedy/lazy matchfinder at compression time.
                                            */
···
     ZSTD_cwksp_move(&cctx->workspace, &ws);
     cctx->staticSize = workspaceSize;
 
-    /* statically sized space. entropyWorkspace never moves (but prev/next block swap places) */
-    if (!ZSTD_cwksp_check_available(&cctx->workspace, ENTROPY_WORKSPACE_SIZE + 2 * sizeof(ZSTD_compressedBlockState_t))) return NULL;
+    /* statically sized space. tmpWorkspace never moves (but prev/next block swap places) */
+    if (!ZSTD_cwksp_check_available(&cctx->workspace, TMP_WORKSPACE_SIZE + 2 * sizeof(ZSTD_compressedBlockState_t))) return NULL;
     cctx->blockState.prevCBlock = (ZSTD_compressedBlockState_t*)ZSTD_cwksp_reserve_object(&cctx->workspace, sizeof(ZSTD_compressedBlockState_t));
     cctx->blockState.nextCBlock = (ZSTD_compressedBlockState_t*)ZSTD_cwksp_reserve_object(&cctx->workspace, sizeof(ZSTD_compressedBlockState_t));
-    cctx->entropyWorkspace = (U32*)ZSTD_cwksp_reserve_object(&cctx->workspace, ENTROPY_WORKSPACE_SIZE);
+    cctx->tmpWorkspace = ZSTD_cwksp_reserve_object(&cctx->workspace, TMP_WORKSPACE_SIZE);
+    cctx->tmpWkspSize = TMP_WORKSPACE_SIZE;
     cctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
     return cctx;
 }
···
 
 size_t ZSTD_freeCCtx(ZSTD_CCtx* cctx)
 {
+    DEBUGLOG(3, "ZSTD_freeCCtx (address: %p)", (void*)cctx);
     if (cctx==NULL) return 0;   /* support free on NULL */
     RETURN_ERROR_IF(cctx->staticSize, memory_allocation,
                     "not compatible with static CCtx");
-    {
-        int cctxInWorkspace = ZSTD_cwksp_owns_buffer(&cctx->workspace, cctx);
+    {   int cctxInWorkspace = ZSTD_cwksp_owns_buffer(&cctx->workspace, cctx);
         ZSTD_freeCCtxContent(cctx);
-        if (!cctxInWorkspace) {
-            ZSTD_customFree(cctx, cctx->customMem);
-        }
+        if (!cctxInWorkspace) ZSTD_customFree(cctx, cctx->customMem);
     }
     return 0;
 }
···
 }
 
 /* private API call, for dictBuilder only */
-const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx) { return &(ctx->seqStore); }
+const SeqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx) { return &(ctx->seqStore); }
 
 /* Returns true if the strategy supports using a row based matchfinder */
 static int ZSTD_rowMatchFinderSupported(const ZSTD_strategy strategy) {
···
 /* Returns true if the strategy and useRowMatchFinder mode indicate that we will use the row based matchfinder
  * for this compression.
  */
-static int ZSTD_rowMatchFinderUsed(const ZSTD_strategy strategy, const ZSTD_paramSwitch_e mode) {
+static int ZSTD_rowMatchFinderUsed(const ZSTD_strategy strategy, const ZSTD_ParamSwitch_e mode) {
     assert(mode != ZSTD_ps_auto);
     return ZSTD_rowMatchFinderSupported(strategy) && (mode == ZSTD_ps_enable);
 }
 
 /* Returns row matchfinder usage given an initial mode and cParams */
-static ZSTD_paramSwitch_e ZSTD_resolveRowMatchFinderMode(ZSTD_paramSwitch_e mode,
+static ZSTD_ParamSwitch_e ZSTD_resolveRowMatchFinderMode(ZSTD_ParamSwitch_e mode,
                                                          const ZSTD_compressionParameters* const cParams) {
-#if defined(ZSTD_ARCH_X86_SSE2) || defined(ZSTD_ARCH_ARM_NEON)
-    int const kHasSIMD128 = 1;
-#else
-    int const kHasSIMD128 = 0;
-#endif
+    /* The Linux Kernel does not use SIMD, and 128KB is a very common size, e.g. in BtrFS.
+     * The row match finder is slower for this size without SIMD, so disable it.
+     */
+    const unsigned kWindowLogLowerBound = 17;
     if (mode != ZSTD_ps_auto) return mode; /* if requested enabled, but no SIMD, we still will use row matchfinder */
     mode = ZSTD_ps_disable;
     if (!ZSTD_rowMatchFinderSupported(cParams->strategy)) return mode;
-    if (kHasSIMD128) {
-        if (cParams->windowLog > 14) mode = ZSTD_ps_enable;
-    } else {
-        if (cParams->windowLog > 17) mode = ZSTD_ps_enable;
-    }
+    if (cParams->windowLog > kWindowLogLowerBound) mode = ZSTD_ps_enable;
     return mode;
 }
 
 /* Returns block splitter usage (generally speaking, when using slower/stronger compression modes) */
-static ZSTD_paramSwitch_e ZSTD_resolveBlockSplitterMode(ZSTD_paramSwitch_e mode,
+static ZSTD_ParamSwitch_e ZSTD_resolveBlockSplitterMode(ZSTD_ParamSwitch_e mode,
                                                         const ZSTD_compressionParameters* const cParams) {
     if (mode != ZSTD_ps_auto) return mode;
     return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 17) ? ZSTD_ps_enable : ZSTD_ps_disable;
···
 
 /* Returns 1 if the arguments indicate that we should allocate a chainTable, 0 otherwise */
 static int ZSTD_allocateChainTable(const ZSTD_strategy strategy,
-                                   const ZSTD_paramSwitch_e useRowMatchFinder,
+                                   const ZSTD_ParamSwitch_e useRowMatchFinder,
                                    const U32 forDDSDict) {
     assert(useRowMatchFinder != ZSTD_ps_auto);
     /* We always should allocate a chaintable if we are allocating a matchstate for a DDS dictionary matchstate.
···
     return forDDSDict || ((strategy != ZSTD_fast) && !ZSTD_rowMatchFinderUsed(strategy, useRowMatchFinder));
 }
 
-/* Returns 1 if compression parameters are such that we should
+/* Returns ZSTD_ps_enable if compression parameters are such that we should
  * enable long distance matching (wlog >= 27, strategy >= btopt).
- * Returns 0 otherwise.
+ * Returns ZSTD_ps_disable otherwise.
  */
-static ZSTD_paramSwitch_e ZSTD_resolveEnableLdm(ZSTD_paramSwitch_e mode,
+static ZSTD_ParamSwitch_e ZSTD_resolveEnableLdm(ZSTD_ParamSwitch_e mode,
                                  const ZSTD_compressionParameters* const cParams) {
     if (mode != ZSTD_ps_auto) return mode;
     return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 27) ? ZSTD_ps_enable : ZSTD_ps_disable;
+}
+
+static int ZSTD_resolveExternalSequenceValidation(int mode) {
+    return mode;
+}
+
+/* Resolves maxBlockSize to the default if no value is present. */
+static size_t ZSTD_resolveMaxBlockSize(size_t maxBlockSize) {
+    if (maxBlockSize == 0) {
+        return ZSTD_BLOCKSIZE_MAX;
+    } else {
+        return maxBlockSize;
+    }
+}
+
+static ZSTD_ParamSwitch_e ZSTD_resolveExternalRepcodeSearch(ZSTD_ParamSwitch_e value, int cLevel) {
+    if (value != ZSTD_ps_auto) return value;
+    if (cLevel < 10) {
+        return ZSTD_ps_disable;
+    } else {
+        return ZSTD_ps_enable;
+    }
+}
+
+/* Returns 1 if compression parameters are such that CDict hashtable and chaintable indices are tagged.
+ * If so, the tags need to be removed in ZSTD_resetCCtx_byCopyingCDict. */
+static int ZSTD_CDictIndicesAreTagged(const ZSTD_compressionParameters* const cParams) {
+    return cParams->strategy == ZSTD_fast || cParams->strategy == ZSTD_dfast;
 }
 
 static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams(
···
         assert(cctxParams.ldmParams.hashLog >= cctxParams.ldmParams.bucketSizeLog);
         assert(cctxParams.ldmParams.hashRateLog < 32);
     }
-    cctxParams.useBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams.useBlockSplitter, &cParams);
+    cctxParams.postBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams.postBlockSplitter, &cParams);
     cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);
+    cctxParams.validateSequences = ZSTD_resolveExternalSequenceValidation(cctxParams.validateSequences);
+    cctxParams.maxBlockSize = ZSTD_resolveMaxBlockSize(cctxParams.maxBlockSize);
+    cctxParams.searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(cctxParams.searchForExternalRepcodes,
+                                                                             cctxParams.compressionLevel);
     assert(!ZSTD_checkCParams(cParams));
     return cctxParams;
 }
···
 #define ZSTD_NO_CLEVEL 0
 
 /*
- * Initializes the cctxParams from params and compressionLevel.
+ * Initializes `cctxParams` from `params` and `compressionLevel`.
  * @param compressionLevel If params are derived from a compression level then that compression level, otherwise ZSTD_NO_CLEVEL.
  */
-static void ZSTD_CCtxParams_init_internal(ZSTD_CCtx_params* cctxParams, ZSTD_parameters const* params, int compressionLevel)
+static void
+ZSTD_CCtxParams_init_internal(ZSTD_CCtx_params* cctxParams,
+                        const ZSTD_parameters* params,
+                              int compressionLevel)
 {
     assert(!ZSTD_checkCParams(params->cParams));
     ZSTD_memset(cctxParams, 0, sizeof(*cctxParams));
···
      */
     cctxParams->compressionLevel = compressionLevel;
     cctxParams->useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams->useRowMatchFinder, &params->cParams);
-    cctxParams->useBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams->useBlockSplitter, &params->cParams);
+    cctxParams->postBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams->postBlockSplitter, &params->cParams);
     cctxParams->ldmParams.enableLdm = ZSTD_resolveEnableLdm(cctxParams->ldmParams.enableLdm, &params->cParams);
+    cctxParams->validateSequences = ZSTD_resolveExternalSequenceValidation(cctxParams->validateSequences);
+    cctxParams->maxBlockSize = ZSTD_resolveMaxBlockSize(cctxParams->maxBlockSize);
+    cctxParams->searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(cctxParams->searchForExternalRepcodes, compressionLevel);
     DEBUGLOG(4, "ZSTD_CCtxParams_init_internal: useRowMatchFinder=%d, useBlockSplitter=%d ldm=%d",
-                cctxParams->useRowMatchFinder, cctxParams->useBlockSplitter, cctxParams->ldmParams.enableLdm);
+                cctxParams->useRowMatchFinder, cctxParams->postBlockSplitter, cctxParams->ldmParams.enableLdm);
 }
 
 size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params)
···
 
 /*
  * Sets cctxParams' cParams and fParams from params, but otherwise leaves them alone.
- * @param param Validated zstd parameters.
+ * @param params Validated zstd parameters.
  */
 static void ZSTD_CCtxParams_setZstdParams(
         ZSTD_CCtx_params* cctxParams, const ZSTD_parameters* params)
···
         return bounds;
 
     case ZSTD_c_enableLongDistanceMatching:
-        bounds.lowerBound = 0;
-        bounds.upperBound = 1;
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
         return bounds;
 
     case ZSTD_c_ldmHashLog:
···
         bounds.upperBound = 1;
         return bounds;
 
-    case ZSTD_c_useBlockSplitter:
+    case ZSTD_c_splitAfterSequences:
         bounds.lowerBound = (int)ZSTD_ps_auto;
         bounds.upperBound = (int)ZSTD_ps_disable;
+        return bounds;
+
+    case ZSTD_c_blockSplitterLevel:
+        bounds.lowerBound = 0;
+        bounds.upperBound = ZSTD_BLOCKSPLITTER_LEVEL_MAX;
         return bounds;
 
     case ZSTD_c_useRowMatchFinder:
···
     case ZSTD_c_deterministicRefPrefix:
         bounds.lowerBound = 0;
         bounds.upperBound = 1;
+        return bounds;
+
+    case ZSTD_c_prefetchCDictTables:
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
+        return bounds;
+
+    case ZSTD_c_enableSeqProducerFallback:
+        bounds.lowerBound = 0;
+        bounds.upperBound = 1;
+        return bounds;
+
+    case ZSTD_c_maxBlockSize:
+        bounds.lowerBound = ZSTD_BLOCKSIZE_MAX_MIN;
+        bounds.upperBound = ZSTD_BLOCKSIZE_MAX;
+        return bounds;
+
+    case ZSTD_c_repcodeResolution:
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
         return bounds;
 
     default:
···
     return 0;
 }
 
-#define BOUNDCHECK(cParam, val) { \
-    RETURN_ERROR_IF(!ZSTD_cParam_withinBounds(cParam,val), \
-                    parameter_outOfBound, "Param out of bounds"); \
-}
+#define BOUNDCHECK(cParam, val)                                       \
+    do {                                                              \
+        RETURN_ERROR_IF(!ZSTD_cParam_withinBounds(cParam,val),        \
+                        parameter_outOfBound, "Param out of bounds"); \
+    } while (0)
 
 
 static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param)
···
     case ZSTD_c_minMatch:
     case ZSTD_c_targetLength:
     case ZSTD_c_strategy:
+    case ZSTD_c_blockSplitterLevel:
         return 1;
 
     case ZSTD_c_format:
···
     case ZSTD_c_stableOutBuffer:
     case ZSTD_c_blockDelimiters:
     case ZSTD_c_validateSequences:
-    case ZSTD_c_useBlockSplitter:
+    case ZSTD_c_splitAfterSequences:
     case ZSTD_c_useRowMatchFinder:
     case ZSTD_c_deterministicRefPrefix:
+    case ZSTD_c_prefetchCDictTables:
+    case ZSTD_c_enableSeqProducerFallback:
+    case ZSTD_c_maxBlockSize:
+    case ZSTD_c_repcodeResolution:
     default:
         return 0;
     }
···
         if (ZSTD_isUpdateAuthorized(param)) {
             cctx->cParamsChanged = 1;
         } else {
-            RETURN_ERROR(stage_wrong, "can only set params in ctx init stage");
+            RETURN_ERROR(stage_wrong, "can only set params in cctx init stage");
     }   }
 
     switch(param)
···
     case ZSTD_c_stableOutBuffer:
     case ZSTD_c_blockDelimiters:
     case ZSTD_c_validateSequences:
-    case ZSTD_c_useBlockSplitter:
+    case ZSTD_c_splitAfterSequences:
+    case ZSTD_c_blockSplitterLevel:
     case ZSTD_c_useRowMatchFinder:
     case ZSTD_c_deterministicRefPrefix:
+    case ZSTD_c_prefetchCDictTables:
+    case ZSTD_c_enableSeqProducerFallback:
+    case ZSTD_c_maxBlockSize:
+    case ZSTD_c_repcodeResolution:
         break;
 
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
···
     case ZSTD_c_minMatch :
         if (value!=0)   /* 0 => use default */
             BOUNDCHECK(ZSTD_c_minMatch, value);
-        CCtxParams->cParams.minMatch = value;
+        CCtxParams->cParams.minMatch = (U32)value;
         return CCtxParams->cParams.minMatch;
 
     case ZSTD_c_targetLength :
         BOUNDCHECK(ZSTD_c_targetLength, value);
-        CCtxParams->cParams.targetLength = value;
+        CCtxParams->cParams.targetLength = (U32)value;
         return CCtxParams->cParams.targetLength;
 
     case ZSTD_c_strategy :
···
         /* Content size written in frame header _when known_ (default:1) */
         DEBUGLOG(4, "set content size flag = %u", (value!=0));
         CCtxParams->fParams.contentSizeFlag = value != 0;
-        return CCtxParams->fParams.contentSizeFlag;
+        return (size_t)CCtxParams->fParams.contentSizeFlag;
 
     case ZSTD_c_checksumFlag :
         /* A 32-bits content checksum will be calculated and written at end of frame (default:0) */
         CCtxParams->fParams.checksumFlag = value != 0;
-        return CCtxParams->fParams.checksumFlag;
+        return (size_t)CCtxParams->fParams.checksumFlag;
 
     case ZSTD_c_dictIDFlag : /* When applicable, dictionary's dictID is provided in frame header (default:1) */
         DEBUGLOG(4, "set dictIDFlag = %u", (value!=0));
···
 
     case ZSTD_c_forceMaxWindow :
         CCtxParams->forceWindow = (value != 0);
-        return CCtxParams->forceWindow;
+        return (size_t)CCtxParams->forceWindow;
 
     case ZSTD_c_forceAttachDict : {
         const ZSTD_dictAttachPref_e pref = (ZSTD_dictAttachPref_e)value;
-        BOUNDCHECK(ZSTD_c_forceAttachDict, pref);
+        BOUNDCHECK(ZSTD_c_forceAttachDict, (int)pref);
         CCtxParams->attachDictPref = pref;
         return CCtxParams->attachDictPref;
     }
 
     case ZSTD_c_literalCompressionMode : {
-        const ZSTD_paramSwitch_e lcm = (ZSTD_paramSwitch_e)value;
-        BOUNDCHECK(ZSTD_c_literalCompressionMode, lcm);
+        const ZSTD_ParamSwitch_e lcm = (ZSTD_ParamSwitch_e)value;
+        BOUNDCHECK(ZSTD_c_literalCompressionMode, (int)lcm);
         CCtxParams->literalCompressionMode = lcm;
         return CCtxParams->literalCompressionMode;
     }
···
 
     case ZSTD_c_enableDedicatedDictSearch :
         CCtxParams->enableDedicatedDictSearch = (value!=0);
-        return CCtxParams->enableDedicatedDictSearch;
+        return (size_t)CCtxParams->enableDedicatedDictSearch;
 
     case ZSTD_c_enableLongDistanceMatching :
-        CCtxParams->ldmParams.enableLdm = (ZSTD_paramSwitch_e)value;
+        BOUNDCHECK(ZSTD_c_enableLongDistanceMatching, value);
+        CCtxParams->ldmParams.enableLdm = (ZSTD_ParamSwitch_e)value;
         return CCtxParams->ldmParams.enableLdm;
 
     case ZSTD_c_ldmHashLog :
         if (value!=0)   /* 0 ==> auto */
             BOUNDCHECK(ZSTD_c_ldmHashLog, value);
-        CCtxParams->ldmParams.hashLog = value;
+        CCtxParams->ldmParams.hashLog = (U32)value;
         return CCtxParams->ldmParams.hashLog;
 
     case ZSTD_c_ldmMinMatch :
         if (value!=0)   /* 0 ==> default */
             BOUNDCHECK(ZSTD_c_ldmMinMatch, value);
-        CCtxParams->ldmParams.minMatchLength = value;
+        CCtxParams->ldmParams.minMatchLength = (U32)value;
         return CCtxParams->ldmParams.minMatchLength;
 
     case ZSTD_c_ldmBucketSizeLog :
         if (value!=0)   /* 0 ==> default */
             BOUNDCHECK(ZSTD_c_ldmBucketSizeLog, value);
-        CCtxParams->ldmParams.bucketSizeLog = value;
+        CCtxParams->ldmParams.bucketSizeLog = (U32)value;
         return CCtxParams->ldmParams.bucketSizeLog;
 
     case ZSTD_c_ldmHashRateLog :
         if (value!=0)   /* 0 ==> default */
             BOUNDCHECK(ZSTD_c_ldmHashRateLog, value);
-        CCtxParams->ldmParams.hashRateLog = value;
+        CCtxParams->ldmParams.hashRateLog = (U32)value;
         return CCtxParams->ldmParams.hashRateLog;
 
     case ZSTD_c_targetCBlockSize :
-        if (value!=0)   /* 0 ==> default */
+        if (value!=0) {  /* 0 ==> default */
+            value = MAX(value, ZSTD_TARGETCBLOCKSIZE_MIN);
             BOUNDCHECK(ZSTD_c_targetCBlockSize, value);
-        CCtxParams->targetCBlockSize = value;
+        }
+        CCtxParams->targetCBlockSize = (U32)value;
         return CCtxParams->targetCBlockSize;
 
     case ZSTD_c_srcSizeHint :
         if (value!=0)    /* 0 ==> default */
             BOUNDCHECK(ZSTD_c_srcSizeHint, value);
         CCtxParams->srcSizeHint = value;
-        return CCtxParams->srcSizeHint;
+        return (size_t)CCtxParams->srcSizeHint;
 
     case ZSTD_c_stableInBuffer:
         BOUNDCHECK(ZSTD_c_stableInBuffer, value);
···
 
     case ZSTD_c_blockDelimiters:
         BOUNDCHECK(ZSTD_c_blockDelimiters, value);
-        CCtxParams->blockDelimiters = (ZSTD_sequenceFormat_e)value;
+        CCtxParams->blockDelimiters = (ZSTD_SequenceFormat_e)value;
         return CCtxParams->blockDelimiters;
 
     case ZSTD_c_validateSequences:
         BOUNDCHECK(ZSTD_c_validateSequences, value);
         CCtxParams->validateSequences = value;
-        return CCtxParams->validateSequences;
+        return (size_t)CCtxParams->validateSequences;
 
-    case ZSTD_c_useBlockSplitter:
-        BOUNDCHECK(ZSTD_c_useBlockSplitter, value);
-        CCtxParams->useBlockSplitter = (ZSTD_paramSwitch_e)value;
-        return CCtxParams->useBlockSplitter;
+    case ZSTD_c_splitAfterSequences:
+        BOUNDCHECK(ZSTD_c_splitAfterSequences, value);
+        CCtxParams->postBlockSplitter = (ZSTD_ParamSwitch_e)value;
+        return CCtxParams->postBlockSplitter;
+
+    case ZSTD_c_blockSplitterLevel:
+        BOUNDCHECK(ZSTD_c_blockSplitterLevel, value);
+        CCtxParams->preBlockSplitter_level = value;
+        return (size_t)CCtxParams->preBlockSplitter_level;
 
     case ZSTD_c_useRowMatchFinder:
         BOUNDCHECK(ZSTD_c_useRowMatchFinder, value);
-        CCtxParams->useRowMatchFinder = (ZSTD_paramSwitch_e)value;
+        CCtxParams->useRowMatchFinder = (ZSTD_ParamSwitch_e)value;
         return CCtxParams->useRowMatchFinder;
 
     case ZSTD_c_deterministicRefPrefix:
         BOUNDCHECK(ZSTD_c_deterministicRefPrefix, value);
         CCtxParams->deterministicRefPrefix = !!value;
-        return CCtxParams->deterministicRefPrefix;
+        return (size_t)CCtxParams->deterministicRefPrefix;
+
+    case ZSTD_c_prefetchCDictTables:
+        BOUNDCHECK(ZSTD_c_prefetchCDictTables, value);
+        CCtxParams->prefetchCDictTables = (ZSTD_ParamSwitch_e)value;
+        return CCtxParams->prefetchCDictTables;
+
+    case ZSTD_c_enableSeqProducerFallback:
+        BOUNDCHECK(ZSTD_c_enableSeqProducerFallback, value);
+        CCtxParams->enableMatchFinderFallback = value;
+        return (size_t)CCtxParams->enableMatchFinderFallback;
+
+    case ZSTD_c_maxBlockSize:
+        if (value!=0)    /* 0 ==> default */
+            BOUNDCHECK(ZSTD_c_maxBlockSize, value);
+        assert(value>=0);
+        CCtxParams->maxBlockSize = (size_t)value;
+        return CCtxParams->maxBlockSize;
+
+    case ZSTD_c_repcodeResolution:
+        BOUNDCHECK(ZSTD_c_repcodeResolution, value);
+        CCtxParams->searchForExternalRepcodes = (ZSTD_ParamSwitch_e)value;
+        return CCtxParams->searchForExternalRepcodes;
 
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
     }
···
     switch(param)
     {
     case ZSTD_c_format :
-        *value = CCtxParams->format;
+        *value = (int)CCtxParams->format;
         break;
     case ZSTD_c_compressionLevel :
         *value = CCtxParams->compressionLevel;
···
         *value = (int)CCtxParams->cParams.chainLog;
         break;
     case ZSTD_c_searchLog :
-        *value = CCtxParams->cParams.searchLog;
+        *value = (int)CCtxParams->cParams.searchLog;
         break;
     case ZSTD_c_minMatch :
-        *value = CCtxParams->cParams.minMatch;
+        *value = (int)CCtxParams->cParams.minMatch;
         break;
     case ZSTD_c_targetLength :
-        *value = CCtxParams->cParams.targetLength;
+        *value = (int)CCtxParams->cParams.targetLength;
         break;
     case ZSTD_c_strategy :
-        *value = (unsigned)CCtxParams->cParams.strategy;
+        *value = (int)CCtxParams->cParams.strategy;
         break;
     case ZSTD_c_contentSizeFlag :
         *value = CCtxParams->fParams.contentSizeFlag;
···
         *value = CCtxParams->forceWindow;
         break;
     case ZSTD_c_forceAttachDict :
-        *value = CCtxParams->attachDictPref;
+        *value = (int)CCtxParams->attachDictPref;
         break;
     case ZSTD_c_literalCompressionMode :
-        *value = CCtxParams->literalCompressionMode;
+        *value = (int)CCtxParams->literalCompressionMode;
         break;
     case ZSTD_c_nbWorkers :
         assert(CCtxParams->nbWorkers == 0);
···
         *value = CCtxParams->enableDedicatedDictSearch;
         break;
     case ZSTD_c_enableLongDistanceMatching :
-        *value = CCtxParams->ldmParams.enableLdm;
+        *value = (int)CCtxParams->ldmParams.enableLdm;
         break;
     case ZSTD_c_ldmHashLog :
-        *value = CCtxParams->ldmParams.hashLog;
+        *value = (int)CCtxParams->ldmParams.hashLog;
         break;
     case ZSTD_c_ldmMinMatch :
-        *value = CCtxParams->ldmParams.minMatchLength;
+        *value = (int)CCtxParams->ldmParams.minMatchLength;
         break;
     case ZSTD_c_ldmBucketSizeLog :
-        *value = CCtxParams->ldmParams.bucketSizeLog;
+        *value = (int)CCtxParams->ldmParams.bucketSizeLog;
         break;
     case ZSTD_c_ldmHashRateLog :
-        *value = CCtxParams->ldmParams.hashRateLog;
+        *value = (int)CCtxParams->ldmParams.hashRateLog;
         break;
     case ZSTD_c_targetCBlockSize :
         *value = (int)CCtxParams->targetCBlockSize;
···
     case ZSTD_c_validateSequences :
         *value = (int)CCtxParams->validateSequences;
         break;
-    case ZSTD_c_useBlockSplitter :
-        *value = (int)CCtxParams->useBlockSplitter;
+    case ZSTD_c_splitAfterSequences :
+        *value = (int)CCtxParams->postBlockSplitter;
+        break;
+    case ZSTD_c_blockSplitterLevel :
+        *value = CCtxParams->preBlockSplitter_level;
         break;
     case ZSTD_c_useRowMatchFinder :
         *value = (int)CCtxParams->useRowMatchFinder;
         break;
     case ZSTD_c_deterministicRefPrefix:
         *value = (int)CCtxParams->deterministicRefPrefix;
+        break;
+    case ZSTD_c_prefetchCDictTables:
+        *value = (int)CCtxParams->prefetchCDictTables;
+        break;
+    case ZSTD_c_enableSeqProducerFallback:
+        *value = CCtxParams->enableMatchFinderFallback;
+        break;
+    case ZSTD_c_maxBlockSize:
+        *value = (int)CCtxParams->maxBlockSize;
+        break;
+    case ZSTD_c_repcodeResolution:
+        *value = (int)CCtxParams->searchForExternalRepcodes;
         break;
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
     }
···
     return 0;
 }
 
+size_t ZSTD_CCtx_setCParams(ZSTD_CCtx* cctx, ZSTD_compressionParameters cparams)
+{
+    ZSTD_STATIC_ASSERT(sizeof(cparams) == 7 * 4 /* all params are listed below */);
+    DEBUGLOG(4, "ZSTD_CCtx_setCParams");
+    /* only update if all parameters are valid */
+    FORWARD_IF_ERROR(ZSTD_checkCParams(cparams), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_windowLog, (int)cparams.windowLog), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_chainLog, (int)cparams.chainLog), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_hashLog, (int)cparams.hashLog), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_searchLog, (int)cparams.searchLog), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_minMatch, (int)cparams.minMatch), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_targetLength, (int)cparams.targetLength), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_strategy, (int)cparams.strategy), "");
+    return 0;
+}
+
+size_t ZSTD_CCtx_setFParams(ZSTD_CCtx* cctx, ZSTD_frameParameters fparams)
+{
+    ZSTD_STATIC_ASSERT(sizeof(fparams) == 3 * 4 /* all params are listed below */);
+    DEBUGLOG(4, "ZSTD_CCtx_setFParams");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_contentSizeFlag, fparams.contentSizeFlag != 0), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_checksumFlag, fparams.checksumFlag != 0), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_dictIDFlag, fparams.noDictIDFlag == 0), "");
+    return 0;
+}
+
+size_t ZSTD_CCtx_setParams(ZSTD_CCtx* cctx, ZSTD_parameters params)
+{
+    DEBUGLOG(4, "ZSTD_CCtx_setParams");
+    /* First check cParams, because we want to update all or none. */
+    FORWARD_IF_ERROR(ZSTD_checkCParams(params.cParams), "");
+    /* Next set fParams, because this could fail if the cctx isn't in init stage. */
+    FORWARD_IF_ERROR(ZSTD_CCtx_setFParams(cctx, params.fParams), "");
+    /* Finally set cParams, which should succeed. */
+    FORWARD_IF_ERROR(ZSTD_CCtx_setCParams(cctx, params.cParams), "");
+    return 0;
+}
+
 size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long long pledgedSrcSize)
 {
-    DEBUGLOG(4, "ZSTD_CCtx_setPledgedSrcSize to %u bytes", (U32)pledgedSrcSize);
+    DEBUGLOG(4, "ZSTD_CCtx_setPledgedSrcSize to %llu bytes", pledgedSrcSize);
     RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                     "Can't set pledgedSrcSize when not in init stage.");
     cctx->pledgedSrcSizePlusOne = pledgedSrcSize+1;
···
         ZSTD_compressionParameters* cParams);
 
 /*
- * Initializes the local dict using the requested parameters.
- * NOTE: This does not use the pledged src size, because it may be used for more
- * than one compression.
+ * Initializes the local dictionary using requested parameters.
+ * NOTE: Initialization does not employ the pledged src size,
+ * because the dictionary may be used for multiple compressions.
  */
 static size_t ZSTD_initLocalDict(ZSTD_CCtx* cctx)
 {
···
         return 0;
     }
     if (dl->cdict != NULL) {
-        assert(cctx->cdict == dl->cdict);
         /* Local dictionary already initialized. */
+        assert(cctx->cdict == dl->cdict);
         return 0;
     }
     assert(dl->dictSize > 0);
···
 }
 
 size_t ZSTD_CCtx_loadDictionary_advanced(
-        ZSTD_CCtx* cctx, const void* dict, size_t dictSize,
-        ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType)
+        ZSTD_CCtx* cctx,
+        const void* dict, size_t dictSize,
+        ZSTD_dictLoadMethod_e dictLoadMethod,
+        ZSTD_dictContentType_e dictContentType)
 {
-    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
-                    "Can't load a dictionary when ctx is not in init stage.");
     DEBUGLOG(4, "ZSTD_CCtx_loadDictionary_advanced (size: %u)", (U32)dictSize);
-    ZSTD_clearAllDicts(cctx);  /* in case one already exists */
-    if (dict == NULL || dictSize == 0)  /* no dictionary mode */
+    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
+                    "Can't load a dictionary when cctx is not in init stage.");
+    ZSTD_clearAllDicts(cctx);  /* erase any previously set dictionary */
+    if (dict == NULL || dictSize == 0)  /* no dictionary */
         return 0;
     if (dictLoadMethod == ZSTD_dlm_byRef) {
         cctx->localDict.dict = dict;
     } else {
+        /* copy dictionary content inside CCtx to own its lifetime */
         void* dictBuffer;
         RETURN_ERROR_IF(cctx->staticSize, memory_allocation,
-                        "no malloc for static CCtx");
+                        "static CCtx can't allocate for an internal copy of dictionary");
         dictBuffer = ZSTD_customMalloc(dictSize, cctx->customMem);
-        RETURN_ERROR_IF(!dictBuffer, memory_allocation, "NULL pointer!");
+        RETURN_ERROR_IF(dictBuffer==NULL, memory_allocation,
+                        "allocation failed for dictionary content");
         ZSTD_memcpy(dictBuffer, dict, dictSize);
-        cctx->localDict.dictBuffer = dictBuffer;
-        cctx->localDict.dict = dictBuffer;
+        cctx->localDict.dictBuffer = dictBuffer;  /* owned ptr to free */
+        cctx->localDict.dict = dictBuffer;        /* read-only reference */
     }
     cctx->localDict.dictSize = dictSize;
     cctx->localDict.dictContentType = dictContentType;
···
     if ( (reset == ZSTD_reset_parameters)
       || (reset == ZSTD_reset_session_and_parameters) ) {
         RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
-                        "Can't reset parameters only when not in init stage.");
+                        "Reset parameters is only possible during init stage.");
         ZSTD_clearAllDicts(cctx);
         return ZSTD_CCtxParams_reset(&cctx->requestedParams);
     }
···
     BOUNDCHECK(ZSTD_c_searchLog, (int)cParams.searchLog);
     BOUNDCHECK(ZSTD_c_minMatch,  (int)cParams.minMatch);
     BOUNDCHECK(ZSTD_c_targetLength,(int)cParams.targetLength);
-    BOUNDCHECK(ZSTD_c_strategy,  cParams.strategy);
+    BOUNDCHECK(ZSTD_c_strategy,  (int)cParams.strategy);
     return 0;
 }
 
···
 static ZSTD_compressionParameters
 ZSTD_clampCParams(ZSTD_compressionParameters cParams)
 {
-#   define CLAMP_TYPE(cParam, val, type) {                                \
-        ZSTD_bounds const bounds = ZSTD_cParam_getBounds(cParam);         \
-        if ((int)val<bounds.lowerBound) val=(type)bounds.lowerBound;      \
-        else if ((int)val>bounds.upperBound) val=(type)bounds.upperBound; \
-    }
+#   define CLAMP_TYPE(cParam, val, type)                                      \
+        do {                                                                  \
+            ZSTD_bounds const bounds = ZSTD_cParam_getBounds(cParam);         \
+            if ((int)val<bounds.lowerBound) val=(type)bounds.lowerBound;      \
+            else if ((int)val>bounds.upperBound) val=(type)bounds.upperBound; \
+        } while (0)
 #   define CLAMP(cParam, val) CLAMP_TYPE(cParam, val, unsigned)
     CLAMP(ZSTD_c_windowLog, cParams.windowLog);
     CLAMP(ZSTD_c_chainLog,  cParams.chainLog);
···
  *  optimize `cPar` for a specified input (`srcSize` and `dictSize`).
  *  mostly downsize to reduce memory consumption and initialization latency.
  * `srcSize` can be ZSTD_CONTENTSIZE_UNKNOWN when not known.
- * `mode` is the mode for parameter adjustment. See docs for `ZSTD_cParamMode_e`.
+ * `mode` is the mode for parameter adjustment. See docs for `ZSTD_CParamMode_e`.
  *  note : `srcSize==0` means 0!
  *  condition : cPar is presumed validated (can be checked using ZSTD_checkCParams()). */
 static ZSTD_compressionParameters
 ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar,
                             unsigned long long srcSize,
                             size_t dictSize,
-                            ZSTD_cParamMode_e mode)
+                            ZSTD_CParamMode_e mode,
+                            ZSTD_ParamSwitch_e useRowMatchFinder)
 {
     const U64 minSrcSize = 513; /* (1<<9) + 1 */
     const U64 maxWindowResize = 1ULL << (ZSTD_WINDOWLOG_MAX-1);
     assert(ZSTD_checkCParams(cPar)==0);
+
+    /* Cascade the selected strategy down to the next-highest one built into
+     * this binary. */
+#ifdef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
+    if (cPar.strategy == ZSTD_btultra2) {
+        cPar.strategy = ZSTD_btultra;
+    }
+    if (cPar.strategy == ZSTD_btultra) {
+        cPar.strategy = ZSTD_btopt;
+    }
+#endif
+#ifdef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
+    if (cPar.strategy == ZSTD_btopt) {
+        cPar.strategy = ZSTD_btlazy2;
+    }
+#endif
+#ifdef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
+    if (cPar.strategy == ZSTD_btlazy2) {
+        cPar.strategy = ZSTD_lazy2;
+    }
+#endif
+#ifdef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
+    if (cPar.strategy == ZSTD_lazy2) {
+        cPar.strategy = ZSTD_lazy;
+    }
+#endif
+#ifdef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
+    if (cPar.strategy == ZSTD_lazy) {
+        cPar.strategy = ZSTD_greedy;
+    }
+#endif
+#ifdef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
+    if (cPar.strategy == ZSTD_greedy) {
+        cPar.strategy = ZSTD_dfast;
+    }
+#endif
+#ifdef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR
+    if (cPar.strategy == ZSTD_dfast) {
+        cPar.strategy = ZSTD_fast;
+        cPar.targetLength = 0;
+    }
+#endif
 
     switch (mode) {
     case ZSTD_cpm_unknown:
···
     }
 
     /* resize windowLog if input is small enough, to use less memory */
-    if ( (srcSize < maxWindowResize)
-      && (dictSize < maxWindowResize) )  {
+    if ( (srcSize <= maxWindowResize)
+      && (dictSize <= maxWindowResize) )  {
         U32 const tSize = (U32)(srcSize + dictSize);
         static U32 const hashSizeMin = 1 << ZSTD_HASHLOG_MIN;
         U32 const srcLog = (tSize < hashSizeMin) ? ZSTD_HASHLOG_MIN :
···
     if (cPar.windowLog < ZSTD_WINDOWLOG_ABSOLUTEMIN)
         cPar.windowLog = ZSTD_WINDOWLOG_ABSOLUTEMIN;  /* minimum wlog required for valid frame header */
 
+    /* We can't use more than 32 bits of hash in total, so that means that we require:
+     * (hashLog + 8) <= 32 && (chainLog + 8) <= 32
+     */
+    if (mode == ZSTD_cpm_createCDict && ZSTD_CDictIndicesAreTagged(&cPar)) {
+        U32 const maxShortCacheHashLog = 32 - ZSTD_SHORT_CACHE_TAG_BITS;
+        if (cPar.hashLog > maxShortCacheHashLog) {
+            cPar.hashLog = maxShortCacheHashLog;
+        }
+        if (cPar.chainLog > maxShortCacheHashLog) {
+            cPar.chainLog = maxShortCacheHashLog;
+        }
+    }
+
+
+    /* At this point, we aren't 100% sure if we are using the row match finder.
+     * Unless it is explicitly disabled, conservatively assume that it is enabled.
+     * In this case it will only be disabled for small sources, so shrinking the
+     * hash log a little bit shouldn't result in any ratio loss.
+     */
+    if (useRowMatchFinder == ZSTD_ps_auto)
+        useRowMatchFinder = ZSTD_ps_enable;
+
+    /* We can't hash more than 32-bits in total. So that means that we require:
+     * (hashLog - rowLog + 8) <= 32
+     */
+    if (ZSTD_rowMatchFinderUsed(cPar.strategy, useRowMatchFinder)) {
+        /* Switch to 32-entry rows if searchLog is 5 (or more) */
+        U32 const rowLog = BOUNDED(4, cPar.searchLog, 6);
+        U32 const maxRowHashLog = 32 - ZSTD_ROW_HASH_TAG_BITS;
+        U32 const maxHashLog = maxRowHashLog + rowLog;
+        assert(cPar.hashLog >= rowLog);
+        if (cPar.hashLog > maxHashLog) {
+            cPar.hashLog = maxHashLog;
+        }
+    }
+
     return cPar;
 }
 
···
 {
     cPar = ZSTD_clampCParams(cPar);   /* resulting cPar is necessarily valid (all parameters within range) */
     if (srcSize == 0) srcSize = ZSTD_CONTENTSIZE_UNKNOWN;
-    return ZSTD_adjustCParams_internal(cPar, srcSize, dictSize, ZSTD_cpm_unknown);
+    return ZSTD_adjustCParams_internal(cPar, srcSize, dictSize, ZSTD_cpm_unknown, ZSTD_ps_auto);
 }
 
-static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode);
-static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode);
+static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode);
+static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode);
 
 static void ZSTD_overrideCParams(
               ZSTD_compressionParameters* cParams,
···
 }
 
 ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
-        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode)
+        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode)
 {
     ZSTD_compressionParameters cParams;
     if (srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN && CCtxParams->srcSizeHint > 0) {
-      srcSizeHint = CCtxParams->srcSizeHint;
+        assert(CCtxParams->srcSizeHint>=0);
+        srcSizeHint = (U64)CCtxParams->srcSizeHint;
     }
     cParams = ZSTD_getCParams_internal(CCtxParams->compressionLevel, srcSizeHint, dictSize, mode);
     if (CCtxParams->ldmParams.enableLdm == ZSTD_ps_enable) cParams.windowLog = ZSTD_LDM_DEFAULT_WINDOW_LOG;
     ZSTD_overrideCParams(&cParams, &CCtxParams->cParams);
     assert(!ZSTD_checkCParams(cParams));
     /* srcSizeHint == 0 means 0 */
-    return ZSTD_adjustCParams_internal(cParams, srcSizeHint, dictSize, mode);
+    return ZSTD_adjustCParams_internal(cParams, srcSizeHint, dictSize, mode, CCtxParams->useRowMatchFinder);
 }
 
 static size_t
 ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams,
-                       const ZSTD_paramSwitch_e useRowMatchFinder,
-                       const U32 enableDedicatedDictSearch,
+                       const ZSTD_ParamSwitch_e useRowMatchFinder,
+                       const int enableDedicatedDictSearch,
                        const U32 forCCtx)
 {
     /* chain table size should be 0 for fast or row-hash strategies */
···
                             + hSize * sizeof(U32)
                             + h3Size * sizeof(U32);
     size_t const optPotentialSpace =
-        ZSTD_cwksp_aligned_alloc_size((MaxML+1) * sizeof(U32))
-      + ZSTD_cwksp_aligned_alloc_size((MaxLL+1) * sizeof(U32))
-      + ZSTD_cwksp_aligned_alloc_size((MaxOff+1) * sizeof(U32))
-      + ZSTD_cwksp_aligned_alloc_size((1<<Litbits) * sizeof(U32))
-      + ZSTD_cwksp_aligned_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_match_t))
-      + ZSTD_cwksp_aligned_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t));
+        ZSTD_cwksp_aligned64_alloc_size((MaxML+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned64_alloc_size((MaxLL+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned64_alloc_size((MaxOff+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned64_alloc_size((1<<Litbits) * sizeof(U32))
+      + ZSTD_cwksp_aligned64_alloc_size(ZSTD_OPT_SIZE * sizeof(ZSTD_match_t))
+      + ZSTD_cwksp_aligned64_alloc_size(ZSTD_OPT_SIZE * sizeof(ZSTD_optimal_t));
     size_t const lazyAdditionalSpace = ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)
-                                            ? ZSTD_cwksp_aligned_alloc_size(hSize*sizeof(U16))
+                                            ? ZSTD_cwksp_aligned64_alloc_size(hSize)
                                             : 0;
     size_t const optSpace = (forCCtx && (cParams->strategy >= ZSTD_btopt))
                                 ? optPotentialSpace
···
     return tableSpace + optSpace + slackSpace + lazyAdditionalSpace;
 }
 
+/* Helper function for calculating memory requirements.
+ * Gives a tighter bound than ZSTD_sequenceBound() by taking minMatch into account. */
+static size_t ZSTD_maxNbSeq(size_t blockSize, unsigned minMatch, int useSequenceProducer) {
+    U32 const divider = (minMatch==3 || useSequenceProducer) ? 3 : 4;
+    return blockSize / divider;
+}
+
 static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
         const ZSTD_compressionParameters* cParams,
         const ldmParams_t* ldmParams,
         const int isStatic,
-        const ZSTD_paramSwitch_e useRowMatchFinder,
+        const ZSTD_ParamSwitch_e useRowMatchFinder,
         const size_t buffInSize,
         const size_t buffOutSize,
-        const U64 pledgedSrcSize)
+        const U64 pledgedSrcSize,
+        int useSequenceProducer,
+        size_t maxBlockSize)
 {
     size_t const windowSize = (size_t) BOUNDED(1ULL, 1ULL << cParams->windowLog, pledgedSrcSize);
-    size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, windowSize);
-    U32    const divider = (cParams->minMatch==3) ? 3 : 4;
-    size_t const maxNbSeq = blockSize / divider;
+    size_t const blockSize = MIN(ZSTD_resolveMaxBlockSize(maxBlockSize), windowSize);
+    size_t const maxNbSeq = ZSTD_maxNbSeq(blockSize, cParams->minMatch, useSequenceProducer);
     size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize)
-                            + ZSTD_cwksp_aligned_alloc_size(maxNbSeq * sizeof(seqDef))
+                            + ZSTD_cwksp_aligned64_alloc_size(maxNbSeq * sizeof(SeqDef))
                             + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE));
-    size_t const entropySpace = ZSTD_cwksp_alloc_size(ENTROPY_WORKSPACE_SIZE);
+    size_t const tmpWorkSpace = ZSTD_cwksp_alloc_size(TMP_WORKSPACE_SIZE);
     size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t));
     size_t const matchStateSize = ZSTD_sizeof_matchState(cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 0, /* forCCtx */ 1);
 
     size_t const ldmSpace = ZSTD_ldm_getTableSize(*ldmParams);
     size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(*ldmParams, blockSize);
     size_t const ldmSeqSpace = ldmParams->enableLdm == ZSTD_ps_enable ?
-        ZSTD_cwksp_aligned_alloc_size(maxNbLdmSeq * sizeof(rawSeq)) : 0;
+        ZSTD_cwksp_aligned64_alloc_size(maxNbLdmSeq * sizeof(rawSeq)) : 0;
 
 
     size_t const bufferSpace = ZSTD_cwksp_alloc_size(buffInSize)
···
 
     size_t const cctxSpace = isStatic ? ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx)) : 0;
 
+    size_t const maxNbExternalSeq = ZSTD_sequenceBound(blockSize);
+    size_t const externalSeqSpace = useSequenceProducer
+        ? ZSTD_cwksp_aligned64_alloc_size(maxNbExternalSeq * sizeof(ZSTD_Sequence))
+        : 0;
+
     size_t const neededSpace =
         cctxSpace +
-        entropySpace +
+        tmpWorkSpace +
         blockStateSpace +
         ldmSpace +
         ldmSeqSpace +
         matchStateSize +
         tokenSpace +
-        bufferSpace;
+        bufferSpace +
+        externalSeqSpace;
 
     DEBUGLOG(5, "estimate workspace : %u", (U32)neededSpace);
     return neededSpace;
···
 {
     ZSTD_compressionParameters const cParams =
                 ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
-    ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder,
+    ZSTD_ParamSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder,
                                                                                &cParams);
 
     RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only.");
···
      * be needed. However, we still allocate two 0-sized buffers, which can
      * take space under ASAN. */
     return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
-        &cParams, &params->ldmParams, 1, useRowMatchFinder, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN);
+        &cParams, &params->ldmParams, 1, useRowMatchFinder, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN, ZSTD_hasExtSeqProd(params), params->maxBlockSize);
 }
 
 size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams)
···
     RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only.");
     {   ZSTD_compressionParameters const cParams =
                 ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
-        size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << cParams.windowLog);
+        size_t const blockSize = MIN(ZSTD_resolveMaxBlockSize(params->maxBlockSize), (size_t)1 << cParams.windowLog);
         size_t const inBuffSize = (params->inBufferMode == ZSTD_bm_buffered)
                 ? ((size_t)1 << cParams.windowLog) + blockSize
                 : 0;
         size_t const outBuffSize = (params->outBufferMode == ZSTD_bm_buffered)
                 ? ZSTD_compressBound(blockSize) + 1
                 : 0;
-        ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, &params->cParams);
+        ZSTD_ParamSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, &params->cParams);
 
         return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
             &cParams, &params->ldmParams, 1, useRowMatchFinder, inBuffSize, outBuffSize,
-            ZSTD_CONTENTSIZE_UNKNOWN);
+            ZSTD_CONTENTSIZE_UNKNOWN, ZSTD_hasExtSeqProd(params), params->maxBlockSize);
     }
 }
 
···
  *  Invalidate all the matches in the match finder tables.
  *  Requires nextSrc and base to be set (can be NULL).
  */
-static void ZSTD_invalidateMatchState(ZSTD_matchState_t* ms)
+static void ZSTD_invalidateMatchState(ZSTD_MatchState_t* ms)
 {
     ZSTD_window_clear(&ms->window);
 
···
     ZSTD_resetTarget_CCtx
 } ZSTD_resetTarget_e;
 
+/* Mixes bits in a 64 bits in a value, based on XXH3_rrmxmx */
+static U64 ZSTD_bitmix(U64 val, U64 len) {
+    val ^= ZSTD_rotateRight_U64(val, 49) ^ ZSTD_rotateRight_U64(val, 24);
+    val *= 0x9FB21C651E98DF25ULL;
+    val ^= (val >> 35) + len ;
+    val *= 0x9FB21C651E98DF25ULL;
+    return val ^ (val >> 28);
+}
+
+/* Mixes in the hashSalt and hashSaltEntropy to create a new hashSalt */
+static void ZSTD_advanceHashSalt(ZSTD_MatchState_t* ms) {
+    ms->hashSalt = ZSTD_bitmix(ms->hashSalt, 8) ^ ZSTD_bitmix((U64) ms->hashSaltEntropy, 4);
+}
 
 static size_t
-ZSTD_reset_matchState(ZSTD_matchState_t* ms,
+ZSTD_reset_matchState(ZSTD_MatchState_t* ms,
                       ZSTD_cwksp* ws,
                 const ZSTD_compressionParameters* cParams,
-                const ZSTD_paramSwitch_e useRowMatchFinder,
+                const ZSTD_ParamSwitch_e useRowMatchFinder,
                 const ZSTD_compResetPolicy_e crp,
                 const ZSTD_indexResetPolicy_e forceResetIndex,
                 const ZSTD_resetTarget_e forWho)
···
     }
 
     ms->hashLog3 = hashLog3;
+    ms->lazySkipping = 0;
 
     ZSTD_invalidateMatchState(ms);
 
···
         ZSTD_cwksp_clean_tables(ws);
     }
 
-    /* opt parser space */
-    if ((forWho == ZSTD_resetTarget_CCtx) && (cParams->strategy >= ZSTD_btopt)) {
-        DEBUGLOG(4, "reserving optimal parser space");
-        ms->opt.litFreq = (unsigned*)ZSTD_cwksp_reserve_aligned(ws, (1<<Litbits) * sizeof(unsigned));
-        ms->opt.litLengthFreq = (unsigned*)ZSTD_cwksp_reserve_aligned(ws, (MaxLL+1) * sizeof(unsigned));
-        ms->opt.matchLengthFreq = (unsigned*)ZSTD_cwksp_reserve_aligned(ws, (MaxML+1) * sizeof(unsigned));
-        ms->opt.offCodeFreq = (unsigned*)ZSTD_cwksp_reserve_aligned(ws, (MaxOff+1) * sizeof(unsigned));
-        ms->opt.matchTable = (ZSTD_match_t*)ZSTD_cwksp_reserve_aligned(ws, (ZSTD_OPT_NUM+1) * sizeof(ZSTD_match_t));
-        ms->opt.priceTable = (ZSTD_optimal_t*)ZSTD_cwksp_reserve_aligned(ws, (ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t));
-    }
-
     if (ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)) {
-        {   /* Row match finder needs an additional table of hashes ("tags") */
-            size_t const tagTableSize = hSize*sizeof(U16);
-            ms->tagTable = (U16*)ZSTD_cwksp_reserve_aligned(ws, tagTableSize);
-            if (ms->tagTable) ZSTD_memset(ms->tagTable, 0, tagTableSize);
+        /* Row match finder needs an additional table of hashes ("tags") */
+        size_t const tagTableSize = hSize;
+        /* We want to generate a new salt in case we reset a Cctx, but we always want to use
+         * 0 when we reset a Cdict */
+        if(forWho == ZSTD_resetTarget_CCtx) {
+            ms->tagTable = (BYTE*) ZSTD_cwksp_reserve_aligned_init_once(ws, tagTableSize);
+            ZSTD_advanceHashSalt(ms);
+        } else {
+            /* When we are not salting we want to always memset the memory */
+            ms->tagTable = (BYTE*) ZSTD_cwksp_reserve_aligned64(ws, tagTableSize);
+            ZSTD_memset(ms->tagTable, 0, tagTableSize);
+            ms->hashSalt = 0;
         }
         {   /* Switch to 32-entry rows if searchLog is 5 (or more) */
             U32 const rowLog = BOUNDED(4, cParams->searchLog, 6);
             assert(cParams->hashLog >= rowLog);
             ms->rowHashLog = cParams->hashLog - rowLog;
         }
+    }
+
+    /* opt parser space */
+    if ((forWho == ZSTD_resetTarget_CCtx) && (cParams->strategy >= ZSTD_btopt)) {
+        DEBUGLOG(4, "reserving optimal parser space");
+        ms->opt.litFreq = (unsigned*)ZSTD_cwksp_reserve_aligned64(ws, (1<<Litbits) * sizeof(unsigned));
+        ms->opt.litLengthFreq = (unsigned*)ZSTD_cwksp_reserve_aligned64(ws, (MaxLL+1) * sizeof(unsigned));
+        ms->opt.matchLengthFreq = (unsigned*)ZSTD_cwksp_reserve_aligned64(ws, (MaxML+1) * sizeof(unsigned));
+        ms->opt.offCodeFreq = (unsigned*)ZSTD_cwksp_reserve_aligned64(ws, (MaxOff+1) * sizeof(unsigned));
+        ms->opt.matchTable = (ZSTD_match_t*)ZSTD_cwksp_reserve_aligned64(ws, ZSTD_OPT_SIZE * sizeof(ZSTD_match_t));
+        ms->opt.priceTable = (ZSTD_optimal_t*)ZSTD_cwksp_reserve_aligned64(ws, ZSTD_OPT_SIZE * sizeof(ZSTD_optimal_t));
     }
 
     ms->cParams = *cParams;
···
 {
     ZSTD_cwksp* const ws = &zc->workspace;
     DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u, useRowMatchFinder=%d useBlockSplitter=%d",
-                (U32)pledgedSrcSize, params->cParams.windowLog, (int)params->useRowMatchFinder, (int)params->useBlockSplitter);
+                (U32)pledgedSrcSize, params->cParams.windowLog, (int)params->useRowMatchFinder, (int)params->postBlockSplitter);
     assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams)));
 
     zc->isFirstBlock = 1;
···
     params = &zc->appliedParams;
 
     assert(params->useRowMatchFinder != ZSTD_ps_auto);
-    assert(params->useBlockSplitter != ZSTD_ps_auto);
+    assert(params->postBlockSplitter != ZSTD_ps_auto);
     assert(params->ldmParams.enableLdm != ZSTD_ps_auto);
+    assert(params->maxBlockSize != 0);
     if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
         /* Adjust long distance matching parameters */
         ZSTD_ldm_adjustParameters(&zc->appliedParams.ldmParams, &params->cParams);
···
     }
 
     {   size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params->cParams.windowLog), pledgedSrcSize));
-        size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, windowSize);
-        U32    const divider = (params->cParams.minMatch==3) ? 3 : 4;
-        size_t const maxNbSeq = blockSize / divider;
+        size_t const blockSize = MIN(params->maxBlockSize, windowSize);
+        size_t const maxNbSeq = ZSTD_maxNbSeq(blockSize, params->cParams.minMatch, ZSTD_hasExtSeqProd(params));
         size_t const buffOutSize = (zbuff == ZSTDb_buffered && params->outBufferMode == ZSTD_bm_buffered)
                 ? ZSTD_compressBound(blockSize) + 1
                 : 0;
···
         size_t const neededSpace =
             ZSTD_estimateCCtxSize_usingCCtxParams_internal(
                 &params->cParams, &params->ldmParams, zc->staticSize != 0, params->useRowMatchFinder,
-                buffInSize, buffOutSize, pledgedSrcSize);
-        int resizeWorkspace;
+                buffInSize, buffOutSize, pledgedSrcSize, ZSTD_hasExtSeqProd(params), params->maxBlockSize);
 
         FORWARD_IF_ERROR(neededSpace, "cctx size estimate failed!");
 
···
         {   /* Check if workspace is large enough, alloc a new one if needed */
             int const workspaceTooSmall = ZSTD_cwksp_sizeof(ws) < neededSpace;
             int const workspaceWasteful = ZSTD_cwksp_check_wasteful(ws, neededSpace);
-            resizeWorkspace = workspaceTooSmall || workspaceWasteful;
+            int resizeWorkspace = workspaceTooSmall || workspaceWasteful;
             DEBUGLOG(4, "Need %zu B workspace", neededSpace);
             DEBUGLOG(4, "windowSize: %zu - blockSize: %zu", windowSize, blockSize);
 
···
 
                 DEBUGLOG(5, "reserving object space");
                 /* Statically sized space.
-                 * entropyWorkspace never moves,
+                 * tmpWorkspace never moves,
                  * though prev/next block swap places */
                 assert(ZSTD_cwksp_check_available(ws, 2 * sizeof(ZSTD_compressedBlockState_t)));
                 zc->blockState.prevCBlock = (ZSTD_compressedBlockState_t*) ZSTD_cwksp_reserve_object(ws, sizeof(ZSTD_compressedBlockState_t));
                 RETURN_ERROR_IF(zc->blockState.prevCBlock == NULL, memory_allocation, "couldn't allocate prevCBlock");
                 zc->blockState.nextCBlock = (ZSTD_compressedBlockState_t*) ZSTD_cwksp_reserve_object(ws, sizeof(ZSTD_compressedBlockState_t));
                 RETURN_ERROR_IF(zc->blockState.nextCBlock == NULL, memory_allocation, "couldn't allocate nextCBlock");
-                zc->entropyWorkspace = (U32*) ZSTD_cwksp_reserve_object(ws, ENTROPY_WORKSPACE_SIZE);
-                RETURN_ERROR_IF(zc->entropyWorkspace == NULL, memory_allocation, "couldn't allocate entropyWorkspace");
+                zc->tmpWorkspace = ZSTD_cwksp_reserve_object(ws, TMP_WORKSPACE_SIZE);
+                RETURN_ERROR_IF(zc->tmpWorkspace == NULL, memory_allocation, "couldn't allocate tmpWorkspace");
+                zc->tmpWkspSize = TMP_WORKSPACE_SIZE;
         }   }
 
         ZSTD_cwksp_clear(ws);
 
         /* init params */
         zc->blockState.matchState.cParams = params->cParams;
+        zc->blockState.matchState.prefetchCDictTables = params->prefetchCDictTables == ZSTD_ps_enable;
         zc->pledgedSrcSizePlusOne = pledgedSrcSize+1;
         zc->consumedSrcSize = 0;
         zc->producedCSize = 0;
···
             zc->appliedParams.fParams.contentSizeFlag = 0;
         DEBUGLOG(4, "pledged content size : %u ; flag : %u",
             (unsigned)pledgedSrcSize, zc->appliedParams.fParams.contentSizeFlag);
-        zc->blockSize = blockSize;
+        zc->blockSizeMax = blockSize;
 
         xxh64_reset(&zc->xxhState, 0);
         zc->stage = ZSTDcs_init;
···
 
         ZSTD_reset_compressedBlockState(zc->blockState.prevCBlock);
 
+        FORWARD_IF_ERROR(ZSTD_reset_matchState(
+                &zc->blockState.matchState,
+                ws,
+                &params->cParams,
+                params->useRowMatchFinder,
+                crp,
+                needsIndexReset,
+                ZSTD_resetTarget_CCtx), "");
+
+        zc->seqStore.sequencesStart = (SeqDef*)ZSTD_cwksp_reserve_aligned64(ws, maxNbSeq * sizeof(SeqDef));
+
+        /* ldm hash table */
+        if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
+            /* TODO: avoid memset? */
+            size_t const ldmHSize = ((size_t)1) << params->ldmParams.hashLog;
+            zc->ldmState.hashTable = (ldmEntry_t*)ZSTD_cwksp_reserve_aligned64(ws, ldmHSize * sizeof(ldmEntry_t));
+            ZSTD_memset(zc->ldmState.hashTable, 0, ldmHSize * sizeof(ldmEntry_t));
+            zc->ldmSequences = (rawSeq*)ZSTD_cwksp_reserve_aligned64(ws, maxNbLdmSeq * sizeof(rawSeq));
+            zc->maxNbLdmSequences = maxNbLdmSeq;
+
+            ZSTD_window_init(&zc->ldmState.window);
+            zc->ldmState.loadedDictEnd = 0;
+        }
+
+        /* reserve space for block-level external sequences */
+        if (ZSTD_hasExtSeqProd(params)) {
+            size_t const maxNbExternalSeq = ZSTD_sequenceBound(blockSize);
+            zc->extSeqBufCapacity = maxNbExternalSeq;
+            zc->extSeqBuf =
+                (ZSTD_Sequence*)ZSTD_cwksp_reserve_aligned64(ws, maxNbExternalSeq * sizeof(ZSTD_Sequence));
+        }
+
+        /* buffers */
+
         /* ZSTD_wildcopy() is used to copy into the literals buffer,
          * so we have to oversize the buffer by WILDCOPY_OVERLENGTH bytes.
          */
         zc->seqStore.litStart = ZSTD_cwksp_reserve_buffer(ws, blockSize + WILDCOPY_OVERLENGTH);
         zc->seqStore.maxNbLit = blockSize;
 
-        /* buffers */
         zc->bufferedPolicy = zbuff;
         zc->inBuffSize = buffInSize;
         zc->inBuff = (char*)ZSTD_cwksp_reserve_buffer(ws, buffInSize);
···
         zc->seqStore.llCode = ZSTD_cwksp_reserve_buffer(ws, maxNbSeq * sizeof(BYTE));
         zc->seqStore.mlCode = ZSTD_cwksp_reserve_buffer(ws, maxNbSeq * sizeof(BYTE));
         zc->seqStore.ofCode = ZSTD_cwksp_reserve_buffer(ws, maxNbSeq * sizeof(BYTE));
-        zc->seqStore.sequencesStart = (seqDef*)ZSTD_cwksp_reserve_aligned(ws, maxNbSeq * sizeof(seqDef));
-
-        FORWARD_IF_ERROR(ZSTD_reset_matchState(
-            &zc->blockState.matchState,
-            ws,
-            &params->cParams,
-            params->useRowMatchFinder,
-            crp,
-            needsIndexReset,
-            ZSTD_resetTarget_CCtx), "");
-
-        /* ldm hash table */
-        if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
-            /* TODO: avoid memset? */
-            size_t const ldmHSize = ((size_t)1) << params->ldmParams.hashLog;
-            zc->ldmState.hashTable = (ldmEntry_t*)ZSTD_cwksp_reserve_aligned(ws, ldmHSize * sizeof(ldmEntry_t));
-            ZSTD_memset(zc->ldmState.hashTable, 0, ldmHSize * sizeof(ldmEntry_t));
-            zc->ldmSequences = (rawSeq*)ZSTD_cwksp_reserve_aligned(ws, maxNbLdmSeq * sizeof(rawSeq));
-            zc->maxNbLdmSequences = maxNbLdmSeq;
-
-            ZSTD_window_init(&zc->ldmState.window);
-            zc->ldmState.loadedDictEnd = 0;
-        }
 
         DEBUGLOG(3, "wksp: finished allocating, %zd bytes remain available", ZSTD_cwksp_available_space(ws));
-        assert(ZSTD_cwksp_estimated_space_within_bounds(ws, neededSpace, resizeWorkspace));
+        assert(ZSTD_cwksp_estimated_space_within_bounds(ws, neededSpace));
 
         zc->initialized = 1;
 
···
         }
 
         params.cParams = ZSTD_adjustCParams_internal(adjusted_cdict_cParams, pledgedSrcSize,
-                                                     cdict->dictContentSize, ZSTD_cpm_attachDict);
+                                                     cdict->dictContentSize, ZSTD_cpm_attachDict,
+                                                     params.useRowMatchFinder);
         params.cParams.windowLog = windowLog;
         params.useRowMatchFinder = cdict->useRowMatchFinder;    /* cdict overrides */
         FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, &params, pledgedSrcSize,
···
     return 0;
 }
 
+static void ZSTD_copyCDictTableIntoCCtx(U32* dst, U32 const* src, size_t tableSize,
+                                        ZSTD_compressionParameters const* cParams) {
+    if (ZSTD_CDictIndicesAreTagged(cParams)){
+        /* Remove tags from the CDict table if they are present.
+         * See docs on "short cache" in zstd_compress_internal.h for context. */
+        size_t i;
+        for (i = 0; i < tableSize; i++) {
+            U32 const taggedIndex = src[i];
+            U32 const index = taggedIndex >> ZSTD_SHORT_CACHE_TAG_BITS;
+            dst[i] = index;
+        }
+    } else {
+        ZSTD_memcpy(dst, src, tableSize * sizeof(U32));
+    }
+}
+
 static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
                             const ZSTD_CDict* cdict,
                             ZSTD_CCtx_params params,
···
                                                             : 0;
         size_t const hSize =  (size_t)1 << cdict_cParams->hashLog;
 
-        ZSTD_memcpy(cctx->blockState.matchState.hashTable,
-               cdict->matchState.hashTable,
-               hSize * sizeof(U32));
+        ZSTD_copyCDictTableIntoCCtx(cctx->blockState.matchState.hashTable,
+                                cdict->matchState.hashTable,
+                                hSize, cdict_cParams);
+
         /* Do not copy cdict's chainTable if cctx has parameters such that it would not use chainTable */
         if (ZSTD_allocateChainTable(cctx->appliedParams.cParams.strategy, cctx->appliedParams.useRowMatchFinder, 0 /* forDDSDict */)) {
-            ZSTD_memcpy(cctx->blockState.matchState.chainTable,
-               cdict->matchState.chainTable,
-               chainSize * sizeof(U32));
+            ZSTD_copyCDictTableIntoCCtx(cctx->blockState.matchState.chainTable,
+                                    cdict->matchState.chainTable,
+                                    chainSize, cdict_cParams);
         }
         /* copy tag table */
         if (ZSTD_rowMatchFinderUsed(cdict_cParams->strategy, cdict->useRowMatchFinder)) {
-            size_t const tagTableSize = hSize*sizeof(U16);
+            size_t const tagTableSize = hSize;
             ZSTD_memcpy(cctx->blockState.matchState.tagTable,
-                cdict->matchState.tagTable,
-                tagTableSize);
+                        cdict->matchState.tagTable,
+                        tagTableSize);
+            cctx->blockState.matchState.hashSalt = cdict->matchState.hashSalt;
         }
     }
 
     /* Zero the hashTable3, since the cdict never fills it */
-    {   int const h3log = cctx->blockState.matchState.hashLog3;
+    assert(cctx->blockState.matchState.hashLog3 <= 31);
+    {   U32 const h3log = cctx->blockState.matchState.hashLog3;
         size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0;
         assert(cdict->matchState.hashLog3 == 0);
         ZSTD_memset(cctx->blockState.matchState.hashTable3, 0, h3Size * sizeof(U32));
···
     ZSTD_cwksp_mark_tables_clean(&cctx->workspace);
 
     /* copy dictionary offsets */
-    {   ZSTD_matchState_t const* srcMatchState = &cdict->matchState;
-        ZSTD_matchState_t* dstMatchState = &cctx->blockState.matchState;
+    {   ZSTD_MatchState_t const* srcMatchState = &cdict->matchState;
+        ZSTD_MatchState_t* dstMatchState = &cctx->blockState.matchState;
         dstMatchState->window       = srcMatchState->window;
         dstMatchState->nextToUpdate = srcMatchState->nextToUpdate;
         dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd;
···
         /* Copy only compression parameters related to tables. */
         params.cParams = srcCCtx->appliedParams.cParams;
         assert(srcCCtx->appliedParams.useRowMatchFinder != ZSTD_ps_auto);
-        assert(srcCCtx->appliedParams.useBlockSplitter != ZSTD_ps_auto);
+        assert(srcCCtx->appliedParams.postBlockSplitter != ZSTD_ps_auto);
         assert(srcCCtx->appliedParams.ldmParams.enableLdm != ZSTD_ps_auto);
         params.useRowMatchFinder = srcCCtx->appliedParams.useRowMatchFinder;
-        params.useBlockSplitter = srcCCtx->appliedParams.useBlockSplitter;
+        params.postBlockSplitter = srcCCtx->appliedParams.postBlockSplitter;
         params.ldmParams = srcCCtx->appliedParams.ldmParams;
         params.fParams = fParams;
+        params.maxBlockSize = srcCCtx->appliedParams.maxBlockSize;
         ZSTD_resetCCtx_internal(dstCCtx, &params, pledgedSrcSize,
                                 /* loadedDictSize */ 0,
                                 ZSTDcrp_leaveDirty, zbuff);
···
                                     ? ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog)
                                     : 0;
         size_t const hSize =  (size_t)1 << srcCCtx->appliedParams.cParams.hashLog;
-        int const h3log = srcCCtx->blockState.matchState.hashLog3;
+        U32 const h3log = srcCCtx->blockState.matchState.hashLog3;
         size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0;
 
         ZSTD_memcpy(dstCCtx->blockState.matchState.hashTable,
···
 
     /* copy dictionary offsets */
     {
-        const ZSTD_matchState_t* srcMatchState = &srcCCtx->blockState.matchState;
-        ZSTD_matchState_t* dstMatchState = &dstCCtx->blockState.matchState;
+        const ZSTD_MatchState_t* srcMatchState = &srcCCtx->blockState.matchState;
+        ZSTD_MatchState_t* dstMatchState = &dstCCtx->blockState.matchState;
         dstMatchState->window       = srcMatchState->window;
         dstMatchState->nextToUpdate = srcMatchState->nextToUpdate;
         dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd;
···
     /* Protect special index values < ZSTD_WINDOW_START_INDEX. */
     U32 const reducerThreshold = reducerValue + ZSTD_WINDOW_START_INDEX;
     assert((size & (ZSTD_ROWSIZE-1)) == 0);  /* multiple of ZSTD_ROWSIZE */
-    assert(size < (1U<<31));   /* can be casted to int */
+    assert(size < (1U<<31));   /* can be cast to int */
 
 
     for (rowNb=0 ; rowNb < nbRows ; rowNb++) {
···
 
 /*! ZSTD_reduceIndex() :
 *   rescale all indexes to avoid future overflow (indexes are U32) */
-static void ZSTD_reduceIndex (ZSTD_matchState_t* ms, ZSTD_CCtx_params const* params, const U32 reducerValue)
+static void ZSTD_reduceIndex (ZSTD_MatchState_t* ms, ZSTD_CCtx_params const* params, const U32 reducerValue)
 {
     {   U32 const hSize = (U32)1 << params->cParams.hashLog;
         ZSTD_reduceTable(ms->hashTable, hSize, reducerValue);
···
 
 /* See doc/zstd_compression_format.md for detailed format description */
 
-void ZSTD_seqToCodes(const seqStore_t* seqStorePtr)
+int ZSTD_seqToCodes(const SeqStore_t* seqStorePtr)
 {
-    const seqDef* const sequences = seqStorePtr->sequencesStart;
+    const SeqDef* const sequences = seqStorePtr->sequencesStart;
     BYTE* const llCodeTable = seqStorePtr->llCode;
     BYTE* const ofCodeTable = seqStorePtr->ofCode;
     BYTE* const mlCodeTable = seqStorePtr->mlCode;
     U32 const nbSeq = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
     U32 u;
+    int longOffsets = 0;
     assert(nbSeq <= seqStorePtr->maxNbSeq);
     for (u=0; u<nbSeq; u++) {
         U32 const llv = sequences[u].litLength;
+        U32 const ofCode = ZSTD_highbit32(sequences[u].offBase);
         U32 const mlv = sequences[u].mlBase;
         llCodeTable[u] = (BYTE)ZSTD_LLcode(llv);
-        ofCodeTable[u] = (BYTE)ZSTD_highbit32(sequences[u].offBase);
+        ofCodeTable[u] = (BYTE)ofCode;
         mlCodeTable[u] = (BYTE)ZSTD_MLcode(mlv);
+        assert(!(MEM_64bits() && ofCode >= STREAM_ACCUMULATOR_MIN));
+        if (MEM_32bits() && ofCode >= STREAM_ACCUMULATOR_MIN)
+            longOffsets = 1;
     }
     if (seqStorePtr->longLengthType==ZSTD_llt_literalLength)
         llCodeTable[seqStorePtr->longLengthPos] = MaxLL;
     if (seqStorePtr->longLengthType==ZSTD_llt_matchLength)
         mlCodeTable[seqStorePtr->longLengthPos] = MaxML;
+    return longOffsets;
 }
 
 /* ZSTD_useTargetCBlockSize():
···
  * Returns 1 if true, 0 otherwise. */
 static int ZSTD_blockSplitterEnabled(ZSTD_CCtx_params* cctxParams)
 {
-    DEBUGLOG(5, "ZSTD_blockSplitterEnabled (useBlockSplitter=%d)", cctxParams->useBlockSplitter);
-    assert(cctxParams->useBlockSplitter != ZSTD_ps_auto);
-    return (cctxParams->useBlockSplitter == ZSTD_ps_enable);
+    DEBUGLOG(5, "ZSTD_blockSplitterEnabled (postBlockSplitter=%d)", cctxParams->postBlockSplitter);
+    assert(cctxParams->postBlockSplitter != ZSTD_ps_auto);
+    return (cctxParams->postBlockSplitter == ZSTD_ps_enable);
 }
 
 /* Type returned by ZSTD_buildSequencesStatistics containing finalized symbol encoding types
···
     U32 MLtype;
     size_t size;
     size_t lastCountSize; /* Accounts for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */
+    int longOffsets;
 } ZSTD_symbolEncodingTypeStats_t;
 
 /* ZSTD_buildSequencesStatistics():
···
  * entropyWkspSize must be of size at least ENTROPY_WORKSPACE_SIZE - (MaxSeq + 1)*sizeof(U32)
  */
 static ZSTD_symbolEncodingTypeStats_t
-ZSTD_buildSequencesStatistics(seqStore_t* seqStorePtr, size_t nbSeq,
-                        const ZSTD_fseCTables_t* prevEntropy, ZSTD_fseCTables_t* nextEntropy,
-                              BYTE* dst, const BYTE* const dstEnd,
-                              ZSTD_strategy strategy, unsigned* countWorkspace,
-                              void* entropyWorkspace, size_t entropyWkspSize) {
+ZSTD_buildSequencesStatistics(
+                const SeqStore_t* seqStorePtr, size_t nbSeq,
+                const ZSTD_fseCTables_t* prevEntropy, ZSTD_fseCTables_t* nextEntropy,
+                      BYTE* dst, const BYTE* const dstEnd,
+                      ZSTD_strategy strategy, unsigned* countWorkspace,
+                      void* entropyWorkspace, size_t entropyWkspSize)
+{
     BYTE* const ostart = dst;
     const BYTE* const oend = dstEnd;
     BYTE* op = ostart;
···
 
     stats.lastCountSize = 0;
     /* convert length/distances into codes */
-    ZSTD_seqToCodes(seqStorePtr);
+    stats.longOffsets = ZSTD_seqToCodes(seqStorePtr);
     assert(op <= oend);
     assert(nbSeq != 0); /* ZSTD_selectEncodingType() divides by nbSeq */
     /* build CTable for Literal Lengths */
···
         assert(!(stats.LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
         {   size_t const countSize = ZSTD_buildCTable(
                 op, (size_t)(oend - op),
-                CTable_LitLength, LLFSELog, (symbolEncodingType_e)stats.LLtype,
+                CTable_LitLength, LLFSELog, (SymbolEncodingType_e)stats.LLtype,
                 countWorkspace, max, llCodeTable, nbSeq,
                 LL_defaultNorm, LL_defaultNormLog, MaxLL,
                 prevEntropy->litlengthCTable,
···
         size_t const mostFrequent = HIST_countFast_wksp(
             countWorkspace, &max, ofCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);  /* can't fail */
         /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
-        ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
+        ZSTD_DefaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
         DEBUGLOG(5, "Building OF table");
         nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
         stats.Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode,
···
         assert(!(stats.Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
         {   size_t const countSize = ZSTD_buildCTable(
                 op, (size_t)(oend - op),
-                CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)stats.Offtype,
+                CTable_OffsetBits, OffFSELog, (SymbolEncodingType_e)stats.Offtype,
                 countWorkspace, max, ofCodeTable, nbSeq,
                 OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
                 prevEntropy->offcodeCTable,
···
         assert(!(stats.MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
         {   size_t const countSize = ZSTD_buildCTable(
                 op, (size_t)(oend - op),
-                CTable_MatchLength, MLFSELog, (symbolEncodingType_e)stats.MLtype,
+                CTable_MatchLength, MLFSELog, (SymbolEncodingType_e)stats.MLtype,
                 countWorkspace, max, mlCodeTable, nbSeq,
                 ML_defaultNorm, ML_defaultNormLog, MaxML,
                 prevEntropy->matchlengthCTable,
···
  */
 #define SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO 20
 MEM_STATIC size_t
-ZSTD_entropyCompressSeqStore_internal(seqStore_t* seqStorePtr,
-                          const ZSTD_entropyCTables_t* prevEntropy,
-                                ZSTD_entropyCTables_t* nextEntropy,
-                          const ZSTD_CCtx_params* cctxParams,
-                                void* dst, size_t dstCapacity,
-                                void* entropyWorkspace, size_t entropyWkspSize,
-                          const int bmi2)
+ZSTD_entropyCompressSeqStore_internal(
+                              void* dst, size_t dstCapacity,
+                        const void* literals, size_t litSize,
+                        const SeqStore_t* seqStorePtr,
+                        const ZSTD_entropyCTables_t* prevEntropy,
+                              ZSTD_entropyCTables_t* nextEntropy,
+                        const ZSTD_CCtx_params* cctxParams,
+                              void* entropyWorkspace, size_t entropyWkspSize,
+                        const int bmi2)
 {
-    const int longOffsets = cctxParams->cParams.windowLog > STREAM_ACCUMULATOR_MIN;
     ZSTD_strategy const strategy = cctxParams->cParams.strategy;
     unsigned* count = (unsigned*)entropyWorkspace;
     FSE_CTable* CTable_LitLength = nextEntropy->fse.litlengthCTable;
     FSE_CTable* CTable_OffsetBits = nextEntropy->fse.offcodeCTable;
     FSE_CTable* CTable_MatchLength = nextEntropy->fse.matchlengthCTable;
-    const seqDef* const sequences = seqStorePtr->sequencesStart;
-    const size_t nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart;
+    const SeqDef* const sequences = seqStorePtr->sequencesStart;
+    const size_t nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
     const BYTE* const ofCodeTable = seqStorePtr->ofCode;
     const BYTE* const llCodeTable = seqStorePtr->llCode;
     const BYTE* const mlCodeTable = seqStorePtr->mlCode;
···
     BYTE* const oend = ostart + dstCapacity;
     BYTE* op = ostart;
     size_t lastCountSize;
+    int longOffsets = 0;
 
     entropyWorkspace = count + (MaxSeq + 1);
     entropyWkspSize -= (MaxSeq + 1) * sizeof(*count);
 
-    DEBUGLOG(4, "ZSTD_entropyCompressSeqStore_internal (nbSeq=%zu)", nbSeq);
+    DEBUGLOG(5, "ZSTD_entropyCompressSeqStore_internal (nbSeq=%zu, dstCapacity=%zu)", nbSeq, dstCapacity);
     ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
     assert(entropyWkspSize >= HUF_WORKSPACE_SIZE);
 
     /* Compress literals */
-    {   const BYTE* const literals = seqStorePtr->litStart;
-        size_t const numSequences = seqStorePtr->sequences - seqStorePtr->sequencesStart;
-        size_t const numLiterals = seqStorePtr->lit - seqStorePtr->litStart;
+    {   size_t const numSequences = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
         /* Base suspicion of uncompressibility on ratio of literals to sequences */
-        unsigned const suspectUncompressible = (numSequences == 0) || (numLiterals / numSequences >= SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO);
-        size_t const litSize = (size_t)(seqStorePtr->lit - literals);
+        int const suspectUncompressible = (numSequences == 0) || (litSize / numSequences >= SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO);
+
         size_t const cSize = ZSTD_compressLiterals(
-                                    &prevEntropy->huf, &nextEntropy->huf,
-                                    cctxParams->cParams.strategy,
-                                    ZSTD_literalsCompressionIsDisabled(cctxParams),
                                     op, dstCapacity,
                                     literals, litSize,
                                     entropyWorkspace, entropyWkspSize,
-                                    bmi2, suspectUncompressible);
+                                    &prevEntropy->huf, &nextEntropy->huf,
+                                    cctxParams->cParams.strategy,
+                                    ZSTD_literalsCompressionIsDisabled(cctxParams),
+                                    suspectUncompressible, bmi2);
         FORWARD_IF_ERROR(cSize, "ZSTD_compressLiterals failed");
         assert(cSize <= dstCapacity);
         op += cSize;
···
         ZSTD_memcpy(&nextEntropy->fse, &prevEntropy->fse, sizeof(prevEntropy->fse));
         return (size_t)(op - ostart);
     }
-    {
-        ZSTD_symbolEncodingTypeStats_t stats;
-        BYTE* seqHead = op++;
+    {   BYTE* const seqHead = op++;
         /* build stats for sequences */
-        stats = ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq,
+        const ZSTD_symbolEncodingTypeStats_t stats =
+                ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq,
                                              &prevEntropy->fse, &nextEntropy->fse,
                                               op, oend,
                                               strategy, count,
···
         *seqHead = (BYTE)((stats.LLtype<<6) + (stats.Offtype<<4) + (stats.MLtype<<2));
         lastCountSize = stats.lastCountSize;
         op += stats.size;
+        longOffsets = stats.longOffsets;
     }
 
     {   size_t const bitstreamSize = ZSTD_encodeSequences(
···
     return (size_t)(op - ostart);
 }
 
-MEM_STATIC size_t
-ZSTD_entropyCompressSeqStore(seqStore_t* seqStorePtr,
-                       const ZSTD_entropyCTables_t* prevEntropy,
-                             ZSTD_entropyCTables_t* nextEntropy,
-                       const ZSTD_CCtx_params* cctxParams,
-                             void* dst, size_t dstCapacity,
-                             size_t srcSize,
-                             void* entropyWorkspace, size_t entropyWkspSize,
-                             int bmi2)
+static size_t
+ZSTD_entropyCompressSeqStore_wExtLitBuffer(
+                          void* dst, size_t dstCapacity,
+                    const void* literals, size_t litSize,
+                          size_t blockSize,
+                    const SeqStore_t* seqStorePtr,
+                    const ZSTD_entropyCTables_t* prevEntropy,
+                          ZSTD_entropyCTables_t* nextEntropy,
+                    const ZSTD_CCtx_params* cctxParams,
+                          void* entropyWorkspace, size_t entropyWkspSize,
+                          int bmi2)
 {
     size_t const cSize = ZSTD_entropyCompressSeqStore_internal(
-                            seqStorePtr, prevEntropy, nextEntropy, cctxParams,
                             dst, dstCapacity,
+                            literals, litSize,
+                            seqStorePtr, prevEntropy, nextEntropy, cctxParams,
                             entropyWorkspace, entropyWkspSize, bmi2);
     if (cSize == 0) return 0;
     /* When srcSize <= dstCapacity, there is enough space to write a raw uncompressed block.
      * Since we ran out of space, block must be not compressible, so fall back to raw uncompressed block.
      */
-    if ((cSize == ERROR(dstSize_tooSmall)) & (srcSize <= dstCapacity))
+    if ((cSize == ERROR(dstSize_tooSmall)) & (blockSize <= dstCapacity)) {
+        DEBUGLOG(4, "not enough dstCapacity (%zu) for ZSTD_entropyCompressSeqStore_internal()=> do not compress block", dstCapacity);
         return 0;  /* block not compressed */
+    }
     FORWARD_IF_ERROR(cSize, "ZSTD_entropyCompressSeqStore_internal failed");
 
     /* Check compressibility */
-    {   size_t const maxCSize = srcSize - ZSTD_minGain(srcSize, cctxParams->cParams.strategy);
+    {   size_t const maxCSize = blockSize - ZSTD_minGain(blockSize, cctxParams->cParams.strategy);
         if (cSize >= maxCSize) return 0;  /* block not compressed */
     }
-    DEBUGLOG(4, "ZSTD_entropyCompressSeqStore() cSize: %zu", cSize);
+    DEBUGLOG(5, "ZSTD_entropyCompressSeqStore() cSize: %zu", cSize);
+    /* libzstd decoder before  > v1.5.4 is not compatible with compressed blocks of size ZSTD_BLOCKSIZE_MAX exactly.
+     * This restriction is indirectly already fulfilled by respecting ZSTD_minGain() condition above.
+     */
+    assert(cSize < ZSTD_BLOCKSIZE_MAX);
     return cSize;
+}
+
+static size_t
+ZSTD_entropyCompressSeqStore(
+                    const SeqStore_t* seqStorePtr,
+                    const ZSTD_entropyCTables_t* prevEntropy,
+                          ZSTD_entropyCTables_t* nextEntropy,
+                    const ZSTD_CCtx_params* cctxParams,
+                          void* dst, size_t dstCapacity,
+                          size_t srcSize,
+                          void* entropyWorkspace, size_t entropyWkspSize,
+                          int bmi2)
+{
+    return ZSTD_entropyCompressSeqStore_wExtLitBuffer(
+                dst, dstCapacity,
+                seqStorePtr->litStart, (size_t)(seqStorePtr->lit - seqStorePtr->litStart),
+                srcSize,
+                seqStorePtr,
+                prevEntropy, nextEntropy,
+                cctxParams,
+                entropyWorkspace, entropyWkspSize,
+                bmi2);
 }
 
 /* ZSTD_selectBlockCompressor() :
  * Not static, but internal use only (used by long distance matcher)
  * assumption : strat is a valid strategy */
-ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_paramSwitch_e useRowMatchFinder, ZSTD_dictMode_e dictMode)
+ZSTD_BlockCompressor_f ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_ParamSwitch_e useRowMatchFinder, ZSTD_dictMode_e dictMode)
 {
-    static const ZSTD_blockCompressor blockCompressor[4][ZSTD_STRATEGY_MAX+1] = {
+    static const ZSTD_BlockCompressor_f blockCompressor[4][ZSTD_STRATEGY_MAX+1] = {
         { ZSTD_compressBlock_fast  /* default for 0 */,
           ZSTD_compressBlock_fast,
-          ZSTD_compressBlock_doubleFast,
-          ZSTD_compressBlock_greedy,
-          ZSTD_compressBlock_lazy,
-          ZSTD_compressBlock_lazy2,
-          ZSTD_compressBlock_btlazy2,
-          ZSTD_compressBlock_btopt,
-          ZSTD_compressBlock_btultra,
-          ZSTD_compressBlock_btultra2 },
+          ZSTD_COMPRESSBLOCK_DOUBLEFAST,
+          ZSTD_COMPRESSBLOCK_GREEDY,
+          ZSTD_COMPRESSBLOCK_LAZY,
+          ZSTD_COMPRESSBLOCK_LAZY2,
+          ZSTD_COMPRESSBLOCK_BTLAZY2,
+          ZSTD_COMPRESSBLOCK_BTOPT,
+          ZSTD_COMPRESSBLOCK_BTULTRA,
+          ZSTD_COMPRESSBLOCK_BTULTRA2
+        },
         { ZSTD_compressBlock_fast_extDict  /* default for 0 */,
           ZSTD_compressBlock_fast_extDict,
-          ZSTD_compressBlock_doubleFast_extDict,
-          ZSTD_compressBlock_greedy_extDict,
-          ZSTD_compressBlock_lazy_extDict,
-          ZSTD_compressBlock_lazy2_extDict,
-          ZSTD_compressBlock_btlazy2_extDict,
-          ZSTD_compressBlock_btopt_extDict,
-          ZSTD_compressBlock_btultra_extDict,
-          ZSTD_compressBlock_btultra_extDict },
+          ZSTD_COMPRESSBLOCK_DOUBLEFAST_EXTDICT,
+          ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT,
+          ZSTD_COMPRESSBLOCK_LAZY_EXTDICT,
+          ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT,
+          ZSTD_COMPRESSBLOCK_BTLAZY2_EXTDICT,
+          ZSTD_COMPRESSBLOCK_BTOPT_EXTDICT,
+          ZSTD_COMPRESSBLOCK_BTULTRA_EXTDICT,
+          ZSTD_COMPRESSBLOCK_BTULTRA_EXTDICT
+        },
         { ZSTD_compressBlock_fast_dictMatchState  /* default for 0 */,
           ZSTD_compressBlock_fast_dictMatchState,
-          ZSTD_compressBlock_doubleFast_dictMatchState,
-          ZSTD_compressBlock_greedy_dictMatchState,
-          ZSTD_compressBlock_lazy_dictMatchState,
-          ZSTD_compressBlock_lazy2_dictMatchState,
-          ZSTD_compressBlock_btlazy2_dictMatchState,
-          ZSTD_compressBlock_btopt_dictMatchState,
-          ZSTD_compressBlock_btultra_dictMatchState,
-          ZSTD_compressBlock_btultra_dictMatchState },
+          ZSTD_COMPRESSBLOCK_DOUBLEFAST_DICTMATCHSTATE,
+          ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE,
+          ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE,
+          ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE,
+          ZSTD_COMPRESSBLOCK_BTLAZY2_DICTMATCHSTATE,
+          ZSTD_COMPRESSBLOCK_BTOPT_DICTMATCHSTATE,
+          ZSTD_COMPRESSBLOCK_BTULTRA_DICTMATCHSTATE,
+          ZSTD_COMPRESSBLOCK_BTULTRA_DICTMATCHSTATE
+        },
         { NULL  /* default for 0 */,
           NULL,
           NULL,
-          ZSTD_compressBlock_greedy_dedicatedDictSearch,
-          ZSTD_compressBlock_lazy_dedicatedDictSearch,
-          ZSTD_compressBlock_lazy2_dedicatedDictSearch,
+          ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH,
+          ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH,
+          ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH,
           NULL,
           NULL,
           NULL,
           NULL }
     };
-    ZSTD_blockCompressor selectedCompressor;
+    ZSTD_BlockCompressor_f selectedCompressor;
     ZSTD_STATIC_ASSERT((unsigned)ZSTD_fast == 1);
 
-    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat));
-    DEBUGLOG(4, "Selected block compressor: dictMode=%d strat=%d rowMatchfinder=%d", (int)dictMode, (int)strat, (int)useRowMatchFinder);
+    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, (int)strat));
+    DEBUGLOG(5, "Selected block compressor: dictMode=%d strat=%d rowMatchfinder=%d", (int)dictMode, (int)strat, (int)useRowMatchFinder);
     if (ZSTD_rowMatchFinderUsed(strat, useRowMatchFinder)) {
-        static const ZSTD_blockCompressor rowBasedBlockCompressors[4][3] = {
-            { ZSTD_compressBlock_greedy_row,
-            ZSTD_compressBlock_lazy_row,
-            ZSTD_compressBlock_lazy2_row },
-            { ZSTD_compressBlock_greedy_extDict_row,
-            ZSTD_compressBlock_lazy_extDict_row,
-            ZSTD_compressBlock_lazy2_extDict_row },
-            { ZSTD_compressBlock_greedy_dictMatchState_row,
-            ZSTD_compressBlock_lazy_dictMatchState_row,
-            ZSTD_compressBlock_lazy2_dictMatchState_row },
-            { ZSTD_compressBlock_greedy_dedicatedDictSearch_row,
-            ZSTD_compressBlock_lazy_dedicatedDictSearch_row,
-            ZSTD_compressBlock_lazy2_dedicatedDictSearch_row }
+        static const ZSTD_BlockCompressor_f rowBasedBlockCompressors[4][3] = {
+            {
+                ZSTD_COMPRESSBLOCK_GREEDY_ROW,
+                ZSTD_COMPRESSBLOCK_LAZY_ROW,
+                ZSTD_COMPRESSBLOCK_LAZY2_ROW
+            },
+            {
+                ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT_ROW,
+                ZSTD_COMPRESSBLOCK_LAZY_EXTDICT_ROW,
+                ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT_ROW
+            },
+            {
+                ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE_ROW,
+                ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE_ROW,
+                ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE_ROW
+            },
+            {
+                ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH_ROW,
+                ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH_ROW,
+                ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH_ROW
+            }
         };
-        DEBUGLOG(4, "Selecting a row-based matchfinder");
+        DEBUGLOG(5, "Selecting a row-based matchfinder");
         assert(useRowMatchFinder != ZSTD_ps_auto);
         selectedCompressor = rowBasedBlockCompressors[(int)dictMode][(int)strat - (int)ZSTD_greedy];
     } else {
···
     return selectedCompressor;
 }
 
-static void ZSTD_storeLastLiterals(seqStore_t* seqStorePtr,
+static void ZSTD_storeLastLiterals(SeqStore_t* seqStorePtr,
                                    const BYTE* anchor, size_t lastLLSize)
 {
     ZSTD_memcpy(seqStorePtr->lit, anchor, lastLLSize);
     seqStorePtr->lit += lastLLSize;
 }
 
-void ZSTD_resetSeqStore(seqStore_t* ssPtr)
+void ZSTD_resetSeqStore(SeqStore_t* ssPtr)
 {
     ssPtr->lit = ssPtr->litStart;
     ssPtr->sequences = ssPtr->sequencesStart;
     ssPtr->longLengthType = ZSTD_llt_none;
 }
 
-typedef enum { ZSTDbss_compress, ZSTDbss_noCompress } ZSTD_buildSeqStore_e;
+/* ZSTD_postProcessSequenceProducerResult() :
+ * Validates and post-processes sequences obtained through the external matchfinder API:
+ *   - Checks whether nbExternalSeqs represents an error condition.
+ *   - Appends a block delimiter to outSeqs if one is not already present.
+ *     See zstd.h for context regarding block delimiters.
+ * Returns the number of sequences after post-processing, or an error code. */
+static size_t ZSTD_postProcessSequenceProducerResult(
+    ZSTD_Sequence* outSeqs, size_t nbExternalSeqs, size_t outSeqsCapacity, size_t srcSize
+) {
+    RETURN_ERROR_IF(
+        nbExternalSeqs > outSeqsCapacity,
+        sequenceProducer_failed,
+        "External sequence producer returned error code %lu",
+        (unsigned long)nbExternalSeqs
+    );
+
+    RETURN_ERROR_IF(
+        nbExternalSeqs == 0 && srcSize > 0,
+        sequenceProducer_failed,
+        "Got zero sequences from external sequence producer for a non-empty src buffer!"
+    );
+
+    if (srcSize == 0) {
+        ZSTD_memset(&outSeqs[0], 0, sizeof(ZSTD_Sequence));
+        return 1;
+    }
+
+    {
+        ZSTD_Sequence const lastSeq = outSeqs[nbExternalSeqs - 1];
+
+        /* We can return early if lastSeq is already a block delimiter. */
+        if (lastSeq.offset == 0 && lastSeq.matchLength == 0) {
+            return nbExternalSeqs;
+        }
+
+        /* This error condition is only possible if the external matchfinder
+         * produced an invalid parse, by definition of ZSTD_sequenceBound(). */
+        RETURN_ERROR_IF(
+            nbExternalSeqs == outSeqsCapacity,
+            sequenceProducer_failed,
+            "nbExternalSeqs == outSeqsCapacity but lastSeq is not a block delimiter!"
+        );
+
+        /* lastSeq is not a block delimiter, so we need to append one. */
+        ZSTD_memset(&outSeqs[nbExternalSeqs], 0, sizeof(ZSTD_Sequence));
+        return nbExternalSeqs + 1;
+    }
+}
+
+/* ZSTD_fastSequenceLengthSum() :
+ * Returns sum(litLen) + sum(matchLen) + lastLits for *seqBuf*.
+ * Similar to another function in zstd_compress.c (determine_blockSize),
+ * except it doesn't check for a block delimiter to end summation.
+ * Removing the early exit allows the compiler to auto-vectorize (https://godbolt.org/z/cY1cajz9P).
+ * This function can be deleted and replaced by determine_blockSize after we resolve issue #3456. */
+static size_t ZSTD_fastSequenceLengthSum(ZSTD_Sequence const* seqBuf, size_t seqBufSize) {
+    size_t matchLenSum, litLenSum, i;
+    matchLenSum = 0;
+    litLenSum = 0;
+    for (i = 0; i < seqBufSize; i++) {
+        litLenSum += seqBuf[i].litLength;
+        matchLenSum += seqBuf[i].matchLength;
+    }
+    return litLenSum + matchLenSum;
+}
+
+/*
+ * Function to validate sequences produced by a block compressor.
+ */
+static void ZSTD_validateSeqStore(const SeqStore_t* seqStore, const ZSTD_compressionParameters* cParams)
+{
+#if DEBUGLEVEL >= 1
+    const SeqDef* seq = seqStore->sequencesStart;
+    const SeqDef* const seqEnd = seqStore->sequences;
+    size_t const matchLenLowerBound = cParams->minMatch == 3 ? 3 : 4;
+    for (; seq < seqEnd; ++seq) {
+        const ZSTD_SequenceLength seqLength = ZSTD_getSequenceLength(seqStore, seq);
+        assert(seqLength.matchLength >= matchLenLowerBound);
+        (void)seqLength;
+        (void)matchLenLowerBound;
+    }
+#else
+    (void)seqStore;
+    (void)cParams;
+#endif
+}
+
+static size_t
+ZSTD_transferSequences_wBlockDelim(ZSTD_CCtx* cctx,
+                                   ZSTD_SequencePosition* seqPos,
+                             const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                             const void* src, size_t blockSize,
+                                   ZSTD_ParamSwitch_e externalRepSearch);
+
+typedef enum { ZSTDbss_compress, ZSTDbss_noCompress } ZSTD_BuildSeqStore_e;
 
 static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
 {
-    ZSTD_matchState_t* const ms = &zc->blockState.matchState;
+    ZSTD_MatchState_t* const ms = &zc->blockState.matchState;
     DEBUGLOG(5, "ZSTD_buildSeqStore (srcSize=%zu)", srcSize);
     assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
     /* Assert that we have correctly flushed the ctx params into the ms's copy */
     ZSTD_assertEqualCParams(zc->appliedParams.cParams, ms->cParams);
-    if (srcSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1) {
+    /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
+     * additional 1. We need to revisit and change this logic to be more consistent */
+    if (srcSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1+1) {
         if (zc->appliedParams.cParams.strategy >= ZSTD_btopt) {
             ZSTD_ldm_skipRawSeqStoreBytes(&zc->externSeqStore, srcSize);
         } else {
···
         }
         if (zc->externSeqStore.pos < zc->externSeqStore.size) {
             assert(zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_disable);
+
+            /* External matchfinder + LDM is technically possible, just not implemented yet.
+             * We need to revisit soon and implement it. */
+            RETURN_ERROR_IF(
+                ZSTD_hasExtSeqProd(&zc->appliedParams),
+                parameter_combination_unsupported,
+                "Long-distance matching with external sequence producer enabled is not currently supported."
+            );
+
             /* Updates ldmSeqStore.pos */
             lastLLSize =
                 ZSTD_ldm_blockCompress(&zc->externSeqStore,
···
                                        src, srcSize);
             assert(zc->externSeqStore.pos <= zc->externSeqStore.size);
         } else if (zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable) {
-            rawSeqStore_t ldmSeqStore = kNullRawSeqStore;
+            RawSeqStore_t ldmSeqStore = kNullRawSeqStore;
+
+            /* External matchfinder + LDM is technically possible, just not implemented yet.
+             * We need to revisit soon and implement it. */
+            RETURN_ERROR_IF(
+                ZSTD_hasExtSeqProd(&zc->appliedParams),
+                parameter_combination_unsupported,
+                "Long-distance matching with external sequence producer enabled is not currently supported."
+            );
 
             ldmSeqStore.seq = zc->ldmSequences;
             ldmSeqStore.capacity = zc->maxNbLdmSequences;
···
                                        zc->appliedParams.useRowMatchFinder,
                                        src, srcSize);
             assert(ldmSeqStore.pos == ldmSeqStore.size);
-        } else {   /* not long range mode */
-            ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy,
-                                                                                    zc->appliedParams.useRowMatchFinder,
-                                                                                    dictMode);
+        } else if (ZSTD_hasExtSeqProd(&zc->appliedParams)) {
+            assert(
+                zc->extSeqBufCapacity >= ZSTD_sequenceBound(srcSize)
+            );
+            assert(zc->appliedParams.extSeqProdFunc != NULL);
+
+            {   U32 const windowSize = (U32)1 << zc->appliedParams.cParams.windowLog;
+
+                size_t const nbExternalSeqs = (zc->appliedParams.extSeqProdFunc)(
+                    zc->appliedParams.extSeqProdState,
+                    zc->extSeqBuf,
+                    zc->extSeqBufCapacity,
+                    src, srcSize,
+                    NULL, 0,  /* dict and dictSize, currently not supported */
+                    zc->appliedParams.compressionLevel,
+                    windowSize
+                );
+
+                size_t const nbPostProcessedSeqs = ZSTD_postProcessSequenceProducerResult(
+                    zc->extSeqBuf,
+                    nbExternalSeqs,
+                    zc->extSeqBufCapacity,
+                    srcSize
+                );
+
+                /* Return early if there is no error, since we don't need to worry about last literals */
+                if (!ZSTD_isError(nbPostProcessedSeqs)) {
+                    ZSTD_SequencePosition seqPos = {0,0,0};
+                    size_t const seqLenSum = ZSTD_fastSequenceLengthSum(zc->extSeqBuf, nbPostProcessedSeqs);
+                    RETURN_ERROR_IF(seqLenSum > srcSize, externalSequences_invalid, "External sequences imply too large a block!");
+                    FORWARD_IF_ERROR(
+                        ZSTD_transferSequences_wBlockDelim(
+                            zc, &seqPos,
+                            zc->extSeqBuf, nbPostProcessedSeqs,
+                            src, srcSize,
+                            zc->appliedParams.searchForExternalRepcodes
+                        ),
+                        "Failed to copy external sequences to seqStore!"
+                    );
+                    ms->ldmSeqStore = NULL;
+                    DEBUGLOG(5, "Copied %lu sequences from external sequence producer to internal seqStore.", (unsigned long)nbExternalSeqs);
+                    return ZSTDbss_compress;
+                }
+
+                /* Propagate the error if fallback is disabled */
+                if (!zc->appliedParams.enableMatchFinderFallback) {
+                    return nbPostProcessedSeqs;
+                }
+
+                /* Fallback to software matchfinder */
+                {   ZSTD_BlockCompressor_f const blockCompressor =
+                        ZSTD_selectBlockCompressor(
+                            zc->appliedParams.cParams.strategy,
+                            zc->appliedParams.useRowMatchFinder,
+                            dictMode);
+                    ms->ldmSeqStore = NULL;
+                    DEBUGLOG(
+                        5,
+                        "External sequence producer returned error code %lu. Falling back to internal parser.",
+                        (unsigned long)nbExternalSeqs
+                    );
+                    lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize);
+            }   }
+        } else {   /* not long range mode and no external matchfinder */
+            ZSTD_BlockCompressor_f const blockCompressor = ZSTD_selectBlockCompressor(
+                    zc->appliedParams.cParams.strategy,
+                    zc->appliedParams.useRowMatchFinder,
+                    dictMode);
             ms->ldmSeqStore = NULL;
             lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize);
         }
         {   const BYTE* const lastLiterals = (const BYTE*)src + srcSize - lastLLSize;
             ZSTD_storeLastLiterals(&zc->seqStore, lastLiterals, lastLLSize);
     }   }
+    ZSTD_validateSeqStore(&zc->seqStore, &zc->appliedParams.cParams);
     return ZSTDbss_compress;
 }
 
-static void ZSTD_copyBlockSequences(ZSTD_CCtx* zc)
+static size_t ZSTD_copyBlockSequences(SeqCollector* seqCollector, const SeqStore_t* seqStore, const U32 prevRepcodes[ZSTD_REP_NUM])
 {
-    const seqStore_t* seqStore = ZSTD_getSeqStore(zc);
-    const seqDef* seqStoreSeqs = seqStore->sequencesStart;
-    size_t seqStoreSeqSize = seqStore->sequences - seqStoreSeqs;
-    size_t seqStoreLiteralsSize = (size_t)(seqStore->lit - seqStore->litStart);
-    size_t literalsRead = 0;
-    size_t lastLLSize;
+    const SeqDef* inSeqs = seqStore->sequencesStart;
+    const size_t nbInSequences = (size_t)(seqStore->sequences - inSeqs);
+    const size_t nbInLiterals = (size_t)(seqStore->lit - seqStore->litStart);
 
-    ZSTD_Sequence* outSeqs = &zc->seqCollector.seqStart[zc->seqCollector.seqIndex];
+    ZSTD_Sequence* outSeqs = seqCollector->seqIndex == 0 ? seqCollector->seqStart : seqCollector->seqStart + seqCollector->seqIndex;
+    const size_t nbOutSequences = nbInSequences + 1;
+    size_t nbOutLiterals = 0;
+    Repcodes_t repcodes;
     size_t i;
-    repcodes_t updatedRepcodes;
 
-    assert(zc->seqCollector.seqIndex + 1 < zc->seqCollector.maxSequences);
-    /* Ensure we have enough space for last literals "sequence" */
-    assert(zc->seqCollector.maxSequences >= seqStoreSeqSize + 1);
-    ZSTD_memcpy(updatedRepcodes.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t));
-    for (i = 0; i < seqStoreSeqSize; ++i) {
-        U32 rawOffset = seqStoreSeqs[i].offBase - ZSTD_REP_NUM;
-        outSeqs[i].litLength = seqStoreSeqs[i].litLength;
-        outSeqs[i].matchLength = seqStoreSeqs[i].mlBase + MINMATCH;
+    /* Bounds check that we have enough space for every input sequence
+     * and the block delimiter
+     */
+    assert(seqCollector->seqIndex <= seqCollector->maxSequences);
+    RETURN_ERROR_IF(
+        nbOutSequences > (size_t)(seqCollector->maxSequences - seqCollector->seqIndex),
+        dstSize_tooSmall,
+        "Not enough space to copy sequences");
+
+    ZSTD_memcpy(&repcodes, prevRepcodes, sizeof(repcodes));
+    for (i = 0; i < nbInSequences; ++i) {
+        U32 rawOffset;
+        outSeqs[i].litLength = inSeqs[i].litLength;
+        outSeqs[i].matchLength = inSeqs[i].mlBase + MINMATCH;
         outSeqs[i].rep = 0;
 
+        /* Handle the possible single length >= 64K
+         * There can only be one because we add MINMATCH to every match length,
+         * and blocks are at most 128K.
+         */
         if (i == seqStore->longLengthPos) {
             if (seqStore->longLengthType == ZSTD_llt_literalLength) {
                 outSeqs[i].litLength += 0x10000;
···
             }
         }
 
-        if (seqStoreSeqs[i].offBase <= ZSTD_REP_NUM) {
-            /* Derive the correct offset corresponding to a repcode */
-            outSeqs[i].rep = seqStoreSeqs[i].offBase;
+        /* Determine the raw offset given the offBase, which may be a repcode. */
+        if (OFFBASE_IS_REPCODE(inSeqs[i].offBase)) {
+            const U32 repcode = OFFBASE_TO_REPCODE(inSeqs[i].offBase);
+            assert(repcode > 0);
+            outSeqs[i].rep = repcode;
             if (outSeqs[i].litLength != 0) {
-                rawOffset = updatedRepcodes.rep[outSeqs[i].rep - 1];
+                rawOffset = repcodes.rep[repcode - 1];
             } else {
-                if (outSeqs[i].rep == 3) {
-                    rawOffset = updatedRepcodes.rep[0] - 1;
+                if (repcode == 3) {
+                    assert(repcodes.rep[0] > 1);
+                    rawOffset = repcodes.rep[0] - 1;
                 } else {
-                    rawOffset = updatedRepcodes.rep[outSeqs[i].rep];
+                    rawOffset = repcodes.rep[repcode];
                 }
             }
+        } else {
+            rawOffset = OFFBASE_TO_OFFSET(inSeqs[i].offBase);
         }
         outSeqs[i].offset = rawOffset;
-        /* seqStoreSeqs[i].offset == offCode+1, and ZSTD_updateRep() expects offCode
-           so we provide seqStoreSeqs[i].offset - 1 */
-        ZSTD_updateRep(updatedRepcodes.rep,
-                       seqStoreSeqs[i].offBase - 1,
-                       seqStoreSeqs[i].litLength == 0);
-        literalsRead += outSeqs[i].litLength;
+
+        /* Update repcode history for the sequence */
+        ZSTD_updateRep(repcodes.rep,
+                       inSeqs[i].offBase,
+                       inSeqs[i].litLength == 0);
+
+        nbOutLiterals += outSeqs[i].litLength;
     }
     /* Insert last literals (if any exist) in the block as a sequence with ml == off == 0.
      * If there are no last literals, then we'll emit (of: 0, ml: 0, ll: 0), which is a marker
      * for the block boundary, according to the API.
      */
-    assert(seqStoreLiteralsSize >= literalsRead);
-    lastLLSize = seqStoreLiteralsSize - literalsRead;
-    outSeqs[i].litLength = (U32)lastLLSize;
-    outSeqs[i].matchLength = outSeqs[i].offset = outSeqs[i].rep = 0;
-    seqStoreSeqSize++;
-    zc->seqCollector.seqIndex += seqStoreSeqSize;
+    assert(nbInLiterals >= nbOutLiterals);
+    {
+        const size_t lastLLSize = nbInLiterals - nbOutLiterals;
+        outSeqs[nbInSequences].litLength = (U32)lastLLSize;
+        outSeqs[nbInSequences].matchLength = 0;
+        outSeqs[nbInSequences].offset = 0;
+        assert(nbOutSequences == nbInSequences + 1);
+    }
+    seqCollector->seqIndex += nbOutSequences;
+    assert(seqCollector->seqIndex <= seqCollector->maxSequences);
+
+    return 0;
+}
+
+size_t ZSTD_sequenceBound(size_t srcSize) {
+    const size_t maxNbSeq = (srcSize / ZSTD_MINMATCH_MIN) + 1;
+    const size_t maxNbDelims = (srcSize / ZSTD_BLOCKSIZE_MAX_MIN) + 1;
+    return maxNbSeq + maxNbDelims;
 }
 
 size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
                               size_t outSeqsSize, const void* src, size_t srcSize)
 {
     const size_t dstCapacity = ZSTD_compressBound(srcSize);
-    void* dst = ZSTD_customMalloc(dstCapacity, ZSTD_defaultCMem);
+    void* dst; /* Make C90 happy. */
     SeqCollector seqCollector;
+    {
+        int targetCBlockSize;
+        FORWARD_IF_ERROR(ZSTD_CCtx_getParameter(zc, ZSTD_c_targetCBlockSize, &targetCBlockSize), "");
+        RETURN_ERROR_IF(targetCBlockSize != 0, parameter_unsupported, "targetCBlockSize != 0");
+    }
+    {
+        int nbWorkers;
+        FORWARD_IF_ERROR(ZSTD_CCtx_getParameter(zc, ZSTD_c_nbWorkers, &nbWorkers), "");
+        RETURN_ERROR_IF(nbWorkers != 0, parameter_unsupported, "nbWorkers != 0");
+    }
 
+    dst = ZSTD_customMalloc(dstCapacity, ZSTD_defaultCMem);
     RETURN_ERROR_IF(dst == NULL, memory_allocation, "NULL pointer!");
 
     seqCollector.collectSequences = 1;
···
     seqCollector.maxSequences = outSeqsSize;
     zc->seqCollector = seqCollector;
 
-    ZSTD_compress2(zc, dst, dstCapacity, src, srcSize);
-    ZSTD_customFree(dst, ZSTD_defaultCMem);
+    {
+        const size_t ret = ZSTD_compress2(zc, dst, dstCapacity, src, srcSize);
+        ZSTD_customFree(dst, ZSTD_defaultCMem);
+        FORWARD_IF_ERROR(ret, "ZSTD_compress2 failed");
+    }
+    assert(zc->seqCollector.seqIndex <= ZSTD_sequenceBound(srcSize));
     return zc->seqCollector.seqIndex;
 }
 
···
     const size_t unrollMask = unrollSize - 1;
     const size_t prefixLength = length & unrollMask;
     size_t i;
-    size_t u;
     if (length == 1) return 1;
     /* Check if prefix is RLE first before using unrolled loop */
     if (prefixLength && ZSTD_count(ip+1, ip, ip+prefixLength) != prefixLength-1) {
         return 0;
     }
     for (i = prefixLength; i != length; i += unrollSize) {
+        size_t u;
         for (u = 0; u < unrollSize; u += sizeof(size_t)) {
             if (MEM_readST(ip + i + u) != valueST) {
                 return 0;
-            }
-        }
-    }
+    }   }   }
     return 1;
 }
 
···
  * This is just a heuristic based on the compressibility.
  * It may return both false positives and false negatives.
  */
-static int ZSTD_maybeRLE(seqStore_t const* seqStore)
+static int ZSTD_maybeRLE(SeqStore_t const* seqStore)
 {
     size_t const nbSeqs = (size_t)(seqStore->sequences - seqStore->sequencesStart);
     size_t const nbLits = (size_t)(seqStore->lit - seqStore->litStart);
···
     return nbSeqs < 4 && nbLits < 10;
 }
 
-static void ZSTD_blockState_confirmRepcodesAndEntropyTables(ZSTD_blockState_t* const bs)
+static void
+ZSTD_blockState_confirmRepcodesAndEntropyTables(ZSTD_blockState_t* const bs)
 {
     ZSTD_compressedBlockState_t* const tmp = bs->prevCBlock;
     bs->prevCBlock = bs->nextCBlock;
···
 }
 
 /* Writes the block header */
-static void writeBlockHeader(void* op, size_t cSize, size_t blockSize, U32 lastBlock) {
+static void
+writeBlockHeader(void* op, size_t cSize, size_t blockSize, U32 lastBlock)
+{
     U32 const cBlockHeader = cSize == 1 ?
                         lastBlock + (((U32)bt_rle)<<1) + (U32)(blockSize << 3) :
                         lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3);
     MEM_writeLE24(op, cBlockHeader);
-    DEBUGLOG(3, "writeBlockHeader: cSize: %zu blockSize: %zu lastBlock: %u", cSize, blockSize, lastBlock);
+    DEBUGLOG(5, "writeBlockHeader: cSize: %zu blockSize: %zu lastBlock: %u", cSize, blockSize, lastBlock);
 }
 
 /* ZSTD_buildBlockEntropyStats_literals() :
···
  *  Stores literals block type (raw, rle, compressed, repeat) and
  *  huffman description table to hufMetadata.
  *  Requires ENTROPY_WORKSPACE_SIZE workspace
- *  @return : size of huffman description table or error code */
-static size_t ZSTD_buildBlockEntropyStats_literals(void* const src, size_t srcSize,
-                                            const ZSTD_hufCTables_t* prevHuf,
-                                                  ZSTD_hufCTables_t* nextHuf,
-                                                  ZSTD_hufCTablesMetadata_t* hufMetadata,
-                                                  const int literalsCompressionIsDisabled,
-                                                  void* workspace, size_t wkspSize)
+ * @return : size of huffman description table, or an error code
+ */
+static size_t
+ZSTD_buildBlockEntropyStats_literals(void* const src, size_t srcSize,
+                               const ZSTD_hufCTables_t* prevHuf,
+                                     ZSTD_hufCTables_t* nextHuf,
+                                     ZSTD_hufCTablesMetadata_t* hufMetadata,
+                               const int literalsCompressionIsDisabled,
+                                     void* workspace, size_t wkspSize,
+                                     int hufFlags)
 {
     BYTE* const wkspStart = (BYTE*)workspace;
     BYTE* const wkspEnd = wkspStart + wkspSize;
···
     unsigned* const countWksp = (unsigned*)workspace;
     const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned);
     BYTE* const nodeWksp = countWkspStart + countWkspSize;
-    const size_t nodeWkspSize = wkspEnd-nodeWksp;
+    const size_t nodeWkspSize = (size_t)(wkspEnd - nodeWksp);
     unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX;
-    unsigned huffLog = HUF_TABLELOG_DEFAULT;
+    unsigned huffLog = LitHufLog;
     HUF_repeat repeat = prevHuf->repeatMode;
     DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_literals (srcSize=%zu)", srcSize);
 
···
 
     /* small ? don't even attempt compression (speed opt) */
 #ifndef COMPRESS_LITERALS_SIZE_MIN
-#define COMPRESS_LITERALS_SIZE_MIN 63
+# define COMPRESS_LITERALS_SIZE_MIN 63  /* heuristic */
 #endif
     {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
         if (srcSize <= minLitSize) {
             DEBUGLOG(5, "set_basic - too small");
             hufMetadata->hType = set_basic;
             return 0;
-        }
-    }
+    }   }
 
     /* Scan input and build symbol stats */
-    {   size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)src, srcSize, workspace, wkspSize);
+    {   size_t const largest =
+            HIST_count_wksp (countWksp, &maxSymbolValue,
+                            (const BYTE*)src, srcSize,
+                            workspace, wkspSize);
         FORWARD_IF_ERROR(largest, "HIST_count_wksp failed");
         if (largest == srcSize) {
+            /* only one literal symbol */
             DEBUGLOG(5, "set_rle");
             hufMetadata->hType = set_rle;
             return 0;
         }
         if (largest <= (srcSize >> 7)+4) {
+            /* heuristic: likely not compressible */
             DEBUGLOG(5, "set_basic - no gain");
             hufMetadata->hType = set_basic;
             return 0;
-        }
-    }
+    }   }
 
     /* Validate the previous Huffman table */
-    if (repeat == HUF_repeat_check && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) {
+    if (repeat == HUF_repeat_check
+      && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) {
         repeat = HUF_repeat_none;
     }
 
     /* Build Huffman Tree */
     ZSTD_memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable));
-    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
+    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue, nodeWksp, nodeWkspSize, nextHuf->CTable, countWksp, hufFlags);
+    assert(huffLog <= LitHufLog);
     {   size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp,
                                                     maxSymbolValue, huffLog,
                                                     nodeWksp, nodeWkspSize);
         FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp");
         huffLog = (U32)maxBits;
-        {   /* Build and write the CTable */
-            size_t const newCSize = HUF_estimateCompressedSize(
-                    (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue);
-            size_t const hSize = HUF_writeCTable_wksp(
-                    hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer),
-                    (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog,
-                    nodeWksp, nodeWkspSize);
-            /* Check against repeating the previous CTable */
-            if (repeat != HUF_repeat_none) {
-                size_t const oldCSize = HUF_estimateCompressedSize(
-                        (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue);
-                if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) {
-                    DEBUGLOG(5, "set_repeat - smaller");
-                    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-                    hufMetadata->hType = set_repeat;
-                    return 0;
-                }
-            }
-            if (newCSize + hSize >= srcSize) {
-                DEBUGLOG(5, "set_basic - no gains");
+    }
+    {   /* Build and write the CTable */
+        size_t const newCSize = HUF_estimateCompressedSize(
+                (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue);
+        size_t const hSize = HUF_writeCTable_wksp(
+                hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer),
+                (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog,
+                nodeWksp, nodeWkspSize);
+        /* Check against repeating the previous CTable */
+        if (repeat != HUF_repeat_none) {
+            size_t const oldCSize = HUF_estimateCompressedSize(
+                    (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue);
+            if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) {
+                DEBUGLOG(5, "set_repeat - smaller");
                 ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-                hufMetadata->hType = set_basic;
+                hufMetadata->hType = set_repeat;
                 return 0;
-            }
-            DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize);
-            hufMetadata->hType = set_compressed;
-            nextHuf->repeatMode = HUF_repeat_check;
-            return hSize;
+        }   }
+        if (newCSize + hSize >= srcSize) {
+            DEBUGLOG(5, "set_basic - no gains");
+            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+            hufMetadata->hType = set_basic;
+            return 0;
         }
+        DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize);
+        hufMetadata->hType = set_compressed;
+        nextHuf->repeatMode = HUF_repeat_check;
+        return hSize;
     }
 }
 
···
  * and updates nextEntropy to the appropriate repeatMode.
  */
 static ZSTD_symbolEncodingTypeStats_t
-ZSTD_buildDummySequencesStatistics(ZSTD_fseCTables_t* nextEntropy) {
-    ZSTD_symbolEncodingTypeStats_t stats = {set_basic, set_basic, set_basic, 0, 0};
+ZSTD_buildDummySequencesStatistics(ZSTD_fseCTables_t* nextEntropy)
+{
+    ZSTD_symbolEncodingTypeStats_t stats = {set_basic, set_basic, set_basic, 0, 0, 0};
     nextEntropy->litlength_repeatMode = FSE_repeat_none;
     nextEntropy->offcode_repeatMode = FSE_repeat_none;
     nextEntropy->matchlength_repeatMode = FSE_repeat_none;
···
  *  Builds entropy for the sequences.
  *  Stores symbol compression modes and fse table to fseMetadata.
  *  Requires ENTROPY_WORKSPACE_SIZE wksp.
- *  @return : size of fse tables or error code */
-static size_t ZSTD_buildBlockEntropyStats_sequences(seqStore_t* seqStorePtr,
-                                              const ZSTD_fseCTables_t* prevEntropy,
-                                                    ZSTD_fseCTables_t* nextEntropy,
-                                              const ZSTD_CCtx_params* cctxParams,
-                                                    ZSTD_fseCTablesMetadata_t* fseMetadata,
-                                                    void* workspace, size_t wkspSize)
+ * @return : size of fse tables or error code */
+static size_t
+ZSTD_buildBlockEntropyStats_sequences(
+                const SeqStore_t* seqStorePtr,
+                const ZSTD_fseCTables_t* prevEntropy,
+                      ZSTD_fseCTables_t* nextEntropy,
+                const ZSTD_CCtx_params* cctxParams,
+                      ZSTD_fseCTablesMetadata_t* fseMetadata,
+                      void* workspace, size_t wkspSize)
 {
     ZSTD_strategy const strategy = cctxParams->cParams.strategy;
-    size_t const nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart;
+    size_t const nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
     BYTE* const ostart = fseMetadata->fseTablesBuffer;
     BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer);
     BYTE* op = ostart;
···
                                           entropyWorkspace, entropyWorkspaceSize)
                        : ZSTD_buildDummySequencesStatistics(nextEntropy);
     FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!");
-    fseMetadata->llType = (symbolEncodingType_e) stats.LLtype;
-    fseMetadata->ofType = (symbolEncodingType_e) stats.Offtype;
-    fseMetadata->mlType = (symbolEncodingType_e) stats.MLtype;
+    fseMetadata->llType = (SymbolEncodingType_e) stats.LLtype;
+    fseMetadata->ofType = (SymbolEncodingType_e) stats.Offtype;
+    fseMetadata->mlType = (SymbolEncodingType_e) stats.MLtype;
     fseMetadata->lastCountSize = stats.lastCountSize;
     return stats.size;
 }
···
 /* ZSTD_buildBlockEntropyStats() :
  *  Builds entropy for the block.
  *  Requires workspace size ENTROPY_WORKSPACE_SIZE
- *
- *  @return : 0 on success or error code
+ * @return : 0 on success, or an error code
+ *  Note : also employed in superblock
  */
-size_t ZSTD_buildBlockEntropyStats(seqStore_t* seqStorePtr,
-                             const ZSTD_entropyCTables_t* prevEntropy,
-                                   ZSTD_entropyCTables_t* nextEntropy,
-                             const ZSTD_CCtx_params* cctxParams,
-                                   ZSTD_entropyCTablesMetadata_t* entropyMetadata,
-                                   void* workspace, size_t wkspSize)
+size_t ZSTD_buildBlockEntropyStats(
+            const SeqStore_t* seqStorePtr,
+            const ZSTD_entropyCTables_t* prevEntropy,
+                  ZSTD_entropyCTables_t* nextEntropy,
+            const ZSTD_CCtx_params* cctxParams,
+                  ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                  void* workspace, size_t wkspSize)
 {
-    size_t const litSize = seqStorePtr->lit - seqStorePtr->litStart;
+    size_t const litSize = (size_t)(seqStorePtr->lit - seqStorePtr->litStart);
+    int const huf_useOptDepth = (cctxParams->cParams.strategy >= HUF_OPTIMAL_DEPTH_THRESHOLD);
+    int const hufFlags = huf_useOptDepth ? HUF_flags_optimalDepth : 0;
+
     entropyMetadata->hufMetadata.hufDesSize =
         ZSTD_buildBlockEntropyStats_literals(seqStorePtr->litStart, litSize,
                                             &prevEntropy->huf, &nextEntropy->huf,
                                             &entropyMetadata->hufMetadata,
                                             ZSTD_literalsCompressionIsDisabled(cctxParams),
-                                            workspace, wkspSize);
+                                            workspace, wkspSize, hufFlags);
+
     FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildBlockEntropyStats_literals failed");
     entropyMetadata->fseMetadata.fseTablesSize =
         ZSTD_buildBlockEntropyStats_sequences(seqStorePtr,
···
 }
 
 /* Returns the size estimate for the literals section (header + content) of a block */
-static size_t ZSTD_estimateBlockSize_literal(const BYTE* literals, size_t litSize,
-                                                const ZSTD_hufCTables_t* huf,
-                                                const ZSTD_hufCTablesMetadata_t* hufMetadata,
-                                                void* workspace, size_t wkspSize,
-                                                int writeEntropy)
+static size_t
+ZSTD_estimateBlockSize_literal(const BYTE* literals, size_t litSize,
+                               const ZSTD_hufCTables_t* huf,
+                               const ZSTD_hufCTablesMetadata_t* hufMetadata,
+                               void* workspace, size_t wkspSize,
+                               int writeEntropy)
 {
     unsigned* const countWksp = (unsigned*)workspace;
     unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX;
···
 }
 
 /* Returns the size estimate for the FSE-compressed symbols (of, ml, ll) of a block */
-static size_t ZSTD_estimateBlockSize_symbolType(symbolEncodingType_e type,
-                        const BYTE* codeTable, size_t nbSeq, unsigned maxCode,
-                        const FSE_CTable* fseCTable,
-                        const U8* additionalBits,
-                        short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
-                        void* workspace, size_t wkspSize)
+static size_t
+ZSTD_estimateBlockSize_symbolType(SymbolEncodingType_e type,
+                    const BYTE* codeTable, size_t nbSeq, unsigned maxCode,
+                    const FSE_CTable* fseCTable,
+                    const U8* additionalBits,
+                    short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
+                    void* workspace, size_t wkspSize)
 {
     unsigned* const countWksp = (unsigned*)workspace;
     const BYTE* ctp = codeTable;
···
 }
 
 /* Returns the size estimate for the sequences section (header + content) of a block */
-static size_t ZSTD_estimateBlockSize_sequences(const BYTE* ofCodeTable,
-                                                  const BYTE* llCodeTable,
-                                                  const BYTE* mlCodeTable,
-                                                  size_t nbSeq,
-                                                  const ZSTD_fseCTables_t* fseTables,
-                                                  const ZSTD_fseCTablesMetadata_t* fseMetadata,
-                                                  void* workspace, size_t wkspSize,
-                                                  int writeEntropy)
+static size_t
+ZSTD_estimateBlockSize_sequences(const BYTE* ofCodeTable,
+                                 const BYTE* llCodeTable,
+                                 const BYTE* mlCodeTable,
+                                 size_t nbSeq,
+                                 const ZSTD_fseCTables_t* fseTables,
+                                 const ZSTD_fseCTablesMetadata_t* fseMetadata,
+                                 void* workspace, size_t wkspSize,
+                                 int writeEntropy)
 {
     size_t sequencesSectionHeaderSize = 1 /* seqHead */ + 1 /* min seqSize size */ + (nbSeq >= 128) + (nbSeq >= LONGNBSEQ);
     size_t cSeqSizeEstimate = 0;
     cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, nbSeq, MaxOff,
-                                         fseTables->offcodeCTable, NULL,
-                                         OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
-                                         workspace, wkspSize);
+                                    fseTables->offcodeCTable, NULL,
+                                    OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
+                                    workspace, wkspSize);
     cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->llType, llCodeTable, nbSeq, MaxLL,
-                                         fseTables->litlengthCTable, LL_bits,
-                                         LL_defaultNorm, LL_defaultNormLog, MaxLL,
-                                         workspace, wkspSize);
+                                    fseTables->litlengthCTable, LL_bits,
+                                    LL_defaultNorm, LL_defaultNormLog, MaxLL,
+                                    workspace, wkspSize);
     cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, nbSeq, MaxML,
-                                         fseTables->matchlengthCTable, ML_bits,
-                                         ML_defaultNorm, ML_defaultNormLog, MaxML,
-                                         workspace, wkspSize);
+                                    fseTables->matchlengthCTable, ML_bits,
+                                    ML_defaultNorm, ML_defaultNormLog, MaxML,
+                                    workspace, wkspSize);
     if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize;
     return cSeqSizeEstimate + sequencesSectionHeaderSize;
 }
 
 /* Returns the size estimate for a given stream of literals, of, ll, ml */
-static size_t ZSTD_estimateBlockSize(const BYTE* literals, size_t litSize,
-                                     const BYTE* ofCodeTable,
-                                     const BYTE* llCodeTable,
-                                     const BYTE* mlCodeTable,
-                                     size_t nbSeq,
-                                     const ZSTD_entropyCTables_t* entropy,
-                                     const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
-                                     void* workspace, size_t wkspSize,
-                                     int writeLitEntropy, int writeSeqEntropy) {
+static size_t
+ZSTD_estimateBlockSize(const BYTE* literals, size_t litSize,
+                       const BYTE* ofCodeTable,
+                       const BYTE* llCodeTable,
+                       const BYTE* mlCodeTable,
+                       size_t nbSeq,
+                       const ZSTD_entropyCTables_t* entropy,
+                       const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                       void* workspace, size_t wkspSize,
+                       int writeLitEntropy, int writeSeqEntropy)
+{
     size_t const literalsSize = ZSTD_estimateBlockSize_literal(literals, litSize,
-                                                         &entropy->huf, &entropyMetadata->hufMetadata,
-                                                         workspace, wkspSize, writeLitEntropy);
+                                    &entropy->huf, &entropyMetadata->hufMetadata,
+                                    workspace, wkspSize, writeLitEntropy);
     size_t const seqSize = ZSTD_estimateBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable,
-                                                         nbSeq, &entropy->fse, &entropyMetadata->fseMetadata,
-                                                         workspace, wkspSize, writeSeqEntropy);
+                                    nbSeq, &entropy->fse, &entropyMetadata->fseMetadata,
+                                    workspace, wkspSize, writeSeqEntropy);
     return seqSize + literalsSize + ZSTD_blockHeaderSize;
 }
 
 /* Builds entropy statistics and uses them for blocksize estimation.
  *
- * Returns the estimated compressed size of the seqStore, or a zstd error.
+ * @return: estimated compressed size of the seqStore, or a zstd error.
  */
-static size_t ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(seqStore_t* seqStore, ZSTD_CCtx* zc) {
-    ZSTD_entropyCTablesMetadata_t* entropyMetadata = &zc->blockSplitCtx.entropyMetadata;
+static size_t
+ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(SeqStore_t* seqStore, ZSTD_CCtx* zc)
+{
+    ZSTD_entropyCTablesMetadata_t* const entropyMetadata = &zc->blockSplitCtx.entropyMetadata;
     DEBUGLOG(6, "ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize()");
     FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(seqStore,
                     &zc->blockState.prevCBlock->entropy,
                     &zc->blockState.nextCBlock->entropy,
                     &zc->appliedParams,
                     entropyMetadata,
-                    zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */), "");
-    return ZSTD_estimateBlockSize(seqStore->litStart, (size_t)(seqStore->lit - seqStore->litStart),
+                    zc->tmpWorkspace, zc->tmpWkspSize), "");
+    return ZSTD_estimateBlockSize(
+                    seqStore->litStart, (size_t)(seqStore->lit - seqStore->litStart),
                     seqStore->ofCode, seqStore->llCode, seqStore->mlCode,
                     (size_t)(seqStore->sequences - seqStore->sequencesStart),
-                    &zc->blockState.nextCBlock->entropy, entropyMetadata, zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE,
+                    &zc->blockState.nextCBlock->entropy,
+                    entropyMetadata,
+                    zc->tmpWorkspace, zc->tmpWkspSize,
                     (int)(entropyMetadata->hufMetadata.hType == set_compressed), 1);
 }
 
 /* Returns literals bytes represented in a seqStore */
-static size_t ZSTD_countSeqStoreLiteralsBytes(const seqStore_t* const seqStore) {
+static size_t ZSTD_countSeqStoreLiteralsBytes(const SeqStore_t* const seqStore)
+{
     size_t literalsBytes = 0;
-    size_t const nbSeqs = seqStore->sequences - seqStore->sequencesStart;
+    size_t const nbSeqs = (size_t)(seqStore->sequences - seqStore->sequencesStart);
     size_t i;
     for (i = 0; i < nbSeqs; ++i) {
-        seqDef seq = seqStore->sequencesStart[i];
+        SeqDef const seq = seqStore->sequencesStart[i];
         literalsBytes += seq.litLength;
         if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_literalLength) {
             literalsBytes += 0x10000;
-        }
-    }
+    }   }
     return literalsBytes;
 }
 
 /* Returns match bytes represented in a seqStore */
-static size_t ZSTD_countSeqStoreMatchBytes(const seqStore_t* const seqStore) {
+static size_t ZSTD_countSeqStoreMatchBytes(const SeqStore_t* const seqStore)
+{
     size_t matchBytes = 0;
-    size_t const nbSeqs = seqStore->sequences - seqStore->sequencesStart;
+    size_t const nbSeqs = (size_t)(seqStore->sequences - seqStore->sequencesStart);
     size_t i;
     for (i = 0; i < nbSeqs; ++i) {
-        seqDef seq = seqStore->sequencesStart[i];
+        SeqDef seq = seqStore->sequencesStart[i];
         matchBytes += seq.mlBase + MINMATCH;
         if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_matchLength) {
             matchBytes += 0x10000;
-        }
-    }
+    }   }
     return matchBytes;
 }
 
 /* Derives the seqStore that is a chunk of the originalSeqStore from [startIdx, endIdx).
  * Stores the result in resultSeqStore.
  */
-static void ZSTD_deriveSeqStoreChunk(seqStore_t* resultSeqStore,
-                               const seqStore_t* originalSeqStore,
-                                     size_t startIdx, size_t endIdx) {
-    BYTE* const litEnd = originalSeqStore->lit;
-    size_t literalsBytes;
-    size_t literalsBytesPreceding = 0;
-
+static void ZSTD_deriveSeqStoreChunk(SeqStore_t* resultSeqStore,
+                               const SeqStore_t* originalSeqStore,
+                                     size_t startIdx, size_t endIdx)
+{
     *resultSeqStore = *originalSeqStore;
     if (startIdx > 0) {
         resultSeqStore->sequences = originalSeqStore->sequencesStart + startIdx;
-        literalsBytesPreceding = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
+        resultSeqStore->litStart += ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
     }
 
     /* Move longLengthPos into the correct position if necessary */
···
     }
     resultSeqStore->sequencesStart = originalSeqStore->sequencesStart + startIdx;
     resultSeqStore->sequences = originalSeqStore->sequencesStart + endIdx;
-    literalsBytes = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
-    resultSeqStore->litStart += literalsBytesPreceding;
     if (endIdx == (size_t)(originalSeqStore->sequences - originalSeqStore->sequencesStart)) {
         /* This accounts for possible last literals if the derived chunk reaches the end of the block */
-        resultSeqStore->lit = litEnd;
+        assert(resultSeqStore->lit == originalSeqStore->lit);
     } else {
-        resultSeqStore->lit = resultSeqStore->litStart+literalsBytes;
+        size_t const literalsBytes = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
+        resultSeqStore->lit = resultSeqStore->litStart + literalsBytes;
     }
     resultSeqStore->llCode += startIdx;
     resultSeqStore->mlCode += startIdx;
···
 }
 
 /*
- * Returns the raw offset represented by the combination of offCode, ll0, and repcode history.
- * offCode must represent a repcode in the numeric representation of ZSTD_storeSeq().
+ * Returns the raw offset represented by the combination of offBase, ll0, and repcode history.
+ * offBase must represent a repcode in the numeric representation of ZSTD_storeSeq().
  */
 static U32
-ZSTD_resolveRepcodeToRawOffset(const U32 rep[ZSTD_REP_NUM], const U32 offCode, const U32 ll0)
+ZSTD_resolveRepcodeToRawOffset(const U32 rep[ZSTD_REP_NUM], const U32 offBase, const U32 ll0)
 {
-    U32 const adjustedOffCode = STORED_REPCODE(offCode) - 1 + ll0;  /* [ 0 - 3 ] */
-    assert(STORED_IS_REPCODE(offCode));
-    if (adjustedOffCode == ZSTD_REP_NUM) {
-        /* litlength == 0 and offCode == 2 implies selection of first repcode - 1 */
-        assert(rep[0] > 0);
+    U32 const adjustedRepCode = OFFBASE_TO_REPCODE(offBase) - 1 + ll0;  /* [ 0 - 3 ] */
+    assert(OFFBASE_IS_REPCODE(offBase));
+    if (adjustedRepCode == ZSTD_REP_NUM) {
+        assert(ll0);
+        /* litlength == 0 and offCode == 2 implies selection of first repcode - 1
+         * This is only valid if it results in a valid offset value, aka > 0.
+         * Note : it may happen that `rep[0]==1` in exceptional circumstances.
+         * In which case this function will return 0, which is an invalid offset.
+         * It's not an issue though, since this value will be
+         * compared and discarded within ZSTD_seqStore_resolveOffCodes().
+         */
         return rep[0] - 1;
     }
-    return rep[adjustedOffCode];
+    return rep[adjustedRepCode];
 }
 
 /*
···
  *        1-3 : repcode 1-3
  *        4+ : real_offset+3
  */
-static void ZSTD_seqStore_resolveOffCodes(repcodes_t* const dRepcodes, repcodes_t* const cRepcodes,
-                                          seqStore_t* const seqStore, U32 const nbSeq) {
+static void
+ZSTD_seqStore_resolveOffCodes(Repcodes_t* const dRepcodes, Repcodes_t* const cRepcodes,
+                        const SeqStore_t* const seqStore, U32 const nbSeq)
+{
     U32 idx = 0;
+    U32 const longLitLenIdx = seqStore->longLengthType == ZSTD_llt_literalLength ? seqStore->longLengthPos : nbSeq;
     for (; idx < nbSeq; ++idx) {
-        seqDef* const seq = seqStore->sequencesStart + idx;
-        U32 const ll0 = (seq->litLength == 0);
-        U32 const offCode = OFFBASE_TO_STORED(seq->offBase);
-        assert(seq->offBase > 0);
-        if (STORED_IS_REPCODE(offCode)) {
-            U32 const dRawOffset = ZSTD_resolveRepcodeToRawOffset(dRepcodes->rep, offCode, ll0);
-            U32 const cRawOffset = ZSTD_resolveRepcodeToRawOffset(cRepcodes->rep, offCode, ll0);
+        SeqDef* const seq = seqStore->sequencesStart + idx;
+        U32 const ll0 = (seq->litLength == 0) && (idx != longLitLenIdx);
+        U32 const offBase = seq->offBase;
+        assert(offBase > 0);
+        if (OFFBASE_IS_REPCODE(offBase)) {
+            U32 const dRawOffset = ZSTD_resolveRepcodeToRawOffset(dRepcodes->rep, offBase, ll0);
+            U32 const cRawOffset = ZSTD_resolveRepcodeToRawOffset(cRepcodes->rep, offBase, ll0);
             /* Adjust simulated decompression repcode history if we come across a mismatch. Replace
              * the repcode with the offset it actually references, determined by the compression
              * repcode history.
              */
             if (dRawOffset != cRawOffset) {
-                seq->offBase = cRawOffset + ZSTD_REP_NUM;
+                seq->offBase = OFFSET_TO_OFFBASE(cRawOffset);
             }
         }
         /* Compression repcode history is always updated with values directly from the unmodified seqStore.
          * Decompression repcode history may use modified seq->offset value taken from compression repcode history.
          */
-        ZSTD_updateRep(dRepcodes->rep, OFFBASE_TO_STORED(seq->offBase), ll0);
-        ZSTD_updateRep(cRepcodes->rep, offCode, ll0);
+        ZSTD_updateRep(dRepcodes->rep, seq->offBase, ll0);
+        ZSTD_updateRep(cRepcodes->rep, offBase, ll0);
     }
 }
 
···
  * Returns the total size of that block (including header) or a ZSTD error code.
  */
 static size_t
-ZSTD_compressSeqStore_singleBlock(ZSTD_CCtx* zc, seqStore_t* const seqStore,
-                                  repcodes_t* const dRep, repcodes_t* const cRep,
+ZSTD_compressSeqStore_singleBlock(ZSTD_CCtx* zc,
+                            const SeqStore_t* const seqStore,
+                                  Repcodes_t* const dRep, Repcodes_t* const cRep,
                                   void* dst, size_t dstCapacity,
-                                  const void* src, size_t srcSize,
+                            const void* src, size_t srcSize,
                                   U32 lastBlock, U32 isPartition)
 {
     const U32 rleMaxLength = 25;
···
     size_t cSeqsSize;
 
     /* In case of an RLE or raw block, the simulated decompression repcode history must be reset */
-    repcodes_t const dRepOriginal = *dRep;
+    Repcodes_t const dRepOriginal = *dRep;
     DEBUGLOG(5, "ZSTD_compressSeqStore_singleBlock");
     if (isPartition)
         ZSTD_seqStore_resolveOffCodes(dRep, cRep, seqStore, (U32)(seqStore->sequences - seqStore->sequencesStart));
···
                 &zc->appliedParams,
                 op + ZSTD_blockHeaderSize, dstCapacity - ZSTD_blockHeaderSize,
                 srcSize,
-                zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
+                zc->tmpWorkspace, zc->tmpWkspSize /* statically allocated in resetCCtx */,
                 zc->bmi2);
     FORWARD_IF_ERROR(cSeqsSize, "ZSTD_entropyCompressSeqStore failed!");
 
···
         cSeqsSize = 1;
     }
 
+    /* Sequence collection not supported when block splitting */
     if (zc->seqCollector.collectSequences) {
-        ZSTD_copyBlockSequences(zc);
+        FORWARD_IF_ERROR(ZSTD_copyBlockSequences(&zc->seqCollector, seqStore, dRepOriginal.rep), "copyBlockSequences failed");
         ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
         return 0;
     }
···
     if (cSeqsSize == 0) {
         cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, srcSize, lastBlock);
         FORWARD_IF_ERROR(cSize, "Nocompress block failed");
-        DEBUGLOG(4, "Writing out nocompress block, size: %zu", cSize);
+        DEBUGLOG(5, "Writing out nocompress block, size: %zu", cSize);
         *dRep = dRepOriginal; /* reset simulated decompression repcode history */
     } else if (cSeqsSize == 1) {
         cSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, srcSize, lastBlock);
         FORWARD_IF_ERROR(cSize, "RLE compress block failed");
-        DEBUGLOG(4, "Writing out RLE block, size: %zu", cSize);
+        DEBUGLOG(5, "Writing out RLE block, size: %zu", cSize);
         *dRep = dRepOriginal; /* reset simulated decompression repcode history */
     } else {
         ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
         writeBlockHeader(op, cSeqsSize, srcSize, lastBlock);
         cSize = ZSTD_blockHeaderSize + cSeqsSize;
-        DEBUGLOG(4, "Writing out compressed block, size: %zu", cSize);
+        DEBUGLOG(5, "Writing out compressed block, size: %zu", cSize);
     }
 
     if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
···
 
 /* Helper function to perform the recursive search for block splits.
  * Estimates the cost of seqStore prior to split, and estimates the cost of splitting the sequences in half.
- * If advantageous to split, then we recurse down the two sub-blocks. If not, or if an error occurred in estimation, then
- * we do not recurse.
+ * If advantageous to split, then we recurse down the two sub-blocks.
+ * If not, or if an error occurred in estimation, then we do not recurse.
  *
- * Note: The recursion depth is capped by a heuristic minimum number of sequences, defined by MIN_SEQUENCES_BLOCK_SPLITTING.
+ * Note: The recursion depth is capped by a heuristic minimum number of sequences,
+ * defined by MIN_SEQUENCES_BLOCK_SPLITTING.
  * In theory, this means the absolute largest recursion depth is 10 == log2(maxNbSeqInBlock/MIN_SEQUENCES_BLOCK_SPLITTING).
  * In practice, recursion depth usually doesn't go beyond 4.
  *
- * Furthermore, the number of splits is capped by ZSTD_MAX_NB_BLOCK_SPLITS. At ZSTD_MAX_NB_BLOCK_SPLITS == 196 with the current existing blockSize
+ * Furthermore, the number of splits is capped by ZSTD_MAX_NB_BLOCK_SPLITS.
+ * At ZSTD_MAX_NB_BLOCK_SPLITS == 196 with the current existing blockSize
  * maximum of 128 KB, this value is actually impossible to reach.
  */
 static void
 ZSTD_deriveBlockSplitsHelper(seqStoreSplits* splits, size_t startIdx, size_t endIdx,
-                             ZSTD_CCtx* zc, const seqStore_t* origSeqStore)
+                             ZSTD_CCtx* zc, const SeqStore_t* origSeqStore)
 {
-    seqStore_t* fullSeqStoreChunk = &zc->blockSplitCtx.fullSeqStoreChunk;
-    seqStore_t* firstHalfSeqStore = &zc->blockSplitCtx.firstHalfSeqStore;
-    seqStore_t* secondHalfSeqStore = &zc->blockSplitCtx.secondHalfSeqStore;
+    SeqStore_t* const fullSeqStoreChunk = &zc->blockSplitCtx.fullSeqStoreChunk;
+    SeqStore_t* const firstHalfSeqStore = &zc->blockSplitCtx.firstHalfSeqStore;
+    SeqStore_t* const secondHalfSeqStore = &zc->blockSplitCtx.secondHalfSeqStore;
     size_t estimatedOriginalSize;
     size_t estimatedFirstHalfSize;
     size_t estimatedSecondHalfSize;
     size_t midIdx = (startIdx + endIdx)/2;
 
+    DEBUGLOG(5, "ZSTD_deriveBlockSplitsHelper: startIdx=%zu endIdx=%zu", startIdx, endIdx);
+    assert(endIdx >= startIdx);
     if (endIdx - startIdx < MIN_SEQUENCES_BLOCK_SPLITTING || splits->idx >= ZSTD_MAX_NB_BLOCK_SPLITS) {
-        DEBUGLOG(6, "ZSTD_deriveBlockSplitsHelper: Too few sequences");
+        DEBUGLOG(6, "ZSTD_deriveBlockSplitsHelper: Too few sequences (%zu)", endIdx - startIdx);
         return;
     }
-    DEBUGLOG(4, "ZSTD_deriveBlockSplitsHelper: startIdx=%zu endIdx=%zu", startIdx, endIdx);
     ZSTD_deriveSeqStoreChunk(fullSeqStoreChunk, origSeqStore, startIdx, endIdx);
     ZSTD_deriveSeqStoreChunk(firstHalfSeqStore, origSeqStore, startIdx, midIdx);
     ZSTD_deriveSeqStoreChunk(secondHalfSeqStore, origSeqStore, midIdx, endIdx);
     estimatedOriginalSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(fullSeqStoreChunk, zc);
     estimatedFirstHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(firstHalfSeqStore, zc);
     estimatedSecondHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(secondHalfSeqStore, zc);
-    DEBUGLOG(4, "Estimated original block size: %zu -- First half split: %zu -- Second half split: %zu",
+    DEBUGLOG(5, "Estimated original block size: %zu -- First half split: %zu -- Second half split: %zu",
              estimatedOriginalSize, estimatedFirstHalfSize, estimatedSecondHalfSize);
     if (ZSTD_isError(estimatedOriginalSize) || ZSTD_isError(estimatedFirstHalfSize) || ZSTD_isError(estimatedSecondHalfSize)) {
         return;
     }
     if (estimatedFirstHalfSize + estimatedSecondHalfSize < estimatedOriginalSize) {
+        DEBUGLOG(5, "split decided at seqNb:%zu", midIdx);
         ZSTD_deriveBlockSplitsHelper(splits, startIdx, midIdx, zc, origSeqStore);
         splits->splitLocations[splits->idx] = (U32)midIdx;
         splits->idx++;
···
     }
 }
 
-/* Base recursive function. Populates a table with intra-block partition indices that can improve compression ratio.
+/* Base recursive function.
+ * Populates a table with intra-block partition indices that can improve compression ratio.
  *
- * Returns the number of splits made (which equals the size of the partition table - 1).
+ * @return: number of splits made (which equals the size of the partition table - 1).
  */
-static size_t ZSTD_deriveBlockSplits(ZSTD_CCtx* zc, U32 partitions[], U32 nbSeq) {
-    seqStoreSplits splits = {partitions, 0};
+static size_t ZSTD_deriveBlockSplits(ZSTD_CCtx* zc, U32 partitions[], U32 nbSeq)
+{
+    seqStoreSplits splits;
+    splits.splitLocations = partitions;
+    splits.idx = 0;
     if (nbSeq <= 4) {
-        DEBUGLOG(4, "ZSTD_deriveBlockSplits: Too few sequences to split");
+        DEBUGLOG(5, "ZSTD_deriveBlockSplits: Too few sequences to split (%u <= 4)", nbSeq);
         /* Refuse to try and split anything with less than 4 sequences */
         return 0;
     }
···
  * Returns combined size of all blocks (which includes headers), or a ZSTD error code.
  */
 static size_t
-ZSTD_compressBlock_splitBlock_internal(ZSTD_CCtx* zc, void* dst, size_t dstCapacity,
-                                       const void* src, size_t blockSize, U32 lastBlock, U32 nbSeq)
+ZSTD_compressBlock_splitBlock_internal(ZSTD_CCtx* zc,
+                                    void* dst, size_t dstCapacity,
+                              const void* src, size_t blockSize,
+                                    U32 lastBlock, U32 nbSeq)
 {
     size_t cSize = 0;
     const BYTE* ip = (const BYTE*)src;
     BYTE* op = (BYTE*)dst;
     size_t i = 0;
     size_t srcBytesTotal = 0;
-    U32* partitions = zc->blockSplitCtx.partitions; /* size == ZSTD_MAX_NB_BLOCK_SPLITS */
-    seqStore_t* nextSeqStore = &zc->blockSplitCtx.nextSeqStore;
-    seqStore_t* currSeqStore = &zc->blockSplitCtx.currSeqStore;
-    size_t numSplits = ZSTD_deriveBlockSplits(zc, partitions, nbSeq);
+    U32* const partitions = zc->blockSplitCtx.partitions; /* size == ZSTD_MAX_NB_BLOCK_SPLITS */
+    SeqStore_t* const nextSeqStore = &zc->blockSplitCtx.nextSeqStore;
+    SeqStore_t* const currSeqStore = &zc->blockSplitCtx.currSeqStore;
+    size_t const numSplits = ZSTD_deriveBlockSplits(zc, partitions, nbSeq);
 
     /* If a block is split and some partitions are emitted as RLE/uncompressed, then repcode history
      * may become invalid. In order to reconcile potentially invalid repcodes, we keep track of two
···
      *
      * See ZSTD_seqStore_resolveOffCodes() for more details.
      */
-    repcodes_t dRep;
-    repcodes_t cRep;
-    ZSTD_memcpy(dRep.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t));
-    ZSTD_memcpy(cRep.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t));
-    ZSTD_memset(nextSeqStore, 0, sizeof(seqStore_t));
+    Repcodes_t dRep;
+    Repcodes_t cRep;
+    ZSTD_memcpy(dRep.rep, zc->blockState.prevCBlock->rep, sizeof(Repcodes_t));
+    ZSTD_memcpy(cRep.rep, zc->blockState.prevCBlock->rep, sizeof(Repcodes_t));
+    ZSTD_memset(nextSeqStore, 0, sizeof(SeqStore_t));
 
-    DEBUGLOG(4, "ZSTD_compressBlock_splitBlock_internal (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u)",
+    DEBUGLOG(5, "ZSTD_compressBlock_splitBlock_internal (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u)",
                 (unsigned)dstCapacity, (unsigned)zc->blockState.matchState.window.dictLimit,
                 (unsigned)zc->blockState.matchState.nextToUpdate);
 
     if (numSplits == 0) {
-        size_t cSizeSingleBlock = ZSTD_compressSeqStore_singleBlock(zc, &zc->seqStore,
-                                                                   &dRep, &cRep,
-                                                                    op, dstCapacity,
-                                                                    ip, blockSize,
-                                                                    lastBlock, 0 /* isPartition */);
+        size_t cSizeSingleBlock =
+            ZSTD_compressSeqStore_singleBlock(zc, &zc->seqStore,
+                                            &dRep, &cRep,
+                                            op, dstCapacity,
+                                            ip, blockSize,
+                                            lastBlock, 0 /* isPartition */);
         FORWARD_IF_ERROR(cSizeSingleBlock, "Compressing single block from splitBlock_internal() failed!");
         DEBUGLOG(5, "ZSTD_compressBlock_splitBlock_internal: No splits");
-        assert(cSizeSingleBlock <= ZSTD_BLOCKSIZE_MAX + ZSTD_blockHeaderSize);
+        assert(zc->blockSizeMax <= ZSTD_BLOCKSIZE_MAX);
+        assert(cSizeSingleBlock <= zc->blockSizeMax + ZSTD_blockHeaderSize);
         return cSizeSingleBlock;
     }
 
     ZSTD_deriveSeqStoreChunk(currSeqStore, &zc->seqStore, 0, partitions[0]);
     for (i = 0; i <= numSplits; ++i) {
-        size_t srcBytes;
         size_t cSizeChunk;
         U32 const lastPartition = (i == numSplits);
         U32 lastBlockEntireSrc = 0;
 
-        srcBytes = ZSTD_countSeqStoreLiteralsBytes(currSeqStore) + ZSTD_countSeqStoreMatchBytes(currSeqStore);
+        size_t srcBytes = ZSTD_countSeqStoreLiteralsBytes(currSeqStore) + ZSTD_countSeqStoreMatchBytes(currSeqStore);
         srcBytesTotal += srcBytes;
         if (lastPartition) {
             /* This is the final partition, need to account for possible last literals */
···
                                                        op, dstCapacity,
                                                        ip, srcBytes,
                                                        lastBlockEntireSrc, 1 /* isPartition */);
-        DEBUGLOG(5, "Estimated size: %zu actual size: %zu", ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(currSeqStore, zc), cSizeChunk);
+        DEBUGLOG(5, "Estimated size: %zu vs %zu : actual size",
+                    ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(currSeqStore, zc), cSizeChunk);
         FORWARD_IF_ERROR(cSizeChunk, "Compressing chunk failed!");
 
         ip += srcBytes;
···
         dstCapacity -= cSizeChunk;
         cSize += cSizeChunk;
         *currSeqStore = *nextSeqStore;
-        assert(cSizeChunk <= ZSTD_BLOCKSIZE_MAX + ZSTD_blockHeaderSize);
+        assert(cSizeChunk <= zc->blockSizeMax + ZSTD_blockHeaderSize);
     }
-    /* cRep and dRep may have diverged during the compression. If so, we use the dRep repcodes
-     * for the next block.
+    /* cRep and dRep may have diverged during the compression.
+     * If so, we use the dRep repcodes for the next block.
      */
-    ZSTD_memcpy(zc->blockState.prevCBlock->rep, dRep.rep, sizeof(repcodes_t));
+    ZSTD_memcpy(zc->blockState.prevCBlock->rep, dRep.rep, sizeof(Repcodes_t));
     return cSize;
 }
 
···
                               void* dst, size_t dstCapacity,
                               const void* src, size_t srcSize, U32 lastBlock)
 {
-    const BYTE* ip = (const BYTE*)src;
-    BYTE* op = (BYTE*)dst;
     U32 nbSeq;
     size_t cSize;
-    DEBUGLOG(4, "ZSTD_compressBlock_splitBlock");
-    assert(zc->appliedParams.useBlockSplitter == ZSTD_ps_enable);
+    DEBUGLOG(5, "ZSTD_compressBlock_splitBlock");
+    assert(zc->appliedParams.postBlockSplitter == ZSTD_ps_enable);
 
     {   const size_t bss = ZSTD_buildSeqStore(zc, src, srcSize);
         FORWARD_IF_ERROR(bss, "ZSTD_buildSeqStore failed");
         if (bss == ZSTDbss_noCompress) {
             if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
                 zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
-            cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, srcSize, lastBlock);
+            RETURN_ERROR_IF(zc->seqCollector.collectSequences, sequenceProducer_failed, "Uncompressible block");
+            cSize = ZSTD_noCompressBlock(dst, dstCapacity, src, srcSize, lastBlock);
             FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed");
-            DEBUGLOG(4, "ZSTD_compressBlock_splitBlock: Nocompress block");
+            DEBUGLOG(5, "ZSTD_compressBlock_splitBlock: Nocompress block");
             return cSize;
         }
         nbSeq = (U32)(zc->seqStore.sequences - zc->seqStore.sequencesStart);
···
                             void* dst, size_t dstCapacity,
                             const void* src, size_t srcSize, U32 frame)
 {
-    /* This the upper bound for the length of an rle block.
-     * This isn't the actual upper bound. Finding the real threshold
-     * needs further investigation.
+    /* This is an estimated upper bound for the length of an rle block.
+     * This isn't the actual upper bound.
+     * Finding the real threshold needs further investigation.
      */
     const U32 rleMaxLength = 25;
     size_t cSize;
···
 
     {   const size_t bss = ZSTD_buildSeqStore(zc, src, srcSize);
         FORWARD_IF_ERROR(bss, "ZSTD_buildSeqStore failed");
-        if (bss == ZSTDbss_noCompress) { cSize = 0; goto out; }
+        if (bss == ZSTDbss_noCompress) {
+            RETURN_ERROR_IF(zc->seqCollector.collectSequences, sequenceProducer_failed, "Uncompressible block");
+            cSize = 0;
+            goto out;
+        }
     }
 
     if (zc->seqCollector.collectSequences) {
-        ZSTD_copyBlockSequences(zc);
+        FORWARD_IF_ERROR(ZSTD_copyBlockSequences(&zc->seqCollector, ZSTD_getSeqStore(zc), zc->blockState.prevCBlock->rep), "copyBlockSequences failed");
         ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
         return 0;
     }
···
             &zc->appliedParams,
             dst, dstCapacity,
             srcSize,
-            zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
+            zc->tmpWorkspace, zc->tmpWkspSize /* statically allocated in resetCCtx */,
             zc->bmi2);
 
     if (frame &&
···
          *   * cSize >= blockBound(srcSize): We have expanded the block too much so
          *     emit an uncompressed block.
          */
-        {
-            size_t const cSize = ZSTD_compressSuperBlock(zc, dst, dstCapacity, src, srcSize, lastBlock);
+        {   size_t const cSize =
+                ZSTD_compressSuperBlock(zc, dst, dstCapacity, src, srcSize, lastBlock);
             if (cSize != ERROR(dstSize_tooSmall)) {
-                size_t const maxCSize = srcSize - ZSTD_minGain(srcSize, zc->appliedParams.cParams.strategy);
+                size_t const maxCSize =
+                    srcSize - ZSTD_minGain(srcSize, zc->appliedParams.cParams.strategy);
                 FORWARD_IF_ERROR(cSize, "ZSTD_compressSuperBlock failed");
                 if (cSize != 0 && cSize < maxCSize + ZSTD_blockHeaderSize) {
                     ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
···
                 }
             }
         }
-    }
+    } /* if (bss == ZSTDbss_compress)*/
 
     DEBUGLOG(6, "Resorting to ZSTD_noCompressBlock()");
     /* Superblock compression failed, attempt to emit a single no compress block.
···
     return cSize;
 }
 
-static void ZSTD_overflowCorrectIfNeeded(ZSTD_matchState_t* ms,
+static void ZSTD_overflowCorrectIfNeeded(ZSTD_MatchState_t* ms,
                                          ZSTD_cwksp* ws,
                                          ZSTD_CCtx_params const* params,
                                          void const* ip,
···
     }
 }
 
+#include "zstd_preSplit.h"
+
+static size_t ZSTD_optimalBlockSize(ZSTD_CCtx* cctx, const void* src, size_t srcSize, size_t blockSizeMax, int splitLevel, ZSTD_strategy strat, S64 savings)
+{
+    /* split level based on compression strategy, from `fast` to `btultra2` */
+    static const int splitLevels[] = { 0, 0, 1, 2, 2, 3, 3, 4, 4, 4 };
+    /* note: conservatively only split full blocks (128 KB) currently.
+     * While it's possible to go lower, let's keep it simple for a first implementation.
+     * Besides, benefits of splitting are reduced when blocks are already small.
+     */
+    if (srcSize < 128 KB || blockSizeMax < 128 KB)
+        return MIN(srcSize, blockSizeMax);
+    /* do not split incompressible data though:
+     * require verified savings to allow pre-splitting.
+     * Note: as a consequence, the first full block is not split.
+     */
+    if (savings < 3) {
+        DEBUGLOG(6, "don't attempt splitting: savings (%i) too low", (int)savings);
+        return 128 KB;
+    }
+    /* apply @splitLevel, or use default value (which depends on @strat).
+     * note that splitting heuristic is still conditioned by @savings >= 3,
+     * so the first block will not reach this code path */
+    if (splitLevel == 1) return 128 KB;
+    if (splitLevel == 0) {
+        assert(ZSTD_fast <= strat && strat <= ZSTD_btultra2);
+        splitLevel = splitLevels[strat];
+    } else {
+        assert(2 <= splitLevel && splitLevel <= 6);
+        splitLevel -= 2;
+    }
+    return ZSTD_splitBlock(src, blockSizeMax, splitLevel, cctx->tmpWorkspace, cctx->tmpWkspSize);
+}
+
 /*! ZSTD_compress_frameChunk() :
 *   Compress a chunk of data into one or multiple blocks.
 *   All blocks will be terminated, all input will be consumed.
 *   Function will issue an error if there is not enough `dstCapacity` to hold the compressed content.
 *   Frame is supposed already started (header already produced)
-*   @return : compressed size, or an error code
+*  @return : compressed size, or an error code
 */
 static size_t ZSTD_compress_frameChunk(ZSTD_CCtx* cctx,
                                      void* dst, size_t dstCapacity,
                                const void* src, size_t srcSize,
                                      U32 lastFrameChunk)
 {
-    size_t blockSize = cctx->blockSize;
+    size_t blockSizeMax = cctx->blockSizeMax;
     size_t remaining = srcSize;
     const BYTE* ip = (const BYTE*)src;
     BYTE* const ostart = (BYTE*)dst;
     BYTE* op = ostart;
     U32 const maxDist = (U32)1 << cctx->appliedParams.cParams.windowLog;
+    S64 savings = (S64)cctx->consumedSrcSize - (S64)cctx->producedCSize;
 
     assert(cctx->appliedParams.cParams.windowLog <= ZSTD_WINDOWLOG_MAX);
 
-    DEBUGLOG(4, "ZSTD_compress_frameChunk (blockSize=%u)", (unsigned)blockSize);
+    DEBUGLOG(5, "ZSTD_compress_frameChunk (srcSize=%u, blockSizeMax=%u)", (unsigned)srcSize, (unsigned)blockSizeMax);
     if (cctx->appliedParams.fParams.checksumFlag && srcSize)
         xxh64_update(&cctx->xxhState, src, srcSize);
 
     while (remaining) {
-        ZSTD_matchState_t* const ms = &cctx->blockState.matchState;
-        U32 const lastBlock = lastFrameChunk & (blockSize >= remaining);
+        ZSTD_MatchState_t* const ms = &cctx->blockState.matchState;
+        size_t const blockSize = ZSTD_optimalBlockSize(cctx,
+                                ip, remaining,
+                                blockSizeMax,
+                                cctx->appliedParams.preBlockSplitter_level,
+                                cctx->appliedParams.cParams.strategy,
+                                savings);
+        U32 const lastBlock = lastFrameChunk & (blockSize == remaining);
+        assert(blockSize <= remaining);
 
-        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize + MIN_CBLOCK_SIZE,
+        /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
+         * additional 1. We need to revisit and change this logic to be more consistent */
+        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize + MIN_CBLOCK_SIZE + 1,
                         dstSize_tooSmall,
                         "not enough space to store compressed block");
-        if (remaining < blockSize) blockSize = remaining;
 
         ZSTD_overflowCorrectIfNeeded(
             ms, &cctx->workspace, &cctx->appliedParams, ip, ip + blockSize);
···
                     MEM_writeLE24(op, cBlockHeader);
                     cSize += ZSTD_blockHeaderSize;
                 }
-            }
+            }  /* if (ZSTD_useTargetCBlockSize(&cctx->appliedParams))*/
 
+            /* @savings is employed to ensure that splitting doesn't worsen expansion of incompressible data.
+             * Without splitting, the maximum expansion is 3 bytes per full block.
+             * An adversarial input could attempt to fudge the split detector,
+             * and make it split incompressible data, resulting in more block headers.
+             * Note that, since ZSTD_COMPRESSBOUND() assumes a worst case scenario of 1KB per block,
+             * and the splitter never creates blocks that small (current lower limit is 8 KB),
+             * there is already no risk to expand beyond ZSTD_COMPRESSBOUND() limit.
+             * But if the goal is to not expand by more than 3-bytes per 128 KB full block,
+             * then yes, it becomes possible to make the block splitter oversplit incompressible data.
+             * Using @savings, we enforce an even more conservative condition,
+             * requiring the presence of enough savings (at least 3 bytes) to authorize splitting,
+             * otherwise only full blocks are used.
+             * But being conservative is fine,
+             * since splitting barely compressible blocks is not fruitful anyway */
+            savings += (S64)blockSize - (S64)cSize;
 
             ip += blockSize;
             assert(remaining >= blockSize);
···
 
 
 static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity,
-                                    const ZSTD_CCtx_params* params, U64 pledgedSrcSize, U32 dictID)
-{   BYTE* const op = (BYTE*)dst;
+                                    const ZSTD_CCtx_params* params,
+                                    U64 pledgedSrcSize, U32 dictID)
+{
+    BYTE* const op = (BYTE*)dst;
     U32   const dictIDSizeCodeLength = (dictID>0) + (dictID>=256) + (dictID>=65536);   /* 0-3 */
     U32   const dictIDSizeCode = params->fParams.noDictIDFlag ? 0 : dictIDSizeCodeLength;   /* 0-3 */
     U32   const checksumFlag = params->fParams.checksumFlag>0;
···
     }
 }
 
-size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSeq)
+void ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSeq)
 {
-    RETURN_ERROR_IF(cctx->stage != ZSTDcs_init, stage_wrong,
-                    "wrong cctx stage");
-    RETURN_ERROR_IF(cctx->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable,
-                    parameter_unsupported,
-                    "incompatible with ldm");
+    assert(cctx->stage == ZSTDcs_init);
+    assert(nbSeq == 0 || cctx->appliedParams.ldmParams.enableLdm != ZSTD_ps_enable);
     cctx->externSeqStore.seq = seq;
     cctx->externSeqStore.size = nbSeq;
     cctx->externSeqStore.capacity = nbSeq;
     cctx->externSeqStore.pos = 0;
     cctx->externSeqStore.posInSequence = 0;
-    return 0;
 }
 
 
···
                         const void* src, size_t srcSize,
                                U32 frame, U32 lastFrameChunk)
 {
-    ZSTD_matchState_t* const ms = &cctx->blockState.matchState;
+    ZSTD_MatchState_t* const ms = &cctx->blockState.matchState;
     size_t fhSize = 0;
 
     DEBUGLOG(5, "ZSTD_compressContinue_internal, stage: %u, srcSize: %u",
···
             src, (BYTE const*)src + srcSize);
     }
 
-    DEBUGLOG(5, "ZSTD_compressContinue_internal (blockSize=%u)", (unsigned)cctx->blockSize);
+    DEBUGLOG(5, "ZSTD_compressContinue_internal (blockSize=%u)", (unsigned)cctx->blockSizeMax);
     {   size_t const cSize = frame ?
                              ZSTD_compress_frameChunk (cctx, dst, dstCapacity, src, srcSize, lastFrameChunk) :
                              ZSTD_compressBlock_internal (cctx, dst, dstCapacity, src, srcSize, 0 /* frame */);
···
     }
 }
 
-size_t ZSTD_compressContinue (ZSTD_CCtx* cctx,
-                              void* dst, size_t dstCapacity,
-                        const void* src, size_t srcSize)
+size_t ZSTD_compressContinue_public(ZSTD_CCtx* cctx,
+                                        void* dst, size_t dstCapacity,
+                                  const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_compressContinue (srcSize=%u)", (unsigned)srcSize);
     return ZSTD_compressContinue_internal(cctx, dst, dstCapacity, src, srcSize, 1 /* frame mode */, 0 /* last chunk */);
 }
 
+/* NOTE: Must just wrap ZSTD_compressContinue_public() */
+size_t ZSTD_compressContinue(ZSTD_CCtx* cctx,
+                             void* dst, size_t dstCapacity,
+                       const void* src, size_t srcSize)
+{
+    return ZSTD_compressContinue_public(cctx, dst, dstCapacity, src, srcSize);
+}
 
-size_t ZSTD_getBlockSize(const ZSTD_CCtx* cctx)
+static size_t ZSTD_getBlockSize_deprecated(const ZSTD_CCtx* cctx)
 {
     ZSTD_compressionParameters const cParams = cctx->appliedParams.cParams;
     assert(!ZSTD_checkCParams(cParams));
-    return MIN (ZSTD_BLOCKSIZE_MAX, (U32)1 << cParams.windowLog);
+    return MIN(cctx->appliedParams.maxBlockSize, (size_t)1 << cParams.windowLog);
 }
 
-size_t ZSTD_compressBlock(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+/* NOTE: Must just wrap ZSTD_getBlockSize_deprecated() */
+size_t ZSTD_getBlockSize(const ZSTD_CCtx* cctx)
+{
+    return ZSTD_getBlockSize_deprecated(cctx);
+}
+
+/* NOTE: Must just wrap ZSTD_compressBlock_deprecated() */
+size_t ZSTD_compressBlock_deprecated(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_compressBlock: srcSize = %u", (unsigned)srcSize);
-    { size_t const blockSizeMax = ZSTD_getBlockSize(cctx);
+    { size_t const blockSizeMax = ZSTD_getBlockSize_deprecated(cctx);
       RETURN_ERROR_IF(srcSize > blockSizeMax, srcSize_wrong, "input is larger than a block"); }
 
     return ZSTD_compressContinue_internal(cctx, dst, dstCapacity, src, srcSize, 0 /* frame mode */, 0 /* last chunk */);
 }
 
+/* NOTE: Must just wrap ZSTD_compressBlock_deprecated() */
+size_t ZSTD_compressBlock(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_deprecated(cctx, dst, dstCapacity, src, srcSize);
+}
+
 /*! ZSTD_loadDictionaryContent() :
  *  @return : 0, or an error code
  */
-static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms,
-                                         ldmState_t* ls,
-                                         ZSTD_cwksp* ws,
-                                         ZSTD_CCtx_params const* params,
-                                         const void* src, size_t srcSize,
-                                         ZSTD_dictTableLoadMethod_e dtlm)
+static size_t
+ZSTD_loadDictionaryContent(ZSTD_MatchState_t* ms,
+                        ldmState_t* ls,
+                        ZSTD_cwksp* ws,
+                        ZSTD_CCtx_params const* params,
+                        const void* src, size_t srcSize,
+                        ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp)
 {
     const BYTE* ip = (const BYTE*) src;
     const BYTE* const iend = ip + srcSize;
     int const loadLdmDict = params->ldmParams.enableLdm == ZSTD_ps_enable && ls != NULL;
 
-    /* Assert that we the ms params match the params we're being given */
+    /* Assert that the ms params match the params we're being given */
     ZSTD_assertEqualCParams(params->cParams, ms->cParams);
 
-    if (srcSize > ZSTD_CHUNKSIZE_MAX) {
+    {   /* Ensure large dictionaries can't cause index overflow */
+
         /* Allow the dictionary to set indices up to exactly ZSTD_CURRENT_MAX.
          * Dictionaries right at the edge will immediately trigger overflow
          * correction, but I don't want to insert extra constraints here.
          */
-        U32 const maxDictSize = ZSTD_CURRENT_MAX - 1;
-        /* We must have cleared our windows when our source is this large. */
-        assert(ZSTD_window_isEmpty(ms->window));
-        if (loadLdmDict)
-            assert(ZSTD_window_isEmpty(ls->window));
+        U32 maxDictSize = ZSTD_CURRENT_MAX - ZSTD_WINDOW_START_INDEX;
+
+        int const CDictTaggedIndices = ZSTD_CDictIndicesAreTagged(&params->cParams);
+        if (CDictTaggedIndices && tfp == ZSTD_tfp_forCDict) {
+            /* Some dictionary matchfinders in zstd use "short cache",
+             * which treats the lower ZSTD_SHORT_CACHE_TAG_BITS of each
+             * CDict hashtable entry as a tag rather than as part of an index.
+             * When short cache is used, we need to truncate the dictionary
+             * so that its indices don't overlap with the tag. */
+            U32 const shortCacheMaxDictSize = (1u << (32 - ZSTD_SHORT_CACHE_TAG_BITS)) - ZSTD_WINDOW_START_INDEX;
+            maxDictSize = MIN(maxDictSize, shortCacheMaxDictSize);
+            assert(!loadLdmDict);
+        }
+
         /* If the dictionary is too large, only load the suffix of the dictionary. */
         if (srcSize > maxDictSize) {
             ip = iend - maxDictSize;
···
         }
     }
 
-    DEBUGLOG(4, "ZSTD_loadDictionaryContent(): useRowMatchFinder=%d", (int)params->useRowMatchFinder);
+    if (srcSize > ZSTD_CHUNKSIZE_MAX) {
+        /* We must have cleared our windows when our source is this large. */
+        assert(ZSTD_window_isEmpty(ms->window));
+        if (loadLdmDict) assert(ZSTD_window_isEmpty(ls->window));
+    }
     ZSTD_window_update(&ms->window, src, srcSize, /* forceNonContiguous */ 0);
-    ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base);
-    ms->forceNonContiguous = params->deterministicRefPrefix;
 
-    if (loadLdmDict) {
+    DEBUGLOG(4, "ZSTD_loadDictionaryContent: useRowMatchFinder=%d", (int)params->useRowMatchFinder);
+
+    if (loadLdmDict) { /* Load the entire dict into LDM matchfinders. */
+        DEBUGLOG(4, "ZSTD_loadDictionaryContent: Trigger loadLdmDict");
         ZSTD_window_update(&ls->window, src, srcSize, /* forceNonContiguous */ 0);
         ls->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ls->window.base);
+        ZSTD_ldm_fillHashTable(ls, ip, iend, &params->ldmParams);
+        DEBUGLOG(4, "ZSTD_loadDictionaryContent: ZSTD_ldm_fillHashTable completes");
     }
+
+    /* If the dict is larger than we can reasonably index in our tables, only load the suffix. */
+    {   U32 maxDictSize = 1U << MIN(MAX(params->cParams.hashLog + 3, params->cParams.chainLog + 1), 31);
+        if (srcSize > maxDictSize) {
+            ip = iend - maxDictSize;
+            src = ip;
+            srcSize = maxDictSize;
+        }
+    }
+
+    ms->nextToUpdate = (U32)(ip - ms->window.base);
+    ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base);
+    ms->forceNonContiguous = params->deterministicRefPrefix;
 
     if (srcSize <= HASH_READ_SIZE) return 0;
 
     ZSTD_overflowCorrectIfNeeded(ms, ws, params, ip, iend);
 
-    if (loadLdmDict)
-        ZSTD_ldm_fillHashTable(ls, ip, iend, &params->ldmParams);
-
     switch(params->cParams.strategy)
     {
     case ZSTD_fast:
-        ZSTD_fillHashTable(ms, iend, dtlm);
+        ZSTD_fillHashTable(ms, iend, dtlm, tfp);
         break;
     case ZSTD_dfast:
-        ZSTD_fillDoubleHashTable(ms, iend, dtlm);
+#ifndef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR
+        ZSTD_fillDoubleHashTable(ms, iend, dtlm, tfp);
+#else
+        assert(0); /* shouldn't be called: cparams should've been adjusted. */
+#endif
         break;
 
     case ZSTD_greedy:
     case ZSTD_lazy:
     case ZSTD_lazy2:
+#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR)
         assert(srcSize >= HASH_READ_SIZE);
         if (ms->dedicatedDictSearch) {
             assert(ms->chainTable != NULL);
···
         } else {
             assert(params->useRowMatchFinder != ZSTD_ps_auto);
             if (params->useRowMatchFinder == ZSTD_ps_enable) {
-                size_t const tagTableSize = ((size_t)1 << params->cParams.hashLog) * sizeof(U16);
+                size_t const tagTableSize = ((size_t)1 << params->cParams.hashLog);
                 ZSTD_memset(ms->tagTable, 0, tagTableSize);
                 ZSTD_row_update(ms, iend-HASH_READ_SIZE);
                 DEBUGLOG(4, "Using row-based hash table for lazy dict");
···
                 DEBUGLOG(4, "Using chain-based hash table for lazy dict");
             }
         }
+#else
+        assert(0); /* shouldn't be called: cparams should've been adjusted. */
+#endif
         break;
 
     case ZSTD_btlazy2:   /* we want the dictionary table fully sorted */
     case ZSTD_btopt:
     case ZSTD_btultra:
     case ZSTD_btultra2:
+#if !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR)
         assert(srcSize >= HASH_READ_SIZE);
+        DEBUGLOG(4, "Fill %u bytes into the Binary Tree", (unsigned)srcSize);
         ZSTD_updateTree(ms, iend-HASH_READ_SIZE, iend);
+#else
+        assert(0); /* shouldn't be called: cparams should've been adjusted. */
+#endif
         break;
 
     default:
···
     {   unsigned maxSymbolValue = 255;
         unsigned hasZeroWeights = 1;
         size_t const hufHeaderSize = HUF_readCTable((HUF_CElt*)bs->entropy.huf.CTable, &maxSymbolValue, dictPtr,
-            dictEnd-dictPtr, &hasZeroWeights);
+            (size_t)(dictEnd-dictPtr), &hasZeroWeights);
 
         /* We only set the loaded table as valid if it contains all non-zero
          * weights. Otherwise, we set it to check */
-        if (!hasZeroWeights)
+        if (!hasZeroWeights && maxSymbolValue == 255)
             bs->entropy.huf.repeatMode = HUF_repeat_valid;
 
         RETURN_ERROR_IF(HUF_isError(hufHeaderSize), dictionary_corrupted, "");
-        RETURN_ERROR_IF(maxSymbolValue < 255, dictionary_corrupted, "");
         dictPtr += hufHeaderSize;
     }
 
     {   unsigned offcodeLog;
-        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr);
+        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, (size_t)(dictEnd-dictPtr));
         RETURN_ERROR_IF(FSE_isError(offcodeHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(offcodeLog > OffFSELog, dictionary_corrupted, "");
         /* fill all offset symbols to avoid garbage at end of table */
···
 
     {   short matchlengthNCount[MaxML+1];
         unsigned matchlengthMaxValue = MaxML, matchlengthLog;
-        size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, dictEnd-dictPtr);
+        size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, (size_t)(dictEnd-dictPtr));
         RETURN_ERROR_IF(FSE_isError(matchlengthHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(matchlengthLog > MLFSELog, dictionary_corrupted, "");
         RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp(
···
 
     {   short litlengthNCount[MaxLL+1];
         unsigned litlengthMaxValue = MaxLL, litlengthLog;
-        size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, dictEnd-dictPtr);
+        size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, (size_t)(dictEnd-dictPtr));
         RETURN_ERROR_IF(FSE_isError(litlengthHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(litlengthLog > LLFSELog, dictionary_corrupted, "");
         RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp(
···
                 RETURN_ERROR_IF(bs->rep[u] > dictContentSize, dictionary_corrupted, "");
     }   }   }
 
-    return dictPtr - (const BYTE*)dict;
+    return (size_t)(dictPtr - (const BYTE*)dict);
 }
 
 /* Dictionary format :
···
  *                dictSize supposed >= 8
  */
 static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs,
-                                      ZSTD_matchState_t* ms,
+                                      ZSTD_MatchState_t* ms,
                                       ZSTD_cwksp* ws,
                                       ZSTD_CCtx_params const* params,
                                       const void* dict, size_t dictSize,
                                       ZSTD_dictTableLoadMethod_e dtlm,
+                                      ZSTD_tableFillPurpose_e tfp,
                                       void* workspace)
 {
     const BYTE* dictPtr = (const BYTE*)dict;
···
     {
         size_t const dictContentSize = (size_t)(dictEnd - dictPtr);
         FORWARD_IF_ERROR(ZSTD_loadDictionaryContent(
-            ms, NULL, ws, params, dictPtr, dictContentSize, dtlm), "");
+            ms, NULL, ws, params, dictPtr, dictContentSize, dtlm, tfp), "");
     }
     return dictID;
 }
···
 *   @return : dictID, or an error code */
 static size_t
 ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs,
-                               ZSTD_matchState_t* ms,
+                               ZSTD_MatchState_t* ms,
                                ldmState_t* ls,
                                ZSTD_cwksp* ws,
                          const ZSTD_CCtx_params* params,
                          const void* dict, size_t dictSize,
                                ZSTD_dictContentType_e dictContentType,
                                ZSTD_dictTableLoadMethod_e dtlm,
+                               ZSTD_tableFillPurpose_e tfp,
                                void* workspace)
 {
     DEBUGLOG(4, "ZSTD_compress_insertDictionary (dictSize=%u)", (U32)dictSize);
···
 
     /* dict restricted modes */
     if (dictContentType == ZSTD_dct_rawContent)
-        return ZSTD_loadDictionaryContent(ms, ls, ws, params, dict, dictSize, dtlm);
+        return ZSTD_loadDictionaryContent(ms, ls, ws, params, dict, dictSize, dtlm, tfp);
 
     if (MEM_readLE32(dict) != ZSTD_MAGIC_DICTIONARY) {
         if (dictContentType == ZSTD_dct_auto) {
             DEBUGLOG(4, "raw content dictionary detected");
             return ZSTD_loadDictionaryContent(
-                ms, ls, ws, params, dict, dictSize, dtlm);
+                ms, ls, ws, params, dict, dictSize, dtlm, tfp);
         }
         RETURN_ERROR_IF(dictContentType == ZSTD_dct_fullDict, dictionary_wrong, "");
         assert(0);   /* impossible */
···
 
     /* dict as full zstd dictionary */
     return ZSTD_loadZstdDictionary(
-        bs, ms, ws, params, dict, dictSize, dtlm, workspace);
+        bs, ms, ws, params, dict, dictSize, dtlm, tfp, workspace);
 }
 
 #define ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF (128 KB)
 #define ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER (6ULL)
 
 /*! ZSTD_compressBegin_internal() :
+ * Assumption : either @dict OR @cdict (or none) is non-NULL, never both
  * @return : 0, or an error code */
 static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
                                     const void* dict, size_t dictSize,
···
                         cctx->blockState.prevCBlock, &cctx->blockState.matchState,
                         &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, cdict->dictContent,
                         cdict->dictContentSize, cdict->dictContentType, dtlm,
-                        cctx->entropyWorkspace)
+                        ZSTD_tfp_forCCtx, cctx->tmpWorkspace)
               : ZSTD_compress_insertDictionary(
                         cctx->blockState.prevCBlock, &cctx->blockState.matchState,
                         &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, dict, dictSize,
-                        dictContentType, dtlm, cctx->entropyWorkspace);
+                        dictContentType, dtlm, ZSTD_tfp_forCCtx, cctx->tmpWorkspace);
         FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed");
         assert(dictID <= UINT_MAX);
         cctx->dictID = (U32)dictID;
···
                                             &cctxParams, pledgedSrcSize);
 }
 
-size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
+static size_t
+ZSTD_compressBegin_usingDict_deprecated(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
 {
     ZSTD_CCtx_params cctxParams;
-    {
-        ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_noAttachDict);
+    {   ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_noAttachDict);
         ZSTD_CCtxParams_init_internal(&cctxParams, &params, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel);
     }
     DEBUGLOG(4, "ZSTD_compressBegin_usingDict (dictSize=%u)", (unsigned)dictSize);
···
                                        &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, ZSTDb_not_buffered);
 }
 
+size_t
+ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
+{
+    return ZSTD_compressBegin_usingDict_deprecated(cctx, dict, dictSize, compressionLevel);
+}
+
 size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel)
 {
-    return ZSTD_compressBegin_usingDict(cctx, NULL, 0, compressionLevel);
+    return ZSTD_compressBegin_usingDict_deprecated(cctx, NULL, 0, compressionLevel);
 }
 
 
···
 {
     BYTE* const ostart = (BYTE*)dst;
     BYTE* op = ostart;
-    size_t fhSize = 0;
 
     DEBUGLOG(4, "ZSTD_writeEpilogue");
     RETURN_ERROR_IF(cctx->stage == ZSTDcs_created, stage_wrong, "init missing");
 
     /* special case : empty frame */
     if (cctx->stage == ZSTDcs_init) {
-        fhSize = ZSTD_writeFrameHeader(dst, dstCapacity, &cctx->appliedParams, 0, 0);
+        size_t fhSize = ZSTD_writeFrameHeader(dst, dstCapacity, &cctx->appliedParams, 0, 0);
         FORWARD_IF_ERROR(fhSize, "ZSTD_writeFrameHeader failed");
         dstCapacity -= fhSize;
         op += fhSize;
···
     if (cctx->stage != ZSTDcs_ending) {
         /* write one last empty block, make it the "last" block */
         U32 const cBlockHeader24 = 1 /* last block */ + (((U32)bt_raw)<<1) + 0;
-        RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "no room for epilogue");
-        MEM_writeLE32(op, cBlockHeader24);
+        ZSTD_STATIC_ASSERT(ZSTD_BLOCKHEADERSIZE == 3);
+        RETURN_ERROR_IF(dstCapacity<3, dstSize_tooSmall, "no room for epilogue");
+        MEM_writeLE24(op, cBlockHeader24);
         op += ZSTD_blockHeaderSize;
         dstCapacity -= ZSTD_blockHeaderSize;
     }
···
     }
 
     cctx->stage = ZSTDcs_created;  /* return to "created but no init" status */
-    return op-ostart;
+    return (size_t)(op-ostart);
 }
 
 void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize)
···
     (void)extraCSize;
 }
 
-size_t ZSTD_compressEnd (ZSTD_CCtx* cctx,
-                         void* dst, size_t dstCapacity,
-                   const void* src, size_t srcSize)
+size_t ZSTD_compressEnd_public(ZSTD_CCtx* cctx,
+                               void* dst, size_t dstCapacity,
+                         const void* src, size_t srcSize)
 {
     size_t endResult;
     size_t const cSize = ZSTD_compressContinue_internal(cctx,
···
     }
     ZSTD_CCtx_trace(cctx, endResult);
     return cSize + endResult;
+}
+
+/* NOTE: Must just wrap ZSTD_compressEnd_public() */
+size_t ZSTD_compressEnd(ZSTD_CCtx* cctx,
+                        void* dst, size_t dstCapacity,
+                  const void* src, size_t srcSize)
+{
+    return ZSTD_compressEnd_public(cctx, dst, dstCapacity, src, srcSize);
 }
 
 size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx,
···
     FORWARD_IF_ERROR( ZSTD_compressBegin_internal(cctx,
                          dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast, NULL,
                          params, srcSize, ZSTDb_not_buffered) , "");
-    return ZSTD_compressEnd(cctx, dst, dstCapacity, src, srcSize);
+    return ZSTD_compressEnd_public(cctx, dst, dstCapacity, src, srcSize);
 }
 
 size_t ZSTD_compress_usingDict(ZSTD_CCtx* cctx,
···
         {   size_t const dictID = ZSTD_compress_insertDictionary(
                     &cdict->cBlockState, &cdict->matchState, NULL, &cdict->workspace,
                     &params, cdict->dictContent, cdict->dictContentSize,
-                    dictContentType, ZSTD_dtlm_full, cdict->entropyWorkspace);
+                    dictContentType, ZSTD_dtlm_full, ZSTD_tfp_forCDict, cdict->entropyWorkspace);
             FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed");
             assert(dictID <= (size_t)(U32)-1);
             cdict->dictID = (U32)dictID;
···
     return 0;
 }
 
-static ZSTD_CDict* ZSTD_createCDict_advanced_internal(size_t dictSize,
-                                      ZSTD_dictLoadMethod_e dictLoadMethod,
-                                      ZSTD_compressionParameters cParams,
-                                      ZSTD_paramSwitch_e useRowMatchFinder,
-                                      U32 enableDedicatedDictSearch,
-                                      ZSTD_customMem customMem)
+static ZSTD_CDict*
+ZSTD_createCDict_advanced_internal(size_t dictSize,
+                                ZSTD_dictLoadMethod_e dictLoadMethod,
+                                ZSTD_compressionParameters cParams,
+                                ZSTD_ParamSwitch_e useRowMatchFinder,
+                                int enableDedicatedDictSearch,
+                                ZSTD_customMem customMem)
 {
     if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
+    DEBUGLOG(3, "ZSTD_createCDict_advanced_internal (dictSize=%u)", (unsigned)dictSize);
 
     {   size_t const workspaceSize =
             ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) +
···
 {
     ZSTD_CCtx_params cctxParams;
     ZSTD_memset(&cctxParams, 0, sizeof(cctxParams));
+    DEBUGLOG(3, "ZSTD_createCDict_advanced, dictSize=%u, mode=%u", (unsigned)dictSize, (unsigned)dictContentType);
     ZSTD_CCtxParams_init(&cctxParams, 0);
     cctxParams.cParams = cParams;
     cctxParams.customMem = customMem;
···
     ZSTD_compressionParameters cParams;
     ZSTD_CDict* cdict;
 
-    DEBUGLOG(3, "ZSTD_createCDict_advanced2, mode %u", (unsigned)dictContentType);
+    DEBUGLOG(3, "ZSTD_createCDict_advanced2, dictSize=%u, mode=%u", (unsigned)dictSize, (unsigned)dictContentType);
     if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
 
     if (cctxParams.enableDedicatedDictSearch) {
···
             &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
     }
 
-    DEBUGLOG(3, "ZSTD_createCDict_advanced2: DDS: %u", cctxParams.enableDedicatedDictSearch);
+    DEBUGLOG(3, "ZSTD_createCDict_advanced2: DedicatedDictSearch=%u", cctxParams.enableDedicatedDictSearch);
     cctxParams.cParams = cParams;
     cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);
 
···
                         dictLoadMethod, cctxParams.cParams,
                         cctxParams.useRowMatchFinder, cctxParams.enableDedicatedDictSearch,
                         customMem);
-    if (!cdict)
-        return NULL;
 
-    if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
+    if (!cdict || ZSTD_isError( ZSTD_initCDict_internal(cdict,
                                     dict, dictSize,
                                     dictLoadMethod, dictContentType,
                                     cctxParams) )) {
···
  *  workspaceSize: Use ZSTD_estimateCDictSize()
  *                 to determine how large workspace must be.
  *  cParams : use ZSTD_getCParams() to transform a compression level
- *            into its relevants cParams.
+ *            into its relevant cParams.
  * @return : pointer to ZSTD_CDict*, or NULL if error (size too small)
  *  Note : there is no corresponding "free" function.
  *         Since workspace was allocated externally, it must be freed externally.
···
                                  ZSTD_dictContentType_e dictContentType,
                                  ZSTD_compressionParameters cParams)
 {
-    ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(ZSTD_ps_auto, &cParams);
+    ZSTD_ParamSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(ZSTD_ps_auto, &cParams);
     /* enableDedicatedDictSearch == 1 ensures matchstate is not too small in case this CDict will be used for DDS + row hash */
     size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0);
     size_t const neededSize = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict))
···
     ZSTD_CDict* cdict;
     ZSTD_CCtx_params params;
 
+    DEBUGLOG(4, "ZSTD_initStaticCDict (dictSize==%u)", (unsigned)dictSize);
     if ((size_t)workspace & 7) return NULL;  /* 8-aligned */
 
     {
···
         ZSTD_cwksp_move(&cdict->workspace, &ws);
     }
 
-    DEBUGLOG(4, "(workspaceSize < neededSize) : (%u < %u) => %u",
-        (unsigned)workspaceSize, (unsigned)neededSize, (unsigned)(workspaceSize < neededSize));
     if (workspaceSize < neededSize) return NULL;
 
     ZSTD_CCtxParams_init(&params, 0);
     params.cParams = cParams;
     params.useRowMatchFinder = useRowMatchFinder;
     cdict->useRowMatchFinder = useRowMatchFinder;
+    cdict->compressionLevel = ZSTD_NO_CLEVEL;
 
     if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
                                               dict, dictSize,
···
 
 /* ZSTD_compressBegin_usingCDict() :
  * cdict must be != NULL */
-size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
+size_t ZSTD_compressBegin_usingCDict_deprecated(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
 {
     ZSTD_frameParameters const fParams = { 0 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
     return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN);
+}
+
+size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
+{
+    return ZSTD_compressBegin_usingCDict_deprecated(cctx, cdict);
 }
 
 /*! ZSTD_compress_usingCDict_internal():
···
                                 const ZSTD_CDict* cdict, ZSTD_frameParameters fParams)
 {
     FORWARD_IF_ERROR(ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, srcSize), ""); /* will check if cdict != NULL */
-    return ZSTD_compressEnd(cctx, dst, dstCapacity, src, srcSize);
+    return ZSTD_compressEnd_public(cctx, dst, dstCapacity, src, srcSize);
 }
 
 /*! ZSTD_compress_usingCDict_advanced():
···
     return ZSTD_compressBound(ZSTD_BLOCKSIZE_MAX) + ZSTD_blockHeaderSize + 4 /* 32-bits hash */ ;
 }
 
-static ZSTD_cParamMode_e ZSTD_getCParamMode(ZSTD_CDict const* cdict, ZSTD_CCtx_params const* params, U64 pledgedSrcSize)
+static ZSTD_CParamMode_e ZSTD_getCParamMode(ZSTD_CDict const* cdict, ZSTD_CCtx_params const* params, U64 pledgedSrcSize)
 {
     if (cdict != NULL && ZSTD_shouldAttachDict(cdict, params, pledgedSrcSize))
         return ZSTD_cpm_attachDict;
···
 
 static size_t ZSTD_nextInputSizeHint(const ZSTD_CCtx* cctx)
 {
-    size_t hintInSize = cctx->inBuffTarget - cctx->inBuffPos;
-    if (hintInSize==0) hintInSize = cctx->blockSize;
-    return hintInSize;
+    if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
+        return cctx->blockSizeMax - cctx->stableIn_notConsumed;
+    }
+    assert(cctx->appliedParams.inBufferMode == ZSTD_bm_buffered);
+    {   size_t hintInSize = cctx->inBuffTarget - cctx->inBuffPos;
+        if (hintInSize==0) hintInSize = cctx->blockSizeMax;
+        return hintInSize;
+    }
 }
 
 /* ZSTD_compressStream_generic():
  *  internal function for all *compressStream*() variants
- *  non-static, because can be called from zstdmt_compress.c
- * @return : hint size for next input */
+ * @return : hint size for next input to complete ongoing block */
 static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
                                           ZSTD_outBuffer* output,
                                           ZSTD_inBuffer* input,
                                           ZSTD_EndDirective const flushMode)
 {
-    const char* const istart = (const char*)input->src;
-    const char* const iend = input->size != 0 ? istart + input->size : istart;
-    const char* ip = input->pos != 0 ? istart + input->pos : istart;
-    char* const ostart = (char*)output->dst;
-    char* const oend = output->size != 0 ? ostart + output->size : ostart;
-    char* op = output->pos != 0 ? ostart + output->pos : ostart;
+    const char* const istart = (assert(input != NULL), (const char*)input->src);
+    const char* const iend = (istart != NULL) ? istart + input->size : istart;
+    const char* ip = (istart != NULL) ? istart + input->pos : istart;
+    char* const ostart = (assert(output != NULL), (char*)output->dst);
+    char* const oend = (ostart != NULL) ? ostart + output->size : ostart;
+    char* op = (ostart != NULL) ? ostart + output->pos : ostart;
     U32 someMoreWork = 1;
 
     /* check expectations */
-    DEBUGLOG(5, "ZSTD_compressStream_generic, flush=%u", (unsigned)flushMode);
+    DEBUGLOG(5, "ZSTD_compressStream_generic, flush=%i, srcSize = %zu", (int)flushMode, input->size - input->pos);
+    assert(zcs != NULL);
+    if (zcs->appliedParams.inBufferMode == ZSTD_bm_stable) {
+        assert(input->pos >= zcs->stableIn_notConsumed);
+        input->pos -= zcs->stableIn_notConsumed;
+        if (ip) ip -= zcs->stableIn_notConsumed;
+        zcs->stableIn_notConsumed = 0;
+    }
     if (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered) {
         assert(zcs->inBuff != NULL);
         assert(zcs->inBuffSize > 0);
···
         assert(zcs->outBuff !=  NULL);
         assert(zcs->outBuffSize > 0);
     }
-    assert(output->pos <= output->size);
+    if (input->src == NULL) assert(input->size == 0);
     assert(input->pos <= input->size);
+    if (output->dst == NULL) assert(output->size == 0);
+    assert(output->pos <= output->size);
     assert((U32)flushMode <= (U32)ZSTD_e_end);
 
     while (someMoreWork) {
···
 
         case zcss_load:
             if ( (flushMode == ZSTD_e_end)
-              && ( (size_t)(oend-op) >= ZSTD_compressBound(iend-ip)     /* Enough output space */
+              && ( (size_t)(oend-op) >= ZSTD_compressBound((size_t)(iend-ip))     /* Enough output space */
                 || zcs->appliedParams.outBufferMode == ZSTD_bm_stable)  /* OR we are allowed to return dstSizeTooSmall */
               && (zcs->inBuffPos == 0) ) {
                 /* shortcut to compression pass directly into output buffer */
-                size_t const cSize = ZSTD_compressEnd(zcs,
-                                                op, oend-op, ip, iend-ip);
+                size_t const cSize = ZSTD_compressEnd_public(zcs,
+                                                op, (size_t)(oend-op),
+                                                ip, (size_t)(iend-ip));
                 DEBUGLOG(4, "ZSTD_compressEnd : cSize=%u", (unsigned)cSize);
                 FORWARD_IF_ERROR(cSize, "ZSTD_compressEnd failed");
                 ip = iend;
···
                 size_t const toLoad = zcs->inBuffTarget - zcs->inBuffPos;
                 size_t const loaded = ZSTD_limitCopy(
                                         zcs->inBuff + zcs->inBuffPos, toLoad,
-                                        ip, iend-ip);
+                                        ip, (size_t)(iend-ip));
                 zcs->inBuffPos += loaded;
-                if (loaded != 0)
-                    ip += loaded;
+                if (ip) ip += loaded;
                 if ( (flushMode == ZSTD_e_continue)
                   && (zcs->inBuffPos < zcs->inBuffTarget) ) {
                     /* not enough input to fill full block : stop here */
···
                     /* empty */
                     someMoreWork = 0; break;
                 }
+            } else {
+                assert(zcs->appliedParams.inBufferMode == ZSTD_bm_stable);
+                if ( (flushMode == ZSTD_e_continue)
+                  && ( (size_t)(iend - ip) < zcs->blockSizeMax) ) {
+                    /* can't compress a full block : stop here */
+                    zcs->stableIn_notConsumed = (size_t)(iend - ip);
+                    ip = iend;  /* pretend to have consumed input */
+                    someMoreWork = 0; break;
+                }
+                if ( (flushMode == ZSTD_e_flush)
+                  && (ip == iend) ) {
+                    /* empty */
+                    someMoreWork = 0; break;
+                }
             }
             /* compress current block (note : this stage cannot be stopped in the middle) */
             DEBUGLOG(5, "stream compression stage (flushMode==%u)", flushMode);
             {   int const inputBuffered = (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered);
                 void* cDst;
                 size_t cSize;
-                size_t oSize = oend-op;
-                size_t const iSize = inputBuffered
-                    ? zcs->inBuffPos - zcs->inToCompress
-                    : MIN((size_t)(iend - ip), zcs->blockSize);
+                size_t oSize = (size_t)(oend-op);
+                size_t const iSize = inputBuffered ? zcs->inBuffPos - zcs->inToCompress
+                                                   : MIN((size_t)(iend - ip), zcs->blockSizeMax);
                 if (oSize >= ZSTD_compressBound(iSize) || zcs->appliedParams.outBufferMode == ZSTD_bm_stable)
                     cDst = op;   /* compress into output buffer, to skip flush stage */
                 else
···
                 if (inputBuffered) {
                     unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip==iend);
                     cSize = lastBlock ?
-                            ZSTD_compressEnd(zcs, cDst, oSize,
+                            ZSTD_compressEnd_public(zcs, cDst, oSize,
                                         zcs->inBuff + zcs->inToCompress, iSize) :
-                            ZSTD_compressContinue(zcs, cDst, oSize,
+                            ZSTD_compressContinue_public(zcs, cDst, oSize,
                                         zcs->inBuff + zcs->inToCompress, iSize);
                     FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed");
                     zcs->frameEnded = lastBlock;
                     /* prepare next block */
-                    zcs->inBuffTarget = zcs->inBuffPos + zcs->blockSize;
+                    zcs->inBuffTarget = zcs->inBuffPos + zcs->blockSizeMax;
                     if (zcs->inBuffTarget > zcs->inBuffSize)
-                        zcs->inBuffPos = 0, zcs->inBuffTarget = zcs->blockSize;
+                        zcs->inBuffPos = 0, zcs->inBuffTarget = zcs->blockSizeMax;
                     DEBUGLOG(5, "inBuffTarget:%u / inBuffSize:%u",
                             (unsigned)zcs->inBuffTarget, (unsigned)zcs->inBuffSize);
                     if (!lastBlock)
                         assert(zcs->inBuffTarget <= zcs->inBuffSize);
                     zcs->inToCompress = zcs->inBuffPos;
-                } else {
-                    unsigned const lastBlock = (ip + iSize == iend);
-                    assert(flushMode == ZSTD_e_end /* Already validated */);
+                } else { /* !inputBuffered, hence ZSTD_bm_stable */
+                    unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip + iSize == iend);
                     cSize = lastBlock ?
-                            ZSTD_compressEnd(zcs, cDst, oSize, ip, iSize) :
-                            ZSTD_compressContinue(zcs, cDst, oSize, ip, iSize);
+                            ZSTD_compressEnd_public(zcs, cDst, oSize, ip, iSize) :
+                            ZSTD_compressContinue_public(zcs, cDst, oSize, ip, iSize);
                     /* Consume the input prior to error checking to mirror buffered mode. */
-                    if (iSize > 0)
-                        ip += iSize;
+                    if (ip) ip += iSize;
                     FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed");
                     zcs->frameEnded = lastBlock;
-                    if (lastBlock)
-                        assert(ip == iend);
+                    if (lastBlock) assert(ip == iend);
                 }
                 if (cDst == op) {  /* no need to flush */
                     op += cSize;
···
         }
     }
 
-    input->pos = ip - istart;
-    output->pos = op - ostart;
+    input->pos = (size_t)(ip - istart);
+    output->pos = (size_t)(op - ostart);
     if (zcs->frameEnded) return 0;
     return ZSTD_nextInputSizeHint(zcs);
 }
···
 /* After a compression call set the expected input/output buffer.
  * This is validated at the start of the next compression call.
  */
-static void ZSTD_setBufferExpectations(ZSTD_CCtx* cctx, ZSTD_outBuffer const* output, ZSTD_inBuffer const* input)
+static void
+ZSTD_setBufferExpectations(ZSTD_CCtx* cctx, const ZSTD_outBuffer* output, const ZSTD_inBuffer* input)
 {
+    DEBUGLOG(5, "ZSTD_setBufferExpectations (for advanced stable in/out modes)");
     if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
         cctx->expectedInBuffer = *input;
     }
···
 {
     if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
         ZSTD_inBuffer const expect = cctx->expectedInBuffer;
-        if (expect.src != input->src || expect.pos != input->pos || expect.size != input->size)
-            RETURN_ERROR(srcBuffer_wrong, "ZSTD_c_stableInBuffer enabled but input differs!");
-        if (endOp != ZSTD_e_end)
-            RETURN_ERROR(srcBuffer_wrong, "ZSTD_c_stableInBuffer can only be used with ZSTD_e_end!");
+        if (expect.src != input->src || expect.pos != input->pos)
+            RETURN_ERROR(stabilityCondition_notRespected, "ZSTD_c_stableInBuffer enabled but input differs!");
     }
+    (void)endOp;
     if (cctx->appliedParams.outBufferMode == ZSTD_bm_stable) {
         size_t const outBufferSize = output->size - output->pos;
         if (cctx->expectedOutBufferSize != outBufferSize)
-            RETURN_ERROR(dstBuffer_wrong, "ZSTD_c_stableOutBuffer enabled but output size differs!");
+            RETURN_ERROR(stabilityCondition_notRespected, "ZSTD_c_stableOutBuffer enabled but output size differs!");
     }
     return 0;
 }
 
+/*
+ * If @endOp == ZSTD_e_end, @inSize becomes pledgedSrcSize.
+ * Otherwise, it's ignored.
+ * @return: 0 on success, or a ZSTD_error code otherwise.
+ */
 static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx,
                                              ZSTD_EndDirective endOp,
-                                             size_t inSize) {
+                                             size_t inSize)
+{
     ZSTD_CCtx_params params = cctx->requestedParams;
     ZSTD_prefixDict const prefixDict = cctx->prefixDict;
     FORWARD_IF_ERROR( ZSTD_initLocalDict(cctx) , ""); /* Init the local dict if present. */
···
          */
         params.compressionLevel = cctx->cdict->compressionLevel;
     }
-    DEBUGLOG(4, "ZSTD_compressStream2 : transparent init stage");
-    if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = inSize + 1;  /* auto-fix pledgedSrcSize */
-    {
-        size_t const dictSize = prefixDict.dict
+    DEBUGLOG(4, "ZSTD_CCtx_init_compressStream2 : transparent init stage");
+    if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = inSize + 1;  /* auto-determine pledgedSrcSize */
+
+    {   size_t const dictSize = prefixDict.dict
                 ? prefixDict.dictSize
                 : (cctx->cdict ? cctx->cdict->dictContentSize : 0);
-        ZSTD_cParamMode_e const mode = ZSTD_getCParamMode(cctx->cdict, &params, cctx->pledgedSrcSizePlusOne - 1);
+        ZSTD_CParamMode_e const mode = ZSTD_getCParamMode(cctx->cdict, &params, cctx->pledgedSrcSizePlusOne - 1);
         params.cParams = ZSTD_getCParamsFromCCtxParams(
                 &params, cctx->pledgedSrcSizePlusOne-1,
                 dictSize, mode);
     }
 
-    params.useBlockSplitter = ZSTD_resolveBlockSplitterMode(params.useBlockSplitter, &params.cParams);
+    params.postBlockSplitter = ZSTD_resolveBlockSplitterMode(params.postBlockSplitter, &params.cParams);
     params.ldmParams.enableLdm = ZSTD_resolveEnableLdm(params.ldmParams.enableLdm, &params.cParams);
     params.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params.useRowMatchFinder, &params.cParams);
+    params.validateSequences = ZSTD_resolveExternalSequenceValidation(params.validateSequences);
+    params.maxBlockSize = ZSTD_resolveMaxBlockSize(params.maxBlockSize);
+    params.searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(params.searchForExternalRepcodes, params.compressionLevel);
 
     {   U64 const pledgedSrcSize = cctx->pledgedSrcSizePlusOne - 1;
         assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
···
             /* for small input: avoid automatic flush on reaching end of block, since
             * it would require to add a 3-bytes null block to end frame
             */
-            cctx->inBuffTarget = cctx->blockSize + (cctx->blockSize == pledgedSrcSize);
+            cctx->inBuffTarget = cctx->blockSizeMax + (cctx->blockSizeMax == pledgedSrcSize);
         } else {
             cctx->inBuffTarget = 0;
         }
···
     return 0;
 }
 
+/* @return provides a minimum amount of data remaining to be flushed from internal buffers
+ */
 size_t ZSTD_compressStream2( ZSTD_CCtx* cctx,
                              ZSTD_outBuffer* output,
                              ZSTD_inBuffer* input,
···
 
     /* transparent initialization stage */
     if (cctx->streamStage == zcss_init) {
-        FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, endOp, input->size), "CompressStream2 initialization failed");
-        ZSTD_setBufferExpectations(cctx, output, input);    /* Set initial buffer expectations now that we've initialized */
+        size_t const inputSize = input->size - input->pos;  /* no obligation to start from pos==0 */
+        size_t const totalInputSize = inputSize + cctx->stableIn_notConsumed;
+        if ( (cctx->requestedParams.inBufferMode == ZSTD_bm_stable) /* input is presumed stable, across invocations */
+          && (endOp == ZSTD_e_continue)                             /* no flush requested, more input to come */
+          && (totalInputSize < ZSTD_BLOCKSIZE_MAX) ) {              /* not even reached one block yet */
+            if (cctx->stableIn_notConsumed) {  /* not the first time */
+                /* check stable source guarantees */
+                RETURN_ERROR_IF(input->src != cctx->expectedInBuffer.src, stabilityCondition_notRespected, "stableInBuffer condition not respected: wrong src pointer");
+                RETURN_ERROR_IF(input->pos != cctx->expectedInBuffer.size, stabilityCondition_notRespected, "stableInBuffer condition not respected: externally modified pos");
+            }
+            /* pretend input was consumed, to give a sense forward progress */
+            input->pos = input->size;
+            /* save stable inBuffer, for later control, and flush/end */
+            cctx->expectedInBuffer = *input;
+            /* but actually input wasn't consumed, so keep track of position from where compression shall resume */
+            cctx->stableIn_notConsumed += inputSize;
+            /* don't initialize yet, wait for the first block of flush() order, for better parameters adaptation */
+            return ZSTD_FRAMEHEADERSIZE_MIN(cctx->requestedParams.format);  /* at least some header to produce */
+        }
+        FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, endOp, totalInputSize), "compressStream2 initialization failed");
+        ZSTD_setBufferExpectations(cctx, output, input);   /* Set initial buffer expectations now that we've initialized */
     }
     /* end of transparent initialization stage */
 
···
                       const void* src, size_t srcSize, size_t* srcPos,
                             ZSTD_EndDirective endOp)
 {
-    ZSTD_outBuffer output = { dst, dstCapacity, *dstPos };
-    ZSTD_inBuffer  input  = { src, srcSize, *srcPos };
+    ZSTD_outBuffer output;
+    ZSTD_inBuffer  input;
+    output.dst = dst;
+    output.size = dstCapacity;
+    output.pos = *dstPos;
+    input.src = src;
+    input.size = srcSize;
+    input.pos = *srcPos;
     /* ZSTD_compressStream2() will check validity of dstPos and srcPos */
-    size_t const cErr = ZSTD_compressStream2(cctx, &output, &input, endOp);
-    *dstPos = output.pos;
-    *srcPos = input.pos;
-    return cErr;
+    {   size_t const cErr = ZSTD_compressStream2(cctx, &output, &input, endOp);
+        *dstPos = output.pos;
+        *srcPos = input.pos;
+        return cErr;
+    }
 }
 
 size_t ZSTD_compress2(ZSTD_CCtx* cctx,
···
         /* Reset to the original values. */
         cctx->requestedParams.inBufferMode = originalInBufferMode;
         cctx->requestedParams.outBufferMode = originalOutBufferMode;
+
         FORWARD_IF_ERROR(result, "ZSTD_compressStream2_simpleArgs failed");
         if (result != 0) {  /* compression not completed, due to lack of output space */
             assert(oPos == dstCapacity);
···
     }
 }
 
-typedef struct {
-    U32 idx;             /* Index in array of ZSTD_Sequence */
-    U32 posInSequence;   /* Position within sequence at idx */
-    size_t posInSrc;        /* Number of bytes given by sequences provided so far */
-} ZSTD_sequencePosition;
-
 /* ZSTD_validateSequence() :
- * @offCode : is presumed to follow format required by ZSTD_storeSeq()
+ * @offBase : must use the format required by ZSTD_storeSeq()
  * @returns a ZSTD error code if sequence is not valid
  */
 static size_t
-ZSTD_validateSequence(U32 offCode, U32 matchLength,
-                      size_t posInSrc, U32 windowLog, size_t dictSize)
+ZSTD_validateSequence(U32 offBase, U32 matchLength, U32 minMatch,
+                      size_t posInSrc, U32 windowLog, size_t dictSize, int useSequenceProducer)
 {
-    U32 const windowSize = 1 << windowLog;
+    U32 const windowSize = 1u << windowLog;
     /* posInSrc represents the amount of data the decoder would decode up to this point.
      * As long as the amount of data decoded is less than or equal to window size, offsets may be
      * larger than the total length of output decoded in order to reference the dict, even larger than
      * window size. After output surpasses windowSize, we're limited to windowSize offsets again.
      */
     size_t const offsetBound = posInSrc > windowSize ? (size_t)windowSize : posInSrc + (size_t)dictSize;
-    RETURN_ERROR_IF(offCode > STORE_OFFSET(offsetBound), corruption_detected, "Offset too large!");
-    RETURN_ERROR_IF(matchLength < MINMATCH, corruption_detected, "Matchlength too small");
+    size_t const matchLenLowerBound = (minMatch == 3 || useSequenceProducer) ? 3 : 4;
+    RETURN_ERROR_IF(offBase > OFFSET_TO_OFFBASE(offsetBound), externalSequences_invalid, "Offset too large!");
+    /* Validate maxNbSeq is large enough for the given matchLength and minMatch */
+    RETURN_ERROR_IF(matchLength < matchLenLowerBound, externalSequences_invalid, "Matchlength too small for the minMatch");
     return 0;
 }
 
 /* Returns an offset code, given a sequence's raw offset, the ongoing repcode array, and whether litLength == 0 */
-static U32 ZSTD_finalizeOffCode(U32 rawOffset, const U32 rep[ZSTD_REP_NUM], U32 ll0)
+static U32 ZSTD_finalizeOffBase(U32 rawOffset, const U32 rep[ZSTD_REP_NUM], U32 ll0)
 {
-    U32 offCode = STORE_OFFSET(rawOffset);
+    U32 offBase = OFFSET_TO_OFFBASE(rawOffset);
 
     if (!ll0 && rawOffset == rep[0]) {
-        offCode = STORE_REPCODE_1;
+        offBase = REPCODE1_TO_OFFBASE;
     } else if (rawOffset == rep[1]) {
-        offCode = STORE_REPCODE(2 - ll0);
+        offBase = REPCODE_TO_OFFBASE(2 - ll0);
     } else if (rawOffset == rep[2]) {
-        offCode = STORE_REPCODE(3 - ll0);
+        offBase = REPCODE_TO_OFFBASE(3 - ll0);
     } else if (ll0 && rawOffset == rep[0] - 1) {
-        offCode = STORE_REPCODE_3;
+        offBase = REPCODE3_TO_OFFBASE;
     }
-    return offCode;
+    return offBase;
 }
 
-/* Returns 0 on success, and a ZSTD_error otherwise. This function scans through an array of
- * ZSTD_Sequence, storing the sequences it finds, until it reaches a block delimiter.
+/* This function scans through an array of ZSTD_Sequence,
+ * storing the sequences it reads, until it reaches a block delimiter.
+ * Note that the block delimiter includes the last literals of the block.
+ * @blockSize must be == sum(sequence_lengths).
+ * @returns @blockSize on success, and a ZSTD_error otherwise.
  */
 static size_t
-ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx,
-                                              ZSTD_sequencePosition* seqPos,
-                                        const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
-                                        const void* src, size_t blockSize)
+ZSTD_transferSequences_wBlockDelim(ZSTD_CCtx* cctx,
+                                   ZSTD_SequencePosition* seqPos,
+                             const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                             const void* src, size_t blockSize,
+                                   ZSTD_ParamSwitch_e externalRepSearch)
 {
     U32 idx = seqPos->idx;
+    U32 const startIdx = idx;
     BYTE const* ip = (BYTE const*)(src);
     const BYTE* const iend = ip + blockSize;
-    repcodes_t updatedRepcodes;
+    Repcodes_t updatedRepcodes;
     U32 dictSize;
+
+    DEBUGLOG(5, "ZSTD_transferSequences_wBlockDelim (blockSize = %zu)", blockSize);
 
     if (cctx->cdict) {
         dictSize = (U32)cctx->cdict->dictContentSize;
···
     } else {
         dictSize = 0;
     }
-    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(repcodes_t));
-    for (; (inSeqs[idx].matchLength != 0 || inSeqs[idx].offset != 0) && idx < inSeqsSize; ++idx) {
+    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(Repcodes_t));
+    for (; idx < inSeqsSize && (inSeqs[idx].matchLength != 0 || inSeqs[idx].offset != 0); ++idx) {
         U32 const litLength = inSeqs[idx].litLength;
-        U32 const ll0 = (litLength == 0);
         U32 const matchLength = inSeqs[idx].matchLength;
-        U32 const offCode = ZSTD_finalizeOffCode(inSeqs[idx].offset, updatedRepcodes.rep, ll0);
-        ZSTD_updateRep(updatedRepcodes.rep, offCode, ll0);
+        U32 offBase;
 
-        DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offCode, matchLength, litLength);
+        if (externalRepSearch == ZSTD_ps_disable) {
+            offBase = OFFSET_TO_OFFBASE(inSeqs[idx].offset);
+        } else {
+            U32 const ll0 = (litLength == 0);
+            offBase = ZSTD_finalizeOffBase(inSeqs[idx].offset, updatedRepcodes.rep, ll0);
+            ZSTD_updateRep(updatedRepcodes.rep, offBase, ll0);
+        }
+
+        DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offBase, matchLength, litLength);
         if (cctx->appliedParams.validateSequences) {
             seqPos->posInSrc += litLength + matchLength;
-            FORWARD_IF_ERROR(ZSTD_validateSequence(offCode, matchLength, seqPos->posInSrc,
-                                                cctx->appliedParams.cParams.windowLog, dictSize),
+            FORWARD_IF_ERROR(ZSTD_validateSequence(offBase, matchLength, cctx->appliedParams.cParams.minMatch,
+                                                seqPos->posInSrc,
+                                                cctx->appliedParams.cParams.windowLog, dictSize,
+                                                ZSTD_hasExtSeqProd(&cctx->appliedParams)),
                                                 "Sequence validation failed");
         }
-        RETURN_ERROR_IF(idx - seqPos->idx > cctx->seqStore.maxNbSeq, memory_allocation,
+        RETURN_ERROR_IF(idx - seqPos->idx >= cctx->seqStore.maxNbSeq, externalSequences_invalid,
                         "Not enough memory allocated. Try adjusting ZSTD_c_minMatch.");
-        ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offCode, matchLength);
+        ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offBase, matchLength);
         ip += matchLength + litLength;
     }
-    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(repcodes_t));
+    RETURN_ERROR_IF(idx == inSeqsSize, externalSequences_invalid, "Block delimiter not found.");
+
+    /* If we skipped repcode search while parsing, we need to update repcodes now */
+    assert(externalRepSearch != ZSTD_ps_auto);
+    assert(idx >= startIdx);
+    if (externalRepSearch == ZSTD_ps_disable && idx != startIdx) {
+        U32* const rep = updatedRepcodes.rep;
+        U32 lastSeqIdx = idx - 1; /* index of last non-block-delimiter sequence */
+
+        if (lastSeqIdx >= startIdx + 2) {
+            rep[2] = inSeqs[lastSeqIdx - 2].offset;
+            rep[1] = inSeqs[lastSeqIdx - 1].offset;
+            rep[0] = inSeqs[lastSeqIdx].offset;
+        } else if (lastSeqIdx == startIdx + 1) {
+            rep[2] = rep[0];
+            rep[1] = inSeqs[lastSeqIdx - 1].offset;
+            rep[0] = inSeqs[lastSeqIdx].offset;
+        } else {
+            assert(lastSeqIdx == startIdx);
+            rep[2] = rep[1];
+            rep[1] = rep[0];
+            rep[0] = inSeqs[lastSeqIdx].offset;
+        }
+    }
+
+    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(Repcodes_t));
 
     if (inSeqs[idx].litLength) {
         DEBUGLOG(6, "Storing last literals of size: %u", inSeqs[idx].litLength);
···
         ip += inSeqs[idx].litLength;
         seqPos->posInSrc += inSeqs[idx].litLength;
     }
-    RETURN_ERROR_IF(ip != iend, corruption_detected, "Blocksize doesn't agree with block delimiter!");
+    RETURN_ERROR_IF(ip != iend, externalSequences_invalid, "Blocksize doesn't agree with block delimiter!");
     seqPos->idx = idx+1;
-    return 0;
+    return blockSize;
 }
 
-/* Returns the number of bytes to move the current read position back by. Only non-zero
- * if we ended up splitting a sequence. Otherwise, it may return a ZSTD error if something
- * went wrong.
+/*
+ * This function attempts to scan through @blockSize bytes in @src
+ * represented by the sequences in @inSeqs,
+ * storing any (partial) sequences.
  *
- * This function will attempt to scan through blockSize bytes represented by the sequences
- * in inSeqs, storing any (partial) sequences.
+ * Occasionally, we may want to reduce the actual number of bytes consumed from @src
+ * to avoid splitting a match, notably if it would produce a match smaller than MINMATCH.
  *
- * Occasionally, we may want to change the actual number of bytes we consumed from inSeqs to
- * avoid splitting a match, or to avoid splitting a match such that it would produce a match
- * smaller than MINMATCH. In this case, we return the number of bytes that we didn't read from this block.
+ * @returns the number of bytes consumed from @src, necessarily <= @blockSize.
+ * Otherwise, it may return a ZSTD error if something went wrong.
  */
 static size_t
-ZSTD_copySequencesToSeqStoreNoBlockDelim(ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos,
-                                   const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
-                                   const void* src, size_t blockSize)
+ZSTD_transferSequences_noDelim(ZSTD_CCtx* cctx,
+                               ZSTD_SequencePosition* seqPos,
+                         const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                         const void* src, size_t blockSize,
+                               ZSTD_ParamSwitch_e externalRepSearch)
 {
     U32 idx = seqPos->idx;
     U32 startPosInSequence = seqPos->posInSequence;
     U32 endPosInSequence = seqPos->posInSequence + (U32)blockSize;
     size_t dictSize;
-    BYTE const* ip = (BYTE const*)(src);
-    BYTE const* iend = ip + blockSize;  /* May be adjusted if we decide to process fewer than blockSize bytes */
-    repcodes_t updatedRepcodes;
+    const BYTE* const istart = (const BYTE*)(src);
+    const BYTE* ip = istart;
+    const BYTE* iend = istart + blockSize;  /* May be adjusted if we decide to process fewer than blockSize bytes */
+    Repcodes_t updatedRepcodes;
     U32 bytesAdjustment = 0;
     U32 finalMatchSplit = 0;
+
+    /* TODO(embg) support fast parsing mode in noBlockDelim mode */
+    (void)externalRepSearch;
 
     if (cctx->cdict) {
         dictSize = cctx->cdict->dictContentSize;
···
     } else {
         dictSize = 0;
     }
-    DEBUGLOG(5, "ZSTD_copySequencesToSeqStore: idx: %u PIS: %u blockSize: %zu", idx, startPosInSequence, blockSize);
+    DEBUGLOG(5, "ZSTD_transferSequences_noDelim: idx: %u PIS: %u blockSize: %zu", idx, startPosInSequence, blockSize);
     DEBUGLOG(5, "Start seq: idx: %u (of: %u ml: %u ll: %u)", idx, inSeqs[idx].offset, inSeqs[idx].matchLength, inSeqs[idx].litLength);
-    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(repcodes_t));
+    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(Repcodes_t));
     while (endPosInSequence && idx < inSeqsSize && !finalMatchSplit) {
         const ZSTD_Sequence currSeq = inSeqs[idx];
         U32 litLength = currSeq.litLength;
         U32 matchLength = currSeq.matchLength;
         U32 const rawOffset = currSeq.offset;
-        U32 offCode;
+        U32 offBase;
 
         /* Modify the sequence depending on where endPosInSequence lies */
         if (endPosInSequence >= currSeq.litLength + currSeq.matchLength) {
···
             /* Move to the next sequence */
             endPosInSequence -= currSeq.litLength + currSeq.matchLength;
             startPosInSequence = 0;
-            idx++;
         } else {
             /* This is the final (partial) sequence we're adding from inSeqs, and endPosInSequence
                does not reach the end of the match. So, we have to split the sequence */
···
         }
         /* Check if this offset can be represented with a repcode */
         {   U32 const ll0 = (litLength == 0);
-            offCode = ZSTD_finalizeOffCode(rawOffset, updatedRepcodes.rep, ll0);
-            ZSTD_updateRep(updatedRepcodes.rep, offCode, ll0);
+            offBase = ZSTD_finalizeOffBase(rawOffset, updatedRepcodes.rep, ll0);
+            ZSTD_updateRep(updatedRepcodes.rep, offBase, ll0);
         }
 
         if (cctx->appliedParams.validateSequences) {
             seqPos->posInSrc += litLength + matchLength;
-            FORWARD_IF_ERROR(ZSTD_validateSequence(offCode, matchLength, seqPos->posInSrc,
-                                                   cctx->appliedParams.cParams.windowLog, dictSize),
+            FORWARD_IF_ERROR(ZSTD_validateSequence(offBase, matchLength, cctx->appliedParams.cParams.minMatch, seqPos->posInSrc,
+                                                   cctx->appliedParams.cParams.windowLog, dictSize, ZSTD_hasExtSeqProd(&cctx->appliedParams)),
                                                    "Sequence validation failed");
         }
-        DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offCode, matchLength, litLength);
-        RETURN_ERROR_IF(idx - seqPos->idx > cctx->seqStore.maxNbSeq, memory_allocation,
+        DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offBase, matchLength, litLength);
+        RETURN_ERROR_IF(idx - seqPos->idx >= cctx->seqStore.maxNbSeq, externalSequences_invalid,
                         "Not enough memory allocated. Try adjusting ZSTD_c_minMatch.");
-        ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offCode, matchLength);
+        ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offBase, matchLength);
         ip += matchLength + litLength;
+        if (!finalMatchSplit)
+            idx++; /* Next Sequence */
     }
     DEBUGLOG(5, "Ending seq: idx: %u (of: %u ml: %u ll: %u)", idx, inSeqs[idx].offset, inSeqs[idx].matchLength, inSeqs[idx].litLength);
     assert(idx == inSeqsSize || endPosInSequence <= inSeqs[idx].litLength + inSeqs[idx].matchLength);
     seqPos->idx = idx;
     seqPos->posInSequence = endPosInSequence;
-    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(repcodes_t));
+    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(Repcodes_t));
 
     iend -= bytesAdjustment;
     if (ip != iend) {
         /* Store any last literals */
-        U32 lastLLSize = (U32)(iend - ip);
+        U32 const lastLLSize = (U32)(iend - ip);
         assert(ip <= iend);
         DEBUGLOG(6, "Storing last literals of size: %u", lastLLSize);
         ZSTD_storeLastLiterals(&cctx->seqStore, ip, lastLLSize);
         seqPos->posInSrc += lastLLSize;
     }
 
-    return bytesAdjustment;
+    return (size_t)(iend-istart);
 }
 
-typedef size_t (*ZSTD_sequenceCopier) (ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos,
-                                       const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
-                                       const void* src, size_t blockSize);
-static ZSTD_sequenceCopier ZSTD_selectSequenceCopier(ZSTD_sequenceFormat_e mode)
+/* @seqPos represents a position within @inSeqs,
+ * it is read and updated by this function,
+ * once the goal to produce a block of size @blockSize is reached.
+ * @return: nb of bytes consumed from @src, necessarily <= @blockSize.
+ */
+typedef size_t (*ZSTD_SequenceCopier_f)(ZSTD_CCtx* cctx,
+                                        ZSTD_SequencePosition* seqPos,
+                                  const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                  const void* src, size_t blockSize,
+                                        ZSTD_ParamSwitch_e externalRepSearch);
+
+static ZSTD_SequenceCopier_f ZSTD_selectSequenceCopier(ZSTD_SequenceFormat_e mode)
 {
-    ZSTD_sequenceCopier sequenceCopier = NULL;
-    assert(ZSTD_cParam_withinBounds(ZSTD_c_blockDelimiters, mode));
+    assert(ZSTD_cParam_withinBounds(ZSTD_c_blockDelimiters, (int)mode));
     if (mode == ZSTD_sf_explicitBlockDelimiters) {
-        return ZSTD_copySequencesToSeqStoreExplicitBlockDelim;
-    } else if (mode == ZSTD_sf_noBlockDelimiters) {
-        return ZSTD_copySequencesToSeqStoreNoBlockDelim;
+        return ZSTD_transferSequences_wBlockDelim;
     }
-    assert(sequenceCopier != NULL);
-    return sequenceCopier;
+    assert(mode == ZSTD_sf_noBlockDelimiters);
+    return ZSTD_transferSequences_noDelim;
 }
 
-/* Compress, block-by-block, all of the sequences given.
+/* Discover the size of next block by searching for the delimiter.
+ * Note that a block delimiter **must** exist in this mode,
+ * otherwise it's an input error.
+ * The block size retrieved will be later compared to ensure it remains within bounds */
+static size_t
+blockSize_explicitDelimiter(const ZSTD_Sequence* inSeqs, size_t inSeqsSize, ZSTD_SequencePosition seqPos)
+{
+    int end = 0;
+    size_t blockSize = 0;
+    size_t spos = seqPos.idx;
+    DEBUGLOG(6, "blockSize_explicitDelimiter : seq %zu / %zu", spos, inSeqsSize);
+    assert(spos <= inSeqsSize);
+    while (spos < inSeqsSize) {
+        end = (inSeqs[spos].offset == 0);
+        blockSize += inSeqs[spos].litLength + inSeqs[spos].matchLength;
+        if (end) {
+            if (inSeqs[spos].matchLength != 0)
+                RETURN_ERROR(externalSequences_invalid, "delimiter format error : both matchlength and offset must be == 0");
+            break;
+        }
+        spos++;
+    }
+    if (!end)
+        RETURN_ERROR(externalSequences_invalid, "Reached end of sequences without finding a block delimiter");
+    return blockSize;
+}
+
+static size_t determine_blockSize(ZSTD_SequenceFormat_e mode,
+                           size_t blockSize, size_t remaining,
+                     const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
+                           ZSTD_SequencePosition seqPos)
+{
+    DEBUGLOG(6, "determine_blockSize : remainingSize = %zu", remaining);
+    if (mode == ZSTD_sf_noBlockDelimiters) {
+        /* Note: more a "target" block size */
+        return MIN(remaining, blockSize);
+    }
+    assert(mode == ZSTD_sf_explicitBlockDelimiters);
+    {   size_t const explicitBlockSize = blockSize_explicitDelimiter(inSeqs, inSeqsSize, seqPos);
+        FORWARD_IF_ERROR(explicitBlockSize, "Error while determining block size with explicit delimiters");
+        if (explicitBlockSize > blockSize)
+            RETURN_ERROR(externalSequences_invalid, "sequences incorrectly define a too large block");
+        if (explicitBlockSize > remaining)
+            RETURN_ERROR(externalSequences_invalid, "sequences define a frame longer than source");
+        return explicitBlockSize;
+    }
+}
+
+/* Compress all provided sequences, block-by-block.
  *
  * Returns the cumulative size of all compressed blocks (including their headers),
  * otherwise a ZSTD error.
···
                           const void* src, size_t srcSize)
 {
     size_t cSize = 0;
-    U32 lastBlock;
-    size_t blockSize;
-    size_t compressedSeqsSize;
     size_t remaining = srcSize;
-    ZSTD_sequencePosition seqPos = {0, 0, 0};
+    ZSTD_SequencePosition seqPos = {0, 0, 0};
 
-    BYTE const* ip = (BYTE const*)src;
+    const BYTE* ip = (BYTE const*)src;
     BYTE* op = (BYTE*)dst;
-    ZSTD_sequenceCopier const sequenceCopier = ZSTD_selectSequenceCopier(cctx->appliedParams.blockDelimiters);
+    ZSTD_SequenceCopier_f const sequenceCopier = ZSTD_selectSequenceCopier(cctx->appliedParams.blockDelimiters);
 
     DEBUGLOG(4, "ZSTD_compressSequences_internal srcSize: %zu, inSeqsSize: %zu", srcSize, inSeqsSize);
     /* Special case: empty frame */
···
     }
 
     while (remaining) {
+        size_t compressedSeqsSize;
         size_t cBlockSize;
-        size_t additionalByteAdjustment;
-        lastBlock = remaining <= cctx->blockSize;
-        blockSize = lastBlock ? (U32)remaining : (U32)cctx->blockSize;
+        size_t blockSize = determine_blockSize(cctx->appliedParams.blockDelimiters,
+                                        cctx->blockSizeMax, remaining,
+                                        inSeqs, inSeqsSize, seqPos);
+        U32 const lastBlock = (blockSize == remaining);
+        FORWARD_IF_ERROR(blockSize, "Error while trying to determine block size");
+        assert(blockSize <= remaining);
         ZSTD_resetSeqStore(&cctx->seqStore);
-        DEBUGLOG(4, "Working on new block. Blocksize: %zu", blockSize);
 
-        additionalByteAdjustment = sequenceCopier(cctx, &seqPos, inSeqs, inSeqsSize, ip, blockSize);
-        FORWARD_IF_ERROR(additionalByteAdjustment, "Bad sequence copy");
-        blockSize -= additionalByteAdjustment;
+        blockSize = sequenceCopier(cctx,
+                                   &seqPos, inSeqs, inSeqsSize,
+                                   ip, blockSize,
+                                   cctx->appliedParams.searchForExternalRepcodes);
+        FORWARD_IF_ERROR(blockSize, "Bad sequence copy");
 
         /* If blocks are too small, emit as a nocompress block */
-        if (blockSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1) {
+        /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
+         * additional 1. We need to revisit and change this logic to be more consistent */
+        if (blockSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1+1) {
             cBlockSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock);
             FORWARD_IF_ERROR(cBlockSize, "Nocompress block failed");
-            DEBUGLOG(4, "Block too small, writing out nocompress block: cSize: %zu", cBlockSize);
+            DEBUGLOG(5, "Block too small (%zu): data remains uncompressed: cSize=%zu", blockSize, cBlockSize);
             cSize += cBlockSize;
             ip += blockSize;
             op += cBlockSize;
···
             continue;
         }
 
+        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "not enough dstCapacity to write a new compressed block");
         compressedSeqsSize = ZSTD_entropyCompressSeqStore(&cctx->seqStore,
                                 &cctx->blockState.prevCBlock->entropy, &cctx->blockState.nextCBlock->entropy,
                                 &cctx->appliedParams,
                                 op + ZSTD_blockHeaderSize /* Leave space for block header */, dstCapacity - ZSTD_blockHeaderSize,
                                 blockSize,
-                                cctx->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
+                                cctx->tmpWorkspace, cctx->tmpWkspSize /* statically allocated in resetCCtx */,
                                 cctx->bmi2);
         FORWARD_IF_ERROR(compressedSeqsSize, "Compressing sequences of block failed");
-        DEBUGLOG(4, "Compressed sequences size: %zu", compressedSeqsSize);
+        DEBUGLOG(5, "Compressed sequences size: %zu", compressedSeqsSize);
 
         if (!cctx->isFirstBlock &&
             ZSTD_maybeRLE(&cctx->seqStore) &&
-            ZSTD_isRLE((BYTE const*)src, srcSize)) {
-            /* We don't want to emit our first block as a RLE even if it qualifies because
-            * doing so will cause the decoder (cli only) to throw a "should consume all input error."
-            * This is only an issue for zstd <= v1.4.3
-            */
+            ZSTD_isRLE(ip, blockSize)) {
+            /* Note: don't emit the first block as RLE even if it qualifies because
+             * doing so will cause the decoder (cli <= v1.4.3 only) to throw an (invalid) error
+             * "should consume all input error."
+             */
             compressedSeqsSize = 1;
         }
 
         if (compressedSeqsSize == 0) {
             /* ZSTD_noCompressBlock writes the block header as well */
             cBlockSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock);
-            FORWARD_IF_ERROR(cBlockSize, "Nocompress block failed");
-            DEBUGLOG(4, "Writing out nocompress block, size: %zu", cBlockSize);
+            FORWARD_IF_ERROR(cBlockSize, "ZSTD_noCompressBlock failed");
+            DEBUGLOG(5, "Writing out nocompress block, size: %zu", cBlockSize);
         } else if (compressedSeqsSize == 1) {
             cBlockSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, blockSize, lastBlock);
-            FORWARD_IF_ERROR(cBlockSize, "RLE compress block failed");
-            DEBUGLOG(4, "Writing out RLE block, size: %zu", cBlockSize);
+            FORWARD_IF_ERROR(cBlockSize, "ZSTD_rleCompressBlock failed");
+            DEBUGLOG(5, "Writing out RLE block, size: %zu", cBlockSize);
         } else {
             U32 cBlockHeader;
             /* Error checking and repcodes update */
···
             cBlockHeader = lastBlock + (((U32)bt_compressed)<<1) + (U32)(compressedSeqsSize << 3);
             MEM_writeLE24(op, cBlockHeader);
             cBlockSize = ZSTD_blockHeaderSize + compressedSeqsSize;
-            DEBUGLOG(4, "Writing out compressed block, size: %zu", cBlockSize);
+            DEBUGLOG(5, "Writing out compressed block, size: %zu", cBlockSize);
         }
 
         cSize += cBlockSize;
-        DEBUGLOG(4, "cSize running total: %zu", cSize);
 
         if (lastBlock) {
             break;
···
             dstCapacity -= cBlockSize;
             cctx->isFirstBlock = 0;
         }
+        DEBUGLOG(5, "cSize running total: %zu (remaining dstCapacity=%zu)", cSize, dstCapacity);
     }
 
+    DEBUGLOG(4, "cSize final total: %zu", cSize);
     return cSize;
 }
 
-size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* dst, size_t dstCapacity,
+size_t ZSTD_compressSequences(ZSTD_CCtx* cctx,
+                              void* dst, size_t dstCapacity,
                               const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
                               const void* src, size_t srcSize)
 {
     BYTE* op = (BYTE*)dst;
     size_t cSize = 0;
-    size_t compressedBlocksSize = 0;
-    size_t frameHeaderSize = 0;
 
     /* Transparent initialization stage, same as compressStream2() */
-    DEBUGLOG(3, "ZSTD_compressSequences()");
+    DEBUGLOG(4, "ZSTD_compressSequences (nbSeqs=%zu,dstCapacity=%zu)", inSeqsSize, dstCapacity);
     assert(cctx != NULL);
     FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, ZSTD_e_end, srcSize), "CCtx initialization failed");
+
     /* Begin writing output, starting with frame header */
-    frameHeaderSize = ZSTD_writeFrameHeader(op, dstCapacity, &cctx->appliedParams, srcSize, cctx->dictID);
-    op += frameHeaderSize;
-    dstCapacity -= frameHeaderSize;
-    cSize += frameHeaderSize;
+    {   size_t const frameHeaderSize = ZSTD_writeFrameHeader(op, dstCapacity,
+                    &cctx->appliedParams, srcSize, cctx->dictID);
+        op += frameHeaderSize;
+        assert(frameHeaderSize <= dstCapacity);
+        dstCapacity -= frameHeaderSize;
+        cSize += frameHeaderSize;
+    }
     if (cctx->appliedParams.fParams.checksumFlag && srcSize) {
         xxh64_update(&cctx->xxhState, src, srcSize);
     }
-    /* cSize includes block header size and compressed sequences size */
-    compressedBlocksSize = ZSTD_compressSequences_internal(cctx,
+
+    /* Now generate compressed blocks */
+    {   size_t const cBlocksSize = ZSTD_compressSequences_internal(cctx,
                                                            op, dstCapacity,
                                                            inSeqs, inSeqsSize,
                                                            src, srcSize);
-    FORWARD_IF_ERROR(compressedBlocksSize, "Compressing blocks failed!");
-    cSize += compressedBlocksSize;
-    dstCapacity -= compressedBlocksSize;
+        FORWARD_IF_ERROR(cBlocksSize, "Compressing blocks failed!");
+        cSize += cBlocksSize;
+        assert(cBlocksSize <= dstCapacity);
+        dstCapacity -= cBlocksSize;
+    }
 
+    /* Complete with frame checksum, if needed */
     if (cctx->appliedParams.fParams.checksumFlag) {
         U32 const checksum = (U32) xxh64_digest(&cctx->xxhState);
         RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "no room for checksum");
···
         cSize += 4;
     }
 
-    DEBUGLOG(3, "Final compressed size: %zu", cSize);
+    DEBUGLOG(4, "Final compressed size: %zu", cSize);
+    return cSize;
+}
+
+
+#if defined(__AVX2__)
+
+#include <immintrin.h>  /* AVX2 intrinsics */
+
+/*
+ * Convert 2 sequences per iteration, using AVX2 intrinsics:
+ *   - offset -> offBase = offset + 2
+ *   - litLength -> (U16) litLength
+ *   - matchLength -> (U16)(matchLength - 3)
+ *   - rep is ignored
+ * Store only 8 bytes per SeqDef (offBase[4], litLength[2], mlBase[2]).
+ *
+ * At the end, instead of extracting two __m128i,
+ * we use _mm256_permute4x64_epi64(..., 0xE8) to move lane2 into lane1,
+ * then store the lower 16 bytes in one go.
+ *
+ * @returns 0 on succes, with no long length detected
+ * @returns > 0 if there is one long length (> 65535),
+ * indicating the position, and type.
+ */
+static size_t convertSequences_noRepcodes(
+    SeqDef* dstSeqs,
+    const ZSTD_Sequence* inSeqs,
+    size_t nbSequences)
+{
+    /*
+     * addition:
+     *   For each 128-bit half: (offset+2, litLength+0, matchLength-3, rep+0)
+     */
+    const __m256i addition = _mm256_setr_epi32(
+        ZSTD_REP_NUM, 0, -MINMATCH, 0,    /* for sequence i */
+        ZSTD_REP_NUM, 0, -MINMATCH, 0     /* for sequence i+1 */
+    );
+
+    /* limit: check if there is a long length */
+    const __m256i limit = _mm256_set1_epi32(65535);
+
+    /*
+     * shuffle mask for byte-level rearrangement in each 128-bit half:
+     *
+     * Input layout (after addition) per 128-bit half:
+     *   [ offset+2 (4 bytes) | litLength (4 bytes) | matchLength (4 bytes) | rep (4 bytes) ]
+     * We only need:
+     *   offBase (4 bytes) = offset+2
+     *   litLength (2 bytes) = low 2 bytes of litLength
+     *   mlBase (2 bytes) = low 2 bytes of (matchLength)
+     * => Bytes [0..3, 4..5, 8..9], zero the rest.
+     */
+    const __m256i mask = _mm256_setr_epi8(
+        /* For the lower 128 bits => sequence i */
+         0, 1, 2, 3,       /* offset+2 */
+         4, 5,             /* litLength (16 bits) */
+         8, 9,             /* matchLength (16 bits) */
+         (BYTE)0x80, (BYTE)0x80, (BYTE)0x80, (BYTE)0x80,
+         (BYTE)0x80, (BYTE)0x80, (BYTE)0x80, (BYTE)0x80,
+
+        /* For the upper 128 bits => sequence i+1 */
+        16,17,18,19,       /* offset+2 */
+        20,21,             /* litLength */
+        24,25,             /* matchLength */
+        (BYTE)0x80, (BYTE)0x80, (BYTE)0x80, (BYTE)0x80,
+        (BYTE)0x80, (BYTE)0x80, (BYTE)0x80, (BYTE)0x80
+    );
+
+    /*
+     * Next, we'll use _mm256_permute4x64_epi64(vshf, 0xE8).
+     * Explanation of 0xE8 = 11101000b => [lane0, lane2, lane2, lane3].
+     * So the lower 128 bits become [lane0, lane2] => combining seq0 and seq1.
+     */
+#define PERM_LANE_0X_E8 0xE8  /* [0,2,2,3] in lane indices */
+
+    size_t longLen = 0, i = 0;
+
+    /* AVX permutation depends on the specific definition of target structures */
+    ZSTD_STATIC_ASSERT(sizeof(ZSTD_Sequence) == 16);
+    ZSTD_STATIC_ASSERT(offsetof(ZSTD_Sequence, offset) == 0);
+    ZSTD_STATIC_ASSERT(offsetof(ZSTD_Sequence, litLength) == 4);
+    ZSTD_STATIC_ASSERT(offsetof(ZSTD_Sequence, matchLength) == 8);
+    ZSTD_STATIC_ASSERT(sizeof(SeqDef) == 8);
+    ZSTD_STATIC_ASSERT(offsetof(SeqDef, offBase) == 0);
+    ZSTD_STATIC_ASSERT(offsetof(SeqDef, litLength) == 4);
+    ZSTD_STATIC_ASSERT(offsetof(SeqDef, mlBase) == 6);
+
+    /* Process 2 sequences per loop iteration */
+    for (; i + 1 < nbSequences; i += 2) {
+        /* Load 2 ZSTD_Sequence (32 bytes) */
+        __m256i vin  = _mm256_loadu_si256((const __m256i*)(const void*)&inSeqs[i]);
+
+        /* Add {2, 0, -3, 0} in each 128-bit half */
+        __m256i vadd = _mm256_add_epi32(vin, addition);
+
+        /* Check for long length */
+        __m256i ll_cmp  = _mm256_cmpgt_epi32(vadd, limit);  /* 0xFFFFFFFF for element > 65535 */
+        int ll_res  = _mm256_movemask_epi8(ll_cmp);
+
+        /* Shuffle bytes so each half gives us the 8 bytes we need */
+        __m256i vshf = _mm256_shuffle_epi8(vadd, mask);
+        /*
+         * Now:
+         *   Lane0 = seq0's 8 bytes
+         *   Lane1 = 0
+         *   Lane2 = seq1's 8 bytes
+         *   Lane3 = 0
+         */
+
+        /* Permute 64-bit lanes => move Lane2 down into Lane1. */
+        __m256i vperm = _mm256_permute4x64_epi64(vshf, PERM_LANE_0X_E8);
+        /*
+         * Now the lower 16 bytes (Lane0+Lane1) = [seq0, seq1].
+         * The upper 16 bytes are [Lane2, Lane3] = [seq1, 0], but we won't use them.
+         */
+
+        /* Store only the lower 16 bytes => 2 SeqDef (8 bytes each) */
+        _mm_storeu_si128((__m128i *)(void*)&dstSeqs[i], _mm256_castsi256_si128(vperm));
+        /*
+         * This writes out 16 bytes total:
+         *   - offset 0..7  => seq0 (offBase, litLength, mlBase)
+         *   - offset 8..15 => seq1 (offBase, litLength, mlBase)
+         */
+
+        /* check (unlikely) long lengths > 65535
+         * indices for lengths correspond to bits [4..7], [8..11], [20..23], [24..27]
+         * => combined mask = 0x0FF00FF0
+         */
+        if (UNLIKELY((ll_res & 0x0FF00FF0) != 0)) {
+            /* long length detected: let's figure out which one*/
+            if (inSeqs[i].matchLength > 65535+MINMATCH) {
+                assert(longLen == 0);
+                longLen = i + 1;
+            }
+            if (inSeqs[i].litLength > 65535) {
+                assert(longLen == 0);
+                longLen = i + nbSequences + 1;
+            }
+            if (inSeqs[i+1].matchLength > 65535+MINMATCH) {
+                assert(longLen == 0);
+                longLen = i + 1 + 1;
+            }
+            if (inSeqs[i+1].litLength > 65535) {
+                assert(longLen == 0);
+                longLen = i + 1 + nbSequences + 1;
+            }
+        }
+    }
+
+    /* Handle leftover if @nbSequences is odd */
+    if (i < nbSequences) {
+        /* process last sequence */
+        assert(i == nbSequences - 1);
+        dstSeqs[i].offBase = OFFSET_TO_OFFBASE(inSeqs[i].offset);
+        dstSeqs[i].litLength = (U16)inSeqs[i].litLength;
+        dstSeqs[i].mlBase = (U16)(inSeqs[i].matchLength - MINMATCH);
+        /* check (unlikely) long lengths > 65535 */
+        if (UNLIKELY(inSeqs[i].matchLength > 65535+MINMATCH)) {
+            assert(longLen == 0);
+            longLen = i + 1;
+        }
+        if (UNLIKELY(inSeqs[i].litLength > 65535)) {
+            assert(longLen == 0);
+            longLen = i + nbSequences + 1;
+        }
+    }
+
+    return longLen;
+}
+
+/* the vector implementation could also be ported to SSSE3,
+ * but since this implementation is targeting modern systems (>= Sapphire Rapid),
+ * it's not useful to develop and maintain code for older pre-AVX2 platforms */
+
+#else /* no AVX2 */
+
+static size_t convertSequences_noRepcodes(
+    SeqDef* dstSeqs,
+    const ZSTD_Sequence* inSeqs,
+    size_t nbSequences)
+{
+    size_t longLen = 0;
+    size_t n;
+    for (n=0; n<nbSequences; n++) {
+        dstSeqs[n].offBase = OFFSET_TO_OFFBASE(inSeqs[n].offset);
+        dstSeqs[n].litLength = (U16)inSeqs[n].litLength;
+        dstSeqs[n].mlBase = (U16)(inSeqs[n].matchLength - MINMATCH);
+        /* check for long length > 65535 */
+        if (UNLIKELY(inSeqs[n].matchLength > 65535+MINMATCH)) {
+            assert(longLen == 0);
+            longLen = n + 1;
+        }
+        if (UNLIKELY(inSeqs[n].litLength > 65535)) {
+            assert(longLen == 0);
+            longLen = n + nbSequences + 1;
+        }
+    }
+    return longLen;
+}
+
+#endif
+
+/*
+ * Precondition: Sequences must end on an explicit Block Delimiter
+ * @return: 0 on success, or an error code.
+ * Note: Sequence validation functionality has been disabled (removed).
+ * This is helpful to generate a lean main pipeline, improving performance.
+ * It may be re-inserted later.
+ */
+size_t ZSTD_convertBlockSequences(ZSTD_CCtx* cctx,
+                const ZSTD_Sequence* const inSeqs, size_t nbSequences,
+                int repcodeResolution)
+{
+    Repcodes_t updatedRepcodes;
+    size_t seqNb = 0;
+
+    DEBUGLOG(5, "ZSTD_convertBlockSequences (nbSequences = %zu)", nbSequences);
+
+    RETURN_ERROR_IF(nbSequences >= cctx->seqStore.maxNbSeq, externalSequences_invalid,
+                    "Not enough memory allocated. Try adjusting ZSTD_c_minMatch.");
+
+    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(Repcodes_t));
+
+    /* check end condition */
+    assert(nbSequences >= 1);
+    assert(inSeqs[nbSequences-1].matchLength == 0);
+    assert(inSeqs[nbSequences-1].offset == 0);
+
+    /* Convert Sequences from public format to internal format */
+    if (!repcodeResolution) {
+        size_t const longl = convertSequences_noRepcodes(cctx->seqStore.sequencesStart, inSeqs, nbSequences-1);
+        cctx->seqStore.sequences = cctx->seqStore.sequencesStart + nbSequences-1;
+        if (longl) {
+            DEBUGLOG(5, "long length");
+            assert(cctx->seqStore.longLengthType == ZSTD_llt_none);
+            if (longl <= nbSequences-1) {
+                DEBUGLOG(5, "long match length detected at pos %zu", longl-1);
+                cctx->seqStore.longLengthType = ZSTD_llt_matchLength;
+                cctx->seqStore.longLengthPos = (U32)(longl-1);
+            } else {
+                DEBUGLOG(5, "long literals length detected at pos %zu", longl-nbSequences);
+                assert(longl <= 2* (nbSequences-1));
+                cctx->seqStore.longLengthType = ZSTD_llt_literalLength;
+                cctx->seqStore.longLengthPos = (U32)(longl-(nbSequences-1)-1);
+            }
+        }
+    } else {
+        for (seqNb = 0; seqNb < nbSequences - 1 ; seqNb++) {
+            U32 const litLength = inSeqs[seqNb].litLength;
+            U32 const matchLength = inSeqs[seqNb].matchLength;
+            U32 const ll0 = (litLength == 0);
+            U32 const offBase = ZSTD_finalizeOffBase(inSeqs[seqNb].offset, updatedRepcodes.rep, ll0);
+
+            DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offBase, matchLength, litLength);
+            ZSTD_storeSeqOnly(&cctx->seqStore, litLength, offBase, matchLength);
+            ZSTD_updateRep(updatedRepcodes.rep, offBase, ll0);
+        }
+    }
+
+    /* If we skipped repcode search while parsing, we need to update repcodes now */
+    if (!repcodeResolution && nbSequences > 1) {
+        U32* const rep = updatedRepcodes.rep;
+
+        if (nbSequences >= 4) {
+            U32 lastSeqIdx = (U32)nbSequences - 2; /* index of last full sequence */
+            rep[2] = inSeqs[lastSeqIdx - 2].offset;
+            rep[1] = inSeqs[lastSeqIdx - 1].offset;
+            rep[0] = inSeqs[lastSeqIdx].offset;
+        } else if (nbSequences == 3) {
+            rep[2] = rep[0];
+            rep[1] = inSeqs[0].offset;
+            rep[0] = inSeqs[1].offset;
+        } else {
+            assert(nbSequences == 2);
+            rep[2] = rep[1];
+            rep[1] = rep[0];
+            rep[0] = inSeqs[0].offset;
+        }
+    }
+
+    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(Repcodes_t));
+
+    return 0;
+}
+
+#if defined(ZSTD_ARCH_X86_AVX2)
+
+BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs)
+{
+    size_t i;
+    __m256i const zeroVec = _mm256_setzero_si256();
+    __m256i sumVec = zeroVec;  /* accumulates match+lit in 32-bit lanes */
+    ZSTD_ALIGNED(32) U32 tmp[8];      /* temporary buffer for reduction */
+    size_t mSum = 0, lSum = 0;
+    ZSTD_STATIC_ASSERT(sizeof(ZSTD_Sequence) == 16);
+
+    /* Process 2 structs (32 bytes) at a time */
+    for (i = 0; i + 2 <= nbSeqs; i += 2) {
+        /* Load two consecutive ZSTD_Sequence (8×4 = 32 bytes) */
+        __m256i data     = _mm256_loadu_si256((const __m256i*)(const void*)&seqs[i]);
+        /* check end of block signal */
+        __m256i cmp      = _mm256_cmpeq_epi32(data, zeroVec);
+        int cmp_res      = _mm256_movemask_epi8(cmp);
+        /* indices for match lengths correspond to bits [8..11], [24..27]
+         * => combined mask = 0x0F000F00 */
+        ZSTD_STATIC_ASSERT(offsetof(ZSTD_Sequence, matchLength) == 8);
+        if (cmp_res & 0x0F000F00) break;
+        /* Accumulate in sumVec */
+        sumVec           = _mm256_add_epi32(sumVec, data);
+    }
+
+    /* Horizontal reduction */
+    _mm256_store_si256((__m256i*)tmp, sumVec);
+    lSum = tmp[1] + tmp[5];
+    mSum = tmp[2] + tmp[6];
+
+    /* Handle the leftover */
+    for (; i < nbSeqs; i++) {
+        lSum += seqs[i].litLength;
+        mSum += seqs[i].matchLength;
+        if (seqs[i].matchLength == 0) break; /* end of block */
+    }
+
+    if (i==nbSeqs) {
+        /* reaching end of sequences: end of block signal was not present */
+        BlockSummary bs;
+        bs.nbSequences = ERROR(externalSequences_invalid);
+        return bs;
+    }
+    {   BlockSummary bs;
+        bs.nbSequences = i+1;
+        bs.blockSize = lSum + mSum;
+        bs.litSize = lSum;
+        return bs;
+    }
+}
+
+#else
+
+BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs)
+{
+    size_t totalMatchSize = 0;
+    size_t litSize = 0;
+    size_t n;
+    assert(seqs);
+    for (n=0; n<nbSeqs; n++) {
+        totalMatchSize += seqs[n].matchLength;
+        litSize += seqs[n].litLength;
+        if (seqs[n].matchLength == 0) {
+            assert(seqs[n].offset == 0);
+            break;
+        }
+    }
+    if (n==nbSeqs) {
+        BlockSummary bs;
+        bs.nbSequences = ERROR(externalSequences_invalid);
+        return bs;
+    }
+    {   BlockSummary bs;
+        bs.nbSequences = n+1;
+        bs.blockSize = litSize + totalMatchSize;
+        bs.litSize = litSize;
+        return bs;
+    }
+}
+#endif
+
+
+static size_t
+ZSTD_compressSequencesAndLiterals_internal(ZSTD_CCtx* cctx,
+                                void* dst, size_t dstCapacity,
+                          const ZSTD_Sequence* inSeqs, size_t nbSequences,
+                          const void* literals, size_t litSize, size_t srcSize)
+{
+    size_t remaining = srcSize;
+    size_t cSize = 0;
+    BYTE* op = (BYTE*)dst;
+    int const repcodeResolution = (cctx->appliedParams.searchForExternalRepcodes == ZSTD_ps_enable);
+    assert(cctx->appliedParams.searchForExternalRepcodes != ZSTD_ps_auto);
+
+    DEBUGLOG(4, "ZSTD_compressSequencesAndLiterals_internal: nbSeqs=%zu, litSize=%zu", nbSequences, litSize);
+    RETURN_ERROR_IF(nbSequences == 0, externalSequences_invalid, "Requires at least 1 end-of-block");
+
+    /* Special case: empty frame */
+    if ((nbSequences == 1) && (inSeqs[0].litLength == 0)) {
+        U32 const cBlockHeader24 = 1 /* last block */ + (((U32)bt_raw)<<1);
+        RETURN_ERROR_IF(dstCapacity<3, dstSize_tooSmall, "No room for empty frame block header");
+        MEM_writeLE24(op, cBlockHeader24);
+        op += ZSTD_blockHeaderSize;
+        dstCapacity -= ZSTD_blockHeaderSize;
+        cSize += ZSTD_blockHeaderSize;
+    }
+
+    while (nbSequences) {
+        size_t compressedSeqsSize, cBlockSize, conversionStatus;
+        BlockSummary const block = ZSTD_get1BlockSummary(inSeqs, nbSequences);
+        U32 const lastBlock = (block.nbSequences == nbSequences);
+        FORWARD_IF_ERROR(block.nbSequences, "Error while trying to determine nb of sequences for a block");
+        assert(block.nbSequences <= nbSequences);
+        RETURN_ERROR_IF(block.litSize > litSize, externalSequences_invalid, "discrepancy: Sequences require more literals than present in buffer");
+        ZSTD_resetSeqStore(&cctx->seqStore);
+
+        conversionStatus = ZSTD_convertBlockSequences(cctx,
+                            inSeqs, block.nbSequences,
+                            repcodeResolution);
+        FORWARD_IF_ERROR(conversionStatus, "Bad sequence conversion");
+        inSeqs += block.nbSequences;
+        nbSequences -= block.nbSequences;
+        remaining -= block.blockSize;
+
+        /* Note: when blockSize is very small, other variant send it uncompressed.
+         * Here, we still send the sequences, because we don't have the original source to send it uncompressed.
+         * One could imagine in theory reproducing the source from the sequences,
+         * but that's complex and costly memory intensive, and goes against the objectives of this variant. */
+
+        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "not enough dstCapacity to write a new compressed block");
+
+        compressedSeqsSize = ZSTD_entropyCompressSeqStore_internal(
+                                op + ZSTD_blockHeaderSize /* Leave space for block header */, dstCapacity - ZSTD_blockHeaderSize,
+                                literals, block.litSize,
+                                &cctx->seqStore,
+                                &cctx->blockState.prevCBlock->entropy, &cctx->blockState.nextCBlock->entropy,
+                                &cctx->appliedParams,
+                                cctx->tmpWorkspace, cctx->tmpWkspSize /* statically allocated in resetCCtx */,
+                                cctx->bmi2);
+        FORWARD_IF_ERROR(compressedSeqsSize, "Compressing sequences of block failed");
+        /* note: the spec forbids for any compressed block to be larger than maximum block size */
+        if (compressedSeqsSize > cctx->blockSizeMax) compressedSeqsSize = 0;
+        DEBUGLOG(5, "Compressed sequences size: %zu", compressedSeqsSize);
+        litSize -= block.litSize;
+        literals = (const char*)literals + block.litSize;
+
+        /* Note: difficult to check source for RLE block when only Literals are provided,
+         * but it could be considered from analyzing the sequence directly */
+
+        if (compressedSeqsSize == 0) {
+            /* Sending uncompressed blocks is out of reach, because the source is not provided.
+             * In theory, one could use the sequences to regenerate the source, like a decompressor,
+             * but it's complex, and memory hungry, killing the purpose of this variant.
+             * Current outcome: generate an error code.
+             */
+            RETURN_ERROR(cannotProduce_uncompressedBlock, "ZSTD_compressSequencesAndLiterals cannot generate an uncompressed block");
+        } else {
+            U32 cBlockHeader;
+            assert(compressedSeqsSize > 1); /* no RLE */
+            /* Error checking and repcodes update */
+            ZSTD_blockState_confirmRepcodesAndEntropyTables(&cctx->blockState);
+            if (cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
+                cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
+
+            /* Write block header into beginning of block*/
+            cBlockHeader = lastBlock + (((U32)bt_compressed)<<1) + (U32)(compressedSeqsSize << 3);
+            MEM_writeLE24(op, cBlockHeader);
+            cBlockSize = ZSTD_blockHeaderSize + compressedSeqsSize;
+            DEBUGLOG(5, "Writing out compressed block, size: %zu", cBlockSize);
+        }
+
+        cSize += cBlockSize;
+        op += cBlockSize;
+        dstCapacity -= cBlockSize;
+        cctx->isFirstBlock = 0;
+        DEBUGLOG(5, "cSize running total: %zu (remaining dstCapacity=%zu)", cSize, dstCapacity);
+
+        if (lastBlock) {
+            assert(nbSequences == 0);
+            break;
+        }
+    }
+
+    RETURN_ERROR_IF(litSize != 0, externalSequences_invalid, "literals must be entirely and exactly consumed");
+    RETURN_ERROR_IF(remaining != 0, externalSequences_invalid, "Sequences must represent a total of exactly srcSize=%zu", srcSize);
+    DEBUGLOG(4, "cSize final total: %zu", cSize);
+    return cSize;
+}
+
+size_t
+ZSTD_compressSequencesAndLiterals(ZSTD_CCtx* cctx,
+                    void* dst, size_t dstCapacity,
+                    const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
+                    const void* literals, size_t litSize, size_t litCapacity,
+                    size_t decompressedSize)
+{
+    BYTE* op = (BYTE*)dst;
+    size_t cSize = 0;
+
+    /* Transparent initialization stage, same as compressStream2() */
+    DEBUGLOG(4, "ZSTD_compressSequencesAndLiterals (dstCapacity=%zu)", dstCapacity);
+    assert(cctx != NULL);
+    if (litCapacity < litSize) {
+        RETURN_ERROR(workSpace_tooSmall, "literals buffer is not large enough: must be at least 8 bytes larger than litSize (risk of read out-of-bound)");
+    }
+    FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, ZSTD_e_end, decompressedSize), "CCtx initialization failed");
+
+    if (cctx->appliedParams.blockDelimiters == ZSTD_sf_noBlockDelimiters) {
+        RETURN_ERROR(frameParameter_unsupported, "This mode is only compatible with explicit delimiters");
+    }
+    if (cctx->appliedParams.validateSequences) {
+        RETURN_ERROR(parameter_unsupported, "This mode is not compatible with Sequence validation");
+    }
+    if (cctx->appliedParams.fParams.checksumFlag) {
+        RETURN_ERROR(frameParameter_unsupported, "this mode is not compatible with frame checksum");
+    }
+
+    /* Begin writing output, starting with frame header */
+    {   size_t const frameHeaderSize = ZSTD_writeFrameHeader(op, dstCapacity,
+                    &cctx->appliedParams, decompressedSize, cctx->dictID);
+        op += frameHeaderSize;
+        assert(frameHeaderSize <= dstCapacity);
+        dstCapacity -= frameHeaderSize;
+        cSize += frameHeaderSize;
+    }
+
+    /* Now generate compressed blocks */
+    {   size_t const cBlocksSize = ZSTD_compressSequencesAndLiterals_internal(cctx,
+                                            op, dstCapacity,
+                                            inSeqs, inSeqsSize,
+                                            literals, litSize, decompressedSize);
+        FORWARD_IF_ERROR(cBlocksSize, "Compressing blocks failed!");
+        cSize += cBlocksSize;
+        assert(cBlocksSize <= dstCapacity);
+        dstCapacity -= cBlocksSize;
+    }
+
+    DEBUGLOG(4, "Final compressed size: %zu", cSize);
     return cSize;
 }
 
 /*======   Finalize   ======*/
 
+static ZSTD_inBuffer inBuffer_forEndFlush(const ZSTD_CStream* zcs)
+{
+    const ZSTD_inBuffer nullInput = { NULL, 0, 0 };
+    const int stableInput = (zcs->appliedParams.inBufferMode == ZSTD_bm_stable);
+    return stableInput ? zcs->expectedInBuffer : nullInput;
+}
+
 /*! ZSTD_flushStream() :
  * @return : amount of data remaining to flush */
 size_t ZSTD_flushStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
 {
-    ZSTD_inBuffer input = { NULL, 0, 0 };
+    ZSTD_inBuffer input = inBuffer_forEndFlush(zcs);
+    input.size = input.pos; /* do not ingest more input during flush */
     return ZSTD_compressStream2(zcs, output, &input, ZSTD_e_flush);
 }
 
-
 size_t ZSTD_endStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
 {
-    ZSTD_inBuffer input = { NULL, 0, 0 };
+    ZSTD_inBuffer input = inBuffer_forEndFlush(zcs);
     size_t const remainingToFlush = ZSTD_compressStream2(zcs, output, &input, ZSTD_e_end);
-    FORWARD_IF_ERROR( remainingToFlush , "ZSTD_compressStream2 failed");
+    FORWARD_IF_ERROR(remainingToFlush , "ZSTD_compressStream2(,,ZSTD_e_end) failed");
     if (zcs->appliedParams.nbWorkers > 0) return remainingToFlush;   /* minimal estimation */
     /* single thread mode : attempt to calculate remaining to flush more precisely */
     {   size_t const lastBlockSize = zcs->frameEnded ? 0 : ZSTD_BLOCKHEADERSIZE;
···
     }
 }
 
-static U64 ZSTD_getCParamRowSize(U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode)
+static U64 ZSTD_getCParamRowSize(U64 srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode)
 {
     switch (mode) {
     case ZSTD_cpm_unknown:
···
  * @return ZSTD_compressionParameters structure for a selected compression level, srcSize and dictSize.
  *  Note: srcSizeHint 0 means 0, use ZSTD_CONTENTSIZE_UNKNOWN for unknown.
  *        Use dictSize == 0 for unknown or unused.
- *  Note: `mode` controls how we treat the `dictSize`. See docs for `ZSTD_cParamMode_e`. */
-static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode)
+ *  Note: `mode` controls how we treat the `dictSize`. See docs for `ZSTD_CParamMode_e`. */
+static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode)
 {
     U64 const rSize = ZSTD_getCParamRowSize(srcSizeHint, dictSize, mode);
     U32 const tableID = (rSize <= 256 KB) + (rSize <= 128 KB) + (rSize <= 16 KB);
···
             cp.targetLength = (unsigned)(-clampedCompressionLevel);
         }
         /* refine parameters based on srcSize & dictSize */
-        return ZSTD_adjustCParams_internal(cp, srcSizeHint, dictSize, mode);
+        return ZSTD_adjustCParams_internal(cp, srcSizeHint, dictSize, mode, ZSTD_ps_auto);
     }
 }
 
···
  *  same idea as ZSTD_getCParams()
  * @return a `ZSTD_parameters` structure (instead of `ZSTD_compressionParameters`).
  *  Fields of `ZSTD_frameParameters` are set to default values */
-static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode) {
+static ZSTD_parameters
+ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode)
+{
     ZSTD_parameters params;
     ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize, mode);
     DEBUGLOG(5, "ZSTD_getParams (cLevel=%i)", compressionLevel);
···
  *  same idea as ZSTD_getCParams()
  * @return a `ZSTD_parameters` structure (instead of `ZSTD_compressionParameters`).
  *  Fields of `ZSTD_frameParameters` are set to default values */
-ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize) {
+ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize)
+{
     if (srcSizeHint == 0) srcSizeHint = ZSTD_CONTENTSIZE_UNKNOWN;
     return ZSTD_getParams_internal(compressionLevel, srcSizeHint, dictSize, ZSTD_cpm_unknown);
+}
+
+void ZSTD_registerSequenceProducer(
+    ZSTD_CCtx* zc,
+    void* extSeqProdState,
+    ZSTD_sequenceProducer_F extSeqProdFunc)
+{
+    assert(zc != NULL);
+    ZSTD_CCtxParams_registerSequenceProducer(
+        &zc->requestedParams, extSeqProdState, extSeqProdFunc
+    );
+}
+
+void ZSTD_CCtxParams_registerSequenceProducer(
+  ZSTD_CCtx_params* params,
+  void* extSeqProdState,
+  ZSTD_sequenceProducer_F extSeqProdFunc)
+{
+    assert(params != NULL);
+    if (extSeqProdFunc != NULL) {
+        params->extSeqProdFunc = extSeqProdFunc;
+        params->extSeqProdState = extSeqProdState;
+    } else {
+        params->extSeqProdFunc = NULL;
+        params->extSeqProdState = NULL;
+    }
 }

lib/zstd/compress/zstd_compress_internal.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 ***************************************/
 #include "../common/zstd_internal.h"
 #include "zstd_cwksp.h"
-
+#include "../common/bits.h" /* ZSTD_highbit32, ZSTD_NbCommonBytes */
+#include "zstd_preSplit.h" /* ZSTD_SLIPBLOCK_WORKSPACESIZE */
 
 /*-*************************************
 *  Constants
···
                                        It's not a big deal though : candidate will just be sorted again.
                                        Additionally, candidate position 1 will be lost.
                                        But candidate 1 cannot hide a large tree of candidates, so it's a minimal loss.
-                                       The benefit is that ZSTD_DUBT_UNSORTED_MARK cannot be mishandled after table re-use with a different strategy.
+                                       The benefit is that ZSTD_DUBT_UNSORTED_MARK cannot be mishandled after table reuse with a different strategy.
                                        This constant is required by ZSTD_compressBlock_btlazy2() and ZSTD_reduceTable_internal() */
 
 
···
 } ZSTD_entropyCTables_t;
 
 /* *********************************************
+*  Sequences *
+***********************************************/
+typedef struct SeqDef_s {
+    U32 offBase;   /* offBase == Offset + ZSTD_REP_NUM, or repcode 1,2,3 */
+    U16 litLength;
+    U16 mlBase;    /* mlBase == matchLength - MINMATCH */
+} SeqDef;
+
+/* Controls whether seqStore has a single "long" litLength or matchLength. See SeqStore_t. */
+typedef enum {
+    ZSTD_llt_none = 0,             /* no longLengthType */
+    ZSTD_llt_literalLength = 1,    /* represents a long literal */
+    ZSTD_llt_matchLength = 2       /* represents a long match */
+} ZSTD_longLengthType_e;
+
+typedef struct {
+    SeqDef* sequencesStart;
+    SeqDef* sequences;      /* ptr to end of sequences */
+    BYTE*  litStart;
+    BYTE*  lit;             /* ptr to end of literals */
+    BYTE*  llCode;
+    BYTE*  mlCode;
+    BYTE*  ofCode;
+    size_t maxNbSeq;
+    size_t maxNbLit;
+
+    /* longLengthPos and longLengthType to allow us to represent either a single litLength or matchLength
+     * in the seqStore that has a value larger than U16 (if it exists). To do so, we increment
+     * the existing value of the litLength or matchLength by 0x10000.
+     */
+    ZSTD_longLengthType_e longLengthType;
+    U32                   longLengthPos;  /* Index of the sequence to apply long length modification to */
+} SeqStore_t;
+
+typedef struct {
+    U32 litLength;
+    U32 matchLength;
+} ZSTD_SequenceLength;
+
+/*
+ * Returns the ZSTD_SequenceLength for the given sequences. It handles the decoding of long sequences
+ * indicated by longLengthPos and longLengthType, and adds MINMATCH back to matchLength.
+ */
+MEM_STATIC ZSTD_SequenceLength ZSTD_getSequenceLength(SeqStore_t const* seqStore, SeqDef const* seq)
+{
+    ZSTD_SequenceLength seqLen;
+    seqLen.litLength = seq->litLength;
+    seqLen.matchLength = seq->mlBase + MINMATCH;
+    if (seqStore->longLengthPos == (U32)(seq - seqStore->sequencesStart)) {
+        if (seqStore->longLengthType == ZSTD_llt_literalLength) {
+            seqLen.litLength += 0x10000;
+        }
+        if (seqStore->longLengthType == ZSTD_llt_matchLength) {
+            seqLen.matchLength += 0x10000;
+        }
+    }
+    return seqLen;
+}
+
+const SeqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx);   /* compress & dictBuilder */
+int ZSTD_seqToCodes(const SeqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
+
+
+/* *********************************************
 *  Entropy buffer statistics structs and funcs *
 ***********************************************/
 /* ZSTD_hufCTablesMetadata_t :
···
  *  hufDesSize refers to the size of huffman tree description in bytes.
  *  This metadata is populated in ZSTD_buildBlockEntropyStats_literals() */
 typedef struct {
-    symbolEncodingType_e hType;
+    SymbolEncodingType_e hType;
     BYTE hufDesBuffer[ZSTD_MAX_HUF_HEADER_SIZE];
     size_t hufDesSize;
 } ZSTD_hufCTablesMetadata_t;
···
  *  fseTablesSize refers to the size of fse tables in bytes.
  *  This metadata is populated in ZSTD_buildBlockEntropyStats_sequences() */
 typedef struct {
-    symbolEncodingType_e llType;
-    symbolEncodingType_e ofType;
-    symbolEncodingType_e mlType;
+    SymbolEncodingType_e llType;
+    SymbolEncodingType_e ofType;
+    SymbolEncodingType_e mlType;
     BYTE fseTablesBuffer[ZSTD_MAX_FSE_HEADERS_SIZE];
     size_t fseTablesSize;
     size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */
···
 /* ZSTD_buildBlockEntropyStats() :
  *  Builds entropy for the block.
  *  @return : 0 on success or error code */
-size_t ZSTD_buildBlockEntropyStats(seqStore_t* seqStorePtr,
-                             const ZSTD_entropyCTables_t* prevEntropy,
-                                   ZSTD_entropyCTables_t* nextEntropy,
-                             const ZSTD_CCtx_params* cctxParams,
-                                   ZSTD_entropyCTablesMetadata_t* entropyMetadata,
-                                   void* workspace, size_t wkspSize);
+size_t ZSTD_buildBlockEntropyStats(
+                    const SeqStore_t* seqStorePtr,
+                    const ZSTD_entropyCTables_t* prevEntropy,
+                          ZSTD_entropyCTables_t* nextEntropy,
+                    const ZSTD_CCtx_params* cctxParams,
+                          ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                          void* workspace, size_t wkspSize);
 
 /* *******************************
 *  Compression internals structs *
···
                            stopped. posInSequence <= seq[pos].litLength + seq[pos].matchLength */
   size_t size;          /* The number of sequences. <= capacity. */
   size_t capacity;      /* The capacity starting from `seq` pointer */
-} rawSeqStore_t;
+} RawSeqStore_t;
 
-UNUSED_ATTR static const rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0, 0};
+UNUSED_ATTR static const RawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0, 0};
 
 typedef struct {
-    int price;
-    U32 off;
-    U32 mlen;
-    U32 litlen;
-    U32 rep[ZSTD_REP_NUM];
+    int price;  /* price from beginning of segment to this position */
+    U32 off;    /* offset of previous match */
+    U32 mlen;   /* length of previous match */
+    U32 litlen; /* nb of literals since previous match */
+    U32 rep[ZSTD_REP_NUM];  /* offset history after previous match */
 } ZSTD_optimal_t;
 
 typedef enum { zop_dynamic=0, zop_predef } ZSTD_OptPrice_e;
 
+#define ZSTD_OPT_SIZE (ZSTD_OPT_NUM+3)
 typedef struct {
     /* All tables are allocated inside cctx->workspace by ZSTD_resetCCtx_internal() */
     unsigned* litFreq;           /* table of literals statistics, of size 256 */
     unsigned* litLengthFreq;     /* table of litLength statistics, of size (MaxLL+1) */
     unsigned* matchLengthFreq;   /* table of matchLength statistics, of size (MaxML+1) */
     unsigned* offCodeFreq;       /* table of offCode statistics, of size (MaxOff+1) */
-    ZSTD_match_t* matchTable;    /* list of found matches, of size ZSTD_OPT_NUM+1 */
-    ZSTD_optimal_t* priceTable;  /* All positions tracked by optimal parser, of size ZSTD_OPT_NUM+1 */
+    ZSTD_match_t* matchTable;    /* list of found matches, of size ZSTD_OPT_SIZE */
+    ZSTD_optimal_t* priceTable;  /* All positions tracked by optimal parser, of size ZSTD_OPT_SIZE */
 
     U32  litSum;                 /* nb of literals */
     U32  litLengthSum;           /* nb of litLength codes */
···
     U32  offCodeSumBasePrice;    /* to compare to log2(offreq)  */
     ZSTD_OptPrice_e priceType;   /* prices can be determined dynamically, or follow a pre-defined cost structure */
     const ZSTD_entropyCTables_t* symbolCosts;  /* pre-calculated dictionary statistics */
-    ZSTD_paramSwitch_e literalCompressionMode;
+    ZSTD_ParamSwitch_e literalCompressionMode;
 } optState_t;
 
 typedef struct {
···
 
 #define ZSTD_WINDOW_START_INDEX 2
 
-typedef struct ZSTD_matchState_t ZSTD_matchState_t;
+typedef struct ZSTD_MatchState_t ZSTD_MatchState_t;
 
 #define ZSTD_ROW_HASH_CACHE_SIZE 8       /* Size of prefetching hash cache for row-based matchfinder */
 
-struct ZSTD_matchState_t {
+struct ZSTD_MatchState_t {
     ZSTD_window_t window;   /* State for window round buffer management */
     U32 loadedDictEnd;      /* index of end of dictionary, within context's referential.
                              * When loadedDictEnd != 0, a dictionary is in use, and still valid.
···
     U32 hashLog3;           /* dispatch table for matches of len==3 : larger == faster, more memory */
 
     U32 rowHashLog;                          /* For row-based matchfinder: Hashlog based on nb of rows in the hashTable.*/
-    U16* tagTable;                           /* For row-based matchFinder: A row-based table containing the hashes and head index. */
+    BYTE* tagTable;                          /* For row-based matchFinder: A row-based table containing the hashes and head index. */
     U32 hashCache[ZSTD_ROW_HASH_CACHE_SIZE]; /* For row-based matchFinder: a cache of hashes to improve speed */
+    U64 hashSalt;                            /* For row-based matchFinder: salts the hash for reuse of tag table */
+    U32 hashSaltEntropy;                     /* For row-based matchFinder: collects entropy for salt generation */
 
     U32* hashTable;
     U32* hashTable3;
     U32* chainTable;
 
-    U32 forceNonContiguous; /* Non-zero if we should force non-contiguous load for the next window update. */
+    int forceNonContiguous; /* Non-zero if we should force non-contiguous load for the next window update. */
 
     int dedicatedDictSearch;  /* Indicates whether this matchState is using the
                                * dedicated dictionary search structure.
                                */
     optState_t opt;         /* optimal parser state */
-    const ZSTD_matchState_t* dictMatchState;
+    const ZSTD_MatchState_t* dictMatchState;
     ZSTD_compressionParameters cParams;
-    const rawSeqStore_t* ldmSeqStore;
+    const RawSeqStore_t* ldmSeqStore;
+
+    /* Controls prefetching in some dictMatchState matchfinders.
+     * This behavior is controlled from the cctx ms.
+     * This parameter has no effect in the cdict ms. */
+    int prefetchCDictTables;
+
+    /* When == 0, lazy match finders insert every position.
+     * When != 0, lazy match finders only insert positions they search.
+     * This allows them to skip much faster over incompressible data,
+     * at a small cost to compression ratio.
+     */
+    int lazySkipping;
 };
 
 typedef struct {
     ZSTD_compressedBlockState_t* prevCBlock;
     ZSTD_compressedBlockState_t* nextCBlock;
-    ZSTD_matchState_t matchState;
+    ZSTD_MatchState_t matchState;
 } ZSTD_blockState_t;
 
 typedef struct {
···
 } ldmState_t;
 
 typedef struct {
-    ZSTD_paramSwitch_e enableLdm; /* ZSTD_ps_enable to enable LDM. ZSTD_ps_auto by default */
+    ZSTD_ParamSwitch_e enableLdm; /* ZSTD_ps_enable to enable LDM. ZSTD_ps_auto by default */
     U32 hashLog;            /* Log size of hashTable */
     U32 bucketSizeLog;      /* Log bucket size for collision resolution, at most 8 */
     U32 minMatchLength;     /* Minimum match length */
···
                                 * There is no guarantee that hint is close to actual source size */
 
     ZSTD_dictAttachPref_e attachDictPref;
-    ZSTD_paramSwitch_e literalCompressionMode;
+    ZSTD_ParamSwitch_e literalCompressionMode;
 
     /* Multithreading: used to pass parameters to mtctx */
     int nbWorkers;
···
     ZSTD_bufferMode_e outBufferMode;
 
     /* Sequence compression API */
-    ZSTD_sequenceFormat_e blockDelimiters;
+    ZSTD_SequenceFormat_e blockDelimiters;
     int validateSequences;
 
-    /* Block splitting */
-    ZSTD_paramSwitch_e useBlockSplitter;
+    /* Block splitting
+     * @postBlockSplitter executes split analysis after sequences are produced,
+     * it's more accurate but consumes more resources.
+     * @preBlockSplitter_level splits before knowing sequences,
+     * it's more approximative but also cheaper.
+     * Valid @preBlockSplitter_level values range from 0 to 6 (included).
+     * 0 means auto, 1 means do not split,
+     * then levels are sorted in increasing cpu budget, from 2 (fastest) to 6 (slowest).
+     * Highest @preBlockSplitter_level combines well with @postBlockSplitter.
+     */
+    ZSTD_ParamSwitch_e postBlockSplitter;
+    int preBlockSplitter_level;
+
+    /* Adjust the max block size*/
+    size_t maxBlockSize;
 
     /* Param for deciding whether to use row-based matchfinder */
-    ZSTD_paramSwitch_e useRowMatchFinder;
+    ZSTD_ParamSwitch_e useRowMatchFinder;
 
     /* Always load a dictionary in ext-dict mode (not prefix mode)? */
     int deterministicRefPrefix;
 
     /* Internal use, for createCCtxParams() and freeCCtxParams() only */
     ZSTD_customMem customMem;
+
+    /* Controls prefetching in some dictMatchState matchfinders */
+    ZSTD_ParamSwitch_e prefetchCDictTables;
+
+    /* Controls whether zstd will fall back to an internal matchfinder
+     * if the external matchfinder returns an error code. */
+    int enableMatchFinderFallback;
+
+    /* Parameters for the external sequence producer API.
+     * Users set these parameters through ZSTD_registerSequenceProducer().
+     * It is not possible to set these parameters individually through the public API. */
+    void* extSeqProdState;
+    ZSTD_sequenceProducer_F extSeqProdFunc;
+
+    /* Controls repcode search in external sequence parsing */
+    ZSTD_ParamSwitch_e searchForExternalRepcodes;
 };  /* typedef'd to ZSTD_CCtx_params within "zstd.h" */
 
 #define COMPRESS_SEQUENCES_WORKSPACE_SIZE (sizeof(unsigned) * (MaxSeq + 2))
 #define ENTROPY_WORKSPACE_SIZE (HUF_WORKSPACE_SIZE + COMPRESS_SEQUENCES_WORKSPACE_SIZE)
+#define TMP_WORKSPACE_SIZE (MAX(ENTROPY_WORKSPACE_SIZE, ZSTD_SLIPBLOCK_WORKSPACESIZE))
 
 /*
  * Indicates whether this compression proceeds directly from user-provided
···
  */
 #define ZSTD_MAX_NB_BLOCK_SPLITS 196
 typedef struct {
-    seqStore_t fullSeqStoreChunk;
-    seqStore_t firstHalfSeqStore;
-    seqStore_t secondHalfSeqStore;
-    seqStore_t currSeqStore;
-    seqStore_t nextSeqStore;
+    SeqStore_t fullSeqStoreChunk;
+    SeqStore_t firstHalfSeqStore;
+    SeqStore_t secondHalfSeqStore;
+    SeqStore_t currSeqStore;
+    SeqStore_t nextSeqStore;
 
     U32 partitions[ZSTD_MAX_NB_BLOCK_SPLITS];
     ZSTD_entropyCTablesMetadata_t entropyMetadata;
···
     size_t dictContentSize;
 
     ZSTD_cwksp workspace; /* manages buffer for dynamic allocations */
-    size_t blockSize;
+    size_t blockSizeMax;
     unsigned long long pledgedSrcSizePlusOne;  /* this way, 0 (default) == unknown */
     unsigned long long consumedSrcSize;
     unsigned long long producedCSize;
···
     int isFirstBlock;
     int initialized;
 
-    seqStore_t seqStore;      /* sequences storage ptrs */
+    SeqStore_t seqStore;      /* sequences storage ptrs */
     ldmState_t ldmState;      /* long distance matching state */
     rawSeq* ldmSequences;     /* Storage for the ldm output sequences */
     size_t maxNbLdmSequences;
-    rawSeqStore_t externSeqStore; /* Mutable reference to external sequences */
+    RawSeqStore_t externSeqStore; /* Mutable reference to external sequences */
     ZSTD_blockState_t blockState;
-    U32* entropyWorkspace;  /* entropy workspace of ENTROPY_WORKSPACE_SIZE bytes */
+    void* tmpWorkspace;  /* used as substitute of stack space - must be aligned for S64 type */
+    size_t tmpWkspSize;
 
     /* Whether we are streaming or not */
     ZSTD_buffered_policy_e bufferedPolicy;
···
 
     /* Stable in/out buffer verification */
     ZSTD_inBuffer expectedInBuffer;
+    size_t stableIn_notConsumed; /* nb bytes within stable input buffer that are said to be consumed but are not */
     size_t expectedOutBufferSize;
 
     /* Dictionary */
···
 
     /* Workspace for block splitter */
     ZSTD_blockSplitCtx blockSplitCtx;
+
+    /* Buffer for output from external sequence producer */
+    ZSTD_Sequence* extSeqBuf;
+    size_t extSeqBufCapacity;
 };
 
 typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
+typedef enum { ZSTD_tfp_forCCtx, ZSTD_tfp_forCDict } ZSTD_tableFillPurpose_e;
 
 typedef enum {
     ZSTD_noDict = 0,
···
                                  * In this mode we take both the source size and the dictionary size
                                  * into account when selecting and adjusting the parameters.
                                  */
-    ZSTD_cpm_unknown = 3,       /* ZSTD_getCParams, ZSTD_getParams, ZSTD_adjustParams.
+    ZSTD_cpm_unknown = 3        /* ZSTD_getCParams, ZSTD_getParams, ZSTD_adjustParams.
                                  * We don't know what these parameters are for. We default to the legacy
                                  * behavior of taking both the source size and the dict size into account
                                  * when selecting and adjusting parameters.
                                  */
-} ZSTD_cParamMode_e;
+} ZSTD_CParamMode_e;
 
-typedef size_t (*ZSTD_blockCompressor) (
-        ZSTD_matchState_t* bs, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+typedef size_t (*ZSTD_BlockCompressor_f) (
+        ZSTD_MatchState_t* bs, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
-ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_paramSwitch_e rowMatchfinderMode, ZSTD_dictMode_e dictMode);
+ZSTD_BlockCompressor_f ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_ParamSwitch_e rowMatchfinderMode, ZSTD_dictMode_e dictMode);
 
 
 MEM_STATIC U32 ZSTD_LLcode(U32 litLength)
···
     return 1;
 }
 
+/* ZSTD_selectAddr:
+ * @return index >= lowLimit ? candidate : backup,
+ * tries to force branchless codegen. */
+MEM_STATIC const BYTE*
+ZSTD_selectAddr(U32 index, U32 lowLimit, const BYTE* candidate, const BYTE* backup)
+{
+#if defined(__x86_64__)
+    __asm__ (
+        "cmp %1, %2\n"
+        "cmova %3, %0\n"
+        : "+r"(candidate)
+        : "r"(index), "r"(lowLimit), "r"(backup)
+        );
+    return candidate;
+#else
+    return index >= lowLimit ? candidate : backup;
+#endif
+}
+
 /* ZSTD_noCompressBlock() :
  * Writes uncompressed block to dst buffer from given src.
  * Returns the size of the block */
-MEM_STATIC size_t ZSTD_noCompressBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
+MEM_STATIC size_t
+ZSTD_noCompressBlock(void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
 {
     U32 const cBlockHeader24 = lastBlock + (((U32)bt_raw)<<1) + (U32)(srcSize << 3);
+    DEBUGLOG(5, "ZSTD_noCompressBlock (srcSize=%zu, dstCapacity=%zu)", srcSize, dstCapacity);
     RETURN_ERROR_IF(srcSize + ZSTD_blockHeaderSize > dstCapacity,
                     dstSize_tooSmall, "dst buf too small for uncompressed block");
     MEM_writeLE24(dst, cBlockHeader24);
···
     return ZSTD_blockHeaderSize + srcSize;
 }
 
-MEM_STATIC size_t ZSTD_rleCompressBlock (void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
+MEM_STATIC size_t
+ZSTD_rleCompressBlock(void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
 {
     BYTE* const op = (BYTE*)dst;
     U32 const cBlockHeader = lastBlock + (((U32)bt_rle)<<1) + (U32)(srcSize << 3);
···
 {
     U32 const minlog = (strat>=ZSTD_btultra) ? (U32)(strat) - 1 : 6;
     ZSTD_STATIC_ASSERT(ZSTD_btultra == 8);
-    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat));
+    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, (int)strat));
     return (srcSize >> minlog) + 2;
 }
 
···
     while (ip < iend) *op++ = *ip++;
 }
 
-#define ZSTD_REP_MOVE     (ZSTD_REP_NUM-1)
-#define STORE_REPCODE_1 STORE_REPCODE(1)
-#define STORE_REPCODE_2 STORE_REPCODE(2)
-#define STORE_REPCODE_3 STORE_REPCODE(3)
-#define STORE_REPCODE(r) (assert((r)>=1), assert((r)<=3), (r)-1)
-#define STORE_OFFSET(o)  (assert((o)>0), o + ZSTD_REP_MOVE)
-#define STORED_IS_OFFSET(o)  ((o) > ZSTD_REP_MOVE)
-#define STORED_IS_REPCODE(o) ((o) <= ZSTD_REP_MOVE)
-#define STORED_OFFSET(o)  (assert(STORED_IS_OFFSET(o)), (o)-ZSTD_REP_MOVE)
-#define STORED_REPCODE(o) (assert(STORED_IS_REPCODE(o)), (o)+1)  /* returns ID 1,2,3 */
-#define STORED_TO_OFFBASE(o) ((o)+1)
-#define OFFBASE_TO_STORED(o) ((o)-1)
 
-/*! ZSTD_storeSeq() :
- *  Store a sequence (litlen, litPtr, offCode and matchLength) into seqStore_t.
- *  @offBase_minus1 : Users should use employ macros STORE_REPCODE_X and STORE_OFFSET().
+#define REPCODE1_TO_OFFBASE REPCODE_TO_OFFBASE(1)
+#define REPCODE2_TO_OFFBASE REPCODE_TO_OFFBASE(2)
+#define REPCODE3_TO_OFFBASE REPCODE_TO_OFFBASE(3)
+#define REPCODE_TO_OFFBASE(r) (assert((r)>=1), assert((r)<=ZSTD_REP_NUM), (r)) /* accepts IDs 1,2,3 */
+#define OFFSET_TO_OFFBASE(o)  (assert((o)>0), o + ZSTD_REP_NUM)
+#define OFFBASE_IS_OFFSET(o)  ((o) > ZSTD_REP_NUM)
+#define OFFBASE_IS_REPCODE(o) ( 1 <= (o) && (o) <= ZSTD_REP_NUM)
+#define OFFBASE_TO_OFFSET(o)  (assert(OFFBASE_IS_OFFSET(o)), (o) - ZSTD_REP_NUM)
+#define OFFBASE_TO_REPCODE(o) (assert(OFFBASE_IS_REPCODE(o)), (o))  /* returns ID 1,2,3 */
+
+/*! ZSTD_storeSeqOnly() :
+ *  Store a sequence (litlen, litPtr, offBase and matchLength) into SeqStore_t.
+ *  Literals themselves are not copied, but @litPtr is updated.
+ *  @offBase : Users should employ macros REPCODE_TO_OFFBASE() and OFFSET_TO_OFFBASE().
  *  @matchLength : must be >= MINMATCH
- *  Allowed to overread literals up to litLimit.
 */
 HINT_INLINE UNUSED_ATTR void
-ZSTD_storeSeq(seqStore_t* seqStorePtr,
-              size_t litLength, const BYTE* literals, const BYTE* litLimit,
-              U32 offBase_minus1,
+ZSTD_storeSeqOnly(SeqStore_t* seqStorePtr,
+              size_t litLength,
+              U32 offBase,
               size_t matchLength)
 {
-    BYTE const* const litLimit_w = litLimit - WILDCOPY_OVERLENGTH;
-    BYTE const* const litEnd = literals + litLength;
-#if defined(DEBUGLEVEL) && (DEBUGLEVEL >= 6)
-    static const BYTE* g_start = NULL;
-    if (g_start==NULL) g_start = (const BYTE*)literals;  /* note : index only works for compression within a single segment */
-    {   U32 const pos = (U32)((const BYTE*)literals - g_start);
-        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offCode%7u",
-               pos, (U32)litLength, (U32)matchLength, (U32)offBase_minus1);
-    }
-#endif
     assert((size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart) < seqStorePtr->maxNbSeq);
-    /* copy Literals */
-    assert(seqStorePtr->maxNbLit <= 128 KB);
-    assert(seqStorePtr->lit + litLength <= seqStorePtr->litStart + seqStorePtr->maxNbLit);
-    assert(literals + litLength <= litLimit);
-    if (litEnd <= litLimit_w) {
-        /* Common case we can use wildcopy.
-	 * First copy 16 bytes, because literals are likely short.
-	 */
-        assert(WILDCOPY_OVERLENGTH >= 16);
-        ZSTD_copy16(seqStorePtr->lit, literals);
-        if (litLength > 16) {
-            ZSTD_wildcopy(seqStorePtr->lit+16, literals+16, (ptrdiff_t)litLength-16, ZSTD_no_overlap);
-        }
-    } else {
-        ZSTD_safecopyLiterals(seqStorePtr->lit, literals, litEnd, litLimit_w);
-    }
-    seqStorePtr->lit += litLength;
 
     /* literal Length */
-    if (litLength>0xFFFF) {
+    assert(litLength <= ZSTD_BLOCKSIZE_MAX);
+    if (UNLIKELY(litLength>0xFFFF)) {
         assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */
         seqStorePtr->longLengthType = ZSTD_llt_literalLength;
         seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
···
     seqStorePtr->sequences[0].litLength = (U16)litLength;
 
     /* match offset */
-    seqStorePtr->sequences[0].offBase = STORED_TO_OFFBASE(offBase_minus1);
+    seqStorePtr->sequences[0].offBase = offBase;
 
     /* match Length */
+    assert(matchLength <= ZSTD_BLOCKSIZE_MAX);
     assert(matchLength >= MINMATCH);
     {   size_t const mlBase = matchLength - MINMATCH;
-        if (mlBase>0xFFFF) {
+        if (UNLIKELY(mlBase>0xFFFF)) {
             assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */
             seqStorePtr->longLengthType = ZSTD_llt_matchLength;
             seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
···
     seqStorePtr->sequences++;
 }
 
+/*! ZSTD_storeSeq() :
+ *  Store a sequence (litlen, litPtr, offBase and matchLength) into SeqStore_t.
+ *  @offBase : Users should employ macros REPCODE_TO_OFFBASE() and OFFSET_TO_OFFBASE().
+ *  @matchLength : must be >= MINMATCH
+ *  Allowed to over-read literals up to litLimit.
+*/
+HINT_INLINE UNUSED_ATTR void
+ZSTD_storeSeq(SeqStore_t* seqStorePtr,
+              size_t litLength, const BYTE* literals, const BYTE* litLimit,
+              U32 offBase,
+              size_t matchLength)
+{
+    BYTE const* const litLimit_w = litLimit - WILDCOPY_OVERLENGTH;
+    BYTE const* const litEnd = literals + litLength;
+#if defined(DEBUGLEVEL) && (DEBUGLEVEL >= 6)
+    static const BYTE* g_start = NULL;
+    if (g_start==NULL) g_start = (const BYTE*)literals;  /* note : index only works for compression within a single segment */
+    {   U32 const pos = (U32)((const BYTE*)literals - g_start);
+        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offBase%7u",
+               pos, (U32)litLength, (U32)matchLength, (U32)offBase);
+    }
+#endif
+    assert((size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart) < seqStorePtr->maxNbSeq);
+    /* copy Literals */
+    assert(seqStorePtr->maxNbLit <= 128 KB);
+    assert(seqStorePtr->lit + litLength <= seqStorePtr->litStart + seqStorePtr->maxNbLit);
+    assert(literals + litLength <= litLimit);
+    if (litEnd <= litLimit_w) {
+        /* Common case we can use wildcopy.
+         * First copy 16 bytes, because literals are likely short.
+         */
+        ZSTD_STATIC_ASSERT(WILDCOPY_OVERLENGTH >= 16);
+        ZSTD_copy16(seqStorePtr->lit, literals);
+        if (litLength > 16) {
+            ZSTD_wildcopy(seqStorePtr->lit+16, literals+16, (ptrdiff_t)litLength-16, ZSTD_no_overlap);
+        }
+    } else {
+        ZSTD_safecopyLiterals(seqStorePtr->lit, literals, litEnd, litLimit_w);
+    }
+    seqStorePtr->lit += litLength;
+
+    ZSTD_storeSeqOnly(seqStorePtr, litLength, offBase, matchLength);
+}
+
 /* ZSTD_updateRep() :
  * updates in-place @rep (array of repeat offsets)
- * @offBase_minus1 : sum-type, with same numeric representation as ZSTD_storeSeq()
+ * @offBase : sum-type, using numeric representation of ZSTD_storeSeq()
  */
 MEM_STATIC void
-ZSTD_updateRep(U32 rep[ZSTD_REP_NUM], U32 const offBase_minus1, U32 const ll0)
+ZSTD_updateRep(U32 rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
 {
-    if (STORED_IS_OFFSET(offBase_minus1)) {  /* full offset */
+    if (OFFBASE_IS_OFFSET(offBase)) {  /* full offset */
         rep[2] = rep[1];
         rep[1] = rep[0];
-        rep[0] = STORED_OFFSET(offBase_minus1);
+        rep[0] = OFFBASE_TO_OFFSET(offBase);
     } else {   /* repcode */
-        U32 const repCode = STORED_REPCODE(offBase_minus1) - 1 + ll0;
+        U32 const repCode = OFFBASE_TO_REPCODE(offBase) - 1 + ll0;
         if (repCode > 0) {  /* note : if repCode==0, no change */
             U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
             rep[2] = (repCode >= 2) ? rep[1] : rep[2];
···
 
 typedef struct repcodes_s {
     U32 rep[3];
-} repcodes_t;
+} Repcodes_t;
 
-MEM_STATIC repcodes_t
-ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase_minus1, U32 const ll0)
+MEM_STATIC Repcodes_t
+ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
 {
-    repcodes_t newReps;
+    Repcodes_t newReps;
     ZSTD_memcpy(&newReps, rep, sizeof(newReps));
-    ZSTD_updateRep(newReps.rep, offBase_minus1, ll0);
+    ZSTD_updateRep(newReps.rep, offBase, ll0);
     return newReps;
 }
 
···
 /*-*************************************
 *  Match length counter
 ***************************************/
-static unsigned ZSTD_NbCommonBytes (size_t val)
-{
-    if (MEM_isLittleEndian()) {
-        if (MEM_64bits()) {
-#       if (__GNUC__ >= 4)
-            return (__builtin_ctzll((U64)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2,
-                                                     0, 3, 1, 3, 1, 4, 2, 7,
-                                                     0, 2, 3, 6, 1, 5, 3, 5,
-                                                     1, 3, 4, 4, 2, 5, 6, 7,
-                                                     7, 0, 1, 2, 3, 3, 4, 6,
-                                                     2, 6, 5, 5, 3, 4, 5, 6,
-                                                     7, 1, 2, 4, 6, 4, 4, 5,
-                                                     7, 2, 6, 5, 7, 6, 7, 7 };
-            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
-#       endif
-        } else { /* 32 bits */
-#       if (__GNUC__ >= 3)
-            return (__builtin_ctz((U32)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0,
-                                                     3, 2, 2, 1, 3, 2, 0, 1,
-                                                     3, 3, 1, 2, 2, 2, 2, 0,
-                                                     3, 1, 2, 0, 1, 0, 1, 1 };
-            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
-#       endif
-        }
-    } else {  /* Big Endian CPU */
-        if (MEM_64bits()) {
-#       if (__GNUC__ >= 4)
-            return (__builtin_clzll(val) >> 3);
-#       else
-            unsigned r;
-            const unsigned n32 = sizeof(size_t)*4;   /* calculate this way due to compiler complaining in 32-bits mode */
-            if (!(val>>n32)) { r=4; } else { r=0; val>>=n32; }
-            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-        } else { /* 32 bits */
-#       if (__GNUC__ >= 3)
-            return (__builtin_clz((U32)val) >> 3);
-#       else
-            unsigned r;
-            if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-    }   }
-}
-
-
 MEM_STATIC size_t ZSTD_count(const BYTE* pIn, const BYTE* pMatch, const BYTE* const pInLimit)
 {
     const BYTE* const pStart = pIn;
···
     size_t const matchLength = ZSTD_count(ip, match, vEnd);
     if (match + matchLength != mEnd) return matchLength;
     DEBUGLOG(7, "ZSTD_count_2segments: found a 2-parts match (current length==%zu)", matchLength);
-    DEBUGLOG(7, "distance from match beginning to end dictionary = %zi", mEnd - match);
-    DEBUGLOG(7, "distance from current pos to end buffer = %zi", iEnd - ip);
+    DEBUGLOG(7, "distance from match beginning to end dictionary = %i", (int)(mEnd - match));
+    DEBUGLOG(7, "distance from current pos to end buffer = %i", (int)(iEnd - ip));
     DEBUGLOG(7, "next byte : ip==%02X, istart==%02X", ip[matchLength], *iStart);
     DEBUGLOG(7, "final match length = %zu", matchLength + ZSTD_count(ip+matchLength, iStart, iEnd));
     return matchLength + ZSTD_count(ip+matchLength, iStart, iEnd);
···
  *  Hashes
  ***************************************/
 static const U32 prime3bytes = 506832829U;
-static U32    ZSTD_hash3(U32 u, U32 h) { return ((u << (32-24)) * prime3bytes)  >> (32-h) ; }
-MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h); } /* only in zstd_opt.h */
+static U32    ZSTD_hash3(U32 u, U32 h, U32 s) { assert(h <= 32); return (((u << (32-24)) * prime3bytes) ^ s)  >> (32-h) ; }
+MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h, 0); } /* only in zstd_opt.h */
+MEM_STATIC size_t ZSTD_hash3PtrS(const void* ptr, U32 h, U32 s) { return ZSTD_hash3(MEM_readLE32(ptr), h, s); }
 
 static const U32 prime4bytes = 2654435761U;
-static U32    ZSTD_hash4(U32 u, U32 h) { return (u * prime4bytes) >> (32-h) ; }
-static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_read32(ptr), h); }
+static U32    ZSTD_hash4(U32 u, U32 h, U32 s) { assert(h <= 32); return ((u * prime4bytes) ^ s) >> (32-h) ; }
+static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_readLE32(ptr), h, 0); }
+static size_t ZSTD_hash4PtrS(const void* ptr, U32 h, U32 s) { return ZSTD_hash4(MEM_readLE32(ptr), h, s); }
 
 static const U64 prime5bytes = 889523592379ULL;
-static size_t ZSTD_hash5(U64 u, U32 h) { return (size_t)(((u  << (64-40)) * prime5bytes) >> (64-h)) ; }
-static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h); }
+static size_t ZSTD_hash5(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-40)) * prime5bytes) ^ s) >> (64-h)) ; }
+static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h, 0); }
+static size_t ZSTD_hash5PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash5(MEM_readLE64(p), h, s); }
 
 static const U64 prime6bytes = 227718039650203ULL;
-static size_t ZSTD_hash6(U64 u, U32 h) { return (size_t)(((u  << (64-48)) * prime6bytes) >> (64-h)) ; }
-static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h); }
+static size_t ZSTD_hash6(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-48)) * prime6bytes) ^ s) >> (64-h)) ; }
+static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h, 0); }
+static size_t ZSTD_hash6PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash6(MEM_readLE64(p), h, s); }
 
 static const U64 prime7bytes = 58295818150454627ULL;
-static size_t ZSTD_hash7(U64 u, U32 h) { return (size_t)(((u  << (64-56)) * prime7bytes) >> (64-h)) ; }
-static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h); }
+static size_t ZSTD_hash7(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-56)) * prime7bytes) ^ s) >> (64-h)) ; }
+static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h, 0); }
+static size_t ZSTD_hash7PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash7(MEM_readLE64(p), h, s); }
 
 static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL;
-static size_t ZSTD_hash8(U64 u, U32 h) { return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
-static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h); }
+static size_t ZSTD_hash8(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u) * prime8bytes)  ^ s) >> (64-h)) ; }
+static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h, 0); }
+static size_t ZSTD_hash8PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash8(MEM_readLE64(p), h, s); }
+
 
 MEM_STATIC FORCE_INLINE_ATTR
 size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
 {
+    /* Although some of these hashes do support hBits up to 64, some do not.
+     * To be on the safe side, always avoid hBits > 32. */
+    assert(hBits <= 32);
+
     switch(mls)
     {
     default:
···
     case 8: return ZSTD_hash8Ptr(p, hBits);
     }
 }
+
+MEM_STATIC FORCE_INLINE_ATTR
+size_t ZSTD_hashPtrSalted(const void* p, U32 hBits, U32 mls, const U64 hashSalt) {
+    /* Although some of these hashes do support hBits up to 64, some do not.
+     * To be on the safe side, always avoid hBits > 32. */
+    assert(hBits <= 32);
+
+    switch(mls)
+    {
+        default:
+        case 4: return ZSTD_hash4PtrS(p, hBits, (U32)hashSalt);
+        case 5: return ZSTD_hash5PtrS(p, hBits, hashSalt);
+        case 6: return ZSTD_hash6PtrS(p, hBits, hashSalt);
+        case 7: return ZSTD_hash7PtrS(p, hBits, hashSalt);
+        case 8: return ZSTD_hash8PtrS(p, hBits, hashSalt);
+    }
+}
+
 
 /* ZSTD_ipow() :
  * Return base^exponent.
···
 /*-*************************************
 *  Round buffer management
 ***************************************/
-#if (ZSTD_WINDOWLOG_MAX_64 > 31)
-# error "ZSTD_WINDOWLOG_MAX is too large : would overflow ZSTD_CURRENT_MAX"
-#endif
-/* Max current allowed */
-#define ZSTD_CURRENT_MAX ((3U << 29) + (1U << ZSTD_WINDOWLOG_MAX))
+/* Max @current value allowed:
+ * In 32-bit mode: we want to avoid crossing the 2 GB limit,
+ *                 reducing risks of side effects in case of signed operations on indexes.
+ * In 64-bit mode: we want to ensure that adding the maximum job size (512 MB)
+ *                 doesn't overflow U32 index capacity (4 GB) */
+#define ZSTD_CURRENT_MAX (MEM_64bits() ? 3500U MB : 2000U MB)
 /* Maximum chunk size before overflow correction needs to be called again */
 #define ZSTD_CHUNKSIZE_MAX                                                     \
     ( ((U32)-1)                  /* Maximum ending current index */            \
···
  * Inspects the provided matchState and figures out what dictMode should be
  * passed to the compressor.
  */
-MEM_STATIC ZSTD_dictMode_e ZSTD_matchState_dictMode(const ZSTD_matchState_t *ms)
+MEM_STATIC ZSTD_dictMode_e ZSTD_matchState_dictMode(const ZSTD_MatchState_t *ms)
 {
     return ZSTD_window_hasExtDict(ms->window) ?
         ZSTD_extDict :
···
  * The least significant cycleLog bits of the indices must remain the same,
  * which may be 0. Every index up to maxDist in the past must be valid.
  */
-MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
                                            U32 maxDist, void const* src)
 {
     /* preemptive overflow correction:
···
                      const void* blockEnd,
                            U32   maxDist,
                            U32*  loadedDictEndPtr,
-                     const ZSTD_matchState_t** dictMatchStatePtr)
+                     const ZSTD_MatchState_t** dictMatchStatePtr)
 {
     U32 const blockEndIdx = (U32)((BYTE const*)blockEnd - window->base);
     U32 const loadedDictEnd = (loadedDictEndPtr != NULL) ? *loadedDictEndPtr : 0;
···
                        const void* blockEnd,
                              U32   maxDist,
                              U32*  loadedDictEndPtr,
-                       const ZSTD_matchState_t** dictMatchStatePtr)
+                       const ZSTD_MatchState_t** dictMatchStatePtr)
 {
     assert(loadedDictEndPtr != NULL);
     assert(dictMatchStatePtr != NULL);
···
                     (unsigned)blockEndIdx, (unsigned)maxDist, (unsigned)loadedDictEnd);
         assert(blockEndIdx >= loadedDictEnd);
 
-        if (blockEndIdx > loadedDictEnd + maxDist) {
+        if (blockEndIdx > loadedDictEnd + maxDist || loadedDictEnd != window->dictLimit) {
             /* On reaching window size, dictionaries are invalidated.
              * For simplification, if window size is reached anywhere within next block,
              * the dictionary is invalidated for the full block.
+             *
+             * We also have to invalidate the dictionary if ZSTD_window_update() has detected
+             * non-contiguous segments, which means that loadedDictEnd != window->dictLimit.
+             * loadedDictEnd may be 0, if forceWindow is true, but in that case we never use
+             * dictMatchState, so setting it to NULL is not a problem.
              */
             DEBUGLOG(6, "invalidating dictionary for current block (distance > windowSize)");
             *loadedDictEndPtr = 0;
···
  * forget about the extDict. Handles overlap of the prefix and extDict.
  * Returns non-zero if the segment is contiguous.
  */
-MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window,
-                                  void const* src, size_t srcSize,
-                                  int forceNonContiguous)
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_window_update(ZSTD_window_t* window,
+                 const void* src, size_t srcSize,
+                       int forceNonContiguous)
 {
     BYTE const* const ip = (BYTE const*)src;
     U32 contiguous = 1;
···
     /* if input and dictionary overlap : reduce dictionary (area presumed modified by input) */
     if ( (ip+srcSize > window->dictBase + window->lowLimit)
        & (ip < window->dictBase + window->dictLimit)) {
-        ptrdiff_t const highInputIdx = (ip + srcSize) - window->dictBase;
-        U32 const lowLimitMax = (highInputIdx > (ptrdiff_t)window->dictLimit) ? window->dictLimit : (U32)highInputIdx;
+        size_t const highInputIdx = (size_t)((ip + srcSize) - window->dictBase);
+        U32 const lowLimitMax = (highInputIdx > (size_t)window->dictLimit) ? window->dictLimit : (U32)highInputIdx;
+        assert(highInputIdx < UINT_MAX);
         window->lowLimit = lowLimitMax;
         DEBUGLOG(5, "Overlapping extDict and input : new lowLimit = %u", window->lowLimit);
     }
···
 /*
  * Returns the lowest allowed match index. It may either be in the ext-dict or the prefix.
  */
-MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_matchState_t* ms, U32 curr, unsigned windowLog)
+MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_MatchState_t* ms, U32 curr, unsigned windowLog)
 {
     U32 const maxDistance = 1U << windowLog;
     U32 const lowestValid = ms->window.lowLimit;
···
 /*
  * Returns the lowest allowed match index in the prefix.
  */
-MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_matchState_t* ms, U32 curr, unsigned windowLog)
+MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_MatchState_t* ms, U32 curr, unsigned windowLog)
 {
     U32    const maxDistance = 1U << windowLog;
     U32    const lowestValid = ms->window.dictLimit;
···
     return matchLowest;
 }
 
+/* index_safety_check:
+ * intentional underflow : ensure repIndex isn't overlapping dict + prefix
+ * @return 1 if values are not overlapping,
+ * 0 otherwise */
+MEM_STATIC int ZSTD_index_overlap_check(const U32 prefixLowestIndex, const U32 repIndex) {
+    return ((U32)((prefixLowestIndex-1)  - repIndex) >= 3);
+}
 
 
 /* debug functions */
···
 
 #endif
 
+/* Short Cache */
 
+/* Normally, zstd matchfinders follow this flow:
+ *     1. Compute hash at ip
+ *     2. Load index from hashTable[hash]
+ *     3. Check if *ip == *(base + index)
+ * In dictionary compression, loading *(base + index) is often an L2 or even L3 miss.
+ *
+ * Short cache is an optimization which allows us to avoid step 3 most of the time
+ * when the data doesn't actually match. With short cache, the flow becomes:
+ *     1. Compute (hash, currentTag) at ip. currentTag is an 8-bit independent hash at ip.
+ *     2. Load (index, matchTag) from hashTable[hash]. See ZSTD_writeTaggedIndex to understand how this works.
+ *     3. Only if currentTag == matchTag, check *ip == *(base + index). Otherwise, continue.
+ *
+ * Currently, short cache is only implemented in CDict hashtables. Thus, its use is limited to
+ * dictMatchState matchfinders.
+ */
+#define ZSTD_SHORT_CACHE_TAG_BITS 8
+#define ZSTD_SHORT_CACHE_TAG_MASK ((1u << ZSTD_SHORT_CACHE_TAG_BITS) - 1)
+
+/* Helper function for ZSTD_fillHashTable and ZSTD_fillDoubleHashTable.
+ * Unpacks hashAndTag into (hash, tag), then packs (index, tag) into hashTable[hash]. */
+MEM_STATIC void ZSTD_writeTaggedIndex(U32* const hashTable, size_t hashAndTag, U32 index) {
+    size_t const hash = hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
+    U32 const tag = (U32)(hashAndTag & ZSTD_SHORT_CACHE_TAG_MASK);
+    assert(index >> (32 - ZSTD_SHORT_CACHE_TAG_BITS) == 0);
+    hashTable[hash] = (index << ZSTD_SHORT_CACHE_TAG_BITS) | tag;
+}
+
+/* Helper function for short cache matchfinders.
+ * Unpacks tag1 and tag2 from lower bits of packedTag1 and packedTag2, then checks if the tags match. */
+MEM_STATIC int ZSTD_comparePackedTags(size_t packedTag1, size_t packedTag2) {
+    U32 const tag1 = packedTag1 & ZSTD_SHORT_CACHE_TAG_MASK;
+    U32 const tag2 = packedTag2 & ZSTD_SHORT_CACHE_TAG_MASK;
+    return tag1 == tag2;
+}
 
 /* ===============================================================
  * Shared internal declarations
···
 
 void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs);
 
+typedef struct {
+    U32 idx;            /* Index in array of ZSTD_Sequence */
+    U32 posInSequence;  /* Position within sequence at idx */
+    size_t posInSrc;    /* Number of bytes given by sequences provided so far */
+} ZSTD_SequencePosition;
+
+/* for benchmark */
+size_t ZSTD_convertBlockSequences(ZSTD_CCtx* cctx,
+                        const ZSTD_Sequence* const inSeqs, size_t nbSequences,
+                        int const repcodeResolution);
+
+typedef struct {
+    size_t nbSequences;
+    size_t blockSize;
+    size_t litSize;
+} BlockSummary;
+
+BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs);
+
 /* ==============================================================
  * Private declarations
  * These prototypes shall only be called from within lib/compress
···
  * Note: srcSizeHint == 0 means 0!
  */
 ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
-        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode);
+        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode);
 
 /*! ZSTD_initCStream_internal() :
  *  Private use only. Init streaming operation.
···
                      const ZSTD_CDict* cdict,
                      const ZSTD_CCtx_params* params, unsigned long long pledgedSrcSize);
 
-void ZSTD_resetSeqStore(seqStore_t* ssPtr);
+void ZSTD_resetSeqStore(SeqStore_t* ssPtr);
 
 /*! ZSTD_getCParamsFromCDict() :
  *  as the name implies */
···
  * This cannot be used when long range matching is enabled.
  * Zstd will use these sequences, and pass the literals to a secondary block
  * compressor.
- * @return : An error code on failure.
  * NOTE: seqs are not verified! Invalid sequences can cause out-of-bounds memory
  * access and data corruption.
  */
-size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSeq);
+void ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSeq);
 
 /* ZSTD_cycleLog() :
  *  condition for correct operation : hashLog > 1 */
···
  *  Trace the end of a compression call.
  */
 void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize);
+
+/* Returns 1 if an external sequence producer is registered, otherwise returns 0. */
+MEM_STATIC int ZSTD_hasExtSeqProd(const ZSTD_CCtx_params* params) {
+    return params->extSeqProdFunc != NULL;
+}
+
+/* ===============================================================
+ * Deprecated definitions that are still used internally to avoid
+ * deprecation warnings. These functions are exactly equivalent to
+ * their public variants, but avoid the deprecation warnings.
+ * =============================================================== */
+
+size_t ZSTD_compressBegin_usingCDict_deprecated(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict);
+
+size_t ZSTD_compressContinue_public(ZSTD_CCtx* cctx,
+                                    void* dst, size_t dstCapacity,
+                              const void* src, size_t srcSize);
+
+size_t ZSTD_compressEnd_public(ZSTD_CCtx* cctx,
+                               void* dst, size_t dstCapacity,
+                         const void* src, size_t srcSize);
+
+size_t ZSTD_compressBlock_deprecated(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
 
 #endif /* ZSTD_COMPRESS_H */

lib/zstd/compress/zstd_compress_literals.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
  ***************************************/
 #include "zstd_compress_literals.h"
 
+
+/* **************************************************************
+*  Debug Traces
+****************************************************************/
+#if DEBUGLEVEL >= 2
+
+static size_t showHexa(const void* src, size_t srcSize)
+{
+    const BYTE* const ip = (const BYTE*)src;
+    size_t u;
+    for (u=0; u<srcSize; u++) {
+        RAWLOG(5, " %02X", ip[u]); (void)ip;
+    }
+    RAWLOG(5, " \n");
+    return srcSize;
+}
+
+#endif
+
+
+/* **************************************************************
+*  Literals compression - special cases
+****************************************************************/
 size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
 {
     BYTE* const ostart = (BYTE*)dst;
     U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
+
+    DEBUGLOG(5, "ZSTD_noCompressLiterals: srcSize=%zu, dstCapacity=%zu", srcSize, dstCapacity);
 
     RETURN_ERROR_IF(srcSize + flSize > dstCapacity, dstSize_tooSmall, "");
 
···
     }
 
     ZSTD_memcpy(ostart + flSize, src, srcSize);
-    DEBUGLOG(5, "Raw literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize));
+    DEBUGLOG(5, "Raw (uncompressed) literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize));
     return srcSize + flSize;
+}
+
+static int allBytesIdentical(const void* src, size_t srcSize)
+{
+    assert(srcSize >= 1);
+    assert(src != NULL);
+    {   const BYTE b = ((const BYTE*)src)[0];
+        size_t p;
+        for (p=1; p<srcSize; p++) {
+            if (((const BYTE*)src)[p] != b) return 0;
+        }
+        return 1;
+    }
 }
 
 size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
···
     BYTE* const ostart = (BYTE*)dst;
     U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
 
-    (void)dstCapacity;  /* dstCapacity already guaranteed to be >=4, hence large enough */
+    assert(dstCapacity >= 4); (void)dstCapacity;
+    assert(allBytesIdentical(src, srcSize));
 
     switch(flSize)
     {
···
     }
 
     ostart[flSize] = *(const BYTE*)src;
-    DEBUGLOG(5, "RLE literals: %u -> %u", (U32)srcSize, (U32)flSize + 1);
+    DEBUGLOG(5, "RLE : Repeated Literal (%02X: %u times) -> %u bytes encoded", ((const BYTE*)src)[0], (U32)srcSize, (U32)flSize + 1);
     return flSize+1;
 }
 
-size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
-                              ZSTD_hufCTables_t* nextHuf,
-                              ZSTD_strategy strategy, int disableLiteralCompression,
-                              void* dst, size_t dstCapacity,
-                        const void* src, size_t srcSize,
-                              void* entropyWorkspace, size_t entropyWorkspaceSize,
-                        const int bmi2,
-                        unsigned suspectUncompressible)
+/* ZSTD_minLiteralsToCompress() :
+ * returns minimal amount of literals
+ * for literal compression to even be attempted.
+ * Minimum is made tighter as compression strategy increases.
+ */
+static size_t
+ZSTD_minLiteralsToCompress(ZSTD_strategy strategy, HUF_repeat huf_repeat)
 {
-    size_t const minGain = ZSTD_minGain(srcSize, strategy);
+    assert((int)strategy >= 0);
+    assert((int)strategy <= 9);
+    /* btultra2 : min 8 bytes;
+     * then 2x larger for each successive compression strategy
+     * max threshold 64 bytes */
+    {   int const shift = MIN(9-(int)strategy, 3);
+        size_t const mintc = (huf_repeat == HUF_repeat_valid) ? 6 : (size_t)8 << shift;
+        DEBUGLOG(7, "minLiteralsToCompress = %zu", mintc);
+        return mintc;
+    }
+}
+
+size_t ZSTD_compressLiterals (
+                  void* dst, size_t dstCapacity,
+            const void* src, size_t srcSize,
+                  void* entropyWorkspace, size_t entropyWorkspaceSize,
+            const ZSTD_hufCTables_t* prevHuf,
+                  ZSTD_hufCTables_t* nextHuf,
+                  ZSTD_strategy strategy,
+                  int disableLiteralCompression,
+                  int suspectUncompressible,
+                  int bmi2)
+{
     size_t const lhSize = 3 + (srcSize >= 1 KB) + (srcSize >= 16 KB);
     BYTE*  const ostart = (BYTE*)dst;
     U32 singleStream = srcSize < 256;
-    symbolEncodingType_e hType = set_compressed;
+    SymbolEncodingType_e hType = set_compressed;
     size_t cLitSize;
 
-    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i srcSize=%u)",
-                disableLiteralCompression, (U32)srcSize);
+    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i, srcSize=%u, dstCapacity=%zu)",
+                disableLiteralCompression, (U32)srcSize, dstCapacity);
+
+    DEBUGLOG(6, "Completed literals listing (%zu bytes)", showHexa(src, srcSize));
 
     /* Prepare nextEntropy assuming reusing the existing table */
     ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
···
     if (disableLiteralCompression)
         return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
 
-    /* small ? don't even attempt compression (speed opt) */
-#   define COMPRESS_LITERALS_SIZE_MIN 63
-    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
-        if (srcSize <= minLitSize) return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
-    }
+    /* if too small, don't even attempt compression (speed opt) */
+    if (srcSize < ZSTD_minLiteralsToCompress(strategy, prevHuf->repeatMode))
+        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
 
     RETURN_ERROR_IF(dstCapacity < lhSize+1, dstSize_tooSmall, "not enough space for compression");
     {   HUF_repeat repeat = prevHuf->repeatMode;
-        int const preferRepeat = strategy < ZSTD_lazy ? srcSize <= 1024 : 0;
+        int const flags = 0
+            | (bmi2 ? HUF_flags_bmi2 : 0)
+            | (strategy < ZSTD_lazy && srcSize <= 1024 ? HUF_flags_preferRepeat : 0)
+            | (strategy >= HUF_OPTIMAL_DEPTH_THRESHOLD ? HUF_flags_optimalDepth : 0)
+            | (suspectUncompressible ? HUF_flags_suspectUncompressible : 0);
+
+        typedef size_t (*huf_compress_f)(void*, size_t, const void*, size_t, unsigned, unsigned, void*, size_t, HUF_CElt*, HUF_repeat*, int);
+        huf_compress_f huf_compress;
         if (repeat == HUF_repeat_valid && lhSize == 3) singleStream = 1;
-        cLitSize = singleStream ?
-            HUF_compress1X_repeat(
-                ostart+lhSize, dstCapacity-lhSize, src, srcSize,
-                HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2, suspectUncompressible) :
-            HUF_compress4X_repeat(
-                ostart+lhSize, dstCapacity-lhSize, src, srcSize,
-                HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2, suspectUncompressible);
+        huf_compress = singleStream ? HUF_compress1X_repeat : HUF_compress4X_repeat;
+        cLitSize = huf_compress(ostart+lhSize, dstCapacity-lhSize,
+                                src, srcSize,
+                                HUF_SYMBOLVALUE_MAX, LitHufLog,
+                                entropyWorkspace, entropyWorkspaceSize,
+                                (HUF_CElt*)nextHuf->CTable,
+                                &repeat, flags);
+        DEBUGLOG(5, "%zu literals compressed into %zu bytes (before header)", srcSize, cLitSize);
         if (repeat != HUF_repeat_none) {
             /* reused the existing table */
-            DEBUGLOG(5, "Reusing previous huffman table");
+            DEBUGLOG(5, "reusing statistics from previous huffman block");
             hType = set_repeat;
         }
     }
 
-    if ((cLitSize==0) || (cLitSize >= srcSize - minGain) || ERR_isError(cLitSize)) {
-        ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
-    }
+    {   size_t const minGain = ZSTD_minGain(srcSize, strategy);
+        if ((cLitSize==0) || (cLitSize >= srcSize - minGain) || ERR_isError(cLitSize)) {
+            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+            return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
+    }   }
     if (cLitSize==1) {
-        ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-        return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize);
-    }
+        /* A return value of 1 signals that the alphabet consists of a single symbol.
+         * However, in some rare circumstances, it could be the compressed size (a single byte).
+         * For that outcome to have a chance to happen, it's necessary that `srcSize < 8`.
+         * (it's also necessary to not generate statistics).
+         * Therefore, in such a case, actively check that all bytes are identical. */
+        if ((srcSize >= 8) || allBytesIdentical(src, srcSize)) {
+            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+            return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize);
+    }   }
 
     if (hType == set_compressed) {
         /* using a newly constructed table */
···
     switch(lhSize)
     {
     case 3: /* 2 - 2 - 10 - 10 */
-        {   U32 const lhc = hType + ((!singleStream) << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<14);
+        if (!singleStream) assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
+        {   U32 const lhc = hType + ((U32)(!singleStream) << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<14);
             MEM_writeLE24(ostart, lhc);
             break;
         }
     case 4: /* 2 - 2 - 14 - 14 */
+        assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
         {   U32 const lhc = hType + (2 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<18);
             MEM_writeLE32(ostart, lhc);
             break;
         }
     case 5: /* 2 - 2 - 18 - 18 */
+        assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
         {   U32 const lhc = hType + (3 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<22);
             MEM_writeLE32(ostart, lhc);
             ostart[4] = (BYTE)(cLitSize >> 10);

lib/zstd/compress/zstd_compress_literals.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 
 size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
+/* ZSTD_compressRleLiteralsBlock() :
+ * Conditions :
+ * - All bytes in @src are identical
+ * - dstCapacity >= 4 */
 size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
-/* If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
-size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
-                              ZSTD_hufCTables_t* nextHuf,
-                              ZSTD_strategy strategy, int disableLiteralCompression,
-                              void* dst, size_t dstCapacity,
+/* ZSTD_compressLiterals():
+ * @entropyWorkspace: must be aligned on 4-bytes boundaries
+ * @entropyWorkspaceSize : must be >= HUF_WORKSPACE_SIZE
+ * @suspectUncompressible: sampling checks, to potentially skip huffman coding
+ */
+size_t ZSTD_compressLiterals (void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize,
                               void* entropyWorkspace, size_t entropyWorkspaceSize,
-                        const int bmi2,
-                        unsigned suspectUncompressible);
+                        const ZSTD_hufCTables_t* prevHuf,
+                              ZSTD_hufCTables_t* nextHuf,
+                              ZSTD_strategy strategy, int disableLiteralCompression,
+                              int suspectUncompressible,
+                              int bmi2);
 
 #endif /* ZSTD_COMPRESS_LITERALS_H */

lib/zstd/compress/zstd_compress_sequences.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 {
     /* Heuristic: This should cover most blocks <= 16K and
      * start to fade out after 16K to about 32K depending on
-     * comprssibility.
+     * compressibility.
      */
     return nbSeq >= 2048;
 }
···
     return cost >> 8;
 }
 
-symbolEncodingType_e
+SymbolEncodingType_e
 ZSTD_selectEncodingType(
         FSE_repeat* repeatMode, unsigned const* count, unsigned const max,
         size_t const mostFrequent, size_t nbSeq, unsigned const FSELog,
         FSE_CTable const* prevCTable,
         short const* defaultNorm, U32 defaultNormLog,
-        ZSTD_defaultPolicy_e const isDefaultAllowed,
+        ZSTD_DefaultPolicy_e const isDefaultAllowed,
         ZSTD_strategy const strategy)
 {
     ZSTD_STATIC_ASSERT(ZSTD_defaultDisallowed == 0 && ZSTD_defaultAllowed != 0);
     if (mostFrequent == nbSeq) {
         *repeatMode = FSE_repeat_none;
         if (isDefaultAllowed && nbSeq <= 2) {
-            /* Prefer set_basic over set_rle when there are 2 or less symbols,
+            /* Prefer set_basic over set_rle when there are 2 or fewer symbols,
              * since RLE uses 1 byte, but set_basic uses 5-6 bits per symbol.
              * If basic encoding isn't possible, always choose RLE.
              */
···
 
 size_t
 ZSTD_buildCTable(void* dst, size_t dstCapacity,
-                FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type,
+                FSE_CTable* nextCTable, U32 FSELog, SymbolEncodingType_e type,
                 unsigned* count, U32 max,
                 const BYTE* codeTable, size_t nbSeq,
                 const S16* defaultNorm, U32 defaultNormLog, U32 defaultMax,
···
             FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
             FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
             FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
-            seqDef const* sequences, size_t nbSeq, int longOffsets)
+            SeqDef const* sequences, size_t nbSeq, int longOffsets)
 {
     BIT_CStream_t blockStream;
     FSE_CState_t  stateMatchLength;
···
             FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
             FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
             FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
-            seqDef const* sequences, size_t nbSeq, int longOffsets)
+            SeqDef const* sequences, size_t nbSeq, int longOffsets)
 {
     return ZSTD_encodeSequences_body(dst, dstCapacity,
                                     CTable_MatchLength, mlCodeTable,
···
             FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
             FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
             FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
-            seqDef const* sequences, size_t nbSeq, int longOffsets)
+            SeqDef const* sequences, size_t nbSeq, int longOffsets)
 {
     return ZSTD_encodeSequences_body(dst, dstCapacity,
                                     CTable_MatchLength, mlCodeTable,
···
             FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
             FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
             FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
-            seqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2)
+            SeqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2)
 {
     DEBUGLOG(5, "ZSTD_encodeSequences: dstCapacity = %u", (unsigned)dstCapacity);
 #if DYNAMIC_BMI2

lib/zstd/compress/zstd_compress_sequences.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #ifndef ZSTD_COMPRESS_SEQUENCES_H
 #define ZSTD_COMPRESS_SEQUENCES_H
 
+#include "zstd_compress_internal.h" /* SeqDef */
 #include "../common/fse.h" /* FSE_repeat, FSE_CTable */
-#include "../common/zstd_internal.h" /* symbolEncodingType_e, ZSTD_strategy */
+#include "../common/zstd_internal.h" /* SymbolEncodingType_e, ZSTD_strategy */
 
 typedef enum {
     ZSTD_defaultDisallowed = 0,
     ZSTD_defaultAllowed = 1
-} ZSTD_defaultPolicy_e;
+} ZSTD_DefaultPolicy_e;
 
-symbolEncodingType_e
+SymbolEncodingType_e
 ZSTD_selectEncodingType(
         FSE_repeat* repeatMode, unsigned const* count, unsigned const max,
         size_t const mostFrequent, size_t nbSeq, unsigned const FSELog,
         FSE_CTable const* prevCTable,
         short const* defaultNorm, U32 defaultNormLog,
-        ZSTD_defaultPolicy_e const isDefaultAllowed,
+        ZSTD_DefaultPolicy_e const isDefaultAllowed,
         ZSTD_strategy const strategy);
 
 size_t
 ZSTD_buildCTable(void* dst, size_t dstCapacity,
-                FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type,
+                FSE_CTable* nextCTable, U32 FSELog, SymbolEncodingType_e type,
                 unsigned* count, U32 max,
                 const BYTE* codeTable, size_t nbSeq,
                 const S16* defaultNorm, U32 defaultNormLog, U32 defaultMax,
···
             FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
             FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
             FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
-            seqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2);
+            SeqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2);
 
 size_t ZSTD_fseBitCost(
     FSE_CTable const* ctable,

lib/zstd/compress/zstd_compress_superblock.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
  *      If it is set_compressed, first sub-block's literals section will be Treeless_Literals_Block
  *      and the following sub-blocks' literals sections will be Treeless_Literals_Block.
  *  @return : compressed size of literals section of a sub-block
- *            Or 0 if it unable to compress.
+ *            Or 0 if unable to compress.
  *            Or error code */
-static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
-                                    const ZSTD_hufCTablesMetadata_t* hufMetadata,
-                                    const BYTE* literals, size_t litSize,
-                                    void* dst, size_t dstSize,
-                                    const int bmi2, int writeEntropy, int* entropyWritten)
+static size_t
+ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
+                              const ZSTD_hufCTablesMetadata_t* hufMetadata,
+                              const BYTE* literals, size_t litSize,
+                              void* dst, size_t dstSize,
+                              const int bmi2, int writeEntropy, int* entropyWritten)
 {
     size_t const header = writeEntropy ? 200 : 0;
     size_t const lhSize = 3 + (litSize >= (1 KB - header)) + (litSize >= (16 KB - header));
···
     BYTE* const oend = ostart + dstSize;
     BYTE* op = ostart + lhSize;
     U32 const singleStream = lhSize == 3;
-    symbolEncodingType_e hType = writeEntropy ? hufMetadata->hType : set_repeat;
+    SymbolEncodingType_e hType = writeEntropy ? hufMetadata->hType : set_repeat;
     size_t cLitSize = 0;
-
-    (void)bmi2; /* TODO bmi2... */
 
     DEBUGLOG(5, "ZSTD_compressSubBlock_literal (litSize=%zu, lhSize=%zu, writeEntropy=%d)", litSize, lhSize, writeEntropy);
 
···
         DEBUGLOG(5, "ZSTD_compressSubBlock_literal (hSize=%zu)", hufMetadata->hufDesSize);
     }
 
-    /* TODO bmi2 */
-    {   const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, oend-op, literals, litSize, hufTable)
-                                          : HUF_compress4X_usingCTable(op, oend-op, literals, litSize, hufTable);
+    {   int const flags = bmi2 ? HUF_flags_bmi2 : 0;
+        const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, (size_t)(oend-op), literals, litSize, hufTable, flags)
+                                          : HUF_compress4X_usingCTable(op, (size_t)(oend-op), literals, litSize, hufTable, flags);
         op += cSize;
         cLitSize += cSize;
         if (cSize == 0 || ERR_isError(cSize)) {
···
     switch(lhSize)
     {
     case 3: /* 2 - 2 - 10 - 10 */
-        {   U32 const lhc = hType + ((!singleStream) << 2) + ((U32)litSize<<4) + ((U32)cLitSize<<14);
+        {   U32 const lhc = hType + ((U32)(!singleStream) << 2) + ((U32)litSize<<4) + ((U32)cLitSize<<14);
             MEM_writeLE24(ostart, lhc);
             break;
         }
···
     }
     *entropyWritten = 1;
     DEBUGLOG(5, "Compressed literals: %u -> %u", (U32)litSize, (U32)(op-ostart));
-    return op-ostart;
+    return (size_t)(op-ostart);
 }
 
-static size_t ZSTD_seqDecompressedSize(seqStore_t const* seqStore, const seqDef* sequences, size_t nbSeq, size_t litSize, int lastSequence) {
-    const seqDef* const sstart = sequences;
-    const seqDef* const send = sequences + nbSeq;
-    const seqDef* sp = sstart;
+static size_t
+ZSTD_seqDecompressedSize(SeqStore_t const* seqStore,
+                   const SeqDef* sequences, size_t nbSeqs,
+                         size_t litSize, int lastSubBlock)
+{
     size_t matchLengthSum = 0;
     size_t litLengthSum = 0;
-    (void)(litLengthSum); /* suppress unused variable warning on some environments */
-    while (send-sp > 0) {
-        ZSTD_sequenceLength const seqLen = ZSTD_getSequenceLength(seqStore, sp);
+    size_t n;
+    for (n=0; n<nbSeqs; n++) {
+        const ZSTD_SequenceLength seqLen = ZSTD_getSequenceLength(seqStore, sequences+n);
         litLengthSum += seqLen.litLength;
         matchLengthSum += seqLen.matchLength;
-        sp++;
     }
-    assert(litLengthSum <= litSize);
-    if (!lastSequence) {
+    DEBUGLOG(5, "ZSTD_seqDecompressedSize: %u sequences from %p: %u literals + %u matchlength",
+                (unsigned)nbSeqs, (const void*)sequences,
+                (unsigned)litLengthSum, (unsigned)matchLengthSum);
+    if (!lastSubBlock)
         assert(litLengthSum == litSize);
-    }
+    else
+        assert(litLengthSum <= litSize);
+    (void)litLengthSum;
     return matchLengthSum + litSize;
 }
 
···
  *  @return : compressed size of sequences section of a sub-block
  *            Or 0 if it is unable to compress
  *            Or error code. */
-static size_t ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables,
-                                              const ZSTD_fseCTablesMetadata_t* fseMetadata,
-                                              const seqDef* sequences, size_t nbSeq,
-                                              const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
-                                              const ZSTD_CCtx_params* cctxParams,
-                                              void* dst, size_t dstCapacity,
-                                              const int bmi2, int writeEntropy, int* entropyWritten)
+static size_t
+ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables,
+                                const ZSTD_fseCTablesMetadata_t* fseMetadata,
+                                const SeqDef* sequences, size_t nbSeq,
+                                const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
+                                const ZSTD_CCtx_params* cctxParams,
+                                void* dst, size_t dstCapacity,
+                                const int bmi2, int writeEntropy, int* entropyWritten)
 {
     const int longOffsets = cctxParams->cParams.windowLog > STREAM_ACCUMULATOR_MIN;
     BYTE* const ostart = (BYTE*)dst;
···
     /* Sequences Header */
     RETURN_ERROR_IF((oend-op) < 3 /*max nbSeq Size*/ + 1 /*seqHead*/,
                     dstSize_tooSmall, "");
-    if (nbSeq < 0x7F)
+    if (nbSeq < 128)
         *op++ = (BYTE)nbSeq;
     else if (nbSeq < LONGNBSEQ)
         op[0] = (BYTE)((nbSeq>>8) + 0x80), op[1] = (BYTE)nbSeq, op+=2;
     else
         op[0]=0xFF, MEM_writeLE16(op+1, (U16)(nbSeq - LONGNBSEQ)), op+=3;
     if (nbSeq==0) {
-        return op - ostart;
+        return (size_t)(op - ostart);
     }
 
     /* seqHead : flags for FSE encoding type */
···
     }
 
     {   size_t const bitstreamSize = ZSTD_encodeSequences(
-                                        op, oend - op,
+                                        op, (size_t)(oend - op),
                                         fseTables->matchlengthCTable, mlCode,
                                         fseTables->offcodeCTable, ofCode,
                                         fseTables->litlengthCTable, llCode,
···
 #endif
 
     *entropyWritten = 1;
-    return op - ostart;
+    return (size_t)(op - ostart);
 }
 
 /* ZSTD_compressSubBlock() :
···
  *            Or 0 if it failed to compress. */
 static size_t ZSTD_compressSubBlock(const ZSTD_entropyCTables_t* entropy,
                                     const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
-                                    const seqDef* sequences, size_t nbSeq,
+                                    const SeqDef* sequences, size_t nbSeq,
                                     const BYTE* literals, size_t litSize,
                                     const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
                                     const ZSTD_CCtx_params* cctxParams,
···
                 litSize, nbSeq, writeLitEntropy, writeSeqEntropy, lastBlock);
     {   size_t cLitSize = ZSTD_compressSubBlock_literal((const HUF_CElt*)entropy->huf.CTable,
                                                         &entropyMetadata->hufMetadata, literals, litSize,
-                                                        op, oend-op, bmi2, writeLitEntropy, litEntropyWritten);
+                                                        op, (size_t)(oend-op),
+                                                        bmi2, writeLitEntropy, litEntropyWritten);
         FORWARD_IF_ERROR(cLitSize, "ZSTD_compressSubBlock_literal failed");
         if (cLitSize == 0) return 0;
         op += cLitSize;
···
                                                   sequences, nbSeq,
                                                   llCode, mlCode, ofCode,
                                                   cctxParams,
-                                                  op, oend-op,
+                                                  op, (size_t)(oend-op),
                                                   bmi2, writeSeqEntropy, seqEntropyWritten);
         FORWARD_IF_ERROR(cSeqSize, "ZSTD_compressSubBlock_sequences failed");
         if (cSeqSize == 0) return 0;
         op += cSeqSize;
     }
     /* Write block header */
-    {   size_t cSize = (op-ostart)-ZSTD_blockHeaderSize;
+    {   size_t cSize = (size_t)(op-ostart) - ZSTD_blockHeaderSize;
         U32 const cBlockHeader24 = lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3);
         MEM_writeLE24(ostart, cBlockHeader24);
     }
-    return op-ostart;
+    return (size_t)(op-ostart);
 }
 
 static size_t ZSTD_estimateSubBlockSize_literal(const BYTE* literals, size_t litSize,
···
     return 0;
 }
 
-static size_t ZSTD_estimateSubBlockSize_symbolType(symbolEncodingType_e type,
+static size_t ZSTD_estimateSubBlockSize_symbolType(SymbolEncodingType_e type,
                         const BYTE* codeTable, unsigned maxCode,
                         size_t nbSeq, const FSE_CTable* fseCTable,
                         const U8* additionalBits,
···
     return cSeqSizeEstimate + sequencesSectionHeaderSize;
 }
 
-static size_t ZSTD_estimateSubBlockSize(const BYTE* literals, size_t litSize,
+typedef struct {
+    size_t estLitSize;
+    size_t estBlockSize;
+} EstimatedBlockSize;
+static EstimatedBlockSize ZSTD_estimateSubBlockSize(const BYTE* literals, size_t litSize,
                                         const BYTE* ofCodeTable,
                                         const BYTE* llCodeTable,
                                         const BYTE* mlCodeTable,
···
                                         const ZSTD_entropyCTables_t* entropy,
                                         const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
                                         void* workspace, size_t wkspSize,
-                                        int writeLitEntropy, int writeSeqEntropy) {
-    size_t cSizeEstimate = 0;
-    cSizeEstimate += ZSTD_estimateSubBlockSize_literal(literals, litSize,
-                                                         &entropy->huf, &entropyMetadata->hufMetadata,
-                                                         workspace, wkspSize, writeLitEntropy);
-    cSizeEstimate += ZSTD_estimateSubBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable,
+                                        int writeLitEntropy, int writeSeqEntropy)
+{
+    EstimatedBlockSize ebs;
+    ebs.estLitSize = ZSTD_estimateSubBlockSize_literal(literals, litSize,
+                                                        &entropy->huf, &entropyMetadata->hufMetadata,
+                                                        workspace, wkspSize, writeLitEntropy);
+    ebs.estBlockSize = ZSTD_estimateSubBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable,
                                                          nbSeq, &entropy->fse, &entropyMetadata->fseMetadata,
                                                          workspace, wkspSize, writeSeqEntropy);
-    return cSizeEstimate + ZSTD_blockHeaderSize;
+    ebs.estBlockSize += ebs.estLitSize + ZSTD_blockHeaderSize;
+    return ebs;
 }
 
 static int ZSTD_needSequenceEntropyTables(ZSTD_fseCTablesMetadata_t const* fseMetadata)
···
     return 0;
 }
 
+static size_t countLiterals(SeqStore_t const* seqStore, const SeqDef* sp, size_t seqCount)
+{
+    size_t n, total = 0;
+    assert(sp != NULL);
+    for (n=0; n<seqCount; n++) {
+        total += ZSTD_getSequenceLength(seqStore, sp+n).litLength;
+    }
+    DEBUGLOG(6, "countLiterals for %zu sequences from %p => %zu bytes", seqCount, (const void*)sp, total);
+    return total;
+}
+
+#define BYTESCALE 256
+
+static size_t sizeBlockSequences(const SeqDef* sp, size_t nbSeqs,
+                size_t targetBudget, size_t avgLitCost, size_t avgSeqCost,
+                int firstSubBlock)
+{
+    size_t n, budget = 0, inSize=0;
+    /* entropy headers */
+    size_t const headerSize = (size_t)firstSubBlock * 120 * BYTESCALE; /* generous estimate */
+    assert(firstSubBlock==0 || firstSubBlock==1);
+    budget += headerSize;
+
+    /* first sequence => at least one sequence*/
+    budget += sp[0].litLength * avgLitCost + avgSeqCost;
+    if (budget > targetBudget) return 1;
+    inSize = sp[0].litLength + (sp[0].mlBase+MINMATCH);
+
+    /* loop over sequences */
+    for (n=1; n<nbSeqs; n++) {
+        size_t currentCost = sp[n].litLength * avgLitCost + avgSeqCost;
+        budget += currentCost;
+        inSize += sp[n].litLength + (sp[n].mlBase+MINMATCH);
+        /* stop when sub-block budget is reached */
+        if ( (budget > targetBudget)
+            /* though continue to expand until the sub-block is deemed compressible */
+          && (budget < inSize * BYTESCALE) )
+            break;
+    }
+
+    return n;
+}
+
 /* ZSTD_compressSubBlock_multi() :
  *  Breaks super-block into multiple sub-blocks and compresses them.
- *  Entropy will be written to the first block.
- *  The following blocks will use repeat mode to compress.
- *  All sub-blocks are compressed blocks (no raw or rle blocks).
- *  @return : compressed size of the super block (which is multiple ZSTD blocks)
- *            Or 0 if it failed to compress. */
-static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr,
+ *  Entropy will be written into the first block.
+ *  The following blocks use repeat_mode to compress.
+ *  Sub-blocks are all compressed, except the last one when beneficial.
+ *  @return : compressed size of the super block (which features multiple ZSTD blocks)
+ *            or 0 if it failed to compress. */
+static size_t ZSTD_compressSubBlock_multi(const SeqStore_t* seqStorePtr,
                             const ZSTD_compressedBlockState_t* prevCBlock,
                             ZSTD_compressedBlockState_t* nextCBlock,
                             const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
···
                             const int bmi2, U32 lastBlock,
                             void* workspace, size_t wkspSize)
 {
-    const seqDef* const sstart = seqStorePtr->sequencesStart;
-    const seqDef* const send = seqStorePtr->sequences;
-    const seqDef* sp = sstart;
+    const SeqDef* const sstart = seqStorePtr->sequencesStart;
+    const SeqDef* const send = seqStorePtr->sequences;
+    const SeqDef* sp = sstart; /* tracks progresses within seqStorePtr->sequences */
+    size_t const nbSeqs = (size_t)(send - sstart);
     const BYTE* const lstart = seqStorePtr->litStart;
     const BYTE* const lend = seqStorePtr->lit;
     const BYTE* lp = lstart;
+    size_t const nbLiterals = (size_t)(lend - lstart);
     BYTE const* ip = (BYTE const*)src;
     BYTE const* const iend = ip + srcSize;
     BYTE* const ostart = (BYTE*)dst;
···
     const BYTE* llCodePtr = seqStorePtr->llCode;
     const BYTE* mlCodePtr = seqStorePtr->mlCode;
     const BYTE* ofCodePtr = seqStorePtr->ofCode;
-    size_t targetCBlockSize = cctxParams->targetCBlockSize;
-    size_t litSize, seqCount;
-    int writeLitEntropy = entropyMetadata->hufMetadata.hType == set_compressed;
+    size_t const minTarget = ZSTD_TARGETCBLOCKSIZE_MIN; /* enforce minimum size, to reduce undesirable side effects */
+    size_t const targetCBlockSize = MAX(minTarget, cctxParams->targetCBlockSize);
+    int writeLitEntropy = (entropyMetadata->hufMetadata.hType == set_compressed);
     int writeSeqEntropy = 1;
-    int lastSequence = 0;
 
-    DEBUGLOG(5, "ZSTD_compressSubBlock_multi (litSize=%u, nbSeq=%u)",
-                (unsigned)(lend-lp), (unsigned)(send-sstart));
+    DEBUGLOG(5, "ZSTD_compressSubBlock_multi (srcSize=%u, litSize=%u, nbSeq=%u)",
+               (unsigned)srcSize, (unsigned)(lend-lstart), (unsigned)(send-sstart));
 
-    litSize = 0;
-    seqCount = 0;
-    do {
-        size_t cBlockSizeEstimate = 0;
-        if (sstart == send) {
-            lastSequence = 1;
-        } else {
-            const seqDef* const sequence = sp + seqCount;
-            lastSequence = sequence == send - 1;
-            litSize += ZSTD_getSequenceLength(seqStorePtr, sequence).litLength;
-            seqCount++;
+        /* let's start by a general estimation for the full block */
+    if (nbSeqs > 0) {
+        EstimatedBlockSize const ebs =
+                ZSTD_estimateSubBlockSize(lp, nbLiterals,
+                                        ofCodePtr, llCodePtr, mlCodePtr, nbSeqs,
+                                        &nextCBlock->entropy, entropyMetadata,
+                                        workspace, wkspSize,
+                                        writeLitEntropy, writeSeqEntropy);
+        /* quick estimation */
+        size_t const avgLitCost = nbLiterals ? (ebs.estLitSize * BYTESCALE) / nbLiterals : BYTESCALE;
+        size_t const avgSeqCost = ((ebs.estBlockSize - ebs.estLitSize) * BYTESCALE) / nbSeqs;
+        const size_t nbSubBlocks = MAX((ebs.estBlockSize + (targetCBlockSize/2)) / targetCBlockSize, 1);
+        size_t n, avgBlockBudget, blockBudgetSupp=0;
+        avgBlockBudget = (ebs.estBlockSize * BYTESCALE) / nbSubBlocks;
+        DEBUGLOG(5, "estimated fullblock size=%u bytes ; avgLitCost=%.2f ; avgSeqCost=%.2f ; targetCBlockSize=%u, nbSubBlocks=%u ; avgBlockBudget=%.0f bytes",
+                    (unsigned)ebs.estBlockSize, (double)avgLitCost/BYTESCALE, (double)avgSeqCost/BYTESCALE,
+                    (unsigned)targetCBlockSize, (unsigned)nbSubBlocks, (double)avgBlockBudget/BYTESCALE);
+        /* simplification: if estimates states that the full superblock doesn't compress, just bail out immediately
+         * this will result in the production of a single uncompressed block covering @srcSize.*/
+        if (ebs.estBlockSize > srcSize) return 0;
+
+        /* compress and write sub-blocks */
+        assert(nbSubBlocks>0);
+        for (n=0; n < nbSubBlocks-1; n++) {
+            /* determine nb of sequences for current sub-block + nbLiterals from next sequence */
+            size_t const seqCount = sizeBlockSequences(sp, (size_t)(send-sp),
+                                        avgBlockBudget + blockBudgetSupp, avgLitCost, avgSeqCost, n==0);
+            /* if reached last sequence : break to last sub-block (simplification) */
+            assert(seqCount <= (size_t)(send-sp));
+            if (sp + seqCount == send) break;
+            assert(seqCount > 0);
+            /* compress sub-block */
+            {   int litEntropyWritten = 0;
+                int seqEntropyWritten = 0;
+                size_t litSize = countLiterals(seqStorePtr, sp, seqCount);
+                const size_t decompressedSize =
+                        ZSTD_seqDecompressedSize(seqStorePtr, sp, seqCount, litSize, 0);
+                size_t const cSize = ZSTD_compressSubBlock(&nextCBlock->entropy, entropyMetadata,
+                                                sp, seqCount,
+                                                lp, litSize,
+                                                llCodePtr, mlCodePtr, ofCodePtr,
+                                                cctxParams,
+                                                op, (size_t)(oend-op),
+                                                bmi2, writeLitEntropy, writeSeqEntropy,
+                                                &litEntropyWritten, &seqEntropyWritten,
+                                                0);
+                FORWARD_IF_ERROR(cSize, "ZSTD_compressSubBlock failed");
+
+                /* check compressibility, update state components */
+                if (cSize > 0 && cSize < decompressedSize) {
+                    DEBUGLOG(5, "Committed sub-block compressing %u bytes => %u bytes",
+                                (unsigned)decompressedSize, (unsigned)cSize);
+                    assert(ip + decompressedSize <= iend);
+                    ip += decompressedSize;
+                    lp += litSize;
+                    op += cSize;
+                    llCodePtr += seqCount;
+                    mlCodePtr += seqCount;
+                    ofCodePtr += seqCount;
+                    /* Entropy only needs to be written once */
+                    if (litEntropyWritten) {
+                        writeLitEntropy = 0;
+                    }
+                    if (seqEntropyWritten) {
+                        writeSeqEntropy = 0;
+                    }
+                    sp += seqCount;
+                    blockBudgetSupp = 0;
+            }   }
+            /* otherwise : do not compress yet, coalesce current sub-block with following one */
         }
-        if (lastSequence) {
-            assert(lp <= lend);
-            assert(litSize <= (size_t)(lend - lp));
-            litSize = (size_t)(lend - lp);
-        }
-        /* I think there is an optimization opportunity here.
-         * Calling ZSTD_estimateSubBlockSize for every sequence can be wasteful
-         * since it recalculates estimate from scratch.
-         * For example, it would recount literal distribution and symbol codes every time.
-         */
-        cBlockSizeEstimate = ZSTD_estimateSubBlockSize(lp, litSize, ofCodePtr, llCodePtr, mlCodePtr, seqCount,
-                                                       &nextCBlock->entropy, entropyMetadata,
-                                                       workspace, wkspSize, writeLitEntropy, writeSeqEntropy);
-        if (cBlockSizeEstimate > targetCBlockSize || lastSequence) {
-            int litEntropyWritten = 0;
-            int seqEntropyWritten = 0;
-            const size_t decompressedSize = ZSTD_seqDecompressedSize(seqStorePtr, sp, seqCount, litSize, lastSequence);
-            const size_t cSize = ZSTD_compressSubBlock(&nextCBlock->entropy, entropyMetadata,
-                                                       sp, seqCount,
-                                                       lp, litSize,
-                                                       llCodePtr, mlCodePtr, ofCodePtr,
-                                                       cctxParams,
-                                                       op, oend-op,
-                                                       bmi2, writeLitEntropy, writeSeqEntropy,
-                                                       &litEntropyWritten, &seqEntropyWritten,
-                                                       lastBlock && lastSequence);
-            FORWARD_IF_ERROR(cSize, "ZSTD_compressSubBlock failed");
-            if (cSize > 0 && cSize < decompressedSize) {
-                DEBUGLOG(5, "Committed the sub-block");
-                assert(ip + decompressedSize <= iend);
-                ip += decompressedSize;
-                sp += seqCount;
-                lp += litSize;
-                op += cSize;
-                llCodePtr += seqCount;
-                mlCodePtr += seqCount;
-                ofCodePtr += seqCount;
-                litSize = 0;
-                seqCount = 0;
-                /* Entropy only needs to be written once */
-                if (litEntropyWritten) {
-                    writeLitEntropy = 0;
-                }
-                if (seqEntropyWritten) {
-                    writeSeqEntropy = 0;
-                }
+    } /* if (nbSeqs > 0) */
+
+    /* write last block */
+    DEBUGLOG(5, "Generate last sub-block: %u sequences remaining", (unsigned)(send - sp));
+    {   int litEntropyWritten = 0;
+        int seqEntropyWritten = 0;
+        size_t litSize = (size_t)(lend - lp);
+        size_t seqCount = (size_t)(send - sp);
+        const size_t decompressedSize =
+                ZSTD_seqDecompressedSize(seqStorePtr, sp, seqCount, litSize, 1);
+        size_t const cSize = ZSTD_compressSubBlock(&nextCBlock->entropy, entropyMetadata,
+                                            sp, seqCount,
+                                            lp, litSize,
+                                            llCodePtr, mlCodePtr, ofCodePtr,
+                                            cctxParams,
+                                            op, (size_t)(oend-op),
+                                            bmi2, writeLitEntropy, writeSeqEntropy,
+                                            &litEntropyWritten, &seqEntropyWritten,
+                                            lastBlock);
+        FORWARD_IF_ERROR(cSize, "ZSTD_compressSubBlock failed");
+
+        /* update pointers, the nb of literals borrowed from next sequence must be preserved */
+        if (cSize > 0 && cSize < decompressedSize) {
+            DEBUGLOG(5, "Last sub-block compressed %u bytes => %u bytes",
+                        (unsigned)decompressedSize, (unsigned)cSize);
+            assert(ip + decompressedSize <= iend);
+            ip += decompressedSize;
+            lp += litSize;
+            op += cSize;
+            llCodePtr += seqCount;
+            mlCodePtr += seqCount;
+            ofCodePtr += seqCount;
+            /* Entropy only needs to be written once */
+            if (litEntropyWritten) {
+                writeLitEntropy = 0;
             }
+            if (seqEntropyWritten) {
+                writeSeqEntropy = 0;
+            }
+            sp += seqCount;
         }
-    } while (!lastSequence);
+    }
+
+
     if (writeLitEntropy) {
-        DEBUGLOG(5, "ZSTD_compressSubBlock_multi has literal entropy tables unwritten");
+        DEBUGLOG(5, "Literal entropy tables were never written");
         ZSTD_memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf));
     }
     if (writeSeqEntropy && ZSTD_needSequenceEntropyTables(&entropyMetadata->fseMetadata)) {
         /* If we haven't written our entropy tables, then we've violated our contract and
          * must emit an uncompressed block.
          */
-        DEBUGLOG(5, "ZSTD_compressSubBlock_multi has sequence entropy tables unwritten");
+        DEBUGLOG(5, "Sequence entropy tables were never written => cancel, emit an uncompressed block");
         return 0;
     }
+
     if (ip < iend) {
-        size_t const cSize = ZSTD_noCompressBlock(op, oend - op, ip, iend - ip, lastBlock);
-        DEBUGLOG(5, "ZSTD_compressSubBlock_multi last sub-block uncompressed, %zu bytes", (size_t)(iend - ip));
+        /* some data left : last part of the block sent uncompressed */
+        size_t const rSize = (size_t)((iend - ip));
+        size_t const cSize = ZSTD_noCompressBlock(op, (size_t)(oend - op), ip, rSize, lastBlock);
+        DEBUGLOG(5, "Generate last uncompressed sub-block of %u bytes", (unsigned)(rSize));
         FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed");
         assert(cSize != 0);
         op += cSize;
         /* We have to regenerate the repcodes because we've skipped some sequences */
         if (sp < send) {
-            seqDef const* seq;
-            repcodes_t rep;
+            const SeqDef* seq;
+            Repcodes_t rep;
             ZSTD_memcpy(&rep, prevCBlock->rep, sizeof(rep));
             for (seq = sstart; seq < sp; ++seq) {
-                ZSTD_updateRep(rep.rep, seq->offBase - 1, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
+                ZSTD_updateRep(rep.rep, seq->offBase, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
             }
             ZSTD_memcpy(nextCBlock->rep, &rep, sizeof(rep));
         }
     }
-    DEBUGLOG(5, "ZSTD_compressSubBlock_multi compressed");
-    return op-ostart;
+
+    DEBUGLOG(5, "ZSTD_compressSubBlock_multi compressed all subBlocks: total compressed size = %u",
+                (unsigned)(op-ostart));
+    return (size_t)(op-ostart);
 }
 
 size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc,
                                void* dst, size_t dstCapacity,
-                               void const* src, size_t srcSize,
-                               unsigned lastBlock) {
+                               const void* src, size_t srcSize,
+                               unsigned lastBlock)
+{
     ZSTD_entropyCTablesMetadata_t entropyMetadata;
 
     FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(&zc->seqStore,
···
           &zc->blockState.nextCBlock->entropy,
           &zc->appliedParams,
           &entropyMetadata,
-          zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */), "");
+          zc->tmpWorkspace, zc->tmpWkspSize /* statically allocated in resetCCtx */), "");
 
     return ZSTD_compressSubBlock_multi(&zc->seqStore,
             zc->blockState.prevCBlock,
···
             dst, dstCapacity,
             src, srcSize,
             zc->bmi2, lastBlock,
-            zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */);
+            zc->tmpWorkspace, zc->tmpWkspSize /* statically allocated in resetCCtx */);
 }

lib/zstd/compress/zstd_compress_superblock.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

lib/zstd/compress/zstd_cwksp.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 /*-*************************************
 *  Dependencies
 ***************************************/
+#include "../common/allocations.h"  /* ZSTD_customMalloc, ZSTD_customFree */
 #include "../common/zstd_internal.h"
-
+#include "../common/portability_macros.h"
+#include "../common/compiler.h" /* ZS2_isPower2 */
 
 /*-*************************************
 *  Constants
···
 ***************************************/
 typedef enum {
     ZSTD_cwksp_alloc_objects,
-    ZSTD_cwksp_alloc_buffers,
-    ZSTD_cwksp_alloc_aligned
+    ZSTD_cwksp_alloc_aligned_init_once,
+    ZSTD_cwksp_alloc_aligned,
+    ZSTD_cwksp_alloc_buffers
 } ZSTD_cwksp_alloc_phase_e;
 
 /*
···
  *
  * Workspace Layout:
  *
- * [                        ... workspace ...                         ]
- * [objects][tables ... ->] free space [<- ... aligned][<- ... buffers]
+ * [                        ... workspace ...                           ]
+ * [objects][tables ->] free space [<- buffers][<- aligned][<- init once]
  *
  * The various objects that live in the workspace are divided into the
  * following categories, and are allocated separately:
···
  *   uint32_t arrays, all of whose values are between 0 and (nextSrc - base).
  *   Their sizes depend on the cparams. These tables are 64-byte aligned.
  *
- * - Aligned: these buffers are used for various purposes that require 4 byte
- *   alignment, but don't require any initialization before they're used. These
- *   buffers are each aligned to 64 bytes.
+ * - Init once: these buffers require to be initialized at least once before
+ *   use. They should be used when we want to skip memory initialization
+ *   while not triggering memory checkers (like Valgrind) when reading from
+ *   from this memory without writing to it first.
+ *   These buffers should be used carefully as they might contain data
+ *   from previous compressions.
+ *   Buffers are aligned to 64 bytes.
+ *
+ * - Aligned: these buffers don't require any initialization before they're
+ *   used. The user of the buffer should make sure they write into a buffer
+ *   location before reading from it.
+ *   Buffers are aligned to 64 bytes.
  *
  * - Buffers: these buffers are used for various purposes that don't require
  *   any alignment or initialization before they're used. This means they can
···
  * correctly packed into the workspace buffer. That order is:
  *
  * 1. Objects
- * 2. Buffers
- * 3. Aligned/Tables
+ * 2. Init once / Tables
+ * 3. Aligned / Tables
+ * 4. Buffers / Tables
  *
  * Attempts to reserve objects of different types out of order will fail.
  */
···
     void* tableEnd;
     void* tableValidEnd;
     void* allocStart;
+    void* initOnceStart;
 
     BYTE allocFailed;
     int workspaceOversizedDuration;
···
 ***************************************/
 
 MEM_STATIC size_t ZSTD_cwksp_available_space(ZSTD_cwksp* ws);
+MEM_STATIC void*  ZSTD_cwksp_initialAllocStart(ZSTD_cwksp* ws);
 
 MEM_STATIC void ZSTD_cwksp_assert_internal_consistency(ZSTD_cwksp* ws) {
     (void)ws;
···
     assert(ws->tableEnd <= ws->allocStart);
     assert(ws->tableValidEnd <= ws->allocStart);
     assert(ws->allocStart <= ws->workspaceEnd);
+    assert(ws->initOnceStart <= ZSTD_cwksp_initialAllocStart(ws));
+    assert(ws->workspace <= ws->initOnceStart);
 }
 
 /*
  * Align must be a power of 2.
  */
-MEM_STATIC size_t ZSTD_cwksp_align(size_t size, size_t const align) {
+MEM_STATIC size_t ZSTD_cwksp_align(size_t size, size_t align) {
     size_t const mask = align - 1;
-    assert((align & mask) == 0);
+    assert(ZSTD_isPower2(align));
     return (size + mask) & ~mask;
 }
 
···
  * to figure out how much space you need for the matchState tables. Everything
  * else is though.
  *
- * Do not use for sizing aligned buffers. Instead, use ZSTD_cwksp_aligned_alloc_size().
+ * Do not use for sizing aligned buffers. Instead, use ZSTD_cwksp_aligned64_alloc_size().
  */
 MEM_STATIC size_t ZSTD_cwksp_alloc_size(size_t size) {
     if (size == 0)
···
     return size;
 }
 
+MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size, size_t alignment) {
+    return ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(size, alignment));
+}
+
 /*
  * Returns an adjusted alloc size that is the nearest larger multiple of 64 bytes.
  * Used to determine the number of bytes required for a given "aligned".
  */
-MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size) {
-    return ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(size, ZSTD_CWKSP_ALIGNMENT_BYTES));
+MEM_STATIC size_t ZSTD_cwksp_aligned64_alloc_size(size_t size) {
+    return ZSTD_cwksp_aligned_alloc_size(size, ZSTD_CWKSP_ALIGNMENT_BYTES);
 }
 
 /*
···
  * for internal purposes (currently only alignment).
  */
 MEM_STATIC size_t ZSTD_cwksp_slack_space_required(void) {
-    /* For alignment, the wksp will always allocate an additional n_1=[1, 64] bytes
-     * to align the beginning of tables section, as well as another n_2=[0, 63] bytes
-     * to align the beginning of the aligned section.
-     *
-     * n_1 + n_2 == 64 bytes if the cwksp is freshly allocated, due to tables and
-     * aligneds being sized in multiples of 64 bytes.
+    /* For alignment, the wksp will always allocate an additional 2*ZSTD_CWKSP_ALIGNMENT_BYTES
+     * bytes to align the beginning of tables section and end of buffers;
      */
-    size_t const slackSpace = ZSTD_CWKSP_ALIGNMENT_BYTES;
+    size_t const slackSpace = ZSTD_CWKSP_ALIGNMENT_BYTES * 2;
     return slackSpace;
 }
 
···
 MEM_STATIC size_t ZSTD_cwksp_bytes_to_align_ptr(void* ptr, const size_t alignBytes) {
     size_t const alignBytesMask = alignBytes - 1;
     size_t const bytes = (alignBytes - ((size_t)ptr & (alignBytesMask))) & alignBytesMask;
-    assert((alignBytes & alignBytesMask) == 0);
-    assert(bytes != ZSTD_CWKSP_ALIGNMENT_BYTES);
+    assert(ZSTD_isPower2(alignBytes));
+    assert(bytes < alignBytes);
     return bytes;
+}
+
+/*
+ * Returns the initial value for allocStart which is used to determine the position from
+ * which we can allocate from the end of the workspace.
+ */
+MEM_STATIC void*  ZSTD_cwksp_initialAllocStart(ZSTD_cwksp* ws)
+{
+    char* endPtr = (char*)ws->workspaceEnd;
+    assert(ZSTD_isPower2(ZSTD_CWKSP_ALIGNMENT_BYTES));
+    endPtr = endPtr - ((size_t)endPtr % ZSTD_CWKSP_ALIGNMENT_BYTES);
+    return (void*)endPtr;
 }
 
 /*
···
 {
     void* const alloc = (BYTE*)ws->allocStart - bytes;
     void* const bottom = ws->tableEnd;
-    DEBUGLOG(5, "cwksp: reserving %p %zd bytes, %zd bytes remaining",
+    DEBUGLOG(5, "cwksp: reserving [0x%p]:%zd bytes; %zd bytes remaining",
         alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
     ZSTD_cwksp_assert_internal_consistency(ws);
     assert(alloc >= bottom);
···
 {
     assert(phase >= ws->phase);
     if (phase > ws->phase) {
-        /* Going from allocating objects to allocating buffers */
-        if (ws->phase < ZSTD_cwksp_alloc_buffers &&
-                phase >= ZSTD_cwksp_alloc_buffers) {
+        /* Going from allocating objects to allocating initOnce / tables */
+        if (ws->phase < ZSTD_cwksp_alloc_aligned_init_once &&
+            phase >= ZSTD_cwksp_alloc_aligned_init_once) {
             ws->tableValidEnd = ws->objectEnd;
-        }
+            ws->initOnceStart = ZSTD_cwksp_initialAllocStart(ws);
 
-        /* Going from allocating buffers to allocating aligneds/tables */
-        if (ws->phase < ZSTD_cwksp_alloc_aligned &&
-                phase >= ZSTD_cwksp_alloc_aligned) {
-            {   /* Align the start of the "aligned" to 64 bytes. Use [1, 64] bytes. */
-                size_t const bytesToAlign =
-                    ZSTD_CWKSP_ALIGNMENT_BYTES - ZSTD_cwksp_bytes_to_align_ptr(ws->allocStart, ZSTD_CWKSP_ALIGNMENT_BYTES);
-                DEBUGLOG(5, "reserving aligned alignment addtl space: %zu", bytesToAlign);
-                ZSTD_STATIC_ASSERT((ZSTD_CWKSP_ALIGNMENT_BYTES & (ZSTD_CWKSP_ALIGNMENT_BYTES - 1)) == 0); /* power of 2 */
-                RETURN_ERROR_IF(!ZSTD_cwksp_reserve_internal_buffer_space(ws, bytesToAlign),
-                                memory_allocation, "aligned phase - alignment initial allocation failed!");
-            }
             {   /* Align the start of the tables to 64 bytes. Use [0, 63] bytes */
-                void* const alloc = ws->objectEnd;
+                void *const alloc = ws->objectEnd;
                 size_t const bytesToAlign = ZSTD_cwksp_bytes_to_align_ptr(alloc, ZSTD_CWKSP_ALIGNMENT_BYTES);
-                void* const objectEnd = (BYTE*)alloc + bytesToAlign;
+                void *const objectEnd = (BYTE *) alloc + bytesToAlign;
                 DEBUGLOG(5, "reserving table alignment addtl space: %zu", bytesToAlign);
                 RETURN_ERROR_IF(objectEnd > ws->workspaceEnd, memory_allocation,
                                 "table phase - alignment initial allocation failed!");
···
                 ws->tableEnd = objectEnd;  /* table area starts being empty */
                 if (ws->tableValidEnd < ws->tableEnd) {
                     ws->tableValidEnd = ws->tableEnd;
-        }   }   }
+                }
+            }
+        }
         ws->phase = phase;
         ZSTD_cwksp_assert_internal_consistency(ws);
     }
···
  */
 MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr)
 {
-    return (ptr != NULL) && (ws->workspace <= ptr) && (ptr <= ws->workspaceEnd);
+    return (ptr != NULL) && (ws->workspace <= ptr) && (ptr < ws->workspaceEnd);
 }
 
 /*
···
 
 /*
  * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
+ * This memory has been initialized at least once in the past.
+ * This doesn't mean it has been initialized this time, and it might contain data from previous
+ * operations.
+ * The main usage is for algorithms that might need read access into uninitialized memory.
+ * The algorithm must maintain safety under these conditions and must make sure it doesn't
+ * leak any of the past data (directly or in side channels).
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes)
+MEM_STATIC void* ZSTD_cwksp_reserve_aligned_init_once(ZSTD_cwksp* ws, size_t bytes)
 {
-    void* ptr = ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES),
-                                            ZSTD_cwksp_alloc_aligned);
-    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
+    size_t const alignedBytes = ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES);
+    void* ptr = ZSTD_cwksp_reserve_internal(ws, alignedBytes, ZSTD_cwksp_alloc_aligned_init_once);
+    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
+    if(ptr && ptr < ws->initOnceStart) {
+        /* We assume the memory following the current allocation is either:
+         * 1. Not usable as initOnce memory (end of workspace)
+         * 2. Another initOnce buffer that has been allocated before (and so was previously memset)
+         * 3. An ASAN redzone, in which case we don't want to write on it
+         * For these reasons it should be fine to not explicitly zero every byte up to ws->initOnceStart.
+         * Note that we assume here that MSAN and ASAN cannot run in the same time. */
+        ZSTD_memset(ptr, 0, MIN((size_t)((U8*)ws->initOnceStart - (U8*)ptr), alignedBytes));
+        ws->initOnceStart = ptr;
+    }
+    return ptr;
+}
+
+/*
+ * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
+ */
+MEM_STATIC void* ZSTD_cwksp_reserve_aligned64(ZSTD_cwksp* ws, size_t bytes)
+{
+    void* const ptr = ZSTD_cwksp_reserve_internal(ws,
+                        ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES),
+                        ZSTD_cwksp_alloc_aligned);
+    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
     return ptr;
 }
 
 /*
  * Aligned on 64 bytes. These buffers have the special property that
- * their values remain constrained, allowing us to re-use them without
+ * their values remain constrained, allowing us to reuse them without
  * memset()-ing them.
  */
 MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes)
 {
-    const ZSTD_cwksp_alloc_phase_e phase = ZSTD_cwksp_alloc_aligned;
+    const ZSTD_cwksp_alloc_phase_e phase = ZSTD_cwksp_alloc_aligned_init_once;
     void* alloc;
     void* end;
     void* top;
 
-    if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase))) {
-        return NULL;
+    /* We can only start allocating tables after we are done reserving space for objects at the
+     * start of the workspace */
+    if(ws->phase < phase) {
+        if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase))) {
+            return NULL;
+        }
     }
     alloc = ws->tableEnd;
     end = (BYTE *)alloc + bytes;
···
 
 
     assert((bytes & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
-    assert(((size_t)alloc & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
+    assert(((size_t)alloc & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
     return alloc;
 }
 
···
 
     return alloc;
 }
+/*
+ * with alignment control
+ * Note : should happen only once, at workspace first initialization
+ */
+MEM_STATIC void* ZSTD_cwksp_reserve_object_aligned(ZSTD_cwksp* ws, size_t byteSize, size_t alignment)
+{
+    size_t const mask = alignment - 1;
+    size_t const surplus = (alignment > sizeof(void*)) ? alignment - sizeof(void*) : 0;
+    void* const start = ZSTD_cwksp_reserve_object(ws, byteSize + surplus);
+    if (start == NULL) return NULL;
+    if (surplus == 0) return start;
+    assert(ZSTD_isPower2(alignment));
+    return (void*)(((size_t)start + surplus) & ~mask);
+}
 
 MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws)
 {
···
     assert(ws->tableValidEnd >= ws->objectEnd);
     assert(ws->tableValidEnd <= ws->allocStart);
     if (ws->tableValidEnd < ws->tableEnd) {
-        ZSTD_memset(ws->tableValidEnd, 0, (BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd);
+        ZSTD_memset(ws->tableValidEnd, 0, (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd));
     }
     ZSTD_cwksp_mark_tables_clean(ws);
 }
···
  * Invalidates table allocations.
  * All other allocations remain valid.
  */
-MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws) {
+MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws)
+{
     DEBUGLOG(4, "cwksp: clearing tables!");
 
 
···
 
 
     ws->tableEnd = ws->objectEnd;
-    ws->allocStart = ws->workspaceEnd;
+    ws->allocStart = ZSTD_cwksp_initialAllocStart(ws);
     ws->allocFailed = 0;
-    if (ws->phase > ZSTD_cwksp_alloc_buffers) {
-        ws->phase = ZSTD_cwksp_alloc_buffers;
+    if (ws->phase > ZSTD_cwksp_alloc_aligned_init_once) {
+        ws->phase = ZSTD_cwksp_alloc_aligned_init_once;
     }
     ZSTD_cwksp_assert_internal_consistency(ws);
+}
+
+MEM_STATIC size_t ZSTD_cwksp_sizeof(const ZSTD_cwksp* ws) {
+    return (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->workspace);
+}
+
+MEM_STATIC size_t ZSTD_cwksp_used(const ZSTD_cwksp* ws) {
+    return (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->workspace)
+         + (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->allocStart);
 }
 
 /*
···
     ws->workspaceEnd = (BYTE*)start + size;
     ws->objectEnd = ws->workspace;
     ws->tableValidEnd = ws->objectEnd;
+    ws->initOnceStart = ZSTD_cwksp_initialAllocStart(ws);
     ws->phase = ZSTD_cwksp_alloc_objects;
     ws->isStatic = isStatic;
     ZSTD_cwksp_clear(ws);
···
     ZSTD_memset(src, 0, sizeof(ZSTD_cwksp));
 }
 
-MEM_STATIC size_t ZSTD_cwksp_sizeof(const ZSTD_cwksp* ws) {
-    return (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->workspace);
-}
-
-MEM_STATIC size_t ZSTD_cwksp_used(const ZSTD_cwksp* ws) {
-    return (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->workspace)
-         + (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->allocStart);
-}
-
 MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) {
     return ws->allocFailed;
 }
···
  * Returns if the estimated space needed for a wksp is within an acceptable limit of the
  * actual amount of space used.
  */
-MEM_STATIC int ZSTD_cwksp_estimated_space_within_bounds(const ZSTD_cwksp* const ws,
-                                                        size_t const estimatedSpace, int resizedWorkspace) {
-    if (resizedWorkspace) {
-        /* Resized/newly allocated wksp should have exact bounds */
-        return ZSTD_cwksp_used(ws) == estimatedSpace;
-    } else {
-        /* Due to alignment, when reusing a workspace, we can actually consume 63 fewer or more bytes
-         * than estimatedSpace. See the comments in zstd_cwksp.h for details.
-         */
-        return (ZSTD_cwksp_used(ws) >= estimatedSpace - 63) && (ZSTD_cwksp_used(ws) <= estimatedSpace + 63);
-    }
+MEM_STATIC int ZSTD_cwksp_estimated_space_within_bounds(const ZSTD_cwksp *const ws, size_t const estimatedSpace) {
+    /* We have an alignment space between objects and tables between tables and buffers, so we can have up to twice
+     * the alignment bytes difference between estimation and actual usage */
+    return (estimatedSpace - ZSTD_cwksp_slack_space_required()) <= ZSTD_cwksp_used(ws) &&
+           ZSTD_cwksp_used(ws) <= estimatedSpace;
 }
 
 
···
         ws->workspaceOversizedDuration = 0;
     }
 }
-
 
 #endif /* ZSTD_CWKSP_H */

lib/zstd/compress/zstd_double_fast.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #include "zstd_compress_internal.h"
 #include "zstd_double_fast.h"
 
+#ifndef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR
 
-void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_fillDoubleHashTableForCDict(ZSTD_MatchState_t* ms,
+                              void const* end, ZSTD_dictTableLoadMethod_e dtlm)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashLarge = ms->hashTable;
+    U32  const hBitsL = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    U32  const mls = cParams->minMatch;
+    U32* const hashSmall = ms->chainTable;
+    U32  const hBitsS = cParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
+
+    /* Always insert every fastHashFillStep position into the hash tables.
+     * Insert the other positions into the large hash table if their entry
+     * is empty.
+     */
+    for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
+        U32 const curr = (U32)(ip - base);
+        U32 i;
+        for (i = 0; i < fastHashFillStep; ++i) {
+            size_t const smHashAndTag = ZSTD_hashPtr(ip + i, hBitsS, mls);
+            size_t const lgHashAndTag = ZSTD_hashPtr(ip + i, hBitsL, 8);
+            if (i == 0) {
+                ZSTD_writeTaggedIndex(hashSmall, smHashAndTag, curr + i);
+            }
+            if (i == 0 || hashLarge[lgHashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {
+                ZSTD_writeTaggedIndex(hashLarge, lgHashAndTag, curr + i);
+            }
+            /* Only load extra positions for ZSTD_dtlm_full */
+            if (dtlm == ZSTD_dtlm_fast)
+                break;
+    }   }
+}
+
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_fillDoubleHashTableForCCtx(ZSTD_MatchState_t* ms,
                               void const* end, ZSTD_dictTableLoadMethod_e dtlm)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
···
             /* Only load extra positions for ZSTD_dtlm_full */
             if (dtlm == ZSTD_dtlm_fast)
                 break;
-    }   }
+        }   }
+}
+
+void ZSTD_fillDoubleHashTable(ZSTD_MatchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp)
+{
+    if (tfp == ZSTD_tfp_forCDict) {
+        ZSTD_fillDoubleHashTableForCDict(ms, end, dtlm);
+    } else {
+        ZSTD_fillDoubleHashTableForCCtx(ms, end, dtlm);
+    }
 }
 
 
 FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_compressBlock_doubleFast_noDict_generic(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize, U32 const mls /* template */)
 {
     ZSTD_compressionParameters const* cParams = &ms->cParams;
···
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
     U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
     size_t mLength;
     U32 offset;
···
     const BYTE* matchl0; /* the long match for ip */
     const BYTE* matchs0; /* the short match for ip */
     const BYTE* matchl1; /* the long match for ip1 */
+    const BYTE* matchs0_safe; /* matchs0 or safe address */
 
     const BYTE* ip = istart; /* the current position */
     const BYTE* ip1; /* the next position */
+    /* Array of ~random data, should have low probability of matching data
+     * we load from here instead of from tables, if matchl0/matchl1 are
+     * invalid indices. Used to avoid unpredictable branches. */
+    const BYTE dummy[] = {0x12,0x34,0x56,0x78,0x9a,0xbc,0xde,0xf0,0xe2,0xb4};
 
     DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_noDict_generic");
 
···
         U32 const current = (U32)(ip - base);
         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
         U32 const maxRep = current - windowLow;
-        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
+        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
     }
 
     /* Outer Loop: one iteration per match found and stored */
···
             if ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1))) {
                 mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
                 ip++;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
                 goto _match_stored;
             }
 
             hl1 = ZSTD_hashPtr(ip1, hBitsL, 8);
 
-            if (idxl0 > prefixLowestIndex) {
+            /* idxl0 > prefixLowestIndex is a (somewhat) unpredictable branch.
+             * However expression below complies into conditional move. Since
+             * match is unlikely and we only *branch* on idxl0 > prefixLowestIndex
+             * if there is a match, all branches become predictable. */
+            {   const BYTE*  const matchl0_safe = ZSTD_selectAddr(idxl0, prefixLowestIndex, matchl0, &dummy[0]);
+
                 /* check prefix long match */
-                if (MEM_read64(matchl0) == MEM_read64(ip)) {
+                if (MEM_read64(matchl0_safe) == MEM_read64(ip) && matchl0_safe == matchl0) {
                     mLength = ZSTD_count(ip+8, matchl0+8, iend) + 8;
                     offset = (U32)(ip-matchl0);
                     while (((ip>anchor) & (matchl0>prefixLowest)) && (ip[-1] == matchl0[-1])) { ip--; matchl0--; mLength++; } /* catch up */
                     goto _match_found;
-                }
-            }
+            }   }
 
             idxl1 = hashLong[hl1];
             matchl1 = base + idxl1;
 
-            if (idxs0 > prefixLowestIndex) {
-                /* check prefix short match */
-                if (MEM_read32(matchs0) == MEM_read32(ip)) {
-                    goto _search_next_long;
-                }
+            /* Same optimization as matchl0 above */
+            matchs0_safe = ZSTD_selectAddr(idxs0, prefixLowestIndex, matchs0, &dummy[0]);
+
+            /* check prefix short match */
+            if(MEM_read32(matchs0_safe) == MEM_read32(ip) && matchs0_safe == matchs0) {
+                  goto _search_next_long;
             }
 
             if (ip1 >= nextStep) {
···
         } while (ip1 <= ilimit);
 
 _cleanup:
+        /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+         * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+        offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
+
         /* save reps for next block */
-        rep[0] = offset_1 ? offset_1 : offsetSaved;
-        rep[1] = offset_2 ? offset_2 : offsetSaved;
+        rep[0] = offset_1 ? offset_1 : offsetSaved1;
+        rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
         /* Return the last literals size */
         return (size_t)(iend - anchor);
 
 _search_next_long:
 
-        /* check prefix long +1 match */
-        if (idxl1 > prefixLowestIndex) {
-            if (MEM_read64(matchl1) == MEM_read64(ip1)) {
+        /* short match found: let's check for a longer one */
+        mLength = ZSTD_count(ip+4, matchs0+4, iend) + 4;
+        offset = (U32)(ip - matchs0);
+
+        /* check long match at +1 position */
+        if ((idxl1 > prefixLowestIndex) && (MEM_read64(matchl1) == MEM_read64(ip1))) {
+            size_t const l1len = ZSTD_count(ip1+8, matchl1+8, iend) + 8;
+            if (l1len > mLength) {
+                /* use the long match instead */
                 ip = ip1;
-                mLength = ZSTD_count(ip+8, matchl1+8, iend) + 8;
+                mLength = l1len;
                 offset = (U32)(ip-matchl1);
-                while (((ip>anchor) & (matchl1>prefixLowest)) && (ip[-1] == matchl1[-1])) { ip--; matchl1--; mLength++; } /* catch up */
-                goto _match_found;
+                matchs0 = matchl1;
             }
         }
 
-        /* if no long +1 match, explore the short match we found */
-        mLength = ZSTD_count(ip+4, matchs0+4, iend) + 4;
-        offset = (U32)(ip - matchs0);
-        while (((ip>anchor) & (matchs0>prefixLowest)) && (ip[-1] == matchs0[-1])) { ip--; matchs0--; mLength++; } /* catch up */
+        while (((ip>anchor) & (matchs0>prefixLowest)) && (ip[-1] == matchs0[-1])) { ip--; matchs0--; mLength++; } /* complete backward */
 
         /* fall-through */
 
···
             hashLong[hl1] = (U32)(ip1 - base);
         }
 
-        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
 _match_stored:
         /* match found */
···
                 U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff;  /* swap offset_2 <=> offset_1 */
                 hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
                 hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, rLength);
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
                 ip += rLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
···
 
 
 FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
         U32 const mls /* template */)
 {
···
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
     U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
 
-    const ZSTD_matchState_t* const dms = ms->dictMatchState;
+    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
     const ZSTD_compressionParameters* const dictCParams = &dms->cParams;
     const U32* const dictHashLong  = dms->hashTable;
     const U32* const dictHashSmall = dms->chainTable;
···
     const BYTE* const dictStart    = dictBase + dictStartIndex;
     const BYTE* const dictEnd      = dms->window.nextSrc;
     const U32 dictIndexDelta       = prefixLowestIndex - (U32)(dictEnd - dictBase);
-    const U32 dictHBitsL           = dictCParams->hashLog;
-    const U32 dictHBitsS           = dictCParams->chainLog;
+    const U32 dictHBitsL           = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    const U32 dictHBitsS           = dictCParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
     const U32 dictAndPrefixLength  = (U32)((ip - prefixLowest) + (dictEnd - dictStart));
 
     DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_dictMatchState_generic");
 
     /* if a dictionary is attached, it must be within window range */
     assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex);
+
+    if (ms->prefetchCDictTables) {
+        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
+        size_t const chainTableBytes = (((size_t)1) << dictCParams->chainLog) * sizeof(U32);
+        PREFETCH_AREA(dictHashLong, hashTableBytes);
+        PREFETCH_AREA(dictHashSmall, chainTableBytes);
+    }
 
     /* init */
     ip += (dictAndPrefixLength == 0);
···
         U32 offset;
         size_t const h2 = ZSTD_hashPtr(ip, hBitsL, 8);
         size_t const h = ZSTD_hashPtr(ip, hBitsS, mls);
-        size_t const dictHL = ZSTD_hashPtr(ip, dictHBitsL, 8);
-        size_t const dictHS = ZSTD_hashPtr(ip, dictHBitsS, mls);
+        size_t const dictHashAndTagL = ZSTD_hashPtr(ip, dictHBitsL, 8);
+        size_t const dictHashAndTagS = ZSTD_hashPtr(ip, dictHBitsS, mls);
+        U32 const dictMatchIndexAndTagL = dictHashLong[dictHashAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS];
+        U32 const dictMatchIndexAndTagS = dictHashSmall[dictHashAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS];
+        int const dictTagsMatchL = ZSTD_comparePackedTags(dictMatchIndexAndTagL, dictHashAndTagL);
+        int const dictTagsMatchS = ZSTD_comparePackedTags(dictMatchIndexAndTagS, dictHashAndTagS);
         U32 const curr = (U32)(ip-base);
         U32 const matchIndexL = hashLong[h2];
         U32 matchIndexS = hashSmall[h];
···
         hashLong[h2] = hashSmall[h] = curr;   /* update hash tables */
 
         /* check repcode */
-        if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
+        if ((ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
             && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
             ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength);
+            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
             goto _match_stored;
         }
 
-        if (matchIndexL > prefixLowestIndex) {
+        if ((matchIndexL >= prefixLowestIndex) && (MEM_read64(matchLong) == MEM_read64(ip))) {
             /* check prefix long match */
-            if (MEM_read64(matchLong) == MEM_read64(ip)) {
-                mLength = ZSTD_count(ip+8, matchLong+8, iend) + 8;
-                offset = (U32)(ip-matchLong);
-                while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
-                goto _match_found;
-            }
-        } else {
+            mLength = ZSTD_count(ip+8, matchLong+8, iend) + 8;
+            offset = (U32)(ip-matchLong);
+            while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
+            goto _match_found;
+        } else if (dictTagsMatchL) {
             /* check dictMatchState long match */
-            U32 const dictMatchIndexL = dictHashLong[dictHL];
+            U32 const dictMatchIndexL = dictMatchIndexAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS;
             const BYTE* dictMatchL = dictBase + dictMatchIndexL;
             assert(dictMatchL < dictEnd);
 
···
         }   }
 
         if (matchIndexS > prefixLowestIndex) {
-            /* check prefix short match */
+            /* short match  candidate */
             if (MEM_read32(match) == MEM_read32(ip)) {
                 goto _search_next_long;
             }
-        } else {
+        } else if (dictTagsMatchS) {
             /* check dictMatchState short match */
-            U32 const dictMatchIndexS = dictHashSmall[dictHS];
+            U32 const dictMatchIndexS = dictMatchIndexAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS;
             match = dictBase + dictMatchIndexS;
             matchIndexS = dictMatchIndexS + dictIndexDelta;
 
···
         continue;
 
 _search_next_long:
-
         {   size_t const hl3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
-            size_t const dictHLNext = ZSTD_hashPtr(ip+1, dictHBitsL, 8);
+            size_t const dictHashAndTagL3 = ZSTD_hashPtr(ip+1, dictHBitsL, 8);
             U32 const matchIndexL3 = hashLong[hl3];
+            U32 const dictMatchIndexAndTagL3 = dictHashLong[dictHashAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS];
+            int const dictTagsMatchL3 = ZSTD_comparePackedTags(dictMatchIndexAndTagL3, dictHashAndTagL3);
             const BYTE* matchL3 = base + matchIndexL3;
             hashLong[hl3] = curr + 1;
 
             /* check prefix long +1 match */
-            if (matchIndexL3 > prefixLowestIndex) {
-                if (MEM_read64(matchL3) == MEM_read64(ip+1)) {
-                    mLength = ZSTD_count(ip+9, matchL3+8, iend) + 8;
-                    ip++;
-                    offset = (U32)(ip-matchL3);
-                    while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
-                    goto _match_found;
-                }
-            } else {
+            if ((matchIndexL3 >= prefixLowestIndex) && (MEM_read64(matchL3) == MEM_read64(ip+1))) {
+                mLength = ZSTD_count(ip+9, matchL3+8, iend) + 8;
+                ip++;
+                offset = (U32)(ip-matchL3);
+                while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
+                goto _match_found;
+            } else if (dictTagsMatchL3) {
                 /* check dict long +1 match */
-                U32 const dictMatchIndexL3 = dictHashLong[dictHLNext];
+                U32 const dictMatchIndexL3 = dictMatchIndexAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS;
                 const BYTE* dictMatchL3 = dictBase + dictMatchIndexL3;
                 assert(dictMatchL3 < dictEnd);
                 if (dictMatchL3 > dictStart && MEM_read64(dictMatchL3) == MEM_read64(ip+1)) {
···
         offset_2 = offset_1;
         offset_1 = offset;
 
-        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
 _match_stored:
         /* match found */
···
                 const BYTE* repMatch2 = repIndex2 < prefixLowestIndex ?
                         dictBase + repIndex2 - dictIndexDelta :
                         base + repIndex2;
-                if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
+                if ( (ZSTD_index_overlap_check(prefixLowestIndex, repIndex2))
                    && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex2 < prefixLowestIndex ? dictEnd : iend;
                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixLowest) + 4;
                     U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, repLength2);
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                     hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
                     hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
                     ip += repLength2;
···
     }   /* while (ip < ilimit) */
 
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[0] = offset_1;
+    rep[1] = offset_2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
···
 
 #define ZSTD_GEN_DFAST_FN(dictMode, mls)                                                                 \
     static size_t ZSTD_compressBlock_doubleFast_##dictMode##_##mls(                                      \
-            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                          \
+            ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                          \
             void const* src, size_t srcSize)                                                             \
     {                                                                                                    \
         return ZSTD_compressBlock_doubleFast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls); \
···
 
 
 size_t ZSTD_compressBlock_doubleFast(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     const U32 mls = ms->cParams.minMatch;
···
 
 
 size_t ZSTD_compressBlock_doubleFast_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     const U32 mls = ms->cParams.minMatch;
···
 }
 
 
-static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_compressBlock_doubleFast_extDict_generic(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
         U32 const mls /* template */)
 {
···
         size_t mLength;
         hashSmall[hSmall] = hashLong[hLong] = curr;   /* update hash table */
 
-        if ((((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex doesn't overlap dict + prefix */
+        if (((ZSTD_index_overlap_check(prefixStartIndex, repIndex))
             & (offset_1 <= curr+1 - dictStartIndex)) /* note: we are searching at curr+1 */
           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
             ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength);
+            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
         } else {
             if ((matchLongIndex > dictStartIndex) && (MEM_read64(matchLong) == MEM_read64(ip))) {
                 const BYTE* const matchEnd = matchLongIndex < prefixStartIndex ? dictEnd : iend;
···
                 while (((ip>anchor) & (matchLong>lowMatchPtr)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; }   /* catch up */
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
             } else if ((matchIndex > dictStartIndex) && (MEM_read32(match) == MEM_read32(ip))) {
                 size_t const h3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
···
                 }
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
             } else {
                 ip += ((ip-anchor) >> kSearchStrength) + 1;
···
                 U32 const current2 = (U32)(ip-base);
                 U32 const repIndex2 = current2 - offset_2;
                 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
-                if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3)   /* intentional overflow : ensure repIndex2 doesn't overlap dict + prefix */
+                if ( ((ZSTD_index_overlap_check(prefixStartIndex, repIndex2))
                     & (offset_2 <= current2 - dictStartIndex))
                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                     U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, repLength2);
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                     hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
                     hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
                     ip += repLength2;
···
 ZSTD_GEN_DFAST_FN(extDict, 7)
 
 size_t ZSTD_compressBlock_doubleFast_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     U32 const mls = ms->cParams.minMatch;
···
         return ZSTD_compressBlock_doubleFast_extDict_7(ms, seqStore, rep, src, srcSize);
     }
 }
+
+#endif /* ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR */

lib/zstd/compress/zstd_double_fast.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #ifndef ZSTD_DOUBLE_FAST_H
 #define ZSTD_DOUBLE_FAST_H
 
-
 #include "../common/mem.h"      /* U32 */
 #include "zstd_compress_internal.h"     /* ZSTD_CCtx, size_t */
 
-void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
-                              void const* end, ZSTD_dictTableLoadMethod_e dtlm);
+#ifndef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR
+
+void ZSTD_fillDoubleHashTable(ZSTD_MatchState_t* ms,
+                              void const* end, ZSTD_dictTableLoadMethod_e dtlm,
+                              ZSTD_tableFillPurpose_e tfp);
+
 size_t ZSTD_compressBlock_doubleFast(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_doubleFast_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_doubleFast_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 
-
+#define ZSTD_COMPRESSBLOCK_DOUBLEFAST ZSTD_compressBlock_doubleFast
+#define ZSTD_COMPRESSBLOCK_DOUBLEFAST_DICTMATCHSTATE ZSTD_compressBlock_doubleFast_dictMatchState
+#define ZSTD_COMPRESSBLOCK_DOUBLEFAST_EXTDICT ZSTD_compressBlock_doubleFast_extDict
+#else
+#define ZSTD_COMPRESSBLOCK_DOUBLEFAST NULL
+#define ZSTD_COMPRESSBLOCK_DOUBLEFAST_DICTMATCHSTATE NULL
+#define ZSTD_COMPRESSBLOCK_DOUBLEFAST_EXTDICT NULL
+#endif /* ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR */
 
 #endif /* ZSTD_DOUBLE_FAST_H */

lib/zstd/compress/zstd_fast.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #include "zstd_compress_internal.h"  /* ZSTD_hashPtr, ZSTD_count, ZSTD_storeSeq */
 #include "zstd_fast.h"
 
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_fillHashTableForCDict(ZSTD_MatchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashTable = ms->hashTable;
+    U32  const hBits = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    U32  const mls = cParams->minMatch;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
 
-void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+    /* Currently, we always use ZSTD_dtlm_full for filling CDict tables.
+     * Feel free to remove this assert if there's a good reason! */
+    assert(dtlm == ZSTD_dtlm_full);
+
+    /* Always insert every fastHashFillStep position into the hash table.
+     * Insert the other positions if their hash entry is empty.
+     */
+    for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
+        U32 const curr = (U32)(ip - base);
+        {   size_t const hashAndTag = ZSTD_hashPtr(ip, hBits, mls);
+            ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr);   }
+
+        if (dtlm == ZSTD_dtlm_fast) continue;
+        /* Only load extra positions for ZSTD_dtlm_full */
+        {   U32 p;
+            for (p = 1; p < fastHashFillStep; ++p) {
+                size_t const hashAndTag = ZSTD_hashPtr(ip + p, hBits, mls);
+                if (hashTable[hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {  /* not yet filled */
+                    ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr + p);
+    }   }   }   }
+}
+
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_fillHashTableForCCtx(ZSTD_MatchState_t* ms,
                         const void* const end,
                         ZSTD_dictTableLoadMethod_e dtlm)
 {
···
     const BYTE* ip = base + ms->nextToUpdate;
     const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
     const U32 fastHashFillStep = 3;
+
+    /* Currently, we always use ZSTD_dtlm_fast for filling CCtx tables.
+     * Feel free to remove this assert if there's a good reason! */
+    assert(dtlm == ZSTD_dtlm_fast);
 
     /* Always insert every fastHashFillStep position into the hash table.
      * Insert the other positions if their hash entry is empty.
···
                 if (hashTable[hash] == 0) {  /* not yet filled */
                     hashTable[hash] = curr + p;
     }   }   }   }
+}
+
+void ZSTD_fillHashTable(ZSTD_MatchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp)
+{
+    if (tfp == ZSTD_tfp_forCDict) {
+        ZSTD_fillHashTableForCDict(ms, end, dtlm);
+    } else {
+        ZSTD_fillHashTableForCCtx(ms, end, dtlm);
+    }
+}
+
+
+typedef int (*ZSTD_match4Found) (const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit);
+
+static int
+ZSTD_match4Found_cmov(const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit)
+{
+    /* Array of ~random data, should have low probability of matching data.
+     * Load from here if the index is invalid.
+     * Used to avoid unpredictable branches. */
+    static const BYTE dummy[] = {0x12,0x34,0x56,0x78};
+
+    /* currentIdx >= lowLimit is a (somewhat) unpredictable branch.
+     * However expression below compiles into conditional move.
+     */
+    const BYTE* mvalAddr = ZSTD_selectAddr(matchIdx, idxLowLimit, matchAddress, dummy);
+    /* Note: this used to be written as : return test1 && test2;
+     * Unfortunately, once inlined, these tests become branches,
+     * in which case it becomes critical that they are executed in the right order (test1 then test2).
+     * So we have to write these tests in a specific manner to ensure their ordering.
+     */
+    if (MEM_read32(currentPtr) != MEM_read32(mvalAddr)) return 0;
+    /* force ordering of these tests, which matters once the function is inlined, as they become branches */
+    __asm__("");
+    return matchIdx >= idxLowLimit;
+}
+
+static int
+ZSTD_match4Found_branch(const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit)
+{
+    /* using a branch instead of a cmov,
+     * because it's faster in scenarios where matchIdx >= idxLowLimit is generally true,
+     * aka almost all candidates are within range */
+    U32 mval;
+    if (matchIdx >= idxLowLimit) {
+        mval = MEM_read32(matchAddress);
+    } else {
+        mval = MEM_read32(currentPtr) ^ 1; /* guaranteed to not match. */
+    }
+
+    return (MEM_read32(currentPtr) == mval);
 }
 
 
···
  *
  * This is also the work we do at the beginning to enter the loop initially.
  */
-FORCE_INLINE_TEMPLATE size_t
-ZSTD_compressBlock_fast_noDict_generic(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_compressBlock_fast_noDict_generic(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
-        U32 const mls, U32 const hasStep)
+        U32 const mls, int useCmov)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
     U32 const hlog = cParams->hashLog;
-    /* support stepSize of 0 */
-    size_t const stepSize = hasStep ? (cParams->targetLength + !(cParams->targetLength) + 1) : 2;
+    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1; /* min 2 */
     const BYTE* const base = ms->window.base;
     const BYTE* const istart = (const BYTE*)src;
     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
···
 
     U32 rep_offset1 = rep[0];
     U32 rep_offset2 = rep[1];
-    U32 offsetSaved = 0;
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
     size_t hash0; /* hash for ip0 */
     size_t hash1; /* hash for ip1 */
-    U32 idx; /* match idx for ip0 */
-    U32 mval; /* src value at match idx */
+    U32 matchIdx; /* match idx for ip0 */
 
     U32 offcode;
     const BYTE* match0;
···
     size_t step;
     const BYTE* nextStep;
     const size_t kStepIncr = (1 << (kSearchStrength - 1));
+    const ZSTD_match4Found matchFound = useCmov ? ZSTD_match4Found_cmov : ZSTD_match4Found_branch;
 
     DEBUGLOG(5, "ZSTD_compressBlock_fast_generic");
     ip0 += (ip0 == prefixStart);
     {   U32 const curr = (U32)(ip0 - base);
         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
         U32 const maxRep = curr - windowLow;
-        if (rep_offset2 > maxRep) offsetSaved = rep_offset2, rep_offset2 = 0;
-        if (rep_offset1 > maxRep) offsetSaved = rep_offset1, rep_offset1 = 0;
+        if (rep_offset2 > maxRep) offsetSaved2 = rep_offset2, rep_offset2 = 0;
+        if (rep_offset1 > maxRep) offsetSaved1 = rep_offset1, rep_offset1 = 0;
+    }
+
+    /* start each op */
+_start: /* Requires: ip0 */
+
+    step = stepSize;
+    nextStep = ip0 + kStepIncr;
+
+    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
+    ip1 = ip0 + 1;
+    ip2 = ip0 + step;
+    ip3 = ip2 + 1;
+
+    if (ip3 >= ilimit) {
+        goto _cleanup;
+    }
+
+    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+    hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+
+    matchIdx = hashTable[hash0];
+
+    do {
+        /* load repcode match for ip[2]*/
+        const U32 rval = MEM_read32(ip2 - rep_offset1);
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        /* check repcode at ip[2] */
+        if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) {
+            ip0 = ip2;
+            match0 = ip0 - rep_offset1;
+            mLength = ip0[-1] == match0[-1];
+            ip0 -= mLength;
+            match0 -= mLength;
+            offcode = REPCODE1_TO_OFFBASE;
+            mLength += 4;
+
+            /* Write next hash table entry: it's already calculated.
+             * This write is known to be safe because ip1 is before the
+             * repcode (ip2). */
+            hashTable[hash1] = (U32)(ip1 - base);
+
+            goto _match;
+        }
+
+         if (matchFound(ip0, base + matchIdx, matchIdx, prefixStartIndex)) {
+            /* Write next hash table entry (it's already calculated).
+            * This write is known to be safe because the ip1 == ip0 + 1,
+            * so searching will resume after ip1 */
+            hashTable[hash1] = (U32)(ip1 - base);
+
+            goto _offset;
+        }
+
+        /* lookup ip[1] */
+        matchIdx = hashTable[hash1];
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip3;
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+         if (matchFound(ip0, base + matchIdx, matchIdx, prefixStartIndex)) {
+            /* Write next hash table entry, since it's already calculated */
+            if (step <= 4) {
+                /* Avoid writing an index if it's >= position where search will resume.
+                * The minimum possible match has length 4, so search can resume at ip0 + 4.
+                */
+                hashTable[hash1] = (U32)(ip1 - base);
+            }
+            goto _offset;
+        }
+
+        /* lookup ip[1] */
+        matchIdx = hashTable[hash1];
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip0 + step;
+        ip3 = ip1 + step;
+
+        /* calculate step */
+        if (ip2 >= nextStep) {
+            step++;
+            PREFETCH_L1(ip1 + 64);
+            PREFETCH_L1(ip1 + 128);
+            nextStep += kStepIncr;
+        }
+    } while (ip3 < ilimit);
+
+_cleanup:
+    /* Note that there are probably still a couple positions one could search.
+     * However, it seems to be a meaningful performance hit to try to search
+     * them. So let's not. */
+
+    /* When the repcodes are outside of the prefix, we set them to zero before the loop.
+     * When the offsets are still zero, we need to restore them after the block to have a correct
+     * repcode history. If only one offset was invalid, it is easy. The tricky case is when both
+     * offsets were invalid. We need to figure out which offset to refill with.
+     *     - If both offsets are zero they are in the same order.
+     *     - If both offsets are non-zero, we won't restore the offsets from `offsetSaved[12]`.
+     *     - If only one is zero, we need to decide which offset to restore.
+     *         - If rep_offset1 is non-zero, then rep_offset2 must be offsetSaved1.
+     *         - It is impossible for rep_offset2 to be non-zero.
+     *
+     * So if rep_offset1 started invalid (offsetSaved1 != 0) and became valid (rep_offset1 != 0), then
+     * set rep[0] = rep_offset1 and rep[1] = offsetSaved1.
+     */
+    offsetSaved2 = ((offsetSaved1 != 0) && (rep_offset1 != 0)) ? offsetSaved1 : offsetSaved2;
+
+    /* save reps for next block */
+    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved1;
+    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved2;
+
+    /* Return the last literals size */
+    return (size_t)(iend - anchor);
+
+_offset: /* Requires: ip0, idx */
+
+    /* Compute the offset code. */
+    match0 = base + matchIdx;
+    rep_offset2 = rep_offset1;
+    rep_offset1 = (U32)(ip0-match0);
+    offcode = OFFSET_TO_OFFBASE(rep_offset1);
+    mLength = 4;
+
+    /* Count the backwards match length. */
+    while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) {
+        ip0--;
+        match0--;
+        mLength++;
+    }
+
+_match: /* Requires: ip0, match0, offcode */
+
+    /* Count the forward length. */
+    mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend);
+
+    ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
+
+    ip0 += mLength;
+    anchor = ip0;
+
+    /* Fill table and check for immediate repcode. */
+    if (ip0 <= ilimit) {
+        /* Fill Table */
+        assert(base+current0+2 > istart);  /* check base overflow */
+        hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
+        hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+
+        if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */
+            while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) {
+                /* store sequence */
+                size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4;
+                { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */
+                hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
+                ip0 += rLength;
+                ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
+                anchor = ip0;
+                continue;   /* faster when present (confirmed on gcc-8) ... (?) */
+    }   }   }
+
+    goto _start;
+}
+
+#define ZSTD_GEN_FAST_FN(dictMode, mml, cmov)                                                       \
+    static size_t ZSTD_compressBlock_fast_##dictMode##_##mml##_##cmov(                              \
+            ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                    \
+            void const* src, size_t srcSize)                                                       \
+    {                                                                                              \
+        return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mml, cmov); \
+    }
+
+ZSTD_GEN_FAST_FN(noDict, 4, 1)
+ZSTD_GEN_FAST_FN(noDict, 5, 1)
+ZSTD_GEN_FAST_FN(noDict, 6, 1)
+ZSTD_GEN_FAST_FN(noDict, 7, 1)
+
+ZSTD_GEN_FAST_FN(noDict, 4, 0)
+ZSTD_GEN_FAST_FN(noDict, 5, 0)
+ZSTD_GEN_FAST_FN(noDict, 6, 0)
+ZSTD_GEN_FAST_FN(noDict, 7, 0)
+
+size_t ZSTD_compressBlock_fast(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    U32 const mml = ms->cParams.minMatch;
+    /* use cmov when "candidate in range" branch is likely unpredictable */
+    int const useCmov = ms->cParams.windowLog < 19;
+    assert(ms->dictMatchState == NULL);
+    if (useCmov) {
+        switch(mml)
+        {
+        default: /* includes case 3 */
+        case 4 :
+            return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize);
+        case 5 :
+            return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize);
+        case 6 :
+            return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize);
+        case 7 :
+            return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize);
+        }
+    } else {
+        /* use a branch instead */
+        switch(mml)
+        {
+        default: /* includes case 3 */
+        case 4 :
+            return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize);
+        case 5 :
+            return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize);
+        case 6 :
+            return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize);
+        case 7 :
+            return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize);
+        }
+    }
+}
+
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_compressBlock_fast_dictMatchState_generic(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashTable = ms->hashTable;
+    U32 const hlog = cParams->hashLog;
+    /* support stepSize of 0 */
+    U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
+    const BYTE* const base = ms->window.base;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* ip0 = istart;
+    const BYTE* ip1 = ip0 + stepSize; /* we assert below that stepSize >= 1 */
+    const BYTE* anchor = istart;
+    const U32   prefixStartIndex = ms->window.dictLimit;
+    const BYTE* const prefixStart = base + prefixStartIndex;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - HASH_READ_SIZE;
+    U32 offset_1=rep[0], offset_2=rep[1];
+
+    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
+    const ZSTD_compressionParameters* const dictCParams = &dms->cParams ;
+    const U32* const dictHashTable = dms->hashTable;
+    const U32 dictStartIndex       = dms->window.dictLimit;
+    const BYTE* const dictBase     = dms->window.base;
+    const BYTE* const dictStart    = dictBase + dictStartIndex;
+    const BYTE* const dictEnd      = dms->window.nextSrc;
+    const U32 dictIndexDelta       = prefixStartIndex - (U32)(dictEnd - dictBase);
+    const U32 dictAndPrefixLength  = (U32)(istart - prefixStart + dictEnd - dictStart);
+    const U32 dictHBits            = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+
+    /* if a dictionary is still attached, it necessarily means that
+     * it is within window size. So we just check it. */
+    const U32 maxDistance = 1U << cParams->windowLog;
+    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
+    assert(endIndex - prefixStartIndex <= maxDistance);
+    (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */
+
+    (void)hasStep; /* not currently specialized on whether it's accelerated */
+
+    /* ensure there will be no underflow
+     * when translating a dict index into a local index */
+    assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
+
+    if (ms->prefetchCDictTables) {
+        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
+        PREFETCH_AREA(dictHashTable, hashTableBytes);
+    }
+
+    /* init */
+    DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic");
+    ip0 += (dictAndPrefixLength == 0);
+    /* dictMatchState repCode checks don't currently handle repCode == 0
+     * disabling. */
+    assert(offset_1 <= dictAndPrefixLength);
+    assert(offset_2 <= dictAndPrefixLength);
+
+    /* Outer search loop */
+    assert(stepSize >= 1);
+    while (ip1 <= ilimit) {   /* repcode check at (ip0 + 1) is safe because ip0 < ip1 */
+        size_t mLength;
+        size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+
+        size_t const dictHashAndTag0 = ZSTD_hashPtr(ip0, dictHBits, mls);
+        U32 dictMatchIndexAndTag = dictHashTable[dictHashAndTag0 >> ZSTD_SHORT_CACHE_TAG_BITS];
+        int dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag0);
+
+        U32 matchIndex = hashTable[hash0];
+        U32 curr = (U32)(ip0 - base);
+        size_t step = stepSize;
+        const size_t kStepIncr = 1 << kSearchStrength;
+        const BYTE* nextStep = ip0 + kStepIncr;
+
+        /* Inner search loop */
+        while (1) {
+            const BYTE* match = base + matchIndex;
+            const U32 repIndex = curr + 1 - offset_1;
+            const BYTE* repMatch = (repIndex < prefixStartIndex) ?
+                                   dictBase + (repIndex - dictIndexDelta) :
+                                   base + repIndex;
+            const size_t hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+            size_t const dictHashAndTag1 = ZSTD_hashPtr(ip1, dictHBits, mls);
+            hashTable[hash0] = curr;   /* update hash table */
+
+            if ((ZSTD_index_overlap_check(prefixStartIndex, repIndex))
+                && (MEM_read32(repMatch) == MEM_read32(ip0 + 1))) {
+                const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
+                mLength = ZSTD_count_2segments(ip0 + 1 + 4, repMatch + 4, iend, repMatchEnd, prefixStart) + 4;
+                ip0++;
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
+                break;
+            }
+
+            if (dictTagsMatch) {
+                /* Found a possible dict match */
+                const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
+                const BYTE* dictMatch = dictBase + dictMatchIndex;
+                if (dictMatchIndex > dictStartIndex &&
+                    MEM_read32(dictMatch) == MEM_read32(ip0)) {
+                    /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */
+                    if (matchIndex <= prefixStartIndex) {
+                        U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta);
+                        mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4;
+                        while (((ip0 > anchor) & (dictMatch > dictStart))
+                            && (ip0[-1] == dictMatch[-1])) {
+                            ip0--;
+                            dictMatch--;
+                            mLength++;
+                        } /* catch up */
+                        offset_2 = offset_1;
+                        offset_1 = offset;
+                        ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                        break;
+                    }
+                }
+            }
+
+            if (ZSTD_match4Found_cmov(ip0, match, matchIndex, prefixStartIndex)) {
+                /* found a regular match of size >= 4 */
+                U32 const offset = (U32) (ip0 - match);
+                mLength = ZSTD_count(ip0 + 4, match + 4, iend) + 4;
+                while (((ip0 > anchor) & (match > prefixStart))
+                       && (ip0[-1] == match[-1])) {
+                    ip0--;
+                    match--;
+                    mLength++;
+                } /* catch up */
+                offset_2 = offset_1;
+                offset_1 = offset;
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                break;
+            }
+
+            /* Prepare for next iteration */
+            dictMatchIndexAndTag = dictHashTable[dictHashAndTag1 >> ZSTD_SHORT_CACHE_TAG_BITS];
+            dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag1);
+            matchIndex = hashTable[hash1];
+
+            if (ip1 >= nextStep) {
+                step++;
+                nextStep += kStepIncr;
+            }
+            ip0 = ip1;
+            ip1 = ip1 + step;
+            if (ip1 > ilimit) goto _cleanup;
+
+            curr = (U32)(ip0 - base);
+            hash0 = hash1;
+        }   /* end inner search loop */
+
+        /* match found */
+        assert(mLength);
+        ip0 += mLength;
+        anchor = ip0;
+
+        if (ip0 <= ilimit) {
+            /* Fill Table */
+            assert(base+curr+2 > istart);  /* check base overflow */
+            hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;  /* here because curr+2 could be > iend-8 */
+            hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+
+            /* check immediate repcode */
+            while (ip0 <= ilimit) {
+                U32 const current2 = (U32)(ip0-base);
+                U32 const repIndex2 = current2 - offset_2;
+                const BYTE* repMatch2 = repIndex2 < prefixStartIndex ?
+                        dictBase - dictIndexDelta + repIndex2 :
+                        base + repIndex2;
+                if ( (ZSTD_index_overlap_check(prefixStartIndex, repIndex2))
+                   && (MEM_read32(repMatch2) == MEM_read32(ip0))) {
+                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
+                    size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+                    U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                    hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = current2;
+                    ip0 += repLength2;
+                    anchor = ip0;
+                    continue;
+                }
+                break;
+            }
+        }
+
+        /* Prepare for next iteration */
+        assert(ip0 == anchor);
+        ip1 = ip0 + stepSize;
+    }
+
+_cleanup:
+    /* save reps for next block */
+    rep[0] = offset_1;
+    rep[1] = offset_2;
+
+    /* Return the last literals size */
+    return (size_t)(iend - anchor);
+}
+
+
+ZSTD_GEN_FAST_FN(dictMatchState, 4, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 5, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 6, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 7, 0)
+
+size_t ZSTD_compressBlock_fast_dictMatchState(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    U32 const mls = ms->cParams.minMatch;
+    assert(ms->dictMatchState != NULL);
+    switch(mls)
+    {
+    default: /* includes case 3 */
+    case 4 :
+        return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize);
+    case 5 :
+        return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize);
+    case 6 :
+        return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize);
+    case 7 :
+        return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize);
+    }
+}
+
+
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_compressBlock_fast_extDict_generic(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashTable = ms->hashTable;
+    U32 const hlog = cParams->hashLog;
+    /* support stepSize of 0 */
+    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* anchor = istart;
+    const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
+    const U32   lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
+    const U32   dictStartIndex = lowLimit;
+    const BYTE* const dictStart = dictBase + dictStartIndex;
+    const U32   dictLimit = ms->window.dictLimit;
+    const U32   prefixStartIndex = dictLimit < lowLimit ? lowLimit : dictLimit;
+    const BYTE* const prefixStart = base + prefixStartIndex;
+    const BYTE* const dictEnd = dictBase + prefixStartIndex;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - 8;
+    U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
+
+    const BYTE* ip0 = istart;
+    const BYTE* ip1;
+    const BYTE* ip2;
+    const BYTE* ip3;
+    U32 current0;
+
+
+    size_t hash0; /* hash for ip0 */
+    size_t hash1; /* hash for ip1 */
+    U32 idx; /* match idx for ip0 */
+    const BYTE* idxBase; /* base pointer for idx */
+
+    U32 offcode;
+    const BYTE* match0;
+    size_t mLength;
+    const BYTE* matchEnd = 0; /* initialize to avoid warning, assert != 0 later */
+
+    size_t step;
+    const BYTE* nextStep;
+    const size_t kStepIncr = (1 << (kSearchStrength - 1));
+
+    (void)hasStep; /* not currently specialized on whether it's accelerated */
+
+    DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);
+
+    /* switch to "regular" variant if extDict is invalidated due to maxDistance */
+    if (prefixStartIndex == dictStartIndex)
+        return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
+
+    {   U32 const curr = (U32)(ip0 - base);
+        U32 const maxRep = curr - dictStartIndex;
+        if (offset_2 >= maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 >= maxRep) offsetSaved1 = offset_1, offset_1 = 0;
     }
 
     /* start each op */
···
     hash1 = ZSTD_hashPtr(ip1, hlog, mls);
 
     idx = hashTable[hash0];
+    idxBase = idx < prefixStartIndex ? dictBase : base;
 
     do {
-        /* load repcode match for ip[2]*/
-        const U32 rval = MEM_read32(ip2 - rep_offset1);
+        {   /* load repcode match for ip[2] */
+            U32 const current2 = (U32)(ip2 - base);
+            U32 const repIndex = current2 - offset_1;
+            const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
+            U32 rval;
+            if ( ((U32)(prefixStartIndex - repIndex) >= 4) /* intentional underflow */
+                 & (offset_1 > 0) ) {
+                rval = MEM_read32(repBase + repIndex);
+            } else {
+                rval = MEM_read32(ip2) ^ 1; /* guaranteed to not match. */
+            }
 
-        /* write back hash table entry */
-        current0 = (U32)(ip0 - base);
-        hashTable[hash0] = current0;
+            /* write back hash table entry */
+            current0 = (U32)(ip0 - base);
+            hashTable[hash0] = current0;
 
-        /* check repcode at ip[2] */
-        if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) {
-            ip0 = ip2;
-            match0 = ip0 - rep_offset1;
-            mLength = ip0[-1] == match0[-1];
-            ip0 -= mLength;
-            match0 -= mLength;
-            offcode = STORE_REPCODE_1;
-            mLength += 4;
-            goto _match;
-        }
+            /* check repcode at ip[2] */
+            if (MEM_read32(ip2) == rval) {
+                ip0 = ip2;
+                match0 = repBase + repIndex;
+                matchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
+                assert((match0 != prefixStart) & (match0 != dictStart));
+                mLength = ip0[-1] == match0[-1];
+                ip0 -= mLength;
+                match0 -= mLength;
+                offcode = REPCODE1_TO_OFFBASE;
+                mLength += 4;
+                goto _match;
+        }   }
 
-        /* load match for ip[0] */
-        if (idx >= prefixStartIndex) {
-            mval = MEM_read32(base + idx);
-        } else {
-            mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
-        }
+        {   /* load match for ip[0] */
+            U32 const mval = idx >= dictStartIndex ?
+                    MEM_read32(idxBase + idx) :
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
 
-        /* check match at ip[0] */
-        if (MEM_read32(ip0) == mval) {
-            /* found a match! */
-            goto _offset;
-        }
+            /* check match at ip[0] */
+            if (MEM_read32(ip0) == mval) {
+                /* found a match! */
+                goto _offset;
+        }   }
 
         /* lookup ip[1] */
         idx = hashTable[hash1];
+        idxBase = idx < prefixStartIndex ? dictBase : base;
 
         /* hash ip[2] */
         hash0 = hash1;
···
         current0 = (U32)(ip0 - base);
         hashTable[hash0] = current0;
 
-        /* load match for ip[0] */
-        if (idx >= prefixStartIndex) {
-            mval = MEM_read32(base + idx);
-        } else {
-            mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
-        }
+        {   /* load match for ip[0] */
+            U32 const mval = idx >= dictStartIndex ?
+                    MEM_read32(idxBase + idx) :
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
 
-        /* check match at ip[0] */
-        if (MEM_read32(ip0) == mval) {
-            /* found a match! */
-            goto _offset;
-        }
+            /* check match at ip[0] */
+            if (MEM_read32(ip0) == mval) {
+                /* found a match! */
+                goto _offset;
+        }   }
 
         /* lookup ip[1] */
         idx = hashTable[hash1];
+        idxBase = idx < prefixStartIndex ? dictBase : base;
 
         /* hash ip[2] */
         hash0 = hash1;
···
      * However, it seems to be a meaningful performance hit to try to search
      * them. So let's not. */
 
+    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
+
     /* save reps for next block */
-    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved;
-    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved;
+    rep[0] = offset_1 ? offset_1 : offsetSaved1;
+    rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 
-_offset: /* Requires: ip0, idx */
+_offset: /* Requires: ip0, idx, idxBase */
 
     /* Compute the offset code. */
-    match0 = base + idx;
-    rep_offset2 = rep_offset1;
-    rep_offset1 = (U32)(ip0-match0);
-    offcode = STORE_OFFSET(rep_offset1);
-    mLength = 4;
+    {   U32 const offset = current0 - idx;
+        const BYTE* const lowMatchPtr = idx < prefixStartIndex ? dictStart : prefixStart;
+        matchEnd = idx < prefixStartIndex ? dictEnd : iend;
+        match0 = idxBase + idx;
+        offset_2 = offset_1;
+        offset_1 = offset;
+        offcode = OFFSET_TO_OFFBASE(offset);
+        mLength = 4;
 
-    /* Count the backwards match length. */
-    while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) {
-        ip0--;
-        match0--;
-        mLength++;
-    }
+        /* Count the backwards match length. */
+        while (((ip0>anchor) & (match0>lowMatchPtr)) && (ip0[-1] == match0[-1])) {
+            ip0--;
+            match0--;
+            mLength++;
+    }   }
 
-_match: /* Requires: ip0, match0, offcode */
+_match: /* Requires: ip0, match0, offcode, matchEnd */
 
     /* Count the forward length. */
-    mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend);
+    assert(matchEnd != 0);
+    mLength += ZSTD_count_2segments(ip0 + mLength, match0 + mLength, iend, matchEnd, prefixStart);
 
     ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
 
···
         hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
         hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
 
-        if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */
-            while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) {
-                /* store sequence */
-                size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4;
-                { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */
+        while (ip0 <= ilimit) {
+            U32 const repIndex2 = (U32)(ip0-base) - offset_2;
+            const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
+            if ( ((ZSTD_index_overlap_check(prefixStartIndex, repIndex2)) & (offset_2 > 0))
+                 && (MEM_read32(repMatch2) == MEM_read32(ip0)) ) {
+                const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
+                size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+                { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
+                ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                 hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
-                ip0 += rLength;
-                ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, STORE_REPCODE_1, rLength);
+                ip0 += repLength2;
                 anchor = ip0;
-                continue;   /* faster when present (confirmed on gcc-8) ... (?) */
-    }   }   }
+                continue;
+            }
+            break;
+    }   }
 
     goto _start;
-}
-
-#define ZSTD_GEN_FAST_FN(dictMode, mls, step)                                                            \
-    static size_t ZSTD_compressBlock_fast_##dictMode##_##mls##_##step(                                      \
-            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                    \
-            void const* src, size_t srcSize)                                                       \
-    {                                                                                              \
-        return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls, step); \
-    }
-
-ZSTD_GEN_FAST_FN(noDict, 4, 1)
-ZSTD_GEN_FAST_FN(noDict, 5, 1)
-ZSTD_GEN_FAST_FN(noDict, 6, 1)
-ZSTD_GEN_FAST_FN(noDict, 7, 1)
-
-ZSTD_GEN_FAST_FN(noDict, 4, 0)
-ZSTD_GEN_FAST_FN(noDict, 5, 0)
-ZSTD_GEN_FAST_FN(noDict, 6, 0)
-ZSTD_GEN_FAST_FN(noDict, 7, 0)
-
-size_t ZSTD_compressBlock_fast(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    U32 const mls = ms->cParams.minMatch;
-    assert(ms->dictMatchState == NULL);
-    if (ms->cParams.targetLength > 1) {
-        switch(mls)
-        {
-        default: /* includes case 3 */
-        case 4 :
-            return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize);
-        case 5 :
-            return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize);
-        case 6 :
-            return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize);
-        case 7 :
-            return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize);
-        }
-    } else {
-        switch(mls)
-        {
-        default: /* includes case 3 */
-        case 4 :
-            return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize);
-        case 5 :
-            return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize);
-        case 6 :
-            return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize);
-        case 7 :
-            return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize);
-        }
-
-    }
-}
-
-FORCE_INLINE_TEMPLATE
-size_t ZSTD_compressBlock_fast_dictMatchState_generic(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
-{
-    const ZSTD_compressionParameters* const cParams = &ms->cParams;
-    U32* const hashTable = ms->hashTable;
-    U32 const hlog = cParams->hashLog;
-    /* support stepSize of 0 */
-    U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
-    const BYTE* const base = ms->window.base;
-    const BYTE* const istart = (const BYTE*)src;
-    const BYTE* ip = istart;
-    const BYTE* anchor = istart;
-    const U32   prefixStartIndex = ms->window.dictLimit;
-    const BYTE* const prefixStart = base + prefixStartIndex;
-    const BYTE* const iend = istart + srcSize;
-    const BYTE* const ilimit = iend - HASH_READ_SIZE;
-    U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
-
-    const ZSTD_matchState_t* const dms = ms->dictMatchState;
-    const ZSTD_compressionParameters* const dictCParams = &dms->cParams ;
-    const U32* const dictHashTable = dms->hashTable;
-    const U32 dictStartIndex       = dms->window.dictLimit;
-    const BYTE* const dictBase     = dms->window.base;
-    const BYTE* const dictStart    = dictBase + dictStartIndex;
-    const BYTE* const dictEnd      = dms->window.nextSrc;
-    const U32 dictIndexDelta       = prefixStartIndex - (U32)(dictEnd - dictBase);
-    const U32 dictAndPrefixLength  = (U32)(ip - prefixStart + dictEnd - dictStart);
-    const U32 dictHLog             = dictCParams->hashLog;
-
-    /* if a dictionary is still attached, it necessarily means that
-     * it is within window size. So we just check it. */
-    const U32 maxDistance = 1U << cParams->windowLog;
-    const U32 endIndex = (U32)((size_t)(ip - base) + srcSize);
-    assert(endIndex - prefixStartIndex <= maxDistance);
-    (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */
-
-    (void)hasStep; /* not currently specialized on whether it's accelerated */
-
-    /* ensure there will be no underflow
-     * when translating a dict index into a local index */
-    assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
-
-    /* init */
-    DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic");
-    ip += (dictAndPrefixLength == 0);
-    /* dictMatchState repCode checks don't currently handle repCode == 0
-     * disabling. */
-    assert(offset_1 <= dictAndPrefixLength);
-    assert(offset_2 <= dictAndPrefixLength);
-
-    /* Main Search Loop */
-    while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
-        size_t mLength;
-        size_t const h = ZSTD_hashPtr(ip, hlog, mls);
-        U32 const curr = (U32)(ip-base);
-        U32 const matchIndex = hashTable[h];
-        const BYTE* match = base + matchIndex;
-        const U32 repIndex = curr + 1 - offset_1;
-        const BYTE* repMatch = (repIndex < prefixStartIndex) ?
-                               dictBase + (repIndex - dictIndexDelta) :
-                               base + repIndex;
-        hashTable[h] = curr;   /* update hash table */
-
-        if ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
-          && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
-            const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
-            mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength);
-        } else if ( (matchIndex <= prefixStartIndex) ) {
-            size_t const dictHash = ZSTD_hashPtr(ip, dictHLog, mls);
-            U32 const dictMatchIndex = dictHashTable[dictHash];
-            const BYTE* dictMatch = dictBase + dictMatchIndex;
-            if (dictMatchIndex <= dictStartIndex ||
-                MEM_read32(dictMatch) != MEM_read32(ip)) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
-            } else {
-                /* found a dict match */
-                U32 const offset = (U32)(curr-dictMatchIndex-dictIndexDelta);
-                mLength = ZSTD_count_2segments(ip+4, dictMatch+4, iend, dictEnd, prefixStart) + 4;
-                while (((ip>anchor) & (dictMatch>dictStart))
-                     && (ip[-1] == dictMatch[-1])) {
-                    ip--; dictMatch--; mLength++;
-                } /* catch up */
-                offset_2 = offset_1;
-                offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
-            }
-        } else if (MEM_read32(match) != MEM_read32(ip)) {
-            /* it's not a match, and we're not going to check the dictionary */
-            assert(stepSize >= 1);
-            ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-            continue;
-        } else {
-            /* found a regular match */
-            U32 const offset = (U32)(ip-match);
-            mLength = ZSTD_count(ip+4, match+4, iend) + 4;
-            while (((ip>anchor) & (match>prefixStart))
-                 && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
-            offset_2 = offset_1;
-            offset_1 = offset;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
-        }
-
-        /* match found */
-        ip += mLength;
-        anchor = ip;
-
-        if (ip <= ilimit) {
-            /* Fill Table */
-            assert(base+curr+2 > istart);  /* check base overflow */
-            hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;  /* here because curr+2 could be > iend-8 */
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
-
-            /* check immediate repcode */
-            while (ip <= ilimit) {
-                U32 const current2 = (U32)(ip-base);
-                U32 const repIndex2 = current2 - offset_2;
-                const BYTE* repMatch2 = repIndex2 < prefixStartIndex ?
-                        dictBase - dictIndexDelta + repIndex2 :
-                        base + repIndex2;
-                if ( ((U32)((prefixStartIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
-                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
-                    U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, repLength2);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
-                    continue;
-                }
-                break;
-            }
-        }
-    }
-
-    /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
-
-    /* Return the last literals size */
-    return (size_t)(iend - anchor);
-}
-
-
-ZSTD_GEN_FAST_FN(dictMatchState, 4, 0)
-ZSTD_GEN_FAST_FN(dictMatchState, 5, 0)
-ZSTD_GEN_FAST_FN(dictMatchState, 6, 0)
-ZSTD_GEN_FAST_FN(dictMatchState, 7, 0)
-
-size_t ZSTD_compressBlock_fast_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    U32 const mls = ms->cParams.minMatch;
-    assert(ms->dictMatchState != NULL);
-    switch(mls)
-    {
-    default: /* includes case 3 */
-    case 4 :
-        return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize);
-    case 5 :
-        return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize);
-    case 6 :
-        return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize);
-    case 7 :
-        return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize);
-    }
-}
-
-
-static size_t ZSTD_compressBlock_fast_extDict_generic(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
-{
-    const ZSTD_compressionParameters* const cParams = &ms->cParams;
-    U32* const hashTable = ms->hashTable;
-    U32 const hlog = cParams->hashLog;
-    /* support stepSize of 0 */
-    U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
-    const BYTE* const base = ms->window.base;
-    const BYTE* const dictBase = ms->window.dictBase;
-    const BYTE* const istart = (const BYTE*)src;
-    const BYTE* ip = istart;
-    const BYTE* anchor = istart;
-    const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
-    const U32   lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
-    const U32   dictStartIndex = lowLimit;
-    const BYTE* const dictStart = dictBase + dictStartIndex;
-    const U32   dictLimit = ms->window.dictLimit;
-    const U32   prefixStartIndex = dictLimit < lowLimit ? lowLimit : dictLimit;
-    const BYTE* const prefixStart = base + prefixStartIndex;
-    const BYTE* const dictEnd = dictBase + prefixStartIndex;
-    const BYTE* const iend = istart + srcSize;
-    const BYTE* const ilimit = iend - 8;
-    U32 offset_1=rep[0], offset_2=rep[1];
-
-    (void)hasStep; /* not currently specialized on whether it's accelerated */
-
-    DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);
-
-    /* switch to "regular" variant if extDict is invalidated due to maxDistance */
-    if (prefixStartIndex == dictStartIndex)
-        return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
-
-    /* Search Loop */
-    while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
-        const size_t h = ZSTD_hashPtr(ip, hlog, mls);
-        const U32    matchIndex = hashTable[h];
-        const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base;
-        const BYTE*  match = matchBase + matchIndex;
-        const U32    curr = (U32)(ip-base);
-        const U32    repIndex = curr + 1 - offset_1;
-        const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
-        const BYTE* const repMatch = repBase + repIndex;
-        hashTable[h] = curr;   /* update hash table */
-        DEBUGLOG(7, "offset_1 = %u , curr = %u", offset_1, curr);
-
-        if ( ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */
-             & (offset_1 <= curr+1 - dictStartIndex) ) /* note: we are searching at curr+1 */
-           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
-            const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
-            size_t const rLength = ZSTD_count_2segments(ip+1 +4, repMatch +4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, rLength);
-            ip += rLength;
-            anchor = ip;
-        } else {
-            if ( (matchIndex < dictStartIndex) ||
-                 (MEM_read32(match) != MEM_read32(ip)) ) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
-            }
-            {   const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
-                const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
-                U32 const offset = curr - matchIndex;
-                size_t mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
-                while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
-                offset_2 = offset_1; offset_1 = offset;  /* update offset history */
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
-                ip += mLength;
-                anchor = ip;
-        }   }
-
-        if (ip <= ilimit) {
-            /* Fill Table */
-            hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
-            /* check immediate repcode */
-            while (ip <= ilimit) {
-                U32 const current2 = (U32)(ip-base);
-                U32 const repIndex2 = current2 - offset_2;
-                const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
-                if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 <= curr - dictStartIndex))  /* intentional overflow */
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
-                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
-                    { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, STORE_REPCODE_1, repLength2);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
-                    continue;
-                }
-                break;
-    }   }   }
-
-    /* save reps for next block */
-    rep[0] = offset_1;
-    rep[1] = offset_2;
-
-    /* Return the last literals size */
-    return (size_t)(iend - anchor);
 }
 
 ZSTD_GEN_FAST_FN(extDict, 4, 0)
···
 ZSTD_GEN_FAST_FN(extDict, 7, 0)
 
 size_t ZSTD_compressBlock_fast_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     U32 const mls = ms->cParams.minMatch;
+    assert(ms->dictMatchState == NULL);
     switch(mls)
     {
     default: /* includes case 3 */

lib/zstd/compress/zstd_fast.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #ifndef ZSTD_FAST_H
 #define ZSTD_FAST_H
 
-
 #include "../common/mem.h"      /* U32 */
 #include "zstd_compress_internal.h"
 
-void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
-                        void const* end, ZSTD_dictTableLoadMethod_e dtlm);
+void ZSTD_fillHashTable(ZSTD_MatchState_t* ms,
+                        void const* end, ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp);
 size_t ZSTD_compressBlock_fast(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_fast_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_fast_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
-
 
 #endif /* ZSTD_FAST_H */

lib/zstd/compress/zstd_lazy.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 
 #include "zstd_compress_internal.h"
 #include "zstd_lazy.h"
+#include "../common/bits.h" /* ZSTD_countTrailingZeros64 */
+
+#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR)
+
+#define kLazySkippingStep 8
 
 
 /*-*************************************
 *  Binary Tree search
 ***************************************/
 
-static void
-ZSTD_updateDUBT(ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_updateDUBT(ZSTD_MatchState_t* ms,
                 const BYTE* ip, const BYTE* iend,
                 U32 mls)
 {
···
  *  sort one already inserted but unsorted position
  *  assumption : curr >= btlow == (curr - btmask)
  *  doesn't fail */
-static void
-ZSTD_insertDUBT1(const ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_insertDUBT1(const ZSTD_MatchState_t* ms,
                  U32 curr, const BYTE* inputEnd,
                  U32 nbCompares, U32 btLow,
                  const ZSTD_dictMode_e dictMode)
···
 }
 
 
-static size_t
-ZSTD_DUBT_findBetterDictMatch (
-        const ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_DUBT_findBetterDictMatch (
+        const ZSTD_MatchState_t* ms,
         const BYTE* const ip, const BYTE* const iend,
         size_t* offsetPtr,
         size_t bestLength,
···
         U32 const mls,
         const ZSTD_dictMode_e dictMode)
 {
-    const ZSTD_matchState_t * const dms = ms->dictMatchState;
+    const ZSTD_MatchState_t * const dms = ms->dictMatchState;
     const ZSTD_compressionParameters* const dmsCParams = &dms->cParams;
     const U32 * const dictHashTable = dms->hashTable;
     U32         const hashLog = dmsCParams->hashLog;
···
             U32 matchIndex = dictMatchIndex + dictIndexDelta;
             if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
                 DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)",
-                    curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, STORE_OFFSET(curr - matchIndex), dictMatchIndex, matchIndex);
-                bestLength = matchLength, *offsetPtr = STORE_OFFSET(curr - matchIndex);
+                    curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, OFFSET_TO_OFFBASE(curr - matchIndex), dictMatchIndex, matchIndex);
+                bestLength = matchLength, *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
             }
             if (ip+matchLength == iend) {   /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */
                 break;   /* drop, to guarantee consistency (miss a little bit of compression) */
···
     }
 
     if (bestLength >= MINMATCH) {
-        U32 const mIndex = curr - (U32)STORED_OFFSET(*offsetPtr); (void)mIndex;
+        U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offsetPtr); (void)mIndex;
         DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
                     curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
     }
···
 }
 
 
-static size_t
-ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_DUBT_findBestMatch(ZSTD_MatchState_t* ms,
                         const BYTE* const ip, const BYTE* const iend,
-                        size_t* offsetPtr,
+                        size_t* offBasePtr,
                         U32 const mls,
                         const ZSTD_dictMode_e dictMode)
 {
···
             if (matchLength > bestLength) {
                 if (matchLength > matchEndIdx - matchIndex)
                     matchEndIdx = matchIndex + (U32)matchLength;
-                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) )
-                    bestLength = matchLength, *offsetPtr = STORE_OFFSET(curr - matchIndex);
+                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr - matchIndex + 1) - ZSTD_highbit32((U32)*offBasePtr)) )
+                    bestLength = matchLength, *offBasePtr = OFFSET_TO_OFFBASE(curr - matchIndex);
                 if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
                     if (dictMode == ZSTD_dictMatchState) {
                         nbCompares = 0; /* in addition to avoiding checking any
···
         if (dictMode == ZSTD_dictMatchState && nbCompares) {
             bestLength = ZSTD_DUBT_findBetterDictMatch(
                     ms, ip, iend,
-                    offsetPtr, bestLength, nbCompares,
+                    offBasePtr, bestLength, nbCompares,
                     mls, dictMode);
         }
 
         assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */
         ms->nextToUpdate = matchEndIdx - 8;   /* skip repetitive patterns */
         if (bestLength >= MINMATCH) {
-            U32 const mIndex = curr - (U32)STORED_OFFSET(*offsetPtr); (void)mIndex;
+            U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offBasePtr); (void)mIndex;
             DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
-                        curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
+                        curr, (U32)bestLength, (U32)*offBasePtr, mIndex);
         }
         return bestLength;
     }
···
 
 
 /* ZSTD_BtFindBestMatch() : Tree updater, providing best match */
-FORCE_INLINE_TEMPLATE size_t
-ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_BtFindBestMatch( ZSTD_MatchState_t* ms,
                 const BYTE* const ip, const BYTE* const iLimit,
-                      size_t* offsetPtr,
+                      size_t* offBasePtr,
                 const U32 mls /* template */,
                 const ZSTD_dictMode_e dictMode)
 {
     DEBUGLOG(7, "ZSTD_BtFindBestMatch");
     if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
     ZSTD_updateDUBT(ms, ip, iLimit, mls);
-    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode);
+    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offBasePtr, mls, dictMode);
 }
 
 /* *********************************
 * Dedicated dict search
 ***********************************/
 
-void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip)
+void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_MatchState_t* ms, const BYTE* const ip)
 {
     const BYTE* const base = ms->window.base;
     U32 const target = (U32)(ip - base);
···
  */
 FORCE_INLINE_TEMPLATE
 size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nbAttempts,
-                                            const ZSTD_matchState_t* const dms,
+                                            const ZSTD_MatchState_t* const dms,
                                             const BYTE* const ip, const BYTE* const iLimit,
                                             const BYTE* const prefixStart, const U32 curr,
                                             const U32 dictLimit, const size_t ddsIdx) {
···
         /* save best solution */
         if (currentMl > ml) {
             ml = currentMl;
-            *offsetPtr = STORE_OFFSET(curr - (matchIndex + ddsIndexDelta));
+            *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
             if (ip+currentMl == iLimit) {
                 /* best possible, avoids read overflow on next attempt */
                 return ml;
···
             /* save best solution */
             if (currentMl > ml) {
                 ml = currentMl;
-                *offsetPtr = STORE_OFFSET(curr - (matchIndex + ddsIndexDelta));
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
             }
         }
···
 
 /* Update chains up to ip (excluded)
    Assumption : always within prefix (i.e. not within extDict) */
-FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal(
-                        ZSTD_matchState_t* ms,
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_insertAndFindFirstIndex_internal(
+                        ZSTD_MatchState_t* ms,
                         const ZSTD_compressionParameters* const cParams,
-                        const BYTE* ip, U32 const mls)
+                        const BYTE* ip, U32 const mls, U32 const lazySkipping)
 {
     U32* const hashTable  = ms->hashTable;
     const U32 hashLog = cParams->hashLog;
···
         NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
         hashTable[h] = idx;
         idx++;
+        /* Stop inserting every position when in the lazy skipping mode. */
+        if (lazySkipping)
+            break;
     }
 
     ms->nextToUpdate = target;
     return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
 }
 
-U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) {
+U32 ZSTD_insertAndFindFirstIndex(ZSTD_MatchState_t* ms, const BYTE* ip) {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
-    return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch);
+    return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch, /* lazySkipping*/ 0);
 }
 
 /* inlining is important to hardwire a hot branch (template emulation) */
 FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_HcFindBestMatch(
-                        ZSTD_matchState_t* ms,
+                        ZSTD_MatchState_t* ms,
                         const BYTE* const ip, const BYTE* const iLimit,
                         size_t* offsetPtr,
                         const U32 mls, const ZSTD_dictMode_e dictMode)
···
     U32 nbAttempts = 1U << cParams->searchLog;
     size_t ml=4-1;
 
-    const ZSTD_matchState_t* const dms = ms->dictMatchState;
+    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
     const U32 ddsHashLog = dictMode == ZSTD_dedicatedDictSearch
                          ? dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
     const size_t ddsIdx = dictMode == ZSTD_dedicatedDictSearch
···
     }
 
     /* HC4 match finder */
-    matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
+    matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls, ms->lazySkipping);
 
     for ( ; (matchIndex>=lowLimit) & (nbAttempts>0) ; nbAttempts--) {
         size_t currentMl=0;
         if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
             const BYTE* const match = base + matchIndex;
             assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
-            if (match[ml] == ip[ml])   /* potentially better */
+            /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
+            if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
                 currentMl = ZSTD_count(ip, match, iLimit);
         } else {
             const BYTE* const match = dictBase + matchIndex;
···
         /* save best solution */
         if (currentMl > ml) {
             ml = currentMl;
-            *offsetPtr = STORE_OFFSET(curr - matchIndex);
+            *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
             if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
         }
 
···
             if (currentMl > ml) {
                 ml = currentMl;
                 assert(curr > matchIndex + dmsIndexDelta);
-                *offsetPtr = STORE_OFFSET(curr - (matchIndex + dmsIndexDelta));
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
             }
 
···
 * (SIMD) Row-based matchfinder
 ***********************************/
 /* Constants for row-based hash */
-#define ZSTD_ROW_HASH_TAG_OFFSET 16     /* byte offset of hashes in the match state's tagTable from the beginning of a row */
-#define ZSTD_ROW_HASH_TAG_BITS 8        /* nb bits to use for the tag */
 #define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1)
 #define ZSTD_ROW_HASH_MAX_ENTRIES 64    /* absolute maximum number of entries per row, for all configurations */
 
···
  * Starting from the LSB, returns the idx of the next non-zero bit.
  * Basically counting the nb of trailing zeroes.
  */
-static U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
-    assert(val != 0);
-#   if (defined(__GNUC__) && ((__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))
-    if (sizeof(size_t) == 4) {
-        U32 mostSignificantWord = (U32)(val >> 32);
-        U32 leastSignificantWord = (U32)val;
-        if (leastSignificantWord == 0) {
-            return 32 + (U32)__builtin_ctz(mostSignificantWord);
-        } else {
-            return (U32)__builtin_ctz(leastSignificantWord);
-        }
-    } else {
-        return (U32)__builtin_ctzll(val);
-    }
-#   else
-    /* Software ctz version: http://aggregate.org/MAGIC/#Trailing%20Zero%20Count
-     * and: https://stackoverflow.com/questions/2709430/count-number-of-bits-in-a-64-bit-long-big-integer
-     */
-    val = ~val & (val - 1ULL); /* Lowest set bit mask */
-    val = val - ((val >> 1) & 0x5555555555555555);
-    val = (val & 0x3333333333333333ULL) + ((val >> 2) & 0x3333333333333333ULL);
-    return (U32)((((val + (val >> 4)) & 0xF0F0F0F0F0F0F0FULL) * 0x101010101010101ULL) >> 56);
-#   endif
-}
-
-/* ZSTD_rotateRight_*():
- * Rotates a bitfield to the right by "count" bits.
- * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
- */
-FORCE_INLINE_TEMPLATE
-U64 ZSTD_rotateRight_U64(U64 const value, U32 count) {
-    assert(count < 64);
-    count &= 0x3F; /* for fickle pattern recognition */
-    return (value >> count) | (U64)(value << ((0U - count) & 0x3F));
-}
-
-FORCE_INLINE_TEMPLATE
-U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
-    assert(count < 32);
-    count &= 0x1F; /* for fickle pattern recognition */
-    return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
-}
-
-FORCE_INLINE_TEMPLATE
-U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
-    assert(count < 16);
-    count &= 0x0F; /* for fickle pattern recognition */
-    return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
+MEM_STATIC U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
+    return ZSTD_countTrailingZeros64(val);
 }
 
 /* ZSTD_row_nextIndex():
  * Returns the next index to insert at within a tagTable row, and updates the "head"
- * value to reflect the update. Essentially cycles backwards from [0, {entries per row})
+ * value to reflect the update. Essentially cycles backwards from [1, {entries per row})
  */
 FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) {
-  U32 const next = (*tagRow - 1) & rowMask;
-  *tagRow = (BYTE)next;
-  return next;
+    U32 next = (*tagRow-1) & rowMask;
+    next += (next == 0) ? rowMask : 0; /* skip first position */
+    *tagRow = (BYTE)next;
+    return next;
 }
 
 /* ZSTD_isAligned():
···
 /* ZSTD_row_prefetch():
  * Performs prefetching for the hashTable and tagTable at a given row.
  */
-FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, U16 const* tagTable, U32 const relRow, U32 const rowLog) {
+FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, BYTE const* tagTable, U32 const relRow, U32 const rowLog) {
     PREFETCH_L1(hashTable + relRow);
     if (rowLog >= 5) {
         PREFETCH_L1(hashTable + relRow + 16);
···
  * Fill up the hash cache starting at idx, prefetching up to ZSTD_ROW_HASH_CACHE_SIZE entries,
  * but not beyond iLimit.
  */
-FORCE_INLINE_TEMPLATE void ZSTD_row_fillHashCache(ZSTD_matchState_t* ms, const BYTE* base,
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_row_fillHashCache(ZSTD_MatchState_t* ms, const BYTE* base,
                                    U32 const rowLog, U32 const mls,
                                    U32 idx, const BYTE* const iLimit)
 {
     U32 const* const hashTable = ms->hashTable;
-    U16 const* const tagTable = ms->tagTable;
+    BYTE const* const tagTable = ms->tagTable;
     U32 const hashLog = ms->rowHashLog;
     U32 const maxElemsToPrefetch = (base + idx) > iLimit ? 0 : (U32)(iLimit - (base + idx) + 1);
     U32 const lim = idx + MIN(ZSTD_ROW_HASH_CACHE_SIZE, maxElemsToPrefetch);
 
     for (; idx < lim; ++idx) {
-        U32 const hash = (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+        U32 const hash = (U32)ZSTD_hashPtrSalted(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt);
         U32 const row = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
         ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
         ms->hashCache[idx & ZSTD_ROW_HASH_CACHE_MASK] = hash;
···
  * Returns the hash of base + idx, and replaces the hash in the hash cache with the byte at
  * base + idx + ZSTD_ROW_HASH_CACHE_SIZE. Also prefetches the appropriate rows from hashTable and tagTable.
  */
-FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable,
-                                                  U16 const* tagTable, BYTE const* base,
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable,
+                                                  BYTE const* tagTable, BYTE const* base,
                                                   U32 idx, U32 const hashLog,
-                                                  U32 const rowLog, U32 const mls)
+                                                  U32 const rowLog, U32 const mls,
+                                                  U64 const hashSalt)
 {
-    U32 const newHash = (U32)ZSTD_hashPtr(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+    U32 const newHash = (U32)ZSTD_hashPtrSalted(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, hashSalt);
     U32 const row = (newHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
     ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
     {   U32 const hash = cache[idx & ZSTD_ROW_HASH_CACHE_MASK];
···
 /* ZSTD_row_update_internalImpl():
  * Updates the hash table with positions starting from updateStartIdx until updateEndIdx.
  */
-FORCE_INLINE_TEMPLATE void ZSTD_row_update_internalImpl(ZSTD_matchState_t* ms,
-                                                        U32 updateStartIdx, U32 const updateEndIdx,
-                                                        U32 const mls, U32 const rowLog,
-                                                        U32 const rowMask, U32 const useCache)
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_row_update_internalImpl(ZSTD_MatchState_t* ms,
+                                  U32 updateStartIdx, U32 const updateEndIdx,
+                                  U32 const mls, U32 const rowLog,
+                                  U32 const rowMask, U32 const useCache)
 {
     U32* const hashTable = ms->hashTable;
-    U16* const tagTable = ms->tagTable;
+    BYTE* const tagTable = ms->tagTable;
     U32 const hashLog = ms->rowHashLog;
     const BYTE* const base = ms->window.base;
 
     DEBUGLOG(6, "ZSTD_row_update_internalImpl(): updateStartIdx=%u, updateEndIdx=%u", updateStartIdx, updateEndIdx);
     for (; updateStartIdx < updateEndIdx; ++updateStartIdx) {
-        U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, updateStartIdx, hashLog, rowLog, mls)
-                                  : (U32)ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+        U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, updateStartIdx, hashLog, rowLog, mls, ms->hashSalt)
+                                  : (U32)ZSTD_hashPtrSalted(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt);
         U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
         U32* const row = hashTable + relRow;
-        BYTE* tagRow = (BYTE*)(tagTable + relRow);  /* Though tagTable is laid out as a table of U16, each tag is only 1 byte.
-                                                       Explicit cast allows us to get exact desired position within each row */
+        BYTE* tagRow = tagTable + relRow;
         U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
 
-        assert(hash == ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls));
-        ((BYTE*)tagRow)[pos + ZSTD_ROW_HASH_TAG_OFFSET] = hash & ZSTD_ROW_HASH_TAG_MASK;
+        assert(hash == ZSTD_hashPtrSalted(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt));
+        tagRow[pos] = hash & ZSTD_ROW_HASH_TAG_MASK;
         row[pos] = updateStartIdx;
     }
 }
···
  * Inserts the byte at ip into the appropriate position in the hash table, and updates ms->nextToUpdate.
  * Skips sections of long matches as is necessary.
  */
-FORCE_INLINE_TEMPLATE void ZSTD_row_update_internal(ZSTD_matchState_t* ms, const BYTE* ip,
-                                                    U32 const mls, U32 const rowLog,
-                                                    U32 const rowMask, U32 const useCache)
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_row_update_internal(ZSTD_MatchState_t* ms, const BYTE* ip,
+                              U32 const mls, U32 const rowLog,
+                              U32 const rowMask, U32 const useCache)
 {
     U32 idx = ms->nextToUpdate;
     const BYTE* const base = ms->window.base;
···
  * External wrapper for ZSTD_row_update_internal(). Used for filling the hashtable during dictionary
  * processing.
  */
-void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) {
+void ZSTD_row_update(ZSTD_MatchState_t* const ms, const BYTE* ip) {
     const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
     const U32 rowMask = (1u << rowLog) - 1;
     const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */);
 
     DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog);
-    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* dont use cache */);
+    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* don't use cache */);
+}
+
+/* Returns the mask width of bits group of which will be set to 1. Given not all
+ * architectures have easy movemask instruction, this helps to iterate over
+ * groups of bits easier and faster.
+ */
+FORCE_INLINE_TEMPLATE U32
+ZSTD_row_matchMaskGroupWidth(const U32 rowEntries)
+{
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+    (void)rowEntries;
+#if defined(ZSTD_ARCH_ARM_NEON)
+    /* NEON path only works for little endian */
+    if (!MEM_isLittleEndian()) {
+        return 1;
+    }
+    if (rowEntries == 16) {
+        return 4;
+    }
+    if (rowEntries == 32) {
+        return 2;
+    }
+    if (rowEntries == 64) {
+        return 1;
+    }
+#endif
+    return 1;
 }
 
 #if defined(ZSTD_ARCH_X86_SSE2)
···
 }
 #endif
 
-/* Returns a ZSTD_VecMask (U32) that has the nth bit set to 1 if the newly-computed "tag" matches
- * the hash at the nth position in a row of the tagTable.
- * Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield
- * to match up with the actual layout of the entries within the hashTable */
+#if defined(ZSTD_ARCH_ARM_NEON)
 FORCE_INLINE_TEMPLATE ZSTD_VecMask
-ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head, const U32 rowEntries)
+ZSTD_row_getNEONMask(const U32 rowEntries, const BYTE* const src, const BYTE tag, const U32 headGrouped)
 {
-    const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET;
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    if (rowEntries == 16) {
+        /* vshrn_n_u16 shifts by 4 every u16 and narrows to 8 lower bits.
+         * After that groups of 4 bits represent the equalMask. We lower
+         * all bits except the highest in these groups by doing AND with
+         * 0x88 = 0b10001000.
+         */
+        const uint8x16_t chunk = vld1q_u8(src);
+        const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
+        const uint8x8_t res = vshrn_n_u16(equalMask, 4);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0);
+        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x8888888888888888ull;
+    } else if (rowEntries == 32) {
+        /* Same idea as with rowEntries == 16 but doing AND with
+         * 0x55 = 0b01010101.
+         */
+        const uint16x8x2_t chunk = vld2q_u16((const uint16_t*)(const void*)src);
+        const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
+        const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
+        const uint8x16_t dup = vdupq_n_u8(tag);
+        const uint8x8_t t0 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk0, dup)), 6);
+        const uint8x8_t t1 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk1, dup)), 6);
+        const uint8x8_t res = vsli_n_u8(t0, t1, 4);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0) ;
+        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x5555555555555555ull;
+    } else { /* rowEntries == 64 */
+        const uint8x16x4_t chunk = vld4q_u8(src);
+        const uint8x16_t dup = vdupq_n_u8(tag);
+        const uint8x16_t cmp0 = vceqq_u8(chunk.val[0], dup);
+        const uint8x16_t cmp1 = vceqq_u8(chunk.val[1], dup);
+        const uint8x16_t cmp2 = vceqq_u8(chunk.val[2], dup);
+        const uint8x16_t cmp3 = vceqq_u8(chunk.val[3], dup);
+
+        const uint8x16_t t0 = vsriq_n_u8(cmp1, cmp0, 1);
+        const uint8x16_t t1 = vsriq_n_u8(cmp3, cmp2, 1);
+        const uint8x16_t t2 = vsriq_n_u8(t1, t0, 2);
+        const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4);
+        const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0);
+        return ZSTD_rotateRight_U64(matches, headGrouped);
+    }
+}
+#endif
+
+/* Returns a ZSTD_VecMask (U64) that has the nth group (determined by
+ * ZSTD_row_matchMaskGroupWidth) of bits set to 1 if the newly-computed "tag"
+ * matches the hash at the nth position in a row of the tagTable.
+ * Each row is a circular buffer beginning at the value of "headGrouped". So we
+ * must rotate the "matches" bitfield to match up with the actual layout of the
+ * entries within the hashTable */
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 headGrouped, const U32 rowEntries)
+{
+    const BYTE* const src = tagRow;
     assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
     assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+    assert(ZSTD_row_matchMaskGroupWidth(rowEntries) * rowEntries <= sizeof(ZSTD_VecMask) * 8);
 
 #if defined(ZSTD_ARCH_X86_SSE2)
 
-    return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, head);
+    return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, headGrouped);
 
 #else /* SW or NEON-LE */
 
 # if defined(ZSTD_ARCH_ARM_NEON)
   /* This NEON path only works for little endian - otherwise use SWAR below */
     if (MEM_isLittleEndian()) {
-        if (rowEntries == 16) {
-            const uint8x16_t chunk = vld1q_u8(src);
-            const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
-            const uint16x8_t t0 = vshlq_n_u16(equalMask, 7);
-            const uint32x4_t t1 = vreinterpretq_u32_u16(vsriq_n_u16(t0, t0, 14));
-            const uint64x2_t t2 = vreinterpretq_u64_u32(vshrq_n_u32(t1, 14));
-            const uint8x16_t t3 = vreinterpretq_u8_u64(vsraq_n_u64(t2, t2, 28));
-            const U16 hi = (U16)vgetq_lane_u8(t3, 8);
-            const U16 lo = (U16)vgetq_lane_u8(t3, 0);
-            return ZSTD_rotateRight_U16((hi << 8) | lo, head);
-        } else if (rowEntries == 32) {
-            const uint16x8x2_t chunk = vld2q_u16((const U16*)(const void*)src);
-            const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
-            const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
-            const uint8x16_t equalMask0 = vceqq_u8(chunk0, vdupq_n_u8(tag));
-            const uint8x16_t equalMask1 = vceqq_u8(chunk1, vdupq_n_u8(tag));
-            const int8x8_t pack0 = vqmovn_s16(vreinterpretq_s16_u8(equalMask0));
-            const int8x8_t pack1 = vqmovn_s16(vreinterpretq_s16_u8(equalMask1));
-            const uint8x8_t t0 = vreinterpret_u8_s8(pack0);
-            const uint8x8_t t1 = vreinterpret_u8_s8(pack1);
-            const uint8x8_t t2 = vsri_n_u8(t1, t0, 2);
-            const uint8x8x2_t t3 = vuzp_u8(t2, t0);
-            const uint8x8_t t4 = vsri_n_u8(t3.val[1], t3.val[0], 4);
-            const U32 matches = vget_lane_u32(vreinterpret_u32_u8(t4), 0);
-            return ZSTD_rotateRight_U32(matches, head);
-        } else { /* rowEntries == 64 */
-            const uint8x16x4_t chunk = vld4q_u8(src);
-            const uint8x16_t dup = vdupq_n_u8(tag);
-            const uint8x16_t cmp0 = vceqq_u8(chunk.val[0], dup);
-            const uint8x16_t cmp1 = vceqq_u8(chunk.val[1], dup);
-            const uint8x16_t cmp2 = vceqq_u8(chunk.val[2], dup);
-            const uint8x16_t cmp3 = vceqq_u8(chunk.val[3], dup);
-
-            const uint8x16_t t0 = vsriq_n_u8(cmp1, cmp0, 1);
-            const uint8x16_t t1 = vsriq_n_u8(cmp3, cmp2, 1);
-            const uint8x16_t t2 = vsriq_n_u8(t1, t0, 2);
-            const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4);
-            const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4);
-            const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0);
-            return ZSTD_rotateRight_U64(matches, head);
-        }
+        return ZSTD_row_getNEONMask(rowEntries, src, tag, headGrouped);
     }
 # endif /* ZSTD_ARCH_ARM_NEON */
     /* SWAR */
-    {   const size_t chunkSize = sizeof(size_t);
+    {   const int chunkSize = sizeof(size_t);
         const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
         const size_t xFF = ~((size_t)0);
         const size_t x01 = xFF / 0xFF;
···
         }
         matches = ~matches;
         if (rowEntries == 16) {
-            return ZSTD_rotateRight_U16((U16)matches, head);
+            return ZSTD_rotateRight_U16((U16)matches, headGrouped);
         } else if (rowEntries == 32) {
-            return ZSTD_rotateRight_U32((U32)matches, head);
+            return ZSTD_rotateRight_U32((U32)matches, headGrouped);
         } else {
-            return ZSTD_rotateRight_U64((U64)matches, head);
+            return ZSTD_rotateRight_U64((U64)matches, headGrouped);
         }
     }
 #endif
···
 
 /* The high-level approach of the SIMD row based match finder is as follows:
  * - Figure out where to insert the new entry:
- *      - Generate a hash from a byte along with an additional 1-byte "short hash". The additional byte is our "tag"
- *      - The hashTable is effectively split into groups or "rows" of 16 or 32 entries of U32, and the hash determines
+ *      - Generate a hash for current input position and split it into a one byte of tag and `rowHashLog` bits of index.
+ *           - The hash is salted by a value that changes on every context reset, so when the same table is used
+ *             we will avoid collisions that would otherwise slow us down by introducing phantom matches.
+ *      - The hashTable is effectively split into groups or "rows" of 15 or 31 entries of U32, and the index determines
  *        which row to insert into.
- *      - Determine the correct position within the row to insert the entry into. Each row of 16 or 32 can
- *        be considered as a circular buffer with a "head" index that resides in the tagTable.
- *      - Also insert the "tag" into the equivalent row and position in the tagTable.
- *          - Note: The tagTable has 17 or 33 1-byte entries per row, due to 16 or 32 tags, and 1 "head" entry.
- *                  The 17 or 33 entry rows are spaced out to occur every 32 or 64 bytes, respectively,
- *                  for alignment/performance reasons, leaving some bytes unused.
- * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte "short hash" and
+ *      - Determine the correct position within the row to insert the entry into. Each row of 15 or 31 can
+ *        be considered as a circular buffer with a "head" index that resides in the tagTable (overall 16 or 32 bytes
+ *        per row).
+ * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte tag calculated for the position and
  *   generate a bitfield that we can cycle through to check the collisions in the hash table.
  * - Pick the longest match.
+ * - Insert the tag into the equivalent row and position in the tagTable.
  */
 FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_RowFindBestMatch(
-                        ZSTD_matchState_t* ms,
+                        ZSTD_MatchState_t* ms,
                         const BYTE* const ip, const BYTE* const iLimit,
                         size_t* offsetPtr,
                         const U32 mls, const ZSTD_dictMode_e dictMode,
                         const U32 rowLog)
 {
     U32* const hashTable = ms->hashTable;
-    U16* const tagTable = ms->tagTable;
+    BYTE* const tagTable = ms->tagTable;
     U32* const hashCache = ms->hashCache;
     const U32 hashLog = ms->rowHashLog;
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
···
     const U32 rowEntries = (1U << rowLog);
     const U32 rowMask = rowEntries - 1;
     const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */
+    const U32 groupWidth = ZSTD_row_matchMaskGroupWidth(rowEntries);
+    const U64 hashSalt = ms->hashSalt;
     U32 nbAttempts = 1U << cappedSearchLog;
     size_t ml=4-1;
+    U32 hash;
 
     /* DMS/DDS variables that may be referenced laster */
-    const ZSTD_matchState_t* const dms = ms->dictMatchState;
+    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
 
     /* Initialize the following variables to satisfy static analyzer */
     size_t ddsIdx = 0;
···
     if (dictMode == ZSTD_dictMatchState) {
         /* Prefetch DMS rows */
         U32* const dmsHashTable = dms->hashTable;
-        U16* const dmsTagTable = dms->tagTable;
+        BYTE* const dmsTagTable = dms->tagTable;
         U32 const dmsHash = (U32)ZSTD_hashPtr(ip, dms->rowHashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
         U32 const dmsRelRow = (dmsHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
         dmsTag = dmsHash & ZSTD_ROW_HASH_TAG_MASK;
···
     }
 
     /* Update the hashTable and tagTable up to (but not including) ip */
-    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */);
+    if (!ms->lazySkipping) {
+        ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */);
+        hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls, hashSalt);
+    } else {
+        /* Stop inserting every position when in the lazy skipping mode.
+         * The hash cache is also not kept up to date in this mode.
+         */
+        hash = (U32)ZSTD_hashPtrSalted(ip, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, hashSalt);
+        ms->nextToUpdate = curr;
+    }
+    ms->hashSaltEntropy += hash; /* collect salt entropy */
+
     {   /* Get the hash for ip, compute the appropriate row */
-        U32 const hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls);
         U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
         U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK;
         U32* const row = hashTable + relRow;
         BYTE* tagRow = (BYTE*)(tagTable + relRow);
-        U32 const head = *tagRow & rowMask;
+        U32 const headGrouped = (*tagRow & rowMask) * groupWidth;
         U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
         size_t numMatches = 0;
         size_t currMatch = 0;
-        ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, head, rowEntries);
+        ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, headGrouped, rowEntries);
 
         /* Cycle through the matches and prefetch */
-        for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
-            U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
+        for (; (matches > 0) && (nbAttempts > 0); matches &= (matches - 1)) {
+            U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
             U32 const matchIndex = row[matchPos];
+            if(matchPos == 0) continue;
             assert(numMatches < rowEntries);
             if (matchIndex < lowLimit)
                 break;
···
                 PREFETCH_L1(dictBase + matchIndex);
             }
             matchBuffer[numMatches++] = matchIndex;
+            --nbAttempts;
         }
 
         /* Speed opt: insert current byte into hashtable too. This allows us to avoid one iteration of the loop
            in ZSTD_row_update_internal() at the next search. */
         {
             U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
-            tagRow[pos + ZSTD_ROW_HASH_TAG_OFFSET] = (BYTE)tag;
+            tagRow[pos] = (BYTE)tag;
             row[pos] = ms->nextToUpdate++;
         }
 
···
             if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
                 const BYTE* const match = base + matchIndex;
                 assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
-                if (match[ml] == ip[ml])   /* potentially better */
+                /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
+                if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
                     currentMl = ZSTD_count(ip, match, iLimit);
             } else {
                 const BYTE* const match = dictBase + matchIndex;
···
             /* Save best solution */
             if (currentMl > ml) {
                 ml = currentMl;
-                *offsetPtr = STORE_OFFSET(curr - matchIndex);
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
             }
         }
···
         const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
         const U32 dmsIndexDelta        = dictLimit - dmsSize;
 
-        {   U32 const head = *dmsTagRow & rowMask;
+        {   U32 const headGrouped = (*dmsTagRow & rowMask) * groupWidth;
             U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
             size_t numMatches = 0;
             size_t currMatch = 0;
-            ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, head, rowEntries);
+            ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, headGrouped, rowEntries);
 
-            for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
-                U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
+            for (; (matches > 0) && (nbAttempts > 0); matches &= (matches - 1)) {
+                U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
                 U32 const matchIndex = dmsRow[matchPos];
+                if(matchPos == 0) continue;
                 if (matchIndex < dmsLowestIndex)
                     break;
                 PREFETCH_L1(dmsBase + matchIndex);
                 matchBuffer[numMatches++] = matchIndex;
+                --nbAttempts;
             }
 
             /* Return the longest match */
···
                 if (currentMl > ml) {
                     ml = currentMl;
                     assert(curr > matchIndex + dmsIndexDelta);
-                    *offsetPtr = STORE_OFFSET(curr - (matchIndex + dmsIndexDelta));
+                    *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
                     if (ip+currentMl == iLimit) break;
                 }
             }
···
  * ZSTD_searchMax() dispatches to the correct implementation function.
  *
  * TODO: The start of the search function involves loading and calculating a
- * bunch of constants from the ZSTD_matchState_t. These computations could be
+ * bunch of constants from the ZSTD_MatchState_t. These computations could be
  * done in an initialization function, and saved somewhere in the match state.
  * Then we could pass a pointer to the saved state instead of the match state,
  * and avoid duplicate computations.
···
 
 #define GEN_ZSTD_BT_SEARCH_FN(dictMode, mls)                                           \
     ZSTD_SEARCH_FN_ATTRS size_t ZSTD_BT_SEARCH_FN(dictMode, mls)(                      \
-            ZSTD_matchState_t* ms,                                                     \
+            ZSTD_MatchState_t* ms,                                                     \
             const BYTE* ip, const BYTE* const iLimit,                                  \
             size_t* offBasePtr)                                                        \
     {                                                                                  \
···
 
 #define GEN_ZSTD_HC_SEARCH_FN(dictMode, mls)                                          \
     ZSTD_SEARCH_FN_ATTRS size_t ZSTD_HC_SEARCH_FN(dictMode, mls)(                     \
-            ZSTD_matchState_t* ms,                                                    \
+            ZSTD_MatchState_t* ms,                                                    \
             const BYTE* ip, const BYTE* const iLimit,                                 \
             size_t* offsetPtr)                                                        \
     {                                                                                 \
···
 
 #define GEN_ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)                                          \
     ZSTD_SEARCH_FN_ATTRS size_t ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(                     \
-            ZSTD_matchState_t* ms,                                                             \
+            ZSTD_MatchState_t* ms,                                                             \
             const BYTE* ip, const BYTE* const iLimit,                                          \
             size_t* offsetPtr)                                                                 \
     {                                                                                          \
···
  * If a match is found its offset is stored in @p offsetPtr.
  */
 FORCE_INLINE_TEMPLATE size_t ZSTD_searchMax(
-    ZSTD_matchState_t* ms,
+    ZSTD_MatchState_t* ms,
     const BYTE* ip,
     const BYTE* iend,
     size_t* offsetPtr,
···
 *  Common parser - lazy strategy
 *********************************/
 
-FORCE_INLINE_TEMPLATE size_t
-ZSTD_compressBlock_lazy_generic(
-                        ZSTD_matchState_t* ms, seqStore_t* seqStore,
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_compressBlock_lazy_generic(
+                        ZSTD_MatchState_t* ms, SeqStore_t* seqStore,
                         U32 rep[ZSTD_REP_NUM],
                         const void* src, size_t srcSize,
                         const searchMethod_e searchMethod, const U32 depth,
···
     const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
     const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
 
-    U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
+    U32 offset_1 = rep[0], offset_2 = rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
     const int isDMS = dictMode == ZSTD_dictMatchState;
     const int isDDS = dictMode == ZSTD_dedicatedDictSearch;
     const int isDxS = isDMS || isDDS;
-    const ZSTD_matchState_t* const dms = ms->dictMatchState;
+    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
     const U32 dictLowestIndex      = isDxS ? dms->window.dictLimit : 0;
     const BYTE* const dictBase     = isDxS ? dms->window.base : NULL;
     const BYTE* const dictLowest   = isDxS ? dictBase + dictLowestIndex : NULL;
···
         U32 const curr = (U32)(ip - base);
         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog);
         U32 const maxRep = curr - windowLow;
-        if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
+        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
     }
     if (isDxS) {
         /* dictMatchState repCode checks don't currently handle repCode == 0
···
         assert(offset_2 <= dictAndPrefixLength);
     }
 
+    /* Reset the lazy skipping state */
+    ms->lazySkipping = 0;
+
     if (searchMethod == search_rowHash) {
-        ZSTD_row_fillHashCache(ms, base, rowLog,
-                            MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
-                            ms->nextToUpdate, ilimit);
+        ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
     }
 
     /* Match Loop */
···
 #endif
     while (ip < ilimit) {
         size_t matchLength=0;
-        size_t offcode=STORE_REPCODE_1;
+        size_t offBase = REPCODE1_TO_OFFBASE;
         const BYTE* start=ip+1;
         DEBUGLOG(7, "search baseline (depth 0)");
 
···
                                 && repIndex < prefixLowestIndex) ?
                                    dictBase + (repIndex - dictIndexDelta) :
                                    base + repIndex;
-            if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
+            if ((ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
                 && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
                 const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                 matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
···
         }
 
         /* first search (depth 0) */
-        {   size_t offsetFound = 999999999;
-            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offsetFound, mls, rowLog, searchMethod, dictMode);
+        {   size_t offbaseFound = 999999999;
+            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offbaseFound, mls, rowLog, searchMethod, dictMode);
             if (ml2 > matchLength)
-                matchLength = ml2, start = ip, offcode=offsetFound;
+                matchLength = ml2, start = ip, offBase = offbaseFound;
         }
 
         if (matchLength < 4) {
-            ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
+            size_t const step = ((size_t)(ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */;
+            ip += step;
+            /* Enter the lazy skipping mode once we are skipping more than 8 bytes at a time.
+             * In this mode we stop inserting every position into our tables, and only insert
+             * positions that we search, which is one in step positions.
+             * The exact cutoff is flexible, I've just chosen a number that is reasonably high,
+             * so we minimize the compression ratio loss in "normal" scenarios. This mode gets
+             * triggered once we've gone 2KB without finding any matches.
+             */
+            ms->lazySkipping = step > kLazySkippingStep;
             continue;
         }
 
···
             DEBUGLOG(7, "search depth 1");
             ip ++;
             if ( (dictMode == ZSTD_noDict)
-              && (offcode) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+              && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                 size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                 int const gain2 = (int)(mlRep * 3);
-                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                 if ((mlRep >= 4) && (gain2 > gain1))
-                    matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
+                    matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
             }
             if (isDxS) {
                 const U32 repIndex = (U32)(ip - base) - offset_1;
                 const BYTE* repMatch = repIndex < prefixLowestIndex ?
                                dictBase + (repIndex - dictIndexDelta) :
                                base + repIndex;
-                if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
+                if ((ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
                     && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                     const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                     size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                     int const gain2 = (int)(mlRep * 3);
-                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((mlRep >= 4) && (gain2 > gain1))
-                        matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
+                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
             }
-            {   size_t offset2=999999999;
-                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offset2, mls, rowLog, searchMethod, dictMode);
-                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
-                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 4);
+            {   size_t ofbCandidate=999999999;
+                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                 if ((ml2 >= 4) && (gain2 > gain1)) {
-                    matchLength = ml2, offcode = offset2, start = ip;
+                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                     continue;   /* search a better one */
             }   }
 
···
                 DEBUGLOG(7, "search depth 2");
                 ip ++;
                 if ( (dictMode == ZSTD_noDict)
-                  && (offcode) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+                  && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                     size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                     int const gain2 = (int)(mlRep * 4);
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((mlRep >= 4) && (gain2 > gain1))
-                        matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
+                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
                 if (isDxS) {
                     const U32 repIndex = (U32)(ip - base) - offset_1;
                     const BYTE* repMatch = repIndex < prefixLowestIndex ?
                                    dictBase + (repIndex - dictIndexDelta) :
                                    base + repIndex;
-                    if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
+                    if ((ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
                         && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                         const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                         size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                         int const gain2 = (int)(mlRep * 4);
-                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                         if ((mlRep >= 4) && (gain2 > gain1))
-                            matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
+                            matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                     }
                 }
-                {   size_t offset2=999999999;
-                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offset2, mls, rowLog, searchMethod, dictMode);
-                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 7);
+                {   size_t ofbCandidate=999999999;
+                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                     if ((ml2 >= 4) && (gain2 > gain1)) {
-                        matchLength = ml2, offcode = offset2, start = ip;
+                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                         continue;
             }   }   }
             break;  /* nothing found : store previous solution */
···
          * notably if `value` is unsigned, resulting in a large positive `-value`.
          */
         /* catch up */
-        if (STORED_IS_OFFSET(offcode)) {
+        if (OFFBASE_IS_OFFSET(offBase)) {
             if (dictMode == ZSTD_noDict) {
-                while ( ((start > anchor) & (start - STORED_OFFSET(offcode) > prefixLowest))
-                     && (start[-1] == (start-STORED_OFFSET(offcode))[-1]) )  /* only search for offset within prefix */
+                while ( ((start > anchor) & (start - OFFBASE_TO_OFFSET(offBase) > prefixLowest))
+                     && (start[-1] == (start-OFFBASE_TO_OFFSET(offBase))[-1]) )  /* only search for offset within prefix */
                     { start--; matchLength++; }
             }
             if (isDxS) {
-                U32 const matchIndex = (U32)((size_t)(start-base) - STORED_OFFSET(offcode));
+                U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
                 const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
                 const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
                 while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
             }
-            offset_2 = offset_1; offset_1 = (U32)STORED_OFFSET(offcode);
+            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
         }
         /* store sequence */
 _storeSequence:
         {   size_t const litLength = (size_t)(start - anchor);
-            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offcode, matchLength);
+            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
             anchor = ip = start + matchLength;
+        }
+        if (ms->lazySkipping) {
+            /* We've found a match, disable lazy skipping mode, and refill the hash cache. */
+            if (searchMethod == search_rowHash) {
+                ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
+            }
+            ms->lazySkipping = 0;
         }
 
         /* check immediate repcode */
···
                 const BYTE* repMatch = repIndex < prefixLowestIndex ?
                         dictBase - dictIndexDelta + repIndex :
                         base + repIndex;
-                if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex) >= 3 /* intentional overflow */)
+                if ( (ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
                    && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
                     matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
-                    offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
+                    offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset_2 <=> offset_1 */
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                     ip += matchLength;
                     anchor = ip;
                     continue;
···
                  && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
                 /* store sequence */
                 matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
-                offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode; /* swap repcodes */
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
+                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase; /* swap repcodes */
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                 ip += matchLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
     }   }   }
 
-    /* Save reps for next block */
-    rep[0] = offset_1 ? offset_1 : savedOffset;
-    rep[1] = offset_2 ? offset_2 : savedOffset;
+    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
+
+    /* save reps for next block */
+    rep[0] = offset_1 ? offset_1 : offsetSaved1;
+    rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
+#endif /* build exclusions */
 
 
-size_t ZSTD_compressBlock_btlazy2(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_noDict);
-}
-
-size_t ZSTD_compressBlock_lazy2(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_noDict);
-}
-
-size_t ZSTD_compressBlock_lazy(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_noDict);
-}
-
+#ifndef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_greedy(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_noDict);
 }
 
-size_t ZSTD_compressBlock_btlazy2_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_dictMatchState);
-}
-
-size_t ZSTD_compressBlock_lazy2_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dictMatchState);
-}
-
-size_t ZSTD_compressBlock_lazy_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dictMatchState);
-}
-
 size_t ZSTD_compressBlock_greedy_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dictMatchState);
 }
 
-
-size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch);
-}
-
-size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch);
-}
-
 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch);
 }
 
-/* Row-based matchfinder */
-size_t ZSTD_compressBlock_lazy2_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict);
-}
-
-size_t ZSTD_compressBlock_lazy_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict);
-}
-
 size_t ZSTD_compressBlock_greedy_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict);
 }
 
-size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState);
-}
-
-size_t ZSTD_compressBlock_lazy_dictMatchState_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState);
-}
-
 size_t ZSTD_compressBlock_greedy_dictMatchState_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState);
 }
 
-
-size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch);
-}
-
-size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch);
-}
-
 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch);
 }
+#endif
 
+#ifndef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_lazy(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_lazy_dictMatchState(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dictMatchState);
+}
+
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch);
+}
+
+size_t ZSTD_compressBlock_lazy_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_lazy_dictMatchState_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState);
+}
+
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_lazy2(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_lazy2_dictMatchState(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dictMatchState);
+}
+
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch);
+}
+
+size_t ZSTD_compressBlock_lazy2_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState);
+}
+
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_btlazy2(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_btlazy2_dictMatchState(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_dictMatchState);
+}
+#endif
+
+#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR)
 FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_compressBlock_lazy_extDict_generic(
-                        ZSTD_matchState_t* ms, seqStore_t* seqStore,
+                        ZSTD_MatchState_t* ms, SeqStore_t* seqStore,
                         U32 rep[ZSTD_REP_NUM],
                         const void* src, size_t srcSize,
                         const searchMethod_e searchMethod, const U32 depth)
···
 
     DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod);
 
+    /* Reset the lazy skipping state */
+    ms->lazySkipping = 0;
+
     /* init */
     ip += (ip == prefixStart);
     if (searchMethod == search_rowHash) {
-        ZSTD_row_fillHashCache(ms, base, rowLog,
-                               MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
-                               ms->nextToUpdate, ilimit);
+        ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
     }
 
     /* Match Loop */
···
 #endif
     while (ip < ilimit) {
         size_t matchLength=0;
-        size_t offcode=STORE_REPCODE_1;
+        size_t offBase = REPCODE1_TO_OFFBASE;
         const BYTE* start=ip+1;
         U32 curr = (U32)(ip-base);
 
···
             const U32 repIndex = (U32)(curr+1 - offset_1);
             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
             const BYTE* const repMatch = repBase + repIndex;
-            if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow */
+            if ( (ZSTD_index_overlap_check(dictLimit, repIndex))
                & (offset_1 <= curr+1 - windowLow) ) /* note: we are searching at curr+1 */
             if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
                 /* repcode detected we should take it */
···
         }   }
 
         /* first search (depth 0) */
-        {   size_t offsetFound = 999999999;
-            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offsetFound, mls, rowLog, searchMethod, ZSTD_extDict);
+        {   size_t ofbCandidate = 999999999;
+            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
             if (ml2 > matchLength)
-                matchLength = ml2, start = ip, offcode=offsetFound;
+                matchLength = ml2, start = ip, offBase = ofbCandidate;
         }
 
         if (matchLength < 4) {
-            ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
+            size_t const step = ((size_t)(ip-anchor) >> kSearchStrength);
+            ip += step + 1;   /* jump faster over incompressible sections */
+            /* Enter the lazy skipping mode once we are skipping more than 8 bytes at a time.
+             * In this mode we stop inserting every position into our tables, and only insert
+             * positions that we search, which is one in step positions.
+             * The exact cutoff is flexible, I've just chosen a number that is reasonably high,
+             * so we minimize the compression ratio loss in "normal" scenarios. This mode gets
+             * triggered once we've gone 2KB without finding any matches.
+             */
+            ms->lazySkipping = step > kLazySkippingStep;
             continue;
         }
 
···
             ip ++;
             curr++;
             /* check repCode */
-            if (offcode) {
+            if (offBase) {
                 const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
                 const U32 repIndex = (U32)(curr - offset_1);
                 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                 const BYTE* const repMatch = repBase + repIndex;
-                if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+                if ( (ZSTD_index_overlap_check(dictLimit, repIndex))
                    & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
                 if (MEM_read32(ip) == MEM_read32(repMatch)) {
                     /* repcode detected */
                     const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                     size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                     int const gain2 = (int)(repLength * 3);
-                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((repLength >= 4) && (gain2 > gain1))
-                        matchLength = repLength, offcode = STORE_REPCODE_1, start = ip;
+                        matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
             }   }
 
             /* search match, depth 1 */
-            {   size_t offset2=999999999;
-                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offset2, mls, rowLog, searchMethod, ZSTD_extDict);
-                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
-                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 4);
+            {   size_t ofbCandidate = 999999999;
+                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                 if ((ml2 >= 4) && (gain2 > gain1)) {
-                    matchLength = ml2, offcode = offset2, start = ip;
+                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                     continue;   /* search a better one */
             }   }
 
···
                 ip ++;
                 curr++;
                 /* check repCode */
-                if (offcode) {
+                if (offBase) {
                     const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
                     const U32 repIndex = (U32)(curr - offset_1);
                     const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                     const BYTE* const repMatch = repBase + repIndex;
-                    if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+                    if ( (ZSTD_index_overlap_check(dictLimit, repIndex))
                        & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
                     if (MEM_read32(ip) == MEM_read32(repMatch)) {
                         /* repcode detected */
                         const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                         size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                         int const gain2 = (int)(repLength * 4);
-                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                         if ((repLength >= 4) && (gain2 > gain1))
-                            matchLength = repLength, offcode = STORE_REPCODE_1, start = ip;
+                            matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }   }
 
                 /* search match, depth 2 */
-                {   size_t offset2=999999999;
-                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offset2, mls, rowLog, searchMethod, ZSTD_extDict);
-                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 7);
+                {   size_t ofbCandidate = 999999999;
+                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                     if ((ml2 >= 4) && (gain2 > gain1)) {
-                        matchLength = ml2, offcode = offset2, start = ip;
+                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                         continue;
             }   }   }
             break;  /* nothing found : store previous solution */
         }
 
         /* catch up */
-        if (STORED_IS_OFFSET(offcode)) {
-            U32 const matchIndex = (U32)((size_t)(start-base) - STORED_OFFSET(offcode));
+        if (OFFBASE_IS_OFFSET(offBase)) {
+            U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
             const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
             const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
             while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
-            offset_2 = offset_1; offset_1 = (U32)STORED_OFFSET(offcode);
+            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
         }
 
         /* store sequence */
 _storeSequence:
         {   size_t const litLength = (size_t)(start - anchor);
-            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offcode, matchLength);
+            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
             anchor = ip = start + matchLength;
+        }
+        if (ms->lazySkipping) {
+            /* We've found a match, disable lazy skipping mode, and refill the hash cache. */
+            if (searchMethod == search_rowHash) {
+                ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
+            }
+            ms->lazySkipping = 0;
         }
 
         /* check immediate repcode */
···
             const U32 repIndex = repCurrent - offset_2;
             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
             const BYTE* const repMatch = repBase + repIndex;
-            if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+            if ( (ZSTD_index_overlap_check(dictLimit, repIndex))
                & (offset_2 <= repCurrent - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
             if (MEM_read32(ip) == MEM_read32(repMatch)) {
                 /* repcode detected we should take it */
                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                 matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
-                offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode;   /* swap offset history */
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
+                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset history */
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                 ip += matchLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
···
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
+#endif /* build exclusions */
 
-
+#ifndef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_greedy_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0);
 }
 
+size_t ZSTD_compressBlock_greedy_extDict_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_lazy_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 
 {
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1);
 }
 
+size_t ZSTD_compressBlock_lazy_extDict_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_lazy2_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 
 {
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2);
 }
 
+size_t ZSTD_compressBlock_lazy2_extDict_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_btlazy2_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 
 {
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
 }
-
-size_t ZSTD_compressBlock_greedy_extDict_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0);
-}
-
-size_t ZSTD_compressBlock_lazy_extDict_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-
-{
-    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1);
-}
-
-size_t ZSTD_compressBlock_lazy2_extDict_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize)
-
-{
-    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2);
-}
+#endif

lib/zstd/compress/zstd_lazy.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #ifndef ZSTD_LAZY_H
 #define ZSTD_LAZY_H
 
-
 #include "zstd_compress_internal.h"
 
 /*
···
  */
 #define ZSTD_LAZY_DDSS_BUCKET_LOG 2
 
-U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip);
-void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip);
+#define ZSTD_ROW_HASH_TAG_BITS 8        /* nb bits to use for the tag */
 
-void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip);
+#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR)
+U32 ZSTD_insertAndFindFirstIndex(ZSTD_MatchState_t* ms, const BYTE* ip);
+void ZSTD_row_update(ZSTD_MatchState_t* const ms, const BYTE* ip);
+
+void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_MatchState_t* ms, const BYTE* const ip);
 
 void ZSTD_preserveUnsortedMark (U32* const table, U32 const size, U32 const reducerValue);  /*! used in ZSTD_reduceIndex(). preemptively increase value of ZSTD_DUBT_UNSORTED_MARK */
+#endif
 
-size_t ZSTD_compressBlock_btlazy2(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy2(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
+#ifndef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_greedy(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy2_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_greedy_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-
-size_t ZSTD_compressBlock_btlazy2_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy2_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_greedy_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy_dictMatchState_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_greedy_dictMatchState_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-
-size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
-
 size_t ZSTD_compressBlock_greedy_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_lazy2_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_greedy_extDict_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+
+#define ZSTD_COMPRESSBLOCK_GREEDY ZSTD_compressBlock_greedy
+#define ZSTD_COMPRESSBLOCK_GREEDY_ROW ZSTD_compressBlock_greedy_row
+#define ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE ZSTD_compressBlock_greedy_dictMatchState
+#define ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE_ROW ZSTD_compressBlock_greedy_dictMatchState_row
+#define ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH ZSTD_compressBlock_greedy_dedicatedDictSearch
+#define ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH_ROW ZSTD_compressBlock_greedy_dedicatedDictSearch_row
+#define ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT ZSTD_compressBlock_greedy_extDict
+#define ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT_ROW ZSTD_compressBlock_greedy_extDict_row
+#else
+#define ZSTD_COMPRESSBLOCK_GREEDY NULL
+#define ZSTD_COMPRESSBLOCK_GREEDY_ROW NULL
+#define ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE NULL
+#define ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE_ROW NULL
+#define ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH NULL
+#define ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH_ROW NULL
+#define ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT NULL
+#define ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT_ROW NULL
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_lazy(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dictMatchState(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dictMatchState_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_extDict(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_lazy_extDict_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+
+#define ZSTD_COMPRESSBLOCK_LAZY ZSTD_compressBlock_lazy
+#define ZSTD_COMPRESSBLOCK_LAZY_ROW ZSTD_compressBlock_lazy_row
+#define ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE ZSTD_compressBlock_lazy_dictMatchState
+#define ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE_ROW ZSTD_compressBlock_lazy_dictMatchState_row
+#define ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH ZSTD_compressBlock_lazy_dedicatedDictSearch
+#define ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH_ROW ZSTD_compressBlock_lazy_dedicatedDictSearch_row
+#define ZSTD_COMPRESSBLOCK_LAZY_EXTDICT ZSTD_compressBlock_lazy_extDict
+#define ZSTD_COMPRESSBLOCK_LAZY_EXTDICT_ROW ZSTD_compressBlock_lazy_extDict_row
+#else
+#define ZSTD_COMPRESSBLOCK_LAZY NULL
+#define ZSTD_COMPRESSBLOCK_LAZY_ROW NULL
+#define ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE NULL
+#define ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE_ROW NULL
+#define ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH NULL
+#define ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH_ROW NULL
+#define ZSTD_COMPRESSBLOCK_LAZY_EXTDICT NULL
+#define ZSTD_COMPRESSBLOCK_LAZY_EXTDICT_ROW NULL
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_lazy2(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dictMatchState(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_extDict(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_lazy2_extDict_row(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+
+#define ZSTD_COMPRESSBLOCK_LAZY2 ZSTD_compressBlock_lazy2
+#define ZSTD_COMPRESSBLOCK_LAZY2_ROW ZSTD_compressBlock_lazy2_row
+#define ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE ZSTD_compressBlock_lazy2_dictMatchState
+#define ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE_ROW ZSTD_compressBlock_lazy2_dictMatchState_row
+#define ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH ZSTD_compressBlock_lazy2_dedicatedDictSearch
+#define ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH_ROW ZSTD_compressBlock_lazy2_dedicatedDictSearch_row
+#define ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT ZSTD_compressBlock_lazy2_extDict
+#define ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT_ROW ZSTD_compressBlock_lazy2_extDict_row
+#else
+#define ZSTD_COMPRESSBLOCK_LAZY2 NULL
+#define ZSTD_COMPRESSBLOCK_LAZY2_ROW NULL
+#define ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE NULL
+#define ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE_ROW NULL
+#define ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH NULL
+#define ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH_ROW NULL
+#define ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT NULL
+#define ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT_ROW NULL
+#endif
+
+#ifndef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_btlazy2(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_btlazy2_dictMatchState(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_btlazy2_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
-        
 
+#define ZSTD_COMPRESSBLOCK_BTLAZY2 ZSTD_compressBlock_btlazy2
+#define ZSTD_COMPRESSBLOCK_BTLAZY2_DICTMATCHSTATE ZSTD_compressBlock_btlazy2_dictMatchState
+#define ZSTD_COMPRESSBLOCK_BTLAZY2_EXTDICT ZSTD_compressBlock_btlazy2_extDict
+#else
+#define ZSTD_COMPRESSBLOCK_BTLAZY2 NULL
+#define ZSTD_COMPRESSBLOCK_BTLAZY2_DICTMATCHSTATE NULL
+#define ZSTD_COMPRESSBLOCK_BTLAZY2_EXTDICT NULL
+#endif
 
 #endif /* ZSTD_LAZY_H */

lib/zstd/compress/zstd_ldm.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #include "zstd_double_fast.h"   /* ZSTD_fillDoubleHashTable() */
 #include "zstd_ldm_geartab.h"
 
-#define LDM_BUCKET_SIZE_LOG 3
+#define LDM_BUCKET_SIZE_LOG 4
 #define LDM_MIN_MATCH_LENGTH 64
 #define LDM_HASH_RLOG 7
 
···
 }
 
 void ZSTD_ldm_adjustParameters(ldmParams_t* params,
-                               ZSTD_compressionParameters const* cParams)
+                        const ZSTD_compressionParameters* cParams)
 {
     params->windowLog = cParams->windowLog;
     ZSTD_STATIC_ASSERT(LDM_BUCKET_SIZE_LOG <= ZSTD_LDM_BUCKETSIZELOG_MAX);
     DEBUGLOG(4, "ZSTD_ldm_adjustParameters");
-    if (!params->bucketSizeLog) params->bucketSizeLog = LDM_BUCKET_SIZE_LOG;
-    if (!params->minMatchLength) params->minMatchLength = LDM_MIN_MATCH_LENGTH;
-    if (params->hashLog == 0) {
-        params->hashLog = MAX(ZSTD_HASHLOG_MIN, params->windowLog - LDM_HASH_RLOG);
-        assert(params->hashLog <= ZSTD_HASHLOG_MAX);
-    }
     if (params->hashRateLog == 0) {
-        params->hashRateLog = params->windowLog < params->hashLog
-                                   ? 0
-                                   : params->windowLog - params->hashLog;
+        if (params->hashLog > 0) {
+            /* if params->hashLog is set, derive hashRateLog from it */
+            assert(params->hashLog <= ZSTD_HASHLOG_MAX);
+            if (params->windowLog > params->hashLog) {
+                params->hashRateLog = params->windowLog - params->hashLog;
+            }
+        } else {
+            assert(1 <= (int)cParams->strategy && (int)cParams->strategy <= 9);
+            /* mapping from [fast, rate7] to [btultra2, rate4] */
+            params->hashRateLog = 7 - (cParams->strategy/3);
+        }
+    }
+    if (params->hashLog == 0) {
+        params->hashLog = BOUNDED(ZSTD_HASHLOG_MIN, params->windowLog - params->hashRateLog, ZSTD_HASHLOG_MAX);
+    }
+    if (params->minMatchLength == 0) {
+        params->minMatchLength = LDM_MIN_MATCH_LENGTH;
+        if (cParams->strategy >= ZSTD_btultra)
+            params->minMatchLength /= 2;
+    }
+    if (params->bucketSizeLog==0) {
+        assert(1 <= (int)cParams->strategy && (int)cParams->strategy <= 9);
+        params->bucketSizeLog = BOUNDED(LDM_BUCKET_SIZE_LOG, (U32)cParams->strategy, ZSTD_LDM_BUCKETSIZELOG_MAX);
     }
     params->bucketSizeLog = MIN(params->bucketSizeLog, params->hashLog);
 }
···
 /* ZSTD_ldm_getBucket() :
  *  Returns a pointer to the start of the bucket associated with hash. */
 static ldmEntry_t* ZSTD_ldm_getBucket(
-        ldmState_t* ldmState, size_t hash, ldmParams_t const ldmParams)
+        const ldmState_t* ldmState, size_t hash, U32 const bucketSizeLog)
 {
-    return ldmState->hashTable + (hash << ldmParams.bucketSizeLog);
+    return ldmState->hashTable + (hash << bucketSizeLog);
 }
 
 /* ZSTD_ldm_insertEntry() :
  *  Insert the entry with corresponding hash into the hash table */
 static void ZSTD_ldm_insertEntry(ldmState_t* ldmState,
                                  size_t const hash, const ldmEntry_t entry,
-                                 ldmParams_t const ldmParams)
+                                 U32 const bucketSizeLog)
 {
     BYTE* const pOffset = ldmState->bucketOffsets + hash;
     unsigned const offset = *pOffset;
 
-    *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + offset) = entry;
-    *pOffset = (BYTE)((offset + 1) & ((1u << ldmParams.bucketSizeLog) - 1));
+    *(ZSTD_ldm_getBucket(ldmState, hash, bucketSizeLog) + offset) = entry;
+    *pOffset = (BYTE)((offset + 1) & ((1u << bucketSizeLog) - 1));
 
 }
 
···
  *
  *  The tables for the other strategies are filled within their
  *  block compressors. */
-static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms,
+static size_t ZSTD_ldm_fillFastTables(ZSTD_MatchState_t* ms,
                                       void const* end)
 {
     const BYTE* const iend = (const BYTE*)end;
···
     switch(ms->cParams.strategy)
     {
     case ZSTD_fast:
-        ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast);
+        ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx);
         break;
 
     case ZSTD_dfast:
-        ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast);
+#ifndef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR
+        ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx);
+#else
+        assert(0); /* shouldn't be called: cparams should've been adjusted. */
+#endif
         break;
 
     case ZSTD_greedy:
···
             const BYTE* iend, ldmParams_t const* params)
 {
     U32 const minMatchLength = params->minMatchLength;
-    U32 const hBits = params->hashLog - params->bucketSizeLog;
+    U32 const bucketSizeLog = params->bucketSizeLog;
+    U32 const hBits = params->hashLog - bucketSizeLog;
     BYTE const* const base = ldmState->window.base;
     BYTE const* const istart = ip;
     ldmRollingHashState_t hashState;
···
         unsigned n;
 
         numSplits = 0;
-        hashed = ZSTD_ldm_gear_feed(&hashState, ip, iend - ip, splits, &numSplits);
+        hashed = ZSTD_ldm_gear_feed(&hashState, ip, (size_t)(iend - ip), splits, &numSplits);
 
         for (n = 0; n < numSplits; n++) {
             if (ip + splits[n] >= istart + minMatchLength) {
···
 
                 entry.offset = (U32)(split - base);
                 entry.checksum = (U32)(xxhash >> 32);
-                ZSTD_ldm_insertEntry(ldmState, hash, entry, *params);
+                ZSTD_ldm_insertEntry(ldmState, hash, entry, params->bucketSizeLog);
             }
         }
 
···
  *  Sets cctx->nextToUpdate to a position corresponding closer to anchor
  *  if it is far way
  *  (after a long match, only update tables a limited amount). */
-static void ZSTD_ldm_limitTableUpdate(ZSTD_matchState_t* ms, const BYTE* anchor)
+static void ZSTD_ldm_limitTableUpdate(ZSTD_MatchState_t* ms, const BYTE* anchor)
 {
     U32 const curr = (U32)(anchor - ms->window.base);
     if (curr > ms->nextToUpdate + 1024) {
···
     }
 }
 
-static size_t ZSTD_ldm_generateSequences_internal(
-        ldmState_t* ldmState, rawSeqStore_t* rawSeqStore,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_ldm_generateSequences_internal(
+        ldmState_t* ldmState, RawSeqStore_t* rawSeqStore,
         ldmParams_t const* params, void const* src, size_t srcSize)
 {
     /* LDM parameters */
···
             candidates[n].split = split;
             candidates[n].hash = hash;
             candidates[n].checksum = (U32)(xxhash >> 32);
-            candidates[n].bucket = ZSTD_ldm_getBucket(ldmState, hash, *params);
+            candidates[n].bucket = ZSTD_ldm_getBucket(ldmState, hash, params->bucketSizeLog);
             PREFETCH_L1(candidates[n].bucket);
         }
 
···
              * the previous one, we merely register it in the hash table and
              * move on */
             if (split < anchor) {
-                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, params->bucketSizeLog);
                 continue;
             }
 
···
             /* No match found -- insert an entry into the hash table
              * and process the next candidate match */
             if (bestEntry == NULL) {
-                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, params->bucketSizeLog);
                 continue;
             }
 
···
 
             /* Insert the current entry into the hash table --- it must be
              * done after the previous block to avoid clobbering bestEntry */
-            ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+            ZSTD_ldm_insertEntry(ldmState, hash, newEntry, params->bucketSizeLog);
 
             anchor = split + forwardMatchLength;
 
···
 }
 
 size_t ZSTD_ldm_generateSequences(
-        ldmState_t* ldmState, rawSeqStore_t* sequences,
+        ldmState_t* ldmState, RawSeqStore_t* sequences,
         ldmParams_t const* params, void const* src, size_t srcSize)
 {
     U32 const maxDist = 1U << params->windowLog;
···
          * the window through early invalidation.
          * TODO: * Test the chunk size.
          *       * Try invalidation after the sequence generation and test the
-         *         the offset against maxDist directly.
+         *         offset against maxDist directly.
          *
          * NOTE: Because of dictionaries + sequence splitting we MUST make sure
          * that any offset used is valid at the END of the sequence, since it may
···
 }
 
 void
-ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch)
+ZSTD_ldm_skipSequences(RawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch)
 {
     while (srcSize > 0 && rawSeqStore->pos < rawSeqStore->size) {
         rawSeq* seq = rawSeqStore->seq + rawSeqStore->pos;
···
  * Returns the current sequence to handle, or if the rest of the block should
  * be literals, it returns a sequence with offset == 0.
  */
-static rawSeq maybeSplitSequence(rawSeqStore_t* rawSeqStore,
+static rawSeq maybeSplitSequence(RawSeqStore_t* rawSeqStore,
                                  U32 const remaining, U32 const minMatch)
 {
     rawSeq sequence = rawSeqStore->seq[rawSeqStore->pos];
···
     return sequence;
 }
 
-void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) {
+void ZSTD_ldm_skipRawSeqStoreBytes(RawSeqStore_t* rawSeqStore, size_t nbBytes) {
     U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
     while (currPos && rawSeqStore->pos < rawSeqStore->size) {
         rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
···
     }
 }
 
-size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
-    ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-    ZSTD_paramSwitch_e useRowMatchFinder,
+size_t ZSTD_ldm_blockCompress(RawSeqStore_t* rawSeqStore,
+    ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+    ZSTD_ParamSwitch_e useRowMatchFinder,
     void const* src, size_t srcSize)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     unsigned const minMatch = cParams->minMatch;
-    ZSTD_blockCompressor const blockCompressor =
+    ZSTD_BlockCompressor_f const blockCompressor =
         ZSTD_selectBlockCompressor(cParams->strategy, useRowMatchFinder, ZSTD_matchState_dictMode(ms));
     /* Input bounds */
     BYTE const* const istart = (BYTE const*)src;
···
         /* maybeSplitSequence updates rawSeqStore->pos */
         rawSeq const sequence = maybeSplitSequence(rawSeqStore,
                                                    (U32)(iend - ip), minMatch);
-        int i;
         /* End signal */
         if (sequence.offset == 0)
             break;
···
         /* Run the block compressor */
         DEBUGLOG(5, "pos %u : calling block compressor on segment of size %u", (unsigned)(ip-istart), sequence.litLength);
         {
+            int i;
             size_t const newLitLength =
                 blockCompressor(ms, seqStore, rep, ip, sequence.litLength);
             ip += sequence.litLength;
···
             rep[0] = sequence.offset;
             /* Store the sequence */
             ZSTD_storeSeq(seqStore, newLitLength, ip - newLitLength, iend,
-                          STORE_OFFSET(sequence.offset),
+                          OFFSET_TO_OFFBASE(sequence.offset),
                           sequence.matchLength);
             ip += sequence.matchLength;
         }

lib/zstd/compress/zstd_ldm.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 
 #ifndef ZSTD_LDM_H
 #define ZSTD_LDM_H
-
 
 #include "zstd_compress_internal.h"   /* ldmParams_t, U32 */
 #include <linux/zstd.h>   /* ZSTD_CCtx, size_t */
···
  *       sequences.
  */
 size_t ZSTD_ldm_generateSequences(
-            ldmState_t* ldms, rawSeqStore_t* sequences,
+            ldmState_t* ldms, RawSeqStore_t* sequences,
             ldmParams_t const* params, void const* src, size_t srcSize);
 
 /*
···
  * two. We handle that case correctly, and update `rawSeqStore` appropriately.
  * NOTE: This function does not return any errors.
  */
-size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
-            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-            ZSTD_paramSwitch_e useRowMatchFinder,
+size_t ZSTD_ldm_blockCompress(RawSeqStore_t* rawSeqStore,
+            ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+            ZSTD_ParamSwitch_e useRowMatchFinder,
             void const* src, size_t srcSize);
 
 /*
···
  * Avoids emitting matches less than `minMatch` bytes.
  * Must be called for data that is not passed to ZSTD_ldm_blockCompress().
  */
-void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize,
+void ZSTD_ldm_skipSequences(RawSeqStore_t* rawSeqStore, size_t srcSize,
     U32 const minMatch);
 
 /* ZSTD_ldm_skipRawSeqStoreBytes():
···
  * Not to be used in conjunction with ZSTD_ldm_skipSequences().
  * Must be called for data with is not passed to ZSTD_ldm_blockCompress().
  */
-void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes);
+void ZSTD_ldm_skipRawSeqStoreBytes(RawSeqStore_t* rawSeqStore, size_t nbBytes);
 
 /* ZSTD_ldm_getTableSize() :
  *  Estimate the space needed for long distance matching tables or 0 if LDM is
···
  */
 void ZSTD_ldm_adjustParameters(ldmParams_t* params,
                                ZSTD_compressionParameters const* cParams);
-
 
 #endif /* ZSTD_FAST_H */

lib/zstd/compress/zstd_ldm_geartab.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

lib/zstd/compress/zstd_opt.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Przemyslaw Skibinski, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #include "hist.h"
 #include "zstd_opt.h"
 
+#if !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR)
 
 #define ZSTD_LITFREQ_ADD    2   /* scaling factor for litFreq, so that frequencies adapt faster to new stats */
 #define ZSTD_MAX_PRICE     (1<<30)
 
-#define ZSTD_PREDEF_THRESHOLD 1024   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */
+#define ZSTD_PREDEF_THRESHOLD 8   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */
 
 
 /*-*************************************
···
 #if 0    /* approximation at bit level (for tests) */
 #  define BITCOST_ACCURACY 0
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat, opt) ((void)opt, ZSTD_bitWeight(stat))
+#  define WEIGHT(stat, opt) ((void)(opt), ZSTD_bitWeight(stat))
 #elif 0  /* fractional bit accuracy (for tests) */
 #  define BITCOST_ACCURACY 8
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat,opt) ((void)opt, ZSTD_fracWeight(stat))
+#  define WEIGHT(stat,opt) ((void)(opt), ZSTD_fracWeight(stat))
 #else    /* opt==approx, ultra==accurate */
 #  define BITCOST_ACCURACY 8
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat,opt) (opt ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
+#  define WEIGHT(stat,opt) ((opt) ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
 #endif
 
+/* ZSTD_bitWeight() :
+ * provide estimated "cost" of a stat in full bits only */
 MEM_STATIC U32 ZSTD_bitWeight(U32 stat)
 {
     return (ZSTD_highbit32(stat+1) * BITCOST_MULTIPLIER);
 }
 
+/* ZSTD_fracWeight() :
+ * provide fractional-bit "cost" of a stat,
+ * using linear interpolation approximation */
 MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat)
 {
     U32 const stat = rawStat + 1;
     U32 const hb = ZSTD_highbit32(stat);
     U32 const BWeight = hb * BITCOST_MULTIPLIER;
+    /* Fweight was meant for "Fractional weight"
+     * but it's effectively a value between 1 and 2
+     * using fixed point arithmetic */
     U32 const FWeight = (stat << BITCOST_ACCURACY) >> hb;
     U32 const weight = BWeight + FWeight;
     assert(hb + BITCOST_ACCURACY < 31);
···
 /* debugging function,
  * @return price in bytes as fractional value
  * for debug messages only */
-MEM_STATIC double ZSTD_fCost(U32 price)
+MEM_STATIC double ZSTD_fCost(int price)
 {
     return (double)price / (BITCOST_MULTIPLIER*8);
 }
···
     return total;
 }
 
-static U32 ZSTD_downscaleStats(unsigned* table, U32 lastEltIndex, U32 shift)
+typedef enum { base_0possible=0, base_1guaranteed=1 } base_directive_e;
+
+static U32
+ZSTD_downscaleStats(unsigned* table, U32 lastEltIndex, U32 shift, base_directive_e base1)
 {
     U32 s, sum=0;
-    DEBUGLOG(5, "ZSTD_downscaleStats (nbElts=%u, shift=%u)", (unsigned)lastEltIndex+1, (unsigned)shift);
+    DEBUGLOG(5, "ZSTD_downscaleStats (nbElts=%u, shift=%u)",
+            (unsigned)lastEltIndex+1, (unsigned)shift );
     assert(shift < 30);
     for (s=0; s<lastEltIndex+1; s++) {
-        table[s] = 1 + (table[s] >> shift);
-        sum += table[s];
+        unsigned const base = base1 ? 1 : (table[s]>0);
+        unsigned const newStat = base + (table[s] >> shift);
+        sum += newStat;
+        table[s] = newStat;
     }
     return sum;
 }
 
 /* ZSTD_scaleStats() :
- * reduce all elements in table is sum too large
+ * reduce all elt frequencies in table if sum too large
  * return the resulting sum of elements */
 static U32 ZSTD_scaleStats(unsigned* table, U32 lastEltIndex, U32 logTarget)
 {
···
     DEBUGLOG(5, "ZSTD_scaleStats (nbElts=%u, target=%u)", (unsigned)lastEltIndex+1, (unsigned)logTarget);
     assert(logTarget < 30);
     if (factor <= 1) return prevsum;
-    return ZSTD_downscaleStats(table, lastEltIndex, ZSTD_highbit32(factor));
+    return ZSTD_downscaleStats(table, lastEltIndex, ZSTD_highbit32(factor), base_1guaranteed);
 }
 
 /* ZSTD_rescaleFreqs() :
···
     DEBUGLOG(5, "ZSTD_rescaleFreqs (srcSize=%u)", (unsigned)srcSize);
     optPtr->priceType = zop_dynamic;
 
-    if (optPtr->litLengthSum == 0) {  /* first block : init */
-        if (srcSize <= ZSTD_PREDEF_THRESHOLD) {  /* heuristic */
-            DEBUGLOG(5, "(srcSize <= ZSTD_PREDEF_THRESHOLD) => zop_predef");
+    if (optPtr->litLengthSum == 0) {  /* no literals stats collected -> first block assumed -> init */
+
+        /* heuristic: use pre-defined stats for too small inputs */
+        if (srcSize <= ZSTD_PREDEF_THRESHOLD) {
+            DEBUGLOG(5, "srcSize <= %i : use predefined stats", ZSTD_PREDEF_THRESHOLD);
             optPtr->priceType = zop_predef;
         }
 
         assert(optPtr->symbolCosts != NULL);
         if (optPtr->symbolCosts->huf.repeatMode == HUF_repeat_valid) {
-            /* huffman table presumed generated by dictionary */
+
+            /* huffman stats covering the full value set : table presumed generated by dictionary */
             optPtr->priceType = zop_dynamic;
 
             if (compressedLiterals) {
+                /* generate literals statistics from huffman table */
                 unsigned lit;
                 assert(optPtr->litFreq != NULL);
                 optPtr->litSum = 0;
···
                     optPtr->offCodeSum += optPtr->offCodeFreq[of];
             }   }
 
-        } else {  /* not a dictionary */
+        } else {  /* first block, no dictionary */
 
             assert(optPtr->litFreq != NULL);
             if (compressedLiterals) {
+                /* base initial cost of literals on direct frequency within src */
                 unsigned lit = MaxLit;
                 HIST_count_simple(optPtr->litFreq, &lit, src, srcSize);   /* use raw first block to init statistics */
-                optPtr->litSum = ZSTD_downscaleStats(optPtr->litFreq, MaxLit, 8);
+                optPtr->litSum = ZSTD_downscaleStats(optPtr->litFreq, MaxLit, 8, base_0possible);
             }
 
             {   unsigned const baseLLfreqs[MaxLL+1] = {
···
                 optPtr->offCodeSum = sum_u32(baseOFCfreqs, MaxOff+1);
             }
 
-
         }
 
-    } else {   /* new block : re-use previous statistics, scaled down */
+    } else {   /* new block : scale down accumulated statistics */
 
         if (compressedLiterals)
             optPtr->litSum = ZSTD_scaleStats(optPtr->litFreq, MaxLit, 12);
···
                                 const optState_t* const optPtr,
                                 int optLevel)
 {
+    DEBUGLOG(8, "ZSTD_rawLiteralsCost (%u literals)", litLength);
     if (litLength == 0) return 0;
 
     if (!ZSTD_compressedLiterals(optPtr))
···
         return (litLength*6) * BITCOST_MULTIPLIER;  /* 6 bit per literal - no statistic used */
 
     /* dynamic statistics */
-    {   U32 price = litLength * optPtr->litSumBasePrice;
+    {   U32 price = optPtr->litSumBasePrice * litLength;
+        U32 const litPriceMax = optPtr->litSumBasePrice - BITCOST_MULTIPLIER;
         U32 u;
+        assert(optPtr->litSumBasePrice >= BITCOST_MULTIPLIER);
         for (u=0; u < litLength; u++) {
-            assert(WEIGHT(optPtr->litFreq[literals[u]], optLevel) <= optPtr->litSumBasePrice);   /* literal cost should never be negative */
-            price -= WEIGHT(optPtr->litFreq[literals[u]], optLevel);
+            U32 litPrice = WEIGHT(optPtr->litFreq[literals[u]], optLevel);
+            if (UNLIKELY(litPrice > litPriceMax)) litPrice = litPriceMax;
+            price -= litPrice;
         }
         return price;
     }
···
     assert(litLength <= ZSTD_BLOCKSIZE_MAX);
     if (optPtr->priceType == zop_predef)
         return WEIGHT(litLength, optLevel);
-    /* We can't compute the litLength price for sizes >= ZSTD_BLOCKSIZE_MAX
-     * because it isn't representable in the zstd format. So instead just
-     * call it 1 bit more than ZSTD_BLOCKSIZE_MAX - 1. In this case the block
-     * would be all literals.
+
+    /* ZSTD_LLcode() can't compute litLength price for sizes >= ZSTD_BLOCKSIZE_MAX
+     * because it isn't representable in the zstd format.
+     * So instead just pretend it would cost 1 bit more than ZSTD_BLOCKSIZE_MAX - 1.
+     * In such a case, the block would be all literals.
      */
     if (litLength == ZSTD_BLOCKSIZE_MAX)
         return BITCOST_MULTIPLIER + ZSTD_litLengthPrice(ZSTD_BLOCKSIZE_MAX - 1, optPtr, optLevel);
···
 }
 
 /* ZSTD_getMatchPrice() :
- * Provides the cost of the match part (offset + matchLength) of a sequence
+ * Provides the cost of the match part (offset + matchLength) of a sequence.
  * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence.
- * @offcode : expects a scale where 0,1,2 are repcodes 1-3, and 3+ are real_offsets+2
+ * @offBase : sumtype, representing an offset or a repcode, and using numeric representation of ZSTD_storeSeq()
  * @optLevel: when <2, favors small offset for decompression speed (improved cache efficiency)
  */
 FORCE_INLINE_TEMPLATE U32
-ZSTD_getMatchPrice(U32 const offcode,
+ZSTD_getMatchPrice(U32 const offBase,
                    U32 const matchLength,
              const optState_t* const optPtr,
                    int const optLevel)
 {
     U32 price;
-    U32 const offCode = ZSTD_highbit32(STORED_TO_OFFBASE(offcode));
+    U32 const offCode = ZSTD_highbit32(offBase);
     U32 const mlBase = matchLength - MINMATCH;
     assert(matchLength >= MINMATCH);
 
-    if (optPtr->priceType == zop_predef)  /* fixed scheme, do not use statistics */
-        return WEIGHT(mlBase, optLevel) + ((16 + offCode) * BITCOST_MULTIPLIER);
+    if (optPtr->priceType == zop_predef)  /* fixed scheme, does not use statistics */
+        return WEIGHT(mlBase, optLevel)
+             + ((16 + offCode) * BITCOST_MULTIPLIER); /* emulated offset cost */
 
     /* dynamic statistics */
     price = (offCode * BITCOST_MULTIPLIER) + (optPtr->offCodeSumBasePrice - WEIGHT(optPtr->offCodeFreq[offCode], optLevel));
···
 }
 
 /* ZSTD_updateStats() :
- * assumption : literals + litLengtn <= iend */
+ * assumption : literals + litLength <= iend */
 static void ZSTD_updateStats(optState_t* const optPtr,
                              U32 litLength, const BYTE* literals,
-                             U32 offsetCode, U32 matchLength)
+                             U32 offBase, U32 matchLength)
 {
     /* literals */
     if (ZSTD_compressedLiterals(optPtr)) {
···
         optPtr->litLengthSum++;
     }
 
-    /* offset code : expected to follow storeSeq() numeric representation */
-    {   U32 const offCode = ZSTD_highbit32(STORED_TO_OFFBASE(offsetCode));
+    /* offset code : follows storeSeq() numeric representation */
+    {   U32 const offCode = ZSTD_highbit32(offBase);
         assert(offCode <= MaxOff);
         optPtr->offCodeFreq[offCode]++;
         optPtr->offCodeSum++;
···
 
 /* Update hashTable3 up to ip (excluded)
    Assumption : always within prefix (i.e. not within extDict) */
-static U32 ZSTD_insertAndFindFirstIndexHash3 (const ZSTD_matchState_t* ms,
-                                              U32* nextToUpdate3,
-                                              const BYTE* const ip)
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_insertAndFindFirstIndexHash3 (const ZSTD_MatchState_t* ms,
+                                       U32* nextToUpdate3,
+                                       const BYTE* const ip)
 {
     U32* const hashTable3 = ms->hashTable3;
     U32 const hashLog3 = ms->hashLog3;
···
  * @param ip assumed <= iend-8 .
  * @param target The target of ZSTD_updateTree_internal() - we are filling to this position
  * @return : nb of positions added */
-static U32 ZSTD_insertBt1(
-                const ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_insertBt1(
+                const ZSTD_MatchState_t* ms,
                 const BYTE* const ip, const BYTE* const iend,
                 U32 const target,
                 U32 const mls, const int extDict)
···
 }
 
 FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 void ZSTD_updateTree_internal(
-                ZSTD_matchState_t* ms,
+                ZSTD_MatchState_t* ms,
                 const BYTE* const ip, const BYTE* const iend,
                 const U32 mls, const ZSTD_dictMode_e dictMode)
 {
     const BYTE* const base = ms->window.base;
     U32 const target = (U32)(ip - base);
     U32 idx = ms->nextToUpdate;
-    DEBUGLOG(6, "ZSTD_updateTree_internal, from %u to %u  (dictMode:%u)",
+    DEBUGLOG(7, "ZSTD_updateTree_internal, from %u to %u  (dictMode:%u)",
                 idx, target, dictMode);
 
     while(idx < target) {
···
     ms->nextToUpdate = target;
 }
 
-void ZSTD_updateTree(ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iend) {
+void ZSTD_updateTree(ZSTD_MatchState_t* ms, const BYTE* ip, const BYTE* iend) {
     ZSTD_updateTree_internal(ms, ip, iend, ms->cParams.minMatch, ZSTD_noDict);
 }
 
 FORCE_INLINE_TEMPLATE
-U32 ZSTD_insertBtAndGetAllMatches (
-                    ZSTD_match_t* matches,   /* store result (found matches) in this table (presumed large enough) */
-                    ZSTD_matchState_t* ms,
-                    U32* nextToUpdate3,
-                    const BYTE* const ip, const BYTE* const iLimit, const ZSTD_dictMode_e dictMode,
-                    const U32 rep[ZSTD_REP_NUM],
-                    U32 const ll0,   /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
-                    const U32 lengthToBeat,
-                    U32 const mls /* template */)
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32
+ZSTD_insertBtAndGetAllMatches (
+                ZSTD_match_t* matches,  /* store result (found matches) in this table (presumed large enough) */
+                ZSTD_MatchState_t* ms,
+                U32* nextToUpdate3,
+                const BYTE* const ip, const BYTE* const iLimit,
+                const ZSTD_dictMode_e dictMode,
+                const U32 rep[ZSTD_REP_NUM],
+                const U32 ll0,  /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
+                const U32 lengthToBeat,
+                const U32 mls /* template */)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
···
     U32 mnum = 0;
     U32 nbCompares = 1U << cParams->searchLog;
 
-    const ZSTD_matchState_t* dms    = dictMode == ZSTD_dictMatchState ? ms->dictMatchState : NULL;
+    const ZSTD_MatchState_t* dms    = dictMode == ZSTD_dictMatchState ? ms->dictMatchState : NULL;
     const ZSTD_compressionParameters* const dmsCParams =
                                       dictMode == ZSTD_dictMatchState ? &dms->cParams : NULL;
     const BYTE* const dmsBase       = dictMode == ZSTD_dictMatchState ? dms->window.base : NULL;
···
                 assert(curr >= windowLow);
                 if ( dictMode == ZSTD_extDict
                   && ( ((repOffset-1) /*intentional overflow*/ < curr - windowLow)  /* equivalent to `curr > repIndex >= windowLow` */
-                     & (((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */)
+                     & (ZSTD_index_overlap_check(dictLimit, repIndex)) )
                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
                     repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dictEnd, prefixStart) + minMatch;
                 }
                 if (dictMode == ZSTD_dictMatchState
                   && ( ((repOffset-1) /*intentional overflow*/ < curr - (dmsLowLimit + dmsIndexDelta))  /* equivalent to `curr > repIndex >= dmsLowLimit` */
-                     & ((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */
+                     & (ZSTD_index_overlap_check(dictLimit, repIndex)) )
                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
                     repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dmsEnd, prefixStart) + minMatch;
             }   }
···
                 DEBUGLOG(8, "found repCode %u (ll0:%u, offset:%u) of length %u",
                             repCode, ll0, repOffset, repLen);
                 bestLength = repLen;
-                matches[mnum].off = STORE_REPCODE(repCode - ll0 + 1);  /* expect value between 1 and 3 */
+                matches[mnum].off = REPCODE_TO_OFFBASE(repCode - ll0 + 1);  /* expect value between 1 and 3 */
                 matches[mnum].len = (U32)repLen;
                 mnum++;
                 if ( (repLen > sufficient_len)
···
                 bestLength = mlen;
                 assert(curr > matchIndex3);
                 assert(mnum==0);  /* no prior solution */
-                matches[0].off = STORE_OFFSET(curr - matchIndex3);
+                matches[0].off = OFFSET_TO_OFFBASE(curr - matchIndex3);
                 matches[0].len = (U32)mlen;
                 mnum = 1;
                 if ( (mlen > sufficient_len) |
···
         }
 
         if (matchLength > bestLength) {
-            DEBUGLOG(8, "found match of length %u at distance %u (offCode=%u)",
-                    (U32)matchLength, curr - matchIndex, STORE_OFFSET(curr - matchIndex));
+            DEBUGLOG(8, "found match of length %u at distance %u (offBase=%u)",
+                    (U32)matchLength, curr - matchIndex, OFFSET_TO_OFFBASE(curr - matchIndex));
             assert(matchEndIdx > matchIndex);
             if (matchLength > matchEndIdx - matchIndex)
                 matchEndIdx = matchIndex + (U32)matchLength;
             bestLength = matchLength;
-            matches[mnum].off = STORE_OFFSET(curr - matchIndex);
+            matches[mnum].off = OFFSET_TO_OFFBASE(curr - matchIndex);
             matches[mnum].len = (U32)matchLength;
             mnum++;
             if ( (matchLength > ZSTD_OPT_NUM)
···
 
             if (matchLength > bestLength) {
                 matchIndex = dictMatchIndex + dmsIndexDelta;
-                DEBUGLOG(8, "found dms match of length %u at distance %u (offCode=%u)",
-                        (U32)matchLength, curr - matchIndex, STORE_OFFSET(curr - matchIndex));
+                DEBUGLOG(8, "found dms match of length %u at distance %u (offBase=%u)",
+                        (U32)matchLength, curr - matchIndex, OFFSET_TO_OFFBASE(curr - matchIndex));
                 if (matchLength > matchEndIdx - matchIndex)
                     matchEndIdx = matchIndex + (U32)matchLength;
                 bestLength = matchLength;
-                matches[mnum].off = STORE_OFFSET(curr - matchIndex);
+                matches[mnum].off = OFFSET_TO_OFFBASE(curr - matchIndex);
                 matches[mnum].len = (U32)matchLength;
                 mnum++;
                 if ( (matchLength > ZSTD_OPT_NUM)
···
 
 typedef U32 (*ZSTD_getAllMatchesFn)(
     ZSTD_match_t*,
-    ZSTD_matchState_t*,
+    ZSTD_MatchState_t*,
     U32*,
     const BYTE*,
     const BYTE*,
···
     U32 const ll0,
     U32 const lengthToBeat);
 
-FORCE_INLINE_TEMPLATE U32 ZSTD_btGetAllMatches_internal(
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_btGetAllMatches_internal(
         ZSTD_match_t* matches,
-        ZSTD_matchState_t* ms,
+        ZSTD_MatchState_t* ms,
         U32* nextToUpdate3,
         const BYTE* ip,
         const BYTE* const iHighLimit,
···
 #define GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, mls)            \
     static U32 ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls)(      \
             ZSTD_match_t* matches,                             \
-            ZSTD_matchState_t* ms,                             \
+            ZSTD_MatchState_t* ms,                             \
             U32* nextToUpdate3,                                \
             const BYTE* ip,                                    \
             const BYTE* const iHighLimit,                      \
···
     }
 
 static ZSTD_getAllMatchesFn
-ZSTD_selectBtGetAllMatches(ZSTD_matchState_t const* ms, ZSTD_dictMode_e const dictMode)
+ZSTD_selectBtGetAllMatches(ZSTD_MatchState_t const* ms, ZSTD_dictMode_e const dictMode)
 {
     ZSTD_getAllMatchesFn const getAllMatchesFns[3][4] = {
         ZSTD_BT_GET_ALL_MATCHES_ARRAY(noDict),
···
 
 /* Struct containing info needed to make decision about ldm inclusion */
 typedef struct {
-    rawSeqStore_t seqStore;   /* External match candidates store for this block */
+    RawSeqStore_t seqStore;   /* External match candidates store for this block */
     U32 startPosInBlock;      /* Start position of the current match candidate */
     U32 endPosInBlock;        /* End position of the current match candidate */
     U32 offset;               /* Offset of the match candidate */
···
  * Moves forward in @rawSeqStore by @nbBytes,
  * which will update the fields 'pos' and 'posInSequence'.
  */
-static void ZSTD_optLdm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes)
+static void ZSTD_optLdm_skipRawSeqStoreBytes(RawSeqStore_t* rawSeqStore, size_t nbBytes)
 {
     U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
     while (currPos && rawSeqStore->pos < rawSeqStore->size) {
···
         return;
     }
 
-    /* Matches may be < MINMATCH by this process. In that case, we will reject them
+    /* Matches may be < minMatch by this process. In that case, we will reject them
        when we are deciding whether or not to add the ldm */
     optLdm->startPosInBlock = currPosInBlock + literalsBytesRemaining;
     optLdm->endPosInBlock = optLdm->startPosInBlock + matchBytesRemaining;
···
  * into 'matches'. Maintains the correct ordering of 'matches'.
  */
 static void ZSTD_optLdm_maybeAddMatch(ZSTD_match_t* matches, U32* nbMatches,
-                                      const ZSTD_optLdm_t* optLdm, U32 currPosInBlock)
+                                      const ZSTD_optLdm_t* optLdm, U32 currPosInBlock,
+                                      U32 minMatch)
 {
     U32 const posDiff = currPosInBlock - optLdm->startPosInBlock;
-    /* Note: ZSTD_match_t actually contains offCode and matchLength (before subtracting MINMATCH) */
+    /* Note: ZSTD_match_t actually contains offBase and matchLength (before subtracting MINMATCH) */
     U32 const candidateMatchLength = optLdm->endPosInBlock - optLdm->startPosInBlock - posDiff;
 
     /* Ensure that current block position is not outside of the match */
     if (currPosInBlock < optLdm->startPosInBlock
       || currPosInBlock >= optLdm->endPosInBlock
-      || candidateMatchLength < MINMATCH) {
+      || candidateMatchLength < minMatch) {
         return;
     }
 
     if (*nbMatches == 0 || ((candidateMatchLength > matches[*nbMatches-1].len) && *nbMatches < ZSTD_OPT_NUM)) {
-        U32 const candidateOffCode = STORE_OFFSET(optLdm->offset);
-        DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offCode: %u matchLength %u) at block position=%u",
-                 candidateOffCode, candidateMatchLength, currPosInBlock);
+        U32 const candidateOffBase = OFFSET_TO_OFFBASE(optLdm->offset);
+        DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offBase: %u matchLength %u) at block position=%u",
+                 candidateOffBase, candidateMatchLength, currPosInBlock);
         matches[*nbMatches].len = candidateMatchLength;
-        matches[*nbMatches].off = candidateOffCode;
+        matches[*nbMatches].off = candidateOffBase;
         (*nbMatches)++;
     }
 }
···
 static void
 ZSTD_optLdm_processMatchCandidate(ZSTD_optLdm_t* optLdm,
                                   ZSTD_match_t* matches, U32* nbMatches,
-                                  U32 currPosInBlock, U32 remainingBytes)
+                                  U32 currPosInBlock, U32 remainingBytes,
+                                  U32 minMatch)
 {
     if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
         return;
···
         }
         ZSTD_opt_getNextMatchAndUpdateSeqStore(optLdm, currPosInBlock, remainingBytes);
     }
-    ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock);
+    ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock, minMatch);
 }
 
 
 /*-*******************************
 *  Optimal parser
 *********************************/
-
-static U32 ZSTD_totalLen(ZSTD_optimal_t sol)
-{
-    return sol.litlen + sol.mlen;
-}
 
 #if 0 /* debug */
 
···
 
 #endif
 
-FORCE_INLINE_TEMPLATE size_t
-ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
-                               seqStore_t* seqStore,
+#define LIT_PRICE(_p) (int)ZSTD_rawLiteralsCost(_p, 1, optStatePtr, optLevel)
+#define LL_PRICE(_l) (int)ZSTD_litLengthPrice(_l, optStatePtr, optLevel)
+#define LL_INCPRICE(_l) (LL_PRICE(_l) - LL_PRICE(_l-1))
+
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t
+ZSTD_compressBlock_opt_generic(ZSTD_MatchState_t* ms,
+                               SeqStore_t* seqStore,
                                U32 rep[ZSTD_REP_NUM],
                          const void* src, size_t srcSize,
                          const int optLevel,
···
 
     ZSTD_optimal_t* const opt = optStatePtr->priceTable;
     ZSTD_match_t* const matches = optStatePtr->matchTable;
-    ZSTD_optimal_t lastSequence;
+    ZSTD_optimal_t lastStretch;
     ZSTD_optLdm_t optLdm;
+
+    ZSTD_memset(&lastStretch, 0, sizeof(ZSTD_optimal_t));
 
     optLdm.seqStore = ms->ldmSeqStore ? *ms->ldmSeqStore : kNullRawSeqStore;
     optLdm.endPosInBlock = optLdm.startPosInBlock = optLdm.offset = 0;
···
             U32 const ll0 = !litlen;
             U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, ip, iend, rep, ll0, minMatch);
             ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
-                                              (U32)(ip-istart), (U32)(iend - ip));
-            if (!nbMatches) { ip++; continue; }
+                                              (U32)(ip-istart), (U32)(iend-ip),
+                                              minMatch);
+            if (!nbMatches) {
+                DEBUGLOG(8, "no match found at cPos %u", (unsigned)(ip-istart));
+                ip++;
+                continue;
+            }
+
+            /* Match found: let's store this solution, and eventually find more candidates.
+             * During this forward pass, @opt is used to store stretches,
+             * defined as "a match followed by N literals".
+             * Note how this is different from a Sequence, which is "N literals followed by a match".
+             * Storing stretches allows us to store different match predecessors
+             * for each literal position part of a literals run. */
 
             /* initialize opt[0] */
-            { U32 i ; for (i=0; i<ZSTD_REP_NUM; i++) opt[0].rep[i] = rep[i]; }
-            opt[0].mlen = 0;  /* means is_a_literal */
+            opt[0].mlen = 0;  /* there are only literals so far */
             opt[0].litlen = litlen;
-            /* We don't need to include the actual price of the literals because
-             * it is static for the duration of the forward pass, and is included
-             * in every price. We include the literal length to avoid negative
-             * prices when we subtract the previous literal length.
+            /* No need to include the actual price of the literals before the first match
+             * because it is static for the duration of the forward pass, and is included
+             * in every subsequent price. But, we include the literal length because
+             * the cost variation of litlen depends on the value of litlen.
              */
-            opt[0].price = (int)ZSTD_litLengthPrice(litlen, optStatePtr, optLevel);
+            opt[0].price = LL_PRICE(litlen);
+            ZSTD_STATIC_ASSERT(sizeof(opt[0].rep[0]) == sizeof(rep[0]));
+            ZSTD_memcpy(&opt[0].rep, rep, sizeof(opt[0].rep));
 
             /* large match -> immediate encoding */
             {   U32 const maxML = matches[nbMatches-1].len;
-                U32 const maxOffcode = matches[nbMatches-1].off;
-                DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffCode=%u at cPos=%u => start new series",
-                            nbMatches, maxML, maxOffcode, (U32)(ip-prefixStart));
+                U32 const maxOffBase = matches[nbMatches-1].off;
+                DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffBase=%u at cPos=%u => start new series",
+                            nbMatches, maxML, maxOffBase, (U32)(ip-prefixStart));
 
                 if (maxML > sufficient_len) {
-                    lastSequence.litlen = litlen;
-                    lastSequence.mlen = maxML;
-                    lastSequence.off = maxOffcode;
-                    DEBUGLOG(6, "large match (%u>%u), immediate encoding",
+                    lastStretch.litlen = 0;
+                    lastStretch.mlen = maxML;
+                    lastStretch.off = maxOffBase;
+                    DEBUGLOG(6, "large match (%u>%u) => immediate encoding",
                                 maxML, sufficient_len);
                     cur = 0;
-                    last_pos = ZSTD_totalLen(lastSequence);
+                    last_pos = maxML;
                     goto _shortestPath;
             }   }
 
             /* set prices for first matches starting position == 0 */
             assert(opt[0].price >= 0);
-            {   U32 const literalsPrice = (U32)opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
-                U32 pos;
+            {   U32 pos;
                 U32 matchNb;
                 for (pos = 1; pos < minMatch; pos++) {
-                    opt[pos].price = ZSTD_MAX_PRICE;   /* mlen, litlen and price will be fixed during forward scanning */
+                    opt[pos].price = ZSTD_MAX_PRICE;
+                    opt[pos].mlen = 0;
+                    opt[pos].litlen = litlen + pos;
                 }
                 for (matchNb = 0; matchNb < nbMatches; matchNb++) {
-                    U32 const offcode = matches[matchNb].off;
+                    U32 const offBase = matches[matchNb].off;
                     U32 const end = matches[matchNb].len;
                     for ( ; pos <= end ; pos++ ) {
-                        U32 const matchPrice = ZSTD_getMatchPrice(offcode, pos, optStatePtr, optLevel);
-                        U32 const sequencePrice = literalsPrice + matchPrice;
+                        int const matchPrice = (int)ZSTD_getMatchPrice(offBase, pos, optStatePtr, optLevel);
+                        int const sequencePrice = opt[0].price + matchPrice;
                         DEBUGLOG(7, "rPos:%u => set initial price : %.2f",
                                     pos, ZSTD_fCost(sequencePrice));
                         opt[pos].mlen = pos;
-                        opt[pos].off = offcode;
-                        opt[pos].litlen = litlen;
-                        opt[pos].price = (int)sequencePrice;
-                }   }
+                        opt[pos].off = offBase;
+                        opt[pos].litlen = 0; /* end of match */
+                        opt[pos].price = sequencePrice + LL_PRICE(0);
+                    }
+                }
                 last_pos = pos-1;
+                opt[pos].price = ZSTD_MAX_PRICE;
             }
         }
 
         /* check further positions */
         for (cur = 1; cur <= last_pos; cur++) {
             const BYTE* const inr = ip + cur;
-            assert(cur < ZSTD_OPT_NUM);
-            DEBUGLOG(7, "cPos:%zi==rPos:%u", inr-istart, cur)
+            assert(cur <= ZSTD_OPT_NUM);
+            DEBUGLOG(7, "cPos:%i==rPos:%u", (int)(inr-istart), cur);
 
             /* Fix current position with one literal if cheaper */
-            {   U32 const litlen = (opt[cur-1].mlen == 0) ? opt[cur-1].litlen + 1 : 1;
+            {   U32 const litlen = opt[cur-1].litlen + 1;
                 int const price = opt[cur-1].price
-                                + (int)ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel)
-                                + (int)ZSTD_litLengthPrice(litlen, optStatePtr, optLevel)
-                                - (int)ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel);
+                                + LIT_PRICE(ip+cur-1)
+                                + LL_INCPRICE(litlen);
                 assert(price < 1000000000); /* overflow check */
                 if (price <= opt[cur].price) {
-                    DEBUGLOG(7, "cPos:%zi==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)",
-                                inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price), litlen,
+                    ZSTD_optimal_t const prevMatch = opt[cur];
+                    DEBUGLOG(7, "cPos:%i==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)",
+                                (int)(inr-istart), cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price), litlen,
                                 opt[cur-1].rep[0], opt[cur-1].rep[1], opt[cur-1].rep[2]);
-                    opt[cur].mlen = 0;
-                    opt[cur].off = 0;
+                    opt[cur] = opt[cur-1];
                     opt[cur].litlen = litlen;
                     opt[cur].price = price;
+                    if ( (optLevel >= 1) /* additional check only for higher modes */
+                      && (prevMatch.litlen == 0) /* replace a match */
+                      && (LL_INCPRICE(1) < 0) /* ll1 is cheaper than ll0 */
+                      && LIKELY(ip + cur < iend)
+                    ) {
+                        /* check next position, in case it would be cheaper */
+                        int with1literal = prevMatch.price + LIT_PRICE(ip+cur) + LL_INCPRICE(1);
+                        int withMoreLiterals = price + LIT_PRICE(ip+cur) + LL_INCPRICE(litlen+1);
+                        DEBUGLOG(7, "then at next rPos %u : match+1lit %.2f vs %ulits %.2f",
+                                cur+1, ZSTD_fCost(with1literal), litlen+1, ZSTD_fCost(withMoreLiterals));
+                        if ( (with1literal < withMoreLiterals)
+                          && (with1literal < opt[cur+1].price) ) {
+                            /* update offset history - before it disappears */
+                            U32 const prev = cur - prevMatch.mlen;
+                            Repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, prevMatch.off, opt[prev].litlen==0);
+                            assert(cur >= prevMatch.mlen);
+                            DEBUGLOG(7, "==> match+1lit is cheaper (%.2f < %.2f) (hist:%u,%u,%u) !",
+                                        ZSTD_fCost(with1literal), ZSTD_fCost(withMoreLiterals),
+                                        newReps.rep[0], newReps.rep[1], newReps.rep[2] );
+                            opt[cur+1] = prevMatch;  /* mlen & offbase */
+                            ZSTD_memcpy(opt[cur+1].rep, &newReps, sizeof(Repcodes_t));
+                            opt[cur+1].litlen = 1;
+                            opt[cur+1].price = with1literal;
+                            if (last_pos < cur+1) last_pos = cur+1;
+                        }
+                    }
                 } else {
-                    DEBUGLOG(7, "cPos:%zi==rPos:%u : literal would cost more (%.2f>%.2f) (hist:%u,%u,%u)",
-                                inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price),
-                                opt[cur].rep[0], opt[cur].rep[1], opt[cur].rep[2]);
+                    DEBUGLOG(7, "cPos:%i==rPos:%u : literal would cost more (%.2f>%.2f)",
+                                (int)(inr-istart), cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price));
                 }
             }
 
-            /* Set the repcodes of the current position. We must do it here
-             * because we rely on the repcodes of the 2nd to last sequence being
-             * correct to set the next chunks repcodes during the backward
-             * traversal.
+            /* Offset history is not updated during match comparison.
+             * Do it here, now that the match is selected and confirmed.
              */
-            ZSTD_STATIC_ASSERT(sizeof(opt[cur].rep) == sizeof(repcodes_t));
+            ZSTD_STATIC_ASSERT(sizeof(opt[cur].rep) == sizeof(Repcodes_t));
             assert(cur >= opt[cur].mlen);
-            if (opt[cur].mlen != 0) {
+            if (opt[cur].litlen == 0) {
+                /* just finished a match => alter offset history */
                 U32 const prev = cur - opt[cur].mlen;
-                repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0);
-                ZSTD_memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t));
-            } else {
-                ZSTD_memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t));
+                Repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, opt[cur].off, opt[prev].litlen==0);
+                ZSTD_memcpy(opt[cur].rep, &newReps, sizeof(Repcodes_t));
             }
 
             /* last match must start at a minimum distance of 8 from oend */
···
 
             if ( (optLevel==0) /*static_test*/
               && (opt[cur+1].price <= opt[cur].price + (BITCOST_MULTIPLIER/2)) ) {
-                DEBUGLOG(7, "move to next rPos:%u : price is <=", cur+1);
+                DEBUGLOG(7, "skip current position : next rPos(%u) price is cheaper", cur+1);
                 continue;  /* skip unpromising positions; about ~+6% speed, -0.01 ratio */
             }
 
             assert(opt[cur].price >= 0);
-            {   U32 const ll0 = (opt[cur].mlen != 0);
-                U32 const litlen = (opt[cur].mlen == 0) ? opt[cur].litlen : 0;
-                U32 const previousPrice = (U32)opt[cur].price;
-                U32 const basePrice = previousPrice + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
+            {   U32 const ll0 = (opt[cur].litlen == 0);
+                int const previousPrice = opt[cur].price;
+                int const basePrice = previousPrice + LL_PRICE(0);
                 U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, inr, iend, opt[cur].rep, ll0, minMatch);
                 U32 matchNb;
 
                 ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
-                                                  (U32)(inr-istart), (U32)(iend-inr));
+                                                  (U32)(inr-istart), (U32)(iend-inr),
+                                                  minMatch);
 
                 if (!nbMatches) {
                     DEBUGLOG(7, "rPos:%u : no match found", cur);
                     continue;
                 }
 
-                {   U32 const maxML = matches[nbMatches-1].len;
-                    DEBUGLOG(7, "cPos:%zi==rPos:%u, found %u matches, of maxLength=%u",
-                                inr-istart, cur, nbMatches, maxML);
+                {   U32 const longestML = matches[nbMatches-1].len;
+                    DEBUGLOG(7, "cPos:%i==rPos:%u, found %u matches, of longest ML=%u",
+                                (int)(inr-istart), cur, nbMatches, longestML);
 
-                    if ( (maxML > sufficient_len)
-                      || (cur + maxML >= ZSTD_OPT_NUM) ) {
-                        lastSequence.mlen = maxML;
-                        lastSequence.off = matches[nbMatches-1].off;
-                        lastSequence.litlen = litlen;
-                        cur -= (opt[cur].mlen==0) ? opt[cur].litlen : 0;  /* last sequence is actually only literals, fix cur to last match - note : may underflow, in which case, it's first sequence, and it's okay */
-                        last_pos = cur + ZSTD_totalLen(lastSequence);
-                        if (cur > ZSTD_OPT_NUM) cur = 0;   /* underflow => first match */
+                    if ( (longestML > sufficient_len)
+                      || (cur + longestML >= ZSTD_OPT_NUM)
+                      || (ip + cur + longestML >= iend) ) {
+                        lastStretch.mlen = longestML;
+                        lastStretch.off = matches[nbMatches-1].off;
+                        lastStretch.litlen = 0;
+                        last_pos = cur + longestML;
                         goto _shortestPath;
                 }   }
 
···
                     U32 const startML = (matchNb>0) ? matches[matchNb-1].len+1 : minMatch;
                     U32 mlen;
 
-                    DEBUGLOG(7, "testing match %u => offCode=%4u, mlen=%2u, llen=%2u",
-                                matchNb, matches[matchNb].off, lastML, litlen);
+                    DEBUGLOG(7, "testing match %u => offBase=%4u, mlen=%2u, llen=%2u",
+                                matchNb, matches[matchNb].off, lastML, opt[cur].litlen);
 
                     for (mlen = lastML; mlen >= startML; mlen--) {  /* scan downward */
                         U32 const pos = cur + mlen;
-                        int const price = (int)basePrice + (int)ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
+                        int const price = basePrice + (int)ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
 
                         if ((pos > last_pos) || (price < opt[pos].price)) {
                             DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)",
                                         pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price));
-                            while (last_pos < pos) { opt[last_pos+1].price = ZSTD_MAX_PRICE; last_pos++; }   /* fill empty positions */
+                            while (last_pos < pos) {
+                                /* fill empty positions, for future comparisons */
+                                last_pos++;
+                                opt[last_pos].price = ZSTD_MAX_PRICE;
+                                opt[last_pos].litlen = !0;  /* just needs to be != 0, to mean "not an end of match" */
+                            }
                             opt[pos].mlen = mlen;
                             opt[pos].off = offset;
-                            opt[pos].litlen = litlen;
+                            opt[pos].litlen = 0;
                             opt[pos].price = price;
                         } else {
                             DEBUGLOG(7, "rPos:%u (ml=%2u) => new price is worse (%.2f>=%.2f)",
···
                             if (optLevel==0) break;  /* early update abort; gets ~+10% speed for about -0.01 ratio loss */
                         }
             }   }   }
+            opt[last_pos+1].price = ZSTD_MAX_PRICE;
         }  /* for (cur = 1; cur <= last_pos; cur++) */
 
-        lastSequence = opt[last_pos];
-        cur = last_pos > ZSTD_totalLen(lastSequence) ? last_pos - ZSTD_totalLen(lastSequence) : 0;  /* single sequence, and it starts before `ip` */
-        assert(cur < ZSTD_OPT_NUM);  /* control overflow*/
+        lastStretch = opt[last_pos];
+        assert(cur >= lastStretch.mlen);
+        cur = last_pos - lastStretch.mlen;
 
 _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
         assert(opt[0].mlen == 0);
+        assert(last_pos >= lastStretch.mlen);
+        assert(cur == last_pos - lastStretch.mlen);
 
-        /* Set the next chunk's repcodes based on the repcodes of the beginning
-         * of the last match, and the last sequence. This avoids us having to
-         * update them while traversing the sequences.
-         */
-        if (lastSequence.mlen != 0) {
-            repcodes_t const reps = ZSTD_newRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0);
-            ZSTD_memcpy(rep, &reps, sizeof(reps));
+        if (lastStretch.mlen==0) {
+            /* no solution : all matches have been converted into literals */
+            assert(lastStretch.litlen == (ip - anchor) + last_pos);
+            ip += last_pos;
+            continue;
+        }
+        assert(lastStretch.off > 0);
+
+        /* Update offset history */
+        if (lastStretch.litlen == 0) {
+            /* finishing on a match : update offset history */
+            Repcodes_t const reps = ZSTD_newRep(opt[cur].rep, lastStretch.off, opt[cur].litlen==0);
+            ZSTD_memcpy(rep, &reps, sizeof(Repcodes_t));
         } else {
-            ZSTD_memcpy(rep, opt[cur].rep, sizeof(repcodes_t));
+            ZSTD_memcpy(rep, lastStretch.rep, sizeof(Repcodes_t));
+            assert(cur >= lastStretch.litlen);
+            cur -= lastStretch.litlen;
         }
 
-        {   U32 const storeEnd = cur + 1;
+        /* Let's write the shortest path solution.
+         * It is stored in @opt in reverse order,
+         * starting from @storeEnd (==cur+2),
+         * effectively partially @opt overwriting.
+         * Content is changed too:
+         * - So far, @opt stored stretches, aka a match followed by literals
+         * - Now, it will store sequences, aka literals followed by a match
+         */
+        {   U32 const storeEnd = cur + 2;
             U32 storeStart = storeEnd;
-            U32 seqPos = cur;
+            U32 stretchPos = cur;
 
             DEBUGLOG(6, "start reverse traversal (last_pos:%u, cur:%u)",
                         last_pos, cur); (void)last_pos;
-            assert(storeEnd < ZSTD_OPT_NUM);
-            DEBUGLOG(6, "last sequence copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
-                        storeEnd, lastSequence.litlen, lastSequence.mlen, lastSequence.off);
-            opt[storeEnd] = lastSequence;
-            while (seqPos > 0) {
-                U32 const backDist = ZSTD_totalLen(opt[seqPos]);
+            assert(storeEnd < ZSTD_OPT_SIZE);
+            DEBUGLOG(6, "last stretch copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
+                        storeEnd, lastStretch.litlen, lastStretch.mlen, lastStretch.off);
+            if (lastStretch.litlen > 0) {
+                /* last "sequence" is unfinished: just a bunch of literals */
+                opt[storeEnd].litlen = lastStretch.litlen;
+                opt[storeEnd].mlen = 0;
+                storeStart = storeEnd-1;
+                opt[storeStart] = lastStretch;
+            } {
+                opt[storeEnd] = lastStretch;  /* note: litlen will be fixed */
+                storeStart = storeEnd;
+            }
+            while (1) {
+                ZSTD_optimal_t nextStretch = opt[stretchPos];
+                opt[storeStart].litlen = nextStretch.litlen;
+                DEBUGLOG(6, "selected sequence (llen=%u,mlen=%u,ofc=%u)",
+                            opt[storeStart].litlen, opt[storeStart].mlen, opt[storeStart].off);
+                if (nextStretch.mlen == 0) {
+                    /* reaching beginning of segment */
+                    break;
+                }
                 storeStart--;
-                DEBUGLOG(6, "sequence from rPos=%u copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
-                            seqPos, storeStart, opt[seqPos].litlen, opt[seqPos].mlen, opt[seqPos].off);
-                opt[storeStart] = opt[seqPos];
-                seqPos = (seqPos > backDist) ? seqPos - backDist : 0;
+                opt[storeStart] = nextStretch; /* note: litlen will be fixed */
+                assert(nextStretch.litlen + nextStretch.mlen <= stretchPos);
+                stretchPos -= nextStretch.litlen + nextStretch.mlen;
             }
 
             /* save sequences */
-            DEBUGLOG(6, "sending selected sequences into seqStore")
+            DEBUGLOG(6, "sending selected sequences into seqStore");
             {   U32 storePos;
                 for (storePos=storeStart; storePos <= storeEnd; storePos++) {
                     U32 const llen = opt[storePos].litlen;
                     U32 const mlen = opt[storePos].mlen;
-                    U32 const offCode = opt[storePos].off;
+                    U32 const offBase = opt[storePos].off;
                     U32 const advance = llen + mlen;
-                    DEBUGLOG(6, "considering seq starting at %zi, llen=%u, mlen=%u",
-                                anchor - istart, (unsigned)llen, (unsigned)mlen);
+                    DEBUGLOG(6, "considering seq starting at %i, llen=%u, mlen=%u",
+                                (int)(anchor - istart), (unsigned)llen, (unsigned)mlen);
 
                     if (mlen==0) {  /* only literals => must be last "sequence", actually starting a new stream of sequences */
                         assert(storePos == storeEnd);   /* must be last sequence */
···
                     }
 
                     assert(anchor + llen <= iend);
-                    ZSTD_updateStats(optStatePtr, llen, anchor, offCode, mlen);
-                    ZSTD_storeSeq(seqStore, llen, anchor, iend, offCode, mlen);
+                    ZSTD_updateStats(optStatePtr, llen, anchor, offBase, mlen);
+                    ZSTD_storeSeq(seqStore, llen, anchor, iend, offBase, mlen);
                     anchor += advance;
                     ip = anchor;
             }   }
+            DEBUGLOG(7, "new offset history : %u, %u, %u", rep[0], rep[1], rep[2]);
+
+            /* update all costs */
             ZSTD_setBasePrices(optStatePtr, optLevel);
         }
     }   /* while (ip < ilimit) */
···
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
+#endif /* build exclusions */
 
+#ifndef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
 static size_t ZSTD_compressBlock_opt0(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode)
 {
     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /* optLevel */, dictMode);
 }
+#endif
 
+#ifndef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
 static size_t ZSTD_compressBlock_opt2(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode)
 {
     return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /* optLevel */, dictMode);
 }
+#endif
 
+#ifndef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_btopt(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_compressBlock_btopt");
     return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
 }
+#endif
 
 
 
 
+#ifndef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
 /* ZSTD_initStats_ultra():
  * make a first compression pass, just to seed stats with more accurate starting values.
  * only works on first block, with no dictionary and no ldm.
- * this function cannot error, hence its contract must be respected.
+ * this function cannot error out, its narrow contract must be respected.
  */
-static void
-ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
-                     seqStore_t* seqStore,
-                     U32 rep[ZSTD_REP_NUM],
-               const void* src, size_t srcSize)
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_initStats_ultra(ZSTD_MatchState_t* ms,
+                          SeqStore_t* seqStore,
+                          U32 rep[ZSTD_REP_NUM],
+                    const void* src, size_t srcSize)
 {
     U32 tmpRep[ZSTD_REP_NUM];  /* updated rep codes will sink here */
     ZSTD_memcpy(tmpRep, rep, sizeof(tmpRep));
···
 
     ZSTD_compressBlock_opt2(ms, seqStore, tmpRep, src, srcSize, ZSTD_noDict);   /* generate stats into ms->opt*/
 
-    /* invalidate first scan from history */
+    /* invalidate first scan from history, only keep entropy stats */
     ZSTD_resetSeqStore(seqStore);
     ms->window.base -= srcSize;
     ms->window.dictLimit += (U32)srcSize;
···
 }
 
 size_t ZSTD_compressBlock_btultra(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_compressBlock_btultra (srcSize=%zu)", srcSize);
···
 }
 
 size_t ZSTD_compressBlock_btultra2(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
     U32 const curr = (U32)((const BYTE*)src - ms->window.base);
     DEBUGLOG(5, "ZSTD_compressBlock_btultra2 (srcSize=%zu)", srcSize);
 
-    /* 2-pass strategy:
+    /* 2-passes strategy:
      * this strategy makes a first pass over first block to collect statistics
-     * and seed next round's statistics with it.
-     * After 1st pass, function forgets everything, and starts a new block.
+     * in order to seed next round's statistics with it.
+     * After 1st pass, function forgets history, and starts a new block.
      * Consequently, this can only work if no data has been previously loaded in tables,
      * aka, no dictionary, no prefix, no ldm preprocessing.
      * The compression ratio gain is generally small (~0.5% on first block),
···
     if ( (ms->opt.litLengthSum==0)   /* first block */
       && (seqStore->sequences == seqStore->sequencesStart)  /* no ldm */
       && (ms->window.dictLimit == ms->window.lowLimit)   /* no dictionary */
-      && (curr == ms->window.dictLimit)   /* start of frame, nothing already loaded nor skipped */
-      && (srcSize > ZSTD_PREDEF_THRESHOLD)
+      && (curr == ms->window.dictLimit)    /* start of frame, nothing already loaded nor skipped */
+      && (srcSize > ZSTD_PREDEF_THRESHOLD) /* input large enough to not employ default stats */
       ) {
         ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize);
     }
 
     return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
 }
+#endif
 
+#ifndef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_btopt_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
     return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState);
 }
 
+size_t ZSTD_compressBlock_btopt_extDict(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        const void* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_btultra_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
     return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState);
 }
 
-size_t ZSTD_compressBlock_btopt_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        const void* src, size_t srcSize)
-{
-    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
-}
-
 size_t ZSTD_compressBlock_btultra_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
     return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
 }
+#endif
 
 /* note : no btultra2 variant for extDict nor dictMatchState,
  * because btultra2 is not meant to work with dictionaries

lib/zstd/compress/zstd_opt.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #ifndef ZSTD_OPT_H
 #define ZSTD_OPT_H
 
-
 #include "zstd_compress_internal.h"
 
+#if !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR)
 /* used in ZSTD_loadDictionaryContent() */
-void ZSTD_updateTree(ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iend);
+void ZSTD_updateTree(ZSTD_MatchState_t* ms, const BYTE* ip, const BYTE* iend);
+#endif
 
+#ifndef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_btopt(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_btultra(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-size_t ZSTD_compressBlock_btultra2(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-
-
 size_t ZSTD_compressBlock_btopt_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_btopt_extDict(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+
+#define ZSTD_COMPRESSBLOCK_BTOPT ZSTD_compressBlock_btopt
+#define ZSTD_COMPRESSBLOCK_BTOPT_DICTMATCHSTATE ZSTD_compressBlock_btopt_dictMatchState
+#define ZSTD_COMPRESSBLOCK_BTOPT_EXTDICT ZSTD_compressBlock_btopt_extDict
+#else
+#define ZSTD_COMPRESSBLOCK_BTOPT NULL
+#define ZSTD_COMPRESSBLOCK_BTOPT_DICTMATCHSTATE NULL
+#define ZSTD_COMPRESSBLOCK_BTOPT_EXTDICT NULL
+#endif
+
+#ifndef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_btultra(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_btultra_dictMatchState(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize);
-
-size_t ZSTD_compressBlock_btopt_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_btultra_extDict(
-        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 
         /* note : no btultra2 variant for extDict nor dictMatchState,
          * because btultra2 is not meant to work with dictionaries
          * and is only specific for the first block (no prefix) */
+size_t ZSTD_compressBlock_btultra2(
+        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 
+#define ZSTD_COMPRESSBLOCK_BTULTRA ZSTD_compressBlock_btultra
+#define ZSTD_COMPRESSBLOCK_BTULTRA_DICTMATCHSTATE ZSTD_compressBlock_btultra_dictMatchState
+#define ZSTD_COMPRESSBLOCK_BTULTRA_EXTDICT ZSTD_compressBlock_btultra_extDict
+#define ZSTD_COMPRESSBLOCK_BTULTRA2 ZSTD_compressBlock_btultra2
+#else
+#define ZSTD_COMPRESSBLOCK_BTULTRA NULL
+#define ZSTD_COMPRESSBLOCK_BTULTRA_DICTMATCHSTATE NULL
+#define ZSTD_COMPRESSBLOCK_BTULTRA_EXTDICT NULL
+#define ZSTD_COMPRESSBLOCK_BTULTRA2 NULL
+#endif
 
 #endif /* ZSTD_OPT_H */

lib/zstd/compress/zstd_preSplit.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#include "../common/compiler.h" /* ZSTD_ALIGNOF */
+#include "../common/mem.h" /* S64 */
+#include "../common/zstd_deps.h" /* ZSTD_memset */
+#include "../common/zstd_internal.h" /* ZSTD_STATIC_ASSERT */
+#include "hist.h" /* HIST_add */
+#include "zstd_preSplit.h"
+
+
+#define BLOCKSIZE_MIN 3500
+#define THRESHOLD_PENALTY_RATE 16
+#define THRESHOLD_BASE (THRESHOLD_PENALTY_RATE - 2)
+#define THRESHOLD_PENALTY 3
+
+#define HASHLENGTH 2
+#define HASHLOG_MAX 10
+#define HASHTABLESIZE (1 << HASHLOG_MAX)
+#define HASHMASK (HASHTABLESIZE - 1)
+#define KNUTH 0x9e3779b9
+
+/* for hashLog > 8, hash 2 bytes.
+ * for hashLog == 8, just take the byte, no hashing.
+ * The speed of this method relies on compile-time constant propagation */
+FORCE_INLINE_TEMPLATE unsigned hash2(const void *p, unsigned hashLog)
+{
+    assert(hashLog >= 8);
+    if (hashLog == 8) return (U32)((const BYTE*)p)[0];
+    assert(hashLog <= HASHLOG_MAX);
+    return (U32)(MEM_read16(p)) * KNUTH >> (32 - hashLog);
+}
+
+
+typedef struct {
+  unsigned events[HASHTABLESIZE];
+  size_t nbEvents;
+} Fingerprint;
+typedef struct {
+    Fingerprint pastEvents;
+    Fingerprint newEvents;
+} FPStats;
+
+static void initStats(FPStats* fpstats)
+{
+    ZSTD_memset(fpstats, 0, sizeof(FPStats));
+}
+
+FORCE_INLINE_TEMPLATE void
+addEvents_generic(Fingerprint* fp, const void* src, size_t srcSize, size_t samplingRate, unsigned hashLog)
+{
+    const char* p = (const char*)src;
+    size_t limit = srcSize - HASHLENGTH + 1;
+    size_t n;
+    assert(srcSize >= HASHLENGTH);
+    for (n = 0; n < limit; n+=samplingRate) {
+        fp->events[hash2(p+n, hashLog)]++;
+    }
+    fp->nbEvents += limit/samplingRate;
+}
+
+FORCE_INLINE_TEMPLATE void
+recordFingerprint_generic(Fingerprint* fp, const void* src, size_t srcSize, size_t samplingRate, unsigned hashLog)
+{
+    ZSTD_memset(fp, 0, sizeof(unsigned) * ((size_t)1 << hashLog));
+    fp->nbEvents = 0;
+    addEvents_generic(fp, src, srcSize, samplingRate, hashLog);
+}
+
+typedef void (*RecordEvents_f)(Fingerprint* fp, const void* src, size_t srcSize);
+
+#define FP_RECORD(_rate) ZSTD_recordFingerprint_##_rate
+
+#define ZSTD_GEN_RECORD_FINGERPRINT(_rate, _hSize)                                 \
+    static void FP_RECORD(_rate)(Fingerprint* fp, const void* src, size_t srcSize) \
+    {                                                                              \
+        recordFingerprint_generic(fp, src, srcSize, _rate, _hSize);                \
+    }
+
+ZSTD_GEN_RECORD_FINGERPRINT(1, 10)
+ZSTD_GEN_RECORD_FINGERPRINT(5, 10)
+ZSTD_GEN_RECORD_FINGERPRINT(11, 9)
+ZSTD_GEN_RECORD_FINGERPRINT(43, 8)
+
+
+static U64 abs64(S64 s64) { return (U64)((s64 < 0) ? -s64 : s64); }
+
+static U64 fpDistance(const Fingerprint* fp1, const Fingerprint* fp2, unsigned hashLog)
+{
+    U64 distance = 0;
+    size_t n;
+    assert(hashLog <= HASHLOG_MAX);
+    for (n = 0; n < ((size_t)1 << hashLog); n++) {
+        distance +=
+            abs64((S64)fp1->events[n] * (S64)fp2->nbEvents - (S64)fp2->events[n] * (S64)fp1->nbEvents);
+    }
+    return distance;
+}
+
+/* Compare newEvents with pastEvents
+ * return 1 when considered "too different"
+ */
+static int compareFingerprints(const Fingerprint* ref,
+                            const Fingerprint* newfp,
+                            int penalty,
+                            unsigned hashLog)
+{
+    assert(ref->nbEvents > 0);
+    assert(newfp->nbEvents > 0);
+    {   U64 p50 = (U64)ref->nbEvents * (U64)newfp->nbEvents;
+        U64 deviation = fpDistance(ref, newfp, hashLog);
+        U64 threshold = p50 * (U64)(THRESHOLD_BASE + penalty) / THRESHOLD_PENALTY_RATE;
+        return deviation >= threshold;
+    }
+}
+
+static void mergeEvents(Fingerprint* acc, const Fingerprint* newfp)
+{
+    size_t n;
+    for (n = 0; n < HASHTABLESIZE; n++) {
+        acc->events[n] += newfp->events[n];
+    }
+    acc->nbEvents += newfp->nbEvents;
+}
+
+static void flushEvents(FPStats* fpstats)
+{
+    size_t n;
+    for (n = 0; n < HASHTABLESIZE; n++) {
+        fpstats->pastEvents.events[n] = fpstats->newEvents.events[n];
+    }
+    fpstats->pastEvents.nbEvents = fpstats->newEvents.nbEvents;
+    ZSTD_memset(&fpstats->newEvents, 0, sizeof(fpstats->newEvents));
+}
+
+static void removeEvents(Fingerprint* acc, const Fingerprint* slice)
+{
+    size_t n;
+    for (n = 0; n < HASHTABLESIZE; n++) {
+        assert(acc->events[n] >= slice->events[n]);
+        acc->events[n] -= slice->events[n];
+    }
+    acc->nbEvents -= slice->nbEvents;
+}
+
+#define CHUNKSIZE (8 << 10)
+static size_t ZSTD_splitBlock_byChunks(const void* blockStart, size_t blockSize,
+                        int level,
+                        void* workspace, size_t wkspSize)
+{
+    static const RecordEvents_f records_fs[] = {
+        FP_RECORD(43), FP_RECORD(11), FP_RECORD(5), FP_RECORD(1)
+    };
+    static const unsigned hashParams[] = { 8, 9, 10, 10 };
+    const RecordEvents_f record_f = (assert(0<=level && level<=3), records_fs[level]);
+    FPStats* const fpstats = (FPStats*)workspace;
+    const char* p = (const char*)blockStart;
+    int penalty = THRESHOLD_PENALTY;
+    size_t pos = 0;
+    assert(blockSize == (128 << 10));
+    assert(workspace != NULL);
+    assert((size_t)workspace % ZSTD_ALIGNOF(FPStats) == 0);
+    ZSTD_STATIC_ASSERT(ZSTD_SLIPBLOCK_WORKSPACESIZE >= sizeof(FPStats));
+    assert(wkspSize >= sizeof(FPStats)); (void)wkspSize;
+
+    initStats(fpstats);
+    record_f(&fpstats->pastEvents, p, CHUNKSIZE);
+    for (pos = CHUNKSIZE; pos <= blockSize - CHUNKSIZE; pos += CHUNKSIZE) {
+        record_f(&fpstats->newEvents, p + pos, CHUNKSIZE);
+        if (compareFingerprints(&fpstats->pastEvents, &fpstats->newEvents, penalty, hashParams[level])) {
+            return pos;
+        } else {
+            mergeEvents(&fpstats->pastEvents, &fpstats->newEvents);
+            if (penalty > 0) penalty--;
+        }
+    }
+    assert(pos == blockSize);
+    return blockSize;
+    (void)flushEvents; (void)removeEvents;
+}
+
+/* ZSTD_splitBlock_fromBorders(): very fast strategy :
+ * compare fingerprint from beginning and end of the block,
+ * derive from their difference if it's preferable to split in the middle,
+ * repeat the process a second time, for finer grained decision.
+ * 3 times did not brought improvements, so I stopped at 2.
+ * Benefits are good enough for a cheap heuristic.
+ * More accurate splitting saves more, but speed impact is also more perceptible.
+ * For better accuracy, use more elaborate variant *_byChunks.
+ */
+static size_t ZSTD_splitBlock_fromBorders(const void* blockStart, size_t blockSize,
+                        void* workspace, size_t wkspSize)
+{
+#define SEGMENT_SIZE 512
+    FPStats* const fpstats = (FPStats*)workspace;
+    Fingerprint* middleEvents = (Fingerprint*)(void*)((char*)workspace + 512 * sizeof(unsigned));
+    assert(blockSize == (128 << 10));
+    assert(workspace != NULL);
+    assert((size_t)workspace % ZSTD_ALIGNOF(FPStats) == 0);
+    ZSTD_STATIC_ASSERT(ZSTD_SLIPBLOCK_WORKSPACESIZE >= sizeof(FPStats));
+    assert(wkspSize >= sizeof(FPStats)); (void)wkspSize;
+
+    initStats(fpstats);
+    HIST_add(fpstats->pastEvents.events, blockStart, SEGMENT_SIZE);
+    HIST_add(fpstats->newEvents.events, (const char*)blockStart + blockSize - SEGMENT_SIZE, SEGMENT_SIZE);
+    fpstats->pastEvents.nbEvents = fpstats->newEvents.nbEvents = SEGMENT_SIZE;
+    if (!compareFingerprints(&fpstats->pastEvents, &fpstats->newEvents, 0, 8))
+        return blockSize;
+
+    HIST_add(middleEvents->events, (const char*)blockStart + blockSize/2 - SEGMENT_SIZE/2, SEGMENT_SIZE);
+    middleEvents->nbEvents = SEGMENT_SIZE;
+    {   U64 const distFromBegin = fpDistance(&fpstats->pastEvents, middleEvents, 8);
+        U64 const distFromEnd = fpDistance(&fpstats->newEvents, middleEvents, 8);
+        U64 const minDistance = SEGMENT_SIZE * SEGMENT_SIZE / 3;
+        if (abs64((S64)distFromBegin - (S64)distFromEnd) < minDistance)
+            return 64 KB;
+        return (distFromBegin > distFromEnd) ? 32 KB : 96 KB;
+    }
+}
+
+size_t ZSTD_splitBlock(const void* blockStart, size_t blockSize,
+                    int level,
+                    void* workspace, size_t wkspSize)
+{
+    DEBUGLOG(6, "ZSTD_splitBlock (level=%i)", level);
+    assert(0<=level && level<=4);
+    if (level == 0)
+        return ZSTD_splitBlock_fromBorders(blockStart, blockSize, workspace, wkspSize);
+    /* level >= 1*/
+    return ZSTD_splitBlock_byChunks(blockStart, blockSize, level-1, workspace, wkspSize);
+}

lib/zstd/compress/zstd_preSplit.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_PRESPLIT_H
+#define ZSTD_PRESPLIT_H
+
+#include <linux/types.h>  /* size_t */
+
+#define ZSTD_SLIPBLOCK_WORKSPACESIZE 8208
+
+/* ZSTD_splitBlock():
+ * @level must be a value between 0 and 4.
+ *        higher levels spend more energy to detect block boundaries.
+ * @workspace must be aligned for size_t.
+ * @wkspSize must be at least >= ZSTD_SLIPBLOCK_WORKSPACESIZE
+ * note:
+ * For the time being, this function only accepts full 128 KB blocks.
+ * Therefore, @blockSize must be == 128 KB.
+ * While this could be extended to smaller sizes in the future,
+ * it is not yet clear if this would be useful. TBD.
+ */
+size_t ZSTD_splitBlock(const void* blockStart, size_t blockSize,
+                    int level,
+                    void* workspace, size_t wkspSize);
+
+#endif /* ZSTD_PRESPLIT_H */

lib/zstd/decompress/huf_decompress.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /* ******************************************************************
  * huff0 huffman decoder,
  * part of Finite State Entropy library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
···
 #include "../common/compiler.h"
 #include "../common/bitstream.h"  /* BIT_* */
 #include "../common/fse.h"        /* to compress headers */
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"
 #include "../common/error_private.h"
 #include "../common/zstd_internal.h"
+#include "../common/bits.h"       /* ZSTD_highbit32, ZSTD_countTrailingZeros64 */
 
 /* **************************************************************
 *  Constants
···
 *  Macros
 ****************************************************************/
 
+#ifdef HUF_DISABLE_FAST_DECODE
+# define HUF_ENABLE_FAST_DECODE 0
+#else
+# define HUF_ENABLE_FAST_DECODE 1
+#endif
+
 /* These two optional macros force the use one way or another of the two
  * Huffman decompression implementations. You can't force in both directions
  * at the same time.
···
 #error "Cannot force the use of the X1 and X2 decoders at the same time!"
 #endif
 
-#if ZSTD_ENABLE_ASM_X86_64_BMI2 && DYNAMIC_BMI2
-# define HUF_ASM_X86_64_BMI2_ATTRS BMI2_TARGET_ATTRIBUTE
+/* When DYNAMIC_BMI2 is enabled, fast decoders are only called when bmi2 is
+ * supported at runtime, so we can add the BMI2 target attribute.
+ * When it is disabled, we will still get BMI2 if it is enabled statically.
+ */
+#if DYNAMIC_BMI2
+# define HUF_FAST_BMI2_ATTRS BMI2_TARGET_ATTRIBUTE
 #else
-# define HUF_ASM_X86_64_BMI2_ATTRS
+# define HUF_FAST_BMI2_ATTRS
 #endif
 
 #define HUF_EXTERN_C
 #define HUF_ASM_DECL HUF_EXTERN_C
 
-#if DYNAMIC_BMI2 || (ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__))
+#if DYNAMIC_BMI2
 # define HUF_NEED_BMI2_FUNCTION 1
 #else
 # define HUF_NEED_BMI2_FUNCTION 0
-#endif
-
-#if !(ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__))
-# define HUF_NEED_DEFAULT_FUNCTION 1
-#else
-# define HUF_NEED_DEFAULT_FUNCTION 0
 #endif
 
 /* **************************************************************
···
 /* **************************************************************
 *  BMI2 Variant Wrappers
 ****************************************************************/
+typedef size_t (*HUF_DecompressUsingDTableFn)(void *dst, size_t dstSize,
+                                              const void *cSrc,
+                                              size_t cSrcSize,
+                                              const HUF_DTable *DTable);
+
 #if DYNAMIC_BMI2
 
 #define HUF_DGEN(fn)                                                        \
···
     }                                                                       \
                                                                             \
     static size_t fn(void* dst, size_t dstSize, void const* cSrc,           \
-                     size_t cSrcSize, HUF_DTable const* DTable, int bmi2)   \
+                     size_t cSrcSize, HUF_DTable const* DTable, int flags)  \
     {                                                                       \
-        if (bmi2) {                                                         \
+        if (flags & HUF_flags_bmi2) {                                       \
             return fn##_bmi2(dst, dstSize, cSrc, cSrcSize, DTable);         \
         }                                                                   \
         return fn##_default(dst, dstSize, cSrc, cSrcSize, DTable);          \
···
 
 #define HUF_DGEN(fn)                                                        \
     static size_t fn(void* dst, size_t dstSize, void const* cSrc,           \
-                     size_t cSrcSize, HUF_DTable const* DTable, int bmi2)   \
+                     size_t cSrcSize, HUF_DTable const* DTable, int flags)  \
     {                                                                       \
-        (void)bmi2;                                                         \
+        (void)flags;                                                        \
         return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
     }
 
···
     return dtd;
 }
 
-#if ZSTD_ENABLE_ASM_X86_64_BMI2
-
-static size_t HUF_initDStream(BYTE const* ip) {
+static size_t HUF_initFastDStream(BYTE const* ip) {
     BYTE const lastByte = ip[7];
-    size_t const bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;
+    size_t const bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
     size_t const value = MEM_readLEST(ip) | 1;
     assert(bitsConsumed <= 8);
+    assert(sizeof(size_t) == 8);
     return value << bitsConsumed;
 }
+
+
+/*
+ * The input/output arguments to the Huffman fast decoding loop:
+ *
+ * ip [in/out] - The input pointers, must be updated to reflect what is consumed.
+ * op [in/out] - The output pointers, must be updated to reflect what is written.
+ * bits [in/out] - The bitstream containers, must be updated to reflect the current state.
+ * dt [in] - The decoding table.
+ * ilowest [in] - The beginning of the valid range of the input. Decoders may read
+ *                down to this pointer. It may be below iend[0].
+ * oend [in] - The end of the output stream. op[3] must not cross oend.
+ * iend [in] - The end of each input stream. ip[i] may cross iend[i],
+ *             as long as it is above ilowest, but that indicates corruption.
+ */
 typedef struct {
     BYTE const* ip[4];
     BYTE* op[4];
     U64 bits[4];
     void const* dt;
-    BYTE const* ilimit;
+    BYTE const* ilowest;
     BYTE* oend;
     BYTE const* iend[4];
-} HUF_DecompressAsmArgs;
+} HUF_DecompressFastArgs;
+
+typedef void (*HUF_DecompressFastLoopFn)(HUF_DecompressFastArgs*);
 
 /*
- * Initializes args for the asm decoding loop.
- * @returns 0 on success
- *          1 if the fallback implementation should be used.
+ * Initializes args for the fast decoding loop.
+ * @returns 1 on success
+ *          0 if the fallback implementation should be used.
  *          Or an error code on failure.
  */
-static size_t HUF_DecompressAsmArgs_init(HUF_DecompressAsmArgs* args, void* dst, size_t dstSize, void const* src, size_t srcSize, const HUF_DTable* DTable)
+static size_t HUF_DecompressFastArgs_init(HUF_DecompressFastArgs* args, void* dst, size_t dstSize, void const* src, size_t srcSize, const HUF_DTable* DTable)
 {
     void const* dt = DTable + 1;
     U32 const dtLog = HUF_getDTableDesc(DTable).tableLog;
 
-    const BYTE* const ilimit = (const BYTE*)src + 6 + 8;
+    const BYTE* const istart = (const BYTE*)src;
 
-    BYTE* const oend = (BYTE*)dst + dstSize;
+    BYTE* const oend = ZSTD_maybeNullPtrAdd((BYTE*)dst, dstSize);
 
-    /* The following condition is false on x32 platform,
-     * but HUF_asm is not compatible with this ABI */
-    if (!(MEM_isLittleEndian() && !MEM_32bits())) return 1;
+    /* The fast decoding loop assumes 64-bit little-endian.
+     * This condition is false on x32.
+     */
+    if (!MEM_isLittleEndian() || MEM_32bits())
+        return 0;
+
+    /* Avoid nullptr addition */
+    if (dstSize == 0)
+        return 0;
+    assert(dst != NULL);
 
     /* strict minimum : jump table + 1 byte per stream */
     if (srcSize < 10)
···
      * On small inputs we don't have enough data to trigger the fast loop, so use the old decoder.
      */
     if (dtLog != HUF_DECODER_FAST_TABLELOG)
-        return 1;
+        return 0;
 
     /* Read the jump table. */
     {
-        const BYTE* const istart = (const BYTE*)src;
         size_t const length1 = MEM_readLE16(istart);
         size_t const length2 = MEM_readLE16(istart+2);
         size_t const length3 = MEM_readLE16(istart+4);
···
         args->iend[2] = args->iend[1] + length2;
         args->iend[3] = args->iend[2] + length3;
 
-        /* HUF_initDStream() requires this, and this small of an input
+        /* HUF_initFastDStream() requires this, and this small of an input
          * won't benefit from the ASM loop anyways.
-         * length1 must be >= 16 so that ip[0] >= ilimit before the loop
-         * starts.
          */
-        if (length1 < 16 || length2 < 8 || length3 < 8 || length4 < 8)
-            return 1;
+        if (length1 < 8 || length2 < 8 || length3 < 8 || length4 < 8)
+            return 0;
         if (length4 > srcSize) return ERROR(corruption_detected);   /* overflow */
     }
     /* ip[] contains the position that is currently loaded into bits[]. */
···
 
     /* No point to call the ASM loop for tiny outputs. */
     if (args->op[3] >= oend)
-        return 1;
+        return 0;
 
     /* bits[] is the bit container.
         * It is read from the MSB down to the LSB.
···
         * set, so that CountTrailingZeros(bits[]) can be used
         * to count how many bits we've consumed.
         */
-    args->bits[0] = HUF_initDStream(args->ip[0]);
-    args->bits[1] = HUF_initDStream(args->ip[1]);
-    args->bits[2] = HUF_initDStream(args->ip[2]);
-    args->bits[3] = HUF_initDStream(args->ip[3]);
+    args->bits[0] = HUF_initFastDStream(args->ip[0]);
+    args->bits[1] = HUF_initFastDStream(args->ip[1]);
+    args->bits[2] = HUF_initFastDStream(args->ip[2]);
+    args->bits[3] = HUF_initFastDStream(args->ip[3]);
 
-    /* If ip[] >= ilimit, it is guaranteed to be safe to
-        * reload bits[]. It may be beyond its section, but is
-        * guaranteed to be valid (>= istart).
-        */
-    args->ilimit = ilimit;
+    /* The decoders must be sure to never read beyond ilowest.
+     * This is lower than iend[0], but allowing decoders to read
+     * down to ilowest can allow an extra iteration or two in the
+     * fast loop.
+     */
+    args->ilowest = istart;
 
     args->oend = oend;
     args->dt = dt;
 
-    return 0;
+    return 1;
 }
 
-static size_t HUF_initRemainingDStream(BIT_DStream_t* bit, HUF_DecompressAsmArgs const* args, int stream, BYTE* segmentEnd)
+static size_t HUF_initRemainingDStream(BIT_DStream_t* bit, HUF_DecompressFastArgs const* args, int stream, BYTE* segmentEnd)
 {
     /* Validate that we haven't overwritten. */
     if (args->op[stream] > segmentEnd)
···
         return ERROR(corruption_detected);
 
     /* Construct the BIT_DStream_t. */
-    bit->bitContainer = MEM_readLE64(args->ip[stream]);
-    bit->bitsConsumed = ZSTD_countTrailingZeros((size_t)args->bits[stream]);
-    bit->start = (const char*)args->iend[0];
+    assert(sizeof(size_t) == 8);
+    bit->bitContainer = MEM_readLEST(args->ip[stream]);
+    bit->bitsConsumed = ZSTD_countTrailingZeros64(args->bits[stream]);
+    bit->start = (const char*)args->ilowest;
     bit->limitPtr = bit->start + sizeof(size_t);
     bit->ptr = (const char*)args->ip[stream];
 
     return 0;
 }
-#endif
+
+/* Calls X(N) for each stream 0, 1, 2, 3. */
+#define HUF_4X_FOR_EACH_STREAM(X) \
+    do {                          \
+        X(0);                     \
+        X(1);                     \
+        X(2);                     \
+        X(3);                     \
+    } while (0)
+
+/* Calls X(N, var) for each stream 0, 1, 2, 3. */
+#define HUF_4X_FOR_EACH_STREAM_WITH_VAR(X, var) \
+    do {                                        \
+        X(0, (var));                            \
+        X(1, (var));                            \
+        X(2, (var));                            \
+        X(3, (var));                            \
+    } while (0)
 
 
 #ifndef HUF_FORCE_DECOMPRESS_X2
···
 static U64 HUF_DEltX1_set4(BYTE symbol, BYTE nbBits) {
     U64 D4;
     if (MEM_isLittleEndian()) {
-        D4 = (symbol << 8) + nbBits;
+        D4 = (U64)((symbol << 8) + nbBits);
     } else {
-        D4 = symbol + (nbBits << 8);
+        D4 = (U64)(symbol + (nbBits << 8));
     }
+    assert(D4 < (1U << 16));
     D4 *= 0x0001000100010001ULL;
     return D4;
 }
···
         BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1];
 } HUF_ReadDTableX1_Workspace;
 
-
-size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize)
-{
-    return HUF_readDTableX1_wksp_bmi2(DTable, src, srcSize, workSpace, wkspSize, /* bmi2 */ 0);
-}
-
-size_t HUF_readDTableX1_wksp_bmi2(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int bmi2)
+size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags)
 {
     U32 tableLog = 0;
     U32 nbSymbols = 0;
···
     DEBUG_STATIC_ASSERT(sizeof(DTableDesc) == sizeof(HUF_DTable));
     /* ZSTD_memset(huffWeight, 0, sizeof(huffWeight)); */   /* is not necessary, even though some analyzer complain ... */
 
-    iSize = HUF_readStats_wksp(wksp->huffWeight, HUF_SYMBOLVALUE_MAX + 1, wksp->rankVal, &nbSymbols, &tableLog, src, srcSize, wksp->statsWksp, sizeof(wksp->statsWksp), bmi2);
+    iSize = HUF_readStats_wksp(wksp->huffWeight, HUF_SYMBOLVALUE_MAX + 1, wksp->rankVal, &nbSymbols, &tableLog, src, srcSize, wksp->statsWksp, sizeof(wksp->statsWksp), flags);
     if (HUF_isError(iSize)) return iSize;
 
 
···
      * rankStart[0] is not filled because there are no entries in the table for
      * weight 0.
      */
-    {
-        int n;
-        int nextRankStart = 0;
+    {   int n;
+        U32 nextRankStart = 0;
         int const unroll = 4;
         int const nLimit = (int)nbSymbols - unroll + 1;
         for (n=0; n<(int)tableLog+1; n++) {
···
      * We can switch based on the length to a different inner loop which is
      * optimized for that particular case.
      */
-    {
-        U32 w;
-        int symbol=wksp->rankVal[0];
-        int rankStart=0;
+    {   U32 w;
+        int symbol = wksp->rankVal[0];
+        int rankStart = 0;
         for (w=1; w<tableLog+1; ++w) {
             int const symbolCount = wksp->rankVal[w];
             int const length = (1 << w) >> 1;
···
 }
 
 #define HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr) \
-    *ptr++ = HUF_decodeSymbolX1(DStreamPtr, dt, dtLog)
+    do { *ptr++ = HUF_decodeSymbolX1(DStreamPtr, dt, dtLog); } while (0)
 
-#define HUF_DECODE_SYMBOLX1_1(ptr, DStreamPtr)  \
-    if (MEM_64bits() || (HUF_TABLELOG_MAX<=12)) \
-        HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr)
+#define HUF_DECODE_SYMBOLX1_1(ptr, DStreamPtr)      \
+    do {                                            \
+        if (MEM_64bits() || (HUF_TABLELOG_MAX<=12)) \
+            HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr); \
+    } while (0)
 
-#define HUF_DECODE_SYMBOLX1_2(ptr, DStreamPtr) \
-    if (MEM_64bits()) \
-        HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr)
+#define HUF_DECODE_SYMBOLX1_2(ptr, DStreamPtr)      \
+    do {                                            \
+        if (MEM_64bits())                           \
+            HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr); \
+    } while (0)
 
 HINT_INLINE size_t
 HUF_decodeStreamX1(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, const HUF_DEltX1* const dt, const U32 dtLog)
···
     while (p < pEnd)
         HUF_DECODE_SYMBOLX1_0(p, bitDPtr);
 
-    return pEnd-pStart;
+    return (size_t)(pEnd-pStart);
 }
 
 FORCE_INLINE_TEMPLATE size_t
···
     const HUF_DTable* DTable)
 {
     BYTE* op = (BYTE*)dst;
-    BYTE* const oend = op + dstSize;
+    BYTE* const oend = ZSTD_maybeNullPtrAdd(op, dstSize);
     const void* dtPtr = DTable + 1;
     const HUF_DEltX1* const dt = (const HUF_DEltX1*)dtPtr;
     BIT_DStream_t bitD;
···
     return dstSize;
 }
 
+/* HUF_decompress4X1_usingDTable_internal_body():
+ * Conditions :
+ * @dstSize >= 6
+ */
 FORCE_INLINE_TEMPLATE size_t
 HUF_decompress4X1_usingDTable_internal_body(
           void* dst,  size_t dstSize,
···
 {
     /* Check */
     if (cSrcSize < 10) return ERROR(corruption_detected);  /* strict minimum : jump table + 1 byte per stream */
+    if (dstSize < 6) return ERROR(corruption_detected);         /* stream 4-split doesn't work */
 
     {   const BYTE* const istart = (const BYTE*) cSrc;
         BYTE* const ostart = (BYTE*) dst;
···
 
         if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
         if (opStart4 > oend) return ERROR(corruption_detected);      /* overflow */
+        assert(dstSize >= 6); /* validated above */
         CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
         CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
         CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
···
 }
 #endif
 
-#if HUF_NEED_DEFAULT_FUNCTION
 static
 size_t HUF_decompress4X1_usingDTable_internal_default(void* dst, size_t dstSize, void const* cSrc,
                     size_t cSrcSize, HUF_DTable const* DTable) {
     return HUF_decompress4X1_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
 }
-#endif
 
 #if ZSTD_ENABLE_ASM_X86_64_BMI2
 
-HUF_ASM_DECL void HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop(HUF_DecompressAsmArgs* args) ZSTDLIB_HIDDEN;
+HUF_ASM_DECL void HUF_decompress4X1_usingDTable_internal_fast_asm_loop(HUF_DecompressFastArgs* args) ZSTDLIB_HIDDEN;
 
-static HUF_ASM_X86_64_BMI2_ATTRS
-size_t
-HUF_decompress4X1_usingDTable_internal_bmi2_asm(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
+#endif
+
+static HUF_FAST_BMI2_ATTRS
+void HUF_decompress4X1_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs* args)
 {
-    void const* dt = DTable + 1;
-    const BYTE* const iend = (const BYTE*)cSrc + 6;
-    BYTE* const oend = (BYTE*)dst + dstSize;
-    HUF_DecompressAsmArgs args;
-    {
-        size_t const ret = HUF_DecompressAsmArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
-        FORWARD_IF_ERROR(ret, "Failed to init asm args");
-        if (ret != 0)
-            return HUF_decompress4X1_usingDTable_internal_bmi2(dst, dstSize, cSrc, cSrcSize, DTable);
+    U64 bits[4];
+    BYTE const* ip[4];
+    BYTE* op[4];
+    U16 const* const dtable = (U16 const*)args->dt;
+    BYTE* const oend = args->oend;
+    BYTE const* const ilowest = args->ilowest;
+
+    /* Copy the arguments to local variables */
+    ZSTD_memcpy(&bits, &args->bits, sizeof(bits));
+    ZSTD_memcpy((void*)(&ip), &args->ip, sizeof(ip));
+    ZSTD_memcpy(&op, &args->op, sizeof(op));
+
+    assert(MEM_isLittleEndian());
+    assert(!MEM_32bits());
+
+    for (;;) {
+        BYTE* olimit;
+        int stream;
+
+        /* Assert loop preconditions */
+#ifndef NDEBUG
+        for (stream = 0; stream < 4; ++stream) {
+            assert(op[stream] <= (stream == 3 ? oend : op[stream + 1]));
+            assert(ip[stream] >= ilowest);
+        }
+#endif
+        /* Compute olimit */
+        {
+            /* Each iteration produces 5 output symbols per stream */
+            size_t const oiters = (size_t)(oend - op[3]) / 5;
+            /* Each iteration consumes up to 11 bits * 5 = 55 bits < 7 bytes
+             * per stream.
+             */
+            size_t const iiters = (size_t)(ip[0] - ilowest) / 7;
+            /* We can safely run iters iterations before running bounds checks */
+            size_t const iters = MIN(oiters, iiters);
+            size_t const symbols = iters * 5;
+
+            /* We can simply check that op[3] < olimit, instead of checking all
+             * of our bounds, since we can't hit the other bounds until we've run
+             * iters iterations, which only happens when op[3] == olimit.
+             */
+            olimit = op[3] + symbols;
+
+            /* Exit fast decoding loop once we reach the end. */
+            if (op[3] == olimit)
+                break;
+
+            /* Exit the decoding loop if any input pointer has crossed the
+             * previous one. This indicates corruption, and a precondition
+             * to our loop is that ip[i] >= ip[0].
+             */
+            for (stream = 1; stream < 4; ++stream) {
+                if (ip[stream] < ip[stream - 1])
+                    goto _out;
+            }
+        }
+
+#ifndef NDEBUG
+        for (stream = 1; stream < 4; ++stream) {
+            assert(ip[stream] >= ip[stream - 1]);
+        }
+#endif
+
+#define HUF_4X1_DECODE_SYMBOL(_stream, _symbol)                 \
+    do {                                                        \
+        int const index = (int)(bits[(_stream)] >> 53);         \
+        int const entry = (int)dtable[index];                   \
+        bits[(_stream)] <<= (entry & 0x3F);                     \
+        op[(_stream)][(_symbol)] = (BYTE)((entry >> 8) & 0xFF); \
+    } while (0)
+
+#define HUF_4X1_RELOAD_STREAM(_stream)                              \
+    do {                                                            \
+        int const ctz = ZSTD_countTrailingZeros64(bits[(_stream)]); \
+        int const nbBits = ctz & 7;                                 \
+        int const nbBytes = ctz >> 3;                               \
+        op[(_stream)] += 5;                                         \
+        ip[(_stream)] -= nbBytes;                                   \
+        bits[(_stream)] = MEM_read64(ip[(_stream)]) | 1;            \
+        bits[(_stream)] <<= nbBits;                                 \
+    } while (0)
+
+        /* Manually unroll the loop because compilers don't consistently
+         * unroll the inner loops, which destroys performance.
+         */
+        do {
+            /* Decode 5 symbols in each of the 4 streams */
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 0);
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 1);
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 2);
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 3);
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 4);
+
+            /* Reload each of the 4 the bitstreams */
+            HUF_4X_FOR_EACH_STREAM(HUF_4X1_RELOAD_STREAM);
+        } while (op[3] < olimit);
+
+#undef HUF_4X1_DECODE_SYMBOL
+#undef HUF_4X1_RELOAD_STREAM
     }
 
-    assert(args.ip[0] >= args.ilimit);
-    HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop(&args);
+_out:
 
-    /* Our loop guarantees that ip[] >= ilimit and that we haven't
+    /* Save the final values of each of the state variables back to args. */
+    ZSTD_memcpy(&args->bits, &bits, sizeof(bits));
+    ZSTD_memcpy((void*)(&args->ip), &ip, sizeof(ip));
+    ZSTD_memcpy(&args->op, &op, sizeof(op));
+}
+
+/*
+ * @returns @p dstSize on success (>= 6)
+ *          0 if the fallback implementation should be used
+ *          An error if an error occurred
+ */
+static HUF_FAST_BMI2_ATTRS
+size_t
+HUF_decompress4X1_usingDTable_internal_fast(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable,
+    HUF_DecompressFastLoopFn loopFn)
+{
+    void const* dt = DTable + 1;
+    BYTE const* const ilowest = (BYTE const*)cSrc;
+    BYTE* const oend = ZSTD_maybeNullPtrAdd((BYTE*)dst, dstSize);
+    HUF_DecompressFastArgs args;
+    {   size_t const ret = HUF_DecompressFastArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
+        FORWARD_IF_ERROR(ret, "Failed to init fast loop args");
+        if (ret == 0)
+            return 0;
+    }
+
+    assert(args.ip[0] >= args.ilowest);
+    loopFn(&args);
+
+    /* Our loop guarantees that ip[] >= ilowest and that we haven't
     * overwritten any op[].
     */
-    assert(args.ip[0] >= iend);
-    assert(args.ip[1] >= iend);
-    assert(args.ip[2] >= iend);
-    assert(args.ip[3] >= iend);
+    assert(args.ip[0] >= ilowest);
+    assert(args.ip[0] >= ilowest);
+    assert(args.ip[1] >= ilowest);
+    assert(args.ip[2] >= ilowest);
+    assert(args.ip[3] >= ilowest);
     assert(args.op[3] <= oend);
-    (void)iend;
+
+    assert(ilowest == args.ilowest);
+    assert(ilowest + 6 == args.iend[0]);
+    (void)ilowest;
 
     /* finish bit streams one by one. */
-    {
-        size_t const segmentSize = (dstSize+3) / 4;
+    {   size_t const segmentSize = (dstSize+3) / 4;
         BYTE* segmentEnd = (BYTE*)dst;
         int i;
         for (i = 0; i < 4; ++i) {
···
     }
 
     /* decoded size */
+    assert(dstSize != 0);
     return dstSize;
 }
-#endif /* ZSTD_ENABLE_ASM_X86_64_BMI2 */
-
-typedef size_t (*HUF_decompress_usingDTable_t)(void *dst, size_t dstSize,
-                                               const void *cSrc,
-                                               size_t cSrcSize,
-                                               const HUF_DTable *DTable);
 
 HUF_DGEN(HUF_decompress1X1_usingDTable_internal)
 
 static size_t HUF_decompress4X1_usingDTable_internal(void* dst, size_t dstSize, void const* cSrc,
-                    size_t cSrcSize, HUF_DTable const* DTable, int bmi2)
+                    size_t cSrcSize, HUF_DTable const* DTable, int flags)
 {
+    HUF_DecompressUsingDTableFn fallbackFn = HUF_decompress4X1_usingDTable_internal_default;
+    HUF_DecompressFastLoopFn loopFn = HUF_decompress4X1_usingDTable_internal_fast_c_loop;
+
 #if DYNAMIC_BMI2
-    if (bmi2) {
+    if (flags & HUF_flags_bmi2) {
+        fallbackFn = HUF_decompress4X1_usingDTable_internal_bmi2;
 # if ZSTD_ENABLE_ASM_X86_64_BMI2
-        return HUF_decompress4X1_usingDTable_internal_bmi2_asm(dst, dstSize, cSrc, cSrcSize, DTable);
-# else
-        return HUF_decompress4X1_usingDTable_internal_bmi2(dst, dstSize, cSrc, cSrcSize, DTable);
+        if (!(flags & HUF_flags_disableAsm)) {
+            loopFn = HUF_decompress4X1_usingDTable_internal_fast_asm_loop;
+        }
 # endif
+    } else {
+        return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
     }
-#else
-    (void)bmi2;
 #endif
 
 #if ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__)
-    return HUF_decompress4X1_usingDTable_internal_bmi2_asm(dst, dstSize, cSrc, cSrcSize, DTable);
-#else
-    return HUF_decompress4X1_usingDTable_internal_default(dst, dstSize, cSrc, cSrcSize, DTable);
+    if (!(flags & HUF_flags_disableAsm)) {
+        loopFn = HUF_decompress4X1_usingDTable_internal_fast_asm_loop;
+    }
 #endif
+
+    if (HUF_ENABLE_FAST_DECODE && !(flags & HUF_flags_disableFast)) {
+        size_t const ret = HUF_decompress4X1_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
+        if (ret != 0)
+            return ret;
+    }
+    return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
 }
 
-
-size_t HUF_decompress1X1_usingDTable(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
-{
-    DTableDesc dtd = HUF_getDTableDesc(DTable);
-    if (dtd.tableType != 0) return ERROR(GENERIC);
-    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-}
-
-size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
+static size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                    const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize)
+                                   void* workSpace, size_t wkspSize, int flags)
 {
     const BYTE* ip = (const BYTE*) cSrc;
 
-    size_t const hSize = HUF_readDTableX1_wksp(DCtx, cSrc, cSrcSize, workSpace, wkspSize);
+    size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize, flags);
     if (HUF_isError(hSize)) return hSize;
     if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
     ip += hSize; cSrcSize -= hSize;
 
-    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, /* bmi2 */ 0);
+    return HUF_decompress4X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
 }
-
-
-size_t HUF_decompress4X1_usingDTable(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
-{
-    DTableDesc dtd = HUF_getDTableDesc(DTable);
-    if (dtd.tableType != 0) return ERROR(GENERIC);
-    return HUF_decompress4X1_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-}
-
-static size_t HUF_decompress4X1_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize,
-                                   const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize, int bmi2)
-{
-    const BYTE* ip = (const BYTE*) cSrc;
-
-    size_t const hSize = HUF_readDTableX1_wksp_bmi2(dctx, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
-    if (HUF_isError(hSize)) return hSize;
-    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
-    ip += hSize; cSrcSize -= hSize;
-
-    return HUF_decompress4X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
-}
-
-size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
-                                   const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize)
-{
-    return HUF_decompress4X1_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, 0);
-}
-
 
 #endif /* HUF_FORCE_DECOMPRESS_X2 */
 
···
 
 static void HUF_fillDTableX2(HUF_DEltX2* DTable, const U32 targetLog,
                            const sortedSymbol_t* sortedList,
-                           const U32* rankStart, rankValCol_t *rankValOrigin, const U32 maxWeight,
+                           const U32* rankStart, rankValCol_t* rankValOrigin, const U32 maxWeight,
                            const U32 nbBitsBaseline)
 {
     U32* const rankVal = rankValOrigin[0];
···
 
 size_t HUF_readDTableX2_wksp(HUF_DTable* DTable,
                        const void* src, size_t srcSize,
-                             void* workSpace, size_t wkspSize)
-{
-    return HUF_readDTableX2_wksp_bmi2(DTable, src, srcSize, workSpace, wkspSize, /* bmi2 */ 0);
-}
-
-size_t HUF_readDTableX2_wksp_bmi2(HUF_DTable* DTable,
-                       const void* src, size_t srcSize,
-                             void* workSpace, size_t wkspSize, int bmi2)
+                             void* workSpace, size_t wkspSize, int flags)
 {
     U32 tableLog, maxW, nbSymbols;
     DTableDesc dtd = HUF_getDTableDesc(DTable);
···
     if (maxTableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
     /* ZSTD_memset(weightList, 0, sizeof(weightList)); */  /* is not necessary, even though some analyzer complain ... */
 
-    iSize = HUF_readStats_wksp(wksp->weightList, HUF_SYMBOLVALUE_MAX + 1, wksp->rankStats, &nbSymbols, &tableLog, src, srcSize, wksp->calleeWksp, sizeof(wksp->calleeWksp), bmi2);
+    iSize = HUF_readStats_wksp(wksp->weightList, HUF_SYMBOLVALUE_MAX + 1, wksp->rankStats, &nbSymbols, &tableLog, src, srcSize, wksp->calleeWksp, sizeof(wksp->calleeWksp), flags);
     if (HUF_isError(iSize)) return iSize;
 
     /* check result */
···
 }
 
 #define HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
-    ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog)
+    do { ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog); } while (0)
 
-#define HUF_DECODE_SYMBOLX2_1(ptr, DStreamPtr) \
-    if (MEM_64bits() || (HUF_TABLELOG_MAX<=12)) \
-        ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog)
+#define HUF_DECODE_SYMBOLX2_1(ptr, DStreamPtr)                     \
+    do {                                                           \
+        if (MEM_64bits() || (HUF_TABLELOG_MAX<=12))                \
+            ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog); \
+    } while (0)
 
-#define HUF_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
-    if (MEM_64bits()) \
-        ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog)
+#define HUF_DECODE_SYMBOLX2_2(ptr, DStreamPtr)                     \
+    do {                                                           \
+        if (MEM_64bits())                                          \
+            ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog); \
+    } while (0)
 
 HINT_INLINE size_t
 HUF_decodeStreamX2(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd,
···
 
     /* decode */
     {   BYTE* const ostart = (BYTE*) dst;
-        BYTE* const oend = ostart + dstSize;
+        BYTE* const oend = ZSTD_maybeNullPtrAdd(ostart, dstSize);
         const void* const dtPtr = DTable+1;   /* force compiler to not use strict-aliasing */
         const HUF_DEltX2* const dt = (const HUF_DEltX2*)dtPtr;
         DTableDesc const dtd = HUF_getDTableDesc(DTable);
···
     /* decoded size */
     return dstSize;
 }
+
+/* HUF_decompress4X2_usingDTable_internal_body():
+ * Conditions:
+ * @dstSize >= 6
+ */
 FORCE_INLINE_TEMPLATE size_t
 HUF_decompress4X2_usingDTable_internal_body(
           void* dst,  size_t dstSize,
···
     const HUF_DTable* DTable)
 {
     if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
+    if (dstSize < 6) return ERROR(corruption_detected);         /* stream 4-split doesn't work */
 
     {   const BYTE* const istart = (const BYTE*) cSrc;
         BYTE* const ostart = (BYTE*) dst;
···
         DTableDesc const dtd = HUF_getDTableDesc(DTable);
         U32 const dtLog = dtd.tableLog;
 
-        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
-        if (opStart4 > oend) return ERROR(corruption_detected);      /* overflow */
+        if (length4 > cSrcSize) return ERROR(corruption_detected);  /* overflow */
+        if (opStart4 > oend) return ERROR(corruption_detected);     /* overflow */
+        assert(dstSize >= 6 /* validated above */);
         CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
         CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
         CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
···
 }
 #endif
 
-#if HUF_NEED_DEFAULT_FUNCTION
 static
 size_t HUF_decompress4X2_usingDTable_internal_default(void* dst, size_t dstSize, void const* cSrc,
                     size_t cSrcSize, HUF_DTable const* DTable) {
     return HUF_decompress4X2_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
 }
-#endif
 
 #if ZSTD_ENABLE_ASM_X86_64_BMI2
 
-HUF_ASM_DECL void HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop(HUF_DecompressAsmArgs* args) ZSTDLIB_HIDDEN;
+HUF_ASM_DECL void HUF_decompress4X2_usingDTable_internal_fast_asm_loop(HUF_DecompressFastArgs* args) ZSTDLIB_HIDDEN;
 
-static HUF_ASM_X86_64_BMI2_ATTRS size_t
-HUF_decompress4X2_usingDTable_internal_bmi2_asm(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable) {
-    void const* dt = DTable + 1;
-    const BYTE* const iend = (const BYTE*)cSrc + 6;
-    BYTE* const oend = (BYTE*)dst + dstSize;
-    HUF_DecompressAsmArgs args;
-    {
-        size_t const ret = HUF_DecompressAsmArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
-        FORWARD_IF_ERROR(ret, "Failed to init asm args");
-        if (ret != 0)
-            return HUF_decompress4X2_usingDTable_internal_bmi2(dst, dstSize, cSrc, cSrcSize, DTable);
+#endif
+
+static HUF_FAST_BMI2_ATTRS
+void HUF_decompress4X2_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs* args)
+{
+    U64 bits[4];
+    BYTE const* ip[4];
+    BYTE* op[4];
+    BYTE* oend[4];
+    HUF_DEltX2 const* const dtable = (HUF_DEltX2 const*)args->dt;
+    BYTE const* const ilowest = args->ilowest;
+
+    /* Copy the arguments to local registers. */
+    ZSTD_memcpy(&bits, &args->bits, sizeof(bits));
+    ZSTD_memcpy((void*)(&ip), &args->ip, sizeof(ip));
+    ZSTD_memcpy(&op, &args->op, sizeof(op));
+
+    oend[0] = op[1];
+    oend[1] = op[2];
+    oend[2] = op[3];
+    oend[3] = args->oend;
+
+    assert(MEM_isLittleEndian());
+    assert(!MEM_32bits());
+
+    for (;;) {
+        BYTE* olimit;
+        int stream;
+
+        /* Assert loop preconditions */
+#ifndef NDEBUG
+        for (stream = 0; stream < 4; ++stream) {
+            assert(op[stream] <= oend[stream]);
+            assert(ip[stream] >= ilowest);
+        }
+#endif
+        /* Compute olimit */
+        {
+            /* Each loop does 5 table lookups for each of the 4 streams.
+             * Each table lookup consumes up to 11 bits of input, and produces
+             * up to 2 bytes of output.
+             */
+            /* We can consume up to 7 bytes of input per iteration per stream.
+             * We also know that each input pointer is >= ip[0]. So we can run
+             * iters loops before running out of input.
+             */
+            size_t iters = (size_t)(ip[0] - ilowest) / 7;
+            /* Each iteration can produce up to 10 bytes of output per stream.
+             * Each output stream my advance at different rates. So take the
+             * minimum number of safe iterations among all the output streams.
+             */
+            for (stream = 0; stream < 4; ++stream) {
+                size_t const oiters = (size_t)(oend[stream] - op[stream]) / 10;
+                iters = MIN(iters, oiters);
+            }
+
+            /* Each iteration produces at least 5 output symbols. So until
+             * op[3] crosses olimit, we know we haven't executed iters
+             * iterations yet. This saves us maintaining an iters counter,
+             * at the expense of computing the remaining # of iterations
+             * more frequently.
+             */
+            olimit = op[3] + (iters * 5);
+
+            /* Exit the fast decoding loop once we reach the end. */
+            if (op[3] == olimit)
+                break;
+
+            /* Exit the decoding loop if any input pointer has crossed the
+             * previous one. This indicates corruption, and a precondition
+             * to our loop is that ip[i] >= ip[0].
+             */
+            for (stream = 1; stream < 4; ++stream) {
+                if (ip[stream] < ip[stream - 1])
+                    goto _out;
+            }
+        }
+
+#ifndef NDEBUG
+        for (stream = 1; stream < 4; ++stream) {
+            assert(ip[stream] >= ip[stream - 1]);
+        }
+#endif
+
+#define HUF_4X2_DECODE_SYMBOL(_stream, _decode3)                      \
+    do {                                                              \
+        if ((_decode3) || (_stream) != 3) {                           \
+            int const index = (int)(bits[(_stream)] >> 53);           \
+            HUF_DEltX2 const entry = dtable[index];                   \
+            MEM_write16(op[(_stream)], entry.sequence); \
+            bits[(_stream)] <<= (entry.nbBits) & 0x3F;                \
+            op[(_stream)] += (entry.length);                          \
+        }                                                             \
+    } while (0)
+
+#define HUF_4X2_RELOAD_STREAM(_stream)                                  \
+    do {                                                                \
+        HUF_4X2_DECODE_SYMBOL(3, 1);                                    \
+        {                                                               \
+            int const ctz = ZSTD_countTrailingZeros64(bits[(_stream)]); \
+            int const nbBits = ctz & 7;                                 \
+            int const nbBytes = ctz >> 3;                               \
+            ip[(_stream)] -= nbBytes;                                   \
+            bits[(_stream)] = MEM_read64(ip[(_stream)]) | 1;            \
+            bits[(_stream)] <<= nbBits;                                 \
+        }                                                               \
+    } while (0)
+
+        /* Manually unroll the loop because compilers don't consistently
+         * unroll the inner loops, which destroys performance.
+         */
+        do {
+            /* Decode 5 symbols from each of the first 3 streams.
+             * The final stream will be decoded during the reload phase
+             * to reduce register pressure.
+             */
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
+            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
+
+            /* Decode one symbol from the final stream */
+            HUF_4X2_DECODE_SYMBOL(3, 1);
+
+            /* Decode 4 symbols from the final stream & reload bitstreams.
+             * The final stream is reloaded last, meaning that all 5 symbols
+             * are decoded from the final stream before it is reloaded.
+             */
+            HUF_4X_FOR_EACH_STREAM(HUF_4X2_RELOAD_STREAM);
+        } while (op[3] < olimit);
     }
 
-    assert(args.ip[0] >= args.ilimit);
-    HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop(&args);
+#undef HUF_4X2_DECODE_SYMBOL
+#undef HUF_4X2_RELOAD_STREAM
+
+_out:
+
+    /* Save the final values of each of the state variables back to args. */
+    ZSTD_memcpy(&args->bits, &bits, sizeof(bits));
+    ZSTD_memcpy((void*)(&args->ip), &ip, sizeof(ip));
+    ZSTD_memcpy(&args->op, &op, sizeof(op));
+}
+
+
+static HUF_FAST_BMI2_ATTRS size_t
+HUF_decompress4X2_usingDTable_internal_fast(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable,
+    HUF_DecompressFastLoopFn loopFn) {
+    void const* dt = DTable + 1;
+    const BYTE* const ilowest = (const BYTE*)cSrc;
+    BYTE* const oend = ZSTD_maybeNullPtrAdd((BYTE*)dst, dstSize);
+    HUF_DecompressFastArgs args;
+    {
+        size_t const ret = HUF_DecompressFastArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
+        FORWARD_IF_ERROR(ret, "Failed to init asm args");
+        if (ret == 0)
+            return 0;
+    }
+
+    assert(args.ip[0] >= args.ilowest);
+    loopFn(&args);
 
     /* note : op4 already verified within main loop */
-    assert(args.ip[0] >= iend);
-    assert(args.ip[1] >= iend);
-    assert(args.ip[2] >= iend);
-    assert(args.ip[3] >= iend);
+    assert(args.ip[0] >= ilowest);
+    assert(args.ip[1] >= ilowest);
+    assert(args.ip[2] >= ilowest);
+    assert(args.ip[3] >= ilowest);
     assert(args.op[3] <= oend);
-    (void)iend;
+
+    assert(ilowest == args.ilowest);
+    assert(ilowest + 6 == args.iend[0]);
+    (void)ilowest;
 
     /* finish bitStreams one by one */
     {
···
     /* decoded size */
     return dstSize;
 }
-#endif /* ZSTD_ENABLE_ASM_X86_64_BMI2 */
 
 static size_t HUF_decompress4X2_usingDTable_internal(void* dst, size_t dstSize, void const* cSrc,
-                    size_t cSrcSize, HUF_DTable const* DTable, int bmi2)
+                    size_t cSrcSize, HUF_DTable const* DTable, int flags)
 {
+    HUF_DecompressUsingDTableFn fallbackFn = HUF_decompress4X2_usingDTable_internal_default;
+    HUF_DecompressFastLoopFn loopFn = HUF_decompress4X2_usingDTable_internal_fast_c_loop;
+
 #if DYNAMIC_BMI2
-    if (bmi2) {
+    if (flags & HUF_flags_bmi2) {
+        fallbackFn = HUF_decompress4X2_usingDTable_internal_bmi2;
 # if ZSTD_ENABLE_ASM_X86_64_BMI2
-        return HUF_decompress4X2_usingDTable_internal_bmi2_asm(dst, dstSize, cSrc, cSrcSize, DTable);
-# else
-        return HUF_decompress4X2_usingDTable_internal_bmi2(dst, dstSize, cSrc, cSrcSize, DTable);
+        if (!(flags & HUF_flags_disableAsm)) {
+            loopFn = HUF_decompress4X2_usingDTable_internal_fast_asm_loop;
+        }
 # endif
+    } else {
+        return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
     }
-#else
-    (void)bmi2;
 #endif
 
 #if ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__)
-    return HUF_decompress4X2_usingDTable_internal_bmi2_asm(dst, dstSize, cSrc, cSrcSize, DTable);
-#else
-    return HUF_decompress4X2_usingDTable_internal_default(dst, dstSize, cSrc, cSrcSize, DTable);
+    if (!(flags & HUF_flags_disableAsm)) {
+        loopFn = HUF_decompress4X2_usingDTable_internal_fast_asm_loop;
+    }
 #endif
+
+    if (HUF_ENABLE_FAST_DECODE && !(flags & HUF_flags_disableFast)) {
+        size_t const ret = HUF_decompress4X2_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
+        if (ret != 0)
+            return ret;
+    }
+    return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
 }
 
 HUF_DGEN(HUF_decompress1X2_usingDTable_internal)
 
-size_t HUF_decompress1X2_usingDTable(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
-{
-    DTableDesc dtd = HUF_getDTableDesc(DTable);
-    if (dtd.tableType != 1) return ERROR(GENERIC);
-    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-}
-
 size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
                                    const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize)
+                                   void* workSpace, size_t wkspSize, int flags)
 {
     const BYTE* ip = (const BYTE*) cSrc;
 
     size_t const hSize = HUF_readDTableX2_wksp(DCtx, cSrc, cSrcSize,
-                                               workSpace, wkspSize);
+                                               workSpace, wkspSize, flags);
     if (HUF_isError(hSize)) return hSize;
     if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
     ip += hSize; cSrcSize -= hSize;
 
-    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, /* bmi2 */ 0);
+    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, flags);
 }
 
-
-size_t HUF_decompress4X2_usingDTable(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
-{
-    DTableDesc dtd = HUF_getDTableDesc(DTable);
-    if (dtd.tableType != 1) return ERROR(GENERIC);
-    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-}
-
-static size_t HUF_decompress4X2_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize,
+static size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                    const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize, int bmi2)
+                                   void* workSpace, size_t wkspSize, int flags)
 {
     const BYTE* ip = (const BYTE*) cSrc;
 
     size_t hSize = HUF_readDTableX2_wksp(dctx, cSrc, cSrcSize,
-                                         workSpace, wkspSize);
+                                         workSpace, wkspSize, flags);
     if (HUF_isError(hSize)) return hSize;
     if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
     ip += hSize; cSrcSize -= hSize;
 
-    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
+    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
 }
-
-size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
-                                   const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize)
-{
-    return HUF_decompress4X2_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, /* bmi2 */ 0);
-}
-
 
 #endif /* HUF_FORCE_DECOMPRESS_X1 */
 
···
 /* ***********************************/
 /* Universal decompression selectors */
 /* ***********************************/
-
-size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize,
-                                    const void* cSrc, size_t cSrcSize,
-                                    const HUF_DTable* DTable)
-{
-    DTableDesc const dtd = HUF_getDTableDesc(DTable);
-#if defined(HUF_FORCE_DECOMPRESS_X1)
-    (void)dtd;
-    assert(dtd.tableType == 0);
-    return HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#elif defined(HUF_FORCE_DECOMPRESS_X2)
-    (void)dtd;
-    assert(dtd.tableType == 1);
-    return HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#else
-    return dtd.tableType ? HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0) :
-                           HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#endif
-}
-
-size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize,
-                                    const void* cSrc, size_t cSrcSize,
-                                    const HUF_DTable* DTable)
-{
-    DTableDesc const dtd = HUF_getDTableDesc(DTable);
-#if defined(HUF_FORCE_DECOMPRESS_X1)
-    (void)dtd;
-    assert(dtd.tableType == 0);
-    return HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#elif defined(HUF_FORCE_DECOMPRESS_X2)
-    (void)dtd;
-    assert(dtd.tableType == 1);
-    return HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#else
-    return dtd.tableType ? HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0) :
-                           HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#endif
-}
 
 
 #if !defined(HUF_FORCE_DECOMPRESS_X1) && !defined(HUF_FORCE_DECOMPRESS_X2)
···
 #endif
 }
 
-
-size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst,
-                                     size_t dstSize, const void* cSrc,
-                                     size_t cSrcSize, void* workSpace,
-                                     size_t wkspSize)
-{
-    /* validation checks */
-    if (dstSize == 0) return ERROR(dstSize_tooSmall);
-    if (cSrcSize == 0) return ERROR(corruption_detected);
-
-    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
-#if defined(HUF_FORCE_DECOMPRESS_X1)
-        (void)algoNb;
-        assert(algoNb == 0);
-        return HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize);
-#elif defined(HUF_FORCE_DECOMPRESS_X2)
-        (void)algoNb;
-        assert(algoNb == 1);
-        return HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize);
-#else
-        return algoNb ? HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                            cSrcSize, workSpace, wkspSize):
-                        HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize);
-#endif
-    }
-}
-
 size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                   const void* cSrc, size_t cSrcSize,
-                                  void* workSpace, size_t wkspSize)
+                                  void* workSpace, size_t wkspSize, int flags)
 {
     /* validation checks */
     if (dstSize == 0) return ERROR(dstSize_tooSmall);
···
         (void)algoNb;
         assert(algoNb == 0);
         return HUF_decompress1X1_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                                cSrcSize, workSpace, wkspSize);
+                                cSrcSize, workSpace, wkspSize, flags);
 #elif defined(HUF_FORCE_DECOMPRESS_X2)
         (void)algoNb;
         assert(algoNb == 1);
         return HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                                cSrcSize, workSpace, wkspSize);
+                                cSrcSize, workSpace, wkspSize, flags);
 #else
         return algoNb ? HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                                cSrcSize, workSpace, wkspSize):
+                                cSrcSize, workSpace, wkspSize, flags):
                         HUF_decompress1X1_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                                cSrcSize, workSpace, wkspSize);
+                                cSrcSize, workSpace, wkspSize, flags);
 #endif
     }
 }
 
 
-size_t HUF_decompress1X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2)
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags)
 {
     DTableDesc const dtd = HUF_getDTableDesc(DTable);
 #if defined(HUF_FORCE_DECOMPRESS_X1)
     (void)dtd;
     assert(dtd.tableType == 0);
-    return HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #elif defined(HUF_FORCE_DECOMPRESS_X2)
     (void)dtd;
     assert(dtd.tableType == 1);
-    return HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #else
-    return dtd.tableType ? HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2) :
-                           HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return dtd.tableType ? HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags) :
+                           HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #endif
 }
 
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2)
+size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags)
 {
     const BYTE* ip = (const BYTE*) cSrc;
 
-    size_t const hSize = HUF_readDTableX1_wksp_bmi2(dctx, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
+    size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize, flags);
     if (HUF_isError(hSize)) return hSize;
     if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
     ip += hSize; cSrcSize -= hSize;
 
-    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
+    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
 }
 #endif
 
-size_t HUF_decompress4X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2)
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags)
 {
     DTableDesc const dtd = HUF_getDTableDesc(DTable);
 #if defined(HUF_FORCE_DECOMPRESS_X1)
     (void)dtd;
     assert(dtd.tableType == 0);
-    return HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #elif defined(HUF_FORCE_DECOMPRESS_X2)
     (void)dtd;
     assert(dtd.tableType == 1);
-    return HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #else
-    return dtd.tableType ? HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2) :
-                           HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return dtd.tableType ? HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags) :
+                           HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #endif
 }
 
-size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2)
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags)
 {
     /* validation checks */
     if (dstSize == 0) return ERROR(dstSize_tooSmall);
···
 #if defined(HUF_FORCE_DECOMPRESS_X1)
         (void)algoNb;
         assert(algoNb == 0);
-        return HUF_decompress4X1_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
+        return HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
 #elif defined(HUF_FORCE_DECOMPRESS_X2)
         (void)algoNb;
         assert(algoNb == 1);
-        return HUF_decompress4X2_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
+        return HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
 #else
-        return algoNb ? HUF_decompress4X2_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2) :
-                        HUF_decompress4X1_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
+        return algoNb ? HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags) :
+                        HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
 #endif
     }
 }
-

lib/zstd/decompress/zstd_ddict.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 /*-*******************************************************
 *  Dependencies
 *********************************************************/
+#include "../common/allocations.h"  /* ZSTD_customMalloc, ZSTD_customFree */
 #include "../common/zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
 #include "../common/cpu.h"         /* bmi2 */
 #include "../common/mem.h"         /* low level memory routines */
 #define FSE_STATIC_LINKING_ONLY
 #include "../common/fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"
 #include "zstd_decompress_internal.h"
 #include "zstd_ddict.h"
···
         ZSTD_memcpy(internalBuffer, dict, dictSize);
     }
     ddict->dictSize = dictSize;
-    ddict->entropy.hufTable[0] = (HUF_DTable)((HufLog)*0x1000001);  /* cover both little and big endian */
+    ddict->entropy.hufTable[0] = (HUF_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);  /* cover both little and big endian */
 
     /* parse dictionary content */
     FORWARD_IF_ERROR( ZSTD_loadEntropy_intoDDict(ddict, dictContentType) , "");
···
 unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict)
 {
     if (ddict==NULL) return 0;
-    return ZSTD_getDictID_fromDict(ddict->dictContent, ddict->dictSize);
+    return ddict->dictID;
 }

lib/zstd/decompress/zstd_ddict.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

lib/zstd/decompress/zstd_decompress.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 *  Dependencies
 *********************************************************/
 #include "../common/zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
+#include "../common/allocations.h"  /* ZSTD_customMalloc, ZSTD_customCalloc, ZSTD_customFree */
+#include "../common/error_private.h"
+#include "../common/zstd_internal.h"  /* blockProperties_t */
 #include "../common/mem.h"         /* low level memory routines */
+#include "../common/bits.h"  /* ZSTD_highbit32 */
 #define FSE_STATIC_LINKING_ONLY
 #include "../common/fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"
 #include <linux/xxhash.h> /* xxh64_reset, xxh64_update, xxh64_digest, XXH64 */
-#include "../common/zstd_internal.h"  /* blockProperties_t */
 #include "zstd_decompress_internal.h"   /* ZSTD_DCtx */
 #include "zstd_ddict.h"  /* ZSTD_DDictDictContent */
 #include "zstd_decompress_block.h"   /* ZSTD_decompressBlock_internal */
···
  *************************************/
 
 #define DDICT_HASHSET_MAX_LOAD_FACTOR_COUNT_MULT 4
-#define DDICT_HASHSET_MAX_LOAD_FACTOR_SIZE_MULT 3   /* These two constants represent SIZE_MULT/COUNT_MULT load factor without using a float.
-                                                     * Currently, that means a 0.75 load factor.
-                                                     * So, if count * COUNT_MULT / size * SIZE_MULT != 0, then we've exceeded
-                                                     * the load factor of the ddict hash set.
-                                                     */
+#define DDICT_HASHSET_MAX_LOAD_FACTOR_SIZE_MULT 3  /* These two constants represent SIZE_MULT/COUNT_MULT load factor without using a float.
+                                                    * Currently, that means a 0.75 load factor.
+                                                    * So, if count * COUNT_MULT / size * SIZE_MULT != 0, then we've exceeded
+                                                    * the load factor of the ddict hash set.
+                                                    */
 
 #define DDICT_HASHSET_TABLE_BASE_SIZE 64
 #define DDICT_HASHSET_RESIZE_FACTOR 2
···
     dctx->outBufferMode = ZSTD_bm_buffered;
     dctx->forceIgnoreChecksum = ZSTD_d_validateChecksum;
     dctx->refMultipleDDicts = ZSTD_rmd_refSingleDDict;
+    dctx->disableHufAsm = 0;
+    dctx->maxBlockSizeParam = 0;
 }
 
 static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
···
     dctx->streamStage = zdss_init;
     dctx->noForwardProgress = 0;
     dctx->oversizedDuration = 0;
+    dctx->isFrameDecompression = 1;
 #if DYNAMIC_BMI2
     dctx->bmi2 = ZSTD_cpuSupportsBmi2();
 #endif
···
  *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless
  * @return : 0, `zfhPtr` is correctly filled,
  *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
- *           or an error code, which can be tested using ZSTD_isError() */
-size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format)
+**           or an error code, which can be tested using ZSTD_isError() */
+size_t ZSTD_getFrameHeader_advanced(ZSTD_FrameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format)
 {
     const BYTE* ip = (const BYTE*)src;
     size_t const minInputSize = ZSTD_startingInputLength(format);
 
-    ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));   /* not strictly necessary, but static analyzer do not understand that zfhPtr is only going to be read only if return value is zero, since they are 2 different signals */
-    if (srcSize < minInputSize) return minInputSize;
-    RETURN_ERROR_IF(src==NULL, GENERIC, "invalid parameter");
+    DEBUGLOG(5, "ZSTD_getFrameHeader_advanced: minInputSize = %zu, srcSize = %zu", minInputSize, srcSize);
 
+    if (srcSize > 0) {
+        /* note : technically could be considered an assert(), since it's an invalid entry */
+        RETURN_ERROR_IF(src==NULL, GENERIC, "invalid parameter : src==NULL, but srcSize>0");
+    }
+    if (srcSize < minInputSize) {
+        if (srcSize > 0 && format != ZSTD_f_zstd1_magicless) {
+            /* when receiving less than @minInputSize bytes,
+             * control these bytes at least correspond to a supported magic number
+             * in order to error out early if they don't.
+            **/
+            size_t const toCopy = MIN(4, srcSize);
+            unsigned char hbuf[4]; MEM_writeLE32(hbuf, ZSTD_MAGICNUMBER);
+            assert(src != NULL);
+            ZSTD_memcpy(hbuf, src, toCopy);
+            if ( MEM_readLE32(hbuf) != ZSTD_MAGICNUMBER ) {
+                /* not a zstd frame : let's check if it's a skippable frame */
+                MEM_writeLE32(hbuf, ZSTD_MAGIC_SKIPPABLE_START);
+                ZSTD_memcpy(hbuf, src, toCopy);
+                if ((MEM_readLE32(hbuf) & ZSTD_MAGIC_SKIPPABLE_MASK) != ZSTD_MAGIC_SKIPPABLE_START) {
+                    RETURN_ERROR(prefix_unknown,
+                                "first bytes don't correspond to any supported magic number");
+        }   }   }
+        return minInputSize;
+    }
+
+    ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));   /* not strictly necessary, but static analyzers may not understand that zfhPtr will be read only if return value is zero, since they are 2 different signals */
     if ( (format != ZSTD_f_zstd1_magicless)
       && (MEM_readLE32(src) != ZSTD_MAGICNUMBER) ) {
         if ((MEM_readLE32(src) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
···
             if (srcSize < ZSTD_SKIPPABLEHEADERSIZE)
                 return ZSTD_SKIPPABLEHEADERSIZE; /* magic number + frame length */
             ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));
-            zfhPtr->frameContentSize = MEM_readLE32((const char *)src + ZSTD_FRAMEIDSIZE);
             zfhPtr->frameType = ZSTD_skippableFrame;
+            zfhPtr->dictID = MEM_readLE32(src) - ZSTD_MAGIC_SKIPPABLE_START;
+            zfhPtr->headerSize = ZSTD_SKIPPABLEHEADERSIZE;
+            zfhPtr->frameContentSize = MEM_readLE32((const char *)src + ZSTD_FRAMEIDSIZE);
             return 0;
         }
         RETURN_ERROR(prefix_unknown, "");
···
  * @return : 0, `zfhPtr` is correctly filled,
  *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
  *           or an error code, which can be tested using ZSTD_isError() */
-size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize)
+size_t ZSTD_getFrameHeader(ZSTD_FrameHeader* zfhPtr, const void* src, size_t srcSize)
 {
     return ZSTD_getFrameHeader_advanced(zfhPtr, src, srcSize, ZSTD_f_zstd1);
 }
···
  *         - ZSTD_CONTENTSIZE_ERROR if an error occurred (e.g. invalid magic number, srcSize too small) */
 unsigned long long ZSTD_getFrameContentSize(const void *src, size_t srcSize)
 {
-    {   ZSTD_frameHeader zfh;
+    {   ZSTD_FrameHeader zfh;
         if (ZSTD_getFrameHeader(&zfh, src, srcSize) != 0)
             return ZSTD_CONTENTSIZE_ERROR;
         if (zfh.frameType == ZSTD_skippableFrame) {
···
     sizeU32 = MEM_readLE32((BYTE const*)src + ZSTD_FRAMEIDSIZE);
     RETURN_ERROR_IF((U32)(sizeU32 + ZSTD_SKIPPABLEHEADERSIZE) < sizeU32,
                     frameParameter_unsupported, "");
-    {
-        size_t const skippableSize = skippableHeaderSize + sizeU32;
+    {   size_t const skippableSize = skippableHeaderSize + sizeU32;
         RETURN_ERROR_IF(skippableSize > srcSize, srcSize_wrong, "");
         return skippableSize;
     }
 }
 
 /*! ZSTD_readSkippableFrame() :
- * Retrieves a zstd skippable frame containing data given by src, and writes it to dst buffer.
+ * Retrieves content of a skippable frame, and writes it to dst buffer.
  *
  * The parameter magicVariant will receive the magicVariant that was supplied when the frame was written,
  * i.e. magicNumber - ZSTD_MAGIC_SKIPPABLE_START.  This can be NULL if the caller is not interested
  * in the magicVariant.
  *
- * Returns an error if destination buffer is not large enough, or if the frame is not skippable.
+ * Returns an error if destination buffer is not large enough, or if this is not a valid skippable frame.
  *
  * @return : number of bytes written or a ZSTD error.
  */
-ZSTDLIB_API size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity, unsigned* magicVariant,
-                                            const void* src, size_t srcSize)
+size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity,
+                               unsigned* magicVariant,  /* optional, can be NULL */
+                         const void* src, size_t srcSize)
 {
-    U32 const magicNumber = MEM_readLE32(src);
-    size_t skippableFrameSize = readSkippableFrameSize(src, srcSize);
-    size_t skippableContentSize = skippableFrameSize - ZSTD_SKIPPABLEHEADERSIZE;
+    RETURN_ERROR_IF(srcSize < ZSTD_SKIPPABLEHEADERSIZE, srcSize_wrong, "");
 
-    /* check input validity */
-    RETURN_ERROR_IF(!ZSTD_isSkippableFrame(src, srcSize), frameParameter_unsupported, "");
-    RETURN_ERROR_IF(skippableFrameSize < ZSTD_SKIPPABLEHEADERSIZE || skippableFrameSize > srcSize, srcSize_wrong, "");
-    RETURN_ERROR_IF(skippableContentSize > dstCapacity, dstSize_tooSmall, "");
+    {   U32 const magicNumber = MEM_readLE32(src);
+        size_t skippableFrameSize = readSkippableFrameSize(src, srcSize);
+        size_t skippableContentSize = skippableFrameSize - ZSTD_SKIPPABLEHEADERSIZE;
 
-    /* deliver payload */
-    if (skippableContentSize > 0  && dst != NULL)
-        ZSTD_memcpy(dst, (const BYTE *)src + ZSTD_SKIPPABLEHEADERSIZE, skippableContentSize);
-    if (magicVariant != NULL)
-        *magicVariant = magicNumber - ZSTD_MAGIC_SKIPPABLE_START;
-    return skippableContentSize;
+        /* check input validity */
+        RETURN_ERROR_IF(!ZSTD_isSkippableFrame(src, srcSize), frameParameter_unsupported, "");
+        RETURN_ERROR_IF(skippableFrameSize < ZSTD_SKIPPABLEHEADERSIZE || skippableFrameSize > srcSize, srcSize_wrong, "");
+        RETURN_ERROR_IF(skippableContentSize > dstCapacity, dstSize_tooSmall, "");
+
+        /* deliver payload */
+        if (skippableContentSize > 0  && dst != NULL)
+            ZSTD_memcpy(dst, (const BYTE *)src + ZSTD_SKIPPABLEHEADERSIZE, skippableContentSize);
+        if (magicVariant != NULL)
+            *magicVariant = magicNumber - ZSTD_MAGIC_SKIPPABLE_START;
+        return skippableContentSize;
+    }
 }
 
 /* ZSTD_findDecompressedSize() :
- *  compatible with legacy mode
  *  `srcSize` must be the exact length of some number of ZSTD compressed and/or
  *      skippable frames
- *  @return : decompressed size of the frames contained */
+ *  note: compatible with legacy mode
+ * @return : decompressed size of the frames contained */
 unsigned long long ZSTD_findDecompressedSize(const void* src, size_t srcSize)
 {
     unsigned long long totalDstSize = 0;
···
 
         if ((magicNumber & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
             size_t const skippableSize = readSkippableFrameSize(src, srcSize);
-            if (ZSTD_isError(skippableSize)) {
-                return ZSTD_CONTENTSIZE_ERROR;
-            }
+            if (ZSTD_isError(skippableSize)) return ZSTD_CONTENTSIZE_ERROR;
             assert(skippableSize <= srcSize);
 
             src = (const BYTE *)src + skippableSize;
···
             continue;
         }
 
-        {   unsigned long long const ret = ZSTD_getFrameContentSize(src, srcSize);
-            if (ret >= ZSTD_CONTENTSIZE_ERROR) return ret;
+        {   unsigned long long const fcs = ZSTD_getFrameContentSize(src, srcSize);
+            if (fcs >= ZSTD_CONTENTSIZE_ERROR) return fcs;
 
-            /* check for overflow */
-            if (totalDstSize + ret < totalDstSize) return ZSTD_CONTENTSIZE_ERROR;
-            totalDstSize += ret;
+            if (totalDstSize + fcs < totalDstSize)
+                return ZSTD_CONTENTSIZE_ERROR; /* check for overflow */
+            totalDstSize += fcs;
         }
+        /* skip to next frame */
         {   size_t const frameSrcSize = ZSTD_findFrameCompressedSize(src, srcSize);
-            if (ZSTD_isError(frameSrcSize)) {
-                return ZSTD_CONTENTSIZE_ERROR;
-            }
+            if (ZSTD_isError(frameSrcSize)) return ZSTD_CONTENTSIZE_ERROR;
+            assert(frameSrcSize <= srcSize);
 
             src = (const BYTE *)src + frameSrcSize;
             srcSize -= frameSrcSize;
···
     return frameSizeInfo;
 }
 
-static ZSTD_frameSizeInfo ZSTD_findFrameSizeInfo(const void* src, size_t srcSize)
+static ZSTD_frameSizeInfo ZSTD_findFrameSizeInfo(const void* src, size_t srcSize, ZSTD_format_e format)
 {
     ZSTD_frameSizeInfo frameSizeInfo;
     ZSTD_memset(&frameSizeInfo, 0, sizeof(ZSTD_frameSizeInfo));
 
 
-    if ((srcSize >= ZSTD_SKIPPABLEHEADERSIZE)
+    if (format == ZSTD_f_zstd1 && (srcSize >= ZSTD_SKIPPABLEHEADERSIZE)
         && (MEM_readLE32(src) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
         frameSizeInfo.compressedSize = readSkippableFrameSize(src, srcSize);
         assert(ZSTD_isError(frameSizeInfo.compressedSize) ||
···
         const BYTE* const ipstart = ip;
         size_t remainingSize = srcSize;
         size_t nbBlocks = 0;
-        ZSTD_frameHeader zfh;
+        ZSTD_FrameHeader zfh;
 
         /* Extract Frame Header */
-        {   size_t const ret = ZSTD_getFrameHeader(&zfh, src, srcSize);
+        {   size_t const ret = ZSTD_getFrameHeader_advanced(&zfh, src, srcSize, format);
             if (ZSTD_isError(ret))
                 return ZSTD_errorFrameSizeInfo(ret);
             if (ret > 0)
···
             ip += 4;
         }
 
+        frameSizeInfo.nbBlocks = nbBlocks;
         frameSizeInfo.compressedSize = (size_t)(ip - ipstart);
         frameSizeInfo.decompressedBound = (zfh.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN)
                                         ? zfh.frameContentSize
-                                        : nbBlocks * zfh.blockSizeMax;
+                                        : (unsigned long long)nbBlocks * zfh.blockSizeMax;
         return frameSizeInfo;
     }
 }
 
+static size_t ZSTD_findFrameCompressedSize_advanced(const void *src, size_t srcSize, ZSTD_format_e format) {
+    ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize, format);
+    return frameSizeInfo.compressedSize;
+}
+
 /* ZSTD_findFrameCompressedSize() :
- *  compatible with legacy mode
- *  `src` must point to the start of a ZSTD frame, ZSTD legacy frame, or skippable frame
- *  `srcSize` must be at least as large as the frame contained
- *  @return : the compressed size of the frame starting at `src` */
+ * See docs in zstd.h
+ * Note: compatible with legacy mode */
 size_t ZSTD_findFrameCompressedSize(const void *src, size_t srcSize)
 {
-    ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize);
-    return frameSizeInfo.compressedSize;
+    return ZSTD_findFrameCompressedSize_advanced(src, srcSize, ZSTD_f_zstd1);
 }
 
 /* ZSTD_decompressBound() :
  *  compatible with legacy mode
- *  `src` must point to the start of a ZSTD frame or a skippeable frame
+ *  `src` must point to the start of a ZSTD frame or a skippable frame
  *  `srcSize` must be at least as large as the frame contained
  *  @return : the maximum decompressed size of the compressed source
  */
···
     unsigned long long bound = 0;
     /* Iterate over each frame */
     while (srcSize > 0) {
-        ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize);
+        ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize, ZSTD_f_zstd1);
         size_t const compressedSize = frameSizeInfo.compressedSize;
         unsigned long long const decompressedBound = frameSizeInfo.decompressedBound;
         if (ZSTD_isError(compressedSize) || decompressedBound == ZSTD_CONTENTSIZE_ERROR)
···
     return bound;
 }
 
+size_t ZSTD_decompressionMargin(void const* src, size_t srcSize)
+{
+    size_t margin = 0;
+    unsigned maxBlockSize = 0;
+
+    /* Iterate over each frame */
+    while (srcSize > 0) {
+        ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize, ZSTD_f_zstd1);
+        size_t const compressedSize = frameSizeInfo.compressedSize;
+        unsigned long long const decompressedBound = frameSizeInfo.decompressedBound;
+        ZSTD_FrameHeader zfh;
+
+        FORWARD_IF_ERROR(ZSTD_getFrameHeader(&zfh, src, srcSize), "");
+        if (ZSTD_isError(compressedSize) || decompressedBound == ZSTD_CONTENTSIZE_ERROR)
+            return ERROR(corruption_detected);
+
+        if (zfh.frameType == ZSTD_frame) {
+            /* Add the frame header to our margin */
+            margin += zfh.headerSize;
+            /* Add the checksum to our margin */
+            margin += zfh.checksumFlag ? 4 : 0;
+            /* Add 3 bytes per block */
+            margin += 3 * frameSizeInfo.nbBlocks;
+
+            /* Compute the max block size */
+            maxBlockSize = MAX(maxBlockSize, zfh.blockSizeMax);
+        } else {
+            assert(zfh.frameType == ZSTD_skippableFrame);
+            /* Add the entire skippable frame size to our margin. */
+            margin += compressedSize;
+        }
+
+        assert(srcSize >= compressedSize);
+        src = (const BYTE*)src + compressedSize;
+        srcSize -= compressedSize;
+    }
+
+    /* Add the max block size back to the margin. */
+    margin += maxBlockSize;
+
+    return margin;
+}
 
 /*-*************************************************************
  *   Frame decoding
···
     return regenSize;
 }
 
-static void ZSTD_DCtx_trace_end(ZSTD_DCtx const* dctx, U64 uncompressedSize, U64 compressedSize, unsigned streaming)
+static void ZSTD_DCtx_trace_end(ZSTD_DCtx const* dctx, U64 uncompressedSize, U64 compressedSize, int streaming)
 {
     (void)dctx;
     (void)uncompressedSize;
···
         ip += frameHeaderSize; remainingSrcSize -= frameHeaderSize;
     }
 
+    /* Shrink the blockSizeMax if enabled */
+    if (dctx->maxBlockSizeParam != 0)
+        dctx->fParams.blockSizeMax = MIN(dctx->fParams.blockSizeMax, (unsigned)dctx->maxBlockSizeParam);
+
     /* Loop on each block */
     while (1) {
         BYTE* oBlockEnd = oend;
···
         switch(blockProperties.blockType)
         {
         case bt_compressed:
-            decodedSize = ZSTD_decompressBlock_internal(dctx, op, (size_t)(oBlockEnd-op), ip, cBlockSize, /* frame */ 1, not_streaming);
+            assert(dctx->isFrameDecompression == 1);
+            decodedSize = ZSTD_decompressBlock_internal(dctx, op, (size_t)(oBlockEnd-op), ip, cBlockSize, not_streaming);
             break;
         case bt_raw :
             /* Use oend instead of oBlockEnd because this function is safe to overlap. It uses memmove. */
···
         default:
             RETURN_ERROR(corruption_detected, "invalid block type");
         }
-
-        if (ZSTD_isError(decodedSize)) return decodedSize;
-        if (dctx->validateChecksum)
+        FORWARD_IF_ERROR(decodedSize, "Block decompression failure");
+        DEBUGLOG(5, "Decompressed block of dSize = %u", (unsigned)decodedSize);
+        if (dctx->validateChecksum) {
             xxh64_update(&dctx->xxhState, op, decodedSize);
-        if (decodedSize != 0)
+        }
+        if (decodedSize) /* support dst = NULL,0 */ {
             op += decodedSize;
+        }
         assert(ip != NULL);
         ip += cBlockSize;
         remainingSrcSize -= cBlockSize;
···
     }
     ZSTD_DCtx_trace_end(dctx, (U64)(op-ostart), (U64)(ip-istart), /* streaming */ 0);
     /* Allow caller to get size read */
+    DEBUGLOG(4, "ZSTD_decompressFrame: decompressed frame of size %i, consuming %i bytes of input", (int)(op-ostart), (int)(ip - (const BYTE*)*srcPtr));
     *srcPtr = ip;
     *srcSizePtr = remainingSrcSize;
     return (size_t)(op-ostart);
 }
 
-static size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
                                         void* dst, size_t dstCapacity,
                                   const void* src, size_t srcSize,
                                   const void* dict, size_t dictSize,
···
     while (srcSize >= ZSTD_startingInputLength(dctx->format)) {
 
 
-        {   U32 const magicNumber = MEM_readLE32(src);
-            DEBUGLOG(4, "reading magic number %08X (expecting %08X)",
-                        (unsigned)magicNumber, ZSTD_MAGICNUMBER);
+        if (dctx->format == ZSTD_f_zstd1 && srcSize >= 4) {
+            U32 const magicNumber = MEM_readLE32(src);
+            DEBUGLOG(5, "reading magic number %08X", (unsigned)magicNumber);
             if ((magicNumber & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
+                /* skippable frame detected : skip it */
                 size_t const skippableSize = readSkippableFrameSize(src, srcSize);
-                FORWARD_IF_ERROR(skippableSize, "readSkippableFrameSize failed");
+                FORWARD_IF_ERROR(skippableSize, "invalid skippable frame");
                 assert(skippableSize <= srcSize);
 
                 src = (const BYTE *)src + skippableSize;
                 srcSize -= skippableSize;
-                continue;
+                continue; /* check next frame */
         }   }
 
         if (ddict) {
···
 size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx) { return dctx->expected; }
 
 /*
- * Similar to ZSTD_nextSrcSizeToDecompress(), but when a block input can be streamed,
- * we allow taking a partial block as the input. Currently only raw uncompressed blocks can
+ * Similar to ZSTD_nextSrcSizeToDecompress(), but when a block input can be streamed, we
+ * allow taking a partial block as the input. Currently only raw uncompressed blocks can
  * be streamed.
  *
  * For blocks that can be streamed, this allows us to reduce the latency until we produce
···
             {
             case bt_compressed:
                 DEBUGLOG(5, "ZSTD_decompressContinue: case bt_compressed");
-                rSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 1, is_streaming);
+                assert(dctx->isFrameDecompression == 1);
+                rSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, is_streaming);
                 dctx->expected = 0;  /* Streaming not supported */
                 break;
             case bt_raw :
···
     case ZSTDds_decodeSkippableHeader:
         assert(src != NULL);
         assert(srcSize <= ZSTD_SKIPPABLEHEADERSIZE);
+        assert(dctx->format != ZSTD_f_zstd1_magicless);
         ZSTD_memcpy(dctx->headerBuffer + (ZSTD_SKIPPABLEHEADERSIZE - srcSize), src, srcSize);   /* complete skippable header */
         dctx->expected = MEM_readLE32(dctx->headerBuffer + ZSTD_FRAMEIDSIZE);   /* note : dctx->expected can grow seriously large, beyond local buffer size */
         dctx->stage = ZSTDds_skipFrame;
···
 
     default:
         assert(0);   /* impossible */
-        RETURN_ERROR(GENERIC, "impossible to reach");   /* some compiler require default to do something */
+        RETURN_ERROR(GENERIC, "impossible to reach");   /* some compilers require default to do something */
     }
 }
 
···
         /* in minimal huffman, we always use X1 variants */
         size_t const hSize = HUF_readDTableX1_wksp(entropy->hufTable,
                                                 dictPtr, dictEnd - dictPtr,
-                                                workspace, workspaceSize);
+                                                workspace, workspaceSize, /* flags */ 0);
 #else
         size_t const hSize = HUF_readDTableX2_wksp(entropy->hufTable,
                                                 dictPtr, (size_t)(dictEnd - dictPtr),
-                                                workspace, workspaceSize);
+                                                workspace, workspaceSize, /* flags */ 0);
 #endif
         RETURN_ERROR_IF(HUF_isError(hSize), dictionary_corrupted, "");
         dictPtr += hSize;
···
     dctx->prefixStart = NULL;
     dctx->virtualStart = NULL;
     dctx->dictEnd = NULL;
-    dctx->entropy.hufTable[0] = (HUF_DTable)((HufLog)*0x1000001);  /* cover both little and big endian */
+    dctx->entropy.hufTable[0] = (HUF_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);  /* cover both little and big endian */
     dctx->litEntropy = dctx->fseEntropy = 0;
     dctx->dictID = 0;
     dctx->bType = bt_reserved;
+    dctx->isFrameDecompression = 1;
     ZSTD_STATIC_ASSERT(sizeof(dctx->entropy.rep) == sizeof(repStartValue));
     ZSTD_memcpy(dctx->entropy.rep, repStartValue, sizeof(repStartValue));  /* initial repcodes */
     dctx->LLTptr = dctx->entropy.LLTable;
···
  *  This could for one of the following reasons :
  *  - The frame does not require a dictionary (most common case).
  *  - The frame was built with dictID intentionally removed.
- *    Needed dictionary is a hidden information.
+ *    Needed dictionary is a hidden piece of information.
  *    Note : this use case also happens when using a non-conformant dictionary.
  *  - `srcSize` is too small, and as a result, frame header could not be decoded.
  *    Note : possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`.
···
  *  ZSTD_getFrameHeader(), which will provide a more precise error code. */
 unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize)
 {
-    ZSTD_frameHeader zfp = { 0, 0, 0, ZSTD_frame, 0, 0, 0 };
+    ZSTD_FrameHeader zfp = { 0, 0, 0, ZSTD_frame, 0, 0, 0, 0, 0 };
     size_t const hError = ZSTD_getFrameHeader(&zfp, src, srcSize);
     if (ZSTD_isError(hError)) return 0;
     return zfp.dictID;
···
 size_t ZSTD_initDStream(ZSTD_DStream* zds)
 {
     DEBUGLOG(4, "ZSTD_initDStream");
-    return ZSTD_initDStream_usingDDict(zds, NULL);
+    FORWARD_IF_ERROR(ZSTD_DCtx_reset(zds, ZSTD_reset_session_only), "");
+    FORWARD_IF_ERROR(ZSTD_DCtx_refDDict(zds, NULL), "");
+    return ZSTD_startingInputLength(zds->format);
 }
 
 /* ZSTD_initDStream_usingDDict() :
···
  * this function cannot fail */
 size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* dctx, const ZSTD_DDict* ddict)
 {
+    DEBUGLOG(4, "ZSTD_initDStream_usingDDict");
     FORWARD_IF_ERROR( ZSTD_DCtx_reset(dctx, ZSTD_reset_session_only) , "");
     FORWARD_IF_ERROR( ZSTD_DCtx_refDDict(dctx, ddict) , "");
     return ZSTD_startingInputLength(dctx->format);
···
  * this function cannot fail */
 size_t ZSTD_resetDStream(ZSTD_DStream* dctx)
 {
+    DEBUGLOG(4, "ZSTD_resetDStream");
     FORWARD_IF_ERROR(ZSTD_DCtx_reset(dctx, ZSTD_reset_session_only), "");
     return ZSTD_startingInputLength(dctx->format);
 }
···
             bounds.lowerBound = (int)ZSTD_rmd_refSingleDDict;
             bounds.upperBound = (int)ZSTD_rmd_refMultipleDDicts;
             return bounds;
+        case ZSTD_d_disableHuffmanAssembly:
+            bounds.lowerBound = 0;
+            bounds.upperBound = 1;
+            return bounds;
+        case ZSTD_d_maxBlockSize:
+            bounds.lowerBound = ZSTD_BLOCKSIZE_MAX_MIN;
+            bounds.upperBound = ZSTD_BLOCKSIZE_MAX;
+            return bounds;
+
         default:;
     }
     bounds.error = ERROR(parameter_unsupported);
···
         case ZSTD_d_refMultipleDDicts:
             *value = (int)dctx->refMultipleDDicts;
             return 0;
+        case ZSTD_d_disableHuffmanAssembly:
+            *value = (int)dctx->disableHufAsm;
+            return 0;
+        case ZSTD_d_maxBlockSize:
+            *value = dctx->maxBlockSizeParam;
+            return 0;
         default:;
     }
     RETURN_ERROR(parameter_unsupported, "");
···
             }
             dctx->refMultipleDDicts = (ZSTD_refMultipleDDicts_e)value;
             return 0;
+        case ZSTD_d_disableHuffmanAssembly:
+            CHECK_DBOUNDS(ZSTD_d_disableHuffmanAssembly, value);
+            dctx->disableHufAsm = value != 0;
+            return 0;
+        case ZSTD_d_maxBlockSize:
+            if (value != 0) CHECK_DBOUNDS(ZSTD_d_maxBlockSize, value);
+            dctx->maxBlockSizeParam = value;
+            return 0;
         default:;
     }
     RETURN_ERROR(parameter_unsupported, "");
···
       || (reset == ZSTD_reset_session_and_parameters) ) {
         dctx->streamStage = zdss_init;
         dctx->noForwardProgress = 0;
+        dctx->isFrameDecompression = 1;
     }
     if ( (reset == ZSTD_reset_parameters)
       || (reset == ZSTD_reset_session_and_parameters) ) {
···
     return ZSTD_sizeof_DCtx(dctx);
 }
 
-size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize)
+static size_t ZSTD_decodingBufferSize_internal(unsigned long long windowSize, unsigned long long frameContentSize, size_t blockSizeMax)
 {
-    size_t const blockSize = (size_t) MIN(windowSize, ZSTD_BLOCKSIZE_MAX);
-    /* space is needed to store the litbuffer after the output of a given block without stomping the extDict of a previous run, as well as to cover both windows against wildcopy*/
-    unsigned long long const neededRBSize = windowSize + blockSize + ZSTD_BLOCKSIZE_MAX + (WILDCOPY_OVERLENGTH * 2);
+    size_t const blockSize = MIN((size_t)MIN(windowSize, ZSTD_BLOCKSIZE_MAX), blockSizeMax);
+    /* We need blockSize + WILDCOPY_OVERLENGTH worth of buffer so that if a block
+     * ends at windowSize + WILDCOPY_OVERLENGTH + 1 bytes, we can start writing
+     * the block at the beginning of the output buffer, and maintain a full window.
+     *
+     * We need another blockSize worth of buffer so that we can store split
+     * literals at the end of the block without overwriting the extDict window.
+     */
+    unsigned long long const neededRBSize = windowSize + (blockSize * 2) + (WILDCOPY_OVERLENGTH * 2);
     unsigned long long const neededSize = MIN(frameContentSize, neededRBSize);
     size_t const minRBSize = (size_t) neededSize;
     RETURN_ERROR_IF((unsigned long long)minRBSize != neededSize,
                     frameParameter_windowTooLarge, "");
     return minRBSize;
+}
+
+size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize)
+{
+    return ZSTD_decodingBufferSize_internal(windowSize, frameContentSize, ZSTD_BLOCKSIZE_MAX);
 }
 
 size_t ZSTD_estimateDStreamSize(size_t windowSize)
···
 size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize)
 {
     U32 const windowSizeMax = 1U << ZSTD_WINDOWLOG_MAX;   /* note : should be user-selectable, but requires an additional parameter (or a dctx) */
-    ZSTD_frameHeader zfh;
+    ZSTD_FrameHeader zfh;
     size_t const err = ZSTD_getFrameHeader(&zfh, src, srcSize);
     if (ZSTD_isError(err)) return err;
     RETURN_ERROR_IF(err>0, srcSize_wrong, "");
···
     U32 someMoreWork = 1;
 
     DEBUGLOG(5, "ZSTD_decompressStream");
+    assert(zds != NULL);
     RETURN_ERROR_IF(
         input->pos > input->size,
         srcSize_wrong,
···
                 if (zds->refMultipleDDicts && zds->ddictSet) {
                     ZSTD_DCtx_selectFrameDDict(zds);
                 }
-                DEBUGLOG(5, "header size : %u", (U32)hSize);
                 if (ZSTD_isError(hSize)) {
                     return hSize;   /* error */
                 }
···
                             zds->lhSize += remainingInput;
                         }
                         input->pos = input->size;
+                        /* check first few bytes */
+                        FORWARD_IF_ERROR(
+                            ZSTD_getFrameHeader_advanced(&zds->fParams, zds->headerBuffer, zds->lhSize, zds->format),
+                            "First few bytes detected incorrect" );
+                        /* return hint input size */
                         return (MAX((size_t)ZSTD_FRAMEHEADERSIZE_MIN(zds->format), hSize) - zds->lhSize) + ZSTD_blockHeaderSize;   /* remaining header bytes + next block header */
                     }
                     assert(ip != NULL);
···
             if (zds->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN
                 && zds->fParams.frameType != ZSTD_skippableFrame
                 && (U64)(size_t)(oend-op) >= zds->fParams.frameContentSize) {
-                size_t const cSize = ZSTD_findFrameCompressedSize(istart, (size_t)(iend-istart));
+                size_t const cSize = ZSTD_findFrameCompressedSize_advanced(istart, (size_t)(iend-istart), zds->format);
                 if (cSize <= (size_t)(iend-istart)) {
                     /* shortcut : using single-pass mode */
                     size_t const decompressedSize = ZSTD_decompress_usingDDict(zds, op, (size_t)(oend-op), istart, cSize, ZSTD_getDDict(zds));
                     if (ZSTD_isError(decompressedSize)) return decompressedSize;
-                    DEBUGLOG(4, "shortcut to single-pass ZSTD_decompress_usingDDict()")
+                    DEBUGLOG(4, "shortcut to single-pass ZSTD_decompress_usingDDict()");
+                    assert(istart != NULL);
                     ip = istart + cSize;
-                    op += decompressedSize;
+                    op = op ? op + decompressedSize : op; /* can occur if frameContentSize = 0 (empty frame) */
                     zds->expected = 0;
                     zds->streamStage = zdss_init;
                     someMoreWork = 0;
···
             DEBUGLOG(4, "Consume header");
             FORWARD_IF_ERROR(ZSTD_decompressBegin_usingDDict(zds, ZSTD_getDDict(zds)), "");
 
-            if ((MEM_readLE32(zds->headerBuffer) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {  /* skippable frame */
+            if (zds->format == ZSTD_f_zstd1
+                && (MEM_readLE32(zds->headerBuffer) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {  /* skippable frame */
                 zds->expected = MEM_readLE32(zds->headerBuffer + ZSTD_FRAMEIDSIZE);
                 zds->stage = ZSTDds_skipFrame;
             } else {
···
             zds->fParams.windowSize = MAX(zds->fParams.windowSize, 1U << ZSTD_WINDOWLOG_ABSOLUTEMIN);
             RETURN_ERROR_IF(zds->fParams.windowSize > zds->maxWindowSize,
                             frameParameter_windowTooLarge, "");
+            if (zds->maxBlockSizeParam != 0)
+                zds->fParams.blockSizeMax = MIN(zds->fParams.blockSizeMax, (unsigned)zds->maxBlockSizeParam);
 
             /* Adapt buffer sizes to frame header instructions */
             {   size_t const neededInBuffSize = MAX(zds->fParams.blockSizeMax, 4 /* frame checksum */);
                 size_t const neededOutBuffSize = zds->outBufferMode == ZSTD_bm_buffered
-                        ? ZSTD_decodingBufferSize_min(zds->fParams.windowSize, zds->fParams.frameContentSize)
+                        ? ZSTD_decodingBufferSize_internal(zds->fParams.windowSize, zds->fParams.frameContentSize, zds->fParams.blockSizeMax)
                         : 0;
 
                 ZSTD_DCtx_updateOversizedDuration(zds, neededInBuffSize, neededOutBuffSize);
···
                 }
                 if ((size_t)(iend-ip) >= neededInSize) {  /* decode directly from src */
                     FORWARD_IF_ERROR(ZSTD_decompressContinueStream(zds, &op, oend, ip, neededInSize), "");
+                    assert(ip != NULL);
                     ip += neededInSize;
                     /* Function modifies the stage so we must break */
                     break;
···
                 int const isSkipFrame = ZSTD_isSkipFrame(zds);
                 size_t loadedSize;
                 /* At this point we shouldn't be decompressing a block that we can stream. */
-                assert(neededInSize == ZSTD_nextSrcSizeToDecompressWithInputSize(zds, iend - ip));
+                assert(neededInSize == ZSTD_nextSrcSizeToDecompressWithInputSize(zds, (size_t)(iend - ip)));
                 if (isSkipFrame) {
                     loadedSize = MIN(toLoad, (size_t)(iend-ip));
                 } else {
···
                                     "should never happen");
                     loadedSize = ZSTD_limitCopy(zds->inBuff + zds->inPos, toLoad, ip, (size_t)(iend-ip));
                 }
-                ip += loadedSize;
-                zds->inPos += loadedSize;
+                if (loadedSize != 0) {
+                    /* ip may be NULL */
+                    ip += loadedSize;
+                    zds->inPos += loadedSize;
+                }
                 if (loadedSize < toLoad) { someMoreWork = 0; break; }   /* not enough input, wait for more */
 
                 /* decode loaded input */
···
                 break;
             }
         case zdss_flush:
-            {   size_t const toFlushSize = zds->outEnd - zds->outStart;
+            {
+                size_t const toFlushSize = zds->outEnd - zds->outStart;
                 size_t const flushedSize = ZSTD_limitCopy(op, (size_t)(oend-op), zds->outBuff + zds->outStart, toFlushSize);
-                op += flushedSize;
+
+                op = op ? op + flushedSize : op;
+
                 zds->outStart += flushedSize;
                 if (flushedSize == toFlushSize) {  /* flush completed */
                     zds->streamStage = zdss_read;
                     if ( (zds->outBuffSize < zds->fParams.frameContentSize)
-                      && (zds->outStart + zds->fParams.blockSizeMax > zds->outBuffSize) ) {
+                        && (zds->outStart + zds->fParams.blockSizeMax > zds->outBuffSize) ) {
                         DEBUGLOG(5, "restart filling outBuff from beginning (left:%i, needed:%u)",
                                 (int)(zds->outBuffSize - zds->outStart),
                                 (U32)zds->fParams.blockSizeMax);
···
 
         default:
             assert(0);    /* impossible */
-            RETURN_ERROR(GENERIC, "impossible to reach");   /* some compiler require default to do something */
+            RETURN_ERROR(GENERIC, "impossible to reach");   /* some compilers require default to do something */
     }   }
 
     /* result */
···
     if ((ip==istart) && (op==ostart)) {  /* no forward progress */
         zds->noForwardProgress ++;
         if (zds->noForwardProgress >= ZSTD_NO_FORWARD_PROGRESS_MAX) {
-            RETURN_ERROR_IF(op==oend, dstSize_tooSmall, "");
-            RETURN_ERROR_IF(ip==iend, srcSize_wrong, "");
+            RETURN_ERROR_IF(op==oend, noForwardProgress_destFull, "");
+            RETURN_ERROR_IF(ip==iend, noForwardProgress_inputEmpty, "");
             assert(0);
         }
     } else {
···
                             void* dst, size_t dstCapacity, size_t* dstPos,
                       const void* src, size_t srcSize, size_t* srcPos)
 {
-    ZSTD_outBuffer output = { dst, dstCapacity, *dstPos };
-    ZSTD_inBuffer  input  = { src, srcSize, *srcPos };
-    /* ZSTD_compress_generic() will check validity of dstPos and srcPos */
-    size_t const cErr = ZSTD_decompressStream(dctx, &output, &input);
-    *dstPos = output.pos;
-    *srcPos = input.pos;
-    return cErr;
+    ZSTD_outBuffer output;
+    ZSTD_inBuffer  input;
+    output.dst = dst;
+    output.size = dstCapacity;
+    output.pos = *dstPos;
+    input.src = src;
+    input.size = srcSize;
+    input.pos = *srcPos;
+    {   size_t const cErr = ZSTD_decompressStream(dctx, &output, &input);
+        *dstPos = output.pos;
+        *srcPos = input.pos;
+        return cErr;
+    }
 }

lib/zstd/decompress/zstd_decompress_block.c

···
+// SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 #include "../common/mem.h"         /* low level memory routines */
 #define FSE_STATIC_LINKING_ONLY
 #include "../common/fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"
 #include "../common/zstd_internal.h"
 #include "zstd_decompress_internal.h"   /* ZSTD_DCtx */
 #include "zstd_ddict.h"  /* ZSTD_DDictDictContent */
 #include "zstd_decompress_block.h"
+#include "../common/bits.h"  /* ZSTD_highbit32 */
 
 /*_*******************************************************
 *  Macros
···
  *   Block decoding
  ***************************************************************/
 
+static size_t ZSTD_blockSizeMax(ZSTD_DCtx const* dctx)
+{
+    size_t const blockSizeMax = dctx->isFrameDecompression ? dctx->fParams.blockSizeMax : ZSTD_BLOCKSIZE_MAX;
+    assert(blockSizeMax <= ZSTD_BLOCKSIZE_MAX);
+    return blockSizeMax;
+}
+
 /*! ZSTD_getcBlockSize() :
  *  Provides the size of compressed block from block header `src` */
 size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
···
 static void ZSTD_allocateLiteralsBuffer(ZSTD_DCtx* dctx, void* const dst, const size_t dstCapacity, const size_t litSize,
     const streaming_operation streaming, const size_t expectedWriteSize, const unsigned splitImmediately)
 {
-    if (streaming == not_streaming && dstCapacity > ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH + litSize + WILDCOPY_OVERLENGTH)
-    {
-        /* room for litbuffer to fit without read faulting */
-        dctx->litBuffer = (BYTE*)dst + ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH;
+    size_t const blockSizeMax = ZSTD_blockSizeMax(dctx);
+    assert(litSize <= blockSizeMax);
+    assert(dctx->isFrameDecompression || streaming == not_streaming);
+    assert(expectedWriteSize <= blockSizeMax);
+    if (streaming == not_streaming && dstCapacity > blockSizeMax + WILDCOPY_OVERLENGTH + litSize + WILDCOPY_OVERLENGTH) {
+        /* If we aren't streaming, we can just put the literals after the output
+         * of the current block. We don't need to worry about overwriting the
+         * extDict of our window, because it doesn't exist.
+         * So if we have space after the end of the block, just put it there.
+         */
+        dctx->litBuffer = (BYTE*)dst + blockSizeMax + WILDCOPY_OVERLENGTH;
         dctx->litBufferEnd = dctx->litBuffer + litSize;
         dctx->litBufferLocation = ZSTD_in_dst;
-    }
-    else if (litSize > ZSTD_LITBUFFEREXTRASIZE)
-    {
-        /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
+    } else if (litSize <= ZSTD_LITBUFFEREXTRASIZE) {
+        /* Literals fit entirely within the extra buffer, put them there to avoid
+         * having to split the literals.
+         */
+        dctx->litBuffer = dctx->litExtraBuffer;
+        dctx->litBufferEnd = dctx->litBuffer + litSize;
+        dctx->litBufferLocation = ZSTD_not_in_dst;
+    } else {
+        assert(blockSizeMax > ZSTD_LITBUFFEREXTRASIZE);
+        /* Literals must be split between the output block and the extra lit
+         * buffer. We fill the extra lit buffer with the tail of the literals,
+         * and put the rest of the literals at the end of the block, with
+         * WILDCOPY_OVERLENGTH of buffer room to allow for overreads.
+         * This MUST not write more than our maxBlockSize beyond dst, because in
+         * streaming mode, that could overwrite part of our extDict window.
+         */
         if (splitImmediately) {
             /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
             dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
             dctx->litBufferEnd = dctx->litBuffer + litSize - ZSTD_LITBUFFEREXTRASIZE;
-        }
-        else {
-            /* initially this will be stored entirely in dst during huffman decoding, it will partially shifted to litExtraBuffer after */
+        } else {
+            /* initially this will be stored entirely in dst during huffman decoding, it will partially be shifted to litExtraBuffer after */
             dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize;
             dctx->litBufferEnd = (BYTE*)dst + expectedWriteSize;
         }
         dctx->litBufferLocation = ZSTD_split;
-    }
-    else
-    {
-        /* fits entirely within litExtraBuffer, so no split is necessary */
-        dctx->litBuffer = dctx->litExtraBuffer;
-        dctx->litBufferEnd = dctx->litBuffer + litSize;
-        dctx->litBufferLocation = ZSTD_not_in_dst;
+        assert(dctx->litBufferEnd <= (BYTE*)dst + expectedWriteSize);
     }
 }
 
-/* Hidden declaration for fullbench */
-size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
-                          const void* src, size_t srcSize,
-                          void* dst, size_t dstCapacity, const streaming_operation streaming);
 /*! ZSTD_decodeLiteralsBlock() :
  * Where it is possible to do so without being stomped by the output during decompression, the literals block will be stored
  * in the dstBuffer.  If there is room to do so, it will be stored in full in the excess dst space after where the current
···
  *
  * @return : nb of bytes read from src (< srcSize )
  *  note : symbol not declared but exposed for fullbench */
-size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
+static size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                           const void* src, size_t srcSize,   /* note : srcSize < BLOCKSIZE */
                           void* dst, size_t dstCapacity, const streaming_operation streaming)
 {
···
     RETURN_ERROR_IF(srcSize < MIN_CBLOCK_SIZE, corruption_detected, "");
 
     {   const BYTE* const istart = (const BYTE*) src;
-        symbolEncodingType_e const litEncType = (symbolEncodingType_e)(istart[0] & 3);
+        SymbolEncodingType_e const litEncType = (SymbolEncodingType_e)(istart[0] & 3);
+        size_t const blockSizeMax = ZSTD_blockSizeMax(dctx);
 
         switch(litEncType)
         {
···
             ZSTD_FALLTHROUGH;
 
         case set_compressed:
-            RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 3; here we need up to 5 for case 3");
+            RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need up to 5 for case 3");
             {   size_t lhSize, litSize, litCSize;
                 U32 singleStream=0;
                 U32 const lhlCode = (istart[0] >> 2) & 3;
                 U32 const lhc = MEM_readLE32(istart);
                 size_t hufSuccess;
-                size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
+                size_t expectedWriteSize = MIN(blockSizeMax, dstCapacity);
+                int const flags = 0
+                    | (ZSTD_DCtx_get_bmi2(dctx) ? HUF_flags_bmi2 : 0)
+                    | (dctx->disableHufAsm ? HUF_flags_disableAsm : 0);
                 switch(lhlCode)
                 {
                 case 0: case 1: default:   /* note : default is impossible, since lhlCode into [0..3] */
···
                     break;
                 }
                 RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
-                RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
+                RETURN_ERROR_IF(litSize > blockSizeMax, corruption_detected, "");
+                if (!singleStream)
+                    RETURN_ERROR_IF(litSize < MIN_LITERALS_FOR_4_STREAMS, literals_headerWrong,
+                        "Not enough literals (%zu) for the 4-streams mode (min %u)",
+                        litSize, MIN_LITERALS_FOR_4_STREAMS);
                 RETURN_ERROR_IF(litCSize + lhSize > srcSize, corruption_detected, "");
                 RETURN_ERROR_IF(expectedWriteSize < litSize , dstSize_tooSmall, "");
                 ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 0);
···
 
                 if (litEncType==set_repeat) {
                     if (singleStream) {
-                        hufSuccess = HUF_decompress1X_usingDTable_bmi2(
+                        hufSuccess = HUF_decompress1X_usingDTable(
                             dctx->litBuffer, litSize, istart+lhSize, litCSize,
-                            dctx->HUFptr, ZSTD_DCtx_get_bmi2(dctx));
+                            dctx->HUFptr, flags);
                     } else {
-                        hufSuccess = HUF_decompress4X_usingDTable_bmi2(
+                        assert(litSize >= MIN_LITERALS_FOR_4_STREAMS);
+                        hufSuccess = HUF_decompress4X_usingDTable(
                             dctx->litBuffer, litSize, istart+lhSize, litCSize,
-                            dctx->HUFptr, ZSTD_DCtx_get_bmi2(dctx));
+                            dctx->HUFptr, flags);
                     }
                 } else {
                     if (singleStream) {
···
                         hufSuccess = HUF_decompress1X_DCtx_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace));
+                            sizeof(dctx->workspace), flags);
 #else
-                        hufSuccess = HUF_decompress1X1_DCtx_wksp_bmi2(
+                        hufSuccess = HUF_decompress1X1_DCtx_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace), ZSTD_DCtx_get_bmi2(dctx));
+                            sizeof(dctx->workspace), flags);
 #endif
                     } else {
-                        hufSuccess = HUF_decompress4X_hufOnly_wksp_bmi2(
+                        hufSuccess = HUF_decompress4X_hufOnly_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace), ZSTD_DCtx_get_bmi2(dctx));
+                            sizeof(dctx->workspace), flags);
                     }
                 }
                 if (dctx->litBufferLocation == ZSTD_split)
                 {
+                    assert(litSize > ZSTD_LITBUFFEREXTRASIZE);
                     ZSTD_memcpy(dctx->litExtraBuffer, dctx->litBufferEnd - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
                     ZSTD_memmove(dctx->litBuffer + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH, dctx->litBuffer, litSize - ZSTD_LITBUFFEREXTRASIZE);
                     dctx->litBuffer += ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
                     dctx->litBufferEnd -= WILDCOPY_OVERLENGTH;
+                    assert(dctx->litBufferEnd <= (BYTE*)dst + blockSizeMax);
                 }
 
                 RETURN_ERROR_IF(HUF_isError(hufSuccess), corruption_detected, "");
···
         case set_basic:
             {   size_t litSize, lhSize;
                 U32 const lhlCode = ((istart[0]) >> 2) & 3;
-                size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
+                size_t expectedWriteSize = MIN(blockSizeMax, dstCapacity);
                 switch(lhlCode)
                 {
                 case 0: case 2: default:   /* note : default is impossible, since lhlCode into [0..3] */
···
                     break;
                 case 3:
                     lhSize = 3;
+                    RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize = 3");
                     litSize = MEM_readLE24(istart) >> 4;
                     break;
                 }
 
                 RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
+                RETURN_ERROR_IF(litSize > blockSizeMax, corruption_detected, "");
                 RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
                 ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
                 if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) {  /* risk reading beyond src buffer with wildcopy */
···
         case set_rle:
             {   U32 const lhlCode = ((istart[0]) >> 2) & 3;
                 size_t litSize, lhSize;
-                size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
+                size_t expectedWriteSize = MIN(blockSizeMax, dstCapacity);
                 switch(lhlCode)
                 {
                 case 0: case 2: default:   /* note : default is impossible, since lhlCode into [0..3] */
···
                     break;
                 case 1:
                     lhSize = 2;
+                    RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 3");
                     litSize = MEM_readLE16(istart) >> 4;
                     break;
                 case 3:
                     lhSize = 3;
+                    RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 4");
                     litSize = MEM_readLE24(istart) >> 4;
-                    RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 3; here we need lhSize+1 = 4");
                     break;
                 }
                 RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
-                RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
+                RETURN_ERROR_IF(litSize > blockSizeMax, corruption_detected, "");
                 RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
                 ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
                 if (dctx->litBufferLocation == ZSTD_split)
···
     }
 }
 
+/* Hidden declaration for fullbench */
+size_t ZSTD_decodeLiteralsBlock_wrapper(ZSTD_DCtx* dctx,
+                          const void* src, size_t srcSize,
+                          void* dst, size_t dstCapacity);
+size_t ZSTD_decodeLiteralsBlock_wrapper(ZSTD_DCtx* dctx,
+                          const void* src, size_t srcSize,
+                          void* dst, size_t dstCapacity)
+{
+    dctx->isFrameDecompression = 0;
+    return ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, not_streaming);
+}
+
 /* Default FSE distribution tables.
  * These are pre-calculated FSE decoding tables using default distributions as defined in specification :
  * https://github.com/facebook/zstd/blob/release/doc/zstd_compression_format.md#default-distributions
···
  * - start from default distributions, present in /lib/common/zstd_internal.h
  * - generate tables normally, using ZSTD_buildFSETable()
  * - printout the content of tables
- * - pretify output, report below, test with fuzzer to ensure it's correct */
+ * - prettify output, report below, test with fuzzer to ensure it's correct */
 
 /* Default FSE distribution table for Literal Lengths */
 static const ZSTD_seqSymbol LL_defaultDTable[(1<<LL_DEFAULTNORMLOG)+1] = {
···
                 for (i = 8; i < n; i += 8) {
                     MEM_write64(spread + pos + i, sv);
                 }
-                pos += n;
+                assert(n>=0);
+                pos += (size_t)n;
             }
         }
         /* Now we spread those positions across the table.
-         * The benefit of doing it in two stages is that we avoid the the
+         * The benefit of doing it in two stages is that we avoid the
          * variable size inner loop, which caused lots of branch misses.
          * Now we can run through all the positions without any branch misses.
-         * We unroll the loop twice, since that is what emperically worked best.
+         * We unroll the loop twice, since that is what empirically worked best.
          */
         {
             size_t position = 0;
···
             for (i=0; i<n; i++) {
                 tableDecode[position].baseValue = s;
                 position = (position + step) & tableMask;
-                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+                while (UNLIKELY(position > highThreshold)) position = (position + step) & tableMask;   /* lowprob area */
         }   }
         assert(position == 0); /* position must reach all cells once, otherwise normalizedCounter is incorrect */
     }
···
         for (u=0; u<tableSize; u++) {
             U32 const symbol = tableDecode[u].baseValue;
             U32 const nextState = symbolNext[symbol]++;
-            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
+            tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
             tableDecode[u].nextState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
             assert(nbAdditionalBits[symbol] < 255);
             tableDecode[u].nbAdditionalBits = nbAdditionalBits[symbol];
···
  * @return : nb bytes read from src,
  *           or an error code if it fails */
 static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymbol** DTablePtr,
-                                 symbolEncodingType_e type, unsigned max, U32 maxLog,
+                                 SymbolEncodingType_e type, unsigned max, U32 maxLog,
                                  const void* src, size_t srcSize,
                                  const U32* baseValue, const U8* nbAdditionalBits,
                                  const ZSTD_seqSymbol* defaultTable, U32 flagRepeatTable,
···
 
     /* SeqHead */
     nbSeq = *ip++;
-    if (!nbSeq) {
-        *nbSeqPtr=0;
-        RETURN_ERROR_IF(srcSize != 1, srcSize_wrong, "");
-        return 1;
-    }
     if (nbSeq > 0x7F) {
         if (nbSeq == 0xFF) {
             RETURN_ERROR_IF(ip+2 > iend, srcSize_wrong, "");
···
     }
     *nbSeqPtr = nbSeq;
 
+    if (nbSeq == 0) {
+        /* No sequence : section ends immediately */
+        RETURN_ERROR_IF(ip != iend, corruption_detected,
+            "extraneous data present in the Sequences section");
+        return (size_t)(ip - istart);
+    }
+
     /* FSE table descriptors */
     RETURN_ERROR_IF(ip+1 > iend, srcSize_wrong, ""); /* minimum possible size: 1 byte for symbol encoding types */
-    {   symbolEncodingType_e const LLtype = (symbolEncodingType_e)(*ip >> 6);
-        symbolEncodingType_e const OFtype = (symbolEncodingType_e)((*ip >> 4) & 3);
-        symbolEncodingType_e const MLtype = (symbolEncodingType_e)((*ip >> 2) & 3);
+    RETURN_ERROR_IF(*ip & 3, corruption_detected, ""); /* The last field, Reserved, must be all-zeroes. */
+    {   SymbolEncodingType_e const LLtype = (SymbolEncodingType_e)(*ip >> 6);
+        SymbolEncodingType_e const OFtype = (SymbolEncodingType_e)((*ip >> 4) & 3);
+        SymbolEncodingType_e const MLtype = (SymbolEncodingType_e)((*ip >> 2) & 3);
         ip++;
 
         /* Build DTables */
···
 /* ZSTD_safecopyDstBeforeSrc():
  * This version allows overlap with dst before src, or handles the non-overlap case with dst after src
  * Kept separate from more common ZSTD_safecopy case to avoid performance impact to the safecopy common case */
-static void ZSTD_safecopyDstBeforeSrc(BYTE* op, BYTE const* ip, ptrdiff_t length) {
+static void ZSTD_safecopyDstBeforeSrc(BYTE* op, const BYTE* ip, ptrdiff_t length) {
     ptrdiff_t const diff = op - ip;
     BYTE* const oend = op + length;
 
···
  * to be optimized for many small sequences, since those fall into ZSTD_execSequence().
  */
 FORCE_NOINLINE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_execSequenceEnd(BYTE* op,
     BYTE* const oend, seq_t sequence,
     const BYTE** litPtr, const BYTE* const litLimit,
···
  * This version is intended to be used during instances where the litBuffer is still split.  It is kept separate to avoid performance impact for the good case.
  */
 FORCE_NOINLINE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_execSequenceEndSplitLitBuffer(BYTE* op,
     BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
     const BYTE** litPtr, const BYTE* const litLimit,
···
 }
 
 HINT_INLINE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_execSequence(BYTE* op,
     BYTE* const oend, seq_t sequence,
     const BYTE** litPtr, const BYTE* const litLimit,
···
 
     assert(op != NULL /* Precondition */);
     assert(oend_w < oend /* No underflow */);
+
+#if defined(__aarch64__)
+    /* prefetch sequence starting from match that will be used for copy later */
+    PREFETCH_L1(match);
+#endif
     /* Handle edge cases in a slow path:
      *   - Read beyond end of literals
      *   - Match end is within WILDCOPY_OVERLIMIT of oend
···
 }
 
 HINT_INLINE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_execSequenceSplitLitBuffer(BYTE* op,
     BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
     const BYTE** litPtr, const BYTE* const litLimit,
···
 }
 
 /* We need to add at most (ZSTD_WINDOWLOG_MAX_32 - 1) bits to read the maximum
- * offset bits. But we can only read at most (STREAM_ACCUMULATOR_MIN_32 - 1)
+ * offset bits. But we can only read at most STREAM_ACCUMULATOR_MIN_32
  * bits before reloading. This value is the maximum number of bytes we read
  * after reloading when we are decoding long offsets.
  */
···
 
 typedef enum { ZSTD_lo_isRegularOffset, ZSTD_lo_isLongOffset=1 } ZSTD_longOffset_e;
 
+/*
+ * ZSTD_decodeSequence():
+ * @p longOffsets : tells the decoder to reload more bit while decoding large offsets
+ *                  only used in 32-bit mode
+ * @return : Sequence (litL + matchL + offset)
+ */
 FORCE_INLINE_TEMPLATE seq_t
-ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
+ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets, const int isLastSeq)
 {
     seq_t seq;
+    /*
+     * ZSTD_seqSymbol is a 64 bits wide structure.
+     * It can be loaded in one operation
+     * and its fields extracted by simply shifting or bit-extracting on aarch64.
+     * GCC doesn't recognize this and generates more unnecessary ldr/ldrb/ldrh
+     * operations that cause performance drop. This can be avoided by using this
+     * ZSTD_memcpy hack.
+     */
+#if defined(__aarch64__) && (defined(__GNUC__) && !defined(__clang__))
+    ZSTD_seqSymbol llDInfoS, mlDInfoS, ofDInfoS;
+    ZSTD_seqSymbol* const llDInfo = &llDInfoS;
+    ZSTD_seqSymbol* const mlDInfo = &mlDInfoS;
+    ZSTD_seqSymbol* const ofDInfo = &ofDInfoS;
+    ZSTD_memcpy(llDInfo, seqState->stateLL.table + seqState->stateLL.state, sizeof(ZSTD_seqSymbol));
+    ZSTD_memcpy(mlDInfo, seqState->stateML.table + seqState->stateML.state, sizeof(ZSTD_seqSymbol));
+    ZSTD_memcpy(ofDInfo, seqState->stateOffb.table + seqState->stateOffb.state, sizeof(ZSTD_seqSymbol));
+#else
     const ZSTD_seqSymbol* const llDInfo = seqState->stateLL.table + seqState->stateLL.state;
     const ZSTD_seqSymbol* const mlDInfo = seqState->stateML.table + seqState->stateML.state;
     const ZSTD_seqSymbol* const ofDInfo = seqState->stateOffb.table + seqState->stateOffb.state;
+#endif
     seq.matchLength = mlDInfo->baseValue;
     seq.litLength = llDInfo->baseValue;
     {   U32 const ofBase = ofDInfo->baseValue;
···
         U32 const llnbBits = llDInfo->nbBits;
         U32 const mlnbBits = mlDInfo->nbBits;
         U32 const ofnbBits = ofDInfo->nbBits;
+
+        assert(llBits <= MaxLLBits);
+        assert(mlBits <= MaxMLBits);
+        assert(ofBits <= MaxOff);
         /*
          * As gcc has better branch and block analyzers, sometimes it is only
-         * valuable to mark likelyness for clang, it gives around 3-4% of
+         * valuable to mark likeliness for clang, it gives around 3-4% of
          * performance.
          */
 
         /* sequence */
         {   size_t offset;
-    #if defined(__clang__)
-            if (LIKELY(ofBits > 1)) {
-    #else
             if (ofBits > 1) {
-    #endif
                 ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
                 ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
-                assert(ofBits <= MaxOff);
+                ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 > LONG_OFFSETS_MAX_EXTRA_BITS_32);
+                ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 - LONG_OFFSETS_MAX_EXTRA_BITS_32 >= MaxMLBits);
                 if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
-                    U32 const extraBits = ofBits - MIN(ofBits, 32 - seqState->DStream.bitsConsumed);
+                    /* Always read extra bits, this keeps the logic simple,
+                     * avoids branches, and avoids accidentally reading 0 bits.
+                     */
+                    U32 const extraBits = LONG_OFFSETS_MAX_EXTRA_BITS_32;
                     offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
                     BIT_reloadDStream(&seqState->DStream);
-                    if (extraBits) offset += BIT_readBitsFast(&seqState->DStream, extraBits);
-                    assert(extraBits <= LONG_OFFSETS_MAX_EXTRA_BITS_32);   /* to avoid another reload */
+                    offset += BIT_readBitsFast(&seqState->DStream, extraBits);
                 } else {
                     offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
                     if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
···
                 } else {
                     offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
                     {   size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
-                        temp += !temp;   /* 0 is not valid; input is corrupted; force offset to 1 */
+                        temp -= !temp; /* 0 is not valid: input corrupted => force offset to -1 => corruption detected at execSequence */
                         if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
                         seqState->prevOffset[1] = seqState->prevOffset[0];
                         seqState->prevOffset[0] = offset = temp;
···
             seq.offset = offset;
         }
 
-    #if defined(__clang__)
-        if (UNLIKELY(mlBits > 0))
-    #else
         if (mlBits > 0)
-    #endif
             seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);
 
         if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
···
         /* Ensure there are enough bits to read the rest of data in 64-bit mode. */
         ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);
 
-    #if defined(__clang__)
-        if (UNLIKELY(llBits > 0))
-    #else
         if (llBits > 0)
-    #endif
             seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);
 
         if (MEM_32bits())
···
         DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
                     (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
 
-        ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llNext, llnbBits);    /* <=  9 bits */
-        ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlNext, mlnbBits);    /* <=  9 bits */
-        if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
-        ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofNext, ofnbBits);  /* <=  8 bits */
+        if (!isLastSeq) {
+            /* don't update FSE state for last Sequence */
+            ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llNext, llnbBits);    /* <=  9 bits */
+            ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlNext, mlnbBits);    /* <=  9 bits */
+            if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
+            ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofNext, ofnbBits);  /* <=  8 bits */
+            BIT_reloadDStream(&seqState->DStream);
+        }
     }
 
     return seq;
 }
 
-#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
-MEM_STATIC int ZSTD_dictionaryIsActive(ZSTD_DCtx const* dctx, BYTE const* prefixStart, BYTE const* oLitEnd)
+#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
+#if DEBUGLEVEL >= 1
+static int ZSTD_dictionaryIsActive(ZSTD_DCtx const* dctx, BYTE const* prefixStart, BYTE const* oLitEnd)
 {
     size_t const windowSize = dctx->fParams.windowSize;
     /* No dictionary used. */
···
     /* Dictionary is active. */
     return 1;
 }
+#endif
 
-MEM_STATIC void ZSTD_assertValidSequence(
+static void ZSTD_assertValidSequence(
         ZSTD_DCtx const* dctx,
         BYTE const* op, BYTE const* oend,
         seq_t const seq,
         BYTE const* prefixStart, BYTE const* virtualStart)
 {
 #if DEBUGLEVEL >= 1
-    size_t const windowSize = dctx->fParams.windowSize;
-    size_t const sequenceSize = seq.litLength + seq.matchLength;
-    BYTE const* const oLitEnd = op + seq.litLength;
-    DEBUGLOG(6, "Checking sequence: litL=%u matchL=%u offset=%u",
-            (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
-    assert(op <= oend);
-    assert((size_t)(oend - op) >= sequenceSize);
-    assert(sequenceSize <= ZSTD_BLOCKSIZE_MAX);
-    if (ZSTD_dictionaryIsActive(dctx, prefixStart, oLitEnd)) {
-        size_t const dictSize = (size_t)((char const*)dctx->dictContentEndForFuzzing - (char const*)dctx->dictContentBeginForFuzzing);
-        /* Offset must be within the dictionary. */
-        assert(seq.offset <= (size_t)(oLitEnd - virtualStart));
-        assert(seq.offset <= windowSize + dictSize);
-    } else {
-        /* Offset must be within our window. */
-        assert(seq.offset <= windowSize);
+    if (dctx->isFrameDecompression) {
+        size_t const windowSize = dctx->fParams.windowSize;
+        size_t const sequenceSize = seq.litLength + seq.matchLength;
+        BYTE const* const oLitEnd = op + seq.litLength;
+        DEBUGLOG(6, "Checking sequence: litL=%u matchL=%u offset=%u",
+                (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
+        assert(op <= oend);
+        assert((size_t)(oend - op) >= sequenceSize);
+        assert(sequenceSize <= ZSTD_blockSizeMax(dctx));
+        if (ZSTD_dictionaryIsActive(dctx, prefixStart, oLitEnd)) {
+            size_t const dictSize = (size_t)((char const*)dctx->dictContentEndForFuzzing - (char const*)dctx->dictContentBeginForFuzzing);
+            /* Offset must be within the dictionary. */
+            assert(seq.offset <= (size_t)(oLitEnd - virtualStart));
+            assert(seq.offset <= windowSize + dictSize);
+        } else {
+            /* Offset must be within our window. */
+            assert(seq.offset <= windowSize);
+        }
     }
 #else
     (void)dctx, (void)op, (void)oend, (void)seq, (void)prefixStart, (void)virtualStart;
···
 ZSTD_decompressSequences_bodySplitLitBuffer( ZSTD_DCtx* dctx,
                                void* dst, size_t maxDstSize,
                          const void* seqStart, size_t seqSize, int nbSeq,
-                         const ZSTD_longOffset_e isLongOffset,
-                         const int frame)
+                         const ZSTD_longOffset_e isLongOffset)
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
     BYTE* const ostart = (BYTE*)dst;
-    BYTE* const oend = ostart + maxDstSize;
+    BYTE* const oend = ZSTD_maybeNullPtrAdd(ostart, maxDstSize);
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
     const BYTE* litBufferEnd = dctx->litBufferEnd;
     const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
     const BYTE* const vBase = (const BYTE*) (dctx->virtualStart);
     const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
-    DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer");
-    (void)frame;
+    DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer (%i seqs)", nbSeq);
 
-    /* Regen sequences */
+    /* Literals are split between internal buffer & output buffer */
     if (nbSeq) {
         seqState_t seqState;
         dctx->fseEntropy = 1;
···
                 BIT_DStream_completed < BIT_DStream_overflow);
 
         /* decompress without overrunning litPtr begins */
-        {
-            seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+        {   seq_t sequence = {0,0,0};  /* some static analyzer believe that @sequence is not initialized (it necessarily is, since for(;;) loop as at least one iteration) */
             /* Align the decompression loop to 32 + 16 bytes.
                 *
                 * zstd compiled with gcc-9 on an Intel i9-9900k shows 10% decompression
···
 #endif
 
             /* Handle the initial state where litBuffer is currently split between dst and litExtraBuffer */
-            for (; litPtr + sequence.litLength <= dctx->litBufferEnd; ) {
-                size_t const oneSeqSize = ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence.litLength - WILDCOPY_OVERLENGTH, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
+            for ( ; nbSeq; nbSeq--) {
+                sequence = ZSTD_decodeSequence(&seqState, isLongOffset, nbSeq==1);
+                if (litPtr + sequence.litLength > dctx->litBufferEnd) break;
+                {   size_t const oneSeqSize = ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence.litLength - WILDCOPY_OVERLENGTH, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
-                assert(!ZSTD_isError(oneSeqSize));
-                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
+                    assert(!ZSTD_isError(oneSeqSize));
+                    ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
 #endif
-                if (UNLIKELY(ZSTD_isError(oneSeqSize)))
-                    return oneSeqSize;
-                DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
-                op += oneSeqSize;
-                if (UNLIKELY(!--nbSeq))
-                    break;
-                BIT_reloadDStream(&(seqState.DStream));
-                sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
-            }
+                    if (UNLIKELY(ZSTD_isError(oneSeqSize)))
+                        return oneSeqSize;
+                    DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
+                    op += oneSeqSize;
+            }   }
+            DEBUGLOG(6, "reached: (litPtr + sequence.litLength > dctx->litBufferEnd)");
 
             /* If there are more sequences, they will need to read literals from litExtraBuffer; copy over the remainder from dst and update litPtr and litEnd */
             if (nbSeq > 0) {
                 const size_t leftoverLit = dctx->litBufferEnd - litPtr;
-                if (leftoverLit)
-                {
+                DEBUGLOG(6, "There are %i sequences left, and %zu/%zu literals left in buffer", nbSeq, leftoverLit, sequence.litLength);
+                if (leftoverLit) {
                     RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
                     ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
                     sequence.litLength -= leftoverLit;
···
                 litPtr = dctx->litExtraBuffer;
                 litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
                 dctx->litBufferLocation = ZSTD_not_in_dst;
-                {
-                    size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
+                {   size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                     assert(!ZSTD_isError(oneSeqSize));
-                    if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
+                    ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
 #endif
                     if (UNLIKELY(ZSTD_isError(oneSeqSize)))
                         return oneSeqSize;
                     DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
                     op += oneSeqSize;
-                    if (--nbSeq)
-                        BIT_reloadDStream(&(seqState.DStream));
                 }
+                nbSeq--;
             }
         }
 
-        if (nbSeq > 0) /* there is remaining lit from extra buffer */
-        {
+        if (nbSeq > 0) {
+            /* there is remaining lit from extra buffer */
 
 #if defined(__x86_64__)
             __asm__(".p2align 6");
···
 #  endif
 #endif
 
-            for (; ; ) {
-                seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+            for ( ; nbSeq ; nbSeq--) {
+                seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, nbSeq==1);
                 size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                 assert(!ZSTD_isError(oneSeqSize));
-                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
+                ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
 #endif
                 if (UNLIKELY(ZSTD_isError(oneSeqSize)))
                     return oneSeqSize;
                 DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
                 op += oneSeqSize;
-                if (UNLIKELY(!--nbSeq))
-                    break;
-                BIT_reloadDStream(&(seqState.DStream));
             }
         }
 
         /* check if reached exact end */
         DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer: after decode loop, remaining nbSeq : %i", nbSeq);
         RETURN_ERROR_IF(nbSeq, corruption_detected, "");
-        RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, corruption_detected, "");
+        DEBUGLOG(5, "bitStream : start=%p, ptr=%p, bitsConsumed=%u", seqState.DStream.start, seqState.DStream.ptr, seqState.DStream.bitsConsumed);
+        RETURN_ERROR_IF(!BIT_endOfDStream(&seqState.DStream), corruption_detected, "");
         /* save reps for next block */
         { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
     }
 
     /* last literal segment */
-    if (dctx->litBufferLocation == ZSTD_split)  /* split hasn't been reached yet, first get dst then copy litExtraBuffer */
-    {
-        size_t const lastLLSize = litBufferEnd - litPtr;
+    if (dctx->litBufferLocation == ZSTD_split) {
+        /* split hasn't been reached yet, first get dst then copy litExtraBuffer */
+        size_t const lastLLSize = (size_t)(litBufferEnd - litPtr);
+        DEBUGLOG(6, "copy last literals from segment : %u", (U32)lastLLSize);
         RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
         if (op != NULL) {
             ZSTD_memmove(op, litPtr, lastLLSize);
···
         litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
         dctx->litBufferLocation = ZSTD_not_in_dst;
     }
-    {   size_t const lastLLSize = litBufferEnd - litPtr;
+    /* copy last literals from internal buffer */
+    {   size_t const lastLLSize = (size_t)(litBufferEnd - litPtr);
+        DEBUGLOG(6, "copy last literals from internal buffer : %u", (U32)lastLLSize);
         RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
         if (op != NULL) {
             ZSTD_memcpy(op, litPtr, lastLLSize);
             op += lastLLSize;
-        }
-    }
+    }   }
 
-    return op-ostart;
+    DEBUGLOG(6, "decoded block of size %u bytes", (U32)(op - ostart));
+    return (size_t)(op - ostart);
 }
 
 FORCE_INLINE_TEMPLATE size_t
···
 ZSTD_decompressSequences_body(ZSTD_DCtx* dctx,
     void* dst, size_t maxDstSize,
     const void* seqStart, size_t seqSize, int nbSeq,
-    const ZSTD_longOffset_e isLongOffset,
-    const int frame)
+    const ZSTD_longOffset_e isLongOffset)
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
     BYTE* const ostart = (BYTE*)dst;
-    BYTE* const oend = dctx->litBufferLocation == ZSTD_not_in_dst ? ostart + maxDstSize : dctx->litBuffer;
+    BYTE* const oend = dctx->litBufferLocation == ZSTD_not_in_dst ? ZSTD_maybeNullPtrAdd(ostart, maxDstSize) : dctx->litBuffer;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
     const BYTE* const litEnd = litPtr + dctx->litSize;
     const BYTE* const prefixStart = (const BYTE*)(dctx->prefixStart);
     const BYTE* const vBase = (const BYTE*)(dctx->virtualStart);
     const BYTE* const dictEnd = (const BYTE*)(dctx->dictEnd);
-    DEBUGLOG(5, "ZSTD_decompressSequences_body");
-    (void)frame;
+    DEBUGLOG(5, "ZSTD_decompressSequences_body: nbSeq = %d", nbSeq);
 
     /* Regen sequences */
     if (nbSeq) {
···
         ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
         ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
         assert(dst != NULL);
-
-        ZSTD_STATIC_ASSERT(
-            BIT_DStream_unfinished < BIT_DStream_completed &&
-            BIT_DStream_endOfBuffer < BIT_DStream_completed &&
-            BIT_DStream_completed < BIT_DStream_overflow);
 
 #if defined(__x86_64__)
             __asm__(".p2align 6");
···
 #  endif
 #endif
 
-        for ( ; ; ) {
-            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+        for ( ; nbSeq ; nbSeq--) {
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, nbSeq==1);
             size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litEnd, prefixStart, vBase, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
             assert(!ZSTD_isError(oneSeqSize));
-            if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
+            ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
 #endif
             if (UNLIKELY(ZSTD_isError(oneSeqSize)))
                 return oneSeqSize;
             DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
             op += oneSeqSize;
-            if (UNLIKELY(!--nbSeq))
-                break;
-            BIT_reloadDStream(&(seqState.DStream));
         }
 
         /* check if reached exact end */
-        DEBUGLOG(5, "ZSTD_decompressSequences_body: after decode loop, remaining nbSeq : %i", nbSeq);
-        RETURN_ERROR_IF(nbSeq, corruption_detected, "");
-        RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, corruption_detected, "");
+        assert(nbSeq == 0);
+        RETURN_ERROR_IF(!BIT_endOfDStream(&seqState.DStream), corruption_detected, "");
         /* save reps for next block */
         { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
     }
 
     /* last literal segment */
-    {   size_t const lastLLSize = litEnd - litPtr;
+    {   size_t const lastLLSize = (size_t)(litEnd - litPtr);
+        DEBUGLOG(6, "copy last literals : %u", (U32)lastLLSize);
         RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
         if (op != NULL) {
             ZSTD_memcpy(op, litPtr, lastLLSize);
             op += lastLLSize;
-        }
-    }
+    }   }
 
-    return op-ostart;
+    DEBUGLOG(6, "decoded block of size %u bytes", (U32)(op - ostart));
+    return (size_t)(op - ostart);
 }
 
 static size_t
 ZSTD_decompressSequences_default(ZSTD_DCtx* dctx,
                                  void* dst, size_t maxDstSize,
                            const void* seqStart, size_t seqSize, int nbSeq,
-                           const ZSTD_longOffset_e isLongOffset,
-                           const int frame)
+                           const ZSTD_longOffset_e isLongOffset)
 {
-    return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 
 static size_t
 ZSTD_decompressSequencesSplitLitBuffer_default(ZSTD_DCtx* dctx,
                                                void* dst, size_t maxDstSize,
                                          const void* seqStart, size_t seqSize, int nbSeq,
-                                         const ZSTD_longOffset_e isLongOffset,
-                                         const int frame)
+                                         const ZSTD_longOffset_e isLongOffset)
 {
-    return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
 
-FORCE_INLINE_TEMPLATE size_t
-ZSTD_prefetchMatch(size_t prefetchPos, seq_t const sequence,
+FORCE_INLINE_TEMPLATE
+
+size_t ZSTD_prefetchMatch(size_t prefetchPos, seq_t const sequence,
                    const BYTE* const prefixStart, const BYTE* const dictEnd)
 {
     prefetchPos += sequence.litLength;
     {   const BYTE* const matchBase = (sequence.offset > prefetchPos) ? dictEnd : prefixStart;
-        const BYTE* const match = matchBase + prefetchPos - sequence.offset; /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
-                                                                              * No consequence though : memory address is only used for prefetching, not for dereferencing */
+        /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
+         * No consequence though : memory address is only used for prefetching, not for dereferencing */
+        const BYTE* const match = ZSTD_wrappedPtrSub(ZSTD_wrappedPtrAdd(matchBase, prefetchPos), sequence.offset);
         PREFETCH_L1(match); PREFETCH_L1(match+CACHELINE_SIZE);   /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
     }
     return prefetchPos + sequence.matchLength;
···
                                ZSTD_DCtx* dctx,
                                void* dst, size_t maxDstSize,
                          const void* seqStart, size_t seqSize, int nbSeq,
-                         const ZSTD_longOffset_e isLongOffset,
-                         const int frame)
+                         const ZSTD_longOffset_e isLongOffset)
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
     BYTE* const ostart = (BYTE*)dst;
-    BYTE* const oend = dctx->litBufferLocation == ZSTD_in_dst ? dctx->litBuffer : ostart + maxDstSize;
+    BYTE* const oend = dctx->litBufferLocation == ZSTD_in_dst ? dctx->litBuffer : ZSTD_maybeNullPtrAdd(ostart, maxDstSize);
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
     const BYTE* litBufferEnd = dctx->litBufferEnd;
     const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
     const BYTE* const dictStart = (const BYTE*) (dctx->virtualStart);
     const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
-    (void)frame;
 
     /* Regen sequences */
     if (nbSeq) {
···
         ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
 
         /* prepare in advance */
-        for (seqNb=0; (BIT_reloadDStream(&seqState.DStream) <= BIT_DStream_completed) && (seqNb<seqAdvance); seqNb++) {
-            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+        for (seqNb=0; seqNb<seqAdvance; seqNb++) {
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, seqNb == nbSeq-1);
             prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
             sequences[seqNb] = sequence;
         }
-        RETURN_ERROR_IF(seqNb<seqAdvance, corruption_detected, "");
 
         /* decompress without stomping litBuffer */
-        for (; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (seqNb < nbSeq); seqNb++) {
-            seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
-            size_t oneSeqSize;
+        for (; seqNb < nbSeq; seqNb++) {
+            seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset, seqNb == nbSeq-1);
 
-            if (dctx->litBufferLocation == ZSTD_split && litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength > dctx->litBufferEnd)
-            {
+            if (dctx->litBufferLocation == ZSTD_split && litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength > dctx->litBufferEnd) {
                 /* lit buffer is reaching split point, empty out the first buffer and transition to litExtraBuffer */
                 const size_t leftoverLit = dctx->litBufferEnd - litPtr;
                 if (leftoverLit)
···
                 litPtr = dctx->litExtraBuffer;
                 litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
                 dctx->litBufferLocation = ZSTD_not_in_dst;
-                oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
+                {   size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
-                assert(!ZSTD_isError(oneSeqSize));
-                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
+                    assert(!ZSTD_isError(oneSeqSize));
+                    ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
 #endif
-                if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+                    if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
 
-                prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
-                sequences[seqNb & STORED_SEQS_MASK] = sequence;
-                op += oneSeqSize;
-            }
+                    prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
+                    sequences[seqNb & STORED_SEQS_MASK] = sequence;
+                    op += oneSeqSize;
+            }   }
             else
             {
                 /* lit buffer is either wholly contained in first or second split, or not split at all*/
-                oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
+                size_t const oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
                     ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength - WILDCOPY_OVERLENGTH, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
                     ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                 assert(!ZSTD_isError(oneSeqSize));
-                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
+                ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
 #endif
                 if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
 
···
                 op += oneSeqSize;
             }
         }
-        RETURN_ERROR_IF(seqNb<nbSeq, corruption_detected, "");
+        RETURN_ERROR_IF(!BIT_endOfDStream(&seqState.DStream), corruption_detected, "");
 
         /* finish queue */
         seqNb -= seqAdvance;
         for ( ; seqNb<nbSeq ; seqNb++) {
             seq_t *sequence = &(sequences[seqNb&STORED_SEQS_MASK]);
-            if (dctx->litBufferLocation == ZSTD_split && litPtr + sequence->litLength > dctx->litBufferEnd)
-            {
+            if (dctx->litBufferLocation == ZSTD_split && litPtr + sequence->litLength > dctx->litBufferEnd) {
                 const size_t leftoverLit = dctx->litBufferEnd - litPtr;
-                if (leftoverLit)
-                {
+                if (leftoverLit) {
                     RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
                     ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
                     sequence->litLength -= leftoverLit;
···
                 litPtr = dctx->litExtraBuffer;
                 litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
                 dctx->litBufferLocation = ZSTD_not_in_dst;
-                {
-                    size_t const oneSeqSize = ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
+                {   size_t const oneSeqSize = ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                     assert(!ZSTD_isError(oneSeqSize));
-                    if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
+                    ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
 #endif
                     if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
                     op += oneSeqSize;
···
                     ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                 assert(!ZSTD_isError(oneSeqSize));
-                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
+                ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
 #endif
                 if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
                 op += oneSeqSize;
···
     }
 
     /* last literal segment */
-    if (dctx->litBufferLocation == ZSTD_split)  /* first deplete literal buffer in dst, then copy litExtraBuffer */
-    {
+    if (dctx->litBufferLocation == ZSTD_split) { /* first deplete literal buffer in dst, then copy litExtraBuffer */
         size_t const lastLLSize = litBufferEnd - litPtr;
         RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
         if (op != NULL) {
···
         }
     }
 
-    return op-ostart;
+    return (size_t)(op - ostart);
 }
 
 static size_t
 ZSTD_decompressSequencesLong_default(ZSTD_DCtx* dctx,
                                  void* dst, size_t maxDstSize,
                            const void* seqStart, size_t seqSize, int nbSeq,
-                           const ZSTD_longOffset_e isLongOffset,
-                           const int frame)
+                           const ZSTD_longOffset_e isLongOffset)
 {
-    return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */
 
···
 ZSTD_decompressSequences_bmi2(ZSTD_DCtx* dctx,
                                  void* dst, size_t maxDstSize,
                            const void* seqStart, size_t seqSize, int nbSeq,
-                           const ZSTD_longOffset_e isLongOffset,
-                           const int frame)
+                           const ZSTD_longOffset_e isLongOffset)
 {
-    return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 static BMI2_TARGET_ATTRIBUTE size_t
 DONT_VECTORIZE
 ZSTD_decompressSequencesSplitLitBuffer_bmi2(ZSTD_DCtx* dctx,
                                  void* dst, size_t maxDstSize,
                            const void* seqStart, size_t seqSize, int nbSeq,
-                           const ZSTD_longOffset_e isLongOffset,
-                           const int frame)
+                           const ZSTD_longOffset_e isLongOffset)
 {
-    return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
 
···
 ZSTD_decompressSequencesLong_bmi2(ZSTD_DCtx* dctx,
                                  void* dst, size_t maxDstSize,
                            const void* seqStart, size_t seqSize, int nbSeq,
-                           const ZSTD_longOffset_e isLongOffset,
-                           const int frame)
+                           const ZSTD_longOffset_e isLongOffset)
 {
-    return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */
 
 #endif /* DYNAMIC_BMI2 */
 
-typedef size_t (*ZSTD_decompressSequences_t)(
-                            ZSTD_DCtx* dctx,
-                            void* dst, size_t maxDstSize,
-                            const void* seqStart, size_t seqSize, int nbSeq,
-                            const ZSTD_longOffset_e isLongOffset,
-                            const int frame);
-
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
 static size_t
 ZSTD_decompressSequences(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
                    const void* seqStart, size_t seqSize, int nbSeq,
-                   const ZSTD_longOffset_e isLongOffset,
-                   const int frame)
+                   const ZSTD_longOffset_e isLongOffset)
 {
     DEBUGLOG(5, "ZSTD_decompressSequences");
 #if DYNAMIC_BMI2
     if (ZSTD_DCtx_get_bmi2(dctx)) {
-        return ZSTD_decompressSequences_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+        return ZSTD_decompressSequences_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
     }
 #endif
-    return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 static size_t
 ZSTD_decompressSequencesSplitLitBuffer(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
                                  const void* seqStart, size_t seqSize, int nbSeq,
-                                 const ZSTD_longOffset_e isLongOffset,
-                                 const int frame)
+                                 const ZSTD_longOffset_e isLongOffset)
 {
     DEBUGLOG(5, "ZSTD_decompressSequencesSplitLitBuffer");
 #if DYNAMIC_BMI2
     if (ZSTD_DCtx_get_bmi2(dctx)) {
-        return ZSTD_decompressSequencesSplitLitBuffer_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+        return ZSTD_decompressSequencesSplitLitBuffer_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
     }
 #endif
-    return ZSTD_decompressSequencesSplitLitBuffer_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequencesSplitLitBuffer_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
 
···
 ZSTD_decompressSequencesLong(ZSTD_DCtx* dctx,
                              void* dst, size_t maxDstSize,
                              const void* seqStart, size_t seqSize, int nbSeq,
-                             const ZSTD_longOffset_e isLongOffset,
-                             const int frame)
+                             const ZSTD_longOffset_e isLongOffset)
 {
     DEBUGLOG(5, "ZSTD_decompressSequencesLong");
 #if DYNAMIC_BMI2
     if (ZSTD_DCtx_get_bmi2(dctx)) {
-        return ZSTD_decompressSequencesLong_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+        return ZSTD_decompressSequencesLong_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
     }
 #endif
-  return ZSTD_decompressSequencesLong_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+  return ZSTD_decompressSequencesLong_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
 }
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */
 
 
+/*
+ * @returns The total size of the history referenceable by zstd, including
+ * both the prefix and the extDict. At @p op any offset larger than this
+ * is invalid.
+ */
+static size_t ZSTD_totalHistorySize(BYTE* op, BYTE const* virtualStart)
+{
+    return (size_t)(op - virtualStart);
+}
 
-#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
-    !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
-/* ZSTD_getLongOffsetsShare() :
+typedef struct {
+    unsigned longOffsetShare;
+    unsigned maxNbAdditionalBits;
+} ZSTD_OffsetInfo;
+
+/* ZSTD_getOffsetInfo() :
  * condition : offTable must be valid
  * @return : "share" of long offsets (arbitrarily defined as > (1<<23))
- *           compared to maximum possible of (1<<OffFSELog) */
-static unsigned
-ZSTD_getLongOffsetsShare(const ZSTD_seqSymbol* offTable)
+ *           compared to maximum possible of (1<<OffFSELog),
+ *           as well as the maximum number additional bits required.
+ */
+static ZSTD_OffsetInfo
+ZSTD_getOffsetInfo(const ZSTD_seqSymbol* offTable, int nbSeq)
 {
-    const void* ptr = offTable;
-    U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog;
-    const ZSTD_seqSymbol* table = offTable + 1;
-    U32 const max = 1 << tableLog;
-    U32 u, total = 0;
-    DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog);
+    ZSTD_OffsetInfo info = {0, 0};
+    /* If nbSeq == 0, then the offTable is uninitialized, but we have
+     * no sequences, so both values should be 0.
+     */
+    if (nbSeq != 0) {
+        const void* ptr = offTable;
+        U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog;
+        const ZSTD_seqSymbol* table = offTable + 1;
+        U32 const max = 1 << tableLog;
+        U32 u;
+        DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog);
 
-    assert(max <= (1 << OffFSELog));  /* max not too large */
-    for (u=0; u<max; u++) {
-        if (table[u].nbAdditionalBits > 22) total += 1;
+        assert(max <= (1 << OffFSELog));  /* max not too large */
+        for (u=0; u<max; u++) {
+            info.maxNbAdditionalBits = MAX(info.maxNbAdditionalBits, table[u].nbAdditionalBits);
+            if (table[u].nbAdditionalBits > 22) info.longOffsetShare += 1;
+        }
+
+        assert(tableLog <= OffFSELog);
+        info.longOffsetShare <<= (OffFSELog - tableLog);  /* scale to OffFSELog */
     }
 
-    assert(tableLog <= OffFSELog);
-    total <<= (OffFSELog - tableLog);  /* scale to OffFSELog */
-
-    return total;
+    return info;
 }
-#endif
+
+/*
+ * @returns The maximum offset we can decode in one read of our bitstream, without
+ * reloading more bits in the middle of the offset bits read. Any offsets larger
+ * than this must use the long offset decoder.
+ */
+static size_t ZSTD_maxShortOffset(void)
+{
+    if (MEM_64bits()) {
+        /* We can decode any offset without reloading bits.
+         * This might change if the max window size grows.
+         */
+        ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX <= 31);
+        return (size_t)-1;
+    } else {
+        /* The maximum offBase is (1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1.
+         * This offBase would require STREAM_ACCUMULATOR_MIN extra bits.
+         * Then we have to subtract ZSTD_REP_NUM to get the maximum possible offset.
+         */
+        size_t const maxOffbase = ((size_t)1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1;
+        size_t const maxOffset = maxOffbase - ZSTD_REP_NUM;
+        assert(ZSTD_highbit32((U32)maxOffbase) == STREAM_ACCUMULATOR_MIN);
+        return maxOffset;
+    }
+}
 
 size_t
 ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                               void* dst, size_t dstCapacity,
-                        const void* src, size_t srcSize, const int frame, const streaming_operation streaming)
+                        const void* src, size_t srcSize, const streaming_operation streaming)
 {   /* blockType == blockCompressed */
     const BYTE* ip = (const BYTE*)src;
-    /* isLongOffset must be true if there are long offsets.
-     * Offsets are long if they are larger than 2^STREAM_ACCUMULATOR_MIN.
-     * We don't expect that to be the case in 64-bit mode.
-     * In block mode, window size is not known, so we have to be conservative.
-     * (note: but it could be evaluated from current-lowLimit)
-     */
-    ZSTD_longOffset_e const isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (!frame || (dctx->fParams.windowSize > (1ULL << STREAM_ACCUMULATOR_MIN))));
-    DEBUGLOG(5, "ZSTD_decompressBlock_internal (size : %u)", (U32)srcSize);
+    DEBUGLOG(5, "ZSTD_decompressBlock_internal (cSize : %u)", (unsigned)srcSize);
 
-    RETURN_ERROR_IF(srcSize >= ZSTD_BLOCKSIZE_MAX, srcSize_wrong, "");
+    /* Note : the wording of the specification
+     * allows compressed block to be sized exactly ZSTD_blockSizeMax(dctx).
+     * This generally does not happen, as it makes little sense,
+     * since an uncompressed block would feature same size and have no decompression cost.
+     * Also, note that decoder from reference libzstd before < v1.5.4
+     * would consider this edge case as an error.
+     * As a consequence, avoid generating compressed blocks of size ZSTD_blockSizeMax(dctx)
+     * for broader compatibility with the deployed ecosystem of zstd decoders */
+    RETURN_ERROR_IF(srcSize > ZSTD_blockSizeMax(dctx), srcSize_wrong, "");
 
     /* Decode literals section */
     {   size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, streaming);
-        DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : %u", (U32)litCSize);
+        DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : cSize=%u, nbLiterals=%zu", (U32)litCSize, dctx->litSize);
         if (ZSTD_isError(litCSize)) return litCSize;
         ip += litCSize;
         srcSize -= litCSize;
···
 
     /* Build Decoding Tables */
     {
+        /* Compute the maximum block size, which must also work when !frame and fParams are unset.
+         * Additionally, take the min with dstCapacity to ensure that the totalHistorySize fits in a size_t.
+         */
+        size_t const blockSizeMax = MIN(dstCapacity, ZSTD_blockSizeMax(dctx));
+        size_t const totalHistorySize = ZSTD_totalHistorySize(ZSTD_maybeNullPtrAdd((BYTE*)dst, blockSizeMax), (BYTE const*)dctx->virtualStart);
+        /* isLongOffset must be true if there are long offsets.
+         * Offsets are long if they are larger than ZSTD_maxShortOffset().
+         * We don't expect that to be the case in 64-bit mode.
+         *
+         * We check here to see if our history is large enough to allow long offsets.
+         * If it isn't, then we can't possible have (valid) long offsets. If the offset
+         * is invalid, then it is okay to read it incorrectly.
+         *
+         * If isLongOffsets is true, then we will later check our decoding table to see
+         * if it is even possible to generate long offsets.
+         */
+        ZSTD_longOffset_e isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (totalHistorySize > ZSTD_maxShortOffset()));
         /* These macros control at build-time which decompressor implementation
          * we use. If neither is defined, we do some inspection and dispatch at
          * runtime.
···
 #if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
     !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
         int usePrefetchDecoder = dctx->ddictIsCold;
+#else
+        /* Set to 1 to avoid computing offset info if we don't need to.
+         * Otherwise this value is ignored.
+         */
+        int usePrefetchDecoder = 1;
 #endif
         int nbSeq;
         size_t const seqHSize = ZSTD_decodeSeqHeaders(dctx, &nbSeq, ip, srcSize);
···
         ip += seqHSize;
         srcSize -= seqHSize;
 
-        RETURN_ERROR_IF(dst == NULL && nbSeq > 0, dstSize_tooSmall, "NULL not handled");
+        RETURN_ERROR_IF((dst == NULL || dstCapacity == 0) && nbSeq > 0, dstSize_tooSmall, "NULL not handled");
+        RETURN_ERROR_IF(MEM_64bits() && sizeof(size_t) == sizeof(void*) && (size_t)(-1) - (size_t)dst < (size_t)(1 << 20), dstSize_tooSmall,
+                "invalid dst");
 
-#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
-    !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
-        if ( !usePrefetchDecoder
-          && (!frame || (dctx->fParams.windowSize > (1<<24)))
-          && (nbSeq>ADVANCED_SEQS) ) {  /* could probably use a larger nbSeq limit */
-            U32 const shareLongOffsets = ZSTD_getLongOffsetsShare(dctx->OFTptr);
-            U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
-            usePrefetchDecoder = (shareLongOffsets >= minShare);
+        /* If we could potentially have long offsets, or we might want to use the prefetch decoder,
+         * compute information about the share of long offsets, and the maximum nbAdditionalBits.
+         * NOTE: could probably use a larger nbSeq limit
+         */
+        if (isLongOffset || (!usePrefetchDecoder && (totalHistorySize > (1u << 24)) && (nbSeq > 8))) {
+            ZSTD_OffsetInfo const info = ZSTD_getOffsetInfo(dctx->OFTptr, nbSeq);
+            if (isLongOffset && info.maxNbAdditionalBits <= STREAM_ACCUMULATOR_MIN) {
+                /* If isLongOffset, but the maximum number of additional bits that we see in our table is small
+                 * enough, then we know it is impossible to have too long an offset in this block, so we can
+                 * use the regular offset decoder.
+                 */
+                isLongOffset = ZSTD_lo_isRegularOffset;
+            }
+            if (!usePrefetchDecoder) {
+                U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
+                usePrefetchDecoder = (info.longOffsetShare >= minShare);
+            }
         }
-#endif
 
         dctx->ddictIsCold = 0;
 
 #if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
     !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
-        if (usePrefetchDecoder)
+        if (usePrefetchDecoder) {
+#else
+        (void)usePrefetchDecoder;
+        {
 #endif
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
-            return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
+            return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset);
 #endif
+        }
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
         /* else */
         if (dctx->litBufferLocation == ZSTD_split)
-            return ZSTD_decompressSequencesSplitLitBuffer(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
+            return ZSTD_decompressSequencesSplitLitBuffer(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset);
         else
-            return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
+            return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset);
 #endif
     }
 }
 
 
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize)
 {
     if (dst != dctx->previousDstEnd && dstSize > 0) {   /* not contiguous */
···
 }
 
 
+size_t ZSTD_decompressBlock_deprecated(ZSTD_DCtx* dctx,
+                                       void* dst, size_t dstCapacity,
+                                 const void* src, size_t srcSize)
+{
+    size_t dSize;
+    dctx->isFrameDecompression = 0;
+    ZSTD_checkContinuity(dctx, dst, dstCapacity);
+    dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, not_streaming);
+    FORWARD_IF_ERROR(dSize, "");
+    dctx->previousDstEnd = (char*)dst + dSize;
+    return dSize;
+}
+
+
+/* NOTE: Must just wrap ZSTD_decompressBlock_deprecated() */
 size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx,
                             void* dst, size_t dstCapacity,
                       const void* src, size_t srcSize)
 {
-    size_t dSize;
-    ZSTD_checkContinuity(dctx, dst, dstCapacity);
-    dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0, not_streaming);
-    dctx->previousDstEnd = (char*)dst + dSize;
-    return dSize;
+    return ZSTD_decompressBlock_deprecated(dctx, dst, dstCapacity, src, srcSize);
 }

lib/zstd/decompress/zstd_decompress_block.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
  */
 size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                                void* dst, size_t dstCapacity,
-                         const void* src, size_t srcSize, const int frame, const streaming_operation streaming);
+                         const void* src, size_t srcSize, const streaming_operation streaming);
 
 /* ZSTD_buildFSETable() :
  * generate FSE decoding table for one symbol (ll, ml or off)
···
              const U32* baseValue, const U8* nbAdditionalBits,
                    unsigned tableLog, void* wksp, size_t wkspSize,
                    int bmi2);
+
+/* Internal definition of ZSTD_decompressBlock() to avoid deprecation warnings. */
+size_t ZSTD_decompressBlock_deprecated(ZSTD_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize);
 
 
 #endif /* ZSTD_DEC_BLOCK_H */

lib/zstd/decompress/zstd_decompress_internal.h

···
+/* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 
 #define ZSTD_BUILD_FSE_TABLE_WKSP_SIZE (sizeof(S16) * (MaxSeq + 1) + (1u << MaxFSELog) + sizeof(U64))
 #define ZSTD_BUILD_FSE_TABLE_WKSP_SIZE_U32 ((ZSTD_BUILD_FSE_TABLE_WKSP_SIZE + sizeof(U32) - 1) / sizeof(U32))
+#define ZSTD_HUFFDTABLE_CAPACITY_LOG 12
 
 typedef struct {
     ZSTD_seqSymbol LLTable[SEQSYMBOL_TABLE_SIZE(LLFSELog)];    /* Note : Space reserved for FSE Tables */
     ZSTD_seqSymbol OFTable[SEQSYMBOL_TABLE_SIZE(OffFSELog)];   /* is also used as temporary workspace while building hufTable during DDict creation */
     ZSTD_seqSymbol MLTable[SEQSYMBOL_TABLE_SIZE(MLFSELog)];    /* and therefore must be at least HUF_DECOMPRESS_WORKSPACE_SIZE large */
-    HUF_DTable hufTable[HUF_DTABLE_SIZE(HufLog)];  /* can accommodate HUF_decompress4X */
+    HUF_DTable hufTable[HUF_DTABLE_SIZE(ZSTD_HUFFDTABLE_CAPACITY_LOG)];  /* can accommodate HUF_decompress4X */
     U32 rep[ZSTD_REP_NUM];
     U32 workspace[ZSTD_BUILD_FSE_TABLE_WKSP_SIZE_U32];
 } ZSTD_entropyDTables_t;
···
     const void* virtualStart;     /* virtual start of previous segment if it was just before current one */
     const void* dictEnd;          /* end of previous segment */
     size_t expected;
-    ZSTD_frameHeader fParams;
+    ZSTD_FrameHeader fParams;
     U64 processedCSize;
     U64 decodedSize;
     blockType_e bType;            /* used in ZSTD_decompressContinue(), store blockType between block header decoding and block decompression stages */
···
     size_t litSize;
     size_t rleSize;
     size_t staticSize;
-#if DYNAMIC_BMI2 != 0
+    int isFrameDecompression;
+#if DYNAMIC_BMI2
     int bmi2;                     /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
 #endif
 
···
     ZSTD_dictUses_e dictUses;
     ZSTD_DDictHashSet* ddictSet;                    /* Hash set for multiple ddicts */
     ZSTD_refMultipleDDicts_e refMultipleDDicts;     /* User specified: if == 1, will allow references to multiple DDicts. Default == 0 (disabled) */
+    int disableHufAsm;
+    int maxBlockSizeParam;
 
     /* streaming */
     ZSTD_dStreamStage streamStage;
···
 };  /* typedef'd to ZSTD_DCtx within "zstd.h" */
 
 MEM_STATIC int ZSTD_DCtx_get_bmi2(const struct ZSTD_DCtx_s *dctx) {
-#if DYNAMIC_BMI2 != 0
-	return dctx->bmi2;
+#if DYNAMIC_BMI2
+    return dctx->bmi2;
 #else
     (void)dctx;
-	return 0;
+    return 0;
 #endif
 }
 

lib/zstd/decompress_sources.h

···
 /* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */
 /*
- * Copyright (c) Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

lib/zstd/zstd_common_module.c

···
 // SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 EXPORT_SYMBOL_GPL(ZSTD_isError);
 EXPORT_SYMBOL_GPL(ZSTD_getErrorName);
 EXPORT_SYMBOL_GPL(ZSTD_getErrorCode);
-EXPORT_SYMBOL_GPL(ZSTD_customMalloc);
-EXPORT_SYMBOL_GPL(ZSTD_customCalloc);
-EXPORT_SYMBOL_GPL(ZSTD_customFree);
 
 MODULE_LICENSE("Dual BSD/GPL");
 MODULE_DESCRIPTION("Zstd Common");

lib/zstd/zstd_compress_module.c

···
 // SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 
 #include "common/zstd_deps.h"
 #include "common/zstd_internal.h"
+#include "compress/zstd_compress_internal.h"
 
 #define ZSTD_FORWARD_IF_ERR(ret)            \
 	do {                                \
···
 }
 EXPORT_SYMBOL(zstd_get_cparams);
 
+size_t zstd_cctx_set_param(zstd_cctx *cctx, ZSTD_cParameter param, int value)
+{
+	return ZSTD_CCtx_setParameter(cctx, param, value);
+}
+EXPORT_SYMBOL(zstd_cctx_set_param);
+
 size_t zstd_cctx_workspace_bound(const zstd_compression_parameters *cparams)
 {
 	return ZSTD_estimateCCtxSize_usingCParams(*cparams);
 }
 EXPORT_SYMBOL(zstd_cctx_workspace_bound);
+
+// Used by zstd_cctx_workspace_bound_with_ext_seq_prod()
+static size_t dummy_external_sequence_producer(
+	void *sequenceProducerState,
+	ZSTD_Sequence *outSeqs, size_t outSeqsCapacity,
+	const void *src, size_t srcSize,
+	const void *dict, size_t dictSize,
+	int compressionLevel,
+	size_t windowSize)
+{
+	(void)sequenceProducerState;
+	(void)outSeqs; (void)outSeqsCapacity;
+	(void)src; (void)srcSize;
+	(void)dict; (void)dictSize;
+	(void)compressionLevel;
+	(void)windowSize;
+	return ZSTD_SEQUENCE_PRODUCER_ERROR;
+}
+
+static void init_cctx_params_from_compress_params(
+	ZSTD_CCtx_params *cctx_params,
+	const zstd_compression_parameters *compress_params)
+{
+	ZSTD_parameters zstd_params;
+	memset(&zstd_params, 0, sizeof(zstd_params));
+	zstd_params.cParams = *compress_params;
+	ZSTD_CCtxParams_init_advanced(cctx_params, zstd_params);
+}
+
+size_t zstd_cctx_workspace_bound_with_ext_seq_prod(const zstd_compression_parameters *compress_params)
+{
+	ZSTD_CCtx_params cctx_params;
+	init_cctx_params_from_compress_params(&cctx_params, compress_params);
+	ZSTD_CCtxParams_registerSequenceProducer(&cctx_params, NULL, dummy_external_sequence_producer);
+	return ZSTD_estimateCCtxSize_usingCCtxParams(&cctx_params);
+}
+EXPORT_SYMBOL(zstd_cctx_workspace_bound_with_ext_seq_prod);
+
+size_t zstd_cstream_workspace_bound_with_ext_seq_prod(const zstd_compression_parameters *compress_params)
+{
+	ZSTD_CCtx_params cctx_params;
+	init_cctx_params_from_compress_params(&cctx_params, compress_params);
+	ZSTD_CCtxParams_registerSequenceProducer(&cctx_params, NULL, dummy_external_sequence_producer);
+	return ZSTD_estimateCStreamSize_usingCCtxParams(&cctx_params);
+}
+EXPORT_SYMBOL(zstd_cstream_workspace_bound_with_ext_seq_prod);
 
 zstd_cctx *zstd_init_cctx(void *workspace, size_t workspace_size)
 {
···
 	return ZSTD_endStream(cstream, output);
 }
 EXPORT_SYMBOL(zstd_end_stream);
+
+void zstd_register_sequence_producer(
+  zstd_cctx *cctx,
+  void* sequence_producer_state,
+  zstd_sequence_producer_f sequence_producer
+) {
+	ZSTD_registerSequenceProducer(cctx, sequence_producer_state, sequence_producer);
+}
+EXPORT_SYMBOL(zstd_register_sequence_producer);
+
+size_t zstd_compress_sequences_and_literals(zstd_cctx *cctx, void* dst, size_t dst_capacity,
+					    const zstd_sequence *in_seqs, size_t in_seqs_size,
+					    const void* literals, size_t lit_size, size_t lit_capacity,
+					    size_t decompressed_size)
+{
+	return ZSTD_compressSequencesAndLiterals(cctx, dst, dst_capacity, in_seqs,
+						 in_seqs_size, literals, lit_size,
+						 lit_capacity, decompressed_size);
+}
+EXPORT_SYMBOL(zstd_compress_sequences_and_literals);
 
 MODULE_LICENSE("Dual BSD/GPL");
 MODULE_DESCRIPTION("Zstd Compressor");

lib/zstd/zstd_decompress_module.c

···
 // SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause
 /*
- * Copyright (c) Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
···
 
 size_t zstd_reset_dstream(zstd_dstream *dstream)
 {
-	return ZSTD_resetDStream(dstream);
+	return ZSTD_DCtx_reset(dstream, ZSTD_reset_session_only);
 }
 EXPORT_SYMBOL(zstd_reset_dstream);