External matchfinder API (#3333)

* First building commit with sample matchfinder

* Set up ZSTD_externalMatchCtx struct

* move seqBuffer to ZSTD_Sequence*

* support non-contiguous dictionary

* clean up parens

* add clearExternalMatchfinder, handle allocation errors

* Add useExternalMatchfinder cParam

* validate useExternalMatchfinder cParam

* Disable LDM + external matchfinder

* Check for static CCtx

* Validate mState and mStateDestructor

* Improve LDM check to cover both branches

* Error API with optional fallback

* handle RLE properly for external matchfinder

* nit

* Move to a CDict-like model for resource ownership

* Add hidden useExternalMatchfinder bool to CCtx_params_s

* Eliminate malloc, move to cwksp allocation

* Handle CCtx reset properly

* Ensure seqStore has enough space for external sequences

* fix capitalization

* Add DEBUGLOG statements

* Add compressionLevel param to matchfinder API

* fix c99 issues and add a param combination error code

* nits

* Test external matchfinder API

* C90 compat for simpleExternalMatchFinder

* Fix some @nocommits and an ASAN bug

* nit

* nit

* nits

* forward declare copySequencesToSeqStore functions in zstd_compress_internal.h

* nit

* nit

* nits

* Update copyright headers

* Fix CMake zstreamtest build

* Fix copyright headers (again)

* typo

* Add externalMatchfinder demo program to make contrib

* Reduce memory consumption for small blockSize

* ZSTD_postProcessExternalMatchFinderResult nits

* test sum(matchlen) + sum(litlen) == srcSize in debug builds

* refExternalMatchFinder -> registerExternalMatchFinder

* C90 nit

* zstreamtest nits

* contrib nits

* contrib nits

* allow block splitter + external matchfinder, refactor

* add windowSize param

* add contrib/externalMatchfinder/README.md

* docs

* go back to old RLE heuristic because of the first block issue

* fix initializer element is not a constant expression

* ref contrib from zstd.h

* extremely pedantic compiler warning fix, meson fix, typo fix

* Additional docs on API limitations

* minor nits

* Refactor maxNbSeq calculation into a helper function

* Fix copyright

Makefile

@@ -123,6 +123,7 @@ contrib: lib
 	$(MAKE) -C contrib/seekable_format/examples all
 	$(MAKE) -C contrib/seekable_format/tests test
 	$(MAKE) -C contrib/largeNbDicts all
+	$(MAKE) -C contrib/externalMatchfinder all
 	cd build/single_file_libs/ ; ./build_decoder_test.sh
 	cd build/single_file_libs/ ; ./build_library_test.sh
 
@@ -142,6 +143,7 @@ clean:
 	$(Q)$(MAKE) -C contrib/seekable_format/examples $@ > $(VOID)
 	$(Q)$(MAKE) -C contrib/seekable_format/tests $@ > $(VOID)
 	$(Q)$(MAKE) -C contrib/largeNbDicts $@ > $(VOID)
+	$(Q)$(MAKE) -C contrib/externalMatchfinder $@ > $(VOID)
 	$(Q)$(RM) zstd$(EXT) zstdmt$(EXT) tmp*
 	$(Q)$(RM) -r lz4
 	@echo Cleaning completed

build/cmake/tests/CMakeLists.txt

@@ -81,7 +81,7 @@ add_test(NAME fuzzer COMMAND fuzzer ${ZSTD_FUZZER_FLAGS})
 #
 # zstreamtest
 #
-add_executable(zstreamtest ${PROGRAMS_DIR}/datagen.c ${PROGRAMS_DIR}/util.c ${PROGRAMS_DIR}/timefn.c ${TESTS_DIR}/seqgen.c ${TESTS_DIR}/zstreamtest.c)
+add_executable(zstreamtest ${PROGRAMS_DIR}/datagen.c ${PROGRAMS_DIR}/util.c ${PROGRAMS_DIR}/timefn.c ${TESTS_DIR}/seqgen.c ${TESTS_DIR}/zstreamtest.c ${TESTS_DIR}/external_matchfinder.c)
 if (NOT MSVC)
     target_compile_options(zstreamtest PRIVATE "-Wno-deprecated-declarations")
 endif()

build/meson/tests/meson.build

@@ -65,8 +65,10 @@ fuzzer = executable('fuzzer',
   dependencies: [ testcommon_dep, thread_dep ],
   install: false)
 
-zstreamtest_sources = [join_paths(zstd_rootdir, 'tests/seqgen.c'),
+zstreamtest_sources = [
-  join_paths(zstd_rootdir, 'tests/zstreamtest.c')]
+  join_paths(zstd_rootdir, 'tests/seqgen.c'),
+  join_paths(zstd_rootdir, 'tests/zstreamtest.c'),
+  join_paths(zstd_rootdir, 'tests/external_matchfinder.c')]
 zstreamtest = executable('zstreamtest',
   zstreamtest_sources,
   include_directories: test_includes,

contrib/externalMatchfinder/.gitignore

@@ -0,0 +1,2 @@
+# build artifacts
+externalMatchfinder

contrib/externalMatchfinder/Makefile

@@ -0,0 +1,40 @@
+# ################################################################
+# Copyright (c) Yann Collet, Meta Platforms, Inc.
+# 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).
+# ################################################################
+
+PROGDIR = ../../programs
+LIBDIR  = ../../lib
+
+LIBZSTD = $(LIBDIR)/libzstd.a
+
+CPPFLAGS+= -I$(LIBDIR) -I$(LIBDIR)/compress -I$(LIBDIR)/common
+
+CFLAGS  ?= -O3
+CFLAGS  += -std=gnu99
+DEBUGFLAGS= -Wall -Wextra -Wcast-qual -Wcast-align -Wshadow \
+            -Wstrict-aliasing=1 -Wswitch-enum \
+            -Wstrict-prototypes -Wundef -Wpointer-arith \
+            -Wvla -Wformat=2 -Winit-self -Wfloat-equal -Wwrite-strings \
+            -Wredundant-decls
+CFLAGS  += $(DEBUGFLAGS) $(MOREFLAGS)
+
+default: externalMatchfinder
+
+all: externalMatchfinder
+
+externalMatchfinder: matchfinder.c main.c $(LIBZSTD)
+	$(CC) $(CPPFLAGS) $(CFLAGS) $^ $(LDFLAGS) -o $@
+
+.PHONY: $(LIBZSTD)
+$(LIBZSTD):
+	$(MAKE) -C $(LIBDIR) libzstd.a CFLAGS="$(CFLAGS)"
+
+clean:
+	$(RM) *.o
+	$(MAKE) -C $(LIBDIR) clean > /dev/null
+	$(RM) externalMatchfinder

contrib/externalMatchfinder/README.md

@@ -0,0 +1,14 @@
+externalMatchfinder
+=====================
+
+`externalMatchfinder` is a test tool for the external matchfinder API.
+It demonstrates how to use the API to perform a simple round-trip test.
+
+A sample matchfinder is provided in matchfinder.c, but the user can swap
+this out with a different one if desired. The sample matchfinder implements
+LZ compression with a 1KB hashtable. Dictionary compression is not currently supported.
+
+Command line :
+```
+externalMatchfinder filename
+```

contrib/externalMatchfinder/main.c

@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) Yann Collet, Meta Platforms, Inc.
+ * 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 <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <assert.h>
+
+#define ZSTD_STATIC_LINKING_ONLY
+#include "zstd.h"
+#include "zstd_errors.h"
+#include "matchfinder.h" // simpleExternalMatchFinder
+
+#define CHECK(res)                                      \
+do {                                                    \
+    if (ZSTD_isError(res)) {                            \
+        printf("ERROR: %s\n", ZSTD_getErrorName(res));  \
+        return 1;                                       \
+    }                                                   \
+} while (0)                                             \
+
+int main(int argc, char *argv[]) {
+    if (argc != 2) {
+        printf("Usage: exampleMatchfinder <file>\n");
+        return 1;
+    }
+
+    ZSTD_CCtx* const zc = ZSTD_createCCtx();
+
+    int simpleExternalMatchState = 0xdeadbeef;
+
+    // Here is the crucial bit of code!
+    ZSTD_registerExternalMatchFinder(
+        zc,
+        &simpleExternalMatchState,
+        simpleExternalMatchFinder
+    );
+
+    {
+        size_t const res = ZSTD_CCtx_setParameter(zc, ZSTD_c_enableMatchFinderFallback, 1);
+        CHECK(res);
+    }
+
+    FILE *f = fopen(argv[1], "rb");
+    assert(f);
+    {
+        int const ret = fseek(f, 0, SEEK_END);
+        assert(ret == 0);
+    }
+    size_t const srcSize = ftell(f);
+    {
+        int const ret = fseek(f, 0, SEEK_SET);
+        assert(ret == 0);
+    }
+
+    char* const src = malloc(srcSize + 1);
+    assert(src);
+    {
+        size_t const ret = fread(src, srcSize, 1, f);
+        assert(ret == 1);
+        int const ret2 = fclose(f);
+        assert(ret2 == 0);
+    }
+
+    size_t const dstSize = ZSTD_compressBound(srcSize);
+    char* const dst = malloc(dstSize);
+    assert(dst);
+
+    size_t const cSize = ZSTD_compress2(zc, dst, dstSize, src, srcSize);
+    CHECK(cSize);
+
+    char* const val = malloc(srcSize);
+    assert(val);
+
+    {
+        size_t const res = ZSTD_decompress(val, srcSize, dst, cSize);
+        CHECK(res);
+    }
+
+    if (memcmp(src, val, srcSize) == 0) {
+        printf("Compression and decompression were successful!\n");
+        printf("Original size: %lu\n", srcSize);
+        printf("Compressed size: %lu\n", cSize);
+    } else {
+        printf("ERROR: input and validation buffers don't match!\n");
+        for (size_t i = 0; i < srcSize; i++) {
+            if (src[i] != val[i]) {
+                printf("First bad index: %zu\n", i);
+                break;
+            }
+        }
+        return 1;
+    }
+
+    ZSTD_freeCCtx(zc);
+    free(src);
+    free(dst);
+    free(val);
+    return 0;
+}

contrib/externalMatchfinder/matchfinder.c

@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) Yann Collet, Meta Platforms, Inc.
+ * 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 "zstd_compress_internal.h"
+#include "matchfinder.h"
+
+#define HSIZE 1024
+static U32 const HLOG = 10;
+static U32 const MLS = 4;
+static U32 const BADIDX = 0xffffffff;
+
+size_t simpleExternalMatchFinder(
+  void* externalMatchState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+) {
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    size_t seqCount = 0;
+    U32 hashTable[HSIZE];
+
+    (void)externalMatchState;
+    (void)dict;
+    (void)dictSize;
+    (void)outSeqsCapacity;
+    (void)compressionLevel;
+
+    {   int i;
+        for (i=0; i < HSIZE; i++) {
+            hashTable[i] = BADIDX;
+    }   }
+
+    while (ip + MLS < iend) {
+        size_t const hash = ZSTD_hashPtr(ip, HLOG, MLS);
+        U32 const matchIndex = hashTable[hash];
+        hashTable[hash] = (U32)(ip - istart);
+
+        if (matchIndex != BADIDX) {
+            const BYTE* const match = istart + matchIndex;
+            U32 const matchLen = (U32)ZSTD_count(ip, match, iend);
+            if (matchLen >= ZSTD_MINMATCH_MIN) {
+                U32 const litLen = (U32)(ip - anchor);
+                U32 const offset = (U32)(ip - match);
+                ZSTD_Sequence const seq = {
+                    offset, litLen, matchLen, 0
+                };
+
+                /* Note: it's crucial to stay within the window size! */
+                if (offset <= windowSize) {
+                    outSeqs[seqCount++] = seq;
+                    ip += matchLen;
+                    anchor = ip;
+                    continue;
+                }
+            }
+        }
+
+        ip++;
+    }
+
+    {   ZSTD_Sequence const finalSeq = {
+            0, (U32)(iend - anchor), 0, 0
+        };
+        outSeqs[seqCount++] = finalSeq;
+    }
+
+    return seqCount;
+}

contrib/externalMatchfinder/matchfinder.h

@@ -0,0 +1,26 @@
+/*
+ * Copyright (c) Yann Collet, Meta Platforms, Inc.
+ * 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 MATCHFINDER_H
+#define MATCHFINDER_H
+
+#define ZSTD_STATIC_LINKING_ONLY
+#include "zstd.h"
+
+size_t simpleExternalMatchFinder(
+  void* externalMatchState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+);
+
+#endif

lib/common/error_private.c

@@ -31,6 +31,7 @@ const char* ERR_getErrorString(ERR_enum code)
     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";
@@ -51,6 +52,7 @@ const char* ERR_getErrorString(ERR_enum code)
     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(externalMatchFinder_failed): return "External matchfinder returned an error code";
     case PREFIX(maxCode):
     default: return notErrorCode;
     }

lib/compress/zstd_compress.c

@@ -278,6 +278,16 @@ static ZSTD_paramSwitch_e ZSTD_resolveEnableLdm(ZSTD_paramSwitch_e mode,
     return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 27) ? ZSTD_ps_enable : ZSTD_ps_disable;
 }
 
+/* Enables validation for external sequences in debug builds. */
+static int ZSTD_resolveExternalSequenceValidation(int mode) {
+#if defined(DEBUGLEVEL) && (DEBUGLEVEL>=2)
+    (void)mode;
+    return 1;
+#else
+    return mode;
+#endif
+}
+
 /* 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) {
@@ -301,6 +311,7 @@ static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams(
     }
     cctxParams.useBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams.useBlockSplitter, &cParams);
     cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);
+    cctxParams.validateSequences = ZSTD_resolveExternalSequenceValidation(cctxParams.validateSequences);
     assert(!ZSTD_checkCParams(cParams));
     return cctxParams;
 }
@@ -362,6 +373,7 @@ static void ZSTD_CCtxParams_init_internal(ZSTD_CCtx_params* cctxParams, ZSTD_par
     cctxParams->useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams->useRowMatchFinder, &params->cParams);
     cctxParams->useBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams->useBlockSplitter, &params->cParams);
     cctxParams->ldmParams.enableLdm = ZSTD_resolveEnableLdm(cctxParams->ldmParams.enableLdm, &params->cParams);
+    cctxParams->validateSequences = ZSTD_resolveExternalSequenceValidation(cctxParams->validateSequences);
     DEBUGLOG(4, "ZSTD_CCtxParams_init_internal: useRowMatchFinder=%d, useBlockSplitter=%d ldm=%d",
                 cctxParams->useRowMatchFinder, cctxParams->useBlockSplitter, cctxParams->ldmParams.enableLdm);
 }
@@ -584,6 +596,11 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param)
         bounds.upperBound = (int)ZSTD_ps_disable;
         return bounds;
 
+    case ZSTD_c_enableMatchFinderFallback:
+        bounds.lowerBound = 0;
+        bounds.upperBound = 1;
+        return bounds;
+
     default:
         bounds.error = ERROR(parameter_unsupported);
         return bounds;
@@ -649,6 +666,7 @@ static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param)
     case ZSTD_c_useRowMatchFinder:
     case ZSTD_c_deterministicRefPrefix:
     case ZSTD_c_prefetchCDictTables:
+    case ZSTD_c_enableMatchFinderFallback:
     default:
         return 0;
     }
@@ -705,6 +723,7 @@ size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int value)
     case ZSTD_c_useRowMatchFinder:
     case ZSTD_c_deterministicRefPrefix:
     case ZSTD_c_prefetchCDictTables:
+    case ZSTD_c_enableMatchFinderFallback:
         break;
 
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
@@ -937,6 +956,11 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams,
         CCtxParams->prefetchCDictTables = (ZSTD_paramSwitch_e)value;
         return CCtxParams->prefetchCDictTables;
 
+    case ZSTD_c_enableMatchFinderFallback:
+        BOUNDCHECK(ZSTD_c_enableMatchFinderFallback, value);
+        CCtxParams->enableMatchFinderFallback = value;
+        return CCtxParams->enableMatchFinderFallback;
+
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
     }
 }
@@ -1072,6 +1096,9 @@ size_t ZSTD_CCtxParams_getParameter(
     case ZSTD_c_prefetchCDictTables:
         *value = (int)CCtxParams->prefetchCDictTables;
         break;
+    case ZSTD_c_enableMatchFinderFallback:
+        *value = CCtxParams->enableMatchFinderFallback;
+        break;
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
     }
     return 0;
@@ -1243,6 +1270,7 @@ size_t ZSTD_CCtx_reset(ZSTD_CCtx* cctx, ZSTD_ResetDirective reset)
         RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                         "Can't reset parameters only when not in init stage.");
         ZSTD_clearAllDicts(cctx);
+        ZSTD_memset(&cctx->externalMatchCtx, 0, sizeof(cctx->externalMatchCtx));
         return ZSTD_CCtxParams_reset(&cctx->requestedParams);
     }
     return 0;
@@ -1485,6 +1513,13 @@ ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams,
     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 useExternalMatchFinder) {
+    U32 const divider = (minMatch==3 || useExternalMatchFinder) ? 3 : 4;
+    return blockSize / divider;
+}
+
 static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
         const ZSTD_compressionParameters* cParams,
         const ldmParams_t* ldmParams,
@@ -1492,12 +1527,12 @@ static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
         const ZSTD_paramSwitch_e useRowMatchFinder,
         const size_t buffInSize,
         const size_t buffOutSize,
-        const U64 pledgedSrcSize)
+        const U64 pledgedSrcSize,
+        int useExternalMatchFinder)
 {
     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 = ZSTD_maxNbSeq(blockSize, cParams->minMatch, useExternalMatchFinder);
-    size_t const maxNbSeq = blockSize / divider;
     size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize)
                             + ZSTD_cwksp_aligned_alloc_size(maxNbSeq * sizeof(seqDef))
                             + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE));
@@ -1516,6 +1551,11 @@ static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
 
     size_t const cctxSpace = isStatic ? ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx)) : 0;
 
+    size_t const maxNbExternalSeq = ZSTD_sequenceBound(blockSize);
+    size_t const externalSeqSpace = useExternalMatchFinder
+        ? ZSTD_cwksp_alloc_size(maxNbExternalSeq * sizeof(ZSTD_Sequence))
+        : 0;
+
     size_t const neededSpace =
         cctxSpace +
         entropySpace +
@@ -1524,7 +1564,8 @@ static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
         ldmSeqSpace +
         matchStateSize +
         tokenSpace +
-        bufferSpace;
+        bufferSpace +
+        externalSeqSpace;
 
     DEBUGLOG(5, "estimate workspace : %u", (U32)neededSpace);
     return neededSpace;
@@ -1542,7 +1583,7 @@ size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params)
      * 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, params->useExternalMatchFinder);
 }
 
 size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams)
@@ -1603,7 +1644,7 @@ size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params)
 
         return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
             &cParams, &params->ldmParams, 1, useRowMatchFinder, inBuffSize, outBuffSize,
-            ZSTD_CONTENTSIZE_UNKNOWN);
+            ZSTD_CONTENTSIZE_UNKNOWN, params->useExternalMatchFinder);
     }
 }
 
@@ -1886,8 +1927,7 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
 
     {   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 = ZSTD_maxNbSeq(blockSize, params->cParams.minMatch, params->useExternalMatchFinder);
-        size_t const maxNbSeq = blockSize / divider;
         size_t const buffOutSize = (zbuff == ZSTDb_buffered && params->outBufferMode == ZSTD_bm_buffered)
                 ? ZSTD_compressBound(blockSize) + 1
                 : 0;
@@ -1904,7 +1944,7 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
         size_t const neededSpace =
             ZSTD_estimateCCtxSize_usingCCtxParams_internal(
                 &params->cParams, &params->ldmParams, zc->staticSize != 0, params->useRowMatchFinder,
-                buffInSize, buffOutSize, pledgedSrcSize);
+                buffInSize, buffOutSize, pledgedSrcSize, params->useExternalMatchFinder);
         int resizeWorkspace;
 
         FORWARD_IF_ERROR(neededSpace, "cctx size estimate failed!");
@@ -2017,6 +2057,14 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
             zc->ldmState.loadedDictEnd = 0;
         }
 
+        /* reserve space for block-level external sequences */
+        if (params->useExternalMatchFinder) {
+            size_t const maxNbExternalSeq = ZSTD_sequenceBound(blockSize);
+            zc->externalMatchCtx.seqBufferCapacity = maxNbExternalSeq;
+            zc->externalMatchCtx.seqBuffer =
+                (ZSTD_Sequence*)ZSTD_cwksp_reserve_aligned(ws, maxNbExternalSeq * sizeof(ZSTD_Sequence));
+        }
+
         DEBUGLOG(3, "wksp: finished allocating, %zd bytes remain available", ZSTD_cwksp_available_space(ws));
         assert(ZSTD_cwksp_estimated_space_within_bounds(ws, neededSpace, resizeWorkspace));
 
@@ -2868,6 +2916,55 @@ void ZSTD_resetSeqStore(seqStore_t* ssPtr)
     ssPtr->longLengthType = ZSTD_llt_none;
 }
 
+/* ZSTD_postProcessExternalMatchFinderResult() :
+ * 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_postProcessExternalMatchFinderResult(
+    ZSTD_Sequence* outSeqs, size_t nbExternalSeqs, size_t outSeqsCapacity, size_t srcSize
+) {
+    RETURN_ERROR_IF(
+        nbExternalSeqs > outSeqsCapacity,
+        externalMatchFinder_failed,
+        "External matchfinder returned error code %lu",
+        (unsigned long)nbExternalSeqs
+    );
+
+    RETURN_ERROR_IF(
+        nbExternalSeqs == 0 && srcSize > 0,
+        externalMatchFinder_failed,
+        "External matchfinder produced zero sequences 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,
+            externalMatchFinder_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;
+    }
+}
+
 typedef enum { ZSTDbss_compress, ZSTDbss_noCompress } ZSTD_buildSeqStore_e;
 
 static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
@@ -2915,6 +3012,15 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
         }
         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(
+                zc->appliedParams.useExternalMatchFinder,
+                parameter_combination_unsupported,
+                "Long-distance matching with external matchfinder enabled is not currently supported."
+            );
+
             /* Updates ldmSeqStore.pos */
             lastLLSize =
                 ZSTD_ldm_blockCompress(&zc->externSeqStore,
@@ -2926,6 +3032,14 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
         } else if (zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable) {
             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(
+                zc->appliedParams.useExternalMatchFinder,
+                parameter_combination_unsupported,
+                "Long-distance matching with external matchfinder enabled is not currently supported."
+            );
+
             ldmSeqStore.seq = zc->ldmSequences;
             ldmSeqStore.capacity = zc->maxNbLdmSequences;
             /* Updates ldmSeqStore.size */
@@ -2940,10 +3054,64 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
                                        zc->appliedParams.useRowMatchFinder,
                                        src, srcSize);
             assert(ldmSeqStore.pos == ldmSeqStore.size);
-        } else {   /* not long range mode */
+        } else if (zc->appliedParams.useExternalMatchFinder) {
+            assert(
+                zc->externalMatchCtx.seqBufferCapacity >= ZSTD_sequenceBound(srcSize)
+            );
+            assert(zc->externalMatchCtx.mFinder != NULL);
+
+            {   U32 const windowSize = (U32)1 << zc->appliedParams.cParams.windowLog;
+
+                size_t const nbExternalSeqs = (zc->externalMatchCtx.mFinder)(
+                    zc->externalMatchCtx.mState,
+                    zc->externalMatchCtx.seqBuffer,
+                    zc->externalMatchCtx.seqBufferCapacity,
+                    src, srcSize,
+                    NULL, 0,  /* dict and dictSize, currently not supported */
+                    zc->appliedParams.compressionLevel,
+                    windowSize
+                );
+
+                size_t const nbPostProcessedSeqs = ZSTD_postProcessExternalMatchFinderResult(
+                    zc->externalMatchCtx.seqBuffer,
+                    nbExternalSeqs,
+                    zc->externalMatchCtx.seqBufferCapacity,
+                    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};
+                    ZSTD_copySequencesToSeqStoreExplicitBlockDelim(
+                        zc, &seqPos, zc->externalMatchCtx.seqBuffer, nbPostProcessedSeqs, src, srcSize
+                    );
+                    ms->ldmSeqStore = NULL;
+                    DEBUGLOG(5, "Copied %lu sequences from external matchfinder 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 const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy,
+                                                                                            zc->appliedParams.useRowMatchFinder,
+                                                                                            dictMode);
+                    ms->ldmSeqStore = NULL;
+                    DEBUGLOG(
+                        5,
+                        "External matchfinder returned error code %lu. Falling back to internal matchfinder.",
+                        (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 const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy,
                                                                                     zc->appliedParams.useRowMatchFinder,
                                                                                     dictMode);
+            assert(zc->externalMatchCtx.mFinder == NULL);
             ms->ldmSeqStore = NULL;
             lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize);
         }
@@ -5726,6 +5894,7 @@ static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx,
     params.useBlockSplitter = ZSTD_resolveBlockSplitterMode(params.useBlockSplitter, &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);
 
 #ifdef ZSTD_MULTITHREAD
     if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) {
@@ -5927,12 +6096,6 @@ size_t ZSTD_compress2(ZSTD_CCtx* cctx,
     }
 }
 
-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()
  * @returns a ZSTD error code if sequence is not valid
@@ -5970,10 +6133,7 @@ static U32 ZSTD_finalizeOffBase(U32 rawOffset, const U32 rep[ZSTD_REP_NUM], U32
     return offBase;
 }
 
-/* Returns 0 on success, and a ZSTD_error otherwise. This function scans through an array of
+size_t
- * ZSTD_Sequence, storing the sequences it finds, until it reaches a block delimiter.
- */
-static size_t
 ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx,
                                               ZSTD_sequencePosition* seqPos,
                                         const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
@@ -6027,19 +6187,7 @@ ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx,
     return 0;
 }
 
-/* Returns the number of bytes to move the current read position back by.
+size_t
- * Only non-zero if we ended up splitting a sequence.
- * Otherwise, it may return a ZSTD error if something went wrong.
- *
- * This function will attempt to scan through blockSize bytes
- * represented by the sequences in @inSeqs,
- * storing any (partial) sequences.
- *
- * 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.
- */
-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)
@@ -6572,3 +6720,19 @@ ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long srcSizeH
     if (srcSizeHint == 0) srcSizeHint = ZSTD_CONTENTSIZE_UNKNOWN;
     return ZSTD_getParams_internal(compressionLevel, srcSizeHint, dictSize, ZSTD_cpm_unknown);
 }
+
+void ZSTD_registerExternalMatchFinder(
+    ZSTD_CCtx* zc, void* mState,
+    ZSTD_externalMatchFinder_F* mFinder
+) {
+    ZSTD_externalMatchCtx emctx = {
+        mState,
+        mFinder,
+
+        /* seqBuffer is allocated later (from the cwskp) */
+        NULL, /* seqBuffer */
+        0 /* seqBufferCapacity */
+    };
+    zc->externalMatchCtx = emctx;
+    zc->requestedParams.useExternalMatchFinder = 1;
+}

lib/compress/zstd_compress_internal.h

@@ -150,6 +150,12 @@ typedef struct {
   size_t capacity;      /* The capacity starting from `seq` pointer */
 } rawSeqStore_t;
 
+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;
+
 UNUSED_ATTR static const rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0, 0};
 
 typedef struct {
@@ -340,6 +346,15 @@ struct ZSTD_CCtx_params_s {
 
     /* 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;
+
+    /* Indicates whether an external matchfinder has been referenced.
+     * Users can't set this externally.
+     * It is set internally in ZSTD_registerExternalMatchFinder(). */
+    int useExternalMatchFinder;
 };  /* typedef'd to ZSTD_CCtx_params within "zstd.h" */
 
 #define COMPRESS_SEQUENCES_WORKSPACE_SIZE (sizeof(unsigned) * (MaxSeq + 2))
@@ -371,6 +386,14 @@ typedef struct {
     ZSTD_entropyCTablesMetadata_t entropyMetadata;
 } ZSTD_blockSplitCtx;
 
+/* Context for block-level external matchfinder API */
+typedef struct {
+  void* mState;
+  ZSTD_externalMatchFinder_F* mFinder;
+  ZSTD_Sequence* seqBuffer;
+  size_t seqBufferCapacity;
+} ZSTD_externalMatchCtx;
+
 struct ZSTD_CCtx_s {
     ZSTD_compressionStage_e stage;
     int cParamsChanged;                  /* == 1 if cParams(except wlog) or compression level are changed in requestedParams. Triggers transmission of new params to ZSTDMT (if available) then reset to 0. */
@@ -440,6 +463,9 @@ struct ZSTD_CCtx_s {
 
     /* Workspace for block splitter */
     ZSTD_blockSplitCtx blockSplitCtx;
+
+    /* Workspace for external matchfinder */
+    ZSTD_externalMatchCtx externalMatchCtx;
 };
 
 typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
@@ -1411,4 +1437,31 @@ U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat);
  */
 void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize);
 
+/* 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.
+ * Note that the block delimiter must include the last literals of the block.
+ */
+size_t
+ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx,
+                                              ZSTD_sequencePosition* seqPos,
+                                        const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                        const void* src, size_t 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 will attempt to scan through blockSize bytes
+ * represented by the sequences in @inSeqs,
+ * storing any (partial) sequences.
+ *
+ * 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.
+ */
+size_t
+ZSTD_copySequencesToSeqStoreNoBlockDelim(ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos,
+                                   const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                   const void* src, size_t blockSize);
+
 #endif /* ZSTD_COMPRESS_H */

lib/zstd.h

@@ -478,6 +478,7 @@ typedef enum {
      * ZSTD_c_useBlockSplitter
      * ZSTD_c_useRowMatchFinder
      * ZSTD_c_prefetchCDictTables
+     * ZSTD_c_enableMatchFinderFallback
      * 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.
@@ -497,7 +498,8 @@ typedef enum {
      ZSTD_c_experimentalParam13=1010,
      ZSTD_c_experimentalParam14=1011,
      ZSTD_c_experimentalParam15=1012,
-     ZSTD_c_experimentalParam16=1013
+     ZSTD_c_experimentalParam16=1013,
+     ZSTD_c_experimentalParam17=1014
 } ZSTD_cParameter;
 
 typedef struct {
@@ -560,7 +562,7 @@ typedef enum {
  *                  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 matchfinder.
  *                  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.
@@ -2044,6 +2046,20 @@ ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const vo
  */
 #define ZSTD_c_prefetchCDictTables ZSTD_c_experimentalParam16
 
+/* ZSTD_c_enableMatchFinderFallback
+ * Allowed values are 0 (disable) and 1 (enable). The default setting is 0.
+ *
+ * Controls whether zstd will fall back to an internal matchfinder if an
+ * external matchfinder is registered and returns an error code. This fallback is
+ * block-by-block: the internal matchfinder will only be called for blocks where
+ * the external matchfinder returns an error code. Fallback compression 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 external matchfinder API
+ * documentation (below) before setting this parameter. */
+#define ZSTD_c_enableMatchFinderFallback ZSTD_c_experimentalParam17
+
 /*! ZSTD_CCtx_getParameter() :
  *  Get the requested compression parameter value, selected by enum ZSTD_cParameter,
  *  and store it into int* value.
@@ -2676,6 +2692,142 @@ ZSTDLIB_STATIC_API size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx, void* dst, size_
 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. */
 
 
+/* ********************** EXTERNAL MATCHFINDER API **********************
+ *
+ * *** OVERVIEW ***
+ * This API allows users to replace the zstd internal block-level matchfinder
+ * with an external matchfinder function. Potential applications of the API
+ * include hardware-accelerated matchfinders and matchfinders specialized to
+ * particular types of data.
+ *
+ * See contrib/externalMatchfinder for an example program employing the
+ * external matchfinder API.
+ *
+ * *** USAGE ***
+ * The user is responsible for implementing a function of type
+ * ZSTD_externalMatchFinder_F. For each block, zstd will pass the following
+ * arguments to the user-provided function:
+ *
+ *   - externalMatchState: a pointer to a user-managed state for the external
+ *     matchfinder.
+ *
+ *   - outSeqs, outSeqsCapacity: an output buffer for sequences produced by the
+ *     external matchfinder. outSeqsCapacity is guaranteed >=
+ *     ZSTD_sequenceBound(srcSize). The memory backing outSeqs is managed by
+ *     the CCtx.
+ *
+ *   - src, srcSize: an input buffer which the external matchfinder must parse
+ *     into sequences. srcSize is guaranteed to be <= ZSTD_BLOCKSIZE_MAX.
+ *
+ *   - dict, dictSize: a history buffer, which may be empty, which the external
+ *     matchfinder may reference as it produces sequences for the src buffer.
+ *     Currently, zstd will always pass dictSize == 0 into external matchfinders,
+ *     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 external matchfinder 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_EXTERNAL_MATCHFINDER_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 is equal to srcSize.
+ *   - All sequences in the parse have matchLength != 0, except for the final
+ *     sequence. matchLength is not constrained for the final sequence.
+ *   - All offsets respect the windowSize parameter as specified in
+ *     doc/zstd_compression_format.md.
+ *
+ * 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 matchfinder or fail the compression operation. The user can
+ * choose between the two behaviors by setting the
+ * ZSTD_c_enableMatchFinderFallback 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_externalMatchFinder_F
+ * function by calling ZSTD_registerExternalMatchFinder(cctx, externalMatchState,
+ * externalMatchFinder). This setting will persist until the next parameter reset
+ * of the CCtx.
+ *
+ * The externalMatchState must be initialized by the user before calling
+ * ZSTD_registerExternalMatchFinder. The user is responsible for destroying the
+ * externalMatchState.
+ *
+ * *** LIMITATIONS ***
+ * External matchfinders are compatible with all zstd compression APIs. There are
+ * only two limitations.
+ *
+ * First, the ZSTD_c_enableLongDistanceMatching cParam is not supported.
+ * COMPRESSION WILL FAIL if it is enabled and the user tries to compress with an
+ * external matchfinder.
+ *   - 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 matchfinder 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 external
+ *     matchfinder API is used in conjunction with advanced settings (like windowLog).
+ *
+ * Second, history buffers are not supported. Concretely, zstd will always pass
+ * dictSize == 0 to the external matchfinder (for now). This has two implications:
+ *   - Dictionaries are not supported. Compression will *not* fail if the user
+ *     references a dictionary, but the dictionary won't have any effect.
+ *   - Stream history is not supported. All compression APIs, including streaming
+ *     APIs, work with the external matchfinder, but the external matchfinder won't
+ *     receive any history from the previous block. Each block is an independent chunk.
+ *
+ * Long-term, we plan to overcome both limitations. There is no technical blocker to
+ * overcoming them. It is purely a question of engineering effort.
+ */
+
+#define ZSTD_EXTERNAL_MATCHFINDER_ERROR ((size_t)(-1))
+
+typedef size_t ZSTD_externalMatchFinder_F (
+  void* externalMatchState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+);
+
+/*! ZSTD_registerExternalMatchFinder() :
+ * Instruct zstd to use an external matchfinder function.
+ *
+ * The externalMatchState 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.
+ *
+ * The user is strongly encouraged to read the full API documentation (above)
+ * before calling this function. */
+ZSTDLIB_STATIC_API void
+ZSTD_registerExternalMatchFinder(
+  ZSTD_CCtx* cctx,
+  void* externalMatchState,
+  ZSTD_externalMatchFinder_F* externalMatchFinder
+);
+
 #endif   /* ZSTD_H_ZSTD_STATIC_LINKING_ONLY */
 
 #if defined (__cplusplus)

lib/zstd_errors.h

@@ -75,6 +75,7 @@ typedef enum {
   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,
@@ -92,6 +93,7 @@ typedef enum {
   ZSTD_error_seekableIO          = 102,
   ZSTD_error_dstBuffer_wrong     = 104,
   ZSTD_error_srcBuffer_wrong     = 105,
+  ZSTD_error_externalMatchFinder_failed = 106,
   ZSTD_error_maxCode = 120  /* never EVER use this value directly, it can change in future versions! Use ZSTD_isError() instead */
 } ZSTD_ErrorCode;
 

tests/Makefile

@@ -169,7 +169,7 @@ fuzzer-dll : $(ZSTDDIR)/common/xxhash.c $(PRGDIR)/util.c $(PRGDIR)/timefn.c $(PR
 	$(CC) $(CPPFLAGS) $(CFLAGS) $(filter %.c,$^) $(LDFLAGS) -o $@$(EXT)
 
 CLEAN += zstreamtest zstreamtest32
-ZSTREAM_LOCAL_FILES := $(PRGDIR)/datagen.c $(PRGDIR)/util.c $(PRGDIR)/timefn.c seqgen.c zstreamtest.c
+ZSTREAM_LOCAL_FILES := $(PRGDIR)/datagen.c $(PRGDIR)/util.c $(PRGDIR)/timefn.c seqgen.c zstreamtest.c external_matchfinder.c
 ZSTREAM_PROPER_FILES := $(ZDICT_FILES) $(ZSTREAM_LOCAL_FILES)
 ZSTREAMFILES := $(ZSTD_FILES) $(ZSTREAM_PROPER_FILES)
 zstreamtest32 : CFLAGS += -m32

tests/external_matchfinder.c

@@ -0,0 +1,117 @@
+/*
+ * Copyright (c) Yann Collet, Meta Platforms, Inc.
+ * 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 "external_matchfinder.h"
+#include <string.h>
+#include "zstd_compress_internal.h"
+
+#define HSIZE 1024
+static U32 const HLOG = 10;
+static U32 const MLS = 4;
+static U32 const BADIDX = 0xffffffff;
+
+static size_t simpleExternalMatchFinder(
+  void* externalMatchState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+) {
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* ip = istart;
+    const BYTE* anchor = istart;
+    size_t seqCount = 0;
+    U32 hashTable[HSIZE];
+
+    (void)externalMatchState;
+    (void)dict;
+    (void)dictSize;
+    (void)outSeqsCapacity;
+    (void)compressionLevel;
+
+    {   int i;
+        for (i=0; i < HSIZE; i++) {
+            hashTable[i] = BADIDX;
+    }   }
+
+    while (ip + MLS < iend) {
+        size_t const hash = ZSTD_hashPtr(ip, HLOG, MLS);
+        U32 const matchIndex = hashTable[hash];
+        hashTable[hash] = (U32)(ip - istart);
+
+        if (matchIndex != BADIDX) {
+            const BYTE* const match = istart + matchIndex;
+            U32 const matchLen = (U32)ZSTD_count(ip, match, iend);
+            if (matchLen >= ZSTD_MINMATCH_MIN) {
+                U32 const litLen = (U32)(ip - anchor);
+                U32 const offset = (U32)(ip - match);
+                ZSTD_Sequence const seq = {
+                    offset, litLen, matchLen, 0
+                };
+
+                /* Note: it's crucial to stay within the window size! */
+                if (offset <= windowSize) {
+                    outSeqs[seqCount++] = seq;
+                    ip += matchLen;
+                    anchor = ip;
+                    continue;
+                }
+            }
+        }
+
+        ip++;
+    }
+
+    {   ZSTD_Sequence const finalSeq = {
+            0, (U32)(iend - anchor), 0, 0
+        };
+        outSeqs[seqCount++] = finalSeq;
+    }
+
+    return seqCount;
+}
+
+size_t zstreamExternalMatchFinder(
+  void* externalMatchState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+) {
+    EMF_testCase const testCase = *((EMF_testCase*)externalMatchState);
+    memset(outSeqs, 0, outSeqsCapacity);
+
+    switch (testCase) {
+        case EMF_ZERO_SEQS:
+            return 0;
+        case EMF_ONE_BIG_SEQ:
+            outSeqs[0].offset = 0;
+            outSeqs[0].matchLength = 0;
+            outSeqs[0].litLength = (U32)(srcSize);
+            return 1;
+         case EMF_LOTS_OF_SEQS:
+            return simpleExternalMatchFinder(
+                externalMatchState,
+                outSeqs, outSeqsCapacity,
+                src, srcSize,
+                dict, dictSize,
+                compressionLevel,
+                windowSize
+            );
+        case EMF_SMALL_ERROR:
+            return outSeqsCapacity + 1;
+        case EMF_BIG_ERROR:
+        default:
+            return ZSTD_EXTERNAL_MATCHFINDER_ERROR;
+    }
+}

tests/external_matchfinder.h

@@ -0,0 +1,35 @@
+/*
+ * Copyright (c) Yann Collet, Meta Platforms, Inc.
+ * 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 EXTERNAL_MATCHFINDER
+#define EXTERNAL_MATCHFINDER
+
+#define ZSTD_STATIC_LINKING_ONLY
+#include "zstd.h"
+
+/* See external_matchfinder.c for details on each test case */
+typedef enum {
+    EMF_ZERO_SEQS = 0,
+    EMF_ONE_BIG_SEQ = 1,
+    EMF_LOTS_OF_SEQS = 2,
+    EMF_BIG_ERROR = 3,
+    EMF_SMALL_ERROR = 4
+} EMF_testCase;
+
+size_t zstreamExternalMatchFinder(
+  void* externalMatchState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+);
+
+#endif // EXTERNAL_MATCHFINDER

tests/zstreamtest.c

@@ -39,7 +39,7 @@
 #include "seqgen.h"
 #include "util.h"
 #include "timefn.h"       /* UTIL_time_t, UTIL_clockSpanMicro, UTIL_getTime */
-
+#include "external_matchfinder.h"   /* zstreamExternalMatchFinder, EMF_testCase */
 
 /*-************************************
  *  Constants
@@ -1834,6 +1834,97 @@ static int basicUnitTests(U32 seed, double compressibility, int bigTests)
     }
     DISPLAYLEVEL(3, "OK \n");
 
+    DISPLAYLEVEL(3, "test%3i : External matchfinder API: ", testNb++);
+    {
+        size_t const dstBufSize = ZSTD_compressBound(CNBufferSize);
+        BYTE* const dstBuf = (BYTE*)malloc(ZSTD_compressBound(dstBufSize));
+        size_t const checkBufSize = CNBufferSize;
+        BYTE* const checkBuf = (BYTE*)malloc(checkBufSize);
+        int enableFallback;
+        EMF_testCase externalMatchState;
+
+        CHECK(dstBuf == NULL || checkBuf == NULL, "allocation failed");
+
+        ZSTD_CCtx_reset(zc, ZSTD_reset_session_and_parameters);
+
+        /* Reference external matchfinder outside the test loop to
+         * check that the reference is preserved across compressions */
+        ZSTD_registerExternalMatchFinder(
+            zc,
+            &externalMatchState,
+            zstreamExternalMatchFinder
+        );
+
+        for (enableFallback = 0; enableFallback < 1; enableFallback++) {
+            size_t testCaseId;
+
+            EMF_testCase const EMF_successCases[] = {
+                EMF_ONE_BIG_SEQ,
+                EMF_LOTS_OF_SEQS,
+            };
+            size_t const EMF_numSuccessCases = 2;
+
+            EMF_testCase const EMF_failureCases[] = {
+                EMF_ZERO_SEQS,
+                EMF_BIG_ERROR,
+                EMF_SMALL_ERROR,
+            };
+            size_t const EMF_numFailureCases = 3;
+
+            /* Test external matchfinder success scenarios */
+            for (testCaseId = 0; testCaseId < EMF_numSuccessCases; testCaseId++) {
+                size_t res;
+                externalMatchState = EMF_successCases[testCaseId];
+                ZSTD_CCtx_reset(zc, ZSTD_reset_session_only);
+                CHECK_Z(ZSTD_CCtx_setParameter(zc, ZSTD_c_enableMatchFinderFallback, enableFallback));
+                res = ZSTD_compress2(zc, dstBuf, dstBufSize, CNBuffer, CNBufferSize);
+                CHECK(ZSTD_isError(res), "EMF: Compression error: %s", ZSTD_getErrorName(res));
+                CHECK_Z(ZSTD_decompress(checkBuf, checkBufSize, dstBuf, res));
+                CHECK(memcmp(CNBuffer, checkBuf, CNBufferSize) != 0, "EMF: Corruption!");
+            }
+
+            /* Test external matchfinder failure scenarios */
+            for (testCaseId = 0; testCaseId < EMF_numFailureCases; testCaseId++) {
+                size_t res;
+                externalMatchState = EMF_failureCases[testCaseId];
+                ZSTD_CCtx_reset(zc, ZSTD_reset_session_only);
+                CHECK_Z(ZSTD_CCtx_setParameter(zc, ZSTD_c_enableMatchFinderFallback, enableFallback));
+                res = ZSTD_compress2(zc, dstBuf, dstBufSize, CNBuffer, CNBufferSize);
+                if (enableFallback) {
+                    CHECK_Z(ZSTD_decompress(checkBuf, checkBufSize, dstBuf, res));
+                    CHECK(memcmp(CNBuffer, checkBuf, CNBufferSize) != 0, "EMF: Corruption!");
+                } else {
+                    CHECK(!ZSTD_isError(res), "EMF: Should have raised an error!");
+                    CHECK(
+                        ZSTD_getErrorCode(res) != ZSTD_error_externalMatchFinder_failed,
+                        "EMF: Wrong error code: %s", ZSTD_getErrorName(res)
+                    );
+                }
+            }
+
+            /* Test compression with external matchfinder + empty src buffer */
+            {
+                size_t res;
+                externalMatchState = EMF_ZERO_SEQS;
+                ZSTD_CCtx_reset(zc, ZSTD_reset_session_only);
+                CHECK_Z(ZSTD_CCtx_setParameter(zc, ZSTD_c_enableMatchFinderFallback, enableFallback));
+                res = ZSTD_compress2(zc, dstBuf, dstBufSize, CNBuffer, 0);
+                CHECK(ZSTD_isError(res), "EMF: Compression error: %s", ZSTD_getErrorName(res));
+                CHECK(ZSTD_decompress(checkBuf, checkBufSize, dstBuf, res) != 0, "EMF: Empty src round trip failed!");
+            }
+        }
+
+        /* Test that reset clears the external matchfinder */
+        ZSTD_CCtx_reset(zc, ZSTD_reset_session_and_parameters);
+        externalMatchState = EMF_BIG_ERROR; /* ensure zstd will fail if the matchfinder wasn't cleared */
+        CHECK_Z(ZSTD_CCtx_setParameter(zc, ZSTD_c_enableMatchFinderFallback, 0));
+        CHECK_Z(ZSTD_compress2(zc, dstBuf, dstBufSize, CNBuffer, CNBufferSize));
+
+        free(dstBuf);
+        free(checkBuf);
+    }
+    DISPLAYLEVEL(3, "OK \n");
+
 _end:
     FUZ_freeDictionary(dictionary);
     ZSTD_freeCStream(zc);