Merge tag 'zstd-for-linus-v5.16' of git://github.com/terrelln/linux

Pull zstd update from Nick Terrell:
 "Update to zstd-1.4.10.

  Add myself as the maintainer of zstd and update the zstd version in
  the kernel, which is now 4 years out of date, to a much more recent
  zstd release. This includes bug fixes, much more extensive fuzzing,
  and performance improvements. And generates the kernel zstd
  automatically from upstream zstd, so it is easier to keep the zstd
  verison up to date, and we don't fall so far out of date again.

  This includes 5 commits that update the zstd library version:

   - Adds a new kernel-style wrapper around zstd.

     This wrapper API is functionally equivalent to the subset of the
     current zstd API that is currently used. The wrapper API changes to
     be kernel style so that the symbols don't collide with zstd's
     symbols. The update to zstd-1.4.10 maintains the same API and
     preserves the semantics, so that none of the callers need to be
     updated. All callers are updated in the commit, because there are
     zero functional changes.

   - Adds an indirection for `lib/decompress_unzstd.c` so it doesn't
     depend on the layout of `lib/zstd/` to include every source file.
     This allows the next patch to be automatically generated.

   - Imports the zstd-1.4.10 source code. This commit is automatically
     generated from upstream zstd (https://github.com/facebook/zstd).

   - Adds me (terrelln@fb.com) as the maintainer of `lib/zstd`.

   - Fixes a newly added build warning for clang.

  The discussion around this patchset has been pretty long, so I've
  included a FAQ-style summary of the history of the patchset, and why
  we are taking this approach.

  Why do we need to update?
  -------------------------

  The zstd version in the kernel is based off of zstd-1.3.1, which is
  was released August 20, 2017. Since then zstd has seen many bug fixes
  and performance improvements. And, importantly, upstream zstd is
  continuously fuzzed by OSS-Fuzz, and bug fixes aren't backported to
  older versions. So the only way to sanely get these fixes is to keep
  up to date with upstream zstd.

  There are no known security issues that affect the kernel, but we need
  to be able to update in case there are. And while there are no known
  security issues, there are relevant bug fixes. For example the problem
  with large kernel decompression has been fixed upstream for over 2
  years [1]

  Additionally the performance improvements for kernel use cases are
  significant. Measured for x86_64 on my Intel i9-9900k @ 3.6 GHz:

   - BtrFS zstd compression at levels 1 and 3 is 5% faster

   - BtrFS zstd decompression+read is 15% faster

   - SquashFS zstd decompression+read is 15% faster

   - F2FS zstd compression+write at level 3 is 8% faster

   - F2FS zstd decompression+read is 20% faster

   - ZRAM decompression+read is 30% faster

   - Kernel zstd decompression is 35% faster

   - Initramfs zstd decompression+build is 5% faster

  On top of this, there are significant performance improvements coming
  down the line in the next zstd release, and the new automated update
  patch generation will allow us to pull them easily.

  How is the update patch generated?
  ----------------------------------

  The first two patches are preparation for updating the zstd version.
  Then the 3rd patch in the series imports upstream zstd into the
  kernel. This patch is automatically generated from upstream. A script
  makes the necessary changes and imports it into the kernel. The
  changes are:

   - Replace all libc dependencies with kernel replacements and rewrite
     includes.

   - Remove unncessary portability macros like: #if defined(_MSC_VER).

   - Use the kernel xxhash instead of bundling it.

  This automation gets tested every commit by upstream's continuous
  integration. When we cut a new zstd release, we will submit a patch to
  the kernel to update the zstd version in the kernel.

  The automated process makes it easy to keep the kernel version of zstd
  up to date. The current zstd in the kernel shares the guts of the
  code, but has a lot of API and minor changes to work in the kernel.
  This is because at the time upstream zstd was not ready to be used in
  the kernel envrionment as-is. But, since then upstream zstd has
  evolved to support being used in the kernel as-is.

  Why are we updating in one big patch?
  -------------------------------------

  The 3rd patch in the series is very large. This is because it is
  restructuring the code, so it both deletes the existing zstd, and
  re-adds the new structure. Future updates will be directly
  proportional to the changes in upstream zstd since the last import.
  They will admittidly be large, as zstd is an actively developed
  project, and has hundreds of commits between every release. However,
  there is no other great alternative.

  One option ruled out is to replay every upstream zstd commit. This is
  not feasible for several reasons:

   - There are over 3500 upstream commits since the zstd version in the
     kernel.

   - The automation to automatically generate the kernel update was only
     added recently, so older commits cannot easily be imported.

   - Not every upstream zstd commit builds.

   - Only zstd releases are "supported", and individual commits may have
     bugs that were fixed before a release.

  Another option to reduce the patch size would be to first reorganize
  to the new file structure, and then apply the patch. However, the
  current kernel zstd is formatted with clang-format to be more
  "kernel-like". But, the new method imports zstd as-is, without
  additional formatting, to allow for closer correlation with upstream,
  and easier debugging. So the patch wouldn't be any smaller.

  It also doesn't make sense to import upstream zstd commit by commit
  going forward. Upstream zstd doesn't support production use cases
  running of the development branch. We have a lot of post-commit
  fuzzing that catches many bugs, so indiviudal commits may be buggy,
  but fixed before a release. So going forward, I intend to import every
  (important) zstd release into the Kernel.

  So, while it isn't ideal, updating in one big patch is the only patch
  I see forward.

  Who is responsible for this code?
  ---------------------------------

  I am. This patchset adds me as the maintainer for zstd. Previously,
  there was no tree for zstd patches. Because of that, there were
  several patches that either got ignored, or took a long time to merge,
  since it wasn't clear which tree should pick them up. I'm officially
  stepping up as maintainer, and setting up my tree as the path through
  which zstd patches get merged. I'll make sure that patches to the
  kernel zstd get ported upstream, so they aren't erased when the next
  version update happens.

  How is this code tested?
  ------------------------

  I tested every caller of zstd on x86_64 (BtrFS, ZRAM, SquashFS, F2FS,
  Kernel, InitRAMFS). I also tested Kernel & InitRAMFS on i386 and
  aarch64. I checked both performance and correctness.

  Also, thanks to many people in the community who have tested these
  patches locally.

  Lastly, this code will bake in linux-next before being merged into
  v5.16.

  Why update to zstd-1.4.10 when zstd-1.5.0 has been released?
  ------------------------------------------------------------

  This patchset has been outstanding since 2020, and zstd-1.4.10 was the
  latest release when it was created. Since the update patch is
  automatically generated from upstream, I could generate it from
  zstd-1.5.0.

  However, there were some large stack usage regressions in zstd-1.5.0,
  and are only fixed in the latest development branch. And the latest
  development branch contains some new code that needs to bake in the
  fuzzer before I would feel comfortable releasing to the kernel.

  Once this patchset has been merged, and we've released zstd-1.5.1, we
  can update the kernel to zstd-1.5.1, and exercise the update process.

  You may notice that zstd-1.4.10 doesn't exist upstream. This release
  is an artifical release based off of zstd-1.4.9, with some fixes for
  the kernel backported from the development branch. I will tag the
  zstd-1.4.10 release after this patchset is merged, so the Linux Kernel
  is running a known version of zstd that can be debugged upstream.

  Why was a wrapper API added?
  ----------------------------

  The first versions of this patchset migrated the kernel to the
  upstream zstd API. It first added a shim API that supported the new
  upstream API with the old code, then updated callers to use the new
  shim API, then transitioned to the new code and deleted the shim API.
  However, Cristoph Hellwig suggested that we transition to a kernel
  style API, and hide zstd's upstream API behind that. This is because
  zstd's upstream API is supports many other use cases, and does not
  follow the kernel style guide, while the kernel API is focused on the
  kernel's use cases, and follows the kernel style guide.

  Where is the previous discussion?
  ---------------------------------

  Links for the discussions of the previous versions of the patch set
  below. The largest changes in the design of the patchset are driven by
  the discussions in v11, v5, and v1. Sorry for the mix of links, I
  couldn't find most of the the threads on lkml.org"

Link: https://lkml.org/lkml/2020/9/29/27 [1]
Link: https://www.spinics.net/lists/linux-crypto/msg58189.html [v12]
Link: https://lore.kernel.org/linux-btrfs/20210430013157.747152-1-nickrterrell@gmail.com/ [v11]
Link: https://lore.kernel.org/lkml/20210426234621.870684-2-nickrterrell@gmail.com/ [v10]
Link: https://lore.kernel.org/linux-btrfs/20210330225112.496213-1-nickrterrell@gmail.com/ [v9]
Link: https://lore.kernel.org/linux-f2fs-devel/20210326191859.1542272-1-nickrterrell@gmail.com/ [v8]
Link: https://lkml.org/lkml/2020/12/3/1195 [v7]
Link: https://lkml.org/lkml/2020/12/2/1245 [v6]
Link: https://lore.kernel.org/linux-btrfs/20200916034307.2092020-1-nickrterrell@gmail.com/ [v5]
Link: https://www.spinics.net/lists/linux-btrfs/msg105783.html [v4]
Link: https://lkml.org/lkml/2020/9/23/1074 [v3]
Link: https://www.spinics.net/lists/linux-btrfs/msg105505.html [v2]
Link: https://lore.kernel.org/linux-btrfs/20200916034307.2092020-1-nickrterrell@gmail.com/ [v1]
Signed-off-by: Nick Terrell <terrelln@fb.com>
Tested By: Paul Jones <paul@pauljones.id.au>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Sedat Dilek <sedat.dilek@gmail.com> # LLVM/Clang v13.0.0 on x86-64
Tested-by: Jean-Denis Girard <jd.girard@sysnux.pf>

* tag 'zstd-for-linus-v5.16' of git://github.com/terrelln/linux:
  lib: zstd: Add cast to silence clang's -Wbitwise-instead-of-logical
  MAINTAINERS: Add maintainer entry for zstd
  lib: zstd: Upgrade to latest upstream zstd version 1.4.10
  lib: zstd: Add decompress_sources.h for decompress_unzstd
  lib: zstd: Add kernel-specific API

1 files changed, 390 insertions, 0 deletions

diff --git a/lib/zstd/common/fse_decompress.c b/lib/zstd/common/fse_decompress.c
new file mode 100644
index 000000000000..2c8bbe3e4c14
--- /dev/null
+++ b/lib/zstd/common/fse_decompress.c
@@ -0,0 +1,390 @@
+/* ******************************************************************
+ * FSE : Finite State Entropy decoder
+ * Copyright (c) Yann Collet, Facebook, Inc.
+ *
+ *  You can contact the author at :
+ *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+ *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
+ *
+ * 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.
+****************************************************************** */
+
+
+/* **************************************************************
+*  Includes
+****************************************************************/
+#include "debug.h"      /* assert */
+#include "bitstream.h"
+#include "compiler.h"
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
+#include "error_private.h"
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h"
+
+
+/* **************************************************************
+*  Error Management
+****************************************************************/
+#define FSE_isError ERR_isError
+#define FSE_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)   /* use only *after* variable declarations */
+
+
+/* **************************************************************
+*  Templates
+****************************************************************/
+/*
+  designed to be included
+  for type-specific functions (template emulation in C)
+  Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSE_FUNCTION_EXTENSION
+#  error "FSE_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSE_FUNCTION_TYPE
+#  error "FSE_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSE_CAT(X,Y) X##Y
+#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 */
+    FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
+    U16* symbolNext = (U16*)workSpace;
+    BYTE* spread = (BYTE*)(symbolNext + maxSymbolValue + 1);
+
+    U32 const maxSV1 = maxSymbolValue + 1;
+    U32 const tableSize = 1 << tableLog;
+    U32 highThreshold = tableSize-1;
+
+    /* Sanity Checks */
+    if (FSE_BUILD_DTABLE_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(maxSymbolValue_tooLarge);
+    if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+    /* Init, lay down lowprob symbols */
+    {   FSE_DTableHeader DTableH;
+        DTableH.tableLog = (U16)tableLog;
+        DTableH.fastMode = 1;
+        {   S16 const largeLimit= (S16)(1 << (tableLog-1));
+            U32 s;
+            for (s=0; s<maxSV1; s++) {
+                if (normalizedCounter[s]==-1) {
+                    tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
+                    symbolNext[s] = 1;
+                } else {
+                    if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
+                    symbolNext[s] = normalizedCounter[s];
+        }   }   }
+        ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
+    }
+
+    /* Spread symbols */
+    if (highThreshold == tableSize - 1) {
+        size_t const tableMask = tableSize-1;
+        size_t const step = FSE_TABLESTEP(tableSize);
+        /* First lay down the symbols in order.
+         * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
+         * misses since small blocks generally have small table logs, so nearly
+         * 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;
+            size_t pos = 0;
+            U64 sv = 0;
+            U32 s;
+            for (s=0; s<maxSV1; ++s, sv += add) {
+                int i;
+                int const n = normalizedCounter[s];
+                MEM_write64(spread + pos, sv);
+                for (i = 8; i < n; i += 8) {
+                    MEM_write64(spread + pos + i, sv);
+                }
+                pos += n;
+            }
+        }
+        /* Now we spread those positions across the table.
+         * The benefit of doing it in two stages is that we avoid the 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.
+         */
+        {
+            size_t position = 0;
+            size_t s;
+            size_t const unroll = 2;
+            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+            for (s = 0; s < (size_t)tableSize; s += unroll) {
+                size_t u;
+                for (u = 0; u < unroll; ++u) {
+                    size_t const uPosition = (position + (u * step)) & tableMask;
+                    tableDecode[uPosition].symbol = spread[s + u];
+                }
+                position = (position + (unroll * step)) & tableMask;
+            }
+            assert(position == 0);
+        }
+    } else {
+        U32 const tableMask = tableSize-1;
+        U32 const step = FSE_TABLESTEP(tableSize);
+        U32 s, position = 0;
+        for (s=0; s<maxSV1; s++) {
+            int i;
+            for (i=0; i<normalizedCounter[s]; i++) {
+                tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
+                position = (position + step) & tableMask;
+                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+        }   }
+        if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+    }
+
+    /* Build Decoding table */
+    {   U32 u;
+        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].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
+    }   }
+
+    return 0;
+}
+
+size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
+{
+    return FSE_buildDTable_internal(dt, normalizedCounter, maxSymbolValue, tableLog, workSpace, wkspSize);
+}
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/*-*******************************************************
+*  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,
+    const void* cSrc, size_t cSrcSize,
+    const FSE_DTable* dt, const unsigned fast)
+{
+    BYTE* const ostart = (BYTE*) dst;
+    BYTE* op = ostart;
+    BYTE* const omax = op + maxDstSize;
+    BYTE* const olimit = omax-3;
+
+    BIT_DStream_t bitD;
+    FSE_DState_t state1;
+    FSE_DState_t state2;
+
+    /* Init */
+    CHECK_F(BIT_initDStream(&bitD, cSrc, cSrcSize));
+
+    FSE_initDState(&state1, &bitD, dt);
+    FSE_initDState(&state2, &bitD, dt);
+
+#define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
+
+    /* 4 symbols per loop */
+    for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) & (op<olimit) ; op+=4) {
+        op[0] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[1] = FSE_GETSYMBOL(&state2);
+
+        if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }
+
+        op[2] = FSE_GETSYMBOL(&state1);
+
+        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
+            BIT_reloadDStream(&bitD);
+
+        op[3] = FSE_GETSYMBOL(&state2);
+    }
+
+    /* tail */
+    /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
+    while (1) {
+        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+        *op++ = FSE_GETSYMBOL(&state1);
+        if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
+            *op++ = FSE_GETSYMBOL(&state2);
+            break;
+        }
+
+        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+        *op++ = FSE_GETSYMBOL(&state2);
+        if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
+            *op++ = FSE_GETSYMBOL(&state1);
+            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);
+}
+
+typedef struct {
+    short ncount[FSE_MAX_SYMBOL_VALUE + 1];
+    FSE_DTable dtable[1]; /* Dynamically sized */
+} FSE_DecompressWksp;
+
+
+FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
+        void* dst, size_t dstCapacity,
+        const void* cSrc, size_t cSrcSize,
+        unsigned maxLog, void* workSpace, size_t wkspSize,
+        int bmi2)
+{
+    const BYTE* const istart = (const BYTE*)cSrc;
+    const BYTE* ip = istart;
+    unsigned tableLog;
+    unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+    FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
+
+    DEBUG_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
+    if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);
+
+    /* normal FSE decoding mode */
+    {
+        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);
+        ip += NCountLength;
+        cSrcSize -= NCountLength;
+    }
+
+    if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
+    workSpace = wksp->dtable + FSE_DTABLE_SIZE_U32(tableLog);
+    wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
+
+    CHECK_F( FSE_buildDTable_internal(wksp->dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
+
+    {
+        const void* ptr = wksp->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);
+    }
+}
+
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t FSE_decompress_wksp_body_default(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
+{
+    return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 0);
+}
+
+#if DYNAMIC_BMI2
+TARGET_ATTRIBUTE("bmi2") static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
+{
+    return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 1);
+}
+#endif
+
+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)
+{
+#if DYNAMIC_BMI2
+    if (bmi2) {
+        return FSE_decompress_wksp_body_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
+    }
+#endif
+    (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 */