zstd-ruby 1.5.2.3 → 1.5.5.0

data/ext/zstdruby/libzstd/compress/zstd_cwksp.h

@@ -1,5 +1,5 @@
 /*
- * 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
@@ -14,7 +14,9 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
+#include "../common/allocations.h"  /* ZSTD_customMalloc, ZSTD_customFree */
 #include "../common/zstd_internal.h"
+#include "../common/portability_macros.h"
 
 #if defined (__cplusplus)
 extern "C" {
@@ -44,8 +46,9 @@ extern "C" {
 ***************************************/
 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;
 
 /**
@@ -98,8 +101,8 @@ typedef enum {
  *
  * 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:
@@ -123,9 +126,18 @@ typedef enum {
  *   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
@@ -137,8 +149,9 @@ typedef enum {
  * 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.
  */
@@ -150,6 +163,7 @@ typedef struct {
     void* tableEnd;
     void* tableValidEnd;
     void* allocStart;
+    void* initOnceStart;
 
     BYTE allocFailed;
     int workspaceOversizedDuration;
@@ -162,6 +176,7 @@ typedef struct {
 ***************************************/
 
 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;
@@ -171,6 +186,20 @@ MEM_STATIC void ZSTD_cwksp_assert_internal_consistency(ZSTD_cwksp* 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);
+#if ZSTD_MEMORY_SANITIZER
+    {
+        intptr_t const offset = __msan_test_shadow(ws->initOnceStart,
+            (U8*)ZSTD_cwksp_initialAllocStart(ws) - (U8*)ws->initOnceStart);
+#if defined(ZSTD_MSAN_PRINT)
+        if(offset!=-1) {
+            __msan_print_shadow((U8*)ws->initOnceStart + offset - 8, 32);
+        }
+#endif
+        assert(offset==-1);
+    };
+#endif
 }
 
 /**
@@ -217,14 +246,10 @@ MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size) {
  * 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;
 }
 
@@ -237,10 +262,18 @@ MEM_STATIC size_t ZSTD_cwksp_bytes_to_align_ptr(void* ptr, const size_t alignByt
     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(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) {
+    return (void*)((size_t)ws->workspaceEnd & ~(ZSTD_CWKSP_ALIGNMENT_BYTES-1));
+}
+
 /**
  * Internal function. Do not use directly.
  * Reserves the given number of bytes within the aligned/buffer segment of the wksp,
@@ -281,27 +314,16 @@ ZSTD_cwksp_internal_advance_phase(ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase
 {
     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!");
@@ -309,7 +331,9 @@ ZSTD_cwksp_internal_advance_phase(ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase
                 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);
     }
@@ -321,7 +345,7 @@ ZSTD_cwksp_internal_advance_phase(ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase
  */
 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);
 }
 
 /**
@@ -348,7 +372,9 @@ ZSTD_cwksp_reserve_internal(ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase
     if (alloc) {
         alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
         if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
-            __asan_unpoison_memory_region(alloc, bytes);
+            /* We need to keep the redzone poisoned while unpoisoning the bytes that
+             * are actually allocated. */
+            __asan_unpoison_memory_region(alloc, bytes - 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE);
         }
     }
 #endif
@@ -364,6 +390,36 @@ MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes)
     return (BYTE*)ZSTD_cwksp_reserve_internal(ws, bytes, ZSTD_cwksp_alloc_buffers);
 }
 
+/**
+ * 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_init_once(ZSTD_cwksp* ws, size_t bytes)
+{
+    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;
+    }
+#if ZSTD_MEMORY_SANITIZER
+    assert(__msan_test_shadow(ptr, bytes) == -1);
+#endif
+    return ptr;
+}
+
 /**
  * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
  */
@@ -382,13 +438,17 @@ MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes)
  */
 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;
@@ -467,11 +527,19 @@ MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws)
 #if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
     /* To validate that the table re-use logic is sound, and that we don't
      * access table space that we haven't cleaned, we re-"poison" the table
-     * space every time we mark it dirty. */
+     * space every time we mark it dirty.
+     * Since tableValidEnd space and initOnce space may overlap we don't poison
+     * the initOnce portion as it break its promise. This means that this poisoning
+     * check isn't always applied fully. */
     {
         size_t size = (BYTE*)ws->tableValidEnd - (BYTE*)ws->objectEnd;
         assert(__msan_test_shadow(ws->objectEnd, size) == -1);
-        __msan_poison(ws->objectEnd, size);
+        if((BYTE*)ws->tableValidEnd < (BYTE*)ws->initOnceStart) {
+            __msan_poison(ws->objectEnd, size);
+        } else {
+            assert(ws->initOnceStart >= ws->objectEnd);
+            __msan_poison(ws->objectEnd, (BYTE*)ws->initOnceStart - (BYTE*)ws->objectEnd);
+        }
     }
 #endif
 
@@ -499,7 +567,7 @@ MEM_STATIC void ZSTD_cwksp_clean_tables(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);
 }
@@ -536,11 +604,14 @@ MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) {
 #if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
     /* To validate that the context re-use logic is sound, and that we don't
      * access stuff that this compression hasn't initialized, we re-"poison"
-     * the workspace (or at least the non-static, non-table parts of it)
-     * every time we start a new compression. */
+     * the workspace except for the areas in which we expect memory re-use
+     * without initialization (objects, valid tables area and init once
+     * memory). */
     {
-        size_t size = (BYTE*)ws->workspaceEnd - (BYTE*)ws->tableValidEnd;
-        __msan_poison(ws->tableValidEnd, size);
+        if((BYTE*)ws->tableValidEnd < (BYTE*)ws->initOnceStart) {
+            size_t size = (BYTE*)ws->initOnceStart - (BYTE*)ws->tableValidEnd;
+            __msan_poison(ws->tableValidEnd, size);
+        }
     }
 #endif
 
@@ -556,10 +627,10 @@ MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) {
 #endif
 
     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);
 }
@@ -576,6 +647,7 @@ MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size, ZSTD_c
     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);
@@ -628,17 +700,11 @@ MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) {
  * 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;
 }
 
 

data/ext/zstdruby/libzstd/compress/zstd_double_fast.c

@@ -1,5 +1,5 @@
 /*
- * 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
@@ -11,8 +11,43 @@
 #include "zstd_compress_internal.h"
 #include "zstd_double_fast.h"
 
+static 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;
 
-void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
+    /* 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 void ZSTD_fillDoubleHashTableForCCtx(ZSTD_matchState_t* ms,
                               void const* end, ZSTD_dictTableLoadMethod_e dtlm)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
@@ -43,7 +78,19 @@ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
             /* 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);
+    }
 }
 
 
@@ -67,7 +114,7 @@ size_t ZSTD_compressBlock_doubleFast_noDict_generic(
     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;
@@ -100,8 +147,8 @@ size_t 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 */
@@ -131,7 +178,7 @@ size_t ZSTD_compressBlock_doubleFast_noDict_generic(
             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;
             }
 
@@ -175,9 +222,13 @@ size_t ZSTD_compressBlock_doubleFast_noDict_generic(
         } 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);
@@ -217,7 +268,7 @@ _match_found: /* requires ip, offset, mLength */
             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 */
@@ -243,7 +294,7 @@ _match_stored:
                 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 ... (?) */
@@ -275,7 +326,6 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
     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;
@@ -286,8 +336,8 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
     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");
@@ -295,6 +345,13 @@ size_t 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);
 
@@ -309,8 +366,12 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
         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];
@@ -328,7 +389,7 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
             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;
         }
 
@@ -340,9 +401,9 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
                 while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
                 goto _match_found;
             }
-        } else {
+        } 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);
 
@@ -358,9 +419,9 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
             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;
 
@@ -375,10 +436,11 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
         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;
 
@@ -391,9 +453,9 @@ _search_next_long:
                     while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
                     goto _match_found;
                 }
-            } else {
+            } 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)) {
@@ -419,7 +481,7 @@ _match_found:
         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 */
@@ -448,7 +510,7 @@ _match_stored:
                     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;
@@ -461,8 +523,8 @@ _match_stored:
     }   /* 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);
@@ -585,7 +647,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
             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;
@@ -596,7 +658,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 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);
@@ -621,7 +683,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 }
                 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;
@@ -653,7 +715,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                     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;