brotli 0.2.3 → 0.5.0

data/vendor/brotli/c/enc/encode.c

@@ -7,6 +7,8 @@
 /* Implementation of Brotli compressor. */
 
 #include <brotli/encode.h>
+#include <brotli/shared_dictionary.h>
+#include <brotli/types.h>
 
 #include <stdlib.h>  /* free, malloc */
 #include <string.h>  /* memcpy, memset */
@@ -15,24 +17,26 @@
 #include "../common/context.h"
 #include "../common/platform.h"
 #include "../common/version.h"
-#include "./backward_references.h"
-#include "./backward_references_hq.h"
-#include "./bit_cost.h"
-#include "./brotli_bit_stream.h"
-#include "./compress_fragment.h"
-#include "./compress_fragment_two_pass.h"
-#include "./encoder_dict.h"
-#include "./entropy_encode.h"
-#include "./fast_log.h"
-#include "./hash.h"
-#include "./histogram.h"
-#include "./memory.h"
-#include "./metablock.h"
-#include "./prefix.h"
-#include "./quality.h"
-#include "./ringbuffer.h"
-#include "./utf8_util.h"
-#include "./write_bits.h"
+#include "backward_references.h"
+#include "backward_references_hq.h"
+#include "bit_cost.h"
+#include "brotli_bit_stream.h"
+#include "compress_fragment.h"
+#include "compress_fragment_two_pass.h"
+#include "dictionary_hash.h"
+#include "encoder_dict.h"
+#include "entropy_encode.h"
+#include "fast_log.h"
+#include "hash.h"
+#include "histogram.h"
+#include "memory.h"
+#include "metablock.h"
+#include "prefix.h"
+#include "state.h"
+#include "quality.h"
+#include "ringbuffer.h"
+#include "utf8_util.h"
+#include "write_bits.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
@@ -40,82 +44,6 @@ extern "C" {
 
 #define COPY_ARRAY(dst, src) memcpy(dst, src, sizeof(src));
 
-typedef enum BrotliEncoderStreamState {
-  /* Default state. */
-  BROTLI_STREAM_PROCESSING = 0,
-  /* Intermediate state; after next block is emitted, byte-padding should be
-     performed before getting back to default state. */
-  BROTLI_STREAM_FLUSH_REQUESTED = 1,
-  /* Last metablock was produced; no more input is acceptable. */
-  BROTLI_STREAM_FINISHED = 2,
-  /* Flushing compressed block and writing meta-data block header. */
-  BROTLI_STREAM_METADATA_HEAD = 3,
-  /* Writing metadata block body. */
-  BROTLI_STREAM_METADATA_BODY = 4
-} BrotliEncoderStreamState;
-
-typedef struct BrotliEncoderStateStruct {
-  BrotliEncoderParams params;
-
-  MemoryManager memory_manager_;
-
-  HasherHandle hasher_;
-  uint64_t input_pos_;
-  RingBuffer ringbuffer_;
-  size_t cmd_alloc_size_;
-  Command* commands_;
-  size_t num_commands_;
-  size_t num_literals_;
-  size_t last_insert_len_;
-  uint64_t last_flush_pos_;
-  uint64_t last_processed_pos_;
-  int dist_cache_[BROTLI_NUM_DISTANCE_SHORT_CODES];
-  int saved_dist_cache_[4];
-  uint16_t last_bytes_;
-  uint8_t last_bytes_bits_;
-  uint8_t prev_byte_;
-  uint8_t prev_byte2_;
-  size_t storage_size_;
-  uint8_t* storage_;
-  /* Hash table for FAST_ONE_PASS_COMPRESSION_QUALITY mode. */
-  int small_table_[1 << 10];  /* 4KiB */
-  int* large_table_;          /* Allocated only when needed */
-  size_t large_table_size_;
-  /* Command and distance prefix codes (each 64 symbols, stored back-to-back)
-     used for the next block in FAST_ONE_PASS_COMPRESSION_QUALITY. The command
-     prefix code is over a smaller alphabet with the following 64 symbols:
-        0 - 15: insert length code 0, copy length code 0 - 15, same distance
-       16 - 39: insert length code 0, copy length code 0 - 23
-       40 - 63: insert length code 0 - 23, copy length code 0
-     Note that symbols 16 and 40 represent the same code in the full alphabet,
-     but we do not use either of them in FAST_ONE_PASS_COMPRESSION_QUALITY. */
-  uint8_t cmd_depths_[128];
-  uint16_t cmd_bits_[128];
-  /* The compressed form of the command and distance prefix codes for the next
-     block in FAST_ONE_PASS_COMPRESSION_QUALITY. */
-  uint8_t cmd_code_[512];
-  size_t cmd_code_numbits_;
-  /* Command and literal buffers for FAST_TWO_PASS_COMPRESSION_QUALITY. */
-  uint32_t* command_buf_;
-  uint8_t* literal_buf_;
-
-  uint8_t* next_out_;
-  size_t available_out_;
-  size_t total_out_;
-  /* Temporary buffer for padding flush bits or metadata block header / body. */
-  union {
-    uint64_t u64[2];
-    uint8_t u8[16];
-  } tiny_buf_;
-  uint32_t remaining_metadata_bytes_;
-  BrotliEncoderStreamState stream_state_;
-
-  BROTLI_BOOL is_last_block_emitted_;
-  BROTLI_BOOL is_initialized_;
-} BrotliEncoderStateStruct;
-
-static BROTLI_BOOL EnsureInitialized(BrotliEncoderState* s);
-
 static size_t InputBlockSize(BrotliEncoderState* s) {
   return (size_t)1 << s->params.lgblock;
 }
@@ -135,7 +63,7 @@ BROTLI_BOOL BrotliEncoderSetParameter(
     BrotliEncoderState* state, BrotliEncoderParameter p, uint32_t value) {
   /* Changing parameters on the fly is not implemented yet. */
   if (state->is_initialized_) return BROTLI_FALSE;
-  /* TODO: Validate/clamp parameters here. */
+  /* TODO(eustas): Validate/clamp parameters here. */
   switch (p) {
     case BROTLI_PARAM_MODE:
       state->params.mode = (BrotliEncoderMode)value;
@@ -174,6 +102,11 @@ BROTLI_BOOL BrotliEncoderSetParameter(
       state->params.dist.num_direct_distance_codes = value;
       return BROTLI_TRUE;
 
+    case BROTLI_PARAM_STREAM_OFFSET:
+      if (value > (1u << 30)) return BROTLI_FALSE;
+      state->params.stream_offset = value;
+      return BROTLI_TRUE;
+
     default: return BROTLI_FALSE;
   }
 }
@@ -195,7 +128,7 @@ static uint8_t* GetBrotliStorage(BrotliEncoderState* s, size_t size) {
   if (s->storage_size_ < size) {
     BROTLI_FREE(m, s->storage_);
     s->storage_ = BROTLI_ALLOC(m, uint8_t, size);
-    if (BROTLI_IS_OOM(m)) return NULL;
+    if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(s->storage_)) return NULL;
     s->storage_size_ = size;
   }
   return s->storage_;
@@ -234,7 +167,7 @@ static int* GetHashTable(BrotliEncoderState* s, int quality,
       s->large_table_size_ = htsize;
       BROTLI_FREE(m, s->large_table_);
       s->large_table_ = BROTLI_ALLOC(m, int, htsize);
-      if (BROTLI_IS_OOM(m)) return 0;
+      if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(s->large_table_)) return 0;
     }
     table = s->large_table_;
   }
@@ -266,11 +199,9 @@ static void EncodeWindowBits(int lgwin, BROTLI_BOOL large_window,
   }
 }
 
+/* TODO(eustas): move to compress_fragment.c? */
 /* Initializes the command and distance prefix codes for the first block. */
-static void InitCommandPrefixCodes(uint8_t cmd_depths[128],
-                                   uint16_t cmd_bits[128],
-                                   uint8_t cmd_code[512],
-                                   size_t* cmd_code_numbits) {
+static void InitCommandPrefixCodes(BrotliOnePassArena* s) {
   static const uint8_t kDefaultCommandDepths[128] = {
     0, 4, 4, 5, 6, 6, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8,
     0, 0, 0, 4, 4, 4, 4, 4, 5, 5, 6, 6, 6, 6, 7, 7,
@@ -303,13 +234,13 @@ static void InitCommandPrefixCodes(uint8_t cmd_depths[128],
     0x46, 0xe1, 0xb0, 0xd0, 0x4e, 0xb2, 0xf7, 0x04, 0x00,
   };
   static const size_t kDefaultCommandCodeNumBits = 448;
-  COPY_ARRAY(cmd_depths, kDefaultCommandDepths);
-  COPY_ARRAY(cmd_bits, kDefaultCommandBits);
+  COPY_ARRAY(s->cmd_depth, kDefaultCommandDepths);
+  COPY_ARRAY(s->cmd_bits, kDefaultCommandBits);
 
   /* Initialize the pre-compressed form of the command and distance prefix
      codes. */
-  COPY_ARRAY(cmd_code, kDefaultCommandCode);
-  *cmd_code_numbits = kDefaultCommandCodeNumBits;
+  COPY_ARRAY(s->cmd_code, kDefaultCommandCode);
+  s->cmd_code_numbits = kDefaultCommandCodeNumBits;
 }
 
 /* Decide about the context map based on the ability of the prediction
@@ -384,7 +315,8 @@ static void ChooseContextMap(int quality,
    first 5 bits of literals. */
 static BROTLI_BOOL ShouldUseComplexStaticContextMap(const uint8_t* input,
     size_t start_pos, size_t length, size_t mask, int quality, size_t size_hint,
-    size_t* num_literal_contexts, const uint32_t** literal_context_map) {
+    size_t* num_literal_contexts, const uint32_t** literal_context_map,
+    uint32_t* arena) {
   static const uint32_t kStaticContextMapComplexUTF8[64] = {
     11, 11, 12, 12, /* 0 special */
     0, 0, 0, 0, /* 4 lf */
@@ -409,16 +341,17 @@ static BROTLI_BOOL ShouldUseComplexStaticContextMap(const uint8_t* input,
     return BROTLI_FALSE;
   } else {
     const size_t end_pos = start_pos + length;
-    /* To make entropy calculations faster and to fit on the stack, we collect
-       histograms over the 5 most significant bits of literals. One histogram
+    /* To make entropy calculations faster, we collect histograms
+       over the 5 most significant bits of literals. One histogram
        without context and 13 additional histograms for each context value. */
-    uint32_t combined_histo[32] = { 0 };
-    uint32_t context_histo[13][32] = { { 0 } };
+    uint32_t* BROTLI_RESTRICT const combined_histo = arena;
+    uint32_t* BROTLI_RESTRICT const context_histo = arena + 32;
     uint32_t total = 0;
     double entropy[3];
     size_t dummy;
     size_t i;
     ContextLut utf8_lut = BROTLI_CONTEXT_LUT(CONTEXT_UTF8);
+    memset(arena, 0, sizeof(arena[0]) * 32 * 14);
     for (; start_pos + 64 <= end_pos; start_pos += 4096) {
       const size_t stride_end_pos = start_pos + 64;
       uint8_t prev2 = input[start_pos & mask];
@@ -432,7 +365,7 @@ static BROTLI_BOOL ShouldUseComplexStaticContextMap(const uint8_t* input,
             BROTLI_CONTEXT(prev1, prev2, utf8_lut)];
         ++total;
         ++combined_histo[literal >> 3];
-        ++context_histo[context][literal >> 3];
+        ++context_histo[(context << 5) + (literal >> 3)];
         prev2 = prev1;
         prev1 = literal;
       }
@@ -440,7 +373,7 @@ static BROTLI_BOOL ShouldUseComplexStaticContextMap(const uint8_t* input,
     entropy[1] = ShannonEntropy(combined_histo, 32, &dummy);
     entropy[2] = 0;
     for (i = 0; i < 13; ++i) {
-      entropy[2] += ShannonEntropy(&context_histo[i][0], 32, &dummy);
+      entropy[2] += ShannonEntropy(context_histo + (i << 5), 32, &dummy);
     }
     entropy[0] = 1.0 / (double)total;
     entropy[1] *= entropy[0];
@@ -464,19 +397,21 @@ static BROTLI_BOOL ShouldUseComplexStaticContextMap(const uint8_t* input,
 
 static void DecideOverLiteralContextModeling(const uint8_t* input,
     size_t start_pos, size_t length, size_t mask, int quality, size_t size_hint,
-    size_t* num_literal_contexts, const uint32_t** literal_context_map) {
+    size_t* num_literal_contexts, const uint32_t** literal_context_map,
+    uint32_t* arena) {
   if (quality < MIN_QUALITY_FOR_CONTEXT_MODELING || length < 64) {
     return;
   } else if (ShouldUseComplexStaticContextMap(
       input, start_pos, length, mask, quality, size_hint,
-      num_literal_contexts, literal_context_map)) {
+      num_literal_contexts, literal_context_map, arena)) {
     /* Context map was already set, nothing else to do. */
   } else {
     /* Gather bi-gram data of the UTF8 byte prefixes. To make the analysis of
        UTF8 data faster we only examine 64 byte long strides at every 4kB
        intervals. */
     const size_t end_pos = start_pos + length;
-    uint32_t bigram_prefix_histo[9] = { 0 };
+    uint32_t* BROTLI_RESTRICT const bigram_prefix_histo = arena;
+    memset(bigram_prefix_histo, 0, sizeof(arena[0]) * 9);
     for (; start_pos + 64 <= end_pos; start_pos += 4096) {
       static const int lut[4] = { 0, 0, 1, 2 };
       const size_t stride_end_pos = start_pos + 64;
@@ -496,10 +431,10 @@ static void DecideOverLiteralContextModeling(const uint8_t* input,
 static BROTLI_BOOL ShouldCompress(
     const uint8_t* data, const size_t mask, const uint64_t last_flush_pos,
     const size_t bytes, const size_t num_literals, const size_t num_commands) {
-  /* TODO: find more precise minimal block overhead. */
+  /* TODO(eustas): find more precise minimal block overhead. */
   if (bytes <= 2) return BROTLI_FALSE;
   if (num_commands < (bytes >> 8) + 2) {
-    if (num_literals > 0.99 * (double)bytes) {
+    if ((double)num_literals > 0.99 * (double)bytes) {
       uint32_t literal_histo[256] = { 0 };
       static const uint32_t kSampleRate = 13;
       static const double kMinEntropy = 7.92;
@@ -596,10 +531,14 @@ static void WriteMetaBlockInternal(MemoryManager* m,
       size_t num_literal_contexts = 1;
       const uint32_t* literal_context_map = NULL;
       if (!params->disable_literal_context_modeling) {
+        /* TODO(eustas): pull to higher level and reuse. */
+        uint32_t* arena = BROTLI_ALLOC(m, uint32_t, 14 * 32);
+        if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(arena)) return;
         DecideOverLiteralContextModeling(
             data, wrapped_last_flush_pos, bytes, mask, params->quality,
             params->size_hint, &num_literal_contexts,
-            &literal_context_map);
+            &literal_context_map, arena);
+        BROTLI_FREE(m, arena);
       }
       BrotliBuildMetaBlockGreedy(m, data, wrapped_last_flush_pos, mask,
           prev_byte, prev_byte2, literal_context_lut, num_literal_contexts,
@@ -617,11 +556,7 @@ static void WriteMetaBlockInternal(MemoryManager* m,
       /* The number of distance symbols effectively used for distance
          histograms. It might be less than distance alphabet size
          for "Large Window Brotli" (32-bit). */
-      uint32_t num_effective_dist_codes = block_params.dist.alphabet_size;
-      if (num_effective_dist_codes > BROTLI_NUM_HISTOGRAM_DISTANCE_SYMBOLS) {
-        num_effective_dist_codes = BROTLI_NUM_HISTOGRAM_DISTANCE_SYMBOLS;
-      }
-      BrotliOptimizeHistograms(num_effective_dist_codes, &mb);
+      BrotliOptimizeHistograms(block_params.dist.alphabet_size_limit, &mb);
     }
     BrotliStoreMetaBlock(m, data, wrapped_last_flush_pos, bytes, mask,
                          prev_byte, prev_byte2,
@@ -668,22 +603,34 @@ static void ChooseDistanceParams(BrotliEncoderParams* params) {
     }
   }
 
-  BrotliInitDistanceParams(
-      params, distance_postfix_bits, num_direct_distance_codes);
+  BrotliInitDistanceParams(&params->dist, distance_postfix_bits,
+                           num_direct_distance_codes, params->large_window);
 }
 
 static BROTLI_BOOL EnsureInitialized(BrotliEncoderState* s) {
-  if (BROTLI_IS_OOM(&s->memory_manager_)) return BROTLI_FALSE;
+  MemoryManager* m = &s->memory_manager_;
+  if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
   if (s->is_initialized_) return BROTLI_TRUE;
 
   s->last_bytes_bits_ = 0;
   s->last_bytes_ = 0;
+  s->flint_ = BROTLI_FLINT_DONE;
   s->remaining_metadata_bytes_ = BROTLI_UINT32_MAX;
 
   SanitizeParams(&s->params);
   s->params.lgblock = ComputeLgBlock(&s->params);
   ChooseDistanceParams(&s->params);
 
+  if (s->params.stream_offset != 0) {
+    s->flint_ = BROTLI_FLINT_NEEDS_2_BYTES;
+    /* Poison the distance cache. -16 +- 3 is still less than zero (invalid). */
+    s->dist_cache_[0] = -16;
+    s->dist_cache_[1] = -16;
+    s->dist_cache_[2] = -16;
+    s->dist_cache_[3] = -16;
+    memcpy(s->saved_dist_cache_, s->dist_cache_, sizeof(s->saved_dist_cache_));
+  }
+
   RingBufferSetup(&s->params, &s->ringbuffer_);
 
   /* Initialize last byte with stream header. */
@@ -693,13 +640,23 @@ static BROTLI_BOOL EnsureInitialized(BrotliEncoderState* s) {
         s->params.quality == FAST_TWO_PASS_COMPRESSION_QUALITY) {
       lgwin = BROTLI_MAX(int, lgwin, 18);
     }
-    EncodeWindowBits(lgwin, s->params.large_window,
-                     &s->last_bytes_, &s->last_bytes_bits_);
+    if (s->params.stream_offset == 0) {
+      EncodeWindowBits(lgwin, s->params.large_window,
+                       &s->last_bytes_, &s->last_bytes_bits_);
+    } else {
+      /* Bigger values have the same effect, but could cause overflows. */
+      s->params.stream_offset = BROTLI_MIN(size_t,
+          s->params.stream_offset, BROTLI_MAX_BACKWARD_LIMIT(lgwin));
+    }
   }
 
   if (s->params.quality == FAST_ONE_PASS_COMPRESSION_QUALITY) {
-    InitCommandPrefixCodes(s->cmd_depths_, s->cmd_bits_,
-                           s->cmd_code_, &s->cmd_code_numbits_);
+    s->one_pass_arena_ = BROTLI_ALLOC(m, BrotliOnePassArena, 1);
+    if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
+    InitCommandPrefixCodes(s->one_pass_arena_);
+  } else if (s->params.quality == FAST_TWO_PASS_COMPRESSION_QUALITY) {
+    s->two_pass_arena_ = BROTLI_ALLOC(m, BrotliTwoPassArena, 1);
+    if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
   }
 
   s->is_initialized_ = BROTLI_TRUE;
@@ -712,16 +669,23 @@ static void BrotliEncoderInitParams(BrotliEncoderParams* params) {
   params->quality = BROTLI_DEFAULT_QUALITY;
   params->lgwin = BROTLI_DEFAULT_WINDOW;
   params->lgblock = 0;
+  params->stream_offset = 0;
   params->size_hint = 0;
   params->disable_literal_context_modeling = BROTLI_FALSE;
-  BrotliInitEncoderDictionary(&params->dictionary);
+  BrotliInitSharedEncoderDictionary(&params->dictionary);
   params->dist.distance_postfix_bits = 0;
   params->dist.num_direct_distance_codes = 0;
-  params->dist.alphabet_size =
+  params->dist.alphabet_size_max =
       BROTLI_DISTANCE_ALPHABET_SIZE(0, 0, BROTLI_MAX_DISTANCE_BITS);
+  params->dist.alphabet_size_limit = params->dist.alphabet_size_max;
   params->dist.max_distance = BROTLI_MAX_DISTANCE;
 }
 
+static void BrotliEncoderCleanupParams(MemoryManager* m,
+    BrotliEncoderParams* params) {
+  BrotliCleanupSharedEncoderDictionary(m, &params->dictionary);
+}
+
 static void BrotliEncoderInitState(BrotliEncoderState* s) {
   BrotliEncoderInitParams(&s->params);
   s->input_pos_ = 0;
@@ -734,12 +698,14 @@ static void BrotliEncoderInitState(BrotliEncoderState* s) {
   s->prev_byte2_ = 0;
   s->storage_size_ = 0;
   s->storage_ = 0;
-  s->hasher_ = NULL;
+  HasherInit(&s->hasher_);
   s->large_table_ = NULL;
   s->large_table_size_ = 0;
-  s->cmd_code_numbits_ = 0;
+  s->one_pass_arena_ = NULL;
+  s->two_pass_arena_ = NULL;
   s->command_buf_ = NULL;
   s->literal_buf_ = NULL;
+  s->total_in_ = 0;
   s->next_out_ = NULL;
   s->available_out_ = 0;
   s->total_out_ = 0;
@@ -764,13 +730,9 @@ static void BrotliEncoderInitState(BrotliEncoderState* s) {
 
 BrotliEncoderState* BrotliEncoderCreateInstance(
     brotli_alloc_func alloc_func, brotli_free_func free_func, void* opaque) {
-  BrotliEncoderState* state = 0;
-  if (!alloc_func && !free_func) {
-    state = (BrotliEncoderState*)malloc(sizeof(BrotliEncoderState));
-  } else if (alloc_func && free_func) {
-    state = (BrotliEncoderState*)alloc_func(opaque, sizeof(BrotliEncoderState));
-  }
-  if (state == 0) {
+  BrotliEncoderState* state = (BrotliEncoderState*)BrotliBootstrapAlloc(
+      sizeof(BrotliEncoderState), alloc_func, free_func, opaque);
+  if (state == NULL) {
     /* BROTLI_DUMP(); */
     return 0;
   }
@@ -780,19 +742,36 @@ BrotliEncoderState* BrotliEncoderCreateInstance(
   return state;
 }
 
+#ifdef BROTLI_REPORTING
+/* When BROTLI_REPORTING is defined extra reporting module have to be linked. */
+void BrotliEncoderOnFinish(const BrotliEncoderState* s);
+#define BROTLI_ENCODER_ON_FINISH(s) BrotliEncoderOnFinish(s);
+#else
+#if !defined(BROTLI_ENCODER_ON_FINISH)
+#define BROTLI_ENCODER_ON_FINISH(s) (void)(s);
+#endif
+#endif
+
 static void BrotliEncoderCleanupState(BrotliEncoderState* s) {
   MemoryManager* m = &s->memory_manager_;
+
+  BROTLI_ENCODER_ON_FINISH(s);
+
   if (BROTLI_IS_OOM(m)) {
     BrotliWipeOutMemoryManager(m);
     return;
   }
+
   BROTLI_FREE(m, s->storage_);
   BROTLI_FREE(m, s->commands_);
   RingBufferFree(m, &s->ringbuffer_);
   DestroyHasher(m, &s->hasher_);
   BROTLI_FREE(m, s->large_table_);
+  BROTLI_FREE(m, s->one_pass_arena_);
+  BROTLI_FREE(m, s->two_pass_arena_);
   BROTLI_FREE(m, s->command_buf_);
   BROTLI_FREE(m, s->literal_buf_);
+  BrotliEncoderCleanupParams(m, &s->params);
 }
 
 /* Deinitializes and frees BrotliEncoderState instance. */
@@ -800,11 +779,8 @@ void BrotliEncoderDestroyInstance(BrotliEncoderState* state) {
   if (!state) {
     return;
   } else {
-    MemoryManager* m = &state->memory_manager_;
-    brotli_free_func free_func = m->free_func;
-    void* opaque = m->opaque;
     BrotliEncoderCleanupState(state);
-    free_func(opaque, state);
+    BrotliBootstrapFree(state, &state->memory_manager_);
   }
 }
 
@@ -893,6 +869,8 @@ static void ExtendLastCommand(BrotliEncoderState* s, uint32_t* bytes,
   uint64_t cmd_dist = (uint64_t)s->dist_cache_[0];
   uint32_t distance_code = CommandRestoreDistanceCode(last_command,
                                                       &s->params.dist);
+  const CompoundDictionary* dict = &s->params.dictionary.compound;
+  size_t compound_dictionary_size = dict->total_size;
   if (distance_code < BROTLI_NUM_DISTANCE_SHORT_CODES ||
       distance_code - (BROTLI_NUM_DISTANCE_SHORT_CODES - 1) == cmd_dist) {
     if (cmd_dist <= max_distance) {
@@ -902,6 +880,39 @@ static void ExtendLastCommand(BrotliEncoderState* s, uint32_t* bytes,
         (*bytes)--;
         (*wrapped_last_processed_pos)++;
       }
+    } else {
+      if ((cmd_dist - max_distance - 1) < compound_dictionary_size &&
+          last_copy_len < cmd_dist - max_distance) {
+        size_t address =
+            compound_dictionary_size - (size_t)(cmd_dist - max_distance) +
+            (size_t)last_copy_len;
+        size_t br_index = 0;
+        size_t br_offset;
+        const uint8_t* chunk;
+        size_t chunk_length;
+        while (address >= dict->chunk_offsets[br_index + 1]) br_index++;
+        br_offset = address - dict->chunk_offsets[br_index];
+        chunk = dict->chunk_source[br_index];
+        chunk_length =
+            dict->chunk_offsets[br_index + 1] - dict->chunk_offsets[br_index];
+        while (*bytes != 0 && data[*wrapped_last_processed_pos & mask] ==
+               chunk[br_offset]) {
+          last_command->copy_len_++;
+          (*bytes)--;
+          (*wrapped_last_processed_pos)++;
+          if (++br_offset == chunk_length) {
+            br_index++;
+            br_offset = 0;
+            if (br_index != dict->num_chunks) {
+              chunk = dict->chunk_source[br_index];
+              chunk_length = dict->chunk_offsets[br_index + 1] -
+                  dict->chunk_offsets[br_index];
+            } else {
+              break;
+            }
+          }
+        }
+      }
     }
     /* The copy length is at most the metablock size, and thus expressible. */
     GetLengthCode(last_command->insert_len_,
@@ -919,8 +930,8 @@ static void ExtendLastCommand(BrotliEncoderState* s, uint32_t* bytes,
    If |*out_size| is positive, |*output| points to the start of the output
    data. If |is_last| or |force_flush| is BROTLI_TRUE, an output meta-block is
    always created. However, until |is_last| is BROTLI_TRUE encoder may retain up
-   to 7 bits of the last byte of output. To force encoder to dump the remaining
-   bits use WriteMetadata() to append an empty meta-data block.
+   to 7 bits of the last byte of output. Byte-alignment could be enforced by
+   emitting an empty meta-data block.
    Returns BROTLI_FALSE if the size of the input data is larger than
    input_block_size().
  */
@@ -934,10 +945,40 @@ static BROTLI_BOOL EncodeData(
   uint32_t mask;
   MemoryManager* m = &s->memory_manager_;
   ContextType literal_context_mode;
+  ContextLut literal_context_lut;
+  BROTLI_BOOL fast_compress =
+      s->params.quality == FAST_ONE_PASS_COMPRESSION_QUALITY ||
+      s->params.quality == FAST_TWO_PASS_COMPRESSION_QUALITY;
 
   data = s->ringbuffer_.buffer_;
   mask = s->ringbuffer_.mask_;
 
+  if (delta == 0) {  /* No new input; still might want to flush or finish. */
+    if (!data) {  /* No input has been processed so far. */
+      if (is_last) {  /* Emit complete finalized stream. */
+        BROTLI_DCHECK(s->last_bytes_bits_ <= 14);
+        s->last_bytes_ |= (uint16_t)(3u << s->last_bytes_bits_);
+        s->last_bytes_bits_ = (uint8_t)(s->last_bytes_bits_ + 2u);
+        s->tiny_buf_.u8[0] = (uint8_t)s->last_bytes_;
+        s->tiny_buf_.u8[1] = (uint8_t)(s->last_bytes_ >> 8);
+        *output = s->tiny_buf_.u8;
+        *out_size = (s->last_bytes_bits_ + 7u) >> 3u;
+        return BROTLI_TRUE;
+      } else {  /* No data, not last -> no-op. */
+        *out_size = 0;
+        return BROTLI_TRUE;
+      }
+    } else {
+      /* Fast compress performs flush every block -> flush is no-op. */
+      if (!is_last && (!force_flush || fast_compress)) {  /* Another no-op. */
+        *out_size = 0;
+        return BROTLI_TRUE;
+      }
+    }
+  }
+  BROTLI_DCHECK(data);
+
+  if (s->params.quality > s->params.dictionary.max_quality) return BROTLI_FALSE;
   /* Adding more blocks after "last" block is forbidden. */
   if (s->is_last_block_emitted_) return BROTLI_FALSE;
   if (is_last) s->is_last_block_emitted_ = BROTLI_TRUE;
@@ -951,22 +992,18 @@ static BROTLI_BOOL EncodeData(
         BROTLI_ALLOC(m, uint32_t, kCompressFragmentTwoPassBlockSize);
     s->literal_buf_ =
         BROTLI_ALLOC(m, uint8_t, kCompressFragmentTwoPassBlockSize);
-    if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
+    if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(s->command_buf_) ||
+        BROTLI_IS_NULL(s->literal_buf_)) {
+      return BROTLI_FALSE;
+    }
   }
 
-  if (s->params.quality == FAST_ONE_PASS_COMPRESSION_QUALITY ||
-      s->params.quality == FAST_TWO_PASS_COMPRESSION_QUALITY) {
+  if (fast_compress) {
     uint8_t* storage;
     size_t storage_ix = s->last_bytes_bits_;
     size_t table_size;
     int* table;
 
-    if (delta == 0 && !is_last) {
-      /* We have no new input data and we don't have to finish the stream, so
-         nothing to do. */
-      *out_size = 0;
-      return BROTLI_TRUE;
-    }
     storage = GetBrotliStorage(s, 2 * bytes + 503);
     if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
     storage[0] = (uint8_t)s->last_bytes_;
@@ -975,16 +1012,14 @@ static BROTLI_BOOL EncodeData(
     if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
     if (s->params.quality == FAST_ONE_PASS_COMPRESSION_QUALITY) {
       BrotliCompressFragmentFast(
-          m, &data[wrapped_last_processed_pos & mask],
+          s->one_pass_arena_, &data[wrapped_last_processed_pos & mask],
           bytes, is_last,
           table, table_size,
-          s->cmd_depths_, s->cmd_bits_,
-          &s->cmd_code_numbits_, s->cmd_code_,
           &storage_ix, storage);
       if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
     } else {
       BrotliCompressFragmentTwoPass(
-          m, &data[wrapped_last_processed_pos & mask],
+          s->two_pass_arena_, &data[wrapped_last_processed_pos & mask],
           bytes, is_last,
           s->command_buf_, s->literal_buf_,
           table, table_size,
@@ -1009,7 +1044,7 @@ static BROTLI_BOOL EncodeData(
       newsize += (bytes / 4) + 16;
       s->cmd_alloc_size_ = newsize;
       new_commands = BROTLI_ALLOC(m, Command, newsize);
-      if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
+      if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(new_commands)) return BROTLI_FALSE;
       if (s->commands_) {
         memcpy(new_commands, s->commands_, sizeof(Command) * s->num_commands_);
         BROTLI_FREE(m, s->commands_);
@@ -1024,6 +1059,7 @@ static BROTLI_BOOL EncodeData(
   literal_context_mode = ChooseContextMode(
       &s->params, data, WrapPosition(s->last_flush_pos_),
       mask, (size_t)(s->input_pos_ - s->last_flush_pos_));
+  literal_context_lut = BROTLI_CONTEXT_LUT(literal_context_mode);
 
   if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
 
@@ -1034,20 +1070,23 @@ static BROTLI_BOOL EncodeData(
   if (s->params.quality == ZOPFLIFICATION_QUALITY) {
     BROTLI_DCHECK(s->params.hasher.type == 10);
     BrotliCreateZopfliBackwardReferences(m, bytes, wrapped_last_processed_pos,
-        data, mask, &s->params, s->hasher_, s->dist_cache_,
+        data, mask, literal_context_lut, &s->params,
+        &s->hasher_, s->dist_cache_,
         &s->last_insert_len_, &s->commands_[s->num_commands_],
         &s->num_commands_, &s->num_literals_);
     if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
   } else if (s->params.quality == HQ_ZOPFLIFICATION_QUALITY) {
     BROTLI_DCHECK(s->params.hasher.type == 10);
     BrotliCreateHqZopfliBackwardReferences(m, bytes, wrapped_last_processed_pos,
-        data, mask, &s->params, s->hasher_, s->dist_cache_,
+        data, mask, literal_context_lut, &s->params,
+        &s->hasher_, s->dist_cache_,
         &s->last_insert_len_, &s->commands_[s->num_commands_],
         &s->num_commands_, &s->num_literals_);
     if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
   } else {
     BrotliCreateBackwardReferences(bytes, wrapped_last_processed_pos,
-        data, mask, &s->params, s->hasher_, s->dist_cache_,
+        data, mask, literal_context_lut, &s->params,
+        &s->hasher_, s->dist_cache_,
         &s->last_insert_len_, &s->commands_[s->num_commands_],
         &s->num_commands_, &s->num_literals_);
   }
@@ -1058,7 +1097,7 @@ static BROTLI_BOOL EncodeData(
     const size_t max_commands = max_length / 8;
     const size_t processed_bytes = (size_t)(s->input_pos_ - s->last_flush_pos_);
     /* If maximal possible additional block doesn't fit metablock, flush now. */
-    /* TODO: Postpone decision until next block arrives? */
+    /* TODO(eustas): Postpone decision until next block arrives? */
     const BROTLI_BOOL next_input_fits_metablock = TO_BROTLI_BOOL(
         processed_bytes + InputBlockSize(s) <= max_length);
     /* If block splitting is not used, then flush as soon as there is some
@@ -1072,7 +1111,7 @@ static BROTLI_BOOL EncodeData(
         s->num_commands_ < max_commands) {
       /* Merge with next input block. Everything will happen later. */
       if (UpdateLastProcessedPos(s)) {
-        HasherReset(s->hasher_);
+        HasherReset(&s->hasher_);
       }
       *out_size = 0;
       return BROTLI_TRUE;
@@ -1113,7 +1152,7 @@ static BROTLI_BOOL EncodeData(
     s->last_bytes_bits_ = storage_ix & 7u;
     s->last_flush_pos_ = s->input_pos_;
     if (UpdateLastProcessedPos(s)) {
-      HasherReset(s->hasher_);
+      HasherReset(&s->hasher_);
     }
     if (s->last_flush_pos_ > 0) {
       s->prev_byte_ = data[((uint32_t)s->last_flush_pos_ - 1) & mask];
@@ -1151,7 +1190,7 @@ static size_t WriteMetadataHeader(
   if (block_size == 0) {
     BrotliWriteBits(2, 0, &storage_ix, header);
   } else {
-    uint32_t nbits = (block_size == 1) ? 0 :
+    uint32_t nbits = (block_size == 1) ? 1 :
         (Log2FloorNonZero((uint32_t)block_size - 1) + 1);
     uint32_t nbytes = (nbits + 7) / 8;
     BrotliWriteBits(2, nbytes, &storage_ix, header);
@@ -1160,229 +1199,6 @@ static size_t WriteMetadataHeader(
   return (storage_ix + 7u) >> 3;
 }
 
-static BROTLI_BOOL BrotliCompressBufferQuality10(
-    int lgwin, size_t input_size, const uint8_t* input_buffer,
-    size_t* encoded_size, uint8_t* encoded_buffer) {
-  MemoryManager memory_manager;
-  MemoryManager* m = &memory_manager;
-
-  const size_t mask = BROTLI_SIZE_MAX >> 1;
-  int dist_cache[4] = { 4, 11, 15, 16 };
-  int saved_dist_cache[4] = { 4, 11, 15, 16 };
-  BROTLI_BOOL ok = BROTLI_TRUE;
-  const size_t max_out_size = *encoded_size;
-  size_t total_out_size = 0;
-  uint16_t last_bytes;
-  uint8_t last_bytes_bits;
-  HasherHandle hasher = NULL;
-
-  const size_t hasher_eff_size = BROTLI_MIN(size_t,
-      input_size, BROTLI_MAX_BACKWARD_LIMIT(lgwin) + BROTLI_WINDOW_GAP);
-
-  BrotliEncoderParams params;
-
-  const int lgmetablock = BROTLI_MIN(int, 24, lgwin + 1);
-  size_t max_block_size;
-  const size_t max_metablock_size = (size_t)1 << lgmetablock;
-  const size_t max_literals_per_metablock = max_metablock_size / 8;
-  const size_t max_commands_per_metablock = max_metablock_size / 8;
-  size_t metablock_start = 0;
-  uint8_t prev_byte = 0;
-  uint8_t prev_byte2 = 0;
-
-  BrotliEncoderInitParams(&params);
-  params.quality = 10;
-  params.lgwin = lgwin;
-  if (lgwin > BROTLI_MAX_WINDOW_BITS) {
-    params.large_window = BROTLI_TRUE;
-  }
-  SanitizeParams(&params);
-  params.lgblock = ComputeLgBlock(&params);
-  ChooseDistanceParams(&params);
-  max_block_size = (size_t)1 << params.lgblock;
-
-  BrotliInitMemoryManager(m, 0, 0, 0);
-
-  BROTLI_DCHECK(input_size <= mask + 1);
-  EncodeWindowBits(lgwin, params.large_window, &last_bytes, &last_bytes_bits);
-  InitOrStitchToPreviousBlock(m, &hasher, input_buffer, mask, &params,
-      0, hasher_eff_size, BROTLI_TRUE);
-  if (BROTLI_IS_OOM(m)) goto oom;
-
-  while (ok && metablock_start < input_size) {
-    const size_t metablock_end =
-        BROTLI_MIN(size_t, input_size, metablock_start + max_metablock_size);
-    const size_t expected_num_commands =
-        (metablock_end - metablock_start) / 12 + 16;
-    Command* commands = 0;
-    size_t num_commands = 0;
-    size_t last_insert_len = 0;
-    size_t num_literals = 0;
-    size_t metablock_size = 0;
-    size_t cmd_alloc_size = 0;
-    BROTLI_BOOL is_last;
-    uint8_t* storage;
-    size_t storage_ix;
-
-    ContextType literal_context_mode = ChooseContextMode(&params,
-        input_buffer, metablock_start, mask, metablock_end - metablock_start);
-
-    size_t block_start;
-    for (block_start = metablock_start; block_start < metablock_end; ) {
-      size_t block_size =
-          BROTLI_MIN(size_t, metablock_end - block_start, max_block_size);
-      ZopfliNode* nodes = BROTLI_ALLOC(m, ZopfliNode, block_size + 1);
-      size_t path_size;
-      size_t new_cmd_alloc_size;
-      if (BROTLI_IS_OOM(m)) goto oom;
-      BrotliInitZopfliNodes(nodes, block_size + 1);
-      StitchToPreviousBlockH10(hasher, block_size, block_start,
-                               input_buffer, mask);
-      path_size = BrotliZopfliComputeShortestPath(m, block_size, block_start,
-          input_buffer, mask, &params, dist_cache, hasher,
-          nodes);
-      if (BROTLI_IS_OOM(m)) goto oom;
-      /* We allocate a command buffer in the first iteration of this loop that
-         will be likely big enough for the whole metablock, so that for most
-         inputs we will not have to reallocate in later iterations. We do the
-         allocation here and not before the loop, because if the input is small,
-         this will be allocated after the Zopfli cost model is freed, so this
-         will not increase peak memory usage.
-         TODO: If the first allocation is too small, increase command
-         buffer size exponentially. */
-      new_cmd_alloc_size = BROTLI_MAX(size_t, expected_num_commands,
-                                      num_commands + path_size + 1);
-      if (cmd_alloc_size != new_cmd_alloc_size) {
-        Command* new_commands = BROTLI_ALLOC(m, Command, new_cmd_alloc_size);
-        if (BROTLI_IS_OOM(m)) goto oom;
-        cmd_alloc_size = new_cmd_alloc_size;
-        if (commands) {
-          memcpy(new_commands, commands, sizeof(Command) * num_commands);
-          BROTLI_FREE(m, commands);
-        }
-        commands = new_commands;
-      }
-      BrotliZopfliCreateCommands(block_size, block_start, &nodes[0], dist_cache,
-          &last_insert_len, &params, &commands[num_commands], &num_literals);
-      num_commands += path_size;
-      block_start += block_size;
-      metablock_size += block_size;
-      BROTLI_FREE(m, nodes);
-      if (num_literals > max_literals_per_metablock ||
-          num_commands > max_commands_per_metablock) {
-        break;
-      }
-    }
-
-    if (last_insert_len > 0) {
-      InitInsertCommand(&commands[num_commands++], last_insert_len);
-      num_literals += last_insert_len;
-    }
-
-    is_last = TO_BROTLI_BOOL(metablock_start + metablock_size == input_size);
-    storage = NULL;
-    storage_ix = last_bytes_bits;
-
-    if (metablock_size == 0) {
-      /* Write the ISLAST and ISEMPTY bits. */
-      storage = BROTLI_ALLOC(m, uint8_t, 16);
-      if (BROTLI_IS_OOM(m)) goto oom;
-      storage[0] = (uint8_t)last_bytes;
-      storage[1] = (uint8_t)(last_bytes >> 8);
-      BrotliWriteBits(2, 3, &storage_ix, storage);
-      storage_ix = (storage_ix + 7u) & ~7u;
-    } else if (!ShouldCompress(input_buffer, mask, metablock_start,
-                               metablock_size, num_literals, num_commands)) {
-      /* Restore the distance cache, as its last update by
-         CreateBackwardReferences is now unused. */
-      memcpy(dist_cache, saved_dist_cache, 4 * sizeof(dist_cache[0]));
-      storage = BROTLI_ALLOC(m, uint8_t, metablock_size + 16);
-      if (BROTLI_IS_OOM(m)) goto oom;
-      storage[0] = (uint8_t)last_bytes;
-      storage[1] = (uint8_t)(last_bytes >> 8);
-      BrotliStoreUncompressedMetaBlock(is_last, input_buffer,
-                                       metablock_start, mask, metablock_size,
-                                       &storage_ix, storage);
-    } else {
-      MetaBlockSplit mb;
-      BrotliEncoderParams block_params = params;
-      InitMetaBlockSplit(&mb);
-      BrotliBuildMetaBlock(m, input_buffer, metablock_start, mask,
-                           &block_params,
-                           prev_byte, prev_byte2,
-                           commands, num_commands,
-                           literal_context_mode,
-                           &mb);
-      if (BROTLI_IS_OOM(m)) goto oom;
-      {
-        /* The number of distance symbols effectively used for distance
-           histograms. It might be less than distance alphabet size
-           for "Large Window Brotli" (32-bit). */
-        uint32_t num_effective_dist_codes = block_params.dist.alphabet_size;
-        if (num_effective_dist_codes > BROTLI_NUM_HISTOGRAM_DISTANCE_SYMBOLS) {
-          num_effective_dist_codes = BROTLI_NUM_HISTOGRAM_DISTANCE_SYMBOLS;
-        }
-        BrotliOptimizeHistograms(num_effective_dist_codes, &mb);
-      }
-      storage = BROTLI_ALLOC(m, uint8_t, 2 * metablock_size + 503);
-      if (BROTLI_IS_OOM(m)) goto oom;
-      storage[0] = (uint8_t)last_bytes;
-      storage[1] = (uint8_t)(last_bytes >> 8);
-      BrotliStoreMetaBlock(m, input_buffer, metablock_start, metablock_size,
-                           mask, prev_byte, prev_byte2,
-                           is_last,
-                           &block_params,
-                           literal_context_mode,
-                           commands, num_commands,
-                           &mb,
-                           &storage_ix, storage);
-      if (BROTLI_IS_OOM(m)) goto oom;
-      if (metablock_size + 4 < (storage_ix >> 3)) {
-        /* Restore the distance cache and last byte. */
-        memcpy(dist_cache, saved_dist_cache, 4 * sizeof(dist_cache[0]));
-        storage[0] = (uint8_t)last_bytes;
-        storage[1] = (uint8_t)(last_bytes >> 8);
-        storage_ix = last_bytes_bits;
-        BrotliStoreUncompressedMetaBlock(is_last, input_buffer,
-                                         metablock_start, mask,
-                                         metablock_size, &storage_ix, storage);
-      }
-      DestroyMetaBlockSplit(m, &mb);
-    }
-    last_bytes = (uint16_t)(storage[storage_ix >> 3]);
-    last_bytes_bits = storage_ix & 7u;
-    metablock_start += metablock_size;
-    if (metablock_start < input_size) {
-      prev_byte = input_buffer[metablock_start - 1];
-      prev_byte2 = input_buffer[metablock_start - 2];
-    }
-    /* Save the state of the distance cache in case we need to restore it for
-       emitting an uncompressed block. */
-    memcpy(saved_dist_cache, dist_cache, 4 * sizeof(dist_cache[0]));
-
-    {
-      const size_t out_size = storage_ix >> 3;
-      total_out_size += out_size;
-      if (total_out_size <= max_out_size) {
-        memcpy(encoded_buffer, storage, out_size);
-        encoded_buffer += out_size;
-      } else {
-        ok = BROTLI_FALSE;
-      }
-    }
-    BROTLI_FREE(m, storage);
-    BROTLI_FREE(m, commands);
-  }
-
-  *encoded_size = total_out_size;
-  DestroyHasher(m, &hasher);
-  return ok;
-
-oom:
-  BrotliWipeOutMemoryManager(m);
-  return BROTLI_FALSE;
-}
-
 size_t BrotliEncoderMaxCompressedSize(size_t input_size) {
   /* [window bits / empty metadata] + N * [uncompressed] + [last empty] */
   size_t num_large_blocks = input_size >> 14;
@@ -1430,8 +1246,9 @@ static size_t MakeUncompressedStream(
 
 BROTLI_BOOL BrotliEncoderCompress(
     int quality, int lgwin, BrotliEncoderMode mode, size_t input_size,
-    const uint8_t* input_buffer, size_t* encoded_size,
-    uint8_t* encoded_buffer) {
+    const uint8_t input_buffer[BROTLI_ARRAY_PARAM(input_size)],
+    size_t* encoded_size,
+    uint8_t encoded_buffer[BROTLI_ARRAY_PARAM(*encoded_size)]) {
   BrotliEncoderState* s;
   size_t out_size = *encoded_size;
   const uint8_t* input_start = input_buffer;
@@ -1447,17 +1264,6 @@ BROTLI_BOOL BrotliEncoderCompress(
     *encoded_buffer = 6;
     return BROTLI_TRUE;
   }
-  if (quality == 10) {
-    /* TODO: Implement this direct path for all quality levels. */
-    const int lg_win = BROTLI_MIN(int, BROTLI_LARGE_MAX_WINDOW_BITS,
-                                       BROTLI_MAX(int, 16, lgwin));
-    int ok = BrotliCompressBufferQuality10(lg_win, input_size, input_buffer,
-                                           encoded_size, encoded_buffer);
-    if (!ok || (max_out_size && *encoded_size > max_out_size)) {
-      goto fallback;
-    }
-    return BROTLI_TRUE;
-  }
 
   s = BrotliEncoderCreateInstance(0, 0, 0);
   if (!s) {
@@ -1469,6 +1275,7 @@ BROTLI_BOOL BrotliEncoderCompress(
     uint8_t* next_out = encoded_buffer;
     size_t total_out = 0;
     BROTLI_BOOL result = BROTLI_FALSE;
+    /* TODO(eustas): check that parameters are sane. */
     BrotliEncoderSetParameter(s, BROTLI_PARAM_QUALITY, (uint32_t)quality);
     BrotliEncoderSetParameter(s, BROTLI_PARAM_LGWIN, (uint32_t)lgwin);
     BrotliEncoderSetParameter(s, BROTLI_PARAM_MODE, (uint32_t)mode);
@@ -1520,6 +1327,18 @@ static void InjectBytePaddingBlock(BrotliEncoderState* s) {
   s->available_out_ += (seal_bits + 7) >> 3;
 }
 
+/* Fills the |total_out|, if it is not NULL. */
+static void SetTotalOut(BrotliEncoderState* s, size_t* total_out) {
+  if (total_out) {
+    /* Saturating conversion uint64_t -> size_t */
+    size_t result = (size_t)-1;
+    if (s->total_out_ < result) {
+      result = (size_t)s->total_out_;
+    }
+    *total_out = result;
+  }
+}
+
 /* Injects padding bits or pushes compressed data to output.
    Returns false if nothing is done. */
 static BROTLI_BOOL InjectFlushOrPushOutput(BrotliEncoderState* s,
@@ -1539,7 +1358,7 @@ static BROTLI_BOOL InjectFlushOrPushOutput(BrotliEncoderState* s,
     s->next_out_ += copy_output_size;
     s->available_out_ -= copy_output_size;
     s->total_out_ += copy_output_size;
-    if (total_out) *total_out = s->total_out_;
+    SetTotalOut(s, total_out);
     return BROTLI_TRUE;
   }
 
@@ -1576,7 +1395,10 @@ static BROTLI_BOOL BrotliEncoderCompressStreamFast(
           BROTLI_ALLOC(m, uint32_t, kCompressFragmentTwoPassBlockSize);
       s->literal_buf_ =
           BROTLI_ALLOC(m, uint8_t, kCompressFragmentTwoPassBlockSize);
-      if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
+      if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(s->command_buf_) ||
+          BROTLI_IS_NULL(s->literal_buf_)) {
+        return BROTLI_FALSE;
+      }
     }
     if (s->command_buf_) {
       command_buf = s->command_buf_;
@@ -1584,7 +1406,10 @@ static BROTLI_BOOL BrotliEncoderCompressStreamFast(
     } else {
       tmp_command_buf = BROTLI_ALLOC(m, uint32_t, buf_size);
       tmp_literal_buf = BROTLI_ALLOC(m, uint8_t, buf_size);
-      if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
+      if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(tmp_command_buf) ||
+          BROTLI_IS_NULL(tmp_literal_buf)) {
+        return BROTLI_FALSE;
+      }
       command_buf = tmp_command_buf;
       literal_buf = tmp_literal_buf;
     }
@@ -1630,18 +1455,20 @@ static BROTLI_BOOL BrotliEncoderCompressStreamFast(
       if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
 
       if (s->params.quality == FAST_ONE_PASS_COMPRESSION_QUALITY) {
-        BrotliCompressFragmentFast(m, *next_in, block_size, is_last, table,
-            table_size, s->cmd_depths_, s->cmd_bits_, &s->cmd_code_numbits_,
-            s->cmd_code_, &storage_ix, storage);
+        BrotliCompressFragmentFast(s->one_pass_arena_, *next_in, block_size,
+            is_last, table, table_size, &storage_ix, storage);
         if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
       } else {
-        BrotliCompressFragmentTwoPass(m, *next_in, block_size, is_last,
-            command_buf, literal_buf, table, table_size,
+        BrotliCompressFragmentTwoPass(s->two_pass_arena_, *next_in, block_size,
+            is_last, command_buf, literal_buf, table, table_size,
             &storage_ix, storage);
         if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
       }
-      *next_in += block_size;
-      *available_in -= block_size;
+      if (block_size != 0) {
+        *next_in += block_size;
+        *available_in -= block_size;
+        s->total_in_ += block_size;
+      }
       if (inplace) {
         size_t out_bytes = storage_ix >> 3;
         BROTLI_DCHECK(out_bytes <= *available_out);
@@ -1649,7 +1476,7 @@ static BROTLI_BOOL BrotliEncoderCompressStreamFast(
         *next_out += out_bytes;
         *available_out -= out_bytes;
         s->total_out_ += out_bytes;
-        if (total_out) *total_out = s->total_out_;
+        SetTotalOut(s, total_out);
       } else {
         size_t out_bytes = storage_ix >> 3;
         s->next_out_ = storage;
@@ -1718,6 +1545,7 @@ static BROTLI_BOOL ProcessMetadata(
         memcpy(*next_out, *next_in, copy);
         *next_in += copy;
         *available_in -= copy;
+        s->total_in_ += copy;  /* not actually data input, though */
         s->remaining_metadata_bytes_ -= copy;
         *next_out += copy;
         *available_out -= copy;
@@ -1728,6 +1556,7 @@ static BROTLI_BOOL ProcessMetadata(
         memcpy(s->next_out_, *next_in, copy);
         *next_in += copy;
         *available_in -= copy;
+        s->total_in_ += copy;  /* not actually data input, though */
         s->remaining_metadata_bytes_ -= copy;
         s->available_out_ = copy;
       }
@@ -1755,7 +1584,7 @@ static void UpdateSizeHint(BrotliEncoderState* s, size_t available_in) {
 
 BROTLI_BOOL BrotliEncoderCompressStream(
     BrotliEncoderState* s, BrotliEncoderOperation op, size_t* available_in,
-    const uint8_t** next_in, size_t* available_out,uint8_t** next_out,
+    const uint8_t** next_in, size_t* available_out, uint8_t** next_out,
     size_t* total_out) {
   if (!EnsureInitialized(s)) return BROTLI_FALSE;
 
@@ -1786,6 +1615,10 @@ BROTLI_BOOL BrotliEncoderCompressStream(
   }
   while (BROTLI_TRUE) {
     size_t remaining_block_size = RemainingInputBlockSize(s);
+    /* Shorten input to flint size. */
+    if (s->flint_ >= 0 && remaining_block_size > (size_t)s->flint_) {
+      remaining_block_size = (size_t)s->flint_;
+    }
 
     if (remaining_block_size != 0 && *available_in != 0) {
       size_t copy_input_size =
@@ -1793,10 +1626,19 @@ BROTLI_BOOL BrotliEncoderCompressStream(
       CopyInputToRingBuffer(s, copy_input_size, *next_in);
       *next_in += copy_input_size;
       *available_in -= copy_input_size;
+      s->total_in_ += copy_input_size;
+      if (s->flint_ > 0) s->flint_ = (int8_t)(s->flint_ - (int)copy_input_size);
       continue;
     }
 
     if (InjectFlushOrPushOutput(s, available_out, next_out, total_out)) {
+      /* Exit the "emit flint" workflow. */
+      if (s->flint_ == BROTLI_FLINT_WAITING_FOR_FLUSHING) {
+        CheckFlushComplete(s);
+        if (s->stream_state_ == BROTLI_STREAM_PROCESSING) {
+          s->flint_ = BROTLI_FLINT_DONE;
+        }
+      }
       continue;
     }
 
@@ -1810,6 +1652,11 @@ BROTLI_BOOL BrotliEncoderCompressStream(
         BROTLI_BOOL force_flush = TO_BROTLI_BOOL(
             (*available_in == 0) && op == BROTLI_OPERATION_FLUSH);
         BROTLI_BOOL result;
+        /* Force emitting (uncompressed) piece containing flint. */
+        if (!is_last && s->flint_ == 0) {
+          s->flint_ = BROTLI_FLINT_WAITING_FOR_FLUSHING;
+          force_flush = BROTLI_TRUE;
+        }
         UpdateSizeHint(s, *available_in);
         result = EncodeData(s, is_last, force_flush,
             &s->available_out_, &s->next_out_);
@@ -1857,6 +1704,293 @@ uint32_t BrotliEncoderVersion(void) {
   return BROTLI_VERSION;
 }
 
+BrotliEncoderPreparedDictionary* BrotliEncoderPrepareDictionary(
+    BrotliSharedDictionaryType type, size_t size,
+    const uint8_t data[BROTLI_ARRAY_PARAM(size)], int quality,
+    brotli_alloc_func alloc_func, brotli_free_func free_func, void* opaque) {
+  ManagedDictionary* managed_dictionary = NULL;
+  BROTLI_BOOL type_is_known = BROTLI_FALSE;
+  type_is_known |= (type == BROTLI_SHARED_DICTIONARY_RAW);
+#if defined(BROTLI_EXPERIMENTAL)
+  type_is_known |= (type == BROTLI_SHARED_DICTIONARY_SERIALIZED);
+#endif  /* BROTLI_EXPERIMENTAL */
+  if (!type_is_known) {
+    return NULL;
+  }
+  managed_dictionary =
+      BrotliCreateManagedDictionary(alloc_func, free_func, opaque);
+  if (managed_dictionary == NULL) {
+    return NULL;
+  }
+  if (type == BROTLI_SHARED_DICTIONARY_RAW) {
+    managed_dictionary->dictionary = (uint32_t*)CreatePreparedDictionary(
+        &managed_dictionary->memory_manager_, data, size);
+  }
+#if defined(BROTLI_EXPERIMENTAL)
+  if (type == BROTLI_SHARED_DICTIONARY_SERIALIZED) {
+    SharedEncoderDictionary* dict = (SharedEncoderDictionary*)BrotliAllocate(
+        &managed_dictionary->memory_manager_, sizeof(SharedEncoderDictionary));
+    managed_dictionary->dictionary = (uint32_t*)dict;
+    if (dict != NULL) {
+      BROTLI_BOOL ok = BrotliInitCustomSharedEncoderDictionary(
+          &managed_dictionary->memory_manager_, data, size, quality, dict);
+      if (!ok) {
+        BrotliFree(&managed_dictionary->memory_manager_, dict);
+        managed_dictionary->dictionary = NULL;
+      }
+    }
+  }
+#else  /* BROTLI_EXPERIMENTAL */
+  (void)quality;
+#endif  /* BROTLI_EXPERIMENTAL */
+  if (managed_dictionary->dictionary == NULL) {
+    BrotliDestroyManagedDictionary(managed_dictionary);
+    return NULL;
+  }
+  return (BrotliEncoderPreparedDictionary*)managed_dictionary;
+}
+
+void BrotliEncoderDestroyPreparedDictionary(
+    BrotliEncoderPreparedDictionary* dictionary) {
+  ManagedDictionary* dict = (ManagedDictionary*)dictionary;
+  if (!dictionary) return;
+  /* First field of dictionary structs. */
+  /* Only managed dictionaries are eligible for destruction by this method. */
+  if (dict->magic != kManagedDictionaryMagic) {
+    return;
+  }
+  if (dict->dictionary == NULL) {
+    /* This should never ever happen. */
+  } else if (*dict->dictionary == kLeanPreparedDictionaryMagic) {
+    DestroyPreparedDictionary(
+        &dict->memory_manager_, (PreparedDictionary*)dict->dictionary);
+  } else if (*dict->dictionary == kSharedDictionaryMagic) {
+    BrotliCleanupSharedEncoderDictionary(&dict->memory_manager_,
+        (SharedEncoderDictionary*)dict->dictionary);
+    BrotliFree(&dict->memory_manager_, dict->dictionary);
+  } else {
+    /* There is also kPreparedDictionaryMagic, but such instances should be
+     * constructed and destroyed by different means. */
+  }
+  dict->dictionary = NULL;
+  BrotliDestroyManagedDictionary(dict);
+}
+
+BROTLI_BOOL BrotliEncoderAttachPreparedDictionary(BrotliEncoderState* state,
+    const BrotliEncoderPreparedDictionary* dictionary) {
+  /* First field of dictionary structs */
+  const BrotliEncoderPreparedDictionary* dict = dictionary;
+  uint32_t magic = *((const uint32_t*)dict);
+  SharedEncoderDictionary* current = NULL;
+  if (magic == kManagedDictionaryMagic) {
+    /* Unwrap managed dictionary. */
+    ManagedDictionary* managed_dictionary = (ManagedDictionary*)dict;
+    magic = *managed_dictionary->dictionary;
+    dict = (BrotliEncoderPreparedDictionary*)managed_dictionary->dictionary;
+  }
+  current = &state->params.dictionary;
+  if (magic == kPreparedDictionaryMagic ||
+      magic == kLeanPreparedDictionaryMagic) {
+    const PreparedDictionary* prepared = (const PreparedDictionary*)dict;
+    if (!AttachPreparedDictionary(&current->compound, prepared)) {
+      return BROTLI_FALSE;
+    }
+  } else if (magic == kSharedDictionaryMagic) {
+    const SharedEncoderDictionary* attached =
+        (const SharedEncoderDictionary*)dict;
+    BROTLI_BOOL was_default = !current->contextual.context_based &&
+        current->contextual.num_dictionaries == 1 &&
+        current->contextual.dict[0]->hash_table_words ==
+        kStaticDictionaryHashWords &&
+        current->contextual.dict[0]->hash_table_lengths ==
+        kStaticDictionaryHashLengths;
+    BROTLI_BOOL new_default = !attached->contextual.context_based &&
+        attached->contextual.num_dictionaries == 1 &&
+        attached->contextual.dict[0]->hash_table_words ==
+        kStaticDictionaryHashWords &&
+        attached->contextual.dict[0]->hash_table_lengths ==
+        kStaticDictionaryHashLengths;
+    size_t i;
+    if (state->is_initialized_) return BROTLI_FALSE;
+    current->max_quality =
+        BROTLI_MIN(int, current->max_quality, attached->max_quality);
+    for (i = 0; i < attached->compound.num_chunks; i++) {
+      if (!AttachPreparedDictionary(&current->compound,
+          attached->compound.chunks[i])) {
+        return BROTLI_FALSE;
+      }
+    }
+    if (!new_default) {
+      if (!was_default) return BROTLI_FALSE;
+      /* Copy by value, but then set num_instances_ to 0 because their memory
+      is managed by attached, not by current */
+      current->contextual = attached->contextual;
+      current->contextual.num_instances_ = 0;
+    }
+  } else {
+    return BROTLI_FALSE;
+  }
+  return BROTLI_TRUE;
+}
+
+size_t BrotliEncoderEstimatePeakMemoryUsage(int quality, int lgwin,
+                                            size_t input_size) {
+  BrotliEncoderParams params;
+  size_t memory_manager_slots = BROTLI_ENCODER_MEMORY_MANAGER_SLOTS;
+  size_t memory_manager_size = memory_manager_slots * sizeof(void*);
+  BrotliEncoderInitParams(&params);
+  params.quality = quality;
+  params.lgwin = lgwin;
+  params.size_hint = input_size;
+  params.large_window = lgwin > BROTLI_MAX_WINDOW_BITS;
+  SanitizeParams(&params);
+  params.lgblock = ComputeLgBlock(&params);
+  ChooseHasher(&params, &params.hasher);
+  if (params.quality == FAST_ONE_PASS_COMPRESSION_QUALITY ||
+      params.quality == FAST_TWO_PASS_COMPRESSION_QUALITY) {
+    size_t state_size = sizeof(BrotliEncoderState);
+    size_t block_size = BROTLI_MIN(size_t, input_size, (1ul << params.lgwin));
+    size_t hash_table_size =
+        HashTableSize(MaxHashTableSize(params.quality), block_size);
+    size_t hash_size =
+        (hash_table_size < (1u << 10)) ? 0 : sizeof(int) * hash_table_size;
+    size_t cmdbuf_size = params.quality == FAST_TWO_PASS_COMPRESSION_QUALITY ?
+        5 * BROTLI_MIN(size_t, block_size, 1ul << 17) : 0;
+    if (params.quality == FAST_ONE_PASS_COMPRESSION_QUALITY) {
+      state_size += sizeof(BrotliOnePassArena);
+    } else {
+      state_size += sizeof(BrotliTwoPassArena);
+    }
+    return hash_size + cmdbuf_size + state_size;
+  } else {
+    size_t short_ringbuffer_size = (size_t)1 << params.lgblock;
+    int ringbuffer_bits = ComputeRbBits(&params);
+    size_t ringbuffer_size = input_size < short_ringbuffer_size ?
+        input_size : (1u << ringbuffer_bits) + short_ringbuffer_size;
+    size_t hash_size[4] = {0};
+    size_t metablock_size =
+        BROTLI_MIN(size_t, input_size, MaxMetablockSize(&params));
+    size_t inputblock_size =
+        BROTLI_MIN(size_t, input_size, (size_t)1 << params.lgblock);
+    size_t cmdbuf_size = metablock_size * 2 + inputblock_size * 6;
+    size_t outbuf_size = metablock_size * 2 + 503;
+    size_t histogram_size = 0;
+    HasherSize(&params, BROTLI_TRUE, input_size, hash_size);
+    if (params.quality < MIN_QUALITY_FOR_BLOCK_SPLIT) {
+      cmdbuf_size = BROTLI_MIN(size_t, cmdbuf_size,
+          MAX_NUM_DELAYED_SYMBOLS * sizeof(Command) + inputblock_size * 12);
+    }
+    if (params.quality >= MIN_QUALITY_FOR_HQ_BLOCK_SPLITTING) {
+      /* Only a very rough estimation, based on enwik8. */
+      histogram_size = 200 << 20;
+    } else if (params.quality >= MIN_QUALITY_FOR_BLOCK_SPLIT) {
+      size_t literal_histograms =
+          BROTLI_MIN(size_t, metablock_size / 6144, 256);
+      size_t command_histograms =
+          BROTLI_MIN(size_t, metablock_size / 6144, 256);
+      size_t distance_histograms =
+          BROTLI_MIN(size_t, metablock_size / 6144, 256);
+      histogram_size = literal_histograms * sizeof(HistogramLiteral) +
+                       command_histograms * sizeof(HistogramCommand) +
+                       distance_histograms * sizeof(HistogramDistance);
+    }
+    return (memory_manager_size + ringbuffer_size +
+            hash_size[0] + hash_size[1] + hash_size[2] + hash_size[3] +
+            cmdbuf_size +
+            outbuf_size +
+            histogram_size);
+  }
+}
+size_t BrotliEncoderGetPreparedDictionarySize(
+    const BrotliEncoderPreparedDictionary* prepared_dictionary) {
+  /* First field of dictionary structs */
+  const BrotliEncoderPreparedDictionary* prepared = prepared_dictionary;
+  uint32_t magic = *((const uint32_t*)prepared);
+  size_t overhead = 0;
+  if (magic == kManagedDictionaryMagic) {
+    const ManagedDictionary* managed = (const ManagedDictionary*)prepared;
+    overhead = sizeof(ManagedDictionary);
+    magic = *managed->dictionary;
+    prepared = (const BrotliEncoderPreparedDictionary*)managed->dictionary;
+  }
+
+  if (magic == kPreparedDictionaryMagic) {
+    const PreparedDictionary* dictionary =
+        (const PreparedDictionary*)prepared;
+    /* Keep in sync with step 3 of CreatePreparedDictionary */
+    return sizeof(PreparedDictionary) + dictionary->source_size +
+        (sizeof(uint32_t) << dictionary->slot_bits) +
+        (sizeof(uint16_t) << dictionary->bucket_bits) +
+        (sizeof(uint32_t) * dictionary->num_items) + overhead;
+  } else if (magic == kLeanPreparedDictionaryMagic) {
+    const PreparedDictionary* dictionary =
+        (const PreparedDictionary*)prepared;
+    /* Keep in sync with step 3 of CreatePreparedDictionary */
+    return sizeof(PreparedDictionary) + sizeof(uint8_t*) +
+        (sizeof(uint32_t) << dictionary->slot_bits) +
+        (sizeof(uint16_t) << dictionary->bucket_bits) +
+        (sizeof(uint32_t) * dictionary->num_items) + overhead;
+  } else if (magic == kSharedDictionaryMagic) {
+    const SharedEncoderDictionary* dictionary =
+        (const SharedEncoderDictionary*)prepared;
+    const CompoundDictionary* compound = &dictionary->compound;
+    const ContextualEncoderDictionary* contextual = &dictionary->contextual;
+    size_t result = sizeof(*dictionary);
+    size_t i;
+    size_t num_instances;
+    const BrotliEncoderDictionary* instances;
+    for (i = 0; i < compound->num_prepared_instances_; i++) {
+      size_t size = BrotliEncoderGetPreparedDictionarySize(
+          (const BrotliEncoderPreparedDictionary*)
+          compound->prepared_instances_[i]);
+      if (!size) return 0;  /* error */
+      result += size;
+    }
+    if (contextual->context_based) {
+      num_instances = contextual->num_instances_;
+      instances = contextual->instances_;
+      result += sizeof(*instances) * num_instances;
+    } else {
+      num_instances = 1;
+      instances = &contextual->instance_;
+    }
+    for (i = 0; i < num_instances; i++) {
+      const BrotliEncoderDictionary* dict = &instances[i];
+      result += dict->trie.pool_capacity * sizeof(BrotliTrieNode);
+      if (dict->hash_table_data_words_) {
+        result += sizeof(kStaticDictionaryHashWords);
+      }
+      if (dict->hash_table_data_lengths_) {
+        result += sizeof(kStaticDictionaryHashLengths);
+      }
+      if (dict->buckets_data_) {
+        result += sizeof(*dict->buckets_data_) * dict->buckets_alloc_size_;
+      }
+      if (dict->dict_words_data_) {
+        result += sizeof(*dict->dict_words) * dict->dict_words_alloc_size_;
+      }
+      if (dict->words_instance_) {
+        result += sizeof(*dict->words_instance_);
+        /* data_size not added here: it is never allocated by the
+           SharedEncoderDictionary, instead it always points to the file
+           already loaded in memory. So if the caller wants to include
+           this memory as well, add the size of the loaded dictionary
+           file to this. */
+      }
+    }
+    return result + overhead;
+  }
+  return 0;  /* error */
+}
+
+#if defined(BROTLI_TEST)
+size_t MakeUncompressedStreamForTest(const uint8_t*, size_t, uint8_t*);
+size_t MakeUncompressedStreamForTest(
+    const uint8_t* input, size_t input_size, uint8_t* output) {
+  return MakeUncompressedStream(input, input_size, output);
+}
+#endif
+
 #if defined(__cplusplus) || defined(c_plusplus)
 }  /* extern "C" */
 #endif