Factor out serialization functions into their own file.

Create a brotli_bit_stream library that is responsible for writing
various structures (headers, Huffman codes, etc.) directly into the
bit-stream.

enc/bit_cost.h

@@ -91,7 +91,7 @@ static inline int HuffmanBitCost(const uint8_t* depth, int length) {
 
   // create huffman tree of huffman tree
   uint8_t cost[kCodeLengthCodes] = { 0 };
-  CreateHuffmanTree(histogram, kCodeLengthCodes, 7, cost);
+  CreateHuffmanTree(histogram, kCodeLengthCodes, 7, 9, cost);
   // account for rle extra bits
   cost[16] += 2;
   cost[17] += 3;
@@ -123,7 +123,7 @@ double PopulationCost(const Histogram<kSize>& histogram) {
     return 20 + histogram.total_count_;
   }
   uint8_t depth[kSize] = { 0 };
-  CreateHuffmanTree(&histogram.data_[0], kSize, 15, depth);
+  CreateHuffmanTree(&histogram.data_[0], kSize, 15, 9, depth);
   int bits = 0;
   for (int i = 0; i < kSize; ++i) {
     bits += histogram.data_[i] * depth[i];

enc/brotli_bit_stream.cc

@@ -0,0 +1,342 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// Brotli bit stream functions to support the low level format. There are no
+// compression algorithms here, just the right ordering of bits to match the
+// specs.
+
+#include "./brotli_bit_stream.h"
+
+#include <vector>
+
+#include "./entropy_encode.h"
+#include "./fast_log.h"
+#include "./write_bits.h"
+
+namespace brotli {
+
+// returns false if fail
+// nibblesbits represents the 2 bits to encode MNIBBLES (0-3)
+bool EncodeMlen(size_t length, int* bits, int* numbits, int* nibblesbits) {
+  length--;  // MLEN - 1 is encoded
+  int lg = length == 0 ? 1 : Log2Floor(length) + 1;
+  if (lg > 28) return false;
+  int mnibbles = (lg < 16 ? 16 : (lg + 3)) / 4;
+  *nibblesbits = mnibbles - 4;
+  *numbits = mnibbles * 4;
+  *bits = length;
+  return true;
+}
+
+void StoreVarLenUint8(int n, int* storage_ix, uint8_t* storage) {
+  if (n == 0) {
+    WriteBits(1, 0, storage_ix, storage);
+  } else {
+    WriteBits(1, 1, storage_ix, storage);
+    int nbits = Log2Floor(n);
+    WriteBits(3, nbits, storage_ix, storage);
+    WriteBits(nbits, n - (1 << nbits), storage_ix, storage);
+  }
+}
+
+void StoreCompressedMetaBlockHeader(bool final_block,
+                                    int length,
+                                    int* storage_ix,
+                                    uint8_t* storage) {
+  // Write ISLAST bit.
+  WriteBits(1, final_block, storage_ix, storage);
+  // Write ISEMPTY bit.
+  if (final_block) WriteBits(1, length == 0, storage_ix, storage);
+
+  int lenbits;
+  int nlenbits;
+  int nibblesbits;
+  EncodeMlen(length, &lenbits, &nlenbits, &nibblesbits);
+
+  WriteBits(2, nibblesbits, storage_ix, storage);
+  WriteBits(nlenbits, lenbits, storage_ix, storage);
+
+  if (!final_block) {
+    // Write ISUNCOMPRESSED bit.
+    WriteBits(1, 0, storage_ix, storage);
+  }
+}
+
+void StoreUncompressedMetaBlockHeader(int length,
+                                      int* storage_ix,
+                                      uint8_t* storage) {
+  // Write ISLAST bit. Uncompressed block cannot be the last one, so set to 0.
+  WriteBits(1, 0, storage_ix, storage);
+  int lenbits;
+  int nlenbits;
+  int nibblesbits;
+  EncodeMlen(length, &lenbits, &nlenbits, &nibblesbits);
+  WriteBits(2, nibblesbits, storage_ix, storage);
+  WriteBits(nlenbits, lenbits, storage_ix, storage);
+  // Write ISUNCOMPRESSED bit.
+  WriteBits(1, 1, storage_ix, storage);
+}
+
+void StoreHuffmanTreeOfHuffmanTreeToBitMask(
+    const int num_codes,
+    const uint8_t *code_length_bitdepth,
+    int *storage_ix,
+    uint8_t *storage) {
+  static const uint8_t kStorageOrder[kCodeLengthCodes] = {
+    1, 2, 3, 4, 0, 5, 17, 6, 16, 7, 8, 9, 10, 11, 12, 13, 14, 15
+  };
+  // The bit lengths of the Huffman code over the code length alphabet
+  // are compressed with the following static Huffman code:
+  //   Symbol   Code
+  //   ------   ----
+  //   0          00
+  //   1        1110
+  //   2         110
+  //   3          01
+  //   4          10
+  //   5        1111
+  static const uint8_t kHuffmanBitLengthHuffmanCodeSymbols[6] = {
+     0, 7, 3, 2, 1, 15
+  };
+  static const uint8_t kHuffmanBitLengthHuffmanCodeBitLengths[6] = {
+    2, 4, 3, 2, 2, 4
+  };
+
+  // Throw away trailing zeros:
+  int codes_to_store = kCodeLengthCodes;
+  if (num_codes > 1) {
+    for (; codes_to_store > 0; --codes_to_store) {
+      if (code_length_bitdepth[kStorageOrder[codes_to_store - 1]] != 0) {
+        break;
+      }
+    }
+  }
+  int skip_some = 0;  // skips none.
+  if (code_length_bitdepth[kStorageOrder[0]] == 0 &&
+      code_length_bitdepth[kStorageOrder[1]] == 0) {
+    skip_some = 2;  // skips two.
+    if (code_length_bitdepth[kStorageOrder[2]] == 0) {
+      skip_some = 3;  // skips three.
+    }
+  }
+  WriteBits(2, skip_some, storage_ix, storage);
+  for (int i = skip_some; i < codes_to_store; ++i) {
+    uint8_t l = code_length_bitdepth[kStorageOrder[i]];
+    WriteBits(kHuffmanBitLengthHuffmanCodeBitLengths[l],
+              kHuffmanBitLengthHuffmanCodeSymbols[l], storage_ix, storage);
+  }
+}
+
+void StoreHuffmanTreeToBitMask(
+    const std::vector<uint8_t> &huffman_tree,
+    const std::vector<uint8_t> &huffman_tree_extra_bits,
+    const uint8_t *code_length_bitdepth,
+    const std::vector<uint16_t> &code_length_bitdepth_symbols,
+    int * __restrict storage_ix,
+    uint8_t * __restrict storage) {
+  for (int i = 0; i < huffman_tree.size(); ++i) {
+    int ix = huffman_tree[i];
+    WriteBits(code_length_bitdepth[ix], code_length_bitdepth_symbols[ix],
+              storage_ix, storage);
+    // Extra bits
+    switch (ix) {
+      case 16:
+        WriteBits(2, huffman_tree_extra_bits[i], storage_ix, storage);
+        break;
+      case 17:
+        WriteBits(3, huffman_tree_extra_bits[i], storage_ix, storage);
+        break;
+    }
+  }
+}
+
+void StoreSimpleHuffmanTree(const uint8_t* depths,
+                            int symbols[4],
+                            int num_symbols,
+                            int max_bits,
+                            int *storage_ix, uint8_t *storage) {
+  // value of 1 indicates a simple Huffman code
+  WriteBits(2, 1, storage_ix, storage);
+  WriteBits(2, num_symbols - 1, storage_ix, storage);  // NSYM - 1
+
+  // Sort
+  for (int i = 0; i < num_symbols; i++) {
+    for (int j = i + 1; j < num_symbols; j++) {
+      if (depths[symbols[j]] < depths[symbols[i]]) {
+        std::swap(symbols[j], symbols[i]);
+      }
+    }
+  }
+
+  if (num_symbols == 2) {
+    WriteBits(max_bits, symbols[0], storage_ix, storage);
+    WriteBits(max_bits, symbols[1], storage_ix, storage);
+  } else if (num_symbols == 3) {
+    WriteBits(max_bits, symbols[0], storage_ix, storage);
+    WriteBits(max_bits, symbols[1], storage_ix, storage);
+    WriteBits(max_bits, symbols[2], storage_ix, storage);
+  } else {
+    WriteBits(max_bits, symbols[0], storage_ix, storage);
+    WriteBits(max_bits, symbols[1], storage_ix, storage);
+    WriteBits(max_bits, symbols[2], storage_ix, storage);
+    WriteBits(max_bits, symbols[3], storage_ix, storage);
+    // tree-select
+    WriteBits(1, depths[symbols[0]] == 1 ? 1 : 0, storage_ix, storage);
+  }
+}
+
+// num = alphabet size
+// depths = symbol depths
+void StoreHuffmanTree(const uint8_t* depths, size_t num,
+                      int quality,
+                      int *storage_ix, uint8_t *storage) {
+  // Write the Huffman tree into the brotli-representation.
+  std::vector<uint8_t> huffman_tree;
+  std::vector<uint8_t> huffman_tree_extra_bits;
+  // TODO(user): Consider allocating these from stack.
+  huffman_tree.reserve(256);
+  huffman_tree_extra_bits.reserve(256);
+  WriteHuffmanTree(depths, num, &huffman_tree, &huffman_tree_extra_bits);
+
+  // Calculate the statistics of the Huffman tree in brotli-representation.
+  int huffman_tree_histogram[kCodeLengthCodes] = { 0 };
+  for (int i = 0; i < huffman_tree.size(); ++i) {
+    ++huffman_tree_histogram[huffman_tree[i]];
+  }
+
+  int num_codes = 0;
+  int code = 0;
+  for (int i = 0; i < kCodeLengthCodes; ++i) {
+    if (huffman_tree_histogram[i]) {
+      if (num_codes == 0) {
+        code = i;
+        num_codes = 1;
+      } else if (num_codes == 1) {
+        num_codes = 2;
+        break;
+      }
+    }
+  }
+
+  // Calculate another Huffman tree to use for compressing both the
+  // earlier Huffman tree with.
+  // TODO(user): Consider allocating these from stack.
+  uint8_t code_length_bitdepth[kCodeLengthCodes] = { 0 };
+  std::vector<uint16_t> code_length_bitdepth_symbols(kCodeLengthCodes);
+  CreateHuffmanTree(&huffman_tree_histogram[0], kCodeLengthCodes,
+                    5, quality, &code_length_bitdepth[0]);
+  ConvertBitDepthsToSymbols(code_length_bitdepth, kCodeLengthCodes,
+                            code_length_bitdepth_symbols.data());
+
+  // Now, we have all the data, let's start storing it
+  StoreHuffmanTreeOfHuffmanTreeToBitMask(num_codes, code_length_bitdepth,
+                                         storage_ix, storage);
+
+  if (num_codes == 1) {
+    code_length_bitdepth[code] = 0;
+  }
+
+  // Store the real huffman tree now.
+  StoreHuffmanTreeToBitMask(huffman_tree,
+                            huffman_tree_extra_bits,
+                            &code_length_bitdepth[0],
+                            code_length_bitdepth_symbols,
+                            storage_ix, storage);
+}
+
+void BuildAndStoreHuffmanTree(const int *histogram,
+                              const int length,
+                              const int quality,
+                              uint8_t* depth,
+                              uint16_t* bits,
+                              int* storage_ix,
+                              uint8_t* storage) {
+  int count = 0;
+  int s4[4] = { 0 };
+  for (size_t i = 0; i < length; i++) {
+    if (histogram[i]) {
+      if (count < 4) {
+        s4[count] = i;
+      } else if (quality < 3 && count > 4) {
+        break;
+      }
+      count++;
+    }
+  }
+
+  int max_bits_counter = length - 1;
+  int max_bits = 0;
+  while (max_bits_counter) {
+    max_bits_counter >>= 1;
+    ++max_bits;
+  }
+
+  if (count <= 1) {
+    WriteBits(4, 1, storage_ix, storage);
+    WriteBits(max_bits, s4[0], storage_ix, storage);
+    return;
+  }
+
+  if (length >= 50 && count >= 16 && quality >= 3) {
+    std::vector<int> counts(length);
+    memcpy(&counts[0], histogram, sizeof(counts[0]) * length);
+    OptimizeHuffmanCountsForRle(length, &counts[0]);
+    CreateHuffmanTree(&counts[0], length, 15, quality, depth);
+  } else {
+    CreateHuffmanTree(histogram, length, 15, quality, depth);
+  }
+  ConvertBitDepthsToSymbols(depth, length, bits);
+
+  if (count <= 4) {
+    StoreSimpleHuffmanTree(depth, s4, count, max_bits, storage_ix, storage);
+  } else {
+    StoreHuffmanTree(depth, length, quality, storage_ix, storage);
+  }
+}
+
+void StoreTrivialContextMap(int num_types,
+                            int context_bits,
+                            int* storage_ix,
+                            uint8_t* storage) {
+  StoreVarLenUint8(num_types - 1, storage_ix, storage);
+  if (num_types > 1) {
+    int repeat_code = context_bits - 1;
+    int repeat_bits = (1 << repeat_code) - 1;
+    int alphabet_size = num_types + repeat_code;
+    std::vector<int> histogram(alphabet_size);
+    std::vector<uint8_t> depths(alphabet_size);
+    std::vector<uint16_t> bits(alphabet_size);
+    // Write RLEMAX.
+    WriteBits(1, 1, storage_ix, storage);
+    WriteBits(4, repeat_code - 1, storage_ix, storage);
+    histogram[repeat_code] = num_types;
+    histogram[0] = 1;
+    for (int i = context_bits; i < alphabet_size; ++i) {
+      histogram[i] = 1;
+    }
+    BuildAndStoreHuffmanTree(&histogram[0], alphabet_size, 1,
+                             &depths[0], &bits[0],
+                             storage_ix, storage);
+    for (int i = 0; i < num_types; ++i) {
+      int code = (i == 0 ? 0 : i + context_bits - 1);
+      WriteBits(depths[code], bits[code], storage_ix, storage);
+      WriteBits(depths[repeat_code], bits[repeat_code], storage_ix, storage);
+      WriteBits(repeat_code, repeat_bits, storage_ix, storage);
+    }
+    // Write IMTF (inverse-move-to-front) bit.
+    WriteBits(1, 1, storage_ix, storage);
+  }
+}
+
+}  // namespace brotli

enc/brotli_bit_stream.h

@@ -0,0 +1,67 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// Functions to convert brotli-related data structures into the
+// brotli bit stream. The functions here operate under
+// assumption that there is enough space in the storage, i.e., there are
+// no out-of-range checks anywhere.
+//
+// These functions do bit addressing into a byte array. The byte array
+// is called "storage" and the index to the bit is called storage_ix
+// in function arguments.
+
+#ifndef BROTLI_ENC_BROTLI_BIT_STREAM_H_
+#define BROTLI_ENC_BROTLI_BIT_STREAM_H_
+
+#include <stddef.h>
+#include <stdint.h>
+
+namespace brotli {
+
+// All Store functions here will use a storage_ix, which is always the bit
+// position for the current storage.
+
+// Stores a number between 0 and 255.
+void StoreVarLenUint8(int n, int* storage_ix, uint8_t* storage);
+
+// Stores the compressed meta-block header.
+void StoreCompressedMetaBlockHeader(bool final_block,
+                                    int length,
+                                    int* storage_ix,
+                                    uint8_t* storage);
+
+// Stores the uncompressed meta-block header.
+void StoreUncompressedMetaBlockHeader(int length,
+                                      int* storage_ix,
+                                      uint8_t* storage);
+
+// Stores a context map where the histogram type is always the block type.
+void StoreTrivialContextMap(int num_types,
+                            int context_bits,
+                            int* storage_ix,
+                            uint8_t* storage);
+
+// Builds a Huffman tree from histogram[0:length] into depth[0:length] and
+// bits[0:length] and stores the encoded tree to the bit stream.
+void BuildAndStoreHuffmanTree(const int *histogram,
+                              const int length,
+                              const int quality,
+                              uint8_t* depth,
+                              uint16_t* bits,
+                              int* storage_ix,
+                              uint8_t* storage);
+
+}  // namespace brotli
+
+#endif  // BROTLI_ENC_BROTLI_BIT_STREAM_H_

enc/encode.cc

@@ -22,6 +22,7 @@
 #include "./backward_references.h"
 #include "./bit_cost.h"
 #include "./block_splitter.h"
+#include "./brotli_bit_stream.h"
 #include "./cluster.h"
 #include "./context.h"
 #include "./transform.h"
@@ -65,19 +66,6 @@ double TotalBitCost(const std::vector<Histogram<kSize> >& histograms) {
   return retval;
 }
 
-void EncodeVarLenUint8(int n, int* storage_ix, uint8_t* storage) {
-  if (n == 0) {
-    WriteBits(1, 0, storage_ix, storage);
-  } else {
-    WriteBits(1, 1, storage_ix, storage);
-    int nbits = Log2Floor(n);
-    WriteBits(3, nbits, storage_ix, storage);
-    if (nbits > 0) {
-      WriteBits(nbits, n - (1 << nbits), storage_ix, storage);
-    }
-  }
-}
-
 int ParseAsUTF8(int* symbol, const uint8_t* input, int size) {
   // ASCII
   if ((input[0] & 0x80) == 0) {
@@ -168,134 +156,6 @@ void EncodeMetaBlockLength(size_t meta_block_size,
   }
 }
 
-void StoreHuffmanTreeOfHuffmanTreeToBitMask(
-    const uint8_t* code_length_bitdepth,
-    int* storage_ix, uint8_t* storage) {
-  static const uint8_t kStorageOrder[kCodeLengthCodes] = {
-    1, 2, 3, 4, 0, 5, 17, 6, 16, 7, 8, 9, 10, 11, 12, 13, 14, 15,
-  };
-  // Throw away trailing zeros:
-  int codes_to_store = kCodeLengthCodes;
-  for (; codes_to_store > 0; --codes_to_store) {
-    if (code_length_bitdepth[kStorageOrder[codes_to_store - 1]] != 0) {
-      break;
-    }
-  }
-  int num_codes = 0;
-  for (int i = 0; i < codes_to_store; ++i) {
-    if (code_length_bitdepth[kStorageOrder[i]] != 0) {
-      ++num_codes;
-    }
-  }
-  if (num_codes == 1) {
-    codes_to_store = kCodeLengthCodes;
-  }
-  int skip_some = 0;  // skips none.
-  if (code_length_bitdepth[kStorageOrder[0]] == 0 &&
-      code_length_bitdepth[kStorageOrder[1]] == 0) {
-    skip_some = 2;  // skips two.
-    if (code_length_bitdepth[kStorageOrder[2]] == 0) {
-      skip_some = 3;  // skips three.
-    }
-  }
-  WriteBits(2, skip_some, storage_ix, storage);
-  for (int i = skip_some; i < codes_to_store; ++i) {
-    uint8_t len[] = { 2, 4, 3, 2, 2, 4 };
-    uint8_t bits[] = { 0, 7, 3, 2, 1, 15 };
-    int v = code_length_bitdepth[kStorageOrder[i]];
-    WriteBits(len[v], bits[v], storage_ix, storage);
-  }
-}
-
-void StoreHuffmanTreeToBitMask(
-    const uint8_t* huffman_tree,
-    const uint8_t* huffman_tree_extra_bits,
-    const int huffman_tree_size,
-    const EntropyCode<kCodeLengthCodes>& entropy,
-    int* storage_ix, uint8_t* storage) {
-  for (int i = 0; i < huffman_tree_size; ++i) {
-    const int ix = huffman_tree[i];
-    const int extra_bits = huffman_tree_extra_bits[i];
-    if (entropy.count_ > 1) {
-      WriteBits(entropy.depth_[ix], entropy.bits_[ix], storage_ix, storage);
-    }
-    switch (ix) {
-      case 16:
-        WriteBits(2, extra_bits, storage_ix, storage);
-        break;
-      case 17:
-        WriteBits(3, extra_bits, storage_ix, storage);
-        break;
-    }
-  }
-}
-
-template<int kSize>
-void StoreHuffmanCodeSimple(
-    const EntropyCode<kSize>& code, int alphabet_size,
-    int max_bits, int* storage_ix, uint8_t* storage) {
-  const uint8_t *depth = &code.depth_[0];
-  int symbols[4];
-  // Quadratic sort.
-  int k, j;
-  for (k = 0; k < code.count_; ++k) {
-    symbols[k] = code.symbols_[k];
-  }
-  for (k = 0; k < code.count_; ++k) {
-    for (j = k + 1; j < code.count_; ++j) {
-      if (depth[symbols[j]] < depth[symbols[k]]) {
-        int t = symbols[k];
-        symbols[k] = symbols[j];
-        symbols[j] = t;
-      }
-    }
-  }
-  // Small tree marker to encode 1-4 symbols.
-  WriteBits(2, 1, storage_ix, storage);
-  WriteBits(2, code.count_ - 1, storage_ix, storage);
-  for (int i = 0; i < code.count_; ++i) {
-    WriteBits(max_bits, symbols[i], storage_ix, storage);
-  }
-  if (code.count_ == 4) {
-    if (depth[symbols[0]] == 2 &&
-        depth[symbols[1]] == 2 &&
-        depth[symbols[2]] == 2 &&
-        depth[symbols[3]] == 2) {
-      WriteBits(1, 0, storage_ix, storage);
-    } else {
-      WriteBits(1, 1, storage_ix, storage);
-    }
-  }
-}
-
-template<int kSize>
-void StoreHuffmanCodeComplex(
-    const EntropyCode<kSize>& code, int alphabet_size,
-    int* storage_ix, uint8_t* storage) {
-  const uint8_t *depth = &code.depth_[0];
-  uint8_t huffman_tree[kSize];
-  uint8_t huffman_tree_extra_bits[kSize];
-  int huffman_tree_size = 0;
-  WriteHuffmanTree(depth,
-                   alphabet_size,
-                   &huffman_tree[0],
-                   &huffman_tree_extra_bits[0],
-                   &huffman_tree_size);
-  Histogram<kCodeLengthCodes> huffman_tree_histogram;
-  memset(huffman_tree_histogram.data_, 0, sizeof(huffman_tree_histogram.data_));
-  for (int i = 0; i < huffman_tree_size; ++i) {
-    huffman_tree_histogram.Add(huffman_tree[i]);
-  }
-  EntropyCode<kCodeLengthCodes> huffman_tree_entropy;
-  BuildEntropyCode(huffman_tree_histogram, 5, kCodeLengthCodes,
-                   &huffman_tree_entropy);
-  StoreHuffmanTreeOfHuffmanTreeToBitMask(
-      &huffman_tree_entropy.depth_[0], storage_ix, storage);
-  StoreHuffmanTreeToBitMask(&huffman_tree[0], &huffman_tree_extra_bits[0],
-                            huffman_tree_size, huffman_tree_entropy,
-                            storage_ix, storage);
-}
-
 template<int kSize>
 void BuildAndStoreEntropyCode(const Histogram<kSize>& histogram,
                               const int tree_limit,
@@ -304,45 +164,8 @@ void BuildAndStoreEntropyCode(const Histogram<kSize>& histogram,
                               int* storage_ix, uint8_t* storage) {
   memset(code->depth_, 0, sizeof(code->depth_));
   memset(code->bits_, 0, sizeof(code->bits_));
-  memset(code->symbols_, 0, sizeof(code->symbols_));
+  BuildAndStoreHuffmanTree(histogram.data_, alphabet_size, 9,
-  code->count_ = 0;
+                           code->depth_, code->bits_, storage_ix, storage);
-
-  int max_bits_counter = alphabet_size - 1;
-  int max_bits = 0;
-  while (max_bits_counter) {
-    max_bits_counter >>= 1;
-    ++max_bits;
-  }
-
-  for (size_t i = 0; i < alphabet_size; i++) {
-    if (histogram.data_[i] > 0) {
-      if (code->count_ < 4) code->symbols_[code->count_] = i;
-      ++code->count_;
-    }
-  }
-
-  if (code->count_ <= 1) {
-    WriteBits(2, 1, storage_ix, storage);
-    WriteBits(2, 0, storage_ix, storage);
-    WriteBits(max_bits, code->symbols_[0], storage_ix, storage);
-    return;
-  }
-
-  if (alphabet_size >= 50 && code->count_ >= 16) {
-    std::vector<int> counts(alphabet_size);
-    memcpy(&counts[0], histogram.data_, sizeof(counts[0]) * alphabet_size);
-    OptimizeHuffmanCountsForRle(alphabet_size, &counts[0]);
-    CreateHuffmanTree(&counts[0], alphabet_size, tree_limit, code->depth_);
-  } else {
-    CreateHuffmanTree(histogram.data_, alphabet_size, tree_limit, code->depth_);
-  }
-  ConvertBitDepthsToSymbols(code->depth_, alphabet_size, code->bits_);
-
-  if (code->count_ <= 4) {
-    StoreHuffmanCodeSimple(*code, alphabet_size, max_bits, storage_ix, storage);
-  } else {
-    StoreHuffmanCodeComplex(*code, alphabet_size, storage_ix, storage);
-  }
 }
 
 template<int kSize>
@@ -575,7 +398,7 @@ int BestMaxZeroRunLengthPrefix(const std::vector<int>& v) {
 void EncodeContextMap(const std::vector<int>& context_map,
                       int num_clusters,
                       int* storage_ix, uint8_t* storage) {
-  EncodeVarLenUint8(num_clusters - 1, storage_ix, storage);
+  StoreVarLenUint8(num_clusters - 1, storage_ix, storage);
 
   if (num_clusters == 1) {
     return;
@@ -656,7 +479,7 @@ void ComputeBlockTypeShortCodes(BlockSplit* split) {
 void BuildAndEncodeBlockSplitCode(const BlockSplit& split,
                                   BlockSplitCode* code,
                                   int* storage_ix, uint8_t* storage) {
-  EncodeVarLenUint8(split.num_types_ - 1, storage_ix, storage);
+  StoreVarLenUint8(split.num_types_ - 1, storage_ix, storage);
 
   if (split.num_types_ == 1) {
     return;
@@ -864,7 +687,6 @@ void StoreMetaBlock(const MetaBlock& mb,
       int context = (distance_it.type_ << 2) +
           ((cmd.copy_length_code_ > 4) ? 3 : cmd.copy_length_code_ - 2);
       int histogram_index = mb.distance_context_map[context];
-      size_t max_distance = std::min(*pos, (size_t)kMaxBackwardDistance);
       EncodeCopyDistance(cmd, distance_codes[histogram_index],
                          storage_ix, storage);
     }
@@ -1015,7 +837,6 @@ void BrotliCompressor::FinishStream(
   WriteMetaBlock(0, NULL, true, encoded_size, encoded_buffer);
 }
 
-
 int BrotliCompressBuffer(BrotliParams params,
                          size_t input_size,
                          const uint8_t* input_buffer,

enc/encode.h

@@ -59,7 +59,6 @@ class BrotliCompressor {
   // sets *encoded_size to the number of bytes written.
   void FinishStream(size_t* encoded_size, uint8_t* encoded_buffer);
 
-
  private:
   // Initializes the hasher with the hashes of dictionary words.
   void StoreDictionaryWordHashes();
@@ -87,7 +86,6 @@ int BrotliCompressBuffer(BrotliParams params,
                          size_t* encoded_size,
                          uint8_t* encoded_buffer);
 
-
 }  // namespace brotli
 
 #endif  // BROTLI_ENC_ENCODE_H_

enc/entropy_encode.cc

@@ -42,7 +42,7 @@ struct HuffmanTree {
 
 HuffmanTree::HuffmanTree() {}
 
-// Sort the root nodes, least popular first.
+// Sort the root nodes, least popular first, break ties by value.
 bool SortHuffmanTree(const HuffmanTree &v0, const HuffmanTree &v1) {
   if (v0.total_count_ == v1.total_count_) {
     return v0.index_right_or_value_ > v1.index_right_or_value_;
@@ -50,6 +50,11 @@ bool SortHuffmanTree(const HuffmanTree &v0, const HuffmanTree &v1) {
   return v0.total_count_ < v1.total_count_;
 }
 
+// Sort the root nodes, least popular first.
+bool SortHuffmanTreeFast(const HuffmanTree &v0, const HuffmanTree &v1) {
+  return v0.total_count_ < v1.total_count_;
+}
+
 void SetDepth(const HuffmanTree &p,
               HuffmanTree *pool,
               uint8_t *depth,
@@ -83,6 +88,7 @@ void SetDepth(const HuffmanTree &p,
 void CreateHuffmanTree(const int *data,
                        const int length,
                        const int tree_limit,
+                       const int quality,
                        uint8_t *depth) {
   // For block sizes below 64 kB, we never need to do a second iteration
   // of this loop. Probably all of our block sizes will be smaller than
@@ -105,8 +111,11 @@ void CreateHuffmanTree(const int *data,
       break;
     }
 
-    std::sort(tree.begin(), tree.end(), SortHuffmanTree);
+    if (quality > 1) {
-
+      std::sort(tree.begin(), tree.end(), SortHuffmanTree);
+    } else {
+      std::sort(tree.begin(), tree.end(), SortHuffmanTreeFast);
+    }
     // The nodes are:
     // [0, n): the sorted leaf nodes that we start with.
     // [n]: we add a sentinel here.
@@ -158,12 +167,12 @@ void CreateHuffmanTree(const int *data,
   }
 }
 
-void Reverse(uint8_t* v, int start, int end) {
+void Reverse(std::vector<uint8_t>* v, int start, int end) {
   --end;
   while (start < end) {
-    int tmp = v[start];
+    int tmp = (*v)[start];
-    v[start] = v[end];
+    (*v)[start] = (*v)[end];
-    v[end] = tmp;
+    (*v)[end] = tmp;
     ++start;
     --end;
   }
@@ -173,75 +182,65 @@ void WriteHuffmanTreeRepetitions(
     const int previous_value,
     const int value,
     int repetitions,
-    uint8_t* tree,
+    std::vector<uint8_t> *tree,
-    uint8_t* extra_bits,
+    std::vector<uint8_t> *extra_bits_data) {
-    int* tree_size) {
   if (previous_value != value) {
-    tree[*tree_size] = value;
+    tree->push_back(value);
-    extra_bits[*tree_size] = 0;
+    extra_bits_data->push_back(0);
-    ++(*tree_size);
     --repetitions;
   }
   if (repetitions == 7) {
-    tree[*tree_size] = value;
+    tree->push_back(value);
-    extra_bits[*tree_size] = 0;
+    extra_bits_data->push_back(0);
-    ++(*tree_size);
     --repetitions;
   }
   if (repetitions < 3) {
     for (int i = 0; i < repetitions; ++i) {
-      tree[*tree_size] = value;
+      tree->push_back(value);
-      extra_bits[*tree_size] = 0;
+      extra_bits_data->push_back(0);
-      ++(*tree_size);
     }
   } else {
     repetitions -= 3;
-    int start = *tree_size;
+    int start = tree->size();
     while (repetitions >= 0) {
-      tree[*tree_size] = 16;
+      tree->push_back(16);
-      extra_bits[*tree_size] = repetitions & 0x3;
+      extra_bits_data->push_back(repetitions & 0x3);
-      ++(*tree_size);
       repetitions >>= 2;
       --repetitions;
     }
-    Reverse(tree, start, *tree_size);
+    Reverse(tree, start, tree->size());
-    Reverse(extra_bits, start, *tree_size);
+    Reverse(extra_bits_data, start, tree->size());
   }
 }
 
 void WriteHuffmanTreeRepetitionsZeros(
     int repetitions,
-    uint8_t* tree,
+    std::vector<uint8_t> *tree,
-    uint8_t* extra_bits,
+    std::vector<uint8_t> *extra_bits_data) {
-    int* tree_size) {
   if (repetitions == 11) {
-    tree[*tree_size] = 0;
+    tree->push_back(0);
-    extra_bits[*tree_size] = 0;
+    extra_bits_data->push_back(0);
-    ++(*tree_size);
     --repetitions;
   }
   if (repetitions < 3) {
     for (int i = 0; i < repetitions; ++i) {
-      tree[*tree_size] = 0;
+      tree->push_back(0);
-      extra_bits[*tree_size] = 0;
+      extra_bits_data->push_back(0);
-      ++(*tree_size);
     }
   } else {
     repetitions -= 3;
-    int start = *tree_size;
+    int start = tree->size();
     while (repetitions >= 0) {
-      tree[*tree_size] = 17;
+      tree->push_back(17);
-      extra_bits[*tree_size] = repetitions & 0x7;
+      extra_bits_data->push_back(repetitions & 0x7);
-      ++(*tree_size);
       repetitions >>= 3;
       --repetitions;
     }
-    Reverse(tree, start, *tree_size);
+    Reverse(tree, start, tree->size());
-    Reverse(extra_bits, start, *tree_size);
+    Reverse(extra_bits_data, start, tree->size());
   }
 }
 
-
 int OptimizeHuffmanCountsForRle(int length, int* counts) {
   int stride;
   int limit;
@@ -371,7 +370,6 @@ int OptimizeHuffmanCountsForRle(int length, int* counts) {
   return 1;
 }
 
-
 static void DecideOverRleUse(const uint8_t* depth, const int length,
                              bool *use_rle_for_non_zero,
                              bool *use_rle_for_zero) {
@@ -379,20 +377,10 @@ static void DecideOverRleUse(const uint8_t* depth, const int length,
   int total_reps_non_zero = 0;
   int count_reps_zero = 0;
   int count_reps_non_zero = 0;
-  int new_length = length;
+  for (uint32_t i = 0; i < length;) {
-  for (int i = 0; i < length; ++i) {
-    if (depth[length - i - 1] == 0) {
-      --new_length;
-    } else {
-      break;
-    }
-  }
-  for (uint32_t i = 0; i < new_length;) {
     const int value = depth[i];
     int reps = 1;
-    // Find rle coding for longer codes.
+    for (uint32_t k = i + 1; k < length && depth[k] == value; ++k) {
-    // Shorter codes seem not to benefit from rle.
-    for (uint32_t k = i + 1; k < new_length && depth[k] == value; ++k) {
       ++reps;
     }
     if (reps >= 3 && value == 0) {
@@ -411,48 +399,51 @@ static void DecideOverRleUse(const uint8_t* depth, const int length,
   *use_rle_for_zero = total_reps_zero > 2;
 }
 
-
+void WriteHuffmanTree(const uint8_t* depth,
-void WriteHuffmanTree(const uint8_t* depth, const int length,
+                      uint32_t length,
-                      uint8_t* tree,
+                      std::vector<uint8_t> *tree,
-                      uint8_t* extra_bits_data,
+                      std::vector<uint8_t> *extra_bits_data) {
-                      int* huffman_tree_size) {
   int previous_value = 8;
 
+  // Throw away trailing zeros.
+  int new_length = length;
+  for (int i = 0; i < length; ++i) {
+    if (depth[length - i - 1] == 0) {
+      --new_length;
+    } else {
+      break;
+    }
+  }
+
   // First gather statistics on if it is a good idea to do rle.
-  bool use_rle_for_non_zero;
+  bool use_rle_for_non_zero = false;
-  bool use_rle_for_zero;
+  bool use_rle_for_zero = false;
-  DecideOverRleUse(depth, length, &use_rle_for_non_zero, &use_rle_for_zero);
+  if (length > 50) {
+    // Find rle coding for longer codes.
+    // Shorter codes seem not to benefit from rle.
+    DecideOverRleUse(depth, new_length,
+                     &use_rle_for_non_zero, &use_rle_for_zero);
+  }
 
   // Actual rle coding.
-  for (uint32_t i = 0; i < length;) {
+  for (uint32_t i = 0; i < new_length;) {
     const int value = depth[i];
     int reps = 1;
-    if (length > 50) {
+    if ((value != 0 && use_rle_for_non_zero) ||
-      // Find rle coding for longer codes.
+        (value == 0 && use_rle_for_zero)) {
-      // Shorter codes seem not to benefit from rle.
+      for (uint32_t k = i + 1; k < new_length && depth[k] == value; ++k) {
-      if ((value != 0 && use_rle_for_non_zero) ||
+        ++reps;
-          (value == 0 && use_rle_for_zero)) {
-        for (uint32_t k = i + 1; k < length && depth[k] == value; ++k) {
-          ++reps;
-        }
       }
     }
     if (value == 0) {
-      WriteHuffmanTreeRepetitionsZeros(reps, tree, extra_bits_data,
+      WriteHuffmanTreeRepetitionsZeros(reps, tree, extra_bits_data);
-                                       huffman_tree_size);
     } else {
-      WriteHuffmanTreeRepetitions(previous_value, value, reps, tree,
+      WriteHuffmanTreeRepetitions(previous_value,
-                                  extra_bits_data, huffman_tree_size);
+                                  value, reps, tree, extra_bits_data);
       previous_value = value;
     }
     i += reps;
   }
-  // Throw away trailing zeros.
-  for (; *huffman_tree_size > 0; --(*huffman_tree_size)) {
-    if (tree[*huffman_tree_size - 1] > 0 && tree[*huffman_tree_size - 1] < 17) {
-      break;
-    }
-  }
 }
 
 namespace {

enc/entropy_encode.h

@@ -19,6 +19,7 @@
 
 #include <stdint.h>
 #include <string.h>
+#include <vector>
 #include "./histogram.h"
 #include "./prefix.h"
 
@@ -36,6 +37,7 @@ namespace brotli {
 void CreateHuffmanTree(const int *data,
                        const int length,
                        const int tree_limit,
+                       const int quality,
                        uint8_t *depth);
 
 // Change the population counts in a way that the consequent
@@ -46,14 +48,13 @@ void CreateHuffmanTree(const int *data,
 // counts contains the population counts.
 int OptimizeHuffmanCountsForRle(int length, int* counts);
 
-
 // Write a huffman tree from bit depths into the bitstream representation
 // of a Huffman tree. The generated Huffman tree is to be compressed once
 // more using a Huffman tree
-void WriteHuffmanTree(const uint8_t* depth, const int length,
+void WriteHuffmanTree(const uint8_t* depth,
-                      uint8_t* tree,
+                      uint32_t num,
-                      uint8_t* extra_bits_data,
+                      std::vector<uint8_t> *tree,
-                      int* huffman_tree_size);
+                      std::vector<uint8_t> *extra_bits_data);
 
 // Get the actual bit values for a tree of bit depths.
 void ConvertBitDepthsToSymbols(const uint8_t *depth, int len, uint16_t *bits);
@@ -70,34 +71,6 @@ struct EntropyCode {
   int symbols_[4];
 };
 
-template<int kSize>
-void BuildEntropyCode(const Histogram<kSize>& histogram,
-                      const int tree_limit,
-                      const int alphabet_size,
-                      EntropyCode<kSize>* code) {
-  memset(code->depth_, 0, sizeof(code->depth_));
-  memset(code->bits_, 0, sizeof(code->bits_));
-  memset(code->symbols_, 0, sizeof(code->symbols_));
-  code->count_ = 0;
-  if (histogram.total_count_ == 0) return;
-  for (int i = 0; i < kSize; ++i) {
-    if (histogram.data_[i] > 0) {
-      if (code->count_ < 4) code->symbols_[code->count_] = i;
-      ++code->count_;
-    }
-  }
-  if (alphabet_size >= 50 && code->count_ >= 16) {
-    int counts[kSize];
-    memcpy(counts, &histogram.data_[0], sizeof(counts[0]) * kSize);
-    OptimizeHuffmanCountsForRle(alphabet_size, counts);
-    CreateHuffmanTree(counts, alphabet_size, tree_limit, &code->depth_[0]);
-  } else {
-    CreateHuffmanTree(&histogram.data_[0], alphabet_size, tree_limit,
-                      &code->depth_[0]);
-  }
-  ConvertBitDepthsToSymbols(&code->depth_[0], alphabet_size, &code->bits_[0]);
-}
-
 static const int kCodeLengthCodes = 18;
 
 // Literal entropy code.