github.com/corona10/go@v0.0.0-20180224231303-7a218942be57/src/compress/bzip2/huffman.go (about) 1 // Copyright 2011 The Go Authors. All rights reserved. 2 // Use of this source code is governed by a BSD-style 3 // license that can be found in the LICENSE file. 4 5 package bzip2 6 7 import "sort" 8 9 // A huffmanTree is a binary tree which is navigated, bit-by-bit to reach a 10 // symbol. 11 type huffmanTree struct { 12 // nodes contains all the non-leaf nodes in the tree. nodes[0] is the 13 // root of the tree and nextNode contains the index of the next element 14 // of nodes to use when the tree is being constructed. 15 nodes []huffmanNode 16 nextNode int 17 } 18 19 // A huffmanNode is a node in the tree. left and right contain indexes into the 20 // nodes slice of the tree. If left or right is invalidNodeValue then the child 21 // is a left node and its value is in leftValue/rightValue. 22 // 23 // The symbols are uint16s because bzip2 encodes not only MTF indexes in the 24 // tree, but also two magic values for run-length encoding and an EOF symbol. 25 // Thus there are more than 256 possible symbols. 26 type huffmanNode struct { 27 left, right uint16 28 leftValue, rightValue uint16 29 } 30 31 // invalidNodeValue is an invalid index which marks a leaf node in the tree. 32 const invalidNodeValue = 0xffff 33 34 // Decode reads bits from the given bitReader and navigates the tree until a 35 // symbol is found. 36 func (t *huffmanTree) Decode(br *bitReader) (v uint16) { 37 nodeIndex := uint16(0) // node 0 is the root of the tree. 38 39 for { 40 node := &t.nodes[nodeIndex] 41 42 var bit uint16 43 if br.bits > 0 { 44 // Get next bit - fast path. 45 br.bits-- 46 bit = 0 - (uint16(br.n>>br.bits) & 1) 47 } else { 48 // Get next bit - slow path. 49 // Use ReadBits to retrieve a single bit 50 // from the underling io.ByteReader. 51 bit = 0 - uint16(br.ReadBits(1)) 52 } 53 // now 54 // bit = 0xffff if the next bit was 1 55 // bit = 0x0000 if the next bit was 0 56 57 // 1 means left, 0 means right. 58 // 59 // if bit == 0xffff { 60 // nodeIndex = node.left 61 // } else { 62 // nodeIndex = node.right 63 // } 64 nodeIndex = (bit & node.left) | (^bit & node.right) 65 66 if nodeIndex == invalidNodeValue { 67 // We found a leaf. Use the value of bit to decide 68 // whether is a left or a right value. 69 return (bit & node.leftValue) | (^bit & node.rightValue) 70 } 71 } 72 } 73 74 // newHuffmanTree builds a Huffman tree from a slice containing the code 75 // lengths of each symbol. The maximum code length is 32 bits. 76 func newHuffmanTree(lengths []uint8) (huffmanTree, error) { 77 // There are many possible trees that assign the same code length to 78 // each symbol (consider reflecting a tree down the middle, for 79 // example). Since the code length assignments determine the 80 // efficiency of the tree, each of these trees is equally good. In 81 // order to minimize the amount of information needed to build a tree 82 // bzip2 uses a canonical tree so that it can be reconstructed given 83 // only the code length assignments. 84 85 if len(lengths) < 2 { 86 panic("newHuffmanTree: too few symbols") 87 } 88 89 var t huffmanTree 90 91 // First we sort the code length assignments by ascending code length, 92 // using the symbol value to break ties. 93 pairs := make([]huffmanSymbolLengthPair, len(lengths)) 94 for i, length := range lengths { 95 pairs[i].value = uint16(i) 96 pairs[i].length = length 97 } 98 99 sort.Slice(pairs, func(i, j int) bool { 100 if pairs[i].length < pairs[j].length { 101 return true 102 } 103 if pairs[i].length > pairs[j].length { 104 return false 105 } 106 if pairs[i].value < pairs[j].value { 107 return true 108 } 109 return false 110 }) 111 112 // Now we assign codes to the symbols, starting with the longest code. 113 // We keep the codes packed into a uint32, at the most-significant end. 114 // So branches are taken from the MSB downwards. This makes it easy to 115 // sort them later. 116 code := uint32(0) 117 length := uint8(32) 118 119 codes := make([]huffmanCode, len(lengths)) 120 for i := len(pairs) - 1; i >= 0; i-- { 121 if length > pairs[i].length { 122 length = pairs[i].length 123 } 124 codes[i].code = code 125 codes[i].codeLen = length 126 codes[i].value = pairs[i].value 127 // We need to 'increment' the code, which means treating |code| 128 // like a |length| bit number. 129 code += 1 << (32 - length) 130 } 131 132 // Now we can sort by the code so that the left half of each branch are 133 // grouped together, recursively. 134 sort.Slice(codes, func(i, j int) bool { 135 return codes[i].code < codes[j].code 136 }) 137 138 t.nodes = make([]huffmanNode, len(codes)) 139 _, err := buildHuffmanNode(&t, codes, 0) 140 return t, err 141 } 142 143 // huffmanSymbolLengthPair contains a symbol and its code length. 144 type huffmanSymbolLengthPair struct { 145 value uint16 146 length uint8 147 } 148 149 // huffmanCode contains a symbol, its code and code length. 150 type huffmanCode struct { 151 code uint32 152 codeLen uint8 153 value uint16 154 } 155 156 // buildHuffmanNode takes a slice of sorted huffmanCodes and builds a node in 157 // the Huffman tree at the given level. It returns the index of the newly 158 // constructed node. 159 func buildHuffmanNode(t *huffmanTree, codes []huffmanCode, level uint32) (nodeIndex uint16, err error) { 160 test := uint32(1) << (31 - level) 161 162 // We have to search the list of codes to find the divide between the left and right sides. 163 firstRightIndex := len(codes) 164 for i, code := range codes { 165 if code.code&test != 0 { 166 firstRightIndex = i 167 break 168 } 169 } 170 171 left := codes[:firstRightIndex] 172 right := codes[firstRightIndex:] 173 174 if len(left) == 0 || len(right) == 0 { 175 // There is a superfluous level in the Huffman tree indicating 176 // a bug in the encoder. However, this bug has been observed in 177 // the wild so we handle it. 178 179 // If this function was called recursively then we know that 180 // len(codes) >= 2 because, otherwise, we would have hit the 181 // "leaf node" case, below, and not recursed. 182 // 183 // However, for the initial call it's possible that len(codes) 184 // is zero or one. Both cases are invalid because a zero length 185 // tree cannot encode anything and a length-1 tree can only 186 // encode EOF and so is superfluous. We reject both. 187 if len(codes) < 2 { 188 return 0, StructuralError("empty Huffman tree") 189 } 190 191 // In this case the recursion doesn't always reduce the length 192 // of codes so we need to ensure termination via another 193 // mechanism. 194 if level == 31 { 195 // Since len(codes) >= 2 the only way that the values 196 // can match at all 32 bits is if they are equal, which 197 // is invalid. This ensures that we never enter 198 // infinite recursion. 199 return 0, StructuralError("equal symbols in Huffman tree") 200 } 201 202 if len(left) == 0 { 203 return buildHuffmanNode(t, right, level+1) 204 } 205 return buildHuffmanNode(t, left, level+1) 206 } 207 208 nodeIndex = uint16(t.nextNode) 209 node := &t.nodes[t.nextNode] 210 t.nextNode++ 211 212 if len(left) == 1 { 213 // leaf node 214 node.left = invalidNodeValue 215 node.leftValue = left[0].value 216 } else { 217 node.left, err = buildHuffmanNode(t, left, level+1) 218 } 219 220 if err != nil { 221 return 222 } 223 224 if len(right) == 1 { 225 // leaf node 226 node.right = invalidNodeValue 227 node.rightValue = right[0].value 228 } else { 229 node.right, err = buildHuffmanNode(t, right, level+1) 230 } 231 232 return 233 }