Scott Baker | ed4efab | 2020-01-13 19:12:25 -0800 | [diff] [blame] | 1 | package fse |
| 2 | |
| 3 | import ( |
| 4 | "errors" |
| 5 | "fmt" |
| 6 | ) |
| 7 | |
| 8 | const ( |
| 9 | tablelogAbsoluteMax = 15 |
| 10 | ) |
| 11 | |
| 12 | // Decompress a block of data. |
| 13 | // You can provide a scratch buffer to avoid allocations. |
| 14 | // If nil is provided a temporary one will be allocated. |
| 15 | // It is possible, but by no way guaranteed that corrupt data will |
| 16 | // return an error. |
| 17 | // It is up to the caller to verify integrity of the returned data. |
| 18 | // Use a predefined Scrach to set maximum acceptable output size. |
| 19 | func Decompress(b []byte, s *Scratch) ([]byte, error) { |
| 20 | s, err := s.prepare(b) |
| 21 | if err != nil { |
| 22 | return nil, err |
| 23 | } |
| 24 | s.Out = s.Out[:0] |
| 25 | err = s.readNCount() |
| 26 | if err != nil { |
| 27 | return nil, err |
| 28 | } |
| 29 | err = s.buildDtable() |
| 30 | if err != nil { |
| 31 | return nil, err |
| 32 | } |
| 33 | err = s.decompress() |
| 34 | if err != nil { |
| 35 | return nil, err |
| 36 | } |
| 37 | |
| 38 | return s.Out, nil |
| 39 | } |
| 40 | |
| 41 | // readNCount will read the symbol distribution so decoding tables can be constructed. |
| 42 | func (s *Scratch) readNCount() error { |
| 43 | var ( |
| 44 | charnum uint16 |
| 45 | previous0 bool |
| 46 | b = &s.br |
| 47 | ) |
| 48 | iend := b.remain() |
| 49 | if iend < 4 { |
| 50 | return errors.New("input too small") |
| 51 | } |
| 52 | bitStream := b.Uint32() |
| 53 | nbBits := uint((bitStream & 0xF) + minTablelog) // extract tableLog |
| 54 | if nbBits > tablelogAbsoluteMax { |
| 55 | return errors.New("tableLog too large") |
| 56 | } |
| 57 | bitStream >>= 4 |
| 58 | bitCount := uint(4) |
| 59 | |
| 60 | s.actualTableLog = uint8(nbBits) |
| 61 | remaining := int32((1 << nbBits) + 1) |
| 62 | threshold := int32(1 << nbBits) |
| 63 | gotTotal := int32(0) |
| 64 | nbBits++ |
| 65 | |
| 66 | for remaining > 1 { |
| 67 | if previous0 { |
| 68 | n0 := charnum |
| 69 | for (bitStream & 0xFFFF) == 0xFFFF { |
| 70 | n0 += 24 |
| 71 | if b.off < iend-5 { |
| 72 | b.advance(2) |
| 73 | bitStream = b.Uint32() >> bitCount |
| 74 | } else { |
| 75 | bitStream >>= 16 |
| 76 | bitCount += 16 |
| 77 | } |
| 78 | } |
| 79 | for (bitStream & 3) == 3 { |
| 80 | n0 += 3 |
| 81 | bitStream >>= 2 |
| 82 | bitCount += 2 |
| 83 | } |
| 84 | n0 += uint16(bitStream & 3) |
| 85 | bitCount += 2 |
| 86 | if n0 > maxSymbolValue { |
| 87 | return errors.New("maxSymbolValue too small") |
| 88 | } |
| 89 | for charnum < n0 { |
| 90 | s.norm[charnum&0xff] = 0 |
| 91 | charnum++ |
| 92 | } |
| 93 | |
| 94 | if b.off <= iend-7 || b.off+int(bitCount>>3) <= iend-4 { |
| 95 | b.advance(bitCount >> 3) |
| 96 | bitCount &= 7 |
| 97 | bitStream = b.Uint32() >> bitCount |
| 98 | } else { |
| 99 | bitStream >>= 2 |
| 100 | } |
| 101 | } |
| 102 | |
| 103 | max := (2*(threshold) - 1) - (remaining) |
| 104 | var count int32 |
| 105 | |
| 106 | if (int32(bitStream) & (threshold - 1)) < max { |
| 107 | count = int32(bitStream) & (threshold - 1) |
| 108 | bitCount += nbBits - 1 |
| 109 | } else { |
| 110 | count = int32(bitStream) & (2*threshold - 1) |
| 111 | if count >= threshold { |
| 112 | count -= max |
| 113 | } |
| 114 | bitCount += nbBits |
| 115 | } |
| 116 | |
| 117 | count-- // extra accuracy |
| 118 | if count < 0 { |
| 119 | // -1 means +1 |
| 120 | remaining += count |
| 121 | gotTotal -= count |
| 122 | } else { |
| 123 | remaining -= count |
| 124 | gotTotal += count |
| 125 | } |
| 126 | s.norm[charnum&0xff] = int16(count) |
| 127 | charnum++ |
| 128 | previous0 = count == 0 |
| 129 | for remaining < threshold { |
| 130 | nbBits-- |
| 131 | threshold >>= 1 |
| 132 | } |
| 133 | if b.off <= iend-7 || b.off+int(bitCount>>3) <= iend-4 { |
| 134 | b.advance(bitCount >> 3) |
| 135 | bitCount &= 7 |
| 136 | } else { |
| 137 | bitCount -= (uint)(8 * (len(b.b) - 4 - b.off)) |
| 138 | b.off = len(b.b) - 4 |
| 139 | } |
| 140 | bitStream = b.Uint32() >> (bitCount & 31) |
| 141 | } |
| 142 | s.symbolLen = charnum |
| 143 | |
| 144 | if s.symbolLen <= 1 { |
| 145 | return fmt.Errorf("symbolLen (%d) too small", s.symbolLen) |
| 146 | } |
| 147 | if s.symbolLen > maxSymbolValue+1 { |
| 148 | return fmt.Errorf("symbolLen (%d) too big", s.symbolLen) |
| 149 | } |
| 150 | if remaining != 1 { |
| 151 | return fmt.Errorf("corruption detected (remaining %d != 1)", remaining) |
| 152 | } |
| 153 | if bitCount > 32 { |
| 154 | return fmt.Errorf("corruption detected (bitCount %d > 32)", bitCount) |
| 155 | } |
| 156 | if gotTotal != 1<<s.actualTableLog { |
| 157 | return fmt.Errorf("corruption detected (total %d != %d)", gotTotal, 1<<s.actualTableLog) |
| 158 | } |
| 159 | b.advance((bitCount + 7) >> 3) |
| 160 | return nil |
| 161 | } |
| 162 | |
| 163 | // decSymbol contains information about a state entry, |
| 164 | // Including the state offset base, the output symbol and |
| 165 | // the number of bits to read for the low part of the destination state. |
| 166 | type decSymbol struct { |
| 167 | newState uint16 |
| 168 | symbol uint8 |
| 169 | nbBits uint8 |
| 170 | } |
| 171 | |
| 172 | // allocDtable will allocate decoding tables if they are not big enough. |
| 173 | func (s *Scratch) allocDtable() { |
| 174 | tableSize := 1 << s.actualTableLog |
| 175 | if cap(s.decTable) < int(tableSize) { |
| 176 | s.decTable = make([]decSymbol, tableSize) |
| 177 | } |
| 178 | s.decTable = s.decTable[:tableSize] |
| 179 | |
| 180 | if cap(s.ct.tableSymbol) < 256 { |
| 181 | s.ct.tableSymbol = make([]byte, 256) |
| 182 | } |
| 183 | s.ct.tableSymbol = s.ct.tableSymbol[:256] |
| 184 | |
| 185 | if cap(s.ct.stateTable) < 256 { |
| 186 | s.ct.stateTable = make([]uint16, 256) |
| 187 | } |
| 188 | s.ct.stateTable = s.ct.stateTable[:256] |
| 189 | } |
| 190 | |
| 191 | // buildDtable will build the decoding table. |
| 192 | func (s *Scratch) buildDtable() error { |
| 193 | tableSize := uint32(1 << s.actualTableLog) |
| 194 | highThreshold := tableSize - 1 |
| 195 | s.allocDtable() |
| 196 | symbolNext := s.ct.stateTable[:256] |
| 197 | |
| 198 | // Init, lay down lowprob symbols |
| 199 | s.zeroBits = false |
| 200 | { |
| 201 | largeLimit := int16(1 << (s.actualTableLog - 1)) |
| 202 | for i, v := range s.norm[:s.symbolLen] { |
| 203 | if v == -1 { |
| 204 | s.decTable[highThreshold].symbol = uint8(i) |
| 205 | highThreshold-- |
| 206 | symbolNext[i] = 1 |
| 207 | } else { |
| 208 | if v >= largeLimit { |
| 209 | s.zeroBits = true |
| 210 | } |
| 211 | symbolNext[i] = uint16(v) |
| 212 | } |
| 213 | } |
| 214 | } |
| 215 | // Spread symbols |
| 216 | { |
| 217 | tableMask := tableSize - 1 |
| 218 | step := tableStep(tableSize) |
| 219 | position := uint32(0) |
| 220 | for ss, v := range s.norm[:s.symbolLen] { |
| 221 | for i := 0; i < int(v); i++ { |
| 222 | s.decTable[position].symbol = uint8(ss) |
| 223 | position = (position + step) & tableMask |
| 224 | for position > highThreshold { |
| 225 | // lowprob area |
| 226 | position = (position + step) & tableMask |
| 227 | } |
| 228 | } |
| 229 | } |
| 230 | if position != 0 { |
| 231 | // position must reach all cells once, otherwise normalizedCounter is incorrect |
| 232 | return errors.New("corrupted input (position != 0)") |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | // Build Decoding table |
| 237 | { |
| 238 | tableSize := uint16(1 << s.actualTableLog) |
| 239 | for u, v := range s.decTable { |
| 240 | symbol := v.symbol |
| 241 | nextState := symbolNext[symbol] |
| 242 | symbolNext[symbol] = nextState + 1 |
| 243 | nBits := s.actualTableLog - byte(highBits(uint32(nextState))) |
| 244 | s.decTable[u].nbBits = nBits |
| 245 | newState := (nextState << nBits) - tableSize |
| 246 | if newState >= tableSize { |
| 247 | return fmt.Errorf("newState (%d) outside table size (%d)", newState, tableSize) |
| 248 | } |
| 249 | if newState == uint16(u) && nBits == 0 { |
| 250 | // Seems weird that this is possible with nbits > 0. |
| 251 | return fmt.Errorf("newState (%d) == oldState (%d) and no bits", newState, u) |
| 252 | } |
| 253 | s.decTable[u].newState = newState |
| 254 | } |
| 255 | } |
| 256 | return nil |
| 257 | } |
| 258 | |
| 259 | // decompress will decompress the bitstream. |
| 260 | // If the buffer is over-read an error is returned. |
| 261 | func (s *Scratch) decompress() error { |
| 262 | br := &s.bits |
| 263 | br.init(s.br.unread()) |
| 264 | |
| 265 | var s1, s2 decoder |
| 266 | // Initialize and decode first state and symbol. |
| 267 | s1.init(br, s.decTable, s.actualTableLog) |
| 268 | s2.init(br, s.decTable, s.actualTableLog) |
| 269 | |
| 270 | // Use temp table to avoid bound checks/append penalty. |
| 271 | var tmp = s.ct.tableSymbol[:256] |
| 272 | var off uint8 |
| 273 | |
| 274 | // Main part |
| 275 | if !s.zeroBits { |
| 276 | for br.off >= 8 { |
| 277 | br.fillFast() |
| 278 | tmp[off+0] = s1.nextFast() |
| 279 | tmp[off+1] = s2.nextFast() |
| 280 | br.fillFast() |
| 281 | tmp[off+2] = s1.nextFast() |
| 282 | tmp[off+3] = s2.nextFast() |
| 283 | off += 4 |
| 284 | // When off is 0, we have overflowed and should write. |
| 285 | if off == 0 { |
| 286 | s.Out = append(s.Out, tmp...) |
| 287 | if len(s.Out) >= s.DecompressLimit { |
| 288 | return fmt.Errorf("output size (%d) > DecompressLimit (%d)", len(s.Out), s.DecompressLimit) |
| 289 | } |
| 290 | } |
| 291 | } |
| 292 | } else { |
| 293 | for br.off >= 8 { |
| 294 | br.fillFast() |
| 295 | tmp[off+0] = s1.next() |
| 296 | tmp[off+1] = s2.next() |
| 297 | br.fillFast() |
| 298 | tmp[off+2] = s1.next() |
| 299 | tmp[off+3] = s2.next() |
| 300 | off += 4 |
| 301 | if off == 0 { |
| 302 | s.Out = append(s.Out, tmp...) |
| 303 | // When off is 0, we have overflowed and should write. |
| 304 | if len(s.Out) >= s.DecompressLimit { |
| 305 | return fmt.Errorf("output size (%d) > DecompressLimit (%d)", len(s.Out), s.DecompressLimit) |
| 306 | } |
| 307 | } |
| 308 | } |
| 309 | } |
| 310 | s.Out = append(s.Out, tmp[:off]...) |
| 311 | |
| 312 | // Final bits, a bit more expensive check |
| 313 | for { |
| 314 | if s1.finished() { |
| 315 | s.Out = append(s.Out, s1.final(), s2.final()) |
| 316 | break |
| 317 | } |
| 318 | br.fill() |
| 319 | s.Out = append(s.Out, s1.next()) |
| 320 | if s2.finished() { |
| 321 | s.Out = append(s.Out, s2.final(), s1.final()) |
| 322 | break |
| 323 | } |
| 324 | s.Out = append(s.Out, s2.next()) |
| 325 | if len(s.Out) >= s.DecompressLimit { |
| 326 | return fmt.Errorf("output size (%d) > DecompressLimit (%d)", len(s.Out), s.DecompressLimit) |
| 327 | } |
| 328 | } |
| 329 | return br.close() |
| 330 | } |
| 331 | |
| 332 | // decoder keeps track of the current state and updates it from the bitstream. |
| 333 | type decoder struct { |
| 334 | state uint16 |
| 335 | br *bitReader |
| 336 | dt []decSymbol |
| 337 | } |
| 338 | |
| 339 | // init will initialize the decoder and read the first state from the stream. |
| 340 | func (d *decoder) init(in *bitReader, dt []decSymbol, tableLog uint8) { |
| 341 | d.dt = dt |
| 342 | d.br = in |
| 343 | d.state = uint16(in.getBits(tableLog)) |
| 344 | } |
| 345 | |
| 346 | // next returns the next symbol and sets the next state. |
| 347 | // At least tablelog bits must be available in the bit reader. |
| 348 | func (d *decoder) next() uint8 { |
| 349 | n := &d.dt[d.state] |
| 350 | lowBits := d.br.getBits(n.nbBits) |
| 351 | d.state = n.newState + lowBits |
| 352 | return n.symbol |
| 353 | } |
| 354 | |
| 355 | // finished returns true if all bits have been read from the bitstream |
| 356 | // and the next state would require reading bits from the input. |
| 357 | func (d *decoder) finished() bool { |
| 358 | return d.br.finished() && d.dt[d.state].nbBits > 0 |
| 359 | } |
| 360 | |
| 361 | // final returns the current state symbol without decoding the next. |
| 362 | func (d *decoder) final() uint8 { |
| 363 | return d.dt[d.state].symbol |
| 364 | } |
| 365 | |
| 366 | // nextFast returns the next symbol and sets the next state. |
| 367 | // This can only be used if no symbols are 0 bits. |
| 368 | // At least tablelog bits must be available in the bit reader. |
| 369 | func (d *decoder) nextFast() uint8 { |
| 370 | n := d.dt[d.state] |
| 371 | lowBits := d.br.getBitsFast(n.nbBits) |
| 372 | d.state = n.newState + lowBits |
| 373 | return n.symbol |
| 374 | } |