khenaidoo | 106c61a | 2021-08-11 18:05:46 -0400 | [diff] [blame] | 1 | // Copyright 2019+ Klaus Post. All rights reserved. |
| 2 | // License information can be found in the LICENSE file. |
| 3 | // Based on work by Yann Collet, released under BSD License. |
| 4 | |
| 5 | package zstd |
| 6 | |
| 7 | import "fmt" |
| 8 | |
| 9 | const ( |
| 10 | dFastLongTableBits = 17 // Bits used in the long match table |
| 11 | dFastLongTableSize = 1 << dFastLongTableBits // Size of the table |
| 12 | dFastLongTableMask = dFastLongTableSize - 1 // Mask for table indices. Redundant, but can eliminate bounds checks. |
| 13 | |
| 14 | dLongTableShardCnt = 1 << (dFastLongTableBits - dictShardBits) // Number of shards in the table |
| 15 | dLongTableShardSize = dFastLongTableSize / tableShardCnt // Size of an individual shard |
| 16 | |
| 17 | dFastShortTableBits = tableBits // Bits used in the short match table |
| 18 | dFastShortTableSize = 1 << dFastShortTableBits // Size of the table |
| 19 | dFastShortTableMask = dFastShortTableSize - 1 // Mask for table indices. Redundant, but can eliminate bounds checks. |
| 20 | ) |
| 21 | |
| 22 | type doubleFastEncoder struct { |
| 23 | fastEncoder |
| 24 | longTable [dFastLongTableSize]tableEntry |
| 25 | } |
| 26 | |
| 27 | type doubleFastEncoderDict struct { |
| 28 | fastEncoderDict |
| 29 | longTable [dFastLongTableSize]tableEntry |
| 30 | dictLongTable []tableEntry |
| 31 | longTableShardDirty [dLongTableShardCnt]bool |
| 32 | } |
| 33 | |
| 34 | // Encode mimmics functionality in zstd_dfast.c |
| 35 | func (e *doubleFastEncoder) Encode(blk *blockEnc, src []byte) { |
| 36 | const ( |
| 37 | // Input margin is the number of bytes we read (8) |
| 38 | // and the maximum we will read ahead (2) |
| 39 | inputMargin = 8 + 2 |
| 40 | minNonLiteralBlockSize = 16 |
| 41 | ) |
| 42 | |
| 43 | // Protect against e.cur wraparound. |
| 44 | for e.cur >= bufferReset { |
| 45 | if len(e.hist) == 0 { |
| 46 | for i := range e.table[:] { |
| 47 | e.table[i] = tableEntry{} |
| 48 | } |
| 49 | for i := range e.longTable[:] { |
| 50 | e.longTable[i] = tableEntry{} |
| 51 | } |
| 52 | e.cur = e.maxMatchOff |
| 53 | break |
| 54 | } |
| 55 | // Shift down everything in the table that isn't already too far away. |
| 56 | minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff |
| 57 | for i := range e.table[:] { |
| 58 | v := e.table[i].offset |
| 59 | if v < minOff { |
| 60 | v = 0 |
| 61 | } else { |
| 62 | v = v - e.cur + e.maxMatchOff |
| 63 | } |
| 64 | e.table[i].offset = v |
| 65 | } |
| 66 | for i := range e.longTable[:] { |
| 67 | v := e.longTable[i].offset |
| 68 | if v < minOff { |
| 69 | v = 0 |
| 70 | } else { |
| 71 | v = v - e.cur + e.maxMatchOff |
| 72 | } |
| 73 | e.longTable[i].offset = v |
| 74 | } |
| 75 | e.cur = e.maxMatchOff |
| 76 | break |
| 77 | } |
| 78 | |
| 79 | s := e.addBlock(src) |
| 80 | blk.size = len(src) |
| 81 | if len(src) < minNonLiteralBlockSize { |
| 82 | blk.extraLits = len(src) |
| 83 | blk.literals = blk.literals[:len(src)] |
| 84 | copy(blk.literals, src) |
| 85 | return |
| 86 | } |
| 87 | |
| 88 | // Override src |
| 89 | src = e.hist |
| 90 | sLimit := int32(len(src)) - inputMargin |
| 91 | // stepSize is the number of bytes to skip on every main loop iteration. |
| 92 | // It should be >= 1. |
| 93 | const stepSize = 1 |
| 94 | |
| 95 | const kSearchStrength = 8 |
| 96 | |
| 97 | // nextEmit is where in src the next emitLiteral should start from. |
| 98 | nextEmit := s |
| 99 | cv := load6432(src, s) |
| 100 | |
| 101 | // Relative offsets |
| 102 | offset1 := int32(blk.recentOffsets[0]) |
| 103 | offset2 := int32(blk.recentOffsets[1]) |
| 104 | |
| 105 | addLiterals := func(s *seq, until int32) { |
| 106 | if until == nextEmit { |
| 107 | return |
| 108 | } |
| 109 | blk.literals = append(blk.literals, src[nextEmit:until]...) |
| 110 | s.litLen = uint32(until - nextEmit) |
| 111 | } |
| 112 | if debug { |
| 113 | println("recent offsets:", blk.recentOffsets) |
| 114 | } |
| 115 | |
| 116 | encodeLoop: |
| 117 | for { |
| 118 | var t int32 |
| 119 | // We allow the encoder to optionally turn off repeat offsets across blocks |
| 120 | canRepeat := len(blk.sequences) > 2 |
| 121 | |
| 122 | for { |
| 123 | if debugAsserts && canRepeat && offset1 == 0 { |
| 124 | panic("offset0 was 0") |
| 125 | } |
| 126 | |
| 127 | nextHashS := hash5(cv, dFastShortTableBits) |
| 128 | nextHashL := hash8(cv, dFastLongTableBits) |
| 129 | candidateL := e.longTable[nextHashL] |
| 130 | candidateS := e.table[nextHashS] |
| 131 | |
| 132 | const repOff = 1 |
| 133 | repIndex := s - offset1 + repOff |
| 134 | entry := tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 135 | e.longTable[nextHashL] = entry |
| 136 | e.table[nextHashS] = entry |
| 137 | |
| 138 | if canRepeat { |
| 139 | if repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>(repOff*8)) { |
| 140 | // Consider history as well. |
| 141 | var seq seq |
| 142 | lenght := 4 + e.matchlen(s+4+repOff, repIndex+4, src) |
| 143 | |
| 144 | seq.matchLen = uint32(lenght - zstdMinMatch) |
| 145 | |
| 146 | // We might be able to match backwards. |
| 147 | // Extend as long as we can. |
| 148 | start := s + repOff |
| 149 | // We end the search early, so we don't risk 0 literals |
| 150 | // and have to do special offset treatment. |
| 151 | startLimit := nextEmit + 1 |
| 152 | |
| 153 | tMin := s - e.maxMatchOff |
| 154 | if tMin < 0 { |
| 155 | tMin = 0 |
| 156 | } |
| 157 | for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-1 { |
| 158 | repIndex-- |
| 159 | start-- |
| 160 | seq.matchLen++ |
| 161 | } |
| 162 | addLiterals(&seq, start) |
| 163 | |
| 164 | // rep 0 |
| 165 | seq.offset = 1 |
| 166 | if debugSequences { |
| 167 | println("repeat sequence", seq, "next s:", s) |
| 168 | } |
| 169 | blk.sequences = append(blk.sequences, seq) |
| 170 | s += lenght + repOff |
| 171 | nextEmit = s |
| 172 | if s >= sLimit { |
| 173 | if debug { |
| 174 | println("repeat ended", s, lenght) |
| 175 | |
| 176 | } |
| 177 | break encodeLoop |
| 178 | } |
| 179 | cv = load6432(src, s) |
| 180 | continue |
| 181 | } |
| 182 | } |
| 183 | // Find the offsets of our two matches. |
| 184 | coffsetL := s - (candidateL.offset - e.cur) |
| 185 | coffsetS := s - (candidateS.offset - e.cur) |
| 186 | |
| 187 | // Check if we have a long match. |
| 188 | if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val { |
| 189 | // Found a long match, likely at least 8 bytes. |
| 190 | // Reference encoder checks all 8 bytes, we only check 4, |
| 191 | // but the likelihood of both the first 4 bytes and the hash matching should be enough. |
| 192 | t = candidateL.offset - e.cur |
| 193 | if debugAsserts && s <= t { |
| 194 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 195 | } |
| 196 | if debugAsserts && s-t > e.maxMatchOff { |
| 197 | panic("s - t >e.maxMatchOff") |
| 198 | } |
| 199 | if debugMatches { |
| 200 | println("long match") |
| 201 | } |
| 202 | break |
| 203 | } |
| 204 | |
| 205 | // Check if we have a short match. |
| 206 | if coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val { |
| 207 | // found a regular match |
| 208 | // See if we can find a long match at s+1 |
| 209 | const checkAt = 1 |
| 210 | cv := load6432(src, s+checkAt) |
| 211 | nextHashL = hash8(cv, dFastLongTableBits) |
| 212 | candidateL = e.longTable[nextHashL] |
| 213 | coffsetL = s - (candidateL.offset - e.cur) + checkAt |
| 214 | |
| 215 | // We can store it, since we have at least a 4 byte match. |
| 216 | e.longTable[nextHashL] = tableEntry{offset: s + checkAt + e.cur, val: uint32(cv)} |
| 217 | if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val { |
| 218 | // Found a long match, likely at least 8 bytes. |
| 219 | // Reference encoder checks all 8 bytes, we only check 4, |
| 220 | // but the likelihood of both the first 4 bytes and the hash matching should be enough. |
| 221 | t = candidateL.offset - e.cur |
| 222 | s += checkAt |
| 223 | if debugMatches { |
| 224 | println("long match (after short)") |
| 225 | } |
| 226 | break |
| 227 | } |
| 228 | |
| 229 | t = candidateS.offset - e.cur |
| 230 | if debugAsserts && s <= t { |
| 231 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 232 | } |
| 233 | if debugAsserts && s-t > e.maxMatchOff { |
| 234 | panic("s - t >e.maxMatchOff") |
| 235 | } |
| 236 | if debugAsserts && t < 0 { |
| 237 | panic("t<0") |
| 238 | } |
| 239 | if debugMatches { |
| 240 | println("short match") |
| 241 | } |
| 242 | break |
| 243 | } |
| 244 | |
| 245 | // No match found, move forward in input. |
| 246 | s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) |
| 247 | if s >= sLimit { |
| 248 | break encodeLoop |
| 249 | } |
| 250 | cv = load6432(src, s) |
| 251 | } |
| 252 | |
| 253 | // A 4-byte match has been found. Update recent offsets. |
| 254 | // We'll later see if more than 4 bytes. |
| 255 | offset2 = offset1 |
| 256 | offset1 = s - t |
| 257 | |
| 258 | if debugAsserts && s <= t { |
| 259 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 260 | } |
| 261 | |
| 262 | if debugAsserts && canRepeat && int(offset1) > len(src) { |
| 263 | panic("invalid offset") |
| 264 | } |
| 265 | |
| 266 | // Extend the 4-byte match as long as possible. |
| 267 | l := e.matchlen(s+4, t+4, src) + 4 |
| 268 | |
| 269 | // Extend backwards |
| 270 | tMin := s - e.maxMatchOff |
| 271 | if tMin < 0 { |
| 272 | tMin = 0 |
| 273 | } |
| 274 | for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength { |
| 275 | s-- |
| 276 | t-- |
| 277 | l++ |
| 278 | } |
| 279 | |
| 280 | // Write our sequence |
| 281 | var seq seq |
| 282 | seq.litLen = uint32(s - nextEmit) |
| 283 | seq.matchLen = uint32(l - zstdMinMatch) |
| 284 | if seq.litLen > 0 { |
| 285 | blk.literals = append(blk.literals, src[nextEmit:s]...) |
| 286 | } |
| 287 | seq.offset = uint32(s-t) + 3 |
| 288 | s += l |
| 289 | if debugSequences { |
| 290 | println("sequence", seq, "next s:", s) |
| 291 | } |
| 292 | blk.sequences = append(blk.sequences, seq) |
| 293 | nextEmit = s |
| 294 | if s >= sLimit { |
| 295 | break encodeLoop |
| 296 | } |
| 297 | |
| 298 | // Index match start+1 (long) and start+2 (short) |
| 299 | index0 := s - l + 1 |
| 300 | // Index match end-2 (long) and end-1 (short) |
| 301 | index1 := s - 2 |
| 302 | |
| 303 | cv0 := load6432(src, index0) |
| 304 | cv1 := load6432(src, index1) |
| 305 | te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)} |
| 306 | te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)} |
| 307 | e.longTable[hash8(cv0, dFastLongTableBits)] = te0 |
| 308 | e.longTable[hash8(cv1, dFastLongTableBits)] = te1 |
| 309 | cv0 >>= 8 |
| 310 | cv1 >>= 8 |
| 311 | te0.offset++ |
| 312 | te1.offset++ |
| 313 | te0.val = uint32(cv0) |
| 314 | te1.val = uint32(cv1) |
| 315 | e.table[hash5(cv0, dFastShortTableBits)] = te0 |
| 316 | e.table[hash5(cv1, dFastShortTableBits)] = te1 |
| 317 | |
| 318 | cv = load6432(src, s) |
| 319 | |
| 320 | if !canRepeat { |
| 321 | continue |
| 322 | } |
| 323 | |
| 324 | // Check offset 2 |
| 325 | for { |
| 326 | o2 := s - offset2 |
| 327 | if load3232(src, o2) != uint32(cv) { |
| 328 | // Do regular search |
| 329 | break |
| 330 | } |
| 331 | |
| 332 | // Store this, since we have it. |
| 333 | nextHashS := hash5(cv, dFastShortTableBits) |
| 334 | nextHashL := hash8(cv, dFastLongTableBits) |
| 335 | |
| 336 | // We have at least 4 byte match. |
| 337 | // No need to check backwards. We come straight from a match |
| 338 | l := 4 + e.matchlen(s+4, o2+4, src) |
| 339 | |
| 340 | entry := tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 341 | e.longTable[nextHashL] = entry |
| 342 | e.table[nextHashS] = entry |
| 343 | seq.matchLen = uint32(l) - zstdMinMatch |
| 344 | seq.litLen = 0 |
| 345 | |
| 346 | // Since litlen is always 0, this is offset 1. |
| 347 | seq.offset = 1 |
| 348 | s += l |
| 349 | nextEmit = s |
| 350 | if debugSequences { |
| 351 | println("sequence", seq, "next s:", s) |
| 352 | } |
| 353 | blk.sequences = append(blk.sequences, seq) |
| 354 | |
| 355 | // Swap offset 1 and 2. |
| 356 | offset1, offset2 = offset2, offset1 |
| 357 | if s >= sLimit { |
| 358 | // Finished |
| 359 | break encodeLoop |
| 360 | } |
| 361 | cv = load6432(src, s) |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | if int(nextEmit) < len(src) { |
| 366 | blk.literals = append(blk.literals, src[nextEmit:]...) |
| 367 | blk.extraLits = len(src) - int(nextEmit) |
| 368 | } |
| 369 | blk.recentOffsets[0] = uint32(offset1) |
| 370 | blk.recentOffsets[1] = uint32(offset2) |
| 371 | if debug { |
| 372 | println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) |
| 373 | } |
| 374 | } |
| 375 | |
| 376 | // EncodeNoHist will encode a block with no history and no following blocks. |
| 377 | // Most notable difference is that src will not be copied for history and |
| 378 | // we do not need to check for max match length. |
| 379 | func (e *doubleFastEncoder) EncodeNoHist(blk *blockEnc, src []byte) { |
| 380 | const ( |
| 381 | // Input margin is the number of bytes we read (8) |
| 382 | // and the maximum we will read ahead (2) |
| 383 | inputMargin = 8 + 2 |
| 384 | minNonLiteralBlockSize = 16 |
| 385 | ) |
| 386 | |
| 387 | // Protect against e.cur wraparound. |
| 388 | if e.cur >= bufferReset { |
| 389 | for i := range e.table[:] { |
| 390 | e.table[i] = tableEntry{} |
| 391 | } |
| 392 | for i := range e.longTable[:] { |
| 393 | e.longTable[i] = tableEntry{} |
| 394 | } |
| 395 | e.cur = e.maxMatchOff |
| 396 | } |
| 397 | |
| 398 | s := int32(0) |
| 399 | blk.size = len(src) |
| 400 | if len(src) < minNonLiteralBlockSize { |
| 401 | blk.extraLits = len(src) |
| 402 | blk.literals = blk.literals[:len(src)] |
| 403 | copy(blk.literals, src) |
| 404 | return |
| 405 | } |
| 406 | |
| 407 | // Override src |
| 408 | sLimit := int32(len(src)) - inputMargin |
| 409 | // stepSize is the number of bytes to skip on every main loop iteration. |
| 410 | // It should be >= 1. |
| 411 | const stepSize = 1 |
| 412 | |
| 413 | const kSearchStrength = 8 |
| 414 | |
| 415 | // nextEmit is where in src the next emitLiteral should start from. |
| 416 | nextEmit := s |
| 417 | cv := load6432(src, s) |
| 418 | |
| 419 | // Relative offsets |
| 420 | offset1 := int32(blk.recentOffsets[0]) |
| 421 | offset2 := int32(blk.recentOffsets[1]) |
| 422 | |
| 423 | addLiterals := func(s *seq, until int32) { |
| 424 | if until == nextEmit { |
| 425 | return |
| 426 | } |
| 427 | blk.literals = append(blk.literals, src[nextEmit:until]...) |
| 428 | s.litLen = uint32(until - nextEmit) |
| 429 | } |
| 430 | if debug { |
| 431 | println("recent offsets:", blk.recentOffsets) |
| 432 | } |
| 433 | |
| 434 | encodeLoop: |
| 435 | for { |
| 436 | var t int32 |
| 437 | for { |
| 438 | |
| 439 | nextHashS := hash5(cv, dFastShortTableBits) |
| 440 | nextHashL := hash8(cv, dFastLongTableBits) |
| 441 | candidateL := e.longTable[nextHashL] |
| 442 | candidateS := e.table[nextHashS] |
| 443 | |
| 444 | const repOff = 1 |
| 445 | repIndex := s - offset1 + repOff |
| 446 | entry := tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 447 | e.longTable[nextHashL] = entry |
| 448 | e.table[nextHashS] = entry |
| 449 | |
| 450 | if len(blk.sequences) > 2 { |
| 451 | if load3232(src, repIndex) == uint32(cv>>(repOff*8)) { |
| 452 | // Consider history as well. |
| 453 | var seq seq |
| 454 | //length := 4 + e.matchlen(s+4+repOff, repIndex+4, src) |
| 455 | length := 4 + int32(matchLen(src[s+4+repOff:], src[repIndex+4:])) |
| 456 | |
| 457 | seq.matchLen = uint32(length - zstdMinMatch) |
| 458 | |
| 459 | // We might be able to match backwards. |
| 460 | // Extend as long as we can. |
| 461 | start := s + repOff |
| 462 | // We end the search early, so we don't risk 0 literals |
| 463 | // and have to do special offset treatment. |
| 464 | startLimit := nextEmit + 1 |
| 465 | |
| 466 | tMin := s - e.maxMatchOff |
| 467 | if tMin < 0 { |
| 468 | tMin = 0 |
| 469 | } |
| 470 | for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] { |
| 471 | repIndex-- |
| 472 | start-- |
| 473 | seq.matchLen++ |
| 474 | } |
| 475 | addLiterals(&seq, start) |
| 476 | |
| 477 | // rep 0 |
| 478 | seq.offset = 1 |
| 479 | if debugSequences { |
| 480 | println("repeat sequence", seq, "next s:", s) |
| 481 | } |
| 482 | blk.sequences = append(blk.sequences, seq) |
| 483 | s += length + repOff |
| 484 | nextEmit = s |
| 485 | if s >= sLimit { |
| 486 | if debug { |
| 487 | println("repeat ended", s, length) |
| 488 | |
| 489 | } |
| 490 | break encodeLoop |
| 491 | } |
| 492 | cv = load6432(src, s) |
| 493 | continue |
| 494 | } |
| 495 | } |
| 496 | // Find the offsets of our two matches. |
| 497 | coffsetL := s - (candidateL.offset - e.cur) |
| 498 | coffsetS := s - (candidateS.offset - e.cur) |
| 499 | |
| 500 | // Check if we have a long match. |
| 501 | if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val { |
| 502 | // Found a long match, likely at least 8 bytes. |
| 503 | // Reference encoder checks all 8 bytes, we only check 4, |
| 504 | // but the likelihood of both the first 4 bytes and the hash matching should be enough. |
| 505 | t = candidateL.offset - e.cur |
| 506 | if debugAsserts && s <= t { |
| 507 | panic(fmt.Sprintf("s (%d) <= t (%d). cur: %d", s, t, e.cur)) |
| 508 | } |
| 509 | if debugAsserts && s-t > e.maxMatchOff { |
| 510 | panic("s - t >e.maxMatchOff") |
| 511 | } |
| 512 | if debugMatches { |
| 513 | println("long match") |
| 514 | } |
| 515 | break |
| 516 | } |
| 517 | |
| 518 | // Check if we have a short match. |
| 519 | if coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val { |
| 520 | // found a regular match |
| 521 | // See if we can find a long match at s+1 |
| 522 | const checkAt = 1 |
| 523 | cv := load6432(src, s+checkAt) |
| 524 | nextHashL = hash8(cv, dFastLongTableBits) |
| 525 | candidateL = e.longTable[nextHashL] |
| 526 | coffsetL = s - (candidateL.offset - e.cur) + checkAt |
| 527 | |
| 528 | // We can store it, since we have at least a 4 byte match. |
| 529 | e.longTable[nextHashL] = tableEntry{offset: s + checkAt + e.cur, val: uint32(cv)} |
| 530 | if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val { |
| 531 | // Found a long match, likely at least 8 bytes. |
| 532 | // Reference encoder checks all 8 bytes, we only check 4, |
| 533 | // but the likelihood of both the first 4 bytes and the hash matching should be enough. |
| 534 | t = candidateL.offset - e.cur |
| 535 | s += checkAt |
| 536 | if debugMatches { |
| 537 | println("long match (after short)") |
| 538 | } |
| 539 | break |
| 540 | } |
| 541 | |
| 542 | t = candidateS.offset - e.cur |
| 543 | if debugAsserts && s <= t { |
| 544 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 545 | } |
| 546 | if debugAsserts && s-t > e.maxMatchOff { |
| 547 | panic("s - t >e.maxMatchOff") |
| 548 | } |
| 549 | if debugAsserts && t < 0 { |
| 550 | panic("t<0") |
| 551 | } |
| 552 | if debugMatches { |
| 553 | println("short match") |
| 554 | } |
| 555 | break |
| 556 | } |
| 557 | |
| 558 | // No match found, move forward in input. |
| 559 | s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) |
| 560 | if s >= sLimit { |
| 561 | break encodeLoop |
| 562 | } |
| 563 | cv = load6432(src, s) |
| 564 | } |
| 565 | |
| 566 | // A 4-byte match has been found. Update recent offsets. |
| 567 | // We'll later see if more than 4 bytes. |
| 568 | offset2 = offset1 |
| 569 | offset1 = s - t |
| 570 | |
| 571 | if debugAsserts && s <= t { |
| 572 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 573 | } |
| 574 | |
| 575 | // Extend the 4-byte match as long as possible. |
| 576 | //l := e.matchlen(s+4, t+4, src) + 4 |
| 577 | l := int32(matchLen(src[s+4:], src[t+4:])) + 4 |
| 578 | |
| 579 | // Extend backwards |
| 580 | tMin := s - e.maxMatchOff |
| 581 | if tMin < 0 { |
| 582 | tMin = 0 |
| 583 | } |
| 584 | for t > tMin && s > nextEmit && src[t-1] == src[s-1] { |
| 585 | s-- |
| 586 | t-- |
| 587 | l++ |
| 588 | } |
| 589 | |
| 590 | // Write our sequence |
| 591 | var seq seq |
| 592 | seq.litLen = uint32(s - nextEmit) |
| 593 | seq.matchLen = uint32(l - zstdMinMatch) |
| 594 | if seq.litLen > 0 { |
| 595 | blk.literals = append(blk.literals, src[nextEmit:s]...) |
| 596 | } |
| 597 | seq.offset = uint32(s-t) + 3 |
| 598 | s += l |
| 599 | if debugSequences { |
| 600 | println("sequence", seq, "next s:", s) |
| 601 | } |
| 602 | blk.sequences = append(blk.sequences, seq) |
| 603 | nextEmit = s |
| 604 | if s >= sLimit { |
| 605 | break encodeLoop |
| 606 | } |
| 607 | |
| 608 | // Index match start+1 (long) and start+2 (short) |
| 609 | index0 := s - l + 1 |
| 610 | // Index match end-2 (long) and end-1 (short) |
| 611 | index1 := s - 2 |
| 612 | |
| 613 | cv0 := load6432(src, index0) |
| 614 | cv1 := load6432(src, index1) |
| 615 | te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)} |
| 616 | te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)} |
| 617 | e.longTable[hash8(cv0, dFastLongTableBits)] = te0 |
| 618 | e.longTable[hash8(cv1, dFastLongTableBits)] = te1 |
| 619 | cv0 >>= 8 |
| 620 | cv1 >>= 8 |
| 621 | te0.offset++ |
| 622 | te1.offset++ |
| 623 | te0.val = uint32(cv0) |
| 624 | te1.val = uint32(cv1) |
| 625 | e.table[hash5(cv0, dFastShortTableBits)] = te0 |
| 626 | e.table[hash5(cv1, dFastShortTableBits)] = te1 |
| 627 | |
| 628 | cv = load6432(src, s) |
| 629 | |
| 630 | if len(blk.sequences) <= 2 { |
| 631 | continue |
| 632 | } |
| 633 | |
| 634 | // Check offset 2 |
| 635 | for { |
| 636 | o2 := s - offset2 |
| 637 | if load3232(src, o2) != uint32(cv) { |
| 638 | // Do regular search |
| 639 | break |
| 640 | } |
| 641 | |
| 642 | // Store this, since we have it. |
| 643 | nextHashS := hash5(cv1>>8, dFastShortTableBits) |
| 644 | nextHashL := hash8(cv, dFastLongTableBits) |
| 645 | |
| 646 | // We have at least 4 byte match. |
| 647 | // No need to check backwards. We come straight from a match |
| 648 | //l := 4 + e.matchlen(s+4, o2+4, src) |
| 649 | l := 4 + int32(matchLen(src[s+4:], src[o2+4:])) |
| 650 | |
| 651 | entry := tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 652 | e.longTable[nextHashL] = entry |
| 653 | e.table[nextHashS] = entry |
| 654 | seq.matchLen = uint32(l) - zstdMinMatch |
| 655 | seq.litLen = 0 |
| 656 | |
| 657 | // Since litlen is always 0, this is offset 1. |
| 658 | seq.offset = 1 |
| 659 | s += l |
| 660 | nextEmit = s |
| 661 | if debugSequences { |
| 662 | println("sequence", seq, "next s:", s) |
| 663 | } |
| 664 | blk.sequences = append(blk.sequences, seq) |
| 665 | |
| 666 | // Swap offset 1 and 2. |
| 667 | offset1, offset2 = offset2, offset1 |
| 668 | if s >= sLimit { |
| 669 | // Finished |
| 670 | break encodeLoop |
| 671 | } |
| 672 | cv = load6432(src, s) |
| 673 | } |
| 674 | } |
| 675 | |
| 676 | if int(nextEmit) < len(src) { |
| 677 | blk.literals = append(blk.literals, src[nextEmit:]...) |
| 678 | blk.extraLits = len(src) - int(nextEmit) |
| 679 | } |
| 680 | if debug { |
| 681 | println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) |
| 682 | } |
| 683 | |
| 684 | // We do not store history, so we must offset e.cur to avoid false matches for next user. |
| 685 | if e.cur < bufferReset { |
| 686 | e.cur += int32(len(src)) |
| 687 | } |
| 688 | } |
| 689 | |
| 690 | // Encode will encode the content, with a dictionary if initialized for it. |
| 691 | func (e *doubleFastEncoderDict) Encode(blk *blockEnc, src []byte) { |
| 692 | const ( |
| 693 | // Input margin is the number of bytes we read (8) |
| 694 | // and the maximum we will read ahead (2) |
| 695 | inputMargin = 8 + 2 |
| 696 | minNonLiteralBlockSize = 16 |
| 697 | ) |
| 698 | |
| 699 | // Protect against e.cur wraparound. |
| 700 | for e.cur >= bufferReset { |
| 701 | if len(e.hist) == 0 { |
| 702 | for i := range e.table[:] { |
| 703 | e.table[i] = tableEntry{} |
| 704 | } |
| 705 | for i := range e.longTable[:] { |
| 706 | e.longTable[i] = tableEntry{} |
| 707 | } |
| 708 | e.markAllShardsDirty() |
| 709 | e.cur = e.maxMatchOff |
| 710 | break |
| 711 | } |
| 712 | // Shift down everything in the table that isn't already too far away. |
| 713 | minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff |
| 714 | for i := range e.table[:] { |
| 715 | v := e.table[i].offset |
| 716 | if v < minOff { |
| 717 | v = 0 |
| 718 | } else { |
| 719 | v = v - e.cur + e.maxMatchOff |
| 720 | } |
| 721 | e.table[i].offset = v |
| 722 | } |
| 723 | for i := range e.longTable[:] { |
| 724 | v := e.longTable[i].offset |
| 725 | if v < minOff { |
| 726 | v = 0 |
| 727 | } else { |
| 728 | v = v - e.cur + e.maxMatchOff |
| 729 | } |
| 730 | e.longTable[i].offset = v |
| 731 | } |
| 732 | e.markAllShardsDirty() |
| 733 | e.cur = e.maxMatchOff |
| 734 | break |
| 735 | } |
| 736 | |
| 737 | s := e.addBlock(src) |
| 738 | blk.size = len(src) |
| 739 | if len(src) < minNonLiteralBlockSize { |
| 740 | blk.extraLits = len(src) |
| 741 | blk.literals = blk.literals[:len(src)] |
| 742 | copy(blk.literals, src) |
| 743 | return |
| 744 | } |
| 745 | |
| 746 | // Override src |
| 747 | src = e.hist |
| 748 | sLimit := int32(len(src)) - inputMargin |
| 749 | // stepSize is the number of bytes to skip on every main loop iteration. |
| 750 | // It should be >= 1. |
| 751 | const stepSize = 1 |
| 752 | |
| 753 | const kSearchStrength = 8 |
| 754 | |
| 755 | // nextEmit is where in src the next emitLiteral should start from. |
| 756 | nextEmit := s |
| 757 | cv := load6432(src, s) |
| 758 | |
| 759 | // Relative offsets |
| 760 | offset1 := int32(blk.recentOffsets[0]) |
| 761 | offset2 := int32(blk.recentOffsets[1]) |
| 762 | |
| 763 | addLiterals := func(s *seq, until int32) { |
| 764 | if until == nextEmit { |
| 765 | return |
| 766 | } |
| 767 | blk.literals = append(blk.literals, src[nextEmit:until]...) |
| 768 | s.litLen = uint32(until - nextEmit) |
| 769 | } |
| 770 | if debug { |
| 771 | println("recent offsets:", blk.recentOffsets) |
| 772 | } |
| 773 | |
| 774 | encodeLoop: |
| 775 | for { |
| 776 | var t int32 |
| 777 | // We allow the encoder to optionally turn off repeat offsets across blocks |
| 778 | canRepeat := len(blk.sequences) > 2 |
| 779 | |
| 780 | for { |
| 781 | if debugAsserts && canRepeat && offset1 == 0 { |
| 782 | panic("offset0 was 0") |
| 783 | } |
| 784 | |
| 785 | nextHashS := hash5(cv, dFastShortTableBits) |
| 786 | nextHashL := hash8(cv, dFastLongTableBits) |
| 787 | candidateL := e.longTable[nextHashL] |
| 788 | candidateS := e.table[nextHashS] |
| 789 | |
| 790 | const repOff = 1 |
| 791 | repIndex := s - offset1 + repOff |
| 792 | entry := tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 793 | e.longTable[nextHashL] = entry |
| 794 | e.markLongShardDirty(nextHashL) |
| 795 | e.table[nextHashS] = entry |
| 796 | e.markShardDirty(nextHashS) |
| 797 | |
| 798 | if canRepeat { |
| 799 | if repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>(repOff*8)) { |
| 800 | // Consider history as well. |
| 801 | var seq seq |
| 802 | lenght := 4 + e.matchlen(s+4+repOff, repIndex+4, src) |
| 803 | |
| 804 | seq.matchLen = uint32(lenght - zstdMinMatch) |
| 805 | |
| 806 | // We might be able to match backwards. |
| 807 | // Extend as long as we can. |
| 808 | start := s + repOff |
| 809 | // We end the search early, so we don't risk 0 literals |
| 810 | // and have to do special offset treatment. |
| 811 | startLimit := nextEmit + 1 |
| 812 | |
| 813 | tMin := s - e.maxMatchOff |
| 814 | if tMin < 0 { |
| 815 | tMin = 0 |
| 816 | } |
| 817 | for repIndex > tMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch-1 { |
| 818 | repIndex-- |
| 819 | start-- |
| 820 | seq.matchLen++ |
| 821 | } |
| 822 | addLiterals(&seq, start) |
| 823 | |
| 824 | // rep 0 |
| 825 | seq.offset = 1 |
| 826 | if debugSequences { |
| 827 | println("repeat sequence", seq, "next s:", s) |
| 828 | } |
| 829 | blk.sequences = append(blk.sequences, seq) |
| 830 | s += lenght + repOff |
| 831 | nextEmit = s |
| 832 | if s >= sLimit { |
| 833 | if debug { |
| 834 | println("repeat ended", s, lenght) |
| 835 | |
| 836 | } |
| 837 | break encodeLoop |
| 838 | } |
| 839 | cv = load6432(src, s) |
| 840 | continue |
| 841 | } |
| 842 | } |
| 843 | // Find the offsets of our two matches. |
| 844 | coffsetL := s - (candidateL.offset - e.cur) |
| 845 | coffsetS := s - (candidateS.offset - e.cur) |
| 846 | |
| 847 | // Check if we have a long match. |
| 848 | if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val { |
| 849 | // Found a long match, likely at least 8 bytes. |
| 850 | // Reference encoder checks all 8 bytes, we only check 4, |
| 851 | // but the likelihood of both the first 4 bytes and the hash matching should be enough. |
| 852 | t = candidateL.offset - e.cur |
| 853 | if debugAsserts && s <= t { |
| 854 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 855 | } |
| 856 | if debugAsserts && s-t > e.maxMatchOff { |
| 857 | panic("s - t >e.maxMatchOff") |
| 858 | } |
| 859 | if debugMatches { |
| 860 | println("long match") |
| 861 | } |
| 862 | break |
| 863 | } |
| 864 | |
| 865 | // Check if we have a short match. |
| 866 | if coffsetS < e.maxMatchOff && uint32(cv) == candidateS.val { |
| 867 | // found a regular match |
| 868 | // See if we can find a long match at s+1 |
| 869 | const checkAt = 1 |
| 870 | cv := load6432(src, s+checkAt) |
| 871 | nextHashL = hash8(cv, dFastLongTableBits) |
| 872 | candidateL = e.longTable[nextHashL] |
| 873 | coffsetL = s - (candidateL.offset - e.cur) + checkAt |
| 874 | |
| 875 | // We can store it, since we have at least a 4 byte match. |
| 876 | e.longTable[nextHashL] = tableEntry{offset: s + checkAt + e.cur, val: uint32(cv)} |
| 877 | e.markLongShardDirty(nextHashL) |
| 878 | if coffsetL < e.maxMatchOff && uint32(cv) == candidateL.val { |
| 879 | // Found a long match, likely at least 8 bytes. |
| 880 | // Reference encoder checks all 8 bytes, we only check 4, |
| 881 | // but the likelihood of both the first 4 bytes and the hash matching should be enough. |
| 882 | t = candidateL.offset - e.cur |
| 883 | s += checkAt |
| 884 | if debugMatches { |
| 885 | println("long match (after short)") |
| 886 | } |
| 887 | break |
| 888 | } |
| 889 | |
| 890 | t = candidateS.offset - e.cur |
| 891 | if debugAsserts && s <= t { |
| 892 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 893 | } |
| 894 | if debugAsserts && s-t > e.maxMatchOff { |
| 895 | panic("s - t >e.maxMatchOff") |
| 896 | } |
| 897 | if debugAsserts && t < 0 { |
| 898 | panic("t<0") |
| 899 | } |
| 900 | if debugMatches { |
| 901 | println("short match") |
| 902 | } |
| 903 | break |
| 904 | } |
| 905 | |
| 906 | // No match found, move forward in input. |
| 907 | s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) |
| 908 | if s >= sLimit { |
| 909 | break encodeLoop |
| 910 | } |
| 911 | cv = load6432(src, s) |
| 912 | } |
| 913 | |
| 914 | // A 4-byte match has been found. Update recent offsets. |
| 915 | // We'll later see if more than 4 bytes. |
| 916 | offset2 = offset1 |
| 917 | offset1 = s - t |
| 918 | |
| 919 | if debugAsserts && s <= t { |
| 920 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 921 | } |
| 922 | |
| 923 | if debugAsserts && canRepeat && int(offset1) > len(src) { |
| 924 | panic("invalid offset") |
| 925 | } |
| 926 | |
| 927 | // Extend the 4-byte match as long as possible. |
| 928 | l := e.matchlen(s+4, t+4, src) + 4 |
| 929 | |
| 930 | // Extend backwards |
| 931 | tMin := s - e.maxMatchOff |
| 932 | if tMin < 0 { |
| 933 | tMin = 0 |
| 934 | } |
| 935 | for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength { |
| 936 | s-- |
| 937 | t-- |
| 938 | l++ |
| 939 | } |
| 940 | |
| 941 | // Write our sequence |
| 942 | var seq seq |
| 943 | seq.litLen = uint32(s - nextEmit) |
| 944 | seq.matchLen = uint32(l - zstdMinMatch) |
| 945 | if seq.litLen > 0 { |
| 946 | blk.literals = append(blk.literals, src[nextEmit:s]...) |
| 947 | } |
| 948 | seq.offset = uint32(s-t) + 3 |
| 949 | s += l |
| 950 | if debugSequences { |
| 951 | println("sequence", seq, "next s:", s) |
| 952 | } |
| 953 | blk.sequences = append(blk.sequences, seq) |
| 954 | nextEmit = s |
| 955 | if s >= sLimit { |
| 956 | break encodeLoop |
| 957 | } |
| 958 | |
| 959 | // Index match start+1 (long) and start+2 (short) |
| 960 | index0 := s - l + 1 |
| 961 | // Index match end-2 (long) and end-1 (short) |
| 962 | index1 := s - 2 |
| 963 | |
| 964 | cv0 := load6432(src, index0) |
| 965 | cv1 := load6432(src, index1) |
| 966 | te0 := tableEntry{offset: index0 + e.cur, val: uint32(cv0)} |
| 967 | te1 := tableEntry{offset: index1 + e.cur, val: uint32(cv1)} |
| 968 | longHash1 := hash8(cv0, dFastLongTableBits) |
| 969 | longHash2 := hash8(cv0, dFastLongTableBits) |
| 970 | e.longTable[longHash1] = te0 |
| 971 | e.longTable[longHash2] = te1 |
| 972 | e.markLongShardDirty(longHash1) |
| 973 | e.markLongShardDirty(longHash2) |
| 974 | cv0 >>= 8 |
| 975 | cv1 >>= 8 |
| 976 | te0.offset++ |
| 977 | te1.offset++ |
| 978 | te0.val = uint32(cv0) |
| 979 | te1.val = uint32(cv1) |
| 980 | hashVal1 := hash5(cv0, dFastShortTableBits) |
| 981 | hashVal2 := hash5(cv1, dFastShortTableBits) |
| 982 | e.table[hashVal1] = te0 |
| 983 | e.markShardDirty(hashVal1) |
| 984 | e.table[hashVal2] = te1 |
| 985 | e.markShardDirty(hashVal2) |
| 986 | |
| 987 | cv = load6432(src, s) |
| 988 | |
| 989 | if !canRepeat { |
| 990 | continue |
| 991 | } |
| 992 | |
| 993 | // Check offset 2 |
| 994 | for { |
| 995 | o2 := s - offset2 |
| 996 | if load3232(src, o2) != uint32(cv) { |
| 997 | // Do regular search |
| 998 | break |
| 999 | } |
| 1000 | |
| 1001 | // Store this, since we have it. |
| 1002 | nextHashS := hash5(cv, dFastShortTableBits) |
| 1003 | nextHashL := hash8(cv, dFastLongTableBits) |
| 1004 | |
| 1005 | // We have at least 4 byte match. |
| 1006 | // No need to check backwards. We come straight from a match |
| 1007 | l := 4 + e.matchlen(s+4, o2+4, src) |
| 1008 | |
| 1009 | entry := tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 1010 | e.longTable[nextHashL] = entry |
| 1011 | e.markLongShardDirty(nextHashL) |
| 1012 | e.table[nextHashS] = entry |
| 1013 | e.markShardDirty(nextHashS) |
| 1014 | seq.matchLen = uint32(l) - zstdMinMatch |
| 1015 | seq.litLen = 0 |
| 1016 | |
| 1017 | // Since litlen is always 0, this is offset 1. |
| 1018 | seq.offset = 1 |
| 1019 | s += l |
| 1020 | nextEmit = s |
| 1021 | if debugSequences { |
| 1022 | println("sequence", seq, "next s:", s) |
| 1023 | } |
| 1024 | blk.sequences = append(blk.sequences, seq) |
| 1025 | |
| 1026 | // Swap offset 1 and 2. |
| 1027 | offset1, offset2 = offset2, offset1 |
| 1028 | if s >= sLimit { |
| 1029 | // Finished |
| 1030 | break encodeLoop |
| 1031 | } |
| 1032 | cv = load6432(src, s) |
| 1033 | } |
| 1034 | } |
| 1035 | |
| 1036 | if int(nextEmit) < len(src) { |
| 1037 | blk.literals = append(blk.literals, src[nextEmit:]...) |
| 1038 | blk.extraLits = len(src) - int(nextEmit) |
| 1039 | } |
| 1040 | blk.recentOffsets[0] = uint32(offset1) |
| 1041 | blk.recentOffsets[1] = uint32(offset2) |
| 1042 | if debug { |
| 1043 | println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) |
| 1044 | } |
| 1045 | // If we encoded more than 64K mark all dirty. |
| 1046 | if len(src) > 64<<10 { |
| 1047 | e.markAllShardsDirty() |
| 1048 | } |
| 1049 | } |
| 1050 | |
| 1051 | // ResetDict will reset and set a dictionary if not nil |
| 1052 | func (e *doubleFastEncoder) Reset(d *dict, singleBlock bool) { |
| 1053 | e.fastEncoder.Reset(d, singleBlock) |
| 1054 | if d != nil { |
| 1055 | panic("doubleFastEncoder: Reset with dict not supported") |
| 1056 | } |
| 1057 | } |
| 1058 | |
| 1059 | // ResetDict will reset and set a dictionary if not nil |
| 1060 | func (e *doubleFastEncoderDict) Reset(d *dict, singleBlock bool) { |
| 1061 | allDirty := e.allDirty |
| 1062 | e.fastEncoderDict.Reset(d, singleBlock) |
| 1063 | if d == nil { |
| 1064 | return |
| 1065 | } |
| 1066 | |
| 1067 | // Init or copy dict table |
| 1068 | if len(e.dictLongTable) != len(e.longTable) || d.id != e.lastDictID { |
| 1069 | if len(e.dictLongTable) != len(e.longTable) { |
| 1070 | e.dictLongTable = make([]tableEntry, len(e.longTable)) |
| 1071 | } |
| 1072 | if len(d.content) >= 8 { |
| 1073 | cv := load6432(d.content, 0) |
| 1074 | e.dictLongTable[hash8(cv, dFastLongTableBits)] = tableEntry{ |
| 1075 | val: uint32(cv), |
| 1076 | offset: e.maxMatchOff, |
| 1077 | } |
| 1078 | end := int32(len(d.content)) - 8 + e.maxMatchOff |
| 1079 | for i := e.maxMatchOff + 1; i < end; i++ { |
| 1080 | cv = cv>>8 | (uint64(d.content[i-e.maxMatchOff+7]) << 56) |
| 1081 | e.dictLongTable[hash8(cv, dFastLongTableBits)] = tableEntry{ |
| 1082 | val: uint32(cv), |
| 1083 | offset: i, |
| 1084 | } |
| 1085 | } |
| 1086 | } |
| 1087 | e.lastDictID = d.id |
| 1088 | e.allDirty = true |
| 1089 | } |
| 1090 | // Reset table to initial state |
| 1091 | e.cur = e.maxMatchOff |
| 1092 | |
| 1093 | dirtyShardCnt := 0 |
| 1094 | if !allDirty { |
| 1095 | for i := range e.longTableShardDirty { |
| 1096 | if e.longTableShardDirty[i] { |
| 1097 | dirtyShardCnt++ |
| 1098 | } |
| 1099 | } |
| 1100 | } |
| 1101 | |
| 1102 | if allDirty || dirtyShardCnt > dLongTableShardCnt/2 { |
| 1103 | copy(e.longTable[:], e.dictLongTable) |
| 1104 | for i := range e.longTableShardDirty { |
| 1105 | e.longTableShardDirty[i] = false |
| 1106 | } |
| 1107 | return |
| 1108 | } |
| 1109 | for i := range e.longTableShardDirty { |
| 1110 | if !e.longTableShardDirty[i] { |
| 1111 | continue |
| 1112 | } |
| 1113 | |
| 1114 | copy(e.longTable[i*dLongTableShardSize:(i+1)*dLongTableShardSize], e.dictLongTable[i*dLongTableShardSize:(i+1)*dLongTableShardSize]) |
| 1115 | e.longTableShardDirty[i] = false |
| 1116 | } |
| 1117 | } |
| 1118 | |
| 1119 | func (e *doubleFastEncoderDict) markLongShardDirty(entryNum uint32) { |
| 1120 | e.longTableShardDirty[entryNum/dLongTableShardSize] = true |
| 1121 | } |