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 ( |
| 8 | "fmt" |
| 9 | "math" |
| 10 | "math/bits" |
| 11 | ) |
| 12 | |
| 13 | const ( |
| 14 | tableBits = 15 // Bits used in the table |
| 15 | tableSize = 1 << tableBits // Size of the table |
| 16 | tableShardCnt = 1 << (tableBits - dictShardBits) // Number of shards in the table |
| 17 | tableShardSize = tableSize / tableShardCnt // Size of an individual shard |
| 18 | tableMask = tableSize - 1 // Mask for table indices. Redundant, but can eliminate bounds checks. |
| 19 | maxMatchLength = 131074 |
| 20 | ) |
| 21 | |
| 22 | type tableEntry struct { |
| 23 | val uint32 |
| 24 | offset int32 |
| 25 | } |
| 26 | |
| 27 | type fastEncoder struct { |
| 28 | fastBase |
| 29 | table [tableSize]tableEntry |
| 30 | } |
| 31 | |
| 32 | type fastEncoderDict struct { |
| 33 | fastEncoder |
| 34 | dictTable []tableEntry |
| 35 | tableShardDirty [tableShardCnt]bool |
| 36 | allDirty bool |
| 37 | } |
| 38 | |
| 39 | // Encode mimmics functionality in zstd_fast.c |
| 40 | func (e *fastEncoder) Encode(blk *blockEnc, src []byte) { |
| 41 | const ( |
| 42 | inputMargin = 8 |
| 43 | minNonLiteralBlockSize = 1 + 1 + inputMargin |
| 44 | ) |
| 45 | |
| 46 | // Protect against e.cur wraparound. |
| 47 | for e.cur >= bufferReset { |
| 48 | if len(e.hist) == 0 { |
| 49 | for i := range e.table[:] { |
| 50 | e.table[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 | e.cur = e.maxMatchOff |
| 67 | break |
| 68 | } |
| 69 | |
| 70 | s := e.addBlock(src) |
| 71 | blk.size = len(src) |
| 72 | if len(src) < minNonLiteralBlockSize { |
| 73 | blk.extraLits = len(src) |
| 74 | blk.literals = blk.literals[:len(src)] |
| 75 | copy(blk.literals, src) |
| 76 | return |
| 77 | } |
| 78 | |
| 79 | // Override src |
| 80 | src = e.hist |
| 81 | sLimit := int32(len(src)) - inputMargin |
| 82 | // stepSize is the number of bytes to skip on every main loop iteration. |
| 83 | // It should be >= 2. |
| 84 | const stepSize = 2 |
| 85 | |
| 86 | // TEMPLATE |
| 87 | const hashLog = tableBits |
| 88 | // seems global, but would be nice to tweak. |
| 89 | const kSearchStrength = 7 |
| 90 | |
| 91 | // nextEmit is where in src the next emitLiteral should start from. |
| 92 | nextEmit := s |
| 93 | cv := load6432(src, s) |
| 94 | |
| 95 | // Relative offsets |
| 96 | offset1 := int32(blk.recentOffsets[0]) |
| 97 | offset2 := int32(blk.recentOffsets[1]) |
| 98 | |
| 99 | addLiterals := func(s *seq, until int32) { |
| 100 | if until == nextEmit { |
| 101 | return |
| 102 | } |
| 103 | blk.literals = append(blk.literals, src[nextEmit:until]...) |
| 104 | s.litLen = uint32(until - nextEmit) |
| 105 | } |
| 106 | if debug { |
| 107 | println("recent offsets:", blk.recentOffsets) |
| 108 | } |
| 109 | |
| 110 | encodeLoop: |
| 111 | for { |
| 112 | // t will contain the match offset when we find one. |
| 113 | // When existing the search loop, we have already checked 4 bytes. |
| 114 | var t int32 |
| 115 | |
| 116 | // We will not use repeat offsets across blocks. |
| 117 | // By not using them for the first 3 matches |
| 118 | canRepeat := len(blk.sequences) > 2 |
| 119 | |
| 120 | for { |
| 121 | if debugAsserts && canRepeat && offset1 == 0 { |
| 122 | panic("offset0 was 0") |
| 123 | } |
| 124 | |
| 125 | nextHash := hash6(cv, hashLog) |
| 126 | nextHash2 := hash6(cv>>8, hashLog) |
| 127 | candidate := e.table[nextHash] |
| 128 | candidate2 := e.table[nextHash2] |
| 129 | repIndex := s - offset1 + 2 |
| 130 | |
| 131 | e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 132 | e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)} |
| 133 | |
| 134 | if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) { |
| 135 | // Consider history as well. |
| 136 | var seq seq |
| 137 | var length int32 |
| 138 | // length = 4 + e.matchlen(s+6, repIndex+4, src) |
| 139 | { |
| 140 | a := src[s+6:] |
| 141 | b := src[repIndex+4:] |
| 142 | endI := len(a) & (math.MaxInt32 - 7) |
| 143 | length = int32(endI) + 4 |
| 144 | for i := 0; i < endI; i += 8 { |
| 145 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 146 | length = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 147 | break |
| 148 | } |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | seq.matchLen = uint32(length - zstdMinMatch) |
| 153 | |
| 154 | // We might be able to match backwards. |
| 155 | // Extend as long as we can. |
| 156 | start := s + 2 |
| 157 | // We end the search early, so we don't risk 0 literals |
| 158 | // and have to do special offset treatment. |
| 159 | startLimit := nextEmit + 1 |
| 160 | |
| 161 | sMin := s - e.maxMatchOff |
| 162 | if sMin < 0 { |
| 163 | sMin = 0 |
| 164 | } |
| 165 | for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch { |
| 166 | repIndex-- |
| 167 | start-- |
| 168 | seq.matchLen++ |
| 169 | } |
| 170 | addLiterals(&seq, start) |
| 171 | |
| 172 | // rep 0 |
| 173 | seq.offset = 1 |
| 174 | if debugSequences { |
| 175 | println("repeat sequence", seq, "next s:", s) |
| 176 | } |
| 177 | blk.sequences = append(blk.sequences, seq) |
| 178 | s += length + 2 |
| 179 | nextEmit = s |
| 180 | if s >= sLimit { |
| 181 | if debug { |
| 182 | println("repeat ended", s, length) |
| 183 | |
| 184 | } |
| 185 | break encodeLoop |
| 186 | } |
| 187 | cv = load6432(src, s) |
| 188 | continue |
| 189 | } |
| 190 | coffset0 := s - (candidate.offset - e.cur) |
| 191 | coffset1 := s - (candidate2.offset - e.cur) + 1 |
| 192 | if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val { |
| 193 | // found a regular match |
| 194 | t = candidate.offset - e.cur |
| 195 | if debugAsserts && s <= t { |
| 196 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 197 | } |
| 198 | if debugAsserts && s-t > e.maxMatchOff { |
| 199 | panic("s - t >e.maxMatchOff") |
| 200 | } |
| 201 | break |
| 202 | } |
| 203 | |
| 204 | if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val { |
| 205 | // found a regular match |
| 206 | t = candidate2.offset - e.cur |
| 207 | s++ |
| 208 | if debugAsserts && s <= t { |
| 209 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 210 | } |
| 211 | if debugAsserts && s-t > e.maxMatchOff { |
| 212 | panic("s - t >e.maxMatchOff") |
| 213 | } |
| 214 | if debugAsserts && t < 0 { |
| 215 | panic("t<0") |
| 216 | } |
| 217 | break |
| 218 | } |
| 219 | s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) |
| 220 | if s >= sLimit { |
| 221 | break encodeLoop |
| 222 | } |
| 223 | cv = load6432(src, s) |
| 224 | } |
| 225 | // A 4-byte match has been found. We'll later see if more than 4 bytes. |
| 226 | offset2 = offset1 |
| 227 | offset1 = s - t |
| 228 | |
| 229 | if debugAsserts && s <= t { |
| 230 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 231 | } |
| 232 | |
| 233 | if debugAsserts && canRepeat && int(offset1) > len(src) { |
| 234 | panic("invalid offset") |
| 235 | } |
| 236 | |
| 237 | // Extend the 4-byte match as long as possible. |
| 238 | //l := e.matchlen(s+4, t+4, src) + 4 |
| 239 | var l int32 |
| 240 | { |
| 241 | a := src[s+4:] |
| 242 | b := src[t+4:] |
| 243 | endI := len(a) & (math.MaxInt32 - 7) |
| 244 | l = int32(endI) + 4 |
| 245 | for i := 0; i < endI; i += 8 { |
| 246 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 247 | l = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 248 | break |
| 249 | } |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | // Extend backwards |
| 254 | tMin := s - e.maxMatchOff |
| 255 | if tMin < 0 { |
| 256 | tMin = 0 |
| 257 | } |
| 258 | for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength { |
| 259 | s-- |
| 260 | t-- |
| 261 | l++ |
| 262 | } |
| 263 | |
| 264 | // Write our sequence. |
| 265 | var seq seq |
| 266 | seq.litLen = uint32(s - nextEmit) |
| 267 | seq.matchLen = uint32(l - zstdMinMatch) |
| 268 | if seq.litLen > 0 { |
| 269 | blk.literals = append(blk.literals, src[nextEmit:s]...) |
| 270 | } |
| 271 | // Don't use repeat offsets |
| 272 | seq.offset = uint32(s-t) + 3 |
| 273 | s += l |
| 274 | if debugSequences { |
| 275 | println("sequence", seq, "next s:", s) |
| 276 | } |
| 277 | blk.sequences = append(blk.sequences, seq) |
| 278 | nextEmit = s |
| 279 | if s >= sLimit { |
| 280 | break encodeLoop |
| 281 | } |
| 282 | cv = load6432(src, s) |
| 283 | |
| 284 | // Check offset 2 |
| 285 | if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) { |
| 286 | // We have at least 4 byte match. |
| 287 | // No need to check backwards. We come straight from a match |
| 288 | //l := 4 + e.matchlen(s+4, o2+4, src) |
| 289 | var l int32 |
| 290 | { |
| 291 | a := src[s+4:] |
| 292 | b := src[o2+4:] |
| 293 | endI := len(a) & (math.MaxInt32 - 7) |
| 294 | l = int32(endI) + 4 |
| 295 | for i := 0; i < endI; i += 8 { |
| 296 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 297 | l = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 298 | break |
| 299 | } |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | // Store this, since we have it. |
| 304 | nextHash := hash6(cv, hashLog) |
| 305 | e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 306 | seq.matchLen = uint32(l) - zstdMinMatch |
| 307 | seq.litLen = 0 |
| 308 | // Since litlen is always 0, this is offset 1. |
| 309 | seq.offset = 1 |
| 310 | s += l |
| 311 | nextEmit = s |
| 312 | if debugSequences { |
| 313 | println("sequence", seq, "next s:", s) |
| 314 | } |
| 315 | blk.sequences = append(blk.sequences, seq) |
| 316 | |
| 317 | // Swap offset 1 and 2. |
| 318 | offset1, offset2 = offset2, offset1 |
| 319 | if s >= sLimit { |
| 320 | break encodeLoop |
| 321 | } |
| 322 | // Prepare next loop. |
| 323 | cv = load6432(src, s) |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | if int(nextEmit) < len(src) { |
| 328 | blk.literals = append(blk.literals, src[nextEmit:]...) |
| 329 | blk.extraLits = len(src) - int(nextEmit) |
| 330 | } |
| 331 | blk.recentOffsets[0] = uint32(offset1) |
| 332 | blk.recentOffsets[1] = uint32(offset2) |
| 333 | if debug { |
| 334 | println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) |
| 335 | } |
| 336 | } |
| 337 | |
| 338 | // EncodeNoHist will encode a block with no history and no following blocks. |
| 339 | // Most notable difference is that src will not be copied for history and |
| 340 | // we do not need to check for max match length. |
| 341 | func (e *fastEncoder) EncodeNoHist(blk *blockEnc, src []byte) { |
| 342 | const ( |
| 343 | inputMargin = 8 |
| 344 | minNonLiteralBlockSize = 1 + 1 + inputMargin |
| 345 | ) |
| 346 | if debug { |
| 347 | if len(src) > maxBlockSize { |
| 348 | panic("src too big") |
| 349 | } |
| 350 | } |
| 351 | |
| 352 | // Protect against e.cur wraparound. |
| 353 | if e.cur >= bufferReset { |
| 354 | for i := range e.table[:] { |
| 355 | e.table[i] = tableEntry{} |
| 356 | } |
| 357 | e.cur = e.maxMatchOff |
| 358 | } |
| 359 | |
| 360 | s := int32(0) |
| 361 | blk.size = len(src) |
| 362 | if len(src) < minNonLiteralBlockSize { |
| 363 | blk.extraLits = len(src) |
| 364 | blk.literals = blk.literals[:len(src)] |
| 365 | copy(blk.literals, src) |
| 366 | return |
| 367 | } |
| 368 | |
| 369 | sLimit := int32(len(src)) - inputMargin |
| 370 | // stepSize is the number of bytes to skip on every main loop iteration. |
| 371 | // It should be >= 2. |
| 372 | const stepSize = 2 |
| 373 | |
| 374 | // TEMPLATE |
| 375 | const hashLog = tableBits |
| 376 | // seems global, but would be nice to tweak. |
| 377 | const kSearchStrength = 8 |
| 378 | |
| 379 | // nextEmit is where in src the next emitLiteral should start from. |
| 380 | nextEmit := s |
| 381 | cv := load6432(src, s) |
| 382 | |
| 383 | // Relative offsets |
| 384 | offset1 := int32(blk.recentOffsets[0]) |
| 385 | offset2 := int32(blk.recentOffsets[1]) |
| 386 | |
| 387 | addLiterals := func(s *seq, until int32) { |
| 388 | if until == nextEmit { |
| 389 | return |
| 390 | } |
| 391 | blk.literals = append(blk.literals, src[nextEmit:until]...) |
| 392 | s.litLen = uint32(until - nextEmit) |
| 393 | } |
| 394 | if debug { |
| 395 | println("recent offsets:", blk.recentOffsets) |
| 396 | } |
| 397 | |
| 398 | encodeLoop: |
| 399 | for { |
| 400 | // t will contain the match offset when we find one. |
| 401 | // When existing the search loop, we have already checked 4 bytes. |
| 402 | var t int32 |
| 403 | |
| 404 | // We will not use repeat offsets across blocks. |
| 405 | // By not using them for the first 3 matches |
| 406 | |
| 407 | for { |
| 408 | nextHash := hash6(cv, hashLog) |
| 409 | nextHash2 := hash6(cv>>8, hashLog) |
| 410 | candidate := e.table[nextHash] |
| 411 | candidate2 := e.table[nextHash2] |
| 412 | repIndex := s - offset1 + 2 |
| 413 | |
| 414 | e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 415 | e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)} |
| 416 | |
| 417 | if len(blk.sequences) > 2 && load3232(src, repIndex) == uint32(cv>>16) { |
| 418 | // Consider history as well. |
| 419 | var seq seq |
| 420 | // length := 4 + e.matchlen(s+6, repIndex+4, src) |
| 421 | // length := 4 + int32(matchLen(src[s+6:], src[repIndex+4:])) |
| 422 | var length int32 |
| 423 | { |
| 424 | a := src[s+6:] |
| 425 | b := src[repIndex+4:] |
| 426 | endI := len(a) & (math.MaxInt32 - 7) |
| 427 | length = int32(endI) + 4 |
| 428 | for i := 0; i < endI; i += 8 { |
| 429 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 430 | length = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 431 | break |
| 432 | } |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | seq.matchLen = uint32(length - zstdMinMatch) |
| 437 | |
| 438 | // We might be able to match backwards. |
| 439 | // Extend as long as we can. |
| 440 | start := s + 2 |
| 441 | // We end the search early, so we don't risk 0 literals |
| 442 | // and have to do special offset treatment. |
| 443 | startLimit := nextEmit + 1 |
| 444 | |
| 445 | sMin := s - e.maxMatchOff |
| 446 | if sMin < 0 { |
| 447 | sMin = 0 |
| 448 | } |
| 449 | for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] { |
| 450 | repIndex-- |
| 451 | start-- |
| 452 | seq.matchLen++ |
| 453 | } |
| 454 | addLiterals(&seq, start) |
| 455 | |
| 456 | // rep 0 |
| 457 | seq.offset = 1 |
| 458 | if debugSequences { |
| 459 | println("repeat sequence", seq, "next s:", s) |
| 460 | } |
| 461 | blk.sequences = append(blk.sequences, seq) |
| 462 | s += length + 2 |
| 463 | nextEmit = s |
| 464 | if s >= sLimit { |
| 465 | if debug { |
| 466 | println("repeat ended", s, length) |
| 467 | |
| 468 | } |
| 469 | break encodeLoop |
| 470 | } |
| 471 | cv = load6432(src, s) |
| 472 | continue |
| 473 | } |
| 474 | coffset0 := s - (candidate.offset - e.cur) |
| 475 | coffset1 := s - (candidate2.offset - e.cur) + 1 |
| 476 | if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val { |
| 477 | // found a regular match |
| 478 | t = candidate.offset - e.cur |
| 479 | if debugAsserts && s <= t { |
| 480 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 481 | } |
| 482 | if debugAsserts && s-t > e.maxMatchOff { |
| 483 | panic("s - t >e.maxMatchOff") |
| 484 | } |
| 485 | if debugAsserts && t < 0 { |
| 486 | panic(fmt.Sprintf("t (%d) < 0, candidate.offset: %d, e.cur: %d, coffset0: %d, e.maxMatchOff: %d", t, candidate.offset, e.cur, coffset0, e.maxMatchOff)) |
| 487 | } |
| 488 | break |
| 489 | } |
| 490 | |
| 491 | if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val { |
| 492 | // found a regular match |
| 493 | t = candidate2.offset - e.cur |
| 494 | s++ |
| 495 | if debugAsserts && s <= t { |
| 496 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 497 | } |
| 498 | if debugAsserts && s-t > e.maxMatchOff { |
| 499 | panic("s - t >e.maxMatchOff") |
| 500 | } |
| 501 | if debugAsserts && t < 0 { |
| 502 | panic("t<0") |
| 503 | } |
| 504 | break |
| 505 | } |
| 506 | s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) |
| 507 | if s >= sLimit { |
| 508 | break encodeLoop |
| 509 | } |
| 510 | cv = load6432(src, s) |
| 511 | } |
| 512 | // A 4-byte match has been found. We'll later see if more than 4 bytes. |
| 513 | offset2 = offset1 |
| 514 | offset1 = s - t |
| 515 | |
| 516 | if debugAsserts && s <= t { |
| 517 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 518 | } |
| 519 | |
| 520 | if debugAsserts && t < 0 { |
| 521 | panic(fmt.Sprintf("t (%d) < 0 ", t)) |
| 522 | } |
| 523 | // Extend the 4-byte match as long as possible. |
| 524 | //l := e.matchlenNoHist(s+4, t+4, src) + 4 |
| 525 | // l := int32(matchLen(src[s+4:], src[t+4:])) + 4 |
| 526 | var l int32 |
| 527 | { |
| 528 | a := src[s+4:] |
| 529 | b := src[t+4:] |
| 530 | endI := len(a) & (math.MaxInt32 - 7) |
| 531 | l = int32(endI) + 4 |
| 532 | for i := 0; i < endI; i += 8 { |
| 533 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 534 | l = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 535 | break |
| 536 | } |
| 537 | } |
| 538 | } |
| 539 | |
| 540 | // Extend backwards |
| 541 | tMin := s - e.maxMatchOff |
| 542 | if tMin < 0 { |
| 543 | tMin = 0 |
| 544 | } |
| 545 | for t > tMin && s > nextEmit && src[t-1] == src[s-1] { |
| 546 | s-- |
| 547 | t-- |
| 548 | l++ |
| 549 | } |
| 550 | |
| 551 | // Write our sequence. |
| 552 | var seq seq |
| 553 | seq.litLen = uint32(s - nextEmit) |
| 554 | seq.matchLen = uint32(l - zstdMinMatch) |
| 555 | if seq.litLen > 0 { |
| 556 | blk.literals = append(blk.literals, src[nextEmit:s]...) |
| 557 | } |
| 558 | // Don't use repeat offsets |
| 559 | seq.offset = uint32(s-t) + 3 |
| 560 | s += l |
| 561 | if debugSequences { |
| 562 | println("sequence", seq, "next s:", s) |
| 563 | } |
| 564 | blk.sequences = append(blk.sequences, seq) |
| 565 | nextEmit = s |
| 566 | if s >= sLimit { |
| 567 | break encodeLoop |
| 568 | } |
| 569 | cv = load6432(src, s) |
| 570 | |
| 571 | // Check offset 2 |
| 572 | if o2 := s - offset2; len(blk.sequences) > 2 && load3232(src, o2) == uint32(cv) { |
| 573 | // We have at least 4 byte match. |
| 574 | // No need to check backwards. We come straight from a match |
| 575 | //l := 4 + e.matchlenNoHist(s+4, o2+4, src) |
| 576 | // l := 4 + int32(matchLen(src[s+4:], src[o2+4:])) |
| 577 | var l int32 |
| 578 | { |
| 579 | a := src[s+4:] |
| 580 | b := src[o2+4:] |
| 581 | endI := len(a) & (math.MaxInt32 - 7) |
| 582 | l = int32(endI) + 4 |
| 583 | for i := 0; i < endI; i += 8 { |
| 584 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 585 | l = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 586 | break |
| 587 | } |
| 588 | } |
| 589 | } |
| 590 | |
| 591 | // Store this, since we have it. |
| 592 | nextHash := hash6(cv, hashLog) |
| 593 | e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 594 | seq.matchLen = uint32(l) - zstdMinMatch |
| 595 | seq.litLen = 0 |
| 596 | // Since litlen is always 0, this is offset 1. |
| 597 | seq.offset = 1 |
| 598 | s += l |
| 599 | nextEmit = s |
| 600 | if debugSequences { |
| 601 | println("sequence", seq, "next s:", s) |
| 602 | } |
| 603 | blk.sequences = append(blk.sequences, seq) |
| 604 | |
| 605 | // Swap offset 1 and 2. |
| 606 | offset1, offset2 = offset2, offset1 |
| 607 | if s >= sLimit { |
| 608 | break encodeLoop |
| 609 | } |
| 610 | // Prepare next loop. |
| 611 | cv = load6432(src, s) |
| 612 | } |
| 613 | } |
| 614 | |
| 615 | if int(nextEmit) < len(src) { |
| 616 | blk.literals = append(blk.literals, src[nextEmit:]...) |
| 617 | blk.extraLits = len(src) - int(nextEmit) |
| 618 | } |
| 619 | if debug { |
| 620 | println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) |
| 621 | } |
| 622 | // We do not store history, so we must offset e.cur to avoid false matches for next user. |
| 623 | if e.cur < bufferReset { |
| 624 | e.cur += int32(len(src)) |
| 625 | } |
| 626 | } |
| 627 | |
| 628 | // Encode will encode the content, with a dictionary if initialized for it. |
| 629 | func (e *fastEncoderDict) Encode(blk *blockEnc, src []byte) { |
| 630 | const ( |
| 631 | inputMargin = 8 |
| 632 | minNonLiteralBlockSize = 1 + 1 + inputMargin |
| 633 | ) |
| 634 | if e.allDirty || len(src) > 32<<10 { |
| 635 | e.fastEncoder.Encode(blk, src) |
| 636 | e.allDirty = true |
| 637 | return |
| 638 | } |
| 639 | // Protect against e.cur wraparound. |
| 640 | for e.cur >= bufferReset { |
| 641 | if len(e.hist) == 0 { |
| 642 | for i := range e.table[:] { |
| 643 | e.table[i] = tableEntry{} |
| 644 | } |
| 645 | e.cur = e.maxMatchOff |
| 646 | break |
| 647 | } |
| 648 | // Shift down everything in the table that isn't already too far away. |
| 649 | minOff := e.cur + int32(len(e.hist)) - e.maxMatchOff |
| 650 | for i := range e.table[:] { |
| 651 | v := e.table[i].offset |
| 652 | if v < minOff { |
| 653 | v = 0 |
| 654 | } else { |
| 655 | v = v - e.cur + e.maxMatchOff |
| 656 | } |
| 657 | e.table[i].offset = v |
| 658 | } |
| 659 | e.cur = e.maxMatchOff |
| 660 | break |
| 661 | } |
| 662 | |
| 663 | s := e.addBlock(src) |
| 664 | blk.size = len(src) |
| 665 | if len(src) < minNonLiteralBlockSize { |
| 666 | blk.extraLits = len(src) |
| 667 | blk.literals = blk.literals[:len(src)] |
| 668 | copy(blk.literals, src) |
| 669 | return |
| 670 | } |
| 671 | |
| 672 | // Override src |
| 673 | src = e.hist |
| 674 | sLimit := int32(len(src)) - inputMargin |
| 675 | // stepSize is the number of bytes to skip on every main loop iteration. |
| 676 | // It should be >= 2. |
| 677 | const stepSize = 2 |
| 678 | |
| 679 | // TEMPLATE |
| 680 | const hashLog = tableBits |
| 681 | // seems global, but would be nice to tweak. |
| 682 | const kSearchStrength = 7 |
| 683 | |
| 684 | // nextEmit is where in src the next emitLiteral should start from. |
| 685 | nextEmit := s |
| 686 | cv := load6432(src, s) |
| 687 | |
| 688 | // Relative offsets |
| 689 | offset1 := int32(blk.recentOffsets[0]) |
| 690 | offset2 := int32(blk.recentOffsets[1]) |
| 691 | |
| 692 | addLiterals := func(s *seq, until int32) { |
| 693 | if until == nextEmit { |
| 694 | return |
| 695 | } |
| 696 | blk.literals = append(blk.literals, src[nextEmit:until]...) |
| 697 | s.litLen = uint32(until - nextEmit) |
| 698 | } |
| 699 | if debug { |
| 700 | println("recent offsets:", blk.recentOffsets) |
| 701 | } |
| 702 | |
| 703 | encodeLoop: |
| 704 | for { |
| 705 | // t will contain the match offset when we find one. |
| 706 | // When existing the search loop, we have already checked 4 bytes. |
| 707 | var t int32 |
| 708 | |
| 709 | // We will not use repeat offsets across blocks. |
| 710 | // By not using them for the first 3 matches |
| 711 | canRepeat := len(blk.sequences) > 2 |
| 712 | |
| 713 | for { |
| 714 | if debugAsserts && canRepeat && offset1 == 0 { |
| 715 | panic("offset0 was 0") |
| 716 | } |
| 717 | |
| 718 | nextHash := hash6(cv, hashLog) |
| 719 | nextHash2 := hash6(cv>>8, hashLog) |
| 720 | candidate := e.table[nextHash] |
| 721 | candidate2 := e.table[nextHash2] |
| 722 | repIndex := s - offset1 + 2 |
| 723 | |
| 724 | e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 725 | e.markShardDirty(nextHash) |
| 726 | e.table[nextHash2] = tableEntry{offset: s + e.cur + 1, val: uint32(cv >> 8)} |
| 727 | e.markShardDirty(nextHash2) |
| 728 | |
| 729 | if canRepeat && repIndex >= 0 && load3232(src, repIndex) == uint32(cv>>16) { |
| 730 | // Consider history as well. |
| 731 | var seq seq |
| 732 | var length int32 |
| 733 | // length = 4 + e.matchlen(s+6, repIndex+4, src) |
| 734 | { |
| 735 | a := src[s+6:] |
| 736 | b := src[repIndex+4:] |
| 737 | endI := len(a) & (math.MaxInt32 - 7) |
| 738 | length = int32(endI) + 4 |
| 739 | for i := 0; i < endI; i += 8 { |
| 740 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 741 | length = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 742 | break |
| 743 | } |
| 744 | } |
| 745 | } |
| 746 | |
| 747 | seq.matchLen = uint32(length - zstdMinMatch) |
| 748 | |
| 749 | // We might be able to match backwards. |
| 750 | // Extend as long as we can. |
| 751 | start := s + 2 |
| 752 | // We end the search early, so we don't risk 0 literals |
| 753 | // and have to do special offset treatment. |
| 754 | startLimit := nextEmit + 1 |
| 755 | |
| 756 | sMin := s - e.maxMatchOff |
| 757 | if sMin < 0 { |
| 758 | sMin = 0 |
| 759 | } |
| 760 | for repIndex > sMin && start > startLimit && src[repIndex-1] == src[start-1] && seq.matchLen < maxMatchLength-zstdMinMatch { |
| 761 | repIndex-- |
| 762 | start-- |
| 763 | seq.matchLen++ |
| 764 | } |
| 765 | addLiterals(&seq, start) |
| 766 | |
| 767 | // rep 0 |
| 768 | seq.offset = 1 |
| 769 | if debugSequences { |
| 770 | println("repeat sequence", seq, "next s:", s) |
| 771 | } |
| 772 | blk.sequences = append(blk.sequences, seq) |
| 773 | s += length + 2 |
| 774 | nextEmit = s |
| 775 | if s >= sLimit { |
| 776 | if debug { |
| 777 | println("repeat ended", s, length) |
| 778 | |
| 779 | } |
| 780 | break encodeLoop |
| 781 | } |
| 782 | cv = load6432(src, s) |
| 783 | continue |
| 784 | } |
| 785 | coffset0 := s - (candidate.offset - e.cur) |
| 786 | coffset1 := s - (candidate2.offset - e.cur) + 1 |
| 787 | if coffset0 < e.maxMatchOff && uint32(cv) == candidate.val { |
| 788 | // found a regular match |
| 789 | t = candidate.offset - e.cur |
| 790 | if debugAsserts && s <= t { |
| 791 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 792 | } |
| 793 | if debugAsserts && s-t > e.maxMatchOff { |
| 794 | panic("s - t >e.maxMatchOff") |
| 795 | } |
| 796 | break |
| 797 | } |
| 798 | |
| 799 | if coffset1 < e.maxMatchOff && uint32(cv>>8) == candidate2.val { |
| 800 | // found a regular match |
| 801 | t = candidate2.offset - e.cur |
| 802 | s++ |
| 803 | if debugAsserts && s <= t { |
| 804 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 805 | } |
| 806 | if debugAsserts && s-t > e.maxMatchOff { |
| 807 | panic("s - t >e.maxMatchOff") |
| 808 | } |
| 809 | if debugAsserts && t < 0 { |
| 810 | panic("t<0") |
| 811 | } |
| 812 | break |
| 813 | } |
| 814 | s += stepSize + ((s - nextEmit) >> (kSearchStrength - 1)) |
| 815 | if s >= sLimit { |
| 816 | break encodeLoop |
| 817 | } |
| 818 | cv = load6432(src, s) |
| 819 | } |
| 820 | // A 4-byte match has been found. We'll later see if more than 4 bytes. |
| 821 | offset2 = offset1 |
| 822 | offset1 = s - t |
| 823 | |
| 824 | if debugAsserts && s <= t { |
| 825 | panic(fmt.Sprintf("s (%d) <= t (%d)", s, t)) |
| 826 | } |
| 827 | |
| 828 | if debugAsserts && canRepeat && int(offset1) > len(src) { |
| 829 | panic("invalid offset") |
| 830 | } |
| 831 | |
| 832 | // Extend the 4-byte match as long as possible. |
| 833 | //l := e.matchlen(s+4, t+4, src) + 4 |
| 834 | var l int32 |
| 835 | { |
| 836 | a := src[s+4:] |
| 837 | b := src[t+4:] |
| 838 | endI := len(a) & (math.MaxInt32 - 7) |
| 839 | l = int32(endI) + 4 |
| 840 | for i := 0; i < endI; i += 8 { |
| 841 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 842 | l = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 843 | break |
| 844 | } |
| 845 | } |
| 846 | } |
| 847 | |
| 848 | // Extend backwards |
| 849 | tMin := s - e.maxMatchOff |
| 850 | if tMin < 0 { |
| 851 | tMin = 0 |
| 852 | } |
| 853 | for t > tMin && s > nextEmit && src[t-1] == src[s-1] && l < maxMatchLength { |
| 854 | s-- |
| 855 | t-- |
| 856 | l++ |
| 857 | } |
| 858 | |
| 859 | // Write our sequence. |
| 860 | var seq seq |
| 861 | seq.litLen = uint32(s - nextEmit) |
| 862 | seq.matchLen = uint32(l - zstdMinMatch) |
| 863 | if seq.litLen > 0 { |
| 864 | blk.literals = append(blk.literals, src[nextEmit:s]...) |
| 865 | } |
| 866 | // Don't use repeat offsets |
| 867 | seq.offset = uint32(s-t) + 3 |
| 868 | s += l |
| 869 | if debugSequences { |
| 870 | println("sequence", seq, "next s:", s) |
| 871 | } |
| 872 | blk.sequences = append(blk.sequences, seq) |
| 873 | nextEmit = s |
| 874 | if s >= sLimit { |
| 875 | break encodeLoop |
| 876 | } |
| 877 | cv = load6432(src, s) |
| 878 | |
| 879 | // Check offset 2 |
| 880 | if o2 := s - offset2; canRepeat && load3232(src, o2) == uint32(cv) { |
| 881 | // We have at least 4 byte match. |
| 882 | // No need to check backwards. We come straight from a match |
| 883 | //l := 4 + e.matchlen(s+4, o2+4, src) |
| 884 | var l int32 |
| 885 | { |
| 886 | a := src[s+4:] |
| 887 | b := src[o2+4:] |
| 888 | endI := len(a) & (math.MaxInt32 - 7) |
| 889 | l = int32(endI) + 4 |
| 890 | for i := 0; i < endI; i += 8 { |
| 891 | if diff := load64(a, i) ^ load64(b, i); diff != 0 { |
| 892 | l = int32(i+bits.TrailingZeros64(diff)>>3) + 4 |
| 893 | break |
| 894 | } |
| 895 | } |
| 896 | } |
| 897 | |
| 898 | // Store this, since we have it. |
| 899 | nextHash := hash6(cv, hashLog) |
| 900 | e.table[nextHash] = tableEntry{offset: s + e.cur, val: uint32(cv)} |
| 901 | e.markShardDirty(nextHash) |
| 902 | seq.matchLen = uint32(l) - zstdMinMatch |
| 903 | seq.litLen = 0 |
| 904 | // Since litlen is always 0, this is offset 1. |
| 905 | seq.offset = 1 |
| 906 | s += l |
| 907 | nextEmit = s |
| 908 | if debugSequences { |
| 909 | println("sequence", seq, "next s:", s) |
| 910 | } |
| 911 | blk.sequences = append(blk.sequences, seq) |
| 912 | |
| 913 | // Swap offset 1 and 2. |
| 914 | offset1, offset2 = offset2, offset1 |
| 915 | if s >= sLimit { |
| 916 | break encodeLoop |
| 917 | } |
| 918 | // Prepare next loop. |
| 919 | cv = load6432(src, s) |
| 920 | } |
| 921 | } |
| 922 | |
| 923 | if int(nextEmit) < len(src) { |
| 924 | blk.literals = append(blk.literals, src[nextEmit:]...) |
| 925 | blk.extraLits = len(src) - int(nextEmit) |
| 926 | } |
| 927 | blk.recentOffsets[0] = uint32(offset1) |
| 928 | blk.recentOffsets[1] = uint32(offset2) |
| 929 | if debug { |
| 930 | println("returning, recent offsets:", blk.recentOffsets, "extra literals:", blk.extraLits) |
| 931 | } |
| 932 | } |
| 933 | |
| 934 | // ResetDict will reset and set a dictionary if not nil |
| 935 | func (e *fastEncoder) Reset(d *dict, singleBlock bool) { |
| 936 | e.resetBase(d, singleBlock) |
| 937 | if d != nil { |
| 938 | panic("fastEncoder: Reset with dict") |
| 939 | } |
| 940 | } |
| 941 | |
| 942 | // ResetDict will reset and set a dictionary if not nil |
| 943 | func (e *fastEncoderDict) Reset(d *dict, singleBlock bool) { |
| 944 | e.resetBase(d, singleBlock) |
| 945 | if d == nil { |
| 946 | return |
| 947 | } |
| 948 | |
| 949 | // Init or copy dict table |
| 950 | if len(e.dictTable) != len(e.table) || d.id != e.lastDictID { |
| 951 | if len(e.dictTable) != len(e.table) { |
| 952 | e.dictTable = make([]tableEntry, len(e.table)) |
| 953 | } |
| 954 | if true { |
| 955 | end := e.maxMatchOff + int32(len(d.content)) - 8 |
| 956 | for i := e.maxMatchOff; i < end; i += 3 { |
| 957 | const hashLog = tableBits |
| 958 | |
| 959 | cv := load6432(d.content, i-e.maxMatchOff) |
| 960 | nextHash := hash6(cv, hashLog) // 0 -> 5 |
| 961 | nextHash1 := hash6(cv>>8, hashLog) // 1 -> 6 |
| 962 | nextHash2 := hash6(cv>>16, hashLog) // 2 -> 7 |
| 963 | e.dictTable[nextHash] = tableEntry{ |
| 964 | val: uint32(cv), |
| 965 | offset: i, |
| 966 | } |
| 967 | e.dictTable[nextHash1] = tableEntry{ |
| 968 | val: uint32(cv >> 8), |
| 969 | offset: i + 1, |
| 970 | } |
| 971 | e.dictTable[nextHash2] = tableEntry{ |
| 972 | val: uint32(cv >> 16), |
| 973 | offset: i + 2, |
| 974 | } |
| 975 | } |
| 976 | } |
| 977 | e.lastDictID = d.id |
| 978 | e.allDirty = true |
| 979 | } |
| 980 | |
| 981 | e.cur = e.maxMatchOff |
| 982 | dirtyShardCnt := 0 |
| 983 | if !e.allDirty { |
| 984 | for i := range e.tableShardDirty { |
| 985 | if e.tableShardDirty[i] { |
| 986 | dirtyShardCnt++ |
| 987 | } |
| 988 | } |
| 989 | } |
| 990 | |
| 991 | const shardCnt = tableShardCnt |
| 992 | const shardSize = tableShardSize |
| 993 | if e.allDirty || dirtyShardCnt > shardCnt*4/6 { |
| 994 | copy(e.table[:], e.dictTable) |
| 995 | for i := range e.tableShardDirty { |
| 996 | e.tableShardDirty[i] = false |
| 997 | } |
| 998 | e.allDirty = false |
| 999 | return |
| 1000 | } |
| 1001 | for i := range e.tableShardDirty { |
| 1002 | if !e.tableShardDirty[i] { |
| 1003 | continue |
| 1004 | } |
| 1005 | |
| 1006 | copy(e.table[i*shardSize:(i+1)*shardSize], e.dictTable[i*shardSize:(i+1)*shardSize]) |
| 1007 | e.tableShardDirty[i] = false |
| 1008 | } |
| 1009 | e.allDirty = false |
| 1010 | } |
| 1011 | |
| 1012 | func (e *fastEncoderDict) markAllShardsDirty() { |
| 1013 | e.allDirty = true |
| 1014 | } |
| 1015 | |
| 1016 | func (e *fastEncoderDict) markShardDirty(entryNum uint32) { |
| 1017 | e.tableShardDirty[entryNum/tableShardSize] = true |
| 1018 | } |