deflate.mx raw

   1  // Copyright 2009 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 flate
   6  
   7  import (
   8  	"errors"
   9  	"fmt"
  10  	"io"
  11  	"math"
  12  )
  13  
  14  const (
  15  	NoCompression      = 0
  16  	BestSpeed          = 1
  17  	BestCompression    = 9
  18  	DefaultCompression = -1
  19  
  20  	// HuffmanOnly disables Lempel-Ziv match searching and only performs Huffman
  21  	// entropy encoding. This mode is useful in compressing data that has
  22  	// already been compressed with an LZ style algorithm (e.g. Snappy or LZ4)
  23  	// that lacks an entropy encoder. Compression gains are achieved when
  24  	// certain bytes in the input stream occur more frequently than others.
  25  	//
  26  	// Note that HuffmanOnly produces a compressed output that is
  27  	// RFC 1951 compliant. That is, any valid DEFLATE decompressor will
  28  	// continue to be able to decompress this output.
  29  	HuffmanOnly = -2
  30  )
  31  
  32  const (
  33  	logWindowSize = 15
  34  	windowSize    = 1 << logWindowSize
  35  	windowMask    = windowSize - 1
  36  
  37  	// The LZ77 step produces a sequence of literal tokens and <length, offset>
  38  	// pair tokens. The offset is also known as distance. The underlying wire
  39  	// format limits the range of lengths and offsets. For example, there are
  40  	// 256 legitimate lengths: those in the range [3, 258]. This package's
  41  	// compressor uses a higher minimum match length, enabling optimizations
  42  	// such as finding matches via 32-bit loads and compares.
  43  	baseMatchLength = 3       // The smallest match length per the RFC section 3.2.5
  44  	minMatchLength  = 4       // The smallest match length that the compressor actually emits
  45  	maxMatchLength  = 258     // The largest match length
  46  	baseMatchOffset = 1       // The smallest match offset
  47  	maxMatchOffset  = 1 << 15 // The largest match offset
  48  
  49  	// The maximum number of tokens we put into a single flate block, just to
  50  	// stop things from getting too large.
  51  	maxFlateBlockTokens = 1 << 14
  52  	maxStoreBlockSize   = 65535
  53  	hashBits            = 17 // After 17 performance degrades
  54  	hashSize            = 1 << hashBits
  55  	hashMask            = (1 << hashBits) - 1
  56  	maxHashOffset       = 1 << 24
  57  
  58  	skipNever = math.MaxInt32
  59  )
  60  
  61  type compressionLevel struct {
  62  	level, good, lazy, nice, chain, fastSkipHashing int
  63  }
  64  
  65  var levels = []compressionLevel{
  66  	{0, 0, 0, 0, 0, 0}, // NoCompression.
  67  	{1, 0, 0, 0, 0, 0}, // BestSpeed uses a custom algorithm; see deflatefast.go.
  68  	// For levels 2-3 we don't bother trying with lazy matches.
  69  	{2, 4, 0, 16, 8, 5},
  70  	{3, 4, 0, 32, 32, 6},
  71  	// Levels 4-9 use increasingly more lazy matching
  72  	// and increasingly stringent conditions for "good enough".
  73  	{4, 4, 4, 16, 16, skipNever},
  74  	{5, 8, 16, 32, 32, skipNever},
  75  	{6, 8, 16, 128, 128, skipNever},
  76  	{7, 8, 32, 128, 256, skipNever},
  77  	{8, 32, 128, 258, 1024, skipNever},
  78  	{9, 32, 258, 258, 4096, skipNever},
  79  }
  80  
  81  type compressor struct {
  82  	compressionLevel
  83  
  84  	w          *huffmanBitWriter
  85  	bulkHasher func([]byte, []uint32)
  86  
  87  	// compression algorithm
  88  	fill      func(*compressor, []byte) int // copy data to window
  89  	step      func(*compressor)             // process window
  90  	bestSpeed *deflateFast                  // Encoder for BestSpeed
  91  
  92  	// Input hash chains
  93  	// hashHead[hashValue] contains the largest inputIndex with the specified hash value
  94  	// If hashHead[hashValue] is within the current window, then
  95  	// hashPrev[hashHead[hashValue] & windowMask] contains the previous index
  96  	// with the same hash value.
  97  	chainHead  int
  98  	hashHead   [hashSize]uint32
  99  	hashPrev   [windowSize]uint32
 100  	hashOffset int
 101  
 102  	// input window: unprocessed data is window[index:windowEnd]
 103  	index         int
 104  	window        []byte
 105  	windowEnd     int
 106  	blockStart    int  // window index where current tokens start
 107  	byteAvailable bool // if true, still need to process window[index-1].
 108  
 109  	sync bool // requesting flush
 110  
 111  	// queued output tokens
 112  	tokens []token
 113  
 114  	// deflate state
 115  	length         int
 116  	offset         int
 117  	maxInsertIndex int
 118  	err            error
 119  
 120  	// hashMatch must be able to contain hashes for the maximum match length.
 121  	hashMatch [maxMatchLength - 1]uint32
 122  }
 123  
 124  func (d *compressor) fillDeflate(b []byte) int {
 125  	if d.index >= 2*windowSize-(minMatchLength+maxMatchLength) {
 126  		// shift the window by windowSize
 127  		copy(d.window, d.window[windowSize:2*windowSize])
 128  		d.index -= windowSize
 129  		d.windowEnd -= windowSize
 130  		if d.blockStart >= windowSize {
 131  			d.blockStart -= windowSize
 132  		} else {
 133  			d.blockStart = math.MaxInt32
 134  		}
 135  		d.hashOffset += windowSize
 136  		if d.hashOffset > maxHashOffset {
 137  			delta := d.hashOffset - 1
 138  			d.hashOffset -= delta
 139  			d.chainHead -= delta
 140  
 141  			// Iterate over slices instead of arrays to avoid copying
 142  			// the entire table onto the stack (Issue #18625).
 143  			for i, v := range d.hashPrev[:] {
 144  				if int(v) > delta {
 145  					d.hashPrev[i] = uint32(int(v) - delta)
 146  				} else {
 147  					d.hashPrev[i] = 0
 148  				}
 149  			}
 150  			for i, v := range d.hashHead[:] {
 151  				if int(v) > delta {
 152  					d.hashHead[i] = uint32(int(v) - delta)
 153  				} else {
 154  					d.hashHead[i] = 0
 155  				}
 156  			}
 157  		}
 158  	}
 159  	n := copy(d.window[d.windowEnd:], b)
 160  	d.windowEnd += n
 161  	return n
 162  }
 163  
 164  func (d *compressor) writeBlock(tokens []token, index int) error {
 165  	if index > 0 {
 166  		var window []byte
 167  		if d.blockStart <= index {
 168  			window = d.window[d.blockStart:index]
 169  		}
 170  		d.blockStart = index
 171  		d.w.writeBlock(tokens, false, window)
 172  		return d.w.err
 173  	}
 174  	return nil
 175  }
 176  
 177  // fillWindow will fill the current window with the supplied
 178  // dictionary and calculate all hashes.
 179  // This is much faster than doing a full encode.
 180  // Should only be used after a reset.
 181  func (d *compressor) fillWindow(b []byte) {
 182  	// Do not fill window if we are in store-only mode.
 183  	if d.compressionLevel.level < 2 {
 184  		return
 185  	}
 186  	if d.index != 0 || d.windowEnd != 0 {
 187  		panic("internal error: fillWindow called with stale data")
 188  	}
 189  
 190  	// If we are given too much, cut it.
 191  	if len(b) > windowSize {
 192  		b = b[len(b)-windowSize:]
 193  	}
 194  	// Add all to window.
 195  	n := copy(d.window, b)
 196  
 197  	// Calculate 256 hashes at the time (more L1 cache hits)
 198  	loops := (n + 256 - minMatchLength) / 256
 199  	for j := 0; j < loops; j++ {
 200  		index := j * 256
 201  		end := index + 256 + minMatchLength - 1
 202  		if end > n {
 203  			end = n
 204  		}
 205  		toCheck := d.window[index:end]
 206  		dstSize := len(toCheck) - minMatchLength + 1
 207  
 208  		if dstSize <= 0 {
 209  			continue
 210  		}
 211  
 212  		dst := d.hashMatch[:dstSize]
 213  		d.bulkHasher(toCheck, dst)
 214  		for i, val := range dst {
 215  			di := i + index
 216  			hh := &d.hashHead[val&hashMask]
 217  			// Get previous value with the same hash.
 218  			// Our chain should point to the previous value.
 219  			d.hashPrev[di&windowMask] = *hh
 220  			// Set the head of the hash chain to us.
 221  			*hh = uint32(di + d.hashOffset)
 222  		}
 223  	}
 224  	// Update window information.
 225  	d.windowEnd = n
 226  	d.index = n
 227  }
 228  
 229  // Try to find a match starting at index whose length is greater than prevSize.
 230  // We only look at chainCount possibilities before giving up.
 231  func (d *compressor) findMatch(pos int, prevHead int, prevLength int, lookahead int) (length, offset int, ok bool) {
 232  	minMatchLook := maxMatchLength
 233  	if lookahead < minMatchLook {
 234  		minMatchLook = lookahead
 235  	}
 236  
 237  	win := d.window[0 : pos+minMatchLook]
 238  
 239  	// We quit when we get a match that's at least nice long
 240  	nice := len(win) - pos
 241  	if d.nice < nice {
 242  		nice = d.nice
 243  	}
 244  
 245  	// If we've got a match that's good enough, only look in 1/4 the chain.
 246  	tries := d.chain
 247  	length = prevLength
 248  	if length >= d.good {
 249  		tries >>= 2
 250  	}
 251  
 252  	wEnd := win[pos+length]
 253  	wPos := win[pos:]
 254  	minIndex := pos - windowSize
 255  
 256  	for i := prevHead; tries > 0; tries-- {
 257  		if wEnd == win[i+length] {
 258  			n := matchLen(win[i:], wPos, minMatchLook)
 259  
 260  			if n > length && (n > minMatchLength || pos-i <= 4096) {
 261  				length = n
 262  				offset = pos - i
 263  				ok = true
 264  				if n >= nice {
 265  					// The match is good enough that we don't try to find a better one.
 266  					break
 267  				}
 268  				wEnd = win[pos+n]
 269  			}
 270  		}
 271  		if i == minIndex {
 272  			// hashPrev[i & windowMask] has already been overwritten, so stop now.
 273  			break
 274  		}
 275  		i = int(d.hashPrev[i&windowMask]) - d.hashOffset
 276  		if i < minIndex || i < 0 {
 277  			break
 278  		}
 279  	}
 280  	return
 281  }
 282  
 283  func (d *compressor) writeStoredBlock(buf []byte) error {
 284  	if d.w.writeStoredHeader(len(buf), false); d.w.err != nil {
 285  		return d.w.err
 286  	}
 287  	d.w.writeBytes(buf)
 288  	return d.w.err
 289  }
 290  
 291  const hashmul = 0x1e35a7bd
 292  
 293  // hash4 returns a hash representation of the first 4 bytes
 294  // of the supplied slice.
 295  // The caller must ensure that len(b) >= 4.
 296  func hash4(b []byte) uint32 {
 297  	return ((uint32(b[3]) | uint32(b[2])<<8 | uint32(b[1])<<16 | uint32(b[0])<<24) * hashmul) >> (32 - hashBits)
 298  }
 299  
 300  // bulkHash4 will compute hashes using the same
 301  // algorithm as hash4.
 302  func bulkHash4(b []byte, dst []uint32) {
 303  	if len(b) < minMatchLength {
 304  		return
 305  	}
 306  	hb := uint32(b[3]) | uint32(b[2])<<8 | uint32(b[1])<<16 | uint32(b[0])<<24
 307  	dst[0] = (hb * hashmul) >> (32 - hashBits)
 308  	end := len(b) - minMatchLength + 1
 309  	for i := 1; i < end; i++ {
 310  		hb = (hb << 8) | uint32(b[i+3])
 311  		dst[i] = (hb * hashmul) >> (32 - hashBits)
 312  	}
 313  }
 314  
 315  // matchLen returns the number of matching bytes in a and b
 316  // up to length 'max'. Both slices must be at least 'max'
 317  // bytes in size.
 318  func matchLen(a, b []byte, max int) int {
 319  	a = a[:max]
 320  	b = b[:len(a)]
 321  	for i, av := range a {
 322  		if b[i] != av {
 323  			return i
 324  		}
 325  	}
 326  	return max
 327  }
 328  
 329  // encSpeed will compress and store the currently added data,
 330  // if enough has been accumulated or we at the end of the stream.
 331  // Any error that occurred will be in d.err
 332  func (d *compressor) encSpeed() {
 333  	// We only compress if we have maxStoreBlockSize.
 334  	if d.windowEnd < maxStoreBlockSize {
 335  		if !d.sync {
 336  			return
 337  		}
 338  
 339  		// Handle small sizes.
 340  		if d.windowEnd < 128 {
 341  			switch {
 342  			case d.windowEnd == 0:
 343  				return
 344  			case d.windowEnd <= 16:
 345  				d.err = d.writeStoredBlock(d.window[:d.windowEnd])
 346  			default:
 347  				d.w.writeBlockHuff(false, d.window[:d.windowEnd])
 348  				d.err = d.w.err
 349  			}
 350  			d.windowEnd = 0
 351  			d.bestSpeed.reset()
 352  			return
 353  		}
 354  
 355  	}
 356  	// Encode the block.
 357  	d.tokens = d.bestSpeed.encode(d.tokens[:0], d.window[:d.windowEnd])
 358  
 359  	// If we removed less than 1/16th, Huffman compress the block.
 360  	if len(d.tokens) > d.windowEnd-(d.windowEnd>>4) {
 361  		d.w.writeBlockHuff(false, d.window[:d.windowEnd])
 362  	} else {
 363  		d.w.writeBlockDynamic(d.tokens, false, d.window[:d.windowEnd])
 364  	}
 365  	d.err = d.w.err
 366  	d.windowEnd = 0
 367  }
 368  
 369  func (d *compressor) initDeflate() {
 370  	d.window = []byte{:2*windowSize}
 371  	d.hashOffset = 1
 372  	d.tokens = []token{:0:maxFlateBlockTokens+1}
 373  	d.length = minMatchLength - 1
 374  	d.offset = 0
 375  	d.byteAvailable = false
 376  	d.index = 0
 377  	d.chainHead = -1
 378  	d.bulkHasher = bulkHash4
 379  }
 380  
 381  func (d *compressor) deflate() {
 382  	if d.windowEnd-d.index < minMatchLength+maxMatchLength && !d.sync {
 383  		return
 384  	}
 385  
 386  	d.maxInsertIndex = d.windowEnd - (minMatchLength - 1)
 387  
 388  Loop:
 389  	for {
 390  		if d.index > d.windowEnd {
 391  			panic("index > windowEnd")
 392  		}
 393  		lookahead := d.windowEnd - d.index
 394  		if lookahead < minMatchLength+maxMatchLength {
 395  			if !d.sync {
 396  				break Loop
 397  			}
 398  			if d.index > d.windowEnd {
 399  				panic("index > windowEnd")
 400  			}
 401  			if lookahead == 0 {
 402  				// Flush current output block if any.
 403  				if d.byteAvailable {
 404  					// There is still one pending token that needs to be flushed
 405  					d.tokens = append(d.tokens, literalToken(uint32(d.window[d.index-1])))
 406  					d.byteAvailable = false
 407  				}
 408  				if len(d.tokens) > 0 {
 409  					if d.err = d.writeBlock(d.tokens, d.index); d.err != nil {
 410  						return
 411  					}
 412  					d.tokens = d.tokens[:0]
 413  				}
 414  				break Loop
 415  			}
 416  		}
 417  		if d.index < d.maxInsertIndex {
 418  			// Update the hash
 419  			hash := hash4(d.window[d.index : d.index+minMatchLength])
 420  			hh := &d.hashHead[hash&hashMask]
 421  			d.chainHead = int(*hh)
 422  			d.hashPrev[d.index&windowMask] = uint32(d.chainHead)
 423  			*hh = uint32(d.index + d.hashOffset)
 424  		}
 425  		prevLength := d.length
 426  		prevOffset := d.offset
 427  		d.length = minMatchLength - 1
 428  		d.offset = 0
 429  		minIndex := d.index - windowSize
 430  		if minIndex < 0 {
 431  			minIndex = 0
 432  		}
 433  
 434  		if d.chainHead-d.hashOffset >= minIndex &&
 435  			(d.fastSkipHashing != skipNever && lookahead > minMatchLength-1 ||
 436  				d.fastSkipHashing == skipNever && lookahead > prevLength && prevLength < d.lazy) {
 437  			if newLength, newOffset, ok := d.findMatch(d.index, d.chainHead-d.hashOffset, minMatchLength-1, lookahead); ok {
 438  				d.length = newLength
 439  				d.offset = newOffset
 440  			}
 441  		}
 442  		if d.fastSkipHashing != skipNever && d.length >= minMatchLength ||
 443  			d.fastSkipHashing == skipNever && prevLength >= minMatchLength && d.length <= prevLength {
 444  			// There was a match at the previous step, and the current match is
 445  			// not better. Output the previous match.
 446  			if d.fastSkipHashing != skipNever {
 447  				d.tokens = append(d.tokens, matchToken(uint32(d.length-baseMatchLength), uint32(d.offset-baseMatchOffset)))
 448  			} else {
 449  				d.tokens = append(d.tokens, matchToken(uint32(prevLength-baseMatchLength), uint32(prevOffset-baseMatchOffset)))
 450  			}
 451  			// Insert in the hash table all strings up to the end of the match.
 452  			// index and index-1 are already inserted. If there is not enough
 453  			// lookahead, the last two strings are not inserted into the hash
 454  			// table.
 455  			if d.length <= d.fastSkipHashing {
 456  				var newIndex int
 457  				if d.fastSkipHashing != skipNever {
 458  					newIndex = d.index + d.length
 459  				} else {
 460  					newIndex = d.index + prevLength - 1
 461  				}
 462  				index := d.index
 463  				for index++; index < newIndex; index++ {
 464  					if index < d.maxInsertIndex {
 465  						hash := hash4(d.window[index : index+minMatchLength])
 466  						// Get previous value with the same hash.
 467  						// Our chain should point to the previous value.
 468  						hh := &d.hashHead[hash&hashMask]
 469  						d.hashPrev[index&windowMask] = *hh
 470  						// Set the head of the hash chain to us.
 471  						*hh = uint32(index + d.hashOffset)
 472  					}
 473  				}
 474  				d.index = index
 475  
 476  				if d.fastSkipHashing == skipNever {
 477  					d.byteAvailable = false
 478  					d.length = minMatchLength - 1
 479  				}
 480  			} else {
 481  				// For matches this long, we don't bother inserting each individual
 482  				// item into the table.
 483  				d.index += d.length
 484  			}
 485  			if len(d.tokens) == maxFlateBlockTokens {
 486  				// The block includes the current character
 487  				if d.err = d.writeBlock(d.tokens, d.index); d.err != nil {
 488  					return
 489  				}
 490  				d.tokens = d.tokens[:0]
 491  			}
 492  		} else {
 493  			if d.fastSkipHashing != skipNever || d.byteAvailable {
 494  				i := d.index - 1
 495  				if d.fastSkipHashing != skipNever {
 496  					i = d.index
 497  				}
 498  				d.tokens = append(d.tokens, literalToken(uint32(d.window[i])))
 499  				if len(d.tokens) == maxFlateBlockTokens {
 500  					if d.err = d.writeBlock(d.tokens, i+1); d.err != nil {
 501  						return
 502  					}
 503  					d.tokens = d.tokens[:0]
 504  				}
 505  			}
 506  			d.index++
 507  			if d.fastSkipHashing == skipNever {
 508  				d.byteAvailable = true
 509  			}
 510  		}
 511  	}
 512  }
 513  
 514  func (d *compressor) fillStore(b []byte) int {
 515  	n := copy(d.window[d.windowEnd:], b)
 516  	d.windowEnd += n
 517  	return n
 518  }
 519  
 520  func (d *compressor) store() {
 521  	if d.windowEnd > 0 && (d.windowEnd == maxStoreBlockSize || d.sync) {
 522  		d.err = d.writeStoredBlock(d.window[:d.windowEnd])
 523  		d.windowEnd = 0
 524  	}
 525  }
 526  
 527  // storeHuff compresses and stores the currently added data
 528  // when the d.window is full or we are at the end of the stream.
 529  // Any error that occurred will be in d.err
 530  func (d *compressor) storeHuff() {
 531  	if d.windowEnd < len(d.window) && !d.sync || d.windowEnd == 0 {
 532  		return
 533  	}
 534  	d.w.writeBlockHuff(false, d.window[:d.windowEnd])
 535  	d.err = d.w.err
 536  	d.windowEnd = 0
 537  }
 538  
 539  func (d *compressor) write(b []byte) (n int, err error) {
 540  	if d.err != nil {
 541  		return 0, d.err
 542  	}
 543  	n = len(b)
 544  	for len(b) > 0 {
 545  		d.step(d)
 546  		b = b[d.fill(d, b):]
 547  		if d.err != nil {
 548  			return 0, d.err
 549  		}
 550  	}
 551  	return n, nil
 552  }
 553  
 554  func (d *compressor) syncFlush() error {
 555  	if d.err != nil {
 556  		return d.err
 557  	}
 558  	d.sync = true
 559  	d.step(d)
 560  	if d.err == nil {
 561  		d.w.writeStoredHeader(0, false)
 562  		d.w.flush()
 563  		d.err = d.w.err
 564  	}
 565  	d.sync = false
 566  	return d.err
 567  }
 568  
 569  func (d *compressor) init(w io.Writer, level int) (err error) {
 570  	d.w = newHuffmanBitWriter(w)
 571  
 572  	switch {
 573  	case level == NoCompression:
 574  		d.window = []byte{:maxStoreBlockSize}
 575  		d.fill = (*compressor).fillStore
 576  		d.step = (*compressor).store
 577  	case level == HuffmanOnly:
 578  		d.window = []byte{:maxStoreBlockSize}
 579  		d.fill = (*compressor).fillStore
 580  		d.step = (*compressor).storeHuff
 581  	case level == BestSpeed:
 582  		d.compressionLevel = levels[level]
 583  		d.window = []byte{:maxStoreBlockSize}
 584  		d.fill = (*compressor).fillStore
 585  		d.step = (*compressor).encSpeed
 586  		d.bestSpeed = newDeflateFast()
 587  		d.tokens = []token{:maxStoreBlockSize}
 588  	case level == DefaultCompression:
 589  		level = 6
 590  		d.compressionLevel = levels[level]
 591  		d.initDeflate()
 592  		d.fill = (*compressor).fillDeflate
 593  		d.step = (*compressor).deflate
 594  	case 2 <= level && level <= 9:
 595  		d.compressionLevel = levels[level]
 596  		d.initDeflate()
 597  		d.fill = (*compressor).fillDeflate
 598  		d.step = (*compressor).deflate
 599  	default:
 600  		return fmt.Errorf("flate: invalid compression level %d: want value in range [-2, 9]", level)
 601  	}
 602  	return nil
 603  }
 604  
 605  func (d *compressor) reset(w io.Writer) {
 606  	d.w.reset(w)
 607  	d.sync = false
 608  	d.err = nil
 609  	switch d.compressionLevel.level {
 610  	case NoCompression:
 611  		d.windowEnd = 0
 612  	case BestSpeed:
 613  		d.windowEnd = 0
 614  		d.tokens = d.tokens[:0]
 615  		d.bestSpeed.reset()
 616  	default:
 617  		d.chainHead = -1
 618  		clear(d.hashHead[:])
 619  		clear(d.hashPrev[:])
 620  		d.hashOffset = 1
 621  		d.index, d.windowEnd = 0, 0
 622  		d.blockStart, d.byteAvailable = 0, false
 623  		d.tokens = d.tokens[:0]
 624  		d.length = minMatchLength - 1
 625  		d.offset = 0
 626  		d.maxInsertIndex = 0
 627  	}
 628  }
 629  
 630  func (d *compressor) close() error {
 631  	if d.err == errWriterClosed {
 632  		return nil
 633  	}
 634  	if d.err != nil {
 635  		return d.err
 636  	}
 637  	d.sync = true
 638  	d.step(d)
 639  	if d.err != nil {
 640  		return d.err
 641  	}
 642  	if d.w.writeStoredHeader(0, true); d.w.err != nil {
 643  		return d.w.err
 644  	}
 645  	d.w.flush()
 646  	if d.w.err != nil {
 647  		return d.w.err
 648  	}
 649  	d.err = errWriterClosed
 650  	return nil
 651  }
 652  
 653  // NewWriter returns a new [Writer] compressing data at the given level.
 654  // Following zlib, levels range from 1 ([BestSpeed]) to 9 ([BestCompression]);
 655  // higher levels typically run slower but compress more. Level 0
 656  // ([NoCompression]) does not attempt any compression; it only adds the
 657  // necessary DEFLATE framing.
 658  // Level -1 ([DefaultCompression]) uses the default compression level.
 659  // Level -2 ([HuffmanOnly]) will use Huffman compression only, giving
 660  // a very fast compression for all types of input, but sacrificing considerable
 661  // compression efficiency.
 662  //
 663  // If level is in the range [-2, 9] then the error returned will be nil.
 664  // Otherwise the error returned will be non-nil.
 665  func NewWriter(w io.Writer, level int) (*Writer, error) {
 666  	var dw Writer
 667  	if err := dw.d.init(w, level); err != nil {
 668  		return nil, err
 669  	}
 670  	return &dw, nil
 671  }
 672  
 673  // NewWriterDict is like [NewWriter] but initializes the new
 674  // [Writer] with a preset dictionary. The returned [Writer] behaves
 675  // as if the dictionary had been written to it without producing
 676  // any compressed output. The compressed data written to w
 677  // can only be decompressed by a reader initialized with the
 678  // same dictionary (see [NewReaderDict]).
 679  func NewWriterDict(w io.Writer, level int, dict []byte) (*Writer, error) {
 680  	dw := &dictWriter{w}
 681  	zw, err := NewWriter(dw, level)
 682  	if err != nil {
 683  		return nil, err
 684  	}
 685  	zw.d.fillWindow(dict)
 686  	zw.dict = append(zw.dict, dict...) // duplicate dictionary for Reset method.
 687  	return zw, nil
 688  }
 689  
 690  type dictWriter struct {
 691  	w io.Writer
 692  }
 693  
 694  func (w *dictWriter) Write(b []byte) (n int, err error) {
 695  	return w.w.Write(b)
 696  }
 697  
 698  var errWriterClosed = errors.New("flate: closed writer")
 699  
 700  // A Writer takes data written to it and writes the compressed
 701  // form of that data to an underlying writer (see [NewWriter]).
 702  type Writer struct {
 703  	d    compressor
 704  	dict []byte
 705  }
 706  
 707  // Write writes data to w, which will eventually write the
 708  // compressed form of data to its underlying writer.
 709  func (w *Writer) Write(data []byte) (n int, err error) {
 710  	return w.d.write(data)
 711  }
 712  
 713  // Flush flushes any pending data to the underlying writer.
 714  // It is useful mainly in compressed network protocols, to ensure that
 715  // a remote reader has enough data to reconstruct a packet.
 716  // Flush does not return until the data has been written.
 717  // Calling Flush when there is no pending data still causes the [Writer]
 718  // to emit a sync marker of at least 4 bytes.
 719  // If the underlying writer returns an error, Flush returns that error.
 720  //
 721  // In the terminology of the zlib library, Flush is equivalent to Z_SYNC_FLUSH.
 722  func (w *Writer) Flush() error {
 723  	// For more about flushing:
 724  	// https://www.bolet.org/~pornin/deflate-flush.html
 725  	return w.d.syncFlush()
 726  }
 727  
 728  // Close flushes and closes the writer.
 729  func (w *Writer) Close() error {
 730  	return w.d.close()
 731  }
 732  
 733  // Reset discards the writer's state and makes it equivalent to
 734  // the result of [NewWriter] or [NewWriterDict] called with dst
 735  // and w's level and dictionary.
 736  func (w *Writer) Reset(dst io.Writer) {
 737  	if dw, ok := w.d.w.writer.(*dictWriter); ok {
 738  		// w was created with NewWriterDict
 739  		dw.w = dst
 740  		w.d.reset(dw)
 741  		w.d.fillWindow(w.dict)
 742  	} else {
 743  		// w was created with NewWriter
 744  		w.d.reset(dst)
 745  	}
 746  }
 747