Black Lives Matter. Support the Equal Justice Initiative.

Source file src/compress/flate/huffman_bit_writer.go

Documentation: compress/flate

     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  	"io"
     9  )
    10  
    11  const (
    12  	// The largest offset code.
    13  	offsetCodeCount = 30
    14  
    15  	// The special code used to mark the end of a block.
    16  	endBlockMarker = 256
    17  
    18  	// The first length code.
    19  	lengthCodesStart = 257
    20  
    21  	// The number of codegen codes.
    22  	codegenCodeCount = 19
    23  	badCode          = 255
    24  
    25  	// bufferFlushSize indicates the buffer size
    26  	// after which bytes are flushed to the writer.
    27  	// Should preferably be a multiple of 6, since
    28  	// we accumulate 6 bytes between writes to the buffer.
    29  	bufferFlushSize = 240
    30  
    31  	// bufferSize is the actual output byte buffer size.
    32  	// It must have additional headroom for a flush
    33  	// which can contain up to 8 bytes.
    34  	bufferSize = bufferFlushSize + 8
    35  )
    36  
    37  // The number of extra bits needed by length code X - LENGTH_CODES_START.
    38  var lengthExtraBits = []int8{
    39  	/* 257 */ 0, 0, 0,
    40  	/* 260 */ 0, 0, 0, 0, 0, 1, 1, 1, 1, 2,
    41  	/* 270 */ 2, 2, 2, 3, 3, 3, 3, 4, 4, 4,
    42  	/* 280 */ 4, 5, 5, 5, 5, 0,
    43  }
    44  
    45  // The length indicated by length code X - LENGTH_CODES_START.
    46  var lengthBase = []uint32{
    47  	0, 1, 2, 3, 4, 5, 6, 7, 8, 10,
    48  	12, 14, 16, 20, 24, 28, 32, 40, 48, 56,
    49  	64, 80, 96, 112, 128, 160, 192, 224, 255,
    50  }
    51  
    52  // offset code word extra bits.
    53  var offsetExtraBits = []int8{
    54  	0, 0, 0, 0, 1, 1, 2, 2, 3, 3,
    55  	4, 4, 5, 5, 6, 6, 7, 7, 8, 8,
    56  	9, 9, 10, 10, 11, 11, 12, 12, 13, 13,
    57  }
    58  
    59  var offsetBase = []uint32{
    60  	0x000000, 0x000001, 0x000002, 0x000003, 0x000004,
    61  	0x000006, 0x000008, 0x00000c, 0x000010, 0x000018,
    62  	0x000020, 0x000030, 0x000040, 0x000060, 0x000080,
    63  	0x0000c0, 0x000100, 0x000180, 0x000200, 0x000300,
    64  	0x000400, 0x000600, 0x000800, 0x000c00, 0x001000,
    65  	0x001800, 0x002000, 0x003000, 0x004000, 0x006000,
    66  }
    67  
    68  // The odd order in which the codegen code sizes are written.
    69  var codegenOrder = []uint32{16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}
    70  
    71  type huffmanBitWriter struct {
    72  	// writer is the underlying writer.
    73  	// Do not use it directly; use the write method, which ensures
    74  	// that Write errors are sticky.
    75  	writer io.Writer
    76  
    77  	// Data waiting to be written is bytes[0:nbytes]
    78  	// and then the low nbits of bits.  Data is always written
    79  	// sequentially into the bytes array.
    80  	bits            uint64
    81  	nbits           uint
    82  	bytes           [bufferSize]byte
    83  	codegenFreq     [codegenCodeCount]int32
    84  	nbytes          int
    85  	literalFreq     []int32
    86  	offsetFreq      []int32
    87  	codegen         []uint8
    88  	literalEncoding *huffmanEncoder
    89  	offsetEncoding  *huffmanEncoder
    90  	codegenEncoding *huffmanEncoder
    91  	err             error
    92  }
    93  
    94  func newHuffmanBitWriter(w io.Writer) *huffmanBitWriter {
    95  	return &huffmanBitWriter{
    96  		writer:          w,
    97  		literalFreq:     make([]int32, maxNumLit),
    98  		offsetFreq:      make([]int32, offsetCodeCount),
    99  		codegen:         make([]uint8, maxNumLit+offsetCodeCount+1),
   100  		literalEncoding: newHuffmanEncoder(maxNumLit),
   101  		codegenEncoding: newHuffmanEncoder(codegenCodeCount),
   102  		offsetEncoding:  newHuffmanEncoder(offsetCodeCount),
   103  	}
   104  }
   105  
   106  func (w *huffmanBitWriter) reset(writer io.Writer) {
   107  	w.writer = writer
   108  	w.bits, w.nbits, w.nbytes, w.err = 0, 0, 0, nil
   109  }
   110  
   111  func (w *huffmanBitWriter) flush() {
   112  	if w.err != nil {
   113  		w.nbits = 0
   114  		return
   115  	}
   116  	n := w.nbytes
   117  	for w.nbits != 0 {
   118  		w.bytes[n] = byte(w.bits)
   119  		w.bits >>= 8
   120  		if w.nbits > 8 { // Avoid underflow
   121  			w.nbits -= 8
   122  		} else {
   123  			w.nbits = 0
   124  		}
   125  		n++
   126  	}
   127  	w.bits = 0
   128  	w.write(w.bytes[:n])
   129  	w.nbytes = 0
   130  }
   131  
   132  func (w *huffmanBitWriter) write(b []byte) {
   133  	if w.err != nil {
   134  		return
   135  	}
   136  	_, w.err = w.writer.Write(b)
   137  }
   138  
   139  func (w *huffmanBitWriter) writeBits(b int32, nb uint) {
   140  	if w.err != nil {
   141  		return
   142  	}
   143  	w.bits |= uint64(b) << w.nbits
   144  	w.nbits += nb
   145  	if w.nbits >= 48 {
   146  		bits := w.bits
   147  		w.bits >>= 48
   148  		w.nbits -= 48
   149  		n := w.nbytes
   150  		bytes := w.bytes[n : n+6]
   151  		bytes[0] = byte(bits)
   152  		bytes[1] = byte(bits >> 8)
   153  		bytes[2] = byte(bits >> 16)
   154  		bytes[3] = byte(bits >> 24)
   155  		bytes[4] = byte(bits >> 32)
   156  		bytes[5] = byte(bits >> 40)
   157  		n += 6
   158  		if n >= bufferFlushSize {
   159  			w.write(w.bytes[:n])
   160  			n = 0
   161  		}
   162  		w.nbytes = n
   163  	}
   164  }
   165  
   166  func (w *huffmanBitWriter) writeBytes(bytes []byte) {
   167  	if w.err != nil {
   168  		return
   169  	}
   170  	n := w.nbytes
   171  	if w.nbits&7 != 0 {
   172  		w.err = InternalError("writeBytes with unfinished bits")
   173  		return
   174  	}
   175  	for w.nbits != 0 {
   176  		w.bytes[n] = byte(w.bits)
   177  		w.bits >>= 8
   178  		w.nbits -= 8
   179  		n++
   180  	}
   181  	if n != 0 {
   182  		w.write(w.bytes[:n])
   183  	}
   184  	w.nbytes = 0
   185  	w.write(bytes)
   186  }
   187  
   188  // RFC 1951 3.2.7 specifies a special run-length encoding for specifying
   189  // the literal and offset lengths arrays (which are concatenated into a single
   190  // array).  This method generates that run-length encoding.
   191  //
   192  // The result is written into the codegen array, and the frequencies
   193  // of each code is written into the codegenFreq array.
   194  // Codes 0-15 are single byte codes. Codes 16-18 are followed by additional
   195  // information. Code badCode is an end marker
   196  //
   197  //  numLiterals      The number of literals in literalEncoding
   198  //  numOffsets       The number of offsets in offsetEncoding
   199  //  litenc, offenc   The literal and offset encoder to use
   200  func (w *huffmanBitWriter) generateCodegen(numLiterals int, numOffsets int, litEnc, offEnc *huffmanEncoder) {
   201  	for i := range w.codegenFreq {
   202  		w.codegenFreq[i] = 0
   203  	}
   204  	// Note that we are using codegen both as a temporary variable for holding
   205  	// a copy of the frequencies, and as the place where we put the result.
   206  	// This is fine because the output is always shorter than the input used
   207  	// so far.
   208  	codegen := w.codegen // cache
   209  	// Copy the concatenated code sizes to codegen. Put a marker at the end.
   210  	cgnl := codegen[:numLiterals]
   211  	for i := range cgnl {
   212  		cgnl[i] = uint8(litEnc.codes[i].len)
   213  	}
   214  
   215  	cgnl = codegen[numLiterals : numLiterals+numOffsets]
   216  	for i := range cgnl {
   217  		cgnl[i] = uint8(offEnc.codes[i].len)
   218  	}
   219  	codegen[numLiterals+numOffsets] = badCode
   220  
   221  	size := codegen[0]
   222  	count := 1
   223  	outIndex := 0
   224  	for inIndex := 1; size != badCode; inIndex++ {
   225  		// INVARIANT: We have seen "count" copies of size that have not yet
   226  		// had output generated for them.
   227  		nextSize := codegen[inIndex]
   228  		if nextSize == size {
   229  			count++
   230  			continue
   231  		}
   232  		// We need to generate codegen indicating "count" of size.
   233  		if size != 0 {
   234  			codegen[outIndex] = size
   235  			outIndex++
   236  			w.codegenFreq[size]++
   237  			count--
   238  			for count >= 3 {
   239  				n := 6
   240  				if n > count {
   241  					n = count
   242  				}
   243  				codegen[outIndex] = 16
   244  				outIndex++
   245  				codegen[outIndex] = uint8(n - 3)
   246  				outIndex++
   247  				w.codegenFreq[16]++
   248  				count -= n
   249  			}
   250  		} else {
   251  			for count >= 11 {
   252  				n := 138
   253  				if n > count {
   254  					n = count
   255  				}
   256  				codegen[outIndex] = 18
   257  				outIndex++
   258  				codegen[outIndex] = uint8(n - 11)
   259  				outIndex++
   260  				w.codegenFreq[18]++
   261  				count -= n
   262  			}
   263  			if count >= 3 {
   264  				// count >= 3 && count <= 10
   265  				codegen[outIndex] = 17
   266  				outIndex++
   267  				codegen[outIndex] = uint8(count - 3)
   268  				outIndex++
   269  				w.codegenFreq[17]++
   270  				count = 0
   271  			}
   272  		}
   273  		count--
   274  		for ; count >= 0; count-- {
   275  			codegen[outIndex] = size
   276  			outIndex++
   277  			w.codegenFreq[size]++
   278  		}
   279  		// Set up invariant for next time through the loop.
   280  		size = nextSize
   281  		count = 1
   282  	}
   283  	// Marker indicating the end of the codegen.
   284  	codegen[outIndex] = badCode
   285  }
   286  
   287  // dynamicSize returns the size of dynamically encoded data in bits.
   288  func (w *huffmanBitWriter) dynamicSize(litEnc, offEnc *huffmanEncoder, extraBits int) (size, numCodegens int) {
   289  	numCodegens = len(w.codegenFreq)
   290  	for numCodegens > 4 && w.codegenFreq[codegenOrder[numCodegens-1]] == 0 {
   291  		numCodegens--
   292  	}
   293  	header := 3 + 5 + 5 + 4 + (3 * numCodegens) +
   294  		w.codegenEncoding.bitLength(w.codegenFreq[:]) +
   295  		int(w.codegenFreq[16])*2 +
   296  		int(w.codegenFreq[17])*3 +
   297  		int(w.codegenFreq[18])*7
   298  	size = header +
   299  		litEnc.bitLength(w.literalFreq) +
   300  		offEnc.bitLength(w.offsetFreq) +
   301  		extraBits
   302  
   303  	return size, numCodegens
   304  }
   305  
   306  // fixedSize returns the size of dynamically encoded data in bits.
   307  func (w *huffmanBitWriter) fixedSize(extraBits int) int {
   308  	return 3 +
   309  		fixedLiteralEncoding.bitLength(w.literalFreq) +
   310  		fixedOffsetEncoding.bitLength(w.offsetFreq) +
   311  		extraBits
   312  }
   313  
   314  // storedSize calculates the stored size, including header.
   315  // The function returns the size in bits and whether the block
   316  // fits inside a single block.
   317  func (w *huffmanBitWriter) storedSize(in []byte) (int, bool) {
   318  	if in == nil {
   319  		return 0, false
   320  	}
   321  	if len(in) <= maxStoreBlockSize {
   322  		return (len(in) + 5) * 8, true
   323  	}
   324  	return 0, false
   325  }
   326  
   327  func (w *huffmanBitWriter) writeCode(c hcode) {
   328  	if w.err != nil {
   329  		return
   330  	}
   331  	w.bits |= uint64(c.code) << w.nbits
   332  	w.nbits += uint(c.len)
   333  	if w.nbits >= 48 {
   334  		bits := w.bits
   335  		w.bits >>= 48
   336  		w.nbits -= 48
   337  		n := w.nbytes
   338  		bytes := w.bytes[n : n+6]
   339  		bytes[0] = byte(bits)
   340  		bytes[1] = byte(bits >> 8)
   341  		bytes[2] = byte(bits >> 16)
   342  		bytes[3] = byte(bits >> 24)
   343  		bytes[4] = byte(bits >> 32)
   344  		bytes[5] = byte(bits >> 40)
   345  		n += 6
   346  		if n >= bufferFlushSize {
   347  			w.write(w.bytes[:n])
   348  			n = 0
   349  		}
   350  		w.nbytes = n
   351  	}
   352  }
   353  
   354  // Write the header of a dynamic Huffman block to the output stream.
   355  //
   356  //  numLiterals  The number of literals specified in codegen
   357  //  numOffsets   The number of offsets specified in codegen
   358  //  numCodegens  The number of codegens used in codegen
   359  func (w *huffmanBitWriter) writeDynamicHeader(numLiterals int, numOffsets int, numCodegens int, isEof bool) {
   360  	if w.err != nil {
   361  		return
   362  	}
   363  	var firstBits int32 = 4
   364  	if isEof {
   365  		firstBits = 5
   366  	}
   367  	w.writeBits(firstBits, 3)
   368  	w.writeBits(int32(numLiterals-257), 5)
   369  	w.writeBits(int32(numOffsets-1), 5)
   370  	w.writeBits(int32(numCodegens-4), 4)
   371  
   372  	for i := 0; i < numCodegens; i++ {
   373  		value := uint(w.codegenEncoding.codes[codegenOrder[i]].len)
   374  		w.writeBits(int32(value), 3)
   375  	}
   376  
   377  	i := 0
   378  	for {
   379  		var codeWord int = int(w.codegen[i])
   380  		i++
   381  		if codeWord == badCode {
   382  			break
   383  		}
   384  		w.writeCode(w.codegenEncoding.codes[uint32(codeWord)])
   385  
   386  		switch codeWord {
   387  		case 16:
   388  			w.writeBits(int32(w.codegen[i]), 2)
   389  			i++
   390  			break
   391  		case 17:
   392  			w.writeBits(int32(w.codegen[i]), 3)
   393  			i++
   394  			break
   395  		case 18:
   396  			w.writeBits(int32(w.codegen[i]), 7)
   397  			i++
   398  			break
   399  		}
   400  	}
   401  }
   402  
   403  func (w *huffmanBitWriter) writeStoredHeader(length int, isEof bool) {
   404  	if w.err != nil {
   405  		return
   406  	}
   407  	var flag int32
   408  	if isEof {
   409  		flag = 1
   410  	}
   411  	w.writeBits(flag, 3)
   412  	w.flush()
   413  	w.writeBits(int32(length), 16)
   414  	w.writeBits(int32(^uint16(length)), 16)
   415  }
   416  
   417  func (w *huffmanBitWriter) writeFixedHeader(isEof bool) {
   418  	if w.err != nil {
   419  		return
   420  	}
   421  	// Indicate that we are a fixed Huffman block
   422  	var value int32 = 2
   423  	if isEof {
   424  		value = 3
   425  	}
   426  	w.writeBits(value, 3)
   427  }
   428  
   429  // writeBlock will write a block of tokens with the smallest encoding.
   430  // The original input can be supplied, and if the huffman encoded data
   431  // is larger than the original bytes, the data will be written as a
   432  // stored block.
   433  // If the input is nil, the tokens will always be Huffman encoded.
   434  func (w *huffmanBitWriter) writeBlock(tokens []token, eof bool, input []byte) {
   435  	if w.err != nil {
   436  		return
   437  	}
   438  
   439  	tokens = append(tokens, endBlockMarker)
   440  	numLiterals, numOffsets := w.indexTokens(tokens)
   441  
   442  	var extraBits int
   443  	storedSize, storable := w.storedSize(input)
   444  	if storable {
   445  		// We only bother calculating the costs of the extra bits required by
   446  		// the length of offset fields (which will be the same for both fixed
   447  		// and dynamic encoding), if we need to compare those two encodings
   448  		// against stored encoding.
   449  		for lengthCode := lengthCodesStart + 8; lengthCode < numLiterals; lengthCode++ {
   450  			// First eight length codes have extra size = 0.
   451  			extraBits += int(w.literalFreq[lengthCode]) * int(lengthExtraBits[lengthCode-lengthCodesStart])
   452  		}
   453  		for offsetCode := 4; offsetCode < numOffsets; offsetCode++ {
   454  			// First four offset codes have extra size = 0.
   455  			extraBits += int(w.offsetFreq[offsetCode]) * int(offsetExtraBits[offsetCode])
   456  		}
   457  	}
   458  
   459  	// Figure out smallest code.
   460  	// Fixed Huffman baseline.
   461  	var literalEncoding = fixedLiteralEncoding
   462  	var offsetEncoding = fixedOffsetEncoding
   463  	var size = w.fixedSize(extraBits)
   464  
   465  	// Dynamic Huffman?
   466  	var numCodegens int
   467  
   468  	// Generate codegen and codegenFrequencies, which indicates how to encode
   469  	// the literalEncoding and the offsetEncoding.
   470  	w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, w.offsetEncoding)
   471  	w.codegenEncoding.generate(w.codegenFreq[:], 7)
   472  	dynamicSize, numCodegens := w.dynamicSize(w.literalEncoding, w.offsetEncoding, extraBits)
   473  
   474  	if dynamicSize < size {
   475  		size = dynamicSize
   476  		literalEncoding = w.literalEncoding
   477  		offsetEncoding = w.offsetEncoding
   478  	}
   479  
   480  	// Stored bytes?
   481  	if storable && storedSize < size {
   482  		w.writeStoredHeader(len(input), eof)
   483  		w.writeBytes(input)
   484  		return
   485  	}
   486  
   487  	// Huffman.
   488  	if literalEncoding == fixedLiteralEncoding {
   489  		w.writeFixedHeader(eof)
   490  	} else {
   491  		w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
   492  	}
   493  
   494  	// Write the tokens.
   495  	w.writeTokens(tokens, literalEncoding.codes, offsetEncoding.codes)
   496  }
   497  
   498  // writeBlockDynamic encodes a block using a dynamic Huffman table.
   499  // This should be used if the symbols used have a disproportionate
   500  // histogram distribution.
   501  // If input is supplied and the compression savings are below 1/16th of the
   502  // input size the block is stored.
   503  func (w *huffmanBitWriter) writeBlockDynamic(tokens []token, eof bool, input []byte) {
   504  	if w.err != nil {
   505  		return
   506  	}
   507  
   508  	tokens = append(tokens, endBlockMarker)
   509  	numLiterals, numOffsets := w.indexTokens(tokens)
   510  
   511  	// Generate codegen and codegenFrequencies, which indicates how to encode
   512  	// the literalEncoding and the offsetEncoding.
   513  	w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, w.offsetEncoding)
   514  	w.codegenEncoding.generate(w.codegenFreq[:], 7)
   515  	size, numCodegens := w.dynamicSize(w.literalEncoding, w.offsetEncoding, 0)
   516  
   517  	// Store bytes, if we don't get a reasonable improvement.
   518  	if ssize, storable := w.storedSize(input); storable && ssize < (size+size>>4) {
   519  		w.writeStoredHeader(len(input), eof)
   520  		w.writeBytes(input)
   521  		return
   522  	}
   523  
   524  	// Write Huffman table.
   525  	w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
   526  
   527  	// Write the tokens.
   528  	w.writeTokens(tokens, w.literalEncoding.codes, w.offsetEncoding.codes)
   529  }
   530  
   531  // indexTokens indexes a slice of tokens, and updates
   532  // literalFreq and offsetFreq, and generates literalEncoding
   533  // and offsetEncoding.
   534  // The number of literal and offset tokens is returned.
   535  func (w *huffmanBitWriter) indexTokens(tokens []token) (numLiterals, numOffsets int) {
   536  	for i := range w.literalFreq {
   537  		w.literalFreq[i] = 0
   538  	}
   539  	for i := range w.offsetFreq {
   540  		w.offsetFreq[i] = 0
   541  	}
   542  
   543  	for _, t := range tokens {
   544  		if t < matchType {
   545  			w.literalFreq[t.literal()]++
   546  			continue
   547  		}
   548  		length := t.length()
   549  		offset := t.offset()
   550  		w.literalFreq[lengthCodesStart+lengthCode(length)]++
   551  		w.offsetFreq[offsetCode(offset)]++
   552  	}
   553  
   554  	// get the number of literals
   555  	numLiterals = len(w.literalFreq)
   556  	for w.literalFreq[numLiterals-1] == 0 {
   557  		numLiterals--
   558  	}
   559  	// get the number of offsets
   560  	numOffsets = len(w.offsetFreq)
   561  	for numOffsets > 0 && w.offsetFreq[numOffsets-1] == 0 {
   562  		numOffsets--
   563  	}
   564  	if numOffsets == 0 {
   565  		// We haven't found a single match. If we want to go with the dynamic encoding,
   566  		// we should count at least one offset to be sure that the offset huffman tree could be encoded.
   567  		w.offsetFreq[0] = 1
   568  		numOffsets = 1
   569  	}
   570  	w.literalEncoding.generate(w.literalFreq, 15)
   571  	w.offsetEncoding.generate(w.offsetFreq, 15)
   572  	return
   573  }
   574  
   575  // writeTokens writes a slice of tokens to the output.
   576  // codes for literal and offset encoding must be supplied.
   577  func (w *huffmanBitWriter) writeTokens(tokens []token, leCodes, oeCodes []hcode) {
   578  	if w.err != nil {
   579  		return
   580  	}
   581  	for _, t := range tokens {
   582  		if t < matchType {
   583  			w.writeCode(leCodes[t.literal()])
   584  			continue
   585  		}
   586  		// Write the length
   587  		length := t.length()
   588  		lengthCode := lengthCode(length)
   589  		w.writeCode(leCodes[lengthCode+lengthCodesStart])
   590  		extraLengthBits := uint(lengthExtraBits[lengthCode])
   591  		if extraLengthBits > 0 {
   592  			extraLength := int32(length - lengthBase[lengthCode])
   593  			w.writeBits(extraLength, extraLengthBits)
   594  		}
   595  		// Write the offset
   596  		offset := t.offset()
   597  		offsetCode := offsetCode(offset)
   598  		w.writeCode(oeCodes[offsetCode])
   599  		extraOffsetBits := uint(offsetExtraBits[offsetCode])
   600  		if extraOffsetBits > 0 {
   601  			extraOffset := int32(offset - offsetBase[offsetCode])
   602  			w.writeBits(extraOffset, extraOffsetBits)
   603  		}
   604  	}
   605  }
   606  
   607  // huffOffset is a static offset encoder used for huffman only encoding.
   608  // It can be reused since we will not be encoding offset values.
   609  var huffOffset *huffmanEncoder
   610  
   611  func init() {
   612  	offsetFreq := make([]int32, offsetCodeCount)
   613  	offsetFreq[0] = 1
   614  	huffOffset = newHuffmanEncoder(offsetCodeCount)
   615  	huffOffset.generate(offsetFreq, 15)
   616  }
   617  
   618  // writeBlockHuff encodes a block of bytes as either
   619  // Huffman encoded literals or uncompressed bytes if the
   620  // results only gains very little from compression.
   621  func (w *huffmanBitWriter) writeBlockHuff(eof bool, input []byte) {
   622  	if w.err != nil {
   623  		return
   624  	}
   625  
   626  	// Clear histogram
   627  	for i := range w.literalFreq {
   628  		w.literalFreq[i] = 0
   629  	}
   630  
   631  	// Add everything as literals
   632  	histogram(input, w.literalFreq)
   633  
   634  	w.literalFreq[endBlockMarker] = 1
   635  
   636  	const numLiterals = endBlockMarker + 1
   637  	w.offsetFreq[0] = 1
   638  	const numOffsets = 1
   639  
   640  	w.literalEncoding.generate(w.literalFreq, 15)
   641  
   642  	// Figure out smallest code.
   643  	// Always use dynamic Huffman or Store
   644  	var numCodegens int
   645  
   646  	// Generate codegen and codegenFrequencies, which indicates how to encode
   647  	// the literalEncoding and the offsetEncoding.
   648  	w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, huffOffset)
   649  	w.codegenEncoding.generate(w.codegenFreq[:], 7)
   650  	size, numCodegens := w.dynamicSize(w.literalEncoding, huffOffset, 0)
   651  
   652  	// Store bytes, if we don't get a reasonable improvement.
   653  	if ssize, storable := w.storedSize(input); storable && ssize < (size+size>>4) {
   654  		w.writeStoredHeader(len(input), eof)
   655  		w.writeBytes(input)
   656  		return
   657  	}
   658  
   659  	// Huffman.
   660  	w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
   661  	encoding := w.literalEncoding.codes[:257]
   662  	n := w.nbytes
   663  	for _, t := range input {
   664  		// Bitwriting inlined, ~30% speedup
   665  		c := encoding[t]
   666  		w.bits |= uint64(c.code) << w.nbits
   667  		w.nbits += uint(c.len)
   668  		if w.nbits < 48 {
   669  			continue
   670  		}
   671  		// Store 6 bytes
   672  		bits := w.bits
   673  		w.bits >>= 48
   674  		w.nbits -= 48
   675  		bytes := w.bytes[n : n+6]
   676  		bytes[0] = byte(bits)
   677  		bytes[1] = byte(bits >> 8)
   678  		bytes[2] = byte(bits >> 16)
   679  		bytes[3] = byte(bits >> 24)
   680  		bytes[4] = byte(bits >> 32)
   681  		bytes[5] = byte(bits >> 40)
   682  		n += 6
   683  		if n < bufferFlushSize {
   684  			continue
   685  		}
   686  		w.write(w.bytes[:n])
   687  		if w.err != nil {
   688  			return // Return early in the event of write failures
   689  		}
   690  		n = 0
   691  	}
   692  	w.nbytes = n
   693  	w.writeCode(encoding[endBlockMarker])
   694  }
   695  
   696  // histogram accumulates a histogram of b in h.
   697  //
   698  // len(h) must be >= 256, and h's elements must be all zeroes.
   699  func histogram(b []byte, h []int32) {
   700  	h = h[:256]
   701  	for _, t := range b {
   702  		h[t]++
   703  	}
   704  }
   705  

View as plain text