...
Run Format

Source file src/compress/flate/huffman_bit_writer.go

Documentation: compress/flate

  // Copyright 2009 The Go Authors. All rights reserved.
  // Use of this source code is governed by a BSD-style
  // license that can be found in the LICENSE file.
  
  package flate
  
  import (
  	"io"
  )
  
  const (
  	// The largest offset code.
  	offsetCodeCount = 30
  
  	// The special code used to mark the end of a block.
  	endBlockMarker = 256
  
  	// The first length code.
  	lengthCodesStart = 257
  
  	// The number of codegen codes.
  	codegenCodeCount = 19
  	badCode          = 255
  
  	// bufferFlushSize indicates the buffer size
  	// after which bytes are flushed to the writer.
  	// Should preferably be a multiple of 6, since
  	// we accumulate 6 bytes between writes to the buffer.
  	bufferFlushSize = 240
  
  	// bufferSize is the actual output byte buffer size.
  	// It must have additional headroom for a flush
  	// which can contain up to 8 bytes.
  	bufferSize = bufferFlushSize + 8
  )
  
  // The number of extra bits needed by length code X - LENGTH_CODES_START.
  var lengthExtraBits = []int8{
  	/* 257 */ 0, 0, 0,
  	/* 260 */ 0, 0, 0, 0, 0, 1, 1, 1, 1, 2,
  	/* 270 */ 2, 2, 2, 3, 3, 3, 3, 4, 4, 4,
  	/* 280 */ 4, 5, 5, 5, 5, 0,
  }
  
  // The length indicated by length code X - LENGTH_CODES_START.
  var lengthBase = []uint32{
  	0, 1, 2, 3, 4, 5, 6, 7, 8, 10,
  	12, 14, 16, 20, 24, 28, 32, 40, 48, 56,
  	64, 80, 96, 112, 128, 160, 192, 224, 255,
  }
  
  // offset code word extra bits.
  var offsetExtraBits = []int8{
  	0, 0, 0, 0, 1, 1, 2, 2, 3, 3,
  	4, 4, 5, 5, 6, 6, 7, 7, 8, 8,
  	9, 9, 10, 10, 11, 11, 12, 12, 13, 13,
  	/* extended window */
  	14, 14, 15, 15, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20,
  }
  
  var offsetBase = []uint32{
  	/* normal deflate */
  	0x000000, 0x000001, 0x000002, 0x000003, 0x000004,
  	0x000006, 0x000008, 0x00000c, 0x000010, 0x000018,
  	0x000020, 0x000030, 0x000040, 0x000060, 0x000080,
  	0x0000c0, 0x000100, 0x000180, 0x000200, 0x000300,
  	0x000400, 0x000600, 0x000800, 0x000c00, 0x001000,
  	0x001800, 0x002000, 0x003000, 0x004000, 0x006000,
  
  	/* extended window */
  	0x008000, 0x00c000, 0x010000, 0x018000, 0x020000,
  	0x030000, 0x040000, 0x060000, 0x080000, 0x0c0000,
  	0x100000, 0x180000, 0x200000, 0x300000,
  }
  
  // The odd order in which the codegen code sizes are written.
  var codegenOrder = []uint32{16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}
  
  type huffmanBitWriter struct {
  	// writer is the underlying writer.
  	// Do not use it directly; use the write method, which ensures
  	// that Write errors are sticky.
  	writer io.Writer
  
  	// Data waiting to be written is bytes[0:nbytes]
  	// and then the low nbits of bits.
  	bits            uint64
  	nbits           uint
  	bytes           [bufferSize]byte
  	codegenFreq     [codegenCodeCount]int32
  	nbytes          int
  	literalFreq     []int32
  	offsetFreq      []int32
  	codegen         []uint8
  	literalEncoding *huffmanEncoder
  	offsetEncoding  *huffmanEncoder
  	codegenEncoding *huffmanEncoder
  	err             error
  }
  
  func newHuffmanBitWriter(w io.Writer) *huffmanBitWriter {
  	return &huffmanBitWriter{
  		writer:          w,
  		literalFreq:     make([]int32, maxNumLit),
  		offsetFreq:      make([]int32, offsetCodeCount),
  		codegen:         make([]uint8, maxNumLit+offsetCodeCount+1),
  		literalEncoding: newHuffmanEncoder(maxNumLit),
  		codegenEncoding: newHuffmanEncoder(codegenCodeCount),
  		offsetEncoding:  newHuffmanEncoder(offsetCodeCount),
  	}
  }
  
  func (w *huffmanBitWriter) reset(writer io.Writer) {
  	w.writer = writer
  	w.bits, w.nbits, w.nbytes, w.err = 0, 0, 0, nil
  	w.bytes = [bufferSize]byte{}
  }
  
  func (w *huffmanBitWriter) flush() {
  	if w.err != nil {
  		w.nbits = 0
  		return
  	}
  	n := w.nbytes
  	for w.nbits != 0 {
  		w.bytes[n] = byte(w.bits)
  		w.bits >>= 8
  		if w.nbits > 8 { // Avoid underflow
  			w.nbits -= 8
  		} else {
  			w.nbits = 0
  		}
  		n++
  	}
  	w.bits = 0
  	w.write(w.bytes[:n])
  	w.nbytes = 0
  }
  
  func (w *huffmanBitWriter) write(b []byte) {
  	if w.err != nil {
  		return
  	}
  	_, w.err = w.writer.Write(b)
  }
  
  func (w *huffmanBitWriter) writeBits(b int32, nb uint) {
  	if w.err != nil {
  		return
  	}
  	w.bits |= uint64(b) << w.nbits
  	w.nbits += nb
  	if w.nbits >= 48 {
  		bits := w.bits
  		w.bits >>= 48
  		w.nbits -= 48
  		n := w.nbytes
  		bytes := w.bytes[n : n+6]
  		bytes[0] = byte(bits)
  		bytes[1] = byte(bits >> 8)
  		bytes[2] = byte(bits >> 16)
  		bytes[3] = byte(bits >> 24)
  		bytes[4] = byte(bits >> 32)
  		bytes[5] = byte(bits >> 40)
  		n += 6
  		if n >= bufferFlushSize {
  			w.write(w.bytes[:n])
  			n = 0
  		}
  		w.nbytes = n
  	}
  }
  
  func (w *huffmanBitWriter) writeBytes(bytes []byte) {
  	if w.err != nil {
  		return
  	}
  	n := w.nbytes
  	if w.nbits&7 != 0 {
  		w.err = InternalError("writeBytes with unfinished bits")
  		return
  	}
  	for w.nbits != 0 {
  		w.bytes[n] = byte(w.bits)
  		w.bits >>= 8
  		w.nbits -= 8
  		n++
  	}
  	if n != 0 {
  		w.write(w.bytes[:n])
  	}
  	w.nbytes = 0
  	w.write(bytes)
  }
  
  // RFC 1951 3.2.7 specifies a special run-length encoding for specifying
  // the literal and offset lengths arrays (which are concatenated into a single
  // array).  This method generates that run-length encoding.
  //
  // The result is written into the codegen array, and the frequencies
  // of each code is written into the codegenFreq array.
  // Codes 0-15 are single byte codes. Codes 16-18 are followed by additional
  // information. Code badCode is an end marker
  //
  //  numLiterals      The number of literals in literalEncoding
  //  numOffsets       The number of offsets in offsetEncoding
  //  litenc, offenc   The literal and offset encoder to use
  func (w *huffmanBitWriter) generateCodegen(numLiterals int, numOffsets int, litEnc, offEnc *huffmanEncoder) {
  	for i := range w.codegenFreq {
  		w.codegenFreq[i] = 0
  	}
  	// Note that we are using codegen both as a temporary variable for holding
  	// a copy of the frequencies, and as the place where we put the result.
  	// This is fine because the output is always shorter than the input used
  	// so far.
  	codegen := w.codegen // cache
  	// Copy the concatenated code sizes to codegen. Put a marker at the end.
  	cgnl := codegen[:numLiterals]
  	for i := range cgnl {
  		cgnl[i] = uint8(litEnc.codes[i].len)
  	}
  
  	cgnl = codegen[numLiterals : numLiterals+numOffsets]
  	for i := range cgnl {
  		cgnl[i] = uint8(offEnc.codes[i].len)
  	}
  	codegen[numLiterals+numOffsets] = badCode
  
  	size := codegen[0]
  	count := 1
  	outIndex := 0
  	for inIndex := 1; size != badCode; inIndex++ {
  		// INVARIANT: We have seen "count" copies of size that have not yet
  		// had output generated for them.
  		nextSize := codegen[inIndex]
  		if nextSize == size {
  			count++
  			continue
  		}
  		// We need to generate codegen indicating "count" of size.
  		if size != 0 {
  			codegen[outIndex] = size
  			outIndex++
  			w.codegenFreq[size]++
  			count--
  			for count >= 3 {
  				n := 6
  				if n > count {
  					n = count
  				}
  				codegen[outIndex] = 16
  				outIndex++
  				codegen[outIndex] = uint8(n - 3)
  				outIndex++
  				w.codegenFreq[16]++
  				count -= n
  			}
  		} else {
  			for count >= 11 {
  				n := 138
  				if n > count {
  					n = count
  				}
  				codegen[outIndex] = 18
  				outIndex++
  				codegen[outIndex] = uint8(n - 11)
  				outIndex++
  				w.codegenFreq[18]++
  				count -= n
  			}
  			if count >= 3 {
  				// count >= 3 && count <= 10
  				codegen[outIndex] = 17
  				outIndex++
  				codegen[outIndex] = uint8(count - 3)
  				outIndex++
  				w.codegenFreq[17]++
  				count = 0
  			}
  		}
  		count--
  		for ; count >= 0; count-- {
  			codegen[outIndex] = size
  			outIndex++
  			w.codegenFreq[size]++
  		}
  		// Set up invariant for next time through the loop.
  		size = nextSize
  		count = 1
  	}
  	// Marker indicating the end of the codegen.
  	codegen[outIndex] = badCode
  }
  
  // dynamicSize returns the size of dynamically encoded data in bits.
  func (w *huffmanBitWriter) dynamicSize(litEnc, offEnc *huffmanEncoder, extraBits int) (size, numCodegens int) {
  	numCodegens = len(w.codegenFreq)
  	for numCodegens > 4 && w.codegenFreq[codegenOrder[numCodegens-1]] == 0 {
  		numCodegens--
  	}
  	header := 3 + 5 + 5 + 4 + (3 * numCodegens) +
  		w.codegenEncoding.bitLength(w.codegenFreq[:]) +
  		int(w.codegenFreq[16])*2 +
  		int(w.codegenFreq[17])*3 +
  		int(w.codegenFreq[18])*7
  	size = header +
  		litEnc.bitLength(w.literalFreq) +
  		offEnc.bitLength(w.offsetFreq) +
  		extraBits
  
  	return size, numCodegens
  }
  
  // fixedSize returns the size of dynamically encoded data in bits.
  func (w *huffmanBitWriter) fixedSize(extraBits int) int {
  	return 3 +
  		fixedLiteralEncoding.bitLength(w.literalFreq) +
  		fixedOffsetEncoding.bitLength(w.offsetFreq) +
  		extraBits
  }
  
  // storedSize calculates the stored size, including header.
  // The function returns the size in bits and whether the block
  // fits inside a single block.
  func (w *huffmanBitWriter) storedSize(in []byte) (int, bool) {
  	if in == nil {
  		return 0, false
  	}
  	if len(in) <= maxStoreBlockSize {
  		return (len(in) + 5) * 8, true
  	}
  	return 0, false
  }
  
  func (w *huffmanBitWriter) writeCode(c hcode) {
  	if w.err != nil {
  		return
  	}
  	w.bits |= uint64(c.code) << w.nbits
  	w.nbits += uint(c.len)
  	if w.nbits >= 48 {
  		bits := w.bits
  		w.bits >>= 48
  		w.nbits -= 48
  		n := w.nbytes
  		bytes := w.bytes[n : n+6]
  		bytes[0] = byte(bits)
  		bytes[1] = byte(bits >> 8)
  		bytes[2] = byte(bits >> 16)
  		bytes[3] = byte(bits >> 24)
  		bytes[4] = byte(bits >> 32)
  		bytes[5] = byte(bits >> 40)
  		n += 6
  		if n >= bufferFlushSize {
  			w.write(w.bytes[:n])
  			n = 0
  		}
  		w.nbytes = n
  	}
  }
  
  // Write the header of a dynamic Huffman block to the output stream.
  //
  //  numLiterals  The number of literals specified in codegen
  //  numOffsets   The number of offsets specified in codegen
  //  numCodegens  The number of codegens used in codegen
  func (w *huffmanBitWriter) writeDynamicHeader(numLiterals int, numOffsets int, numCodegens int, isEof bool) {
  	if w.err != nil {
  		return
  	}
  	var firstBits int32 = 4
  	if isEof {
  		firstBits = 5
  	}
  	w.writeBits(firstBits, 3)
  	w.writeBits(int32(numLiterals-257), 5)
  	w.writeBits(int32(numOffsets-1), 5)
  	w.writeBits(int32(numCodegens-4), 4)
  
  	for i := 0; i < numCodegens; i++ {
  		value := uint(w.codegenEncoding.codes[codegenOrder[i]].len)
  		w.writeBits(int32(value), 3)
  	}
  
  	i := 0
  	for {
  		var codeWord int = int(w.codegen[i])
  		i++
  		if codeWord == badCode {
  			break
  		}
  		w.writeCode(w.codegenEncoding.codes[uint32(codeWord)])
  
  		switch codeWord {
  		case 16:
  			w.writeBits(int32(w.codegen[i]), 2)
  			i++
  			break
  		case 17:
  			w.writeBits(int32(w.codegen[i]), 3)
  			i++
  			break
  		case 18:
  			w.writeBits(int32(w.codegen[i]), 7)
  			i++
  			break
  		}
  	}
  }
  
  func (w *huffmanBitWriter) writeStoredHeader(length int, isEof bool) {
  	if w.err != nil {
  		return
  	}
  	var flag int32
  	if isEof {
  		flag = 1
  	}
  	w.writeBits(flag, 3)
  	w.flush()
  	w.writeBits(int32(length), 16)
  	w.writeBits(int32(^uint16(length)), 16)
  }
  
  func (w *huffmanBitWriter) writeFixedHeader(isEof bool) {
  	if w.err != nil {
  		return
  	}
  	// Indicate that we are a fixed Huffman block
  	var value int32 = 2
  	if isEof {
  		value = 3
  	}
  	w.writeBits(value, 3)
  }
  
  // writeBlock will write a block of tokens with the smallest encoding.
  // The original input can be supplied, and if the huffman encoded data
  // is larger than the original bytes, the data will be written as a
  // stored block.
  // If the input is nil, the tokens will always be Huffman encoded.
  func (w *huffmanBitWriter) writeBlock(tokens []token, eof bool, input []byte) {
  	if w.err != nil {
  		return
  	}
  
  	tokens = append(tokens, endBlockMarker)
  	numLiterals, numOffsets := w.indexTokens(tokens)
  
  	var extraBits int
  	storedSize, storable := w.storedSize(input)
  	if storable {
  		// We only bother calculating the costs of the extra bits required by
  		// the length of offset fields (which will be the same for both fixed
  		// and dynamic encoding), if we need to compare those two encodings
  		// against stored encoding.
  		for lengthCode := lengthCodesStart + 8; lengthCode < numLiterals; lengthCode++ {
  			// First eight length codes have extra size = 0.
  			extraBits += int(w.literalFreq[lengthCode]) * int(lengthExtraBits[lengthCode-lengthCodesStart])
  		}
  		for offsetCode := 4; offsetCode < numOffsets; offsetCode++ {
  			// First four offset codes have extra size = 0.
  			extraBits += int(w.offsetFreq[offsetCode]) * int(offsetExtraBits[offsetCode])
  		}
  	}
  
  	// Figure out smallest code.
  	// Fixed Huffman baseline.
  	var literalEncoding = fixedLiteralEncoding
  	var offsetEncoding = fixedOffsetEncoding
  	var size = w.fixedSize(extraBits)
  
  	// Dynamic Huffman?
  	var numCodegens int
  
  	// Generate codegen and codegenFrequencies, which indicates how to encode
  	// the literalEncoding and the offsetEncoding.
  	w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, w.offsetEncoding)
  	w.codegenEncoding.generate(w.codegenFreq[:], 7)
  	dynamicSize, numCodegens := w.dynamicSize(w.literalEncoding, w.offsetEncoding, extraBits)
  
  	if dynamicSize < size {
  		size = dynamicSize
  		literalEncoding = w.literalEncoding
  		offsetEncoding = w.offsetEncoding
  	}
  
  	// Stored bytes?
  	if storable && storedSize < size {
  		w.writeStoredHeader(len(input), eof)
  		w.writeBytes(input)
  		return
  	}
  
  	// Huffman.
  	if literalEncoding == fixedLiteralEncoding {
  		w.writeFixedHeader(eof)
  	} else {
  		w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
  	}
  
  	// Write the tokens.
  	w.writeTokens(tokens, literalEncoding.codes, offsetEncoding.codes)
  }
  
  // writeBlockDynamic encodes a block using a dynamic Huffman table.
  // This should be used if the symbols used have a disproportionate
  // histogram distribution.
  // If input is supplied and the compression savings are below 1/16th of the
  // input size the block is stored.
  func (w *huffmanBitWriter) writeBlockDynamic(tokens []token, eof bool, input []byte) {
  	if w.err != nil {
  		return
  	}
  
  	tokens = append(tokens, endBlockMarker)
  	numLiterals, numOffsets := w.indexTokens(tokens)
  
  	// Generate codegen and codegenFrequencies, which indicates how to encode
  	// the literalEncoding and the offsetEncoding.
  	w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, w.offsetEncoding)
  	w.codegenEncoding.generate(w.codegenFreq[:], 7)
  	size, numCodegens := w.dynamicSize(w.literalEncoding, w.offsetEncoding, 0)
  
  	// Store bytes, if we don't get a reasonable improvement.
  	if ssize, storable := w.storedSize(input); storable && ssize < (size+size>>4) {
  		w.writeStoredHeader(len(input), eof)
  		w.writeBytes(input)
  		return
  	}
  
  	// Write Huffman table.
  	w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
  
  	// Write the tokens.
  	w.writeTokens(tokens, w.literalEncoding.codes, w.offsetEncoding.codes)
  }
  
  // indexTokens indexes a slice of tokens, and updates
  // literalFreq and offsetFreq, and generates literalEncoding
  // and offsetEncoding.
  // The number of literal and offset tokens is returned.
  func (w *huffmanBitWriter) indexTokens(tokens []token) (numLiterals, numOffsets int) {
  	for i := range w.literalFreq {
  		w.literalFreq[i] = 0
  	}
  	for i := range w.offsetFreq {
  		w.offsetFreq[i] = 0
  	}
  
  	for _, t := range tokens {
  		if t < matchType {
  			w.literalFreq[t.literal()]++
  			continue
  		}
  		length := t.length()
  		offset := t.offset()
  		w.literalFreq[lengthCodesStart+lengthCode(length)]++
  		w.offsetFreq[offsetCode(offset)]++
  	}
  
  	// get the number of literals
  	numLiterals = len(w.literalFreq)
  	for w.literalFreq[numLiterals-1] == 0 {
  		numLiterals--
  	}
  	// get the number of offsets
  	numOffsets = len(w.offsetFreq)
  	for numOffsets > 0 && w.offsetFreq[numOffsets-1] == 0 {
  		numOffsets--
  	}
  	if numOffsets == 0 {
  		// We haven't found a single match. If we want to go with the dynamic encoding,
  		// we should count at least one offset to be sure that the offset huffman tree could be encoded.
  		w.offsetFreq[0] = 1
  		numOffsets = 1
  	}
  	w.literalEncoding.generate(w.literalFreq, 15)
  	w.offsetEncoding.generate(w.offsetFreq, 15)
  	return
  }
  
  // writeTokens writes a slice of tokens to the output.
  // codes for literal and offset encoding must be supplied.
  func (w *huffmanBitWriter) writeTokens(tokens []token, leCodes, oeCodes []hcode) {
  	if w.err != nil {
  		return
  	}
  	for _, t := range tokens {
  		if t < matchType {
  			w.writeCode(leCodes[t.literal()])
  			continue
  		}
  		// Write the length
  		length := t.length()
  		lengthCode := lengthCode(length)
  		w.writeCode(leCodes[lengthCode+lengthCodesStart])
  		extraLengthBits := uint(lengthExtraBits[lengthCode])
  		if extraLengthBits > 0 {
  			extraLength := int32(length - lengthBase[lengthCode])
  			w.writeBits(extraLength, extraLengthBits)
  		}
  		// Write the offset
  		offset := t.offset()
  		offsetCode := offsetCode(offset)
  		w.writeCode(oeCodes[offsetCode])
  		extraOffsetBits := uint(offsetExtraBits[offsetCode])
  		if extraOffsetBits > 0 {
  			extraOffset := int32(offset - offsetBase[offsetCode])
  			w.writeBits(extraOffset, extraOffsetBits)
  		}
  	}
  }
  
  // huffOffset is a static offset encoder used for huffman only encoding.
  // It can be reused since we will not be encoding offset values.
  var huffOffset *huffmanEncoder
  
  func init() {
  	w := newHuffmanBitWriter(nil)
  	w.offsetFreq[0] = 1
  	huffOffset = newHuffmanEncoder(offsetCodeCount)
  	huffOffset.generate(w.offsetFreq, 15)
  }
  
  // writeBlockHuff encodes a block of bytes as either
  // Huffman encoded literals or uncompressed bytes if the
  // results only gains very little from compression.
  func (w *huffmanBitWriter) writeBlockHuff(eof bool, input []byte) {
  	if w.err != nil {
  		return
  	}
  
  	// Clear histogram
  	for i := range w.literalFreq {
  		w.literalFreq[i] = 0
  	}
  
  	// Add everything as literals
  	histogram(input, w.literalFreq)
  
  	w.literalFreq[endBlockMarker] = 1
  
  	const numLiterals = endBlockMarker + 1
  	const numOffsets = 1
  
  	w.literalEncoding.generate(w.literalFreq, 15)
  
  	// Figure out smallest code.
  	// Always use dynamic Huffman or Store
  	var numCodegens int
  
  	// Generate codegen and codegenFrequencies, which indicates how to encode
  	// the literalEncoding and the offsetEncoding.
  	w.generateCodegen(numLiterals, numOffsets, w.literalEncoding, huffOffset)
  	w.codegenEncoding.generate(w.codegenFreq[:], 7)
  	size, numCodegens := w.dynamicSize(w.literalEncoding, huffOffset, 0)
  
  	// Store bytes, if we don't get a reasonable improvement.
  	if ssize, storable := w.storedSize(input); storable && ssize < (size+size>>4) {
  		w.writeStoredHeader(len(input), eof)
  		w.writeBytes(input)
  		return
  	}
  
  	// Huffman.
  	w.writeDynamicHeader(numLiterals, numOffsets, numCodegens, eof)
  	encoding := w.literalEncoding.codes[:257]
  	n := w.nbytes
  	for _, t := range input {
  		// Bitwriting inlined, ~30% speedup
  		c := encoding[t]
  		w.bits |= uint64(c.code) << w.nbits
  		w.nbits += uint(c.len)
  		if w.nbits < 48 {
  			continue
  		}
  		// Store 6 bytes
  		bits := w.bits
  		w.bits >>= 48
  		w.nbits -= 48
  		bytes := w.bytes[n : n+6]
  		bytes[0] = byte(bits)
  		bytes[1] = byte(bits >> 8)
  		bytes[2] = byte(bits >> 16)
  		bytes[3] = byte(bits >> 24)
  		bytes[4] = byte(bits >> 32)
  		bytes[5] = byte(bits >> 40)
  		n += 6
  		if n < bufferFlushSize {
  			continue
  		}
  		w.write(w.bytes[:n])
  		if w.err != nil {
  			return // Return early in the event of write failures
  		}
  		n = 0
  	}
  	w.nbytes = n
  	w.writeCode(encoding[endBlockMarker])
  }
  
  // histogram accumulates a histogram of b in h.
  //
  // len(h) must be >= 256, and h's elements must be all zeroes.
  func histogram(b []byte, h []int32) {
  	h = h[:256]
  	for _, t := range b {
  		h[t]++
  	}
  }
  

View as plain text