deflatefast.go

     1  // Copyright 2016 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 "math"
     8  
     9  // This encoding algorithm, which prioritizes speed over output size, is
    10  // based on Snappy's LZ77-style encoder: github.com/golang/snappy
    11  
    12  const (
    13  	tableBits  = 14             // Bits used in the table.
    14  	tableSize  = 1 << tableBits // Size of the table.
    15  	tableMask  = tableSize - 1  // Mask for table indices. Redundant, but can eliminate bounds checks.
    16  	tableShift = 32 - tableBits // Right-shift to get the tableBits most significant bits of a uint32.
    17  
    18  	// Reset the buffer offset when reaching this.
    19  	// Offsets are stored between blocks as int32 values.
    20  	// Since the offset we are checking against is at the beginning
    21  	// of the buffer, we need to subtract the current and input
    22  	// buffer to not risk overflowing the int32.
    23  	bufferReset = math.MaxInt32 - maxStoreBlockSize*2
    24  )
    25  
    26  func load32(b []byte, i int32) uint32 {
    27  	b = b[i : i+4 : len(b)] // Help the compiler eliminate bounds checks on the next line.
    28  	return uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24
    29  }
    30  
    31  func load64(b []byte, i int32) uint64 {
    32  	b = b[i : i+8 : len(b)] // Help the compiler eliminate bounds checks on the next line.
    33  	return uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 |
    34  		uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56
    35  }
    36  
    37  func hash(u uint32) uint32 {
    38  	return (u * 0x1e35a7bd) >> tableShift
    39  }
    40  
    41  // These constants are defined by the Snappy implementation so that its
    42  // assembly implementation can fast-path some 16-bytes-at-a-time copies. They
    43  // aren't necessary in the pure Go implementation, as we don't use those same
    44  // optimizations, but using the same thresholds doesn't really hurt.
    45  const (
    46  	inputMargin            = 16 - 1
    47  	minNonLiteralBlockSize = 1 + 1 + inputMargin
    48  )
    49  
    50  type tableEntry struct {
    51  	val    uint32 // Value at destination
    52  	offset int32
    53  }
    54  
    55  // deflateFast maintains the table for matches,
    56  // and the previous byte block for cross block matching.
    57  type deflateFast struct {
    58  	table [tableSize]tableEntry
    59  	prev  []byte // Previous block, zero length if unknown.
    60  	cur   int32  // Current match offset.
    61  }
    62  
    63  func newDeflateFast() *deflateFast {
    64  	return &deflateFast{cur: maxStoreBlockSize, prev: make([]byte, 0, maxStoreBlockSize)}
    65  }
    66  
    67  // encode encodes a block given in src and appends tokens
    68  // to dst and returns the result.
    69  func (e *deflateFast) encode(dst []token, src []byte) []token {
    70  	// Ensure that e.cur doesn't wrap.
    71  	if e.cur >= bufferReset {
    72  		e.shiftOffsets()
    73  	}
    74  
    75  	// This check isn't in the Snappy implementation, but there, the caller
    76  	// instead of the callee handles this case.
    77  	if len(src) < minNonLiteralBlockSize {
    78  		e.cur += maxStoreBlockSize
    79  		e.prev = e.prev[:0]
    80  		return emitLiteral(dst, src)
    81  	}
    82  
    83  	// sLimit is when to stop looking for offset/length copies. The inputMargin
    84  	// lets us use a fast path for emitLiteral in the main loop, while we are
    85  	// looking for copies.
    86  	sLimit := int32(len(src) - inputMargin)
    87  
    88  	// nextEmit is where in src the next emitLiteral should start from.
    89  	nextEmit := int32(0)
    90  	s := int32(0)
    91  	cv := load32(src, s)
    92  	nextHash := hash(cv)
    93  
    94  	for {
    95  		// Copied from the C++ snappy implementation:
    96  		//
    97  		// Heuristic match skipping: If 32 bytes are scanned with no matches
    98  		// found, start looking only at every other byte. If 32 more bytes are
    99  		// scanned (or skipped), look at every third byte, etc.. When a match
   100  		// is found, immediately go back to looking at every byte. This is a
   101  		// small loss (~5% performance, ~0.1% density) for compressible data
   102  		// due to more bookkeeping, but for non-compressible data (such as
   103  		// JPEG) it's a huge win since the compressor quickly "realizes" the
   104  		// data is incompressible and doesn't bother looking for matches
   105  		// everywhere.
   106  		//
   107  		// The "skip" variable keeps track of how many bytes there are since
   108  		// the last match; dividing it by 32 (ie. right-shifting by five) gives
   109  		// the number of bytes to move ahead for each iteration.
   110  		skip := int32(32)
   111  
   112  		nextS := s
   113  		var candidate tableEntry
   114  		for {
   115  			s = nextS
   116  			bytesBetweenHashLookups := skip >> 5
   117  			nextS = s + bytesBetweenHashLookups
   118  			skip += bytesBetweenHashLookups
   119  			if nextS > sLimit {
   120  				goto emitRemainder
   121  			}
   122  			candidate = e.table[nextHash&tableMask]
   123  			now := load32(src, nextS)
   124  			e.table[nextHash&tableMask] = tableEntry{offset: s + e.cur, val: cv}
   125  			nextHash = hash(now)
   126  
   127  			offset := s - (candidate.offset - e.cur)
   128  			if offset > maxMatchOffset || cv != candidate.val {
   129  				// Out of range or not matched.
   130  				cv = now
   131  				continue
   132  			}
   133  			break
   134  		}
   135  
   136  		// A 4-byte match has been found. We'll later see if more than 4 bytes
   137  		// match. But, prior to the match, src[nextEmit:s] are unmatched. Emit
   138  		// them as literal bytes.
   139  		dst = emitLiteral(dst, src[nextEmit:s])
   140  
   141  		// Call emitCopy, and then see if another emitCopy could be our next
   142  		// move. Repeat until we find no match for the input immediately after
   143  		// what was consumed by the last emitCopy call.
   144  		//
   145  		// If we exit this loop normally then we need to call emitLiteral next,
   146  		// though we don't yet know how big the literal will be. We handle that
   147  		// by proceeding to the next iteration of the main loop. We also can
   148  		// exit this loop via goto if we get close to exhausting the input.
   149  		for {
   150  			// Invariant: we have a 4-byte match at s, and no need to emit any
   151  			// literal bytes prior to s.
   152  
   153  			// Extend the 4-byte match as long as possible.
   154  			//
   155  			s += 4
   156  			t := candidate.offset - e.cur + 4
   157  			l := e.matchLen(s, t, src)
   158  
   159  			// matchToken is flate's equivalent of Snappy's emitCopy. (length,offset)
   160  			dst = append(dst, matchToken(uint32(l+4-baseMatchLength), uint32(s-t-baseMatchOffset)))
   161  			s += l
   162  			nextEmit = s
   163  			if s >= sLimit {
   164  				goto emitRemainder
   165  			}
   166  
   167  			// We could immediately start working at s now, but to improve
   168  			// compression we first update the hash table at s-1 and at s. If
   169  			// another emitCopy is not our next move, also calculate nextHash
   170  			// at s+1. At least on GOARCH=amd64, these three hash calculations
   171  			// are faster as one load64 call (with some shifts) instead of
   172  			// three load32 calls.
   173  			x := load64(src, s-1)
   174  			prevHash := hash(uint32(x))
   175  			e.table[prevHash&tableMask] = tableEntry{offset: e.cur + s - 1, val: uint32(x)}
   176  			x >>= 8
   177  			currHash := hash(uint32(x))
   178  			candidate = e.table[currHash&tableMask]
   179  			e.table[currHash&tableMask] = tableEntry{offset: e.cur + s, val: uint32(x)}
   180  
   181  			offset := s - (candidate.offset - e.cur)
   182  			if offset > maxMatchOffset || uint32(x) != candidate.val {
   183  				cv = uint32(x >> 8)
   184  				nextHash = hash(cv)
   185  				s++
   186  				break
   187  			}
   188  		}
   189  	}
   190  
   191  emitRemainder:
   192  	if int(nextEmit) < len(src) {
   193  		dst = emitLiteral(dst, src[nextEmit:])
   194  	}
   195  	e.cur += int32(len(src))
   196  	e.prev = e.prev[:len(src)]
   197  	copy(e.prev, src)
   198  	return dst
   199  }
   200  
   201  func emitLiteral(dst []token, lit []byte) []token {
   202  	for _, v := range lit {
   203  		dst = append(dst, literalToken(uint32(v)))
   204  	}
   205  	return dst
   206  }
   207  
   208  // matchLen returns the match length between src[s:] and src[t:].
   209  // t can be negative to indicate the match is starting in e.prev.
   210  // We assume that src[s-4:s] and src[t-4:t] already match.
   211  func (e *deflateFast) matchLen(s, t int32, src []byte) int32 {
   212  	s1 := int(s) + maxMatchLength - 4
   213  	if s1 > len(src) {
   214  		s1 = len(src)
   215  	}
   216  
   217  	// If we are inside the current block
   218  	if t >= 0 {
   219  		b := src[t:]
   220  		a := src[s:s1]
   221  		b = b[:len(a)]
   222  		// Extend the match to be as long as possible.
   223  		for i := range a {
   224  			if a[i] != b[i] {
   225  				return int32(i)
   226  			}
   227  		}
   228  		return int32(len(a))
   229  	}
   230  
   231  	// We found a match in the previous block.
   232  	tp := int32(len(e.prev)) + t
   233  	if tp < 0 {
   234  		return 0
   235  	}
   236  
   237  	// Extend the match to be as long as possible.
   238  	a := src[s:s1]
   239  	b := e.prev[tp:]
   240  	if len(b) > len(a) {
   241  		b = b[:len(a)]
   242  	}
   243  	a = a[:len(b)]
   244  	for i := range b {
   245  		if a[i] != b[i] {
   246  			return int32(i)
   247  		}
   248  	}
   249  
   250  	// If we reached our limit, we matched everything we are
   251  	// allowed to in the previous block and we return.
   252  	n := int32(len(b))
   253  	if int(s+n) == s1 {
   254  		return n
   255  	}
   256  
   257  	// Continue looking for more matches in the current block.
   258  	a = src[s+n : s1]
   259  	b = src[:len(a)]
   260  	for i := range a {
   261  		if a[i] != b[i] {
   262  			return int32(i) + n
   263  		}
   264  	}
   265  	return int32(len(a)) + n
   266  }
   267  
   268  // Reset resets the encoding history.
   269  // This ensures that no matches are made to the previous block.
   270  func (e *deflateFast) reset() {
   271  	e.prev = e.prev[:0]
   272  	// Bump the offset, so all matches will fail distance check.
   273  	// Nothing should be >= e.cur in the table.
   274  	e.cur += maxMatchOffset
   275  
   276  	// Protect against e.cur wraparound.
   277  	if e.cur >= bufferReset {
   278  		e.shiftOffsets()
   279  	}
   280  }
   281  
   282  // shiftOffsets will shift down all match offset.
   283  // This is only called in rare situations to prevent integer overflow.
   284  //
   285  // See https://golang.org/issue/18636 and https://github.com/golang/go/issues/34121.
   286  func (e *deflateFast) shiftOffsets() {
   287  	if len(e.prev) == 0 {
   288  		// We have no history; just clear the table.
   289  		clear(e.table[:])
   290  		e.cur = maxMatchOffset + 1
   291  		return
   292  	}
   293  
   294  	// Shift down everything in the table that isn't already too far away.
   295  	for i := range e.table[:] {
   296  		v := e.table[i].offset - e.cur + maxMatchOffset + 1
   297  		if v < 0 {
   298  			// We want to reset e.cur to maxMatchOffset + 1, so we need to shift
   299  			// all table entries down by (e.cur - (maxMatchOffset + 1)).
   300  			// Because we ignore matches > maxMatchOffset, we can cap
   301  			// any negative offsets at 0.
   302  			v = 0
   303  		}
   304  		e.table[i].offset = v
   305  	}
   306  	e.cur = maxMatchOffset + 1
   307  }
   308
View as plain text