deflatefast.mx raw

   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: []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