blob: 5755cda246007db751ca12a7c249b96b966a3b1c [file] [log] [blame]
Takahiro Suzukid7bf8202020-12-17 20:21:59 +09001package lz4
2
3import (
4 "encoding/binary"
5 "fmt"
6 "math/bits"
7)
8
9// blockHash hashes the lower 6 bytes into a value < htSize.
10func blockHash(x uint64) uint32 {
11 const prime6bytes = 227718039650203
12 return uint32(((x << (64 - 48)) * prime6bytes) >> (64 - hashLog))
13}
14
15// CompressBlockBound returns the maximum size of a given buffer of size n, when not compressible.
16func CompressBlockBound(n int) int {
17 return n + n/255 + 16
18}
19
20// UncompressBlock uncompresses the source buffer into the destination one,
21// and returns the uncompressed size.
22//
23// The destination buffer must be sized appropriately.
24//
25// An error is returned if the source data is invalid or the destination buffer is too small.
26func UncompressBlock(src, dst []byte) (int, error) {
27 if len(src) == 0 {
28 return 0, nil
29 }
30 if di := decodeBlock(dst, src); di >= 0 {
31 return di, nil
32 }
33 return 0, ErrInvalidSourceShortBuffer
34}
35
36// CompressBlock compresses the source buffer into the destination one.
37// This is the fast version of LZ4 compression and also the default one.
38// The size of hashTable must be at least 64Kb.
39//
40// The size of the compressed data is returned. If it is 0 and no error, then the data is incompressible.
41//
42// An error is returned if the destination buffer is too small.
43func CompressBlock(src, dst []byte, hashTable []int) (di int, err error) {
44 defer recoverBlock(&err)
45
46 // adaptSkipLog sets how quickly the compressor begins skipping blocks when data is incompressible.
47 // This significantly speeds up incompressible data and usually has very small impact on compresssion.
48 // bytes to skip = 1 + (bytes since last match >> adaptSkipLog)
49 const adaptSkipLog = 7
50 sn, dn := len(src)-mfLimit, len(dst)
51 if sn <= 0 || dn == 0 {
52 return 0, nil
53 }
54 if len(hashTable) < htSize {
55 return 0, fmt.Errorf("hash table too small, should be at least %d in size", htSize)
56 }
57 // Prove to the compiler the table has at least htSize elements.
58 // The compiler can see that "uint32() >> hashShift" cannot be out of bounds.
59 hashTable = hashTable[:htSize]
60
61 // si: Current position of the search.
62 // anchor: Position of the current literals.
63 var si, anchor int
64
65 // Fast scan strategy: the hash table only stores the last 4 bytes sequences.
66 for si < sn {
67 // Hash the next 6 bytes (sequence)...
68 match := binary.LittleEndian.Uint64(src[si:])
69 h := blockHash(match)
70 h2 := blockHash(match >> 8)
71
72 // We check a match at s, s+1 and s+2 and pick the first one we get.
73 // Checking 3 only requires us to load the source one.
74 ref := hashTable[h]
75 ref2 := hashTable[h2]
76 hashTable[h] = si
77 hashTable[h2] = si + 1
78 offset := si - ref
79
80 // If offset <= 0 we got an old entry in the hash table.
81 if offset <= 0 || offset >= winSize || // Out of window.
82 uint32(match) != binary.LittleEndian.Uint32(src[ref:]) { // Hash collision on different matches.
83 // No match. Start calculating another hash.
84 // The processor can usually do this out-of-order.
85 h = blockHash(match >> 16)
86 ref = hashTable[h]
87
88 // Check the second match at si+1
89 si += 1
90 offset = si - ref2
91
92 if offset <= 0 || offset >= winSize ||
93 uint32(match>>8) != binary.LittleEndian.Uint32(src[ref2:]) {
94 // No match. Check the third match at si+2
95 si += 1
96 offset = si - ref
97 hashTable[h] = si
98
99 if offset <= 0 || offset >= winSize ||
100 uint32(match>>16) != binary.LittleEndian.Uint32(src[ref:]) {
101 // Skip one extra byte (at si+3) before we check 3 matches again.
102 si += 2 + (si-anchor)>>adaptSkipLog
103 continue
104 }
105 }
106 }
107
108 // Match found.
109 lLen := si - anchor // Literal length.
110 // We already matched 4 bytes.
111 mLen := 4
112
113 // Extend backwards if we can, reducing literals.
114 tOff := si - offset - 1
115 for lLen > 0 && tOff >= 0 && src[si-1] == src[tOff] {
116 si--
117 tOff--
118 lLen--
119 mLen++
120 }
121
122 // Add the match length, so we continue search at the end.
123 // Use mLen to store the offset base.
124 si, mLen = si+mLen, si+minMatch
125
126 // Find the longest match by looking by batches of 8 bytes.
127 for si < sn {
128 x := binary.LittleEndian.Uint64(src[si:]) ^ binary.LittleEndian.Uint64(src[si-offset:])
129 if x == 0 {
130 si += 8
131 } else {
132 // Stop is first non-zero byte.
133 si += bits.TrailingZeros64(x) >> 3
134 break
135 }
136 }
137
138 mLen = si - mLen
139 if mLen < 0xF {
140 dst[di] = byte(mLen)
141 } else {
142 dst[di] = 0xF
143 }
144
145 // Encode literals length.
146 if lLen < 0xF {
147 dst[di] |= byte(lLen << 4)
148 } else {
149 dst[di] |= 0xF0
150 di++
151 l := lLen - 0xF
152 for ; l >= 0xFF; l -= 0xFF {
153 dst[di] = 0xFF
154 di++
155 }
156 dst[di] = byte(l)
157 }
158 di++
159
160 // Literals.
161 copy(dst[di:di+lLen], src[anchor:anchor+lLen])
162 di += lLen + 2
163 anchor = si
164
165 // Encode offset.
166 _ = dst[di] // Bound check elimination.
167 dst[di-2], dst[di-1] = byte(offset), byte(offset>>8)
168
169 // Encode match length part 2.
170 if mLen >= 0xF {
171 for mLen -= 0xF; mLen >= 0xFF; mLen -= 0xFF {
172 dst[di] = 0xFF
173 di++
174 }
175 dst[di] = byte(mLen)
176 di++
177 }
178 // Check if we can load next values.
179 if si >= sn {
180 break
181 }
182 // Hash match end-2
183 h = blockHash(binary.LittleEndian.Uint64(src[si-2:]))
184 hashTable[h] = si - 2
185 }
186
187 if anchor == 0 {
188 // Incompressible.
189 return 0, nil
190 }
191
192 // Last literals.
193 lLen := len(src) - anchor
194 if lLen < 0xF {
195 dst[di] = byte(lLen << 4)
196 } else {
197 dst[di] = 0xF0
198 di++
199 for lLen -= 0xF; lLen >= 0xFF; lLen -= 0xFF {
200 dst[di] = 0xFF
201 di++
202 }
203 dst[di] = byte(lLen)
204 }
205 di++
206
207 // Write the last literals.
208 if di >= anchor {
209 // Incompressible.
210 return 0, nil
211 }
212 di += copy(dst[di:di+len(src)-anchor], src[anchor:])
213 return di, nil
214}
215
216// blockHash hashes 4 bytes into a value < winSize.
217func blockHashHC(x uint32) uint32 {
218 const hasher uint32 = 2654435761 // Knuth multiplicative hash.
219 return x * hasher >> (32 - winSizeLog)
220}
221
222// CompressBlockHC compresses the source buffer src into the destination dst
223// with max search depth (use 0 or negative value for no max).
224//
225// CompressBlockHC compression ratio is better than CompressBlock but it is also slower.
226//
227// The size of the compressed data is returned. If it is 0 and no error, then the data is not compressible.
228//
229// An error is returned if the destination buffer is too small.
230func CompressBlockHC(src, dst []byte, depth int) (di int, err error) {
231 defer recoverBlock(&err)
232
233 // adaptSkipLog sets how quickly the compressor begins skipping blocks when data is incompressible.
234 // This significantly speeds up incompressible data and usually has very small impact on compresssion.
235 // bytes to skip = 1 + (bytes since last match >> adaptSkipLog)
236 const adaptSkipLog = 7
237
238 sn, dn := len(src)-mfLimit, len(dst)
239 if sn <= 0 || dn == 0 {
240 return 0, nil
241 }
242 var si int
243
244 // hashTable: stores the last position found for a given hash
245 // chainTable: stores previous positions for a given hash
246 var hashTable, chainTable [winSize]int
247
248 if depth <= 0 {
249 depth = winSize
250 }
251
252 anchor := si
253 for si < sn {
254 // Hash the next 4 bytes (sequence).
255 match := binary.LittleEndian.Uint32(src[si:])
256 h := blockHashHC(match)
257
258 // Follow the chain until out of window and give the longest match.
259 mLen := 0
260 offset := 0
261 for next, try := hashTable[h], depth; try > 0 && next > 0 && si-next < winSize; next = chainTable[next&winMask] {
262 // The first (mLen==0) or next byte (mLen>=minMatch) at current match length
263 // must match to improve on the match length.
264 if src[next+mLen] != src[si+mLen] {
265 continue
266 }
267 ml := 0
268 // Compare the current position with a previous with the same hash.
269 for ml < sn-si {
270 x := binary.LittleEndian.Uint64(src[next+ml:]) ^ binary.LittleEndian.Uint64(src[si+ml:])
271 if x == 0 {
272 ml += 8
273 } else {
274 // Stop is first non-zero byte.
275 ml += bits.TrailingZeros64(x) >> 3
276 break
277 }
278 }
279 if ml < minMatch || ml <= mLen {
280 // Match too small (<minMath) or smaller than the current match.
281 continue
282 }
283 // Found a longer match, keep its position and length.
284 mLen = ml
285 offset = si - next
286 // Try another previous position with the same hash.
287 try--
288 }
289 chainTable[si&winMask] = hashTable[h]
290 hashTable[h] = si
291
292 // No match found.
293 if mLen == 0 {
294 si += 1 + (si-anchor)>>adaptSkipLog
295 continue
296 }
297
298 // Match found.
299 // Update hash/chain tables with overlapping bytes:
300 // si already hashed, add everything from si+1 up to the match length.
301 winStart := si + 1
302 if ws := si + mLen - winSize; ws > winStart {
303 winStart = ws
304 }
305 for si, ml := winStart, si+mLen; si < ml; {
306 match >>= 8
307 match |= uint32(src[si+3]) << 24
308 h := blockHashHC(match)
309 chainTable[si&winMask] = hashTable[h]
310 hashTable[h] = si
311 si++
312 }
313
314 lLen := si - anchor
315 si += mLen
316 mLen -= minMatch // Match length does not include minMatch.
317
318 if mLen < 0xF {
319 dst[di] = byte(mLen)
320 } else {
321 dst[di] = 0xF
322 }
323
324 // Encode literals length.
325 if lLen < 0xF {
326 dst[di] |= byte(lLen << 4)
327 } else {
328 dst[di] |= 0xF0
329 di++
330 l := lLen - 0xF
331 for ; l >= 0xFF; l -= 0xFF {
332 dst[di] = 0xFF
333 di++
334 }
335 dst[di] = byte(l)
336 }
337 di++
338
339 // Literals.
340 copy(dst[di:di+lLen], src[anchor:anchor+lLen])
341 di += lLen
342 anchor = si
343
344 // Encode offset.
345 di += 2
346 dst[di-2], dst[di-1] = byte(offset), byte(offset>>8)
347
348 // Encode match length part 2.
349 if mLen >= 0xF {
350 for mLen -= 0xF; mLen >= 0xFF; mLen -= 0xFF {
351 dst[di] = 0xFF
352 di++
353 }
354 dst[di] = byte(mLen)
355 di++
356 }
357 }
358
359 if anchor == 0 {
360 // Incompressible.
361 return 0, nil
362 }
363
364 // Last literals.
365 lLen := len(src) - anchor
366 if lLen < 0xF {
367 dst[di] = byte(lLen << 4)
368 } else {
369 dst[di] = 0xF0
370 di++
371 lLen -= 0xF
372 for ; lLen >= 0xFF; lLen -= 0xFF {
373 dst[di] = 0xFF
374 di++
375 }
376 dst[di] = byte(lLen)
377 }
378 di++
379
380 // Write the last literals.
381 if di >= anchor {
382 // Incompressible.
383 return 0, nil
384 }
385 di += copy(dst[di:di+len(src)-anchor], src[anchor:])
386 return di, nil
387}