| Abhay Kumar | a2ae599 | 2025-11-10 14:02:24 +0000 | [diff] [blame^] | 1 | package flate |
| 2 | |
| 3 | import ( |
| 4 | "fmt" |
| 5 | |
| 6 | "github.com/klauspost/compress/internal/le" |
| 7 | ) |
| 8 | |
| 9 | // fastGen maintains the table for matches, |
| 10 | // and the previous byte block for level 2. |
| 11 | // This is the generic implementation. |
| 12 | type fastEncL1 struct { |
| 13 | fastGen |
| 14 | table [tableSize]tableEntry |
| 15 | } |
| 16 | |
| 17 | // EncodeL1 uses a similar algorithm to level 1 |
| 18 | func (e *fastEncL1) Encode(dst *tokens, src []byte) { |
| 19 | const ( |
| 20 | inputMargin = 12 - 1 |
| 21 | minNonLiteralBlockSize = 1 + 1 + inputMargin |
| 22 | hashBytes = 5 |
| 23 | ) |
| 24 | if debugDeflate && e.cur < 0 { |
| 25 | panic(fmt.Sprint("e.cur < 0: ", e.cur)) |
| 26 | } |
| 27 | |
| 28 | // Protect against e.cur wraparound. |
| 29 | for e.cur >= bufferReset { |
| 30 | if len(e.hist) == 0 { |
| 31 | for i := range e.table[:] { |
| 32 | e.table[i] = tableEntry{} |
| 33 | } |
| 34 | e.cur = maxMatchOffset |
| 35 | break |
| 36 | } |
| 37 | // Shift down everything in the table that isn't already too far away. |
| 38 | minOff := e.cur + int32(len(e.hist)) - maxMatchOffset |
| 39 | for i := range e.table[:] { |
| 40 | v := e.table[i].offset |
| 41 | if v <= minOff { |
| 42 | v = 0 |
| 43 | } else { |
| 44 | v = v - e.cur + maxMatchOffset |
| 45 | } |
| 46 | e.table[i].offset = v |
| 47 | } |
| 48 | e.cur = maxMatchOffset |
| 49 | } |
| 50 | |
| 51 | s := e.addBlock(src) |
| 52 | |
| 53 | // This check isn't in the Snappy implementation, but there, the caller |
| 54 | // instead of the callee handles this case. |
| 55 | if len(src) < minNonLiteralBlockSize { |
| 56 | // We do not fill the token table. |
| 57 | // This will be picked up by caller. |
| 58 | dst.n = uint16(len(src)) |
| 59 | return |
| 60 | } |
| 61 | |
| 62 | // Override src |
| 63 | src = e.hist |
| 64 | nextEmit := s |
| 65 | |
| 66 | // sLimit is when to stop looking for offset/length copies. The inputMargin |
| 67 | // lets us use a fast path for emitLiteral in the main loop, while we are |
| 68 | // looking for copies. |
| 69 | sLimit := int32(len(src) - inputMargin) |
| 70 | |
| 71 | // nextEmit is where in src the next emitLiteral should start from. |
| 72 | cv := load6432(src, s) |
| 73 | |
| 74 | for { |
| 75 | const skipLog = 5 |
| 76 | const doEvery = 2 |
| 77 | |
| 78 | nextS := s |
| 79 | var candidate tableEntry |
| 80 | var t int32 |
| 81 | for { |
| 82 | nextHash := hashLen(cv, tableBits, hashBytes) |
| 83 | candidate = e.table[nextHash] |
| 84 | nextS = s + doEvery + (s-nextEmit)>>skipLog |
| 85 | if nextS > sLimit { |
| 86 | goto emitRemainder |
| 87 | } |
| 88 | |
| 89 | now := load6432(src, nextS) |
| 90 | e.table[nextHash] = tableEntry{offset: s + e.cur} |
| 91 | nextHash = hashLen(now, tableBits, hashBytes) |
| 92 | t = candidate.offset - e.cur |
| 93 | if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) { |
| 94 | e.table[nextHash] = tableEntry{offset: nextS + e.cur} |
| 95 | break |
| 96 | } |
| 97 | |
| 98 | // Do one right away... |
| 99 | cv = now |
| 100 | s = nextS |
| 101 | nextS++ |
| 102 | candidate = e.table[nextHash] |
| 103 | now >>= 8 |
| 104 | e.table[nextHash] = tableEntry{offset: s + e.cur} |
| 105 | |
| 106 | t = candidate.offset - e.cur |
| 107 | if s-t < maxMatchOffset && uint32(cv) == load3232(src, t) { |
| 108 | e.table[nextHash] = tableEntry{offset: nextS + e.cur} |
| 109 | break |
| 110 | } |
| 111 | cv = now |
| 112 | s = nextS |
| 113 | } |
| 114 | |
| 115 | // A 4-byte match has been found. We'll later see if more than 4 bytes |
| 116 | // match. But, prior to the match, src[nextEmit:s] are unmatched. Emit |
| 117 | // them as literal bytes. |
| 118 | for { |
| 119 | // Invariant: we have a 4-byte match at s, and no need to emit any |
| 120 | // literal bytes prior to s. |
| 121 | |
| 122 | // Extend the 4-byte match as long as possible. |
| 123 | l := e.matchlenLong(int(s+4), int(t+4), src) + 4 |
| 124 | |
| 125 | // Extend backwards |
| 126 | for t > 0 && s > nextEmit && le.Load8(src, t-1) == le.Load8(src, s-1) { |
| 127 | s-- |
| 128 | t-- |
| 129 | l++ |
| 130 | } |
| 131 | if nextEmit < s { |
| 132 | if false { |
| 133 | emitLiteral(dst, src[nextEmit:s]) |
| 134 | } else { |
| 135 | for _, v := range src[nextEmit:s] { |
| 136 | dst.tokens[dst.n] = token(v) |
| 137 | dst.litHist[v]++ |
| 138 | dst.n++ |
| 139 | } |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | // Save the match found |
| 144 | if false { |
| 145 | dst.AddMatchLong(l, uint32(s-t-baseMatchOffset)) |
| 146 | } else { |
| 147 | // Inlined... |
| 148 | xoffset := uint32(s - t - baseMatchOffset) |
| 149 | xlength := l |
| 150 | oc := offsetCode(xoffset) |
| 151 | xoffset |= oc << 16 |
| 152 | for xlength > 0 { |
| 153 | xl := xlength |
| 154 | if xl > 258 { |
| 155 | if xl > 258+baseMatchLength { |
| 156 | xl = 258 |
| 157 | } else { |
| 158 | xl = 258 - baseMatchLength |
| 159 | } |
| 160 | } |
| 161 | xlength -= xl |
| 162 | xl -= baseMatchLength |
| 163 | dst.extraHist[lengthCodes1[uint8(xl)]]++ |
| 164 | dst.offHist[oc]++ |
| 165 | dst.tokens[dst.n] = token(matchType | uint32(xl)<<lengthShift | xoffset) |
| 166 | dst.n++ |
| 167 | } |
| 168 | } |
| 169 | s += l |
| 170 | nextEmit = s |
| 171 | if nextS >= s { |
| 172 | s = nextS + 1 |
| 173 | } |
| 174 | if s >= sLimit { |
| 175 | // Index first pair after match end. |
| 176 | if int(s+l+8) < len(src) { |
| 177 | cv := load6432(src, s) |
| 178 | e.table[hashLen(cv, tableBits, hashBytes)] = tableEntry{offset: s + e.cur} |
| 179 | } |
| 180 | goto emitRemainder |
| 181 | } |
| 182 | |
| 183 | // We could immediately start working at s now, but to improve |
| 184 | // compression we first update the hash table at s-2 and at s. If |
| 185 | // another emitCopy is not our next move, also calculate nextHash |
| 186 | // at s+1. At least on GOARCH=amd64, these three hash calculations |
| 187 | // are faster as one load64 call (with some shifts) instead of |
| 188 | // three load32 calls. |
| 189 | x := load6432(src, s-2) |
| 190 | o := e.cur + s - 2 |
| 191 | prevHash := hashLen(x, tableBits, hashBytes) |
| 192 | e.table[prevHash] = tableEntry{offset: o} |
| 193 | x >>= 16 |
| 194 | currHash := hashLen(x, tableBits, hashBytes) |
| 195 | candidate = e.table[currHash] |
| 196 | e.table[currHash] = tableEntry{offset: o + 2} |
| 197 | |
| 198 | t = candidate.offset - e.cur |
| 199 | if s-t > maxMatchOffset || uint32(x) != load3232(src, t) { |
| 200 | cv = x >> 8 |
| 201 | s++ |
| 202 | break |
| 203 | } |
| 204 | } |
| 205 | } |
| 206 | |
| 207 | emitRemainder: |
| 208 | if int(nextEmit) < len(src) { |
| 209 | // If nothing was added, don't encode literals. |
| 210 | if dst.n == 0 { |
| 211 | return |
| 212 | } |
| 213 | emitLiteral(dst, src[nextEmit:]) |
| 214 | } |
| 215 | } |