| Abhay Kumar | a2ae599 | 2025-11-10 14:02:24 +0000 | [diff] [blame^] | 1 | package flate |
| 2 | |
| 3 | import "fmt" |
| 4 | |
| 5 | // fastEncL3 |
| 6 | type fastEncL3 struct { |
| 7 | fastGen |
| 8 | table [1 << 16]tableEntryPrev |
| 9 | } |
| 10 | |
| 11 | // Encode uses a similar algorithm to level 2, will check up to two candidates. |
| 12 | func (e *fastEncL3) Encode(dst *tokens, src []byte) { |
| 13 | const ( |
| 14 | inputMargin = 12 - 1 |
| 15 | minNonLiteralBlockSize = 1 + 1 + inputMargin |
| 16 | tableBits = 16 |
| 17 | tableSize = 1 << tableBits |
| 18 | hashBytes = 5 |
| 19 | ) |
| 20 | |
| 21 | if debugDeflate && e.cur < 0 { |
| 22 | panic(fmt.Sprint("e.cur < 0: ", e.cur)) |
| 23 | } |
| 24 | |
| 25 | // Protect against e.cur wraparound. |
| 26 | for e.cur >= bufferReset { |
| 27 | if len(e.hist) == 0 { |
| 28 | for i := range e.table[:] { |
| 29 | e.table[i] = tableEntryPrev{} |
| 30 | } |
| 31 | e.cur = maxMatchOffset |
| 32 | break |
| 33 | } |
| 34 | // Shift down everything in the table that isn't already too far away. |
| 35 | minOff := e.cur + int32(len(e.hist)) - maxMatchOffset |
| 36 | for i := range e.table[:] { |
| 37 | v := e.table[i] |
| 38 | if v.Cur.offset <= minOff { |
| 39 | v.Cur.offset = 0 |
| 40 | } else { |
| 41 | v.Cur.offset = v.Cur.offset - e.cur + maxMatchOffset |
| 42 | } |
| 43 | if v.Prev.offset <= minOff { |
| 44 | v.Prev.offset = 0 |
| 45 | } else { |
| 46 | v.Prev.offset = v.Prev.offset - e.cur + maxMatchOffset |
| 47 | } |
| 48 | e.table[i] = v |
| 49 | } |
| 50 | e.cur = maxMatchOffset |
| 51 | } |
| 52 | |
| 53 | s := e.addBlock(src) |
| 54 | |
| 55 | // Skip if too small. |
| 56 | if len(src) < minNonLiteralBlockSize { |
| 57 | // We do not fill the token table. |
| 58 | // This will be picked up by caller. |
| 59 | dst.n = uint16(len(src)) |
| 60 | return |
| 61 | } |
| 62 | |
| 63 | // Override src |
| 64 | src = e.hist |
| 65 | nextEmit := s |
| 66 | |
| 67 | // sLimit is when to stop looking for offset/length copies. The inputMargin |
| 68 | // lets us use a fast path for emitLiteral in the main loop, while we are |
| 69 | // looking for copies. |
| 70 | sLimit := int32(len(src) - inputMargin) |
| 71 | |
| 72 | // nextEmit is where in src the next emitLiteral should start from. |
| 73 | cv := load6432(src, s) |
| 74 | for { |
| 75 | const skipLog = 7 |
| 76 | nextS := s |
| 77 | var candidate tableEntry |
| 78 | for { |
| 79 | nextHash := hashLen(cv, tableBits, hashBytes) |
| 80 | s = nextS |
| 81 | nextS = s + 1 + (s-nextEmit)>>skipLog |
| 82 | if nextS > sLimit { |
| 83 | goto emitRemainder |
| 84 | } |
| 85 | candidates := e.table[nextHash] |
| 86 | now := load6432(src, nextS) |
| 87 | |
| 88 | // Safe offset distance until s + 4... |
| 89 | minOffset := e.cur + s - (maxMatchOffset - 4) |
| 90 | e.table[nextHash] = tableEntryPrev{Prev: candidates.Cur, Cur: tableEntry{offset: s + e.cur}} |
| 91 | |
| 92 | // Check both candidates |
| 93 | candidate = candidates.Cur |
| 94 | if candidate.offset < minOffset { |
| 95 | cv = now |
| 96 | // Previous will also be invalid, we have nothing. |
| 97 | continue |
| 98 | } |
| 99 | |
| 100 | if uint32(cv) == load3232(src, candidate.offset-e.cur) { |
| 101 | if candidates.Prev.offset < minOffset || uint32(cv) != load3232(src, candidates.Prev.offset-e.cur) { |
| 102 | break |
| 103 | } |
| 104 | // Both match and are valid, pick longest. |
| 105 | offset := s - (candidate.offset - e.cur) |
| 106 | o2 := s - (candidates.Prev.offset - e.cur) |
| 107 | l1, l2 := matchLen(src[s+4:], src[s-offset+4:]), matchLen(src[s+4:], src[s-o2+4:]) |
| 108 | if l2 > l1 { |
| 109 | candidate = candidates.Prev |
| 110 | } |
| 111 | break |
| 112 | } else { |
| 113 | // We only check if value mismatches. |
| 114 | // Offset will always be invalid in other cases. |
| 115 | candidate = candidates.Prev |
| 116 | if candidate.offset > minOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) { |
| 117 | break |
| 118 | } |
| 119 | } |
| 120 | cv = now |
| 121 | } |
| 122 | |
| 123 | // Call emitCopy, and then see if another emitCopy could be our next |
| 124 | // move. Repeat until we find no match for the input immediately after |
| 125 | // what was consumed by the last emitCopy call. |
| 126 | // |
| 127 | // If we exit this loop normally then we need to call emitLiteral next, |
| 128 | // though we don't yet know how big the literal will be. We handle that |
| 129 | // by proceeding to the next iteration of the main loop. We also can |
| 130 | // exit this loop via goto if we get close to exhausting the input. |
| 131 | for { |
| 132 | // Invariant: we have a 4-byte match at s, and no need to emit any |
| 133 | // literal bytes prior to s. |
| 134 | |
| 135 | // Extend the 4-byte match as long as possible. |
| 136 | // |
| 137 | t := candidate.offset - e.cur |
| 138 | l := e.matchlenLong(int(s+4), int(t+4), src) + 4 |
| 139 | |
| 140 | // Extend backwards |
| 141 | for t > 0 && s > nextEmit && src[t-1] == src[s-1] { |
| 142 | s-- |
| 143 | t-- |
| 144 | l++ |
| 145 | } |
| 146 | if nextEmit < s { |
| 147 | if false { |
| 148 | emitLiteral(dst, src[nextEmit:s]) |
| 149 | } else { |
| 150 | for _, v := range src[nextEmit:s] { |
| 151 | dst.tokens[dst.n] = token(v) |
| 152 | dst.litHist[v]++ |
| 153 | dst.n++ |
| 154 | } |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | dst.AddMatchLong(l, uint32(s-t-baseMatchOffset)) |
| 159 | s += l |
| 160 | nextEmit = s |
| 161 | if nextS >= s { |
| 162 | s = nextS + 1 |
| 163 | } |
| 164 | |
| 165 | if s >= sLimit { |
| 166 | t += l |
| 167 | // Index first pair after match end. |
| 168 | if int(t+8) < len(src) && t > 0 { |
| 169 | cv = load6432(src, t) |
| 170 | nextHash := hashLen(cv, tableBits, hashBytes) |
| 171 | e.table[nextHash] = tableEntryPrev{ |
| 172 | Prev: e.table[nextHash].Cur, |
| 173 | Cur: tableEntry{offset: e.cur + t}, |
| 174 | } |
| 175 | } |
| 176 | goto emitRemainder |
| 177 | } |
| 178 | |
| 179 | // Store every 5th hash in-between. |
| 180 | for i := s - l + 2; i < s-5; i += 6 { |
| 181 | nextHash := hashLen(load6432(src, i), tableBits, hashBytes) |
| 182 | e.table[nextHash] = tableEntryPrev{ |
| 183 | Prev: e.table[nextHash].Cur, |
| 184 | Cur: tableEntry{offset: e.cur + i}} |
| 185 | } |
| 186 | // We could immediately start working at s now, but to improve |
| 187 | // compression we first update the hash table at s-2 to s. |
| 188 | x := load6432(src, s-2) |
| 189 | prevHash := hashLen(x, tableBits, hashBytes) |
| 190 | |
| 191 | e.table[prevHash] = tableEntryPrev{ |
| 192 | Prev: e.table[prevHash].Cur, |
| 193 | Cur: tableEntry{offset: e.cur + s - 2}, |
| 194 | } |
| 195 | x >>= 8 |
| 196 | prevHash = hashLen(x, tableBits, hashBytes) |
| 197 | |
| 198 | e.table[prevHash] = tableEntryPrev{ |
| 199 | Prev: e.table[prevHash].Cur, |
| 200 | Cur: tableEntry{offset: e.cur + s - 1}, |
| 201 | } |
| 202 | x >>= 8 |
| 203 | currHash := hashLen(x, tableBits, hashBytes) |
| 204 | candidates := e.table[currHash] |
| 205 | cv = x |
| 206 | e.table[currHash] = tableEntryPrev{ |
| 207 | Prev: candidates.Cur, |
| 208 | Cur: tableEntry{offset: s + e.cur}, |
| 209 | } |
| 210 | |
| 211 | // Check both candidates |
| 212 | candidate = candidates.Cur |
| 213 | minOffset := e.cur + s - (maxMatchOffset - 4) |
| 214 | |
| 215 | if candidate.offset > minOffset { |
| 216 | if uint32(cv) == load3232(src, candidate.offset-e.cur) { |
| 217 | // Found a match... |
| 218 | continue |
| 219 | } |
| 220 | candidate = candidates.Prev |
| 221 | if candidate.offset > minOffset && uint32(cv) == load3232(src, candidate.offset-e.cur) { |
| 222 | // Match at prev... |
| 223 | continue |
| 224 | } |
| 225 | } |
| 226 | cv = x >> 8 |
| 227 | s++ |
| 228 | break |
| 229 | } |
| 230 | } |
| 231 | |
| 232 | emitRemainder: |
| 233 | if int(nextEmit) < len(src) { |
| 234 | // If nothing was added, don't encode literals. |
| 235 | if dst.n == 0 { |
| 236 | return |
| 237 | } |
| 238 | |
| 239 | emitLiteral(dst, src[nextEmit:]) |
| 240 | } |
| 241 | } |