|
| 1 | +// This is a very baby version of LZ4 compression algorithm |
| 2 | + |
| 3 | +#include "compress.h" |
| 4 | + |
| 5 | +#include "bool.h" |
| 6 | +#include "libc/string.h" |
| 7 | + |
| 8 | +#define PACKED __attribute__((packed)) |
| 9 | +#define __GET_UNALIGNED_T(type, ptr) \ |
| 10 | + ({ \ |
| 11 | + const struct { \ |
| 12 | + type x; \ |
| 13 | + } PACKED *__pptr = (typeof(__pptr)) (ptr); \ |
| 14 | + __pptr->x; \ |
| 15 | + }) |
| 16 | + |
| 17 | +#define __PUT_UNALIGNED_T(type, val, ptr) \ |
| 18 | + do { \ |
| 19 | + struct { \ |
| 20 | + type x; \ |
| 21 | + } PACKED *__pptr = (typeof(__pptr)) (ptr); \ |
| 22 | + __pptr->x = (val); \ |
| 23 | + } while (0) |
| 24 | + |
| 25 | +#define GET_UNALIGNED4(ptr) __GET_UNALIGNED_T(uint32_t, (ptr)) |
| 26 | +#define PUT_UNALIGNED4(val, ptr) __PUT_UNALIGNED_T(uint32_t, (val), (ptr)) |
| 27 | + |
| 28 | +#define MINMATCH 3 |
| 29 | + |
| 30 | +typedef struct |
| 31 | +{ |
| 32 | + uint32_t match3; |
| 33 | + uint32_t match4; |
| 34 | +} MinMatch; |
| 35 | + |
| 36 | +static MinMatch readMinMatch(const uint8_t* buf) |
| 37 | +{ |
| 38 | + const uint32_t mask = 0xffffff00; |
| 39 | + const uint32_t match4 = GET_UNALIGNED4(buf); |
| 40 | + const uint32_t match3 = match4 & mask; |
| 41 | + return (MinMatch) { match3, match4 }; |
| 42 | +} |
| 43 | + |
| 44 | +static uint32_t readMinMatch3(const uint8_t* buf) |
| 45 | +{ |
| 46 | + const uint32_t mask = 0xffffff00; |
| 47 | + const uint32_t match4 = GET_UNALIGNED4(buf); |
| 48 | + return match4 & mask; |
| 49 | +} |
| 50 | + |
| 51 | +#define HASH_CHAINS_BITS 12 |
| 52 | +#define HASH_CHAINS_SIZE (1 << HASH_CHAINS_BITS) |
| 53 | + |
| 54 | +typedef struct |
| 55 | +{ |
| 56 | + uint32_t hash3; |
| 57 | + uint32_t hash4; |
| 58 | +} MinMatcHash; |
| 59 | + |
| 60 | +static MinMatcHash hashMinMatch(MinMatch match) |
| 61 | +{ |
| 62 | + return (MinMatcHash) { ((match.match3 * 506832829U) << 8) >> (32 - HASH_CHAINS_BITS), (match.match4 * 2654435761U) >> (32 - HASH_CHAINS_BITS)}; |
| 63 | +} |
| 64 | + |
| 65 | +static void wildCopy4(uint8_t* dst, const uint8_t* src, const uint8_t* dstEnd) |
| 66 | +{ |
| 67 | + do |
| 68 | + { |
| 69 | + PUT_UNALIGNED4(GET_UNALIGNED4(src), dst); |
| 70 | + src += 4; |
| 71 | + dst += 4; |
| 72 | + } while (dst < dstEnd); |
| 73 | +} |
| 74 | + |
| 75 | +#define WILD_COPY_MOVE(dst, src, amount) do{ uint8_t* start = dst; dst += (amount); wildCopy4(start, src, dst); }while(0) |
| 76 | + |
| 77 | +static void encodeLargeInt(uint8_t** _out, int leftToEncode) |
| 78 | +{ |
| 79 | +#define out (*_out) |
| 80 | + bool keepEncoding = true; |
| 81 | + while (keepEncoding) |
| 82 | + { |
| 83 | + keepEncoding = leftToEncode >= 255; |
| 84 | + *(out++) = (uint8_t)(keepEncoding ? 255 : leftToEncode); |
| 85 | + leftToEncode -= 255; |
| 86 | + } |
| 87 | +#undef out |
| 88 | +} |
| 89 | + |
| 90 | +typedef struct |
| 91 | +{ |
| 92 | + int matchLen; |
| 93 | + const uint8_t* inMatch; |
| 94 | +} MinMatchPtr; |
| 95 | + |
| 96 | +static MinMatchPtr tryMinMatch(uint32_t potentialMinMatchOffset, MinMatch minMatch, const uint8_t* inCursor, const uint8_t* inLimit, int inCursorOffset) |
| 97 | +{ |
| 98 | + const uint8_t* potentialMinMatch = inCursor - inCursorOffset + potentialMinMatchOffset; |
| 99 | + |
| 100 | + // need to do some fun wraparound stuff |
| 101 | + if (potentialMinMatch > inCursor - 4) |
| 102 | + potentialMinMatch -= 256; |
| 103 | + |
| 104 | + if (minMatch.match3 == readMinMatch3(potentialMinMatch)) |
| 105 | + { |
| 106 | + // find the longest match we have with the potential match |
| 107 | + int matchLen = MINMATCH; |
| 108 | + while (1) |
| 109 | + { |
| 110 | + if (inCursor + matchLen >= inLimit) |
| 111 | + break; |
| 112 | + if (inCursor[matchLen] != potentialMinMatch[matchLen]) |
| 113 | + break; |
| 114 | + |
| 115 | + matchLen++; |
| 116 | + } |
| 117 | + |
| 118 | + return (MinMatchPtr) { matchLen, potentialMinMatch }; |
| 119 | + } |
| 120 | + else |
| 121 | + { |
| 122 | + return (MinMatchPtr) { 0, NULL }; |
| 123 | + } |
| 124 | +} |
| 125 | + |
| 126 | +void mlz4_compress(const uint8_t* in, const uint32_t inSize, uint8_t* out, uint32_t* outSize) |
| 127 | +{ |
| 128 | + uint8_t* compressedData = out; |
| 129 | + |
| 130 | + uint8_t* outCursor = compressedData; |
| 131 | +#ifndef TARGET_N64 |
| 132 | + uint8_t hashChains3[HASH_CHAINS_SIZE] = {}; |
| 133 | + uint8_t hashChains4[HASH_CHAINS_SIZE] = {}; |
| 134 | +#else |
| 135 | + // placed in gDecompressionHeap |
| 136 | + uint8_t* hashChains3 = (uint8_t*) 0x801c1000; |
| 137 | + uint8_t* hashChains4 = (uint8_t*) 0x801c1000 + HASH_CHAINS_SIZE; |
| 138 | +#endif |
| 139 | + |
| 140 | + // should be safe to do |
| 141 | + *(uint32_t*)outCursor = inSize; |
| 142 | + outCursor += 4; |
| 143 | + |
| 144 | + // uncompressible data, just write it out as literals |
| 145 | + if (inSize <= 8) |
| 146 | + { |
| 147 | + outCursor[0] = ((uint8_t)inSize) << 4; |
| 148 | + outCursor++; |
| 149 | + WILD_COPY_MOVE(outCursor, in, inSize); |
| 150 | + *outSize = (size_t) (outCursor - compressedData); |
| 151 | + return; |
| 152 | + } |
| 153 | + |
| 154 | + const uint8_t* inEnd = in + inSize; |
| 155 | + const uint8_t* inCursor = in + 4; |
| 156 | + const uint8_t* inLimit = inEnd - 4; |
| 157 | + const uint8_t* inLiteralStart = in; |
| 158 | + |
| 159 | + // start the loop |
| 160 | + while (1) |
| 161 | + { |
| 162 | + // push match we took last time |
| 163 | + { |
| 164 | + MinMatch prevMinMatch = readMinMatch(inCursor - 4); |
| 165 | + MinMatcHash hashValue = hashMinMatch(prevMinMatch); |
| 166 | + uint8_t offset = (uint8_t)(inCursor - 4 - in); |
| 167 | + hashChains3[hashValue.hash3] = offset; |
| 168 | + hashChains4[hashValue.hash4] = offset; |
| 169 | + } |
| 170 | + |
| 171 | + if (inCursor >= inLimit) |
| 172 | + { |
| 173 | + // printf("Final literals: %lld\n", inEnd - inLiteralStart); |
| 174 | + const int literalsCount = inEnd - inLiteralStart; |
| 175 | + if (literalsCount >= 15) |
| 176 | + { |
| 177 | + *(outCursor++) = 0xf0; |
| 178 | + encodeLargeInt(&outCursor, literalsCount - 15); |
| 179 | + } |
| 180 | + else |
| 181 | + { |
| 182 | + *(outCursor++) = ((uint8_t) literalsCount) << 4; |
| 183 | + } |
| 184 | + |
| 185 | + WILD_COPY_MOVE(outCursor, inLiteralStart, literalsCount); |
| 186 | + *outSize = (uint32_t)(outCursor - compressedData); |
| 187 | + return; |
| 188 | + } |
| 189 | + |
| 190 | + MinMatch minMatch = readMinMatch(inCursor); |
| 191 | + MinMatcHash hashValue = hashMinMatch(minMatch); |
| 192 | + uint8_t inCursorOffset = (uint8_t) (inCursor - in); |
| 193 | + |
| 194 | + uint8_t potentialMinMatchOffset3 = hashChains3[hashValue.hash3]; |
| 195 | + uint8_t potentialMinMatchOffset4 = hashChains4[hashValue.hash4]; |
| 196 | + |
| 197 | + MinMatchPtr match3 = tryMinMatch(potentialMinMatchOffset3, minMatch, inCursor, inLimit, inCursorOffset); |
| 198 | + MinMatchPtr match4 = tryMinMatch(potentialMinMatchOffset4, minMatch, inCursor, inLimit, inCursorOffset); |
| 199 | + |
| 200 | + MinMatchPtr match = match3.matchLen > match4.matchLen ? match3 : match4; |
| 201 | + if (match.matchLen) |
| 202 | + { |
| 203 | + // write the match... |
| 204 | + uint8_t* pnibble = outCursor++; |
| 205 | + const int literalsCount = (int) (inCursor - inLiteralStart); |
| 206 | + if (literalsCount >= 15) |
| 207 | + { |
| 208 | + *pnibble = 0xf0; |
| 209 | + encodeLargeInt(&outCursor, literalsCount - 15); |
| 210 | + } |
| 211 | + else |
| 212 | + { |
| 213 | + *pnibble = ((uint8_t)literalsCount) << 4; |
| 214 | + } |
| 215 | + |
| 216 | + WILD_COPY_MOVE(outCursor, inLiteralStart, literalsCount); |
| 217 | + *(outCursor++) = (uint8_t)(inCursor - match.inMatch - 4); |
| 218 | + |
| 219 | + const int matchesCount = match.matchLen - MINMATCH; |
| 220 | + if (matchesCount >= 15) |
| 221 | + { |
| 222 | + *pnibble |= 0x0f; |
| 223 | + encodeLargeInt(&outCursor, matchesCount - 15); |
| 224 | + } |
| 225 | + else |
| 226 | + { |
| 227 | + *pnibble |= (uint8_t)matchesCount; |
| 228 | + } |
| 229 | + |
| 230 | + // printf("Literals: %lld, Offset: %lld, Match: %d\n", inCursor - inLiteralStart, inCursor - match.inMatch, match.matchLen); |
| 231 | + inCursor += match.matchLen; |
| 232 | + inLiteralStart = inCursor; |
| 233 | + } |
| 234 | + else |
| 235 | + { |
| 236 | + inCursor++; |
| 237 | + } |
| 238 | + } |
| 239 | +} |
| 240 | + |
| 241 | +void mlz4_decompress(const uint8_t* in, uint8_t* out) |
| 242 | +{ |
| 243 | + const uint8_t* inCursor = in; |
| 244 | + uint32_t originalSize = *(uint32_t*)inCursor; |
| 245 | + inCursor += 4; |
| 246 | + |
| 247 | + uint8_t* outCursor = out; |
| 248 | + uint8_t* outEnd = out + originalSize; |
| 249 | + uint8_t* outSafetyMargin = outEnd - 4; |
| 250 | + while (1) |
| 251 | + { |
| 252 | + uint8_t nibble = *(inCursor++); |
| 253 | + int literalsCount = nibble >> 4; |
| 254 | + if (literalsCount == 15) |
| 255 | + { |
| 256 | + while (1) |
| 257 | + { |
| 258 | + uint8_t nextNibble = *(inCursor++); |
| 259 | + literalsCount += nextNibble; |
| 260 | + if (nextNibble != 255) |
| 261 | + break; |
| 262 | + } |
| 263 | + } |
| 264 | + |
| 265 | + if (literalsCount) |
| 266 | + { |
| 267 | + uint8_t* literalsEnd = outCursor + literalsCount; |
| 268 | + if (literalsEnd >= outSafetyMargin) |
| 269 | + { |
| 270 | + // last 4 literals must be in safety margin |
| 271 | + memcpy(outCursor, inCursor, literalsCount); |
| 272 | + outCursor += literalsCount; |
| 273 | + break; |
| 274 | + } |
| 275 | + else |
| 276 | + { |
| 277 | + wildCopy4(outCursor, inCursor, literalsEnd); |
| 278 | + outCursor = literalsEnd; |
| 279 | + inCursor += literalsCount; |
| 280 | + } |
| 281 | + } |
| 282 | + |
| 283 | + int offset = 4 + *(inCursor++); |
| 284 | + int matchesCount = nibble & 0x0f; |
| 285 | + if (matchesCount == 15) |
| 286 | + { |
| 287 | + while (1) |
| 288 | + { |
| 289 | + uint8_t nextNibble = *(inCursor++); |
| 290 | + matchesCount += nextNibble; |
| 291 | + if (nextNibble != 255) |
| 292 | + break; |
| 293 | + } |
| 294 | + } |
| 295 | + matchesCount += MINMATCH; |
| 296 | + |
| 297 | + const uint8_t* matchStart = outCursor - offset; |
| 298 | + WILD_COPY_MOVE(outCursor, matchStart, matchesCount); |
| 299 | + } |
| 300 | +} |
0 commit comments