| 1 | /* inffast.c -- fast decoding |
| 2 | * Copyright (C) 1995-2017 Mark Adler |
| 3 | * For conditions of distribution and use, see copyright notice in zlib.h |
| 4 | */ |
| 5 | |
| 6 | #include "zbuild.h" |
| 7 | #include "zutil.h" |
| 8 | #include "inftrees.h" |
| 9 | #include "inflate.h" |
| 10 | #include "inffast.h" |
| 11 | #include "inflate_p.h" |
| 12 | #include "memcopy.h" |
| 13 | |
| 14 | /* |
| 15 | Decode literal, length, and distance codes and write out the resulting |
| 16 | literal and match bytes until either not enough input or output is |
| 17 | available, an end-of-block is encountered, or a data error is encountered. |
| 18 | When large enough input and output buffers are supplied to inflate(), for |
| 19 | example, a 16K input buffer and a 64K output buffer, more than 95% of the |
| 20 | inflate execution time is spent in this routine. |
| 21 | |
| 22 | Entry assumptions: |
| 23 | |
| 24 | state->mode == LEN |
| 25 | strm->avail_in >= INFLATE_FAST_MIN_HAVE |
| 26 | strm->avail_out >= INFLATE_FAST_MIN_LEFT |
| 27 | start >= strm->avail_out |
| 28 | state->bits < 8 |
| 29 | |
| 30 | On return, state->mode is one of: |
| 31 | |
| 32 | LEN -- ran out of enough output space or enough available input |
| 33 | TYPE -- reached end of block code, inflate() to interpret next block |
| 34 | BAD -- error in block data |
| 35 | |
| 36 | Notes: |
| 37 | |
| 38 | - The maximum input bits used by a length/distance pair is 15 bits for the |
| 39 | length code, 5 bits for the length extra, 15 bits for the distance code, |
| 40 | and 13 bits for the distance extra. This totals 48 bits, or six bytes. |
| 41 | Therefore if strm->avail_in >= 6, then there is enough input to avoid |
| 42 | checking for available input while decoding. |
| 43 | |
| 44 | - On some architectures, it can be significantly faster (e.g. up to 1.2x |
| 45 | faster on x86_64) to load from strm->next_in 64 bits, or 8 bytes, at a |
| 46 | time, so INFLATE_FAST_MIN_HAVE == 8. |
| 47 | |
| 48 | - The maximum bytes that a single length/distance pair can output is 258 |
| 49 | bytes, which is the maximum length that can be coded. inflate_fast() |
| 50 | requires strm->avail_out >= 258 for each loop to avoid checking for |
| 51 | output space. |
| 52 | */ |
| 53 | void ZLIB_INTERNAL zng_inflate_fast(PREFIX3(stream) *strm, unsigned long start) { |
| 54 | /* start: inflate()'s starting value for strm->avail_out */ |
| 55 | struct inflate_state *state; |
| 56 | const unsigned char *in; /* local strm->next_in */ |
| 57 | const unsigned char *last; /* have enough input while in < last */ |
| 58 | unsigned char *out; /* local strm->next_out */ |
| 59 | unsigned char *beg; /* inflate()'s initial strm->next_out */ |
| 60 | unsigned char *end; /* while out < end, enough space available */ |
| 61 | #ifdef INFFAST_CHUNKSIZE |
| 62 | unsigned char *safe; /* can use chunkcopy provided out < safe */ |
| 63 | #endif |
| 64 | #ifdef INFLATE_STRICT |
| 65 | unsigned dmax; /* maximum distance from zlib header */ |
| 66 | #endif |
| 67 | unsigned wsize; /* window size or zero if not using window */ |
| 68 | unsigned whave; /* valid bytes in the window */ |
| 69 | unsigned wnext; /* window write index */ |
| 70 | unsigned char *window; /* allocated sliding window, if wsize != 0 */ |
| 71 | |
| 72 | /* hold is a local copy of strm->hold. By default, hold satisfies the same |
| 73 | invariants that strm->hold does, namely that (hold >> bits) == 0. This |
| 74 | invariant is kept by loading bits into hold one byte at a time, like: |
| 75 | |
| 76 | hold |= next_byte_of_input << bits; in++; bits += 8; |
| 77 | |
| 78 | If we need to ensure that bits >= 15 then this code snippet is simply |
| 79 | repeated. Over one iteration of the outermost do/while loop, this |
| 80 | happens up to six times (48 bits of input), as described in the NOTES |
| 81 | above. |
| 82 | |
| 83 | However, on some little endian architectures, it can be significantly |
| 84 | faster to load 64 bits once instead of 8 bits six times: |
| 85 | |
| 86 | if (bits <= 16) { |
| 87 | hold |= next_8_bytes_of_input << bits; in += 6; bits += 48; |
| 88 | } |
| 89 | |
| 90 | Unlike the simpler one byte load, shifting the next_8_bytes_of_input |
| 91 | by bits will overflow and lose those high bits, up to 2 bytes' worth. |
| 92 | The conservative estimate is therefore that we have read only 6 bytes |
| 93 | (48 bits). Again, as per the NOTES above, 48 bits is sufficient for the |
| 94 | rest of the iteration, and we will not need to load another 8 bytes. |
| 95 | |
| 96 | Inside this function, we no longer satisfy (hold >> bits) == 0, but |
| 97 | this is not problematic, even if that overflow does not land on an 8 bit |
| 98 | byte boundary. Those excess bits will eventually shift down lower as the |
| 99 | Huffman decoder consumes input, and when new input bits need to be loaded |
| 100 | into the bits variable, the same input bits will be or'ed over those |
| 101 | existing bits. A bitwise or is idempotent: (a | b | b) equals (a | b). |
| 102 | Note that we therefore write that load operation as "hold |= etc" and not |
| 103 | "hold += etc". |
| 104 | |
| 105 | Outside that loop, at the end of the function, hold is bitwise and'ed |
| 106 | with (1<<bits)-1 to drop those excess bits so that, on function exit, we |
| 107 | keep the invariant that (state->hold >> state->bits) == 0. |
| 108 | */ |
| 109 | uint64_t hold; /* local strm->hold */ |
| 110 | unsigned bits; /* local strm->bits */ |
| 111 | code const *lcode; /* local strm->lencode */ |
| 112 | code const *dcode; /* local strm->distcode */ |
| 113 | unsigned lmask; /* mask for first level of length codes */ |
| 114 | unsigned dmask; /* mask for first level of distance codes */ |
| 115 | const code *here; /* retrieved table entry */ |
| 116 | unsigned op; /* code bits, operation, extra bits, or */ |
| 117 | /* window position, window bytes to copy */ |
| 118 | unsigned len; /* match length, unused bytes */ |
| 119 | unsigned dist; /* match distance */ |
| 120 | unsigned char *from; /* where to copy match from */ |
| 121 | |
| 122 | /* copy state to local variables */ |
| 123 | state = (struct inflate_state *)strm->state; |
| 124 | in = strm->next_in; |
| 125 | last = in + (strm->avail_in - (INFLATE_FAST_MIN_HAVE - 1)); |
| 126 | out = strm->next_out; |
| 127 | beg = out - (start - strm->avail_out); |
| 128 | end = out + (strm->avail_out - (INFLATE_FAST_MIN_LEFT - 1)); |
| 129 | |
| 130 | #ifdef INFFAST_CHUNKSIZE |
| 131 | safe = out + strm->avail_out; |
| 132 | #endif |
| 133 | #ifdef INFLATE_STRICT |
| 134 | dmax = state->dmax; |
| 135 | #endif |
| 136 | wsize = state->wsize; |
| 137 | whave = state->whave; |
| 138 | wnext = state->wnext; |
| 139 | window = state->window; |
| 140 | hold = state->hold; |
| 141 | bits = state->bits; |
| 142 | lcode = state->lencode; |
| 143 | dcode = state->distcode; |
| 144 | lmask = (1U << state->lenbits) - 1; |
| 145 | dmask = (1U << state->distbits) - 1; |
| 146 | |
| 147 | /* decode literals and length/distances until end-of-block or not enough |
| 148 | input data or output space */ |
| 149 | do { |
| 150 | if (bits < 15) { |
| 151 | hold |= load_64_bits(in, bits); |
| 152 | in += 6; |
| 153 | bits += 48; |
| 154 | } |
| 155 | here = lcode + (hold & lmask); |
| 156 | dolen: |
| 157 | DROPBITS(here->bits); |
| 158 | op = here->op; |
| 159 | if (op == 0) { /* literal */ |
| 160 | Tracevv((stderr, here->val >= 0x20 && here->val < 0x7f ? |
| 161 | "inflate: literal '%c'\n" : |
| 162 | "inflate: literal 0x%02x\n" , here->val)); |
| 163 | *out++ = (unsigned char)(here->val); |
| 164 | } else if (op & 16) { /* length base */ |
| 165 | len = here->val; |
| 166 | op &= 15; /* number of extra bits */ |
| 167 | if (op) { |
| 168 | if (bits < op) { |
| 169 | hold |= load_64_bits(in, bits); |
| 170 | in += 6; |
| 171 | bits += 48; |
| 172 | } |
| 173 | len += BITS(op); |
| 174 | DROPBITS(op); |
| 175 | } |
| 176 | Tracevv((stderr, "inflate: length %u\n" , len)); |
| 177 | if (bits < 15) { |
| 178 | hold |= load_64_bits(in, bits); |
| 179 | in += 6; |
| 180 | bits += 48; |
| 181 | } |
| 182 | here = dcode + (hold & dmask); |
| 183 | dodist: |
| 184 | DROPBITS(here->bits); |
| 185 | op = here->op; |
| 186 | if (op & 16) { /* distance base */ |
| 187 | dist = here->val; |
| 188 | op &= 15; /* number of extra bits */ |
| 189 | if (bits < op) { |
| 190 | hold |= load_64_bits(in, bits); |
| 191 | in += 6; |
| 192 | bits += 48; |
| 193 | } |
| 194 | dist += BITS(op); |
| 195 | #ifdef INFLATE_STRICT |
| 196 | if (dist > dmax) { |
| 197 | strm->msg = (char *)"invalid distance too far back" ; |
| 198 | state->mode = BAD; |
| 199 | break; |
| 200 | } |
| 201 | #endif |
| 202 | DROPBITS(op); |
| 203 | Tracevv((stderr, "inflate: distance %u\n" , dist)); |
| 204 | op = (unsigned)(out - beg); /* max distance in output */ |
| 205 | if (dist > op) { /* see if copy from window */ |
| 206 | op = dist - op; /* distance back in window */ |
| 207 | if (op > whave) { |
| 208 | if (state->sane) { |
| 209 | strm->msg = (char *)"invalid distance too far back" ; |
| 210 | state->mode = BAD; |
| 211 | break; |
| 212 | } |
| 213 | #ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR |
| 214 | if (len <= op - whave) { |
| 215 | do { |
| 216 | *out++ = 0; |
| 217 | } while (--len); |
| 218 | continue; |
| 219 | } |
| 220 | len -= op - whave; |
| 221 | do { |
| 222 | *out++ = 0; |
| 223 | } while (--op > whave); |
| 224 | if (op == 0) { |
| 225 | from = out - dist; |
| 226 | do { |
| 227 | *out++ = *from++; |
| 228 | } while (--len); |
| 229 | continue; |
| 230 | } |
| 231 | #endif |
| 232 | } |
| 233 | #ifdef INFFAST_CHUNKSIZE |
| 234 | from = window; |
| 235 | if (wnext == 0) { /* very common case */ |
| 236 | from += wsize - op; |
| 237 | } else if (wnext >= op) { /* contiguous in window */ |
| 238 | from += wnext - op; |
| 239 | } else { /* wrap around window */ |
| 240 | op -= wnext; |
| 241 | from += wsize - op; |
| 242 | if (op < len) { /* some from end of window */ |
| 243 | len -= op; |
| 244 | out = chunkcopysafe(out, from, op, safe); |
| 245 | from = window; /* more from start of window */ |
| 246 | op = wnext; |
| 247 | /* This (rare) case can create a situation where |
| 248 | the first chunkcopy below must be checked. |
| 249 | */ |
| 250 | } |
| 251 | } |
| 252 | if (op < len) { /* still need some from output */ |
| 253 | len -= op; |
| 254 | out = chunkcopysafe(out, from, op, safe); |
| 255 | out = chunkunroll(out, &dist, &len); |
| 256 | out = chunkcopysafe(out, out - dist, len, safe); |
| 257 | } else { |
| 258 | out = chunkcopysafe(out, from, len, safe); |
| 259 | } |
| 260 | #else |
| 261 | from = window; |
| 262 | if (wnext == 0) { /* very common case */ |
| 263 | from += wsize - op; |
| 264 | if (op < len) { /* some from window */ |
| 265 | len -= op; |
| 266 | do { |
| 267 | *out++ = *from++; |
| 268 | } while (--op); |
| 269 | from = out - dist; /* rest from output */ |
| 270 | } |
| 271 | } else if (wnext < op) { /* wrap around window */ |
| 272 | from += wsize + wnext - op; |
| 273 | op -= wnext; |
| 274 | if (op < len) { /* some from end of window */ |
| 275 | len -= op; |
| 276 | do { |
| 277 | *out++ = *from++; |
| 278 | } while (--op); |
| 279 | from = window; |
| 280 | if (wnext < len) { /* some from start of window */ |
| 281 | op = wnext; |
| 282 | len -= op; |
| 283 | do { |
| 284 | *out++ = *from++; |
| 285 | } while (--op); |
| 286 | from = out - dist; /* rest from output */ |
| 287 | } |
| 288 | } |
| 289 | } else { /* contiguous in window */ |
| 290 | from += wnext - op; |
| 291 | if (op < len) { /* some from window */ |
| 292 | len -= op; |
| 293 | do { |
| 294 | *out++ = *from++; |
| 295 | } while (--op); |
| 296 | from = out - dist; /* rest from output */ |
| 297 | } |
| 298 | } |
| 299 | |
| 300 | out = chunk_copy(out, from, (int) (out - from), len); |
| 301 | #endif |
| 302 | } else { |
| 303 | #ifdef INFFAST_CHUNKSIZE |
| 304 | /* Whole reference is in range of current output. No |
| 305 | range checks are necessary because we start with room |
| 306 | for at least 258 bytes of output, so unroll and roundoff |
| 307 | operations can write beyond `out+len` so long as they |
| 308 | stay within 258 bytes of `out`. |
| 309 | */ |
| 310 | if (dist >= len || dist >= INFFAST_CHUNKSIZE) |
| 311 | out = chunkcopy(out, out - dist, len); |
| 312 | else |
| 313 | out = chunkmemset(out, dist, len); |
| 314 | #else |
| 315 | if (len < sizeof(uint64_t)) |
| 316 | out = set_bytes(out, out - dist, dist, len); |
| 317 | else if (dist == 1) |
| 318 | out = byte_memset(out, len); |
| 319 | else |
| 320 | out = chunk_memset(out, out - dist, dist, len); |
| 321 | #endif |
| 322 | } |
| 323 | } else if ((op & 64) == 0) { /* 2nd level distance code */ |
| 324 | here = dcode + here->val + BITS(op); |
| 325 | goto dodist; |
| 326 | } else { |
| 327 | strm->msg = (char *)"invalid distance code" ; |
| 328 | state->mode = BAD; |
| 329 | break; |
| 330 | } |
| 331 | } else if ((op & 64) == 0) { /* 2nd level length code */ |
| 332 | here = lcode + here->val + BITS(op); |
| 333 | goto dolen; |
| 334 | } else if (op & 32) { /* end-of-block */ |
| 335 | Tracevv((stderr, "inflate: end of block\n" )); |
| 336 | state->mode = TYPE; |
| 337 | break; |
| 338 | } else { |
| 339 | strm->msg = (char *)"invalid literal/length code" ; |
| 340 | state->mode = BAD; |
| 341 | break; |
| 342 | } |
| 343 | } while (in < last && out < end); |
| 344 | |
| 345 | /* return unused bytes (on entry, bits < 8, so in won't go too far back) */ |
| 346 | len = bits >> 3; |
| 347 | in -= len; |
| 348 | bits -= len << 3; |
| 349 | hold &= (UINT64_C(1) << bits) - 1; |
| 350 | |
| 351 | /* update state and return */ |
| 352 | strm->next_in = in; |
| 353 | strm->next_out = out; |
| 354 | strm->avail_in = |
| 355 | (unsigned)(in < last ? (INFLATE_FAST_MIN_HAVE - 1) + (last - in) |
| 356 | : (INFLATE_FAST_MIN_HAVE - 1) - (in - last)); |
| 357 | strm->avail_out = |
| 358 | (unsigned)(out < end ? (INFLATE_FAST_MIN_LEFT - 1) + (end - out) |
| 359 | : (INFLATE_FAST_MIN_LEFT - 1) - (out - end)); |
| 360 | |
| 361 | Assert(bits <= 32, "Remaining bits greater than 32" ); |
| 362 | state->hold = (uint32_t)hold; |
| 363 | state->bits = bits; |
| 364 | return; |
| 365 | } |
| 366 | |
| 367 | /* |
| 368 | inflate_fast() speedups that turned out slower (on a PowerPC G3 750CXe): |
| 369 | - Using bit fields for code structure |
| 370 | - Different op definition to avoid & for extra bits (do & for table bits) |
| 371 | - Three separate decoding do-loops for direct, window, and wnext == 0 |
| 372 | - Special case for distance > 1 copies to do overlapped load and store copy |
| 373 | - Explicit branch predictions (based on measured branch probabilities) |
| 374 | - Deferring match copy and interspersed it with decoding subsequent codes |
| 375 | - Swapping literal/length else |
| 376 | - Swapping window/direct else |
| 377 | - Larger unrolled copy loops (three is about right) |
| 378 | - Moving len -= 3 statement into middle of loop |
| 379 | */ |
| 380 | |