Thu, 22 Jan 2015 13:21:57 +0100
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
michael@0 | 1 | /* inffast.c -- fast decoding |
michael@0 | 2 | * Copyright (C) 1995-2008, 2010 Mark Adler |
michael@0 | 3 | * For conditions of distribution and use, see copyright notice in zlib.h |
michael@0 | 4 | */ |
michael@0 | 5 | |
michael@0 | 6 | #include "zutil.h" |
michael@0 | 7 | #include "inftrees.h" |
michael@0 | 8 | #include "inflate.h" |
michael@0 | 9 | #include "inffast.h" |
michael@0 | 10 | |
michael@0 | 11 | #ifndef ASMINF |
michael@0 | 12 | |
michael@0 | 13 | /* Allow machine dependent optimization for post-increment or pre-increment. |
michael@0 | 14 | Based on testing to date, |
michael@0 | 15 | Pre-increment preferred for: |
michael@0 | 16 | - PowerPC G3 (Adler) |
michael@0 | 17 | - MIPS R5000 (Randers-Pehrson) |
michael@0 | 18 | Post-increment preferred for: |
michael@0 | 19 | - none |
michael@0 | 20 | No measurable difference: |
michael@0 | 21 | - Pentium III (Anderson) |
michael@0 | 22 | - M68060 (Nikl) |
michael@0 | 23 | */ |
michael@0 | 24 | #ifdef POSTINC |
michael@0 | 25 | # define OFF 0 |
michael@0 | 26 | # define PUP(a) *(a)++ |
michael@0 | 27 | #else |
michael@0 | 28 | # define OFF 1 |
michael@0 | 29 | # define PUP(a) *++(a) |
michael@0 | 30 | #endif |
michael@0 | 31 | |
michael@0 | 32 | /* |
michael@0 | 33 | Decode literal, length, and distance codes and write out the resulting |
michael@0 | 34 | literal and match bytes until either not enough input or output is |
michael@0 | 35 | available, an end-of-block is encountered, or a data error is encountered. |
michael@0 | 36 | When large enough input and output buffers are supplied to inflate(), for |
michael@0 | 37 | example, a 16K input buffer and a 64K output buffer, more than 95% of the |
michael@0 | 38 | inflate execution time is spent in this routine. |
michael@0 | 39 | |
michael@0 | 40 | Entry assumptions: |
michael@0 | 41 | |
michael@0 | 42 | state->mode == LEN |
michael@0 | 43 | strm->avail_in >= 6 |
michael@0 | 44 | strm->avail_out >= 258 |
michael@0 | 45 | start >= strm->avail_out |
michael@0 | 46 | state->bits < 8 |
michael@0 | 47 | |
michael@0 | 48 | On return, state->mode is one of: |
michael@0 | 49 | |
michael@0 | 50 | LEN -- ran out of enough output space or enough available input |
michael@0 | 51 | TYPE -- reached end of block code, inflate() to interpret next block |
michael@0 | 52 | BAD -- error in block data |
michael@0 | 53 | |
michael@0 | 54 | Notes: |
michael@0 | 55 | |
michael@0 | 56 | - The maximum input bits used by a length/distance pair is 15 bits for the |
michael@0 | 57 | length code, 5 bits for the length extra, 15 bits for the distance code, |
michael@0 | 58 | and 13 bits for the distance extra. This totals 48 bits, or six bytes. |
michael@0 | 59 | Therefore if strm->avail_in >= 6, then there is enough input to avoid |
michael@0 | 60 | checking for available input while decoding. |
michael@0 | 61 | |
michael@0 | 62 | - The maximum bytes that a single length/distance pair can output is 258 |
michael@0 | 63 | bytes, which is the maximum length that can be coded. inflate_fast() |
michael@0 | 64 | requires strm->avail_out >= 258 for each loop to avoid checking for |
michael@0 | 65 | output space. |
michael@0 | 66 | */ |
michael@0 | 67 | void ZLIB_INTERNAL inflate_fast(strm, start) |
michael@0 | 68 | z_streamp strm; |
michael@0 | 69 | unsigned start; /* inflate()'s starting value for strm->avail_out */ |
michael@0 | 70 | { |
michael@0 | 71 | struct inflate_state FAR *state; |
michael@0 | 72 | unsigned char FAR *in; /* local strm->next_in */ |
michael@0 | 73 | unsigned char FAR *last; /* while in < last, enough input available */ |
michael@0 | 74 | unsigned char FAR *out; /* local strm->next_out */ |
michael@0 | 75 | unsigned char FAR *beg; /* inflate()'s initial strm->next_out */ |
michael@0 | 76 | unsigned char FAR *end; /* while out < end, enough space available */ |
michael@0 | 77 | #ifdef INFLATE_STRICT |
michael@0 | 78 | unsigned dmax; /* maximum distance from zlib header */ |
michael@0 | 79 | #endif |
michael@0 | 80 | unsigned wsize; /* window size or zero if not using window */ |
michael@0 | 81 | unsigned whave; /* valid bytes in the window */ |
michael@0 | 82 | unsigned wnext; /* window write index */ |
michael@0 | 83 | unsigned char FAR *window; /* allocated sliding window, if wsize != 0 */ |
michael@0 | 84 | unsigned long hold; /* local strm->hold */ |
michael@0 | 85 | unsigned bits; /* local strm->bits */ |
michael@0 | 86 | code const FAR *lcode; /* local strm->lencode */ |
michael@0 | 87 | code const FAR *dcode; /* local strm->distcode */ |
michael@0 | 88 | unsigned lmask; /* mask for first level of length codes */ |
michael@0 | 89 | unsigned dmask; /* mask for first level of distance codes */ |
michael@0 | 90 | code here; /* retrieved table entry */ |
michael@0 | 91 | unsigned op; /* code bits, operation, extra bits, or */ |
michael@0 | 92 | /* window position, window bytes to copy */ |
michael@0 | 93 | unsigned len; /* match length, unused bytes */ |
michael@0 | 94 | unsigned dist; /* match distance */ |
michael@0 | 95 | unsigned char FAR *from; /* where to copy match from */ |
michael@0 | 96 | |
michael@0 | 97 | /* copy state to local variables */ |
michael@0 | 98 | state = (struct inflate_state FAR *)strm->state; |
michael@0 | 99 | in = strm->next_in - OFF; |
michael@0 | 100 | last = in + (strm->avail_in - 5); |
michael@0 | 101 | out = strm->next_out - OFF; |
michael@0 | 102 | beg = out - (start - strm->avail_out); |
michael@0 | 103 | end = out + (strm->avail_out - 257); |
michael@0 | 104 | #ifdef INFLATE_STRICT |
michael@0 | 105 | dmax = state->dmax; |
michael@0 | 106 | #endif |
michael@0 | 107 | wsize = state->wsize; |
michael@0 | 108 | whave = state->whave; |
michael@0 | 109 | wnext = state->wnext; |
michael@0 | 110 | window = state->window; |
michael@0 | 111 | hold = state->hold; |
michael@0 | 112 | bits = state->bits; |
michael@0 | 113 | lcode = state->lencode; |
michael@0 | 114 | dcode = state->distcode; |
michael@0 | 115 | lmask = (1U << state->lenbits) - 1; |
michael@0 | 116 | dmask = (1U << state->distbits) - 1; |
michael@0 | 117 | |
michael@0 | 118 | /* decode literals and length/distances until end-of-block or not enough |
michael@0 | 119 | input data or output space */ |
michael@0 | 120 | do { |
michael@0 | 121 | if (bits < 15) { |
michael@0 | 122 | hold += (unsigned long)(PUP(in)) << bits; |
michael@0 | 123 | bits += 8; |
michael@0 | 124 | hold += (unsigned long)(PUP(in)) << bits; |
michael@0 | 125 | bits += 8; |
michael@0 | 126 | } |
michael@0 | 127 | here = lcode[hold & lmask]; |
michael@0 | 128 | dolen: |
michael@0 | 129 | op = (unsigned)(here.bits); |
michael@0 | 130 | hold >>= op; |
michael@0 | 131 | bits -= op; |
michael@0 | 132 | op = (unsigned)(here.op); |
michael@0 | 133 | if (op == 0) { /* literal */ |
michael@0 | 134 | Tracevv((stderr, here.val >= 0x20 && here.val < 0x7f ? |
michael@0 | 135 | "inflate: literal '%c'\n" : |
michael@0 | 136 | "inflate: literal 0x%02x\n", here.val)); |
michael@0 | 137 | PUP(out) = (unsigned char)(here.val); |
michael@0 | 138 | } |
michael@0 | 139 | else if (op & 16) { /* length base */ |
michael@0 | 140 | len = (unsigned)(here.val); |
michael@0 | 141 | op &= 15; /* number of extra bits */ |
michael@0 | 142 | if (op) { |
michael@0 | 143 | if (bits < op) { |
michael@0 | 144 | hold += (unsigned long)(PUP(in)) << bits; |
michael@0 | 145 | bits += 8; |
michael@0 | 146 | } |
michael@0 | 147 | len += (unsigned)hold & ((1U << op) - 1); |
michael@0 | 148 | hold >>= op; |
michael@0 | 149 | bits -= op; |
michael@0 | 150 | } |
michael@0 | 151 | Tracevv((stderr, "inflate: length %u\n", len)); |
michael@0 | 152 | if (bits < 15) { |
michael@0 | 153 | hold += (unsigned long)(PUP(in)) << bits; |
michael@0 | 154 | bits += 8; |
michael@0 | 155 | hold += (unsigned long)(PUP(in)) << bits; |
michael@0 | 156 | bits += 8; |
michael@0 | 157 | } |
michael@0 | 158 | here = dcode[hold & dmask]; |
michael@0 | 159 | dodist: |
michael@0 | 160 | op = (unsigned)(here.bits); |
michael@0 | 161 | hold >>= op; |
michael@0 | 162 | bits -= op; |
michael@0 | 163 | op = (unsigned)(here.op); |
michael@0 | 164 | if (op & 16) { /* distance base */ |
michael@0 | 165 | dist = (unsigned)(here.val); |
michael@0 | 166 | op &= 15; /* number of extra bits */ |
michael@0 | 167 | if (bits < op) { |
michael@0 | 168 | hold += (unsigned long)(PUP(in)) << bits; |
michael@0 | 169 | bits += 8; |
michael@0 | 170 | if (bits < op) { |
michael@0 | 171 | hold += (unsigned long)(PUP(in)) << bits; |
michael@0 | 172 | bits += 8; |
michael@0 | 173 | } |
michael@0 | 174 | } |
michael@0 | 175 | dist += (unsigned)hold & ((1U << op) - 1); |
michael@0 | 176 | #ifdef INFLATE_STRICT |
michael@0 | 177 | if (dist > dmax) { |
michael@0 | 178 | strm->msg = (char *)"invalid distance too far back"; |
michael@0 | 179 | state->mode = BAD; |
michael@0 | 180 | break; |
michael@0 | 181 | } |
michael@0 | 182 | #endif |
michael@0 | 183 | hold >>= op; |
michael@0 | 184 | bits -= op; |
michael@0 | 185 | Tracevv((stderr, "inflate: distance %u\n", dist)); |
michael@0 | 186 | op = (unsigned)(out - beg); /* max distance in output */ |
michael@0 | 187 | if (dist > op) { /* see if copy from window */ |
michael@0 | 188 | op = dist - op; /* distance back in window */ |
michael@0 | 189 | if (op > whave) { |
michael@0 | 190 | if (state->sane) { |
michael@0 | 191 | strm->msg = |
michael@0 | 192 | (char *)"invalid distance too far back"; |
michael@0 | 193 | state->mode = BAD; |
michael@0 | 194 | break; |
michael@0 | 195 | } |
michael@0 | 196 | #ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR |
michael@0 | 197 | if (len <= op - whave) { |
michael@0 | 198 | do { |
michael@0 | 199 | PUP(out) = 0; |
michael@0 | 200 | } while (--len); |
michael@0 | 201 | continue; |
michael@0 | 202 | } |
michael@0 | 203 | len -= op - whave; |
michael@0 | 204 | do { |
michael@0 | 205 | PUP(out) = 0; |
michael@0 | 206 | } while (--op > whave); |
michael@0 | 207 | if (op == 0) { |
michael@0 | 208 | from = out - dist; |
michael@0 | 209 | do { |
michael@0 | 210 | PUP(out) = PUP(from); |
michael@0 | 211 | } while (--len); |
michael@0 | 212 | continue; |
michael@0 | 213 | } |
michael@0 | 214 | #endif |
michael@0 | 215 | } |
michael@0 | 216 | from = window - OFF; |
michael@0 | 217 | if (wnext == 0) { /* very common case */ |
michael@0 | 218 | from += wsize - op; |
michael@0 | 219 | if (op < len) { /* some from window */ |
michael@0 | 220 | len -= op; |
michael@0 | 221 | do { |
michael@0 | 222 | PUP(out) = PUP(from); |
michael@0 | 223 | } while (--op); |
michael@0 | 224 | from = out - dist; /* rest from output */ |
michael@0 | 225 | } |
michael@0 | 226 | } |
michael@0 | 227 | else if (wnext < op) { /* wrap around window */ |
michael@0 | 228 | from += wsize + wnext - op; |
michael@0 | 229 | op -= wnext; |
michael@0 | 230 | if (op < len) { /* some from end of window */ |
michael@0 | 231 | len -= op; |
michael@0 | 232 | do { |
michael@0 | 233 | PUP(out) = PUP(from); |
michael@0 | 234 | } while (--op); |
michael@0 | 235 | from = window - OFF; |
michael@0 | 236 | if (wnext < len) { /* some from start of window */ |
michael@0 | 237 | op = wnext; |
michael@0 | 238 | len -= op; |
michael@0 | 239 | do { |
michael@0 | 240 | PUP(out) = PUP(from); |
michael@0 | 241 | } while (--op); |
michael@0 | 242 | from = out - dist; /* rest from output */ |
michael@0 | 243 | } |
michael@0 | 244 | } |
michael@0 | 245 | } |
michael@0 | 246 | else { /* contiguous in window */ |
michael@0 | 247 | from += wnext - op; |
michael@0 | 248 | if (op < len) { /* some from window */ |
michael@0 | 249 | len -= op; |
michael@0 | 250 | do { |
michael@0 | 251 | PUP(out) = PUP(from); |
michael@0 | 252 | } while (--op); |
michael@0 | 253 | from = out - dist; /* rest from output */ |
michael@0 | 254 | } |
michael@0 | 255 | } |
michael@0 | 256 | while (len > 2) { |
michael@0 | 257 | PUP(out) = PUP(from); |
michael@0 | 258 | PUP(out) = PUP(from); |
michael@0 | 259 | PUP(out) = PUP(from); |
michael@0 | 260 | len -= 3; |
michael@0 | 261 | } |
michael@0 | 262 | if (len) { |
michael@0 | 263 | PUP(out) = PUP(from); |
michael@0 | 264 | if (len > 1) |
michael@0 | 265 | PUP(out) = PUP(from); |
michael@0 | 266 | } |
michael@0 | 267 | } |
michael@0 | 268 | else { |
michael@0 | 269 | from = out - dist; /* copy direct from output */ |
michael@0 | 270 | do { /* minimum length is three */ |
michael@0 | 271 | PUP(out) = PUP(from); |
michael@0 | 272 | PUP(out) = PUP(from); |
michael@0 | 273 | PUP(out) = PUP(from); |
michael@0 | 274 | len -= 3; |
michael@0 | 275 | } while (len > 2); |
michael@0 | 276 | if (len) { |
michael@0 | 277 | PUP(out) = PUP(from); |
michael@0 | 278 | if (len > 1) |
michael@0 | 279 | PUP(out) = PUP(from); |
michael@0 | 280 | } |
michael@0 | 281 | } |
michael@0 | 282 | } |
michael@0 | 283 | else if ((op & 64) == 0) { /* 2nd level distance code */ |
michael@0 | 284 | here = dcode[here.val + (hold & ((1U << op) - 1))]; |
michael@0 | 285 | goto dodist; |
michael@0 | 286 | } |
michael@0 | 287 | else { |
michael@0 | 288 | strm->msg = (char *)"invalid distance code"; |
michael@0 | 289 | state->mode = BAD; |
michael@0 | 290 | break; |
michael@0 | 291 | } |
michael@0 | 292 | } |
michael@0 | 293 | else if ((op & 64) == 0) { /* 2nd level length code */ |
michael@0 | 294 | here = lcode[here.val + (hold & ((1U << op) - 1))]; |
michael@0 | 295 | goto dolen; |
michael@0 | 296 | } |
michael@0 | 297 | else if (op & 32) { /* end-of-block */ |
michael@0 | 298 | Tracevv((stderr, "inflate: end of block\n")); |
michael@0 | 299 | state->mode = TYPE; |
michael@0 | 300 | break; |
michael@0 | 301 | } |
michael@0 | 302 | else { |
michael@0 | 303 | strm->msg = (char *)"invalid literal/length code"; |
michael@0 | 304 | state->mode = BAD; |
michael@0 | 305 | break; |
michael@0 | 306 | } |
michael@0 | 307 | } while (in < last && out < end); |
michael@0 | 308 | |
michael@0 | 309 | /* return unused bytes (on entry, bits < 8, so in won't go too far back) */ |
michael@0 | 310 | len = bits >> 3; |
michael@0 | 311 | in -= len; |
michael@0 | 312 | bits -= len << 3; |
michael@0 | 313 | hold &= (1U << bits) - 1; |
michael@0 | 314 | |
michael@0 | 315 | /* update state and return */ |
michael@0 | 316 | strm->next_in = in + OFF; |
michael@0 | 317 | strm->next_out = out + OFF; |
michael@0 | 318 | strm->avail_in = (unsigned)(in < last ? 5 + (last - in) : 5 - (in - last)); |
michael@0 | 319 | strm->avail_out = (unsigned)(out < end ? |
michael@0 | 320 | 257 + (end - out) : 257 - (out - end)); |
michael@0 | 321 | state->hold = hold; |
michael@0 | 322 | state->bits = bits; |
michael@0 | 323 | return; |
michael@0 | 324 | } |
michael@0 | 325 | |
michael@0 | 326 | /* |
michael@0 | 327 | inflate_fast() speedups that turned out slower (on a PowerPC G3 750CXe): |
michael@0 | 328 | - Using bit fields for code structure |
michael@0 | 329 | - Different op definition to avoid & for extra bits (do & for table bits) |
michael@0 | 330 | - Three separate decoding do-loops for direct, window, and wnext == 0 |
michael@0 | 331 | - Special case for distance > 1 copies to do overlapped load and store copy |
michael@0 | 332 | - Explicit branch predictions (based on measured branch probabilities) |
michael@0 | 333 | - Deferring match copy and interspersed it with decoding subsequent codes |
michael@0 | 334 | - Swapping literal/length else |
michael@0 | 335 | - Swapping window/direct else |
michael@0 | 336 | - Larger unrolled copy loops (three is about right) |
michael@0 | 337 | - Moving len -= 3 statement into middle of loop |
michael@0 | 338 | */ |
michael@0 | 339 | |
michael@0 | 340 | #endif /* !ASMINF */ |