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 "functable.h" |
13 | |
14 | |
15 | /* Load 64 bits from IN and place the bytes at offset BITS in the result. */ |
16 | static inline uint64_t load_64_bits(const unsigned char *in, unsigned bits) { |
17 | uint64_t chunk; |
18 | memcpy(dest: &chunk, src: in, n: sizeof(chunk)); |
19 | |
20 | #if BYTE_ORDER == LITTLE_ENDIAN |
21 | return chunk << bits; |
22 | #else |
23 | return ZSWAP64(chunk) << bits; |
24 | #endif |
25 | } |
26 | /* |
27 | Decode literal, length, and distance codes and write out the resulting |
28 | literal and match bytes until either not enough input or output is |
29 | available, an end-of-block is encountered, or a data error is encountered. |
30 | When large enough input and output buffers are supplied to inflate(), for |
31 | example, a 16K input buffer and a 64K output buffer, more than 95% of the |
32 | inflate execution time is spent in this routine. |
33 | |
34 | Entry assumptions: |
35 | |
36 | state->mode == LEN |
37 | strm->avail_in >= INFLATE_FAST_MIN_HAVE |
38 | strm->avail_out >= INFLATE_FAST_MIN_LEFT |
39 | start >= strm->avail_out |
40 | state->bits < 8 |
41 | |
42 | On return, state->mode is one of: |
43 | |
44 | LEN -- ran out of enough output space or enough available input |
45 | TYPE -- reached end of block code, inflate() to interpret next block |
46 | BAD -- error in block data |
47 | |
48 | Notes: |
49 | |
50 | - The maximum input bits used by a length/distance pair is 15 bits for the |
51 | length code, 5 bits for the length extra, 15 bits for the distance code, |
52 | and 13 bits for the distance extra. This totals 48 bits, or six bytes. |
53 | Therefore if strm->avail_in >= 6, then there is enough input to avoid |
54 | checking for available input while decoding. |
55 | |
56 | - On some architectures, it can be significantly faster (e.g. up to 1.2x |
57 | faster on x86_64) to load from strm->next_in 64 bits, or 8 bytes, at a |
58 | time, so INFLATE_FAST_MIN_HAVE == 8. |
59 | |
60 | - The maximum bytes that a single length/distance pair can output is 258 |
61 | bytes, which is the maximum length that can be coded. inflate_fast() |
62 | requires strm->avail_out >= 258 for each loop to avoid checking for |
63 | output space. |
64 | */ |
65 | void Z_INTERNAL zng_inflate_fast(PREFIX3(stream) *strm, unsigned long start) { |
66 | /* start: inflate()'s starting value for strm->avail_out */ |
67 | struct inflate_state *state; |
68 | z_const unsigned char *in; /* local strm->next_in */ |
69 | const unsigned char *last; /* have enough input while in < last */ |
70 | unsigned char *out; /* local strm->next_out */ |
71 | unsigned char *beg; /* inflate()'s initial strm->next_out */ |
72 | unsigned char *end; /* while out < end, enough space available */ |
73 | unsigned char *safe; /* can use chunkcopy provided out < safe */ |
74 | #ifdef INFLATE_STRICT |
75 | unsigned dmax; /* maximum distance from zlib header */ |
76 | #endif |
77 | unsigned wsize; /* window size or zero if not using window */ |
78 | unsigned whave; /* valid bytes in the window */ |
79 | unsigned wnext; /* window write index */ |
80 | unsigned char *window; /* allocated sliding window, if wsize != 0 */ |
81 | |
82 | /* hold is a local copy of strm->hold. By default, hold satisfies the same |
83 | invariants that strm->hold does, namely that (hold >> bits) == 0. This |
84 | invariant is kept by loading bits into hold one byte at a time, like: |
85 | |
86 | hold |= next_byte_of_input << bits; in++; bits += 8; |
87 | |
88 | If we need to ensure that bits >= 15 then this code snippet is simply |
89 | repeated. Over one iteration of the outermost do/while loop, this |
90 | happens up to six times (48 bits of input), as described in the NOTES |
91 | above. |
92 | |
93 | However, on some little endian architectures, it can be significantly |
94 | faster to load 64 bits once instead of 8 bits six times: |
95 | |
96 | if (bits <= 16) { |
97 | hold |= next_8_bytes_of_input << bits; in += 6; bits += 48; |
98 | } |
99 | |
100 | Unlike the simpler one byte load, shifting the next_8_bytes_of_input |
101 | by bits will overflow and lose those high bits, up to 2 bytes' worth. |
102 | The conservative estimate is therefore that we have read only 6 bytes |
103 | (48 bits). Again, as per the NOTES above, 48 bits is sufficient for the |
104 | rest of the iteration, and we will not need to load another 8 bytes. |
105 | |
106 | Inside this function, we no longer satisfy (hold >> bits) == 0, but |
107 | this is not problematic, even if that overflow does not land on an 8 bit |
108 | byte boundary. Those excess bits will eventually shift down lower as the |
109 | Huffman decoder consumes input, and when new input bits need to be loaded |
110 | into the bits variable, the same input bits will be or'ed over those |
111 | existing bits. A bitwise or is idempotent: (a | b | b) equals (a | b). |
112 | Note that we therefore write that load operation as "hold |= etc" and not |
113 | "hold += etc". |
114 | |
115 | Outside that loop, at the end of the function, hold is bitwise and'ed |
116 | with (1<<bits)-1 to drop those excess bits so that, on function exit, we |
117 | keep the invariant that (state->hold >> state->bits) == 0. |
118 | */ |
119 | uint64_t hold; /* local strm->hold */ |
120 | unsigned bits; /* local strm->bits */ |
121 | code const *lcode; /* local strm->lencode */ |
122 | code const *dcode; /* local strm->distcode */ |
123 | unsigned lmask; /* mask for first level of length codes */ |
124 | unsigned dmask; /* mask for first level of distance codes */ |
125 | const code *here; /* retrieved table entry */ |
126 | unsigned op; /* code bits, operation, extra bits, or */ |
127 | /* window position, window bytes to copy */ |
128 | unsigned len; /* match length, unused bytes */ |
129 | unsigned dist; /* match distance */ |
130 | unsigned char *from; /* where to copy match from */ |
131 | unsigned ; /* copy chunks safely in all cases */ |
132 | |
133 | /* copy state to local variables */ |
134 | state = (struct inflate_state *)strm->state; |
135 | in = strm->next_in; |
136 | last = in + (strm->avail_in - (INFLATE_FAST_MIN_HAVE - 1)); |
137 | out = strm->next_out; |
138 | beg = out - (start - strm->avail_out); |
139 | end = out + (strm->avail_out - (INFLATE_FAST_MIN_LEFT - 1)); |
140 | safe = out + strm->avail_out; |
141 | #ifdef INFLATE_STRICT |
142 | dmax = state->dmax; |
143 | #endif |
144 | wsize = state->wsize; |
145 | whave = state->whave; |
146 | wnext = state->wnext; |
147 | window = state->window; |
148 | hold = state->hold; |
149 | bits = state->bits; |
150 | lcode = state->lencode; |
151 | dcode = state->distcode; |
152 | lmask = (1U << state->lenbits) - 1; |
153 | dmask = (1U << state->distbits) - 1; |
154 | |
155 | /* Detect if out and window point to the same memory allocation. In this instance it is |
156 | necessary to use safe chunk copy functions to prevent overwriting the window. If the |
157 | window is overwritten then future matches with far distances will fail to copy correctly. */ |
158 | extra_safe = (wsize != 0 && out >= window && out + INFLATE_FAST_MIN_LEFT <= window + wsize); |
159 | |
160 | /* decode literals and length/distances until end-of-block or not enough |
161 | input data or output space */ |
162 | do { |
163 | if (bits < 15) { |
164 | hold |= load_64_bits(in, bits); |
165 | in += 6; |
166 | bits += 48; |
167 | } |
168 | here = lcode + (hold & lmask); |
169 | dolen: |
170 | DROPBITS(here->bits); |
171 | op = here->op; |
172 | if (op == 0) { /* literal */ |
173 | Tracevv((stderr, here->val >= 0x20 && here->val < 0x7f ? |
174 | "inflate: literal '%c'\n" : |
175 | "inflate: literal 0x%02x\n" , here->val)); |
176 | *out++ = (unsigned char)(here->val); |
177 | } else if (op & 16) { /* length base */ |
178 | len = here->val; |
179 | op &= 15; /* number of extra bits */ |
180 | if (bits < op) { |
181 | hold |= load_64_bits(in, bits); |
182 | in += 6; |
183 | bits += 48; |
184 | } |
185 | len += BITS(op); |
186 | DROPBITS(op); |
187 | Tracevv((stderr, "inflate: length %u\n" , len)); |
188 | if (bits < 15) { |
189 | hold |= load_64_bits(in, bits); |
190 | in += 6; |
191 | bits += 48; |
192 | } |
193 | here = dcode + (hold & dmask); |
194 | dodist: |
195 | DROPBITS(here->bits); |
196 | op = here->op; |
197 | if (op & 16) { /* distance base */ |
198 | dist = here->val; |
199 | op &= 15; /* number of extra bits */ |
200 | if (bits < op) { |
201 | hold |= load_64_bits(in, bits); |
202 | in += 6; |
203 | bits += 48; |
204 | } |
205 | dist += BITS(op); |
206 | #ifdef INFLATE_STRICT |
207 | if (dist > dmax) { |
208 | SET_BAD("invalid distance too far back" ); |
209 | break; |
210 | } |
211 | #endif |
212 | DROPBITS(op); |
213 | Tracevv((stderr, "inflate: distance %u\n" , dist)); |
214 | op = (unsigned)(out - beg); /* max distance in output */ |
215 | if (dist > op) { /* see if copy from window */ |
216 | op = dist - op; /* distance back in window */ |
217 | if (op > whave) { |
218 | if (state->sane) { |
219 | SET_BAD("invalid distance too far back" ); |
220 | break; |
221 | } |
222 | #ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR |
223 | if (len <= op - whave) { |
224 | do { |
225 | *out++ = 0; |
226 | } while (--len); |
227 | continue; |
228 | } |
229 | len -= op - whave; |
230 | do { |
231 | *out++ = 0; |
232 | } while (--op > whave); |
233 | if (op == 0) { |
234 | from = out - dist; |
235 | do { |
236 | *out++ = *from++; |
237 | } while (--len); |
238 | continue; |
239 | } |
240 | #endif |
241 | } |
242 | from = window; |
243 | if (wnext == 0) { /* very common case */ |
244 | from += wsize - op; |
245 | } else if (wnext >= op) { /* contiguous in window */ |
246 | from += wnext - op; |
247 | } else { /* wrap around window */ |
248 | op -= wnext; |
249 | from += wsize - op; |
250 | if (op < len) { /* some from end of window */ |
251 | len -= op; |
252 | out = functable.chunkcopy_safe(out, from, op, safe); |
253 | from = window; /* more from start of window */ |
254 | op = wnext; |
255 | /* This (rare) case can create a situation where |
256 | the first chunkcopy below must be checked. |
257 | */ |
258 | } |
259 | } |
260 | if (op < len) { /* still need some from output */ |
261 | len -= op; |
262 | out = functable.chunkcopy_safe(out, from, op, safe); |
263 | out = functable.chunkunroll(out, &dist, &len); |
264 | out = functable.chunkcopy_safe(out, out - dist, len, safe); |
265 | } else { |
266 | out = functable.chunkcopy_safe(out, from, len, safe); |
267 | } |
268 | } else if (extra_safe) { |
269 | /* Whole reference is in range of current output. */ |
270 | if (dist >= len || dist >= state->chunksize) |
271 | out = functable.chunkcopy_safe(out, out - dist, len, safe); |
272 | else |
273 | out = functable.chunkmemset_safe(out, dist, len, (unsigned)((safe - out) + 1)); |
274 | } else { |
275 | /* Whole reference is in range of current output. No range checks are |
276 | necessary because we start with room for at least 258 bytes of output, |
277 | so unroll and roundoff operations can write beyond `out+len` so long |
278 | as they stay within 258 bytes of `out`. |
279 | */ |
280 | if (dist >= len || dist >= state->chunksize) |
281 | out = functable.chunkcopy(out, out - dist, len); |
282 | else |
283 | out = functable.chunkmemset(out, dist, len); |
284 | } |
285 | } else if ((op & 64) == 0) { /* 2nd level distance code */ |
286 | here = dcode + here->val + BITS(op); |
287 | goto dodist; |
288 | } else { |
289 | SET_BAD("invalid distance code" ); |
290 | break; |
291 | } |
292 | } else if ((op & 64) == 0) { /* 2nd level length code */ |
293 | here = lcode + here->val + BITS(op); |
294 | goto dolen; |
295 | } else if (op & 32) { /* end-of-block */ |
296 | Tracevv((stderr, "inflate: end of block\n" )); |
297 | state->mode = TYPE; |
298 | break; |
299 | } else { |
300 | SET_BAD("invalid literal/length code" ); |
301 | break; |
302 | } |
303 | } while (in < last && out < end); |
304 | |
305 | /* return unused bytes (on entry, bits < 8, so in won't go too far back) */ |
306 | len = bits >> 3; |
307 | in -= len; |
308 | bits -= len << 3; |
309 | hold &= (UINT64_C(1) << bits) - 1; |
310 | |
311 | /* update state and return */ |
312 | strm->next_in = in; |
313 | strm->next_out = out; |
314 | strm->avail_in = (unsigned)(in < last ? (INFLATE_FAST_MIN_HAVE - 1) + (last - in) |
315 | : (INFLATE_FAST_MIN_HAVE - 1) - (in - last)); |
316 | strm->avail_out = (unsigned)(out < end ? (INFLATE_FAST_MIN_LEFT - 1) + (end - out) |
317 | : (INFLATE_FAST_MIN_LEFT - 1) - (out - end)); |
318 | |
319 | Assert(bits <= 32, "Remaining bits greater than 32" ); |
320 | state->hold = (uint32_t)hold; |
321 | state->bits = bits; |
322 | return; |
323 | } |
324 | |
325 | /* |
326 | inflate_fast() speedups that turned out slower (on a PowerPC G3 750CXe): |
327 | - Using bit fields for code structure |
328 | - Different op definition to avoid & for extra bits (do & for table bits) |
329 | - Three separate decoding do-loops for direct, window, and wnext == 0 |
330 | - Special case for distance > 1 copies to do overlapped load and store copy |
331 | - Explicit branch predictions (based on measured branch probabilities) |
332 | - Deferring match copy and interspersed it with decoding subsequent codes |
333 | - Swapping literal/length else |
334 | - Swapping window/direct else |
335 | - Larger unrolled copy loops (three is about right) |
336 | - Moving len -= 3 statement into middle of loop |
337 | */ |
338 | |