/src/zlib-ng/inffast_tpl.h
Line | Count | Source |
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 "inflate_p.h" |
11 | | #include "functable.h" |
12 | | |
13 | | /* |
14 | | Decode literal, length, and distance codes and write out the resulting |
15 | | literal and match bytes until either not enough input or output is |
16 | | available, an end-of-block is encountered, or a data error is encountered. |
17 | | When large enough input and output buffers are supplied to inflate(), for |
18 | | example, a 16K input buffer and a 64K output buffer, more than 95% of the |
19 | | inflate execution time is spent in this routine. |
20 | | |
21 | | Entry assumptions: |
22 | | |
23 | | state->mode == LEN |
24 | | strm->avail_in >= INFLATE_FAST_MIN_HAVE |
25 | | strm->avail_out >= INFLATE_FAST_MIN_LEFT |
26 | | start >= strm->avail_out |
27 | | state->bits < 8 |
28 | | |
29 | | On return, state->mode is one of: |
30 | | |
31 | | LEN -- ran out of enough output space or enough available input |
32 | | TYPE -- reached end of block code, inflate() to interpret next block |
33 | | BAD -- error in block data |
34 | | |
35 | | Notes: |
36 | | |
37 | | - The maximum input bits used by a length/distance pair is 15 bits for the |
38 | | length code, 5 bits for the length extra, 15 bits for the distance code, |
39 | | and 13 bits for the distance extra. This totals 48 bits, or six bytes. |
40 | | Therefore if strm->avail_in >= 6, then there is enough input to avoid |
41 | | checking for available input while decoding. |
42 | | |
43 | | - On some architectures, it can be significantly faster (e.g. up to 1.2x |
44 | | faster on x86_64) to load from strm->next_in 64 bits, or 8 bytes, at a |
45 | | time, so INFLATE_FAST_MIN_HAVE == 8. |
46 | | |
47 | | - The maximum bytes that a single length/distance pair can output is 258 |
48 | | bytes, which is the maximum length that can be coded. inflate_fast() |
49 | | requires strm->avail_out >= 258 for each loop to avoid checking for |
50 | | output space. |
51 | | */ |
52 | 0 | void Z_INTERNAL INFLATE_FAST(PREFIX3(stream) *strm, uint32_t start, int safe_mode) { |
53 | | /* start: inflate()'s starting value for strm->avail_out */ |
54 | 0 | struct inflate_state *state; |
55 | 0 | z_const unsigned char *in; /* local strm->next_in */ |
56 | 0 | const unsigned char *last; /* have enough input while in < last */ |
57 | 0 | unsigned char *out; /* local strm->next_out */ |
58 | 0 | unsigned char *beg; /* inflate()'s initial strm->next_out */ |
59 | 0 | unsigned char *end; /* while out < end, enough space available */ |
60 | 0 | unsigned char *safe; /* can use chunkcopy provided out < safe */ |
61 | 0 | unsigned char *window; /* allocated sliding window, if wsize != 0 */ |
62 | 0 | unsigned wsize; /* window size or zero if not using window */ |
63 | 0 | unsigned whave; /* valid bytes in the window */ |
64 | 0 | unsigned wnext; /* window write index */ |
65 | | |
66 | | /* hold is a local copy of strm->hold. By default, hold satisfies the same |
67 | | invariants that strm->hold does, namely that (hold >> bits) == 0. This |
68 | | invariant is kept by loading bits into hold one byte at a time, like: |
69 | | |
70 | | hold |= next_byte_of_input << bits; in++; bits += 8; |
71 | | |
72 | | If we need to ensure that bits >= 15 then this code snippet is simply |
73 | | repeated. Over one iteration of the outermost do/while loop, this |
74 | | happens up to six times (48 bits of input), as described in the NOTES |
75 | | above. |
76 | | |
77 | | However, on some little endian architectures, it can be significantly |
78 | | faster to load 64 bits once instead of 8 bits six times: |
79 | | |
80 | | if (bits <= 16) { |
81 | | hold |= next_8_bytes_of_input << bits; in += 6; bits += 48; |
82 | | } |
83 | | |
84 | | Unlike the simpler one byte load, shifting the next_8_bytes_of_input |
85 | | by bits will overflow and lose those high bits, up to 2 bytes' worth. |
86 | | The conservative estimate is therefore that we have read only 6 bytes |
87 | | (48 bits). Again, as per the NOTES above, 48 bits is sufficient for the |
88 | | rest of the iteration, and we will not need to load another 8 bytes. |
89 | | |
90 | | Inside this function, we no longer satisfy (hold >> bits) == 0, but |
91 | | this is not problematic, even if that overflow does not land on an 8 bit |
92 | | byte boundary. Those excess bits will eventually shift down lower as the |
93 | | Huffman decoder consumes input, and when new input bits need to be loaded |
94 | | into the bits variable, the same input bits will be or'ed over those |
95 | | existing bits. A bitwise or is idempotent: (a | b | b) equals (a | b). |
96 | | Note that we therefore write that load operation as "hold |= etc" and not |
97 | | "hold += etc". |
98 | | |
99 | | Outside that loop, at the end of the function, hold is bitwise and'ed |
100 | | with (1<<bits)-1 to drop those excess bits so that, on function exit, we |
101 | | keep the invariant that (state->hold >> state->bits) == 0. |
102 | | */ |
103 | 0 | bits_t bits; /* local strm->bits */ |
104 | 0 | uint64_t hold; /* local strm->hold */ |
105 | 0 | unsigned lmask; /* mask for first level of length codes */ |
106 | 0 | unsigned dmask; /* mask for first level of distance codes */ |
107 | 0 | code const *lcode; /* local strm->lencode */ |
108 | 0 | code const *dcode; /* local strm->distcode */ |
109 | 0 | code here; /* retrieved table entry */ |
110 | 0 | unsigned op; /* code bits, operation, extra bits, or */ |
111 | | /* window position, window bytes to copy */ |
112 | 0 | unsigned len; /* match length, unused bytes */ |
113 | 0 | unsigned char *from; /* where to copy match from */ |
114 | 0 | unsigned dist; /* match distance */ |
115 | 0 | uint64_t old; /* look-behind buffer for extra bits */ |
116 | | |
117 | | /* copy state to local variables */ |
118 | 0 | state = (struct inflate_state *)strm->state; |
119 | 0 | in = strm->next_in; |
120 | 0 | last = in + (strm->avail_in - (INFLATE_FAST_MIN_HAVE - 1)); |
121 | 0 | out = strm->next_out; |
122 | 0 | beg = out - (start - strm->avail_out); |
123 | 0 | safe = out + strm->avail_out; |
124 | 0 | end = safe - (safe_mode ? INFLATE_FAST_MIN_SAFE : INFLATE_FAST_MIN_LEFT) + 1; |
125 | 0 | wsize = state->wsize; |
126 | 0 | whave = state->whave; |
127 | 0 | wnext = state->wnext; |
128 | 0 | window = state->window; |
129 | 0 | hold = state->hold; |
130 | 0 | bits = (bits_t)state->bits; |
131 | 0 | lcode = state->lencode; |
132 | 0 | dcode = state->distcode; |
133 | 0 | lmask = (1U << state->lenbits) - 1; |
134 | 0 | dmask = (1U << state->distbits) - 1; |
135 | | |
136 | | /* decode literals and length/distances until end-of-block or not enough |
137 | | input data or output space */ |
138 | 0 | do { |
139 | 0 | REFILL(); |
140 | 0 | here = lcode[hold & lmask]; |
141 | 0 | Z_TOUCH(here); |
142 | 0 | old = hold; |
143 | 0 | DROPBITS(here.bits); |
144 | 0 | if (LIKELY(here.op == 0)) { |
145 | 0 | TRACE_LITERAL(here.val); |
146 | 0 | *out++ = (unsigned char)(here.val); |
147 | 0 | here = lcode[hold & lmask]; |
148 | 0 | Z_TOUCH(here); |
149 | 0 | old = hold; |
150 | 0 | DROPBITS(here.bits); |
151 | 0 | if (LIKELY(here.op == 0)) { |
152 | 0 | TRACE_LITERAL(here.val); |
153 | 0 | *out++ = (unsigned char)(here.val); |
154 | 0 | here = lcode[hold & lmask]; |
155 | 0 | Z_TOUCH(here); |
156 | 0 | dolen: |
157 | 0 | old = hold; |
158 | 0 | DROPBITS(here.bits); |
159 | 0 | if (LIKELY(here.op == 0)) { |
160 | 0 | TRACE_LITERAL(here.val); |
161 | 0 | *out++ = (unsigned char)(here.val); |
162 | 0 | continue; |
163 | 0 | } |
164 | 0 | } |
165 | 0 | } |
166 | 0 | op = here.op; |
167 | 0 | if (LIKELY(op & 16)) { /* length base */ |
168 | 0 | len = here.val + EXTRA_BITS(old, here, op); |
169 | 0 | TRACE_LENGTH(len); |
170 | 0 | here = dcode[hold & dmask]; |
171 | 0 | Z_TOUCH(here); |
172 | 0 | if (UNLIKELY(bits < MAX_BITS + MAX_DIST_EXTRA_BITS)) { |
173 | 0 | REFILL(); |
174 | 0 | } |
175 | 0 | dodist: |
176 | 0 | old = hold; |
177 | 0 | DROPBITS(here.bits); |
178 | 0 | op = here.op; |
179 | 0 | if (LIKELY(op & 16)) { /* distance base */ |
180 | 0 | dist = here.val + EXTRA_BITS(old, here, op); |
181 | | #ifdef INFLATE_STRICT |
182 | | if (UNLIKELY(dist > state->dmax)) { |
183 | | SET_BAD("invalid distance too far back"); |
184 | | break; |
185 | | } |
186 | | #endif |
187 | 0 | TRACE_DISTANCE(dist); |
188 | | |
189 | | /* In safe mode, if there isn't enough output space for the full copy, |
190 | | bail to the slow path's MATCH state which handles partial copies. */ |
191 | 0 | if (UNLIKELY(safe_mode && len > (unsigned)(safe - out))) { |
192 | 0 | state->mode = MATCH; |
193 | 0 | state->length = len; |
194 | 0 | state->offset = dist; |
195 | 0 | break; |
196 | 0 | } |
197 | | |
198 | 0 | op = (unsigned)(out - beg); /* max distance in output */ |
199 | 0 | if (UNLIKELY(dist > op)) { /* see if copy from window */ |
200 | 0 | op = dist - op; /* distance back in window */ |
201 | 0 | if (UNLIKELY(op > whave)) { |
202 | | #ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR |
203 | | if (LIKELY(state->sane)) { |
204 | | SET_BAD("invalid distance too far back"); |
205 | | break; |
206 | | } |
207 | | unsigned gap = op - whave; |
208 | | unsigned zeros = MIN(len, gap); |
209 | | memset(out, 0, zeros); /* fill missing bytes with zeros */ |
210 | | out += zeros; |
211 | | len -= zeros; |
212 | | if (UNLIKELY(len == 0)) |
213 | | continue; |
214 | | op = whave; |
215 | | if (UNLIKELY(op == 0)) {/* copy from already-decoded output */ |
216 | | out = chunkcopy_safe(out, out - dist, len, safe); |
217 | | continue; |
218 | | } |
219 | | #else |
220 | 0 | SET_BAD("invalid distance too far back"); |
221 | 0 | break; |
222 | 0 | #endif |
223 | 0 | } |
224 | 0 | from = window; |
225 | 0 | if (LIKELY(wnext == 0)) { /* very common case */ |
226 | 0 | from += wsize - op; |
227 | 0 | } else if (LIKELY(wnext >= op)) { /* contiguous in window */ |
228 | 0 | from += wnext - op; |
229 | 0 | } else { /* wrap around window */ |
230 | 0 | op -= wnext; |
231 | 0 | from += wsize - op; |
232 | 0 | if (UNLIKELY(op < len)) { /* some from end of window */ |
233 | 0 | len -= op; |
234 | 0 | out = CHUNKCOPY_SAFE(out, from, op, safe); |
235 | 0 | from = window; /* more from start of window */ |
236 | 0 | op = wnext; |
237 | | /* This (rare) case can create a situation where |
238 | | the first chunkcopy below must be checked. |
239 | | */ |
240 | 0 | } |
241 | 0 | } |
242 | 0 | if (UNLIKELY(op < len)) { /* still need some from output */ |
243 | 0 | len -= op; |
244 | 0 | if (LIKELY(!safe_mode)) { |
245 | 0 | out = CHUNKCOPY_SAFE(out, from, op, safe); |
246 | 0 | out = CHUNKUNROLL(out, &dist, &len); |
247 | 0 | out = CHUNKCOPY_SAFE(out, out - dist, len, safe); |
248 | 0 | } else { |
249 | | #ifdef HAVE_MASKED_READWRITE |
250 | | out = CHUNKCOPY_SAFE(out, from, op, safe); |
251 | | out = CHUNKCOPY_SAFE(out, out - dist, len, safe); |
252 | | #else |
253 | | out = chunkcopy_safe(out, from, op, safe); |
254 | | out = chunkcopy_safe(out, out - dist, len, safe); |
255 | | #endif |
256 | 0 | } |
257 | 0 | } else { |
258 | | #ifdef HAVE_MASKED_READWRITE |
259 | | out = CHUNKCOPY_SAFE(out, from, len, safe); |
260 | | #else |
261 | 0 | if (LIKELY(!safe_mode)) |
262 | 0 | out = CHUNKCOPY_SAFE(out, from, len, safe); |
263 | 0 | else |
264 | 0 | out = chunkcopy_safe(out, from, len, safe); |
265 | | #endif |
266 | 0 | } |
267 | 0 | } else if (LIKELY(!safe_mode)) { |
268 | | /* Whole reference is in range of current output. No range checks are |
269 | | necessary because we start with room for at least 258 bytes of output, |
270 | | so unroll and roundoff operations can write beyond `out+len` so long |
271 | | as they stay within 258 bytes of `out`. |
272 | | */ |
273 | 0 | if (LIKELY(dist >= len || dist >= CHUNKSIZE())) |
274 | 0 | out = CHUNKCOPY(out, out - dist, len); |
275 | 0 | else |
276 | 0 | out = CHUNKMEMSET(out, out - dist, len); |
277 | 0 | } else { |
278 | | #ifdef HAVE_MASKED_READWRITE |
279 | | out = CHUNKCOPY_SAFE(out, out - dist, len, safe); |
280 | | #else |
281 | | out = chunkcopy_safe(out, out - dist, len, safe); |
282 | | #endif |
283 | 0 | } |
284 | 0 | } else if (UNLIKELY((op & 64) == 0)) { /* 2nd level distance code */ |
285 | 0 | here = dcode[here.val + BITS(op)]; |
286 | 0 | Z_TOUCH(here); |
287 | 0 | goto dodist; |
288 | 0 | } else { |
289 | 0 | SET_BAD("invalid distance code"); |
290 | 0 | break; |
291 | 0 | } |
292 | 0 | } else if (UNLIKELY((op & 64) == 0)) { /* 2nd level length code */ |
293 | 0 | here = lcode[here.val + BITS(op)]; |
294 | 0 | Z_TOUCH(here); |
295 | 0 | goto dolen; |
296 | 0 | } else if (UNLIKELY(op & 32)) { /* end-of-block */ |
297 | 0 | TRACE_END_OF_BLOCK(); |
298 | 0 | state->mode = TYPE; |
299 | 0 | break; |
300 | 0 | } else { |
301 | 0 | SET_BAD("invalid literal/length code"); |
302 | 0 | break; |
303 | 0 | } |
304 | 0 | } while (in < last && out < end); |
305 | | |
306 | | /* return unused bytes (on entry, bits < 8, so in won't go too far back) */ |
307 | 0 | len = bits >> 3; |
308 | 0 | in -= len; |
309 | 0 | bits -= (bits_t)(len << 3); |
310 | 0 | hold &= (UINT64_C(1) << bits) - 1; |
311 | | |
312 | | /* update state and return */ |
313 | 0 | strm->next_in = in; |
314 | 0 | strm->next_out = out; |
315 | 0 | strm->avail_in = (unsigned)(in < last ? (INFLATE_FAST_MIN_HAVE - 1) + (last - in) |
316 | 0 | : (INFLATE_FAST_MIN_HAVE - 1) - (in - last)); |
317 | 0 | strm->avail_out = (unsigned)(safe - out); |
318 | |
|
319 | 0 | Assert(bits <= 32, "Remaining bits greater than 32"); |
320 | 0 | state->hold = (uint32_t)hold; |
321 | 0 | state->bits = bits; |
322 | 0 | return; |
323 | 0 | } Unexecuted instantiation: inflate_fast_sse2 Unexecuted instantiation: inflate_fast_ssse3 Unexecuted instantiation: inflate_fast_avx2 Unexecuted instantiation: inflate_fast_avx512 |
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 | | |
339 | | // Cleanup |
340 | | #undef CHUNKCOPY |
341 | | #undef CHUNKMEMSET |
342 | | #undef CHUNKMEMSET_SAFE |
343 | | #undef CHUNKSIZE |
344 | | #undef CHUNKUNROLL |
345 | | #undef HAVE_CHUNKCOPY |
346 | | #undef HAVE_CHUNKMEMSET_2 |
347 | | #undef HAVE_CHUNKMEMSET_4 |
348 | | #undef HAVE_CHUNKMEMSET_8 |
349 | | #undef HAVE_CHUNKMEMSET_16 |
350 | | #undef HAVE_CHUNK_MAG |
351 | | #undef HAVE_HALFCHUNKCOPY |
352 | | #undef HAVE_HALF_CHUNK |
353 | | #undef HAVE_MASKED_READWRITE |
354 | | #undef INFLATE_FAST |