Line | Count | Source |
1 | | /* match_tpl.h -- find longest match template for compare256 variants |
2 | | * |
3 | | * Copyright (C) 1995-2024 Jean-loup Gailly and Mark Adler |
4 | | * For conditions of distribution and use, see copyright notice in zlib.h |
5 | | * |
6 | | * Portions copyright (C) 2014-2021 Konstantin Nosov |
7 | | * Fast-zlib optimized longest_match |
8 | | * https://github.com/gildor2/fast_zlib |
9 | | */ |
10 | | |
11 | | #include "insert_string_p.h" |
12 | | |
13 | 20.7M | #define EARLY_EXIT_TRIGGER_LEVEL 5 |
14 | | |
15 | | #define GOTO_NEXT_CHAIN \ |
16 | 65.4M | if (--chain_length && (cur_match = prev[cur_match & wmask]) > limit) \ |
17 | 62.8M | continue; \ |
18 | 2.56M | return best_len; |
19 | | |
20 | | /* Set match_start to the longest match starting at the given string and |
21 | | * return its length. Matches shorter or equal to prev_length are discarded, |
22 | | * in which case the result is equal to prev_length and match_start is garbage. |
23 | | * |
24 | | * IN assertions: cur_match is the head of the hash chain for the current |
25 | | * string (strstart) and its distance is <= MAX_DIST, and prev_length >=1 |
26 | | * OUT assertion: the match length is not greater than s->lookahead |
27 | | */ |
28 | 60.5M | Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, uint32_t cur_match) { |
29 | 60.5M | const unsigned wmask = W_MASK(s); |
30 | 60.5M | unsigned int strstart = s->strstart; |
31 | 60.5M | const unsigned char *window = s->window; |
32 | 60.5M | const Pos *prev = s->prev; |
33 | | #ifdef LONGEST_MATCH_SLOW |
34 | | const Pos *head = s->head; |
35 | | #endif |
36 | 60.5M | const unsigned char *scan; |
37 | 60.5M | const unsigned char *mbase_start = window; |
38 | 60.5M | const unsigned char *mbase_end; |
39 | 60.5M | uint32_t limit; |
40 | | #ifdef LONGEST_MATCH_SLOW |
41 | | uint32_t limit_base; |
42 | | #endif |
43 | | #ifndef LONGEST_MATCH_SLOW |
44 | | int32_t early_exit; |
45 | | #endif |
46 | 60.5M | uint32_t chain_length = s->max_chain_length; |
47 | 60.5M | uint32_t nice_match = (uint32_t)s->nice_match; |
48 | 60.5M | uint32_t best_len, offset; |
49 | 60.5M | uint32_t lookahead = s->lookahead; |
50 | 60.5M | uint32_t match_offset = 0; |
51 | 60.5M | uint64_t scan_start; |
52 | 60.5M | uint64_t scan_end; |
53 | | |
54 | | /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */ |
55 | 60.5M | Assert(STD_MAX_MATCH == 258, "Code too clever"); |
56 | | |
57 | 60.5M | best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1; |
58 | 60.5M | if (UNLIKELY(best_len >= lookahead)) |
59 | 599 | return lookahead; |
60 | | #ifdef LONGEST_MATCH_SLOW |
61 | | # ifdef LONGEST_MATCH_SLOW_ROLL |
62 | | /* Rolling-hash variant always runs the post-match offset search; the |
63 | | * compiler folds the constant away below. |
64 | | */ |
65 | 34.7M | const int offset_search = 1; |
66 | | # else |
67 | | /* The post-match offset search only pays off when we entered with a |
68 | | * prior best_len from lazy evaluation, so gate on it at function entry. |
69 | | */ |
70 | 5.05M | const int offset_search = (best_len >= STD_MIN_MATCH); |
71 | | # endif |
72 | 5.05M | #endif |
73 | | |
74 | | /* Calculate read offset which should only extend an extra byte to find the |
75 | | * next best match length. When best_len is shorter than the read width, we |
76 | | * diff the mismatched bytes instead. |
77 | | */ |
78 | 60.5M | offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0; |
79 | | |
80 | 20.7M | scan = window + strstart; |
81 | 20.7M | scan_start = zng_memread_8(scan); |
82 | 20.7M | scan_end = zng_memread_8(scan+offset); |
83 | 20.7M | mbase_end = (mbase_start+offset); |
84 | | |
85 | | /* Do not waste too much time if we already have a good match */ |
86 | 60.5M | if (UNLIKELY(best_len >= s->good_match)) |
87 | 118k | chain_length >>= 2; |
88 | | |
89 | | /* Stop when cur_match becomes <= limit. To simplify the code, |
90 | | * we prevent matches with the string of window index 0 |
91 | | */ |
92 | 60.5M | limit = strstart > MAX_DIST(s) ? (strstart - MAX_DIST(s)) : 0; |
93 | | #ifndef LONGEST_MATCH_SLOW |
94 | 20.7M | early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL; |
95 | | #endif |
96 | | #ifdef LONGEST_MATCH_SLOW |
97 | | limit_base = limit; |
98 | 39.7M | if (best_len >= STD_MIN_MATCH) { |
99 | | /* We're continuing search (lazy evaluation). Find a most distant |
100 | | * chain by hashing substrings within the match area. We cannot use |
101 | | * s->prev[strstart+1,...] immediately because those strings are not |
102 | | * yet inserted into the hash table. |
103 | | */ |
104 | | # ifdef LONGEST_MATCH_SLOW_ROLL |
105 | | uint32_t hash; |
106 | | uint32_t pos; |
107 | | |
108 | | hash = update_hash_roll(0, scan[1]); |
109 | | hash = update_hash_roll(hash, scan[2]); |
110 | | |
111 | 25.8M | for (uint32_t i = 3; i <= best_len; i++) { |
112 | 19.3M | hash = update_hash_roll(hash, scan[i]); |
113 | 19.3M | pos = head[hash]; |
114 | 19.3M | if (UNLIKELY(pos < cur_match)) { |
115 | 5.24M | match_offset = i - 2; |
116 | 5.24M | cur_match = pos; |
117 | 5.24M | } |
118 | 19.3M | } |
119 | | # else /* 4-byte Knuth hash, fresh lookup per offset */ |
120 | 1.93M | for (uint32_t i = 1; i + (WANT_MIN_MATCH - 1) <= best_len; i++) { |
121 | 1.70M | uint32_t val = Z_U32_FROM_LE(zng_memread_4(scan + i)); |
122 | 1.70M | uint32_t hash; |
123 | 1.70M | UPDATE_HASH_KNUTH(hash, val); |
124 | 1.70M | uint32_t pos = head[hash]; |
125 | 1.70M | if (pos < cur_match) { |
126 | 276k | match_offset = i; |
127 | 276k | cur_match = pos; |
128 | 276k | } |
129 | 1.70M | } |
130 | | # endif |
131 | | |
132 | | /* Update offset-dependent variables */ |
133 | 6.67M | limit = limit_base+match_offset; |
134 | 6.67M | if (UNLIKELY(cur_match <= limit)) |
135 | 1.59M | return best_len; |
136 | 5.07M | mbase_start -= match_offset; |
137 | 5.07M | mbase_end -= match_offset; |
138 | 5.07M | } |
139 | 38.1M | #endif |
140 | 38.1M | Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead"); |
141 | 73.6M | for (;;) { |
142 | 73.6M | if (UNLIKELY(cur_match >= strstart)) |
143 | 82 | break; |
144 | | |
145 | | /* Skip to next match if the match length cannot increase or if the match length is |
146 | | * less than 2. Note that the checks below for insufficient lookahead only occur |
147 | | * occasionally for performance reasons. |
148 | | * Therefore uninitialized memory will be accessed and conditional jumps will be made |
149 | | * that depend on those values. However the length of the match is limited to the |
150 | | * lookahead, so the output of deflate is not affected by the uninitialized values. |
151 | | */ |
152 | 73.6M | uint32_t len; |
153 | 73.6M | if (best_len < sizeof(uint64_t)) { |
154 | 69.0M | uint64_t cand_start = zng_memread_8(mbase_start + cur_match); |
155 | 69.0M | if (scan_start != cand_start) { |
156 | | /* Peel the first candidate out of the loop. A full 8-byte match falls straight |
157 | | * through to compare256, and single-candidate chains (barely-compressible data) |
158 | | * run with no loop overhead. */ |
159 | 67.9M | uint64_t first_mask = zng_first_bytes_mask64(best_len + 1); |
160 | 67.9M | uint64_t diff = scan_start ^ cand_start; |
161 | | /* A candidate beats best_len only when its first best_len+1 bytes match, i.e. |
162 | | * those bytes of the XOR are zero. The masked test rejects without running ctz. */ |
163 | 67.9M | if (UNLIKELY((diff & first_mask) == 0)) { |
164 | 10.1M | len = zng_first_diff_byte64(diff); |
165 | 10.1M | goto short_match_accept; |
166 | 10.1M | } |
167 | 57.8M | if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit) |
168 | 30.4M | return best_len; |
169 | 27.3M | cand_start = zng_memread_8(mbase_start + cur_match); |
170 | 27.3M | if (scan_start != cand_start) { |
171 | | /* Walk the remaining candidates with the chain advance kept inline. */ |
172 | 156M | for (;;) { |
173 | 156M | diff = scan_start ^ cand_start; |
174 | 156M | if (UNLIKELY((diff & first_mask) == 0)) { |
175 | 3.49M | len = zng_first_diff_byte64(diff); |
176 | 3.49M | goto short_match_accept; |
177 | 3.49M | } |
178 | 152M | if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit) |
179 | 23.7M | return best_len; |
180 | 129M | cand_start = zng_memread_8(mbase_start + cur_match); |
181 | 129M | if (scan_start == cand_start) |
182 | 98.4k | break; |
183 | 129M | } |
184 | 27.2M | } |
185 | 27.3M | } |
186 | | /* All 8 bytes match, fallthrough to compare256 for the tail. */ |
187 | 69.0M | } else { |
188 | | /* Pre-filter the candidate on the start and end sentinels before compare256 using |
189 | | * simple 8-byte comparison since best_len >= 8. */ |
190 | 56.1M | for (;;) { |
191 | | /* First check the end of the candidate at best_len+1 due to the higher |
192 | | * likelihood of a mismatch. */ |
193 | 56.1M | if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 && |
194 | 5.46M | zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0) |
195 | 3.91M | break; |
196 | 52.1M | GOTO_NEXT_CHAIN; |
197 | 52.1M | } |
198 | 4.55M | } |
199 | 5.15M | len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2; |
200 | 5.15M | Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan"); |
201 | | |
202 | 5.15M | if (len > best_len) |
203 | 19.6M | short_match_accept: |
204 | 19.6M | { |
205 | 19.6M | uint32_t match_start = cur_match - match_offset; |
206 | 19.6M | s->match_start = match_start; |
207 | | |
208 | | /* Do not look for better matches if the current match reaches |
209 | | * or exceeds the end of the input. |
210 | | */ |
211 | 19.6M | if (UNLIKELY(len >= lookahead)) |
212 | 1.29k | return lookahead; |
213 | 16.6M | if (UNLIKELY(len >= nice_match)) |
214 | 441k | return len; |
215 | | |
216 | 16.2M | best_len = len; |
217 | | |
218 | 16.2M | offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0; |
219 | | |
220 | 16.2M | scan_end = zng_memread_8(scan+offset); |
221 | | |
222 | | #ifdef LONGEST_MATCH_SLOW |
223 | | /* Look for a better string offset */ |
224 | 11.9M | if (UNLIKELY(offset_search && len > STD_MIN_MATCH && match_start + len < strstart)) { |
225 | 5.14M | const unsigned char *scan_endstr; |
226 | 5.14M | uint32_t hash; |
227 | 5.14M | uint32_t pos, next_pos; |
228 | | |
229 | | /* Go back to offset 0 */ |
230 | | cur_match -= match_offset; |
231 | | match_offset = 0; |
232 | | next_pos = cur_match; |
233 | | |
234 | | /* Walk prev[] for positions inside the match. The bound keeps |
235 | | * the hash window within the match, so we follow chains that |
236 | | * share bytes with the current match, not data past its end. |
237 | | */ |
238 | | # ifdef LONGEST_MATCH_SLOW_ROLL |
239 | 80.2M | for (uint32_t i = 0; i <= len - STD_MIN_MATCH; i++) { |
240 | | # else /* 4-byte Knuth hash needs len - 4 bound */ |
241 | 770k | for (uint32_t i = 0; i <= len - WANT_MIN_MATCH; i++) { |
242 | 740k | # endif |
243 | 740k | pos = prev[(cur_match + i) & wmask]; |
244 | 77.2M | if (UNLIKELY(pos < next_pos)) { |
245 | | /* Hash chain is more distant, use it */ |
246 | 6.92M | if (UNLIKELY(pos <= limit_base + i)) |
247 | 1.38M | return best_len; |
248 | 5.54M | next_pos = pos; |
249 | 5.54M | match_offset = i; |
250 | 5.54M | } |
251 | 740k | } |
252 | | /* Switch cur_match to next_pos chain */ |
253 | 3.76M | cur_match = next_pos; |
254 | | |
255 | | /* Try hash head at a window covering the tail of the current |
256 | | * match to find chains that extend farther back. The window |
257 | | * width differs per variant: 3 bytes for rolling, 4 for Knuth. |
258 | | */ |
259 | | # ifdef LONGEST_MATCH_SLOW_ROLL |
260 | 3.73M | scan_endstr = scan + len - (STD_MIN_MATCH-1); |
261 | | |
262 | | hash = update_hash_roll(0, scan_endstr[0]); |
263 | | hash = update_hash_roll(hash, scan_endstr[1]); |
264 | | hash = update_hash_roll(hash, scan_endstr[2]); |
265 | | |
266 | | pos = head[hash]; |
267 | 3.73M | if (UNLIKELY(pos < cur_match)) { |
268 | 800k | match_offset = len - (STD_MIN_MATCH-1); |
269 | 800k | if (pos <= limit_base + match_offset) |
270 | 370k | return best_len; |
271 | 430k | cur_match = pos; |
272 | 430k | } |
273 | | # else |
274 | 30.1k | scan_endstr = scan + len - WANT_MIN_MATCH; |
275 | 30.1k | uint32_t val = Z_U32_FROM_LE(zng_memread_4(scan_endstr)); |
276 | 30.1k | UPDATE_HASH_KNUTH(hash, val); |
277 | | |
278 | | pos = head[hash]; |
279 | 30.1k | if (pos < cur_match) { |
280 | 0 | match_offset = len - WANT_MIN_MATCH; |
281 | 0 | if (pos <= limit_base + match_offset) |
282 | 0 | return best_len; |
283 | 0 | cur_match = pos; |
284 | 0 | } |
285 | 30.1k | # endif |
286 | | |
287 | | /* Update offset-dependent variables */ |
288 | 3.39M | limit = limit_base+match_offset; |
289 | 3.39M | mbase_start = window-match_offset; |
290 | 3.39M | mbase_end = (mbase_start+offset); |
291 | 3.39M | continue; |
292 | 3.76M | } |
293 | 6.79M | #endif |
294 | 6.79M | mbase_end = (mbase_start+offset); |
295 | 6.79M | } |
296 | | #ifndef LONGEST_MATCH_SLOW |
297 | 16.6k | else if (UNLIKELY(early_exit)) { |
298 | | /* The probability of finding a match later if we here is pretty low, so for |
299 | | * performance it's best to outright stop here for the lower compression levels |
300 | | */ |
301 | 482 | break; |
302 | 482 | } |
303 | 4.29M | #endif |
304 | 13.2M | GOTO_NEXT_CHAIN; |
305 | 13.2M | } |
306 | 564 | return best_len; |
307 | 38.1M | } Unexecuted instantiation: longest_match_sse2 Unexecuted instantiation: longest_match_slow_knuth_sse2 Unexecuted instantiation: longest_match_slow_roll_sse2 Line | Count | Source | 28 | 20.7M | Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, uint32_t cur_match) { | 29 | 20.7M | const unsigned wmask = W_MASK(s); | 30 | 20.7M | unsigned int strstart = s->strstart; | 31 | 20.7M | const unsigned char *window = s->window; | 32 | 20.7M | const Pos *prev = s->prev; | 33 | | #ifdef LONGEST_MATCH_SLOW | 34 | | const Pos *head = s->head; | 35 | | #endif | 36 | 20.7M | const unsigned char *scan; | 37 | 20.7M | const unsigned char *mbase_start = window; | 38 | 20.7M | const unsigned char *mbase_end; | 39 | 20.7M | uint32_t limit; | 40 | | #ifdef LONGEST_MATCH_SLOW | 41 | | uint32_t limit_base; | 42 | | #endif | 43 | 20.7M | #ifndef LONGEST_MATCH_SLOW | 44 | 20.7M | int32_t early_exit; | 45 | 20.7M | #endif | 46 | 20.7M | uint32_t chain_length = s->max_chain_length; | 47 | 20.7M | uint32_t nice_match = (uint32_t)s->nice_match; | 48 | 20.7M | uint32_t best_len, offset; | 49 | 20.7M | uint32_t lookahead = s->lookahead; | 50 | 20.7M | uint32_t match_offset = 0; | 51 | 20.7M | uint64_t scan_start; | 52 | 20.7M | uint64_t scan_end; | 53 | | | 54 | | /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */ | 55 | 20.7M | Assert(STD_MAX_MATCH == 258, "Code too clever"); | 56 | | | 57 | 20.7M | best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1; | 58 | 20.7M | if (UNLIKELY(best_len >= lookahead)) | 59 | 0 | return lookahead; | 60 | | #ifdef LONGEST_MATCH_SLOW | 61 | | # ifdef LONGEST_MATCH_SLOW_ROLL | 62 | | /* Rolling-hash variant always runs the post-match offset search; the | 63 | | * compiler folds the constant away below. | 64 | | */ | 65 | | const int offset_search = 1; | 66 | | # else | 67 | | /* The post-match offset search only pays off when we entered with a | 68 | | * prior best_len from lazy evaluation, so gate on it at function entry. | 69 | | */ | 70 | | const int offset_search = (best_len >= STD_MIN_MATCH); | 71 | | # endif | 72 | | #endif | 73 | | | 74 | | /* Calculate read offset which should only extend an extra byte to find the | 75 | | * next best match length. When best_len is shorter than the read width, we | 76 | | * diff the mismatched bytes instead. | 77 | | */ | 78 | 20.7M | offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0; | 79 | | | 80 | 20.7M | scan = window + strstart; | 81 | 20.7M | scan_start = zng_memread_8(scan); | 82 | 20.7M | scan_end = zng_memread_8(scan+offset); | 83 | 20.7M | mbase_end = (mbase_start+offset); | 84 | | | 85 | | /* Do not waste too much time if we already have a good match */ | 86 | 20.7M | if (UNLIKELY(best_len >= s->good_match)) | 87 | 0 | chain_length >>= 2; | 88 | | | 89 | | /* Stop when cur_match becomes <= limit. To simplify the code, | 90 | | * we prevent matches with the string of window index 0 | 91 | | */ | 92 | 20.7M | limit = strstart > MAX_DIST(s) ? (strstart - MAX_DIST(s)) : 0; | 93 | 20.7M | #ifndef LONGEST_MATCH_SLOW | 94 | 20.7M | early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL; | 95 | 20.7M | #endif | 96 | | #ifdef LONGEST_MATCH_SLOW | 97 | | limit_base = limit; | 98 | | if (best_len >= STD_MIN_MATCH) { | 99 | | /* We're continuing search (lazy evaluation). Find a most distant | 100 | | * chain by hashing substrings within the match area. We cannot use | 101 | | * s->prev[strstart+1,...] immediately because those strings are not | 102 | | * yet inserted into the hash table. | 103 | | */ | 104 | | # ifdef LONGEST_MATCH_SLOW_ROLL | 105 | | uint32_t hash; | 106 | | uint32_t pos; | 107 | | | 108 | | hash = update_hash_roll(0, scan[1]); | 109 | | hash = update_hash_roll(hash, scan[2]); | 110 | | | 111 | | for (uint32_t i = 3; i <= best_len; i++) { | 112 | | hash = update_hash_roll(hash, scan[i]); | 113 | | pos = head[hash]; | 114 | | if (UNLIKELY(pos < cur_match)) { | 115 | | match_offset = i - 2; | 116 | | cur_match = pos; | 117 | | } | 118 | | } | 119 | | # else /* 4-byte Knuth hash, fresh lookup per offset */ | 120 | | for (uint32_t i = 1; i + (WANT_MIN_MATCH - 1) <= best_len; i++) { | 121 | | uint32_t val = Z_U32_FROM_LE(zng_memread_4(scan + i)); | 122 | | uint32_t hash; | 123 | | UPDATE_HASH_KNUTH(hash, val); | 124 | | uint32_t pos = head[hash]; | 125 | | if (pos < cur_match) { | 126 | | match_offset = i; | 127 | | cur_match = pos; | 128 | | } | 129 | | } | 130 | | # endif | 131 | | | 132 | | /* Update offset-dependent variables */ | 133 | | limit = limit_base+match_offset; | 134 | | if (UNLIKELY(cur_match <= limit)) | 135 | | return best_len; | 136 | | mbase_start -= match_offset; | 137 | | mbase_end -= match_offset; | 138 | | } | 139 | | #endif | 140 | 20.7M | Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead"); | 141 | 24.1M | for (;;) { | 142 | 24.1M | if (UNLIKELY(cur_match >= strstart)) | 143 | 82 | break; | 144 | | | 145 | | /* Skip to next match if the match length cannot increase or if the match length is | 146 | | * less than 2. Note that the checks below for insufficient lookahead only occur | 147 | | * occasionally for performance reasons. | 148 | | * Therefore uninitialized memory will be accessed and conditional jumps will be made | 149 | | * that depend on those values. However the length of the match is limited to the | 150 | | * lookahead, so the output of deflate is not affected by the uninitialized values. | 151 | | */ | 152 | 24.1M | uint32_t len; | 153 | 24.1M | if (best_len < sizeof(uint64_t)) { | 154 | 23.2M | uint64_t cand_start = zng_memread_8(mbase_start + cur_match); | 155 | 23.2M | if (scan_start != cand_start) { | 156 | | /* Peel the first candidate out of the loop. A full 8-byte match falls straight | 157 | | * through to compare256, and single-candidate chains (barely-compressible data) | 158 | | * run with no loop overhead. */ | 159 | 22.5M | uint64_t first_mask = zng_first_bytes_mask64(best_len + 1); | 160 | 22.5M | uint64_t diff = scan_start ^ cand_start; | 161 | | /* A candidate beats best_len only when its first best_len+1 bytes match, i.e. | 162 | | * those bytes of the XOR are zero. The masked test rejects without running ctz. */ | 163 | 22.5M | if (UNLIKELY((diff & first_mask) == 0)) { | 164 | 3.04M | len = zng_first_diff_byte64(diff); | 165 | 3.04M | goto short_match_accept; | 166 | 3.04M | } | 167 | 19.4M | if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit) | 168 | 13.8M | return best_len; | 169 | 5.64M | cand_start = zng_memread_8(mbase_start + cur_match); | 170 | 5.64M | if (scan_start != cand_start) { | 171 | | /* Walk the remaining candidates with the chain advance kept inline. */ | 172 | 13.4M | for (;;) { | 173 | 13.4M | diff = scan_start ^ cand_start; | 174 | 13.4M | if (UNLIKELY((diff & first_mask) == 0)) { | 175 | 189k | len = zng_first_diff_byte64(diff); | 176 | 189k | goto short_match_accept; | 177 | 189k | } | 178 | 13.2M | if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit) | 179 | 5.39M | return best_len; | 180 | 7.84M | cand_start = zng_memread_8(mbase_start + cur_match); | 181 | 7.84M | if (scan_start == cand_start) | 182 | 33.5k | break; | 183 | 7.84M | } | 184 | 5.62M | } | 185 | 5.64M | } | 186 | | /* All 8 bytes match, fallthrough to compare256 for the tail. */ | 187 | 23.2M | } else { | 188 | | /* Pre-filter the candidate on the start and end sentinels before compare256 using | 189 | | * simple 8-byte comparison since best_len >= 8. */ | 190 | 3.32M | for (;;) { | 191 | | /* First check the end of the candidate at best_len+1 due to the higher | 192 | | * likelihood of a mismatch. */ | 193 | 3.32M | if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 && | 194 | 657k | zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0) | 195 | 642k | break; | 196 | 2.67M | GOTO_NEXT_CHAIN; | 197 | 2.67M | } | 198 | 864k | } | 199 | 1.41M | len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2; | 200 | 1.41M | Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan"); | 201 | | | 202 | 1.41M | if (len > best_len) | 203 | 6.02M | short_match_accept: | 204 | 6.02M | { | 205 | 6.02M | uint32_t match_start = cur_match - match_offset; | 206 | 6.02M | s->match_start = match_start; | 207 | | | 208 | | /* Do not look for better matches if the current match reaches | 209 | | * or exceeds the end of the input. | 210 | | */ | 211 | 6.02M | if (UNLIKELY(len >= lookahead)) | 212 | 651 | return lookahead; | 213 | 4.62M | if (UNLIKELY(len >= nice_match)) | 214 | 351k | return len; | 215 | | | 216 | 4.27M | best_len = len; | 217 | | | 218 | 4.27M | offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0; | 219 | | | 220 | 4.27M | scan_end = zng_memread_8(scan+offset); | 221 | | | 222 | | #ifdef LONGEST_MATCH_SLOW | 223 | | /* Look for a better string offset */ | 224 | | if (UNLIKELY(offset_search && len > STD_MIN_MATCH && match_start + len < strstart)) { | 225 | | const unsigned char *scan_endstr; | 226 | | uint32_t hash; | 227 | | uint32_t pos, next_pos; | 228 | | | 229 | | /* Go back to offset 0 */ | 230 | | cur_match -= match_offset; | 231 | | match_offset = 0; | 232 | | next_pos = cur_match; | 233 | | | 234 | | /* Walk prev[] for positions inside the match. The bound keeps | 235 | | * the hash window within the match, so we follow chains that | 236 | | * share bytes with the current match, not data past its end. | 237 | | */ | 238 | | # ifdef LONGEST_MATCH_SLOW_ROLL | 239 | | for (uint32_t i = 0; i <= len - STD_MIN_MATCH; i++) { | 240 | | # else /* 4-byte Knuth hash needs len - 4 bound */ | 241 | | for (uint32_t i = 0; i <= len - WANT_MIN_MATCH; i++) { | 242 | | # endif | 243 | | pos = prev[(cur_match + i) & wmask]; | 244 | | if (UNLIKELY(pos < next_pos)) { | 245 | | /* Hash chain is more distant, use it */ | 246 | | if (UNLIKELY(pos <= limit_base + i)) | 247 | | return best_len; | 248 | | next_pos = pos; | 249 | | match_offset = i; | 250 | | } | 251 | | } | 252 | | /* Switch cur_match to next_pos chain */ | 253 | | cur_match = next_pos; | 254 | | | 255 | | /* Try hash head at a window covering the tail of the current | 256 | | * match to find chains that extend farther back. The window | 257 | | * width differs per variant: 3 bytes for rolling, 4 for Knuth. | 258 | | */ | 259 | | # ifdef LONGEST_MATCH_SLOW_ROLL | 260 | | scan_endstr = scan + len - (STD_MIN_MATCH-1); | 261 | | | 262 | | hash = update_hash_roll(0, scan_endstr[0]); | 263 | | hash = update_hash_roll(hash, scan_endstr[1]); | 264 | | hash = update_hash_roll(hash, scan_endstr[2]); | 265 | | | 266 | | pos = head[hash]; | 267 | | if (UNLIKELY(pos < cur_match)) { | 268 | | match_offset = len - (STD_MIN_MATCH-1); | 269 | | if (pos <= limit_base + match_offset) | 270 | | return best_len; | 271 | | cur_match = pos; | 272 | | } | 273 | | # else | 274 | | scan_endstr = scan + len - WANT_MIN_MATCH; | 275 | | uint32_t val = Z_U32_FROM_LE(zng_memread_4(scan_endstr)); | 276 | | UPDATE_HASH_KNUTH(hash, val); | 277 | | | 278 | | pos = head[hash]; | 279 | | if (pos < cur_match) { | 280 | | match_offset = len - WANT_MIN_MATCH; | 281 | | if (pos <= limit_base + match_offset) | 282 | | return best_len; | 283 | | cur_match = pos; | 284 | | } | 285 | | # endif | 286 | | | 287 | | /* Update offset-dependent variables */ | 288 | | limit = limit_base+match_offset; | 289 | | mbase_start = window-match_offset; | 290 | | mbase_end = (mbase_start+offset); | 291 | | continue; | 292 | | } | 293 | | #endif | 294 | 4.27M | mbase_end = (mbase_start+offset); | 295 | 4.27M | } | 296 | 16.6k | #ifndef LONGEST_MATCH_SLOW | 297 | 16.6k | else if (UNLIKELY(early_exit)) { | 298 | | /* The probability of finding a match later if we here is pretty low, so for | 299 | | * performance it's best to outright stop here for the lower compression levels | 300 | | */ | 301 | 482 | break; | 302 | 482 | } | 303 | 4.29M | #endif | 304 | 4.29M | GOTO_NEXT_CHAIN; | 305 | 4.29M | } | 306 | 564 | return best_len; | 307 | 20.7M | } |
longest_match_slow_knuth_avx2 Line | Count | Source | 28 | 5.06M | Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, uint32_t cur_match) { | 29 | 5.06M | const unsigned wmask = W_MASK(s); | 30 | 5.06M | unsigned int strstart = s->strstart; | 31 | 5.06M | const unsigned char *window = s->window; | 32 | 5.06M | const Pos *prev = s->prev; | 33 | 5.06M | #ifdef LONGEST_MATCH_SLOW | 34 | 5.06M | const Pos *head = s->head; | 35 | 5.06M | #endif | 36 | 5.06M | const unsigned char *scan; | 37 | 5.06M | const unsigned char *mbase_start = window; | 38 | 5.06M | const unsigned char *mbase_end; | 39 | 5.06M | uint32_t limit; | 40 | 5.06M | #ifdef LONGEST_MATCH_SLOW | 41 | 5.06M | uint32_t limit_base; | 42 | 5.06M | #endif | 43 | | #ifndef LONGEST_MATCH_SLOW | 44 | | int32_t early_exit; | 45 | | #endif | 46 | 5.06M | uint32_t chain_length = s->max_chain_length; | 47 | 5.06M | uint32_t nice_match = (uint32_t)s->nice_match; | 48 | 5.06M | uint32_t best_len, offset; | 49 | 5.06M | uint32_t lookahead = s->lookahead; | 50 | 5.06M | uint32_t match_offset = 0; | 51 | 5.06M | uint64_t scan_start; | 52 | 5.06M | uint64_t scan_end; | 53 | | | 54 | | /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */ | 55 | 5.06M | Assert(STD_MAX_MATCH == 258, "Code too clever"); | 56 | | | 57 | 5.06M | best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1; | 58 | 5.06M | if (UNLIKELY(best_len >= lookahead)) | 59 | 165 | return lookahead; | 60 | 5.05M | #ifdef LONGEST_MATCH_SLOW | 61 | | # ifdef LONGEST_MATCH_SLOW_ROLL | 62 | | /* Rolling-hash variant always runs the post-match offset search; the | 63 | | * compiler folds the constant away below. | 64 | | */ | 65 | | const int offset_search = 1; | 66 | | # else | 67 | | /* The post-match offset search only pays off when we entered with a | 68 | | * prior best_len from lazy evaluation, so gate on it at function entry. | 69 | | */ | 70 | 5.05M | const int offset_search = (best_len >= STD_MIN_MATCH); | 71 | 5.05M | # endif | 72 | 5.05M | #endif | 73 | | | 74 | | /* Calculate read offset which should only extend an extra byte to find the | 75 | | * next best match length. When best_len is shorter than the read width, we | 76 | | * diff the mismatched bytes instead. | 77 | | */ | 78 | 5.05M | offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0; | 79 | | | 80 | 5.05M | scan = window + strstart; | 81 | 5.05M | scan_start = zng_memread_8(scan); | 82 | 5.05M | scan_end = zng_memread_8(scan+offset); | 83 | 5.05M | mbase_end = (mbase_start+offset); | 84 | | | 85 | | /* Do not waste too much time if we already have a good match */ | 86 | 5.05M | if (UNLIKELY(best_len >= s->good_match)) | 87 | 47.7k | chain_length >>= 2; | 88 | | | 89 | | /* Stop when cur_match becomes <= limit. To simplify the code, | 90 | | * we prevent matches with the string of window index 0 | 91 | | */ | 92 | 5.05M | limit = strstart > MAX_DIST(s) ? (strstart - MAX_DIST(s)) : 0; | 93 | | #ifndef LONGEST_MATCH_SLOW | 94 | | early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL; | 95 | | #endif | 96 | 5.05M | #ifdef LONGEST_MATCH_SLOW | 97 | 5.05M | limit_base = limit; | 98 | 5.05M | if (best_len >= STD_MIN_MATCH) { | 99 | | /* We're continuing search (lazy evaluation). Find a most distant | 100 | | * chain by hashing substrings within the match area. We cannot use | 101 | | * s->prev[strstart+1,...] immediately because those strings are not | 102 | | * yet inserted into the hash table. | 103 | | */ | 104 | | # ifdef LONGEST_MATCH_SLOW_ROLL | 105 | | uint32_t hash; | 106 | | uint32_t pos; | 107 | | | 108 | | hash = update_hash_roll(0, scan[1]); | 109 | | hash = update_hash_roll(hash, scan[2]); | 110 | | | 111 | | for (uint32_t i = 3; i <= best_len; i++) { | 112 | | hash = update_hash_roll(hash, scan[i]); | 113 | | pos = head[hash]; | 114 | | if (UNLIKELY(pos < cur_match)) { | 115 | | match_offset = i - 2; | 116 | | cur_match = pos; | 117 | | } | 118 | | } | 119 | | # else /* 4-byte Knuth hash, fresh lookup per offset */ | 120 | 1.93M | for (uint32_t i = 1; i + (WANT_MIN_MATCH - 1) <= best_len; i++) { | 121 | 1.70M | uint32_t val = Z_U32_FROM_LE(zng_memread_4(scan + i)); | 122 | 1.70M | uint32_t hash; | 123 | 1.70M | UPDATE_HASH_KNUTH(hash, val); | 124 | 1.70M | uint32_t pos = head[hash]; | 125 | 1.70M | if (pos < cur_match) { | 126 | 276k | match_offset = i; | 127 | 276k | cur_match = pos; | 128 | 276k | } | 129 | 1.70M | } | 130 | 230k | # endif | 131 | | | 132 | | /* Update offset-dependent variables */ | 133 | 230k | limit = limit_base+match_offset; | 134 | 230k | if (UNLIKELY(cur_match <= limit)) | 135 | 122k | return best_len; | 136 | 107k | mbase_start -= match_offset; | 137 | 107k | mbase_end -= match_offset; | 138 | 107k | } | 139 | 4.93M | #endif | 140 | 4.93M | Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead"); | 141 | 5.51M | for (;;) { | 142 | 5.51M | if (UNLIKELY(cur_match >= strstart)) | 143 | 0 | break; | 144 | | | 145 | | /* Skip to next match if the match length cannot increase or if the match length is | 146 | | * less than 2. Note that the checks below for insufficient lookahead only occur | 147 | | * occasionally for performance reasons. | 148 | | * Therefore uninitialized memory will be accessed and conditional jumps will be made | 149 | | * that depend on those values. However the length of the match is limited to the | 150 | | * lookahead, so the output of deflate is not affected by the uninitialized values. | 151 | | */ | 152 | 5.51M | uint32_t len; | 153 | 5.51M | if (best_len < sizeof(uint64_t)) { | 154 | 5.08M | uint64_t cand_start = zng_memread_8(mbase_start + cur_match); | 155 | 5.08M | if (scan_start != cand_start) { | 156 | | /* Peel the first candidate out of the loop. A full 8-byte match falls straight | 157 | | * through to compare256, and single-candidate chains (barely-compressible data) | 158 | | * run with no loop overhead. */ | 159 | 4.97M | uint64_t first_mask = zng_first_bytes_mask64(best_len + 1); | 160 | 4.97M | uint64_t diff = scan_start ^ cand_start; | 161 | | /* A candidate beats best_len only when its first best_len+1 bytes match, i.e. | 162 | | * those bytes of the XOR are zero. The masked test rejects without running ctz. */ | 163 | 4.97M | if (UNLIKELY((diff & first_mask) == 0)) { | 164 | 254k | len = zng_first_diff_byte64(diff); | 165 | 254k | goto short_match_accept; | 166 | 254k | } | 167 | 4.71M | if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit) | 168 | 3.50M | return best_len; | 169 | 1.21M | cand_start = zng_memread_8(mbase_start + cur_match); | 170 | 1.21M | if (scan_start != cand_start) { | 171 | | /* Walk the remaining candidates with the chain advance kept inline. */ | 172 | 5.13M | for (;;) { | 173 | 5.13M | diff = scan_start ^ cand_start; | 174 | 5.13M | if (UNLIKELY((diff & first_mask) == 0)) { | 175 | 33.4k | len = zng_first_diff_byte64(diff); | 176 | 33.4k | goto short_match_accept; | 177 | 33.4k | } | 178 | 5.09M | if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit) | 179 | 1.16M | return best_len; | 180 | 3.92M | cand_start = zng_memread_8(mbase_start + cur_match); | 181 | 3.92M | if (scan_start == cand_start) | 182 | 10.2k | break; | 183 | 3.92M | } | 184 | 1.21M | } | 185 | 1.21M | } | 186 | | /* All 8 bytes match, fallthrough to compare256 for the tail. */ | 187 | 5.08M | } else { | 188 | | /* Pre-filter the candidate on the start and end sentinels before compare256 using | 189 | | * simple 8-byte comparison since best_len >= 8. */ | 190 | 7.86M | for (;;) { | 191 | | /* First check the end of the candidate at best_len+1 due to the higher | 192 | | * likelihood of a mismatch. */ | 193 | 7.86M | if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 && | 194 | 445k | zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0) | 195 | 312k | break; | 196 | 7.54M | GOTO_NEXT_CHAIN; | 197 | 7.54M | } | 198 | 430k | } | 199 | 441k | len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2; | 200 | 441k | Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan"); | 201 | | | 202 | 441k | if (len > best_len) | 203 | 924k | short_match_accept: | 204 | 924k | { | 205 | 924k | uint32_t match_start = cur_match - match_offset; | 206 | 924k | s->match_start = match_start; | 207 | | | 208 | | /* Do not look for better matches if the current match reaches | 209 | | * or exceeds the end of the input. | 210 | | */ | 211 | 924k | if (UNLIKELY(len >= lookahead)) | 212 | 205 | return lookahead; | 213 | 606k | if (UNLIKELY(len >= nice_match)) | 214 | 28.7k | return len; | 215 | | | 216 | 577k | best_len = len; | 217 | | | 218 | 577k | offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0; | 219 | | | 220 | 577k | scan_end = zng_memread_8(scan+offset); | 221 | | | 222 | 577k | #ifdef LONGEST_MATCH_SLOW | 223 | | /* Look for a better string offset */ | 224 | 577k | if (UNLIKELY(offset_search && len > STD_MIN_MATCH && match_start + len < strstart)) { | 225 | 40.7k | const unsigned char *scan_endstr; | 226 | 40.7k | uint32_t hash; | 227 | 40.7k | uint32_t pos, next_pos; | 228 | | | 229 | | /* Go back to offset 0 */ | 230 | 40.7k | cur_match -= match_offset; | 231 | 40.7k | match_offset = 0; | 232 | 40.7k | next_pos = cur_match; | 233 | | | 234 | | /* Walk prev[] for positions inside the match. The bound keeps | 235 | | * the hash window within the match, so we follow chains that | 236 | | * share bytes with the current match, not data past its end. | 237 | | */ | 238 | | # ifdef LONGEST_MATCH_SLOW_ROLL | 239 | | for (uint32_t i = 0; i <= len - STD_MIN_MATCH; i++) { | 240 | | # else /* 4-byte Knuth hash needs len - 4 bound */ | 241 | 770k | for (uint32_t i = 0; i <= len - WANT_MIN_MATCH; i++) { | 242 | 740k | # endif | 243 | 740k | pos = prev[(cur_match + i) & wmask]; | 244 | 740k | if (UNLIKELY(pos < next_pos)) { | 245 | | /* Hash chain is more distant, use it */ | 246 | 60.0k | if (UNLIKELY(pos <= limit_base + i)) | 247 | 10.5k | return best_len; | 248 | 49.4k | next_pos = pos; | 249 | 49.4k | match_offset = i; | 250 | 49.4k | } | 251 | 740k | } | 252 | | /* Switch cur_match to next_pos chain */ | 253 | 30.1k | cur_match = next_pos; | 254 | | | 255 | | /* Try hash head at a window covering the tail of the current | 256 | | * match to find chains that extend farther back. The window | 257 | | * width differs per variant: 3 bytes for rolling, 4 for Knuth. | 258 | | */ | 259 | | # ifdef LONGEST_MATCH_SLOW_ROLL | 260 | | scan_endstr = scan + len - (STD_MIN_MATCH-1); | 261 | | | 262 | | hash = update_hash_roll(0, scan_endstr[0]); | 263 | | hash = update_hash_roll(hash, scan_endstr[1]); | 264 | | hash = update_hash_roll(hash, scan_endstr[2]); | 265 | | | 266 | | pos = head[hash]; | 267 | | if (UNLIKELY(pos < cur_match)) { | 268 | | match_offset = len - (STD_MIN_MATCH-1); | 269 | | if (pos <= limit_base + match_offset) | 270 | | return best_len; | 271 | | cur_match = pos; | 272 | | } | 273 | | # else | 274 | 30.1k | scan_endstr = scan + len - WANT_MIN_MATCH; | 275 | 30.1k | uint32_t val = Z_U32_FROM_LE(zng_memread_4(scan_endstr)); | 276 | 30.1k | UPDATE_HASH_KNUTH(hash, val); | 277 | | | 278 | 30.1k | pos = head[hash]; | 279 | 30.1k | if (pos < cur_match) { | 280 | 0 | match_offset = len - WANT_MIN_MATCH; | 281 | 0 | if (pos <= limit_base + match_offset) | 282 | 0 | return best_len; | 283 | 0 | cur_match = pos; | 284 | 0 | } | 285 | 30.1k | # endif | 286 | | | 287 | | /* Update offset-dependent variables */ | 288 | 30.1k | limit = limit_base+match_offset; | 289 | 30.1k | mbase_start = window-match_offset; | 290 | 30.1k | mbase_end = (mbase_start+offset); | 291 | 30.1k | continue; | 292 | 30.1k | } | 293 | 536k | #endif | 294 | 536k | mbase_end = (mbase_start+offset); | 295 | 536k | } | 296 | | #ifndef LONGEST_MATCH_SLOW | 297 | | else if (UNLIKELY(early_exit)) { | 298 | | /* The probability of finding a match later if we here is pretty low, so for | 299 | | * performance it's best to outright stop here for the lower compression levels | 300 | | */ | 301 | | break; | 302 | | } | 303 | | #endif | 304 | 660k | GOTO_NEXT_CHAIN; | 305 | 660k | } | 306 | 0 | return best_len; | 307 | 4.93M | } |
longest_match_slow_roll_avx2 Line | Count | Source | 28 | 34.7M | Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, uint32_t cur_match) { | 29 | 34.7M | const unsigned wmask = W_MASK(s); | 30 | 34.7M | unsigned int strstart = s->strstart; | 31 | 34.7M | const unsigned char *window = s->window; | 32 | 34.7M | const Pos *prev = s->prev; | 33 | 34.7M | #ifdef LONGEST_MATCH_SLOW | 34 | 34.7M | const Pos *head = s->head; | 35 | 34.7M | #endif | 36 | 34.7M | const unsigned char *scan; | 37 | 34.7M | const unsigned char *mbase_start = window; | 38 | 34.7M | const unsigned char *mbase_end; | 39 | 34.7M | uint32_t limit; | 40 | 34.7M | #ifdef LONGEST_MATCH_SLOW | 41 | 34.7M | uint32_t limit_base; | 42 | 34.7M | #endif | 43 | | #ifndef LONGEST_MATCH_SLOW | 44 | | int32_t early_exit; | 45 | | #endif | 46 | 34.7M | uint32_t chain_length = s->max_chain_length; | 47 | 34.7M | uint32_t nice_match = (uint32_t)s->nice_match; | 48 | 34.7M | uint32_t best_len, offset; | 49 | 34.7M | uint32_t lookahead = s->lookahead; | 50 | 34.7M | uint32_t match_offset = 0; | 51 | 34.7M | uint64_t scan_start; | 52 | 34.7M | uint64_t scan_end; | 53 | | | 54 | | /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */ | 55 | 34.7M | Assert(STD_MAX_MATCH == 258, "Code too clever"); | 56 | | | 57 | 34.7M | best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1; | 58 | 34.7M | if (UNLIKELY(best_len >= lookahead)) | 59 | 434 | return lookahead; | 60 | 34.7M | #ifdef LONGEST_MATCH_SLOW | 61 | 34.7M | # ifdef LONGEST_MATCH_SLOW_ROLL | 62 | | /* Rolling-hash variant always runs the post-match offset search; the | 63 | | * compiler folds the constant away below. | 64 | | */ | 65 | 34.7M | const int offset_search = 1; | 66 | | # else | 67 | | /* The post-match offset search only pays off when we entered with a | 68 | | * prior best_len from lazy evaluation, so gate on it at function entry. | 69 | | */ | 70 | | const int offset_search = (best_len >= STD_MIN_MATCH); | 71 | | # endif | 72 | 34.7M | #endif | 73 | | | 74 | | /* Calculate read offset which should only extend an extra byte to find the | 75 | | * next best match length. When best_len is shorter than the read width, we | 76 | | * diff the mismatched bytes instead. | 77 | | */ | 78 | 34.7M | offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0; | 79 | | | 80 | 34.7M | scan = window + strstart; | 81 | 34.7M | scan_start = zng_memread_8(scan); | 82 | 34.7M | scan_end = zng_memread_8(scan+offset); | 83 | 34.7M | mbase_end = (mbase_start+offset); | 84 | | | 85 | | /* Do not waste too much time if we already have a good match */ | 86 | 34.7M | if (UNLIKELY(best_len >= s->good_match)) | 87 | 70.8k | chain_length >>= 2; | 88 | | | 89 | | /* Stop when cur_match becomes <= limit. To simplify the code, | 90 | | * we prevent matches with the string of window index 0 | 91 | | */ | 92 | 34.7M | limit = strstart > MAX_DIST(s) ? (strstart - MAX_DIST(s)) : 0; | 93 | | #ifndef LONGEST_MATCH_SLOW | 94 | | early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL; | 95 | | #endif | 96 | 34.7M | #ifdef LONGEST_MATCH_SLOW | 97 | 34.7M | limit_base = limit; | 98 | 34.7M | if (best_len >= STD_MIN_MATCH) { | 99 | | /* We're continuing search (lazy evaluation). Find a most distant | 100 | | * chain by hashing substrings within the match area. We cannot use | 101 | | * s->prev[strstart+1,...] immediately because those strings are not | 102 | | * yet inserted into the hash table. | 103 | | */ | 104 | 6.44M | # ifdef LONGEST_MATCH_SLOW_ROLL | 105 | 6.44M | uint32_t hash; | 106 | 6.44M | uint32_t pos; | 107 | | | 108 | 6.44M | hash = update_hash_roll(0, scan[1]); | 109 | 6.44M | hash = update_hash_roll(hash, scan[2]); | 110 | | | 111 | 25.8M | for (uint32_t i = 3; i <= best_len; i++) { | 112 | 19.3M | hash = update_hash_roll(hash, scan[i]); | 113 | 19.3M | pos = head[hash]; | 114 | 19.3M | if (UNLIKELY(pos < cur_match)) { | 115 | 5.24M | match_offset = i - 2; | 116 | 5.24M | cur_match = pos; | 117 | 5.24M | } | 118 | 19.3M | } | 119 | | # else /* 4-byte Knuth hash, fresh lookup per offset */ | 120 | | for (uint32_t i = 1; i + (WANT_MIN_MATCH - 1) <= best_len; i++) { | 121 | | uint32_t val = Z_U32_FROM_LE(zng_memread_4(scan + i)); | 122 | | uint32_t hash; | 123 | | UPDATE_HASH_KNUTH(hash, val); | 124 | | uint32_t pos = head[hash]; | 125 | | if (pos < cur_match) { | 126 | | match_offset = i; | 127 | | cur_match = pos; | 128 | | } | 129 | | } | 130 | | # endif | 131 | | | 132 | | /* Update offset-dependent variables */ | 133 | 6.44M | limit = limit_base+match_offset; | 134 | 6.44M | if (UNLIKELY(cur_match <= limit)) | 135 | 1.47M | return best_len; | 136 | 4.96M | mbase_start -= match_offset; | 137 | 4.96M | mbase_end -= match_offset; | 138 | 4.96M | } | 139 | 33.2M | #endif | 140 | 33.2M | Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead"); | 141 | 44.0M | for (;;) { | 142 | 44.0M | if (UNLIKELY(cur_match >= strstart)) | 143 | 0 | break; | 144 | | | 145 | | /* Skip to next match if the match length cannot increase or if the match length is | 146 | | * less than 2. Note that the checks below for insufficient lookahead only occur | 147 | | * occasionally for performance reasons. | 148 | | * Therefore uninitialized memory will be accessed and conditional jumps will be made | 149 | | * that depend on those values. However the length of the match is limited to the | 150 | | * lookahead, so the output of deflate is not affected by the uninitialized values. | 151 | | */ | 152 | 44.0M | uint32_t len; | 153 | 44.0M | if (best_len < sizeof(uint64_t)) { | 154 | 40.7M | uint64_t cand_start = zng_memread_8(mbase_start + cur_match); | 155 | 40.7M | if (scan_start != cand_start) { | 156 | | /* Peel the first candidate out of the loop. A full 8-byte match falls straight | 157 | | * through to compare256, and single-candidate chains (barely-compressible data) | 158 | | * run with no loop overhead. */ | 159 | 40.4M | uint64_t first_mask = zng_first_bytes_mask64(best_len + 1); | 160 | 40.4M | uint64_t diff = scan_start ^ cand_start; | 161 | | /* A candidate beats best_len only when its first best_len+1 bytes match, i.e. | 162 | | * those bytes of the XOR are zero. The masked test rejects without running ctz. */ | 163 | 40.4M | if (UNLIKELY((diff & first_mask) == 0)) { | 164 | 6.87M | len = zng_first_diff_byte64(diff); | 165 | 6.87M | goto short_match_accept; | 166 | 6.87M | } | 167 | 33.6M | if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit) | 168 | 13.1M | return best_len; | 169 | 20.4M | cand_start = zng_memread_8(mbase_start + cur_match); | 170 | 20.4M | if (scan_start != cand_start) { | 171 | | /* Walk the remaining candidates with the chain advance kept inline. */ | 172 | 137M | for (;;) { | 173 | 137M | diff = scan_start ^ cand_start; | 174 | 137M | if (UNLIKELY((diff & first_mask) == 0)) { | 175 | 3.27M | len = zng_first_diff_byte64(diff); | 176 | 3.27M | goto short_match_accept; | 177 | 3.27M | } | 178 | 134M | if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit) | 179 | 17.1M | return best_len; | 180 | 117M | cand_start = zng_memread_8(mbase_start + cur_match); | 181 | 117M | if (scan_start == cand_start) | 182 | 54.6k | break; | 183 | 117M | } | 184 | 20.4M | } | 185 | 20.4M | } | 186 | | /* All 8 bytes match, fallthrough to compare256 for the tail. */ | 187 | 40.7M | } else { | 188 | | /* Pre-filter the candidate on the start and end sentinels before compare256 using | 189 | | * simple 8-byte comparison since best_len >= 8. */ | 190 | 44.9M | for (;;) { | 191 | | /* First check the end of the candidate at best_len+1 due to the higher | 192 | | * likelihood of a mismatch. */ | 193 | 44.9M | if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 && | 194 | 4.36M | zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0) | 195 | 2.96M | break; | 196 | 41.9M | GOTO_NEXT_CHAIN; | 197 | 41.9M | } | 198 | 3.25M | } | 199 | 3.29M | len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2; | 200 | 3.29M | Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan"); | 201 | | | 202 | 3.29M | if (len > best_len) | 203 | 12.7M | short_match_accept: | 204 | 12.7M | { | 205 | 12.7M | uint32_t match_start = cur_match - match_offset; | 206 | 12.7M | s->match_start = match_start; | 207 | | | 208 | | /* Do not look for better matches if the current match reaches | 209 | | * or exceeds the end of the input. | 210 | | */ | 211 | 12.7M | if (UNLIKELY(len >= lookahead)) | 212 | 440 | return lookahead; | 213 | 11.4M | if (UNLIKELY(len >= nice_match)) | 214 | 60.8k | return len; | 215 | | | 216 | 11.3M | best_len = len; | 217 | | | 218 | 11.3M | offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0; | 219 | | | 220 | 11.3M | scan_end = zng_memread_8(scan+offset); | 221 | | | 222 | 11.3M | #ifdef LONGEST_MATCH_SLOW | 223 | | /* Look for a better string offset */ | 224 | 11.3M | if (UNLIKELY(offset_search && len > STD_MIN_MATCH && match_start + len < strstart)) { | 225 | 5.10M | const unsigned char *scan_endstr; | 226 | 5.10M | uint32_t hash; | 227 | 5.10M | uint32_t pos, next_pos; | 228 | | | 229 | | /* Go back to offset 0 */ | 230 | 5.10M | cur_match -= match_offset; | 231 | 5.10M | match_offset = 0; | 232 | 5.10M | next_pos = cur_match; | 233 | | | 234 | | /* Walk prev[] for positions inside the match. The bound keeps | 235 | | * the hash window within the match, so we follow chains that | 236 | | * share bytes with the current match, not data past its end. | 237 | | */ | 238 | 5.10M | # ifdef LONGEST_MATCH_SLOW_ROLL | 239 | 80.2M | for (uint32_t i = 0; i <= len - STD_MIN_MATCH; i++) { | 240 | | # else /* 4-byte Knuth hash needs len - 4 bound */ | 241 | | for (uint32_t i = 0; i <= len - WANT_MIN_MATCH; i++) { | 242 | | # endif | 243 | 76.5M | pos = prev[(cur_match + i) & wmask]; | 244 | 76.5M | if (UNLIKELY(pos < next_pos)) { | 245 | | /* Hash chain is more distant, use it */ | 246 | 6.86M | if (UNLIKELY(pos <= limit_base + i)) | 247 | 1.37M | return best_len; | 248 | 5.49M | next_pos = pos; | 249 | 5.49M | match_offset = i; | 250 | 5.49M | } | 251 | 76.5M | } | 252 | | /* Switch cur_match to next_pos chain */ | 253 | 3.73M | cur_match = next_pos; | 254 | | | 255 | | /* Try hash head at a window covering the tail of the current | 256 | | * match to find chains that extend farther back. The window | 257 | | * width differs per variant: 3 bytes for rolling, 4 for Knuth. | 258 | | */ | 259 | 3.73M | # ifdef LONGEST_MATCH_SLOW_ROLL | 260 | 3.73M | scan_endstr = scan + len - (STD_MIN_MATCH-1); | 261 | | | 262 | 3.73M | hash = update_hash_roll(0, scan_endstr[0]); | 263 | 3.73M | hash = update_hash_roll(hash, scan_endstr[1]); | 264 | 3.73M | hash = update_hash_roll(hash, scan_endstr[2]); | 265 | | | 266 | 3.73M | pos = head[hash]; | 267 | 3.73M | if (UNLIKELY(pos < cur_match)) { | 268 | 800k | match_offset = len - (STD_MIN_MATCH-1); | 269 | 800k | if (pos <= limit_base + match_offset) | 270 | 370k | return best_len; | 271 | 430k | cur_match = pos; | 272 | 430k | } | 273 | | # else | 274 | | scan_endstr = scan + len - WANT_MIN_MATCH; | 275 | | uint32_t val = Z_U32_FROM_LE(zng_memread_4(scan_endstr)); | 276 | | UPDATE_HASH_KNUTH(hash, val); | 277 | | | 278 | | pos = head[hash]; | 279 | | if (pos < cur_match) { | 280 | | match_offset = len - WANT_MIN_MATCH; | 281 | | if (pos <= limit_base + match_offset) | 282 | | return best_len; | 283 | | cur_match = pos; | 284 | | } | 285 | | # endif | 286 | | | 287 | | /* Update offset-dependent variables */ | 288 | 3.36M | limit = limit_base+match_offset; | 289 | 3.36M | mbase_start = window-match_offset; | 290 | 3.36M | mbase_end = (mbase_start+offset); | 291 | 3.36M | continue; | 292 | 3.73M | } | 293 | 6.26M | #endif | 294 | 6.26M | mbase_end = (mbase_start+offset); | 295 | 6.26M | } | 296 | | #ifndef LONGEST_MATCH_SLOW | 297 | | else if (UNLIKELY(early_exit)) { | 298 | | /* The probability of finding a match later if we here is pretty low, so for | 299 | | * performance it's best to outright stop here for the lower compression levels | 300 | | */ | 301 | | break; | 302 | | } | 303 | | #endif | 304 | 8.28M | GOTO_NEXT_CHAIN; | 305 | 8.28M | } | 306 | 0 | return best_len; | 307 | 33.2M | } |
Unexecuted instantiation: longest_match_avx512 Unexecuted instantiation: longest_match_slow_knuth_avx512 Unexecuted instantiation: longest_match_slow_roll_avx512 |
308 | | |
309 | | #undef LONGEST_MATCH_SLOW |
310 | | #undef LONGEST_MATCH_SLOW_ROLL |
311 | | #undef LONGEST_MATCH |