Coverage Report

Created: 2025-09-05 06:28

/src/zlib-ng/match_tpl.h
Line
Count
Source (jump to first uncovered line)
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
#ifndef MATCH_TPL_H
12
#define MATCH_TPL_H
13
14
209M
#define EARLY_EXIT_TRIGGER_LEVEL 5
15
16
#endif
17
18
/* Set match_start to the longest match starting at the given string and
19
 * return its length. Matches shorter or equal to prev_length are discarded,
20
 * in which case the result is equal to prev_length and match_start is garbage.
21
 *
22
 * IN assertions: cur_match is the head of the hash chain for the current
23
 * string (strstart) and its distance is <= MAX_DIST, and prev_length >=1
24
 * OUT assertion: the match length is not greater than s->lookahead
25
 *
26
 * The LONGEST_MATCH_SLOW variant spends more time to attempt to find longer
27
 * matches once a match has already been found.
28
 */
29
226M
Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, Pos cur_match) {
30
226M
    unsigned int strstart = s->strstart;
31
226M
    const unsigned wmask = s->w_mask;
32
226M
    unsigned char *window = s->window;
33
226M
    unsigned char *scan = window + strstart;
34
226M
    Z_REGISTER unsigned char *mbase_start = window;
35
226M
    Z_REGISTER unsigned char *mbase_end;
36
226M
    const Pos *prev = s->prev;
37
226M
    Pos limit;
38
#ifdef LONGEST_MATCH_SLOW
39
    Pos limit_base;
40
#else
41
    int32_t early_exit;
42
#endif
43
226M
    uint32_t chain_length, nice_match, best_len, offset;
44
226M
    uint32_t lookahead = s->lookahead;
45
226M
    Pos match_offset = 0;
46
226M
    uint64_t scan_start;
47
226M
    uint64_t scan_end;
48
49
226M
#define GOTO_NEXT_CHAIN \
50
914M
    if (--chain_length && (cur_match = prev[cur_match & wmask]) > limit) \
51
692M
        continue; \
52
221M
    return best_len;
53
54
    /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */
55
226M
    Assert(STD_MAX_MATCH == 258, "Code too clever");
56
57
226M
    best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1;
58
59
    /* Calculate read offset which should only extend an extra byte
60
     * to find the next best match length.
61
     */
62
226M
    offset = best_len-1;
63
226M
    if (best_len >= sizeof(uint32_t)) {
64
5.83M
        offset -= 2;
65
5.83M
        if (best_len >= sizeof(uint64_t))
66
1.91M
            offset -= 4;
67
5.83M
    }
68
69
226M
    scan_start = zng_memread_8(scan);
70
226M
    scan_end = zng_memread_8(scan+offset);
71
226M
    mbase_end  = (mbase_start+offset);
72
73
    /* Do not waste too much time if we already have a good match */
74
226M
    chain_length = s->max_chain_length;
75
226M
    if (best_len >= s->good_match)
76
1.57M
        chain_length >>= 2;
77
226M
    nice_match = (uint32_t)s->nice_match;
78
79
    /* Stop when cur_match becomes <= limit. To simplify the code,
80
     * we prevent matches with the string of window index 0
81
     */
82
226M
    limit = strstart > MAX_DIST(s) ? (Pos)(strstart - MAX_DIST(s)) : 0;
83
#ifdef LONGEST_MATCH_SLOW
84
    limit_base = limit;
85
17.8M
    if (best_len >= STD_MIN_MATCH) {
86
        /* We're continuing search (lazy evaluation). */
87
3.35M
        uint32_t i, hash;
88
3.35M
        Pos pos;
89
90
        /* Find a most distant chain starting from scan with index=1 (index=0 corresponds
91
         * to cur_match). We cannot use s->prev[strstart+1,...] immediately, because
92
         * these strings are not yet inserted into the hash table.
93
         */
94
        hash = s->update_hash(0, scan[1]);
95
        hash = s->update_hash(hash, scan[2]);
96
97
18.0M
        for (i = 3; i <= best_len; i++) {
98
14.7M
            hash = s->update_hash(hash, scan[i]);
99
100
            /* If we're starting with best_len >= 3, we can use offset search. */
101
14.7M
            pos = s->head[hash];
102
14.7M
            if (pos < cur_match) {
103
3.02M
                match_offset = (Pos)(i - 2);
104
3.02M
                cur_match = pos;
105
3.02M
            }
106
14.7M
        }
107
108
        /* Update offset-dependent variables */
109
3.35M
        limit = limit_base+match_offset;
110
3.35M
        if (cur_match <= limit)
111
807k
            goto break_matching;
112
2.55M
        mbase_start -= match_offset;
113
2.55M
        mbase_end -= match_offset;
114
2.55M
    }
115
#else
116
209M
    early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL;
117
#endif
118
17.0M
    Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead");
119
264M
    for (;;) {
120
264M
        if (cur_match >= strstart)
121
26
            break;
122
123
        /* Skip to next match if the match length cannot increase or if the match length is
124
         * less than 2. Note that the checks below for insufficient lookahead only occur
125
         * occasionally for performance reasons.
126
         * Therefore uninitialized memory will be accessed and conditional jumps will be made
127
         * that depend on those values. However the length of the match is limited to the
128
         * lookahead, so the output of deflate is not affected by the uninitialized values.
129
         */
130
264M
        if (best_len < sizeof(uint32_t)) {
131
325M
            for (;;) {
132
325M
                if (zng_memcmp_2(mbase_end+cur_match, &scan_end) == 0 &&
133
325M
                    zng_memcmp_2(mbase_start+cur_match, &scan_start) == 0)
134
32.3M
                    break;
135
293M
                GOTO_NEXT_CHAIN;
136
293M
            }
137
223M
        } else if (best_len >= sizeof(uint64_t)) {
138
265M
            for (;;) {
139
265M
                if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 &&
140
265M
                    zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0)
141
10.8M
                    break;
142
254M
                GOTO_NEXT_CHAIN;
143
254M
            }
144
22.5M
        } else {
145
331M
            for (;;) {
146
331M
                if (zng_memcmp_4(mbase_end+cur_match, &scan_end) == 0 &&
147
331M
                    zng_memcmp_4(mbase_start+cur_match, &scan_start) == 0)
148
7.73M
                    break;
149
323M
                GOTO_NEXT_CHAIN;
150
323M
            }
151
22.5M
        }
152
50.9M
        uint32_t len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2;
153
50.9M
        Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan");
154
155
50.9M
        if (len > best_len) {
156
48.5M
            uint32_t match_start = cur_match - match_offset;
157
48.5M
            s->match_start = match_start;
158
159
            /* Do not look for matches beyond the end of the input. */
160
48.5M
            if (len > lookahead)
161
9.62k
                return lookahead;
162
48.5M
            best_len = len;
163
48.5M
            if (best_len >= nice_match)
164
3.75M
                return best_len;
165
166
44.7M
            offset = best_len-1;
167
44.7M
            if (best_len >= sizeof(uint32_t)) {
168
41.3M
                offset -= 2;
169
41.3M
                if (best_len >= sizeof(uint64_t))
170
15.4M
                    offset -= 4;
171
41.3M
            }
172
173
44.7M
            scan_end = zng_memread_8(scan+offset);
174
175
#ifdef LONGEST_MATCH_SLOW
176
            /* Look for a better string offset */
177
7.54M
            if (UNLIKELY(len > STD_MIN_MATCH && match_start + len < strstart)) {
178
3.95M
                Pos pos, next_pos;
179
3.95M
                uint32_t i, hash;
180
3.95M
                unsigned char *scan_endstr;
181
182
                /* Go back to offset 0 */
183
                cur_match -= match_offset;
184
                match_offset = 0;
185
                next_pos = cur_match;
186
108M
                for (i = 0; i <= len - STD_MIN_MATCH; i++) {
187
105M
                    pos = prev[(cur_match + i) & wmask];
188
105M
                    if (pos < next_pos) {
189
                        /* Hash chain is more distant, use it */
190
5.14M
                        if (pos <= limit_base + i)
191
880k
                            goto break_matching;
192
4.26M
                        next_pos = pos;
193
4.26M
                        match_offset = (Pos)i;
194
4.26M
                    }
195
105M
                }
196
                /* Switch cur_match to next_pos chain */
197
3.07M
                cur_match = next_pos;
198
199
                /* Try hash head at len-(STD_MIN_MATCH-1) position to see if we could get
200
                 * a better cur_match at the end of string. Using (STD_MIN_MATCH-1) lets
201
                 * us include one more byte into hash - the byte which will be checked
202
                 * in main loop now, and which allows to grow match by 1.
203
                 */
204
3.07M
                scan_endstr = scan + len - (STD_MIN_MATCH+1);
205
206
3.07M
                hash = s->update_hash(0, scan_endstr[0]);
207
3.07M
                hash = s->update_hash(hash, scan_endstr[1]);
208
3.07M
                hash = s->update_hash(hash, scan_endstr[2]);
209
210
3.07M
                pos = s->head[hash];
211
3.07M
                if (pos < cur_match) {
212
0
                    match_offset = (Pos)(len - (STD_MIN_MATCH+1));
213
0
                    if (pos <= limit_base + match_offset)
214
0
                        goto break_matching;
215
0
                    cur_match = pos;
216
0
                }
217
218
                /* Update offset-dependent variables */
219
3.07M
                limit = limit_base+match_offset;
220
3.07M
                mbase_start = window-match_offset;
221
3.07M
                mbase_end = (mbase_start+offset);
222
3.07M
                continue;
223
3.07M
            }
224
3.59M
#endif
225
3.59M
            mbase_end = (mbase_start+offset);
226
3.59M
        }
227
#ifndef LONGEST_MATCH_SLOW
228
1.08M
        else if (UNLIKELY(early_exit)) {
229
            /* The probability of finding a match later if we here is pretty low, so for
230
             * performance it's best to outright stop here for the lower compression levels
231
             */
232
866
            break;
233
866
        }
234
38.3M
#endif
235
43.2M
        GOTO_NEXT_CHAIN;
236
43.2M
    }
237
892
    return best_len;
238
239
#ifdef LONGEST_MATCH_SLOW
240
1.68M
break_matching:
241
242
1.68M
    if (best_len < s->lookahead)
243
1.68M
        return best_len;
244
245
1.10k
    return s->lookahead;
246
#endif
247
1.68M
}
Unexecuted instantiation: longest_match_sse2
Unexecuted instantiation: longest_match_slow_sse2
longest_match_avx2
Line
Count
Source
29
209M
Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, Pos cur_match) {
30
209M
    unsigned int strstart = s->strstart;
31
209M
    const unsigned wmask = s->w_mask;
32
209M
    unsigned char *window = s->window;
33
209M
    unsigned char *scan = window + strstart;
34
209M
    Z_REGISTER unsigned char *mbase_start = window;
35
209M
    Z_REGISTER unsigned char *mbase_end;
36
209M
    const Pos *prev = s->prev;
37
209M
    Pos limit;
38
#ifdef LONGEST_MATCH_SLOW
39
    Pos limit_base;
40
#else
41
209M
    int32_t early_exit;
42
209M
#endif
43
209M
    uint32_t chain_length, nice_match, best_len, offset;
44
209M
    uint32_t lookahead = s->lookahead;
45
209M
    Pos match_offset = 0;
46
209M
    uint64_t scan_start;
47
209M
    uint64_t scan_end;
48
49
209M
#define GOTO_NEXT_CHAIN \
50
209M
    if (--chain_length && (cur_match = prev[cur_match & wmask]) > limit) \
51
209M
        continue; \
52
209M
    return best_len;
53
54
    /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */
55
209M
    Assert(STD_MAX_MATCH == 258, "Code too clever");
56
57
209M
    best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1;
58
59
    /* Calculate read offset which should only extend an extra byte
60
     * to find the next best match length.
61
     */
62
209M
    offset = best_len-1;
63
209M
    if (best_len >= sizeof(uint32_t)) {
64
3.77M
        offset -= 2;
65
3.77M
        if (best_len >= sizeof(uint64_t))
66
1.51M
            offset -= 4;
67
3.77M
    }
68
69
209M
    scan_start = zng_memread_8(scan);
70
209M
    scan_end = zng_memread_8(scan+offset);
71
209M
    mbase_end  = (mbase_start+offset);
72
73
    /* Do not waste too much time if we already have a good match */
74
209M
    chain_length = s->max_chain_length;
75
209M
    if (best_len >= s->good_match)
76
1.50M
        chain_length >>= 2;
77
209M
    nice_match = (uint32_t)s->nice_match;
78
79
    /* Stop when cur_match becomes <= limit. To simplify the code,
80
     * we prevent matches with the string of window index 0
81
     */
82
209M
    limit = strstart > MAX_DIST(s) ? (Pos)(strstart - MAX_DIST(s)) : 0;
83
#ifdef LONGEST_MATCH_SLOW
84
    limit_base = limit;
85
    if (best_len >= STD_MIN_MATCH) {
86
        /* We're continuing search (lazy evaluation). */
87
        uint32_t i, hash;
88
        Pos pos;
89
90
        /* Find a most distant chain starting from scan with index=1 (index=0 corresponds
91
         * to cur_match). We cannot use s->prev[strstart+1,...] immediately, because
92
         * these strings are not yet inserted into the hash table.
93
         */
94
        hash = s->update_hash(0, scan[1]);
95
        hash = s->update_hash(hash, scan[2]);
96
97
        for (i = 3; i <= best_len; i++) {
98
            hash = s->update_hash(hash, scan[i]);
99
100
            /* If we're starting with best_len >= 3, we can use offset search. */
101
            pos = s->head[hash];
102
            if (pos < cur_match) {
103
                match_offset = (Pos)(i - 2);
104
                cur_match = pos;
105
            }
106
        }
107
108
        /* Update offset-dependent variables */
109
        limit = limit_base+match_offset;
110
        if (cur_match <= limit)
111
            goto break_matching;
112
        mbase_start -= match_offset;
113
        mbase_end -= match_offset;
114
    }
115
#else
116
209M
    early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL;
117
209M
#endif
118
209M
    Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead");
119
239M
    for (;;) {
120
239M
        if (cur_match >= strstart)
121
26
            break;
122
123
        /* Skip to next match if the match length cannot increase or if the match length is
124
         * less than 2. Note that the checks below for insufficient lookahead only occur
125
         * occasionally for performance reasons.
126
         * Therefore uninitialized memory will be accessed and conditional jumps will be made
127
         * that depend on those values. However the length of the match is limited to the
128
         * lookahead, so the output of deflate is not affected by the uninitialized values.
129
         */
130
239M
        if (best_len < sizeof(uint32_t)) {
131
262M
            for (;;) {
132
262M
                if (zng_memcmp_2(mbase_end+cur_match, &scan_end) == 0 &&
133
262M
                    zng_memcmp_2(mbase_start+cur_match, &scan_start) == 0)
134
24.8M
                    break;
135
237M
                GOTO_NEXT_CHAIN;
136
237M
            }
137
205M
        } else if (best_len >= sizeof(uint64_t)) {
138
219M
            for (;;) {
139
219M
                if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 &&
140
219M
                    zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0)
141
8.31M
                    break;
142
211M
                GOTO_NEXT_CHAIN;
143
211M
            }
144
19.6M
        } else {
145
292M
            for (;;) {
146
292M
                if (zng_memcmp_4(mbase_end+cur_match, &scan_end) == 0 &&
147
292M
                    zng_memcmp_4(mbase_start+cur_match, &scan_start) == 0)
148
7.18M
                    break;
149
285M
                GOTO_NEXT_CHAIN;
150
285M
            }
151
19.6M
        }
152
40.3M
        uint32_t len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2;
153
40.3M
        Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan");
154
155
40.3M
        if (len > best_len) {
156
39.2M
            uint32_t match_start = cur_match - match_offset;
157
39.2M
            s->match_start = match_start;
158
159
            /* Do not look for matches beyond the end of the input. */
160
39.2M
            if (len > lookahead)
161
3.07k
                return lookahead;
162
39.2M
            best_len = len;
163
39.2M
            if (best_len >= nice_match)
164
2.04M
                return best_len;
165
166
37.2M
            offset = best_len-1;
167
37.2M
            if (best_len >= sizeof(uint32_t)) {
168
37.2M
                offset -= 2;
169
37.2M
                if (best_len >= sizeof(uint64_t))
170
13.8M
                    offset -= 4;
171
37.2M
            }
172
173
37.2M
            scan_end = zng_memread_8(scan+offset);
174
175
#ifdef LONGEST_MATCH_SLOW
176
            /* Look for a better string offset */
177
            if (UNLIKELY(len > STD_MIN_MATCH && match_start + len < strstart)) {
178
                Pos pos, next_pos;
179
                uint32_t i, hash;
180
                unsigned char *scan_endstr;
181
182
                /* Go back to offset 0 */
183
                cur_match -= match_offset;
184
                match_offset = 0;
185
                next_pos = cur_match;
186
                for (i = 0; i <= len - STD_MIN_MATCH; i++) {
187
                    pos = prev[(cur_match + i) & wmask];
188
                    if (pos < next_pos) {
189
                        /* Hash chain is more distant, use it */
190
                        if (pos <= limit_base + i)
191
                            goto break_matching;
192
                        next_pos = pos;
193
                        match_offset = (Pos)i;
194
                    }
195
                }
196
                /* Switch cur_match to next_pos chain */
197
                cur_match = next_pos;
198
199
                /* Try hash head at len-(STD_MIN_MATCH-1) position to see if we could get
200
                 * a better cur_match at the end of string. Using (STD_MIN_MATCH-1) lets
201
                 * us include one more byte into hash - the byte which will be checked
202
                 * in main loop now, and which allows to grow match by 1.
203
                 */
204
                scan_endstr = scan + len - (STD_MIN_MATCH+1);
205
206
                hash = s->update_hash(0, scan_endstr[0]);
207
                hash = s->update_hash(hash, scan_endstr[1]);
208
                hash = s->update_hash(hash, scan_endstr[2]);
209
210
                pos = s->head[hash];
211
                if (pos < cur_match) {
212
                    match_offset = (Pos)(len - (STD_MIN_MATCH+1));
213
                    if (pos <= limit_base + match_offset)
214
                        goto break_matching;
215
                    cur_match = pos;
216
                }
217
218
                /* Update offset-dependent variables */
219
                limit = limit_base+match_offset;
220
                mbase_start = window-match_offset;
221
                mbase_end = (mbase_start+offset);
222
                continue;
223
            }
224
#endif
225
37.2M
            mbase_end = (mbase_start+offset);
226
37.2M
        }
227
1.08M
#ifndef LONGEST_MATCH_SLOW
228
1.08M
        else if (UNLIKELY(early_exit)) {
229
            /* The probability of finding a match later if we here is pretty low, so for
230
             * performance it's best to outright stop here for the lower compression levels
231
             */
232
866
            break;
233
866
        }
234
38.3M
#endif
235
38.3M
        GOTO_NEXT_CHAIN;
236
38.3M
    }
237
892
    return best_len;
238
239
#ifdef LONGEST_MATCH_SLOW
240
break_matching:
241
242
    if (best_len < s->lookahead)
243
        return best_len;
244
245
    return s->lookahead;
246
#endif
247
209M
}
longest_match_slow_avx2
Line
Count
Source
29
17.8M
Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, Pos cur_match) {
30
17.8M
    unsigned int strstart = s->strstart;
31
17.8M
    const unsigned wmask = s->w_mask;
32
17.8M
    unsigned char *window = s->window;
33
17.8M
    unsigned char *scan = window + strstart;
34
17.8M
    Z_REGISTER unsigned char *mbase_start = window;
35
17.8M
    Z_REGISTER unsigned char *mbase_end;
36
17.8M
    const Pos *prev = s->prev;
37
17.8M
    Pos limit;
38
17.8M
#ifdef LONGEST_MATCH_SLOW
39
17.8M
    Pos limit_base;
40
#else
41
    int32_t early_exit;
42
#endif
43
17.8M
    uint32_t chain_length, nice_match, best_len, offset;
44
17.8M
    uint32_t lookahead = s->lookahead;
45
17.8M
    Pos match_offset = 0;
46
17.8M
    uint64_t scan_start;
47
17.8M
    uint64_t scan_end;
48
49
17.8M
#define GOTO_NEXT_CHAIN \
50
17.8M
    if (--chain_length && (cur_match = prev[cur_match & wmask]) > limit) \
51
17.8M
        continue; \
52
17.8M
    return best_len;
53
54
    /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */
55
17.8M
    Assert(STD_MAX_MATCH == 258, "Code too clever");
56
57
17.8M
    best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1;
58
59
    /* Calculate read offset which should only extend an extra byte
60
     * to find the next best match length.
61
     */
62
17.8M
    offset = best_len-1;
63
17.8M
    if (best_len >= sizeof(uint32_t)) {
64
2.05M
        offset -= 2;
65
2.05M
        if (best_len >= sizeof(uint64_t))
66
395k
            offset -= 4;
67
2.05M
    }
68
69
17.8M
    scan_start = zng_memread_8(scan);
70
17.8M
    scan_end = zng_memread_8(scan+offset);
71
17.8M
    mbase_end  = (mbase_start+offset);
72
73
    /* Do not waste too much time if we already have a good match */
74
17.8M
    chain_length = s->max_chain_length;
75
17.8M
    if (best_len >= s->good_match)
76
72.6k
        chain_length >>= 2;
77
17.8M
    nice_match = (uint32_t)s->nice_match;
78
79
    /* Stop when cur_match becomes <= limit. To simplify the code,
80
     * we prevent matches with the string of window index 0
81
     */
82
17.8M
    limit = strstart > MAX_DIST(s) ? (Pos)(strstart - MAX_DIST(s)) : 0;
83
17.8M
#ifdef LONGEST_MATCH_SLOW
84
17.8M
    limit_base = limit;
85
17.8M
    if (best_len >= STD_MIN_MATCH) {
86
        /* We're continuing search (lazy evaluation). */
87
3.35M
        uint32_t i, hash;
88
3.35M
        Pos pos;
89
90
        /* Find a most distant chain starting from scan with index=1 (index=0 corresponds
91
         * to cur_match). We cannot use s->prev[strstart+1,...] immediately, because
92
         * these strings are not yet inserted into the hash table.
93
         */
94
3.35M
        hash = s->update_hash(0, scan[1]);
95
3.35M
        hash = s->update_hash(hash, scan[2]);
96
97
18.0M
        for (i = 3; i <= best_len; i++) {
98
14.7M
            hash = s->update_hash(hash, scan[i]);
99
100
            /* If we're starting with best_len >= 3, we can use offset search. */
101
14.7M
            pos = s->head[hash];
102
14.7M
            if (pos < cur_match) {
103
3.02M
                match_offset = (Pos)(i - 2);
104
3.02M
                cur_match = pos;
105
3.02M
            }
106
14.7M
        }
107
108
        /* Update offset-dependent variables */
109
3.35M
        limit = limit_base+match_offset;
110
3.35M
        if (cur_match <= limit)
111
807k
            goto break_matching;
112
2.55M
        mbase_start -= match_offset;
113
2.55M
        mbase_end -= match_offset;
114
2.55M
    }
115
#else
116
    early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL;
117
#endif
118
17.0M
    Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead");
119
24.4M
    for (;;) {
120
24.4M
        if (cur_match >= strstart)
121
0
            break;
122
123
        /* Skip to next match if the match length cannot increase or if the match length is
124
         * less than 2. Note that the checks below for insufficient lookahead only occur
125
         * occasionally for performance reasons.
126
         * Therefore uninitialized memory will be accessed and conditional jumps will be made
127
         * that depend on those values. However the length of the match is limited to the
128
         * lookahead, so the output of deflate is not affected by the uninitialized values.
129
         */
130
24.4M
        if (best_len < sizeof(uint32_t)) {
131
63.0M
            for (;;) {
132
63.0M
                if (zng_memcmp_2(mbase_end+cur_match, &scan_end) == 0 &&
133
63.0M
                    zng_memcmp_2(mbase_start+cur_match, &scan_start) == 0)
134
7.49M
                    break;
135
55.5M
                GOTO_NEXT_CHAIN;
136
55.5M
            }
137
18.5M
        } else if (best_len >= sizeof(uint64_t)) {
138
45.2M
            for (;;) {
139
45.2M
                if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 &&
140
45.2M
                    zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0)
141
2.49M
                    break;
142
42.7M
                GOTO_NEXT_CHAIN;
143
42.7M
            }
144
2.97M
        } else {
145
38.3M
            for (;;) {
146
38.3M
                if (zng_memcmp_4(mbase_end+cur_match, &scan_end) == 0 &&
147
38.3M
                    zng_memcmp_4(mbase_start+cur_match, &scan_start) == 0)
148
548k
                    break;
149
37.7M
                GOTO_NEXT_CHAIN;
150
37.7M
            }
151
2.94M
        }
152
10.5M
        uint32_t len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2;
153
10.5M
        Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan");
154
155
10.5M
        if (len > best_len) {
156
9.26M
            uint32_t match_start = cur_match - match_offset;
157
9.26M
            s->match_start = match_start;
158
159
            /* Do not look for matches beyond the end of the input. */
160
9.26M
            if (len > lookahead)
161
6.54k
                return lookahead;
162
9.26M
            best_len = len;
163
9.26M
            if (best_len >= nice_match)
164
1.71M
                return best_len;
165
166
7.54M
            offset = best_len-1;
167
7.54M
            if (best_len >= sizeof(uint32_t)) {
168
4.07M
                offset -= 2;
169
4.07M
                if (best_len >= sizeof(uint64_t))
170
1.57M
                    offset -= 4;
171
4.07M
            }
172
173
7.54M
            scan_end = zng_memread_8(scan+offset);
174
175
7.54M
#ifdef LONGEST_MATCH_SLOW
176
            /* Look for a better string offset */
177
7.54M
            if (UNLIKELY(len > STD_MIN_MATCH && match_start + len < strstart)) {
178
3.95M
                Pos pos, next_pos;
179
3.95M
                uint32_t i, hash;
180
3.95M
                unsigned char *scan_endstr;
181
182
                /* Go back to offset 0 */
183
3.95M
                cur_match -= match_offset;
184
3.95M
                match_offset = 0;
185
3.95M
                next_pos = cur_match;
186
108M
                for (i = 0; i <= len - STD_MIN_MATCH; i++) {
187
105M
                    pos = prev[(cur_match + i) & wmask];
188
105M
                    if (pos < next_pos) {
189
                        /* Hash chain is more distant, use it */
190
5.14M
                        if (pos <= limit_base + i)
191
880k
                            goto break_matching;
192
4.26M
                        next_pos = pos;
193
4.26M
                        match_offset = (Pos)i;
194
4.26M
                    }
195
105M
                }
196
                /* Switch cur_match to next_pos chain */
197
3.07M
                cur_match = next_pos;
198
199
                /* Try hash head at len-(STD_MIN_MATCH-1) position to see if we could get
200
                 * a better cur_match at the end of string. Using (STD_MIN_MATCH-1) lets
201
                 * us include one more byte into hash - the byte which will be checked
202
                 * in main loop now, and which allows to grow match by 1.
203
                 */
204
3.07M
                scan_endstr = scan + len - (STD_MIN_MATCH+1);
205
206
3.07M
                hash = s->update_hash(0, scan_endstr[0]);
207
3.07M
                hash = s->update_hash(hash, scan_endstr[1]);
208
3.07M
                hash = s->update_hash(hash, scan_endstr[2]);
209
210
3.07M
                pos = s->head[hash];
211
3.07M
                if (pos < cur_match) {
212
0
                    match_offset = (Pos)(len - (STD_MIN_MATCH+1));
213
0
                    if (pos <= limit_base + match_offset)
214
0
                        goto break_matching;
215
0
                    cur_match = pos;
216
0
                }
217
218
                /* Update offset-dependent variables */
219
3.07M
                limit = limit_base+match_offset;
220
3.07M
                mbase_start = window-match_offset;
221
3.07M
                mbase_end = (mbase_start+offset);
222
3.07M
                continue;
223
3.07M
            }
224
3.59M
#endif
225
3.59M
            mbase_end = (mbase_start+offset);
226
3.59M
        }
227
#ifndef LONGEST_MATCH_SLOW
228
        else if (UNLIKELY(early_exit)) {
229
            /* The probability of finding a match later if we here is pretty low, so for
230
             * performance it's best to outright stop here for the lower compression levels
231
             */
232
            break;
233
        }
234
#endif
235
4.86M
        GOTO_NEXT_CHAIN;
236
4.86M
    }
237
0
    return best_len;
238
239
0
#ifdef LONGEST_MATCH_SLOW
240
1.68M
break_matching:
241
242
1.68M
    if (best_len < s->lookahead)
243
1.68M
        return best_len;
244
245
1.10k
    return s->lookahead;
246
1.68M
#endif
247
1.68M
}
Unexecuted instantiation: longest_match_avx512
Unexecuted instantiation: longest_match_slow_avx512
248
249
#undef LONGEST_MATCH_SLOW
250
#undef LONGEST_MATCH