/src/zlib-ng/deflate_quick.c
Line | Count | Source |
1 | | /* |
2 | | * The deflate_quick deflate strategy, designed to be used when cycles are |
3 | | * at a premium. |
4 | | * |
5 | | * Copyright (C) 2013 Intel Corporation. All rights reserved. |
6 | | * Authors: |
7 | | * Wajdi Feghali <wajdi.k.feghali@intel.com> |
8 | | * Jim Guilford <james.guilford@intel.com> |
9 | | * Vinodh Gopal <vinodh.gopal@intel.com> |
10 | | * Erdinc Ozturk <erdinc.ozturk@intel.com> |
11 | | * Jim Kukunas <james.t.kukunas@linux.intel.com> |
12 | | * |
13 | | * Portions are Copyright (C) 2016 12Sided Technology, LLC. |
14 | | * Author: |
15 | | * Phil Vachon <pvachon@12sidedtech.com> |
16 | | * |
17 | | * For conditions of distribution and use, see copyright notice in zlib.h |
18 | | */ |
19 | | |
20 | | #include "zbuild.h" |
21 | | #include "zmemory.h" |
22 | | #include "deflate.h" |
23 | | #include "deflate_p.h" |
24 | | #include "functable.h" |
25 | | #include "trees_emit.h" |
26 | | #include "insert_string_p.h" |
27 | | |
28 | | extern const ct_data static_ltree[L_CODES+2]; |
29 | | extern const ct_data static_dtree[D_CODES]; |
30 | | |
31 | 707 | #define QUICK_START_BLOCK(s, last) { \ |
32 | 707 | zng_tr_emit_tree(s, STATIC_TREES, last); \ |
33 | 707 | s->block_open = 1 + last; \ |
34 | 707 | s->block_start = (int)s->strstart; \ |
35 | 707 | } |
36 | | |
37 | 1.23k | #define QUICK_END_BLOCK(s, last) { \ |
38 | 1.23k | if (s->block_open) { \ |
39 | 707 | zng_tr_emit_end_block(s, static_ltree, last); \ |
40 | 707 | s->block_open = 0; \ |
41 | 707 | s->block_start = (int)s->strstart; \ |
42 | 707 | PREFIX(flush_pending)(s->strm); \ |
43 | 707 | if (s->strm->avail_out == 0) \ |
44 | 707 | return (last) ? finish_started : need_more; \ |
45 | 707 | } \ |
46 | 1.23k | } |
47 | | |
48 | 1.10k | Z_INTERNAL block_state deflate_quick(deflate_state *s, int flush) { |
49 | 1.10k | unsigned char *window; |
50 | 1.10k | unsigned last = (flush == Z_FINISH) ? 1 : 0; |
51 | | |
52 | 1.10k | if (UNLIKELY(last && s->block_open != 2)) { |
53 | | /* Emit end of previous block */ |
54 | 618 | QUICK_END_BLOCK(s, 0); |
55 | | /* Emit start of last block */ |
56 | 616 | QUICK_START_BLOCK(s, last); |
57 | 616 | } else if (UNLIKELY(s->block_open == 0 && s->lookahead > 0)) { |
58 | | /* Start new block only when we have lookahead data, so that if no |
59 | | input data is given an empty block will not be written */ |
60 | 0 | QUICK_START_BLOCK(s, last); |
61 | 0 | } |
62 | | |
63 | 1.09k | window = s->window; |
64 | | |
65 | 28.4M | for (;;) { |
66 | 28.4M | uint8_t lc; |
67 | | |
68 | 28.4M | if (UNLIKELY(s->pending + ((BIT_BUF_SIZE + 7) >> 3) >= s->pending_buf_size)) { |
69 | 368 | PREFIX(flush_pending)(s->strm); |
70 | 368 | if (s->strm->avail_out == 0) { |
71 | 77 | return (last && s->strm->avail_in == 0 && s->bi_valid == 0 && s->block_open == 0) ? finish_started : need_more; |
72 | 77 | } |
73 | 368 | } |
74 | | |
75 | 28.4M | if (UNLIKELY(s->lookahead < MIN_LOOKAHEAD)) { |
76 | 40.5k | PREFIX(fill_window)(s); |
77 | 40.5k | if (UNLIKELY(s->lookahead < MIN_LOOKAHEAD && flush == Z_NO_FLUSH)) { |
78 | 405 | return need_more; |
79 | 405 | } |
80 | 40.1k | if (UNLIKELY(s->lookahead == 0)) |
81 | 616 | break; |
82 | | |
83 | 39.5k | if (UNLIKELY(s->block_open == 0)) { |
84 | | /* Start new block when we have lookahead data, so that if no |
85 | | input data is given an empty block will not be written */ |
86 | 91 | QUICK_START_BLOCK(s, last); |
87 | 91 | } |
88 | 39.5k | } |
89 | | |
90 | 28.4M | if (LIKELY(s->lookahead >= WANT_MIN_MATCH)) { |
91 | 28.4M | #if BYTE_ORDER == LITTLE_ENDIAN |
92 | 28.4M | uint32_t str_val = zng_memread_4(window + s->strstart); |
93 | | #else |
94 | | uint32_t str_val = ZSWAP32(zng_memread_4(window + s->strstart)); |
95 | | #endif |
96 | 28.4M | uint32_t hash_head = quick_insert_value(s, s->strstart, str_val); |
97 | 28.4M | int64_t dist = (int64_t)s->strstart - hash_head; |
98 | 28.4M | lc = (uint8_t)str_val; |
99 | | |
100 | 28.4M | if (dist <= MAX_DIST(s) && dist > 0) { |
101 | 13.4M | const uint8_t *match_start = window + hash_head; |
102 | 13.4M | #if BYTE_ORDER == LITTLE_ENDIAN |
103 | 13.4M | uint32_t match_val = zng_memread_4(match_start); |
104 | | #else |
105 | | uint32_t match_val = ZSWAP32(zng_memread_4(match_start)); |
106 | | #endif |
107 | | |
108 | 13.4M | if (str_val == match_val) { |
109 | 1.26M | const uint8_t *str_start = window + s->strstart; |
110 | 1.26M | uint32_t match_len = FUNCTABLE_CALL(compare256)(str_start+2, match_start+2) + 2; |
111 | | |
112 | 1.26M | if (match_len >= WANT_MIN_MATCH) { |
113 | 1.26M | if (UNLIKELY(match_len > s->lookahead)) |
114 | 84 | match_len = s->lookahead; |
115 | | |
116 | 1.26M | Assert(match_len <= STD_MAX_MATCH, "match too long"); |
117 | 1.26M | Assert(s->strstart <= UINT16_MAX, "strstart should fit in uint16_t"); |
118 | 1.26M | check_match(s, s->strstart, hash_head, match_len); |
119 | | |
120 | 1.26M | zng_tr_emit_dist(s, static_ltree, static_dtree, match_len - STD_MIN_MATCH, (uint32_t)dist); |
121 | 1.26M | s->lookahead -= match_len; |
122 | 1.26M | s->strstart += match_len; |
123 | 1.26M | continue; |
124 | 1.26M | } |
125 | 1.26M | } |
126 | 13.4M | } |
127 | 28.4M | } else { |
128 | 1.08k | lc = window[s->strstart]; |
129 | 1.08k | } |
130 | 27.1M | zng_tr_emit_lit(s, static_ltree, lc); |
131 | 27.1M | s->strstart++; |
132 | 27.1M | s->lookahead--; |
133 | 27.1M | } |
134 | | |
135 | 616 | s->insert = s->strstart < (STD_MIN_MATCH - 1) ? s->strstart : (STD_MIN_MATCH - 1); |
136 | 616 | if (UNLIKELY(last)) { |
137 | 616 | QUICK_END_BLOCK(s, 1); |
138 | 538 | return finish_done; |
139 | 616 | } |
140 | | |
141 | 0 | QUICK_END_BLOCK(s, 0); |
142 | 0 | return block_done; |
143 | 0 | } |