|
4 | 4 | */
|
5 | 5 |
|
6 | 6 | #include "zbuild.h"
|
7 |
| -#include "zmemory.h" |
8 |
| -#include "deflate.h" |
9 |
| -#include "fallback_builtins.h" |
10 |
| - |
11 |
| -/* ALIGNED, byte comparison */ |
12 |
| -static inline uint32_t compare256_c_static(const uint8_t *src0, const uint8_t *src1) { |
13 |
| - uint32_t len = 0; |
14 |
| - |
15 |
| - do { |
16 |
| - if (*src0 != *src1) |
17 |
| - return len; |
18 |
| - src0 += 1, src1 += 1, len += 1; |
19 |
| - if (*src0 != *src1) |
20 |
| - return len; |
21 |
| - src0 += 1, src1 += 1, len += 1; |
22 |
| - if (*src0 != *src1) |
23 |
| - return len; |
24 |
| - src0 += 1, src1 += 1, len += 1; |
25 |
| - if (*src0 != *src1) |
26 |
| - return len; |
27 |
| - src0 += 1, src1 += 1, len += 1; |
28 |
| - if (*src0 != *src1) |
29 |
| - return len; |
30 |
| - src0 += 1, src1 += 1, len += 1; |
31 |
| - if (*src0 != *src1) |
32 |
| - return len; |
33 |
| - src0 += 1, src1 += 1, len += 1; |
34 |
| - if (*src0 != *src1) |
35 |
| - return len; |
36 |
| - src0 += 1, src1 += 1, len += 1; |
37 |
| - if (*src0 != *src1) |
38 |
| - return len; |
39 |
| - src0 += 1, src1 += 1, len += 1; |
40 |
| - } while (len < 256); |
41 |
| - |
42 |
| - return 256; |
43 |
| -} |
| 7 | +#include "compare256_p.h" |
| 8 | + |
| 9 | +// Set optimal COMPARE256 function variant |
| 10 | +#if OPTIMAL_CMP == 8 |
| 11 | +# define COMPARE256 compare256_8 |
| 12 | +#elif defined(HAVE_BUILTIN_CTZLL) |
| 13 | +# define COMPARE256 compare256_64 |
| 14 | +#elif defined(HAVE_BUILTIN_CTZ) |
| 15 | +# define COMPARE256 compare256_32 |
| 16 | +#else |
| 17 | +# define COMPARE256 compare256_16 |
| 18 | +#endif |
44 | 19 |
|
45 | 20 | Z_INTERNAL uint32_t compare256_c(const uint8_t *src0, const uint8_t *src1) {
|
46 |
| - return compare256_c_static(src0, src1); |
| 21 | + return COMPARE256(src0, src1); |
47 | 22 | }
|
48 | 23 |
|
| 24 | +// Generate longest_match_c |
49 | 25 | #define LONGEST_MATCH longest_match_c
|
50 |
| -#define COMPARE256 compare256_c_static |
51 |
| - |
52 | 26 | #include "match_tpl.h"
|
53 | 27 |
|
| 28 | +// Generate longest_match_slow_c |
54 | 29 | #define LONGEST_MATCH_SLOW
|
55 | 30 | #define LONGEST_MATCH longest_match_slow_c
|
56 |
| -#define COMPARE256 compare256_c_static |
57 |
| - |
58 |
| -#include "match_tpl.h" |
59 |
| - |
60 |
| -#if OPTIMAL_CMP >= 32 |
61 |
| - |
62 |
| -/* 16-bit unaligned integer comparison */ |
63 |
| -static inline uint32_t compare256_16_static(const uint8_t *src0, const uint8_t *src1) { |
64 |
| - uint32_t len = 0; |
65 |
| - |
66 |
| - do { |
67 |
| - if (zng_memcmp_2(src0, src1) != 0) |
68 |
| - return len + (*src0 == *src1); |
69 |
| - src0 += 2, src1 += 2, len += 2; |
70 |
| - |
71 |
| - if (zng_memcmp_2(src0, src1) != 0) |
72 |
| - return len + (*src0 == *src1); |
73 |
| - src0 += 2, src1 += 2, len += 2; |
74 |
| - |
75 |
| - if (zng_memcmp_2(src0, src1) != 0) |
76 |
| - return len + (*src0 == *src1); |
77 |
| - src0 += 2, src1 += 2, len += 2; |
78 |
| - |
79 |
| - if (zng_memcmp_2(src0, src1) != 0) |
80 |
| - return len + (*src0 == *src1); |
81 |
| - src0 += 2, src1 += 2, len += 2; |
82 |
| - } while (len < 256); |
83 |
| - |
84 |
| - return 256; |
85 |
| -} |
86 |
| - |
87 |
| -Z_INTERNAL uint32_t compare256_16(const uint8_t *src0, const uint8_t *src1) { |
88 |
| - return compare256_16_static(src0, src1); |
89 |
| -} |
90 |
| - |
91 |
| -#define LONGEST_MATCH longest_match_16 |
92 |
| -#define COMPARE256 compare256_16_static |
93 |
| - |
94 | 31 | #include "match_tpl.h"
|
95 |
| - |
96 |
| -#define LONGEST_MATCH_SLOW |
97 |
| -#define LONGEST_MATCH longest_match_slow_16 |
98 |
| -#define COMPARE256 compare256_16_static |
99 |
| - |
100 |
| -#include "match_tpl.h" |
101 |
| - |
102 |
| -#ifdef HAVE_BUILTIN_CTZ |
103 |
| -/* 32-bit unaligned integer comparison */ |
104 |
| -static inline uint32_t compare256_32_static(const uint8_t *src0, const uint8_t *src1) { |
105 |
| - uint32_t len = 0; |
106 |
| - |
107 |
| - do { |
108 |
| - uint32_t sv, mv, diff; |
109 |
| - |
110 |
| - sv = zng_memread_4(src0); |
111 |
| - mv = zng_memread_4(src1); |
112 |
| - |
113 |
| - diff = sv ^ mv; |
114 |
| - if (diff) { |
115 |
| -#if BYTE_ORDER == LITTLE_ENDIAN |
116 |
| - uint32_t match_byte = __builtin_ctz(diff) / 8; |
117 |
| -#else |
118 |
| - uint32_t match_byte = __builtin_clz(diff) / 8; |
119 |
| -#endif |
120 |
| - return len + match_byte; |
121 |
| - } |
122 |
| - |
123 |
| - src0 += 4, src1 += 4, len += 4; |
124 |
| - } while (len < 256); |
125 |
| - |
126 |
| - return 256; |
127 |
| -} |
128 |
| - |
129 |
| -Z_INTERNAL uint32_t compare256_32(const uint8_t *src0, const uint8_t *src1) { |
130 |
| - return compare256_32_static(src0, src1); |
131 |
| -} |
132 |
| - |
133 |
| -#define LONGEST_MATCH longest_match_32 |
134 |
| -#define COMPARE256 compare256_32_static |
135 |
| - |
136 |
| -#include "match_tpl.h" |
137 |
| - |
138 |
| -#define LONGEST_MATCH_SLOW |
139 |
| -#define LONGEST_MATCH longest_match_slow_32 |
140 |
| -#define COMPARE256 compare256_32_static |
141 |
| - |
142 |
| -#include "match_tpl.h" |
143 |
| - |
144 |
| -#endif |
145 |
| - |
146 |
| -#if defined(HAVE_BUILTIN_CTZLL) && OPTIMAL_CMP >= 64 |
147 |
| -/* 64-bit integer comparison */ |
148 |
| -static inline uint32_t compare256_64_static(const uint8_t *src0, const uint8_t *src1) { |
149 |
| - uint32_t len = 0; |
150 |
| - |
151 |
| - do { |
152 |
| - uint64_t sv, mv, diff; |
153 |
| - |
154 |
| - sv = zng_memread_8(src0); |
155 |
| - mv = zng_memread_8(src1); |
156 |
| - |
157 |
| - diff = sv ^ mv; |
158 |
| - if (diff) { |
159 |
| -#if BYTE_ORDER == LITTLE_ENDIAN |
160 |
| - uint64_t match_byte = __builtin_ctzll(diff) / 8; |
161 |
| -#else |
162 |
| - uint64_t match_byte = __builtin_clzll(diff) / 8; |
163 |
| -#endif |
164 |
| - return len + (uint32_t)match_byte; |
165 |
| - } |
166 |
| - |
167 |
| - src0 += 8, src1 += 8, len += 8; |
168 |
| - } while (len < 256); |
169 |
| - |
170 |
| - return 256; |
171 |
| -} |
172 |
| - |
173 |
| -Z_INTERNAL uint32_t compare256_64(const uint8_t *src0, const uint8_t *src1) { |
174 |
| - return compare256_64_static(src0, src1); |
175 |
| -} |
176 |
| - |
177 |
| -#define LONGEST_MATCH longest_match_64 |
178 |
| -#define COMPARE256 compare256_64_static |
179 |
| - |
180 |
| -#include "match_tpl.h" |
181 |
| - |
182 |
| -#define LONGEST_MATCH_SLOW |
183 |
| -#define LONGEST_MATCH longest_match_slow_64 |
184 |
| -#define COMPARE256 compare256_64_static |
185 |
| - |
186 |
| -#include "match_tpl.h" |
187 |
| - |
188 |
| -#endif |
189 |
| - |
190 |
| -#endif |
0 commit comments