| 1 |
|
| /* utf8_range.c -- SIMD UTF-8 validation, vendored as explained in |
| 2 |
|
| * utf8_range.h. Compiled as its own object (see Makefile.PL) so the SIMD |
| 3 |
|
| * bodies stay out of the parser's translation unit. |
| 4 |
|
| */ |
| 5 |
|
| #include "utf8_range.h" |
| 6 |
|
| |
| 7 |
|
| #include <stdint.h> |
| 8 |
|
| #include <limits.h> |
| 9 |
|
| |
| 10 |
|
| /* x86 needs the target attribute plus __builtin_cpu_supports(), i.e. GCC 4.9 |
| 11 |
|
| * or clang 4; MSVC and older compilers fall back to the scalar path. NEON is |
| 12 |
|
| * baseline on aarch64, so no dispatch is needed there. */ |
| 13 |
|
| #if defined(__x86_64__) || defined(__i386__) |
| 14 |
|
| # if defined(__clang__) |
| 15 |
|
| # if __clang_major__ >= 4 && (!defined __apple_build_version__ || __apple_build_version__ > 5030040) |
| 16 |
|
| # define CJSON_HAVE_X86_SIMD 1 |
| 17 |
|
| # endif |
| 18 |
|
| # elif defined(__GNUC__) |
| 19 |
|
| # if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9) |
| 20 |
|
| # define CJSON_HAVE_X86_SIMD 1 |
| 21 |
|
| # endif |
| 22 |
|
| # endif |
| 23 |
|
| #endif |
| 24 |
|
| |
| 25 |
|
| #if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__)) |
| 26 |
|
| # define CJSON_HAVE_NEON 1 |
| 27 |
|
| #endif |
| 28 |
|
| |
| 29 |
|
| /* Portable scalar validation. Returns 0 if valid, else the 1-based index of |
| 30 |
|
| * the first offending character (only used as a boolean here). */ |
| 31 |
|
| |
| 32 |
|
| |
| 33 |
|
| /* |
| 34 |
|
| * http://www.unicode.org/versions/Unicode6.0.0/ch03.pdf - page 94 |
| 35 |
|
| * |
| 36 |
|
| * Table 3-7. Well-Formed UTF-8 Byte Sequences |
| 37 |
|
| * |
| 38 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 39 |
|
| * | Code Points | First Byte | Second Byte | Third Byte | Fourth Byte | |
| 40 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 41 |
|
| * | U+0000..U+007F | 00..7F | | | | |
| 42 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 43 |
|
| * | U+0080..U+07FF | C2..DF | 80..BF | | | |
| 44 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 45 |
|
| * | U+0800..U+0FFF | E0 | A0..BF | 80..BF | | |
| 46 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 47 |
|
| * | U+1000..U+CFFF | E1..EC | 80..BF | 80..BF | | |
| 48 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 49 |
|
| * | U+D000..U+D7FF | ED | 80..9F | 80..BF | | |
| 50 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 51 |
|
| * | U+E000..U+FFFF | EE..EF | 80..BF | 80..BF | | |
| 52 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 53 |
|
| * | U+10000..U+3FFFF | F0 | 90..BF | 80..BF | 80..BF | |
| 54 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 55 |
|
| * | U+40000..U+FFFFF | F1..F3 | 80..BF | 80..BF | 80..BF | |
| 56 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 57 |
|
| * | U+100000..U+10FFFF | F4 | 80..8F | 80..BF | 80..BF | |
| 58 |
|
| * +--------------------+------------+-------------+------------+-------------+ |
| 59 |
|
| */ |
| 60 |
|
| |
| 61 |
|
| /* Return 0 - success, >0 - index(1 based) of first error char */ |
| 62 |
70 |
| static int cjson_utf8_naive(const unsigned char *data, int len) |
| 63 |
|
| { |
| 64 |
70 |
| int err_pos = 1; |
| 65 |
|
| |
| 66 |
690 |
▶ | while (len) { |
Branch: 1 branch
| Branch | True | False | Total |
| while (len) { |
621 |
69 |
690 |
|
| 67 |
|
| int bytes; |
| 68 |
621 |
| const unsigned char byte1 = data[0]; |
| 69 |
|
| |
| 70 |
|
| /* 00..7F */ |
| 71 |
621 |
▶ | if (byte1 <= 0x7F) { |
Branch: 1 branch
| Branch | True | False | Total |
| if (byte1 <= 0x7F) { |
317 |
304 |
621 |
|
| 72 |
317 |
| bytes = 1; |
| 73 |
|
| /* C2..DF, 80..BF */ |
| 74 |
304 |
▶ | } else if (len >= 2 && byte1 >= 0xC2 && byte1 <= 0xDF && |
Branch: 3 branches
| Branch | True | False | Total |
| } else if (len >= 2 && byte1 >= 0xC2 && byte1 <= 0xDF && |
304 |
0 |
304 |
| } else if (len >= 2 && byte1 >= 0xC2 && byte1 <= 0xDF && |
304 |
0 |
304 |
| } else if (len >= 2 && byte1 >= 0xC2 && byte1 <= 0xDF && |
206 |
98 |
304 |
|
| 75 |
206 |
▶ | (signed char)data[1] <= (signed char)0xBF) { |
Branch: 1 branch
| Branch | True | False | Total |
| (signed char)data[1] <= (signed char)0xBF) { |
206 |
0 |
206 |
|
| 76 |
206 |
| bytes = 2; |
| 77 |
98 |
▶ | } else if (len >= 3) { |
Branch: 1 branch
| Branch | True | False | Total |
| } else if (len >= 3) { |
98 |
0 |
98 |
|
| 78 |
98 |
| const unsigned char byte2 = data[1]; |
| 79 |
|
| |
| 80 |
|
| /* Is byte2, byte3 between 0x80 ~ 0xBF */ |
| 81 |
98 |
| const int byte2_ok = (signed char)byte2 <= (signed char)0xBF; |
| 82 |
98 |
| const int byte3_ok = (signed char)data[2] <= (signed char)0xBF; |
| 83 |
|
| |
| 84 |
98 |
▶ | if (byte2_ok && byte3_ok && |
Branch: 3 branches
| Branch | True | False | Total |
| if (byte2_ok && byte3_ok && |
97 |
1 |
98 |
| if (byte2_ok && byte3_ok && |
97 |
0 |
97 |
| if (byte2_ok && byte3_ok && |
10 |
87 |
97 |
|
| 85 |
|
| /* E0, A0..BF, 80..BF */ |
| 86 |
97 |
▶ | ((byte1 == 0xE0 && byte2 >= 0xA0) || |
Branch: 2 branches
| Branch | True | False | Total |
| ((byte1 == 0xE0 && byte2 >= 0xA0) || |
0 |
10 |
10 |
| ((byte1 == 0xE0 && byte2 >= 0xA0) || |
87 |
0 |
87 |
|
| 87 |
|
| /* E1..EC, 80..BF, 80..BF */ |
| 88 |
87 |
▶ | (byte1 >= 0xE1 && byte1 <= 0xEC) || |
Branch: 2 branches
| Branch | True | False | Total |
| (byte1 >= 0xE1 && byte1 <= 0xEC) || |
79 |
8 |
87 |
| (byte1 >= 0xE1 && byte1 <= 0xEC) || |
0 |
79 |
79 |
|
| 89 |
|
| /* ED, 80..9F, 80..BF */ |
| 90 |
79 |
▶ | (byte1 == 0xED && byte2 <= 0x9F) || |
Branch: 2 branches
| Branch | True | False | Total |
| (byte1 == 0xED && byte2 <= 0x9F) || |
0 |
0 |
0 |
| (byte1 == 0xED && byte2 <= 0x9F) || |
79 |
0 |
79 |
|
| 91 |
|
| /* EE..EF, 80..BF, 80..BF */ |
| 92 |
79 |
▶ | (byte1 >= 0xEE && byte1 <= 0xEF))) { |
Branch: 1 branch
| Branch | True | False | Total |
| (byte1 >= 0xEE && byte1 <= 0xEF))) { |
0 |
79 |
79 |
|
| 93 |
18 |
| bytes = 3; |
| 94 |
80 |
▶ | } else if (len >= 4) { |
Branch: 1 branch
| Branch | True | False | Total |
| } else if (len >= 4) { |
80 |
0 |
80 |
|
| 95 |
|
| /* Is byte4 between 0x80 ~ 0xBF */ |
| 96 |
80 |
| const int byte4_ok = (signed char)data[3] <= (signed char)0xBF; |
| 97 |
|
| |
| 98 |
80 |
▶ | if (byte2_ok && byte3_ok && byte4_ok && |
Branch: 4 branches
| Branch | True | False | Total |
| if (byte2_ok && byte3_ok && byte4_ok && |
79 |
1 |
80 |
| if (byte2_ok && byte3_ok && byte4_ok && |
79 |
0 |
79 |
| if (byte2_ok && byte3_ok && byte4_ok && |
79 |
0 |
79 |
| if (byte2_ok && byte3_ok && byte4_ok && |
18 |
61 |
79 |
|
| 99 |
|
| /* F0, 90..BF, 80..BF, 80..BF */ |
| 100 |
79 |
▶ | ((byte1 == 0xF0 && byte2 >= 0x90) || |
Branch: 2 branches
| Branch | True | False | Total |
| ((byte1 == 0xF0 && byte2 >= 0x90) || |
0 |
18 |
18 |
| ((byte1 == 0xF0 && byte2 >= 0x90) || |
61 |
0 |
61 |
|
| 101 |
|
| /* F1..F3, 80..BF, 80..BF, 80..BF */ |
| 102 |
61 |
▶ | (byte1 >= 0xF1 && byte1 <= 0xF3) || |
Branch: 2 branches
| Branch | True | False | Total |
| (byte1 >= 0xF1 && byte1 <= 0xF3) || |
0 |
61 |
61 |
| (byte1 >= 0xF1 && byte1 <= 0xF3) || |
0 |
0 |
0 |
|
| 103 |
|
| /* F4, 80..8F, 80..BF, 80..BF */ |
| 104 |
0 |
▶ | (byte1 == 0xF4 && byte2 <= 0x8F))) { |
Branch: 1 branch
| Branch | True | False | Total |
| (byte1 == 0xF4 && byte2 <= 0x8F))) { |
0 |
0 |
0 |
|
| 105 |
79 |
| bytes = 4; |
| 106 |
|
| } else { |
| 107 |
1 |
| return err_pos; |
| 108 |
|
| } |
| 109 |
|
| } else { |
| 110 |
0 |
| return err_pos; |
| 111 |
|
| } |
| 112 |
|
| } else { |
| 113 |
0 |
| return err_pos; |
| 114 |
|
| } |
| 115 |
|
| |
| 116 |
620 |
| len -= bytes; |
| 117 |
620 |
| err_pos += bytes; |
| 118 |
620 |
| data += bytes; |
| 119 |
|
| } |
| 120 |
|
| |
| 121 |
69 |
| return 0; |
| 122 |
|
| } |
| 123 |
|
| |
| 124 |
|
| /* ------------------------------------------------------------------ x86 --- */ |
| 125 |
|
| #ifdef CJSON_HAVE_X86_SIMD |
| 126 |
|
| # include <immintrin.h> |
| 127 |
|
| # define CJSON_SIMD_FN(opts) __attribute__((target(opts))) |
| 128 |
|
| #endif |
| 129 |
|
| |
| 130 |
|
| #ifdef CJSON_HAVE_X86_SIMD |
| 131 |
|
| |
| 132 |
|
| |
| 133 |
|
| |
| 134 |
|
| |
| 135 |
|
| /* |
| 136 |
|
| * Map high nibble of "First Byte" to legal character length minus 1 |
| 137 |
|
| * 0x00 ~ 0xBF --> 0 |
| 138 |
|
| * 0xC0 ~ 0xDF --> 1 |
| 139 |
|
| * 0xE0 ~ 0xEF --> 2 |
| 140 |
|
| * 0xF0 ~ 0xFF --> 3 |
| 141 |
|
| */ |
| 142 |
|
| static const int8_t cjson_sse_first_len_tbl[] = { |
| 143 |
|
| 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 2, 3, |
| 144 |
|
| }; |
| 145 |
|
| |
| 146 |
|
| /* Map "First Byte" to 8-th item of range table (0xC2 ~ 0xF4) */ |
| 147 |
|
| static const int8_t cjson_sse_first_range_tbl[] = { |
| 148 |
|
| 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, |
| 149 |
|
| }; |
| 150 |
|
| |
| 151 |
|
| /* |
| 152 |
|
| * Range table, map range index to min and max values |
| 153 |
|
| * Index 0 : 00 ~ 7F (First Byte, ascii) |
| 154 |
|
| * Index 1,2,3: 80 ~ BF (Second, Third, Fourth Byte) |
| 155 |
|
| * Index 4 : A0 ~ BF (Second Byte after E0) |
| 156 |
|
| * Index 5 : 80 ~ 9F (Second Byte after ED) |
| 157 |
|
| * Index 6 : 90 ~ BF (Second Byte after F0) |
| 158 |
|
| * Index 7 : 80 ~ 8F (Second Byte after F4) |
| 159 |
|
| * Index 8 : C2 ~ F4 (First Byte, non ascii) |
| 160 |
|
| * Index 9~15 : illegal: i >= 127 && i <= -128 |
| 161 |
|
| */ |
| 162 |
|
| static const int8_t cjson_sse_range_min_tbl[] = { |
| 163 |
|
| 0x00, 0x80, 0x80, 0x80, 0xA0, 0x80, 0x90, 0x80, |
| 164 |
|
| 0xC2, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, |
| 165 |
|
| }; |
| 166 |
|
| static const int8_t cjson_sse_range_max_tbl[] = { |
| 167 |
|
| 0x7F, 0xBF, 0xBF, 0xBF, 0xBF, 0x9F, 0xBF, 0x8F, |
| 168 |
|
| 0xF4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, |
| 169 |
|
| }; |
| 170 |
|
| |
| 171 |
|
| /* |
| 172 |
|
| * Tables for fast handling of four special First Bytes(E0,ED,F0,F4), after |
| 173 |
|
| * which the Second Byte are not 80~BF. It contains "range index adjustment". |
| 174 |
|
| * +------------+---------------+------------------+----------------+ |
| 175 |
|
| * | First Byte | original range| range adjustment | adjusted range | |
| 176 |
|
| * +------------+---------------+------------------+----------------+ |
| 177 |
|
| * | E0 | 2 | 2 | 4 | |
| 178 |
|
| * +------------+---------------+------------------+----------------+ |
| 179 |
|
| * | ED | 2 | 3 | 5 | |
| 180 |
|
| * +------------+---------------+------------------+----------------+ |
| 181 |
|
| * | F0 | 3 | 3 | 6 | |
| 182 |
|
| * +------------+---------------+------------------+----------------+ |
| 183 |
|
| * | F4 | 4 | 4 | 8 | |
| 184 |
|
| * +------------+---------------+------------------+----------------+ |
| 185 |
|
| */ |
| 186 |
|
| /* index1 -> E0, index14 -> ED */ |
| 187 |
|
| static const int8_t cjson_sse_df_ee_tbl[] = { |
| 188 |
|
| 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, |
| 189 |
|
| }; |
| 190 |
|
| /* index1 -> F0, index5 -> F4 */ |
| 191 |
|
| static const int8_t cjson_sse_ef_fe_tbl[] = { |
| 192 |
|
| 0, 3, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 193 |
|
| }; |
| 194 |
|
| |
| 195 |
|
| #define RET_ERR_IDX 0 /* Define 1 to return index of first error char */ |
| 196 |
|
| |
| 197 |
|
| /* 5x faster than naive method */ |
| 198 |
|
| /* Return 0 - success, -1 - error, >0 - first error char(if RET_ERR_IDX = 1) */ |
| 199 |
|
| CJSON_SIMD_FN("ssse3,sse4.1") |
| 200 |
0 |
| static int cjson_utf8_range_sse(const unsigned char *data, int len) |
| 201 |
|
| { |
| 202 |
|
| #if RET_ERR_IDX |
| 203 |
|
| int err_pos = 1; |
| 204 |
|
| #endif |
| 205 |
|
| |
| 206 |
0 |
▶ | if (len >= 16) { |
Branch: 1 branch
| Branch | True | False | Total |
| if (len >= 16) { |
0 |
0 |
0 |
|
| 207 |
0 |
| __m128i prev_input = _mm_set1_epi8(0); |
| 208 |
0 |
| __m128i prev_first_len = _mm_set1_epi8(0); |
| 209 |
|
| |
| 210 |
|
| /* Cached tables */ |
| 211 |
|
| const __m128i first_len_tbl = |
| 212 |
0 |
| _mm_loadu_si128((const __m128i *)cjson_sse_first_len_tbl); |
| 213 |
|
| const __m128i first_range_tbl = |
| 214 |
0 |
| _mm_loadu_si128((const __m128i *)cjson_sse_first_range_tbl); |
| 215 |
|
| const __m128i range_min_tbl = |
| 216 |
0 |
| _mm_loadu_si128((const __m128i *)cjson_sse_range_min_tbl); |
| 217 |
|
| const __m128i range_max_tbl = |
| 218 |
0 |
| _mm_loadu_si128((const __m128i *)cjson_sse_range_max_tbl); |
| 219 |
|
| const __m128i df_ee_tbl = |
| 220 |
0 |
| _mm_loadu_si128((const __m128i *)cjson_sse_df_ee_tbl); |
| 221 |
|
| const __m128i ef_fe_tbl = |
| 222 |
0 |
| _mm_loadu_si128((const __m128i *)cjson_sse_ef_fe_tbl); |
| 223 |
|
| |
| 224 |
0 |
| __m128i error = _mm_set1_epi8(0); |
| 225 |
|
| |
| 226 |
0 |
▶ | while (len >= 16) { |
Branch: 1 branch
| Branch | True | False | Total |
| while (len >= 16) { |
0 |
0 |
0 |
|
| 227 |
|
| /* declarations hoisted to the block top: perl built before 5.36 |
| 228 |
|
| passes -Werror=declaration-after-statement */ |
| 229 |
|
| __m128i tmp, shift1, pos, range2, minv, maxv; |
| 230 |
0 |
| const __m128i input = _mm_loadu_si128((const __m128i *)data); |
| 231 |
|
| |
| 232 |
|
| /* high_nibbles = input >> 4 */ |
| 233 |
|
| const __m128i high_nibbles = |
| 234 |
0 |
| _mm_and_si128(_mm_srli_epi16(input, 4), _mm_set1_epi8(0x0F)); |
| 235 |
|
| |
| 236 |
|
| /* first_len = legal character length minus 1 */ |
| 237 |
|
| /* 0 for 00~7F, 1 for C0~DF, 2 for E0~EF, 3 for F0~FF */ |
| 238 |
|
| /* first_len = first_len_tbl[high_nibbles] */ |
| 239 |
0 |
| __m128i first_len = _mm_shuffle_epi8(first_len_tbl, high_nibbles); |
| 240 |
|
| |
| 241 |
|
| /* First Byte: set range index to 8 for bytes within 0xC0 ~ 0xFF */ |
| 242 |
|
| /* range = first_range_tbl[high_nibbles] */ |
| 243 |
0 |
| __m128i range = _mm_shuffle_epi8(first_range_tbl, high_nibbles); |
| 244 |
|
| |
| 245 |
|
| /* Second Byte: set range index to first_len */ |
| 246 |
|
| /* 0 for 00~7F, 1 for C0~DF, 2 for E0~EF, 3 for F0~FF */ |
| 247 |
|
| /* range |= (first_len, prev_first_len) << 1 byte */ |
| 248 |
0 |
| range = _mm_or_si128( |
| 249 |
0 |
| range, _mm_alignr_epi8(first_len, prev_first_len, 15)); |
| 250 |
|
| |
| 251 |
|
| /* Third Byte: set range index to saturate_sub(first_len, 1) */ |
| 252 |
|
| /* 0 for 00~7F, 0 for C0~DF, 1 for E0~EF, 2 for F0~FF */ |
| 253 |
|
| /* tmp = (first_len, prev_first_len) << 2 bytes */ |
| 254 |
0 |
| tmp = _mm_alignr_epi8(first_len, prev_first_len, 14); |
| 255 |
|
| /* tmp = saturate_sub(tmp, 1) */ |
| 256 |
0 |
| tmp = _mm_subs_epu8(tmp, _mm_set1_epi8(1)); |
| 257 |
|
| /* range |= tmp */ |
| 258 |
0 |
| range = _mm_or_si128(range, tmp); |
| 259 |
|
| |
| 260 |
|
| /* Fourth Byte: set range index to saturate_sub(first_len, 2) */ |
| 261 |
|
| /* 0 for 00~7F, 0 for C0~DF, 0 for E0~EF, 1 for F0~FF */ |
| 262 |
|
| /* tmp = (first_len, prev_first_len) << 3 bytes */ |
| 263 |
0 |
| tmp = _mm_alignr_epi8(first_len, prev_first_len, 13); |
| 264 |
|
| /* tmp = saturate_sub(tmp, 2) */ |
| 265 |
0 |
| tmp = _mm_subs_epu8(tmp, _mm_set1_epi8(2)); |
| 266 |
|
| /* range |= tmp */ |
| 267 |
0 |
| range = _mm_or_si128(range, tmp); |
| 268 |
|
| |
| 269 |
|
| /* |
| 270 |
|
| * Now we have below range indices caluclated |
| 271 |
|
| * Correct cases: |
| 272 |
|
| * - 8 for C0~FF |
| 273 |
|
| * - 3 for 1st byte after F0~FF |
| 274 |
|
| * - 2 for 1st byte after E0~EF or 2nd byte after F0~FF |
| 275 |
|
| * - 1 for 1st byte after C0~DF or 2nd byte after E0~EF or |
| 276 |
|
| * 3rd byte after F0~FF |
| 277 |
|
| * - 0 for others |
| 278 |
|
| * Error cases: |
| 279 |
|
| * 9,10,11 if non ascii First Byte overlaps |
| 280 |
|
| * E.g., F1 80 C2 90 --> 8 3 10 2, where 10 indicates error |
| 281 |
|
| */ |
| 282 |
|
| |
| 283 |
|
| /* Adjust Second Byte range for special First Bytes(E0,ED,F0,F4) */ |
| 284 |
|
| /* Overlaps lead to index 9~15, which are illegal in range table */ |
| 285 |
|
| /* shift1 = (input, prev_input) << 1 byte */ |
| 286 |
0 |
| shift1 = _mm_alignr_epi8(input, prev_input, 15); |
| 287 |
0 |
| pos = _mm_sub_epi8(shift1, _mm_set1_epi8(0xEF)); |
| 288 |
|
| /* |
| 289 |
|
| * shift1: | EF F0 ... FE | FF 00 ... ... DE | DF E0 ... EE | |
| 290 |
|
| * pos: | 0 1 15 | 16 17 239| 240 241 255| |
| 291 |
|
| * pos-240: | 0 0 0 | 0 0 0 | 0 1 15 | |
| 292 |
|
| * pos+112: | 112 113 127| >= 128 | >= 128 | |
| 293 |
|
| */ |
| 294 |
0 |
| tmp = _mm_subs_epu8(pos, _mm_set1_epi8(0xF0)); |
| 295 |
0 |
| range2 = _mm_shuffle_epi8(df_ee_tbl, tmp); |
| 296 |
0 |
| tmp = _mm_adds_epu8(pos, _mm_set1_epi8(0x70)); |
| 297 |
0 |
| range2 = _mm_add_epi8(range2, _mm_shuffle_epi8(ef_fe_tbl, tmp)); |
| 298 |
|
| |
| 299 |
0 |
| range = _mm_add_epi8(range, range2); |
| 300 |
|
| |
| 301 |
|
| /* Load min and max values per calculated range index */ |
| 302 |
0 |
| minv = _mm_shuffle_epi8(range_min_tbl, range); |
| 303 |
0 |
| maxv = _mm_shuffle_epi8(range_max_tbl, range); |
| 304 |
|
| |
| 305 |
|
| /* Check value range */ |
| 306 |
|
| #if RET_ERR_IDX |
| 307 |
|
| error = _mm_cmplt_epi8(input, minv); |
| 308 |
|
| error = _mm_or_si128(error, _mm_cmpgt_epi8(input, maxv)); |
| 309 |
|
| /* 5% performance drop from this conditional branch */ |
| 310 |
|
| if (!_mm_testz_si128(error, error)) |
| 311 |
|
| break; |
| 312 |
|
| #else |
| 313 |
|
| /* error |= (input < minv) | (input > maxv) */ |
| 314 |
0 |
| tmp = _mm_or_si128( |
| 315 |
|
| _mm_cmplt_epi8(input, minv), |
| 316 |
|
| _mm_cmpgt_epi8(input, maxv) |
| 317 |
|
| ); |
| 318 |
0 |
| error = _mm_or_si128(error, tmp); |
| 319 |
|
| #endif |
| 320 |
|
| |
| 321 |
0 |
| prev_input = input; |
| 322 |
0 |
| prev_first_len = first_len; |
| 323 |
|
| |
| 324 |
0 |
| data += 16; |
| 325 |
0 |
| len -= 16; |
| 326 |
|
| #if RET_ERR_IDX |
| 327 |
|
| err_pos += 16; |
| 328 |
|
| #endif |
| 329 |
|
| } |
| 330 |
|
| |
| 331 |
|
| #if RET_ERR_IDX |
| 332 |
|
| /* Error in first 16 bytes */ |
| 333 |
|
| if (err_pos == 1) |
| 334 |
|
| goto do_naive; |
| 335 |
|
| #else |
| 336 |
0 |
▶ | if (!_mm_testz_si128(error, error)) |
Branch: 1 branch
| Branch | True | False | Total |
| if (!_mm_testz_si128(error, error)) |
0 |
0 |
0 |
|
| 337 |
0 |
| return -1; |
| 338 |
|
| #endif |
| 339 |
|
| |
| 340 |
|
| /* Find previous token (not 80~BF) */ |
| 341 |
|
| { |
| 342 |
0 |
| int32_t token4 = _mm_extract_epi32(prev_input, 3); |
| 343 |
0 |
| const int8_t *token = (const int8_t *)&token4; |
| 344 |
0 |
| int lookahead = 0; |
| 345 |
0 |
▶ | if (token[3] > (int8_t)0xBF) |
Branch: 1 branch
| Branch | True | False | Total |
| if (token[3] > (int8_t)0xBF) |
0 |
0 |
0 |
|
| 346 |
0 |
| lookahead = 1; |
| 347 |
0 |
▶ | else if (token[2] > (int8_t)0xBF) |
Branch: 1 branch
| Branch | True | False | Total |
| else if (token[2] > (int8_t)0xBF) |
0 |
0 |
0 |
|
| 348 |
0 |
| lookahead = 2; |
| 349 |
0 |
▶ | else if (token[1] > (int8_t)0xBF) |
Branch: 1 branch
| Branch | True | False | Total |
| else if (token[1] > (int8_t)0xBF) |
0 |
0 |
0 |
|
| 350 |
0 |
| lookahead = 3; |
| 351 |
|
| |
| 352 |
0 |
| data -= lookahead; |
| 353 |
0 |
| len += lookahead; |
| 354 |
|
| #if RET_ERR_IDX |
| 355 |
|
| err_pos -= lookahead; |
| 356 |
|
| #endif |
| 357 |
|
| } |
| 358 |
|
| } |
| 359 |
|
| |
| 360 |
|
| /* Check remaining bytes with naive method */ |
| 361 |
|
| #if RET_ERR_IDX |
| 362 |
|
| int err_pos2; |
| 363 |
|
| do_naive: |
| 364 |
|
| err_pos2 = cjson_utf8_naive(data, len); |
| 365 |
|
| if (err_pos2) |
| 366 |
|
| return err_pos + err_pos2 - 1; |
| 367 |
|
| return 0; |
| 368 |
|
| #else |
| 369 |
0 |
| return cjson_utf8_naive(data, len); |
| 370 |
|
| #endif |
| 371 |
|
| } |
| 372 |
|
| #endif /* CJSON_HAVE_X86_SIMD */ |
| 373 |
|
| #ifdef CJSON_HAVE_X86_SIMD |
| 374 |
|
| |
| 375 |
|
| |
| 376 |
|
| |
| 377 |
|
| |
| 378 |
|
| /* |
| 379 |
|
| * Map high nibble of "First Byte" to legal character length minus 1 |
| 380 |
|
| * 0x00 ~ 0xBF --> 0 |
| 381 |
|
| * 0xC0 ~ 0xDF --> 1 |
| 382 |
|
| * 0xE0 ~ 0xEF --> 2 |
| 383 |
|
| * 0xF0 ~ 0xFF --> 3 |
| 384 |
|
| */ |
| 385 |
|
| static const int8_t cjson_avx2_first_len_tbl[] = { |
| 386 |
|
| 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 2, 3, |
| 387 |
|
| 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 2, 3, |
| 388 |
|
| }; |
| 389 |
|
| |
| 390 |
|
| /* Map "First Byte" to 8-th item of range table (0xC2 ~ 0xF4) */ |
| 391 |
|
| static const int8_t cjson_avx2_first_range_tbl[] = { |
| 392 |
|
| 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, |
| 393 |
|
| 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, |
| 394 |
|
| }; |
| 395 |
|
| |
| 396 |
|
| /* |
| 397 |
|
| * Range table, map range index to min and max values |
| 398 |
|
| * Index 0 : 00 ~ 7F (First Byte, ascii) |
| 399 |
|
| * Index 1,2,3: 80 ~ BF (Second, Third, Fourth Byte) |
| 400 |
|
| * Index 4 : A0 ~ BF (Second Byte after E0) |
| 401 |
|
| * Index 5 : 80 ~ 9F (Second Byte after ED) |
| 402 |
|
| * Index 6 : 90 ~ BF (Second Byte after F0) |
| 403 |
|
| * Index 7 : 80 ~ 8F (Second Byte after F4) |
| 404 |
|
| * Index 8 : C2 ~ F4 (First Byte, non ascii) |
| 405 |
|
| * Index 9~15 : illegal: i >= 127 && i <= -128 |
| 406 |
|
| */ |
| 407 |
|
| static const int8_t cjson_avx2_range_min_tbl[] = { |
| 408 |
|
| 0x00, 0x80, 0x80, 0x80, 0xA0, 0x80, 0x90, 0x80, |
| 409 |
|
| 0xC2, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, |
| 410 |
|
| 0x00, 0x80, 0x80, 0x80, 0xA0, 0x80, 0x90, 0x80, |
| 411 |
|
| 0xC2, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, 0x7F, |
| 412 |
|
| }; |
| 413 |
|
| static const int8_t cjson_avx2_range_max_tbl[] = { |
| 414 |
|
| 0x7F, 0xBF, 0xBF, 0xBF, 0xBF, 0x9F, 0xBF, 0x8F, |
| 415 |
|
| 0xF4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, |
| 416 |
|
| 0x7F, 0xBF, 0xBF, 0xBF, 0xBF, 0x9F, 0xBF, 0x8F, |
| 417 |
|
| 0xF4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, |
| 418 |
|
| }; |
| 419 |
|
| |
| 420 |
|
| /* |
| 421 |
|
| * Tables for fast handling of four special First Bytes(E0,ED,F0,F4), after |
| 422 |
|
| * which the Second Byte are not 80~BF. It contains "range index adjustment". |
| 423 |
|
| * +------------+---------------+------------------+----------------+ |
| 424 |
|
| * | First Byte | original range| range adjustment | adjusted range | |
| 425 |
|
| * +------------+---------------+------------------+----------------+ |
| 426 |
|
| * | E0 | 2 | 2 | 4 | |
| 427 |
|
| * +------------+---------------+------------------+----------------+ |
| 428 |
|
| * | ED | 2 | 3 | 5 | |
| 429 |
|
| * +------------+---------------+------------------+----------------+ |
| 430 |
|
| * | F0 | 3 | 3 | 6 | |
| 431 |
|
| * +------------+---------------+------------------+----------------+ |
| 432 |
|
| * | F4 | 4 | 4 | 8 | |
| 433 |
|
| * +------------+---------------+------------------+----------------+ |
| 434 |
|
| */ |
| 435 |
|
| /* index1 -> E0, index14 -> ED */ |
| 436 |
|
| static const int8_t cjson_avx2_df_ee_tbl[] = { |
| 437 |
|
| 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, |
| 438 |
|
| 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, |
| 439 |
|
| }; |
| 440 |
|
| /* index1 -> F0, index5 -> F4 */ |
| 441 |
|
| static const int8_t cjson_avx2_ef_fe_tbl[] = { |
| 442 |
|
| 0, 3, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 443 |
|
| 0, 3, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 444 |
|
| }; |
| 445 |
|
| |
| 446 |
|
| #define RET_ERR_IDX 0 /* Define 1 to return index of first error char */ |
| 447 |
|
| |
| 448 |
|
| CJSON_SIMD_FN("avx2") |
| 449 |
210 |
| static inline __m256i push_last_byte_of_a_to_b(__m256i a, __m256i b) { |
| 450 |
210 |
| return _mm256_alignr_epi8(b, _mm256_permute2x128_si256(a, b, 0x21), 15); |
| 451 |
|
| } |
| 452 |
|
| |
| 453 |
|
| CJSON_SIMD_FN("avx2") |
| 454 |
105 |
| static inline __m256i push_last_2bytes_of_a_to_b(__m256i a, __m256i b) { |
| 455 |
105 |
| return _mm256_alignr_epi8(b, _mm256_permute2x128_si256(a, b, 0x21), 14); |
| 456 |
|
| } |
| 457 |
|
| |
| 458 |
|
| CJSON_SIMD_FN("avx2") |
| 459 |
105 |
| static inline __m256i push_last_3bytes_of_a_to_b(__m256i a, __m256i b) { |
| 460 |
105 |
| return _mm256_alignr_epi8(b, _mm256_permute2x128_si256(a, b, 0x21), 13); |
| 461 |
|
| } |
| 462 |
|
| |
| 463 |
|
| /* 5x faster than naive method */ |
| 464 |
|
| /* Return 0 - success, -1 - error, >0 - first error char(if RET_ERR_IDX = 1) */ |
| 465 |
|
| CJSON_SIMD_FN("avx2") |
| 466 |
74 |
| static int cjson_utf8_range_avx2(const unsigned char *data, int len) |
| 467 |
|
| { |
| 468 |
|
| #if RET_ERR_IDX |
| 469 |
|
| int err_pos = 1; |
| 470 |
|
| #endif |
| 471 |
|
| |
| 472 |
74 |
▶ | if (len >= 32) { |
Branch: 1 branch
| Branch | True | False | Total |
| if (len >= 32) { |
53 |
21 |
74 |
|
| 473 |
53 |
| __m256i prev_input = _mm256_set1_epi8(0); |
| 474 |
53 |
| __m256i prev_first_len = _mm256_set1_epi8(0); |
| 475 |
|
| |
| 476 |
|
| /* Cached tables */ |
| 477 |
|
| const __m256i first_len_tbl = |
| 478 |
53 |
| _mm256_loadu_si256((const __m256i *)cjson_avx2_first_len_tbl); |
| 479 |
|
| const __m256i first_range_tbl = |
| 480 |
53 |
| _mm256_loadu_si256((const __m256i *)cjson_avx2_first_range_tbl); |
| 481 |
|
| const __m256i range_min_tbl = |
| 482 |
53 |
| _mm256_loadu_si256((const __m256i *)cjson_avx2_range_min_tbl); |
| 483 |
|
| const __m256i range_max_tbl = |
| 484 |
53 |
| _mm256_loadu_si256((const __m256i *)cjson_avx2_range_max_tbl); |
| 485 |
|
| const __m256i df_ee_tbl = |
| 486 |
53 |
| _mm256_loadu_si256((const __m256i *)cjson_avx2_df_ee_tbl); |
| 487 |
|
| const __m256i ef_fe_tbl = |
| 488 |
53 |
| _mm256_loadu_si256((const __m256i *)cjson_avx2_ef_fe_tbl); |
| 489 |
|
| |
| 490 |
|
| #if !RET_ERR_IDX |
| 491 |
53 |
| __m256i error1 = _mm256_set1_epi8(0); |
| 492 |
53 |
| __m256i error2 = _mm256_set1_epi8(0); |
| 493 |
|
| __m256i error; |
| 494 |
|
| #endif |
| 495 |
|
| |
| 496 |
158 |
▶ | while (len >= 32) { |
Branch: 1 branch
| Branch | True | False | Total |
| while (len >= 32) { |
105 |
53 |
158 |
|
| 497 |
|
| /* declarations hoisted to the block top, see the SSE version */ |
| 498 |
|
| __m256i tmp1, tmp2, shift1, pos, range2, minv, maxv; |
| 499 |
105 |
| const __m256i input = _mm256_loadu_si256((const __m256i *)data); |
| 500 |
|
| |
| 501 |
|
| /* high_nibbles = input >> 4 */ |
| 502 |
|
| const __m256i high_nibbles = |
| 503 |
315 |
| _mm256_and_si256(_mm256_srli_epi16(input, 4), _mm256_set1_epi8(0x0F)); |
| 504 |
|
| |
| 505 |
|
| /* first_len = legal character length minus 1 */ |
| 506 |
|
| /* 0 for 00~7F, 1 for C0~DF, 2 for E0~EF, 3 for F0~FF */ |
| 507 |
|
| /* first_len = first_len_tbl[high_nibbles] */ |
| 508 |
105 |
| __m256i first_len = _mm256_shuffle_epi8(first_len_tbl, high_nibbles); |
| 509 |
|
| |
| 510 |
|
| /* First Byte: set range index to 8 for bytes within 0xC0 ~ 0xFF */ |
| 511 |
|
| /* range = first_range_tbl[high_nibbles] */ |
| 512 |
105 |
| __m256i range = _mm256_shuffle_epi8(first_range_tbl, high_nibbles); |
| 513 |
|
| |
| 514 |
|
| /* Second Byte: set range index to first_len */ |
| 515 |
|
| /* 0 for 00~7F, 1 for C0~DF, 2 for E0~EF, 3 for F0~FF */ |
| 516 |
|
| /* range |= (first_len, prev_first_len) << 1 byte */ |
| 517 |
105 |
| range = _mm256_or_si256( |
| 518 |
|
| range, push_last_byte_of_a_to_b(prev_first_len, first_len)); |
| 519 |
|
| |
| 520 |
|
| /* Third Byte: set range index to saturate_sub(first_len, 1) */ |
| 521 |
|
| /* 0 for 00~7F, 0 for C0~DF, 1 for E0~EF, 2 for F0~FF */ |
| 522 |
|
| |
| 523 |
|
| /* tmp1 = (first_len, prev_first_len) << 2 bytes */ |
| 524 |
105 |
| tmp1 = push_last_2bytes_of_a_to_b(prev_first_len, first_len); |
| 525 |
|
| /* tmp2 = saturate_sub(tmp1, 1) */ |
| 526 |
210 |
| tmp2 = _mm256_subs_epu8(tmp1, _mm256_set1_epi8(1)); |
| 527 |
|
| |
| 528 |
|
| /* range |= tmp2 */ |
| 529 |
105 |
| range = _mm256_or_si256(range, tmp2); |
| 530 |
|
| |
| 531 |
|
| /* Fourth Byte: set range index to saturate_sub(first_len, 2) */ |
| 532 |
|
| /* 0 for 00~7F, 0 for C0~DF, 0 for E0~EF, 1 for F0~FF */ |
| 533 |
|
| /* tmp1 = (first_len, prev_first_len) << 3 bytes */ |
| 534 |
105 |
| tmp1 = push_last_3bytes_of_a_to_b(prev_first_len, first_len); |
| 535 |
|
| /* tmp2 = saturate_sub(tmp1, 2) */ |
| 536 |
210 |
| tmp2 = _mm256_subs_epu8(tmp1, _mm256_set1_epi8(2)); |
| 537 |
|
| /* range |= tmp2 */ |
| 538 |
105 |
| range = _mm256_or_si256(range, tmp2); |
| 539 |
|
| |
| 540 |
|
| /* |
| 541 |
|
| * Now we have below range indices caluclated |
| 542 |
|
| * Correct cases: |
| 543 |
|
| * - 8 for C0~FF |
| 544 |
|
| * - 3 for 1st byte after F0~FF |
| 545 |
|
| * - 2 for 1st byte after E0~EF or 2nd byte after F0~FF |
| 546 |
|
| * - 1 for 1st byte after C0~DF or 2nd byte after E0~EF or |
| 547 |
|
| * 3rd byte after F0~FF |
| 548 |
|
| * - 0 for others |
| 549 |
|
| * Error cases: |
| 550 |
|
| * 9,10,11 if non ascii First Byte overlaps |
| 551 |
|
| * E.g., F1 80 C2 90 --> 8 3 10 2, where 10 indicates error |
| 552 |
|
| */ |
| 553 |
|
| |
| 554 |
|
| /* Adjust Second Byte range for special First Bytes(E0,ED,F0,F4) */ |
| 555 |
|
| /* Overlaps lead to index 9~15, which are illegal in range table */ |
| 556 |
|
| /* shift1 = (input, prev_input) << 1 byte */ |
| 557 |
105 |
| shift1 = push_last_byte_of_a_to_b(prev_input, input); |
| 558 |
210 |
| pos = _mm256_sub_epi8(shift1, _mm256_set1_epi8(0xEF)); |
| 559 |
|
| /* |
| 560 |
|
| * shift1: | EF F0 ... FE | FF 00 ... ... DE | DF E0 ... EE | |
| 561 |
|
| * pos: | 0 1 15 | 16 17 239| 240 241 255| |
| 562 |
|
| * pos-240: | 0 0 0 | 0 0 0 | 0 1 15 | |
| 563 |
|
| * pos+112: | 112 113 127| >= 128 | >= 128 | |
| 564 |
|
| */ |
| 565 |
210 |
| tmp1 = _mm256_subs_epu8(pos, _mm256_set1_epi8(240)); |
| 566 |
105 |
| range2 = _mm256_shuffle_epi8(df_ee_tbl, tmp1); |
| 567 |
210 |
| tmp2 = _mm256_adds_epu8(pos, _mm256_set1_epi8(112)); |
| 568 |
210 |
| range2 = _mm256_add_epi8(range2, _mm256_shuffle_epi8(ef_fe_tbl, tmp2)); |
| 569 |
|
| |
| 570 |
105 |
| range = _mm256_add_epi8(range, range2); |
| 571 |
|
| |
| 572 |
|
| /* Load min and max values per calculated range index */ |
| 573 |
105 |
| minv = _mm256_shuffle_epi8(range_min_tbl, range); |
| 574 |
105 |
| maxv = _mm256_shuffle_epi8(range_max_tbl, range); |
| 575 |
|
| |
| 576 |
|
| /* Check value range */ |
| 577 |
|
| #if RET_ERR_IDX |
| 578 |
|
| __m256i error = _mm256_cmpgt_epi8(minv, input); |
| 579 |
|
| error = _mm256_or_si256(error, _mm256_cmpgt_epi8(input, maxv)); |
| 580 |
|
| /* 5% performance drop from this conditional branch */ |
| 581 |
|
| if (!_mm256_testz_si256(error, error)) |
| 582 |
|
| break; |
| 583 |
|
| #else |
| 584 |
210 |
| error1 = _mm256_or_si256(error1, _mm256_cmpgt_epi8(minv, input)); |
| 585 |
105 |
| error2 = _mm256_or_si256(error2, _mm256_cmpgt_epi8(input, maxv)); |
| 586 |
|
| #endif |
| 587 |
|
| |
| 588 |
105 |
| prev_input = input; |
| 589 |
105 |
| prev_first_len = first_len; |
| 590 |
|
| |
| 591 |
105 |
| data += 32; |
| 592 |
105 |
| len -= 32; |
| 593 |
|
| #if RET_ERR_IDX |
| 594 |
|
| err_pos += 32; |
| 595 |
|
| #endif |
| 596 |
|
| } |
| 597 |
|
| |
| 598 |
|
| #if RET_ERR_IDX |
| 599 |
|
| /* Error in first 16 bytes */ |
| 600 |
|
| if (err_pos == 1) |
| 601 |
|
| goto do_naive; |
| 602 |
|
| #else |
| 603 |
53 |
| error = _mm256_or_si256(error1, error2); |
| 604 |
53 |
▶ | if (!_mm256_testz_si256(error, error)) |
Branch: 1 branch
| Branch | True | False | Total |
| if (!_mm256_testz_si256(error, error)) |
4 |
49 |
53 |
|
| 605 |
4 |
| return -1; |
| 606 |
|
| #endif |
| 607 |
|
| |
| 608 |
|
| /* Find previous token (not 80~BF) */ |
| 609 |
|
| { |
| 610 |
49 |
| int32_t token4 = _mm256_extract_epi32(prev_input, 7); |
| 611 |
49 |
| const int8_t *token = (const int8_t *)&token4; |
| 612 |
49 |
| int lookahead = 0; |
| 613 |
49 |
▶ | if (token[3] > (int8_t)0xBF) |
Branch: 1 branch
| Branch | True | False | Total |
| if (token[3] > (int8_t)0xBF) |
18 |
31 |
49 |
|
| 614 |
18 |
| lookahead = 1; |
| 615 |
31 |
▶ | else if (token[2] > (int8_t)0xBF) |
Branch: 1 branch
| Branch | True | False | Total |
| else if (token[2] > (int8_t)0xBF) |
20 |
11 |
31 |
|
| 616 |
20 |
| lookahead = 2; |
| 617 |
11 |
▶ | else if (token[1] > (int8_t)0xBF) |
Branch: 1 branch
| Branch | True | False | Total |
| else if (token[1] > (int8_t)0xBF) |
2 |
9 |
11 |
|
| 618 |
2 |
| lookahead = 3; |
| 619 |
|
| |
| 620 |
49 |
| data -= lookahead; |
| 621 |
49 |
| len += lookahead; |
| 622 |
|
| #if RET_ERR_IDX |
| 623 |
|
| err_pos -= lookahead; |
| 624 |
|
| #endif |
| 625 |
|
| } |
| 626 |
|
| } |
| 627 |
|
| |
| 628 |
|
| /* Check remaining bytes with naive method */ |
| 629 |
|
| #if RET_ERR_IDX |
| 630 |
|
| int err_pos2; |
| 631 |
|
| do_naive: |
| 632 |
|
| err_pos2 = cjson_utf8_naive(data, len); |
| 633 |
|
| if (err_pos2) |
| 634 |
|
| return err_pos + err_pos2 - 1; |
| 635 |
|
| return 0; |
| 636 |
|
| #else |
| 637 |
70 |
| return cjson_utf8_naive(data, len); |
| 638 |
|
| #endif |
| 639 |
|
| } |
| 640 |
|
| #endif /* CJSON_HAVE_X86_SIMD */ |
| 641 |
|
| |
| 642 |
|
| /* ------------------------------------------------------------------ ARM --- */ |
| 643 |
|
| #ifdef CJSON_HAVE_NEON |
| 644 |
|
| # include <arm_neon.h> |
| 645 |
|
| # define CJSON_SIMD_FN(opts) |
| 646 |
|
| #endif |
| 647 |
|
| |
| 648 |
|
| #ifdef CJSON_HAVE_NEON |
| 649 |
|
| |
| 650 |
|
| |
| 651 |
|
| |
| 652 |
|
| |
| 653 |
|
| /* |
| 654 |
|
| * Map high nibble of "First Byte" to legal character length minus 1 |
| 655 |
|
| * 0x00 ~ 0xBF --> 0 |
| 656 |
|
| * 0xC0 ~ 0xDF --> 1 |
| 657 |
|
| * 0xE0 ~ 0xEF --> 2 |
| 658 |
|
| * 0xF0 ~ 0xFF --> 3 |
| 659 |
|
| */ |
| 660 |
|
| static const uint8_t cjson_neon_first_len_tbl[] = { |
| 661 |
|
| 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 2, 3, |
| 662 |
|
| }; |
| 663 |
|
| |
| 664 |
|
| /* Map "First Byte" to 8-th item of range table (0xC2 ~ 0xF4) */ |
| 665 |
|
| static const uint8_t cjson_neon_first_range_tbl[] = { |
| 666 |
|
| 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, |
| 667 |
|
| }; |
| 668 |
|
| |
| 669 |
|
| /* |
| 670 |
|
| * Range table, map range index to min and max values |
| 671 |
|
| * Index 0 : 00 ~ 7F (First Byte, ascii) |
| 672 |
|
| * Index 1,2,3: 80 ~ BF (Second, Third, Fourth Byte) |
| 673 |
|
| * Index 4 : A0 ~ BF (Second Byte after E0) |
| 674 |
|
| * Index 5 : 80 ~ 9F (Second Byte after ED) |
| 675 |
|
| * Index 6 : 90 ~ BF (Second Byte after F0) |
| 676 |
|
| * Index 7 : 80 ~ 8F (Second Byte after F4) |
| 677 |
|
| * Index 8 : C2 ~ F4 (First Byte, non ascii) |
| 678 |
|
| * Index 9~15 : illegal: u >= 255 && u <= 0 |
| 679 |
|
| */ |
| 680 |
|
| static const uint8_t cjson_neon_range_min_tbl[] = { |
| 681 |
|
| 0x00, 0x80, 0x80, 0x80, 0xA0, 0x80, 0x90, 0x80, |
| 682 |
|
| 0xC2, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, |
| 683 |
|
| }; |
| 684 |
|
| static const uint8_t cjson_neon_range_max_tbl[] = { |
| 685 |
|
| 0x7F, 0xBF, 0xBF, 0xBF, 0xBF, 0x9F, 0xBF, 0x8F, |
| 686 |
|
| 0xF4, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, |
| 687 |
|
| }; |
| 688 |
|
| |
| 689 |
|
| /* |
| 690 |
|
| * This table is for fast handling four special First Bytes(E0,ED,F0,F4), after |
| 691 |
|
| * which the Second Byte are not 80~BF. It contains "range index adjustment". |
| 692 |
|
| * - The idea is to minus byte with E0, use the result(0~31) as the index to |
| 693 |
|
| * lookup the "range index adjustment". Then add the adjustment to original |
| 694 |
|
| * range index to get the correct range. |
| 695 |
|
| * - Range index adjustment |
| 696 |
|
| * +------------+---------------+------------------+----------------+ |
| 697 |
|
| * | First Byte | original range| range adjustment | adjusted range | |
| 698 |
|
| * +------------+---------------+------------------+----------------+ |
| 699 |
|
| * | E0 | 2 | 2 | 4 | |
| 700 |
|
| * +------------+---------------+------------------+----------------+ |
| 701 |
|
| * | ED | 2 | 3 | 5 | |
| 702 |
|
| * +------------+---------------+------------------+----------------+ |
| 703 |
|
| * | F0 | 3 | 3 | 6 | |
| 704 |
|
| * +------------+---------------+------------------+----------------+ |
| 705 |
|
| * | F4 | 4 | 4 | 8 | |
| 706 |
|
| * +------------+---------------+------------------+----------------+ |
| 707 |
|
| * - Below is a uint8x16x2 table, data is interleaved in NEON register. So I'm |
| 708 |
|
| * putting it vertically. 1st column is for E0~EF, 2nd column for F0~FF. |
| 709 |
|
| */ |
| 710 |
|
| static const uint8_t cjson_neon_range_adjust_tbl[] = { |
| 711 |
|
| /* index -> 0~15 16~31 <- index */ |
| 712 |
|
| /* E0 -> */ 2, 3, /* <- F0 */ |
| 713 |
|
| 0, 0, |
| 714 |
|
| 0, 0, |
| 715 |
|
| 0, 0, |
| 716 |
|
| 0, 4, /* <- F4 */ |
| 717 |
|
| 0, 0, |
| 718 |
|
| 0, 0, |
| 719 |
|
| 0, 0, |
| 720 |
|
| 0, 0, |
| 721 |
|
| 0, 0, |
| 722 |
|
| 0, 0, |
| 723 |
|
| 0, 0, |
| 724 |
|
| 0, 0, |
| 725 |
|
| /* ED -> */ 3, 0, |
| 726 |
|
| 0, 0, |
| 727 |
|
| 0, 0, |
| 728 |
|
| }; |
| 729 |
|
| |
| 730 |
|
| /* 2x ~ 4x faster than naive method */ |
| 731 |
|
| /* Return 0 on success, -1 on error */ |
| 732 |
|
| static int cjson_utf8_range_neon(const unsigned char *data, int len) |
| 733 |
|
| { |
| 734 |
|
| if (len >= 16) { |
| 735 |
|
| uint32_t token4; |
| 736 |
|
| uint8x16_t prev_input = vdupq_n_u8(0); |
| 737 |
|
| uint8x16_t prev_first_len = vdupq_n_u8(0); |
| 738 |
|
| |
| 739 |
|
| /* Cached tables */ |
| 740 |
|
| const uint8x16_t first_len_tbl = vld1q_u8(cjson_neon_first_len_tbl); |
| 741 |
|
| const uint8x16_t first_range_tbl = vld1q_u8(cjson_neon_first_range_tbl); |
| 742 |
|
| const uint8x16_t range_min_tbl = vld1q_u8(cjson_neon_range_min_tbl); |
| 743 |
|
| const uint8x16_t range_max_tbl = vld1q_u8(cjson_neon_range_max_tbl); |
| 744 |
|
| const uint8x16x2_t range_adjust_tbl = vld2q_u8(cjson_neon_range_adjust_tbl); |
| 745 |
|
| |
| 746 |
|
| /* Cached values */ |
| 747 |
|
| const uint8x16_t const_1 = vdupq_n_u8(1); |
| 748 |
|
| const uint8x16_t const_2 = vdupq_n_u8(2); |
| 749 |
|
| const uint8x16_t const_e0 = vdupq_n_u8(0xE0); |
| 750 |
|
| |
| 751 |
|
| /* We use two error registers to remove a dependency. */ |
| 752 |
|
| uint8x16_t error1 = vdupq_n_u8(0); |
| 753 |
|
| uint8x16_t error2 = vdupq_n_u8(0); |
| 754 |
|
| |
| 755 |
|
| while (len >= 16) { |
| 756 |
|
| uint8x16_t tmp1, tmp2; |
| 757 |
|
| const uint8x16_t input = vld1q_u8(data); |
| 758 |
|
| uint8x16_t shift1, pos, minv, maxv; |
| 759 |
|
| |
| 760 |
|
| /* high_nibbles = input >> 4 */ |
| 761 |
|
| const uint8x16_t high_nibbles = vshrq_n_u8(input, 4); |
| 762 |
|
| |
| 763 |
|
| /* first_len = legal character length minus 1 */ |
| 764 |
|
| /* 0 for 00~7F, 1 for C0~DF, 2 for E0~EF, 3 for F0~FF */ |
| 765 |
|
| /* first_len = first_len_tbl[high_nibbles] */ |
| 766 |
|
| const uint8x16_t first_len = |
| 767 |
|
| vqtbl1q_u8(first_len_tbl, high_nibbles); |
| 768 |
|
| |
| 769 |
|
| /* First Byte: set range index to 8 for bytes within 0xC0 ~ 0xFF */ |
| 770 |
|
| /* range = first_range_tbl[high_nibbles] */ |
| 771 |
|
| uint8x16_t range = vqtbl1q_u8(first_range_tbl, high_nibbles); |
| 772 |
|
| |
| 773 |
|
| /* Second Byte: set range index to first_len */ |
| 774 |
|
| /* 0 for 00~7F, 1 for C0~DF, 2 for E0~EF, 3 for F0~FF */ |
| 775 |
|
| /* range |= (first_len, prev_first_len) << 1 byte */ |
| 776 |
|
| range = |
| 777 |
|
| vorrq_u8(range, vextq_u8(prev_first_len, first_len, 15)); |
| 778 |
|
| |
| 779 |
|
| /* Third Byte: set range index to saturate_sub(first_len, 1) */ |
| 780 |
|
| /* 0 for 00~7F, 0 for C0~DF, 1 for E0~EF, 2 for F0~FF */ |
| 781 |
|
| /* tmp1 = (first_len, prev_first_len) << 2 bytes */ |
| 782 |
|
| tmp1 = vextq_u8(prev_first_len, first_len, 14); |
| 783 |
|
| /* tmp1 = saturate_sub(tmp1, 1) */ |
| 784 |
|
| tmp1 = vqsubq_u8(tmp1, const_1); |
| 785 |
|
| /* range |= tmp1 */ |
| 786 |
|
| range = vorrq_u8(range, tmp1); |
| 787 |
|
| |
| 788 |
|
| /* Fourth Byte: set range index to saturate_sub(first_len, 2) */ |
| 789 |
|
| /* 0 for 00~7F, 0 for C0~DF, 0 for E0~EF, 1 for F0~FF */ |
| 790 |
|
| /* tmp2 = (first_len, prev_first_len) << 3 bytes */ |
| 791 |
|
| tmp2 = vextq_u8(prev_first_len, first_len, 13); |
| 792 |
|
| /* tmp2 = saturate_sub(tmp2, 2) */ |
| 793 |
|
| tmp2 = vqsubq_u8(tmp2, const_2); |
| 794 |
|
| /* range |= tmp2 */ |
| 795 |
|
| range = vorrq_u8(range, tmp2); |
| 796 |
|
| |
| 797 |
|
| /* |
| 798 |
|
| * Now we have below range indices caluclated |
| 799 |
|
| * Correct cases: |
| 800 |
|
| * - 8 for C0~FF |
| 801 |
|
| * - 3 for 1st byte after F0~FF |
| 802 |
|
| * - 2 for 1st byte after E0~EF or 2nd byte after F0~FF |
| 803 |
|
| * - 1 for 1st byte after C0~DF or 2nd byte after E0~EF or |
| 804 |
|
| * 3rd byte after F0~FF |
| 805 |
|
| * - 0 for others |
| 806 |
|
| * Error cases: |
| 807 |
|
| * 9,10,11 if non ascii First Byte overlaps |
| 808 |
|
| * E.g., F1 80 C2 90 --> 8 3 10 2, where 10 indicates error |
| 809 |
|
| */ |
| 810 |
|
| |
| 811 |
|
| /* Adjust Second Byte range for special First Bytes(E0,ED,F0,F4) */ |
| 812 |
|
| /* See cjson_neon_range_adjust_tbl[] definition for details */ |
| 813 |
|
| /* Overlaps lead to index 9~15, which are illegal in range table */ |
| 814 |
|
| shift1 = vextq_u8(prev_input, input, 15); |
| 815 |
|
| pos = vsubq_u8(shift1, const_e0); |
| 816 |
|
| range = vaddq_u8(range, vqtbl2q_u8(range_adjust_tbl, pos)); |
| 817 |
|
| |
| 818 |
|
| /* Load min and max values per calculated range index */ |
| 819 |
|
| minv = vqtbl1q_u8(range_min_tbl, range); |
| 820 |
|
| maxv = vqtbl1q_u8(range_max_tbl, range); |
| 821 |
|
| |
| 822 |
|
| /* Check value range */ |
| 823 |
|
| error1 = vorrq_u8(error1, vcltq_u8(input, minv)); |
| 824 |
|
| error2 = vorrq_u8(error2, vcgtq_u8(input, maxv)); |
| 825 |
|
| |
| 826 |
|
| prev_input = input; |
| 827 |
|
| prev_first_len = first_len; |
| 828 |
|
| |
| 829 |
|
| data += 16; |
| 830 |
|
| len -= 16; |
| 831 |
|
| } |
| 832 |
|
| /* Merge our error counters together */ |
| 833 |
|
| error1 = vorrq_u8(error1, error2); |
| 834 |
|
| |
| 835 |
|
| /* Delay error check till loop ends */ |
| 836 |
|
| if (vmaxvq_u8(error1)) |
| 837 |
|
| return -1; |
| 838 |
|
| |
| 839 |
|
| /* Find previous token (not 80~BF) */ |
| 840 |
|
| token4; |
| 841 |
|
| vst1q_lane_u32(&token4, vreinterpretq_u32_u8(prev_input), 3); |
| 842 |
|
| |
| 843 |
|
| { |
| 844 |
|
| const int8_t *token = (const int8_t *)&token4; |
| 845 |
|
| int lookahead = 0; |
| 846 |
|
| if (token[3] > (int8_t)0xBF) |
| 847 |
|
| lookahead = 1; |
| 848 |
|
| else if (token[2] > (int8_t)0xBF) |
| 849 |
|
| lookahead = 2; |
| 850 |
|
| else if (token[1] > (int8_t)0xBF) |
| 851 |
|
| lookahead = 3; |
| 852 |
|
| data -= lookahead; |
| 853 |
|
| len += lookahead; |
| 854 |
|
| } |
| 855 |
|
| } |
| 856 |
|
| |
| 857 |
|
| /* Check remaining bytes with naive method */ |
| 858 |
|
| return cjson_utf8_naive(data, len); |
| 859 |
|
| } |
| 860 |
|
| #endif /* CJSON_HAVE_NEON */ |
| 861 |
|
| |
| 862 |
|
| /* ------------------------------------------------------------- dispatch --- */ |
| 863 |
|
| /* Which SIMD backend this CPU has; set from a constructor so the hot path |
| 864 |
|
| * never pays for a cpuid. Stays 0 (scalar) if no constructor ran. */ |
| 865 |
|
| #if defined(CJSON_HAVE_X86_SIMD) |
| 866 |
|
| static int cjson_utf8_backend = 0; /* 0 = scalar, 1 = sse4.1, 2 = avx2 */ |
| 867 |
|
| |
| 868 |
|
| __attribute__((constructor)) |
| 869 |
69 |
| static void cjson_utf8_backend_init (void) |
| 870 |
|
| { |
| 871 |
69 |
| __builtin_cpu_init (); |
| 872 |
69 |
▶ | if (__builtin_cpu_supports ("avx2")) |
Branch: 1 branch
| Branch | True | False | Total |
| if (__builtin_cpu_supports ("avx2")) |
69 |
0 |
69 |
|
| 873 |
69 |
| cjson_utf8_backend = 2; |
| 874 |
0 |
▶ | else if (__builtin_cpu_supports ("sse4.1")) |
Branch: 1 branch
| Branch | True | False | Total |
| else if (__builtin_cpu_supports ("sse4.1")) |
0 |
0 |
0 |
|
| 875 |
0 |
| cjson_utf8_backend = 1; |
| 876 |
69 |
| } |
| 877 |
|
| #endif |
| 878 |
|
| |
| 879 |
|
| /* Offset of the first byte >= 0x80, or len if the input is pure ascii. Lets a |
| 880 |
|
| * pure-ascii document -- the common case, and one that never reaches the |
| 881 |
|
| * decoder -- cost a movemask pass instead of the whole range algorithm. */ |
| 882 |
|
| #ifdef CJSON_HAVE_X86_SIMD |
| 883 |
|
| CJSON_SIMD_FN("avx2") |
| 884 |
|
| static size_t |
| 885 |
74 |
| cjson_first_high_avx2 (const unsigned char *d, size_t n) |
| 886 |
|
| { |
| 887 |
74 |
| size_t i = 0; |
| 888 |
74 |
▶ | for (; i + 32 <= n; i += 32) |
Branch: 1 branch
| Branch | True | False | Total |
| for (; i + 32 <= n; i += 32) |
53 |
21 |
74 |
|
| 889 |
|
| { |
| 890 |
106 |
| __m256i v = _mm256_loadu_si256 ((const __m256i *)(d + i)); |
| 891 |
53 |
| unsigned m = (unsigned)_mm256_movemask_epi8 (v); |
| 892 |
53 |
▶ | if (m) |
Branch: 1 branch
| Branch | True | False | Total |
| if (m) |
53 |
0 |
53 |
|
| 893 |
53 |
| return i + (size_t)__builtin_ctz (m); |
| 894 |
|
| } |
| 895 |
21 |
▶ | for (; i < n; i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (; i < n; i++) |
21 |
0 |
21 |
|
| 896 |
21 |
▶ | if (d[i] >= 0x80) |
Branch: 1 branch
| Branch | True | False | Total |
| if (d[i] >= 0x80) |
21 |
0 |
21 |
|
| 897 |
21 |
| return i; |
| 898 |
0 |
| return n; |
| 899 |
|
| } |
| 900 |
|
| |
| 901 |
|
| CJSON_SIMD_FN("sse2") |
| 902 |
|
| static size_t |
| 903 |
0 |
| cjson_first_high_sse2 (const unsigned char *d, size_t n) |
| 904 |
|
| { |
| 905 |
0 |
| size_t i = 0; |
| 906 |
0 |
▶ | for (; i + 16 <= n; i += 16) |
Branch: 1 branch
| Branch | True | False | Total |
| for (; i + 16 <= n; i += 16) |
0 |
0 |
0 |
|
| 907 |
|
| { |
| 908 |
0 |
| __m128i v = _mm_loadu_si128 ((const __m128i *)(d + i)); |
| 909 |
0 |
| unsigned m = (unsigned)_mm_movemask_epi8 (v); |
| 910 |
0 |
▶ | if (m) |
Branch: 1 branch
| Branch | True | False | Total |
| if (m) |
0 |
0 |
0 |
|
| 911 |
0 |
| return i + (size_t)__builtin_ctz (m); |
| 912 |
|
| } |
| 913 |
0 |
▶ | for (; i < n; i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (; i < n; i++) |
0 |
0 |
0 |
|
| 914 |
0 |
▶ | if (d[i] >= 0x80) |
Branch: 1 branch
| Branch | True | False | Total |
| if (d[i] >= 0x80) |
0 |
0 |
0 |
|
| 915 |
0 |
| return i; |
| 916 |
0 |
| return n; |
| 917 |
|
| } |
| 918 |
|
| #endif /* CJSON_HAVE_X86_SIMD */ |
| 919 |
|
| |
| 920 |
|
| static size_t |
| 921 |
0 |
| cjson_first_high_scalar (const unsigned char *d, size_t n) |
| 922 |
|
| { |
| 923 |
0 |
| size_t i = 0; |
| 924 |
0 |
▶ | for (; i < n; i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (; i < n; i++) |
0 |
0 |
0 |
|
| 925 |
0 |
▶ | if (d[i] >= 0x80) |
Branch: 1 branch
| Branch | True | False | Total |
| if (d[i] >= 0x80) |
0 |
0 |
0 |
|
| 926 |
0 |
| return i; |
| 927 |
0 |
| return n; |
| 928 |
|
| } |
| 929 |
|
| |
| 930 |
|
| static size_t |
| 931 |
74 |
| cjson_utf8_first_high (const unsigned char *data, size_t len) |
| 932 |
|
| { |
| 933 |
|
| #if defined(CJSON_HAVE_X86_SIMD) |
| 934 |
74 |
▶ | if (cjson_utf8_backend == 2) |
Branch: 1 branch
| Branch | True | False | Total |
| if (cjson_utf8_backend == 2) |
74 |
0 |
74 |
|
| 935 |
74 |
| return cjson_first_high_avx2 (data, len); |
| 936 |
0 |
▶ | if (cjson_utf8_backend == 1) |
Branch: 1 branch
| Branch | True | False | Total |
| if (cjson_utf8_backend == 1) |
0 |
0 |
0 |
|
| 937 |
0 |
| return cjson_first_high_sse2 (data, len); |
| 938 |
|
| #endif |
| 939 |
0 |
| return cjson_first_high_scalar (data, len); |
| 940 |
|
| } |
| 941 |
|
| |
| 942 |
|
| int |
| 943 |
74 |
| cjson_utf8_validate (const unsigned char *data, size_t len) |
| 944 |
|
| { |
| 945 |
|
| int n; |
| 946 |
|
| size_t off; |
| 947 |
|
| |
| 948 |
|
| /* >2GB of input: don't bother, let the caller use its safe path */ |
| 949 |
74 |
▶ | if (len > (size_t)INT_MAX) |
Branch: 1 branch
| Branch | True | False | Total |
| if (len > (size_t)INT_MAX) |
0 |
74 |
74 |
|
| 950 |
0 |
| return 0; |
| 951 |
|
| |
| 952 |
|
| /* skip a pure-ascii prefix: those bytes are valid by definition and each is |
| 953 |
|
| * a complete character, so the range algorithm may start right after it */ |
| 954 |
74 |
| off = cjson_utf8_first_high (data, len); |
| 955 |
74 |
▶ | if (off == len) |
Branch: 1 branch
| Branch | True | False | Total |
| if (off == len) |
0 |
74 |
74 |
|
| 956 |
0 |
| return 1; |
| 957 |
74 |
| data += off; |
| 958 |
74 |
| len -= off; |
| 959 |
|
| |
| 960 |
74 |
| n = (int)len; |
| 961 |
74 |
▶ | if (n < 16) |
Branch: 1 branch
| Branch | True | False | Total |
| if (n < 16) |
0 |
74 |
74 |
|
| 962 |
0 |
| return cjson_utf8_naive (data, n) == 0; |
| 963 |
|
| #if defined(CJSON_HAVE_X86_SIMD) |
| 964 |
74 |
▶ | if (cjson_utf8_backend == 2) |
Branch: 1 branch
| Branch | True | False | Total |
| if (cjson_utf8_backend == 2) |
74 |
0 |
74 |
|
| 965 |
74 |
| return cjson_utf8_range_avx2 (data, n) == 0; |
| 966 |
0 |
▶ | if (cjson_utf8_backend == 1) |
Branch: 1 branch
| Branch | True | False | Total |
| if (cjson_utf8_backend == 1) |
0 |
0 |
0 |
|
| 967 |
0 |
| return cjson_utf8_range_sse (data, n) == 0; |
| 968 |
0 |
| return cjson_utf8_naive (data, n) == 0; |
| 969 |
|
| #elif defined(CJSON_HAVE_NEON) |
| 970 |
|
| return cjson_utf8_range_neon (data, n) == 0; |
| 971 |
|
| #else |
| 972 |
|
| return cjson_utf8_naive (data, n) == 0; |
| 973 |
|
| #endif |
| 974 |
|
| } |