| 1 |
#ifdef __aarch64__ |
| 2 |
|
| 3 |
#include <stdio.h> |
| 4 |
#include <stdint.h> |
| 5 |
#include <arm_neon.h> |
| 6 |
|
| 7 |
int utf8_naive(const unsigned char *data, int len); |
| 8 |
|
| 9 |
#if 0 |
| 10 |
static void print128(const char *s, const uint8x16_t v128) |
| 11 |
{ |
| 12 |
unsigned char v8[16]; |
| 13 |
vst1q_u8(v8, v128); |
| 14 |
|
| 15 |
if (s) |
| 16 |
printf("%s:\t", s); |
| 17 |
for (int i = 0; i < 16; ++i) |
| 18 |
printf("%02x ", v8[i]); |
| 19 |
printf("\n"); |
| 20 |
} |
| 21 |
#endif |
| 22 |
|
| 23 |
/* |
| 24 |
* Map high nibble of "First Byte" to legal character length minus 1 |
| 25 |
* 0x00 ~ 0xBF --> 0 |
| 26 |
* 0xC0 ~ 0xDF --> 1 |
| 27 |
* 0xE0 ~ 0xEF --> 2 |
| 28 |
* 0xF0 ~ 0xFF --> 3 |
| 29 |
*/ |
| 30 |
static const uint8_t _first_len_tbl[] = { |
| 31 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 2, 3, |
| 32 |
}; |
| 33 |
|
| 34 |
/* Map "First Byte" to 8-th item of range table (0xC2 ~ 0xF4) */ |
| 35 |
static const uint8_t _first_range_tbl[] = { |
| 36 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, |
| 37 |
}; |
| 38 |
|
| 39 |
/* |
| 40 |
* Range table, map range index to min and max values |
| 41 |
* Index 0 : 00 ~ 7F (First Byte, ascii) |
| 42 |
* Index 1,2,3: 80 ~ BF (Second, Third, Fourth Byte) |
| 43 |
* Index 4 : A0 ~ BF (Second Byte after E0) |
| 44 |
* Index 5 : 80 ~ 9F (Second Byte after ED) |
| 45 |
* Index 6 : 90 ~ BF (Second Byte after F0) |
| 46 |
* Index 7 : 80 ~ 8F (Second Byte after F4) |
| 47 |
* Index 8 : C2 ~ F4 (First Byte, non ascii) |
| 48 |
* Index 9~15 : illegal: u >= 255 && u <= 0 |
| 49 |
*/ |
| 50 |
static const uint8_t _range_min_tbl[] = { |
| 51 |
0x00, 0x80, 0x80, 0x80, 0xA0, 0x80, 0x90, 0x80, |
| 52 |
0xC2, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, |
| 53 |
}; |
| 54 |
static const uint8_t _range_max_tbl[] = { |
| 55 |
0x7F, 0xBF, 0xBF, 0xBF, 0xBF, 0x9F, 0xBF, 0x8F, |
| 56 |
0xF4, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, |
| 57 |
}; |
| 58 |
|
| 59 |
/* |
| 60 |
* This table is for fast handling four special First Bytes(E0,ED,F0,F4), after |
| 61 |
* which the Second Byte are not 80~BF. It contains "range index adjustment". |
| 62 |
* - The idea is to minus byte with E0, use the result(0~31) as the index to |
| 63 |
* lookup the "range index adjustment". Then add the adjustment to original |
| 64 |
* range index to get the correct range. |
| 65 |
* - Range index adjustment |
| 66 |
* +------------+---------------+------------------+----------------+ |
| 67 |
* | First Byte | original range| range adjustment | adjusted range | |
| 68 |
* +------------+---------------+------------------+----------------+ |
| 69 |
* | E0 | 2 | 2 | 4 | |
| 70 |
* +------------+---------------+------------------+----------------+ |
| 71 |
* | ED | 2 | 3 | 5 | |
| 72 |
* +------------+---------------+------------------+----------------+ |
| 73 |
* | F0 | 3 | 3 | 6 | |
| 74 |
* +------------+---------------+------------------+----------------+ |
| 75 |
* | F4 | 4 | 4 | 8 | |
| 76 |
* +------------+---------------+------------------+----------------+ |
| 77 |
* - Below is a uint8x16x2 table, data is interleaved in NEON register. So I'm |
| 78 |
* putting it vertically. 1st column is for E0~EF, 2nd column for F0~FF. |
| 79 |
*/ |
| 80 |
static const uint8_t _range_adjust_tbl[] = { |
| 81 |
/* index -> 0~15 16~31 <- index */ |
| 82 |
/* E0 -> */ 2, 3, /* <- F0 */ |
| 83 |
0, 0, |
| 84 |
0, 0, |
| 85 |
0, 0, |
| 86 |
0, 4, /* <- F4 */ |
| 87 |
0, 0, |
| 88 |
0, 0, |
| 89 |
0, 0, |
| 90 |
0, 0, |
| 91 |
0, 0, |
| 92 |
0, 0, |
| 93 |
0, 0, |
| 94 |
0, 0, |
| 95 |
/* ED -> */ 3, 0, |
| 96 |
0, 0, |
| 97 |
0, 0, |
| 98 |
}; |
| 99 |
|
| 100 |
/* 2x ~ 4x faster than naive method */ |
| 101 |
/* Return 0 on success, -1 on error */ |
| 102 |
int utf8_range(const unsigned char *data, int len) |
| 103 |
{ |
| 104 |
if (len >= 16) { |
| 105 |
uint8x16_t prev_input = vdupq_n_u8(0); |
| 106 |
uint8x16_t prev_first_len = vdupq_n_u8(0); |
| 107 |
|
| 108 |
/* Cached tables */ |
| 109 |
const uint8x16_t first_len_tbl = vld1q_u8(_first_len_tbl); |
| 110 |
const uint8x16_t first_range_tbl = vld1q_u8(_first_range_tbl); |
| 111 |
const uint8x16_t range_min_tbl = vld1q_u8(_range_min_tbl); |
| 112 |
const uint8x16_t range_max_tbl = vld1q_u8(_range_max_tbl); |
| 113 |
const uint8x16x2_t range_adjust_tbl = vld2q_u8(_range_adjust_tbl); |
| 114 |
|
| 115 |
/* Cached values */ |
| 116 |
const uint8x16_t const_1 = vdupq_n_u8(1); |
| 117 |
const uint8x16_t const_2 = vdupq_n_u8(2); |
| 118 |
const uint8x16_t const_e0 = vdupq_n_u8(0xE0); |
| 119 |
|
| 120 |
/* We use two error registers to remove a dependency. */ |
| 121 |
uint8x16_t error1 = vdupq_n_u8(0); |
| 122 |
uint8x16_t error2 = vdupq_n_u8(0); |
| 123 |
|
| 124 |
while (len >= 16) { |
| 125 |
const uint8x16_t input = vld1q_u8(data); |
| 126 |
|
| 127 |
/* high_nibbles = input >> 4 */ |
| 128 |
const uint8x16_t high_nibbles = vshrq_n_u8(input, 4); |
| 129 |
|
| 130 |
/* first_len = legal character length minus 1 */ |
| 131 |
/* 0 for 00~7F, 1 for C0~DF, 2 for E0~EF, 3 for F0~FF */ |
| 132 |
/* first_len = first_len_tbl[high_nibbles] */ |
| 133 |
const uint8x16_t first_len = |
| 134 |
vqtbl1q_u8(first_len_tbl, high_nibbles); |
| 135 |
|
| 136 |
/* First Byte: set range index to 8 for bytes within 0xC0 ~ 0xFF */ |
| 137 |
/* range = first_range_tbl[high_nibbles] */ |
| 138 |
uint8x16_t range = vqtbl1q_u8(first_range_tbl, high_nibbles); |
| 139 |
|
| 140 |
/* Second Byte: set range index to first_len */ |
| 141 |
/* 0 for 00~7F, 1 for C0~DF, 2 for E0~EF, 3 for F0~FF */ |
| 142 |
/* range |= (first_len, prev_first_len) << 1 byte */ |
| 143 |
range = |
| 144 |
vorrq_u8(range, vextq_u8(prev_first_len, first_len, 15)); |
| 145 |
|
| 146 |
/* Third Byte: set range index to saturate_sub(first_len, 1) */ |
| 147 |
/* 0 for 00~7F, 0 for C0~DF, 1 for E0~EF, 2 for F0~FF */ |
| 148 |
uint8x16_t tmp1, tmp2; |
| 149 |
/* tmp1 = (first_len, prev_first_len) << 2 bytes */ |
| 150 |
tmp1 = vextq_u8(prev_first_len, first_len, 14); |
| 151 |
/* tmp1 = saturate_sub(tmp1, 1) */ |
| 152 |
tmp1 = vqsubq_u8(tmp1, const_1); |
| 153 |
/* range |= tmp1 */ |
| 154 |
range = vorrq_u8(range, tmp1); |
| 155 |
|
| 156 |
/* Fourth Byte: set range index to saturate_sub(first_len, 2) */ |
| 157 |
/* 0 for 00~7F, 0 for C0~DF, 0 for E0~EF, 1 for F0~FF */ |
| 158 |
/* tmp2 = (first_len, prev_first_len) << 3 bytes */ |
| 159 |
tmp2 = vextq_u8(prev_first_len, first_len, 13); |
| 160 |
/* tmp2 = saturate_sub(tmp2, 2) */ |
| 161 |
tmp2 = vqsubq_u8(tmp2, const_2); |
| 162 |
/* range |= tmp2 */ |
| 163 |
range = vorrq_u8(range, tmp2); |
| 164 |
|
| 165 |
/* |
| 166 |
* Now we have below range indices caluclated |
| 167 |
* Correct cases: |
| 168 |
* - 8 for C0~FF |
| 169 |
* - 3 for 1st byte after F0~FF |
| 170 |
* - 2 for 1st byte after E0~EF or 2nd byte after F0~FF |
| 171 |
* - 1 for 1st byte after C0~DF or 2nd byte after E0~EF or |
| 172 |
* 3rd byte after F0~FF |
| 173 |
* - 0 for others |
| 174 |
* Error cases: |
| 175 |
* 9,10,11 if non ascii First Byte overlaps |
| 176 |
* E.g., F1 80 C2 90 --> 8 3 10 2, where 10 indicates error |
| 177 |
*/ |
| 178 |
|
| 179 |
/* Adjust Second Byte range for special First Bytes(E0,ED,F0,F4) */ |
| 180 |
/* See _range_adjust_tbl[] definition for details */ |
| 181 |
/* Overlaps lead to index 9~15, which are illegal in range table */ |
| 182 |
uint8x16_t shift1 = vextq_u8(prev_input, input, 15); |
| 183 |
uint8x16_t pos = vsubq_u8(shift1, const_e0); |
| 184 |
range = vaddq_u8(range, vqtbl2q_u8(range_adjust_tbl, pos)); |
| 185 |
|
| 186 |
/* Load min and max values per calculated range index */ |
| 187 |
uint8x16_t minv = vqtbl1q_u8(range_min_tbl, range); |
| 188 |
uint8x16_t maxv = vqtbl1q_u8(range_max_tbl, range); |
| 189 |
|
| 190 |
/* Check value range */ |
| 191 |
error1 = vorrq_u8(error1, vcltq_u8(input, minv)); |
| 192 |
error2 = vorrq_u8(error2, vcgtq_u8(input, maxv)); |
| 193 |
|
| 194 |
prev_input = input; |
| 195 |
prev_first_len = first_len; |
| 196 |
|
| 197 |
data += 16; |
| 198 |
len -= 16; |
| 199 |
} |
| 200 |
/* Merge our error counters together */ |
| 201 |
error1 = vorrq_u8(error1, error2); |
| 202 |
|
| 203 |
/* Delay error check till loop ends */ |
| 204 |
if (vmaxvq_u8(error1)) |
| 205 |
return -1; |
| 206 |
|
| 207 |
/* Find previous token (not 80~BF) */ |
| 208 |
uint32_t token4; |
| 209 |
vst1q_lane_u32(&token4, vreinterpretq_u32_u8(prev_input), 3); |
| 210 |
|
| 211 |
const int8_t *token = (const int8_t *)&token4; |
| 212 |
int lookahead = 0; |
| 213 |
if (token[3] > (int8_t)0xBF) |
| 214 |
lookahead = 1; |
| 215 |
else if (token[2] > (int8_t)0xBF) |
| 216 |
lookahead = 2; |
| 217 |
else if (token[1] > (int8_t)0xBF) |
| 218 |
lookahead = 3; |
| 219 |
|
| 220 |
data -= lookahead; |
| 221 |
len += lookahead; |
| 222 |
} |
| 223 |
|
| 224 |
/* Check remaining bytes with naive method */ |
| 225 |
return utf8_naive(data, len); |
| 226 |
} |
| 227 |
|
| 228 |
#endif |