-
Notifications
You must be signed in to change notification settings - Fork 4
/
fromutf8-sse.cc
296 lines (235 loc) · 10.7 KB
/
fromutf8-sse.cc
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
/****************************************************************************
*
* Copyright (C) 2012 Olivier Goffart <[email protected]>
* http://woboq.com
*
* This is an experiment to process UTF-8 using SSE4 intrinscis.
* Read: http://woboq.com/blog/utf-8-processing-using-simd.html
*
* This file may be used under the terms of the GNU Lesser General Public
* License version 2.1 as published by the Free Software Foundation
* http://www.gnu.org/licenses/old-licenses/lgpl-2.1.html.
*
* For any question, please contact [email protected]
*
****************************************************************************/
#include <cstddef>
#include <algorithm>
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef unsigned int uint;
#ifdef __SSE4_1__
#include <smmintrin.h>
#endif
#ifdef __SSE4_2__
#include <nmmintrin.h>
#endif
void fromUtf8_sse(const char *&src, int &len, ushort * &dst) {
#if defined(__SSE4_1__)
const char *end = src + len;
while(src + 16 < end) {
__m128i chunk = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src));
#if 1 //ASCII optim
int asciiMask = _mm_movemask_epi8(chunk);
if (!asciiMask) {
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst), _mm_unpacklo_epi8(chunk, _mm_set1_epi8(0)));
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst+8) , _mm_unpackhi_epi8(chunk, _mm_set1_epi8(0)));
dst+=16;
src+=16;
continue;
}
#endif
__m128i chunk_signed = _mm_add_epi8(chunk, _mm_set1_epi8(0x80));
__m128i cond2 = _mm_cmplt_epi8( _mm_set1_epi8(0xc2-1 -0x80), chunk_signed);
__m128i state = _mm_set1_epi8(0x0 | 0x80);
state = _mm_blendv_epi8(state , _mm_set1_epi8(0x2 | 0xc0), cond2);
__m128i cond3 = _mm_cmplt_epi8( _mm_set1_epi8(0xe0-1 -0x80), chunk_signed);
// Possible improvement: create a separate processing when there are
// only 2b ytes sequences
//if (!_mm_movemask_epi8(cond3)) { /*process 2 max*/ }
state = _mm_blendv_epi8(state , _mm_set1_epi8(0x3 | 0xe0), cond3);
__m128i mask3 = _mm_slli_si128(cond3, 1);
__m128i cond4 = _mm_cmplt_epi8(_mm_set1_epi8(0xf0-1 -0x80), chunk_signed);
// 4 bytes sequences are not vectorize. Fall back to the scalar processing
if (_mm_movemask_epi8(cond4)) { break; }
__m128i count = _mm_and_si128(state, _mm_set1_epi8(0x7));
__m128i count_sub1 = _mm_subs_epu8(count, _mm_set1_epi8(0x1));
__m128i counts = _mm_add_epi8(count, _mm_slli_si128(count_sub1, 1));
__m128i shifts = count_sub1;
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 1));
counts = _mm_add_epi8(counts, _mm_slli_si128(_mm_subs_epu8(counts, _mm_set1_epi8(0x2)), 2));
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 2));
if (asciiMask ^ _mm_movemask_epi8(_mm_cmpgt_epi8(counts, _mm_set1_epi8(0))))
break; // error
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 4));
if (_mm_movemask_epi8(_mm_cmpgt_epi8(_mm_sub_epi8(_mm_slli_si128(counts, 1), counts), _mm_set1_epi8(1))))
break; //error
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 8));
__m128i mask = _mm_and_si128( state, _mm_set1_epi8(0xf8));
shifts = _mm_and_si128 (shifts , _mm_cmplt_epi8(counts, _mm_set1_epi8(2))); // <=1
chunk = _mm_andnot_si128(mask , chunk); // from now on, we only have useful bits
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 1),
_mm_srli_si128(_mm_slli_epi16(shifts, 7) , 1));
__m128i chunk_right = _mm_slli_si128(chunk, 1);
__m128i chunk_low = _mm_blendv_epi8(chunk,
_mm_or_si128(chunk, _mm_and_si128( _mm_slli_epi16(chunk_right, 6), _mm_set1_epi8(0xc0))) ,
_mm_cmpeq_epi8(counts, _mm_set1_epi8(1)) );
__m128i chunk_high = _mm_and_si128(chunk , _mm_cmpeq_epi8(counts, _mm_set1_epi8(2)) );
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 2),
_mm_srli_si128(_mm_slli_epi16(shifts, 6) , 2));
chunk_high = _mm_srli_epi32(chunk_high, 2);
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 4),
_mm_srli_si128(_mm_slli_epi16(shifts, 5) , 4));
chunk_high = _mm_or_si128(chunk_high,
_mm_and_si128(_mm_and_si128(_mm_slli_epi32(chunk_right, 4), _mm_set1_epi8(0xf0)),
mask3));
int c = _mm_extract_epi16(counts, 7);
int source_advance = !(c & 0x0200) ? 16 : !(c & 0x02) ? 15 : 14;
__m128i high_bits = _mm_and_si128(chunk_high, _mm_set1_epi8(0xf8));
if (!_mm_testz_si128(mask3,
_mm_or_si128(_mm_cmpeq_epi8(high_bits,_mm_set1_epi8(0x00)) ,
_mm_cmpeq_epi8(high_bits,_mm_set1_epi8(0xd8))) )) break;
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 8),
_mm_srli_si128(_mm_slli_epi16(shifts, 4) , 8));
chunk_high = _mm_slli_si128(chunk_high, 1);
__m128i shuf = _mm_add_epi8(shifts, _mm_set_epi8(15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0));
chunk_low = _mm_shuffle_epi8(chunk_low, shuf);
chunk_high = _mm_shuffle_epi8(chunk_high, shuf);
__m128i utf16_low = _mm_unpacklo_epi8(chunk_low, chunk_high);
__m128i utf16_high = _mm_unpackhi_epi8(chunk_low, chunk_high);
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst), utf16_low);
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst+8) , utf16_high);
int s = _mm_extract_epi32(shifts, 3);
int dst_advance = source_advance - (0xff & (s >> 8*(3 - 16 + source_advance)));
#if defined(__SSE4_2__)
const int check_mode = 5 /*_SIDD_UWORD_OPS | _SIDD_CMP_RANGES*/;
if (_mm_cmpestrc( _mm_cvtsi64_si128(0xfdeffdd0fffffffe), 4, utf16_high, 8, check_mode) |
_mm_cmpestrc( _mm_cvtsi64_si128(0xfdeffdd0fffffffe), 4, utf16_low, 8, check_mode)) {
break; }
#else
if (
!_mm_testz_si128(_mm_cmpeq_epi8(_mm_set1_epi8(0xfd), chunk_high),
_mm_and_si128(_mm_cmplt_epi8(_mm_set1_epi8(0xd0), chunk_low),
_mm_cmpgt_epi8(_mm_set1_epi8(0xef), chunk_low)))
||
!_mm_testz_si128(_mm_cmpeq_epi8(_mm_set1_epi8(0xff), chunk_high),
_mm_or_si128(_mm_cmpeq_epi8(_mm_set1_epi8(0xfe), chunk_low),
_mm_cmpeq_epi8(_mm_set1_epi8(0xff), chunk_low)))
) break;
#endif
dst += dst_advance;
src += source_advance;
}
len = end - src;
//The rest will be handled sequentially.
// Possible improvement: go back to the vectorized processing after the error or the 4 byte sequence
#endif
}
// same signature as match iconv
size_t fromUtf8(const char **inbuf, size_t *inbytesleft, char **outbuf, size_t *outbytesleft)
{
ushort *&qch = *reinterpret_cast<ushort **>(outbuf);
const char *&chars = *const_cast<const char**>(inbuf);
int len = std::min<int>(*inbytesleft , *outbytesleft/2);
//First process the bytes using SSE
fromUtf8_sse(chars, len, qch);
//Then handle the remaining bytes using scalar algorithm.
// Basically extracted from from QUtf8::convertToUnicode in qutfcodec.c
// required QChar API
class QChar {
public:
enum SpecialCharacter {
ReplacementCharacter = 0xfffd,
ObjectReplacementCharacter = 0xfffc,
LastValidCodePoint = 0x10ffff
};
static inline bool isNonCharacter(uint ucs4)
{ return ucs4 >= 0xfdd0 && (ucs4 <= 0xfdef || (ucs4 & 0xfffe) == 0xfffe); }
static inline bool isSurrogate(uint ucs4)
{ return (ucs4 - 0xd800u < 2048u); }
static inline bool requiresSurrogates(uint ucs4)
{ return (ucs4 >= 0x10000); }
static inline ushort highSurrogate(uint ucs4)
{ return ushort((ucs4>>10) + 0xd7c0); }
static inline ushort lowSurrogate(uint ucs4)
{ return ushort(ucs4%0x400 + 0xdc00); }
};
bool headerdone = false;
ushort replacement = QChar::ReplacementCharacter;
int need = 0;
int error = -1;
uint uc = 0;
uint min_uc = 0;
uchar ch;
int invalid = 0;
ushort *start =qch;
int i;
for (i = 0; i < len - need; ++i) {
ch = chars[i];
if (need) {
if ((ch&0xc0) == 0x80) {
uc = (uc << 6) | (ch & 0x3f);
--need;
if (!need) {
// utf-8 bom composes into 0xfeff code point
bool nonCharacter;
if (!headerdone && uc == 0xfeff) {
// don't do anything, just skip the BOM
} else if (!(nonCharacter = QChar::isNonCharacter(uc)) && QChar::requiresSurrogates(uc) && uc <= QChar::LastValidCodePoint) {
// surrogate pair
*qch++ = QChar::highSurrogate(uc);
*qch++ = QChar::lowSurrogate(uc);
} else if ((uc < min_uc) || QChar::isSurrogate(uc) || nonCharacter || uc > QChar::LastValidCodePoint) {
// error: overlong sequence, UTF16 surrogate or non-character
*qch++ = replacement;
++invalid;
} else {
*qch++ = ((uc & 0xff) << 8) | ((uc & 0xff00) >> 8);
}
headerdone = true;
}
} else {
// error
i = error;
*qch++ = replacement;
++invalid;
need = 0;
headerdone = true;
}
} else {
if (ch < 128) {
*qch++ = ushort(ch) << 8;
headerdone = true;
} else if ((ch & 0xe0) == 0xc0) {
uc = ch & 0x1f;
need = 1;
error = i;
min_uc = 0x80;
headerdone = true;
} else if ((ch & 0xf0) == 0xe0) {
uc = ch & 0x0f;
need = 2;
error = i;
min_uc = 0x800;
} else if ((ch&0xf8) == 0xf0) {
uc = ch & 0x07;
need = 3;
error = i;
min_uc = 0x10000;
headerdone = true;
} else {
// error
*qch++ = replacement;
++invalid;
headerdone = true;
}
}
}
if (need)
i--;
*inbytesleft = len - i;
chars += i;
size_t r = (qch - start);
*outbytesleft -= r;
return need ? -1 : 0;
}