| line |
stmt |
bran |
cond |
sub |
pod |
time |
code |
|
1
|
|
|
|
|
|
|
/* src/simd/mds_simd_scalar.c — reference implementation. |
|
2
|
|
|
|
|
|
|
* |
|
3
|
|
|
|
|
|
|
* Defines the semantics; every other backend must produce identical |
|
4
|
|
|
|
|
|
|
* output for identical input. AVX2/SSE2/NEON backends may delegate |
|
5
|
|
|
|
|
|
|
* here for operations they have not specialised. |
|
6
|
|
|
|
|
|
|
*/ |
|
7
|
|
|
|
|
|
|
#include "mds_simd.h" |
|
8
|
|
|
|
|
|
|
#include "mds_classifier_lut.h" |
|
9
|
|
|
|
|
|
|
#include |
|
10
|
|
|
|
|
|
|
|
|
11
|
|
|
|
|
|
|
/* Bytes that "start something" in CommonMark/GFM. Kept in sync with |
|
12
|
|
|
|
|
|
|
* src/mds_block.c (block recognisers) and src/mds_inline.c (inline). |
|
13
|
|
|
|
|
|
|
* The set of structural bytes is defined by mds_classifier_truth() in |
|
14
|
|
|
|
|
|
|
* mds_classifier_lut.h — the scalar path uses that exact same lookup |
|
15
|
|
|
|
|
|
|
* so the SIMD backends can be fuzz-checked against it. */ |
|
16
|
|
|
|
|
|
|
|
|
17
|
601
|
|
|
|
|
|
static void classify_structural_scalar(const char* in, size_t len, |
|
18
|
|
|
|
|
|
|
uint64_t* out_bitmap) |
|
19
|
|
|
|
|
|
|
{ |
|
20
|
601
|
|
|
|
|
|
const uint8_t* p = (const uint8_t*)in; |
|
21
|
|
|
|
|
|
|
size_t i; |
|
22
|
515395
|
100
|
|
|
|
|
for (i = 0; i < len; i++) { |
|
23
|
514794
|
|
|
|
|
|
uint8_t b = p[i]; |
|
24
|
514794
|
|
|
|
|
|
uint8_t m = (uint8_t)(MDS_CLASSIFIER_LO[b & 0xF] & |
|
25
|
514794
|
|
|
|
|
|
MDS_CLASSIFIER_HI[b >> 4]); |
|
26
|
514794
|
100
|
|
|
|
|
if (m) out_bitmap[i >> 6] |= (uint64_t)1 << (i & 63); |
|
27
|
|
|
|
|
|
|
} |
|
28
|
601
|
|
|
|
|
|
} |
|
29
|
|
|
|
|
|
|
|
|
30
|
|
|
|
|
|
|
/* Minimal-correctness UTF-8 validator. Spec: RFC 3629. Rejects |
|
31
|
|
|
|
|
|
|
* overlong sequences and surrogate halves. */ |
|
32
|
575
|
|
|
|
|
|
static int validate_utf8_scalar(const char* in, size_t len) |
|
33
|
|
|
|
|
|
|
{ |
|
34
|
575
|
|
|
|
|
|
const uint8_t* p = (const uint8_t*)in; |
|
35
|
575
|
|
|
|
|
|
const uint8_t* e = p + len; |
|
36
|
|
|
|
|
|
|
unsigned need; |
|
37
|
|
|
|
|
|
|
uint32_t cp; |
|
38
|
|
|
|
|
|
|
uint32_t lo, hi; |
|
39
|
|
|
|
|
|
|
unsigned i; |
|
40
|
216424
|
100
|
|
|
|
|
while (p < e) { |
|
41
|
216283
|
|
|
|
|
|
uint8_t c = *p; |
|
42
|
216283
|
100
|
|
|
|
|
if (c < 0x80) { p++; continue; } |
|
43
|
691
|
100
|
|
|
|
|
if ((c & 0xE0) == 0xC0) { need = 1; cp = c & 0x1F; lo = 0x80; hi = 0x7FF; } |
|
44
|
552
|
100
|
|
|
|
|
else if ((c & 0xF0) == 0xE0) { need = 2; cp = c & 0x0F; lo = 0x800; hi = 0xFFFF; } |
|
45
|
482
|
100
|
|
|
|
|
else if ((c & 0xF8) == 0xF0) { need = 3; cp = c & 0x07; lo = 0x10000; hi = 0x10FFFF; } |
|
46
|
250
|
|
|
|
|
|
else return 0; |
|
47
|
441
|
100
|
|
|
|
|
if (p + 1 + need > e) return 0; |
|
48
|
1178
|
100
|
|
|
|
|
for (i = 0; i < need; i++) { |
|
49
|
901
|
|
|
|
|
|
uint8_t cc = p[1 + i]; |
|
50
|
901
|
100
|
|
|
|
|
if ((cc & 0xC0) != 0x80) return 0; |
|
51
|
745
|
|
|
|
|
|
cp = (cp << 6) | (cc & 0x3F); |
|
52
|
|
|
|
|
|
|
} |
|
53
|
277
|
100
|
|
|
|
|
if (cp < lo || cp > hi) return 0; |
|
|
|
100
|
|
|
|
|
|
|
54
|
261
|
100
|
|
|
|
|
if (cp >= 0xD800 && cp <= 0xDFFF) return 0; |
|
|
|
100
|
|
|
|
|
|
|
55
|
257
|
|
|
|
|
|
p += 1 + need; |
|
56
|
|
|
|
|
|
|
} |
|
57
|
141
|
|
|
|
|
|
return 1; |
|
58
|
|
|
|
|
|
|
} |
|
59
|
|
|
|
|
|
|
|
|
60
|
905
|
|
|
|
|
|
static size_t find_newlines_scalar(const char* in, size_t len, |
|
61
|
|
|
|
|
|
|
uint32_t* out_offsets, size_t cap) |
|
62
|
|
|
|
|
|
|
{ |
|
63
|
905
|
|
|
|
|
|
const char* p = in; |
|
64
|
905
|
|
|
|
|
|
const char* end = in + len; |
|
65
|
905
|
|
|
|
|
|
size_t k = 0; |
|
66
|
|
|
|
|
|
|
const char* nl; |
|
67
|
43424
|
100
|
|
|
|
|
while (p < end) { |
|
68
|
42700
|
|
|
|
|
|
nl = (const char*)memchr(p, '\n', (size_t)(end - p)); |
|
69
|
42700
|
100
|
|
|
|
|
if (!nl) break; |
|
70
|
42519
|
50
|
|
|
|
|
if (k >= cap) return (size_t)-1; /* overflow sentinel */ |
|
71
|
42519
|
|
|
|
|
|
out_offsets[k++] = (uint32_t)(nl - in); |
|
72
|
42519
|
|
|
|
|
|
p = nl + 1; |
|
73
|
|
|
|
|
|
|
} |
|
74
|
905
|
|
|
|
|
|
return k; |
|
75
|
|
|
|
|
|
|
} |
|
76
|
|
|
|
|
|
|
|
|
77
|
0
|
|
|
|
|
|
static const char* next_structural_scalar(const char* p, const char* end) |
|
78
|
|
|
|
|
|
|
{ |
|
79
|
0
|
|
|
|
|
|
const uint8_t* q = (const uint8_t*)p; |
|
80
|
0
|
|
|
|
|
|
const uint8_t* e = (const uint8_t*)end; |
|
81
|
0
|
0
|
|
|
|
|
while (q < e) { |
|
82
|
0
|
|
|
|
|
|
uint8_t b = *q; |
|
83
|
0
|
0
|
|
|
|
|
if (MDS_CLASSIFIER_LO[b & 0xF] & MDS_CLASSIFIER_HI[b >> 4]) break; |
|
84
|
0
|
|
|
|
|
|
q++; |
|
85
|
|
|
|
|
|
|
} |
|
86
|
0
|
|
|
|
|
|
return (const char*)q; |
|
87
|
|
|
|
|
|
|
} |
|
88
|
|
|
|
|
|
|
|
|
89
|
|
|
|
|
|
|
/* Bitmap-aware variant: ctzll walk over uint64_t words. O(set-bits). */ |
|
90
|
0
|
|
|
|
|
|
static const char* next_structural_bm_scalar(const char* base, size_t bm_len, |
|
91
|
|
|
|
|
|
|
const uint64_t* bm, size_t p_off) |
|
92
|
|
|
|
|
|
|
{ |
|
93
|
|
|
|
|
|
|
size_t word; |
|
94
|
|
|
|
|
|
|
size_t off; |
|
95
|
|
|
|
|
|
|
uint64_t w; |
|
96
|
|
|
|
|
|
|
size_t nwords; |
|
97
|
|
|
|
|
|
|
size_t i; |
|
98
|
0
|
0
|
|
|
|
|
if (p_off >= bm_len) return base + bm_len; |
|
99
|
0
|
|
|
|
|
|
word = p_off >> 6; |
|
100
|
0
|
|
|
|
|
|
off = p_off & 63u; |
|
101
|
|
|
|
|
|
|
/* First (possibly partial) word: mask out bits before p_off. */ |
|
102
|
0
|
|
|
|
|
|
w = bm[word] & (~(uint64_t)0 << off); |
|
103
|
0
|
0
|
|
|
|
|
if (w) { |
|
104
|
0
|
|
|
|
|
|
size_t bit = (size_t)__builtin_ctzll(w); |
|
105
|
0
|
|
|
|
|
|
size_t cand = (word << 6) + bit; |
|
106
|
0
|
|
|
|
|
|
return base + (cand < bm_len ? cand : bm_len); |
|
107
|
|
|
|
|
|
|
} |
|
108
|
|
|
|
|
|
|
/* Full words after. */ |
|
109
|
0
|
|
|
|
|
|
nwords = (bm_len + 63) >> 6; |
|
110
|
0
|
0
|
|
|
|
|
for (i = word + 1; i < nwords; i++) { |
|
111
|
0
|
|
|
|
|
|
uint64_t v = bm[i]; |
|
112
|
0
|
0
|
|
|
|
|
if (v) { |
|
113
|
0
|
|
|
|
|
|
size_t bit = (size_t)__builtin_ctzll(v); |
|
114
|
0
|
|
|
|
|
|
size_t cand = (i << 6) + bit; |
|
115
|
0
|
|
|
|
|
|
return base + (cand < bm_len ? cand : bm_len); |
|
116
|
|
|
|
|
|
|
} |
|
117
|
|
|
|
|
|
|
} |
|
118
|
0
|
|
|
|
|
|
return base + bm_len; |
|
119
|
|
|
|
|
|
|
} |
|
120
|
|
|
|
|
|
|
|
|
121
|
|
|
|
|
|
|
static const mds_simd_ops k_ops_scalar = { |
|
122
|
|
|
|
|
|
|
classify_structural_scalar, |
|
123
|
|
|
|
|
|
|
validate_utf8_scalar, |
|
124
|
|
|
|
|
|
|
find_newlines_scalar, |
|
125
|
|
|
|
|
|
|
next_structural_bm_scalar, |
|
126
|
|
|
|
|
|
|
next_structural_scalar, |
|
127
|
|
|
|
|
|
|
}; |
|
128
|
|
|
|
|
|
|
|
|
129
|
2083
|
|
|
|
|
|
const mds_simd_ops* mds_simd_ops_scalar(void) { return &k_ops_scalar; } |