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
use core::mem;
use crate::char::{KernelFn, Wide};
mod packed;
use self::packed::{simd_eq, NonZeroPacked, Pack, Packed};
#[inline(always)]
pub fn wmemchr<T: Wide>(needle: T, haystack: &[T]) -> Option<usize> {
T::wmemchr_fallback(needle, haystack)
}
pub(crate) struct Kernel;
impl<T: Pack> KernelFn<T> for Kernel {
fn kernel(needle: T, haystack: &[T]) -> Option<usize> {
const VECTOR_SIZE: usize = mem::size_of::<Packed>();
const VECTOR_ALIGN: usize = VECTOR_SIZE - 1;
const LOOP_SIZE: usize = 4 * VECTOR_SIZE;
let loop_elements = 4 * T::LANES;
let start = haystack.as_ptr();
let mut ptr = start;
unsafe {
let end = start.add(haystack.len());
debug_assert!(start <= end);
if haystack.len() < T::LANES {
while ptr < end {
if *ptr == needle {
return Some(ptr.offset_from(start) as usize);
}
ptr = ptr.add(1);
}
return None;
}
debug_assert!(end.offset_from(start) as usize >= T::LANES);
let v_needle = needle.broadcast();
if let Some(pos) = forward_search_unaligned(start, end, ptr, v_needle) {
return Some(pos);
}
ptr = {
let align_offset = VECTOR_SIZE - ((start as usize) & VECTOR_ALIGN);
(start as *const u8).add(align_offset) as *const T
};
debug_assert!(start < ptr);
debug_assert!(ptr.offset_from(start) as usize <= T::LANES);
if let Some(loop_end) = (end as usize).checked_sub(LOOP_SIZE) {
while (ptr as usize) <= loop_end {
debug_assert_eq!((ptr as usize) % VECTOR_SIZE, 0);
let p = ptr as *const Packed;
let a = *p;
let b = *p.add(1);
let c = *p.add(2);
let d = *p.add(3);
let eq_a = simd_eq::<T>(a, v_needle);
let eq_b = simd_eq::<T>(b, v_needle);
let eq_c = simd_eq::<T>(c, v_needle);
let eq_d = simd_eq::<T>(d, v_needle);
let or_ab = eq_a | eq_b;
let or_cd = eq_c | eq_d;
let or = or_ab | or_cd;
if or != 0 {
let mut offset = ptr.offset_from(start) as usize;
if let Some(mask) = NonZeroPacked::new(eq_a) {
return Some(offset + forward_pos::<T>(mask));
}
offset += T::LANES;
if let Some(mask) = NonZeroPacked::new(eq_b) {
return Some(offset + forward_pos::<T>(mask));
}
offset += T::LANES;
if let Some(mask) = NonZeroPacked::new(eq_c) {
return Some(offset + forward_pos::<T>(mask));
}
offset += T::LANES;
debug_assert_ne!(eq_d, 0);
let mask = NonZeroPacked::new_unchecked(eq_d);
return Some(offset + forward_pos::<T>(mask));
}
ptr = ptr.add(loop_elements);
}
}
if let Some(loop_end) = (end as usize).checked_sub(VECTOR_SIZE) {
while (ptr as usize) <= loop_end {
debug_assert_eq!((ptr as usize) % VECTOR_SIZE, 0);
let chunk = *(ptr as *const Packed);
let eq = simd_eq::<T>(chunk, v_needle);
if let Some(mask) = NonZeroPacked::new(eq) {
let offset = ptr.offset_from(start) as usize;
return Some(offset + forward_pos::<T>(mask));
}
ptr = ptr.add(T::LANES);
}
}
if ptr < end {
let remaining = end.offset_from(ptr) as usize;
debug_assert!(remaining < T::LANES);
ptr = ptr.sub(T::LANES - remaining);
debug_assert_eq!(end.offset_from(ptr) as usize, T::LANES);
return forward_search_unaligned(start, end, ptr, v_needle);
}
None
}
}
}
#[inline]
unsafe fn forward_search_unaligned<T: Pack>(
start: *const T,
end: *const T,
ptr: *const T,
v_needle: Packed,
) -> Option<usize> {
debug_assert!(start <= ptr);
debug_assert!(end.offset_from(ptr) as usize >= T::LANES);
let chunk = (ptr as *const Packed).read_unaligned();
let eq = simd_eq::<T>(chunk, v_needle);
if let Some(mask) = NonZeroPacked::new(eq) {
let offset = ptr.offset_from(start) as usize;
Some(offset + forward_pos::<T>(mask))
} else {
None
}
}
#[inline(always)]
fn forward_pos<T: Pack>(mask: NonZeroPacked) -> usize {
(bsf!(mask) as usize) / T::BITS
}