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
|
/* Copyright (C) 2012-2016 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library. If not, see
<http://www.gnu.org/licenses/>. */
#include <sysdep.h>
/* Assumptions:
*
* ARMv8-a, AArch64, unaligned accesses, min page size 4k.
*/
/* To test the page crossing code path more thoroughly, compile with
-DTEST_PAGE_CROSS - this will force all calls through the slower
entry path. This option is not intended for production use. */
/* Arguments and results. */
#define srcin x0
#define len x0
/* Locals and temporaries. */
#define src x1
#define data1 x2
#define data2 x3
#define has_nul1 x4
#define has_nul2 x5
#define tmp1 x4
#define tmp2 x5
#define tmp3 x6
#define tmp4 x7
#define zeroones x8
/* NUL detection works on the principle that (X - 1) & (~X) & 0x80
(=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and
can be done in parallel across the entire word. A faster check
(X - 1) & 0x80 is zero for non-NUL ASCII characters, but gives
false hits for characters 129..255. */
#define REP8_01 0x0101010101010101
#define REP8_7f 0x7f7f7f7f7f7f7f7f
#define REP8_80 0x8080808080808080
#ifdef TEST_PAGE_CROSS
# define MIN_PAGE_SIZE 15
#else
# define MIN_PAGE_SIZE 4096
#endif
/* Since strings are short on average, we check the first 16 bytes
of the string for a NUL character. In order to do an unaligned ldp
safely we have to do a page cross check first. If there is a NUL
byte we calculate the length from the 2 8-byte words using
conditional select to reduce branch mispredictions (it is unlikely
strlen will be repeatedly called on strings with the same length).
If the string is longer than 16 bytes, we align src so don't need
further page cross checks, and process 32 bytes per iteration
using the fast NUL check. If we encounter non-ASCII characters,
fallback to a second loop using the full NUL check.
If the page cross check fails, we read 16 bytes from an aligned
address, remove any characters before the string, and continue
in the main loop using aligned loads. Since strings crossing a
page in the first 16 bytes are rare (probability of
16/MIN_PAGE_SIZE ~= 0.4%), this case does not need to be optimized.
AArch64 systems have a minimum page size of 4k. We don't bother
checking for larger page sizes - the cost of setting up the correct
page size is just not worth the extra gain from a small reduction in
the cases taking the slow path. Note that we only care about
whether the first fetch, which may be misaligned, crosses a page
boundary. */
ENTRY_ALIGN (__strlen, 6)
DELOUSE (0)
DELOUSE (1)
and tmp1, srcin, MIN_PAGE_SIZE - 1
mov zeroones, REP8_01
cmp tmp1, MIN_PAGE_SIZE - 16
b.gt L(page_cross)
ldp data1, data2, [srcin]
#ifdef __AARCH64EB__
/* For big-endian, carry propagation (if the final byte in the
string is 0x01) means we cannot use has_nul1/2 directly.
Since we expect strings to be small and early-exit,
byte-swap the data now so has_null1/2 will be correct. */
rev data1, data1
rev data2, data2
#endif
sub tmp1, data1, zeroones
orr tmp2, data1, REP8_7f
sub tmp3, data2, zeroones
orr tmp4, data2, REP8_7f
bics has_nul1, tmp1, tmp2
bic has_nul2, tmp3, tmp4
ccmp has_nul2, 0, 0, eq
beq L(main_loop_entry)
/* Enter with C = has_nul1 == 0. */
csel has_nul1, has_nul1, has_nul2, cc
mov len, 8
rev has_nul1, has_nul1
clz tmp1, has_nul1
csel len, xzr, len, cc
add len, len, tmp1, lsr 3
ret
/* The inner loop processes 32 bytes per iteration and uses the fast
NUL check. If we encounter non-ASCII characters, use a second
loop with the accurate NUL check. */
.p2align 4
L(main_loop_entry):
bic src, srcin, 15
sub src, src, 16
L(main_loop):
ldp data1, data2, [src, 32]!
L(page_cross_entry):
sub tmp1, data1, zeroones
sub tmp3, data2, zeroones
orr tmp2, tmp1, tmp3
tst tmp2, zeroones, lsl 7
bne 1f
ldp data1, data2, [src, 16]
sub tmp1, data1, zeroones
sub tmp3, data2, zeroones
orr tmp2, tmp1, tmp3
tst tmp2, zeroones, lsl 7
beq L(main_loop)
add src, src, 16
1:
/* The fast check failed, so do the slower, accurate NUL check. */
orr tmp2, data1, REP8_7f
orr tmp4, data2, REP8_7f
bics has_nul1, tmp1, tmp2
bic has_nul2, tmp3, tmp4
ccmp has_nul2, 0, 0, eq
beq L(nonascii_loop)
/* Enter with C = has_nul1 == 0. */
L(tail):
#ifdef __AARCH64EB__
/* For big-endian, carry propagation (if the final byte in the
string is 0x01) means we cannot use has_nul1/2 directly. The
easiest way to get the correct byte is to byte-swap the data
and calculate the syndrome a second time. */
csel data1, data1, data2, cc
rev data1, data1
sub tmp1, data1, zeroones
orr tmp2, data1, REP8_7f
bic has_nul1, tmp1, tmp2
#else
csel has_nul1, has_nul1, has_nul2, cc
#endif
sub len, src, srcin
rev has_nul1, has_nul1
add tmp2, len, 8
clz tmp1, has_nul1
csel len, len, tmp2, cc
add len, len, tmp1, lsr 3
ret
L(nonascii_loop):
ldp data1, data2, [src, 16]!
sub tmp1, data1, zeroones
orr tmp2, data1, REP8_7f
sub tmp3, data2, zeroones
orr tmp4, data2, REP8_7f
bics has_nul1, tmp1, tmp2
bic has_nul2, tmp3, tmp4
ccmp has_nul2, 0, 0, eq
bne L(tail)
ldp data1, data2, [src, 16]!
sub tmp1, data1, zeroones
orr tmp2, data1, REP8_7f
sub tmp3, data2, zeroones
orr tmp4, data2, REP8_7f
bics has_nul1, tmp1, tmp2
bic has_nul2, tmp3, tmp4
ccmp has_nul2, 0, 0, eq
beq L(nonascii_loop)
b L(tail)
/* Load 16 bytes from [srcin & ~15] and force the bytes that precede
srcin to 0x7f, so we ignore any NUL bytes before the string.
Then continue in the aligned loop. */
L(page_cross):
bic src, srcin, 15
ldp data1, data2, [src]
lsl tmp1, srcin, 3
mov tmp4, -1
#ifdef __AARCH64EB__
/* Big-endian. Early bytes are at MSB. */
lsr tmp1, tmp4, tmp1 /* Shift (tmp1 & 63). */
#else
/* Little-endian. Early bytes are at LSB. */
lsl tmp1, tmp4, tmp1 /* Shift (tmp1 & 63). */
#endif
orr tmp1, tmp1, REP8_80
orn data1, data1, tmp1
orn tmp2, data2, tmp1
tst srcin, 8
csel data1, data1, tmp4, eq
csel data2, data2, tmp2, eq
b L(page_cross_entry)
END (__strlen)
weak_alias (__strlen, strlen)
libc_hidden_builtin_def (strlen)
|
