1 | /* |
2 | * strlen - calculate the length of a string |
3 | * |
4 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
5 | * See https://llvm.org/LICENSE.txt for license information. |
6 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
7 | */ |
8 | |
9 | /* Assumptions: |
10 | * |
11 | * ARMv8-a, AArch64, unaligned accesses, min page size 4k. |
12 | */ |
13 | |
14 | #include "../asmdefs.h" |
15 | |
16 | /* To test the page crossing code path more thoroughly, compile with |
17 | -DTEST_PAGE_CROSS - this will force all calls through the slower |
18 | entry path. This option is not intended for production use. */ |
19 | |
20 | /* Arguments and results. */ |
21 | #define srcin x0 |
22 | #define len x0 |
23 | |
24 | /* Locals and temporaries. */ |
25 | #define src x1 |
26 | #define data1 x2 |
27 | #define data2 x3 |
28 | #define has_nul1 x4 |
29 | #define has_nul2 x5 |
30 | #define tmp1 x4 |
31 | #define tmp2 x5 |
32 | #define tmp3 x6 |
33 | #define tmp4 x7 |
34 | #define zeroones x8 |
35 | |
36 | /* NUL detection works on the principle that (X - 1) & (~X) & 0x80 |
37 | (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and |
38 | can be done in parallel across the entire word. A faster check |
39 | (X - 1) & 0x80 is zero for non-NUL ASCII characters, but gives |
40 | false hits for characters 129..255. */ |
41 | |
42 | #define REP8_01 0x0101010101010101 |
43 | #define REP8_7f 0x7f7f7f7f7f7f7f7f |
44 | #define REP8_80 0x8080808080808080 |
45 | |
46 | #ifdef TEST_PAGE_CROSS |
47 | # define MIN_PAGE_SIZE 15 |
48 | #else |
49 | # define MIN_PAGE_SIZE 4096 |
50 | #endif |
51 | |
52 | /* Since strings are short on average, we check the first 16 bytes |
53 | of the string for a NUL character. In order to do an unaligned ldp |
54 | safely we have to do a page cross check first. If there is a NUL |
55 | byte we calculate the length from the 2 8-byte words using |
56 | conditional select to reduce branch mispredictions (it is unlikely |
57 | __strlen_aarch64 will be repeatedly called on strings with the same length). |
58 | |
59 | If the string is longer than 16 bytes, we align src so don't need |
60 | further page cross checks, and process 32 bytes per iteration |
61 | using the fast NUL check. If we encounter non-ASCII characters, |
62 | fallback to a second loop using the full NUL check. |
63 | |
64 | If the page cross check fails, we read 16 bytes from an aligned |
65 | address, remove any characters before the string, and continue |
66 | in the main loop using aligned loads. Since strings crossing a |
67 | page in the first 16 bytes are rare (probability of |
68 | 16/MIN_PAGE_SIZE ~= 0.4%), this case does not need to be optimized. |
69 | |
70 | AArch64 systems have a minimum page size of 4k. We don't bother |
71 | checking for larger page sizes - the cost of setting up the correct |
72 | page size is just not worth the extra gain from a small reduction in |
73 | the cases taking the slow path. Note that we only care about |
74 | whether the first fetch, which may be misaligned, crosses a page |
75 | boundary. */ |
76 | |
77 | ENTRY (__strlen_aarch64) |
78 | and tmp1, srcin, MIN_PAGE_SIZE - 1 |
79 | mov zeroones, REP8_01 |
80 | cmp tmp1, MIN_PAGE_SIZE - 16 |
81 | b.gt L(page_cross) |
82 | ldp data1, data2, [srcin] |
83 | #ifdef __AARCH64EB__ |
84 | /* For big-endian, carry propagation (if the final byte in the |
85 | string is 0x01) means we cannot use has_nul1/2 directly. |
86 | Since we expect strings to be small and early-exit, |
87 | byte-swap the data now so has_null1/2 will be correct. */ |
88 | rev data1, data1 |
89 | rev data2, data2 |
90 | #endif |
91 | sub tmp1, data1, zeroones |
92 | orr tmp2, data1, REP8_7f |
93 | sub tmp3, data2, zeroones |
94 | orr tmp4, data2, REP8_7f |
95 | bics has_nul1, tmp1, tmp2 |
96 | bic has_nul2, tmp3, tmp4 |
97 | ccmp has_nul2, 0, 0, eq |
98 | beq L(main_loop_entry) |
99 | |
100 | /* Enter with C = has_nul1 == 0. */ |
101 | csel has_nul1, has_nul1, has_nul2, cc |
102 | mov len, 8 |
103 | rev has_nul1, has_nul1 |
104 | clz tmp1, has_nul1 |
105 | csel len, xzr, len, cc |
106 | add len, len, tmp1, lsr 3 |
107 | ret |
108 | |
109 | /* The inner loop processes 32 bytes per iteration and uses the fast |
110 | NUL check. If we encounter non-ASCII characters, use a second |
111 | loop with the accurate NUL check. */ |
112 | .p2align 4 |
113 | L(main_loop_entry): |
114 | bic src, srcin, 15 |
115 | sub src, src, 16 |
116 | L(main_loop): |
117 | ldp data1, data2, [src, 32]! |
118 | L(page_cross_entry): |
119 | sub tmp1, data1, zeroones |
120 | sub tmp3, data2, zeroones |
121 | orr tmp2, tmp1, tmp3 |
122 | tst tmp2, zeroones, lsl 7 |
123 | bne 1f |
124 | ldp data1, data2, [src, 16] |
125 | sub tmp1, data1, zeroones |
126 | sub tmp3, data2, zeroones |
127 | orr tmp2, tmp1, tmp3 |
128 | tst tmp2, zeroones, lsl 7 |
129 | beq L(main_loop) |
130 | add src, src, 16 |
131 | 1: |
132 | /* The fast check failed, so do the slower, accurate NUL check. */ |
133 | orr tmp2, data1, REP8_7f |
134 | orr tmp4, data2, REP8_7f |
135 | bics has_nul1, tmp1, tmp2 |
136 | bic has_nul2, tmp3, tmp4 |
137 | ccmp has_nul2, 0, 0, eq |
138 | beq L(nonascii_loop) |
139 | |
140 | /* Enter with C = has_nul1 == 0. */ |
141 | L(tail): |
142 | #ifdef __AARCH64EB__ |
143 | /* For big-endian, carry propagation (if the final byte in the |
144 | string is 0x01) means we cannot use has_nul1/2 directly. The |
145 | easiest way to get the correct byte is to byte-swap the data |
146 | and calculate the syndrome a second time. */ |
147 | csel data1, data1, data2, cc |
148 | rev data1, data1 |
149 | sub tmp1, data1, zeroones |
150 | orr tmp2, data1, REP8_7f |
151 | bic has_nul1, tmp1, tmp2 |
152 | #else |
153 | csel has_nul1, has_nul1, has_nul2, cc |
154 | #endif |
155 | sub len, src, srcin |
156 | rev has_nul1, has_nul1 |
157 | add tmp2, len, 8 |
158 | clz tmp1, has_nul1 |
159 | csel len, len, tmp2, cc |
160 | add len, len, tmp1, lsr 3 |
161 | ret |
162 | |
163 | L(nonascii_loop): |
164 | ldp data1, data2, [src, 16]! |
165 | sub tmp1, data1, zeroones |
166 | orr tmp2, data1, REP8_7f |
167 | sub tmp3, data2, zeroones |
168 | orr tmp4, data2, REP8_7f |
169 | bics has_nul1, tmp1, tmp2 |
170 | bic has_nul2, tmp3, tmp4 |
171 | ccmp has_nul2, 0, 0, eq |
172 | bne L(tail) |
173 | ldp data1, data2, [src, 16]! |
174 | sub tmp1, data1, zeroones |
175 | orr tmp2, data1, REP8_7f |
176 | sub tmp3, data2, zeroones |
177 | orr tmp4, data2, REP8_7f |
178 | bics has_nul1, tmp1, tmp2 |
179 | bic has_nul2, tmp3, tmp4 |
180 | ccmp has_nul2, 0, 0, eq |
181 | beq L(nonascii_loop) |
182 | b L(tail) |
183 | |
184 | /* Load 16 bytes from [srcin & ~15] and force the bytes that precede |
185 | srcin to 0x7f, so we ignore any NUL bytes before the string. |
186 | Then continue in the aligned loop. */ |
187 | L(page_cross): |
188 | bic src, srcin, 15 |
189 | ldp data1, data2, [src] |
190 | lsl tmp1, srcin, 3 |
191 | mov tmp4, -1 |
192 | #ifdef __AARCH64EB__ |
193 | /* Big-endian. Early bytes are at MSB. */ |
194 | lsr tmp1, tmp4, tmp1 /* Shift (tmp1 & 63). */ |
195 | #else |
196 | /* Little-endian. Early bytes are at LSB. */ |
197 | lsl tmp1, tmp4, tmp1 /* Shift (tmp1 & 63). */ |
198 | #endif |
199 | orr tmp1, tmp1, REP8_80 |
200 | orn data1, data1, tmp1 |
201 | orn tmp2, data2, tmp1 |
202 | tst srcin, 8 |
203 | csel data1, data1, tmp4, eq |
204 | csel data2, data2, tmp2, eq |
205 | b L(page_cross_entry) |
206 | |
207 | END (__strlen_aarch64) |
208 | |