1 | /* |
2 | * strnlen - calculate the length of a string with limit. |
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 |
12 | */ |
13 | |
14 | #include "../asmdefs.h" |
15 | |
16 | /* Arguments and results. */ |
17 | #define srcin x0 |
18 | #define len x0 |
19 | #define limit x1 |
20 | |
21 | /* Locals and temporaries. */ |
22 | #define src x2 |
23 | #define data1 x3 |
24 | #define data2 x4 |
25 | #define data2a x5 |
26 | #define has_nul1 x6 |
27 | #define has_nul2 x7 |
28 | #define tmp1 x8 |
29 | #define tmp2 x9 |
30 | #define tmp3 x10 |
31 | #define tmp4 x11 |
32 | #define zeroones x12 |
33 | #define pos x13 |
34 | #define limit_wd x14 |
35 | |
36 | #define REP8_01 0x0101010101010101 |
37 | #define REP8_7f 0x7f7f7f7f7f7f7f7f |
38 | #define REP8_80 0x8080808080808080 |
39 | |
40 | .text |
41 | .p2align 6 |
42 | L(start): |
43 | /* Pre-pad to ensure critical loop begins an icache line. */ |
44 | .rep 7 |
45 | nop |
46 | .endr |
47 | /* Put this code here to avoid wasting more space with pre-padding. */ |
48 | L(hit_limit): |
49 | mov len, limit |
50 | ret |
51 | |
52 | ENTRY_ALIGN (__strnlen_aarch64, 0) |
53 | cbz limit, L(hit_limit) |
54 | mov zeroones, #REP8_01 |
55 | bic src, srcin, #15 |
56 | ands tmp1, srcin, #15 |
57 | b.ne L(misaligned) |
58 | /* Calculate the number of full and partial words -1. */ |
59 | sub limit_wd, limit, #1 /* Limit != 0, so no underflow. */ |
60 | lsr limit_wd, limit_wd, #4 /* Convert to Qwords. */ |
61 | |
62 | /* NUL detection works on the principle that (X - 1) & (~X) & 0x80 |
63 | (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and |
64 | can be done in parallel across the entire word. */ |
65 | /* The inner loop deals with two Dwords at a time. This has a |
66 | slightly higher start-up cost, but we should win quite quickly, |
67 | especially on cores with a high number of issue slots per |
68 | cycle, as we get much better parallelism out of the operations. */ |
69 | |
70 | /* Start of critical section -- keep to one 64Byte cache line. */ |
71 | L(loop): |
72 | ldp data1, data2, [src], #16 |
73 | L(realigned): |
74 | sub tmp1, data1, zeroones |
75 | orr tmp2, data1, #REP8_7f |
76 | sub tmp3, data2, zeroones |
77 | orr tmp4, data2, #REP8_7f |
78 | bic has_nul1, tmp1, tmp2 |
79 | bic has_nul2, tmp3, tmp4 |
80 | subs limit_wd, limit_wd, #1 |
81 | orr tmp1, has_nul1, has_nul2 |
82 | ccmp tmp1, #0, #0, pl /* NZCV = 0000 */ |
83 | b.eq L(loop) |
84 | /* End of critical section -- keep to one 64Byte cache line. */ |
85 | |
86 | orr tmp1, has_nul1, has_nul2 |
87 | cbz tmp1, L(hit_limit) /* No null in final Qword. */ |
88 | |
89 | /* We know there's a null in the final Qword. The easiest thing |
90 | to do now is work out the length of the string and return |
91 | MIN (len, limit). */ |
92 | |
93 | sub len, src, srcin |
94 | cbz has_nul1, L(nul_in_data2) |
95 | #ifdef __AARCH64EB__ |
96 | mov data2, data1 |
97 | #endif |
98 | sub len, len, #8 |
99 | mov has_nul2, has_nul1 |
100 | L(nul_in_data2): |
101 | #ifdef __AARCH64EB__ |
102 | /* For big-endian, carry propagation (if the final byte in the |
103 | string is 0x01) means we cannot use has_nul directly. The |
104 | easiest way to get the correct byte is to byte-swap the data |
105 | and calculate the syndrome a second time. */ |
106 | rev data2, data2 |
107 | sub tmp1, data2, zeroones |
108 | orr tmp2, data2, #REP8_7f |
109 | bic has_nul2, tmp1, tmp2 |
110 | #endif |
111 | sub len, len, #8 |
112 | rev has_nul2, has_nul2 |
113 | clz pos, has_nul2 |
114 | add len, len, pos, lsr #3 /* Bits to bytes. */ |
115 | cmp len, limit |
116 | csel len, len, limit, ls /* Return the lower value. */ |
117 | ret |
118 | |
119 | L(misaligned): |
120 | /* Deal with a partial first word. |
121 | We're doing two things in parallel here; |
122 | 1) Calculate the number of words (but avoiding overflow if |
123 | limit is near ULONG_MAX) - to do this we need to work out |
124 | limit + tmp1 - 1 as a 65-bit value before shifting it; |
125 | 2) Load and mask the initial data words - we force the bytes |
126 | before the ones we are interested in to 0xff - this ensures |
127 | early bytes will not hit any zero detection. */ |
128 | sub limit_wd, limit, #1 |
129 | neg tmp4, tmp1 |
130 | cmp tmp1, #8 |
131 | |
132 | and tmp3, limit_wd, #15 |
133 | lsr limit_wd, limit_wd, #4 |
134 | mov tmp2, #~0 |
135 | |
136 | ldp data1, data2, [src], #16 |
137 | lsl tmp4, tmp4, #3 /* Bytes beyond alignment -> bits. */ |
138 | add tmp3, tmp3, tmp1 |
139 | |
140 | #ifdef __AARCH64EB__ |
141 | /* Big-endian. Early bytes are at MSB. */ |
142 | lsl tmp2, tmp2, tmp4 /* Shift (tmp1 & 63). */ |
143 | #else |
144 | /* Little-endian. Early bytes are at LSB. */ |
145 | lsr tmp2, tmp2, tmp4 /* Shift (tmp1 & 63). */ |
146 | #endif |
147 | add limit_wd, limit_wd, tmp3, lsr #4 |
148 | |
149 | orr data1, data1, tmp2 |
150 | orr data2a, data2, tmp2 |
151 | |
152 | csinv data1, data1, xzr, le |
153 | csel data2, data2, data2a, le |
154 | b L(realigned) |
155 | |
156 | END (__strnlen_aarch64) |
157 | |