1 /* SPDX-License-Identifier: GPL-2.0 */ 1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* $Id: strlen.S,v 1.2 2001/06/29 14:07:15 gni !! 2 /*
>> 3 * strlen.S (c) 1995 David Mosberger (davidm@cs.arizona.edu)
3 * 4 *
4 * "strlen" implementation of SuperH !! 5 * Finds length of a 0-terminated string. Optimized for the
5 * !! 6 * Alpha architecture:
6 * Copyright (C) 1999 Kaz Kojima <<
7 * 7 *
>> 8 * - memory accessed as aligned quadwords only
>> 9 * - uses bcmpge to compare 8 bytes in parallel
>> 10 * - does binary search to find 0 byte in last
>> 11 * quadword (HAKMEM needed 12 instructions to
>> 12 * do this instead of the 9 instructions that
>> 13 * binary search needs).
8 */ 14 */
>> 15 #include <asm/export.h>
>> 16 .set noreorder
>> 17 .set noat
>> 18
>> 19 .align 3
>> 20
>> 21 .globl strlen
>> 22 .ent strlen
>> 23
>> 24 strlen:
>> 25 ldq_u $1, 0($16) # load first quadword ($16 may be misaligned)
>> 26 lda $2, -1($31)
>> 27 insqh $2, $16, $2
>> 28 andnot $16, 7, $0
>> 29 or $2, $1, $1
>> 30 cmpbge $31, $1, $2 # $2 <- bitmask: bit i == 1 <==> i-th byte == 0
>> 31 bne $2, found
>> 32
>> 33 loop: ldq $1, 8($0)
>> 34 addq $0, 8, $0 # addr += 8
>> 35 nop # helps dual issue last two insns
>> 36 cmpbge $31, $1, $2
>> 37 beq $2, loop
>> 38
>> 39 found: blbs $2, done # make aligned case fast
>> 40 negq $2, $3
>> 41 and $2, $3, $2
>> 42
>> 43 and $2, 0x0f, $1
>> 44 addq $0, 4, $3
>> 45 cmoveq $1, $3, $0
>> 46
>> 47 and $2, 0x33, $1
>> 48 addq $0, 2, $3
>> 49 cmoveq $1, $3, $0
>> 50
>> 51 and $2, 0x55, $1
>> 52 addq $0, 1, $3
>> 53 cmoveq $1, $3, $0
9 54
10 /* size_t strlen (const char *s) */ !! 55 done: subq $0, $16, $0
>> 56 ret $31, ($26)
11 57
12 #include <linux/linkage.h> !! 58 .end strlen
13 ENTRY(strlen) !! 59 EXPORT_SYMBOL(strlen)
14 mov r4,r0 <<
15 and #3,r0 <<
16 tst r0,r0 <<
17 bt/s 1f <<
18 mov #0,r2 <<
19 <<
20 add #-1,r0 <<
21 shll2 r0 <<
22 shll r0 <<
23 braf r0 <<
24 nop <<
25 <<
26 mov.b @r4+,r1 <<
27 tst r1,r1 <<
28 bt 8f <<
29 add #1,r2 <<
30 <<
31 mov.b @r4+,r1 <<
32 tst r1,r1 <<
33 bt 8f <<
34 add #1,r2 <<
35 <<
36 mov.b @r4+,r1 <<
37 tst r1,r1 <<
38 bt 8f <<
39 add #1,r2 <<
40 <<
41 1: <<
42 mov #0,r3 <<
43 2: <<
44 mov.l @r4+,r1 <<
45 cmp/str r3,r1 <<
46 bf/s 2b <<
47 add #4,r2 <<
48 <<
49 add #-4,r2 <<
50 #ifndef __LITTLE_ENDIAN__ <<
51 swap.b r1,r1 <<
52 swap.w r1,r1 <<
53 swap.b r1,r1 <<
54 #endif <<
55 extu.b r1,r0 <<
56 tst r0,r0 <<
57 bt/s 8f <<
58 shlr8 r1 <<
59 add #1,r2 <<
60 extu.b r1,r0 <<
61 tst r0,r0 <<
62 bt/s 8f <<
63 shlr8 r1 <<
64 add #1,r2 <<
65 extu.b r1,r0 <<
66 tst r0,r0 <<
67 bt 8f <<
68 add #1,r2 <<
69 8: <<
70 rts <<
71 mov r2,r0 <<
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.