blob: e9ba1920d818e2c714c1beb8ed75010224563b2d [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/* strlen.S: Sparc64 optimized strlen code
2 * Hand optimized from GNU libc's strlen
3 * Copyright (C) 1991,1996 Free Software Foundation
4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
5 * Copyright (C) 1996, 1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
6 */
7
8#define LO_MAGIC 0x01010101
9#define HI_MAGIC 0x80808080
10
11 .align 32
12 .globl strlen
13 .type strlen,#function
14strlen:
15 mov %o0, %o1
16 andcc %o0, 3, %g0
17 be,pt %icc, 9f
18 sethi %hi(HI_MAGIC), %o4
19 ldub [%o0], %o5
20 brz,pn %o5, 11f
21 add %o0, 1, %o0
22 andcc %o0, 3, %g0
23 be,pn %icc, 4f
24 or %o4, %lo(HI_MAGIC), %o3
25 ldub [%o0], %o5
26 brz,pn %o5, 12f
27 add %o0, 1, %o0
28 andcc %o0, 3, %g0
29 be,pt %icc, 5f
30 sethi %hi(LO_MAGIC), %o4
31 ldub [%o0], %o5
32 brz,pn %o5, 13f
33 add %o0, 1, %o0
34 ba,pt %icc, 8f
35 or %o4, %lo(LO_MAGIC), %o2
369:
37 or %o4, %lo(HI_MAGIC), %o3
384:
39 sethi %hi(LO_MAGIC), %o4
405:
41 or %o4, %lo(LO_MAGIC), %o2
428:
43 ld [%o0], %o5
442:
45 sub %o5, %o2, %o4
46 andcc %o4, %o3, %g0
47 be,pt %icc, 8b
48 add %o0, 4, %o0
49
50 /* Check every byte. */
51 srl %o5, 24, %g7
52 andcc %g7, 0xff, %g0
53 be,pn %icc, 1f
54 add %o0, -4, %o4
55 srl %o5, 16, %g7
56 andcc %g7, 0xff, %g0
57 be,pn %icc, 1f
58 add %o4, 1, %o4
59 srl %o5, 8, %g7
60 andcc %g7, 0xff, %g0
61 be,pn %icc, 1f
62 add %o4, 1, %o4
63 andcc %o5, 0xff, %g0
64 bne,a,pt %icc, 2b
65 ld [%o0], %o5
66 add %o4, 1, %o4
671:
68 retl
69 sub %o4, %o1, %o0
7011:
71 retl
72 mov 0, %o0
7312:
74 retl
75 mov 1, %o0
7613:
77 retl
78 mov 2, %o0
79
80 .size strlen, .-strlen