Loading...
1/*
2 * memscan.S: Optimized memscan for the Sparc.
3 *
4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
5 */
6
7/* In essence, this is just a fancy strlen. */
8
9#define LO_MAGIC 0x01010101
10#define HI_MAGIC 0x80808080
11
12 .text
13 .align 4
14 .globl __memscan_zero, __memscan_generic
15 .globl memscan
16__memscan_zero:
17 /* %o0 = addr, %o1 = size */
18 cmp %o1, 0
19 bne,a 1f
20 andcc %o0, 3, %g0
21
22 retl
23 nop
24
251:
26 be mzero_scan_word
27 sethi %hi(HI_MAGIC), %g2
28
29 ldsb [%o0], %g3
30mzero_still_not_word_aligned:
31 cmp %g3, 0
32 bne 1f
33 add %o0, 1, %o0
34
35 retl
36 sub %o0, 1, %o0
37
381:
39 subcc %o1, 1, %o1
40 bne,a 1f
41 andcc %o0, 3, %g0
42
43 retl
44 nop
45
461:
47 bne,a mzero_still_not_word_aligned
48 ldsb [%o0], %g3
49
50 sethi %hi(HI_MAGIC), %g2
51mzero_scan_word:
52 or %g2, %lo(HI_MAGIC), %o3
53 sethi %hi(LO_MAGIC), %g3
54 or %g3, %lo(LO_MAGIC), %o2
55mzero_next_word:
56 ld [%o0], %g2
57mzero_next_word_preloaded:
58 sub %g2, %o2, %g2
59mzero_next_word_preloaded_next:
60 andcc %g2, %o3, %g0
61 bne mzero_byte_zero
62 add %o0, 4, %o0
63
64mzero_check_out_of_fuel:
65 subcc %o1, 4, %o1
66 bg,a 1f
67 ld [%o0], %g2
68
69 retl
70 nop
71
721:
73 b mzero_next_word_preloaded_next
74 sub %g2, %o2, %g2
75
76 /* Check every byte. */
77mzero_byte_zero:
78 ldsb [%o0 - 4], %g2
79 cmp %g2, 0
80 bne mzero_byte_one
81 sub %o0, 4, %g3
82
83 retl
84 mov %g3, %o0
85
86mzero_byte_one:
87 ldsb [%o0 - 3], %g2
88 cmp %g2, 0
89 bne,a mzero_byte_two_and_three
90 ldsb [%o0 - 2], %g2
91
92 retl
93 sub %o0, 3, %o0
94
95mzero_byte_two_and_three:
96 cmp %g2, 0
97 bne,a 1f
98 ldsb [%o0 - 1], %g2
99
100 retl
101 sub %o0, 2, %o0
102
1031:
104 cmp %g2, 0
105 bne,a mzero_next_word_preloaded
106 ld [%o0], %g2
107
108 retl
109 sub %o0, 1, %o0
110
111mzero_found_it:
112 retl
113 sub %o0, 2, %o0
114
115memscan:
116__memscan_generic:
117 /* %o0 = addr, %o1 = c, %o2 = size */
118 cmp %o2, 0
119 bne,a 0f
120 ldub [%o0], %g2
121
122 b,a 2f
1231:
124 ldub [%o0], %g2
1250:
126 cmp %g2, %o1
127 be 2f
128 addcc %o2, -1, %o2
129 bne 1b
130 add %o0, 1, %o0
1312:
132 retl
133 nop
1/*
2 * memscan.S: Optimized memscan for the Sparc.
3 *
4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
5 */
6
7#include <asm/export.h>
8
9/* In essence, this is just a fancy strlen. */
10
11#define LO_MAGIC 0x01010101
12#define HI_MAGIC 0x80808080
13
14 .text
15 .align 4
16 .globl __memscan_zero, __memscan_generic
17 .globl memscan
18EXPORT_SYMBOL(__memscan_zero)
19EXPORT_SYMBOL(__memscan_generic)
20__memscan_zero:
21 /* %o0 = addr, %o1 = size */
22 cmp %o1, 0
23 bne,a 1f
24 andcc %o0, 3, %g0
25
26 retl
27 nop
28
291:
30 be mzero_scan_word
31 sethi %hi(HI_MAGIC), %g2
32
33 ldsb [%o0], %g3
34mzero_still_not_word_aligned:
35 cmp %g3, 0
36 bne 1f
37 add %o0, 1, %o0
38
39 retl
40 sub %o0, 1, %o0
41
421:
43 subcc %o1, 1, %o1
44 bne,a 1f
45 andcc %o0, 3, %g0
46
47 retl
48 nop
49
501:
51 bne,a mzero_still_not_word_aligned
52 ldsb [%o0], %g3
53
54 sethi %hi(HI_MAGIC), %g2
55mzero_scan_word:
56 or %g2, %lo(HI_MAGIC), %o3
57 sethi %hi(LO_MAGIC), %g3
58 or %g3, %lo(LO_MAGIC), %o2
59mzero_next_word:
60 ld [%o0], %g2
61mzero_next_word_preloaded:
62 sub %g2, %o2, %g2
63mzero_next_word_preloaded_next:
64 andcc %g2, %o3, %g0
65 bne mzero_byte_zero
66 add %o0, 4, %o0
67
68mzero_check_out_of_fuel:
69 subcc %o1, 4, %o1
70 bg,a 1f
71 ld [%o0], %g2
72
73 retl
74 nop
75
761:
77 b mzero_next_word_preloaded_next
78 sub %g2, %o2, %g2
79
80 /* Check every byte. */
81mzero_byte_zero:
82 ldsb [%o0 - 4], %g2
83 cmp %g2, 0
84 bne mzero_byte_one
85 sub %o0, 4, %g3
86
87 retl
88 mov %g3, %o0
89
90mzero_byte_one:
91 ldsb [%o0 - 3], %g2
92 cmp %g2, 0
93 bne,a mzero_byte_two_and_three
94 ldsb [%o0 - 2], %g2
95
96 retl
97 sub %o0, 3, %o0
98
99mzero_byte_two_and_three:
100 cmp %g2, 0
101 bne,a 1f
102 ldsb [%o0 - 1], %g2
103
104 retl
105 sub %o0, 2, %o0
106
1071:
108 cmp %g2, 0
109 bne,a mzero_next_word_preloaded
110 ld [%o0], %g2
111
112 retl
113 sub %o0, 1, %o0
114
115mzero_found_it:
116 retl
117 sub %o0, 2, %o0
118
119memscan:
120__memscan_generic:
121 /* %o0 = addr, %o1 = c, %o2 = size */
122 cmp %o2, 0
123 bne,a 0f
124 ldub [%o0], %g2
125
126 b,a 2f
1271:
128 ldub [%o0], %g2
1290:
130 cmp %g2, %o1
131 be 2f
132 addcc %o2, -1, %o2
133 bne 1b
134 add %o0, 1, %o0
1352:
136 retl
137 nop