blob: 99c017be8719609bbc6e99ada83610e2d38763a1 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/* linux/arch/sparc/lib/memset.S: Sparc optimized memset, bzero and clear_user code
2 * Copyright (C) 1991,1996 Free Software Foundation
3 * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
5 *
6 * Returns 0, if ok, and number of bytes not yet set if exception
7 * occurs and we were called as clear_user.
8 */
9
10#include <asm/ptrace.h>
11
12/* Work around cpp -rob */
13#define ALLOC #alloc
14#define EXECINSTR #execinstr
15#define EX(x,y,a,b) \
1698: x,y; \
17 .section .fixup,ALLOC,EXECINSTR; \
18 .align 4; \
1999: ba 30f; \
20 a, b, %o0; \
21 .section __ex_table,ALLOC; \
22 .align 4; \
23 .word 98b, 99b; \
24 .text; \
25 .align 4
26
27#define EXT(start,end,handler) \
28 .section __ex_table,ALLOC; \
29 .align 4; \
30 .word start, 0, end, handler; \
31 .text; \
32 .align 4
33
34/* Please don't change these macros, unless you change the logic
35 * in the .fixup section below as well.
36 * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
37#define ZERO_BIG_BLOCK(base, offset, source) \
38 std source, [base + offset + 0x00]; \
39 std source, [base + offset + 0x08]; \
40 std source, [base + offset + 0x10]; \
41 std source, [base + offset + 0x18]; \
42 std source, [base + offset + 0x20]; \
43 std source, [base + offset + 0x28]; \
44 std source, [base + offset + 0x30]; \
45 std source, [base + offset + 0x38];
46
47#define ZERO_LAST_BLOCKS(base, offset, source) \
48 std source, [base - offset - 0x38]; \
49 std source, [base - offset - 0x30]; \
50 std source, [base - offset - 0x28]; \
51 std source, [base - offset - 0x20]; \
52 std source, [base - offset - 0x18]; \
53 std source, [base - offset - 0x10]; \
54 std source, [base - offset - 0x08]; \
55 std source, [base - offset - 0x00];
56
57 .text
58 .align 4
59
60 .globl __bzero_begin
61__bzero_begin:
62
David S. Miller4d14a452009-12-10 23:32:10 -080063 .globl __bzero
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 .globl memset
65 .globl __memset_start, __memset_end
66__memset_start:
Linus Torvalds1da177e2005-04-16 15:20:36 -070067memset:
68 and %o1, 0xff, %g3
69 sll %g3, 8, %g2
70 or %g3, %g2, %g3
71 sll %g3, 16, %g2
72 or %g3, %g2, %g3
73 b 1f
74 mov %o2, %o1
753:
76 cmp %o2, 3
77 be 2f
78 EX(stb %g3, [%o0], sub %o1, 0)
79
80 cmp %o2, 2
81 be 2f
82 EX(stb %g3, [%o0 + 0x01], sub %o1, 1)
83
84 EX(stb %g3, [%o0 + 0x02], sub %o1, 2)
852:
86 sub %o2, 4, %o2
87 add %o1, %o2, %o1
88 b 4f
89 sub %o0, %o2, %o0
90
91__bzero:
92 mov %g0, %g3
931:
94 cmp %o1, 7
95 bleu 7f
96 andcc %o0, 3, %o2
97
98 bne 3b
994:
100 andcc %o0, 4, %g0
101
102 be 2f
103 mov %g3, %g2
104
105 EX(st %g3, [%o0], sub %o1, 0)
106 sub %o1, 4, %o1
107 add %o0, 4, %o0
1082:
109 andcc %o1, 0xffffff80, %o3 ! Now everything is 8 aligned and o1 is len to run
110 be 9f
111 andcc %o1, 0x78, %o2
11210:
113 ZERO_BIG_BLOCK(%o0, 0x00, %g2)
114 subcc %o3, 128, %o3
115 ZERO_BIG_BLOCK(%o0, 0x40, %g2)
11611:
117 EXT(10b, 11b, 20f)
118 bne 10b
119 add %o0, 128, %o0
120
121 orcc %o2, %g0, %g0
1229:
123 be 13f
124 andcc %o1, 7, %o1
125
126 srl %o2, 1, %o3
127 set 13f, %o4
128 sub %o4, %o3, %o4
129 jmp %o4
130 add %o0, %o2, %o0
131
13212:
133 ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
134 ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
13513:
136 be 8f
137 andcc %o1, 4, %g0
138
139 be 1f
140 andcc %o1, 2, %g0
141
142 EX(st %g3, [%o0], and %o1, 7)
143 add %o0, 4, %o0
1441:
145 be 1f
146 andcc %o1, 1, %g0
147
148 EX(sth %g3, [%o0], and %o1, 3)
149 add %o0, 2, %o0
1501:
151 bne,a 8f
152 EX(stb %g3, [%o0], and %o1, 1)
1538:
154 retl
155 clr %o0
1567:
157 be 13b
158 orcc %o1, 0, %g0
159
160 be 0f
1618:
162 add %o0, 1, %o0
163 subcc %o1, 1, %o1
Alexander Shmelevf61698e2007-07-24 13:41:44 -0700164 bne 8b
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 EX(stb %g3, [%o0 - 1], add %o1, 1)
1660:
167 retl
168 clr %o0
169__memset_end:
170
171 .section .fixup,#alloc,#execinstr
172 .align 4
17320:
174 cmp %g2, 8
175 bleu 1f
176 and %o1, 0x7f, %o1
177 sub %g2, 9, %g2
178 add %o3, 64, %o3
1791:
180 sll %g2, 3, %g2
181 add %o3, %o1, %o0
182 b 30f
183 sub %o0, %g2, %o0
18421:
185 mov 8, %o0
186 and %o1, 7, %o1
187 sub %o0, %g2, %o0
188 sll %o0, 3, %o0
189 b 30f
190 add %o0, %o1, %o0
19130:
192/* %o4 is faulting address, %o5 is %pc where fault occurred */
193 save %sp, -104, %sp
194 mov %i5, %o0
195 mov %i7, %o1
196 call lookup_fault
197 mov %i4, %o2
198 ret
199 restore
200
201 .globl __bzero_end
202__bzero_end: