1 /* SPDX-License-Identifier: GPL-2.0 */ <<
2 /* linux/arch/sparc/lib/memset.S: Sparc optimi 1 /* linux/arch/sparc/lib/memset.S: Sparc optimized memset, bzero and clear_user code
3 * Copyright (C) 1991,1996 Free Software Found 2 * Copyright (C) 1991,1996 Free Software Foundation
4 * Copyright (C) 1996,1997 Jakub Jelinek (jj@s 3 * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
5 * Copyright (C) 1996 David S. Miller (davem@c 4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
6 * 5 *
7 * Calls to memset returns initial %o0. Calls 6 * Calls to memset returns initial %o0. Calls to bzero returns 0, if ok, and
8 * number of bytes not yet set if exception oc 7 * number of bytes not yet set if exception occurs and we were called as
9 * clear_user. 8 * clear_user.
10 */ 9 */
11 10
12 #include <linux/export.h> <<
13 #include <asm/ptrace.h> 11 #include <asm/ptrace.h>
>> 12 #include <asm/export.h>
14 13
15 /* Work around cpp -rob */ 14 /* Work around cpp -rob */
16 #define ALLOC #alloc 15 #define ALLOC #alloc
17 #define EXECINSTR #execinstr 16 #define EXECINSTR #execinstr
18 #define EX(x,y,a,b) 17 #define EX(x,y,a,b) \
19 98: x,y; 18 98: x,y; \
20 .section .fixup,ALLOC,EXECINSTR; 19 .section .fixup,ALLOC,EXECINSTR; \
21 .align 4; 20 .align 4; \
22 99: retl; !! 21 99: ba 30f; \
23 a, b, %o0; 22 a, b, %o0; \
24 .section __ex_table,ALLOC; 23 .section __ex_table,ALLOC; \
25 .align 4; 24 .align 4; \
26 .word 98b, 99b; 25 .word 98b, 99b; \
27 .text; 26 .text; \
28 .align 4 27 .align 4
29 28
30 #define STORE(source, base, offset, n) !! 29 #define EXT(start,end,handler) \
31 98: std source, [base + offset + n]; <<
32 .section .fixup,ALLOC,EXECINSTR; <<
33 .align 4; <<
34 99: ba 30f; <<
35 sub %o3, n - offset, %o3; <<
36 .section __ex_table,ALLOC; 30 .section __ex_table,ALLOC; \
37 .align 4; 31 .align 4; \
38 .word 98b, 99b; !! 32 .word start, 0, end, handler; \
39 .text; 33 .text; \
40 .align 4; !! 34 .align 4
41 <<
42 #define STORE_LAST(source, base, offset, n) <<
43 EX(std source, [base - offset - n], <<
44 add %o1, offset + n); <<
45 35
46 /* Please don't change these macros, unless yo 36 /* Please don't change these macros, unless you change the logic
47 * in the .fixup section below as well. 37 * in the .fixup section below as well.
48 * Store 64 bytes at (BASE + OFFSET) using val 38 * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
49 #define ZERO_BIG_BLOCK(base, offset, source) !! 39 #define ZERO_BIG_BLOCK(base, offset, source) \
50 STORE(source, base, offset, 0x00); !! 40 std source, [base + offset + 0x00]; \
51 STORE(source, base, offset, 0x08); !! 41 std source, [base + offset + 0x08]; \
52 STORE(source, base, offset, 0x10); !! 42 std source, [base + offset + 0x10]; \
53 STORE(source, base, offset, 0x18); !! 43 std source, [base + offset + 0x18]; \
54 STORE(source, base, offset, 0x20); !! 44 std source, [base + offset + 0x20]; \
55 STORE(source, base, offset, 0x28); !! 45 std source, [base + offset + 0x28]; \
56 STORE(source, base, offset, 0x30); !! 46 std source, [base + offset + 0x30]; \
57 STORE(source, base, offset, 0x38); !! 47 std source, [base + offset + 0x38];
58 48
59 #define ZERO_LAST_BLOCKS(base, offset, source) 49 #define ZERO_LAST_BLOCKS(base, offset, source) \
60 STORE_LAST(source, base, offset, 0x38) !! 50 std source, [base - offset - 0x38]; \
61 STORE_LAST(source, base, offset, 0x30) !! 51 std source, [base - offset - 0x30]; \
62 STORE_LAST(source, base, offset, 0x28) !! 52 std source, [base - offset - 0x28]; \
63 STORE_LAST(source, base, offset, 0x20) !! 53 std source, [base - offset - 0x20]; \
64 STORE_LAST(source, base, offset, 0x18) !! 54 std source, [base - offset - 0x18]; \
65 STORE_LAST(source, base, offset, 0x10) !! 55 std source, [base - offset - 0x10]; \
66 STORE_LAST(source, base, offset, 0x08) !! 56 std source, [base - offset - 0x08]; \
67 STORE_LAST(source, base, offset, 0x00) !! 57 std source, [base - offset - 0x00];
68 58
69 .text 59 .text
70 .align 4 60 .align 4
71 61
72 .globl __bzero_begin 62 .globl __bzero_begin
73 __bzero_begin: 63 __bzero_begin:
74 64
75 .globl __bzero 65 .globl __bzero
76 .type __bzero,#function 66 .type __bzero,#function
77 .globl memset 67 .globl memset
78 EXPORT_SYMBOL(__bzero) 68 EXPORT_SYMBOL(__bzero)
79 EXPORT_SYMBOL(memset) 69 EXPORT_SYMBOL(memset)
>> 70 .globl __memset_start, __memset_end
>> 71 __memset_start:
80 memset: 72 memset:
81 mov %o0, %g1 73 mov %o0, %g1
82 mov 1, %g4 74 mov 1, %g4
83 and %o1, 0xff, %g3 75 and %o1, 0xff, %g3
84 sll %g3, 8, %g2 76 sll %g3, 8, %g2
85 or %g3, %g2, %g3 77 or %g3, %g2, %g3
86 sll %g3, 16, %g2 78 sll %g3, 16, %g2
87 or %g3, %g2, %g3 79 or %g3, %g2, %g3
88 b 1f 80 b 1f
89 mov %o2, %o1 81 mov %o2, %o1
90 3: 82 3:
91 cmp %o2, 3 83 cmp %o2, 3
92 be 2f 84 be 2f
93 EX(stb %g3, [%o0], sub %o1, 0) 85 EX(stb %g3, [%o0], sub %o1, 0)
94 86
95 cmp %o2, 2 87 cmp %o2, 2
96 be 2f 88 be 2f
97 EX(stb %g3, [%o0 + 0x01], sub %o1, 1) 89 EX(stb %g3, [%o0 + 0x01], sub %o1, 1)
98 90
99 EX(stb %g3, [%o0 + 0x02], sub %o1, 2) 91 EX(stb %g3, [%o0 + 0x02], sub %o1, 2)
100 2: 92 2:
101 sub %o2, 4, %o2 93 sub %o2, 4, %o2
102 add %o1, %o2, %o1 94 add %o1, %o2, %o1
103 b 4f 95 b 4f
104 sub %o0, %o2, %o0 96 sub %o0, %o2, %o0
105 97
106 __bzero: 98 __bzero:
107 clr %g4 99 clr %g4
108 mov %g0, %g3 100 mov %g0, %g3
109 1: 101 1:
110 cmp %o1, 7 102 cmp %o1, 7
111 bleu 7f 103 bleu 7f
112 andcc %o0, 3, %o2 104 andcc %o0, 3, %o2
113 105
114 bne 3b 106 bne 3b
115 4: 107 4:
116 andcc %o0, 4, %g0 108 andcc %o0, 4, %g0
117 109
118 be 2f 110 be 2f
119 mov %g3, %g2 111 mov %g3, %g2
120 112
121 EX(st %g3, [%o0], sub %o1, 0) 113 EX(st %g3, [%o0], sub %o1, 0)
122 sub %o1, 4, %o1 114 sub %o1, 4, %o1
123 add %o0, 4, %o0 115 add %o0, 4, %o0
124 2: 116 2:
125 andcc %o1, 0xffffff80, %o3 ! Now 117 andcc %o1, 0xffffff80, %o3 ! Now everything is 8 aligned and o1 is len to run
126 be 9f 118 be 9f
127 andcc %o1, 0x78, %o2 119 andcc %o1, 0x78, %o2
128 10: 120 10:
129 ZERO_BIG_BLOCK(%o0, 0x00, %g2) 121 ZERO_BIG_BLOCK(%o0, 0x00, %g2)
130 subcc %o3, 128, %o3 122 subcc %o3, 128, %o3
131 ZERO_BIG_BLOCK(%o0, 0x40, %g2) 123 ZERO_BIG_BLOCK(%o0, 0x40, %g2)
>> 124 11:
>> 125 EXT(10b, 11b, 20f)
132 bne 10b 126 bne 10b
133 add %o0, 128, %o0 127 add %o0, 128, %o0
134 128
135 orcc %o2, %g0, %g0 129 orcc %o2, %g0, %g0
136 9: 130 9:
137 be 13f 131 be 13f
138 andcc %o1, 7, %o1 132 andcc %o1, 7, %o1
139 133
140 srl %o2, 1, %o3 134 srl %o2, 1, %o3
141 set 13f, %o4 135 set 13f, %o4
142 sub %o4, %o3, %o4 136 sub %o4, %o3, %o4
143 jmp %o4 137 jmp %o4
144 add %o0, %o2, %o0 138 add %o0, %o2, %o0
145 139
>> 140 12:
146 ZERO_LAST_BLOCKS(%o0, 0x48, %g2) 141 ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
147 ZERO_LAST_BLOCKS(%o0, 0x08, %g2) 142 ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
148 13: 143 13:
149 be 8f 144 be 8f
150 andcc %o1, 4, %g0 145 andcc %o1, 4, %g0
151 146
152 be 1f 147 be 1f
153 andcc %o1, 2, %g0 148 andcc %o1, 2, %g0
154 149
155 EX(st %g3, [%o0], and %o1, 7) 150 EX(st %g3, [%o0], and %o1, 7)
156 add %o0, 4, %o0 151 add %o0, 4, %o0
157 1: 152 1:
158 be 1f 153 be 1f
159 andcc %o1, 1, %g0 154 andcc %o1, 1, %g0
160 155
161 EX(sth %g3, [%o0], and %o1, 3) 156 EX(sth %g3, [%o0], and %o1, 3)
162 add %o0, 2, %o0 157 add %o0, 2, %o0
163 1: 158 1:
164 bne,a 8f 159 bne,a 8f
165 EX(stb %g3, [%o0], and %o1, 1) 160 EX(stb %g3, [%o0], and %o1, 1)
166 8: 161 8:
167 b 0f 162 b 0f
168 nop 163 nop
169 7: 164 7:
170 be 13b 165 be 13b
171 orcc %o1, 0, %g0 166 orcc %o1, 0, %g0
172 167
173 be 0f 168 be 0f
174 8: 169 8:
175 add %o0, 1, %o0 170 add %o0, 1, %o0
176 subcc %o1, 1, %o1 171 subcc %o1, 1, %o1
177 bne 8b 172 bne 8b
178 EX(stb %g3, [%o0 - 1], add %o1, 1) 173 EX(stb %g3, [%o0 - 1], add %o1, 1)
179 0: 174 0:
180 andcc %g4, 1, %g0 175 andcc %g4, 1, %g0
181 be 5f 176 be 5f
182 nop 177 nop
183 retl 178 retl
184 mov %g1, %o0 179 mov %g1, %o0
185 5: 180 5:
186 retl 181 retl
187 clr %o0 182 clr %o0
>> 183 __memset_end:
188 184
189 .section .fixup,#alloc,#execinstr 185 .section .fixup,#alloc,#execinstr
190 .align 4 186 .align 4
>> 187 20:
>> 188 cmp %g2, 8
>> 189 bleu 1f
>> 190 and %o1, 0x7f, %o1
>> 191 sub %g2, 9, %g2
>> 192 add %o3, 64, %o3
>> 193 1:
>> 194 sll %g2, 3, %g2
>> 195 add %o3, %o1, %o0
>> 196 b 30f
>> 197 sub %o0, %g2, %o0
>> 198 21:
>> 199 mov 8, %o0
>> 200 and %o1, 7, %o1
>> 201 sub %o0, %g2, %o0
>> 202 sll %o0, 3, %o0
>> 203 b 30f
>> 204 add %o0, %o1, %o0
191 30: 205 30:
192 and %o1, 0x7f, %o1 !! 206 /* %o4 is faulting address, %o5 is %pc where fault occurred */
193 retl !! 207 save %sp, -104, %sp
194 add %o3, %o1, %o0 !! 208 mov %i5, %o0
>> 209 mov %i7, %o1
>> 210 call lookup_fault
>> 211 mov %i4, %o2
>> 212 ret
>> 213 restore
195 214
196 .globl __bzero_end 215 .globl __bzero_end
197 __bzero_end: 216 __bzero_end:
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.