1 /* SPDX-License-Identifier: GPL-2.0 */ 1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* linux/arch/sparc/lib/memset.S: Sparc optimi 2 /* linux/arch/sparc/lib/memset.S: Sparc optimized memset, bzero and clear_user code
3 * Copyright (C) 1991,1996 Free Software Found 3 * Copyright (C) 1991,1996 Free Software Foundation
4 * Copyright (C) 1996,1997 Jakub Jelinek (jj@s 4 * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
5 * Copyright (C) 1996 David S. Miller (davem@c 5 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
6 * 6 *
7 * Calls to memset returns initial %o0. Calls 7 * Calls to memset returns initial %o0. Calls to bzero returns 0, if ok, and
8 * number of bytes not yet set if exception oc 8 * number of bytes not yet set if exception occurs and we were called as
9 * clear_user. 9 * clear_user.
10 */ 10 */
11 11
12 #include <linux/export.h> <<
13 #include <asm/ptrace.h> 12 #include <asm/ptrace.h>
>> 13 #include <asm/export.h>
14 14
15 /* Work around cpp -rob */ 15 /* Work around cpp -rob */
16 #define ALLOC #alloc 16 #define ALLOC #alloc
17 #define EXECINSTR #execinstr 17 #define EXECINSTR #execinstr
18 #define EX(x,y,a,b) 18 #define EX(x,y,a,b) \
19 98: x,y; 19 98: x,y; \
20 .section .fixup,ALLOC,EXECINSTR; 20 .section .fixup,ALLOC,EXECINSTR; \
21 .align 4; 21 .align 4; \
22 99: retl; !! 22 99: ba 30f; \
23 a, b, %o0; 23 a, b, %o0; \
24 .section __ex_table,ALLOC; 24 .section __ex_table,ALLOC; \
25 .align 4; 25 .align 4; \
26 .word 98b, 99b; 26 .word 98b, 99b; \
27 .text; 27 .text; \
28 .align 4 28 .align 4
29 29
30 #define STORE(source, base, offset, n) !! 30 #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; 31 .section __ex_table,ALLOC; \
37 .align 4; 32 .align 4; \
38 .word 98b, 99b; !! 33 .word start, 0, end, handler; \
39 .text; 34 .text; \
40 .align 4; !! 35 .align 4
41 <<
42 #define STORE_LAST(source, base, offset, n) <<
43 EX(std source, [base - offset - n], <<
44 add %o1, offset + n); <<
45 36
46 /* Please don't change these macros, unless yo 37 /* Please don't change these macros, unless you change the logic
47 * in the .fixup section below as well. 38 * in the .fixup section below as well.
48 * Store 64 bytes at (BASE + OFFSET) using val 39 * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
49 #define ZERO_BIG_BLOCK(base, offset, source) !! 40 #define ZERO_BIG_BLOCK(base, offset, source) \
50 STORE(source, base, offset, 0x00); !! 41 std source, [base + offset + 0x00]; \
51 STORE(source, base, offset, 0x08); !! 42 std source, [base + offset + 0x08]; \
52 STORE(source, base, offset, 0x10); !! 43 std source, [base + offset + 0x10]; \
53 STORE(source, base, offset, 0x18); !! 44 std source, [base + offset + 0x18]; \
54 STORE(source, base, offset, 0x20); !! 45 std source, [base + offset + 0x20]; \
55 STORE(source, base, offset, 0x28); !! 46 std source, [base + offset + 0x28]; \
56 STORE(source, base, offset, 0x30); !! 47 std source, [base + offset + 0x30]; \
57 STORE(source, base, offset, 0x38); !! 48 std source, [base + offset + 0x38];
58 49
59 #define ZERO_LAST_BLOCKS(base, offset, source) 50 #define ZERO_LAST_BLOCKS(base, offset, source) \
60 STORE_LAST(source, base, offset, 0x38) !! 51 std source, [base - offset - 0x38]; \
61 STORE_LAST(source, base, offset, 0x30) !! 52 std source, [base - offset - 0x30]; \
62 STORE_LAST(source, base, offset, 0x28) !! 53 std source, [base - offset - 0x28]; \
63 STORE_LAST(source, base, offset, 0x20) !! 54 std source, [base - offset - 0x20]; \
64 STORE_LAST(source, base, offset, 0x18) !! 55 std source, [base - offset - 0x18]; \
65 STORE_LAST(source, base, offset, 0x10) !! 56 std source, [base - offset - 0x10]; \
66 STORE_LAST(source, base, offset, 0x08) !! 57 std source, [base - offset - 0x08]; \
67 STORE_LAST(source, base, offset, 0x00) !! 58 std source, [base - offset - 0x00];
68 59
69 .text 60 .text
70 .align 4 61 .align 4
71 62
72 .globl __bzero_begin 63 .globl __bzero_begin
73 __bzero_begin: 64 __bzero_begin:
74 65
75 .globl __bzero 66 .globl __bzero
76 .type __bzero,#function 67 .type __bzero,#function
77 .globl memset 68 .globl memset
78 EXPORT_SYMBOL(__bzero) 69 EXPORT_SYMBOL(__bzero)
79 EXPORT_SYMBOL(memset) 70 EXPORT_SYMBOL(memset)
>> 71 .globl __memset_start, __memset_end
>> 72 __memset_start:
80 memset: 73 memset:
81 mov %o0, %g1 74 mov %o0, %g1
82 mov 1, %g4 75 mov 1, %g4
83 and %o1, 0xff, %g3 76 and %o1, 0xff, %g3
84 sll %g3, 8, %g2 77 sll %g3, 8, %g2
85 or %g3, %g2, %g3 78 or %g3, %g2, %g3
86 sll %g3, 16, %g2 79 sll %g3, 16, %g2
87 or %g3, %g2, %g3 80 or %g3, %g2, %g3
88 b 1f 81 b 1f
89 mov %o2, %o1 82 mov %o2, %o1
90 3: 83 3:
91 cmp %o2, 3 84 cmp %o2, 3
92 be 2f 85 be 2f
93 EX(stb %g3, [%o0], sub %o1, 0) 86 EX(stb %g3, [%o0], sub %o1, 0)
94 87
95 cmp %o2, 2 88 cmp %o2, 2
96 be 2f 89 be 2f
97 EX(stb %g3, [%o0 + 0x01], sub %o1, 1) 90 EX(stb %g3, [%o0 + 0x01], sub %o1, 1)
98 91
99 EX(stb %g3, [%o0 + 0x02], sub %o1, 2) 92 EX(stb %g3, [%o0 + 0x02], sub %o1, 2)
100 2: 93 2:
101 sub %o2, 4, %o2 94 sub %o2, 4, %o2
102 add %o1, %o2, %o1 95 add %o1, %o2, %o1
103 b 4f 96 b 4f
104 sub %o0, %o2, %o0 97 sub %o0, %o2, %o0
105 98
106 __bzero: 99 __bzero:
107 clr %g4 100 clr %g4
108 mov %g0, %g3 101 mov %g0, %g3
109 1: 102 1:
110 cmp %o1, 7 103 cmp %o1, 7
111 bleu 7f 104 bleu 7f
112 andcc %o0, 3, %o2 105 andcc %o0, 3, %o2
113 106
114 bne 3b 107 bne 3b
115 4: 108 4:
116 andcc %o0, 4, %g0 109 andcc %o0, 4, %g0
117 110
118 be 2f 111 be 2f
119 mov %g3, %g2 112 mov %g3, %g2
120 113
121 EX(st %g3, [%o0], sub %o1, 0) 114 EX(st %g3, [%o0], sub %o1, 0)
122 sub %o1, 4, %o1 115 sub %o1, 4, %o1
123 add %o0, 4, %o0 116 add %o0, 4, %o0
124 2: 117 2:
125 andcc %o1, 0xffffff80, %o3 ! Now 118 andcc %o1, 0xffffff80, %o3 ! Now everything is 8 aligned and o1 is len to run
126 be 9f 119 be 9f
127 andcc %o1, 0x78, %o2 120 andcc %o1, 0x78, %o2
128 10: 121 10:
129 ZERO_BIG_BLOCK(%o0, 0x00, %g2) 122 ZERO_BIG_BLOCK(%o0, 0x00, %g2)
130 subcc %o3, 128, %o3 123 subcc %o3, 128, %o3
131 ZERO_BIG_BLOCK(%o0, 0x40, %g2) 124 ZERO_BIG_BLOCK(%o0, 0x40, %g2)
>> 125 11:
>> 126 EXT(10b, 11b, 20f)
132 bne 10b 127 bne 10b
133 add %o0, 128, %o0 128 add %o0, 128, %o0
134 129
135 orcc %o2, %g0, %g0 130 orcc %o2, %g0, %g0
136 9: 131 9:
137 be 13f 132 be 13f
138 andcc %o1, 7, %o1 133 andcc %o1, 7, %o1
139 134
140 srl %o2, 1, %o3 135 srl %o2, 1, %o3
141 set 13f, %o4 136 set 13f, %o4
142 sub %o4, %o3, %o4 137 sub %o4, %o3, %o4
143 jmp %o4 138 jmp %o4
144 add %o0, %o2, %o0 139 add %o0, %o2, %o0
145 140
>> 141 12:
146 ZERO_LAST_BLOCKS(%o0, 0x48, %g2) 142 ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
147 ZERO_LAST_BLOCKS(%o0, 0x08, %g2) 143 ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
148 13: 144 13:
149 be 8f 145 be 8f
150 andcc %o1, 4, %g0 146 andcc %o1, 4, %g0
151 147
152 be 1f 148 be 1f
153 andcc %o1, 2, %g0 149 andcc %o1, 2, %g0
154 150
155 EX(st %g3, [%o0], and %o1, 7) 151 EX(st %g3, [%o0], and %o1, 7)
156 add %o0, 4, %o0 152 add %o0, 4, %o0
157 1: 153 1:
158 be 1f 154 be 1f
159 andcc %o1, 1, %g0 155 andcc %o1, 1, %g0
160 156
161 EX(sth %g3, [%o0], and %o1, 3) 157 EX(sth %g3, [%o0], and %o1, 3)
162 add %o0, 2, %o0 158 add %o0, 2, %o0
163 1: 159 1:
164 bne,a 8f 160 bne,a 8f
165 EX(stb %g3, [%o0], and %o1, 1) 161 EX(stb %g3, [%o0], and %o1, 1)
166 8: 162 8:
167 b 0f 163 b 0f
168 nop 164 nop
169 7: 165 7:
170 be 13b 166 be 13b
171 orcc %o1, 0, %g0 167 orcc %o1, 0, %g0
172 168
173 be 0f 169 be 0f
174 8: 170 8:
175 add %o0, 1, %o0 171 add %o0, 1, %o0
176 subcc %o1, 1, %o1 172 subcc %o1, 1, %o1
177 bne 8b 173 bne 8b
178 EX(stb %g3, [%o0 - 1], add %o1, 1) 174 EX(stb %g3, [%o0 - 1], add %o1, 1)
179 0: 175 0:
180 andcc %g4, 1, %g0 176 andcc %g4, 1, %g0
181 be 5f 177 be 5f
182 nop 178 nop
183 retl 179 retl
184 mov %g1, %o0 180 mov %g1, %o0
185 5: 181 5:
186 retl 182 retl
187 clr %o0 183 clr %o0
>> 184 __memset_end:
188 185
189 .section .fixup,#alloc,#execinstr 186 .section .fixup,#alloc,#execinstr
190 .align 4 187 .align 4
>> 188 20:
>> 189 cmp %g2, 8
>> 190 bleu 1f
>> 191 and %o1, 0x7f, %o1
>> 192 sub %g2, 9, %g2
>> 193 add %o3, 64, %o3
>> 194 1:
>> 195 sll %g2, 3, %g2
>> 196 add %o3, %o1, %o0
>> 197 b 30f
>> 198 sub %o0, %g2, %o0
>> 199 21:
>> 200 mov 8, %o0
>> 201 and %o1, 7, %o1
>> 202 sub %o0, %g2, %o0
>> 203 sll %o0, 3, %o0
>> 204 b 30f
>> 205 add %o0, %o1, %o0
191 30: 206 30:
192 and %o1, 0x7f, %o1 !! 207 /* %o4 is faulting address, %o5 is %pc where fault occurred */
193 retl !! 208 save %sp, -104, %sp
194 add %o3, %o1, %o0 !! 209 mov %i5, %o0
>> 210 mov %i7, %o1
>> 211 call lookup_fault
>> 212 mov %i4, %o2
>> 213 ret
>> 214 restore
195 215
196 .globl __bzero_end 216 .globl __bzero_end
197 __bzero_end: 217 __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.