1 /* SPDX-License-Identifier: GPL-2.0 */ !! 1 /* linux/arch/sparc/lib/memset.S: Sparc optimized memset, bzero and clear_user code
2 /* $Id: memset.S,v 1.1 2000/04/14 16:49:01 mjd !! 2 * Copyright (C) 1991,1996 Free Software Foundation
3 * !! 3 * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
4 * "memset" implementation of SuperH !! 4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
5 * <<
6 * Copyright (C) 1999 Niibe Yutaka <<
7 * 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 */ 8 */
9 9
10 /* !! 10 #include <asm/cprefix.h>
11 * void *memset(void *s, int c, siz !! 11 #include <asm/ptrace.h>
12 */ <<
13 12
14 #include <linux/linkage.h> !! 13 /* Work around cpp -rob */
>> 14 #define ALLOC #alloc
>> 15 #define EXECINSTR #execinstr
>> 16 #define EX(x,y,a,b) \
>> 17 98: x,y; \
>> 18 .section .fixup,ALLOC,EXECINSTR; \
>> 19 .align 4; \
>> 20 99: ba 30f; \
>> 21 a, b, %o0; \
>> 22 .section __ex_table,ALLOC; \
>> 23 .align 4; \
>> 24 .word 98b, 99b; \
>> 25 .text; \
>> 26 .align 4
15 27
16 ENTRY(memset) !! 28 #define EXT(start,end,handler) \
17 tst r6,r6 !! 29 .section __ex_table,ALLOC; \
18 bt/s 5f ! if n=0, do n !! 30 .align 4; \
19 add r6,r4 !! 31 .word start, 0, end, handler; \
20 mov #12,r0 !! 32 .text; \
21 cmp/gt r6,r0 !! 33 .align 4
22 bt/s 4f ! if it's too !! 34
23 mov r4,r0 !! 35 /* Please don't change these macros, unless you change the logic
24 and #3,r0 !! 36 * in the .fixup section below as well.
25 cmp/eq #0,r0 !! 37 * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
26 bt/s 2f ! It's aligned !! 38 #define ZERO_BIG_BLOCK(base, offset, source) \
27 sub r0,r6 !! 39 std source, [base + offset + 0x00]; \
28 1: !! 40 std source, [base + offset + 0x08]; \
29 dt r0 !! 41 std source, [base + offset + 0x10]; \
30 bf/s 1b !! 42 std source, [base + offset + 0x18]; \
31 mov.b r5,@-r4 !! 43 std source, [base + offset + 0x20]; \
32 2: ! make VVVV !! 44 std source, [base + offset + 0x28]; \
33 extu.b r5,r5 !! 45 std source, [base + offset + 0x30]; \
34 swap.b r5,r0 ! V0 !! 46 std source, [base + offset + 0x38];
35 or r0,r5 ! VV !! 47
36 swap.w r5,r0 ! VV00 !! 48 #define ZERO_LAST_BLOCKS(base, offset, source) \
37 or r0,r5 ! VVVV !! 49 std source, [base - offset - 0x38]; \
38 ! !! 50 std source, [base - offset - 0x30]; \
39 mov r6,r0 !! 51 std source, [base - offset - 0x28]; \
40 shlr2 r0 !! 52 std source, [base - offset - 0x20]; \
41 shlr r0 ! r0 = r6 >> 3 !! 53 std source, [base - offset - 0x18]; \
>> 54 std source, [base - offset - 0x10]; \
>> 55 std source, [base - offset - 0x08]; \
>> 56 std source, [base - offset - 0x00];
>> 57
>> 58 .text
>> 59 .align 4
>> 60
>> 61 .globl __bzero_begin
>> 62 __bzero_begin:
>> 63
>> 64 .globl C_LABEL(__bzero), C_LABEL(__memset),
>> 65 .globl C_LABEL(memset)
>> 66 .globl C_LABEL(__memset_start), C_LABEL(__memset_end)
>> 67 C_LABEL(__memset_start):
>> 68 C_LABEL(__memset):
>> 69 C_LABEL(memset):
>> 70 and %o1, 0xff, %g3
>> 71 sll %g3, 8, %g2
>> 72 or %g3, %g2, %g3
>> 73 sll %g3, 16, %g2
>> 74 or %g3, %g2, %g3
>> 75 b 1f
>> 76 mov %o2, %o1
42 3: 77 3:
43 dt r0 !! 78 cmp %o2, 3
44 mov.l r5,@-r4 ! set 8-byte a !! 79 be 2f
45 bf/s 3b !! 80 EX(stb %g3, [%o0], sub %o1, 0)
46 mov.l r5,@-r4 !! 81
47 ! !! 82 cmp %o2, 2
48 mov #7,r0 !! 83 be 2f
49 and r0,r6 !! 84 EX(stb %g3, [%o0 + 0x01], sub %o1, 1)
50 tst r6,r6 !! 85
51 bt 5f !! 86 EX(stb %g3, [%o0 + 0x02], sub %o1, 2)
52 ! fill bytes !! 87 2:
>> 88 sub %o2, 4, %o2
>> 89 add %o1, %o2, %o1
>> 90 b 4f
>> 91 sub %o0, %o2, %o0
>> 92
>> 93 C_LABEL(__bzero):
>> 94 mov %g0, %g3
>> 95 1:
>> 96 cmp %o1, 7
>> 97 bleu 7f
>> 98 andcc %o0, 3, %o2
>> 99
>> 100 bne 3b
53 4: 101 4:
54 dt r6 !! 102 andcc %o0, 4, %g0
55 bf/s 4b !! 103
56 mov.b r5,@-r4 !! 104 be 2f
57 5: !! 105 mov %g3, %g2
58 rts !! 106
59 mov r4,r0 !! 107 EX(st %g3, [%o0], sub %o1, 0)
>> 108 sub %o1, 4, %o1
>> 109 add %o0, 4, %o0
>> 110 2:
>> 111 andcc %o1, 0xffffff80, %o3 ! Now everything is 8 aligned and o1 is len to run
>> 112 be 9f
>> 113 andcc %o1, 0x78, %o2
>> 114 10:
>> 115 ZERO_BIG_BLOCK(%o0, 0x00, %g2)
>> 116 subcc %o3, 128, %o3
>> 117 ZERO_BIG_BLOCK(%o0, 0x40, %g2)
>> 118 11:
>> 119 EXT(10b, 11b, 20f)
>> 120 bne 10b
>> 121 add %o0, 128, %o0
>> 122
>> 123 orcc %o2, %g0, %g0
>> 124 9:
>> 125 be 13f
>> 126 andcc %o1, 7, %o1
>> 127
>> 128 srl %o2, 1, %o3
>> 129 set 13f, %o4
>> 130 sub %o4, %o3, %o4
>> 131 jmp %o4
>> 132 add %o0, %o2, %o0
>> 133
>> 134 12:
>> 135 ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
>> 136 ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
>> 137 13:
>> 138 be 8f
>> 139 andcc %o1, 4, %g0
>> 140
>> 141 be 1f
>> 142 andcc %o1, 2, %g0
>> 143
>> 144 EX(st %g3, [%o0], and %o1, 7)
>> 145 add %o0, 4, %o0
>> 146 1:
>> 147 be 1f
>> 148 andcc %o1, 1, %g0
>> 149
>> 150 EX(sth %g3, [%o0], and %o1, 3)
>> 151 add %o0, 2, %o0
>> 152 1:
>> 153 bne,a 8f
>> 154 EX(stb %g3, [%o0], and %o1, 1)
>> 155 8:
>> 156 retl
>> 157 clr %o0
>> 158 7:
>> 159 be 13b
>> 160 orcc %o1, 0, %g0
>> 161
>> 162 be 0f
>> 163 8:
>> 164 add %o0, 1, %o0
>> 165 subcc %o1, 1, %o1
>> 166 bne,a 8b
>> 167 EX(stb %g3, [%o0 - 1], add %o1, 1)
>> 168 0:
>> 169 retl
>> 170 clr %o0
>> 171 C_LABEL(__memset_end):
>> 172
>> 173 .section .fixup,#alloc,#execinstr
>> 174 .align 4
>> 175 20:
>> 176 cmp %g2, 8
>> 177 bleu 1f
>> 178 and %o1, 0x7f, %o1
>> 179 sub %g2, 9, %g2
>> 180 add %o3, 64, %o3
>> 181 1:
>> 182 sll %g2, 3, %g2
>> 183 add %o3, %o1, %o0
>> 184 b 30f
>> 185 sub %o0, %g2, %o0
>> 186 21:
>> 187 mov 8, %o0
>> 188 and %o1, 7, %o1
>> 189 sub %o0, %g2, %o0
>> 190 sll %o0, 3, %o0
>> 191 b 30f
>> 192 add %o0, %o1, %o0
>> 193 30:
>> 194 /* %o4 is faulting address, %o5 is %pc where fault occurred */
>> 195 save %sp, -104, %sp
>> 196 mov %i5, %o0
>> 197 mov %i7, %o1
>> 198 call C_LABEL(lookup_fault)
>> 199 mov %i4, %o2
>> 200 ret
>> 201 restore
>> 202
>> 203 .globl __bzero_end
>> 204 __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.