1 /* SPDX-License-Identifier: GPL-2.0 */ !! 1 /* linux/arch/sparc/lib/memset.S: Sparc optimized memset, bzero and clear_user code
2 /* !! 2 * Copyright (C) 1991,1996 Free Software Foundation
3 * linux/arch/alpha/lib/memset.S !! 3 * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
>> 4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
4 * 5 *
5 * This is an efficient (and small) implementa !! 6 * Returns 0, if ok, and number of bytes not yet set if exception
6 * function for the alpha. !! 7 * occurs and we were called as clear_user.
7 * <<
8 * (C) Copyright 1996 Linus Torvalds <<
9 * <<
10 * This routine is "moral-ware": you are free <<
11 * the only obligation I put on you is a moral <<
12 * to the routine, please send me your improve <<
13 * <<
14 * The scheduling comments are according to th <<
15 * hand, so they might well be incorrect, plea <<
16 */ 8 */
17 #include <linux/export.h> !! 9
18 .set noat !! 10 #include <asm/ptrace.h>
19 .set noreorder !! 11
20 .text !! 12 /* Work around cpp -rob */
21 .globl memset !! 13 #define ALLOC #alloc
22 .globl __memset !! 14 #define EXECINSTR #execinstr
23 .globl ___memset !! 15 #define EX(x,y,a,b) \
24 .globl __memset16 !! 16 98: x,y; \
25 .globl __constant_c_memset !! 17 .section .fixup,ALLOC,EXECINSTR; \
26 !! 18 .align 4; \
27 .ent ___memset !! 19 99: ba 30f; \
28 .align 5 !! 20 a, b, %o0; \
29 ___memset: !! 21 .section __ex_table,ALLOC; \
30 .frame $30,0,$26,0 !! 22 .align 4; \
31 .prologue 0 !! 23 .word 98b, 99b; \
32 !! 24 .text; \
33 and $17,255,$1 /* E1 */ !! 25 .align 4
34 insbl $17,1,$17 /* .. E0 */ !! 26
35 bis $17,$1,$17 /* E0 (p-c lat !! 27 #define EXT(start,end,handler) \
36 sll $17,16,$1 /* E1 (p-c lat !! 28 .section __ex_table,ALLOC; \
37 !! 29 .align 4; \
38 bis $17,$1,$17 /* E0 (p-c lat !! 30 .word start, 0, end, handler; \
39 sll $17,32,$1 /* E1 (p-c lat !! 31 .text; \
40 bis $17,$1,$17 /* E0 (p-c lat !! 32 .align 4
41 ldq_u $31,0($30) /* .. E1 */ !! 33
42 !! 34 /* Please don't change these macros, unless you change the logic
43 .align 5 !! 35 * in the .fixup section below as well.
44 __constant_c_memset: !! 36 * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
45 addq $18,$16,$6 /* E0 */ !! 37 #define ZERO_BIG_BLOCK(base, offset, source) \
46 bis $16,$16,$0 /* .. E1 */ !! 38 std source, [base + offset + 0x00]; \
47 xor $16,$6,$1 /* E0 */ !! 39 std source, [base + offset + 0x08]; \
48 ble $18,end /* .. E1 */ !! 40 std source, [base + offset + 0x10]; \
49 !! 41 std source, [base + offset + 0x18]; \
50 bic $1,7,$1 /* E0 */ !! 42 std source, [base + offset + 0x20]; \
51 beq $1,within_one_quad /* .. E1 (note !! 43 std source, [base + offset + 0x28]; \
52 and $16,7,$3 /* E0 */ !! 44 std source, [base + offset + 0x30]; \
53 beq $3,aligned /* .. E1 (note !! 45 std source, [base + offset + 0x38];
54 !! 46
55 ldq_u $4,0($16) /* E0 */ !! 47 #define ZERO_LAST_BLOCKS(base, offset, source) \
56 bis $16,$16,$5 /* .. E1 */ !! 48 std source, [base - offset - 0x38]; \
57 insql $17,$16,$2 /* E0 */ !! 49 std source, [base - offset - 0x30]; \
58 subq $3,8,$3 /* .. E1 */ !! 50 std source, [base - offset - 0x28]; \
59 !! 51 std source, [base - offset - 0x20]; \
60 addq $18,$3,$18 /* E0 $18 is !! 52 std source, [base - offset - 0x18]; \
61 mskql $4,$16,$4 /* .. E1 (and !! 53 std source, [base - offset - 0x10]; \
62 subq $16,$3,$16 /* E0 $16 is !! 54 std source, [base - offset - 0x08]; \
63 bis $2,$4,$1 /* .. E1 */ !! 55 std source, [base - offset - 0x00];
64 !! 56
65 bis $31,$31,$31 /* E0 */ !! 57 .text
66 ldq_u $31,0($30) /* .. E1 */ !! 58 .align 4
67 stq_u $1,0($5) /* E0 */ !! 59
68 bis $31,$31,$31 /* .. E1 */ !! 60 .globl __bzero_begin
69 !! 61 __bzero_begin:
70 .align 4 !! 62
71 aligned: !! 63 .globl __bzero
72 sra $18,3,$3 /* E0 */ !! 64 .globl memset
73 and $18,7,$18 /* .. E1 */ !! 65 .globl __memset_start, __memset_end
74 bis $16,$16,$5 /* E0 */ !! 66 __memset_start:
75 beq $3,no_quad /* .. E1 */ !! 67 memset:
76 !! 68 and %o1, 0xff, %g3
77 .align 3 !! 69 sll %g3, 8, %g2
78 loop: !! 70 or %g3, %g2, %g3
79 stq $17,0($5) /* E0 */ !! 71 sll %g3, 16, %g2
80 subq $3,1,$3 /* .. E1 */ !! 72 or %g3, %g2, %g3
81 addq $5,8,$5 /* E0 */ !! 73 b 1f
82 bne $3,loop /* .. E1 */ !! 74 mov %o2, %o1
83 !! 75 3:
84 no_quad: !! 76 cmp %o2, 3
85 bis $31,$31,$31 /* E0 */ !! 77 be 2f
86 beq $18,end /* .. E1 */ !! 78 EX(stb %g3, [%o0], sub %o1, 0)
87 ldq $7,0($5) /* E0 */ !! 79
88 mskqh $7,$6,$2 /* .. E1 (and !! 80 cmp %o2, 2
89 !! 81 be 2f
90 insqh $17,$6,$4 /* E0 */ !! 82 EX(stb %g3, [%o0 + 0x01], sub %o1, 1)
91 bis $2,$4,$1 /* .. E1 */ !! 83
92 stq $1,0($5) /* E0 */ !! 84 EX(stb %g3, [%o0 + 0x02], sub %o1, 2)
93 ret $31,($26),1 /* .. E1 */ !! 85 2:
94 !! 86 sub %o2, 4, %o2
95 .align 3 !! 87 add %o1, %o2, %o1
96 within_one_quad: !! 88 b 4f
97 ldq_u $1,0($16) /* E0 */ !! 89 sub %o0, %o2, %o0
98 insql $17,$16,$2 /* E1 */ !! 90
99 mskql $1,$16,$4 /* E0 (after l !! 91 __bzero:
100 bis $2,$4,$2 /* E0 */ !! 92 mov %g0, %g3
101 !! 93 1:
102 mskql $2,$6,$4 /* E0 */ !! 94 cmp %o1, 7
103 mskqh $1,$6,$2 /* .. E1 */ !! 95 bleu 7f
104 bis $2,$4,$1 /* E0 */ !! 96 andcc %o0, 3, %o2
105 stq_u $1,0($16) /* E0 */ !! 97
106 !! 98 bne 3b
107 end: !! 99 4:
108 ret $31,($26),1 /* E1 */ !! 100 andcc %o0, 4, %g0
109 .end ___memset !! 101
110 EXPORT_SYMBOL(___memset) !! 102 be 2f
111 EXPORT_SYMBOL(__constant_c_memset) !! 103 mov %g3, %g2
112 !! 104
113 .align 5 !! 105 EX(st %g3, [%o0], sub %o1, 0)
114 .ent __memset16 !! 106 sub %o1, 4, %o1
115 __memset16: !! 107 add %o0, 4, %o0
116 .prologue 0 !! 108 2:
117 !! 109 andcc %o1, 0xffffff80, %o3 ! Now everything is 8 aligned and o1 is len to run
118 inswl $17,0,$1 /* E0 */ !! 110 be 9f
119 inswl $17,2,$2 /* E0 */ !! 111 andcc %o1, 0x78, %o2
120 inswl $17,4,$3 /* E0 */ !! 112 10:
121 or $1,$2,$1 /* .. E1 */ !! 113 ZERO_BIG_BLOCK(%o0, 0x00, %g2)
122 inswl $17,6,$4 /* E0 */ !! 114 subcc %o3, 128, %o3
123 or $1,$3,$1 /* .. E1 */ !! 115 ZERO_BIG_BLOCK(%o0, 0x40, %g2)
124 or $1,$4,$17 /* E0 */ !! 116 11:
125 br __constant_c_memset /* .. E1 */ !! 117 EXT(10b, 11b, 20f)
126 !! 118 bne 10b
127 .end __memset16 !! 119 add %o0, 128, %o0
128 EXPORT_SYMBOL(__memset16) !! 120
129 !! 121 orcc %o2, %g0, %g0
130 memset = ___memset !! 122 9:
131 __memset = ___memset !! 123 be 13f
132 EXPORT_SYMBOL(memset) !! 124 andcc %o1, 7, %o1
133 EXPORT_SYMBOL(__memset) !! 125
>> 126 srl %o2, 1, %o3
>> 127 set 13f, %o4
>> 128 sub %o4, %o3, %o4
>> 129 jmp %o4
>> 130 add %o0, %o2, %o0
>> 131
>> 132 12:
>> 133 ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
>> 134 ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
>> 135 13:
>> 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
>> 144 1:
>> 145 be 1f
>> 146 andcc %o1, 1, %g0
>> 147
>> 148 EX(sth %g3, [%o0], and %o1, 3)
>> 149 add %o0, 2, %o0
>> 150 1:
>> 151 bne,a 8f
>> 152 EX(stb %g3, [%o0], and %o1, 1)
>> 153 8:
>> 154 retl
>> 155 clr %o0
>> 156 7:
>> 157 be 13b
>> 158 orcc %o1, 0, %g0
>> 159
>> 160 be 0f
>> 161 8:
>> 162 add %o0, 1, %o0
>> 163 subcc %o1, 1, %o1
>> 164 bne 8b
>> 165 EX(stb %g3, [%o0 - 1], add %o1, 1)
>> 166 0:
>> 167 retl
>> 168 clr %o0
>> 169 __memset_end:
>> 170
>> 171 .section .fixup,#alloc,#execinstr
>> 172 .align 4
>> 173 20:
>> 174 cmp %g2, 8
>> 175 bleu 1f
>> 176 and %o1, 0x7f, %o1
>> 177 sub %g2, 9, %g2
>> 178 add %o3, 64, %o3
>> 179 1:
>> 180 sll %g2, 3, %g2
>> 181 add %o3, %o1, %o0
>> 182 b 30f
>> 183 sub %o0, %g2, %o0
>> 184 21:
>> 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
>> 191 30:
>> 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:
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.