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 * Calls to memset returns initial %o0. Calls to bzero returns 0, if ok, and
6 * function for the alpha. !! 7 * number of bytes not yet set if exception occurs and we were called as
7 * !! 8 * clear_user.
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 */ 9 */
17 #include <linux/export.h> <<
18 .set noat <<
19 .set noreorder <<
20 .text <<
21 .globl memset <<
22 .globl __memset <<
23 .globl ___memset <<
24 .globl __memset16 <<
25 .globl __constant_c_memset <<
26 <<
27 .ent ___memset <<
28 .align 5 <<
29 ___memset: <<
30 .frame $30,0,$26,0 <<
31 .prologue 0 <<
32 <<
33 and $17,255,$1 /* E1 */ <<
34 insbl $17,1,$17 /* .. E0 */ <<
35 bis $17,$1,$17 /* E0 (p-c lat <<
36 sll $17,16,$1 /* E1 (p-c lat <<
37 <<
38 bis $17,$1,$17 /* E0 (p-c lat <<
39 sll $17,32,$1 /* E1 (p-c lat <<
40 bis $17,$1,$17 /* E0 (p-c lat <<
41 ldq_u $31,0($30) /* .. E1 */ <<
42 <<
43 .align 5 <<
44 __constant_c_memset: <<
45 addq $18,$16,$6 /* E0 */ <<
46 bis $16,$16,$0 /* .. E1 */ <<
47 xor $16,$6,$1 /* E0 */ <<
48 ble $18,end /* .. E1 */ <<
49 <<
50 bic $1,7,$1 /* E0 */ <<
51 beq $1,within_one_quad /* .. E1 (note <<
52 and $16,7,$3 /* E0 */ <<
53 beq $3,aligned /* .. E1 (note <<
54 <<
55 ldq_u $4,0($16) /* E0 */ <<
56 bis $16,$16,$5 /* .. E1 */ <<
57 insql $17,$16,$2 /* E0 */ <<
58 subq $3,8,$3 /* .. E1 */ <<
59 <<
60 addq $18,$3,$18 /* E0 $18 is <<
61 mskql $4,$16,$4 /* .. E1 (and <<
62 subq $16,$3,$16 /* E0 $16 is <<
63 bis $2,$4,$1 /* .. E1 */ <<
64 <<
65 bis $31,$31,$31 /* E0 */ <<
66 ldq_u $31,0($30) /* .. E1 */ <<
67 stq_u $1,0($5) /* E0 */ <<
68 bis $31,$31,$31 /* .. E1 */ <<
69 <<
70 .align 4 <<
71 aligned: <<
72 sra $18,3,$3 /* E0 */ <<
73 and $18,7,$18 /* .. E1 */ <<
74 bis $16,$16,$5 /* E0 */ <<
75 beq $3,no_quad /* .. E1 */ <<
76 <<
77 .align 3 <<
78 loop: <<
79 stq $17,0($5) /* E0 */ <<
80 subq $3,1,$3 /* .. E1 */ <<
81 addq $5,8,$5 /* E0 */ <<
82 bne $3,loop /* .. E1 */ <<
83 <<
84 no_quad: <<
85 bis $31,$31,$31 /* E0 */ <<
86 beq $18,end /* .. E1 */ <<
87 ldq $7,0($5) /* E0 */ <<
88 mskqh $7,$6,$2 /* .. E1 (and <<
89 <<
90 insqh $17,$6,$4 /* E0 */ <<
91 bis $2,$4,$1 /* .. E1 */ <<
92 stq $1,0($5) /* E0 */ <<
93 ret $31,($26),1 /* .. E1 */ <<
94 <<
95 .align 3 <<
96 within_one_quad: <<
97 ldq_u $1,0($16) /* E0 */ <<
98 insql $17,$16,$2 /* E1 */ <<
99 mskql $1,$16,$4 /* E0 (after l <<
100 bis $2,$4,$2 /* E0 */ <<
101 <<
102 mskql $2,$6,$4 /* E0 */ <<
103 mskqh $1,$6,$2 /* .. E1 */ <<
104 bis $2,$4,$1 /* E0 */ <<
105 stq_u $1,0($16) /* E0 */ <<
106 <<
107 end: <<
108 ret $31,($26),1 /* E1 */ <<
109 .end ___memset <<
110 EXPORT_SYMBOL(___memset) <<
111 EXPORT_SYMBOL(__constant_c_memset) <<
112 <<
113 .align 5 <<
114 .ent __memset16 <<
115 __memset16: <<
116 .prologue 0 <<
117 <<
118 inswl $17,0,$1 /* E0 */ <<
119 inswl $17,2,$2 /* E0 */ <<
120 inswl $17,4,$3 /* E0 */ <<
121 or $1,$2,$1 /* .. E1 */ <<
122 inswl $17,6,$4 /* E0 */ <<
123 or $1,$3,$1 /* .. E1 */ <<
124 or $1,$4,$17 /* E0 */ <<
125 br __constant_c_memset /* .. E1 */ <<
126 10
127 .end __memset16 !! 11 #include <asm/ptrace.h>
128 EXPORT_SYMBOL(__memset16) !! 12 #include <asm/export.h>
129 13
130 memset = ___memset !! 14 /* Work around cpp -rob */
131 __memset = ___memset !! 15 #define ALLOC #alloc
>> 16 #define EXECINSTR #execinstr
>> 17 #define EX(x,y,a,b) \
>> 18 98: x,y; \
>> 19 .section .fixup,ALLOC,EXECINSTR; \
>> 20 .align 4; \
>> 21 99: ba 30f; \
>> 22 a, b, %o0; \
>> 23 .section __ex_table,ALLOC; \
>> 24 .align 4; \
>> 25 .word 98b, 99b; \
>> 26 .text; \
>> 27 .align 4
>> 28
>> 29 #define EXT(start,end,handler) \
>> 30 .section __ex_table,ALLOC; \
>> 31 .align 4; \
>> 32 .word start, 0, end, handler; \
>> 33 .text; \
>> 34 .align 4
>> 35
>> 36 /* Please don't change these macros, unless you change the logic
>> 37 * in the .fixup section below as well.
>> 38 * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
>> 39 #define ZERO_BIG_BLOCK(base, offset, source) \
>> 40 std source, [base + offset + 0x00]; \
>> 41 std source, [base + offset + 0x08]; \
>> 42 std source, [base + offset + 0x10]; \
>> 43 std source, [base + offset + 0x18]; \
>> 44 std source, [base + offset + 0x20]; \
>> 45 std source, [base + offset + 0x28]; \
>> 46 std source, [base + offset + 0x30]; \
>> 47 std source, [base + offset + 0x38];
>> 48
>> 49 #define ZERO_LAST_BLOCKS(base, offset, source) \
>> 50 std source, [base - offset - 0x38]; \
>> 51 std source, [base - offset - 0x30]; \
>> 52 std source, [base - offset - 0x28]; \
>> 53 std source, [base - offset - 0x20]; \
>> 54 std source, [base - offset - 0x18]; \
>> 55 std source, [base - offset - 0x10]; \
>> 56 std source, [base - offset - 0x08]; \
>> 57 std source, [base - offset - 0x00];
>> 58
>> 59 .text
>> 60 .align 4
>> 61
>> 62 .globl __bzero_begin
>> 63 __bzero_begin:
>> 64
>> 65 .globl __bzero
>> 66 .globl memset
>> 67 EXPORT_SYMBOL(__bzero)
132 EXPORT_SYMBOL(memset) 68 EXPORT_SYMBOL(memset)
133 EXPORT_SYMBOL(__memset) !! 69 .globl __memset_start, __memset_end
>> 70 __memset_start:
>> 71 memset:
>> 72 mov %o0, %g1
>> 73 mov 1, %g4
>> 74 and %o1, 0xff, %g3
>> 75 sll %g3, 8, %g2
>> 76 or %g3, %g2, %g3
>> 77 sll %g3, 16, %g2
>> 78 or %g3, %g2, %g3
>> 79 b 1f
>> 80 mov %o2, %o1
>> 81 3:
>> 82 cmp %o2, 3
>> 83 be 2f
>> 84 EX(stb %g3, [%o0], sub %o1, 0)
>> 85
>> 86 cmp %o2, 2
>> 87 be 2f
>> 88 EX(stb %g3, [%o0 + 0x01], sub %o1, 1)
>> 89
>> 90 EX(stb %g3, [%o0 + 0x02], sub %o1, 2)
>> 91 2:
>> 92 sub %o2, 4, %o2
>> 93 add %o1, %o2, %o1
>> 94 b 4f
>> 95 sub %o0, %o2, %o0
>> 96
>> 97 __bzero:
>> 98 clr %g4
>> 99 mov %g0, %g3
>> 100 1:
>> 101 cmp %o1, 7
>> 102 bleu 7f
>> 103 andcc %o0, 3, %o2
>> 104
>> 105 bne 3b
>> 106 4:
>> 107 andcc %o0, 4, %g0
>> 108
>> 109 be 2f
>> 110 mov %g3, %g2
>> 111
>> 112 EX(st %g3, [%o0], sub %o1, 0)
>> 113 sub %o1, 4, %o1
>> 114 add %o0, 4, %o0
>> 115 2:
>> 116 andcc %o1, 0xffffff80, %o3 ! Now everything is 8 aligned and o1 is len to run
>> 117 be 9f
>> 118 andcc %o1, 0x78, %o2
>> 119 10:
>> 120 ZERO_BIG_BLOCK(%o0, 0x00, %g2)
>> 121 subcc %o3, 128, %o3
>> 122 ZERO_BIG_BLOCK(%o0, 0x40, %g2)
>> 123 11:
>> 124 EXT(10b, 11b, 20f)
>> 125 bne 10b
>> 126 add %o0, 128, %o0
>> 127
>> 128 orcc %o2, %g0, %g0
>> 129 9:
>> 130 be 13f
>> 131 andcc %o1, 7, %o1
>> 132
>> 133 srl %o2, 1, %o3
>> 134 set 13f, %o4
>> 135 sub %o4, %o3, %o4
>> 136 jmp %o4
>> 137 add %o0, %o2, %o0
>> 138
>> 139 12:
>> 140 ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
>> 141 ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
>> 142 13:
>> 143 be 8f
>> 144 andcc %o1, 4, %g0
>> 145
>> 146 be 1f
>> 147 andcc %o1, 2, %g0
>> 148
>> 149 EX(st %g3, [%o0], and %o1, 7)
>> 150 add %o0, 4, %o0
>> 151 1:
>> 152 be 1f
>> 153 andcc %o1, 1, %g0
>> 154
>> 155 EX(sth %g3, [%o0], and %o1, 3)
>> 156 add %o0, 2, %o0
>> 157 1:
>> 158 bne,a 8f
>> 159 EX(stb %g3, [%o0], and %o1, 1)
>> 160 8:
>> 161 b 0f
>> 162 nop
>> 163 7:
>> 164 be 13b
>> 165 orcc %o1, 0, %g0
>> 166
>> 167 be 0f
>> 168 8:
>> 169 add %o0, 1, %o0
>> 170 subcc %o1, 1, %o1
>> 171 bne 8b
>> 172 EX(stb %g3, [%o0 - 1], add %o1, 1)
>> 173 0:
>> 174 andcc %g4, 1, %g0
>> 175 be 5f
>> 176 nop
>> 177 retl
>> 178 mov %g1, %o0
>> 179 5:
>> 180 retl
>> 181 clr %o0
>> 182 __memset_end:
>> 183
>> 184 .section .fixup,#alloc,#execinstr
>> 185 .align 4
>> 186 20:
>> 187 cmp %g2, 8
>> 188 bleu 1f
>> 189 and %o1, 0x7f, %o1
>> 190 sub %g2, 9, %g2
>> 191 add %o3, 64, %o3
>> 192 1:
>> 193 sll %g2, 3, %g2
>> 194 add %o3, %o1, %o0
>> 195 b 30f
>> 196 sub %o0, %g2, %o0
>> 197 21:
>> 198 mov 8, %o0
>> 199 and %o1, 7, %o1
>> 200 sub %o0, %g2, %o0
>> 201 sll %o0, 3, %o0
>> 202 b 30f
>> 203 add %o0, %o1, %o0
>> 204 30:
>> 205 /* %o4 is faulting address, %o5 is %pc where fault occurred */
>> 206 save %sp, -104, %sp
>> 207 mov %i5, %o0
>> 208 mov %i7, %o1
>> 209 call lookup_fault
>> 210 mov %i4, %o2
>> 211 ret
>> 212 restore
>> 213
>> 214 .globl __bzero_end
>> 215 __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.