1 /* SPDX-License-Identifier: GPL-2.0 */ !! 1 /*
2 /* linux/arch/sparc/lib/memset.S: Sparc optimi !! 2 * linux/arch/alpha/memset.S
3 * Copyright (C) 1991,1996 Free Software Found <<
4 * Copyright (C) 1996,1997 Jakub Jelinek (jj@s <<
5 * Copyright (C) 1996 David S. Miller (davem@c <<
6 * 3 *
7 * Calls to memset returns initial %o0. Calls !! 4 * This is an efficient (and small) implementation of the C library "memset()"
8 * number of bytes not yet set if exception oc !! 5 * function for the alpha.
9 * clear_user. !! 6 *
>> 7 * (C) Copyright 1996 Linus Torvalds
>> 8 *
>> 9 * This routine is "moral-ware": you are free to use it any way you wish, and
>> 10 * the only obligation I put on you is a moral one: if you make any improvements
>> 11 * to the routine, please send me your improvements for me to use similarly.
>> 12 *
>> 13 * The scheduling comments are according to the EV5 documentation (and done by
>> 14 * hand, so they might well be incorrect, please do tell me about it..)
10 */ 15 */
11 16
12 #include <linux/export.h> !! 17 .set noat
13 #include <asm/ptrace.h> !! 18 .set noreorder
14 !! 19 .text
15 /* Work around cpp -rob */ !! 20 .globl __memset
16 #define ALLOC #alloc !! 21 .globl __memsetw
17 #define EXECINSTR #execinstr !! 22 .globl __constant_c_memset
18 #define EX(x,y,a,b) !! 23 .ent __memset
19 98: x,y; !! 24 .align 5
20 .section .fixup,ALLOC,EXECINSTR; !! 25 __memset:
21 .align 4; !! 26 .frame $30,0,$26,0
22 99: retl; !! 27 .prologue 0
23 a, b, %o0; !! 28
24 .section __ex_table,ALLOC; !! 29 and $17,255,$1 /* E1 */
25 .align 4; !! 30 insbl $17,1,$17 /* .. E0 */
26 .word 98b, 99b; !! 31 bis $17,$1,$17 /* E0 (p-c latency, next cycle) */
27 .text; !! 32 sll $17,16,$1 /* E1 (p-c latency, next cycle) */
28 .align 4 !! 33
29 !! 34 bis $17,$1,$17 /* E0 (p-c latency, next cycle) */
30 #define STORE(source, base, offset, n) !! 35 sll $17,32,$1 /* E1 (p-c latency, next cycle) */
31 98: std source, [base + offset + n]; !! 36 bis $17,$1,$17 /* E0 (p-c latency, next cycle) */
32 .section .fixup,ALLOC,EXECINSTR; !! 37 ldq_u $31,0($30) /* .. E1 */
33 .align 4; !! 38
34 99: ba 30f; !! 39 .align 5
35 sub %o3, n - offset, %o3; !! 40 __constant_c_memset:
36 .section __ex_table,ALLOC; !! 41 addq $18,$16,$6 /* E0 */
37 .align 4; !! 42 bis $16,$16,$0 /* .. E1 */
38 .word 98b, 99b; !! 43 xor $16,$6,$1 /* E0 */
39 .text; !! 44 ble $18,end /* .. E1 */
40 .align 4; !! 45
41 !! 46 bic $1,7,$1 /* E0 */
42 #define STORE_LAST(source, base, offset, n) !! 47 beq $1,within_one_quad /* .. E1 (note EV5 zero-latency forwarding) */
43 EX(std source, [base - offset - n], !! 48 and $16,7,$3 /* E0 */
44 add %o1, offset + n); !! 49 beq $3,aligned /* .. E1 (note EV5 zero-latency forwarding) */
45 !! 50
46 /* Please don't change these macros, unless yo !! 51 ldq_u $4,0($16) /* E0 */
47 * in the .fixup section below as well. !! 52 bis $16,$16,$5 /* .. E1 */
48 * Store 64 bytes at (BASE + OFFSET) using val !! 53 insql $17,$16,$2 /* E0 */
49 #define ZERO_BIG_BLOCK(base, offset, source) !! 54 subq $3,8,$3 /* .. E1 */
50 STORE(source, base, offset, 0x00); !! 55
51 STORE(source, base, offset, 0x08); !! 56 addq $18,$3,$18 /* E0 $18 is new count ($3 is negative) */
52 STORE(source, base, offset, 0x10); !! 57 mskql $4,$16,$4 /* .. E1 (and possible load stall) */
53 STORE(source, base, offset, 0x18); !! 58 subq $16,$3,$16 /* E0 $16 is new aligned destination */
54 STORE(source, base, offset, 0x20); !! 59 bis $2,$4,$1 /* .. E1 */
55 STORE(source, base, offset, 0x28); !! 60
56 STORE(source, base, offset, 0x30); !! 61 bis $31,$31,$31 /* E0 */
57 STORE(source, base, offset, 0x38); !! 62 ldq_u $31,0($30) /* .. E1 */
58 !! 63 stq_u $1,0($5) /* E0 */
59 #define ZERO_LAST_BLOCKS(base, offset, source) !! 64 bis $31,$31,$31 /* .. E1 */
60 STORE_LAST(source, base, offset, 0x38) !! 65
61 STORE_LAST(source, base, offset, 0x30) !! 66 .align 4
62 STORE_LAST(source, base, offset, 0x28) !! 67 aligned:
63 STORE_LAST(source, base, offset, 0x20) !! 68 sra $18,3,$3 /* E0 */
64 STORE_LAST(source, base, offset, 0x18) !! 69 and $18,7,$18 /* .. E1 */
65 STORE_LAST(source, base, offset, 0x10) !! 70 bis $16,$16,$5 /* E0 */
66 STORE_LAST(source, base, offset, 0x08) !! 71 beq $3,no_quad /* .. E1 */
67 STORE_LAST(source, base, offset, 0x00) !! 72
68 !! 73 .align 3
69 .text !! 74 loop:
70 .align 4 !! 75 stq $17,0($5) /* E0 */
71 !! 76 subq $3,1,$3 /* .. E1 */
72 .globl __bzero_begin !! 77 addq $5,8,$5 /* E0 */
73 __bzero_begin: !! 78 bne $3,loop /* .. E1 */
74 !! 79
75 .globl __bzero !! 80 no_quad:
76 .type __bzero,#function !! 81 bis $31,$31,$31 /* E0 */
77 .globl memset !! 82 beq $18,end /* .. E1 */
78 EXPORT_SYMBOL(__bzero) !! 83 ldq $7,0($5) /* E0 */
79 EXPORT_SYMBOL(memset) !! 84 mskqh $7,$6,$2 /* .. E1 (and load stall) */
80 memset: !! 85
81 mov %o0, %g1 !! 86 insqh $17,$6,$4 /* E0 */
82 mov 1, %g4 !! 87 bis $2,$4,$1 /* .. E1 */
83 and %o1, 0xff, %g3 !! 88 stq $1,0($5) /* E0 */
84 sll %g3, 8, %g2 !! 89 ret $31,($26),1 /* .. E1 */
85 or %g3, %g2, %g3 !! 90
86 sll %g3, 16, %g2 !! 91 .align 3
87 or %g3, %g2, %g3 !! 92 within_one_quad:
88 b 1f !! 93 ldq_u $1,0($16) /* E0 */
89 mov %o2, %o1 !! 94 insql $17,$16,$2 /* E1 */
90 3: !! 95 mskql $1,$16,$4 /* E0 (after load stall) */
91 cmp %o2, 3 !! 96 bis $2,$4,$2 /* E0 */
92 be 2f !! 97
93 EX(stb %g3, [%o0], sub %o1, 0) !! 98 mskql $2,$6,$4 /* E0 */
94 !! 99 mskqh $1,$6,$2 /* .. E1 */
95 cmp %o2, 2 !! 100 bis $2,$4,$1 /* E0 */
96 be 2f !! 101 stq_u $1,0($16) /* E0 */
97 EX(stb %g3, [%o0 + 0x01], sub %o1, 1) !! 102
98 !! 103 end:
99 EX(stb %g3, [%o0 + 0x02], sub %o1, 2) !! 104 ret $31,($26),1 /* E1 */
100 2: !! 105 .end __memset
101 sub %o2, 4, %o2 !! 106
102 add %o1, %o2, %o1 !! 107 .align 5
103 b 4f !! 108 .ent __memsetw
104 sub %o0, %o2, %o0 !! 109 __memsetw:
105 !! 110 .prologue 0
106 __bzero: !! 111
107 clr %g4 !! 112 inswl $17,0,$1 /* E0 */
108 mov %g0, %g3 !! 113 inswl $17,2,$2 /* E0 */
109 1: !! 114 inswl $17,4,$3 /* E0 */
110 cmp %o1, 7 !! 115 or $1,$2,$1 /* .. E1 */
111 bleu 7f !! 116 inswl $17,6,$4 /* E0 */
112 andcc %o0, 3, %o2 !! 117 or $1,$3,$1 /* .. E1 */
113 !! 118 or $1,$4,$17 /* E0 */
114 bne 3b !! 119 br __constant_c_memset /* .. E1 */
115 4: !! 120
116 andcc %o0, 4, %g0 !! 121 .end __memsetw
117 !! 122
118 be 2f !! 123 #ifdef __ELF__
119 mov %g3, %g2 !! 124 .weak memset; memset = __memset
120 !! 125 #else
121 EX(st %g3, [%o0], sub %o1, 0) !! 126 .weakext memset, __memset
122 sub %o1, 4, %o1 !! 127 #endif
123 add %o0, 4, %o0 <<
124 2: <<
125 andcc %o1, 0xffffff80, %o3 ! Now <<
126 be 9f <<
127 andcc %o1, 0x78, %o2 <<
128 10: <<
129 ZERO_BIG_BLOCK(%o0, 0x00, %g2) <<
130 subcc %o3, 128, %o3 <<
131 ZERO_BIG_BLOCK(%o0, 0x40, %g2) <<
132 bne 10b <<
133 add %o0, 128, %o0 <<
134 <<
135 orcc %o2, %g0, %g0 <<
136 9: <<
137 be 13f <<
138 andcc %o1, 7, %o1 <<
139 <<
140 srl %o2, 1, %o3 <<
141 set 13f, %o4 <<
142 sub %o4, %o3, %o4 <<
143 jmp %o4 <<
144 add %o0, %o2, %o0 <<
145 <<
146 ZERO_LAST_BLOCKS(%o0, 0x48, %g2) <<
147 ZERO_LAST_BLOCKS(%o0, 0x08, %g2) <<
148 13: <<
149 be 8f <<
150 andcc %o1, 4, %g0 <<
151 <<
152 be 1f <<
153 andcc %o1, 2, %g0 <<
154 <<
155 EX(st %g3, [%o0], and %o1, 7) <<
156 add %o0, 4, %o0 <<
157 1: <<
158 be 1f <<
159 andcc %o1, 1, %g0 <<
160 <<
161 EX(sth %g3, [%o0], and %o1, 3) <<
162 add %o0, 2, %o0 <<
163 1: <<
164 bne,a 8f <<
165 EX(stb %g3, [%o0], and %o1, 1) <<
166 8: <<
167 b 0f <<
168 nop <<
169 7: <<
170 be 13b <<
171 orcc %o1, 0, %g0 <<
172 <<
173 be 0f <<
174 8: <<
175 add %o0, 1, %o0 <<
176 subcc %o1, 1, %o1 <<
177 bne 8b <<
178 EX(stb %g3, [%o0 - 1], add %o1, 1) <<
179 0: <<
180 andcc %g4, 1, %g0 <<
181 be 5f <<
182 nop <<
183 retl <<
184 mov %g1, %o0 <<
185 5: <<
186 retl <<
187 clr %o0 <<
188 <<
189 .section .fixup,#alloc,#execinstr <<
190 .align 4 <<
191 30: <<
192 and %o1, 0x7f, %o1 <<
193 retl <<
194 add %o3, %o1, %o0 <<
195 <<
196 .globl __bzero_end <<
197 __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.