1 /* SPDX-License-Identifier: GPL-2.0 */ !! 1 /*
2 /* linux/arch/sparc/lib/memset.S: Sparc optimi !! 2 * This file is subject to the terms and conditions of the GNU General Public
3 * Copyright (C) 1991,1996 Free Software Found !! 3 * License. See the file "COPYING" in the main directory of this archive
4 * Copyright (C) 1996,1997 Jakub Jelinek (jj@s !! 4 * for more details.
5 * Copyright (C) 1996 David S. Miller (davem@c <<
6 * 5 *
7 * Calls to memset returns initial %o0. Calls !! 6 * Copyright (C) 1998, 1999, 2000 by Ralf Baechle
8 * number of bytes not yet set if exception oc !! 7 * Copyright (C) 1999, 2000 Silicon Graphics, Inc.
9 * clear_user. !! 8 * Copyright (C) 2007 by Maciej W. Rozycki
>> 9 * Copyright (C) 2011, 2012 MIPS Technologies, Inc.
10 */ 10 */
>> 11 #include <asm/asm.h>
>> 12 #include <asm/asm-offsets.h>
>> 13 #include <asm/export.h>
>> 14 #include <asm/regdef.h>
>> 15
>> 16 #if LONGSIZE == 4
>> 17 #define LONG_S_L swl
>> 18 #define LONG_S_R swr
>> 19 #else
>> 20 #define LONG_S_L sdl
>> 21 #define LONG_S_R sdr
>> 22 #endif
>> 23
>> 24 #ifdef CONFIG_CPU_MICROMIPS
>> 25 #define STORSIZE (LONGSIZE * 2)
>> 26 #define STORMASK (STORSIZE - 1)
>> 27 #define FILL64RG t8
>> 28 #define FILLPTRG t7
>> 29 #undef LONG_S
>> 30 #define LONG_S LONG_SP
>> 31 #else
>> 32 #define STORSIZE LONGSIZE
>> 33 #define STORMASK LONGMASK
>> 34 #define FILL64RG a1
>> 35 #define FILLPTRG t0
>> 36 #endif
>> 37
>> 38 #define LEGACY_MODE 1
>> 39 #define EVA_MODE 2
>> 40
>> 41 /*
>> 42 * No need to protect it with EVA #ifdefery. The generated block of code
>> 43 * will never be assembled if EVA is not enabled.
>> 44 */
>> 45 #define __EVAFY(insn, reg, addr) __BUILD_EVA_INSN(insn##e, reg, addr)
>> 46 #define ___BUILD_EVA_INSN(insn, reg, addr) __EVAFY(insn, reg, addr)
11 47
12 #include <linux/export.h> !! 48 #define EX(insn,reg,addr,handler) \
13 #include <asm/ptrace.h> !! 49 .if \mode == LEGACY_MODE; \
14 !! 50 9: insn reg, addr; \
15 /* Work around cpp -rob */ !! 51 .else; \
16 #define ALLOC #alloc !! 52 9: ___BUILD_EVA_INSN(insn, reg, addr); \
17 #define EXECINSTR #execinstr !! 53 .endif; \
18 #define EX(x,y,a,b) !! 54 .section __ex_table,"a"; \
19 98: x,y; !! 55 PTR 9b, handler; \
20 .section .fixup,ALLOC,EXECINSTR; !! 56 .previous
21 .align 4; !! 57
22 99: retl; !! 58 .macro f_fill64 dst, offset, val, fixup, mode
23 a, b, %o0; !! 59 EX(LONG_S, \val, (\offset + 0 * STORSIZE)(\dst), \fixup)
24 .section __ex_table,ALLOC; !! 60 EX(LONG_S, \val, (\offset + 1 * STORSIZE)(\dst), \fixup)
25 .align 4; !! 61 EX(LONG_S, \val, (\offset + 2 * STORSIZE)(\dst), \fixup)
26 .word 98b, 99b; !! 62 EX(LONG_S, \val, (\offset + 3 * STORSIZE)(\dst), \fixup)
27 .text; !! 63 #if ((defined(CONFIG_CPU_MICROMIPS) && (LONGSIZE == 4)) || !defined(CONFIG_CPU_MICROMIPS))
28 .align 4 !! 64 EX(LONG_S, \val, (\offset + 4 * STORSIZE)(\dst), \fixup)
29 !! 65 EX(LONG_S, \val, (\offset + 5 * STORSIZE)(\dst), \fixup)
30 #define STORE(source, base, offset, n) !! 66 EX(LONG_S, \val, (\offset + 6 * STORSIZE)(\dst), \fixup)
31 98: std source, [base + offset + n]; !! 67 EX(LONG_S, \val, (\offset + 7 * STORSIZE)(\dst), \fixup)
32 .section .fixup,ALLOC,EXECINSTR; !! 68 #endif
33 .align 4; !! 69 #if (!defined(CONFIG_CPU_MICROMIPS) && (LONGSIZE == 4))
34 99: ba 30f; !! 70 EX(LONG_S, \val, (\offset + 8 * STORSIZE)(\dst), \fixup)
35 sub %o3, n - offset, %o3; !! 71 EX(LONG_S, \val, (\offset + 9 * STORSIZE)(\dst), \fixup)
36 .section __ex_table,ALLOC; !! 72 EX(LONG_S, \val, (\offset + 10 * STORSIZE)(\dst), \fixup)
37 .align 4; !! 73 EX(LONG_S, \val, (\offset + 11 * STORSIZE)(\dst), \fixup)
38 .word 98b, 99b; !! 74 EX(LONG_S, \val, (\offset + 12 * STORSIZE)(\dst), \fixup)
39 .text; !! 75 EX(LONG_S, \val, (\offset + 13 * STORSIZE)(\dst), \fixup)
40 .align 4; !! 76 EX(LONG_S, \val, (\offset + 14 * STORSIZE)(\dst), \fixup)
41 !! 77 EX(LONG_S, \val, (\offset + 15 * STORSIZE)(\dst), \fixup)
42 #define STORE_LAST(source, base, offset, n) !! 78 #endif
43 EX(std source, [base - offset - n], !! 79 .endm
44 add %o1, offset + n); !! 80
45 !! 81 .align 5
46 /* Please don't change these macros, unless yo !! 82
47 * in the .fixup section below as well. !! 83 /*
48 * Store 64 bytes at (BASE + OFFSET) using val !! 84 * Macro to generate the __bzero{,_user} symbol
49 #define ZERO_BIG_BLOCK(base, offset, source) !! 85 * Arguments:
50 STORE(source, base, offset, 0x00); !! 86 * mode: LEGACY_MODE or EVA_MODE
51 STORE(source, base, offset, 0x08); !! 87 */
52 STORE(source, base, offset, 0x10); !! 88 .macro __BUILD_BZERO mode
53 STORE(source, base, offset, 0x18); !! 89 /* Initialize __memset if this is the first time we call this macro */
54 STORE(source, base, offset, 0x20); !! 90 .ifnotdef __memset
55 STORE(source, base, offset, 0x28); !! 91 .set __memset, 1
56 STORE(source, base, offset, 0x30); !! 92 .hidden __memset /* Make sure it does not leak */
57 STORE(source, base, offset, 0x38); !! 93 .endif
58 !! 94
59 #define ZERO_LAST_BLOCKS(base, offset, source) !! 95 sltiu t0, a2, STORSIZE /* very small region? */
60 STORE_LAST(source, base, offset, 0x38) !! 96 .set noreorder
61 STORE_LAST(source, base, offset, 0x30) !! 97 bnez t0, .Lsmall_memset\@
62 STORE_LAST(source, base, offset, 0x28) !! 98 andi t0, a0, STORMASK /* aligned? */
63 STORE_LAST(source, base, offset, 0x20) !! 99 .set reorder
64 STORE_LAST(source, base, offset, 0x18) !! 100
65 STORE_LAST(source, base, offset, 0x10) !! 101 #ifdef CONFIG_CPU_MICROMIPS
66 STORE_LAST(source, base, offset, 0x08) !! 102 move t8, a1 /* used by 'swp' instruction */
67 STORE_LAST(source, base, offset, 0x00) !! 103 move t9, a1
68 !! 104 #endif
69 .text !! 105 .set noreorder
70 .align 4 !! 106 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
71 !! 107 beqz t0, 1f
72 .globl __bzero_begin !! 108 PTR_SUBU t0, STORSIZE /* alignment in bytes */
73 __bzero_begin: !! 109 #else
74 !! 110 .set noat
75 .globl __bzero !! 111 li AT, STORSIZE
76 .type __bzero,#function !! 112 beqz t0, 1f
77 .globl memset !! 113 PTR_SUBU t0, AT /* alignment in bytes */
78 EXPORT_SYMBOL(__bzero) !! 114 .set at
79 EXPORT_SYMBOL(memset) !! 115 #endif
80 memset: !! 116 .set reorder
81 mov %o0, %g1 !! 117
82 mov 1, %g4 !! 118 #ifdef CONFIG_CPU_HAS_LOAD_STORE_LR
83 and %o1, 0xff, %g3 !! 119 R10KCBARRIER(0(ra))
84 sll %g3, 8, %g2 !! 120 #ifdef __MIPSEB__
85 or %g3, %g2, %g3 !! 121 EX(LONG_S_L, a1, (a0), .Lfirst_fixup\@) /* make word/dword aligned */
86 sll %g3, 16, %g2 !! 122 #else
87 or %g3, %g2, %g3 !! 123 EX(LONG_S_R, a1, (a0), .Lfirst_fixup\@) /* make word/dword aligned */
88 b 1f !! 124 #endif
89 mov %o2, %o1 !! 125 PTR_SUBU a0, t0 /* long align ptr */
90 3: !! 126 PTR_ADDU a2, t0 /* correct size */
91 cmp %o2, 3 !! 127
92 be 2f !! 128 #else /* !CONFIG_CPU_HAS_LOAD_STORE_LR */
93 EX(stb %g3, [%o0], sub %o1, 0) !! 129 #define STORE_BYTE(N) \
94 !! 130 EX(sb, a1, N(a0), .Lbyte_fixup\@); \
95 cmp %o2, 2 !! 131 .set noreorder; \
96 be 2f !! 132 beqz t0, 0f; \
97 EX(stb %g3, [%o0 + 0x01], sub %o1, 1) !! 133 PTR_ADDU t0, 1; \
98 !! 134 .set reorder;
99 EX(stb %g3, [%o0 + 0x02], sub %o1, 2) !! 135
100 2: !! 136 PTR_ADDU a2, t0 /* correct size */
101 sub %o2, 4, %o2 !! 137 PTR_ADDU t0, 1
102 add %o1, %o2, %o1 !! 138 STORE_BYTE(0)
103 b 4f !! 139 STORE_BYTE(1)
104 sub %o0, %o2, %o0 !! 140 #if LONGSIZE == 4
105 !! 141 EX(sb, a1, 2(a0), .Lbyte_fixup\@)
106 __bzero: !! 142 #else
107 clr %g4 !! 143 STORE_BYTE(2)
108 mov %g0, %g3 !! 144 STORE_BYTE(3)
109 1: !! 145 STORE_BYTE(4)
110 cmp %o1, 7 !! 146 STORE_BYTE(5)
111 bleu 7f !! 147 EX(sb, a1, 6(a0), .Lbyte_fixup\@)
112 andcc %o0, 3, %o2 !! 148 #endif
113 !! 149 0:
114 bne 3b !! 150 ori a0, STORMASK
115 4: !! 151 xori a0, STORMASK
116 andcc %o0, 4, %g0 !! 152 PTR_ADDIU a0, STORSIZE
117 !! 153 #endif /* !CONFIG_CPU_HAS_LOAD_STORE_LR */
118 be 2f !! 154 1: ori t1, a2, 0x3f /* # of full blocks */
119 mov %g3, %g2 !! 155 xori t1, 0x3f
120 !! 156 andi t0, a2, 0x40-STORSIZE
121 EX(st %g3, [%o0], sub %o1, 0) !! 157 beqz t1, .Lmemset_partial\@ /* no block to fill */
122 sub %o1, 4, %o1 !! 158
123 add %o0, 4, %o0 !! 159 PTR_ADDU t1, a0 /* end address */
124 2: !! 160 1: PTR_ADDIU a0, 64
125 andcc %o1, 0xffffff80, %o3 ! Now !! 161 R10KCBARRIER(0(ra))
126 be 9f !! 162 f_fill64 a0, -64, FILL64RG, .Lfwd_fixup\@, \mode
127 andcc %o1, 0x78, %o2 !! 163 bne t1, a0, 1b
128 10: !! 164
129 ZERO_BIG_BLOCK(%o0, 0x00, %g2) !! 165 .Lmemset_partial\@:
130 subcc %o3, 128, %o3 !! 166 R10KCBARRIER(0(ra))
131 ZERO_BIG_BLOCK(%o0, 0x40, %g2) !! 167 PTR_LA t1, 2f /* where to start */
132 bne 10b !! 168 #ifdef CONFIG_CPU_MICROMIPS
133 add %o0, 128, %o0 !! 169 LONG_SRL t7, t0, 1
134 !! 170 #endif
135 orcc %o2, %g0, %g0 !! 171 #if LONGSIZE == 4
136 9: !! 172 PTR_SUBU t1, FILLPTRG
137 be 13f !! 173 #else
138 andcc %o1, 7, %o1 !! 174 .set noat
139 !! 175 LONG_SRL AT, FILLPTRG, 1
140 srl %o2, 1, %o3 !! 176 PTR_SUBU t1, AT
141 set 13f, %o4 !! 177 .set at
142 sub %o4, %o3, %o4 !! 178 #endif
143 jmp %o4 !! 179 PTR_ADDU a0, t0 /* dest ptr */
144 add %o0, %o2, %o0 !! 180 jr t1
145 !! 181
146 ZERO_LAST_BLOCKS(%o0, 0x48, %g2) !! 182 /* ... but first do longs ... */
147 ZERO_LAST_BLOCKS(%o0, 0x08, %g2) !! 183 f_fill64 a0, -64, FILL64RG, .Lpartial_fixup\@, \mode
148 13: !! 184 2: andi a2, STORMASK /* At most one long to go */
149 be 8f !! 185
150 andcc %o1, 4, %g0 !! 186 .set noreorder
>> 187 beqz a2, 1f
>> 188 #ifdef CONFIG_CPU_HAS_LOAD_STORE_LR
>> 189 PTR_ADDU a0, a2 /* What's left */
>> 190 .set reorder
>> 191 R10KCBARRIER(0(ra))
>> 192 #ifdef __MIPSEB__
>> 193 EX(LONG_S_R, a1, -1(a0), .Llast_fixup\@)
>> 194 #else
>> 195 EX(LONG_S_L, a1, -1(a0), .Llast_fixup\@)
>> 196 #endif
>> 197 #else
>> 198 PTR_SUBU t0, $0, a2
>> 199 .set reorder
>> 200 move a2, zero /* No remaining longs */
>> 201 PTR_ADDIU t0, 1
>> 202 STORE_BYTE(0)
>> 203 STORE_BYTE(1)
>> 204 #if LONGSIZE == 4
>> 205 EX(sb, a1, 2(a0), .Lbyte_fixup\@)
>> 206 #else
>> 207 STORE_BYTE(2)
>> 208 STORE_BYTE(3)
>> 209 STORE_BYTE(4)
>> 210 STORE_BYTE(5)
>> 211 EX(sb, a1, 6(a0), .Lbyte_fixup\@)
>> 212 #endif
>> 213 0:
>> 214 #endif
>> 215 1: move a2, zero
>> 216 jr ra
>> 217
>> 218 .Lsmall_memset\@:
>> 219 PTR_ADDU t1, a0, a2
>> 220 beqz a2, 2f
>> 221
>> 222 1: PTR_ADDIU a0, 1 /* fill bytewise */
>> 223 R10KCBARRIER(0(ra))
>> 224 .set noreorder
>> 225 bne t1, a0, 1b
>> 226 EX(sb, a1, -1(a0), .Lsmall_fixup\@)
>> 227 .set reorder
>> 228
>> 229 2: move a2, zero
>> 230 jr ra /* done */
>> 231 .if __memset == 1
>> 232 END(memset)
>> 233 .set __memset, 0
>> 234 .hidden __memset
>> 235 .endif
>> 236
>> 237 #ifndef CONFIG_CPU_HAS_LOAD_STORE_LR
>> 238 .Lbyte_fixup\@:
>> 239 /*
>> 240 * unset_bytes = (#bytes - (#unaligned bytes)) - (-#unaligned bytes remaining + 1) + 1
>> 241 * a2 = a2 - t0 + 1
>> 242 */
>> 243 PTR_SUBU a2, t0
>> 244 PTR_ADDIU a2, 1
>> 245 jr ra
>> 246 #endif /* !CONFIG_CPU_HAS_LOAD_STORE_LR */
>> 247
>> 248 .Lfirst_fixup\@:
>> 249 /* unset_bytes already in a2 */
>> 250 jr ra
>> 251
>> 252 .Lfwd_fixup\@:
>> 253 /*
>> 254 * unset_bytes = partial_start_addr + #bytes - fault_addr
>> 255 * a2 = t1 + (a2 & 3f) - $28->task->BUADDR
>> 256 */
>> 257 PTR_L t0, TI_TASK($28)
>> 258 andi a2, 0x3f
>> 259 LONG_L t0, THREAD_BUADDR(t0)
>> 260 LONG_ADDU a2, t1
>> 261 LONG_SUBU a2, t0
>> 262 jr ra
>> 263
>> 264 .Lpartial_fixup\@:
>> 265 /*
>> 266 * unset_bytes = partial_end_addr + #bytes - fault_addr
>> 267 * a2 = a0 + (a2 & STORMASK) - $28->task->BUADDR
>> 268 */
>> 269 PTR_L t0, TI_TASK($28)
>> 270 andi a2, STORMASK
>> 271 LONG_L t0, THREAD_BUADDR(t0)
>> 272 LONG_ADDU a2, a0
>> 273 LONG_SUBU a2, t0
>> 274 jr ra
>> 275
>> 276 .Llast_fixup\@:
>> 277 /* unset_bytes already in a2 */
>> 278 jr ra
>> 279
>> 280 .Lsmall_fixup\@:
>> 281 /*
>> 282 * unset_bytes = end_addr - current_addr + 1
>> 283 * a2 = t1 - a0 + 1
>> 284 */
>> 285 PTR_SUBU a2, t1, a0
>> 286 PTR_ADDIU a2, 1
>> 287 jr ra
151 288
152 be 1f !! 289 .endm
153 andcc %o1, 2, %g0 <<
154 290
155 EX(st %g3, [%o0], and %o1, 7) !! 291 /*
156 add %o0, 4, %o0 !! 292 * memset(void *s, int c, size_t n)
157 1: !! 293 *
158 be 1f !! 294 * a0: start of area to clear
159 andcc %o1, 1, %g0 !! 295 * a1: char to fill with
>> 296 * a2: size of area to clear
>> 297 */
160 298
161 EX(sth %g3, [%o0], and %o1, 3) !! 299 LEAF(memset)
162 add %o0, 2, %o0 !! 300 EXPORT_SYMBOL(memset)
>> 301 move v0, a0 /* result */
>> 302 beqz a1, 1f
>> 303
>> 304 andi a1, 0xff /* spread fillword */
>> 305 LONG_SLL t1, a1, 8
>> 306 or a1, t1
>> 307 LONG_SLL t1, a1, 16
>> 308 #if LONGSIZE == 8
>> 309 or a1, t1
>> 310 LONG_SLL t1, a1, 32
>> 311 #endif
>> 312 or a1, t1
163 1: 313 1:
164 bne,a 8f !! 314 #ifndef CONFIG_EVA
165 EX(stb %g3, [%o0], and %o1, 1) !! 315 FEXPORT(__bzero)
166 8: !! 316 EXPORT_SYMBOL(__bzero)
167 b 0f !! 317 #else
168 nop !! 318 FEXPORT(__bzero_kernel)
169 7: !! 319 EXPORT_SYMBOL(__bzero_kernel)
170 be 13b !! 320 #endif
171 orcc %o1, 0, %g0 !! 321 __BUILD_BZERO LEGACY_MODE
172 !! 322
173 be 0f !! 323 #ifdef CONFIG_EVA
174 8: !! 324 LEAF(__bzero)
175 add %o0, 1, %o0 !! 325 EXPORT_SYMBOL(__bzero)
176 subcc %o1, 1, %o1 !! 326 __BUILD_BZERO EVA_MODE
177 bne 8b !! 327 END(__bzero)
178 EX(stb %g3, [%o0 - 1], add %o1, 1) !! 328 #endif
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.