1 /* SPDX-License-Identifier: GPL-2.0 */ !! 1 /*
2 /* $Id: memset.S,v 1.1 2000/04/14 16:49:01 mjd !! 2 * This file is subject to the terms and conditions of the GNU General Public
3 * !! 3 * License. See the file "COPYING" in the main directory of this archive
4 * "memset" implementation of SuperH !! 4 * for more details.
5 * <<
6 * Copyright (C) 1999 Niibe Yutaka <<
7 * 5 *
>> 6 * Copyright (C) 1998, 1999, 2000 by Ralf Baechle
>> 7 * Copyright (C) 1999, 2000 Silicon Graphics, Inc.
>> 8 * Copyright (C) 2007 by Maciej W. Rozycki
>> 9 * Copyright (C) 2011, 2012 MIPS Technologies, Inc.
>> 10 */
>> 11 #include <linux/export.h>
>> 12 #include <asm/asm.h>
>> 13 #include <asm/asm-offsets.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.
8 */ 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)
>> 47
>> 48 #define EX(insn,reg,addr,handler) \
>> 49 .if \mode == LEGACY_MODE; \
>> 50 9: insn reg, addr; \
>> 51 .else; \
>> 52 9: ___BUILD_EVA_INSN(insn, reg, addr); \
>> 53 .endif; \
>> 54 .section __ex_table,"a"; \
>> 55 PTR_WD 9b, handler; \
>> 56 .previous
>> 57
>> 58 .macro f_fill64 dst, offset, val, fixup, mode
>> 59 EX(LONG_S, \val, (\offset + 0 * STORSIZE)(\dst), \fixup)
>> 60 EX(LONG_S, \val, (\offset + 1 * STORSIZE)(\dst), \fixup)
>> 61 EX(LONG_S, \val, (\offset + 2 * STORSIZE)(\dst), \fixup)
>> 62 EX(LONG_S, \val, (\offset + 3 * STORSIZE)(\dst), \fixup)
>> 63 #if ((defined(CONFIG_CPU_MICROMIPS) && (LONGSIZE == 4)) || !defined(CONFIG_CPU_MICROMIPS))
>> 64 EX(LONG_S, \val, (\offset + 4 * STORSIZE)(\dst), \fixup)
>> 65 EX(LONG_S, \val, (\offset + 5 * STORSIZE)(\dst), \fixup)
>> 66 EX(LONG_S, \val, (\offset + 6 * STORSIZE)(\dst), \fixup)
>> 67 EX(LONG_S, \val, (\offset + 7 * STORSIZE)(\dst), \fixup)
>> 68 #endif
>> 69 #if (!defined(CONFIG_CPU_MICROMIPS) && (LONGSIZE == 4))
>> 70 EX(LONG_S, \val, (\offset + 8 * STORSIZE)(\dst), \fixup)
>> 71 EX(LONG_S, \val, (\offset + 9 * STORSIZE)(\dst), \fixup)
>> 72 EX(LONG_S, \val, (\offset + 10 * STORSIZE)(\dst), \fixup)
>> 73 EX(LONG_S, \val, (\offset + 11 * STORSIZE)(\dst), \fixup)
>> 74 EX(LONG_S, \val, (\offset + 12 * STORSIZE)(\dst), \fixup)
>> 75 EX(LONG_S, \val, (\offset + 13 * STORSIZE)(\dst), \fixup)
>> 76 EX(LONG_S, \val, (\offset + 14 * STORSIZE)(\dst), \fixup)
>> 77 EX(LONG_S, \val, (\offset + 15 * STORSIZE)(\dst), \fixup)
>> 78 #endif
>> 79 .endm
>> 80
>> 81 .align 5
>> 82
>> 83 /*
>> 84 * Macro to generate the __bzero{,_user} symbol
>> 85 * Arguments:
>> 86 * mode: LEGACY_MODE or EVA_MODE
>> 87 */
>> 88 .macro __BUILD_BZERO mode
>> 89 /* Initialize __memset if this is the first time we call this macro */
>> 90 .ifnotdef __memset
>> 91 .set __memset, 1
>> 92 .hidden __memset /* Make sure it does not leak */
>> 93 .endif
>> 94
>> 95 sltiu t0, a2, STORSIZE /* very small region? */
>> 96 .set noreorder
>> 97 bnez t0, .Lsmall_memset\@
>> 98 andi t0, a0, STORMASK /* aligned? */
>> 99 .set reorder
>> 100
>> 101 #ifdef CONFIG_CPU_MICROMIPS
>> 102 move t8, a1 /* used by 'swp' instruction */
>> 103 move t9, a1
>> 104 #endif
>> 105 .set noreorder
>> 106 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
>> 107 beqz t0, 1f
>> 108 PTR_SUBU t0, STORSIZE /* alignment in bytes */
>> 109 #else
>> 110 .set noat
>> 111 li AT, STORSIZE
>> 112 beqz t0, 1f
>> 113 PTR_SUBU t0, AT /* alignment in bytes */
>> 114 .set at
>> 115 #endif
>> 116 .set reorder
>> 117
>> 118 #ifndef CONFIG_CPU_NO_LOAD_STORE_LR
>> 119 R10KCBARRIER(0(ra))
>> 120 #ifdef __MIPSEB__
>> 121 EX(LONG_S_L, a1, (a0), .Lfirst_fixup\@) /* make word/dword aligned */
>> 122 #else
>> 123 EX(LONG_S_R, a1, (a0), .Lfirst_fixup\@) /* make word/dword aligned */
>> 124 #endif
>> 125 PTR_SUBU a0, t0 /* long align ptr */
>> 126 PTR_ADDU a2, t0 /* correct size */
>> 127
>> 128 #else /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 129 #define STORE_BYTE(N) \
>> 130 EX(sb, a1, N(a0), .Lbyte_fixup\@); \
>> 131 .set noreorder; \
>> 132 beqz t0, 0f; \
>> 133 PTR_ADDU t0, 1; \
>> 134 .set reorder;
>> 135
>> 136 PTR_ADDU a2, t0 /* correct size */
>> 137 PTR_ADDU t0, 1
>> 138 STORE_BYTE(0)
>> 139 STORE_BYTE(1)
>> 140 #if LONGSIZE == 4
>> 141 EX(sb, a1, 2(a0), .Lbyte_fixup\@)
>> 142 #else
>> 143 STORE_BYTE(2)
>> 144 STORE_BYTE(3)
>> 145 STORE_BYTE(4)
>> 146 STORE_BYTE(5)
>> 147 EX(sb, a1, 6(a0), .Lbyte_fixup\@)
>> 148 #endif
>> 149 0:
>> 150 ori a0, STORMASK
>> 151 xori a0, STORMASK
>> 152 PTR_ADDIU a0, STORSIZE
>> 153 #endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 154 1: ori t1, a2, 0x3f /* # of full blocks */
>> 155 xori t1, 0x3f
>> 156 andi t0, a2, 0x40-STORSIZE
>> 157 beqz t1, .Lmemset_partial\@ /* no block to fill */
>> 158
>> 159 PTR_ADDU t1, a0 /* end address */
>> 160 1: PTR_ADDIU a0, 64
>> 161 R10KCBARRIER(0(ra))
>> 162 f_fill64 a0, -64, FILL64RG, .Lfwd_fixup\@, \mode
>> 163 bne t1, a0, 1b
>> 164
>> 165 .Lmemset_partial\@:
>> 166 R10KCBARRIER(0(ra))
>> 167 PTR_LA t1, 2f /* where to start */
>> 168 #ifdef CONFIG_CPU_MICROMIPS
>> 169 LONG_SRL t7, t0, 1
>> 170 #endif
>> 171 #if LONGSIZE == 4
>> 172 PTR_SUBU t1, FILLPTRG
>> 173 #else
>> 174 .set noat
>> 175 LONG_SRL AT, FILLPTRG, 1
>> 176 PTR_SUBU t1, AT
>> 177 .set at
>> 178 #endif
>> 179 PTR_ADDU a0, t0 /* dest ptr */
>> 180 jr t1
>> 181
>> 182 /* ... but first do longs ... */
>> 183 f_fill64 a0, -64, FILL64RG, .Lpartial_fixup\@, \mode
>> 184 2: andi a2, STORMASK /* At most one long to go */
>> 185
>> 186 .set noreorder
>> 187 beqz a2, 1f
>> 188 #ifndef CONFIG_CPU_NO_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 /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 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 /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 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 #ifdef CONFIG_CPU_NO_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_NO_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
>> 288
>> 289 .endm
9 290
10 /* 291 /*
11 * void *memset(void *s, int c, siz !! 292 * memset(void *s, int c, size_t n)
>> 293 *
>> 294 * a0: start of area to clear
>> 295 * a1: char to fill with
>> 296 * a2: size of area to clear
12 */ 297 */
13 298
14 #include <linux/linkage.h> !! 299 LEAF(memset)
>> 300 EXPORT_SYMBOL(memset)
>> 301 move v0, a0 /* result */
>> 302 beqz a1, 1f
15 303
16 ENTRY(memset) !! 304 andi a1, 0xff /* spread fillword */
17 tst r6,r6 !! 305 LONG_SLL t1, a1, 8
18 bt/s 5f ! if n=0, do n !! 306 or a1, t1
19 add r6,r4 !! 307 LONG_SLL t1, a1, 16
20 mov #12,r0 !! 308 #if LONGSIZE == 8
21 cmp/gt r6,r0 !! 309 or a1, t1
22 bt/s 4f ! if it's too !! 310 LONG_SLL t1, a1, 32
23 mov r4,r0 !! 311 #endif
24 and #3,r0 !! 312 or a1, t1
25 cmp/eq #0,r0 <<
26 bt/s 2f ! It's aligned <<
27 sub r0,r6 <<
28 1: 313 1:
29 dt r0 !! 314 #ifndef CONFIG_EVA
30 bf/s 1b !! 315 FEXPORT(__bzero)
31 mov.b r5,@-r4 !! 316 EXPORT_SYMBOL(__bzero)
32 2: ! make VVVV !! 317 #endif
33 extu.b r5,r5 !! 318 __BUILD_BZERO LEGACY_MODE
34 swap.b r5,r0 ! V0 !! 319
35 or r0,r5 ! VV !! 320 #ifdef CONFIG_EVA
36 swap.w r5,r0 ! VV00 !! 321 LEAF(__bzero)
37 or r0,r5 ! VVVV !! 322 EXPORT_SYMBOL(__bzero)
38 ! !! 323 __BUILD_BZERO EVA_MODE
39 mov r6,r0 !! 324 END(__bzero)
40 shlr2 r0 !! 325 #endif
41 shlr r0 ! r0 = r6 >> 3 <<
42 3: <<
43 dt r0 <<
44 mov.l r5,@-r4 ! set 8-byte a <<
45 bf/s 3b <<
46 mov.l r5,@-r4 <<
47 ! <<
48 mov #7,r0 <<
49 and r0,r6 <<
50 tst r6,r6 <<
51 bt 5f <<
52 ! fill bytes <<
53 4: <<
54 dt r6 <<
55 bf/s 4b <<
56 mov.b r5,@-r4 <<
57 5: <<
58 rts <<
59 mov r4,r0 <<
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.