| 1 | /* |
| 2 | * arch/ubicom32/lib/mem_ubicom32.c |
| 3 | * String functions. |
| 4 | * |
| 5 | * (C) Copyright 2009, Ubicom, Inc. |
| 6 | * |
| 7 | * This file is part of the Ubicom32 Linux Kernel Port. |
| 8 | * |
| 9 | * The Ubicom32 Linux Kernel Port is free software: you can redistribute |
| 10 | * it and/or modify it under the terms of the GNU General Public License |
| 11 | * as published by the Free Software Foundation, either version 2 of the |
| 12 | * License, or (at your option) any later version. |
| 13 | * |
| 14 | * The Ubicom32 Linux Kernel Port is distributed in the hope that it |
| 15 | * will be useful, but WITHOUT ANY WARRANTY; without even the implied |
| 16 | * warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See |
| 17 | * the GNU General Public License for more details. |
| 18 | * |
| 19 | * You should have received a copy of the GNU General Public License |
| 20 | * along with the Ubicom32 Linux Kernel Port. If not, |
| 21 | * see <http://www.gnu.org/licenses/>. |
| 22 | * |
| 23 | * Ubicom32 implementation derived from (with many thanks): |
| 24 | * arch/m68knommu |
| 25 | * arch/blackfin |
| 26 | * arch/parisc |
| 27 | */ |
| 28 | |
| 29 | #include <linux/module.h> |
| 30 | #include <linux/types.h> |
| 31 | #include <linux/compiler.h> |
| 32 | |
| 33 | #define LIKELY likely |
| 34 | #define UNLIKELY unlikely |
| 35 | |
| 36 | typedef u32_t addr_t; |
| 37 | |
| 38 | /* |
| 39 | * memcpy() |
| 40 | */ |
| 41 | void *memcpy(void *dest, const void *src, size_t n) |
| 42 | { |
| 43 | void *dest_ret = dest; |
| 44 | |
| 45 | if (LIKELY((((addr_t)dest ^ (addr_t)src) & 3) == 0) && LIKELY(n > 6)) { |
| 46 | size_t m; |
| 47 | n -= (4 - (addr_t)dest) & 0x03; |
| 48 | m = n >> 2; |
| 49 | asm volatile ( |
| 50 | " sub.4 d15, #0, %2 \n\t" // set up for jump table |
| 51 | " and.4 d15, #(32-1), d15 \n\t" // d15 = (-m) & (32 - 1) |
| 52 | " moveai a3, #%%hi(1f) \n\t" |
| 53 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 54 | " lea.4 a3, (a3,d15) \n\t" |
| 55 | |
| 56 | " bfextu d15, %0, #2 \n\t" // d15 = (dest & 3) |
| 57 | " jmpne.w.f 100f \n\t" |
| 58 | " calli a3, 0(a3) \n\t" // 4-byte alignment |
| 59 | |
| 60 | "100: cmpi d15, #2 \n\t" |
| 61 | " jmpne.s.f 101f \n\t" |
| 62 | " move.2 (%0)2++, (%1)2++ \n\t" |
| 63 | " calli a3, 0(a3) \n\t" // 2-byte alignment |
| 64 | |
| 65 | "101: move.1 (%0)1++, (%1)1++ \n\t" |
| 66 | " jmpgt.s.f 102f \n\t" // 3-byte alignment |
| 67 | " move.2 (%0)2++, (%1)2++ \n\t" // 1-byte alignment |
| 68 | "102: calli a3, 0(a3) \n\t" |
| 69 | |
| 70 | "200: cmpi %3, #2 \n\t" |
| 71 | " jmplt.s.f 201f \n\t" |
| 72 | " move.2 (%0)2++, (%1)2++ \n\t" |
| 73 | " jmpeq.s.t 2f \n\t" |
| 74 | "201: move.1 (%0)1++, (%1)1++ \n\t" |
| 75 | " jmpt.w.t 2f \n\t" |
| 76 | |
| 77 | "1: .rept 25 \n\t" |
| 78 | " movea (%0)4++, (%1)4++ \n\t" |
| 79 | " .endr \n\t" |
| 80 | " .rept 7 \n\t" |
| 81 | " move.4 (%0)4++, (%1)4++ \n\t" |
| 82 | " .endr \n\t" |
| 83 | " add.4 %2, #-32, %2 \n\t" |
| 84 | " jmpgt.w.f 1b \n\t" |
| 85 | |
| 86 | " and.4 %3, #3, %3 \n\t" // check n |
| 87 | " jmpne.w.f 200b \n\t" |
| 88 | "2: \n\t" |
| 89 | : "+a" (dest), "+a" (src), "+d" (m), "+d" (n) |
| 90 | : |
| 91 | : "d15", "a3", "memory", "cc" |
| 92 | ); |
| 93 | |
| 94 | return dest_ret; |
| 95 | } |
| 96 | |
| 97 | if (LIKELY((((addr_t)dest ^ (addr_t)src) & 1) == 0) && LIKELY(n > 2)) { |
| 98 | size_t m; |
| 99 | n -= (addr_t)dest & 0x01; |
| 100 | m = n >> 1; |
| 101 | asm volatile ( |
| 102 | " sub.4 d15, #0, %2 \n\t" // set up for jump table |
| 103 | " and.4 d15, #(32-1), d15 \n\t" // d15 = (-m) & (32 - 1) |
| 104 | " moveai a3, #%%hi(1f) \n\t" |
| 105 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 106 | " lea.4 a3, (a3,d15) \n\t" |
| 107 | |
| 108 | " btst %0, #0 \n\t" // check bit 0 |
| 109 | " jmpne.w.f 100f \n\t" |
| 110 | " calli a3, 0(a3) \n\t" // 4-byte alignment |
| 111 | |
| 112 | "100: move.1 (%0)1++, (%1)1++ \n\t" |
| 113 | " calli a3, 0(a3) \n\t" |
| 114 | |
| 115 | "200: move.1 (%0)1++, (%1)1++ \n\t" |
| 116 | " jmpt.w.t 2f \n\t" |
| 117 | |
| 118 | "1: .rept 32 \n\t" |
| 119 | " move.2 (%0)2++, (%1)2++ \n\t" |
| 120 | " .endr \n\t" |
| 121 | " add.4 %2, #-32, %2 \n\t" |
| 122 | " jmpgt.w.f 1b \n\t" |
| 123 | |
| 124 | " and.4 %3, #1, %3 \n\t" // check n |
| 125 | " jmpne.w.f 200b \n\t" |
| 126 | "2: \n\t" |
| 127 | |
| 128 | : "+a" (dest), "+a" (src), "+d" (m), "+d" (n) |
| 129 | : |
| 130 | : "d15", "a3", "memory", "cc" |
| 131 | ); |
| 132 | |
| 133 | return dest_ret; |
| 134 | } |
| 135 | |
| 136 | asm volatile ( |
| 137 | " sub.4 d15, #0, %2 \n\t" |
| 138 | " jmpeq.w.f 2f \n\t" |
| 139 | " and.4 d15, #(16-1), d15 \n\t" // d15 = (-n) & (16 - 1) |
| 140 | " moveai a3, #%%hi(1f) \n\t" |
| 141 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 142 | " lea.4 a3, (a3,d15) \n\t" |
| 143 | " calli a3, 0(a3) \n\t" |
| 144 | |
| 145 | "1: .rept 16 \n\t" |
| 146 | " move.1 (%0)1++, (%1)1++ \n\t" |
| 147 | " .endr \n\t" |
| 148 | " add.4 %2, #-16, %2 \n\t" |
| 149 | " jmpgt.w.f 1b \n\t" |
| 150 | "2: \n\t" |
| 151 | |
| 152 | : "+a" (dest), "+a" (src), "+d" (n) |
| 153 | : |
| 154 | : "d15", "a3", "memory", "cc" |
| 155 | ); |
| 156 | |
| 157 | return dest_ret; |
| 158 | } |
| 159 | |
| 160 | /* |
| 161 | * memset() |
| 162 | */ |
| 163 | void *memset(void *s, int c, size_t n) |
| 164 | { |
| 165 | void *s_ret = s; |
| 166 | |
| 167 | if (LIKELY(n > 6)) { |
| 168 | size_t m; |
| 169 | n -= (4 - (addr_t)s) & 0x03; |
| 170 | m = n >> 2; |
| 171 | asm volatile ( |
| 172 | " sub.4 d15, #0, %2 \n\t" // set up for jump table |
| 173 | " and.4 d15, #(32-1), d15 \n\t" // d15 = (-m) & (32 - 1) |
| 174 | " shmrg.1 %1, %1, %1 \n\t" |
| 175 | " shmrg.2 %1, %1, %1 \n\t" // %1 = (c<<24)|(c<<16)|(c<<8)|c |
| 176 | " moveai a3, #%%hi(1f) \n\t" |
| 177 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 178 | " lea.4 a3, (a3,d15) \n\t" |
| 179 | |
| 180 | " bfextu d15, %0, #2 \n\t" // d15 = (s & 3) |
| 181 | " jmpne.w.f 100f \n\t" |
| 182 | " calli a3, 0(a3) \n\t" // 4-byte alignment |
| 183 | |
| 184 | "100: cmpi d15, #2 \n\t" |
| 185 | " jmpne.s.f 101f \n\t" |
| 186 | " move.2 (%0)2++, %1 \n\t" |
| 187 | " calli a3, 0(a3) \n\t" // 2-byte alignment |
| 188 | |
| 189 | "101: move.1 (%0)1++, %1 \n\t" |
| 190 | " jmpgt.s.f 102f \n\t" // 3-byte alignment |
| 191 | " move.2 (%0)2++, %1 \n\t" // 1-byte alignment |
| 192 | "102: calli a3, 0(a3) \n\t" |
| 193 | |
| 194 | "200: cmpi %3, #2 \n\t" |
| 195 | " jmplt.s.f 201f \n\t" |
| 196 | " move.2 (%0)2++, %1 \n\t" |
| 197 | " jmpeq.s.t 2f \n\t" |
| 198 | "201: move.1 (%0)1++, %1 \n\t" |
| 199 | " jmpt.w.t 2f \n\t" |
| 200 | |
| 201 | "1: .rept 25 \n\t" |
| 202 | " movea (%0)4++, %1 \n\t" |
| 203 | " .endr \n\t" |
| 204 | " .rept 7 \n\t" |
| 205 | " move.4 (%0)4++, %1 \n\t" |
| 206 | " .endr \n\t" |
| 207 | " add.4 %2, #-32, %2 \n\t" |
| 208 | " jmpgt.w.f 1b \n\t" |
| 209 | |
| 210 | " and.4 %3, #3, %3 \n\t" // test bit 1 of n |
| 211 | " jmpne.w.f 200b \n\t" |
| 212 | "2: \n\t" |
| 213 | |
| 214 | : "+a" (s), "+d" (c), "+d" (m), "+d" (n) |
| 215 | : |
| 216 | : "d15", "a3", "memory", "cc" |
| 217 | ); |
| 218 | |
| 219 | return s_ret; |
| 220 | } |
| 221 | |
| 222 | asm volatile ( |
| 223 | " sub.4 d15, #0, %2 \n\t" |
| 224 | " jmpeq.w.f 2f \n\t" |
| 225 | " and.4 d15, #(8-1), d15 \n\t" // d15 = (-%2) & (16 - 1) |
| 226 | " moveai a3, #%%hi(1f) \n\t" |
| 227 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 228 | " lea.4 a3, (a3,d15) \n\t" |
| 229 | " calli a3, 0(a3) \n\t" |
| 230 | |
| 231 | "1: .rept 8 \n\t" |
| 232 | " move.1 (%0)1++, %1 \n\t" |
| 233 | " .endr \n\t" |
| 234 | "2: \n\t" |
| 235 | |
| 236 | : "+a" (s), "+d" (c), "+d" (n) |
| 237 | : |
| 238 | : "d15", "a3", "memory", "cc" |
| 239 | ); |
| 240 | |
| 241 | return s_ret; |
| 242 | } |
| 243 | |
| 244 | void *memmove(void *dest, const void *src, size_t n) |
| 245 | { |
| 246 | char *tmp; |
| 247 | const char *s; |
| 248 | |
| 249 | if (n == 0) |
| 250 | return dest; |
| 251 | |
| 252 | tmp = dest; |
| 253 | s = src; |
| 254 | |
| 255 | /* |
| 256 | * Will perform 16-bit move if possible |
| 257 | */ |
| 258 | if (likely((((u32)dest | (u32)src | n) & 1) == 0)) { |
| 259 | if (dest <= src) { |
| 260 | asm volatile ( |
| 261 | " sub.4 d15, #0, %2 \n\t" // set up for jump table |
| 262 | " and.4 d15, #(32-2), d15 \n\t" // d15 = (- count) & (32 - 2) |
| 263 | " moveai a3, #%%hi(1f) \n\t" |
| 264 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 265 | " lea.2 a3, (a3,d15) \n\t" |
| 266 | " calli a3, 0(a3) \n\t" |
| 267 | |
| 268 | "1: .rept 16 \n\t" |
| 269 | " move.2 (%0)2++, (%1)2++ \n\t" |
| 270 | " .endr \n\t" |
| 271 | " add.4 %2, #-32, %2 \n\t" |
| 272 | " jmpgt.w.f 1b \n\t" |
| 273 | |
| 274 | : "+a" (tmp), "+a" (s), "+d" (n) |
| 275 | : |
| 276 | : "d15", "a3", "memory", "cc" |
| 277 | ); |
| 278 | } else { |
| 279 | tmp += n; |
| 280 | s += n; |
| 281 | asm volatile ( |
| 282 | " sub.4 d15, #0, %2 \n\t" // set up for jump table |
| 283 | " and.4 d15, #(32-2), d15 \n\t" // d15 = (- count) & (32 - 2) |
| 284 | " moveai a3, #%%hi(1f) \n\t" |
| 285 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 286 | " lea.2 a3, (a3,d15) \n\t" |
| 287 | " calli a3, 0(a3) \n\t" |
| 288 | |
| 289 | "1: .rept 16 \n\t" |
| 290 | " move.2 -2(%0)++, -2(%1)++ \n\t" |
| 291 | " .endr \n\t" |
| 292 | " add.4 %2, #-32, %2 \n\t" |
| 293 | " jmpgt.w.f 1b \n\t" |
| 294 | |
| 295 | : "+a" (tmp), "+a" (s), "+d" (n) |
| 296 | : |
| 297 | : "d15", "a3", "memory", "cc" |
| 298 | ); |
| 299 | } |
| 300 | return dest; |
| 301 | } |
| 302 | |
| 303 | if (dest <= src) { |
| 304 | asm volatile ( |
| 305 | " sub.4 d15, #0, %2 \n\t" // set up for jump table |
| 306 | " and.4 d15, #(16-1), d15 \n\t" // d15 = (- count) & (16 - 1) |
| 307 | " moveai a3, #%%hi(1f) \n\t" |
| 308 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 309 | " lea.4 a3, (a3,d15) \n\t" |
| 310 | " calli a3, 0(a3) \n\t" |
| 311 | |
| 312 | "1: .rept 16 \n\t" |
| 313 | " move.1 (%0)1++, (%1)1++ \n\t" |
| 314 | " .endr \n\t" |
| 315 | " add.4 %2, #-16, %2 \n\t" |
| 316 | " jmpgt.w.f 1b \n\t" |
| 317 | : "+a" (tmp), "+a" (s), "+d" (n) |
| 318 | : |
| 319 | : "d15", "a3", "memory", "cc" |
| 320 | ); |
| 321 | } else { |
| 322 | tmp += n; |
| 323 | s += n; |
| 324 | asm volatile ( |
| 325 | " sub.4 d15, #0, %2 \n\t" // set up for jump table |
| 326 | " and.4 d15, #(16-1), d15 \n\t" // d15 = (- count) & (16 - 1) |
| 327 | " moveai a3, #%%hi(1f) \n\t" |
| 328 | " lea.1 a3, %%lo(1f)(a3) \n\t" |
| 329 | " lea.4 a3, (a3,d15) \n\t" |
| 330 | " calli a3, 0(a3) \n\t" |
| 331 | |
| 332 | "1: .rept 16 \n\t" |
| 333 | " move.1 -1(%0)++, -1(%1)++ \n\t" |
| 334 | " .endr \n\t" |
| 335 | " add.4 %2, #-16, %2 \n\t" |
| 336 | " jmpgt.w.f 1b \n\t" |
| 337 | : "+a" (tmp), "+a" (s), "+d" (n) |
| 338 | : |
| 339 | : "d15", "a3", "memory", "cc" |
| 340 | ); |
| 341 | } |
| 342 | return dest; |
| 343 | } |
| 344 | |
