bzero.S revision 1.3.2.3
1 1.3.2.3 nathanw /* $NetBSD: bzero.S,v 1.3.2.3 2002/03/22 20:41:54 nathanw Exp $ */ 2 1.3.2.2 nathanw 3 1.3.2.2 nathanw /*- 4 1.3.2.2 nathanw * Copyright (C) 2001 Martin J. Laubach <mjl (at) netbsd.org> 5 1.3.2.2 nathanw * All rights reserved. 6 1.3.2.2 nathanw * 7 1.3.2.2 nathanw * Redistribution and use in source and binary forms, with or without 8 1.3.2.2 nathanw * modification, are permitted provided that the following conditions 9 1.3.2.2 nathanw * are met: 10 1.3.2.2 nathanw * 1. Redistributions of source code must retain the above copyright 11 1.3.2.2 nathanw * notice, this list of conditions and the following disclaimer. 12 1.3.2.2 nathanw * 2. Redistributions in binary form must reproduce the above copyright 13 1.3.2.2 nathanw * notice, this list of conditions and the following disclaimer in the 14 1.3.2.2 nathanw * documentation and/or other materials provided with the distribution. 15 1.3.2.2 nathanw * 3. The name of the author may not be used to endorse or promote products 16 1.3.2.2 nathanw * derived from this software without specific prior written permission. 17 1.3.2.2 nathanw * 18 1.3.2.2 nathanw * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 19 1.3.2.2 nathanw * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 20 1.3.2.2 nathanw * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 21 1.3.2.2 nathanw * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 22 1.3.2.2 nathanw * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 23 1.3.2.2 nathanw * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 24 1.3.2.2 nathanw * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 25 1.3.2.2 nathanw * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 26 1.3.2.2 nathanw * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 27 1.3.2.2 nathanw * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 28 1.3.2.2 nathanw */ 29 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 30 1.3.2.2 nathanw 31 1.3.2.2 nathanw #include <machine/asm.h> 32 1.3.2.2 nathanw #ifdef _KERNEL 33 1.3.2.2 nathanw #include <assym.h> 34 1.3.2.2 nathanw #endif 35 1.3.2.2 nathanw 36 1.3.2.2 nathanw #define USE_STSWX 0 /* don't. slower than trivial copy loop */ 37 1.3.2.2 nathanw 38 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 39 1.3.2.2 nathanw /* 40 1.3.2.2 nathanw void bzero(void *b r3, size_t len r4); 41 1.3.2.2 nathanw void * memset(void *b r3, int c r4, size_t len r5); 42 1.3.2.2 nathanw */ 43 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 44 1.3.2.2 nathanw 45 1.3.2.2 nathanw #define r_dst r3 46 1.3.2.2 nathanw #define r_len r4 47 1.3.2.2 nathanw #define r_val r0 48 1.3.2.2 nathanw 49 1.3.2.2 nathanw .text 50 1.3.2.2 nathanw .align 4 51 1.3.2.2 nathanw ENTRY(bzero) 52 1.3.2.2 nathanw li r_val, 0 /* Value to stuff in */ 53 1.3.2.2 nathanw b cb_memset 54 1.3.2.2 nathanw 55 1.3.2.2 nathanw ENTRY(memset) 56 1.3.2.2 nathanw cmplwi cr1, r5, 0 57 1.3.2.2 nathanw mr. r0, r4 58 1.3.2.2 nathanw mr r8, r3 59 1.3.2.2 nathanw beqlr- cr1 /* Nothing to do */ 60 1.3.2.2 nathanw 61 1.3.2.2 nathanw rlwimi r0, r4, 8, 16, 23 /* word extend fill value */ 62 1.3.2.2 nathanw rlwimi r0, r0, 16, 0, 15 63 1.3.2.2 nathanw mr r4, r5 64 1.3.2.2 nathanw bne- simple_fill /* =! 0, use trivial fill */ 65 1.3.2.2 nathanw cb_memset: 66 1.3.2.2 nathanw 67 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 68 1.3.2.2 nathanw #ifndef _KERNEL 69 1.3.2.2 nathanw /* First find out cache line size */ 70 1.3.2.2 nathanw #ifdef PIC 71 1.3.2.2 nathanw mflr r9 72 1.3.2.2 nathanw bl _GLOBAL_OFFSET_TABLE_@local-4 73 1.3.2.2 nathanw mflr r10 74 1.3.2.2 nathanw mtlr r9 75 1.3.2.3 nathanw lwz r5,cache_info@got(r10) 76 1.3.2.2 nathanw #else 77 1.3.2.3 nathanw lis r5,cache_info@h 78 1.3.2.3 nathanw ori r5,r5,cache_info@l 79 1.3.2.2 nathanw #endif 80 1.3.2.3 nathanw lwz r6, 4(r5) 81 1.3.2.2 nathanw cmpwi r6, -1 82 1.3.2.2 nathanw bne+ cb_cacheline_known 83 1.3.2.2 nathanw 84 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 85 1.3.2.2 nathanw #define CTL_MACHDEP 7 86 1.3.2.2 nathanw #define CPU_CACHELINE 1 87 1.3.2.3 nathanw #define CPU_CACHEINFO 5 88 1.3.2.2 nathanw 89 1.3.2.2 nathanw #define STKFRAME_SZ 48 90 1.3.2.2 nathanw #define MIB 8 91 1.3.2.2 nathanw #define OLDPLEN 16 92 1.3.2.2 nathanw #define R3_SAVE 20 93 1.3.2.2 nathanw #define R4_SAVE 24 94 1.3.2.2 nathanw #define R0_SAVE 28 95 1.3.2.2 nathanw #define R8_SAVE 32 96 1.3.2.2 nathanw 97 1.3.2.2 nathanw mflr r6 98 1.3.2.2 nathanw stw r6, 4(r1) 99 1.3.2.2 nathanw stwu r1, -STKFRAME_SZ(r1) 100 1.3.2.2 nathanw 101 1.3.2.2 nathanw stw r8, R8_SAVE(r1) 102 1.3.2.2 nathanw stw r3, R3_SAVE(r1) 103 1.3.2.2 nathanw stw r4, R4_SAVE(r1) 104 1.3.2.2 nathanw stw r0, R0_SAVE(r1) 105 1.3.2.2 nathanw 106 1.3.2.3 nathanw 107 1.3.2.3 nathanw 108 1.3.2.3 nathanw li r0, CTL_MACHDEP /* Construct MIB */ 109 1.3.2.3 nathanw stw r0, MIB(r1) 110 1.3.2.3 nathanw li r0, CPU_CACHEINFO 111 1.3.2.3 nathanw stw r0, MIB+4(r1) 112 1.3.2.3 nathanw 113 1.3.2.3 nathanw li r0, 4*4 /* Oldlenp := 4*4 */ 114 1.3.2.3 nathanw stw r0, OLDPLEN(r1) 115 1.3.2.3 nathanw 116 1.3.2.3 nathanw addi r3, r1, MIB 117 1.3.2.3 nathanw li r4, 2 /* namelen */ 118 1.3.2.3 nathanw /* r5 already contains &cache_info */ 119 1.3.2.3 nathanw addi r6, r1, OLDPLEN 120 1.3.2.3 nathanw li r7, 0 121 1.3.2.3 nathanw li r8, 0 122 1.3.2.3 nathanw bl PIC_PLT(_C_LABEL(sysctl)) 123 1.3.2.3 nathanw 124 1.3.2.3 nathanw cmpwi r3, 0 /* Check result */ 125 1.3.2.3 nathanw beq 1f 126 1.3.2.3 nathanw 127 1.3.2.3 nathanw /* Failure, try older sysctl */ 128 1.3.2.3 nathanw 129 1.3.2.2 nathanw li r0, CTL_MACHDEP /* Construct MIB */ 130 1.3.2.2 nathanw stw r0, MIB(r1) 131 1.3.2.2 nathanw li r0, CPU_CACHELINE 132 1.3.2.2 nathanw stw r0, MIB+4(r1) 133 1.3.2.2 nathanw 134 1.3.2.2 nathanw li r0, 4 /* Oldlenp := 4 */ 135 1.3.2.2 nathanw stw r0, OLDPLEN(r1) 136 1.3.2.2 nathanw 137 1.3.2.2 nathanw addi r3, r1, MIB 138 1.3.2.2 nathanw li r4, 2 /* namelen */ 139 1.3.2.3 nathanw #ifdef PIC 140 1.3.2.3 nathanw mflr r9 141 1.3.2.3 nathanw bl _GLOBAL_OFFSET_TABLE_@local-4 142 1.3.2.3 nathanw mflr r10 143 1.3.2.3 nathanw mtlr r9 144 1.3.2.3 nathanw lwz r5,cache_info@got(r10) 145 1.3.2.3 nathanw addi r5, r5, 4 146 1.3.2.3 nathanw #else 147 1.3.2.3 nathanw lis r5,cache_info+4@h 148 1.3.2.3 nathanw ori r5,r5,cache_info+4@l 149 1.3.2.3 nathanw #endif 150 1.3.2.2 nathanw addi r6, r1, OLDPLEN 151 1.3.2.2 nathanw li r7, 0 152 1.3.2.2 nathanw li r8, 0 153 1.3.2.2 nathanw bl PIC_PLT(_C_LABEL(sysctl)) 154 1.3.2.3 nathanw 1: 155 1.3.2.2 nathanw lwz r8, R8_SAVE(r1) 156 1.3.2.2 nathanw lwz r3, R3_SAVE(r1) 157 1.3.2.2 nathanw lwz r4, R4_SAVE(r1) 158 1.3.2.2 nathanw lwz r0, R0_SAVE(r1) 159 1.3.2.2 nathanw 160 1.3.2.2 nathanw #ifdef PIC 161 1.3.2.2 nathanw bl _GLOBAL_OFFSET_TABLE_@local-4 162 1.3.2.2 nathanw mflr r10 163 1.3.2.3 nathanw lwz r9, cache_info@got(r10) 164 1.3.2.3 nathanw lwz r9, 4(r9) 165 1.3.2.2 nathanw #else 166 1.3.2.3 nathanw lis r5, cache_info+4@ha 167 1.3.2.3 nathanw lwz r9, cache_info+4@l(r5) 168 1.3.2.2 nathanw #endif 169 1.3.2.2 nathanw la r1, STKFRAME_SZ(r1) 170 1.3.2.2 nathanw lwz r5, 4(r1) 171 1.3.2.2 nathanw mtlr r5 172 1.3.2.2 nathanw 173 1.3.2.2 nathanw cntlzw r6, r9 /* compute shift value */ 174 1.3.2.2 nathanw li r5, 31 175 1.3.2.2 nathanw subf r5, r6, r5 176 1.3.2.2 nathanw 177 1.3.2.2 nathanw #ifdef PIC 178 1.3.2.2 nathanw lwz r6, cache_sh@got(r10) 179 1.3.2.2 nathanw stw r5, 0(r6) 180 1.3.2.2 nathanw #else 181 1.3.2.2 nathanw lis r6, cache_sh@ha 182 1.3.2.2 nathanw stw r5, cache_sh@l(r6) 183 1.3.2.2 nathanw #endif 184 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 185 1.3.2.2 nathanw /* Okay, we know the cache line size (r9) and shift value (r10) */ 186 1.3.2.2 nathanw cb_cacheline_known: 187 1.3.2.2 nathanw #ifdef PIC 188 1.3.2.3 nathanw lwz r5, cache_info@got(r10) 189 1.3.2.3 nathanw lwz r9, 4(r5) 190 1.3.2.2 nathanw lwz r5, cache_sh@got(r10) 191 1.3.2.2 nathanw lwz r10, 0(r5) 192 1.3.2.2 nathanw #else 193 1.3.2.3 nathanw lis r9, cache_info+4@ha 194 1.3.2.3 nathanw lwz r9, cache_info+4@l(r9) 195 1.3.2.2 nathanw lis r10, cache_sh@ha 196 1.3.2.2 nathanw lwz r10, cache_sh@l(r10) 197 1.3.2.2 nathanw #endif 198 1.3.2.2 nathanw 199 1.3.2.2 nathanw #else /* _KERNEL */ 200 1.3.2.3 nathanw #ifdef MULTIPROCESSOR 201 1.3.2.3 nathanw mfspr r10, 0 /* Get cpu_info pointer */ 202 1.3.2.2 nathanw #else 203 1.3.2.3 nathanw lis r10, cpu_info_store@ha 204 1.3.2.3 nathanw addi r10, r10, cpu_info_store@l 205 1.3.2.2 nathanw #endif 206 1.3.2.3 nathanw lwz r9, CPU_CI+4(r10) /* Load D$ line size */ 207 1.3.2.3 nathanw cntlzw r10, r9 /* Calculate shift.. */ 208 1.3.2.3 nathanw li r6, 31 209 1.3.2.3 nathanw subf r10, r10, r6 210 1.3.2.2 nathanw #endif /* _KERNEL */ 211 1.3.2.2 nathanw /* Back in memory filling business */ 212 1.3.2.2 nathanw 213 1.3.2.2 nathanw cmplwi cr1, r_len, 0 /* Nothing to do? */ 214 1.3.2.2 nathanw add r5, r9, r9 215 1.3.2.2 nathanw cmplw r_len, r5 /* <= 2*CL bytes to move? */ 216 1.3.2.2 nathanw beqlr- cr1 /* then do nothing */ 217 1.3.2.2 nathanw 218 1.3.2.2 nathanw blt+ simple_fill /* a trivial fill routine */ 219 1.3.2.2 nathanw 220 1.3.2.2 nathanw /* Word align the block, fill bytewise until dst even*/ 221 1.3.2.2 nathanw 222 1.3.2.2 nathanw andi. r5, r_dst, 0x03 223 1.3.2.2 nathanw li r6, 4 224 1.3.2.2 nathanw beq+ cb_aligned_w /* already aligned to word? */ 225 1.3.2.2 nathanw 226 1.3.2.2 nathanw subf r5, r5, r6 /* bytes to fill to align4 */ 227 1.3.2.2 nathanw #if USE_STSWX 228 1.3.2.2 nathanw mtxer r5 229 1.3.2.2 nathanw stswx r0, 0, r_dst 230 1.3.2.2 nathanw add r_dst, r5, r_dst 231 1.3.2.2 nathanw #else 232 1.3.2.2 nathanw mtctr r5 233 1.3.2.2 nathanw 234 1.3.2.2 nathanw subi r_dst, r_dst, 1 235 1.3.2.2 nathanw 1: stbu r_val, 1(r_dst) /* Fill bytewise */ 236 1.3.2.2 nathanw bdnz 1b 237 1.3.2.2 nathanw 238 1.3.2.2 nathanw addi r_dst, r_dst, 1 239 1.3.2.2 nathanw #endif 240 1.3.2.2 nathanw subf r_len, r5, r_len 241 1.3.2.2 nathanw 242 1.3.2.2 nathanw cb_aligned_w: /* Cache block align, fill wordwise until dst aligned */ 243 1.3.2.2 nathanw 244 1.3.2.2 nathanw /* I know I have something to do since we had > 2*CL initially */ 245 1.3.2.2 nathanw /* so no need to check for r_len = 0 */ 246 1.3.2.2 nathanw 247 1.3.2.2 nathanw rlwinm. r5, r_dst, 30, 29, 31 248 1.3.2.2 nathanw srwi r6, r9, 2 249 1.3.2.2 nathanw beq cb_aligned_cb /* already on CL boundary? */ 250 1.3.2.2 nathanw 251 1.3.2.2 nathanw subf r5, r5, r6 /* words to fill to alignment */ 252 1.3.2.2 nathanw mtctr r5 253 1.3.2.2 nathanw slwi r5, r5, 2 254 1.3.2.2 nathanw subf r_len, r5, r_len 255 1.3.2.2 nathanw 256 1.3.2.2 nathanw subi r_dst, r_dst, 4 257 1.3.2.2 nathanw 1: stwu r_val, 4(r_dst) /* Fill wordwise */ 258 1.3.2.2 nathanw bdnz 1b 259 1.3.2.2 nathanw addi r_dst, r_dst, 4 260 1.3.2.2 nathanw 261 1.3.2.2 nathanw cb_aligned_cb: /* no need to check r_len, see above */ 262 1.3.2.2 nathanw 263 1.3.2.2 nathanw srw. r5, r_len, r10 /* Number of cache blocks */ 264 1.3.2.2 nathanw mtctr r5 265 1.3.2.2 nathanw beq cblocks_done 266 1.3.2.2 nathanw 267 1.3.2.2 nathanw slw r5, r5, r10 268 1.3.2.2 nathanw subf r_len, r5, r_len 269 1.3.2.2 nathanw 270 1.3.2.2 nathanw 1: dcbz 0, r_dst /* Clear blockwise */ 271 1.3.2.2 nathanw add r_dst, r_dst, r9 272 1.3.2.2 nathanw bdnz 1b 273 1.3.2.2 nathanw 274 1.3.2.2 nathanw cblocks_done: /* still CL aligned, but less than CL bytes left */ 275 1.3.2.2 nathanw cmplwi cr1, r_len, 0 276 1.3.2.2 nathanw cmplwi r_len, 8 277 1.3.2.2 nathanw beq- cr1, sf_return 278 1.3.2.2 nathanw 279 1.3.2.2 nathanw blt- sf_bytewise /* <8 remaining? */ 280 1.3.2.2 nathanw b sf_aligned_w 281 1.3.2.2 nathanw 282 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 283 1.3.2.2 nathanw wbzero: li r_val, 0 284 1.3.2.2 nathanw 285 1.3.2.2 nathanw cmplwi r_len, 0 286 1.3.2.2 nathanw beqlr- /* Nothing to do */ 287 1.3.2.2 nathanw 288 1.3.2.2 nathanw simple_fill: 289 1.3.2.2 nathanw #if USE_STSWX 290 1.3.2.2 nathanw cmplwi cr1, r_len, 12 /* < 12 bytes to move? */ 291 1.3.2.2 nathanw #else 292 1.3.2.2 nathanw cmplwi cr1, r_len, 8 /* < 8 bytes to move? */ 293 1.3.2.2 nathanw #endif 294 1.3.2.2 nathanw andi. r5, r_dst, 0x03 /* bytes to fill to align4 */ 295 1.3.2.2 nathanw blt cr1, sf_bytewise /* trivial byte mover */ 296 1.3.2.2 nathanw 297 1.3.2.2 nathanw li r6, 4 298 1.3.2.2 nathanw subf r5, r5, r6 299 1.3.2.2 nathanw beq+ sf_aligned_w /* dest is word aligned */ 300 1.3.2.2 nathanw 301 1.3.2.2 nathanw #if USE_STSWX 302 1.3.2.2 nathanw mtxer r5 303 1.3.2.2 nathanw stswx r0, 0, r_dst 304 1.3.2.2 nathanw add r_dst, r5, r_dst 305 1.3.2.2 nathanw #else 306 1.3.2.2 nathanw mtctr r5 /* nope, then fill bytewise */ 307 1.3.2.2 nathanw subi r_dst, r_dst, 1 /* until it is */ 308 1.3.2.2 nathanw 1: stbu r_val, 1(r_dst) 309 1.3.2.2 nathanw bdnz 1b 310 1.3.2.2 nathanw 311 1.3.2.2 nathanw addi r_dst, r_dst, 1 312 1.3.2.2 nathanw #endif 313 1.3.2.2 nathanw subf r_len, r5, r_len 314 1.3.2.2 nathanw 315 1.3.2.2 nathanw sf_aligned_w: /* no need to check r_len since it were >= 8 bytes initially */ 316 1.3.2.2 nathanw #if USE_STSWX 317 1.3.2.2 nathanw mr r6, r0 318 1.3.2.2 nathanw mr r7, r0 319 1.3.2.2 nathanw 320 1.3.2.2 nathanw srwi r5, r_len, 3 321 1.3.2.2 nathanw mtctr r5 322 1.3.2.2 nathanw 323 1.3.2.2 nathanw slwi r5, r5, 3 /* adjust len */ 324 1.3.2.2 nathanw subf. r_len, r5, r_len 325 1.3.2.2 nathanw 326 1.3.2.2 nathanw 1: stswi r6, r_dst, 8 327 1.3.2.2 nathanw addi r_dst, r_dst, 8 328 1.3.2.2 nathanw bdnz 1b 329 1.3.2.2 nathanw #else 330 1.3.2.2 nathanw srwi r5, r_len, 2 /* words to fill */ 331 1.3.2.2 nathanw mtctr r5 332 1.3.2.2 nathanw 333 1.3.2.2 nathanw slwi r5, r5, 2 334 1.3.2.2 nathanw subf. r_len, r5, r_len /* adjust len for fill */ 335 1.3.2.2 nathanw 336 1.3.2.2 nathanw subi r_dst, r_dst, 4 337 1.3.2.2 nathanw 1: stwu r_val, 4(r_dst) 338 1.3.2.2 nathanw bdnz 1b 339 1.3.2.2 nathanw addi r_dst, r_dst, 4 340 1.3.2.2 nathanw #endif 341 1.3.2.2 nathanw 342 1.3.2.2 nathanw sf_word_done: bne- sf_bytewise 343 1.3.2.2 nathanw 344 1.3.2.2 nathanw sf_return: mr r3, r8 /* restore orig ptr */ 345 1.3.2.2 nathanw blr /* for memset functionality */ 346 1.3.2.2 nathanw 347 1.3.2.2 nathanw sf_bytewise: 348 1.3.2.2 nathanw #if USE_STSWX 349 1.3.2.2 nathanw mr r5, r0 350 1.3.2.2 nathanw mr r6, r0 351 1.3.2.2 nathanw mr r7, r0 352 1.3.2.2 nathanw 353 1.3.2.2 nathanw mtxer r_len 354 1.3.2.2 nathanw stswx r5, 0, r_dst 355 1.3.2.2 nathanw #else 356 1.3.2.2 nathanw mtctr r_len 357 1.3.2.2 nathanw 358 1.3.2.2 nathanw subi r_dst, r_dst, 1 359 1.3.2.2 nathanw 1: stbu r_val, 1(r_dst) 360 1.3.2.2 nathanw bdnz 1b 361 1.3.2.2 nathanw #endif 362 1.3.2.2 nathanw mr r3, r8 /* restore orig ptr */ 363 1.3.2.2 nathanw blr /* for memset functionality */ 364 1.3.2.2 nathanw 365 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 366 1.3.2.2 nathanw #ifndef _KERNEL 367 1.3.2.2 nathanw .data 368 1.3.2.3 nathanw cache_info: .long -1, -1, -1, -1 369 1.3.2.2 nathanw cache_sh: .long 0 370 1.3.2.2 nathanw 371 1.3.2.2 nathanw #endif 372 1.3.2.2 nathanw /*----------------------------------------------------------------------*/ 373
Indexes created Fri Sep 26 20:09:58 GMT 2025