include/asm-x86/xor_32.h

   1 /*
   2  * include/asm-i386/xor.h
   3  *
   4  * Optimized RAID-5 checksumming functions for MMX and SSE.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2, or (at your option)
   9  * any later version.
  10  *
  11  * You should have received a copy of the GNU General Public License
  12  * (for example /usr/src/linux/COPYING); if not, write to the Free
  13  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  14  */
  15
  16 /*
  17  * High-speed RAID5 checksumming functions utilizing MMX instructions.
  18  * Copyright (C) 1998 Ingo Molnar.
  19  */
  20
  21 #define LD(x,y)         "       movq   8*("#x")(%1), %%mm"#y"   ;\n"
  22 #define ST(x,y)         "       movq %%mm"#y",   8*("#x")(%1)   ;\n"
  23 #define XO1(x,y)        "       pxor   8*("#x")(%2), %%mm"#y"   ;\n"
  24 #define XO2(x,y)        "       pxor   8*("#x")(%3), %%mm"#y"   ;\n"
  25 #define XO3(x,y)        "       pxor   8*("#x")(%4), %%mm"#y"   ;\n"
  26 #define XO4(x,y)        "       pxor   8*("#x")(%5), %%mm"#y"   ;\n"
  27
  28 #include <asm/i387.h>
  29
  30 static void
  31 xor_pII_mmx_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
  32 {
  33         unsigned long lines = bytes >> 7;
  34
  35         kernel_fpu_begin();
  36
  37         __asm__ __volatile__ (
  38 #undef BLOCK
  39 #define BLOCK(i) \
  40         LD(i,0)                                 \
  41                 LD(i+1,1)                       \
  42                         LD(i+2,2)               \
  43                                 LD(i+3,3)       \
  44         XO1(i,0)                                \
  45         ST(i,0)                                 \
  46                 XO1(i+1,1)                      \
  47                 ST(i+1,1)                       \
  48                         XO1(i+2,2)              \
  49                         ST(i+2,2)               \
  50                                 XO1(i+3,3)      \
  51                                 ST(i+3,3)
  52
  53         " .align 32                     ;\n"
  54         " 1:                            ;\n"
  55
  56         BLOCK(0)
  57         BLOCK(4)
  58         BLOCK(8)
  59         BLOCK(12)
  60
  61         "       addl $128, %1         ;\n"
  62         "       addl $128, %2         ;\n"
  63         "       decl %0               ;\n"
  64         "       jnz 1b                ;\n"
  65         : "+r" (lines),
  66           "+r" (p1), "+r" (p2)
  67         :
  68         : "memory");
  69
  70         kernel_fpu_end();
  71 }
  72
  73 static void
  74 xor_pII_mmx_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
  75               unsigned long *p3)
  76 {
  77         unsigned long lines = bytes >> 7;
  78
  79         kernel_fpu_begin();
  80
  81         __asm__ __volatile__ (
  82 #undef BLOCK
  83 #define BLOCK(i) \
  84         LD(i,0)                                 \
  85                 LD(i+1,1)                       \
  86                         LD(i+2,2)               \
  87                                 LD(i+3,3)       \
  88         XO1(i,0)                                \
  89                 XO1(i+1,1)                      \
  90                         XO1(i+2,2)              \
  91                                 XO1(i+3,3)      \
  92         XO2(i,0)                                \
  93         ST(i,0)                                 \
  94                 XO2(i+1,1)                      \
  95                 ST(i+1,1)                       \
  96                         XO2(i+2,2)              \
  97                         ST(i+2,2)               \
  98                                 XO2(i+3,3)      \
  99                                 ST(i+3,3)
 100
 101         " .align 32                     ;\n"
 102         " 1:                            ;\n"
 103
 104         BLOCK(0)
 105         BLOCK(4)
 106         BLOCK(8)
 107         BLOCK(12)
 108
 109         "       addl $128, %1         ;\n"
 110         "       addl $128, %2         ;\n"
 111         "       addl $128, %3         ;\n"
 112         "       decl %0               ;\n"
 113         "       jnz 1b                ;\n"
 114         : "+r" (lines),
 115           "+r" (p1), "+r" (p2), "+r" (p3)
 116         :
 117         : "memory");
 118
 119         kernel_fpu_end();
 120 }
 121
 122 static void
 123 xor_pII_mmx_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 124               unsigned long *p3, unsigned long *p4)
 125 {
 126         unsigned long lines = bytes >> 7;
 127
 128         kernel_fpu_begin();
 129
 130         __asm__ __volatile__ (
 131 #undef BLOCK
 132 #define BLOCK(i) \
 133         LD(i,0)                                 \
 134                 LD(i+1,1)                       \
 135                         LD(i+2,2)               \
 136                                 LD(i+3,3)       \
 137         XO1(i,0)                                \
 138                 XO1(i+1,1)                      \
 139                         XO1(i+2,2)              \
 140                                 XO1(i+3,3)      \
 141         XO2(i,0)                                \
 142                 XO2(i+1,1)                      \
 143                         XO2(i+2,2)              \
 144                                 XO2(i+3,3)      \
 145         XO3(i,0)                                \
 146         ST(i,0)                                 \
 147                 XO3(i+1,1)                      \
 148                 ST(i+1,1)                       \
 149                         XO3(i+2,2)              \
 150                         ST(i+2,2)               \
 151                                 XO3(i+3,3)      \
 152                                 ST(i+3,3)
 153
 154         " .align 32                     ;\n"
 155         " 1:                            ;\n"
 156
 157         BLOCK(0)
 158         BLOCK(4)
 159         BLOCK(8)
 160         BLOCK(12)
 161
 162         "       addl $128, %1         ;\n"
 163         "       addl $128, %2         ;\n"
 164         "       addl $128, %3         ;\n"
 165         "       addl $128, %4         ;\n"
 166         "       decl %0               ;\n"
 167         "       jnz 1b                ;\n"
 168         : "+r" (lines),
 169           "+r" (p1), "+r" (p2), "+r" (p3), "+r" (p4)
 170         :
 171         : "memory");
 172
 173         kernel_fpu_end();
 174 }
 175
 176
 177 static void
 178 xor_pII_mmx_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 179               unsigned long *p3, unsigned long *p4, unsigned long *p5)
 180 {
 181         unsigned long lines = bytes >> 7;
 182
 183         kernel_fpu_begin();
 184
 185         /* Make sure GCC forgets anything it knows about p4 or p5,
 186            such that it won't pass to the asm volatile below a
 187            register that is shared with any other variable.  That's
 188            because we modify p4 and p5 there, but we can't mark them
 189            as read/write, otherwise we'd overflow the 10-asm-operands
 190            limit of GCC < 3.1.  */
 191         __asm__ ("" : "+r" (p4), "+r" (p5));
 192
 193         __asm__ __volatile__ (
 194 #undef BLOCK
 195 #define BLOCK(i) \
 196         LD(i,0)                                 \
 197                 LD(i+1,1)                       \
 198                         LD(i+2,2)               \
 199                                 LD(i+3,3)       \
 200         XO1(i,0)                                \
 201                 XO1(i+1,1)                      \
 202                         XO1(i+2,2)              \
 203                                 XO1(i+3,3)      \
 204         XO2(i,0)                                \
 205                 XO2(i+1,1)                      \
 206                         XO2(i+2,2)              \
 207                                 XO2(i+3,3)      \
 208         XO3(i,0)                                \
 209                 XO3(i+1,1)                      \
 210                         XO3(i+2,2)              \
 211                                 XO3(i+3,3)      \
 212         XO4(i,0)                                \
 213         ST(i,0)                                 \
 214                 XO4(i+1,1)                      \
 215                 ST(i+1,1)                       \
 216                         XO4(i+2,2)              \
 217                         ST(i+2,2)               \
 218                                 XO4(i+3,3)      \
 219                                 ST(i+3,3)
 220
 221         " .align 32                     ;\n"
 222         " 1:                            ;\n"
 223
 224         BLOCK(0)
 225         BLOCK(4)
 226         BLOCK(8)
 227         BLOCK(12)
 228
 229         "       addl $128, %1         ;\n"
 230         "       addl $128, %2         ;\n"
 231         "       addl $128, %3         ;\n"
 232         "       addl $128, %4         ;\n"
 233         "       addl $128, %5         ;\n"
 234         "       decl %0               ;\n"
 235         "       jnz 1b                ;\n"
 236         : "+r" (lines),
 237           "+r" (p1), "+r" (p2), "+r" (p3)
 238         : "r" (p4), "r" (p5)
 239         : "memory");
 240
 241         /* p4 and p5 were modified, and now the variables are dead.
 242            Clobber them just to be sure nobody does something stupid
 243            like assuming they have some legal value.  */
 244         __asm__ ("" : "=r" (p4), "=r" (p5));
 245
 246         kernel_fpu_end();
 247 }
 248
 249 #undef LD
 250 #undef XO1
 251 #undef XO2
 252 #undef XO3
 253 #undef XO4
 254 #undef ST
 255 #undef BLOCK
 256
 257 static void
 258 xor_p5_mmx_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
 259 {
 260         unsigned long lines = bytes >> 6;
 261
 262         kernel_fpu_begin();
 263
 264         __asm__ __volatile__ (
 265         " .align 32                  ;\n"
 266         " 1:                         ;\n"
 267         "       movq   (%1), %%mm0   ;\n"
 268         "       movq  8(%1), %%mm1   ;\n"
 269         "       pxor   (%2), %%mm0   ;\n"
 270         "       movq 16(%1), %%mm2   ;\n"
 271         "       movq %%mm0,   (%1)   ;\n"
 272         "       pxor  8(%2), %%mm1   ;\n"
 273         "       movq 24(%1), %%mm3   ;\n"
 274         "       movq %%mm1,  8(%1)   ;\n"
 275         "       pxor 16(%2), %%mm2   ;\n"
 276         "       movq 32(%1), %%mm4   ;\n"
 277         "       movq %%mm2, 16(%1)   ;\n"
 278         "       pxor 24(%2), %%mm3   ;\n"
 279         "       movq 40(%1), %%mm5   ;\n"
 280         "       movq %%mm3, 24(%1)   ;\n"
 281         "       pxor 32(%2), %%mm4   ;\n"
 282         "       movq 48(%1), %%mm6   ;\n"
 283         "       movq %%mm4, 32(%1)   ;\n"
 284         "       pxor 40(%2), %%mm5   ;\n"
 285         "       movq 56(%1), %%mm7   ;\n"
 286         "       movq %%mm5, 40(%1)   ;\n"
 287         "       pxor 48(%2), %%mm6   ;\n"
 288         "       pxor 56(%2), %%mm7   ;\n"
 289         "       movq %%mm6, 48(%1)   ;\n"
 290         "       movq %%mm7, 56(%1)   ;\n"
 291
 292         "       addl $64, %1         ;\n"
 293         "       addl $64, %2         ;\n"
 294         "       decl %0              ;\n"
 295         "       jnz 1b               ;\n"
 296         : "+r" (lines),
 297           "+r" (p1), "+r" (p2)
 298         :
 299         : "memory");
 300
 301         kernel_fpu_end();
 302 }
 303
 304 static void
 305 xor_p5_mmx_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 306              unsigned long *p3)
 307 {
 308         unsigned long lines = bytes >> 6;
 309
 310         kernel_fpu_begin();
 311
 312         __asm__ __volatile__ (
 313         " .align 32,0x90             ;\n"
 314         " 1:                         ;\n"
 315         "       movq   (%1), %%mm0   ;\n"
 316         "       movq  8(%1), %%mm1   ;\n"
 317         "       pxor   (%2), %%mm0   ;\n"
 318         "       movq 16(%1), %%mm2   ;\n"
 319         "       pxor  8(%2), %%mm1   ;\n"
 320         "       pxor   (%3), %%mm0   ;\n"
 321         "       pxor 16(%2), %%mm2   ;\n"
 322         "       movq %%mm0,   (%1)   ;\n"
 323         "       pxor  8(%3), %%mm1   ;\n"
 324         "       pxor 16(%3), %%mm2   ;\n"
 325         "       movq 24(%1), %%mm3   ;\n"
 326         "       movq %%mm1,  8(%1)   ;\n"
 327         "       movq 32(%1), %%mm4   ;\n"
 328         "       movq 40(%1), %%mm5   ;\n"
 329         "       pxor 24(%2), %%mm3   ;\n"
 330         "       movq %%mm2, 16(%1)   ;\n"
 331         "       pxor 32(%2), %%mm4   ;\n"
 332         "       pxor 24(%3), %%mm3   ;\n"
 333         "       pxor 40(%2), %%mm5   ;\n"
 334         "       movq %%mm3, 24(%1)   ;\n"
 335         "       pxor 32(%3), %%mm4   ;\n"
 336         "       pxor 40(%3), %%mm5   ;\n"
 337         "       movq 48(%1), %%mm6   ;\n"
 338         "       movq %%mm4, 32(%1)   ;\n"
 339         "       movq 56(%1), %%mm7   ;\n"
 340         "       pxor 48(%2), %%mm6   ;\n"
 341         "       movq %%mm5, 40(%1)   ;\n"
 342         "       pxor 56(%2), %%mm7   ;\n"
 343         "       pxor 48(%3), %%mm6   ;\n"
 344         "       pxor 56(%3), %%mm7   ;\n"
 345         "       movq %%mm6, 48(%1)   ;\n"
 346         "       movq %%mm7, 56(%1)   ;\n"
 347
 348         "       addl $64, %1         ;\n"
 349         "       addl $64, %2         ;\n"
 350         "       addl $64, %3         ;\n"
 351         "       decl %0              ;\n"
 352         "       jnz 1b               ;\n"
 353         : "+r" (lines),
 354           "+r" (p1), "+r" (p2), "+r" (p3)
 355         :
 356         : "memory" );
 357
 358         kernel_fpu_end();
 359 }
 360
 361 static void
 362 xor_p5_mmx_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 363              unsigned long *p3, unsigned long *p4)
 364 {
 365         unsigned long lines = bytes >> 6;
 366
 367         kernel_fpu_begin();
 368
 369         __asm__ __volatile__ (
 370         " .align 32,0x90             ;\n"
 371         " 1:                         ;\n"
 372         "       movq   (%1), %%mm0   ;\n"
 373         "       movq  8(%1), %%mm1   ;\n"
 374         "       pxor   (%2), %%mm0   ;\n"
 375         "       movq 16(%1), %%mm2   ;\n"
 376         "       pxor  8(%2), %%mm1   ;\n"
 377         "       pxor   (%3), %%mm0   ;\n"
 378         "       pxor 16(%2), %%mm2   ;\n"
 379         "       pxor  8(%3), %%mm1   ;\n"
 380         "       pxor   (%4), %%mm0   ;\n"
 381         "       movq 24(%1), %%mm3   ;\n"
 382         "       pxor 16(%3), %%mm2   ;\n"
 383         "       pxor  8(%4), %%mm1   ;\n"
 384         "       movq %%mm0,   (%1)   ;\n"
 385         "       movq 32(%1), %%mm4   ;\n"
 386         "       pxor 24(%2), %%mm3   ;\n"
 387         "       pxor 16(%4), %%mm2   ;\n"
 388         "       movq %%mm1,  8(%1)   ;\n"
 389         "       movq 40(%1), %%mm5   ;\n"
 390         "       pxor 32(%2), %%mm4   ;\n"
 391         "       pxor 24(%3), %%mm3   ;\n"
 392         "       movq %%mm2, 16(%1)   ;\n"
 393         "       pxor 40(%2), %%mm5   ;\n"
 394         "       pxor 32(%3), %%mm4   ;\n"
 395         "       pxor 24(%4), %%mm3   ;\n"
 396         "       movq %%mm3, 24(%1)   ;\n"
 397         "       movq 56(%1), %%mm7   ;\n"
 398         "       movq 48(%1), %%mm6   ;\n"
 399         "       pxor 40(%3), %%mm5   ;\n"
 400         "       pxor 32(%4), %%mm4   ;\n"
 401         "       pxor 48(%2), %%mm6   ;\n"
 402         "       movq %%mm4, 32(%1)   ;\n"
 403         "       pxor 56(%2), %%mm7   ;\n"
 404         "       pxor 40(%4), %%mm5   ;\n"
 405         "       pxor 48(%3), %%mm6   ;\n"
 406         "       pxor 56(%3), %%mm7   ;\n"
 407         "       movq %%mm5, 40(%1)   ;\n"
 408         "       pxor 48(%4), %%mm6   ;\n"
 409         "       pxor 56(%4), %%mm7   ;\n"
 410         "       movq %%mm6, 48(%1)   ;\n"
 411         "       movq %%mm7, 56(%1)   ;\n"
 412
 413         "       addl $64, %1         ;\n"
 414         "       addl $64, %2         ;\n"
 415         "       addl $64, %3         ;\n"
 416         "       addl $64, %4         ;\n"
 417         "       decl %0              ;\n"
 418         "       jnz 1b               ;\n"
 419         : "+r" (lines),
 420           "+r" (p1), "+r" (p2), "+r" (p3), "+r" (p4)
 421         :
 422         : "memory");
 423
 424         kernel_fpu_end();
 425 }
 426
 427 static void
 428 xor_p5_mmx_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 429              unsigned long *p3, unsigned long *p4, unsigned long *p5)
 430 {
 431         unsigned long lines = bytes >> 6;
 432
 433         kernel_fpu_begin();
 434
 435         /* Make sure GCC forgets anything it knows about p4 or p5,
 436            such that it won't pass to the asm volatile below a
 437            register that is shared with any other variable.  That's
 438            because we modify p4 and p5 there, but we can't mark them
 439            as read/write, otherwise we'd overflow the 10-asm-operands
 440            limit of GCC < 3.1.  */
 441         __asm__ ("" : "+r" (p4), "+r" (p5));
 442
 443         __asm__ __volatile__ (
 444         " .align 32,0x90             ;\n"
 445         " 1:                         ;\n"
 446         "       movq   (%1), %%mm0   ;\n"
 447         "       movq  8(%1), %%mm1   ;\n"
 448         "       pxor   (%2), %%mm0   ;\n"
 449         "       pxor  8(%2), %%mm1   ;\n"
 450         "       movq 16(%1), %%mm2   ;\n"
 451         "       pxor   (%3), %%mm0   ;\n"
 452         "       pxor  8(%3), %%mm1   ;\n"
 453         "       pxor 16(%2), %%mm2   ;\n"
 454         "       pxor   (%4), %%mm0   ;\n"
 455         "       pxor  8(%4), %%mm1   ;\n"
 456         "       pxor 16(%3), %%mm2   ;\n"
 457         "       movq 24(%1), %%mm3   ;\n"
 458         "       pxor   (%5), %%mm0   ;\n"
 459         "       pxor  8(%5), %%mm1   ;\n"
 460         "       movq %%mm0,   (%1)   ;\n"
 461         "       pxor 16(%4), %%mm2   ;\n"
 462         "       pxor 24(%2), %%mm3   ;\n"
 463         "       movq %%mm1,  8(%1)   ;\n"
 464         "       pxor 16(%5), %%mm2   ;\n"
 465         "       pxor 24(%3), %%mm3   ;\n"
 466         "       movq 32(%1), %%mm4   ;\n"
 467         "       movq %%mm2, 16(%1)   ;\n"
 468         "       pxor 24(%4), %%mm3   ;\n"
 469         "       pxor 32(%2), %%mm4   ;\n"
 470         "       movq 40(%1), %%mm5   ;\n"
 471         "       pxor 24(%5), %%mm3   ;\n"
 472         "       pxor 32(%3), %%mm4   ;\n"
 473         "       pxor 40(%2), %%mm5   ;\n"
 474         "       movq %%mm3, 24(%1)   ;\n"
 475         "       pxor 32(%4), %%mm4   ;\n"
 476         "       pxor 40(%3), %%mm5   ;\n"
 477         "       movq 48(%1), %%mm6   ;\n"
 478         "       movq 56(%1), %%mm7   ;\n"
 479         "       pxor 32(%5), %%mm4   ;\n"
 480         "       pxor 40(%4), %%mm5   ;\n"
 481         "       pxor 48(%2), %%mm6   ;\n"
 482         "       pxor 56(%2), %%mm7   ;\n"
 483         "       movq %%mm4, 32(%1)   ;\n"
 484         "       pxor 48(%3), %%mm6   ;\n"
 485         "       pxor 56(%3), %%mm7   ;\n"
 486         "       pxor 40(%5), %%mm5   ;\n"
 487         "       pxor 48(%4), %%mm6   ;\n"
 488         "       pxor 56(%4), %%mm7   ;\n"
 489         "       movq %%mm5, 40(%1)   ;\n"
 490         "       pxor 48(%5), %%mm6   ;\n"
 491         "       pxor 56(%5), %%mm7   ;\n"
 492         "       movq %%mm6, 48(%1)   ;\n"
 493         "       movq %%mm7, 56(%1)   ;\n"
 494
 495         "       addl $64, %1         ;\n"
 496         "       addl $64, %2         ;\n"
 497         "       addl $64, %3         ;\n"
 498         "       addl $64, %4         ;\n"
 499         "       addl $64, %5         ;\n"
 500         "       decl %0              ;\n"
 501         "       jnz 1b               ;\n"
 502         : "+r" (lines),
 503           "+r" (p1), "+r" (p2), "+r" (p3)
 504         : "r" (p4), "r" (p5)
 505         : "memory");
 506
 507         /* p4 and p5 were modified, and now the variables are dead.
 508            Clobber them just to be sure nobody does something stupid
 509            like assuming they have some legal value.  */
 510         __asm__ ("" : "=r" (p4), "=r" (p5));
 511
 512         kernel_fpu_end();
 513 }
 514
 515 static struct xor_block_template xor_block_pII_mmx = {
 516         .name = "pII_mmx",
 517         .do_2 = xor_pII_mmx_2,
 518         .do_3 = xor_pII_mmx_3,
 519         .do_4 = xor_pII_mmx_4,
 520         .do_5 = xor_pII_mmx_5,
 521 };
 522
 523 static struct xor_block_template xor_block_p5_mmx = {
 524         .name = "p5_mmx",
 525         .do_2 = xor_p5_mmx_2,
 526         .do_3 = xor_p5_mmx_3,
 527         .do_4 = xor_p5_mmx_4,
 528         .do_5 = xor_p5_mmx_5,
 529 };
 530
 531 /*
 532  * Cache avoiding checksumming functions utilizing KNI instructions
 533  * Copyright (C) 1999 Zach Brown (with obvious credit due Ingo)
 534  */
 535
 536 #define XMMS_SAVE do {                          \
 537         preempt_disable();                      \
 538         cr0 = read_cr0();                       \
 539         clts();                                 \
 540         __asm__ __volatile__ (                  \
 541                 "movups %%xmm0,(%0)     ;\n\t"  \
 542                 "movups %%xmm1,0x10(%0) ;\n\t"  \
 543                 "movups %%xmm2,0x20(%0) ;\n\t"  \
 544                 "movups %%xmm3,0x30(%0) ;\n\t"  \
 545                 :                               \
 546                 : "r" (xmm_save)                \
 547                 : "memory");                    \
 548 } while(0)
 549
 550 #define XMMS_RESTORE do {                       \
 551         __asm__ __volatile__ (                  \
 552                 "sfence                 ;\n\t"  \
 553                 "movups (%0),%%xmm0     ;\n\t"  \
 554                 "movups 0x10(%0),%%xmm1 ;\n\t"  \
 555                 "movups 0x20(%0),%%xmm2 ;\n\t"  \
 556                 "movups 0x30(%0),%%xmm3 ;\n\t"  \
 557                 :                               \
 558                 : "r" (xmm_save)                \
 559                 : "memory");                    \
 560         write_cr0(cr0);                         \
 561         preempt_enable();                       \
 562 } while(0)
 563
 564 #define ALIGN16 __attribute__((aligned(16)))
 565
 566 #define OFFS(x)         "16*("#x")"
 567 #define PF_OFFS(x)      "256+16*("#x")"
 568 #define PF0(x)          "       prefetchnta "PF_OFFS(x)"(%1)            ;\n"
 569 #define LD(x,y)         "       movaps   "OFFS(x)"(%1), %%xmm"#y"       ;\n"
 570 #define ST(x,y)         "       movaps %%xmm"#y",   "OFFS(x)"(%1)       ;\n"
 571 #define PF1(x)          "       prefetchnta "PF_OFFS(x)"(%2)            ;\n"
 572 #define PF2(x)          "       prefetchnta "PF_OFFS(x)"(%3)            ;\n"
 573 #define PF3(x)          "       prefetchnta "PF_OFFS(x)"(%4)            ;\n"
 574 #define PF4(x)          "       prefetchnta "PF_OFFS(x)"(%5)            ;\n"
 575 #define PF5(x)          "       prefetchnta "PF_OFFS(x)"(%6)            ;\n"
 576 #define XO1(x,y)        "       xorps   "OFFS(x)"(%2), %%xmm"#y"        ;\n"
 577 #define XO2(x,y)        "       xorps   "OFFS(x)"(%3), %%xmm"#y"        ;\n"
 578 #define XO3(x,y)        "       xorps   "OFFS(x)"(%4), %%xmm"#y"        ;\n"
 579 #define XO4(x,y)        "       xorps   "OFFS(x)"(%5), %%xmm"#y"        ;\n"
 580 #define XO5(x,y)        "       xorps   "OFFS(x)"(%6), %%xmm"#y"        ;\n"
 581
 582
 583 static void
 584 xor_sse_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
 585 {
 586         unsigned long lines = bytes >> 8;
 587         char xmm_save[16*4] ALIGN16;
 588         int cr0;
 589
 590         XMMS_SAVE;
 591
 592         __asm__ __volatile__ (
 593 #undef BLOCK
 594 #define BLOCK(i) \
 595                 LD(i,0)                                 \
 596                         LD(i+1,1)                       \
 597                 PF1(i)                                  \
 598                                 PF1(i+2)                \
 599                                 LD(i+2,2)               \
 600                                         LD(i+3,3)       \
 601                 PF0(i+4)                                \
 602                                 PF0(i+6)                \
 603                 XO1(i,0)                                \
 604                         XO1(i+1,1)                      \
 605                                 XO1(i+2,2)              \
 606                                         XO1(i+3,3)      \
 607                 ST(i,0)                                 \
 608                         ST(i+1,1)                       \
 609                                 ST(i+2,2)               \
 610                                         ST(i+3,3)       \
 611
 612
 613                 PF0(0)
 614                                 PF0(2)
 615
 616         " .align 32                     ;\n"
 617         " 1:                            ;\n"
 618
 619                 BLOCK(0)
 620                 BLOCK(4)
 621                 BLOCK(8)
 622                 BLOCK(12)
 623
 624         "       addl $256, %1           ;\n"
 625         "       addl $256, %2           ;\n"
 626         "       decl %0                 ;\n"
 627         "       jnz 1b                  ;\n"
 628         : "+r" (lines),
 629           "+r" (p1), "+r" (p2)
 630         :
 631         : "memory");
 632
 633         XMMS_RESTORE;
 634 }
 635
 636 static void
 637 xor_sse_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 638           unsigned long *p3)
 639 {
 640         unsigned long lines = bytes >> 8;
 641         char xmm_save[16*4] ALIGN16;
 642         int cr0;
 643
 644         XMMS_SAVE;
 645
 646         __asm__ __volatile__ (
 647 #undef BLOCK
 648 #define BLOCK(i) \
 649                 PF1(i)                                  \
 650                                 PF1(i+2)                \
 651                 LD(i,0)                                 \
 652                         LD(i+1,1)                       \
 653                                 LD(i+2,2)               \
 654                                         LD(i+3,3)       \
 655                 PF2(i)                                  \
 656                                 PF2(i+2)                \
 657                 PF0(i+4)                                \
 658                                 PF0(i+6)                \
 659                 XO1(i,0)                                \
 660                         XO1(i+1,1)                      \
 661                                 XO1(i+2,2)              \
 662                                         XO1(i+3,3)      \
 663                 XO2(i,0)                                \
 664                         XO2(i+1,1)                      \
 665                                 XO2(i+2,2)              \
 666                                         XO2(i+3,3)      \
 667                 ST(i,0)                                 \
 668                         ST(i+1,1)                       \
 669                                 ST(i+2,2)               \
 670                                         ST(i+3,3)       \
 671
 672
 673                 PF0(0)
 674                                 PF0(2)
 675
 676         " .align 32                     ;\n"
 677         " 1:                            ;\n"
 678
 679                 BLOCK(0)
 680                 BLOCK(4)
 681                 BLOCK(8)
 682                 BLOCK(12)
 683
 684         "       addl $256, %1           ;\n"
 685         "       addl $256, %2           ;\n"
 686         "       addl $256, %3           ;\n"
 687         "       decl %0                 ;\n"
 688         "       jnz 1b                  ;\n"
 689         : "+r" (lines),
 690           "+r" (p1), "+r"(p2), "+r"(p3)
 691         :
 692         : "memory" );
 693
 694         XMMS_RESTORE;
 695 }
 696
 697 static void
 698 xor_sse_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 699           unsigned long *p3, unsigned long *p4)
 700 {
 701         unsigned long lines = bytes >> 8;
 702         char xmm_save[16*4] ALIGN16;
 703         int cr0;
 704
 705         XMMS_SAVE;
 706
 707         __asm__ __volatile__ (
 708 #undef BLOCK
 709 #define BLOCK(i) \
 710                 PF1(i)                                  \
 711                                 PF1(i+2)                \
 712                 LD(i,0)                                 \
 713                         LD(i+1,1)                       \
 714                                 LD(i+2,2)               \
 715                                         LD(i+3,3)       \
 716                 PF2(i)                                  \
 717                                 PF2(i+2)                \
 718                 XO1(i,0)                                \
 719                         XO1(i+1,1)                      \
 720                                 XO1(i+2,2)              \
 721                                         XO1(i+3,3)      \
 722                 PF3(i)                                  \
 723                                 PF3(i+2)                \
 724                 PF0(i+4)                                \
 725                                 PF0(i+6)                \
 726                 XO2(i,0)                                \
 727                         XO2(i+1,1)                      \
 728                                 XO2(i+2,2)              \
 729                                         XO2(i+3,3)      \
 730                 XO3(i,0)                                \
 731                         XO3(i+1,1)                      \
 732                                 XO3(i+2,2)              \
 733                                         XO3(i+3,3)      \
 734                 ST(i,0)                                 \
 735                         ST(i+1,1)                       \
 736                                 ST(i+2,2)               \
 737                                         ST(i+3,3)       \
 738
 739
 740                 PF0(0)
 741                                 PF0(2)
 742
 743         " .align 32                     ;\n"
 744         " 1:                            ;\n"
 745
 746                 BLOCK(0)
 747                 BLOCK(4)
 748                 BLOCK(8)
 749                 BLOCK(12)
 750
 751         "       addl $256, %1           ;\n"
 752         "       addl $256, %2           ;\n"
 753         "       addl $256, %3           ;\n"
 754         "       addl $256, %4           ;\n"
 755         "       decl %0                 ;\n"
 756         "       jnz 1b                  ;\n"
 757         : "+r" (lines),
 758           "+r" (p1), "+r" (p2), "+r" (p3), "+r" (p4)
 759         :
 760         : "memory" );
 761
 762         XMMS_RESTORE;
 763 }
 764
 765 static void
 766 xor_sse_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
 767           unsigned long *p3, unsigned long *p4, unsigned long *p5)
 768 {
 769         unsigned long lines = bytes >> 8;
 770         char xmm_save[16*4] ALIGN16;
 771         int cr0;
 772
 773         XMMS_SAVE;
 774
 775         /* Make sure GCC forgets anything it knows about p4 or p5,
 776            such that it won't pass to the asm volatile below a
 777            register that is shared with any other variable.  That's
 778            because we modify p4 and p5 there, but we can't mark them
 779            as read/write, otherwise we'd overflow the 10-asm-operands
 780            limit of GCC < 3.1.  */
 781         __asm__ ("" : "+r" (p4), "+r" (p5));
 782
 783         __asm__ __volatile__ (
 784 #undef BLOCK
 785 #define BLOCK(i) \
 786                 PF1(i)                                  \
 787                                 PF1(i+2)                \
 788                 LD(i,0)                                 \
 789                         LD(i+1,1)                       \
 790                                 LD(i+2,2)               \
 791                                         LD(i+3,3)       \
 792                 PF2(i)                                  \
 793                                 PF2(i+2)                \
 794                 XO1(i,0)                                \
 795                         XO1(i+1,1)                      \
 796                                 XO1(i+2,2)              \
 797                                         XO1(i+3,3)      \
 798                 PF3(i)                                  \
 799                                 PF3(i+2)                \
 800                 XO2(i,0)                                \
 801                         XO2(i+1,1)                      \
 802                                 XO2(i+2,2)              \
 803                                         XO2(i+3,3)      \
 804                 PF4(i)                                  \
 805                                 PF4(i+2)                \
 806                 PF0(i+4)                                \
 807                                 PF0(i+6)                \
 808                 XO3(i,0)                                \
 809                         XO3(i+1,1)                      \
 810                                 XO3(i+2,2)              \
 811                                         XO3(i+3,3)      \
 812                 XO4(i,0)                                \
 813                         XO4(i+1,1)                      \
 814                                 XO4(i+2,2)              \
 815                                         XO4(i+3,3)      \
 816                 ST(i,0)                                 \
 817                         ST(i+1,1)                       \
 818                                 ST(i+2,2)               \
 819                                         ST(i+3,3)       \
 820
 821
 822                 PF0(0)
 823                                 PF0(2)
 824
 825         " .align 32                     ;\n"
 826         " 1:                            ;\n"
 827
 828                 BLOCK(0)
 829                 BLOCK(4)
 830                 BLOCK(8)
 831                 BLOCK(12)
 832
 833         "       addl $256, %1           ;\n"
 834         "       addl $256, %2           ;\n"
 835         "       addl $256, %3           ;\n"
 836         "       addl $256, %4           ;\n"
 837         "       addl $256, %5           ;\n"
 838         "       decl %0                 ;\n"
 839         "       jnz 1b                  ;\n"
 840         : "+r" (lines),
 841           "+r" (p1), "+r" (p2), "+r" (p3)
 842         : "r" (p4), "r" (p5)
 843         : "memory");
 844
 845         /* p4 and p5 were modified, and now the variables are dead.
 846            Clobber them just to be sure nobody does something stupid
 847            like assuming they have some legal value.  */
 848         __asm__ ("" : "=r" (p4), "=r" (p5));
 849
 850         XMMS_RESTORE;
 851 }
 852
 853 static struct xor_block_template xor_block_pIII_sse = {
 854         .name = "pIII_sse",
 855         .do_2 =  xor_sse_2,
 856         .do_3 =  xor_sse_3,
 857         .do_4 =  xor_sse_4,
 858         .do_5 = xor_sse_5,
 859 };
 860
 861 /* Also try the generic routines.  */
 862 #include <asm-generic/xor.h>
 863
 864 #undef XOR_TRY_TEMPLATES
 865 #define XOR_TRY_TEMPLATES                               \
 866         do {                                            \
 867                 xor_speed(&xor_block_8regs);            \
 868                 xor_speed(&xor_block_8regs_p);          \
 869                 xor_speed(&xor_block_32regs);           \
 870                 xor_speed(&xor_block_32regs_p);         \
 871                 if (cpu_has_xmm)                        \
 872                         xor_speed(&xor_block_pIII_sse); \
 873                 if (cpu_has_mmx) {                      \
 874                         xor_speed(&xor_block_pII_mmx);  \
 875                         xor_speed(&xor_block_p5_mmx);   \
 876                 }                                       \
 877         } while (0)
 878
 879 /* We force the use of the SSE xor block because it can write around L2.
 880    We may also be able to load into the L1 only depending on how the cpu
 881    deals with a load to a line that is being prefetched.  */
 882 #define XOR_SELECT_TEMPLATE(FASTEST) \
 883         (cpu_has_xmm ? &xor_block_pIII_sse : FASTEST)