1#ifndef _I386_STRING_I486_H_ 2#define _I386_STRING_I486_H_ 3 4/* 5 * This string-include defines all string functions as inline 6 * functions. Use gcc. It also assumes ds=es=data space, this should be 7 * normal. Most of the string-functions are rather heavily hand-optimized, 8 * see especially strtok,strstr,str[c]spn. They should work, but are not 9 * very easy to understand. Everything is done entirely within the register 10 * set, making the functions fast and clean. 11 * 12 * Copyright (C) 1991, 1992 Linus Torvalds 13 * Revised and optimized for i486/pentium 14 * 1994/03/15 by Alberto Vignani/Davide Parodi @crf.it 15 * 16 * Split into 2 CPU specific files by Alan Cox to keep #ifdef noise down. 17 * 18 * 1999/10/5 Proper register args for newer GCCs and minor bugs 19 * fixed - Petko Manolov (petkan@spct.net) 20 * 1999/10/14 3DNow memscpy() added - Petkan 21 * 2000/05/09 extern changed to static in function definitions 22 * and a few cleanups - Petkan 23 */ 24 25#define __HAVE_ARCH_STRCPY 26static inline char * strcpy(char * dest,const char *src) 27{ 28register char *tmp= (char *)dest; 29register char dummy; 30__asm__ __volatile__( 31 "\n1:\t" 32 "movb (%0),%2\n\t" 33 "incl %0\n\t" 34 "movb %2,(%1)\n\t" 35 "incl %1\n\t" 36 "testb %2,%2\n\t" 37 "jne 1b" 38 :"=r" (src), "=r" (tmp), "=q" (dummy) 39 :"0" (src), "1" (tmp) 40 :"memory"); 41return dest; 42} 43 44#define __HAVE_ARCH_STRNCPY 45static inline char * strncpy(char * dest,const char *src,size_t count) 46{ 47register char *tmp= (char *)dest; 48register char dummy; 49if (count) { 50__asm__ __volatile__( 51 "\n1:\t" 52 "movb (%0),%2\n\t" 53 "incl %0\n\t" 54 "movb %2,(%1)\n\t" 55 "incl %1\n\t" 56 "decl %3\n\t" 57 "je 3f\n\t" 58 "testb %2,%2\n\t" 59 "jne 1b\n\t" 60 "2:\tmovb %2,(%1)\n\t" 61 "incl %1\n\t" 62 "decl %3\n\t" 63 "jne 2b\n\t" 64 "3:" 65 :"=r" (src), "=r" (tmp), "=q" (dummy), "=r" (count) 66 :"0" (src), "1" (tmp), "3" (count) 67 :"memory"); 68 } /* if (count) */ 69return dest; 70} 71 72#define __HAVE_ARCH_STRCAT 73static inline char * strcat(char * dest,const char * src) 74{ 75register char *tmp = (char *)(dest-1); 76register char dummy; 77__asm__ __volatile__( 78 "\n1:\tincl %1\n\t" 79 "cmpb $0,(%1)\n\t" 80 "jne 1b\n" 81 "2:\tmovb (%2),%b0\n\t" 82 "incl %2\n\t" 83 "movb %b0,(%1)\n\t" 84 "incl %1\n\t" 85 "testb %b0,%b0\n\t" 86 "jne 2b\n" 87 :"=q" (dummy), "=r" (tmp), "=r" (src) 88 :"1" (tmp), "2" (src) 89 :"memory"); 90return dest; 91} 92 93#define __HAVE_ARCH_STRNCAT 94static inline char * strncat(char * dest,const char * src,size_t count) 95{ 96register char *tmp = (char *)(dest-1); 97register char dummy; 98__asm__ __volatile__( 99 "\n1:\tincl %1\n\t" 100 "cmpb $0,(%1)\n\t" 101 "jne 1b\n" 102 "2:\tdecl %3\n\t" 103 "js 3f\n\t" 104 "movb (%2),%b0\n\t" 105 "incl %2\n\t" 106 "movb %b0,(%1)\n\t" 107 "incl %1\n\t" 108 "testb %b0,%b0\n\t" 109 "jne 2b\n" 110 "3:\txorb %0,%0\n\t" 111 "movb %b0,(%1)\n\t" 112 :"=q" (dummy), "=r" (tmp), "=r" (src), "=r" (count) 113 :"1" (tmp), "2" (src), "3" (count) 114 :"memory"); 115return dest; 116} 117 118#define __HAVE_ARCH_STRCMP 119static inline int strcmp(const char * cs,const char * ct) 120{ 121register int __res; 122__asm__ __volatile__( 123 "\n1:\tmovb (%1),%b0\n\t" 124 "incl %1\n\t" 125 "cmpb %b0,(%2)\n\t" 126 "jne 2f\n\t" 127 "incl %2\n\t" 128 "testb %b0,%b0\n\t" 129 "jne 1b\n\t" 130 "xorl %0,%0\n\t" 131 "jmp 3f\n" 132 "2:\tmovl $1,%0\n\t" 133 "jb 3f\n\t" 134 "negl %0\n" 135 "3:" 136 :"=q" (__res), "=r" (cs), "=r" (ct) 137 :"1" (cs), "2" (ct) 138 : "memory" ); 139return __res; 140} 141 142#define __HAVE_ARCH_STRNCMP 143static inline int strncmp(const char * cs,const char * ct,size_t count) 144{ 145register int __res; 146__asm__ __volatile__( 147 "\n1:\tdecl %3\n\t" 148 "js 2f\n\t" 149 "movb (%1),%b0\n\t" 150 "incl %1\n\t" 151 "cmpb %b0,(%2)\n\t" 152 "jne 3f\n\t" 153 "incl %2\n\t" 154 "testb %b0,%b0\n\t" 155 "jne 1b\n" 156 "2:\txorl %0,%0\n\t" 157 "jmp 4f\n" 158 "3:\tmovl $1,%0\n\t" 159 "jb 4f\n\t" 160 "negl %0\n" 161 "4:" 162 :"=q" (__res), "=r" (cs), "=r" (ct), "=r" (count) 163 :"1" (cs), "2" (ct), "3" (count)); 164return __res; 165} 166 167#define __HAVE_ARCH_STRCHR 168static inline char * strchr(const char * s, int c) 169{ 170register char * __res; 171__asm__ __volatile__( 172 "movb %%al,%%ah\n" 173 "1:\tmovb (%1),%%al\n\t" 174 "cmpb %%ah,%%al\n\t" 175 "je 2f\n\t" 176 "incl %1\n\t" 177 "testb %%al,%%al\n\t" 178 "jne 1b\n\t" 179 "xorl %1,%1\n" 180 "2:\tmovl %1,%0\n\t" 181 :"=a" (__res), "=r" (s) 182 :"0" (c), "1" (s)); 183return __res; 184} 185 186#define __HAVE_ARCH_STRRCHR 187static inline char * strrchr(const char * s, int c) 188{ 189int d0, d1; 190register char * __res; 191__asm__ __volatile__( 192 "movb %%al,%%ah\n" 193 "1:\tlodsb\n\t" 194 "cmpb %%ah,%%al\n\t" 195 "jne 2f\n\t" 196 "leal -1(%%esi),%0\n" 197 "2:\ttestb %%al,%%al\n\t" 198 "jne 1b" 199 :"=d" (__res), "=&S" (d0), "=&a" (d1) 200 :"0" (0), "1" (s), "2" (c)); 201return __res; 202} 203 204 205#define __HAVE_ARCH_STRCSPN 206static inline size_t strcspn(const char * cs, const char * ct) 207{ 208int d0, d1; 209register char * __res; 210__asm__ __volatile__( 211 "movl %6,%%edi\n\t" 212 "repne\n\t" 213 "scasb\n\t" 214 "notl %%ecx\n\t" 215 "decl %%ecx\n\t" 216 "movl %%ecx,%%edx\n" 217 "1:\tlodsb\n\t" 218 "testb %%al,%%al\n\t" 219 "je 2f\n\t" 220 "movl %6,%%edi\n\t" 221 "movl %%edx,%%ecx\n\t" 222 "repne\n\t" 223 "scasb\n\t" 224 "jne 1b\n" 225 "2:\tdecl %0" 226 :"=S" (__res), "=&a" (d0), "=&c" (d1) 227 :"0" (cs), "1" (0), "2" (0xffffffff), "g" (ct) 228 :"dx", "di"); 229return __res-cs; 230} 231 232 233#define __HAVE_ARCH_STRLEN 234static inline size_t strlen(const char * s) 235{ 236/* 237 * slightly slower on a 486, but with better chances of 238 * register allocation 239 */ 240register char dummy, *tmp= (char *)s; 241__asm__ __volatile__( 242 "\n1:\t" 243 "movb\t(%0),%1\n\t" 244 "incl\t%0\n\t" 245 "testb\t%1,%1\n\t" 246 "jne\t1b" 247 :"=r" (tmp),"=q" (dummy) 248 :"0" (s) 249 : "memory" ); 250return (tmp-s-1); 251} 252 253/* Added by Gertjan van Wingerde to make minix and sysv module work */ 254#define __HAVE_ARCH_STRNLEN 255static inline size_t strnlen(const char * s, size_t count) 256{ 257int d0; 258register int __res; 259__asm__ __volatile__( 260 "movl %3,%0\n\t" 261 "jmp 2f\n" 262 "1:\tcmpb $0,(%0)\n\t" 263 "je 3f\n\t" 264 "incl %0\n" 265 "2:\tdecl %2\n\t" 266 "cmpl $-1,%2\n\t" 267 "jne 1b\n" 268 "3:\tsubl %3,%0" 269 :"=a" (__res), "=&d" (d0) 270 :"1" (count), "c" (s)); 271return __res; 272} 273/* end of additional stuff */ 274 275 276/* 277 * These ought to get tweaked to do some cache priming. 278 */ 279 280static inline void * __memcpy_by4(void * to, const void * from, size_t n) 281{ 282register void *tmp = (void *)to; 283register int dummy1,dummy2; 284__asm__ __volatile__ ( 285 "\n1:\tmovl (%2),%0\n\t" 286 "addl $4,%2\n\t" 287 "movl %0,(%1)\n\t" 288 "addl $4,%1\n\t" 289 "decl %3\n\t" 290 "jnz 1b" 291 :"=r" (dummy1), "=r" (tmp), "=r" (from), "=r" (dummy2) 292 :"1" (tmp), "2" (from), "3" (n/4) 293 :"memory"); 294return (to); 295} 296 297static inline void * __memcpy_by2(void * to, const void * from, size_t n) 298{ 299register void *tmp = (void *)to; 300register int dummy1,dummy2; 301__asm__ __volatile__ ( 302 "shrl $1,%3\n\t" 303 "jz 2f\n" /* only a word */ 304 "1:\tmovl (%2),%0\n\t" 305 "addl $4,%2\n\t" 306 "movl %0,(%1)\n\t" 307 "addl $4,%1\n\t" 308 "decl %3\n\t" 309 "jnz 1b\n" 310 "2:\tmovw (%2),%w0\n\t" 311 "movw %w0,(%1)" 312 :"=r" (dummy1), "=r" (tmp), "=r" (from), "=r" (dummy2) 313 :"1" (tmp), "2" (from), "3" (n/2) 314 :"memory"); 315return (to); 316} 317 318static inline void * __memcpy_g(void * to, const void * from, size_t n) 319{ 320int d0, d1, d2; 321register void *tmp = (void *)to; 322__asm__ __volatile__ ( 323 "shrl $1,%%ecx\n\t" 324 "jnc 1f\n\t" 325 "movsb\n" 326 "1:\tshrl $1,%%ecx\n\t" 327 "jnc 2f\n\t" 328 "movsw\n" 329 "2:\trep\n\t" 330 "movsl" 331 :"=&c" (d0), "=&D" (d1), "=&S" (d2) 332 :"0" (n), "1" ((long) tmp), "2" ((long) from) 333 :"memory"); 334return (to); 335} 336 337#define __memcpy_c(d,s,count) \ 338((count%4==0) ? \ 339 __memcpy_by4((d),(s),(count)) : \ 340 ((count%2==0) ? \ 341 __memcpy_by2((d),(s),(count)) : \ 342 __memcpy_g((d),(s),(count)))) 343 344#define __memcpy(d,s,count) \ 345(__builtin_constant_p(count) ? \ 346 __memcpy_c((d),(s),(count)) : \ 347 __memcpy_g((d),(s),(count))) 348 349#define __HAVE_ARCH_MEMCPY 350 351#include <linux/config.h> 352 353#ifdef CONFIG_X86_USE_3DNOW 354 355#include <asm/mmx.h> 356 357/* 358** This CPU favours 3DNow strongly (eg AMD K6-II, K6-III, Athlon) 359*/ 360 361static inline void * __constant_memcpy3d(void * to, const void * from, size_t len) 362{ 363 if (len < 512) 364 return __memcpy_c(to, from, len); 365 return _mmx_memcpy(to, from, len); 366} 367 368static inline void *__memcpy3d(void *to, const void *from, size_t len) 369{ 370 if(len < 512) 371 return __memcpy_g(to, from, len); 372 return _mmx_memcpy(to, from, len); 373} 374 375#define memcpy(d, s, count) \ 376(__builtin_constant_p(count) ? \ 377 __constant_memcpy3d((d),(s),(count)) : \ 378 __memcpy3d((d),(s),(count))) 379 380#else /* CONFIG_X86_USE_3DNOW */ 381 382/* 383** Generic routines 384*/ 385 386 387#define memcpy(d, s, count) __memcpy(d, s, count) 388 389#endif /* CONFIG_X86_USE_3DNOW */ 390 391 392extern void __struct_cpy_bug( void ); 393 394#define struct_cpy(x,y) \ 395({ \ 396 if (sizeof(*(x)) != sizeof(*(y))) \ 397 __struct_cpy_bug; \ 398 memcpy(x, y, sizeof(*(x))); \ 399}) 400 401 402#define __HAVE_ARCH_MEMMOVE 403static inline void * memmove(void * dest,const void * src, size_t n) 404{ 405int d0, d1, d2; 406register void *tmp = (void *)dest; 407if (dest<src) 408__asm__ __volatile__ ( 409 "rep\n\t" 410 "movsb" 411 :"=&c" (d0), "=&S" (d1), "=&D" (d2) 412 :"0" (n), "1" (src), "2" (tmp) 413 :"memory"); 414else 415__asm__ __volatile__ ( 416 "std\n\t" 417 "rep\n\t" 418 "movsb\n\t" 419 "cld" 420 :"=&c" (d0), "=&S" (d1), "=&D" (d2) 421 :"0" (n), "1" (n-1+(const char *)src), "2" (n-1+(char *)tmp) 422 :"memory"); 423return dest; 424} 425 426 427#define __HAVE_ARCH_MEMCMP 428static inline int memcmp(const void * cs,const void * ct,size_t count) 429{ 430int d0, d1, d2; 431register int __res; 432__asm__ __volatile__( 433 "repe\n\t" 434 "cmpsb\n\t" 435 "je 1f\n\t" 436 "sbbl %0,%0\n\t" 437 "orb $1,%b0\n" 438 "1:" 439 :"=a" (__res), "=&S" (d0), "=&D" (d1), "=&c" (d2) 440 :"0" (0), "1" (cs), "2" (ct), "3" (count)); 441return __res; 442} 443 444 445#define __HAVE_ARCH_MEMCHR 446static inline void * memchr(const void * cs,int c,size_t count) 447{ 448int d0; 449register void * __res; 450if (!count) 451 return NULL; 452__asm__ __volatile__( 453 "repne\n\t" 454 "scasb\n\t" 455 "je 1f\n\t" 456 "movl $1,%0\n" 457 "1:\tdecl %0" 458 :"=D" (__res), "=&c" (d0) 459 :"a" (c), "0" (cs), "1" (count)); 460return __res; 461} 462 463#define __memset_cc(s,c,count) \ 464((count%4==0) ? \ 465 __memset_cc_by4((s),(c),(count)) : \ 466 ((count%2==0) ? \ 467 __memset_cc_by2((s),(c),(count)) : \ 468 __memset_cg((s),(c),(count)))) 469 470#define __memset_gc(s,c,count) \ 471((count%4==0) ? \ 472 __memset_gc_by4((s),(c),(count)) : \ 473 ((count%2==0) ? \ 474 __memset_gc_by2((s),(c),(count)) : \ 475 __memset_gg((s),(c),(count)))) 476 477#define __HAVE_ARCH_MEMSET 478#define memset(s,c,count) \ 479(__builtin_constant_p(c) ? \ 480 (__builtin_constant_p(count) ? \ 481 __memset_cc((s),(c),(count)) : \ 482 __memset_cg((s),(c),(count))) : \ 483 (__builtin_constant_p(count) ? \ 484 __memset_gc((s),(c),(count)) : \ 485 __memset_gg((s),(c),(count)))) 486 487static inline void * __memset_cc_by4(void * s, char c, size_t count) 488{ 489/* 490 * register char *tmp = s; 491 */ 492register char *tmp = (char *)s; 493register int dummy; 494__asm__ __volatile__ ( 495 "\n1:\tmovl %2,(%0)\n\t" 496 "addl $4,%0\n\t" 497 "decl %1\n\t" 498 "jnz 1b" 499 :"=r" (tmp), "=r" (dummy) 500 :"q" (0x01010101UL * (unsigned char) c), "0" (tmp), "1" (count/4) 501 :"memory"); 502return s; 503} 504 505static inline void * __memset_cc_by2(void * s, char c, size_t count) 506{ 507register void *tmp = (void *)s; 508register int dummy; 509__asm__ __volatile__ ( 510 "shrl $1,%1\n\t" /* may be divisible also by 4 */ 511 "jz 2f\n" 512 "\n1:\tmovl %2,(%0)\n\t" 513 "addl $4,%0\n\t" 514 "decl %1\n\t" 515 "jnz 1b\n" 516 "2:\tmovw %w2,(%0)" 517 :"=r" (tmp), "=r" (dummy) 518 :"q" (0x01010101UL * (unsigned char) c), "0" (tmp), "1" (count/2) 519 :"memory"); 520return s; 521} 522 523static inline void * __memset_gc_by4(void * s, char c, size_t count) 524{ 525register void *tmp = (void *)s; 526register int dummy; 527__asm__ __volatile__ ( 528 "movb %b0,%h0\n" 529 "pushw %w0\n\t" 530 "shll $16,%0\n\t" 531 "popw %w0\n" 532 "1:\tmovl %0,(%1)\n\t" 533 "addl $4,%1\n\t" 534 "decl %2\n\t" 535 "jnz 1b\n" 536 :"=q" (c), "=r" (tmp), "=r" (dummy) 537 :"0" ((unsigned) c), "1" (tmp), "2" (count/4) 538 :"memory"); 539return s; 540} 541 542static inline void * __memset_gc_by2(void * s, char c, size_t count) 543{ 544register void *tmp = (void *)s; 545register int dummy1,dummy2; 546__asm__ __volatile__ ( 547 "movb %b0,%h0\n\t" 548 "shrl $1,%2\n\t" /* may be divisible also by 4 */ 549 "jz 2f\n\t" 550 "pushw %w0\n\t" 551 "shll $16,%0\n\t" 552 "popw %w0\n" 553 "1:\tmovl %0,(%1)\n\t" 554 "addl $4,%1\n\t" 555 "decl %2\n\t" 556 "jnz 1b\n" 557 "2:\tmovw %w0,(%1)" 558 :"=q" (dummy1), "=r" (tmp), "=r" (dummy2) 559 :"0" ((unsigned) c), "1" (tmp), "2" (count/2) 560 :"memory"); 561return s; 562} 563 564static inline void * __memset_cg(void * s, char c, size_t count) 565{ 566int d0, d1; 567register void *tmp = (void *)s; 568__asm__ __volatile__ ( 569 "shrl $1,%%ecx\n\t" 570 "rep\n\t" 571 "stosw\n\t" 572 "jnc 1f\n\t" 573 "movb %%al,(%%edi)\n" 574 "1:" 575 :"=&c" (d0), "=&D" (d1) 576 :"a" (0x0101U * (unsigned char) c), "0" (count), "1" (tmp) 577 :"memory"); 578return s; 579} 580 581static inline void * __memset_gg(void * s,char c,size_t count) 582{ 583int d0, d1, d2; 584register void *tmp = (void *)s; 585__asm__ __volatile__ ( 586 "movb %%al,%%ah\n\t" 587 "shrl $1,%%ecx\n\t" 588 "rep\n\t" 589 "stosw\n\t" 590 "jnc 1f\n\t" 591 "movb %%al,(%%edi)\n" 592 "1:" 593 :"=&c" (d0), "=&D" (d1), "=&D" (d2) 594 :"0" (count), "1" (tmp), "2" (c) 595 :"memory"); 596return s; 597} 598 599 600/* 601 * find the first occurrence of byte 'c', or 1 past the area if none 602 */ 603#define __HAVE_ARCH_MEMSCAN 604static inline void * memscan(void * addr, int c, size_t size) 605{ 606 if (!size) 607 return addr; 608 __asm__("repnz; scasb 609 jnz 1f 610 dec %%edi 6111: " 612 : "=D" (addr), "=c" (size) 613 : "0" (addr), "1" (size), "a" (c)); 614 return addr; 615} 616 617#endif 618