URL https://opencores.org/ocsvn/or1k_old/or1k_old/trunk

Subversion Repositories or1k_old

[/] [or1k_old/] [trunk/] [rc203soc/] [sw/] [uClinux/] [include/] [asm-i960/] [string.h] - Blame information for rev 1782

Details | Compare with Previous | View Log


#ifndef _I960_STRING_H_
#define _I960_STRING_H_
 
/* FIXME we should probably do this, since the i960 can move 16 bytes in
   only a few cycles if in cache, whereas stock strcpy is going to SUCK */
 
#include <linux/config.h>
#include <asm/page.h>
 
#if 0
 
#define __HAVE_ARCH_STRCPY
extern inline char * strcpy(char * dest,const char *src)
{
  char *xdest = dest;
 
  __asm__ __volatile__
       ("1:\tmoveb %1@+,%0@+\n\t"
        "jne 1b"
        : "=a" (dest), "=a" (src)
        : "0" (dest), "1" (src) : "memory");
  return xdest;
}
 
#define __HAVE_ARCH_STRNCPY
extern inline char * strncpy(char *dest, const char *src, size_t n)
{
  char *xdest = dest;
 
  if (n == 0)
    return xdest;
 
  __asm__ __volatile__
       ("1:\tmoveb %1@+,%0@+\n\t"
        "jeq 2f\n\t"
        "subql #1,%2\n\t"
        "jne 1b\n\t"
        "2:"
        : "=a" (dest), "=a" (src), "=d" (n)
        : "0" (dest), "1" (src), "2" (n)
        : "memory");
  return xdest;
}
 
#define __HAVE_ARCH_STRCAT
extern inline char * strcat(char * dest, const char * src)
{
        char *tmp = dest;
 
        while (*dest)
                dest++;
        while ((*dest++ = *src++))
                ;
 
        return tmp;
}
 
#define __HAVE_ARCH_STRNCAT
extern inline char * strncat(char *dest, const char *src, size_t count)
{
        char *tmp = dest;
 
        if (count) {
                while (*dest)
                        dest++;
                while ((*dest++ = *src++)) {
                        if (--count == 0) {
                                *dest++='\0';
                                break;
                        }
                }
        }
 
        return tmp;
}
 
#define __HAVE_ARCH_STRCHR
extern inline char * strchr(const char * s, int c)
{
  const char ch = c;
 
  for(; *s != ch; ++s)
    if (*s == '\0')
      return( NULL );
  return( (char *) s);
}
 
#define __HAVE_ARCH_STRPBRK
extern inline char * strpbrk(const char * cs,const char * ct)
{
  const char *sc1,*sc2;
 
  for( sc1 = cs; *sc1 != '\0'; ++sc1)
    for( sc2 = ct; *sc2 != '\0'; ++sc2)
      if (*sc1 == *sc2)
        return((char *) sc1);
  return( NULL );
}
 
#define __HAVE_ARCH_STRSPN
extern inline size_t strspn(const char *s, const char *accept)
{
  const char *p;
  const char *a;
  size_t count = 0;
 
  for (p = s; *p != '\0'; ++p)
    {
      for (a = accept; *a != '\0'; ++a)
        if (*p == *a)
          break;
      if (*a == '\0')
        return count;
      else
        ++count;
    }
 
  return count;
}
 
#define __HAVE_ARCH_STRTOK
extern inline char * strtok(char * s,const char * ct)
{
  char *sbegin, *send;
 
  sbegin  = s ? s : ___strtok;
  if (!sbegin) {
          return NULL;
  }
  sbegin += strspn(sbegin,ct);
  if (*sbegin == '\0') {
    ___strtok = NULL;
    return( NULL );
  }
  send = strpbrk( sbegin, ct);
  if (send && *send != '\0')
    *send++ = '\0';
  ___strtok = send;
  return (sbegin);
}
 
/* strstr !! */
 
#define __HAVE_ARCH_STRLEN
extern inline size_t strlen(const char * s)
{
  const char *sc;
  for (sc = s; *sc != '\0'; ++sc) ;
  return(sc - s);
}
 
/* strnlen !! */
 
#define __HAVE_ARCH_STRCMP
extern inline int strcmp(const char * cs,const char * ct)
{
  char __res;
 
  __asm__
       ("1:\tmoveb %0@+,%2\n\t" /* get *cs */
        "cmpb %1@+,%2\n\t"      /* compare a byte */
        "jne  2f\n\t"           /* not equal, break out */
        "tstb %2\n\t"           /* at end of cs? */
        "jne  1b\n\t"           /* no, keep going */
        "jra  3f\n\t"           /* strings are equal */
        "2:\tsubb %1@-,%2\n\t"  /* *cs - *ct */
        "3:"
        : "=a" (cs), "=a" (ct), "=d" (__res)
        : "0" (cs), "1" (ct));
  return __res;
}
 
#define __HAVE_ARCH_STRNCMP
extern inline int strncmp(const char * cs,const char * ct,size_t count)
{
  char __res;
 
  if (!count)
    return 0;
  __asm__
       ("1:\tmovb %0@+,%3\n\t"          /* get *cs */
        "cmpb   %1@+,%3\n\t"            /* compare a byte */
        "jne    3f\n\t"                 /* not equal, break out */
        "tstb   %3\n\t"                 /* at end of cs? */
        "jeq    4f\n\t"                 /* yes, all done */
        "subql  #1,%2\n\t"              /* no, adjust count */
        "jne    1b\n\t"                 /* more to do, keep going */
        "2:\tmoveq #0,%3\n\t"           /* strings are equal */
        "jra    4f\n\t"
        "3:\tsubb %1@-,%3\n\t"          /* *cs - *ct */
        "4:"
        : "=a" (cs), "=a" (ct), "=d" (count), "=d" (__res)
        : "0" (cs), "1" (ct), "2" (count));
  return __res;
}
 
#define __HAVE_ARCH_MEMSET
/*
 * This is really ugly, but its highly optimizatiable by the
 * compiler and is meant as compensation for gcc's missing
 * __builtin_memset(). For the 680[23]0 it might be worth considering
 * the optimal number of misaligned writes compared to the number of
 * tests'n'branches needed to align the destination address. The
 * 680[46]0 doesn't really care due to their copy-back caches.
 *                                              10/09/96 - Jes Sorensen
 */
extern inline void * __memset_g(void * s, int c, size_t count)
{
  void *xs = s;
  size_t temp;
 
  if (!count)
    return xs;
 
  c &= 0xff;
  c |= c << 8;
  c |= c << 16;
 
  if (count < 36){
          long *ls = s;
 
          switch(count){
          case 32: case 33: case 34: case 35:
                  *ls++ = c;
          case 28: case 29: case 30: case 31:
                  *ls++ = c;
          case 24: case 25: case 26: case 27:
                  *ls++ = c;
          case 20: case 21: case 22: case 23:
                  *ls++ = c;
          case 16: case 17: case 18: case 19:
                  *ls++ = c;
          case 12: case 13: case 14: case 15:
                  *ls++ = c;
          case 8: case 9: case 10: case 11:
                  *ls++ = c;
          case 4: case 5: case 6: case 7:
                  *ls++ = c;
                  break;
          default:
                  break;
          }
          s = ls;
          if (count & 0x02){
                  short *ss = s;
                  *ss++ = c;
                  s = ss;
          }
          if (count & 0x01){
                  char *cs = s;
                  *cs++ = c;
                  s = cs;
          }
          return xs;
  }
 
  if ((long) s & 1)
    {
      char *cs = s;
      *cs++ = c;
      s = cs;
      count--;
    }
  if (count > 2 && (long) s & 2)
    {
      short *ss = s;
      *ss++ = c;
      s = ss;
      count -= 2;
    }
  temp = count >> 2;
  if (temp)
    {
      long *ls = s;
      temp--;
      do
        *ls++ = c;
      while (temp--);
      s = ls;
    }
  if (count & 2)
    {
      short *ss = s;
      *ss++ = c;
      s = ss;
    }
  if (count & 1)
    {
      char *cs = s;
      *cs = c;
    }
  return xs;
}
 
/*
 * __memset_page assumes that data is longword aligned. Most, if not
 * all, of these page sized memsets are performed on page aligned
 * areas, thus we do not need to check if the destination is longword
 * aligned. Of course we suffer a serious performance loss if this is
 * not the case but I think the risk of this ever happening is
 * extremely small. We spend a lot of time clearing pages in
 * get_empty_page() so I think it is worth it anyway. Besides, the
 * 680[46]0 do not really care about misaligned writes due to their
 * copy-back cache.
 *
 * The optimized case for the 680[46]0 is implemented using the move16
 * instruction. My tests showed that this implementation is 35-45%
 * faster than the original implementation using movel, the only
 * caveat is that the destination address must be 16-byte aligned.
 *                                            01/09/96 - Jes Sorensen
 */
extern inline void * __memset_page(void * s,int c,size_t count)
{
  unsigned long data, tmp;
  void *xs, *sp;
 
  xs = sp = s;
 
  c = c & 255;
  data = c | (c << 8);
  data |= data << 16;
 
#if defined(CONFIG_OPTIMIZE_040) || defined(CONFIG_OPTIMIZE_060)
 
  if (((unsigned long) s) & 0x0f)
          memset(s, c, count);
  else{
          *((unsigned long *)(s))++ = data;
          *((unsigned long *)(s))++ = data;
          *((unsigned long *)(s))++ = data;
          *((unsigned long *)(s))++ = data;
 
          __asm__ __volatile__("1:\t"
                               "move16 %2@+,%0@+\n\t"
                               "subqw  #8,%2\n\t"
                               "subqw  #8,%2\n\t"
                               "dbra   %1,1b\n\t"
                               : "=a" (s), "=d" (tmp)
                               : "a" (sp), "0" (s), "1" ((count - 16) / 16 - 1)
                               );
  }
 
#else
  __asm__ __volatile__("1:\t"
                       "movel %2,%0@+\n\t"
                       "movel %2,%0@+\n\t"
                       "movel %2,%0@+\n\t"
                       "movel %2,%0@+\n\t"
                       "movel %2,%0@+\n\t"
                       "movel %2,%0@+\n\t"
                       "movel %2,%0@+\n\t"
                       "movel %2,%0@+\n\t"
                       "dbra  %1,1b\n\t"
                       : "=a" (s), "=d" (tmp)
                       : "d" (data), "0" (s), "1" (count / 32 - 1)
                       );
#endif
 
  return xs;
}
 
#define __memset_const(s,c,count) \
((count==PAGE_SIZE) ? \
  __memset_page((s),(c),(count)) : \
  __memset_g((s),(c),(count)))
 
#define memset(s, c, count) \
(__builtin_constant_p(count) ? \
 __memset_const((s),(c),(count)) : \
 memset((s),(c),(count)))
 
#define __HAVE_ARCH_MEMCPY
/*
 * __builtin_memcpy() does not handle page-sized memcpys very well,
 * thus following the same assumptions as for page-sized memsets, this
 * function copies page-sized areas using an unrolled loop, without
 * considering alignment.
 *
 * For the 680[46]0 only kernels we use the move16 instruction instead
 * as it writes through the data-cache, invalidating the cache-lines
 * touched. In this way we do not use up the entire data-cache (well,
 * half of it on the 68060) by copying a page. An unrolled loop of two
 * move16 instructions seem to the fastest. The only caveat is that
 * both source and destination must be 16-byte aligned, if not we fall
 * back to the generic memcpy function.  - Jes
 */
extern inline void * __memcpy_page(void * to, const void * from, size_t count)
{
  unsigned long tmp;
  void *xto = to;
 
#if defined(CONFIG_OPTIMIZE_040) || defined(CONFIG_OPTIMIZE_060)
 
  if (((unsigned long) to | (unsigned long) from) & 0x0f)
          return memcpy(to, from, count);
 
  __asm__ __volatile__("1:\t"
                       "move16 %1@+,%0@+\n\t"
                       "move16 %1@+,%0@+\n\t"
                       "dbra  %2,1b\n\t"
                       : "=a" (to), "=a" (from), "=d" (tmp)
                       : "0" (to), "1" (from) , "2" (count / 32 - 1)
                       );
#else
  __asm__ __volatile__("1:\t"
                       "movel %1@+,%0@+\n\t"
                       "movel %1@+,%0@+\n\t"
                       "movel %1@+,%0@+\n\t"
                       "movel %1@+,%0@+\n\t"
                       "movel %1@+,%0@+\n\t"
                       "movel %1@+,%0@+\n\t"
                       "movel %1@+,%0@+\n\t"
                       "movel %1@+,%0@+\n\t"
                       "dbra  %2,1b\n\t"
                       : "=a" (to), "=a" (from), "=d" (tmp)
                       : "0" (to), "1" (from) , "2" (count / 32 - 1)
                       );
#endif
  return xto;
}
 
#define __memcpy_const(to, from, n) \
((n==PAGE_SIZE) ? \
  __memcpy_page((to),(from),(n)) : \
  __builtin_memcpy((to),(from),(n)))
 
#define memcpy(to, from, n) \
(__builtin_constant_p(n) ? \
 __memcpy_const((to),(from),(n)) : \
 memcpy((to),(from),(n)))
 
#define __HAVE_ARCH_MEMMOVE
extern inline void * memmove(void * dest,const void * src, size_t n)
{
  void *xdest = dest;
  size_t temp;
 
  if (!n)
    return xdest;
 
  if (dest < src)
    {
      if ((long) dest & 1)
        {
          char *cdest = dest;
          const char *csrc = src;
          *cdest++ = *csrc++;
          dest = cdest;
          src = csrc;
          n--;
        }
      if (n > 2 && (long) dest & 2)
        {
          short *sdest = dest;
          const short *ssrc = src;
          *sdest++ = *ssrc++;
          dest = sdest;
          src = ssrc;
          n -= 2;
        }
      temp = n >> 2;
      if (temp)
        {
          long *ldest = dest;
          const long *lsrc = src;
          temp--;
          do
            *ldest++ = *lsrc++;
          while (temp--);
          dest = ldest;
          src = lsrc;
        }
      if (n & 2)
        {
          short *sdest = dest;
          const short *ssrc = src;
          *sdest++ = *ssrc++;
          dest = sdest;
          src = ssrc;
        }
      if (n & 1)
        {
          char *cdest = dest;
          const char *csrc = src;
          *cdest = *csrc;
        }
    }
  else
    {
      dest = (char *) dest + n;
      src = (const char *) src + n;
      if ((long) dest & 1)
        {
          char *cdest = dest;
          const char *csrc = src;
          *--cdest = *--csrc;
          dest = cdest;
          src = csrc;
          n--;
        }
      if (n > 2 && (long) dest & 2)
        {
          short *sdest = dest;
          const short *ssrc = src;
          *--sdest = *--ssrc;
          dest = sdest;
          src = ssrc;
          n -= 2;
        }
      temp = n >> 2;
      if (temp)
        {
          long *ldest = dest;
          const long *lsrc = src;
          temp--;
          do
            *--ldest = *--lsrc;
          while (temp--);
          dest = ldest;
          src = lsrc;
        }
      if (n & 2)
        {
          short *sdest = dest;
          const short *ssrc = src;
          *--sdest = *--ssrc;
          dest = sdest;
          src = ssrc;
        }
      if (n & 1)
        {
          char *cdest = dest;
          const char *csrc = src;
          *--cdest = *--csrc;
        }
    }
  return xdest;
}
 
#define __HAVE_ARCH_MEMCMP
#define memcmp(cs, ct, n) \
(__builtin_constant_p(n) ? \
 __builtin_memcmp((cs),(ct),(n)) : \
 memcmp((cs),(ct),(n)))
 
#endif
 
#endif /* _I960_STRING_H_ */

Browse

Tools

Subversion Repositories or1k_old

[/] [or1k_old/] [trunk/] [rc203soc/] [sw/] [uClinux/] [include/] [asm-i960/] [string.h] - Blame information for rev 1782

Line No.	Rev	Author	Line
1	1633	jcastillo	`#ifndef _I960_STRING_H_`
2			`#define _I960_STRING_H_`
3
4			`/* FIXME we should probably do this, since the i960 can move 16 bytes in`
5			`only a few cycles if in cache, whereas stock strcpy is going to SUCK */`
6
7			`#include <linux/config.h>`
8			`#include <asm/page.h>`
9
10			`#if 0`
11
12			`#define __HAVE_ARCH_STRCPY`
13			`extern inline char * strcpy(char * dest,const char *src)`
14			`{`
15			`char *xdest = dest;`
16
17			`__asm__ __volatile__`
18			`("1:\tmoveb %1@+,%0@+\n\t"`
19			`"jne 1b"`
20			`: "=a" (dest), "=a" (src)`
21			`: "0" (dest), "1" (src) : "memory");`
22			`return xdest;`
23			`}`
24
25			`#define __HAVE_ARCH_STRNCPY`
26			`extern inline char * strncpy(char dest, const char src, size_t n)`
27			`{`
28			`char *xdest = dest;`
29
30			`if (n == 0)`
31			`return xdest;`
32
33			`__asm__ __volatile__`
34			`("1:\tmoveb %1@+,%0@+\n\t"`
35			`"jeq 2f\n\t"`
36			`"subql #1,%2\n\t"`
37			`"jne 1b\n\t"`
38			`"2:"`
39			`: "=a" (dest), "=a" (src), "=d" (n)`
40			`: "0" (dest), "1" (src), "2" (n)`
41			`: "memory");`
42			`return xdest;`
43			`}`
44
45			`#define __HAVE_ARCH_STRCAT`
46			`extern inline char * strcat(char * dest, const char * src)`
47			`{`
48			`char *tmp = dest;`
49
50			`while (*dest)`
51			`dest++;`
52			`while ((dest++ = src++))`
53			`;`
54
55			`return tmp;`
56			`}`
57
58			`#define __HAVE_ARCH_STRNCAT`
59			`extern inline char * strncat(char dest, const char src, size_t count)`
60			`{`
61			`char *tmp = dest;`
62
63			`if (count) {`
64			`while (*dest)`
65			`dest++;`
66			`while ((dest++ = src++)) {`
67			`if (--count == 0) {`
68			`*dest++='\0';`
69			`break;`
70			`}`
71			`}`
72			`}`
73
74			`return tmp;`
75			`}`
76
77			`#define __HAVE_ARCH_STRCHR`
78			`extern inline char * strchr(const char * s, int c)`
79			`{`
80			`const char ch = c;`
81
82			`for(; *s != ch; ++s)`
83			`if (*s == '\0')`
84			`return( NULL );`
85			`return( (char *) s);`
86			`}`
87
88			`#define __HAVE_ARCH_STRPBRK`
89			`extern inline char * strpbrk(const char * cs,const char * ct)`
90			`{`
91			`const char sc1,sc2;`
92
93			`for( sc1 = cs; *sc1 != '\0'; ++sc1)`
94			`for( sc2 = ct; *sc2 != '\0'; ++sc2)`
95			`if (sc1 == sc2)`
96			`return((char *) sc1);`
97			`return( NULL );`
98			`}`
99
100			`#define __HAVE_ARCH_STRSPN`
101			`extern inline size_t strspn(const char s, const char accept)`
102			`{`
103			`const char *p;`
104			`const char *a;`
105			`size_t count = 0;`
106
107			`for (p = s; *p != '\0'; ++p)`
108			`{`
109			`for (a = accept; *a != '\0'; ++a)`
110			`if (p == a)`
111			`break;`
112			`if (*a == '\0')`
113			`return count;`
114			`else`
115			`++count;`
116			`}`
117
118			`return count;`
119			`}`
120
121			`#define __HAVE_ARCH_STRTOK`
122			`extern inline char * strtok(char * s,const char * ct)`
123			`{`
124			`char sbegin, send;`
125
126			`sbegin = s ? s : ___strtok;`
127			`if (!sbegin) {`
128			`return NULL;`
129			`}`
130			`sbegin += strspn(sbegin,ct);`
131			`if (*sbegin == '\0') {`
132			`___strtok = NULL;`
133			`return( NULL );`
134			`}`
135			`send = strpbrk( sbegin, ct);`
136			`if (send && *send != '\0')`
137			`*send++ = '\0';`
138			`___strtok = send;`
139			`return (sbegin);`
140			`}`
141
142			`/* strstr !! */`
143
144			`#define __HAVE_ARCH_STRLEN`
145			`extern inline size_t strlen(const char * s)`
146			`{`
147			`const char *sc;`
148			`for (sc = s; *sc != '\0'; ++sc) ;`
149			`return(sc - s);`
150			`}`
151
152			`/* strnlen !! */`
153
154			`#define __HAVE_ARCH_STRCMP`
155			`extern inline int strcmp(const char * cs,const char * ct)`
156			`{`
157			`char __res;`
158
159			`__asm__`
160			`("1:\tmoveb %0@+,%2\n\t" /* get cs /`
161			`"cmpb %1@+,%2\n\t" /* compare a byte */`
162			`"jne 2f\n\t" /* not equal, break out */`
163			`"tstb %2\n\t" /* at end of cs? */`
164			`"jne 1b\n\t" /* no, keep going */`
165			`"jra 3f\n\t" /* strings are equal */`
166			`"2:\tsubb %1@-,%2\n\t" /* cs - ct */`
167			`"3:"`
168			`: "=a" (cs), "=a" (ct), "=d" (__res)`
169			`: "0" (cs), "1" (ct));`
170			`return __res;`
171			`}`
172
173			`#define __HAVE_ARCH_STRNCMP`
174			`extern inline int strncmp(const char * cs,const char * ct,size_t count)`
175			`{`
176			`char __res;`
177
178			`if (!count)`
179			`return 0;`
180			`__asm__`
181			`("1:\tmovb %0@+,%3\n\t" /* get cs /`
182			`"cmpb %1@+,%3\n\t" /* compare a byte */`
183			`"jne 3f\n\t" /* not equal, break out */`
184			`"tstb %3\n\t" /* at end of cs? */`
185			`"jeq 4f\n\t" /* yes, all done */`
186			`"subql #1,%2\n\t" /* no, adjust count */`
187			`"jne 1b\n\t" /* more to do, keep going */`
188			`"2:\tmoveq #0,%3\n\t" /* strings are equal */`
189			`"jra 4f\n\t"`
190			`"3:\tsubb %1@-,%3\n\t" /* cs - ct */`
191			`"4:"`
192			`: "=a" (cs), "=a" (ct), "=d" (count), "=d" (__res)`
193			`: "0" (cs), "1" (ct), "2" (count));`
194			`return __res;`
195			`}`
196
197			`#define __HAVE_ARCH_MEMSET`
198			`/*`
199			`* This is really ugly, but its highly optimizatiable by the`
200			`* compiler and is meant as compensation for gcc's missing`
201			`* __builtin_memset(). For the 680[23]0 it might be worth considering`
202			`* the optimal number of misaligned writes compared to the number of`
203			`* tests'n'branches needed to align the destination address. The`
204			`* 680[46]0 doesn't really care due to their copy-back caches.`
205			`* 10/09/96 - Jes Sorensen`
206			`*/`
207			`extern inline void * __memset_g(void * s, int c, size_t count)`
208			`{`
209			`void *xs = s;`
210			`size_t temp;`
211
212			`if (!count)`
213			`return xs;`
214
215			`c &= 0xff;`
216			`c \|= c << 8;`
217			`c \|= c << 16;`
218
219			`if (count < 36){`
220			`long *ls = s;`
221
222			`switch(count){`
223			`case 32: case 33: case 34: case 35:`
224			`*ls++ = c;`
225			`case 28: case 29: case 30: case 31:`
226			`*ls++ = c;`
227			`case 24: case 25: case 26: case 27:`
228			`*ls++ = c;`
229			`case 20: case 21: case 22: case 23:`
230			`*ls++ = c;`
231			`case 16: case 17: case 18: case 19:`
232			`*ls++ = c;`
233			`case 12: case 13: case 14: case 15:`
234			`*ls++ = c;`
235			`case 8: case 9: case 10: case 11:`
236			`*ls++ = c;`
237			`case 4: case 5: case 6: case 7:`
238			`*ls++ = c;`
239			`break;`
240			`default:`
241			`break;`
242			`}`
243			`s = ls;`
244			`if (count & 0x02){`
245			`short *ss = s;`
246			`*ss++ = c;`
247			`s = ss;`
248			`}`
249			`if (count & 0x01){`
250			`char *cs = s;`
251			`*cs++ = c;`
252			`s = cs;`
253			`}`
254			`return xs;`
255			`}`
256
257			`if ((long) s & 1)`
258			`{`
259			`char *cs = s;`
260			`*cs++ = c;`
261			`s = cs;`
262			`count--;`
263			`}`
264			`if (count > 2 && (long) s & 2)`
265			`{`
266			`short *ss = s;`
267			`*ss++ = c;`
268			`s = ss;`
269			`count -= 2;`
270			`}`
271			`temp = count >> 2;`
272			`if (temp)`
273			`{`
274			`long *ls = s;`
275			`temp--;`
276			`do`
277			`*ls++ = c;`
278			`while (temp--);`
279			`s = ls;`
280			`}`
281			`if (count & 2)`
282			`{`
283			`short *ss = s;`
284			`*ss++ = c;`
285			`s = ss;`
286			`}`
287			`if (count & 1)`
288			`{`
289			`char *cs = s;`
290			`*cs = c;`
291			`}`
292			`return xs;`
293			`}`
294
295			`/*`
296			`* __memset_page assumes that data is longword aligned. Most, if not`
297			`* all, of these page sized memsets are performed on page aligned`
298			`* areas, thus we do not need to check if the destination is longword`
299			`* aligned. Of course we suffer a serious performance loss if this is`
300			`* not the case but I think the risk of this ever happening is`
301			`* extremely small. We spend a lot of time clearing pages in`
302			`* get_empty_page() so I think it is worth it anyway. Besides, the`
303			`* 680[46]0 do not really care about misaligned writes due to their`
304			`* copy-back cache.`
305			`*`
306			`* The optimized case for the 680[46]0 is implemented using the move16`
307			`* instruction. My tests showed that this implementation is 35-45%`
308			`* faster than the original implementation using movel, the only`
309			`* caveat is that the destination address must be 16-byte aligned.`
310			`* 01/09/96 - Jes Sorensen`
311			`*/`
312			`extern inline void * __memset_page(void * s,int c,size_t count)`
313			`{`
314			`unsigned long data, tmp;`
315			`void xs, sp;`
316
317			`xs = sp = s;`
318
319			`c = c & 255;`
320			`data = c \| (c << 8);`
321			`data \|= data << 16;`
322
323			`#if defined(CONFIG_OPTIMIZE_040) \|\| defined(CONFIG_OPTIMIZE_060)`
324
325			`if (((unsigned long) s) & 0x0f)`
326			`memset(s, c, count);`
327			`else{`
328			`((unsigned long )(s))++ = data;`
329			`((unsigned long )(s))++ = data;`
330			`((unsigned long )(s))++ = data;`
331			`((unsigned long )(s))++ = data;`
332
333			`__asm__ __volatile__("1:\t"`
334			`"move16 %2@+,%0@+\n\t"`
335			`"subqw #8,%2\n\t"`
336			`"subqw #8,%2\n\t"`
337			`"dbra %1,1b\n\t"`
338			`: "=a" (s), "=d" (tmp)`
339			`: "a" (sp), "0" (s), "1" ((count - 16) / 16 - 1)`
340			`);`
341			`}`
342
343			`#else`
344			`__asm__ __volatile__("1:\t"`
345			`"movel %2,%0@+\n\t"`
346			`"movel %2,%0@+\n\t"`
347			`"movel %2,%0@+\n\t"`
348			`"movel %2,%0@+\n\t"`
349			`"movel %2,%0@+\n\t"`
350			`"movel %2,%0@+\n\t"`
351			`"movel %2,%0@+\n\t"`
352			`"movel %2,%0@+\n\t"`
353			`"dbra %1,1b\n\t"`
354			`: "=a" (s), "=d" (tmp)`
355			`: "d" (data), "0" (s), "1" (count / 32 - 1)`
356			`);`
357			`#endif`
358
359			`return xs;`
360			`}`
361
362			`#define __memset_const(s,c,count) \`
363			`((count==PAGE_SIZE) ? \`
364			`__memset_page((s),(c),(count)) : \`
365			`__memset_g((s),(c),(count)))`
366
367			`#define memset(s, c, count) \`
368			`(__builtin_constant_p(count) ? \`
369			`__memset_const((s),(c),(count)) : \`
370			`memset((s),(c),(count)))`
371
372			`#define __HAVE_ARCH_MEMCPY`
373			`/*`
374			`* __builtin_memcpy() does not handle page-sized memcpys very well,`
375			`* thus following the same assumptions as for page-sized memsets, this`
376			`* function copies page-sized areas using an unrolled loop, without`
377			`* considering alignment.`
378			`*`
379			`* For the 680[46]0 only kernels we use the move16 instruction instead`
380			`* as it writes through the data-cache, invalidating the cache-lines`
381			`* touched. In this way we do not use up the entire data-cache (well,`
382			`* half of it on the 68060) by copying a page. An unrolled loop of two`
383			`* move16 instructions seem to the fastest. The only caveat is that`
384			`* both source and destination must be 16-byte aligned, if not we fall`
385			`* back to the generic memcpy function. - Jes`
386			`*/`
387			`extern inline void * __memcpy_page(void * to, const void * from, size_t count)`
388			`{`
389			`unsigned long tmp;`
390			`void *xto = to;`
391
392			`#if defined(CONFIG_OPTIMIZE_040) \|\| defined(CONFIG_OPTIMIZE_060)`
393
394			`if (((unsigned long) to \| (unsigned long) from) & 0x0f)`
395			`return memcpy(to, from, count);`
396
397			`__asm__ __volatile__("1:\t"`
398			`"move16 %1@+,%0@+\n\t"`
399			`"move16 %1@+,%0@+\n\t"`
400			`"dbra %2,1b\n\t"`
401			`: "=a" (to), "=a" (from), "=d" (tmp)`
402			`: "0" (to), "1" (from) , "2" (count / 32 - 1)`
403			`);`
404			`#else`
405			`__asm__ __volatile__("1:\t"`
406			`"movel %1@+,%0@+\n\t"`
407			`"movel %1@+,%0@+\n\t"`
408			`"movel %1@+,%0@+\n\t"`
409			`"movel %1@+,%0@+\n\t"`
410			`"movel %1@+,%0@+\n\t"`
411			`"movel %1@+,%0@+\n\t"`
412			`"movel %1@+,%0@+\n\t"`
413			`"movel %1@+,%0@+\n\t"`
414			`"dbra %2,1b\n\t"`
415			`: "=a" (to), "=a" (from), "=d" (tmp)`
416			`: "0" (to), "1" (from) , "2" (count / 32 - 1)`
417			`);`
418			`#endif`
419			`return xto;`
420			`}`
421
422			`#define __memcpy_const(to, from, n) \`
423			`((n==PAGE_SIZE) ? \`
424			`__memcpy_page((to),(from),(n)) : \`
425			`__builtin_memcpy((to),(from),(n)))`
426
427			`#define memcpy(to, from, n) \`
428			`(__builtin_constant_p(n) ? \`
429			`__memcpy_const((to),(from),(n)) : \`
430			`memcpy((to),(from),(n)))`
431
432			`#define __HAVE_ARCH_MEMMOVE`
433			`extern inline void * memmove(void * dest,const void * src, size_t n)`
434			`{`
435			`void *xdest = dest;`
436			`size_t temp;`
437
438			`if (!n)`
439			`return xdest;`
440
441			`if (dest < src)`
442			`{`
443			`if ((long) dest & 1)`
444			`{`
445			`char *cdest = dest;`
446			`const char *csrc = src;`
447			`cdest++ = csrc++;`
448			`dest = cdest;`
449			`src = csrc;`
450			`n--;`
451			`}`
452			`if (n > 2 && (long) dest & 2)`
453			`{`
454			`short *sdest = dest;`
455			`const short *ssrc = src;`
456			`sdest++ = ssrc++;`
457			`dest = sdest;`
458			`src = ssrc;`
459			`n -= 2;`
460			`}`
461			`temp = n >> 2;`
462			`if (temp)`
463			`{`
464			`long *ldest = dest;`
465			`const long *lsrc = src;`
466			`temp--;`
467			`do`
468			`ldest++ = lsrc++;`
469			`while (temp--);`
470			`dest = ldest;`
471			`src = lsrc;`
472			`}`
473			`if (n & 2)`
474			`{`
475			`short *sdest = dest;`
476			`const short *ssrc = src;`
477			`sdest++ = ssrc++;`
478			`dest = sdest;`
479			`src = ssrc;`
480			`}`
481			`if (n & 1)`
482			`{`
483			`char *cdest = dest;`
484			`const char *csrc = src;`
485			`cdest = csrc;`
486			`}`
487			`}`
488			`else`
489			`{`
490			`dest = (char *) dest + n;`
491			`src = (const char *) src + n;`
492			`if ((long) dest & 1)`
493			`{`
494			`char *cdest = dest;`
495			`const char *csrc = src;`
496			`--cdest = --csrc;`
497			`dest = cdest;`
498			`src = csrc;`
499			`n--;`
500			`}`
501			`if (n > 2 && (long) dest & 2)`
502			`{`
503			`short *sdest = dest;`
504			`const short *ssrc = src;`
505			`--sdest = --ssrc;`
506			`dest = sdest;`
507			`src = ssrc;`
508			`n -= 2;`
509			`}`
510			`temp = n >> 2;`
511			`if (temp)`
512			`{`
513			`long *ldest = dest;`
514			`const long *lsrc = src;`
515			`temp--;`
516			`do`
517			`--ldest = --lsrc;`
518			`while (temp--);`
519			`dest = ldest;`
520			`src = lsrc;`
521			`}`
522			`if (n & 2)`
523			`{`
524			`short *sdest = dest;`
525			`const short *ssrc = src;`
526			`--sdest = --ssrc;`
527			`dest = sdest;`
528			`src = ssrc;`
529			`}`
530			`if (n & 1)`
531			`{`
532			`char *cdest = dest;`
533			`const char *csrc = src;`
534			`--cdest = --csrc;`
535			`}`
536			`}`
537			`return xdest;`
538			`}`
539
540			`#define __HAVE_ARCH_MEMCMP`
541			`#define memcmp(cs, ct, n) \`
542			`(__builtin_constant_p(n) ? \`
543			`__builtin_memcmp((cs),(ct),(n)) : \`
544			`memcmp((cs),(ct),(n)))`
545
546			`#endif`
547
548			`#endif /* _I960_STRING_H_ */`