toolchain/gcc/newlib/libc/machine/mips/memcpy.c

   1 /*
   2 FUNCTION
   3         <<memcpy>>---copy memory regions, optimized for the mips processors
   4
   5 ANSI_SYNOPSIS
   6         #include <string.h>
   7         void* memcpy(void *<[out]>, const void *<[in]>, size_t <[n]>);
   8
   9 TRAD_SYNOPSIS
  10         void *memcpy(<[out]>, <[in]>, <[n]>
  11         void *<[out]>;
  12         void *<[in]>;
  13         size_t <[n]>;
  14
  15 DESCRIPTION
  16         This function copies <[n]> bytes from the memory region
  17         pointed to by <[in]> to the memory region pointed to by
  18         <[out]>.
  19
  20         If the regions overlap, the behavior is undefined.
  21
  22 RETURNS
  23         <<memcpy>> returns a pointer to the first byte of the <[out]>
  24         region.
  25
  26 PORTABILITY
  27 <<memcpy>> is ANSI C.
  28
  29 <<memcpy>> requires no supporting OS subroutines.
  30
  31 QUICKREF
  32         memcpy ansi pure
  33         */
  34
  35 #include <_ansi.h>
  36 #include <stddef.h>
  37 #include <limits.h>
  38
  39 #ifdef __mips64
  40 #define wordtype long long
  41 #else
  42 #define wordtype long
  43 #endif
  44
  45 /* Nonzero if either X or Y is not aligned on a "long" boundary.  */
  46 #define UNALIGNED(X, Y) \
  47   (((long)X & (sizeof (wordtype) - 1)) | ((long)Y & (sizeof (wordtype) - 1)))
  48
  49 /* How many bytes are copied each iteration of the 4X unrolled loop.  */
  50 #define BIGBLOCKSIZE    (sizeof (wordtype) << 2)
  51
  52 /* How many bytes are copied each iteration of the word copy loop.  */
  53 #define LITTLEBLOCKSIZE (sizeof (wordtype))
  54
  55 /* Threshhold for punting to the byte copier.  */
  56 #define TOO_SMALL(LEN)  ((LEN) < BIGBLOCKSIZE)
  57
  58 _PTR
  59 _DEFUN (memcpy, (dst0, src0, len0),
  60         _PTR dst0 _AND
  61         _CONST _PTR src0 _AND
  62         size_t len0)
  63 {
  64 #if defined(PREFER_SIZE_OVER_SPEED) || defined(__OPTIMIZE_SIZE__) || defined(__mips16)
  65   char *dst = (char *) dst0;
  66   char *src = (char *) src0;
  67
  68   _PTR save = dst0;
  69
  70   while (len0--)
  71     {
  72       *dst++ = *src++;
  73     }
  74
  75   return save;
  76 #else
  77   char *dst = dst0;
  78   _CONST char *src = src0;
  79   wordtype *aligned_dst;
  80   _CONST wordtype *aligned_src;
  81   int   len =  len0;
  82   size_t iter;
  83
  84   /* Handle aligned moves here.  */
  85   if (!UNALIGNED (src, dst))
  86     {
  87       iter = len / BIGBLOCKSIZE;
  88       len = len % BIGBLOCKSIZE;
  89       aligned_dst = (wordtype *)dst;
  90       aligned_src = (wordtype *)src;
  91
  92           /* Copy 4X long or long long words at a time if possible.  */
  93       while (iter > 0)
  94         {
  95           wordtype tmp0 = aligned_src[0];
  96           wordtype tmp1 = aligned_src[1];
  97           wordtype tmp2 = aligned_src[2];
  98           wordtype tmp3 = aligned_src[3];
  99
 100           aligned_dst[0] = tmp0;
 101           aligned_dst[1] = tmp1;
 102           aligned_dst[2] = tmp2;
 103           aligned_dst[3] = tmp3;
 104           aligned_src += 4;
 105           aligned_dst += 4;
 106           iter--;
 107         }
 108
 109       /* Copy one long or long long word at a time if possible.  */
 110       iter = len / LITTLEBLOCKSIZE;
 111       len = len % LITTLEBLOCKSIZE;
 112
 113       while (iter > 0)
 114         {
 115           *aligned_dst++ = *aligned_src++;
 116           iter--;
 117         }
 118
 119       /* Pick up any residual with a byte copier.  */
 120       dst = (char*)aligned_dst;
 121       src = (char*)aligned_src;
 122
 123       while (len > 0)
 124         {
 125           *dst++ = *src++;
 126           len--;
 127         }
 128
 129       return dst0;
 130     }
 131
 132   /* Handle unaligned moves here, using lwr/lwl and swr/swl where possible */
 133   else
 134     {
 135 #ifndef NO_UNALIGNED_LOADSTORE
 136       int tmp;
 137       int *int_src = (int *)src;
 138       int *int_dst = (int *)dst;
 139       iter = len / 4;
 140       len = len % 4;
 141       while (iter > 0)
 142         {
 143           __asm__ ("ulw %0,%1" : "=r" (tmp) : "m" (*int_src));
 144           iter--;
 145           int_src++;
 146           __asm__ ("usw %1,%0" : "=m" (*int_dst) : "r" (tmp));
 147           int_dst++;
 148         }
 149
 150       /* Pick up any residual with a byte copier.  */
 151       dst = (char*)int_dst;
 152       src = (char*)int_src;
 153 #endif
 154
 155       while (len > 0)
 156         {
 157           *dst++ = *src++;
 158           len--;
 159         }
 160
 161       return dst0;
 162     }
 163 #endif /* not PREFER_SIZE_OVER_SPEED */
 164 }