Merge "Use more optimized version of memmove"
This commit is contained in:
commit
2282ecc295
1 changed files with 3 additions and 34 deletions
|
@ -26,6 +26,7 @@
|
|||
* SUCH DAMAGE.
|
||||
*/
|
||||
#include <string.h>
|
||||
#include <strings.h>
|
||||
|
||||
void *memmove(void *dst, const void *src, size_t n)
|
||||
{
|
||||
|
@ -37,39 +38,7 @@ void *memmove(void *dst, const void *src, size_t n)
|
|||
if (__builtin_expect((q < p) || ((size_t)(q - p) >= n), 1)) {
|
||||
return memcpy(dst, src, n);
|
||||
} else {
|
||||
#define PRELOAD_DISTANCE 64
|
||||
/* a semi-optimized memmove(). we're preloading the src and dst buffers
|
||||
* as we go */
|
||||
size_t c0, c1, i;
|
||||
p += n;
|
||||
q += n;
|
||||
/* note: we preload the destination as well, because the 1-byte at a time
|
||||
* copy below doesn't take advantage of the write-buffer, we need
|
||||
* to use the cache instead as a poor man's write-combiner */
|
||||
__builtin_prefetch(p-1);
|
||||
__builtin_prefetch(q-1);
|
||||
if (PRELOAD_DISTANCE > 32) {
|
||||
__builtin_prefetch(p-(32+1));
|
||||
__builtin_prefetch(q-(32+1));
|
||||
}
|
||||
/* do the prefetech as soon as possible, prevent the compiler to
|
||||
* reorder the instructions above the prefetch */
|
||||
asm volatile("":::"memory");
|
||||
c0 = n & 0x1F; /* cache-line is 32 bytes */
|
||||
c1 = n >> 5;
|
||||
while ( c1-- ) {
|
||||
/* ARMv6 can have up to 3 memory access outstanding */
|
||||
__builtin_prefetch(p - (PRELOAD_DISTANCE+1));
|
||||
__builtin_prefetch(q - (PRELOAD_DISTANCE+1));
|
||||
asm volatile("":::"memory");
|
||||
for (i=0 ; i<32 ; i++) {
|
||||
*--q = *--p;
|
||||
}
|
||||
}
|
||||
while ( c0-- ) {
|
||||
*--q = *--p;
|
||||
}
|
||||
bcopy(src, dst, n);
|
||||
return dst;
|
||||
}
|
||||
|
||||
return dst;
|
||||
}
|
||||
|
|
Loading…
Reference in a new issue