#include <stdio.h>
#include "memcpy_SSE2.h"

// SSE2のmovdqu命令を用いて128bitずつデータを移動する
void memcpy_sse2(void *src, void *dst, size_t size)
{
  asm ("\n"
    "    mov      %0, %%eax\n"
    "    lea      (%%esi, %0), %%ebx\n"
    "  ssecalc:\n"
    "    movdqu   (%%esi),%%xmm0\n"
    "    lea      16(%%esi), %%esi \n"
    "    movdqu   %%xmm0,(%%edi)\n"
    "    lea      16(%%edi), %%edi\n"
    "    sub      $16, %%eax\n"
    "    cmp      $16, %%eax\n"
    "    jge      ssecalc\n"
    "    cmp      %%ebx, %%esi\n"
    "    je       end\n"

    " normalcalc:\n"
    "    mov      (%%esi), %%al\n"
    "    lea      1(%%esi), %%esi\n"
    "    mov      %%al, (%%edi)\n"
    "    lea      1(%%edi), %%edi\n"
    "    cmp      %%ebx, %%esi\n"
    "    jne      normalcalc\n"

    " end:\n"
    "    emms" :: "r" (size), "S" (src), "D" (dst));
}