void
lshift_s16_mmx_2(int16_t *d1, int16_t *s1, int16_t *s3_1, int n)
{
  while (n&7) {
    d1[0] = s1[0]<<s3_1[0];
    d1++;
    s1++;
    n--;
  }
  n >>= 3;
  if (n == 0) return;
  __asm__ __volatile__ ("\n"
      "  movzwl 0(%[s3_1]), %%ecx\n"
      "  movd %%ecx, %%mm1\n"
      "1:\n"
      "  movq 0(%[s1]), %%mm0\n"
      "  psllw %%mm1, %%mm0\n"
      "  movq %%mm0, 0(%[d1])\n"
      "  movq 8(%[s1]), %%mm0\n"
      "  psllw %%mm1, %%mm0\n"
      "  movq %%mm0, 8(%[d1])\n"
      "  add $16, %[d1]\n"
      "  add $16, %[s1]\n"
      "  decl %[n]\n"
      "  jnz 1b\n"
      "  emms"
      : [d1] "+r" (d1),
        [s1] "+r" (s1),
        [n] "+r" (n)
      : [s3_1] "r" (s3_1)
      : "ecx");
}
OIL_DEFINE_IMPL_FULL (lshift_s16_mmx_2, lshift_s16, OIL_IMPL_FLAG_MMX);