代码

void my_memset(char *a, char c, unsigned n) {
	char s[16]={c,c,c,c,c,c,c,c,c,c,c,c,c,c,c,c}; while(n&127)a[--n]=c;
	__asm("movups (%1), %%xmm0; _LAR: movaps %%xmm0, (%0); movaps %%xmm0, 16(%0); movaps %%xmm0, 32(%0); movaps %%xmm0, 48(%0); movaps %%xmm0, 64(%0); movaps %%xmm0, 80(%0); movaps %%xmm0, 96(%0); movaps %%xmm0, 112(%0); addl $128, %0; subl $1, %2; jnz _LAR; ": : "r"(a), "r"(&s), "r"(n/128): "memory");
}

评测结果