代码

#pragma GCC optimize("Ofast,-funroll-loops")
#pragma target("sse,sse2,tune=native")
#include<cstdio>
#include<immintrin.h>
#include<emmintrin.h>
#include<xmmintrin.h>
void my_memset(char *a, char c, unsigned n){
	n--;
	while((n&15)!=15&&n>=0)a[n]=c,n--;
	if(n==-1)return ;
	__m128 fa;
	for(int i=0;i<16;++i)((char*)&fa)[i]=c;
	__m64* _a=(__m64*)a;
	for(int i=0;i<n;i+=16){
		_mm_storeh_pi(_a,fa);
		_a++;
		_mm_storeh_pi(_a,fa);
		_a++;
	}
}

评测结果