代码

unsigned b[100000000],*rs[0x100],t0[0x100],t1[0x100],t2[0x100],t3[0x100];
void sort(unsigned *a, int n){
    #define cal(w,w2,tw,op)\
    p=w2-1;\
    for(i=0;i<0x100;++i)rs[i]=p,p+=tw[i];\
    for(i=0;i<100000000;i+=8){\
    	p=w+i;\
    	*++rs[p[0]op]=p[0];\
    	*++rs[p[1]op]=p[1];\
    	*++rs[p[2]op]=p[2];\
    	*++rs[p[3]op]=p[3];\
    	*++rs[p[4]op]=p[4];\
    	*++rs[p[5]op]=p[5];\
    	*++rs[p[6]op]=p[6];\
    	*++rs[p[7]op]=p[7];\
    }
    register unsigned*p,i,x;
    for(i=0;i<100000000;){
    #define A x=a[i];++t0[x&0xFF];++t1[x>>8&0xFF];++t2[x>>16&0xFF];++t3[x>>24];++i;
    A A A A
    }
    cal(a,b,t0,&0xFF);
    cal(b,a,t1,>>8&0xFF);
    cal(a,b,t2,>>16&0xFF);
    cal(b,a,t3,>>24);
}

评测结果