代码

unsigned b[100000000], c0[256], c1[256], c2[256], c3[256];
unsigned *pos[256];

#define gao(cnt,op) \
		int sum=-1;\
		for (i=0;i<256;++i) pos[i]=B+sum, sum+=cnt[i];\
		unsigned *now=A;\
		for (i=0;i<n;i+=4,now+=4){\
			*(++pos[A[i] op])=*now;\
			*(++pos[A[i+1] op])=now[1];\
			*(++pos[A[i+2] op])=now[2];\
			*(++pos[A[i+3] op])=now[3];\
		}\
		T=A; A=B; B=T;

void songsongsong(unsigned *a,unsigned n){
	{
		register unsigned i;
		for (i=0;i<n;++i){
			++c0[a[i]&255];
			++c1[a[i]>>8&255];
			++c2[a[i]>>16&255];
			++c3[a[i]>>24&255];
		}
	}
	
	unsigned *A=a, *B=b, *T;
	
	{
		register unsigned i;
		gao(c0,&255);
	}
	{
		register unsigned i;
		gao(c1,>>8&255);
	}
	{
		register unsigned i;
		gao(c2,>>16&255);
	}
	{
		register unsigned i;
		gao(c3,>>24&255);
	}
}

void sort(unsigned *a, int n)
{
	songsongsong(a,n);
}

评测结果