代码

#include <cstdlib>
#include <cstdio>

unsigned cnt1[256], cnt2[256], cnt3[256], cnt4[256], b[100000008], a[100000008];

void sort(unsigned *a, int n) {
	for (register int i = 0; i < n; i++) {
		cnt1[(a[i]      ) & 255]++;
		cnt2[(a[i] >> 8 ) & 255]++;
		cnt3[(a[i] >> 16) & 255]++;
		cnt4[(a[i] >> 24) & 255]++;
	}
	for (register int i = 1; i < 256; i++) {
		cnt1[i] += cnt1[i-1];
		cnt2[i] += cnt2[i-1];
		cnt3[i] += cnt3[i-1];
		cnt4[i] += cnt4[i-1];
	}
	for (register int i = n-1; ~i; i--) b[--cnt1[(a[i]      ) & 255]] = a[i];
	for (register int i = n-1; ~i; i--) a[--cnt2[(b[i] >> 8 ) & 255]] = b[i];
	for (register int i = n-1; ~i; i--) b[--cnt3[(a[i] >> 16) & 255]] = a[i];
	for (register int i = n-1; ~i; i--) a[--cnt4[(b[i] >> 24) & 255]] = b[i];
}

评测结果