代码

#include <bits/stdc++.h>

typedef unsigned int uint;

void sort(uint *a, int n) {
	uint t[100000000];
	register uint b0[256], b1[256], b2[256], b3[256], *ed = a + n;
	for (register uint *i = a; i != ed; i++) {
		b0[*i & 255]++, b1[(*i >> 8) & 255]++, b2[(*i >> 16) & 255]++, b3[*i >> 24]++;
	}
	for (register int i = 1; i < 256; i++) {
		b0[i] += b0[i - 1], b1[i] += b1[i - 1], b2[i] += b2[i - 1], b3[i] += b3[i - 1];
	}
	for (register uint *i = a + n - 1; i != a; i -= 8) {
		t[--b0[i[-0] & 255]] = i[-0];
		t[--b0[i[-1] & 255]] = i[-1];
		t[--b0[i[-2] & 255]] = i[-2];
		t[--b0[i[-3] & 255]] = i[-3];
		t[--b0[i[-4] & 255]] = i[-4];
		t[--b0[i[-5] & 255]] = i[-5];
		t[--b0[i[-6] & 255]] = i[-6];
		t[--b0[i[-7] & 255]] = i[-7];
	}
	for (register uint *i = t + n - 1; i != t; i -= 8) {
		a[--b1[(i[-0] >> 8) & 255]] = i[-0];
		a[--b1[(i[-1] >> 8) & 255]] = i[-1];
		a[--b1[(i[-2] >> 8) & 255]] = i[-2];
		a[--b1[(i[-3] >> 8) & 255]] = i[-3];
		a[--b1[(i[-4] >> 8) & 255]] = i[-4];
		a[--b1[(i[-5] >> 8) & 255]] = i[-5];
		a[--b1[(i[-6] >> 8) & 255]] = i[-6];
		a[--b1[(i[-7] >> 8) & 255]] = i[-7];
	}
	for (register uint *i = a + n - 1; i != a; i -= 8) {
		t[--b2[(i[-0] >> 16) & 255]] = i[-0];
		t[--b2[(i[-1] >> 16) & 255]] = i[-1];
		t[--b2[(i[-2] >> 16) & 255]] = i[-2];
		t[--b2[(i[-3] >> 16) & 255]] = i[-3];
		t[--b2[(i[-4] >> 16) & 255]] = i[-4];
		t[--b2[(i[-5] >> 16) & 255]] = i[-5];
		t[--b2[(i[-6] >> 16) & 255]] = i[-6];
		t[--b2[(i[-7] >> 16) & 255]] = i[-7];
	}
	for (register uint *i = t + n - 1; i != t; i -= 8) {
		a[--b3[(i[-0] >> 24)]] = i[-0];
		a[--b3[(i[-1] >> 24)]] = i[-1];
		a[--b3[(i[-2] >> 24)]] = i[-2];
		a[--b3[(i[-3] >> 24)]] = i[-3];
		a[--b3[(i[-4] >> 24)]] = i[-4];
		a[--b3[(i[-5] >> 24)]] = i[-5];
		a[--b3[(i[-6] >> 24)]] = i[-6];
		a[--b3[(i[-7] >> 24)]] = i[-7];
	}
}

评测结果