代码

#include <string.h>
typedef unsigned int u32;

static u32 b[200000000];
static u32 cnt[1024];

void sort(unsigned *a, int n) {
  u32 *c0 = cnt, *c1 = cnt + 256, *c2 = cnt + 512, *c3 = cnt + 768;
  u32 *t = b;

  memset(c0, 0, 256*4);
  for (int i = 0; i < n; i++) c0[a[i] & 255]++;
  u32 s = 0;
  for (int i = 0; i < 256; i++) { u32 v = c0[i]; c0[i] = s; s += v; }

  memset(c1, 0, 256*4);
  int i = 0;
  for (; i + 1 < n; i += 2) {
    u32 x0 = a[i], x1 = a[i+1];
    t[c0[x0 & 255]++] = x0;
    c1[(x0 >> 8) & 255]++;
    t[c0[x1 & 255]++] = x1;
    c1[(x1 >> 8) & 255]++;
  }
  for (; i < n; i++) { u32 x = a[i]; t[c0[x & 255]++] = x; c1[(x >> 8) & 255]++; }
  s = 0;
  for (int j = 0; j < 256; j++) { u32 v = c1[j]; c1[j] = s; s += v; }

  memset(c2, 0, 256*4);
  i = 0;
  for (; i + 1 < n; i += 2) {
    u32 x0 = t[i], x1 = t[i+1];
    a[c1[(x0 >> 8) & 255]++] = x0;
    c2[(x0 >> 16) & 255]++;
    a[c1[(x1 >> 8) & 255]++] = x1;
    c2[(x1 >> 16) & 255]++;
  }
  for (; i < n; i++) { u32 x = t[i]; a[c1[(x >> 8) & 255]++] = x; c2[(x >> 16) & 255]++; }
  s = 0;
  for (int j = 0; j < 256; j++) { u32 v = c2[j]; c2[j] = s; s += v; }

  memset(c3, 0, 256*4);
  i = 0;
  for (; i + 1 < n; i += 2) {
    u32 x0 = a[i], x1 = a[i+1];
    t[c2[(x0 >> 16) & 255]++] = x0;
    c3[x0 >> 24]++;
    t[c2[(x1 >> 16) & 255]++] = x1;
    c3[x1 >> 24]++;
  }
  for (; i < n; i++) { u32 x = a[i]; t[c2[(x >> 16) & 255]++] = x; c3[x >> 24]++; }
  s = 0;
  for (int j = 0; j < 256; j++) { u32 v = c3[j]; c3[j] = s; s += v; }

  for (int k = 0; k < n; k++) { u32 x = t[k]; a[c3[x >> 24]++] = x; }
}

评测结果