代码

// phase probe: only the fixed-stride top-byte scatter (deliberately WA)
#include <stdlib.h>
#include <string.h>
static unsigned fill[256];
void sort(unsigned *a, int n) {
    unsigned stride = (unsigned)((n + 255) / 256) + (unsigned)(n >> 9) + 4096;
    unsigned *tmp = (unsigned *)malloc((size_t)stride * 256u * 4u);
    if (!tmp) return;
    memset(fill, 0, sizeof(fill));
    for (int i = 0; i < n; i++) {
        unsigned v = a[i];
        unsigned b = v >> 24;
        tmp[(size_t)b * stride + fill[b]] = v;
        fill[b]++;
    }
    /* leave a untouched -> WA, we only want the timing */
}

评测结果