代码

#include "router.h"
#include <stdlib.h>
#include <string.h>
#include <stdint.h>

// /20 cut + FULL /24 u32 table (directly indexed, lazily materialized).
//   A[2^20] u32 = /20 answer (raw nexthop, 0 = no match)
//   hasDeep bitset = /20 slice has any len>=21 entry
//   L2[2^24] u32 = /24 answer (raw nexthop); only deep /20's 16 /24s written.
//   len>=25: L2 slot = SENTINEL -> per-/24 L3 (binary search, rare).
// Deep read L2[ip>>8] is DIRECT (no offset dependency) -> MLP-friendly.
#define SENTINEL 0xFFFFFFFFu

static uint32_t *A;
static uint8_t  *hasDeep;
static uint32_t *L2;
static uint32_t *L3;
static uint32_t *l3key;
static uint32_t *l3off;
static uint32_t l3cnt;

void init(int n, int q, const RoutingTableEntry *a){
    (void)q;
    A = (uint32_t*)calloc(1u << 20, 4);
    hasDeep = (uint8_t*)calloc((1u << 20) >> 3, 1);
    L2 = (uint32_t*)malloc((size_t)(1u << 24) * 4);
    L3 = (uint32_t*)malloc((size_t)n * 256 * 4);
    l3key = (uint32_t*)malloc((size_t)n * 4);
    l3off = (uint32_t*)malloc((size_t)n * 4);
    l3cnt = 0;

    uint32_t l3cur = 0;

    for (int i = 0; i < n; i++) {
        uint32_t v = __builtin_bswap32(a[i].addr);
        uint32_t len = a[i].len;
        uint32_t nh = a[i].nexthop;
        if (len <= 20) {
            uint32_t shift = 20 - len;
            uint32_t count = 1u << shift;
            uint32_t mask = count - 1;
            uint32_t base = (v >> 12) & ~mask;
            uint32_t end = base | mask;
            for (uint32_t x = base; x <= end; x++) A[x] = nh;
        } else {
            uint32_t s = v >> 12;
            uint32_t idx = s >> 3;
            uint32_t bit = 1u << (s & 7);
            if (!(hasDeep[idx] & bit)) {
                hasDeep[idx] |= bit;
                uint32_t base24 = s << 4;
                for (uint32_t c = 0; c < 16; c++) L2[base24 + c] = A[s];
            }
            uint32_t cell = (v >> 8) & 15;
            uint32_t i24 = (s << 4) + cell;
            if (len <= 24) {
                uint32_t shift = 24 - len;
                uint32_t count = 1u << shift;
                uint32_t mask = count - 1;
                uint32_t base = cell & ~mask;
                uint32_t end = base | mask;
                for (uint32_t x = base; x <= end; x++) L2[(s << 4) + x] = nh;
            } else {
                uint32_t *e = &L2[i24];
                if (*e != SENTINEL) {
                    uint32_t l3 = l3cur; l3cur += 256;
                    for (uint32_t c = 0; c < 256; c++) L3[l3 + c] = *e;
                    *e = SENTINEL;
                    l3key[l3cnt] = v >> 8;
                    l3off[l3cnt] = l3;
                    l3cnt++;
                }
                uint32_t l3 = l3off[l3cnt - 1];
                uint32_t shift = 32 - len;
                uint32_t count = 1u << shift;
                uint32_t mask = count - 1;
                uint32_t base = v & ~mask;
                uint32_t end = base | mask;
                for (uint32_t x = base; x <= end; x++) L3[l3 + (x & 255)] = nh;
            }
        }
    }
}

static inline uint32_t lookup(uint32_t h){
    uint32_t s = h >> 12;
    uint32_t ans = A[s];
    if (hasDeep[s >> 3] & (1u << (s & 7))) {
        ans = L2[h >> 8];
        if (ans == SENTINEL) {
            uint32_t k = h >> 8;
            uint32_t lo = 0, hi = l3cnt;
            while (lo < hi) {
                uint32_t mid = (lo + hi) >> 1;
                if (l3key[mid] < k) lo = mid + 1; else hi = mid;
            }
            ans = L3[l3off[lo] + (h & 255)];
        }
    }
    return ans;
}

unsigned query(unsigned addr){ return lookup(__builtin_bswap32(addr)); }

评测结果