代码

#include <stdint.h>
#include <arpa/inet.h>
#include <algorithm>
#include "router.h"

/* HashMap */

const int HASHMAP_SIZE = 1 << 16;
const uint32_t HASHMAP_MOD = 65537;

uint32_t hashmap_key[HASHMAP_SIZE];
uint32_t hashmap_next[HASHMAP_SIZE];
uint32_t hashmap_first[HASHMAP_MOD];
int hashmap_size;

inline uint32_t hashmap_get(uint32_t x) {
	uint32_t hash = x % HASHMAP_MOD;
	uint32_t &first_entry_id = hashmap_first[hash];
	uint32_t entry_id = first_entry_id;
	while (1) {
		const uint32_t &key = hashmap_key[entry_id];
		if (key == x) {
			return entry_id;
		} else if (key) {
			entry_id = hashmap_next[entry_id];
		} else {
			break;
		}
	}
	++hashmap_size;
	hashmap_key[hashmap_size] = x;
	hashmap_next[hashmap_size] = first_entry_id;
	first_entry_id = hashmap_size;
}

/* 3-level tree */

const uint32_t TABLE_32_SIZE = 16384;
const uint32_t TABLE_24_SIZE = 32768;

uint32_t table_32_cnt = 0;
uint32_t table_24_cnt = 0;
uint32_t table_32[TABLE_32_SIZE][1 << 8] __attribute__((aligned(4096)));
uint32_t table_24[TABLE_24_SIZE][1 << 8] __attribute__((aligned(4096)));
uint32_t table_16[1 << 16] __attribute__((aligned(4096)));

inline void fill(uint32_t *a, int n, uint32_t val) {
	while (n >= 4) {
		a[0] = val;
		a[1] = val;
		a[2] = val;
		a[3] = val;
		n -= 4;
		a += 4;
	}
	while (n) {
		a[0] = val;
		n--;
		a++;
	}
}

inline void ins(uint32_t addr, int len, uint32_t nexthop) {
	if (len <= 16) {
		fill(table_16 + (addr >> 16), 1u << (16 - len), nexthop);
	} else if (len <= 24) {
		uint32_t &t16 = table_16[addr >> 16];
		addr = (addr & 65535u) >> 8;
		
		uint32_t *tmp;
		if (t16 < -TABLE_24_SIZE) {
			tmp = table_24[--table_24_cnt + TABLE_24_SIZE];
			fill(tmp, addr, t16);
			fill(tmp + addr, 1u << (24 - len), nexthop);
			fill(tmp + addr + (1u << (24 - len)), 256 - addr - (1u << (24 - len)), t16);
			t16 = table_24_cnt;
		} else {
			tmp = table_24[t16 + TABLE_24_SIZE];
			fill(tmp + addr, 1u << (24 - len), nexthop);
		}
	} else {
		uint32_t &t16 = table_16[addr >> 16];
		addr &= 65535u;
		
		uint32_t *tmp;
		if (t16 < -TABLE_24_SIZE) {
			tmp = table_24[--table_24_cnt + TABLE_24_SIZE];
			fill(tmp, 256, t16);
			t16 = table_24_cnt;
		} else {
			tmp = table_24[t16 + TABLE_24_SIZE];
		}
		
		uint32_t &t24 = tmp[addr >> 8];
		addr &= 255u;
		
		if (t24 < -TABLE_32_SIZE) {
			tmp = table_32[--table_32_cnt + TABLE_32_SIZE];
			fill(tmp, addr, t24);
			fill(tmp + addr, 1u << (32 - len), nexthop);
			fill(tmp + addr + (1u << (32 - len)), 256 - addr - (1u << (32 - len)), t24);
			t24 = table_32_cnt;
		} else {
			tmp = table_32[t24 + TABLE_32_SIZE];
			fill(tmp + addr, 1u << (32 - len), nexthop);
		}
	}
}

/* Bit sets */

const int MAX_N_LEVEL3_POINTERS = 1000000;
uint32_t level3_bits[TABLE_32_SIZE][(1 << 8) / 32];
uint32_t level3_bit_sums[TABLE_32_SIZE][(1 << 8) / 32];
uint16_t level3_offsets[TABLE_32_SIZE];
uint16_t level3_pointers[MAX_N_LEVEL3_POINTERS];
uint32_t n_level3_pointers;

inline void walk_level3(uint32_t level3_index) {
	uint32_t *tmp = table_32[level3_index];
	uint32_t *bits = level3_bits[level3_index];
	uint32_t *bit_sums = level3_bit_sums[level3_index];
	level3_offsets[level3_index] = n_level3_pointers;
	uint16_t *pointers = level3_pointers + n_level3_pointers;
	
	uint32_t cur = tmp[0];
	int cnt = 0;
	bit_sums[0] = 0;
	
	// index 0
	pointers[cnt++] = cur;
	
	for (int i = 1; i < 1 << 8; i++) {
		if (i % 32 == 0) {
			bit_sums[i / 32] = cnt;
		}
		if (tmp[i] != cur) {
			bits[i >> 5] |= 1u << (i & 31);
			pointers[cnt++] = cur;
			cur = tmp[i];
		}
	}
	
	n_level3_pointers += cnt;
}

const int MAX_N_LEVEL2_POINTERS = 1000000;
uint32_t level2_bits[TABLE_24_SIZE][(1 << 8) / 32];
uint32_t level2_bit_sums[TABLE_24_SIZE][(1 << 8) / 32];
uint16_t level2_offsets[TABLE_24_SIZE];
uint16_t level2_pointers[MAX_N_LEVEL2_POINTERS];
uint32_t n_level2_pointers;

inline void walk_level2(uint32_t level2_index) {
	uint32_t *tmp = table_24[level2_index];
	uint32_t *bits = level2_bits[level2_index];
	uint32_t *bit_sums = level2_bit_sums[level2_index];
	level2_offsets[level2_index] = n_level2_pointers;
	uint16_t *pointers = level2_pointers + n_level2_pointers;
	
	uint32_t cur = tmp[0];
	int cnt = 0;
	bit_sums[0] = 0;
	
	// index 0
	pointers[cnt++] = cur >= -TABLE_32_SIZE ? cur : hashmap_get(cur);
	
	for (int i = 1; i < 1 << 8; i++) {
		if (i % 32 == 0) {
			bit_sums[i / 32] = cnt;
		}
		if (tmp[i] != cur) {
			bits[i >> 5] |= 1u << (i & 31);
			pointers[cnt++] = cur >= -TABLE_32_SIZE ? cur : hashmap_get(cur);
			cur = tmp[i];
		}
	}
	
	n_level2_pointers += cnt;
	
	for (int i = 0; i < 1 << 8; i++) {
		if (tmp[i] >= -TABLE_32_SIZE) {
			walk_level3(tmp[i] + TABLE_32_SIZE);
		}
	}
}

inline void walk_level1() {
	for (uint32_t i = 0; i < 1 << 16; i++) {
		uint32_t t16 = table_16[i];
		if (t16 >= -TABLE_24_SIZE) {
			walk_level2(t16 + TABLE_24_SIZE);
		}
	}
}

void init(int n, int q, const RoutingTableEntry *tbl) {
	for (int i = 0; i < n; i++) {
		ins(htonl(tbl[i].addr), tbl[i].len, tbl[i].nexthop);
	}
	
	walk_level1();
}

unsigned query(unsigned addr) {
	addr = htonl(addr);
	
	uint32_t tmp = table_16[addr >> 16];
	if (tmp >= -TABLE_24_SIZE) {
		uint32_t level2_index = tmp + TABLE_24_SIZE;
		uint32_t addr_l2 = addr << 16 >> 24;
		uint32_t off = level2_offsets[level2_index] + level2_bit_sums[level2_index][addr_l2 / 32]
			+ __builtin_popcount(level2_bits[level2_index][addr_l2 / 32] & ((2u << (addr_l2 & 31)) - 1));
		
		tmp = (int32_t) (int16_t) level2_pointers[off];
		if (tmp >= -TABLE_32_SIZE) {
			uint32_t level3_index = tmp + TABLE_32_SIZE;
			uint32_t addr_l3 = addr & 255u;
			uint32_t off = level3_offsets[level3_index] + level3_bit_sums[level3_index][addr_l3 / 32]
				+ __builtin_popcount(level3_bits[level3_index][addr_l3 / 32] & ((2u << (addr_l3 & 31)) - 1));
			
			tmp = (int32_t) (int16_t) level3_pointers[off];
			return hashmap_key[tmp];
		} else {
			return hashmap_key[tmp];
		}
	} else {
		return tmp;
	}
}

评测结果