// NOI2017 蚯蚓排队 (noi17b)
// k<=50 hash tables (unified, key = (k, base-6 value mod 2^128)), open addressing, power-of-2.
#include <sys/auxv.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
typedef unsigned __int128 u128;
struct DuckInfo {
uint64_t abi_version;
const char *stdin_ptr; uint64_t stdin_size;
char *stdout_ptr; uint64_t stdout_limit; uint64_t stdout_size;
char *stderr_ptr; uint64_t stderr_limit; uint64_t stderr_size;
const char *IB_ptr; uint64_t IB_limit;
char *OB_ptr; uint64_t OB_limit;
uint64_t tsc_frequency;
} __attribute__((packed));
// ---- worm linked list ----
static int n, m;
static int *prv, *nxt, *dg; // dg = length-1 (0..5)
// ---- hash table ----
struct Slot {
uint64_t lo, hi;
uint32_t cnt;
uint8_t k, state; // state: 0 empty, 1 occupied, 2 tombstone
};
#define TABSZ (1u<<22)
#define TABMASK (TABSZ-1)
static Slot tbl[TABSZ];
static uint64_t kc[51]; // k*CONST, precomputed
static u128 pow6[50]; // 6^0 .. 6^49
static const uint64_t H1 = 0x9E3779B97F4A7C15ull;
static inline uint32_t hash_idx(uint64_t lo, uint64_t hi, uint64_t kseed) {
uint64_t h = lo ^ (hi * H1);
h += kseed;
return (uint32_t)(h & TABMASK);
}
static inline void tinc(u128 val, uint32_t k) {
uint64_t lo = (uint64_t)val;
uint64_t hi = (uint64_t)(val >> 64);
uint64_t h = lo ^ (hi * H1) + kc[k];
uint32_t idx = (uint32_t)(h & TABMASK);
uint32_t first_del = 0xFFFFFFFFu;
for (;;) {
Slot &s = tbl[idx];
if (s.state == 1 && s.lo == lo && s.hi == hi && s.k == k) { s.cnt++; return; }
if (s.state == 2 && first_del == 0xFFFFFFFFu) first_del = idx;
if (s.state == 0) {
if (first_del != 0xFFFFFFFFu) idx = first_del;
Slot &t = tbl[idx];
t.lo = lo; t.hi = hi; t.k = (uint8_t)k; t.cnt = 1; t.state = 1;
return;
}
idx = (idx + 1) & TABMASK;
}
}
static inline void tdec(u128 val, uint32_t k) {
uint64_t lo = (uint64_t)val;
uint64_t hi = (uint64_t)(val >> 64);
uint64_t h = lo ^ (hi * H1) + kc[k];
uint32_t idx = (uint32_t)(h & TABMASK);
for (;;) {
Slot &s = tbl[idx];
if (s.state == 0) return;
if (s.state == 1 && s.lo == lo && s.hi == hi && s.k == k) {
if (--s.cnt == 0) s.state = 2;
return;
}
idx = (idx + 1) & TABMASK;
}
}
static inline uint32_t tlookup(u128 val, uint32_t k, uint64_t kseed) {
uint64_t lo = (uint64_t)val;
uint64_t hi = (uint64_t)(val >> 64);
uint64_t h = lo ^ (hi * H1) + kseed;
uint32_t idx = (uint32_t)(h & TABMASK);
for (;;) {
Slot &s = tbl[idx];
if (s.state == 0) return 0;
if (s.state == 1 && s.lo == lo && s.hi == hi && s.k == k) return s.cnt;
idx = (idx + 1) & TABMASK;
}
}
// collect A-suffix (from i backward, up to 49) and B-prefix (from j forward, up to 49)
static inline void collect(int i, int j, uint8_t *sa, int *na, uint8_t *sb, int *nb) {
int x = i;
sa[0] = (uint8_t)dg[i]; *na = 1;
while (*na < 49 && prv[x]) { x = prv[x]; sa[(*na)++] = (uint8_t)dg[x]; }
x = j;
sb[0] = (uint8_t)dg[j]; *nb = 1;
while (*nb < 49 && nxt[x]) { x = nxt[x]; sb[(*nb)++] = (uint8_t)dg[x]; }
}
static u128 RA[50];
static inline void do_merge(int i, int j) {
uint8_t sa[50], sb[50];
int na, nb;
collect(i, j, sa, &na, sb, &nb);
u128 ra = 0, p6 = 1;
for (int t = 0; t < na; t++) { ra += (u128)sa[t] * p6; RA[t] = ra; p6 *= 6; }
int dmax = na - 1; if (dmax > 48) dmax = 48;
for (int d = 0; d <= dmax; d++) {
u128 val = RA[d];
int mmax = 49 - d; if (mmax > nb) mmax = nb;
for (int m2 = 1; m2 <= mmax; m2++) {
val = val * 6 + sb[m2 - 1];
tinc(val, (uint32_t)(d + 1 + m2));
}
}
nxt[i] = j; prv[j] = i;
}
static inline void do_split(int i) {
int j = nxt[i];
uint8_t sa[50], sb[50];
int na, nb;
collect(i, j, sa, &na, sb, &nb);
u128 ra = 0, p6 = 1;
for (int t = 0; t < na; t++) { ra += (u128)sa[t] * p6; RA[t] = ra; p6 *= 6; }
int dmax = na - 1; if (dmax > 48) dmax = 48;
for (int d = 0; d <= dmax; d++) {
u128 val = RA[d];
int mmax = 49 - d; if (mmax > nb) mmax = nb;
for (int m2 = 1; m2 <= mmax; m2++) {
val = val * 6 + sb[m2 - 1];
tdec(val, (uint32_t)(d + 1 + m2));
}
}
nxt[i] = 0; prv[j] = 0;
}
static char *optr;
static const uint64_t MOD = 998244353ull;
static inline void out_u64(uint64_t x) {
char tmp[24]; int len = 0;
if (x == 0) { *optr++ = '0'; *optr++ = '\n'; return; }
while (x) { tmp[len++] = (char)('0' + (x % 10)); x /= 10; }
while (len) *optr++ = tmp[--len];
*optr++ = '\n';
}
static inline void query(const char *s, uint32_t slen, uint32_t k) {
uint64_t ans = 1;
uint64_t kseed = kc[k];
u128 powk = pow6[k - 1];
u128 val = 0;
for (uint32_t i = 0; i < k; i++) {
int c = s[i] - '1';
if (c < 0 || c > 5) { *optr++ = '0'; *optr++ = '\n'; return; }
val = val * 6 + (u128)c;
}
uint32_t cnt = tlookup(val, k, kseed);
ans = ans * cnt % MOD;
if (ans == 0) { *optr++ = '0'; *optr++ = '\n'; return; }
for (uint32_t i = k; i < slen; i++) {
int c = s[i] - '1';
if (c < 0 || c > 5) { *optr++ = '0'; *optr++ = '\n'; return; }
u128 old = (u128)(s[i - k] - '1');
val = (val - old * powk) * 6 + (u128)c;
cnt = tlookup(val, k, kseed);
ans = ans * cnt % MOD;
if (ans == 0) { *optr++ = '0'; *optr++ = '\n'; return; }
}
out_u64(ans);
}
static inline uint32_t rd_u32(const char *&p) {
while (*p < '0' || *p > '9') p++;
uint32_t v = 0;
while (*p >= '0' && *p <= '9') { v = v * 10 + (uint32_t)(*p - '0'); p++; }
return v;
}
int main() {
unsigned long duck = getauxval(0x6b637564UL);
const char *inp; uint64_t insz;
char *outp; uint64_t outlim;
DuckInfo *di = 0;
bool use_duck = (duck != 0);
static char local_in[1u<<26];
static char local_out[1u<<21];
if (use_duck) {
di = (DuckInfo*)duck;
inp = di->stdin_ptr; insz = di->stdin_size;
outp = di->stdout_ptr; outlim = di->stdout_limit;
} else {
insz = fread(local_in, 1, sizeof(local_in), stdin);
inp = local_in;
outp = local_out; outlim = sizeof(local_out);
}
(void)outlim;
optr = outp;
const char *p = inp;
n = (int)rd_u32(p);
m = (int)rd_u32(p);
prv = (int*)malloc((size_t)(n + 1) * sizeof(int));
nxt = (int*)malloc((size_t)(n + 1) * sizeof(int));
dg = (int*)malloc((size_t)(n + 1) * sizeof(int));
memset(prv, 0, (size_t)(n + 1) * sizeof(int));
memset(nxt, 0, (size_t)(n + 1) * sizeof(int));
for (int i = 1; i <= n; i++) { uint32_t L = rd_u32(p); dg[i] = (int)(L - 1); }
for (int k = 1; k <= 50; k++) kc[k] = (uint64_t)k * 0xBF58476D1CE4E5B9ull;
pow6[0] = 1;
for (int i = 1; i < 50; i++) pow6[i] = pow6[i - 1] * 6;
// initial k=1 counts
for (int i = 1; i <= n; i++) tinc((u128)dg[i], 1);
for (int op = 0; op < m; op++) {
uint32_t t = rd_u32(p);
if (t == 1) {
uint32_t i = rd_u32(p), j = rd_u32(p);
do_merge((int)i, (int)j);
} else if (t == 2) {
uint32_t i = rd_u32(p);
do_split((int)i);
} else {
while (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t') p++;
const char *s = p;
while (*p >= '0' && *p <= '9') p++;
uint32_t slen = (uint32_t)(p - s);
uint32_t k = rd_u32(p);
query(s, slen, k);
}
}
if (use_duck) {
di->stdout_size = (uint64_t)(optr - outp);
__asm__ volatile("mov $60, %%rax; xor %%rdi, %%rdi; syscall" ::: "rax","rdi","rcx","r11","memory");
} else {
fwrite(outp, 1, (size_t)(optr - outp), stdout);
}
return 0;
}
| Compilation | N/A | N/A | Compile OK | Score: N/A | 显示更多 |
| Testcase #1 | 21.39 us | 32 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #2 | 46.22 us | 268 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #3 | 9.324 ms | 4 MB + 888 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #4 | 1.304 ms | 9 MB + 1020 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #5 | 6.486 ms | 13 MB + 20 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #6 | 104.747 ms | 96 MB + 616 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #7 | 13.231 ms | 820 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #8 | 72.248 ms | 96 MB + 608 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #9 | 108.769 ms | 96 MB + 608 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #10 | 134.123 ms | 96 MB + 696 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #11 | 217.188 ms | 96 MB + 696 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #12 | 121.569 ms | 97 MB + 188 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #13 | 27.373 ms | 1 MB + 388 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #14 | 161.929 ms | 97 MB + 176 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #15 | 193.806 ms | 97 MB + 176 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #16 | 331.829 ms | 97 MB + 360 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #17 | 460.308 ms | 97 MB + 356 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #18 | 668.712 ms | 100 MB + 188 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #19 | 1.107 s | 100 MB + 188 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #20 | 109.28 ms | 98 MB + 352 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #21 | 56.496 ms | 2 MB + 540 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #22 | 490.15 ms | 98 MB + 324 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #23 | 508.275 ms | 98 MB + 324 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #24 | 1.444 s | 98 MB + 728 KB | Accepted | Score: 4 | 显示更多 |
| Testcase #25 | 2 s | 98 MB + 692 KB | Time Limit Exceeded | Score: 0 | 显示更多 |