CP-Algorithms Library
This documentation is automatically generated by competitive-verifier/competitive-verifier
Suffix Array (verify/string/suffix_array.test.cpp)
- View this file on GitHub
- Last update: 2026-09-13 14:14:06+02:00
- Problem: https://judge.yosupo.jp/problem/suffixarray
Depends on
Code
// @brief Suffix Array
#define PROBLEM "https://judge.yosupo.jp/problem/suffixarray"
#pragma GCC optimize("O3,unroll-loops")
#define CP_ALGO_CHECKPOINT
#include "cp-algo/util/checkpoint.hpp"
#include <iostream>
#include "blazingio/blazingio.min.hpp"
#include <algorithm>
#include <array>
#include <cassert>
#include <cstdint>
#include <numeric>
#include <vector>
int main() {
std::ios::sync_with_stdio(false);
std::cin.tie(nullptr);
std::string s;
std::cin >> s;
std::ranges::reverse(s);
int n = (int)s.size();
// Ordinary state i represents the prefix of length i; clones follow state n.
struct clone_data { int len, pos; };
std::vector<clone_data> clones;
clones.reserve(n);
cp_algo::big_vector<int> link(2 * n + 1);
auto length = [&](int v) { return v <= n ? v : clones[v - n - 1].len; };
auto position = [&](int v) { return v <= n ? v : clones[v - n - 1].pos; };
// Pack two (letter + 1, state) pairs inline; bit 31 marks a dense row.
// State IDs fit in 24 bits under the problem's n <= 500000 bound.
struct transitions { uint32_t a = 0, b = 0; };
cp_algo::big_vector<std::array<int, 26>> dense;
// At most one dense row per state; avoid growth on branching structured strings.
dense.reserve(link.size());
cp_algo::big_vector<transitions> to;
to.reserve(link.size());
to.resize(n + 1);
auto get = [&](int p, int x) {
auto t = to[p];
if(t.a & 0x80000000) {return dense[t.a & 0x7FFFFFFF][x];}
if((t.a >> 24) == unsigned(x + 1)) {return int(t.a & 0xFFFFFF);}
if((t.b >> 24) == unsigned(x + 1)) {return int(t.b & 0xFFFFFF);}
return 0;
};
auto set = [&](int p, int x, int v) {
auto &t = to[p];
auto encoded = uint32_t((x + 1) << 24) | v;
if(t.a & 0x80000000) {dense[t.a & 0x7FFFFFFF][x] = v;}
else if(!t.a || (t.a >> 24) == unsigned(x + 1)) {t.a = encoded;}
else if(!t.b || (t.b >> 24) == unsigned(x + 1)) {t.b = encoded;}
else {
std::array<int, 26> row{};
row[(t.a >> 24) - 1] = t.a & 0xFFFFFF;
row[(t.b >> 24) - 1] = t.b & 0xFFFFFF;
row[x] = v;
t.a = 0x80000000 | uint32_t(dense.size());
dense.push_back(row);
}
};
cp_algo::checkpoint("init");
int last = 0;
for(char c: s) {
int x = c - 'a', p = last;
++last;
for(; !get(p, x); p = link[p]) {
set(p, x, last);
}
int q = get(p, x);
if(q != last) {
if(length(q) == length(p) + 1) {
link[last] = q;
} else {
int clone = n + 1 + (int)clones.size();
clones.push_back({length(p) + 1, position(q)});
link[clone] = link[q];
auto row = to[q];
if(row.a & 0x80000000) {
auto copy = dense[row.a & 0x7FFFFFFF];
row.a = 0x80000000 | uint32_t(dense.size());
dense.push_back(copy);
}
to.push_back(row);
link[last] = link[q] = clone;
for(; get(p, x) == q; p = link[p]) {
set(p, x, clone);
}
}
}
}
int size = n + 1 + (int)clones.size();
to = decltype(to){};
dense = decltype(dense){};
cp_algo::checkpoint("build");
// Store the suffix-link tree compactly, with each node's children in letter order.
// Child IDs fit in 24 bits; the high byte stores the incoming letter.
cp_algo::big_vector<int> offset(size + 1);
cp_algo::big_vector<uint32_t> edges(size - 1);
for(int i = 1; i < size; i++) {
offset[link[i]]++;
}
std::partial_sum(offset.begin(), offset.end(), offset.begin());
for(int i = 1; i < size; i++) {
int p = link[i];
edges[--offset[p]] = (uint32_t(s[position(i) - length(p) - 1] - 'a') << 24) | i;
}
for(int i = 0; i < size; i++) {
std::sort(edges.begin() + offset[i], edges.begin() + offset[i + 1]);
}
cp_algo::checkpoint("tree");
std::vector<int> stack{0}, answer;
answer.reserve(n);
while(!stack.empty()) {
int u = stack.back();
stack.pop_back();
if(u && u <= n) {
answer.push_back(n - u);
}
for(int j = offset[u + 1]; j > offset[u];) {
stack.push_back(edges[--j] & 0xFFFFFF);
}
}
cp_algo::checkpoint("dfs");
for(int i: answer) {
std::cout << i << ' ';
}
std::cout << '\n';
cp_algo::checkpoint("write");
cp_algo::checkpoint<1>();
}
#line 1 "verify/string/suffix_array.test.cpp"
// @brief Suffix Array
#define PROBLEM "https://judge.yosupo.jp/problem/suffixarray"
#pragma GCC optimize("O3,unroll-loops")
#define CP_ALGO_CHECKPOINT
#line 1 "cp-algo/util/checkpoint.hpp"
#line 1 "cp-algo/util/big_alloc.hpp"
#include <set>
#include <map>
#include <deque>
#include <stack>
#include <queue>
#include <vector>
#include <string>
#include <cstddef>
#include <iostream>
#include <forward_list>
// Single macro to detect POSIX platforms (Linux, Unix, macOS)
#if defined(__linux__) || defined(__unix__) || (defined(__APPLE__) && defined(__MACH__))
# define CP_ALGO_USE_MMAP 1
# include <sys/mman.h>
#else
# define CP_ALGO_USE_MMAP 0
#endif
namespace cp_algo {
template <typename T, size_t Align = 32>
class big_alloc {
static_assert( Align >= alignof(void*), "Align must be at least pointer-size");
static_assert(std::popcount(Align) == 1, "Align must be a power of two");
public:
using value_type = T;
template <class U> struct rebind { using other = big_alloc<U, Align>; };
constexpr bool operator==(const big_alloc&) const = default;
constexpr bool operator!=(const big_alloc&) const = default;
big_alloc() noexcept = default;
template <typename U, std::size_t A>
big_alloc(const big_alloc<U, A>&) noexcept {}
[[nodiscard]] T* allocate(std::size_t n) {
std::size_t padded = round_up(n * sizeof(T));
std::size_t align = std::max<std::size_t>(alignof(T), Align);
#if CP_ALGO_USE_MMAP
if (padded >= MEGABYTE) {
void* raw = mmap(nullptr, padded,
PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
madvise(raw, padded, MADV_HUGEPAGE);
return static_cast<T*>(raw);
}
#endif
return static_cast<T*>(::operator new(padded, std::align_val_t(align)));
}
void deallocate(T* p, std::size_t n) noexcept {
if (!p) return;
std::size_t padded = round_up(n * sizeof(T));
std::size_t align = std::max<std::size_t>(alignof(T), Align);
#if CP_ALGO_USE_MMAP
if (padded >= MEGABYTE) { munmap(p, padded); return; }
#endif
::operator delete(p, padded, std::align_val_t(align));
}
private:
static constexpr std::size_t MEGABYTE = 1 << 20;
static constexpr std::size_t round_up(std::size_t x) noexcept {
return (x + Align - 1) / Align * Align;
}
};
template<typename T> using big_vector = std::vector<T, big_alloc<T>>;
template<typename T> using big_basic_string = std::basic_string<T, std::char_traits<T>, big_alloc<T>>;
template<typename T> using big_deque = std::deque<T, big_alloc<T>>;
template<typename T> using big_stack = std::stack<T, big_deque<T>>;
template<typename T> using big_queue = std::queue<T, big_deque<T>>;
template<typename T> using big_priority_queue = std::priority_queue<T, big_vector<T>>;
template<typename T> using big_forward_list = std::forward_list<T, big_alloc<T>>;
using big_string = big_basic_string<char>;
template<typename Key, typename Value, typename Compare = std::less<Key>>
using big_map = std::map<Key, Value, Compare, big_alloc<std::pair<const Key, Value>>>;
template<typename T, typename Compare = std::less<T>>
using big_multiset = std::multiset<T, Compare, big_alloc<T>>;
template<typename T, typename Compare = std::less<T>>
using big_set = std::set<T, Compare, big_alloc<T>>;
}
#line 5 "cp-algo/util/checkpoint.hpp"
#include <chrono>
#line 8 "cp-algo/util/checkpoint.hpp"
namespace cp_algo {
#ifdef CP_ALGO_CHECKPOINT
big_map<big_string, double> checkpoints;
double last;
#endif
template<bool final = false>
void checkpoint([[maybe_unused]] auto const& _msg) {
#ifdef CP_ALGO_CHECKPOINT
big_string msg = _msg;
double now = (double)clock() / CLOCKS_PER_SEC;
double delta = now - last;
last = now;
if(msg.size() && !final) {
checkpoints[msg] += delta;
}
if(final) {
for(auto const& [key, value] : checkpoints) {
std::cerr << key << ": " << value * 1000 << " ms\n";
}
std::cerr << "Total: " << now * 1000 << " ms\n";
}
#endif
}
template<bool final = false>
void checkpoint() {
checkpoint<final>("");
}
}
#line 1 "blazingio/blazingio.min.hpp"
// NOLINTBEGIN
// clang-format off
// DO NOT REMOVE THIS MESSAGE. The mess that follows is a minified build of
// https://github.com/purplesyringa/blazingio. Refer to the repository for
// a human-readable version and documentation.
// Options: cbfoiedrhWLMXaIaAn
#define M$(x,...)_mm256_##x##_epi8(__VA_ARGS__)
#define $u(...)__VA_ARGS__
#if __APPLE__
#define $m(A,B)A
#else
#define $m(A,B)B
#endif
#if _WIN32
#define $w(A,B)A
#else
#define $w(A,B)B
#endif
#if __i386__|_M_IX86
#define $H(A,B)A
#else
#define $H(A,B)B
#endif
#if __aarch64__
#define $a(A,B)A
#else
#define $a(A,B)B
#endif
#define $P(x)void F(x K){
#define $T template<$c T
#define $c class
#define $C constexpr
#define $R return
#define $O operator
#define u$ uint64_t
#define $r $R*this;
#include<array>
#include<bitset>
#include<complex>
#include<cstring>
#include $a(<arm_neon.h>,<immintrin.h>)
#include<stdint.h>
#include $w(<windows.h>,<sys/mman.h>)
#include<sys/stat.h>
#include $w(<io.h>,<unistd.h>)
#include $w(<ios>,<sys/resource.h>)
#if _MSC_VER
#define __builtin_add_overflow(a,b,c)_addcarry_u64(0,a,b,c)
#define $s
#else
$H(,u$ _umul128(u$ a,u$ b,u$*D){auto x=(__uint128_t)a*b;*D=u$(x>>64);$R(u$)x;})
#define $s $a(,__attribute__((target("avx2"))))
#endif
#define $z $a(16,32)
#define $t $a(uint8x16_t,__m256i)
#define $I $w(__forceinline,__attribute__((always_inline)))
#define $F M(),
#define E$(x)if(!(x))abort();
$w(LONG WINAPI $x(_EXCEPTION_POINTERS*);,)namespace $f{using namespace std;struct B{enum $c A:char{}c;B&$O=(char x){c=A{x};$r}$O char(){$R(char)c;}};$C u$ C=~0ULL/255;struct D{string&K;};static B E[65568];template<int F>struct G{B*H,*S;void K(off_t C){$w(char*D=(char*)VirtualAlloc(0,(C+8191)&-4096,8192,1);E$(D)E$(VirtualFree(D,0,32768))DWORD A=C&-65536;E$(!A||MapViewOfFileEx(CreateFileMapping(GetStdHandle(-10),0,2,0,A,0),4,0,0,0,D)==D)E$(VirtualAlloc(D+A,65536,12288,4)==D+A)E$(~_lseek(0,A,0))DWORD E=0;ReadFile(GetStdHandle(-10),D+A,65536,&E,0);,int A=getpagesize();char*D=(char*)mmap(0,C+A,3,2,0,0);E$(D!=(void*)-1)E$(mmap(D+((C+A-1)&-A),A,3,$m(4114,50),-1,0)!=(void*)-1))H=(B*)D+C;*H=10;H[1]=48;H[2]=0;S=(B*)D;}void L(){H=S=E;}$I void M(){if(F&&S==H){$w(DWORD A=0;ReadFile(GetStdHandle(-10),S=E,65536,&A,0);,$a($u(register long A asm("x0")=0,D asm("x1")=(long)E,G asm("x2")=65536,C asm($m("x16","x8"))=$m(3,63);asm volatile("svc 0" $m("x80",):"+r"(A),"+r"(D):"r"(C),"r"(G));S=launder(E);),off_t A=$H(3,$m(33554435,0));B*D=E;asm volatile($H("int $128","syscall"):"+a"(A),$H("+c"(D):"b","+S"(D):"D")(0),"d"(65536)$H(,$u(:"rcx","r11")));S=D;))H=S+A;*H=10;if(!A)E[1]=48,E[2]=0;}}$T>$I void N(T&x){while($F(*S&240)==48)x=T(x*10+(*S++-48));}$T>$I decltype((void)~T{1})O(T&x){M();int A=is_signed_v<T>&&*S==45;S+=A;N(x=0);x=A?1+~x:x;}$T>$I decltype((void)T{1.})O(T&x){M();int A=*S==45;S+=A;$F S+=*S==43;u$ n=0;int i=0;for(;i<18&&($F*S&240)==48;i++)n=n*10+*S++-48;int B=20;int C=*S==46;S+=C;for(;i<18&&($F*S&240)==48;i++)n=n*10+*S++-48,B-=C;x=(T)n;while(($F*S&240)==48)x=x*10+*S++-48,B-=C;if(*S==46)S++,C=1;while(($F*S&240)==48)x=x*10+*S++-48,B-=C;int D;if((*S|32)==101)S++,$F S+=*S==43,O(D),B+=D;static $C auto E=[](){array<T,41>E{};T x=1;for(int i=21;i--;)E[40-i]=x,E[i]=1/x,x*=10;$R E;}();while(B>40)x*=(T)1e10,B-=10;while(B<0)x*=(T)1e-10,B+=10;x*=E[B];x=A?-x:x;}$I void O(bool&x){$F x=*S++==49;}$I void O(char&x){$F x=*S++;}$I void O(uint8_t&x){$F x=*S++;}$I void O(int8_t&x){$F x=*S++;}$T>$s void P(string&K,T C){M();B*G=S;C();K.assign((char*)G,S-G);while(F&&S==H&&($F H!=E)){C();K.append(E,S);}}$s void O(string&K){P(K,[&]()$s{B*p=S;$w(ULONG R;,)$t x;$a(uint64x2_t A;while(memcpy(&x,p,16),A=uint64x2_t(x<33),!(A[0]|A[1]))p+=16;S=p+(A[0]?0:8)+$w((_BitScanForward64(&R,A[0]?A[0]:A[1]),R),__builtin_ctzll(A[0]?A[0]:A[1]))/8;,int J;$t C=M$(set1,32);while(memcpy(&x,p,32),!(J=M$(movemask,M$(cmpeq,C,_mm256_max_epu8(C,x)))))p+=32;S=p+$w((_BitScanForward(&R,J),R),__builtin_ctz(J));)});}$s void O(D&A){P(A.K,[&](){S=(B*)memchr(S,10,H-S+1);});if(A.K.size()&&A.K.back()==13)A.K.pop_back();if(A.K.empty()||S<H)S+=*S==10;}$T>$I void O(complex<T>&K){T A,B{};if($F*S==40){S++;O(A);if($F*S++==44)Q(B),S++;}else O(A);K={A,B};}template<size_t N>$s void O(bitset<N>&K){if(N>4095&&!*this)$R;ptrdiff_t i=N;while(i)if($F i%$z||H-S<$z)K[--i]=*S++==49;else{B*p=S;for(int64_t j=0;j<min(i,H-S)/$z;j++){i-=$z;$t x;memcpy(&x,p,$z);$a(auto B=(uint8x16_t)vdupq_n_u64(~2ULL/254)&(48-x);auto C=vzip_u8(vget_high_u8(B),vget_low_u8(B));auto y=vaddvq_u16((uint16x8_t)vcombine_u8(C.val[0],C.val[1]));,u$ a=~0ULL/65025;auto y=$w(_byteswap_ulong,__builtin_bswap32)(M$(movemask,M$(shuffle,_mm256_slli_epi32(x,7),_mm256_set_epi64x(a+C*24,a+C*16,a+C*8,a))));)p+=$z;memcpy((char*)&K+i/8,&y,$z/8);}S=p;}}$T>$I void Q(T&K){if(!is_same_v<T,D>)while($F(uint8_t)*S<33)S++;O(K);}$O bool(){$R!!*this;}bool $O!(){$R S>H;}};struct U{G<0>A;G<1>B;U(){struct stat D;E$(~fstat(0,&D))(D.st_mode>>12)==8?A.K(D.st_size):B.L();}U*tie(nullptr_t){$R this;}void sync_with_stdio(bool){}$T>$I U&$O>>(T&K){A.S?A.Q(K):B.Q(K);$r}$O bool(){$R!!*this;}bool $O!(){$R A.S?!A:!B;}};short A[100];char L[64]{1};struct
V{char*D;B*S;int J;V(){$w(E$(D=(char*)VirtualAlloc(0,536870912,8192,4))E$(VirtualAlloc(D,4096,4096,260))AddVectoredExceptionHandler(1,$x);,size_t C=536870912;$m(,rlimit E;getrlimit(RLIMIT_AS,&E);if(~E.rlim_cur)C=25165824;)D=(char*)mmap(0,C,3,$m(4162,16418),-1,0);E$(D!=(void*)-1))S=(B*)D;for(int i=0;i<100;i++)A[i]=short((48+i/10)|((48+i%10)<<8));for(int i=1;i<64;i++)L[i]=L[i-1]+(0x8922489224892249>>i&1);}~V(){flush($w(!J,));}void flush($w(int F=0,)){$w(J=1;auto E=GetStdHandle(-11);auto C=F?ReOpenFile(E,1073741824,7,2684354560):(void*)-1;DWORD A;E$(C==(void*)-1?WriteFile(E,D,DWORD((char*)S-D),&A,0):(WriteFile(C,D,DWORD(((char*)S-D+4095)&-4096),&A,0)&&~_chsize(1,int((char*)S-D)))),auto G=D;ssize_t A;while((A=write(1,G,(char*)S-G))>0)G+=A;E$(~A))S=(B*)D;}$P(char)*S++=K;}$P(uint8_t)*S++=K;}$P(int8_t)*S++=K;}$P(bool)*S++=48+K;}$T>decltype((void)~T{1})F(T K){using D=make_unsigned_t<T>;D C=K;if(K<0)F('-'),C=1+~C;static $C auto N=[](){array<D,5*sizeof(T)/2>N{};D n=1;for(size_t i=1;i<N.size();i++)n*=10,N[i]=n;$R N;}();$w(ULONG M;,)int G=L[$w(($H(_BitScanReverse(&M,ULONG((int64_t)C>>32))?M+=32:_BitScanReverse(&M,(ULONG)C|1),_BitScanReverse64(&M,C|1)),M),63^__builtin_clzll(C|1))];G-=C<N[G-1];short H[20];if $C(sizeof(T)==2){auto n=33555U*C-C/2;u$ H=A[n>>25];n=(n&33554431)*25;H|=A[n>>23]<<16;H|=u$(48+((n&8388607)*5>>22))<<32;H>>=40-G*8;memcpy(S,&H,8);}else if $C(sizeof(T)==4){auto n=1441151881ULL*C;$H(n>>=25;n++;for(int i=0;i<5;i++){H[i]=A[n>>32];n=(n&~0U)*100;},int K=57;auto J=~0ULL>>7;for(int i=0;i<5;i++){H[i]=A[n>>K];n=(n&J)*25;K-=2;J/=4;})memcpy(S,(B*)H+10-G,16);}else{$H($u(if(C<(1ULL<<32)){$R F((uint32_t)C);}auto J=(u$)1e10;auto x=C/J,y=C%J;int K=100000,b[]{int(x/K),int(x%K),int(y/K),int(y%K)};B H[40];for(int i=0;i<4;i++){int n=int((429497ULL*b[i]>>7)+1);B*p=H+i*5;*p=48+char(n>>25);n=(n&~0U>>7)*25;memcpy(p+1,A+(n>>23),2);memcpy(p+3,A+((n&~0U>>9)*25>>21),2);}),$u(u$ D,E=_umul128(18,C,&D),F;_umul128(0x725dd1d243aba0e8,C,&F);D+=__builtin_add_overflow(E,F+1,&E);for(int i=0;i<10;i++)H[i]=A[D],E=_umul128(100,E,&D);))memcpy(S,(B*)H+20-G,20);}S+=G;}$T>decltype((void)T{1.})F(T K){if(K<0)F('-'),K=-K;auto G=[&](){auto x=u$(K*1e12);$H($u(x-=x>999999999999;uint32_t n[]{uint32_t(x/1000000*429497>>7)+1,uint32_t(x%1000000*429497>>7)+1};int K=25,J=~0U>>7;for(int i=0;i<3;i++){for(int j=0;j<2;j++)memcpy(S+i*2+j*6,A+(n[j]>>K),2),n[j]=(n[j]&J)*25;K-=2;J/=4;}S+=12;),$u(u$ D,E=_umul128(472236648287,x,&D)>>8;E|=D<<56;D>>=8;E++;for(int i=0;i<6;i++)memcpy(S,A+D,2),S+=2,E=_umul128(100,E,&D);))};if(K==0)$R F('0');if(K>=1e16){K*=(T)1e-16;int B=16;while(K>=1)K*=(T).1,B++;F("0.");G();F('e');F(B);}else if(K>=1){auto B=(u$)K;F(B);if((K-=(T)B)>0)F('.'),G();}else F("0."),G();}$P(const char*)$w(size_t A=strlen(K);memcpy((char*)S,K,A);S+=A;,S=(B*)stpcpy((char*)S,K);)}$P(const uint8_t*)F((char*)K);}$P(const int8_t*)F((char*)K);}$P(string_view)memcpy(S,K.data(),K.size());S+=K.size();}$T>$P(complex<T>)*this<<'('<<K.real()<<','<<K.imag()<<')';}template<size_t N>$s $P(const bitset<N>&)auto i=N;while(i%$z)*S++=48+K[--i];B*p=S;while(i){i-=$z;$a(short,int)x;memcpy(&x,(char*)&K+i/8,$z/8);$a(auto A=(uint8x8_t)vdup_n_u16(x);vst1q_u8((uint8_t*)p,48-vtstq_u8(vcombine_u8(vuzp2_u8(A,A),vuzp1_u8(A,A)),(uint8x16_t)vdupq_n_u64(~2ULL/254)));,auto b=_mm256_set1_epi64x(~2ULL/254);_mm256_storeu_si256(($t*)p,M$(sub,M$(set1,48),M$(cmpeq,_mm256_and_si256(M$(shuffle,_mm256_set1_epi32(x),_mm256_set_epi64x(0,C,C*2,C*3)),b),b)));)p+=$z;}S=p;}$T>V&$O<<(const T&K){F(K);$r}V&$O<<(V&(*A)(V&)){$R A(*this);}};struct W{$T>W&$O<<(const T&K){$r}W&$O<<(W&(*A)(W&)){$R A(*this);}};}namespace std{$f::U i$;$f::V o$;$f::W e$;$f::U&getline($f::U&B,string&K){$f::D A{K};$R B>>A;}$f::V&flush($f::V&B){if(!i$.A.S)B.flush();$R B;}$f::V&endl($f::V&B){$R B<<'\n'<<flush;}$f::W&endl($f::W&B){$R B;}$f::W&flush($f::W&B){$R B;}}$w(LONG WINAPI $x(_EXCEPTION_POINTERS*A){auto C=A->ExceptionRecord;auto B=C->ExceptionInformation[1];if(C->ExceptionCode==2147483649&&B-(ULONG_PTR)std::o$.D<0x40000000){E$(VirtualAlloc((char*)B,16777216,4096,4)&&VirtualAlloc((char*)(B+16777216),4096,4096,260))$R-1;}$R 0;},)
#define freopen(...)if(freopen(__VA_ARGS__)==stdin)std::i$=$f::U{}
#define cin i$
#define cout o$
#ifdef ONLINE_JUDGE
#define cerr e$
#define clog e$
#endif
// End of blazingio
// NOLINTEND
// clang-format on
#line 8 "verify/string/suffix_array.test.cpp"
#include <algorithm>
#line 10 "verify/string/suffix_array.test.cpp"
#include <cassert>
#include <cstdint>
#include <numeric>
#line 14 "verify/string/suffix_array.test.cpp"
int main() {
std::ios::sync_with_stdio(false);
std::cin.tie(nullptr);
std::string s;
std::cin >> s;
std::ranges::reverse(s);
int n = (int)s.size();
// Ordinary state i represents the prefix of length i; clones follow state n.
struct clone_data { int len, pos; };
std::vector<clone_data> clones;
clones.reserve(n);
cp_algo::big_vector<int> link(2 * n + 1);
auto length = [&](int v) { return v <= n ? v : clones[v - n - 1].len; };
auto position = [&](int v) { return v <= n ? v : clones[v - n - 1].pos; };
// Pack two (letter + 1, state) pairs inline; bit 31 marks a dense row.
// State IDs fit in 24 bits under the problem's n <= 500000 bound.
struct transitions { uint32_t a = 0, b = 0; };
cp_algo::big_vector<std::array<int, 26>> dense;
// At most one dense row per state; avoid growth on branching structured strings.
dense.reserve(link.size());
cp_algo::big_vector<transitions> to;
to.reserve(link.size());
to.resize(n + 1);
auto get = [&](int p, int x) {
auto t = to[p];
if(t.a & 0x80000000) {return dense[t.a & 0x7FFFFFFF][x];}
if((t.a >> 24) == unsigned(x + 1)) {return int(t.a & 0xFFFFFF);}
if((t.b >> 24) == unsigned(x + 1)) {return int(t.b & 0xFFFFFF);}
return 0;
};
auto set = [&](int p, int x, int v) {
auto &t = to[p];
auto encoded = uint32_t((x + 1) << 24) | v;
if(t.a & 0x80000000) {dense[t.a & 0x7FFFFFFF][x] = v;}
else if(!t.a || (t.a >> 24) == unsigned(x + 1)) {t.a = encoded;}
else if(!t.b || (t.b >> 24) == unsigned(x + 1)) {t.b = encoded;}
else {
std::array<int, 26> row{};
row[(t.a >> 24) - 1] = t.a & 0xFFFFFF;
row[(t.b >> 24) - 1] = t.b & 0xFFFFFF;
row[x] = v;
t.a = 0x80000000 | uint32_t(dense.size());
dense.push_back(row);
}
};
cp_algo::checkpoint("init");
int last = 0;
for(char c: s) {
int x = c - 'a', p = last;
++last;
for(; !get(p, x); p = link[p]) {
set(p, x, last);
}
int q = get(p, x);
if(q != last) {
if(length(q) == length(p) + 1) {
link[last] = q;
} else {
int clone = n + 1 + (int)clones.size();
clones.push_back({length(p) + 1, position(q)});
link[clone] = link[q];
auto row = to[q];
if(row.a & 0x80000000) {
auto copy = dense[row.a & 0x7FFFFFFF];
row.a = 0x80000000 | uint32_t(dense.size());
dense.push_back(copy);
}
to.push_back(row);
link[last] = link[q] = clone;
for(; get(p, x) == q; p = link[p]) {
set(p, x, clone);
}
}
}
}
int size = n + 1 + (int)clones.size();
to = decltype(to){};
dense = decltype(dense){};
cp_algo::checkpoint("build");
// Store the suffix-link tree compactly, with each node's children in letter order.
// Child IDs fit in 24 bits; the high byte stores the incoming letter.
cp_algo::big_vector<int> offset(size + 1);
cp_algo::big_vector<uint32_t> edges(size - 1);
for(int i = 1; i < size; i++) {
offset[link[i]]++;
}
std::partial_sum(offset.begin(), offset.end(), offset.begin());
for(int i = 1; i < size; i++) {
int p = link[i];
edges[--offset[p]] = (uint32_t(s[position(i) - length(p) - 1] - 'a') << 24) | i;
}
for(int i = 0; i < size; i++) {
std::sort(edges.begin() + offset[i], edges.begin() + offset[i + 1]);
}
cp_algo::checkpoint("tree");
std::vector<int> stack{0}, answer;
answer.reserve(n);
while(!stack.empty()) {
int u = stack.back();
stack.pop_back();
if(u && u <= n) {
answer.push_back(n - u);
}
for(int j = offset[u + 1]; j > offset[u];) {
stack.push_back(edges[--j] & 0xFFFFFF);
}
}
cp_algo::checkpoint("dfs");
for(int i: answer) {
std::cout << i << ' ';
}
std::cout << '\n';
cp_algo::checkpoint("write");
cp_algo::checkpoint<1>();
}
Test cases
| Env | Name | Status | Elapsed | Memory |
|---|---|---|---|---|
| g++ | all_same_00 |
AC |
26 ms | 18 MB |
| g++ | all_same_01 |
AC |
25 ms | 19 MB |
| g++ | all_same_02 |
AC |
26 ms | 19 MB |
| g++ | all_same_03 |
AC |
24 ms | 18 MB |
| g++ | all_same_04 |
AC |
24 ms | 18 MB |
| g++ | almost_single_00 |
AC |
32 ms | 33 MB |
| g++ | almost_single_01 |
AC |
32 ms | 25 MB |
| g++ | almost_single_02 |
AC |
32 ms | 35 MB |
| g++ | almost_single_03 |
AC |
33 ms | 27 MB |
| g++ | almost_single_04 |
AC |
28 ms | 27 MB |
| g++ | almost_single_05 |
AC |
29 ms | 28 MB |
| g++ | binary_carry_00 |
AC |
32 ms | 21 MB |
| g++ | binary_carry_01 |
AC |
33 ms | 21 MB |
| g++ | example_00 |
AC |
3 ms | 6 MB |
| g++ | example_01 |
AC |
2 ms | 6 MB |
| g++ | example_02 |
AC |
2 ms | 6 MB |
| g++ | example_03 |
AC |
2 ms | 6 MB |
| g++ | fib_str_00 |
AC |
30 ms | 19 MB |
| g++ | fib_str_01 |
AC |
23 ms | 16 MB |
| g++ | fib_str_02 |
AC |
22 ms | 15 MB |
| g++ | fib_str_03 |
AC |
20 ms | 15 MB |
| g++ | fib_str_04 |
AC |
39 ms | 22 MB |
| g++ | hack_00 |
AC |
3 ms | 6 MB |
| g++ | hack_01 |
AC |
2 ms | 6 MB |
| g++ | hack_02 |
AC |
2 ms | 6 MB |
| g++ | max_random_00 |
AC |
55 ms | 24 MB |
| g++ | max_random_01 |
AC |
56 ms | 23 MB |
| g++ | max_random_02 |
AC |
55 ms | 23 MB |
| g++ | max_random_03 |
AC |
54 ms | 25 MB |
| g++ | max_random_04 |
AC |
54 ms | 24 MB |
| g++ | near_power_of_2_max_random_00 |
AC |
27 ms | 16 MB |
| g++ | near_power_of_2_max_random_01 |
AC |
25 ms | 16 MB |
| g++ | near_power_of_2_max_same_00 |
AC |
14 ms | 12 MB |
| g++ | near_power_of_2_max_same_01 |
AC |
13 ms | 12 MB |
| g++ | one_00 |
AC |
3 ms | 6 MB |
| g++ | random_00 |
AC |
44 ms | 20 MB |
| g++ | random_01 |
AC |
52 ms | 23 MB |
| g++ | random_02 |
AC |
8 ms | 7 MB |
| g++ | random_03 |
AC |
48 ms | 21 MB |
| g++ | random_04 |
AC |
29 ms | 16 MB |
| g++ | small_random_00 |
AC |
3 ms | 6 MB |
| g++ | small_random_01 |
AC |
2 ms | 6 MB |
| g++ | small_random_02 |
AC |
2 ms | 6 MB |
| g++ | small_random_03 |
AC |
2 ms | 6 MB |
| g++ | small_random_04 |
AC |
2 ms | 6 MB |
| g++ | small_random_05 |
AC |
2 ms | 6 MB |
| g++ | small_random_06 |
AC |
2 ms | 6 MB |
| g++ | small_random_07 |
AC |
2 ms | 6 MB |
| g++ | small_random_08 |
AC |
2 ms | 6 MB |
| g++ | small_random_09 |
AC |
2 ms | 6 MB |