cp-library
競技プログラミング用 C++ ライブラリ
test/suffix_array_public.test.cpp
test/suffix_array_public.test.cpp · GitHub で見る
依存ライブラリ
コード
#define PROBLEM "https://judge.yosupo.jp/problem/suffixarray"
#include "../string/suffix_array.hpp"
#include "../components/fast_io.hpp"
int main() {
FastInput input;
FastOutput output;
std::string s;
input.read(s);
for (int x : suffix_array(s)) { output.write(x); output.write(' '); }
output.write('\n');
}
#line 1 "test/suffix_array_public.test.cpp"
#define PROBLEM "https://judge.yosupo.jp/problem/suffixarray"
#line 1 "string/suffix_array.hpp"
#line 1 "components/suffix_array.hpp"
#include <algorithm>
#include <array>
#include <bit>
#include <cstdint>
#include <cassert>
#include <limits>
#include <numeric>
#include <string_view>
#include <vector>
#include <type_traits>
#if defined(__x86_64__) && (defined(__GNUC__) || defined(__clang__))
#include <immintrin.h>
#endif
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC push_options
#pragma GCC optimize("O3")
#endif
namespace suffix_array_detail {
#if defined(__x86_64__) && (defined(__GNUC__) || defined(__clang__))
__attribute__((target("avx2"))) inline void byte_seeds(const unsigned char *s, int n, std::vector<int> &seeds, std::vector<unsigned char> &marked) {
seeds.push_back(n), marked[n] = 1;
bool right = true;
int end = n;
const auto sign = _mm256_set1_epi8(static_cast<char>(0x80));
while (end >= 33) {
int begin = end - 32;
auto a = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(s + begin));
auto next = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(s + begin + 1));
auto previous = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(s + begin - 1));
auto biased = _mm256_xor_si256(a, sign);
std::uint32_t equal = _mm256_movemask_epi8(_mm256_cmpeq_epi8(a, next));
std::uint32_t small = _mm256_movemask_epi8(_mm256_cmpgt_epi8(_mm256_xor_si256(next, sign), biased));
if (right) small |= equal & 0x80000000U;
for (int shift : {1, 2, 4, 8, 16}) {
small |= equal & (small >> shift);
equal &= equal >> shift;
}
std::uint32_t selected = small & std::uint32_t(_mm256_movemask_epi8(_mm256_cmpgt_epi8(_mm256_xor_si256(previous, sign), biased)));
while (selected) {
int bit = 31 - std::countl_zero(selected), p = begin + bit;
seeds.push_back(p), marked[p] = 1;
selected ^= std::uint32_t(1) << bit;
}
right = small & 1;
end = begin;
}
for (int i = end - 1; i > 0; --i) {
bool current = s[i] < s[i + 1] || (s[i] == s[i + 1] && right);
if (current && s[i - 1] > s[i]) seeds.push_back(i), marked[i] = 1;
right = current;
}
}
#endif
template<class Sequence>
bool prefix_order(const Sequence &s, int n, int alphabet, const std::vector<int> &seeds, std::vector<int> &ordered) {
if (seeds.size() <= 64 || alphabet > 257) return false;
int bits = std::bit_width(unsigned(alphabet - 1)), position_bits = std::bit_width(unsigned(n));
int length = (64 - position_bits) / bits;
std::uint64_t position_mask = (std::uint64_t(1) << position_bits) - 1;
auto key_at = [&](int p) {
std::uint64_t key = 0;
for (int k = 0; k < length; ++k) key = (key << bits) | (k < n - p ? unsigned(s[p + k]) : 0);
return key;
};
std::array<std::uint64_t, 128> sample;
int sample_size = std::min(128, static_cast<int>(seeds.size()));
for (int i = 0; i < sample_size; ++i) sample[i] = key_at(seeds[i]);
std::sort(sample.begin(), sample.begin() + sample_size);
for (int i = 1; i < sample_size; ++i) if (sample[i] == sample[i - 1]) return false;
std::vector<std::uint64_t> entries(seeds.size()), scratch(seeds.size());
for (std::size_t i = 0; i < seeds.size(); ++i) entries[i] = (key_at(seeds[i]) << position_bits) | unsigned(n - seeds[i]);
int first_shift = std::max(position_bits, bits * length + position_bits - 33);
for (int shift = first_shift; shift < bits * length + position_bits; shift += 11) {
std::array<int, 2048> count{};
for (auto entry : entries) ++count[(entry >> shift) & 2047];
int total = 0;
for (int &value : count) { int next = total + value; value = total; total = next; }
for (auto entry : entries) scratch[count[(entry >> shift) & 2047]++] = entry;
entries.swap(scratch);
}
for (std::size_t first = 0; first < entries.size();) {
std::size_t last = first + 1;
while (last < entries.size() && (entries[last] >> first_shift) == (entries[first] >> first_shift)) ++last;
if (last - first > 64) return false;
std::sort(entries.begin() + first, entries.begin() + last);
for (std::size_t i = first + 1; i < last; ++i) if ((entries[i] >> position_bits) == (entries[i - 1] >> position_bits)) return false;
first = last;
}
ordered.resize(seeds.size());
for (std::size_t i = 0; i < entries.size(); ++i) ordered[i] = n - static_cast<int>(entries[i] & position_mask);
return true;
}
template<class Sequence>
std::vector<int> encoded_sort(const Sequence &s, int n, int alphabet) {
std::vector<int> sa(std::size_t(n) + 1);
if (n < 16) {
std::iota(sa.begin(), sa.end(), 0);
std::sort(sa.begin(), sa.end(), [&](int a, int b) {
while (a < n && b < n && s[a] == s[b]) ++a, ++b;
return a != n ? b != n && s[a] < s[b] : b != n;
});
return sa;
}
std::vector<unsigned char> is_seed(std::size_t(n) + 1);
std::vector<int> seeds;
seeds.reserve(n / 2 + 1);
bool gathered = false;
#if defined(__x86_64__) && (defined(__GNUC__) || defined(__clang__))
if constexpr (std::is_same_v<Sequence, std::vector<unsigned char>>) {
if (n >= 64 && __builtin_cpu_supports("avx2")) {
byte_seeds(s.data(), n, seeds, is_seed);
gathered = true;
}
}
#endif
if (!gathered) {
bool next = true;
for (int i = n - 1; i >= 0; --i) {
bool current = s[i] == s[i + 1] ? next : s[i] < s[i + 1];
if (!current && next) seeds.push_back(i + 1), is_seed[i + 1] = 1;
next = current;
}
}
std::reverse(seeds.begin(), seeds.end());
std::vector<std::size_t> ends(alphabet), work(alphabet);
for (std::size_t i = 0; i < sa.size(); ++i) ++ends[s[i]];
std::partial_sum(ends.begin(), ends.end(), ends.begin());
auto induce = [&](const std::vector<int> &ordered) {
std::fill(sa.begin(), sa.end(), 0);
work = ends;
for (auto i = ordered.rbegin(); i != ordered.rend(); ++i) if (*i != n) sa[--work[s[*i]]] = *i + 1;
work[0] = 0;
std::copy(ends.begin(), ends.end() - 1, work.begin() + 1);
int last = n;
if (s[n - 2] < s[n - 1]) last = -last;
sa[work[s[n - 1]]++] = last;
for (std::size_t i = 1; i < sa.size(); ++i) {
int p = sa[i];
if (p > 1) {
int j = p - 2, c = s[j], v = j + 1;
if (!j || s[j - 1] < c) v = -v;
sa[work[c]++] = v;
}
}
work = ends;
for (int i = n; i > 0; --i) {
int p = sa[i];
sa[i] = (p < 0 ? -p : p) - 1;
if (p < -1) {
int j = -p - 2, c = s[j], v = j + 1;
if (j && s[j - 1] <= c) v = -v;
sa[--work[c]] = v;
}
}
sa[0] = n;
};
std::vector<int> ordered;
int used = alphabet;
while (used > 1 && ends[used - 1] == ends[used - 2]) --used;
if (prefix_order(s, n, used, seeds, ordered)) { induce(ordered); return sa; }
induce(seeds);
int m = static_cast<int>(seeds.size());
std::vector<int> sorted;
sorted.reserve(m);
for (int p : sa) if (is_seed[p]) sorted.push_back(p);
std::vector<int> names(std::size_t(n) / 2 + 1);
for (int i = 0; i + 1 < m; ++i) names[seeds[i] >> 1] = seeds[i + 1] - seeds[i] + 1;
names[n >> 1] = 1;
int rank = -1, previous = -1, previous_length = 0;
for (int p : sorted) {
int length = names[p >> 1];
bool different = previous < 0 || length != previous_length;
if (!different) for (int k = 0; k < length; ++k) if (s[p + k] != s[previous + k]) { different = true; break; }
rank += different;
names[p >> 1] = rank;
previous = p, previous_length = length;
}
std::vector<int> reduced(m), order(m);
for (int i = 0; i < m; ++i) reduced[i] = names[seeds[i] >> 1];
if (rank + 1 == m) for (int i = 0; i < m; ++i) order[reduced[i]] = i;
else order = encoded_sort(reduced, m - 1, rank + 1);
for (int i = 0; i < m; ++i) sorted[i] = seeds[order[i]];
induce(sorted);
return sa;
}
template<class Sequence>
std::vector<int> induced_sort(const Sequence &s, int upper) {
struct Encoded {
const Sequence &value;
int n;
int operator[](int i) const { return i == n ? 0 : value[i] + 1; }
};
int n = static_cast<int>(s.size());
if (n < 2) return n ? std::vector<int>{0} : std::vector<int>{};
int common = 1;
while (common < n && s[common] == s[0]) ++common;
if (common == n) {
std::vector<int> sa(n);
std::iota(sa.rbegin(), sa.rend(), 0);
return sa;
}
auto sa = encoded_sort(Encoded{s, n}, n, upper + 2);
sa.erase(sa.begin());
return sa;
}
inline std::vector<int> bytes(std::string_view s) {
assert(s.size() <= std::size_t(std::numeric_limits<int>::max()));
if (s.empty()) return {};
std::array<int, 256> codes{};
for (unsigned char c : s) codes[c] = 1;
int alphabet = 0;
for (int &code : codes) if (code) code = ++alphabet;
int n = static_cast<int>(s.size());
if (alphabet == 1) {
std::vector<int> sa(n);
std::iota(sa.rbegin(), sa.rend(), 0);
return sa;
}
auto build = [&]<class Symbol>() {
std::vector<Symbol> text(s.size() + 1);
for (int i = 0; i < n; ++i) text[i] = static_cast<Symbol>(codes[static_cast<unsigned char>(s[i])]);
auto sa = encoded_sort(text, n, alphabet + 1);
sa.erase(sa.begin());
return sa;
};
return alphabet < 256 ? build.template operator()<unsigned char>() : build.template operator()<std::uint16_t>();
}
}
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC pop_options
#endif
#line 5 "string/suffix_array.hpp"
inline std::vector<int> suffix_array(std::string_view s) { return suffix_array_detail::bytes(s); }
template <class T>
std::vector<int> suffix_array(const std::vector<T> &s) {
assert(s.size() <= std::size_t(std::numeric_limits<int>::max()));
std::vector<int> order(s.size()), ranks(s.size());
std::iota(order.begin(), order.end(), 0);
std::sort(order.begin(), order.end(), [&](int a, int b) { return s[a] < s[b]; });
int rank = 0;
for (std::size_t i = 0; i < order.size(); ++i) {
if (i && s[order[i - 1]] < s[order[i]]) ++rank;
ranks[order[i]] = rank;
}
return suffix_array_detail::induced_sort(ranks, rank);
}
inline std::vector<int> suffix_array(const std::vector<int> &s, int upper) {
assert(upper >= 0 && s.size() <= std::size_t(std::numeric_limits<int>::max()));
for ([[maybe_unused]] int x : s) assert(0 <= x && x <= upper);
if (upper > std::numeric_limits<int>::max() - 2) return suffix_array(s);
return suffix_array_detail::induced_sort(s, upper);
}
#line 1 "components/fast_io.hpp"
#line 8 "components/fast_io.hpp"
#include <charconv>
#include <cstddef>
#line 11 "components/fast_io.hpp"
#include <cstdio>
#include <cstring>
#include <memory>
#include <stdexcept>
#include <string>
#line 18 "components/fast_io.hpp"
#if defined(__unix__) || defined(__APPLE__)
#include <sys/mman.h>
#include <sys/stat.h>
#include <unistd.h>
#endif
namespace fast_io_detail { struct BulkInput; struct BulkOutput; }
class FastInput {
friend struct fast_io_detail::BulkInput;
static constexpr std::size_t capacity = 1 << 16;
FILE *file = nullptr;
std::unique_ptr<char[]> buffer;
const char *current = "", *end = current;
std::string scratch;
void *mapping = nullptr;
std::size_t mapped_size = 0;
bool refill() {
if (!file) return false;
std::size_t n = std::fread(buffer.get(), 1, capacity, file);
current = buffer.get();
end = current + n;
return n;
}
static const char *scan(const char *p, const char *end) {
while (end - p >= 8) {
std::uint64_t x;
std::memcpy(&x, p, 8);
if ((x - 0x2121212121212121ULL) & ~x & 0x8080808080808080ULL) break;
p += 8;
}
while (p != end && static_cast<unsigned char>(*p) > ' ') ++p;
return p;
}
public:
explicit FastInput(FILE *input = stdin) {
#if defined(__unix__) || defined(__APPLE__)
struct stat status;
int fd = fileno(input);
auto offset = ftello(input);
if (fd >= 0 && offset >= 0 && !fstat(fd, &status) && S_ISREG(status.st_mode) && status.st_size > offset) {
mapped_size = status.st_size;
mapping = mmap(nullptr, mapped_size, PROT_READ, MAP_PRIVATE, fd, 0);
if (mapping != MAP_FAILED) {
current = static_cast<const char *>(mapping) + offset;
end = static_cast<const char *>(mapping) + mapped_size;
return;
}
mapping = nullptr;
}
#endif
file = input;
buffer = std::make_unique_for_overwrite<char[]>(capacity);
}
explicit FastInput(std::string_view input) : current(input.empty() ? "" : input.data()), end(current + input.size()) {}
FastInput(const FastInput &) = delete;
FastInput &operator=(const FastInput &) = delete;
~FastInput() {
#if defined(__unix__) || defined(__APPLE__)
if (mapping) munmap(mapping, mapped_size);
#endif
}
std::string_view token() {
while (true) {
while (current != end && static_cast<unsigned char>(*current) <= ' ') ++current;
if (current != end) break;
if (!refill()) return {};
}
const char *begin = current;
current = scan(current, end);
if (current != end || !file) return {begin, std::size_t(current - begin)};
scratch.assign(begin, end);
while (refill()) {
current = scan(current, end);
scratch.append(buffer.get(), std::size_t(current - buffer.get()));
if (current != end) break;
}
return scratch;
}
template <class T>
bool read(T &value) {
if constexpr (std::is_integral_v<T> && !std::is_same_v<T, bool>) {
while (true) {
while (current != end && static_cast<unsigned char>(*current) <= ' ') ++current;
if (current != end) break;
if (!refill()) return false;
}
if constexpr (std::is_unsigned_v<T> && sizeof(T) == 4 && std::endian::native == std::endian::little) {
if (end - current >= 11) {
std::uint64_t x;
std::memcpy(&x, current, 8);
if (!(((x + 0x4646464646464646ULL) | (x - 0x3030303030303030ULL)) & 0x8080808080808080ULL)) {
x &= 0x0f0f0f0f0f0f0f0fULL;
x = (x * 10 + (x >> 8)) & 0x00ff00ff00ff00ffULL;
x = (x * 100 + (x >> 16)) & 0x0000ffff0000ffffULL;
std::uint64_t v = std::uint32_t(x * 10000 + (x >> 32));
int n = 8;
for (; n < 10 && static_cast<unsigned char>(current[n]) - unsigned('0') < 10; ++n)
v = v * 10 + (current[n] - '0');
if (static_cast<unsigned char>(current[n]) <= ' ' && v <= 0xffffffffULL) {
current += n;
value = T(v);
return true;
}
}
}
}
T parsed{};
auto result = std::from_chars(current, end, parsed);
if (result.ec == std::errc{} && (result.ptr != end ? static_cast<unsigned char>(*result.ptr) <= ' ' : !file)) {
current = result.ptr;
value = parsed;
return true;
}
}
auto s = token();
if (s.empty()) return false;
if constexpr (std::is_same_v<T, bool>) {
assert(s == "0" || s == "1");
value = s[0] == '1';
} else if constexpr (std::is_integral_v<T>) {
auto result = std::from_chars(s.data(), s.data() + s.size(), value);
assert(result.ec == std::errc{} && result.ptr == s.data() + s.size());
(void)result;
} else value = T(s);
return true;
}
};
class FastOutput {
friend struct fast_io_detail::BulkOutput;
inline static constexpr auto quads = [] {
std::array<std::array<char, 4>, 10000> table{};
for (int i = 0; i < 10000; ++i)
table[i] = {char('0' + i / 1000), char('0' + i / 100 % 10), char('0' + i / 10 % 10), char('0' + i % 10)};
return table;
}();
FILE *file;
std::unique_ptr<char[]> buffer = std::make_unique_for_overwrite<char[]>(1 << 16);
std::size_t capacity = 1 << 16, length = 0;
char *reserve(std::size_t n) {
if (n > capacity - length) {
if (!flush()) throw std::runtime_error("output failed");
if (n > capacity) {
std::size_t next = std::max(n, 2 * capacity);
buffer = std::make_unique_for_overwrite<char[]>(next);
capacity = next;
}
}
return buffer.get() + length;
}
public:
explicit FastOutput(FILE *output = stdout) : file(output) {}
FastOutput(const FastOutput &) = delete;
FastOutput &operator=(const FastOutput &) = delete;
~FastOutput() { flush(); }
bool flush() {
std::size_t offset = 0;
while (offset < length) {
auto n = std::fwrite(buffer.get() + offset, 1, length - offset, file);
if (!n) {
std::memmove(buffer.get(), buffer.get() + offset, length - offset);
length -= offset;
return false;
}
offset += n;
}
length = 0;
return std::fflush(file) == 0;
}
void write(char c) { *reserve(1) = c; ++length; }
void write(__uint128_t value) {
char digits[40];
char *end = digits + sizeof(digits), *begin = end;
do { *--begin = char('0' + value % 10); value /= 10; } while (value);
write(std::string_view(begin, end - begin));
}
void write(__int128_t value) {
__uint128_t magnitude = __uint128_t(value);
if (value < 0) write('-'), magnitude = 0 - magnitude;
write(magnitude);
}
void write(std::string_view s) {
char *out = reserve(s.size());
if (!s.empty()) std::memcpy(out, s.data(), s.size());
length += s.size();
}
template <class T> requires std::is_integral_v<T>
void write(T value) {
char *out = reserve(32);
if constexpr (std::is_same_v<T, bool>) *out = '0' + value, ++length;
else if constexpr (std::is_unsigned_v<T> && sizeof(T) == 4) {
char *p = out;
if (value >= 100000000) {
auto high = value / 100000000, low = value % 100000000;
p = std::to_chars(p, p + 2, high).ptr;
std::memcpy(p, quads[low / 10000].data(), 4);
std::memcpy(p + 4, quads[low % 10000].data(), 4);
p += 8;
} else if (value >= 10000) {
p = std::to_chars(p, p + 4, value / 10000).ptr;
std::memcpy(p, quads[value % 10000].data(), 4);
p += 4;
} else p = std::to_chars(p, p + 4, value).ptr;
length += p - out;
} else length += std::to_chars(out, out + 32, value).ptr - out;
}
template <class T> requires requires(const T &value, char *out) { value.decimal_capacity(); value.to_chars(out); }
void write(const T &value) {
char *out = reserve(value.decimal_capacity());
length += value.to_chars(out) - out;
}
template <class T>
void println(const T &value) { write(value); write('\n'); }
};
#line 4 "test/suffix_array_public.test.cpp"
int main() {
FastInput input;
FastOutput output;
std::string s;
input.read(s);
for (int x : suffix_array(s)) { output.write(x); output.write(' '); }
output.write('\n');
}