cp-library
競技プログラミング用 C++ ライブラリ
test/z_algorithm.test.cpp
test/z_algorithm.test.cpp · GitHub で見る
依存ライブラリ
コード
#define PROBLEM "https://judge.yosupo.jp/problem/zalgorithm"
#include "string/z_algorithm.hpp"
#include "components/fast_io.hpp"
int main() {
FastInput input;
FastOutput output;
std::string s;
input.read(s);
auto z = z_algorithm(s);
for (int i = 0; i < int(z.size()); ++i) {
if (i) output.write(' ');
output.write(z[i]);
}
output.write('\n');
}
#line 1 "test/z_algorithm.test.cpp"
#define PROBLEM "https://judge.yosupo.jp/problem/zalgorithm"
#line 1 "string/z_algorithm.hpp"
#line 1 "components/sequence_match.hpp"
#line 1 "components/sequence_compare.hpp"
#include <bit>
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <type_traits>
#if defined(__x86_64__)
#include <immintrin.h>
#endif
namespace sequence_compare_detail {
#if defined(__x86_64__)
__attribute__((target("avx2"))) inline std::size_t forward_avx2(const unsigned char *a, const unsigned char *b, std::size_t size) {
std::size_t i = 0;
for (; size - i >= 32; i += 32) {
auto x = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(a + i));
auto y = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(b + i));
auto difference = ~std::uint32_t(_mm256_movemask_epi8(_mm256_cmpeq_epi8(x, y)));
if (difference) return i + std::countr_zero(difference);
}
while (i < size && a[i] == b[i]) ++i;
return i;
}
__attribute__((target("avx2"))) inline std::size_t backward_avx2(const unsigned char *a, const unsigned char *b, std::size_t size) {
std::size_t i = 0;
for (; size - i >= 32; i += 32) {
auto x = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(a - i - 32));
auto y = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(b - i - 32));
auto difference = ~std::uint32_t(_mm256_movemask_epi8(_mm256_cmpeq_epi8(x, y)));
if (difference) return i + std::countl_zero(difference);
}
while (i < size && a[-std::ptrdiff_t(i) - 1] == b[-std::ptrdiff_t(i) - 1]) ++i;
return i;
}
#endif
template<bool Vectorize = true, class T>
inline int prefix(const T *first, const T *second, int limit) {
if (!limit) return 0;
int matched = 0;
if constexpr (std::is_integral_v<T> && std::has_unique_object_representations_v<T>) {
const auto *a = reinterpret_cast<const unsigned char *>(first);
const auto *b = reinterpret_cast<const unsigned char *>(second);
std::size_t offset = 0, size = std::size_t(limit) * sizeof(T);
while (size - offset >= 8) {
std::uint64_t x, y;
std::memcpy(&x, a + offset, 8);
std::memcpy(&y, b + offset, 8);
auto difference = x ^ y;
if (difference) {
int bits = std::endian::native == std::endian::little ? std::countr_zero(difference) : std::countl_zero(difference);
return static_cast<int>((offset + bits / 8) / sizeof(T));
}
offset += 8;
#if defined(__x86_64__)
if constexpr (Vectorize) if (size - offset >= 64 && __builtin_cpu_supports("avx2"))
return static_cast<int>((offset + forward_avx2(a + offset, b + offset, size - offset)) / sizeof(T));
#endif
}
matched = static_cast<int>(offset / sizeof(T));
}
while (matched < limit && first[matched] == second[matched]) ++matched;
return matched;
}
template<bool Vectorize = true, class T>
inline int suffix(const T *first_end, const T *second_end, int limit) {
if (!limit) return 0;
int matched = 0;
if constexpr (std::is_integral_v<T> && std::has_unique_object_representations_v<T>) {
const auto *a = reinterpret_cast<const unsigned char *>(first_end);
const auto *b = reinterpret_cast<const unsigned char *>(second_end);
std::size_t offset = 0, size = std::size_t(limit) * sizeof(T);
while (size - offset >= 8) {
std::uint64_t x, y;
std::memcpy(&x, a - offset - 8, 8);
std::memcpy(&y, b - offset - 8, 8);
auto difference = x ^ y;
if (difference) {
int bits = std::endian::native == std::endian::little ? std::countl_zero(difference) : std::countr_zero(difference);
return static_cast<int>((offset + bits / 8) / sizeof(T));
}
offset += 8;
#if defined(__x86_64__)
if constexpr (Vectorize) if (size - offset >= 64 && __builtin_cpu_supports("avx2"))
return static_cast<int>((offset + backward_avx2(a - offset, b - offset, size - offset)) / sizeof(T));
#endif
}
matched = static_cast<int>(offset / sizeof(T));
}
while (matched < limit && first_end[-matched - 1] == second_end[-matched - 1]) ++matched;
return matched;
}
}
#line 1 "components/z_algorithm.hpp"
namespace sequence_match_detail {
template <class F>
void z_table(int n, int *z, F match) {
if (!n) return;
z[0] = n;
for (int i = 1, base = 0, end = 0; i < n; ++i) {
int length = 0;
if (i < end) {
length = z[i - base];
if (length < end - i) {
z[i] = length;
continue;
}
length = end - i;
}
length += match(length, i + length, n - i - length);
z[i] = length;
if (i + length > end) base = i, end = i + length;
}
}
template <class F, class Emit>
void prefix_matches(int n, int m, const int *z, F match, Emit emit) {
for (int i = 0, base = 0, end = 0; i < n; ++i) {
int length = 0;
if (i < end) {
length = z[i - base];
if (length < end - i) {
emit(i, length);
continue;
}
length = end - i;
}
length += match(length, i + length, m - length);
if (i + length > end) base = i, end = i + length;
emit(i, length);
}
}
}
#line 6 "components/sequence_match.hpp"
#include <algorithm>
#include <array>
#include <span>
#include <utility>
#include <vector>
namespace sequence_match_detail {
#if defined(__x86_64__)
using sequence_compare_detail::forward_avx2, sequence_compare_detail::backward_avx2;
#endif
template <class T, bool Vectorize = true>
class Matcher {
const T *data;
std::vector<std::array<int, 2>> runs;
public:
explicit Matcher(std::span<const T> values) : data(values.data()) {}
void index_runs(int n) {
runs.resize(n);
for (int first = 0, last; first < n; first = last) {
last = first + 1;
while (last < n && data[first] == data[last]) ++last;
for (int i = first; i < last; ++i) runs[i] = {first, last};
}
}
std::pair<int, int> forward_counted(int i, int j, int limit) const {
if (runs.empty() || !limit || data[i] != data[j]) {
int length = forward(i, j, limit);
return {length, length};
}
int a = runs[i][1] - i, b = runs[j][1] - j, skip = std::min({limit, a, b});
if (a != b || skip == limit) return {skip, 1};
int length = forward(i + skip, j + skip, limit - skip);
return {skip + length, length + 1};
}
std::pair<int, int> backward_counted(int i, int j, int limit) const {
if (runs.empty() || !limit || data[i - 1] != data[j - 1]) {
int length = backward(i, j, limit);
return {length, length};
}
int a = i - runs[i - 1][0], b = j - runs[j - 1][0], skip = std::min({limit, a, b});
if (a != b || skip == limit) return {skip, 1};
int length = backward(i - skip, j - skip, limit - skip);
return {skip + length, length + 1};
}
int forward(int i, int j, int limit) const {
if (!limit) return 0;
return sequence_compare_detail::prefix<Vectorize>(data + i, data + j, limit);
}
int backward(int i, int j, int limit) const {
if (!limit) return 0;
return sequence_compare_detail::suffix<Vectorize>(data + i, data + j, limit);
}
};
}
#line 5 "string/z_algorithm.hpp"
#include <cassert>
#include <limits>
#line 8 "string/z_algorithm.hpp"
#include <string_view>
#line 10 "string/z_algorithm.hpp"
template <class T>
std::vector<int> z_algorithm(std::span<const T> sequence) {
assert(sequence.size() <= std::size_t(std::numeric_limits<int>::max()));
std::vector<int> result(sequence.size());
sequence_match_detail::Matcher<T> match(sequence);
sequence_match_detail::z_table(int(sequence.size()), result.data(), [&](int a, int b, int limit) { return match.forward(a, b, limit); });
return result;
}
inline std::vector<int> z_algorithm(std::string_view sequence) {
return z_algorithm(std::span<const char>(sequence.data(), sequence.size()));
}
template <class T, class Allocator>
std::vector<int> z_algorithm(const std::vector<T, Allocator> &sequence) {
return z_algorithm(std::span<const T>(sequence.data(), sequence.size()));
}
#line 1 "components/fast_io.hpp"
#line 8 "components/fast_io.hpp"
#include <charconv>
#line 11 "components/fast_io.hpp"
#include <cstdio>
#line 13 "components/fast_io.hpp"
#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/z_algorithm.test.cpp"
int main() {
FastInput input;
FastOutput output;
std::string s;
input.read(s);
auto z = z_algorithm(s);
for (int i = 0; i < int(z.size()); ++i) {
if (i) output.write(' ');
output.write(z[i]);
}
output.write('\n');
}