#include <cstdint>
#include <vector>
#include <iostream>
#include <cstring>

#pragma GCC target("avx2")
#pragma GCC optimize("O3,unroll-loops")

#include <immintrin.h>

using namespace std;

#define INF 1e9

int minfunc(int *arr, int n) {
	int ans = INF;
	int i = 0;
	__m256i rans = _mm256_set1_epi32(INF);
	for (; (uintptr_t)&arr[i] % 32; ++i)
		ans = min(ans, arr[i]);
	for (; i + 24 <= n; i += 24) {
		__m256i ra0 = _mm256_load_si256((__m256i*)&arr[i]);
		__m256i ra8 = _mm256_load_si256((__m256i*)&arr[i + 8]);
		__m256i ra16 = _mm256_load_si256((__m256i*)&arr[i + 16]);
		__m256i rm08 = _mm256_min_epi32(ra0, ra8);
		rans = _mm256_min_epi32(rans, ra16);
		rans = _mm256_min_epi32(rans, rm08);
		// rans = _mm256_min_epi32(rans, ra0);
		// rans = _mm256_min_epi32(rans, ra8);
		// rans = _mm256_min_epi32(rans, ra16);
	}
	int res[8];
	_mm256_storeu_si256((__m256i*)res, rans);
	ans = min(ans, res[0]);
	ans = min(ans, res[1]);
	ans = min(ans, res[2]);
	ans = min(ans, res[3]);
	ans = min(ans, res[4]);
	ans = min(ans, res[5]);
	ans = min(ans, res[6]);
	ans = min(ans, res[7]);
	for (; i < n; ++i)
		ans = min(ans, arr[i]);
	return ans;
}

bool iseq(char* s, char* t, int n) {
	int i = 0;
	__m256i rans = _mm256_set1_epi32(-1);
	for (; i + 32 <= n; i += 32) {
		__m256i rs = _mm256_lddqu_si256((__m256i*)&s[i]);
		__m256i rt = _mm256_lddqu_si256((__m256i*)&t[i]);
		__m256i rmask = _mm256_cmpeq_epi8(rs, rt);
		rans = _mm256_and_si256(rans, rmask);
	}
	long long res[4];
	_mm256_storeu_si256((__m256i*)res, rans);
	if ((res[0] & res[1] & res[2] & res[3]) != -1LL)
		return false;
	for (; i < n; ++i)
		if (s[i] != t[i]) return false;
	// return memcmp(&s[i], &t[i], n - i) == 0;
	return true;
}

bool iseq2(char* s, char* t, int n) {
	int i = 0;
	__m256i r1 = _mm256_set1_epi32(-1);
	for (; i + 32 <= n; i += 32) {
		__m256i rs = _mm256_lddqu_si256((__m256i*)&s[i]);
		__m256i rt = _mm256_lddqu_si256((__m256i*)&t[i]);
		__m256i rmask = _mm256_cmpeq_epi8(rs, rt);
		if (_mm256_testc_si256(rmask, r1) == 0) return false;
	}
	for (; i < n; ++i)
		if (s[i] != t[i]) return false;
	// return memcmp(&s[i], &t[i], n - i) == 0;
	return true;
}

void change3(int *arr, int n, int k, int x) {
	int i = 0;
	__m256i rk = _mm256_set1_epi32(k);
	__m256i rx = _mm256_set1_epi32(x);
	for (; i + 8 <= n; i += 8) {
		__m256i ra = _mm256_lddqu_si256((__m256i*)&arr[i]);
		__m256i rmask = _mm256_cmpgt_epi32(ra, rk);
		ra = _mm256_or_si256(
				_mm256_and_si256(rmask, rx),
				_mm256_andnot_si256(rmask, ra)
			);
		_mm256_storeu_si256((__m256i*)&arr[i], ra);
	}
	for (; i < n; ++i)
		if (arr[i] > k) arr[i] = x;
}

int main() {
	vector<int> arr = {4, 1, 2,3, 4, 6, 7, 2, 4, 687, 8, -10, 9, 123, 4123, 123, 5432, 123, 532, 123};
	cout << minfunc(&arr[0], arr.size()) << "\n";

	string s(100, 'a');
	string t(100, 'a');
	t[40] = 'r';
	cout << iseq(&s[0], &t[0], s.size()) << "\n";
}