// g++ -std=c++17 -march=native -O3 -lcrypto -lpthread gewalt.cpp -o gewalt

#include <cassert>
#include <iostream>
#include <iomanip>
#include <string>
#include <functional>
#include <random>
#include <unordered_map>
#include <algorithm>
#include <thread>
#include <atomic>
#include <mutex>

const unsigned num_threads = std::thread::hardware_concurrency();

#include <openssl/sha.h>

static std::string hash(std::string const& s)
{
    SHA_CTX ctx;
    if (!SHA1_Init(&ctx)) throw;
    if (!SHA1_Update(&ctx, s.data(), s.length())) throw;
    std::string d(SHA_DIGEST_LENGTH, 0);
    if (!SHA1_Final((uint8_t *) &d[0], &ctx)) throw;
    return d;
}

static std::u32string kapot(std::string const& s)
{
    std::u32string r(s.size(), 0);
    size_t o = 0;

    for (size_t i = 0; i < s.length(); ) {

        auto T = [](uint8_t c) {
            return (c < 0x80)         ? 1   /* ASCII */
                 : (c & 0xc0) == 0x80 ? 0   /* continuation */
                 : (c & 0xe0) == 0xc0 ? 2   /* 2-byte chunk */
                 : (c & 0xf0) == 0xe0 ? 3   /* 3-byte chunk */
                 : (c & 0xf8) == 0xf0 ? 4   /* 4-byte chunk */
                 : -1;
        };

        uint32_t c = s[i++];
        auto cont = [&]() { c = (c << 6) | (s[i++] & 0x3f); };

        switch (T(c)) {

        case -1:
        case  0:
        invalid: c = 0xfffd; /* fall through */

        case  1:
        valid:   r[o++] = c; break;

        case  2:
                 if (c &= 0x1f, i+0 >= s.size() || T(s[i+0]))
                     goto invalid;
                 goto one;

        case  3:
                 if (c &= 0x1f, i+1 >= s.size() || T(s[i+0]) || T(s[i+1]))
                     goto invalid;
                 goto two;

        case  4:
                 if (c &= 0x1f, i+2 >= s.size() || T(s[i+0]) || T(s[i+1]) || T(s[i+2]))
                     goto invalid;
                 cont();
        two:     cont();
        one:     cont();
                 goto valid;

        }

    }

    r.resize(o);

    return r;
}

std::atomic<uint64_t> hcount = 0, kcount = 0;
typedef std::unordered_map<std::u32string, std::string> tab_t;
tab_t tab0, tab1;
std::mutex mtx;

std::array<uint8_t,4> ip;
std::string cmd0, cmd1;

class stuffer_t
{
    private:
        std::array<size_t,4> cnts;
        size_t step;
        std::string cmd;
    public:
        stuffer_t(size_t t, size_t s, std::string c) : cnts{t}, step(s), cmd(c) {}
        std::string operator()()
        {
            //XXX this is by far not the most efficient way of doing this, but yeah
            if (++cnts[3] >= cnts[0]) {
                cnts[3] = 0;
                if (++cnts[2] >= cnts[0]) {
                    cnts[2] = 0;
                    if (++cnts[1] >= cnts[0]) {
                        cnts[1] = 0;
                        cnts[0] += step;
                    }
                }
            }
            std::stringstream o;
            for (size_t i = 0; i < 4; ++i)
                o << (i ? "." : "")
                  << std::string(cnts[i], '0')
                  << (unsigned) ip[i];
            o << "|" << cmd;
            return o.str();
        }
};

void go(size_t tid)
{
    //XXX tid stuff is a hack, but YOLO

    bool one = tid & 1;

    stuffer_t next(tid >> 1, (num_threads + 1) >> 1, one ? cmd1 : cmd0);

    tab_t& mytab = one ? tab1 : tab0;
    tab_t& thtab = one ? tab0 : tab1;

    uint64_t myhcount = 0, mykcount = 0;

    while (1) {

        std::string r = next();

        {

            ++myhcount;

            auto h = hash(r);
            if ((h.size()+3)/4 < (size_t) std::count_if(h.begin(), h.end(),
                                            [](unsigned char c) { return c < 0x80; }))
                continue;

            ++mykcount;

            auto k = kapot(h);
            if (k.size() > 3 + (size_t) std::count(k.begin(), k.end(), 0xfffd))
                continue;

            std::lock_guard<std::mutex> lck(mtx);

            hcount += myhcount, myhcount = 0;
            kcount += mykcount, mykcount = 0;

            if (thtab.find(k) != thtab.end()) {

                mytab[k] = r;

                std::cerr << "\r\x1b[K"
                          << "\x1b[32m";
                std::cout << tab0[k] << std::endl
                          << tab1[k] << std::endl;
                std::cerr << "\x1b[0m";

                std::cerr << std::hex;
                bool first = true;
                for (uint32_t c: k)
                    std::cerr << (first ? first = false, "" : " ") << c;
                std::cerr << std::endl;

                std::cerr << std::dec << "hash count:  \x1b[35m" << hcount << "\x1b[0m";
                {
                    std::stringstream s;
                    s << std::fixed << std::setprecision(2) << log(hcount|1)/log(2);
                    std::cerr << " (2^\x1b[35m" << std::setw(5) << s.str() << "\x1b[0m" << ")" << std::endl;
                }
                std::cerr << "kapot count: " << "\x1b[35m" << kcount << "\x1b[0m";
                {
                    std::stringstream s;
                    s << std::fixed << std::setprecision(2) << log(kcount|1)/log(2);
                    std::cerr << " (2^\x1b[35m" << std::setw(5) << s.str() << "\x1b[0m)" << std::endl;
                }
                std::cerr << "table sizes: \x1b[35m"
                          << tab0.size() << "\x1b[0m \x1b[35m"
                          << tab1.size() << "\x1b[0m" << std::endl;

                exit(0);

            }

            if (mytab.size() < (1 << 20))
                mytab[k] = r;

        }

        hcount += myhcount;
        kcount += mykcount;

    }
}

void status()
{
    while (1) {

        {
            std::lock_guard<std::mutex> lck(mtx);

            std::cerr << "\r\x1b[K";
            std::cerr << "hash count: \x1b[35m" << std::setw(12) << hcount << "\x1b[0m ";
            {
                std::stringstream s;
                s << std::fixed << std::setprecision(2) << log(hcount|1)/log(2);
                std::cerr << "(2^\x1b[35m" << std::setw(5) << s.str() << "\x1b[0m) | ";
            }
            std::cerr << "kapot count: \x1b[35m" << std::setw(12) << kcount << "\x1b[0m ";
            {
                std::stringstream s;
                s << std::fixed << std::setprecision(2) << log(kcount|1)/log(2);
                std::cerr << "(2^\x1b[35m" << std::setw(5) << s.str() << "\x1b[0m) | ";
            }
            std::cerr << "tables: \x1b[35m"
                      << std::setw(9) << tab0.size() << " "
                      << std::setw(9) << tab1.size() << "\x1b[0m "
                      << std::flush;
        }

        std::this_thread::sleep_for(std::chrono::milliseconds(100));
    }
}

int main(int argc, char **argv)
{

    if (argc < 2) {
        std::cerr << "\x1b[31mneed IPv4 in argv[1]\x1b[0m" << std::endl;
        exit(1);
    }
    {
        std::stringstream ss(argv[1]);
        for (auto& v: ip) {
            std::string s;
            std::getline(ss, s, '.');
            int n = std::atoi(s.c_str());
            if (n < std::numeric_limits<uint8_t>::min() || n > std::numeric_limits<uint8_t>::max())
                goto bad_ip;
            v = n;
        }
        if (!ss) {
bad_ip:
            std::cerr << "\x1b[31mbad IPv4 given?\x1b[0m" << std::endl;
            exit(2);
        }
    }


    if (argc < 4) {
        std::cerr << "\x1b[31mneed commands in argv[2] and argv[3]\x1b[0m" << std::endl;
        exit(2);
    }
    cmd0 = argv[2];
    cmd1 = argv[3];


    std::thread status_thread(status);
    std::vector<std::thread> ts;
    for (unsigned i = 0; i < num_threads; ++i)
        ts.push_back(std::thread(go, i));
    for (auto& t: ts)
        t.join();

}