#include <algorithm>
#include <stdexcept>
#include <array>
#include <vector>
#include <iostream>

#define _byteswap_uint64 __builtin_bswap64
#define _rotr(x,k) (((x)>>(k)) | ((x)<<(32-(k))))

template<typename T> struct Output {
    std::array<T, 8> h;
};
template<typename T> struct Input {
    std::array<T, 16> c;
};
template<typename T> Output<T> sha2(Input<T> in) {
    T w[64];
    for(int i = 0; i < 16; i++)
        w[i] = in.c[i];

    for(int i = 16; i < 64; i++) {
        auto s0 = _rotr(w[i - 15], 7) ^ _rotr(w[i - 15], 18) ^ (w[i - 15] >> 3);

        auto s1 = _rotr(w[i - 2], 17) ^ _rotr(w[i - 2], 19) ^ (w[i - 2] >> 10);
        w[i] = w[i - 16] + s0 + w[i - 7] + s1;
    }

    static const T k[] = {
        0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5,
        0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
        0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3,
        0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
        0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc,
        0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
        0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7,
        0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
        0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13,
        0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
        0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3,
        0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
        0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5,
        0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
        0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208,
        0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
    };

    static const T h[] = {
        0x6a09e667,
        0xbb67ae85,
        0x3c6ef372,
        0xa54ff53a,
        0x510e527f,
        0x9b05688c,
        0x1f83d9ab,
        0x5be0cd19
    };

    T loopvars[8];
    for(int i = 0; i < 8; i++)
        loopvars[i] = h[i];

    for(int i = 0; i < 64; i++) {
        auto&& la = loopvars[0];
        auto&& lb = loopvars[1];
        auto&& lc = loopvars[2];
        auto&& ld = loopvars[3];
        auto&& le = loopvars[4];
        auto&& lf = loopvars[5];
        auto&& lg = loopvars[6];
        auto&& lh = loopvars[7];

        auto s0 = _rotr(la, 2) ^ _rotr(la, 13) ^ _rotr(la, 22);
        auto maj = (la & lb) ^ (la & lc) ^ (lb & lc);
        auto t2 = s0 + maj;

        auto s1 = _rotr(le, 6) ^ _rotr(le, 11) ^ _rotr(le, 25);
        auto ch = (le & lf) ^ ((~le) & lg);
        auto t1 = lh + s1 + ch + k[i] + w[i];

        lh = lg;
        lg = lf;
        lf = le;
        le = ld + t1;
        ld = lc;
        lc = lb;
        lb = la;
        la = t1 + t2;
    }
    Output<T> output;
    for(int i = 0; i < 8; i++) {
        output.h[i] = h[i] + loopvars[i];
    }
    return output;
}
Output<unsigned int> SHA2(std::vector<char> bytes) {
    auto bitlen = bytes.size() * 8;
    auto big_endian_bitlen = ::_byteswap_uint64(bitlen);
    if (bitlen > 440)
        throw std::runtime_error("Epic fail!");
    Input<unsigned int> in;
    for(int i = 0; i < 16; i++) {
        in.c[i] = 0;
    }
    for(int i = 0; i < bitlen; i++) {
        in.c[i / 32] |= ((bytes[i / 8] >> (i % 8))) << (i % 32);
    }
    in.c[bitlen / 32] |= (1 << (bitlen % 32));
    // all zero by default, so no need to append the extra bits
    in.c[14] = (big_endian_bitlen >> 32);
    in.c[15] = big_endian_bitlen;
    return sha2(in);
}

int main(int argc, const char *argv[])
{
    std::vector<char> bytes;
    for(auto o : SHA2(bytes).h)
        std::cout << std::hex << o << ' ';
    std::cout << '\n';
}