#include <string.h>
#include "core.h"

void process_data(state* s, unsigned char* out, const unsigned char* in,
                  unsigned long long len, u8 mode) {
  u32_2 t0;
  u64 tmp0;

  while (len >= RATE) {
    tmp0 = (*(u64*)in);
    t0 = to_big(tmp0);
    s->x0.h ^= t0.h;
    s->x0.l ^= t0.l;
    if (mode != ASCON_AD) {
      tmp0 = from_big(s->x0);
      *(u64*)out = (tmp0);
    }
    if (mode == ASCON_DEC) {
      s->x0 = t0;
    }
    P(s, START_ROUND(PB_ROUNDS));
    in += RATE;
    out += RATE;
    len -= RATE;
  }

  u8 bytes[8];
  memcpy(bytes, in, len);
  memset(bytes + len, 0, 8 - len);
  bytes[len] ^= 0x80;

  u64* tmp = (u64*)bytes;

  t0 = to_big(tmp[0]);
  s->x0.h ^= t0.h;
  s->x0.l ^= t0.l;

  if (mode != ASCON_AD) {
    tmp[0] = from_big(s->x0);
    memcpy(out, bytes, len);
  }

  if (mode == ASCON_DEC) {
    memcpy(bytes, in, len);
    s->x0 = to_big(tmp[0]);
  }
}

void ascon_core(state* s, unsigned char* out, const unsigned char* in,
                unsigned long long tlen, const unsigned char* ad,
                unsigned long long adlen, const unsigned char* npub,
                const unsigned char* k, u8 mode) {
  u32_2 K0, K1, K2, N0, N1;

  // load key and nonce
  K0 = to_big((*(u64*)k));
  K1 = to_big((*(u64*)(k + 4)));
  K2 = to_big((*(u64*)(k + 12)));
  N0 = to_big((*(u64*)npub));
  N1 = to_big((*(u64*)(npub + 8)));

  // initialization
  to_big_immediate(s->x0, IV);
  s->x0.l = K0.h;
  s->x1.h = K1.h;
  s->x1.l = K1.l;
  s->x2.h = K2.h;
  s->x2.l = K2.l;
  s->x3.h = N0.h;
  s->x3.l = N0.l;
  s->x4.h = N1.h;
  s->x4.l = N1.l;
  P(s, START_ROUND(PA_ROUNDS));
  s->x2.l ^= K0.h;
  s->x3.h ^= K1.h;
  s->x3.l ^= K1.l;
  s->x4.h ^= K2.h;
  s->x4.l ^= K2.l;

  // process associated data
  if (adlen) {
    process_data(s, (void*)0, ad, adlen, ASCON_AD);
    P(s, START_ROUND(PB_ROUNDS));
  }
  s->x4.l ^= 1;

  // process plaintext/ciphertext
  process_data(s, out, in, tlen, mode);

  // finalization
  s->x1.h ^= K0.h;
  s->x1.l ^= K1.h;
  s->x2.h ^= K1.l;
  s->x2.l ^= K2.h;
  s->x3.h ^= K2.l;
  P(s, START_ROUND(PA_ROUNDS));
  s->x3.h ^= K1.h;
  s->x3.l ^= K1.l;
  s->x4.h ^= K2.h;
  s->x4.l ^= K2.l;
}