#include <stdbool.h>
#include <string.h>
#include <inttypes.h>           /* debug */

#include "constants.h"
#include "parameters.h"
#include "tweakey.h"


#define LANE_BITS  64
#define LANE_BYTES (LANE_BITS/8)
#define LANES_NB (TWEAKEY_BYTES/LANE_BYTES)


static void _dump_buffer(FILE *output, const char *header, size_t len, const uint8_t buf[len], int indent)
{
    if (!output)
    {
        return;
    }

    fprintf(output, "%s\n", header);

    for (size_t line=0; line<len/8; line++)
    {
        fprintf(output, "%*s", indent, "");
        for (size_t b=0; b<8; b++)
        {
            /* start with MSB */
            size_t byte_index = len-(1+line*8+b);
            fprintf(output, "%*s", 5, "");
            fprintf(output, "%02x", buf[byte_index]);
        }
        fprintf(output, "\n");
    }
    fprintf(output, "\n");
}


void tweakey_state_init(
    tweakey_state *TK,
    const uint8_t key[KEY_BYTES],
    const uint8_t tweak[TWEAK_BYTES],
    FILE *debug
)
{
    memcpy(TK->TK,             tweak, TWEAK_BYTES);
    memcpy(TK->TK+TWEAK_BYTES, key,   KEY_BYTES);

    TK->debug = debug;
    _dump_buffer(debug, "     Tweak is :", TWEAK_BYTES, tweak, 5);
    _dump_buffer(debug, "     Key is :", KEY_BYTES, key, 5);
    _dump_buffer(debug, "     Tweakey is :", sizeof(TK->TK), TK->TK, 5);
}


void tweakey_state_extract(
    const tweakey_state *TK,
    uint8_t round_tweakey[ROUND_TWEAKEY_BYTES], /* output */
    uint8_t i                                   /* round constant */
)
{
    memset(round_tweakey, 0, ROUND_TWEAKEY_BYTES);

    for (const uint8_t *lane=TK->TK; lane<TK->TK+TWEAKEY_BYTES; lane+=LANE_BYTES)
    {
        for (size_t j=0; j<LANE_BYTES; j++)
        {
            round_tweakey[j] ^= lane[j];
        }
    }

    round_tweakey[0] ^= i;

    char debug[512];
    snprintf(debug, sizeof(debug), "     Extracting Subtweakey round  %"PRIu8, i);
    _dump_buffer(TK->debug, debug, sizeof(TK->TK), TK->TK, 5);
    _dump_buffer(TK->debug, "     Subtweakey :", ROUND_TWEAKEY_BYTES, round_tweakey, 5);
}


static void _permute_state(tweakey_state *TK)
{
    uint8_t TK_old[TWEAKEY_BYTES];
    memcpy(TK_old, TK->TK, sizeof(TK_old));

    /* TODO: homogenize indices; here j=lane; k=byte */

    for (size_t j=0; j<TWEAKEY_BYTES; j+=LANE_BYTES)
    {
        for (size_t k=0; k<LANE_BYTES; k++)
        {
            TK->TK[j+k] = TK_old[j+h[k]];
        }
    }
}

static void _multiply_state(tweakey_state *TK)
{
    /* Lane 0 is multiplied by Id; lane 1 by P_0, lane 2 by P_1... */

    for (size_t lane=1; lane<LANES_NB; lane++)
    {
        const uint8_t* P_lane = P[lane-1];

        /* TODO: homogenize indices; here b=byte */

        for (size_t b=0; b<LANE_BYTES; b++)
        {
            size_t offset = lane*LANE_BYTES + b;
            TK->TK[offset] = P_lane[TK->TK[offset]];
        }
    }
}

void tweakey_state_update(tweakey_state *TK)
{
    _dump_buffer(TK->debug, "          Input Tweakey :", sizeof(TK->TK), TK->TK, 10);

    _permute_state(TK);

    _dump_buffer(TK->debug, "          Post permutation Tweakey :", sizeof(TK->TK), TK->TK, 10);

    _multiply_state(TK);

    _dump_buffer(TK->debug, "          Post multiplication Tweakey :", sizeof(TK->TK), TK->TK, 10);
}