#include "owi.h"
#include "r_cg_wdt.h"
#include "delay.h"
#include "r_cg_timer.h"
#include "r_cg_cgc.h"
#include <string.h>

/* 내부 상태 */
#ifndef OWI_USE_TAU1_CAPTURE
#define OWI_USE_TAU1_CAPTURE       1u
#endif

/* Renesas CG: count-clock period is half the selected 0.063 us minimum.
   Therefore TI15 capture runs at approximately 32 timer ticks per us. */
#define OWI_TAU1_TICKS_PER_US      32UL
#define OWI_CAPTURE_QUEUE_WAIT_US  (bit_period_us * 4u)
#define OWI_CAPTURE_SYNC_BITS      16u

/* 내부 상태 */
static uint32_t bit_period_us = OWI_BIT_PERIOD_US;
static uint8_t  g_owi_timeout_latched = 0;
/* 0=none, 1=no pulse, 2=queue overflow, 3=header mismatch. */
static uint8_t g_capture_error = 0U;
static uint16_t g_capture_sync = 0U;
uint8_t OWI_GetCaptureError(void) { return g_capture_error; }
uint16_t OWI_GetCaptureSync(void) { return g_capture_sync; }
#if OWI_USE_TAU1_CAPTURE
static uint32_t g_owi_capture_threshold_ticks = 0UL;
static uint8_t  g_owi_capture_pending_bits[OWI_CAPTURE_SYNC_BITS];
static uint8_t  g_owi_capture_pending_index = 0U;
static uint8_t  g_owi_capture_pending_count = 0U;
#endif


static uint16_t g_owi_last_timeout_byte_index = 0xFFFFu;
static uint8_t  g_owi_last_timeout_bit_index  = 0xFFu;

static uint16_t g_owi_current_read_byte_index = 0u;
static uint8_t  g_owi_current_read_bit_index  = 0u;

static void OWI_KickWdt(void)
{
    R_WDT_Restart();
    NVM_TraceStack();
}

/* =========================================================
 * GPIO helpers (P70 / HW Open-Drain)
 * - LOW  : 출력 모드 + 0
 * - HIGH : 입력(Hi-Z)로 release
 * ========================================================= */
void GPIO_Clear(void)
{
    OWI_PORT_POM |= (uint8_t)OWI_PIN_MASK;
    OWI_PORT_P   &= (uint8_t)~OWI_PIN_MASK;
    OWI_PORT_PM  &= (uint8_t)~OWI_PIN_MASK;
}

void GPIO_Input(void)
{
    OWI_PORT_POM |= (uint8_t)OWI_PIN_MASK;
    OWI_PORT_P   |= (uint8_t)OWI_PIN_MASK;
    OWI_PORT_PM  |= (uint8_t)OWI_PIN_MASK;
}

int GPIO_Read(void)
{
    return (OWI_PORT_P & (uint8_t)OWI_PIN_MASK) ? 1 : 0;
}

static void OWI_Release(void)
{
    OWI_PORT_POM |= (uint8_t)OWI_PIN_MASK;
    OWI_PORT_P   |= (uint8_t)OWI_PIN_MASK;
    OWI_PORT_PM  |= (uint8_t)OWI_PIN_MASK;
}

static void OWI_DriveLow(void)
{
    OWI_PORT_POM |= (uint8_t)OWI_PIN_MASK;
    OWI_PORT_P   &= (uint8_t)~OWI_PIN_MASK;
    OWI_PORT_PM  &= (uint8_t)~OWI_PIN_MASK;
}

/* =========================================================
 * Diagnostics
 * ========================================================= */
uint8_t OWI_HasTimeout(void)
{
    return g_owi_timeout_latched;
}

void OWI_ClearTimeout(void)
{
    g_owi_timeout_latched = 0;
    g_capture_error = 0U;
    g_capture_sync = 0U;
    g_owi_last_timeout_byte_index = 0xFFFFu;
    g_owi_last_timeout_bit_index  = 0xFFu;
}

uint16_t OWI_GetLastTimeoutByteIndex(void)
{
    return g_owi_last_timeout_byte_index;
}

uint8_t OWI_GetLastTimeoutBitIndex(void)
{
    return g_owi_last_timeout_bit_index;
}

/* =========================================================
 * OWI Init
 * ========================================================= */
void OWI_Init(uint32_t bit_time_us)
{
    bit_period_us = bit_time_us;
    (void)bit_period_us;
    PMC7 &= (uint8_t)~OWI_PIN_MASK;
    PIM7 &= (uint8_t)~OWI_PIN_MASK;

    OWI_PORT_POM |= (uint8_t)OWI_PIN_MASK;
    OWI_PORT_PU &= (uint8_t)~OWI_PIN_MASK;

    OWI_ClearTimeout();
    GPIO_Input();
    OWI_KickWdt();
}

/* =========================================================
 * Start/Stop/Secure
 * ========================================================= */
void OWI_Start(void)
{
    GPIO_Clear();
    delay_us(TSTART_HOLD);

    GPIO_Input();
    delay_us(TBIT);
}

void OWI_Stop(void)
{
    GPIO_Input();
    delay_us(TSTOP_LOW);
    delay_us(TIDLE);
}

static void OWI_StopRead(void)
{
    OWI_DriveLow();
    delay_us(TSTOP_LOW);
    OWI_Release();
    delay_us(TIDLE);
}

static uint8_t OWI_WaitForFallingEdge(uint16_t byte_index, uint8_t bit_index)
{
    int timeout = (int)(bit_period_us * 4u);

    while (GPIO_Read() && timeout-- > 0) {
        delay_us(1u);
    }

    if (timeout <= 0) {
        if (!g_owi_timeout_latched) {
            g_owi_timeout_latched = 1;
            g_owi_last_timeout_byte_index = byte_index;
            g_owi_last_timeout_bit_index  = bit_index;
        }
        return 0u;
    }

    return 1u;
}

static uint8_t OWI_ReadByte_StreamSynced(uint16_t byte_index)
{
    uint8_t data = 0;
    int b;

    for (b = 7; b >= 0; b--) {
        uint8_t bit_index = (uint8_t)(7 - b);
        uint8_t bit;

        g_owi_current_read_byte_index = byte_index;
        g_owi_current_read_bit_index  = bit_index;

        /* 슬롯 시작 후 중앙 샘플 */
        delay_us(bit_period_us / 2u);
        bit = GPIO_Read() ? 1u : 0u;
        data |= (uint8_t)(bit << b);

        /* 남은 반 슬롯 대기 */
        delay_us(bit_period_us / 2u);
    }

    return data;
}

#if OWI_USE_TAU1_CAPTURE
static void OWI_CapturePrepare(void)
{
    /* P70 remains a normal open-drain GPIO at boot.  Configure TI15 only
       immediately before a captured receive transaction. */
    /* R_Systeminit configured TAU1; do not reset unrelated channels. */
    R_TAU1_Channel5_Stop();
    /* Slow RC rising edges on P70 benefit from the TI15 digital noise filter.
       Give capture precedence over UART interrupts while receiving a frame. */
    NFEN2 |= _20_TAU_CH5_NOISE_ON;
    TMPR115 = 0U;
    TMPR015 = 0U;
    GPIO_Input();
}

static void OWI_CaptureBegin(void)
{
    /* Arm while the master still holds the final address LOW.
       Its release is the first edge; the first complete slave LOW follows. */
    g_owi_capture_threshold_ticks = 0UL;
    g_owi_capture_pending_index = 0U;
    g_owi_capture_pending_count = 0U;
    R_TAU1_Channel5_CaptureReset();
    g_trace_stage = 7U;
    R_TAU1_Channel5_Start();
}

/* Send the read address and arm capture BEFORE releasing its final LOW.
   Otherwise the GPIO/function-return overhead can lose the first response bit. */
static void OWI_WriteCapturedReadAddress(uint8_t address)
{
    int b;
    uint32_t t_low = (address & 1U) ? (uint32_t)TLOW_1 : (uint32_t)TLOW_0;
    for (b = 7; b > 0; b--) OWI_WriteBit((address >> b) & 1U);
    OWI_Release();
    delay_us((uint32_t)TBIT - t_low);
    OWI_DriveLow();
    delay_us(t_low);
    OWI_CaptureBegin();
    OWI_Release();
}

static void OWI_CaptureEnd(void)
{
    R_TAU1_Channel5_Stop();
    GPIO_Input();
}

static uint8_t OWI_CaptureWaitLowWidth(uint16_t byte_index,
                                       uint8_t bit_index,
                                       uint32_t * const low_ticks)
{
    uint32_t timeout_us = OWI_CAPTURE_QUEUE_WAIT_US;

    while (timeout_us-- > 0UL)
    {
        /* Keep the bounded wait alive even when the input stays idle. */
        if ((timeout_us & 63UL) == 0UL) OWI_KickWdt();
        if (R_TAU1_Channel5_CaptureOverflowed()) break;

        if (R_TAU1_Channel5_GetLowWidthTicks(low_ticks))
        {
            return 1U;
        }

        if (R_TAU1_Channel5_CaptureOverflowed())
        {
            break;
        }

        /* This delay is only a timeout yield; pulse timing is captured by TAU1. */
        delay_us(1U);
    }

    if (!g_owi_timeout_latched)
    {
        g_capture_error = R_TAU1_Channel5_CaptureOverflowed() ? 2U : 1U;
        g_owi_timeout_latched = 1U;
        g_owi_last_timeout_byte_index = byte_index;
        g_owi_last_timeout_bit_index = bit_index;
    }

    return 0U;
}

static uint32_t OWI_CaptureFindThreshold(const uint32_t *widths,
                                         uint8_t width_count)
{
    uint32_t sorted[OWI_CAPTURE_SYNC_BITS];
    uint32_t max_gap = 0UL;
    uint32_t threshold =
        ((uint32_t)bit_period_us * OWI_TAU1_TICKS_PER_US) / 2UL;
    uint32_t gap;
    uint32_t temp;
    uint8_t i;
    uint8_t j;

    for (i = 0U; i < width_count; i++) {
        sorted[i] = widths[i];
    }

    for (i = 0U; i < (uint8_t)(width_count - 1U); i++) {
        for (j = (uint8_t)(i + 1U); j < width_count; j++) {
            if (sorted[j] < sorted[i]) {
                temp = sorted[i];
                sorted[i] = sorted[j];
                sorted[j] = temp;
            }
        }
    }

    /* The 0x26 header contains both short and long LOW pulses.  The largest
       gap separates the two measured clusters without assuming timer scale. */
    for (i = 0U; i < (uint8_t)(width_count - 1U); i++) {
        gap = sorted[i + 1U] - sorted[i];
        if (gap > max_gap) {
            max_gap = gap;
            threshold = sorted[i] + (gap / 2UL);
        }
    }

    return threshold;
}

static uint8_t OWI_CaptureExpectedHeaderBit(uint8_t header_bit_index)
{
    return (uint8_t)((0x26U >> (7U - header_bit_index)) & 0x01U);
}

/* Find a complete header in the bounded startup window. The read-address
   release may add a leading capture (observed SYNC=9343: 1|00100110|1000011).
   Preserve every bit after the header; never reconstruct a partial header. */
static uint8_t OWI_CaptureFindHeaderEnd(const uint8_t *bits,
                                     uint8_t bit_count,
                                     uint8_t * const header_end)
{
    uint8_t start;
    uint8_t i;
    if (!bits || !header_end || bit_count < 8U ||
        bit_count > OWI_CAPTURE_SYNC_BITS) return 0U;
    for (start = 0U; start <= (uint8_t)(bit_count - 8U); start++) {
        for (i = 0U; i < 8U; i++) {
            if (bits[start + i] != OWI_CaptureExpectedHeaderBit(i)) break;
        }
        if (i == 8U) {
            *header_end = (uint8_t)(start + 8U);
            return 1U;
        }
    }
    return 0U;
}

static uint8_t OWI_CaptureGetDecodedBit(uint16_t byte_index,
                                        uint8_t bit_index,
                                        uint8_t * const decoded_bit)
{
    uint32_t low_ticks;

    if (decoded_bit == 0) {
        return 0U;
    }

    if (g_owi_capture_pending_index < g_owi_capture_pending_count) {
        *decoded_bit = g_owi_capture_pending_bits[g_owi_capture_pending_index++];
        return 1U;
    }

    if (!OWI_CaptureWaitLowWidth(byte_index, bit_index, &low_ticks)) {
        return 0U;
    }

    *decoded_bit = (low_ticks < g_owi_capture_threshold_ticks) ? 1U : 0U;
    return 1U;
}

static uint8_t OWI_ReadByte_Captured(uint16_t byte_index)
{
    uint8_t data = 0U;
    int b;
    uint32_t sync_widths[OWI_CAPTURE_SYNC_BITS];

    if (byte_index == 0U)
    {
        uint8_t sync_bits[OWI_CAPTURE_SYNC_BITS];
        uint8_t header_end;
        uint8_t i;

        for (i = 0U; i < OWI_CAPTURE_SYNC_BITS; i++)
        {
            OWI_KickWdt();
            g_owi_current_read_byte_index = 0U;
            g_owi_current_read_bit_index = i;

            if (!OWI_CaptureWaitLowWidth(0U, i, &sync_widths[i]))
            {
                return 0U;
            }
        }

        g_owi_capture_threshold_ticks =
            OWI_CaptureFindThreshold(sync_widths, OWI_CAPTURE_SYNC_BITS);

        for (i = 0U; i < OWI_CAPTURE_SYNC_BITS; i++) {
            sync_bits[i] =
                (sync_widths[i] < g_owi_capture_threshold_ticks) ? 1U : 0U;
        }

        g_capture_sync = 0U;
        for (i = 0U; i < OWI_CAPTURE_SYNC_BITS; i++) {
            g_capture_sync = (uint16_t)((g_capture_sync << 1) | sync_bits[i]);
        }
        if (!OWI_CaptureFindHeaderEnd(sync_bits,
                                      OWI_CAPTURE_SYNC_BITS,
                                      &header_end)) {
            if (!g_owi_timeout_latched) {
                g_owi_timeout_latched = 1U;
                g_capture_error = 3U;
                g_owi_last_timeout_byte_index = 0U;
                g_owi_last_timeout_bit_index = 0U;
            }
            return 0U;
        }

        g_owi_capture_pending_index = 0U;
        g_owi_capture_pending_count =
            (uint8_t)(OWI_CAPTURE_SYNC_BITS - header_end);
        for (i = 0U; i < g_owi_capture_pending_count; i++) {
            g_owi_capture_pending_bits[i] = sync_bits[header_end + i];
        }

        return 0x26U;
    }

    for (b = 7; b >= 0; b--)
    {
        uint8_t bit_index = (uint8_t)(7 - b);
        uint8_t bit;

        g_owi_current_read_byte_index = byte_index;
        g_owi_current_read_bit_index = bit_index;

        if (!OWI_CaptureGetDecodedBit(byte_index, bit_index, &bit))
        {
            break;
        }

        if (bit)
        {
            data |= (uint8_t)(1U << b);
        }
    }

    return data;
}
#endif

void OWI_SecureStop(void)
{
    int i;

    GPIO_Clear();
    delay_us(SECURE_HIGH);

    for (i = 0; i < (int)SECURE_TOGGLE_COUNT; i++) {
        GPIO_Input();
        delay_us(SECURE_TOGGLE_HIGH);

        GPIO_Clear();
        delay_us(SECURE_TOGGLE_LOW);
    }

    GPIO_Input();
    delay_us(SECURE_HIGH);

    GPIO_Clear();
    delay_us(TSTART_HOLD);
    GPIO_Input();
}

/* =========================================================
 * Bit/Byte IO
 * ========================================================= */
void OWI_WriteBit(int bit)
{
    uint32_t t_low  = bit ? (uint32_t)TLOW_1 : (uint32_t)TLOW_0;
    uint32_t t_high = (uint32_t)TBIT - t_low;

    /* HIGH 구간 */
    OWI_Release();
    delay_us(t_high);

    /* LOW 구간 */
    OWI_DriveLow();
    delay_us(t_low);

    /* 다음 슬롯 시작 전 반드시 release */
    OWI_Release();
}

void OWI_WriteByte(uint8_t data)
{
    int i;
    for (i = 7; i >= 0; i--) {
        OWI_WriteBit((data >> i) & 0x01u);
    }
    GPIO_Input();
}

uint8_t OWI_ReadBit(void)
{
    int timeout;
    int low_width = 0;

    /* 1) 라인이 HIGH가 될 때까지 정리 */
    timeout = (int)(bit_period_us * 2u);
    while (!GPIO_Read() && timeout-- > 0) {
        delay_us(1u);
    }

    if (timeout <= 0) {
        if (!g_owi_timeout_latched) {
            g_owi_timeout_latched = 1;
            g_owi_last_timeout_byte_index = g_owi_current_read_byte_index;
            g_owi_last_timeout_bit_index  = g_owi_current_read_bit_index;
        }
        return 0u;
    }

    /* 2) 다음 LOW 시작 대기 */
    timeout = (int)(bit_period_us * 2u);
    while (GPIO_Read() && timeout-- > 0) {
        delay_us(1u);
    }

    if (timeout <= 0) {
        if (!g_owi_timeout_latched) {
            g_owi_timeout_latched = 1;
            g_owi_last_timeout_byte_index = g_owi_current_read_byte_index;
            g_owi_last_timeout_bit_index  = g_owi_current_read_bit_index;
        }
        return 0u;
    }

    /* 3) LOW가 유지되는 시간을 직접 측정 */
    while (!GPIO_Read() && low_width < (int)(bit_period_us * 2u)) {
        delay_us(1u);
        low_width++;
    }

    /* 4) LOW 폭 기준으로 판정
       - 짧으면 1
       - 길면 0
       threshold ~= TBIT/2
    */
    return (low_width < (int)(bit_period_us / 2u)) ? 1u : 0u;
}

uint8_t OWI_ReadByte(void)
{
    uint8_t data = 0;
    int i;

    for (i = 7; i >= 0; i--) {
        g_owi_current_read_bit_index = (uint8_t)(7 - i);
        data |= (uint8_t)(OWI_ReadBit() << i);
        if (OWI_HasTimeout()) {
            break;
        }
    }
    return data;
}

/* =========================================================
 * Raw helpers
 * ========================================================= */
static void owi_result_init(owi_io_result_t *r)
{
    if (!r) return;
    memset(r, 0, sizeof(*r));
    r->ok = 1;
    r->timeout_byte_index = 0xFFFFu;
    r->timeout_bit_index  = 0xFFu;
}

static void owi_result_latch_timeout(owi_io_result_t *r)
{
    if (!r) return;

    if (OWI_HasTimeout()) {
        r->ok = 0;
        r->timeout = 1;
        r->timeout_byte_index = OWI_GetLastTimeoutByteIndex();
        r->timeout_bit_index  = OWI_GetLastTimeoutBitIndex();
    }
}

void OWI_T_CommandModeRaw(const uint8_t *tx_data, uint8_t tx_len, uint8_t id, owi_io_result_t *r)
{
    int i;

    owi_result_init(r);

    //delay_us(200);
    OWI_Init(OWI_BIT_PERIOD_US);
    OWI_ClearTimeout();

    OWI_KickWdt();
    OWI_SecureStop();
    OWI_KickWdt();
    OWI_WriteByte((uint8_t)(id << 1));
    OWI_KickWdt();

    for (i = 0; i < (int)tx_len; i++) {
        OWI_KickWdt();
        OWI_WriteByte(tx_data[i]);
        OWI_KickWdt();
    }

    OWI_Stop();
    OWI_KickWdt();
    owi_result_latch_timeout(r);
}

void OWI_CommandModeRaw(const uint8_t *tx_data, uint8_t tx_len, uint8_t id, owi_io_result_t *r)
{
    int i;

    owi_result_init(r);
    OWI_Init(OWI_BIT_PERIOD_US);
    OWI_ClearTimeout();

    OWI_KickWdt();
    OWI_SecureStop();
    OWI_KickWdt();
    OWI_WriteByte((uint8_t)(id << 1));
    OWI_KickWdt();

    for (i = 0; i < (int)tx_len; i++) {
        OWI_KickWdt();
        OWI_WriteByte(tx_data[i]);
        OWI_KickWdt();
    }

    OWI_Stop();
    OWI_KickWdt();
    owi_result_latch_timeout(r);
}

void OWI_ReadBytesRaw(int length, uint8_t id, owi_io_result_t *r)
{
    int i;
    int max_len;
    uint8_t is_long_read;

    if (!r) return;

    owi_result_init(r);
    OWI_Init(OWI_BIT_PERIOD_US);

    if (length <= 0) {
        r->ok = 0;
        return;
    }

    max_len = (int)OWI_IO_MAX_BYTES;
    if (length > max_len) {
        length = max_len;
    }

    r->read_len = 0U;
    OWI_ClearTimeout();

    is_long_read = (length >= 119) ? 1u : 0u;

    OWI_KickWdt();
#if OWI_USE_TAU1_CAPTURE
    if (length == 127) OWI_CapturePrepare();
#endif
    OWI_SecureStop();
#if OWI_USE_TAU1_CAPTURE
    if (length == 127) OWI_WriteCapturedReadAddress((uint8_t)((id << 1) | 1u));
    else
#endif
    OWI_WriteByte((uint8_t)((id << 1) | 1u));

    /* 기존 잘 되던 값 유지 */
#if OWI_USE_TAU1_CAPTURE
    if (length == 127) {
        for (i = 0; i < length; i++) {
            uint8_t value;
            OWI_KickWdt();
            g_trace_stage = 8U;
            g_trace_byte = (uint16_t)i;
            value = OWI_ReadByte_Captured((uint16_t)i);
            if (R_TAU1_Channel5_CaptureOverflowed() && !OWI_HasTimeout()) {
                g_owi_timeout_latched = 1U;
                g_owi_last_timeout_byte_index = (uint16_t)i;
                g_owi_last_timeout_bit_index = 0xFFU;
                g_capture_error = 2U;
            }
            if (OWI_HasTimeout()) {
                owi_result_latch_timeout(r);
                break;
            }
            r->data[r->read_len++] = value;
        }
        g_trace_stage = 9U;
        OWI_CaptureEnd();
        OWI_StopRead();
        OWI_KickWdt();
        return;
    }

#endif
    delay_us(100u);
    OWI_KickWdt();

    /* 첫 falling edge를 한 번 잡는다 */
    if (!OWI_WaitForFallingEdge(0u, 0u)) {
        owi_result_latch_timeout(r);
        OWI_StopRead();
        OWI_KickWdt();
        return;
    }

    for (i = 0; i < length; i++) {
        OWI_KickWdt();

        /* long read만 주기적 재동기화 */
        if (is_long_read && i == (length - 50)) {
            OWI_KickWdt();
	    if (!OWI_WaitForFallingEdge((uint16_t)i, 0u)) {
	        owi_result_latch_timeout(r);
	        break;
	    }
	}

        r->data[i] = OWI_ReadByte_StreamSynced((uint16_t)i);

        if (OWI_HasTimeout()) {
            owi_result_latch_timeout(r);
            break;
        }
        r->read_len++;
    }

    OWI_StopRead();
    OWI_KickWdt();
}