#include <c_FastByteRingBuffer.h>
#include <c_Memory.h>
#include <stdlib.h>
#include <ctype.h>
#include <errno.h>

/* ------------------------------------------------------------------------------------------------------------------ */
/*  */

C_STATIC_FORCE_INLINE
c_size_t round_up_to_pow2(c_size_t v) {
    v--;
    v |= v >> 1;
    v |= v >> 2;
    v |= v >> 4;
    v |= v >> 8;
    v |= v >> 16;
#if (defined(__WORDSIZE) && __WORDSIZE == 64) || defined(_WIN64) || defined(__x86_64__) || defined(__aarch64__)
    v |= v >> 32;
#endif
    v++;
    return v;
}

/* ------------------------------------------------------------------------------------------------------------------ */
/*  */

c_err_t c_FastByteRingBuffer_Init(c_FastByteRingBuffer_t* self, c_size_t capacity) {
    if (!self || capacity == 0) return C_ERR_PARAM;

    // 自動向上對齊，確保符合 Power of Two
    self->capacity = round_up_to_pow2(capacity);
    self->mask = self->capacity - 1; // 建立遮罩
    self->head = 0;
    self->tail = 0;
    self->is_full = C_FALSE;

    self->buffer = (uint8_t*)malloc(self->capacity);
    if (!self->buffer) {
        self->capacity = 0;
        self->mask = 0;
        return C_ERR_NOMEM;
    }

    return C_ERR_SUCCESS;
}

void c_FastByteRingBuffer_Destroy(c_FastByteRingBuffer_t* self) {
    if (!self) return;

    C_FREE(self->buffer);
    self->capacity = 0;
    self->mask = 0;
    self->head = 0;
    self->tail = 0;
    self->is_full = C_FALSE;
}

// 寫入單一單元組 (極速 O(1))
c_err_t c_FastByteRingBuffer_WriteByte(c_FastByteRingBuffer_t* self, uint8_t byte) {
    if (!self || !self->buffer) return C_ERR_PARAM;
    if (self->is_full) return C_ERR_FULL;

    self->buffer[self->tail] = byte;

    // 使用高速位元與運算取代模除 %
    self->tail = (self->tail + 1) & self->mask;

    if (self->tail == self->head) {
        self->is_full = C_TRUE;
    }

    return C_ERR_SUCCESS;
}

// 讀取單一單元組 (極速 O(1))
c_err_t c_FastByteRingBuffer_ReadByte(c_FastByteRingBuffer_t* self, uint8_t* out_byte) {
    if (!self || !self->buffer || !out_byte) return C_ERR_PARAM;
    if (c_FastByteRingBuffer_IsEmpty(self)) return C_ERR_EMPTY;

    *out_byte = self->buffer[self->head];

    // 使用高速位元與運算取代模除 %
    self->head = (self->head + 1) & self->mask;
    self->is_full = C_FALSE;

    return C_ERR_SUCCESS;
}


c_size_t c_FastByteRingBuffer_WriteBuffer(c_FastByteRingBuffer_t* self, const uint8_t* src, c_size_t len) {
    if (!self || !self->buffer || !src || len == 0) return 0;
    if (self->is_full) return 0;

    // Calculate free space directly using fast size validation tracking
    c_size_t current_size = c_FastByteRingBuffer_GetSize(self);
    c_size_t free_space = self->capacity - current_size;

    // Clamp write operations to avoid overrunning existing readable elements
    if (len > free_space) {
        len = free_space;
    }

    if (len == 0) return 0;

    // Segment 1: Write from tail index up to the physical end boundary of the backing array
    c_size_t space_to_end = self->capacity - self->tail;
    c_size_t first_chunk = (len < space_to_end) ? len : space_to_end;
    memcpy(self->buffer + self->tail, src, first_chunk);

    // Segment 2: Wrap around to index 0 using bitwise optimizations if a split block configuration is required
    c_size_t second_chunk = len - first_chunk;
    if (second_chunk > 0) {
        memcpy(self->buffer, src + first_chunk, second_chunk);
        self->tail = second_chunk; // The wrapped tail calculation reduces cleanly to second_chunk
    } else {
        // Fast tail stepping path utilizing the mask constant
        self->tail = (self->tail + first_chunk) & self->mask;
    }

    // Set the full status flag if the cursors perfectly intersect
    if (self->tail == self->head) {
        self->is_full = C_TRUE;
    }

    return len;
}


// 區塊讀取
c_size_t c_FastByteRingBuffer_ReadBuffer(c_FastByteRingBuffer_t* self, uint8_t* dest, c_size_t len) {
    if (!self || !self->buffer || !dest || len == 0) return 0;

    c_size_t bytes_read = 0;
    while (bytes_read < len && !c_FastByteRingBuffer_IsEmpty(self)) {
        dest[bytes_read] = self->buffer[self->head];
        self->head = (self->head + 1) & self->mask;
        self->is_full = C_FALSE;
        bytes_read++;
    }
    return bytes_read;
}

c_size_t c_FastByteRingBuffer_GetSize(const c_FastByteRingBuffer_t* self) {
    if (!self || !self->buffer) return 0;
    if (self->is_full) return self->capacity;

    if (self->tail >= self->head) {
        return self->tail - self->head;
    } else {
        return self->capacity + self->tail - self->head;
    }
}

c_bool_t c_FastByteRingBuffer_IsEmpty(const c_FastByteRingBuffer_t* self) {
    if (!self) return C_TRUE;
    return (self->head == self->tail) && !self->is_full;
}

c_bool_t c_FastByteRingBuffer_IsFull(const c_FastByteRingBuffer_t* self) {
    if (!self) return C_FALSE;
    return self->is_full;
}


/* ------------------------------------------------------------------------------------------------------------------ */
/*  */



C_STATIC_FORCE_INLINE
uint8_t c_FastByteRingBuffer_GetAtRelativeInternal(const c_FastByteRingBuffer_t* self, c_size_t relative_offset) {
    c_size_t absolute_index = (self->head + relative_offset) & self->mask;
    return self->buffer[absolute_index];
}

/* ------------------------------------------------------------------------------------------------------------------ */

void c_FastByteRingBuffer_WriteByteOverwrite(c_FastByteRingBuffer_t* self, uint8_t byte) {
    if (!self || !self->buffer) return;

    if (self->is_full) {
        // 環形陣列已滿時，強制將讀取指標向前推一格，拋棄最舊數據
        self->head = (self->head + 1) & self->mask;
    }

    self->buffer[self->tail] = byte;
    self->tail = (self->tail + 1) & self->mask;

    if (self->tail == self->head) {
        self->is_full = C_TRUE;
    }
}

c_size_t c_FastByteRingBuffer_WriteBufferOverwrite(c_FastByteRingBuffer_t* self, const uint8_t* src, c_size_t len) {
    if (!self || !self->buffer || !src || len == 0) return 0;

    // 邊界極端優化：如果寫入長度超過總容量，只有最後符合容量大小的數據能存活
    if (len >= self->capacity) {
        src += (len - self->capacity);
        len = self->capacity;

        memcpy(self->buffer, src, len);
        self->head = 0;
        self->tail = 0;
        self->is_full = C_TRUE;
        return len;
    }

    c_size_t size = c_FastByteRingBuffer_GetSize(self);
    c_size_t free_space = self->capacity - size;

    // 若寫入長度大於剩餘空間，計算溢出量並自動同步前推 head 指標
    if (len > free_space) {
        c_size_t overwrite_count = len - free_space;
        self->head = (self->head + overwrite_count) & self->mask;
    }

    // 分段一：從 tail 寫入到物理內存陣列末尾
    c_size_t space_to_end = self->capacity - self->tail;
    c_size_t first_chunk = (len < space_to_end) ? len : space_to_end;
    memcpy(self->buffer + self->tail, src, first_chunk);

    // 分段二：折返到內存陣列開頭寫入剩餘數據
    c_size_t second_chunk = len - first_chunk;
    if (second_chunk > 0) {
        memcpy(self->buffer, src + first_chunk, second_chunk);
        self->tail = second_chunk;
    } else {
        self->tail = (self->tail + first_chunk) & self->mask;
    }

    if (self->tail == self->head) {
        self->is_full = C_TRUE;
    } else {
        self->is_full = C_FALSE;
    }

    return len;
}

c_err_t c_FastByteRingBuffer_PeekByte(const c_FastByteRingBuffer_t* self, uint8_t* out_byte) {
    if (!self || !self->buffer || !out_byte) return C_ERR_INVALID_PARAM;
    if (c_FastByteRingBuffer_IsEmpty(self)) return C_ERR_OUT_OF_BOUNDS;

    *out_byte = self->buffer[self->head];
    return C_SUCCESS;
}

c_size_t c_FastByteRingBuffer_PeekBuffer(const c_FastByteRingBuffer_t* self, uint8_t* dest, c_size_t len) {
    if (!self || !self->buffer || !dest || len == 0) return 0;

    c_size_t available_bytes = c_FastByteRingBuffer_GetSize(self);
    if (len > available_bytes) {
        len = available_bytes;
    }

    if (len == 0) return 0;

    // 複製標準讀取邏輯，但完全不變動真實核心 head 指標狀態
    c_size_t bytes_to_end = self->capacity - self->head;
    c_size_t first_chunk = (len < bytes_to_end) ? len : bytes_to_end;
    memcpy(dest, self->buffer + self->head, first_chunk);

    c_size_t second_chunk = len - first_chunk;
    if (second_chunk > 0) {
        memcpy(dest + first_chunk, self->buffer, second_chunk);
    }

    return len;
}

c_size_t c_FastByteRingBuffer_Discard(c_FastByteRingBuffer_t* self, c_size_t len) {
    if (!self || !self->buffer || len == 0) return 0;

    c_size_t available_bytes = c_FastByteRingBuffer_GetSize(self);
    if (len > available_bytes) {
        len = available_bytes;
    }

    if (len == 0) return 0;

    self->head = (self->head + len) & self->mask;
    self->is_full = C_FALSE;

    return len;
}

const uint8_t* c_FastByteRingBuffer_GetReadPtr(const c_FastByteRingBuffer_t* self, c_size_t* out_contiguous_len) {
    if (!self || !self->buffer || !out_contiguous_len) return NULL;

    *out_contiguous_len = 0;
    if (c_FastByteRingBuffer_IsEmpty(self)) return NULL;

    if (self->tail > self->head) {
        *out_contiguous_len = self->tail - self->head;
    } else {
        *out_contiguous_len = self->capacity - self->head;
    }

    return self->buffer + self->head;
}

uint8_t* c_FastByteRingBuffer_GetWritePtr(const c_FastByteRingBuffer_t* self, c_size_t* out_contiguous_len) {
    if (!self || !self->buffer || !out_contiguous_len) return NULL;

    *out_contiguous_len = 0;
    if (self->is_full) return NULL;

    if (self->tail >= self->head) {
        *out_contiguous_len = self->capacity - self->tail;
    } else {
        *out_contiguous_len = self->head - self->tail;
    }

    return self->buffer + self->tail;
}

c_index_t c_FastByteRingBuffer_IndexOfByte(const c_FastByteRingBuffer_t* self, uint8_t target) {
    if (!self || !self->buffer) return C_ERR_NOT_FOUND;

    c_size_t total_size = c_FastByteRingBuffer_GetSize(self);
    for (c_size_t offset = 0; offset < total_size; offset++) {
        if (c_FastByteRingBuffer_GetAtRelativeInternal(self, offset) == target) {
            return (c_index_t)offset;
        }
    }
    return C_ERR_NOT_FOUND;
}

c_index_t c_FastByteRingBuffer_IndexOfBuffer(const c_FastByteRingBuffer_t* self, const uint8_t* pattern, c_size_t pattern_len) {
    if (!self || !self->buffer || !pattern || pattern_len == 0) return C_ERR_NOT_FOUND;

    c_size_t total_size = c_FastByteRingBuffer_GetSize(self);
    if (pattern_len > total_size) return C_ERR_NOT_FOUND;

    c_size_t max_search_offset = total_size - pattern_len;

    for (c_size_t offset = 0; offset <= max_search_offset; offset++) {
        c_bool_t match_found = C_TRUE;

        for (c_size_t p_idx = 0; p_idx < pattern_len; p_idx++) {
            if (c_FastByteRingBuffer_GetAtRelativeInternal(self, offset + p_idx) != pattern[p_idx]) {
                match_found = C_FALSE;
                break;
            }
        }

        if (match_found) {
            return (c_index_t)offset;
        }
    }
    return C_ERR_NOT_FOUND;
}

c_index_t c_FastByteRingBuffer_LastIndexOfBuffer(const c_FastByteRingBuffer_t* self, const uint8_t* pattern, c_size_t pattern_len) {
    if (!self || !self->buffer || !pattern || pattern_len == 0) return C_ERR_NOT_FOUND;

    c_size_t total_size = c_FastByteRingBuffer_GetSize(self);
    if (pattern_len > total_size) return C_ERR_NOT_FOUND;

    c_size_t max_search_offset = total_size - pattern_len;

    for (c_size_t offset = max_search_offset; ; offset--) {
        c_bool_t match_found = C_TRUE;

        for (c_size_t p_idx = 0; p_idx < pattern_len; p_idx++) {
            if (c_FastByteRingBuffer_GetAtRelativeInternal(self, offset + p_idx) != pattern[p_idx]) {
                match_found = C_FALSE;
                break;
            }
        }

        if (match_found) {
            return (c_index_t)offset;
        }

        if (offset == 0) break;
    }
    return C_ERR_NOT_FOUND;
}

c_size_t c_FastByteRingBuffer_ReadUntilToken(c_FastByteRingBuffer_t* self, const uint8_t* token, c_size_t token_len, uint8_t* dest, c_size_t dest_max_len) {
    if (!self || !self->buffer || !token || token_len == 0 || !dest || dest_max_len == 0) return 0;

    c_index_t match_offset = c_FastByteRingBuffer_IndexOfBuffer(self, token, token_len);
    if (match_offset == C_ERR_NOT_FOUND) {
        return 0;
    }

    c_size_t aggregate_bytes = (c_size_t)match_offset + token_len;
    if (aggregate_bytes > dest_max_len) {
        return 0; // 避免目標緩衝區溢出
    }

    return c_FastByteRingBuffer_ReadBuffer(self, dest, aggregate_bytes);
}

c_err_t c_FastByteRingBuffer_GetAtRelative(const c_FastByteRingBuffer_t* self, c_size_t relative_offset, uint8_t* out_byte) {
    if (!self || !self->buffer || !out_byte) return C_ERR_INVALID_PARAM;

    c_size_t active_size = c_FastByteRingBuffer_GetSize(self);
    if (relative_offset >= active_size) {
        return C_ERR_OUT_OF_BOUNDS;
    }

    *out_byte = c_FastByteRingBuffer_GetAtRelativeInternal(self, relative_offset);
    return C_SUCCESS;
}

c_bool_t c_FastByteRingBuffer_Is(const c_FastByteRingBuffer_t* self, c_index_t offset, uint8_t value) {
    if (!self || !self->buffer || offset < 0) return C_FALSE;

    c_size_t active_size = c_FastByteRingBuffer_GetSize(self);
    if ((c_size_t)offset >= active_size) return C_FALSE;

    return (c_FastByteRingBuffer_GetAtRelativeInternal(self, (c_size_t)offset) == value) ? C_TRUE : C_FALSE;
}

int c_FastByteRingBuffer_Memcmp(const c_FastByteRingBuffer_t* self, c_size_t offset, const uint8_t* buffer, c_size_t len) {
    if (!self || !self->buffer || !buffer) return C_ERR_INVALID_PARAM;
    if (len == 0) return 0;

    c_size_t active_size = c_FastByteRingBuffer_GetSize(self);
    if (offset >= active_size || (offset + len) > active_size) {
        return C_ERR_OUT_OF_BOUNDS;
    }

    c_size_t absolute_start = (self->head + offset) & self->mask;
    c_size_t bytes_to_end = self->capacity - absolute_start;

    if (len <= bytes_to_end) {
        return memcmp(self->buffer + absolute_start, buffer, len);
    } else {
        int first_segment_match = memcmp(self->buffer + absolute_start, buffer, bytes_to_end);
        if (first_segment_match != 0) {
            return first_segment_match;
        }
        return memcmp(self->buffer, buffer + bytes_to_end, len - bytes_to_end);
    }
}



c_err_t c_FastByteRingBuffer_Strtoul(const c_FastByteRingBuffer_t* self, c_size_t offset, int base, unsigned long* out_value, c_size_t* out_end_offset) {
    if (!self || !self->buffer || !out_value) return C_ERR_INVALID_PARAM;

    c_size_t total_size = c_FastByteRingBuffer_GetSize(self);
    if (offset >= total_size) return C_ERR_OUT_OF_BOUNDS;

    // 1. Skip leading spaces safely using fast inline relative scans
    c_size_t scan_idx = offset;
    while (scan_idx < total_size) {
        uint8_t byte = c_FastByteRingBuffer_GetAtRelativeInternal(self, scan_idx);
        if (!isspace(byte)) break;
        scan_idx++;
    }

    if (scan_idx == total_size) return C_ERR_INVALID_PARAM; // Contains only whitespace

    // 2. Locate trailing delimiters to calculate numeric chunk span width
    c_size_t start_numeric_offset = scan_idx;
    c_size_t numeric_len = 0;
    while (scan_idx < total_size) {
        uint8_t byte = c_FastByteRingBuffer_GetAtRelativeInternal(self, scan_idx);
        // Include numerical modifiers (+, -), hex identifiers (x, X) and alphanumeric bases
        if (!isalnum(byte) && byte != '+' && byte != '-') {
            break;
        }
        numeric_len++;
        scan_idx++;
    }

    if (numeric_len == 0) return C_ERR_INVALID_PARAM;

    // 3. Resolve the starting index using bitwise mask instead of slow % operators
    c_size_t absolute_start = (self->head + start_numeric_offset) & self->mask;
    c_size_t bytes_to_end = self->capacity - absolute_start;

    unsigned long result = 0;
    char* parse_end = NULL;
    int current_errno = errno;
    errno = 0;

    // Fast-path: Segment runs contiguous without wrapping lines
    if (numeric_len <= bytes_to_end) {
        const char* flat_ptr = (const char*)(self->buffer + absolute_start);
        result = strtoul(flat_ptr, &parse_end, base);

        c_size_t parsed_bytes = (c_size_t)(parse_end - flat_ptr);
        if (parsed_bytes == 0 || parse_end == flat_ptr) {
            errno = current_errno;
            return C_ERR_INVALID_PARAM;
        }

        if (errno == ERANGE) return C_ERR_OUT_OF_BOUNDS;

        *out_value = result;
        if (out_end_offset) {
            *out_end_offset = start_numeric_offset + parsed_bytes;
        }
    } else {
        // Slow-path: Structural wrap configuration requires localized stack flattening
        if (numeric_len >= 64) return C_ERR_OUT_OF_BOUNDS;

        char stack_scratch[64];
        for (c_size_t i = 0; i < numeric_len; i++) {
            stack_scratch[i] = (char)c_FastByteRingBuffer_GetAtRelativeInternal(self, start_numeric_offset + i);
        }
        stack_scratch[numeric_len] = '\0'; // Guarantee absolute zero string termination

        result = strtoul(stack_scratch, &parse_end, base);
        c_size_t parsed_bytes = (c_size_t)(parse_end - stack_scratch);

        if (parsed_bytes == 0 || parse_end == stack_scratch) {
            errno = current_errno;
            return C_ERR_INVALID_PARAM;
        }

        if (errno == ERANGE) return C_ERR_OUT_OF_BOUNDS;

        *out_value = result;
        if (out_end_offset) {
            *out_end_offset = start_numeric_offset + parsed_bytes;
        }
    }

    errno = current_errno; // Preserve runtime environment variables smoothly
    return C_SUCCESS;
}