/***
    Copyright (C) 2022 J Reece Wilson (a/k/a "Reece"). All rights reserved.

    File: auBitsUtils.hpp
    Date: 2022-2-1
    Author: Reece
***/
#pragma once

template <class T>
static auline bool AuTestBit(T value, AuUInt8 idx)
{
    return value & (T(1) << T(idx));
}

template <class T>
static auline bool AuBitTest(T value, AuUInt8 idx)
{
    return value & (T(1) << T(idx));
}

template <class T>
static auline void AuSetBit(T &value, AuUInt8 idx)
{
    value |= T(1) << T(idx);
}

template <class T>
static auline void AuBitSet(T &value, AuUInt8 idx)
{
    value |= T(1) << T(idx);
}

template <class T>
static auline void  AuClearBit(T &value, AuUInt8 idx)
{
    value &= ~(T(1) << T(idx));
}

template <class T>
static auline void  AuBitClear(T &value, AuUInt8 idx)
{
    value &= ~(T(1) << T(idx));
}

template <class T>
static auline bool AuBitScanForward(AuUInt8 &index, T value)
{
    unsigned long ret;
    bool success;

    success = false;
    index = 0;

#if defined(AURORA_COMPILER_MSVC)
    if constexpr (sizeof(T) == sizeof(AuUInt64))
    {
    #if defined(AURORA_IS_32BIT)
        if (!(success = _BitScanForward(&ret, static_cast<AuUInt32>(value & 0xffffffff))))
        {
            if (!_BitScanForward(&ret, static_cast<AuUInt32>((value >> 32) & 0xffffffff)))
            {
                return false;
            }
            ret += 32;
        }
    #else
        success = _BitScanForward64(&ret, static_cast<AuUInt64>(value));
    #endif
    }
    else
    {
        success = _BitScanForward(&ret, static_cast<unsigned long>(value));
    }
#elif defined(AURORA_COMPILER_GCC) ||  defined(AURORA_COMPILER_CLANG)
    if (value == 0)
    {
        return false;
    }

    if constexpr (sizeof(T) == sizeof(AuUInt64))
    {
    #if defined(AURORA_IS_32BIT)
        auto lower = static_cast<AuUInt32>(value & 0xffffffff));
        if (lower == 0)
        {
            ret = __builtin_ctzl(static_cast<AuUInt32>((value >> 32) & 0xffffffff));
            ret += 32;
        }
        else
        {
            ret = __builtin_ctzl(static_cast<AuUInt32>(lower));
        }
    #else
        ret = __builtin_ctzll(static_cast<AuUInt64>(value));
    #endif
    }
    else if constexpr (sizeof(T) == sizeof(unsigned long))
    {
        ret = __builtin_ctzl(static_cast<unsigned long>(value));
    }
    else if constexpr (sizeof(T) == sizeof(unsigned int))
    {
        ret = __builtin_ctz(static_cast<unsigned int>(value));
    }

    success = true;
#endif
    index = ret;
    return success;
}

template <class T>
static auline bool AuBitScanReverse(AuUInt8 &index, T value)
{
    unsigned long ret;
    bool success;

    success = false;
    index = 0;

#if defined(AURORA_COMPILER_MSVC)
    if constexpr (sizeof(T) == sizeof(AuUInt64))
    {
    #if defined(AURORA_IS_32BIT)
        if (!(success = _BitScanReverse(&ret, static_cast<AuUInt32>((value >> 32) & 0xffffffff))))
        {
            if (!_BitScanReverse(&ret, static_cast<AuUInt32>(value & 0xffffffff)))
            {
                return false;
            }
            ret += 32;
        }
    #else
        success = _BitScanReverse64(&ret, static_cast<AuUInt64>(value));
    #endif
    }
    else
    {
        success = _BitScanReverse(&ret, static_cast<unsigned long>(value));
    }
#elif defined(AURORA_COMPILER_GCC) ||  defined(AURORA_COMPILER_CLANG)
    if (value == 0)
    {
        return false;
    }

    if constexpr (sizeof(T) == sizeof(AuUInt64))
    {
    #if defined(AURORA_IS_32BIT)
        auto lower = static_cast<AuUInt32>((value >> 32) & 0xffffffff);
        if (lower == 0)
        {
            ret = __builtin_clzl(static_cast<AuUInt32>(value & 0xffffffff)));
            ret = 63 - ret;
        }
        else
        {
            ret = __builtin_clzl(static_cast<AuUInt32>(lower));
            ret = 31 - ret;
        }
    #else
        ret = __builtin_clzll(static_cast<AuUInt64>(value));
        ret = 63 - ret;
    #endif
    }
    else if constexpr (sizeof(T) == sizeof(unsigned long))
    {
        ret = __builtin_clzl(static_cast<unsigned long>(value));
        if constexpr (sizeof(unsigned long) == 4)
        {
            ret = 31 - ret;
        }
        else if constexpr (sizeof(unsigned long) == 8)
        {
            ret = 63 - ret;
        }
    }
    else// if constexpr (sizeof(T) == sizeof(unsigned int))
    {
        ret = __builtin_clz(static_cast<unsigned int>(value));
        if constexpr (sizeof(unsigned int) == 4)
        {
            ret = 31 - ret;
        }
        else if constexpr (sizeof(unsigned int) == 8)
        {
            ret = 63 - ret;
        }
    }

    success = true;
#endif
    index = ret;
    return success;
}

template <class T>
struct AuHalfWord
{
    using ReturnType_t = AuConditional_t<AuIsSame_v<T, AuUInt64>, AuUInt32, AuConditional_t<AuIsSame_v<T, AuUInt32>, AuUInt32, AuConditional_t<AuIsSame_v<T, AuUInt16>, AuUInt8, AuFalseType>>>;

    static ReturnType_t ToLower(T in)
    {
        if constexpr (AuIsSame_v<T, AuUInt64>)
        {
            return in & AuUInt64(0xFFFFFFFF);
        }
        else if constexpr (AuIsSame_v<T, AuUInt32>)
        {
            return in & 0xFFFF;
        }
        else if constexpr (AuIsSame_v<T, AuUInt16>)
        {
            return in & 0xFF;
        }
        else
        {
            return {};
        }
    }

    static ReturnType_t ToHigher(T in)
    {
        if constexpr (AuIsSame_v<T, AuUInt64>)
        {
            return (in >> AuUInt64(32)) & AuUInt64(0xFFFFFFFF);
        }
        else if constexpr (AuIsSame_v<T, AuUInt32>)
        {
            return (in >> 16) & 0xFFFF;
        }
        else if constexpr (AuIsSame_v<T, AuUInt16>)
        {
            return (in >> 8) & 0xFF;
        }
        else
        {
            return {};
        }
    }
};

template <class T>
static auto AuBitsToLower(T in)
{
    return AuHalfWord<T>::ToLower(in);
}

template <class T>
static auto AuBitsToHigher(T in)
{
    return AuHalfWord<T>::ToHigher(in);
}

template <class T>
static AuUInt8 AuPopCnt(T in)
{
#if defined(AURORA_COMPILER_MSVC)
    #if defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86)
        #if defined(AURORA_ARCH_X64)
            if constexpr (sizeof(T) == sizeof(AuUInt64))
            {
                return _mm_popcnt_u64(static_cast<AuUInt64>(in));
            }
            else
        #endif
        if constexpr (sizeof(T) == sizeof(unsigned int))
        {
            return __popcnt(static_cast<unsigned int>(in));
        }
        else if constexpr (sizeof(T) <= sizeof(AuUInt16))
        {
            return __popcnt16(static_cast<AuUInt16>(in));
        }
    #endif
#else
    if constexpr (sizeof(T) == sizeof(unsigned long long))
    {
        return __builtin_popcountll(static_cast<unsigned long long>(in));
    }
    else if constexpr (sizeof(T) == sizeof(unsigned long))
    {
        return __builtin_popcountl(static_cast<unsigned long>(in));
    }
    else if constexpr (sizeof(T) == sizeof(unsigned int))
    {
        return __builtin_popcount(static_cast<unsigned int>(in));
    }
#endif

#if defined(AU_CPU_ENDIAN_LITTLE)
    if constexpr (sizeof(T) == sizeof(AuUInt64))
    {
        const AuUInt64 m1 = 0x5555555555555555ll;
        const AuUInt64 m2 = 0x3333333333333333ll;
        const AuUInt64 m4 = 0x0F0F0F0F0F0F0F0Fll;
        const AuUInt64 h01 = 0x0101010101010101ll;

        in -= (in >> 1) & m1;
        in = (in & m2) + ((in >> 2) & m2);
        in = (in + (in >> 4)) & m4;

        return (in * h01) >> 56;
    }
    else if constexpr (sizeof(T) == sizeof(AuUInt32))
    {
        const AuUInt32 m1  = 0x55555555l;
        const AuUInt32 m2  = 0x33333333l;
        const AuUInt32 m4  = 0x0F0F0F0Fl;
        const AuUInt32 h01 = 0x01010101l;

        in -= (in >> 1) & m1;
        in = (in & m2) + ((in >> 2) & m2);
        in = (in + (in >> 4)) & m4;

        return (in * h01) >> 24;
    }
#endif

    if constexpr ((sizeof(T) == sizeof(AuUInt16)) ||
                  (sizeof(T) == sizeof(AuUInt8)))
    {
        return AuPopCnt<AuUInt32>(AuUInt32(in));
    }
    
    return {};
}