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

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

#if defined(AURORA_COMPILER_CLANG)
    #include <stdatomic.h> 
#endif

#if defined(AURORA_COMPILER_MSVC) || 1
    using AuAInt8 = volatile AuInt8;
    using AuAUInt8 = volatile AuUInt8;

    using AuAInt16 = volatile AuInt16;
    using AuAUInt16 = volatile AuUInt16;

    using AuAInt32 = volatile AuInt32;
    using AuAUInt32 = volatile AuUInt32;

    #if defined(AURORA_IS_64BIT)
        using AuAInt64 = volatile AuInt64;
        using AuAUInt64 = volatile AuUInt64;
    #endif

    #define AU_ATOMIC_INIT(n) n
#else

    using AuAInt8 = _Atomic(int8_t);
    using AuAUInt8 = _Atomic(uint8_t);

    using AuAInt16 = _Atomic(int16_t);
    using AuAUInt16 = _Atomic(uint16_t);


    using AuAInt32 = _Atomic(int32_t);
    using AuAUInt32 = _Atomic(uint32_t);

    #if defined(AURORA_IS_64BIT)
        using AuAInt64 = _Atomic(int64_t);
        using AuAUInt64 = _Atomic(uint64_t);
    #endif

    #define AU_ATOMIC_INIT ATOMIC_VAR_INIT
#endif


// Defines:
//     AuAtomicCompareExchange (no weak variants yet)
//     -
//     AuAtomicTestAndSet
//     AuAtomicClearU8Lock
//     AuAtomicSet
//     AuAtomicUnset (ret: bool)
//     AuAtomicAndUnsetBit (ret: T)
//     -
//     AuAtomicLoadWeak
//     AuAtomicStoreWeak
//     -
//     AuAtomicLoad
//     AuAtomicStore
//     -
//     AuAtomicOrSetBit
//     AuAtomicOr
//     AuAtomicAnd
//     AuAtomicAdd (warn: returns post op atomic value)
//     AuAtomicSub (warn: returns post op atomic value)

template <class T>
struct AuAtomicUtils
{
    /**
     * @brief Generic bitwise (1 << offset)
     * @return original value 
     * @warning T is bound by platform and compiler constraints
    */
    static T    Set(T *in, AuUInt8 offset);

    /**
     * @brief
     * @param in
     * @param orValue
     * @return original value
    */
    static T    Or(T *in, T orValue);

    /**
     * @brief
     * @param in
     * @param andValue
     * @return original value
    */
    static T    And(T *in, T andValue);

    /**
     * @brief Adds addend to in
     * @return updated value 
     * @warning T is bound by platform and compiler constraints
    */
    static T    Add(T *in, T addend);

    /**
     * @brief Subtracts the minuend from in
     * @return updated value
     * @warning T is bound by platform and compiler constraints
    */
    static T    Sub(T *in, T minuend);

    /**
     * @brief Generic compare exchange
     * @param replace replacement value for in if in matches compare
     * @param compare required reference value
     * @return original value 
     * @warning T is bound by platform and compiler constraints
    */
    static T    CompareExchange(T *in, T replace, T compare);

    /**
     * @brief { return *in & (1 << offset); in |= (1 << offset) }
     * @param in 
     * @param offset Bit index
     * @return *in & (1 << offset)
     * @warning T is bound by platform and compiler constraints
    */
    static bool TestAndSet(T *in, const AuUInt8 offset);

    // 
    static T Load(T *in);

    //
    static void Store(T *in, T value);

    //
    static void ClearU8Lock(T *in);

    //
    static T LoadWeak(T *in);

    static void StoreWeak(T *in, T value);
};

#if defined(AURORA_COMPILER_MSVC)

template <>
inline auline AuUInt64 AuAtomicUtils<AuUInt64>::CompareExchange(AuUInt64 *in, AuUInt64 replace, AuUInt64 compare)
{
    return static_cast<AuUInt64>(_InterlockedCompareExchange64(reinterpret_cast<long long volatile *>(in), static_cast<long long>(replace), static_cast<long long>(compare)));
}

template <>
inline auline AuUInt32 AuAtomicUtils<AuUInt32>::CompareExchange(AuUInt32 *in, AuUInt32 replace, AuUInt32 compare)
{
    return static_cast<AuUInt32>(_InterlockedCompareExchange(reinterpret_cast<long volatile *>(in), static_cast<long>(replace), static_cast<long>(compare)));
}

template <>
inline auline AuUInt16 AuAtomicUtils<AuUInt16>::CompareExchange(AuUInt16 *in, AuUInt16 replace, AuUInt16 compare)
{
    return static_cast<AuUInt16>(_InterlockedCompareExchange16(reinterpret_cast<short volatile *>(in), static_cast<short>(replace), static_cast<short>(compare)));
}

template <>
inline auline AuUInt8 AuAtomicUtils<AuUInt8>::CompareExchange(AuUInt8 *in, AuUInt8 replace, AuUInt8 compare)
{
    return static_cast<AuUInt8>(_InterlockedCompareExchange8(reinterpret_cast<char volatile *>(in), static_cast<char>(replace), static_cast<char>(compare)));
}

template <>
inline auline AuInt64 AuAtomicUtils<AuInt64>::CompareExchange(AuInt64 *in, AuInt64 replace, AuInt64 compare)
{
    return _InterlockedCompareExchange64(reinterpret_cast<long long volatile *>(in), static_cast<long long>(replace), static_cast<long long>(compare));
}

template <>
inline auline AuInt32 AuAtomicUtils<AuInt32>::CompareExchange(AuInt32 *in, AuInt32 replace, AuInt32 compare)
{
    return _InterlockedCompareExchange(reinterpret_cast<long volatile *>(in), static_cast<long>(replace), static_cast<long>(compare));
}

template <>
inline auline AuInt16 AuAtomicUtils<AuInt16>::CompareExchange(AuInt16 *in, AuInt16 replace, AuInt16 compare)
{
    return _InterlockedCompareExchange16(reinterpret_cast<short volatile *>(in), static_cast<short>(replace), static_cast<short>(compare));
}

template <>
inline auline AuInt8 AuAtomicUtils<AuInt8>::CompareExchange(AuInt8 *in, AuInt8 replace, AuInt8 compare)
{
    return _InterlockedCompareExchange8(reinterpret_cast<char volatile *>(in), static_cast<char>(replace), static_cast<char>(compare));
}

#if !defined(AURORA_IS_32BIT)
template <>
inline auline AuUInt64 AuAtomicUtils<AuUInt64>::Add(AuUInt64 *in, AuUInt64 addend)
{
    return static_cast<AuUInt64>(_InterlockedExchangeAdd64(reinterpret_cast<long long volatile *>(in), static_cast<long long>(addend)) + static_cast<long long>(addend));
}
#endif

template <>
inline auline AuUInt32 AuAtomicUtils<AuUInt32>::Add(AuUInt32 *in, AuUInt32 addend)
{
    return static_cast<AuUInt32>(_InterlockedExchangeAdd(reinterpret_cast<long volatile *>(in), static_cast<long>(addend)) + static_cast<long>(addend));
}

#if !defined(AURORA_IS_32BIT)
template <>
inline auline AuInt64 AuAtomicUtils<AuInt64>::Add(AuInt64 *in, AuInt64 addend)
{
    return _InterlockedExchangeAdd64(reinterpret_cast<long long volatile *>(in), static_cast<long long>(addend)) + static_cast<long long>(addend);
}
#endif

template <>
inline auline AuInt32 AuAtomicUtils<AuInt32>::Add(AuInt32 *in, AuInt32 addend)
{
    return _InterlockedExchangeAdd(reinterpret_cast<long volatile *>(in), static_cast<long>(addend)) + static_cast<long>(addend);
}

// TODO:
#if 0
template <>
inline auline AuUInt16 AuAtomicUtils<AuUInt16>::Add(AuUInt16 *in, AuUInt16 addend)
{
    return {};
}
#endif


template <>
inline auline AuUInt64 AuAtomicUtils<AuUInt64>::Sub(AuUInt64 *in, AuUInt64 minuend)
{
    return Add(in, AuUInt64(0) - minuend);
}

template <>
inline auline AuUInt32 AuAtomicUtils<AuUInt32>::Sub(AuUInt32 *in, AuUInt32 minuend)
{
    return Add(in, AuUInt32(0) - minuend);
}

template <>
inline auline AuInt64 AuAtomicUtils<AuInt64>::Sub(AuInt64 *in, AuInt64 minuend)
{
    return Add(in, AuInt64(0) - minuend);
}

template <>
inline auline AuInt32 AuAtomicUtils<AuInt32>::Sub(AuInt32 *in, AuInt32 minuend)
{
    return Add(in, AuInt32(0) - minuend);
}

// TODO:
#if 0
template <>
inline auline AuUInt16 AuAtomicUtils<AuUInt16>::Sub(AuUInt16 *in, AuUInt16 minuend)
{
    return {};
}

#endif

#if !defined(AURORA_IS_32BIT)
template <>
inline auline AuUInt64 AuAtomicUtils<AuUInt64>::Or(AuUInt64 *in, AuUInt64 orValue)
{
    return _InterlockedOr64(reinterpret_cast<long long volatile *>(in), orValue);
}
#endif

template <>
inline auline AuUInt32 AuAtomicUtils<AuUInt32>::Or(AuUInt32 *in, AuUInt32 orValue)
{
    return _InterlockedOr(reinterpret_cast<long volatile *>(in), orValue);
}

template <>
inline auline AuUInt16 AuAtomicUtils<AuUInt16>::Or(AuUInt16 *in, AuUInt16 orValue)
{
    return _InterlockedOr16(reinterpret_cast<short volatile *>(in), orValue);
}

#if !defined(AURORA_IS_32BIT)
template <>
inline auline AuInt64 AuAtomicUtils<AuInt64>::Or(AuInt64 *in, AuInt64 orValue)
{
    return _InterlockedOr64(reinterpret_cast<long long volatile *>(in), orValue);
}
#endif

template <>
inline auline AuInt32 AuAtomicUtils<AuInt32>::Or(AuInt32 *in, AuInt32 orValue)
{
    return _InterlockedOr(reinterpret_cast<long volatile *>(in), orValue);
}

template <>
inline auline long AuAtomicUtils<long>::Or(long *in, long orValue)
{
    return _InterlockedOr(reinterpret_cast<long volatile *>(in), orValue);
}

template <>
inline auline unsigned long AuAtomicUtils<unsigned long>::Or(unsigned long *in, unsigned long orValue)
{
    return _InterlockedOr(reinterpret_cast<long volatile *>(in), orValue);
}

template <>
inline auline AuInt16 AuAtomicUtils<AuInt16>::Or(AuInt16 *in, AuInt16 orValue)
{
    return _InterlockedOr16(reinterpret_cast<short volatile *>(in), orValue);
}

template <class T>
inline auline T AuAtomicUtils<T>::Set(T *in, AuUInt8 offset)
{
    return AuAtomicUtils<T>::Or(in, T(1) << offset);
}

#if !defined(AURORA_IS_32BIT)
template <>
inline auline AuUInt64 AuAtomicUtils<AuUInt64>::And(AuUInt64 *in, AuUInt64 AndValue)
{
    return _InterlockedAnd64(reinterpret_cast<long long volatile *>(in), AndValue);
}
#endif

template <>
inline auline AuUInt32 AuAtomicUtils<AuUInt32>::And(AuUInt32 *in, AuUInt32 AndValue)
{
    return _InterlockedAnd(reinterpret_cast<long volatile *>(in), AndValue);
}

template <>
inline auline AuUInt16 AuAtomicUtils<AuUInt16>::And(AuUInt16 *in, AuUInt16 AndValue)
{
    return _InterlockedAnd16(reinterpret_cast<short volatile *>(in), AndValue);
}

#if !defined(AURORA_IS_32BIT)
template <>
inline auline AuInt64 AuAtomicUtils<AuInt64>::And(AuInt64 *in, AuInt64 AndValue)
{
    return _InterlockedAnd64(reinterpret_cast<long long volatile *>(in), AndValue);
}
#endif

template <>
inline auline AuInt32 AuAtomicUtils<AuInt32>::And(AuInt32 *in, AuInt32 AndValue)
{
    return _InterlockedAnd(reinterpret_cast<long volatile *>(in), AndValue);
}

template <>
inline auline long AuAtomicUtils<long>::And(long *in, long AndValue)
{
    return _InterlockedAnd(reinterpret_cast<long volatile *>(in), AndValue);
}

template <>
inline auline unsigned long AuAtomicUtils<unsigned long>::And(unsigned long *in, unsigned long AndValue)
{
    return _InterlockedAnd(reinterpret_cast<long volatile *>(in), AndValue);
}

template <>
inline auline AuInt16 AuAtomicUtils<AuInt16>::And(AuInt16 *in, AuInt16 AndValue)
{
    return _InterlockedAnd16(reinterpret_cast<short volatile *>(in), AndValue);
}


#elif defined(AURORA_COMPILER_CLANG) || defined(AURORA_COMPILER_GCC)

template <class T>
inline auline T AuAtomicUtils<T>::CompareExchange(T *in, T replace, T compare)
{
    return __sync_val_compare_and_swap(in, compare, replace);
}

template <class T>
inline auline T AuAtomicUtils<T>::Add(T *in, T addend)
{
    return __sync_add_and_fetch(in, addend);
}

template <class T>
inline auline T AuAtomicUtils<T>::Sub(T *in, T minuend)
{
    return __sync_sub_and_fetch(in, minuend);
}

template <class T>
inline auline T AuAtomicUtils<T>::Set(T *in, AuUInt8 offset)
{
    return __sync_fetch_and_or(in, T(1) << offset);
}

template <class T>
inline auline T AuAtomicUtils<T>::Or(T *in, T value)
{
    return __sync_fetch_and_or(in, value);
}

template <class T>
inline auline T AuAtomicUtils<T>::And(T *in, T value)
{
     return __sync_fetch_and_and(in, value);
}

#endif

template <class T>
inline auline T AuAtomicUtils<T>::LoadWeak(T *in)
{
    return AuAtomicUtils<T>::Load(in);
}

template <class T>
inline void AuAtomicUtils<T>::StoreWeak(T *in, T val)
{
    *in = val;
}

#if defined(AURORA_COMPILER_CLANG)
    #define ATOMIC_PREFIX_HAX(name) __c11_ ## name
#else
    #define ATOMIC_PREFIX_HAX(name) __ ## name ## _explicit
#endif

template <class T>
inline auline T AuAtomicUtils<T>::Load(T *in)
{
    if constexpr (AuIsPointer_v<T>)
    {
        return (T)AuAtomicUtils<AuUInt>::Load((AuUInt *)in);
    }
    else
    {
#if defined(AURORA_COMPILER_MSVC)
    return *(volatile T*)in;
#else

#if defined(AURORA_COMPILER_CLANG)
    if constexpr (AuIsSame_v<AuUInt8, T>)
    {
        return ATOMIC_PREFIX_HAX(atomic_load)((_Atomic(AuUInt8) *)(in), __ATOMIC_ACQUIRE);
    }
    else if constexpr (AuIsSame_v<AuInt8, T>)
    {
        return ATOMIC_PREFIX_HAX(atomic_load)((_Atomic(AuInt8) *)(in), __ATOMIC_ACQUIRE);
    }
    else if constexpr (AuIsSame_v<AuUInt16, T>)
    {
        return ATOMIC_PREFIX_HAX(atomic_load)((_Atomic(AuUInt16) *)(in), __ATOMIC_ACQUIRE);
    }
    else if constexpr (AuIsSame_v<AuInt16, T>)
    {
        return ATOMIC_PREFIX_HAX(atomic_load)((_Atomic(AuInt16) *)(in), __ATOMIC_ACQUIRE);
    }
    else if constexpr (AuIsSame_v<AuUInt32, T>)
    {
        return ATOMIC_PREFIX_HAX(atomic_load)((_Atomic(AuUInt32) *)(in), __ATOMIC_ACQUIRE);
    }
    else if constexpr (AuIsSame_v<AuInt32, T>)
    {
        return ATOMIC_PREFIX_HAX(atomic_load)((_Atomic(AuInt32) *)(in), __ATOMIC_ACQUIRE);
    }
    else if constexpr (AuIsSame_v<AuUInt64, T>)
    {
        return ATOMIC_PREFIX_HAX(atomic_load)((_Atomic(AuUInt64) *)(in), __ATOMIC_ACQUIRE);
    }
    else if constexpr (AuIsSame_v<AuInt64, T>)
    {
        return ATOMIC_PREFIX_HAX(atomic_load)((_Atomic(AuInt64) *)(in), __ATOMIC_ACQUIRE);
    }
    else
    {
        static_assert(AuIsVoid_v<T>, "T");
    }
#else
    return __sync_val_compare_and_swap(*in, 0, 0);
#endif
#endif
    }
}

template <class T>
inline auline void AuAtomicUtils<T>::Store(T *in, T val)
{
    if constexpr (AuIsPointer_v<T>)
    {
        AuAtomicUtils<AuUInt>::Store((AuUInt *)in, (AuUInt)val);
        return;
    }
    else
    {
#if defined(AURORA_COMPILER_MSVC) && (defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86))
    *(volatile T*)in = val;
#elif defined(AURORA_COMPILER_MSVC)
#if 0
    ::MemoryBarrier();
    *(volatile T *)in = val;
#else
    // assume /volatile:iso 
    MemoryBarrier();

    if constexpr (AuIsSame_v<AuUInt8, T>)
    {
        ::InterlockedExchange8((CHAR volatile *)in, (CHAR)val);
    }
    else if constexpr (AuIsSame_v<AuInt8, T>)
    {
        ::InterlockedExchange8((CHAR volatile *)in, (CHAR)val);
    }
    else if constexpr (AuIsSame_v<AuUInt16, T>)
    {
        ::InterlockedExchange16((SHORT volatile *)in, (SHORT)val);
    }
    else if constexpr (AuIsSame_v<AuInt16, T>)
    {
        ::InterlockedExchange16((SHORT volatile *)in, (SHORT)val);
    }
    else if constexpr (AuIsSame_v<AuUInt32, T>)
    {
        ::InterlockedExchange32((LONG32 volatile *)in, (LONG32)val);
    }
    else if constexpr (AuIsSame_v<AuInt32, T>)
    {
        ::InterlockedExchange32((LONG32 volatile *)in, (LONG32)val);
    }
    else if constexpr (AuIsSame_v<AuUInt64, T>)
    {
        ::InterlockedExchange64((LONG64 volatile *)in, (LONG64)val);
    }
    else if constexpr (AuIsSame_v<AuInt64, T>)
    {
        ::InterlockedExchange64((LONG64 volatile *)in, (LONG64)val);
    }
    else
    {
        static_assert(AuIsVoid_v<T>, "T");
    }

#endif
#else
    #if !(defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86))
        __sync_synchronize();
    #endif
    if constexpr (AuIsSame_v<AuUInt8, T>)
    {
        ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuUInt8) *)(in), val, __ATOMIC_RELEASE);
    }
    else if constexpr (AuIsSame_v<AuInt8, T>)
    {
        ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuInt8) *)(in), val, __ATOMIC_RELEASE);
    }
    else if constexpr (AuIsSame_v<AuUInt16, T>)
    {
        ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuUInt16) *)(in), val, __ATOMIC_RELEASE);
    }
    else if constexpr (AuIsSame_v<AuInt16, T>)
    {
        ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuInt16) *)(in), val, __ATOMIC_RELEASE);
    }
    else if constexpr (AuIsSame_v<AuUInt32, T>)
    {
        ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuUInt32) *)(in), val, __ATOMIC_RELEASE);
    }
    else if constexpr (AuIsSame_v<AuInt32, T>)
    {
        ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuInt32) *)(in), val, __ATOMIC_RELEASE);
    }
    else if constexpr (AuIsSame_v<AuUInt64, T>)
    {
        ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuUInt64) *)(in), val, __ATOMIC_RELEASE);
    }
    else if constexpr (AuIsSame_v<AuInt64, T>)
    {
        ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuInt64) *)(in), val, __ATOMIC_RELEASE);
    }
    else
    {
        static_assert(AuIsVoid_v<T>, "T");
    }
#endif
    }
}

template <>
inline auline 
void AuAtomicUtils<AuUInt8>::ClearU8Lock(AuUInt8 *in)
{
#if defined(AURORA_COMPILER_MSVC) && (defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86))
    *(volatile AuUInt8 *)in = 0;
#elif defined(AURORA_COMPILER_MSVC)
    // assume /volatile:iso 
    MemoryBarrier();
    ::_InterlockedExchange8((volatile char *)in, 0);
    // i think this will work on aarch64 and most risc architectures
    //InterlockedAndRelease((volatile LONG *)in, ~0xFF);
#else
    // Flush non-atomic operations within L1 or other store/load queue
#if !(defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86))
    __sync_synchronize();
#endif
    
    // Clear the lock
    ATOMIC_PREFIX_HAX(atomic_store)((_Atomic(AuUInt8) *)(in), 0, __ATOMIC_RELEASE);
#endif
}

template <>
inline auline
void AuAtomicUtils<AuUInt32>::ClearU8Lock(AuUInt32 *in)
{
#if defined(AU_CPU_ENDIAN_LITTLE)
    AuAtomicUtils<AuUInt8>::ClearU8Lock((AuUInt8 *)in);
#else
    AuAtomicUtils<AuUInt8>::ClearU8Lock(((AuUInt8 *)in) + 3);
#endif
}

template <class T>
inline auline bool AuAtomicUtils<T>::TestAndSet(T *in, const AuUInt8 offset)
{
    return AuAtomicUtils<T>::Set(in, offset) & (1 << offset);
}

#if defined(AURORA_COMPILER_MSVC) && (defined(AURORA_ARCH_X64) || defined(AURORA_ARCH_X86))

template <>
inline auline bool AuAtomicUtils<unsigned long>::TestAndSet(unsigned long *in, const AuUInt8 offset)
{
    return _interlockedbittestandset(reinterpret_cast<volatile long *>(in), offset);
}

template <>
inline auline bool AuAtomicUtils<long>::TestAndSet(long *in, const AuUInt8 offset)
{
    return _interlockedbittestandset(reinterpret_cast<volatile long *>(in), offset);
}

template <>
inline auline bool AuAtomicUtils<AuUInt32>::TestAndSet(AuUInt32 *in, const AuUInt8 offset)
{
    return _interlockedbittestandset(reinterpret_cast<volatile long *>(in), offset);
}

template <>
inline auline bool AuAtomicUtils<AuInt32>::TestAndSet(AuInt32 *in, const AuUInt8 offset)
{
    return _interlockedbittestandset(reinterpret_cast<volatile long *>(in), offset);
}

#if !defined(AURORA_IS_32BIT)
template <>
inline auline bool AuAtomicUtils<AuUInt64>::TestAndSet(AuUInt64 *in, const AuUInt8 offset)
{
    return _interlockedbittestandset64(reinterpret_cast<volatile long long *>(in), offset);
}

template <>
inline auline bool AuAtomicUtils<AuInt64>::TestAndSet(AuInt64 *in, const AuUInt8 offset)
{
    return _interlockedbittestandset64(reinterpret_cast<volatile long long *>(in), offset);
}
#endif

#endif

template <class T>
auline
void     AuAtomicStore(T *in, T value)
{
    AuAtomicUtils<T>::Store(in, value);
}

template <class T>
auline
T        AuAtomicLoad(T *in)
{
    return AuAtomicUtils<T>::Load(in);
}

template <class T>
auline
T        AuAtomicLoadWeak(T *in)
{
    return AuAtomicUtils<T>::LoadWeak(in);
}

template <class T>
auline
void      AuAtomicStoreWeak(T *in, T value)
{
    AuAtomicUtils<T>::StoreWeak(in, value);
}

template <class T>
auline
void     AuAtomicClearU8Lock(T *in)
{
    AuAtomicUtils<T>::ClearU8Lock(in);
}

template <class T>
auline
T    AuAtomicOrSetBit(T *in, AuUInt8 offset)
{
    return AuAtomicUtils<T>::Set(in, offset);
}

template <class T>
auline
bool AuAtomicSet(T *in, AuUInt8 offset)
{
    return AuAtomicOrSetBit(in, offset) & (T(1) << offset);
}

template <class T>
auline
T    AuAtomicAndUnsetBit(T *in, AuUInt8 offset)
{
    return AuAtomicUtils<T>::And(in, ~(T(1) << T(offset)));
}

template <class T>
auline
bool AuAtomicUnset(T *in, AuUInt8 offset)
{
    auto uBit = T(1) << T(offset);
    return AuAtomicUtils<T>::And(in, ~(uBit)) & uBit;
}

template <class T>
auline
T    AuAtomicOr(T *in, T value)
{
    return AuAtomicUtils<T>::Or(in, value);
}

template <class T>
auline
T    AuAtomicAnd(T *in, T value)
{
    return AuAtomicUtils<T>::And(in, value);
}

template <class T>
auline
T    AuAtomicAdd(T *in, T addend)
{
    return AuAtomicUtils<T>::Add(in, addend);
}

template <class T>
auline
T    AuAtomicSub(T *in, T minuend)
{
    return AuAtomicUtils<T>::Sub(in, minuend);
}

template <class T>
auline
T    AuAtomicCompareExchange(T *in, T replace, T compare)
{
    return AuAtomicUtils<T>::CompareExchange(in, replace, compare);
}

template <class T>
auline
bool AuAtomicTestAndSet(T *in, AuUInt8 offset)
{
    return AuAtomicUtils<T>::TestAndSet(in, offset);
}

template <class T>
auline
void     AuAtomicStore(volatile T *in, T value)
{
    AuAtomicUtils<T>::Store((T *)in, value);
}

template <class T>
auline
T        AuAtomicLoad(volatile T *in)
{
    return AuAtomicUtils<T>::Load((T *)in);
}

template <class T>
auline
T        AuAtomicLoadWeak(volatile T *in)
{
    return AuAtomicUtils<T>::LoadWeak((T *)in);
}

template <class T>
auline
void     AuAtomicStoreWeak(volatile T *in, T value)
{
    AuAtomicUtils<T>::StoreWeak((T *)in, value);
}

template <class T>
auline
void     AuAtomicClearU8Lock(volatile T *in)
{
    AuAtomicUtils<T>::ClearU8Lock((T *)in);
}

template <class T>
auline
T    AuAtomicOrSetBit(volatile T *in, AuUInt8 offset)
{
    return AuAtomicOrSetBit((T *)in, offset);
}

template <class T>
auline
bool AuAtomicSet(volatile T *in, AuUInt8 offset)
{
    return AuAtomicSet((T *)in, offset);
}

template <class T>
auline
T    AuAtomicAndUnsetBit(volatile T *in, AuUInt8 offset)
{
    return AuAtomicAndUnsetBit((T *)in, offset);
}

template <class T>
auline
bool AuAtomicUnset(volatile T *in, AuUInt8 offset)
{
    return AuAtomicUnset((T *)in, offset);
}

template <class T>
auline
T    AuAtomicOr(volatile T *in, T value)
{
    return AuAtomicOr((T *)in, value);
}

template <class T>
auline
T    AuAtomicAnd(volatile T *in, T value)
{
    return AuAtomicAnd((T *)in, value);
}

template <class T>
auline
T    AuAtomicAdd(volatile T *in, T addend)
{
    return AuAtomicUtils<T>::Add((T *)(in), addend);
}

template <class T>
auline
T    AuAtomicSub(volatile T *in, T minuend)
{
    return AuAtomicUtils<T>::Sub((T *)(in), minuend);
}

template <class T>
auline
T    AuAtomicCompareExchange(volatile T *in, T replace, T compare)
{
    return AuAtomicUtils<T>::CompareExchange((T *)(in), replace, compare);
}

template <class T>
auline
bool AuAtomicTestAndSet(volatile T *in, AuUInt8 offset)
{
    return AuAtomicUtils<T>::TestAndSet((T *)(in), offset);
}