/**
 * A D implementation of the C stdatomic.h header.
 *
 * $(NOTE If it compiles it should produce similar assembly to the system C toolchain
 *   and should not introduce when optimizing unnecessary behaviors,
 *   if you do not care about this guarantee use the _impl suffix.)
 *
 * $(NOTE The D shared type qualifier is the closest to the _Atomic type qualifier from C. It may be changed from shared in the future.)
 *
 * $(NOTE Fail memory order is currently ignored due to limitations with internal implementation of atomics.)
 *
 * Copyright: Copyright Richard (Rikki) Andrew Cattermole 2023.
 * License:   $(LINK2 http://www.boost.org/LICENSE_1_0.txt, Boost License 1.0)
 * Authors:   Richard (Rikki) Andrew cattermole
 * Source:    $(DRUNTIMESRC core/stdc/stdatomic.d)
 */
module core.stdc.stdatomic;
import core.atomic : MemoryOrder;
import core.internal.atomic;
import core.stdc.config;
import core.stdc.stddef;
import core.stdc.stdint;

@safe nothrow @nogc:

///
enum memory_order
{
    /// No ordering provided
    memory_order_relaxed = MemoryOrder.raw,
    /// As per cppreference.com circa 2015 no compiler supports consume memory order and in practice it devolves to acquire.
    memory_order_consume = MemoryOrder.acq,
    /// Prevent reordering before operation
    memory_order_acquire = MemoryOrder.acq,
    /// Prevent reordering after operation
    memory_order_release = MemoryOrder.rel,
    /// Prevent reordering before and after operation
    memory_order_acq_rel = MemoryOrder.acq_rel,
    /// Prevent reordering before for read operations and after for writes.
    memory_order_seq_cst = MemoryOrder.seq
}

///
enum
{
    ///
    __STDC_VERSION_STDATOMIC_H__ = 202311,

    ///
    ATOMIC_BOOL_LOCK_FREE = IsAtomicLockFree!bool ? 2 : 0,
    ///
    ATOMIC_CHAR_LOCK_FREE = IsAtomicLockFree!char ? 2 : 0,
    ///
    ATOMIC_CHAR16_T_LOCK_FREE = IsAtomicLockFree!wchar ? 2 : 0,
    ///
    ATOMIC_CHAR32_T_LOCK_FREE = IsAtomicLockFree!dchar ? 2 : 0,
    ///
    ATOMIC_WCHAR_T_LOCK_FREE = IsAtomicLockFree!wchar_t ? 2 : 0,
    ///
    ATOMIC_SHORT_LOCK_FREE = IsAtomicLockFree!short ? 2 : 0,
    ///
    ATOMIC_INT_LOCK_FREE = IsAtomicLockFree!int ? 2 : 0,
    ///
    ATOMIC_LONG_LOCK_FREE = IsAtomicLockFree!c_long ? 2 : 0,
    ///
    ATOMIC_LLONG_LOCK_FREE = IsAtomicLockFree!ulong ? 2 : 0,
    ///
    ATOMIC_POINTER_LOCK_FREE = IsAtomicLockFree!(void*) ? 2 : 0,
    ///
    ATOMIC_CHAR8_T_LOCK_FREE = ATOMIC_CHAR_LOCK_FREE,
}

version (DigitalMars)
{
    alias atomic_signal_fence = atomic_signal_fence_impl; ///

    // these all use inline assembly, so will unlikely produce the codegen a user will expect
    version(none)
    {
        alias atomic_flag_clear = atomic_flag_clear_impl; ///
        alias atomic_flag_clear_explicit = atomic_flag_clear_explicit_impl; ///
        alias atomic_flag_test_and_set = atomic_flag_test_and_set_impl; ///
        alias atomic_flag_test_and_set_explicit = atomic_flag_test_and_set_explicit_impl; ///
        alias atomic_thread_fence = atomic_thread_fence_impl; ///
        alias atomic_store = atomic_store_impl; ///
        alias atomic_store_explicit = atomic_store_explicit_impl; ///
        alias atomic_load = atomic_load_impl; ///
        alias atomic_load_explicit = atomic_load_explicit_impl; ///
        alias atomic_exchange = atomic_exchange_impl; ///
        alias atomic_exchange_explicit = atomic_exchange_explicit_impl; ///
        alias atomic_compare_exchange_strong = atomic_compare_exchange_strong_impl; ///
        alias atomic_compare_exchange_weak = atomic_compare_exchange_weak_impl; ///
        alias atomic_compare_exchange_strong_explicit = atomic_compare_exchange_strong_explicit_impl; ///
        alias atomic_compare_exchange_weak_explicit = atomic_compare_exchange_weak_explicit_impl; ///
        alias atomic_fetch_add = atomic_fetch_add_impl; ///
        alias atomic_fetch_add_explicit = atomic_fetch_add_explicit_impl; ///
        alias atomic_fetch_sub = atomic_fetch_sub_impl; ///
        alias atomic_fetch_sub_explicit = atomic_fetch_sub_explicit_impl; ///
        alias atomic_fetch_or = atomic_fetch_or_impl; ///
        alias atomic_fetch_or_explicit = atomic_fetch_or_explicit_impl; ///
        alias atomic_fetch_xor = atomic_fetch_xor_impl; ///
        alias atomic_fetch_xor_explicit = atomic_fetch_xor_explicit_impl; ///
        alias atomic_fetch_and = atomic_fetch_and_impl; ///
        alias atomic_fetch_and_explicit = atomic_fetch_and_explicit_impl; ///
    }
}
else
{
    alias atomic_flag_clear = atomic_flag_clear_impl; ///
    alias atomic_flag_clear_explicit = atomic_flag_clear_explicit_impl; ///
    alias atomic_flag_test_and_set = atomic_flag_test_and_set_impl; ///
    alias atomic_flag_test_and_set_explicit = atomic_flag_test_and_set_explicit_impl; ///
    alias atomic_signal_fence = atomic_signal_fence_impl; ///
    alias atomic_thread_fence = atomic_thread_fence_impl; ///
    alias atomic_store = atomic_store_impl; ///
    alias atomic_store_explicit = atomic_store_explicit_impl; ///
    alias atomic_load = atomic_load_impl; ///
    alias atomic_load_explicit = atomic_load_explicit_impl; ///
    alias atomic_exchange = atomic_exchange_impl; ///
    alias atomic_exchange_explicit = atomic_exchange_explicit_impl; ///
    alias atomic_compare_exchange_strong = atomic_compare_exchange_strong_impl; ///
    alias atomic_compare_exchange_weak = atomic_compare_exchange_weak_impl; ///
    alias atomic_compare_exchange_strong_explicit = atomic_compare_exchange_strong_explicit_impl; ///
    alias atomic_compare_exchange_weak_explicit = atomic_compare_exchange_weak_explicit_impl; ///
    alias atomic_fetch_add = atomic_fetch_add_impl; ///
    alias atomic_fetch_add_explicit = atomic_fetch_add_explicit_impl; ///
    alias atomic_fetch_sub = atomic_fetch_sub_impl; ///
    alias atomic_fetch_sub_explicit = atomic_fetch_sub_explicit_impl; ///
    alias atomic_fetch_or = atomic_fetch_or_impl; ///
    alias atomic_fetch_or_explicit = atomic_fetch_or_explicit_impl; ///
    alias atomic_fetch_xor = atomic_fetch_xor_impl; ///
    alias atomic_fetch_xor_explicit = atomic_fetch_xor_explicit_impl; ///
    alias atomic_fetch_and = atomic_fetch_and_impl; ///
    alias atomic_fetch_and_explicit = atomic_fetch_and_explicit_impl; ///
}

///
pragma(inline, true)
bool atomic_is_lock_free(A)(const shared(A)* obj)
{
    return IsAtomicLockFree!A;
}

/// Guaranteed to be a atomic boolean type
struct atomic_flag
{
    private bool b;
}

///
enum ATOMIC_FLAG_INIT = atomic_flag.init;

///
pragma(inline, true)
void atomic_flag_clear_impl()(atomic_flag* obj)
{
    assert(obj !is null);

    atomicStore(&obj.b, false);
}

///
pragma(inline, true)
void atomic_flag_clear_explicit_impl()(atomic_flag* obj, memory_order order)
{
    assert(obj !is null);

    // use series of ternary expressions instead of switch to help the dmd inliner
    order == memory_order.memory_order_relaxed ?
        atomicStore!(memory_order.memory_order_relaxed)(&obj.b, false) :

    order == memory_order.memory_order_acquire ||
    order == memory_order.memory_order_acq_rel ?
        // Ideally this would error at compile time but alas it is not an intrinsic.
        // Note: this is not a valid memory order for this operation.
        atomicStore!(memory_order.memory_order_seq_cst)(&obj.b, false) :

    order == memory_order.memory_order_release ?
        atomicStore!(memory_order.memory_order_release)(&obj.b, false) :

    order == memory_order.memory_order_seq_cst ?
        atomicStore(&obj.b, false) :

    cast(void) assert(0);
}

///
pragma(inline, true)
bool atomic_flag_test_and_set_impl()(atomic_flag* obj)
{
    assert(obj !is null);
    return atomicExchange(&obj.b, true);
}

///
@trusted unittest
{
    atomic_flag flag;
    assert(!atomic_flag_test_and_set_impl(&flag));
    atomic_flag_clear_impl(&flag);
}

///
pragma(inline, true)
bool atomic_flag_test_and_set_explicit_impl()(atomic_flag* obj, memory_order order)
{
    assert(obj !is null);

    // use series of ternary expressions instead of switch to help the dmd inliner
    return
        order == memory_order.memory_order_relaxed ?
            atomicExchange!(memory_order.memory_order_relaxed)(&obj.b, true) :

        order == memory_order.memory_order_acquire ?
            // Ideally this would error at compile time but alas it is not an intrinsic.
            // Note: this is not a valid memory order for this operation.
            atomicExchange!(memory_order.memory_order_seq_cst)(&obj.b, true) :

         order == memory_order.memory_order_release ?
            atomicExchange!(memory_order.memory_order_release)(&obj.b, true) :

         order == memory_order.memory_order_acq_rel ?
            atomicExchange!(memory_order.memory_order_acq_rel)(&obj.b, true) :

         order == memory_order.memory_order_seq_cst ?
            atomicExchange(&obj.b, true) :

         assert(0);
}

///
@trusted unittest
{
    atomic_flag flag;
    assert(!atomic_flag_test_and_set_explicit_impl(&flag, memory_order.memory_order_seq_cst));
    atomic_flag_clear_explicit_impl(&flag, memory_order.memory_order_seq_cst);
}

/**
 * Initializes an atomic variable, the destination should not have any expression associated with it prior to this call.
 *
 * We use an out parameter instead of a pointer for destination in an attempt to communicate to the compiler that it initializers.
 */
pragma(inline, true)
void atomic_init(A, C)(out shared(A) obj, C desired) @trusted
{
    obj = cast(shared) desired;
}

///
unittest
{
    shared int val;
    atomic_init(val, 2);

    shared float valF;
    atomic_init(valF, 3.2);
}

/// No-op function, doesn't apply to D
pragma(inline, true)
A kill_dependency(A)(A y) @trusted
{
    return y;
}

/// Don't allow reordering, does not emit any instructions.
pragma(inline, true)
void atomic_signal_fence_impl()(memory_order order)
{
    // use series of ternary expressions instead of switch to help the dmd inliner
    order == memory_order.memory_order_relaxed ?
        // This is a no-op operation for relaxed memory orders.
        cast(void)0 :

    order == memory_order.memory_order_acquire ?
        atomicSignalFence!(memory_order.memory_order_acquire) :

    order == memory_order.memory_order_release ?
        atomicSignalFence!(memory_order.memory_order_release) :

    order == memory_order.memory_order_acq_rel ?
        atomicSignalFence!(memory_order.memory_order_acq_rel) :

    order == memory_order.memory_order_seq_cst ?
        atomicSignalFence!(memory_order.memory_order_seq_cst) :

    cast(void) assert(0);
}

///
unittest
{
    atomic_signal_fence_impl(memory_order.memory_order_seq_cst);
}

/// Don't allow reordering, and emit a fence instruction.
pragma(inline, true)
void atomic_thread_fence_impl()(memory_order order)
{
    // use series of ternary expressions instead of switch to help the dmd inliner
    order == memory_order.memory_order_relaxed ?
        // This is a no-op operation for relaxed memory orders.
        cast(void)0 :

    order == memory_order.memory_order_acquire ?
        atomicFence!(memory_order.memory_order_acquire) :

    order == memory_order.memory_order_release ?
        atomicFence!(memory_order.memory_order_release) :

    order == memory_order.memory_order_acq_rel ?
        atomicFence!(memory_order.memory_order_acq_rel) :

    order == memory_order.memory_order_seq_cst ?
        atomicFence!(memory_order.memory_order_seq_cst) :

    cast(void) assert(0);
}

///
unittest
{
    atomic_thread_fence_impl(memory_order.memory_order_seq_cst);
}

///
alias atomic_bool = shared(bool);
///
alias atomic_char = shared(char);
///
alias atomic_schar = shared(byte);
///
alias atomic_uchar = shared(ubyte);
///
alias atomic_short = shared(short);
///
alias atomic_ushort = shared(ushort);
///
alias atomic_int = shared(int);
///
alias atomic_uint = shared(uint);
///
alias atomic_long = shared(c_long);
///
alias atomic_ulong = shared(c_ulong);
///
alias atomic_llong = shared(long);
///
alias atomic_ullong = shared(ulong);
///
alias atomic_char8_t = shared(char);
///
alias atomic_char16_t = shared(wchar);
///
alias atomic_char32_t = shared(dchar);
///
alias atomic_wchar_t = shared(wchar_t);

///
alias atomic_int_least8_t = shared(int_least8_t);
///
alias atomic_uint_least8_t = shared(uint_least8_t);
///
alias atomic_int_least16_t = shared(int_least16_t);
///
alias atomic_uint_least16_t = shared(uint_least16_t);
///
alias atomic_int_least32_t = shared(int_least32_t);
///
alias atomic_uint_least32_t = shared(uint_least32_t);
///
alias atomic_int_least64_t = shared(int_least64_t);
///
alias atomic_uint_least64_t = shared(uint_least64_t);
///
alias atomic_int_fast8_t = shared(int_fast8_t);
///
alias atomic_uint_fast8_t = shared(uint_fast8_t);
///
alias atomic_int_fast16_t = shared(int_fast16_t);
///
alias atomic_uint_fast16_t = shared(uint_fast16_t);
///
alias atomic_int_fast32_t = shared(int_fast32_t);
///
alias atomic_uint_fast32_t = shared(uint_fast32_t);
///
alias atomic_int_fast64_t = shared(int_fast64_t);
///
alias atomic_uint_fast64_t = shared(uint_fast64_t);
///
alias atomic_intptr_t = shared(intptr_t);
///
alias atomic_uintptr_t = shared(uintptr_t);
///
alias atomic_size_t = shared(size_t);
///
alias atomic_ptrdiff_t = shared(ptrdiff_t);
///
alias atomic_intmax_t = shared(intmax_t);
///
alias atomic_uintmax_t = shared(uintmax_t);

///
pragma(inline, true)
void atomic_store_impl(A, C)(shared(A)* obj, C desired) @trusted
{
    assert(obj !is null);
    atomicStore(cast(A*)obj, cast(A)desired);
}

///
@trusted unittest
{
    shared(int) obj;
    atomic_store_impl(&obj, 3);

    shared(float) objF;
    atomic_store_impl(&objF, 3.21);
}

///
pragma(inline, true)
void atomic_store_explicit_impl(A, C)(shared(A)* obj, C desired, memory_order order) @trusted
{
    assert(obj !is null);

    // use series of ternary expressions instead of switch to help the dmd inliner
    order == memory_order.memory_order_relaxed ?
        atomicStore!(memory_order.memory_order_relaxed)(cast(A*)obj, cast(A)desired) :

    order == memory_order.memory_order_acquire ||
    order == memory_order.memory_order_acq_rel ?
        // Ideally this would error at compile time but alas it is not an intrinsic.
        // Note: this is not a valid memory order for this operation.
        atomicStore!(memory_order.memory_order_release)(cast(A*)obj, cast(A)desired) :

    order == memory_order.memory_order_release ?
        atomicStore!(memory_order.memory_order_release)(cast(A*)obj, cast(A)desired) :

    order == memory_order.memory_order_seq_cst ?
        atomicStore!(memory_order.memory_order_seq_cst)(cast(A*)obj, cast(A)desired) :

    cast(void) assert(0);
}

///
@trusted unittest
{
    shared(int) obj;
    atomic_store_explicit_impl(&obj, 3, memory_order.memory_order_seq_cst);

    shared(float) objF;
    atomic_store_explicit_impl(&objF, 3.21, memory_order.memory_order_seq_cst);
}

///
pragma(inline, true)
A atomic_load_impl(A)(const shared(A)* obj) @trusted
{
    assert(obj !is null);
    return atomicLoad(cast(A*)obj);
}

///
@trusted unittest
{
    shared(int) obj = 3;
    assert(atomic_load_impl(&obj) == 3);

    shared(float) objF = 3.5;
    assert(atomic_load_impl(&objF) > 3);

    static struct S2
    {
        size_t[2] values;
    }
    align(S2.sizeof) shared(S2) objS2 = {[1, 2]};
    assert(atomic_load_impl(&objS2).values == [1, 2]);
}

///
pragma(inline, true)
A atomic_load_explicit_impl(A)(const shared(A)* obj, memory_order order) @trusted
{
    assert(obj !is null);

    // use series of ternary expressions instead of switch to help the dmd inliner
    return
        order == memory_order.memory_order_relaxed ?
            atomicLoad!(memory_order.memory_order_relaxed)(cast(A*)obj) :

        order == memory_order.memory_order_acquire ?
            atomicLoad!(memory_order.memory_order_acquire)(cast(A*)obj) :

        order == memory_order.memory_order_release ||
        order == memory_order.memory_order_acq_rel ?
            // Ideally this would error at compile time but alas it is not an intrinsic.
            // Note: this is not a valid memory order for this operation.
            atomicLoad!(memory_order.memory_order_acquire)(cast(A*)obj) :

        order == memory_order.memory_order_seq_cst ?
            atomicLoad!(memory_order.memory_order_seq_cst)(cast(A*)obj) :

        assert(0);
}

///
@trusted unittest
{
    shared(int) obj = 3;
    assert(atomic_load_explicit_impl(&obj, memory_order.memory_order_seq_cst) == 3);

    shared(float) objF = 3.5;
    assert(atomic_load_explicit_impl(&objF, memory_order.memory_order_seq_cst) > 3);
}

///
pragma(inline, true)
A atomic_exchange_impl(A, C)(shared(A)* obj, C desired) @trusted
{
    assert(obj !is null);
    return atomicExchange(cast(A*)obj, cast(A)desired);
}

///
@trusted unittest
{
    shared(int) obj = 3;
    assert(atomic_exchange_impl(&obj, 2) == 3);

    shared(float) objF = 3;
    assert(atomic_exchange_impl(&objF, 2.1) > 2.5);
}

///
pragma(inline, true)
A atomic_exchange_explicit_impl(A, C)(shared(A)* obj, C desired, memory_order order) @trusted
{
    assert(obj !is null);

    // use series of ternary expressions instead of switch to help the dmd inliner
    return
        order == memory_order.memory_order_relaxed ?
            atomicExchange!(memory_order.memory_order_relaxed)(cast(A*)obj, cast(A)desired) :

        order == memory_order.memory_order_acquire ?
            // Ideally this would error at compile time but alas it is not an intrinsic.
            // Note: this is not a valid memory order for this operation.
            atomicExchange!(memory_order.memory_order_seq_cst)(cast(A*)obj, cast(A)desired) :

        order == memory_order.memory_order_release ?
            atomicExchange!(memory_order.memory_order_release)(cast(A*)obj, cast(A)desired) :

        order == memory_order.memory_order_acq_rel ?
            atomicExchange!(memory_order.memory_order_acq_rel)(cast(A*)obj, cast(A)desired) :

        order == memory_order.memory_order_seq_cst ?
            atomicExchange!(memory_order.memory_order_seq_cst)(cast(A*)obj, cast(A)desired) :

        assert(0);
}

///
@trusted unittest
{
    shared(int) obj = 3;
    assert(atomic_exchange_explicit_impl(&obj, 2, memory_order.memory_order_seq_cst) == 3);

    shared(float) objF = 1.5;
    assert(atomic_exchange_explicit_impl(&objF, 2.1, memory_order.memory_order_seq_cst) < 2);
}

///
pragma(inline, true)
bool atomic_compare_exchange_strong_impl(A, B, C)(shared(A)* obj, B* expected, C desired) @trusted
{
    static assert(is(shared(B) : A), "Both expected and object must be the same type");
    return atomicCompareExchangeStrong(cast(B*)obj, expected, cast(B)desired);
}

///
@trusted unittest
{
    shared(int) obj = 3;
    int expected = 3;
    assert(atomic_compare_exchange_strong_impl(&obj, &expected, 2));
}

///
@trusted unittest
{
    shared(float) obj = 3;
    float expected = 3;
    assert(atomic_compare_exchange_strong_impl(&obj, &expected, 2.1));
}

///
pragma(inline, true)
bool atomic_compare_exchange_weak_impl(A, B, C)(shared(A)* obj, B* expected, C desired) @trusted
{
    static assert(is(shared(B) : A), "Both expected and object must be the same type");
    return atomicCompareExchangeWeak(cast(B*)obj, expected, cast(B)desired);
}

///
@trusted unittest
{
    shared(int) obj = 3;
    int expected = 3;
    static assert(__traits(compiles, {atomic_compare_exchange_weak_impl(&obj, &expected, 2);}));
}

///
@trusted unittest
{
    shared(float) obj = 3;
    float expected = 3;
    static assert(__traits(compiles, {atomic_compare_exchange_weak_impl(&obj, &expected, 2.1);}));
}

///
pragma(inline, true)
bool atomic_compare_exchange_strong_explicit_impl(A, B, C)(shared(A)* obj, B* expected, C desired, memory_order succ, memory_order fail) @trusted
{
    static assert(is(shared(B) : A), "Both expected and object must be the same type");
    assert(obj !is null);
    // NOTE: To not have to deal with all invalid cases, the failure model is ignored for now.

    // use series of ternary expressions instead of switch to help the dmd inliner
    return
        succ == memory_order.memory_order_relaxed ?
            atomicCompareExchangeStrong!(memory_order.memory_order_relaxed, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        succ == memory_order.memory_order_acquire ?
            atomicCompareExchangeStrong!(memory_order.memory_order_acquire, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        succ == memory_order.memory_order_release ?
            atomicCompareExchangeStrong!(memory_order.memory_order_release, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        succ == memory_order.memory_order_acq_rel ?
            atomicCompareExchangeStrong!(memory_order.memory_order_acq_rel, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        succ == memory_order.memory_order_seq_cst ?
            atomicCompareExchangeStrong!(memory_order.memory_order_seq_cst, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        assert(0);
}

///
@trusted unittest
{
    shared(int) obj = 3;
    int expected = 3;
    assert(atomic_compare_exchange_strong_explicit_impl(&obj, &expected, 2, memory_order.memory_order_seq_cst, memory_order.memory_order_seq_cst));
}

///
@trusted unittest
{
    align(size_t[2].sizeof) shared(size_t[2]) obj = [3, 4];
    size_t[2] expected = [3, 4];
    size_t[2] toSwap = [1, 2];
    assert(atomic_compare_exchange_strong_explicit_impl(&obj, &expected, toSwap, memory_order.memory_order_seq_cst, memory_order.memory_order_seq_cst));
}

///
@trusted unittest
{
    shared(float) obj = 3;
    float expected = 3;
    assert(atomic_compare_exchange_strong_explicit_impl(&obj, &expected, 2.1, memory_order.memory_order_seq_cst, memory_order.memory_order_seq_cst));
}

///
pragma(inline, true)
bool atomic_compare_exchange_weak_explicit_impl(A, B, C)(shared(A)* obj, B* expected, C desired, memory_order succ, memory_order fail) @trusted
{
    static assert(is(shared(B) : A), "Both expected and object must be the same type");
    assert(obj !is null);
    // NOTE: To not have to deal with all invalid cases, the failure model is ignored for now.

    // use series of ternary expressions instead of switch to help the dmd inliner
    return
        succ == memory_order.memory_order_relaxed ?
            atomicCompareExchangeWeak!(memory_order.memory_order_relaxed, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        succ == memory_order.memory_order_acquire ?
            atomicCompareExchangeWeak!(memory_order.memory_order_acquire, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        succ == memory_order.memory_order_release ?
            atomicCompareExchangeWeak!(memory_order.memory_order_release, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        succ == memory_order.memory_order_acq_rel ?
            atomicCompareExchangeWeak!(memory_order.memory_order_acq_rel, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        succ == memory_order.memory_order_seq_cst ?
            atomicCompareExchangeWeak!(memory_order.memory_order_seq_cst, memory_order.memory_order_relaxed)(cast(B*)obj, expected, cast(B)desired) :
        assert(0);
}

///
@trusted unittest
{
    shared(int) obj = 3;
    int expected = 3;
    atomic_compare_exchange_weak_explicit_impl(&obj, &expected, 2, memory_order.memory_order_seq_cst, memory_order.memory_order_seq_cst);
}

///
@trusted unittest
{
    align(size_t[2].sizeof) shared(size_t[2]) obj = [3, 4];
    size_t[2] expected = [3, 4];
    size_t[2] toSwap = [1, 2];
    assert(atomic_compare_exchange_weak_explicit_impl(&obj, &expected, toSwap, memory_order.memory_order_seq_cst, memory_order.memory_order_seq_cst));
}

///
@trusted unittest
{
    shared(float) obj = 3;
    float expected = 3;
    atomic_compare_exchange_weak_explicit_impl(&obj, &expected, 2, memory_order.memory_order_seq_cst, memory_order.memory_order_seq_cst);
}

///
pragma(inline, true)
A atomic_fetch_add_impl(A, M)(shared(A)* obj, M arg) @trusted
{
    assert(obj !is null);
    return atomic_fetch_op!(memory_order.memory_order_seq_cst, "+=")(cast(A*)obj, arg);
}

///
@trusted unittest
{
    shared(int) val;
    atomic_fetch_add_impl(&val, 3);
    assert(atomic_load_impl(&val) == 3);

    shared(float) valF = 0.5;
    atomic_fetch_add_impl(&valF, 3);
    assert(atomic_load_impl(&valF) > 3);
}

pragma(inline, true)
A atomic_fetch_sub_impl(A, M)(shared(A)* obj, M arg) @trusted
{
    assert(obj !is null);
    return atomic_fetch_op!(memory_order.memory_order_seq_cst, "-=")(cast(A*)obj, arg);
}

///
@trusted unittest
{
    shared(int) val = 3;
    atomic_fetch_sub_impl(&val, 3);
    assert(atomic_load_impl(&val) == 0);

    shared(float) valF = 3;
    atomic_fetch_sub_impl(&valF, 1);
    assert(atomic_load_impl(&valF) < 3);
}

///
pragma(inline, true)
A atomic_fetch_add_explicit_impl(A, M)(shared(A)* obj, M arg, memory_order order) @trusted
{
    assert(obj !is null);

    // use series of ternary expressions instead of switch to help the dmd inliner
    return
        order == memory_order.memory_order_relaxed ?
            atomic_fetch_op!(memory_order.memory_order_relaxed, "+=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acquire ?
            atomic_fetch_op!(memory_order.memory_order_acquire, "+=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_release ?
            atomic_fetch_op!(memory_order.memory_order_release, "+=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acq_rel ?
            atomic_fetch_op!(memory_order.memory_order_acq_rel, "+=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_seq_cst ?
            atomic_fetch_op!(memory_order.memory_order_seq_cst, "+=")(cast(A*)obj, arg) :
        assert(0);
}

///
@trusted unittest
{
    shared(int) val;
    atomic_fetch_add_explicit_impl(&val, 3, memory_order.memory_order_seq_cst);
    assert(atomic_load_impl(&val) == 3);

    shared(float) valF = 3;
    atomic_fetch_add_explicit_impl(&valF, 3, memory_order.memory_order_seq_cst);
    assert(atomic_load_impl(&valF) > 3);
}

///
pragma(inline, true)
A atomic_fetch_sub_explicit_impl(A, M)(shared(A)* obj, M arg, memory_order order) @trusted
{
    assert(obj !is null);

    return
        order == memory_order.memory_order_relaxed ?
            atomic_fetch_op!(memory_order.memory_order_relaxed, "-=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acquire ?
            atomic_fetch_op!(memory_order.memory_order_acquire, "-=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_release ?
            atomic_fetch_op!(memory_order.memory_order_release, "-=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acq_rel ?
            atomic_fetch_op!(memory_order.memory_order_acq_rel, "-=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_seq_cst ?
            atomic_fetch_op!(memory_order.memory_order_seq_cst, "-=")(cast(A*)obj, arg) :
        assert(0);
}

///
@trusted unittest
{
    shared(int) val = 3;
    atomic_fetch_sub_explicit_impl(&val, 3, memory_order.memory_order_seq_cst);
    assert(atomic_load_impl(&val) == 0);

    shared(float) valF = 4;
    atomic_fetch_sub_explicit_impl(&valF, 3, memory_order.memory_order_seq_cst);
    assert(atomic_load_impl(&valF) < 4);
}

///
pragma(inline, true)
A atomic_fetch_or_impl(A, M)(shared(A)* obj, M arg) @trusted
{
    assert(obj !is null);
    return atomic_fetch_op!(memory_order.memory_order_seq_cst, "|=")(cast(A*)obj, arg);
}

///
@trusted unittest
{
    shared(int) val = 5;
    atomic_fetch_or_impl(&val, 3);
    assert(atomic_load_impl(&val) == 7);
}

///
pragma(inline, true)
A atomic_fetch_or_explicit_impl(A, M)(shared(A)* obj, M arg, memory_order order) @trusted
{
    assert(obj !is null);

    return
        order == memory_order.memory_order_relaxed ?
            atomic_fetch_op!(memory_order.memory_order_relaxed, "|=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acquire ?
            atomic_fetch_op!(memory_order.memory_order_acquire, "|=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_release ?
            atomic_fetch_op!(memory_order.memory_order_release, "|=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acq_rel ?
            atomic_fetch_op!(memory_order.memory_order_acq_rel, "|=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_seq_cst ?
            atomic_fetch_op!(memory_order.memory_order_seq_cst, "|=")(cast(A*)obj, arg) :
        assert(0);
}

///
@trusted unittest
{
    shared(int) val = 5;
    atomic_fetch_or_explicit_impl(&val, 3, memory_order.memory_order_seq_cst);
    assert(atomic_load_impl(&val) == 7);
}

///
pragma(inline, true)
A atomic_fetch_xor_impl(A, M)(shared(A)* obj, M arg) @trusted
{
    assert(obj !is null);
    return atomic_fetch_op!(memory_order.memory_order_seq_cst, "^=")(cast(A*)obj, arg);
}

///
@trusted unittest
{
    shared(int) val = 5;
    atomic_fetch_xor_impl(&val, 3);
    assert(atomic_load_impl(&val) == 6);
}

///
pragma(inline, true)
A atomic_fetch_xor_explicit_impl(A, M)(shared(A)* obj, M arg, memory_order order) @trusted
{
    assert(obj !is null);

    return
        order == memory_order.memory_order_relaxed ?
            atomic_fetch_op!(memory_order.memory_order_relaxed, "^=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acquire ?
            atomic_fetch_op!(memory_order.memory_order_acquire, "^=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_release ?
            atomic_fetch_op!(memory_order.memory_order_release, "^=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acq_rel ?
            atomic_fetch_op!(memory_order.memory_order_acq_rel, "^=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_seq_cst ?
            atomic_fetch_op!(memory_order.memory_order_seq_cst, "^=")(cast(A*)obj, arg) :
        assert(0);
}

///
@trusted unittest
{
    shared(int) val = 5;
    atomic_fetch_xor_explicit_impl(&val, 3, memory_order.memory_order_seq_cst);
    assert(atomic_load_impl(&val) == 6);
}

///
pragma(inline, true)
A atomic_fetch_and_impl(A, M)(shared(A)* obj, M arg) @trusted
{
    assert(obj !is null);
    return atomic_fetch_op!(memory_order.memory_order_seq_cst, "&=")(cast(A*)obj, arg);
}

///
@trusted unittest
{
    shared(int) val = 5;
    atomic_fetch_and_impl(&val, 3);
    assert(atomic_load_impl(&val) == 1);
}

///
pragma(inline, true)
A atomic_fetch_and_explicit_impl(A, M)(shared(A)* obj, M arg, memory_order order) @trusted
{
    assert(obj !is null);

    return
        order == memory_order.memory_order_relaxed ?
            atomic_fetch_op!(memory_order.memory_order_relaxed, "&=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acquire ?
            atomic_fetch_op!(memory_order.memory_order_acquire, "&=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_release ?
            atomic_fetch_op!(memory_order.memory_order_release, "&=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_acq_rel ?
            atomic_fetch_op!(memory_order.memory_order_acq_rel, "&=")(cast(A*)obj, arg) :
        order == memory_order.memory_order_seq_cst ?
            atomic_fetch_op!(memory_order.memory_order_seq_cst, "&=")(cast(A*)obj, arg) :
        assert(0);
}

///
unittest
{
    shared(int) val = 5;
    atomic_fetch_and_explicit_impl(&val, 3, memory_order.memory_order_seq_cst);
    assert(atomic_load_impl(&val) == 1);
}

private:

A atomic_fetch_op(memory_order order, string op, A, M)(A* obj, M arg) @trusted
{
    static if (is(A : ulong) && (op == "+=" || op == "-="))
    {
        pragma(inline, true)
        // these cannot handle floats
        static if (op == "+=")
        {
            return atomicFetchAdd!order(obj, arg);
        }
        else static if (op == "-=")
        {
            return atomicFetchSub!order(obj, arg);
        }
    }
    else
    {
        // copied from core.atomic
        A set, get = atomicLoad!(MemoryOrder.raw, A)(obj);
        do
        {
            set = get;
            mixin("set " ~ op ~ " arg;"); // will error if op (which is not exposed to user) is invalid
        } while (!atomicCompareExchangeWeak!(order, MemoryOrder.raw)(obj, &get, set));
        return get; // unlike core.atomic we return the prior value, not the new one.
    }
}