Question

I am trying to use the MS VC++ Intrinsic InterlockedCompareExchange128 function.

As a hello-world, I am trying to compare a 16byte address with itself, and replace it with something else. This compiles, but it’s not working – the address is not exchanged with new values. The const_cast is used to make it compile (otherwise it cries for not being able to cast volatile).

typedef struct t_node
{
    volatile __int64 arr[2];
}node;

int main()
{   

    node *a = new node();

    a->arr[0] = 100;
    a->arr[1] = 1;

    __int64 i = 200;
    __int64 j = 500;

    char r = _InterlockedCompareExchange128(a->arr, i,j, const_cast<__int64*>(&a->arr[0]));

    cout<<endl<<"Interlocked Compare Res: "<<r;

    cin.get();
    return 0;
}

Answer 1

From the documentation:

unsigned char _InterlockedCompareExchange128(
   __int64 volatile * Destination,
   __int64 ExchangeHigh,
   __int64 ExchangeLow,
   __int64 * ComparandResult
);

[in, out] ComparandResult
Pointer to an array of two 64-bit integers (considered as a 128-bit field) to compare with the destination. On output, this is overwritten with the original value of the destination.

Thus, what happens is in pseudocode:

if(ComparandResult != Destination)
{
    temp = Destination
    Destination = ExchangeHigh:ExchangeLow
    ComparandResult = temp
}

What happens when Destination == ComparandResult (your case) is:

if(ComparandResult != Destination)
{
    temp = Destination
    Destination = ExchangeHigh:ExchangeLow
    Destination = temp
}

Which is a nop.
In addition there is a note for that in the same page:

Note
The value of ComparandResult is always overwritten. After the
lock instruction, this intrinsic immediately copies the initial value
of Destination to ComparandResult. For this reason, ComparandResult
and Destination should point to separate memory locations to avoid
unexpected behavior.