Question

I was wondering if there’s a windows API to set a 16-byte array atomically?

I would really hate to introduce critical sections or mutexes into this just to implement this operations, so I’m trying to find an easier solution.

PS. I need this because this 16-byte array can be written into from a worker thread, and mostly read from a main thread.

Answer 1

All of the following assumes a 64-bit x86_64 architecture. I believe that in general what you asked for is not possible on 32-bit x86.

There are really two options. The first one is _InterlockedCompareExchange128, which translates to LOCK CMPXCHG16B. To duplicate the functionality of a 16-byte store with it, you’d need to do something like this:

__int64* dest = ...;
__int64* orig = ...;
unsigned char ok;
do
{
    __int64 high = dest[0];
    __int64 low  = dest[1];
    ok = _InterlockedCompareExchange128(dest, high, low, orig);
}
while (!ok);

Note that because this is an interlocked operation, it implies the equivalent of a _ReadWriteBarrier. This is normally what is meant by an “atomic” operation.

If you want a pure store, not a compare-exchange, then you can use anything that translates to an aligned 16-byte store on the hardware. For example, the MOVDQA instruction would qualify. In C++, that’d look something like this:

__m128i* dest = ...;
_mm_mfence();
_mm_store_si128(dest, _mm_loadu_si128((__m128i*)&orig));
_mm_mfence();

Unless you have a reason to avoid SSE instructions, I think the second version might perform better. You also may want to decide if you want to relax the two memory barriers depending on your actual needs.