Question

I’m implementing a non-locking FIFO using a linked list.

The Enqueue of the FIFO is basically:

void Enqueue(CNode node)
{
  m_tail->m_next = node;

  // Do I need a memory barrier here?

  m_tail = node;
}

I’m wondering if there is any need to add a memory barrier if it’s single threaded(i.e., Could compiler/processor rearrange the order of the two lines above?). And what if it’s multi-threaded(i.e., as simple as single read single writer case)?

Edit: According to here, this is a case of data anti-dependency and statements should not be reordered. So I assume CPU should always access memory in the given order. Is that right?

Answer 1

The compiler must not rearrange your m_tail and m_tail->next assignment such that m_tail is assigned with node before m_tail->next has been set. However, for a multi-threaded solution, you could have to worry about:

temp = m_tail;
m_tail = node;
temp->next = node;
node->next = NULL; 

With a memory barrier, the compiler and/or the processor must complete the m_tail->next = node; (and node->next = NULL;) before it writes the m_tail = node;. Whether this is sufficent to guarantee correct execution is not certain, it depends a bit on what the code reading at the other end does.