05-02-13 - Simple C++0x style LF structs and adapter for x86-Windows

I've seen a lot of lockfree libraries out there that are total crap. Really weird non-standard ways of doing things, or overly huge and complex.

I thought I'd make a super simple one in the correct modern style. Download : cbliblf.zip

(If you want a big fully functional much-more-complete library, Intel TBB is the best I've seen. The problem with TBB is that it's huge and entangled, and the license is not clearly free for all use).

There are two pieces here :

"cblibCpp0x.h" provides atomic and such in C++0x style for MSVC/Windows/x86 compilers that don't have real C++0x yet. I have made zero attempt to make this header syntatically identical to C++0x, there are various intentional and unintentional differences.

"cblibLF.h" provides some simple lockfree utilities (mostly queues) built on C++0x atomics.

"cblibCpp0x.h" is kind of by design not at all portable. "cblibLF.h" should be portable to any C++0x platform.

WARNING : this stuff is not super well tested because it's not what I use in Oodle. I've mostly copy-pasted this from my Relacy test code, so it should be pretty strong but there may have been some copy-paste errors.

ADDENDUM : In case it's not clear, you do not *want* to use "cblibCpp0x.h". You want to use real Cpp0x atomics provided by your compiler. This is a temporary band-aid so that people like me who use old compilers can get a cpp0x stand-in, so that they can do work using the modern syntax. If you're on a gcc platform that has the __atomic extensions but not C1X, use that.

You should be able to take any of the C++0x-style lockfree code I've posted over the years and use it with "cblibCpp0x.h" , perhaps with some minor syntactic fixes. eg. you could take the fastsemaphore wrapper and put the "semaphore" from "cblibCpp0x.h" in there as the base semaphore.

Here's an example of what the objects in "cblibLF.h" look like :

//=================================================================         
// spsc fifo
//  lock free for single producer, single consumer
//  requires an allocator
//  and a dummy node so the fifo is never empty
template <typename t_data>
struct lf_spsc_fifo_t
{
public:

    lf_spsc_fifo_t()
    {
        // initialize with one dummy node :
        node * dummy = new node;
        m_head = dummy;
        m_tail = dummy;
    }

    ~lf_spsc_fifo_t()
    {
        // should be one node left :
        LF_OS_ASSERT( m_head == m_tail );
        delete m_head;
    }

    void push(const t_data & data)
    {
        node * n = new node(data);
        // n->next == NULL from constructor
        m_head->next.store(n, memory_order_release); 
        m_head = n;
    }

    // returns true if a node was popped
    //  fills *pdata only if the return value is true
    bool pop(t_data * pdata)
    {
        // we're going to take the data from m_tail->next
        //  and free m_tail
        node* t = m_tail;
        node* n = t->next.load(memory_order_acquire);
        if ( n == NULL )
            return false;
        *pdata = n->data; // could be a swap
        m_tail = n;
        delete t;
        return true;
    }

private:

    struct node
    {
        atomic<node *>      next;
        nonatomic<t_data>   data;
        
        node() : next(NULL) { }
        node(const t_data & d) : next(NULL), data(d) { }
    };

    // head and tail are owned by separate threads,
    //  make sure there's no false sharing :
    nonatomic<node *>   m_head;
    char                m_pad[LF_OS_CACHE_LINE_SIZE];
    nonatomic<node *>   m_tail;
};

Download : cbliblf.zip