load/store pair is never being executed out-of-order, even in relaxed mode

[aleh-kazakevich]

Hi, Dmitry!

First of all, thanks for the great tool (one of a kind)!!

---

Probably I missed something, but...
In my tests, 'load' is never being reordered with subsequent 'store', even when both tagged as 'relaxed'.

Example (see comment in 'after' method):

#include <relacy/relacy_std.hpp>

const size_t k_numberOfThreads = 2;

struct LoadStore : public rl::test_suite<LoadStore, k_numberOfThreads>
{
    rl::atomic<int> a = 0;
    rl::atomic<int> b = 0;
    int x = 0;
    int y = 0; 

    void thread(unsigned int index)
    {
        if (0 == index)
        {
            // Thread 1
            x = a.load(rl::memory_order_relaxed);
            b.store(1, rl::memory_order_relaxed);
        }
        else
        {
            // Thread 2
            y = b.load(rl::memory_order_relaxed);
            a.store(1, rl::memory_order_relaxed);
        }
    }
    
    void after()
    {
        const bool outOfOrder = ((1 == x) && (1 == y)); // <--- Never fired!
        RL_ASSERT(!outOfOrder);
    }
};

int main()
{
    rl::test_params p;
    p.search_type = rl::sched_full;
    rl::simulate<LoadStore>(p);
    return 0;
}

Tested on Clang 18/21 (Linux Fedora 40/43) with Relacy/Jan8'2026
С++20

I also found 2 similar cases with 'suspicious' results:

load-relaxed / load-seq_cst
store-seq_cst / store-relaxed
(both never being executed out-of-order)

---

Could you please confirm whether it is a bug?

Or, may be, these cases can't be easily handled by technical reasons (I know that Relacy is not verifier of C++ formal model / abstract machine)?

Thanks in advance!

[dvyukov]

Hi Aleh,

Yes, it's a bug (or not implemented feature).
Currently Relacy achieves non-sequentially consistent execution by only allowing loads to be satisfied by older/non-current stores. How to implement other types of relaxation is not obvious to me.