Can we do something like Java's lazySet in Haskell?

[symbiont-stevan-andjelkovic]

In the Java version they have one optimisation where they use lazySet which provides a store-store barrier rather than set which implies a full memory barrier.

atomic-primops exposes some memory barriers from the RTS, could they be used together with IORefs to achieve better performance than atomics (while retaining correctness)?