sync/atomic: make load and store fully sequentially consistent on s390x

[mundaym]

As discussed in https://groups.google.com/forum/#!topic/golang-dev/KKfRLcTeb0E loads and stores are not fully sequentially consistent on s390x since loads may be moved before stores if they access independent memory locations. This is similar to the memory model used by amd64.

We can force full sequential consistency of memory accesses by using fast serialization instructions (SYNC in Go asm). From discussion in #5045 it seems it is probably worth adding serialization between loads and stores in the sync/atomic package at least.

As far as I can tell the runtime doesn't currently require full sequential consistency, so I'm inclined to leave the atomic implementations as they are until a requirement to change them arises. (Does this sound reasonable @aclements?)

[aclements]

As far as I can tell the runtime doesn't currently require full sequential consistency

I don't know that the runtime requires sequential consistency, but we've always assumed that atomics are sequentially consistent, so I strongly suspect there are parts that do require sequential consistency. We explicitly use atomic.LoadAcq/atomic.StoreRel in the runtime where we've considered it worth the engineering and maintenance effort to reason about using weaker atomics. On a TSO-like model, those can just be regular loads and stores (with the appropriate compiler barriers).

But this isn't just about the runtime. Go atomics on all other platforms are sequentially consistent. That may be not be documented, but there is certainly code out there that relies on that.

So, we do need to fix this.

[gopherbot]

Change https://golang.org/cl/180439 mentions this issue: cmd/compile, runtime: make atomic loads/stores sequentially consistent on s390x