Revert "Fix fence on non-x86 arch and miri (#16)"

[taiki-e]

Based on the discussion in #17:

• Revert Fix fence on non-x86 arch and miri #16 for now (v1.2.3 is already yanked)
• Use compiler_fence in full_fence on x86
• Do not use x86 specific fence on Miri (because Miri doesn't understand it)

The fence for the unbounded queue needs to be fixed again in some way, but for now, I'm going to revert all of them.

@sbarral Any thoughts?

[sbarral]

@sbarral Any thoughts?

Nope, looks good to me :-)

[RalfJung]

This is still doing the fence before the load though, how does that make sense?

[sbarral]

This is still doing the fence before the load though, how does that make sense?

I do not know the full context of your reply (I understand this has been discussed someplace else) so sorry if my comment is irrelevant, but in general there certainly are legit use cases for SeqCst fences before a load. For instance:

// Thread A:
x.store(true, Ordering::relaxed);
fence(Ordering::SeqCst);
let y_val = y.load(Ordering::relaxed);

// Thread B:
y.store(true, Ordering::relaxed);
fence(Ordering::SeqCst);
let x_val = x.load(Ordering::relaxed);

will ensure that either the x store will be visible in thread B or the y store will be visible in thread A, or both.