[Bug middle-end/63273] atomic operations lead to inefficient code

["dvyukov at google dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63273

--- Comment #2 from Dmitry Vyukov <dvyukov at google dot com> ---
> Which is because to me it's not exactly clear as for what other operations an
atomic load/store is a barrier for.

That's trivial to answer -- memory_order_relaxed is barrier for nothing.

> But eventually that's by design to make locking primitives work
(where we may not move any loads/stores into/out of the protected region).

That's a very bad design for a production compiler. Locking primitives will
necessary include memory_order_acquire/release atomic, and they will prevent
unwanted code movement. Relaxed atomic must not prevent any code movement and
also don't clobber anything. Similarly, memory_order_release should generally
not prevent hoisting of loads above it.

For the record, this is just fixed in clang compiler (did not test yet,
though):
http://llvm.org/bugs/show_bug.cgi?id=17281