--- Comment #17 from Peter Cordes <peter at cordes dot ca> ---
(In reply to Jonathan Wakely from comment #16)
> But what we do care about is comment 2, i.e. _Atomic(T) and std::atomic<T>
> should have the same alignment (both in an out of structs). Maybe that needs
> the C front-end to change how _Atomic works, or maybe it needs the C++
> library to change how std::atomic works, but I want to keep this bug open
> while comment 2 gives different answers for C and C++.
Right, gcc's C _Atomic ABI is still broken for long long on 32-bit x86. It
only aligned _Atomic long long to 32 bits (inside structs), but then assumes
that 8-byte loads / stores (with x87 or SSE1/2) are atomic.
It also leads to abysmal performance for LOCK CMPXCHG or other RMW operations
if the atomic object is split across a cache line.
That's bug 65146, so we can close this one. (I never got around to posting in
the google group for the ABI. By far the best good solution is giving _Atomic
long long (and other 8-byte objects) a boost to their _Alignof, up to 8 byte
alignment even inside structs.)