[PATCH v2 3/5] include: Prefer GCC __atomic builtins if possible.

[Rémi Bernon]

On 1/15/21 5:46 PM, Zebediah Figura (she/her) wrote:
> On 1/15/21 10:34 AM, Alexandre Julliard wrote:
>> Zebediah Figura <z.figura12 at gmail.com> writes:
>>
>>> Signed-off-by: Zebediah Figura <z.figura12 at gmail.com>
>>> ---
>>> v2: avoid 16-bit __atomic_compare_exchange_n()
>>>
>>>   include/winnt.h | 53 +++++++++++++++++++++++++++++++++++++++++++++++++
>>>   1 file changed, 53 insertions(+)
>>
>> How are these better than the __sync ones in practice?
>>
> 
> I think the only real benefit is atomic exchange; it allows the compiler
> to maybe yield better code for x86 than our inline assembly (although it
> surely would never make a difference in practice), and presumably better
> code for other platforms than a compare-and-swap loop.
> 
> The actual practical benefit is probably negligible, but I figured we
> might as well use a proper compiler intrinsic for atomic swap if there
> was one available.
> 

__sync GCC builtins are supposedly legacy, but I believe they may also 
imply a slightly stronger memory model than __atomic.

For instance, notice the additional "dmb ish" on ARM64 for __sync, and 
the use of "ldaxr" instead of "ldxr" for __atomic here:

   https://gcc.godbolt.org/z/vq5Gzv

(Note it may also depend on the compiler, Clang generates the same code 
with "ldaxr" and no "dmb ish", so either GCC implementation is 
sub-optimal, or they have different semantics for their builtins).

On X86 of course this won't change anything, because the CPU model is 
already implying an even stronger model.
-- 
Rémi Bernon <rbernon at codeweavers.com>