basic_string atomicity

[Andrew MacLeod]

On 01/08/2012 07:50 AM, Gabriel Dos Reis wrote:
> On Fri, Jan 6, 2012 at 8:12 PM, Jonathan Wakely<jwakely.gcc@gmail.com>  wrote:
>> On 6 January 2012 14:34, David Edelsohn wrote:
>>> Unfortunately the changes to __sync_fetch_and_add because of the C++11
>>> memory model changes to GCC ALREADY HAVE CAUSED A REGRESSION that
>>> affects libstdc++.  G++ now produces worse code, at least for
>>> architectures that can make a distinction among the various memory
>>> models.
>> We probably want to CC Andrew and Aldy here, as I can't speak for the
>> other libstdc++ maintainers but personally I usually try not to touch
>> the existing code using atomic ops unless they're broken, and even
>> then I prefer to ask an expert!
>>
>> Andrew, Aldy, please see
>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51766 for the sync vs
>> lwsync change, and the thread starting at
>> http://gcc.gnu.org/ml/libstdc++/2012-01/msg00037.html
> I agree that we need to change std::string implementation for C++11
> to move away from refcounting.
>
> I also think that we need to unbreak libstdc++ on powerpc.
> Hence I am leaning towards David's patch.  The real question
> is whether it does harm on other architectures that are less
> "fine grained" than power.
Thats the danger.  the __sync routines were suppose to be seq-cst unless 
otherwise specified, and I believe the fetch_and_add family in 
particular should have always been treated as seq-cst...

We can change libstdc++ to use a less stringent memory model, but we'd 
have to be sure that its correct for *all* platforms... ie, that 
'exchange_and_add' is really acquire/release and not seq-cst sometimes. 
     I expect for the next release to go through everything and remove 
all references to the old __sync routines and replace them with __atomic 
calls.

Im not really familiar with the atomicity.h stuff,  bkoz has been taking 
care of that, maybe he can chip in here on what the model expected for 
exchange_and_add and others really should be.

In any case, the patch shouldn't be using numbers for the memory model, 
they should be using the CPP predefines like
    __atomic_fetch_sub(__mem, __val, __ATOMIC_RELEASE);

Andrew