[cpp-threads] Alternatives to SC

[Peter Dimov]

Herb Sutter wrote:

>> Towards TSO, maybe.  In light of the example above, I have a hard
>> time believing they will really get all the way to SC.
>
> One reference I had in mind was:
>
>   "Multiprocessors Should Support Simple Memory Consistency Models"
>   Mark D. Hill, IEEE Computer, August 1998
>   http://citeseer.ist.psu.edu/hill98multiprocessors.html
>
> Mark argues that hardware should implement (not only enable) SC, "or
> in some cases" read reordering.

To which extent this also implies that our compilers need to be SC as well? 
To give you a concrete example:

// (B1)

atomic_fetchadd_raw( &x, 1 );

// (B2)

atomic_fetchadd_raw( &x, 1 );

Here B1 and B2 are blocks of ordinary operations that don't touch x. This 
can be transformed by the compiler to:

// (B1)
// (B2)

atomic_fetchadd_raw( &x, 2 );

which then may enable further optimizations due to B1 and B2 sharing a basic 
block. Imposing sequentially consistent (or CCCC, for that matter) semantics 
on x precludes this trasformation.

And if we agree that we need the option to be able to grant the compiler a 
license to do this sort of thing, why deny the hardware the opportunity to 
also perform similar optimizations? The end result for the program is 
exactly the same.