[cpp-threads] pthreads (was: RE: C++ Connections proposal)

[Doug Lea]

Boehm, Hans wrote:
> I'd actually like to get more comments on this.  The more I think
> about this, the less I'm convinced that we should change the proposal.


I agree with both of these sentiments.


> 
> If I understand correctly, the current proposal, with this ABI,
> does require compiler changes, but it does not require an ABI
> change.  And the compiler changes generally involve generating
> something like the following for bit-field updates:
> 
> 1. Possibly read old value of misaligned field, as before.
> 2. Compute halves/thirds/whatever of new field value.
> 3. Store pieces back into only the field being updated.
> 
> This probably takes slightly more instructions than the old
> version.  But I wouldn't be surprised if it shortens the critical
> path in some cases, since you often no longer need the initial
> load.
> 
> Thus my guess would be that this is a small penalty that affects
> a very small portion of the code, where the alternative is to
> significantly complicate the programming model, and leave some
> very subtly broken client code broken.
> 

It would be great if the compiler developers now on this list
could hit us with their best shots about such issues now,
so we can proceed knowing exactly what the performance impact
will be.

-Doug