[PATCH v1] nptl: Replace CAS with cheaper atomics for pthread_cancel logic

[Wilco Dijkstra]

Hi Noah,

In general I agree with replacing any compare_exchange loops which can be done
using simpler atomics to make code clearer and possibly faster. However we have
to be careful, consider this:

   if ((cancelhandling & CANCELTYPE_BITMASK) == 0)
     {
-      int newval;
-      do
-	{
-	  newval = cancelhandling | CANCELTYPE_BITMASK;
-	}
-      while (!atomic_compare_exchange_weak_acquire (&self->cancelhandling,
-						    &cancelhandling, newval));
-
+      atomic_fetch_or_acquire (&self->cancelhandling, CANCELTYPE_BITMASK);
       if (cancel_enabled_and_canceled (cancelhandling))

This now uses an stale read of cancelhandling - this might be bad if it was modified
by another thread. It seems you want something like:

cancelhandling = atomic_fetch_or_acquire (&self->cancelhandling, CANCELTYPE_BITMASK); 
if (cancel_enabled_and_canceled (cancelhandling)

Using or_fetch may also be feasible. Also assuming we only need atomicity and no
synchronization of global data then these can become relaxed atomics which should
be a lot faster on most targets. We should consider removing some of the pre-reads
before atomics as well since this is a premature optimization and counterproductive
on modern cores.

Cheers,
Wilco