71b
Subj : Re: A question about atomic_ptr
To   : comp.programming.threads
From : Alexander Terekhov
Date : Mon Apr 11 2005 03:47 pm


Peter Dimov wrote:
> 
> Alexander Terekhov wrote:
> >>>>  if ( ! atomic_dec_acquire( &shared_ptr<X>::refs ) )
> >
> > That must be upcoming Apple's OSAtomicDecrementRelAcq32() or
> > OSAtomicDecrementSlbHsb32() (when X is immutable while managed
> > by shared_ptr<>). ;-)
> 
> Speaking of which, is it possible to implement the ...SlbHsb variant in a
> more efficient manner on a PPC? 

SlbHsb version doesn't seem to need trailing isync, and maybe can 
achieve the effect of leading {lw}sync with respect to loads->store 
ordering using "branch never taken" trick. Consider: (excerpt from 
Book II)

----
B.2.3 Safe Fetch

If a load must be performed before a subsequent store (e.g., the 
store that releases a lock protecting a shared data structure), a 
technique similar to the following can be used.

In this example it is assumed that the address of the storage 
operand to be loaded is in GPR 3, the contents of the storage 
operand are returned in GPR 4, and the address of the storage 
operand to be stored is in GPR 5.

  lwz r4,0(r3) #load shared data
  cmpw r4,r4 #set CR0 to "equal"
  bne- $-8 #branch never taken
  stw r7,0(r5) #store other shared data
----

In our case, we load and check the count and do take a conditional 
branch which leads to stores which we want to order with respect to 
load and check. That's trailing isync. As for leading {lw}sync, we 
can add similar "branch never taken" trick to order it with respect 
to preceding loads, but I'm not sure whether register dependency
(r4 for both load and compare above) can play some role here in 
addition to "branch never taken".

regards,
alexander.

.

0