Subj : Re: A question about atomic_ptr
To   : comp.programming.threads
From : Joe Seigh
Date : Thu Apr 14 2005 12:06 pm

On Thu, 14 Apr 2005 16:53:13 +0300, Peter Dimov <pdimov@gmail.com> wrote:

> Joe Seigh wrote:
>> The atomic swap is one part.  The other part is safely incrementing
>> the reference count without "owning" it first.  There's the DWCAS
>> solution, the PPC solution, and various GC based solutions using RCU,
>> SMR, etc... And of course the 2CAS solution which only works on
>> MC68020 and MC68030 processors.
>
> You mean the race between
>
>     atomic_ptr<T> p2( p1 );
>
> and
>
>     p1 = 0;
>
> ?

Yes.

>
> I think that you have a bug in your PPC implementation, but I may be missing
> something. In inc_ref, you reload the pointer and check for change, however,
> the pointer is a local function argument (int* refcnt) and will never
> change.
>
> And you are right that Reinholz uses the same algorithm, except he passes an
> int** (the offset hack aside).
>
>
You're right, that's a bug.  Good catch.  It should be int **.  I think I'll throw
in a extra temp register.  I don't trust GCC to form the memory operand
for that correctly.  As far as the offset goes, I hardcoded the offset
in my version since I have a non-intrusive refcount and I control the
format and there didn't seem to be an offset_of macro around.  It's
not really a finished version, just a pre-beta proof of concept so
there's some quick and dirty stuff in it still.


-- 
Joe Seigh

.