Subj : Re: A question about atomic_ptr
To   : comp.programming.threads
From : Chris Thomasson
Date : Sun Apr 10 2005 08:52 pm

> I was thinking DWCAS or mutex pool... I'm new to these lock-free things.

I would go for a strict pthread solution "first"; Quickly address the 
portability factor. I guess you could simply hash the address of a 
shared_ptr reference count into an index of an array of pthread_mutex_t's. 
There is a simple solution for this:

http://msdn.microsoft.com/msdnmag/issues/01/08/Concur/

You can take advantage of the fact that the mutex table would be never be 
destroyed.

DWCAS solutions will work fine, just remember that once you use DWCAS with 
pointers you fall into the portability trap. Most 64-bit processors do not 
have DWCAS.




> Need to read about SMR, I guess.

SMR needs a rather extensive underlying support system in order to be 
useful. You need to heavily optimize the allocation scheme that provides the 
nodes you use for your collections and/or reference counts. Some key SMR 
API's need to be coded in assembly language; they are too sensitive to an 
over aggressive optimizing compiler. Also, you need to examine the method in 
which SMR utilizes Thread-Local-Storage. My AppCore library has an efficient 
hazard pointer scheme that you might want to take a quick look at. Any 
questions are welcome...

.