Subj : Re: double-checked locking in C
To   : comp.programming.threads
From : Laurent Deniau
Date : Thu Jul 07 2005 02:55 pm

Giancarlo Niccolai wrote:
> Laurent Deniau wrote:
> 
> 
>>David Hopwood wrote:
> 
> 
>>>You're mistaken; processors can reorder memory accesses across jumps.
>>
>>I was abusively talking of a jump while in fact the goal was to have a
>>call. I don't know if processor can do tentative evaluation of a call,
>>but the value it has to assigned is returned by value from the call.
> 
> Actually, in parallel architectures it is possible to perform tentative
> evaluation of calls. 
> 
> However, I would not suggest to rely on such idioms in C, and more notably
> in C++, as the compiler may even decide on its own to inline calls, or your

I guess that you talk about the linker optimizer. The compiler sees only 
one translation unit at the same time.

> programmers team may turn a function into a macro/inline later on, and you

I was explicitly talking of function with *external* linkage.

> would end up with a bug that would be deadly difficult to track down.

Clear if it happens.

> Please notice that I know that what I am saying here is not directly
> related with the efficacy of this method, but it's however worth to be
> considered when choosing to use this idiom instead of others.

As far as I know, there is only three idioms:

DCL (not portable)
pthread_once (fast)
mutex (slow)

When I say slow I am talking of a factor 20 comparing to the same 
semantic without threads. This is a real bottleneck.

Do you know portable (C+POSIX) alternatives?

Regards,

ld.

.