Subj : Re: Adding thread support for C...
To   : comp.programming.threads
From : Mayan Moudgill
Date : Fri Jul 22 2005 04:10 pm

Joe Seigh wrote:
> Mayan Moudgill wrote:
> 
>>
>> I was wondering what features people wish "C" had that would make life 
>> easier. [Obviously, we're talking about extensions to C, not standard C].
>>
>> For instance, we could add a statement __barrier() {don't reorder 
>> memory operations past this statement}, or a non-stack scope variable 
>> qualifier __thread {this variable is thread private}.
>>
>> Anything else people wish they had in the language?
>>
> 
> Actually, we're probably better off without it.  Once they put thread 
> support
> in C we'll be in the same situation Java is in where you can't add anything
> without adding support directly in Java itself, e.g. JSR 166.   Where do 
> you
> think the stuff they added in JSR 166 came from?
> 

Interesting.

Our original motivation for extensions came from the desire to write 
various atomic operations in C. We wanted to do:
   void
   atomic_increment(
	volatile int * p
	)
   {
     do {
       int x = __load_locked(p);
       x++;
     } while( !__store_conditional(x,p);
   }
[We've already added two extensions, which can be modelled as functions].

We found that the compiler was:
a) inlining functions of this form.
b) then moving code around in somewat unpleasent ways.

We introducted the __barrier(); "call" to inhibit the compiler from 
doing this. [__load_locked/__store_conditional are also __barriers 
before/after the operation].

Another problem came from spin-lock wait loops.
   void
   spin(unsigned n)
   {
     unsigned i;
     for( i = 0; i < n; i++ ) {
     }
   }
The compiler was eliminating the wait-loops [as it should]. We had to 
introduce a mechanism to inhibit this optimization.

It turns out that, as the compiler starts optimizing more, and worse, 
doing whole program optimizations, it becomes critical with 
multi-threaded programs to make sure that the compiler can distinguish 
between optimizations that would be correct in single-threaded C but are 
invalid in multi-threaded codes.

.