Have you ever wondered why operating system kernels are compiled without full compiler optimizations turned on? The reason is that it’s impossible to implement proper threading in C/C++ in a library, because the execution model of C/C++ is single-threaded. If the compiler doesn’t know about threads, it might optimize a global variable access into a register access and if now another thread writes to that global variable we have a race. That’s just one example what can happen. A good introduction into the problem is given here (in German) or in the video Getting C++ Threads Right by Hans Boehm.

The next version of C++ called C++0x will include approaches to overcome this limitation. That would be a nice addition to the C standard as well, as C is used quite extensivly in systems programming where this matters even more.

If you are into high-performance number crunching then you should consider Unified Parallel C (UPC). It unifies both message passing (as used by MPI) and the shared memory model (as used by OpenMP).

What I am dreaming of is a combination of features from UPC (unified message passing/shared memory), Cyclone (abstract data types, region analysis, fat pointers), C (performance) and some concepts from D (modules, templates, closures). I’d call this language ASYL - Advanced Systems Language.