[openib-general] Re: [PATCH] opensm: make cl_atomic functions atomic

[Jean-Christophe Hugly]

On Mon, 2006-03-06 at 19:29 +0200, Michael S. Tsirkin wrote:

> The right thing to do is to look at the implementation
> and write your own for architectures that matter.
> Use lock for others.

If performance at that point is *that* critical...

I used the following in the past for x86. Tested on 32bit only.

That's my own code, not my employers. So, I donate it, fwiw.

inline void AtomicInc(int volatile& x) {
    __asm__ volatile ("lock\nincl %0" : : "m" (x));
}

inline void AtomicDec(int volatile& x) {
    __asm__ volatile ("lock\ndecl %0" : : "m" (x));
}


-- 
Jean-Christophe Hugly <jice at pantasys.com>
PANTA