Oh, did the mb() only generate lock on Linux?
Then you should use x86_lfence() everywhere so
you get the lock-prefixed instructions.

	-- Lennart

Manuel Bouyer wrote:
> On Thu, Mar 24, 2005 at 05:14:49PM +0100, Lennart Augustsson wrote:
> 
>>The x86 can do some reordering, but nothing should ever
>>move across an instruction with a lock prefix.  So with
>>enough mb() in your code you should be totally safe.
> 
> 
> And is __insn_barrier() equivalent to x86_lfence() in NetBSD ?
> The generated assembly doesn't change when I use one or the other.
> Using mb() didn't change anything either.
> There's no "lock" in the generated assembly file, or nothing looking different
> from normal instructions.
>