cpu: enforce a memory barrier inside cpu_relax()

cpu: enforce a memory barrier inside cpu_relax()

Put a memory barrier inside cpu_relax() to make sure the compiler
doesn't cache variables used outside cpu_relax() in registers.

Suppose to have the following case (used in many device drivers):

static int done = false;

void irq_handler(void)
{
	done = true;
}

void wait_for_completion(void)
{
	while (done == false)
		cpu_relax();
}

If cpu_relax() is just considered a nop the compiler may decide to not
reload the value in the variable "done" from memory, causing a
neverending loop.

The presence of an explicit memory barrier fixes this case.

git-svn-id: https://src.develer.com/svnoss/bertos/trunk@4344 38d2e660-2303-0410-9eaa-f027e97ec537