Atomic swap

[Daniel Eischen]

On Thu, 7 Aug 2003, Andrew Gallatin wrote:
> 
> Daniel Eischen writes:
>  > [ I'm not subscribed to alpha@; please keep me on the CC ]
>  > 
>  > I need an atomic swap function for libpthread.  Here's my hack
>  > of an implementation:
>  > 
>  > /*
>  >  * Atomic swap:
>  >  *   Atomic (tmp = *dst, *dst = val), then *res = tmp
>  >  *
>  >  * void atomic_swap_long(long *dst, long val, long *res);
>  >  */
>  > static __inline 
>  > void atomic_swap_long(volatile long *dst, long val, long *res)
>  > {
>  > 	u_int64_t result;
>  > 
>  > 	__asm __volatile (
>  > 		"1:\tldq_l %0,%1\n\t"
>  > 		"stq_c %2,%1\n\t"
>  > 		"beq %2,2f\n\t"		/* Why is this beq instead of bne 1b? */
>  > 		"br 3f\n"
>  > 		"2:\tbr 1b\n"
>  > 		"3:\n"
>  > 		: "=&r" (result)
>  > 		: "m" (*dst), "r" (val)
>  > 		: "memory");
>  > 
>  > 	*res = result;
>  > }
>  > 
>  > As annotated above, there seems to be one more branch than
>  > necessary.
> 
> Its actually an optimization.   Alphas predict that backward branches
> will always be taken (think loops).   If you were to branch directly
> back to 1:, then if the store succeeds (which it nearly always
> should), then the cpu would have been betting on taking the branch,
> and that would slow things down.

OK.

> 
>  > Can someone look this over for me?  I really don't quite
>  > know what I'm doing when it comes to inline assembly.
> 
> I think it looks OK, but I'm also terrible at inline asm.

Yeah, me too.  It took me quite a few tries to hit upon
something that seemed to work.

-- 
Dan Eischen