Why does ia64 not use fetchadd in atomic.h?

Why does ia64 not use fetchadd in atomic.h?

[Christoph Lameter]

I just looked at at the asm generated by some of the use of atomic for 
statistics and saw that there were complicated compxchg constructs where I 
would have expected a simple fetchadd. Why is this?

F.e.

static __inline__ int
ia64_atomic_add (int i, atomic_t *v)
{
        __s32 old, new;
        CMPXCHG_BUGCHECK_DECL

        do {
                CMPXCHG_BUGCHECK(v);
                old = atomic_read(v);
                new = old + i;
        } while (ia64_cmpxchg(acq, v, old, new, sizeof(atomic_t)) != old);
        return new;
}

Why not

#define ia64_atomic_add(__i, __v) ia64_fetchadd(__i, &v->counter, acq)

?

Re: Why does ia64 not use fetchadd in atomic.h?

[Matthew Wilcox]

On Thu, Jun 15, 2006 at 01:35:42PM -0700, Christoph Lameter wrote:
> Why not
> 
> #define ia64_atomic_add(__i, __v) ia64_fetchadd(__i, &v->counter, acq)

Because ia64_fetchadd() only supports 8 different constants?

RE: Why does ia64 not use fetchadd in atomic.h?

[Chen, Kenneth W]

Christoph Lameter wrote on Thursday, June 15, 2006 1:36 PM
> I just looked at at the asm generated by some of the use of atomic for 
> statistics and saw that there were complicated compxchg constructs where I 
> would have expected a simple fetchadd. Why is this?
> 
> F.e.
> 
> static __inline__ int
> ia64_atomic_add (int i, atomic_t *v)
> {
>         __s32 old, new;
>         CMPXCHG_BUGCHECK_DECL
> 
>         do {
>                 CMPXCHG_BUGCHECK(v);
>                 old = atomic_read(v);
>                 new = old + i;
>         } while (ia64_cmpxchg(acq, v, old, new, sizeof(atomic_t)) != old);
>         return new;
> }
> 
> Why not
> 
> #define ia64_atomic_add(__i, __v) ia64_fetchadd(__i, &v->counter, acq)


The immediate operand of fetchadd only takes limited value, so it would have
to be a combination of fetchadd and cmpxchg.  If the constant is known at
compile time, it can be optimized for those constants that fetchadd can
operate on.

Back to your original question, I guess whoever implements ia64_atomic_add
didn't bother to optimize it like it's close cousin of atomic_add().

- Ken

Re: Why does ia64 not use fetchadd in atomic.h?

[Christoph Lameter]

On Thu, 15 Jun 2006, Matthew Wilcox wrote:

> On Thu, Jun 15, 2006 at 01:35:42PM -0700, Christoph Lameter wrote:
> > Why not
> > 
> > #define ia64_atomic_add(__i, __v) ia64_fetchadd(__i, &v->counter, acq)
> 
> Because ia64_fetchadd() only supports 8 different constants?

Right. I forgot about that. Only a call to atomic_add does the fetchadd 
trickery. Thanks.

RE: Why does ia64 not use fetchadd in atomic.h?

[Chen, Kenneth W]

Christoph Lameter wrote on Thursday, June 15, 2006 2:29 PM
> > On Thu, Jun 15, 2006 at 01:35:42PM -0700, Christoph Lameter wrote:
> > > Why not
> > > 
> > > #define ia64_atomic_add(__i, __v) ia64_fetchadd(__i, &v->counter, acq)
> > 
> > Because ia64_fetchadd() only supports 8 different constants?
> 
> Right. I forgot about that. Only a call to atomic_add does the fetchadd 
> trickery. Thanks.


I guess I got confused, are you complaining about ia64_atomic_add or
ia64_atomic64_add?  If the former, there is not much can be done, but if
the latter, yeah, it can be optimized for certain constant.

RE: Why does ia64 not use fetchadd in atomic.h?

[Christoph Lameter]

On Thu, 15 Jun 2006, Chen, Kenneth W wrote:

> I guess I got confused, are you complaining about ia64_atomic_add or
> ia64_atomic64_add?  If the former, there is not much can be done, but if
> the latter, yeah, it can be optimized for certain constant.

No complaints. I was just looking for an explanation.