Re: [PATCH RFC] x86: enforce inlining for atomics

From: Ingo Molnar
Date: Wed Apr 22 2015 - 05:29:06 EST



* Markus Trippelsdorf <markus@xxxxxxxxxxxxxxx> wrote:

> On 2015.04.22 at 07:58 +0200, Hagen Paul Pfeifer wrote:
> > * Markus Trippelsdorf | 2015-04-22 07:24:41 [+0200]:
> > >I cannot reproduce this issue with my config with 4.8, 4.9 or 5. Could
> > >you please come up with a small testcase and open a gcc bug (with full
> > >gcc command line)?
> >
> > the attached config file should do the trick. Can you try it?
>
> It is definitely a -O2 vs. -Os issue. With your config:
>
> -Os:
> % nm vmlinux | grep " atomic_" | wc -l
> 736
>
> -O2 is fine:
> % nm vmlinux | grep " atomic_"
> ffffffff852d8300 r atomic_counters_ops
> ffffffff83d66bf0 t atomic_counters_read
> ffffffff8119cd30 T atomic_dec_and_mutex_lock
> ffffffff81172e10 T atomic_notifier_call_chain
> ffffffff81172b90 T atomic_notifier_chain_register
> ffffffff81172d70 T atomic_notifier_chain_unregister
> ffffffff844b0ef0 t atomic_read_file
> ffffffff85a0c260 r atomic_ro_fops
> ffffffff87616000 d atomic_rw
> ffffffff852d8400 r atomic_stats_ops
> ffffffff83d66c50 t atomic_stats_read
>
> The easiest fix would be to force always-inline in
> include/linux/compiler-gcc.h when CONFIG_CC_OPTIMIZE_FOR_SIZE is set.

That pretty much defeats the purpose of any inlining size calculations
on -Os - and would make it impossible for better or bugfixed compilers
to improve the situation.

So I think the original patch makes sense (and I already applied it),
we want known-simple and performance critical methods (such as atomic
ops) always inlined.

Thanks,

Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/