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

From: Hagen Paul Pfeifer
Date: Tue Apr 21 2015 - 06:56:59 EST


On 21 April 2015 at 09:42, Ingo Molnar <mingo@xxxxxxxxxx> wrote:

Hey Ingo, Peter, Boris,

the environment is Debian with gcc version 4.9.2. I tried to reproduce
the results under Ubuntu 14.04 but their version (4.8.2!) seems fine:
at max one duplicate for each atomic_* function.

> So the thing is that allyesconfig turns on -Os:
>
> CONFIG_CC_OPTIMIZE_FOR_SIZE=y
>
> which is known to make bad decisions in other areas as well ...

I can recompile with "CONFIG_CC_OPTIMIZE_FOR_SIZE=n" and check the
results again!?

> If -Os does such bad inlining decisions (and the inlining examples you
> cited are horrible!) then I guess a lot of the other 'inline'
> functions are handled by it badly as well.

Assumption is correct - I see duplicates all over the place.

> I'm not sure we should start fighting the compiler: if a compiler does
> not take 'inline' seriously then the solution is to use another
> compiler, or at least to use different compiler flags.
>
> If inlining decisions are bad even with saner compiler options then we
> can use __always_inline, and we actually do that for locking
> primitives and some other low level primitives: which are typically
> larger than atomics, so even reasonable compilers might uninline them.

Probably we should check this on a wider gcc front! If the behavior is an
exception, or is this common behavior?

Hagen
--
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/