Re: [PATCH] x86: Introduce ASM flags to bitops

[Ingo Molnar <mingo@kernel.org>]

* Uros Bizjak <ubizjak@gmail.com> wrote:

> On Mon, Jul 27, 2015 at 5:54 PM, Ingo Molnar <mingo@kernel.org> wrote:
> >
> > * Uros Bizjak <ubizjak@gmail.com> wrote:
> >
> >> On Mon, Jul 27, 2015 at 5:40 PM, Ingo Molnar <mingo@kernel.org> wrote:
> >> >
> >> > * Uros Bizjak <uros_bizjak1@t-2.net> wrote:
> >> >
> >> >> From: Uros Bizjak <ubizjak@gmail.com>
> >> >>
> >> >> This patch introduces GCC ASM flags to bitops. Instead of e.g.
> >> >>
> >> >>    136d7:     48 0f a3 3d 00 00 00    bt     %rdi,0x0(%rip)
> >> >>    136de:     00
> >> >>    136df:     19 ff                   sbb    %edi,%edi
> >> >>    136e1:     85 ff                   test   %edi,%edi
> >> >>    136e3:     0f 95 c0                setne  %al
> >> >>
> >> >> following code is generated:
> >> >>
> >> >>    13767:     48 0f a3 3d 00 00 00    bt     %rdi,0x0(%rip)
> >> >>    1376e:     00
> >> >>    1376f:     0f 92 c0                setb   %al
> >> >>
> >> >> Similar improvement can be seen in following code:
> >> >>
> >> >>     7a6c:     48 0f a3 11             bt     %rdx,(%rcx)
> >> >>     7a70:     19 d2                   sbb    %edx,%edx
> >> >>     7a72:     85 d2                   test   %edx,%edx
> >> >>     7a74:     74 eb                   je     7a61
> >> >>
> >> >> which becomes:
> >> >>
> >> >>     7a8c:     48 0f a3 11             bt     %rdx,(%rcx)
> >> >>     7a90:     73 ef                   jae    7a81
> >> >>
> >> >> Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
> >> >> ---
> >> >>  arch/x86/include/asm/bitops.h      | 26 ++++++++++++++++++++++++--
> >> >>  arch/x86/include/asm/percpu.h      | 18 +++++++++++++++++-
> >> >>  arch/x86/include/asm/signal.h      |  6 ++++++
> >> >>  arch/x86/include/asm/sync_bitops.h | 18 ++++++++++++++++++
> >> >>  4 files changed, 65 insertions(+), 3 deletions(-)
> >> >
> >> > Nothing in your patch seems to be setting __GCC_ASM_FLAG_OUTPUTS__, and the patch
> >> > does not seem to be mailed as part of a larger series ...
> >> >
> >> > So how is this supposed to work?
> >>
> >> GCC version 6+ will automatically define __GCC_ASM_FLAG_OUTPUTS__ when
> >> this feature is supported. Please see [1] for RFC GCC patch series and
> >> [2] for final committed patch.
> >>
> >> [1] https://gcc.gnu.org/ml/gcc-patches/2015-05/msg00594.html
> >> [2] https://gcc.gnu.org/ml/gcc-patches/2015-06/msg02087.html
> >
> > Ok, great. This information should be part of the changelog and such, as it's not
> > obvious.
> 
> No problem, I'll add this information and send a v2 patch.
> 
> > Does the GCC project treat this as an ABI kind of thing, i.e. can the kernel rely
> > on it from now on, without the GCC side semantics of this feature not ever
> > changing and breaking the kernel?
> 
> Yes. It was discussed and agreed between GCC and kernel people (HPA)
> on GCC mailing list. Please see [3].
> 
> [3] https://gcc.gnu.org/ml/gcc-patches/2015-05/msg00725.html

Ok, great!

One more request: please make it similar to how we handle other, similar compiler 
features and introduce an interim macro like:

 #ifdef __GCC_ASM_FLAG_OUTPUTS__
 # define CC_HAVE_ASM_FLAG_OUTPUTS
 #endif

in a suitable spot in include/linux/compiler-gcc.h, and use 
CC_HAVE_ASM_FLAG_OUTPUTS in the other places.

this makes it easier to disable/enable this feature centrally, and it might enable 
other compilers to introduce a similar feature.

Thanks,

	Ingo