Re: [PATCH] compiler: enable CONFIG_OPTIMIZE_INLINING forcibly

[Masahiro Yamada]

Hi Geert,

On Thu, Sep 26, 2019 at 6:26 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>
> Hi Yamada-san,
>
> On Thu, Sep 26, 2019 at 11:03 AM Masahiro Yamada
> <yamada.masahiro@xxxxxxxxxxxxx> wrote:
> > On Thu, Sep 26, 2019 at 5:54 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> > > On Fri, Aug 30, 2019 at 5:44 AM Masahiro Yamada
> > > <yamada.masahiro@xxxxxxxxxxxxx> wrote:
> > > > Commit 9012d011660e ("compiler: allow all arches to enable
> > > > CONFIG_OPTIMIZE_INLINING") allowed all architectures to enable
> > > > this option. A couple of build errors were reported by randconfig,
> > > > but all of them have been ironed out.
> > > >
> > > > Towards the goal of removing CONFIG_OPTIMIZE_INLINING entirely
> > > > (and it will simplify the 'inline' macro in compiler_types.h),
> > > > this commit changes it to always-on option. Going forward, the
> > > > compiler will always be allowed to not inline functions marked
> > > > 'inline'.
> > > >
> > > > This is not a problem for x86 since it has been long used by
> > > > arch/x86/configs/{x86_64,i386}_defconfig.
> > > >
> > > > I am keeping the config option just in case any problem crops up for
> > > > other architectures.
> > > >
> > > > The code clean-up will be done after confirming this is solid.
> > > >
> > > > Signed-off-by: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
> > >
> > > This breaks compiling drivers/video/fbdev/c2p*, as the functions in
> > > drivers/video/fbdev/c2p_core.h are no longer inlined, leading to calls
> > > to the non-existent function c2p_unsupported(), as reported by KISSKB:
> > >
> > > On Thu, Sep 26, 2019 at 5:02 AM <noreply@xxxxxxxxxxxxxx> wrote:
> > > > FAILED linux-next/m68k-defconfig/m68k Thu Sep 26, 12:58
> > > >
> > > > http://kisskb.ellerman.id.au/kisskb/buildresult/13973194/
> > > >
> > > > Commit:   Add linux-next specific files for 20190925
> > > >           d47175169c28eedd2cc2ab8c01f38764cb0269cc
> > > > Compiler: m68k-linux-gcc (GCC) 4.6.3 / GNU ld (GNU Binutils) 2.22
> > > >
> > > > Possible errors
> > > > ---------------
> > > >
> > > > c2p_planar.c:(.text+0xd6): undefined reference to `c2p_unsupported'
> > > > c2p_planar.c:(.text+0x1dc): undefined reference to `c2p_unsupported'
> > > > c2p_iplan2.c:(.text+0xc4): undefined reference to `c2p_unsupported'
> > > > c2p_iplan2.c:(.text+0x150): undefined reference to `c2p_unsupported'
> > > > make[1]: *** [Makefile:1074: vmlinux] Error 1
> > > > make: *** [Makefile:179: sub-make] Error 2
> > >
> > > I managed to reproduce this with gcc version 8.3.0 (Ubuntu
> > > 8.3.0-6ubuntu1~18.04.1) , and bisected the failure to commit
> > > 025f072e5823947c ("compiler: enable CONFIG_OPTIMIZE_INLINING forcibly") .
> > >
> > > Marking the functions __always_inline instead of inline fixes that.
> > > Shall I send a patch to do that?
> >
> >
> > Yes, please.
>
> OK, will do.
>
> > But you do not need to touch _transp() or comp().
> > Only functions that call c2p_unsupported().
>
> However, the inlining of these functions is very performance-sensitive too.
> So perhaps I should mark all of them __always_inline?


I want to leave as much compiler's freedom as possible,
and the optimization criteria depends on CONFIG option.

When CONFIG_CC_OPTIMIZE_FOR_SIZE is chosen,
the optimization should respect the code size over performance,
and the compiler may prefer not inlining it.


So, the basic idea is, use __always_inline
only when otherwise the builds would be broken.


Masahiro Yamada




> > BTW, c2p_unsupported() can be replaced with BUILD_BUG().
> > This will break the build earlier in case
> > it cannot be optimized out.
>
> Right. Will do, too.
>
> Thanks!
>
> Gr{oetje,eeting}s,
>
>                         Geert
>
> --
> Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
>
> In personal conversations with technical people, I call myself a hacker. But
> when I'm talking to journalists I just say "programmer" or something like that.
>                                 -- Linus Torvalds



-- 
Best Regards
Masahiro Yamada