Re: Surprising code generated for vdso_read_begin()

[Arnd Bergmann <arnd@arndb.de>]

On Wed, Feb 19, 2020 at 9:45 AM Christophe Leroy
<christophe.leroy@c-s.fr> wrote:
> Le 16/02/2020 à 19:10, Arnd Bergmann a écrit :
> > On Sat, Jan 11, 2020 at 12:33 PM Segher Boessenkool
> > <segher@kernel.crashing.org> wrote:
> >>
> >> On Fri, Jan 10, 2020 at 07:45:44AM +0100, Christophe Leroy wrote:
> >>> Le 09/01/2020 à 21:07, Segher Boessenkool a écrit :
> >>>> It looks like the compiler did loop peeling.  What GCC version is this?
> >>>> Please try current trunk (to become GCC 10), or at least GCC 9?
> >>>
> >>> It is with GCC 5.5
> >>>
> >>> https://mirrors.edge.kernel.org/pub/tools/crosstool/ doesn't have more
> >>> recent than 8.1
> >>
> >> Arnd, can you update the tools?  We are at 8.3 and 9.2 now :-)  Or is
> >> this hard and/or painful to do?
> >
> > To follow up on this older thread, I have now uploaded 6.5, 7.5, 8.3 and 9.2
> > binaries, as well as a recent 10.0 snapshot.
> >
>
> Thanks Arnd,
>
> I have built the VDSO with 9.2, I get less performant result than with
> 8.2 (same performance as with 5.5).
>
> After a quick look, I see:
> - Irrelevant NOPs to align loops and stuff, allthough -mpcu=860 should
> avoid that.
> - A stack frame is set for saving r31 in __c_kernel_clock_gettime. GCC
> 8.1 don't need that, all VDSO functions are frameless with 8.1

If you think it should be fixed in gcc, maybe try to reproduce it in
https://godbolt.org/ and open a gcc bug against that.

Also, please try the gcc-10 snapshot, which has the highest chance
of getting fixes if it shows the same issue (or worse).

     Arnd