Re: [tip: x86/build] x86/build: Remove stale cc-option checks

From: Nick Desaulniers
Date: Thu Aug 26 2021 - 14:55:55 EST


On Sun, Aug 22, 2021 at 1:31 AM Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
>
> * tip-bot2 for Nick Desaulniers <tip-bot2@xxxxxxxxxxxxx> wrote:
>
> > The following commit has been merged into the x86/build branch of tip:
> >
> > Commit-ID: 1463c2a27d59c69358ad1cbd869d3a8649695d8c
> > Gitweb: https://git.kernel.org/tip/1463c2a27d59c69358ad1cbd869d3a8649695d8c
> > Author: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
> > AuthorDate: Thu, 12 Aug 2021 11:38:48 -07:00
> > Committer: Borislav Petkov <bp@xxxxxxx>
> > CommitterDate: Sun, 15 Aug 2021 10:32:52 +02:00
> >
> > x86/build: Remove stale cc-option checks
> >
> > -mpreferred-stack-boundary= is specific to GCC, while -mstack-alignment=
> > is specific to Clang. Rather than test for this three times via
> > cc-option and __cc-option, rely on CONFIG_CC_IS_* from Kconfig.
> >
> > GCC did not support values less than 4 for -mpreferred-stack-boundary=
> > until GCC 7+. Change the cc-option test to check for a value of 2,
> > rather than 4.
>
> > --- a/arch/x86/Makefile
> > +++ b/arch/x86/Makefile
> > @@ -14,10 +14,13 @@ endif
> >
> > # For gcc stack alignment is specified with -mpreferred-stack-boundary,
> > # clang has the option -mstack-alignment for that purpose.
> > -ifneq ($(call cc-option, -mpreferred-stack-boundary=4),)
> > +ifdef CONFIG_CC_IS_GCC
> > +ifneq ($(call cc-option, -mpreferred-stack-boundary=2),)
> > cc_stack_align4 := -mpreferred-stack-boundary=2
> > cc_stack_align8 := -mpreferred-stack-boundary=3
> > -else ifneq ($(call cc-option, -mstack-alignment=16),)
> > +endif
> > +endif
> > +ifdef CONFIG_CC_IS_CLANG
> > cc_stack_align4 := -mstack-alignment=4
> > cc_stack_align8 := -mstack-alignment=8
>
> So I spent most of yesterday bisecting a hard to diagnose bug that looked
> like a GPU driver bug - but the bisect somewhat surprisingly ended up at
> this commit.

I'm genuinely sorry about that. Let me guess, GPF on SSE instruction
with stack based operand from AMDGPU? (I've seen that twice so far
related to these options.)

I see now what went wrong....
GCC only supports a 4B stack alignment for ***32b*** (or 16b) x86;
`-mpreferred-stack-boundary=2` will produce an error unless -m32 or
-m16 is set; but `-mpreferred-stack-boundary=2 -m32` has been long
supported. It's -mpreferred-stack-boundary=3 for -m64 that wasn't
supported until the gcc-7 release. So the cc-option test should
instead test -mpreferred-stack-boundary=3.

>
> Doing the partial revert below solves the regression - as the above hunk is
> not obviously an identity transformation. I have a pretty usual GCC 10.3.0
> build environment with nothing exotic.
>
> I amdended the commit with the partial revert in tip:x86/build.

No worries. I'll send a follow up.

--
Thanks,
~Nick Desaulniers