Re: [PATCH] x86: Fix detection of GCC -mpreferred-stack-boundary support

[Ingo Molnar <mingo@kernel.org>]

* Andy Lutomirski <luto@kernel.org> wrote:

> As per https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383, GCC only
> allows -mpreferred-stack-boundary=3 on x86_64 if -mno-sse is set.
> That means that cc-option will not detect
> -mpreferred-stack-boundary=3 support, because we test for it before
> setting -mno-sse.
> 
> Fix it by reordering the Makefile bits.
> 
> Compile-tested only.  This should help avoid code generation issues
> such as the one that was worked around in b96fecbfa8c8 ("x86/fpu:
> Fix boot crash in the early FPU code").
> 
> I'm a bit concerned that we could still have problems on older GCC
> versions given that our asm code does not respect GCC's idea of the
> ABI-required stack alignment.
> 
> Cc: Linus Torvalds <torvalds@linux-foundation.org>
> Cc: Jan Kara <jack@suse.cz>
> Cc: Ingo Molnar <mingo@kernel.org>
> Cc: Borislav Petkov <bp@alien8.de>
> Signed-off-by: Andy Lutomirski <luto@kernel.org>
> ---
>  arch/x86/Makefile | 9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/x86/Makefile b/arch/x86/Makefile
> index 118e6debc483..344dd2110b2a 100644
> --- a/arch/x86/Makefile
> +++ b/arch/x86/Makefile
> @@ -39,6 +39,12 @@ ifdef CONFIG_X86_NEED_RELOCS
>          LDFLAGS_vmlinux := --emit-relocs
>  endif
>  
> +# prevent gcc from generating any FP code by mistake
> +# This must be before we try -mpreferred-stack-boundary -- see
> +# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53383
> +KBUILD_CFLAGS += -mno-sse -mno-mmx -mno-sse2 -mno-3dnow
> +KBUILD_CFLAGS += $(call cc-option,-mno-avx,)
> +

So the 'stack boundary' is the RSP that GCC generates before it calls another 
function from within an existing function, right?

So looking at this I question the choice of -mpreferred-stack-boundary=3. Why not 
do -mpreferred-stack-boundary=2?

My reasoning: on modern uarchs there's no penalty for 32-bit misalignment of 
64-bit variables, only if they cross 64-byte cache lines, which should be rare 
with a chance of 1:16. This small penalty (of at most +1 cycle in some 
circumstances IIRC) should be more than counterbalanced by the compression of the 
stack by 5% on average.

... using stack-boundary=1 or stack-boundary=0 would probably be 
counterproductive, as these more exotic misalignments get treated progressively 
worse by x86 CPUs.

... but I have not measured any of this and even the 5% is just a possibly 
overoptimistic guess.

Thanks,

	Ingo