Re: [PATCH 7/7] x86: simplify 32-bit instruction set selection

["H. Peter Anvin" <hpa@zytor.com>]

On November 5, 2026 2:01:40 PM PST, Arnd Bergmann <arnd@arndb.de> wrote:
>On Thu, Nov 5, 2026, at 21:17, H. Peter Anvin wrote:
>> On November 5, 2026 8:34:38 AM PST, Arnd Bergmann <arnd@kernel.org> wrote:
>>>From: Arnd Bergmann <arnd@arndb.de>
>>>
>>>The -march= compiler flags select both instruction set and tuning for
>>>a particular CPU, which works well when building a kernel that can only
>>>run on this one type, but has some counterintuitive effects because it
>>>is not always clear which models are compatible.
>>>
>>>E.g. building for Geode LX results in a kernel that can work on all on
>>>all 686-class CPUs but crashes on CPUs without the CMOV instructions.
>>>Building a kernel for 686 could work on Geode LX and Crusoe but fails
>>>because of the CPU generation check that detects these as 585-class.
>...
>>
>> Crusoe (TM3xxx/TM5xxx) do have a handful of compiler/user space visible 
>> items that are missing from i686, as I recall notably the long nop 
>> instructions and *possibly* fcomi (I *think* we caught that one, but 
>> I'm not 100% sure.)
>>
>> As a result, it reports correctly as family 5.
>
>I don't know about fcomi, but the P6 long NOP hasn't been used
>in 32-bit kernels since your 2008 commit 14469a8dd236 ("x86:
>disable static NOPLs on 32 bits"). From that point, I believe
>we effectively treated Geode LX and Crusoe as 686-compatible.
>
>binutils apparently uses the same definition and emits long
>nops for -march=pentiumpro but not for -march=i686 in modern
>versions.
>
>You are of course correct that this remains inconsistent:
>a kernel with CONFIG_MCRUSOE is still built with -march=i686
>and CONFIG_X86_MINIMUM_CPU_FAMILY=5, so it will fail to run
>on 586-class CPUs without CMOV.
>
>I'll fold in this patch to make a MCRUSOE kernel actually compatible
>with all i586-class CPUs, unless you have a better idea:
>
>diff --git a/arch/x86/Makefile_32.cpu b/arch/x86/Makefile_32.cpu
>index 1735a5f95d33..9a3a1b1c2452 100644
>--- a/arch/x86/Makefile_32.cpu
>+++ b/arch/x86/Makefile_32.cpu
>@@ -21,7 +21,7 @@ cflags-$(CONFIG_MK6)		+= -march=i586 $(call tune,k6)
> # Please note, that patches that add -march=athlon-xp and friends are pointless.
> # They make zero difference whatsosever to performance at this time.
> cflags-$(CONFIG_MK7)		+= -march=i686 $(call tune,athlon)
>-cflags-$(CONFIG_MCRUSOE)	+= -march=i686 $(align)
>+cflags-$(CONFIG_MCRUSOE)	+= -march=i586 $(call tune,i686) $(align)
> cflags-$(CONFIG_MEFFICEON)	+= -march=i686 $(call tune,pentium3) $(align)
> cflags-$(CONFIG_MCYRIXIII)	+= -march=i586 $(call tune,i486) $(align)
> cflags-$(CONFIG_MVIAC3_2)	+= -march=i686
>
>     Arnd

Yeah, the problem isn't the kernel, but user space.