Re: [PATCH] [RFC] x86/cpu: rework instruction set selection

["H. Peter Anvin" <hpa@xxxxxxxxx>]

On April 26, 2025 2:08:17 AM PDT, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
>* Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
>
>> From: Arnd Bergmann <arnd@xxxxxxxx>
>> 
>> With cx8 and tsc being mandatory features, the only important
>> architectural features are now cmov and pae.
>> 
>> Change the large list of target CPUs to no longer pick the instruction set
>> itself but only the mtune= optimization level and in-kernel optimizations
>> that remain compatible with all cores.
>> 
>> The CONFIG_X86_CMOV instead becomes user-selectable and is now how
>> Kconfig picks between 586-class (Pentium, Pentium MMX, K6, C3, GeodeGX)
>> and 686-class (everything else) targets.
>> 
>> In order to allow running on late 32-bit cores (Athlon, Pentium-M,
>> Pentium 4, ...), the X86_L1_CACHE_SHIFT can no longer be set to anything
>> lower than 6 (i.e. 64 byte cache lines).
>> 
>> The optimization options now depend on X86_CMOV and X86_PAE instead
>> of the other way round, while other compile-time conditionals that
>> checked for MATOM/MGEODEGX1 instead now check for CPU_SUP_* options
>> that enable support for a particular CPU family.
>> 
>> Link: 
>> https://lore.kernel.org/lkml/dd29df0c-0b4f-44e6-b71b-2a358ea76fb4@xxxxxxxxxxxxxxxx/
>> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
>> ---
>> This is what I had in mind as mentioned in the earlier thread on
>> cx8/tsc removal. I based this on top of the Ingo's [RFC 15/15]
>> patch.
>> ---
>>  arch/x86/Kconfig                |   2 +-
>>  arch/x86/Kconfig.cpu            | 100 ++++++++++++++------------------
>>  arch/x86/Makefile_32.cpu        |  48 +++++++--------
>>  arch/x86/include/asm/vermagic.h |  36 +-----------
>>  arch/x86/kernel/tsc.c           |   2 +-
>>  arch/x86/xen/Kconfig            |   1 -
>>  drivers/misc/mei/Kconfig        |   2 +-
>>  7 files changed, 74 insertions(+), 117 deletions(-)
>
>While the simplification is nice on its face, this looks messy:
>
>> diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
>> index a9d717558972..1e33f88c9b97 100644
>> --- a/arch/x86/Kconfig
>> +++ b/arch/x86/Kconfig
>> @@ -1438,7 +1438,7 @@ config HIGHMEM
>>  
>>  config X86_PAE
>>      bool "PAE (Physical Address Extension) Support"
>> -    depends on X86_32 && X86_HAVE_PAE
>> +    depends on X86_32 && X86_CMOV
>
>Coupling CMOV to PAE ... :-/
>
>> +config X86_CMOV
>> +    bool "Require 686-class CMOV instructions" if X86_32
>> +    default y
>>      help
>> -      This is the processor type of your CPU. This information is
>> -      used for optimizing purposes. In order to compile a kernel
>> -      that can run on all supported x86 CPU types (albeit not
>> -      optimally fast), you can specify "586" here.
>> +      Most x86-32 processor implementations are compatible with
>> +      the the CMOV instruction originally added in the Pentium Pro,
>> +      and they perform much better when using it.
>> +
>> +      Disable this option to build for 586-class CPUs without this
>> +      instruction. This is only required for the original Intel
>> +      Pentium (P5, P54, P55), AMD K6/K6-II/K6-3D, Geode GX1 and Via
>> +      CyrixIII/C3 CPUs.
>
>Very few users will know anything about CMOV.
>
>I'd argue the right path forward is to just bite the bullet and remove 
>non-CMOV support as well, that would be the outcome *anyway* in a few 
>years. That would allow basically a single 'modern' 32-bit kernel that 
>is supposed to boot on every supported CPU. People might even end up 
>testing it ... ;-)
>
>Thanks,
>
>       Ingo

Dropping CMOV would mean dropping P5 support.