Potential issue with --enable-fat

[Marc Glisse]

On Mon, 10 Feb 2025, Niels Möller wrote:

> Marc Glisse <marc.glisse at inria.fr> writes:
>
>> Let's see. If I configure GMP with just --enable-fat, I still see in
>> config.h
>>
>> #define HAVE_HOST_CPU_nehalem 1
>>
>> and longlong.h enables mulx based on
>>
>> #if X86_ASM_MULX \
>>    && (HAVE_HOST_CPU_haswell || HAVE_HOST_CPU_broadwell \
>>        || HAVE_HOST_CPU_skylake || HAVE_HOST_CPU_bd4 || HAVE_HOST_CPU_zen)
>
> Torbjörn Granlund <tg at gmplib.org> writes:
>
>> Emmanuel Lonca <emmanuel.lonca at univ-artois.fr> writes:
>>
>>   To what I understand from the GMP documentation, compiling with the
>>   --enable-fat flag should produce a library that can be used on any x86
>>   architecture.
>>
>> Yes, that's the idea IIRC.
>
> It seems the above use in longlong.h is the only use of X86_ASM_MULX?
> Then it seems right to completely disable that in fat builds, like
>
> diff -r 1a2ad0e32507 configure.ac
> --- a/configure.ac	Mon Feb 03 17:52:54 2025 +0100
> +++ b/configure.ac	Mon Feb 10 16:21:35 2025 +0100
> @@ -3843,7 +3843,9 @@ yes
>     X86_PATTERN | X86_64_PATTERN)
>       GMP_ASM_ALIGN_FILL_0x90
>       if test "$x86_have_mulx" = yes; then
> -        GMP_ASM_X86_MULX
> +        if test "$enable_fat" = no; then
> +	  GMP_ASM_X86_MULX
> +	fi
>       fi
>       GMP_ASM_COFF_TYPE
>       case $ABI in

I don't like it, GMP_ASM_X86_MULX is about the assembler supporting the 
mulx instruction, not about the host processor being able to execute it.

Something simple might be in configure.ac to hide

tmp_host=`echo $host_cpu | sed 's/\./_/'`
AC_DEFINE_UNQUOTED(HAVE_HOST_CPU_$tmp_host)
GMP_DEFINE_RAW("define_not_for_expansion(\`HAVE_HOST_CPU_$tmp_host')", POST)

behind a test $enable_fat = no

since any use of HAVE_HOST_CPU_* (as opposed to HAVE_HOST_CPU_FAMILY_*) is 
suspect in a fat build, but that prevents from having a fat build with a 
non-basic baseline (say only support skylake or later).

I feel that somehow, with --enable-fat, we should avoid GMP's config.guess 
and stick to the FSF version, but that doesn't look convenient.

-- 
Marc Glisse