Re: [PATCH] x86-64: Exclude FMA4 IFUNC functions for -mapxf

[Sunil Pandey <skpgkp2@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 6516 bytes --]

On Thu, Apr 4, 2024 at 3:44 PM H.J. Lu <hjl.tools@gmail.com> wrote:

> When -mapxf is used to build glibc, the resulting glibc will never run
> on FMA4 machines.  Exclude FMA4 IFUNC functions when -mapxf is used.
> This requires GCC which defines __APX_F__ for -mapxf with commit:
>
> 1df56719bd8 x86: Define __APX_F__ for -mapxf
> ---
>  config.h.in                                   |  3 ++
>  sysdeps/x86_64/configure                      | 32 +++++++++++++++++++
>  sysdeps/x86_64/configure.ac                   | 18 +++++++++++
>  sysdeps/x86_64/fpu/multiarch/Makefile         | 22 +++++++------
>  sysdeps/x86_64/fpu/multiarch/ifunc-avx-fma4.h |  2 ++
>  sysdeps/x86_64/fpu/multiarch/ifunc-fma4.h     |  2 ++
>  6 files changed, 70 insertions(+), 9 deletions(-)
>
> diff --git a/config.h.in b/config.h.in
> index 4108d9d460..7b008420dc 100644
> --- a/config.h.in
> +++ b/config.h.in
> @@ -298,4 +298,7 @@
>  /* Define if -mmovbe is enabled by default on x86.  */
>  #undef HAVE_X86_MOVBE
>
> +/* Define if -mapxf is enabled by default on x86.  */
> +#undef HAVE_X86_APX
> +
>  #endif
> diff --git a/sysdeps/x86_64/configure b/sysdeps/x86_64/configure
> index 04a534fa12..07bdd40a37 100755
> --- a/sysdeps/x86_64/configure
> +++ b/sysdeps/x86_64/configure
> @@ -162,6 +162,38 @@ printf "%s\n" "$libc_cv_x86_have_amx_tile" >&6; }
>  config_vars="$config_vars
>  have-mamx-tile = $libc_cv_x86_have_amx_tile"
>
> +# Check if -mapxf is enabled.
> +{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking whether -mapxf is
> enabled" >&5
> +printf %s "checking whether -mapxf is enabled... " >&6; }
> +if test ${libc_cv_x86_have_apx+y}
> +then :
> +  printf %s "(cached) " >&6
> +else $as_nop
> +  cat > conftest.c <<EOF
> +#ifndef __APX_F__
> +# error APX isn't enabled
> +#endif
> +EOF
> +              libc_cv_x86_have_apx=no
> +              if { ac_try='${CC-cc} -c $CFLAGS conftest.c -o conftest.o
> 1>&5'
> +  { { eval echo "\"\$as_me\":${as_lineno-$LINENO}: \"$ac_try\""; } >&5
> +  (eval $ac_try) 2>&5
> +  ac_status=$?
> +  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
> +  test $ac_status = 0; }; }; then
> +                  libc_cv_x86_have_apx=yes
> +              fi
> +              rm -rf conftest*
> +fi
> +{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result:
> $libc_cv_x86_have_apx" >&5
> +printf "%s\n" "$libc_cv_x86_have_apx" >&6; }
> +if test $libc_cv_x86_have_apx = yes; then
> +  printf "%s\n" "#define HAVE_X86_APX 1" >>confdefs.h
> +
> +fi
> +config_vars="$config_vars
> +have-x86-apx = $libc_cv_x86_have_apx"
> +
>  test -n "$critic_missing" && as_fn_error $? "
>  *** $critic_missing" "$LINENO" 5
>
> diff --git a/sysdeps/x86_64/configure.ac b/sysdeps/x86_64/configure.ac
> index c714c47351..c7b68544a2 100644
> --- a/sysdeps/x86_64/configure.ac
> +++ b/sysdeps/x86_64/configure.ac
> @@ -76,5 +76,23 @@ EOF
>                rm -rf conftest*])
>  LIBC_CONFIG_VAR([have-mamx-tile], [$libc_cv_x86_have_amx_tile])
>
> +# Check if -mapxf is enabled.
> +AC_CACHE_CHECK(whether -mapxf is enabled,
> +              libc_cv_x86_have_apx, [dnl
> +cat > conftest.c <<EOF
> +#ifndef __APX_F__
> +# error APX isn't enabled
> +#endif
> +EOF
> +              libc_cv_x86_have_apx=no
> +              if AC_TRY_COMMAND(${CC-cc} -c $CFLAGS conftest.c -o
> conftest.o 1>&AS_MESSAGE_LOG_FD); then
> +                  libc_cv_x86_have_apx=yes
> +              fi
> +              rm -rf conftest*])
> +if test $libc_cv_x86_have_apx = yes; then
> +  AC_DEFINE(HAVE_X86_APX)
> +fi
> +LIBC_CONFIG_VAR([have-x86-apx], [$libc_cv_x86_have_apx])
> +
>  test -n "$critic_missing" && AC_MSG_ERROR([
>  *** $critic_missing])
> diff --git a/sysdeps/x86_64/fpu/multiarch/Makefile
> b/sysdeps/x86_64/fpu/multiarch/Makefile
> index 6ddd50240c..cbe09d49f4 100644
> --- a/sysdeps/x86_64/fpu/multiarch/Makefile
> +++ b/sysdeps/x86_64/fpu/multiarch/Makefile
> @@ -38,29 +38,36 @@ libm-sysdep_routines += \
>    s_truncf-avx \
>  # libm-sysdep_routines
>  else
> +ifeq (no,$(have-x86-apx))
>  libm-sysdep_routines += \
> -  e_asin-fma \
>    e_asin-fma4 \
> +  e_atan2-fma4 \
> +  e_exp-fma4 \
> +  e_log-fma4 \
> +  e_pow-fma4 \
> +  s_atan-fma4 \
> +  s_sin-fma4 \
> +  s_sincos-fma4 \
> +  s_tan-fma4 \
> +# libm-sysdep_routines
> +endif
> +libm-sysdep_routines += \
> +  e_asin-fma \
>    e_atan2-avx \
>    e_atan2-fma \
> -  e_atan2-fma4 \
>    e_exp-avx \
>    e_exp-fma \
> -  e_exp-fma4 \
>    e_exp2f-fma \
>    e_expf-fma \
>    e_log-avx \
>    e_log-fma \
> -  e_log-fma4 \
>    e_log2-fma \
>    e_log2f-fma \
>    e_logf-fma \
>    e_pow-fma \
> -  e_pow-fma4 \
>    e_powf-fma \
>    s_atan-avx \
>    s_atan-fma \
> -  s_atan-fma4 \
>    s_ceil-sse4_1 \
>    s_ceilf-sse4_1 \
>    s_cosf-fma \
> @@ -77,17 +84,14 @@ libm-sysdep_routines += \
>    s_roundevenf-sse4_1 \
>    s_sin-avx \
>    s_sin-fma \
> -  s_sin-fma4 \
>    s_sincos-avx \
>    s_sincos-fma \
> -  s_sincos-fma4 \
>    s_sincosf-fma \
>    s_sincosf-sse2 \
>    s_sinf-fma \
>    s_sinf-sse2 \
>    s_tan-avx \
>    s_tan-fma \
> -  s_tan-fma4 \
>    s_trunc-sse4_1 \
>    s_truncf-sse4_1 \
>  # libm-sysdep_routines
> diff --git a/sysdeps/x86_64/fpu/multiarch/ifunc-avx-fma4.h
> b/sysdeps/x86_64/fpu/multiarch/ifunc-avx-fma4.h
> index 7719188888..d126cf9cd5 100644
> --- a/sysdeps/x86_64/fpu/multiarch/ifunc-avx-fma4.h
> +++ b/sysdeps/x86_64/fpu/multiarch/ifunc-avx-fma4.h
> @@ -33,8 +33,10 @@ IFUNC_SELECTOR (void)
>        && CPU_FEATURE_USABLE_P (cpu_features, AVX2))
>      return OPTIMIZE (fma);
>
> +#ifndef HAVE_X86_APX
>    if (CPU_FEATURE_USABLE_P (cpu_features, FMA4))
>      return OPTIMIZE (fma4);
> +#endif
>
>    if (CPU_FEATURE_USABLE_P (cpu_features, AVX))
>      return OPTIMIZE (avx);
> diff --git a/sysdeps/x86_64/fpu/multiarch/ifunc-fma4.h
> b/sysdeps/x86_64/fpu/multiarch/ifunc-fma4.h
> index c35ba13845..18d372d25c 100644
> --- a/sysdeps/x86_64/fpu/multiarch/ifunc-fma4.h
> +++ b/sysdeps/x86_64/fpu/multiarch/ifunc-fma4.h
> @@ -32,8 +32,10 @@ IFUNC_SELECTOR (void)
>        && CPU_FEATURE_USABLE_P (cpu_features, AVX2))
>      return OPTIMIZE (fma);
>
> +#ifndef HAVE_X86_APX
>    if (CPU_FEATURE_USABLE_P (cpu_features, FMA4))
>      return OPTIMIZE (fma4);
> +#endif
>
>    return OPTIMIZE (sse2);
>  }
> --
> 2.44.0
>
>
LGTM
Reviewed-by: Sunil K Pandey <skpgkp2@gmail.com>

--Sunil