benchtests: Run _Float128 tests only on architectures that support it

  From: Arjun Shankar <arjun@redhat.com>

__float128 is a non-standard name and is not available on some architectures
(like aarch64 or s390x) even though they may support the standard _Float128
type.  Other architectures (like armv7) don't support quad-precision
floating-point operations at all.

This commit adds a configure check for quad-precision FP support and runs
the corresponding benchtests only on architectures that support it.  It also
replaces references to __float128 with _Float128.
---
 Tested on armv7hl, i686, x86_64, aarch64, ppc64le, and s390x.

 benchtests/Makefile       |  6 +++++-
 benchtests/expf128-inputs |  4 ++--
 benchtests/powf128-inputs |  4 ++--
 benchtests/sinf128-inputs |  4 ++--
 configure                 | 24 ++++++++++++++++++++++++
 configure.ac              | 15 +++++++++++++++
 6 files changed, 50 insertions(+), 7 deletions(-)
  

Dear Arjun,

some comments on your patch:

* the changes to benchtests/Makefile and benchtests/*f128-inputs look ok to me
* for configure.ac I'm not fluent enough in the autotools
* I wonder why you also patch configure, isn't this file automatically
  generated?
* finally "find . -type f -exec grep \-l __float128 {} \;" finds several other
  instances of __float128 (sysdeps/x86_64/fpu/math-tests-snan.h for example).
  Should some of those be changed to _Float128 too?

Best regards,
Paul
  

* Paul Zimmermann:

> * I wonder why you also patch configure, isn't this file automatically
>   generated?

It's checked into the repository, so it shows up in the diff.

> * finally "find . -type f -exec grep \-l __float128 {} \;" finds several other
>   instances of __float128 (sysdeps/x86_64/fpu/math-tests-snan.h for example).
>   Should some of those be changed to _Float128 too?

I think it's mostly comments or architecture-specific code.  For x86, we
know that __float128 exists, so that's not actually a problem.

Thanks,
Florian
  

* Arjun Shankar:

> diff --git a/configure.ac b/configure.ac
> index 93e68fb696..7ee79162be 100644
> --- a/configure.ac
> +++ b/configure.ac
> @@ -1821,6 +1821,21 @@ libc_cv_pie_default=$libc_cv_cc_pie_default
>  AC_SUBST(libc_cv_cc_pie_default)
>  AC_SUBST(libc_cv_pie_default)
>  
> +AC_CACHE_CHECK([whether quad-precision floating-point is supported],
> +libc_cv_have_qp_float, [libc_cv_have_qp_float=yes
> +cat > conftest.c <<EOF
> +#include <math.h>
> +#if __HAVE_FLOAT128
> +# error quad-precision floating-point is supported
> +#endif
> +EOF
> +if eval "${CC-cc} -S conftest.c 2>&AS_MESSAGE_LOG_FD 1>&AS_MESSAGE_LOG_FD"; then
> +  libc_cv_have_qp_float=no
> +fi
> +rm -f conftest.*])
> +AC_SUBST(libc_cv_have_qp_float)
> +LIBC_CONFIG_VAR([have-qp-float], [$libc_cv_have_qp_float])

This configure test does not run against <math.h> from the glibc tree,
so its result is not correct.

Please try adding the benchtests via a fragment in
sysdeps/ieee754/float128/Makefile.  I think this won't need additional
configure checks.

Thanks,
Florian
  

On Thu, 17 Sep 2020, Florian Weimer via Libc-alpha wrote:

> Please try adding the benchtests via a fragment in
> sysdeps/ieee754/float128/Makefile.  I think this won't need additional
> configure checks.

If you want these benchmarks only if _Float128 has a different format from 
long double, you can use the existing makefile variable $(float128-fcts).  
If you want to run them also when _Float128 has the same format as long 
double, use $(float128-alias-fcts) as well.  See math/Makefile for example 
usage.
  

* Joseph Myers:

> On Thu, 17 Sep 2020, Florian Weimer via Libc-alpha wrote:
>
>> Please try adding the benchtests via a fragment in
>> sysdeps/ieee754/float128/Makefile.  I think this won't need additional
>> configure checks.
>
> If you want these benchmarks only if _Float128 has a different format from 
> long double, you can use the existing makefile variable $(float128-fcts).  
> If you want to run them also when _Float128 has the same format as long 
> double, use $(float128-alias-fcts) as well.  See math/Makefile for example 
> usage.

In this case, the ifeq Makefile conditional should be added to
benchtests/Makefile, right?

Thanks,
Florian
  

On Thu, 17 Sep 2020, Florian Weimer via Libc-alpha wrote:

> * Joseph Myers:
> 
> > On Thu, 17 Sep 2020, Florian Weimer via Libc-alpha wrote:
> >
> >> Please try adding the benchtests via a fragment in
> >> sysdeps/ieee754/float128/Makefile.  I think this won't need additional
> >> configure checks.
> >
> > If you want these benchmarks only if _Float128 has a different format from 
> > long double, you can use the existing makefile variable $(float128-fcts).  
> > If you want to run them also when _Float128 has the same format as long 
> > double, use $(float128-alias-fcts) as well.  See math/Makefile for example 
> > usage.
> 
> In this case, the ifeq Makefile conditional should be added to
> benchtests/Makefile, right?

ifeq, or any of the other mechanisms used such as $(var-$(float128-fcts)).