This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] update libm-test-ulps for new exp
On Thu, 19 Jul 2018, Wilco Dijkstra wrote:
> Would it be feasible to include a generic ulp file which sets the
> expected ulps into the target ulp files that add or override such
> settings? That way most ulp files could be significantly reduced and
> only targets with special formats may need to have more entries. Also
> this would make it clearer when there are anomalies and possible ulp
> errors that need to be fixed.
One could construct various such schemes, yes, at least for float and
double ulps (but being able to share ldbl-128 ulps would be a good idea as
well, including the same ulps being used for long double on some platforms
and _Float128 on others, it's not as if there are architecture-specific
ldbl-128 functions beyond fully-determined ones such as sqrtl and fmal -
though there are architecture variations in whether multiply and add might
get fused for float128).
I'd suggest a requirement for any such scheme should be code sharing
between ulps file parsing for libm tests (gen-libm-test.pl) and the manual
(libm-err-tab.pl). We probably want to finish getting consensus on the
use of Python to replace Perl (and possibly awk) in the build first.
--
Joseph S. Myers
joseph@codesourcery.com