Arch maintainers: please regenerate libm-test-ulps for 2.19

[Joseph S. Myers]

On Sat, 25 Jan 2014, kkojima@rr.iij4u.or.jp wrote:

> "Joseph S. Myers" <joseph@codesourcery.com> wrote:
> > The following architectures do not appear to have had libm-test-ulps 
> > regenerated for 2.19:
> [snip]
> > sh (I also suggest moving the file from sysdeps/sh/sh4/fpu to sysdeps/sh, 
> >   as ulps for SH4 with FPU should be a superset of those for other 
> >   variants)
> 
> I've committed the patch below.

The errors for fma indicate a (presumably SH-specific) bug (which should 
be filed in Bugzilla before investigating it).

Those for float

+Test "fma (0x1.7ff8p+13, 0x1.000002p+0, 0x1.ffffp-24)":
+float: 1
+ifloat: 1

are undoubtedly the easiest to investigate.  First make sure that the case 
in sysdeps/sh/s_fmaf.c that includes sysdeps/ieee754/dbl-64/s_fmaf.c is 
indeed the one being used.  Then disassemble fmaf and see if all the 
operations are in the right order - nothing's been scheduled across a call 
to feholdexcept / fesetround / feupdateenv / fetestexcept.  If there's a 
scheduling problem, we can work out how to prevent it, otherwise it's 
likely to be a problem with one of those <fenv.h> functions.


-- 
Joseph S. Myers
joseph@codesourcery.com