[PATCH] libm-test.inc: Correctly implement ulp().

[Joseph S. Myers]

On Fri, 24 May 2013, Chris Metcalf wrote:

> On 5/17/2013 3:18 PM, Carlos O'Donell wrote:
> > On 05/17/2013 03:09 PM, Andreas Jaeger wrote:
> >> Does this change mean we should regenerate the ULPs for all
> >> architectures from scratch?
> > As part of the final 2.18 release process everyone should be
> > regenerating ulps from scratch.
> 
> Are we actually at that stage now?  It does seem that Joseph is still 
> pouring in his cool code-to-data rework changes for the math tests, and 
> presumably it makes sense to hold off regenerating until that's 
> complete?

David will say exactly when we're freezing for 2.18 release, at which 
point changes requiring ulps updates should no longer be going in and 
architecture maintainers can truncate and regenerate their ulps (taking 
the usual care about not checking in excessively high ulps, or ulps for 
functions such as rint/sqrt/fma/... that shouldn't have any, at least in 
the absence of a relevant bug filed in Bugzilla ... note there are such 
ulps checked in for fma, fmod, sqrt and cproj at present, but I hope the 
latter three will go away on regeneration).

It is of course possible that during the freeze architecture-specific 
problems may be fixed to eliminate errors that show up in the regeneration 
of ulps.

-- 
Joseph S. Myers
joseph@codesourcery.com