Fix e_logl (128ibm) spurious underflow

[Joseph S. Myers]

On Wed, 27 Feb 2013, Richard Henderson wrote:

> On 2013-02-27 12:11, Adhemerval Zanella wrote:
> > +static const long double
> > +  ldbl_epsilon = 0x1p-106L;
> > +
> 
> Can't you use the <float.h> constant LDBL_EPSILON?

The semantics of LDBL_EPSILON for IBM long double (the least positive 
subnormal double value, because of the discontiguous mantissa bits) are 
wrong for uses such as this.

-- 
Joseph S. Myers
joseph@codesourcery.com