Fix exp in non-default rounding modes (bug 3976)

[David Miller]

From: "Joseph S. Myers" <joseph@codesourcery.com>
Date: Wed, 29 Feb 2012 01:10:01 +0000 (UTC)

> Bug 3976 discusses how some libm functions produce wildly incorrect
> results in rounding modes other than round-to-nearest (they don't need
> to be correctly rounded for each mode, but they should still produce
> some reasonable approximation to the correct value of the function).

I don't think we need to be so accomodating.

Calling libm functions in any rounding mode other than
round-to-nearest is completely undefined.

And if that is the case, I seriously question the value of doing
expensive control word changes in any of these routines.