Dear all,

I intend to commit the attached obvious patch within 24h unless there
are comments.

The issue is an off-by-one error in setting up the maximum exponent
of the real kind that is passed to mpfr, so that model numbers between
huge(x)/2 and huge(x), when given as an argument to NEAREST(arg,-1.0),
are rounded down to huge(x)/2 during compile-time simplification.
As no such issue is observed at run-time, the testcase compares the
compile-time and run-time results for corner cases.

Regtested on x86_64-pc-linux-gnu.

As this is sort of a wrong-code issue, I intend to backport to all
open branches.

(The issue was apparently introduced in r0-84566-gb6f63e898498e6
without noticing, so it is technically a regression.)

Thanks,
Harald