real

Real floating point numbers

Quiet NaN

We might choose to use NaN as a sentinal value, where instead of returning separate "OK" logical variable from a function or subroutine, if a failure happens, we return NaN in one of the important variables. There was a classical way to do this that was type specific, by setting the NaN bit pattern for your data type. For example, for single-precision real you'd type

nan_bit = transfer(Z'7FF80000',1.)

For a standards-based way to handle all floating point types, consider:

use, intrinsic :: ieee_arithmetic, only: ieee_value, ieee_quiet_nan

nan = ieee_value(1.,ieee_quiet_nan)

This is in program real/nans.f90

NaN Notes

• must NOT use -Ofast or -ffast-math because IEEE standards are broken by these options--NaN detection will intermittently fail!
• gfortran ≥ 6 needed for ieee_arithmetic: ieee_is_nan