Returns the mantissa of X in the real model for that kind of floating-point number; if X is finite, this is a value between 1/RADIX(X) and 1.
If the value of X is representable in that model, the result satisfies the relationship X = FRACTION(X)*(REAL(RADIX(X),KIND(X))**EXPONENT(X),
or if X is an IEEE infinity or NaN, the result is an IEEE NaN.
Note that if X is double-double precision, FRACTION(X) might not be accurate but the result of EXPONENT(X) is.
Like many intrinsic functions, this is elemental; the result has the same shape as X, with the operation being applied elementwise.