Fortran Intrinsics (2018)

Beware of the examples in the Fortran standard for intrinsic functions like TINY() -- the example results are within the framework of a 32-bit floating-point model that looks an awful lot like IEEE-754 finite numbers (24-bit precision including an implicit normalization bit, 8-bit exponent biased by 127), except for an interpretation that is off by a factor of two.

That's right. If you take any of those example calls from the Fortran standard and compile them on any modern processor, your results will differ by a factor of two one way or another.

Does the standard warn you about this? Nowhere. Does the standard committee know about this? Yes. Will they fix their documentary floating-point model to conform with the one truly ubiquitous 32-bit floating-point representation left to us? Of course not.

I still remember back in 2004 when I needed to use the bessel j functions for a C++ electromagnetics simulation for the university. Back then, these functions were only available on a 1986 Fortran implementation by DE Amos (https://dl.acm.org/doi/10.1145/7921.214331).

I was so delighted when I was able to compile the fortran sources as .obj files and then call them from my C++ program to receive the correct results!

Interesting that the website advertise fortran as natively parallel, was it the case in the 60s too or just some marketing trick to get people to pay attention to this old friend ? (honest question)

I know that matrix inversion was deliberately left out of the fortan built in functions