Feature Request: Detection of NaN values

[dominikbell]

I want to detect NaN values in a file that is to be "pyccelized" (translated to fortran and compiled). As far as I know and understood, there is currently no way of doing that. When trying to import the function isnan, or the types NaN or nan from numpy, the compilation command throws the fatal error

Function 'isnan' from module 'numpy' is not currently supported by pyccel (from numpy import isnan)

when e.g. trying to import isnan.

I am not sure if this counts as a "non-trivial feature" but on the issue creation of a "Feature Request" it told me to link to a discussion which should take place here. I have not found a discussion, PR, or issue concerning this feature.

[EmilyBourne]

I think NaN detection can be non-trivial in Fortran.
In C it is simple. There is a isnan function in math.h (https://en.cppreference.com/w/c/numeric/math/isnan).
However in Fortran it can be compiler-dependent. GNU provides an isnan function (https://gcc.gnu.org/onlinedocs/gcc-4.8.5/gfortran/ISNAN.html) as does Intel (https://www.intel.com/content/www/us/en/docs/fortran-compiler/developer-guide-reference/2023-0/isnan.html). I have also seen people use a /= a (returns True if a is NaN but is often optimised out by the compiler).

That said I have just discovered the function ieee_is_nan from the ieee_arithmetic intrinsic module which was introduced in Fortran 2003. In my opinion this module could be used to implement the NumPy functions:

• isnan
• isfinite
• isinf

@yguclu What is your opinion on this solution?

[yguclu]

@EmilyBourne I am absolutely in favor of using an intrinsic module for this. How much do you know about the compilers' support for ieee_arithmetic?

[EmilyBourne]

According to this: https://fortranwiki.org/fortran/show/Fortran+2003+status it is supported by all major compilers

[EmilyBourne]

This is the earliest versions I can see in the source documents. These seem early enough for ieee_arithmetic to not be a problem

[yguclu]

Excellent! I think we can go forward and create dedicated issues.