Ondřej Čertík

@ghwilliams thanks a lot for creating the page and suggesting an API. My main feedback is that it looks like this would be the high level API -- or middle...

This is really important. I don't think we should use allocatable at all in the lowest level API, unless the size of the output is truly not known ahead of...

We should setup a project to carefully benchmark all these options. Note that gfortran and Intel can put an allocatable on a stack as well. On Mon, Jan 20, 2020,...

Also I want to setup benchmarks for operations on arrays (`y(:) = a(:) + b(:)`) compared to an explicit loop. The loop seems to be faster some times and I...

@zerothi indeed, that seems to be what is needed now. It seems very unfortunate that the natural syntax `a = [1, 2, 3]` has a runtime overhead, and one must...

> I have seen the same behaviour on both Intel and GCC, but yes, compilers could handle this differently. @zerothi how could compilers handle it differently?

Both GFortran and Intel have such an option, see, e.g.: https://software.intel.com/en-us/fortran-compiler-developer-guide-and-reference-standard-realloc-lhs. I didn't know if that's what you were talking about, or something else.

Yes, Fortran definitely needs an arbitrary precision floating point, as either a library or part of stdlib. Arbitrary precision lapack/blas is also occasionally useful. I've used quadruple precision Lapack in...

CC @awvwgk, @milancurcic, @LKedward, @arjenmarkus What would be the best approach?

@nshaffer you described the "nested functions" approach: https://www.fortran90.org/src/best-practices.html#iv-nested-functions. It works well, and it is my method of choice also, but there is one caveat that we recently discovered and discussed...