f2py fails for functions returning character(len=len(input_string))

[chatcannon]

I've written simple toupper() and tolower() functions in Fortran that start like this:

pure function toupper(str)
  character(len=*) intent(in) :: str
  character(len=len(str)) :: toupper
  ...

When wrapping these with f2py, I get an error looking like this:

/tmp/tmpmrZ5Ra/src.linux-x86_64-2.7/RMCfunctionsmodule.c: In function ‘f2py_rout_RMCfunctions_mod_functions_toupper’:
/tmp/tmpmrZ5Ra/src.linux-x86_64-2.7/RMCfunctionsmodule.c:926:23: error: ‘str_Dims’ undeclared (first use in this function)

slen(toupper) = len(str);
                      ^

I guess the len(str) needs to be replaced by slen(str).

This problem occurs with both f2py-2.7 and f2py-3.3
It does not depend on the name of the variable str - I tried a couple of other names.

[chatcannon]

Sorry, I couldn't work out how to tag this with component: numpy.f2py

[chatcannon]

P.S. I'm running numpy version 1.7.1 on Gentoo GNU/Linux

[charris]

@chatcannon I suspect this problem persists, but could you try with numpy 1.8 or current development?

[Goutte]

I can confirm the issue is still here in numpy 1.20.
Barring a fix (it's been 8 years), any tip on how to bypass this issue would be welcome!

[melissawm]

Here's what I've managed to get so far:

1. Wrapping the following:

  pure function test(str)
    character(len=*), intent(in) :: str
    character(len=len(str)) :: test
    integer :: i

    test(1:1) = 'A'
    do i = 2, len(str)
       test(i:i) = str(i:i)
    enddo
    
  end function test

I get:

/tmp/tmpdg8zrhbc/src.linux-x86_64-3.9/stringsmodule.c: In function 'f2py_rout_strings_strings_test':
/tmp/tmpdg8zrhbc/src.linux-x86_64-3.9/stringsmodule.c:279:20: error: 'str_Dims' undeclared (first use in this function)
  279 |   slen(test) = len(str);
      |                    ^~~

2. I tried using a subroutine, with the following signature file:

python module strings ! in 
    interface  ! in :strings
        module strings ! in :strings:strings.f90
            subroutine test(str, a) ! in :strings:strings.f90:strings
                character*(*) intent(in) :: str
                character*len(str) intent(out) :: a
            end subroutine test
        end module strings
    end interface 
end python module strings

Running $ f2py -c strings.f90 strings.pyf compiles fine, but

>>> import strings
>>> print(strings.strings.test.__doc__)  # should return a as output, but returns
test(str,a)

Wrapper for ``test``.

Parameters
----------
str : input string(len=-1)
a : input string(len=1)

(note the incorrect intent for a)

3. Changing strings.pyf to use character(len=slen(str)), intent(out) :: a runs fine and works.

So the problem seems to be when using a function but not a subroutine.

When inspecting the code, IIUC the problem comes from the fact that we need to use slen for the C function wrapper, but len for the fortran function wrapper (both generated by f2py).

My question to @pearu is if we should require the use of the slen expression in the pyf file in this case, or try to fix this from f2py itself, considering the two separate cases of the fortran and C function wrappers.

[pearu]

My first reaction (without looking at code) is that f2py should be fixed by replacing len(..) call with slen(..) when len is used in the context of character len specification. Note that there are three kinds of length functions involved here: slen and len defined by f2py and to be used withing .pyf files, and len that is Fortran intrinsic. f2py len should be used in the context of arrays while f2py slen in the context of character types.