runtime failure due to useing character*8 in DECLARE_CCTK_ARGUMENTS_CHECKED macro

Create issue
Issue #2571 new
Roland Haas created an issue

I just learned (via Skype) that Cactus' use of character*8 variables can cause runtime errors in Fortran:

At line 13 of file Cactus/arrangements/EinsteinBase/TmunuBase/src/SetStressEnergyState.F90
Fortran runtime error: Actual string length is shorter than the declared one for dummy argument 'dtbetax_p' (0/8)

and digging down a bit and looking at the processed Fortran file (configs/sim/build/TmunuBase/SetStressEnergyState.f90) one finds:

characTer*8, intent(IN) :: dtbetax_p
integer, parameter :: cctki_use_dtbetax_p = kind(dtbetax_p)

and https://stackoverflow.com/questions/4780069/passing-a-string-as-an-argument-when-the-dummy-has-specified-length where the Fortran 2003 standard gets quoted with an explanation:

If a scalar dummy argument is of type default character, the length len of the dummy argument shall be less than or equal to the length of the actual argument. The dummy argument becomes associated with the leftmost len characters of the actual argument.

This all used gfortran 11.

Safest would be if we could declare as characTer*0 if that is allowed. Of course it’s still wrong in that the actual argument passed to dtbetax_p is never a character at all (it’s either a double precision real or NULL) and we only use the characTer construct to avoid assignment and also unused actual argument warnings.

Comments (16)

  1. Erik Schnetter

    A declaration character*8 as input argument is really strange. one would only use that if one knows that the length is 8, similar to passing a character array in C (instead of a character pointer).

    Usually one would write character*(*), and the compiler then passes the actual string length as additional (hidden) argument, and uses that length in the callee.

    In this case I rather think that the Fortran code is a mistranslation from C, and that the code generating the _CHECKED macro is wrong. dtbetax_p should be declared as CCTK_REAL dtbetax_p(ash0, ash1, ash2) instead. None of the Cactus grid function or parameter passing arguments or declaration use Fortran character variables.

  2. Roland Haas reporter

    We are using characTer*8, intent(IN) :: dtbetax_p for grid functions that are neither read nor written by a scheduled function that is that should not even be accessible. Since there is only a single C to Fortran wrapper per thorn, they are still being passed to the Fortran subroutine but we want to make sure that the compiler aborts with an error if the code actually tries to access it.

    Declaring it as a nonsensical type achieves that since the compiler will not let you assign a real or integer number to a charcater*8. We can only use character*(*) if that does not add an extra integer “length of string” to the end of the argument list.

    We may however have to do something like:

    CCTK_REAL :: dtbetax_p
    integer, parameter :: cctki_use_dtbetax_p = kind(dtbetax_p)
    #define dtbetax_p dtbetax_p_IS_NOT_ACCESSIBLE
    

  3. Erik Schnetter

    Clever!

    Character types are a bit magical in Fortran; they don’t really behave like strings in C. You could use logical*8 instead; assignments between integer/`real` and logical are also disallowed in Fortran.

  4. Roland Haas reporter

    Changing from charater*8 to logical*8 may not be not be enough to avoid assignment. The intel compiler (ifort (IFORT) 18.0.3 20180410) compiles this without warning:

    program foo
      implicit none
    
      logical*8 :: bar
      integer*4 :: baz
    
      bar = baz
    
    end program
    

    On the other hand we are not worried about assignment to the pointer itself. We want to fails is things like: bar(1,2,3) for reading and writing and passing bar to a subroutine (which we likely cannot prevent no matter what).

    So it should be enough for our purposes.

  5. Roland Haas reporter

    Ah, one exception: we do worry about assignment without parenthesis in the case of grid scalars eg ADMBase::shift_state would be accessed like this if(shift_state .ne. 0) then STOP endif. Though I am not sure if either charater*8 or logical*8 would prevent this (I suspect not).

  6. Erik Schnetter

    Which line triggers the original error about the string length? Is it the call to the kind function, or is it the call to that routine?

  7. Roland Haas reporter

    It apparently happened in a simulation of Kiki’s (Kyriaki Dionysopoulou). Apparently this happened (the error messages in the ticket description) in a regular run and “with different parameter files” and “completely different setups” (quoting Kiki). I have no idea why this never showed up before in the time since it has been introduced.

    Not sure if this matters but apparently this happened on (at least) an older, Intel macOS laptop using:

    CPP = cpp-11
    FPP = cpp-11
    CC = gcc-11
    CXX = g++-11
    F90 = gfortran-11

    GNU Fortran (Homebrew GCC 11.2.0) 11.2.0
    Copyright (C) 2021 Free Software Foundation, Inc.
    This is free software; see the source for copying conditions. There is NO
    warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

  8. Roland Haas reporter

    The error message mentions a line number (13) in Cactus/arrangements/EinsteinBase/TmunuBase/src/SetStressEnergyState.F90 which is the line:

    subroutine TmunuBase_SetStressEnergyState (CCTK_ARGUMENTS)
    

    The error message also gives the dummy argument in question and the found and expected lengths (0 and 8 respectively): 'dtbetax_p' (0/8)

  9. Erik Schnetter

    The problem is caused by bounds checking, in particular the -fcheck=bounds option that is used in debug mode.

    See <https://godbolt.org/z/zhGv3bjj6> for a shortened example. The error message and the comparison (line 8 ) are clearly visible.

    I assume what happens is that gfortran stores the actual string length in memory, presumably just before the actual string. With bounds checking enabled, the string length is checked. Cactus never sets up the string length since it just reinterprets pointers, assuming that gfortran’s string layout mimics that of C.

    You could define a new type (“struct”) instead of using character*8:

    subroutine sub(beta)
      implicit none
      type empty
      end type
      type(empty) beta
    end subroutine sub
    

    However, this wouldn’t allow the kind trick to make the function argument appear used.

  10. Roland Haas reporter

    Interesting. A logical*8 would still be viable (there is no bounds check for it). I’d rather risk cases of grid scalars not being found (read only at least, write access is caught due to INTENT(IN)) than having lots of “unused actual argument” warnings.

  11. Roland Haas reporter

    Something like this would work:

      type inaccessible
      integer :: dummy
      end type
      type(inaccessible), intent(IN):: alp
      integer, parameter :: cctki_use_alp = kind(alp%dummy)
    

    Trying to access it via eg foo = alp produces an error:

    Error: Cannot convert TYPE(inaccessible) to INTEGER(4) at (1)

    One could even make dummy of the actual CCTK_XXX type of the grid variable.

  12. Log in to comment