C_NULL_PTR is passed incorrectly with -standard-semantics option

[Neil_Carlson]

The following example calls a C function with a void* argument, first passing an associated TYPE(C_PTR) actual, and then passing C_NULL_PTR.  When compiled with ifort 13.1.2 and the -standard-semantics option, the example executes as expected, but with 14.0.1 the C function is receiving a non NULL pointer in the latter case, which is incorrect.  Note that the example runs correctly when the -standard-semantics option is omitted.

I'll attach the C and Fortran source files, but here is the Fortran code:

[fortran]program main

  use :: iso_c_binding

  interface

    subroutine c_function(arg) bind(c)

      import c_ptr

      type(c_ptr), value :: arg

    end subroutine

  end interface

  integer, target :: arg = 42

  print *, 'passing c_loc(arg) to C_function'

  call c_function (c_loc(arg))

  print *, 'passing c_null_ptr to C_function'

  call c_function (c_null_ptr)

end program[/fortran]

and the C code:

[cpp]#include <stdio.h>

void c_function (void *arg) {

  if (arg == NULL)

    printf("\\targ pointer is NULL in c_function\\n");

  else

    printf("\\targ pointer is not NULL in c_function; arg=%d\\n",*(int*)arg);

}[/cpp]

The (correct) output using the 13.1.2 compiler is:

passing c_loc(arg) to C_function

arg pointer is not NULL in c_function; arg=42

passing c_null_ptr to C_function

arg pointer is NULL in c_function <=== THIS IS RIGHT

The incorrect output using the 14.0.1 compiler is:

passing c_loc(arg) to C_function

arg pointer is not NULL in c_function; arg=42

passing c_null_ptr to C_function

arg pointer is not NULL in c_function; arg=0  <=== THIS IS WRONG

[Steven_L_Intel1]

This should be fixed in 14.0.2, which I expect to be available Wednesday or Thursday this week.

[FortranFan]

Steve,

Is it possible for you to explain how the -standard_semantics option adversely affected the behavior in 14.0.1?  I'm curious since the code posted appears to use standard Fortran 2003 stuff for C-interoperability.  Hence I would have thought including or omitting the -standard_semantics option would make no difference.

Thanks,

[Steven_L_Intel1]

It was related to the treatment of VALUE for non-BIND(C) routines. We had misimplemented this back in 10.0 causing it to act like ATTRIBUTES VALUE, where the standard says to make a copy. But when we tried to fix this late in the 14.0 beta cycle, it caused problems elsewhere, so we came up with -assume std_value and made this part of -standard-semantics.

The problem here was that pass-by-value for "small structures" requires special treatment and we missed making the appropriate changes for that case, when we backed out the original fix, resulting in the inconsistency noted here.

[Steven_L_Intel1]

Update 2 is now available at the Intel Registration Center.