[flang] Flang fails to interpret `-x f95-cpp-input` and related modes

[inaki-amatria]

• flang version is: flang version 21.0.0git (https://github.com/llvm/llvm-project.git a3dc77c00a012bb613cb08e669dab4fadf88e935).
• Reduced test case:

! foo.f90

program main
  print *, __FILE__, __LINE__
end

• flang invocation is: flang -x f95-cpp-input -c foo.f90.
• Expected behavior is: The code compiles successfully.
• Actual behavior: Compilation aborts with the following error:

error: Semantic errors foo.f90
foo.f90:4:12: error: No explicit type declared for '__file__'
    print *, __FILE__, __LINE__
             ^^^^^^^^
foo.f90:4:22: error: No explicit type declared for '__line__'
    print *, __FILE__, __LINE__
                       ^^^^^^^^

• Analysis: It appears that flang does not correctly interpret -x f95-cpp-input. This option should enable preprocessing (equivalent to flang -x f95 -cpp -c foo.f90), but instead, it is not recognized correctly. Additionally, flang does not properly recognize the following language modes:

  • -x f95 should map to whatever the file extension is or command-line option is passed.
  • -x f77 should map to -ffixed-form.
  • -x f77-cpp-input should map to -x f77 -cpp (i.e., -ffixed-form -cpp).

  flang incorrectly applies -ffixed-form when given -x f95 (see [flang] Flang fails to interpret -x f95-cpp-input and related modes #127617 (comment)):

  $ cat foo.f90
  program main
    print *, __FILE__, __LINE__
  end
  $ flang -x f95 -cpp foo.f90
  foo.f90:1:1: warning: Character in fixed-form label field must be a digit
    program main
    ^
  foo.f90:2:3: warning: Character in fixed-form label field must be a digit
    print *, __FILE__, __LINE__
    ^
  foo.f90:3:1: warning: Character in fixed-form label field must be a digit
    end
    ^

  gfortran correctly applies -ffixed-form -cpp when given -x f77-cpp-input:

  $ cat foo.f90
  #ifdef __FILE__
  subroutine foo()
  end
  #endif
  $ gfortran -x f77-cpp-input foo.f90 
  foo.f90:2:2:
  
      2 | subroutine foo()
        |  1
  Error: Non-numeric character in statement label at (1)
  foo.f90:3:2:
  
      3 | end
        |  1
  Error: Non-numeric character in statement label at (1)

• Request: The -x flag should correctly recognize and apply the appropiate Fortran language modes. This is required to compile codes such as MESA (MESA's Makefile) with flang. It is unclear whether flang aims to mimic gfortran's behavior in its driver. However, since flang explicitly supports f95 and f95-cpp-input, we believe it should at least align with gfortran's behavior for this option.

[llvmbot]

@llvm/issue-subscribers-flang-frontend

Author: Iñaki Amatria Barral (inaki-amatria)

- `flang` version is: `flang version 21.0.0git (https://github.com/llvm/llvm-project.git a3dc77c)`. - Reduced test case: ```f90 ! foo.f90

program main
print *, FILE, LINE
end

- `flang` invocation is: `flang -x f95-cpp-input -c foo.f90`.
- Expected behavior is: The code compiles successfully.
- Actual behavior: Compilation aborts with the following error:

error: Semantic errors foo.f90
foo.f90:4:12: error: No explicit type declared for 'file'
print *, FILE, LINE
^^^^^^^^
foo.f90:4:22: error: No explicit type declared for 'line'
print *, FILE, LINE
^^^^^^^^

- Analysis: It appears that `flang` does not correctly interpret `-x f95-cpp-input`. This option should enable preprocessing (equivalent to `flang -x f95 -cpp -c foo.f90`), but instead, it is not recognized correctly. Additionally, `flang` does not properly recognize the following language modes:
  - `-x f77` should map to `-ffixed-form`.
  - `-x f77-cpp-input` should map to `-x f77 -cpp` (i.e., `-ffixed-form -cpp`).

  `gfortran` correctly applies `-ffixed-form -cpp` when given `-x f77-cpp-input`:
```bash
$ cat foo.f90
#ifdef __FILE__
subroutine foo()
end
#endif
$ gfortran -x f77-cpp-input foo.f90 
foo.f90:2:2:

    2 | subroutine foo()
      |  1
Error: Non-numeric character in statement label at (1)
foo.f90:3:2:

    3 | end
      |  1
Error: Non-numeric character in statement label at (1)

• Request: The -x flag should correctly recognize and apply the appropiate Fortran language modes. This is required to compile codes such as MESA (MESA's Makefile) with flang.

[inaki-amatria]

It looks like the issue with flang applying -ffixed-form when given -x f95 was introduced by #117563. This PR resolves #117712 by always appending -ffixed-form when processing a -x f95 input. However, the real issue appears to be that OP did not pass -ffixed-form when compiling their preprocessed file. Since flang cannot infer whether a .i file uses fixed or free-form Fortran, the correct form must be explicitly specified.