Re: [PATCH] Fortran: passing of optional scalar arguments with VALUE attribute [PR113377]

[Mikael Morin <morin-mikael@orange.fr>]

Hello,

Le 20/01/2024 à 22:58, Harald Anlauf a écrit :
> Dear all,
> 
> here's the first part of an attempt to fix issues with optional
> dummy arguments as actual arguments to optional dummies.  This patch
> rectifies the case of scalar dummies with the VALUE attribute,
> which in gfortran's argument passing convention are passed on the
> stack when they are of intrinsic type, and have a hidden variable
> for the presence status.
> 
> The testcase tries to cover valid combinations of actual and dummy
> argument.  A few tests that are not standard-conforming but would
> still work with gfortran (due to the argument passing convention)
> are left there but commented out with a pointer to the standard
> (thanks, Mikael!).
> 
> Regtested on x86_64-pc-linux-gnu.  OK for mainline?
> 
Well, not yet.

> 
> diff --git a/gcc/fortran/trans-expr.cc b/gcc/fortran/trans-expr.cc
> index 9dd1f4086f4..2f47a75955c 100644
> --- a/gcc/fortran/trans-expr.cc
> +++ b/gcc/fortran/trans-expr.cc
> @@ -6526,6 +6526,10 @@ gfc_conv_procedure_call (gfc_se * se, gfc_symbol * sym,
>  			    gfc_init_se (&argse, NULL);
>  			    argse.want_pointer = 1;
>  			    gfc_conv_expr (&argse, e);
> +			    if (e->symtree->n.sym->attr.dummy
> +				&& POINTER_TYPE_P (TREE_TYPE (argse.expr)))
> +			      argse.expr = gfc_build_addr_expr (NULL_TREE,
> +								argse.expr);

The second part of the condition looks superfluous: if 
argse.want_pointer was set, we can expect to get a pointer result.

But more important, I don't understand the need for this whole part, the 
new test seems to pass without it.
And here is an example that regresses with it.

program p
   type :: t
     integer, allocatable :: c
   end type
   call s2(t())
contains
   subroutine s1(a)
     integer, value, optional :: a
     if (present(a)) stop 1
   end subroutine
   subroutine s2(a)
     type(t) :: a
     call s1(a%c)
   end subroutine
end program


>  			    cond = fold_convert (TREE_TYPE (argse.expr),
>  						 null_pointer_node);
>  			    cond = fold_build2_loc (input_location, NE_EXPR,
> @@ -7256,6 +7260,7 @@ gfc_conv_procedure_call (gfc_se * se, gfc_symbol * sym,
>  	      && e->symtree->n.sym->attr.optional
>  	      && (((e->rank != 0 && elemental_proc)
>  		   || e->representation.length || e->ts.type == BT_CHARACTER
> +		   || (e->rank == 0 && e->symtree->n.sym->attr.value)

This looks good.

>  		   || (e->rank != 0
>  		       && (fsym == NULL
>  			   || (fsym->as
> diff --git a/gcc/testsuite/gfortran.dg/optional_absent_9.f90 b/gcc/testsuite/gfortran.dg/optional_absent_9.f90
> new file mode 100644
> index 00000000000..495a6c00d7f
> --- /dev/null
> +++ b/gcc/testsuite/gfortran.dg/optional_absent_9.f90
> @@ -0,0 +1,324 @@
> +! { dg-do run }
> +! PR fortran/113377
> +!
> +! Test passing of missing optional scalar dummies of intrinsic type
> +
> +module m_int
> +  implicit none
> +contains
> +  subroutine test_int ()
> +    integer :: k = 1
> +    call one     (k)
> +    call one_val (k)
> +    call one_all (k)
> +    call one_ptr (k)
> +  end
> +
> +  subroutine one (i, j)
> +    integer, intent(in)           :: i
> +    integer             ,optional :: j
> +    integer, allocatable :: aa
> +    integer, pointer     :: pp => NULL()
> +    if (present (j)) error stop "j is present"
> +    call two     (i, j)
> +    call two_val (i, j)
> +    call two     (i, aa)
> +    call two     (i, pp)

To be complete, you could check two_val(i, aa) and two_val(i, pp) as well.
Both seem to pass already without the patch, so not absolutely needed.
Your call.

> +  end
> +

I think the patch is OK with the first trans-expr.cc hunk removed.
Thanks.

Mikael