With this variant:

module pdfs
  type :: pdf
  end type pdf

  abstract interface
    pure function getdims(this)
      import pdf
      class(pdf), intent(in) :: this
      integer getdims
    end function getdims
  end interface
  
contains

  pure function getx(this)
    class(pdf), intent(in) :: this
    real, dimension(getdims(this)) :: getx
  end function getx

end module pdfs


I get the correct error message (unfortunately twice):

    real, dimension(getdims(this)) :: getx
                    1
Error: ABSTRACT INTERFACE 'getdims' must not be referenced at (1)


I'm not completely sure if the usage of 'getdims' is valid in comment 0.

Yet another variant:


module pdfs
  type, abstract :: pdf
  contains
    procedure(getdims), deferred :: getdims
    procedure(getx), deferred :: getx
  end type pdf

  abstract interface
    pure function getdims(this)
      import pdf
      class(pdf), intent(in) :: this
      integer getdims
    end function getdims

    pure function getx(this)
      import pdf
      class(pdf), intent(in) :: this
      real, dimension(this%getdims()) :: getx
    end function getx
  end interface

end module pdfs



gives me (four times!):

      real, dimension(this%getdims()) :: getx
                      1
Error: Function 'this' at (1) must be PURE


which is of course bogus, since 'this' is not even a function, and getdims is actually pure.

This fixes the ICE in comment 0:

Index: gcc/fortran/expr.c
===================================================================
--- gcc/fortran/expr.c	(revision 195125)
+++ gcc/fortran/expr.c	(working copy)
@@ -4335,9 +4335,11 @@ replace_comp (gfc_expr *expr, gfc_symbol *sym, int
 	 the symtree rather than create a new one (and probably fail later).  */
       stree = gfc_find_symtree (ns ? ns->sym_root : gfc_current_ns->sym_root,
 		      		expr->symtree->n.sym->name);
-      gcc_assert (stree);
-      stree->n.sym->attr = expr->symtree->n.sym->attr;
-      expr->symtree = stree;
+      if (stree)
+	{
+	  stree->n.sym->attr = expr->symtree->n.sym->attr;
+	  expr->symtree = stree;
+	}
     }
   return false;
 }

(In reply to comment #2)
> Yet another variant:
> [...]
>
> gives me (four times!):
> 
>       real, dimension(this%getdims()) :: getx
>                       1
> Error: Function 'this' at (1) must be PURE
> 
> 
> which is of course bogus, since 'this' is not even a function, and getdims is
> actually pure.

The wording of the error message can be fixed by this patch:


Index: gcc/fortran/expr.c
===================================================================
--- gcc/fortran/expr.c	(revision 195310)
+++ gcc/fortran/expr.c	(working copy)
@@ -2962,18 +2962,24 @@ gfc_specification_expr (gfc_expr *e)
       return FAILURE;
     }
 
-  comp = gfc_get_proc_ptr_comp (e);
-  if (e->expr_type == EXPR_FUNCTION
-      && !e->value.function.isym
-      && !e->value.function.esym
-      && !gfc_pure (e->symtree->n.sym)
-      && (!comp || !comp->attr.pure))
+  if (e->expr_type == EXPR_FUNCTION)
     {
-      gfc_error ("Function '%s' at %L must be PURE",
-		 e->symtree->n.sym->name, &e->where);
-      /* Prevent repeat error messages.  */
-      e->symtree->n.sym->attr.pure = 1;
-      return FAILURE;
+      comp = gfc_get_proc_ptr_comp (e);
+      if (comp && !comp->attr.pure)
+	{
+	  gfc_error ("Procedure pointer component '%s' at %L must be PURE",
+		     comp->name, &e->where);
+	  return FAILURE;
+	}
+      else if (!comp && !e->value.function.isym && !e->value.function.esym
+	       && !gfc_pure (e->symtree->n.sym))
+	{
+	  gfc_error ("Function '%s' at %L must be PURE",
+		     e->symtree->n.sym->name, &e->where);
+	  /* Prevent repeat error messages.  */
+	  e->symtree->n.sym->attr.pure = 1;
+	  return FAILURE;
+	}
     }
 
   if (e->rank != 0)



With this ones gets:

      real, dimension(this%getdims()) :: getx
                      1
Error: Procedure pointer component 'getdims' at (1) must be PURE


This still leaves us with some problems:
1) That we get the error at all. The interface of getdims is clearly PURE, but it seems that its resolution happens too late, so that the attributes have not yet been copied from the interface when the pureness check is done.
2) getdims is identified as a PPC in the error message, though it is a type-bound procedure. The reason for this is that internally "this%getdims()" is resolved to "this->_vptr->getdims" (where getdims is a PPC of _vptr).
3) The error message still comes four times (not sure why).

Note that 'replace_comp' has been removed in r195562 (cf. PR 54107), with the consequence that the original ICE on comment 0 is fixed in 4.8.0 and above.

Also the double error message is reduced to a single one in 4.8.

For comment 2, 4.8 actually gives an ICE in mio_expr, but it compiles cleanly with trunk, so I'm closing this whole PR as fixed.

Well, a very similar case, with a non-abstract type throws an ICE with gfortran 4.8.1:

module Foo_class
    type :: Foo
        integer :: n
    contains
        procedure :: n2  => Foo_n2
        procedure :: vec => Foo_vec
    end type

contains

    pure function Foo_n2(self)
        class(Foo), intent(in) :: self
        integer                :: Foo_n2

        Foo_n2 = self%n**2
    end function

    function Foo_vec(self)
        class(Foo) :: self
        real       :: Foo_vec(self%n2())

        Foo_vec=1.0
    end function

end module

(In reply to Sergio Losilla from comment #6)
> Well, a very similar case, with a non-abstract type throws an ICE with
> gfortran 4.8.1:

Right. This comes from mio_expr and seems to be similar or identical to the ICE that 4.8 throws on comment 2. But both of them are gone with current trunk.