Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
Beginner
23 Views

Compilation error #6355: This binary operation is invalid for this data type.

Hello,

Again, my environment:

$ ifort --version
ifort (IFORT) 16.0.1 20151021
Copyright (C) 1985-2015 Intel Corporation.  All rights reserved.

$ uname -a
Linux tfe03 2.6.32-431.29.2.el6.x86_64 #1 SMP Sun Jul 27 15:55:46 EDT 2014 x86_64 x86_64 x86_64 GNU/Linux

I believe the linux is a flavour of Red Hat (RHEL6 I think).

I discovered what I think may be a minor compiler bug due to my overuse of PRIVATE statements in object definitions. Note the use of the PRIVATE statement in the CONTAINS section of the iVar_type object definition below:

MODULE my_define
  IMPLICIT NONE
  PRIVATE
  PUBLIC :: my_type

  ! Contained object definition
  TYPE, PRIVATE :: iVar_type
    INTEGER :: i = 0
  CONTAINS
    PRIVATE  ! *** This is the statement in question ***
    PROCEDURE :: iVar_Equal
    PROCEDURE :: iVar_NotEqual
    PROCEDURE :: iVar_Compare
    GENERIC :: OPERATOR(==) => iVar_Equal
    GENERIC :: OPERATOR(/=) => iVar_NotEqual
    GENERIC :: OPERATOR(.Compare.) => iVar_Compare
  END TYPE iVar_type

  ! The main object definition
  TYPE :: my_type
    INTEGER :: j = 0
    ! Contained object
    TYPE(iVar_type) :: iVar
  CONTAINS
    PRIVATE
    PROCEDURE :: Equal
    PROCEDURE :: NotEqual
    PROCEDURE :: Compare_
    GENERIC, PUBLIC :: OPERATOR(==) => Equal
    GENERIC, PUBLIC :: OPERATOR(/=) => NotEqual
    GENERIC, PUBLIC :: OPERATOR(.Compare.) => Compare_
  END TYPE my_type

CONTAINS

  ELEMENTAL FUNCTION Equal( x, y ) RESULT( is_equal )
    CLASS(my_type), INTENT(IN) :: x, y
    LOGICAL :: is_equal
    is_equal = (x%j == y%j) .AND. (x%iVar == y%iVar)
  END FUNCTION Equal

  ELEMENTAL FUNCTION iVar_Equal( x, y ) RESULT( is_equal )
    CLASS(iVar_type), INTENT(IN) :: x, y
    LOGICAL :: is_equal
    is_equal = (x%i == y%i)
  END FUNCTION iVar_Equal


  ELEMENTAL FUNCTION NotEqual( x, y ) RESULT( not_equal )
    CLASS(my_type), INTENT(IN) :: x, y
    LOGICAL :: not_equal
    not_equal = .NOT. (x == y)
  END FUNCTION NotEqual

  ELEMENTAL FUNCTION iVar_NotEqual( x, y ) RESULT( not_equal )
    CLASS(iVar_type), INTENT(IN) :: x, y
    LOGICAL :: not_equal
    not_equal = .NOT. (x == y)
  END FUNCTION iVar_NotEqual


  FUNCTION Compare_( x, y ) RESULT( is_equal )
    CLASS(my_type), INTENT(IN) :: x, y
    LOGICAL :: is_equal
    is_equal = .TRUE.
    IF ( x%j /= y%j ) THEN
      print *, 'J component of my objects are different'
      is_equal = .FALSE.
    END IF
    IF ( x%iVar /= y%iVar ) THEN
      print *, 'iVar component of my objects are different'
      is_equal = .FALSE.
    END IF
  END FUNCTION Compare_

  FUNCTION iVar_Compare( x, y ) RESULT( is_equal )
    CLASS(iVar_type), INTENT(IN) :: x, y
    LOGICAL :: is_equal
    is_equal = .TRUE.
    IF ( x%i /= y%i ) THEN
      print *, 'I component of iVar objects are different'
      is_equal = .FALSE.
    END IF
  END FUNCTION iVar_Compare
  
END MODULE my_define



! =======================
! Test program for module
! =======================

PROGRAM Test_my
  USE my_define, ONLY: my_type
  IMPLICIT NONE

  LOGICAL :: is_equal
  TYPE(my_type) :: my, my_copy

  my%j = 1
  my_copy = my
  IF ( my /= my_copy ) THEN
    is_equal = my .Compare. my_copy
    print *, 'objects are not equal'
  ELSE
    print *, 'objects are equal'
  END IF
END PROGRAM Test_my

When I compile the above I get:

$ ifort two_types.f90 

two_types.f90(39): error #6355: This binary operation is invalid for this data type.   [IVAR]
    is_equal = (x%j == y%j) .AND. (x%iVar == y%iVar)
-------------------------------------^
two_types.f90(39): error #6355: This binary operation is invalid for this data type.   [IVAR]
    is_equal = (x%j == y%j) .AND. (x%iVar == y%iVar)
-----------------------------------------------^
two_types.f90(58): error #6355: This binary operation is invalid for this data type.   
    not_equal = .NOT. (x == y)
-----------------------^
two_types.f90(58): error #6355: This binary operation is invalid for this data type.   
    not_equal = .NOT. (x == y)
----------------------------^
two_types.f90(70): error #6355: This binary operation is invalid for this data type.   [IVAR]
    IF ( x%iVar /= y%iVar ) THEN
-----------^
two_types.f90(70): error #6355: This binary operation is invalid for this data type.   [IVAR]
    IF ( x%iVar /= y%iVar ) THEN
---------------------^
compilation aborted for two_types.f90 (code 1)

When I comment out the PRIVATE statement in question the test code compiles and runs fine:

$ diff two_types.f90 two_types_commented.f90 
--- two_types.f90	2016-01-25 13:19:42.382898000 +0000
+++ two_types_commented.f90	2016-01-25 13:19:31.705164000 +0000
@@ -7,7 +7,7 @@
   TYPE, PRIVATE :: iVar_type
     INTEGER :: i = 0
   CONTAINS
-    PRIVATE  ! *** This is the statement in question ***
+!    PRIVATE  ! *** This is the statement in question ***
     PROCEDURE :: iVar_Equal
     PROCEDURE :: iVar_NotEqual
     PROCEDURE :: iVar_Compare

$ ifort two_types_commented.f90 
$ ./a.out
 objects are equal

So, while I have a solution to the problem in my production code, should the ifort compiler toss an error due to this extra PRIVATE statement?

Note that the test case compiles and runs using gofrtran 4.9.x

Thanks for any info.

cheers,

paulv

0 Kudos
7 Replies
Highlighted
23 Views

What happens with

What happens with

is_equal = (x%j == y%j) .AND. ((x%iVar) == (y%iVar)) ! add ()'s

There was another thread on idz forum relating to a constructor issue where the additional ()'s resolved the problem.
I cannot recall if the additional ()'s were not required by the specification. The use of the additional ()'s may (stressed) work as a work-around, but will also add some additional overhead in creating temporaries.

Jim Dempsey

0 Kudos
Highlighted
Black Belt
23 Views

I've tripped over cases where

I've tripped over cases where fpp requires parentheses not normally required in Fortran, so I won't rule out possible strange effects of parentheses.  Reminder : ifort doesn't necessarily follow standard unless protect_parents or standard-semantics is set.  I don't agree with Jim about parentheses necessarily creating a temporary.

0 Kudos
Highlighted
Beginner
23 Views

Quote:jimdempseyatthecove

jimdempseyatthecove wrote:

What happens with

is_equal = (x%j == y%j) .AND. ((x%iVar) == (y%iVar)) ! add ()'s

There was another thread on idz forum relating to a constructor issue where the additional ()'s resolved the problem.
I cannot recall if the additional ()'s were not required by the specification. The use of the additional ()'s may (stressed) work as a work-around, but will also add some additional overhead in creating temporaries.

Hi Jim,

Thanks for the info but, nope, no difference:

tfe03:/home/Paul.Vandelst/scratch/intel-bug : ifort two_types.f90 
two_types.f90(39): error #6355: This binary operation is invalid for this data type.   [IVAR]
    is_equal = (x%j == y%j) .AND. ((x%iVar) == (y%iVar))
--------------------------------------^
two_types.f90(39): error #6355: This binary operation is invalid for this data type.   [IVAR]
    is_equal = (x%j == y%j) .AND. ((x%iVar) == (y%iVar))
--------------------------------------------------^

I need to check but I would be very surprised if the standard required the extra parentheses. That, to me at least, would suggest the structure separator, %, has a lower precedence than the operator. Which doesn't make sense...or does it? I'm having a hard time wrapping my head around how that would work? (Not enough coffee this morning....after being trapped inside most of the weekend due to snow)

cheers,

paulv

0 Kudos
Highlighted
23 Views

Thanks for the report - we'll

Thanks for the report - we'll check it out.

Retired 12/31/2016
0 Kudos
Highlighted
23 Views

Paul, thanks for this one -

Paul, thanks for this one - seems very odd that accessibility changes would affect uses in the same module! Escalated as issue DPD200381114.

Retired 12/31/2016
0 Kudos
Highlighted
Beginner
23 Views

Quote:Steve Lionel (Intel)

Steve Lionel (Intel) wrote:

Paul, thanks for this one - seems very odd that accessibility changes would affect uses in the same module! Escalated as issue DPD200381114.

And thanks again.

0 Kudos
Highlighted
23 Views

I expect the fix for this to

I expect the fix for this to appear in Parallel Studio XE 2016 Update 3.

Retired 12/31/2016
0 Kudos