- 17.1 New features
- 17.2 Deleted features
- 17.3 Obsolescent features
- 17.4 Schedule
- 17.5 Description of the new features

Fortran 95 was published as ISO/IEC 1539-1:1997 on December 15, 1997.

The preliminary revision has been published as "Special Issue,
Fortran 95, Committee Draft, May 1995" in the **Fortran Forum**,
Vol. 12, No. 2, June 1995. The draft is available on the net,
see my Fortran page.

Two important issues, **exception handling**, especially for floating
point, and **interoperability between languages**
(mixed language programming)
are not addressed in Fortran 95. Exception handling is discussed in some papers by
John Reid, see the
proceedings
of the Kyoto Workshop on *Current Directions in
Numerical Software and High Performance Computing.*

- The statement
`FORALL`as an alternative to the`DO`-statement - Partial
nesting of
`FORALL`and`WHERE`statements - Masked
`ELSEWHERE` - Pure procedures
- Elemental procedures
- Pure procedures in specification expressions
- Revised
`MINLOC`and`MAXLOC` - Extensions to
`CEILING`and`FLOOR`with the`KIND`keyword argument - Pointer initialization
- Default initialization of derived type objects
- Increased compatibility with IEEE arithmetic
- A
`CPU_TIME`intrinsic subroutine - A function
`NULL`to nullify a pointer - Automatic deallocation of allocatable arrays at exit of scoping unit
- Comments in
`NAMELIST`at input - Minimal field at input
- Complete version of
`END INTERFACE`

- real and double precision
`DO`loop index variables - branching to
`END IF`from an outer block `PAUSE`statements`ASSIGN`statements and assigned`GO TO`statements and the use of an assigned integer as a`FORMAT`specification- Hollerith editing in
`FORMAT`

It is permitted for a compiler to have extensions to the standard, provided
that these *can* be flagged by the system. It is very common for compilers
to include the two features
(see Appendix 4)
that were removed from Fortran when Fortran 77 was
introduced, and it is expected that this tradition will continue.

The most probable candidates, in our opinion, for removal at the next revision (in the next century) are the first six.

- Arithmetic
`IF`-statement - Terminating several
`DO`-loops on the same statement or terminating the`DO`-loop in some other way than with`CONTINUE`or`END DO` - Alternate return
- Computed
`GO TO`statement - Statement functions
`DATA`statements*among executable statements*- Assumed character length functions
- Fixed form source code
`CHARACTER*`form of`CHARACTER`declaration

To replace `CHARACTER*LENGTH ` with `CHARACTER(LEN=LENGTH) ` or
`CHARACTER(LENGTH) ` is simple, as are the other items above. For example,
item 6 is just to move all `DATA ` statements to the top of the program unit,
before the executable statements.
Statement functions are of course replaced with
internal functions.

- April 1995:
**Technical completion** - End of May 1995:
**Document available** - July - September 1995:
**Review and ISO CD ballot** - November 1995:
**Revision based on ballot comments** - March 1996:
**Revised document available** - 26 January 1997:
**End of the ISO DIS international ballot** - 15 December, 1997:
**Fortran 95 published as ISO/IEC 1539-1:1997**(replacing Fortran 90, ISO/IEC 1539:1991).

#### The statement FORALL

The statement`FORALL`is introduced as an alternative to the`DO`-statement. The main difference is that while the execution order of the various parts of the`DO`-loop is very strict, the execution order of the`FORALL`is less strict, thus permitting parallel execution. For further information we refer to our HPF Appendix or directly to the HPFF home page.#### Partial nesting of FORALL and WHERE statements

A`FORALL`statement can include a`WHERE`statement.#### Masked ELSEWHERE

It is now permitted to mask not only the`WHERE`statement of the`WHERE`construct, but also its`ELSEWHERE`, which now may be repeated.WHERE (condition_1) ... ELSEWHERE (condition_2) ... ELSEWHERE ... END WHERE

#### Pure procedures

Pure functions are functions without side effects. That a function has no side effects can be indicated with the new`PURE`prefix.A long list of constraints is given in the standard proposal, section 12.6.

Pure subroutines are defined in a similar way. The only major difference is that "side effects" are of course permitted for arguments associated with dummy arguments specified

`INTENT(OUT)`or`INTENT(INOUT)`.The advantage with knowing that a function is pure is that this fact simplifies parallel execution.

#### Elemental procedures

Elemental functions are pure functions with only scalar dummy arguments (not pointers or procedures) and with scalar result (not pointer). That a function is elemental can be indicated with the new`ELEMENTAL`prefix. The`RECURSIVE`prefix my not be combined with the`ELEMENTAL`prefix. The`PURE`prefix is automatically implied by the`ELEMENTAL`prefix.An elemental function may be used with arrays as actual arguments. These must then be conformable. The result is the same array as if the function had been used individually with each of the array elements, in any order.

Elemental subroutines are defined in a similar way. The only major difference is that "side effects" are of course permitted for arguments associated with dummy arguments specified

`INTENT(OUT)`or`INTENT(INOUT)`.The advantage with knowing that a function is elemental is that this fact simplifies parallel execution, even more than if it is only pure.

#### Pure procedures in specification expressions

Pure functions can be used in specification expressions if certain conditions are fulfilled, see the standard proposal, section 7.1.6.2.Specification expressions can be used to specify array bounds and character lengths of data objects in a subprogram.

#### Revised MINLOC and MAXLOC

The array location functions`MINLOC`and`MAXLOC`are extended with the optional argument`DIM`corresponding to those for the array functions`MINVAL`and`MAXVAL`.#### Extensions to CEILING and FLOOR

The new numerical functions`CEILING`and`FLOOR`are extended with the`KIND`keyword argument, in the same way as for`INT`and`NINT`. The result is an integer, but of the specified`KIND`or subtype, not necessarily the standard (default) integer subtype.#### Pointer initialization

The new function`NULL`can be used at specification to define a pointer to be initially disassociated, see below.#### Default initialization of derived type objects

Means are now available to specify default initial values for derived type components. It is the usual way of using the equal sign (or pointer assignment) followed by the value (perhaps an array constructor). Initialization does not have to apply to all components of a certain derived type.A simple example. In Appendix 3, section 12, we introduced a sparse matrix.

A numerically interesting example is a sparse matrix

`A`with at most one hundred non-zero elements, which can be specified with the following statement, where now initialization is done to 2.0 for all elements.TYPE NONZERO REAL :: VALUE = 2.0 INTEGER :: ROW, COLUMN END TYPE

andTYPE (NONZERO) :: A(100)

You then get the value (which will be 2.0) of`A(10)`by writing`A(10)%VALUE`. The default value can be individually changed with an assignment, for exampleA(15) = NONZERO(17.0,3,7)

#### Increased compatibility with IEEE arithmetic

The IEEE arithmetic has for floating point numbers one bit pattern for*plus zero*and another one for*minus zero*. Processors that distinguish between them shall treat them as identical- In all relational operations
- As input arguments to all intrinsics except
`SIGN` - As the scalar expression in the arithmetic
`IF`-statement

`SIGN`has to be used. It is generalized so that the sign of the second argument is considered also if the value is a floating-point zero.#### A CPU_TIME intrinsic subroutine

The subroutine

`CPU_TIME(TIME)`belongs of course to intrinsic subroutines. In the scalar real variable`TIME`the present processor time is returned in seconds. If the processor is unable to provide a timing, a negative value is returned instead. As usual, the time for a certain computation is obtained by subtracting two different calls to the timing routine.The exact nature of the timing is implementation dependent, a parallel processor might reurn an array of times corresponding to the various processors. The difference between CPU-time and system time is also implementation dependent.

#### A function NULL to nullify a pointer

This function can be used at specification to define a pointer to be initially disassociated, in the example below the array`VECTOR`.REAL, POINTER, DIMENSION(:) :: VECTOR => NULL()

The argument is not necessary, if present it determines the characteristics of the pointer, if not present, the characteristics are determined from context.The function belongs to section 20, pointer inquiry functions, which is now renamed "Pointer association status functions".

#### Automatic deallocation of allocatable arrays at exit of scoping unit

If the user has not explicitly deallocated local allocatable arrays at the exit of the scoping unit, this deallocation is now performed automatically, thus decreasing the memory required.#### Comments in NAMELIST at input

It is now permitted to use comments in the usual way with`!`#### Minimal field at output

In order to obtain an optimized use of the available positions it is now possible to only give the number of decimals, if any, and not the total field width, at the`FORMAT`s`B, F, I, O`, and`Z`. Examples are`I0`and`F0.6`. The result is of course not a field with zero digits, but with a suitable number of digits.#### Complete version of END INTERFACE

Also`END INTERFACE`can now be given in a complete variant, so the second interface in chapter 10 can be given either in the old version asINTERFACE SWAP MODULE PROCEDURE SWAP_R, SWAP_I, SWAP_C END INTERFACE

or in the new preferred way asINTERFACE SWAP MODULE PROCEDURE SWAP_R, SWAP_I, SWAP_C END INTERFACE SWAP

This can also include an optional generic specification.

Last modified: 11 September 1998