Fortran For C Programmers

This note is limited to essential information about Fortran so that a C or C++ programmer can get started more quickly with the language, at least as a reader, and avoid some common pitfalls when starting to write or modify Fortran code. Please see other sources to learn about Fortran’s rich history, current applications, and modern best practices in new code.

Know This At Least

  • There have been many implementations of Fortran, often from competing vendors, and the standard language has been defined by U.S. and international standards organizations. The various editions of the standard are known as the ‘66, ‘77, ‘90, ‘95, 2003, 2008, and (now) 2018 standards.

  • Forward compatibility is important. Fortran has outlasted many generations of computer systems hardware and software. Standard compliance notwithstanding, Fortran programmers generally expect that code that has compiled successfully in the past will continue to compile and work indefinitely. The standards sometimes designate features as being deprecated, obsolescent, or even deleted, but that can be read only as discouraging their use in new code – they’ll probably always work in any serious implementation.

  • Fortran has two source forms, which are typically distinguished by filename suffixes. foo.f is old-style “fixed-form” source, and foo.f90 is new-style “free-form” source. All language features are available in both source forms. Neither form has reserved words in the sense that C does. Spaces are not required between tokens in fixed form, and case is not significant in either form.

  • Variable declarations are optional by default. Variables whose names begin with the letters I through N are implicitly INTEGER, and others are implicitly REAL. These implicit typing rules can be changed in the source.

  • Fortran uses parentheses in both array references and function calls. All arrays must be declared as such; other names followed by parenthesized expressions are assumed to be function calls.

  • Fortran has a lot of built-in “intrinsic” functions. They are always available without a need to declare or import them. Their names reflect the implicit typing rules, so you will encounter names that have been modified so that they have the right type (e.g., AIMAG has a leading A so that it’s REAL rather than INTEGER).

  • The modern language has means for declaring types, data, and subprogram interfaces in compiled “modules”, as well as legacy mechanisms for sharing data and interconnecting subprograms.

A Rosetta Stone

Fortran’s language standard and other documentation uses some terminology in particular ways that might be unfamiliar.

Fortran

English

Association

Making a name refer to something else

Assumed

Some attribute of an argument or interface that is not known until a call is made

Companion processor

A C compiler

Component

Class member

Deferred

Some attribute of a variable that is not known until an allocation or assignment

Derived type

C++ class

Dummy argument

C++ reference argument

Final procedure

C++ destructor

Generic

Overloaded function, resolved by actual arguments

Host procedure

The subprogram that contains a nested one

Implied DO

There’s a loop inside a statement

Interface

Prototype

Internal I/O

sscanf and snprintf

Intrinsic

Built-in type or function

Polymorphic

Dynamically typed

Processor

Fortran compiler

Rank

Number of dimensions that an array has

SAVE attribute

Statically allocated

Type-bound procedure

Kind of a C++ member function but not really

Unformatted

Raw binary

Data Types

There are five built-in (“intrinsic”) types: INTEGER, REAL, COMPLEX, LOGICAL, and CHARACTER. They are parameterized with “kind” values, which should be treated as non-portable integer codes, although in practice today these are the byte sizes of the data. (For COMPLEX, the kind type parameter value is the byte size of one of the two REAL components, or half of the total size.) The legacy DOUBLE PRECISION intrinsic type is an alias for a kind of REAL that should be more precise, and bigger, than the default REAL.

COMPLEX is a simple structure that comprises two REAL components.

CHARACTER data also have length, which may or may not be known at compilation time. CHARACTER variables are fixed-length strings and they get padded out with space characters when not completely assigned.

User-defined (“derived”) data types can be synthesized from the intrinsic types and from previously-defined user types, much like a C struct. Derived types can be parameterized with integer values that either have to be constant at compilation time (“kind” parameters) or deferred to execution (“len” parameters).

Derived types can inherit (“extend”) from at most one other derived type. They can have user-defined destructors (FINAL procedures). They can specify default initial values for their components. With some work, one can also specify a general constructor function, since Fortran allows a generic interface to have the same name as that of a derived type.

Last, there are “typeless” binary constants that can be used in a few situations, like static data initialization or immediate conversion, where type is not necessary.

Arrays

Arrays are not types in Fortran. Being an array is a property of an object or function, not of a type. Unlike C, one cannot have an array of arrays or an array of pointers, although can can have an array of a derived type that has arrays or pointers as components. Arrays are multidimensional, and the number of dimensions is called the rank of the array. In storage, arrays are stored such that the last subscript has the largest stride in memory, e.g. A(1,1) is followed by A(2,1), not A(1,2). And yes, the default lower bound on each dimension is 1, not 0.

Expressions can manipulate arrays as multidimensional values, and the compiler will create the necessary loops.

Allocatables

Modern Fortran programs use ALLOCATABLE data extensively. Such variables and derived type components are allocated dynamically. They are automatically deallocated when they go out of scope, much like C++’s std::vector<> class template instances are. The array bounds, derived type LEN parameters, and even the type of an allocatable can all be deferred to run time. (If you really want to learn all about modern Fortran, I suggest that you study everything that can be done with ALLOCATABLE data, and follow up all the references that are made in the documentation from the description of ALLOCATABLE to other topics; it’s a feature that interacts with much of the rest of the language.)

I/O

Fortran’s input/output features are built into the syntax of the language, rather than being defined by library interfaces as in C and C++. There are means for raw binary I/O and for “formatted” transfers to character representations. There are means for random-access I/O using fixed-size records as well as for sequential I/O. One can scan data from or format data into CHARACTER variables via “internal” formatted I/O. I/O from and to files uses a scheme of integer “unit” numbers that is similar to the open file descriptors of UNIX; i.e., one opens a file and assigns it a unit number, then uses that unit number in subsequent READ and WRITE statements.

Formatted I/O relies on format specifications to map values to fields of characters, similar to the format strings used with C’s printf family of standard library functions. These format specifications can appear in FORMAT statements and be referenced by their labels, in character literals directly in I/O statements, or in character variables.

One can also use compiler-generated formatting in “list-directed” I/O, in which the compiler derives reasonable default formats based on data types.

Subprograms

Fortran has both FUNCTION and SUBROUTINE subprograms. They share the same name space, but functions cannot be called as subroutines or vice versa. Subroutines are called with the CALL statement, while functions are invoked with function references in expressions.

There is one level of subprogram nesting. A function, subroutine, or main program can have functions and subroutines nested within it, but these “internal” procedures cannot themselves have their own internal procedures. As is the case with C++ lambda expressions, internal procedures can reference names from their host subprograms.

Modules

Modern Fortran has good support for separate compilation and namespace management. The module is the basic unit of compilation, although independent subprograms still exist, of course, as well as the main program. Modules define types, constants, interfaces, and nested subprograms.

Objects from a module are made available for use in other compilation units via the USE statement, which has options for limiting the objects that are made available as well as for renaming them. All references to objects in modules are done with direct names or aliases that have been added to the local scope, as Fortran has no means of qualifying references with module names.

Arguments

Functions and subroutines have “dummy” arguments that are dynamically associated with actual arguments during calls. Essentially, all argument passing in Fortran is by reference, not value. One may restrict access to argument data by declaring that dummy arguments have INTENT(IN), but that corresponds to the use of a const reference in C++ and does not imply that the data are copied; use VALUE for that.

When it is not possible to pass a reference to an object, or a sparse regular array section of an object, as an actual argument, Fortran compilers must allocate temporary space to hold the actual argument across the call. This is always guaranteed to happen when an actual argument is enclosed in parentheses.

The compiler is free to assume that any aliasing between dummy arguments and other data is safe. In other words, if some object can be written to under one name, it’s never going to be read or written using some other name in that same scope.

  SUBROUTINE FOO(X,Y,Z)
  X = 3.14159
  Y = 2.1828
  Z = 2 * X ! CAN BE FOLDED AT COMPILE TIME
  END

This is the opposite of the assumptions under which a C or C++ compiler must labor when trying to optimize code with pointers.

Overloading

Fortran supports a form of overloading via its interface feature. By default, an interface is a means for specifying prototypes for a set of subroutines and functions. But when an interface is named, that name becomes a generic name for its specific subprograms, and calls via the generic name are mapped at compile time to one of the specific subprograms based on the types, kinds, and ranks of the actual arguments. A similar feature can be used for generic type-bound procedures.

This feature can be used to overload the built-in operators and some I/O statements, too.

Polymorphism

Fortran code can be written to accept data of some derived type or any extension thereof using CLASS, deferring the actual type to execution, rather than the usual TYPE syntax. This is somewhat similar to the use of virtual functions in c++.

Fortran’s SELECT TYPE construct is used to distinguish between possible specific types dynamically, when necessary. It’s a little like C++17’s std::visit() on a discriminated union.

Pointers

Pointers are objects in Fortran, not data types. Pointers can point to data, arrays, and subprograms. A pointer can only point to data that has the TARGET attribute. Outside of the pointer assignment statement (P=>X) and some intrinsic functions and cases with pointer dummy arguments, pointers are implicitly dereferenced, and the use of their name is a reference to the data to which they point instead.

Unlike C, a pointer cannot point to a pointer per se, nor can they be used to implement a level of indirection to the management structure of an allocatable. If you assign to a Fortran pointer to make it point at another pointer, you are making the pointer point to the data (if any) to which the other pointer points. Similarly, if you assign to a Fortran pointer to make it point to an allocatable, you are making the pointer point to the current content of the allocatable, not to the metadata that manages the allocatable.

Unlike allocatables, pointers do not deallocate their data when they go out of scope.

A legacy feature, “Cray pointers”, implements dynamic base addressing of one variable using an address stored in another.

Preprocessing

There is no standard preprocessing feature, but every real Fortran implementation has some support for passing Fortran source code through a variant of the standard C source preprocessor. Since Fortran is very different from C at the lexical level (e.g., line continuations, Hollerith literals, no reserved words, fixed form), using a stock modern C preprocessor on Fortran source can be difficult. Preprocessing behavior varies across implementations and one should not depend on much portability. Preprocessing is typically requested by the use of a capitalized filename suffix (e.g., “foo.F90”) or a compiler command line option. (Since the F18 compiler always runs its built-in preprocessing stage, no special option or filename suffix is required.)

“Object Oriented” Programming

Fortran doesn’t have member functions (or subroutines) in the sense that C++ does, in which a function has immediate access to the members of a specific instance of a derived type. But Fortran does have an analog to C++’s this via type-bound procedures. This is a means of binding a particular subprogram name to a derived type, possibly with aliasing, in such a way that the subprogram can be called as if it were a component of the type (e.g., X%F(Y)) and receive the object to the left of the % as an additional actual argument, exactly as if the call had been written F(X,Y). The object is passed as the first argument by default, but that can be changed; indeed, the same specific subprogram can be used for multiple type-bound procedures by choosing different dummy arguments to serve as the passed object. The equivalent of a static member function is also available by saying that no argument is to be associated with the object via NOPASS.

There’s a lot more that can be said about type-bound procedures (e.g., how they support overloading) but this should be enough to get you started with the most common usage.

Pitfalls

Variable initializers, e.g. INTEGER :: J=123, are static initializers! They imply that the variable is stored in static storage, not on the stack, and the initialized value lasts only until the variable is assigned. One must use an assignment statement to implement a dynamic initializer that will apply to every fresh instance of the variable. Be especially careful when using initializers in the newish BLOCK construct, which perpetuates the interpretation as static data. (Derived type component initializers, however, do work as expected.)

If you see an assignment to an array that’s never been declared as such, it’s probably a definition of a statement function, which is like a parameterized macro definition, e.g. A(X)=SQRT(X)**3. In the original Fortran language, this was the only means for user function definitions. Today, of course, one should use an external or internal function instead.

Fortran expressions don’t bind exactly like C’s do. Watch out for exponentiation with **, which of course C lacks; it binds more tightly than negation does (e.g., -2**2 is -4), and it binds to the right, unlike what any other Fortran and most C operators do; e.g., 2**2**3 is 256, not 64. Logical values must be compared with special logical equivalence relations (.EQV. and .NEQV.) rather than the usual equality operators.

A Fortran compiler is allowed to short-circuit expression evaluation, but not required to do so. If one needs to protect a use of an OPTIONAL argument or possibly disassociated pointer, use an IF statement, not a logical .AND. operation. In fact, Fortran can remove function calls from expressions if their values are not required to determine the value of the expression’s result; e.g., if there is a PRINT statement in function F, it may or may not be executed by the assignment statement X=0*F(). (Well, it probably will be, in practice, but compilers always reserve the right to optimize better.)

Unless they have an explicit suffix (1.0_8, 2.0_8) or a D exponent (3.0D0), real literal constants in Fortran have the default REAL type – not double as in the case in C and C++. If you’re not careful, you can lose precision at compilation time from your constant values and never know it.