CHAPTER 5: Numeric Types

5.1 Introduction

Fortran is widely used for scientific computation and compilers are available on almost all machines. Several large packages of numerical routines of a high professional standard, such as the library of subroutines of the Numerical Algorithms Group (NAG), have been implemented in Fortran and made available on a wide range of computers. Nevertheless, numerous defects can easily trap the unwary. For example, when a floating point value is assigned to an integer variable the value is truncated; this obvious trap is compounded by the lack of any definition of this effect - the standard does not say whether -3.8 truncates to -3 or to -4, that is, whether the sign is considered after truncation or with it. Moreover Fortran provides no facilities for fixed point arithmetic, for which there is a particular need on computers without floating point hardware.

5.1.1 Floating Point: The Problems

The most difficult area is the control of floating point precision, for which no entirely adequate solution is available. Fortran does not define the accuracy of single precision values. Consequently, the number of bits in the mantissa of a single precision value can be 48 on one system and 24 on another; to achieve a given precision, say 30 bits, one would have to specify single precision on the first system but double precision on the second.

To change the precision for a Fortran 66 program is extremely awkward, and requires a careful review of the program text: the exponents of floating point literals must be changed, all intrinsic functions must be altered, and so on; some functions such as FLOAT have no double length counterpart. The Numerical Algorithms Group overcame this by an elaborate text processing package [HF 76]. By adopting suitable programming conventions, most of the changes can be made with a simple text edit, but there is no simple complete solution. For instance, use of double length throughout is not effective because of its excessive cost, changing the type by IMPLICIT is not standard Fortran, and in any case IMPLICIT cannot be used for literals.

Changing precision is much easier for a Fortran 77 program, because many of the problems identified above have been eliminated. Some problems remain, however: thus it is still not possible to specify the precision of a type explicitly - say in decimal digits. Moreover the change from single to double precision is sometimes difficult: for instance single length COMPLEX is a defined data type but DOUBLE COMPLEX is not. It should be noted that the proposed standard Fortran 8X attempts to overcome all of these problems, and others, and in consequence has features similar to those of Ada.

Several languages in the Algol 60 tradition, such as Pascal, Coral 66 and RTL/2, admit only one floating point data type. In some cases this simple solution meets the users' requirements better than Fortran does. The two Algol 60 compilers for the IBM 360 provide a directive to specify 32 or 64 bit precision - substantially easier to change than the corresponding precision in Fortran. In essence, unless the declarations can determine different precisions, it is best to use the same precision for all floating point quantities, and therefore to have only one floating-point data type in the language.

Control of precision in Algol 68 is by declaration of types real, long real, or even long long real. Although the precision of real is implementation-dependent, so that declarative changes to a program may still be needed in order to maintain the required accuracy when moving it from one implementation to another, these changes are rather easy.

Any language that has user-defined types, and some method of controlling precision, has the essential mechanism for an effective solution of this problem. It is, of course, imperative that the programmer use the typing facility in such a way that the floating point declarations can easily be remapped when a change of precision is needed.

5.1.2 Fixed Point: The Problems

... fixed point numbers shall be treated as exact numeric values. ... The scale or step size (i.e., the minimal representable difference between values) of each fixed point variable must be ... determinable during translation. Scales shall not be restricted to powers of two.

To illustrate these problems, let us consider the case of calculations on electrical insulation, using Ohm's Law: current multiplied by resistance equals voltage. Suppose that we measure the leakage current to an accuracy of one milliampere, and adopt this as the step size or scale of a variable LEAKAGE. This means that only whole numbers of milliamperes can be represented: the value of LEAKAGE will always be an integer L times the scale of 0.001 amperes. In like fashion we may measure the resistance of the insulation to an accuracy of 1000 ohms, and use a variable RESISTANCE whose value will always be an integer R times the scale of 1000 ohms, that is, R kilo-ohms.

Now the potential supported by the insulation is the value of LEAKAGE*RESISTANCE, and because the scale factors happen to cancel it will be LR volts. This is again an integer, but we cannot simply assign it to a third fixed point variable POTENTIAL having scale factor one volt, and treat this variable in the same way as the others, because only a subset of the possible values of POTENTIAL can arise in this way. Thus a given value of POTENTIAL, say P volts (an integer) cannot be divided by a given value of RESISTANCE, say R kilo- ohms, to get L milliamperes exactly (which must be an integer) because P/R will usually not be an integer. In addition there are size problems because single length factors give a double length product. The Ironman requirements [DoD 77] recognized this, and required built- in operations for integer and fixed point division with remainder. This would allow a double length representation of P to be divided by R to yield an integer quotient L1 and integer remainder L2, each single length:

    P = R*L1 + L2

and it would be in the hands of the programmer to ignore or use L2 as he wished. The operation would be exact, and L1 could be assigned to LEAKAGE for further use, as a quantity whose inaccuracy was known.

Cobol apparently meets the Ironman requirements, but only by using decimal scales, which are not adequate for two reasons. First, this is not necessarily the scaling required by the application, and secondly, 10 is too coarse for the standard 16-bit minicomputer. A glance at a Cobol manual will also indicate that explaining the implicit decimal point to the programmer is not easy.

In view of the difficulty of providing exact fixed point computation to meet the Steelman requirements, we considered what was really needed by the users. An analysis of actual applications in many real- time situations revealed that there was a need for cheap approximate computation. Small but frequently executed computations are performed upon digital input signals. Simple machines do not have floating point hardware, and emulation of floating point operations by software or firmware is not fast enough, hence some other means is required to perform approximate computations rapidly on such machines. To say that in the future floating point hardware will always be available may not be the answer: source data input is inevitably captured in fixed point representation, and floating point representation requires more space. Hence approximate fixed point is better matched than floating point to the needs of common applications.

It must be admitted that, as we shall see, programming with fixed point is much more difficult than with floating point. On the other hand, fixed point is potentially more reliable because effective numerical error analysis requires tight bounds to be placed upon data values.

It is concluded that approximate fixed point is generally the most useful arithmetic capability to provide that will complement integer and floating point facilities. However, Ada fixed point also provides some exact operations such as addition and subtraction, and these are invaluable, for example for the manipulation of intervals of time.

5.1.3 Overview of Numerics in Ada

The semantics of each numeric operation is determined by the type of its operands. The facility for numerics is based upon three types that cannot be named in a program (and hence are said to be anonymous - no variable of such a type can be declared). These types are referred to as universal_integer, universal_real, and universal_fixed. Any specific type in a given implementation is a partial representation of a universal type.

(1) universal_integer

The type universal_integer is an integer type with a range large enough to encompass every conceivable integer type of the implementation. Integer literals are of type universal_integer.

(2) universal_real

The type universal_real is a real type with a precision that is high enough to encompass any implemented real type. Real literals are of type universal_real.

(3) universal_fixed

The type universal_fixed is introduced as the result type of the unscaled fixed point operations of multiplication and division to be detailed later. It is essentially a type for intermediate results. The universal_fixed type has a finer delta than any implemented fixed point type.

(1) Integer types

There is an implementation-dependent type INTEGER, defined in the package STANDARD. The range of type INTEGER reflects the properties of the underlying hardware, in that the most efficiently handled integer size is used, with a range symmetric about zero (apart from possibly one negative value - see RM 3.5.4(7)). It would have been possible to have designed a language that had no predefined type such as INTEGER, but this would have meant that in order to obtain a type that would give as large a range of intÂger values as possible without losing efficiency, the programmer would Ëave had to use language facilities that were highly system dependent. So to avoid this dependence, it is desirable to have a predefined type that maps onto the integer type that is most efficiently handled by the target computer.

Additionally, an implementation may provide types LONG_INTEGER and SHORT_INTEGER with larger and smaller ranges than INTEGER. The user can define an integer type by stating the required range; this defines a subtype of a type derived from one of these predefined types - the predefined type being suitably selected by the implementation. The most positive and most negative integer values that are supported by an implementation are SYSTEM.MAX_INT and SYSTEM.MIN_INT respectively.

(2) Floating point types

The implementation-defined types FLOAT, LONG_FLOAT (and so on) can be thought of as copies of the type universal_real but with restricted relative accuracy. These types are predefined so as to approximate closely to the intrinsic floating-point types of the target computer. The user may define other floating point types in terms of these machine types, or simply by stating the required precision; in the latter case this defines a subtype of a type derived from one of the predefined types - the predefined type being suitably selected by the implementation.

A hardware floating point representation has two independent parameters - the length of the mantissa and the range of the exponent. The mantissa length defines the relative accuracy, and the exponent range defines the range of the floating point values. In Ada, the user defines a floating point type by stating only the required precision, as a number of decimal digits; this defines the mantissa length. The language requires each floating point type to have an exponent range that is workable in relation to its mantissa length. In this way the two parameters are reduced to one, with a gain in simplicity for the user.

(3) Fixed point types

In contrast to floating point, a fixed point type has an absolute rather than a relative accuracy. This absolute bound on the errors is called the delta. The user can define other fixed point types by specifying the required delta, together with the range of magnitudes to be encompassed.

Constants

LINE_LENGTH          :  constant :=  80;
MID_LINE             :  constant :=  LINE_LENGTH/2;
PI                   :  constant :=  3.14159_26536;
RADIANS_PER_DEGREE   :  constant :=  PI/180;

Without this facility, a change to the program to modify a constant would involve a search for all occurrences of the constant as well as of related constants. This would be both tedious and risky: for example the constant 40 might or might not be intended to signify half the line length, and even with a corresponding comment the process would be error prone.

The type of such a named number depends on the primaries used in the expression on the right; if these yield real values, the type of the constant is universal_real, if they yield integer values, the type is universal_integer; a mixture of real and integer values gives universal_real. Thus the first two examples above are of type universal_integer and the second two are of type universal_real.

A numeric literal is either an integer literal or a real literal. Thus 80 and 2 above are integer literals because they contain no decimal point, while 3.14159_26536 is a real literal. Within an expression, a numeric literal will be implicitly converted to the required type determined by the context - an integer literal to an integer type and a real literal to a real type. For example, implicit conversion is performed in these cases:

J : INTEGER := 2; P : INTEGER := 4*J; A : REAL := REAL(P) - 0.23;

In the second case, because J is of type INTEGER, the integer literal 4 is implicitly converted to INTEGER as an operand of the multiplication, which yields a product of type INTEGER. In the last case, the subtraction must deliver a REAL result, so P needs explicit conversion to the type REAL, but conversion of the real literal 0.23 to REAL is implicit. It should be noted that no accuracy can be lost by such implicit conversion of numeric literals - the accuracy required by the target type is always provided.