[Ada Information Clearinghouse]

"Rationale for the Design of the
Ada® Programming Language"

[Ada '83 Rationale, HTML Version]

Copyright ©1986 owned by the United States Government. All rights reserved.
Direct inquiries to the Ada Information Clearinghouse at adainfo@sw-eng.falls-church.va.us.

CHAPTER 5: Numeric Types

5.4 Implementation Considerations

Fixed point types can be represented on most machines with one or two machine words. Implementations should not support fixed point types in excess of this length if credible performance is required (and cannot be provided). Note that the predefined type DURATION requires 23 bits of accuracy (plus one sign bit), because the reference manual requires that intervals of at least twenty milliseconds be accommodated, as well as an interval of up to a day. Such an implementation would be as efficient in time and space as conventional assembler coding (assuming a good register allocation algorithm). As with subranges of integers, tight packing is possible, and this could result in a major advantage over floating point.

Good performance depends largely upon the proper specification of the range and delta. For machines with limited arithmetic shifts, a value in a range that excludes negative values could have its scale converted by the use of logical shifts. All type conversions with fixed types could be accomplished with simple arithmetic shifts and masking. All the operations are likewise straightforward.

With real types there is a problem about the end points of the range with a range constraint. Ordinarily, such values would be in the set of values permitted. However, a fixed point range 0.0 .. 1.0 on a two's complement machine would not usually want to include 1.0. To avoid giving ranges as 0.0 .. 0.999999999, it seems best not to require that nonzero end values be within the specified range. See the wording in RM 3.5.9(6) for details.

A lazy implementation of numeric types which could be used by a diagnostic compiler is as follows. Every value is stored in long floating point format together with a flag indicating if it is integer or real. The long format must be sufficiently long to encompass the longest integer, fixed point and floating point types supported by the implementation. Operations can now be applied to these values, the flag being used to ensure that integer results are correctly rounded to integers (if the floating point hardware does not give integer results from integer values). This implementation method is clearly inefficient, especially for fixed point types, which are often used as a method of avoiding expensive floating point. However, it illustrates the concept of the abstract value and the fact that the operators have the same meaning for each type.

Although it is theoretically feasible, it is not practical to implement floating point types as fixed point quantities. This is because of the potentially large dynamic range of floating point values - a floating point mantissa length B would need small = 2**(- 5*B) and hence fixed point mantissa length 9*B.

With the real types, the language does not specify rounding or truncation, since either choice could be excessively expensive on some machines. The user can control its effect by increasing the digits or decreasing the delta in the type declaration, but should note that a small decrease in the delta could require going from 1 word to 2 words, with consequent performance degradation. With multiplication and division, rounding may be required in order to preserve the relational inequalities. Exact conversion can only occur between integer types (although many other conversions may not require any rounding). No conversions are significantly more troublesome than are integer to real and real to integer in (say) Algol 60.

Consider a function ROOT for taking the square root of an argument, where the argument and the result are of type FRACTION:

    type FRACTION is delta D range 0.0 .. 1.0;

By declaring the fixed point quantities X and Y to have a type with larger delta:
type SIXTEEN is delta 16.0*D range 0.0 .. 16.0;

one can then take the square root of X by

    Y :=  SIXTEEN(4 * ROOT(FRACTION (X/16)));

and here the division by 16 is a shift that corresponds to the converse of the FRACTION type conversion and hence produces no code (assuming reasonable peephole optimization). (Note that the literals 4 and 16 are integers; real literals would not be allowed here.)

The body of ROOT (for argument range 0.5 to 1) could be:
function ROOT(X :  FRACTION) return FRACTION is
  HALF :  constant FRACTION :=  0.5;
  APPROX :  FRACTION  :=  0.7;           -- a starting value
  end loop;
  return APPROX;
end ROOT;

The machine dependence is largely restricted to the declaration of FRACTION, whose range relative to the accuracy would reflect the word length of the machine. Note that since the declared range is 0.0 .. 1.0 the algorithm may give values equal to 1.0 for arguments near 1.0. This would cause overflow on a two's complement machine. The check for negative arguments is implicit in the type definition.

In evaluating an expression at compilation time, the identification of the operators must be performed. Then expressions involving only literals, constants, other evaluated expressions and predefined operators can be evaluated. The accuracy of real arithmetic may here be different from that of the target machine, although both are within that specified in the type declarations.

The efficient implementation of some mathematical algorithms requires access to the component parts of a floating point value. For instance, a square root routine typically starts by estimating a value by dividing the exponent by 2. We therefore need to be able to access and update the exponent. This can elegantly be done in Ada by the use of a record type and associated record representation clause defining the internal structure of the floating point value, and then using UNCHECKED_CONVERSION to convert between the floating point type and the record type. Such operations are best performed by subprograms in the mathematical library; in this way the implementation dependent operations can be encapsulated so that they are hidden from the normal user.

The above example illustrates the essential dilemma between efficiency and portability that intrudes into certain sensitive areas of numerical computation; there are occasions where the demands for a very efficient implementation outweigh those for complete portability. The facilities in Ada enable the non-portable parts to be readily identified and encapsulated so that a proper balance between the conflicting aims can be obtained.

Address any questions or comments to adainfo@sw-eng.falls-church.va.us.