Silicon Graphics IRIS Workstation, User Manual, Page: 94 / 170

90 IRIS WORKSTATION GUIDE
Version 1.0
Appendix E:
In addition, the standard specifies that the user should have control over how
rounding is to occur. The user may opt to round toward nearest, toward zero,
∞ or toward -∞.
Because of the user-control provided, the IEEE standard is quite difficult to
implement and can suffer from unnecessary inefficiency due to the large
number of special cases. The IRIS implementation provides user-control only
over exception conditions. The other modes specified as user-selectable are
implemented as follows:
• Affine mapping for infinities is used. In this mode, -∞ < any
number < +∞. (The default specified in the standard is projective ,
in which infinities compare equal regardless of sign, and do not
compare to other numbers.)
• Normalizing mode is used for arithmetic. In this mode,
denormalized numbers are normalized (in extended precision) before
operations are performed on them. The default specified in the
IEEE standard is warning mode, in which NaN is generated when a
denormalized number is used in an arithmetic operation. Warning
mode is not implemented.
• Rounding to nearest is implemented using at least seven guard bits.
No sticky bit is used. Directed rounding modes are not
implemented.
The implementation of exception handling is treated in Section E.3.
E.3 The C Floating Point Implementation
IRIS Floating Point Types
The C language specification
2
includes two types for floating point - float
(single-precision floating point) and double . It specifies that both of these types
must be implemented, although they may be synonymous. In some floating
point formats (notably IEEE), conversion between these types is simple,
involving only the addition or removal of mantissa bits. Such conversion can
easily be done in software with in-line code. In the IEEE floating point format,
however, conversion between these types is expensive, as their formats differ
considerably.
Numbers of type
float in C are perhaps best thought of as abbreviated forms of
the true floating point format— double . This is seen in the following two rules:
2. Kernighan, B. W. and Ritchie. D.M. The C Programming Language , Prentice-Hall, 1978.