C Pocket Reference

Optimizing Runtime Efficiency

ANSI C99 has introduced features to optimize the efficiency of floating-point operations.

The types float_t and double_t , defined in math.h, represent the types used internally in floating-point arithmetic. When these types are used in a program, no conversions are necessary before arithmetic operations are performed. The macro FLT_EVAL_METHOD indicates what the equivalent basic types are, and returns one of the values described in Table 1-33.

Table 1-33. Interpretation of float_t and double_t

FLT_EVAL_METHOD	Type represented by float_t	Type represented by double_t
0	float	double
1	double	double
2	long double	long double

CPUs may have special machine instructions to perform standard arithmetic operations quickly. Rounding and error conditions may also be ignored. Optimizations of these kinds can be enabled by the pragma FP_CONTRACT . For example:

#pragma  STDC FP_CONTRACT ON

The same pragma with the switch OFF rather than ON disables such optimizations.

Furthermore, the macro FP_FAST_FMA is defined if the “floating-point multiply-add” function fma( x, y, z ), which returns x*y+z, is implemented as a special instruction, and is thus faster than separate multiplication and addition operations. The macros FP_FAST_FMAF and FP_FAST_FMAL are analogous indicators for the functions fmaf() and fmal().