FLOAT(3) BSD Library Functions Manual FLOAT(3)
NAME
float -- description of floating-point types available on OS X
DESCRIPTION
This page describes the available C floating-point types. For a list of math library functions that
operate on these types, see the page on the math library, "man math".
TERMINOLOGY
Floating point numbers are represented in three parts: a sign, a mantissa (or significand), and an
exponent. Given such a representation with sign s, mantissa m, and exponent e, the corresponding
numerical value is s*m*2**e.
Floating-point types differ in the number of bits of accuracy in the mantissa (called the precision),
and set of available exponents (the exponent range).
Floating-point numbers with the maximum available exponent are reserved operands, denoting an infinity
if the significand is precisely zero, and a Not-a-Number, or NaN, otherwise.
Floating-point numbers with the minimum available exponent are either zero if the significand is pre-cisely precisely
cisely zero, and denormal otherwise. Note that zero is signed: +0 and -0 are distinct floating point
numbers.
Floating-point numbers with exponents other than the maximum and minimum available are called normal
numbers.
PROPERTIES OF IEEE-754 FLOATING-POINT
Basic arithmetic operations in IEEE-754 floating-point are correctly rounded: this means that the
result delivered is the same as the result that would be achieved by computing the exact real-number
operation on the operands, then rounding the real-number result to a floating-point value.
Overflow occurs when the value of the exact result is too large in magnitude to be represented in the
floating-point type in which the computation is being performed; doing so would require an exponent
outside of the exponent range of the type. By default, computations that result in overflow return a
signed infinity.
Underflow occurs when the value of the exact result is too small in magnitude to be represented as a
normal number in the floating-point type in which the computation is being performed. By default,
underflow is gradual, and produces a denormal number or a zero.
All floating-points number of a given type are integer multiples of the smallest non-zero floating-point floatingpoint
point number of that type; however, the converse is not true. This means that, in the default mode,
(x-y) = 0 only if x = y.
The sign of zero transforms correctly through multiplication and division, and is preserved by addition
of zeros with like signs, but x - x yields +0 for every finite floating-point number x. The only oper-ations operations
ations that reveal the sign of a zero are x/(+-0) and copysign(x,+-0). In particular, comparisons (x >
y, x != y, etc) are not affected by the sign of zero.
The sign of infinity transforms correctly through multiplication and division, and infinities are unaf-fected unaffected
fected by addition or subtraction of any finite floating-point number. But Inf-Inf, Inf*0, and Inf/Inf
are, like 0/0 or sqrt(-3), invalid operations that produce NaN.
NaNs are the default results of invalid operations, and they propagate through subsequent arithmetic
operations. If x is a NaN, then x != x is TRUE, and every other comparison predicate (x > y, x = y, x
<= y, etc) evaluates to FALSE, regardless of the value of y. Additionally, predicates that entail an
ordered comparison (rather than mere equality or inequality) signal Invalid Operation when one of the
arguments is NaN.
IEEE-754 provides five kinds of floating-point exceptions, listed below:
Exception Default Result
__________________________________________
Invalid Operation NaN or FALSE
Overflow +-Infinity
Divide by Zero +-Infinity
Underflow Gradual Underflow
Inexact Rounded Value
NOTE: An exception is not an error unless it is handled incorrectly. What makes a class of exceptions
exceptional is that no single default response can be satisfactory in every instance. On the other
hand, because a default response will serve most instances of the exception satisfactorily, simply
aborting the computation cannot be justified.
For each kind of floating-point exception, IEEE-754 provides a flag that is raised each time its excep-
tion is signaled, and remains raised until the program resets it. Programs may test, save, and restore
the flags, or a subset thereof.
PRECISION AND EXPONENT RANGE OF SPECIFIC FLOATING-POINT TYPES
On both Intel and PPC macs, the type float corresponds to IEEE-754 single precision. A single-preci-sion single-precision
sion number is represented in 32 bits, and has a precision of 24 significant bits, roughly like 7 sig-nificant significant
nificant decimal digits. 8 bits are used to encode the exponent, which gives an exponent range from
-126 to 127, inclusive.
The header <float.h> defines several useful constants for the float type:
FLT_MANT_DIG - The number of binary digits in the significand of a float.
FLT_MIN_EXP - One more than the smallest exponent available in the float type.
FLT_MAX_EXP - One more than the largest exponent available in the float type.
FLT_DIG - the precision in decimal digits of a float. A decimal value with this many digits, stored as
a float, always yields the same value up to this many digits when converted back to decimal notation.
FLT_MIN_10_EXP - the smallest n such that 10**n is a non-zero normal number as a float.
FLT_MAX_10_EXP - the largest n such that 10**n is finite as a float.
FLT_MIN - the smallest positive normal float.
FLT_MAX - the largest finite float.
FLT_EPSILON - the difference between 1.0 and the smallest float bigger than 1.0.
On both Intel and PPC macs, the type double corresponds to IEEE-754 double precision. A double-preci-sion double-precision
sion number is represented in 64 bits, and has a precision of 53 significant bits, roughly like 16 sig-nificant significant
nificant decimal digits. 11 bits are used to encode the exponent, which gives an exponent range from
-1022 to 1023, inclusive.
The header <float.h> defines several useful constants for the double type:
DBL_MANT_DIG - The number of binary digits in the significand of a double.
DBL_MIN_EXP - One more than the smallest exponent available in the double type.
DBL_MAX_EXP - One more than the exponent available in the double type.
DBL_DIG - the precision in decimal digits of a double. A decimal value with this many digits, stored
as a double, always yields the same value up to this many digits when converted back to decimal nota-tion. notation.
tion.
DBL_MIN_10_EXP - the smallest n such that 10**n is a non-zero normal number as a double.
DBL_MAX_10_EXP - the largest n such that 10**n is finite as a double.
DBL_MIN - the smallest positive normal double.
DBL_MAX - the largest finite double.
DBL_EPSILON - the difference between 1.0 and the smallest double bigger than 1.0.
On Intel macs, the type long double corresponds to IEEE-754 double extended precision. A double
extended number is represented in 80 bits, and has a precision of 64 significant bits, roughly like 19
significant decimal digits. 15 bits are used to encode the exponent, which gives an exponent range
from -16383 to 16384, inclusive.
The header <float.h> defines several useful constants for the long double type:
LDBL_MANT_DIG - The number of binary digits in the significand of a long double.
LDBL_MIN_EXP - One more than the smallest exponent available in the long double type.
LDBL_MAX_EXP - One more than the exponent available in the long double type.
LDBL_DIG - the precision in decimal digits of a long double. A decimal value with this many digits,
stored as a long double, always yields the same value up to this many digits when converted back to
decimal notation.
LDBL_MIN_10_EXP - the smallest n such that 10**n is a non-zero normal number as a long double.
LDBL_MAX_10_EXP - the largest n such that 10**n is finite as a long double.
LDBL_MIN - the smallest positive normal long double.
LDBL_MAX - the largest finite long double.
LDBL_EPSILON - the difference between 1.0 and the smallest long double bigger than 1.0.
LONG DOUBLE ON POWERPC MACS
On PowerPC macs, by default the type long double is mapped to IEEE-754 double precision, described
above. If additional precision is required, a non-IEEE-754 128 bit long double format is also avail-able. available.
able. To use this format, compile with the -mlong-double-128 flag. If you wish to use the <math.h>
functions, you must include the linker flag -lmx as well as the usual -lm. The -mlong-double-128 flag
is only valid when the target architecture is ppc or ppc64.
Each 128-bit long double is made up of two IEEE doubles (head and tail). The value of the long double
is the sum of the values of the two parts (unless the head double has value -0.0, in which case the
value of the long double is -0.0). The value of the head is required to be the value of the long dou-ble double
ble rounded to the nearest double. If the head is an infinity, the value of the long double is the
value of the head, and the tail must be +-0.0. The tail of a NaN can be any double value. There are
many 128-bit bit patterns that are not valid as long doubles. These do not represet any value.
The 128-bit long double format provides 106 significant bits, which is roughly like 31 significant dec-imal decimal
imal digits. It has the same exponent range as double, from -1022 to 1023, inclusive. The usual con-stants constants
stants are provided from <float.h>, as described above.
In the 128-bit long double format, addition and subtraction have a relative error bound of one ulp, or
2**-106. Multiplication has a relative error within 2 ulps, and division a relative error within 3
ulps.
SEE ALSO
math(3), complex(3)
STANDARDS
Floating-point arithmetic conforms to the ISO/IEC 9899:1999(E) standard.
BSD March 20, 2007 BSD
|