Developer Guide and Reference

Contents

Special Values

This is a list and a brief description of the special values that the Intel® C++ Compiler supports.

Signed Zero

The sign of zero is the same as the sign of a nonzero number. Comparisons consider +0 to be equal to -0. A signed zero is useful in certain numerical analysis algorithms, but in most applications the sign of zero is invisible.

Denormalized Numbers

Denormalized numbers (denormals) fill the gap between the smallest positive and the smallest negative normalized number, otherwise only (+/-) 0 occurs in the interval. Denormalized numbers extend the range of computable results by allowing for gradual underflow.
Systems based on the IA-32 architecture support a Denormal Operand status flag. When this is set, at least one of the input operands to a floating-point operation is a denormal. The Underflow status flag is set when a number loses precision and becomes a denormal.

Signed Infinity

Infinities are the result of arithmetic in the limiting case of operands with arbitrarily large magnitude. They provide a way to continue when an overflow occurs. The sign of an infinity is simply the sign you obtain for a finite number in the same operation as the finite number approaches an infinite value.
By retrieving the status flags, you can differentiate between an infinity that results from an overflow and one that results from division by zero. The Intel® C++ Compiler treats infinity as signed by default. The output value of infinity is +Infinity or -Infinity.

Not a Number

Not a Number (NaN) results from an invalid operation. For example,
0/0
and
SQRT(-1)
result in NaN. In general, an operation involving a NaN produces another NaN. Because the fraction of a NaN is unspecified, there are many possible NaNs.
The Intel® C++ Compiler treats all NaNs identically, but provides two different types:
  • Signaling NaN, which has an initial mantissa bit of 0 (zero). This usually raises an invalid exception when used in an operation.
  • Quiet NaN, which has an initial mantissa bit of 1.
The floating-point hardware changes a signaling NaN into a quiet NaN during many arithmetic operations, including the assignment operation. An invalid exception may be raised but the resulting floating-point value is a quiet NaN.
The output value of NaN is NaN.