Редагувати

Поділитися через


IEEE Floating-Point Representation

Microsoft C++ (MSVC) is consistent with the IEEE numeric standards. The IEEE-754 standard describes floating-point formats, a way to represent real numbers in hardware. There are at least five internal formats for floating-point numbers that are representable in hardware targeted by the MSVC compiler. The compiler only uses two of them. The single-precision (4-byte) and double-precision (8-byte) formats are used in MSVC. Single-precision is declared using the keyword float. Double-precision is declared using the keyword double. The IEEE standard also specifies half-precision (2-byte) and quadruple-precision (16-byte) formats, and a double-extended-precision (10-byte) format, which some C and C++ compilers implement as the long double data type. In the MSVC compiler, the long double data type is treated as a distinct type, but the storage type maps to double. There is, however, intrinsic and assembly language support for computations using the other formats, including the double-extended-precision format, where supported by hardware.

The values are stored as follows:

Value Stored as
single-precision sign bit, 8-bit exponent, 23-bit significand
double-precision sign bit, 11-bit exponent, 52-bit significand

In single-precision and double-precision formats, there's an assumed leading 1 in the fractional part. The fractional part is called the significand (sometimes known as the mantissa). This leading 1 isn't stored in memory, so the significands are actually 24 or 53 bits, even though one less bit gets stored. The double-extended-precision format actually stores this bit.

The exponents are biased by half of their possible value. It means you subtract this bias from the stored exponent to get the actual exponent. If the stored exponent is less than the bias, it's actually a negative exponent.

The exponents are biased as follows:

Exponent Biased by
8-bit (single-precision) 127
11-bit (double-precision) 1023

These exponents aren't powers of ten; they're powers of two. That is, 8-bit stored exponents can range from -127 to 127, stored as 0 to 254. The value 2127 is roughly equivalent to 1038, which is the actual limit of single-precision.

The significand is stored as a binary fraction of the form 1.XXX... . This fraction has a value greater than or equal to 1 and less than 2. Real numbers are always stored in normalized form. That is, the significand is left-shifted such that the high-order bit of the significand is always 1. Because this bit is always 1, it's assumed (not stored) in the single-precision and double-precision formats. The binary (not decimal) point is assumed to be just to the right of the leading 1.

The format for floating-point representation is as follows:

Format byte 1 byte 2 byte 3 byte 4 ... byte n
single-precision SXXXXXXX XMMMMMMM MMMMMMMM MMMMMMMM
double-precision SXXXXXXX XXXXMMMM MMMMMMMM MMMMMMMM ... MMMMMMMM

S represents the sign bit, the X's are the biased exponent bits, and the M's are the significand bits. The leftmost bit is assumed in single-precision and double-precision formats.

To shift the binary point properly, you first unbias the exponent and then move the binary point to the right or left the appropriate number of bits.

Special values

The floating-point formats include some values that are treated specially.

Zero

Zero can't be normalized, which makes it unrepresentable in the normalized form of a single-precision or double-precision value. A special bit pattern of all zeroes represents 0. It's also possible to represent -0 as zero with the sign bit set, but -0 and 0 always compare as equal.

Infinities

The +∞ and −∞ values are represented by an exponent of all ones, and a significand that's all zeroes. Positive and negative are represented by using the sign bit.

Subnormals

It's possible to represent numbers of smaller magnitude than the smallest number in normalized form. They're called subnormal or denormal numbers. If the exponent is all zeroes and the significand is non-zero, then implicit leading bit of the significand is considered to be zero, not one. The precision of subnormal numbers goes down as the number of leading zeroes in the significand goes up.

NaN - Not a Number

It's possible to represent values that aren't real numbers, such as 0 / 0, in the IEEE floating-point format. A value of this kind is called a NaN. A NaN is represented by an exponent of all ones and a non-zero significand. There are two kinds of NaNs, quiet NaNs, or QNaNs, and signaling NaNs, or SNaNs. Quiet NaNs have a leading one in the significand, and get propagated through an expression. They represent an indeterminate value, such as the result of dividing by infinity, or multiplying an infinity by zero. Signaling NaNs have a leading zero in the significand. They're used for operations that aren't valid, to signal a floating-point hardware exception.

Examples

The following are some examples in single-precision format:

  • For the value 2, the sign bit is zero. The stored exponent is 128, or 1000 0000 in binary, which is 127 plus 1. The stored binary significand is (1.) 000 0000 0000 0000 0000 0000, which has an implied leading 1 and binary point, so the actual significand is one.

    Value Formula Binary representation Hexadecimal
    2 1 * 21 0100 0000 0000 0000 0000 0000 0000 0000 0x40000000
  • The value -2. Same as +2 except that the sign bit is set. The same thing is true for the negative of all IEEE format floating-point numbers.

    Value Formula Binary representation Hexadecimal
    -2 -1 * 21 1100 0000 0000 0000 0000 0000 0000 0000 0xC0000000
  • The value 4. Same significand, exponent increases by one (biased value is 129, or 100 0000 1 in binary.

    Value Formula Binary representation Hexadecimal
    4 1 * 22 0100 0000 1000 0000 0000 0000 0000 0000 0x40800000
  • The value 6. Same exponent, significand is larger by half. It's (1.) 100 0000 ... 0000 0000, which, since it's a binary fraction, is 1 1/2 because the values of the fractional digits are 1/2, 1/4, 1/8, and so forth.

    Value Formula Binary representation Hexadecimal
    6 1.5 * 22 0100 0000 1100 0000 0000 0000 0000 0000 0x40C00000
  • The value 1. Same significand as other powers of two, the biased exponent is one less than two at 127, or 011 1111 1 in binary.

    Value Formula Binary representation Hexadecimal
    1 1 * 20 0011 1111 1000 0000 0000 0000 0000 0000 0x3F800000
  • The value 0.75. The biased exponent is 126, 011 1111 0 in binary, and the significand is (1.) 100 0000 ... 0000 0000, which is 1 1/2.

    Value Formula Binary representation Hexadecimal
    0.75 1.5 * 2-1 0011 1111 0100 0000 0000 0000 0000 0000 0x3F400000
  • The value 2.5. Exactly the same as two except that the bit that represents 1/4 is set in the significand.

    Value Formula Binary representation Hexadecimal
    2.5 1.25 * 21 0100 0000 0010 0000 0000 0000 0000 0000 0x40200000
  • 1/10 is a repeating fraction in binary. The significand is a little less than 1.6, and the biased exponent says that 1.6 is to be divided by 16. (It's 011 1101 1 in binary, which is 123 in decimal.) The true exponent is 123 - 127 = -4, which means that the factor by which to multiply is 2-4 = 1/16. The stored significand is rounded up in the last bit in an attempt to represent the unrepresentable number as accurately as possible. (The reason that 1/10 and 1/100 aren't exactly representable in binary is similar to the reason that 1/3 isn't exactly representable in decimal.)

    Value Formula Binary representation Hexadecimal
    0.1 1.6 * 2-4 0011 1101 1100 1100 1100 1100 1100 1101 0x3DCCCCCD
  • Zero is a special case. It uses the formula for the minimum possible representable positive value, which is all zeroes.

    Value Formula Binary representation Hexadecimal
    0 1 * 2-128 0000 0000 0000 0000 0000 0000 0000 0000 0x00000000

See also

Why Floating-Point Numbers May Lose Precision