Intel 80287, 80286 manual Rounding Control

OVERVIEW OF NUMERIC PROCESSING

The NPX usually carries the digits of the significand in normalized form. This means that, except for the value zero, the significand is an integer and a fraction as follows:

l~fff.. .ff

where ~ indicates an assumed binary point. The number of fraction bits varies according to the real format: 23 for short, 52 for long, and 63 for temporary real. By normalizing real numbers so that their integer bit is always a 1, the 80287 eliminates leading zeros in small values (Ix! < 1). This technique maximizes the number of significant digits that can be accommodated in a significand of a given width. Note that, in the short and long real formats, the integer bit is implicit and is not actually stored; the integer bit is physically present in the temporary real format only.

If one were to examine only the signficand with its assumed binary point, all normalized real numbers would have values between 1 and 2. The exponent field locates the actual binary point in the significant digits. Just as in decimal scientific notation, a positive exponent has the effect of moving the binary point to the right, and a negative exponent effectively moves the binary point to the left, inserting leading zeros as necessary. An unbiased exponent of zero indicates that the position of the assumed binary point is also the position of the actual binary point. The exponent field, then, determines a real number's magnitude.

In order to simplify comparing real numbers (e.g., for sorting), the 80287 stores exponents in a biased form. This means that a constant is added to the true exponent described above. The value of this bias is different for each real format (see figure 1-9). It has been chosen so as to force the biased exponent to be a positive value. This allows two real numbers (of the same format and sign) to be compared as if they are unsigned binary integers. That is, when comparing them bitwise from left to right (begin- ning with the leftmost exponent bit), the first bit position that differs orders the numbers; there is no need to proceed further with the comparison. A number's true exponent can be determined simply by subtracting the bias value of its format.

The short and long real formats exist in memory only. If a number in one of these formats is loaded into an 80287 register, it is automatically converted to temporary real, the format used for all internal operations. Likewise, data in registers can be converted to short or long real for storage in memory. The temporary real format may be used in memory also, typically to store intermediate results that cannot be held in registers.

Most applications should use the long real form to store real number data and results; it provides sufficient range and precision to return correct results with a minimum of programmer attention. The short real format is appropriate for applications that are constrained by memory, but it should be recognized that this format provides a smaller margin of safety. It is also useful for debugging algorithms, because roundoff problems will manifest themselves more quickly in this format. The temporary real format should normally be reserved for holding intermediate results, loop accumulations, and constants. Its extra length is designed to shield final results from the effects of rounding and overflow/underflow in intermediate calculations. However, the range and precision of the long real form are adequate for most microcomputer applications.

Rounding Control

Internally, the 80287 employ~ three extra bits (guard, round, and sticky bits) that enable it to represent the infinitely precise true result of a computation; these bits are not accessible to programmers. Whenever the destination can represent the infinitely precise true result, the 80287 delivers it. Rounding occurs

1-18