In the Language Reference documentation for the (kinda new) MINEXP builtin, it lists the returned values for the various floating-pt formats of various precisions.

 

In particular, for the HFP floating-pt formats it says:

 

   Example (z/OS Hexadecimal Values)

       minexp(x) = -64 for x float bin(p), p <= 21

       minexp(x) = -64 for x float bin(p), 21 <p <= 53

       minexp(x) = -50 for x float bin(p), 53 < p

 

The values of -64 for precisions less than or equal to 53 make sense; but that -50 for precisions greater than 53 is curious.

In the IBM HFP floating-pt format; shouldn't that also be -64?

I tested on zOS, and true to the documentation it returns -50.  But I really don't understand where that -50 comes from?   The exponent range of a 128-bit HFP value is the same as that of a 64-bit HFP value (and the same as that of a 32-bit HFP value.)

 

The -50 comes from the following description of the low-order characteristic in the Principles of Operation, the low-order characteristic being a true representation of the exponent of the low-order half of the mantissa, 14 the number of hexadecimal digits in the upper (and incidentally the lower) half of the number. If the low-order characteristic underflows, then effectively the low-order half of the number is no longer useful, you're left with a double float, not the extended float you think you have...

 

When a result is generated in the extended format
and placed in a register pair, the sign of the low-order
part is made the same as that of the high-order part,
and, unless the result is a true zero, the low-order
characteristic is made 14 less than the high-order
characteristic. When the subtraction of 14 would
cause the low-order characteristic to become less
than zero, the characteristic is made 128 greater than
its correct value. (Thus, the subtraction is performed
modulo 128.) HFP exponent underflow is indicated
only when the high-order characteristic underflows.

 

While that is true - that the low-order characteristic does underflow; it is not true that the low-order portion is unusable and you are only left with the 64-bit value.   

 

From the principles of operation in the section on "Hexadecimal-Floating-Point":

 

    "The exponent range is the same for the short, long, and extended formats."

 

Furthermore, from the section on "Equivalent Floating-Point Number Represenations":

 

   "The exponent of an HFP number is represented in the format as an unsigned seven-bit

     binary interger called the characteristic.  The characteristic is obtained by adding 64

     to the exponent value (excess-64 notation).  The range of the characteristic is 0 to 127,

     which corresponds to an exponent range of -64 to +63."

 

And - again in chapter 18 - we have "HFP Data Formats"

   

 

Thus - while the low-order characteristic does underflow during normalization; it is ignored for any operation; which

means that the range of an extended-precision floating-pt value is the same as the range of a long-precision or

short-precision.

 

The low-order characteristic is not considered during any of the extended floating-pt operations; it is only the number of digits in the fraction

that changes (from 14 to 28.)

 

 

So - the proper value for MINEXP of an extended-precision HFP floating point value, I think, is -64, not -50.

 

 

Interestingly, the values for MAXEXP appear to be correct.