denormals/subnormals are heading for extinction

[Stephen L Moshier <moshier at mediaone dot net>]

> Theoretically speaking, I think denormals are just fine.
> But there is a major disconnect between theory and practice.

Here is a concrete example of a useful application of denormal
numbers.  It is the telephone codec that encodes audio signals for
digital switching and transmission.  The codes go by the name of A law
or Mu law.  The representation has a sign bit, a three-bit exponent,
and a four-bit significand.  There is an understood high order bit in
the significand.  When the exponent goes to zero, the denormal segment
descends linearly to zero.  In other words, the code is an eight bit
floating point number.  It is not hard to demonstrate (for example by
amplifying some low level coded audio) that the denormal
representation gives better audio fidelity than a flush to zero rule.

I do not know any anecdotes about useful applications for IEEE
denormals.  Perhaps the IEEE enthusiasts here can provide one?

Whatever your estimate may be of their statistical frequency,
denormals do occur often in standards documents.  By policy, GCC has
to support them.