An exploration of Inexact FP Arithmetic
In case anybody was wondering: I'm NOT proposing anything! I just wanted
to explore this notion.
I should also mention that formats with explicit unnormal representation
support, such as DFP and even good old HFP (of S/360 vintage), can also
mostly distinguish between exact and inexact results, if by convention
all normalized representations are interpreted as Inexact.
Indeed, this feature is exploited when 754-2008 DFP is used to emulate
variable-precision fixed-point decimal arithmetic. This does involve
giving up one high-order digit, although the latest IBM mainframe (z196)
supports Quantum Exceptions, which allow the full precision to be used,
as violation of fixed-point rules can be trapped.
The unnormalized-arithmetic approach does not help with inexact zeros
and infinities however, which was the main issue addressed by my IFP
approach.
Michel.
---Sent: 2012-03-22 19:41:44 UTC