Handling of floating point exceptions
Thomas W. Eggers, Judson S. Leonard, Mary H. Payne · ACM SIGNUM Newsletter · 1978
An IEEE subcommittee on the standardization of microprocessor floating point arithmetic has a proposal under discussion. Part of that proposal concerns overflow and underflow exceptions. The proposal calls for a “gradual” underflow implemented with demoralized numbers. For a sequence of addition/subtraction operations, the gradual underflow works very well: it almost has the effect of a machine with infinite range numbers. But if an addition/subtraction sequence is interrupted by a multiply or divide, things don't work nearly as well, and a fallback to symbolic information is likely. The proposal helps overflow hardly at all.