loss of floating-point precision with transition from 5.21.1 to 5.22.1 ?!
Subject: loss of floating-point precision with transition from 5.21.1 to 5.22.1 ?!
From: Raymond Toy
Date: Sun, 10 Oct 2010 13:46:08 -0400
On 10/10/10 12:01 PM, Oliver Kullmann wrote:
> Hello,
>
> in several examples (all involving sums of logarithms) I noticed
> a substantial loss of precisions regarding floating-point arithmetic,
> comparing 5.22.1 with 5.21.1, using ECL 10.4.1.
> Is this expected and are there good reasons for this?
> Or shall I provide more details, so that the issue can be considered further?
Ah, this is probably caused by a change I made to allow log(<bigint>) to
work when bigint won't fit in a double. Because ecl doesn't signal an
error in these cases, the code always uses the new method. With other
lisps, an error is signaled, so the CL log function can be used in most
cases, and the error is caught so the new method is used.
Some simple examples, however, would certainly help me track this down.
Ray