Xref: news2.mv.net comp.os.msdos.djgpp:1756
From: terra AT diku DOT dk (Morten Welinder)
Newsgroups: comp.os.msdos.djgpp
Subject: Re: paranoia
Date: 6 Mar 1996 13:01:09 GMT
Organization: Department of Computer Science, U of Copenhagen
Lines: 26
Sender: terra AT tyr DOT diku DOT dk
Message-ID: <4hk2al$q82@odin.diku.dk>
References: <9603051656 DOT AA06324 AT tntnhb3 DOT tn DOT tudelft DOT nl>
NNTP-Posting-Host: odin.diku.dk
To: djgpp AT delorie DOT com
DJ-Gateway: from newsgroup comp.os.msdos.djgpp

erik AT tntnhb3 DOT tn DOT tudelft DOT nl (Erik Luijten) writes:

>The recent messages about DJGPP's floating point library have somewhat
>disturbed me. Therefore I have compiled paranoia (from netlib), to see whether
>this program reveals some problems. Well, I have compiled it without
>optimization (the program completely choked when I tried -O2, why?) and
>paranoia said the following:

>No chopping nor rounding on * / + - sqrt: sticky bit incorrectly used or not
>used at all. 


This is a bug in the paranoia program.  The job that paranoia seeks
to perform cannot be done within well-defined C.  The assumptions
that paranoia puts on the compiler are not present in the C standard.

Paranoia assumes that there is a one-to-one correspondence between
operations ("+", "-", ...) at source level and operations at the
target language level ("FADD", ...).  This is not required by the
C standard and it is not so for gcc, probably not even with
"optimizations off".

Things are different in fortran where parentheses have more than
syntactic meaning.

Morten