Path: utzoo!utgpu!news-server.csri.toronto.edu!clyde.concordia.ca!uunet!mcsun!hp4nl!charon!dik
From: dik@cwi.nl (Dik T. Winter)
Newsgroups: comp.lang.fortran
Subject: Re: Explain this output to me
Message-ID: <1549@charon.cwi.nl>
Date: 26 May 90 00:32:48 GMT
References: <30499@ut-emx.UUCP>
Sender: news@cwi.nl
Organization: CWI, Amsterdam
Lines: 26

In article <30499@ut-emx.UUCP> vaughan@ut-emx.UUCP (Curt Vaughan) writes:
(About 0.1 etc.)
 > It may also be dependent upon the machine architecture.  Although
 > most machines now use two's complement arithmetic, the Cyber 170,
 > which uses the most significant bit as a sign bit (60-bit word)
 > produces the correct result with the original code.  The Cray X-MP,
 > VAX 780, and Convex, all of which use two's complement arithmetic,
 > yield varying, but incorrect values.
 > 
This is not correct.  There are only a few machines that use two's
complement for floating-point.  I know only: CDC-205, ETA-10,
Modcomp (does it still exist?) and Gould (really 16's complement).
All the others use sign/magnitude, except the Cyber 170 which uses
one's complement.  There is not really a difference between one's
complement and sign/magnitude (except representation of course).
The rounding mode remains the same.  The reason you get different
results on those machines is:
1.  Probably the code on the Cyber 170 was compiled using rounding
    instructions.
2.  The Cray (with the same mantissa size as the Cyber) does truncate.
3.  The Vax has a different mantissa size (53 vs. 48 bits).
4.  Convex has two floating point formats, both approximately the
    same format as the Vax, but rounding is slightly different.
--
dik t. winter, cwi, amsterdam, nederland
dik@cwi.nl