home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!mcsun!sun4nl!moene!moene.indiv.nluug.nl
- From: toon@moene.indiv.nluug.nl (Toon Moene)
- Newsgroups: comp.arch
- Subject: Re: Unnormalized FP and a PDP-11 bug (Re: Algorithm for bit finding)
- Message-ID: <346@moene.indiv.nluug.nl>
- Date: 14 Aug 92 22:02:42 GMT
- References: <1992Aug12.075553.1@zodiac.rutgers.edu>
- Sender: toon@moene.indiv.nluug.nl
- Organization: Moene Computational Physics, Amsterdam, The Netherlands
- Lines: 33
-
- In article <1992Aug12.075553.1@zodiac.rutgers.edu>
- leichter@zodiac.rutgers.edu writes:
- [... Interesting story about FP faults generated by copying bytes
- with FP move instructions deleted ... ]
-
- > (My favorite story of a simliar problem is on the CDC 6000 series. They
- > use 1's complement arithmetic, and when you load a B register, you can't
- > really load a constant - you load contant plus some other B register.
- > No problem: B0 is always 0, so you add it. Except that in one's
- > complement there is also -0 (all 1 bits), and the arithmetic unit tries
- > very hard never to generate -0 as an output. In fact, 0+(-0) is 0. So
- > you can't load -0 that way! The trick is to realize that the REAL
- > identity for addition on this machine is -0, and get that into a
- > register (using bit-wise operations) for use in place of 0 when loading
- > data that isn't really numbers.)
-
- Hm, I wonder why you would want -0 in a 18-bit index register - masking
- instructions only operated on the 60-bit registers ...
-
- I've another horror story on one's complement 'arithmetic'. On the old
- UNIVAC 1100 series computers (36-bit one's complement machines), there
- once was a FORTRAN compiler that stored the 'no-dead-code-in-this-basic-
- block' flag in a double (72-bit) integer. Now for optimization, they first
- tested this integer against zero, before examining every bit to determine
- the basic blocks to remove :-) It was certainly interesting hearing this
- customer talk about his large FORTRAN subroutine diminished to a few
- instructions by this compiler ...
-
- --
- Toon Moene (toon@moene.indiv.nluug.nl)
- Kantershof 269, 1104 GN Amsterdam, The Netherlands
- Tel.: + 31 20 6982029; Fax: + 31 20 6003411
- No Disclaimers; a NeXT at home protects against this occupational hazard.
-