Date: Wed, 11 Mar 1998 17:31:29 +0100 From: Martin Cracauer <cracauer@cons.org> To: "Neil A. Carson" <neil@causality.com> Cc: bugs@FreeBSD.ORG Subject: Re: Floating point errors Message-ID: <19980311173129.43427@cons.org> In-Reply-To: <34EE57FE.5B1234BB@causality.com>; from Neil A. Carson on Sat, Feb 21, 1998 at 04:28:46AM %2B0000 References: <199802210421.PAA18491@godzilla.zeta.org.au> <34EE57FE.5B1234BB@causality.com>
next in thread | previous in thread | raw e-mail | index | archive | help
In <34EE57FE.5B1234BB@causality.com>, Neil A. Carson wrote: > Bruce Evans wrote: > > > Wrong. I feel that it is more useful to die rather than possibly > > run incorrectly. Programmers who actually understand FP exceptions > > can easily change the default exception mask (I changed it 3 years > > ago on my systems). The technical correctness of this is debatable. > > It prevents the math libraries from being ANSI conformant, but the math > > libraries have more serious ANSI conformance bugs. I don't plan to > > change the exception handling until the other bugs are fixed. > > Are the conformance bugs quite major? We had to write a collection of FP > emulator code that was completely ANSI conformant (for one reason or another, > to pass the Java compliance tests here at Oracle) for the ARM, as it doesn't > have an FPU of any sort. Of course, programmers who understand FP exceptions > can easily change stuff, but the mathematicians/physicians (who are in this > case quite computer-illiterate and only like to do the bear minimumof code to > get something to 'work') in my case are unaware of the set mask call. Of > course, as a reverse, one could argue that as a result their stuff shouldn't > work anyway! Not that of course this is of any direct relevance to my small > evaluation here anyway, as I also have a number of NetBSD machines around; > maybe I'm just waffling so I'll shut up now. If you want to change the default behaviour of the running system for all processes, you can set #define __INITIAL_NPXCW__ 0x127f in npx.h (change both in kernel sources and /usr/include/machine) and build a new kernel. > If I can be of assistance with any stuff, let me know. Well, reproduceable test cases of faulty behaviour with gcc/i386 would be useful. This applies to NetBSD as well as FreeBSD. I don't think Linux is different, either. Martin -- %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Martin Cracauer <cracauer@cons.org> http://www.cons.org/cracauer/ BSD User Group Hamburg/Germany http://www.bsdhh.org/ To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-bugs" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?19980311173129.43427>