[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: SIGFPE and -mieee

I'll read the documentation I was pointed to, but I'd like to mention
a few things as well:
On Thu, Jun 19, 2003 at 10:58:20PM -0400, Tyson Whitehead wrote:
> I wrote the original email after a week that went something like this.
> Day #1:
>   Hummm, let's browse the web.  Oh, Konqueor SIGFPEs.
>   A good portion of the rest of the day spent recompiling.

Well, I use konquerer (Debian stable) only for file management, but I
cannot remember it segfaulting; even when (rarly) using it for actual
browsing. I don't know if mozilla uses -mieee, though.

> Day #2:
>   Hummm, let's import an Excel spreadsheet.  Oh, KSpread SIGFPEs.
>   A good portion of the rest of the day spent recompiling.

I used gnumeric, also importing Excel spreadsheets. Again, I don't
know if -mieee is used there, but it works fine and does not segault.

> Day #4:
>   Hummm, let's listen to some music.  Oh, mpg321 SIGFPEs.
>   Hummm, let's view that video.  Oh, Xine SIGFPEs.

Ok, ogg123 sometimes segfaults, this could be related. mpg123 however,
I've never seen to segfault.

> etc...

Well, I never really did a research, but presuming that only few
applications do use -mieee on woody, I rarly experience segfaults
regarding this. I have some open bugs in respect to this, but they are
mostly games and simulations. Both cases, which I IMHO rather get
"fixed" (at least in the long run) than masked. If you tell me, that
fixing is impossible (I am not a experienced enough to really evaluate
that) than this would be sad, and in this case -mieee should be made
the default.

> By the term default, I was referring to the 95% (probably more like 99%) of 
> the (11858) Debian packages that you won't be using to perform your 30 day 
> numerical simulations.  The user/desktop applications.

Well, ok. For this applications, especially if fixes are highly
unlikly in the best case, this might make sense. 

> If you guys have figured out that ATLAS/BLAS, LAPACK, etc, runs fine without 
> -mieee, that's fine by me.  In my books those apps/libs are all highly 
> specialized.  I expect the package maintainers to use non-default compilation 
> flags (i.e. to specify things like -O6 *grin*, -fstrict-aliasing, and not 
> things like -mieee).

Well, I don't exactly know how "default" would be realizied. Would it
be something a package maintainer could easily overrule? In this case
I'd suggest something along:

*Broadcast in the appropriate mailing lists (maybe with a note in DWN)
 the change; noting the speed penalties for performance related
 applications and asking maintainers to check if they do not need
 -mieee (e.g. because it is fixed already).

Do you know if there is an easy way of accessing the rules files of
all packages? If that is possible, I'd compile a list of "known to
work" apps, so that those packages would override the -mieee. I'd post
the list here, so anyone could add/object to entries.

> PS:  As was mentioned earlier, isn't the point of not using -mieee kind of 
> irrelevant anyway, as modern Alapha architectures (i.e. 21264 [ev6] and 
> later) aren't slowed down by imprecise exception handling anyway (trapb 
> instructions are simply dropped)?

If you consider my machine irrelevant, because it is not the latest
alpha generation but only a (very nice IMHO) 20164a, then yes. One of
the reasons I chose Debian was btw. that "irrelevant" machines are
supported and not "work by chance". 



Helge Kreutzmann, Dipl.-Phys.               Helge.Kreutzmann@itp.uni-hannover.de
  gpg signed mail preferred    gpg-key: finger kreutzm@rigel.itp.uni-hannover.de
    64bit GNU powered                  http://www.itp.uni-hannover.de/~kreutzm
       Help keep free software "libre": http://www.freepatents.org/

Attachment: pgprvG1xdojXR.pgp
Description: PGP signature

Reply to: