Tim,

> > There's another possibility, though. I found on Windows 
> > that under some circumstances (which I could never figure
> > out completely) the FPU flags get changed to signal exceptions
> > instead of silently over/underflowing.
>
> Good grief, this surely can't still be a problem in Windows;

Well, maybe not, but would I want to put my money on this bet? ;-)

> we (as in PPS, in particular Jan Bottorff, Daniel Lanovaz & I)
> found this afflicting VW back in 93 or there abouts. Certain
> library calls (all vis dlls IIRC) would change the FPU flags
> and _not restore them_ on return.
> Wham, bang, floating pint arithmetic goes weird on you. I suppose it
> might be worth emailing Eliot to ask if an answer was ever found. 

I actually solved this problem by merely establishing an FPU exception
filter and reset the flags if they got changed. Note that question was about
Linux, so I was mostly wondering if there might be a similar problem here.

Cheers,
  - Andreas