Re: Floating point problems

[Greg Lindahl <lindahl cs virginia edu>]

> Actually this is a good idea.  And the compiler already defines
> -D_IEEE_FP.  And the define can even be replaced by
> 
>    extern double sqrt (double) __asm__("_sqrt_no_denorm");

Yes, that's all clean. We just have to incite the library writers to
produce the *_no_denorm routines.

> The question, then, is what to do with the other problematic
> return values, Inf and NaN?  Should Inf be replaced by the 
> largest finite value?

No. The object here is to allow code to run fast while changing its
semantics as little as possible. Messing with Inf and NaN would not
really help anyone.

There are users who would like their code to crash whenever Inf or NaN
is produced; we should document how to do that.

-- g