Re: erfc() & Alphas

[Stephen L Moshier]

> I noticed that the erfc() implementation on the Alpha seems to return
> denormalized results (i.e. erfc(27) -> 5.237046e-319) when -mieee is
> not specified to GCC.  This is somewhat curious and problematic.

Newlib (sources.redhat.com/newlib) has a policy that favors support of
variant floating point types.  It seems, though, that the only
version of erf there is the same IEEE one from fdlibm that glibc uses.
You could try the GAMS index (gams.nist.gov) for various other sources
of math routines.


> Anyone have any idea of why erfc() is twiddling with the bits directly
> instead of using the FPU (or some other explanation for what is going
> on)?  Could something be done to fix this?

That is supposed to be a feature, not a bug.  Fdlibm uses integer operations
on floating point values extensively, because they are fast.