Re: [Help-gsl] Double vs Long Double

[dave]

On Thu, Oct 04, 2007 at 07:53:39PM +0100, Brian Gough wrote:
> At Tue, 2 Oct 2007 12:09:00 +0000 (GMT),
> dave wrote:
> > I tried computing J0(5) using the Bessel function in the gnu math
> > library and I got the exact result listed in the book. So I'm wondering
> > if the difference in the final 3 digits is the result of using double 
> > (64-bit xmm registers vs long double (80-bit temp real in  8087 registers).
> 
> You can use the GSL_IEEE_MODE=double-precision setting (described in
> the IEEE Chapter of the manual) to turn off the 80-bit mode and see if
> it makes a difference.
> 
> -- 
> Brian Gough
> (GSL Maintainer)

I tried setting GSL_IEEE_MODE='extended-precision' and the OpenBSD version
dumped core after printing an error message.

gsl: fp-openbsd.c:41: ERROR: OpenBSD only supports default precision rounding
Default GSL error handler invoked.
Abort trap (core dumped)

This also happened with GSL_IEEE_MODE='double-precision'. It appears that this
variable affects only rounding mode. It would be extremely handy if this 
technique
could be used to switch between 80 bit temporary real  (8087) math and 64 bit 
xmm math
operations.