[Bug-gsl] gsl_cdf_fdist_P() nan return value

[Daniel Neilson]

Hi there,
 I just spent the entire day tracking down the source of a nan in some code
of mine, and tracked it back to gsl_cdf_fdist_P(x, nu1, nu2). When called
with:

 double x = 0x1.bd5290e3772c6p+1 (printf("%a", x) format to get the full
binary representation as hex)
 int nu1 = 1
 int nu2 = 4040712

this function returns nan. Furthermore, for Y \in {100,10000} calling:

gsl_cdf_fdist_P( (double)((unsigned int)x * Y + 0.5) / Y, nu1, nu2)

also returns nan.

 However, for Y \in {1000, 100000} it returns a proper non-nan value.

 This is all with the version of GSL in the Fedora 7 packages gsl-1.8-3.fc7&
gsl-devel-1.8-3.fc7.

-Daniel