On Friday, September 6, 2002, at 02:01  PM, Brian Gough wrote:
> Try compiling that directory without optimisation, to see if it is
> compiler-related.  I think it is likely to be, since the code works on
> other platforms.

I have recompiled with less optimization, and with -O0 the matrix 
decompositions work. This, however, allowed me to find another mistake. 
Compiling the tests with no optimization, they stop further down, in 
the specfunc directory. Only this failure is reported:

FAIL: gsl_sf_conicalP_0_e(0.0, 100.0, &r)
   expected:   0.3009174858819926
   obtained:    0.300917485881992   7.027199282265925e-16  2.33526e-15
   fracdiff: 1.014601497407329e-15
   value not within tolerance of expected value
      0.30091748588199202  7.02719928226592486e-16

This is the only reported failure when the full make check is compiled 
with no optimization. And it is awfully close to machine epsilon.

I must retract a statement I made in an earlier posting about the 
beneficial effect of Apple's upgrade of gcc. It seems I must 
inadvertently have compiled the checks with less than -O2 when the 
internal compiler error first disappeared. I am sorry for that.

--
Erik Ø. Sørensen,
phd student, Economics
Norwegian School of Economics and Business Administration