Re: secnds.f / secnds-1.f failures

[Brooks Moses <bmoses at stanford dot edu>]

At 10:53 AM 4/27/2007, Steve Ellcey wrote:
> I have taken a look at the sporadic secnds.f / secnds-1.f failures
> and I have a theory about what the problem is.
>
> In the libgfortran function secnds we have C code containing:
>
>   temp1 = 3600.0 * (GFC_REAL_4)values[4] +
>             60.0 * (GFC_REAL_4)values[5] +
>                    (GFC_REAL_4)values[6] +
>            0.001 * (GFC_REAL_4)values[7];
[...]
> We can change the library to do things in single precision or change
> the test to do things in double precision and I think that will fix
> the test failures.  Any preference on which way to go?

Aha!  Yes, that sounds like the sort of thing I'd expect to be causing this 
problem.  Thanks for tracking it down.

I think it probably makes more sense to leave the library code as 
double-precision, and change the testcase to match.

The difference in results is fairly minimal, but we do have experimental 
evidence that the different precisions _do_ occasionally get different 
results, and this function isn't ever likely to be used in a place where 
speed is of critical importance, so I think there's no good reason to 
change it to single precision.

- Brooks