Re: Test suite failure on Cygwin: atan2_1.f90

[Steve Kargl <sgk at troutmask dot apl dot washington dot edu>]

On Mon, Sep 14, 2009 at 08:50:50PM -0700, Jerry DeLisle wrote:
> 
> I have found that atan2_1.f90 fails on Cygwin on the following code:
> 
> do i = 1, 10
>   if(atan(1.0,  i/10.0)  -atan2(1.0,  i/10.)    /= 0.0)   call abort()
>   if(atan(1.0d0,i/10.0d0)-atan2(1.0d0,i/10.0d0) /= 0.0d0) call abort()
> end do
> 
> Notice that the test case is looking for an exact answer.  This test can 
> pass if we use epsilon() in the criteria.
> 
> Further, I noticed that if I use -ffloat-store (ye ole pr323) I do get 
> the exact results.  Should we change the test to accept some error or 
> add a target specific dg-option -ffloat-store for the Cygwin target?
> 

It's probably acceptable to accept some multiple
of epsilon in this case.  The Fortran standard is
fairly clear that an intrinsic function returns a
processor-dependent approximation to these functions.
If cygwins are concerned about the quality of the
cygwin/windows math library, let them run ucbtest
and fix it.

-- 
Steve