Re: Puzzle computing arrays

[Jerry DeLisle <jvdelisle at verizon dot net>]

Brian Dessent wrote:
> Angelo Graziosi wrote:
>
> > Instead the SSE-math does not work on Cygwin:
> >
> > $ gfortran -mfpmath=sse test_case.f90 -o test_case
> > f951: warning: SSE instruction set disabled, using 387 arithmetics
>
> I seriously doubt that this has anything to do with Cygwin.  You didn't
> specify any -march, and the default if none was specified is the
> original 386 instruction set which does not contain any vector
> instructions.  You can also explicitly enable those instructions with
> -msse or -msse2, but specifying the proper -march is a cleaner way to do
> it than indicating piecewise what is available, IMO.  See also
> -march=native.
>
> Brian
I can reproduce nearly the exact behavior with the test case on x86-64-Linux 
using -m32.  The problem also disappears with -m32 -ffloat-store.  So I think 
this is the age old 80 bit floating point issue. See pr323.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=323

Cheers,

Jerry

]$ gfc -m32 testcase.f90
[jerry@lenova test]$ ./a.out
           0   0.0000000000000000        0.0000000000000000 
0.0000000000000000
           1 -7.10542735760100186E-015 -7.10542735760100186E-015 
0.0000000000000000
           2   0.0000000000000000        0.0000000000000000 
0.0000000000000000
[jerry@lenova test]$ gfc -m32 -ffloat-store testcase.f90
[jerry@lenova test]$ ./a.out
           0   0.0000000000000000        0.0000000000000000 
0.0000000000000000
           1   0.0000000000000000        0.0000000000000000 
0.0000000000000000
           2   0.0000000000000000        0.0000000000000000 
0.0000000000000000