[PATCH, rs6000, testsuite] Skip gfortran.dg/round_4.f90 for PowerPC Linux

[PATCH, rs6000, testsuite] Skip gfortran.dg/round_4.f90 for PowerPC Linux

[Bill Schmidt]

Hi,

The test in gfortran.dg/round_4.f90, introduced in GCC 4.9, checks for
correct behavior of different rounding modes.  However, for
quad-precision floating-point, it requires that the number
0.10000000000000000000000000000000000481 be exactly represented.  Since
the PowerPC long double implementation (double-double) only guarantees
31 bits of precision, the test fails for the real(qp) portions of the
test.  Thus this patch marks the test invalid for PowerPC Linux for now.
(We may want to also do this for other subtargets; let me know if so.)

At such time as IEEE 128-bit floating-point is supported by the PowerPC
port, we should revisit this.

Is this ok for trunk and 4.9?

Thanks,
Bill


2014-06-29  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

	* gfortran.dg/round_4.f90: Skip for powerpc*-*-linux* since the
	test requires greater precision than the current PowerPC long
	double implementation supports.


Index: gcc/testsuite/gfortran.dg/round_4.f90
===================================================================
--- gcc/testsuite/gfortran.dg/round_4.f90	(revision 212074)
+++ gcc/testsuite/gfortran.dg/round_4.f90	(working copy)
@@ -1,6 +1,7 @@
 ! { dg-do run }
 ! { dg-add-options ieee }
 ! { dg-skip-if "PR libfortran/58015" { hppa*-*-hpux* } }
+! { dg-skip-if "IBM long double 31 bits of precision, test requires 38" { powerpc*-*-linux* } }
 !
 ! PR fortran/35862
 !

Re: [PATCH, rs6000, testsuite] Skip gfortran.dg/round_4.f90 for PowerPC Linux

[David Edelsohn]

On Sun, Jun 29, 2014 at 9:48 PM, Bill Schmidt
<wschmidt@linux.vnet.ibm.com> wrote:
> Hi,
>
> The test in gfortran.dg/round_4.f90, introduced in GCC 4.9, checks for
> correct behavior of different rounding modes.  However, for
> quad-precision floating-point, it requires that the number
> 0.10000000000000000000000000000000000481 be exactly represented.  Since
> the PowerPC long double implementation (double-double) only guarantees
> 31 bits of precision, the test fails for the real(qp) portions of the
> test.  Thus this patch marks the test invalid for PowerPC Linux for now.
> (We may want to also do this for other subtargets; let me know if so.)
>
> At such time as IEEE 128-bit floating-point is supported by the PowerPC
> port, we should revisit this.
>
> Is this ok for trunk and 4.9?
>
> Thanks,
> Bill
>
>
> 2014-06-29  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
>
>         * gfortran.dg/round_4.f90: Skip for powerpc*-*-linux* since the
>         test requires greater precision than the current PowerPC long
>         double implementation supports.

Okay.

I wonder if this is passing on AIX because AIX defaults to double?

Thanks, David

Re: [PATCH, rs6000, testsuite] Skip gfortran.dg/round_4.f90 for PowerPC Linux

[Mike Stump]

On Jun 29, 2014, at 6:48 PM, Bill Schmidt <wschmidt@linux.vnet.ibm.com> wrote:
> +! { dg-skip-if "IBM long double 31 bits of precision, test requires 38" { powerpc*-*-linux* } }

Thanks for the comment.  Can someone confirm that this fails on darwin and should be skipped there as well?

Re: [PATCH, rs6000, testsuite] Skip gfortran.dg/round_4.f90 for PowerPC Linux

[Tobias Burnus]

Hi Bill,

Bill Schmidt wrote:
> The test in gfortran.dg/round_4.f90, introduced in GCC 4.9, checks for
> correct behavior of different rounding modes.  However, for
> quad-precision floating-point, it requires that the number
> 0.10000000000000000000000000000000000481 be exactly represented.  Since
> the PowerPC long double implementation (double-double) only guarantees
> 31 bits of precision, the test fails for the real(qp) portions of the
> test.  Thus this patch marks the test invalid for PowerPC Linux for now.
> (We may want to also do this for other subtargets; let me know if so.)
> 
> At such time as IEEE 128-bit floating-point is supported by the PowerPC
> port, we should revisit this.
> 
> Is this ok for trunk and 4.9?

Looks good to me. Thanks for digging into test-suite failures!

However, it probably wouldn't harm either to add a test case for
powerpc*-*-* (which might also work with hppa-*-*-hpux) for the rounding.
I worry less about the quad-precision (qp) rounding, but I think it would
be useful to test that the single and double precision rounding work. But
as it requires an extra file, one could also check for the 31bit QP as
well. Thus, if you have some spare cycles and agree on the usefulness ...

Tobias

> 2014-06-29  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
> 
> 	* gfortran.dg/round_4.f90: Skip for powerpc*-*-linux* since the
> 	test requires greater precision than the current PowerPC long
> 	double implementation supports.