[testsuite] Fix gfortran.dg/vect/fast-math-pr38968.f90 timeout.

Submitted by Steve Ellcey on Aug. 20, 2010, 9:33 p.m.

Details

Message ID 201008202133.o7KLX4E09618@lucas.cup.hp.com
State New
Headers show

Commit Message

Steve Ellcey Aug. 20, 2010, 9:33 p.m.
On some slow (or overloaded) IA64 machines that I use this test still
times out even with the increased timeout factor.  Since the loops
aren't vectorized on IA64 (or SPARC) I was wondering if it would make
sense to just skip this test on those platforms instead of running it
and then having an xfail on the vectorization check.  This would avoid
having to increase the dg-timeout on the test even more.  If the 
platform isn't vectorizing it is not clear to me that the test has
any value anyway.

Tested on IA64 (now shows as UNSUPPORTED) and on X86 (still passes).

OK for checkin?

Steve Ellcey
sje@cup.hp.com


2010-08-20  Steve Ellcey  <sje@cup.hp.com>

	* gfortran.dg/vect/fast-math-pr38969.f90: Skip if not vectorizing.

Comments

Ira Rosen Aug. 24, 2010, 10:50 a.m.
gcc-patches-owner@gcc.gnu.org wrote on 21/08/2010 12:33:04 AM:

>
> On some slow (or overloaded) IA64 machines that I use this test still
> times out even with the increased timeout factor.  Since the loops
> aren't vectorized on IA64 (or SPARC) I was wondering if it would make
> sense to just skip this test on those platforms instead of running it
> and then having an xfail on the vectorization check.  This would avoid
> having to increase the dg-timeout on the test even more.  If the
> platform isn't vectorizing it is not clear to me that the test has
> any value anyway.
>
> Tested on IA64 (now shows as UNSUPPORTED) and on X86 (still passes).
>
> OK for checkin?

I am not sure I am authorized to review testsuite patches, but I am OK with
this change.

Thanks,
Ira

>
> Steve Ellcey
> sje@cup.hp.com
>
>
> 2010-08-20  Steve Ellcey  <sje@cup.hp.com>
>
>    * gfortran.dg/vect/fast-math-pr38969.f90: Skip if not vectorizing.
>
>
> Index: gfortran.dg/vect/fast-math-pr38968.f90
> ===================================================================
> --- gfortran.dg/vect/fast-math-pr38968.f90   (revision 163393)
> +++ gfortran.dg/vect/fast-math-pr38968.f90   (working copy)
> @@ -1,4 +1,5 @@
>  ! { dg-timeout-factor 4.0 }
> +! { dg-skip-if "" { ia64-*-* sparc*-*-* } { "*" } { "" } }
>  program mymatmul
>    implicit none
>    integer, parameter :: kp = 4
> @@ -18,5 +19,5 @@ program mymatmul
>
>  end program mymatmul
>
> -! { dg-final { scan-tree-dump "vectorized 1 loops" "vect"  { xfail
> { ia64-*-* sparc*-*-* } } } }
> +! { dg-final { scan-tree-dump "vectorized 1 loops" "vect" } }
>  ! { dg-final { cleanup-tree-dump "vect" } }
Tobias Burnus Aug. 24, 2010, 2:08 p.m.
On 08/24/2010 12:50 PM, Ira Rosen wrote:
 >
>> times out even with the increased timeout factor.  Since the loops
>> aren't vectorized on IA64 (or SPARC) I was wondering if it would make
>> sense to just skip this test on those platforms instead of running it
>> and then having an xfail on the vectorization check.

>>   ! { dg-timeout-factor 4.0 }
>> +! { dg-skip-if "" { ia64-*-* sparc*-*-* } { "*" } { "" } }

I am also fine with that change but I wonder about "loops aren't 
vectorized": Can't they be vectorized, shouldn't they be vectorized, or 
does it just happen that they are not vectorized?
And: How about adding a small comment why the test is skipped?

Tobias
Ira Rosen Aug. 25, 2010, 7:18 a.m.
Tobias Burnus <burnus@net-b.de> wrote on 24/08/2010 05:08:23 PM:

> On 08/24/2010 12:50 PM, Ira Rosen wrote:
>  >
> >> times out even with the increased timeout factor.  Since the loops
> >> aren't vectorized on IA64 (or SPARC) I was wondering if it would make
> >> sense to just skip this test on those platforms instead of running it
> >> and then having an xfail on the vectorization check.
>
> >>   ! { dg-timeout-factor 4.0 }
> >> +! { dg-skip-if "" { ia64-*-* sparc*-*-* } { "*" } { "" } }
>
> I am also fine with that change but I wonder about "loops aren't
> vectorized": Can't they be vectorized, shouldn't they be vectorized, or
> does it just happen that they are not vectorized?

According to this thread
http://gcc.gnu.org/ml/gcc-patches/2009-04/msg02459.html, the loop can't be
vectorized because of missing instruction.

Ira

> And: How about adding a small comment why the test is skipped?
>
> Tobias
Steve Ellcey Sept. 1, 2010, 9:18 p.m.
On Wed, 2010-08-25 at 10:18 +0300, Ira Rosen wrote:
> 
> Tobias Burnus <burnus@net-b.de> wrote on 24/08/2010 05:08:23 PM:
> 
> > On 08/24/2010 12:50 PM, Ira Rosen wrote:
> >  >
> > >> times out even with the increased timeout factor.  Since the loops
> > >> aren't vectorized on IA64 (or SPARC) I was wondering if it would make
> > >> sense to just skip this test on those platforms instead of running it
> > >> and then having an xfail on the vectorization check.
> >
> > >>   ! { dg-timeout-factor 4.0 }
> > >> +! { dg-skip-if "" { ia64-*-* sparc*-*-* } { "*" } { "" } }
> >
> > I am also fine with that change but I wonder about "loops aren't
> > vectorized": Can't they be vectorized, shouldn't they be vectorized, or
> > does it just happen that they are not vectorized?
> 
> According to this thread
> http://gcc.gnu.org/ml/gcc-patches/2009-04/msg02459.html, the loop can't be
> vectorized because of missing instruction.
> 
> Ira
> 
> > And: How about adding a small comment why the test is skipped?
> >
> > Tobias

I have gone ahead and checked in the patch after adding a comment about
why it isn't vectorized.

Steve Ellcey
sje@cup.hp.com

Patch hide | download patch | download mbox

Index: gfortran.dg/vect/fast-math-pr38968.f90
===================================================================
--- gfortran.dg/vect/fast-math-pr38968.f90	(revision 163393)
+++ gfortran.dg/vect/fast-math-pr38968.f90	(working copy)
@@ -1,4 +1,5 @@ 
 ! { dg-timeout-factor 4.0 }
+! { dg-skip-if "" { ia64-*-* sparc*-*-* } { "*" } { "" } }
 program mymatmul
   implicit none
   integer, parameter :: kp = 4
@@ -18,5 +19,5 @@  program mymatmul
 
 end program mymatmul
 
-! { dg-final { scan-tree-dump "vectorized 1 loops" "vect"  { xfail { ia64-*-* sparc*-*-* } } } }
+! { dg-final { scan-tree-dump "vectorized 1 loops" "vect" } }
 ! { dg-final { cleanup-tree-dump "vect" } }