[Bug testsuite/51057] New: FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[Bug testsuite/51057] New: FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

             Bug #: 51057
           Summary: FAIL: gfortran.dg/quad_2.f90  -O0  execution test on
                    powerpc*-*-*
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: testsuite
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: dominiq@lps.ens.fr
                CC: burnus@net-b.de
              Host: powerpc*-*-*
            Target: powerpc*-*-*
             Build: powerpc*-*-*


gfortran.dg/quad_2.f90 fails on powerpc*-*-* (see
http://gcc.gnu.org/ml/gcc-testresults/2011-11/msg00620.html ). This is due to a
less precise representation of REAL(16) on ppc (i.e., two REAL(8)) than on
Intel platforms. The following patch fixes the failure

--- /opt/gcc/work/gcc/testsuite/gfortran.dg/quad_2.f90    2011-11-06
20:55:43.000000000 +0100
+++ quad_2_db.f90    2011-11-09 14:06:48.000000000 +0100
@@ -48,8 +48,8 @@ program test_qp
      case (16)
        if (str1 /= "   1.00000000000000000000000000000000000") call abort()
        if (str2 /= "1.00000000000000000000000000000000000") call abort()
-       if (str3 /= "   1.41421356237309504880168872420969798") call abort()
-       if (str4 /= "1.41421356237309504880168872420969798") call abort()
+       if (str3(1:37) /= "   1.41421356237309504880168872420969") call abort()
+       if (str4(1:34) /= "1.41421356237309504880168872420969") call abort()
        block
          real(qp), volatile :: fp2a
          fp2a = 2.0_qp

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[kargl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

kargl at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |kargl at gcc dot gnu.org

--- Comment #1 from kargl at gcc dot gnu.org 2011-11-09 18:51:10 UTC ---
(In reply to comment #0)
> gfortran.dg/quad_2.f90 fails on powerpc*-*-* (see
> http://gcc.gnu.org/ml/gcc-testresults/2011-11/msg00620.html ). This is due to a
> less precise representation of REAL(16) on ppc (i.e., two REAL(8)) than on
> Intel platforms. The following patch fixes the failure
> 
> --- /opt/gcc/work/gcc/testsuite/gfortran.dg/quad_2.f90    2011-11-06
> 20:55:43.000000000 +0100
> +++ quad_2_db.f90    2011-11-09 14:06:48.000000000 +0100
> @@ -48,8 +48,8 @@ program test_qp
>       case (16)
>         if (str1 /= "   1.00000000000000000000000000000000000") call abort()
>         if (str2 /= "1.00000000000000000000000000000000000") call abort()
> -       if (str3 /= "   1.41421356237309504880168872420969798") call abort()
> -       if (str4 /= "1.41421356237309504880168872420969798") call abort()
> +       if (str3(1:37) /= "   1.41421356237309504880168872420969") call abort()
> +       if (str4(1:34) /= "1.41421356237309504880168872420969") call abort()
>         block
>           real(qp), volatile :: fp2a
>           fp2a = 2.0_qp

Please XFAIL this testcase on powerpc, or make this
testcase i686/x64_64 specific.  The problem runs rather
deep.  See PR 46481 and 47032.

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[iains at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

Iain Sandoe <iains at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011-12-01
                 CC|                            |iains at gcc dot gnu.org
     Ever Confirmed|0                           |1

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[iains at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #2 from Iain Sandoe <iains at gcc dot gnu.org> 2011-12-03 20:36:43 UTC ---
whilst I appreciate that there are wider issues with IBM long double (and its
availability) 
[I read the two PRs you cross-referenced].

In this case, it appears to be a different situation.
- what Dominique is indicating is that the tests are likely failing on an IBM
long double target because the number of mantissa bits is 106 rather than the
112 for float128.

So I wonder what opportunities exist for adjusting the test expectation
depending on the target.
(since this, presumably, will affect all IBM long double targets).

from 'c' we have some #defines that would allow us to detect the IBM case on
ppc  (although, presumably because it's provided by a library, we don't get the
same for x86).  It would even be a reasonable first approximation to assume IBM
long double on ppc and float128 on x86.

---

This would allow some kind of target capability test in tcl - which could
thence regulate some conditional compilation of the Fortran, I guess?  I assume
the Fortran can't see the 'c' defines (ISTR a PR about that somewhen).

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[sgk at troutmask dot apl.washington.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #3 from Steve Kargl <sgk at troutmask dot apl.washington.edu> 2011-12-03 20:46:36 UTC ---
On Sat, Dec 03, 2011 at 08:36:43PM +0000, iains at gcc dot gnu.org wrote:
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057
> 
> --- Comment #2 from Iain Sandoe <iains at gcc dot gnu.org> 2011-12-03 20:36:43 UTC ---
> whilst I appreciate that there are wider issues with IBM long double (and its
> availability) 
> [I read the two PRs you cross-referenced].
> 
> In this case, it appears to be a different situation.
> - what Dominique is indicating is that the tests are likely failing on an IBM
> long double target because the number of mantissa bits is 106 rather than the
> 112 for float128.

113 bits for IEEE binary 128 format.


> So I wonder what opportunities exist for adjusting the test expectation
> depending on the target.
> (since this, presumably, will affect all IBM long double targets).
> 
> from 'c' we have some #defines that would allow us to detect the IBM case on
> ppc  (although, presumably because it's provided by a library, we don't get the
> same for x86).  It would even be a reasonable first approximation to assume IBM
> long double on ppc and float128 on x86.

This test came into being when libquadmath came into the tree.
The correct, and only, solution for targets whose long double
representation is double-double math is to XFAIL the test on those
targets.

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[iains at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #4 from Iain Sandoe <iains at gcc dot gnu.org> 2011-12-03 21:08:26 UTC ---
(In reply to comment #3)
> On Sat, Dec 03, 2011 at 08:36:43PM +0000, iains at gcc dot gnu.org wrote:

> This test came into being when libquadmath came into the tree.
> The correct, and only, solution for targets whose long double
> representation is double-double math is to XFAIL the test on those
> targets.

Elsewhere in the test suite, if a test is meaningless or impossible to complete
in the absence of a capability - the test is made conditional on that
capability, rather than having a long list of XFAILs.

If this is really not possible here, then I'll prepare an XFAIL patch, but it's
not an ideal solution.

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[sgk at troutmask dot apl.washington.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #5 from Steve Kargl <sgk at troutmask dot apl.washington.edu> 2011-12-03 21:22:39 UTC ---
On Sat, Dec 03, 2011 at 09:08:26PM +0000, iains at gcc dot gnu.org wrote:
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057
> 
> --- Comment #4 from Iain Sandoe <iains at gcc dot gnu.org> 2011-12-03 21:08:26 UTC ---
> (In reply to comment #3)
> > On Sat, Dec 03, 2011 at 08:36:43PM +0000, iains at gcc dot gnu.org wrote:
> 
> > This test came into being when libquadmath came into the tree.
> > The correct, and only, solution for targets whose long double
> > representation is double-double math is to XFAIL the test on those
> > targets.
> 
> Elsewhere in the test suite, if a test is meaningless or impossible
> to complete in the absence of a capability - the test is made
> conditional on that capability, rather than having a long list of XFAILs.

Well, there is 

! { dg-require-effective-target fortran_large_real }

but I believe that that is insufficient in detecting ppc's use
of a double-double representation for its long double.  I also
don't know if there is an equivalent effective-target to provide
the required differentiation.

> If this is really not possible here, then I'll prepare an XFAIL
> patch, but it's not an ideal solution.

It's not a long list if you XFAIL this for all ppc systems.
You need something along the lines

! { dg-do run { target { ! { ppc*-*-* } } } }

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

Dominique d'Humieres <dominiq at lps dot ens.fr> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at libertysurf
                   |                            |dot fr, ro at CeBiTec dot
                   |                            |Uni-Bielefeld.DE

--- Comment #6 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-12-03 22:01:22 UTC ---
The main problem with this test is not ppc platform implementing real(16) as
two real(8), but rather platforms that don't have sqrtl (see
http://gcc.gnu.org/ml/fortran/2011-11/msg00079.html ).

Now this test has been motivated by the libquadmath, but is not specific and I
don't see why it should be xfailed if real(16) is available on a platform as
long as it gives a sensible answer as it is the case for ppc.

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[sgk at troutmask dot apl.washington.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #7 from Steve Kargl <sgk at troutmask dot apl.washington.edu> 2011-12-03 22:24:30 UTC ---
On Sat, Dec 03, 2011 at 10:01:22PM +0000, dominiq at lps dot ens.fr wrote:
> 
> --- Comment #6 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-12-03 22:01:22 UTC ---
> The main problem with this test is not ppc platform implementing real(16) as
> two real(8), but rather platforms that don't have sqrtl (see
> http://gcc.gnu.org/ml/fortran/2011-11/msg00079.html ).

Did you read the thread?  I specifically recommended that
the testcase should be XFAILed for the missing sqrtl, too;
because no one is going to implement a suitable sqrtl in
c99_functions.c.

> Now this test has been motivated by the libquadmath, but is not specific and I
> don't see why it should be xfailed if real(16) is available on a platform as
> long as it gives a sensible answer as it is the case for ppc.

The test is designed to test float128 (aka IEEE 754 binary 128 format).
Double-double math gives you 106 bits of precision, which means it is
impossible to meet the requirements of a 113 bit precision type.
Lowering the accuracy of the test to allow ppc to pass opens up the
possibility of missing a bug in libquadmath when someone makes a
change to the library.

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #8 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-12-03 23:15:32 UTC ---
(In reply to comment #5)
> It's not a long list if you XFAIL this for all ppc systems.
>You need something along the lines
>
> ! { dg-do run { target { ! { ppc*-*-* } } } }

(In reply to comment #7)
> Did you read the thread?  I specifically recommended that
> the testcase should be XFAILed for the missing sqrtl, too;
> because no one is going to implement a suitable sqrtl in
> c99_functions.c.

So the list will be longer than ppc*-*-*.

> Lowering the accuracy of the test to allow ppc to pass opens up the
> possibility of missing a bug in libquadmath when someone makes a
> change to the library.

I am impressed by the coverage! what about

--- /opt/gcc/_gcc_clean/gcc/testsuite/gfortran.dg/quad_2.f90    2011-11-06
20:24:56.000000000 +0100
+++ /opt/gcc/work/gcc/testsuite/gfortran.dg/quad_2.f90    2011-12-04
00:08:16.000000000 +0100
@@ -48,8 +48,13 @@ program test_qp
      case (16)
        if (str1 /= "   1.00000000000000000000000000000000000") call abort()
        if (str2 /= "1.00000000000000000000000000000000000") call abort()
-       if (str3 /= "   1.41421356237309504880168872420969798") call abort()
-       if (str4 /= "1.41421356237309504880168872420969798") call abort()
+       if (size (real_kinds) >= 4) then
+         if (str3 /= "   1.41421356237309504880168872420969798") call abort()
+         if (str4 /= "1.41421356237309504880168872420969798") call abort()
+       else
+         if (str3(1:37) /= "   1.41421356237309504880168872420969") call
abort()
+         if (str4(1:34) /= "1.41421356237309504880168872420969") call abort()
+       end if
        block
          real(qp), volatile :: fp2a
          fp2a = 2.0_qp

?

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[sgk at troutmask dot apl.washington.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #9 from Steve Kargl <sgk at troutmask dot apl.washington.edu> 2011-12-03 23:36:01 UTC ---
On Sat, Dec 03, 2011 at 11:15:32PM +0000, dominiq at lps dot ens.fr wrote:
> (In reply to comment #5)
> > It's not a long list if you XFAIL this for all ppc systems.
> >You need something along the lines
> >
> > ! { dg-do run { target { ! { ppc*-*-* } } } }
> 
> (In reply to comment #7)
> > Did you read the thread?  I specifically recommended that
> > the testcase should be XFAILed for the missing sqrtl, too;
> > because no one is going to implement a suitable sqrtl in
> > c99_functions.c.
> 
> So the list will be longer than ppc*-*-*.
> 

As others report failures, then I suppose the answer is yes.
If you don't like this option, implement a suitable sqrtl().
Go re-read my responses in the other and then read

http://svnweb.freebsd.org/base/head/lib/msun/src/e_sqrtl.c?revision=176720&view=markup

It's only 159 lines of code. 

> > Lowering the accuracy of the test to allow ppc to pass opens up the
> > possibility of missing a bug in libquadmath when someone makes a
> > change to the library.
> 
> I am impressed by the coverage! what about
> 
> --- /opt/gcc/_gcc_clean/gcc/testsuite/gfortran.dg/quad_2.f90    2011-11-06
> 20:24:56.000000000 +0100
> +++ /opt/gcc/work/gcc/testsuite/gfortran.dg/quad_2.f90    2011-12-04
> 00:08:16.000000000 +0100
> @@ -48,8 +48,13 @@ program test_qp
>       case (16)
>         if (str1 /= "   1.00000000000000000000000000000000000") call abort()
>         if (str2 /= "1.00000000000000000000000000000000000") call abort()
> -       if (str3 /= "   1.41421356237309504880168872420969798") call abort()
> -       if (str4 /= "1.41421356237309504880168872420969798") call abort()
> +       if (size (real_kinds) >= 4) then
> +         if (str3 /= "   1.41421356237309504880168872420969798") call abort()
> +         if (str4 /= "1.41421356237309504880168872420969798") call abort()
> +       else
> +         if (str3(1:37) /= "   1.41421356237309504880168872420969") call
> abort()
> +         if (str4(1:34) /= "1.41421356237309504880168872420969") call abort()
> +       end if
>         block
>           real(qp), volatile :: fp2a
>           fp2a = 2.0_qp
> 

Why is it so difficult to accept that the test can be XFAILed.
The above is an ugly hack.  Are you going to fix all other
REAL(16) examples with a similar approach?

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[burnus at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #10 from Tobias Burnus <burnus at gcc dot gnu.org> 2012-01-12 06:54:03 UTC ---
Author: burnus
Date: Thu Jan 12 06:53:59 2012
New Revision: 183121

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=183121
Log:
2012-01-12  Dominique d'Humieres  <dominiq@lps.ens.fr>
            Tobias Burnus  <burnus@net-b.de>

        PR fortran/51057
        PR fortran/51616
        * lib/target-supports.exp
        (check_effective_target_fortran_largest_fp_has_sqrt): New.
        * gfortran.dg/quad_2.f90: Use it, add pattern for IBM's
        real(16).


Modified:
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/gfortran.dg/quad_2.f90
    trunk/gcc/testsuite/lib/target-supports.exp

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[burnus at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |burnus at gcc dot gnu.org

--- Comment #11 from Tobias Burnus <burnus at gcc dot gnu.org> 2012-01-12 07:59:04 UTC ---
Please confirm that this now passes on PowerPC - and, if so, please close.
Thanks for the patch Dominique and thanks to all for the patience.

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[iains at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

--- Comment #12 from Iain Sandoe <iains at gcc dot gnu.org> 2012-01-12 21:27:57 UTC ---
passes now on powerpc-darwin9.
thanks for the patch.

[Bug testsuite/51057] FAIL: gfortran.dg/quad_2.f90 -O0 execution test on powerpc*-*-*

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51057

Dominique d'Humieres <dominiq at lps dot ens.fr> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED

--- Comment #13 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2012-01-12 21:45:27 UTC ---
> passes now on powerpc-darwin9.
> thanks for the patch.

So closing as fixed. Thanks Tobias!