[Bug fortran/51250] New: [4.7 Regression] Bug with SUM(,dim,mask)

[Bug fortran/51250] New: [4.7 Regression] Bug with SUM(,dim,mask)

[anlauf at gmx dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51250

             Bug #: 51250
           Summary: [4.7 Regression] Bug with SUM(,dim,mask)
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: anlauf@gmx.de


The following program produces a wrong result:

program gfcbug115
  implicit none
  integer :: n_obstype = 2
  integer :: nboxes = 1
  integer :: nprocs = 1
  integer :: nbox, j
  integer, allocatable :: nbx(:,:), pes(:)

  allocate (pes(nboxes))
  allocate (nbx(n_obstype,nboxes))
  nbx(:,:) = 1
  do j = 1, nboxes
     pes(j) = modulo (j-1, nprocs)
  end do
  do nbox=1, nboxes
     print '(20i4)', nbox, nbx(:,nbox)
  end do
  do j = 0, nprocs-1
     print '(i4,20l4)', j, spread (pes==j,dim=1,ncopies=n_obstype)
     print '(20i4)', j, &
          sum (nbx,dim=2,mask=spread (pes==j,dim=1,ncopies=n_obstype)), &
          shape (sum (nbx,dim=2,mask=spread (pes==j,dim=1,ncopies=n_obstype)))
  end do
end program gfcbug115


gfortran 4.7 svn rev. 181390:

   1   1   1
   0   T   T
   0   1   1

gfortran 4.6 and earlier:
   1   1   1
   0   T   T
   0   1   1   2

The latter is correct.

[Bug fortran/51250] [4.7 Regression] Bug with SUM(,dim,mask)

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51250

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011-11-20
     Ever Confirmed|0                           |1

--- Comment #1 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-11-20 23:22:08 UTC ---
Revision 180683 also gives

   1   1   1
   0   T   T
   0   1   1   2

[Bug fortran/51250] [4.7 Regression] Bug with SUM(,dim,mask)

[burnus at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51250

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
                 CC|                            |burnus at gcc dot gnu.org
   Target Milestone|---                         |4.7.0

--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> 2011-11-21 07:46:46 UTC ---
Workaround: -Os  (optimize for code size)

I would assume that it is due to Mikael's scalarizer patch set (Rev. 180842 to
Rev. 180922), cf. http://gcc.gnu.org/ml/fortran/2011-11/msg00046.html

Similar issue exists for PRODUCT, which is not surprising as they mostly share
the same code path.

(The expr->shape is correctly set in iresolve.c - but only if the shape and dim
is known at compile time, thus not in this case.)

(I couldn't quickly find the right spot in trans*.c)

[Bug fortran/51250] [4.7 Regression] Bug with SUM(,dim,mask)

[mikael at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51250

Mikael Morin <mikael at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mikael at gcc dot gnu.org
         AssignedTo|unassigned at gcc dot       |mikael at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #3 from Mikael Morin <mikael at gcc dot gnu.org> 2011-11-22 19:02:27 UTC ---
I will take care of this one after mid-december.

Beside the workaround proposed by Tobias, you can use a variable instead of a
constant for SUM's DIM argument.
Or you can patch gfc_inline_intrinsic_function_p (in trans-intrinsic.c) to
return false in the SUM case.

[Bug fortran/51250] [4.7 Regression] Bug with SUM(,dim,mask)

[mikael at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51250

--- Comment #4 from Mikael Morin <mikael at gcc dot gnu.org> 2011-11-23 23:20:16 UTC ---
(In reply to comment #3)
> I will take care of this one after mid-december.
> 
I couldn't help having a look.

diff --git a/gcc/fortran/trans-array.c b/gcc/fortran/trans-array.c
index 2fb2d34..ee8f896 100644
--- a/gcc/fortran/trans-array.c
+++ b/gcc/fortran/trans-array.c
@@ -1087,7 +1087,7 @@ gfc_trans_create_temp_array (stmtblock_t * pre,
stmtblock_t * post, gfc_ss * ss,
     for (s = ss; s; s = s->parent)
       for (n = 0; n < s->loop->dimen; n++)
     {
-      dim = get_scalarizer_dim_for_array_dim (ss, ss->dim[n]);
+      dim = get_scalarizer_dim_for_array_dim (ss, s->dim[n]);

       /* For a callee allocated array express the loop bounds in terms
          of the descriptor fields.  */



This gives the following change in the dump. Then one gets the expected result.

--- comment_0.f90.003t.original.bad    2011-11-24 00:15:42.045565178 +0100
+++ comment_0.f90.003t.original.good    2011-11-24 00:12:16.409796545 +0100
@@ -473,7 +473,7 @@
                           S.18 = 0;
                           while (1)
                             {
-                              if (atmp.17.dim[1].ubound -
atmp.17.dim[1].lbound < S.18) goto L.14;
+                              if (atmp.17.dim[0].ubound -
atmp.17.dim[0].lbound < S.18) goto L.14;
                               {
                                 integer(kind=4) D.2013;
                                 integer(kind=4) val.19;
@@ -608,7 +608,7 @@
                               __builtin_free (D.2035);
                             }
                         }
-                        D.2051 = atmp.23.dim[1].ubound -
atmp.23.dim[1].lbound;
+                        D.2051 = atmp.23.dim[0].ubound -
atmp.23.dim[0].lbound;
                         atmp.24.dtype = 265;
                         atmp.24.dim[0].stride = 1;
                         atmp.24.dim[0].lbound = 0;

[Bug fortran/51250] [4.7 Regression] Bug with SUM(,dim,mask)

[anlauf at gmx dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51250

--- Comment #5 from Harald Anlauf <anlauf at gmx dot de> 2011-11-24 22:39:10 UTC ---
(In reply to comment #4)

The patch in comment #4 works for me without regressions!

Thanks,
Harald

[Bug fortran/51250] [4.7 Regression] Bug with SUM(,dim,mask)

[mikael at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51250

--- Comment #6 from Mikael Morin <mikael at gcc dot gnu.org> 2011-11-25 20:18:32 UTC ---
Author: mikael
Date: Fri Nov 25 20:18:21 2011
New Revision: 181730

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=181730
Log:
fortran/
    PR fortran/51250
    PR fortran/43829
    * trans-array.c (gfc_trans_create_temp_array): Get dimension from
    the right gfc_ss struct.

testsuite/
    PR fortran/51250
    PR fortran/43829
    * gfortran.dg/inline_sum_3.f90: New test.


Added:
    trunk/gcc/testsuite/gfortran.dg/inline_sum_3.f90
Modified:
    trunk/gcc/fortran/ChangeLog
    trunk/gcc/fortran/trans-array.c
    trunk/gcc/testsuite/ChangeLog

[Bug fortran/51250] [4.7 Regression] Bug with SUM(,dim,mask)

[mikael at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51250

Mikael Morin <mikael at gcc dot gnu.org> changed:

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

--- Comment #7 from Mikael Morin <mikael at gcc dot gnu.org> 2011-11-25 21:17:12 UTC ---
Fixed.
Thanks for the report.