Re: [Libmesh-users] Crash when computing parallel

[John P. <pet...@cf...>]

Hi Tim,

I'm almost afraid to ask... how did you ever find that one!?
Do you have a link to a GCC-3.3.5 bug report for this issue?

-John





Tim Kr=F6ger writes:
 > Dear John and all,
 >=20
 > On Tue, 29 Aug 2006, Tim Kr=F6ger wrote:
 >=20
 > > I have been running in optimized mode so far.  I now tried debug m=
ode
 > > and got the same results as you.  After switching back to optimize=
d
 > > mode, I again got the results I got before.
 > >
 > > Maximum number of processes without crash (n.t. =3D not tested):
 > >
 > > =09=09debug=09optimized
 > > grid 1*1*1=091=091
 > > grid 2*1*1=093=093
 > > grid 3*1*1=095=095
 > > grid 4*1*1=097=097=09// seems to be no problem for x*1*1
 > > grid 3*3*3=09n.t.=091
 > > grid 4*4*4=0927=092=09// this is not a typo
 > > grid 5*5*5=09n.t.=094
 > > grid 6*6*6=09n.t.=095
 >=20
 > I now found out what the problem is: gcc-3.3.5 seems to have an=20
 > optimization bug that makes DofMap::sort_sparsity_row() work wrong, =
so=20
 > that the rows are sometimes not sorted correctly and consequently so=
me=20
 > indices are inserted more than once.  The computed number of nonzero=
=20
 > local matrix row entries is then too large and sometimes larger than=
=20
 > the total number of columns of the matrix, in which case PETSc's=20
 > function MatCreateMPIAIJ() crashes.
 >=20
 > As a workaround, I have inserted a call to a dummy function, so that=
=20
 > the compiler does not optimize (see attached patch).  My test progra=
m=20
 > now works fine.
 >=20
 > Since other users might also be using gcc-3.3.5, it might be=20
 > appropriate to commit this change to cvs.  If you know a better way =
of=20
 > avoiding optimization at a certain place, please let me know.
 >=20
 > Best Regards,
 >=20
 > Tim? patch
 > Index: include/base/dof_map.h
 > =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
 > RCS file: /cvsroot/libmesh/libmesh/include/base/dof_map.h,v
 > retrieving revision 1.17
 > diff -c -r1.17 dof_map.h
 > *** include/base/dof_map.h=0928 Jun 2006 21:32:21 -0000=091.17
 > --- include/base/dof_map.h=0915 Sep 2006 07:20:56 -0000
 > ***************
 > *** 504,509 ****
 > --- 504,516 ----
 >     DofConstraints _dof_constraints;
 >  =20
 >   #endif
 > +=20
 > +   /**
 > +    * Dummy function that does nothing but can be used to prohibit
 > +    * compiler optimization in some situations where some compilers=

 > +    * have optimization bugs.
 > +    */
 > +   static void _dummy_function(void);
 >   };
 >  =20
 >  =20
 > ***************
 > *** 568,573 ****
 > --- 575,584 ----
 >   =09{
 >   =09  std::swap (*a, *b);
 >  =20
 > + =09  /* Prohibit optimization at this point since gcc 3.3.5 seems
 > + =09     to have a bug.  */
 > + =09  this->_dummy_function();
 > +=20
 >   =09  if (a =3D=3D begin) break;
 >   =09 =20
 >   =09  b=3Da;
 > Index: src/base/dof_map.C
 > =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
 > RCS file: /cvsroot/libmesh/libmesh/src/base/dof_map.C,v
 > retrieving revision 1.94
 > diff -c -r1.94 dof_map.C
 > *** src/base/dof_map.C=0912 Jun 2006 22:51:29 -0000=091.94
 > --- src/base/dof_map.C=0915 Sep 2006 07:20:56 -0000
 > ***************
 > *** 1554,1556 ****
 > --- 1554,1561 ----
 >   #endif // #ifdef ENABLE_AMR
 >  =20
 >   }
 > +=20
 > + void DofMap::_dummy_function(void)
 > + {
 > + }
 > +=20