http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57096
--- Comment #3 from janus at gcc dot gnu.org 2013-04-28 13:19:03 UTC ---
The simplest test case I found contains two files:
ModA.f03:
---------
module ModA
implicit none
integer, allocatable :: gA
end module ModA
test.f03:
---------
subroutine init()
use moda
implicit none
print *,ga
end subroutine
To trigger the bug, they need to be compiled with:
gfortran -std=f2003 -c ModA.f03
gfortran -fdump-tree-original -c test.f03
The "-std=f2003" is required to reproduce, and also two separate files (though
I'm not sure why). One then gets the following dump:
init ()
{
extern integer(kind=4) * ga;
try
{
ga = 0B;
{
struct __st_parameter_dt dt_parm.0;
dt_parm.0.common.filename = &"test.f03"[1]{lb: 1 sz: 1};
dt_parm.0.common.line = 5;
dt_parm.0.common.flags = 128;
dt_parm.0.common.unit = 6;
_gfortran_st_write (&dt_parm.0);
_gfortran_transfer_integer_write (&dt_parm.0, ga, 4);
_gfortran_st_write_done (&dt_parm.0);
}
}
finally
{
if (ga != 0B)
{
__builtin_free ((void *) ga);
}
}
}
which shows that 'gA' is both nullified at the beginning and deallocated at the
end. Same behavior with 4.7, 4.8 and trunk.
Removing the -std=f2003 or putting the module into the same file as the
subroutine makes the auto-deallocation (and nullification) go away. Puzzling.
The patch in comment 2 fixes it and regtests cleanly.
