[Bug fortran/51529] [OOP] gfortran.dg/class_to_type_1.f03 is miscompiled: Uninitialized variable used

["burnus at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51529

--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> 2011-12-13 22:18:44 UTC ---
The problem is the following, for:
  allocate(t2 :: x(10))
one has:
  x._data.data = (void * restrict) __builtin_malloc (640);

I had now expected that one would do:
  for (i = 1; i <= 10; i++)
    __builtin_memcpy ((void *) x._data,
                      (void *) &__vtab_MAIN___T2._def_init,
                      __vtab_MAIN___T2._size)

However, the current code does:
  struct t2 D.1921;
  struct t2 t2.2;
  integer(kind=8) D.1919;
  struct t[0:] * restrict D.1918;
  D.1918 = (struct t[0:] * restrict) x._data.data;
  D.1919 = x._data.offset;
  t2.2.t.b.data = 0B;
  t2.2.z = __complex__ (3.2999999523162841796875e+0,
4.400000095367431640625e+0);
  D.1921 = t2.2;
So far so good. That's the same as __vtab_MAIN___T2._def_init. Disadvantage:
Code duplication. Advantage: The information in in the same file (translation
unit).

However, instead of just doing a __builtin_memcpy in a loop, one calls:
  __vtab_MAIN___T2._copy (&D.1921,
                          (struct t *) D.1918
                          + (sizetype) ((S.3 + D.1919) * (integer(kind=8))
                          x._vptr->_size));

This has several disadvantages: First, makes the advantage of having all data
in the same translation unit void as one calls a function, located in another
translation unit. It is also much slower as _copy does many checks which we
know shouldn't matter. For MOLD= or a type-spec we know that the destination
does not have allocated allocatable components.


However, I do now understand why one needs for SOURCE= to memset the source to
NULL - at least as long _copy not only copies the data but also frees it. The
latter could be also left to _free. - Actually, I am in favour of separating
_copy and _free. As this issue shows, there are cases where one does not want
to combine them, leading to work around actions (memset). I think only for
polymorphic assignment, one needs _free + _copy, for allocate with SOURCE= a
_copy should be enough.


The reason for the crash is:

__copy_MAIN___T2 (struct t2 & restrict src, struct t2 & restrict dst)
{
  if (dst->t.b.data != 0B)
      __builtin_free ((void *) dst->t.b.data);

where dst == x._data.data, where the latter and thus also
x._data.data->t.b.data is filled with random memory.