https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128735

            Bug ID: 128735
           Summary: [OOP] END ASSOCIATE restores a CLASS array selector's
                    old descriptor
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jvdelisle at gcc dot gnu.org
  Target Milestone: ---

If the selector of ASSOCIATE (or SELECT TYPE) is a polymorphic array
pointer or allocatable, END ASSOCIATE copies the selector's descriptor
from before the block back into it.  This undoes a pointer assignment
to the selector made inside the block:

program p
  class(*), pointer :: q(:)
  integer, target :: a(2) = [1, 2], b(3) = [3, 4, 5]
  q => a
  associate (x => q)
    q => b
  end associate
  print *, size (q)    ! expect 3
end program

gfortran prints 2; flang prints 3.  For an allocatable selector it
brings back a freed array, so this gives a double free:

program p
  call s
contains
  subroutine s
    class(*), allocatable :: q(:)
    allocate (q, source=[1, 2])
    associate (x => q)
      deallocate (q)
    end associate
  end subroutine
end program

free(): double free detected in tcache 2

CLASS(*) and CLASS(t) selectors, both variables and components (obj%p),
are affected, with or without a SELECT TYPE on the associate name.
Scalar CLASS selectors and non-polymorphic arrays are not affected.

trans_associate_var (trans-stmt.cc) makes a temporary class container
for the associate name with gfc_conv_class_to_class (..., copyback=true).
-fdump-tree-original for the first example shows the copy-back:

    finally
      {
        q._data = class.0._data;
        q._vptr = class.0._vptr;
        q._len = class.0._len;
        q._len = class.0._len;
      }

An associate name is neither POINTER nor ALLOCATABLE (F2023 11.1.3.3;
see PR54322), so the block cannot change _data or _vptr through it,
and no copy-back is needed.
  • [Bug fortran/128735] New: [OOP] ... jvdelisle at gcc dot gnu.org via Gcc-bugs

Reply via email to