Re: [Patch, fortran] PR121972 - [PDT] Fix for the pdt_86 leak.

Hello Jerry and Harald,

The patch does what it says on the box and regression tests correctly.

Concerning, Haralds suggestion: I don't see any problem in principle,
as long as it is recognised that both calls to has_class_alloc_comp
are negated and so types with class pointer components get through the
gate. gfc_deallocate_alloc_comp_no_caf should ignore them but I
haven't checked.

In chunks 2 and 3, please move expr->ts.type == BT_DERIVED to the
start of the condition to guard the use of the typespec derived field
in the call to has_class_alloc_comp. It should be OK as it is but it
is worthwhile taking that extra little bit of care :-)

Finally, the leaking part of pdt_86.f03 has been isolated as an
additional test. If you think that it is useful, please add it to the
patch. (Renamed to pdt_86_leak_test.f90?)

OK for mainline and backporting.

Cheers

Paul

On Tue, 23 Jun 2026 at 00:05, Jerry D <[email protected]> wrote:
>
> On 6/22/26 1:35 PM, Harald Anlauf wrote:
> > Hi Jerry!
> >
> > Am 22.06.26 um 8:37 PM schrieb Jerry D:
> >> See attached patch.
> >>
> >> Regression tested on x86_64.
> >>
> >> A new test case is added to asan to catch this in the future.
> >
> > I'd rather leave the review to Paul, but I have one question:
> > As is, the new helper has_class_alloc_comp might return true for
> > pointer components, but I think it should not.
> >
> > Can you please check?
> >
>
> Harald, I think you are right. I changed it to this:
>
> static bool
> has_class_alloc_comp (gfc_symbol *der)
> {
>    for (gfc_component *c = der->components; c; c = c->next)
>      if (c->ts.type == BT_CLASS && !c->attr.pointer)
>        return true;
>    return false;
> }
>
> I will wait for Paul to review.
>
> Thanks..
>
>
>
>
> > Cheers,
> > Harald
> >
> >> OK for mainline.
> >>
> >> Regards,
> >>
> >> Jerry
> >> ---
> >> fortran: Fix memory leak for array structure-constructor element [PR121972]
> >>
> >> Fix the remaining leak in pdt_86.f03. Added a new test in asan to avoid it.
> >>
> >>    PR fortran/121972
> >>
> >> gcc/fortran/ChangeLog:
> >>
> >>    * trans-array.cc (has_class_alloc_comp): New helper; returns true if
> >>    derived type DER has any CLASS component.
> >>    (gfc_trans_array_ctor_element): Also free allocatable components when
> >>    the element expression is EXPR_STRUCTURE, skipping types that contain
> >>    CLASS components to avoid freeing stack-allocated _data pointers.
> >>    (gfc_constructor_is_owned_alloc_comp): Likewise treat EXPR_STRUCTURE
> >>    elements as owned only when the derived type has no CLASS components.
> >>    * trans-expr.cc (gfc_trans_alloc_subarray_assign): Free allocatable
> >>    components of the component's temporary descriptor before nulling its
> >>    data pointer, for non-variable source expressions.
> >>
> >> gcc/testsuite/ChangeLog:
> >>
> >>    * gfortran.dg/class_array_15.f03: Adjust expected free count.
> >>    * gfortran.dg/derived_constructor_comps_6.f90: Adjust expected free count.
> >>    * gfortran.dg/asan/structure_constructor_alloc_comp_leak_1.f90: New test.
> >> ---
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >
>

! { dg-do compile }
! { dg-options "-fdump-tree-original" }
!
! Check the fix for the memory leak in pdt_77.f03. This test is a reduced
! version of the original that isolates the source of the leak. Note that
! the leak is not PDT specific.
!
module input_output_pair_m
  implicit none
  type tensor_t                             ! No longer a PDT!
    real(kind(1e0)), allocatable :: values_(:)
  end type
end module

  use input_output_pair_m
  call test
contains
  subroutine test
    implicit none
    type(tensor_t), allocatable :: inputs(:)
    integer :: i
    inputs = [(tensor_t([real(i)]), i=1,2)] ! Leaked 8 bytes in 2 blocks.
  end
end
! { dg-final { scan-tree-dump-times "__builtin_malloc" 7 "original" } }
! { dg-final { scan-tree-dump-times "__builtin_free" 8 "original" } }