https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110626
Bug ID: 110626
Summary: Duplicated finalization in derived
Product: gcc
Version: 13.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: fortran
Assignee: unassigned at gcc dot gnu.org
Reporter: habbit89 at hotmail dot es
Target Milestone: ---
Created attachment 55520
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55520&action=edit
Example module and test program
After the changes to finalization in gfortran 13 there seems to be an issue
under the following circumstances:
* A derived type A has custom assignment and finalization
* It is used as a component (or as a base) of another type B
Then, assignment of type B will call the assignment of A only once, but the
final subroutine twice, which breaks resource holding code such as reference
counting.
The example has two subroutines, one where two objects of type A are used
directly, and one where objects of type B are used. In both cases, o1 is
initialized to some value, then it is copied to o2, then o2 is overwritten
again. The expected result would, in both cases, be a finalization of the
target before the assignment, then the assignment call.
Compiling and running the example will give the following correct results on
gfortran 11
> $ gfortran-11 -Wall -Wextra -o a-11.out testmod.f90
> $ ./a-11.out
> o1: 7FFFA5E1BD2C
> o2: 7FFFA5E1BD28
> dtor: -1 7FFFA5E1BD28
> copy: 15 from 7FFFA5E1BD2C to 7FFFA5E1BD28
> dtor: 16 7FFFA5E1BD28
> copy: 15 from 7FFFA5E1BD2C to 7FFFA5E1BD28
> dtor: 16 7FFFA5E1BD28
> dtor: 15 7FFFA5E1BD2C
> objects of type B in subroutine
> o1: 7FFFA5E1BD2C
> o2: 7FFFA5E1BD28
> dtor: 15 7FFFA5E1BD28
> copy: 15 from 7FFFA5E1BD2C to 7FFFA5E1BD28
> dtor: 15 7FFFA5E1BD28
> copy: 15 from 7FFFA5E1BD2C to 7FFFA5E1BD28
> dtor: 16 7FFFA5E1BD28
> dtor: 15 7FFFA5E1BD2C
But the following *invalid* results in gfortran 13:
> $ gfortran-13 -Wall -Wextra -o a-13.out testmod.f90
> $ ./a-13.out
> objects of type A in subroutine
> o1: 7FFFCEEDBC2C
> o2: 7FFFCEEDBC28
> dtor: -1 7FFFCEEDBC28
> copy: 15 from 7FFFCEEDBC2C to 7FFFCEEDBC28
> dtor: 16 7FFFCEEDBC28
> copy: 15 from 7FFFCEEDBC2C to 7FFFCEEDBC28
> dtor: 16 7FFFCEEDBC28
> dtor: 15 7FFFCEEDBC2C
> objects of type B in subroutine
> o1: 7FFFCEEDBC24
> o2: 7FFFCEEDBC20
> dtor: -1 7FFFCEEDBC20
> dtor: -1 7FFFCEEDBC2C
> copy: 15 from 7FFFCEEDBC24 to 7FFFCEEDBC2C
> dtor: 16 7FFFCEEDBC20
> dtor: 16 7FFFCEEDBC28
> copy: 15 from 7FFFCEEDBC24 to 7FFFCEEDBC28
> dtor: 16 7FFFCEEDBC20
> dtor: 15 7FFFCEEDBC24
The part where objects of type A are used directly works in both versions.
However, when objects of type *B* are used, gfortran 13 shows the following
behaviour:
* There seems to be a "shadow"/temporary object created at a different location
which is neither o1 nor o2, probably at a stack address.
* The assignment operator runs only once, from o1 to this shadow object.
* The value is then apparently blitted onto/used for o2, which might be okay
except that...
* Before the next assignment, the final subroutine of A runs *twice*, once with
the actual o2 and once with this shadow object.
Thus, given that the assignment code runs once but the finalization runs
*twice*, using this scheme to hold resources (e.g. via ref counting) breaks. In
particular, it is very weird that
I am _assuming_ that the two separate finalizations may be conceptually come
from 1. the overall finalization of B before the assignment, and 2. the
intent(out) for A in subroutine copy. However, both calls use the values
_prior_ to the finalization (since it sets the value to -2 but the calls both
print 16)
