Intel's response:


Hello Andrew,

 

Our development team investigated the issue and determined the following:

 

---------------

Finalization is a tricky thing - there are many different events that trigger a call to a finalizer. Although it may not seem so at first, this program *should* call the finalizer 3 times, not once.

 

The first finalization occurs as a result of this line of code: u = new_t() Here, 'u' is finalized as a result of the following text from the F2008 standard:

   "When an intrinsic assignment statement is executed, the variable is finalized after evaluation of expr and before the definition of the variable."

'Variable' in this line of code is 'u'. Since this is an intrinsic assignment statement, 'u' is the first object to be finalized.

 

The second finalization occurs because of the same line of user code, but this time it's the result of 'new_t' that is finalized according to the following text:

 " If an executable construct references a function, the result is finalized after execution of the innermost executable construct containing the reference." 

Because of this rule, the result of the call to 'new_t' is finalized next.

 

The third finalization occurs at the end of routine 's'. The standard says:

"A nonpointer, nonallocatable object that is not a dummy argument or function result is finalized immediately before it would become undefined due to execution of a RETURN or END statement"  

 

The variable 'u' is a nonpointer, nonallocatable object that is not a dummy argument or function result, so it is finalized at the end of the routine 's'.

 

So the three calls to the finalizer are correct and the results from PGI Fortran are incorrect.

-------------------

 

Since this is not a Intel Fortran Compiler defect, I am closing this case. 

Do not hesitate to contact me if you have any other questions or need further assistance with Intel® Software Development Products. 

 

Thank you for contacting Intel® Online Service Center,

 

Devorah

Intel® Developer Support

Confirmed from 4.9 up to trunk (8.0).

I think this depends a lot on the compiler implementation. Declaring a type does not define an instance of it. I can see that since new_t is returning a function result which could be a pointer, the compiler could end up only creating one instance.

(In reply to Jerry DeLisle from comment #3)
> I think this depends a lot on the compiler implementation.

I don't actually think the calling of finalization routines is supposed to depend on the compiler implementation. F08 is rather clear on where it should happen, I guess.

For the missing cases, see also:
* PR 64290 (No finalization at deallocation of LHS)
* PR 65347 (Final subroutine not called for function result)

and PR 37336 comment 27 which includes both of them.

(In reply to janus from comment #4)
> (In reply to Jerry DeLisle from comment #3)
> > I think this depends a lot on the compiler implementation.
> 
> I don't actually think the calling of finalization routines is supposed to
> depend on the compiler implementation. F08 is rather clear on where it
> should happen, I guess.
> 
> For the missing cases, see also:
> * PR 64290 (No finalization at deallocation of LHS)
> * PR 65347 (Final subroutine not called for function result)
> 
> and PR 37336 comment 27 which includes both of them.

My only thought here is that those conditions may be A or B or C vs A and B and C

If the finalization is to for example deallocate something, would we end up trying to deallocate more than once? It seems to me that it depends on how many actual instances of the object were created.

For info, the actual code I was developing did use the finalization subroutine to deallocate pointers, which resulted in seg faults when compiled with ifort. This was how I discovered the problem.

I also reported the problem to PGI. The latest version of pgfortran (17.7) now produces the same behaviour as ifort:

> pgfortran -V

pgfortran 17.5-0 64-bit target on x86-64 Linux -tp haswell 
PGI Compilers and Tools
Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.

> pgfortran prog.f90 && ./a.out 
 In SUBROUTINE t_final




> pgfortran -V

pgfortran 17.7-0 64-bit target on x86-64 Linux -tp haswell 
PGI Compilers and Tools
Copyright (c) 2017, NVIDIA CORPORATION.  All rights reserved.

> pgfortran prog.f90 && ./a.out 
 In SUBROUTINE t_final
 In SUBROUTINE t_final
 In SUBROUTINE t_final

This is still present in the actual trunk.

(In reply to Jürgen Reuter from comment #7)
> This is still present in the actual trunk.

I can confirm that Intel's interpretation is completely correct. I have a patch in progress that issues the correct number of final calls but which is ordering the evaluation of the function result and the finalization of the lhs in the wrong order.

Paul

The master branch has been updated by Paul Thomas <pault@gcc.gnu.org>:

https://gcc.gnu.org/g:d7caf313525a46f200d7f5db1ba893f853774aee

commit r13-6747-gd7caf313525a46f200d7f5db1ba893f853774aee
Author: Paul Thomas <pault@gcc.gnu.org>
Date:   Sat Mar 18 07:56:23 2023 +0000

    Fortran: Fix bugs and missing features in finalization [PR37336]
    
    2023-03-18  Paul Thomas  <pault@gcc.gnu.org>
    
    gcc/fortran
            PR fortran/103854
            PR fortran/96122
            PR fortran/37336
            * class.cc (finalize_component): Include the missing arguments
            in the call to the component's finalizer wrapper.
            (has_finalizer_component): Do not return true for procedure
            pointer components.
            (finalizer_insert_packed_call): Remove the redundant argument
            in the call to the final subroutine.
            (generate_finalization_wrapper): Add support for assumed rank
            finalizers.
            (gfc_may_be_finalized): New helper function.
            * dump-parse-tree.cc (write_proc): Whitespace.
            * gfortran.h : Add prototype for gfc_may_be_finalized.
            * resolve.cc (resolve_function): Correct derived types that
            have an incomplete namespace.
            (resolve_where, gfc_resolve_where_code_in_forall,
            gfc_resolve_forall_body, gfc_resolve_code): Check that the op
            code is still EXEC_ASSIGN. If it is set lhs to must finalize.
            (is_finalizable_type): New function.
            (generate_component_assignments): Set must_finalize if needed.
            (gfc_resolve_finalizers): Error if assumed rank finalizer is
            not the only one. Warning on lack of scalar finalizer modified
            to account for assumed rank finalizers.
            (generate_final_call): New function.
            (generate_component_assignments): Enclose the outermost call in
            a block to capture automatic deallocation and final calls.
            Set must_finalize as required to satisfy the standards. Use an
            explicit pointer assignment for pointer components to capture
            finalization of the target. Likewise use explicit assignment
            for allocatable components. Do not use the temporary copy of
            the lhs in defined assignment if the component is allocatable.
            Put the temporary in the same namespace as the lhs symbol if
            the component may be finalized. Remove the leading assignment
            from the expansion of assignment of components that have their
            own defined assignment components. Suppress finalization of
            assignment of temporary components to the lhs. Make an explicit
            final call for the rhs function temporary if it exists.
            (gfc_resolve_code): Set must_finalize for assignments with an
            array constructor on the rhs.
            (gfc_resolve_finalizers): Ensure that an assumed rank finalizer
            is the only finalizer for that type and correct the surprising
            warning for the lack of a scalar finalizer.
            (check_defined_assignments): Handle allocatable components.
            (resolve_fl_derived): Set referenced the vtab for use
            associated symbols.
            (resolve_symbol): Set referenced an unreferenced symbol that
            will be finalized.
            * trans-array.cc (gfc_trans_array_constructor_value): Add code
            to finalize the constructor result. Warn that this feature was
            removed in F2018 and that it is suppressed by -std=2018.
            (trans_array_constructor): Add finalblock, pass to previous
            and apply to loop->post if filled.
            (gfc_add_loop_ss_code): Add se finalblock to outer loop post.
            (gfc_trans_array_cobounds, gfc_trans_array_bounds): Add any
            generated finalization code to the main block.
            (structure_alloc_comps): Add boolean argument to suppress
            finalization and use it for calls from
            gfc_deallocate_alloc_comp_no_caf. Otherwise it defaults to
            false.
            (gfc_copy_alloc_comp_no_fini): New wrapper for
            structure_alloc_comps.
            (gfc_alloc_allocatable_for_assignment): Suppress finalization
            by setting new arg in call to gfc_deallocate_alloc_comp_no_caf.
            (gfc_trans_deferred_array): Use gfc_may_be_finalized and do not
            deallocate the components of entities with a leading '_' in the
            name that are also marked as artificial.
            * trans-array.h : Add the new boolean argument to the prototype
            of gfc_deallocate_alloc_comp_no_caf with a default of false.
            Add prototype for gfc_copy_alloc_comp_no_fini.
            * trans-decl.cc(init_intent_out_dt): Tidy up the code.
            * trans-expr.cc (gfc_init_se): Initialize finalblock.
            (gfc_conv_procedure_call): Use gfc_finalize_tree_expr to
            finalize function results. Replace in-line block for class
            results with call to new function.
            (gfc_conv_expr): Finalize structure constructors for F2003 and
            F2008. Warn that this feature was deleted in F2018 and, unlike
            array constructors, is not default. Add array constructor
            finalblock to the post block.
            (gfc_trans_scalar_assign): Suppress finalization by setting new
            argument in call to gfc_deallocate_alloc_comp_no_caf. Add the
            finalization blocks to the main block.
            (gfc_trans_arrayfunc_assign): Use gfc_assignment_finalizer_call
            and ensure that finalization occurs after the evaluation of the
            rhs but using the initial value for the lhs. Finalize rhs
            function results using gfc_finalize_tree_expr.
            (trans_class_assignment, gfc_trans_assignment_1): As previous
            function, taking care to order evaluation, assignment and
            finalization correctly.
            * trans-io.cc (gfc_trans_transfer): Add the final block.
            * trans-stmt.cc (gfc_trans_call, gfc_trans_allocate): likewise.
            (trans_associate_var): Nullify derived allocatable components
            and finalize function targets with defined assignment
            components on leaving the block scope.
            (trans_allocate): Finalize source expressions, if required,
            and set init_expr artificial temporarily to suppress the
            finalization in gfc_trans_assignment.
            * trans.cc (gfc_add_finalizer_call): Do not finalize the
            temporaries generated in type assignment with defined
            assignment components.
            (gfc_assignment_finalizer_call): New function.
            (gfc_finalize_tree_expr): New function.
            * trans.h: Add finalblock to gfc_se. Add the prototypes for
            gfc_finalize_tree_expr and gfc_assignment_finalizer_call.
    
    gcc/testsuite/
            PR fortran/64290
            * gfortran.dg/finalize_38.f90 : New test.
            * gfortran.dg/finalize_38a.f90 : New test.
            * gfortran.dg/allocate_with_source_25.f90 : The number of final
            calls goes down from 6 to 4.
            * gfortran.dg/associate_25.f90 : Remove the incorrect comment.
            * gfortran.dg/auto_dealloc_2.f90 : Change the tree dump expr
            but the final count remains the same.
            * gfortran.dg/unlimited_polymorphic_8.f90 : Tree dump reveals
            foo.1.x rather than foo.0.x
    
            PR fortran/67444
            * gfortran.dg/finalize_39.f90 : New test.
    
            PR fortran/67471
            * gfortran.dg/finalize_40.f90 : New test.
    
            PR fortran/69298
            PR fortran/70863
            * gfortran.dg/finalize_41.f90 : New test.
    
            PR fortran/71798
            * gfortran.dg/finalize_42.f90 : New test.
    
            PR fortran/80524
            * gfortran.dg/finalize_43.f90 : New test.
    
            PR fortran/82996
            * gfortran.dg/finalize_44.f90 : New test.
    
            PR fortran/84472
            * gfortran.dg/finalize_45.f90 : New test.
    
            PR fortran/88735
            PR fortran/93691
            * gfortran.dg/finalize_46.f90 : New test.
    
            PR fortran/91316
            * gfortran.dg/finalize_47.f90 : New test.
    
            PR fortran/106576
            * gfortran.dg/finalize_48.f90 : New test.
    
            PR fortran/37336
            * gfortran.dg/finalize_49.f90 : New test.
            * gfortran.dg/finalize_50.f90 : New test.
            * gfortran.dg/finalize_51.f90 : New test.

Mainline now behaves in the same way as nagfor and ifort with this testcase. I'll ask to backport to 12-branch in a few weeks.

Cheers

Paul