Hi Salvatore,
My reading of F2018: 7.5.6.3 "When finalization occurs" is that three calls
is correct. (i) Paragraph 1 stipulates that the 'var' expression be
evaluated before the reallocation and intrinsic assignment; (ii)stipulates
that the function result be finalised "the result is finalized after
execution of the innermost
executable construct containing the reference." ; and (iii) Finalisation
occurs on going out of scope.
That is what is implemented in the attached. I am working my way through
the testcase dependencies of PR37336 making sure that finalizat
ion occurs as required by 7.5.6.3 and in the order defined in the previous
section. It will all be done in the next few days.
What will remain is finalization of function results within array
constructors and one or two other corner cases. Following that, the
existing finalization calls will be brought into the framework as the new
calls.
Best regards
Paul
On Thu, 27 Jan 2022 at 07:17, Salvatore Filippone <
filippone.salvat...@gmail.com> wrote:
> One more data point: Cray FTN issues TWO calls to the FINAL.
> Which begs the question: what is the correct number of calls one, two or
> three?
> Salvatore
>
> fsalvato@daint102:/project/prce01/fsalvato/NUMERICAL/PSBLAS/V4/psblas4/test/newstuff>
> ftn --version
> Cray Fortran : Version 11.0.0
> fsalvato@daint102:/project/prce01/fsalvato/NUMERICAL/PSBLAS/V4/psblas4/test/newstuff>
> ftn -o testfinal testfinal.f90
> fsalvato@daint102:/project/prce01/fsalvato/NUMERICAL/PSBLAS/V4/psblas4/test/newstuff>
> ./testfinal
> Allocating wrapper
> Calling new_outer_type
> Assigning outer%test_item
> Called delete_test_type
> End of new_outer_type
> DeAllocating wrapper
> Called delete_test_type
>
> On Wed, Jan 26, 2022 at 11:59 PM Paul Richard Thomas <
> paul.richard.tho...@gmail.com> wrote:
>
>> Hi Jerry,
>>
>> I am trying to fix the failure of my latest patch with this very test
>> case. Otherwise it fixes most of the remaining dependencies in PR37336.
>>
>> At a pinch, I could submit the earlier patch that Andrew mentions and
>> work from there. However, you will note that it does miss one of the
>> finalizations. This is critical because function results, which should be
>> finalized, are not.
>>
>> I'll keep an eye on the state of the branch. By and large, release occurs
>> 3-4 months after the start of stage 4. I will leave 2 months maximum.
>>
>> Best regards
>>
>> Paul
>>
>>
>> On Wed, 26 Jan 2022 at 21:29, Jerry D via Fortran
>> wrote:
>>
>>> Is there any reason these patches can not be applied and use this test
>>> as a test case?
>>>
>>> Regards,
>>>
>>> Jerry
>>>
>>> On 1/24/22 8:11 AM, Salvatore Filippone via Fortran wrote:
>>> > Thanks a lot
>>> > (yes, I suspected both gfortran and intel were wrong, precisely
>>> because I
>>> > could see why you'd need two FINAL calls, but not three).
>>> >
>>> > Salvatore
>>> >
>>> > On Mon, Jan 24, 2022 at 4:45 PM Andrew Benson <
>>> aben...@carnegiescience.edu>
>>> > wrote:
>>> >
>>> >> Hi Salvatore,
>>> >>
>>> >> This looks like it's related to some of the missing finalization
>>> >> functionality
>>> >> (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=37336). Paul has some
>>> >> patches
>>> >> (e.g. https://gcc.gnu.org/pipermail/fortran/2022-January/057415.html)
>>> >> which
>>> >> implement most of the missing functionality. With those patches
>>> >> incorporated
>>> >> your code gives the following output with gfortran:
>>> >>
>>> >> $ ./testfinal
>>> >> Allocating wrapper
>>> >> Calling new_outer_type
>>> >> Assigning outer%test_item
>>> >> Called delete_test_type
>>> >> End of new_outer_type
>>> >> DeAllocating wrapper
>>> >> Called delete_test_type
>>> >>
>>> >> So there is one more call to the finalizer than you found - I haven't
>>> >> checked
>>> >> carefully but I would guess this is a deallocation of LHS on
>>> assignment.
>>> >>
>>> >> In testing these patches using the Intel compiler we found that it
>>> seems
>>> >> to
>>> >> call the finalization wrapper more than it should, sometimes on
>>> objects
>>> >> that
>>> >> have already been deallocated. Your code, compiled with the Intel
>>> compiler
>>> >> and
>>> >> then run under valgrind shows the following:
>>> >>
>>> >> $ valgrind ./testfinal
>>> >> ==7340== Memcheck, a memory error detector
>>> >> ==7340== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et
>>> al.
>>> >> ==7340== Using Valgrind-3.13.0 and LibVEX; rerun with -h for
>>> copyright info
>>> >> ==7340== Command: ./testfinal
>>> >> ==7340==
>>> >> ==7340== Conditional jump or move depends on uninitialised value(s)
>>> >> ==7340==at 0x493A51: __intel_sse2_strcpy (in
>>> /home/abensonca/Scratch/
>>> >> ifortTests/testfinal)
>>> >> ==7340==by 0x45D70E: for__add_to_lf_table (in
>>> /home/abensonca/Scratch/
>>> >> ifortTests/testfinal)
>>> >> ==7340==by 0x4410CB: for__open_proc (in /home/abensonca/Scratch/
>>> >> ifortTests/testfinal)
>>> >> ==7340==by 0x423A64: for__open_default (in
>>> /h
