Re: [Patch, Fortran, committed] free gfc_code of EXEC_END_PROCEDURE

[Mikael Morin <mikael dot morin at sfr dot fr>]

On 22/08/2012 09:47, Tobias Burnus wrote:
> Dear Thomas, dear all,
> 
> On 08/22/2012 07:28 AM, Thomas Koenig wrote:
>> Am 22.08.2012 00:26, schrieb Tobias Burnus:
>>
>>> I think we should join in an try to remove some leakage - and try to not
>>> introduce new ones.
>>
>> What about using C++ constructors/destructors that allocate/deallocate
>> the necessary memory?  This should simplify the code for memory
>> allocation a lot.
> 
> I don't see how. GCC operates on lists - and the main problem is to keep
> track about them to free them when they aren't needed. Only when the
> scope is limited, it could help to free pointer components; still, one
> needs to be careful whether the same pointer is used elsewhere.
> 
> If you have time, maybe you are able to find out why the following
> program leaks memory:
> 
> !subroutine foo()
>   call bar()
> contains
>   subroutine bar()
>   end subroutine bar
> end
> 
> It doesn't if one comments in the "subroutine  foo()" line. According to
> valgrind, the memory is allocated in parse_contained:
>       gfc_current_ns = gfc_get_namespace (parent_ns, 1);
> However, my impression is that the namespace for ST_END_PROGRAM is
> properly freed in that function ("gfc_free_namespace (ns);") and the
> rest (namespace for MAIN__ and its contained "bar" namespace) is
> properly freed via translate_all_program_units's call to gfc_done_2
> which calls gfc_symbol_done_2 and in turn gfc_free_namespace.
> 
I think it is the bar namespace that is not freed, because the bar
symbol's reference count doesn't drop to zero.
There is this code in gfc_release_symbol which is not entered because
sym->formal_ns == NULL (sym->refs == 2 at this point for `bar'):

  if (sym->formal_ns != NULL && sym->refs == 2)
    {
      /* As formal_ns contains a reference to sym, delete formal_ns just
	 before the deletion of sym.  */
      gfc_namespace *ns = sym->formal_ns;
      sym->formal_ns = NULL;
      gfc_free_namespace (ns);
    }
With a dummy argument for `bar', the code above is entered, but it
doesn't help either as sym->formal_ns->refs == 2, so that the formal
namespace (and the symbol as a result) is not freed anyway.
Maybe we could take this hunk from one of the old pr43829 patches:
http://gcc.gnu.org/bugzilla/attachment.cgi?id=21016&action=diff#a/symbol.c_sec13
I don't know if it would work as the patch this is extracted from
contains many more memory fixes, and it may depend on them. To update to
current trunk I would have to rewrite them from scratch unfortunately.


Actually, I think the uncommented variant leaks more memory. From
running it in a debugger the `foo' namespace's reference count is 2 when
entering gfc_free_namespace, compared to 1 for the `__MAIN' case.
So the __MAIN namespace is freed, but foo isn't. I have the feeling that
the namespace remains reachable somehow, so that valgrind doesn't regard
the attached pointers as lost memory.

Unfortunately, valgrind tells me nothing more than `unhandled
instruction', so it's difficult to test patches.

Mikael