Re: [PATCH] fortran: Remove reference count update [PR108957]

[Paul Richard Thomas <paul.richard.thomas@gmail.com>]

Hi Mikael,

The comment is very welcome! Looks good to me. OK for mainline.

Thanks for the patch.

Paul

On Fri, 15 Sept 2023 at 08:19, Mikael Morin via Fortran
<fortran@gcc.gnu.org> wrote:
>
> Hello,
>
> Harald reminded me recently that there was a working patch attached to the PR.
> I added a documentation comment with the hope that it may help avoid
> making the same mistake in the future.
> Regression tested on x86_64-pc-linux-gnu.
> OK for master?
>
> -- >8 --
>
> Remove one reference count incrementation following the assignment of a
> symbol pointer to a local variable.  Most symbol pointers are "weak" pointer
> and don't need any reference count update when they are assigned, and it is
> especially the case of local variables.
>
> This fixes a memory leak with the testcase from the PR (not included).
>
>         PR fortran/108957
>
> gcc/fortran/ChangeLog:
>
>         * gfortran.h (gfc_symbol): Add comment documenting reference counting.
>         * parse.cc (parse_interface): Remove reference count incrementation.
> ---
>  gcc/fortran/gfortran.h | 20 ++++++++++++++++++++
>  gcc/fortran/parse.cc   |  3 ---
>  2 files changed, 20 insertions(+), 3 deletions(-)
>
> diff --git a/gcc/fortran/gfortran.h b/gcc/fortran/gfortran.h
> index f4a1c106cea..6caf7765ac6 100644
> --- a/gcc/fortran/gfortran.h
> +++ b/gcc/fortran/gfortran.h
> @@ -1944,7 +1944,27 @@ typedef struct gfc_symbol
>       according to the Fortran standard.  */
>    unsigned pass_as_value:1;
>
> +  /* Reference counter, used for memory management.
> +
> +     Some symbols may be present in more than one namespace, for example
> +     function and subroutine symbols are present both in the outer namespace and
> +     the procedure body namespace.  Freeing symbols with the namespaces they are
> +     in would result in double free for those symbols.  This field counts
> +     references and is used to delay the memory release until the last reference
> +     to the symbol is removed.
> +
> +     Not every symbol pointer is accounted for reference counting.  Fields
> +     gfc_symtree::n::sym are, and gfc_finalizer::proc_sym as well.  But most of
> +     them (dummy arguments, generic list elements, etc) are "weak" pointers;
> +     the reference count isn't updated when they are assigned, and they are
> +     ignored when the surrounding structure memory is released.  This is not a
> +     problem because there is always a namespace as surrounding context and
> +     symbols have a name they can be referred with in that context, so the
> +     namespace keeps the symbol from being freed, keeping the pointer valid.
> +     When the namespace ceases to exist, and the symbols with it, the other
> +     structures referencing symbols cease to exist as well.  */
>    int refs;
> +
>    struct gfc_namespace *ns;    /* namespace containing this symbol */
>
>    tree backend_decl;
> diff --git a/gcc/fortran/parse.cc b/gcc/fortran/parse.cc
> index 8f09ddf753c..58386805ffe 100644
> --- a/gcc/fortran/parse.cc
> +++ b/gcc/fortran/parse.cc
> @@ -4064,9 +4064,6 @@ loop:
>    accept_statement (st);
>    prog_unit = gfc_new_block;
>    prog_unit->formal_ns = gfc_current_ns;
> -  if (prog_unit == prog_unit->formal_ns->proc_name
> -      && prog_unit->ns != prog_unit->formal_ns)
> -    prog_unit->refs++;
>
>  decl:
>    /* Read data declaration statements.  */
> --
> 2.40.1
>