[Patch, Fortran] PR 41714: [OOP] ALLOCATE SOURCE= does not properly copy the value from SOURCE

[Richard Guenther]

On Sun, Oct 25, 2009 at 6:56 PM, Janus Weil <janus@gcc.gnu.org> wrote:
>>>> It looks OK to me except:
>>>>
>>>>>        PR fortran/41714
>>>>>        * trans-expr.c (gfc_build_memcpy_call): Take care of the case that the
>>>>>        call to '__builtin_memcpy' is optimized away (replaced by a direct
>>>>>        assignment).
>>>>
>>>> How the heck does that work?  It comes out as a NOP_EXPR and yet it's
>>>> really an assignment..... Is that documented somewhere?
>>>
>>> That patch looks indeed dubious.  It tests for an implementation detail
>>> (the memcpy folder returns (void *) ({ *dst = *src; dst; })).  You should
>>> be able to unconditionally fold-convert to void_type_node as in the
>>> original code.  Instead tree_annotate_all_with_location should be fixed.
>>
>> Or rather the FE should not call this function - it assumes that the code
>> is already gimplified.
>
> Ok, so you mean one should instead just do the stuff which this
> function does, but without the extra checks? Like here:
>
> Index: gcc/fortran/trans.c
> ===================================================================
> --- gcc/fortran/trans.c (Revision 153538)
> +++ gcc/fortran/trans.c (Arbeitskopie)
> @@ -1282,7 +1282,11 @@ gfc_trans_code (gfc_code * code)
>       if (res != NULL_TREE && ! IS_EMPTY_STMT (res))
>        {
>          if (TREE_CODE (res) == STATEMENT_LIST)
> -           tree_annotate_all_with_location (&res, input_location);
> +           {
> +             tree_stmt_iterator i;
> +             for (i = tsi_start (res); !tsi_end_p (i); tsi_next (&i))
> +               SET_EXPR_LOCATION (tsi_stmt (i), input_location);
> +           }
>          else
>            SET_EXPR_LOCATION (res, input_location);

No.  I think the above should just be dropped (as well as the other
call in the Fortran frontend).  The location should have been set
by the various stmt builders (like build_call_expr_loc in the
memcpy case).  For the folding of memcpy case
the folder will have distributed the locations appropriately.

The middle-end function can then be removed completely (the Fortran
FE is the only caller).  A patch to do so is pre-approved.

Thanks,
Richard.