[Bug fortran/47839] ICE in dwarf2out.c:add_AT_specification

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47839

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011.02.21 17:04:46
     Ever Confirmed|0                           |1

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-02-21 17:04:46 UTC ---
We merge the decls during symtab merging.  The prevailing one looks for example
like

(gdb) call debug_tree (prevailing->decl)
 <var_decl 0x7ffff5b3b0a0 zstart
    type <integer_type 0x7ffff7ee3498 int public SI
        size <integer_cst 0x7ffff7ed36e0 constant 32>
        unit size <integer_cst 0x7ffff7ed33e8 constant 4>
        align 32 symtab 0 alias set -1 canonical type 0x7ffff7ee3498 precision
32 min <integer_cst 0x7ffff7ed3668 -2147483648> max <integer_cst 0x7ffff7ed3690
2147483647>
        pointer_to_this <pointer_type 0x7ffff7ef03f0>>
    used public static SI file mod.f90 line 2 col 0 size <integer_cst
0x7ffff7ed36e0 32> unit size <integer_cst 0x7ffff7ed33e8 4>
    align 32 context <namespace_decl 0x7ffff5b31678 globalvar_mod>>

while the non-prevailing one is

 <var_decl 0x7ffff5b3b640 zstart
    type <integer_type 0x7ffff7ee3498 int public SI
        size <integer_cst 0x7ffff7ed36e0 constant 32>
        unit size <integer_cst 0x7ffff7ed33e8 constant 4>
        align 32 symtab 0 alias set -1 canonical type 0x7ffff7ee3498 precision
32 min <integer_cst 0x7ffff7ed3668 -2147483648> max <integer_cst 0x7ffff7ed3690
2147483647>
        pointer_to_this <pointer_type 0x7ffff7ef03f0>>
    used public ignored external SI file t.f90 line 4 col 0 size <integer_cst
0x7ffff7ed36e0 32> unit size <integer_cst 0x7ffff7ed33e8 4>
    align 32>

note that it is public and external and has a NULL DECL_CONTEXT.

The C frontend for a local extern declaration has

 <var_decl 0x7ffff5b481e0 i
    type <integer_type 0x7ffff7ee6498 int public SI
        size <integer_cst 0x7ffff7ed36e0 constant 32>
        unit size <integer_cst 0x7ffff7ed33e8 constant 4>
        align 32 symtab 0 alias set -1 canonical type 0x7ffff7ee6498 precision
32 min <integer_cst 0x7ffff7ed3668 -2147483648> max <integer_cst 0x7ffff7ed3690
2147483647>
        pointer_to_this <pointer_type 0x7ffff7ef7540>>
    used public external common SI defer-output file t.c line 8 col 15 size
<integer_cst 0x7ffff7ed36e0 32> unit size <integer_cst 0x7ffff7ed33e8 4>
    align 32 context <function_decl 0x7ffff5b49000 foo>>

thus puts it into function context.  This decl is solely used for the BLOCK
tree, in the function a public external global var is used (which is
then merged with the static one from the other TU).

void foo (void)
{
  extern int i;
  i = 0;
}

---

int i;



Simplified Fortran testcase:

MODULE globalvar_mod
integer        :: xstop
CONTAINS
END MODULE globalvar_mod

---

MODULE PEC_mod
CONTAINS
SUBROUTINE PECapply(Ex)
USE globalvar_mod, ONLY : xstop
real(kind=8), dimension(1:xstop), intent(inout) :: Ex
END SUBROUTINE PECapply
END MODULE PEC_mod

it's important that PECapply is inside a module.

The decl is built by gfc_get_symbol_decl and put into the function via
gfc_add_decl_to_function - which is I think in general bogus for
imported decls.  Its context is later cleared in pushdecl, but the
variable isn't removed from BLOCK_VARS.

I think we want to avoid gfc_add_decl_to_function in the first place.