------- Comment #9 from davek at gcc dot gnu dot org 2009-09-19 10:56 -------
(In reply to comment #8)
> The "ret" string is shared between some attribute and a value from
> CONST_STRING.
Sorry, as I just figured out I was off by one on that:
(gdb) fin
Run till exit from #0 gen_label_for_indirect_string (node=0x7e945910)
at /gnu/gcc/gcc-unpatched/gcc/dwarf2out.c:6849
get_debug_string_label (str=0x7f15e0b8 "ret")
at /gnu/gcc/gcc-unpatched/gcc/dwarf2out.c:6864
6864 return gen_rtx_SYMBOL_REF (Pmode, node->label);
(gdb) print ((struct indirect_string_node *) 0x7e945910)[0]
$5 = {str = 0x7ff2f498 "ret", refcount = 2, form = 0,
label = 0x8284dc0 "*LASF74"}
(gdb)
> But prune_unused_types_walk_attribs resets the count to 0:
> /* Set the string's refcount to 0 so that prune_unused_types_mark
> accounts properly for it. */
> if (AT_class (a) == dw_val_class_str)
> a->dw_attr_val.v.val_str->refcount = 0;
> and while attributes are walked, location lists are not and thus nothing
> notices it is still used. I'll have to read the prune_unused_types stuff
> carefully to understand how this can be fixed, ideally if we could just
> decrement refcount in prune_unused_types_walk_attribs, as long as we are sure
> it is walked for all attributes, it could fix this. Alternatively we'd have
> to
> walk all location lists as well, looking for labels for strings.
But we're looking at the same suspect here anyway :)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41404
