http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43808
--- Comment #7 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-11-08
15:32:03 UTC ---
(In reply to comment #6)
> Indeed, for
> x:
> (mem/s:DI (reg/f:DI 117 [ D.1692 ]) [6 MEM[(struct a[2] *)D.1692_63]+0 S8
> A64])
> mem:
> (mem/s/f/c:DI (plus:DI (reg/f:DI 20 frame)
> (const_int -384 [0xfffffffffffffe80])) [3 D.1548.i.data+0 S8 A64])
> rtx_refs_may_alias_p (x, mem, 0)
> returns 0.
> indirect_ref_may_alias_decl_p is called with base1 the MEM_REF and base2
> VAR_DECL D.1548.
>
> And the problem is that may_be_aliased is false for D.1548. Sure, it doesn't
> have address taken. But as expansion decided to share its slot together with
> other variables and those are may_be_aliased, this of course breaks.
>
> I wonder whether expansion shouldn't somehow make sure that if at least one
> var
> in the partition is may_be_aliased, all of them are. Be it by not merging
> vars
> that have different may_be_aliased, or by say marking the non-addressable vars
> TREE_ADDRESSABLE if anything in the partition is addressable.
>
> Richard?
We do this already, see cfgexpand.c:add_partitioned_vars_to_ptset and
update_alias_info_with_stack_vars and alias.c:ao_ref_from_mem:
/* If this is a reference based on a partitioned decl replace the
base with an INDIRECT_REF of the pointer representative we
created during stack slot partitioning. */
if (TREE_CODE (base) == VAR_DECL
&& ! TREE_STATIC (base)
&& cfun->gimple_df->decls_to_pointers != NULL)
{
void *namep;
namep = pointer_map_contains (cfun->gimple_df->decls_to_pointers, base);
if (namep)
ref->base = build_simple_mem_ref (*(tree *)namep);
}
and we try to handle this particular case in update_alias_info_with_stack_vars:
/* Make all points-to sets that contain one member of a partition
contain all members of the partition. */
if (decls_to_partitions)
{
unsigned i;
struct pointer_set_t *visited = pointer_set_create ();
bitmap temp = BITMAP_ALLOC (NULL);
for (i = 1; i < num_ssa_names; i++)
{
...
So you say that for some reason this doesn't work.
