On Thu, Mar 15, 2018 at 4:32 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> On Wed, Mar 14, 2018 at 08:55:47PM -0400, Jason Merrill wrote:
>> > @@ -3192,16 +3198,70 @@ replace_placeholders (tree exp, tree obj
>> > return exp;
>> >
>> > tree *tp = &exp;
>> > - hash_set<tree> pset;
>> > - replace_placeholders_t data = { obj, false, &pset };
>> > + /* Use exp instead of *(type *)&exp. */
>> > + while (TREE_CODE (exp) == INDIRECT_REF)
>> > + {
>> > + tree t = TREE_OPERAND (exp, 0);
>> > + STRIP_NOPS (t);
>> > + if (TREE_CODE (t) == ADDR_EXPR
>> > + && (t = TREE_OPERAND (t, 0))
>> > + && same_type_ignoring_top_level_qualifiers_p (TREE_TYPE (exp),
>> > + TREE_TYPE (t)))
>> > + exp = t;
>> > + else
>> > + break;
>> > + }
>>
>> What's the motivation for this?
>
> g++.dg/cpp0x/nsdmi13.C ICEs without that, we have there:
> a = A({});
> and build_over_call does:
> 8163 else if (tree_int_cst_equal (TYPE_SIZE (type), TYPE_SIZE
> (as_base)))
> 8164 {
> 8165 arg = cp_build_fold_indirect_ref (arg);
> 8166 val = build2 (MODIFY_EXPR, TREE_TYPE (to), to, arg);
> 8167 /* Handle NSDMI that refer to the object being initialized.
> */
> 8168 replace_placeholders (arg, to);
> 8169 }
> where arg is after the cp_build_fold_indirect_ref
> *(struct A &) &TARGET_EXPR <D.2403, {.p=&<PLACEHOLDER_EXPR struct A>}>
> This is in MODIFY_EXPR rather than INIT_EXPR and the gimplifier through
> gimple_fold_indirect_ref_rhs folds the *(struct A &) & away and so there is
> no further temporary and thus cp_gimplify_init_expr isn't called to
> replace_placeholders, so if we don't replace it here (with to being the a
> VAR_DECL), we don't replace it ever.
Ah. That's a problem: the language says there's a temporary, so after
the assignment a.p should not point to a.
So the existing call to replace_placeholders in build_over_call is wrong.
Seems like if the initializer for a TARGET_EXPR involves
PLACEHOLDER_EXPR, the gimple_fold_modify_expr_rhs treatment of
TARGET_EXPR isn't safe for MODIFY_EXPR, for the same reason that
cp_gimplify_expr treats INIT_EXPR and MODIFY_EXPR differently.
Jason
