https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86485
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I've tried:
--- gcc/cp/cp-gimplify.c.jj 2019-02-18 20:48:37.666649912 +0100
+++ gcc/cp/cp-gimplify.c 2019-03-05 19:03:17.076410253 +0100
@@ -709,6 +709,16 @@ cp_gimplify_expr (tree *expr_p, gimple_s
is_gimple_lvalue, fb_lvalue);
*expr_p = TREE_OPERAND (*expr_p, 0);
}
+ else if (TREE_CODE (op1) == TARGET_EXPR
+ && is_really_empty_class (TREE_TYPE (op0))
+ && !TREE_THIS_VOLATILE (op1)
+ && !TREE_ADDRESSABLE (TREE_TYPE (op0)))
+ {
+ gimplify_and_add (op1, pre_p);
+ gimplify_expr (&TREE_OPERAND (*expr_p, 0), pre_p, post_p,
+ is_gimple_lvalue, fb_lvalue);
+ *expr_p = TREE_OPERAND (*expr_p, 0);
+ }
/* P0145 says that the RHS is sequenced before the LHS.
gimplify_modify_expr gimplifies the RHS before the LHS, but that
isn't quite strong enough in two cases:
(poor attempt to gimplify the TARGET_EXPRs on the rhs in that case always and
also the lhs, but not actually copy any data in between) but that unfortunately
causes ICEs on g++.dg/cpp0x/initlist-new1.C test (apparently CLEANUP_POINT_EXPR
then makes it through into the IL).
If the empty class has a destructor, then we actually share the same temporary
through gimplify_modify_expr_rhs
case COND_EXPR:
/* If we're assigning to a non-register type, push the assignment
down into the branches. This is mandatory for ADDRESSABLE types,
since we cannot generate temporaries for such, but it saves a
copy in other cases as well. */
if (!is_gimple_reg_type (TREE_TYPE (*from_p)))
handling, but we don't do that in this case.
Jason, any ideas?
