===================================================================
@@ -5693,7 +5693,7 @@ check_initializer (tree decl, tree init, int flags
&& (!init || TREE_CODE (init) == TREE_LIST))
{
init = build_functional_cast (type, init, tf_none);
- if (init != error_mark_node)
+ if (TREE_CODE (init) == TARGET_EXPR)
TARGET_EXPR_DIRECT_INIT_P (init) = true;
}
init_code = NULL_TREE;
===================================================================
@@ -0,0 +1,19 @@
+// PR c++/55944
+// { dg-options -std=c++11 }
+
+template<class T>
+struct Test
+{
+ constexpr Test(T val) : value(val) {}
+ static void test()
+ {
+ static constexpr Test<int> x(42); // ICE
+ }
+ T value;
+};
+
+int main()
+{
+ static constexpr Test<int> x(42); // OK
+ Test<double>::test();
+}
Hi, today I had a look to this PR, maybe it's simple enough to be fixed at this time. The issue is that we ICE in check_initializer in: init = build_functional_cast (type, init, tf_none); - if (init != error_mark_node) + if (TREE_CODE (init) == TARGET_EXPR) TARGET_EXPR_DIRECT_INIT_P (init) = true; because init is a CAST_EXPR, not a TARGET_EXPR. The reason is simple: we return very early from build_functional_cast, because processing_template_decl is true and we simply build and return a CAST_EXPR. Is it then Ok to tweak the code per the above? It works for the testcase and passes the testsuite. The other possibility I have in mind is doing something entirely different in check_initializer itself when processing_template_decl is true?!? Thanks! Paolo. /////////////////////////