https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89512
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Slightly cleaned up testcase:
struct A {
template <typename T>
static const int a = 0;
};
struct B {
template <typename T>
static int foo ()
{
return T::a;
}
};
int bar ()
{
return B::foo<A> ();
}
Outside of a template this is caught by finish_id_expression_1 doing:
/* If we didn't find anything, or what we found was a type,
then this wasn't really an id-expression. */
if (TREE_CODE (decl) == TEMPLATE_DECL
&& !DECL_FUNCTION_TEMPLATE_P (decl))
{
*error_msg = G_("missing template arguments");
return error_mark_node;
}
but in this case, we don't call finish_id_expression, but go through
tsubst_qualified_id -> finish_qualified_id_expr
So, I wonder if we shouldn't handle this in the latter, something like:
--- gcc/cp/semantics.c.jj 2019-03-08 11:45:27.556465237 +0100
+++ gcc/cp/semantics.c 2019-03-13 17:24:46.260345284 +0100
@@ -2112,6 +2112,14 @@ finish_qualified_id_expr (tree qualifyin
expr = build_offset_ref (qualifying_class, expr, /*address_p=*/false,
complain);
}
+ else if (!template_p
+ && TREE_CODE (expr) == TEMPLATE_DECL
+ && !DECL_FUNCTION_TEMPLATE_P (expr))
+ {
+ if (complain & tf_error)
+ error ("%qE missing template arguments", DECL_TEMPLATE_RESULT (expr));
+ return error_mark_node;
+ }
else
{
/* In a template, return a SCOPE_REF for most qualified-ids
