Re: C++ PATCH to fix an ICE on invalid with OVERLOADs (PR c++/84854)

[Marek Polacek <polacek at redhat dot com>]

On Wed, Mar 14, 2018 at 02:06:36PM -0400, Jason Merrill wrote:
> On Wed, Mar 14, 2018 at 11:59 AM, Marek Polacek <polacek@redhat.com> wrote:
> > cxx_constant_value doesn't understand template codes, and neither it
> > understands OVERLOADs, so if we pass an OVERLOAD to it, we crash.  Here
> > instantiate_non_dependent_expr got an OVERLOAD, but since it calls
> > is_nondependent_constant_expression which checks type_unknown_p, it left the
> > expression as it was.  We can't use is_nondependent_constant_expression in
> > finish_if_stmt_cond because i_n_c_e checks is_constant_expression and that is
> > not suitable here; we'd miss diagnostics.  So I did the following; I think we
> > should reject the testcase with an error.
> >
> > Bootstrapped/regtested on x86_64-linux, ok for trunk?
> >
> > 2018-03-14  Marek Polacek  <polacek@redhat.com>
> >
> >         PR c++/84854
> >         * semantics.c (finish_if_stmt_cond): Give error if the condition
> >         is an overloaded function with no contextual type information.
> >
> >         * g++.dg/cpp1z/constexpr-if15.C: New test.
> >
> > diff --git gcc/cp/semantics.c gcc/cp/semantics.c
> > index fdf37bea770..a056e9445e9 100644
> > --- gcc/cp/semantics.c
> > +++ gcc/cp/semantics.c
> > @@ -735,8 +735,16 @@ finish_if_stmt_cond (tree cond, tree if_stmt)
> >        && require_constant_expression (cond)
> >        && !value_dependent_expression_p (cond))
> >      {
> > -      cond = instantiate_non_dependent_expr (cond);
> > -      cond = cxx_constant_value (cond, NULL_TREE);
> > +      if (type_unknown_p (cond))
> > +       {
> > +         cxx_incomplete_type_error (cond, TREE_TYPE (cond));
> > +         cond = error_mark_node;
> 
> I think I'd prefer to skip this block when type_unknown_p, and leave
> error handling up to the code shared with regular if.

Like this?  It was my first version, but I thought we wanted the error;
the following rejects the testcase only when instantiating the template
function.  Which is fine by me, too.

Bootstrapped/regtested on x86_64-linux, ok for trunk?

2018-03-15  Marek Polacek  <polacek@redhat.com>

	PR c++/84854
	* semantics.c (finish_if_stmt_cond): Check type_unknown_p.

	* g++.dg/cpp1z/constexpr-if15.C: New test.

diff --git gcc/cp/semantics.c gcc/cp/semantics.c
index fdf37bea770..753a5f4d4f8 100644
--- gcc/cp/semantics.c
+++ gcc/cp/semantics.c
@@ -733,7 +733,8 @@ finish_if_stmt_cond (tree cond, tree if_stmt)
   if (IF_STMT_CONSTEXPR_P (if_stmt)
       && !type_dependent_expression_p (cond)
       && require_constant_expression (cond)
-      && !value_dependent_expression_p (cond))
+      && !value_dependent_expression_p (cond)
+      && !type_unknown_p (cond))
     {
       cond = instantiate_non_dependent_expr (cond);
       cond = cxx_constant_value (cond, NULL_TREE);
diff --git gcc/testsuite/g++.dg/cpp1z/constexpr-if15.C gcc/testsuite/g++.dg/cpp1z/constexpr-if15.C
index e69de29bb2d..c819b3e3a07 100644
--- gcc/testsuite/g++.dg/cpp1z/constexpr-if15.C
+++ gcc/testsuite/g++.dg/cpp1z/constexpr-if15.C
@@ -0,0 +1,11 @@
+// PR c++/84854
+// { dg-options -std=c++17 }
+
+constexpr int foo () { return 1; }
+constexpr int foo (int) { return 2; }
+
+template <typename>
+void a()
+{
+  if constexpr(foo) { };
+}

	Marek