[PATCH] c++: wrong std::is_convertible with cv-qual fn [PR109680]

[Marek Polacek <polacek@redhat.com>]

This PR points out that std::is_convertible has given the wrong answer
in

  static_assert (!std::is_convertible_v <int () const, int (*) ()>, "");

since r13-2822 implemented __is_{,nothrow_}convertible.

std::is_convertible uses the imaginary

  To test() { return std::declval<From>(); }

to do its job.  Here, From is 'int () const'.  std::declval is defined as:

  template<class T>
  typename std::add_rvalue_reference<T>::type declval() noexcept;

std::add_rvalue_reference is defined as "If T is a function type that
has no cv- or ref- qualifier or an object type, provides a member typedef
type which is T&&, otherwise type is T."

In our case, T is cv-qualified, so the result is T, so we end up with

  int () const declval() noexcept;

which is invalid.  In other words, this is pretty much like:

  using T = int () const;
  T fn1(); // bad, fn returning a fn
  T& fn2(); // bad, cannot declare reference to qualified function type
  T* fn3(); // bad, cannot declare pointer to qualified function type

  using U = int ();
  U fn4(); // bad, fn returning a fn
  U& fn5(); // OK
  U* fn6(); // OK

I think is_convertible_helper needs to simulate std::declval better.
I wouldn't be surprised if other type traits needed a similar fix.

Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk/13?

I've tested the new test with G++12 and clang++ as well (with
std::is_convertible).

	PR c++/109680

gcc/cp/ChangeLog:

	* method.cc (is_convertible_helper): Correct simulating std::declval.

gcc/testsuite/ChangeLog:

	* g++.dg/ext/is_convertible6.C: New test.
---
 gcc/cp/method.cc                           | 20 ++++++++++++++++++++
 gcc/testsuite/g++.dg/ext/is_convertible6.C | 16 ++++++++++++++++
 2 files changed, 36 insertions(+)
 create mode 100644 gcc/testsuite/g++.dg/ext/is_convertible6.C

diff --git a/gcc/cp/method.cc b/gcc/cp/method.cc
index 00eae56eb5b..38eb7520312 100644
--- a/gcc/cp/method.cc
+++ b/gcc/cp/method.cc
@@ -2245,6 +2245,26 @@ is_convertible_helper (tree from, tree to)
 {
   if (VOID_TYPE_P (from) && VOID_TYPE_P (to))
     return integer_one_node;
+  /* std::is_{,nothrow_}convertible test whether the imaginary function
+     definition
+
+       To test() { return std::declval<From>(); }
+
+     is well-formed.  A function can't return a function...  */
+  if (FUNC_OR_METHOD_TYPE_P (to)
+      /* ...neither can From be a function with cv-/ref-qualifiers:
+	 std::declval is defined as
+
+	  template<class T>
+	  typename std::add_rvalue_reference<T>::type declval() noexcept;
+
+	and std::add_rvalue_reference yields T when T is a function with
+	cv- or ref-qualifiers, making the definition ill-formed.
+	??? Should we check this in other uses of build_stub_object too?  */
+      || (FUNC_OR_METHOD_TYPE_P (from)
+	  && (type_memfn_quals (from) != TYPE_UNQUALIFIED
+	      || type_memfn_rqual (from) != REF_QUAL_NONE)))
+    return error_mark_node;
   cp_unevaluated u;
   tree expr = build_stub_object (from);
   deferring_access_check_sentinel acs (dk_no_deferred);
diff --git a/gcc/testsuite/g++.dg/ext/is_convertible6.C b/gcc/testsuite/g++.dg/ext/is_convertible6.C
new file mode 100644
index 00000000000..180582663e8
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ext/is_convertible6.C
@@ -0,0 +1,16 @@
+// PR c++/109680
+// { dg-do compile { target c++11 } }
+
+#define SA(X) static_assert((X),#X)
+
+SA(!__is_convertible(int () const, int (*)()));
+SA(!__is_convertible(int (*)(), int () const));
+
+SA( __is_convertible(int (), int (*)()));
+SA(!__is_convertible(int (*)(), int ()));
+
+SA( __is_convertible(int (int), int (*) (int)));
+SA(!__is_convertible(int (*) (int), int (int)));
+
+SA(!__is_convertible(int (int) const, int (*) (int)));
+SA(!__is_convertible(int (*) (int), int (int) const));

base-commit: 33020780a9699f1146eeed61783cec89fde337a0
-- 
2.40.1