Improve our handling of user-defined conversions when computing

implicit conversion sequences. In particular, model the "standard
conversion" from a class to its own type (or a base type) directly as
a standard conversion in the normal path *without* trying to determine
if there is a valid copy constructor. This appears to match the intent
of C++ [over.best.ics]p6 and more closely matches GCC and EDG.

As part of this, model non-lvalue reference initialization via
user-defined conversion in overloading the same way we handle it in
InitializationSequence, separating the "general user-defined
conversion" and "conversion to compatible class type" cases.

The churn in the overload-call-copycon.cpp test case is because the
test case was originally wrong; it assumed that we should do more
checking for copy constructors that we actually should, which affected
overload resolution.

Fixes PR7055. Bootstrapped okay.

llvm-svn: 110773

1 files changed, 34 insertions, 0 deletions

diff --git a/clang/test/SemaCXX/conversion-function.cpp b/clang/test/SemaCXX/conversion-function.cpp
index d7d41a40311..73b78f414a5 100644
--- a/clang/test/SemaCXX/conversion-function.cpp
+++ b/clang/test/SemaCXX/conversion-function.cpp
@@ -214,3 +214,37 @@ struct Other {
 void test_any() {
   Any any = Other(); // expected-error{{cannot pass object of non-POD type 'Other' through variadic constructor; call will abort at runtime}}
 }
+
+namespace PR7055 {
+  // Make sure that we don't allow too many conversions in an
+  // auto_ptr-like template. In particular, we can't create multiple
+  // temporary objects when binding to a reference.
+  struct auto_ptr {
+    struct auto_ptr_ref { };
+
+    auto_ptr(auto_ptr&);
+    auto_ptr(auto_ptr_ref);
+    explicit auto_ptr(int *);
+
+    operator auto_ptr_ref();
+  };
+
+  struct X {
+    X(auto_ptr);
+  };
+
+  X f() {
+    X x(auto_ptr(new int));
+    return X(auto_ptr(new int));
+  }
+
+  auto_ptr foo();
+
+  X e(foo());
+
+  struct Y {
+    Y(X);
+  };
+  
+  Y f2(foo());
+}