Move support for return values of called Python functions

Currently pybind11 always translates values returned by Python functions
invoked from C++ code by copying, even when moving is feasible--and,
more importantly, even when moving is required.

The first, and relatively minor, concern is that moving may be
considerably more efficient for some types.  The second problem,
however, is more serious: there's currently no way python code can
return a non-copyable type to C++ code.

I ran into this while trying to add a PYBIND11_OVERLOAD of a virtual
method that returns just such a type: it simply fails to compile because
this:

    overload = ...
    overload(args).template cast<ret_type>();

involves a copy: overload(args) returns an object instance, and the
invoked object::cast() loads the returned value, then returns a copy of
the loaded value.

We can, however, safely move that returned value *if* the object has the
only reference to it (i.e. if ref_count() == 1) and the object is
itself temporary (i.e. if it's an rvalue).

This commit does that by adding an rvalue-qualified object::cast()
method that allows the returned value to be move-constructed out of the
stored instance when feasible.

This basically comes down to three cases:

- For objects that are movable but not copyable, we always try the move,
  with a runtime exception raised if this would involve moving a value
  with multiple references.
- When the type is both movable and non-trivially copyable, the move
  happens only if the invoked object has a ref_count of 1, otherwise the
  object is copied.  (Trivially copyable types are excluded from this
  case because they are typically just collections of primitive types,
  which can be copied just as easily as they can be moved.)
- Non-movable and trivially copy constructible objects are simply
  copied.

This also adds examples to example-virtual-functions that shows both a
non-copyable object and a movable/copyable object in action: the former
raises an exception if returned while holding a reference, the latter
invokes a move constructor if unreferenced, or a copy constructor if
referenced.

Basically this allows code such as:

    class MyClass(Pybind11Class):
        def somemethod(self, whatever):
            mt = MovableType(whatever)
            # ...
            return mt

which allows the MovableType instance to be returned to the C++ code
via its move constructor.

Of course if you attempt to violate this by doing something like:

    self.value = MovableType(whatever)
    return self.value

you get an exception--but right now, the pybind11-side of that code
won't compile at all.

example/example-virtual-functions.cpp

@@ -69,6 +69,53 @@ public:
     }
 };
 
+class NonCopyable {
+public:
+    NonCopyable(int a, int b) : value{new int(a*b)} {}
+    NonCopyable(NonCopyable &&) = default;
+    NonCopyable(const NonCopyable &) = delete;
+    NonCopyable() = delete;
+    void operator=(const NonCopyable &) = delete;
+    void operator=(NonCopyable &&) = delete;
+    std::string get_value() const {
+        if (value) return std::to_string(*value); else return "(null)";
+    }
+    ~NonCopyable() { std::cout << "NonCopyable destructor @ " << this << "; value = " << get_value() << std::endl; }
+
+private:
+    std::unique_ptr<int> value;
+};
+
+// This is like the above, but is both copy and movable.  In effect this means it should get moved
+// when it is not referenced elsewhere, but copied if it is still referenced.
+class Movable {
+public:
+    Movable(int a, int b) : value{a+b} {}
+    Movable(const Movable &m) { value = m.value; std::cout << "Movable @ " << this << " copy constructor" << std::endl; }
+    Movable(Movable &&m) { value = std::move(m.value); std::cout << "Movable @ " << this << " move constructor" << std::endl; }
+    int get_value() const { return value; }
+    ~Movable() { std::cout << "Movable destructor @ " << this << "; value = " << get_value() << std::endl; }
+private:
+    int value;
+};
+
+class NCVirt {
+public:
+    virtual NonCopyable get_noncopyable(int a, int b) { return NonCopyable(a, b); }
+    virtual Movable get_movable(int a, int b) = 0;
+
+    void print_nc(int a, int b) { std::cout << get_noncopyable(a, b).get_value() << std::endl; }
+    void print_movable(int a, int b) { std::cout << get_movable(a, b).get_value() << std::endl; }
+};
+class NCVirtTrampoline : public NCVirt {
+    virtual NonCopyable get_noncopyable(int a, int b) {
+        PYBIND11_OVERLOAD(NonCopyable, NCVirt, get_noncopyable, a, b);
+    }
+    virtual Movable get_movable(int a, int b) {
+        PYBIND11_OVERLOAD_PURE(Movable, NCVirt, get_movable, a, b);
+    }
+};
+
 int runExampleVirt(ExampleVirt *ex, int value) {
     return ex->run(value);
 }
@@ -240,6 +287,20 @@ void init_ex_virtual_functions(py::module &m) {
         .def("run_bool", &ExampleVirt::run_bool)
         .def("pure_virtual", &ExampleVirt::pure_virtual);
 
+    py::class_<NonCopyable>(m, "NonCopyable")
+        .def(py::init<int, int>())
+        ;
+    py::class_<Movable>(m, "Movable")
+        .def(py::init<int, int>())
+        ;
+    py::class_<NCVirt, std::unique_ptr<NCVirt>, NCVirtTrampoline>(m, "NCVirt")
+        .def(py::init<>())
+        .def("get_noncopyable", &NCVirt::get_noncopyable)
+        .def("get_movable", &NCVirt::get_movable)
+        .def("print_nc", &NCVirt::print_nc)
+        .def("print_movable", &NCVirt::print_movable)
+        ;
+
     m.def("runExampleVirt", &runExampleVirt);
     m.def("runExampleVirtBool", &runExampleVirtBool);
     m.def("runExampleVirtVirtual", &runExampleVirtVirtual);

example/example-virtual-functions.py

@@ -5,6 +5,8 @@ sys.path.append('.')
 
 from example import ExampleVirt, runExampleVirt, runExampleVirtVirtual, runExampleVirtBool
 from example import A_Repeat, B_Repeat, C_Repeat, D_Repeat, A_Tpl, B_Tpl, C_Tpl, D_Tpl
+from example import NCVirt, NonCopyable, Movable
+
 
 class ExtendedExampleVirt(ExampleVirt):
     def __init__(self, state):
@@ -87,3 +89,36 @@ for cl in classes:
     if hasattr(obj, "lucky_number"):
         print("Lucky = %.2f" % obj.lucky_number())
 
+class NCVirtExt(NCVirt):
+    def get_noncopyable(self, a, b):
+        # Constructs and returns a new instance:
+        nc = NonCopyable(a*a, b*b)
+        return nc
+    def get_movable(self, a, b):
+        # Return a referenced copy
+        self.movable = Movable(a, b)
+        return self.movable
+
+class NCVirtExt2(NCVirt):
+    def get_noncopyable(self, a, b):
+        # Keep a reference: this is going to throw an exception
+        self.nc = NonCopyable(a, b)
+        return self.nc
+    def get_movable(self, a, b):
+        # Return a new instance without storing it
+        return Movable(a, b)
+
+ncv1 = NCVirtExt()
+print("2^2 * 3^2 =")
+ncv1.print_nc(2, 3)
+print("4 + 5 =")
+ncv1.print_movable(4, 5)
+ncv2 = NCVirtExt2()
+print("7 + 7 =")
+ncv2.print_movable(7, 7)
+try:
+    ncv2.print_nc(9, 9)
+    print("Something's wrong: exception not raised!")
+except RuntimeError as e:
+    # Don't print the exception message here because it differs under debug/non-debug mode
+    print("Caught expected exception")

example/example-virtual-functions.ref

@@ -77,5 +77,21 @@ VI_DT:
 VI_DT says: quack quack quack
 Unlucky = 1234
 Lucky = -4.25
+2^2 * 3^2 =
+NonCopyable destructor @ 0x1a6c3f0; value = (null)
+36
+NonCopyable destructor @ 0x7ffc6d1fbaa8; value = 36
+4 + 5 =
+Movable @ 0x7ffc6d1fbacc copy constructor
+9
+Movable destructor @ 0x7ffc6d1fbacc; value = 9
+7 + 7 =
+Movable @ 0x7ffc6d1fbacc move constructor
+Movable destructor @ 0x1a6c4d0; value = 14
+14
+Movable destructor @ 0x7ffc6d1fbacc; value = 14
+Caught expected exception
+NonCopyable destructor @ 0x29a64b0; value = 81
+Movable destructor @ 0x1a6c410; value = 9
 Destructing ExampleVirt..
 Destructing ExampleVirt..

include/pybind11/cast.h

@@ -809,9 +809,36 @@ public:
     PYBIND11_TYPE_CASTER(type, handle_type_name<type>::name());
 };
 
+// Our conditions for enabling moving are quite restrictive:
+// At compile time:
+// - T needs to be a non-const, non-pointer, non-reference type
+// - type_caster<T>::operator T&() must exist
+// - the type must be move constructible (obviously)
+// At run-time:
+// - if the type is non-copy-constructible, the object must be the sole owner of the type (i.e. it
+//   must have ref_count() == 1)h
+// If any of the above are not satisfied, we fall back to copying.
+template <typename T, typename SFINAE = void> struct move_is_plain_type : std::false_type {};
+template <typename T> struct move_is_plain_type<T, typename std::enable_if<
+        !std::is_void<T>::value && !std::is_pointer<T>::value && !std::is_reference<T>::value && !std::is_const<T>::value
+    >::type> : std::true_type {};
+template <typename T, typename SFINAE = void> struct move_always : std::false_type {};
+template <typename T> struct move_always<T, typename std::enable_if<
+        move_is_plain_type<T>::value &&
+        !std::is_copy_constructible<T>::value && std::is_move_constructible<T>::value &&
+        std::is_same<decltype(std::declval<type_caster<T>>().operator T&()), T&>::value
+    >::type> : std::true_type {};
+template <typename T, typename SFINAE = void> struct move_if_unreferenced : std::false_type {};
+template <typename T> struct move_if_unreferenced<T, typename std::enable_if<
+        move_is_plain_type<T>::value &&
+        !move_always<T>::value && std::is_move_constructible<T>::value &&
+        std::is_same<decltype(std::declval<type_caster<T>>().operator T&()), T&>::value
+    >::type> : std::true_type {};
+template <typename T> using move_never = std::integral_constant<bool, !move_always<T>::value && !move_if_unreferenced<T>::value>;
+
 NAMESPACE_END(detail)
 
-template <typename T> T cast(handle handle) {
+template <typename T> T cast(const handle &handle) {
     typedef detail::type_caster<typename detail::intrinsic_type<T>::type> type_caster;
     type_caster conv;
     if (!conv.load(handle, true)) {
@@ -838,6 +865,57 @@ template <typename T> object cast(const T &value,
 template <typename T> T handle::cast() const { return pybind11::cast<T>(*this); }
 template <> inline void handle::cast() const { return; }
 
+template <typename T>
+typename std::enable_if<detail::move_always<T>::value || detail::move_if_unreferenced<T>::value, T>::type move(object &&obj) {
+    if (obj.ref_count() > 1)
+#if defined(NDEBUG)
+        throw cast_error("Unable to cast Python instance to C++ rvalue: instance has multiple references"
+            " (compile in debug mode for details)");
+#else
+        throw cast_error("Unable to move from Python " + (std::string) obj.get_type().str() +
+                " instance to C++ " + type_id<T>() + " instance: instance has multiple references");
+#endif
+
+    typedef detail::type_caster<T> type_caster;
+    type_caster conv;
+    if (!conv.load(obj, true))
+#if defined(NDEBUG)
+        throw cast_error("Unable to cast Python instance to C++ type (compile in debug mode for details)");
+#else
+        throw cast_error("Unable to cast Python instance of type " +
+            (std::string) obj.get_type().str() + " to C++ type '" + type_id<T>() + "''");
+#endif
+
+    // Move into a temporary and return that, because the reference may be a local value of `conv`
+    T ret = std::move(conv.operator T&());
+    return ret;
+}
+
+// Calling cast() on an rvalue calls pybind::cast with the object rvalue, which does:
+// - If we have to move (because T has no copy constructor), do it.  This will fail if the moved
+//   object has multiple references, but trying to copy will fail to compile.
+// - If both movable and copyable, check ref count: if 1, move; otherwise copy
+// - Otherwise (not movable), copy.
+template <typename T> typename std::enable_if<detail::move_always<T>::value, T>::type cast(object &&object) {
+    return move<T>(std::move(object));
+}
+template <typename T> typename std::enable_if<detail::move_if_unreferenced<T>::value, T>::type cast(object &&object) {
+    if (object.ref_count() > 1)
+        return cast<T>(object);
+    else
+        return move<T>(std::move(object));
+}
+template <typename T> typename std::enable_if<detail::move_never<T>::value, T>::type cast(object &&object) {
+    return cast<T>(object);
+}
+
+template <typename T> T object::cast() const & { return pybind11::cast<T>(*this); }
+template <typename T> T object::cast() && { return pybind11::cast<T>(std::move(*this)); }
+template <> inline void object::cast() const & { return; }
+template <> inline void object::cast() && { return; }
+
+
+
 template <return_value_policy policy = return_value_policy::automatic_reference,
           typename... Args> tuple make_tuple(Args&&... args_) {
     const size_t size = sizeof...(Args);

include/pybind11/pytypes.h

@@ -89,6 +89,11 @@ public:
         }
         return *this;
     }
+
+    // Calling cast() on an object lvalue just copies (via handle::cast)
+    template <typename T> T cast() const &;
+    // Calling on an object rvalue does a move, if needed and/or possible
+    template <typename T> T cast() &&;
 };
 
 NAMESPACE_BEGIN(detail)