Merge branch 'master' into sh_merge_master

include/pybind11/stl.h

@@ -13,6 +13,7 @@
 #include "detail/common.h"
 
 #include <deque>
+#include <initializer_list>
 #include <list>
 #include <map>
 #include <ostream>
@@ -35,6 +36,89 @@
 PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
 PYBIND11_NAMESPACE_BEGIN(detail)
 
+//
+// Begin: Equivalent of
+//        https://github.com/google/clif/blob/ae4eee1de07cdf115c0c9bf9fec9ff28efce6f6c/clif/python/runtime.cc#L388-L438
+/*
+The three `PyObjectTypeIsConvertibleTo*()` functions below are
+the result of converging the behaviors of pybind11 and PyCLIF
+(http://github.com/google/clif).
+
+Originally PyCLIF was extremely far on the permissive side of the spectrum,
+while pybind11 was very far on the strict side. Originally PyCLIF accepted any
+Python iterable as input for a C++ `vector`/`set`/`map` argument, as long as
+the elements were convertible. The obvious (in hindsight) problem was that
+any empty Python iterable could be passed to any of these C++ types, e.g. `{}`
+was accepted for C++ `vector`/`set` arguments, or `[]` for C++ `map` arguments.
+
+The functions below strike a practical permissive-vs-strict compromise,
+informed by tens of thousands of use cases in the wild. A main objective is
+to prevent accidents and improve readability:
+
+- Python literals must match the C++ types.
+
+- For C++ `set`: The potentially reducing conversion from a Python sequence
+  (e.g. Python `list` or `tuple`) to a C++ `set` must be explicit, by going
+  through a Python `set`.
+
+- However, a Python `set` can still be passed to a C++ `vector`. The rationale
+  is that this conversion is not reducing. Implicit conversions of this kind
+  are also fairly commonly used, therefore enforcing explicit conversions
+  would have an unfavorable cost : benefit ratio; more sloppily speaking,
+  such an enforcement would be more annoying than helpful.
+*/
+
+inline bool PyObjectIsInstanceWithOneOfTpNames(PyObject *obj,
+                                               std::initializer_list<const char *> tp_names) {
+    if (PyType_Check(obj)) {
+        return false;
+    }
+    const char *obj_tp_name = Py_TYPE(obj)->tp_name;
+    for (const auto *tp_name : tp_names) {
+        if (std::strcmp(obj_tp_name, tp_name) == 0) {
+            return true;
+        }
+    }
+    return false;
+}
+
+inline bool PyObjectTypeIsConvertibleToStdVector(PyObject *obj) {
+    if (PySequence_Check(obj) != 0) {
+        return !PyUnicode_Check(obj) && !PyBytes_Check(obj);
+    }
+    return (PyGen_Check(obj) != 0) || (PyAnySet_Check(obj) != 0)
+           || PyObjectIsInstanceWithOneOfTpNames(
+               obj, {"dict_keys", "dict_values", "dict_items", "map", "zip"});
+}
+
+inline bool PyObjectTypeIsConvertibleToStdSet(PyObject *obj) {
+    return (PyAnySet_Check(obj) != 0) || PyObjectIsInstanceWithOneOfTpNames(obj, {"dict_keys"});
+}
+
+inline bool PyObjectTypeIsConvertibleToStdMap(PyObject *obj) {
+    if (PyDict_Check(obj)) {
+        return true;
+    }
+    // Implicit requirement in the conditions below:
+    // A type with `.__getitem__()` & `.items()` methods must implement these
+    // to be compatible with https://docs.python.org/3/c-api/mapping.html
+    if (PyMapping_Check(obj) == 0) {
+        return false;
+    }
+    PyObject *items = PyObject_GetAttrString(obj, "items");
+    if (items == nullptr) {
+        PyErr_Clear();
+        return false;
+    }
+    bool is_convertible = (PyCallable_Check(items) != 0);
+    Py_DECREF(items);
+    return is_convertible;
+}
+
+//
+// End: Equivalent of clif/python/runtime.cc
+//
+
 /// Extracts an const lvalue reference or rvalue reference for U based on the type of T (e.g. for
 /// forwarding a container element).  Typically used indirect via forwarded_type(), below.
 template <typename T, typename U>
@@ -66,17 +150,10 @@ private:
     }
     void reserve_maybe(const anyset &, void *) {}
 
-public:
-    bool load(handle src, bool convert) {
-        if (!isinstance<anyset>(src)) {
-            return false;
-        }
-        auto s = reinterpret_borrow<anyset>(src);
-        value.clear();
-        reserve_maybe(s, &value);
-        for (auto entry : s) {
+    bool convert_iterable(const iterable &itbl, bool convert) {
+        for (const auto &it : itbl) {
             key_conv conv;
-            if (!conv.load(entry, convert)) {
+            if (!conv.load(it, convert)) {
                 return false;
             }
             value.insert(cast_op<Key &&>(std::move(conv)));
@@ -84,6 +161,29 @@ public:
         return true;
     }
 
+    bool convert_anyset(anyset s, bool convert) {
+        value.clear();
+        reserve_maybe(s, &value);
+        return convert_iterable(s, convert);
+    }
+
+public:
+    bool load(handle src, bool convert) {
+        if (!PyObjectTypeIsConvertibleToStdSet(src.ptr())) {
+            return false;
+        }
+        if (isinstance<anyset>(src)) {
+            value.clear();
+            return convert_anyset(reinterpret_borrow<anyset>(src), convert);
+        }
+        if (!convert) {
+            return false;
+        }
+        assert(isinstance<iterable>(src));
+        value.clear();
+        return convert_iterable(reinterpret_borrow<iterable>(src), convert);
+    }
+
     template <typename T>
     static handle cast(T &&src, return_value_policy policy, handle parent) {
         if (!std::is_lvalue_reference<T>::value) {
@@ -115,15 +215,10 @@ private:
     }
     void reserve_maybe(const dict &, void *) {}
 
-public:
-    bool load(handle src, bool convert) {
-        if (!isinstance<dict>(src)) {
-            return false;
-        }
-        auto d = reinterpret_borrow<dict>(src);
+    bool convert_elements(const dict &d, bool convert) {
         value.clear();
         reserve_maybe(d, &value);
-        for (auto it : d) {
+        for (const auto &it : d) {
             key_conv kconv;
             value_conv vconv;
             if (!kconv.load(it.first.ptr(), convert) || !vconv.load(it.second.ptr(), convert)) {
@@ -134,6 +229,25 @@ public:
         return true;
     }
 
+public:
+    bool load(handle src, bool convert) {
+        if (!PyObjectTypeIsConvertibleToStdMap(src.ptr())) {
+            return false;
+        }
+        if (isinstance<dict>(src)) {
+            return convert_elements(reinterpret_borrow<dict>(src), convert);
+        }
+        if (!convert) {
+            return false;
+        }
+        auto items = reinterpret_steal<object>(PyMapping_Items(src.ptr()));
+        if (!items) {
+            throw error_already_set();
+        }
+        assert(isinstance<iterable>(items));
+        return convert_elements(dict(reinterpret_borrow<iterable>(items)), convert);
+    }
+
     template <typename T>
     static handle cast(T &&src, return_value_policy policy, handle parent) {
         dict d;
@@ -166,20 +280,21 @@ struct list_caster {
     using value_conv = make_caster<Value>;
 
     bool load(handle src, bool convert) {
-        if (!isinstance<sequence>(src) || isinstance<bytes>(src) || isinstance<str>(src)) {
+        if (!PyObjectTypeIsConvertibleToStdVector(src.ptr())) {
             return false;
         }
-        auto s = reinterpret_borrow<sequence>(src);
-        value.clear();
-        reserve_maybe(s, &value);
-        for (const auto &it : s) {
-            value_conv conv;
-            if (!conv.load(it, convert)) {
-                return false;
-            }
-            value.push_back(cast_op<Value &&>(std::move(conv)));
+        if (isinstance<sequence>(src)) {
+            return convert_elements(src, convert);
         }
-        return true;
+        if (!convert) {
+            return false;
+        }
+        // Designed to be behavior-equivalent to passing tuple(src) from Python:
+        // The conversion to a tuple will first exhaust the generator object, to ensure that
+        // the generator is not left in an unpredictable (to the caller) partially-consumed
+        // state.
+        assert(isinstance<iterable>(src));
+        return convert_elements(tuple(reinterpret_borrow<iterable>(src)), convert);
     }
 
 private:
@@ -189,6 +304,20 @@ private:
     }
     void reserve_maybe(const sequence &, void *) {}
 
+    bool convert_elements(handle seq, bool convert) {
+        auto s = reinterpret_borrow<sequence>(seq);
+        value.clear();
+        reserve_maybe(s, &value);
+        for (const auto &it : seq) {
+            value_conv conv;
+            if (!conv.load(it, convert)) {
+                return false;
+            }
+            value.push_back(cast_op<Value &&>(std::move(conv)));
+        }
+        return true;
+    }
+
 public:
     template <typename T>
     static handle cast(T &&src, return_value_policy policy, handle parent) {
@@ -237,12 +366,8 @@ private:
         return size == Size;
     }
 
-public:
-    bool load(handle src, bool convert) {
-        if (!isinstance<sequence>(src)) {
-            return false;
-        }
-        auto l = reinterpret_borrow<sequence>(src);
+    bool convert_elements(handle seq, bool convert) {
+        auto l = reinterpret_borrow<sequence>(seq);
         if (!require_size(l.size())) {
             return false;
         }
@@ -257,6 +382,25 @@ public:
         return true;
     }
 
+public:
+    bool load(handle src, bool convert) {
+        if (!PyObjectTypeIsConvertibleToStdVector(src.ptr())) {
+            return false;
+        }
+        if (isinstance<sequence>(src)) {
+            return convert_elements(src, convert);
+        }
+        if (!convert) {
+            return false;
+        }
+        // Designed to be behavior-equivalent to passing tuple(src) from Python:
+        // The conversion to a tuple will first exhaust the generator object, to ensure that
+        // the generator is not left in an unpredictable (to the caller) partially-consumed
+        // state.
+        assert(isinstance<iterable>(src));
+        return convert_elements(tuple(reinterpret_borrow<iterable>(src)), convert);
+    }
+
     template <typename T>
     static handle cast(T &&src, return_value_policy policy, handle parent) {
         list l(src.size());