One of the points in Armin Ronacher's The Python I Would Like To See is that slots are slow. That is, A() + A() is slower than A().__add__(A()) in the context of the following:

class A(object):
    def __add__(self, other):
        return 42

I'd like to investigate this claim for myself. To begin, let us repeat the experiment and see whether we get the same result:

$ cat x.py
class A(object):
    def __add__(self, other):
        return 42
$ ./python.exe
Python 3.5.0a4+ (default, Apr 25 2015, 21:57:28) 
[GCC 4.2.1 Compatible Apple LLVM 6.1.0 (clang-602.0.49)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> from x import A
>>> a = A()
>>> b = A()
>>> a + b
42
>>> quit()
$ ./python.exe -mtimeit -s 'from x import A; a = A(); b = A()' 'a + b'
1000000 loops, best of 3: 0.215 usec per loop
$ ./python.exe -mtimeit -s 'from x import A; a = A(); b = A()' 'a.__add__(b)'
10000000 loops, best of 3: 0.113 usec per loop

It would seem that Armin's claim stands up; a + b is indeed considerably slower than a.__add__(b).

First of all, an implicit assumption of Armin's claim is that a + b should be equivalent to a.__add__(b). Let us check this assumption by asking what does a + b mean in Python? The documentation for + is probably a good place to start:

The + (addition) operator yields the sum of its arguments. The arguments must either both be numbers or both be sequences of the same type. In the former case, the numbers are converted to a common type and then added together. In the latter case, the sequences are concatenated.

Well, uh, that doesn't explain the observed behaviour of a + b giving 42. Perhaps the documentation for __add__ will shed some light on the situation:

These methods are called to implement the binary arithmetic operations (+, [...]). For instance, to evaluate the expression x + y, where x is an instance of a class that has an __add__() method, x.__add__(y) is called. [...] If one of those methods does not support the operation with the supplied arguments, it should return NotImplemented.

Well, that explains the observed behaviour, and seems to pretty much straight up say that a + b means a.__add__(b). However, let's not get ahead of ourselves. On the off chance that it is relevant, let's consider the documentation for __radd__:

These methods are called to implement the binary arithmetic operations (+, [...]) with reflected (swapped) operands. These functions are only called if the left operand does not support the corresponding operation and the operands are of different types. For instance, to evaluate the expression x - y, where y is an instance of a class that has an __rsub__() method, y.__rsub__(x) is called if x.__sub__(y) returns NotImplemented.

Well, whad'ya know, it was relevant. With this extra bit of information, it seems like a + b is equivalent to something like:

if __add__ in a:
  tmp = a.__add__(b)
else:
  tmp = NotImplemented
if tmp is NotImplemented and type(a) != type(b):
  return b.__radd__(a)
else:
  return tmp

Of course, the story doesn't end there; immediately after the piece of documentation quoted above is the following gem:

Note: If the right operand’s type is a subclass of the left operand’s type and that subclass provides the reflected method for the operation, this method will be called before the left operand’s non-reflected method. This behavior allows subclasses to override their ancestors’ operations.

Bearing this in mind, maybe a + b is equivalent to something like:

if issubclass(type(b), type(a)) and __radd__ in b:
  tmp = b.__radd__(a)
  if tmp is not NotImplemented:
    return tmp
if __add__ in a:
  tmp = a.__add__(b)
else:
  tmp = NotImplemented
if tmp is NotImplemented and type(a) != type(b):
  return b.__radd__(a)
else:
  return tmp

I wish that the above were the full story, but alas it is not. Let us pluck another link out of thin air, this time to the documentation on special method lookup:

For custom classes, implicit invocations of special methods are only guaranteed to work correctly if defined on an object’s type, not in the object’s instance dictionary. [...] In addition to bypassing any instance attributes in the interest of correctness, implicit special method lookup generally also bypasses the __getattribute__() method even of the object’s metaclass.

I like to interpret this paragraph as saying bugger it, a + b means whatever the CPython interpreter does for a + b . Having studied the interpreter, the meaning of a + b is equivalent to something along the lines of the following:

def get(x, field):
  try:
    return getattr(type(x), field) # Doesn't call __getattribute__
  except AttributeError:
    return None

def has(x, field):
  return get(x, field) is not None

# From now on, `x.__yzw__` means `get(x, '__yzw__')`
# and `__abc__ in d` means `has(d, '__abc__')`

def tp_add_slot(x):
  if x is a builtin type or a type from a C extension:
    return ?
  elif __add__ in x or __radd__ in x:
    return slot_nb_add
  else:
    return None

def sq_concat_slot(x):
  return ?

def slot_nb_add(x, y):
  do_other = type(x) != type(y) and tp_add_slot(y) == slot_nb_add and __radd__ in y
  if tp_add_slot(x) == slot_nb_add:
    if do_other and issubclass(type(y), type(x)) and (__radd__ not in x or x.__radd__ != y.__radd__):
      tmp = y.__radd__(x)
      if tmp is NotImplemented:
        do_other = False
      else:
        return tmp
    if __add__ in x:
      tmp = x.__add__(y)
      if tmp is not NotImplemented:
        return tmp
  if do_other:
    return y.__radd__(x)
  return NotImplemented

slota = tp_add_slot(a)
slotb = tp_add_slot(b)
slotc = sq_concat_slot(a)
if slota == slotb:
  return slota(a, b)
if slota is not None and slotb is not None and issubclass(type(b), type(a)):
  tmp = slotb(a, b)
  if tmp is NotImplemented:
    slotb = None
  else:
    return tmp
if slota is not None:
  tmp = slota(a, b)
  if tmp is not NotImplementd:
    return tmp
if slotb is not None:
  tmp = slotb(a, b)
  if tmp is not NotImplementd:
    return tmp
if slotc is None:
  raise error
else:
  return slotc(a, b)

The conclusion of the above exploration is that a + b has a rather more nuanced meaning than just a.__add__(b). If we accept this conclusion, then perhaps it shouldn't be surprising that a + b is slower than a.__add__(b). However, in our case, a and b are the same type, so the above pseudo-code should pretty quickly conclude that the meaning of a + b, in our case, is just a.__add__(b).

Let us consider an alternative conclusion: the people behind the CPython interpreter have spent more time optimising a.__add__(b) than they have spent optimising a + b. To test this hypothesis, we need to dig into the bytecode of these two expressions. If we ignore the bytecode which is common to both expressions, then we can say that a.__add__(b) consists of two bytecode instructions (LOAD_ATTR and CALL_FUNCTION), while a + b consists of just a single bytecode instruction (BINARY_ADD).

Let's begin with a.__add__(b) and look at what happens when the bytecode is executed:

  1. Begin LOAD_ATTR instruction.
  2. Call PyObject_GetAttr.
  3. Call PyObject_GenericGetAttr (via the tp_getattro slot in type(a)).
  4. Call _PyObject_GenericGetAttrWithDict.
  5. Call _PyType_Lookup.
  6. Successfully find __add__ in the method cache.
  7. Return a function object to _PyObject_GenericGetAttrWithDict.
  8. Call func_descr_get (via the tp_descr_get slot in the type of the function object).
  9. Call PyMethod_New (to bind a to the first argument of the function).
  10. Return a method object from PyObject_GetAttr.
  11. Push the method object onto the stack.
  12. End LOAD_ATTR instruction.
  13. Begin CALL_FUNCTION instruction.
  14. Call call_function.
  15. Realise that we have a method object.
  16. Replace the stack entry underneath b with the bound argument from the method object.
  17. Call fast_function with the function from the method object.
  18. Call PyFrame_New to create a new stack frame.
  19. Call PyEval_EvalFrameEx to actually evaluate our __add__ code.
  20. Return 42 from call_function.
  21. Free the method object.
  22. End CALL_FUNCTION instruction.

On the other hand, for a + b, we have:

  1. Begin BINARY_ADD instruction.
  2. Call PyNumber_Add.
  3. Call binary_op1.
  4. Call slot_nb_add (via the tp_add slot in type(a)) [slot_nb_add is defined via a SLOT1BIN macro].
  5. Call call_maybe(a, &Id__add__, "(O)", b) [call_maybe is variadic after the string parameter].
  6. Call lookup_maybe [lookup_maybe is like PyObject_GetAttr, but only looks in the type, and doesn't invoke __getattribute__].
  7. Call _PyType_LookupId.
  8. Call _PyUnicode_FromId [this converts Id__add__ into a PyObject representing "__add__", but this conversion is cached, and therefore effectively free].
  9. Call _PyType_Lookup [as in step 5 above].
  10. Successfully find __add__ in the method cache [as in step 6 above].
  11. Return a function object from _PyType_LookupId.
  12. Call func_descr_get (via the tp_descr_get slot in the type of the function object) [as in step 8 above].
  13. Call PyMethod_New (to bind a to the first argument of the function) [as in step 9 above].
  14. Return a method object from lookup_maybe.
  15. Call Py_VaBuildValue, passing the string literal "(O)" and a reference to call_maybe's variadic arguments.
  16. Do a tonne of string literal parsing and variadic argument fetching and tuple construction, resulting in Py_VaBuildValue eventually returning a singleton tuple containing b.
  17. Call method_call (via the tp_call slot in the type of the method object).
  18. Construct a new two-element tuple, filling it with the bound argument from the method object, and the contents of the previously constructed singleton tuple. In other words, we now have the tuple (a, b).
  19. Call function_call (via the tp_call slot in the type of the function from the method object).
  20. Call PyEval_EvalCodeEx.
  21. Call _PyEval_EvalCodeWithName.
  22. Call PyFrame_New to create a new stack frame [as in step 18 above].
  23. Call PyEval_EvalFrameEx to actually evaluate our __add__ code [as in step 19 above].
  24. Return 42 from function_call.
  25. Free the two-element tuple.
  26. Return 42 from method_call.
  27. Free the singleton tuple.
  28. Free the method object.
  29. Return 42 from PyNumber_Add.
  30. End BINARY_ADD instruction.

One obvious diference is that a + b does far more manipulation of tuples and of variadic arguments. Given that call_maybe is always called with a format of "(O)", let's acknowledge this by changing its signature to be fixed-arg rather than vararg, and also construct an argument tuple via PyTuple_New / PyTuple_SET_ITEM rather than Py_VaBuildValue:

diff --git a/Objects/typeobject.c b/Objects/typeobject.c
index 4b99287..d27cc07 100644
--- a/Objects/typeobject.c
+++ b/Objects/typeobject.c
@@ -1465,29 +1465,22 @@ call_method(PyObject *o, _Py_Identifier *nameid, char *format, ...)
 /* Clone of call_method() that returns NotImplemented when the lookup fails. */
 
 static PyObject *
-call_maybe(PyObject *o, _Py_Identifier *nameid, char *format, ...)
+call_maybe(PyObject *o, _Py_Identifier *nameid, PyObject* p)
 {
-    va_list va;
     PyObject *args, *func = 0, *retval;
-    va_start(va, format);
 
     func = lookup_maybe(o, nameid);
     if (func == NULL) {
-        va_end(va);
         if (!PyErr_Occurred())
             Py_RETURN_NOTIMPLEMENTED;
         return NULL;
     }
 
-    if (format && *format)
-        args = Py_VaBuildValue(format, va);
-    else
-        args = PyTuple_New(0);
-
-    va_end(va);
-
+    args = PyTuple_New(1);
     if (args == NULL)
         return NULL;
+    PyTuple_SET_ITEM(args, 0, p);
+    Py_XINCREF(p);
 
     assert(PyTuple_Check(args));
     retval = PyObject_Call(func, args, NULL);
@@ -5624,20 +5617,20 @@ FUNCNAME(PyObject *self, PyObject *other) \
         if (do_other && \
             PyType_IsSubtype(Py_TYPE(other), Py_TYPE(self)) && \
             method_is_overloaded(self, other, &rop_id)) { \
-            r = call_maybe(other, &rop_id, "(O)", self); \
+            r = call_maybe(other, &rop_id, self); \
             if (r != Py_NotImplemented) \
                 return r; \
             Py_DECREF(r); \
             do_other = 0; \
         } \
-        r = call_maybe(self, &op_id, "(O)", other); \
+        r = call_maybe(self, &op_id, other); \
         if (r != Py_NotImplemented || \
             Py_TYPE(other) == Py_TYPE(self)) \
             return r; \
         Py_DECREF(r); \
     } \
     if (do_other) { \
-        return call_maybe(other, &rop_id, "(O)", self); \
+        return call_maybe(other, &rop_id, self); \
     } \
     Py_RETURN_NOTIMPLEMENTED; \
 }

This gives a nice little speedup; we're down from 0.215 usec to 0.176 usec:

$ make python.exe
...
$ ./python.exe -mtimeit -s 'from x import A; a = A(); b = A()' 'a + b'
10000000 loops, best of 3: 0.176 usec per loop

We're still falling somewhat short the of 0.113 usec time set by a.__add__(b), so let's copy step 15 of a.__add__(b) and special-case method objects:

diff --git a/Objects/typeobject.c b/Objects/typeobject.c
index 4b99287..2cd8e23 100644
--- a/Objects/typeobject.c
+++ b/Objects/typeobject.c
@@ -1465,36 +1465,43 @@ call_method(PyObject *o, _Py_Identifier *nameid, char *format, ...)
 /* Clone of call_method() that returns NotImplemented when the lookup fails. */
 
 static PyObject *
-call_maybe(PyObject *o, _Py_Identifier *nameid, char *format, ...)
+call_maybe(PyObject *o, _Py_Identifier *nameid, PyObject* p)
 {
-    va_list va;
-    PyObject *args, *func = 0, *retval;
-    va_start(va, format);
+    PyObject *args[2], *func = 0, *retval, *tuple;
+    int na = 1;
 
     func = lookup_maybe(o, nameid);
     if (func == NULL) {
-        va_end(va);
         if (!PyErr_Occurred())
             Py_RETURN_NOTIMPLEMENTED;
         return NULL;
     }
 
-    if (format && *format)
-        args = Py_VaBuildValue(format, va);
-    else
-        args = PyTuple_New(0);
-
-    va_end(va);
-
-    if (args == NULL)
-        return NULL;
+    args[1] = p;
+    if (PyMethod_Check(func) && PyMethod_GET_SELF(func) != NULL) {
+        PyObject *mself = PyMethod_GET_SELF(func);
+        PyObject *mfunc = PyMethod_GET_FUNCTION(func);
+        args[0] = mself;
+        na = 2;
+        Py_INCREF(mfunc);
+        Py_DECREF(func);
+        func = mfunc;
+    } else {
+        args[0] = NULL;
+    }
 
-    assert(PyTuple_Check(args));
-    retval = PyObject_Call(func, args, NULL);
+    tuple = PyTuple_New(na);
+    if (tuple == NULL) {
+        retval = NULL;
+    } else {
+        memcpy(((PyTupleObject *)tuple)->ob_item, args, sizeof(PyObject*) * na);
+        Py_XINCREF(args[0]);
+        Py_XINCREF(args[1]);
+        retval = PyObject_Call(func, tuple, NULL);
+        Py_DECREF(tuple);
+    }
 
-    Py_DECREF(args);
     Py_DECREF(func);
-
     return retval;
 }
 
@@ -5624,20 +5631,20 @@ FUNCNAME(PyObject *self, PyObject *other) \
         if (do_other && \
             PyType_IsSubtype(Py_TYPE(other), Py_TYPE(self)) && \
             method_is_overloaded(self, other, &rop_id)) { \
-            r = call_maybe(other, &rop_id, "(O)", self); \
+            r = call_maybe(other, &rop_id, self); \
             if (r != Py_NotImplemented) \
                 return r; \
             Py_DECREF(r); \
             do_other = 0; \
         } \
-        r = call_maybe(self, &op_id, "(O)", other); \
+        r = call_maybe(self, &op_id, other); \
         if (r != Py_NotImplemented || \
             Py_TYPE(other) == Py_TYPE(self)) \
             return r; \
         Py_DECREF(r); \
     } \
     if (do_other) { \
-        return call_maybe(other, &rop_id, "(O)", self); \
+        return call_maybe(other, &rop_id, self); \
     } \
     Py_RETURN_NOTIMPLEMENTED; \
 }

This gives another nice little speedup; we're down from 0.176 usec to 0.155 usec:

$ make python.exe
...
$ ./python.exe -mtimeit -s 'from x import A; a = A(); b = A()' 'a + b'
10000000 loops, best of 3: 0.155 usec per loop

Even better would be to also pull the fast_function trick that the interpreter does at step 17 in order to call a function without creating any argument tuples at all:

diff --git a/Include/ceval.h b/Include/ceval.h
index 6811367..f0997ac 100644
--- a/Include/ceval.h
+++ b/Include/ceval.h
@@ -10,6 +10,9 @@ extern "C" {
 PyAPI_FUNC(PyObject *) PyEval_CallObjectWithKeywords(
     PyObject *, PyObject *, PyObject *);
 
+PyAPI_FUNC(PyObject *)
+PyEval_FastFunction(PyObject *func, PyObject **stack, int n);
+
 /* Inline this */
 #define PyEval_CallObject(func,arg) \
     PyEval_CallObjectWithKeywords(func, arg, (PyObject *)NULL)
diff --git a/Objects/typeobject.c b/Objects/typeobject.c
index 4b99287..6419ea2 100644
--- a/Objects/typeobject.c
+++ b/Objects/typeobject.c
@@ -1465,36 +1465,47 @@ call_method(PyObject *o, _Py_Identifier *nameid, char *format, ...)
 /* Clone of call_method() that returns NotImplemented when the lookup fails. */
 
 static PyObject *
-call_maybe(PyObject *o, _Py_Identifier *nameid, char *format, ...)
+call_maybe(PyObject *o, _Py_Identifier *nameid, PyObject* p)
 {
-    va_list va;
-    PyObject *args, *func = 0, *retval;
-    va_start(va, format);
+    PyObject *args[2], *func = 0, *retval;
+    int na = 1;
 
     func = lookup_maybe(o, nameid);
     if (func == NULL) {
-        va_end(va);
         if (!PyErr_Occurred())
             Py_RETURN_NOTIMPLEMENTED;
         return NULL;
     }
 
-    if (format && *format)
-        args = Py_VaBuildValue(format, va);
-    else
-        args = PyTuple_New(0);
-
-    va_end(va);
-
-    if (args == NULL)
-        return NULL;
+    args[1] = p;
+    if (PyMethod_Check(func) && PyMethod_GET_SELF(func) != NULL) {
+        PyObject *mself = PyMethod_GET_SELF(func);
+        PyObject *mfunc = PyMethod_GET_FUNCTION(func);
+        args[0] = mself;
+        na = 2;
+        Py_INCREF(mfunc);
+        Py_DECREF(func);
+        func = mfunc;
+    } else {
+        args[0] = NULL;
+    }
 
-    assert(PyTuple_Check(args));
-    retval = PyObject_Call(func, args, NULL);
+    if (PyFunction_Check(func)) {
+        retval = PyEval_FastFunction(func, &args[2], na);
+    } else {
+        PyObject* tuple = PyTuple_New(na);
+        if (tuple == NULL) {
+            retval = NULL;
+        } else {
+            memcpy(((PyTupleObject *)tuple)->ob_item, args, sizeof(PyObject*) * na);
+            Py_XINCREF(args[0]);
+            Py_XINCREF(args[1]);
+            retval = PyObject_Call(func, tuple, NULL);
+            Py_DECREF(tuple);
+        }
+    }
 
-    Py_DECREF(args);
     Py_DECREF(func);
-
     return retval;
 }
 
@@ -5624,20 +5635,20 @@ FUNCNAME(PyObject *self, PyObject *other) \
         if (do_other && \
             PyType_IsSubtype(Py_TYPE(other), Py_TYPE(self)) && \
             method_is_overloaded(self, other, &rop_id)) { \
-            r = call_maybe(other, &rop_id, "(O)", self); \
+            r = call_maybe(other, &rop_id, self); \
             if (r != Py_NotImplemented) \
                 return r; \
             Py_DECREF(r); \
             do_other = 0; \
         } \
-        r = call_maybe(self, &op_id, "(O)", other); \
+        r = call_maybe(self, &op_id, other); \
         if (r != Py_NotImplemented || \
             Py_TYPE(other) == Py_TYPE(self)) \
             return r; \
         Py_DECREF(r); \
     } \
     if (do_other) { \
-        return call_maybe(other, &rop_id, "(O)", self); \
+        return call_maybe(other, &rop_id, self); \
     } \
     Py_RETURN_NOTIMPLEMENTED; \
 }
diff --git a/Python/ceval.c b/Python/ceval.c
index 2f3d3ad..bf6aedc 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -4329,6 +4329,12 @@ call_function(PyObject ***pp_stack, int oparg
     return x;
 }
 
+PyAPI_FUNC(PyObject *)
+PyEval_FastFunction(PyObject *func, PyObject **stack, int n)
+{
+  return fast_function(func, &stack, n, n, 0);
+}
+
 /* The fast_function() function optimize calls for which no argument
    tuple is necessary; the objects are passed directly from the stack.
    For the simplest case -- a function that takes only positional

And with that, we're down from 0.155 usec to 0.113 usec:

$ make python.exe
...
$ ./python.exe -mtimeit -s 'from x import A; a = A(); b = A()' 'a + b'
10000000 loops, best of 3: 0.113 usec per loop

So, it seems that slots aren't intrinsically slow. Provided that the implementation of slots in typeobject.c is taught to use the exact same tricks that the interpreter does, then they are the exact same speed as non-slots. We could even go further and elide construction of the method object entirely:

diff --git a/Include/ceval.h b/Include/ceval.h
index 6811367..f0997ac 100644
--- a/Include/ceval.h
+++ b/Include/ceval.h
@@ -10,6 +10,9 @@ extern "C" {
 PyAPI_FUNC(PyObject *) PyEval_CallObjectWithKeywords(
     PyObject *, PyObject *, PyObject *);
 
+PyAPI_FUNC(PyObject *)
+PyEval_FastFunction(PyObject *func, PyObject **stack, int n);
+
 /* Inline this */
 #define PyEval_CallObject(func,arg) \
     PyEval_CallObjectWithKeywords(func, arg, (PyObject *)NULL)
diff --git a/Objects/typeobject.c b/Objects/typeobject.c
index 4b99287..c4ffa70 100644
--- a/Objects/typeobject.c
+++ b/Objects/typeobject.c
@@ -1465,36 +1465,64 @@ call_method(PyObject *o, _Py_Identifier *nameid, char *format, ...)
 /* Clone of call_method() that returns NotImplemented when the lookup fails. */
 
 static PyObject *
-call_maybe(PyObject *o, _Py_Identifier *nameid, char *format, ...)
+call_maybe(PyObject *o, _Py_Identifier *nameid, PyObject* p)
 {
-    va_list va;
-    PyObject *args, *func = 0, *retval;
-    va_start(va, format);
+    PyObject *args[2], *func = 0, *retval;
+    int na = 2;
 
-    func = lookup_maybe(o, nameid);
+    args[1] = p;
+    func = _PyType_LookupId(Py_TYPE(o), nameid);
     if (func == NULL) {
-        va_end(va);
         if (!PyErr_Occurred())
             Py_RETURN_NOTIMPLEMENTED;
         return NULL;
     }
+    if (PyFunction_Check(func)) {
+        Py_INCREF(func);
+        args[0] = o;
+        retval = PyEval_FastFunction(func, &args[2], na);
+    } else {
+        descrgetfunc f = Py_TYPE(func)->tp_descr_get;
+        if (f == NULL) {
+            Py_INCREF(func);
+        } else {
+            func = f(func, o, (PyObject *)(Py_TYPE(o)));
+            if (func == NULL) {
+                if (!PyErr_Occurred())
+                    Py_RETURN_NOTIMPLEMENTED;
+                return NULL;
+            }
+        }
 
-    if (format && *format)
-        args = Py_VaBuildValue(format, va);
-    else
-        args = PyTuple_New(0);
-
-    va_end(va);
-
-    if (args == NULL)
-        return NULL;
-
-    assert(PyTuple_Check(args));
-    retval = PyObject_Call(func, args, NULL);
+        if (PyMethod_Check(func) && PyMethod_GET_SELF(func) != NULL) {
+            PyObject *mself = PyMethod_GET_SELF(func);
+            PyObject *mfunc = PyMethod_GET_FUNCTION(func);
+            args[0] = mself;
+            Py_INCREF(mfunc);
+            Py_DECREF(func);
+            func = mfunc;
+        } else {
+            args[0] = NULL;
+            na = 1;
+        }
+    
+        if (PyFunction_Check(func)) {
+            retval = PyEval_FastFunction(func, &args[2], na);
+        } else {
+            PyObject* tuple = PyTuple_New(na);
+            if (tuple == NULL) {
+                retval = NULL;
+            } else {
+                memcpy(((PyTupleObject *)tuple)->ob_item, args, sizeof(PyObject*) * na);
+                Py_XINCREF(args[0]);
+                Py_XINCREF(args[1]);
+                retval = PyObject_Call(func, tuple, NULL);
+                Py_DECREF(tuple);
+            }
+        }
+    }
 
-    Py_DECREF(args);
     Py_DECREF(func);
-
     return retval;
 }
 
@@ -5624,20 +5652,20 @@ FUNCNAME(PyObject *self, PyObject *other) \
         if (do_other && \
             PyType_IsSubtype(Py_TYPE(other), Py_TYPE(self)) && \
             method_is_overloaded(self, other, &rop_id)) { \
-            r = call_maybe(other, &rop_id, "(O)", self); \
+            r = call_maybe(other, &rop_id, self); \
             if (r != Py_NotImplemented) \
                 return r; \
             Py_DECREF(r); \
             do_other = 0; \
         } \
-        r = call_maybe(self, &op_id, "(O)", other); \
+        r = call_maybe(self, &op_id, other); \
         if (r != Py_NotImplemented || \
             Py_TYPE(other) == Py_TYPE(self)) \
             return r; \
         Py_DECREF(r); \
     } \
     if (do_other) { \
-        return call_maybe(other, &rop_id, "(O)", self); \
+        return call_maybe(other, &rop_id, self); \
     } \
     Py_RETURN_NOTIMPLEMENTED; \
 }
diff --git a/Python/ceval.c b/Python/ceval.c
index 2f3d3ad..bf6aedc 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -4329,6 +4329,12 @@ call_function(PyObject ***pp_stack, int oparg
     return x;
 }
 
+PyAPI_FUNC(PyObject *)
+PyEval_FastFunction(PyObject *func, PyObject **stack, int n)
+{
+  return fast_function(func, &stack, n, n, 0);
+}
+
 /* The fast_function() function optimize calls for which no argument
    tuple is necessary; the objects are passed directly from the stack.
    For the simplest case -- a function that takes only positional

With this extra optimisation, we're down from 0.113 usec to 0.0972 usec:

$ make python.exe
...
$ ./python.exe -mtimeit -s 'from x import A; a = A(); b = A()' 'a + b'
10000000 loops, best of 3: 0.0972 usec per loop

In conclusion, slots don't need to be slow - the above diff makes them fast (at least for some binary operators; applying similar transformations to other slots is left as an exercise to the reader).