Fix speed regression due to PySequence_Fast in for loop.

PySequence_Fast isn't so fast after all (due to multiple branching),
was a 35% slowdown for looping over lists. Now it handles tuples as
well with only about a 4% slowdown to lists compared to before.

Cython/Compiler/ExprNodes.py

@@ -1169,20 +1169,24 @@ class NextNode(AtomicExprNode):
         self.is_temp = 1
     
     def generate_result_code(self, code):
+        for py_type in ["List", "Tuple"]:
             code.putln(
-            "if (likely(%s != -1)) {" % self.iterator.counter.result_code)
+                "if (likely(Py%s_CheckExact(%s))) {" % (py_type, self.iterator.py_result()))
             code.putln(
-            "if (%s >= PySequence_Fast_GET_SIZE(%s)) break;" % (
+                "if (%s >= Py%s_GET_SIZE(%s)) break;" % (
                     self.iterator.counter.result_code,
+                    py_type,
                     self.iterator.py_result()))
             code.putln(
-            "%s = PySequence_Fast_GET_ITEM(%s, %s); Py_INCREF(%s); %s++;" % (
+                "%s = Py%s_GET_ITEM(%s, %s); Py_INCREF(%s); %s++;" % (
                     self.result_code,
+                    py_type,
                     self.iterator.py_result(),
                     self.iterator.counter.result_code,
                     self.result_code,
                     self.iterator.counter.result_code))
-        code.putln("} else {")
+            code.put("} else ")
+        code.putln("{")
         code.putln(
             "%s = PyIter_Next(%s);" % (
                 self.result_code,