add tuple free list

this uses a slightly modified version of cpythons free list.
I removed the special handling for the empty tuple and fixed some C/C++ type mismatches.

src/runtime/tuple.cpp

@@ -27,6 +27,11 @@
 
 namespace pyston {
 
+#if PyTuple_MAXSAVESIZE > 0
+BoxedTuple* BoxedTuple::free_list[PyTuple_MAXSAVESIZE];
+int BoxedTuple::numfree[PyTuple_MAXSAVESIZE];
+#endif
+
 extern "C" Box* createTuple(int64_t nelts, Box** elts) {
     return BoxedTuple::create(nelts, elts);
 }
@@ -725,7 +730,22 @@ extern "C" void _PyTuple_MaybeUntrack(PyObject* op) noexcept {
 }
 
 extern "C" int PyTuple_ClearFreeList() noexcept {
-    return 0; // number of entries cleared
+    int freelist_size = 0;
+#if PyTuple_MAXSAVESIZE > 0
+    int i;
+    for (i = 1; i < PyTuple_MAXSAVESIZE; i++) {
+        PyTupleObject* p = (PyTupleObject*)BoxedTuple::free_list[i];
+        freelist_size += BoxedTuple::numfree[i];
+        BoxedTuple::free_list[i] = NULL;
+        BoxedTuple::numfree[i] = 0;
+        while (p) {
+            PyTupleObject* q = p;
+            p = (PyTupleObject*)(p->ob_item[0]);
+            PyObject_GC_Del(q);
+        }
+    }
+#endif
+    return freelist_size;
 }
 
 void setupTuple() {

src/runtime/types.cpp

@@ -3780,7 +3780,7 @@ void HiddenClass::dump() noexcept {
     }
 }
 
-static void tupledealloc(PyTupleObject* op) noexcept {
+void BoxedTuple::dealloc(PyTupleObject* op) noexcept {
     Py_ssize_t i;
     Py_ssize_t len = Py_SIZE(op);
     PyObject_GC_UnTrack(op);
@@ -3789,10 +3789,11 @@ static void tupledealloc(PyTupleObject* op) noexcept {
         while (--i >= 0)
             Py_XDECREF(op->ob_item[i]);
 #if PyTuple_MAXSAVESIZE > 0
-        if (len < PyTuple_MAXSAVESIZE && numfree[len] < PyTuple_MAXFREELIST && Py_TYPE(op) == &PyTuple_Type) {
+        if (likely(len < PyTuple_MAXSAVESIZE && BoxedTuple::numfree[len] < PyTuple_MAXFREELIST
+                   && ((BoxedTuple*)op)->cls == tuple_cls)) {
             op->ob_item[0] = (PyObject*)free_list[len];
             numfree[len]++;
-            free_list[len] = op;
+            free_list[len] = (BoxedTuple*)op;
             goto done; /* return */
         }
 #endif
@@ -4247,7 +4248,7 @@ void setupRuntime() {
     // Not sure why CPython defines sizeof(PyTupleObject) to include one element,
     // but we copy that, which means we have to subtract that extra pointer to get the tp_basicsize:
     tuple_cls = new (0) BoxedClass(object_cls, 0, 0, sizeof(BoxedTuple) - sizeof(Box*), false, "tuple", true,
-                                   (destructor)tupledealloc, NULL, true, (traverseproc)tupletraverse, NOCLEAR);
+                                   (destructor)BoxedTuple::dealloc, NULL, true, (traverseproc)tupletraverse, NOCLEAR);
 
     tuple_cls->tp_flags |= Py_TPFLAGS_TUPLE_SUBCLASS;
     tuple_cls->tp_itemsize = sizeof(Box*);

src/runtime/types.h

@@ -695,7 +695,16 @@ static_assert(offsetof(BoxedList, elts) == offsetof(PyListObject, ob_item), "");
 static_assert(offsetof(GCdArray, elts) == 0, "");
 static_assert(offsetof(BoxedList, capacity) == offsetof(PyListObject, allocated), "");
 
+#define PyTuple_MAXSAVESIZE 20   /* Largest tuple to save on free list */
+#define PyTuple_MAXFREELIST 2000 /* Maximum number of tuples of each size to save */
+extern "C" int PyTuple_ClearFreeList() noexcept;
 class BoxedTuple : public BoxVar {
+private:
+#if PyTuple_MAXSAVESIZE > 0
+    static BoxedTuple* free_list[PyTuple_MAXSAVESIZE];
+    static int numfree[PyTuple_MAXSAVESIZE];
+#endif
+
 public:
     static BoxedTuple* create(int64_t size) {
         if (size == 0) {
@@ -834,7 +843,7 @@ public:
         return BoxVar::operator new(size, cls, nitems);
     }
 
-    void* operator new(size_t size, size_t nitems) __attribute__((visibility("default"))) {
+    void* operator new(size_t /*size*/, size_t nitems) __attribute__((visibility("default"))) {
         ALLOC_STATS_VAR(tuple_cls)
 
         assert(tuple_cls->tp_alloc == PyType_GenericAlloc);
@@ -843,11 +852,31 @@ public:
         assert(tuple_cls->is_pyston_class);
         assert(tuple_cls->attrs_offset == 0);
 
-        BoxVar* rtn = static_cast<BoxVar*>(PyObject_GC_NewVar(BoxedTuple, &PyTuple_Type, nitems));
-        assert(rtn);
-        _PyObject_GC_TRACK(rtn);
-
-        return rtn;
+        BoxedTuple* op = NULL;
+#if PyTuple_MAXSAVESIZE > 0
+        if (likely(nitems < PyTuple_MAXSAVESIZE && (op = free_list[nitems]) != NULL)) {
+            free_list[nitems] = (BoxedTuple*)op->elts[0];
+            numfree[nitems]--;
+/* Inline PyObject_InitVar */
+#ifdef Py_TRACE_REFS
+            Py_SIZE(op) = nitems;
+            Py_TYPE(op) = &PyTuple_Type;
+#endif
+            _Py_NewReference((PyObject*)op);
+        } else
+#endif
+        {
+            Py_ssize_t nbytes = nitems * sizeof(PyObject*);
+            /* Check for overflow */
+            if (unlikely(nbytes / sizeof(PyObject*) != (size_t)nitems
+                         || (nbytes > PY_SSIZE_T_MAX - sizeof(PyTupleObject) - sizeof(PyObject*)))) {
+                return PyErr_NoMemory();
+            }
+
+            op = PyObject_GC_NewVar(BoxedTuple, &PyTuple_Type, nitems);
+        }
+        _PyObject_GC_TRACK(op);
+        return (PyObject*)op;
     }
 
 private:
@@ -871,6 +900,9 @@ public:
         Box* elts[0];
         Box* _elts[1];
     };
+
+    static void dealloc(PyTupleObject* op) noexcept;
+    friend int PyTuple_ClearFreeList() noexcept;
 };
 static_assert(sizeof(BoxedTuple) == sizeof(PyTupleObject), "");
 static_assert(offsetof(BoxedTuple, ob_size) == offsetof(PyTupleObject, ob_size), "");