Merge pull request #1187 from kmod/perf9

Extract the logic from DEFAULT_CLASS_SIMPLE

src/runtime/inline/dict.cpp

@@ -26,7 +26,7 @@ BoxedDictIterator::BoxedDictIterator(BoxedDict* d) : d(d), it(d->d.begin()), itE
 Box* dict_iter(Box* s) noexcept {
     assert(PyDict_Check(s));
     BoxedDict* self = static_cast<BoxedDict*>(s);
-    return new (&PyDictIterKey_Type) BoxedDictIterator(self);
+    return new (&PyDictIterKey_Type, FAST_GC) BoxedDictIterator(self);
 }
 
 Box* dictIterKeys(Box* s) {
@@ -36,13 +36,13 @@ Box* dictIterKeys(Box* s) {
 Box* dictIterValues(Box* s) {
     assert(PyDict_Check(s));
     BoxedDict* self = static_cast<BoxedDict*>(s);
-    return new (&PyDictIterValue_Type) BoxedDictIterator(self);
+    return new (&PyDictIterValue_Type, FAST_GC) BoxedDictIterator(self);
 }
 
 Box* dictIterItems(Box* s) {
     assert(PyDict_Check(s));
     BoxedDict* self = static_cast<BoxedDict*>(s);
-    return new (&PyDictIterItem_Type) BoxedDictIterator(self);
+    return new (&PyDictIterItem_Type, FAST_GC) BoxedDictIterator(self);
 }
 
 Box* dictIterIter(Box* s) {

src/runtime/inline/list.cpp

@@ -33,7 +33,7 @@ Box* listIterIter(Box* s) {
 Box* listIter(Box* s) noexcept {
     assert(PyList_Check(s));
     BoxedList* self = static_cast<BoxedList*>(s);
-    return new BoxedListIterator(self, 0);
+    return new (list_iterator_cls, FAST_GC) BoxedListIterator(self, 0);
 }
 
 Box* listiterHasnext(Box* s) {
@@ -89,7 +89,7 @@ Box* listiter_next(Box* s) noexcept {
 Box* listReversed(Box* s) {
     assert(PyList_Check(s));
     BoxedList* self = static_cast<BoxedList*>(s);
-    return new (list_reverse_iterator_cls) BoxedListIterator(self, self->size - 1);
+    return new (list_reverse_iterator_cls, FAST_GC) BoxedListIterator(self, self->size - 1);
 }
 
 Box* listreviterHasnext(Box* s) {

src/runtime/list.h

@@ -28,8 +28,6 @@ public:
     int pos;
     BoxedListIterator(BoxedList* l, int start);
 
-    DEFAULT_CLASS(list_iterator_cls);
-
     static void dealloc(BoxedListIterator* o) noexcept {
         PyObject_GC_UnTrack(o);
         Py_XDECREF(o->l);

src/runtime/long.h

@@ -33,7 +33,7 @@ public:
 
     BoxedLong() __attribute__((visibility("default"))) {}
 
-    DEFAULT_CLASS(long_cls);
+    DEFAULT_CLASS_SIMPLE(long_cls, false);
 };
 
 extern "C" Box* createLong(llvm::StringRef s);

src/runtime/types.cpp

@@ -2217,7 +2217,7 @@ public:
         b = NULL;
     }
 
-    DEFAULT_CLASS(attrwrapper_cls);
+    DEFAULT_CLASS_SIMPLE(attrwrapper_cls, true);
 
 
     BORROWED(Box*) getUnderlying() {

src/runtime/types.h

@@ -281,6 +281,42 @@ protected:
     friend void setupThread();
 };
 
+template <bool is_gc> void* Box::newFast(size_t size, BoxedClass* cls) {
+    ALLOC_STATS(cls);
+    assert(cls->tp_alloc == PyType_GenericAlloc);
+    assert(cls->tp_itemsize == 0);
+    assert(cls->tp_basicsize == size);
+    assert(cls->is_pyston_class);
+    assert(cls->attrs_offset == 0);
+    assert(is_gc == PyType_IS_GC(cls));
+    bool is_heaptype = false;
+    assert(is_heaptype == (bool)(cls->tp_flags & Py_TPFLAGS_HEAPTYPE));
+
+    /* Don't allocate classes through this -- we need to keep track of all class objects. */
+    assert(cls != type_cls);
+
+    /* note: we want to use size instead of tp_basicsize, since size is a compile-time constant */
+    void* mem;
+    if (is_gc)
+        mem = _PyObject_GC_Malloc(size);
+    else
+        mem = PyObject_MALLOC(size);
+    assert(mem);
+
+    Box* rtn = static_cast<Box*>(mem);
+
+    if (is_heaptype)
+        Py_INCREF(cls);
+
+    PyObject_INIT(rtn, cls);
+
+    if (is_gc)
+        _PyObject_GC_TRACK(rtn);
+
+    return rtn;
+    /* TODO: there should be a way to not have to do this nested inlining by hand */
+}
+
 // Corresponds to PyHeapTypeObject.  Very similar to BoxedClass, but allocates some extra space for
 // structures that otherwise might get allocated statically.  For instance, tp_as_number for builtin
 // types will usually point to a `static PyNumberMethods` object, but for a heap-allocated class it