Merge pull request #873 from kmod/perf3

Some random runtime function optimizations

src/runtime/builtin_modules/builtins.cpp

@@ -250,15 +250,35 @@ extern "C" Box* max(Box* arg0, BoxedTuple* args, BoxedDict* kwargs) {
     return maxElement;
 }
 
-extern "C" Box* next(Box* iterator, Box* _default) {
-    try {
-        static BoxedString* next_str = internStringImmortal("next");
-        CallattrFlags callattr_flags{.cls_only = true, .null_on_nonexistent = false, .argspec = ArgPassSpec(0) };
-        return callattr(iterator, next_str, callattr_flags, NULL, NULL, NULL, NULL, NULL);
-    } catch (ExcInfo e) {
-        if (_default && e.matches(StopIteration))
-            return _default;
-        throw e;
+extern "C" Box* next(Box* iterator, Box* _default) noexcept {
+    if (!PyIter_Check(iterator)) {
+        PyErr_Format(PyExc_TypeError, "%.200s object is not an iterator", iterator->cls->tp_name);
+        return NULL;
+    }
+
+    Box* rtn;
+
+    if (iterator->cls->tp_iternext == slot_tp_iternext) {
+        rtn = iterator->cls->call_nextIC(iterator);
+    } else {
+        rtn = iterator->cls->tp_iternext(iterator);
+    }
+
+    if (rtn != NULL) {
+        return rtn;
+    } else if (_default != NULL) {
+        if (PyErr_Occurred()) {
+            if (!PyErr_ExceptionMatches(PyExc_StopIteration))
+                return NULL;
+            PyErr_Clear();
+        }
+        Py_INCREF(_default);
+        return _default;
+    } else if (PyErr_Occurred()) {
+        return NULL;
+    } else {
+        PyErr_SetNone(PyExc_StopIteration);
+        return NULL;
     }
 }
 
@@ -1316,6 +1336,11 @@ Box* print(BoxedTuple* args, BoxedDict* kwargs) {
 Box* getreversed(Box* o) {
     static BoxedString* reversed_str = internStringImmortal("__reversed__");
 
+    // common case:
+    if (o->cls == list_cls) {
+        return listReversed(o);
+    }
+
     // TODO add rewriting to this?  probably want to try to avoid this path though
     CallattrFlags callattr_flags{.cls_only = true, .null_on_nonexistent = true, .argspec = ArgPassSpec(0) };
     Box* r = callattr(o, reversed_str, callattr_flags, NULL, NULL, NULL, NULL, NULL);
@@ -1582,8 +1607,9 @@ void setupBuiltins() {
     max_obj = new BoxedBuiltinFunctionOrMethod(boxRTFunction((void*)max, UNKNOWN, 1, 1, true, true), "max", { None });
     builtins_module->giveAttr("max", max_obj);
 
-    builtins_module->giveAttr("next", new BoxedBuiltinFunctionOrMethod(
-                                          boxRTFunction((void*)next, UNKNOWN, 2, 1, false, false), "next", { NULL }));
+    builtins_module->giveAttr("next", new BoxedBuiltinFunctionOrMethod(boxRTFunction((void*)next, UNKNOWN, 2, 1, false,
+                                                                                     false, ParamNames::empty(), CAPI),
+                                                                       "next", { NULL }));
 
     builtins_module->giveAttr("sum", new BoxedBuiltinFunctionOrMethod(
                                          boxRTFunction((void*)sum, UNKNOWN, 2, 1, false, false), "sum", { boxInt(0) }));

src/runtime/dict.cpp

@@ -20,6 +20,7 @@
 #include "core/stats.h"
 #include "core/types.h"
 #include "gc/collector.h"
+#include "runtime/ics.h"
 #include "runtime/inline/list.h"
 #include "runtime/objmodel.h"
 #include "runtime/types.h"
@@ -238,11 +239,24 @@ template <enum ExceptionStyle S> Box* dictGetitem(BoxedDict* self, Box* k) noexc
     if (it == self->d.end()) {
         // Try calling __missing__ if this is a subclass
         if (self->cls != dict_cls) {
+            // Special-case defaultdict, assuming that it's the main time we will actually hit this.
+            // We could just use a single runtime IC here, or have a small cache that maps type->runtimeic.
+            // Or use a polymorphic runtime ic.
+            static BoxedClass* defaultdict_cls = NULL;
+            static CallattrIC defaultdict_ic;
+            if (defaultdict_cls == NULL && strcmp(self->cls->tp_name, "collections.defaultdict") == 0) {
+                defaultdict_cls = self->cls;
+            }
+
             static BoxedString* missing_str = internStringImmortal("__missing__");
             CallattrFlags callattr_flags{.cls_only = true, .null_on_nonexistent = true, .argspec = ArgPassSpec(1) };
             Box* r;
             try {
-                r = callattr(self, missing_str, callattr_flags, k, NULL, NULL, NULL, NULL);
+                if (self->cls == defaultdict_cls) {
+                    r = defaultdict_ic.call(self, missing_str, callattr_flags, k, NULL, NULL, NULL, NULL);
+                } else {
+                    r = callattr(self, missing_str, callattr_flags, k, NULL, NULL, NULL, NULL);
+                }
             } catch (ExcInfo e) {
                 if (S == CAPI) {
                     setCAPIException(e);
@@ -603,9 +617,14 @@ void dictMerge(BoxedDict* self, Box* other) {
         return;
     }
 
-    static BoxedString* keys_str = internStringImmortal("keys");
-    CallattrFlags callattr_flags{.cls_only = false, .null_on_nonexistent = true, .argspec = ArgPassSpec(0) };
-    Box* keys = callattr(other, keys_str, callattr_flags, NULL, NULL, NULL, NULL, NULL);
+    Box* keys;
+    if (other->cls == attrwrapper_cls) {
+        keys = attrwrapperKeys(other);
+    } else {
+        static BoxedString* keys_str = internStringImmortal("keys");
+        CallattrFlags callattr_flags{.cls_only = false, .null_on_nonexistent = true, .argspec = ArgPassSpec(0) };
+        keys = callattr(other, keys_str, callattr_flags, NULL, NULL, NULL, NULL, NULL);
+    }
     assert(keys);
 
     for (Box* k : keys->pyElements()) {