Reduce allocations in a few places

- Check for the allocation of empty tuples and just return the singleton
- Try to avoid creating the kwargs dict since it might end up being empty
- Let unicode-creation special case apply to all argument types
- Fix type(obj) to be fast again (got superceded by a different special case)
- Do fewer allocations in int()

from_cpython/Include/unicodeobject.h

@@ -1409,7 +1409,7 @@ PyAPI_FUNC(int) _PyUnicode_IsAlpha(
     ) PYSTON_NOEXCEPT;
 
 // Pyston addition:
-PyAPI_FUNC(PyObject*) unicode_new_inner(PyObject* x, char* encoding, char* errors) PYSTON_NOEXCEPT;
+PyAPI_FUNC(PyObject*) unicode_new_inner(PyTypeObject* type, PyObject* x, char* encoding, char* errors) PYSTON_NOEXCEPT;
 
 #ifdef __cplusplus
 }

from_cpython/Objects/unicodeobject.c

@@ -8745,61 +8745,63 @@ static PyBufferProcs unicode_as_buffer = {
 static PyObject *
 unicode_subtype_new(PyTypeObject *type, PyObject *args, PyObject *kwds);
 
-PyObject* unicode_new_inner(PyObject* x, char* encoding, char* errors) {
+PyObject* unicode_new_inner(PyTypeObject* type, PyObject* x, char* encoding, char* errors) {
+    PyObject* tmp;
     if (x == NULL)
-        return (PyObject *)_PyUnicode_New(0);
-    if (encoding == NULL && errors == NULL)
-        return PyObject_Unicode(x);
+        tmp = (PyObject *)_PyUnicode_New(0);
+    else if (encoding == NULL && errors == NULL)
+        tmp = PyObject_Unicode(x);
     else
-        return PyUnicode_FromEncodedObject(x, encoding, errors);
-}
+        tmp = PyUnicode_FromEncodedObject(x, encoding, errors);
 
-static PyObject *
-unicode_new(PyTypeObject *type, PyObject *args, PyObject *kwds)
-{
-    PyObject *x = NULL;
-    static char *kwlist[] = {"string", "encoding", "errors", 0};
-    char *encoding = NULL;
-    char *errors = NULL;
+    if (type == &PyUnicode_Type)
+        return tmp;
 
-    if (type != &PyUnicode_Type)
-        return unicode_subtype_new(type, args, kwds);
-    if (!PyArg_ParseTupleAndKeywords(args, kwds, "|Oss:unicode",
-                                     kwlist, &x, &encoding, &errors))
+    if (tmp == NULL)
         return NULL;
-    return unicode_new_inner(x, encoding, errors);
-}
 
-static PyObject *
-unicode_subtype_new(PyTypeObject *type, PyObject *args, PyObject *kwds)
-{
-    PyUnicodeObject *tmp, *pnew;
+    // Inlined unicode_subtype_new:
+    PyUnicodeObject *tmp2;
+    PyUnicodeObject *pnew;
     Py_ssize_t n;
 
-    assert(PyType_IsSubtype(type, &PyUnicode_Type));
-    tmp = (PyUnicodeObject *)unicode_new(&PyUnicode_Type, args, kwds);
-    if (tmp == NULL)
-        return NULL;
     assert(PyUnicode_Check(tmp));
-    pnew = (PyUnicodeObject *) type->tp_alloc(type, n = tmp->length);
+    tmp2 = (PyUnicodeObject*)tmp;
+
+    pnew = (PyUnicodeObject *) type->tp_alloc(type, n = tmp2->length);
     if (pnew == NULL) {
-        Py_DECREF(tmp);
+        Py_DECREF(tmp2);
         return NULL;
     }
     pnew->str = (Py_UNICODE*) PyObject_MALLOC(sizeof(Py_UNICODE) * (n+1));
     if (pnew->str == NULL) {
         _Py_ForgetReference((PyObject *)pnew);
         PyObject_Del(pnew);
-        Py_DECREF(tmp);
+        Py_DECREF(tmp2);
         return PyErr_NoMemory();
     }
-    Py_UNICODE_COPY(pnew->str, tmp->str, n+1);
+    Py_UNICODE_COPY(pnew->str, tmp2->str, n+1);
     pnew->length = n;
-    pnew->hash = tmp->hash;
-    Py_DECREF(tmp);
+    pnew->hash = tmp2->hash;
+    Py_DECREF(tmp2);
     return (PyObject *)pnew;
 }
 
+static PyObject *
+unicode_new(PyTypeObject *type, PyObject *args, PyObject *kwds)
+{
+    PyObject *x = NULL;
+    static char *kwlist[] = {"string", "encoding", "errors", 0};
+    char *encoding = NULL;
+    char *errors = NULL;
+
+    if (!PyArg_ParseTupleAndKeywords(args, kwds, "|Oss:unicode",
+                                     kwlist, &x, &encoding, &errors))
+        return NULL;
+
+    return unicode_new_inner(type, x, encoding, errors);
+}
+
 PyDoc_STRVAR(unicode_doc,
              "unicode(object='') -> unicode object\n\
 unicode(string[, encoding[, errors]]) -> unicode object\n\

src/capi/abstract.cpp

@@ -1986,7 +1986,11 @@ extern "C" PyObject* PyNumber_Int(PyObject* o) noexcept {
         return PyInt_FromLong(io->n);
     }
 
-    PyObject* trunc_func = PyObject_GetAttrString(o, "__trunc__");
+    // Pyston change: this should be an optimization
+    // PyObject* trunc_func = PyObject_GetAttrString(o, "__trunc__");
+    static BoxedString* trunc_str = internStringImmortal("__trunc__");
+    PyObject* trunc_func = getattrInternal<ExceptionStyle::CAPI>(o, trunc_str, NULL);
+
     if (trunc_func) {
         PyObject* truncated = PyEval_CallObject(trunc_func, NULL);
         Py_DECREF(trunc_func);

src/core/stats.h

@@ -29,6 +29,7 @@ namespace pyston {
 #define DISABLE_STATS 0
 
 #define STAT_ALLOCATIONS (0 && !DISABLE_STATS)
+#define STAT_ALLOCATION_TYPES (0 && !DISABLE_STATS)
 #define STAT_CALLATTR_DESCR_ABORTS (0 && !DISABLE_STATS)
 #define STAT_EXCEPTIONS (0 && !DISABLE_STATS)
 #define STAT_EXCEPTIONS_LOCATION (0 && STAT_EXCEPTIONS)

src/core/types.h

@@ -581,7 +581,7 @@ extern "C" PyObject* PystonType_GenericAlloc(BoxedClass* cls, Py_ssize_t nitems)
         return Box::operator new(size, default_cls);                                                                   \
     }
 
-#if STAT_ALLOCATIONS
+#if STAT_ALLOCATION_TYPES
 #define ALLOC_STATS(cls)                                                                                               \
     if (cls->tp_name) {                                                                                                \
         std::string per_name_alloc_name = "alloc." + std::string(cls->tp_name);                                        \

src/gc/gc_alloc.h

@@ -112,8 +112,8 @@ extern "C" inline void* gc_alloc(size_t bytes, GCKind kind_id) {
 #endif
 
 #if STAT_ALLOCATIONS
-    gc_alloc_bytes.log(bytes);
-    gc_alloc_bytes_typed[(int)kind_id].log(bytes);
+    gc_alloc_bytes.log(alloc_bytes);
+    gc_alloc_bytes_typed[(int)kind_id].log(alloc_bytes);
 #endif
 
     return r;

src/runtime/int.cpp

@@ -221,6 +221,21 @@ extern "C" PyObject* PyInt_FromString(const char* s, char** pend, int base) noex
     if (*end != '\0') {
     bad:
         slen = strlen(s) < 200 ? strlen(s) : 200;
+
+        // Pyston addition: try to avoid doing the string repr if possible.
+        bool use_fast = true;
+        for (int i = 0; i < slen; i++) {
+            char c = s[i];
+            if (c == '\'' || c == '\\' || c == '\t' || c == '\n' || c == '\r' || c < ' ' || c >= 0x7f) {
+                use_fast = false;
+                break;
+            }
+        }
+        if (use_fast) {
+            PyErr_Format(PyExc_ValueError, "invalid literal for int() with base %d: '%.200s'", base, s);
+            return NULL;
+        }
+
         sobj = PyString_FromStringAndSize(s, slen);
         if (sobj == NULL)
             return NULL;

src/runtime/objmodel.cpp

@@ -3254,7 +3254,11 @@ void rearrangeArguments(ParamReceiveSpec paramspec, const ParamNames* param_name
     ////
     // Second, apply any keywords:
 
-    BoxedDict* okwargs = NULL;
+    // Speed hack: we try to not create the kwargs dictionary if it will end up being empty.
+    // So if we see that we need to pass something, first set it to NULL, and then store the
+    // pointer here so that if we need to we can instantiate the dict and store it here.
+    // If you need to access the dict, you should call get_okwargs()
+    BoxedDict** _okwargs = NULL;
     if (paramspec.takes_kwargs) {
         int kwargs_idx = paramspec.num_args + (paramspec.takes_varargs ? 1 : 0);
         if (rewrite_args) {
@@ -3270,33 +3274,43 @@ void rearrangeArguments(ParamReceiveSpec paramspec, const ParamNames* param_name
                 rewrite_args->args->setAttr((kwargs_idx - 3) * sizeof(Box*), r_kwargs);
         }
 
-        okwargs = new BoxedDict();
-        getArg(kwargs_idx, oarg1, oarg2, oarg3, oargs) = okwargs;
+        _okwargs = (BoxedDict**)&getArg(kwargs_idx, oarg1, oarg2, oarg3, oargs);
+        *_okwargs = NULL;
     }
+    auto get_okwargs = [=]() {
+        if (!paramspec.takes_kwargs)
+            return (BoxedDict*)nullptr;
+        BoxedDict* okw = *_okwargs;
+        if (okw)
+            return okw;
+        okw = *_okwargs = new BoxedDict();
+        return okw;
+    };
 
     if ((!param_names || !param_names->takes_param_names) && argspec.num_keywords && !paramspec.takes_kwargs) {
         raiseExcHelper(TypeError, "%s() doesn't take keyword arguments", func_name);
     }
 
-    if (argspec.num_keywords)
+    if (argspec.num_keywords) {
         assert(argspec.num_keywords == keyword_names->size());
 
-    for (int i = 0; i < argspec.num_keywords; i++) {
-        assert(!rewrite_args && "would need to be handled here");
+        BoxedDict* okwargs = get_okwargs();
+        for (int i = 0; i < argspec.num_keywords; i++) {
+            assert(!rewrite_args && "would need to be handled here");
 
-        int arg_idx = i + argspec.num_args;
-        Box* kw_val = getArg(arg_idx, arg1, arg2, arg3, args);
+            int arg_idx = i + argspec.num_args;
+            Box* kw_val = getArg(arg_idx, arg1, arg2, arg3, args);
 
-        if (!param_names || !param_names->takes_param_names) {
-            assert(okwargs);
-            assert(!rewrite_args); // would need to add it to r_kwargs
-            okwargs->d[(*keyword_names)[i]] = kw_val;
-            continue;
-        }
+            if (!param_names || !param_names->takes_param_names) {
+                assert(!rewrite_args); // would need to add it to r_kwargs
+                okwargs->d[(*keyword_names)[i]] = kw_val;
+                continue;
+            }
 
-        auto dest = placeKeyword(param_names, params_filled, (*keyword_names)[i], kw_val, oarg1, oarg2, oarg3, oargs,
-                                 okwargs, func_name);
-        assert(!rewrite_args);
+            auto dest = placeKeyword(param_names, params_filled, (*keyword_names)[i], kw_val, oarg1, oarg2, oarg3,
+                                     oargs, okwargs, func_name);
+            assert(!rewrite_args);
+        }
     }
 
     if (argspec.has_kwargs) {
@@ -3316,6 +3330,10 @@ void rearrangeArguments(ParamReceiveSpec paramspec, const ParamNames* param_name
         assert(PyDict_Check(kwargs));
         BoxedDict* d_kwargs = static_cast<BoxedDict*>(kwargs);
 
+        BoxedDict* okwargs = NULL;
+        if (d_kwargs->d.size())
+            okwargs = get_okwargs();
+
         for (const auto& p : *d_kwargs) {
             auto k = coerceUnicodeToStr(p.first);
 

src/runtime/types.cpp

@@ -712,7 +712,8 @@ static PyObject* cpythonTypeCall(BoxedClass* type, PyObject* args, PyObject* kwd
 static Box* unicodeNewHelper(BoxedClass* type, Box* string, Box* encoding_obj, Box** _args) {
     Box* errors_obj = _args[0];
 
-    assert(type == unicode_cls);
+    assert(isSubclass(type, unicode_cls));
+    assert(type->tp_new == unicode_cls->tp_new);
 
     char* encoding = NULL;
     char* errors = NULL;
@@ -723,10 +724,44 @@ static Box* unicodeNewHelper(BoxedClass* type, Box* string, Box* encoding_obj, B
         if (!PyArg_ParseSingle(errors_obj, 1, "unicode", "s", &errors))
             throwCAPIException();
 
-    Box* r = unicode_new_inner(string, encoding, errors);
+    Box* r = unicode_new_inner(type, string, encoding, errors);
     if (!r)
         throwCAPIException();
-    assert(r->cls == unicode_cls); // otherwise we'd need to call this object's init
+
+    if (type == unicode_cls) {
+        if (r->cls == unicode_cls) // no init
+            return r;
+        if (!PyUnicode_Check(r)) // skip init
+            return r;
+    } else {
+        if (!isSubclass(r->cls, type)) // skip init
+            return r;
+    }
+
+    // Call tp_init:
+
+    if (r->cls->tp_init == object_cls->tp_init)
+        return r;
+
+    if (!string)
+        assert(!encoding_obj);
+    if (!encoding_obj)
+        assert(!errors_obj);
+
+    Box* args;
+    if (!string)
+        args = EmptyTuple;
+    else if (!encoding_obj)
+        args = BoxedTuple::create1(string);
+    else if (!errors_obj)
+        args = BoxedTuple::create2(string, encoding_obj);
+    else
+        args = BoxedTuple::create3(string, encoding_obj, errors_obj);
+
+    int init_code = r->cls->tp_init(r, args, NULL);
+    if (init_code == -1)
+        throwCAPIException();
+
     return r;
 }
 
@@ -762,19 +797,14 @@ static Box* typeCallInner(CallRewriteArgs* rewrite_args, ArgPassSpec argspec, Bo
 
     BoxedClass* cls = static_cast<BoxedClass*>(_cls);
 
-    if (cls == unicode_cls && !argspec.has_kwargs && !argspec.has_starargs
-        && (argspec.num_args == 1 || (argspec.num_args == 2 && arg2->cls == str_cls)) && S == CXX) {
-        // unicode() takes an "encoding" parameter which can cause the constructor to return unicode subclasses.
-
+    // Special-case unicode for now, maybe there's something about this that can eventually be generalized:
+    if (cls->tp_new == unicode_cls->tp_new) {
         assert(S == CXX && "implement me");
 
         if (rewrite_args) {
             rewrite_args->arg1->addGuard((intptr_t)cls);
-            if (argspec.num_args >= 2)
-                rewrite_args->arg2->addGuard((intptr_t)arg2->cls);
         }
 
-        // Special-case unicode for now, maybe there's something about this that can eventually be generalized:
         ParamReceiveSpec paramspec(4, 3, false, false);
         bool rewrite_success = false;
         static ParamNames param_names({ "string", "encoding", "errors" }, "", "");
@@ -795,12 +825,11 @@ static Box* typeCallInner(CallRewriteArgs* rewrite_args, ArgPassSpec argspec, Bo
             rewrite_args->out_success = true;
         }
 
-        // TODO other encodings could return non-unicode?
         return unicodeNewHelper(cls, arg2, arg3, oargs);
     }
 
     if (cls->tp_new != object_cls->tp_new && cls->tp_new != slot_tp_new && cls->tp_new != BaseException->tp_new
-        && S == CXX) {
+        && cls->tp_new != type_cls->tp_new && S == CXX) {
         // Looks like we're calling an extension class and we're not going to be able to
         // separately rewrite the new + init calls.  But we can rewrite the fact that we
         // should just call the cpython version, which will end up working pretty well.

src/runtime/types.h

@@ -83,6 +83,9 @@ BoxedDict* getSysModulesDict();
 BoxedList* getSysPath();
 extern "C" Box* getSysStdout();
 
+extern "C" BoxedTuple* EmptyTuple;
+extern "C" BoxedString* EmptyString;
+
 extern "C" {
 extern BoxedClass* object_cls, *type_cls, *bool_cls, *int_cls, *long_cls, *float_cls, *str_cls, *function_cls,
     *none_cls, *instancemethod_cls, *list_cls, *slice_cls, *module_cls, *dict_cls, *tuple_cls, *file_cls,
@@ -492,11 +495,15 @@ static_assert(offsetof(BoxedList, capacity) == offsetof(PyListObject, allocated)
 class BoxedTuple : public BoxVar {
 public:
     static BoxedTuple* create(int64_t size) {
+        if (size == 0)
+            return EmptyTuple;
         BoxedTuple* rtn = new (size) BoxedTuple();
         memset(rtn->elts, 0, size * sizeof(Box*)); // TODO not all callers want this (but some do)
         return rtn;
     }
     static BoxedTuple* create(int64_t nelts, Box** elts) {
+        if (nelts == 0)
+            return EmptyTuple;
         BoxedTuple* rtn = new (nelts) BoxedTuple();
         for (int i = 0; i < nelts; i++)
             assert(gc::isValidGCObject(elts[i]));
@@ -643,8 +650,6 @@ static_assert(sizeof(BoxedTuple) == sizeof(PyTupleObject), "");
 static_assert(offsetof(BoxedTuple, ob_size) == offsetof(PyTupleObject, ob_size), "");
 static_assert(offsetof(BoxedTuple, elts) == offsetof(PyTupleObject, ob_item), "");
 
-extern "C" BoxedTuple* EmptyTuple;
-extern "C" BoxedString* EmptyString;
 extern BoxedString* characters[UCHAR_MAX + 1];
 
 struct PyHasher {