Merge remote-tracking branch 'upstream/master' into intern

.travis.yml

@@ -23,6 +23,7 @@ addons:
     - clang-3.5
     - cmake
     - g++-4.8
+    - gdb
     - libgmp3-dev
     - liblzma-dev
     - libncurses5-dev
@@ -51,7 +52,7 @@ script:
   - ccache -z
   - ninja -j4 pyston
   - ccache -s
-  - ninja check-pyston
+  - PYSTON_RUN_ARGS=G ninja check-pyston
 
 os:
   - linux

Makefile

@@ -112,12 +112,25 @@ ifeq ($(NEED_OLD_JIT),1)
 	LLVM_LINK_LIBS += jit
 endif
 
+LLVM_CONFIG_DBG := $(LLVM_BUILD)/Release+Asserts/bin/llvm-config
+ifneq ($(wildcard $(LLVM_CONFIG_DBG)),)
 LLVM_CXXFLAGS := $(shell $(LLVM_BUILD)/Release+Asserts/bin/llvm-config --cxxflags)
 LLVM_LDFLAGS := $(shell $(LLVM_BUILD)/Release+Asserts/bin/llvm-config --ldflags --system-libs --libs $(LLVM_LINK_LIBS))
 LLVM_LIB_DEPS := $(wildcard $(LLVM_BUILD)/Release+Asserts/lib/*)
+else
+LLVM_CXXFLAGS := DBG_NOT_BUILT
+LLVM_LDFLAGS := DBG_NOT_BUILT
+LLVM_LIB_DEPS := DBG_NOT_BUILT
+endif
 
+LLVM_CONFIG_DEBUG := $(LLVM_BUILD)/Debug+Asserts/bin/llvm-config
+ifneq ($(wildcard $(LLVM_CONFIG_DBG)),)
 LLVM_DEBUG_LDFLAGS := $(shell $(LLVM_BUILD)/Debug+Asserts/bin/llvm-config --ldflags --system-libs --libs $(LLVM_LINK_LIBS))
 LLVM_DEBUG_LIB_DEPS := $(wildcard $(LLVM_BUILD)/Debug+Asserts/lib/*)
+else
+LLVM_DEBUG_LDFLAGS := DEBUG_NOT_BUILT
+LLVM_DEBUG_LIB_DEPS := DEBUG_NOT_BUILT
+endif
 
 LLVM_CONFIG_RELEASE := $(LLVM_BUILD)/Release/bin/llvm-config
 ifneq ($(wildcard $(LLVM_CONFIG_RELEASE)),)

from_cpython/Include/pyconfig.h

@@ -27,6 +27,8 @@
 #define SIZEOF_INT 4
 #define SIZEOF_LONG 8
 #define SIZEOF_LONG_LONG 8
+#define SIZEOF_DOUBLE 8
+#define SIZEOF_FLOAT 4
 #define SIZEOF_OFF_T 8
 #define SIZEOF_PTHREAD_T 8
 #define HAVE_COPYSIGN 1
@@ -59,6 +61,9 @@
 #define HAVE_STRUCT_TM_TM_ZONE 1
 #define HAVE_MKTIME 1
 
+#define TIME_WITH_SYS_TIME
+#define HAVE_GETTIMEOFDAY 1
+
 #define PY_FORMAT_LONG_LONG "ll"
 #define PY_FORMAT_SIZE_T "z"
 

from_cpython/Modules/_collectionsmodule.c

@@ -55,7 +55,9 @@ typedef struct BLOCK {
     struct BLOCK *leftlink;
 } block;
 
-#define MAXFREEBLOCKS 10
+// Pyston change: disable free block cache
+// #define MAXFREEBLOCKS 10
+#define MAXFREEBLOCKS 0
 static Py_ssize_t numfreeblocks = 0;
 static block *freeblocks[MAXFREEBLOCKS];
 

minibenchmarks/django-template.py

+from django.template.base import Origin, Template, Context, TemplateDoesNotExist
+from django.conf import settings
+from django.apps import apps
+import time
+
+try:
+    import __pyston__
+    pyston_loaded = True
+except:
+    pyston_loaded = False
+
+template_source = """
+{% extends "admin/base_site.html" %}
+{% load i18n admin_static %}
+
+{% block extrastyle %}{{ block.super }}<link rel="stylesheet" type="text/css" href="{% static "admin/css/dashboard.css" %}" />{% endblock %}
+
+{% block coltype %}colMS{% endblock %}
+
+{% block bodyclass %}{{ block.super }} dashboard{% endblock %}
+
+{% block breadcrumbs %}{% endblock %}
+
+{% block content %}
+<div id="content-main">
+
+{% if app_list %}
+    {% for app in app_list %}
+        <div class="app-{{ app.app_label }} module">
+        <table>
+        <caption>
+            <a href="{{ app.app_url }}" class="section" title="{% blocktrans with name=app.name %}Models in the {{ name }} application{% endblocktrans %}">{{ app.name }}</a>
+        </caption>
+        {% for model in app.models %}
+            <tr class="model-{{ model.object_name|lower }}">
+            {% if model.admin_url %}
+                <th scope="row"><a href="{{ model.admin_url }}">{{ model.name }}</a></th>
+            {% else %}
+                <th scope="row">{{ model.name }}</th>
+            {% endif %}
+
+            {% if model.add_url %}
+                <td><a href="{{ model.add_url }}" class="addlink">{% trans 'Add' %}</a></td>
+            {% else %}
+                <td>&nbsp;</td>
+            {% endif %}
+
+            {% if model.admin_url %}
+                <td><a href="{{ model.admin_url }}" class="changelink">{% trans 'Change' %}</a></td>
+            {% else %}
+                <td>&nbsp;</td>
+            {% endif %}
+            </tr>
+        {% endfor %}
+        </table>
+        </div>
+    {% endfor %}
+{% else %}
+    <p>{% trans "You don't have permission to edit anything." %}</p>
+{% endif %}
+</div>
+{% endblock %}
+
+{% block sidebar %}
+<div id="content-related">
+    <div class="module" id="recent-actions-module">
+        <h2>{% trans 'Recent Actions' %}</h2>
+        <h3>{% trans 'My Actions' %}</h3>
+            {% load log %}
+            {% get_admin_log 10 as admin_log for_user user %}
+            {% if not admin_log %}
+            <p>{% trans 'None available' %}</p>
+            {% else %}
+            <ul class="actionlist">
+            {% for entry in admin_log %}
+            <li class="{% if entry.is_addition %}addlink{% endif %}{% if entry.is_change %}changelink{% endif %}{% if entry.is_deletion %}deletelink{% endif %}">
+                {% if entry.is_deletion or not entry.get_admin_url %}
+                    {{ entry.object_repr }}
+                {% else %}
+                    <a href="{{ entry.get_admin_url }}">{{ entry.object_repr }}</a>
+                {% endif %}
+                <br/>
+                {% if entry.content_type %}
+                    <span class="mini quiet">{% filter capfirst %}{% trans entry.content_type.name %}{% endfilter %}</span>
+                {% else %}
+                    <span class="mini quiet">{% trans 'Unknown content' %}</span>
+                {% endif %}
+            </li>
+            {% endfor %}
+            </ul>
+            {% endif %}
+    </div>
+</div>
+{% endblock %}
+"""
+
+settings.configure()
+
+apps.populate((
+    'django.contrib.admin',
+    'django.contrib.auth',
+    'django.contrib.contenttypes',
+    'django.contrib.sessions',
+    'django.contrib.messages',
+    'django.contrib.staticfiles',
+))
+
+elapsed = 0
+for i in xrange(500):
+    #if pyston_loaded:
+    #    __pyston__.clearStats()
+    start = time.time()
+    template = Template(template_source, None, "admin/index.html")
+    elapsed = time.time() - start
+print "took %4.1fms for last iteration" % (elapsed * 1000.0,)

src/CMakeLists.txt

@@ -30,7 +30,6 @@ add_library(PYSTON_OBJECTS OBJECT ${OPTIONAL_SRCS}
 		capi/codecs.cpp
 		capi/descrobject.cpp
 		capi/errors.cpp
-		capi/float.cpp
 		capi/modsupport.cpp
 		capi/object.cpp
 		capi/typeobject.cpp

src/capi/abstract.cpp

@@ -557,8 +557,29 @@ extern "C" int PyObject_CheckReadBuffer(PyObject* obj) noexcept {
 }
 
 extern "C" int PyObject_AsReadBuffer(PyObject* obj, const void** buffer, Py_ssize_t* buffer_len) noexcept {
-    fatalOrError(PyExc_NotImplementedError, "unimplemented");
-    return -1;
+    PyBufferProcs* pb;
+    void* pp;
+    Py_ssize_t len;
+
+    if (obj == NULL || buffer == NULL || buffer_len == NULL) {
+        null_error();
+        return -1;
+    }
+    pb = obj->cls->tp_as_buffer;
+    if (pb == NULL || pb->bf_getreadbuffer == NULL || pb->bf_getsegcount == NULL) {
+        PyErr_SetString(PyExc_TypeError, "expected a readable buffer object");
+        return -1;
+    }
+    if ((*pb->bf_getsegcount)(obj, NULL) != 1) {
+        PyErr_SetString(PyExc_TypeError, "expected a single-segment buffer object");
+        return -1;
+    }
+    len = (*pb->bf_getreadbuffer)(obj, 0, &pp);
+    if (len < 0)
+        return -1;
+    *buffer = pp;
+    *buffer_len = len;
+    return 0;
 }
 
 static PyObject* call_function_tail(PyObject* callable, PyObject* args) {
@@ -1818,6 +1839,21 @@ extern "C" PyObject* PyNumber_Long(PyObject* o) noexcept {
 }
 
 extern "C" PyObject* PyNumber_Float(PyObject* o) noexcept {
+    if (o == NULL)
+        return null_error();
+
+    if (o->cls == float_cls)
+        return o;
+
+    if (PyInt_Check(o))
+        return boxFloat(((BoxedInt*)o)->n);
+    else if (PyLong_Check(o)) {
+        double result = PyLong_AsDouble(o);
+        if (result == -1.0 && PyErr_Occurred())
+            return NULL;
+        return boxFloat(result);
+    }
+
     fatalOrError(PyExc_NotImplementedError, "unimplemented");
     return nullptr;
 }

src/capi/descrobject.cpp

@@ -18,6 +18,7 @@
 namespace pyston {
 
 Box* BoxedMethodDescriptor::__call__(BoxedMethodDescriptor* self, Box* obj, BoxedTuple* varargs, Box** _args) {
+    STAT_TIMER(t0, "us_timer_boxedmethoddescriptor__call__");
     BoxedDict* kwargs = static_cast<BoxedDict*>(_args[0]);
 
     assert(self->cls == method_cls);

src/capi/typeobject.cpp

@@ -58,6 +58,7 @@ static int hackcheck(PyObject* self, setattrofunc func, const char* what) noexce
 }
 
 static PyObject* wrap_setattr(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_setattr");
     setattrofunc func = (setattrofunc)wrapped;
     int res;
     PyObject* name, *value;
@@ -74,6 +75,7 @@ static PyObject* wrap_setattr(PyObject* self, PyObject* args, void* wrapped) noe
 }
 
 static PyObject* wrap_delattr(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_delattr");
     setattrofunc func = (setattrofunc)wrapped;
     int res;
     PyObject* name;
@@ -91,6 +93,7 @@ static PyObject* wrap_delattr(PyObject* self, PyObject* args, void* wrapped) noe
 }
 
 static PyObject* wrap_hashfunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_hashfunc");
     hashfunc func = (hashfunc)wrapped;
     long res;
 
@@ -103,12 +106,14 @@ static PyObject* wrap_hashfunc(PyObject* self, PyObject* args, void* wrapped) no
 }
 
 static PyObject* wrap_call(PyObject* self, PyObject* args, void* wrapped, PyObject* kwds) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_call");
     ternaryfunc func = (ternaryfunc)wrapped;
 
     return (*func)(self, args, kwds);
 }
 
 static PyObject* wrap_richcmpfunc(PyObject* self, PyObject* args, void* wrapped, int op) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_richcmpfunc");
     richcmpfunc func = (richcmpfunc)wrapped;
     PyObject* other;
 
@@ -132,6 +137,7 @@ RICHCMP_WRAPPER(gt, Py_GT)
 RICHCMP_WRAPPER(ge, Py_GE)
 
 static PyObject* wrap_next(PyObject* self, PyObject* args, void* wrapped) {
+    STAT_TIMER(t0, "us_timer_wrap_next");
     unaryfunc func = (unaryfunc)wrapped;
     PyObject* res;
 
@@ -144,6 +150,7 @@ static PyObject* wrap_next(PyObject* self, PyObject* args, void* wrapped) {
 }
 
 static PyObject* wrap_descr_get(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_descr_get");
     descrgetfunc func = (descrgetfunc)wrapped;
     PyObject* obj;
     PyObject* type = NULL;
@@ -162,6 +169,7 @@ static PyObject* wrap_descr_get(PyObject* self, PyObject* args, void* wrapped) n
 }
 
 static PyObject* wrap_coercefunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_coercefunc");
     coercion func = (coercion)wrapped;
     PyObject* other, *res;
     int ok;
@@ -188,6 +196,7 @@ static PyObject* wrap_coercefunc(PyObject* self, PyObject* args, void* wrapped)
 }
 
 static PyObject* wrap_ternaryfunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_ternaryfunc");
     ternaryfunc func = (ternaryfunc)wrapped;
     PyObject* other;
     PyObject* third = Py_None;
@@ -200,6 +209,7 @@ static PyObject* wrap_ternaryfunc(PyObject* self, PyObject* args, void* wrapped)
 }
 
 static PyObject* wrap_ternaryfunc_r(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_ternaryfunc_r");
     ternaryfunc func = (ternaryfunc)wrapped;
     PyObject* other;
     PyObject* third = Py_None;
@@ -212,6 +222,7 @@ static PyObject* wrap_ternaryfunc_r(PyObject* self, PyObject* args, void* wrappe
 }
 
 static PyObject* wrap_unaryfunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_unaryfunc");
     unaryfunc func = (unaryfunc)wrapped;
 
     if (!check_num_args(args, 0))
@@ -220,6 +231,7 @@ static PyObject* wrap_unaryfunc(PyObject* self, PyObject* args, void* wrapped) n
 }
 
 static PyObject* wrap_inquirypred(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_inquirypred");
     inquiry func = (inquiry)wrapped;
     int res;
 
@@ -232,6 +244,7 @@ static PyObject* wrap_inquirypred(PyObject* self, PyObject* args, void* wrapped)
 }
 
 static PyObject* wrapInquirypred(PyObject* self, PyObject* args, void* wrapped) {
+    STAT_TIMER(t0, "us_timer_wrapInquirypred");
     inquiry func = (inquiry)wrapped;
     int res;
 
@@ -244,6 +257,7 @@ static PyObject* wrapInquirypred(PyObject* self, PyObject* args, void* wrapped)
 }
 
 static PyObject* wrap_binaryfunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_binaryfunc");
     binaryfunc func = (binaryfunc)wrapped;
     PyObject* other;
 
@@ -254,6 +268,7 @@ static PyObject* wrap_binaryfunc(PyObject* self, PyObject* args, void* wrapped)
 }
 
 static PyObject* wrap_binaryfunc_l(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_binaryfunc_l");
     binaryfunc func = (binaryfunc)wrapped;
     PyObject* other;
 
@@ -268,6 +283,7 @@ static PyObject* wrap_binaryfunc_l(PyObject* self, PyObject* args, void* wrapped
 }
 
 static PyObject* wrap_binaryfunc_r(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_binaryfunc_r");
     binaryfunc func = (binaryfunc)wrapped;
     PyObject* other;
 
@@ -300,6 +316,7 @@ static Py_ssize_t getindex(PyObject* self, PyObject* arg) noexcept {
 }
 
 static PyObject* wrap_lenfunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_lenfunc");
     lenfunc func = (lenfunc)wrapped;
     Py_ssize_t res;
 
@@ -312,6 +329,7 @@ static PyObject* wrap_lenfunc(PyObject* self, PyObject* args, void* wrapped) noe
 }
 
 static PyObject* wrap_indexargfunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_indexargfunc");
     ssizeargfunc func = (ssizeargfunc)wrapped;
     PyObject* o;
     Py_ssize_t i;
@@ -325,6 +343,7 @@ static PyObject* wrap_indexargfunc(PyObject* self, PyObject* args, void* wrapped
 }
 
 static PyObject* wrap_sq_item(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_sq_item");
     ssizeargfunc func = (ssizeargfunc)wrapped;
     PyObject* arg;
     Py_ssize_t i;
@@ -342,6 +361,7 @@ static PyObject* wrap_sq_item(PyObject* self, PyObject* args, void* wrapped) noe
 }
 
 static PyObject* wrap_ssizessizeargfunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_ssizessizeargfunc");
     ssizessizeargfunc func = (ssizessizeargfunc)wrapped;
     Py_ssize_t i, j;
 
@@ -351,6 +371,7 @@ static PyObject* wrap_ssizessizeargfunc(PyObject* self, PyObject* args, void* wr
 }
 
 static PyObject* wrap_sq_setitem(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_sq_setitem");
     ssizeobjargproc func = (ssizeobjargproc)wrapped;
     Py_ssize_t i;
     int res;
@@ -369,6 +390,7 @@ static PyObject* wrap_sq_setitem(PyObject* self, PyObject* args, void* wrapped)
 }
 
 static PyObject* wrap_sq_delitem(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_sq_delitem");
     ssizeobjargproc func = (ssizeobjargproc)wrapped;
     Py_ssize_t i;
     int res;
@@ -388,6 +410,7 @@ static PyObject* wrap_sq_delitem(PyObject* self, PyObject* args, void* wrapped)
 }
 
 static PyObject* wrap_ssizessizeobjargproc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_ssizessizeobjargproc");
     ssizessizeobjargproc func = (ssizessizeobjargproc)wrapped;
     Py_ssize_t i, j;
     int res;
@@ -403,6 +426,7 @@ static PyObject* wrap_ssizessizeobjargproc(PyObject* self, PyObject* args, void*
 }
 
 static PyObject* wrap_delslice(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_delslice");
     ssizessizeobjargproc func = (ssizessizeobjargproc)wrapped;
     Py_ssize_t i, j;
     int res;
@@ -418,6 +442,7 @@ static PyObject* wrap_delslice(PyObject* self, PyObject* args, void* wrapped) no
 
 /* XXX objobjproc is a misnomer; should be objargpred */
 static PyObject* wrap_objobjproc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_objobjproc");
     objobjproc func = (objobjproc)wrapped;
     int res;
     PyObject* value;
@@ -433,6 +458,7 @@ static PyObject* wrap_objobjproc(PyObject* self, PyObject* args, void* wrapped)
 }
 
 static PyObject* wrap_objobjargproc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_objobjargproc");
     objobjargproc func = (objobjargproc)wrapped;
     int res;
     PyObject* key, *value;
@@ -447,6 +473,7 @@ static PyObject* wrap_objobjargproc(PyObject* self, PyObject* args, void* wrappe
 }
 
 static PyObject* wrap_delitem(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_delitem");
     objobjargproc func = (objobjargproc)wrapped;
     int res;
     PyObject* key;
@@ -462,6 +489,7 @@ static PyObject* wrap_delitem(PyObject* self, PyObject* args, void* wrapped) noe
 }
 
 static PyObject* wrap_cmpfunc(PyObject* self, PyObject* args, void* wrapped) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_cmpfunc");
     cmpfunc func = (cmpfunc)wrapped;
     int res;
     PyObject* other;
@@ -482,6 +510,7 @@ static PyObject* wrap_cmpfunc(PyObject* self, PyObject* args, void* wrapped) noe
 
 
 static PyObject* wrap_init(PyObject* self, PyObject* args, void* wrapped, PyObject* kwds) noexcept {
+    STAT_TIMER(t0, "us_timer_wrap_init");
     initproc func = (initproc)wrapped;
 
     if (func(self, args, kwds) < 0)

src/capi/types.h

@@ -54,6 +54,7 @@ public:
     }
 
     static Box* __call__(BoxedCApiFunction* self, BoxedTuple* varargs, BoxedDict* kwargs) {
+        STAT_TIMER(t0, "us_timer_boxedcapifunction__call__");
         assert(self->cls == capifunc_cls);
         assert(varargs->cls == tuple_cls);
         assert(kwargs->cls == dict_cls);
@@ -111,6 +112,8 @@ public:
     DEFAULT_CLASS(wrapperobject_cls);
 
     static Box* __call__(BoxedWrapperObject* self, Box* args, Box* kwds) {
+        STAT_TIMER(t0, "us_timer_boxedwrapperobject__call__");
+
         assert(self->cls == wrapperobject_cls);
         assert(args->cls == tuple_cls);
         assert(kwds->cls == dict_cls);

src/codegen/ast_interpreter.cpp

@@ -284,6 +284,8 @@ public:
 Value ASTInterpreter::execute(ASTInterpreter& interpreter, CFGBlock* start_block, AST_stmt* start_at) {
     threading::allowGLReadPreemption();
 
+    STAT_TIMER(t0, "us_timer_astinterpreter_execute");
+
     void* frame_addr = __builtin_frame_address(0);
     RegisterHelper frame_registerer(&interpreter, frame_addr);
 
@@ -340,12 +342,7 @@ Value ASTInterpreter::doBinOp(Box* left, Box* right, int op, BinExpType exp_type
 void ASTInterpreter::doStore(InternedString name, Value value) {
     ScopeInfo::VarScopeType vst = scope_info->getScopeTypeOfName(name);
     if (vst == ScopeInfo::VarScopeType::GLOBAL) {
-        if (globals->cls == module_cls) {
-            setattr(static_cast<BoxedModule*>(globals), name.c_str(), value.o);
-        } else {
-            assert(globals->cls == dict_cls);
-            static_cast<BoxedDict*>(globals)->d[boxString(name.str())] = value.o;
-        }
+        setGlobal(globals, name, value.o);
     } else if (vst == ScopeInfo::VarScopeType::NAME) {
         assert(frame_info.boxedLocals != NULL);
         // TODO should probably pre-box the names when it's a scope that usesNameLookup
@@ -540,6 +537,7 @@ Value ASTInterpreter::visit_jump(AST_Jump* node) {
                 arg_array.push_back(it.second);
             }
 
+            STAT_TIMER(t0, "us_timer_astinterpreter_jump_osrexit");
             CompiledFunction* partial_func = compilePartialFuncInternal(&exit);
             auto arg_tuple = getTupleFromArgsArray(&arg_array[0], arg_array.size());
             Box* r = partial_func->call(std::get<0>(arg_tuple), std::get<1>(arg_tuple), std::get<2>(arg_tuple),
@@ -876,7 +874,6 @@ Value ASTInterpreter::visit_delete(AST_Delete* node) {
 
                 ScopeInfo::VarScopeType vst = scope_info->getScopeTypeOfName(target->id);
                 if (vst == ScopeInfo::VarScopeType::GLOBAL) {
-                    // Can't use delattr since the errors are different:
                     delGlobal(globals, &target->id.str());
                     continue;
                 } else if (vst == ScopeInfo::VarScopeType::NAME) {
@@ -927,6 +924,8 @@ Value ASTInterpreter::visit_print(AST_Print* node) {
     static const std::string newline_str("\n");
     static const std::string space_str(" ");
 
+    STAT_TIMER(t0, "us_timer_visit_print");
+
     Box* dest = node->dest ? visit_expr(node->dest).o : getSysStdout();
     int nvals = node->values.size();
     assert(nvals <= 1 && "cfg should have lowered it to 0 or 1 values");

src/codegen/entry.cpp

@@ -362,7 +362,7 @@ static void handle_sigint(int signum) {
     // For now, just call abort(), so that we get a traceback at least.
     fprintf(stderr, "SIGINT!\n");
     joinRuntime();
-    Stats::dump();
+    Stats::dump(false);
     abort();
 }
 

src/codegen/irgen/hooks.cpp

@@ -182,6 +182,7 @@ static void compileIR(CompiledFunction* cf, EffortLevel effort) {
 // The codegen_lock needs to be held in W mode before calling this function:
 CompiledFunction* compileFunction(CLFunction* f, FunctionSpecialization* spec, EffortLevel effort,
                                   const OSREntryDescriptor* entry_descriptor) {
+    STAT_TIMER(t0, "us_timer_compileFunction");
     Timer _t("for compileFunction()", 1000);
 
     assert((entry_descriptor != NULL) + (spec != NULL) == 1);
@@ -315,10 +316,13 @@ void compileAndRunModule(AST_Module* m, BoxedModule* bm) {
         assert(cf->clfunc->versions.size());
     }
 
-    if (cf->is_interpreted)
+    if (cf->is_interpreted) {
+        STAT_TIMER(t0, "us_timer_interpreted_module_toplevel");
         astInterpretFunction(cf, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL);
-    else
+    } else {
+        STAT_TIMER(t1, "us_timer_jitted_module_toplevel");
         ((void (*)())cf->code)();
+    }
 }
 
 Box* evalOrExec(CLFunction* cl, Box* globals, Box* boxedLocals) {

src/codegen/parser.cpp

@@ -982,6 +982,7 @@ AST_Module* parse_string(const char* code) {
 }
 
 AST_Module* parse_file(const char* fn) {
+    STAT_TIMER(t0, "us_timer_cpyton_parsing");
     Timer _t("parsing");
 
     if (ENABLE_PYPA_PARSER) {
@@ -1071,9 +1072,10 @@ static ParseResult _reparse(const char* fn, const std::string& cache_fn, AST_Mod
 // it's not a huge deal right now, but this caching version can significantly cut down
 // on the startup time (40ms -> 10ms).
 AST_Module* caching_parse_file(const char* fn) {
+    STAT_TIMER(t0, "us_timer_caching_parse_file");
     static StatCounter us_parsing("us_parsing");
     Timer _t("parsing");
-    _t.setExitCallback([](long t) { us_parsing.log(t); });
+    _t.setExitCallback([](uint64_t t) { us_parsing.log(t); });
 
     int code;
     std::string cache_fn = std::string(fn) + "c";

src/codegen/unwinding.cpp

@@ -494,6 +494,7 @@ void unwindPythonStack(std::function<bool(std::unique_ptr<PythonFrameIteratorImp
 }
 
 static std::unique_ptr<PythonFrameIteratorImpl> getTopPythonFrame() {
+    STAT_TIMER(t0, "us_timer_getTopPythonFrame");
     std::unique_ptr<PythonFrameIteratorImpl> rtn(nullptr);
     unwindPythonStack([&](std::unique_ptr<PythonFrameIteratorImpl> iter) {
         rtn = std::move(iter);
@@ -514,6 +515,7 @@ static const LineInfo* lineInfoForFrame(PythonFrameIteratorImpl& frame_it) {
 
 static StatCounter us_gettraceback("us_gettraceback");
 BoxedTraceback* getTraceback() {
+    STAT_TIMER(t0, "us_timer_gettraceback");
     if (!ENABLE_FRAME_INTROSPECTION) {
         static bool printed_warning = false;
         if (!printed_warning) {
@@ -597,6 +599,8 @@ CompiledFunction* getTopCompiledFunction() {
 
 Box* getGlobals() {
     auto it = getTopPythonFrame();
+    if (!it)
+        return NULL;
     return it->getGlobals();
 }
 

src/core/stats.cpp

@@ -22,9 +22,105 @@
 namespace pyston {
 
 #if !DISABLE_STATS
-std::vector<long>* Stats::counts;
+#if STAT_TIMERS
+extern "C" const char* getStatTimerNameById(int id) {
+    return Stats::getStatName(id).c_str();
+}
+
+extern "C" int getStatTimerId() {
+    return StatTimer::getStack()->getId();
+}
+
+extern "C" const char* getStatTimerName() {
+    return getStatTimerNameById(getStatTimerId());
+}
+
+__thread StatTimer* StatTimer::stack;
+StatTimer::StatTimer(int statid, bool push) {
+    uint64_t at_time = getCPUTicks();
+    _start_time = 0;
+    _statid = statid;
+
+    if (!push) {
+        _prev = NULL;
+        return;
+    }
+
+    _prev = stack;
+    stack = this;
+    if (_prev) {
+        _prev->pause(at_time);
+    }
+    resume(at_time);
+}
+
+StatTimer::StatTimer(int statid, uint64_t at_time) {
+    _start_time = 0;
+    _statid = statid;
+    _prev = stack;
+    stack = this;
+    if (_prev) {
+        _prev->pause(at_time);
+    }
+    resume(at_time);
+}
+
+StatTimer::~StatTimer() {
+    assert(stack == this);
+
+    uint64_t at_time;
+    if (!isPaused()) {
+        at_time = getCPUTicks();
+        pause(at_time);
+    } else {
+        // fprintf (stderr, "WARNING: timer was paused.\n");
+        at_time = _last_pause_time;
+    }
+
+    stack = _prev;
+    if (stack) {
+        stack->resume(at_time);
+    }
+}
+
+void StatTimer::pause(uint64_t at_time) {
+    assert(!isPaused());
+    assert(at_time > _start_time);
+
+    uint64_t _duration = at_time - _start_time;
+    Stats::log(_statid, _duration);
+
+    _start_time = 0;
+    _last_pause_time = at_time;
+    // fprintf (stderr, "paused  %d at %lu\n", _statid, at_time);
+}
+
+void StatTimer::resume(uint64_t at_time) {
+    assert(isPaused());
+    _start_time = at_time;
+    // fprintf (stderr, "resumed %d at %lu\n", _statid, at_time);
+}
+
+StatTimer* StatTimer::swapStack(StatTimer* s, uint64_t at_time) {
+    StatTimer* prev_stack = stack;
+    if (stack) {
+        stack->pause(at_time);
+    }
+    stack = s;
+    if (stack) {
+        stack->resume(at_time);
+    }
+    return prev_stack;
+}
+#endif
+
+std::vector<uint64_t>* Stats::counts;
 std::unordered_map<int, std::string>* Stats::names;
 bool Stats::enabled;
+
+timespec Stats::start_ts;
+uint64_t Stats::start_tick;
+
 StatCounter::StatCounter(const std::string& name) : id(Stats::getStatId(name)) {
 }
 
@@ -38,7 +134,7 @@ int Stats::getStatId(const std::string& name) {
     // hacky but easy way of getting around static constructor ordering issues for now:
     static std::unordered_map<int, std::string> names;
     Stats::names = &names;
-    static std::vector<long> counts;
+    static std::vector<uint64_t> counts;
     Stats::counts = &counts;
     static std::unordered_map<std::string, int> made;
 
@@ -52,11 +148,36 @@ int Stats::getStatId(const std::string& name) {
     return rtn;
 }
 
+std::string Stats::getStatName(int id) {
+    return (*names)[id];
+}
+
+void Stats::startEstimatingCPUFreq() {
+    if (!Stats::enabled)
+        return;
+
+    clock_gettime(CLOCK_REALTIME, &Stats::start_ts);
+    Stats::start_tick = getCPUTicks();
+}
+
+// returns our estimate of the MHz of the cpu.  MHz is handy because we're mostly interested in microsoecond-resolution
+// timing.
+double Stats::estimateCPUFreq() {
+    timespec dump_ts;
+    clock_gettime(CLOCK_REALTIME, &dump_ts);
+    uint64_t end_tick = getCPUTicks();
+
+    uint64_t wall_clock_ns = (dump_ts.tv_sec - start_ts.tv_sec) * 1000000000 + (dump_ts.tv_nsec - start_ts.tv_nsec);
+    return (double)(end_tick - Stats::start_tick) * 1000 / wall_clock_ns;
+}
+
 void Stats::dump(bool includeZeros) {
     if (!Stats::enabled)
         return;
 
+    double cycles_per_us = Stats::estimateCPUFreq();
     fprintf(stderr, "Stats:\n");
+    fprintf(stderr, "estimated_cpu_mhz: %5.5f\n", cycles_per_us);
 
     gc::dumpHeapStatistics(0);
 
@@ -69,11 +190,41 @@ void Stats::dump(bool includeZeros) {
 
     std::sort(pairs.begin(), pairs.end());
 
+    uint64_t ticks_in_main = 0;
+    uint64_t accumulated_stat_timer_ticks = 0;
     for (int i = 0; i < pairs.size(); i++) {
-        if (includeZeros || (*counts)[pairs[i].second] > 0)
-            fprintf(stderr, "%s: %ld\n", pairs[i].first.c_str(), (*counts)[pairs[i].second]);
+        if (includeZeros || (*counts)[pairs[i].second] > 0) {
+            if (startswith(pairs[i].first, "us_") || startswith(pairs[i].first, "_init_us_")) {
+                fprintf(stderr, "%s: %lu\n", pairs[i].first.c_str(),
+                        (uint64_t)((*counts)[pairs[i].second] / cycles_per_us));
+
+            } else
+                fprintf(stderr, "%s: %lu\n", pairs[i].first.c_str(), (*counts)[pairs[i].second]);
+
+            if (startswith(pairs[i].first, "us_timer_"))
+                accumulated_stat_timer_ticks += (*counts)[pairs[i].second];
+
+            if (pairs[i].first == "ticks_in_main")
+                ticks_in_main = (*counts)[pairs[i].second];
+        }
     }
 
+    if (includeZeros || accumulated_stat_timer_ticks > 0)
+        fprintf(stderr, "ticks_all_timers: %lu\n", accumulated_stat_timer_ticks);
+
+#if 0
+    // I want to enable this, but am leaving it disabled for the time
+    // being because it causes test failures due to:
+    //
+    // 1) some tests exit from main from inside catch blocks, without
+    //    going through the logic to stop the timers.
+    // 2) some tests create multiple threads which causes problems
+    //    with our non-per thread stat timers.
+
+    if (ticks_in_main && ticks_in_main != accumulated_stat_timer_ticks) {
+        fprintf(stderr, "WARNING: accumulated stat timer ticks != ticks in main - don't trust timer output.");
+    }
+#endif
     fprintf(stderr, "(End of stats)\n");
 }
 

src/core/stats.h

@@ -22,23 +22,48 @@
 #include <vector>
 
 #include "core/options.h"
+#include "core/util.h"
 
 namespace pyston {
 
 #define DISABLE_STATS 0
 
+#define STAT_ALLOCATIONS 0 && !DISABLE_STATS
+#define STAT_TIMERS 0 && !DISABLE_STATS
+
+#if STAT_TIMERS
+#define STAT_TIMER(id, name)                                                                                           \
+    static int _stid##id = Stats::getStatId(name);                                                                     \
+    StatTimer _st##id(_stid##id)
+#define STAT_TIMER2(id, name, at_time)                                                                                 \
+    static int _stid##id = Stats::getStatId(name);                                                                     \
+    StatTimer _st##id(_stid##id, at_time)
+#else
+#define STAT_TIMER(id, name) StatTimer _st##id(0);
+#define STAT_TIMER2(id, name, at_time) StatTimer _st##id(0);
+#endif
+
+#define STAT_TIMER_NAME(id) _st##id
+
 #if !DISABLE_STATS
 struct Stats {
 private:
-    static std::vector<long>* counts;
+    static std::vector<uint64_t>* counts;
     static std::unordered_map<int, std::string>* names;
     static bool enabled;
 
+    static timespec start_ts;
+    static uint64_t start_tick;
+
 public:
+    static void startEstimatingCPUFreq();
+    static double estimateCPUFreq();
+
     static int getStatId(const std::string& name);
+    static std::string getStatName(int id);
 
     static void setEnabled(bool enabled) { Stats::enabled = enabled; }
-    static void log(int id, int count = 1) { (*counts)[id] += count; }
+    static void log(int id, uint64_t count = 1) { (*counts)[id] += count; }
 
     static void clear() { std::fill(counts->begin(), counts->end(), 0); }
     static void dump(bool includeZeros = true);
@@ -52,7 +77,7 @@ private:
 public:
     StatCounter(const std::string& name);
 
-    void log(int count = 1) { Stats::log(id, count); }
+    void log(uint64_t count = 1) { Stats::log(id, count); }
 };
 
 struct StatPerThreadCounter {
@@ -62,24 +87,74 @@ private:
 public:
     StatPerThreadCounter(const std::string& name);
 
-    void log(int count = 1) { Stats::log(id, count); }
+    void log(uint64_t count = 1) { Stats::log(id, count); }
 };
 
 #else
 struct Stats {
+    static void startEstimatingCPUFreq() {}
+    static double estimateCPUFreq() { return 0; }
     static void setEnabled(bool enabled) {}
-    static void dump() { printf("(Stats disabled)\n"); }
+    static void dump(bool includeZeros = true) { printf("(Stats disabled)\n"); }
+    static void clear() {}
     static void log(int id, int count = 1) {}
     static int getStatId(const std::string& name) { return 0; }
     static void endOfInit() {}
 };
 struct StatCounter {
     StatCounter(const char* name) {}
-    void log(int count = 1){};
+    void log(uint64_t count = 1){};
 };
 struct StatPerThreadCounter {
     StatPerThreadCounter(const char* name) {}
-    void log(int count = 1){};
+    void log(uint64_t count = 1){};
+};
+#endif
+
+#if STAT_TIMERS
+class StatTimer {
+private:
+    static __thread StatTimer* stack;
+
+    // the start time of the current active segment (0 == paused)
+    uint64_t _start_time;
+
+    uint64_t _last_pause_time;
+
+    StatTimer* _prev;
+
+    int _statid;
+
+public:
+    StatTimer(int statid, bool push = true);
+    StatTimer(int statid, uint64_t at_time);
+    ~StatTimer();
+
+    void pause(uint64_t at_time);
+    void resume(uint64_t at_time);
+
+    bool isPaused() const { return _start_time == 0; }
+    int getId() const { return _statid; }
+
+    static StatTimer* getStack() { return stack; }
+
+    static StatTimer* swapStack(StatTimer* s, uint64_t at_time);
+
+    static void assertActive() { RELEASE_ASSERT(stack && !stack->isPaused(), ""); }
+};
+#else
+struct StatTimer {
+    StatTimer(int statid, bool push = true) {}
+    StatTimer(int statid, uint64_t at_time) {}
+    ~StatTimer() {}
+    bool isPaused() const { return false; }
+
+    void pause(uint64_t at_time) {}
+    void resume(uint64_t at_time) {}
+
+    static StatTimer* getStack() { return NULL; }
+    static StatTimer* swapStack(StatTimer* s, uint64_t at_time) { return NULL; }
+    static void assertActive() {}
 };
 #endif
 }

src/core/stringpool.h

@@ -80,6 +80,8 @@ public:
         return *this->_str < *rhs._str;
     }
 
+    operator llvm::StringRef() { return llvm::StringRef(*_str); }
+
     friend class InternedStringPool;
     friend struct std::hash<InternedString>;
     friend struct std::less<InternedString>;

src/core/types.h

@@ -485,13 +485,26 @@ extern "C" PyObject* PystonType_GenericAlloc(BoxedClass* cls, Py_ssize_t nitems)
         return Box::operator new(size, default_cls);                                                                   \
     }
 
+#if STAT_ALLOCATIONS
+#define ALLOC_STATS(cls)                                                                                               \
+    std::string per_name_alloc_name = "alloc." + std::string(cls->tp_name);                                            \
+    std::string per_name_allocsize_name = "allocsize." + std::string(cls->tp_name);                                    \
+    Stats::log(Stats::getStatId(per_name_alloc_name));                                                                 \
+    Stats::log(Stats::getStatId(per_name_allocsize_name), size);
+#else
+#define ALLOC_STATS(cls)
+#endif
+
+
 // The restrictions on when you can use the SIMPLE (ie fast) variant are encoded as
 // asserts in the 1-arg operator new function:
 #define DEFAULT_CLASS_SIMPLE(default_cls)                                                                              \
     void* operator new(size_t size, BoxedClass * cls) __attribute__((visibility("default"))) {                         \
+        ALLOC_STATS(cls);                                                                                              \
         return Box::operator new(size, cls);                                                                           \
     }                                                                                                                  \
     void* operator new(size_t size) __attribute__((visibility("default"))) {                                           \
+        ALLOC_STATS(default_cls);                                                                                      \
         /* In the simple cases, we can inline the following methods and simplify things a lot:                         \
          * - Box::operator new                                                                                         \
          * - cls->tp_alloc                                                                                             \

src/core/util.cpp

@@ -16,6 +16,9 @@
 
 #include <cstdio>
 #include <cstdlib>
+#include <fstream>
+#include <iostream>
+#include <string>
 
 #include "llvm/Support/FileSystem.h"
 #include "llvm/Support/FormattedStream.h"
@@ -26,6 +29,18 @@
 
 namespace pyston {
 
+static inline uint64_t rdtsc() {
+    unsigned long lo, hi;
+    asm("rdtsc" : "=a"(lo), "=d"(hi));
+    return (lo | (hi << 32));
+}
+
+uint64_t getCPUTicks() {
+    // unsigned int _unused;
+    // return __rdtscp(&_unused);
+    return rdtsc();
+}
+
 int Timer::level = 0;
 
 Timer::Timer(const char* desc, long min_usec) : min_usec(min_usec), ended(true) {
@@ -36,7 +51,7 @@ void Timer::restart(const char* newdesc) {
     assert(ended);
 
     desc = newdesc;
-    gettimeofday(&start_time, NULL);
+    start_time = getCPUTicks();
     Timer::level++;
     ended = false;
 }
@@ -46,14 +61,14 @@ void Timer::restart(const char* newdesc, long new_min_usec) {
     restart(newdesc);
 }
 
-long Timer::end() {
+uint64_t Timer::end(uint64_t* ended_at) {
     if (!ended) {
-        timeval end;
-        gettimeofday(&end, NULL);
-        long us = 1000000L * (end.tv_sec - start_time.tv_sec) + (end.tv_usec - start_time.tv_usec);
-
+        uint64_t end = getCPUTicks();
+        uint64_t duration = end - start_time;
         Timer::level--;
         if (VERBOSITY("time") >= 2 && desc) {
+            uint64_t us = (uint64_t)(duration / Stats::estimateCPUFreq());
+
             if (us > min_usec) {
                 for (int i = 0; i < Timer::level; i++) {
                     putchar(' ');
@@ -70,15 +85,17 @@ long Timer::end() {
                 fflush(stdout);
             }
         }
+        if (ended_at)
+            *ended_at = end;
         ended = true;
-        return us;
+        return duration;
     }
     return -1;
 }
 
 Timer::~Timer() {
     if (!ended) {
-        long t = end();
+        uint64_t t = end();
         if (exit_callback)
             exit_callback(t);
     }

src/core/util.h

@@ -24,30 +24,36 @@
 
 namespace pyston {
 
+uint64_t getCPUTicks();
+
 class Timer {
 private:
     static int level;
-    timeval start_time;
+    uint64_t start_time;
     const char* desc;
     long min_usec;
     bool ended;
-    std::function<void(long)> exit_callback;
+    std::function<void(uint64_t)> exit_callback;
 
 public:
     Timer(const char* desc = NULL, long min_usec = -1);
     ~Timer();
 
-    void setExitCallback(std::function<void(long)> _exit_callback) { exit_callback = _exit_callback; }
+    void setExitCallback(std::function<void(uint64_t)> _exit_callback) { exit_callback = _exit_callback; }
 
     void restart(const char* newdesc, long new_min_usec);
     void restart(const char* newdesc = NULL);
 
-    long end();
-    long split(const char* newdesc = NULL) {
-        long rtn = end();
+    // returns the duration.  if @ended_at is non-null, it's filled in
+    // with the tick the timer stopped at.
+    uint64_t end(uint64_t* ended_at = NULL);
+    uint64_t split(const char* newdesc = NULL) {
+        uint64_t rtn = end();
         restart(newdesc);
         return rtn;
     }
+
+    uint64_t getStartTime() const { return start_time; }
 };
 
 bool startswith(const std::string& s, const std::string& pattern);

src/gc/collector.cpp

@@ -339,6 +339,8 @@ void runCollection() {
     static StatCounter sc("gc_collections");
     sc.log();
 
+    STAT_TIMER(t0, "us_timer_gc_collection");
+
     ncollections++;
 
     if (VERBOSITY("gc") >= 2)

src/gc/gc_alloc.h

@@ -28,7 +28,9 @@ namespace pyston {
 
 namespace gc {
 
+static StatCounter gc_alloc_bytes("zzz_gc_alloc_bytes");
 extern "C" inline void* gc_alloc(size_t bytes, GCKind kind_id) {
+    STAT_TIMER(t0, "us_timer_gc_alloc");
     size_t alloc_bytes = bytes + sizeof(GCAllocation);
 
 #ifndef NVALGRIND
@@ -89,8 +91,10 @@ extern "C" inline void* gc_alloc(size_t bytes, GCKind kind_id) {
 
 // if (VERBOSITY()) printf("Allocated %ld bytes at [%p, %p)\n", bytes, r, (char*)r + bytes);
 #endif
-    // printf("Allocated %p\n", r);
 
+#if STAT_ALLOCATIONS
+    gc_alloc_bytes.log(bytes);
+#endif
 
     return r;
 }
@@ -128,6 +132,10 @@ extern "C" inline void* gc_realloc(void* ptr, size_t bytes) {
         alloc->kind_data = bytes;
     }
 
+#if STAT_ALLOCATIONS
+    gc_alloc_bytes.log(bytes);
+#endif
+
     return rtn;
 }
 

src/runtime/builtin_modules/builtins.cpp

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include <algorithm>
+#include <cfloat>
 #include <cstddef>
 #include <err.h>
 
@@ -303,6 +304,7 @@ extern "C" Box* ord(Box* obj) {
 }
 
 Box* range(Box* start, Box* stop, Box* step) {
+    STAT_TIMER(t0, "us_timer_builtin_range");
     i64 istart, istop, istep;
     if (stop == NULL) {
         istart = 0;
@@ -346,6 +348,7 @@ Box* notimplementedRepr(Box* self) {
 }
 
 Box* sorted(Box* obj, Box* cmp, Box* key, Box** args) {
+    STAT_TIMER(t0, "us_timer_builtin_sorted");
     Box* reverse = args[0];
 
     BoxedList* rtn = new BoxedList();
@@ -358,6 +361,7 @@ Box* sorted(Box* obj, Box* cmp, Box* key, Box** args) {
 }
 
 Box* isinstance_func(Box* obj, Box* cls) {
+    STAT_TIMER(t0, "us_timer_builtin_isinstance");
     int rtn = PyObject_IsInstance(obj, cls);
     if (rtn < 0)
         checkAndThrowCAPIException();
@@ -365,16 +369,29 @@ Box* isinstance_func(Box* obj, Box* cls) {
 }
 
 Box* issubclass_func(Box* child, Box* parent) {
+    STAT_TIMER(t0, "us_timer_builtin_issubclass");
     int rtn = PyObject_IsSubclass(child, parent);
     if (rtn < 0)
         checkAndThrowCAPIException();
     return boxBool(rtn);
 }
 
+Box* intern_func(Box* str) {
+    if (!PyString_CheckExact(str)) // have to use exact check!
+        raiseExcHelper(TypeError, "can't intern subclass of string");
+    PyString_InternInPlace(&str);
+    checkAndThrowCAPIException();
+    return str;
+}
+
 Box* bltinImport(Box* name, Box* globals, Box* locals, Box** args) {
     Box* fromlist = args[0];
     Box* level = args[1];
 
+    // __import__ takes a 'locals' argument, but it doesn't get used in CPython.
+    // Well, it gets passed to PyImport_ImportModuleLevel() and then import_module_level(),
+    // which ignores it.  So we don't even pass it through.
+
     name = coerceUnicodeToStr(name);
 
     if (name->cls != str_cls) {
@@ -586,7 +603,6 @@ public:
     static Box* __reduce__(Box* self) {
         RELEASE_ASSERT(isSubclass(self->cls, BaseException), "");
         BoxedException* exc = static_cast<BoxedException*>(self);
-
         return BoxedTuple::create({ self->cls, EmptyTuple, self->getAttrWrapper() });
     }
 };
@@ -973,16 +989,38 @@ Box* builtinRound(Box* _number, Box* _ndigits) {
         raiseExcHelper(TypeError, "a float is required");
 
     BoxedFloat* number = (BoxedFloat*)_number;
+    double x = number->d;
 
-    if (isSubclass(_ndigits->cls, int_cls)) {
-        BoxedInt* ndigits = (BoxedInt*)_ndigits;
+    /* interpret 2nd argument as a Py_ssize_t; clip on overflow */
+    Py_ssize_t ndigits = PyNumber_AsSsize_t(_ndigits, NULL);
+    if (ndigits == -1 && PyErr_Occurred())
+        throwCAPIException();
 
-        if (ndigits->n == 0)
-            return boxFloat(round(number->d));
+    /* nans, infinities and zeros round to themselves */
+    if (!std::isfinite(x) || x == 0.0)
+        return boxFloat(x);
+
+/* Deal with extreme values for ndigits. For ndigits > NDIGITS_MAX, x
+   always rounds to itself.  For ndigits < NDIGITS_MIN, x always
+   rounds to +-0.0.  Here 0.30103 is an upper bound for log10(2). */
+#define NDIGITS_MAX ((int)((DBL_MANT_DIG - DBL_MIN_EXP) * 0.30103))
+#define NDIGITS_MIN (-(int)((DBL_MAX_EXP + 1) * 0.30103))
+    if (ndigits > NDIGITS_MAX)
+        /* return x */
+        return boxFloat(x);
+    else if (ndigits < NDIGITS_MIN)
+        /* return 0.0, but with sign of x */
+        return boxFloat(0.0 * x);
+    else {
+        /* finite x, and ndigits is not unreasonably large */
+        /* _Py_double_round is defined in floatobject.c */
+        Box* rtn = _Py_double_round(x, (int)ndigits);
+        if (!rtn)
+            throwCAPIException();
+        return rtn;
     }
-
-    fatalOrError(PyExc_NotImplementedError, "unimplemented");
-    throwCAPIException();
+#undef NDIGITS_MAX
+#undef NDIGITS_MIN
 }
 
 Box* builtinCmp(Box* a, Box* b) {
@@ -1109,6 +1147,9 @@ void setupBuiltins() {
         = new BoxedBuiltinFunctionOrMethod(boxRTFunction((void*)issubclass_func, BOXED_BOOL, 2), "issubclass");
     builtins_module->giveAttr("issubclass", issubclass_obj);
 
+    Box* intern_obj = new BoxedBuiltinFunctionOrMethod(boxRTFunction((void*)intern_func, UNKNOWN, 1), "intern");
+    builtins_module->giveAttr("intern", intern_obj);
+
     CLFunction* import_func = boxRTFunction((void*)bltinImport, UNKNOWN, 5, 4, false, false,
                                             ParamNames({ "name", "globals", "locals", "fromlist", "level" }, "", ""));
     builtins_module->giveAttr("__import__", new BoxedBuiltinFunctionOrMethod(import_func, "__import__",

src/runtime/dict.cpp

@@ -48,6 +48,7 @@ Box* dictRepr(BoxedDict* self) {
 }
 
 Box* dictClear(BoxedDict* self) {
+    STAT_TIMER(t0, "us_timer_dictClear");
     if (!isSubclass(self->cls, dict_cls))
         raiseExcHelper(TypeError, "descriptor 'clear' requires a 'dict' object but received a '%s'", getTypeName(self));
 
@@ -56,6 +57,7 @@ Box* dictClear(BoxedDict* self) {
 }
 
 Box* dictCopy(BoxedDict* self) {
+    STAT_TIMER(t0, "us_timer_dictCopy");
     if (!isSubclass(self->cls, dict_cls))
         raiseExcHelper(TypeError, "descriptor 'copy' requires a 'dict' object but received a '%s'", getTypeName(self));
 
@@ -65,6 +67,7 @@ Box* dictCopy(BoxedDict* self) {
 }
 
 Box* dictItems(BoxedDict* self) {
+    STAT_TIMER(t0, "us_timer_dictItems");
     BoxedList* rtn = new BoxedList();
 
     rtn->ensure(self->d.size());
@@ -77,6 +80,7 @@ Box* dictItems(BoxedDict* self) {
 }
 
 Box* dictValues(BoxedDict* self) {
+    STAT_TIMER(t0, "us_timer_dictValues");
     BoxedList* rtn = new BoxedList();
     rtn->ensure(self->d.size());
     for (const auto& p : self->d) {
@@ -86,6 +90,7 @@ Box* dictValues(BoxedDict* self) {
 }
 
 Box* dictKeys(BoxedDict* self) {
+    STAT_TIMER(t0, "us_timer_dictKeys");
     RELEASE_ASSERT(isSubclass(self->cls, dict_cls), "");
 
     BoxedList* rtn = new BoxedList();
@@ -182,6 +187,7 @@ extern "C" int PyDict_Update(PyObject* a, PyObject* b) noexcept {
 }
 
 Box* dictGetitem(BoxedDict* self, Box* k) {
+    STAT_TIMER(t0, "us_timer_dictGetitem");
     if (!isSubclass(self->cls, dict_cls))
         raiseExcHelper(TypeError, "descriptor '__getitem__' requires a 'dict' object but received a '%s'",
                        getTypeName(self));
@@ -306,6 +312,7 @@ extern "C" PyObject* PyDict_GetItemString(PyObject* dict, const char* key) noexc
 }
 
 Box* dictSetitem(BoxedDict* self, Box* k, Box* v) {
+    STAT_TIMER(t0, "us_timer_dictSetitem");
     // printf("Starting setitem\n");
     Box*& pos = self->d[k];
     // printf("Got the pos\n");
@@ -320,6 +327,7 @@ Box* dictSetitem(BoxedDict* self, Box* k, Box* v) {
 }
 
 Box* dictDelitem(BoxedDict* self, Box* k) {
+    STAT_TIMER(t0, "us_timer_dictDelitem");
     if (!isSubclass(self->cls, dict_cls))
         raiseExcHelper(TypeError, "descriptor '__delitem__' requires a 'dict' object but received a '%s'",
                        getTypeName(self));

src/runtime/generator.cpp

@@ -86,28 +86,36 @@ Context* getReturnContextForGeneratorFrame(void* frame_addr) {
 }
 
 void generatorEntry(BoxedGenerator* g) {
-    assert(g->cls == generator_cls);
-    assert(g->function->cls == function_cls);
+    {
+        STAT_TIMER2(t0, "us_timer_generator_toplevel", g->timer_time);
 
-    threading::pushGenerator(g, g->stack_begin, g->returnContext);
+        assert(g->cls == generator_cls);
+        assert(g->function->cls == function_cls);
 
-    try {
-        RegisterHelper context_registerer(g, __builtin_frame_address(0));
+        threading::pushGenerator(g, g->stack_begin, g->returnContext);
+        try {
+            RegisterHelper context_registerer(g, __builtin_frame_address(0));
 
-        // call body of the generator
-        BoxedFunctionBase* func = g->function;
+            // call body of the generator
+            BoxedFunctionBase* func = g->function;
 
-        Box** args = g->args ? &g->args->elts[0] : nullptr;
-        callCLFunc(func->f, nullptr, func->f->numReceivedArgs(), func->closure, g, func->globals, g->arg1, g->arg2,
-                   g->arg3, args);
-    } catch (ExcInfo e) {
-        // unhandled exception: propagate the exception to the caller
-        g->exception = e;
-    }
+            Box** args = g->args ? &g->args->elts[0] : nullptr;
+            callCLFunc(func->f, nullptr, func->f->numReceivedArgs(), func->closure, g, func->globals, g->arg1, g->arg2,
+                       g->arg3, args);
+        } catch (ExcInfo e) {
+            // unhandled exception: propagate the exception to the caller
+            g->exception = e;
+        }
 
-    // we returned from the body of the generator. next/send/throw will notify the caller
-    g->entryExited = true;
-    threading::popGenerator();
+        // we returned from the body of the generator. next/send/throw will notify the caller
+        g->entryExited = true;
+        threading::popGenerator();
+
+#if STAT_TIMERS
+        g->timer_time = getCPUTicks(); // store off the timer that our caller (in g->returnContext) will resume at
+        STAT_TIMER_NAME(t0).pause(g->timer_time);
+#endif
+    }
     swapContext(&g->context, g->returnContext, 0);
 }
 
@@ -130,7 +138,21 @@ Box* generatorSend(Box* s, Box* v) {
 
     self->returnValue = v;
     self->running = true;
+
+#if STAT_TIMERS
+    // store off the time that the generator will use to initialize its toplevel timer
+    self->timer_time = getCPUTicks();
+    StatTimer* current_timers = StatTimer::swapStack(self->statTimers, self->timer_time);
+#endif
+
     swapContext(&self->returnContext, self->context, (intptr_t)self);
+
+#if STAT_TIMERS
+    // if the generator exited we use the time that generatorEntry stored in self->timer_time (the same time it paused
+    // its timer at).
+    self->statTimers = StatTimer::swapStack(current_timers, self->entryExited ? self->timer_time : getCPUTicks());
+#endif
+
     self->running = false;
 
     // propagate exception to the caller

src/runtime/import.cpp

@@ -21,6 +21,7 @@
 
 #include "codegen/irgen/hooks.h"
 #include "codegen/parser.h"
+#include "codegen/unwinding.h"
 #include "runtime/capi.h"
 #include "runtime/objmodel.h"
 
@@ -267,7 +268,7 @@ static Box* getParent(Box* globals, int level, std::string& buf) {
     if (globals == NULL || globals == None || level == 0)
         return None;
 
-    BoxedString* pkgname = static_cast<BoxedString*>(getattrInternal(globals, package_str, NULL));
+    BoxedString* pkgname = static_cast<BoxedString*>(getFromGlobals(globals, package_str));
     if (pkgname != NULL && pkgname != None) {
         /* __package__ is set, so use it */
         if (pkgname->cls != str_cls) {
@@ -286,17 +287,11 @@ static Box* getParent(Box* globals, int level, std::string& buf) {
         buf += pkgname->s;
     } else {
         /* __package__ not set, so figure it out and set it */
-        BoxedString* modname = static_cast<BoxedString*>(getattrInternal(globals, name_str, NULL));
+        BoxedString* modname = static_cast<BoxedString*>(getFromGlobals(globals, name_str));
         if (modname == NULL || modname->cls != str_cls)
             return None;
 
-        Box* modpath = NULL;
-        try {
-            modpath = getattrInternal(globals, path_str, NULL);
-        } catch (ExcInfo e) {
-            if (!e.matches(AttributeError))
-                raiseRaw(e);
-        }
+        Box* modpath = getFromGlobals(globals, path_str);
 
         if (modpath != NULL) {
             /* __path__ is set, so modname is already the package name */
@@ -304,7 +299,7 @@ static Box* getParent(Box* globals, int level, std::string& buf) {
                 raiseExcHelper(ValueError, "Module name too long");
             }
             buf += modname->s;
-            globals->setattr(package_str, modname, NULL);
+            setGlobal(globals, package_str, modname);
         } else {
             /* Normal module, so work out the package name if any */
             size_t lastdot = modname->s.rfind('.');
@@ -312,7 +307,7 @@ static Box* getParent(Box* globals, int level, std::string& buf) {
                 raiseExcHelper(ValueError, "Attempted relative import in non-package");
             }
             if (lastdot == std::string::npos) {
-                globals->setattr(package_str, None, NULL);
+                setGlobal(globals, package_str, None);
                 return None;
             }
             if (lastdot >= PATH_MAX) {
@@ -320,7 +315,7 @@ static Box* getParent(Box* globals, int level, std::string& buf) {
             }
 
             buf = std::string(modname->s, 0, lastdot);
-            globals->setattr(package_str, boxStringPtr(&buf), NULL);
+            setGlobal(globals, package_str, boxStringPtr(&buf));
         }
     }
 
@@ -471,7 +466,6 @@ static bool loadNext(Box* mod, Box* altmod, std::string& name, std::string& buf,
 
 static void ensureFromlist(Box* module, Box* fromlist, std::string& buf, bool recursive);
 Box* importModuleLevel(const std::string& name, Box* globals, Box* from_imports, int level) {
-    RELEASE_ASSERT(!globals || globals == None || isSubclass(globals->cls, module_cls), "");
     bool return_first = from_imports == None;
 
     static StatCounter slowpath_import("slowpath_import");
@@ -665,7 +659,7 @@ Box* nullImporterFindModule(Box* self, Box* fullname, Box* path) {
 
 extern "C" Box* import(int level, Box* from_imports, const std::string* module_name) {
     std::string _module_name(*module_name);
-    return importModuleLevel(_module_name, getCurrentModule(), from_imports, level);
+    return importModuleLevel(_module_name, getGlobals(), from_imports, level);
 }
 
 Box* impFindModule(Box* _name, BoxedList* path) {

src/runtime/int.cpp

@@ -200,7 +200,7 @@ extern "C" PyObject* PyInt_FromString(const char* s, char** pend, int base) noex
         if (x < 0)
             return PyLong_FromString(s, pend, base);
     } else
-        x = strtoul(s, &end, base);
+        x = strtol(s, &end, base);
     if (end == s || !isalnum(Py_CHARMASK(end[-1])))
         goto bad;
     while (*end && isspace(Py_CHARMASK(*end)))
@@ -672,8 +672,13 @@ extern "C" Box* intLShiftInt(BoxedInt* lhs, BoxedInt* rhs) {
     if (rhs->n < 0)
         raiseExcHelper(ValueError, "negative shift count");
 
-    // TODO overflow?
-    return boxInt(lhs->n << rhs->n);
+    bool undefined = rhs->n >= sizeof(rhs->n) * 8;
+    if (!undefined) {
+        int64_t res = lhs->n << rhs->n;
+        if ((res >> rhs->n) == lhs->n)
+            return boxInt(lhs->n << rhs->n);
+    }
+    return longLshift(boxLong(lhs->n), rhs);
 }
 
 extern "C" Box* intLShift(BoxedInt* lhs, Box* rhs) {

src/runtime/iterobject.cpp

@@ -109,7 +109,7 @@ Box* seqiterNext(Box* s) {
 }
 
 static void seqiterGCVisit(GCVisitor* v, Box* b) {
-    assert(b->cls == seqiter_cls);
+    assert(b->cls == seqiter_cls || b->cls == seqreviter_cls);
     boxGCHandler(v, b);
 
     BoxedSeqIter* si = static_cast<BoxedSeqIter*>(b);
@@ -182,7 +182,8 @@ void setupIter() {
     seqiter_cls->freeze();
     seqiter_cls->tpp_hasnext = seqiterHasnextUnboxed;
 
-    seqreviter_cls = BoxedHeapClass::create(type_cls, object_cls, NULL, 0, 0, sizeof(BoxedSeqIter), false, "reversed");
+    seqreviter_cls
+        = BoxedHeapClass::create(type_cls, object_cls, seqiterGCVisit, 0, 0, sizeof(BoxedSeqIter), false, "reversed");
 
     seqreviter_cls->giveAttr("next", new BoxedFunction(boxRTFunction((void*)seqiterNext, UNKNOWN, 1)));
     seqreviter_cls->giveAttr("__hasnext__", new BoxedFunction(boxRTFunction((void*)seqreviterHasnext, BOXED_BOOL, 1)));

src/runtime/list.cpp

@@ -25,6 +25,7 @@
 #include "core/stats.h"
 #include "core/types.h"
 #include "gc/collector.h"
+#include "gc/roots.h"
 #include "runtime/objmodel.h"
 #include "runtime/types.h"
 #include "runtime/util.h"
@@ -423,23 +424,21 @@ extern "C" Box* listSetitemSlice(BoxedList* self, BoxedSlice* slice, Box* v) {
     size_t v_size;
     Box** v_elts;
 
+    RootedBox v_as_seq((Box*)nullptr);
     if (!v) {
         v_size = 0;
         v_elts = NULL;
-    } else if (v->cls == list_cls || isSubclass(v->cls, list_cls)) {
-        BoxedList* lv = static_cast<BoxedList*>(v);
-        v_size = lv->size;
+    } else {
+        v_as_seq = RootedBox(PySequence_Fast(v, "can only assign an iterable"));
+        if (v_as_seq == NULL)
+            throwCAPIException();
+
+        v_size = PySequence_Fast_GET_SIZE(v_as_seq);
         // If lv->size is 0, lv->elts->elts is garbage
         if (v_size)
-            v_elts = lv->elts->elts;
+            v_elts = PySequence_Fast_ITEMS(v_as_seq);
         else
             v_elts = NULL;
-    } else if (v->cls == tuple_cls || isSubclass(v->cls, tuple_cls)) {
-        BoxedTuple* tv = static_cast<BoxedTuple*>(v);
-        v_size = tv->size();
-        v_elts = &tv->elts[0];
-    } else {
-        RELEASE_ASSERT(0, "unsupported type for list slice assignment: '%s'", getTypeName(v));
     }
 
     // If self->size is 0, self->elts->elts is garbage
@@ -542,6 +541,8 @@ extern "C" Box* listInsert(BoxedList* self, Box* idx, Box* v) {
 }
 
 Box* listMul(BoxedList* self, Box* rhs) {
+    STAT_TIMER(t0, "us_timer_listMul");
+
     if (rhs->cls != int_cls) {
         raiseExcHelper(TypeError, "can't multiply sequence by non-int of type '%s'", getTypeName(rhs));
     }
@@ -961,7 +962,7 @@ Box* listNe(BoxedList* self, Box* rhs) {
 }
 
 Box* listLt(BoxedList* self, Box* rhs) {
-    if (rhs->cls != list_cls) {
+    if (!isSubclass(rhs->cls, list_cls)) {
         return NotImplemented;
     }
 
@@ -971,7 +972,7 @@ Box* listLt(BoxedList* self, Box* rhs) {
 }
 
 Box* listLe(BoxedList* self, Box* rhs) {
-    if (rhs->cls != list_cls) {
+    if (!isSubclass(rhs->cls, list_cls)) {
         return NotImplemented;
     }
 
@@ -981,7 +982,7 @@ Box* listLe(BoxedList* self, Box* rhs) {
 }
 
 Box* listGt(BoxedList* self, Box* rhs) {
-    if (rhs->cls != list_cls) {
+    if (!isSubclass(rhs->cls, list_cls)) {
         return NotImplemented;
     }
 
@@ -991,7 +992,7 @@ Box* listGt(BoxedList* self, Box* rhs) {
 }
 
 Box* listGe(BoxedList* self, Box* rhs) {
-    if (rhs->cls != list_cls) {
+    if (!isSubclass(rhs->cls, list_cls)) {
         return NotImplemented;
     }
 

src/runtime/long.cpp

@@ -116,7 +116,48 @@ extern "C" unsigned PY_LONG_LONG PyLong_AsUnsignedLongLongMask(PyObject* vv) noe
 }
 
 extern "C" PY_LONG_LONG PyLong_AsLongLong(PyObject* vv) noexcept {
-    Py_FatalError("unimplemented");
+    PY_LONG_LONG bytes;
+    int one = 1;
+    int res;
+
+    if (vv == NULL) {
+        PyErr_BadInternalCall();
+        return -1;
+    }
+    if (!PyLong_Check(vv)) {
+        PyNumberMethods* nb;
+        PyObject* io;
+        if (PyInt_Check(vv))
+            return (PY_LONG_LONG)PyInt_AsLong(vv);
+        if ((nb = vv->cls->tp_as_number) == NULL || nb->nb_int == NULL) {
+            PyErr_SetString(PyExc_TypeError, "an integer is required");
+            return -1;
+        }
+        io = (*nb->nb_int)(vv);
+        if (io == NULL)
+            return -1;
+        if (PyInt_Check(io)) {
+            bytes = PyInt_AsLong(io);
+            Py_DECREF(io);
+            return bytes;
+        }
+        if (PyLong_Check(io)) {
+            bytes = PyLong_AsLongLong(io);
+            Py_DECREF(io);
+            return bytes;
+        }
+        Py_DECREF(io);
+        PyErr_SetString(PyExc_TypeError, "integer conversion failed");
+        return -1;
+    }
+
+    res = _PyLong_AsByteArray((PyLongObject*)vv, (unsigned char*)&bytes, SIZEOF_LONG_LONG, IS_LITTLE_ENDIAN, 1);
+
+    /* Plan 9 can't handle PY_LONG_LONG in ? : expressions */
+    if (res < 0)
+        return (PY_LONG_LONG)-1;
+    else
+        return bytes;
 }
 
 extern "C" PY_LONG_LONG PyLong_AsLongLongAndOverflow(PyObject* obj, int* overflow) noexcept {
@@ -442,7 +483,12 @@ extern "C" void* PyLong_AsVoidPtr(PyObject* vv) noexcept {
 
 extern "C" int _PyLong_AsByteArray(PyLongObject* v, unsigned char* bytes, size_t n, int little_endian,
                                    int is_signed) noexcept {
-    Py_FatalError("unimplemented");
+    RELEASE_ASSERT(little_endian == 1, "not implemented");
+    RELEASE_ASSERT(n == 8, "did not yet check if the behaviour is correct for sizes other than 8");
+    size_t count = 0;
+    mpz_export(bytes, &count, -1, n, 0, 0, ((BoxedLong*)v)->n);
+    RELEASE_ASSERT(count <= n, "overflow handling is not yet implemented");
+    return 0;
 }
 
 extern "C" PyObject* _PyLong_FromByteArray(const unsigned char* bytes, size_t n, int little_endian,

src/runtime/objmodel.h

@@ -183,7 +183,13 @@ inline std::tuple<Box*, Box*, Box*, Box**> getTupleFromArgsArray(Box** args, int
 
 // The `globals` argument can be either a BoxedModule or a BoxedDict
 
+// Corresponds to a name lookup with GLOBAL scope.  Checks the passed globals object, then the builtins,
+// and if not found raises an exception.
 extern "C" Box* getGlobal(Box* globals, const std::string* name);
+// Checks for the name just in the passed globals object, and returns NULL if it is not found.
+// This includes if the globals object defined a custom __getattr__ method that threw an AttributeError.
+Box* getFromGlobals(Box* globals, llvm::StringRef name);
+void setGlobal(Box* globals, llvm::StringRef name, Box* value);
 extern "C" void delGlobal(Box* globals, const std::string* name);
 
 extern "C" void boxedLocalsSet(Box* boxedLocals, const char* attr, Box* val);

src/runtime/stacktrace.cpp

@@ -96,6 +96,7 @@ void unwindExc(Box* exc_obj) {
 
 void raiseRaw(const ExcInfo& e) __attribute__((__noreturn__));
 void raiseRaw(const ExcInfo& e) {
+    STAT_TIMER(t0, "us_timer_raiseraw");
     // Should set these to None before getting here:
     assert(e.type);
     assert(e.value);
@@ -197,6 +198,7 @@ extern "C" void abort() {
     __builtin_unreachable();
 }
 
+#if 0
 extern "C" void exit(int code) {
     static void (*libc_exit)(int) = (void (*)(int))dlsym(RTLD_NEXT, "exit");
 
@@ -218,6 +220,7 @@ extern "C" void exit(int code) {
     libc_exit(code);
     __builtin_unreachable();
 }
+#endif
 
 extern "C" void raise0() {
     ExcInfo* exc_info = getFrameExcInfo();

src/runtime/str.cpp

@@ -1112,6 +1112,7 @@ extern "C" Box* strMod(BoxedString* lhs, Box* rhs) {
 }
 
 extern "C" Box* strMul(BoxedString* lhs, Box* rhs) {
+    STAT_TIMER(t0, "us_timer_strMul");
     assert(isSubclass(lhs->cls, str_cls));
 
     int n;
@@ -1514,6 +1515,7 @@ failed:
 }
 
 extern "C" size_t unicodeHashUnboxed(PyUnicodeObject* self) {
+    STAT_TIMER(t0, "us_timer_unicodeHashUnboxed");
     if (self->hash != -1)
         return self->hash;
 
@@ -1524,6 +1526,7 @@ extern "C" size_t unicodeHashUnboxed(PyUnicodeObject* self) {
 }
 
 extern "C" Box* strHash(BoxedString* self) {
+    STAT_TIMER(t0, "us_timer_strHash");
     assert(isSubclass(self->cls, str_cls));
 
     StringHash<char> H;
@@ -2302,8 +2305,17 @@ extern "C" PyObject* PyString_FromStringAndSize(const char* s, ssize_t n) noexce
     return boxStrConstantSize(s, n);
 }
 
+static /*const*/ char* string_getbuffer(register PyObject* op) noexcept {
+    char* s;
+    Py_ssize_t len;
+    if (PyString_AsStringAndSize(op, &s, &len))
+        return NULL;
+    return s;
+}
+
 extern "C" char* PyString_AsString(PyObject* o) noexcept {
-    RELEASE_ASSERT(isSubclass(o->cls, str_cls), "");
+    if (!PyString_Check(o))
+        return string_getbuffer(o);
 
     BoxedString* s = static_cast<BoxedString*>(o);
     return getWriteableStringContents(s);

src/runtime/tuple.cpp

@@ -168,6 +168,7 @@ Box* tupleAdd(BoxedTuple* self, Box* rhs) {
 }
 
 Box* tupleMul(BoxedTuple* self, Box* rhs) {
+    STAT_TIMER(t0, "us_timer_tupleMul");
     if (rhs->cls != int_cls) {
         raiseExcHelper(TypeError, "can't multiply sequence by non-int of type '%s'", getTypeName(rhs));
     }
@@ -337,6 +338,7 @@ Box* tupleIndex(BoxedTuple* self, Box* elt) {
 }
 
 Box* tupleHash(BoxedTuple* self) {
+    STAT_TIMER(t0, "us_timer_tupleHash");
     assert(isSubclass(self->cls, tuple_cls));
 
     int64_t rtn = 3527539;

src/runtime/types.cpp

@@ -978,7 +978,39 @@ Box* sliceRepr(BoxedSlice* self) {
 
 extern "C" int PySlice_GetIndices(PySliceObject* r, Py_ssize_t length, Py_ssize_t* start, Py_ssize_t* stop,
                                   Py_ssize_t* step) noexcept {
-    Py_FatalError("unimplemented");
+    /* XXX support long ints */
+    if (r->step == Py_None) {
+        *step = 1;
+    } else {
+        if (!PyInt_Check(r->step) && !PyLong_Check(r->step))
+            return -1;
+        *step = PyInt_AsSsize_t(r->step);
+    }
+    if (r->start == Py_None) {
+        *start = *step < 0 ? length - 1 : 0;
+    } else {
+        if (!PyInt_Check(r->start) && !PyLong_Check(r->step))
+            return -1;
+        *start = PyInt_AsSsize_t(r->start);
+        if (*start < 0)
+            *start += length;
+    }
+    if (r->stop == Py_None) {
+        *stop = *step < 0 ? -1 : length;
+    } else {
+        if (!PyInt_Check(r->stop) && !PyLong_Check(r->step))
+            return -1;
+        *stop = PyInt_AsSsize_t(r->stop);
+        if (*stop < 0)
+            *stop += length;
+    }
+    if (*stop > length)
+        return -1;
+    if (*start >= length)
+        return -1;
+    if (*step == 0)
+        return -1;
+    return 0;
 }
 
 extern "C" int PySlice_GetIndicesEx(PySliceObject* _r, Py_ssize_t length, Py_ssize_t* start, Py_ssize_t* stop,
@@ -1238,6 +1270,8 @@ public:
     }
 
     static Box* setitem(Box* _self, Box* _key, Box* value) {
+        STAT_TIMER(t0, "us_timer_AttrWrapper_setitem");
+
         RELEASE_ASSERT(_self->cls == attrwrapper_cls, "");
         AttrWrapper* self = static_cast<AttrWrapper*>(_self);
 
@@ -1250,6 +1284,7 @@ public:
     }
 
     static Box* setdefault(Box* _self, Box* _key, Box* value) {
+        STAT_TIMER(t0, "us_timer_AttrWrapper_setdefault");
         RELEASE_ASSERT(_self->cls == attrwrapper_cls, "");
         AttrWrapper* self = static_cast<AttrWrapper*>(_self);
 
@@ -1265,6 +1300,7 @@ public:
     }
 
     static Box* get(Box* _self, Box* _key, Box* def) {
+        STAT_TIMER(t0, "us_timer_AttrWrapper_get");
         RELEASE_ASSERT(_self->cls == attrwrapper_cls, "");
         AttrWrapper* self = static_cast<AttrWrapper*>(_self);
 
@@ -1279,6 +1315,7 @@ public:
     }
 
     static Box* getitem(Box* _self, Box* _key) {
+        STAT_TIMER(t0, "us_timer_AttrWrapper_getitem");
         RELEASE_ASSERT(_self->cls == attrwrapper_cls, "");
         AttrWrapper* self = static_cast<AttrWrapper*>(_self);
 
@@ -1312,6 +1349,7 @@ public:
     }
 
     static Box* delitem(Box* _self, Box* _key) {
+        STAT_TIMER(t0, "us_timer_AttrWrapper_delitem");
         RELEASE_ASSERT(_self->cls == attrwrapper_cls, "");
         AttrWrapper* self = static_cast<AttrWrapper*>(_self);
 
@@ -1377,6 +1415,7 @@ public:
     }
 
     static Box* values(Box* _self) {
+        STAT_TIMER(t0, "us_timer_AttrWrapper_values");
         RELEASE_ASSERT(_self->cls == attrwrapper_cls, "");
         AttrWrapper* self = static_cast<AttrWrapper*>(_self);
 
@@ -1391,6 +1430,7 @@ public:
     }
 
     static Box* items(Box* _self) {
+        STAT_TIMER(t0, "us_timer_AttrWrapper_items");
         RELEASE_ASSERT(_self->cls == attrwrapper_cls, "");
         AttrWrapper* self = static_cast<AttrWrapper*>(_self);
 
@@ -1429,6 +1469,7 @@ public:
     }
 
     static Box* update(Box* _self, Box* _container) {
+        STAT_TIMER(t0, "us_timer_AttrWrapper_update");
         RELEASE_ASSERT(_self->cls == attrwrapper_cls, "");
         AttrWrapper* self = static_cast<AttrWrapper*>(_self);
 

src/runtime/types.h

@@ -410,12 +410,37 @@ public:
     size_t size() { return s.size(); }
 
     void* operator new(size_t size, size_t ssize) __attribute__((visibility("default"))) {
+#if STAT_ALLOCATIONS
+        static StatCounter alloc_str("alloc.str");
+        static StatCounter alloc_str1("alloc.str(1)");
+        static StatCounter allocsize_str("allocsize.str");
+
+        if (ssize == 1)
+            alloc_str1.log();
+        else
+            alloc_str.log();
+
+        allocsize_str.log(str_cls->tp_basicsize + ssize + 1);
+#endif
         Box* rtn = static_cast<Box*>(gc_alloc(str_cls->tp_basicsize + ssize + 1, gc::GCKind::PYTHON));
         rtn->cls = str_cls;
         return rtn;
     }
 
     void* operator new(size_t size, BoxedClass* cls, size_t ssize) __attribute__((visibility("default"))) {
+#if STAT_ALLOCATIONS
+        static StatCounter alloc_str("alloc.str");
+        static StatCounter alloc_str1("alloc.str(1)");
+
+        static StatCounter allocsize_str("allocsize.str");
+
+        if (ssize == 1)
+            alloc_str1.log();
+        else
+            alloc_str.log();
+
+        allocsize_str.log(cls->tp_basicsize + ssize + 1);
+#endif
         Box* rtn = static_cast<Box*>(cls->tp_alloc(cls, ssize + 1));
         rtn->cls = cls;
         return rtn;
@@ -517,12 +542,43 @@ public:
     Box** elts;
 
     void* operator new(size_t size, size_t nelts) __attribute__((visibility("default"))) {
+#if STAT_ALLOCATIONS
+        static StatCounter alloc_tuple("alloc.tuple");
+        static StatCounter alloc_tuple0("alloc.tuple(0)");
+
+        static StatCounter allocsize_tuple("allocsize.tuple");
+        static StatCounter allocsize_tuple0("allocsize.tuple(0)");
+
+        if (nelts == 0) {
+            alloc_tuple0.log();
+            allocsize_tuple0.log(_PyObject_VAR_SIZE(tuple_cls, nelts + 1));
+        } else {
+            alloc_tuple.log();
+            allocsize_tuple.log(_PyObject_VAR_SIZE(tuple_cls, nelts + 1));
+        }
+#endif
+
         Box* rtn = static_cast<Box*>(gc_alloc(_PyObject_VAR_SIZE(tuple_cls, nelts + 1), gc::GCKind::PYTHON));
         rtn->cls = tuple_cls;
         return rtn;
     }
 
     void* operator new(size_t size, BoxedClass* cls, size_t nelts) __attribute__((visibility("default"))) {
+#if STAT_ALLOCATIONS
+        static StatCounter alloc_tuple("alloc.tuple");
+        static StatCounter alloc_tuple0("alloc.tuple(0)");
+
+        static StatCounter allocsize_tuple("allocsize.tuple");
+        static StatCounter allocsize_tuple0("allocsize.tuple(0)");
+
+        if (nelts == 0) {
+            alloc_tuple0.log();
+            allocsize_tuple0.log(_PyObject_VAR_SIZE(cls, nelts + 1));
+        } else {
+            alloc_tuple.log();
+            allocsize_tuple.log(_PyObject_VAR_SIZE(cls, nelts + 1));
+        }
+#endif
         Box* rtn = static_cast<Box*>(cls->tp_alloc(cls, nelts));
         rtn->cls = cls;
         return rtn;
@@ -817,6 +873,11 @@ public:
     struct Context* context, *returnContext;
     void* stack_begin;
 
+#if STAT_TIMERS
+    StatTimer* statTimers;
+    uint64_t timer_time;
+#endif
+
     BoxedGenerator(BoxedFunctionBase* function, Box* arg1, Box* arg2, Box* arg3, Box** args);
 
     DEFAULT_CLASS(generator_cls);

test/integration/virtualenv_test.py

@@ -21,9 +21,10 @@ set -e
 set -ux
 python -c 'import __future__'
 python -c 'import sys; print sys.executable'
-pip install bcrypt==1.1.0 python-gflags==2.0
+pip install bcrypt==1.1.0 python-gflags==2.0 sqlalchemy==1.0.0
 python -c 'import bcrypt; assert bcrypt.__version__ == "1.1.0"; assert bcrypt.hashpw("password1", "$2a$12$0123456789012345678901").endswith("I1hdtg4K"); print "bcrypt seems to work"'
 python -c 'import gflags; print "gflags imports"'
+python -c 'import sqlalchemy; print "sqlalchemy imports"'
 """.strip()
 
 # print sh_script

test/tests/builtins.py

@@ -105,8 +105,8 @@ print callable(lambda: 1)
 
 print range(5L, 7L)
 
-print round(-1.1), round(-1.9)
-print round(0.5), round(-0.5)
+for n in [0, 1, 2, 3, 4, 5]:
+    print round(-1.1, n), round(-1.9, n), round(0.5, n), round(-0.5, n), round(-0.123456789, n)
 
 print list(iter(xrange(100).__iter__().next, 20))
 

test/tests/import_globals.py

+import test_package.import_target
+import import_target
+print
+
+s = """
+import import_target
+print import_target.__file__.replace(".pyc", ".py")
+"""
+
+def test(g):
+    print "Testing with globals:", "None" if g is None else ("globals" if g is globals() else sorted(g.items()))
+    print __import__("import_target", g).__file__.replace(".pyc", ".py")
+    exec s in g
+    print "Resulting globals:", "None" if g is None else ("globals" if g is globals() else sorted(g.keys()))
+    print
+
+test(None)
+test(globals())
+
+test({"__package__":"test_package"})
+
+test({"__name__":"test_package.foo"})
+test({"__name__":"test_package"})
+test({"__name__":"test_package", "__path__":"foo"})

test/tests/intmethods.py

@@ -66,6 +66,9 @@ class L(object):
 print type(int(L()))
 
 print int(u'123')
+print int("9223372036854775808", 0)
+print 1 << 63, 1 << 64, -1 << 63, -1 << 64, 2 << 63
+print type(1 << 63), type(1 << 64), type(-1 << 63), type(-1 << 64), type(2 << 63)
 
 for b in range(26):
     try:

test/tests/list.py

@@ -124,7 +124,13 @@ for i in xrange(3):
             l[j:k] = ["added"]
             print i, j, k, l
 
-
+def G():
+    yield "a"
+    yield "b"
+    yield "c"
+l = [0, 1, 2, 3, 4, 5]
+l[1:] = G()
+print l
 
 l = [1, 3, 5, 7, 2, 4]
 print l.sort(key=lambda x:x%3)

test/tests/list_subclassing.py

@@ -13,5 +13,26 @@ print len(MyList.__new__(MyList))
 l[:] = l[:]
 print l
 
-print [1,2,3] == MyList((1,2,3))
-print [1,2,3] != MyList((1,2,3))
+print [1,2,3] == MyList((1,2,3,4))
+print [1,2,3] != MyList((1,2,3,4))
+
+print [1,2,3,4] > MyList((1,2,3))
+print [1,2,3,4] < MyList((1,2,3))
+
+print [1,2,3] > MyList((1,2,3,4))
+print [1,2,3] < MyList((1,2,3,4))
+
+print [1,2,3] >= MyList((1,2,3))
+print [1,2,3] <= MyList((1,2,3))
+
+print MyList((1,2,3)) == MyList((1,2,3,4))
+print MyList((1,2,3)) != MyList((1,2,3,4))
+
+print MyList((1,2,3,4)) > MyList((1,2,3))
+print MyList((1,2,3,4)) < MyList((1,2,3))
+
+print MyList((1,2,3)) > MyList((1,2,3,4))
+print MyList((1,2,3)) < MyList((1,2,3,4))
+
+print MyList((1,2,3)) >= MyList((1,2,3))
+print MyList((1,2,3)) <= MyList((1,2,3))

tools/tester.py

@@ -157,6 +157,9 @@ def run_test(fn, check_stats, run_memcheck):
     code = p.wait()
     elapsed = time.time() - start
 
+    if code >= 128:
+        code -= 256
+
     return determine_test_result(fn, opts, code, out, stderr, elapsed)
 
 def get_test_options(fn, check_stats, run_memcheck):