Merge branch 'master' into HEAD

fb62db02 · Kevin Modzelewski · f2632e45 · 0defcec2 · fb62db02 · fb62db02
Commit fb62db02 authored May 07, 2015 by Kevin Modzelewski
50 changed files
--- a/Makefile
+++ b/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)),)
@@ -779,11 +792,16 @@ endef
 PASS_SRCS := codegen/opt/aa.cpp
 PASS_OBJS := $(PASS_SRCS:.cpp=.standalone.o)

+ifneq ($(USE_CMAKE),1)
 $(call make_compile_config,,$(CXXFLAGS_DBG))
 $(call make_compile_config,.release,$(CXXFLAGS_RELEASE))
 $(call make_compile_config,.grwl,$(CXXFLAGS_RELEASE) -DTHREADING_USE_GRWL=1 -DTHREADING_USE_GIL=0 -UBINARY_SUFFIX -DBINARY_SUFFIX=_grwl)
 $(call make_compile_config,.grwl_dbg,$(CXXFLAGS_DBG) -DTHREADING_USE_GRWL=1 -DTHREADING_USE_GIL=0 -UBINARY_SUFFIX -DBINARY_SUFFIX=_grwl_dbg -UBINARY_STRIPPED_SUFFIX -DBINARY_STRIPPED_SUFFIX=)
 $(call make_compile_config,.nosync,$(CXXFLAGS_RELEASE) -DTHREADING_USE_GRWL=0 -DTHREADING_USE_GIL=0 -UBINARY_SUFFIX -DBINARY_SUFFIX=_nosync)
+else
+%.o: %.cpp $(CMAKE_SETUP_DBG)
+	$(NINJA) -C $(HOME)/pyston-build-dbg src/CMakeFiles/PYSTON_OBJECTS.dir/$(patsubst src/%.o,%.cpp.o,$@) $(NINJAFLAGS)
+endif

 $(UNITTEST_SRCS:.cpp=.o): CXXFLAGS += -isystem $(GTEST_DIR)/include


--- a/from_cpython/Include/pyconfig.h
+++ b/from_cpython/Include/pyconfig.h
@@ -61,6 +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"


--- a/minibenchmarks/django-template.py
+++ b/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,)
--- a/src/analysis/function_analysis.cpp
+++ b/src/analysis/function_analysis.cpp
@@ -76,14 +76,21 @@ private:
        }
    }

+    Status::Usage getStatusFirst(InternedString name) const {
+        auto it = statuses.find(name);
+        if (it == statuses.end())
+            return Status::NONE;
+        return it->second.first;
+    }
+
 public:
    LivenessBBVisitor(LivenessAnalysis* analysis) : analysis(analysis) {}

-    bool firstIsUse(InternedString name) { return statuses[name].first == Status::USED; }
+    bool firstIsUse(InternedString name) const { return getStatusFirst(name) == Status::USED; }

-    bool firstIsDef(InternedString name) { return statuses[name].first == Status::DEFINED; }
+    bool firstIsDef(InternedString name) const { return getStatusFirst(name) == Status::DEFINED; }

-    bool isKilledAt(AST_Name* node, bool is_live_at_end) {
+    bool isKilledAt(AST_Name* node, bool is_live_at_end) const {
        if (kills.count(node))
            return true;

@@ -163,6 +170,9 @@ LivenessAnalysis::LivenessAnalysis(CFG* cfg) : cfg(cfg) {
    us_liveness.log(_t.end());
 }

+LivenessAnalysis::~LivenessAnalysis() {
+}
+
 bool LivenessAnalysis::isKill(AST_Name* node, CFGBlock* parent_block) {
    if (node->id.str()[0] != '#')
        return false;
@@ -487,11 +497,18 @@ bool PhiAnalysis::isPotentiallyUndefinedAt(InternedString name, CFGBlock* block)
    return definedness.defined_at_beginning[block][name] != DefinednessAnalysis::Defined;
 }

-LivenessAnalysis* computeLivenessInfo(CFG* cfg) {
-    return new LivenessAnalysis(cfg);
+std::unique_ptr<LivenessAnalysis> computeLivenessInfo(CFG* cfg) {
+    static StatCounter counter("num_liveness_analysis");
+    counter.log();
+
+    return std::unique_ptr<LivenessAnalysis>(new LivenessAnalysis(cfg));
 }

-PhiAnalysis* computeRequiredPhis(const ParamNames& args, CFG* cfg, LivenessAnalysis* liveness, ScopeInfo* scope_info) {
+std::unique_ptr<PhiAnalysis> computeRequiredPhis(const ParamNames& args, CFG* cfg, LivenessAnalysis* liveness,
+                                                 ScopeInfo* scope_info) {
+    static StatCounter counter("num_phi_analysis");
+    counter.log();
+
    llvm::DenseMap<InternedString, DefinednessAnalysis::DefinitionLevel> initial_map;

    for (auto e : args.args)
@@ -501,11 +518,15 @@ PhiAnalysis* computeRequiredPhis(const ParamNames& args, CFG* cfg, LivenessAnaly
    if (args.kwarg.size())
        initial_map[scope_info->internString(args.kwarg)] = DefinednessAnalysis::Defined;

-    return new PhiAnalysis(std::move(initial_map), cfg->getStartingBlock(), false, liveness, scope_info);
+    return std::unique_ptr<PhiAnalysis>(
+        new PhiAnalysis(std::move(initial_map), cfg->getStartingBlock(), false, liveness, scope_info));
 }

-PhiAnalysis* computeRequiredPhis(const OSREntryDescriptor* entry_descriptor, LivenessAnalysis* liveness,
-                                 ScopeInfo* scope_info) {
+std::unique_ptr<PhiAnalysis> computeRequiredPhis(const OSREntryDescriptor* entry_descriptor, LivenessAnalysis* liveness,
+                                                 ScopeInfo* scope_info) {
+    static StatCounter counter("num_phi_analysis");
+    counter.log();
+
    llvm::DenseMap<InternedString, DefinednessAnalysis::DefinitionLevel> initial_map;

    llvm::StringSet<> potentially_undefined;
@@ -524,6 +545,7 @@ PhiAnalysis* computeRequiredPhis(const OSREntryDescriptor* entry_descriptor, Liv
            initial_map[p.first] = DefinednessAnalysis::Defined;
    }

-    return new PhiAnalysis(std::move(initial_map), entry_descriptor->backedge->target, true, liveness, scope_info);
+    return std::unique_ptr<PhiAnalysis>(
+        new PhiAnalysis(std::move(initial_map), entry_descriptor->backedge->target, true, liveness, scope_info));
 }
 }
--- a/src/analysis/function_analysis.h
+++ b/src/analysis/function_analysis.h
@@ -45,6 +45,7 @@ private:

 public:
    LivenessAnalysis(CFG* cfg);
+    ~LivenessAnalysis();

    // we don't keep track of node->parent_block relationships, so you have to pass both:
    bool isKill(AST_Name* node, CFGBlock* parent_block);
@@ -105,9 +106,9 @@ public:
    bool isPotentiallyUndefinedAt(InternedString name, CFGBlock* block);
 };

-LivenessAnalysis* computeLivenessInfo(CFG*);
-PhiAnalysis* computeRequiredPhis(const ParamNames&, CFG*, LivenessAnalysis*, ScopeInfo* scope_info);
-PhiAnalysis* computeRequiredPhis(const OSREntryDescriptor*, LivenessAnalysis*, ScopeInfo* scope_info);
+std::unique_ptr<LivenessAnalysis> computeLivenessInfo(CFG*);
+std::unique_ptr<PhiAnalysis> computeRequiredPhis(const ParamNames&, CFG*, LivenessAnalysis*, ScopeInfo* scope_info);
+std::unique_ptr<PhiAnalysis> computeRequiredPhis(const OSREntryDescriptor*, LivenessAnalysis*, ScopeInfo* scope_info);
 }

 #endif
--- a/src/capi/descrobject.cpp
+++ b/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);

--- a/src/capi/typeobject.cpp
+++ b/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)

--- a/src/capi/types.h
+++ b/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);

--- a/src/codegen/ast_interpreter.cpp
+++ b/src/codegen/ast_interpreter.cpp
@@ -83,7 +83,6 @@ private:
    Value doBinOp(Box* left, Box* right, int op, BinExpType exp_type);
    void doStore(AST_expr* node, Value value);
    void doStore(InternedString name, Value value);
-    void eraseDeadSymbols();

    Value visit_assert(AST_Assert* node);
    Value visit_assign(AST_Assign* node);
@@ -285,6 +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);

@@ -321,31 +322,6 @@ Value ASTInterpreter::execute(ASTInterpreter& interpreter, CFGBlock* start_block
    return v;
 }

-void ASTInterpreter::eraseDeadSymbols() {
-    if (source_info->liveness == NULL)
-        source_info->liveness = computeLivenessInfo(source_info->cfg);
-
-    if (this->phis == NULL) {
-        PhiAnalysis*& phis = source_info->phis[/* entry_descriptor = */ NULL];
-        if (!phis)
-            phis = computeRequiredPhis(compiled_func->clfunc->param_names, source_info->cfg, source_info->liveness,
-                                       scope_info);
-        this->phis = phis;
-    }
-
-    std::vector<InternedString> dead_symbols;
-    for (auto& it : sym_table) {
-        if (!source_info->liveness->isLiveAtEnd(it.first, current_block)) {
-            dead_symbols.push_back(it.first);
-        } else if (phis->isRequiredAfter(it.first, current_block)) {
-            assert(scope_info->getScopeTypeOfName(it.first) != ScopeInfo::VarScopeType::GLOBAL);
-        } else {
-        }
-    }
-    for (auto&& dead : dead_symbols)
-        sym_table.erase(dead);
-}
-
 Value ASTInterpreter::doBinOp(Box* left, Box* right, int op, BinExpType exp_type) {
    if (op == AST_TYPE::Div && (source_info->parent_module->future_flags & FF_DIVISION)) {
        op = AST_TYPE::TrueDiv;
@@ -463,7 +439,26 @@ Value ASTInterpreter::visit_jump(AST_Jump* node) {
    if (ENABLE_OSR && backedge && (globals->cls == module_cls)) {
        bool can_osr = !FORCE_INTERPRETER && (globals->cls == module_cls);
        if (can_osr && edgecount++ == OSR_THRESHOLD_INTERPRETER) {
-            eraseDeadSymbols();
+            static StatCounter ast_osrs("num_ast_osrs");
+            ast_osrs.log();
+
+            // TODO: we will immediately want the liveness info again in the jit, we should pass
+            // it through.
+            std::unique_ptr<LivenessAnalysis> liveness = computeLivenessInfo(source_info->cfg);
+            std::unique_ptr<PhiAnalysis> phis
+                = computeRequiredPhis(compiled_func->clfunc->param_names, source_info->cfg, liveness.get(), scope_info);
+
+            std::vector<InternedString> dead_symbols;
+            for (auto& it : sym_table) {
+                if (!liveness->isLiveAtEnd(it.first, current_block)) {
+                    dead_symbols.push_back(it.first);
+                } else if (phis->isRequiredAfter(it.first, current_block)) {
+                    assert(scope_info->getScopeTypeOfName(it.first) != ScopeInfo::VarScopeType::GLOBAL);
+                } else {
+                }
+            }
+            for (auto&& dead : dead_symbols)
+                sym_table.erase(dead);

            const OSREntryDescriptor* found_entry = nullptr;
            for (auto& p : compiled_func->clfunc->osr_versions) {
@@ -479,7 +474,7 @@ Value ASTInterpreter::visit_jump(AST_Jump* node) {

            for (auto& name : phis->definedness.getDefinedNamesAtEnd(current_block)) {
                auto it = sym_table.find(name);
-                if (!source_info->liveness->isLiveAtEnd(name, current_block))
+                if (!liveness->isLiveAtEnd(name, current_block))
                    continue;

                if (phis->isPotentiallyUndefinedAfter(name, current_block)) {
@@ -495,10 +490,14 @@ Value ASTInterpreter::visit_jump(AST_Jump* node) {
                }
            }

+            // Manually free these here, since we might not return from this scope for a long time.
+            liveness.reset(nullptr);
+            phis.reset(nullptr);
+
            // LLVM has a limit on the number of operands a machine instruction can have (~255),
            // in order to not hit the limit with the patchpoints cancel OSR when we have a high number of symbols.
            if (sorted_symbol_table.size() > 225) {
-                static StatCounter times_osr_cancel("num_osr_cancel_to_many_syms");
+                static StatCounter times_osr_cancel("num_osr_cancel_too_many_syms");
                times_osr_cancel.log();
                next_block = node->target;
                return Value();
@@ -543,6 +542,7 @@ 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),
@@ -930,6 +930,8 @@ 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");

--- a/src/codegen/codegen.cpp
+++ b/src/codegen/codegen.cpp
@@ -35,8 +35,7 @@ namespace pyston {
 DS_DEFINE_RWLOCK(codegen_rwlock);

 SourceInfo::SourceInfo(BoxedModule* m, ScopingAnalysis* scoping, AST* ast, std::vector<AST_stmt*> body, std::string fn)
-    : parent_module(m), scoping(scoping), ast(ast), cfg(NULL), liveness(NULL), fn(std::move(fn)),
-      body(std::move(body)) {
+    : parent_module(m), scoping(scoping), ast(ast), cfg(NULL), fn(std::move(fn)), body(std::move(body)) {
    assert(this->fn.size());

    switch (ast->type) {

--- a/src/codegen/entry.cpp
+++ b/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();
 }


--- a/src/codegen/irgen.cpp
+++ b/src/codegen/irgen.cpp
@@ -342,7 +342,7 @@ static void emitBBs(IRGenState* irstate, TypeAnalysis* types, const OSREntryDesc
    CompiledFunction* cf = irstate->getCurFunction();
    ConcreteCompilerType* rtn_type = irstate->getReturnType();
    // llvm::MDNode* func_info = irstate->getFuncDbgInfo();
-    PhiAnalysis* phi_analysis = source->phis[entry_descriptor];
+    PhiAnalysis* phi_analysis = irstate->getPhis();
    assert(phi_analysis);

    if (entry_descriptor != NULL)
@@ -1055,7 +1055,15 @@ CompiledFunction* doCompile(SourceInfo* source, ParamNames* param_names, const O
        computeBlockSetClosure(blocks);
    }

-    IRGenState irstate(cf, source, source->phis[entry_descriptor], param_names, getGCBuilder(), dbg_funcinfo);
+    std::unique_ptr<LivenessAnalysis> liveness = computeLivenessInfo(source->cfg);
+    std::unique_ptr<PhiAnalysis> phis;
+
+    if (entry_descriptor)
+        phis = computeRequiredPhis(entry_descriptor, liveness.get(), source->getScopeInfo());
+    else
+        phis = computeRequiredPhis(*param_names, source->cfg, liveness.get(), source->getScopeInfo());
+
+    IRGenState irstate(cf, source, std::move(liveness), std::move(phis), param_names, getGCBuilder(), dbg_funcinfo);

    emitBBs(&irstate, types, entry_descriptor, blocks);


--- a/src/codegen/irgen/hooks.cpp
+++ b/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);
@@ -228,19 +229,6 @@ CompiledFunction* compileFunction(CLFunction* f, FunctionSpecialization* spec, E
        source->cfg = computeCFG(source, source->body);
    }

-    if (effort != EffortLevel::INTERPRETED) {
-        if (source->liveness == NULL)
-            source->liveness = computeLivenessInfo(source->cfg);
-
-        PhiAnalysis*& phis = source->phis[entry_descriptor];
-        if (!phis) {
-            if (entry_descriptor)
-                phis = computeRequiredPhis(entry_descriptor, source->liveness, source->getScopeInfo());
-            else
-                phis = computeRequiredPhis(f->param_names, source->cfg, source->liveness, source->getScopeInfo());
-        }
-    }
-


    CompiledFunction* cf = 0;
@@ -328,10 +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) {

--- a/src/codegen/irgen/irgenerator.cpp
+++ b/src/codegen/irgen/irgenerator.cpp
@@ -42,6 +42,19 @@ extern "C" void dumpLLVM(llvm::Value* v) {
    v->dump();
 }

+IRGenState::IRGenState(CompiledFunction* cf, SourceInfo* source_info, std::unique_ptr<LivenessAnalysis> liveness,
+                       std::unique_ptr<PhiAnalysis> phis, ParamNames* param_names, GCBuilder* gc,
+                       llvm::MDNode* func_dbg_info)
+    : cf(cf), source_info(source_info), liveness(std::move(liveness)), phis(std::move(phis)), param_names(param_names),
+      gc(gc), func_dbg_info(func_dbg_info), scratch_space(NULL), frame_info(NULL), frame_info_arg(NULL),
+      scratch_size(0) {
+    assert(cf->func);
+    assert(!cf->clfunc); // in this case don't need to pass in sourceinfo
+}
+
+IRGenState::~IRGenState() {
+}
+
 llvm::Value* IRGenState::getScratchSpace(int min_bytes) {
    llvm::BasicBlock& entry_block = getLLVMFunction()->getEntryBlock();

@@ -944,7 +957,7 @@ private:
    CompilerVariable* evalName(AST_Name* node, UnwindInfo unw_info) {
        auto scope_info = irstate->getScopeInfo();

-        bool is_kill = irstate->getSourceInfo()->liveness->isKill(node, myblock);
+        bool is_kill = irstate->getLiveness()->isKill(node, myblock);
        assert(!is_kill || node->id.str()[0] == '#');

        ScopeInfo::VarScopeType vst = scope_info->getScopeTypeOfName(node->id);
@@ -2271,7 +2284,7 @@ private:
            // ASSERT(p.first[0] != '!' || isIsDefinedName(p.first), "left a fake variable in the real
            // symbol table? '%s'", p.first.c_str());

-            if (!source->liveness->isLiveAtEnd(p.first, myblock)) {
+            if (!irstate->getLiveness()->isLiveAtEnd(p.first, myblock)) {
                // printf("%s dead at end of %d; grabbed = %d, %d vrefs\n", p.first.c_str(), myblock->idx,
                //        p.second->isGrabbed(), p.second->getVrefs());


--- a/src/codegen/irgen/irgenerator.h
+++ b/src/codegen/irgen/irgenerator.h
@@ -56,7 +56,8 @@ class IRGenState {
 private:
    CompiledFunction* cf;
    SourceInfo* source_info;
-    PhiAnalysis* phis;
+    std::unique_ptr<LivenessAnalysis> liveness;
+    std::unique_ptr<PhiAnalysis> phis;
    ParamNames* param_names;
    GCBuilder* gc;
    llvm::MDNode* func_dbg_info;
@@ -69,13 +70,9 @@ private:


 public:
-    IRGenState(CompiledFunction* cf, SourceInfo* source_info, PhiAnalysis* phis, ParamNames* param_names, GCBuilder* gc,
-               llvm::MDNode* func_dbg_info)
-        : cf(cf), source_info(source_info), phis(phis), param_names(param_names), gc(gc), func_dbg_info(func_dbg_info),
-          scratch_space(NULL), frame_info(NULL), frame_info_arg(NULL), scratch_size(0) {
-        assert(cf->func);
-        assert(!cf->clfunc); // in this case don't need to pass in sourceinfo
-    }
+    IRGenState(CompiledFunction* cf, SourceInfo* source_info, std::unique_ptr<LivenessAnalysis> liveness,
+               std::unique_ptr<PhiAnalysis> phis, ParamNames* param_names, GCBuilder* gc, llvm::MDNode* func_dbg_info);
+    ~IRGenState();

    CompiledFunction* getCurFunction() { return cf; }

@@ -93,7 +90,8 @@ public:

    SourceInfo* getSourceInfo() { return source_info; }

-    PhiAnalysis* getPhis() { return phis; }
+    LivenessAnalysis* getLiveness() { return liveness.get(); }
+    PhiAnalysis* getPhis() { return phis.get(); }

    ScopeInfo* getScopeInfo();
    ScopeInfo* getScopeInfoForNode(AST* node);

--- a/src/codegen/parser.cpp
+++ b/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";

--- a/src/codegen/unwinding.cpp
+++ b/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) {

--- a/src/core/stats.cpp
+++ b/src/core/stats.cpp
@@ -17,13 +17,115 @@
 #include <algorithm>

 #include "core/thread_utils.h"
+#include "gc/heap.h"

 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();
+    _duration = 0;
+    _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) {
+    _duration = 0;
+    _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);
+    auto cur_duration = _duration;
+    _duration = at_time - _start_time;
+    assert(_duration > cur_duration);
+
+    Stats::log(_statid, _duration);
+
+    _duration = 0;
+    _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)) {
 }

@@ -37,7 +139,7 @@ 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;

@@ -51,11 +153,40 @@ 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;

-    printf("Stats:\n");
+    double cycles_per_us = Stats::estimateCPUFreq();
+    fprintf(stderr, "Stats:\n");
+    fprintf(stderr, "estimated_cpu_mhz: %5.5f\n", cycles_per_us);
+
+    gc::dumpHeapStatistics(0);
+
+    fprintf(stderr, "Counters:\n");

    std::vector<std::pair<std::string, int>> pairs;
    for (const auto& p : *names) {
@@ -64,10 +195,42 @@ 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)
-            printf("%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");
 }

 void Stats::endOfInit() {

--- a/src/core/stats.h
+++ b/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 +87,77 @@ 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 accumulated active duration of this timer
+    uint64_t _duration;
+
+    // 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
 }

--- a/src/core/types.h
+++ b/src/core/types.h
@@ -243,8 +243,6 @@ public:
    ScopingAnalysis* scoping;
    AST* ast;
    CFG* cfg;
-    LivenessAnalysis* liveness;
-    std::unordered_map<const OSREntryDescriptor*, PhiAnalysis*> phis;
    bool is_generator;
    std::string fn; // equivalent of code.co_filename

@@ -487,13 +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                                                                                             \

--- a/src/core/util.cpp
+++ b/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);
    }

--- a/src/core/util.h
+++ b/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);

--- a/src/gc/collector.cpp
+++ b/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)

--- a/src/gc/gc_alloc.h
+++ b/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;
 }


--- a/src/gc/heap.cpp
+++ b/src/gc/heap.cpp
@@ -185,16 +185,32 @@ struct HeapStatistics {

        void print(const char* name) const {
            if (nbytes > (1 << 20))
-                printf("%s: %ld allocations for %.1f MB\n", name, nallocs, nbytes * 1.0 / (1 << 20));
+                fprintf(stderr, "%s: %ld allocations for %.1f MB\n", name, nallocs, nbytes * 1.0 / (1 << 20));
            else if (nbytes > (1 << 10))
-                printf("%s: %ld allocations for %.1f KB\n", name, nallocs, nbytes * 1.0 / (1 << 10));
+                fprintf(stderr, "%s: %ld allocations for %.1f KB\n", name, nallocs, nbytes * 1.0 / (1 << 10));
            else
-                printf("%s: %ld allocations for %ld bytes\n", name, nallocs, nbytes);
+                fprintf(stderr, "%s: %ld allocations for %ld bytes\n", name, nallocs, nbytes);
        }
    };
+
+    bool collect_cls_stats, collect_hcls_stats;
+
+    // For use if collect_cls_stats == true:
    std::unordered_map<BoxedClass*, TypeStats> by_cls;
-    TypeStats conservative, untracked, hcls;
+
+    // For use if collect_hcls_stats == true:
+    std::unordered_map<HiddenClass*, int> hcls_uses;
+#define HCLS_ATTRS_STAT_MAX 20
+    int num_hcls_by_attrs[HCLS_ATTRS_STAT_MAX + 1];
+    int num_hcls_by_attrs_exceed;
+
+    TypeStats python, conservative, untracked, hcls, precise;
    TypeStats total;
+
+    HeapStatistics(bool collect_cls_stats, bool collect_hcls_stats)
+        : collect_cls_stats(collect_cls_stats), collect_hcls_stats(collect_hcls_stats) {
+        memset(num_hcls_by_attrs, 0, sizeof(num_hcls_by_attrs));
+    }
 };

 void addStatistic(HeapStatistics* stats, GCAllocation* al, int nbytes) {
@@ -202,11 +218,24 @@ void addStatistic(HeapStatistics* stats, GCAllocation* al, int nbytes) {
    stats->total.nbytes += nbytes;

    if (al->kind_id == GCKind::PYTHON) {
-        Box* b = (Box*)al->user_data;
-        auto& t = stats->by_cls[b->cls];
+        stats->python.nallocs++;
+        stats->python.nbytes += nbytes;

-        t.nallocs++;
-        t.nbytes += nbytes;
+        if (stats->collect_cls_stats) {
+            Box* b = (Box*)al->user_data;
+            auto& t = stats->by_cls[b->cls];
+
+            t.nallocs++;
+            t.nbytes += nbytes;
+        }
+
+        if (stats->collect_hcls_stats) {
+            Box* b = (Box*)al->user_data;
+            if (b->cls->instancesHaveHCAttrs()) {
+                HCAttrs* attrs = b->getHCAttrsPtr();
+                stats->hcls_uses[attrs->hcls]++;
+            }
+        }
    } else if (al->kind_id == GCKind::CONSERVATIVE) {
        stats->conservative.nallocs++;
        stats->conservative.nbytes += nbytes;
@@ -216,6 +245,18 @@ void addStatistic(HeapStatistics* stats, GCAllocation* al, int nbytes) {
    } else if (al->kind_id == GCKind::HIDDEN_CLASS) {
        stats->hcls.nallocs++;
        stats->hcls.nbytes += nbytes;
+
+        if (stats->collect_hcls_stats) {
+            HiddenClass* hcls = (HiddenClass*)al->user_data;
+            int numattrs = hcls->attributeArraySize();
+            if (numattrs <= HCLS_ATTRS_STAT_MAX)
+                stats->num_hcls_by_attrs[numattrs]++;
+            else
+                stats->num_hcls_by_attrs_exceed++;
+        }
+    } else if (al->kind_id == GCKind::PRECISE) {
+        stats->precise.nallocs++;
+        stats->precise.nbytes += nbytes;
    } else {
        RELEASE_ASSERT(0, "%d", (int)al->kind_id);
    }
@@ -223,27 +264,49 @@ void addStatistic(HeapStatistics* stats, GCAllocation* al, int nbytes) {



-void Heap::dumpHeapStatistics() {
+void Heap::dumpHeapStatistics(int level) {
+    bool collect_cls_stats = (level >= 1);
+    bool collect_hcls_stats = (level >= 1);
+
    threading::GLPromoteRegion _lock;

-    HeapStatistics stats;
+    fprintf(stderr, "\nCollecting heap stats for pid %d...\n", getpid());
+
+    HeapStatistics stats(collect_cls_stats, collect_hcls_stats);

    small_arena.getStatistics(&stats);
    large_arena.getStatistics(&stats);
    huge_arena.getStatistics(&stats);

+    stats.python.print("python");
    stats.conservative.print("conservative");
    stats.untracked.print("untracked");
    stats.hcls.print("hcls");
-    for (const auto& p : stats.by_cls) {
-        p.second.print(getFullNameOfClass(p.first).c_str());
+    stats.precise.print("precise");
+
+    if (collect_cls_stats) {
+        for (const auto& p : stats.by_cls) {
+            p.second.print(getFullNameOfClass(p.first).c_str());
+        }
    }
+
    stats.total.print("Total");
-    printf("\n");
+
+    if (collect_hcls_stats) {
+        fprintf(stderr, "%ld hidden classes currently alive\n", stats.hcls.nallocs);
+        fprintf(stderr, "%ld have at least one Box that uses them\n", stats.hcls_uses.size());
+
+        for (int i = 0; i <= HCLS_ATTRS_STAT_MAX; i++) {
+            fprintf(stderr, "With % 3d attributes: %d\n", i, stats.num_hcls_by_attrs[i]);
+        }
+        fprintf(stderr, "With >% 2d attributes: %d\n", HCLS_ATTRS_STAT_MAX, stats.num_hcls_by_attrs_exceed);
+    }
+
+    fprintf(stderr, "\n");
 }

-void dumpHeapStatistics() {
-    global_heap.dumpHeapStatistics();
+void dumpHeapStatistics(int level) {
+    global_heap.dumpHeapStatistics(level);
 }

 //////

--- a/src/gc/heap.h
+++ b/src/gc/heap.h
@@ -22,6 +22,7 @@

 #include "core/common.h"
 #include "core/threading.h"
+#include "core/types.h"

 namespace pyston {

@@ -535,13 +536,13 @@ public:
        huge_arena.freeUnmarked(weakly_referenced);
    }

-    void dumpHeapStatistics();
+    void dumpHeapStatistics(int level);

    friend void markPhase();
 };

 extern Heap global_heap;
-void dumpHeapStatistics();
+void dumpHeapStatistics(int level);

 } // namespace gc
 } // namespace pyston

--- a/src/jit.cpp
+++ b/src/jit.cpp
--- a/src/runtime/builtin_modules/builtins.cpp
+++ b/src/runtime/builtin_modules/builtins.cpp
@@ -304,6 +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;
@@ -347,6 +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();
@@ -359,6 +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();
@@ -366,6 +369,7 @@ 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();

--- a/src/runtime/descr.cpp
+++ b/src/runtime/descr.cpp
@@ -19,7 +19,7 @@
 namespace pyston {

 static Box* memberGet(BoxedMemberDescriptor* self, Box* inst, Box* owner) {
-    RELEASE_ASSERT(self->cls == member_cls, "");
+    RELEASE_ASSERT(self->cls == member_descriptor_cls, "");

    if (inst == None)
        return self;
@@ -169,8 +169,8 @@ static Box* classmethodGet(Box* self, Box* obj, Box* type) {
 }

 void setupDescr() {
-    member_cls->giveAttr("__get__", new BoxedFunction(boxRTFunction((void*)memberGet, UNKNOWN, 3)));
-    member_cls->freeze();
+    member_descriptor_cls->giveAttr("__get__", new BoxedFunction(boxRTFunction((void*)memberGet, UNKNOWN, 3)));
+    member_descriptor_cls->freeze();

    property_cls->giveAttr("__init__",
                           new BoxedFunction(boxRTFunction((void*)propertyInit, UNKNOWN, 5, 4, false, false,

--- a/src/runtime/dict.cpp
+++ b/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));

--- a/src/runtime/float.cpp
+++ b/src/runtime/float.cpp
@@ -741,6 +741,98 @@ static Box* float0(Box*, void*) {
    return boxFloat(0.0);
 }

+// __getformat__
+// ported pretty directly from cpython Objects/floatobject.c
+
+typedef enum { unknown_format, ieee_big_endian_format, ieee_little_endian_format } float_format_type;
+
+static float_format_type double_format, float_format;
+static float_format_type detected_double_format, detected_float_format;
+
+static void floatFormatInit() {
+    /* We attempt to determine if this machine is using IEEE
+       floating point formats by peering at the bits of some
+       carefully chosen values.  If it looks like we are on an
+       IEEE platform, the float packing/unpacking routines can
+       just copy bits, if not they resort to arithmetic & shifts
+       and masks.  The shifts & masks approach works on all finite
+       values, but what happens to infinities, NaNs and signed
+       zeroes on packing is an accident, and attempting to unpack
+       a NaN or an infinity will raise an exception.
+
+       Note that if we're on some whacked-out platform which uses
+       IEEE formats but isn't strictly little-endian or big-
+       endian, we will fall back to the portable shifts & masks
+       method. */
+
+    if (sizeof(double) == 8) {
+        double x = 9006104071832581.0;
+        if (memcmp(&x, "\x43\x3f\xff\x01\x02\x03\x04\x05", 8) == 0)
+            detected_double_format = ieee_big_endian_format;
+        else if (memcmp(&x, "\x05\x04\x03\x02\x01\xff\x3f\x43", 8) == 0)
+            detected_double_format = ieee_little_endian_format;
+        else
+            detected_double_format = unknown_format;
+    } else {
+        detected_double_format = unknown_format;
+    }
+
+    if (sizeof(float) == 4) {
+        float y = 16711938.0;
+        if (memcmp(&y, "\x4b\x7f\x01\x02", 4) == 0)
+            detected_float_format = ieee_big_endian_format;
+        else if (memcmp(&y, "\x02\x01\x7f\x4b", 4) == 0)
+            detected_float_format = ieee_little_endian_format;
+        else
+            detected_float_format = unknown_format;
+    } else {
+        detected_float_format = unknown_format;
+    }
+
+    double_format = detected_double_format;
+    float_format = detected_float_format;
+}
+
+// ported pretty directly from cpython
+Box* floatGetFormat(BoxedClass* v, Box* arg) {
+    char* s;
+    float_format_type r;
+
+    if (!PyString_Check(arg)) {
+        raiseExcHelper(TypeError, "__getformat__() argument must be string, not %s", arg->cls->tp_name);
+    }
+    s = PyString_AS_STRING(arg);
+    if (strcmp(s, "double") == 0) {
+        r = double_format;
+    } else if (strcmp(s, "float") == 0) {
+        r = float_format;
+    } else {
+        raiseExcHelper(ValueError, "__getformat__() argument 1 must be "
+                                   "'double' or 'float'");
+    }
+
+    switch (r) {
+        case unknown_format:
+            return boxString("unknown");
+        case ieee_little_endian_format:
+            return boxString("IEEE, little-endian");
+        case ieee_big_endian_format:
+            return boxString("IEEE, big-endian");
+        default:
+            RELEASE_ASSERT(false, "insane float_format or double_format");
+            return NULL;
+    }
+}
+
+const char* floatGetFormatDoc = "float.__getformat__(typestr) -> string\n"
+                                "\n"
+                                "You probably don't want to use this function.  It exists mainly to be\n"
+                                "used in Python's test suite.\n"
+                                "\n"
+                                "typestr must be 'double' or 'float'.  This function returns whichever of\n"
+                                "'unknown', 'IEEE, big-endian' or 'IEEE, little-endian' best describes the\n"
+                                "format of floating point numbers used by the C type named by typestr.";
+
 void setupFloat() {
    _addFunc("__add__", BOXED_FLOAT, (void*)floatAddFloat, (void*)floatAddInt, (void*)floatAdd);
    float_cls->giveAttr("__radd__", float_cls->getattr("__add__"));
@@ -785,7 +877,13 @@ void setupFloat() {
    float_cls->giveAttr("imag", new (pyston_getset_cls) BoxedGetsetDescriptor(float0, NULL, NULL));
    float_cls->giveAttr("conjugate", new BoxedFunction(boxRTFunction((void*)floatFloat, BOXED_FLOAT, 1)));

+    float_cls->giveAttr("__getformat__",
+                        new BoxedClassmethod(new BoxedBuiltinFunctionOrMethod(
+                            boxRTFunction((void*)floatGetFormat, STR, 2), "__getformat__", floatGetFormatDoc)));
+
    float_cls->freeze();
+
+    floatFormatInit();
 }

 void teardownFloat() {

--- a/src/runtime/generator.cpp
+++ b/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

--- a/src/runtime/list.cpp
+++ b/src/runtime/list.cpp
@@ -541,6 +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));
    }
@@ -940,7 +942,7 @@ Box* _listCmp(BoxedList* lhs, BoxedList* rhs, AST_TYPE::AST_TYPE op_type) {
 }

 Box* listEq(BoxedList* self, Box* rhs) {
-    if (rhs->cls != list_cls) {
+    if (!isSubclass(rhs->cls, list_cls)) {
        return NotImplemented;
    }

@@ -950,7 +952,7 @@ Box* listEq(BoxedList* self, Box* rhs) {
 }

 Box* listNe(BoxedList* self, Box* rhs) {
-    if (rhs->cls != list_cls) {
+    if (!isSubclass(rhs->cls, list_cls)) {
        return NotImplemented;
    }

@@ -960,7 +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;
    }

@@ -970,7 +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;
    }

@@ -980,7 +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;
    }

@@ -990,7 +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;
    }


--- a/src/runtime/objmodel.cpp
+++ b/src/runtime/objmodel.cpp
--- a/src/runtime/stacktrace.cpp
+++ b/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);
@@ -160,28 +161,31 @@ extern "C" void abort() {
    // In case displaying the traceback recursively calls abort:
    static bool recursive = false;

-    // If traceback_cls is NULL, then we somehow died early on, and won't be able to display a traceback.
-    if (!recursive && traceback_cls) {
+    if (!recursive) {
        recursive = true;
-
+        Stats::dump();
        fprintf(stderr, "Someone called abort!\n");

-        // If we call abort(), things may be seriously wrong.  Set an alarm() to
-        // try to handle cases that we would just hang.
-        // (Ex if we abort() from a static constructor, and _printStackTrace uses
-        // that object, _printStackTrace will hang waiting for the first construction
-        // to finish.)
-        alarm(1);
-        try {
-            _printStacktrace();
-        } catch (ExcInfo) {
-            fprintf(stderr, "error printing stack trace during abort()");
-        }
+        // If traceback_cls is NULL, then we somehow died early on, and won't be able to display a traceback.
+        if (traceback_cls) {
+
+            // If we call abort(), things may be seriously wrong.  Set an alarm() to
+            // try to handle cases that we would just hang.
+            // (Ex if we abort() from a static constructor, and _printStackTrace uses
+            // that object, _printStackTrace will hang waiting for the first construction
+            // to finish.)
+            alarm(1);
+            try {
+                _printStacktrace();
+            } catch (ExcInfo) {
+                fprintf(stderr, "error printing stack trace during abort()");
+            }

-        // Cancel the alarm.
-        // This is helpful for when running in a debugger, since otherwise the debugger will catch the
-        // abort and let you investigate, but the alarm will still come back to kill the program.
-        alarm(0);
+            // Cancel the alarm.
+            // This is helpful for when running in a debugger, since otherwise the debugger will catch the
+            // abort and let you investigate, but the alarm will still come back to kill the program.
+            alarm(0);
+        }
    }

    if (PAUSE_AT_ABORT) {

--- a/src/runtime/str.cpp
+++ b/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;

--- a/src/runtime/tuple.cpp
+++ b/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;

--- a/src/runtime/types.cpp
+++ b/src/runtime/types.cpp
@@ -655,7 +655,7 @@ extern "C" void closureGCHandler(GCVisitor* v, Box* b) {
 extern "C" {
 BoxedClass* object_cls, *type_cls, *none_cls, *bool_cls, *int_cls, *float_cls,
    * str_cls = NULL, *function_cls, *instancemethod_cls, *list_cls, *slice_cls, *module_cls, *dict_cls, *tuple_cls,
-      *file_cls, *member_cls, *closure_cls, *generator_cls, *complex_cls, *basestring_cls, *property_cls,
+      *file_cls, *member_descriptor_cls, *closure_cls, *generator_cls, *complex_cls, *basestring_cls, *property_cls,
      *staticmethod_cls, *classmethod_cls, *attrwrapper_cls, *pyston_getset_cls, *capi_getset_cls,
      *builtin_function_or_method_cls, *attrwrapperiter_cls, *set_cls, *frozenset_cls;

@@ -1270,6 +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);

@@ -1282,6 +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);

@@ -1297,6 +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);

@@ -1311,6 +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);

@@ -1344,6 +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);

@@ -1409,6 +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);

@@ -1423,6 +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);

@@ -1461,6 +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);

@@ -2101,8 +2110,8 @@ void setupRuntime() {

    module_cls = new BoxedHeapClass(object_cls, &moduleGCHandler, offsetof(BoxedModule, attrs), 0, sizeof(BoxedModule),
                                    false, static_cast<BoxedString*>(boxStrConstant("module")));
-    member_cls = new BoxedHeapClass(object_cls, NULL, 0, 0, sizeof(BoxedMemberDescriptor), false,
-                                    static_cast<BoxedString*>(boxStrConstant("member")));
+    member_descriptor_cls = new BoxedHeapClass(object_cls, NULL, 0, 0, sizeof(BoxedMemberDescriptor), false,
+                                               static_cast<BoxedString*>(boxStrConstant("member_descriptor")));
    capifunc_cls = new BoxedHeapClass(object_cls, NULL, 0, 0, sizeof(BoxedCApiFunction), false,
                                      static_cast<BoxedString*>(boxStrConstant("capifunc")));
    method_cls = new BoxedHeapClass(object_cls, NULL, 0, 0, sizeof(BoxedMethodDescriptor), false,
@@ -2132,7 +2141,7 @@ void setupRuntime() {
    float_cls->tp_mro = BoxedTuple::create({ float_cls, object_cls });
    function_cls->tp_mro = BoxedTuple::create({ function_cls, object_cls });
    builtin_function_or_method_cls->tp_mro = BoxedTuple::create({ builtin_function_or_method_cls, object_cls });
-    member_cls->tp_mro = BoxedTuple::create({ member_cls, object_cls });
+    member_descriptor_cls->tp_mro = BoxedTuple::create({ member_descriptor_cls, object_cls });
    capifunc_cls->tp_mro = BoxedTuple::create({ capifunc_cls, object_cls });
    module_cls->tp_mro = BoxedTuple::create({ module_cls, object_cls });
    method_cls->tp_mro = BoxedTuple::create({ method_cls, object_cls });
@@ -2187,7 +2196,7 @@ void setupRuntime() {
    float_cls->finishInitialization();
    function_cls->finishInitialization();
    builtin_function_or_method_cls->finishInitialization();
-    member_cls->finishInitialization();
+    member_descriptor_cls->finishInitialization();
    module_cls->finishInitialization();
    capifunc_cls->finishInitialization();
    method_cls->finishInitialization();

--- a/src/runtime/types.h
+++ b/src/runtime/types.h
@@ -85,9 +85,9 @@ extern "C" Box* getSysStdout();
 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,
-    *enumerate_cls, *xrange_cls, *member_cls, *method_cls, *closure_cls, *generator_cls, *complex_cls, *basestring_cls,
-    *property_cls, *staticmethod_cls, *classmethod_cls, *attrwrapper_cls, *pyston_getset_cls, *capi_getset_cls,
-    *builtin_function_or_method_cls, *set_cls, *frozenset_cls, *code_cls;
+    *enumerate_cls, *xrange_cls, *member_descriptor_cls, *method_cls, *closure_cls, *generator_cls, *complex_cls,
+    *basestring_cls, *property_cls, *staticmethod_cls, *classmethod_cls, *attrwrapper_cls, *pyston_getset_cls,
+    *capi_getset_cls, *builtin_function_or_method_cls, *set_cls, *frozenset_cls, *code_cls;
 }
 #define unicode_cls (&PyUnicode_Type)
 #define memoryview_cls (&PyMemoryView_Type)
@@ -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;
@@ -732,7 +788,7 @@ public:
    BoxedMemberDescriptor(PyMemberDef* member)
        : type((MemberType)member->type), offset(member->offset), readonly(member->flags & READONLY) {}

-    DEFAULT_CLASS_SIMPLE(member_cls);
+    DEFAULT_CLASS_SIMPLE(member_descriptor_cls);
 };

 class BoxedGetsetDescriptor : public Box {
@@ -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);

--- a/test/cpython/exception_hierarchy.txt
+++ b/test/cpython/exception_hierarchy.txt
+BaseException
+ +-- SystemExit
+ +-- KeyboardInterrupt
+ +-- GeneratorExit
+ +-- Exception
+      +-- StopIteration
+      +-- StandardError
+      |    +-- BufferError
+      |    +-- ArithmeticError
+      |    |    +-- FloatingPointError
+      |    |    +-- OverflowError
+      |    |    +-- ZeroDivisionError
+      |    +-- AssertionError
+      |    +-- AttributeError
+      |    +-- EnvironmentError
+      |    |    +-- IOError
+      |    |    +-- OSError
+      |    |         +-- WindowsError (Windows)
+      |    |         +-- VMSError (VMS)
+      |    +-- EOFError
+      |    +-- ImportError
+      |    +-- LookupError
+      |    |    +-- IndexError
+      |    |    +-- KeyError
+      |    +-- MemoryError
+      |    +-- NameError
+      |    |    +-- UnboundLocalError
+      |    +-- ReferenceError
+      |    +-- RuntimeError
+      |    |    +-- NotImplementedError
+      |    +-- SyntaxError
+      |    |    +-- IndentationError
+      |    |         +-- TabError
+      |    +-- SystemError
+      |    +-- TypeError
+      |    +-- ValueError
+      |         +-- UnicodeError
+      |              +-- UnicodeDecodeError
+      |              +-- UnicodeEncodeError
+      |              +-- UnicodeTranslateError
+      +-- Warning
+           +-- DeprecationWarning
+           +-- PendingDeprecationWarning
+           +-- RuntimeWarning
+           +-- SyntaxWarning
+           +-- UserWarning
+           +-- FutureWarning
+	   +-- ImportWarning
+	   +-- UnicodeWarning
+	   +-- BytesWarning
--- a/test/tests/exec_syntax_error.py
+++ b/test/tests/exec_syntax_error.py
-
 s = """
 def f():
    a = 1
@@ -11,7 +10,8 @@ f()
 try:
    exec s
 except Exception as e:
-    print repr(e)
+    # avoid microrevision changes to Python error messages
+    print repr(e).replace("because ", "")

 s = """
 def f():
@@ -27,4 +27,5 @@ f()
 try:
    exec s
 except Exception as e:
-    print repr(e)
+    # avoid microrevision changes to Python error messages
+    print repr(e).replace("because ", "")
--- a/test/tests/float.py
+++ b/test/tests/float.py
@@ -50,3 +50,11 @@ except ValueError as e:
    pass
    # We don't print the right thing yet:
    # print e
+
+print '__getformat__ test'
+print float.__getformat__('double')
+print float.__getformat__('float')
+try:
+    float.__getformat__('oooga booga boooga')
+except Exception as e:
+    print e.message
--- a/test/tests/list_subclassing.py
+++ b/test/tests/list_subclassing.py
@@ -12,3 +12,27 @@ print l
 print len(MyList.__new__(MyList))
 l[:] = l[:]
 print l
+
+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))
--- a/test/tests/name_forcing_syntax_error.py
+++ b/test/tests/name_forcing_syntax_error.py
@@ -208,6 +208,6 @@ for case in cases:
    try:
        exec case
    except SyntaxError as se:
-        print se.message
+        print se.message.replace("because ", "")
        # TODO uncomment this
        # traceback.print_exc()
--- a/test/tests/package_test.py
+++ b/test/tests/package_test.py
 import os

 import test_package
-print 1, test_package.__name__, os.path.normpath(test_package.__file__)
+print 1, test_package.__name__, os.path.normpath(test_package.__file__).replace(".pyc", ".py")

 import test_package.intrapackage_import
 import test_package.absolute_import

--- a/test/tests/test_package/absolute_import.py
+++ b/test/tests/test_package/absolute_import.py
 import os
 import import_target

-print 3, import_target.__name__, os.path.normpath(import_target.__file__)
+print 3, import_target.__name__, os.path.normpath(import_target.__file__).replace(".pyc", ".py")
--- a/test/tests/test_package/absolute_import_with_future.py
+++ b/test/tests/test_package/absolute_import_with_future.py
@@ -2,4 +2,4 @@ from __future__ import absolute_import
 import os
 import import_target

-print 4, import_target.__name__, os.path.normpath(import_target.__file__)
+print 4, import_target.__name__, os.path.normpath(import_target.__file__).replace(".pyc", ".py")
--- a/test/tests/test_package/intrapackage_import.py
+++ b/test/tests/test_package/intrapackage_import.py
 import os
 from . import import_target
-print 2, import_target.__name__, os.path.normpath(import_target.__file__)
+print 2, import_target.__name__, os.path.normpath(import_target.__file__).replace(".pyc", ".py")
--- a/test/unittests/analysis.cpp
+++ b/test/unittests/analysis.cpp
@@ -39,7 +39,7 @@ TEST_F(AnalysisTest, augassign) {
    SourceInfo* si = new SourceInfo(createModule("augassign", fn), scoping, func, func->body, fn);

    CFG* cfg = computeCFG(si, func->body);
-    LivenessAnalysis* liveness = computeLivenessInfo(cfg);
+    std::unique_ptr<LivenessAnalysis> liveness = computeLivenessInfo(cfg);

    //cfg->print();

@@ -49,7 +49,7 @@ TEST_F(AnalysisTest, augassign) {
            ASSERT_TRUE(liveness->isLiveAtEnd(module->interned_strings->get("a"), block));
    }

-    PhiAnalysis* phis = computeRequiredPhis(ParamNames(func), cfg, liveness, scope_info);
+    std::unique_ptr<PhiAnalysis> phis = computeRequiredPhis(ParamNames(func), cfg, liveness.get(), scope_info);
 }

 void doOsrTest(bool is_osr, bool i_maybe_undefined) {
@@ -67,7 +67,7 @@ void doOsrTest(bool is_osr, bool i_maybe_undefined) {
            scoping, func, func->body, fn);

    CFG* cfg = computeCFG(si, func->body);
-    LivenessAnalysis* liveness = computeLivenessInfo(cfg);
+    std::unique_ptr<LivenessAnalysis> liveness = computeLivenessInfo(cfg);

    // cfg->print();

@@ -83,7 +83,7 @@ void doOsrTest(bool is_osr, bool i_maybe_undefined) {
    AST_Jump* backedge = ast_cast<AST_Jump>(loop_backedge->body[0]);
    ASSERT_LE(backedge->target->idx, loop_backedge->idx);

-    PhiAnalysis* phis;
+    std::unique_ptr<PhiAnalysis> phis;

    if (is_osr) {
        OSREntryDescriptor* entry_descriptor = OSREntryDescriptor::create(NULL, backedge);
@@ -91,9 +91,9 @@ void doOsrTest(bool is_osr, bool i_maybe_undefined) {
        if (i_maybe_undefined)
            entry_descriptor->args[idi_str] = NULL;
        entry_descriptor->args[iter_str] = NULL;
-        phis = computeRequiredPhis(entry_descriptor, liveness, scope_info);
+        phis = computeRequiredPhis(entry_descriptor, liveness.get(), scope_info);
    } else {
-        phis = computeRequiredPhis(ParamNames(func), cfg, liveness, scope_info);
+        phis = computeRequiredPhis(ParamNames(func), cfg, liveness.get(), scope_info);
    }

    // First, verify that we require phi nodes for the block we enter into.

--- a/tools/tester.py
+++ b/tools/tester.py
@@ -125,6 +125,8 @@ def canonicalize_stderr(stderr):
            ("AttributeError: '(\w+)' object attribute '(\w+)' is read-only", "AttributeError: \\2"),
            (r"TypeError: object.__new__\(\) takes no parameters", "TypeError: object() takes no parameters"),
            ("IndexError: list assignment index out of range", "IndexError: list index out of range"),
+            (r"unqualified exec is not allowed in function '(\w+)' it (.*)",
+             r"unqualified exec is not allowed in function '\1' because it \2"),
            ]

    for pattern, subst_with in substitutions:
@@ -206,9 +208,6 @@ def get_test_options(fn, check_stats, run_memcheck):
        elif os.path.basename(fn).split('.')[0] in TESTS_TO_SKIP:
            opts.skip = 'command line option'

-    if opts.collect_stats:
-        opts.jit_args = ['-s'] + opts.jit_args
-
    assert opts.expected in ("success", "fail", "statfail"), opts.expected

    if TEST_PYPY:
@@ -217,11 +216,12 @@ def get_test_options(fn, check_stats, run_memcheck):
        opts.check_stats = False
        opts.expected = "success"

+    if opts.collect_stats:
+        opts.jit_args = ['-s'] + opts.jit_args
+
    return opts

 def determine_test_result(fn, opts, code, out, stderr, elapsed):
-    last_stderr_line = stderr.strip().split('\n')[-1]
-
    if opts.allow_warnings:
        out_lines = []
        for l in out.split('\n'):
@@ -233,13 +233,26 @@ def determine_test_result(fn, opts, code, out, stderr, elapsed):
        out = "\n".join(out_lines)

    stats = None
-    if code >= 0 and opts.collect_stats:
+    if opts.collect_stats:
        stats = {}
-        assert out.count("Stats:") == 1
-        out, stats_str = out.split("Stats:")
-        for l in stats_str.strip().split('\n'):
-            k, v = l.split(':')
-            stats[k.strip()] = int(v)
+        have_stats = (stderr.count("Stats:") == 1 and stderr.count("(End of stats)") == 1)
+
+        if code >= 0:
+            assert have_stats
+
+        if have_stats:
+            assert stderr.count("Stats:") == 1
+            stderr, stats_str = stderr.split("Stats:")
+            stats_str, stderr_tail = stats_str.split("(End of stats)\n")
+            stderr += stderr_tail
+
+            other_stats_str, counter_str = stats_str.split("Counters:")
+            for l in counter_str.strip().split('\n'):
+                assert l.count(':') == 1, l
+                k, v = l.split(':')
+                stats[k.strip()] = int(v)
+
+    last_stderr_line = stderr.strip().split('\n')[-1]

    if EXIT_CODE_ONLY:
        # fools the rest of this function into thinking the output is OK & just checking the exit code.