[pypy-commit] pypy reflex-support: merge default

wlav noreply at buildbot.pypy.org
Fri Jul 1 01:15:52 CEST 2011


Author: Wim Lavrijsen <WLavrijsen at lbl.gov>
Branch: reflex-support
Changeset: r45199:33ebb9235c26
Date: 2011-06-29 16:35 -0700
http://bitbucket.org/pypy/pypy/changeset/33ebb9235c26/

Log:	merge default

diff --git a/_pytest/__init__.py b/_pytest/__init__.py
--- a/_pytest/__init__.py
+++ b/_pytest/__init__.py
@@ -1,2 +1,2 @@
 #
-__version__ = '2.0.3'
+__version__ = '2.1.0.dev4'
diff --git a/_pytest/assertion.py b/_pytest/assertion.py
deleted file mode 100644
--- a/_pytest/assertion.py
+++ /dev/null
@@ -1,177 +0,0 @@
-"""
-support for presented detailed information in failing assertions.
-"""
-import py
-import sys
-from _pytest.monkeypatch import monkeypatch
-
-def pytest_addoption(parser):
-    group = parser.getgroup("debugconfig")
-    group._addoption('--no-assert', action="store_true", default=False,
-        dest="noassert",
-        help="disable python assert expression reinterpretation."),
-
-def pytest_configure(config):
-    # The _reprcompare attribute on the py.code module is used by
-    # py._code._assertionnew to detect this plugin was loaded and in
-    # turn call the hooks defined here as part of the
-    # DebugInterpreter.
-    m = monkeypatch()
-    config._cleanup.append(m.undo)
-    warn_about_missing_assertion()
-    if not config.getvalue("noassert") and not config.getvalue("nomagic"):
-        def callbinrepr(op, left, right):
-            hook_result = config.hook.pytest_assertrepr_compare(
-                config=config, op=op, left=left, right=right)
-            for new_expl in hook_result:
-                if new_expl:
-                    return '\n~'.join(new_expl)
-        m.setattr(py.builtin.builtins,
-                  'AssertionError', py.code._AssertionError)
-        m.setattr(py.code, '_reprcompare', callbinrepr)
-
-def warn_about_missing_assertion():
-    try:
-        assert False
-    except AssertionError:
-        pass
-    else:
-        sys.stderr.write("WARNING: failing tests may report as passing because "
-        "assertions are turned off!  (are you using python -O?)\n")
-
-# Provide basestring in python3
-try:
-    basestring = basestring
-except NameError:
-    basestring = str
-
-
-def pytest_assertrepr_compare(op, left, right):
-    """return specialised explanations for some operators/operands"""
-    width = 80 - 15 - len(op) - 2 # 15 chars indentation, 1 space around op
-    left_repr = py.io.saferepr(left, maxsize=int(width/2))
-    right_repr = py.io.saferepr(right, maxsize=width-len(left_repr))
-    summary = '%s %s %s' % (left_repr, op, right_repr)
-
-    issequence = lambda x: isinstance(x, (list, tuple))
-    istext = lambda x: isinstance(x, basestring)
-    isdict = lambda x: isinstance(x, dict)
-    isset = lambda x: isinstance(x, set)
-
-    explanation = None
-    try:
-        if op == '==':
-            if istext(left) and istext(right):
-                explanation = _diff_text(left, right)
-            elif issequence(left) and issequence(right):
-                explanation = _compare_eq_sequence(left, right)
-            elif isset(left) and isset(right):
-                explanation = _compare_eq_set(left, right)
-            elif isdict(left) and isdict(right):
-                explanation = _diff_text(py.std.pprint.pformat(left),
-                                         py.std.pprint.pformat(right))
-        elif op == 'not in':
-            if istext(left) and istext(right):
-                explanation = _notin_text(left, right)
-    except py.builtin._sysex:
-        raise
-    except:
-        excinfo = py.code.ExceptionInfo()
-        explanation = ['(pytest_assertion plugin: representation of '
-            'details failed. Probably an object has a faulty __repr__.)',
-            str(excinfo)
-            ]
-
-
-    if not explanation:
-        return None
-
-    # Don't include pageloads of data, should be configurable
-    if len(''.join(explanation)) > 80*8:
-        explanation = ['Detailed information too verbose, truncated']
-
-    return [summary] + explanation
-
-
-def _diff_text(left, right):
-    """Return the explanation for the diff between text
-
-    This will skip leading and trailing characters which are
-    identical to keep the diff minimal.
-    """
-    explanation = []
-    i = 0 # just in case left or right has zero length
-    for i in range(min(len(left), len(right))):
-        if left[i] != right[i]:
-            break
-    if i > 42:
-        i -= 10                 # Provide some context
-        explanation = ['Skipping %s identical '
-                       'leading characters in diff' % i]
-        left = left[i:]
-        right = right[i:]
-    if len(left) == len(right):
-        for i in range(len(left)):
-            if left[-i] != right[-i]:
-                break
-        if i > 42:
-            i -= 10     # Provide some context
-            explanation += ['Skipping %s identical '
-                            'trailing characters in diff' % i]
-            left = left[:-i]
-            right = right[:-i]
-    explanation += [line.strip('\n')
-                    for line in py.std.difflib.ndiff(left.splitlines(),
-                                                     right.splitlines())]
-    return explanation
-
-
-def _compare_eq_sequence(left, right):
-    explanation = []
-    for i in range(min(len(left), len(right))):
-        if left[i] != right[i]:
-            explanation += ['At index %s diff: %r != %r' %
-                            (i, left[i], right[i])]
-            break
-    if len(left) > len(right):
-        explanation += ['Left contains more items, '
-            'first extra item: %s' % py.io.saferepr(left[len(right)],)]
-    elif len(left) < len(right):
-        explanation += ['Right contains more items, '
-            'first extra item: %s' % py.io.saferepr(right[len(left)],)]
-    return explanation # + _diff_text(py.std.pprint.pformat(left),
-                       #             py.std.pprint.pformat(right))
-
-
-def _compare_eq_set(left, right):
-    explanation = []
-    diff_left = left - right
-    diff_right = right - left
-    if diff_left:
-        explanation.append('Extra items in the left set:')
-        for item in diff_left:
-            explanation.append(py.io.saferepr(item))
-    if diff_right:
-        explanation.append('Extra items in the right set:')
-        for item in diff_right:
-            explanation.append(py.io.saferepr(item))
-    return explanation
-
-
-def _notin_text(term, text):
-    index = text.find(term)
-    head = text[:index]
-    tail = text[index+len(term):]
-    correct_text = head + tail
-    diff = _diff_text(correct_text, text)
-    newdiff = ['%s is contained here:' % py.io.saferepr(term, maxsize=42)]
-    for line in diff:
-        if line.startswith('Skipping'):
-            continue
-        if line.startswith('- '):
-            continue
-        if line.startswith('+ '):
-            newdiff.append('  ' + line[2:])
-        else:
-            newdiff.append(line)
-    return newdiff
diff --git a/_pytest/assertion/__init__.py b/_pytest/assertion/__init__.py
new file mode 100644
--- /dev/null
+++ b/_pytest/assertion/__init__.py
@@ -0,0 +1,128 @@
+"""
+support for presenting detailed information in failing assertions.
+"""
+import py
+import imp
+import marshal
+import struct
+import sys
+import pytest
+from _pytest.monkeypatch import monkeypatch
+from _pytest.assertion import reinterpret, util
+
+try:
+    from _pytest.assertion.rewrite import rewrite_asserts
+except ImportError:
+    rewrite_asserts = None
+else:
+    import ast
+
+def pytest_addoption(parser):
+    group = parser.getgroup("debugconfig")
+    group.addoption('--assertmode', action="store", dest="assertmode",
+                    choices=("on", "old", "off", "default"), default="default",
+                    metavar="on|old|off",
+                    help="""control assertion debugging tools.
+'off' performs no assertion debugging.
+'old' reinterprets the expressions in asserts to glean information.
+'on' (the default) rewrites the assert statements in test modules to provide
+sub-expression results.""")
+    group.addoption('--no-assert', action="store_true", default=False,
+        dest="noassert", help="DEPRECATED equivalent to --assertmode=off")
+    group.addoption('--nomagic', action="store_true", default=False,
+        dest="nomagic", help="DEPRECATED equivalent to --assertmode=off")
+
+class AssertionState:
+    """State for the assertion plugin."""
+
+    def __init__(self, config, mode):
+        self.mode = mode
+        self.trace = config.trace.root.get("assertion")
+
+def pytest_configure(config):
+    warn_about_missing_assertion()
+    mode = config.getvalue("assertmode")
+    if config.getvalue("noassert") or config.getvalue("nomagic"):
+        if mode not in ("off", "default"):
+            raise pytest.UsageError("assertion options conflict")
+        mode = "off"
+    elif mode == "default":
+        mode = "on"
+    if mode != "off":
+        def callbinrepr(op, left, right):
+            hook_result = config.hook.pytest_assertrepr_compare(
+                config=config, op=op, left=left, right=right)
+            for new_expl in hook_result:
+                if new_expl:
+                    return '\n~'.join(new_expl)
+        m = monkeypatch()
+        config._cleanup.append(m.undo)
+        m.setattr(py.builtin.builtins, 'AssertionError',
+                  reinterpret.AssertionError)
+        m.setattr(util, '_reprcompare', callbinrepr)
+    if mode == "on" and rewrite_asserts is None:
+        mode = "old"
+    config._assertstate = AssertionState(config, mode)
+    config._assertstate.trace("configured with mode set to %r" % (mode,))
+
+def _write_pyc(co, source_path):
+    if hasattr(imp, "cache_from_source"):
+        # Handle PEP 3147 pycs.
+        pyc = py.path.local(imp.cache_from_source(str(source_path)))
+        pyc.ensure()
+    else:
+        pyc = source_path + "c"
+    mtime = int(source_path.mtime())
+    fp = pyc.open("wb")
+    try:
+        fp.write(imp.get_magic())
+        fp.write(struct.pack("<l", mtime))
+        marshal.dump(co, fp)
+    finally:
+        fp.close()
+    return pyc
+
+def before_module_import(mod):
+    if mod.config._assertstate.mode != "on":
+        return
+    # Some deep magic: load the source, rewrite the asserts, and write a
+    # fake pyc, so that it'll be loaded when the module is imported.
+    source = mod.fspath.read()
+    try:
+        tree = ast.parse(source)
+    except SyntaxError:
+        # Let this pop up again in the real import.
+        mod.config._assertstate.trace("failed to parse: %r" % (mod.fspath,))
+        return
+    rewrite_asserts(tree)
+    try:
+        co = compile(tree, str(mod.fspath), "exec")
+    except SyntaxError:
+        # It's possible that this error is from some bug in the assertion
+        # rewriting, but I don't know of a fast way to tell.
+        mod.config._assertstate.trace("failed to compile: %r" % (mod.fspath,))
+        return
+    mod._pyc = _write_pyc(co, mod.fspath)
+    mod.config._assertstate.trace("wrote pyc: %r" % (mod._pyc,))
+
+def after_module_import(mod):
+    if not hasattr(mod, "_pyc"):
+        return
+    state = mod.config._assertstate
+    try:
+        mod._pyc.remove()
+    except py.error.ENOENT:
+        state.trace("couldn't find pyc: %r" % (mod._pyc,))
+    else:
+        state.trace("removed pyc: %r" % (mod._pyc,))
+
+def warn_about_missing_assertion():
+    try:
+        assert False
+    except AssertionError:
+        pass
+    else:
+        sys.stderr.write("WARNING: failing tests may report as passing because "
+        "assertions are turned off!  (are you using python -O?)\n")
+
+pytest_assertrepr_compare = util.assertrepr_compare
diff --git a/_pytest/assertion/newinterpret.py b/_pytest/assertion/newinterpret.py
new file mode 100644
--- /dev/null
+++ b/_pytest/assertion/newinterpret.py
@@ -0,0 +1,333 @@
+"""
+Find intermediate evalutation results in assert statements through builtin AST.
+This should replace oldinterpret.py eventually.
+"""
+
+import sys
+import ast
+
+import py
+from _pytest.assertion import util
+from _pytest.assertion.reinterpret import BuiltinAssertionError
+
+
+if sys.platform.startswith("java") and sys.version_info < (2, 5, 2):
+    # See http://bugs.jython.org/issue1497
+    _exprs = ("BoolOp", "BinOp", "UnaryOp", "Lambda", "IfExp", "Dict",
+              "ListComp", "GeneratorExp", "Yield", "Compare", "Call",
+              "Repr", "Num", "Str", "Attribute", "Subscript", "Name",
+              "List", "Tuple")
+    _stmts = ("FunctionDef", "ClassDef", "Return", "Delete", "Assign",
+              "AugAssign", "Print", "For", "While", "If", "With", "Raise",
+              "TryExcept", "TryFinally", "Assert", "Import", "ImportFrom",
+              "Exec", "Global", "Expr", "Pass", "Break", "Continue")
+    _expr_nodes = set(getattr(ast, name) for name in _exprs)
+    _stmt_nodes = set(getattr(ast, name) for name in _stmts)
+    def _is_ast_expr(node):
+        return node.__class__ in _expr_nodes
+    def _is_ast_stmt(node):
+        return node.__class__ in _stmt_nodes
+else:
+    def _is_ast_expr(node):
+        return isinstance(node, ast.expr)
+    def _is_ast_stmt(node):
+        return isinstance(node, ast.stmt)
+
+
+class Failure(Exception):
+    """Error found while interpreting AST."""
+
+    def __init__(self, explanation=""):
+        self.cause = sys.exc_info()
+        self.explanation = explanation
+
+
+def interpret(source, frame, should_fail=False):
+    mod = ast.parse(source)
+    visitor = DebugInterpreter(frame)
+    try:
+        visitor.visit(mod)
+    except Failure:
+        failure = sys.exc_info()[1]
+        return getfailure(failure)
+    if should_fail:
+        return ("(assertion failed, but when it was re-run for "
+                "printing intermediate values, it did not fail.  Suggestions: "
+                "compute assert expression before the assert or use --no-assert)")
+
+def run(offending_line, frame=None):
+    if frame is None:
+        frame = py.code.Frame(sys._getframe(1))
+    return interpret(offending_line, frame)
+
+def getfailure(e):
+    explanation = util.format_explanation(e.explanation)
+    value = e.cause[1]
+    if str(value):
+        lines = explanation.split('\n')
+        lines[0] += "  << %s" % (value,)
+        explanation = '\n'.join(lines)
+    text = "%s: %s" % (e.cause[0].__name__, explanation)
+    if text.startswith('AssertionError: assert '):
+        text = text[16:]
+    return text
+
+operator_map = {
+    ast.BitOr : "|",
+    ast.BitXor : "^",
+    ast.BitAnd : "&",
+    ast.LShift : "<<",
+    ast.RShift : ">>",
+    ast.Add : "+",
+    ast.Sub : "-",
+    ast.Mult : "*",
+    ast.Div : "/",
+    ast.FloorDiv : "//",
+    ast.Mod : "%",
+    ast.Eq : "==",
+    ast.NotEq : "!=",
+    ast.Lt : "<",
+    ast.LtE : "<=",
+    ast.Gt : ">",
+    ast.GtE : ">=",
+    ast.Pow : "**",
+    ast.Is : "is",
+    ast.IsNot : "is not",
+    ast.In : "in",
+    ast.NotIn : "not in"
+}
+
+unary_map = {
+    ast.Not : "not %s",
+    ast.Invert : "~%s",
+    ast.USub : "-%s",
+    ast.UAdd : "+%s"
+}
+
+
+class DebugInterpreter(ast.NodeVisitor):
+    """Interpret AST nodes to gleam useful debugging information. """
+
+    def __init__(self, frame):
+        self.frame = frame
+
+    def generic_visit(self, node):
+        # Fallback when we don't have a special implementation.
+        if _is_ast_expr(node):
+            mod = ast.Expression(node)
+            co = self._compile(mod)
+            try:
+                result = self.frame.eval(co)
+            except Exception:
+                raise Failure()
+            explanation = self.frame.repr(result)
+            return explanation, result
+        elif _is_ast_stmt(node):
+            mod = ast.Module([node])
+            co = self._compile(mod, "exec")
+            try:
+                self.frame.exec_(co)
+            except Exception:
+                raise Failure()
+            return None, None
+        else:
+            raise AssertionError("can't handle %s" %(node,))
+
+    def _compile(self, source, mode="eval"):
+        return compile(source, "<assertion interpretation>", mode)
+
+    def visit_Expr(self, expr):
+        return self.visit(expr.value)
+
+    def visit_Module(self, mod):
+        for stmt in mod.body:
+            self.visit(stmt)
+
+    def visit_Name(self, name):
+        explanation, result = self.generic_visit(name)
+        # See if the name is local.
+        source = "%r in locals() is not globals()" % (name.id,)
+        co = self._compile(source)
+        try:
+            local = self.frame.eval(co)
+        except Exception:
+            # have to assume it isn't
+            local = None
+        if local is None or not self.frame.is_true(local):
+            return name.id, result
+        return explanation, result
+
+    def visit_Compare(self, comp):
+        left = comp.left
+        left_explanation, left_result = self.visit(left)
+        for op, next_op in zip(comp.ops, comp.comparators):
+            next_explanation, next_result = self.visit(next_op)
+            op_symbol = operator_map[op.__class__]
+            explanation = "%s %s %s" % (left_explanation, op_symbol,
+                                        next_explanation)
+            source = "__exprinfo_left %s __exprinfo_right" % (op_symbol,)
+            co = self._compile(source)
+            try:
+                result = self.frame.eval(co, __exprinfo_left=left_result,
+                                         __exprinfo_right=next_result)
+            except Exception:
+                raise Failure(explanation)
+            try:
+                if not self.frame.is_true(result):
+                    break
+            except KeyboardInterrupt:
+                raise
+            except:
+                break
+            left_explanation, left_result = next_explanation, next_result
+
+        if util._reprcompare is not None:
+            res = util._reprcompare(op_symbol, left_result, next_result)
+            if res:
+                explanation = res
+        return explanation, result
+
+    def visit_BoolOp(self, boolop):
+        is_or = isinstance(boolop.op, ast.Or)
+        explanations = []
+        for operand in boolop.values:
+            explanation, result = self.visit(operand)
+            explanations.append(explanation)
+            if result == is_or:
+                break
+        name = is_or and " or " or " and "
+        explanation = "(" + name.join(explanations) + ")"
+        return explanation, result
+
+    def visit_UnaryOp(self, unary):
+        pattern = unary_map[unary.op.__class__]
+        operand_explanation, operand_result = self.visit(unary.operand)
+        explanation = pattern % (operand_explanation,)
+        co = self._compile(pattern % ("__exprinfo_expr",))
+        try:
+            result = self.frame.eval(co, __exprinfo_expr=operand_result)
+        except Exception:
+            raise Failure(explanation)
+        return explanation, result
+
+    def visit_BinOp(self, binop):
+        left_explanation, left_result = self.visit(binop.left)
+        right_explanation, right_result = self.visit(binop.right)
+        symbol = operator_map[binop.op.__class__]
+        explanation = "(%s %s %s)" % (left_explanation, symbol,
+                                      right_explanation)
+        source = "__exprinfo_left %s __exprinfo_right" % (symbol,)
+        co = self._compile(source)
+        try:
+            result = self.frame.eval(co, __exprinfo_left=left_result,
+                                     __exprinfo_right=right_result)
+        except Exception:
+            raise Failure(explanation)
+        return explanation, result
+
+    def visit_Call(self, call):
+        func_explanation, func = self.visit(call.func)
+        arg_explanations = []
+        ns = {"__exprinfo_func" : func}
+        arguments = []
+        for arg in call.args:
+            arg_explanation, arg_result = self.visit(arg)
+            arg_name = "__exprinfo_%s" % (len(ns),)
+            ns[arg_name] = arg_result
+            arguments.append(arg_name)
+            arg_explanations.append(arg_explanation)
+        for keyword in call.keywords:
+            arg_explanation, arg_result = self.visit(keyword.value)
+            arg_name = "__exprinfo_%s" % (len(ns),)
+            ns[arg_name] = arg_result
+            keyword_source = "%s=%%s" % (keyword.arg)
+            arguments.append(keyword_source % (arg_name,))
+            arg_explanations.append(keyword_source % (arg_explanation,))
+        if call.starargs:
+            arg_explanation, arg_result = self.visit(call.starargs)
+            arg_name = "__exprinfo_star"
+            ns[arg_name] = arg_result
+            arguments.append("*%s" % (arg_name,))
+            arg_explanations.append("*%s" % (arg_explanation,))
+        if call.kwargs:
+            arg_explanation, arg_result = self.visit(call.kwargs)
+            arg_name = "__exprinfo_kwds"
+            ns[arg_name] = arg_result
+            arguments.append("**%s" % (arg_name,))
+            arg_explanations.append("**%s" % (arg_explanation,))
+        args_explained = ", ".join(arg_explanations)
+        explanation = "%s(%s)" % (func_explanation, args_explained)
+        args = ", ".join(arguments)
+        source = "__exprinfo_func(%s)" % (args,)
+        co = self._compile(source)
+        try:
+            result = self.frame.eval(co, **ns)
+        except Exception:
+            raise Failure(explanation)
+        pattern = "%s\n{%s = %s\n}"
+        rep = self.frame.repr(result)
+        explanation = pattern % (rep, rep, explanation)
+        return explanation, result
+
+    def _is_builtin_name(self, name):
+        pattern = "%r not in globals() and %r not in locals()"
+        source = pattern % (name.id, name.id)
+        co = self._compile(source)
+        try:
+            return self.frame.eval(co)
+        except Exception:
+            return False
+
+    def visit_Attribute(self, attr):
+        if not isinstance(attr.ctx, ast.Load):
+            return self.generic_visit(attr)
+        source_explanation, source_result = self.visit(attr.value)
+        explanation = "%s.%s" % (source_explanation, attr.attr)
+        source = "__exprinfo_expr.%s" % (attr.attr,)
+        co = self._compile(source)
+        try:
+            result = self.frame.eval(co, __exprinfo_expr=source_result)
+        except Exception:
+            raise Failure(explanation)
+        explanation = "%s\n{%s = %s.%s\n}" % (self.frame.repr(result),
+                                              self.frame.repr(result),
+                                              source_explanation, attr.attr)
+        # Check if the attr is from an instance.
+        source = "%r in getattr(__exprinfo_expr, '__dict__', {})"
+        source = source % (attr.attr,)
+        co = self._compile(source)
+        try:
+            from_instance = self.frame.eval(co, __exprinfo_expr=source_result)
+        except Exception:
+            from_instance = None
+        if from_instance is None or self.frame.is_true(from_instance):
+            rep = self.frame.repr(result)
+            pattern = "%s\n{%s = %s\n}"
+            explanation = pattern % (rep, rep, explanation)
+        return explanation, result
+
+    def visit_Assert(self, assrt):
+        test_explanation, test_result = self.visit(assrt.test)
+        explanation = "assert %s" % (test_explanation,)
+        if not self.frame.is_true(test_result):
+            try:
+                raise BuiltinAssertionError
+            except Exception:
+                raise Failure(explanation)
+        return explanation, test_result
+
+    def visit_Assign(self, assign):
+        value_explanation, value_result = self.visit(assign.value)
+        explanation = "... = %s" % (value_explanation,)
+        name = ast.Name("__exprinfo_expr", ast.Load(),
+                        lineno=assign.value.lineno,
+                        col_offset=assign.value.col_offset)
+        new_assign = ast.Assign(assign.targets, name, lineno=assign.lineno,
+                                col_offset=assign.col_offset)
+        mod = ast.Module([new_assign])
+        co = self._compile(mod, "exec")
+        try:
+            self.frame.exec_(co, __exprinfo_expr=value_result)
+        except Exception:
+            raise Failure(explanation)
+        return explanation, value_result
diff --git a/_pytest/assertion/oldinterpret.py b/_pytest/assertion/oldinterpret.py
new file mode 100644
--- /dev/null
+++ b/_pytest/assertion/oldinterpret.py
@@ -0,0 +1,552 @@
+import py
+import sys, inspect
+from compiler import parse, ast, pycodegen
+from _pytest.assertion.util import format_explanation
+from _pytest.assertion.reinterpret import BuiltinAssertionError
+
+passthroughex = py.builtin._sysex
+
+class Failure:
+    def __init__(self, node):
+        self.exc, self.value, self.tb = sys.exc_info()
+        self.node = node
+
+class View(object):
+    """View base class.
+
+    If C is a subclass of View, then C(x) creates a proxy object around
+    the object x.  The actual class of the proxy is not C in general,
+    but a *subclass* of C determined by the rules below.  To avoid confusion
+    we call view class the class of the proxy (a subclass of C, so of View)
+    and object class the class of x.
+
+    Attributes and methods not found in the proxy are automatically read on x.
+    Other operations like setting attributes are performed on the proxy, as
+    determined by its view class.  The object x is available from the proxy
+    as its __obj__ attribute.
+
+    The view class selection is determined by the __view__ tuples and the
+    optional __viewkey__ method.  By default, the selected view class is the
+    most specific subclass of C whose __view__ mentions the class of x.
+    If no such subclass is found, the search proceeds with the parent
+    object classes.  For example, C(True) will first look for a subclass
+    of C with __view__ = (..., bool, ...) and only if it doesn't find any
+    look for one with __view__ = (..., int, ...), and then ..., object,...
+    If everything fails the class C itself is considered to be the default.
+
+    Alternatively, the view class selection can be driven by another aspect
+    of the object x, instead of the class of x, by overriding __viewkey__.
+    See last example at the end of this module.
+    """
+
+    _viewcache = {}
+    __view__ = ()
+
+    def __new__(rootclass, obj, *args, **kwds):
+        self = object.__new__(rootclass)
+        self.__obj__ = obj
+        self.__rootclass__ = rootclass
+        key = self.__viewkey__()
+        try:
+            self.__class__ = self._viewcache[key]
+        except KeyError:
+            self.__class__ = self._selectsubclass(key)
+        return self
+
+    def __getattr__(self, attr):
+        # attributes not found in the normal hierarchy rooted on View
+        # are looked up in the object's real class
+        return getattr(self.__obj__, attr)
+
+    def __viewkey__(self):
+        return self.__obj__.__class__
+
+    def __matchkey__(self, key, subclasses):
+        if inspect.isclass(key):
+            keys = inspect.getmro(key)
+        else:
+            keys = [key]
+        for key in keys:
+            result = [C for C in subclasses if key in C.__view__]
+            if result:
+                return result
+        return []
+
+    def _selectsubclass(self, key):
+        subclasses = list(enumsubclasses(self.__rootclass__))
+        for C in subclasses:
+            if not isinstance(C.__view__, tuple):
+                C.__view__ = (C.__view__,)
+        choices = self.__matchkey__(key, subclasses)
+        if not choices:
+            return self.__rootclass__
+        elif len(choices) == 1:
+            return choices[0]
+        else:
+            # combine the multiple choices
+            return type('?', tuple(choices), {})
+
+    def __repr__(self):
+        return '%s(%r)' % (self.__rootclass__.__name__, self.__obj__)
+
+
+def enumsubclasses(cls):
+    for subcls in cls.__subclasses__():
+        for subsubclass in enumsubclasses(subcls):
+            yield subsubclass
+    yield cls
+
+
+class Interpretable(View):
+    """A parse tree node with a few extra methods."""
+    explanation = None
+
+    def is_builtin(self, frame):
+        return False
+
+    def eval(self, frame):
+        # fall-back for unknown expression nodes
+        try:
+            expr = ast.Expression(self.__obj__)
+            expr.filename = '<eval>'
+            self.__obj__.filename = '<eval>'
+            co = pycodegen.ExpressionCodeGenerator(expr).getCode()
+            result = frame.eval(co)
+        except passthroughex:
+            raise
+        except:
+            raise Failure(self)
+        self.result = result
+        self.explanation = self.explanation or frame.repr(self.result)
+
+    def run(self, frame):
+        # fall-back for unknown statement nodes
+        try:
+            expr = ast.Module(None, ast.Stmt([self.__obj__]))
+            expr.filename = '<run>'
+            co = pycodegen.ModuleCodeGenerator(expr).getCode()
+            frame.exec_(co)
+        except passthroughex:
+            raise
+        except:
+            raise Failure(self)
+
+    def nice_explanation(self):
+        return format_explanation(self.explanation)
+
+
+class Name(Interpretable):
+    __view__ = ast.Name
+
+    def is_local(self, frame):
+        source = '%r in locals() is not globals()' % self.name
+        try:
+            return frame.is_true(frame.eval(source))
+        except passthroughex:
+            raise
+        except:
+            return False
+
+    def is_global(self, frame):
+        source = '%r in globals()' % self.name
+        try:
+            return frame.is_true(frame.eval(source))
+        except passthroughex:
+            raise
+        except:
+            return False
+
+    def is_builtin(self, frame):
+        source = '%r not in locals() and %r not in globals()' % (
+            self.name, self.name)
+        try:
+            return frame.is_true(frame.eval(source))
+        except passthroughex:
+            raise
+        except:
+            return False
+
+    def eval(self, frame):
+        super(Name, self).eval(frame)
+        if not self.is_local(frame):
+            self.explanation = self.name
+
+class Compare(Interpretable):
+    __view__ = ast.Compare
+
+    def eval(self, frame):
+        expr = Interpretable(self.expr)
+        expr.eval(frame)
+        for operation, expr2 in self.ops:
+            if hasattr(self, 'result'):
+                # shortcutting in chained expressions
+                if not frame.is_true(self.result):
+                    break
+            expr2 = Interpretable(expr2)
+            expr2.eval(frame)
+            self.explanation = "%s %s %s" % (
+                expr.explanation, operation, expr2.explanation)
+            source = "__exprinfo_left %s __exprinfo_right" % operation
+            try:
+                self.result = frame.eval(source,
+                                         __exprinfo_left=expr.result,
+                                         __exprinfo_right=expr2.result)
+            except passthroughex:
+                raise
+            except:
+                raise Failure(self)
+            expr = expr2
+
+class And(Interpretable):
+    __view__ = ast.And
+
+    def eval(self, frame):
+        explanations = []
+        for expr in self.nodes:
+            expr = Interpretable(expr)
+            expr.eval(frame)
+            explanations.append(expr.explanation)
+            self.result = expr.result
+            if not frame.is_true(expr.result):
+                break
+        self.explanation = '(' + ' and '.join(explanations) + ')'
+
+class Or(Interpretable):
+    __view__ = ast.Or
+
+    def eval(self, frame):
+        explanations = []
+        for expr in self.nodes:
+            expr = Interpretable(expr)
+            expr.eval(frame)
+            explanations.append(expr.explanation)
+            self.result = expr.result
+            if frame.is_true(expr.result):
+                break
+        self.explanation = '(' + ' or '.join(explanations) + ')'
+
+
+# == Unary operations ==
+keepalive = []
+for astclass, astpattern in {
+    ast.Not    : 'not __exprinfo_expr',
+    ast.Invert : '(~__exprinfo_expr)',
+    }.items():
+
+    class UnaryArith(Interpretable):
+        __view__ = astclass
+
+        def eval(self, frame, astpattern=astpattern):
+            expr = Interpretable(self.expr)
+            expr.eval(frame)
+            self.explanation = astpattern.replace('__exprinfo_expr',
+                                                  expr.explanation)
+            try:
+                self.result = frame.eval(astpattern,
+                                         __exprinfo_expr=expr.result)
+            except passthroughex:
+                raise
+            except:
+                raise Failure(self)
+
+    keepalive.append(UnaryArith)
+
+# == Binary operations ==
+for astclass, astpattern in {
+    ast.Add    : '(__exprinfo_left + __exprinfo_right)',
+    ast.Sub    : '(__exprinfo_left - __exprinfo_right)',
+    ast.Mul    : '(__exprinfo_left * __exprinfo_right)',
+    ast.Div    : '(__exprinfo_left / __exprinfo_right)',
+    ast.Mod    : '(__exprinfo_left % __exprinfo_right)',
+    ast.Power  : '(__exprinfo_left ** __exprinfo_right)',
+    }.items():
+
+    class BinaryArith(Interpretable):
+        __view__ = astclass
+
+        def eval(self, frame, astpattern=astpattern):
+            left = Interpretable(self.left)
+            left.eval(frame)
+            right = Interpretable(self.right)
+            right.eval(frame)
+            self.explanation = (astpattern
+                                .replace('__exprinfo_left',  left .explanation)
+                                .replace('__exprinfo_right', right.explanation))
+            try:
+                self.result = frame.eval(astpattern,
+                                         __exprinfo_left=left.result,
+                                         __exprinfo_right=right.result)
+            except passthroughex:
+                raise
+            except:
+                raise Failure(self)
+
+    keepalive.append(BinaryArith)
+
+
+class CallFunc(Interpretable):
+    __view__ = ast.CallFunc
+
+    def is_bool(self, frame):
+        source = 'isinstance(__exprinfo_value, bool)'
+        try:
+            return frame.is_true(frame.eval(source,
+                                            __exprinfo_value=self.result))
+        except passthroughex:
+            raise
+        except:
+            return False
+
+    def eval(self, frame):
+        node = Interpretable(self.node)
+        node.eval(frame)
+        explanations = []
+        vars = {'__exprinfo_fn': node.result}
+        source = '__exprinfo_fn('
+        for a in self.args:
+            if isinstance(a, ast.Keyword):
+                keyword = a.name
+                a = a.expr
+            else:
+                keyword = None
+            a = Interpretable(a)
+            a.eval(frame)
+            argname = '__exprinfo_%d' % len(vars)
+            vars[argname] = a.result
+            if keyword is None:
+                source += argname + ','
+                explanations.append(a.explanation)
+            else:
+                source += '%s=%s,' % (keyword, argname)
+                explanations.append('%s=%s' % (keyword, a.explanation))
+        if self.star_args:
+            star_args = Interpretable(self.star_args)
+            star_args.eval(frame)
+            argname = '__exprinfo_star'
+            vars[argname] = star_args.result
+            source += '*' + argname + ','
+            explanations.append('*' + star_args.explanation)
+        if self.dstar_args:
+            dstar_args = Interpretable(self.dstar_args)
+            dstar_args.eval(frame)
+            argname = '__exprinfo_kwds'
+            vars[argname] = dstar_args.result
+            source += '**' + argname + ','
+            explanations.append('**' + dstar_args.explanation)
+        self.explanation = "%s(%s)" % (
+            node.explanation, ', '.join(explanations))
+        if source.endswith(','):
+            source = source[:-1]
+        source += ')'
+        try:
+            self.result = frame.eval(source, **vars)
+        except passthroughex:
+            raise
+        except:
+            raise Failure(self)
+        if not node.is_builtin(frame) or not self.is_bool(frame):
+            r = frame.repr(self.result)
+            self.explanation = '%s\n{%s = %s\n}' % (r, r, self.explanation)
+
+class Getattr(Interpretable):
+    __view__ = ast.Getattr
+
+    def eval(self, frame):
+        expr = Interpretable(self.expr)
+        expr.eval(frame)
+        source = '__exprinfo_expr.%s' % self.attrname
+        try:
+            self.result = frame.eval(source, __exprinfo_expr=expr.result)
+        except passthroughex:
+            raise
+        except:
+            raise Failure(self)
+        self.explanation = '%s.%s' % (expr.explanation, self.attrname)
+        # if the attribute comes from the instance, its value is interesting
+        source = ('hasattr(__exprinfo_expr, "__dict__") and '
+                  '%r in __exprinfo_expr.__dict__' % self.attrname)
+        try:
+            from_instance = frame.is_true(
+                frame.eval(source, __exprinfo_expr=expr.result))
+        except passthroughex:
+            raise
+        except:
+            from_instance = True
+        if from_instance:
+            r = frame.repr(self.result)
+            self.explanation = '%s\n{%s = %s\n}' % (r, r, self.explanation)
+
+# == Re-interpretation of full statements ==
+
+class Assert(Interpretable):
+    __view__ = ast.Assert
+
+    def run(self, frame):
+        test = Interpretable(self.test)
+        test.eval(frame)
+        # print the result as  'assert <explanation>'
+        self.result = test.result
+        self.explanation = 'assert ' + test.explanation
+        if not frame.is_true(test.result):
+            try:
+                raise BuiltinAssertionError
+            except passthroughex:
+                raise
+            except:
+                raise Failure(self)
+
+class Assign(Interpretable):
+    __view__ = ast.Assign
+
+    def run(self, frame):
+        expr = Interpretable(self.expr)
+        expr.eval(frame)
+        self.result = expr.result
+        self.explanation = '... = ' + expr.explanation
+        # fall-back-run the rest of the assignment
+        ass = ast.Assign(self.nodes, ast.Name('__exprinfo_expr'))
+        mod = ast.Module(None, ast.Stmt([ass]))
+        mod.filename = '<run>'
+        co = pycodegen.ModuleCodeGenerator(mod).getCode()
+        try:
+            frame.exec_(co, __exprinfo_expr=expr.result)
+        except passthroughex:
+            raise
+        except:
+            raise Failure(self)
+
+class Discard(Interpretable):
+    __view__ = ast.Discard
+
+    def run(self, frame):
+        expr = Interpretable(self.expr)
+        expr.eval(frame)
+        self.result = expr.result
+        self.explanation = expr.explanation
+
+class Stmt(Interpretable):
+    __view__ = ast.Stmt
+
+    def run(self, frame):
+        for stmt in self.nodes:
+            stmt = Interpretable(stmt)
+            stmt.run(frame)
+
+
+def report_failure(e):
+    explanation = e.node.nice_explanation()
+    if explanation:
+        explanation = ", in: " + explanation
+    else:
+        explanation = ""
+    sys.stdout.write("%s: %s%s\n" % (e.exc.__name__, e.value, explanation))
+
+def check(s, frame=None):
+    if frame is None:
+        frame = sys._getframe(1)
+        frame = py.code.Frame(frame)
+    expr = parse(s, 'eval')
+    assert isinstance(expr, ast.Expression)
+    node = Interpretable(expr.node)
+    try:
+        node.eval(frame)
+    except passthroughex:
+        raise
+    except Failure:
+        e = sys.exc_info()[1]
+        report_failure(e)
+    else:
+        if not frame.is_true(node.result):
+            sys.stderr.write("assertion failed: %s\n" % node.nice_explanation())
+
+
+###########################################################
+# API / Entry points
+# #########################################################
+
+def interpret(source, frame, should_fail=False):
+    module = Interpretable(parse(source, 'exec').node)
+    #print "got module", module
+    if isinstance(frame, py.std.types.FrameType):
+        frame = py.code.Frame(frame)
+    try:
+        module.run(frame)
+    except Failure:
+        e = sys.exc_info()[1]
+        return getfailure(e)
+    except passthroughex:
+        raise
+    except:
+        import traceback
+        traceback.print_exc()
+    if should_fail:
+        return ("(assertion failed, but when it was re-run for "
+                "printing intermediate values, it did not fail.  Suggestions: "
+                "compute assert expression before the assert or use --nomagic)")
+    else:
+        return None
+
+def getmsg(excinfo):
+    if isinstance(excinfo, tuple):
+        excinfo = py.code.ExceptionInfo(excinfo)
+    #frame, line = gettbline(tb)
+    #frame = py.code.Frame(frame)
+    #return interpret(line, frame)
+
+    tb = excinfo.traceback[-1]
+    source = str(tb.statement).strip()
+    x = interpret(source, tb.frame, should_fail=True)
+    if not isinstance(x, str):
+        raise TypeError("interpret returned non-string %r" % (x,))
+    return x
+
+def getfailure(e):
+    explanation = e.node.nice_explanation()
+    if str(e.value):
+        lines = explanation.split('\n')
+        lines[0] += "  << %s" % (e.value,)
+        explanation = '\n'.join(lines)
+    text = "%s: %s" % (e.exc.__name__, explanation)
+    if text.startswith('AssertionError: assert '):
+        text = text[16:]
+    return text
+
+def run(s, frame=None):
+    if frame is None:
+        frame = sys._getframe(1)
+        frame = py.code.Frame(frame)
+    module = Interpretable(parse(s, 'exec').node)
+    try:
+        module.run(frame)
+    except Failure:
+        e = sys.exc_info()[1]
+        report_failure(e)
+
+
+if __name__ == '__main__':
+    # example:
+    def f():
+        return 5
+    def g():
+        return 3
+    def h(x):
+        return 'never'
+    check("f() * g() == 5")
+    check("not f()")
+    check("not (f() and g() or 0)")
+    check("f() == g()")
+    i = 4
+    check("i == f()")
+    check("len(f()) == 0")
+    check("isinstance(2+3+4, float)")
+
+    run("x = i")
+    check("x == 5")
+
+    run("assert not f(), 'oops'")
+    run("a, b, c = 1, 2")
+    run("a, b, c = f()")
+
+    check("max([f(),g()]) == 4")
+    check("'hello'[g()] == 'h'")
+    run("'guk%d' % h(f())")
diff --git a/_pytest/assertion/reinterpret.py b/_pytest/assertion/reinterpret.py
new file mode 100644
--- /dev/null
+++ b/_pytest/assertion/reinterpret.py
@@ -0,0 +1,48 @@
+import sys
+import py
+
+BuiltinAssertionError = py.builtin.builtins.AssertionError
+
+class AssertionError(BuiltinAssertionError):
+    def __init__(self, *args):
+        BuiltinAssertionError.__init__(self, *args)
+        if args:
+            try:
+                self.msg = str(args[0])
+            except py.builtin._sysex:
+                raise
+            except:
+                self.msg = "<[broken __repr__] %s at %0xd>" %(
+                    args[0].__class__, id(args[0]))
+        else:
+            f = py.code.Frame(sys._getframe(1))
+            try:
+                source = f.code.fullsource
+                if source is not None:
+                    try:
+                        source = source.getstatement(f.lineno, assertion=True)
+                    except IndexError:
+                        source = None
+                    else:
+                        source = str(source.deindent()).strip()
+            except py.error.ENOENT:
+                source = None
+                # this can also occur during reinterpretation, when the
+                # co_filename is set to "<run>".
+            if source:
+                self.msg = reinterpret(source, f, should_fail=True)
+            else:
+                self.msg = "<could not determine information>"
+            if not self.args:
+                self.args = (self.msg,)
+
+if sys.version_info > (3, 0):
+    AssertionError.__module__ = "builtins"
+    reinterpret_old = "old reinterpretation not available for py3"
+else:
+    from _pytest.assertion.oldinterpret import interpret as reinterpret_old
+if sys.version_info >= (2, 6) or (sys.platform.startswith("java")):
+    from _pytest.assertion.newinterpret import interpret as reinterpret
+else:
+    reinterpret = reinterpret_old
+
diff --git a/_pytest/assertion/rewrite.py b/_pytest/assertion/rewrite.py
new file mode 100644
--- /dev/null
+++ b/_pytest/assertion/rewrite.py
@@ -0,0 +1,340 @@
+"""Rewrite assertion AST to produce nice error messages"""
+
+import ast
+import collections
+import itertools
+import sys
+
+import py
+from _pytest.assertion import util
+
+
+def rewrite_asserts(mod):
+    """Rewrite the assert statements in mod."""
+    AssertionRewriter().run(mod)
+
+
+_saferepr = py.io.saferepr
+from _pytest.assertion.util import format_explanation as _format_explanation
+
+def _format_boolop(operands, explanations, is_or):
+    show_explanations = []
+    for operand, expl in zip(operands, explanations):
+        show_explanations.append(expl)
+        if operand == is_or:
+            break
+    return "(" + (is_or and " or " or " and ").join(show_explanations) + ")"
+
+def _call_reprcompare(ops, results, expls, each_obj):
+    for i, res, expl in zip(range(len(ops)), results, expls):
+        try:
+            done = not res
+        except Exception:
+            done = True
+        if done:
+            break
+    if util._reprcompare is not None:
+        custom = util._reprcompare(ops[i], each_obj[i], each_obj[i + 1])
+        if custom is not None:
+            return custom
+    return expl
+
+
+unary_map = {
+    ast.Not : "not %s",
+    ast.Invert : "~%s",
+    ast.USub : "-%s",
+    ast.UAdd : "+%s"
+}
+
+binop_map = {
+    ast.BitOr : "|",
+    ast.BitXor : "^",
+    ast.BitAnd : "&",
+    ast.LShift : "<<",
+    ast.RShift : ">>",
+    ast.Add : "+",
+    ast.Sub : "-",
+    ast.Mult : "*",
+    ast.Div : "/",
+    ast.FloorDiv : "//",
+    ast.Mod : "%",
+    ast.Eq : "==",
+    ast.NotEq : "!=",
+    ast.Lt : "<",
+    ast.LtE : "<=",
+    ast.Gt : ">",
+    ast.GtE : ">=",
+    ast.Pow : "**",
+    ast.Is : "is",
+    ast.IsNot : "is not",
+    ast.In : "in",
+    ast.NotIn : "not in"
+}
+
+
+def set_location(node, lineno, col_offset):
+    """Set node location information recursively."""
+    def _fix(node, lineno, col_offset):
+        if "lineno" in node._attributes:
+            node.lineno = lineno
+        if "col_offset" in node._attributes:
+            node.col_offset = col_offset
+        for child in ast.iter_child_nodes(node):
+            _fix(child, lineno, col_offset)
+    _fix(node, lineno, col_offset)
+    return node
+
+
+class AssertionRewriter(ast.NodeVisitor):
+
+    def run(self, mod):
+        """Find all assert statements in *mod* and rewrite them."""
+        if not mod.body:
+            # Nothing to do.
+            return
+        # Insert some special imports at the top of the module but after any
+        # docstrings and __future__ imports.
+        aliases = [ast.alias(py.builtin.builtins.__name__, "@py_builtins"),
+                   ast.alias("_pytest.assertion.rewrite", "@pytest_ar")]
+        expect_docstring = True
+        pos = 0
+        lineno = 0
+        for item in mod.body:
+            if (expect_docstring and isinstance(item, ast.Expr) and
+                isinstance(item.value, ast.Str)):
+                doc = item.value.s
+                if "PYTEST_DONT_REWRITE" in doc:
+                    # The module has disabled assertion rewriting.
+                    return
+                lineno += len(doc) - 1
+                expect_docstring = False
+            elif (not isinstance(item, ast.ImportFrom) or item.level > 0 and
+                  item.identifier != "__future__"):
+                lineno = item.lineno
+                break
+            pos += 1
+        imports = [ast.Import([alias], lineno=lineno, col_offset=0)
+                   for alias in aliases]
+        mod.body[pos:pos] = imports
+        # Collect asserts.
+        nodes = collections.deque([mod])
+        while nodes:
+            node = nodes.popleft()
+            for name, field in ast.iter_fields(node):
+                if isinstance(field, list):
+                    new = []
+                    for i, child in enumerate(field):
+                        if isinstance(child, ast.Assert):
+                            # Transform assert.
+                            new.extend(self.visit(child))
+                        else:
+                            new.append(child)
+                            if isinstance(child, ast.AST):
+                                nodes.append(child)
+                    setattr(node, name, new)
+                elif (isinstance(field, ast.AST) and
+                      # Don't recurse into expressions as they can't contain
+                      # asserts.
+                      not isinstance(field, ast.expr)):
+                    nodes.append(field)
+
+    def variable(self):
+        """Get a new variable."""
+        # Use a character invalid in python identifiers to avoid clashing.
+        name = "@py_assert" + str(next(self.variable_counter))
+        self.variables.add(name)
+        return name
+
+    def assign(self, expr):
+        """Give *expr* a name."""
+        name = self.variable()
+        self.statements.append(ast.Assign([ast.Name(name, ast.Store())], expr))
+        return ast.Name(name, ast.Load())
+
+    def display(self, expr):
+        """Call py.io.saferepr on the expression."""
+        return self.helper("saferepr", expr)
+
+    def helper(self, name, *args):
+        """Call a helper in this module."""
+        py_name = ast.Name("@pytest_ar", ast.Load())
+        attr = ast.Attribute(py_name, "_" + name, ast.Load())
+        return ast.Call(attr, list(args), [], None, None)
+
+    def builtin(self, name):
+        """Return the builtin called *name*."""
+        builtin_name = ast.Name("@py_builtins", ast.Load())
+        return ast.Attribute(builtin_name, name, ast.Load())
+
+    def explanation_param(self, expr):
+        specifier = "py" + str(next(self.variable_counter))
+        self.explanation_specifiers[specifier] = expr
+        return "%(" + specifier + ")s"
+
+    def push_format_context(self):
+        self.explanation_specifiers = {}
+        self.stack.append(self.explanation_specifiers)
+
+    def pop_format_context(self, expl_expr):
+        current = self.stack.pop()
+        if self.stack:
+            self.explanation_specifiers = self.stack[-1]
+        keys = [ast.Str(key) for key in current.keys()]
+        format_dict = ast.Dict(keys, list(current.values()))
+        form = ast.BinOp(expl_expr, ast.Mod(), format_dict)
+        name = "@py_format" + str(next(self.variable_counter))
+        self.on_failure.append(ast.Assign([ast.Name(name, ast.Store())], form))
+        return ast.Name(name, ast.Load())
+
+    def generic_visit(self, node):
+        """Handle expressions we don't have custom code for."""
+        assert isinstance(node, ast.expr)
+        res = self.assign(node)
+        return res, self.explanation_param(self.display(res))
+
+    def visit_Assert(self, assert_):
+        if assert_.msg:
+            # There's already a message. Don't mess with it.
+            return [assert_]
+        self.statements = []
+        self.variables = set()
+        self.variable_counter = itertools.count()
+        self.stack = []
+        self.on_failure = []
+        self.push_format_context()
+        # Rewrite assert into a bunch of statements.
+        top_condition, explanation = self.visit(assert_.test)
+        # Create failure message.
+        body = self.on_failure
+        negation = ast.UnaryOp(ast.Not(), top_condition)
+        self.statements.append(ast.If(negation, body, []))
+        explanation = "assert " + explanation
+        template = ast.Str(explanation)
+        msg = self.pop_format_context(template)
+        fmt = self.helper("format_explanation", msg)
+        err_name = ast.Name("AssertionError", ast.Load())
+        exc = ast.Call(err_name, [fmt], [], None, None)
+        if sys.version_info[0] >= 3:
+            raise_ = ast.Raise(exc, None)
+        else:
+            raise_ = ast.Raise(exc, None, None)
+        body.append(raise_)
+        # Delete temporary variables.
+        names = [ast.Name(name, ast.Del()) for name in self.variables]
+        if names:
+            delete = ast.Delete(names)
+            self.statements.append(delete)
+        # Fix line numbers.
+        for stmt in self.statements:
+            set_location(stmt, assert_.lineno, assert_.col_offset)
+        return self.statements
+
+    def visit_Name(self, name):
+        # Check if the name is local or not.
+        locs = ast.Call(self.builtin("locals"), [], [], None, None)
+        globs = ast.Call(self.builtin("globals"), [], [], None, None)
+        ops = [ast.In(), ast.IsNot()]
+        test = ast.Compare(ast.Str(name.id), ops, [locs, globs])
+        expr = ast.IfExp(test, self.display(name), ast.Str(name.id))
+        return name, self.explanation_param(expr)
+
+    def visit_BoolOp(self, boolop):
+        operands = []
+        explanations = []
+        self.push_format_context()
+        for operand in boolop.values:
+            res, explanation = self.visit(operand)
+            operands.append(res)
+            explanations.append(explanation)
+        expls = ast.Tuple([ast.Str(expl) for expl in explanations], ast.Load())
+        is_or = ast.Num(isinstance(boolop.op, ast.Or))
+        expl_template = self.helper("format_boolop",
+                                    ast.Tuple(operands, ast.Load()), expls,
+                                    is_or)
+        expl = self.pop_format_context(expl_template)
+        res = self.assign(ast.BoolOp(boolop.op, operands))
+        return res, self.explanation_param(expl)
+
+    def visit_UnaryOp(self, unary):
+        pattern = unary_map[unary.op.__class__]
+        operand_res, operand_expl = self.visit(unary.operand)
+        res = self.assign(ast.UnaryOp(unary.op, operand_res))
+        return res, pattern % (operand_expl,)
+
+    def visit_BinOp(self, binop):
+        symbol = binop_map[binop.op.__class__]
+        left_expr, left_expl = self.visit(binop.left)
+        right_expr, right_expl = self.visit(binop.right)
+        explanation = "(%s %s %s)" % (left_expl, symbol, right_expl)
+        res = self.assign(ast.BinOp(left_expr, binop.op, right_expr))
+        return res, explanation
+
+    def visit_Call(self, call):
+        new_func, func_expl = self.visit(call.func)
+        arg_expls = []
+        new_args = []
+        new_kwargs = []
+        new_star = new_kwarg = None
+        for arg in call.args:
+            res, expl = self.visit(arg)
+            new_args.append(res)
+            arg_expls.append(expl)
+        for keyword in call.keywords:
+            res, expl = self.visit(keyword.value)
+            new_kwargs.append(ast.keyword(keyword.arg, res))
+            arg_expls.append(keyword.arg + "=" + expl)
+        if call.starargs:
+            new_star, expl = self.visit(call.starargs)
+            arg_expls.append("*" + expl)
+        if call.kwargs:
+            new_kwarg, expl = self.visit(call.kwarg)
+            arg_expls.append("**" + expl)
+        expl = "%s(%s)" % (func_expl, ', '.join(arg_expls))
+        new_call = ast.Call(new_func, new_args, new_kwargs, new_star, new_kwarg)
+        res = self.assign(new_call)
+        res_expl = self.explanation_param(self.display(res))
+        outer_expl = "%s\n{%s = %s\n}" % (res_expl, res_expl, expl)
+        return res, outer_expl
+
+    def visit_Attribute(self, attr):
+        if not isinstance(attr.ctx, ast.Load):
+            return self.generic_visit(attr)
+        value, value_expl = self.visit(attr.value)
+        res = self.assign(ast.Attribute(value, attr.attr, ast.Load()))
+        res_expl = self.explanation_param(self.display(res))
+        pat = "%s\n{%s = %s.%s\n}"
+        expl = pat % (res_expl, res_expl, value_expl, attr.attr)
+        return res, expl
+
+    def visit_Compare(self, comp):
+        self.push_format_context()
+        left_res, left_expl = self.visit(comp.left)
+        res_variables = [self.variable() for i in range(len(comp.ops))]
+        load_names = [ast.Name(v, ast.Load()) for v in res_variables]
+        store_names = [ast.Name(v, ast.Store()) for v in res_variables]
+        it = zip(range(len(comp.ops)), comp.ops, comp.comparators)
+        expls = []
+        syms = []
+        results = [left_res]
+        for i, op, next_operand in it:
+            next_res, next_expl = self.visit(next_operand)
+            results.append(next_res)
+            sym = binop_map[op.__class__]
+            syms.append(ast.Str(sym))
+            expl = "%s %s %s" % (left_expl, sym, next_expl)
+            expls.append(ast.Str(expl))
+            res_expr = ast.Compare(left_res, [op], [next_res])
+            self.statements.append(ast.Assign([store_names[i]], res_expr))
+            left_res, left_expl = next_res, next_expl
+        # Use py.code._reprcompare if that's available.
+        expl_call = self.helper("call_reprcompare",
+                                ast.Tuple(syms, ast.Load()),
+                                ast.Tuple(load_names, ast.Load()),
+                                ast.Tuple(expls, ast.Load()),
+                                ast.Tuple(results, ast.Load()))
+        if len(comp.ops) > 1:
+            res = ast.BoolOp(ast.And(), load_names)
+        else:
+            res = load_names[0]
+        return res, self.explanation_param(self.pop_format_context(expl_call))
diff --git a/_pytest/assertion/util.py b/_pytest/assertion/util.py
new file mode 100644
--- /dev/null
+++ b/_pytest/assertion/util.py
@@ -0,0 +1,213 @@
+"""Utilities for assertion debugging"""
+
+import py
+
+
+# The _reprcompare attribute on the util module is used by the new assertion
+# interpretation code and assertion rewriter to detect this plugin was
+# loaded and in turn call the hooks defined here as part of the
+# DebugInterpreter.
+_reprcompare = None
+
+def format_explanation(explanation):
+    """This formats an explanation
+
+    Normally all embedded newlines are escaped, however there are
+    three exceptions: \n{, \n} and \n~.  The first two are intended
+    cover nested explanations, see function and attribute explanations
+    for examples (.visit_Call(), visit_Attribute()).  The last one is
+    for when one explanation needs to span multiple lines, e.g. when
+    displaying diffs.
+    """
+    # simplify 'assert False where False = ...'
+    where = 0
+    while True:
+        start = where = explanation.find("False\n{False = ", where)
+        if where == -1:
+            break
+        level = 0
+        for i, c in enumerate(explanation[start:]):
+            if c == "{":
+                level += 1
+            elif c == "}":
+                level -= 1
+                if not level:
+                    break
+        else:
+            raise AssertionError("unbalanced braces: %r" % (explanation,))
+        end = start + i
+        where = end
+        if explanation[end - 1] == '\n':
+            explanation = (explanation[:start] + explanation[start+15:end-1] +
+                           explanation[end+1:])
+            where -= 17
+    raw_lines = (explanation or '').split('\n')
+    # escape newlines not followed by {, } and ~
+    lines = [raw_lines[0]]
+    for l in raw_lines[1:]:
+        if l.startswith('{') or l.startswith('}') or l.startswith('~'):
+            lines.append(l)
+        else:
+            lines[-1] += '\\n' + l
+
+    result = lines[:1]
+    stack = [0]
+    stackcnt = [0]
+    for line in lines[1:]:
+        if line.startswith('{'):
+            if stackcnt[-1]:
+                s = 'and   '
+            else:
+                s = 'where '
+            stack.append(len(result))
+            stackcnt[-1] += 1
+            stackcnt.append(0)
+            result.append(' +' + '  '*(len(stack)-1) + s + line[1:])
+        elif line.startswith('}'):
+            assert line.startswith('}')
+            stack.pop()
+            stackcnt.pop()
+            result[stack[-1]] += line[1:]
+        else:
+            assert line.startswith('~')
+            result.append('  '*len(stack) + line[1:])
+    assert len(stack) == 1
+    return '\n'.join(result)
+
+
+# Provide basestring in python3
+try:
+    basestring = basestring
+except NameError:
+    basestring = str
+
+
+def assertrepr_compare(op, left, right):
+    """return specialised explanations for some operators/operands"""
+    width = 80 - 15 - len(op) - 2 # 15 chars indentation, 1 space around op
+    left_repr = py.io.saferepr(left, maxsize=int(width/2))
+    right_repr = py.io.saferepr(right, maxsize=width-len(left_repr))
+    summary = '%s %s %s' % (left_repr, op, right_repr)
+
+    issequence = lambda x: isinstance(x, (list, tuple))
+    istext = lambda x: isinstance(x, basestring)
+    isdict = lambda x: isinstance(x, dict)
+    isset = lambda x: isinstance(x, set)
+
+    explanation = None
+    try:
+        if op == '==':
+            if istext(left) and istext(right):
+                explanation = _diff_text(left, right)
+            elif issequence(left) and issequence(right):
+                explanation = _compare_eq_sequence(left, right)
+            elif isset(left) and isset(right):
+                explanation = _compare_eq_set(left, right)
+            elif isdict(left) and isdict(right):
+                explanation = _diff_text(py.std.pprint.pformat(left),
+                                         py.std.pprint.pformat(right))
+        elif op == 'not in':
+            if istext(left) and istext(right):
+                explanation = _notin_text(left, right)
+    except py.builtin._sysex:
+        raise
+    except:
+        excinfo = py.code.ExceptionInfo()
+        explanation = ['(pytest_assertion plugin: representation of '
+            'details failed. Probably an object has a faulty __repr__.)',
+            str(excinfo)
+            ]
+
+
+    if not explanation:
+        return None
+
+    # Don't include pageloads of data, should be configurable
+    if len(''.join(explanation)) > 80*8:
+        explanation = ['Detailed information too verbose, truncated']
+
+    return [summary] + explanation
+
+
+def _diff_text(left, right):
+    """Return the explanation for the diff between text
+
+    This will skip leading and trailing characters which are
+    identical to keep the diff minimal.
+    """
+    explanation = []
+    i = 0 # just in case left or right has zero length
+    for i in range(min(len(left), len(right))):
+        if left[i] != right[i]:
+            break
+    if i > 42:
+        i -= 10                 # Provide some context
+        explanation = ['Skipping %s identical '
+                       'leading characters in diff' % i]
+        left = left[i:]
+        right = right[i:]
+    if len(left) == len(right):
+        for i in range(len(left)):
+            if left[-i] != right[-i]:
+                break
+        if i > 42:
+            i -= 10     # Provide some context
+            explanation += ['Skipping %s identical '
+                            'trailing characters in diff' % i]
+            left = left[:-i]
+            right = right[:-i]
+    explanation += [line.strip('\n')
+                    for line in py.std.difflib.ndiff(left.splitlines(),
+                                                     right.splitlines())]
+    return explanation
+
+
+def _compare_eq_sequence(left, right):
+    explanation = []
+    for i in range(min(len(left), len(right))):
+        if left[i] != right[i]:
+            explanation += ['At index %s diff: %r != %r' %
+                            (i, left[i], right[i])]
+            break
+    if len(left) > len(right):
+        explanation += ['Left contains more items, '
+            'first extra item: %s' % py.io.saferepr(left[len(right)],)]
+    elif len(left) < len(right):
+        explanation += ['Right contains more items, '
+            'first extra item: %s' % py.io.saferepr(right[len(left)],)]
+    return explanation # + _diff_text(py.std.pprint.pformat(left),
+                       #             py.std.pprint.pformat(right))
+
+
+def _compare_eq_set(left, right):
+    explanation = []
+    diff_left = left - right
+    diff_right = right - left
+    if diff_left:
+        explanation.append('Extra items in the left set:')
+        for item in diff_left:
+            explanation.append(py.io.saferepr(item))
+    if diff_right:
+        explanation.append('Extra items in the right set:')
+        for item in diff_right:
+            explanation.append(py.io.saferepr(item))
+    return explanation
+
+
+def _notin_text(term, text):
+    index = text.find(term)
+    head = text[:index]
+    tail = text[index+len(term):]
+    correct_text = head + tail
+    diff = _diff_text(correct_text, text)
+    newdiff = ['%s is contained here:' % py.io.saferepr(term, maxsize=42)]
+    for line in diff:
+        if line.startswith('Skipping'):
+            continue
+        if line.startswith('- '):
+            continue
+        if line.startswith('+ '):
+            newdiff.append('  ' + line[2:])
+        else:
+            newdiff.append(line)
+    return newdiff
diff --git a/_pytest/doctest.py b/_pytest/doctest.py
--- a/_pytest/doctest.py
+++ b/_pytest/doctest.py
@@ -59,7 +59,7 @@
                 inner_excinfo = py.code.ExceptionInfo(excinfo.value.exc_info)
                 lines += ["UNEXPECTED EXCEPTION: %s" %
                             repr(inner_excinfo.value)]
-
+                lines += py.std.traceback.format_exception(*excinfo.value.exc_info)
             return ReprFailDoctest(reprlocation, lines)
         else:
             return super(DoctestItem, self).repr_failure(excinfo)
diff --git a/_pytest/helpconfig.py b/_pytest/helpconfig.py
--- a/_pytest/helpconfig.py
+++ b/_pytest/helpconfig.py
@@ -16,9 +16,6 @@
     group.addoption('--traceconfig',
                action="store_true", dest="traceconfig", default=False,
                help="trace considerations of conftest.py files."),
-    group._addoption('--nomagic',
-               action="store_true", dest="nomagic", default=False,
-               help="don't reinterpret asserts, no traceback cutting. ")
     group.addoption('--debug',
                action="store_true", dest="debug", default=False,
                help="generate and show internal debugging information.")
diff --git a/_pytest/junitxml.py b/_pytest/junitxml.py
--- a/_pytest/junitxml.py
+++ b/_pytest/junitxml.py
@@ -65,7 +65,8 @@
 
 class LogXML(object):
     def __init__(self, logfile, prefix):
-        self.logfile = logfile
+        logfile = os.path.expanduser(os.path.expandvars(logfile))
+        self.logfile = os.path.normpath(logfile)
         self.prefix = prefix
         self.test_logs = []
         self.passed = self.skipped = 0
@@ -76,7 +77,7 @@
         names = report.nodeid.split("::")
         names[0] = names[0].replace("/", '.')
         names = tuple(names)
-        d = {'time': self._durations.pop(names, "0")}
+        d = {'time': self._durations.pop(report.nodeid, "0")}
         names = [x.replace(".py", "") for x in names if x != "()"]
         classnames = names[:-1]
         if self.prefix:
@@ -170,12 +171,11 @@
             self.append_skipped(report)
 
     def pytest_runtest_call(self, item, __multicall__):
-        names = tuple(item.listnames())
         start = time.time()
         try:
             return __multicall__.execute()
         finally:
-            self._durations[names] = time.time() - start
+            self._durations[item.nodeid] = time.time() - start
 
     def pytest_collectreport(self, report):
         if not report.passed:
diff --git a/_pytest/main.py b/_pytest/main.py
--- a/_pytest/main.py
+++ b/_pytest/main.py
@@ -46,23 +46,25 @@
 
 
 def pytest_namespace():
-    return dict(collect=dict(Item=Item, Collector=Collector, File=File))
+    collect = dict(Item=Item, Collector=Collector, File=File, Session=Session)
+    return dict(collect=collect)
         
 def pytest_configure(config):
     py.test.config = config # compatibiltiy
     if config.option.exitfirst:
         config.option.maxfail = 1
 
-def pytest_cmdline_main(config):
-    """ default command line protocol for initialization, session,
-    running tests and reporting. """
+def wrap_session(config, doit):
+    """Skeleton command line program"""
     session = Session(config)
     session.exitstatus = EXIT_OK
+    initstate = 0
     try:
         config.pluginmanager.do_configure(config)
+        initstate = 1
         config.hook.pytest_sessionstart(session=session)
-        config.hook.pytest_collection(session=session)
-        config.hook.pytest_runtestloop(session=session)
+        initstate = 2
+        doit(config, session)
     except pytest.UsageError:
         raise
     except KeyboardInterrupt:
@@ -77,18 +79,24 @@
             sys.stderr.write("mainloop: caught Spurious SystemExit!\n")
     if not session.exitstatus and session._testsfailed:
         session.exitstatus = EXIT_TESTSFAILED
-    config.hook.pytest_sessionfinish(session=session,
-        exitstatus=session.exitstatus)
-    config.pluginmanager.do_unconfigure(config)
+    if initstate >= 2:
+        config.hook.pytest_sessionfinish(session=session,
+            exitstatus=session.exitstatus)
+    if initstate >= 1:
+        config.pluginmanager.do_unconfigure(config)
     return session.exitstatus
 
+def pytest_cmdline_main(config):
+    return wrap_session(config, _main)
+
+def _main(config, session):
+    """ default command line protocol for initialization, session,
+    running tests and reporting. """
+    config.hook.pytest_collection(session=session)
+    config.hook.pytest_runtestloop(session=session)
+
 def pytest_collection(session):
-    session.perform_collect()
-    hook = session.config.hook
-    hook.pytest_collection_modifyitems(session=session,
-        config=session.config, items=session.items)
-    hook.pytest_collection_finish(session=session)
-    return True
+    return session.perform_collect()
 
 def pytest_runtestloop(session):
     if session.config.option.collectonly:
@@ -374,6 +382,16 @@
         return HookProxy(fspath, self.config)
 
     def perform_collect(self, args=None, genitems=True):
+        hook = self.config.hook
+        try:
+            items = self._perform_collect(args, genitems)
+            hook.pytest_collection_modifyitems(session=self,
+                config=self.config, items=items)
+        finally:
+            hook.pytest_collection_finish(session=self)
+        return items
+
+    def _perform_collect(self, args, genitems):
         if args is None:
             args = self.config.args
         self.trace("perform_collect", self, args)
diff --git a/_pytest/mark.py b/_pytest/mark.py
--- a/_pytest/mark.py
+++ b/_pytest/mark.py
@@ -153,7 +153,7 @@
 
     def __repr__(self):
         return "<MarkInfo %r args=%r kwargs=%r>" % (
-                self._name, self.args, self.kwargs)
+                self.name, self.args, self.kwargs)
 
 def pytest_itemcollected(item):
     if not isinstance(item, pytest.Function):
diff --git a/_pytest/pytester.py b/_pytest/pytester.py
--- a/_pytest/pytester.py
+++ b/_pytest/pytester.py
@@ -6,7 +6,7 @@
 import inspect
 import time
 from fnmatch import fnmatch
-from _pytest.main import Session
+from _pytest.main import Session, EXIT_OK
 from py.builtin import print_
 from _pytest.core import HookRelay
 
@@ -292,13 +292,19 @@
         assert '::' not in str(arg)
         p = py.path.local(arg)
         x = session.fspath.bestrelpath(p)
-        return session.perform_collect([x], genitems=False)[0]
+        config.hook.pytest_sessionstart(session=session)
+        res = session.perform_collect([x], genitems=False)[0]
+        config.hook.pytest_sessionfinish(session=session, exitstatus=EXIT_OK)
+        return res
 
     def getpathnode(self, path):
-        config = self.parseconfig(path)
+        config = self.parseconfigure(path)
         session = Session(config)
         x = session.fspath.bestrelpath(path)
-        return session.perform_collect([x], genitems=False)[0]
+        config.hook.pytest_sessionstart(session=session)
+        res = session.perform_collect([x], genitems=False)[0]
+        config.hook.pytest_sessionfinish(session=session, exitstatus=EXIT_OK)
+        return res
 
     def genitems(self, colitems):
         session = colitems[0].session
@@ -312,7 +318,9 @@
         config = self.parseconfigure(*args)
         rec = self.getreportrecorder(config)
         session = Session(config)
+        config.hook.pytest_sessionstart(session=session)
         session.perform_collect()
+        config.hook.pytest_sessionfinish(session=session, exitstatus=EXIT_OK)
         return session.items, rec
 
     def runitem(self, source):
@@ -382,6 +390,8 @@
             c.basetemp = py.path.local.make_numbered_dir(prefix="reparse",
                 keep=0, rootdir=self.tmpdir, lock_timeout=None)
             c.parse(args)
+            c.pluginmanager.do_configure(c)
+            self.request.addfinalizer(lambda: c.pluginmanager.do_unconfigure(c))
             return c
         finally:
             py.test.config = oldconfig
diff --git a/_pytest/python.py b/_pytest/python.py
--- a/_pytest/python.py
+++ b/_pytest/python.py
@@ -226,8 +226,13 @@
 
     def _importtestmodule(self):
         # we assume we are only called once per module
+        from _pytest import assertion
+        assertion.before_module_import(self)
         try:
-            mod = self.fspath.pyimport(ensuresyspath=True)
+            try:
+                mod = self.fspath.pyimport(ensuresyspath=True)
+            finally:
+                assertion.after_module_import(self)
         except SyntaxError:
             excinfo = py.code.ExceptionInfo()
             raise self.CollectError(excinfo.getrepr(style="short"))
@@ -374,7 +379,7 @@
         # test generators are seen as collectors but they also
         # invoke setup/teardown on popular request
         # (induced by the common "test_*" naming shared with normal tests)
-        self.config._setupstate.prepare(self)
+        self.session._setupstate.prepare(self)
         # see FunctionMixin.setup and test_setupstate_is_preserved_134
         self._preservedparent = self.parent.obj
         l = []
@@ -721,7 +726,7 @@
 
     def _addfinalizer(self, finalizer, scope):
         colitem = self._getscopeitem(scope)
-        self.config._setupstate.addfinalizer(
+        self._pyfuncitem.session._setupstate.addfinalizer(
             finalizer=finalizer, colitem=colitem)
 
     def __repr__(self):
@@ -742,8 +747,10 @@
         raise self.LookupError(msg)
 
 def showfuncargs(config):
-    from _pytest.main import Session
-    session = Session(config)
+    from _pytest.main import wrap_session
+    return wrap_session(config, _showfuncargs_main)
+
+def _showfuncargs_main(config, session):
     session.perform_collect()
     if session.items:
         plugins = session.items[0].getplugins()
diff --git a/_pytest/runner.py b/_pytest/runner.py
--- a/_pytest/runner.py
+++ b/_pytest/runner.py
@@ -14,17 +14,15 @@
 #
 # pytest plugin hooks
 
-# XXX move to pytest_sessionstart and fix py.test owns tests
-def pytest_configure(config):
-    config._setupstate = SetupState()
+def pytest_sessionstart(session):
+    session._setupstate = SetupState()
 
 def pytest_sessionfinish(session, exitstatus):
-    if hasattr(session.config, '_setupstate'):
-        hook = session.config.hook
-        rep = hook.pytest__teardown_final(session=session)
-        if rep:
-            hook.pytest__teardown_final_logerror(session=session, report=rep)
-            session.exitstatus = 1
+    hook = session.config.hook
+    rep = hook.pytest__teardown_final(session=session)
+    if rep:
+        hook.pytest__teardown_final_logerror(session=session, report=rep)
+        session.exitstatus = 1
 
 class NodeInfo:
     def __init__(self, location):
@@ -46,16 +44,16 @@
     return reports
 
 def pytest_runtest_setup(item):
-    item.config._setupstate.prepare(item)
+    item.session._setupstate.prepare(item)
 
 def pytest_runtest_call(item):
     item.runtest()
 
 def pytest_runtest_teardown(item):
-    item.config._setupstate.teardown_exact(item)
+    item.session._setupstate.teardown_exact(item)
 
 def pytest__teardown_final(session):
-    call = CallInfo(session.config._setupstate.teardown_all, when="teardown")
+    call = CallInfo(session._setupstate.teardown_all, when="teardown")
     if call.excinfo:
         ntraceback = call.excinfo.traceback .cut(excludepath=py._pydir)
         call.excinfo.traceback = ntraceback.filter()
diff --git a/lib-python/modified-2.7/test/test_descr.py b/lib-python/modified-2.7/test/test_descr.py
--- a/lib-python/modified-2.7/test/test_descr.py
+++ b/lib-python/modified-2.7/test/test_descr.py
@@ -4399,14 +4399,8 @@
         self.assertTrue(l.__add__ != [5].__add__)
         self.assertTrue(l.__add__ != l.__mul__)
         self.assertTrue(l.__add__.__name__ == '__add__')
-        if hasattr(l.__add__, '__self__'):
-            # CPython
-            self.assertTrue(l.__add__.__self__ is l)
-            self.assertTrue(l.__add__.__objclass__ is list)
-        else:
-            # Python implementations where [].__add__ is a normal bound method
-            self.assertTrue(l.__add__.im_self is l)
-            self.assertTrue(l.__add__.im_class is list)
+        self.assertTrue(l.__add__.__self__ is l)
+        self.assertTrue(l.__add__.__objclass__ is list)
         self.assertEqual(l.__add__.__doc__, list.__add__.__doc__)
         try:
             hash(l.__add__)
diff --git a/lib-python/modified-2.7/test/test_extcall.py b/lib-python/modified-2.7/test/test_extcall.py
--- a/lib-python/modified-2.7/test/test_extcall.py
+++ b/lib-python/modified-2.7/test/test_extcall.py
@@ -299,7 +299,7 @@
         def f(a):
             return a
         self.assertEqual(f(**{u'a': 4}), 4)
-        self.assertRaises(TypeError, lambda: f(**{u'st&#246;ren': 4}))
+        self.assertRaises(TypeError, f, **{u'st&#246;ren': 4})
         self.assertRaises(TypeError, f, **{u'someLongString':2})
         try:
             f(a=4, **{u'a': 4})
diff --git a/lib_pypy/_sqlite3.py b/lib_pypy/_sqlite3.py
--- a/lib_pypy/_sqlite3.py
+++ b/lib_pypy/_sqlite3.py
@@ -275,7 +275,8 @@
     return unicode(x, 'utf-8')
 
 class Connection(object):
-    def __init__(self, database, isolation_level="", detect_types=0, timeout=None, cached_statements=None, factory=None):
+    def __init__(self, database, timeout=5.0, detect_types=0, isolation_level="",
+                 check_same_thread=True, factory=None, cached_statements=100):
         self.db = c_void_p()
         if sqlite.sqlite3_open(database, byref(self.db)) != SQLITE_OK:
             raise OperationalError("Could not open database")
@@ -308,7 +309,8 @@
         self._aggregates = {}
         self.aggregate_instances = {}
         self._collations = {}
-        self.thread_ident = thread_get_ident()
+        if check_same_thread:
+            self.thread_ident = thread_get_ident()
 
     def _get_exception(self, error_code = None):
         if error_code is None:
diff --git a/lib_pypy/datetime.py b/lib_pypy/datetime.py
--- a/lib_pypy/datetime.py
+++ b/lib_pypy/datetime.py
@@ -1422,12 +1422,17 @@
             converter = _time.localtime
         else:
             converter = _time.gmtime
-        if 1 - (t % 1.0) < 0.000001:
-            t = float(int(t)) + 1
-        if t < 0:
-            t -= 1
+        if t < 0.0:
+            us = int(round(((-t) % 1.0) * 1000000))
+            if us > 0:
+                us = 1000000 - us
+                t -= 1.0
+        else:
+            us = int(round((t % 1.0) * 1000000))
+            if us == 1000000:
+                us = 0
+                t += 1.0
         y, m, d, hh, mm, ss, weekday, jday, dst = converter(t)
-        us = int((t % 1.0) * 1000000)
         ss = min(ss, 59)    # clamp out leap seconds if the platform has them
         result = cls(y, m, d, hh, mm, ss, us, tz)
         if tz is not None:
diff --git a/lib_pypy/msvcrt.py b/lib_pypy/msvcrt.py
--- a/lib_pypy/msvcrt.py
+++ b/lib_pypy/msvcrt.py
@@ -46,4 +46,42 @@
         e = get_errno()
         raise IOError(e, errno.errorcode[e])
 
+# Console I/O routines
+
+kbhit = _c._kbhit
+kbhit.argtypes = []
+kbhit.restype = ctypes.c_int
+
+getch = _c._getch
+getch.argtypes = []
+getch.restype = ctypes.c_char
+
+getwch = _c._getwch
+getwch.argtypes = []
+getwch.restype = ctypes.c_wchar
+
+getche = _c._getche
+getche.argtypes = []
+getche.restype = ctypes.c_char
+
+getwche = _c._getwche
+getwche.argtypes = []
+getwche.restype = ctypes.c_wchar
+
+putch = _c._putch
+putch.argtypes = [ctypes.c_char]
+putch.restype = None
+
+putwch = _c._putwch
+putwch.argtypes = [ctypes.c_wchar]
+putwch.restype = None
+
+ungetch = _c._ungetch
+ungetch.argtypes = [ctypes.c_char]
+ungetch.restype = None
+
+ungetwch = _c._ungetwch
+ungetwch.argtypes = [ctypes.c_wchar]
+ungetwch.restype = None
+
 del ctypes
diff --git a/lib_pypy/pypy_test/test_datetime.py b/lib_pypy/pypy_test/test_datetime.py
--- a/lib_pypy/pypy_test/test_datetime.py
+++ b/lib_pypy/pypy_test/test_datetime.py
@@ -32,4 +32,28 @@
     assert datetime.datetime.utcfromtimestamp(a).microsecond == 0
     assert datetime.datetime.utcfromtimestamp(a).second == 1
 
-    
+def test_more_datetime_rounding():
+    # this test verified on top of CPython 2.7 (using a plain
+    # "import datetime" above)
+    expected_results = {
+        -1000.0: 'datetime.datetime(1970, 1, 1, 0, 43, 20)',
+        -999.9999996: 'datetime.datetime(1970, 1, 1, 0, 43, 20)',
+        -999.4: 'datetime.datetime(1970, 1, 1, 0, 43, 20, 600000)',
+        -999.0000004: 'datetime.datetime(1970, 1, 1, 0, 43, 21)',
+        -1.0: 'datetime.datetime(1970, 1, 1, 0, 59, 59)',
+        -0.9999996: 'datetime.datetime(1970, 1, 1, 0, 59, 59)',
+        -0.4: 'datetime.datetime(1970, 1, 1, 0, 59, 59, 600000)',
+        -0.0000004: 'datetime.datetime(1970, 1, 1, 1, 0)',
+        0.0: 'datetime.datetime(1970, 1, 1, 1, 0)',
+        0.0000004: 'datetime.datetime(1970, 1, 1, 1, 0)',
+        0.4: 'datetime.datetime(1970, 1, 1, 1, 0, 0, 400000)',
+        0.9999996: 'datetime.datetime(1970, 1, 1, 1, 0, 1)',
+        1000.0: 'datetime.datetime(1970, 1, 1, 1, 16, 40)',
+        1000.0000004: 'datetime.datetime(1970, 1, 1, 1, 16, 40)',
+        1000.4: 'datetime.datetime(1970, 1, 1, 1, 16, 40, 400000)',
+        1000.9999996: 'datetime.datetime(1970, 1, 1, 1, 16, 41)',
+        1293843661.191: 'datetime.datetime(2011, 1, 1, 2, 1, 1, 191000)',
+        }
+    for t in sorted(expected_results):
+        dt = datetime.datetime.fromtimestamp(t)
+        assert repr(dt) == expected_results[t]
diff --git a/lib_pypy/resource.py b/lib_pypy/resource.py
--- a/lib_pypy/resource.py
+++ b/lib_pypy/resource.py
@@ -7,7 +7,7 @@
 
 from ctypes_support import standard_c_lib as libc
 from ctypes_support import get_errno
-from ctypes import Structure, c_int, c_long, byref, sizeof
+from ctypes import Structure, c_int, c_long, byref, sizeof, POINTER
 from errno import EINVAL, EPERM
 import _structseq
 
@@ -25,6 +25,8 @@
 _setrlimit = libc.setrlimit
 try:
     _getpagesize = libc.getpagesize
+    _getpagesize.argtypes = ()
+    _getpagesize.restype = c_int
 except AttributeError:
     from os import sysconf
     _getpagesize = None
@@ -61,6 +63,10 @@
         ("ru_nivcsw", c_long),
     )
 
+_getrusage.argtypes = (c_int, POINTER(_struct_rusage))
+_getrusage.restype = c_int
+
+
 class struct_rusage:
     __metaclass__ = _structseq.structseqtype
 
@@ -94,6 +100,12 @@
         ("rlim_max", rlim_t),
     )
 
+_getrlimit.argtypes = (c_int, POINTER(rlimit))
+_getrlimit.restype = c_int
+_setrlimit.argtypes = (c_int, POINTER(rlimit))
+_setrlimit.restype = c_int
+
+
 @builtinify
 def getrusage(who):
     ru = _struct_rusage()
diff --git a/py/__init__.py b/py/__init__.py
--- a/py/__init__.py
+++ b/py/__init__.py
@@ -8,7 +8,7 @@
 
 (c) Holger Krekel and others, 2004-2010
 """
-__version__ = '1.4.3'
+__version__ = '1.4.4.dev1'
 
 from py import _apipkg
 
@@ -70,10 +70,6 @@
         'getrawcode'        : '._code.code:getrawcode',
         'patch_builtins'    : '._code.code:patch_builtins',
         'unpatch_builtins'  : '._code.code:unpatch_builtins',
-        '_AssertionError'   : '._code.assertion:AssertionError',
-        '_reinterpret_old'  : '._code.assertion:reinterpret_old',
-        '_reinterpret'      : '._code.assertion:reinterpret',
-        '_reprcompare'      : '._code.assertion:_reprcompare',
     },
 
     # backports and additions of builtins
diff --git a/py/_code/_assertionnew.py b/py/_code/_assertionnew.py
deleted file mode 100644
--- a/py/_code/_assertionnew.py
+++ /dev/null
@@ -1,339 +0,0 @@
-"""
-Find intermediate evalutation results in assert statements through builtin AST.
-This should replace _assertionold.py eventually.
-"""
-
-import sys
-import ast
-
-import py
-from py._code.assertion import _format_explanation, BuiltinAssertionError
-
-
-if sys.platform.startswith("java") and sys.version_info < (2, 5, 2):
-    # See http://bugs.jython.org/issue1497
-    _exprs = ("BoolOp", "BinOp", "UnaryOp", "Lambda", "IfExp", "Dict",
-              "ListComp", "GeneratorExp", "Yield", "Compare", "Call",
-              "Repr", "Num", "Str", "Attribute", "Subscript", "Name",
-              "List", "Tuple")
-    _stmts = ("FunctionDef", "ClassDef", "Return", "Delete", "Assign",
-              "AugAssign", "Print", "For", "While", "If", "With", "Raise",
-              "TryExcept", "TryFinally", "Assert", "Import", "ImportFrom",
-              "Exec", "Global", "Expr", "Pass", "Break", "Continue")
-    _expr_nodes = set(getattr(ast, name) for name in _exprs)
-    _stmt_nodes = set(getattr(ast, name) for name in _stmts)
-    def _is_ast_expr(node):
-        return node.__class__ in _expr_nodes
-    def _is_ast_stmt(node):
-        return node.__class__ in _stmt_nodes
-else:
-    def _is_ast_expr(node):
-        return isinstance(node, ast.expr)
-    def _is_ast_stmt(node):
-        return isinstance(node, ast.stmt)
-
-
-class Failure(Exception):
-    """Error found while interpreting AST."""
-
-    def __init__(self, explanation=""):
-        self.cause = sys.exc_info()
-        self.explanation = explanation
-
-
-def interpret(source, frame, should_fail=False):
-    mod = ast.parse(source)
-    visitor = DebugInterpreter(frame)
-    try:
-        visitor.visit(mod)
-    except Failure:
-        failure = sys.exc_info()[1]
-        return getfailure(failure)
-    if should_fail:
-        return ("(assertion failed, but when it was re-run for "
-                "printing intermediate values, it did not fail.  Suggestions: "
-                "compute assert expression before the assert or use --no-assert)")
-
-def run(offending_line, frame=None):
-    if frame is None:
-        frame = py.code.Frame(sys._getframe(1))
-    return interpret(offending_line, frame)
-
-def getfailure(failure):
-    explanation = _format_explanation(failure.explanation)
-    value = failure.cause[1]
-    if str(value):
-        lines = explanation.splitlines()
-        if not lines:
-            lines.append("")
-        lines[0] += " << %s" % (value,)
-        explanation = "\n".join(lines)
-    text = "%s: %s" % (failure.cause[0].__name__, explanation)
-    if text.startswith("AssertionError: assert "):
-        text = text[16:]
-    return text
-
-
-operator_map = {
-    ast.BitOr : "|",
-    ast.BitXor : "^",
-    ast.BitAnd : "&",
-    ast.LShift : "<<",
-    ast.RShift : ">>",
-    ast.Add : "+",
-    ast.Sub : "-",
-    ast.Mult : "*",
-    ast.Div : "/",
-    ast.FloorDiv : "//",
-    ast.Mod : "%",
-    ast.Eq : "==",
-    ast.NotEq : "!=",
-    ast.Lt : "<",
-    ast.LtE : "<=",
-    ast.Gt : ">",
-    ast.GtE : ">=",
-    ast.Pow : "**",
-    ast.Is : "is",
-    ast.IsNot : "is not",
-    ast.In : "in",
-    ast.NotIn : "not in"
-}
-
-unary_map = {
-    ast.Not : "not %s",
-    ast.Invert : "~%s",
-    ast.USub : "-%s",
-    ast.UAdd : "+%s"
-}
-
-
-class DebugInterpreter(ast.NodeVisitor):
-    """Interpret AST nodes to gleam useful debugging information. """
-
-    def __init__(self, frame):
-        self.frame = frame
-
-    def generic_visit(self, node):
-        # Fallback when we don't have a special implementation.
-        if _is_ast_expr(node):
-            mod = ast.Expression(node)
-            co = self._compile(mod)
-            try:
-                result = self.frame.eval(co)
-            except Exception:
-                raise Failure()
-            explanation = self.frame.repr(result)
-            return explanation, result
-        elif _is_ast_stmt(node):
-            mod = ast.Module([node])
-            co = self._compile(mod, "exec")
-            try:
-                self.frame.exec_(co)
-            except Exception:
-                raise Failure()
-            return None, None
-        else:
-            raise AssertionError("can't handle %s" %(node,))
-
-    def _compile(self, source, mode="eval"):
-        return compile(source, "<assertion interpretation>", mode)
-
-    def visit_Expr(self, expr):
-        return self.visit(expr.value)
-
-    def visit_Module(self, mod):
-        for stmt in mod.body:
-            self.visit(stmt)
-
-    def visit_Name(self, name):
-        explanation, result = self.generic_visit(name)
-        # See if the name is local.
-        source = "%r in locals() is not globals()" % (name.id,)
-        co = self._compile(source)
-        try:
-            local = self.frame.eval(co)
-        except Exception:
-            # have to assume it isn't
-            local = False
-        if not local:
-            return name.id, result
-        return explanation, result
-
-    def visit_Compare(self, comp):
-        left = comp.left
-        left_explanation, left_result = self.visit(left)
-        for op, next_op in zip(comp.ops, comp.comparators):
-            next_explanation, next_result = self.visit(next_op)
-            op_symbol = operator_map[op.__class__]
-            explanation = "%s %s %s" % (left_explanation, op_symbol,
-                                        next_explanation)
-            source = "__exprinfo_left %s __exprinfo_right" % (op_symbol,)
-            co = self._compile(source)
-            try:
-                result = self.frame.eval(co, __exprinfo_left=left_result,
-                                         __exprinfo_right=next_result)
-            except Exception:
-                raise Failure(explanation)
-            try:
-                if not result:
-                    break
-            except KeyboardInterrupt:
-                raise
-            except:
-                break
-            left_explanation, left_result = next_explanation, next_result
-
-        rcomp = py.code._reprcompare
-        if rcomp:
-            res = rcomp(op_symbol, left_result, next_result)
-            if res:
-                explanation = res
-        return explanation, result
-
-    def visit_BoolOp(self, boolop):
-        is_or = isinstance(boolop.op, ast.Or)
-        explanations = []
-        for operand in boolop.values:
-            explanation, result = self.visit(operand)
-            explanations.append(explanation)
-            if result == is_or:
-                break
-        name = is_or and " or " or " and "
-        explanation = "(" + name.join(explanations) + ")"
-        return explanation, result
-
-    def visit_UnaryOp(self, unary):
-        pattern = unary_map[unary.op.__class__]
-        operand_explanation, operand_result = self.visit(unary.operand)
-        explanation = pattern % (operand_explanation,)
-        co = self._compile(pattern % ("__exprinfo_expr",))
-        try:
-            result = self.frame.eval(co, __exprinfo_expr=operand_result)
-        except Exception:
-            raise Failure(explanation)
-        return explanation, result
-
-    def visit_BinOp(self, binop):
-        left_explanation, left_result = self.visit(binop.left)
-        right_explanation, right_result = self.visit(binop.right)
-        symbol = operator_map[binop.op.__class__]
-        explanation = "(%s %s %s)" % (left_explanation, symbol,
-                                      right_explanation)
-        source = "__exprinfo_left %s __exprinfo_right" % (symbol,)
-        co = self._compile(source)
-        try:
-            result = self.frame.eval(co, __exprinfo_left=left_result,
-                                     __exprinfo_right=right_result)
-        except Exception:
-            raise Failure(explanation)
-        return explanation, result
-
-    def visit_Call(self, call):
-        func_explanation, func = self.visit(call.func)
-        arg_explanations = []
-        ns = {"__exprinfo_func" : func}
-        arguments = []
-        for arg in call.args:
-            arg_explanation, arg_result = self.visit(arg)
-            arg_name = "__exprinfo_%s" % (len(ns),)
-            ns[arg_name] = arg_result
-            arguments.append(arg_name)
-            arg_explanations.append(arg_explanation)
-        for keyword in call.keywords:
-            arg_explanation, arg_result = self.visit(keyword.value)
-            arg_name = "__exprinfo_%s" % (len(ns),)
-            ns[arg_name] = arg_result
-            keyword_source = "%s=%%s" % (keyword.arg)
-            arguments.append(keyword_source % (arg_name,))
-            arg_explanations.append(keyword_source % (arg_explanation,))
-        if call.starargs:
-            arg_explanation, arg_result = self.visit(call.starargs)
-            arg_name = "__exprinfo_star"
-            ns[arg_name] = arg_result
-            arguments.append("*%s" % (arg_name,))
-            arg_explanations.append("*%s" % (arg_explanation,))
-        if call.kwargs:
-            arg_explanation, arg_result = self.visit(call.kwargs)
-            arg_name = "__exprinfo_kwds"
-            ns[arg_name] = arg_result
-            arguments.append("**%s" % (arg_name,))
-            arg_explanations.append("**%s" % (arg_explanation,))
-        args_explained = ", ".join(arg_explanations)
-        explanation = "%s(%s)" % (func_explanation, args_explained)
-        args = ", ".join(arguments)
-        source = "__exprinfo_func(%s)" % (args,)
-        co = self._compile(source)
-        try:
-            result = self.frame.eval(co, **ns)
-        except Exception:
-            raise Failure(explanation)
-        pattern = "%s\n{%s = %s\n}"
-        rep = self.frame.repr(result)
-        explanation = pattern % (rep, rep, explanation)
-        return explanation, result
-
-    def _is_builtin_name(self, name):
-        pattern = "%r not in globals() and %r not in locals()"
-        source = pattern % (name.id, name.id)
-        co = self._compile(source)
-        try:
-            return self.frame.eval(co)
-        except Exception:
-            return False
-
-    def visit_Attribute(self, attr):
-        if not isinstance(attr.ctx, ast.Load):
-            return self.generic_visit(attr)
-        source_explanation, source_result = self.visit(attr.value)
-        explanation = "%s.%s" % (source_explanation, attr.attr)
-        source = "__exprinfo_expr.%s" % (attr.attr,)
-        co = self._compile(source)
-        try:
-            result = self.frame.eval(co, __exprinfo_expr=source_result)
-        except Exception:
-            raise Failure(explanation)
-        explanation = "%s\n{%s = %s.%s\n}" % (self.frame.repr(result),
-                                              self.frame.repr(result),
-                                              source_explanation, attr.attr)
-        # Check if the attr is from an instance.
-        source = "%r in getattr(__exprinfo_expr, '__dict__', {})"
-        source = source % (attr.attr,)
-        co = self._compile(source)
-        try:
-            from_instance = self.frame.eval(co, __exprinfo_expr=source_result)
-        except Exception:
-            from_instance = True
-        if from_instance:
-            rep = self.frame.repr(result)
-            pattern = "%s\n{%s = %s\n}"
-            explanation = pattern % (rep, rep, explanation)
-        return explanation, result
-
-    def visit_Assert(self, assrt):
-        test_explanation, test_result = self.visit(assrt.test)
-        if test_explanation.startswith("False\n{False =") and \
-                test_explanation.endswith("\n"):
-            test_explanation = test_explanation[15:-2]
-        explanation = "assert %s" % (test_explanation,)
-        if not test_result:
-            try:
-                raise BuiltinAssertionError
-            except Exception:
-                raise Failure(explanation)
-        return explanation, test_result
-
-    def visit_Assign(self, assign):
-        value_explanation, value_result = self.visit(assign.value)
-        explanation = "... = %s" % (value_explanation,)
-        name = ast.Name("__exprinfo_expr", ast.Load(),
-                        lineno=assign.value.lineno,
-                        col_offset=assign.value.col_offset)
-        new_assign = ast.Assign(assign.targets, name, lineno=assign.lineno,
-                                col_offset=assign.col_offset)
-        mod = ast.Module([new_assign])
-        co = self._compile(mod, "exec")
-        try:
-            self.frame.exec_(co, __exprinfo_expr=value_result)
-        except Exception:
-            raise Failure(explanation)
-        return explanation, value_result
diff --git a/py/_code/_assertionold.py b/py/_code/_assertionold.py
deleted file mode 100644
--- a/py/_code/_assertionold.py
+++ /dev/null
@@ -1,555 +0,0 @@
-import py
-import sys, inspect
-from compiler import parse, ast, pycodegen
-from py._code.assertion import BuiltinAssertionError, _format_explanation
-
-passthroughex = py.builtin._sysex
-
-class Failure:
-    def __init__(self, node):
-        self.exc, self.value, self.tb = sys.exc_info()
-        self.node = node
-
-class View(object):
-    """View base class.
-
-    If C is a subclass of View, then C(x) creates a proxy object around
-    the object x.  The actual class of the proxy is not C in general,
-    but a *subclass* of C determined by the rules below.  To avoid confusion
-    we call view class the class of the proxy (a subclass of C, so of View)
-    and object class the class of x.
-
-    Attributes and methods not found in the proxy are automatically read on x.
-    Other operations like setting attributes are performed on the proxy, as
-    determined by its view class.  The object x is available from the proxy
-    as its __obj__ attribute.
-
-    The view class selection is determined by the __view__ tuples and the
-    optional __viewkey__ method.  By default, the selected view class is the
-    most specific subclass of C whose __view__ mentions the class of x.
-    If no such subclass is found, the search proceeds with the parent
-    object classes.  For example, C(True) will first look for a subclass
-    of C with __view__ = (..., bool, ...) and only if it doesn't find any
-    look for one with __view__ = (..., int, ...), and then ..., object,...
-    If everything fails the class C itself is considered to be the default.
-
-    Alternatively, the view class selection can be driven by another aspect
-    of the object x, instead of the class of x, by overriding __viewkey__.
-    See last example at the end of this module.
-    """
-
-    _viewcache = {}
-    __view__ = ()
-
-    def __new__(rootclass, obj, *args, **kwds):
-        self = object.__new__(rootclass)
-        self.__obj__ = obj
-        self.__rootclass__ = rootclass
-        key = self.__viewkey__()
-        try:
-            self.__class__ = self._viewcache[key]
-        except KeyError:
-            self.__class__ = self._selectsubclass(key)
-        return self
-
-    def __getattr__(self, attr):
-        # attributes not found in the normal hierarchy rooted on View
-        # are looked up in the object's real class
-        return getattr(self.__obj__, attr)
-
-    def __viewkey__(self):
-        return self.__obj__.__class__
-
-    def __matchkey__(self, key, subclasses):
-        if inspect.isclass(key):
-            keys = inspect.getmro(key)
-        else:
-            keys = [key]
-        for key in keys:
-            result = [C for C in subclasses if key in C.__view__]
-            if result:
-                return result
-        return []
-
-    def _selectsubclass(self, key):
-        subclasses = list(enumsubclasses(self.__rootclass__))
-        for C in subclasses:
-            if not isinstance(C.__view__, tuple):
-                C.__view__ = (C.__view__,)
-        choices = self.__matchkey__(key, subclasses)
-        if not choices:
-            return self.__rootclass__
-        elif len(choices) == 1:
-            return choices[0]
-        else:
-            # combine the multiple choices
-            return type('?', tuple(choices), {})
-
-    def __repr__(self):
-        return '%s(%r)' % (self.__rootclass__.__name__, self.__obj__)
-
-
-def enumsubclasses(cls):
-    for subcls in cls.__subclasses__():
-        for subsubclass in enumsubclasses(subcls):
-            yield subsubclass
-    yield cls
-
-
-class Interpretable(View):
-    """A parse tree node with a few extra methods."""
-    explanation = None
-
-    def is_builtin(self, frame):
-        return False
-
-    def eval(self, frame):
-        # fall-back for unknown expression nodes
-        try:
-            expr = ast.Expression(self.__obj__)
-            expr.filename = '<eval>'
-            self.__obj__.filename = '<eval>'
-            co = pycodegen.ExpressionCodeGenerator(expr).getCode()
-            result = frame.eval(co)
-        except passthroughex:
-            raise
-        except:
-            raise Failure(self)
-        self.result = result
-        self.explanation = self.explanation or frame.repr(self.result)
-
-    def run(self, frame):
-        # fall-back for unknown statement nodes
-        try:
-            expr = ast.Module(None, ast.Stmt([self.__obj__]))
-            expr.filename = '<run>'
-            co = pycodegen.ModuleCodeGenerator(expr).getCode()
-            frame.exec_(co)
-        except passthroughex:
-            raise
-        except:
-            raise Failure(self)
-
-    def nice_explanation(self):
-        return _format_explanation(self.explanation)
-
-
-class Name(Interpretable):
-    __view__ = ast.Name
-
-    def is_local(self, frame):
-        source = '%r in locals() is not globals()' % self.name
-        try:
-            return frame.is_true(frame.eval(source))
-        except passthroughex:
-            raise
-        except:
-            return False
-
-    def is_global(self, frame):
-        source = '%r in globals()' % self.name
-        try:
-            return frame.is_true(frame.eval(source))
-        except passthroughex:
-            raise
-        except:
-            return False
-
-    def is_builtin(self, frame):
-        source = '%r not in locals() and %r not in globals()' % (
-            self.name, self.name)
-        try:
-            return frame.is_true(frame.eval(source))
-        except passthroughex:
-            raise
-        except:
-            return False
-
-    def eval(self, frame):
-        super(Name, self).eval(frame)
-        if not self.is_local(frame):
-            self.explanation = self.name
-
-class Compare(Interpretable):
-    __view__ = ast.Compare
-
-    def eval(self, frame):
-        expr = Interpretable(self.expr)
-        expr.eval(frame)
-        for operation, expr2 in self.ops:
-            if hasattr(self, 'result'):
-                # shortcutting in chained expressions
-                if not frame.is_true(self.result):
-                    break
-            expr2 = Interpretable(expr2)
-            expr2.eval(frame)
-            self.explanation = "%s %s %s" % (
-                expr.explanation, operation, expr2.explanation)
-            source = "__exprinfo_left %s __exprinfo_right" % operation
-            try:
-                self.result = frame.eval(source,
-                                         __exprinfo_left=expr.result,
-                                         __exprinfo_right=expr2.result)
-            except passthroughex:
-                raise
-            except:
-                raise Failure(self)
-            expr = expr2
-
-class And(Interpretable):
-    __view__ = ast.And
-
-    def eval(self, frame):
-        explanations = []
-        for expr in self.nodes:
-            expr = Interpretable(expr)
-            expr.eval(frame)
-            explanations.append(expr.explanation)
-            self.result = expr.result
-            if not frame.is_true(expr.result):
-                break
-        self.explanation = '(' + ' and '.join(explanations) + ')'
-
-class Or(Interpretable):
-    __view__ = ast.Or
-
-    def eval(self, frame):
-        explanations = []
-        for expr in self.nodes:
-            expr = Interpretable(expr)
-            expr.eval(frame)
-            explanations.append(expr.explanation)
-            self.result = expr.result
-            if frame.is_true(expr.result):
-                break
-        self.explanation = '(' + ' or '.join(explanations) + ')'
-
-
-# == Unary operations ==
-keepalive = []
-for astclass, astpattern in {
-    ast.Not    : 'not __exprinfo_expr',
-    ast.Invert : '(~__exprinfo_expr)',
-    }.items():
-
-    class UnaryArith(Interpretable):
-        __view__ = astclass
-
-        def eval(self, frame, astpattern=astpattern):
-            expr = Interpretable(self.expr)
-            expr.eval(frame)
-            self.explanation = astpattern.replace('__exprinfo_expr',
-                                                  expr.explanation)
-            try:
-                self.result = frame.eval(astpattern,
-                                         __exprinfo_expr=expr.result)
-            except passthroughex:
-                raise
-            except:
-                raise Failure(self)
-
-    keepalive.append(UnaryArith)
-
-# == Binary operations ==
-for astclass, astpattern in {
-    ast.Add    : '(__exprinfo_left + __exprinfo_right)',
-    ast.Sub    : '(__exprinfo_left - __exprinfo_right)',
-    ast.Mul    : '(__exprinfo_left * __exprinfo_right)',
-    ast.Div    : '(__exprinfo_left / __exprinfo_right)',
-    ast.Mod    : '(__exprinfo_left % __exprinfo_right)',
-    ast.Power  : '(__exprinfo_left ** __exprinfo_right)',
-    }.items():
-
-    class BinaryArith(Interpretable):
-        __view__ = astclass
-
-        def eval(self, frame, astpattern=astpattern):
-            left = Interpretable(self.left)
-            left.eval(frame)
-            right = Interpretable(self.right)
-            right.eval(frame)
-            self.explanation = (astpattern
-                                .replace('__exprinfo_left',  left .explanation)
-                                .replace('__exprinfo_right', right.explanation))
-            try:
-                self.result = frame.eval(astpattern,
-                                         __exprinfo_left=left.result,
-                                         __exprinfo_right=right.result)
-            except passthroughex:
-                raise
-            except:
-                raise Failure(self)
-
-    keepalive.append(BinaryArith)
-
-
-class CallFunc(Interpretable):
-    __view__ = ast.CallFunc
-
-    def is_bool(self, frame):
-        source = 'isinstance(__exprinfo_value, bool)'
-        try:
-            return frame.is_true(frame.eval(source,
-                                            __exprinfo_value=self.result))
-        except passthroughex:
-            raise
-        except:
-            return False
-
-    def eval(self, frame):
-        node = Interpretable(self.node)
-        node.eval(frame)
-        explanations = []
-        vars = {'__exprinfo_fn': node.result}
-        source = '__exprinfo_fn('
-        for a in self.args:
-            if isinstance(a, ast.Keyword):
-                keyword = a.name
-                a = a.expr
-            else:
-                keyword = None
-            a = Interpretable(a)
-            a.eval(frame)
-            argname = '__exprinfo_%d' % len(vars)
-            vars[argname] = a.result
-            if keyword is None:
-                source += argname + ','
-                explanations.append(a.explanation)
-            else:
-                source += '%s=%s,' % (keyword, argname)
-                explanations.append('%s=%s' % (keyword, a.explanation))
-        if self.star_args:
-            star_args = Interpretable(self.star_args)
-            star_args.eval(frame)
-            argname = '__exprinfo_star'
-            vars[argname] = star_args.result
-            source += '*' + argname + ','
-            explanations.append('*' + star_args.explanation)
-        if self.dstar_args:
-            dstar_args = Interpretable(self.dstar_args)
-            dstar_args.eval(frame)
-            argname = '__exprinfo_kwds'
-            vars[argname] = dstar_args.result
-            source += '**' + argname + ','
-            explanations.append('**' + dstar_args.explanation)
-        self.explanation = "%s(%s)" % (
-            node.explanation, ', '.join(explanations))
-        if source.endswith(','):
-            source = source[:-1]
-        source += ')'
-        try:
-            self.result = frame.eval(source, **vars)
-        except passthroughex:
-            raise
-        except:
-            raise Failure(self)
-        if not node.is_builtin(frame) or not self.is_bool(frame):
-            r = frame.repr(self.result)
-            self.explanation = '%s\n{%s = %s\n}' % (r, r, self.explanation)
-
-class Getattr(Interpretable):
-    __view__ = ast.Getattr
-
-    def eval(self, frame):
-        expr = Interpretable(self.expr)
-        expr.eval(frame)
-        source = '__exprinfo_expr.%s' % self.attrname
-        try:
-            self.result = frame.eval(source, __exprinfo_expr=expr.result)
-        except passthroughex:
-            raise
-        except:
-            raise Failure(self)
-        self.explanation = '%s.%s' % (expr.explanation, self.attrname)
-        # if the attribute comes from the instance, its value is interesting
-        source = ('hasattr(__exprinfo_expr, "__dict__") and '
-                  '%r in __exprinfo_expr.__dict__' % self.attrname)
-        try:
-            from_instance = frame.is_true(
-                frame.eval(source, __exprinfo_expr=expr.result))
-        except passthroughex:
-            raise
-        except:
-            from_instance = True
-        if from_instance:
-            r = frame.repr(self.result)
-            self.explanation = '%s\n{%s = %s\n}' % (r, r, self.explanation)
-
-# == Re-interpretation of full statements ==
-
-class Assert(Interpretable):
-    __view__ = ast.Assert
-
-    def run(self, frame):
-        test = Interpretable(self.test)
-        test.eval(frame)
-        # simplify 'assert False where False = ...'
-        if (test.explanation.startswith('False\n{False = ') and
-            test.explanation.endswith('\n}')):
-            test.explanation = test.explanation[15:-2]
-        # print the result as  'assert <explanation>'
-        self.result = test.result
-        self.explanation = 'assert ' + test.explanation
-        if not frame.is_true(test.result):
-            try:
-                raise BuiltinAssertionError
-            except passthroughex:
-                raise
-            except:
-                raise Failure(self)
-
-class Assign(Interpretable):
-    __view__ = ast.Assign
-
-    def run(self, frame):
-        expr = Interpretable(self.expr)
-        expr.eval(frame)
-        self.result = expr.result
-        self.explanation = '... = ' + expr.explanation
-        # fall-back-run the rest of the assignment
-        ass = ast.Assign(self.nodes, ast.Name('__exprinfo_expr'))
-        mod = ast.Module(None, ast.Stmt([ass]))
-        mod.filename = '<run>'
-        co = pycodegen.ModuleCodeGenerator(mod).getCode()
-        try:
-            frame.exec_(co, __exprinfo_expr=expr.result)
-        except passthroughex:
-            raise
-        except:
-            raise Failure(self)
-
-class Discard(Interpretable):
-    __view__ = ast.Discard
-
-    def run(self, frame):
-        expr = Interpretable(self.expr)
-        expr.eval(frame)
-        self.result = expr.result
-        self.explanation = expr.explanation
-
-class Stmt(Interpretable):
-    __view__ = ast.Stmt
-
-    def run(self, frame):
-        for stmt in self.nodes:
-            stmt = Interpretable(stmt)
-            stmt.run(frame)
-
-
-def report_failure(e):
-    explanation = e.node.nice_explanation()
-    if explanation:
-        explanation = ", in: " + explanation
-    else:
-        explanation = ""
-    sys.stdout.write("%s: %s%s\n" % (e.exc.__name__, e.value, explanation))
-
-def check(s, frame=None):
-    if frame is None:
-        frame = sys._getframe(1)
-        frame = py.code.Frame(frame)
-    expr = parse(s, 'eval')
-    assert isinstance(expr, ast.Expression)
-    node = Interpretable(expr.node)
-    try:
-        node.eval(frame)
-    except passthroughex:
-        raise
-    except Failure:
-        e = sys.exc_info()[1]
-        report_failure(e)
-    else:
-        if not frame.is_true(node.result):
-            sys.stderr.write("assertion failed: %s\n" % node.nice_explanation())
-
-
-###########################################################
-# API / Entry points
-# #########################################################
-
-def interpret(source, frame, should_fail=False):
-    module = Interpretable(parse(source, 'exec').node)
-    #print "got module", module
-    if isinstance(frame, py.std.types.FrameType):
-        frame = py.code.Frame(frame)
-    try:
-        module.run(frame)
-    except Failure:
-        e = sys.exc_info()[1]
-        return getfailure(e)
-    except passthroughex:
-        raise
-    except:
-        import traceback
-        traceback.print_exc()
-    if should_fail:
-        return ("(assertion failed, but when it was re-run for "
-                "printing intermediate values, it did not fail.  Suggestions: "
-                "compute assert expression before the assert or use --nomagic)")
-    else:
-        return None
-
-def getmsg(excinfo):
-    if isinstance(excinfo, tuple):
-        excinfo = py.code.ExceptionInfo(excinfo)
-    #frame, line = gettbline(tb)
-    #frame = py.code.Frame(frame)
-    #return interpret(line, frame)
-
-    tb = excinfo.traceback[-1]
-    source = str(tb.statement).strip()
-    x = interpret(source, tb.frame, should_fail=True)
-    if not isinstance(x, str):
-        raise TypeError("interpret returned non-string %r" % (x,))
-    return x
-
-def getfailure(e):
-    explanation = e.node.nice_explanation()
-    if str(e.value):
-        lines = explanation.split('\n')
-        lines[0] += "  << %s" % (e.value,)
-        explanation = '\n'.join(lines)
-    text = "%s: %s" % (e.exc.__name__, explanation)
-    if text.startswith('AssertionError: assert '):
-        text = text[16:]
-    return text
-
-def run(s, frame=None):
-    if frame is None:
-        frame = sys._getframe(1)
-        frame = py.code.Frame(frame)
-    module = Interpretable(parse(s, 'exec').node)
-    try:
-        module.run(frame)
-    except Failure:
-        e = sys.exc_info()[1]
-        report_failure(e)
-
-
-if __name__ == '__main__':
-    # example:
-    def f():
-        return 5
-    def g():
-        return 3
-    def h(x):
-        return 'never'
-    check("f() * g() == 5")
-    check("not f()")
-    check("not (f() and g() or 0)")
-    check("f() == g()")
-    i = 4
-    check("i == f()")
-    check("len(f()) == 0")
-    check("isinstance(2+3+4, float)")
-
-    run("x = i")
-    check("x == 5")
-
-    run("assert not f(), 'oops'")
-    run("a, b, c = 1, 2")
-    run("a, b, c = f()")
-
-    check("max([f(),g()]) == 4")
-    check("'hello'[g()] == 'h'")
-    run("'guk%d' % h(f())")
diff --git a/py/_code/assertion.py b/py/_code/assertion.py
deleted file mode 100644
--- a/py/_code/assertion.py
+++ /dev/null
@@ -1,94 +0,0 @@
-import sys
-import py
-
-BuiltinAssertionError = py.builtin.builtins.AssertionError
-
-_reprcompare = None # if set, will be called by assert reinterp for comparison ops
-
-def _format_explanation(explanation):
-    """This formats an explanation
-
-    Normally all embedded newlines are escaped, however there are
-    three exceptions: \n{, \n} and \n~.  The first two are intended
-    cover nested explanations, see function and attribute explanations
-    for examples (.visit_Call(), visit_Attribute()).  The last one is
-    for when one explanation needs to span multiple lines, e.g. when
-    displaying diffs.
-    """
-    raw_lines = (explanation or '').split('\n')
-    # escape newlines not followed by {, } and ~
-    lines = [raw_lines[0]]
-    for l in raw_lines[1:]:
-        if l.startswith('{') or l.startswith('}') or l.startswith('~'):
-            lines.append(l)
-        else:
-            lines[-1] += '\\n' + l
-
-    result = lines[:1]
-    stack = [0]
-    stackcnt = [0]
-    for line in lines[1:]:
-        if line.startswith('{'):
-            if stackcnt[-1]:
-                s = 'and   '
-            else:
-                s = 'where '
-            stack.append(len(result))
-            stackcnt[-1] += 1
-            stackcnt.append(0)
-            result.append(' +' + '  '*(len(stack)-1) + s + line[1:])
-        elif line.startswith('}'):
-            assert line.startswith('}')
-            stack.pop()
-            stackcnt.pop()
-            result[stack[-1]] += line[1:]
-        else:
-            assert line.startswith('~')
-            result.append('  '*len(stack) + line[1:])
-    assert len(stack) == 1
-    return '\n'.join(result)
-
-
-class AssertionError(BuiltinAssertionError):
-    def __init__(self, *args):
-        BuiltinAssertionError.__init__(self, *args)
-        if args:
-            try:
-                self.msg = str(args[0])
-            except py.builtin._sysex:
-                raise
-            except:
-                self.msg = "<[broken __repr__] %s at %0xd>" %(
-                    args[0].__class__, id(args[0]))
-        else:
-            f = py.code.Frame(sys._getframe(1))
-            try:
-                source = f.code.fullsource
-                if source is not None:
-                    try:
-                        source = source.getstatement(f.lineno, assertion=True)
-                    except IndexError:
-                        source = None
-                    else:
-                        source = str(source.deindent()).strip()
-            except py.error.ENOENT:
-                source = None
-                # this can also occur during reinterpretation, when the
-                # co_filename is set to "<run>".
-            if source:
-                self.msg = reinterpret(source, f, should_fail=True)
-            else:
-                self.msg = "<could not determine information>"
-            if not self.args:
-                self.args = (self.msg,)
-
-if sys.version_info > (3, 0):
-    AssertionError.__module__ = "builtins"
-    reinterpret_old = "old reinterpretation not available for py3"
-else:
-    from py._code._assertionold import interpret as reinterpret_old
-if sys.version_info >= (2, 6) or (sys.platform.startswith("java")):
-    from py._code._assertionnew import interpret as reinterpret
-else:
-    reinterpret = reinterpret_old
-
diff --git a/py/_code/code.py b/py/_code/code.py
--- a/py/_code/code.py
+++ b/py/_code/code.py
@@ -145,17 +145,6 @@
         return self.frame.f_locals
     locals = property(getlocals, None, None, "locals of underlaying frame")
 
-    def reinterpret(self):
-        """Reinterpret the failing statement and returns a detailed information
-           about what operations are performed."""
-        if self.exprinfo is None:
-            source = str(self.statement).strip()
-            x = py.code._reinterpret(source, self.frame, should_fail=True)
-            if not isinstance(x, str):
-                raise TypeError("interpret returned non-string %r" % (x,))
-            self.exprinfo = x
-        return self.exprinfo
-
     def getfirstlinesource(self):
         # on Jython this firstlineno can be -1 apparently
         return max(self.frame.code.firstlineno, 0)
@@ -310,7 +299,7 @@
         #     ExceptionInfo-like classes may have different attributes.
         if tup is None:
             tup = sys.exc_info()
-            if exprinfo is None and isinstance(tup[1], py.code._AssertionError):
+            if exprinfo is None and isinstance(tup[1], AssertionError):
                 exprinfo = getattr(tup[1], 'msg', None)
                 if exprinfo is None:
                     exprinfo = str(tup[1])
@@ -690,22 +679,15 @@
 
 oldbuiltins = {}
 
-def patch_builtins(assertion=True, compile=True):
-    """ put compile and AssertionError builtins to Python's builtins. """
-    if assertion:
-        from py._code import assertion
-        l = oldbuiltins.setdefault('AssertionError', [])
-        l.append(py.builtin.builtins.AssertionError)
-        py.builtin.builtins.AssertionError = assertion.AssertionError
+def patch_builtins(compile=True):
+    """ put compile builtins to Python's builtins. """
     if compile:
         l = oldbuiltins.setdefault('compile', [])
         l.append(py.builtin.builtins.compile)
         py.builtin.builtins.compile = py.code.compile
 
-def unpatch_builtins(assertion=True, compile=True):
+def unpatch_builtins(compile=True):
     """ remove compile and AssertionError builtins from Python builtins. """
-    if assertion:
-        py.builtin.builtins.AssertionError = oldbuiltins['AssertionError'].pop()
     if compile:
         py.builtin.builtins.compile = oldbuiltins['compile'].pop()
 
diff --git a/pypy/annotation/bookkeeper.py b/pypy/annotation/bookkeeper.py
--- a/pypy/annotation/bookkeeper.py
+++ b/pypy/annotation/bookkeeper.py
@@ -299,12 +299,13 @@
             listdef.generalize_range_step(flags['range_step'])
         return SomeList(listdef)
 
-    def getdictdef(self, is_r_dict=False):
+    def getdictdef(self, is_r_dict=False, force_non_null=False):
         """Get the DictDef associated with the current position."""
         try:
             dictdef = self.dictdefs[self.position_key]
         except KeyError:
-            dictdef = DictDef(self, is_r_dict=is_r_dict)
+            dictdef = DictDef(self, is_r_dict=is_r_dict,
+                              force_non_null=force_non_null)
             self.dictdefs[self.position_key] = dictdef
         return dictdef
 
diff --git a/pypy/annotation/builtin.py b/pypy/annotation/builtin.py
--- a/pypy/annotation/builtin.py
+++ b/pypy/annotation/builtin.py
@@ -311,8 +311,14 @@
 def robjmodel_we_are_translated():
     return immutablevalue(True)
 
-def robjmodel_r_dict(s_eqfn, s_hashfn):
-    dictdef = getbookkeeper().getdictdef(is_r_dict=True)
+def robjmodel_r_dict(s_eqfn, s_hashfn, s_force_non_null=None):
+    if s_force_non_null is None:
+        force_non_null = False
+    else:
+        assert s_force_non_null.is_constant()
+        force_non_null = s_force_non_null.const
+    dictdef = getbookkeeper().getdictdef(is_r_dict=True,
+                                         force_non_null=force_non_null)
     dictdef.dictkey.update_rdict_annotations(s_eqfn, s_hashfn)
     return SomeDict(dictdef)
 
diff --git a/pypy/annotation/dictdef.py b/pypy/annotation/dictdef.py
--- a/pypy/annotation/dictdef.py
+++ b/pypy/annotation/dictdef.py
@@ -85,12 +85,14 @@
 
     def __init__(self, bookkeeper, s_key = s_ImpossibleValue,
                                  s_value = s_ImpossibleValue,
-                               is_r_dict = False):
+                               is_r_dict = False,
+                           force_non_null = False):
         self.dictkey = DictKey(bookkeeper, s_key, is_r_dict)
         self.dictkey.itemof[self] = True
         self.dictvalue = DictValue(bookkeeper, s_value)
         self.dictvalue.itemof[self] = True
         self.bookkeeper = bookkeeper
+        self.force_non_null = force_non_null
 
     def read_key(self, position_key=None):
         if position_key is None:
diff --git a/pypy/config/pypyoption.py b/pypy/config/pypyoption.py
--- a/pypy/config/pypyoption.py
+++ b/pypy/config/pypyoption.py
@@ -84,6 +84,7 @@
     "_rawffi": [("objspace.usemodules.struct", True)],
     "cpyext": [("translation.secondaryentrypoints", "cpyext"),
                ("translation.shared", sys.platform == "win32")],
+    "_ffi":    [("translation.jit_ffi", True)],
 }
 
 module_import_dependencies = {
diff --git a/pypy/config/test/test_pypyoption.py b/pypy/config/test/test_pypyoption.py
--- a/pypy/config/test/test_pypyoption.py
+++ b/pypy/config/test/test_pypyoption.py
@@ -73,3 +73,7 @@
             fn = prefix + "." + path + ".txt"
             yield check_file_exists, fn
 
+def test__ffi_opt():
+    config = get_pypy_config(translating=True)
+    config.objspace.usemodules._ffi = True
+    assert config.translation.jit_ffi
diff --git a/pypy/config/translationoption.py b/pypy/config/translationoption.py
--- a/pypy/config/translationoption.py
+++ b/pypy/config/translationoption.py
@@ -117,6 +117,8 @@
     ChoiceOption("jit_profiler", "integrate profiler support into the JIT",
                  ["off", "oprofile"],
                  default="off"),
+    # jit_ffi is automatically turned on by withmod-_ffi (which is enabled by default)
+    BoolOption("jit_ffi", "optimize libffi calls", default=False, cmdline=None),
 
     # misc
     BoolOption("verbose", "Print extra information", default=False),
diff --git a/pypy/doc/garbage_collection.rst b/pypy/doc/garbage_collection.rst
--- a/pypy/doc/garbage_collection.rst
+++ b/pypy/doc/garbage_collection.rst
@@ -212,90 +212,4 @@
   becomes free garbage, to be collected at the next major collection.
 
 
-Minimark GC
------------
-
-This is a simplification and rewrite of the ideas from the Hybrid GC.
-It uses a nursery for the young objects, and mark-and-sweep for the old
-objects.  This is a moving GC, but objects may only move once (from
-the nursery to the old stage).
-
-The main difference with the Hybrid GC is that the mark-and-sweep
-objects (the "old stage") are directly handled by the GC's custom
-allocator, instead of being handled by malloc() calls.  The gain is that
-it is then possible, during a major collection, to walk through all old
-generation objects without needing to store a list of pointers to them.
-So as a first approximation, when compared to the Hybrid GC, the
-Minimark GC saves one word of memory per old object.
-
-There are a number of environment variables that can be tweaked to
-influence the GC.  (Their default value should be ok for most usages.)
-You can read more about them at the start of
-`pypy/rpython/memory/gc/minimark.py`_.
-
-In more details:
-
-- The small newly malloced objects are allocated in the nursery (case 1).
-  All objects living in the nursery are "young".
-
-- The big objects are always handled directly by the system malloc().
-  But the big newly malloced objects are still "young" when they are
-  allocated (case 2), even though they don't live in the nursery.
-
-- When the nursery is full, we do a minor collection, i.e. we find
-  which "young" objects are still alive (from cases 1 and 2).  The
-  "young" flag is then removed.  The surviving case 1 objects are moved
-  to the old stage. The dying case 2 objects are immediately freed.
-
-- The old stage is an area of memory containing old (small) objects.  It
-  is handled by `pypy/rpython/memory/gc/minimarkpage.py`_.  It is organized
-  as "arenas" of 256KB or 512KB, subdivided into "pages" of 4KB or 8KB.
-  Each page can either be free, or contain small objects of all the same
-  size.  Furthermore at any point in time each object location can be
-  either allocated or freed.  The basic design comes from ``obmalloc.c``
-  from CPython (which itself comes from the same source as the Linux
-  system malloc()).
-
-- New objects are added to the old stage at every minor collection.
-  Immediately after a minor collection, when we reach some threshold, we
-  trigger a major collection.  This is the mark-and-sweep step.  It walks
-  over *all* objects (mark), and then frees some fraction of them (sweep).
-  This means that the only time when we want to free objects is while
-  walking over all of them; we never ask to free an object given just its
-  address.  This allows some simplifications and memory savings when
-  compared to ``obmalloc.c``.
-
-- As with all generational collectors, this GC needs a write barrier to
-  record which old objects have a reference to young objects.
-
-- Additionally, we found out that it is useful to handle the case of
-  big arrays specially: when we allocate a big array (with the system
-  malloc()), we reserve a small number of bytes before.  When the array
-  grows old, we use the extra bytes as a set of bits.  Each bit
-  represents 128 entries in the array.  Whenever the write barrier is
-  called to record a reference from the Nth entry of the array to some
-  young object, we set the bit number ``(N/128)`` to 1.  This can
-  considerably speed up minor collections, because we then only have to
-  scan 128 entries of the array instead of all of them.
-
-- As usual, we need special care about weak references, and objects with
-  finalizers.  Weak references are allocated in the nursery, and if they
-  survive they move to the old stage, as usual for all objects; the
-  difference is that the reference they contain must either follow the
-  object, or be set to NULL if the object dies.  And the objects with
-  finalizers, considered rare enough, are immediately allocated old to
-  simplify the design.  In particular their ``__del__`` method can only
-  be called just after a major collection.
-
-- The objects move once only, so we can use a trick to implement id()
-  and hash().  If the object is not in the nursery, it won't move any
-  more, so its id() and hash() are the object's address, cast to an
-  integer.  If the object is in the nursery, and we ask for its id()
-  or its hash(), then we pre-reserve a location in the old stage, and
-  return the address of that location.  If the object survives the
-  next minor collection, we move it there, and so its id() and hash()
-  are preserved.  If the object dies then the pre-reserved location
-  becomes free garbage, to be collected at the next major collection.
-
-
 .. include:: _ref.txt
diff --git a/pypy/doc/project-ideas.rst b/pypy/doc/project-ideas.rst
--- a/pypy/doc/project-ideas.rst
+++ b/pypy/doc/project-ideas.rst
@@ -91,7 +91,7 @@
 Remove the GIL
 --------------
 
-This is a major task that requiers lots of thinking. However, few subprojects
+This is a major task that requires lots of thinking. However, few subprojects
 can be potentially specified, unless a better plan can be thought out:
 
 * A thread-aware garbage collector
@@ -124,6 +124,25 @@
 for our needs. It's possible that this has changed, reviving the LLVM backend
 (or writing new from scratch) for static compilation would be a good project.
 
+(On the other hand, just generating C code and using clang might be enough.
+The issue with that is the so-called "asmgcc GC root finder", which has tons
+of issues of this own.  In my opinion (arigo), it would be definitely a
+better project to try to optimize the alternative, the "shadowstack" GC root
+finder, which is nicely portable.  So far it gives a pypy that is around
+7% slower.)
+
+Embedding PyPy
+----------------------------------------
+
+Being able to embed PyPy, say with its own limited C API, would be
+useful.  But here is the most interesting variant, straight from
+EuroPython live discussion :-)  We can have a generic "libpypy.so" that
+can be used as a placeholder dynamic library, and when it gets loaded,
+it runs a .py module that installs (via ctypes) the interface it wants
+exported.  This would give us a one-size-fits-all generic .so file to be
+imported by any application that wants to load .so files :-)
+
+
 .. _`issue tracker`: http://bugs.pypy.org
 .. _`mailing list`: http://mail.python.org/mailman/listinfo/pypy-dev
 .. _`jitviewer`: http://bitbucket.org/pypy/jitviewer
diff --git a/pypy/interpreter/argument.py b/pypy/interpreter/argument.py
--- a/pypy/interpreter/argument.py
+++ b/pypy/interpreter/argument.py
@@ -90,15 +90,18 @@
     ###  Construction  ###
 
     def __init__(self, space, args_w, keywords=None, keywords_w=None,
-                 w_stararg=None, w_starstararg=None):
+                 w_stararg=None, w_starstararg=None, keyword_names_w=None):
         self.space = space
         assert isinstance(args_w, list)
         self.arguments_w = args_w
         self.keywords = keywords
         self.keywords_w = keywords_w
+        self.keyword_names_w = keyword_names_w  # matches the tail of .keywords
         if keywords is not None:
             assert keywords_w is not None
             assert len(keywords_w) == len(keywords)
+            assert (keyword_names_w is None or
+                    len(keyword_names_w) <= len(keywords))
             make_sure_not_resized(self.keywords)
             make_sure_not_resized(self.keywords_w)
 
@@ -132,7 +135,8 @@
 
     def replace_arguments(self, args_w):
         "Return a new Arguments with a args_w as positional arguments."
-        return Arguments(self.space, args_w, self.keywords, self.keywords_w)
+        return Arguments(self.space, args_w, self.keywords, self.keywords_w,
+                         keyword_names_w = self.keyword_names_w)
 
     def prepend(self, w_firstarg):
         "Return a new Arguments with a new argument inserted first."
@@ -201,15 +205,16 @@
                         space.w_TypeError,
                         space.wrap("keywords must be strings"))
                 if e.match(space, space.w_UnicodeEncodeError):
-                    raise OperationError(
-                        space.w_TypeError,
-                        space.wrap("keyword cannot be encoded to ascii"))
-                raise
-            if self.keywords and key in self.keywords:
-                raise operationerrfmt(self.space.w_TypeError,
-                                      "got multiple values "
-                                      "for keyword argument "
-                                      "'%s'", key)
+                    # Allow this to pass through
+                    key = None
+                else:
+                    raise
+            else:
+                if self.keywords and key in self.keywords:
+                    raise operationerrfmt(self.space.w_TypeError,
+                                          "got multiple values "
+                                          "for keyword argument "
+                                          "'%s'", key)
             keywords[i] = key
             keywords_w[i] = space.getitem(w_starstararg, w_key)
             i += 1
@@ -219,6 +224,7 @@
         else:
             self.keywords = self.keywords + keywords
             self.keywords_w = self.keywords_w + keywords_w
+        self.keyword_names_w = keys_w
 
     def fixedunpack(self, argcount):
         """The simplest argument parsing: get the 'argcount' arguments,
@@ -339,6 +345,10 @@
             used_keywords = [False] * num_kwds
             for i in range(num_kwds):
                 name = keywords[i]
+                # If name was not encoded as a string, it could be None. In that
+                # case, it's definitely not going to be in the signature.
+                if name is None:
+                    continue
                 j = signature.find_argname(name)
                 if j < 0:
                     continue
@@ -374,17 +384,26 @@
         if has_kwarg:
             w_kwds = self.space.newdict()
             if num_remainingkwds:
+                #
+                limit = len(keywords)
+                if self.keyword_names_w is not None:
+                    limit -= len(self.keyword_names_w)
                 for i in range(len(keywords)):
                     if not used_keywords[i]:
-                        key = keywords[i]
-                        self.space.setitem(w_kwds, self.space.wrap(key), keywords_w[i])
+                        if i < limit:
+                            w_key = self.space.wrap(keywords[i])
+                        else:
+                            w_key = self.keyword_names_w[i - limit]
+                        self.space.setitem(w_kwds, w_key, keywords_w[i])
+                #
             scope_w[co_argcount + has_vararg] = w_kwds
         elif num_remainingkwds:
             if co_argcount == 0:
                 raise ArgErrCount(avail, num_kwds,
                               co_argcount, has_vararg, has_kwarg,
                               defaults_w, missing)
-            raise ArgErrUnknownKwds(num_remainingkwds, keywords, used_keywords)
+            raise ArgErrUnknownKwds(self.space, num_remainingkwds, keywords,
+                                    used_keywords, self.keyword_names_w)
 
         if missing:
             raise ArgErrCount(avail, num_kwds,
@@ -443,9 +462,15 @@
         w_args = space.newtuple(self.arguments_w)
         w_kwds = space.newdict()
         if self.keywords is not None:
+            limit = len(self.keywords)
+            if self.keyword_names_w is not None:
+                limit -= len(self.keyword_names_w)
             for i in range(len(self.keywords)):
-                space.setitem(w_kwds, space.wrap(self.keywords[i]),
-                                      self.keywords_w[i])
+                if i < limit:
+                    w_key = space.wrap(self.keywords[i])
+                else:
+                    w_key = self.keyword_names_w[i - limit]
+                space.setitem(w_kwds, w_key, self.keywords_w[i])
         return w_args, w_kwds
 
 class ArgumentsForTranslation(Arguments):
@@ -666,14 +691,33 @@
 
 class ArgErrUnknownKwds(ArgErr):
 
-    def __init__(self, num_remainingkwds, keywords, used_keywords):
-        self.kwd_name = ''
+    def __init__(self, space, num_remainingkwds, keywords, used_keywords,
+                 keyword_names_w):
+        name = ''
         self.num_kwds = num_remainingkwds
         if num_remainingkwds == 1:
             for i in range(len(keywords)):
                 if not used_keywords[i]:
-                    self.kwd_name = keywords[i]
+                    name = keywords[i]
+                    if name is None:
+                        # We'll assume it's unicode. Encode it.
+                        # Careful, I *think* it should not be possible to
+                        # get an IndexError here but you never know.
+                        try:
+                            if keyword_names_w is None:
+                                raise IndexError
+                            # note: negative-based indexing from the end
+                            w_name = keyword_names_w[i - len(keywords)]
+                        except IndexError:
+                            name = '?'
+                        else:
+                            w_enc = space.wrap(space.sys.defaultencoding)
+                            w_err = space.wrap("replace")
+                            w_name = space.call_method(w_name, "encode", w_enc,
+                                                       w_err)
+                            name = space.str_w(w_name)
                     break
+        self.kwd_name = name
 
     def getmsg(self, fnname):
         if self.num_kwds == 1:
diff --git a/pypy/interpreter/astcompiler/codegen.py b/pypy/interpreter/astcompiler/codegen.py
--- a/pypy/interpreter/astcompiler/codegen.py
+++ b/pypy/interpreter/astcompiler/codegen.py
@@ -134,7 +134,7 @@
 
     def accept_comp_iteration(self, codegen, index):
         self.elt.walkabout(codegen)
-        codegen.emit_op_arg(ops.SET_ADD, index)
+        codegen.emit_op_arg(ops.SET_ADD, index + 1)
 
 
 class __extend__(ast.DictComp):
@@ -148,7 +148,7 @@
     def accept_comp_iteration(self, codegen, index):
         self.value.walkabout(codegen)
         self.key.walkabout(codegen)
-        codegen.emit_op_arg(ops.MAP_ADD, index)
+        codegen.emit_op_arg(ops.MAP_ADD, index + 1)
 
 
 # These are frame blocks.
diff --git a/pypy/interpreter/astcompiler/misc.py b/pypy/interpreter/astcompiler/misc.py
--- a/pypy/interpreter/astcompiler/misc.py
+++ b/pypy/interpreter/astcompiler/misc.py
@@ -92,7 +92,10 @@
         return name
     if len(name) + 2 >= MANGLE_LEN:
         return name
-    if name.endswith('__'):
+    # Don't mangle __id__ or names with dots. The only time a name with a dot
+    # can occur is when we are compiling an import statement that has a package
+    # name.
+    if name.endswith('__') or '.' in name:
         return name
     try:
         i = 0
diff --git a/pypy/interpreter/astcompiler/test/test_compiler.py b/pypy/interpreter/astcompiler/test/test_compiler.py
--- a/pypy/interpreter/astcompiler/test/test_compiler.py
+++ b/pypy/interpreter/astcompiler/test/test_compiler.py
@@ -55,7 +55,7 @@
         co_expr = compile(evalexpr, '<evalexpr>', 'eval')
         space = self.space
         pyco_expr = PyCode._from_code(space, co_expr)
-        w_res = pyco_expr.exec_host_bytecode(space, w_dict, w_dict)
+        w_res = pyco_expr.exec_host_bytecode(w_dict, w_dict)
         res = space.str_w(space.repr(w_res))
         if not isinstance(expected, float):
             assert res == repr(expected)
@@ -308,6 +308,15 @@
                "p.__name__", os.path.__name__)
         yield (self.st, 'from os import *',
                "path.__name__, sep", (os.path.__name__, os.sep))
+        yield (self.st, '''
+            class A(object):
+                def m(self):
+                    from __foo__.bar import x
+            try:
+                A().m()
+            except ImportError, e:
+                msg = str(e)
+            ''', "msg", "No module named __foo__")
 
     def test_if_stmts(self):
         yield self.st, "a = 42\nif a > 10: a += 2", "a", 44
diff --git a/pypy/interpreter/baseobjspace.py b/pypy/interpreter/baseobjspace.py
--- a/pypy/interpreter/baseobjspace.py
+++ b/pypy/interpreter/baseobjspace.py
@@ -989,10 +989,7 @@
             compiler = self.createcompiler()
             expression = compiler.compile(expression, '?', 'eval', 0,
                                          hidden_applevel=hidden_applevel)
-        if isinstance(expression, types.CodeType):
-            # XXX only used by appsupport
-            expression = PyCode._from_code(self, expression)
-        if not isinstance(expression, PyCode):
+        else:
             raise TypeError, 'space.eval(): expected a string, code or PyCode object'
         return expression.exec_code(self, w_globals, w_locals)
 
@@ -1007,9 +1004,6 @@
             compiler = self.createcompiler()
             statement = compiler.compile(statement, filename, 'exec', 0,
                                          hidden_applevel=hidden_applevel)
-        if isinstance(statement, types.CodeType):
-            # XXX only used by appsupport
-            statement = PyCode._from_code(self, statement)
         if not isinstance(statement, PyCode):
             raise TypeError, 'space.exec_(): expected a string, code or PyCode object'
         w_key = self.wrap('__builtins__')
diff --git a/pypy/interpreter/eval.py b/pypy/interpreter/eval.py
--- a/pypy/interpreter/eval.py
+++ b/pypy/interpreter/eval.py
@@ -100,12 +100,12 @@
 
     @jit.dont_look_inside
     def fast2locals(self):
-        # Copy values from self.fastlocals_w to self.w_locals
+        # Copy values from the fastlocals to self.w_locals
         if self.w_locals is None:
             self.w_locals = self.space.newdict()
         varnames = self.getcode().getvarnames()
         fastscope_w = self.getfastscope()
-        for i in range(min(len(varnames), len(fastscope_w))):
+        for i in range(min(len(varnames), self.getfastscopelength())):
             name = varnames[i]
             w_value = fastscope_w[i]
             if w_value is not None:
@@ -114,7 +114,7 @@
 
     @jit.dont_look_inside
     def locals2fast(self):
-        # Copy values from self.w_locals to self.fastlocals_w
+        # Copy values from self.w_locals to the fastlocals
         assert self.w_locals is not None
         varnames = self.getcode().getvarnames()
         numlocals = self.getfastscopelength()
diff --git a/pypy/interpreter/function.py b/pypy/interpreter/function.py
--- a/pypy/interpreter/function.py
+++ b/pypy/interpreter/function.py
@@ -98,7 +98,7 @@
                                                    self.closure)
                 for i in funccallunrolling:
                     if i < nargs:
-                        new_frame.fastlocals_w[i] = args_w[i]
+                        new_frame.locals_stack_w[i] = args_w[i]
                 return new_frame.run()
         elif nargs >= 1 and fast_natural_arity == Code.PASSTHROUGHARGS1:
             assert isinstance(code, gateway.BuiltinCodePassThroughArguments1)
@@ -158,7 +158,7 @@
                                                    self.closure)
         for i in xrange(nargs):
             w_arg = frame.peekvalue(nargs-1-i)
-            new_frame.fastlocals_w[i] = w_arg
+            new_frame.locals_stack_w[i] = w_arg
 
         return new_frame.run()
 
@@ -169,13 +169,13 @@
                                                    self.closure)
         for i in xrange(nargs):
             w_arg = frame.peekvalue(nargs-1-i)
-            new_frame.fastlocals_w[i] = w_arg
+            new_frame.locals_stack_w[i] = w_arg
 
         ndefs = len(self.defs_w)
         start = ndefs - defs_to_load
         i = nargs
         for j in xrange(start, ndefs):
-            new_frame.fastlocals_w[i] = self.defs_w[j]
+            new_frame.locals_stack_w[i] = self.defs_w[j]
             i += 1
         return new_frame.run()
 
diff --git a/pypy/interpreter/nestedscope.py b/pypy/interpreter/nestedscope.py
--- a/pypy/interpreter/nestedscope.py
+++ b/pypy/interpreter/nestedscope.py
@@ -170,7 +170,7 @@
         for i in range(len(args_to_copy)):
             argnum = args_to_copy[i]
             if argnum >= 0:
-                self.cells[i].set(self.fastlocals_w[argnum])
+                self.cells[i].set(self.locals_stack_w[argnum])
 
     def getfreevarname(self, index):
         freevarnames = self.pycode.co_cellvars + self.pycode.co_freevars
diff --git a/pypy/interpreter/pycode.py b/pypy/interpreter/pycode.py
--- a/pypy/interpreter/pycode.py
+++ b/pypy/interpreter/pycode.py
@@ -63,6 +63,7 @@
         the pypy compiler"""
         self.space = space
         eval.Code.__init__(self, name)
+        assert nlocals >= 0
         self.co_argcount = argcount
         self.co_nlocals = nlocals
         self.co_stacksize = stacksize
@@ -95,7 +96,7 @@
             if self.co_flags & CO_VARKEYWORDS:
                 argcount += 1
             # Cell vars could shadow already-set arguments.
-            # astcompiler.pyassem used to be clever about the order of
+            # The compiler used to be clever about the order of
             # the variables in both co_varnames and co_cellvars, but
             # it no longer is for the sake of simplicity.  Moreover
             # code objects loaded from CPython don't necessarily follow
@@ -202,7 +203,7 @@
         # speed hack
         fresh_frame = jit.hint(frame, access_directly=True,
                                       fresh_virtualizable=True)
-        args_matched = args.parse_into_scope(None, fresh_frame.fastlocals_w,
+        args_matched = args.parse_into_scope(None, fresh_frame.locals_stack_w,
                                              func.name,
                                              sig, func.defs_w)
         fresh_frame.init_cells()
@@ -215,7 +216,7 @@
         # speed hack
         fresh_frame = jit.hint(frame, access_directly=True,
                                       fresh_virtualizable=True)
-        args_matched = args.parse_into_scope(w_obj, fresh_frame.fastlocals_w,
+        args_matched = args.parse_into_scope(w_obj, fresh_frame.locals_stack_w,
                                              func.name,
                                              sig, func.defs_w)
         fresh_frame.init_cells()
@@ -256,7 +257,7 @@
                          tuple(self.co_freevars),
                          tuple(self.co_cellvars) )
 
-    def exec_host_bytecode(self, w_dict, w_globals, w_locals):
+    def exec_host_bytecode(self, w_globals, w_locals):
         from pypy.interpreter.pyframe import CPythonFrame
         frame = CPythonFrame(self.space, self, w_globals, None)
         frame.setdictscope(w_locals)
diff --git a/pypy/interpreter/pyframe.py b/pypy/interpreter/pyframe.py
--- a/pypy/interpreter/pyframe.py
+++ b/pypy/interpreter/pyframe.py
@@ -9,7 +9,7 @@
 from pypy.interpreter import pytraceback
 from pypy.rlib.objectmodel import we_are_translated, instantiate
 from pypy.rlib.jit import hint
-from pypy.rlib.debug import make_sure_not_resized
+from pypy.rlib.debug import make_sure_not_resized, check_nonneg
 from pypy.rlib.rarithmetic import intmask
 from pypy.rlib import jit
 from pypy.tool import stdlib_opcode
@@ -56,16 +56,18 @@
         assert isinstance(code, pycode.PyCode)
         self.pycode = code
         eval.Frame.__init__(self, space, w_globals)
-        self.valuestack_w = [None] * code.co_stacksize
-        self.valuestackdepth = 0
+        self.locals_stack_w = [None] * (code.co_nlocals + code.co_stacksize)
+        self.nlocals = code.co_nlocals
+        self.valuestackdepth = code.co_nlocals
         self.lastblock = None
+        make_sure_not_resized(self.locals_stack_w)
+        check_nonneg(self.nlocals)
+        #
         if space.config.objspace.honor__builtins__:
             self.builtin = space.builtin.pick_builtin(w_globals)
         # regular functions always have CO_OPTIMIZED and CO_NEWLOCALS.
         # class bodies only have CO_NEWLOCALS.
         self.initialize_frame_scopes(closure, code)
-        self.fastlocals_w = [None] * code.co_nlocals
-        make_sure_not_resized(self.fastlocals_w)
         self.f_lineno = code.co_firstlineno
 
     def mark_as_escaped(self):
@@ -184,14 +186,14 @@
     # stack manipulation helpers
     def pushvalue(self, w_object):
         depth = self.valuestackdepth
-        self.valuestack_w[depth] = w_object
+        self.locals_stack_w[depth] = w_object
         self.valuestackdepth = depth + 1
 
     def popvalue(self):
         depth = self.valuestackdepth - 1
-        assert depth >= 0, "pop from empty value stack"
-        w_object = self.valuestack_w[depth]
-        self.valuestack_w[depth] = None
+        assert depth >= self.nlocals, "pop from empty value stack"
+        w_object = self.locals_stack_w[depth]
+        self.locals_stack_w[depth] = None
         self.valuestackdepth = depth
         return w_object
 
@@ -217,24 +219,24 @@
     def peekvalues(self, n):
         values_w = [None] * n
         base = self.valuestackdepth - n
-        assert base >= 0
+        assert base >= self.nlocals
         while True:
             n -= 1
             if n < 0:
                 break
-            values_w[n] = self.valuestack_w[base+n]
+            values_w[n] = self.locals_stack_w[base+n]
         return values_w
 
     @jit.unroll_safe
     def dropvalues(self, n):
         n = hint(n, promote=True)
         finaldepth = self.valuestackdepth - n
-        assert finaldepth >= 0, "stack underflow in dropvalues()"        
+        assert finaldepth >= self.nlocals, "stack underflow in dropvalues()"
         while True:
             n -= 1
             if n < 0:
                 break
-            self.valuestack_w[finaldepth+n] = None
+            self.locals_stack_w[finaldepth+n] = None
         self.valuestackdepth = finaldepth
 
     @jit.unroll_safe
@@ -261,30 +263,30 @@
         # Contrast this with CPython where it's PEEK(-1).
         index_from_top = hint(index_from_top, promote=True)
         index = self.valuestackdepth + ~index_from_top
-        assert index >= 0, "peek past the bottom of the stack"
-        return self.valuestack_w[index]
+        assert index >= self.nlocals, "peek past the bottom of the stack"
+        return self.locals_stack_w[index]
 
     def settopvalue(self, w_object, index_from_top=0):
         index_from_top = hint(index_from_top, promote=True)
         index = self.valuestackdepth + ~index_from_top
-        assert index >= 0, "settop past the bottom of the stack"
-        self.valuestack_w[index] = w_object
+        assert index >= self.nlocals, "settop past the bottom of the stack"
+        self.locals_stack_w[index] = w_object
 
     @jit.unroll_safe
     def dropvaluesuntil(self, finaldepth):
         depth = self.valuestackdepth - 1
         finaldepth = hint(finaldepth, promote=True)
         while depth >= finaldepth:
-            self.valuestack_w[depth] = None
+            self.locals_stack_w[depth] = None
             depth -= 1
         self.valuestackdepth = finaldepth
 
-    def savevaluestack(self):
-        return self.valuestack_w[:self.valuestackdepth]
+    def save_locals_stack(self):
+        return self.locals_stack_w[:self.valuestackdepth]
 
-    def restorevaluestack(self, items_w):
-        assert None not in items_w
-        self.valuestack_w[:len(items_w)] = items_w
+    def restore_locals_stack(self, items_w):
+        self.locals_stack_w[:len(items_w)] = items_w
+        self.init_cells()
         self.dropvaluesuntil(len(items_w))
 
     def make_arguments(self, nargs):
@@ -314,11 +316,12 @@
         else:
             f_lineno = self.f_lineno
 
-        values_w = self.valuestack_w[0:self.valuestackdepth]
+        values_w = self.locals_stack_w[self.nlocals:self.valuestackdepth]
         w_valuestack = maker.slp_into_tuple_with_nulls(space, values_w)
         
         w_blockstack = nt([block._get_state_(space) for block in self.get_blocklist()])
-        w_fastlocals = maker.slp_into_tuple_with_nulls(space, self.fastlocals_w)
+        w_fastlocals = maker.slp_into_tuple_with_nulls(
+            space, self.locals_stack_w[:self.nlocals])
         if self.last_exception is None:
             w_exc_value = space.w_None
             w_tb = space.w_None
@@ -399,7 +402,8 @@
         new_frame.last_instr = space.int_w(w_last_instr)
         new_frame.frame_finished_execution = space.is_true(w_finished)
         new_frame.f_lineno = space.int_w(w_f_lineno)
-        new_frame.fastlocals_w = maker.slp_from_tuple_with_nulls(space, w_fastlocals)
+        fastlocals_w = maker.slp_from_tuple_with_nulls(space, w_fastlocals)
+        new_frame.locals_stack_w[:len(fastlocals_w)] = fastlocals_w
 
         if space.is_w(w_f_trace, space.w_None):
             new_frame.w_f_trace = None
@@ -423,28 +427,28 @@
     @jit.dont_look_inside
     def getfastscope(self):
         "Get the fast locals as a list."
-        return self.fastlocals_w
+        return self.locals_stack_w
 
     @jit.dont_look_inside
     def setfastscope(self, scope_w):
         """Initialize the fast locals from a list of values,
         where the order is according to self.pycode.signature()."""
         scope_len = len(scope_w)
-        if scope_len > len(self.fastlocals_w):
+        if scope_len > self.nlocals:
             raise ValueError, "new fastscope is longer than the allocated area"
-        # don't assign directly to 'fastlocals_w[:scope_len]' to be
+        # don't assign directly to 'locals_stack_w[:scope_len]' to be
         # virtualizable-friendly
         for i in range(scope_len):
-            self.fastlocals_w[i] = scope_w[i]
+            self.locals_stack_w[i] = scope_w[i]
         self.init_cells()
 
     def init_cells(self):
-        """Initialize cellvars from self.fastlocals_w
+        """Initialize cellvars from self.locals_stack_w.
         This is overridden in nestedscope.py"""
         pass
 
     def getfastscopelength(self):
-        return self.pycode.co_nlocals
+        return self.nlocals
 
     def getclosure(self):
         return None
diff --git a/pypy/interpreter/pyopcode.py b/pypy/interpreter/pyopcode.py
--- a/pypy/interpreter/pyopcode.py
+++ b/pypy/interpreter/pyopcode.py
@@ -324,7 +324,7 @@
 
     def LOAD_FAST(self, varindex, next_instr):
         # access a local variable directly
-        w_value = self.fastlocals_w[varindex]
+        w_value = self.locals_stack_w[varindex]
         if w_value is None:
             self._load_fast_failed(varindex)
         self.pushvalue(w_value)
@@ -343,7 +343,7 @@
     def STORE_FAST(self, varindex, next_instr):
         w_newvalue = self.popvalue()
         assert w_newvalue is not None
-        self.fastlocals_w[varindex] = w_newvalue
+        self.locals_stack_w[varindex] = w_newvalue
 
     def POP_TOP(self, oparg, next_instr):
         self.popvalue()
@@ -696,12 +696,12 @@
     LOAD_GLOBAL._always_inline_ = True
 
     def DELETE_FAST(self, varindex, next_instr):
-        if self.fastlocals_w[varindex] is None:
+        if self.locals_stack_w[varindex] is None:
             varname = self.getlocalvarname(varindex)
             message = "local variable '%s' referenced before assignment"
             raise operationerrfmt(self.space.w_UnboundLocalError, message,
                                   varname)
-        self.fastlocals_w[varindex] = None
+        self.locals_stack_w[varindex] = None
 
     def BUILD_TUPLE(self, itemcount, next_instr):
         items = self.popvalues(itemcount)
@@ -1048,13 +1048,13 @@
 
     def SET_ADD(self, oparg, next_instr):
         w_value = self.popvalue()
-        w_set = self.peekvalue(oparg)
+        w_set = self.peekvalue(oparg - 1)
         self.space.call_method(w_set, 'add', w_value)
 
     def MAP_ADD(self, oparg, next_instr):
         w_key = self.popvalue()
         w_value = self.popvalue()
-        w_dict = self.peekvalue(oparg)
+        w_dict = self.peekvalue(oparg - 1)
         self.space.setitem(w_dict, w_key, w_value)
 
     def SET_LINENO(self, lineno, next_instr):
@@ -1091,12 +1091,10 @@
 
     @jit.unroll_safe
     def BUILD_SET(self, itemcount, next_instr):
-        w_set = self.space.call_function(self.space.w_set)
-        if itemcount:
-            w_add = self.space.getattr(w_set, self.space.wrap("add"))
-            for i in range(itemcount):
-                w_item = self.popvalue()
-                self.space.call_function(w_add, w_item)
+        w_set = self.space.newset()
+        for i in range(itemcount):
+            w_item = self.popvalue()
+            self.space.call_method(w_set, 'add', w_item)
         self.pushvalue(w_set)
 
     def STORE_MAP(self, oparg, next_instr):
diff --git a/pypy/interpreter/test/test_argument.py b/pypy/interpreter/test/test_argument.py
--- a/pypy/interpreter/test/test_argument.py
+++ b/pypy/interpreter/test/test_argument.py
@@ -1,3 +1,4 @@
+# -*- coding: utf-8 -*-
 import py
 from pypy.interpreter.argument import (Arguments, ArgumentsForTranslation,
     ArgErr, ArgErrUnknownKwds, ArgErrMultipleValues, ArgErrCount, rawshape,
@@ -126,6 +127,7 @@
     w_AttributeError = AttributeError
     w_UnicodeEncodeError = UnicodeEncodeError
     w_dict = dict
+    w_str = str
 
 class TestArgumentsNormal(object):
 
@@ -485,26 +487,6 @@
         args._match_signature(None, l, Signature(['abc']))
         assert len(l) == 1
         assert l[0] == space.wrap(5)
-        #
-        def str_w(w):
-            try:
-                return str(w)
-            except UnicodeEncodeError:
-                raise OperationError(space.w_UnicodeEncodeError,
-                                     space.wrap("oups"))
-        space.str_w = str_w
-        w_starstar = space.wrap({u'\u1234': 5})
-        err = py.test.raises(OperationError, Arguments,
-                             space, [], w_starstararg=w_starstar)
-        # Check that we get a TypeError.  On CPython it is because of
-        # "no argument called '?'".  On PyPy we get a TypeError too, but
-        # earlier: "keyword cannot be encoded to ascii".  The
-        # difference, besides the error message, is only apparent if the
-        # receiver also takes a **arg.  Then CPython passes the
-        # non-ascii unicode unmodified, whereas PyPy complains.  We will
-        # not care until someone has a use case for that.
-        assert not err.value.match(space, space.w_UnicodeEncodeError)
-        assert     err.value.match(space, space.w_TypeError)
 
 class TestErrorHandling(object):
     def test_missing_args(self):
@@ -559,13 +541,26 @@
             assert 0, "did not raise"
 
     def test_unknown_keywords(self):
-        err = ArgErrUnknownKwds(1, ['a', 'b'], [True, False])
+        space = DummySpace()
+        err = ArgErrUnknownKwds(space, 1, ['a', 'b'], [True, False], None)
         s = err.getmsg('foo')
         assert s == "foo() got an unexpected keyword argument 'b'"
-        err = ArgErrUnknownKwds(2, ['a', 'b', 'c'], [True, False, False])
+        err = ArgErrUnknownKwds(space, 2, ['a', 'b', 'c'],
+                                [True, False, False], None)
         s = err.getmsg('foo')
         assert s == "foo() got 2 unexpected keyword arguments"
 
+    def test_unknown_unicode_keyword(self):
+        class DummySpaceUnicode(DummySpace):
+            class sys:
+                defaultencoding = 'utf-8'
+        space = DummySpaceUnicode()
+        err = ArgErrUnknownKwds(space, 1, ['a', None, 'b', 'c'],
+                                [True, False, True, True],
+                                [unichr(0x1234), u'b', u'c'])
+        s = err.getmsg('foo')
+        assert s == "foo() got an unexpected keyword argument '\xe1\x88\xb4'"
+
     def test_multiple_values(self):
         err = ArgErrMultipleValues('bla')
         s = err.getmsg('foo')
@@ -592,6 +587,14 @@
         exc = raises(TypeError, (lambda a, b, **kw: 0), a=1)
         assert exc.value.message == "<lambda>() takes exactly 2 non-keyword arguments (0 given)"
 
+    def test_unicode_keywords(self):
+        def f(**kwargs):
+            assert kwargs[u"&#32654;"] == 42
+        f(**{u"&#32654;" : 42})
+        def f(x): pass
+        e = raises(TypeError, "f(**{u'&#252;' : 19})")
+        assert "?" in str(e.value)
+
 def make_arguments_for_translation(space, args_w, keywords_w={},
                                    w_stararg=None, w_starstararg=None):
     return ArgumentsForTranslation(space, args_w, keywords_w.keys(),
diff --git a/pypy/interpreter/test/test_eval.py b/pypy/interpreter/test/test_eval.py
--- a/pypy/interpreter/test/test_eval.py
+++ b/pypy/interpreter/test/test_eval.py
@@ -15,16 +15,16 @@
                 self.code = code
                 Frame.__init__(self, space)
                 self.numlocals = numlocals
-                self.fastlocals_w = [None] * self.numlocals
+                self._fastlocals_w = [None] * self.numlocals
 
             def getcode(self):
                 return self.code
 
             def setfastscope(self, scope_w):
-                self.fastlocals_w = scope_w
+                self._fastlocals_w = scope_w
 
             def getfastscope(self):
-                return self.fastlocals_w
+                return self._fastlocals_w
 
             def getfastscopelength(self):
                 return self.numlocals
@@ -38,11 +38,11 @@
         self.f.fast2locals()
         assert space.eq_w(self.f.w_locals, self.space.wrap({}))
         
-        self.f.fastlocals_w[0] = w(5)
+        self.f._fastlocals_w[0] = w(5)
         self.f.fast2locals()
         assert space.eq_w(self.f.w_locals, self.space.wrap({'x': 5}))
 
-        self.f.fastlocals_w[2] = w(7)
+        self.f._fastlocals_w[2] = w(7)
         self.f.fast2locals()
         assert space.eq_w(self.f.w_locals, self.space.wrap({'x': 5, 'args': 7}))
 
@@ -57,13 +57,13 @@
         w = self.space.wrap
         self.f.w_locals = self.space.wrap({})
         self.f.locals2fast()
-        self.sameList(self.f.fastlocals_w, [None]*5)
+        self.sameList(self.f._fastlocals_w, [None]*5)
 
         self.f.w_locals = self.space.wrap({'x': 5})
         self.f.locals2fast()
-        self.sameList(self.f.fastlocals_w, [w(5)] + [None]*4)
+        self.sameList(self.f._fastlocals_w, [w(5)] + [None]*4)
 
         self.f.w_locals = self.space.wrap({'x':5, 'args':7})
         self.f.locals2fast()
-        self.sameList(self.f.fastlocals_w, [w(5), None, w(7),
-                                            None, None])
+        self.sameList(self.f._fastlocals_w, [w(5), None, w(7),
+                                             None, None])
diff --git a/pypy/interpreter/test/test_typedef.py b/pypy/interpreter/test/test_typedef.py
--- a/pypy/interpreter/test/test_typedef.py
+++ b/pypy/interpreter/test/test_typedef.py
@@ -203,3 +203,26 @@
         lst = seen[:]
         assert lst == [5, 10, 2]
         raises(OSError, os.lseek, fd, 7, 0)
+
+    def test_method_attrs(self):
+        import sys
+        class A(object):
+            def m(self):
+                "aaa"
+            m.x = 3
+
+        bm = A().m
+        assert bm.__func__ is bm.im_func
+        assert bm.__self__ is bm.im_self
+        assert bm.im_class is A
+        if '__pypy__' in sys.builtin_module_names:
+            assert bm.__objclass__ is A
+        assert bm.__doc__ == "aaa"
+        assert bm.x == 3
+        raises(AttributeError, setattr, bm, 'x', 15)
+        l = []
+        assert l.append.__self__ is l
+        if '__pypy__' in sys.builtin_module_names:
+            assert l.append.__objclass__ is list
+        assert l.__add__.__self__ is l
+        assert l.__add__.__objclass__ is list
diff --git a/pypy/interpreter/typedef.py b/pypy/interpreter/typedef.py
--- a/pypy/interpreter/typedef.py
+++ b/pypy/interpreter/typedef.py
@@ -761,13 +761,17 @@
     )
 Function.typedef.acceptable_as_base_class = False
 
-Method.typedef = TypeDef("method",
+Method.typedef = TypeDef(
+    "method",
     __new__ = interp2app(Method.descr_method__new__.im_func),
     __call__ = interp2app(Method.descr_method_call),
     __get__ = interp2app(Method.descr_method_get),
     im_func  = interp_attrproperty_w('w_function', cls=Method),
+    __func__ = interp_attrproperty_w('w_function', cls=Method),
     im_self  = interp_attrproperty_w('w_instance', cls=Method),
+    __self__ = interp_attrproperty_w('w_instance', cls=Method),
     im_class = interp_attrproperty_w('w_class', cls=Method),
+    __objclass__ = interp_attrproperty_w('w_class', cls=Method),
     __getattribute__ = interp2app(Method.descr_method_getattribute),
     __eq__ = interp2app(Method.descr_method_eq),
     __ne__ = descr_generic_ne,
diff --git a/pypy/jit/backend/llgraph/llimpl.py b/pypy/jit/backend/llgraph/llimpl.py
--- a/pypy/jit/backend/llgraph/llimpl.py
+++ b/pypy/jit/backend/llgraph/llimpl.py
@@ -136,6 +136,7 @@
     'call'            : (('ref', 'varargs'), 'intorptr'),
     'call_assembler'  : (('varargs',), 'intorptr'),
     'cond_call_gc_wb' : (('ptr', 'ptr'), None),
+    'cond_call_gc_wb_array': (('ptr', 'int', 'ptr'), None),
     'oosend'          : (('varargs',), 'intorptr'),
     'oosend_pure'     : (('varargs',), 'intorptr'),
     'guard_true'      : (('bool',), None),
@@ -857,6 +858,9 @@
     def op_cond_call_gc_wb(self, descr, a, b):
         py.test.skip("cond_call_gc_wb not supported")
 
+    def op_cond_call_gc_wb_array(self, descr, a, b, c):
+        py.test.skip("cond_call_gc_wb_array not supported")
+
     def op_oosend(self, descr, obj, *args):
         raise NotImplementedError("oosend for lltype backend??")
 
diff --git a/pypy/jit/backend/llsupport/descr.py b/pypy/jit/backend/llsupport/descr.py
--- a/pypy/jit/backend/llsupport/descr.py
+++ b/pypy/jit/backend/llsupport/descr.py
@@ -1,5 +1,6 @@
 import py
 from pypy.rpython.lltypesystem import lltype, rffi, llmemory, rclass
+from pypy.rpython.lltypesystem.lloperation import llop
 from pypy.jit.backend.llsupport import symbolic, support
 from pypy.jit.metainterp.history import AbstractDescr, getkind, BoxInt, BoxPtr
 from pypy.jit.metainterp.history import BasicFailDescr, LoopToken, BoxFloat
@@ -45,6 +46,8 @@
     size = 0      # help translation
     is_immutable = False
 
+    tid = llop.combine_ushort(lltype.Signed, 0, 0)
+
     def __init__(self, size, count_fields_if_immut=-1):
         self.size = size
         self.count_fields_if_immut = count_fields_if_immut
@@ -149,6 +152,7 @@
 
 class BaseArrayDescr(AbstractDescr):
     _clsname = ''
+    tid = llop.combine_ushort(lltype.Signed, 0, 0)
 
     def get_base_size(self, translate_support_code):
         basesize, _, _ = symbolic.get_array_token(_A, translate_support_code)
@@ -263,6 +267,9 @@
 
     def __repr__(self):
         res = '%s(%s)' % (self.__class__.__name__, self.arg_classes)
+        extraeffect = getattr(self.extrainfo, 'extraeffect', None)
+        if extraeffect is not None:
+            res += ' EF=%r' % extraeffect
         oopspecindex = getattr(self.extrainfo, 'oopspecindex', 0)
         if oopspecindex:
             from pypy.jit.codewriter.effectinfo import EffectInfo
diff --git a/pypy/jit/backend/llsupport/gc.py b/pypy/jit/backend/llsupport/gc.py
--- a/pypy/jit/backend/llsupport/gc.py
+++ b/pypy/jit/backend/llsupport/gc.py
@@ -476,6 +476,7 @@
         return cpu.cast_adr_to_int(funcaddr)
 
     def get_write_barrier_from_array_fn(self, cpu):
+        # returns a function with arguments [array, index, newvalue]
         llop1 = self.llop1
         funcptr = llop1.get_write_barrier_from_array_failing_case(
             self.WB_ARRAY_FUNCPTR)
@@ -552,7 +553,7 @@
         self.WB_FUNCPTR = lltype.Ptr(lltype.FuncType(
             [llmemory.Address, llmemory.Address], lltype.Void))
         self.WB_ARRAY_FUNCPTR = lltype.Ptr(lltype.FuncType(
-            [llmemory.Address, lltype.Signed], lltype.Void))
+            [llmemory.Address, lltype.Signed, llmemory.Address], lltype.Void))
         self.write_barrier_descr = WriteBarrierDescr(self)
         #
         def malloc_array(itemsize, tid, num_elem):
@@ -763,10 +764,8 @@
             newops.append(op)
         return newops
 
-    def _gen_write_barrier(self, newops, v_base, v_value_or_index):
-        # NB. the 2nd argument of COND_CALL_GC_WB is either a pointer
-        # (regular case), or an index (case of write_barrier_from_array)
-        args = [v_base, v_value_or_index]
+    def _gen_write_barrier(self, newops, v_base, v_value):
+        args = [v_base, v_value]
         newops.append(ResOperation(rop.COND_CALL_GC_WB, args, None,
                                    descr=self.write_barrier_descr))
 
@@ -780,7 +779,10 @@
             length = known_lengths.get(v_base, LARGE)
             if length >= LARGE:
                 # unknown or too big: produce a write_barrier_from_array
-                self._gen_write_barrier(newops, v_base, v_index)
+                args = [v_base, v_index, v_value]
+                newops.append(ResOperation(rop.COND_CALL_GC_WB_ARRAY, args,
+                                           None,
+                                           descr=self.write_barrier_descr))
                 return
         # fall-back case: produce a write_barrier
         self._gen_write_barrier(newops, v_base, v_value)
diff --git a/pypy/jit/backend/llsupport/test/test_gc.py b/pypy/jit/backend/llsupport/test/test_gc.py
--- a/pypy/jit/backend/llsupport/test/test_gc.py
+++ b/pypy/jit/backend/llsupport/test/test_gc.py
@@ -9,7 +9,7 @@
 from pypy.jit.metainterp.resoperation import get_deep_immutable_oplist
 from pypy.jit.tool.oparser import parse
 from pypy.rpython.lltypesystem.rclass import OBJECT, OBJECT_VTABLE
-from pypy.jit.metainterp.test.test_optimizeopt import equaloplists
+from pypy.jit.metainterp.optimizeopt.util import equaloplists
 
 def test_boehm():
     gc_ll_descr = GcLLDescr_boehm(None, None, None)
@@ -553,12 +553,15 @@
                 del operations[:2]
             assert len(operations) == 2
             #
-            assert operations[0].getopnum() == rop.COND_CALL_GC_WB
-            assert operations[0].getarg(0) == v_base
             if isinstance(v_new_length, ConstInt) and v_new_length.value < 130:
+                assert operations[0].getopnum() == rop.COND_CALL_GC_WB
+                assert operations[0].getarg(0) == v_base
                 assert operations[0].getarg(1) == v_value
             else:
+                assert operations[0].getopnum() == rop.COND_CALL_GC_WB_ARRAY
+                assert operations[0].getarg(0) == v_base
                 assert operations[0].getarg(1) == v_index
+                assert operations[0].getarg(2) == v_value
             assert operations[0].result is None
             #
             assert operations[1].getopnum() == rop.SETARRAYITEM_RAW
diff --git a/pypy/jit/backend/test/runner_test.py b/pypy/jit/backend/test/runner_test.py
--- a/pypy/jit/backend/test/runner_test.py
+++ b/pypy/jit/backend/test/runner_test.py
@@ -1694,12 +1694,13 @@
                 assert record == []
 
     def test_cond_call_gc_wb_array(self):
-        def func_void(a, b):
-            record.append((a, b))
+        def func_void(a, b, c):
+            record.append((a, b, c))
         record = []
         #
         S = lltype.GcStruct('S', ('tid', lltype.Signed))
-        FUNC = self.FuncType([lltype.Ptr(S), lltype.Signed], lltype.Void)
+        FUNC = self.FuncType([lltype.Ptr(S), lltype.Signed, lltype.Ptr(S)],
+                             lltype.Void)
         func_ptr = llhelper(lltype.Ptr(FUNC), func_void)
         funcbox = self.get_funcbox(self.cpu, func_ptr)
         class WriteBarrierDescr(AbstractDescr):
@@ -1719,11 +1720,11 @@
             s.tid = value
             sgcref = lltype.cast_opaque_ptr(llmemory.GCREF, s)
             del record[:]
-            self.execute_operation(rop.COND_CALL_GC_WB,
-                                   [BoxPtr(sgcref), ConstInt(123)],
-                                   'void', descr=WriteBarrierDescr())
+            self.execute_operation(rop.COND_CALL_GC_WB_ARRAY,
+                       [BoxPtr(sgcref), ConstInt(123), BoxPtr(sgcref)],
+                       'void', descr=WriteBarrierDescr())
             if cond:
-                assert record == [(s, 123)]
+                assert record == [(s, 123, s)]
             else:
                 assert record == []
 
diff --git a/pypy/jit/backend/x86/assembler.py b/pypy/jit/backend/x86/assembler.py
--- a/pypy/jit/backend/x86/assembler.py
+++ b/pypy/jit/backend/x86/assembler.py
@@ -703,22 +703,28 @@
         # we need to put two words into the shadowstack: the MARKER
         # and the address of the frame (ebp, actually)
         rst = gcrootmap.get_root_stack_top_addr()
-        assert rx86.fits_in_32bits(rst)
-        if IS_X86_64:
-            # cannot use rdx here, it's used to pass arguments!
-            tmp = X86_64_SCRATCH_REG
+        if rx86.fits_in_32bits(rst):
+            self.mc.MOV_rj(eax.value, rst)            # MOV eax, [rootstacktop]
         else:
-            tmp = edx
-        self.mc.MOV_rj(eax.value, rst)                # MOV eax, [rootstacktop]
-        self.mc.LEA_rm(tmp.value, (eax.value, 2*WORD))  # LEA edx, [eax+2*WORD]
+            self.mc.MOV_ri(r13.value, rst)            # MOV r13, rootstacktop
+            self.mc.MOV_rm(eax.value, (r13.value, 0)) # MOV eax, [r13]
+        #
+        self.mc.LEA_rm(ebx.value, (eax.value, 2*WORD))  # LEA ebx, [eax+2*WORD]
         self.mc.MOV_mi((eax.value, 0), gcrootmap.MARKER)    # MOV [eax], MARKER
         self.mc.MOV_mr((eax.value, WORD), ebp.value)      # MOV [eax+WORD], ebp
-        self.mc.MOV_jr(rst, tmp.value)                # MOV [rootstacktop], edx
+        #
+        if rx86.fits_in_32bits(rst):
+            self.mc.MOV_jr(rst, ebx.value)            # MOV [rootstacktop], ebx
+        else:
+            self.mc.MOV_mr((r13.value, 0), ebx.value) # MOV [r13], ebx
 
     def _call_footer_shadowstack(self, gcrootmap):
         rst = gcrootmap.get_root_stack_top_addr()
-        assert rx86.fits_in_32bits(rst)
-        self.mc.SUB_ji8(rst, 2*WORD)       # SUB [rootstacktop], 2*WORD
+        if rx86.fits_in_32bits(rst):
+            self.mc.SUB_ji8(rst, 2*WORD)       # SUB [rootstacktop], 2*WORD
+        else:
+            self.mc.MOV_ri(ebx.value, rst)           # MOV ebx, rootstacktop
+            self.mc.SUB_mi8((ebx.value, 0), 2*WORD)  # SUB [ebx], 2*WORD
 
     def _assemble_bootstrap_direct_call(self, arglocs, jmppos, stackdepth):
         if IS_X86_64:
@@ -889,7 +895,7 @@
 
     def regalloc_push(self, loc):
         if isinstance(loc, RegLoc) and loc.is_xmm:
-            self.mc.SUB_ri(esp.value, 2*WORD)
+            self.mc.SUB_ri(esp.value, 8)   # = size of doubles
             self.mc.MOVSD_sx(0, loc.value)
         elif WORD == 4 and isinstance(loc, StackLoc) and loc.width == 8:
             # XXX evil trick
@@ -901,7 +907,7 @@
     def regalloc_pop(self, loc):
         if isinstance(loc, RegLoc) and loc.is_xmm:
             self.mc.MOVSD_xs(loc.value, 0)
-            self.mc.ADD_ri(esp.value, 2*WORD)
+            self.mc.ADD_ri(esp.value, 8)   # = size of doubles
         elif WORD == 4 and isinstance(loc, StackLoc) and loc.width == 8:
             # XXX evil trick
             self.mc.POP_b(get_ebp_ofs(loc.position + 1))
@@ -2223,15 +2229,26 @@
     def genop_discard_cond_call_gc_wb(self, op, arglocs):
         # Write code equivalent to write_barrier() in the GC: it checks
         # a flag in the object at arglocs[0], and if set, it calls the
-        # function remember_young_pointer() from the GC.  The two arguments
-        # to the call are in arglocs[:2].  The rest, arglocs[2:], contains
+        # function remember_young_pointer() from the GC.  The arguments
+        # to the call are in arglocs[:N].  The rest, arglocs[N:], contains
         # registers that need to be saved and restored across the call.
-        # If op.getarg(1) is a int, it is an array index and we must call
-        # instead remember_young_pointer_from_array().
+        # N is either 2 (regular write barrier) or 3 (array write barrier).
         descr = op.getdescr()
         if we_are_translated():
             cls = self.cpu.gc_ll_descr.has_write_barrier_class()
             assert cls is not None and isinstance(descr, cls)
+        #
+        opnum = op.getopnum()
+        if opnum == rop.COND_CALL_GC_WB:
+            N = 2
+            func = descr.get_write_barrier_fn(self.cpu)
+        elif opnum == rop.COND_CALL_GC_WB_ARRAY:
+            N = 3
+            func = descr.get_write_barrier_from_array_fn(self.cpu)
+            assert func != 0
+        else:
+            raise AssertionError(opnum)
+        #
         loc_base = arglocs[0]
         self.mc.TEST8(addr_add_const(loc_base, descr.jit_wb_if_flag_byteofs),
                       imm(descr.jit_wb_if_flag_singlebyte))
@@ -2242,29 +2259,27 @@
         if IS_X86_32:
             limit = -1      # push all arglocs on the stack
         elif IS_X86_64:
-            limit = 1       # push only arglocs[2:] on the stack
+            limit = N - 1   # push only arglocs[N:] on the stack
         for i in range(len(arglocs)-1, limit, -1):
             loc = arglocs[i]
             if isinstance(loc, RegLoc):
                 self.mc.PUSH_r(loc.value)
             else:
-                assert not IS_X86_64 # there should only be regs in arglocs[2:]
+                assert not IS_X86_64 # there should only be regs in arglocs[N:]
                 self.mc.PUSH_i32(loc.getint())
         if IS_X86_64:
             # We clobber these registers to pass the arguments, but that's
             # okay, because consider_cond_call_gc_wb makes sure that any
             # caller-save registers with values in them are present in
-            # arglocs[2:] too, so they are saved on the stack above and
+            # arglocs[N:] too, so they are saved on the stack above and
             # restored below.
-            remap_frame_layout(self, arglocs[:2], [edi, esi],
+            if N == 2:
+                callargs = [edi, esi]
+            else:
+                callargs = [edi, esi, edx]
+            remap_frame_layout(self, arglocs[:N], callargs,
                                X86_64_SCRATCH_REG)
-
-        if op.getarg(1).type == INT:
-            func = descr.get_write_barrier_from_array_fn(self.cpu)
-            assert func != 0
-        else:
-            func = descr.get_write_barrier_fn(self.cpu)
-
+        #
         # misaligned stack in the call, but it's ok because the write barrier
         # is not going to call anything more.  Also, this assumes that the
         # write barrier does not touch the xmm registers.  (Slightly delicate
@@ -2273,8 +2288,8 @@
         # be done properly)
         self.mc.CALL(imm(func))
         if IS_X86_32:
-            self.mc.ADD_ri(esp.value, 2*WORD)
-        for i in range(2, len(arglocs)):
+            self.mc.ADD_ri(esp.value, N*WORD)
+        for i in range(N, len(arglocs)):
             loc = arglocs[i]
             assert isinstance(loc, RegLoc)
             self.mc.POP_r(loc.value)
@@ -2283,6 +2298,8 @@
         assert 0 < offset <= 127
         self.mc.overwrite(jz_location-1, chr(offset))
 
+    genop_discard_cond_call_gc_wb_array = genop_discard_cond_call_gc_wb
+
     def genop_force_token(self, op, arglocs, resloc):
         # RegAlloc.consider_force_token ensures this:
         assert isinstance(resloc, RegLoc)
diff --git a/pypy/jit/backend/x86/regalloc.py b/pypy/jit/backend/x86/regalloc.py
--- a/pypy/jit/backend/x86/regalloc.py
+++ b/pypy/jit/backend/x86/regalloc.py
@@ -884,12 +884,12 @@
     def consider_cond_call_gc_wb(self, op):
         assert op.result is None
         args = op.getarglist()
-        loc_newvalue_or_index= self.rm.make_sure_var_in_reg(op.getarg(1), args)
-        # ^^^ we force loc_newvalue_or_index in a reg (unless it's a Const),
-        # because it will be needed anyway by the following setfield_gc.
-        # It avoids loading it twice from the memory.
-        loc_base = self.rm.make_sure_var_in_reg(op.getarg(0), args)
-        arglocs = [loc_base, loc_newvalue_or_index]
+        N = len(args)
+        # we force all arguments in a reg (unless they are Consts),
+        # because it will be needed anyway by the following setfield_gc
+        # or setarrayitem_gc. It avoids loading it twice from the memory.
+        arglocs = [self.rm.make_sure_var_in_reg(op.getarg(i), args)
+                   for i in range(N)]
         # add eax, ecx and edx as extra "arguments" to ensure they are
         # saved and restored.  Fish in self.rm to know which of these
         # registers really need to be saved (a bit of a hack).  Moreover,
@@ -903,6 +903,8 @@
         self.PerformDiscard(op, arglocs)
         self.rm.possibly_free_vars_for_op(op)
 
+    consider_cond_call_gc_wb_array = consider_cond_call_gc_wb
+
     def fastpath_malloc_fixedsize(self, op, descr):
         assert isinstance(descr, BaseSizeDescr)
         self._do_fastpath_malloc(op, descr.size, descr.tid)
diff --git a/pypy/jit/backend/x86/regloc.py b/pypy/jit/backend/x86/regloc.py
--- a/pypy/jit/backend/x86/regloc.py
+++ b/pypy/jit/backend/x86/regloc.py
@@ -318,7 +318,9 @@
             # must be careful not to combine it with location types that
             # might need to use the scratch register themselves.
             if loc2 is X86_64_SCRATCH_REG:
-                assert code1 != 'j'
+                if code1 == 'j':
+                    assert (name.startswith("MOV") and
+                            rx86.fits_in_32bits(loc1.value_j()))
             if loc1 is X86_64_SCRATCH_REG and not name.startswith("MOV"):
                 assert code2 not in ('j', 'i')
 
diff --git a/pypy/jit/backend/x86/rx86.py b/pypy/jit/backend/x86/rx86.py
--- a/pypy/jit/backend/x86/rx86.py
+++ b/pypy/jit/backend/x86/rx86.py
@@ -283,7 +283,7 @@
 # with immediate(argnum)).
 
 def encode_abs(mc, _1, _2, orbyte):
-    # expands to either '\x05' on 32-bit, or '\x04\x25' or 64-bit
+    # expands to either '\x05' on 32-bit, or '\x04\x25' on 64-bit
     if mc.WORD == 8:
         mc.writechar(chr(0x04 | orbyte))
         mc.writechar(chr(0x25))
@@ -370,6 +370,8 @@
     INSN_rj = insn(rex_w, chr(base+3), register(1,8), abs_, immediate(2))
     INSN_ji8 = insn(rex_w, '\x83', orbyte(base), abs_, immediate(1),
                     immediate(2,'b'))
+    INSN_mi8 = insn(rex_w, '\x83', orbyte(base), mem_reg_plus_const(1),
+                    immediate(2,'b'))
     INSN_bi8 = insn(rex_w, '\x83', orbyte(base), stack_bp(1), immediate(2,'b'))
     INSN_bi32= insn(rex_w, '\x81', orbyte(base), stack_bp(1), immediate(2))
 
@@ -388,7 +390,7 @@
     INSN_bi._always_inline_ = True      # try to constant-fold single_byte()
 
     return (INSN_ri, INSN_rr, INSN_rb, INSN_bi, INSN_br, INSN_rm, INSN_rj,
-            INSN_ji8)
+            INSN_ji8, INSN_mi8)
 
 def select_8_or_32_bit_immed(insn_8, insn_32):
     def INSN(*args):
@@ -467,13 +469,13 @@
 
     # ------------------------------ Arithmetic ------------------------------
 
-    ADD_ri, ADD_rr, ADD_rb, _, _, ADD_rm, ADD_rj, _ = common_modes(0)
-    OR_ri,  OR_rr,  OR_rb,  _, _, OR_rm,  OR_rj,  _ = common_modes(1)
-    AND_ri, AND_rr, AND_rb, _, _, AND_rm, AND_rj, _ = common_modes(4)
-    SUB_ri, SUB_rr, SUB_rb, _, _, SUB_rm, SUB_rj, SUB_ji8 = common_modes(5)
-    SBB_ri, SBB_rr, SBB_rb, _, _, SBB_rm, SBB_rj, _ = common_modes(3)
-    XOR_ri, XOR_rr, XOR_rb, _, _, XOR_rm, XOR_rj, _ = common_modes(6)
-    CMP_ri, CMP_rr, CMP_rb, CMP_bi, CMP_br, CMP_rm, CMP_rj, _ = common_modes(7)
+    ADD_ri,ADD_rr,ADD_rb,_,_,ADD_rm,ADD_rj,_,_ = common_modes(0)
+    OR_ri, OR_rr, OR_rb, _,_,OR_rm, OR_rj, _,_ = common_modes(1)
+    AND_ri,AND_rr,AND_rb,_,_,AND_rm,AND_rj,_,_ = common_modes(4)
+    SUB_ri,SUB_rr,SUB_rb,_,_,SUB_rm,SUB_rj,SUB_ji8,SUB_mi8 = common_modes(5)
+    SBB_ri,SBB_rr,SBB_rb,_,_,SBB_rm,SBB_rj,_,_ = common_modes(3)
+    XOR_ri,XOR_rr,XOR_rb,_,_,XOR_rm,XOR_rj,_,_ = common_modes(6)
+    CMP_ri,CMP_rr,CMP_rb,CMP_bi,CMP_br,CMP_rm,CMP_rj,_,_ = common_modes(7)
 
     CMP_mi8 = insn(rex_w, '\x83', orbyte(7<<3), mem_reg_plus_const(1), immediate(2, 'b'))
     CMP_mi32 = insn(rex_w, '\x81', orbyte(7<<3), mem_reg_plus_const(1), immediate(2))
diff --git a/pypy/jit/backend/x86/test/test_assembler.py b/pypy/jit/backend/x86/test/test_assembler.py
--- a/pypy/jit/backend/x86/test/test_assembler.py
+++ b/pypy/jit/backend/x86/test/test_assembler.py
@@ -1,13 +1,15 @@
 from pypy.jit.backend.x86.regloc import *
 from pypy.jit.backend.x86.assembler import Assembler386
 from pypy.jit.backend.x86.regalloc import X86FrameManager, get_ebp_ofs
-from pypy.jit.metainterp.history import BoxInt, BoxPtr, BoxFloat, INT, REF, FLOAT
+from pypy.jit.metainterp.history import BoxInt, BoxPtr, BoxFloat, ConstFloat
+from pypy.jit.metainterp.history import INT, REF, FLOAT
 from pypy.rlib.rarithmetic import intmask
 from pypy.rpython.lltypesystem import lltype, llmemory, rffi
 from pypy.jit.backend.x86.arch import WORD, IS_X86_32, IS_X86_64
 from pypy.jit.backend.detect_cpu import getcpuclass 
 from pypy.jit.backend.x86.regalloc import X86RegisterManager, X86_64_RegisterManager, X86XMMRegisterManager, X86_64_XMMRegisterManager
 from pypy.jit.codewriter import longlong
+import ctypes
 
 ACTUAL_CPU = getcpuclass()
 
@@ -238,3 +240,103 @@
         assert assembler.fail_boxes_int.getitem(i) == expected_ints[i]
         assert assembler.fail_boxes_ptr.getitem(i) == expected_ptrs[i]
         assert assembler.fail_boxes_float.getitem(i) == expected_floats[i]
+
+# ____________________________________________________________
+
+class TestRegallocPushPop(object):
+
+    def do_test(self, callback):
+        from pypy.jit.backend.x86.regalloc import X86FrameManager
+        from pypy.jit.backend.x86.regalloc import X86XMMRegisterManager
+        class FakeToken:
+            class compiled_loop_token:
+                asmmemmgr_blocks = None
+        cpu = ACTUAL_CPU(None, None)
+        cpu.setup()
+        looptoken = FakeToken()
+        asm = cpu.assembler
+        asm.setup_once()
+        asm.setup(looptoken)
+        self.fm = X86FrameManager()
+        self.xrm = X86XMMRegisterManager(None, frame_manager=self.fm,
+                                         assembler=asm)
+        callback(asm)
+        asm.mc.RET()
+        rawstart = asm.materialize_loop(looptoken)
+        #
+        F = ctypes.CFUNCTYPE(ctypes.c_long)
+        fn = ctypes.cast(rawstart, F)
+        res = fn()
+        return res
+
+    def test_simple(self):
+        def callback(asm):
+            asm.mov(imm(42), edx)
+            asm.regalloc_push(edx)
+            asm.regalloc_pop(eax)
+        res = self.do_test(callback)
+        assert res == 42
+
+    def test_push_stack(self):
+        def callback(asm):
+            loc = self.fm.frame_pos(5, INT)
+            asm.mc.SUB_ri(esp.value, 64)
+            asm.mov(imm(42), loc)
+            asm.regalloc_push(loc)
+            asm.regalloc_pop(eax)
+            asm.mc.ADD_ri(esp.value, 64)
+        res = self.do_test(callback)
+        assert res == 42
+
+    def test_pop_stack(self):
+        def callback(asm):
+            loc = self.fm.frame_pos(5, INT)
+            asm.mc.SUB_ri(esp.value, 64)
+            asm.mov(imm(42), edx)
+            asm.regalloc_push(edx)
+            asm.regalloc_pop(loc)
+            asm.mov(loc, eax)
+            asm.mc.ADD_ri(esp.value, 64)
+        res = self.do_test(callback)
+        assert res == 42
+
+    def test_simple_xmm(self):
+        def callback(asm):
+            c = ConstFloat(longlong.getfloatstorage(-42.5))
+            loc = self.xrm.convert_to_imm(c)
+            asm.mov(loc, xmm5)
+            asm.regalloc_push(xmm5)
+            asm.regalloc_pop(xmm0)
+            asm.mc.CVTTSD2SI(eax, xmm0)
+        res = self.do_test(callback)
+        assert res == -42
+
+    def test_push_stack_xmm(self):
+        def callback(asm):
+            c = ConstFloat(longlong.getfloatstorage(-42.5))
+            loc = self.xrm.convert_to_imm(c)
+            loc2 = self.fm.frame_pos(4, FLOAT)
+            asm.mc.SUB_ri(esp.value, 64)
+            asm.mov(loc, xmm5)
+            asm.mov(xmm5, loc2)
+            asm.regalloc_push(loc2)
+            asm.regalloc_pop(xmm0)
+            asm.mc.ADD_ri(esp.value, 64)
+            asm.mc.CVTTSD2SI(eax, xmm0)
+        res = self.do_test(callback)
+        assert res == -42
+
+    def test_pop_stack_xmm(self):
+        def callback(asm):
+            c = ConstFloat(longlong.getfloatstorage(-42.5))
+            loc = self.xrm.convert_to_imm(c)
+            loc2 = self.fm.frame_pos(4, FLOAT)
+            asm.mc.SUB_ri(esp.value, 64)
+            asm.mov(loc, xmm5)
+            asm.regalloc_push(xmm5)
+            asm.regalloc_pop(loc2)
+            asm.mov(loc2, xmm0)
+            asm.mc.ADD_ri(esp.value, 64)
+            asm.mc.CVTTSD2SI(eax, xmm0)
+        res = self.do_test(callback)
+        assert res == -42
diff --git a/pypy/jit/backend/x86/test/test_runner.py b/pypy/jit/backend/x86/test/test_runner.py
--- a/pypy/jit/backend/x86/test/test_runner.py
+++ b/pypy/jit/backend/x86/test/test_runner.py
@@ -6,6 +6,7 @@
                                          ConstPtr, Box, BoxFloat, BasicFailDescr)
 from pypy.jit.backend.detect_cpu import getcpuclass
 from pypy.jit.backend.x86.arch import WORD
+from pypy.jit.backend.x86.rx86 import fits_in_32bits
 from pypy.jit.backend.llsupport import symbolic
 from pypy.jit.metainterp.resoperation import rop
 from pypy.jit.metainterp.executor import execute
@@ -241,6 +242,23 @@
         c = self.execute_operation(rop.GETFIELD_GC, [res], 'int', ofsc3)
         assert c.value == 3
 
+    def test_bug_setfield_64bit(self):
+        if WORD == 4:
+            py.test.skip("only for 64 bits")
+        TP = lltype.GcStruct('S', ('i', lltype.Signed))
+        ofsi = self.cpu.fielddescrof(TP, 'i')
+        for i in range(500):
+            p = lltype.malloc(TP)
+            addr = rffi.cast(lltype.Signed, p)
+            if fits_in_32bits(addr):
+                break    # fitting in 32 bits, good
+        else:
+            py.test.skip("cannot get a 32-bit pointer")
+        res = ConstPtr(rffi.cast(llmemory.GCREF, addr))
+        self.execute_operation(rop.SETFIELD_RAW, [res, ConstInt(3**33)],
+                               'void', ofsi)
+        assert p.i == 3**33
+
     def test_nullity_with_guard(self):
         allops = [rop.INT_IS_TRUE]
         guards = [rop.GUARD_TRUE, rop.GUARD_FALSE]
diff --git a/pypy/jit/backend/x86/test/test_rx86.py b/pypy/jit/backend/x86/test/test_rx86.py
--- a/pypy/jit/backend/x86/test/test_rx86.py
+++ b/pypy/jit/backend/x86/test/test_rx86.py
@@ -185,6 +185,13 @@
     cb = CodeBuilder32
     assert_encodes_as(cb, 'PUSH_i32', (9,), '\x68\x09\x00\x00\x00')
 
+def test_sub_ji8():
+    cb = CodeBuilder32
+    assert_encodes_as(cb, 'SUB_ji8', (11223344, 55),
+                      '\x83\x2D\x30\x41\xAB\x00\x37')
+    assert_encodes_as(cb, 'SUB_mi8', ((edx, 16), 55),
+                      '\x83\x6A\x10\x37')
+
 class CodeBuilder64(CodeBuilderMixin, X86_64_CodeBuilder):
     pass
 
diff --git a/pypy/jit/codewriter/effectinfo.py b/pypy/jit/codewriter/effectinfo.py
--- a/pypy/jit/codewriter/effectinfo.py
+++ b/pypy/jit/codewriter/effectinfo.py
@@ -75,20 +75,22 @@
     #
     OS_MATH_SQRT                = 100
 
-    def __new__(cls, readonly_descrs_fields,
+    def __new__(cls, readonly_descrs_fields, readonly_descrs_arrays,
                 write_descrs_fields, write_descrs_arrays,
                 extraeffect=EF_CAN_RAISE,
                 oopspecindex=OS_NONE,
                 can_invalidate=False):
-        key = (_frozenset_or_none(readonly_descrs_fields),
-               _frozenset_or_none(write_descrs_fields),
-               _frozenset_or_none(write_descrs_arrays),
+        key = (frozenset(readonly_descrs_fields),
+               frozenset(readonly_descrs_arrays),
+               frozenset(write_descrs_fields),
+               frozenset(write_descrs_arrays),
                extraeffect,
                oopspecindex)
         if key in cls._cache:
             return cls._cache[key]
         result = object.__new__(cls)
         result.readonly_descrs_fields = readonly_descrs_fields
+        result.readonly_descrs_arrays = readonly_descrs_arrays
         if extraeffect == EffectInfo.EF_LOOPINVARIANT or \
            extraeffect == EffectInfo.EF_PURE:            
             result.write_descrs_fields = []
@@ -124,7 +126,7 @@
     if effects is top_set:
         return EffectInfo(None, None, None, extraeffect)
     readonly_descrs_fields = []
-    # readonly_descrs_arrays = [] --- not enabled for now
+    readonly_descrs_arrays = []
     write_descrs_fields = []
     write_descrs_arrays = []
 
@@ -150,10 +152,13 @@
         elif tup[0] == "array":
             add_array(write_descrs_arrays, tup)
         elif tup[0] == "readarray":
-            pass
+            tupw = ("array",) + tup[1:]
+            if tupw not in effects:
+                add_array(readonly_descrs_arrays, tup)
         else:
             assert 0
     return EffectInfo(readonly_descrs_fields,
+                      readonly_descrs_arrays,
                       write_descrs_fields,
                       write_descrs_arrays,
                       extraeffect,
diff --git a/pypy/jit/codewriter/policy.py b/pypy/jit/codewriter/policy.py
--- a/pypy/jit/codewriter/policy.py
+++ b/pypy/jit/codewriter/policy.py
@@ -44,10 +44,6 @@
             return True
         if mod.startswith('pypy.translator.'): # XXX wtf?
             return True
-        # string builder interface
-        if mod == 'pypy.rpython.lltypesystem.rbuilder':
-            return True
-        
         return False
 
     def look_inside_graph(self, graph):
diff --git a/pypy/jit/codewriter/test/test_effectinfo.py b/pypy/jit/codewriter/test/test_effectinfo.py
--- a/pypy/jit/codewriter/test/test_effectinfo.py
+++ b/pypy/jit/codewriter/test/test_effectinfo.py
@@ -35,6 +35,15 @@
     assert not effectinfo.readonly_descrs_fields
     assert not effectinfo.write_descrs_arrays
 
+def test_include_read_array():
+    A = lltype.GcArray(lltype.Signed)
+    effects = frozenset([("readarray", lltype.Ptr(A))])
+    effectinfo = effectinfo_from_writeanalyze(effects, FakeCPU())
+    assert not effectinfo.readonly_descrs_fields
+    assert list(effectinfo.readonly_descrs_arrays) == [('arraydescr', A)]
+    assert not effectinfo.write_descrs_fields
+    assert not effectinfo.write_descrs_arrays
+
 def test_include_write_array():
     A = lltype.GcArray(lltype.Signed)
     effects = frozenset([("array", lltype.Ptr(A))])
@@ -52,6 +61,16 @@
     assert list(effectinfo.write_descrs_fields) == [('fielddescr', S, "a")]
     assert not effectinfo.write_descrs_arrays
 
+def test_dont_include_read_and_write_array():
+    A = lltype.GcArray(lltype.Signed)
+    effects = frozenset([("readarray", lltype.Ptr(A)),
+                         ("array", lltype.Ptr(A))])
+    effectinfo = effectinfo_from_writeanalyze(effects, FakeCPU())
+    assert not effectinfo.readonly_descrs_fields
+    assert not effectinfo.readonly_descrs_arrays
+    assert not effectinfo.write_descrs_fields
+    assert list(effectinfo.write_descrs_arrays) == [('arraydescr', A)]
+
 
 def test_filter_out_typeptr():
     effects = frozenset([("struct", lltype.Ptr(OBJECT), "typeptr")])
diff --git a/pypy/jit/metainterp/blackhole.py b/pypy/jit/metainterp/blackhole.py
--- a/pypy/jit/metainterp/blackhole.py
+++ b/pypy/jit/metainterp/blackhole.py
@@ -3,7 +3,7 @@
 from pypy.rlib.rarithmetic import intmask, LONG_BIT, r_uint, ovfcheck
 from pypy.rlib.objectmodel import we_are_translated
 from pypy.rlib.debug import debug_start, debug_stop
-from pypy.rlib.debug import make_sure_not_resized, fatalerror
+from pypy.rlib.debug import make_sure_not_resized
 from pypy.rpython.lltypesystem import lltype, llmemory, rclass
 from pypy.rpython.lltypesystem.lloperation import llop
 from pypy.rpython.llinterp import LLException
diff --git a/pypy/jit/metainterp/compile.py b/pypy/jit/metainterp/compile.py
--- a/pypy/jit/metainterp/compile.py
+++ b/pypy/jit/metainterp/compile.py
@@ -3,7 +3,7 @@
 from pypy.rpython.ootypesystem import ootype
 from pypy.objspace.flow.model import Constant, Variable
 from pypy.rlib.objectmodel import we_are_translated
-from pypy.rlib.debug import debug_start, debug_stop
+from pypy.rlib.debug import debug_start, debug_stop, debug_print
 from pypy.rlib import rstack
 from pypy.conftest import option
 from pypy.tool.sourcetools import func_with_new_name
@@ -14,8 +14,8 @@
 from pypy.jit.metainterp.history import BoxPtr, BoxObj, BoxFloat, Const
 from pypy.jit.metainterp import history
 from pypy.jit.metainterp.typesystem import llhelper, oohelper
-from pypy.jit.metainterp.optimizeutil import InvalidLoop
-from pypy.jit.metainterp.resume import NUMBERING
+from pypy.jit.metainterp.optimize import InvalidLoop
+from pypy.jit.metainterp.resume import NUMBERING, PENDINGFIELDSP
 from pypy.jit.codewriter import heaptracker, longlong
 
 def giveup():
@@ -119,6 +119,7 @@
         old_loop_token = optimize_loop(metainterp_sd, old_loop_tokens, loop,
                                        jitdriver_sd.warmstate.enable_opts)
     except InvalidLoop:
+        debug_print("compile_new_loop: got an InvalidLoop")
         return None
     if old_loop_token is not None:
         metainterp.staticdata.log("reusing old loop")
@@ -302,7 +303,7 @@
     rd_numb = lltype.nullptr(NUMBERING)
     rd_consts = None
     rd_virtuals = None
-    rd_pendingfields = None
+    rd_pendingfields = lltype.nullptr(PENDINGFIELDSP.TO)
 
     CNT_INT   = -0x20000000
     CNT_REF   = -0x40000000
@@ -633,6 +634,7 @@
                                             new_loop, state.enable_opts,
                                             inline_short_preamble, retraced)
     except InvalidLoop:
+        debug_print("compile_new_bridge: got an InvalidLoop")
         # XXX I am fairly convinced that optimize_bridge cannot actually raise
         # InvalidLoop
         return None
diff --git a/pypy/jit/metainterp/executor.py b/pypy/jit/metainterp/executor.py
--- a/pypy/jit/metainterp/executor.py
+++ b/pypy/jit/metainterp/executor.py
@@ -316,6 +316,7 @@
             if value in (rop.FORCE_TOKEN,
                          rop.CALL_ASSEMBLER,
                          rop.COND_CALL_GC_WB,
+                         rop.COND_CALL_GC_WB_ARRAY,
                          rop.DEBUG_MERGE_POINT,
                          rop.JIT_DEBUG,
                          rop.SETARRAYITEM_RAW,
diff --git a/pypy/jit/metainterp/history.py b/pypy/jit/metainterp/history.py
--- a/pypy/jit/metainterp/history.py
+++ b/pypy/jit/metainterp/history.py
@@ -4,7 +4,7 @@
 from pypy.rpython.ootypesystem import ootype
 from pypy.rlib.objectmodel import we_are_translated, r_dict, Symbolic
 from pypy.rlib.objectmodel import compute_unique_id
-from pypy.rlib.rarithmetic import intmask, r_int64
+from pypy.rlib.rarithmetic import r_int64
 from pypy.conftest import option
 
 from pypy.jit.metainterp.resoperation import ResOperation, rop
@@ -791,6 +791,7 @@
 
     def dump(self):
         self.compiled_loop_token.cpu.dump_loop_token(self)
+
 class TreeLoop(object):
     inputargs = None
     operations = None
diff --git a/pypy/jit/metainterp/logger.py b/pypy/jit/metainterp/logger.py
--- a/pypy/jit/metainterp/logger.py
+++ b/pypy/jit/metainterp/logger.py
@@ -103,6 +103,7 @@
         if op.getopnum() == rop.DEBUG_MERGE_POINT:
             jd_sd = self.metainterp_sd.jitdrivers_sd[op.getarg(0).getint()]
             s = jd_sd.warmstate.get_location_str(op.getarglist()[2:])
+            s = s.replace(',', '.') # we use comma for argument splitting
             return "debug_merge_point(%d, '%s')" % (op.getarg(1).getint(), s)
         if ops_offset is None:
             offset = -1
diff --git a/pypy/jit/metainterp/optimize.py b/pypy/jit/metainterp/optimize.py
--- a/pypy/jit/metainterp/optimize.py
+++ b/pypy/jit/metainterp/optimize.py
@@ -1,9 +1,20 @@
 from pypy.rlib.debug import debug_start, debug_stop
+from pypy.jit.metainterp.jitexc import JitException
+
+class InvalidLoop(JitException):
+    """Raised when the optimize*.py detect that the loop that
+    we are trying to build cannot possibly make sense as a
+    long-running loop (e.g. it cannot run 2 complete iterations)."""
+
+class RetraceLoop(JitException):
+    """ Raised when inlining a short preamble resulted in an
+        InvalidLoop. This means the optimized loop is too specialized
+        to be useful here, so we trace it again and produced a second
+        copy specialized in some different way.
+    """
 
 # ____________________________________________________________
 
-from pypy.jit.metainterp.optimizeopt import optimize_loop_1, optimize_bridge_1
-
 def optimize_loop(metainterp_sd, old_loop_tokens, loop, enable_opts):
     debug_start("jit-optimize")
     try:
@@ -13,7 +24,7 @@
         debug_stop("jit-optimize")
 
 def _optimize_loop(metainterp_sd, old_loop_tokens, loop, enable_opts):
-    cpu = metainterp_sd.cpu
+    from pypy.jit.metainterp.optimizeopt import optimize_loop_1
     loop.logops = metainterp_sd.logger_noopt.log_loop(loop.inputargs,
                                                       loop.operations)
     # XXX do we really still need a list?
@@ -36,7 +47,7 @@
 
 def _optimize_bridge(metainterp_sd, old_loop_tokens, bridge, enable_opts,
                      inline_short_preamble, retraced=False):
-    cpu = metainterp_sd.cpu
+    from pypy.jit.metainterp.optimizeopt import optimize_bridge_1
     bridge.logops = metainterp_sd.logger_noopt.log_loop(bridge.inputargs,
                                                         bridge.operations)
     if old_loop_tokens:
diff --git a/pypy/jit/metainterp/optimizeopt/__init__.py b/pypy/jit/metainterp/optimizeopt/__init__.py
--- a/pypy/jit/metainterp/optimizeopt/__init__.py
+++ b/pypy/jit/metainterp/optimizeopt/__init__.py
@@ -15,7 +15,7 @@
             ('virtualize', OptVirtualize),
             ('string', OptString),
             ('heap', OptHeap),
-            ('ffi', OptFfiCall),
+            ('ffi', None),
             ('unroll', None)]
 # no direct instantiation of unroll
 unroll_all_opts = unrolling_iterable(ALL_OPTS)
@@ -25,10 +25,9 @@
 ALL_OPTS_NAMES = ':'.join([name for name, _ in ALL_OPTS])
 PARAMETERS['enable_opts'] = ALL_OPTS_NAMES
 
-def optimize_loop_1(metainterp_sd, loop, enable_opts,
+def build_opt_chain(metainterp_sd, enable_opts,
                     inline_short_preamble=True, retraced=False):
-    """Optimize loop.operations to remove internal overheadish operations.
-    """
+    config = metainterp_sd.config
     optimizations = []
     unroll = 'unroll' in enable_opts
     for name, opt in unroll_all_opts:
@@ -40,6 +39,11 @@
                 # FIXME: Workaround to disable string optimisation
                 # during preamble but to keep it during the loop
                 optimizations.append(o)
+            elif name == 'ffi' and config.translation.jit_ffi:
+                # we cannot put the class directly in the unrolling_iterable,
+                # because we do not want it to be seen at all (to avoid to
+                # introduce a dependency on libffi in case we do not need it)
+                optimizations.append(OptFfiCall())
 
     if ('rewrite' not in enable_opts or 'virtualize' not in enable_opts
         or 'heap' not in enable_opts):
@@ -48,6 +52,17 @@
     if inline_short_preamble:
         optimizations = [OptInlineShortPreamble(retraced)] + optimizations
 
+    return optimizations, unroll
+
+
+def optimize_loop_1(metainterp_sd, loop, enable_opts,
+                    inline_short_preamble=True, retraced=False):
+    """Optimize loop.operations to remove internal overheadish operations.
+    """
+
+    optimizations, unroll = build_opt_chain(metainterp_sd, enable_opts,
+                                            inline_short_preamble, retraced)
+
     if unroll:
         optimize_unroll(metainterp_sd, loop, optimizations)
     else:
diff --git a/pypy/jit/metainterp/optimizeopt/fficall.py b/pypy/jit/metainterp/optimizeopt/fficall.py
--- a/pypy/jit/metainterp/optimizeopt/fficall.py
+++ b/pypy/jit/metainterp/optimizeopt/fficall.py
@@ -4,7 +4,7 @@
 from pypy.rlib.debug import debug_start, debug_stop, debug_print, have_debug_prints
 from pypy.jit.codewriter.effectinfo import EffectInfo
 from pypy.jit.metainterp.resoperation import rop, ResOperation
-from pypy.jit.metainterp.optimizeutil import _findall
+from pypy.jit.metainterp.optimizeopt.util import _findall
 from pypy.jit.metainterp.optimizeopt.optimizer import Optimization
 from pypy.jit.backend.llsupport.ffisupport import UnsupportedKind
 
diff --git a/pypy/jit/metainterp/optimizeopt/heap.py b/pypy/jit/metainterp/optimizeopt/heap.py
--- a/pypy/jit/metainterp/optimizeopt/heap.py
+++ b/pypy/jit/metainterp/optimizeopt/heap.py
@@ -1,5 +1,5 @@
 import os
-from pypy.jit.metainterp.optimizeutil import _findall
+from pypy.jit.metainterp.optimizeopt.util import _findall
 from pypy.jit.metainterp.resoperation import rop, ResOperation
 from pypy.rlib.objectmodel import we_are_translated
 from pypy.jit.metainterp.jitexc import JitException
@@ -8,8 +8,8 @@
 
 class CachedField(object):
     def __init__(self):
-        # Cache information for a field descr.  It can be in one
-        # of two states:
+        # Cache information for a field descr, or for an (array descr, index)
+        # pair.  It can be in one of two states:
         #
         #   1. 'cached_fields' is a dict mapping OptValues of structs
         #      to OptValues of fields.  All fields on-heap are
@@ -27,19 +27,19 @@
         self._lazy_setfield_registered = False
 
     def do_setfield(self, optheap, op):
-        # Update the state with the SETFIELD_GC operation 'op'.
+        # Update the state with the SETFIELD_GC/SETARRAYITEM_GC operation 'op'.
         structvalue = optheap.getvalue(op.getarg(0))
-        fieldvalue  = optheap.getvalue(op.getarg(1))
+        fieldvalue  = optheap.getvalue(op.getarglist()[-1])
         if self.possible_aliasing(optheap, structvalue):
             self.force_lazy_setfield(optheap)
             assert not self.possible_aliasing(optheap, structvalue)
         cached_fieldvalue = self._cached_fields.get(structvalue, None)
         if cached_fieldvalue is not fieldvalue:
             # common case: store the 'op' as lazy_setfield, and register
-            # myself in the optheap's _lazy_setfields list
+            # myself in the optheap's _lazy_setfields_and_arrayitems list
             self._lazy_setfield = op
             if not self._lazy_setfield_registered:
-                optheap._lazy_setfields.append(self)
+                optheap._lazy_setfields_and_arrayitems.append(self)
                 self._lazy_setfield_registered = True
         else:
             # this is the case where the pending setfield ends up
@@ -65,7 +65,7 @@
         if self._lazy_setfield is not None:
             op = self._lazy_setfield
             assert optheap.getvalue(op.getarg(0)) is structvalue
-            return optheap.getvalue(op.getarg(1))
+            return optheap.getvalue(op.getarglist()[-1])
         else:
             return self._cached_fields.get(structvalue, None)
 
@@ -87,7 +87,7 @@
             # back in the cache: the value of this particular structure's
             # field.
             structvalue = optheap.getvalue(op.getarg(0))
-            fieldvalue  = optheap.getvalue(op.getarg(1))
+            fieldvalue  = optheap.getvalue(op.getarglist()[-1])
             self.remember_field_value(structvalue, fieldvalue)
 
     def get_reconstructed(self, optimizer, valuemap):
@@ -100,25 +100,20 @@
         return cf
 
 
-class CachedArrayItems(object):
-    def __init__(self):
-        self.fixed_index_items = {}
-        self.var_index_item = None
-        self.var_index_indexvalue = None
-
 class BogusPureField(JitException):
     pass
 
 
 class OptHeap(Optimization):
     """Cache repeated heap accesses"""
-    
+
     def __init__(self):
         # cached fields:  {descr: CachedField}
         self.cached_fields = {}
-        self._lazy_setfields = []
-        # cached array items:  {descr: CachedArrayItems}
+        # cached array items:  {array descr: {index: CachedField}}
         self.cached_arrayitems = {}
+        #
+        self._lazy_setfields_and_arrayitems = []
         self._remove_guard_not_invalidated = False
         self._seen_guard_not_invalidated = False
 
@@ -126,34 +121,23 @@
         new = OptHeap()
 
         if True:
-            self.force_all_lazy_setfields()
+            self.force_all_lazy_setfields_and_arrayitems()
         else:
             assert 0   # was: new.lazy_setfields = self.lazy_setfields
-        
+
         for descr, d in self.cached_fields.items():
             new.cached_fields[descr] = d.get_reconstructed(optimizer, valuemap)
 
-        new.cached_arrayitems = {}
-        for descr, d in self.cached_arrayitems.items():
-            newd = {}
-            new.cached_arrayitems[descr] = newd
-            for value, cache in d.items():
-                newcache = CachedArrayItems()
-                newd[value.get_reconstructed(optimizer, valuemap)] = newcache
-                if cache.var_index_item:
-                    newcache.var_index_item = \
-                          cache.var_index_item.get_reconstructed(optimizer, valuemap)
-                if cache.var_index_indexvalue:
-                    newcache.var_index_indexvalue = \
-                          cache.var_index_indexvalue.get_reconstructed(optimizer, valuemap)
-                for index, fieldvalue in cache.fixed_index_items.items():
-                    newcache.fixed_index_items[index] = \
-                           fieldvalue.get_reconstructed(optimizer, valuemap)
+        for descr, submap in self.cached_arrayitems.items():
+            newdict = {}
+            for index, d in submap.items():
+                newdict[index] = d.get_reconstructed(optimizer, valuemap)
+            new.cached_arrayitems[descr] = newdict
 
         return new
 
     def clean_caches(self):
-        del self._lazy_setfields[:]
+        del self._lazy_setfields_and_arrayitems[:]
         self.cached_fields.clear()
         self.cached_arrayitems.clear()
 
@@ -164,50 +148,16 @@
             cf = self.cached_fields[descr] = CachedField()
         return cf
 
-    def cache_arrayitem_value(self, descr, value, indexvalue, fieldvalue, write=False):
-        d = self.cached_arrayitems.get(descr, None)
-        if d is None:
-            d = self.cached_arrayitems[descr] = {}
-        cache = d.get(value, None)
-        if cache is None:
-            cache = d[value] = CachedArrayItems()
-        indexbox = self.get_constant_box(indexvalue.box)
-        if indexbox is not None:
-            index = indexbox.getint()
-            if write:
-                for value, othercache in d.iteritems():
-                    # fixed index, clean the variable index cache, in case the
-                    # index is the same
-                    othercache.var_index_indexvalue = None
-                    othercache.var_index_item = None
-                    try:
-                        del othercache.fixed_index_items[index]
-                    except KeyError:
-                        pass
-            cache.fixed_index_items[index] = fieldvalue
-        else:
-            if write:
-                for value, othercache in d.iteritems():
-                    # variable index, clear all caches for this descr
-                    othercache.var_index_indexvalue = None
-                    othercache.var_index_item = None
-                    othercache.fixed_index_items.clear()
-            cache.var_index_indexvalue = indexvalue
-            cache.var_index_item = fieldvalue
-
-    def read_cached_arrayitem(self, descr, value, indexvalue):
-        d = self.cached_arrayitems.get(descr, None)
-        if d is None:
-            return None
-        cache = d.get(value, None)
-        if cache is None:
-            return None
-        indexbox = self.get_constant_box(indexvalue.box)
-        if indexbox is not None:
-            return cache.fixed_index_items.get(indexbox.getint(), None)
-        elif cache.var_index_indexvalue is indexvalue:
-            return cache.var_index_item
-        return None
+    def arrayitem_cache(self, descr, index):
+        try:
+            submap = self.cached_arrayitems[descr]
+        except KeyError:
+            submap = self.cached_arrayitems[descr] = {}
+        try:
+            cf = submap[index]
+        except KeyError:
+            cf = submap[index] = CachedField()
+        return cf
 
     def emit_operation(self, op):
         self.emitting_operation(op)
@@ -219,7 +169,8 @@
         if op.is_ovf():
             return
         if op.is_guard():
-            self.optimizer.pendingfields = self.force_lazy_setfields_for_guard()
+            self.optimizer.pendingfields = (
+                self.force_lazy_setfields_and_arrayitems_for_guard())
             return
         opnum = op.getopnum()
         if (opnum == rop.SETFIELD_GC or        # handled specially
@@ -248,6 +199,8 @@
                 # XXX stored on effectinfo are large
                 for fielddescr in effectinfo.readonly_descrs_fields:
                     self.force_lazy_setfield(fielddescr)
+                for arraydescr in effectinfo.readonly_descrs_arrays:
+                    self.force_lazy_setarrayitem(arraydescr)
                 for fielddescr in effectinfo.write_descrs_fields:
                     self.force_lazy_setfield(fielddescr)
                     try:
@@ -256,8 +209,11 @@
                     except KeyError:
                         pass
                 for arraydescr in effectinfo.write_descrs_arrays:
+                    self.force_lazy_setarrayitem(arraydescr)
                     try:
-                        del self.cached_arrayitems[arraydescr]
+                        submap = self.cached_arrayitems[arraydescr]
+                        for cf in submap.itervalues():
+                            cf._cached_fields.clear()
                     except KeyError:
                         pass
                 if effectinfo.check_forces_virtual_or_virtualizable():
@@ -266,7 +222,7 @@
                     # ^^^ we only need to force this field; the other fields
                     # of virtualref_info and virtualizable_info are not gcptrs.
                 return
-        self.force_all_lazy_setfields()
+        self.force_all_lazy_setfields_and_arrayitems()
         self.clean_caches()
 
 
@@ -277,6 +233,10 @@
             for cf in self.cached_fields.itervalues():
                 if value in cf._cached_fields:
                     cf._cached_fields[newvalue] = cf._cached_fields[value]
+            for submap in self.cached_arrayitems.itervalues():
+                for cf in submap.itervalues():
+                    if value in cf._cached_fields:
+                        cf._cached_fields[newvalue] = cf._cached_fields[value]
 
     def force_lazy_setfield(self, descr):
         try:
@@ -285,6 +245,14 @@
             return
         cf.force_lazy_setfield(self)
 
+    def force_lazy_setarrayitem(self, arraydescr):
+        try:
+            submap = self.cached_arrayitems[arraydescr]
+        except KeyError:
+            return
+        for cf in submap.values():
+            cf.force_lazy_setfield(self)
+
     def fixup_guard_situation(self):
         # hackish: reverse the order of the last two operations if it makes
         # sense to avoid a situation like "int_eq/setfield_gc/guard_true",
@@ -309,30 +277,49 @@
         newoperations[-2] = lastop
         newoperations[-1] = prevop
 
-    def force_all_lazy_setfields(self):
-        for cf in self._lazy_setfields:
-            if not we_are_translated():
-                assert cf in self.cached_fields.values()
+    def _assert_valid_cf(self, cf):
+        # check that 'cf' is in cached_fields or cached_arrayitems
+        if not we_are_translated():
+            if cf not in self.cached_fields.values():
+                for submap in self.cached_arrayitems.values():
+                    if cf in submap.values():
+                        break
+                else:
+                    assert 0, "'cf' not in cached_fields/cached_arrayitems"
+
+    def force_all_lazy_setfields_and_arrayitems(self):
+        for cf in self._lazy_setfields_and_arrayitems:
+            self._assert_valid_cf(cf)
             cf.force_lazy_setfield(self)
 
-    def force_lazy_setfields_for_guard(self):
+    def force_lazy_setfields_and_arrayitems_for_guard(self):
         pendingfields = []
-        for cf in self._lazy_setfields:
-            if not we_are_translated():
-                assert cf in self.cached_fields.values()
+        for cf in self._lazy_setfields_and_arrayitems:
+            self._assert_valid_cf(cf)
             op = cf._lazy_setfield
             if op is None:
                 continue
             # the only really interesting case that we need to handle in the
             # guards' resume data is that of a virtual object that is stored
-            # into a field of a non-virtual object.
+            # into a field of a non-virtual object.  Here, 'op' in either
+            # SETFIELD_GC or SETARRAYITEM_GC.
             value = self.getvalue(op.getarg(0))
             assert not value.is_virtual()      # it must be a non-virtual
-            fieldvalue = self.getvalue(op.getarg(1))
+            fieldvalue = self.getvalue(op.getarglist()[-1])
             if fieldvalue.is_virtual():
                 # this is the case that we leave to resume.py
+                opnum = op.getopnum()
+                if opnum == rop.SETFIELD_GC:
+                    itemindex = -1
+                elif opnum == rop.SETARRAYITEM_GC:
+                    indexvalue = self.getvalue(op.getarg(1))
+                    assert indexvalue.is_constant()
+                    itemindex = indexvalue.box.getint()
+                    assert itemindex >= 0
+                else:
+                    assert 0
                 pendingfields.append((op.getdescr(), value.box,
-                                      fieldvalue.get_key_box()))
+                                      fieldvalue.get_key_box(), itemindex))
             else:
                 cf.force_lazy_setfield(self)
                 self.fixup_guard_situation()
@@ -364,24 +351,45 @@
         cf.do_setfield(self, op)
 
     def optimize_GETARRAYITEM_GC(self, op):
-        value = self.getvalue(op.getarg(0))
+        arrayvalue = self.getvalue(op.getarg(0))
         indexvalue = self.getvalue(op.getarg(1))
-        fieldvalue = self.read_cached_arrayitem(op.getdescr(), value, indexvalue)
-        if fieldvalue is not None:
-            self.make_equal_to(op.result, fieldvalue)
-            return
-        ###self.optimizer.optimize_default(op)
+        cf = None
+        if indexvalue.is_constant():
+            # use the cache on (arraydescr, index), which is a constant
+            cf = self.arrayitem_cache(op.getdescr(), indexvalue.box.getint())
+            fieldvalue = cf.getfield_from_cache(self, arrayvalue)
+            if fieldvalue is not None:
+                self.make_equal_to(op.result, fieldvalue)
+                return
+        else:
+            # variable index, so make sure the lazy setarrayitems are done
+            self.force_lazy_setarrayitem(op.getdescr())
+        # default case: produce the operation
+        arrayvalue.ensure_nonnull()
         self.emit_operation(op)
-        fieldvalue = self.getvalue(op.result)
-        self.cache_arrayitem_value(op.getdescr(), value, indexvalue, fieldvalue)
+        # the remember the result of reading the array item
+        if cf is not None:
+            fieldvalue = self.getvalue(op.result)
+            cf.remember_field_value(arrayvalue, fieldvalue)
 
     def optimize_SETARRAYITEM_GC(self, op):
-        self.emit_operation(op)
-        value = self.getvalue(op.getarg(0))
-        fieldvalue = self.getvalue(op.getarg(2))
+        if self.has_pure_result(rop.GETARRAYITEM_GC_PURE, [op.getarg(0),
+                                                           op.getarg(1)],
+                                op.getdescr()):
+            os.write(2, '[bogus immutable array declaration: %s]\n' %
+                     (op.getdescr().repr_of_descr()))
+            raise BogusPureField
+        #
         indexvalue = self.getvalue(op.getarg(1))
-        self.cache_arrayitem_value(op.getdescr(), value, indexvalue, fieldvalue,
-                                   write=True)
+        if indexvalue.is_constant():
+            # use the cache on (arraydescr, index), which is a constant
+            cf = self.arrayitem_cache(op.getdescr(), indexvalue.box.getint())
+            cf.do_setfield(self, op)
+        else:
+            # variable index, so make sure the lazy setarrayitems are done
+            self.force_lazy_setarrayitem(op.getdescr())
+            # and then emit the operation
+            self.emit_operation(op)
 
     def optimize_QUASIIMMUT_FIELD(self, op):
         # Pattern: QUASIIMMUT_FIELD(s, descr=QuasiImmutDescr)
diff --git a/pypy/jit/metainterp/optimizeopt/intbounds.py b/pypy/jit/metainterp/optimizeopt/intbounds.py
--- a/pypy/jit/metainterp/optimizeopt/intbounds.py
+++ b/pypy/jit/metainterp/optimizeopt/intbounds.py
@@ -1,7 +1,7 @@
 from pypy.jit.metainterp.optimizeopt.optimizer import Optimization, CONST_1, CONST_0
-from pypy.jit.metainterp.optimizeutil import _findall
-from pypy.jit.metainterp.optimizeopt.intutils import IntBound, IntUnbounded, \
-    IntLowerBound, IntUpperBound
+from pypy.jit.metainterp.optimizeopt.util import _findall
+from pypy.jit.metainterp.optimizeopt.intutils import (IntBound, IntUnbounded,
+    IntLowerBound, IntUpperBound)
 from pypy.jit.metainterp.history import Const, ConstInt
 from pypy.jit.metainterp.resoperation import rop, ResOperation
 
@@ -23,7 +23,7 @@
 
     def reconstruct_for_next_iteration(self, optimizer, valuemap):
         assert self.posponedop is None
-        return self 
+        return self
 
     def propagate_forward(self, op):
         if op.is_ovf():
@@ -194,7 +194,7 @@
                 # Synthesize the reverse ops for optimize_default to reuse
                 self.pure(rop.INT_ADD, [op.result, op.getarg(1)], op.getarg(0))
                 self.pure(rop.INT_SUB, [op.getarg(0), op.result], op.getarg(1))
-                
+
 
     def optimize_INT_MUL_OVF(self, op):
         v1 = self.getvalue(op.getarg(0))
@@ -292,6 +292,11 @@
         v1.intbound.make_ge(IntLowerBound(0))
         v1.intbound.make_lt(IntUpperBound(256))
 
+    def optimize_UNICODEGETITEM(self, op):
+        self.emit_operation(op)
+        v1 = self.getvalue(op.result)
+        v1.intbound.make_ge(IntLowerBound(0))
+
     def make_int_lt(self, box1, box2):
         v1 = self.getvalue(box1)
         v2 = self.getvalue(box2)
@@ -368,6 +373,15 @@
                 if v2.intbound.intersect(v1.intbound):
                     self.propagate_bounds_backward(op.getarg(1))
 
+    def propagate_bounds_INT_IS_TRUE(self, op):
+        r = self.getvalue(op.result)
+        if r.is_constant():
+            if r.box.same_constant(CONST_1):
+                v1 = self.getvalue(op.getarg(0))
+                if v1.intbound.known_ge(IntBound(0, 0)):
+                    v1.intbound.make_gt(IntBound(0, 0))
+                    self.propagate_bounds_backward(op.getarg(0))
+
     def propagate_bounds_INT_ADD(self, op):
         v1 = self.getvalue(op.getarg(0))
         v2 = self.getvalue(op.getarg(1))
@@ -413,5 +427,6 @@
     propagate_bounds_INT_SUB_OVF  = propagate_bounds_INT_SUB
     propagate_bounds_INT_MUL_OVF  = propagate_bounds_INT_MUL
 
+
 optimize_ops = _findall(OptIntBounds, 'optimize_')
 propagate_bounds_ops = _findall(OptIntBounds, 'propagate_bounds_')
diff --git a/pypy/jit/metainterp/optimizeopt/optimizer.py b/pypy/jit/metainterp/optimizeopt/optimizer.py
--- a/pypy/jit/metainterp/optimizeopt/optimizer.py
+++ b/pypy/jit/metainterp/optimizeopt/optimizer.py
@@ -4,9 +4,9 @@
 from pypy.jit.metainterp.resoperation import rop, ResOperation
 from pypy.jit.metainterp import jitprof
 from pypy.jit.metainterp.executor import execute_nonspec
-from pypy.jit.metainterp.optimizeutil import _findall, sort_descrs
-from pypy.jit.metainterp.optimizeutil import descrlist_dict
-from pypy.jit.metainterp.optimizeutil import InvalidLoop, args_dict
+from pypy.jit.metainterp.optimizeopt.util import _findall, sort_descrs
+from pypy.jit.metainterp.optimizeopt.util import descrlist_dict, args_dict
+from pypy.jit.metainterp.optimize import InvalidLoop
 from pypy.jit.metainterp import resume, compile
 from pypy.jit.metainterp.typesystem import llhelper, oohelper
 from pypy.rpython.lltypesystem import lltype
@@ -141,6 +141,9 @@
         # meaning it has been forced.
         return self.box is None
 
+    def is_forced_virtual(self):
+        return False
+
     def getfield(self, ofs, default):
         raise NotImplementedError
 
diff --git a/pypy/jit/metainterp/optimizeopt/rewrite.py b/pypy/jit/metainterp/optimizeopt/rewrite.py
--- a/pypy/jit/metainterp/optimizeopt/rewrite.py
+++ b/pypy/jit/metainterp/optimizeopt/rewrite.py
@@ -1,7 +1,7 @@
 from pypy.jit.metainterp.optimizeopt.optimizer import *
 from pypy.jit.metainterp.resoperation import opboolinvers, opboolreflex
 from pypy.jit.metainterp.history import ConstInt
-from pypy.jit.metainterp.optimizeutil import _findall
+from pypy.jit.metainterp.optimizeopt.util import _findall
 from pypy.jit.metainterp.resoperation import rop, ResOperation
 from pypy.jit.codewriter.effectinfo import EffectInfo
 from pypy.jit.metainterp.optimizeopt.intutils import IntBound
@@ -184,6 +184,32 @@
         else:
             self.emit_operation(op)
 
+    def optimize_FLOAT_MUL(self, op):
+        arg1 = op.getarg(0)
+        arg2 = op.getarg(1)
+
+        # Constant fold f0 * 1.0 and turn f0 * -1.0 into a FLOAT_NEG, these
+        # work in all cases, including NaN and inf
+        for lhs, rhs in [(arg1, arg2), (arg2, arg1)]:
+            v1 = self.getvalue(lhs)
+            v2 = self.getvalue(rhs)
+
+            if v1.is_constant():
+                if v1.box.getfloat() == 1.0:
+                    self.make_equal_to(op.result, v2)
+                    return
+                elif v1.box.getfloat() == -1.0:
+                    self.emit_operation(ResOperation(
+                        rop.FLOAT_NEG, [rhs], op.result
+                    ))
+                    return
+        self.emit_operation(op)
+
+    def optimize_FLOAT_NEG(self, op):
+        v1 = op.getarg(0)
+        self.emit_operation(op)
+        self.pure(rop.FLOAT_NEG, [op.result], v1)
+
     def optimize_CALL_PURE(self, op):
         arg_consts = []
         for i in range(op.numargs()):
@@ -193,7 +219,7 @@
                 break
             arg_consts.append(const)
         else:
-            # all constant arguments: check if we already know the reslut
+            # all constant arguments: check if we already know the result
             try:
                 result = self.optimizer.call_pure_results[arg_consts]
             except KeyError:
diff --git a/pypy/jit/metainterp/optimizeopt/simplify.py b/pypy/jit/metainterp/optimizeopt/simplify.py
--- a/pypy/jit/metainterp/optimizeopt/simplify.py
+++ b/pypy/jit/metainterp/optimizeopt/simplify.py
@@ -1,7 +1,7 @@
 
 from pypy.jit.metainterp.resoperation import ResOperation, rop
 from pypy.jit.metainterp.optimizeopt.optimizer import Optimization
-from pypy.jit.metainterp.optimizeutil import _findall
+from pypy.jit.metainterp.optimizeopt.util import _findall
 
 class OptSimplify(Optimization):
     def optimize_CALL_PURE(self, op):
diff --git a/pypy/jit/metainterp/optimizeopt/string.py b/pypy/jit/metainterp/optimizeopt/string.py
--- a/pypy/jit/metainterp/optimizeopt/string.py
+++ b/pypy/jit/metainterp/optimizeopt/string.py
@@ -8,7 +8,7 @@
 from pypy.jit.metainterp.optimizeopt import optimizer, virtualize
 from pypy.jit.metainterp.optimizeopt.optimizer import CONST_0, CONST_1
 from pypy.jit.metainterp.optimizeopt.optimizer import llhelper
-from pypy.jit.metainterp.optimizeutil import _findall
+from pypy.jit.metainterp.optimizeopt.util import _findall
 from pypy.jit.codewriter.effectinfo import EffectInfo
 from pypy.jit.codewriter import heaptracker
 from pypy.rlib.unroll import unrolling_iterable
@@ -348,7 +348,7 @@
     optimizer.emit_operation(ResOperation(rop.INT_SUB, [box1, box2], resbox))
     return resbox
 
-def _strgetitem(optimizer, strbox, indexbox, mode):
+def _strgetitem(optimization, strbox, indexbox, mode):
     if isinstance(strbox, ConstPtr) and isinstance(indexbox, ConstInt):
         if mode is mode_string:
             s = strbox.getref(lltype.Ptr(rstr.STR))
@@ -357,7 +357,7 @@
             s = strbox.getref(lltype.Ptr(rstr.UNICODE))
             return ConstInt(ord(s.chars[indexbox.getint()]))
     resbox = BoxInt()
-    optimizer.emit_operation(ResOperation(mode.STRGETITEM, [strbox, indexbox],
+    optimization.emit_operation(ResOperation(mode.STRGETITEM, [strbox, indexbox],
                                       resbox))
     return resbox
 
@@ -440,8 +440,7 @@
             if vindex.is_constant():
                 return value.getitem(vindex.box.getint())
         #
-        resbox = _strgetitem(self.optimizer,
-                             value.force_box(),vindex.force_box(), mode)
+        resbox = _strgetitem(self, value.force_box(), vindex.force_box(), mode)
         return self.getvalue(resbox)
 
     def optimize_STRLEN(self, op):
diff --git a/pypy/jit/metainterp/optimizeopt/test/__init__.py b/pypy/jit/metainterp/optimizeopt/test/__init__.py
new file mode 100644
diff --git a/pypy/jit/metainterp/test/test_optimizebasic.py b/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
rename from pypy/jit/metainterp/test/test_optimizebasic.py
rename to pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
--- a/pypy/jit/metainterp/test/test_optimizebasic.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
@@ -1,40 +1,15 @@
 import py
 from pypy.rlib.objectmodel import instantiate
-from pypy.jit.metainterp.test.test_optimizeutil import (LLtypeMixin,
-                                                        #OOtypeMixin,
-                                                        BaseTest)
+from pypy.jit.metainterp.optimizeopt.test.test_util import (
+    LLtypeMixin, BaseTest, FakeMetaInterpStaticData)
 from pypy.jit.metainterp.test.test_compile import FakeLogger
 import pypy.jit.metainterp.optimizeopt.optimizer as optimizeopt
 import pypy.jit.metainterp.optimizeopt.virtualize as virtualize
-from pypy.jit.metainterp.optimizeutil import InvalidLoop
+from pypy.jit.metainterp.optimize import InvalidLoop
 from pypy.jit.metainterp.history import AbstractDescr, ConstInt, BoxInt
-from pypy.jit.metainterp.jitprof import EmptyProfiler
 from pypy.jit.metainterp import executor, compile, resume, history
 from pypy.jit.metainterp.resoperation import rop, opname, ResOperation
-from pypy.jit.tool.oparser import pure_parse
-from pypy.jit.metainterp.optimizeutil import args_dict
-
-##class FakeFrame(object):
-##    parent_resumedata_snapshot = None
-##    parent_resumedata_frame_info_list = None
-
-##    def __init__(self, code="", pc=0):
-##        self.jitcode = code
-##        self.pc = pc
-
-class Fake(object):
-    failargs_limit = 1000
-    storedebug = None
-
-class FakeMetaInterpStaticData(object):
-
-    def __init__(self, cpu):
-        self.cpu = cpu
-        self.profiler = EmptyProfiler()
-        self.options = Fake()
-        self.globaldata = Fake()
-        self.logger_ops = FakeLogger()
-        self.logger_noopt = FakeLogger()
+
 
 def test_store_final_boxes_in_guard():
     from pypy.jit.metainterp.compile import ResumeGuardDescr
@@ -104,7 +79,7 @@
     assert vinfo3 is vinfo4
 
 def test_descrlist_dict():
-    from pypy.jit.metainterp import optimizeutil
+    from pypy.jit.metainterp.optimizeopt import util as optimizeutil
     h1 = optimizeutil.descrlist_hash([])
     h2 = optimizeutil.descrlist_hash([LLtypeMixin.valuedescr])
     h3 = optimizeutil.descrlist_hash(
@@ -133,159 +108,55 @@
 
 # ____________________________________________________________
 
-def equaloplists(oplist1, oplist2, strict_fail_args=True, remap={},
-                 text_right=None):
-    # try to use the full width of the terminal to display the list
-    # unfortunately, does not work with the default capture method of py.test
-    # (which is fd), you you need to use either -s or --capture=sys, else you
-    # get the standard 80 columns width
-    totwidth = py.io.get_terminal_width()
-    width = totwidth / 2 - 1
-    print ' Comparing lists '.center(totwidth, '-')
-    text_right = text_right or 'expected'
-    print '%s| %s' % ('optimized'.center(width), text_right.center(width))
-    for op1, op2 in zip(oplist1, oplist2):
-        txt1 = str(op1)
-        txt2 = str(op2)
-        while txt1 or txt2:
-            print '%s| %s' % (txt1[:width].ljust(width), txt2[:width])
-            txt1 = txt1[width:]
-            txt2 = txt2[width:]
-        assert op1.getopnum() == op2.getopnum()
-        assert op1.numargs() == op2.numargs()
-        for i in range(op1.numargs()):
-            x = op1.getarg(i)
-            y = op2.getarg(i)
-            assert x == remap.get(y, y)
-        if op2.result in remap:
-            assert op1.result == remap[op2.result]
-        else:
-            remap[op2.result] = op1.result
-        if op1.getopnum() != rop.JUMP:      # xxx obscure
-            assert op1.getdescr() == op2.getdescr()
-        if op1.getfailargs() or op2.getfailargs():
-            assert len(op1.getfailargs()) == len(op2.getfailargs())
-            if strict_fail_args:
-                for x, y in zip(op1.getfailargs(), op2.getfailargs()):
-                    assert x == remap.get(y, y)
-            else:
-                fail_args1 = set(op1.getfailargs())
-                fail_args2 = set([remap.get(y, y) for y in op2.getfailargs()])
-                assert fail_args1 == fail_args2
-    assert len(oplist1) == len(oplist2)
-    print '-'*totwidth
-    return True
-
-def test_equaloplists():
-    ops = """
-    [i0]
-    i1 = int_add(i0, 1)
-    i2 = int_add(i1, 1)
-    guard_true(i1) [i2]
-    jump(i1)
-    """
-    namespace = {}
-    loop1 = pure_parse(ops, namespace=namespace)
-    loop2 = pure_parse(ops, namespace=namespace)
-    loop3 = pure_parse(ops.replace("i2 = int_add", "i2 = int_sub"),
-                       namespace=namespace)
-    assert equaloplists(loop1.operations, loop2.operations)
-    py.test.raises(AssertionError,
-                   "equaloplists(loop1.operations, loop3.operations)")
-
-def test_equaloplists_fail_args():
-    ops = """
-    [i0]
-    i1 = int_add(i0, 1)
-    i2 = int_add(i1, 1)
-    guard_true(i1) [i2, i1]
-    jump(i1)
-    """
-    namespace = {}
-    loop1 = pure_parse(ops, namespace=namespace)
-    loop2 = pure_parse(ops.replace("[i2, i1]", "[i1, i2]"),
-                       namespace=namespace)
-    py.test.raises(AssertionError,
-                   "equaloplists(loop1.operations, loop2.operations)")
-    assert equaloplists(loop1.operations, loop2.operations,
-                        strict_fail_args=False)
-    loop3 = pure_parse(ops.replace("[i2, i1]", "[i2, i0]"),
-                       namespace=namespace)
-    py.test.raises(AssertionError,
-                   "equaloplists(loop1.operations, loop3.operations)")
-
-# ____________________________________________________________
-
-class Storage(compile.ResumeGuardDescr):
-    "for tests."
-    def __init__(self, metainterp_sd=None, original_greenkey=None):
-        self.metainterp_sd = metainterp_sd
-        self.original_greenkey = original_greenkey
-    def store_final_boxes(self, op, boxes):
-        op.setfailargs(boxes)
-    def __eq__(self, other):
-        return type(self) is type(other)      # xxx obscure
-
-def _sortboxes(boxes):
-    _kind2count = {history.INT: 1, history.REF: 2, history.FLOAT: 3}
-    return sorted(boxes, key=lambda box: _kind2count[box.type])
 
 class BaseTestBasic(BaseTest):
 
-    def invent_fail_descr(self, fail_args):
-        if fail_args is None:
-            return None
-        descr = Storage()
-        descr.rd_frame_info_list = resume.FrameInfo(None, "code", 11)
-        descr.rd_snapshot = resume.Snapshot(None, _sortboxes(fail_args))
-        return descr
-
-    def assert_equal(self, optimized, expected):
-        assert len(optimized.inputargs) == len(expected.inputargs)
-        remap = {}
-        for box1, box2 in zip(optimized.inputargs, expected.inputargs):
-            assert box1.__class__ == box2.__class__
-            remap[box2] = box1
-        assert equaloplists(optimized.operations,
-                            expected.operations, False, remap)
+    enable_opts = "intbounds:rewrite:virtualize:string:heap"
 
     def optimize_loop(self, ops, optops, call_pure_results=None):
+
         loop = self.parse(ops)
-        #
-        self.loop = loop
-        loop.call_pure_results = args_dict()
-        if call_pure_results is not None:
-            for k, v in call_pure_results.items():
-                loop.call_pure_results[list(k)] = v
-        metainterp_sd = FakeMetaInterpStaticData(self.cpu)
-        if hasattr(self, 'vrefinfo'):
-            metainterp_sd.virtualref_info = self.vrefinfo
-        if hasattr(self, 'callinfocollection'):
-            metainterp_sd.callinfocollection = self.callinfocollection
-        #
-        # XXX list the exact optimizations that are needed for each test
-        from pypy.jit.metainterp.optimizeopt import (OptIntBounds,
-                                                     OptRewrite,
-                                                     OptVirtualize,
-                                                     OptString,
-                                                     OptHeap,
-                                                     Optimizer)
-        from pypy.jit.metainterp.optimizeopt.fficall import OptFfiCall
-
-        optimizations = [OptIntBounds(),
-                         OptRewrite(),
-                         OptVirtualize(),
-                         OptString(),
-                         OptHeap(),
-                         OptFfiCall(),
-                         ]
-        optimizer = Optimizer(metainterp_sd, loop, optimizations)
-        optimizer.propagate_all_forward()
-        #
         expected = self.parse(optops)
+        self._do_optimize_loop(loop, call_pure_results)
         print '\n'.join([str(o) for o in loop.operations])
         self.assert_equal(loop, expected)
 
+    def setup_method(self, meth=None):
+        class FailDescr(compile.ResumeGuardDescr):
+            oparse = None
+            def _oparser_uses_descr_of_guard(self, oparse, fail_args):
+                # typically called 3 times: once when parsing 'ops',
+                # once when parsing 'preamble', once when parsing 'expected'.
+                self.oparse = oparse
+                self.rd_frame_info_list, self.rd_snapshot = snapshot(fail_args)
+            def _clone_if_mutable(self):
+                assert self is fdescr
+                return fdescr2
+            def __repr__(self):
+                if self is fdescr:
+                    return 'fdescr'
+                if self is fdescr2:
+                    return 'fdescr2'
+                return compile.ResumeGuardDescr.__repr__(self)
+        #
+        def snapshot(fail_args, got=[]):
+            if not got:    # only the first time, i.e. when parsing 'ops'
+                rd_frame_info_list = resume.FrameInfo(None, "code", 11)
+                rd_snapshot = resume.Snapshot(None, fail_args)
+                got.append(rd_frame_info_list)
+                got.append(rd_snapshot)
+            return got
+        #
+        fdescr = instantiate(FailDescr)
+        self.namespace['fdescr'] = fdescr
+        fdescr2 = instantiate(FailDescr)
+        self.namespace['fdescr2'] = fdescr2
+
+    def teardown_method(self, meth):
+        self.namespace.pop('fdescr', None)
+        self.namespace.pop('fdescr2', None)
+
+
 
 class BaseTestOptimizeBasic(BaseTestBasic):
 
@@ -1234,8 +1105,8 @@
         """
         expected = """
         [i1, p0]
+        p1 = new_array(i1, descr=arraydescr)
         setarrayitem_gc(p0, 0, i1, descr=arraydescr)
-        p1 = new_array(i1, descr=arraydescr)
         jump(i1, p1)
         """
         self.optimize_loop(ops, expected)
@@ -1600,9 +1471,9 @@
         i3 = getarrayitem_gc_pure(p3, 1, descr=arraydescr)
         i4 = getarrayitem_gc(p3, i3, descr=arraydescr)
         i5 = int_add(i3, i4)
-        setarrayitem_gc(p3, 0, i5, descr=arraydescr)
         #
         setfield_gc(p1, i2, descr=valuedescr)
+        setarrayitem_gc(p3, 0, i5, descr=arraydescr)
         setfield_gc(p1, i4, descr=nextdescr)
         jump(p1, i1, i2, p3)
         """
@@ -1776,6 +1647,7 @@
         self.optimize_loop(ops, expected)
 
     def test_duplicate_getarrayitem_after_setarrayitem_2(self):
+        py.test.skip("setarrayitem with variable index")
         ops = """
         [p1, p2, p3, i1]
         setarrayitem_gc(p1, 0, p2, descr=arraydescr2)
@@ -2038,7 +1910,6 @@
         self.optimize_loop(ops, expected)
 
     def test_merge_guard_nonnull_guard_class(self):
-        self.make_fail_descr()
         ops = """
         [p1, i0, i1, i2, p2]
         guard_nonnull(p1, descr=fdescr) [i0]
@@ -2056,7 +1927,6 @@
         self.check_expanded_fail_descr("i0", rop.GUARD_NONNULL_CLASS)
 
     def test_merge_guard_nonnull_guard_value(self):
-        self.make_fail_descr()
         ops = """
         [p1, i0, i1, i2, p2]
         guard_nonnull(p1, descr=fdescr) [i0]
@@ -2074,7 +1944,6 @@
         self.check_expanded_fail_descr("i0", rop.GUARD_VALUE)
 
     def test_merge_guard_nonnull_guard_class_guard_value(self):
-        self.make_fail_descr()
         ops = """
         [p1, i0, i1, i2, p2]
         guard_nonnull(p1, descr=fdescr) [i0]
@@ -2290,25 +2159,83 @@
         """
         self.optimize_loop(ops, expected)
 
+    def test_fold_constant_partial_ops_float(self):
+        ops = """
+        [f0]
+        f1 = float_mul(f0, 1.0)
+        f2 = escape(f1)
+        jump(f2)
+        """
+        expected = """
+        [f0]
+        f2 = escape(f0)
+        jump(f2)
+        """
+        self.optimize_loop(ops, expected)
+
+        ops = """
+        [f0]
+        f1 = float_mul(1.0, f0)
+        f2 = escape(f1)
+        jump(f2)
+        """
+        expected = """
+        [f0]
+        f2 = escape(f0)
+        jump(f2)
+        """
+        self.optimize_loop(ops, expected)
+
+
+        ops = """
+        [f0]
+        f1 = float_mul(f0, -1.0)
+        f2 = escape(f1)
+        jump(f2)
+        """
+        expected = """
+        [f0]
+        f1 = float_neg(f0)
+        f2 = escape(f1)
+        jump(f2)
+        """
+        self.optimize_loop(ops, expected)
+
+        ops = """
+        [f0]
+        f1 = float_mul(-1.0, f0)
+        f2 = escape(f1)
+        jump(f2)
+        """
+        expected = """
+        [f0]
+        f1 = float_neg(f0)
+        f2 = escape(f1)
+        jump(f2)
+        """
+        self.optimize_loop(ops, expected)
+
+    def test_fold_repeated_float_neg(self):
+        ops = """
+        [f0]
+        f1 = float_neg(f0)
+        f2 = float_neg(f1)
+        f3 = float_neg(f2)
+        f4 = float_neg(f3)
+        escape(f4)
+        jump(f4)
+        """
+        expected = """
+        [f0]
+        # The backend removes this dead op.
+        f1 = float_neg(f0)
+        escape(f0)
+        jump(f0)
+        """
+        self.optimize_loop(ops, expected)
+
     # ----------
 
-    def make_fail_descr(self):
-        class FailDescr(compile.ResumeGuardDescr):
-            oparse = None
-            def _oparser_uses_descr_of_guard(self, oparse, fail_args):
-                # typically called twice, before and after optimization
-                if self.oparse is None:
-                    fdescr.rd_frame_info_list = resume.FrameInfo(None,
-                                                                 "code", 11)
-                    fdescr.rd_snapshot = resume.Snapshot(None, fail_args)
-                self.oparse = oparse
-        #
-        fdescr = instantiate(FailDescr)
-        self.namespace['fdescr'] = fdescr
-
-    def teardown_method(self, meth):
-        self.namespace.pop('fdescr', None)
-
     def _verify_fail_args(self, boxes, oparse, text):
         import re
         r = re.compile(r"\bwhere\s+(\w+)\s+is a\s+(\w+)")
@@ -2417,7 +2344,6 @@
         self._verify_fail_args(boxes, fdescr.oparse, expectedtext)
 
     def test_expand_fail_1(self):
-        self.make_fail_descr()
         ops = """
         [i1, i3]
         # first rename i3 into i4
@@ -2438,7 +2364,6 @@
         self.check_expanded_fail_descr('15, i3', rop.GUARD_TRUE)
 
     def test_expand_fail_2(self):
-        self.make_fail_descr()
         ops = """
         [i1, i2]
         p1 = new_with_vtable(ConstClass(node_vtable))
@@ -2458,7 +2383,6 @@
             ''', rop.GUARD_TRUE)
 
     def test_expand_fail_3(self):
-        self.make_fail_descr()
         ops = """
         [i1, i2, i3, p3]
         p1 = new_with_vtable(ConstClass(node_vtable))
@@ -2484,7 +2408,7 @@
     def test_expand_fail_4(self):
         for arg in ['p1', 'i2,p1', 'p1,p2', 'p2,p1',
                     'i2,p1,p2', 'i2,p2,p1']:
-            self.make_fail_descr()
+            self.setup_method() # humpf
             ops = """
             [i1, i2, i3]
             p1 = new_with_vtable(ConstClass(node_vtable))
@@ -2509,7 +2433,6 @@
                                            rop.GUARD_TRUE)
 
     def test_expand_fail_5(self):
-        self.make_fail_descr()
         ops = """
         [i1, i2, i3, i4]
         p1 = new_with_vtable(ConstClass(node_vtable))
@@ -2533,7 +2456,6 @@
             ''', rop.GUARD_TRUE)
 
     def test_expand_fail_6(self):
-        self.make_fail_descr()
         ops = """
         [p0, i0, i1]
         guard_true(i0, descr=fdescr) [p0]
@@ -2554,7 +2476,6 @@
             ''', rop.GUARD_TRUE)
 
     def test_expand_fail_varray(self):
-        self.make_fail_descr()
         ops = """
         [i1]
         p1 = new_array(3, descr=arraydescr)
@@ -2575,7 +2496,6 @@
             ''', rop.GUARD_TRUE)
 
     def test_expand_fail_vstruct(self):
-        self.make_fail_descr()
         ops = """
         [i1, p1]
         p2 = new(descr=ssize)
@@ -2597,7 +2517,6 @@
             ''', rop.GUARD_TRUE)
 
     def test_expand_fail_v_all_1(self):
-        self.make_fail_descr()
         ops = """
         [i1, p1a, i2]
         p6s = getarrayitem_gc(p1a, 0, descr=arraydescr2)
@@ -2639,7 +2558,6 @@
             ''', rop.GUARD_TRUE)
 
     def test_expand_fail_lazy_setfield_1(self):
-        self.make_fail_descr()
         ops = """
         [p1, i2, i3]
         p2 = new_with_vtable(ConstClass(node_vtable))
@@ -2665,7 +2583,6 @@
             ''', rop.GUARD_TRUE)
 
     def test_expand_fail_lazy_setfield_2(self):
-        self.make_fail_descr()
         ops = """
         [i2, i3]
         p2 = new_with_vtable(ConstClass(node_vtable))
@@ -2689,9 +2606,6 @@
             where p2 is a node_vtable, valuedescr=i2
             ''', rop.GUARD_TRUE)
 
-
-class TestLLtype(BaseTestOptimizeBasic, LLtypeMixin):
-
     def test_residual_call_does_not_invalidate_caches(self):
         ops = """
         [p1, p2]
@@ -2983,7 +2897,6 @@
         self.optimize_loop(ops, expected)
 
     def test_vref_virtual_2(self):
-        self.make_fail_descr()
         ops = """
         [p0, i1]
         #
@@ -3029,7 +2942,6 @@
             ''', rop.GUARD_NOT_FORCED)
 
     def test_vref_virtual_and_lazy_setfield(self):
-        self.make_fail_descr()
         ops = """
         [p0, i1]
         #
@@ -3068,7 +2980,6 @@
             ''', rop.GUARD_NO_EXCEPTION)
 
     def test_vref_virtual_after_finish(self):
-        self.make_fail_descr()
         ops = """
         [i1]
         p1 = new_with_vtable(ConstClass(node_vtable))
@@ -3095,7 +3006,6 @@
         self.optimize_loop(ops, expected)
 
     def test_vref_nonvirtual_and_lazy_setfield(self):
-        self.make_fail_descr()
         ops = """
         [i1, p1]
         p2 = virtual_ref(p1, 23)
@@ -4569,6 +4479,47 @@
         # not obvious, because of the exception UnicodeDecodeError that
         # can be raised by ll_str2unicode()
 
+    def test_strgetitem_repeated(self):
+        ops = """
+        [p0, i0]
+        i1 = strgetitem(p0, i0)
+        i2 = strgetitem(p0, i0)
+        i3 = int_eq(i1, i2)
+        guard_true(i3) []
+        escape(i2)
+        jump(p0, i0)
+        """
+        expected = """
+        [p0, i0]
+        i1 = strgetitem(p0, i0)
+        escape(i1)
+        jump(p0, i0)
+        """
+        self.optimize_loop(ops, expected)
+
+    def test_int_is_true_bounds(self):
+        ops = """
+        [p0]
+        i0 = strlen(p0)
+        i1 = int_is_true(i0)
+        guard_true(i1) []
+        i2 = int_ge(0, i0)
+        guard_false(i2) []
+        jump(p0)
+        """
+        expected = """
+        [p0]
+        i0 = strlen(p0)
+        i1 = int_is_true(i0)
+        guard_true(i1) []
+        jump(p0)
+        """
+        self.optimize_loop(ops, expected)
+
+
+class TestLLtype(BaseTestOptimizeBasic, LLtypeMixin):
+    pass
+
 
 ##class TestOOtype(BaseTestOptimizeBasic, OOtypeMixin):
 
diff --git a/pypy/jit/metainterp/test/test_optimizefficall.py b/pypy/jit/metainterp/optimizeopt/test/test_optimizefficall.py
rename from pypy/jit/metainterp/test/test_optimizefficall.py
rename to pypy/jit/metainterp/optimizeopt/test/test_optimizefficall.py
--- a/pypy/jit/metainterp/test/test_optimizefficall.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_optimizefficall.py
@@ -2,8 +2,8 @@
 from pypy.rlib.libffi import Func, types
 from pypy.jit.metainterp.history import AbstractDescr
 from pypy.jit.codewriter.effectinfo import EffectInfo
-from pypy.jit.metainterp.test.test_optimizebasic import BaseTestBasic
-from pypy.jit.metainterp.test.test_optimizebasic import LLtypeMixin
+from pypy.jit.metainterp.optimizeopt.test.test_optimizebasic import BaseTestBasic
+from pypy.jit.metainterp.optimizeopt.test.test_optimizebasic import LLtypeMixin
 
 class MyCallDescr(AbstractDescr):
     """
@@ -32,7 +32,8 @@
 
 
 class TestFfiCall(BaseTestBasic, LLtypeMixin):
-    jit_ffi = True
+
+    enable_opts = "intbounds:rewrite:virtualize:string:heap:ffi"
 
     class namespace:
         cpu = LLtypeMixin.cpu
@@ -50,7 +51,7 @@
                              restype=types.sint)
         #
         def calldescr(cpu, FUNC, oopspecindex, extraeffect=None):
-            einfo = EffectInfo([], [], [], oopspecindex=oopspecindex,
+            einfo = EffectInfo([], [], [], [], oopspecindex=oopspecindex,
                                extraeffect=extraeffect)
             return cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT, einfo)
         #
diff --git a/pypy/jit/metainterp/test/test_optimizeopt.py b/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
rename from pypy/jit/metainterp/test/test_optimizeopt.py
rename to pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
--- a/pypy/jit/metainterp/test/test_optimizeopt.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
@@ -1,202 +1,88 @@
 import py
 from pypy.rlib.objectmodel import instantiate
-from pypy.jit.metainterp.test.test_optimizeutil import (LLtypeMixin,
-                                                        #OOtypeMixin,
-                                                        BaseTest)
+from pypy.jit.metainterp.optimizeopt.test.test_util import (
+    LLtypeMixin, BaseTest, Storage, _sortboxes)
 import pypy.jit.metainterp.optimizeopt.optimizer as optimizeopt
 import pypy.jit.metainterp.optimizeopt.virtualize as virtualize
-from pypy.jit.metainterp.optimizeopt import optimize_loop_1, ALL_OPTS_DICT
-from pypy.jit.metainterp.optimizeutil import InvalidLoop
+from pypy.jit.metainterp.optimizeopt import optimize_loop_1, ALL_OPTS_DICT, build_opt_chain
+from pypy.jit.metainterp.optimize import InvalidLoop
 from pypy.jit.metainterp.history import AbstractDescr, ConstInt, BoxInt
 from pypy.jit.metainterp.history import TreeLoop, LoopToken
 from pypy.jit.metainterp.jitprof import EmptyProfiler
 from pypy.jit.metainterp import executor, compile, resume, history
 from pypy.jit.metainterp.resoperation import rop, opname, ResOperation
 from pypy.jit.tool.oparser import pure_parse
-from pypy.jit.metainterp.test.test_optimizebasic import equaloplists
-from pypy.jit.metainterp.optimizeutil import args_dict
-
-class Fake(object):
-    failargs_limit = 1000
-    storedebug = None
-
-class FakeMetaInterpStaticData(object):
-
-    def __init__(self, cpu, jit_ffi=False):
-        self.cpu = cpu
-        self.profiler = EmptyProfiler()
-        self.options = Fake()
-        self.globaldata = Fake()
-        self.jit_ffi = jit_ffi
-
-def test_store_final_boxes_in_guard():
-    from pypy.jit.metainterp.compile import ResumeGuardDescr
-    from pypy.jit.metainterp.resume import tag, TAGBOX
-    b0 = BoxInt()
-    b1 = BoxInt()
-    opt = optimizeopt.Optimizer(FakeMetaInterpStaticData(LLtypeMixin.cpu),
-                                None)
-    fdescr = ResumeGuardDescr()
-    op = ResOperation(rop.GUARD_TRUE, ['dummy'], None, descr=fdescr)
-    # setup rd data
-    fi0 = resume.FrameInfo(None, "code0", 11)
-    fdescr.rd_frame_info_list = resume.FrameInfo(fi0, "code1", 33)
-    snapshot0 = resume.Snapshot(None, [b0])
-    fdescr.rd_snapshot = resume.Snapshot(snapshot0, [b1])
+from pypy.jit.metainterp.optimizeopt.util import args_dict
+from pypy.jit.metainterp.optimizeopt.test.test_optimizebasic import FakeMetaInterpStaticData
+from pypy.config.pypyoption import get_pypy_config
+
+
+def test_build_opt_chain():
+    def check(chain, expected_names):
+        names = [opt.__class__.__name__ for opt in chain]
+        assert names == expected_names
     #
-    opt.store_final_boxes_in_guard(op)
-    if op.getfailargs() == [b0, b1]:
-        assert list(fdescr.rd_numb.nums)      == [tag(1, TAGBOX)]
-        assert list(fdescr.rd_numb.prev.nums) == [tag(0, TAGBOX)]
-    else:
-        assert op.getfailargs() == [b1, b0]
-        assert list(fdescr.rd_numb.nums)      == [tag(0, TAGBOX)]
-        assert list(fdescr.rd_numb.prev.nums) == [tag(1, TAGBOX)]
-    assert fdescr.rd_virtuals is None
-    assert fdescr.rd_consts == []
-
-def test_sharing_field_lists_of_virtual():
-    class FakeOptimizer(object):
-        class cpu(object):
-            pass
-    opt = FakeOptimizer()
-    virt1 = virtualize.AbstractVirtualStructValue(opt, None)
-    lst1 = virt1._get_field_descr_list()
-    assert lst1 == []
-    lst2 = virt1._get_field_descr_list()
-    assert lst1 is lst2
-    virt1.setfield(LLtypeMixin.valuedescr, optimizeopt.OptValue(None))
-    lst3 = virt1._get_field_descr_list()
-    assert lst3 == [LLtypeMixin.valuedescr]
-    lst4 = virt1._get_field_descr_list()
-    assert lst3 is lst4
-
-    virt2 = virtualize.AbstractVirtualStructValue(opt, None)
-    lst5 = virt2._get_field_descr_list()
-    assert lst5 is lst1
-    virt2.setfield(LLtypeMixin.valuedescr, optimizeopt.OptValue(None))
-    lst6 = virt1._get_field_descr_list()
-    assert lst6 is lst3
-
-def test_reuse_vinfo():
-    class FakeVInfo(object):
-        def set_content(self, fieldnums):
-            self.fieldnums = fieldnums
-        def equals(self, fieldnums):
-            return self.fieldnums == fieldnums
-    class FakeVirtualValue(virtualize.AbstractVirtualValue):
-        def _make_virtual(self, *args):
-            return FakeVInfo()
-    v1 = FakeVirtualValue(None, None, None)
-    vinfo1 = v1.make_virtual_info(None, [1, 2, 4])
-    vinfo2 = v1.make_virtual_info(None, [1, 2, 4])
-    assert vinfo1 is vinfo2
-    vinfo3 = v1.make_virtual_info(None, [1, 2, 6])
-    assert vinfo3 is not vinfo2
-    vinfo4 = v1.make_virtual_info(None, [1, 2, 6])
-    assert vinfo3 is vinfo4
-
-def test_descrlist_dict():
-    from pypy.jit.metainterp import optimizeutil
-    h1 = optimizeutil.descrlist_hash([])
-    h2 = optimizeutil.descrlist_hash([LLtypeMixin.valuedescr])
-    h3 = optimizeutil.descrlist_hash(
-            [LLtypeMixin.valuedescr, LLtypeMixin.nextdescr])
-    assert h1 != h2
-    assert h2 != h3
-    assert optimizeutil.descrlist_eq([], [])
-    assert not optimizeutil.descrlist_eq([], [LLtypeMixin.valuedescr])
-    assert optimizeutil.descrlist_eq([LLtypeMixin.valuedescr],
-                                     [LLtypeMixin.valuedescr])
-    assert not optimizeutil.descrlist_eq([LLtypeMixin.valuedescr],
-                                         [LLtypeMixin.nextdescr])
-    assert optimizeutil.descrlist_eq([LLtypeMixin.valuedescr, LLtypeMixin.nextdescr],
-                                     [LLtypeMixin.valuedescr, LLtypeMixin.nextdescr])
-    assert not optimizeutil.descrlist_eq([LLtypeMixin.nextdescr, LLtypeMixin.valuedescr],
-                                         [LLtypeMixin.valuedescr, LLtypeMixin.nextdescr])
-
-    # descrlist_eq should compare by identity of the descrs, not by the result
-    # of sort_key
-    class FakeDescr(object):
-        def sort_key(self):
-            return 1
-
-    assert not optimizeutil.descrlist_eq([FakeDescr()], [FakeDescr()])
+    metainterp_sd = FakeMetaInterpStaticData(None)
+    chain, _ = build_opt_chain(metainterp_sd, "", inline_short_preamble=False)
+    check(chain, ["OptSimplify"])
+    #
+    chain, _ = build_opt_chain(metainterp_sd, "")
+    check(chain, ["OptInlineShortPreamble", "OptSimplify"])
+    #
+    chain, _ = build_opt_chain(metainterp_sd, "")
+    check(chain, ["OptInlineShortPreamble", "OptSimplify"])
+    #
+    chain, _ = build_opt_chain(metainterp_sd, "heap:intbounds")
+    check(chain, ["OptInlineShortPreamble", "OptIntBounds", "OptHeap", "OptSimplify"])
+    #
+    chain, unroll = build_opt_chain(metainterp_sd, "unroll")
+    check(chain, ["OptInlineShortPreamble", "OptSimplify"])
+    assert unroll
+    #
+    chain, _ = build_opt_chain(metainterp_sd, "aaa:bbb", inline_short_preamble=False)
+    check(chain, ["OptSimplify"])
+    #
+    chain, _ = build_opt_chain(metainterp_sd, "ffi", inline_short_preamble=False)
+    check(chain, ["OptFfiCall", "OptSimplify"])
+    #
+    metainterp_sd.config = get_pypy_config(translating=True)
+    assert not metainterp_sd.config.translation.jit_ffi
+    chain, _ = build_opt_chain(metainterp_sd, "ffi", inline_short_preamble=False)
+    check(chain, ["OptSimplify"])
+
 
 # ____________________________________________________________
-class Storage(compile.ResumeGuardDescr):
-    "for tests."
-    def __init__(self, metainterp_sd=None, original_greenkey=None):
-        self.metainterp_sd = metainterp_sd
-        self.original_greenkey = original_greenkey
-    def store_final_boxes(self, op, boxes):
-        op.setfailargs(boxes)
-    def __eq__(self, other):
-        return type(self) is type(other)      # xxx obscure
+
+
+class FakeDescr(compile.ResumeGuardDescr):
+    class rd_snapshot:
+        class prev:
+            prev = None
+            boxes = []
+        boxes = []
     def clone_if_mutable(self):
-        res = Storage(self.metainterp_sd, self.original_greenkey)
-        self.copy_all_attributes_into(res)
-        return res
-
-def _sortboxes(boxes):
-    _kind2count = {history.INT: 1, history.REF: 2, history.FLOAT: 3}
-    return sorted(boxes, key=lambda box: _kind2count[box.type])
-
-class BaseTestOptimizeOpt(BaseTest):
-    jit_ffi = False
-
-    def invent_fail_descr(self, fail_args):
-        if fail_args is None:
-            return None
-        descr = Storage()
-        descr.rd_frame_info_list = resume.FrameInfo(None, "code", 11)
-        descr.rd_snapshot = resume.Snapshot(None, _sortboxes(fail_args))
-        return descr
-
-    def assert_equal(self, optimized, expected, text_right=None):
-        assert len(optimized.inputargs) == len(expected.inputargs)
-        remap = {}
-        for box1, box2 in zip(optimized.inputargs, expected.inputargs):
-            assert box1.__class__ == box2.__class__
-            remap[box2] = box1
-        assert equaloplists(optimized.operations,
-                            expected.operations, False, remap, text_right)
-
-    def optimize_loop(self, ops, optops, expected_preamble=None,
+        return self
+
+
+class BaseTestWithUnroll(BaseTest):
+
+    enable_opts = "intbounds:rewrite:virtualize:string:heap:unroll"
+
+    def optimize_loop(self, ops, expected, expected_preamble=None,
                       call_pure_results=None):
         loop = self.parse(ops)
-        if optops != "crash!":
-            expected = self.parse(optops)
-        else:
-            expected = "crash!"
+        if expected != "crash!":
+            expected = self.parse(expected)
         if expected_preamble:
             expected_preamble = self.parse(expected_preamble)
-        #
-        self.loop = loop
-        loop.call_pure_results = args_dict()
-        if call_pure_results is not None:
-            for k, v in call_pure_results.items():
-                loop.call_pure_results[list(k)] = v
+
         loop.preamble = TreeLoop('preamble')
         loop.preamble.inputargs = loop.inputargs
         loop.preamble.token = LoopToken()
-        metainterp_sd = FakeMetaInterpStaticData(self.cpu, self.jit_ffi)
-        if hasattr(self, 'vrefinfo'):
-            metainterp_sd.virtualref_info = self.vrefinfo
-        if hasattr(self, 'callinfocollection'):
-            metainterp_sd.callinfocollection = self.callinfocollection
-        class FakeDescr(compile.ResumeGuardDescr):
-            class rd_snapshot:
-                class prev:
-                    prev = None
-                    boxes = []
-                boxes = []
-            def clone_if_mutable(self):
-                return self
         loop.preamble.start_resumedescr = FakeDescr()
-        optimize_loop_1(metainterp_sd, loop, ALL_OPTS_DICT)
         #
-
+        self._do_optimize_loop(loop, call_pure_results)
+        #
         print
         print loop.preamble.inputargs
         print '\n'.join([str(o) for o in loop.preamble.operations])
@@ -204,16 +90,14 @@
         print loop.inputargs
         print '\n'.join([str(o) for o in loop.operations])
         print
-
         assert expected != "crash!", "should have raised an exception"
         self.assert_equal(loop, expected)
         if expected_preamble:
             self.assert_equal(loop.preamble, expected_preamble,
                               text_right='expected preamble')
-
         return loop
 
-class OptimizeOptTest(BaseTestOptimizeOpt):
+class OptimizeOptTest(BaseTestWithUnroll):
 
     def setup_method(self, meth=None):
         class FailDescr(compile.ResumeGuardDescr):
@@ -1497,8 +1381,8 @@
         """
         expected = """
         [i1, p0]
+        p1 = new_array(i1, descr=arraydescr)
         setarrayitem_gc(p0, 0, i1, descr=arraydescr)
-        p1 = new_array(i1, descr=arraydescr)
         jump(i1, p1)
         """
         self.optimize_loop(ops, expected)
@@ -1922,9 +1806,9 @@
         i3 = getarrayitem_gc_pure(p3, 1, descr=arraydescr)
         i4 = getarrayitem_gc(p3, i3, descr=arraydescr)
         i5 = int_add(i3, i4)
-        setarrayitem_gc(p3, 0, i5, descr=arraydescr)
         #
         setfield_gc(p1, i2, descr=valuedescr)
+        setarrayitem_gc(p3, 0, i5, descr=arraydescr)
         setfield_gc(p1, i4, descr=nextdescr)
         escape()
         jump(p1, i1, i2, p3, i3)
@@ -1934,9 +1818,9 @@
         #
         i4 = getarrayitem_gc(p3, i3, descr=arraydescr)
         i5 = int_add(i3, i4)
-        setarrayitem_gc(p3, 0, i5, descr=arraydescr)
         #
         setfield_gc(p1, i2, descr=valuedescr)
+        setarrayitem_gc(p3, 0, i5, descr=arraydescr)
         setfield_gc(p1, i4, descr=nextdescr)
         escape()
         jump(p1, i1, i2, p3, i3)
@@ -2171,6 +2055,7 @@
         self.optimize_loop(ops, expected)
 
     def test_duplicate_getarrayitem_after_setarrayitem_2(self):
+        py.test.skip("setarrayitem with variable index")
         ops = """
         [p1, p2, p3, i1]
         setarrayitem_gc(p1, 0, p2, descr=arraydescr2)
@@ -2857,8 +2742,6 @@
 
     # ----------
 
-class TestLLtype(OptimizeOptTest, LLtypeMixin):
-
     def test_residual_call_does_not_invalidate_caches(self):
         ops = """
         [p1, p2]
@@ -5427,7 +5310,7 @@
         """
         self.optimize_strunicode_loop(ops, expected)
 
-    def test_strgetitem_small(self):
+    def test_strgetitem_bounds(self):
         ops = """
         [p0, i0]
         i1 = strgetitem(p0, i0)
@@ -5439,7 +5322,20 @@
         """
         expected = """
         [p0, i0]
-        i1 = strgetitem(p0, i0)
+        jump(p0, i0)
+        """
+        self.optimize_loop(ops, expected)
+
+    def test_unicodegetitem_bounds(self):
+        ops = """
+        [p0, i0]
+        i1 = unicodegetitem(p0, i0)
+        i2 = int_lt(i1, 0)
+        guard_false(i2) []
+        jump(p0, i0)
+        """
+        expected = """
+        [p0, i0]
         jump(p0, i0)
         """
         self.optimize_loop(ops, expected)
@@ -5953,3 +5849,54 @@
         jump(i3, i4)
         """
         self.optimize_loop(ops, expected)
+
+    def test_forced_virtual_pure_getfield(self):
+        ops = """
+        [p0]
+        p1 = getfield_gc_pure(p0, descr=valuedescr)
+        jump(p1)
+        """
+        self.optimize_loop(ops, ops)
+
+        ops = """
+        [p0]
+        p1 = new_with_vtable(ConstClass(node_vtable))
+        setfield_gc(p1, p0, descr=valuedescr)
+        escape(p1)
+        p2 = getfield_gc_pure(p1, descr=valuedescr)
+        escape(p2)
+        jump(p0)
+        """
+        expected = """
+        [p0]
+        p1 = new_with_vtable(ConstClass(node_vtable))
+        setfield_gc(p1, p0, descr=valuedescr)
+        escape(p1)
+        escape(p0)
+        jump(p0)
+        """
+        self.optimize_loop(ops, expected)
+
+    def test_setarrayitem_lazy(self):
+        ops = """
+        [i0, i1]
+        p0 = escape()
+        i2 = escape()
+        p1 = new_with_vtable(ConstClass(node_vtable))
+        setarrayitem_gc(p0, 2, p1, descr=arraydescr)
+        guard_true(i2) []
+        setarrayitem_gc(p0, 2, p0, descr=arraydescr)
+        jump(i0, i1)
+        """
+        expected = """
+        [i0, i1]
+        p0 = escape()
+        i2 = escape()
+        guard_true(i2) [p0]
+        setarrayitem_gc(p0, 2, p0, descr=arraydescr)
+        jump(i0, i1)
+        """
+        self.optimize_loop(ops, expected)
+
+class TestLLtype(OptimizeOptTest, LLtypeMixin):
+    pass
diff --git a/pypy/jit/metainterp/test/test_optimizeutil.py b/pypy/jit/metainterp/optimizeopt/test/test_util.py
rename from pypy/jit/metainterp/test/test_optimizeutil.py
rename to pypy/jit/metainterp/optimizeopt/test/test_util.py
--- a/pypy/jit/metainterp/test/test_optimizeutil.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_util.py
@@ -9,11 +9,15 @@
 from pypy.jit.metainterp.history import (BoxInt, BoxPtr, ConstInt, ConstPtr,
                                          Const, TreeLoop, BoxObj,
                                          ConstObj, AbstractDescr)
-from pypy.jit.metainterp.optimizeutil import sort_descrs, InvalidLoop
+from pypy.jit.metainterp.optimizeopt.util import sort_descrs, equaloplists
+from pypy.jit.metainterp.optimize import InvalidLoop
 from pypy.jit.codewriter.effectinfo import EffectInfo
 from pypy.jit.codewriter.heaptracker import register_known_gctype, adr2int
-from pypy.jit.tool.oparser import parse
+from pypy.jit.tool.oparser import parse, pure_parse
 from pypy.jit.metainterp.quasiimmut import QuasiImmutDescr
+from pypy.jit.metainterp import compile, resume, history
+from pypy.jit.metainterp.jitprof import EmptyProfiler
+from pypy.config.pypyoption import get_pypy_config
 
 def test_sort_descrs():
     class PseudoDescr(AbstractDescr):
@@ -28,6 +32,44 @@
         sort_descrs(lst2)
         assert lst2 == lst
 
+def test_equaloplists():
+    ops = """
+    [i0]
+    i1 = int_add(i0, 1)
+    i2 = int_add(i1, 1)
+    guard_true(i1) [i2]
+    jump(i1)
+    """
+    namespace = {}
+    loop1 = pure_parse(ops, namespace=namespace)
+    loop2 = pure_parse(ops, namespace=namespace)
+    loop3 = pure_parse(ops.replace("i2 = int_add", "i2 = int_sub"),
+                       namespace=namespace)
+    assert equaloplists(loop1.operations, loop2.operations)
+    py.test.raises(AssertionError,
+                   "equaloplists(loop1.operations, loop3.operations)")
+
+def test_equaloplists_fail_args():
+    ops = """
+    [i0]
+    i1 = int_add(i0, 1)
+    i2 = int_add(i1, 1)
+    guard_true(i1) [i2, i1]
+    jump(i1)
+    """
+    namespace = {}
+    loop1 = pure_parse(ops, namespace=namespace)
+    loop2 = pure_parse(ops.replace("[i2, i1]", "[i1, i2]"),
+                       namespace=namespace)
+    py.test.raises(AssertionError,
+                   "equaloplists(loop1.operations, loop2.operations)")
+    assert equaloplists(loop1.operations, loop2.operations,
+                        strict_fail_args=False)
+    loop3 = pure_parse(ops.replace("[i2, i1]", "[i2, i0]"),
+                       namespace=namespace)
+    py.test.raises(AssertionError,
+                   "equaloplists(loop1.operations, loop3.operations)")
+
 # ____________________________________________________________
 
 class LLtypeMixin(object):
@@ -124,19 +166,19 @@
     FUNC = lltype.FuncType([lltype.Signed], lltype.Signed)
     plaincalldescr = cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT)
     nonwritedescr = cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                                    EffectInfo([], [], []))
+                                    EffectInfo([], [], [], []))
     writeadescr = cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                                  EffectInfo([], [adescr], []))
+                                  EffectInfo([], [], [adescr], []))
     writearraydescr = cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                                      EffectInfo([], [adescr], [arraydescr]))
+                                  EffectInfo([], [], [adescr], [arraydescr]))
     readadescr = cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                                 EffectInfo([adescr], [], []))
+                                 EffectInfo([adescr], [], [], []))
     mayforcevirtdescr = cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                 EffectInfo([nextdescr], [], [],
+                 EffectInfo([nextdescr], [], [], [],
                             EffectInfo.EF_FORCES_VIRTUAL_OR_VIRTUALIZABLE,
                             can_invalidate=True))
     arraycopydescr = cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                 EffectInfo([], [], [], oopspecindex=EffectInfo.OS_ARRAYCOPY))
+             EffectInfo([], [], [], [], oopspecindex=EffectInfo.OS_ARRAYCOPY))
 
     for _name, _os in [
         ('strconcatdescr',               'OS_STR_CONCAT'),
@@ -153,15 +195,15 @@
         _oopspecindex = getattr(EffectInfo, _os)
         locals()[_name] = \
             cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                EffectInfo([], [], [], oopspecindex=_oopspecindex))
+                EffectInfo([], [], [], [], oopspecindex=_oopspecindex))
         #
         _oopspecindex = getattr(EffectInfo, _os.replace('STR', 'UNI'))
         locals()[_name.replace('str', 'unicode')] = \
             cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                EffectInfo([], [], [], oopspecindex=_oopspecindex))
+                EffectInfo([], [], [], [], oopspecindex=_oopspecindex))
 
     s2u_descr = cpu.calldescrof(FUNC, FUNC.ARGS, FUNC.RESULT,
-                EffectInfo([], [], [], oopspecindex=EffectInfo.OS_STR2UNICODE))
+            EffectInfo([], [], [], [], oopspecindex=EffectInfo.OS_STR2UNICODE))
     #
 
     class LoopToken(AbstractDescr):
@@ -256,8 +298,45 @@
 ##                       u_vtable_adr: cpu.typedescrof(U)}
 ##    namespace = locals()
 
+# ____________________________________________________________
+
+
+
+class Fake(object):
+    failargs_limit = 1000
+    storedebug = None
+
+
+class FakeMetaInterpStaticData(object):
+
+    def __init__(self, cpu):
+        self.cpu = cpu
+        self.profiler = EmptyProfiler()
+        self.options = Fake()
+        self.globaldata = Fake()
+        self.config = get_pypy_config(translating=True)
+        self.config.translation.jit_ffi = True
+
+
+class Storage(compile.ResumeGuardDescr):
+    "for tests."
+    def __init__(self, metainterp_sd=None, original_greenkey=None):
+        self.metainterp_sd = metainterp_sd
+        self.original_greenkey = original_greenkey
+    def store_final_boxes(self, op, boxes):
+        op.setfailargs(boxes)
+    def __eq__(self, other):
+        return type(self) is type(other)      # xxx obscure
+    def clone_if_mutable(self):
+        res = Storage(self.metainterp_sd, self.original_greenkey)
+        self.copy_all_attributes_into(res)
+        return res
+
+def _sortboxes(boxes):
+    _kind2count = {history.INT: 1, history.REF: 2, history.FLOAT: 3}
+    return sorted(boxes, key=lambda box: _kind2count[box.type])
+
 class BaseTest(object):
-    invent_fail_descr = None
 
     def parse(self, s, boxkinds=None):
         return parse(s, self.cpu, self.namespace,
@@ -265,5 +344,40 @@
                      boxkinds=boxkinds,
                      invent_fail_descr=self.invent_fail_descr)
 
+    def invent_fail_descr(self, model, fail_args):
+        if fail_args is None:
+            return None
+        descr = Storage()
+        descr.rd_frame_info_list = resume.FrameInfo(None, "code", 11)
+        descr.rd_snapshot = resume.Snapshot(None, _sortboxes(fail_args))
+        return descr
+
+    def assert_equal(self, optimized, expected, text_right=None):
+        from pypy.jit.metainterp.optimizeopt.util import equaloplists
+        assert len(optimized.inputargs) == len(expected.inputargs)
+        remap = {}
+        for box1, box2 in zip(optimized.inputargs, expected.inputargs):
+            assert box1.__class__ == box2.__class__
+            remap[box2] = box1
+        assert equaloplists(optimized.operations,
+                            expected.operations, False, remap, text_right)
+
+    def _do_optimize_loop(self, loop, call_pure_results):
+        from pypy.jit.metainterp.optimizeopt import optimize_loop_1
+        from pypy.jit.metainterp.optimizeopt.util import args_dict
+
+        self.loop = loop
+        loop.call_pure_results = args_dict()
+        if call_pure_results is not None:
+            for k, v in call_pure_results.items():
+                loop.call_pure_results[list(k)] = v
+        metainterp_sd = FakeMetaInterpStaticData(self.cpu)
+        if hasattr(self, 'vrefinfo'):
+            metainterp_sd.virtualref_info = self.vrefinfo
+        if hasattr(self, 'callinfocollection'):
+            metainterp_sd.callinfocollection = self.callinfocollection
+        #
+        optimize_loop_1(metainterp_sd, loop, self.enable_opts)
+
 # ____________________________________________________________
 
diff --git a/pypy/jit/metainterp/optimizeopt/unroll.py b/pypy/jit/metainterp/optimizeopt/unroll.py
--- a/pypy/jit/metainterp/optimizeopt/unroll.py
+++ b/pypy/jit/metainterp/optimizeopt/unroll.py
@@ -5,7 +5,7 @@
 from pypy.jit.metainterp.resume import Snapshot
 from pypy.jit.metainterp.history import TreeLoop, LoopToken
 from pypy.rlib.debug import debug_start, debug_stop, debug_print
-from pypy.jit.metainterp.optimizeutil import InvalidLoop, RetraceLoop
+from pypy.jit.metainterp.optimize import InvalidLoop, RetraceLoop
 from pypy.jit.metainterp.jitexc import JitException
 from pypy.jit.metainterp.history import make_hashable_int
 from pypy.jit.codewriter.effectinfo import EffectInfo
diff --git a/pypy/jit/metainterp/optimizeutil.py b/pypy/jit/metainterp/optimizeopt/util.py
rename from pypy/jit/metainterp/optimizeutil.py
rename to pypy/jit/metainterp/optimizeopt/util.py
--- a/pypy/jit/metainterp/optimizeutil.py
+++ b/pypy/jit/metainterp/optimizeopt/util.py
@@ -1,21 +1,10 @@
+import py
 from pypy.rlib.objectmodel import r_dict, compute_identity_hash
 from pypy.rlib.rarithmetic import intmask
 from pypy.rlib.unroll import unrolling_iterable
 from pypy.jit.metainterp import resoperation, history
-from pypy.jit.metainterp.jitexc import JitException
 from pypy.rlib.debug import make_sure_not_resized
-
-class InvalidLoop(JitException):
-    """Raised when the optimize*.py detect that the loop that
-    we are trying to build cannot possibly make sense as a
-    long-running loop (e.g. it cannot run 2 complete iterations)."""
-
-class RetraceLoop(JitException):
-    """ Raised when inlining a short preamble resulted in an
-        InvalidLoop. This means the optimized loop is too specialized
-        to be useful here, so we trace it again and produced a second
-        copy specialized in some different way.
-    """
+from pypy.jit.metainterp.resoperation import rop
 
 # ____________________________________________________________
 # Misc. utilities
@@ -113,3 +102,49 @@
 
 def args_dict_box():
     return r_dict(args_eq, args_hash)
+
+
+# ____________________________________________________________
+
+def equaloplists(oplist1, oplist2, strict_fail_args=True, remap={},
+                 text_right=None):
+    # try to use the full width of the terminal to display the list
+    # unfortunately, does not work with the default capture method of py.test
+    # (which is fd), you you need to use either -s or --capture=sys, else you
+    # get the standard 80 columns width
+    totwidth = py.io.get_terminal_width()
+    width = totwidth / 2 - 1
+    print ' Comparing lists '.center(totwidth, '-')
+    text_right = text_right or 'expected'
+    print '%s| %s' % ('optimized'.center(width), text_right.center(width))
+    for op1, op2 in zip(oplist1, oplist2):
+        txt1 = str(op1)
+        txt2 = str(op2)
+        while txt1 or txt2:
+            print '%s| %s' % (txt1[:width].ljust(width), txt2[:width])
+            txt1 = txt1[width:]
+            txt2 = txt2[width:]
+        assert op1.getopnum() == op2.getopnum()
+        assert op1.numargs() == op2.numargs()
+        for i in range(op1.numargs()):
+            x = op1.getarg(i)
+            y = op2.getarg(i)
+            assert x == remap.get(y, y)
+        if op2.result in remap:
+            assert op1.result == remap[op2.result]
+        else:
+            remap[op2.result] = op1.result
+        if op1.getopnum() != rop.JUMP:      # xxx obscure
+            assert op1.getdescr() == op2.getdescr()
+        if op1.getfailargs() or op2.getfailargs():
+            assert len(op1.getfailargs()) == len(op2.getfailargs())
+            if strict_fail_args:
+                for x, y in zip(op1.getfailargs(), op2.getfailargs()):
+                    assert x == remap.get(y, y)
+            else:
+                fail_args1 = set(op1.getfailargs())
+                fail_args2 = set([remap.get(y, y) for y in op2.getfailargs()])
+                assert fail_args1 == fail_args2
+    assert len(oplist1) == len(oplist2)
+    print '-'*totwidth
+    return True
diff --git a/pypy/jit/metainterp/optimizeopt/virtualize.py b/pypy/jit/metainterp/optimizeopt/virtualize.py
--- a/pypy/jit/metainterp/optimizeopt/virtualize.py
+++ b/pypy/jit/metainterp/optimizeopt/virtualize.py
@@ -1,7 +1,7 @@
 from pypy.jit.metainterp.history import Const, ConstInt, BoxInt
 from pypy.jit.metainterp.resoperation import rop, ResOperation
-from pypy.jit.metainterp.optimizeutil import _findall, sort_descrs
-from pypy.jit.metainterp.optimizeutil import descrlist_dict
+from pypy.jit.metainterp.optimizeopt.util import _findall, sort_descrs
+from pypy.jit.metainterp.optimizeopt.util import descrlist_dict
 from pypy.rlib.objectmodel import we_are_translated
 from pypy.jit.metainterp.optimizeopt import optimizer
 from pypy.jit.metainterp.executor import execute
@@ -20,6 +20,9 @@
         self.source_op = source_op  # the NEW_WITH_VTABLE/NEW_ARRAY operation
                                     # that builds this box
 
+    def is_forced_virtual(self):
+        return self.box is not None
+
     def get_key_box(self):
         if self.box is None:
             return self.keybox
@@ -120,7 +123,6 @@
                 op = ResOperation(rop.SETFIELD_GC, [box, subbox], None,
                                   descr=ofs)
                 newoperations.append(op)
-            self._fields = None
 
     def _get_field_descr_list(self):
         _cached_sorted_fields = self._cached_sorted_fields
@@ -351,7 +353,7 @@
         if not self.optimizer.cpu.ts.CONST_NULL.same_constant(objbox):
             seo(ResOperation(rop.SETFIELD_GC, op.getarglist(), None,
                              descr = vrefinfo.descr_forced))
-        
+
         # - set 'virtual_token' to TOKEN_NONE
         args = [op.getarg(0), ConstInt(vrefinfo.TOKEN_NONE)]
         seo(ResOperation(rop.SETFIELD_GC, args, None,
@@ -365,6 +367,14 @@
 
     def optimize_GETFIELD_GC(self, op):
         value = self.getvalue(op.getarg(0))
+        # If this is an immutable field (as indicated by op.is_always_pure())
+        # then it's safe to reuse the virtual's field, even if it has been
+        # forced, because it should never be written to again.
+        if value.is_forced_virtual() and op.is_always_pure():
+            fieldvalue = value.getfield(op.getdescr(), None)
+            if fieldvalue is not None:
+                self.make_equal_to(op.result, fieldvalue)
+                return
         if value.is_virtual():
             assert isinstance(value, AbstractVirtualValue)
             fieldvalue = value.getfield(op.getdescr(), None)
@@ -382,6 +392,7 @@
 
     def optimize_SETFIELD_GC(self, op):
         value = self.getvalue(op.getarg(0))
+
         if value.is_virtual():
             fieldvalue = self.getvalue(op.getarg(1))
             value.setfield(op.getdescr(), fieldvalue)
diff --git a/pypy/jit/metainterp/pyjitpl.py b/pypy/jit/metainterp/pyjitpl.py
--- a/pypy/jit/metainterp/pyjitpl.py
+++ b/pypy/jit/metainterp/pyjitpl.py
@@ -1,5 +1,5 @@
-import py, os, sys
-from pypy.rpython.lltypesystem import lltype, llmemory, rclass
+import py, sys
+from pypy.rpython.lltypesystem import lltype, rclass
 from pypy.rlib.objectmodel import we_are_translated
 from pypy.rlib.unroll import unrolling_iterable
 from pypy.rlib.debug import debug_start, debug_stop, debug_print
@@ -15,13 +15,13 @@
 from pypy.jit.metainterp.jitprof import EmptyProfiler
 from pypy.jit.metainterp.jitprof import GUARDS, RECORDED_OPS, ABORT_ESCAPE
 from pypy.jit.metainterp.jitprof import ABORT_TOO_LONG, ABORT_BRIDGE, \
-                                        ABORT_BAD_LOOP, ABORT_FORCE_QUASIIMMUT
+                                        ABORT_FORCE_QUASIIMMUT
 from pypy.jit.metainterp.jitexc import JitException, get_llexception
-from pypy.rlib.rarithmetic import intmask
 from pypy.rlib.objectmodel import specialize
-from pypy.jit.codewriter.jitcode import JitCode, SwitchDictDescr, MissingLiveness
-from pypy.jit.codewriter import heaptracker, longlong
-from pypy.jit.metainterp.optimizeutil import RetraceLoop, args_dict_box, args_dict
+from pypy.jit.codewriter.jitcode import JitCode, SwitchDictDescr
+from pypy.jit.codewriter import heaptracker
+from pypy.jit.metainterp.optimizeopt.util import args_dict_box
+from pypy.jit.metainterp.optimize import RetraceLoop
 
 # ____________________________________________________________
 
@@ -867,7 +867,7 @@
         any_operation = len(self.metainterp.history.operations) > 0
         jitdriver_sd = self.metainterp.staticdata.jitdrivers_sd[jdindex]
         self.verify_green_args(jitdriver_sd, greenboxes)
-        self.debug_merge_point(jdindex, self.metainterp.in_recursion,
+        self.debug_merge_point(jitdriver_sd, jdindex, self.metainterp.in_recursion,
                                greenboxes)
 
         if self.metainterp.seen_loop_header_for_jdindex < 0:
@@ -914,8 +914,10 @@
                                     assembler_call=True)
             raise ChangeFrame
 
-    def debug_merge_point(self, jd_index, in_recursion, greenkey):
+    def debug_merge_point(self, jitdriver_sd, jd_index, in_recursion, greenkey):
         # debugging: produce a DEBUG_MERGE_POINT operation
+        loc = jitdriver_sd.warmstate.get_location_str(greenkey)
+        debug_print(loc)
         args = [ConstInt(jd_index), ConstInt(in_recursion)] + greenkey
         self.metainterp.history.record(rop.DEBUG_MERGE_POINT, args, None)
 
@@ -1262,8 +1264,7 @@
     logger_ops = None
 
     def __init__(self, cpu, options,
-                 ProfilerClass=EmptyProfiler, warmrunnerdesc=None,
-                 jit_ffi=True):
+                 ProfilerClass=EmptyProfiler, warmrunnerdesc=None):
         self.cpu = cpu
         self.stats = self.cpu.stats
         self.options = options
@@ -1273,7 +1274,11 @@
         self.profiler = ProfilerClass()
         self.profiler.cpu = cpu
         self.warmrunnerdesc = warmrunnerdesc
-        self.jit_ffi = jit_ffi
+        if warmrunnerdesc:
+            self.config = warmrunnerdesc.translator.config
+        else:
+            from pypy.config.pypyoption import get_pypy_config
+            self.config = get_pypy_config(translating=True)
 
         backendmodule = self.cpu.__module__
         backendmodule = backendmodule.split('.')[-2]
@@ -1924,7 +1929,6 @@
 
         self.history.inputargs = original_inputargs
         self.history.operations.pop()     # remove the JUMP
-        # FIXME: Why is self.history.inputargs not restored?
 
     def compile_bridge(self, live_arg_boxes):
         num_green_args = self.jitdriver_sd.num_green_args
@@ -1960,6 +1964,8 @@
                                               start_resumedescr, False)
         self.history.operations.pop()     # remove the JUMP
         if loop_token is None:
+            self.history.inputargs = original_inputargs
+            self.history.operations = original_operations
             return
 
         if loop_token.short_preamble:
@@ -2114,7 +2120,6 @@
     def vrefs_after_residual_call(self):
         vrefinfo = self.staticdata.virtualref_info
         for i in range(0, len(self.virtualref_boxes), 2):
-            virtualbox = self.virtualref_boxes[i]
             vrefbox = self.virtualref_boxes[i+1]
             vref = vrefbox.getref_base()
             if vrefinfo.tracing_after_residual_call(vref):
diff --git a/pypy/jit/metainterp/resoperation.py b/pypy/jit/metainterp/resoperation.py
--- a/pypy/jit/metainterp/resoperation.py
+++ b/pypy/jit/metainterp/resoperation.py
@@ -191,9 +191,15 @@
         # of the operation.  It must inherit from AbstractDescr.  The
         # backend provides it with cpu.fielddescrof(), cpu.arraydescrof(),
         # cpu.calldescrof(), and cpu.typedescrof().
+        self._check_descr(descr)
+        self._descr = descr
+
+    def _check_descr(self, descr):
+        if not we_are_translated() and getattr(descr, 'I_am_a_descr', False):
+            return # needed for the mock case in oparser_model
         from pypy.jit.metainterp.history import check_descr
         check_descr(descr)
-        self._descr = descr
+
 
 class GuardResOp(ResOpWithDescr):
 
@@ -471,8 +477,8 @@
     'STRSETITEM/3',
     'UNICODESETITEM/3',
     #'RUNTIMENEW/1',     # ootype operation
-    'COND_CALL_GC_WB/2d', # [objptr, newvalue] or [arrayptr, index]
-                          # (for the write barrier, latter is in an array)
+    'COND_CALL_GC_WB/2d', # [objptr, newvalue] (for the write barrier)
+    'COND_CALL_GC_WB_ARRAY/3d', # [objptr, arrayindex, newvalue] (write barr.)
     'DEBUG_MERGE_POINT/*',      # debugging only
     'JIT_DEBUG/*',              # debugging only
     'VIRTUAL_REF_FINISH/2',   # removed before it's passed to the backend
diff --git a/pypy/jit/metainterp/resume.py b/pypy/jit/metainterp/resume.py
--- a/pypy/jit/metainterp/resume.py
+++ b/pypy/jit/metainterp/resume.py
@@ -2,15 +2,17 @@
 from pypy.jit.metainterp.history import Box, Const, ConstInt, getkind
 from pypy.jit.metainterp.history import BoxInt, BoxPtr, BoxFloat
 from pypy.jit.metainterp.history import INT, REF, FLOAT, HOLE
+from pypy.jit.metainterp.history import AbstractDescr
 from pypy.jit.metainterp.resoperation import rop
 from pypy.jit.metainterp import jitprof
 from pypy.jit.codewriter.effectinfo import EffectInfo
 from pypy.rpython.lltypesystem import lltype, llmemory, rffi, rstr
+from pypy.rpython import annlowlevel
 from pypy.rlib import rarithmetic, rstack
 from pypy.rlib.objectmodel import we_are_translated, specialize
 from pypy.rlib.debug import have_debug_prints, ll_assert
 from pypy.rlib.debug import debug_start, debug_stop, debug_print
-from pypy.jit.metainterp.optimizeutil import InvalidLoop
+from pypy.jit.metainterp.optimize import InvalidLoop
 
 # Logic to encode the chain of frames and the state of the boxes at a
 # guard operation, and to decode it again.  This is a bit advanced,
@@ -82,6 +84,13 @@
                             ('nums', lltype.Array(rffi.SHORT)))
 NUMBERINGP.TO.become(NUMBERING)
 
+PENDINGFIELDSTRUCT = lltype.Struct('PendingField',
+                                   ('lldescr', annlowlevel.base_ptr_lltype()),
+                                   ('num', rffi.SHORT),
+                                   ('fieldnum', rffi.SHORT),
+                                   ('itemindex', rffi.INT))
+PENDINGFIELDSP = lltype.Ptr(lltype.GcArray(PENDINGFIELDSTRUCT))
+
 TAGMASK = 3
 
 def tag(value, tagbits):
@@ -329,7 +338,7 @@
                 value = values[box]
                 value.get_args_for_fail(self)
 
-        for _, box, fieldbox in pending_setfields:
+        for _, box, fieldbox, _ in pending_setfields:
             self.register_box(box)
             self.register_box(fieldbox)
             value = values[fieldbox]
@@ -405,13 +414,25 @@
         return False
 
     def _add_pending_fields(self, pending_setfields):
-        rd_pendingfields = None
+        rd_pendingfields = lltype.nullptr(PENDINGFIELDSP.TO)
         if pending_setfields:
-            rd_pendingfields = []
-            for descr, box, fieldbox in pending_setfields:
+            n = len(pending_setfields)
+            rd_pendingfields = lltype.malloc(PENDINGFIELDSP.TO, n)
+            for i in range(n):
+                descr, box, fieldbox, itemindex = pending_setfields[i]
+                lldescr = annlowlevel.cast_instance_to_base_ptr(descr)
                 num = self._gettagged(box)
                 fieldnum = self._gettagged(fieldbox)
-                rd_pendingfields.append((descr, num, fieldnum))
+                # the index is limited to 2147483647 (64-bit machines only)
+                if itemindex > 2147483647:
+                    from pypy.jit.metainterp import compile
+                    compile.giveup()
+                itemindex = rffi.cast(rffi.INT, itemindex)
+                #
+                rd_pendingfields[i].lldescr  = lldescr
+                rd_pendingfields[i].num      = num
+                rd_pendingfields[i].fieldnum = fieldnum
+                rd_pendingfields[i].itemindex= itemindex
         self.storage.rd_pendingfields = rd_pendingfields
 
     def _gettagged(self, box):
@@ -727,10 +748,28 @@
             self.virtuals_cache = [self.virtual_default] * len(virtuals)
 
     def _prepare_pendingfields(self, pendingfields):
-        if pendingfields is not None:
-            for descr, num, fieldnum in pendingfields:
+        if pendingfields:
+            for i in range(len(pendingfields)):
+                lldescr  = pendingfields[i].lldescr
+                num      = pendingfields[i].num
+                fieldnum = pendingfields[i].fieldnum
+                itemindex= pendingfields[i].itemindex
+                descr = annlowlevel.cast_base_ptr_to_instance(AbstractDescr,
+                                                              lldescr)
                 struct = self.decode_ref(num)
-                self.setfield(descr, struct, fieldnum)
+                itemindex = rffi.cast(lltype.Signed, itemindex)
+                if itemindex < 0:
+                    self.setfield(descr, struct, fieldnum)
+                else:
+                    self.setarrayitem(descr, struct, itemindex, fieldnum)
+
+    def setarrayitem(self, arraydescr, array, index, fieldnum):
+        if arraydescr.is_array_of_pointers():
+            self.setarrayitem_ref(arraydescr, array, index, fieldnum)
+        elif arraydescr.is_array_of_floats():
+            self.setarrayitem_float(arraydescr, array, index, fieldnum)
+        else:
+            self.setarrayitem_int(arraydescr, array, index, fieldnum)
 
     def _prepare_next_section(self, info):
         # Use info.enumerate_vars(), normally dispatching to
@@ -903,15 +942,15 @@
                                            structbox, fieldbox)
 
     def setarrayitem_int(self, arraydescr, arraybox, index, fieldnum):
-        self.setarrayitem(arraydescr, arraybox, index, fieldnum, INT)
+        self._setarrayitem(arraydescr, arraybox, index, fieldnum, INT)
 
     def setarrayitem_ref(self, arraydescr, arraybox, index, fieldnum):
-        self.setarrayitem(arraydescr, arraybox, index, fieldnum, REF)
+        self._setarrayitem(arraydescr, arraybox, index, fieldnum, REF)
 
     def setarrayitem_float(self, arraydescr, arraybox, index, fieldnum):
-        self.setarrayitem(arraydescr, arraybox, index, fieldnum, FLOAT)
+        self._setarrayitem(arraydescr, arraybox, index, fieldnum, FLOAT)
 
-    def setarrayitem(self, arraydescr, arraybox, index, fieldnum, kind):
+    def _setarrayitem(self, arraydescr, arraybox, index, fieldnum, kind):
         itembox = self.decode_box(fieldnum, kind)
         self.metainterp.execute_and_record(rop.SETARRAYITEM_GC,
                                            arraydescr, arraybox,
diff --git a/pypy/jit/metainterp/test/test_ajit.py b/pypy/jit/metainterp/test/test_ajit.py
--- a/pypy/jit/metainterp/test/test_ajit.py
+++ b/pypy/jit/metainterp/test/test_ajit.py
@@ -500,7 +500,7 @@
                 y -= x
             return y
         #
-        res = self.meta_interp(f, [3, 6], repeat=7)
+        res = self.meta_interp(f, [3, 6], repeat=7, function_threshold=0)
         assert res == 6 - 4 - 5
         self.check_history(call=0)   # because the trace starts in the middle
         #
@@ -1677,6 +1677,8 @@
         res = self.meta_interp(g, [6, 14])
         assert res == g(6, 14)
         self.check_loop_count(9)
+        self.check_loops(getarrayitem_gc=8, everywhere=True)
+        py.test.skip("for the following, we need setarrayitem(varindex)")
         self.check_loops(getarrayitem_gc=6, everywhere=True)
 
     def test_multiple_specialied_versions_bridge(self):
@@ -2230,6 +2232,87 @@
         self.check_loops(getfield_gc_pure=0)
         self.check_loops(getfield_gc_pure=2, everywhere=True)
         
+    def test_frame_finished_during_retrace(self):
+        class Base(object):
+            pass
+        class A(Base):
+            def __init__(self, a):
+                self.val = a
+                self.num = 1
+            def inc(self):
+                return A(self.val + 1)
+        class B(Base):
+            def __init__(self, a):
+                self.val = a
+                self.num = 1000
+            def inc(self):
+                return B(self.val + 1)
+        myjitdriver = JitDriver(greens = [], reds = ['sa', 'a'])
+        def f():
+            myjitdriver.set_param('threshold', 3)
+            myjitdriver.set_param('trace_eagerness', 2)
+            a = A(0)
+            sa = 0
+            while a.val < 8:
+                myjitdriver.jit_merge_point(a=a, sa=sa)
+                a = a.inc()
+                if a.val > 4:
+                    a = B(a.val)
+                sa += a.num
+            return sa
+        res = self.meta_interp(f, [])
+        assert res == f()
+        
+    def test_frame_finished_during_continued_retrace(self):
+        class Base(object):
+            pass
+        class A(Base):
+            def __init__(self, a):
+                self.val = a
+                self.num = 100
+            def inc(self):
+                return A(self.val + 1)
+        class B(Base):
+            def __init__(self, a):
+                self.val = a
+                self.num = 10000
+            def inc(self):
+                return B(self.val + 1)
+        myjitdriver = JitDriver(greens = [], reds = ['sa', 'b', 'a'])
+        def f(b):
+            myjitdriver.set_param('threshold', 6)
+            myjitdriver.set_param('trace_eagerness', 4)
+            a = A(0)
+            sa = 0
+            while a.val < 15:
+                myjitdriver.jit_merge_point(a=a, b=b, sa=sa)
+                a = a.inc()
+                if a.val > 8:
+                    a = B(a.val)
+                if b == 1:
+                    b = 2
+                else:
+                    b = 1
+                sa += a.num + b
+            return sa
+        res = self.meta_interp(f, [1])
+        assert res == f(1)
+
+    def test_remove_array_operations(self):
+        myjitdriver = JitDriver(greens = [], reds = ['a'])
+        class W_Int:
+            def __init__(self, intvalue):
+                self.intvalue = intvalue
+        def f(x):
+            a = [W_Int(x)]
+            while a[0].intvalue > 0:
+                myjitdriver.jit_merge_point(a=a)
+                a[0] = W_Int(a[0].intvalue - 3)
+            return a[0].intvalue
+        res = self.meta_interp(f, [100])
+        assert res == -2
+        #self.check_loops(getarrayitem_gc=0, setarrayitem_gc=0) -- xxx?
+
 class TestOOtype(BasicTests, OOJitMixin):
 
     def test_oohash(self):
diff --git a/pypy/jit/metainterp/test/test_compile.py b/pypy/jit/metainterp/test/test_compile.py
--- a/pypy/jit/metainterp/test/test_compile.py
+++ b/pypy/jit/metainterp/test/test_compile.py
@@ -1,3 +1,4 @@
+from pypy.config.pypyoption import get_pypy_config
 from pypy.jit.metainterp.history import LoopToken, ConstInt, History, Stats
 from pypy.jit.metainterp.history import BoxInt, INT
 from pypy.jit.metainterp.compile import insert_loop_token, compile_new_loop
@@ -5,7 +6,7 @@
 from pypy.jit.metainterp.compile import ResumeGuardCountersInt
 from pypy.jit.metainterp.compile import compile_tmp_callback
 from pypy.jit.metainterp import jitprof, typesystem, compile
-from pypy.jit.metainterp.test.test_optimizeutil import LLtypeMixin
+from pypy.jit.metainterp.optimizeopt.test.test_util import LLtypeMixin
 from pypy.jit.tool.oparser import parse
 from pypy.jit.metainterp.optimizeopt import ALL_OPTS_DICT
 
@@ -57,11 +58,11 @@
     
     logger_noopt = FakeLogger()
     logger_ops = FakeLogger()
+    config = get_pypy_config(translating=True)
 
     stats = Stats()
     profiler = jitprof.EmptyProfiler()
     warmrunnerdesc = None
-    jit_ffi = False
     def log(self, msg, event_kind=None):
         pass
 
diff --git a/pypy/jit/metainterp/test/test_dict.py b/pypy/jit/metainterp/test/test_dict.py
--- a/pypy/jit/metainterp/test/test_dict.py
+++ b/pypy/jit/metainterp/test/test_dict.py
@@ -130,6 +130,38 @@
         assert res == 50
         self.check_loops(int_mod=1)
 
+    def test_repeated_lookup(self):
+        myjitdriver = JitDriver(greens = [], reds = ['n', 'd'])
+        class Wrapper(object):
+            _immutable_fields_ = ["value"]
+            def __init__(self, value):
+                self.value = value
+        def eq_func(a, b):
+            return a.value == b.value
+        def hash_func(x):
+            return objectmodel.compute_hash(x.value)
+
+        def f(n):
+            d = None
+            while n > 0:
+                myjitdriver.jit_merge_point(n=n, d=d)
+                d = objectmodel.r_dict(eq_func, hash_func)
+                y = Wrapper(str(n))
+                d[y] = n - 1
+                n = d[y]
+            return d[Wrapper(str(n + 1))]
+
+        res = self.meta_interp(f, [100], listops=True)
+        assert res == f(50)
+        # XXX: ideally there would be 7 calls here, but repeated CALL_PURE with
+        # the same arguments are not folded, because we have conflicting
+        # definitions of pure, once strhash can be appropriately folded
+        # this should be decreased to seven.
+        self.check_loops({"call": 8, "guard_false": 1, "guard_no_exception": 5,
+                          "guard_true": 1, "int_and": 1, "int_gt": 1,
+                          "int_is_true": 1, "int_sub": 1, "jump": 1,
+                          "new_with_vtable": 1, "setfield_gc": 1})
+
 
 class TestOOtype(DictTests, OOJitMixin):
     pass
diff --git a/pypy/jit/metainterp/test/test_jitdriver.py b/pypy/jit/metainterp/test/test_jitdriver.py
--- a/pypy/jit/metainterp/test/test_jitdriver.py
+++ b/pypy/jit/metainterp/test/test_jitdriver.py
@@ -113,6 +113,7 @@
             return n
         #
         def loop2(g, r):
+            myjitdriver1.set_param('function_threshold', 0)
             while r > 0:
                 myjitdriver2.can_enter_jit(g=g, r=r)
                 myjitdriver2.jit_merge_point(g=g, r=r)
diff --git a/pypy/jit/metainterp/test/test_list.py b/pypy/jit/metainterp/test/test_list.py
--- a/pypy/jit/metainterp/test/test_list.py
+++ b/pypy/jit/metainterp/test/test_list.py
@@ -49,7 +49,7 @@
                 x = l[n]
                 l = [3] * 100
                 l[3] = x
-                l[3] = x + 1
+                l[4] = x + 1
                 n -= 1
             return l[0]
 
diff --git a/pypy/jit/metainterp/test/test_logger.py b/pypy/jit/metainterp/test/test_logger.py
--- a/pypy/jit/metainterp/test/test_logger.py
+++ b/pypy/jit/metainterp/test/test_logger.py
@@ -4,7 +4,7 @@
 from pypy.jit.metainterp import logger
 from pypy.jit.metainterp.typesystem import llhelper
 from StringIO import StringIO
-from pypy.jit.metainterp.test.test_optimizeopt import equaloplists
+from pypy.jit.metainterp.optimizeopt.util import equaloplists
 from pypy.jit.metainterp.history import AbstractDescr, LoopToken, BasicFailDescr
 from pypy.jit.backend.model import AbstractCPU
 
@@ -53,7 +53,7 @@
     def make_metainterp_sd(self):
         class FakeJitDriver(object):
             class warmstate(object):
-                get_location_str = staticmethod(lambda args: args[0]._get_str())
+                get_location_str = staticmethod(lambda args: "dupa")
         
         class FakeMetaInterpSd:
             cpu = AbstractCPU()
@@ -116,10 +116,10 @@
     def test_debug_merge_point(self):
         inp = '''
         []
-        debug_merge_point(0, 0, "dupa")
+        debug_merge_point(0, 0)
         '''
         _, loop, oloop = self.reparse(inp, check_equal=False)
-        assert loop.operations[0].getarg(2)._get_str() == "dupa"
+        assert loop.operations[0].getarg(1).getint() == 0
         assert oloop.operations[0].getarg(1)._get_str() == "dupa"
         
     def test_floats(self):
diff --git a/pypy/jit/metainterp/test/test_pyjitpl.py b/pypy/jit/metainterp/test/test_pyjitpl.py
--- a/pypy/jit/metainterp/test/test_pyjitpl.py
+++ b/pypy/jit/metainterp/test/test_pyjitpl.py
@@ -6,7 +6,7 @@
 from pypy.jit.metainterp.history import BoxInt, ConstInt
 from pypy.jit.metainterp.history import History
 from pypy.jit.metainterp.resoperation import ResOperation, rop
-from pypy.jit.metainterp.test.test_optimizeopt import equaloplists
+from pypy.jit.metainterp.optimizeopt.util import equaloplists
 from pypy.jit.codewriter.jitcode import JitCode
 
 
diff --git a/pypy/jit/metainterp/test/test_recursive.py b/pypy/jit/metainterp/test/test_recursive.py
--- a/pypy/jit/metainterp/test/test_recursive.py
+++ b/pypy/jit/metainterp/test/test_recursive.py
@@ -483,6 +483,7 @@
  
         def main(inline):
             myjitdriver.set_param("threshold", 10)
+            myjitdriver.set_param('function_threshold', 60)
             if inline:
                 myjitdriver.set_param('inlining', True)
             else:
@@ -1193,6 +1194,51 @@
                 i -= 1
         self.meta_interp(portal, [0, 10], inline=True)
 
+    def test_trace_from_start_always(self):
+        from pypy.rlib.nonconst import NonConstant
+        
+        driver = JitDriver(greens = ['c'], reds = ['i', 'v'])
+
+        def portal(c, i, v):
+            while i > 0:
+                driver.jit_merge_point(c=c, i=i, v=v)
+                portal(c, i - 1, v)
+                if v:
+                    driver.can_enter_jit(c=c, i=i, v=v)
+                break
+
+        def main(c, i, set_param, v):
+            if set_param:
+                driver.set_param('function_threshold', 0)
+            portal(c, i, v)
+
+        self.meta_interp(main, [10, 10, False, False], inline=True)
+        self.check_tree_loop_count(1)
+        self.check_loop_count(0)
+        self.meta_interp(main, [3, 10, True, False], inline=True)
+        self.check_tree_loop_count(0)
+        self.check_loop_count(0)
+
+    def test_trace_from_start_does_not_prevent_inlining(self):
+        driver = JitDriver(greens = ['c', 'bc'], reds = ['i'])
+        
+        def portal(bc, c, i):
+            while True:
+                driver.jit_merge_point(c=c, bc=bc, i=i)
+                if bc == 0:
+                    portal(1, 8, 0)
+                    c += 1
+                else:
+                    return
+                if c == 10: # bc == 0                    
+                    c = 0
+                    if i >= 100:
+                        return
+                    driver.can_enter_jit(c=c, bc=bc, i=i)
+                i += 1
+
+        self.meta_interp(portal, [0, 0, 0], inline=True)
+        self.check_loops(call=0, call_may_force=0)
 
 class TestLLtype(RecursiveTests, LLJitMixin):
     pass
diff --git a/pypy/jit/metainterp/test/test_resume.py b/pypy/jit/metainterp/test/test_resume.py
--- a/pypy/jit/metainterp/test/test_resume.py
+++ b/pypy/jit/metainterp/test/test_resume.py
@@ -6,7 +6,7 @@
 from pypy.jit.metainterp.resume import *
 from pypy.jit.metainterp.history import BoxInt, BoxPtr, ConstInt
 from pypy.jit.metainterp.history import ConstPtr, ConstFloat
-from pypy.jit.metainterp.test.test_optimizeutil import LLtypeMixin
+from pypy.jit.metainterp.optimizeopt.test.test_util import LLtypeMixin
 from pypy.jit.metainterp import executor
 from pypy.jit.codewriter import heaptracker, longlong
 
@@ -1238,7 +1238,7 @@
     liveboxes = []
     modifier._number_virtuals(liveboxes, values, 0)
     assert liveboxes == [b2s, b4s] or liveboxes == [b4s, b2s]
-    modifier._add_pending_fields([(LLtypeMixin.nextdescr, b2s, b4s)])
+    modifier._add_pending_fields([(LLtypeMixin.nextdescr, b2s, b4s, -1)])
     storage.rd_consts = memo.consts[:]
     storage.rd_numb = None
     # resume
@@ -1259,6 +1259,106 @@
     assert len(expected) == len(trace)
     assert demo55.next == demo66
 
+def test_virtual_adder_pending_fields_and_arrayitems():
+    class Storage(object):
+        pass
+    storage = Storage()
+    modifier = ResumeDataVirtualAdder(storage, None)
+    modifier._add_pending_fields([])
+    assert not storage.rd_pendingfields
+    #
+    class FieldDescr(object):
+        pass
+    field_a = FieldDescr()
+    storage = Storage()
+    modifier = ResumeDataVirtualAdder(storage, None)
+    modifier.liveboxes_from_env = {42: rffi.cast(rffi.SHORT, 1042),
+                                   61: rffi.cast(rffi.SHORT, 1061)}
+    modifier._add_pending_fields([(field_a, 42, 61, -1)])
+    pf = storage.rd_pendingfields
+    assert len(pf) == 1
+    assert (annlowlevel.cast_base_ptr_to_instance(FieldDescr, pf[0].lldescr)
+            is field_a)
+    assert rffi.cast(lltype.Signed, pf[0].num) == 1042
+    assert rffi.cast(lltype.Signed, pf[0].fieldnum) == 1061
+    assert rffi.cast(lltype.Signed, pf[0].itemindex) == -1
+    #
+    array_a = FieldDescr()
+    storage = Storage()
+    modifier = ResumeDataVirtualAdder(storage, None)
+    modifier.liveboxes_from_env = {42: rffi.cast(rffi.SHORT, 1042),
+                                   61: rffi.cast(rffi.SHORT, 1061),
+                                   62: rffi.cast(rffi.SHORT, 1062),
+                                   63: rffi.cast(rffi.SHORT, 1063)}
+    modifier._add_pending_fields([(array_a, 42, 61, 0),
+                                  (array_a, 42, 62, 2147483647)])
+    pf = storage.rd_pendingfields
+    assert len(pf) == 2
+    assert (annlowlevel.cast_base_ptr_to_instance(FieldDescr, pf[0].lldescr)
+            is array_a)
+    assert rffi.cast(lltype.Signed, pf[0].num) == 1042
+    assert rffi.cast(lltype.Signed, pf[0].fieldnum) == 1061
+    assert rffi.cast(lltype.Signed, pf[0].itemindex) == 0
+    assert (annlowlevel.cast_base_ptr_to_instance(FieldDescr, pf[1].lldescr)
+            is array_a)
+    assert rffi.cast(lltype.Signed, pf[1].num) == 1042
+    assert rffi.cast(lltype.Signed, pf[1].fieldnum) == 1062
+    assert rffi.cast(lltype.Signed, pf[1].itemindex) == 2147483647
+    #
+    from pypy.jit.metainterp.pyjitpl import SwitchToBlackhole
+    py.test.raises(SwitchToBlackhole, modifier._add_pending_fields,
+                   [(array_a, 42, 63, 2147483648)])
+
+def test_resume_reader_fields_and_arrayitems():
+    class ResumeReader(AbstractResumeDataReader):
+        def __init__(self, got=None, got_array=None):
+            self.got = got
+            self.got_array = got_array
+        def setfield(self, descr, struct, fieldnum):
+            assert lltype.typeOf(struct) is lltype.Signed
+            assert lltype.typeOf(fieldnum) is rffi.SHORT
+            fieldnum = rffi.cast(lltype.Signed, fieldnum)
+            self.got.append((descr, struct, fieldnum))
+        def setarrayitem(self, arraydescr, array, index, fieldnum):
+            assert lltype.typeOf(array) is lltype.Signed
+            assert lltype.typeOf(index) is lltype.Signed
+            assert lltype.typeOf(fieldnum) is rffi.SHORT
+            fieldnum = rffi.cast(lltype.Signed, fieldnum)
+            self.got_array.append((arraydescr, array, index, fieldnum))
+        def decode_ref(self, num):
+            return rffi.cast(lltype.Signed, num) * 100
+    got = []
+    pf = lltype.nullptr(PENDINGFIELDSP.TO)
+    ResumeReader(got)._prepare_pendingfields(pf)
+    assert got == []
+    #
+    class FieldDescr(AbstractDescr):
+        pass
+    field_a = FieldDescr()
+    field_b = FieldDescr()
+    pf = lltype.malloc(PENDINGFIELDSP.TO, 2)
+    pf[0].lldescr = annlowlevel.cast_instance_to_base_ptr(field_a)
+    pf[0].num = rffi.cast(rffi.SHORT, 1042)
+    pf[0].fieldnum = rffi.cast(rffi.SHORT, 1061)
+    pf[0].itemindex = rffi.cast(rffi.INT, -1)
+    pf[1].lldescr = annlowlevel.cast_instance_to_base_ptr(field_b)
+    pf[1].num = rffi.cast(rffi.SHORT, 2042)
+    pf[1].fieldnum = rffi.cast(rffi.SHORT, 2061)
+    pf[1].itemindex = rffi.cast(rffi.INT, -1)
+    got = []
+    ResumeReader(got)._prepare_pendingfields(pf)
+    assert got == [(field_a, 104200, 1061), (field_b, 204200, 2061)]
+    #
+    array_a = FieldDescr()
+    pf = lltype.malloc(PENDINGFIELDSP.TO, 1)
+    pf[0].lldescr = annlowlevel.cast_instance_to_base_ptr(array_a)
+    pf[0].num = rffi.cast(rffi.SHORT, 1042)
+    pf[0].fieldnum = rffi.cast(rffi.SHORT, 1063)
+    pf[0].itemindex = rffi.cast(rffi.INT, 123)
+    got_array = []
+    ResumeReader(got_array=got_array)._prepare_pendingfields(pf)
+    assert got_array == [(array_a, 104200, 123, 1063)]
+
 
 def test_invalidation_needed():
     class options:
diff --git a/pypy/jit/metainterp/test/test_virtualizable.py b/pypy/jit/metainterp/test/test_virtualizable.py
--- a/pypy/jit/metainterp/test/test_virtualizable.py
+++ b/pypy/jit/metainterp/test/test_virtualizable.py
@@ -11,7 +11,7 @@
 from pypy.rpython.rclass import FieldListAccessor
 from pypy.jit.metainterp.warmspot import get_stats, get_translator
 from pypy.jit.metainterp import history
-from pypy.jit.metainterp.test.test_optimizeutil import LLtypeMixin
+from pypy.jit.metainterp.optimizeopt.test.test_util import LLtypeMixin
 
 def promote_virtualizable(*args):
     pass
diff --git a/pypy/jit/metainterp/virtualref.py b/pypy/jit/metainterp/virtualref.py
--- a/pypy/jit/metainterp/virtualref.py
+++ b/pypy/jit/metainterp/virtualref.py
@@ -1,5 +1,5 @@
 from pypy.rpython.rmodel import inputconst, log
-from pypy.rpython.lltypesystem import lltype, llmemory, rffi, rclass
+from pypy.rpython.lltypesystem import lltype, llmemory, rclass
 from pypy.jit.metainterp import history
 from pypy.jit.codewriter import heaptracker
 from pypy.rlib.jit import InvalidVirtualRef
diff --git a/pypy/jit/metainterp/warmspot.py b/pypy/jit/metainterp/warmspot.py
--- a/pypy/jit/metainterp/warmspot.py
+++ b/pypy/jit/metainterp/warmspot.py
@@ -1,6 +1,5 @@
 import sys, py
-from pypy.rpython.lltypesystem import lltype, llmemory, rclass, rstr
-from pypy.rpython.ootypesystem import ootype
+from pypy.rpython.lltypesystem import lltype, llmemory
 from pypy.rpython.annlowlevel import llhelper, MixLevelHelperAnnotator,\
      cast_base_ptr_to_instance, hlstr
 from pypy.annotation import model as annmodel
@@ -10,16 +9,12 @@
 from pypy.objspace.flow.model import checkgraph, Link, copygraph
 from pypy.rlib.objectmodel import we_are_translated
 from pypy.rlib.unroll import unrolling_iterable
-from pypy.rlib.rarithmetic import r_uint, intmask
-from pypy.rlib.debug import debug_print, fatalerror
-from pypy.rlib.debug import debug_start, debug_stop
-from pypy.rpython.lltypesystem.lloperation import llop
-from pypy.translator.simplify import get_funcobj, get_functype
+from pypy.rlib.debug import fatalerror
+from pypy.translator.simplify import get_functype
 from pypy.translator.unsimplify import call_final_function
 
 from pypy.jit.metainterp import history, pyjitpl, gc, memmgr
-from pypy.jit.metainterp.pyjitpl import MetaInterpStaticData, MetaInterp
-from pypy.jit.metainterp.typesystem import LLTypeHelper, OOTypeHelper
+from pypy.jit.metainterp.pyjitpl import MetaInterpStaticData
 from pypy.jit.metainterp.jitprof import Profiler, EmptyProfiler
 from pypy.jit.metainterp.jitexc import JitException
 from pypy.jit.metainterp.jitdriver import JitDriverStaticData
@@ -66,6 +61,7 @@
 def jittify_and_run(interp, graph, args, repeat=1,
                     backendopt=False, trace_limit=sys.maxint,
                     inline=False, loop_longevity=0, retrace_limit=5,
+                    function_threshold=4,
                     enable_opts=ALL_OPTS_NAMES, **kwds):
     from pypy.config.config import ConfigError
     translator = interp.typer.annotator.translator
@@ -77,9 +73,14 @@
         translator.config.translation.list_comprehension_operations = True
     except ConfigError:
         pass
+    try:
+        translator.config.translation.jit_ffi = True
+    except ConfigError:
+        pass
     warmrunnerdesc = WarmRunnerDesc(translator, backendopt=backendopt, **kwds)
     for jd in warmrunnerdesc.jitdrivers_sd:
         jd.warmstate.set_param_threshold(3)          # for tests
+        jd.warmstate.set_param_function_threshold(function_threshold)
         jd.warmstate.set_param_trace_eagerness(2)    # for tests
         jd.warmstate.set_param_trace_limit(trace_limit)
         jd.warmstate.set_param_inlining(inline)
@@ -291,9 +292,6 @@
         self.stats = stats
         if translate_support_code:
             self.annhelper = MixLevelHelperAnnotator(self.translator.rtyper)
-            annhelper = self.annhelper
-        else:
-            annhelper = None
         cpu = CPUClass(self.translator.rtyper, self.stats, self.opt,
                        translate_support_code, gcdescr=self.gcdescr)
         self.cpu = cpu
@@ -422,7 +420,7 @@
         if self.translator.rtyper.type_system.name == 'lltypesystem':
             def maybe_enter_jit(*args):
                 try:
-                    maybe_compile_and_run(*args)
+                    maybe_compile_and_run(state.increment_threshold, *args)
                 except JitException:
                     raise     # go through
                 except Exception, e:
@@ -430,15 +428,12 @@
             maybe_enter_jit._always_inline_ = True
         else:
             def maybe_enter_jit(*args):
-                maybe_compile_and_run(*args)
+                maybe_compile_and_run(state.increment_threshold, *args)
             maybe_enter_jit._always_inline_ = True
         jd._maybe_enter_jit_fn = maybe_enter_jit
 
-        can_inline = state.can_inline_greenargs
-        num_green_args = jd.num_green_args
         def maybe_enter_from_start(*args):
-            if not can_inline(*args[:num_green_args]):
-                maybe_compile_and_run(*args)
+            maybe_compile_and_run(state.increment_function_threshold, *args)
         maybe_enter_from_start._always_inline_ = True
         jd._maybe_enter_from_start_fn = maybe_enter_from_start
 
@@ -549,7 +544,6 @@
             self.rewrite_can_enter_jit(jd, sublist)
 
     def rewrite_can_enter_jit(self, jd, can_enter_jits):
-        FUNC = jd._JIT_ENTER_FUNCTYPE
         FUNCPTR = jd._PTR_JIT_ENTER_FUNCTYPE
         jit_enter_fnptr = self.helper_func(FUNCPTR, jd._maybe_enter_jit_fn)
 
diff --git a/pypy/jit/metainterp/warmstate.py b/pypy/jit/metainterp/warmstate.py
--- a/pypy/jit/metainterp/warmstate.py
+++ b/pypy/jit/metainterp/warmstate.py
@@ -1,7 +1,7 @@
 import sys, weakref
 from pypy.rpython.lltypesystem import lltype, llmemory, rstr, rffi
 from pypy.rpython.ootypesystem import ootype
-from pypy.rpython.annlowlevel import hlstr, llstr, cast_base_ptr_to_instance
+from pypy.rpython.annlowlevel import hlstr, cast_base_ptr_to_instance
 from pypy.rpython.annlowlevel import cast_object_to_ptr
 from pypy.rlib.objectmodel import specialize, we_are_translated, r_dict
 from pypy.rlib.rarithmetic import intmask
@@ -208,15 +208,20 @@
             meth = getattr(self, 'set_param_' + name)
             meth(default_value)
 
-    def set_param_threshold(self, threshold):
+    def _compute_threshold(self, threshold):
         if threshold <= 0:
-            self.increment_threshold = 0   # never reach the THRESHOLD_LIMIT
-            return
+            return 0 # never reach the THRESHOLD_LIMIT
         if threshold < 2:
             threshold = 2
-        self.increment_threshold = (self.THRESHOLD_LIMIT // threshold) + 1
+        return (self.THRESHOLD_LIMIT // threshold) + 1
         # the number is at least 1, and at most about half THRESHOLD_LIMIT
 
+    def set_param_threshold(self, threshold):
+        self.increment_threshold = self._compute_threshold(threshold)
+
+    def set_param_function_threshold(self, threshold):
+        self.increment_function_threshold = self._compute_threshold(threshold)
+
     def set_param_trace_eagerness(self, value):
         self.trace_eagerness = value
 
@@ -291,7 +296,7 @@
         self.make_jitdriver_callbacks()
         confirm_enter_jit = self.confirm_enter_jit
 
-        def maybe_compile_and_run(*args):
+        def maybe_compile_and_run(threshold, *args):
             """Entry point to the JIT.  Called at the point with the
             can_enter_jit() hint.
             """
@@ -307,7 +312,7 @@
 
             if cell.counter >= 0:
                 # update the profiling counter
-                n = cell.counter + self.increment_threshold
+                n = cell.counter + threshold
                 if n <= self.THRESHOLD_LIMIT:       # bound not reached
                     cell.counter = n
                     return
@@ -497,7 +502,6 @@
         if hasattr(self, 'set_future_values'):
             return self.set_future_values
 
-        warmrunnerdesc = self.warmrunnerdesc
         jitdriver_sd   = self.jitdriver_sd
         cpu = self.cpu
         vinfo = jitdriver_sd.virtualizable_info
@@ -513,7 +517,6 @@
         #
         if vinfo is not None:
             i0 = len(jitdriver_sd._red_args_types)
-            num_green_args = jitdriver_sd.num_green_args
             index_of_virtualizable = jitdriver_sd.index_of_virtualizable
             vable_static_fields = unrolling_iterable(
                 zip(vinfo.static_extra_types, vinfo.static_fields))
diff --git a/pypy/jit/tool/oparser.py b/pypy/jit/tool/oparser.py
--- a/pypy/jit/tool/oparser.py
+++ b/pypy/jit/tool/oparser.py
@@ -3,24 +3,15 @@
 in a nicer fashion
 """
 
-from pypy.jit.metainterp.history import TreeLoop, BoxInt, ConstInt,\
-     ConstObj, ConstPtr, Box, BasicFailDescr, BoxFloat, ConstFloat,\
-     LoopToken, get_const_ptr_for_string, get_const_ptr_for_unicode
+from pypy.jit.tool.oparser_model import get_model
+
 from pypy.jit.metainterp.resoperation import rop, ResOperation, \
                                             ResOpWithDescr, N_aryOp, \
                                             UnaryOp, PlainResOp
-from pypy.jit.metainterp.typesystem import llhelper
-from pypy.jit.codewriter.heaptracker import adr2int
-from pypy.jit.codewriter import longlong
-from pypy.rpython.lltypesystem import lltype, llmemory
-from pypy.rpython.ootypesystem import ootype
 
 class ParseError(Exception):
     pass
 
-class Boxes(object):
-    pass
-
 class ESCAPE_OP(N_aryOp, ResOpWithDescr):
 
     OPNUM = -123
@@ -54,37 +45,15 @@
     def clone(self):
         return FORCE_SPILL(self.OPNUM, self.getarglist()[:])
 
-class ExtendedTreeLoop(TreeLoop):
 
-    def getboxes(self):
-        def opboxes(operations):
-            for op in operations:
-                yield op.result
-                for box in op.getarglist():
-                    yield box
-        def allboxes():
-            for box in self.inputargs:
-                yield box
-            for box in opboxes(self.operations):
-                yield box
-
-        boxes = Boxes()
-        for box in allboxes():
-            if isinstance(box, Box):
-                name = str(box)
-                setattr(boxes, name, box)
-        return boxes
-
-    def setvalues(self, **kwds):
-        boxes = self.getboxes()
-        for name, value in kwds.iteritems():
-            getattr(boxes, name).value = value
-
-def default_fail_descr(fail_args=None):
-    return BasicFailDescr()
+def default_fail_descr(model, fail_args=None):
+    return model.BasicFailDescr()
 
 
 class OpParser(object):
+
+    use_mock_model = False
+    
     def __init__(self, input, cpu, namespace, type_system, boxkinds,
                  invent_fail_descr=default_fail_descr,
                  nonstrict=False):
@@ -100,7 +69,8 @@
             self._cache = {}
         self.invent_fail_descr = invent_fail_descr
         self.nonstrict = nonstrict
-        self.looptoken = LoopToken()
+        self.model = get_model(self.use_mock_model)
+        self.looptoken = self.model.LoopToken()
 
     def get_const(self, name, typ):
         if self._consts is None:
@@ -108,16 +78,16 @@
         obj = self._consts[name]
         if self.type_system == 'lltype':
             if typ == 'ptr':
-                return ConstPtr(obj)
+                return self.model.ConstPtr(obj)
             else:
                 assert typ == 'class'
-                return ConstInt(adr2int(llmemory.cast_ptr_to_adr(obj)))
+                return self.model.ConstInt(self.model.ptr_to_int(obj))
         else:
             if typ == 'ptr':
-                return ConstObj(obj)
+                return self.model.ConstObj(obj)
             else:
                 assert typ == 'class'
-                return ConstObj(ootype.cast_to_object(obj))
+                return self.model.ConstObj(ootype.cast_to_object(obj))
 
     def get_descr(self, poss_descr):
         if poss_descr.startswith('<'):
@@ -132,16 +102,16 @@
             pass
         if elem.startswith('i'):
             # integer
-            box = BoxInt()
-            _box_counter_more_than(elem[1:])
+            box = self.model.BoxInt()
+            _box_counter_more_than(self.model, elem[1:])
         elif elem.startswith('f'):
-            box = BoxFloat()
-            _box_counter_more_than(elem[1:])
+            box = self.model.BoxFloat()
+            _box_counter_more_than(self.model, elem[1:])
         elif elem.startswith('p'):
             # pointer
-            ts = getattr(self.cpu, 'ts', llhelper)
+            ts = getattr(self.cpu, 'ts', self.model.llhelper)
             box = ts.BoxRef()
-            _box_counter_more_than(elem[1:])
+            _box_counter_more_than(self.model, elem[1:])
         else:
             for prefix, boxclass in self.boxkinds.iteritems():
                 if elem.startswith(prefix):
@@ -175,21 +145,21 @@
 
     def getvar(self, arg):
         if not arg:
-            return ConstInt(0)
+            return self.model.ConstInt(0)
         try:
-            return ConstInt(int(arg))
+            return self.model.ConstInt(int(arg))
         except ValueError:
             if self.is_float(arg):
-                return ConstFloat(longlong.getfloatstorage(float(arg)))
+                return self.model.ConstFloat(self.model.convert_to_floatstorage(arg))
             if (arg.startswith('"') or arg.startswith("'") or
                 arg.startswith('s"')):
                 # XXX ootype
                 info = arg[1:].strip("'\"")
-                return get_const_ptr_for_string(info)
+                return self.model.get_const_ptr_for_string(info)
             if arg.startswith('u"'):
                 # XXX ootype
                 info = arg[1:].strip("'\"")
-                return get_const_ptr_for_unicode(info)
+                return self.model.get_const_ptr_for_unicode(info)
             if arg.startswith('ConstClass('):
                 name = arg[len('ConstClass('):-1]
                 return self.get_const(name, 'class')
@@ -197,9 +167,9 @@
                 return None
             elif arg == 'NULL':
                 if self.type_system == 'lltype':
-                    return ConstPtr(ConstPtr.value)
+                    return self.model.ConstPtr(self.model.ConstPtr.value)
                 else:
-                    return ConstObj(ConstObj.value)
+                    return self.model.ConstObj(self.model.ConstObj.value)
             elif arg.startswith('ConstPtr('):
                 name = arg[len('ConstPtr('):-1]
                 return self.get_const(name, 'ptr')
@@ -211,11 +181,8 @@
         args = []
         descr = None
         if argspec.strip():
-            if opname == 'debug_merge_point':
-                allargs = argspec.split(',', 2)
-            else:
-                allargs = [arg for arg in argspec.split(",")
-                           if arg != '']
+            allargs = [arg for arg in argspec.split(",")
+                       if arg != '']
 
             poss_descr = allargs[-1].strip()
             if poss_descr.startswith('descr='):
@@ -266,14 +233,14 @@
                                 "Unknown var in fail_args: %s" % arg)
                     fail_args.append(fail_arg)
             if descr is None and self.invent_fail_descr:
-                descr = self.invent_fail_descr(fail_args)
+                descr = self.invent_fail_descr(self.model, fail_args)
             if hasattr(descr, '_oparser_uses_descr_of_guard'):
                 descr._oparser_uses_descr_of_guard(self, fail_args)
         else:
             fail_args = None
             if opnum == rop.FINISH:
                 if descr is None and self.invent_fail_descr:
-                    descr = self.invent_fail_descr()
+                    descr = self.invent_fail_descr(self.model)
             elif opnum == rop.JUMP:
                 if descr is None and self.invent_fail_descr:
                     descr = self.looptoken
@@ -338,7 +305,7 @@
         num, ops, last_offset = self.parse_ops(base_indent, newlines, 0)
         if num < len(newlines):
             raise ParseError("unexpected dedent at line: %s" % newlines[num])
-        loop = ExtendedTreeLoop("loop")
+        loop = self.model.ExtendedTreeLoop("loop")
         loop.comment = first_comment
         loop.token = self.looptoken
         loop.operations = ops
@@ -394,7 +361,7 @@
 
 def parse(input, cpu=None, namespace=None, type_system='lltype',
           boxkinds=None, invent_fail_descr=default_fail_descr,
-          no_namespace=False, nonstrict=False):
+          no_namespace=False, nonstrict=False, OpParser=OpParser):
     if namespace is None and not no_namespace:
         namespace = {}
     return OpParser(input, cpu, namespace, type_system, boxkinds,
@@ -405,6 +372,6 @@
     return parse(*args, **kwds)
 
 
-def _box_counter_more_than(s):
+def _box_counter_more_than(model, s):
     if s.isdigit():
-        Box._counter = max(Box._counter, int(s)+1)
+        model.Box._counter = max(model.Box._counter, int(s)+1)
diff --git a/pypy/jit/tool/oparser_model.py b/pypy/jit/tool/oparser_model.py
new file mode 100644
--- /dev/null
+++ b/pypy/jit/tool/oparser_model.py
@@ -0,0 +1,148 @@
+class Boxes(object):
+    pass
+
+def get_real_model():
+    class LoopModel(object):
+        from pypy.jit.metainterp.history import TreeLoop, LoopToken
+        from pypy.jit.metainterp.history import Box, BoxInt, BoxFloat
+        from pypy.jit.metainterp.history import ConstInt, ConstObj, ConstPtr, ConstFloat
+        from pypy.jit.metainterp.history import BasicFailDescr
+        from pypy.jit.metainterp.typesystem import llhelper
+
+        from pypy.jit.metainterp.history import get_const_ptr_for_string
+        from pypy.jit.metainterp.history import get_const_ptr_for_unicode
+        get_const_ptr_for_string = staticmethod(get_const_ptr_for_string)
+        get_const_ptr_for_unicode = staticmethod(get_const_ptr_for_unicode)
+
+        @staticmethod
+        def convert_to_floatstorage(arg):
+            from pypy.jit.codewriter import longlong
+            return longlong.getfloatstorage(float(arg))
+
+        @staticmethod
+        def ptr_to_int(obj):
+            from pypy.jit.codewriter.heaptracker import adr2int
+            from pypy.rpython.lltypesystem import llmemory
+            return adr2int(llmemory.cast_ptr_to_adr(obj))
+
+        @staticmethod
+        def ootype_cast_to_object(obj):
+            from pypy.rpython.ootypesystem import ootype
+            return ootype.cast_to_object(obj)
+
+    return LoopModel
+
+def get_mock_model():
+    class LoopModel(object):
+
+        class TreeLoop(object):
+            def __init__(self, name):
+                self.name = name
+
+        class LoopToken(object):
+            I_am_a_descr = True
+
+        class BasicFailDescr(object):
+            I_am_a_descr = True
+
+        class Box(object):
+            _counter = 0
+            type = 'b'
+
+            def __init__(self, value=0):
+                self.value = value
+
+            def __repr__(self):
+                result = str(self)
+                result += '(%s)' % self.value
+                return result
+
+            def __str__(self):
+                if not hasattr(self, '_str'):
+                    self._str = '%s%d' % (self.type, Box._counter)
+                    Box._counter += 1
+                return self._str
+
+        class BoxInt(Box):
+            type = 'i'
+
+        class BoxFloat(Box):
+            type = 'f'
+
+        class BoxRef(Box):
+            type = 'p'
+
+        class Const(object):
+            def __init__(self, value=None):
+                self.value = value
+
+            def _get_str(self):
+                return str(self.value)
+
+        class ConstInt(Const):
+            pass
+
+        class ConstPtr(Const):
+            pass
+
+        class ConstFloat(Const):
+            pass
+
+        @classmethod
+        def get_const_ptr_for_string(cls, s):
+            return cls.ConstPtr(s)
+
+        @classmethod
+        def get_const_ptr_for_unicode(cls, s):
+            return cls.ConstPtr(s)
+
+        @staticmethod
+        def convert_to_floatstorage(arg):
+            return float(arg)
+
+        @staticmethod
+        def ptr_to_int(obj):
+            return id(obj)
+
+        class llhelper(object):
+            pass
+
+    LoopModel.llhelper.BoxRef = LoopModel.BoxRef
+
+    return LoopModel
+
+
+def get_model(use_mock):
+    if use_mock:
+        model = get_mock_model()
+    else:
+        model = get_real_model()
+
+    class ExtendedTreeLoop(model.TreeLoop):
+
+        def getboxes(self):
+            def opboxes(operations):
+                for op in operations:
+                    yield op.result
+                    for box in op.getarglist():
+                        yield box
+            def allboxes():
+                for box in self.inputargs:
+                    yield box
+                for box in opboxes(self.operations):
+                    yield box
+
+            boxes = Boxes()
+            for box in allboxes():
+                if isinstance(box, model.Box):
+                    name = str(box)
+                    setattr(boxes, name, box)
+            return boxes
+
+        def setvalues(self, **kwds):
+            boxes = self.getboxes()
+            for name, value in kwds.iteritems():
+                getattr(boxes, name).value = value
+
+    model.ExtendedTreeLoop = ExtendedTreeLoop
+    return model
diff --git a/pypy/jit/tool/pypytrace-mode.el b/pypy/jit/tool/pypytrace-mode.el
--- a/pypy/jit/tool/pypytrace-mode.el
+++ b/pypy/jit/tool/pypytrace-mode.el
@@ -32,7 +32,7 @@
     ("<.*FieldDescr \\([^ ]*\\)" (1 'font-lock-variable-name-face))
     ;; comment out debug_merge_point, but then highlight specific part of it
     ("^debug_merge_point.*" . font-lock-comment-face)
-    ("^\\(debug_merge_point\\).*code object\\(.*\\), file \\('.*'\\), \\(line .*\\)> \\(.*\\)"
+    ("^\\(debug_merge_point\\).*code object\\(.*\\). file \\('.*'\\). \\(line .*\\)> \\(.*\\)"
      (1 'compilation-warning t)
      (2 'escape-glyph t)
      (3 'font-lock-string-face t)
diff --git a/pypy/jit/tool/test/test_oparser.py b/pypy/jit/tool/test/test_oparser.py
--- a/pypy/jit/tool/test/test_oparser.py
+++ b/pypy/jit/tool/test/test_oparser.py
@@ -1,227 +1,274 @@
 import py
+import sys
 from pypy.rpython.lltypesystem import lltype, llmemory
 
-from pypy.jit.tool.oparser import parse, ParseError
+from pypy.jit.tool.oparser import parse, OpParser
 from pypy.jit.metainterp.resoperation import rop
-from pypy.jit.metainterp.history import AbstractDescr, BoxInt, LoopToken,\
-     BoxFloat
+from pypy.jit.metainterp.history import AbstractDescr, BoxInt, LoopToken
 
-def test_basic_parse():
-    x = """
-    [i0, i1]
-    # a comment
-    i2 = int_add(i0, i1)
-    i3 = int_sub(i2, 3) # another comment
-    finish() # (tricky)
-    """
-    loop = parse(x)
-    assert len(loop.operations) == 3
-    assert [op.getopnum() for op in loop.operations] == [rop.INT_ADD, rop.INT_SUB,
-                                                    rop.FINISH]
-    assert len(loop.inputargs) == 2
-    assert loop.operations[-1].getdescr()
+class BaseTestOparser(object):
 
-def test_const_ptr_subops():
-    x = """
-    [p0]
-    guard_class(p0, ConstClass(vtable)) []
-    """
-    S = lltype.Struct('S')
-    vtable = lltype.nullptr(S)
-    loop = parse(x, None, locals())
-    assert len(loop.operations) == 1
-    assert loop.operations[0].getdescr()
-    assert loop.operations[0].getfailargs() == []
+    OpParser = None
 
-def test_descr():
-    class Xyz(AbstractDescr):
-        pass
-    
-    x = """
-    [p0]
-    i1 = getfield_gc(p0, descr=stuff)
-    """
-    stuff = Xyz()
-    loop = parse(x, None, locals())
-    assert loop.operations[0].getdescr() is stuff
+    def parse(self, *args, **kwds):
+        kwds['OpParser'] = self.OpParser
+        return parse(*args, **kwds)
 
-def test_after_fail():
-    x = """
-    [i0]
-    guard_value(i0, 3) []
-    i1 = int_add(1, 2)
-    """
-    loop = parse(x, None, {})
-    assert len(loop.operations) == 2
+    def test_basic_parse(self):
+        x = """
+        [i0, i1]
+        # a comment
+        i2 = int_add(i0, i1)
+        i3 = int_sub(i2, 3) # another comment
+        finish() # (tricky)
+        """
+        loop = self.parse(x)
+        assert len(loop.operations) == 3
+        assert [op.getopnum() for op in loop.operations] == [rop.INT_ADD, rop.INT_SUB,
+                                                        rop.FINISH]
+        assert len(loop.inputargs) == 2
+        assert loop.operations[-1].getdescr()
 
-def test_descr_setfield():
-    class Xyz(AbstractDescr):
-        pass
-    
-    x = """
-    [p0]
-    setfield_gc(p0, 3, descr=stuff)
-    """
-    stuff = Xyz()
-    loop = parse(x, None, locals())
-    assert loop.operations[0].getdescr() is stuff
+    def test_const_ptr_subops(self):
+        x = """
+        [p0]
+        guard_class(p0, ConstClass(vtable)) []
+        """
+        S = lltype.Struct('S')
+        vtable = lltype.nullptr(S)
+        loop = self.parse(x, None, locals())
+        assert len(loop.operations) == 1
+        assert loop.operations[0].getdescr()
+        assert loop.operations[0].getfailargs() == []
 
-def test_boxname():
-    x = """
-    [i42]
-    i50 = int_add(i42, 1)
-    """
-    loop = parse(x, None, {})
-    assert str(loop.inputargs[0]) == 'i42'
-    assert str(loop.operations[0].result) == 'i50'
+    def test_descr(self):
+        class Xyz(AbstractDescr):
+            I_am_a_descr = True # for the mock case
 
-def test_getboxes():
-    x = """
-    [i0]
-    i1 = int_add(i0, 10)
-    """
-    loop = parse(x, None, {})
-    boxes = loop.getboxes()
-    assert boxes.i0 is loop.inputargs[0]
-    assert boxes.i1 is loop.operations[0].result
-    
-def test_setvalues():
-    x = """
-    [i0]
-    i1 = int_add(i0, 10)
-    """
-    loop = parse(x, None, {})
-    loop.setvalues(i0=32, i1=42)
-    assert loop.inputargs[0].value == 32
-    assert loop.operations[0].result.value == 42
+        x = """
+        [p0]
+        i1 = getfield_gc(p0, descr=stuff)
+        """
+        stuff = Xyz()
+        loop = self.parse(x, None, locals())
+        assert loop.operations[0].getdescr() is stuff
 
-def test_boxkind():
-    x = """
-    [sum0]
-    """
-    loop = parse(x, None, {}, boxkinds={'sum': BoxInt})
-    b = loop.getboxes()
-    assert isinstance(b.sum0, BoxInt)
-    
-def test_getvar_const_ptr():
-    x = '''
-    []
-    call(ConstPtr(func_ptr))
+    def test_after_fail(self):
+        x = """
+        [i0]
+        guard_value(i0, 3) []
+        i1 = int_add(1, 2)
+        """
+        loop = self.parse(x, None, {})
+        assert len(loop.operations) == 2
+
+    def test_descr_setfield(self):
+        class Xyz(AbstractDescr):
+            I_am_a_descr = True # for the mock case
+
+        x = """
+        [p0]
+        setfield_gc(p0, 3, descr=stuff)
+        """
+        stuff = Xyz()
+        loop = self.parse(x, None, locals())
+        assert loop.operations[0].getdescr() is stuff
+
+    def test_boxname(self):
+        x = """
+        [i42]
+        i50 = int_add(i42, 1)
+        """
+        loop = self.parse(x, None, {})
+        assert str(loop.inputargs[0]) == 'i42'
+        assert str(loop.operations[0].result) == 'i50'
+
+    def test_getboxes(self):
+        x = """
+        [i0]
+        i1 = int_add(i0, 10)
+        """
+        loop = self.parse(x, None, {})
+        boxes = loop.getboxes()
+        assert boxes.i0 is loop.inputargs[0]
+        assert boxes.i1 is loop.operations[0].result
+
+    def test_setvalues(self):
+        x = """
+        [i0]
+        i1 = int_add(i0, 10)
+        """
+        loop = self.parse(x, None, {})
+        loop.setvalues(i0=32, i1=42)
+        assert loop.inputargs[0].value == 32
+        assert loop.operations[0].result.value == 42
+
+    def test_getvar_const_ptr(self):
+        x = '''
+        []
+        call(ConstPtr(func_ptr))
+        '''
+        TP = lltype.GcArray(lltype.Signed)
+        NULL = lltype.cast_opaque_ptr(llmemory.GCREF, lltype.nullptr(TP))
+        loop = self.parse(x, None, {'func_ptr' : NULL})
+        assert loop.operations[0].getarg(0).value == NULL
+
+    def test_jump_target(self):
+        x = '''
+        []
+        jump()
+        '''
+        loop = self.parse(x)
+        assert loop.operations[0].getdescr() is loop.token
+
+    def test_jump_target_other(self):
+        looptoken = LoopToken()
+        looptoken.I_am_a_descr = True # for the mock case
+        x = '''
+        []
+        jump(descr=looptoken)
+        '''
+        loop = self.parse(x, namespace=locals())
+        assert loop.operations[0].getdescr() is looptoken
+
+    def test_floats(self):
+        x = '''
+        [f0]
+        f1 = float_add(f0, 3.5)
+        '''
+        loop = self.parse(x)
+        box = loop.operations[0].getarg(0)
+        # we cannot use isinstance, because in case of mock the class will be
+        # constructed on the fly
+        assert box.__class__.__name__ == 'BoxFloat'
+
+    def test_debug_merge_point(self):
+        x = '''
+        []
+        debug_merge_point(0, "info")
+        debug_merge_point(0, 'info')
+        debug_merge_point(1, '<some ('other.')> info')
+        debug_merge_point(0, '(stuff) #1')
+        '''
+        loop = self.parse(x)
+        assert loop.operations[0].getarg(1)._get_str() == 'info'
+        assert loop.operations[1].getarg(1)._get_str() == 'info'
+        assert loop.operations[2].getarg(1)._get_str() == "<some ('other.')> info"
+        assert loop.operations[3].getarg(1)._get_str() == "(stuff) #1"
+
+
+    def test_descr_with_obj_print(self):
+        x = '''
+        [p0]
+        setfield_gc(p0, 1, descr=<SomeDescr>)
+        '''
+        loop = self.parse(x)
+        # assert did not explode
+
+    example_loop_log = '''\
+    # bridge out of Guard12, 6 ops
+    [i0, i1, i2]
+    i4 = int_add(i0, 2)
+    i6 = int_sub(i1, 1)
+    i8 = int_gt(i6, 3)
+    guard_true(i8, descr=<Guard15>) [i4, i6]
+    debug_merge_point('(no jitdriver.get_printable_location!)', 0)
+    jump(i6, i4, descr=<Loop0>)
     '''
-    TP = lltype.GcArray(lltype.Signed)
-    NULL = lltype.cast_opaque_ptr(llmemory.GCREF, lltype.nullptr(TP))
-    loop = parse(x, None, {'func_ptr' : NULL})
-    assert loop.operations[0].getarg(0).value == NULL
 
-def test_jump_target():
-    x = '''
-    []
-    jump()
-    '''
-    loop = parse(x)
-    assert loop.operations[0].getdescr() is loop.token
+    def test_parse_no_namespace(self):
+        loop = self.parse(self.example_loop_log, no_namespace=True)
 
-def test_jump_target_other():
-    looptoken = LoopToken()
-    x = '''
-    []
-    jump(descr=looptoken)
-    '''
-    loop = parse(x, namespace=locals())
-    assert loop.operations[0].getdescr() is looptoken
+    def test_attach_comment_to_loop(self):
+        loop = self.parse(self.example_loop_log, no_namespace=True)
+        assert loop.comment == '    # bridge out of Guard12, 6 ops'
 
-def test_floats():
-    x = '''
-    [f0]
-    f1 = float_add(f0, 3.5)
-    '''
-    loop = parse(x)
-    assert isinstance(loop.operations[0].getarg(0), BoxFloat)
-    
-def test_debug_merge_point():
-    x = '''
-    []
-    debug_merge_point(0, "info")
-    debug_merge_point(0, 'info')
-    debug_merge_point(1, '<some ('other,')> info')
-    debug_merge_point(0, '(stuff) #1')
-    '''
-    loop = parse(x)
-    assert loop.operations[0].getarg(1)._get_str() == 'info'
-    assert loop.operations[1].getarg(1)._get_str() == 'info'
-    assert loop.operations[2].getarg(1)._get_str() == "<some ('other,')> info"
-    assert loop.operations[3].getarg(1)._get_str() == "(stuff) #1"
-    
+    def test_parse_new_with_comma(self):
+        # this is generated by PYPYJITLOG, check that we can handle it
+        x = '''
+        []
+        p0 = new(, descr=<SizeDescr 12>)
+        '''
+        loop = self.parse(x)
+        assert loop.operations[0].getopname() == 'new'
 
-def test_descr_with_obj_print():
-    x = '''
-    [p0]
-    setfield_gc(p0, 1, descr=<SomeDescr>)
-    '''
-    loop = parse(x)
-    # assert did not explode
+    def test_no_fail_args(self):
+        x = '''
+        [i0]
+        guard_true(i0, descr=<Guard0>)
+        '''
+        loop = self.parse(x, nonstrict=True)
+        assert loop.operations[0].getfailargs() == []
 
-example_loop_log = '''\
-# bridge out of Guard12, 6 ops
-[i0, i1, i2]
-i4 = int_add(i0, 2)
-i6 = int_sub(i1, 1)
-i8 = int_gt(i6, 3)
-guard_true(i8, descr=<Guard15>) [i4, i6]
-debug_merge_point('(no jitdriver.get_printable_location!)', 0)
-jump(i6, i4, descr=<Loop0>)
-'''
+    def test_no_inputargs(self):
+        x = '''
+        i2 = int_add(i0, i1)
+        '''
+        loop = self.parse(x, nonstrict=True)
+        assert loop.inputargs == []
+        assert loop.operations[0].getopname() == 'int_add'
 
-def test_parse_no_namespace():
-    loop = parse(example_loop_log, no_namespace=True)
+    def test_offsets(self):
+        x = """
+        [i0, i1]
+        +10: i2 = int_add(i0, i1)
+        i3 = int_add(i2, 3)
+        """
+        #    +30: --end of the loop--
+        loop = self.parse(x)
+        assert loop.operations[0].offset == 10
+        assert not hasattr(loop.operations[1], 'offset')
 
-def test_attach_comment_to_loop():
-    loop = parse(example_loop_log, no_namespace=True)
-    assert loop.comment == '# bridge out of Guard12, 6 ops'
+    def test_last_offset(self):
+        x = """
+        [i0, i1]
+        +10: i2 = int_add(i0, i1)
+        i3 = int_add(i2, 3)
+        +30: --end of the loop--
+        """
+        loop = self.parse(x)
+        assert len(loop.operations) == 2
+        assert loop.last_offset == 30
 
-def test_parse_new_with_comma():
-    # this is generated by PYPYJITLOG, check that we can handle it
-    x = '''
-    []
-    p0 = new(, descr=<SizeDescr 12>)
-    '''
-    loop = parse(x)
-    assert loop.operations[0].getopname() == 'new'
 
-def test_no_fail_args():
-    x = '''
-    [i0]
-    guard_true(i0, descr=<Guard0>)
-    '''
-    loop = parse(x, nonstrict=True)
-    assert loop.operations[0].getfailargs() == []
+class TestOpParser(BaseTestOparser):
 
-def test_no_inputargs():
-    x = '''
-    i2 = int_add(i0, i1)
-    '''
-    loop = parse(x, nonstrict=True)
-    assert loop.inputargs == []
-    assert loop.operations[0].getopname() == 'int_add'
+    OpParser = OpParser
 
-def test_offsets():
-    x = """
-    [i0, i1]
-    +10: i2 = int_add(i0, i1)
-    i3 = int_add(i2, 3)
-    """
-    #    +30: --end of the loop--
-    loop = parse(x)
-    assert loop.operations[0].offset == 10
-    assert not hasattr(loop.operations[1], 'offset')
+    def test_boxkind(self):
+        x = """
+        [sum0]
+        """
+        loop = self.parse(x, None, {}, boxkinds={'sum': BoxInt})
+        b = loop.getboxes()
+        assert isinstance(b.sum0, BoxInt)
 
-def test_last_offset():
-    x = """
-    [i0, i1]
-    +10: i2 = int_add(i0, i1)
-    i3 = int_add(i2, 3)
-    +30: --end of the loop--
-    """
-    loop = parse(x)
-    assert len(loop.operations) == 2
-    assert loop.last_offset == 30
+
+class ForbiddenModule(object):
+    def __init__(self, name, old_mod):
+        self.name = name
+        self.old_mod = old_mod
+
+    def __getattr__(self, attr):
+        assert False, "You should not import module %s" % self.name
+
+
+class TestOpParserWithMock(BaseTestOparser):
+
+    class OpParser(OpParser):
+        use_mock_model = True
+
+    def setup_class(cls):
+        forbidden_mods = [
+            'pypy.jit.metainterp.history',
+            'pypy.rpython.lltypesystem.lltype',
+            ]
+        for modname in forbidden_mods:
+            if modname in sys.modules:
+                newmod = ForbiddenModule(modname, sys.modules[modname])
+                sys.modules[modname] = newmod
+
+    def teardown_class(cls):
+        for modname, mod in sys.modules.iteritems():
+            if isinstance(mod, ForbiddenModule):
+                sys.modules[modname] = mod.old_mod
diff --git a/pypy/module/__builtin__/functional.py b/pypy/module/__builtin__/functional.py
--- a/pypy/module/__builtin__/functional.py
+++ b/pypy/module/__builtin__/functional.py
@@ -294,7 +294,7 @@
             break
         new_frame = space.createframe(code, w_func.w_func_globals,
                                       w_func.closure)
-        new_frame.fastlocals_w[0] = w_item
+        new_frame.locals_stack_w[0] = w_item
         w_res = new_frame.run()
         result_w.append(w_res)
     return result_w
diff --git a/pypy/module/__builtin__/test/test_builtin.py b/pypy/module/__builtin__/test/test_builtin.py
--- a/pypy/module/__builtin__/test/test_builtin.py
+++ b/pypy/module/__builtin__/test/test_builtin.py
@@ -1,5 +1,6 @@
 import autopath
 import sys
+from pypy import conftest
 
 class AppTestBuiltinApp:
     def setup_class(cls):
@@ -15,6 +16,15 @@
             cls.w_sane_lookup = cls.space.wrap(True)
         except KeyError:
             cls.w_sane_lookup = cls.space.wrap(False)
+        # starting with CPython 2.6, when the stack is almost out, we
+        # can get a random error, instead of just a RuntimeError.
+        # For example if an object x has a __getattr__, we can get
+        # AttributeError if attempting to call x.__getattr__ runs out
+        # of stack.  That's annoying, so we just work around it.
+        if conftest.option.runappdirect:
+            cls.w_safe_runtimerror = cls.space.wrap(True)
+        else:
+            cls.w_safe_runtimerror = cls.space.wrap(sys.version_info < (2, 6))
 
     def test_bytes_alias(self):
         assert bytes is str
@@ -399,6 +409,8 @@
     def test_cmp_cyclic(self):
         if not self.sane_lookup:
             skip("underlying Python implementation has insane dict lookup")
+        if not self.safe_runtimerror:
+            skip("underlying Python may raise random exceptions on stack ovf")
         a = []; a.append(a)
         b = []; b.append(b)
         from UserList import UserList
diff --git a/pypy/module/__pypy__/__init__.py b/pypy/module/__pypy__/__init__.py
--- a/pypy/module/__pypy__/__init__.py
+++ b/pypy/module/__pypy__/__init__.py
@@ -3,6 +3,14 @@
 from pypy.interpreter.mixedmodule import MixedModule
 from pypy.module.imp.importing import get_pyc_magic
 
+
+class BuildersModule(MixedModule):
+    appleveldefs = {}
+
+    interpleveldefs = {
+        "UnicodeBuilder": "interp_builders.W_UnicodeBuilder",
+    }
+
 class Module(MixedModule):
     appleveldefs = {
     }
@@ -19,6 +27,10 @@
         'lookup_special'            : 'interp_magic.lookup_special',
     }
 
+    submodules = {
+        "builders": BuildersModule,
+    }
+
     def setup_after_space_initialization(self):
         """NOT_RPYTHON"""
         if not self.space.config.translating:
diff --git a/pypy/module/__pypy__/interp_builders.py b/pypy/module/__pypy__/interp_builders.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/__pypy__/interp_builders.py
@@ -0,0 +1,50 @@
+from pypy.interpreter.baseobjspace import Wrappable
+from pypy.interpreter.error import OperationError
+from pypy.interpreter.gateway import interp2app, unwrap_spec
+from pypy.interpreter.typedef import TypeDef
+from pypy.rlib.rstring import UnicodeBuilder
+
+
+class W_UnicodeBuilder(Wrappable):
+    def __init__(self, space, size):
+        if size == -1:
+            self.builder = UnicodeBuilder()
+        else:
+            self.builder = UnicodeBuilder(size)
+        self.done = False
+
+    def _check_done(self, space):
+        if self.done:
+            raise OperationError(space.w_ValueError, space.wrap("Can't operate on a done builder"))
+
+    @unwrap_spec(size=int)
+    def descr__new__(space, w_subtype, size=-1):
+        return W_UnicodeBuilder(space, size)
+
+    @unwrap_spec(s=unicode)
+    def descr_append(self, space, s):
+        self._check_done(space)
+        self.builder.append(s)
+
+    @unwrap_spec(s=unicode, start=int, end=int)
+    def descr_append_slice(self, space, s, start, end):
+        self._check_done(space)
+        if not 0 <= start <= end <= len(s):
+            raise OperationError(space.w_ValueError, space.wrap("bad start/stop"))
+        self.builder.append_slice(s, start, end)
+
+    def descr_build(self, space):
+        self._check_done(space)
+        w_s = space.wrap(self.builder.build())
+        self.done = True
+        return w_s
+
+
+W_UnicodeBuilder.typedef = TypeDef("UnicodeBuilder",
+    __new__ = interp2app(W_UnicodeBuilder.descr__new__.im_func),
+
+    append = interp2app(W_UnicodeBuilder.descr_append),
+    append_slice = interp2app(W_UnicodeBuilder.descr_append_slice),
+    build = interp2app(W_UnicodeBuilder.descr_build),
+)
+W_UnicodeBuilder.typedef.acceptable_as_base_class = False
\ No newline at end of file
diff --git a/pypy/module/__pypy__/interp_debug.py b/pypy/module/__pypy__/interp_debug.py
--- a/pypy/module/__pypy__/interp_debug.py
+++ b/pypy/module/__pypy__/interp_debug.py
@@ -1,15 +1,19 @@
 from pypy.interpreter.gateway import interp2app, NoneNotWrapped, unwrap_spec
 from pypy.interpreter.error import OperationError
-from pypy.rlib import debug
+from pypy.rlib import debug, jit
 
+
+ at jit.dont_look_inside
 @unwrap_spec(category=str)
 def debug_start(space, category):
     debug.debug_start(category)
 
+ at jit.dont_look_inside
 def debug_print(space, args_w):
     parts = [space.str_w(space.str(w_item)) for w_item in args_w]
     debug.debug_print(' '.join(parts))
 
+ at jit.dont_look_inside
 @unwrap_spec(category=str)
 def debug_stop(space, category):
     debug.debug_stop(category)
diff --git a/pypy/module/__pypy__/test/test_builders.py b/pypy/module/__pypy__/test/test_builders.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/__pypy__/test/test_builders.py
@@ -0,0 +1,34 @@
+from pypy.conftest import gettestobjspace
+
+
+class AppTestBuilders(object):
+    def setup_class(cls):
+        cls.space = gettestobjspace(usemodules=['__pypy__'])
+
+    def test_simple(self):
+        from __pypy__.builders import UnicodeBuilder
+        b = UnicodeBuilder()
+        b.append(u"abc")
+        b.append(u"123")
+        b.append(u"1")
+        s = b.build()
+        assert s == u"abc1231"
+        raises(ValueError, b.build)
+        raises(ValueError, b.append, u"123")
+
+    def test_preallocate(self):
+        from __pypy__.builders import UnicodeBuilder
+        b = UnicodeBuilder(10)
+        b.append(u"abc")
+        b.append(u"123")
+        s = b.build()
+        assert s == u"abc123"
+
+    def test_append_slice(self):
+        from __pypy__.builders import UnicodeBuilder
+        b = UnicodeBuilder()
+        b.append_slice(u"abcdefgh", 2, 5)
+        raises(ValueError, b.append_slice, u"1", 2, 1)
+        s = b.build()
+        assert s == "cde"
+        raises(ValueError, b.append_slice, u"abc", 1, 2)
\ No newline at end of file
diff --git a/pypy/module/_rawffi/interp_rawffi.py b/pypy/module/_rawffi/interp_rawffi.py
--- a/pypy/module/_rawffi/interp_rawffi.py
+++ b/pypy/module/_rawffi/interp_rawffi.py
@@ -176,7 +176,7 @@
             except KeyError:
                 raise operationerrfmt(space.w_AttributeError,
                     "No symbol %s found in library %s", name, self.name)
-        
+
         elif (_MS_WINDOWS and
               space.is_true(space.isinstance(w_name, space.w_int))):
             ordinal = space.int_w(w_name)
@@ -261,7 +261,7 @@
     def descr_size_alignment(self, space, n=1):
         return space.newtuple([space.wrap(self.size * n),
                                space.wrap(self.alignment)])
-    
+
 
 class W_DataInstance(Wrappable):
     def __init__(self, space, size, address=r_uint(0)):
@@ -427,7 +427,7 @@
                     if not (argletter in TYPEMAP_PTR_LETTERS and
                             letter in TYPEMAP_PTR_LETTERS):
                         msg = "Argument %d should be typecode %s, got %s"
-                        raise operationerrfmt(space.w_TypeError, msg, 
+                        raise operationerrfmt(space.w_TypeError, msg,
                                               i+1, argletter, letter)
             args_ll.append(arg.ll_buffer)
             # XXX we could avoid the intermediate list args_ll
@@ -480,17 +480,25 @@
 alignment = _create_new_accessor('alignment', 'c_alignment')
 
 @unwrap_spec(address=r_uint, maxlength=int)
-def charp2string(space, address, maxlength=sys.maxint):
+def charp2string(space, address, maxlength=-1):
     if address == 0:
         return space.w_None
-    s = rffi.charp2strn(rffi.cast(rffi.CCHARP, address), maxlength)
+    charp_addr = rffi.cast(rffi.CCHARP, address)
+    if maxlength == -1:
+        s = rffi.charp2str(charp_addr)
+    else:
+        s = rffi.charp2strn(charp_addr, maxlength)
     return space.wrap(s)
 
 @unwrap_spec(address=r_uint, maxlength=int)
-def wcharp2unicode(space, address, maxlength=sys.maxint):
+def wcharp2unicode(space, address, maxlength=-1):
     if address == 0:
         return space.w_None
-    s = rffi.wcharp2unicoden(rffi.cast(rffi.CWCHARP, address), maxlength)
+    wcharp_addr = rffi.cast(rffi.CWCHARP, address)
+    if maxlength == -1:
+        s = rffi.wcharp2unicode(wcharp_addr)
+    else:
+        s = rffi.wcharp2unicoden(wcharp_addr, maxlength)
     return space.wrap(s)
 
 @unwrap_spec(address=r_uint, maxlength=int)
diff --git a/pypy/module/_ssl/interp_ssl.py b/pypy/module/_ssl/interp_ssl.py
--- a/pypy/module/_ssl/interp_ssl.py
+++ b/pypy/module/_ssl/interp_ssl.py
@@ -1,3 +1,4 @@
+from __future__ import with_statement
 from pypy.rpython.lltypesystem import rffi, lltype
 from pypy.interpreter.error import OperationError
 from pypy.interpreter.baseobjspace import W_Root, ObjSpace, Wrappable
@@ -899,7 +900,7 @@
 
 def _ssl_thread_id_function():
     from pypy.module.thread import ll_thread
-    return rffi.cast(rffi.INT, ll_thread.get_ident())
+    return rffi.cast(rffi.LONG, ll_thread.get_ident())
 
 def setup_ssl_threads():
     from pypy.module.thread import ll_thread
diff --git a/pypy/module/_stackless/test/test_greenlet.py b/pypy/module/_stackless/test/test_greenlet.py
--- a/pypy/module/_stackless/test/test_greenlet.py
+++ b/pypy/module/_stackless/test/test_greenlet.py
@@ -72,6 +72,23 @@
         g1 = greenlet(f)
         raises(ValueError, g2.switch)
 
+
+    def test_exc_info_save_restore(self):
+        from _stackless import greenlet
+        import sys
+        def f():
+            try:
+                raise ValueError('fun')
+            except:
+                exc_info = sys.exc_info()
+                greenlet(h).switch()
+                assert exc_info == sys.exc_info()
+
+        def h():
+            assert sys.exc_info() == (None, None, None)
+
+        greenlet(f).switch()
+
     def test_exception(self):
         from _stackless import greenlet
         import sys
diff --git a/pypy/module/array/interp_array.py b/pypy/module/array/interp_array.py
--- a/pypy/module/array/interp_array.py
+++ b/pypy/module/array/interp_array.py
@@ -526,15 +526,7 @@
 
     def array_tostring__Array(space, self):
         cbuf = self.charbuf()
-        s = ''.join([cbuf[i] for i in xrange(self.len * mytype.bytes)])
-        return self.space.wrap(s)
-##
-##         s = ''
-##         i = 0
-##         while i < self.len * mytype.bytes:
-##             s += cbuf[i]
-##             i += 1
-##         return self.space.wrap(s)
+        return self.space.wrap(rffi.charpsize2str(cbuf, self.len * mytype.bytes))
 
     def array_fromfile__Array_ANY_ANY(space, self, w_f, w_n):
         if not isinstance(w_f, W_File):
diff --git a/pypy/module/cpyext/__init__.py b/pypy/module/cpyext/__init__.py
--- a/pypy/module/cpyext/__init__.py
+++ b/pypy/module/cpyext/__init__.py
@@ -39,6 +39,7 @@
 import pypy.module.cpyext.object
 import pypy.module.cpyext.stringobject
 import pypy.module.cpyext.tupleobject
+import pypy.module.cpyext.setobject
 import pypy.module.cpyext.dictobject
 import pypy.module.cpyext.intobject
 import pypy.module.cpyext.longobject
@@ -64,6 +65,7 @@
 import pypy.module.cpyext.memoryobject
 import pypy.module.cpyext.codecs
 import pypy.module.cpyext.pyfile
+import pypy.module.cpyext.pystrtod
 
 # now that all rffi_platform.Struct types are registered, configure them
 api.configure_types()
diff --git a/pypy/module/cpyext/api.py b/pypy/module/cpyext/api.py
--- a/pypy/module/cpyext/api.py
+++ b/pypy/module/cpyext/api.py
@@ -562,7 +562,8 @@
             elif callable.api_func.restype is not lltype.Void:
                 retval = rffi.cast(callable.api_func.restype, result)
         except Exception, e:
-            print 'Fatal error in cpyext, calling', callable.__name__
+            print 'Fatal error in cpyext, CPython compatibility layer, calling', callable.__name__
+            print 'Either report a bug or consider not using this particular extension'
             if not we_are_translated():
                 import traceback
                 traceback.print_exc()
diff --git a/pypy/module/cpyext/pystrtod.py b/pypy/module/cpyext/pystrtod.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/cpyext/pystrtod.py
@@ -0,0 +1,68 @@
+import errno
+from pypy.interpreter.error import OperationError
+from pypy.module.cpyext.api import cpython_api
+from pypy.module.cpyext.pyobject import PyObject
+from pypy.rlib import rdtoa
+from pypy.rlib import rfloat
+from pypy.rlib import rposix
+from pypy.rpython.lltypesystem import lltype
+from pypy.rpython.lltypesystem import rffi
+
+
+ at cpython_api([rffi.CCHARP, rffi.CCHARPP, PyObject], rffi.DOUBLE, error=-1.0)
+def PyOS_string_to_double(space, s, endptr, w_overflow_exception):
+    """Convert a string s to a double, raising a Python
+    exception on failure.  The set of accepted strings corresponds to
+    the set of strings accepted by Python's float() constructor,
+    except that s must not have leading or trailing whitespace.
+    The conversion is independent of the current locale.
+
+    If endptr is NULL, convert the whole string.  Raise
+    ValueError and return -1.0 if the string is not a valid
+    representation of a floating-point number.
+
+    If endptr is not NULL, convert as much of the string as
+    possible and set *endptr to point to the first unconverted
+    character.  If no initial segment of the string is the valid
+    representation of a floating-point number, set *endptr to point
+    to the beginning of the string, raise ValueError, and return
+    -1.0.
+
+    If s represents a value that is too large to store in a float
+    (for example, "1e500" is such a string on many platforms) then
+    if overflow_exception is NULL return Py_HUGE_VAL (with
+    an appropriate sign) and don't set any exception.  Otherwise,
+    overflow_exception must point to a Python exception object;
+    raise that exception and return -1.0.  In both cases, set
+    *endptr to point to the first character after the converted value.
+
+    If any other error occurs during the conversion (for example an
+    out-of-memory error), set the appropriate Python exception and
+    return -1.0.
+    """
+    user_endptr = True
+    try:
+        if not endptr:
+            endptr = lltype.malloc(rffi.CCHARPP.TO, 1, flavor='raw')
+            user_endptr = False
+        result = rdtoa.dg_strtod(s, endptr)
+        endpos = (rffi.cast(rffi.LONG, endptr[0]) -
+                  rffi.cast(rffi.LONG, s))
+        if endpos == 0 or (not user_endptr and not endptr[0][0] == '\0'):
+            raise OperationError(
+                space.w_ValueError,
+                space.wrap('invalid input at position %s' % endpos))
+        if rposix.get_errno() == errno.ERANGE:
+            rposix.set_errno(0)
+            if w_overflow_exception is None:
+                if result > 0:
+                    return rfloat.INFINITY
+                else:
+                    return -rfloat.INFINITY
+            else:
+                raise OperationError(w_overflow_exception,
+                                     space.wrap('value too large'))
+        return result
+    finally:
+        if not user_endptr:
+            lltype.free(endptr, flavor='raw')
diff --git a/pypy/module/cpyext/setobject.py b/pypy/module/cpyext/setobject.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/cpyext/setobject.py
@@ -0,0 +1,46 @@
+from pypy.interpreter.error import OperationError
+from pypy.rpython.lltypesystem import rffi, lltype
+from pypy.module.cpyext.api import (cpython_api, Py_ssize_t, CANNOT_FAIL,
+                                    build_type_checkers)
+from pypy.module.cpyext.pyobject import (PyObject, PyObjectP, Py_DecRef,
+    borrow_from, make_ref, from_ref)
+from pypy.module.cpyext.pyerrors import PyErr_BadInternalCall
+from pypy.objspace.std.setobject import W_SetObject, newset
+from pypy.objspace.std.smalltupleobject import W_SmallTupleObject
+
+
+PySet_Check, PySet_CheckExact = build_type_checkers("Set")
+
+
+ at cpython_api([PyObject], PyObject)
+def PySet_New(space, w_iterable):
+    if w_iterable is None:
+        return space.call_function(space.w_set)
+    else:
+        return space.call_function(space.w_set, w_iterable)
+
+ at cpython_api([PyObject, PyObject], rffi.INT_real, error=-1)
+def PySet_Add(space, w_s, w_obj):
+    if not PySet_Check(space, w_s):
+        PyErr_BadInternalCall(space)
+    space.call_method(w_s, 'add', w_obj)
+    return 0
+
+ at cpython_api([PyObject, PyObject], rffi.INT_real, error=-1)
+def PySet_Discard(space, w_s, w_obj):
+    if not PySet_Check(space, w_s):
+        PyErr_BadInternalCall(space)
+    space.call_method(w_s, 'discard', w_obj)
+    return 0
+
+
+ at cpython_api([PyObject], Py_ssize_t, error=CANNOT_FAIL)
+def PySet_GET_SIZE(space, w_s):
+    return space.int_w(space.len(w_s))
+
+ at cpython_api([PyObject], Py_ssize_t, error=-1)
+def PySet_Size(space, ref):
+    if not PySet_Check(space, ref):
+        raise OperationError(space.w_TypeError,
+                             space.wrap("expected set object"))
+    return PySet_GET_SIZE(space, ref)
diff --git a/pypy/module/cpyext/stubs.py b/pypy/module/cpyext/stubs.py
--- a/pypy/module/cpyext/stubs.py
+++ b/pypy/module/cpyext/stubs.py
@@ -480,39 +480,6 @@
     """Create a new Python complex number object from a C Py_complex value."""
     raise NotImplementedError
 
- at cpython_api([rffi.CCHARP, rffi.CCHARPP, PyObject], rffi.DOUBLE, error=-1.0)
-def PyOS_string_to_double(space, s, endptr, overflow_exception):
-    """Convert a string s to a double, raising a Python
-    exception on failure.  The set of accepted strings corresponds to
-    the set of strings accepted by Python's float() constructor,
-    except that s must not have leading or trailing whitespace.
-    The conversion is independent of the current locale.
-
-    If endptr is NULL, convert the whole string.  Raise
-    ValueError and return -1.0 if the string is not a valid
-    representation of a floating-point number.
-
-    If endptr is not NULL, convert as much of the string as
-    possible and set *endptr to point to the first unconverted
-    character.  If no initial segment of the string is the valid
-    representation of a floating-point number, set *endptr to point
-    to the beginning of the string, raise ValueError, and return
-    -1.0.
-
-    If s represents a value that is too large to store in a float
-    (for example, "1e500" is such a string on many platforms) then
-    if overflow_exception is NULL return Py_HUGE_VAL (with
-    an appropriate sign) and don't set any exception.  Otherwise,
-    overflow_exception must point to a Python exception object;
-    raise that exception and return -1.0.  In both cases, set
-    *endptr to point to the first character after the converted value.
-
-    If any other error occurs during the conversion (for example an
-    out-of-memory error), set the appropriate Python exception and
-    return -1.0.
-    """
-    raise NotImplementedError
-
 @cpython_api([rffi.CCHARP, rffi.CCHARPP], rffi.DOUBLE, error=CANNOT_FAIL)
 def PyOS_ascii_strtod(space, nptr, endptr):
     """Convert a string to a double. This function behaves like the Standard C
diff --git a/pypy/module/cpyext/test/test_pystrtod.py b/pypy/module/cpyext/test/test_pystrtod.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/cpyext/test/test_pystrtod.py
@@ -0,0 +1,93 @@
+import math
+
+from pypy.module.cpyext.test.test_api import BaseApiTest
+from pypy.rpython.lltypesystem import rffi
+from pypy.rpython.lltypesystem import lltype
+
+
+class TestPyOS_string_to_double(BaseApiTest):
+
+    def test_simple_float(self, api):
+        s = rffi.str2charp('0.4')
+        null = lltype.nullptr(rffi.CCHARPP.TO)
+        r = api.PyOS_string_to_double(s, null, None)
+        assert r == 0.4
+        rffi.free_charp(s)
+
+    def test_empty_string(self, api):
+        s = rffi.str2charp('')
+        null = lltype.nullptr(rffi.CCHARPP.TO)
+        r = api.PyOS_string_to_double(s, null, None)
+        assert r == -1.0
+        raises(ValueError)
+        api.PyErr_Clear()
+        rffi.free_charp(s)
+
+    def test_bad_string(self, api):
+        s = rffi.str2charp(' 0.4')
+        null = lltype.nullptr(rffi.CCHARPP.TO)
+        r = api.PyOS_string_to_double(s, null, None)
+        assert r == -1.0
+        raises(ValueError)
+        api.PyErr_Clear()
+        rffi.free_charp(s)
+
+    def test_overflow_pos(self, api):
+        s = rffi.str2charp('1e500')
+        null = lltype.nullptr(rffi.CCHARPP.TO)
+        r = api.PyOS_string_to_double(s, null, None)
+        assert math.isinf(r)
+        assert r > 0
+        rffi.free_charp(s)
+
+    def test_overflow_neg(self, api):
+        s = rffi.str2charp('-1e500')
+        null = lltype.nullptr(rffi.CCHARPP.TO)
+        r = api.PyOS_string_to_double(s, null, None)
+        assert math.isinf(r)
+        assert r < 0
+        rffi.free_charp(s)
+
+    def test_overflow_exc(self, space, api):
+        s = rffi.str2charp('1e500')
+        null = lltype.nullptr(rffi.CCHARPP.TO)
+        r = api.PyOS_string_to_double(s, null, space.w_ValueError)
+        assert r == -1.0
+        raises(ValueError)
+        api.PyErr_Clear()
+        rffi.free_charp(s)
+
+    def test_endptr_number(self, api):
+        s = rffi.str2charp('0.4')
+        endp = lltype.malloc(rffi.CCHARPP.TO, 1, flavor='raw')
+        r = api.PyOS_string_to_double(s, endp, None)
+        assert r == 0.4
+        endp_addr = rffi.cast(rffi.LONG, endp[0])
+        s_addr = rffi.cast(rffi.LONG, s)
+        assert endp_addr == s_addr + 3
+        rffi.free_charp(s)
+        lltype.free(endp, flavor='raw')
+
+    def test_endptr_tail(self, api):
+        s = rffi.str2charp('0.4 foo')
+        endp = lltype.malloc(rffi.CCHARPP.TO, 1, flavor='raw')
+        r = api.PyOS_string_to_double(s, endp, None)
+        assert r == 0.4
+        endp_addr = rffi.cast(rffi.LONG, endp[0])
+        s_addr = rffi.cast(rffi.LONG, s)
+        assert endp_addr == s_addr + 3
+        rffi.free_charp(s)
+        lltype.free(endp, flavor='raw')
+
+    def test_endptr_no_conversion(self, api):
+        s = rffi.str2charp('foo')
+        endp = lltype.malloc(rffi.CCHARPP.TO, 1, flavor='raw')
+        r = api.PyOS_string_to_double(s, endp, None)
+        assert r == -1.0
+        raises(ValueError)
+        endp_addr = rffi.cast(rffi.LONG, endp[0])
+        s_addr = rffi.cast(rffi.LONG, s)
+        assert endp_addr == s_addr
+        api.PyErr_Clear()
+        rffi.free_charp(s)
+        lltype.free(endp, flavor='raw')
diff --git a/pypy/module/cpyext/test/test_setobject.py b/pypy/module/cpyext/test/test_setobject.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/cpyext/test/test_setobject.py
@@ -0,0 +1,29 @@
+import py
+
+from pypy.module.cpyext.pyobject import PyObject, PyObjectP, make_ref, from_ref
+from pypy.module.cpyext.test.test_api import BaseApiTest
+from pypy.rpython.lltypesystem import rffi, lltype
+from pypy.conftest import gettestobjspace
+
+
+class TestTupleObject(BaseApiTest):
+    def test_setobj(self, space, api):
+        assert not api.PySet_Check(space.w_None)
+        assert api.PySet_Add(space.w_None, space.w_None) == -1
+        api.PyErr_Clear()
+        w_set = space.call_function(space.w_set)
+        space.call_method(w_set, 'update', space.wrap([1,2,3,4]))
+        assert api.PySet_Size(w_set) == 4
+        assert api.PySet_GET_SIZE(w_set) == 4
+        raises(TypeError, api.PySet_Size(space.newlist([])))
+        api.PyErr_Clear()
+
+    def test_set_add_discard(self, space, api):
+        w_set = api.PySet_New(None)
+        assert api.PySet_Size(w_set) == 0
+        w_set = api.PySet_New(space.wrap([1,2,3,4]))
+        assert api.PySet_Size(w_set) == 4
+        api.PySet_Add(w_set, space.wrap(6))
+        assert api.PySet_Size(w_set) == 5
+        api.PySet_Discard(w_set, space.wrap(6))
+        assert api.PySet_Size(w_set) == 4
diff --git a/pypy/module/cpyext/test/test_unicodeobject.py b/pypy/module/cpyext/test/test_unicodeobject.py
--- a/pypy/module/cpyext/test/test_unicodeobject.py
+++ b/pypy/module/cpyext/test/test_unicodeobject.py
@@ -367,3 +367,14 @@
                     data, len(u), lltype.nullptr(rffi.CCHARP.TO))
         rffi.free_wcharp(data)
 
+    def test_format(self, space, api):
+        w_format = space.wrap(u'hi %s')
+        w_args = space.wrap((u'test',))
+        w_formated = api.PyUnicode_Format(w_format, w_args)
+        assert space.unwrap(w_formated) == space.unwrap(space.mod(w_format, w_args))
+
+    def test_join(self, space, api):
+        w_sep = space.wrap(u'<sep>')
+        w_seq = space.wrap([u'a', u'b'])
+        w_joined = api.PyUnicode_Join(w_sep, w_seq)
+        assert space.unwrap(w_joined) == u'a<sep>b'
diff --git a/pypy/module/cpyext/test/test_weakref.py b/pypy/module/cpyext/test/test_weakref.py
--- a/pypy/module/cpyext/test/test_weakref.py
+++ b/pypy/module/cpyext/test/test_weakref.py
@@ -7,6 +7,7 @@
         w_ref = api.PyWeakref_NewRef(w_obj, space.w_None)
         assert w_ref is not None
         assert space.is_w(api.PyWeakref_GetObject(w_ref), w_obj)
+        assert space.is_w(api.PyWeakref_GET_OBJECT(w_ref), w_obj)
         assert space.is_w(api.PyWeakref_LockObject(w_ref), w_obj)
 
         w_obj = space.newtuple([])
diff --git a/pypy/module/cpyext/unicodeobject.py b/pypy/module/cpyext/unicodeobject.py
--- a/pypy/module/cpyext/unicodeobject.py
+++ b/pypy/module/cpyext/unicodeobject.py
@@ -523,3 +523,11 @@
     copies sizeof(Py_UNICODE) * length bytes from source to target"""
     for i in range(0, length):
         target[i] = source[i]
+
+ at cpython_api([PyObject, PyObject], PyObject)
+def PyUnicode_Format(space, w_format, w_args):
+    return space.mod(w_format, w_args)
+
+ at cpython_api([PyObject, PyObject], PyObject)
+def PyUnicode_Join(space, w_sep, w_seq):
+    return space.call_method(w_sep, 'join', w_seq)
diff --git a/pypy/module/cpyext/weakrefobject.py b/pypy/module/cpyext/weakrefobject.py
--- a/pypy/module/cpyext/weakrefobject.py
+++ b/pypy/module/cpyext/weakrefobject.py
@@ -21,6 +21,10 @@
     """Return the referenced object from a weak reference.  If the referent is
     no longer live, returns None. This function returns a borrowed reference.
     """
+    return PyWeakref_GET_OBJECT(space, w_ref)
+
+ at cpython_api([PyObject], PyObject)
+def PyWeakref_GET_OBJECT(space, w_ref):
     return borrow_from(w_ref, space.call_function(w_ref))
 
 @cpython_api([PyObject], PyObject)
diff --git a/pypy/module/imp/importing.py b/pypy/module/imp/importing.py
--- a/pypy/module/imp/importing.py
+++ b/pypy/module/imp/importing.py
@@ -622,7 +622,13 @@
         try:
             if find_info:
                 w_mod = load_module(space, w_modulename, find_info)
-                w_mod = space.getitem(space.sys.get("modules"), w_modulename)
+                try:
+                    w_mod = space.getitem(space.sys.get("modules"),
+                                          w_modulename)
+                except OperationError, oe:
+                    if not oe.match(space, space.w_KeyError):
+                        raise
+                    raise OperationError(space.w_ImportError, w_modulename)
                 if w_parent is not None:
                     space.setattr(w_parent, space.wrap(partname), w_mod)
                 return w_mod
diff --git a/pypy/module/imp/test/test_import.py b/pypy/module/imp/test/test_import.py
--- a/pypy/module/imp/test/test_import.py
+++ b/pypy/module/imp/test/test_import.py
@@ -37,6 +37,7 @@
                     ambig = "imamodule = 1",
                     test_reload = "def test():\n    raise ValueError\n",
                     infinite_reload = "import infinite_reload; reload(infinite_reload)",
+                    del_sys_module = "import sys\ndel sys.modules['del_sys_module']\n",
                     )
     root.ensure("notapackage", dir=1)    # empty, no __init__.py
     setuppkg("pkg",
@@ -562,6 +563,14 @@
         except ImportError:
             pass
 
+    def test_del_from_sys_modules(self):
+        try:
+            import del_sys_module
+        except ImportError:
+            pass    # ok
+        else:
+            assert False, 'should not work'
+
 class TestAbi:
     def test_abi_tag(self):
         space1 = gettestobjspace(soabi='TEST')
diff --git a/pypy/module/math/__init__.py b/pypy/module/math/__init__.py
--- a/pypy/module/math/__init__.py
+++ b/pypy/module/math/__init__.py
@@ -4,6 +4,7 @@
 
 class Module(MixedModule):
     appleveldefs = {
+       'factorial' : 'app_math.factorial'
     }
 
     interpleveldefs = {
@@ -40,7 +41,6 @@
        'isnan'          : 'interp_math.isnan',
        'trunc'          : 'interp_math.trunc',
        'fsum'           : 'interp_math.fsum',
-       'factorial'      : 'interp_math.factorial',
        'asinh'          : 'interp_math.asinh',
        'acosh'          : 'interp_math.acosh',
        'atanh'          : 'interp_math.atanh',
diff --git a/pypy/module/math/app_math.py b/pypy/module/math/app_math.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/math/app_math.py
@@ -0,0 +1,13 @@
+def factorial(x):
+    """Find x!."""
+    if isinstance(x, float):
+        fl = int(x)
+        if fl != x:
+            raise ValueError("float arguments must be integral")
+        x = fl
+    if x < 0:
+        raise ValueError("x must be >= 0")
+    res = 1
+    for i in range(1, x + 1):
+        res *= i
+    return res
diff --git a/pypy/module/math/interp_math.py b/pypy/module/math/interp_math.py
--- a/pypy/module/math/interp_math.py
+++ b/pypy/module/math/interp_math.py
@@ -373,22 +373,6 @@
                 hi = v
     return space.wrap(hi)
 
-def factorial(space, w_x):
-    """Find x!."""
-    if space.isinstance_w(w_x, space.w_float):
-        fl = space.float_w(w_x)
-        if math.floor(fl) != fl:
-            raise OperationError(space.w_ValueError,
-                                 space.wrap("float arguments must be integral"))
-        w_x = space.long(w_x)
-    x = space.int_w(w_x)
-    if x < 0:
-        raise OperationError(space.w_ValueError, space.wrap("x must be >= 0"))
-    w_res = space.wrap(1)
-    for i in range(1, x + 1):
-        w_res = space.mul(w_res, space.wrap(i))
-    return w_res
-
 def log1p(space, w_x):
     """Find log(x + 1)."""
     return math1(space, rfloat.log1p, w_x)
diff --git a/pypy/module/micronumpy/__init__.py b/pypy/module/micronumpy/__init__.py
--- a/pypy/module/micronumpy/__init__.py
+++ b/pypy/module/micronumpy/__init__.py
@@ -8,11 +8,15 @@
     interpleveldefs = {
         'array': 'interp_numarray.SingleDimArray',
         'zeros': 'interp_numarray.zeros',
+        'empty': 'interp_numarray.zeros',
+        'ones': 'interp_numarray.ones',
 
         # ufuncs
+        'abs': 'interp_ufuncs.absolute',
         'absolute': 'interp_ufuncs.absolute',
         'copysign': 'interp_ufuncs.copysign',
         'exp': 'interp_ufuncs.exp',
+        'floor': 'interp_ufuncs.floor',
         'maximum': 'interp_ufuncs.maximum',
         'minimum': 'interp_ufuncs.minimum',
         'negative': 'interp_ufuncs.negative',
@@ -20,4 +24,7 @@
         'sign': 'interp_ufuncs.sign',
     }
 
-    appleveldefs = {}
+    appleveldefs = {
+        'average': 'app_numpy.average',
+        'mean': 'app_numpy.mean',
+    }
diff --git a/pypy/module/micronumpy/app_numpy.py b/pypy/module/micronumpy/app_numpy.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/micronumpy/app_numpy.py
@@ -0,0 +1,11 @@
+import numpy
+
+def average(a):
+    # This implements a weighted average, for now we don't implement the
+    # weighting, just the average part!
+    return mean(a)
+
+def mean(a):
+    if not hasattr(a, "mean"):
+        a = numpy.array(a)
+    return a.mean()
\ No newline at end of file
diff --git a/pypy/module/micronumpy/compile.py b/pypy/module/micronumpy/compile.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/micronumpy/compile.py
@@ -0,0 +1,49 @@
+
+""" This is a set of tools for standalone compiling of numpy expressions.
+It should not be imported by the module itself
+"""
+
+from pypy.module.micronumpy.interp_numarray import FloatWrapper, SingleDimArray
+
+class BogusBytecode(Exception):
+    pass
+
+def create_array(size):
+    a = SingleDimArray(size)
+    for i in range(size):
+        a.storage[i] = float(i % 10)
+    return a
+
+class TrivialSpace(object):
+    def wrap(self, x):
+        return x
+
+def numpy_compile(bytecode, array_size):
+    space = TrivialSpace()
+    stack = []
+    i = 0
+    for b in bytecode:
+        if b == 'a':
+            stack.append(create_array(array_size))
+            i += 1
+        elif b == 'f':
+            stack.append(FloatWrapper(1.2))
+        elif b == '+':
+            right = stack.pop()
+            stack.append(stack.pop().descr_add(space, right))
+        elif b == '-':
+            right = stack.pop()
+            stack.append(stack.pop().descr_sub(space, right))
+        elif b == '*':
+            right = stack.pop()
+            stack.append(stack.pop().descr_mul(space, right))
+        elif b == '/':
+            right = stack.pop()
+            stack.append(stack.pop().descr_div(space, right))
+        else:
+            print "Unknown opcode: %s" % b
+            raise BogusBytecode()
+    if len(stack) != 1:
+        print "Bogus bytecode, uneven stack length"
+        raise BogusBytecode()
+    return stack[0]
diff --git a/pypy/module/micronumpy/interp_numarray.py b/pypy/module/micronumpy/interp_numarray.py
--- a/pypy/module/micronumpy/interp_numarray.py
+++ b/pypy/module/micronumpy/interp_numarray.py
@@ -1,7 +1,7 @@
 from pypy.interpreter.baseobjspace import ObjSpace, W_Root, Wrappable
 from pypy.interpreter.error import operationerrfmt
 from pypy.interpreter.gateway import interp2app, unwrap_spec
-from pypy.interpreter.typedef import TypeDef
+from pypy.interpreter.typedef import TypeDef, GetSetProperty
 from pypy.rlib import jit
 from pypy.rpython.lltypesystem import lltype
 from pypy.tool.sourcetools import func_with_new_name
@@ -44,9 +44,13 @@
         self.invalidates = []
 
     def invalidated(self):
+        if self.invalidates:
+            self._invalidated()
+
+    def _invalidated(self):
         for arr in self.invalidates:
             arr.force_if_needed()
-        self.invalidates = []
+        del self.invalidates[:]
 
     def _binop_impl(function):
         signature = Signature()
@@ -80,18 +84,36 @@
     def get_concrete(self):
         raise NotImplementedError
 
+    def descr_get_shape(self, space):
+        return space.newtuple([self.descr_len(space)])
+
     def descr_len(self, space):
         return self.get_concrete().descr_len(space)
 
-    @unwrap_spec(item=int)
-    def descr_getitem(self, space, item):
-        return self.get_concrete().descr_getitem(space, item)
+    def descr_getitem(self, space, w_idx):
+        # TODO: indexing by tuples
+        start, stop, step, slice_length = space.decode_index4(w_idx, self.find_size())
+        if step == 0:
+            # Single index
+            return space.wrap(self.get_concrete().getitem(start))
+        else:
+            # Slice
+            res = SingleDimSlice(start, stop, step, slice_length, self, self.signature.transition(SingleDimSlice.static_signature))
+            return space.wrap(res)
 
     @unwrap_spec(item=int, value=float)
     def descr_setitem(self, space, item, value):
         self.invalidated()
         return self.get_concrete().descr_setitem(space, item, value)
 
+    def descr_mean(self, space):
+        s = 0
+        concrete = self.get_concrete()
+        size = concrete.find_size()
+        for i in xrange(size):
+            s += concrete.getitem(i)
+        return space.wrap(s / size)
+
 
 class FloatWrapper(BaseArray):
     """
@@ -119,6 +141,10 @@
         self.forced_result = None
         self.signature = signature
 
+    def _del_sources(self):
+        # Function for deleting references to source arrays, to allow garbage-collecting them
+        raise NotImplementedError
+
     def compute(self):
         i = 0
         signature = self.signature
@@ -135,6 +161,7 @@
     def force_if_needed(self):
         if self.forced_result is None:
             self.forced_result = self.compute()
+            self._del_sources()
 
     def get_concrete(self):
         self.force_if_needed()
@@ -145,6 +172,13 @@
             return self.forced_result.eval(i)
         return self._eval(i)
 
+    def find_size(self):
+        if self.forced_result is not None:
+            # The result has been computed and sources may be unavailable
+            return self.forced_result.find_size()
+        return self._find_size()
+
+
 class Call1(VirtualArray):
     _immutable_fields_ = ["function", "values"]
 
@@ -153,7 +187,10 @@
         self.function = function
         self.values = values
 
-    def find_size(self):
+    def _del_sources(self):
+        self.values = None
+
+    def _find_size(self):
         return self.values.find_size()
 
     def _eval(self, i):
@@ -164,13 +201,18 @@
     Intermediate class for performing binary operations.
     """
     _immutable_fields_ = ["function", "left", "right"]
+
     def __init__(self, function, left, right, signature):
         VirtualArray.__init__(self, signature)
         self.function = function
         self.left = left
         self.right = right
 
-    def find_size(self):
+    def _del_sources(self):
+        self.left = None
+        self.right = None
+
+    def _find_size(self):
         try:
             return self.left.find_size()
         except ValueError:
@@ -181,6 +223,58 @@
         lhs, rhs = self.left.eval(i), self.right.eval(i)
         return self.function(lhs, rhs)
 
+class ViewArray(BaseArray):
+    """
+    Class for representing views of arrays, they will reflect changes of parent
+    arrays. Example: slices
+    """
+    _immutable_fields_ = ["parent"]
+
+    def __init__(self, parent, signature):
+        BaseArray.__init__(self)
+        self.signature = signature
+        self.parent = parent
+        self.invalidates = parent.invalidates
+
+    def get_concrete(self):
+        # in fact, ViewArray never gets "concrete" as it never stores data.
+        # This implementation is needed for BaseArray getitem/setitem to work,
+        # can be refactored.
+        return self
+
+    def eval(self, i):
+        return self.parent.eval(self.calc_index(i))
+
+    def getitem(self, item):
+        return self.parent.getitem(self.calc_index(item))
+
+    @unwrap_spec(item=int, value=float)
+    def descr_setitem(self, space, item, value):
+        return self.parent.descr_setitem(space, self.calc_index(item), value)
+
+    def descr_len(self, space):
+        return space.wrap(self.find_size())
+
+    def calc_index(self, item):
+        raise NotImplementedError
+
+class SingleDimSlice(ViewArray):
+    _immutable_fields_ = ["start", "stop", "step", "size"]
+    static_signature = Signature()
+
+    def __init__(self, start, stop, step, slice_length, parent, signature):
+        ViewArray.__init__(self, parent, signature)
+        self.start = start
+        self.stop = stop
+        self.step = step
+        self.size = slice_length
+
+    def find_size(self):
+        return self.size
+
+    def calc_index(self, item):
+        return (self.start + item * self.step)
+
 
 class SingleDimArray(BaseArray):
     signature = Signature()
@@ -215,10 +309,8 @@
     def descr_len(self, space):
         return space.wrap(self.size)
 
-    @unwrap_spec(item=int)
-    def descr_getitem(self, space, item):
-        item = self.getindex(space, item)
-        return space.wrap(self.storage[item])
+    def getitem(self, item):
+        return self.storage[item]
 
     @unwrap_spec(item=int, value=float)
     def descr_setitem(self, space, item, value):
@@ -238,14 +330,23 @@
         i += 1
     return space.wrap(arr)
 
- at unwrap_spec(ObjSpace, int)
+ at unwrap_spec(size=int)
 def zeros(space, size):
     return space.wrap(SingleDimArray(size))
 
+ at unwrap_spec(size=int)
+def ones(space, size):
+    arr = SingleDimArray(size)
+    for i in xrange(size):
+        arr.storage[i] = 1.0
+    return space.wrap(arr)
 
 BaseArray.typedef = TypeDef(
     'numarray',
     __new__ = interp2app(descr_new_numarray),
+
+    shape = GetSetProperty(BaseArray.descr_get_shape),
+
     __len__ = interp2app(BaseArray.descr_len),
     __getitem__ = interp2app(BaseArray.descr_getitem),
     __setitem__ = interp2app(BaseArray.descr_setitem),
@@ -254,4 +355,6 @@
     __sub__ = interp2app(BaseArray.descr_sub),
     __mul__ = interp2app(BaseArray.descr_mul),
     __div__ = interp2app(BaseArray.descr_div),
-)
\ No newline at end of file
+
+    mean = interp2app(BaseArray.descr_mean),
+)
diff --git a/pypy/module/micronumpy/interp_ufuncs.py b/pypy/module/micronumpy/interp_ufuncs.py
--- a/pypy/module/micronumpy/interp_ufuncs.py
+++ b/pypy/module/micronumpy/interp_ufuncs.py
@@ -8,22 +8,24 @@
 
 def ufunc(func):
     signature = Signature()
-    @unwrap_spec(array=BaseArray)
-    def impl(space, array):
-        w_res = Call1(func, array, array.signature.transition(signature))
-        array.invalidates.append(w_res)
-        return w_res
+    def impl(space, w_obj):
+        if isinstance(w_obj, BaseArray):
+            w_res = Call1(func, w_obj, w_obj.signature.transition(signature))
+            w_obj.invalidates.append(w_res)
+            return w_res
+        return space.wrap(func(space.float_w(w_obj)))
     return func_with_new_name(impl, "%s_dispatcher" % func.__name__)
 
 def ufunc2(func):
     signature = Signature()
-    @unwrap_spec(larray=BaseArray, rarray=BaseArray)
-    def impl(space, larray, rarray):
-        new_sig = larray.signature.transition(signature).transition(rarray.signature)
-        w_res = Call2(func, larray, rarray, new_sig)
-        larray.invalidates.append(w_res)
-        rarray.invalidates.append(w_res)
-        return w_res
+    def impl(space, w_lhs, w_rhs):
+        if isinstance(w_lhs, BaseArray) and isinstance(w_rhs, BaseArray):
+            new_sig = w_lhs.signature.transition(signature).transition(w_rhs.signature)
+            w_res = Call2(func, w_lhs, w_rhs, new_sig)
+            w_lhs.invalidates.append(w_res)
+            w_rhs.invalidates.append(w_res)
+            return w_res
+        return space.wrap(func(space.float_w(w_lhs), space.float_w(w_rhs)))
     return func_with_new_name(impl, "%s_dispatcher" % func.__name__)
 
 @ufunc
@@ -60,6 +62,10 @@
     return 1.0 / value
 
 @ufunc
+def floor(value):
+    return math.floor(value)
+
+ at ufunc
 def sign(value):
     if value == 0.0:
         return 0.0
diff --git a/pypy/module/micronumpy/test/test_base.py b/pypy/module/micronumpy/test/test_base.py
--- a/pypy/module/micronumpy/test/test_base.py
+++ b/pypy/module/micronumpy/test/test_base.py
@@ -16,4 +16,14 @@
         v3 = ar.descr_add(space, FloatWrapper(1.0))
         assert v2.signature is v3.signature
         v4 = ar.descr_add(space, ar)
-        assert v1.signature is v4.signature
\ No newline at end of file
+        assert v1.signature is v4.signature
+
+    def test_slice_signature(self, space):
+        ar = SingleDimArray(10)
+        v1 = ar.descr_getitem(space, space.wrap(slice(1, 5, 1)))
+        v2 = ar.descr_getitem(space, space.wrap(slice(4, 6, 1)))
+        assert v1.signature is v2.signature
+
+        v3 = ar.descr_add(space, v1)
+        v4 = ar.descr_add(space, v2)
+        assert v3.signature is v4.signature
\ No newline at end of file
diff --git a/pypy/module/micronumpy/test/test_module.py b/pypy/module/micronumpy/test/test_module.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/micronumpy/test/test_module.py
@@ -0,0 +1,13 @@
+from pypy.module.micronumpy.test.test_base import BaseNumpyAppTest
+
+
+class AppTestNumPyModule(BaseNumpyAppTest):
+    def test_mean(self):
+        from numpy import array, mean
+        assert mean(array(range(5))) == 2.0
+        assert mean(range(5)) == 2.0
+
+    def test_average(self):
+        from numpy import array, average
+        assert average(range(10)) == 4.5
+        assert average(array(range(10))) == 4.5
\ No newline at end of file
diff --git a/pypy/module/micronumpy/test/test_numarray.py b/pypy/module/micronumpy/test/test_numarray.py
--- a/pypy/module/micronumpy/test/test_numarray.py
+++ b/pypy/module/micronumpy/test/test_numarray.py
@@ -18,6 +18,25 @@
         a[13] = 5.3
         assert a[13] == 5.3
 
+    def test_empty(self):
+        """
+        Test that empty() works.
+        """
+
+        from numpy import empty
+        a = empty(2)
+        a[1] = 1.0
+        assert a[1] == 1.0
+
+    def test_ones(self):
+        from numpy import ones
+        a = ones(3)
+        assert len(a) == 3
+        assert a[0] == 1
+        raises(IndexError, "a[3]")
+        a[2] = 4
+        assert a[2] == 4
+
     def test_iterator_init(self):
         from numpy import array
         a = array(range(5))
@@ -46,6 +65,15 @@
         assert len(a) == 5
         assert len(a + a) == 5
 
+    def test_shape(self):
+        from numpy import array
+        a = array(range(5))
+        assert a.shape == (5,)
+        b = a + a
+        assert b.shape == (5,)
+        c = a[:3]
+        assert c.shape == (3,)
+
     def test_add(self):
         from numpy import array
         a = array(range(5))
@@ -138,4 +166,51 @@
         b = a + a
         c = b + b
         b[1] = 5
-        assert c[1] == 4
\ No newline at end of file
+        assert c[1] == 4
+
+    def test_getslice(self):
+        from numpy import array
+        a = array(range(5))
+        s = a[1:5]
+        assert len(s) == 4
+        for i in range(4):
+            assert s[i] == a[i+1]
+
+    def test_getslice_step(self):
+        from numpy import array
+        a = array(range(10))
+        s = a[1:9:2]
+        assert len(s) == 4
+        for i in range(4):
+            assert s[i] == a[2*i+1]
+
+    def test_slice_update(self):
+        from numpy import array
+        a = array(range(5))
+        s = a[0:3]
+        s[1] = 10
+        assert a[1] == 10
+        a[2] = 20
+        assert s[2] == 20
+
+
+    def test_slice_invaidate(self):
+        # check that slice shares invalidation list with
+        from numpy import array
+        a = array(range(5))
+        s = a[0:2]
+        b = array([10,11])
+        c = s + b
+        a[0] = 100
+        assert c[0] == 10
+        assert c[1] == 12
+        d = s + b
+        a[1] = 101
+        assert d[0] == 110
+        assert d[1] == 12
+
+    def test_mean(self):
+        from numpy import array, mean
+        a = array(range(5))
+        assert a.mean() == 2.0
+        assert a[:4].mean() == 1.5
\ No newline at end of file
diff --git a/pypy/module/micronumpy/test/test_ufuncs.py b/pypy/module/micronumpy/test/test_ufuncs.py
--- a/pypy/module/micronumpy/test/test_ufuncs.py
+++ b/pypy/module/micronumpy/test/test_ufuncs.py
@@ -3,6 +3,13 @@
 
 
 class AppTestUfuncs(BaseNumpyAppTest):
+    def test_single_item(self):
+        from numpy import negative, sign, minimum
+
+        assert negative(5.0) == -5.0
+        assert sign(-0.0) == 0.0
+        assert minimum(2.0, 3.0) == 2.0
+
     def test_negative(self):
         from numpy import array, negative
 
@@ -60,6 +67,15 @@
         for i in range(4):
             assert b[i] == reference[i]
 
+    def test_floor(self):
+        from numpy import array, floor
+
+        reference = [-2.0, -1.0, 0.0, 1.0, 1.0]
+        a = array([-1.4, -1.0, 0.0, 1.0, 1.4])
+        b = floor(a)
+        for i in range(5):
+            assert b[i] == reference[i]
+
     def test_copysign(self):
         from numpy import array, copysign
 
diff --git a/pypy/module/micronumpy/test/test_zjit.py b/pypy/module/micronumpy/test/test_zjit.py
--- a/pypy/module/micronumpy/test/test_zjit.py
+++ b/pypy/module/micronumpy/test/test_zjit.py
@@ -1,8 +1,9 @@
 from pypy.jit.metainterp.test.support import LLJitMixin
+from pypy.rpython.test.test_llinterp import interpret
 from pypy.module.micronumpy.interp_numarray import (SingleDimArray, Signature,
-    FloatWrapper, Call1, Call2, add, mul)
+    FloatWrapper, Call1, Call2, SingleDimSlice, add, mul)
 from pypy.module.micronumpy.interp_ufuncs import negative
-
+from pypy.module.micronumpy.compile import numpy_compile
 
 class FakeSpace(object):
     pass
@@ -91,4 +92,54 @@
 
         self.meta_interp(f, [5], listops=True, backendopt=True)
         # This is 3, not 2 because there is a bridge for the exit.
-        self.check_loop_count(3)
\ No newline at end of file
+        self.check_loop_count(3)
+
+    def test_slice(self):
+        space = self.space
+
+        def f(i):
+            step = 3
+            ar = SingleDimArray(step*i)
+            s = SingleDimSlice(0, step*i, step, i, ar, ar.signature.transition(SingleDimSlice.static_signature))
+            v = Call2(add, s, s, Signature())
+            return v.get_concrete().storage[3]
+
+        result = self.meta_interp(f, [5], listops=True, backendopt=True)
+        self.check_loops({'int_mul': 1, 'getarrayitem_raw': 2, 'float_add': 1,
+                          'setarrayitem_raw': 1, 'int_add': 1,
+                          'int_lt': 1, 'guard_true': 1, 'jump': 1})
+        assert result == f(5)
+
+    def test_slice2(self):
+        space = self.space
+
+        def f(i):
+            step1 = 2
+            step2 = 3
+            ar = SingleDimArray(step2*i)
+            s1 = SingleDimSlice(0, step1*i, step1, i, ar, ar.signature.transition(SingleDimSlice.static_signature))
+            s2 = SingleDimSlice(0, step2*i, step2, i, ar, ar.signature.transition(SingleDimSlice.static_signature))
+            v = Call2(add, s1, s2, Signature())
+            return v.get_concrete().storage[3]
+
+        result = self.meta_interp(f, [5], listops=True, backendopt=True)
+        self.check_loops({'int_mul': 2, 'getarrayitem_raw': 2, 'float_add': 1,
+                          'setarrayitem_raw': 1, 'int_add': 1,
+                          'int_lt': 1, 'guard_true': 1, 'jump': 1})
+        assert result == f(5)
+
+class TestTranslation(object):
+    def test_compile(self):
+        x = numpy_compile('aa+f*f/a-', 10)
+        x = x.compute()
+        assert isinstance(x, SingleDimArray)
+        assert x.size == 10
+        assert x.storage[0] == 0
+        assert x.storage[1] == ((1 + 1) * 1.2) / 1.2 - 1
+    
+    def test_translation(self):
+        # we import main to check if the target compiles
+        from pypy.translator.goal.targetnumpystandalone import main
+        from pypy.rpython.annlowlevel import llstr
+        
+        interpret(main, [llstr('af+'), 100])
diff --git a/pypy/module/operator/app_operator.py b/pypy/module/operator/app_operator.py
--- a/pypy/module/operator/app_operator.py
+++ b/pypy/module/operator/app_operator.py
@@ -4,6 +4,7 @@
 This module exports a set of operators as functions. E.g. operator.add(x,y) is
 equivalent to x+y.
 '''
+from __pypy__ import builtinify
 
 def countOf(a,b): 
     'countOf(a, b) -- Return the number of times b occurs in a.'
@@ -66,50 +67,56 @@
     a[b:c] = d 
 __setslice__ = setslice
 
-class attrgetter(object):
 
-    def __init__(self, attr, *attrs):
-        self.attrs = (attr,) + attrs
+def attrgetter(attr, *attrs):
+    if attrs:
+        getters = [single_attr_getter(a) for a in (attr,) + attrs]
+        def getter(obj):
+            return tuple([getter(obj) for getter in getters])
+    else:
+        getter = single_attr_getter(attr)
+    return builtinify(getter)
 
-    def _resolve_attr(self, obj, attr):
-        last = 0
-        while True:
-            try:
-                dot = attr.find(".", last)
-            except AttributeError:
-                raise TypeError
-            if dot > 0:
-                obj = getattr(obj, attr[last:dot])
-                last = dot + 1
-            else:
-                return getattr(obj, attr[last:])
+def single_attr_getter(attr):
+    if not isinstance(attr, str):
+        if not isinstance(attr, unicode):
+            def _raise_typeerror(obj):
+                raise TypeError("argument must be a string, not %r" %
+                                (type(attr).__name__,))
+            return _raise_typeerror
+        attr = attr.encode('ascii')
+    #
+    def make_getter(name, prevfn=None):
+        if prevfn is None:
+            def getter(obj):
+                return getattr(obj, name)
+        else:
+            def getter(obj):
+                return getattr(prevfn(obj), name)
+        return getter
+    #
+    last = 0
+    getter = None
+    while True:
+        dot = attr.find(".", last)
+        if dot < 0: break
+        getter = make_getter(attr[last:dot], getter)
+        last = dot + 1
+    return make_getter(attr[last:], getter)
 
-    def __call__(self, obj):
-        if len(self.attrs) == 1:
-            return self._resolve_attr(obj, self.attrs[0])
-        return tuple(self._resolve_attr(obj, attr) for attr in self.attrs)
 
-class itemgetter(object):
+def itemgetter(item, *items):
+    if items:
+        list_of_indices = [item] + list(items)
+        def getter(obj):
+            return tuple([obj[i] for i in list_of_indices])
+    else:
+        def getter(obj):
+            return obj[item]
+    return builtinify(getter)
 
-    def __init__(self, item, *args):
-        self.items = args
-        self.item = item
 
-    def __call__(self, obj):
-        result = obj[self.item]
-
-        if self.items:
-            list = [result] + [obj[item] for item in self.items]
-            return tuple(list)
-
-        return result
-
-class methodcaller(object):
-
-    def __init__(self, method_name, *args, **kwargs):
-        self.method_name = method_name
-        self.args = args
-        self.kwargs = kwargs
-
-    def __call__(self, obj):
-        return getattr(obj, self.method_name)(*self.args, **self.kwargs)
+def methodcaller(method_name, *args, **kwargs):
+    def call(obj):
+        return getattr(obj, method_name)(*args, **kwargs)
+    return builtinify(call)
diff --git a/pypy/module/posix/app_posix.py b/pypy/module/posix/app_posix.py
--- a/pypy/module/posix/app_posix.py
+++ b/pypy/module/posix/app_posix.py
@@ -107,6 +107,9 @@
 def tmpnam():
     """Return an absolute pathname of a file that did not exist at the
     time the call is made."""
+    from warnings import warn
+    warn(RuntimeWarning("tmpnam is a potential security risk to your program"))
+
     import tempfile
     return tempfile.mktemp()
 
@@ -114,6 +117,9 @@
     """Return an absolute pathname of a file that did not exist at the
     time the call is made.  The directory and a prefix may be specified
     as strings; they may be omitted or None if not needed."""
+    from warnings import warn
+    warn(RuntimeWarning("tempnam is a potential security risk to your program"))
+
     import tempfile
     return tempfile.mktemp('', prefix or 'tmp', dir)
 
diff --git a/pypy/module/posix/test/test_posix2.py b/pypy/module/posix/test/test_posix2.py
--- a/pypy/module/posix/test/test_posix2.py
+++ b/pypy/module/posix/test/test_posix2.py
@@ -847,6 +847,21 @@
                 assert os.path.basename(s1).startswith(prefix or 'tmp')
                 assert os.path.basename(s2).startswith(prefix or 'tmp')
 
+    def test_tmpnam_warning(self):
+        import warnings, os
+        #
+        def f_tmpnam_warning(): os.tmpnam()    # a single line
+        #
+        with warnings.catch_warnings(record=True) as w:
+            warnings.simplefilter("always")
+            f_tmpnam_warning()
+            assert len(w) == 1
+            assert issubclass(w[-1].category, RuntimeWarning)
+            assert "potential security risk" in str(w[-1].message)
+            # check that the warning points to the call to os.tmpnam(),
+            # not to some code inside app_posix.py
+            assert w[-1].lineno == f_tmpnam_warning.func_code.co_firstlineno
+
 
 class AppTestEnvironment(object):
     def setup_class(cls):
diff --git a/pypy/module/pypyjit/__init__.py b/pypy/module/pypyjit/__init__.py
--- a/pypy/module/pypyjit/__init__.py
+++ b/pypy/module/pypyjit/__init__.py
@@ -8,6 +8,7 @@
         'set_param':    'interp_jit.set_param',
         'residual_call': 'interp_jit.residual_call',
         'set_compile_hook': 'interp_jit.set_compile_hook',
+        'DebugMergePoint': 'interp_resop.W_DebugMergePoint',
     }
 
     def setup_after_space_initialization(self):
diff --git a/pypy/module/pypyjit/interp_jit.py b/pypy/module/pypyjit/interp_jit.py
--- a/pypy/module/pypyjit/interp_jit.py
+++ b/pypy/module/pypyjit/interp_jit.py
@@ -17,10 +17,11 @@
 from opcode import opmap
 from pypy.rlib.objectmodel import we_are_translated
 from pypy.rlib.nonconst import NonConstant
+from pypy.jit.metainterp.resoperation import rop
+from pypy.module.pypyjit.interp_resop import debug_merge_point_from_boxes
 
 PyFrame._virtualizable2_ = ['last_instr', 'pycode',
-                            'valuestackdepth', 'valuestack_w[*]',
-                            'fastlocals_w[*]',
+                            'valuestackdepth', 'locals_stack_w[*]',
                             'last_exception',
                             'lastblock',
                             'is_being_profiled',
@@ -47,6 +48,16 @@
     return (bytecode.co_flags & CO_GENERATOR) != 0
 
 
+def wrap_oplist(space, logops, operations):
+    list_w = []
+    for op in operations:
+        if op.getopnum() == rop.DEBUG_MERGE_POINT:
+            list_w.append(space.wrap(debug_merge_point_from_boxes(
+                op.getarglist())))
+        else:
+            list_w.append(space.wrap(logops.repr_of_resop(op)))
+    return list_w
+
 class PyPyJitDriver(JitDriver):
     reds = ['frame', 'ec']
     greens = ['next_instr', 'is_being_profiled', 'pycode']
@@ -62,8 +73,7 @@
             return
         if space.is_true(cache.w_compile_hook):
             logops = logger._make_log_operations()
-            list_w = [space.wrap(logops.repr_of_resop(op))
-                      for op in operations]
+            list_w = wrap_oplist(space, logops, operations)
             pycode = cast_base_ptr_to_instance(PyCode, ll_pycode)
             cache.in_recursion = True
             try:
@@ -85,8 +95,7 @@
             return
         if space.is_true(cache.w_compile_hook):
             logops = logger._make_log_operations()
-            list_w = [space.wrap(logops.repr_of_resop(op))
-                      for op in operations]
+            list_w = wrap_oplist(space, logops, operations)
             cache.in_recursion = True
             try:
                 space.call_function(cache.w_compile_hook,
diff --git a/pypy/module/pypyjit/interp_resop.py b/pypy/module/pypyjit/interp_resop.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/pypyjit/interp_resop.py
@@ -0,0 +1,41 @@
+
+from pypy.interpreter.typedef import TypeDef, interp_attrproperty
+from pypy.interpreter.baseobjspace import Wrappable, ObjSpace, W_Root
+from pypy.interpreter.gateway import unwrap_spec, interp2app
+from pypy.interpreter.pycode import PyCode
+from pypy.rpython.lltypesystem import lltype, llmemory
+from pypy.rpython.annlowlevel import cast_base_ptr_to_instance
+from pypy.rpython.lltypesystem.rclass import OBJECT
+
+class W_DebugMergePoint(Wrappable):
+    """ A class representing debug_merge_point JIT operation
+    """
+
+    def __init__(self, mp_no, offset, pycode):
+        self.mp_no = mp_no
+        self.offset = offset
+        self.pycode = pycode
+
+    def descr_repr(self, space):
+        return space.wrap('DebugMergePoint()')
+
+ at unwrap_spec(mp_no=int, offset=int, pycode=PyCode)
+def new_debug_merge_point(space, w_tp, mp_no, offset, pycode):
+    return W_DebugMergePoint(mp_no, offset, pycode)
+
+def debug_merge_point_from_boxes(boxes):
+    mp_no = boxes[0].getint()
+    offset = boxes[2].getint()
+    llcode = lltype.cast_opaque_ptr(lltype.Ptr(OBJECT),
+                                    boxes[4].getref_base())
+    pycode = cast_base_ptr_to_instance(PyCode, llcode)
+    assert pycode is not None
+    return W_DebugMergePoint(mp_no, offset, pycode)
+
+W_DebugMergePoint.typedef = TypeDef(
+    'DebugMergePoint',
+    __new__ = interp2app(new_debug_merge_point),
+    __doc__ = W_DebugMergePoint.__doc__,
+    __repr__ = interp2app(W_DebugMergePoint.descr_repr),
+    code = interp_attrproperty('pycode', W_DebugMergePoint),
+)
diff --git a/pypy/module/pypyjit/policy.py b/pypy/module/pypyjit/policy.py
--- a/pypy/module/pypyjit/policy.py
+++ b/pypy/module/pypyjit/policy.py
@@ -14,7 +14,8 @@
             modname, _ = modname.split('.', 1)
         if modname in ['pypyjit', 'signal', 'micronumpy', 'math', 'exceptions',
                        'imp', 'sys', 'array', '_ffi', 'itertools', 'operator',
-                       'posix', '_socket', '_sre', '_lsprof', 'cppyy', '_weakref']:
+                       'posix', '_socket', '_sre', '_lsprof', 'cppyy', '_weakref',
+                       '__pypy__']:
             return True
         return False
 
diff --git a/pypy/module/pypyjit/test/test_jit_hook.py b/pypy/module/pypyjit/test/test_jit_hook.py
--- a/pypy/module/pypyjit/test/test_jit_hook.py
+++ b/pypy/module/pypyjit/test/test_jit_hook.py
@@ -8,12 +8,13 @@
 from pypy.jit.metainterp.logger import Logger
 from pypy.rpython.annlowlevel import (cast_instance_to_base_ptr,
                                       cast_base_ptr_to_instance)
+from pypy.rpython.lltypesystem import lltype, llmemory
 from pypy.module.pypyjit.interp_jit import pypyjitdriver
 from pypy.jit.tool.oparser import parse
 from pypy.jit.metainterp.typesystem import llhelper
 
 class MockSD(object):
-    class cpu:
+    class cpu(object):
         ts = llhelper
 
 class AppTestJitHook(object):
@@ -27,14 +28,17 @@
             pass
         return f
         """)
+        cls.w_f = w_f
         ll_code = cast_instance_to_base_ptr(w_f.code)
+        code_gcref = lltype.cast_opaque_ptr(llmemory.GCREF, ll_code)
         logger = Logger(MockSD())
 
         oplist = parse("""
         [i1, i2]
         i3 = int_add(i1, i2)
+        debug_merge_point(0, 0, 0, 0, ConstPtr(ptr0))
         guard_true(i3) []
-        """).operations
+        """, namespace={'ptr0': code_gcref}).operations
 
         def interp_on_compile():
             pypyjitdriver.on_compile(logger, LoopToken(), oplist, 'loop',
@@ -63,7 +67,7 @@
         assert all[0][0][0].co_name == 'f'
         assert all[0][0][1] == 0
         assert all[0][0][2] == False
-        assert len(all[0][1]) == 2
+        assert len(all[0][1]) == 3
         assert 'int_add' in all[0][1][0]
         self.on_compile_bridge()
         assert len(all) == 2
@@ -103,3 +107,20 @@
         self.on_compile_bridge()
         assert len(l) == 2 # and did not crash
         
+    def test_on_compile_types(self):
+        import pypyjit
+        l = []
+
+        def hook(*args):
+            l.append(args)
+
+        pypyjit.set_compile_hook(hook)
+        self.on_compile()
+        dmp = l[0][3][1]
+        assert isinstance(dmp, pypyjit.DebugMergePoint)
+        assert dmp.code is self.f.func_code
+
+    def test_creation(self):
+        import pypyjit
+        dmp = pypyjit.DebugMergePoint(0, 0, self.f.func_code)
+        assert dmp.code is self.f.func_code 
diff --git a/pypy/module/pypyjit/test_pypy_c/model.py b/pypy/module/pypyjit/test_pypy_c/model.py
--- a/pypy/module/pypyjit/test_pypy_c/model.py
+++ b/pypy/module/pypyjit/test_pypy_c/model.py
@@ -2,6 +2,7 @@
 import sys
 import re
 import os.path
+from _pytest.assertion import newinterpret
 from pypy.tool.jitlogparser.parser import SimpleParser, Function, TraceForOpcode
 from pypy.tool.jitlogparser.storage import LoopStorage
 
@@ -194,7 +195,7 @@
             # transform self._assert(x, 'foo') into assert x, 'foo'
             source = source.replace('self._assert(', 'assert ')
             source = source[:-1] # remove the trailing ')'
-            self.msg = py.code._reinterpret(source, f, should_fail=True)
+            self.msg = newinterpret.interpret(source, f, should_fail=True)
         else:
             self.msg = "<could not determine information>"
 
diff --git a/pypy/module/pypyjit/test_pypy_c/test_00_model.py b/pypy/module/pypyjit/test_pypy_c/test_00_model.py
--- a/pypy/module/pypyjit/test_pypy_c/test_00_model.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_00_model.py
@@ -58,6 +58,8 @@
         stdout, stderr = pipe.communicate()
         if stderr.startswith('SKIP:'):
             py.test.skip(stderr)
+        if stderr.startswith('debug_alloc.h:'):   # lldebug builds
+            stderr = ''
         assert not stderr
         #
         # parse the JIT log
diff --git a/pypy/module/pypyjit/test_pypy_c/test_array.py b/pypy/module/pypyjit/test_pypy_c/test_array.py
--- a/pypy/module/pypyjit/test_pypy_c/test_array.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_array.py
@@ -46,7 +46,7 @@
             guard_no_overflow(descr=<Guard4>)
             i18 = int_add(i7, 1)
             --TICK--
-            jump(p0, p1, p2, p3, p4, p5, p6, i18, i16, i9, i10, descr=<Loop0>)
+            jump(p0, p1, p2, p3, p4, p5, i18, i16, p8, i9, i10, descr=<Loop0>)
         """)
 
     def test_array_intimg(self):
@@ -83,7 +83,7 @@
             setarrayitem_raw(i11, i8, _, descr=<.*ArrayNoLengthDescr>)
             i28 = int_add(i8, 1)
             --TICK--
-            jump(p0, p1, p2, p3, p4, p5, p6, p7, i28, i15, i10, i11, descr=<Loop0>)
+            jump(p0, p1, p2, p3, p4, p5, p6, i28, i15, p9, i10, i11, descr=<Loop0>)
         """)
 
 
diff --git a/pypy/module/pypyjit/test_pypy_c/test_call.py b/pypy/module/pypyjit/test_pypy_c/test_call.py
--- a/pypy/module/pypyjit/test_pypy_c/test_call.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_call.py
@@ -11,21 +11,14 @@
                 return 1 + rec(n-1)
             #
             # this loop is traced and then aborted, because the trace is too
-            # long. But then "rec" is marked as "don't inline"
-            i = 0
-            j = 0
-            while i < 20:
-                i += 1
-                j += rec(100)
-            #
-            # next time we try to trace "rec", instead of inlining we compile
-            # it separately and generate a call_assembler
+            # long. But then "rec" is marked as "don't inline". Since we
+            # already traced function from the start (because of number),
+            # now we can inline it as call assembler
             i = 0
             j = 0
             while i < 20:
                 i += 1
                 j += rec(100) # ID: call_rec
-                a = 0
             return j
         #
         log = self.run(fn, [], threshold=18)
@@ -38,6 +31,20 @@
             ...
         """)
 
+    def test_fib(self):
+        def fib(n):
+            if n == 0 or n == 1:
+                return 1
+            return fib(n - 1) + fib(n - 2) # ID: call_rec
+
+        log = self.run(fib, [7], function_threshold=15)
+        loop, = log.loops_by_filename(self.filepath, is_entry_bridge='*')
+        #assert loop.match_by_id('call_rec', '''
+        #...
+        #p1 = call_assembler(..., descr=...)
+        #...
+        #''')
+
     def test_simple_call(self):
         src = """
             OFFSET = 0
@@ -180,7 +187,7 @@
             guard_no_overflow(descr=<Guard5>)
             i18 = force_token()
             --TICK--
-            jump(p0, p1, p2, p3, p4, p5, i8, p7, i17, i9, p10, p11, p12, descr=<Loop0>)
+            jump(p0, p1, p2, p3, p4, i8, p7, i17, p8, i9, p10, p11, p12, descr=<Loop0>)
         """)
 
     def test_default_and_kw(self):
diff --git a/pypy/module/pypyjit/test_pypy_c/test_instance.py b/pypy/module/pypyjit/test_pypy_c/test_instance.py
--- a/pypy/module/pypyjit/test_pypy_c/test_instance.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_instance.py
@@ -115,7 +115,6 @@
         # ----------------------
         loop, = log.loops_by_filename(self.filepath)
         assert loop.match("""
-            i8 = getfield_gc_pure(p5, descr=<SignedFieldDescr .*W_IntObject.inst_intval.*>)
             i9 = int_lt(i8, i7)
             guard_true(i9, descr=.*)
             guard_not_invalidated(descr=.*)
@@ -125,7 +124,7 @@
             p20 = new_with_vtable(ConstClass(W_IntObject))
             setfield_gc(p20, i11, descr=<SignedFieldDescr.*W_IntObject.inst_intval .*>)
             setfield_gc(ConstPtr(ptr21), p20, descr=<GcPtrFieldDescr .*TypeCell.inst_w_value .*>)
-            jump(p0, p1, p2, p3, p4, p20, p6, i7, descr=<Loop.>)
+            jump(p0, p1, p2, p3, p4, p20, p6, i11, i7, descr=<Loop.>)
         """)
 
     def test_oldstyle_newstyle_mix(self):
diff --git a/pypy/module/pypyjit/test_pypy_c/test_intbound.py b/pypy/module/pypyjit/test_pypy_c/test_intbound.py
--- a/pypy/module/pypyjit/test_pypy_c/test_intbound.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_intbound.py
@@ -97,7 +97,7 @@
             guard_no_overflow(descr=...)
             i17 = int_add(i8, 1)
             --TICK--
-            jump(p0, p1, p2, p3, p4, p5, i14, i12, i17, i9, descr=<Loop0>)
+            jump(p0, p1, p2, p3, p4, i14, i12, i17, p8, i9, descr=<Loop0>)
         """)
 
     def test_intbound_sub_lt(self):
@@ -149,7 +149,7 @@
             guard_no_overflow(descr=...)
             i19 = int_add(i8, 1)
             --TICK--
-            jump(p0, p1, p2, p3, p4, p5, i16, i14, i19, i9, descr=<Loop0>)
+            jump(p0, p1, p2, p3, p4, i16, i14, i19, p8, i9, descr=<Loop0>)
         """)
 
     def test_intbound_addmul_ge(self):
@@ -177,7 +177,7 @@
             guard_no_overflow(descr=...)
             i21 = int_add(i8, 1)
             --TICK--
-            jump(p0, p1, p2, p3, p4, p5, i18, i14, i21, descr=<Loop0>)
+            jump(p0, p1, p2, p3, p4, i18, i14, i21, p8, descr=<Loop0>)
         """)
 
     def test_intbound_eq(self):
@@ -209,7 +209,7 @@
             guard_no_overflow(descr=...)
             i16 = int_add(i8, 1)
             --TICK--
-            jump(p0, p1, p2, p3, p4, p5, p6, i14, i16, descr=<Loop0>)
+            jump(p0, p1, p2, p3, p4, p6, i14, i16, p8, descr=<Loop0>)
         """)
 
     def test_intbound_mul(self):
diff --git a/pypy/module/pypyjit/test_pypy_c/test_misc.py b/pypy/module/pypyjit/test_pypy_c/test_misc.py
--- a/pypy/module/pypyjit/test_pypy_c/test_misc.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_misc.py
@@ -167,7 +167,7 @@
             guard_false(i16, descr=<Guard3>)
             p17 = getarrayitem_gc(p15, i12, descr=<GcPtrArrayDescr>)
             i19 = int_add(i12, 1)
-            setfield_gc(p4, i19, descr=<SignedFieldDescr .*W_AbstractSeqIterObject.inst_index .*>)
+            setfield_gc(p9, i19, descr=<SignedFieldDescr .*W_AbstractSeqIterObject.inst_index .*>)
             guard_nonnull_class(p17, 146982464, descr=<Guard4>)
             i21 = getfield_gc(p17, descr=<SignedFieldDescr .*W_ArrayTypei.inst_len .*>)
             i23 = int_lt(0, i21)
@@ -179,7 +179,7 @@
             i28 = int_add_ovf(i10, i25)
             guard_no_overflow(descr=<Guard7>)
             --TICK--
-            jump(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, i28, i25, i19, i13, p14, p15, descr=<Loop0>)
+            jump(p0, p1, p2, p3, p4, p5, p6, i28, i25, p9, p10, p11, i19, i13, p14, p15, descr=<Loop0>)
         """)
 
 
diff --git a/pypy/module/pypyjit/test_pypy_c/test_string.py b/pypy/module/pypyjit/test_pypy_c/test_string.py
new file mode 100644
--- /dev/null
+++ b/pypy/module/pypyjit/test_pypy_c/test_string.py
@@ -0,0 +1,42 @@
+from pypy.module.pypyjit.test_pypy_c.test_00_model import BaseTestPyPyC
+
+class TestString(BaseTestPyPyC):
+    def test_lookup_default_encoding(self):
+        def main(n):
+            import string
+            i = 0
+            letters = string.letters
+            uletters = unicode(string.letters)
+            while i < n:
+                i += letters[i % len(letters)] == uletters[i % len(letters)]
+            return i
+
+        log = self.run(main, [300])
+        assert log.result == 300
+        loop, = log.loops_by_filename(self.filepath)
+        assert loop.match("""
+            i14 = int_lt(i6, i9)
+            guard_true(i14, descr=<Guard42>)
+            i15 = int_mod(i6, i10)
+            i17 = int_rshift(i15, 63)
+            i18 = int_and(i10, i17)
+            i19 = int_add(i15, i18)
+            i21 = int_lt(i19, 0)
+            guard_false(i21, descr=<Guard43>)
+            i22 = int_ge(i19, i10)
+            guard_false(i22, descr=<Guard44>)
+            i23 = strgetitem(p11, i19)
+            i24 = int_ge(i19, i12)
+            guard_false(i24, descr=<Guard45>)
+            i25 = unicodegetitem(p13, i19)
+            guard_not_invalidated(descr=<Guard46>)
+            p27 = newstr(1)
+            strsetitem(p27, 0, i23)
+            p30 = call(ConstClass(ll_str2unicode__rpy_stringPtr), p27, descr=<GcPtrCallDescr>)
+            guard_no_exception(descr=<Guard47>)
+            i32 = call(ConstClass(_ll_2_str_eq_checknull_char__rpy_unicodePtr_UniChar), p30, i25, descr=<SignedCallDescr>)
+            guard_true(i32, descr=<Guard48>)
+            i34 = int_add(i6, 1)
+            --TICK--
+            jump(p0, p1, p2, p3, p4, p5, i34, p7, p8, i9, i10, p11, i12, p13, descr=<Loop4>)
+        """)
\ No newline at end of file
diff --git a/pypy/module/sys/__init__.py b/pypy/module/sys/__init__.py
--- a/pypy/module/sys/__init__.py
+++ b/pypy/module/sys/__init__.py
@@ -7,6 +7,8 @@
 
 class Module(MixedModule):
     """Sys Builtin Module. """
+    _immutable_fields_ = ["defaultencoding?"]
+
     def __init__(self, space, w_name):
         """NOT_RPYTHON""" # because parent __init__ isn't
         if space.config.translating:
diff --git a/pypy/objspace/flow/flowcontext.py b/pypy/objspace/flow/flowcontext.py
--- a/pypy/objspace/flow/flowcontext.py
+++ b/pypy/objspace/flow/flowcontext.py
@@ -384,8 +384,9 @@
     # hack for unrolling iterables, don't use this
     def replace_in_stack(self, oldvalue, newvalue):
         w_new = Constant(newvalue)
-        stack_items_w = self.crnt_frame.valuestack_w
-        for i in range(self.crnt_frame.valuestackdepth-1, -1, -1):
+        f = self.crnt_frame
+        stack_items_w = f.locals_stack_w
+        for i in range(f.valuestackdepth-1, f.nlocals-1, -1):
             w_v = stack_items_w[i]
             if isinstance(w_v, Constant):
                 if w_v.value is oldvalue:
diff --git a/pypy/objspace/flow/framestate.py b/pypy/objspace/flow/framestate.py
--- a/pypy/objspace/flow/framestate.py
+++ b/pypy/objspace/flow/framestate.py
@@ -10,7 +10,7 @@
     def __init__(self, state):
         if isinstance(state, PyFrame):
             # getfastscope() can return real None, for undefined locals
-            data = state.getfastscope() + state.savevaluestack()
+            data = state.save_locals_stack()
             if state.last_exception is None:
                 data.append(Constant(None))
                 data.append(Constant(None))
@@ -36,11 +36,9 @@
 
     def restoreframe(self, frame):
         if isinstance(frame, PyFrame):
-            fastlocals = len(frame.fastlocals_w)
             data = self.mergeable[:]
             recursively_unflatten(frame.space, data)
-            frame.setfastscope(data[:fastlocals])  # Nones == undefined locals
-            frame.restorevaluestack(data[fastlocals:-2])
+            frame.restore_locals_stack(data[:-2])  # Nones == undefined locals
             if data[-2] == Constant(None):
                 assert data[-1] == Constant(None)
                 frame.last_exception = None
diff --git a/pypy/objspace/flow/test/test_framestate.py b/pypy/objspace/flow/test/test_framestate.py
--- a/pypy/objspace/flow/test/test_framestate.py
+++ b/pypy/objspace/flow/test/test_framestate.py
@@ -25,7 +25,7 @@
         dummy = Constant(None)
         #dummy.dummy = True
         arg_list = ([Variable() for i in range(formalargcount)] +
-                    [dummy] * (len(frame.fastlocals_w) - formalargcount))
+                    [dummy] * (frame.nlocals - formalargcount))
         frame.setfastscope(arg_list)
         return frame
 
@@ -42,7 +42,7 @@
     def test_neq_hacked_framestate(self):
         frame = self.getframe(self.func_simple)
         fs1 = FrameState(frame)
-        frame.fastlocals_w[-1] = Variable()
+        frame.locals_stack_w[frame.nlocals-1] = Variable()
         fs2 = FrameState(frame)
         assert fs1 != fs2
 
@@ -55,7 +55,7 @@
     def test_union_on_hacked_framestates(self):
         frame = self.getframe(self.func_simple)
         fs1 = FrameState(frame)
-        frame.fastlocals_w[-1] = Variable()
+        frame.locals_stack_w[frame.nlocals-1] = Variable()
         fs2 = FrameState(frame)
         assert fs1.union(fs2) == fs2  # fs2 is more general
         assert fs2.union(fs1) == fs2  # fs2 is more general
@@ -63,7 +63,7 @@
     def test_restore_frame(self):
         frame = self.getframe(self.func_simple)
         fs1 = FrameState(frame)
-        frame.fastlocals_w[-1] = Variable()
+        frame.locals_stack_w[frame.nlocals-1] = Variable()
         fs1.restoreframe(frame)
         assert fs1 == FrameState(frame)
 
@@ -82,25 +82,26 @@
     def test_getoutputargs(self):
         frame = self.getframe(self.func_simple)
         fs1 = FrameState(frame)
-        frame.fastlocals_w[-1] = Variable()
+        frame.locals_stack_w[frame.nlocals-1] = Variable()
         fs2 = FrameState(frame)
         outputargs = fs1.getoutputargs(fs2)
         # 'x' -> 'x' is a Variable
-        # fastlocals_w[-1] -> fastlocals_w[-1] is Constant(None)
-        assert outputargs == [frame.fastlocals_w[0], Constant(None)]
+        # locals_w[n-1] -> locals_w[n-1] is Constant(None)
+        assert outputargs == [frame.locals_stack_w[0], Constant(None)]
 
     def test_union_different_constants(self):
         frame = self.getframe(self.func_simple)
         fs1 = FrameState(frame)
-        frame.fastlocals_w[-1] = Constant(42)
+        frame.locals_stack_w[frame.nlocals-1] = Constant(42)
         fs2 = FrameState(frame)
         fs3 = fs1.union(fs2)
         fs3.restoreframe(frame)
-        assert isinstance(frame.fastlocals_w[-1], Variable) # generalized
+        assert isinstance(frame.locals_stack_w[frame.nlocals-1], Variable)
+                                 # ^^^ generalized
 
     def test_union_spectag(self):
         frame = self.getframe(self.func_simple)
         fs1 = FrameState(frame)
-        frame.fastlocals_w[-1] = Constant(SpecTag())
+        frame.locals_stack_w[frame.nlocals-1] = Constant(SpecTag())
         fs2 = FrameState(frame)
         assert fs1.union(fs2) is None   # UnionError
diff --git a/pypy/objspace/std/dictmultiobject.py b/pypy/objspace/std/dictmultiobject.py
--- a/pypy/objspace/std/dictmultiobject.py
+++ b/pypy/objspace/std/dictmultiobject.py
@@ -1,13 +1,14 @@
 import py, sys
 from pypy.objspace.std.model import registerimplementation, W_Object
 from pypy.objspace.std.register_all import register_all
+from pypy.objspace.std.settype import set_typedef as settypedef
 from pypy.interpreter import gateway
 from pypy.interpreter.argument import Signature
 from pypy.interpreter.error import OperationError, operationerrfmt
 from pypy.module.__builtin__.__init__ import BUILTIN_TO_INDEX, OPTIMIZED_BUILTINS
 
 from pypy.rlib.objectmodel import r_dict, we_are_translated
-from pypy.objspace.std.settype import set_typedef as settypedef
+from pypy.rlib.debug import mark_dict_non_null
 
 def _is_str(space, w_key):
     return space.is_w(space.type(w_key), space.w_str)
@@ -59,7 +60,8 @@
 
     def initialize_as_rdict(self):
         assert self.r_dict_content is None
-        self.r_dict_content = r_dict(self.space.eq_w, self.space.hash_w)
+        self.r_dict_content = r_dict(self.space.eq_w, self.space.hash_w,
+                                     force_non_null=True)
         return self.r_dict_content
 
 
@@ -308,6 +310,7 @@
     def __init__(self, space):
         self.space = space
         self.content = {}
+        mark_dict_non_null(self.content)
 
     def impl_setitem(self, w_key, w_value):
         space = self.space
@@ -317,6 +320,7 @@
             self._as_rdict().impl_fallback_setitem(w_key, w_value)
 
     def impl_setitem_str(self, key, w_value):
+        assert key is not None
         self.content[key] = w_value
 
     def impl_setdefault(self, w_key, w_default):
@@ -342,6 +346,7 @@
         return len(self.content)
 
     def impl_getitem_str(self, key):
+        assert key is not None
         return self.content.get(key, None)
 
     def impl_getitem(self, w_key):
diff --git a/pypy/objspace/std/objspace.py b/pypy/objspace/std/objspace.py
--- a/pypy/objspace/std/objspace.py
+++ b/pypy/objspace/std/objspace.py
@@ -311,6 +311,10 @@
                 classofinstance=classofinstance,
                 strdict=strdict)
 
+    def newset(self):
+        from pypy.objspace.std.setobject import newset
+        return W_SetObject(self, newset(self))
+
     def newslice(self, w_start, w_end, w_step):
         return W_SliceObject(w_start, w_end, w_step)
 
diff --git a/pypy/objspace/std/setobject.py b/pypy/objspace/std/setobject.py
--- a/pypy/objspace/std/setobject.py
+++ b/pypy/objspace/std/setobject.py
@@ -112,7 +112,7 @@
 # some helper functions
 
 def newset(space):
-    return r_dict(space.eq_w, space.hash_w)
+    return r_dict(space.eq_w, space.hash_w, force_non_null=True)
 
 def make_setdata_from_w_iterable(space, w_iterable=None):
     """Return a new r_dict with the content of w_iterable."""
diff --git a/pypy/objspace/std/test/test_setobject.py b/pypy/objspace/std/test/test_setobject.py
--- a/pypy/objspace/std/test/test_setobject.py
+++ b/pypy/objspace/std/test/test_setobject.py
@@ -50,6 +50,10 @@
         u = self.space.wrap(set('simsalabim'))
         assert self.space.eq_w(s,u)
 
+    def test_space_newset(self):
+        s = self.space.newset()
+        assert self.space.str_w(self.space.repr(s)) == 'set([])'
+
 class AppTestAppSetTest:
     def test_subtype(self):
         class subset(set):pass
diff --git a/pypy/pytest.ini b/pypy/pytest.ini
new file mode 100644
--- /dev/null
+++ b/pypy/pytest.ini
@@ -0,0 +1,2 @@
+[pytest]
+addopts = --assertmode=old
\ No newline at end of file
diff --git a/pypy/rlib/clibffi.py b/pypy/rlib/clibffi.py
--- a/pypy/rlib/clibffi.py
+++ b/pypy/rlib/clibffi.py
@@ -10,6 +10,7 @@
 from pypy.rlib.rmmap import alloc
 from pypy.rlib.rdynload import dlopen, dlclose, dlsym, dlsym_byordinal
 from pypy.rlib.rdynload import DLOpenError, DLLHANDLE
+from pypy.rlib import jit
 from pypy.tool.autopath import pypydir
 from pypy.translator.tool.cbuild import ExternalCompilationInfo
 from pypy.translator.platform import platform
@@ -18,6 +19,10 @@
 import sys
 import ctypes.util
 
+from pypy.tool.ansi_print import ansi_log
+log = py.log.Producer("libffi")
+py.log.setconsumer("libffi", ansi_log)
+
 # maaaybe isinstance here would be better. Think
 _MSVC = platform.name == "msvc"
 _MINGW = platform.name == "mingw32"
@@ -67,12 +72,17 @@
             result = os.path.join(dir, 'libffi.a')
             if os.path.exists(result):
                 return result
-        raise ImportError("'libffi.a' not found in %s" % (dirlist,))
+        log.WARNING("'libffi.a' not found in %s" % (dirlist,))
+        log.WARNING("trying to use the dynamic library instead...")
+        return None
 
+    path_libffi_a = None
     if hasattr(platform, 'library_dirs_for_libffi_a'):
+        path_libffi_a = find_libffi_a()
+    if path_libffi_a is not None:
         # platforms on which we want static linking
         libraries = []
-        link_files = [find_libffi_a()]
+        link_files = [path_libffi_a]
     else:
         # platforms on which we want dynamic linking
         libraries = ['ffi']
@@ -261,6 +271,7 @@
 elif _MSVC:
     get_libc_handle = external('pypy_get_libc_handle', [], DLLHANDLE)
 
+    @jit.dont_look_inside
     def get_libc_name():
         return rwin32.GetModuleFileName(get_libc_handle())
 
diff --git a/pypy/rlib/debug.py b/pypy/rlib/debug.py
--- a/pypy/rlib/debug.py
+++ b/pypy/rlib/debug.py
@@ -262,6 +262,28 @@
         return hop.inputarg(hop.args_r[0], arg=0)
 
 
+def mark_dict_non_null(d):
+    """ Mark dictionary as having non-null keys and values. A warning would
+    be emitted (not an error!) in case annotation disagrees.
+    """
+    assert isinstance(d, dict)
+    return d
+
+
+class DictMarkEntry(ExtRegistryEntry):
+    _about_ = mark_dict_non_null
+    
+    def compute_result_annotation(self, s_dict):
+        from pypy.annotation.model import SomeDict, s_None
+
+        assert isinstance(s_dict, SomeDict)
+        s_dict.dictdef.force_non_null = True
+        return s_dict
+
+    def specialize_call(self, hop):
+        hop.exception_cannot_occur()
+        return hop.inputarg(hop.args_r[0], arg=0)
+
 class IntegerCanBeNegative(Exception):
     pass
 
diff --git a/pypy/rlib/jit.py b/pypy/rlib/jit.py
--- a/pypy/rlib/jit.py
+++ b/pypy/rlib/jit.py
@@ -273,7 +273,8 @@
 class JitHintError(Exception):
     """Inconsistency in the JIT hints."""
 
-PARAMETERS = {'threshold': 1000,
+PARAMETERS = {'threshold': 1032, # just above 1024
+              'function_threshold': 1617, # slightly more than one above 
               'trace_eagerness': 200,
               'trace_limit': 12000,
               'inlining': 0,
diff --git a/pypy/rlib/objectmodel.py b/pypy/rlib/objectmodel.py
--- a/pypy/rlib/objectmodel.py
+++ b/pypy/rlib/objectmodel.py
@@ -448,10 +448,11 @@
     The functions key_eq() and key_hash() are used by the key comparison
     algorithm."""
 
-    def __init__(self, key_eq, key_hash):
+    def __init__(self, key_eq, key_hash, force_non_null=False):
         self._dict = {}
         self.key_eq = key_eq
         self.key_hash = key_hash
+        self.force_non_null = force_non_null
 
     def __getitem__(self, key):
         return self._dict[_r_dictkey(self, key)]
diff --git a/pypy/rlib/rgc.py b/pypy/rlib/rgc.py
--- a/pypy/rlib/rgc.py
+++ b/pypy/rlib/rgc.py
@@ -272,7 +272,9 @@
     if isinstance(TP.OF, lltype.Ptr) and TP.OF.TO._gckind == 'gc':
         # perform a write barrier that copies necessary flags from
         # source to dest
-        if not llop.gc_writebarrier_before_copy(lltype.Bool, source, dest):
+        if not llop.gc_writebarrier_before_copy(lltype.Bool, source, dest,
+                                                source_start, dest_start,
+                                                length):
             # if the write barrier is not supported, copy by hand
             for i in range(length):
                 dest[i + dest_start] = source[i + source_start]
diff --git a/pypy/rlib/ropenssl.py b/pypy/rlib/ropenssl.py
--- a/pypy/rlib/ropenssl.py
+++ b/pypy/rlib/ropenssl.py
@@ -134,7 +134,8 @@
 
 def external(name, argtypes, restype, **kw):
     kw['compilation_info'] = eci
-    eci.export_symbols += (name,)
+    if not kw.get('macro', False):
+        eci.export_symbols += (name,)
     return rffi.llexternal(
         name, argtypes, restype, **kw)
 
@@ -150,7 +151,7 @@
                 [rffi.INT, rffi.INT, rffi.CCHARP, rffi.INT], lltype.Void))],
              lltype.Void)
 ssl_external('CRYPTO_set_id_callback',
-             [lltype.Ptr(lltype.FuncType([], rffi.INT))],
+             [lltype.Ptr(lltype.FuncType([], rffi.LONG))],
              lltype.Void)
              
 if HAVE_OPENSSL_RAND:
diff --git a/pypy/rlib/rrandom.py b/pypy/rlib/rrandom.py
--- a/pypy/rlib/rrandom.py
+++ b/pypy/rlib/rrandom.py
@@ -24,8 +24,7 @@
     def __init__(self, seed=r_uint(0)):
         self.state = [r_uint(0)] * N
         self.index = 0
-        if seed:
-            self.init_genrand(seed)
+        self.init_genrand(seed)
 
     def init_genrand(self, s):
         mt = self.state
diff --git a/pypy/rlib/rsdl/RMix.py b/pypy/rlib/rsdl/RMix.py
--- a/pypy/rlib/rsdl/RMix.py
+++ b/pypy/rlib/rsdl/RMix.py
@@ -52,7 +52,8 @@
                                ChunkPtr)
 
 def LoadWAV(filename_ccharp):
-    return LoadWAV_RW(RSDL.RWFromFile(filename_ccharp, rffi.str2charp('rb')), 1)
+    with rffi.scoped_str2charp('rb') as mode:
+        return LoadWAV_RW(RSDL.RWFromFile(filename_ccharp, mode), 1)
 
 
 PlayChannelTimed    = external('Mix_PlayChannelTimed',
@@ -64,4 +65,4 @@
 
 """Returns zero if the channel is not playing. 
 Otherwise if you passed in -1, the number of channels playing is returned"""
-ChannelPlaying  = external('Mix_Playing', [ rffi.INT])
\ No newline at end of file
+ChannelPlaying  = external('Mix_Playing', [rffi.INT], rffi.INT)
diff --git a/pypy/rlib/test/test_debug.py b/pypy/rlib/test/test_debug.py
--- a/pypy/rlib/test/test_debug.py
+++ b/pypy/rlib/test/test_debug.py
@@ -1,11 +1,12 @@
 
 import py
-from pypy.rlib.debug import check_annotation, make_sure_not_resized
-from pypy.rlib.debug import debug_print, debug_start, debug_stop
-from pypy.rlib.debug import have_debug_prints, debug_offset, debug_flush
-from pypy.rlib.debug import check_nonneg, IntegerCanBeNegative
+from pypy.rlib.debug import (check_annotation, make_sure_not_resized,
+                             debug_print, debug_start, debug_stop,
+                             have_debug_prints, debug_offset, debug_flush,
+                             check_nonneg, IntegerCanBeNegative,
+                             mark_dict_non_null)
 from pypy.rlib import debug
-from pypy.rpython.test.test_llinterp import interpret
+from pypy.rpython.test.test_llinterp import interpret, gengraph
 
 def test_check_annotation():
     class Error(Exception):
@@ -52,8 +53,17 @@
     py.test.raises(ListChangeUnallowed, interpret, f, [], 
                    list_comprehension_operations=True)
 
+def test_mark_dict_non_null():
+    def f():
+        d = {"ac": "bx"}
+        mark_dict_non_null(d)
+        return d
 
-class DebugTests:
+    t, typer, graph = gengraph(f, [])
+    assert sorted(graph.returnblock.inputargs[0].concretetype.TO.entries.TO.OF._flds.keys()) == ['key', 'value']
+
+
+class DebugTests(object):
 
     def test_debug_print_start_stop(self):
         def f(x):
diff --git a/pypy/rlib/test/test_rrandom.py b/pypy/rlib/test/test_rrandom.py
--- a/pypy/rlib/test/test_rrandom.py
+++ b/pypy/rlib/test/test_rrandom.py
@@ -3,6 +3,12 @@
 
 # the numbers were created by using CPython's _randommodule.c
 
+def test_init_from_zero():
+    rnd = Random(0)
+    assert rnd.state[:14] == [0, 1, 1812433255, 1900727105, 1208447044,
+            2481403966, 4042607538, 337614300, 3232553940,
+            1018809052, 3202401494, 1775180719, 3192392114, 594215549]
+
 def test_init_from_seed():
     rnd = Random(1000)
     assert rnd.state[:14] == [1000, 4252021385, 1724402292, 571538732,
diff --git a/pypy/rpython/llinterp.py b/pypy/rpython/llinterp.py
--- a/pypy/rpython/llinterp.py
+++ b/pypy/rpython/llinterp.py
@@ -737,9 +737,12 @@
     def op_zero_gc_pointers_inside(self, obj):
         raise NotImplementedError("zero_gc_pointers_inside")
 
-    def op_gc_writebarrier_before_copy(self, source, dest):
+    def op_gc_writebarrier_before_copy(self, source, dest,
+                                       source_start, dest_start, length):
         if hasattr(self.heap, 'writebarrier_before_copy'):
-            return self.heap.writebarrier_before_copy(source, dest)
+            return self.heap.writebarrier_before_copy(source, dest,
+                                                      source_start, dest_start,
+                                                      length)
         else:
             return True
 
diff --git a/pypy/rpython/lltypesystem/ll2ctypes.py b/pypy/rpython/lltypesystem/ll2ctypes.py
--- a/pypy/rpython/lltypesystem/ll2ctypes.py
+++ b/pypy/rpython/lltypesystem/ll2ctypes.py
@@ -37,7 +37,9 @@
     if far_regions:
         import random
         pieces = far_regions._ll2ctypes_pieces
-        num = random.randrange(len(pieces))
+        num = random.randrange(len(pieces)+1)
+        if num == len(pieces):
+            return ctype()
         i1, stop = pieces[num]
         i2 = i1 + ((ctypes.sizeof(ctype) or 1) + 7) & ~7
         if i2 > stop:
diff --git a/pypy/rpython/lltypesystem/opimpl.py b/pypy/rpython/lltypesystem/opimpl.py
--- a/pypy/rpython/lltypesystem/opimpl.py
+++ b/pypy/rpython/lltypesystem/opimpl.py
@@ -473,12 +473,16 @@
     checkadr(addr2)
     return addr1 - addr2
 
-def op_gc_writebarrier_before_copy(source, dest):
+def op_gc_writebarrier_before_copy(source, dest,
+                                   source_start, dest_start, length):
     A = lltype.typeOf(source)
     assert A == lltype.typeOf(dest)
     assert isinstance(A.TO, lltype.GcArray)
     assert isinstance(A.TO.OF, lltype.Ptr)
     assert A.TO.OF.TO._gckind == 'gc'
+    assert type(source_start) is int
+    assert type(dest_start) is int
+    assert type(length) is int
     return True
 
 def op_getfield(p, name):
diff --git a/pypy/rpython/lltypesystem/rdict.py b/pypy/rpython/lltypesystem/rdict.py
--- a/pypy/rpython/lltypesystem/rdict.py
+++ b/pypy/rpython/lltypesystem/rdict.py
@@ -9,6 +9,7 @@
 from pypy.rpython import robject
 from pypy.rlib import objectmodel, jit
 from pypy.rpython import rmodel
+from pypy.rpython.error import TyperError
 
 HIGHEST_BIT = intmask(1 << (LONG_BIT - 1))
 MASK = intmask(HIGHEST_BIT - 1)
@@ -42,7 +43,7 @@
 class DictRepr(AbstractDictRepr):
 
     def __init__(self, rtyper, key_repr, value_repr, dictkey, dictvalue,
-                 custom_eq_hash=None):
+                 custom_eq_hash=None, force_non_null=False):
         self.rtyper = rtyper
         self.DICT = lltype.GcForwardReference()
         self.lowleveltype = lltype.Ptr(self.DICT)
@@ -61,6 +62,7 @@
         self.dictvalue = dictvalue
         self.dict_cache = {}
         self._custom_eq_hash_repr = custom_eq_hash
+        self.force_non_null = force_non_null
         # setup() needs to be called to finish this initialization
 
     def _externalvsinternal(self, rtyper, item_repr):
@@ -97,6 +99,13 @@
             s_value = self.dictvalue.s_value
             nullkeymarker = not self.key_repr.can_ll_be_null(s_key)
             nullvaluemarker = not self.value_repr.can_ll_be_null(s_value)
+            if self.force_non_null:
+                if not nullkeymarker:
+                    rmodel.warning("%s can be null, but forcing non-null in dict key" % s_key)
+                    nullkeymarker = True
+                if not nullvaluemarker:
+                    rmodel.warning("%s can be null, but forcing non-null in dict value" % s_value)
+                    nullvaluemarker = True
             dummykeyobj = self.key_repr.get_ll_dummyval_obj(self.rtyper,
                                                             s_key)
             dummyvalueobj = self.value_repr.get_ll_dummyval_obj(self.rtyper,
@@ -206,7 +215,7 @@
         if dictobj is None:
             return lltype.nullptr(self.DICT)
         if not isinstance(dictobj, (dict, objectmodel.r_dict)):
-            raise TyperError("expected a dict: %r" % (dictobj,))
+            raise TypeError("expected a dict: %r" % (dictobj,))
         try:
             key = Constant(dictobj)
             return self.dict_cache[key]
@@ -640,12 +649,15 @@
     pass
 
 
-def rtype_r_dict(hop):
+def rtype_r_dict(hop, i_force_non_null=None):
     r_dict = hop.r_result
     if not r_dict.custom_eq_hash:
         raise TyperError("r_dict() call does not return an r_dict instance")
-    v_eqfn, v_hashfn = hop.inputargs(r_dict.r_rdict_eqfn,
-                                     r_dict.r_rdict_hashfn)
+    v_eqfn = hop.inputarg(r_dict.r_rdict_eqfn, arg=0)
+    v_hashfn = hop.inputarg(r_dict.r_rdict_hashfn, arg=1)
+    if i_force_non_null is not None:
+        assert i_force_non_null == 2
+        hop.inputarg(lltype.Void, arg=2)
     cDICT = hop.inputconst(lltype.Void, r_dict.DICT)
     hop.exception_cannot_occur()
     v_result = hop.gendirectcall(ll_newdict, cDICT)
diff --git a/pypy/rpython/lltypesystem/rffi.py b/pypy/rpython/lltypesystem/rffi.py
--- a/pypy/rpython/lltypesystem/rffi.py
+++ b/pypy/rpython/lltypesystem/rffi.py
@@ -139,10 +139,10 @@
         source = py.code.Source("""
             def call_external_function(%(argnames)s):
                 before = aroundstate.before
-                after = aroundstate.after
                 if before: before()
                 # NB. it is essential that no exception checking occurs here!
                 res = funcptr(%(argnames)s)
+                after = aroundstate.after
                 if after: after()
                 return res
         """ % locals())
@@ -253,7 +253,7 @@
     if hasattr(callable, '_errorcode_'):
         errorcode = callable._errorcode_
     else:
-        errorcode = TP.TO.RESULT._example()
+        errorcode = TP.TO.RESULT._defl()
     callable_name = getattr(callable, '__name__', '?')
     if callbackholder is not None:
         callbackholder.callbacks[callable] = True
@@ -262,13 +262,9 @@
         def wrapper(%s):    # no *args - no GIL for mallocing the tuple
             llop.gc_stack_bottom(lltype.Void)   # marker for trackgcroot.py
             if aroundstate is not None:
-                before = aroundstate.before
                 after = aroundstate.after
-            else:
-                before = None
-                after = None
-            if after:
-                after()
+                if after:
+                    after()
             # from now on we hold the GIL
             stackcounter.stacks_counter += 1
             try:
@@ -282,8 +278,10 @@
                     traceback.print_exc()
                 result = errorcode
             stackcounter.stacks_counter -= 1
-            if before:
-                before()
+            if aroundstate is not None:
+                before = aroundstate.before
+                if before:
+                    before()
             # here we don't hold the GIL any more. As in the wrapper() produced
             # by llexternal, it is essential that no exception checking occurs
             # after the call to before().
diff --git a/pypy/rpython/lltypesystem/rlist.py b/pypy/rpython/lltypesystem/rlist.py
--- a/pypy/rpython/lltypesystem/rlist.py
+++ b/pypy/rpython/lltypesystem/rlist.py
@@ -250,12 +250,11 @@
     length = l.length
     l.length = length + 1
     l.ll_setitem_fast(length, newitem)
-ll_append_noresize.oopspec = 'list.append(l, newitem)'
 
 
 def ll_both_none(lst1, lst2):
     return not lst1 and not lst2
-        
+
 
 # ____________________________________________________________
 #
diff --git a/pypy/rpython/lltypesystem/rstr.py b/pypy/rpython/lltypesystem/rstr.py
--- a/pypy/rpython/lltypesystem/rstr.py
+++ b/pypy/rpython/lltypesystem/rstr.py
@@ -4,7 +4,7 @@
 from pypy.rlib.objectmodel import malloc_zero_filled, we_are_translated
 from pypy.rlib.objectmodel import _hash_string, enforceargs
 from pypy.rlib.debug import ll_assert
-from pypy.rlib.jit import purefunction, we_are_jitted
+from pypy.rlib.jit import purefunction, we_are_jitted, dont_look_inside
 from pypy.rlib.rarithmetic import ovfcheck
 from pypy.rpython.robject import PyObjRepr, pyobj_repr
 from pypy.rpython.rmodel import inputconst, IntegerRepr
@@ -57,6 +57,8 @@
                 llmemory.itemoffsetof(TP.chars, 0) +
                 llmemory.sizeof(CHAR_TP) * item)
 
+    # It'd be nice to be able to look inside this function.
+    @dont_look_inside
     @enforceargs(None, None, int, int, int)
     def copy_string_contents(src, dst, srcstart, dststart, length):
         assert srcstart >= 0
@@ -323,6 +325,8 @@
         return s
     ll_str2unicode.oopspec = 'str.str2unicode(str)'
 
+    # it's pure but it does not look like it
+    @purefunction
     def ll_strhash(s):
         # unlike CPython, there is no reason to avoid to return -1
         # but our malloc initializes the memory to zero, so we use zero as the
@@ -334,7 +338,6 @@
                 x = 29872897
             s.hash = x
         return x
-    ll_strhash._pure_function_ = True # it's pure but it does not look like it
 
     def ll_strfasthash(s):
         return s.hash     # assumes that the hash is already computed
diff --git a/pypy/rpython/memory/gc/generation.py b/pypy/rpython/memory/gc/generation.py
--- a/pypy/rpython/memory/gc/generation.py
+++ b/pypy/rpython/memory/gc/generation.py
@@ -517,7 +517,8 @@
             objhdr.tid &= ~GCFLAG_NO_HEAP_PTRS
             self.last_generation_root_objects.append(addr_struct)
 
-    def writebarrier_before_copy(self, source_addr, dest_addr):
+    def writebarrier_before_copy(self, source_addr, dest_addr,
+                                 source_start, dest_start, length):
         """ This has the same effect as calling writebarrier over
         each element in dest copied from source, except it might reset
         one of the following flags a bit too eagerly, which means we'll have
diff --git a/pypy/rpython/memory/gc/minimark.py b/pypy/rpython/memory/gc/minimark.py
--- a/pypy/rpython/memory/gc/minimark.py
+++ b/pypy/rpython/memory/gc/minimark.py
@@ -75,10 +75,16 @@
 
 first_gcflag = 1 << (LONG_BIT//2)
 
-# The following flag is never set on young objects.  It is initially set
-# on all prebuilt and old objects, and gets cleared by the write_barrier()
-# when we write in them a pointer to a young object.
-GCFLAG_NO_YOUNG_PTRS = first_gcflag << 0
+# The following flag is set on objects if we need to do something to
+# track the young pointers that it might contain.  The flag is not set
+# on young objects (unless they are large arrays, see below), and we
+# simply assume that any young object can point to any other young object.
+# For old and prebuilt objects, the flag is usually set, and is cleared
+# when we write a young pointer to it.  For large arrays with
+# GCFLAG_HAS_CARDS, we rely on card marking to track where the
+# young pointers are; the flag GCFLAG_TRACK_YOUNG_PTRS is set in this
+# case too, to speed up the write barrier.
+GCFLAG_TRACK_YOUNG_PTRS = first_gcflag << 0
 
 # The following flag is set on some prebuilt objects.  The flag is set
 # unless the object is already listed in 'prebuilt_root_objects'.
@@ -246,17 +252,23 @@
         self.ac = ArenaCollectionClass(arena_size, page_size,
                                        small_request_threshold)
         #
-        # Used by minor collection: a list of non-young objects that
+        # Used by minor collection: a list of (mostly non-young) objects that
         # (may) contain a pointer to a young object.  Populated by
-        # the write barrier.
-        self.old_objects_pointing_to_young = self.AddressStack()
+        # the write barrier: when we clear GCFLAG_TRACK_YOUNG_PTRS, we
+        # add it to this list.
+        class Cls(self.AddressStack):
+            def append(self2, addr):
+                assert addr not in self2.tolist()
+                self.AddressStack.append(self2, addr)
+        self.objects_pointing_to_young = self.AddressStack()
         #
-        # Similar to 'old_objects_pointing_to_young', but lists objects
+        # Similar to 'objects_pointing_to_young', but lists objects
         # that have the GCFLAG_CARDS_SET bit.  For large arrays.  Note
         # that it is possible for an object to be listed both in here
-        # and in 'old_objects_pointing_to_young', in which case we
+        # and in 'objects_pointing_to_young', in which case we
         # should just clear the cards and trace it fully, as usual.
-        self.old_objects_with_cards_set = self.AddressStack()
+        # Note also that young array objects may be added to this list.
+        self.objects_with_cards_set = self.AddressStack()
         #
         # A list of all prebuilt GC objects that contain pointers to the heap
         self.prebuilt_root_objects = self.AddressStack()
@@ -625,7 +637,7 @@
             # if 'can_make_young'.  The interesting case of 'can_make_young'
             # is for large objects, bigger than the 'large_objects' threshold,
             # which are raw-malloced but still young.
-            extra_flags = GCFLAG_NO_YOUNG_PTRS
+            extra_flags = GCFLAG_TRACK_YOUNG_PTRS
             #
         else:
             # No, so proceed to allocate it externally with raw_malloc().
@@ -643,7 +655,7 @@
                 # Reserve N extra words containing card bits before the object.
                 extra_words = self.card_marking_words_for_length(length)
                 cardheadersize = WORD * extra_words
-                extra_flags = GCFLAG_HAS_CARDS
+                extra_flags = GCFLAG_HAS_CARDS | GCFLAG_TRACK_YOUNG_PTRS
                 # note that if 'can_make_young', then card marking will only
                 # be used later, after (and if) the object becomes old
             #
@@ -686,7 +698,7 @@
                 self.young_rawmalloced_objects.add(result + size_gc_header)
             else:
                 self.old_rawmalloced_objects.append(result + size_gc_header)
-                extra_flags |= GCFLAG_NO_YOUNG_PTRS
+                extra_flags |= GCFLAG_TRACK_YOUNG_PTRS
         #
         # Common code to fill the header and length of the object.
         self.init_gc_object(result, typeid, extra_flags)
@@ -777,7 +789,7 @@
     def init_gc_object_immortal(self, addr, typeid16, flags=0):
         # For prebuilt GC objects, the flags must contain
         # GCFLAG_NO_xxx_PTRS, at least initially.
-        flags |= GCFLAG_NO_HEAP_PTRS | GCFLAG_NO_YOUNG_PTRS
+        flags |= GCFLAG_NO_HEAP_PTRS | GCFLAG_TRACK_YOUNG_PTRS
         self.init_gc_object(addr, typeid16, flags)
 
     def is_in_nursery(self, addr):
@@ -870,8 +882,8 @@
         ll_assert(not self.is_in_nursery(obj),
                   "object in nursery after collection")
         # similarily, all objects should have this flag:
-        ll_assert(self.header(obj).tid & GCFLAG_NO_YOUNG_PTRS,
-                  "missing GCFLAG_NO_YOUNG_PTRS")
+        ll_assert(self.header(obj).tid & GCFLAG_TRACK_YOUNG_PTRS,
+                  "missing GCFLAG_TRACK_YOUNG_PTRS")
         # the GCFLAG_VISITED should not be set between collections
         ll_assert(self.header(obj).tid & GCFLAG_VISITED == 0,
                   "unexpected GCFLAG_VISITED")
@@ -910,7 +922,7 @@
     # for the JIT: a minimal description of the write_barrier() method
     # (the JIT assumes it is of the shape
     #  "if addr_struct.int0 & JIT_WB_IF_FLAG: remember_young_pointer()")
-    JIT_WB_IF_FLAG = GCFLAG_NO_YOUNG_PTRS
+    JIT_WB_IF_FLAG = GCFLAG_TRACK_YOUNG_PTRS
 
     @classmethod
     def JIT_max_size_of_young_obj(cls):
@@ -921,13 +933,13 @@
         return cls.minimal_size_in_nursery
 
     def write_barrier(self, newvalue, addr_struct):
-        if self.header(addr_struct).tid & GCFLAG_NO_YOUNG_PTRS:
+        if self.header(addr_struct).tid & GCFLAG_TRACK_YOUNG_PTRS:
             self.remember_young_pointer(addr_struct, newvalue)
 
     def write_barrier_from_array(self, newvalue, addr_array, index):
-        if self.header(addr_array).tid & GCFLAG_NO_YOUNG_PTRS:
+        if self.header(addr_array).tid & GCFLAG_TRACK_YOUNG_PTRS:
             if self.card_page_indices > 0:     # <- constant-folded
-                self.remember_young_pointer_from_array(addr_array, index)
+                self.remember_young_pointer_from_array2(addr_array, index)
             else:
                 self.remember_young_pointer(addr_array, newvalue)
 
@@ -943,20 +955,23 @@
         def remember_young_pointer(addr_struct, newvalue):
             # 'addr_struct' is the address of the object in which we write.
             # 'newvalue' is the address that we are going to write in there.
+            # We know that 'addr_struct' has GCFLAG_TRACK_YOUNG_PTRS so far.
+            #
             if DEBUG:   # note: PYPY_GC_DEBUG=1 does not enable this
-                ll_assert(self.debug_is_old_object(addr_struct),
-                          "young object with GCFLAG_NO_YOUNG_PTRS")
+                ll_assert(self.debug_is_old_object(addr_struct) or
+                          self.header(addr_struct).tid & GCFLAG_HAS_CARDS != 0,
+                      "young object with GCFLAG_TRACK_YOUNG_PTRS and no cards")
             #
-            # If it seems that what we are writing is a pointer to the nursery
+            # If it seems that what we are writing is a pointer to a young obj
             # (as checked with appears_to_be_young()), then we need
-            # to remove the flag GCFLAG_NO_YOUNG_PTRS and add the old object
-            # to the list 'old_objects_pointing_to_young'.  We know that
+            # to remove the flag GCFLAG_TRACK_YOUNG_PTRS and add the object
+            # to the list 'objects_pointing_to_young'.  We know that
             # 'addr_struct' cannot be in the nursery, because nursery objects
-            # never have the flag GCFLAG_NO_YOUNG_PTRS to start with.
+            # never have the flag GCFLAG_TRACK_YOUNG_PTRS to start with.
             objhdr = self.header(addr_struct)
             if self.appears_to_be_young(newvalue):
-                self.old_objects_pointing_to_young.append(addr_struct)
-                objhdr.tid &= ~GCFLAG_NO_YOUNG_PTRS
+                self.objects_pointing_to_young.append(addr_struct)
+                objhdr.tid &= ~GCFLAG_TRACK_YOUNG_PTRS
             #
             # Second part: if 'addr_struct' is actually a prebuilt GC
             # object and it's the first time we see a write to it, we
@@ -976,20 +991,22 @@
 
     def _init_writebarrier_with_card_marker(self):
         DEBUG = self.DEBUG
-        def remember_young_pointer_from_array(addr_array, index):
+        def remember_young_pointer_from_array2(addr_array, index):
             # 'addr_array' is the address of the object in which we write,
             # which must have an array part;  'index' is the index of the
             # item that is (or contains) the pointer that we write.
-            if DEBUG:   # note: PYPY_GC_DEBUG=1 does not enable this
-                ll_assert(self.debug_is_old_object(addr_array),
-                          "young array with GCFLAG_NO_YOUNG_PTRS")
+            # We know that 'addr_array' has GCFLAG_TRACK_YOUNG_PTRS so far.
+            #
             objhdr = self.header(addr_array)
             if objhdr.tid & GCFLAG_HAS_CARDS == 0:
                 #
+                if DEBUG:   # note: PYPY_GC_DEBUG=1 does not enable this
+                    ll_assert(self.debug_is_old_object(addr_array),
+                        "young array with no card but GCFLAG_TRACK_YOUNG_PTRS")
+                #
                 # no cards, use default logic.  Mostly copied from above.
-                self.old_objects_pointing_to_young.append(addr_array)
-                objhdr = self.header(addr_array)
-                objhdr.tid &= ~GCFLAG_NO_YOUNG_PTRS
+                self.objects_pointing_to_young.append(addr_array)
+                objhdr.tid &= ~GCFLAG_TRACK_YOUNG_PTRS
                 if objhdr.tid & GCFLAG_NO_HEAP_PTRS:
                     objhdr.tid &= ~GCFLAG_NO_HEAP_PTRS
                     self.prebuilt_root_objects.append(addr_array)
@@ -1002,27 +1019,85 @@
             bitmask = 1 << (bitindex & 7)
             #
             # If the bit is already set, leave now.
-            size_gc_header = self.gcheaderbuilder.size_gc_header
-            addr_byte = addr_array - size_gc_header
-            addr_byte = llarena.getfakearenaaddress(addr_byte) + (~byteindex)
+            addr_byte = self.get_card(addr_array, byteindex)
             byte = ord(addr_byte.char[0])
             if byte & bitmask:
                 return
             #
             # We set the flag (even if the newly written address does not
             # actually point to the nursery, which seems to be ok -- actually
-            # it seems more important that remember_young_pointer_from_array()
+            # it seems more important that remember_young_pointer_from_array2()
             # does not take 3 arguments).
             addr_byte.char[0] = chr(byte | bitmask)
             #
             if objhdr.tid & GCFLAG_CARDS_SET == 0:
-                self.old_objects_with_cards_set.append(addr_array)
+                self.objects_with_cards_set.append(addr_array)
                 objhdr.tid |= GCFLAG_CARDS_SET
 
-        remember_young_pointer_from_array._dont_inline_ = True
+        remember_young_pointer_from_array2._dont_inline_ = True
         assert self.card_page_indices > 0
-        self.remember_young_pointer_from_array = (
-            remember_young_pointer_from_array)
+        self.remember_young_pointer_from_array2 = (
+            remember_young_pointer_from_array2)
+
+        # xxx trying it out for the JIT: a 3-arguments version of the above
+        def remember_young_pointer_from_array3(addr_array, index, newvalue):
+            objhdr = self.header(addr_array)
+            #
+            # a single check for the common case of neither GCFLAG_HAS_CARDS
+            # nor GCFLAG_NO_HEAP_PTRS
+            if objhdr.tid & (GCFLAG_HAS_CARDS | GCFLAG_NO_HEAP_PTRS) == 0:
+                # common case: fast path, jump to the end of the function
+                pass
+            elif objhdr.tid & GCFLAG_HAS_CARDS == 0:
+                # no cards, but GCFLAG_NO_HEAP_PTRS is set.
+                objhdr.tid &= ~GCFLAG_NO_HEAP_PTRS
+                self.prebuilt_root_objects.append(addr_array)
+                # jump to the end of the function
+            else:
+                # case with cards.
+                #
+                # If the newly written address does not actually point to a
+                # young object, leave now.
+                if not self.appears_to_be_young(newvalue):
+                    return
+                #
+                # 'addr_array' is a raw_malloc'ed array with card markers
+                # in front.  Compute the index of the bit to set:
+                bitindex = index >> self.card_page_shift
+                byteindex = bitindex >> 3
+                bitmask = 1 << (bitindex & 7)
+                #
+                # If the bit is already set, leave now.
+                addr_byte = self.get_card(addr_array, byteindex)
+                byte = ord(addr_byte.char[0])
+                if byte & bitmask:
+                    return
+                addr_byte.char[0] = chr(byte | bitmask)
+                #
+                if objhdr.tid & GCFLAG_CARDS_SET == 0:
+                    self.objects_with_cards_set.append(addr_array)
+                    objhdr.tid |= GCFLAG_CARDS_SET
+                return
+            #
+            # Logic for the no-cards case, put here to minimize the number
+            # of checks done at the start of the function
+            if DEBUG:   # note: PYPY_GC_DEBUG=1 does not enable this
+                ll_assert(self.debug_is_old_object(addr_array),
+                        "young array with no card but GCFLAG_TRACK_YOUNG_PTRS")
+            #
+            if self.appears_to_be_young(newvalue):
+                self.objects_pointing_to_young.append(addr_array)
+                objhdr.tid &= ~GCFLAG_TRACK_YOUNG_PTRS
+
+        remember_young_pointer_from_array3._dont_inline_ = True
+        assert self.card_page_indices > 0
+        self.remember_young_pointer_from_array3 = (
+            remember_young_pointer_from_array3)
+
+    def get_card(self, obj, byteindex):
+        size_gc_header = self.gcheaderbuilder.size_gc_header
+        addr_byte = obj - size_gc_header
+        return llarena.getfakearenaaddress(addr_byte) + (~byteindex)
 
 
     def assume_young_pointers(self, addr_struct):
@@ -1030,15 +1105,16 @@
         may now contain young pointers.''
         """
         objhdr = self.header(addr_struct)
-        if objhdr.tid & GCFLAG_NO_YOUNG_PTRS:
-            self.old_objects_pointing_to_young.append(addr_struct)
-            objhdr.tid &= ~GCFLAG_NO_YOUNG_PTRS
+        if objhdr.tid & GCFLAG_TRACK_YOUNG_PTRS:
+            self.objects_pointing_to_young.append(addr_struct)
+            objhdr.tid &= ~GCFLAG_TRACK_YOUNG_PTRS
             #
             if objhdr.tid & GCFLAG_NO_HEAP_PTRS:
                 objhdr.tid &= ~GCFLAG_NO_HEAP_PTRS
                 self.prebuilt_root_objects.append(addr_struct)
 
-    def writebarrier_before_copy(self, source_addr, dest_addr):
+    def writebarrier_before_copy(self, source_addr, dest_addr,
+                                 source_start, dest_start, length):
         """ This has the same effect as calling writebarrier over
         each element in dest copied from source, except it might reset
         one of the following flags a bit too eagerly, which means we'll have
@@ -1046,15 +1122,36 @@
         """
         source_hdr = self.header(source_addr)
         dest_hdr = self.header(dest_addr)
-        if dest_hdr.tid & GCFLAG_NO_YOUNG_PTRS == 0:
+        if dest_hdr.tid & GCFLAG_TRACK_YOUNG_PTRS == 0:
             return True
         # ^^^ a fast path of write-barrier
         #
-        if (source_hdr.tid & GCFLAG_NO_YOUNG_PTRS == 0 or
-            source_hdr.tid & GCFLAG_CARDS_SET != 0):
+        if source_hdr.tid & GCFLAG_HAS_CARDS != 0:
+            #
+            if source_hdr.tid & GCFLAG_TRACK_YOUNG_PTRS == 0:
+                # The source object may have random young pointers.
+                # Return False to mean "do it manually in ll_arraycopy".
+                return False
+            #
+            if source_hdr.tid & GCFLAG_CARDS_SET == 0:
+                # The source object has no young pointers at all.  Done.
+                return True
+            #
+            if dest_hdr.tid & GCFLAG_HAS_CARDS == 0:
+                # The dest object doesn't have cards.  Do it manually.
+                return False
+            #
+            if source_start != 0 or dest_start != 0:
+                # Misaligned.  Do it manually.
+                return False
+            #
+            self.manually_copy_card_bits(source_addr, dest_addr, length)
+            return True
+        #
+        if source_hdr.tid & GCFLAG_TRACK_YOUNG_PTRS == 0:
             # there might be in source a pointer to a young object
-            self.old_objects_pointing_to_young.append(dest_addr)
-            dest_hdr.tid &= ~GCFLAG_NO_YOUNG_PTRS
+            self.objects_pointing_to_young.append(dest_addr)
+            dest_hdr.tid &= ~GCFLAG_TRACK_YOUNG_PTRS
         #
         if dest_hdr.tid & GCFLAG_NO_HEAP_PTRS:
             if source_hdr.tid & GCFLAG_NO_HEAP_PTRS == 0:
@@ -1062,6 +1159,22 @@
                 self.prebuilt_root_objects.append(dest_addr)
         return True
 
+    def manually_copy_card_bits(self, source_addr, dest_addr, length):
+        # manually copy the individual card marks from source to dest
+        bytes = self.card_marking_bytes_for_length(length)
+        #
+        i = 0
+        while i < bytes:
+            addr_srcbyte = self.get_card(source_addr, i)
+            addr_dstbyte = self.get_card(dest_addr, i)
+            byte = ord(addr_srcbyte.char[0])
+            addr_dstbyte.char[0] = chr(ord(addr_dstbyte.char[0]) | byte)
+            i += 1
+        #
+        dest_hdr = self.header(dest_addr)
+        if dest_hdr.tid & GCFLAG_CARDS_SET == 0:
+            self.objects_with_cards_set.append(dest_addr)
+            dest_hdr.tid |= GCFLAG_CARDS_SET
 
     # ----------
     # Nursery collection
@@ -1078,20 +1191,28 @@
         # Note that during this step, we ignore references to further
         # young objects; only objects directly referenced by roots
         # are copied out or flagged.  They are also added to the list
-        # 'old_objects_pointing_to_young'.
+        # 'objects_pointing_to_young'.
         self.collect_roots_in_nursery()
         #
-        # If we are using card marking, do a partial trace of the arrays
-        # that are flagged with GCFLAG_CARDS_SET.
-        if self.card_page_indices > 0:
-            self.collect_cardrefs_to_nursery()
-        #
-        # Now trace objects from 'old_objects_pointing_to_young'.
-        # All nursery objects they reference are copied out of the
-        # nursery, and again added to 'old_objects_pointing_to_young'.
-        # All young raw-malloced object found is flagged GCFLAG_VISITED.
-        # We proceed until 'old_objects_pointing_to_young' is empty.
-        self.collect_oldrefs_to_nursery()
+        while True:
+            # If we are using card marking, do a partial trace of the arrays
+            # that are flagged with GCFLAG_CARDS_SET.
+            if self.card_page_indices > 0:
+                self.collect_cardrefs_to_nursery()
+            #
+            # Now trace objects from 'objects_pointing_to_young'.
+            # All nursery objects they reference are copied out of the
+            # nursery, and again added to 'objects_pointing_to_young'.
+            # All young raw-malloced object found is flagged GCFLAG_VISITED.
+            # We proceed until 'objects_pointing_to_young' is empty.
+            self.collect_oldrefs_to_nursery()
+            #
+            # We have to loop back if collect_oldrefs_to_nursery caused
+            # new objects to show up in objects_with_cards_set
+            if self.card_page_indices > 0:
+                if self.objects_with_cards_set.non_empty():
+                    continue
+            break
         #
         # Now all live nursery objects should be out.  Update the young
         # weakrefs' targets.
@@ -1124,7 +1245,7 @@
         # we don't need to trace prebuilt GcStructs during a minor collect:
         # if a prebuilt GcStruct contains a pointer to a young object,
         # then the write_barrier must have ensured that the prebuilt
-        # GcStruct is in the list self.old_objects_pointing_to_young.
+        # GcStruct is in the list self.objects_pointing_to_young.
         self.root_walker.walk_roots(
             MiniMarkGC._trace_drag_out1,  # stack roots
             MiniMarkGC._trace_drag_out1,  # static in prebuilt non-gc
@@ -1132,7 +1253,7 @@
 
     def collect_cardrefs_to_nursery(self):
         size_gc_header = self.gcheaderbuilder.size_gc_header
-        oldlist = self.old_objects_with_cards_set
+        oldlist = self.objects_with_cards_set
         while oldlist.non_empty():
             obj = oldlist.pop()
             #
@@ -1148,11 +1269,11 @@
             bytes = self.card_marking_bytes_for_length(length)
             p = llarena.getfakearenaaddress(obj - size_gc_header)
             #
-            # If the object doesn't have GCFLAG_NO_YOUNG_PTRS, then it
-            # means that it is in 'old_objects_pointing_to_young' and
+            # If the object doesn't have GCFLAG_TRACK_YOUNG_PTRS, then it
+            # means that it is in 'objects_pointing_to_young' and
             # will be fully traced by collect_oldrefs_to_nursery() just
             # afterwards.
-            if self.header(obj).tid & GCFLAG_NO_YOUNG_PTRS == 0:
+            if self.header(obj).tid & GCFLAG_TRACK_YOUNG_PTRS == 0:
                 #
                 # In that case, we just have to reset all card bits.
                 while bytes > 0:
@@ -1188,19 +1309,30 @@
 
 
     def collect_oldrefs_to_nursery(self):
-        # Follow the old_objects_pointing_to_young list and move the
+        # Follow the objects_pointing_to_young list and move the
         # young objects they point to out of the nursery.
-        oldlist = self.old_objects_pointing_to_young
+        oldlist = self.objects_pointing_to_young
         while oldlist.non_empty():
             obj = oldlist.pop()
             #
-            # Add the flag GCFLAG_NO_YOUNG_PTRS.  All live objects should have
-            # this flag set after a nursery collection.
-            self.header(obj).tid |= GCFLAG_NO_YOUNG_PTRS
+            # Check (somehow) that the flags are correct: we must not have
+            # GCFLAG_TRACK_YOUNG_PTRS so far.  But in a rare case, it's
+            # possible that the same obj is appended twice to the list
+            # (see _trace_drag_out, GCFLAG_VISITED case).  Filter it out
+            # here.
+            if self.header(obj).tid & GCFLAG_TRACK_YOUNG_PTRS != 0:
+                ll_assert(self.header(obj).tid & GCFLAG_VISITED != 0,
+                          "objects_pointing_to_young contains obj with "
+                          "GCFLAG_TRACK_YOUNG_PTRS and not GCFLAG_VISITED")
+                continue
+            #
+            # Add the flag GCFLAG_TRACK_YOUNG_PTRS.  All live objects should
+            # have this flag set after a nursery collection.
+            self.header(obj).tid |= GCFLAG_TRACK_YOUNG_PTRS
             #
             # Trace the 'obj' to replace pointers to nursery with pointers
             # outside the nursery, possibly forcing nursery objects out
-            # and adding them to 'old_objects_pointing_to_young' as well.
+            # and adding them to 'objects_pointing_to_young' as well.
             self.trace_and_drag_out_of_nursery(obj)
 
     def trace_and_drag_out_of_nursery(self, obj):
@@ -1239,7 +1371,19 @@
                 # 'obj' points to a young, raw-malloced object
                 if (self.header(obj).tid & GCFLAG_VISITED) == 0:
                     self.header(obj).tid |= GCFLAG_VISITED
-                    self.old_objects_pointing_to_young.append(obj)
+                    #
+                    # we just made 'obj' old, so we may need to add it
+                    # in the correct list:
+                    if self.header(obj).tid & GCFLAG_TRACK_YOUNG_PTRS == 0:
+                        # common case: GCFLAG_TRACK_YOUNG_PTRS is not set, so
+                        # the object may contain young pointers anywhere
+                        self.objects_pointing_to_young.append(obj)
+                    else:
+                        # large array case: the object contains card marks
+                        # that tell us where young pointers are, and it
+                        # is already in objects_with_cards_set.
+                        ll_assert(self.header(obj).tid & GCFLAG_HAS_CARDS != 0,
+                                  "neither YOUNG_PTRS nor HAS_CARDS??")
             return
         #
         # If 'obj' was already forwarded, change it to its forwarding address.
@@ -1286,11 +1430,11 @@
         # Change the original pointer to this object.
         root.address[0] = newobj
         #
-        # Add the newobj to the list 'old_objects_pointing_to_young',
+        # Add the newobj to the list 'objects_pointing_to_young',
         # because it can contain further pointers to other young objects.
         # We will fix such references to point to the copy of the young
-        # objects when we walk 'old_objects_pointing_to_young'.
-        self.old_objects_pointing_to_young.append(newobj)
+        # objects when we walk 'objects_pointing_to_young'.
+        self.objects_pointing_to_young.append(newobj)
 
 
     def _malloc_out_of_nursery(self, totalsize):
diff --git a/pypy/rpython/memory/gc/test/test_direct.py b/pypy/rpython/memory/gc/test/test_direct.py
--- a/pypy/rpython/memory/gc/test/test_direct.py
+++ b/pypy/rpython/memory/gc/test/test_direct.py
@@ -522,5 +522,78 @@
             self.stackroots.pop()
     test_card_marker.GC_PARAMS = {"card_page_indices": 4}
 
+    def test_writebarrier_before_copy(self):
+        from pypy.rpython.memory.gc import minimark
+        largeobj_size =  self.gc.nonlarge_max + 1
+        p_src = self.malloc(VAR, largeobj_size)
+        p_dst = self.malloc(VAR, largeobj_size)
+        # make them old
+        self.stackroots.append(p_src)
+        self.stackroots.append(p_dst)
+        self.gc.collect()
+        p_dst = self.stackroots.pop()
+        p_src = self.stackroots.pop()
+        #
+        addr_src = llmemory.cast_ptr_to_adr(p_src)
+        addr_dst = llmemory.cast_ptr_to_adr(p_dst)
+        hdr_src = self.gc.header(addr_src)
+        hdr_dst = self.gc.header(addr_dst)
+        #
+        assert hdr_src.tid & minimark.GCFLAG_TRACK_YOUNG_PTRS
+        assert hdr_dst.tid & minimark.GCFLAG_TRACK_YOUNG_PTRS
+        #
+        res = self.gc.writebarrier_before_copy(addr_src, addr_dst, 0, 0, 10)
+        assert res
+        assert hdr_dst.tid & minimark.GCFLAG_TRACK_YOUNG_PTRS
+        #
+        hdr_src.tid &= ~minimark.GCFLAG_TRACK_YOUNG_PTRS  # pretend we have young ptrs
+        res = self.gc.writebarrier_before_copy(addr_src, addr_dst, 0, 0, 10)
+        assert res # we optimized it
+        assert hdr_dst.tid & minimark.GCFLAG_TRACK_YOUNG_PTRS == 0 # and we copied the flag
+        #
+        hdr_src.tid |= minimark.GCFLAG_TRACK_YOUNG_PTRS
+        hdr_dst.tid |= minimark.GCFLAG_TRACK_YOUNG_PTRS
+        hdr_src.tid |= minimark.GCFLAG_HAS_CARDS
+        hdr_src.tid |= minimark.GCFLAG_CARDS_SET
+        # hdr_dst.tid does not have minimark.GCFLAG_HAS_CARDS
+        res = self.gc.writebarrier_before_copy(addr_src, addr_dst, 0, 0, 10)
+        assert not res # there might be young ptrs, let ll_arraycopy to find them
+
+    def test_writebarrier_before_copy_preserving_cards(self):
+        from pypy.rpython.lltypesystem import llarena
+        from pypy.rpython.memory.gc import minimark
+        tid = self.get_type_id(VAR)
+        largeobj_size =  self.gc.nonlarge_max + 1
+        addr_src = self.gc.external_malloc(tid, largeobj_size)
+        addr_dst = self.gc.external_malloc(tid, largeobj_size)
+        hdr_src = self.gc.header(addr_src)
+        hdr_dst = self.gc.header(addr_dst)
+        #
+        assert hdr_src.tid & minimark.GCFLAG_HAS_CARDS
+        assert hdr_dst.tid & minimark.GCFLAG_HAS_CARDS
+        #
+        young_p = self.malloc(S)
+        self.gc.write_barrier_from_array(young_p, addr_src, 0)
+        index_in_third_page = int(2.5 * self.gc.card_page_indices)
+        assert index_in_third_page < largeobj_size
+        self.gc.write_barrier_from_array(young_p, addr_src,
+                                         index_in_third_page)
+        #
+        assert hdr_src.tid & minimark.GCFLAG_CARDS_SET
+        addr_byte = self.gc.get_card(addr_src, 0)
+        assert ord(addr_byte.char[0]) == 0x01 | 0x04  # bits 0 and 2
+        #
+        res = self.gc.writebarrier_before_copy(addr_src, addr_dst,
+                                             0, 0, 2*self.gc.card_page_indices)
+        assert res
+        #
+        assert hdr_dst.tid & minimark.GCFLAG_CARDS_SET
+        addr_byte = self.gc.get_card(addr_dst, 0)
+        assert ord(addr_byte.char[0]) == 0x01 | 0x04  # bits 0 and 2
+
+    test_writebarrier_before_copy_preserving_cards.GC_PARAMS = {
+        "card_page_indices": 4}
+
+
 class TestMiniMarkGCFull(DirectGCTest):
     from pypy.rpython.memory.gc.minimark import MiniMarkGC as GCClass
diff --git a/pypy/rpython/memory/gctransform/framework.py b/pypy/rpython/memory/gctransform/framework.py
--- a/pypy/rpython/memory/gctransform/framework.py
+++ b/pypy/rpython/memory/gctransform/framework.py
@@ -322,7 +322,8 @@
         if hasattr(GCClass, 'writebarrier_before_copy'):
             self.wb_before_copy_ptr = \
                     getfn(GCClass.writebarrier_before_copy.im_func,
-                    [s_gc] + [annmodel.SomeAddress()] * 2, annmodel.SomeBool())
+                    [s_gc] + [annmodel.SomeAddress()] * 2 +
+                    [annmodel.SomeInteger()] * 3, annmodel.SomeBool())
         elif GCClass.needs_write_barrier:
             raise NotImplementedError("GC needs write barrier, but does not provide writebarrier_before_copy functionality")
 
@@ -463,7 +464,7 @@
                                             annmodel.SomeInteger()],
                                            annmodel.s_None,
                                            inline=True)
-                func = getattr(gcdata.gc, 'remember_young_pointer_from_array',
+                func = getattr(gcdata.gc, 'remember_young_pointer_from_array3',
                                None)
                 if func is not None:
                     # func should not be a bound method, but a real function
@@ -471,7 +472,8 @@
                     self.write_barrier_from_array_failing_case_ptr = \
                                              getfn(func,
                                                    [annmodel.SomeAddress(),
-                                                    annmodel.SomeInteger()],
+                                                    annmodel.SomeInteger(),
+                                                    annmodel.SomeAddress()],
                                                    annmodel.s_None)
         self.statistics_ptr = getfn(GCClass.statistics.im_func,
                                     [s_gc, annmodel.SomeInteger()],
@@ -883,7 +885,7 @@
         dest_addr = hop.genop('cast_ptr_to_adr', [op.args[1]],
                                 resulttype=llmemory.Address)
         hop.genop('direct_call', [self.wb_before_copy_ptr, self.c_const_gc,
-                                  source_addr, dest_addr],
+                                  source_addr, dest_addr] + op.args[2:],
                   resultvar=op.result)
 
     def gct_weakref_create(self, hop):
diff --git a/pypy/rpython/memory/gctransform/test/test_framework.py b/pypy/rpython/memory/gctransform/test/test_framework.py
--- a/pypy/rpython/memory/gctransform/test/test_framework.py
+++ b/pypy/rpython/memory/gctransform/test/test_framework.py
@@ -163,7 +163,8 @@
     GC_PARAMS = {}
     class GCClass(MarkSweepGC):
         needs_write_barrier = True
-        def writebarrier_before_copy(self, source, dest):
+        def writebarrier_before_copy(self, source, dest,
+                                     source_start, dest_start, length):
             return True
 
 def write_barrier_check(spaceop, needs_write_barrier=True):
diff --git a/pypy/rpython/memory/gcwrapper.py b/pypy/rpython/memory/gcwrapper.py
--- a/pypy/rpython/memory/gcwrapper.py
+++ b/pypy/rpython/memory/gcwrapper.py
@@ -136,11 +136,14 @@
         ptr = lltype.cast_opaque_ptr(llmemory.GCREF, ptr)
         return self.gc.id(ptr)
 
-    def writebarrier_before_copy(self, source, dest):
+    def writebarrier_before_copy(self, source, dest,
+                                 source_start, dest_start, length):
         if self.gc.needs_write_barrier:
             source_addr = llmemory.cast_ptr_to_adr(source)
             dest_addr   = llmemory.cast_ptr_to_adr(dest)
-            return self.gc.writebarrier_before_copy(source_addr, dest_addr)
+            return self.gc.writebarrier_before_copy(source_addr, dest_addr,
+                                                    source_start, dest_start,
+                                                    length)
         else:
             return True
 
diff --git a/pypy/rpython/memory/support.py b/pypy/rpython/memory/support.py
--- a/pypy/rpython/memory/support.py
+++ b/pypy/rpython/memory/support.py
@@ -140,6 +140,14 @@
             self.foreach(_add_in_dict, result)
             return result
 
+        def tolist(self):
+            """NOT_RPYTHON.  Returns the content as a list."""
+            lst = []
+            def _add(obj, lst):
+                lst.append(obj)
+            self.foreach(_add, lst)
+            return lst
+
         def remove(self, addr):
             """Remove 'addr' from the stack.  The addr *must* be in the list,
             and preferrably near the top.
diff --git a/pypy/rpython/ootypesystem/rdict.py b/pypy/rpython/ootypesystem/rdict.py
--- a/pypy/rpython/ootypesystem/rdict.py
+++ b/pypy/rpython/ootypesystem/rdict.py
@@ -18,7 +18,7 @@
 
 class DictRepr(AbstractDictRepr):
     def __init__(self, rtyper, key_repr, value_repr, dictkey, dictvalue,
-                 custom_eq_hash=None):
+                 custom_eq_hash=None, force_non_null=False):
         self.rtyper = rtyper
         self.custom_eq_hash = custom_eq_hash is not None
 
diff --git a/pypy/rpython/rdict.py b/pypy/rpython/rdict.py
--- a/pypy/rpython/rdict.py
+++ b/pypy/rpython/rdict.py
@@ -15,6 +15,7 @@
         dictvalue = self.dictdef.dictvalue
         s_key     = dictkey  .s_value
         s_value   = dictvalue.s_value
+        force_non_null = self.dictdef.force_non_null
         if (s_key.__class__ is annmodel.SomeObject and s_key.knowntype == object and
             s_value.__class__ is annmodel.SomeObject and s_value.knowntype == object):
             return robject.pyobj_repr
@@ -29,7 +30,8 @@
                                                      lambda: rtyper.getrepr(s_value),
                                                      dictkey,
                                                      dictvalue,
-                                                     custom_eq_hash)
+                                                     custom_eq_hash,
+                                                     force_non_null)
 
     def rtyper_makekey(self):
         self.dictdef.dictkey  .dont_change_any_more = True
diff --git a/pypy/rpython/test/test_rdict.py b/pypy/rpython/test/test_rdict.py
--- a/pypy/rpython/test/test_rdict.py
+++ b/pypy/rpython/test/test_rdict.py
@@ -598,7 +598,6 @@
         res = self.interpret(func, [])
         assert res in [5263, 6352]
 
-
 class TestLLtype(BaseTestRdict, LLRtypeMixin):
     def test_dict_but_not_with_char_keys(self):
         def func(i):
@@ -860,6 +859,25 @@
         res = f()
         assert res == 1
 
+    def test_nonnull_hint(self):
+        def eq(a, b):
+            return a == b
+        def rhash(a):
+            return 3
+        
+        def func(i):
+            d = r_dict(eq, rhash, force_non_null=True)
+            if not i:
+                d[None] = i
+            else:
+                d[str(i)] = i
+            return "12" in d, d
+
+        llres = self.interpret(func, [12])
+        assert llres.item0 == 1
+        DICT = lltype.typeOf(llres.item1)
+        assert sorted(DICT.TO.entries.TO.OF._flds) == ['f_hash', 'key', 'value']
+
     # ____________________________________________________________
 
 
diff --git a/pypy/tool/gcc_cache.py b/pypy/tool/gcc_cache.py
--- a/pypy/tool/gcc_cache.py
+++ b/pypy/tool/gcc_cache.py
@@ -39,7 +39,16 @@
         data = ''
     if not (data.startswith('True') or data.startswith('FAIL\n')):
         try:
-            platform.compile(c_files, eci)
+            _previous = platform.log_errors
+            try:
+                platform.log_errors = False
+                platform.compile(c_files, eci)
+            finally:
+                del platform.log_errors
+                # ^^^remove from the instance --- needed so that it can
+                # compare equal to another instance without it
+                if platform.log_errors != _previous:
+                    platform.log_errors = _previous
             data = 'True'
             path.write(data)
         except CompilationError, e:
diff --git a/pypy/tool/jitlogparser/parser.py b/pypy/tool/jitlogparser/parser.py
--- a/pypy/tool/jitlogparser/parser.py
+++ b/pypy/tool/jitlogparser/parser.py
@@ -1,4 +1,5 @@
 import re, sys
+
 from pypy.jit.metainterp.resoperation import rop, opname
 from pypy.jit.tool.oparser import OpParser
 
@@ -51,6 +52,7 @@
 
     # factory method
     Op = Op
+    use_mock_model = True
 
     @classmethod
     def parse_from_input(cls, input):
@@ -96,7 +98,7 @@
     def __init__(self, operations, storage):
         if operations[0].name == 'debug_merge_point':
             self.inline_level = int(operations[0].args[0])
-            m = re.search('<code object ([<>\w]+), file \'(.+?)\', line (\d+)> #(\d+) (\w+)',
+            m = re.search('<code object ([<>\w]+)\. file \'(.+?)\'\. line (\d+)> #(\d+) (\w+)',
                          operations[0].getarg(1))
             if m is None:
                 # a non-code loop, like StrLiteralSearch or something
@@ -119,6 +121,9 @@
     def getcode(self):
         return self.code
 
+    def has_valid_code(self):
+        return self.code is not None
+
     def getopcode(self):
         return self.code.map[self.bytecode_no]
 
@@ -218,6 +223,12 @@
         return self._lineset
     lineset = property(getlineset)
 
+    def has_valid_code(self):
+        for chunk in self.chunks:
+            if not chunk.has_valid_code():
+                return False
+        return True
+
     def _compute_linerange(self):
         self._lineset = set()
         minline = sys.maxint
diff --git a/pypy/tool/jitlogparser/test/test_parser.py b/pypy/tool/jitlogparser/test/test_parser.py
--- a/pypy/tool/jitlogparser/test/test_parser.py
+++ b/pypy/tool/jitlogparser/test/test_parser.py
@@ -38,10 +38,10 @@
 def test_split():
     ops = parse('''
     [i0]
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 200> #10 ADD")
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 200> #11 SUB")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 200> #10 ADD")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 200> #11 SUB")
     i1 = int_add(i0, 1)
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 200> #11 SUB")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 200> #11 SUB")
     i2 = int_add(i1, 1)
     ''')
     res = Function.from_operations(ops.operations, LoopStorage())
@@ -54,12 +54,12 @@
 def test_inlined_call():
     ops = parse("""
     []
-    debug_merge_point(0, '<code object inlined_call, file 'source.py', line 12> #28 CALL_FUNCTION')
+    debug_merge_point(0, '<code object inlined_call. file 'source.py'. line 12> #28 CALL_FUNCTION')
     i18 = getfield_gc(p0, descr=<BoolFieldDescr pypy.interpreter.pyframe.PyFrame.inst_is_being_profiled 89>)
-    debug_merge_point(1, '<code object inner, file 'source.py', line 9> #0 LOAD_FAST')
-    debug_merge_point(1, '<code object inner, file 'source.py', line 9> #3 LOAD_CONST')
-    debug_merge_point(1, '<code object inner, file 'source.py', line 9> #7 RETURN_VALUE')
-    debug_merge_point(0, '<code object inlined_call, file 'source.py', line 12> #31 STORE_FAST')
+    debug_merge_point(1, '<code object inner. file 'source.py'. line 9> #0 LOAD_FAST')
+    debug_merge_point(1, '<code object inner. file 'source.py'. line 9> #3 LOAD_CONST')
+    debug_merge_point(1, '<code object inner. file 'source.py'. line 9> #7 RETURN_VALUE')
+    debug_merge_point(0, '<code object inlined_call. file 'source.py'. line 12> #31 STORE_FAST')
     """)
     res = Function.from_operations(ops.operations, LoopStorage())
     assert len(res.chunks) == 3 # two chunks + inlined call
@@ -72,10 +72,10 @@
 def test_name():
     ops = parse('''
     [i0]
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 200> #10 ADD")
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 201> #11 SUB")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 200> #10 ADD")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 201> #11 SUB")
     i1 = int_add(i0, 1)
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 202> #11 SUB")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 202> #11 SUB")
     i2 = int_add(i1, 1)
     ''')
     res = Function.from_operations(ops.operations, LoopStorage())
@@ -89,10 +89,10 @@
     ops = parse('''
     [i0]
     i3 = int_add(i0, 1)
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 200> #10 ADD")
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 201> #11 SUB")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 200> #10 ADD")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 201> #11 SUB")
     i1 = int_add(i0, 1)
-    debug_merge_point(0, "<code object stuff, file '/I/dont/exist.py', line 202> #11 SUB")
+    debug_merge_point(0, "<code object stuff. file '/I/dont/exist.py'. line 202> #11 SUB")
     i2 = int_add(i1, 1)
     ''')
     res = Function.from_operations(ops.operations, LoopStorage())
@@ -102,10 +102,10 @@
     fname = str(py.path.local(__file__).join('..', 'x.py'))
     ops = parse('''
     [i0, i1]
-    debug_merge_point(0, "<code object f, file '%(fname)s', line 2> #0 LOAD_FAST")
-    debug_merge_point(0, "<code object f, file '%(fname)s', line 2> #3 LOAD_FAST")
-    debug_merge_point(0, "<code object f, file '%(fname)s', line 2> #6 BINARY_ADD")
-    debug_merge_point(0, "<code object f, file '%(fname)s', line 2> #7 RETURN_VALUE")
+    debug_merge_point(0, "<code object f. file '%(fname)s'. line 2> #0 LOAD_FAST")
+    debug_merge_point(0, "<code object f. file '%(fname)s'. line 2> #3 LOAD_FAST")
+    debug_merge_point(0, "<code object f. file '%(fname)s'. line 2> #6 BINARY_ADD")
+    debug_merge_point(0, "<code object f. file '%(fname)s'. line 2> #7 RETURN_VALUE")
     ''' % locals())
     res = Function.from_operations(ops.operations, LoopStorage())
     assert res.chunks[1].lineno == 3
@@ -114,11 +114,11 @@
     fname = str(py.path.local(__file__).join('..', 'x.py'))
     ops = parse('''
     [i0, i1]
-    debug_merge_point(0, "<code object g, file '%(fname)s', line 5> #9 LOAD_FAST")
-    debug_merge_point(0, "<code object g, file '%(fname)s', line 5> #12 LOAD_CONST")
-    debug_merge_point(0, "<code object g, file '%(fname)s', line 5> #22 LOAD_CONST")
-    debug_merge_point(0, "<code object g, file '%(fname)s', line 5> #28 LOAD_CONST")
-    debug_merge_point(0, "<code object g, file '%(fname)s', line 5> #6 SETUP_LOOP")
+    debug_merge_point(0, "<code object g. file '%(fname)s'. line 5> #9 LOAD_FAST")
+    debug_merge_point(0, "<code object g. file '%(fname)s'. line 5> #12 LOAD_CONST")
+    debug_merge_point(0, "<code object g. file '%(fname)s'. line 5> #22 LOAD_CONST")
+    debug_merge_point(0, "<code object g. file '%(fname)s'. line 5> #28 LOAD_CONST")
+    debug_merge_point(0, "<code object g. file '%(fname)s'. line 5> #6 SETUP_LOOP")
     ''' % locals())
     res = Function.from_operations(ops.operations, LoopStorage())
     assert res.linerange == (7, 9)
@@ -128,7 +128,7 @@
     fname = str(py.path.local(__file__).join('..', 'x.py'))
     ops = parse("""
     [p6, p1]
-    debug_merge_point(0, '<code object h, file '%(fname)s', line 11> #17 FOR_ITER')
+    debug_merge_point(0, '<code object h. file '%(fname)s'. line 11> #17 FOR_ITER')
     guard_class(p6, 144264192, descr=<Guard2>)
     p12 = getfield_gc(p6, descr=<GcPtrFieldDescr pypy.objspace.std.iterobject.W_AbstractSeqIterObject.inst_w_seq 12>)
     """ % locals())
@@ -168,7 +168,7 @@
     []
     int_add(0, 1)
     ''')
-    loops = LoopStorage().reconnect_loops([main, bridge])
+    LoopStorage().reconnect_loops([main, bridge])
     assert adjust_bridges(main, {})[1].name == 'guard_true'
     assert adjust_bridges(main, {'loop-13': True})[1].name == 'int_add'
 
diff --git a/pypy/tool/pytest/appsupport.py b/pypy/tool/pytest/appsupport.py
--- a/pypy/tool/pytest/appsupport.py
+++ b/pypy/tool/pytest/appsupport.py
@@ -1,8 +1,13 @@
 import autopath
 import py
-from pypy.interpreter import gateway
+from pypy.interpreter import gateway, pycode
 from pypy.interpreter.error import OperationError
 
+try:
+    from _pytest.assertion.newinterpret import interpret
+except ImportError:
+    from _pytest.assertion.oldinterpret import interpret
+
 # ____________________________________________________________
 
 class AppCode(object):
@@ -51,13 +56,11 @@
         space = self.space
         for key, w_value in vars.items():
             space.setitem(self.w_locals, space.wrap(key), w_value)
-        return space.eval(code, self.w_globals, self.w_locals)
-
-    def exec_(self, code, **vars):
-        space = self.space
-        for key, w_value in vars.items():
-            space.setitem(self.w_locals, space.wrap(key), w_value)
-        space.exec_(code, self.w_globals, self.w_locals)
+        if isinstance(code, str):
+            return space.eval(code, self.w_globals, self.w_locals)
+        pyc = pycode.PyCode._from_code(space, code)
+        return pyc.exec_host_bytecode(self.w_globals, self.w_locals)
+    exec_ = eval
 
     def repr(self, w_value):
         return self.space.unwrap(self.space.repr(w_value))
@@ -163,8 +166,8 @@
             except py.error.ENOENT: 
                 source = None
             from pypy import conftest
-            if source and not py.test.config.option.nomagic:
-                msg = py.code._reinterpret_old(source, runner, should_fail=True)
+            if source and py.test.config._assertstate.mode != "off":
+                msg = interpret(source, runner, should_fail=True)
                 space.setattr(w_self, space.wrap('args'),
                             space.newtuple([space.wrap(msg)]))
                 w_msg = space.wrap(msg)
diff --git a/pypy/tool/pytest/test/test_pytestsupport.py b/pypy/tool/pytest/test/test_pytestsupport.py
--- a/pypy/tool/pytest/test/test_pytestsupport.py
+++ b/pypy/tool/pytest/test/test_pytestsupport.py
@@ -4,7 +4,7 @@
 from pypy.interpreter.pycode import PyCode
 from pypy.interpreter.pyframe import PyFrame
 from pypy.tool.pytest.appsupport import (AppFrame, build_pytest_assertion,
-    AppExceptionInfo)
+    AppExceptionInfo, interpret)
 import py
 from pypy.tool.udir import udir
 import os
@@ -22,8 +22,8 @@
     co = PyCode._from_code(space, somefunc.func_code)
     pyframe = PyFrame(space, co, space.newdict(), None)
     runner = AppFrame(space, pyframe)
-    py.code._reinterpret_old("f = lambda x: x+1", runner, should_fail=False)
-    msg = py.code._reinterpret_old("assert isinstance(f(2), float)", runner)
+    interpret("f = lambda x: x+1", runner, should_fail=False)
+    msg = interpret("assert isinstance(f(2), float)", runner)
     assert msg.startswith("assert isinstance(3, float)\n"
                           " +  where 3 = ")
 
@@ -58,6 +58,12 @@
     except AssertionError, e:
         assert e.msg == "Failed"
 
+def app_test_comparison():
+    try:
+        assert 3 > 4
+    except AssertionError, e:
+        assert "3 > 4" in e.msg
+
 
 def test_appexecinfo(space):
     try:
diff --git a/pypy/tool/release/package.py b/pypy/tool/release/package.py
--- a/pypy/tool/release/package.py
+++ b/pypy/tool/release/package.py
@@ -3,9 +3,9 @@
 It uses 'pypy/translator/goal/pypy-c' and parts of the rest of the working
 copy.  Usage:
 
-    package.py root-pypy-dir [name-of-archive] [name-of-pypy-c]
+    package.py root-pypy-dir [name-of-archive] [name-of-pypy-c] [destination-for-tarball] [pypy-c-path]
 
-Usually you would do:   package.py ../../.. pypy-VER-PLATFORM.
+Usually you would do:   package.py ../../.. pypy-VER-PLATFORM
 The output is found in the directory /tmp/usession-YOURNAME/build/.
 """
 
@@ -122,7 +122,10 @@
             zf.close()
         else:
             archive = str(builddir.join(name + '.tar.bz2'))
-            e = os.system('tar --owner=root --group=root --numeric-owner -cvjf ' + archive + " " + name)
+            if sys.platform == 'darwin':
+                e = os.system('tar --numeric-owner -cvjf ' + archive + " " + name)
+            else:
+                e = os.system('tar --owner=root --group=root --numeric-owner -cvjf ' + archive + " " + name)
             if e:
                 raise OSError('"tar" returned exit status %r' % e)
     finally:
diff --git a/pypy/tool/test/test_gcc_cache.py b/pypy/tool/test/test_gcc_cache.py
--- a/pypy/tool/test/test_gcc_cache.py
+++ b/pypy/tool/test/test_gcc_cache.py
@@ -1,11 +1,13 @@
-
+import sys
 from pypy.tool.gcc_cache import *
 from pypy.tool.udir import udir
-import md5
+import md5, cStringIO
 from pypy.translator.tool.cbuild import ExternalCompilationInfo
 
+localudir = udir.join('test_gcc_cache').ensure(dir=1)
+
 def test_gcc_exec():
-    f = udir.join("x.c")
+    f = localudir.join("x.c")
     f.write("""
     #include <stdio.h>
     #include <test_gcc_exec.h>
@@ -15,8 +17,8 @@
        return 0;
     }
     """)
-    dir1 = udir.join('test_gcc_exec_dir1').ensure(dir=1)
-    dir2 = udir.join('test_gcc_exec_dir2').ensure(dir=1)
+    dir1 = localudir.join('test_gcc_exec_dir1').ensure(dir=1)
+    dir2 = localudir.join('test_gcc_exec_dir2').ensure(dir=1)
     dir1.join('test_gcc_exec.h').write('#define ANSWER 3\n')
     dir2.join('test_gcc_exec.h').write('#define ANSWER 42\n')
     eci = ExternalCompilationInfo(include_dirs=[str(dir1)])
@@ -36,7 +38,7 @@
     print '>>>'
 
 def test_gcc_ask():
-    f = udir.join("y.c")
+    f = localudir.join("y.c")
     f.write("""
     #include <stdio.h>
     #include <test_gcc_ask.h>
@@ -46,8 +48,8 @@
        return 0;
     }
     """)
-    dir1 = udir.join('test_gcc_ask_dir1').ensure(dir=1)
-    dir2 = udir.join('test_gcc_ask_dir2').ensure(dir=1)
+    dir1 = localudir.join('test_gcc_ask_dir1').ensure(dir=1)
+    dir2 = localudir.join('test_gcc_ask_dir2').ensure(dir=1)
     dir1.join('test_gcc_ask.h').write('/* hello world */\n')
     dir2.join('test_gcc_ask.h').write('#error boom\n')
     eci = ExternalCompilationInfo(include_dirs=[str(dir1)])
@@ -63,3 +65,15 @@
     print '<<<'
     print err
     print '>>>'
+
+def test_gcc_ask_doesnt_log_errors():
+    f = localudir.join('z.c')
+    f.write("""this file is not valid C code\n""")
+    eci = ExternalCompilationInfo()
+    oldstderr = sys.stderr
+    try:
+        sys.stderr = capture = cStringIO.StringIO()
+        py.test.raises(CompilationError, try_compile_cache, [f], eci)
+    finally:
+        sys.stderr = oldstderr
+    assert 'ERROR' not in capture.getvalue().upper()
diff --git a/pypy/translator/c/gc.py b/pypy/translator/c/gc.py
--- a/pypy/translator/c/gc.py
+++ b/pypy/translator/c/gc.py
@@ -297,6 +297,13 @@
 
     gc_startup_code = RefcountingGcPolicy.gc_startup_code.im_func
 
+    def compilation_info(self):
+        eci = BasicGcPolicy.compilation_info(self)
+        eci = eci.merge(ExternalCompilationInfo(
+            post_include_bits=['#define USING_NO_GC_AT_ALL'],
+            ))
+        return eci
+
 
 class FrameworkGcPolicy(BasicGcPolicy):
     transformerclass = framework.FrameworkGCTransformer
diff --git a/pypy/translator/c/gcc/test/elf/track12.s b/pypy/translator/c/gcc/test/elf/track12.s
new file mode 100644
--- /dev/null
+++ b/pypy/translator/c/gcc/test/elf/track12.s
@@ -0,0 +1,9 @@
+	.type	pypy_f, @function
+pypy_f:
+	pushl   4(%esp)
+	call    pypy_other
+	;; expected {4(%esp) | %ebx, %esi, %edi, %ebp | (%esp)}
+	popl    %eax
+	/* GCROOT %eax */
+	ret
+	.size	pypy_f, .-pypy_f
diff --git a/pypy/translator/c/gcc/test/elf/track13.s b/pypy/translator/c/gcc/test/elf/track13.s
new file mode 100644
--- /dev/null
+++ b/pypy/translator/c/gcc/test/elf/track13.s
@@ -0,0 +1,9 @@
+	.type	pypy_f, @function
+pypy_f:
+	call    pypy_other
+	;; expected {(%esp) | %ebx, %esi, %edi, %ebp | 8(%esp)}
+	pushl   8(%esp)
+	popl    %eax
+	/* GCROOT %eax */
+	ret
+	.size	pypy_f, .-pypy_f
diff --git a/pypy/translator/c/gcc/test/msvc/track_and_esp.s b/pypy/translator/c/gcc/test/msvc/track_and_esp.s
new file mode 100644
--- /dev/null
+++ b/pypy/translator/c/gcc/test/msvc/track_and_esp.s
@@ -0,0 +1,474 @@
+PUBLIC	??_C at _0BN@BIPHFGBC at pypy_g_ll_math_ll_math_frexp?$AA@ ; `string'
+PUBLIC	_pypy_g_ll_math_ll_math_frexp
+;	COMDAT ??_C at _0BN@BIPHFGBC at pypy_g_ll_math_ll_math_frexp?$AA@
+CONST	SEGMENT
+??_C at _0BN@BIPHFGBC at pypy_g_ll_math_ll_math_frexp?$AA@ DB 'pypy_g_ll_math_l'
+	DB	'l_math_frexp', 00H				; `string'
+; Function compile flags: /Ogtpy
+CONST	ENDS
+;	COMDAT _pypy_g_ll_math_ll_math_frexp
+_TEXT	SEGMENT
+_l_mantissa_0$ = -8					; size = 8
+_l_v21638$ = -8						; size = 8
+_l_x_14$ = 8						; size = 8
+_pypy_g_ll_math_ll_math_frexp PROC			; COMDAT
+
+; 58245: struct pypy_tuple2_0 *pypy_g_ll_math_ll_math_frexp(double l_x_14) {
+
+	push	ebp
+	mov	ebp, esp
+	and	esp, -64				; ffffffc0H
+
+; 58246: 	long *l_exp_p_0; double l_mantissa_0; bool_t l_v21641;
+; 58247: 	bool_t l_v21643; bool_t l_v21644; bool_t l_v21646; bool_t l_v21647;
+; 58248: 	bool_t l_v21652; bool_t l_v21653; bool_t l_v21660; bool_t l_v21666;
+; 58249: 	bool_t l_v21670; bool_t l_v21674; bool_t l_v21676; double l_v21638;
+; 58250: 	long l_v21637; long l_v21649; long l_v21651; long l_v21677;
+; 58251: 	long l_v21678; struct pypy_exceptions_Exception0 *l_v21687;
+; 58252: 	struct pypy_header0 *l_v21654; struct pypy_object0 *l_v21682;
+; 58253: 	struct pypy_object0 *l_v21691; struct pypy_object_vtable0 *l_v21665;
+; 58254: 	struct pypy_object_vtable0 *l_v21669;
+; 58255: 	struct pypy_object_vtable0 *l_v21675;
+; 58256: 	struct pypy_object_vtable0 *l_v21683; struct pypy_tuple2_0 *l_v21640;
+; 58257: 	struct pypy_tuple2_0 *l_v21695; void* l_v21639; void* l_v21648;
+; 58258: 	void* l_v21650; void* l_v21656; void* l_v21658; void* l_v21659;
+; 58259: 	void* l_v21668; void* l_v21672; void* l_v21679; void* l_v21688;
+; 58260: 	void* l_v21696;
+; 58261: 	goto block0;
+; 58262: 
+; 58263:     block0:
+; 58264: 	l_v21641 = pypy_g_ll_math_ll_math_isnan(l_x_14);
+
+	fld	QWORD PTR _l_x_14$[ebp]
+	sub	esp, 52					; 00000034H
+	push	ebx
+	push	esi
+	push	edi
+	sub	esp, 8
+	fstp	QWORD PTR [esp]
+$block0$88239:
+	call	_pypy_g_ll_math_ll_math_isnan
+
+; 58265: 	pypy_asm_gc_nocollect(pypy_g_ll_math_ll_math_isnan);
+; 58266: 	l_v21643 = l_v21641;
+; 58267: 	if (l_v21643) {
+; 58268: 		l_v21637 = 0L;
+; 58269: 		l_v21638 = l_x_14;
+
+	fld	QWORD PTR _l_x_14$[ebp]
+	add	esp, 8
+	test	al, al
+
+; 58270: 		goto block3;
+
+	jne	SHORT $LN10 at pypy_g_ll_@159
+
+; 58271: 	}
+; 58272: 	goto block1;
+; 58273: 
+; 58274:     block1:
+; 58275: 	l_v21644 = pypy_g_ll_math_ll_math_isinf(l_x_14);
+
+	sub	esp, 8
+	fstp	QWORD PTR [esp]
+$block1$88243:
+	call	_pypy_g_ll_math_ll_math_isinf
+	add	esp, 8
+
+; 58276: 	pypy_asm_gc_nocollect(pypy_g_ll_math_ll_math_isinf);
+; 58277: 	l_v21646 = l_v21644;
+; 58278: 	if (l_v21646) {
+
+	test	al, al
+	je	SHORT $block2$88245
+
+; 58279: 		l_v21637 = 0L;
+; 58280: 		l_v21638 = l_x_14;
+
+	fld	QWORD PTR _l_x_14$[ebp]
+$LN10 at pypy_g_ll_@159:
+
+; 58288: 		goto block14;
+; 58289: 	}
+; 58290: 	l_v21637 = 0L;
+
+	xor	edi, edi
+$LN30 at pypy_g_ll_@159:
+
+; 58291: 	l_v21638 = l_x_14;
+; 58292: 	goto block3;
+; 58293: 
+; 58294:     block3:
+; 58295: 	l_v21648 = (&pypy_g_pypy_rpython_memory_gc_semispace_SemiSpaceGC)->ssgc_inst_free;
+
+	mov	esi, DWORD PTR _pypy_g_pypy_rpython_memory_gc_semispace_SemiSpaceGC+4
+	fstp	QWORD PTR _l_v21638$[esp+64]
+
+; 58296: 	OP_RAW_MALLOC_USAGE((0 + ROUND_UP_FOR_ALLOCATION(sizeof(struct pypy_tuple2_0), sizeof(struct pypy_forwarding_stub0))), l_v21649);
+; 58297: 	l_v21650 = (&pypy_g_pypy_rpython_memory_gc_semispace_SemiSpaceGC)->ssgc_inst_top_of_space;
+; 58298: 	OP_ADR_DELTA(l_v21650, l_v21648, l_v21651);
+
+	mov	eax, DWORD PTR _pypy_g_pypy_rpython_memory_gc_semispace_SemiSpaceGC+12
+	sub	eax, esi
+
+; 58299: 	OP_INT_GT(l_v21649, l_v21651, l_v21652);
+
+	cmp	eax, 24					; 00000018H
+$block3$88242:
+
+; 58300: 	if (l_v21652) {
+
+	jge	$block4$88260
+
+; 58334: 	l_v21695 = l_v21640;
+; 58335: 	goto block8;
+; 58336: 
+; 58337:     block8:
+; 58338: 	RPY_DEBUG_RETURN();
+; 58339: 	return l_v21695;
+; 58340: 
+; 58341:     block9:
+; 58342: 	PYPY_DEBUG_RECORD_TRACEBACK("ll_math_ll_math_frexp");
+; 58343: 	l_v21695 = ((struct pypy_tuple2_0 *) NULL);
+; 58344: 	goto block8;
+; 58345: 
+; 58346:     block10:
+; 58347: 	abort();  /* debug_llinterpcall should be unreachable */
+; 58348: 	l_v21665 = (&pypy_g_ExcData)->ed_exc_type;
+; 58349: 	l_v21666 = (l_v21665 == NULL);
+; 58350: 	if (!l_v21666) {
+; 58351: 		goto block11;
+; 58352: 	}
+; 58353: 	goto block5;
+; 58354: 
+; 58355:     block11:
+; 58356: 	PYPY_DEBUG_RECORD_TRACEBACK("ll_math_ll_math_frexp");
+; 58357: 	l_v21696 = NULL;
+; 58358: 	goto block6;
+; 58359: 
+; 58360:     block12:
+; 58361: 	l_v21668 = pypy_g_SemiSpaceGC_obtain_free_space((&pypy_g_pypy_rpython_memory_gc_semispace_SemiSpaceGC), (0 + ROUND_UP_FOR_ALLOCATION(sizeof(struct pypy_tuple2_0), sizeof(struct pypy_forwarding_stub0))));
+
+	push	24					; 00000018H
+	push	OFFSET _pypy_g_pypy_rpython_memory_gc_semispace_SemiSpaceGC
+$block12$88259:
+	call	_pypy_g_SemiSpaceGC_obtain_free_space
+    ;; expected {4(%ebp) | 16(%esp), 12(%esp), 8(%esp), (%ebp) | }
+
+; 58362: 	l_v21669 = (&pypy_g_ExcData)->ed_exc_type;
+; 58363: 	l_v21670 = (l_v21669 == NULL);
+
+	xor	ecx, ecx
+	add	esp, 8
+	cmp	DWORD PTR _pypy_g_ExcData, ecx
+
+; 58364: 	if (!l_v21670) {
+
+	je	$LN5 at pypy_g_ll_@159
+
+; 58368: 	goto block4;
+; 58369: 
+; 58370:     block13:
+; 58371: 	PYPY_DEBUG_RECORD_TRACEBACK("ll_math_ll_math_frexp");
+
+	mov	eax, DWORD PTR _pypydtcount
+	mov	DWORD PTR _pypy_debug_tracebacks[eax*8], OFFSET ?loc@?N@??pypy_g_ll_math_ll_math_frexp@@9 at 9
+	mov	DWORD PTR _pypy_debug_tracebacks[eax*8+4], ecx
+	inc	eax
+	and	eax, 8191				; 00001fffH
+	mov	DWORD PTR _pypy_debug_tracebacks[eax*8], OFFSET ?loc@?8??pypy_g_ll_math_ll_math_frexp@@9 at 9
+	mov	DWORD PTR _pypy_debug_tracebacks[eax*8+4], ecx
+	inc	eax
+	and	eax, 8191				; 00001fffH
+	mov	DWORD PTR _pypydtcount, eax
+$block13$88313:
+$block9$88285:
+	xor	eax, eax
+
+; 58423: 	goto block8;
+; 58424: }
+
+	pop	edi
+	pop	esi
+	pop	ebx
+	mov	esp, ebp
+	pop	ebp
+	ret	0
+$block2$88245:
+
+; 58281: 		goto block3;
+; 58282: 	}
+; 58283: 	goto block2;
+; 58284: 
+; 58285:     block2:
+; 58286: 	OP_FLOAT_IS_TRUE(l_x_14, l_v21647);
+
+	fldz
+	fld	QWORD PTR _l_x_14$[ebp]
+	fucom	ST(1)
+	fnstsw	ax
+	fstp	ST(1)
+	test	ah, 68					; 00000044H
+
+; 58287: 	if (l_v21647) {
+
+	jnp	$LN10 at pypy_g_ll_@159
+
+; 58372: 	l_v21696 = NULL;
+; 58373: 	goto block6;
+; 58374: 
+; 58375:     block14:
+; 58376: 	l_v21672 = pypy_g__ll_malloc_varsize_no_length__Signed_Signed_Sign(1L, (0 + 0), sizeof(long));
+
+	push	4
+	fstp	ST(0)
+	push	0
+	push	1
+$block14$88247:
+	call	_pypy_g__ll_malloc_varsize_no_length__Signed_Signed_Sign
+    ;; expected {4(%ebp) | 20(%esp), 16(%esp), 12(%esp), (%ebp) | }
+	mov	esi, eax
+
+; 58377: 	OP_TRACK_ALLOC_START(l_v21672, /* nothing */);
+
+	push	OFFSET ??_C at _0BN@BIPHFGBC at pypy_g_ll_math_ll_math_frexp?$AA@
+	push	esi
+	call	_pypy_debug_alloc_start
+    ;; expected {4(%ebp) | 28(%esp), 24(%esp), 20(%esp), (%ebp) | }
+	add	esp, 20					; 00000014H
+
+; 58378: 	l_exp_p_0 = (long *)l_v21672;
+; 58379: 	l_v21674 = (l_exp_p_0 != NULL);
+
+	test	esi, esi
+
+; 58380: 	if (!l_v21674) {
+
+	jne	SHORT $block15$88324
+
+; 58418: 	goto block8;
+; 58419: 
+; 58420:     block18:
+; 58421: 	PYPY_DEBUG_RECORD_TRACEBACK("ll_math_ll_math_frexp");
+
+	mov	eax, DWORD PTR _pypydtcount
+	mov	DWORD PTR _pypy_debug_tracebacks[eax*8], OFFSET ?loc@?BB@??pypy_g_ll_math_ll_math_frexp@@9 at 9
+	mov	DWORD PTR _pypy_debug_tracebacks[eax*8+4], esi
+	inc	eax
+	and	eax, 8191				; 00001fffH
+	mov	DWORD PTR _pypydtcount, eax
+$block18$88323:
+
+; 58422: 	l_v21695 = ((struct pypy_tuple2_0 *) NULL);
+
+	xor	eax, eax
+
+; 58423: 	goto block8;
+; 58424: }
+
+	pop	edi
+	pop	esi
+	pop	ebx
+	mov	esp, ebp
+	pop	ebp
+	ret	0
+$block15$88324:
+
+; 58381: 		goto block18;
+; 58382: 	}
+; 58383: 	goto block15;
+; 58384: 
+; 58385:     block15:
+; 58386: 	l_mantissa_0 = pypy_g_frexp__Float_arrayPtr_star_2(l_x_14, l_exp_p_0);
+
+	fld	QWORD PTR _l_x_14$[ebp]
+	push	esi
+	sub	esp, 8
+	fstp	QWORD PTR [esp]
+	call	_pypy_g_frexp__Float_arrayPtr_star_2
+    ;; expected {4(%ebp) | 20(%esp), 16(%esp), 12(%esp), (%ebp) | }
+
+; 58387: 	l_v21675 = (&pypy_g_ExcData)->ed_exc_type;
+; 58388: 	l_v21676 = (l_v21675 == NULL);
+
+	mov	edi, DWORD PTR _pypy_g_ExcData
+	fstp	QWORD PTR _l_mantissa_0$[esp+76]
+	add	esp, 12					; 0000000cH
+	test	edi, edi
+
+; 58389: 	if (!l_v21676) {
+
+	je	SHORT $block16$88328
+
+; 58403: 
+; 58404:     block17:
+; 58405: 	l_v21682 = (&pypy_g_ExcData)->ed_exc_value;
+; 58406: 	l_v21683 = (&pypy_g_ExcData)->ed_exc_type;
+; 58407: 	PYPY_DEBUG_CATCH_EXCEPTION("ll_math_ll_math_frexp", l_v21683, l_v21683 == (&pypy_g_py__code_assertion_AssertionError_vtable.ae_super.ae_super.se_super.e_super) || l_v21683 == (&pypy_g_exceptions_NotImplementedError_vtable.nie_super.re_super.se_super.e_super));
+
+	mov	eax, DWORD PTR _pypydtcount
+	mov	ebx, DWORD PTR _pypy_g_ExcData+4
+	mov	DWORD PTR _pypy_debug_tracebacks[eax*8], OFFSET ?loc@?BA@??pypy_g_ll_math_ll_math_frexp@@9 at 9
+	mov	DWORD PTR _pypy_debug_tracebacks[eax*8+4], edi
+	inc	eax
+	and	eax, 8191				; 00001fffH
+$block17$88327:
+	mov	DWORD PTR _pypydtcount, eax
+	cmp	edi, OFFSET _pypy_g_py__code_assertion_AssertionError_vtable
+	je	SHORT $LN1 at pypy_g_ll_@159
+	cmp	edi, OFFSET _pypy_g_exceptions_NotImplementedError_vtable
+	jne	SHORT $LN2 at pypy_g_ll_@159
+$LN1 at pypy_g_ll_@159:
+	call	_pypy_debug_catch_fatal_exception
+$LN2 at pypy_g_ll_@159:
+
+; 58408: 	(&pypy_g_ExcData)->ed_exc_value = ((struct pypy_object0 *) NULL);
+
+	xor	eax, eax
+
+; 58409: 	(&pypy_g_ExcData)->ed_exc_type = ((struct pypy_object_vtable0 *) NULL);
+; 58410: 	l_v21687 = (struct pypy_exceptions_Exception0 *)l_v21682;
+; 58411: 	l_v21688 = (void*)l_exp_p_0;
+; 58412: 	OP_TRACK_ALLOC_STOP(l_v21688, /* nothing */);
+
+	push	esi
+	mov	DWORD PTR _pypy_g_ExcData+4, eax
+	mov	DWORD PTR _pypy_g_ExcData, eax
+	call	_pypy_debug_alloc_stop
+    ;; expected {4(%ebp) | 12(%esp), 8(%esp), 4(%esp), (%ebp) | }
+
+; 58413: 	OP_RAW_FREE(l_v21688, /* nothing */);
+
+	push	esi
+	call	_PyObject_Free
+    ;; expected {4(%ebp) | 16(%esp), 12(%esp), 8(%esp), (%ebp) | }
+
+; 58414: 	l_v21691 = (struct pypy_object0 *)l_v21687;
+; 58415: 	pypy_g_RPyReRaiseException(l_v21683, l_v21691);
+
+	push	ebx
+	push	edi
+	call	_pypy_g_RPyReRaiseException
+	add	esp, 16					; 00000010H
+
+; 58416: 	pypy_asm_gc_nocollect(pypy_g_RPyReRaiseException);
+; 58417: 	l_v21695 = ((struct pypy_tuple2_0 *) NULL);
+
+	xor	eax, eax
+
+; 58423: 	goto block8;
+; 58424: }
+
+	pop	edi
+	pop	esi
+	pop	ebx
+	mov	esp, ebp
+	pop	ebp
+	ret	0
+$block16$88328:
+
+; 58390: 		goto block17;
+; 58391: 	}
+; 58392: 	goto block16;
+; 58393: 
+; 58394:     block16:
+; 58395: 	l_v21677 = RPyBareItem(l_exp_p_0, 0L);
+; 58396: 	l_v21678 = (long)(l_v21677);
+
+	mov	edi, DWORD PTR [esi]
+
+; 58397: 	l_v21679 = (void*)l_exp_p_0;
+; 58398: 	OP_TRACK_ALLOC_STOP(l_v21679, /* nothing */);
+
+	push	esi
+	call	_pypy_debug_alloc_stop
+    ;; expected {4(%ebp) | 12(%esp), 8(%esp), 4(%esp), (%ebp) | }
+
+; 58399: 	OP_RAW_FREE(l_v21679, /* nothing */);
+
+	push	esi
+	call	_PyObject_Free
+    ;; expected {4(%ebp) | 16(%esp), 12(%esp), 8(%esp), (%ebp) | }
+
+; 58400: 	l_v21637 = l_v21678;
+; 58401: 	l_v21638 = l_mantissa_0;
+
+	fld	QWORD PTR _l_mantissa_0$[esp+72]
+	add	esp, 8
+
+; 58402: 	goto block3;
+
+	jmp	$LN30 at pypy_g_ll_@159
+$LN5 at pypy_g_ll_@159:
+
+; 58365: 		goto block13;
+; 58366: 	}
+; 58367: 	l_v21639 = l_v21668;
+
+	mov	esi, eax
+$block4$88260:
+$block5$88263:
+
+; 58301: 		goto block12;
+; 58302: 	}
+; 58303: 	l_v21639 = l_v21648;
+; 58304: 	goto block4;
+; 58305: 
+; 58306:     block4:
+; 58307: 	OP_INT_IS_TRUE(RUNNING_ON_LLINTERP, l_v21653);
+; 58308: 	if (l_v21653) {
+; 58309: 		goto block10;
+; 58310: 	}
+; 58311: 	goto block5;
+; 58312: 
+; 58313:     block5:
+; 58314: 	l_v21654 = (struct pypy_header0 *)l_v21639;
+; 58315: 	RPyField(l_v21654, h_tid) = (GROUP_MEMBER_OFFSET(struct group_pypy_g_typeinfo_s, member20)+0L);
+
+	test	esi, esi
+	jne	SHORT $LN18 at pypy_g_ll_@159
+	call	_RPyAbort
+$LN18 at pypy_g_ll_@159:
+
+; 58316: 	OP_ADR_ADD(l_v21639, (0 + ROUND_UP_FOR_ALLOCATION(sizeof(struct pypy_tuple2_0), sizeof(struct pypy_forwarding_stub0))), l_v21656);
+; 58317: 	(&pypy_g_pypy_rpython_memory_gc_semispace_SemiSpaceGC)->ssgc_inst_free = l_v21656;
+; 58318: 	OP_ADR_ADD(l_v21639, 0, l_v21658);
+; 58319: 	l_v21659 = (void*)l_v21658;
+; 58320: 	l_v21696 = l_v21659;
+; 58321: 	goto block6;
+; 58322: 
+; 58323:     block6:
+; 58324: 	l_v21640 = (struct pypy_tuple2_0 *)l_v21696;
+; 58325: 	l_v21660 = (l_v21640 != NULL);
+; 58326: 	if (!l_v21660) {
+; 58327: 		goto block9;
+; 58328: 	}
+; 58329: 	goto block7;
+; 58330: 
+; 58331:     block7:
+; 58332: 	RPyField(l_v21640, t_item0) = l_v21638;
+
+	fld	QWORD PTR _l_v21638$[esp+64]
+	mov	DWORD PTR [esi], 81			; 00000051H
+	lea	ecx, DWORD PTR [esi+24]
+	mov	DWORD PTR _pypy_g_pypy_rpython_memory_gc_semispace_SemiSpaceGC+4, ecx
+	fstp	QWORD PTR [esi+8]
+
+; 58333: 	RPyField(l_v21640, t_item1) = l_v21637;
+
+	mov	DWORD PTR [esi+16], edi
+
+; 58423: 	goto block8;
+; 58424: }
+
+	pop	edi
+	mov	eax, esi
+	pop	esi
+$block6$88281:
+$block8$88289:
+	pop	ebx
+	mov	esp, ebp
+	pop	ebp
+	ret	0
+_pypy_g_ll_math_ll_math_frexp ENDP
+_TEXT	ENDS
diff --git a/pypy/translator/c/gcc/trackgcroot.py b/pypy/translator/c/gcc/trackgcroot.py
--- a/pypy/translator/c/gcc/trackgcroot.py
+++ b/pypy/translator/c/gcc/trackgcroot.py
@@ -271,7 +271,8 @@
 
             match = self.r_localvar_esp.match(localvar)
             if match:
-                if localvar == self.TOP_OF_STACK: # for pushl and popl, by
+                if localvar == self.TOP_OF_STACK_MINUS_WORD:
+                                                  # for pushl and popl, by
                     hint = None                   # default ebp addressing is
                 else:                             # a bit nicer
                     hint = 'esp'
@@ -526,8 +527,9 @@
         target = match.group("target")
         if target == self.ESP:
             # only for  andl $-16, %esp  used to align the stack in main().
-            # main() should not be seen at all.
-            raise AssertionError("instruction unexpected outside of main()")
+            # main() should not be seen at all.  But on e.g. MSVC we see
+            # the instruction somewhere else too...
+            return InsnCannotFollowEsp()
         else:
             return self.binary_insn(line)
 
@@ -591,10 +593,12 @@
     def _visit_push(self, line):
         match = self.r_unaryinsn.match(line)
         source = match.group(1)
-        return [InsnStackAdjust(-self.WORD)] + self.insns_for_copy(source, self.TOP_OF_STACK)
+        return self.insns_for_copy(source, self.TOP_OF_STACK_MINUS_WORD) + \
+               [InsnStackAdjust(-self.WORD)]
 
     def _visit_pop(self, target):
-        return self.insns_for_copy(self.TOP_OF_STACK, target) + [InsnStackAdjust(+self.WORD)]
+        return [InsnStackAdjust(+self.WORD)] + \
+               self.insns_for_copy(self.TOP_OF_STACK_MINUS_WORD, target)
 
     def _visit_prologue(self):
         # for the prologue of functions that use %ebp as frame pointer
@@ -986,15 +990,15 @@
     OPERAND = r'(?:[-\w$%+.:@"]+(?:[(][\w%,]+[)])?|[(][\w%,]+[)])'
     LABEL   = r'([a-zA-Z_$.][a-zA-Z0-9_$@.]*)'
     OFFSET_LABELS   = 2**30
-    TOP_OF_STACK = '0(%esp)'
+    TOP_OF_STACK_MINUS_WORD = '-4(%esp)'
 
     r_functionstart = re.compile(r"\t.type\s+"+LABEL+",\s*[@]function\s*$")
     r_functionend   = re.compile(r"\t.size\s+"+LABEL+",\s*[.]-"+LABEL+"\s*$")
-    LOCALVAR        = r"%eax|%edx|%ecx|%ebx|%esi|%edi|%ebp|\d*[(]%esp[)]"
+    LOCALVAR        = r"%eax|%edx|%ecx|%ebx|%esi|%edi|%ebp|-?\d*[(]%esp[)]"
     LOCALVARFP      = LOCALVAR + r"|-?\d*[(]%ebp[)]"
     r_localvarnofp  = re.compile(LOCALVAR)
     r_localvarfp    = re.compile(LOCALVARFP)
-    r_localvar_esp  = re.compile(r"(\d*)[(]%esp[)]")
+    r_localvar_esp  = re.compile(r"(-?\d*)[(]%esp[)]")
     r_localvar_ebp  = re.compile(r"(-?\d*)[(]%ebp[)]")
 
     r_rel_label      = re.compile(r"(\d+):\s*$")
@@ -1047,7 +1051,7 @@
     OPERAND = r'(?:[-\w$%+.:@"]+(?:[(][\w%,]+[)])?|[(][\w%,]+[)])'
     LABEL   = r'([a-zA-Z_$.][a-zA-Z0-9_$@.]*)'
     OFFSET_LABELS   = 2**30
-    TOP_OF_STACK = '0(%rsp)'
+    TOP_OF_STACK_MINUS_WORD = '-8(%rsp)'
 
     r_functionstart = re.compile(r"\t.type\s+"+LABEL+",\s*[@]function\s*$")
     r_functionend   = re.compile(r"\t.size\s+"+LABEL+",\s*[.]-"+LABEL+"\s*$")
@@ -1143,7 +1147,7 @@
     CALLEE_SAVE_REGISTERS = ['ebx', 'esi', 'edi', 'ebp']
     REG2LOC = dict((_reg, LOC_REG | ((_i+1)<<2))
                    for _i, _reg in enumerate(CALLEE_SAVE_REGISTERS))
-    TOP_OF_STACK = 'DWORD PTR [esp]'
+    TOP_OF_STACK_MINUS_WORD = 'DWORD PTR [esp-4]'
 
     OPERAND = r'(?:(:?WORD|DWORD|BYTE) PTR |OFFSET )?[_\w?:@$]*(?:[-+0-9]+)?(:?\[[-+*\w0-9]+\])?'
     LABEL   = r'([a-zA-Z_$@.][a-zA-Z0-9_$@.]*)'
@@ -1173,7 +1177,7 @@
     r_gcroot_marker = re.compile(r"$1") # never matches
     r_gcroot_marker_var = re.compile(r"DWORD PTR .+_constant_always_one_.+pypy_asm_gcroot")
     r_gcnocollect_marker = re.compile(r"\spypy_asm_gc_nocollect\(("+OPERAND+")\);")
-    r_bottom_marker = re.compile(r"; .+\tpypy_asm_stack_bottom\(\);")
+    r_bottom_marker = re.compile(r"; .+\spypy_asm_stack_bottom\(\);")
 
     FUNCTIONS_NOT_RETURNING = {
         '__exit': None,
diff --git a/pypy/translator/c/genc.py b/pypy/translator/c/genc.py
--- a/pypy/translator/c/genc.py
+++ b/pypy/translator/c/genc.py
@@ -569,7 +569,10 @@
             gcmapfiles = ['%s.gcmap' % (c,) for c in trackgcfiles]
             mk.definition('ASMFILES', sfiles)
             mk.definition('GCMAPFILES', gcmapfiles)
-            mk.definition('DEBUGFLAGS', '-O2 -fomit-frame-pointer -g')
+            if sys.platform == 'win32':
+                mk.definition('DEBUGFLAGS', '/Zi')
+            else:
+                mk.definition('DEBUGFLAGS', '-O2 -fomit-frame-pointer -g')
 
             if self.config.translation.shared:
                 mk.definition('PYPY_MAIN_FUNCTION', "pypy_main_startup")
@@ -644,7 +647,10 @@
                             '$(CC) $(CFLAGS) $(CFLAGSEXTRA) -o $@ -c $< $(INCLUDEDIRS)')
 
         else:
-            mk.definition('DEBUGFLAGS', '-O1 -g')
+            if sys.platform == 'win32':
+                mk.definition('DEBUGFLAGS', '/Zi')
+            else:
+                mk.definition('DEBUGFLAGS', '-O1 -g')
         mk.write()
         #self.translator.platform,
         #                           ,
diff --git a/pypy/translator/c/node.py b/pypy/translator/c/node.py
--- a/pypy/translator/c/node.py
+++ b/pypy/translator/c/node.py
@@ -1031,7 +1031,7 @@
             if (issubclass(value, BaseException) and
                 value.__module__ == 'exceptions'):
                 return 'PyExc_' + value.__name__
-            if value is py.code._AssertionError:
+            if issubclass(value, AssertionError):
                 return 'PyExc_AssertionError'
             if value is _StackOverflow:
                 return 'PyExc_RuntimeError'
diff --git a/pypy/translator/c/src/main.h b/pypy/translator/c/src/main.h
--- a/pypy/translator/c/src/main.h
+++ b/pypy/translator/c/src/main.h
@@ -79,6 +79,7 @@
     fprintf(stderr, "Fatal error during initialization: %s\n", errmsg);
 #endif
     abort();
+    return 1;
 }
 
 int PYPY_MAIN_FUNCTION(int argc, char *argv[])
diff --git a/pypy/translator/c/src/mem.h b/pypy/translator/c/src/mem.h
--- a/pypy/translator/c/src/mem.h
+++ b/pypy/translator/c/src/mem.h
@@ -222,6 +222,15 @@
 
 #endif /* USING_BOEHM_GC */
 
+
+#ifdef USING_NO_GC_AT_ALL
+#define OP_BOEHM_ZERO_MALLOC(size, r, restype, is_atomic, is_varsize)  \
+  r = (restype) calloc(1, size);
+#define OP_BOEHM_DISAPPEARING_LINK(link, obj, r)  /* nothing */
+#define OP_GC__DISABLE_FINALIZERS(r)  /* nothing */
+#define OP_GC__ENABLE_FINALIZERS(r)  /* nothing */
+#endif
+
 /************************************************************/
 /* weakref support */
 
diff --git a/pypy/translator/c/test/test_newgc.py b/pypy/translator/c/test/test_newgc.py
--- a/pypy/translator/c/test/test_newgc.py
+++ b/pypy/translator/c/test/test_newgc.py
@@ -1117,6 +1117,7 @@
         S = lltype.GcStruct('S', ('u', lltype.Ptr(U)))
         A = lltype.GcArray(lltype.Ptr(S))
         filename = self.filename_dump_typeids_z
+        open_flags = os.O_WRONLY | os.O_CREAT | getattr(os, 'O_BINARY', 0)
 
         def fn():
             s = lltype.malloc(S)
@@ -1128,7 +1129,7 @@
             #
             p = rgc.get_typeids_z()
             s = ''.join([p[i] for i in range(len(p))])
-            fd = os.open(filename, os.O_WRONLY | os.O_CREAT, 0666)
+            fd = os.open(filename, open_flags, 0666)
             os.write(fd, s)
             os.close(fd)
             return 0
@@ -1137,7 +1138,7 @@
 
     def test_write_typeids_z(self):
         self.run("write_typeids_z")
-        f = open(self.filename_dump_typeids_z)
+        f = open(self.filename_dump_typeids_z, 'rb')
         data_z = f.read()
         f.close()
         import zlib
diff --git a/pypy/translator/goal/targetnumpystandalone.py b/pypy/translator/goal/targetnumpystandalone.py
--- a/pypy/translator/goal/targetnumpystandalone.py
+++ b/pypy/translator/goal/targetnumpystandalone.py
@@ -10,46 +10,32 @@
 """
 
 import time
-from pypy.module.micronumpy.numarray import SingleDimArray, Code, compute
+from pypy.module.micronumpy.compile import numpy_compile
 from pypy.jit.codewriter.policy import JitPolicy
-
-def create_array(size):
-    a = SingleDimArray(size)
-    for i in range(size):
-        a.storage[i] = float(i % 10)
-    return a
+from pypy.rpython.annlowlevel import hlstr
 
 def entry_point(argv):
     if len(argv) != 3:
         print __doc__
         return 1
-    bytecode = argv[1]
-    for b in bytecode:
-        if b not in 'alf':
-            print "WRONG BYTECODE"
-            print __doc__
-            return 2
     try:
         size = int(argv[2])
     except ValueError:
         print "INVALID LITERAL FOR INT:", argv[2]
         print __doc__
         return 3
-    no_arrays = bytecode.count('l')
-    no_floats = bytecode.count('f')
-    arrays = []
-    floats = []
-    for i in range(no_arrays):
-        arrays.append(create_array(size))
-    for i in range(no_floats):
-        floats.append(float(i + 1))
-    code = Code(bytecode, arrays, floats)
     t0 = time.time()
-    compute(code)
-    print "bytecode:", bytecode, "size:", size
+    main(argv[0], size)
+    print "bytecode:", argv[0], "size:", size
     print "took:", time.time() - t0
     return 0
 
+def main(bc, size):
+    if not isinstance(bc, str):
+        bc = hlstr(bc) # for tests
+    a = numpy_compile(bc, size)
+    a = a.compute()
+
 def target(*args):
     return entry_point, None
 
diff --git a/pypy/translator/goal/translate.py b/pypy/translator/goal/translate.py
--- a/pypy/translator/goal/translate.py
+++ b/pypy/translator/goal/translate.py
@@ -103,6 +103,8 @@
     specname = os.path.splitext(os.path.basename(targetspec))[0]
     sys.path.insert(0, os.path.dirname(targetspec))
     mod = __import__(specname)
+    if 'target' not in mod.__dict__:
+        raise Exception("file %r is not a valid targetxxx.py." % (targetspec,))
     return mod.__dict__
 
 def parse_options_and_load_target():
@@ -149,6 +151,9 @@
             log.ERROR("Could not find target %r" % (arg, ))
             sys.exit(1)
 
+    # apply the platform settings
+    set_platform(config)
+
     targetspec = translateconfig.targetspec
     targetspec_dic = load_target(targetspec)
 
@@ -164,9 +169,6 @@
                 existing_config=config,
                 translating=True)
 
-    # apply the platform settings
-    set_platform(config)
-
     # apply the optimization level settings
     set_opt_level(config, translateconfig.opt)
 
diff --git a/pypy/translator/platform/__init__.py b/pypy/translator/platform/__init__.py
--- a/pypy/translator/platform/__init__.py
+++ b/pypy/translator/platform/__init__.py
@@ -38,6 +38,7 @@
     c_environ = None
 
     relevant_environ = ()
+    log_errors = True
 
     so_prefixes = ('',)
 
@@ -120,11 +121,12 @@
         if returncode != 0:
             errorfile = outname.new(ext='errors')
             errorfile.write(stderr, 'wb')
-            stderrlines = stderr.splitlines()
-            for line in stderrlines:
-                log.Error(line)
-            # ^^^ don't use ERROR, because it might actually be fine.
-            # Also, ERROR confuses lib-python/conftest.py.
+            if self.log_errors:
+                stderrlines = stderr.splitlines()
+                for line in stderrlines:
+                    log.Error(line)
+                # ^^^ don't use ERROR, because it might actually be fine.
+                # Also, ERROR confuses lib-python/conftest.py.
             raise CompilationError(stdout, stderr)
         else:
             for line in stderr.splitlines():
diff --git a/pypy/translator/platform/darwin.py b/pypy/translator/platform/darwin.py
--- a/pypy/translator/platform/darwin.py
+++ b/pypy/translator/platform/darwin.py
@@ -68,12 +68,10 @@
 
 class Darwin_i386(Darwin):
     name = "darwin_i386"
-    link_flags = ('-arch', 'i386', '-mmacosx-version-min=10.4')
-    cflags = ('-arch', 'i386', '-O3', '-fomit-frame-pointer',
-              '-mmacosx-version-min=10.4')
+    link_flags = ('-arch', 'i386')
+    cflags = ('-arch', 'i386', '-O3', '-fomit-frame-pointer')
 
 class Darwin_x86_64(Darwin):
     name = "darwin_x86_64"
-    link_flags = ('-arch', 'x86_64', '-mmacosx-version-min=10.4')
-    cflags = ('-arch', 'x86_64', '-O3', '-fomit-frame-pointer',
-              '-mmacosx-version-min=10.4')
+    link_flags = ('-arch', 'x86_64')
+    cflags = ('-arch', 'x86_64', '-O3', '-fomit-frame-pointer')
diff --git a/pytest.py b/pytest.py
old mode 100644
new mode 100755
--- a/pytest.py
+++ b/pytest.py
@@ -1,7 +1,6 @@
+#!/usr/bin/env python
 """
 unit and functional testing with Python.
-(pypy version of startup script)
-see http://pytest.org for details.
 """
 __all__ = ['main']
 
@@ -9,23 +8,6 @@
 from _pytest import core as cmdline
 from _pytest import __version__
 
-# This pytest.py script is located in the pypy source tree
-# which has a copy of pytest and py within its source tree.
-# If the environment also has an installed version of pytest/py
-# we are bound to get warnings so we disable them.
-# XXX eventually pytest and py should not be inlined shipped
-# with the pypy source code but become a requirement for installation.
-
-import warnings
-warnings.filterwarnings("ignore",
-    "Module py was already imported", category=UserWarning)
-warnings.filterwarnings("ignore",
-    "Module _pytest was already imported",
-    category=UserWarning)
-warnings.filterwarnings("ignore",
-    "Module pytest was already imported",
-    category=UserWarning)
-
 if __name__ == '__main__': # if run as a script or by 'python -m pytest'
     raise SystemExit(main())
 else:


More information about the pypy-commit mailing list