[pypy-commit] pypy default: Merged in unroll-if-alt:

alex_gaynor noreply at buildbot.pypy.org
Wed Sep 21 01:19:14 CEST 2011


Author: Alex Gaynor <alex.gaynor at gmail.com>
Branch: 
Changeset: r47381:c2297d1b6a6b
Date: 2011-09-20 19:19 -0400
http://bitbucket.org/pypy/pypy/changeset/c2297d1b6a6b/

Log:	Merged in unroll-if-alt:

	This adds @jit.look_inside_iff(lambda *args) which allows for
	conditional unrolling, as well as jit.isconstant and jit.isvirtual
	for use in these predicates. It also makes use of these to optimize
	struct.pack, struct.unpack, and str.__mod__. There's still plenty
	of room for improvements though.

	Thanks to cfbolz and fijal for review.

diff --git a/pypy/annotation/annrpython.py b/pypy/annotation/annrpython.py
--- a/pypy/annotation/annrpython.py
+++ b/pypy/annotation/annrpython.py
@@ -149,7 +149,7 @@
         desc = olddesc.bind_self(classdef)
         args = self.bookkeeper.build_args("simple_call", args_s[:])
         desc.consider_call_site(self.bookkeeper, desc.getcallfamily(), [desc],
-            args, annmodel.s_ImpossibleValue)
+            args, annmodel.s_ImpossibleValue, None)
         result = []
         def schedule(graph, inputcells):
             result.append((graph, inputcells))
diff --git a/pypy/annotation/bookkeeper.py b/pypy/annotation/bookkeeper.py
--- a/pypy/annotation/bookkeeper.py
+++ b/pypy/annotation/bookkeeper.py
@@ -209,8 +209,8 @@
                 self.consider_call_site(call_op)
 
             for pbc, args_s in self.emulated_pbc_calls.itervalues():
-                self.consider_call_site_for_pbc(pbc, 'simple_call', 
-                                                args_s, s_ImpossibleValue)
+                self.consider_call_site_for_pbc(pbc, 'simple_call',
+                                                args_s, s_ImpossibleValue, None)
             self.emulated_pbc_calls = {}
         finally:
             self.leave()
@@ -257,18 +257,18 @@
             args_s = [lltype_to_annotation(adtmeth.ll_ptrtype)] + args_s
         if isinstance(s_callable, SomePBC):
             s_result = binding(call_op.result, s_ImpossibleValue)
-            self.consider_call_site_for_pbc(s_callable,
-                                            call_op.opname,
-                                            args_s, s_result)
+            self.consider_call_site_for_pbc(s_callable, call_op.opname, args_s,
+                                            s_result, call_op)
 
-    def consider_call_site_for_pbc(self, s_callable, opname, args_s, s_result):
+    def consider_call_site_for_pbc(self, s_callable, opname, args_s, s_result,
+                                   call_op):
         descs = list(s_callable.descriptions)
         if not descs:
             return
         family = descs[0].getcallfamily()
         args = self.build_args(opname, args_s)
         s_callable.getKind().consider_call_site(self, family, descs, args,
-                                                s_result)
+                                                s_result, call_op)
 
     def getuniqueclassdef(self, cls):
         """Get the ClassDef associated with the given user cls.
@@ -656,6 +656,7 @@
                 whence = None
             else:
                 whence = emulated # callback case
+            op = None
             s_previous_result = s_ImpossibleValue
 
         def schedule(graph, inputcells):
@@ -663,7 +664,7 @@
 
         results = []
         for desc in descs:
-            results.append(desc.pycall(schedule, args, s_previous_result))
+            results.append(desc.pycall(schedule, args, s_previous_result, op))
         s_result = unionof(*results)
         return s_result
 
diff --git a/pypy/annotation/description.py b/pypy/annotation/description.py
--- a/pypy/annotation/description.py
+++ b/pypy/annotation/description.py
@@ -255,7 +255,11 @@
             raise TypeError, "signature mismatch: %s" % e.getmsg(self.name)
         return inputcells
 
-    def specialize(self, inputcells):
+    def specialize(self, inputcells, op=None):
+        if (op is None and
+            getattr(self.bookkeeper, "position_key", None) is not None):
+            _, block, i = self.bookkeeper.position_key
+            op = block.operations[i]
         if self.specializer is None:
             # get the specializer based on the tag of the 'pyobj'
             # (if any), according to the current policy
@@ -269,11 +273,14 @@
                 enforceargs = Sig(*enforceargs)
                 self.pyobj._annenforceargs_ = enforceargs
             enforceargs(self, inputcells) # can modify inputcells in-place
-        return self.specializer(self, inputcells)
+        if getattr(self.pyobj, '_annspecialcase_', '').endswith("call_location"):
+            return self.specializer(self, inputcells, op)
+        else:
+            return self.specializer(self, inputcells)
 
-    def pycall(self, schedule, args, s_previous_result):
+    def pycall(self, schedule, args, s_previous_result, op=None):
         inputcells = self.parse_arguments(args)
-        result = self.specialize(inputcells)
+        result = self.specialize(inputcells, op)
         if isinstance(result, FunctionGraph):
             graph = result         # common case
             # if that graph has a different signature, we need to re-parse
@@ -296,17 +303,17 @@
                                              None,       # selfclassdef
                                              name)
 
-    def consider_call_site(bookkeeper, family, descs, args, s_result):
+    def consider_call_site(bookkeeper, family, descs, args, s_result, op):
         shape = rawshape(args)
-        row = FunctionDesc.row_to_consider(descs, args)
+        row = FunctionDesc.row_to_consider(descs, args, op)
         family.calltable_add_row(shape, row)
     consider_call_site = staticmethod(consider_call_site)
 
-    def variant_for_call_site(bookkeeper, family, descs, args):
+    def variant_for_call_site(bookkeeper, family, descs, args, op):
         shape = rawshape(args)
         bookkeeper.enter(None)
         try:
-            row = FunctionDesc.row_to_consider(descs, args)
+            row = FunctionDesc.row_to_consider(descs, args, op)
         finally:
             bookkeeper.leave()
         index = family.calltable_lookup_row(shape, row)
@@ -316,7 +323,7 @@
     def rowkey(self):
         return self
 
-    def row_to_consider(descs, args):
+    def row_to_consider(descs, args, op):
         # see comments in CallFamily
         from pypy.annotation.model import s_ImpossibleValue
         row = {}
@@ -324,7 +331,7 @@
             def enlist(graph, ignore):
                 row[desc.rowkey()] = graph
                 return s_ImpossibleValue   # meaningless
-            desc.pycall(enlist, args, s_ImpossibleValue)
+            desc.pycall(enlist, args, s_ImpossibleValue, op)
         return row
     row_to_consider = staticmethod(row_to_consider)
 
@@ -521,7 +528,7 @@
                             "specialization" % (self.name,))
         return self.getclassdef(None)
 
-    def pycall(self, schedule, args, s_previous_result):
+    def pycall(self, schedule, args, s_previous_result, op=None):
         from pypy.annotation.model import SomeInstance, SomeImpossibleValue
         if self.specialize:
             if self.specialize == 'specialize:ctr_location':
@@ -664,7 +671,7 @@
             cdesc = cdesc.basedesc
         return s_result     # common case
 
-    def consider_call_site(bookkeeper, family, descs, args, s_result):
+    def consider_call_site(bookkeeper, family, descs, args, s_result, op):
         from pypy.annotation.model import SomeInstance, SomePBC, s_None
         if len(descs) == 1:
             # call to a single class, look at the result annotation
@@ -709,7 +716,7 @@
             initdescs[0].mergecallfamilies(*initdescs[1:])
             initfamily = initdescs[0].getcallfamily()
             MethodDesc.consider_call_site(bookkeeper, initfamily, initdescs,
-                                          args, s_None)
+                                          args, s_None, op)
     consider_call_site = staticmethod(consider_call_site)
 
     def getallbases(self):
@@ -782,13 +789,13 @@
     def getuniquegraph(self):
         return self.funcdesc.getuniquegraph()
 
-    def pycall(self, schedule, args, s_previous_result):
+    def pycall(self, schedule, args, s_previous_result, op=None):
         from pypy.annotation.model import SomeInstance
         if self.selfclassdef is None:
             raise Exception("calling %r" % (self,))
         s_instance = SomeInstance(self.selfclassdef, flags = self.flags)
         args = args.prepend(s_instance)
-        return self.funcdesc.pycall(schedule, args, s_previous_result)
+        return self.funcdesc.pycall(schedule, args, s_previous_result, op)
 
     def bind_under(self, classdef, name):
         self.bookkeeper.warning("rebinding an already bound %r" % (self,))
@@ -801,10 +808,10 @@
                                              self.name,
                                              flags)
 
-    def consider_call_site(bookkeeper, family, descs, args, s_result):
+    def consider_call_site(bookkeeper, family, descs, args, s_result, op):
         shape = rawshape(args, nextra=1)     # account for the extra 'self'
         funcdescs = [methoddesc.funcdesc for methoddesc in descs]
-        row = FunctionDesc.row_to_consider(descs, args)
+        row = FunctionDesc.row_to_consider(descs, args, op)
         family.calltable_add_row(shape, row)
     consider_call_site = staticmethod(consider_call_site)
 
@@ -956,16 +963,16 @@
         return '<MethodOfFrozenDesc %r of %r>' % (self.funcdesc,
                                                   self.frozendesc)
 
-    def pycall(self, schedule, args, s_previous_result):
+    def pycall(self, schedule, args, s_previous_result, op=None):
         from pypy.annotation.model import SomePBC
         s_self = SomePBC([self.frozendesc])
         args = args.prepend(s_self)
-        return self.funcdesc.pycall(schedule, args, s_previous_result)
+        return self.funcdesc.pycall(schedule, args, s_previous_result, op)
 
-    def consider_call_site(bookkeeper, family, descs, args, s_result):
+    def consider_call_site(bookkeeper, family, descs, args, s_result, op):
         shape = rawshape(args, nextra=1)    # account for the extra 'self'
         funcdescs = [mofdesc.funcdesc for mofdesc in descs]
-        row = FunctionDesc.row_to_consider(descs, args)
+        row = FunctionDesc.row_to_consider(descs, args, op)
         family.calltable_add_row(shape, row)
     consider_call_site = staticmethod(consider_call_site)
 
diff --git a/pypy/annotation/policy.py b/pypy/annotation/policy.py
--- a/pypy/annotation/policy.py
+++ b/pypy/annotation/policy.py
@@ -1,7 +1,7 @@
 # base annotation policy for specialization
 from pypy.annotation.specialize import default_specialize as default
 from pypy.annotation.specialize import specialize_argvalue, specialize_argtype, specialize_arglistitemtype
-from pypy.annotation.specialize import memo
+from pypy.annotation.specialize import memo, specialize_call_location
 # for some reason, model must be imported first,
 # or we create a cycle.
 from pypy.annotation import model as annmodel
@@ -75,6 +75,7 @@
     specialize__arg = staticmethod(specialize_argvalue) # specialize:arg(N)
     specialize__argtype = staticmethod(specialize_argtype) # specialize:argtype(N)
     specialize__arglistitemtype = staticmethod(specialize_arglistitemtype)
+    specialize__call_location = staticmethod(specialize_call_location)
 
     def specialize__ll(pol, *args):
         from pypy.rpython.annlowlevel import LowLevelAnnotatorPolicy
diff --git a/pypy/annotation/specialize.py b/pypy/annotation/specialize.py
--- a/pypy/annotation/specialize.py
+++ b/pypy/annotation/specialize.py
@@ -370,3 +370,7 @@
     else:
         key = s.listdef.listitem.s_value.knowntype
     return maybe_star_args(funcdesc, key, args_s)
+
+def specialize_call_location(funcdesc, args_s, op):
+    assert op is not None
+    return maybe_star_args(funcdesc, op, args_s)
diff --git a/pypy/annotation/test/test_annrpython.py b/pypy/annotation/test/test_annrpython.py
--- a/pypy/annotation/test/test_annrpython.py
+++ b/pypy/annotation/test/test_annrpython.py
@@ -1099,8 +1099,8 @@
         allocdesc = a.bookkeeper.getdesc(alloc)
         s_C1 = a.bookkeeper.immutablevalue(C1)
         s_C2 = a.bookkeeper.immutablevalue(C2)
-        graph1 = allocdesc.specialize([s_C1])
-        graph2 = allocdesc.specialize([s_C2])
+        graph1 = allocdesc.specialize([s_C1], None)
+        graph2 = allocdesc.specialize([s_C2], None)
         assert a.binding(graph1.getreturnvar()).classdef == C1df
         assert a.binding(graph2.getreturnvar()).classdef == C2df
         assert graph1 in a.translator.graphs
@@ -1135,8 +1135,8 @@
         allocdesc = a.bookkeeper.getdesc(alloc)
         s_C1 = a.bookkeeper.immutablevalue(C1)
         s_C2 = a.bookkeeper.immutablevalue(C2)
-        graph1 = allocdesc.specialize([s_C1, s_C2])
-        graph2 = allocdesc.specialize([s_C2, s_C2])
+        graph1 = allocdesc.specialize([s_C1, s_C2], None)
+        graph2 = allocdesc.specialize([s_C2, s_C2], None)
         assert a.binding(graph1.getreturnvar()).classdef == C1df
         assert a.binding(graph2.getreturnvar()).classdef == C2df
         assert graph1 in a.translator.graphs
@@ -1194,6 +1194,19 @@
         assert len(executedesc._cache[(0, 'star', 2)].startblock.inputargs) == 4
         assert len(executedesc._cache[(1, 'star', 3)].startblock.inputargs) == 5
 
+    def test_specialize_call_location(self):
+        def g(a):
+            return a
+        g._annspecialcase_ = "specialize:call_location"
+        def f(x):
+            return g(x)
+        f._annspecialcase_ = "specialize:argtype(0)"
+        def h(y):
+            w = f(y)
+            return int(f(str(y))) + w
+        a = self.RPythonAnnotator()
+        assert a.build_types(h, [int]) == annmodel.SomeInteger()
+
     def test_assert_list_doesnt_lose_info(self):
         class T(object):
             pass
diff --git a/pypy/interpreter/argument.py b/pypy/interpreter/argument.py
--- a/pypy/interpreter/argument.py
+++ b/pypy/interpreter/argument.py
@@ -125,6 +125,7 @@
 
     ###  Manipulation  ###
 
+    @jit.look_inside_iff(lambda self: not self._dont_jit)
     def unpack(self): # slowish
         "Return a ([w1,w2...], {'kw':w3...}) pair."
         kwds_w = {}
@@ -245,6 +246,8 @@
 
     ###  Parsing for function calls  ###
 
+    # XXX: this should be @jit.look_inside_iff, but we need key word arguments,
+    # and it doesn't support them for now.
     def _match_signature(self, w_firstarg, scope_w, signature, defaults_w=None,
                          blindargs=0):
         """Parse args and kwargs according to the signature of a code object,
diff --git a/pypy/jit/backend/llsupport/llmodel.py b/pypy/jit/backend/llsupport/llmodel.py
--- a/pypy/jit/backend/llsupport/llmodel.py
+++ b/pypy/jit/backend/llsupport/llmodel.py
@@ -496,6 +496,16 @@
         u = lltype.cast_opaque_ptr(lltype.Ptr(rstr.UNICODE), string)
         u.chars[index] = unichr(newvalue)
 
+    def bh_copystrcontent(self, src, dst, srcstart, dststart, length):
+        src = lltype.cast_opaque_ptr(lltype.Ptr(rstr.STR), src)
+        dst = lltype.cast_opaque_ptr(lltype.Ptr(rstr.STR), dst)
+        rstr.copy_string_contents(src, dst, srcstart, dststart, length)
+
+    def bh_copyunicodecontent(self, src, dst, srcstart, dststart, length):
+        src = lltype.cast_opaque_ptr(lltype.Ptr(rstr.UNICODE), src)
+        dst = lltype.cast_opaque_ptr(lltype.Ptr(rstr.UNICODE), dst)
+        rstr.copy_unicode_contents(src, dst, srcstart, dststart, length)
+
     def bh_call_i(self, func, calldescr, args_i, args_r, args_f):
         assert isinstance(calldescr, BaseIntCallDescr)
         if not we_are_translated():
diff --git a/pypy/jit/backend/model.py b/pypy/jit/backend/model.py
--- a/pypy/jit/backend/model.py
+++ b/pypy/jit/backend/model.py
@@ -78,7 +78,7 @@
         Optionally, return a ``ops_offset`` dictionary.  See the docstring of
         ``compiled_loop`` for more informations about it.
         """
-        raise NotImplementedError    
+        raise NotImplementedError
 
     def dump_loop_token(self, looptoken):
         """Print a disassembled version of looptoken to stdout"""
@@ -298,6 +298,10 @@
         raise NotImplementedError
     def bh_unicodesetitem(self, string, index, newvalue):
         raise NotImplementedError
+    def bh_copystrcontent(self, src, dst, srcstart, dststart, length):
+        raise NotImplementedError
+    def bh_copyunicodecontent(self, src, dst, srcstart, dststart, length):
+        raise NotImplementedError
 
     def force(self, force_token):
         raise NotImplementedError
diff --git a/pypy/jit/codewriter/jtransform.py b/pypy/jit/codewriter/jtransform.py
--- a/pypy/jit/codewriter/jtransform.py
+++ b/pypy/jit/codewriter/jtransform.py
@@ -1158,6 +1158,12 @@
             return SpaceOperation('%s_assert_green' % kind, args, None)
         elif oopspec_name == 'jit.current_trace_length':
             return SpaceOperation('current_trace_length', [], op.result)
+        elif oopspec_name == 'jit.isconstant':
+            kind = getkind(args[0].concretetype)
+            return SpaceOperation('%s_isconstant' % kind, args, op.result)
+        elif oopspec_name == 'jit.isvirtual':
+            kind = getkind(args[0].concretetype)
+            return SpaceOperation('%s_isvirtual' % kind, args, op.result)
         else:
             raise AssertionError("missing support for %r" % oopspec_name)
 
@@ -1415,6 +1421,14 @@
         else:
             assert 0, "args[0].concretetype must be STR or UNICODE"
         #
+        if oopspec_name == 'stroruni.copy_contents':
+            if SoU.TO == rstr.STR:
+                new_op = 'copystrcontent'
+            elif SoU.TO == rstr.UNICODE:
+                new_op = 'copyunicodecontent'
+            else:
+                assert 0
+            return SpaceOperation(new_op, args, op.result)
         if oopspec_name == "stroruni.equal":
             for otherindex, othername, argtypes, resulttype in [
                 (EffectInfo.OS_STREQ_SLICE_CHECKNULL,
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
@@ -835,6 +835,18 @@
     def bhimpl_current_trace_length():
         return -1
 
+    @arguments("i", returns="i")
+    def bhimpl_int_isconstant(x):
+        return False
+
+    @arguments("r", returns="i")
+    def bhimpl_ref_isconstant(x):
+        return False
+
+    @arguments("r", returns="i")
+    def bhimpl_ref_isvirtual(x):
+        return False
+
     # ----------
     # the main hints and recursive calls
 
@@ -1224,6 +1236,9 @@
     @arguments("cpu", "r", "i", "i")
     def bhimpl_strsetitem(cpu, string, index, newchr):
         cpu.bh_strsetitem(string, index, newchr)
+    @arguments("cpu", "r", "r", "i", "i", "i")
+    def bhimpl_copystrcontent(cpu, src, dst, srcstart, dststart, length):
+        cpu.bh_copystrcontent(src, dst, srcstart, dststart, length)
 
     @arguments("cpu", "i", returns="r")
     def bhimpl_newunicode(cpu, length):
@@ -1237,6 +1252,9 @@
     @arguments("cpu", "r", "i", "i")
     def bhimpl_unicodesetitem(cpu, unicode, index, newchr):
         cpu.bh_unicodesetitem(unicode, index, newchr)
+    @arguments("cpu", "r", "r", "i", "i", "i")
+    def bhimpl_copyunicodecontent(cpu, src, dst, srcstart, dststart, length):
+        cpu.bh_copyunicodecontent(src, dst, srcstart, dststart, length)
 
     @arguments(returns=(longlong.is_64_bit and "i" or "f"))
     def bhimpl_ll_read_timestamp():
diff --git a/pypy/jit/metainterp/heapcache.py b/pypy/jit/metainterp/heapcache.py
--- a/pypy/jit/metainterp/heapcache.py
+++ b/pypy/jit/metainterp/heapcache.py
@@ -11,9 +11,13 @@
         self.known_class_boxes = {}
         # store the boxes that contain newly allocated objects, this maps the
         # boxes to a bool, the bool indicates whether or not the object has
-        # escaped the trace or not, its presences in the mapping shows that it
-        # was allocated inside the trace
+        # escaped the trace or not (True means the box never escaped, False
+        # means it did escape), its presences in the mapping shows that it was
+        # allocated inside the trace
         self.new_boxes = {}
+        # Tracks which boxes should be marked as escaped when the key box
+        # escapes.
+        self.dependencies = {}
         # contains frame boxes that are not virtualizables
         self.nonstandard_virtualizables = {}
         # heap cache
@@ -31,12 +35,28 @@
 
     def mark_escaped(self, opnum, argboxes):
         idx = 0
-        for box in argboxes:
-            # setfield_gc and setarrayitem_gc don't escape their first argument
-            if not (idx == 0 and opnum in [rop.SETFIELD_GC, rop.SETARRAYITEM_GC]):
-                if box in self.new_boxes:
-                    self.new_boxes[box] = False
-            idx += 1
+        if opnum == rop.SETFIELD_GC:
+            assert len(argboxes) == 2
+            box, valuebox = argboxes
+            if self.is_unescaped(box) and self.is_unescaped(valuebox):
+                self.dependencies.setdefault(box, []).append(valuebox)
+            else:
+                self._escape(valuebox)
+        # GETFIELD_GC doesn't escape it's argument
+        elif opnum != rop.GETFIELD_GC:
+            for box in argboxes:
+                # setarrayitem_gc don't escape its first argument
+                if not (idx == 0 and opnum in [rop.SETARRAYITEM_GC]):
+                    self._escape(box)
+                idx += 1
+
+    def _escape(self, box):
+        if box in self.new_boxes:
+            self.new_boxes[box] = False
+        if box in self.dependencies:
+            for dep in self.dependencies[box]:
+                self._escape(dep)
+            del self.dependencies[box]
 
     def clear_caches(self, opnum, descr, argboxes):
         if opnum == rop.SETFIELD_GC:
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
@@ -71,7 +71,7 @@
             guards.append(op)
         elif self.level == LEVEL_KNOWNCLASS:
             op = ResOperation(rop.GUARD_NONNULL, [box], None)
-            guards.append(op)            
+            guards.append(op)
             op = ResOperation(rop.GUARD_CLASS, [box, self.known_class], None)
             guards.append(op)
         else:
@@ -112,7 +112,7 @@
                     self.lenbound.bound.intersect(other.lenbound.bound)
                 else:
                     self.lenbound = other.lenbound.clone()
-                    
+
 
     def force_box(self):
         return self.box
@@ -146,7 +146,7 @@
         assert isinstance(constbox, Const)
         self.box = constbox
         self.level = LEVEL_CONSTANT
-        
+
         if isinstance(constbox, ConstInt):
             val = constbox.getint()
             self.intbound = IntBound(val, val)
@@ -378,7 +378,7 @@
         new.set_optimizations(optimizations)
         new.quasi_immutable_deps = self.quasi_immutable_deps
         return new
-        
+
     def produce_potential_short_preamble_ops(self, sb):
         raise NotImplementedError('This is implemented in unroll.UnrollableOptimizer')
 
@@ -505,9 +505,9 @@
         if op.returns_bool_result():
             self.bool_boxes[self.getvalue(op.result)] = None
         self._emit_operation(op)
-        
+
     @specialize.argtype(0)
-    def _emit_operation(self, op):        
+    def _emit_operation(self, op):
         for i in range(op.numargs()):
             arg = op.getarg(i)
             try:
@@ -568,7 +568,7 @@
                 arg = value.get_key_box()
             args[i] = arg
         args[n] = ConstInt(op.getopnum())
-        args[n+1] = op.getdescr()
+        args[n + 1] = op.getdescr()
         return args
 
     @specialize.argtype(0)
@@ -616,7 +616,7 @@
 
     def remember_emitting_pure(self, op):
         pass
-    
+
     def constant_fold(self, op):
         argboxes = [self.get_constant_box(op.getarg(i))
                     for i in range(op.numargs())]
@@ -658,9 +658,9 @@
             arrayvalue = self.getvalue(op.getarg(0))
             arrayvalue.make_len_gt(MODE_UNICODE, op.getdescr(), indexvalue.box.getint())
         self.optimize_default(op)
-        
 
-    
+
+
 
 dispatch_opt = make_dispatcher_method(Optimizer, 'optimize_',
         default=Optimizer.optimize_default)
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
@@ -19,7 +19,7 @@
 
     def new(self):
         return OptRewrite()
-        
+
     def produce_potential_short_preamble_ops(self, sb):
         for op in self.loop_invariant_producer.values():
             sb.add_potential(op)
@@ -231,6 +231,17 @@
             else:
                 self.make_constant(op.result, result)
                 return
+
+        args = self.optimizer.make_args_key(op)
+        oldop = self.optimizer.pure_operations.get(args, None)
+        if oldop is not None and oldop.getdescr() is op.getdescr():
+            assert oldop.getopnum() == op.getopnum()
+            self.make_equal_to(op.result, self.getvalue(oldop.result))
+            return
+        else:
+            self.optimizer.pure_operations[args] = op
+            self.optimizer.remember_emitting_pure(op)
+
         # replace CALL_PURE with just CALL
         args = op.getarglist()
         self.emit_operation(ResOperation(rop.CALL, args, op.result,
@@ -351,7 +362,7 @@
         # expects a compile-time constant
         assert isinstance(arg, Const)
         key = make_hashable_int(arg.getint())
-        
+
         resvalue = self.loop_invariant_results.get(key, None)
         if resvalue is not None:
             self.make_equal_to(op.result, resvalue)
diff --git a/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py b/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
--- a/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
@@ -4711,6 +4711,35 @@
         """
         self.optimize_loop(ops, expected)
 
+    def test_empty_copystrunicontent(self):
+        ops = """
+        [p0, p1, i0, i2, i3]
+        i4 = int_eq(i3, 0)
+        guard_true(i4) []
+        copystrcontent(p0, p1, i0, i2, i3)
+        jump(p0, p1, i0, i2, i3)
+        """
+        expected = """
+        [p0, p1, i0, i2, i3]
+        i4 = int_eq(i3, 0)
+        guard_true(i4) []
+        jump(p0, p1, i0, i2, 0)
+        """
+        self.optimize_strunicode_loop(ops, expected)
+
+    def test_empty_copystrunicontent_virtual(self):
+        ops = """
+        [p0]
+        p1 = newstr(23)
+        copystrcontent(p0, p1, 0, 0, 0)
+        jump(p0)
+        """
+        expected = """
+        [p0]
+        jump(p0)
+        """
+        self.optimize_strunicode_loop(ops, expected)
+
     def test_forced_virtuals_aliasing(self):
         ops = """
         [i0, i1]
@@ -4739,6 +4768,7 @@
         self.optimize_loop(ops, expected)
 
 
+
 class TestLLtype(BaseTestOptimizeBasic, LLtypeMixin):
     pass
 
diff --git a/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py b/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
--- a/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
@@ -102,9 +102,9 @@
             print "Short Preamble:"
             short = loop.preamble.token.short_preamble[0]
             print short.inputargs
-            print '\n'.join([str(o) for o in short.operations])        
+            print '\n'.join([str(o) for o in short.operations])
             print
-        
+
         assert expected != "crash!", "should have raised an exception"
         self.assert_equal(loop, expected)
         if expected_preamble:
@@ -113,7 +113,7 @@
         if expected_short:
             self.assert_equal(short, expected_short,
                               text_right='expected short preamble')
-            
+
         return loop
 
 class OptimizeOptTest(BaseTestWithUnroll):
@@ -866,10 +866,10 @@
         setfield_gc(p3sub, i1, descr=valuedescr)
         setfield_gc(p1, p3sub, descr=nextdescr)
         # XXX: We get two extra operations here because the setfield
-        #      above is the result of forcing p1 and thus not 
+        #      above is the result of forcing p1 and thus not
         #      registered with the heap optimizer. I've makred tests
         #      below with VIRTUALHEAP if they suffer from this issue
-        p3sub2 = getfield_gc(p1, descr=nextdescr) 
+        p3sub2 = getfield_gc(p1, descr=nextdescr)
         guard_nonnull_class(p3sub2, ConstClass(node_vtable2)) []
         jump(i1, p1, p3sub2)
         """
@@ -1411,7 +1411,7 @@
         guard_isnull(p18) [p0, p8]
         p31 = new(descr=ssize)
         p35 = new_with_vtable(ConstClass(node_vtable))
-        setfield_gc(p35, p31, descr=valuedescr)        
+        setfield_gc(p35, p31, descr=valuedescr)
         jump(p0, p35)
         """
         expected = """
@@ -1426,7 +1426,7 @@
         guard_isnull(p18) [p0, p8]
         p31 = new(descr=ssize)
         p35 = new_with_vtable(ConstClass(node_vtable))
-        setfield_gc(p35, p31, descr=valuedescr)        
+        setfield_gc(p35, p31, descr=valuedescr)
         jump(p0, p35, p19, p18)
         """
         expected = """
@@ -1435,7 +1435,7 @@
         jump(p0, NULL)
         """
         self.optimize_loop(ops, expected)
-        
+
     def test_varray_1(self):
         ops = """
         [i1]
@@ -2181,7 +2181,7 @@
         jump(p1)
         """
         self.optimize_loop(ops, expected)
-        
+
     def test_duplicate_getarrayitem_2(self):
         ops = """
         [p1, i0]
@@ -2199,7 +2199,7 @@
         jump(p1, i7, i6)
         """
         self.optimize_loop(ops, expected)
-        
+
     def test_duplicate_getarrayitem_after_setarrayitem_1(self):
         ops = """
         [p1, p2]
@@ -2812,14 +2812,14 @@
         guard_no_overflow() []
         i3b = int_is_true(i3)
         guard_true(i3b) []
-        setfield_gc(p1, i1, descr=valuedescr)        
+        setfield_gc(p1, i1, descr=valuedescr)
         escape(i3)
         escape(i3)
         jump(i1, p1, i3)
         """
         expected = """
         [i1, p1, i3]
-        setfield_gc(p1, i1, descr=valuedescr)        
+        setfield_gc(p1, i1, descr=valuedescr)
         escape(i3)
         escape(i3)
         jump(i1, p1, i3)
@@ -2830,7 +2830,7 @@
         ops = """
         [p8, p11, i24]
         p26 = new_with_vtable(ConstClass(node_vtable))
-        setfield_gc(p26, i24, descr=adescr)        
+        setfield_gc(p26, i24, descr=adescr)
         i34 = getfield_gc_pure(p11, descr=valuedescr)
         i35 = getfield_gc_pure(p26, descr=adescr)
         i36 = int_add_ovf(i34, i35)
@@ -2839,10 +2839,10 @@
         """
         expected = """
         [p8, p11, i26]
-        jump(p8, p11, i26)        
-        """
-        self.optimize_loop(ops, expected)
-        
+        jump(p8, p11, i26)
+        """
+        self.optimize_loop(ops, expected)
+
     def test_ovf_guard_in_short_preamble2(self):
         ops = """
         [p8, p11, p12]
@@ -3191,13 +3191,18 @@
         jump(p1, i4, i3)
         '''
         expected = '''
+        [p1, i4, i3, i5]
+        setfield_gc(p1, i5, descr=valuedescr)
+        jump(p1, i3, i5, i5)
+        '''
+        preamble = '''
         [p1, i1, i4]
         setfield_gc(p1, i1, descr=valuedescr)
         i3 = call(p1, descr=plaincalldescr)
         setfield_gc(p1, i3, descr=valuedescr)
-        jump(p1, i4, i3)
+        jump(p1, i4, i3, i3)
         '''
-        self.optimize_loop(ops, expected, expected)
+        self.optimize_loop(ops, expected, preamble)
 
     def test_call_pure_invalidates_heap_knowledge(self):
         # CALL_PURE should still force the setfield_gc() to occur before it
@@ -3209,21 +3214,20 @@
         jump(p1, i4, i3)
         '''
         expected = '''
+        [p1, i4, i3, i5]
+        setfield_gc(p1, i4, descr=valuedescr)
+        jump(p1, i3, i5, i5)
+        '''
+        preamble = '''
         [p1, i1, i4]
         setfield_gc(p1, i1, descr=valuedescr)
         i3 = call(p1, descr=plaincalldescr)
         setfield_gc(p1, i1, descr=valuedescr)
-        jump(p1, i4, i3)
+        jump(p1, i4, i3, i3)
         '''
-        self.optimize_loop(ops, expected, expected)
+        self.optimize_loop(ops, expected, preamble)
 
     def test_call_pure_constant_folding(self):
-        # CALL_PURE is not marked as is_always_pure(), because it is wrong
-        # to call the function arbitrary many times at arbitrary points in
-        # time.  Check that it is either constant-folded (and replaced by
-        # the result of the call, recorded as the first arg), or turned into
-        # a regular CALL.
-        # XXX can this test be improved with unrolling?
         arg_consts = [ConstInt(i) for i in (123456, 4, 5, 6)]
         call_pure_results = {tuple(arg_consts): ConstInt(42)}
         ops = '''
@@ -3239,14 +3243,13 @@
         escape(i1)
         escape(i2)
         i4 = call(123456, 4, i0, 6, descr=plaincalldescr)
-        jump(i0, i4)
+        jump(i0, i4, i4)
         '''
         expected = '''
-        [i0, i2]
+        [i0, i4, i5]
         escape(42)
-        escape(i2)
-        i4 = call(123456, 4, i0, 6, descr=plaincalldescr)
-        jump(i0, i4)
+        escape(i4)
+        jump(i0, i5, i5)
         '''
         self.optimize_loop(ops, expected, preamble, call_pure_results)
 
@@ -3270,18 +3273,43 @@
         escape(i2)
         i4 = call(123456, 4, i0, 6, descr=plaincalldescr)
         guard_no_exception() []
-        jump(i0, i4)
+        jump(i0, i4, i4)
         '''
         expected = '''
-        [i0, i2]
+        [i0, i2, i3]
         escape(42)
         escape(i2)
-        i4 = call(123456, 4, i0, 6, descr=plaincalldescr)
-        guard_no_exception() []
-        jump(i0, i4)
+        jump(i0, i3, i3)
         '''
         self.optimize_loop(ops, expected, preamble, call_pure_results)
 
+    def test_call_pure_returning_virtual(self):
+        # XXX: This kind of loop invaraint call_pure will be forced
+        #      both in the preamble and in the peeled loop
+        ops = '''
+        [p1, i1, i2]
+        p2 = call_pure(0, p1, i1, i2, descr=strslicedescr)
+        escape(p2)
+        jump(p1, i1, i2)
+        '''
+        preamble = '''
+        [p1, i1, i2]
+        i6 = int_sub(i2, i1)
+        p2 = newstr(i6)
+        copystrcontent(p1, p2, i1, 0, i6)
+        escape(p2)
+        jump(p1, i1, i2, i6)
+        '''
+        expected = '''
+        [p1, i1, i2, i6]
+        p2 = newstr(i6)
+        copystrcontent(p1, p2, i1, 0, i6)
+        escape(p2)
+        jump(p1, i1, i2, i6)
+        '''
+        self.optimize_loop(ops, expected, preamble)
+        
+
     # ----------
 
     def test_vref_nonvirtual_nonescape(self):
@@ -5150,14 +5178,14 @@
         [i0, i1, i10, i11, i2, i3, i4]
         escape(i2)
         escape(i3)
-        escape(i4)        
+        escape(i4)
         i24 = int_mul_ovf(i10, i11)
         guard_no_overflow() []
         i23 = int_sub_ovf(i10, i11)
         guard_no_overflow() []
         i22 = int_add_ovf(i10, i11)
         guard_no_overflow() []
-        jump(i0, i1, i10, i11, i2, i3, i4) 
+        jump(i0, i1, i10, i11, i2, i3, i4)
         """
         self.optimize_loop(ops, expected)
 
@@ -5366,6 +5394,8 @@
         """
         self.optimize_strunicode_loop(ops, expected, expected)
 
+    # XXX Should some of the call's below now be call_pure?
+
     def test_str_concat_1(self):
         ops = """
         [p1, p2]
@@ -5699,14 +5729,14 @@
         ops = """
         [p0, i0]
         i1 = unicodegetitem(p0, i0)
-        i10 = unicodegetitem(p0, i0)        
+        i10 = unicodegetitem(p0, i0)
         i2 = int_lt(i1, 0)
         guard_false(i2) []
         jump(p0, i0)
         """
         expected = """
         [p0, i0]
-        i1 = unicodegetitem(p0, i0)        
+        i1 = unicodegetitem(p0, i0)
         jump(p0, i0)
         """
         self.optimize_loop(ops, expected)
@@ -5865,7 +5895,7 @@
         """
         preamble = """
         [p1, i1, i2, p3]
-        guard_nonnull(p3) []        
+        guard_nonnull(p3) []
         i4 = int_sub(i2, i1)
         i0 = call(0, p1, i1, i4, p3, descr=streq_slice_nonnull_descr)
         escape(i0)
@@ -6474,7 +6504,7 @@
         setfield_gc(p3, i1, descr=adescr)
         setfield_gc(p3, i2, descr=bdescr)
         i5 = int_gt(ii, 42)
-        guard_true(i5) []        
+        guard_true(i5) []
         jump(p0, p1, p3, ii2, ii, i1, i2)
         """
         self.optimize_loop(ops, expected)
@@ -6500,7 +6530,7 @@
         p1 = getfield_gc(p0, descr=nextdescr)
         guard_nonnull_class(p1, ConstClass(node_vtable)) []
         p2 = getfield_gc(p1, descr=nextdescr)
-        guard_nonnull_class(p2, ConstClass(node_vtable)) []        
+        guard_nonnull_class(p2, ConstClass(node_vtable)) []
         jump(p0)
         """
         expected = """
@@ -6514,11 +6544,11 @@
         guard_class(p1, ConstClass(node_vtable)) []
         p2 = getfield_gc(p1, descr=nextdescr)
         guard_nonnull(p2) []
-        guard_class(p2, ConstClass(node_vtable)) []        
+        guard_class(p2, ConstClass(node_vtable)) []
         jump(p0)
         """
         self.optimize_loop(ops, expected, expected_short=short)
-        
+
     def test_forced_virtual_pure_getfield(self):
         ops = """
         [p0]
@@ -6582,7 +6612,7 @@
         jump(p1, i2)
         """
         self.optimize_loop(ops, expected)
-        
+
     def test_loopinvariant_strlen(self):
         ops = """
         [p9]
@@ -6715,7 +6745,7 @@
         [p0, p1]
         p2 = new_with_vtable(ConstClass(node_vtable))
         p3 = new_with_vtable(ConstClass(node_vtable))
-        setfield_gc(p2, p3, descr=nextdescr) 
+        setfield_gc(p2, p3, descr=nextdescr)
         jump(p2, p3)
         """
         expected = """
@@ -6734,7 +6764,7 @@
         jump(p2, i2)
         """
         expected = """
-        [p1]        
+        [p1]
         p2 = getarrayitem_gc(p1, 7, descr=<GcPtrArrayDescr>)
         i1 = arraylen_gc(p1)
         jump(p2)
@@ -6775,8 +6805,8 @@
         jump(p0, p2, p1)
         """
         self.optimize_loop(ops, expected, expected_short=short)
-        
-        
+
+
     def test_loopinvariant_constant_strgetitem(self):
         ops = """
         [p0]
@@ -6830,11 +6860,11 @@
         expected = """
         [p0, i22, p1]
         call(i22, descr=nonwritedescr)
-        i3 = unicodelen(p1) # Should be killed by backend        
+        i3 = unicodelen(p1) # Should be killed by backend
         jump(p0, i22, p1)
         """
         self.optimize_loop(ops, expected, expected_short=short)
-        
+
     def test_propagate_virtual_arryalen(self):
         ops = """
         [p0]
@@ -6903,7 +6933,7 @@
         [p0, p1, p10, p11]
         i1 = arraylen_gc(p10, descr=arraydescr)
         getarrayitem_gc(p11, 1, descr=arraydescr)
-        call(i1, descr=nonwritedescr)        
+        call(i1, descr=nonwritedescr)
         jump(p1, p0, p11, p10)
         """
         self.optimize_loop(ops, expected)
@@ -6912,20 +6942,20 @@
         ops = """
         [p5]
         i10 = getfield_gc(p5, descr=valuedescr)
-        call(i10, descr=nonwritedescr) 
+        call(i10, descr=nonwritedescr)
         setfield_gc(p5, 1, descr=valuedescr)
         jump(p5)
         """
         preamble = """
         [p5]
         i10 = getfield_gc(p5, descr=valuedescr)
-        call(i10, descr=nonwritedescr) 
+        call(i10, descr=nonwritedescr)
         setfield_gc(p5, 1, descr=valuedescr)
         jump(p5)
         """
         expected = """
         [p5]
-        call(1, descr=nonwritedescr) 
+        call(1, descr=nonwritedescr)
         jump(p5)
         """
         self.optimize_loop(ops, expected, preamble)
@@ -6963,7 +6993,7 @@
         [p9]
         call_assembler(0, descr=asmdescr)
         i18 = getfield_gc(p9, descr=valuedescr)
-        guard_value(i18, 0) []        
+        guard_value(i18, 0) []
         jump(p9)
         """
         self.optimize_loop(ops, expected)
@@ -6992,17 +7022,17 @@
         i10 = getfield_gc(p5, descr=valuedescr)
         i11 = getfield_gc(p6, descr=nextdescr)
         call(i10, i11, descr=nonwritedescr)
-        setfield_gc(p6, i10, descr=nextdescr)        
+        setfield_gc(p6, i10, descr=nextdescr)
         jump(p5, p6)
         """
         expected = """
         [p5, p6, i10, i11]
         call(i10, i11, descr=nonwritedescr)
-        setfield_gc(p6, i10, descr=nextdescr)        
+        setfield_gc(p6, i10, descr=nextdescr)
         jump(p5, p6, i10, i10)
         """
         self.optimize_loop(ops, expected)
-        
+
     def test_cached_pure_func_of_equal_fields(self):
         ops = """
         [p5, p6]
@@ -7011,18 +7041,18 @@
         i12 = int_add(i10, 7)
         i13 = int_add(i11, 7)
         call(i12, i13, descr=nonwritedescr)
-        setfield_gc(p6, i10, descr=nextdescr)        
+        setfield_gc(p6, i10, descr=nextdescr)
         jump(p5, p6)
         """
         expected = """
         [p5, p6, i14, i12, i10]
         i13 = int_add(i14, 7)
         call(i12, i13, descr=nonwritedescr)
-        setfield_gc(p6, i10, descr=nextdescr)        
+        setfield_gc(p6, i10, descr=nextdescr)
         jump(p5, p6, i10, i12, i10)
         """
         self.optimize_loop(ops, expected)
-        
+
     def test_forced_counter(self):
         # XXX: VIRTUALHEAP (see above)
         py.test.skip("would be fixed by make heap optimizer aware of virtual setfields")
@@ -7165,7 +7195,7 @@
         expected = """
         [p1, p2, i2, i1]
         call(i2, descr=nonwritedescr)
-        setfield_gc(p2, i1, descr=nextdescr)        
+        setfield_gc(p2, i1, descr=nextdescr)
         jump(p1, p2, i2, i1)
         """
         self.optimize_loop(ops, expected)
@@ -7185,11 +7215,11 @@
         expected = """
         [p1, p2, i2, i1]
         call(i2, descr=nonwritedescr)
-        setfield_gc(p2, i1, descr=valuedescr)        
+        setfield_gc(p2, i1, descr=valuedescr)
         jump(p1, p2, i2, i1)
         """
         self.optimize_loop(ops, expected)
-        
+
 class TestLLtype(OptimizeOptTest, LLtypeMixin):
     pass
-        
+
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
@@ -226,8 +226,9 @@
                 if op and op.result:
                     preamble_value = preamble_optimizer.getvalue(op.result)
                     value = self.optimizer.getvalue(op.result)
-                    imp = ValueImporter(self, preamble_value, op)
-                    self.optimizer.importable_values[value] = imp
+                    if not value.is_virtual():
+                        imp = ValueImporter(self, preamble_value, op)
+                        self.optimizer.importable_values[value] = imp
                     newresult = self.optimizer.getvalue(op.result).get_key_box()
                     if newresult is not op.result:
                         self.short_boxes.alias(newresult, op.result)
diff --git a/pypy/jit/metainterp/optimizeopt/vstring.py b/pypy/jit/metainterp/optimizeopt/vstring.py
--- a/pypy/jit/metainterp/optimizeopt/vstring.py
+++ b/pypy/jit/metainterp/optimizeopt/vstring.py
@@ -296,7 +296,7 @@
 
 
 def copy_str_content(optimizer, srcbox, targetbox,
-                     srcoffsetbox, offsetbox, lengthbox, mode):
+                     srcoffsetbox, offsetbox, lengthbox, mode, need_next_offset=True):
     if isinstance(srcbox, ConstPtr) and isinstance(srcoffsetbox, Const):
         M = 5
     else:
@@ -313,7 +313,10 @@
                                               None))
             offsetbox = _int_add(optimizer, offsetbox, CONST_1)
     else:
-        nextoffsetbox = _int_add(optimizer, offsetbox, lengthbox)
+        if need_next_offset:
+            nextoffsetbox = _int_add(optimizer, offsetbox, lengthbox)
+        else:
+            nextoffsetbox = None
         op = ResOperation(mode.COPYSTRCONTENT, [srcbox, targetbox,
                                                 srcoffsetbox, offsetbox,
                                                 lengthbox], None)
@@ -450,6 +453,30 @@
         lengthbox = value.getstrlen(self.optimizer, mode)
         self.make_equal_to(op.result, self.getvalue(lengthbox))
 
+    def optimize_COPYSTRCONTENT(self, op):
+        self._optimize_COPYSTRCONTENT(op, mode_string)
+    def optimize_COPYUNICODECONTENT(self, op):
+        self._optimize_COPYSTRCONTENT(op, mode_unicode)
+
+    def _optimize_COPYSTRCONTENT(self, op, mode):
+        # args: src dst srcstart dststart length
+        src = self.getvalue(op.getarg(0))
+        dst = self.getvalue(op.getarg(1))
+        srcstart = self.getvalue(op.getarg(2))
+        dststart = self.getvalue(op.getarg(3))
+        length = self.getvalue(op.getarg(4))
+
+        if length.is_constant() and length.box.getint() == 0:
+            return
+        copy_str_content(self.optimizer,
+            src.force_box(),
+            dst.force_box(),
+            srcstart.force_box(),
+            dststart.force_box(),
+            length.force_box(),
+            mode, need_next_offset=False
+        )
+
     def optimize_CALL(self, op):
         # dispatch based on 'oopspecindex' to a method that handles
         # specifically the given oopspec call.  For non-oopspec calls,
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
@@ -210,7 +210,8 @@
                 self.metainterp.clear_exception()
                 resbox = self.execute(rop.%s, b1, b2)
                 self.make_result_of_lastop(resbox)  # same as execute_varargs()
-                self.metainterp.handle_possible_overflow_error()
+                if not isinstance(resbox, Const):
+                    self.metainterp.handle_possible_overflow_error()
                 return resbox
         ''' % (_opimpl, _opimpl.upper())).compile()
 
@@ -401,23 +402,25 @@
         self.metainterp.heapcache.new_array(resbox, lengthbox)
         return resbox
 
-    @arguments("box", "descr", "box")
-    def _opimpl_getarrayitem_gc_any(self, arraybox, arraydescr, indexbox):
+    @specialize.arg(1)
+    def _do_getarrayitem_gc_any(self, op, arraybox, arraydescr, indexbox):
         tobox = self.metainterp.heapcache.getarrayitem(
                 arraybox, arraydescr, indexbox)
         if tobox:
             # sanity check: see whether the current array value
             # corresponds to what the cache thinks the value is
-            resbox = executor.execute(self.metainterp.cpu, self.metainterp,
-                                      rop.GETARRAYITEM_GC, arraydescr, arraybox, indexbox)
+            resbox = executor.execute(self.metainterp.cpu, self.metainterp, op,
+                                      arraydescr, arraybox, indexbox)
             assert resbox.constbox().same_constant(tobox.constbox())
             return tobox
-        resbox = self.execute_with_descr(rop.GETARRAYITEM_GC,
-                                         arraydescr, arraybox, indexbox)
+        resbox = self.execute_with_descr(op, arraydescr, arraybox, indexbox)
         self.metainterp.heapcache.getarrayitem_now_known(
                 arraybox, arraydescr, indexbox, resbox)
         return resbox
 
+    @arguments("box", "descr", "box")
+    def _opimpl_getarrayitem_gc_any(self, arraybox, arraydescr, indexbox):
+        return self._do_getarrayitem_gc_any(rop.GETARRAYITEM_GC, arraybox, arraydescr, indexbox)
 
     opimpl_getarrayitem_gc_i = _opimpl_getarrayitem_gc_any
     opimpl_getarrayitem_gc_r = _opimpl_getarrayitem_gc_any
@@ -433,8 +436,7 @@
 
     @arguments("box", "descr", "box")
     def _opimpl_getarrayitem_gc_pure_any(self, arraybox, arraydescr, indexbox):
-        return self.execute_with_descr(rop.GETARRAYITEM_GC_PURE,
-                                       arraydescr, arraybox, indexbox)
+        return self._do_getarrayitem_gc_any(rop.GETARRAYITEM_GC_PURE, arraybox, arraydescr, indexbox)
 
     opimpl_getarrayitem_gc_pure_i = _opimpl_getarrayitem_gc_pure_any
     opimpl_getarrayitem_gc_pure_r = _opimpl_getarrayitem_gc_pure_any
@@ -866,6 +868,14 @@
     def opimpl_newunicode(self, lengthbox):
         return self.execute(rop.NEWUNICODE, lengthbox)
 
+    @arguments("box", "box", "box", "box", "box")
+    def opimpl_copystrcontent(self, srcbox, dstbox, srcstartbox, dststartbox, lengthbox):
+        return self.execute(rop.COPYSTRCONTENT, srcbox, dstbox, srcstartbox, dststartbox, lengthbox)
+
+    @arguments("box", "box", "box", "box", "box")
+    def opimpl_copyunicodecontent(self, srcbox, dstbox, srcstartbox, dststartbox, lengthbox):
+        return self.execute(rop.COPYUNICODECONTENT, srcbox, dstbox, srcstartbox, dststartbox, lengthbox)
+
 ##    @FixME  #arguments("descr", "varargs")
 ##    def opimpl_residual_oosend_canraise(self, methdescr, varargs):
 ##        return self.execute_varargs(rop.OOSEND, varargs, descr=methdescr,
@@ -1060,6 +1070,18 @@
         return ConstInt(trace_length)
 
     @arguments("box")
+    def _opimpl_isconstant(self, box):
+        return ConstInt(isinstance(box, Const))
+
+    opimpl_int_isconstant = opimpl_ref_isconstant = _opimpl_isconstant
+
+    @arguments("box")
+    def _opimpl_isvirtual(self, box):
+        return ConstInt(self.metainterp.heapcache.is_unescaped(box))
+
+    opimpl_ref_isvirtual = _opimpl_isvirtual
+
+    @arguments("box")
     def opimpl_virtual_ref(self, box):
         # Details on the content of metainterp.virtualref_boxes:
         #
@@ -1673,6 +1695,10 @@
 
     def _record_helper_nonpure_varargs(self, opnum, resbox, descr, argboxes):
         assert resbox is None or isinstance(resbox, Box)
+        if (rop._OVF_FIRST <= opnum <= rop._OVF_LAST and
+            self.last_exc_value_box is None and
+            self._all_constants_varargs(argboxes)):
+            return resbox.constbox()
         # record the operation
         profiler = self.staticdata.profiler
         profiler.count_ops(opnum, RECORDED_OPS)
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
@@ -1,23 +1,25 @@
+import sys
+
 import py
-import sys
-from pypy.rlib.jit import JitDriver, we_are_jitted, hint, dont_look_inside
-from pypy.rlib.jit import loop_invariant, elidable, promote
-from pypy.rlib.jit import jit_debug, assert_green, AssertGreenFailed
-from pypy.rlib.jit import unroll_safe, current_trace_length
+
+from pypy import conftest
+from pypy.jit.codewriter.policy import JitPolicy, StopAtXPolicy
 from pypy.jit.metainterp import pyjitpl, history
+from pypy.jit.metainterp.optimizeopt import ALL_OPTS_DICT
+from pypy.jit.metainterp.test.support import LLJitMixin, OOJitMixin, noConst
+from pypy.jit.metainterp.typesystem import LLTypeHelper, OOTypeHelper
+from pypy.jit.metainterp.warmspot import get_stats
 from pypy.jit.metainterp.warmstate import set_future_value
-from pypy.jit.metainterp.warmspot import get_stats
-from pypy.jit.codewriter.policy import JitPolicy, StopAtXPolicy
-from pypy import conftest
+from pypy.rlib.jit import (JitDriver, we_are_jitted, hint, dont_look_inside,
+    loop_invariant, elidable, promote, jit_debug, assert_green,
+    AssertGreenFailed, unroll_safe, current_trace_length, look_inside_iff,
+    isconstant, isvirtual)
 from pypy.rlib.rarithmetic import ovfcheck
-from pypy.jit.metainterp.typesystem import LLTypeHelper, OOTypeHelper
 from pypy.rpython.lltypesystem import lltype, llmemory, rffi
 from pypy.rpython.ootypesystem import ootype
-from pypy.jit.metainterp.optimizeopt import ALL_OPTS_DICT
-from pypy.jit.metainterp.test.support import LLJitMixin, OOJitMixin, noConst
+
 
 class BasicTests:
-
     def test_basic(self):
         def f(x, y):
             return x + y
@@ -99,14 +101,14 @@
                 myjitdriver.jit_merge_point(x=x, y=y, res=res)
                 res += x * x
                 x += 1
-                res += x * x                
+                res += x * x
                 y -= 1
             return res
         res = self.meta_interp(f, [6, 7])
         assert res == 1323
         self.check_loop_count(1)
         self.check_loops(int_mul=1)
-        
+
     def test_loop_variant_mul_ovf(self):
         myjitdriver = JitDriver(greens = [], reds = ['y', 'res', 'x'])
         def f(x, y):
@@ -1372,7 +1374,7 @@
             return x
         res = self.meta_interp(f, [299], listops=True)
         assert res == f(299)
-        self.check_loops(guard_class=0, guard_value=3)        
+        self.check_loops(guard_class=0, guard_value=3)
         self.check_loops(guard_class=0, guard_value=6, everywhere=True)
 
     def test_merge_guardnonnull_guardclass(self):
@@ -2118,7 +2120,7 @@
             return sa
         res = self.meta_interp(f, [32, 7])
         assert res == f(32, 7)
-        
+
     def test_caching_setarrayitem_fixed(self):
         myjitdriver = JitDriver(greens = [], reds = ['sa', 'i', 'n', 'a', 'node'])
         def f(n, a):
@@ -2138,7 +2140,7 @@
             return sa
         res = self.meta_interp(f, [32, 7])
         assert res == f(32, 7)
-        
+
     def test_caching_setarrayitem_var(self):
         myjitdriver = JitDriver(greens = [], reds = ['sa', 'i', 'n', 'a', 'b', 'node'])
         def f(n, a, b):
@@ -2668,7 +2670,7 @@
             myjitdriver.set_param('threshold', 3)
             myjitdriver.set_param('trace_eagerness', 1)
             myjitdriver.set_param('retrace_limit', 5)
-            myjitdriver.set_param('function_threshold', -1)            
+            myjitdriver.set_param('function_threshold', -1)
             pc = sa = i = 0
             while pc < len(bytecode):
                 myjitdriver.jit_merge_point(pc=pc, n=n, sa=sa, i=i)
@@ -2693,12 +2695,12 @@
         def g(n1, n2):
             for i in range(10):
                 f(n1)
-            for i in range(10):                
+            for i in range(10):
                 f(n2)
 
         nn = [10, 3]
         assert self.meta_interp(g, nn) == g(*nn)
-        
+
         # The attempts of retracing first loop will end up retracing the
         # second and thus fail 5 times, saturating the retrace_count. Instead a
         # bridge back to the preamble of the first loop is produced. A guard in
@@ -2709,7 +2711,7 @@
         self.check_tree_loop_count(2 + 3)
 
         # FIXME: Add a gloabl retrace counter and test that we are not trying more than 5 times.
-        
+
         def g(n):
             for i in range(n):
                 for j in range(10):
@@ -2945,15 +2947,15 @@
             a = [0, 1, 2, 3, 4]
             while i < n:
                 myjitdriver.jit_merge_point(sa=sa, n=n, a=a, i=i)
-                if i < n/2:
+                if i < n / 2:
                     sa += a[4]
-                elif i == n/2:
+                elif i == n / 2:
                     a.pop()
                 i += 1
         res = self.meta_interp(f, [32])
         assert res == f(32)
         self.check_loops(arraylen_gc=2)
-        
+
 class TestOOtype(BasicTests, OOJitMixin):
 
     def test_oohash(self):
@@ -3173,7 +3175,7 @@
         res = self.meta_interp(f, [32])
         assert res == f(32)
         self.check_tree_loop_count(3)
-        
+
     def test_two_loopinvariant_arrays3(self):
         from pypy.rpython.lltypesystem import lltype, llmemory, rffi
         myjitdriver = JitDriver(greens = [], reds = ['sa', 'n', 'i', 'a'])
@@ -3197,7 +3199,7 @@
         res = self.meta_interp(f, [32])
         assert res == f(32)
         self.check_tree_loop_count(2)
-        
+
     def test_two_loopinvariant_arrays_boxed(self):
         class A(object):
             def __init__(self, a):
@@ -3222,7 +3224,7 @@
         res = self.meta_interp(f, [32])
         assert res == f(32)
         self.check_loops(arraylen_gc=2, everywhere=True)
-        
+
     def test_release_gil_flush_heap_cache(self):
         if sys.platform == "win32":
             py.test.skip("needs 'time'")
@@ -3298,5 +3300,114 @@
 
         self.meta_interp(main, [10])
 
+    def test_look_inside_iff_const(self):
+        @look_inside_iff(lambda arg: isconstant(arg))
+        def f(arg):
+            s = 0
+            while arg > 0:
+                s += arg
+                arg -= 1
+            return s
+
+        driver = JitDriver(greens = ['code'], reds = ['n', 'arg', 's'])
+
+        def main(code, n, arg):
+            s = 0
+            while n > 0:
+                driver.jit_merge_point(code=code, n=n, arg=arg, s=s)
+                if code == 0:
+                    s += f(arg)
+                else:
+                    s += f(1)
+                n -= 1
+            return s
+
+        res = self.meta_interp(main, [0, 10, 2], enable_opts='')
+        assert res == main(0, 10, 2)
+        self.check_loops(call=1)
+        res = self.meta_interp(main, [1, 10, 2], enable_opts='')
+        assert res == main(1, 10, 2)
+        self.check_loops(call=0)
+
+    def test_look_inside_iff_virtual(self):
+        # There's no good reason for this to be look_inside_iff, but it's a test!
+        @look_inside_iff(lambda arg, n: isvirtual(arg))
+        def f(arg, n):
+            if n == 100:
+                for i in xrange(n):
+                    n += i
+            return arg.x
+        class A(object):
+            def __init__(self, x):
+                self.x = x
+        driver = JitDriver(greens=['n'], reds=['i', 'a'])
+        def main(n):
+            i = 0
+            a = A(3)
+            while i < 20:
+                driver.jit_merge_point(i=i, n=n, a=a)
+                if n == 0:
+                    i += f(a, n)
+                else:
+                    i += f(A(2), n)
+        res = self.meta_interp(main, [0], enable_opts='')
+        assert res == main(0)
+        self.check_loops(call=1, getfield_gc=0)
+        res = self.meta_interp(main, [1], enable_opts='')
+        assert res == main(1)
+        self.check_loops(call=0, getfield_gc=0)
+
+    def test_reuse_elidable_result(self):
+        driver = JitDriver(reds=['n', 's'], greens = [])
+        def main(n):
+            s = 0
+            while n > 0:
+                driver.jit_merge_point(s=s, n=n)
+                s += len(str(n)) + len(str(n))
+                n -= 1
+            return s
+        res = self.meta_interp(main, [10])
+        assert res == main(10)
+        self.check_loops({
+            'call': 1, 'guard_no_exception': 1, 'guard_true': 1, 'int_add': 2,
+            'int_gt': 1, 'int_sub': 1, 'strlen': 1, 'jump': 1,
+        })
+
+    def test_look_inside_iff_const_getarrayitem_gc_pure(self):
+        driver = JitDriver(greens=['unroll'], reds=['s', 'n'])
+
+        class A(object):
+            _immutable_fields_ = ["x[*]"]
+            def __init__(self, x):
+                self.x = [x]
+
+        @look_inside_iff(lambda x: isconstant(x))
+        def f(x):
+            i = 0
+            for c in x:
+                i += 1
+            return i
+
+        def main(unroll, n):
+            s = 0
+            while n > 0:
+                driver.jit_merge_point(s=s, n=n, unroll=unroll)
+                if unroll:
+                    x = A("xx")
+                else:
+                    x = A("x" * n)
+                s += f(x.x[0])
+                n -= 1
+            return s
+
+        res = self.meta_interp(main, [0, 10])
+        assert res == main(0, 10)
+        # 2 calls, one for f() and one for char_mul
+        self.check_loops(call=2)
+        res = self.meta_interp(main, [1, 10])
+        assert res == main(1, 10)
+        self.check_loops(call=0)
+
+
 class TestLLtype(BaseLLtypeTests, LLJitMixin):
     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
@@ -153,11 +153,7 @@
 
         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": 6,
+        self.check_loops({"call": 7, "guard_false": 1, "guard_no_exception": 6,
                           "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})
diff --git a/pypy/jit/metainterp/test/test_heapcache.py b/pypy/jit/metainterp/test/test_heapcache.py
--- a/pypy/jit/metainterp/test/test_heapcache.py
+++ b/pypy/jit/metainterp/test/test_heapcache.py
@@ -337,6 +337,24 @@
         h.invalidate_caches(rop.SETFIELD_GC, None, [box1, box2])
         assert not h.is_unescaped(box2)
 
+    def test_unescaped_testing(self):
+        h = HeapCache()
+        h.new(box1)
+        h.new(box2)
+        assert h.is_unescaped(box1)
+        assert h.is_unescaped(box2)
+        # Putting a virtual inside of another virtual doesn't escape it.
+        h.invalidate_caches(rop.SETFIELD_GC, None, [box1, box2])
+        assert h.is_unescaped(box2)
+        # Reading a field from a virtual doesn't escape it.
+        h.invalidate_caches(rop.GETFIELD_GC, None, [box1])
+        assert h.is_unescaped(box1)
+        # Escaping a virtual transitively escapes anything inside of it.
+        assert not h.is_unescaped(box3)
+        h.invalidate_caches(rop.SETFIELD_GC, None, [box3, box1])
+        assert not h.is_unescaped(box1)
+        assert not h.is_unescaped(box2)
+
     def test_unescaped_array(self):
         h = HeapCache()
         h.new_array(box1, lengthbox1)
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
@@ -34,7 +34,7 @@
                 l = [x + 1]
                 n -= 1
             return l[0]
-        
+
         res = self.meta_interp(f, [10], listops=True)
         assert res == f(10)
         self.check_all_virtualized()
@@ -60,7 +60,7 @@
 
     def test_ll_fixed_setitem_fast(self):
         jitdriver = JitDriver(greens = [], reds = ['n', 'l'])
-        
+
         def f(n):
             l = [1, 2, 3]
 
@@ -116,7 +116,7 @@
         assert res == f(10)
         py.test.skip("'[non-null] * n' gives a residual call so far")
         self.check_loops(setarrayitem_gc=0, getarrayitem_gc=0, call=0)
-    
+
     def test_arraycopy_simpleoptimize(self):
         def f():
             l = [1, 2, 3, 4]
@@ -208,6 +208,26 @@
         assert res == f(15)
         self.check_loops(guard_exception=0)
 
+    def test_virtual_resize(self):
+        jitdriver = JitDriver(greens = [], reds = ['n', 's'])
+        def f(n):
+            s = 0
+            while n > 0:
+                jitdriver.jit_merge_point(n=n, s=s)
+                lst = []
+                lst += [1]
+                n -= len(lst)
+                s += lst[0]
+                lst.pop()
+                lst.append(1)
+                s /= lst.pop()
+            return s
+        res = self.meta_interp(f, [15], listops=True)
+        assert res == f(15)
+        self.check_loops({"int_add": 1, "int_sub": 1, "int_gt": 1,
+                          "guard_true": 1, "jump": 1})
+
+
 class TestOOtype(ListTests, OOJitMixin):
     pass
 
@@ -236,8 +256,6 @@
             return a * b
         res = self.meta_interp(f, [37])
         assert res == f(37)
-        # There is the one actual field on a, plus 2 getfield's from the list
-        # itself, 1 to get the length (which is then incremented and passed to
-        # the resize func), and then a read of the items field to actually
-        # perform the setarrayitem on
-        self.check_loops(getfield_gc=5, everywhere=True)
+        # There is the one actual field on a, plus several fields on the list
+        # itself
+        self.check_loops(getfield_gc=10, everywhere=True)
diff --git a/pypy/jit/metainterp/test/test_slist.py b/pypy/jit/metainterp/test/test_slist.py
--- a/pypy/jit/metainterp/test/test_slist.py
+++ b/pypy/jit/metainterp/test/test_slist.py
@@ -5,7 +5,6 @@
 class ListTests(object):
 
     def test_basic_list(self):
-        py.test.skip("not yet")
         myjitdriver = JitDriver(greens = [], reds = ['n', 'lst'])
         def f(n):
             lst = []
@@ -34,7 +33,7 @@
             return m
         res = self.interp_operations(f, [11], listops=True)
         assert res == 49
-        self.check_operations_history(call=5)
+        self.check_operations_history(call=3)
 
     def test_list_of_voids(self):
         myjitdriver = JitDriver(greens = [], reds = ['n', 'lst'])
@@ -93,7 +92,7 @@
             return x
         res = self.meta_interp(f, [-2], listops=True)
         assert res == 41
-        self.check_loops(call=1, guard_value=0)
+        self.check_loops(call=0, guard_value=0)
 
 # we don't support resizable lists on ootype
 #class TestOOtype(ListTests, OOJitMixin):
diff --git a/pypy/jit/metainterp/test/test_string.py b/pypy/jit/metainterp/test/test_string.py
--- a/pypy/jit/metainterp/test/test_string.py
+++ b/pypy/jit/metainterp/test/test_string.py
@@ -27,7 +27,7 @@
             return i
         res = self.meta_interp(f, [10, True, _str('h')], listops=True)
         assert res == 5
-        self.check_loops(**{self.CALL: 1, self.CALL_PURE: 0})
+        self.check_loops(**{self.CALL: 1, self.CALL_PURE: 0, 'everywhere': True})
 
     def test_eq_folded(self):
         _str = self._str
@@ -327,7 +327,7 @@
     def test_str_slice_len_surviving(self):
         _str = self._str
         longstring = _str("Unrolling Trouble")
-        mydriver = JitDriver(reds = ['i', 'a', 'sa'], greens = []) 
+        mydriver = JitDriver(reds = ['i', 'a', 'sa'], greens = [])
         def f(a):
             i = sa = a
             while i < len(longstring):
@@ -343,7 +343,7 @@
         fillers = _str("abcdefghijklmnopqrstuvwxyz")
         data = _str("ABCDEFGHIJKLMNOPQRSTUVWXYZ")
 
-        mydriver = JitDriver(reds = ['line', 'noise', 'res'], greens = []) 
+        mydriver = JitDriver(reds = ['line', 'noise', 'res'], greens = [])
         def f():
             line = data
             noise = fillers
@@ -370,7 +370,7 @@
             def __init__(self, value):
                 self.value = value
         mydriver = JitDriver(reds = ['ratio', 'line', 'noise', 'res'],
-                             greens = []) 
+                             greens = [])
         def f():
             line = Str(data)
             noise = Str(fillers)
@@ -408,7 +408,7 @@
             return len(sa)
         assert self.meta_interp(f, [16]) == f(16)
 
-    def test_loop_invariant_string_slize(self):
+    def test_loop_invariant_string_slice(self):
         _str = self._str
         mydriver = JitDriver(reds = ['i', 'n', 'sa', 's', 's1'], greens = [])
         def f(n, c):
@@ -425,7 +425,7 @@
             return sa
         assert self.meta_interp(f, [16, 'a']) == f(16, 'a')
 
-    def test_loop_invariant_string_slize_boxed(self):
+    def test_loop_invariant_string_slice_boxed(self):
         class Str(object):
             def __init__(self, value):
                 self.value = value
@@ -445,7 +445,7 @@
             return sa
         assert self.meta_interp(f, [16, 'a']) == f(16, 'a')
 
-    def test_loop_invariant_string_slize_in_array(self):
+    def test_loop_invariant_string_slice_in_array(self):
         _str = self._str
         mydriver = JitDriver(reds = ['i', 'n', 'sa', 's', 's1'], greens = [])
         def f(n, c):
@@ -513,7 +513,7 @@
                 m -= 1
             return 42
         self.meta_interp(f, [6, 7])
-        self.check_loops(call=3,    # str(), _str(), escape()
+        self.check_loops(call=1,    # escape()
                          newunicode=1, unicodegetitem=0,
                          unicodesetitem=1, copyunicodecontent=1)
 
@@ -536,3 +536,27 @@
         self.check_loops(call_pure=0, call=1,
                          newunicode=0, unicodegetitem=0,
                          unicodesetitem=0, copyunicodecontent=0)
+
+    def test_join_chars(self):
+        jitdriver = JitDriver(reds=['a', 'b', 'c', 'i'], greens=[])
+        def f(a, b, c):
+            i = 0
+            while i < 10:
+                jitdriver.jit_merge_point(a=a, b=b, c=c, i=i)
+                x = []
+                if a:
+                    x.append("a")
+                if b:
+                    x.append("b")
+                if c:
+                    x.append("c")
+                i += len("".join(x))
+            return i
+        res = self.meta_interp(f, [1, 1, 1])
+        assert res == f(True, True, True)
+        # The "".join should be unrolled, since the length of x is known since
+        # it is virtual, ensure there are no calls to ll_join_chars, or
+        # allocations.
+        self.check_loops({
+            "guard_true": 5, "int_is_true": 3, "int_lt": 2, "int_add": 2, "jump": 2,
+        }, everywhere=True)
diff --git a/pypy/jit/metainterp/test/test_tracingopts.py b/pypy/jit/metainterp/test/test_tracingopts.py
--- a/pypy/jit/metainterp/test/test_tracingopts.py
+++ b/pypy/jit/metainterp/test/test_tracingopts.py
@@ -1,7 +1,10 @@
+import sys
+
+from pypy.jit.metainterp.test.support import LLJitMixin
+from pypy.rlib import jit
+from pypy.rlib.rarithmetic import ovfcheck
+
 import py
-import sys
-from pypy.rlib import jit
-from pypy.jit.metainterp.test.support import LLJitMixin
 
 
 class TestLLtype(LLJitMixin):
@@ -573,3 +576,18 @@
         res = self.interp_operations(fn, [3])
         assert res == 24
         self.check_operations_history(getarrayitem_gc=0)
+
+    def test_fold_int_add_ovf(self):
+        def fn(n):
+            jit.promote(n)
+            try:
+                n = ovfcheck(n + 1)
+            except OverflowError:
+                return 12
+            else:
+                return n
+        res = self.interp_operations(fn, [3])
+        assert res == 4
+        self.check_operations_history(int_add_ovf=0)
+        res = self.interp_operations(fn, [sys.maxint])
+        assert res == 12
\ No newline at end of file
diff --git a/pypy/jit/tl/pypyjit.py b/pypy/jit/tl/pypyjit.py
--- a/pypy/jit/tl/pypyjit.py
+++ b/pypy/jit/tl/pypyjit.py
@@ -40,7 +40,7 @@
 config.objspace.usemodules.array = False
 config.objspace.usemodules._weakref = True
 config.objspace.usemodules._sre = False
-config.objspace.usemodules._lsprof = True
+config.objspace.usemodules._lsprof = False
 #
 config.objspace.usemodules._ffi = True
 config.objspace.usemodules.micronumpy = False
@@ -77,7 +77,7 @@
 
 def read_code():
     from pypy.module.marshal.interp_marshal import dumps
-    
+
     filename = 'pypyjit_demo.py'
     source = readfile(filename)
     ec = space.getexecutioncontext()
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
@@ -16,7 +16,7 @@
         if modname in ['pypyjit', 'signal', 'micronumpy', 'math', 'exceptions',
                        'imp', 'sys', 'array', '_ffi', 'itertools', 'operator',
                        'posix', '_socket', '_sre', '_lsprof', '_weakref',
-                       '__pypy__', 'cStringIO', '_collections']:
+                       '__pypy__', 'cStringIO', '_collections', 'struct']:
             return True
         return False
 
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
@@ -337,7 +337,9 @@
         assert loop.match_by_id('append', """
             i13 = getfield_gc(p8, descr=<SignedFieldDescr list.length .*>)
             i15 = int_add(i13, 1)
-            call(ConstClass(_ll_list_resize_ge__listPtr_Signed), p8, i15, descr=<VoidCallDescr>)
+            # Will be killed by the backend
+            i17 = arraylen_gc(p7, descr=<GcPtrArrayDescr>)
+            call(ConstClass(_ll_list_resize_ge), p8, i15, descr=<VoidCallDescr>)
             guard_no_exception(descr=...)
             p17 = getfield_gc(p8, descr=<GcPtrFieldDescr list.items .*>)
             p19 = new_with_vtable(ConstClass(W_IntObject))
diff --git a/pypy/module/pypyjit/test_pypy_c/test_containers.py b/pypy/module/pypyjit/test_pypy_c/test_containers.py
--- a/pypy/module/pypyjit/test_pypy_c/test_containers.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_containers.py
@@ -40,10 +40,10 @@
         log = self.run(fn, [1000])
         assert log.result == 300
         loop, = log.loops_by_filename(self.filepath)
-        # check that the call to ll_dict_lookup is not a call_may_force
+        # check that the call to ll_dict_lookup is not a call_may_force, the
+        # gc_id call is hoisted out of the loop, the id of a value obviously
+        # can't change ;)
         assert loop.match_by_id("getitem", """
-            i25 = call(ConstClass(_ll_1_gc_identityhash__objectPtr), p6, descr=...)
-            ...
             i28 = call(ConstClass(ll_dict_lookup__dicttablePtr_objectPtr_Signed), p18, p6, i25, descr=...)
             ...
             p33 = call(ConstClass(ll_get_value__dicttablePtr_Signed), p18, i28, descr=...)
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
@@ -92,7 +92,7 @@
         """)
 
 
-    def test_cached_pure_func_of_equal_fields(self):            
+    def test_cached_pure_func_of_equal_fields(self):
         def main(n):
             class A(object):
                 def __init__(self, val):
@@ -285,3 +285,48 @@
 
         loop, = log.loops_by_id("globalread", is_entry_bridge=True)
         assert len(loop.ops_by_id("globalread")) == 0
+
+    def test_struct_module(self):
+        def main():
+            import struct
+            i = 1
+            while i < 1000:
+                x = struct.unpack("i", struct.pack("i", i))[0] # ID: struct
+                i += x / i
+            return i
+
+        log = self.run(main)
+        assert log.result == main()
+
+        loop, = log.loops_by_id("struct")
+        if sys.maxint == 2 ** 63 - 1:
+            extra = """
+            i8 = int_lt(i4, -2147483648)
+            guard_false(i8, descr=...)
+            """
+        else:
+            extra = ""
+        # This could, of course stand some improvement, to remove all these
+        # arithmatic ops, but we've removed all the core overhead.
+        assert loop.match_by_id("struct", """
+            guard_not_invalidated(descr=...)
+            # struct.pack
+            %(32_bit_only)s
+            i11 = int_and(i4, 255)
+            i13 = int_rshift(i4, 8)
+            i14 = int_and(i13, 255)
+            i16 = int_rshift(i13, 8)
+            i17 = int_and(i16, 255)
+            i19 = int_rshift(i16, 8)
+            i20 = int_and(i19, 255)
+
+            # struct.unpack
+            i22 = int_lshift(i14, 8)
+            i23 = int_or(i11, i22)
+            i25 = int_lshift(i17, 16)
+            i26 = int_or(i23, i25)
+            i28 = int_ge(i20, 128)
+            guard_false(i28, descr=...)
+            i30 = int_lshift(i20, 24)
+            i31 = int_or(i26, i30)
+        """ % {"32_bit_only": extra})
\ No newline at end of file
diff --git a/pypy/module/pypyjit/test_pypy_c/test_string.py b/pypy/module/pypyjit/test_pypy_c/test_string.py
--- a/pypy/module/pypyjit/test_pypy_c/test_string.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_string.py
@@ -1,5 +1,6 @@
 from pypy.module.pypyjit.test_pypy_c.test_00_model import BaseTestPyPyC
 
+
 class TestString(BaseTestPyPyC):
     def test_lookup_default_encoding(self):
         def main(n):
@@ -107,3 +108,52 @@
             --TICK--
             jump(p0, p1, p2, p3, p4, p5, i58, i7, descr=<Loop4>)
         """)
+
+    def test_str_mod(self):
+        def main(n):
+            s = 0
+            while n > 0:
+                s += len('%d %d' % (n, n))
+                n -= 1
+            return s
+
+        log = self.run(main, [1000])
+        assert log.result == main(1000)
+        loop, = log.loops_by_filename(self.filepath)
+        assert loop.match("""
+            i7 = int_gt(i4, 0)
+            guard_true(i7, descr=...)
+            guard_not_invalidated(descr=...)
+            p9 = call(ConstClass(ll_int2dec__Signed), i4, descr=<GcPtrCallDescr>)
+            guard_no_exception(descr=...)
+            i10 = strlen(p9)
+            i11 = int_is_true(i10)
+            guard_true(i11, descr=...)
+            i13 = strgetitem(p9, 0)
+            i15 = int_eq(i13, 45)
+            guard_false(i15, descr=...)
+            i17 = int_sub(0, i10)
+            i19 = int_gt(i10, 23)
+            guard_false(i19, descr=...)
+            p21 = newstr(23)
+            copystrcontent(p9, p21, 0, 0, i10)
+            i25 = int_add(1, i10)
+            i26 = int_gt(i25, 23)
+            guard_false(i26, descr=...)
+            strsetitem(p21, i10, 32)
+            i29 = int_add(i10, 1)
+            i30 = int_add(i10, i25)
+            i31 = int_gt(i30, 23)
+            guard_false(i31, descr=...)
+            copystrcontent(p9, p21, 0, i25, i10)
+            i33 = int_eq(i30, 23)
+            guard_false(i33, descr=...)
+            p35 = call(ConstClass(ll_shrink_array__rpy_stringPtr_Signed), p21, i30, descr=<GcPtrCallDescr>)
+            guard_no_exception(descr=...)
+            i37 = strlen(p35)
+            i38 = int_add_ovf(i5, i37)
+            guard_no_overflow(descr=...)
+            i40 = int_sub(i4, 1)
+            --TICK--
+            jump(p0, p1, p2, p3, i40, i38, descr=<Loop0>)
+        """)
\ No newline at end of file
diff --git a/pypy/module/struct/formatiterator.py b/pypy/module/struct/formatiterator.py
--- a/pypy/module/struct/formatiterator.py
+++ b/pypy/module/struct/formatiterator.py
@@ -1,9 +1,9 @@
-from pypy.interpreter.error import OperationError
-
+from pypy.rlib import jit
 from pypy.rlib.objectmodel import specialize
 from pypy.rlib.rstruct.error import StructError
+from pypy.rlib.rstruct.formatiterator import FormatIterator
 from pypy.rlib.rstruct.standardfmttable import PACK_ACCEPTS_BROKEN_INPUT
-from pypy.rlib.rstruct.formatiterator import FormatIterator
+from pypy.interpreter.error import OperationError
 
 
 class PackFormatIterator(FormatIterator):
@@ -14,15 +14,20 @@
         self.args_index = 0
         self.result = []      # list of characters
 
+    # This *should* be always unroll safe, the only way to get here is by
+    # unroll the interpret function, which means the fmt is const, and thus
+    # this should be const (in theory ;)
+    @jit.unroll_safe
+    @specialize.arg(1)
     def operate(self, fmtdesc, repetitions):
         if fmtdesc.needcount:
             fmtdesc.pack(self, repetitions)
         else:
             for i in range(repetitions):
                 fmtdesc.pack(self)
-    operate._annspecialcase_ = 'specialize:arg(1)'
     _operate_is_specialized_ = True
 
+    @jit.unroll_safe
     def align(self, mask):
         pad = (-len(self.result)) & mask
         for i in range(pad):
@@ -130,13 +135,15 @@
         self.inputpos = 0
         self.result_w = []     # list of wrapped objects
 
+    # See above comment on operate.
+    @jit.unroll_safe
+    @specialize.arg(1)
     def operate(self, fmtdesc, repetitions):
         if fmtdesc.needcount:
             fmtdesc.unpack(self, repetitions)
         else:
             for i in range(repetitions):
                 fmtdesc.unpack(self)
-    operate._annspecialcase_ = 'specialize:arg(1)'
     _operate_is_specialized_ = True
 
     def align(self, mask):
@@ -154,7 +161,6 @@
         self.inputpos = end
         return s
 
+    @specialize.argtype(1)
     def appendobj(self, value):
         self.result_w.append(self.space.wrap(value))
-    appendobj._annspecialcase_ = 'specialize:argtype(1)'
-
diff --git a/pypy/module/struct/interp_struct.py b/pypy/module/struct/interp_struct.py
--- a/pypy/module/struct/interp_struct.py
+++ b/pypy/module/struct/interp_struct.py
@@ -3,6 +3,7 @@
 from pypy.rlib.rstruct.error import StructError
 from pypy.rlib.rstruct.formatiterator import CalcSizeFormatIterator
 
+
 @unwrap_spec(format=str)
 def calcsize(space, format):
     fmtiter = CalcSizeFormatIterator()
diff --git a/pypy/objspace/std/bytearrayobject.py b/pypy/objspace/std/bytearrayobject.py
--- a/pypy/objspace/std/bytearrayobject.py
+++ b/pypy/objspace/std/bytearrayobject.py
@@ -250,7 +250,8 @@
 def repr__Bytearray(space, w_bytearray):
     s = w_bytearray.data
 
-    buf = StringBuilder(50)
+    # Good default if there are no replacements.
+    buf = StringBuilder(len("bytearray(b'')") + len(s))
 
     buf.append("bytearray(b'")
 
diff --git a/pypy/objspace/std/formatting.py b/pypy/objspace/std/formatting.py
--- a/pypy/objspace/std/formatting.py
+++ b/pypy/objspace/std/formatting.py
@@ -1,13 +1,15 @@
 """
 String formatting routines.
 """
-from pypy.rlib.unroll import unrolling_iterable
+from pypy.interpreter.error import OperationError
+from pypy.objspace.std.unicodetype import unicode_from_object
+from pypy.rlib import jit
 from pypy.rlib.rarithmetic import ovfcheck
 from pypy.rlib.rfloat import formatd, DTSF_ALT, isnan, isinf
-from pypy.interpreter.error import OperationError
+from pypy.rlib.rstring import StringBuilder, UnicodeBuilder
+from pypy.rlib.unroll import unrolling_iterable
 from pypy.tool.sourcetools import func_with_new_name
-from pypy.rlib.rstring import StringBuilder, UnicodeBuilder
-from pypy.objspace.std.unicodetype import unicode_from_object
+
 
 class BaseStringFormatter(object):
     def __init__(self, space, values_w, w_valuedict):
@@ -173,6 +175,9 @@
                 raise OperationError(space.w_ValueError,
                                      space.wrap("incomplete format"))
 
+        # Only shows up if we've already started inlining format(), so just
+        # unconditionally unroll this.
+        @jit.unroll_safe
         def getmappingkey(self):
             # return the mapping key in a '%(key)s' specifier
             fmt = self.fmt
@@ -233,6 +238,8 @@
 
             return w_value
 
+        # Same as getmappingkey
+        @jit.unroll_safe
         def peel_flags(self):
             self.f_ljust = False
             self.f_sign  = False
@@ -255,6 +262,8 @@
                     break
                 self.forward()
 
+        # Same as getmappingkey
+        @jit.unroll_safe
         def peel_num(self):
             space = self.space
             c = self.peekchr()
@@ -276,6 +285,7 @@
                 c = self.peekchr()
             return result
 
+        @jit.look_inside_iff(lambda self: jit.isconstant(self.fmt))
         def format(self):
             lgt = len(self.fmt) + 4 * len(self.values_w) + 10
             if do_unicode:
diff --git a/pypy/objspace/std/mapdict.py b/pypy/objspace/std/mapdict.py
--- a/pypy/objspace/std/mapdict.py
+++ b/pypy/objspace/std/mapdict.py
@@ -132,7 +132,10 @@
             cache[selector] = attr
         return attr
 
-    @jit.unroll_safe
+    @jit.look_inside_iff(lambda self, obj, selector, w_value:
+            jit.isconstant(self) and
+            jit.isconstant(selector[0]) and
+            jit.isconstant(selector[1]))
     def add_attr(self, obj, selector, w_value):
         # grumble, jit needs this
         attr = self._get_new_attr(selector[0], selector[1])
diff --git a/pypy/objspace/std/newformat.py b/pypy/objspace/std/newformat.py
--- a/pypy/objspace/std/newformat.py
+++ b/pypy/objspace/std/newformat.py
@@ -3,9 +3,10 @@
 import string
 
 from pypy.interpreter.error import OperationError
-from pypy.rlib import rstring, runicode, rlocale, rarithmetic, rfloat
+from pypy.rlib import rstring, runicode, rlocale, rarithmetic, rfloat, jit
 from pypy.rlib.objectmodel import specialize
 from pypy.rlib.rfloat import copysign, formatd
+from pypy.tool import sourcetools
 
 
 @specialize.argtype(1)
@@ -36,314 +37,321 @@
 ANS_MANUAL = 3
 
 
-class TemplateFormatter(object):
+def make_template_formatting_class():
+    class TemplateFormatter(object):
 
-    _annspecialcase_ = "specialize:ctr_location"
+        parser_list_w = None
 
-    parser_list_w = None
+        def __init__(self, space, is_unicode, template):
+            self.space = space
+            self.is_unicode = is_unicode
+            self.empty = u"" if is_unicode else ""
+            self.template = template
 
-    def __init__(self, space, is_unicode, template):
-        self.space = space
-        self.is_unicode = is_unicode
-        self.empty = u"" if is_unicode else ""
-        self.template = template
+        def build(self, args):
+            self.args, self.kwargs = args.unpack()
+            self.auto_numbering = 0
+            self.auto_numbering_state = ANS_INIT
+            return self._build_string(0, len(self.template), 2)
 
-    def build(self, args):
-        self.args, self.kwargs = args.unpack()
-        self.auto_numbering = 0
-        self.auto_numbering_state = ANS_INIT
-        return self._build_string(0, len(self.template), 2)
+        def _build_string(self, start, end, level):
+            space = self.space
+            if self.is_unicode:
+                out = rstring.UnicodeBuilder()
+            else:
+                out = rstring.StringBuilder()
+            if not level:
+                raise OperationError(space.w_ValueError,
+                                     space.wrap("Recursion depth exceeded"))
+            level -= 1
+            s = self.template
+            return self._do_build_string(start, end, level, out, s)
 
-    def _build_string(self, start, end, level):
-        space = self.space
-        if self.is_unicode:
-            out = rstring.UnicodeBuilder()
-        else:
-            out = rstring.StringBuilder()
-        if not level:
-            raise OperationError(space.w_ValueError,
-                                 space.wrap("Recursion depth exceeded"))
-        level -= 1
-        s = self.template
-        last_literal = i = start
-        while i < end:
-            c = s[i]
-            i += 1
-            if c == "{" or c == "}":
-                at_end = i == end
-                # Find escaped "{" and "}"
-                markup_follows = True
-                if c == "}":
-                    if at_end or s[i] != "}":
-                        raise OperationError(space.w_ValueError,
-                                             space.wrap("Single '}'"))
-                    i += 1
-                    markup_follows = False
-                if c == "{":
-                    if at_end:
-                        raise OperationError(space.w_ValueError,
-                                             space.wrap("Single '{'"))
-                    if s[i] == "{":
+        @jit.look_inside_iff(lambda self, start, end, level, out, s: jit.isconstant(s))
+        def _do_build_string(self, start, end, level, out, s):
+            space = self.space
+            last_literal = i = start
+            while i < end:
+                c = s[i]
+                i += 1
+                if c == "{" or c == "}":
+                    at_end = i == end
+                    # Find escaped "{" and "}"
+                    markup_follows = True
+                    if c == "}":
+                        if at_end or s[i] != "}":
+                            raise OperationError(space.w_ValueError,
+                                                 space.wrap("Single '}'"))
                         i += 1
                         markup_follows = False
-                # Attach literal data
-                out.append_slice(s, last_literal, i - 1)
-                if not markup_follows:
+                    if c == "{":
+                        if at_end:
+                            raise OperationError(space.w_ValueError,
+                                                 space.wrap("Single '{'"))
+                        if s[i] == "{":
+                            i += 1
+                            markup_follows = False
+                    # Attach literal data
+                    out.append_slice(s, last_literal, i - 1)
+                    if not markup_follows:
+                        last_literal = i
+                        continue
+                    nested = 1
+                    field_start = i
+                    recursive = False
+                    while i < end:
+                        c = s[i]
+                        if c == "{":
+                            recursive = True
+                            nested += 1
+                        elif c == "}":
+                            nested -= 1
+                            if not nested:
+                                break
+                        i += 1
+                    if nested:
+                        raise OperationError(space.w_ValueError,
+                                             space.wrap("Unmatched '{'"))
+                    rendered = self._render_field(field_start, i, recursive, level)
+                    out.append(rendered)
+                    i += 1
                     last_literal = i
-                    continue
-                nested = 1
-                field_start = i
-                recursive = False
-                while i < end:
-                    c = s[i]
-                    if c == "{":
-                        recursive = True
-                        nested += 1
-                    elif c == "}":
-                        nested -= 1
-                        if not nested:
-                            break
-                    i += 1
-                if nested:
-                    raise OperationError(space.w_ValueError,
-                                         space.wrap("Unmatched '{'"))
-                rendered = self._render_field(field_start, i, recursive, level)
-                out.append(rendered)
+
+            out.append_slice(s, last_literal, end)
+            return out.build()
+
+        def _parse_field(self, start, end):
+            s = self.template
+            # Find ":" or "!"
+            i = start
+            while i < end:
+                c = s[i]
+                if c == ":" or c == "!":
+                    end_name = i
+                    if c == "!":
+                        i += 1
+                        if i == end:
+                            w_msg = self.space.wrap("expected conversion")
+                            raise OperationError(self.space.w_ValueError, w_msg)
+                        conversion = s[i]
+                        i += 1
+                        if i < end:
+                            if s[i] != ':':
+                                w_msg = self.space.wrap("expected ':' after"
+                                                        " format specifier")
+                                raise OperationError(self.space.w_ValueError,
+                                                     w_msg)
+                            i += 1
+                    else:
+                        conversion = None
+                        i += 1
+                    return s[start:end_name], conversion, i
                 i += 1
-                last_literal = i
+            return s[start:end], None, end
 
-        out.append_slice(s, last_literal, end)
-        return out.build()
-
-    def _parse_field(self, start, end):
-        s = self.template
-        # Find ":" or "!"
-        i = start
-        while i < end:
-            c = s[i]
-            if c == ":" or c == "!":
-                end_name = i
-                if c == "!":
-                    i += 1
-                    if i == end:
-                        w_msg = self.space.wrap("expected conversion")
-                        raise OperationError(self.space.w_ValueError, w_msg)
-                    conversion = s[i]
-                    i += 1
-                    if i < end:
-                        if s[i] != ':':
-                            w_msg = self.space.wrap("expected ':' after"
-                                                    " format specifier")
-                            raise OperationError(self.space.w_ValueError,
-                                                 w_msg)
-                        i += 1
+        def _get_argument(self, name):
+            # First, find the argument.
+            space = self.space
+            i = 0
+            end = len(name)
+            while i < end:
+                c = name[i]
+                if c == "[" or c == ".":
+                    break
+                i += 1
+            empty = not i
+            if empty:
+                index = -1
+            else:
+                index, stop = _parse_int(self.space, name, 0, i)
+                if stop != i:
+                    index = -1
+            use_numeric = empty or index != -1
+            if self.auto_numbering_state == ANS_INIT and use_numeric:
+                if empty:
+                    self.auto_numbering_state = ANS_AUTO
                 else:
-                    conversion = None
-                    i += 1
-                return s[start:end_name], conversion, i
-            i += 1
-        return s[start:end], None, end
-
-    def _get_argument(self, name):
-        # First, find the argument.
-        space = self.space
-        i = 0
-        end = len(name)
-        while i < end:
-            c = name[i]
-            if c == "[" or c == ".":
-                break
-            i += 1
-        empty = not i
-        if empty:
-            index = -1
-        else:
-            index, stop = _parse_int(self.space, name, 0, i)
-            if stop != i:
-                index = -1
-        use_numeric = empty or index != -1
-        if self.auto_numbering_state == ANS_INIT and use_numeric:
-            if empty:
-                self.auto_numbering_state = ANS_AUTO
-            else:
-                self.auto_numbering_state = ANS_MANUAL
-        if use_numeric:
-            if self.auto_numbering_state == ANS_MANUAL:
-                if empty:
-                    msg = "switching from manual to automatic numbering"
+                    self.auto_numbering_state = ANS_MANUAL
+            if use_numeric:
+                if self.auto_numbering_state == ANS_MANUAL:
+                    if empty:
+                        msg = "switching from manual to automatic numbering"
+                        raise OperationError(space.w_ValueError,
+                                             space.wrap(msg))
+                elif not empty:
+                    msg = "switching from automatic to manual numbering"
                     raise OperationError(space.w_ValueError,
                                          space.wrap(msg))
-            elif not empty:
-                msg = "switching from automatic to manual numbering"
-                raise OperationError(space.w_ValueError,
-                                     space.wrap(msg))
-        if empty:
-            index = self.auto_numbering
-            self.auto_numbering += 1
-        if index == -1:
-            kwarg = name[:i]
-            if self.is_unicode:
+            if empty:
+                index = self.auto_numbering
+                self.auto_numbering += 1
+            if index == -1:
+                kwarg = name[:i]
+                if self.is_unicode:
+                    try:
+                        arg_key = kwarg.encode("latin-1")
+                    except UnicodeEncodeError:
+                        # Not going to be found in a dict of strings.
+                        raise OperationError(space.w_KeyError, space.wrap(kwarg))
+                else:
+                    arg_key = kwarg
                 try:
-                    arg_key = kwarg.encode("latin-1")
-                except UnicodeEncodeError:
-                    # Not going to be found in a dict of strings.
-                    raise OperationError(space.w_KeyError, space.wrap(kwarg))
+                    w_arg = self.kwargs[arg_key]
+                except KeyError:
+                    raise OperationError(space.w_KeyError, space.wrap(arg_key))
             else:
-                arg_key = kwarg
-            try:
-                w_arg = self.kwargs[arg_key]
-            except KeyError:
-                raise OperationError(space.w_KeyError, space.wrap(arg_key))
-        else:
-            try:
-                w_arg = self.args[index]
-            except IndexError:
-                w_msg = space.wrap("index out of range")
-                raise OperationError(space.w_IndexError, w_msg)
-        return self._resolve_lookups(w_arg, name, i, end)
+                try:
+                    w_arg = self.args[index]
+                except IndexError:
+                    w_msg = space.wrap("index out of range")
+                    raise OperationError(space.w_IndexError, w_msg)
+            return self._resolve_lookups(w_arg, name, i, end)
 
-    def _resolve_lookups(self, w_obj, name, start, end):
-        # Resolve attribute and item lookups.
-        space = self.space
-        i = start
-        while i < end:
-            c = name[i]
-            if c == ".":
+        def _resolve_lookups(self, w_obj, name, start, end):
+            # Resolve attribute and item lookups.
+            space = self.space
+            i = start
+            while i < end:
+                c = name[i]
+                if c == ".":
+                    i += 1
+                    start = i
+                    while i < end:
+                        c = name[i]
+                        if c == "[" or c == ".":
+                            break
+                        i += 1
+                    if start == i:
+                        w_msg = space.wrap("Empty attribute in format string")
+                        raise OperationError(space.w_ValueError, w_msg)
+                    w_attr = space.wrap(name[start:i])
+                    if w_obj is not None:
+                        w_obj = space.getattr(w_obj, w_attr)
+                    else:
+                        self.parser_list_w.append(space.newtuple([
+                            space.w_True, w_attr]))
+                elif c == "[":
+                    got_bracket = False
+                    i += 1
+                    start = i
+                    while i < end:
+                        c = name[i]
+                        if c == "]":
+                            got_bracket = True
+                            break
+                        i += 1
+                    if not got_bracket:
+                        raise OperationError(space.w_ValueError,
+                                             space.wrap("Missing ']'"))
+                    index, reached = _parse_int(self.space, name, start, i)
+                    if index != -1 and reached == i:
+                        w_item = space.wrap(index)
+                    else:
+                        w_item = space.wrap(name[start:i])
+                    i += 1 # Skip "]"
+                    if w_obj is not None:
+                        w_obj = space.getitem(w_obj, w_item)
+                    else:
+                        self.parser_list_w.append(space.newtuple([
+                            space.w_False, w_item]))
+                else:
+                    msg = "Only '[' and '.' may follow ']'"
+                    raise OperationError(space.w_ValueError, space.wrap(msg))
+            return w_obj
+
+        def formatter_field_name_split(self):
+            space = self.space
+            name = self.template
+            i = 0
+            end = len(name)
+            while i < end:
+                c = name[i]
+                if c == "[" or c == ".":
+                    break
                 i += 1
-                start = i
-                while i < end:
-                    c = name[i]
-                    if c == "[" or c == ".":
-                        break
-                    i += 1
-                if start == i:
-                    w_msg = space.wrap("Empty attribute in format string")
-                    raise OperationError(space.w_ValueError, w_msg)
-                w_attr = space.wrap(name[start:i])
-                if w_obj is not None:
-                    w_obj = space.getattr(w_obj, w_attr)
-                else:
-                    self.parser_list_w.append(space.newtuple([
-                        space.w_True, w_attr]))
-            elif c == "[":
-                got_bracket = False
-                i += 1
-                start = i
-                while i < end:
-                    c = name[i]
-                    if c == "]":
-                        got_bracket = True
-                        break
-                    i += 1
-                if not got_bracket:
-                    raise OperationError(space.w_ValueError,
-                                         space.wrap("Missing ']'"))
-                index, reached = _parse_int(self.space, name, start, i)
-                if index != -1 and reached == i:
-                    w_item = space.wrap(index)
-                else:
-                    w_item = space.wrap(name[start:i])
-                i += 1 # Skip "]"
-                if w_obj is not None:
-                    w_obj = space.getitem(w_obj, w_item)
-                else:
-                    self.parser_list_w.append(space.newtuple([
-                        space.w_False, w_item]))
+            if i == 0:
+                index = -1
             else:
-                msg = "Only '[' and '.' may follow ']'"
-                raise OperationError(space.w_ValueError, space.wrap(msg))
-        return w_obj
+                index, stop = _parse_int(self.space, name, 0, i)
+                if stop != i:
+                    index = -1
+            if index >= 0:
+                w_first = space.wrap(index)
+            else:
+                w_first = space.wrap(name[:i])
+            #
+            self.parser_list_w = []
+            self._resolve_lookups(None, name, i, end)
+            #
+            return space.newtuple([w_first,
+                                   space.iter(space.newlist(self.parser_list_w))])
 
-    def formatter_field_name_split(self):
-        space = self.space
-        name = self.template
-        i = 0
-        end = len(name)
-        while i < end:
-            c = name[i]
-            if c == "[" or c == ".":
-                break
-            i += 1
-        if i == 0:
-            index = -1
-        else:
-            index, stop = _parse_int(self.space, name, 0, i)
-            if stop != i:
-                index = -1
-        if index >= 0:
-            w_first = space.wrap(index)
-        else:
-            w_first = space.wrap(name[:i])
-        #
-        self.parser_list_w = []
-        self._resolve_lookups(None, name, i, end)
-        #
-        return space.newtuple([w_first,
-                               space.iter(space.newlist(self.parser_list_w))])
+        def _convert(self, w_obj, conversion):
+            space = self.space
+            conv = conversion[0]
+            if conv == "r":
+                return space.repr(w_obj)
+            elif conv == "s":
+                if self.is_unicode:
+                    return space.call_function(space.w_unicode, w_obj)
+                return space.str(w_obj)
+            else:
+                raise OperationError(self.space.w_ValueError,
+                                     self.space.wrap("invalid conversion"))
 
-    def _convert(self, w_obj, conversion):
-        space = self.space
-        conv = conversion[0]
-        if conv == "r":
-            return space.repr(w_obj)
-        elif conv == "s":
-            if self.is_unicode:
-                return space.call_function(space.w_unicode, w_obj)
-            return space.str(w_obj)
-        else:
-            raise OperationError(self.space.w_ValueError,
-                                 self.space.wrap("invalid conversion"))
+        def _render_field(self, start, end, recursive, level):
+            name, conversion, spec_start = self._parse_field(start, end)
+            spec = self.template[spec_start:end]
+            #
+            if self.parser_list_w is not None:
+                # used from formatter_parser()
+                if level == 1:    # ignore recursive calls
+                    space = self.space
+                    startm1 = start - 1
+                    assert startm1 >= self.last_end
+                    w_entry = space.newtuple([
+                        space.wrap(self.template[self.last_end:startm1]),
+                        space.wrap(name),
+                        space.wrap(spec),
+                        space.wrap(conversion)])
+                    self.parser_list_w.append(w_entry)
+                    self.last_end = end + 1
+                return self.empty
+            #
+            w_obj = self._get_argument(name)
+            if conversion is not None:
+                w_obj = self._convert(w_obj, conversion)
+            if recursive:
+                spec = self._build_string(spec_start, end, level)
+            w_rendered = self.space.format(w_obj, self.space.wrap(spec))
+            unwrapper = "unicode_w" if self.is_unicode else "str_w"
+            to_interp = getattr(self.space, unwrapper)
+            return to_interp(w_rendered)
 
-    def _render_field(self, start, end, recursive, level):
-        name, conversion, spec_start = self._parse_field(start, end)
-        spec = self.template[spec_start:end]
-        #
-        if self.parser_list_w is not None:
-            # used from formatter_parser()
-            if level == 1:    # ignore recursive calls
-                space = self.space
-                startm1 = start - 1
-                assert startm1 >= self.last_end
-                w_entry = space.newtuple([
-                    space.wrap(self.template[self.last_end:startm1]),
-                    space.wrap(name),
-                    space.wrap(spec),
-                    space.wrap(conversion)])
-                self.parser_list_w.append(w_entry)
-                self.last_end = end + 1
-            return self.empty
-        #
-        w_obj = self._get_argument(name)
-        if conversion is not None:
-            w_obj = self._convert(w_obj, conversion)
-        if recursive:
-            spec = self._build_string(spec_start, end, level)
-        w_rendered = self.space.format(w_obj, self.space.wrap(spec))
-        unwrapper = "unicode_w" if self.is_unicode else "str_w"
-        to_interp = getattr(self.space, unwrapper)
-        return to_interp(w_rendered)
+        def formatter_parser(self):
+            self.parser_list_w = []
+            self.last_end = 0
+            self._build_string(0, len(self.template), 2)
+            #
+            space = self.space
+            if self.last_end < len(self.template):
+                w_lastentry = space.newtuple([
+                    space.wrap(self.template[self.last_end:]),
+                    space.w_None,
+                    space.w_None,
+                    space.w_None])
+                self.parser_list_w.append(w_lastentry)
+            return space.iter(space.newlist(self.parser_list_w))
+    return TemplateFormatter
 
-    def formatter_parser(self):
-        self.parser_list_w = []
-        self.last_end = 0
-        self._build_string(0, len(self.template), 2)
-        #
-        space = self.space
-        if self.last_end < len(self.template):
-            w_lastentry = space.newtuple([
-                space.wrap(self.template[self.last_end:]),
-                space.w_None,
-                space.w_None,
-                space.w_None])
-            self.parser_list_w.append(w_lastentry)
-        return space.iter(space.newlist(self.parser_list_w))
-
+StrTemplateFormatter = make_template_formatting_class()
+UnicodeTemplateFormatter = make_template_formatting_class()
 
 def str_template_formatter(space, template):
-    return TemplateFormatter(space, False, template)
+    return StrTemplateFormatter(space, False, template)
 
 def unicode_template_formatter(space, template):
-    return TemplateFormatter(space, True, template)
+    return UnicodeTemplateFormatter(space, True, template)
 
 
 def format_method(space, w_string, args, is_unicode):
@@ -380,756 +388,759 @@
 
 LONG_DIGITS = string.digits + string.ascii_lowercase
 
-class Formatter(BaseFormatter):
-    """__format__ implementation for builtin types."""
+def make_formatting_class():
+    class Formatter(BaseFormatter):
+        """__format__ implementation for builtin types."""
 
-    _annspecialcase_ = "specialize:ctr_location"
-    _grouped_digits = None
+        _grouped_digits = None
 
-    def __init__(self, space, is_unicode, spec):
-        self.space = space
-        self.is_unicode = is_unicode
-        self.empty = u"" if is_unicode else ""
-        self.spec = spec
+        def __init__(self, space, is_unicode, spec):
+            self.space = space
+            self.is_unicode = is_unicode
+            self.empty = u"" if is_unicode else ""
+            self.spec = spec
 
-    def _is_alignment(self, c):
-        return (c == "<" or
-                c == ">" or
-                c == "=" or
-                c == "^")
+        def _is_alignment(self, c):
+            return (c == "<" or
+                    c == ">" or
+                    c == "=" or
+                    c == "^")
 
-    def _is_sign(self, c):
-        return (c == " " or
-                c == "+" or
-                c == "-")
+        def _is_sign(self, c):
+            return (c == " " or
+                    c == "+" or
+                    c == "-")
 
-    def _parse_spec(self, default_type, default_align):
-        space = self.space
-        self._fill_char = self._lit("\0")[0]
-        self._align = default_align
-        self._alternate = False
-        self._sign = "\0"
-        self._thousands_sep = False
-        self._precision = -1
-        the_type = default_type
-        spec = self.spec
-        if not spec:
-            return True
-        length = len(spec)
-        i = 0
-        got_align = True
-        if length - i >= 2 and self._is_alignment(spec[i + 1]):
-            self._align = spec[i + 1]
-            self._fill_char = spec[i]
-            i += 2
-        elif length - i >= 1 and self._is_alignment(spec[i]):
-            self._align = spec[i]
-            i += 1
-        else:
-            got_align = False
-        if length - i >= 1 and self._is_sign(spec[i]):
-            self._sign = spec[i]
-            i += 1
-        if length - i >= 1 and spec[i] == "#":
-            self._alternate = True
-            i += 1
-        if self._fill_char == "\0" and length - i >= 1 and spec[i] == "0":
-            self._fill_char = self._lit("0")[0]
-            if not got_align:
-                self._align = "="
-            i += 1
-        start_i = i
-        self._width, i = _parse_int(self.space, spec, i, length)
-        if length != i and spec[i] == ",":
-            self._thousands_sep = True
-            i += 1
-        if length != i and spec[i] == ".":
-            i += 1
-            self._precision, i = _parse_int(self.space, spec, i, length)
-            if self._precision == -1:
+        def _parse_spec(self, default_type, default_align):
+            space = self.space
+            self._fill_char = self._lit("\0")[0]
+            self._align = default_align
+            self._alternate = False
+            self._sign = "\0"
+            self._thousands_sep = False
+            self._precision = -1
+            the_type = default_type
+            spec = self.spec
+            if not spec:
+                return True
+            length = len(spec)
+            i = 0
+            got_align = True
+            if length - i >= 2 and self._is_alignment(spec[i + 1]):
+                self._align = spec[i + 1]
+                self._fill_char = spec[i]
+                i += 2
+            elif length - i >= 1 and self._is_alignment(spec[i]):
+                self._align = spec[i]
+                i += 1
+            else:
+                got_align = False
+            if length - i >= 1 and self._is_sign(spec[i]):
+                self._sign = spec[i]
+                i += 1
+            if length - i >= 1 and spec[i] == "#":
+                self._alternate = True
+                i += 1
+            if self._fill_char == "\0" and length - i >= 1 and spec[i] == "0":
+                self._fill_char = self._lit("0")[0]
+                if not got_align:
+                    self._align = "="
+                i += 1
+            start_i = i
+            self._width, i = _parse_int(self.space, spec, i, length)
+            if length != i and spec[i] == ",":
+                self._thousands_sep = True
+                i += 1
+            if length != i and spec[i] == ".":
+                i += 1
+                self._precision, i = _parse_int(self.space, spec, i, length)
+                if self._precision == -1:
+                    raise OperationError(space.w_ValueError,
+                                         space.wrap("no precision given"))
+            if length - i > 1:
                 raise OperationError(space.w_ValueError,
-                                     space.wrap("no precision given"))
-        if length - i > 1:
-            raise OperationError(space.w_ValueError,
-                                 space.wrap("invalid format spec"))
-        if length - i == 1:
-            presentation_type = spec[i]
-            if self.is_unicode:
-                try:
-                    the_type = spec[i].encode("ascii")[0]
-                except UnicodeEncodeError:
+                                     space.wrap("invalid format spec"))
+            if length - i == 1:
+                presentation_type = spec[i]
+                if self.is_unicode:
+                    try:
+                        the_type = spec[i].encode("ascii")[0]
+                    except UnicodeEncodeError:
+                        raise OperationError(space.w_ValueError,
+                                             space.wrap("invalid presentation type"))
+                else:
+                    the_type = presentation_type
+                i += 1
+            self._type = the_type
+            if self._thousands_sep:
+                tp = self._type
+                if (tp == "d" or
+                    tp == "e" or
+                    tp == "f" or
+                    tp == "g" or
+                    tp == "E" or
+                    tp == "G" or
+                    tp == "%" or
+                    tp == "F" or
+                    tp == "\0"):
+                    # ok
+                    pass
+                else:
                     raise OperationError(space.w_ValueError,
-                                         space.wrap("invalid presentation type"))
+                                         space.wrap("invalid type with ','"))
+            return False
+
+        def _calc_padding(self, string, length):
+            """compute left and right padding, return total width of string"""
+            if self._width != -1 and length < self._width:
+                total = self._width
             else:
-                the_type = presentation_type
-            i += 1
-        self._type = the_type
-        if self._thousands_sep:
-            tp = self._type
-            if (tp == "d" or
-                tp == "e" or
-                tp == "f" or
-                tp == "g" or
-                tp == "E" or
-                tp == "G" or
-                tp == "%" or
-                tp == "F" or
-                tp == "\0"):
-                # ok
-                pass
+                total = length
+            align = self._align
+            if align == ">":
+                left = total - length
+            elif align == "^":
+                left = (total - length) / 2
+            elif align == "<" or align == "=":
+                left = 0
             else:
-                raise OperationError(space.w_ValueError,
-                                     space.wrap("invalid type with ','"))
-        return False
+                raise AssertionError("shouldn't be here")
+            right = total - length - left
+            self._left_pad = left
+            self._right_pad = right
+            return total
 
-    def _calc_padding(self, string, length):
-        """compute left and right padding, return total width of string"""
-        if self._width != -1 and length < self._width:
-            total = self._width
-        else:
-            total = length
-        align = self._align
-        if align == ">":
-            left = total - length
-        elif align == "^":
-            left = (total - length) / 2
-        elif align == "<" or align == "=":
-            left = 0
-        else:
-            raise AssertionError("shouldn't be here")
-        right = total - length - left
-        self._left_pad = left
-        self._right_pad = right
-        return total
-
-    def _lit(self, s):
-        if self.is_unicode:
-            return s.decode("ascii")
-        else:
-            return s
-
-    def _pad(self, string):
-        builder = self._builder()
-        builder.append_multiple_char(self._fill_char, self._left_pad)
-        builder.append(string)
-        builder.append_multiple_char(self._fill_char, self._right_pad)
-        return builder.build()
-
-    def _builder(self):
-        if self.is_unicode:
-            return rstring.UnicodeBuilder()
-        else:
-            return rstring.StringBuilder()
-
-    def _unknown_presentation(self, tp):
-        msg = "unknown presentation for %s: '%s'"
-        w_msg = self.space.wrap(msg  % (tp, self._type))
-        raise OperationError(self.space.w_ValueError, w_msg)
-
-    def format_string(self, string):
-        space = self.space
-        if self._parse_spec("s", "<"):
-            return space.wrap(string)
-        if self._type != "s":
-            self._unknown_presentation("string")
-        if self._sign != "\0":
-            msg = "Sign not allowed in string format specifier"
-            raise OperationError(space.w_ValueError, space.wrap(msg))
-        if self._alternate:
-            msg = "Alternate form not allowed in string format specifier"
-            raise OperationError(space.w_ValueError, space.wrap(msg))
-        if self._align == "=":
-            msg = "'=' alignment not allowed in string format specifier"
-            raise OperationError(space.w_ValueError, space.wrap(msg))
-        length = len(string)
-        precision = self._precision
-        if precision != -1 and length >= precision:
-            assert precision >= 0
-            length = precision
-            string = string[:precision]
-        if self._fill_char == "\0":
-            self._fill_char = self._lit(" ")[0]
-        self._calc_padding(string, length)
-        return space.wrap(self._pad(string))
-
-    def _get_locale(self, tp):
-        space = self.space
-        if tp == "n":
-            dec, thousands, grouping = rlocale.numeric_formatting()
-        elif self._thousands_sep:
-            dec = "."
-            thousands = ","
-            grouping = "\3\0"
-        else:
-            dec = "."
-            thousands = ""
-            grouping = "\256"
-        if self.is_unicode:
-            self._loc_dec = dec.decode("ascii")
-            self._loc_thousands = thousands.decode("ascii")
-        else:
-            self._loc_dec = dec
-            self._loc_thousands = thousands
-        self._loc_grouping = grouping
-
-    def _calc_num_width(self, n_prefix, sign_char, to_number, n_number,
-                        n_remainder, has_dec, digits):
-        """Calculate widths of all parts of formatted number.
-
-        Output will look like:
-
-            <lpadding> <sign> <prefix> <spadding> <grouped_digits> <decimal>
-            <remainder> <rpadding>
-
-        sign is computed from self._sign, and the sign of the number
-        prefix is given
-        digits is known
-        """
-        spec = NumberSpec()
-        spec.n_digits = n_number - n_remainder - has_dec
-        spec.n_prefix = n_prefix
-        spec.n_lpadding = 0
-        spec.n_decimal = int(has_dec)
-        spec.n_remainder = n_remainder
-        spec.n_spadding = 0
-        spec.n_rpadding = 0
-        spec.n_min_width = 0
-        spec.n_total = 0
-        spec.sign = "\0"
-        spec.n_sign = 0
-        sign = self._sign
-        if sign == "+":
-            spec.n_sign = 1
-            spec.sign = "-" if sign_char == "-" else "+"
-        elif sign == " ":
-            spec.n_sign = 1
-            spec.sign = "-" if sign_char == "-" else " "
-        elif sign_char == "-":
-            spec.n_sign = 1
-            spec.sign = "-"
-        extra_length = (spec.n_sign + spec.n_prefix + spec.n_decimal +
-                        spec.n_remainder) # Not padding or digits
-        if self._fill_char == "0" and self._align == "=":
-            spec.n_min_width = self._width - extra_length
-        if self._loc_thousands:
-            self._group_digits(spec, digits[to_number:])
-            n_grouped_digits = len(self._grouped_digits)
-        else:
-            n_grouped_digits = spec.n_digits
-        n_padding = self._width - (extra_length + n_grouped_digits)
-        if n_padding > 0:
-            align = self._align
-            if align == "<":
-                spec.n_rpadding = n_padding
-            elif align == ">":
-                spec.n_lpadding = n_padding
-            elif align == "^":
-                spec.n_lpadding = n_padding // 2
-                spec.n_rpadding = n_padding - spec.n_lpadding
-            elif align == "=":
-                spec.n_spadding = n_padding
-            else:
-                raise AssertionError("shouldn't reach")
-        spec.n_total = spec.n_lpadding + spec.n_sign + spec.n_prefix + \
-                       spec.n_spadding + n_grouped_digits + \
-                       spec.n_decimal + spec.n_remainder + spec.n_rpadding
-        return spec
-
-    def _fill_digits(self, buf, digits, d_state, n_chars, n_zeros,
-                     thousands_sep):
-        if thousands_sep:
-            for c in thousands_sep:
-                buf.append(c)
-        for i in range(d_state - 1, d_state - n_chars - 1, -1):
-            buf.append(digits[i])
-        for i in range(n_zeros):
-            buf.append("0")
-
-    def _group_digits(self, spec, digits):
-        buf = []
-        grouping = self._loc_grouping
-        min_width = spec.n_min_width
-        grouping_state = 0
-        count = 0
-        left = spec.n_digits
-        n_ts = len(self._loc_thousands)
-        need_separator = False
-        done = False
-        groupings = len(grouping)
-        previous = 0
-        while True:
-            group = ord(grouping[grouping_state])
-            if group > 0:
-                if group == 256:
-                    break
-                grouping_state += 1
-                previous = group
-            else:
-                group = previous
-            final_grouping = min(group, max(left, max(min_width, 1)))
-            n_zeros = max(0, final_grouping - left)
-            n_chars = max(0, min(left, final_grouping))
-            ts = self._loc_thousands if need_separator else None
-            self._fill_digits(buf, digits, left, n_chars, n_zeros, ts)
-            need_separator = True
-            left -= n_chars
-            min_width -= final_grouping
-            if left <= 0 and min_width <= 0:
-                done = True
-                break
-            min_width -= n_ts
-        if not done:
-            group = max(max(left, min_width), 1)
-            n_zeros = max(0, group - left)
-            n_chars = max(0, min(left, group))
-            ts = self._loc_thousands if need_separator else None
-            self._fill_digits(buf, digits, left, n_chars, n_zeros, ts)
-        buf.reverse()
-        self._grouped_digits = self.empty.join(buf)
-
-    def _upcase_string(self, s):
-        buf = []
-        for c in s:
-            index = ord(c)
-            if ord("a") <= index <= ord("z"):
-                c = chr(index - 32)
-            buf.append(c)
-        return self.empty.join(buf)
-
-
-    def _fill_number(self, spec, num, to_digits, to_prefix, fill_char,
-                     to_remainder, upper, grouped_digits=None):
-        out = self._builder()
-        if spec.n_lpadding:
-            out.append_multiple_char(fill_char[0], spec.n_lpadding)
-        if spec.n_sign:
-            if self.is_unicode:
-                sign = spec.sign.decode("ascii")
-            else:
-                sign = spec.sign
-            out.append(sign)
-        if spec.n_prefix:
-            pref = num[to_prefix:to_prefix + spec.n_prefix]
-            if upper:
-                pref = self._upcase_string(pref)
-            out.append(pref)
-        if spec.n_spadding:
-            out.append_multiple_char(fill_char[0], spec.n_spadding)
-        if spec.n_digits != 0:
-            if self._loc_thousands:
-                if grouped_digits is not None:
-                    digits = grouped_digits
-                else:
-                    digits = self._grouped_digits
-                    assert digits is not None
-            else:
-                stop = to_digits + spec.n_digits
-                assert stop >= 0
-                digits = num[to_digits:stop]
-            if upper:
-                digits = self._upcase_string(digits)
-            out.append(digits)
-        if spec.n_decimal:
-            out.append(self._lit(".")[0])
-        if spec.n_remainder:
-            out.append(num[to_remainder:])
-        if spec.n_rpadding:
-            out.append_multiple_char(fill_char[0], spec.n_rpadding)
-        #if complex, need to call twice - just retun the buffer
-        return out.build()
-
-    def _format_int_or_long(self, w_num, kind):
-        space = self.space
-        if self._precision != -1:
-            msg = "precision not allowed in integer type"
-            raise OperationError(space.w_ValueError, space.wrap(msg))
-        sign_char = "\0"
-        tp = self._type
-        if tp == "c":
-            if self._sign != "\0":
-                msg = "sign not allowed with 'c' presentation type"
-                raise OperationError(space.w_ValueError, space.wrap(msg))
-            value = space.int_w(w_num)
-            if self.is_unicode:
-                result = runicode.UNICHR(value)
-            else:
-                result = chr(value)
-            n_digits = 1
-            n_remainder = 1
-            to_remainder = 0
-            n_prefix = 0
-            to_prefix = 0
-            to_numeric = 0
-        else:
-            if tp == "b":
-                base = 2
-                skip_leading = 2
-            elif tp == "o":
-                base = 8
-                skip_leading = 2
-            elif tp == "x" or tp == "X":
-                base = 16
-                skip_leading = 2
-            elif tp == "n" or tp == "d":
-                base = 10
-                skip_leading = 0
-            else:
-                raise AssertionError("shouldn't reach")
-            if kind == INT_KIND:
-                result = self._int_to_base(base, space.int_w(w_num))
-            else:
-                result = self._long_to_base(base, space.bigint_w(w_num))
-            n_prefix = skip_leading if self._alternate else 0
-            to_prefix = 0
-            if result[0] == "-":
-                sign_char = "-"
-                skip_leading += 1
-                to_prefix += 1
-            n_digits = len(result) - skip_leading
-            n_remainder = 0
-            to_remainder = 0
-            to_numeric = skip_leading
-        self._get_locale(tp)
-        spec = self._calc_num_width(n_prefix, sign_char, to_numeric, n_digits,
-                                    n_remainder, False, result)
-        fill = self._lit(" ") if self._fill_char == "\0" else self._fill_char
-        upper = self._type == "X"
-        return self.space.wrap(self._fill_number(spec, result, to_numeric,
-                                 to_prefix, fill, to_remainder, upper))
-
-    def _long_to_base(self, base, value):
-        prefix = ""
-        if base == 2:
-            prefix = "0b"
-        elif base == 8:
-            prefix = "0o"
-        elif base == 16:
-            prefix = "0x"
-        as_str = value.format(LONG_DIGITS[:base], prefix)
-        if self.is_unicode:
-            return as_str.decode("ascii")
-        return as_str
-
-    def _int_to_base(self, base, value):
-        if base == 10:
-            s = str(value)
+        def _lit(self, s):
             if self.is_unicode:
                 return s.decode("ascii")
-            return s
-        # This part is slow.
-        negative = value < 0
-        value = abs(value)
-        buf = ["\0"] * (8 * 8 + 6) # Too much on 32 bit, but who cares?
-        i = len(buf) - 1
-        while True:
-            div = value // base
-            mod = value - div * base
-            digit = abs(mod)
-            digit += ord("0") if digit < 10 else ord("a") - 10
-            buf[i] = chr(digit)
-            value = div
+            else:
+                return s
+
+        def _pad(self, string):
+            builder = self._builder()
+            builder.append_multiple_char(self._fill_char, self._left_pad)
+            builder.append(string)
+            builder.append_multiple_char(self._fill_char, self._right_pad)
+            return builder.build()
+
+        def _builder(self):
+            if self.is_unicode:
+                return rstring.UnicodeBuilder()
+            else:
+                return rstring.StringBuilder()
+
+        def _unknown_presentation(self, tp):
+            msg = "unknown presentation for %s: '%s'"
+            w_msg = self.space.wrap(msg  % (tp, self._type))
+            raise OperationError(self.space.w_ValueError, w_msg)
+
+        def format_string(self, string):
+            space = self.space
+            if self._parse_spec("s", "<"):
+                return space.wrap(string)
+            if self._type != "s":
+                self._unknown_presentation("string")
+            if self._sign != "\0":
+                msg = "Sign not allowed in string format specifier"
+                raise OperationError(space.w_ValueError, space.wrap(msg))
+            if self._alternate:
+                msg = "Alternate form not allowed in string format specifier"
+                raise OperationError(space.w_ValueError, space.wrap(msg))
+            if self._align == "=":
+                msg = "'=' alignment not allowed in string format specifier"
+                raise OperationError(space.w_ValueError, space.wrap(msg))
+            length = len(string)
+            precision = self._precision
+            if precision != -1 and length >= precision:
+                assert precision >= 0
+                length = precision
+                string = string[:precision]
+            if self._fill_char == "\0":
+                self._fill_char = self._lit(" ")[0]
+            self._calc_padding(string, length)
+            return space.wrap(self._pad(string))
+
+        def _get_locale(self, tp):
+            space = self.space
+            if tp == "n":
+                dec, thousands, grouping = rlocale.numeric_formatting()
+            elif self._thousands_sep:
+                dec = "."
+                thousands = ","
+                grouping = "\3\0"
+            else:
+                dec = "."
+                thousands = ""
+                grouping = "\256"
+            if self.is_unicode:
+                self._loc_dec = dec.decode("ascii")
+                self._loc_thousands = thousands.decode("ascii")
+            else:
+                self._loc_dec = dec
+                self._loc_thousands = thousands
+            self._loc_grouping = grouping
+
+        def _calc_num_width(self, n_prefix, sign_char, to_number, n_number,
+                            n_remainder, has_dec, digits):
+            """Calculate widths of all parts of formatted number.
+
+            Output will look like:
+
+                <lpadding> <sign> <prefix> <spadding> <grouped_digits> <decimal>
+                <remainder> <rpadding>
+
+            sign is computed from self._sign, and the sign of the number
+            prefix is given
+            digits is known
+            """
+            spec = NumberSpec()
+            spec.n_digits = n_number - n_remainder - has_dec
+            spec.n_prefix = n_prefix
+            spec.n_lpadding = 0
+            spec.n_decimal = int(has_dec)
+            spec.n_remainder = n_remainder
+            spec.n_spadding = 0
+            spec.n_rpadding = 0
+            spec.n_min_width = 0
+            spec.n_total = 0
+            spec.sign = "\0"
+            spec.n_sign = 0
+            sign = self._sign
+            if sign == "+":
+                spec.n_sign = 1
+                spec.sign = "-" if sign_char == "-" else "+"
+            elif sign == " ":
+                spec.n_sign = 1
+                spec.sign = "-" if sign_char == "-" else " "
+            elif sign_char == "-":
+                spec.n_sign = 1
+                spec.sign = "-"
+            extra_length = (spec.n_sign + spec.n_prefix + spec.n_decimal +
+                            spec.n_remainder) # Not padding or digits
+            if self._fill_char == "0" and self._align == "=":
+                spec.n_min_width = self._width - extra_length
+            if self._loc_thousands:
+                self._group_digits(spec, digits[to_number:])
+                n_grouped_digits = len(self._grouped_digits)
+            else:
+                n_grouped_digits = spec.n_digits
+            n_padding = self._width - (extra_length + n_grouped_digits)
+            if n_padding > 0:
+                align = self._align
+                if align == "<":
+                    spec.n_rpadding = n_padding
+                elif align == ">":
+                    spec.n_lpadding = n_padding
+                elif align == "^":
+                    spec.n_lpadding = n_padding // 2
+                    spec.n_rpadding = n_padding - spec.n_lpadding
+                elif align == "=":
+                    spec.n_spadding = n_padding
+                else:
+                    raise AssertionError("shouldn't reach")
+            spec.n_total = spec.n_lpadding + spec.n_sign + spec.n_prefix + \
+                           spec.n_spadding + n_grouped_digits + \
+                           spec.n_decimal + spec.n_remainder + spec.n_rpadding
+            return spec
+
+        def _fill_digits(self, buf, digits, d_state, n_chars, n_zeros,
+                         thousands_sep):
+            if thousands_sep:
+                for c in thousands_sep:
+                    buf.append(c)
+            for i in range(d_state - 1, d_state - n_chars - 1, -1):
+                buf.append(digits[i])
+            for i in range(n_zeros):
+                buf.append("0")
+
+        def _group_digits(self, spec, digits):
+            buf = []
+            grouping = self._loc_grouping
+            min_width = spec.n_min_width
+            grouping_state = 0
+            count = 0
+            left = spec.n_digits
+            n_ts = len(self._loc_thousands)
+            need_separator = False
+            done = False
+            groupings = len(grouping)
+            previous = 0
+            while True:
+                group = ord(grouping[grouping_state])
+                if group > 0:
+                    if group == 256:
+                        break
+                    grouping_state += 1
+                    previous = group
+                else:
+                    group = previous
+                final_grouping = min(group, max(left, max(min_width, 1)))
+                n_zeros = max(0, final_grouping - left)
+                n_chars = max(0, min(left, final_grouping))
+                ts = self._loc_thousands if need_separator else None
+                self._fill_digits(buf, digits, left, n_chars, n_zeros, ts)
+                need_separator = True
+                left -= n_chars
+                min_width -= final_grouping
+                if left <= 0 and min_width <= 0:
+                    done = True
+                    break
+                min_width -= n_ts
+            if not done:
+                group = max(max(left, min_width), 1)
+                n_zeros = max(0, group - left)
+                n_chars = max(0, min(left, group))
+                ts = self._loc_thousands if need_separator else None
+                self._fill_digits(buf, digits, left, n_chars, n_zeros, ts)
+            buf.reverse()
+            self._grouped_digits = self.empty.join(buf)
+
+        def _upcase_string(self, s):
+            buf = []
+            for c in s:
+                index = ord(c)
+                if ord("a") <= index <= ord("z"):
+                    c = chr(index - 32)
+                buf.append(c)
+            return self.empty.join(buf)
+
+
+        def _fill_number(self, spec, num, to_digits, to_prefix, fill_char,
+                         to_remainder, upper, grouped_digits=None):
+            out = self._builder()
+            if spec.n_lpadding:
+                out.append_multiple_char(fill_char[0], spec.n_lpadding)
+            if spec.n_sign:
+                if self.is_unicode:
+                    sign = spec.sign.decode("ascii")
+                else:
+                    sign = spec.sign
+                out.append(sign)
+            if spec.n_prefix:
+                pref = num[to_prefix:to_prefix + spec.n_prefix]
+                if upper:
+                    pref = self._upcase_string(pref)
+                out.append(pref)
+            if spec.n_spadding:
+                out.append_multiple_char(fill_char[0], spec.n_spadding)
+            if spec.n_digits != 0:
+                if self._loc_thousands:
+                    if grouped_digits is not None:
+                        digits = grouped_digits
+                    else:
+                        digits = self._grouped_digits
+                        assert digits is not None
+                else:
+                    stop = to_digits + spec.n_digits
+                    assert stop >= 0
+                    digits = num[to_digits:stop]
+                if upper:
+                    digits = self._upcase_string(digits)
+                out.append(digits)
+            if spec.n_decimal:
+                out.append(self._lit(".")[0])
+            if spec.n_remainder:
+                out.append(num[to_remainder:])
+            if spec.n_rpadding:
+                out.append_multiple_char(fill_char[0], spec.n_rpadding)
+            #if complex, need to call twice - just retun the buffer
+            return out.build()
+
+        def _format_int_or_long(self, w_num, kind):
+            space = self.space
+            if self._precision != -1:
+                msg = "precision not allowed in integer type"
+                raise OperationError(space.w_ValueError, space.wrap(msg))
+            sign_char = "\0"
+            tp = self._type
+            if tp == "c":
+                if self._sign != "\0":
+                    msg = "sign not allowed with 'c' presentation type"
+                    raise OperationError(space.w_ValueError, space.wrap(msg))
+                value = space.int_w(w_num)
+                if self.is_unicode:
+                    result = runicode.UNICHR(value)
+                else:
+                    result = chr(value)
+                n_digits = 1
+                n_remainder = 1
+                to_remainder = 0
+                n_prefix = 0
+                to_prefix = 0
+                to_numeric = 0
+            else:
+                if tp == "b":
+                    base = 2
+                    skip_leading = 2
+                elif tp == "o":
+                    base = 8
+                    skip_leading = 2
+                elif tp == "x" or tp == "X":
+                    base = 16
+                    skip_leading = 2
+                elif tp == "n" or tp == "d":
+                    base = 10
+                    skip_leading = 0
+                else:
+                    raise AssertionError("shouldn't reach")
+                if kind == INT_KIND:
+                    result = self._int_to_base(base, space.int_w(w_num))
+                else:
+                    result = self._long_to_base(base, space.bigint_w(w_num))
+                n_prefix = skip_leading if self._alternate else 0
+                to_prefix = 0
+                if result[0] == "-":
+                    sign_char = "-"
+                    skip_leading += 1
+                    to_prefix += 1
+                n_digits = len(result) - skip_leading
+                n_remainder = 0
+                to_remainder = 0
+                to_numeric = skip_leading
+            self._get_locale(tp)
+            spec = self._calc_num_width(n_prefix, sign_char, to_numeric, n_digits,
+                                        n_remainder, False, result)
+            fill = self._lit(" ") if self._fill_char == "\0" else self._fill_char
+            upper = self._type == "X"
+            return self.space.wrap(self._fill_number(spec, result, to_numeric,
+                                     to_prefix, fill, to_remainder, upper))
+
+        def _long_to_base(self, base, value):
+            prefix = ""
+            if base == 2:
+                prefix = "0b"
+            elif base == 8:
+                prefix = "0o"
+            elif base == 16:
+                prefix = "0x"
+            as_str = value.format(LONG_DIGITS[:base], prefix)
+            if self.is_unicode:
+                return as_str.decode("ascii")
+            return as_str
+
+        def _int_to_base(self, base, value):
+            if base == 10:
+                s = str(value)
+                if self.is_unicode:
+                    return s.decode("ascii")
+                return s
+            # This part is slow.
+            negative = value < 0
+            value = abs(value)
+            buf = ["\0"] * (8 * 8 + 6) # Too much on 32 bit, but who cares?
+            i = len(buf) - 1
+            while True:
+                div = value // base
+                mod = value - div * base
+                digit = abs(mod)
+                digit += ord("0") if digit < 10 else ord("a") - 10
+                buf[i] = chr(digit)
+                value = div
+                i -= 1
+                if not value:
+                    break
+            if base == 2:
+                buf[i] = "b"
+                buf[i - 1] = "0"
+            elif base == 8:
+                buf[i] = "o"
+                buf[i - 1] = "0"
+            elif base == 16:
+                buf[i] = "x"
+                buf[i - 1] = "0"
+            else:
+                buf[i] = "#"
+                buf[i - 1] = chr(ord("0") + base % 10)
+                if base > 10:
+                    buf[i - 2] = chr(ord("0") + base // 10)
+                    i -= 1
             i -= 1
-            if not value:
-                break
-        if base == 2:
-            buf[i] = "b"
-            buf[i - 1] = "0"
-        elif base == 8:
-            buf[i] = "o"
-            buf[i - 1] = "0"
-        elif base == 16:
-            buf[i] = "x"
-            buf[i - 1] = "0"
-        else:
-            buf[i] = "#"
-            buf[i - 1] = chr(ord("0") + base % 10)
-            if base > 10:
-                buf[i - 2] = chr(ord("0") + base // 10)
+            if negative:
                 i -= 1
-        i -= 1
-        if negative:
-            i -= 1
-            buf[i] = "-"
-        assert i >= 0
-        return self.empty.join(buf[i:])
+                buf[i] = "-"
+            assert i >= 0
+            return self.empty.join(buf[i:])
 
-    def format_int_or_long(self, w_num, kind):
-        space = self.space
-        if self._parse_spec("d", ">"):
+        def format_int_or_long(self, w_num, kind):
+            space = self.space
+            if self._parse_spec("d", ">"):
+                if self.is_unicode:
+                    return space.call_function(space.w_unicode, w_num)
+                return self.space.str(w_num)
+            tp = self._type
+            if (tp == "b" or
+                tp == "c" or
+                tp == "d" or
+                tp == "o" or
+                tp == "x" or
+                tp == "X" or
+                tp == "n"):
+                return self._format_int_or_long(w_num, kind)
+            elif (tp == "e" or
+                  tp == "E" or
+                  tp == "f" or
+                  tp == "F" or
+                  tp == "g" or
+                  tp == "G" or
+                  tp == "%"):
+                w_float = space.float(w_num)
+                return self._format_float(w_float)
+            else:
+                self._unknown_presentation("int" if kind == INT_KIND else "long")
+
+        def _parse_number(self, s, i):
+            """Determine if s has a decimal point, and the index of the first #
+            after the decimal, or the end of the number."""
+            length = len(s)
+            while i < length and "0" <= s[i] <= "9":
+                i += 1
+            rest = i
+            dec_point = i < length and s[i] == "."
+            if dec_point:
+                rest += 1
+            #differs from CPython method - CPython sets n_remainder
+            return dec_point, rest
+
+        def _format_float(self, w_float):
+            """helper for format_float"""
+            space = self.space
+            flags = 0
+            default_precision = 6
+            if self._alternate:
+                msg = "alternate form not allowed in float formats"
+                raise OperationError(space.w_ValueError, space.wrap(msg))
+            tp = self._type
+            self._get_locale(tp)
+            if tp == "\0":
+                tp = "g"
+                default_precision = 12
+                flags |= rfloat.DTSF_ADD_DOT_0
+            elif tp == "n":
+                tp = "g"
+            value = space.float_w(w_float)
+            if tp == "%":
+                tp = "f"
+                value *= 100
+                add_pct = True
+            else:
+                add_pct = False
+            if self._precision == -1:
+                self._precision = default_precision
+            result, special = rfloat.double_to_string(value, tp,
+                                                      self._precision, flags)
+            if add_pct:
+                result += "%"
+            n_digits = len(result)
+            if result[0] == "-":
+                sign = "-"
+                to_number = 1
+                n_digits -= 1
+            else:
+                sign = "\0"
+                to_number = 0
+            have_dec_point, to_remainder = self._parse_number(result, to_number)
+            n_remainder = len(result) - to_remainder
             if self.is_unicode:
-                return space.call_function(space.w_unicode, w_num)
-            return self.space.str(w_num)
-        tp = self._type
-        if (tp == "b" or
-            tp == "c" or
-            tp == "d" or
-            tp == "o" or
-            tp == "x" or
-            tp == "X" or
-            tp == "n"):
-            return self._format_int_or_long(w_num, kind)
-        elif (tp == "e" or
-              tp == "E" or
-              tp == "f" or
-              tp == "F" or
-              tp == "g" or
-              tp == "G" or
-              tp == "%"):
-            w_float = space.float(w_num)
-            return self._format_float(w_float)
-        else:
-            self._unknown_presentation("int" if kind == INT_KIND else "long")
+                digits = result.decode("ascii")
+            else:
+                digits = result
+            spec = self._calc_num_width(0, sign, to_number, n_digits,
+                                        n_remainder, have_dec_point, digits)
+            fill = self._lit(" ") if self._fill_char == "\0" else self._fill_char
+            return self.space.wrap(self._fill_number(spec, digits, to_number, 0,
+                                      fill, to_remainder, False))
 
-    def _parse_number(self, s, i):
-        """Determine if s has a decimal point, and the index of the first #
-        after the decimal, or the end of the number."""
-        length = len(s)
-        while i < length and "0" <= s[i] <= "9":
-            i += 1
-        rest = i
-        dec_point = i < length and s[i] == "."
-        if dec_point:
-            rest += 1
-        #differs from CPython method - CPython sets n_remainder
-        return dec_point, rest
+        def format_float(self, w_float):
+            space = self.space
+            if self._parse_spec("\0", ">"):
+                if self.is_unicode:
+                    return space.call_function(space.w_unicode, w_float)
+                return space.str(w_float)
+            tp = self._type
+            if (tp == "\0" or
+                tp == "e" or
+                tp == "E" or
+                tp == "f" or
+                tp == "F" or
+                tp == "g" or
+                tp == "G" or
+                tp == "n" or
+                tp == "%"):
+                return self._format_float(w_float)
+            self._unknown_presentation("float")
 
-    def _format_float(self, w_float):
-        """helper for format_float"""
-        space = self.space
-        flags = 0
-        default_precision = 6
-        if self._alternate:
-            msg = "alternate form not allowed in float formats"
-            raise OperationError(space.w_ValueError, space.wrap(msg))
-        tp = self._type
-        self._get_locale(tp)
-        if tp == "\0":
-            tp = "g"
-            default_precision = 12
-            flags |= rfloat.DTSF_ADD_DOT_0
-        elif tp == "n":
-            tp = "g"
-        value = space.float_w(w_float)
-        if tp == "%":
-            tp = "f"
-            value *= 100
-            add_pct = True
-        else:
-            add_pct = False
-        if self._precision == -1:
-            self._precision = default_precision
-        result, special = rfloat.double_to_string(value, tp,
-                                                  self._precision, flags)
-        if add_pct:
-            result += "%"
-        n_digits = len(result)
-        if result[0] == "-":
-            sign = "-"
-            to_number = 1
-            n_digits -= 1
-        else:
-            sign = "\0"
-            to_number = 0
-        have_dec_point, to_remainder = self._parse_number(result, to_number)
-        n_remainder = len(result) - to_remainder
-        if self.is_unicode:
-            digits = result.decode("ascii")
-        else:
-            digits = result
-        spec = self._calc_num_width(0, sign, to_number, n_digits,
-                                    n_remainder, have_dec_point, digits)
-        fill = self._lit(" ") if self._fill_char == "\0" else self._fill_char
-        return self.space.wrap(self._fill_number(spec, digits, to_number, 0,
-                                  fill, to_remainder, False))
+        def _format_complex(self, w_complex):
+            space = self.space
+            tp = self._type
+            self._get_locale(tp)
+            default_precision = 6
+            if self._align == "=":
+                # '=' alignment is invalid
+                msg = ("'=' alignment flag is not allowed in"
+                       " complex format specifier")
+                raise OperationError(space.w_ValueError, space.wrap(msg))
+            if self._fill_char == "0":
+                #zero padding is invalid
+                msg = "Zero padding is not allowed in complex format specifier"
+                raise OperationError(space.w_ValueError, space.wrap(msg))
+            if self._alternate:
+                #alternate is invalid
+                msg = "Alternate form %s not allowed in complex format specifier"
+                raise OperationError(space.w_ValueError,
+                                     space.wrap(msg % (self._alternate)))
+            skip_re = 0
+            add_parens = 0
+            if tp == "\0":
+                #should mirror str() output
+                tp = "g"
+                default_precision = 12
+                #test if real part is non-zero
+                if (w_complex.realval == 0 and
+                    copysign(1., w_complex.realval) == 1.):
+                    skip_re = 1
+                else:
+                    add_parens = 1
 
-    def format_float(self, w_float):
-        space = self.space
-        if self._parse_spec("\0", ">"):
+            if tp == "n":
+                #same as 'g' except for locale, taken care of later
+                tp = "g"
+
+            #check if precision not set
+            if self._precision == -1:
+                self._precision = default_precision
+
+            #might want to switch to double_to_string from formatd
+            #in CPython it's named 're' - clashes with re module
+            re_num = formatd(w_complex.realval, tp, self._precision)
+            im_num = formatd(w_complex.imagval, tp, self._precision)
+            n_re_digits = len(re_num)
+            n_im_digits = len(im_num)
+
+            to_real_number = 0
+            to_imag_number = 0
+            re_sign = im_sign = ''
+            #if a sign character is in the output, remember it and skip
+            if re_num[0] == "-":
+                re_sign = "-"
+                to_real_number = 1
+                n_re_digits -= 1
+            if im_num[0] == "-":
+                im_sign = "-"
+                to_imag_number = 1
+                n_im_digits -= 1
+
+            #turn off padding - do it after number composition
+            #calc_num_width uses self._width, so assign to temporary variable,
+            #calculate width of real and imag parts, then reassign padding, align
+            tmp_fill_char = self._fill_char
+            tmp_align = self._align
+            tmp_width = self._width
+            self._fill_char = "\0"
+            self._align = "<"
+            self._width = -1
+
+            #determine if we have remainder, might include dec or exponent or both
+            re_have_dec, re_remainder_ptr = self._parse_number(re_num,
+                                                               to_real_number)
+            im_have_dec, im_remainder_ptr = self._parse_number(im_num,
+                                                               to_imag_number)
+
             if self.is_unicode:
-                return space.call_function(space.w_unicode, w_float)
-            return space.str(w_float)
-        tp = self._type
-        if (tp == "\0" or
-            tp == "e" or
-            tp == "E" or
-            tp == "f" or
-            tp == "F" or
-            tp == "g" or
-            tp == "G" or
-            tp == "n" or
-            tp == "%"):
-            return self._format_float(w_float)
-        self._unknown_presentation("float")
+                re_num = re_num.decode("ascii")
+                im_num = im_num.decode("ascii")
 
-    def _format_complex(self, w_complex):
-        space = self.space
-        tp = self._type
-        self._get_locale(tp)
-        default_precision = 6
-        if self._align == "=":
-            # '=' alignment is invalid
-            msg = ("'=' alignment flag is not allowed in"
-                   " complex format specifier")
-            raise OperationError(space.w_ValueError, space.wrap(msg))
-        if self._fill_char == "0":
-            #zero padding is invalid
-            msg = "Zero padding is not allowed in complex format specifier"
-            raise OperationError(space.w_ValueError, space.wrap(msg))
-        if self._alternate:
-            #alternate is invalid
-            msg = "Alternate form %s not allowed in complex format specifier"
-            raise OperationError(space.w_ValueError,
-                                 space.wrap(msg % (self._alternate)))
-        skip_re = 0
-        add_parens = 0
-        if tp == "\0":
-            #should mirror str() output
-            tp = "g"
-            default_precision = 12
-            #test if real part is non-zero
-            if (w_complex.realval == 0 and
-                copysign(1., w_complex.realval) == 1.):
-                skip_re = 1
-            else:
-                add_parens = 1
+            #set remainder, in CPython _parse_number sets this
+            #using n_re_digits causes tests to fail
+            re_n_remainder = len(re_num) - re_remainder_ptr
+            im_n_remainder = len(im_num) - im_remainder_ptr
+            re_spec = self._calc_num_width(0, re_sign, to_real_number, n_re_digits,
+                                           re_n_remainder, re_have_dec,
+                                           re_num)
 
-        if tp == "n":
-            #same as 'g' except for locale, taken care of later
-            tp = "g"
+            #capture grouped digits b/c _fill_number reads from self._grouped_digits
+            #self._grouped_digits will get overwritten in imaginary calc_num_width
+            re_grouped_digits = self._grouped_digits
+            if not skip_re:
+                self._sign = "+"
+            im_spec = self._calc_num_width(0, im_sign, to_imag_number, n_im_digits,
+                                           im_n_remainder, im_have_dec,
+                                           im_num)
 
-        #check if precision not set
-        if self._precision == -1:
-            self._precision = default_precision
+            im_grouped_digits = self._grouped_digits
+            if skip_re:
+                re_spec.n_total = 0
 
-        #might want to switch to double_to_string from formatd
-        #in CPython it's named 're' - clashes with re module
-        re_num = formatd(w_complex.realval, tp, self._precision)
-        im_num = formatd(w_complex.imagval, tp, self._precision)
-        n_re_digits = len(re_num)
-        n_im_digits = len(im_num)
+            #reassign width, alignment, fill character
+            self._align = tmp_align
+            self._width = tmp_width
+            self._fill_char = tmp_fill_char
 
-        to_real_number = 0
-        to_imag_number = 0
-        re_sign = im_sign = ''
-        #if a sign character is in the output, remember it and skip
-        if re_num[0] == "-":
-            re_sign = "-"
-            to_real_number = 1
-            n_re_digits -= 1
-        if im_num[0] == "-":
-            im_sign = "-"
-            to_imag_number = 1
-            n_im_digits -= 1
+            #compute L and R padding - stored in self._left_pad and self._right_pad
+            self._calc_padding(self.empty, re_spec.n_total + im_spec.n_total + 1 +
+                                           add_parens * 2)
 
-        #turn off padding - do it after number composition
-        #calc_num_width uses self._width, so assign to temporary variable,
-        #calculate width of real and imag parts, then reassign padding, align
-        tmp_fill_char = self._fill_char
-        tmp_align = self._align
-        tmp_width = self._width
-        self._fill_char = "\0"
-        self._align = "<"
-        self._width = -1
+            out = self._builder()
+            fill = self._fill_char
+            if fill == "\0":
+                fill = self._lit(" ")[0]
 
-        #determine if we have remainder, might include dec or exponent or both
-        re_have_dec, re_remainder_ptr = self._parse_number(re_num,
-                                                           to_real_number)
-        im_have_dec, im_remainder_ptr = self._parse_number(im_num,
-                                                           to_imag_number)
+            #compose the string
+            #add left padding
+            out.append_multiple_char(fill, self._left_pad)
+            if add_parens:
+                out.append(self._lit('(')[0])
 
-        if self.is_unicode:
-            re_num = re_num.decode("ascii")
-            im_num = im_num.decode("ascii")
+            #if the no. has a real component, add it
+            if not skip_re:
+                out.append(self._fill_number(re_spec, re_num, to_real_number, 0,
+                                             fill, re_remainder_ptr, False,
+                                             re_grouped_digits))
 
-        #set remainder, in CPython _parse_number sets this
-        #using n_re_digits causes tests to fail
-        re_n_remainder = len(re_num) - re_remainder_ptr
-        im_n_remainder = len(im_num) - im_remainder_ptr
-        re_spec = self._calc_num_width(0, re_sign, to_real_number, n_re_digits,
-                                       re_n_remainder, re_have_dec,
-                                       re_num)
+            #add imaginary component
+            out.append(self._fill_number(im_spec, im_num, to_imag_number, 0,
+                                         fill, im_remainder_ptr, False,
+                                         im_grouped_digits))
 
-        #capture grouped digits b/c _fill_number reads from self._grouped_digits
-        #self._grouped_digits will get overwritten in imaginary calc_num_width
-        re_grouped_digits = self._grouped_digits
-        if not skip_re:
-            self._sign = "+"
-        im_spec = self._calc_num_width(0, im_sign, to_imag_number, n_im_digits,
-                                       im_n_remainder, im_have_dec,
-                                       im_num)
+            #add 'j' character
+            out.append(self._lit('j')[0])
 
-        im_grouped_digits = self._grouped_digits
-        if skip_re:
-            re_spec.n_total = 0
+            if add_parens:
+                out.append(self._lit(')')[0])
 
-        #reassign width, alignment, fill character
-        self._align = tmp_align
-        self._width = tmp_width
-        self._fill_char = tmp_fill_char
+            #add right padding
+            out.append_multiple_char(fill, self._right_pad)
 
-        #compute L and R padding - stored in self._left_pad and self._right_pad
-        self._calc_padding(self.empty, re_spec.n_total + im_spec.n_total + 1 +
-                                       add_parens * 2)
+            return self.space.wrap(out.build())
 
-        out = self._builder()
-        fill = self._fill_char
-        if fill == "\0":
-            fill = self._lit(" ")[0]
 
-        #compose the string
-        #add left padding
-        out.append_multiple_char(fill, self._left_pad)
-        if add_parens:
-            out.append(self._lit('(')[0])
+        def format_complex(self, w_complex):
+            """return the string representation of a complex number"""
+            space = self.space
+            #parse format specification, set associated variables
+            if self._parse_spec("\0", ">"):
+                return space.str(w_complex)
+            tp = self._type
+            if (tp == "\0" or
+                tp == "e" or
+                tp == "E" or
+                tp == "f" or
+                tp == "F" or
+                tp == "g" or
+                tp == "G" or
+                tp == "n"):
+                return self._format_complex(w_complex)
+            self._unknown_presentation("complex")
+    return Formatter
 
-        #if the no. has a real component, add it
-        if not skip_re:
-            out.append(self._fill_number(re_spec, re_num, to_real_number, 0,
-                                         fill, re_remainder_ptr, False,
-                                         re_grouped_digits))
-
-        #add imaginary component
-        out.append(self._fill_number(im_spec, im_num, to_imag_number, 0,
-                                     fill, im_remainder_ptr, False,
-                                     im_grouped_digits))
-
-        #add 'j' character
-        out.append(self._lit('j')[0])
-
-        if add_parens:
-            out.append(self._lit(')')[0])
-
-        #add right padding
-        out.append_multiple_char(fill, self._right_pad)
-
-        return self.space.wrap(out.build())
-
-
-    def format_complex(self, w_complex):
-        """return the string representation of a complex number"""
-        space = self.space
-        #parse format specification, set associated variables
-        if self._parse_spec("\0", ">"):
-            return space.str(w_complex)
-        tp = self._type
-        if (tp == "\0" or
-            tp == "e" or
-            tp == "E" or
-            tp == "f" or
-            tp == "F" or
-            tp == "g" or
-            tp == "G" or
-            tp == "n"):
-            return self._format_complex(w_complex)
-        self._unknown_presentation("complex")
+StrFormatter = make_formatting_class()
+UnicodeFormatter = make_formatting_class()
 
 
 def unicode_formatter(space, spec):
-    return Formatter(space, True, spec)
-
+    return StrFormatter(space, True, spec)
 
 def str_formatter(space, spec):
-    return Formatter(space, False, spec)
+    return UnicodeFormatter(space, False, spec)
 
 
 @specialize.arg(2)
diff --git a/pypy/rlib/jit.py b/pypy/rlib/jit.py
--- a/pypy/rlib/jit.py
+++ b/pypy/rlib/jit.py
@@ -1,10 +1,13 @@
+import sys
+
 import py
-import sys
+
+from pypy.rlib.nonconst import NonConstant
+from pypy.rlib.objectmodel import CDefinedIntSymbolic, keepalive_until_here, specialize
+from pypy.rlib.unroll import unrolling_iterable
 from pypy.rpython.extregistry import ExtRegistryEntry
-from pypy.rlib.objectmodel import CDefinedIntSymbolic
-from pypy.rlib.objectmodel import keepalive_until_here, specialize
-from pypy.rlib.unroll import unrolling_iterable
-from pypy.rlib.nonconst import NonConstant
+from pypy.tool.sourcetools import func_with_new_name
+
 
 def elidable(func):
     """ Decorate a function as "trace-elidable". This means precisely that:
@@ -72,17 +75,22 @@
     func._jit_loop_invariant_ = True
     return func
 
+def _get_args(func):
+    import inspect
+
+    args, varargs, varkw, defaults = inspect.getargspec(func)
+    args = ["v%s" % (i, ) for i in range(len(args))]
+    assert varargs is None and varkw is None
+    assert not defaults
+    return args
+
 def elidable_promote(promote_args='all'):
     """ A decorator that promotes all arguments and then calls the supplied
     function
     """
     def decorator(func):
-        import inspect
         elidable(func)
-        args, varargs, varkw, defaults = inspect.getargspec(func)
-        args = ["v%s" % (i, ) for i in range(len(args))]
-        assert varargs is None and varkw is None
-        assert not defaults
+        args = _get_args(func)
         argstring = ", ".join(args)
         code = ["def f(%s):\n" % (argstring, )]
         if promote_args != 'all':
@@ -102,6 +110,46 @@
     warnings.warn("purefunction_promote is deprecated, use elidable_promote instead", DeprecationWarning)
     return elidable_promote(*args, **kwargs)
 
+def look_inside_iff(predicate):
+    """
+    look inside (including unrolling loops) the target function, if and only if
+    predicate(*args) returns True
+    """
+    def inner(func):
+        func = unroll_safe(func)
+        # When we return the new function, it might be specialized in some
+        # way. We "propogate" this specialization by using
+        # specialize:call_location on relevant functions.
+        for thing in [func, predicate]:
+            thing._annspecialcase_ = "specialize:call_location"
+
+        args = _get_args(func)
+        d = {
+            "dont_look_inside": dont_look_inside,
+            "predicate": predicate,
+            "func": func,
+            "we_are_jitted": we_are_jitted,
+        }
+        exec py.code.Source("""
+            @dont_look_inside
+            def trampoline(%(arguments)s):
+                return func(%(arguments)s)
+            if hasattr(func, "oopspec"):
+                # XXX: This seems like it should be here, but it causes errors.
+                # trampoline.oopspec = func.oopspec
+                del func.oopspec
+            trampoline.__name__ = func.__name__ + "_trampoline"
+            trampoline._annspecialcase_ = "specialize:call_location"
+
+            def f(%(arguments)s):
+                if not we_are_jitted() or predicate(%(arguments)s):
+                    return func(%(arguments)s)
+                else:
+                    return trampoline(%(arguments)s)
+            f.__name__ = func.__name__ + "_look_inside_iff"
+        """ % {"arguments": ", ".join(args)}).compile() in d
+        return d["f"]
+    return inner
 
 def oopspec(spec):
     def decorator(func):
@@ -109,6 +157,34 @@
         return func
     return decorator
 
+ at oopspec("jit.isconstant(value)")
+ at specialize.argtype(0)
+def isconstant(value):
+    """
+    While tracing, returns whether or not the value is currently known to be
+    constant. This is not perfect, values can become constant later. Mostly for
+    use with @look_inside_iff.
+
+    This is for advanced usage only.
+    """
+    # I hate the annotator so much.
+    if NonConstant(False):
+        return True
+    return False
+
+ at oopspec("jit.isvirtual(value)")
+ at specialize.ll()
+def isvirtual(value):
+    """
+    Returns if this value is virtual, while tracing, it's relatively
+    conservative and will miss some cases.
+
+    This is for advanced usage only.
+    """
+    if NonConstant(False):
+        return True
+    return False
+
 class Entry(ExtRegistryEntry):
     _about_ = hint
 
diff --git a/pypy/rlib/rgc.py b/pypy/rlib/rgc.py
--- a/pypy/rlib/rgc.py
+++ b/pypy/rlib/rgc.py
@@ -1,6 +1,9 @@
-import gc, types
+import gc
+import types
+
+from pypy.rlib import jit
+from pypy.rlib.objectmodel import we_are_translated, enforceargs, specialize
 from pypy.rpython.extregistry import ExtRegistryEntry
-from pypy.rlib.objectmodel import we_are_translated
 from pypy.rpython.lltypesystem import lltype, llmemory
 
 # ____________________________________________________________
@@ -32,7 +35,7 @@
         if len(hop.args_s) == 1:
             args_v = hop.inputargs(lltype.Signed)
         return hop.genop('gc__collect', args_v, resulttype=hop.r_result)
-    
+
 class SetMaxHeapSizeEntry(ExtRegistryEntry):
     _about_ = set_max_heap_size
 
@@ -133,6 +136,9 @@
         hop.exception_cannot_occur()
         return hop.genop(opname, vlist, resulttype = hop.r_result.lowleveltype)
 
+ at jit.oopspec('list.ll_arraycopy(source, dest, source_start, dest_start, length)')
+ at specialize.ll()
+ at enforceargs(None, None, int, int, int)
 def ll_arraycopy(source, dest, source_start, dest_start, length):
     from pypy.rpython.lltypesystem.lloperation import llop
     from pypy.rlib.objectmodel import keepalive_until_here
@@ -161,14 +167,11 @@
                       llmemory.sizeof(TP.OF) * source_start)
     cp_dest_addr = (dest_addr + llmemory.itemoffsetof(TP, 0) +
                     llmemory.sizeof(TP.OF) * dest_start)
-    
+
     llmemory.raw_memcopy(cp_source_addr, cp_dest_addr,
                          llmemory.sizeof(TP.OF) * length)
     keepalive_until_here(source)
     keepalive_until_here(dest)
-ll_arraycopy._annenforceargs_ = [None, None, int, int, int]
-ll_arraycopy._annspecialcase_ = 'specialize:ll'
-ll_arraycopy.oopspec = 'list.ll_arraycopy(source, dest, source_start, dest_start, length)'
 
 def ll_shrink_array(p, smallerlength):
     from pypy.rpython.lltypesystem.lloperation import llop
@@ -192,7 +195,7 @@
               llmemory.itemoffsetof(ARRAY, 0))
     source_addr = llmemory.cast_ptr_to_adr(p)    + offset
     dest_addr   = llmemory.cast_ptr_to_adr(newp) + offset
-    llmemory.raw_memcopy(source_addr, dest_addr, 
+    llmemory.raw_memcopy(source_addr, dest_addr,
                          llmemory.sizeof(ARRAY.OF) * smallerlength)
 
     keepalive_until_here(p)
diff --git a/pypy/rlib/rstruct/formatiterator.py b/pypy/rlib/rstruct/formatiterator.py
--- a/pypy/rlib/rstruct/formatiterator.py
+++ b/pypy/rlib/rstruct/formatiterator.py
@@ -1,10 +1,10 @@
-
-from pypy.rlib.rstruct.nativefmttable import native_is_bigendian
-from pypy.rlib.unroll import unrolling_iterable
+from pypy.rlib import jit
 from pypy.rlib.rarithmetic import ovfcheck
 from pypy.rlib.rstruct.error import StructError
+from pypy.rlib.rstruct.nativefmttable import native_is_bigendian, native_fmttable
 from pypy.rlib.rstruct.standardfmttable import standard_fmttable
-from pypy.rlib.rstruct.nativefmttable import native_fmttable
+from pypy.rlib.unroll import unrolling_iterable
+
 
 class FormatIterator(object):
     """
@@ -16,6 +16,7 @@
     _mixin_ = True
     _operate_is_specialized_ = False
 
+    @jit.look_inside_iff(lambda self, fmt: jit.isconstant(fmt))
     def interpret(self, fmt):
         # decode the byte order, size and alignment based on the 1st char
         table = unroll_native_fmtdescs
diff --git a/pypy/rpython/lltypesystem/ll_str.py b/pypy/rpython/lltypesystem/ll_str.py
--- a/pypy/rpython/lltypesystem/ll_str.py
+++ b/pypy/rpython/lltypesystem/ll_str.py
@@ -16,34 +16,31 @@
         return r_uint(i)
 
 @jit.elidable
-def ll_int2dec(i):
+def ll_int2dec(val):
     from pypy.rpython.lltypesystem.rstr import mallocstr
-    temp = malloc(CHAR_ARRAY, 20)
+
+    sign = int(val < 0)
+    if sign:
+        val = ll_unsigned(-val)
+    else:
+        val = ll_unsigned(val)
     len = 0
-    sign = 0
-    if i < 0:
-        sign = 1
-        i = ll_unsigned(-i)
-    else:
-        i = ll_unsigned(i)
-    if i == 0:
-        len = 1
-        temp[0] = '0'
-    else:
-        while i:
-            temp[len] = chr(i%10+ord('0'))
-            i //= 10
-            len += 1
-    len += sign
-    result = mallocstr(len)
-    result.hash = 0
+    i = val
+    while i:
+        len += 1
+        i //= 10
+
+    total_len = sign + len + int(val == 0)
+    result = mallocstr(total_len)
     if sign:
         result.chars[0] = '-'
-        j = 1
-    else:
-        j = 0
+    elif val == 0:
+        result.chars[0] = '0'
+
+    j = 0
     while j < len:
-        result.chars[j] = temp[len-j-1]
+        result.chars[total_len - j - 1] = chr(val % 10 + ord('0'))
+        val //= 10
         j += 1
     return result
 
diff --git a/pypy/rpython/lltypesystem/rbuilder.py b/pypy/rpython/lltypesystem/rbuilder.py
--- a/pypy/rpython/lltypesystem/rbuilder.py
+++ b/pypy/rpython/lltypesystem/rbuilder.py
@@ -1,4 +1,4 @@
-from pypy.rlib import rgc
+from pypy.rlib import rgc, jit
 from pypy.rlib.objectmodel import enforceargs
 from pypy.rlib.rarithmetic import ovfcheck
 from pypy.rpython.annlowlevel import llstr
@@ -95,6 +95,7 @@
         ll_builder.used = needed + used
 
     @staticmethod
+    @jit.look_inside_iff(lambda ll_builder, char, times: jit.isconstant(times) and times <= 4)
     def ll_append_multiple_char(ll_builder, char, times):
         used = ll_builder.used
         if times + used > ll_builder.allocated:
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
@@ -9,7 +9,7 @@
      GcStruct, Void, Signed, malloc, typeOf, nullptr, typeMethod
 from pypy.rpython.lltypesystem import rstr
 from pypy.rlib.debug import ll_assert
-from pypy.rlib import rgc
+from pypy.rlib import rgc, jit
 
 # ____________________________________________________________
 #
@@ -225,20 +225,22 @@
     else:
         _ll_list_resize_really(l, newsize)
 
+ at jit.look_inside_iff(lambda l, newsize: jit.isconstant(len(l.items)) and jit.isconstant(newsize))
+ at jit.oopspec("list._resize_ge(l, newsize)")
 def _ll_list_resize_ge(l, newsize):
     if len(l.items) >= newsize:
         l.length = newsize
     else:
         _ll_list_resize_really(l, newsize)
-_ll_list_resize_ge.oopspec = 'list._resize_ge(l, newsize)'
 
+ at jit.look_inside_iff(lambda l, newsize: jit.isconstant(len(l.items)) and jit.isconstant(newsize))
+ at jit.oopspec("list._resize_le(l, newsize)")
 def _ll_list_resize_le(l, newsize):
     if newsize >= (len(l.items) >> 1) - 5:
         l.length = newsize
     else:
         _ll_list_resize_really(l, newsize)
 
-
 def ll_append_noresize(l, newitem):
     length = l.length
     l.length = length + 1
diff --git a/pypy/rpython/lltypesystem/rpbc.py b/pypy/rpython/lltypesystem/rpbc.py
--- a/pypy/rpython/lltypesystem/rpbc.py
+++ b/pypy/rpython/lltypesystem/rpbc.py
@@ -230,7 +230,8 @@
         args = bk.build_args(opname, hop.args_s[1:])
         s_pbc = hop.args_s[0]   # possibly more precise than self.s_pbc
         descs = list(s_pbc.descriptions)
-        shape, index = description.FunctionDesc.variant_for_call_site(bk, self.callfamily, descs, args)
+        vfcs = description.FunctionDesc.variant_for_call_site
+        shape, index = vfcs(bk, self.callfamily, descs, args, hop.spaceop)
         row_of_graphs = self.callfamily.calltables[shape][index]
         anygraph = row_of_graphs.itervalues().next()  # pick any witness
         vlist = [hop.inputarg(self, arg=0)]
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
@@ -5,7 +5,7 @@
 from pypy.rlib.objectmodel import _hash_string, enforceargs
 from pypy.rlib.objectmodel import keepalive_until_here
 from pypy.rlib.debug import ll_assert
-from pypy.rlib.jit import elidable, we_are_jitted, dont_look_inside
+from pypy.rlib import jit
 from pypy.rlib.rarithmetic import ovfcheck
 from pypy.rpython.robject import PyObjRepr, pyobj_repr
 from pypy.rpython.rmodel import inputconst, IntegerRepr
@@ -58,8 +58,7 @@
                 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
+    @jit.oopspec('stroruni.copy_contents(src, dst, srcstart, dststart, length)')
     @enforceargs(None, None, int, int, int)
     def copy_string_contents(src, dst, srcstart, dststart, length):
         assert srcstart >= 0
@@ -71,8 +70,6 @@
         keepalive_until_here(src)
         keepalive_until_here(dst)
     copy_string_contents._always_inline_ = True
-    #copy_string_contents.oopspec = (
-    #    '%s.copy_contents(src, dst, srcstart, dststart, length)' % name)
     return func_with_new_name(copy_string_contents, 'copy_%s_contents' % name)
 
 copy_string_contents = _new_copy_contents_fun(STR, Char, 'string')
@@ -147,7 +144,7 @@
         self.ll = LLHelpers
         self.malloc = mallocunicode
 
-    @elidable
+    @jit.elidable
     def ll_str(self, s):
         # XXX crazy that this is here, but I don't want to break
         #     rmodel logic
@@ -162,7 +159,7 @@
             result.chars[i] = cast_primitive(Char, c)
         return result
 
-    @elidable
+    @jit.elidable
     def ll_encode_latin1(self, s):
         length = len(s.chars)
         result = mallocstr(length)
@@ -261,7 +258,7 @@
 
 
 class LLHelpers(AbstractLLHelpers):
-    @elidable
+    @jit.elidable
     def ll_str_mul(s, times):
         if times < 0:
             times = 0
@@ -283,7 +280,7 @@
             i += j
         return newstr
 
-    @elidable
+    @jit.elidable
     def ll_char_mul(ch, times):
         if typeOf(ch) is Char:
             malloc = mallocstr
@@ -328,7 +325,7 @@
         return s
     ll_str2unicode.oopspec = 'str.str2unicode(str)'
 
-    @elidable
+    @jit.elidable
     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
@@ -344,7 +341,7 @@
     def ll_strfasthash(s):
         return s.hash     # assumes that the hash is already computed
 
-    @elidable
+    @jit.elidable
     def ll_strconcat(s1, s2):
         len1 = len(s1.chars)
         len2 = len(s2.chars)
@@ -356,7 +353,7 @@
         return newstr
     ll_strconcat.oopspec = 'stroruni.concat(s1, s2)'
 
-    @elidable
+    @jit.elidable
     def ll_strip(s, ch, left, right):
         s_len = len(s.chars)
         if s_len == 0:
@@ -374,7 +371,7 @@
         s.copy_contents(s, result, lpos, 0, r_len)
         return result
 
-    @elidable
+    @jit.elidable
     def ll_upper(s):
         s_chars = s.chars
         s_len = len(s_chars)
@@ -391,7 +388,7 @@
             i += 1
         return result
 
-    @elidable
+    @jit.elidable
     def ll_lower(s):
         s_chars = s.chars
         s_len = len(s_chars)
@@ -441,7 +438,7 @@
             i += 1
         return result
 
-    @elidable
+    @jit.elidable
     def ll_strcmp(s1, s2):
         if not s1 and not s2:
             return True
@@ -464,7 +461,7 @@
             i += 1
         return len1 - len2
 
-    @elidable
+    @jit.elidable
     def ll_streq(s1, s2):
         if s1 == s2:       # also if both are NULLs
             return True
@@ -484,7 +481,7 @@
         return True
     ll_streq.oopspec = 'stroruni.equal(s1, s2)'
 
-    @elidable
+    @jit.elidable
     def ll_startswith(s1, s2):
         len1 = len(s1.chars)
         len2 = len(s2.chars)
@@ -505,7 +502,7 @@
             return False
         return s.chars[0] == ch
 
-    @elidable
+    @jit.elidable
     def ll_endswith(s1, s2):
         len1 = len(s1.chars)
         len2 = len(s2.chars)
@@ -527,7 +524,7 @@
             return False
         return s.chars[len(s.chars) - 1] == ch
 
-    @elidable
+    @jit.elidable
     def ll_find_char(s, ch, start, end):
         i = start
         if end > len(s.chars):
@@ -539,7 +536,7 @@
         return -1
     ll_find_char._annenforceargs_ = [None, None, int, int]
 
-    @elidable
+    @jit.elidable
     def ll_rfind_char(s, ch, start, end):
         if end > len(s.chars):
             end = len(s.chars)
@@ -550,7 +547,7 @@
                 return i
         return -1
 
-    @elidable
+    @jit.elidable
     def ll_count_char(s, ch, start, end):
         count = 0
         i = start
@@ -618,7 +615,7 @@
             res = 0
         return res
 
-    @elidable
+    @jit.elidable
     def ll_search(s1, s2, start, end, mode):
         count = 0
         n = end - start
@@ -697,7 +694,13 @@
             return -1
         return count
 
+    @jit.look_inside_iff(lambda length, items: jit.isconstant(length) and length <= 2)
+    @enforceargs(int, None)
     def ll_join_strs(length, items):
+        # Special case for length 1 items, helps both the JIT and other code
+        if length == 1:
+            return items[0]
+
         num_items = length
         itemslen = 0
         i = 0
@@ -724,8 +727,8 @@
             res_index += item_len
             i += 1
         return result
-    ll_join_strs._annenforceargs_ = [int, None]
 
+    @jit.look_inside_iff(lambda length, chars, RES: jit.isconstant(length) and jit.isvirtual(chars))
     def ll_join_chars(length, chars, RES):
         # no need to optimize this, will be replaced by string builder
         # at some point soon
@@ -744,7 +747,7 @@
             i += 1
         return result
 
-    @elidable
+    @jit.elidable
     def _ll_stringslice(s1, start, stop):
         lgt = stop - start
         assert start >= 0
@@ -759,7 +762,7 @@
         return LLHelpers._ll_stringslice(s1, start, len(s1.chars))
 
     def ll_stringslice_startstop(s1, start, stop):
-        if we_are_jitted():
+        if jit.we_are_jitted():
             if stop > len(s1.chars):
                 stop = len(s1.chars)
         else:
@@ -842,7 +845,7 @@
         item.copy_contents(s, item, j, 0, i - j)
         return res
 
-    @elidable
+    @jit.elidable
     def ll_replace_chr_chr(s, c1, c2):
         length = len(s.chars)
         newstr = s.malloc(length)
@@ -857,7 +860,7 @@
             j += 1
         return newstr
 
-    @elidable
+    @jit.elidable
     def ll_contains(s, c):
         chars = s.chars
         strlen = len(chars)
@@ -868,7 +871,7 @@
             i += 1
         return False
 
-    @elidable
+    @jit.elidable
     def ll_int(s, base):
         if not 2 <= base <= 36:
             raise ValueError
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
@@ -247,7 +247,7 @@
         fn = None
         v_obj = hop.inputarg(r_func, arg=arg)
         s_pbc_fn = hop.args_s[arg]
-        methodname = r_func._get_method_name("simple_call", s_pbc_fn, params_annotation)
+        methodname = r_func._get_method_name("simple_call", s_pbc_fn, params_annotation, hop)
     elif isinstance(r_func, MethodOfFrozenPBCRepr):
         r_impl, nimplicitarg = r_func.get_r_implfunc()
         fn = r_impl.get_unique_llfn().value
diff --git a/pypy/rpython/ootypesystem/rpbc.py b/pypy/rpython/ootypesystem/rpbc.py
--- a/pypy/rpython/ootypesystem/rpbc.py
+++ b/pypy/rpython/ootypesystem/rpbc.py
@@ -130,14 +130,14 @@
     def call(self, opname, hop):
         s_pbc = hop.args_s[0]   # possibly more precise than self.s_pbc        
         args_s = hop.args_s[1:]
-        shape, index, callfamily = self._get_shape_index_callfamily(opname, s_pbc, args_s)
+        shape, index, callfamily = self._get_shape_index_callfamily(opname, s_pbc, args_s, hop)
         row_of_graphs = callfamily.calltables[shape][index]
         anygraph = row_of_graphs.itervalues().next()  # pick any witness
         hop2 = self.add_instance_arg_to_hop(hop, opname == "call_args")
         vlist = callparse.callparse(self.rtyper, anygraph, hop2, opname,
                                     r_self = self.r_im_self)
         rresult = callparse.getrresult(self.rtyper, anygraph)
-        derived_mangled = self._get_method_name(opname, s_pbc, args_s)
+        derived_mangled = self._get_method_name(opname, s_pbc, args_s, hop)
         cname = hop.inputconst(ootype.Void, derived_mangled)
         hop.exception_is_here()
         # sanity check: make sure that INSTANCE has the method
@@ -151,18 +151,18 @@
         else:
             return hop.llops.convertvar(v, rresult, hop.r_result)
 
-    def _get_shape_index_callfamily(self, opname, s_pbc, args_s):
+    def _get_shape_index_callfamily(self, opname, s_pbc, args_s, hop):
         bk = self.rtyper.annotator.bookkeeper
         args = bk.build_args(opname, args_s)
         args = args.prepend(self.s_im_self)
         descs = [desc.funcdesc for desc in s_pbc.descriptions]
         callfamily = descs[0].getcallfamily()
         shape, index = description.FunctionDesc.variant_for_call_site(
-                bk, callfamily, descs, args)
+                bk, callfamily, descs, args, hop.spaceop)
         return shape, index, callfamily
 
-    def _get_method_name(self, opname, s_pbc, args_s):
-        shape, index, callfamily = self._get_shape_index_callfamily(opname, s_pbc, args_s)
+    def _get_method_name(self, opname, s_pbc, args_s, hop):
+        shape, index, callfamily = self._get_shape_index_callfamily(opname, s_pbc, args_s, hop)
         mangled = mangle(self.methodname, self.rtyper.getconfig())
         row = self.concretetable[shape, index]
         derived_mangled = row_method_name(mangled, row.attrname)
diff --git a/pypy/rpython/rlist.py b/pypy/rpython/rlist.py
--- a/pypy/rpython/rlist.py
+++ b/pypy/rpython/rlist.py
@@ -116,7 +116,7 @@
         v_lst = hop.inputarg(self, 0)
         cRESLIST = hop.inputconst(Void, hop.r_result.LIST)
         return hop.gendirectcall(ll_copy, cRESLIST, v_lst)
-    
+
     def rtype_len(self, hop):
         v_lst, = hop.inputargs(self)
         if hop.args_s[0].listdef.listitem.resized:
@@ -132,7 +132,7 @@
         else:
             ll_func = ll_list_is_true_foldable
         return hop.gendirectcall(ll_func, v_lst)
-    
+
     def rtype_method_reverse(self, hop):
         v_lst, = hop.inputargs(self)
         hop.exception_cannot_occur()
@@ -273,7 +273,7 @@
         return pair(r_lst, r_int).rtype_getitem(hop, checkidx=True)
 
     rtype_getitem_idx_key = rtype_getitem_idx
-    
+
     def rtype_setitem((r_lst, r_int), hop):
         if hop.has_implicit_exception(IndexError):
             spec = dum_checkidx
@@ -331,7 +331,7 @@
 ##            return hop.gendirectcall(ll_both_none, v_lst1, v_lst2)
 
 ##        return pairtype(Repr, Repr).rtype_is_(pair(r_lst1, r_lst2), hop)
- 
+
     def rtype_eq((r_lst1, r_lst2), hop):
         assert r_lst1.item_repr == r_lst2.item_repr
         v_lst1, v_lst2 = hop.inputargs(r_lst1, r_lst2)
@@ -499,7 +499,7 @@
     else:
         check = item
     if (not malloc_zero_filled) or check: # as long as malloc it is known to zero the allocated memory avoid zeroing twice
-    
+
         i = 0
         while i < count:
             l.ll_setitem_fast(i, item)
@@ -633,7 +633,6 @@
         l.ll_setitem_fast(index, null)
     l._ll_resize_le(newlength)
     return res
-ll_pop_default.oopspec = 'list.pop(l)'
 
 def ll_pop_zero(func, l):
     length = l.ll_length()
diff --git a/pypy/rpython/rpbc.py b/pypy/rpython/rpbc.py
--- a/pypy/rpython/rpbc.py
+++ b/pypy/rpython/rpbc.py
@@ -322,7 +322,8 @@
         args = bk.build_args(opname, hop.args_s[1:])
         s_pbc = hop.args_s[0]   # possibly more precise than self.s_pbc
         descs = list(s_pbc.descriptions)
-        shape, index = description.FunctionDesc.variant_for_call_site(bk, self.callfamily, descs, args)
+        vfcs = description.FunctionDesc.variant_for_call_site
+        shape, index = vfcs(bk, self.callfamily, descs, args, hop.spaceop)
         row_of_graphs = self.callfamily.calltables[shape][index]
         anygraph = row_of_graphs.itervalues().next()  # pick any witness
         vfn = hop.inputarg(self, arg=0)
diff --git a/pypy/rpython/test/test_rint.py b/pypy/rpython/test/test_rint.py
--- a/pypy/rpython/test/test_rint.py
+++ b/pypy/rpython/test/test_rint.py
@@ -18,8 +18,8 @@
         t = TranslationContext()
         t.buildannotator().build_types(func, types)
         t.buildrtyper().specialize()
-        t.checkgraphs()    
-     
+        t.checkgraphs()
+
     def test_not1(self):
         self._test(snippet.not1, [int])
 
@@ -44,7 +44,7 @@
 
 
 class BaseTestRint(BaseRtypingTest):
-    
+
     def test_char_constant(self):
         def dummyfn(i):
             return chr(i)


More information about the pypy-commit mailing list