jtransform.py

import py

from rpython.jit.codewriter import support, heaptracker, longlong
from rpython.jit.codewriter.effectinfo import EffectInfo
from rpython.jit.codewriter.flatten import ListOfKind, IndirectCallTargets
from rpython.jit.codewriter.policy import log
from rpython.jit.metainterp import quasiimmut
from rpython.jit.metainterp.history import getkind
from rpython.jit.metainterp.typesystem import deref, arrayItem
from rpython.jit.metainterp.blackhole import BlackholeInterpreter
from rpython.flowspace.model import SpaceOperation, Variable, Constant,\
     c_last_exception
from rpython.rlib import objectmodel
from rpython.rlib.jit import _we_are_jitted
from rpython.rlib.rgc import lltype_is_gc
from rpython.rtyper.lltypesystem import lltype, llmemory, rstr, rffi
from rpython.rtyper.lltypesystem import rbytearray
from rpython.rtyper import rclass
from rpython.rtyper.rclass import IR_QUASIIMMUTABLE, IR_QUASIIMMUTABLE_ARRAY
from rpython.translator.unsimplify import varoftype

class UnsupportedMallocFlags(Exception):
    pass

def transform_graph(graph, cpu=None, callcontrol=None, portal_jd=None):
    """Transform a control flow graph to make it suitable for
    being flattened in a JitCode.
    """
    constant_fold_ll_issubclass(graph, cpu)
    t = Transformer(cpu, callcontrol, portal_jd)
    t.transform(graph)

def constant_fold_ll_issubclass(graph, cpu):
    # ll_issubclass can be inserted by the inliner to check exception types.
    # See corner case metainterp.test.test_exception:test_catch_different_class
    if cpu is None:
        return
    excmatch = cpu.rtyper.exceptiondata.fn_exception_match
    for block in list(graph.iterblocks()):
        for i, op in enumerate(block.operations):
            if (op.opname == 'direct_call' and
                    all(isinstance(a, Constant) for a in op.args) and
                    op.args[0].value._obj is excmatch._obj):
                constant_result = excmatch(*[a.value for a in op.args[1:]])
                block.operations[i] = SpaceOperation(
                    'same_as',
                    [Constant(constant_result, lltype.Bool)],
                    op.result)
                if block.exitswitch is op.result:
                    block.exitswitch = None
                    block.recloseblock(*[link for link in block.exits
                                         if link.exitcase == constant_result])

def integer_bounds(size, unsigned):
    if unsigned:
        return 0, 1 << (8 * size)
    else:
        return -(1 << (8 * size - 1)), 1 << (8 * size - 1)

class Transformer(object):
    vable_array_vars = None

    def __init__(self, cpu=None, callcontrol=None, portal_jd=None):
        self.cpu = cpu
        self.callcontrol = callcontrol
        self.portal_jd = portal_jd   # non-None only for the portal graph(s)

    def transform(self, graph):
        self.graph = graph
        for block in list(graph.iterblocks()):
            self.optimize_block(block)

    def optimize_block(self, block):
        if block.operations == ():
            return
        self.vable_array_vars = {}
        self.vable_flags = {}
        renamings = {}
        renamings_constants = {}    # subset of 'renamings', {Var:Const} only
        newoperations = []
        #
        def do_rename(var, var_or_const):
            if var.concretetype is lltype.Void:
                renamings[var] = Constant(None, lltype.Void)
                return
            renamings[var] = var_or_const
            if isinstance(var_or_const, Constant):
                value = var_or_const.value
                value = lltype._cast_whatever(var.concretetype, value)
                renamings_constants[var] = Constant(value, var.concretetype)
        #
        for op in block.operations:
            if renamings_constants:
                op = self._do_renaming(renamings_constants, op)
            oplist = self.rewrite_operation(op)
            #
            count_before_last_operation = len(newoperations)
            if not isinstance(oplist, list):
                oplist = [oplist]
            for op1 in oplist:
                if isinstance(op1, SpaceOperation):
                    newoperations.append(self._do_renaming(renamings, op1))
                elif op1 is None:
                    # rewrite_operation() returns None to mean "has no real
                    # effect, the result should just be renamed to args[0]"
                    if op.result is not None:
                        do_rename(op.result, renamings.get(op.args[0],
                                                           op.args[0]))
                elif isinstance(op1, Constant):
                    do_rename(op.result, op1)
                else:
                    raise TypeError(repr(op1))
        #
        if block.canraise:
            if len(newoperations) == count_before_last_operation:
                self._killed_exception_raising_operation(block)
        block.operations = newoperations
        block.exitswitch = renamings.get(block.exitswitch, block.exitswitch)
        self.follow_constant_exit(block)
        self.optimize_goto_if_not(block)
        if isinstance(block.exitswitch, tuple):
            self._check_no_vable_array(block.exitswitch)
        for link in block.exits:
            self._check_no_vable_array(link.args)
            self._do_renaming_on_link(renamings, link)

    def _do_renaming(self, rename, op):
        op = SpaceOperation(op.opname, op.args[:], op.result)
        for i, v in enumerate(op.args):
            if isinstance(v, Variable):
                if v in rename:
                    op.args[i] = rename[v]
            elif isinstance(v, ListOfKind):
                newlst = []
                for x in v:
                    if x in rename:
                        x = rename[x]
                    newlst.append(x)
                op.args[i] = ListOfKind(v.kind, newlst)
        return op

    def _check_no_vable_array(self, list):
        if not self.vable_array_vars:
            return
        for v in list:
            if v in self.vable_array_vars:
                vars = self.vable_array_vars[v]
                (v_base, arrayfielddescr, arraydescr) = vars
                raise AssertionError(
                    "A virtualizable array is passed around; it should\n"
                    "only be used immediately after being read.  Note\n"
                    "that a possible cause is indexing with an index not\n"
                    "known non-negative, or catching IndexError, or\n"
                    "not inlining at all (for tests: use listops=True).\n"
                    "This is about: %r\n"
                    "Occurred in: %r" % (arrayfielddescr, self.graph))
            # extra explanation: with the way things are organized in
            # rpython/rlist.py, the ll_getitem becomes a function call
            # that is typically meant to be inlined by the JIT, but
            # this does not work with vable arrays because
            # jtransform.py expects the getfield and the getarrayitem
            # to be in the same basic block.  It works a bit as a hack
            # for simple cases where we performed the backendopt
            # inlining before (even with a very low threshold, because
            # there is _always_inline_ on the relevant functions).

    def _do_renaming_on_link(self, rename, link):
        for i, v in enumerate(link.args):
            if isinstance(v, Variable):
                if v in rename:
                    link.args[i] = rename[v]

    def _killed_exception_raising_operation(self, block):
        assert block.exits[0].exitcase is None
        block.exits = block.exits[:1]
        block.exitswitch = None

    # ----------

    def follow_constant_exit(self, block):
        v = block.exitswitch
        if isinstance(v, Constant) and not block.canraise:
            llvalue = v.value
            for link in block.exits:
                if link.llexitcase == llvalue:
                    break
            else:
                assert link.exitcase == 'default'
            block.exitswitch = None
            link.exitcase = link.llexitcase = None
            block.recloseblock(link)

    def optimize_goto_if_not(self, block):
        """Replace code like 'v = int_gt(x,y); exitswitch = v'
           with just 'exitswitch = ('int_gt',x,y)'."""
        if len(block.exits) != 2:
            return False
        v = block.exitswitch
        if (block.canraise or isinstance(v, tuple)
            or v.concretetype != lltype.Bool):
            return False
        for op in block.operations[::-1]:
            # check if variable is used in block
            for arg in op.args:
                if arg == v:
                    return False
                if isinstance(arg, ListOfKind) and v in arg.content:
                    return False
            if v is op.result:
                if op.opname not in ('int_lt', 'int_le', 'int_eq', 'int_ne',
                                     'int_gt', 'int_ge',
                                     'float_lt', 'float_le', 'float_eq',
                                     'float_ne', 'float_gt', 'float_ge',
                                     'int_is_zero', 'int_is_true',
                                     'ptr_eq', 'ptr_ne',
                                     'ptr_iszero', 'ptr_nonzero'):
                    return False    # not a supported operation
                # ok! optimize this case
                block.operations.remove(op)
                block.exitswitch = (op.opname,) + tuple(op.args)
                #if op.opname in ('ptr_iszero', 'ptr_nonzero'):
                block.exitswitch += ('-live-before',)
                # if the variable escape to the next block along a link,
                # replace it with a constant, because we know its value
                for link in block.exits:
                    while v in link.args:
                        index = link.args.index(v)
                        link.args[index] = Constant(link.llexitcase,
                                                    lltype.Bool)
                return True
        return False

    # ----------

    def rewrite_operation(self, op):
        try:
            rewrite = _rewrite_ops[op.opname]
        except KeyError:
            raise Exception("the JIT doesn't support the operation %r"
                            " in %r" % (op, getattr(self, 'graph', '?')))
        return rewrite(self, op)

    def rewrite_op_same_as(self, op):
        if op.args[0] in self.vable_array_vars:
            self.vable_array_vars[op.result]= self.vable_array_vars[op.args[0]]

    def rewrite_op_cast_ptr_to_adr(self, op):
        if lltype_is_gc(op.args[0].concretetype):
            raise Exception("cast_ptr_to_adr for GC types unsupported")

    def rewrite_op_cast_pointer(self, op):
        newop = self.rewrite_op_same_as(op)
        assert newop is None
        return
        # disabled for now
        if (self._is_rclass_instance(op.args[0]) and
                self._is_rclass_instance(op.result)):
            FROM = op.args[0].concretetype.TO
            TO = op.result.concretetype.TO
            if lltype._castdepth(TO, FROM) > 0:
                vtable = heaptracker.get_vtable_for_gcstruct(self.cpu, TO)
                if vtable.subclassrange_max - vtable.subclassrange_min == 1:
                    # it's a precise class check
                    const_vtable = Constant(vtable, lltype.typeOf(vtable))
                    return [None, # hack, do the right renaming from op.args[0] to op.result
                            SpaceOperation("record_exact_class", [op.args[0], const_vtable], None)]

    def rewrite_op_likely(self, op):
        return None   # "no real effect"

    def rewrite_op_unlikely(self, op):
        return None   # "no real effect"

    def rewrite_op_raw_malloc_usage(self, op):
        if self.cpu.translate_support_code or isinstance(op.args[0], Variable):
            return   # the operation disappears
        else:
            # only for untranslated tests: get a real integer estimate
            arg = op.args[0].value
            arg = llmemory.raw_malloc_usage(arg)
            return [Constant(arg, lltype.Signed)]

    def rewrite_op_jit_record_exact_class(self, op):
        return SpaceOperation("record_exact_class", [op.args[0], op.args[1]], None)

    def rewrite_op_debug_assert_not_none(self, op):
        if isinstance(op.args[0], Variable):
            return SpaceOperation('assert_not_none', [op.args[0]], None)
        else:
            return []

    def rewrite_op_cast_bool_to_int(self, op): pass
    def rewrite_op_cast_bool_to_uint(self, op): pass
    def rewrite_op_cast_char_to_int(self, op): pass
    def rewrite_op_cast_unichar_to_int(self, op): pass
    def rewrite_op_cast_int_to_char(self, op): pass
    def rewrite_op_cast_int_to_unichar(self, op): pass
    def rewrite_op_cast_int_to_uint(self, op): pass
    def rewrite_op_cast_uint_to_int(self, op): pass

    def _rewrite_symmetric(self, op):
        """Rewrite 'c1+v2' into 'v2+c1' in an attempt to avoid generating
        too many variants of the bytecode."""
        if (isinstance(op.args[0], Constant) and
            isinstance(op.args[1], Variable)):
            reversename = {'int_lt': 'int_gt',
                           'int_le': 'int_ge',
                           'int_gt': 'int_lt',
                           'int_ge': 'int_le',
                           'uint_lt': 'uint_gt',
                           'uint_le': 'uint_ge',
                           'uint_gt': 'uint_lt',
                           'uint_ge': 'uint_le',
                           'float_lt': 'float_gt',
                           'float_le': 'float_ge',
                           'float_gt': 'float_lt',
                           'float_ge': 'float_le',
                           }.get(op.opname, op.opname)
            return SpaceOperation(reversename,
                                  [op.args[1], op.args[0]] + op.args[2:],
                                  op.result)
        else:
            return op

    rewrite_op_int_add = _rewrite_symmetric
    rewrite_op_int_mul = _rewrite_symmetric
    rewrite_op_int_and = _rewrite_symmetric
    rewrite_op_int_or  = _rewrite_symmetric
    rewrite_op_int_xor = _rewrite_symmetric
    rewrite_op_int_lt  = _rewrite_symmetric
    rewrite_op_int_le  = _rewrite_symmetric
    rewrite_op_int_gt  = _rewrite_symmetric
    rewrite_op_int_ge  = _rewrite_symmetric
    rewrite_op_uint_lt = _rewrite_symmetric
    rewrite_op_uint_le = _rewrite_symmetric
    rewrite_op_uint_gt = _rewrite_symmetric
    rewrite_op_uint_ge = _rewrite_symmetric

    rewrite_op_float_add = _rewrite_symmetric
    rewrite_op_float_mul = _rewrite_symmetric
    rewrite_op_float_lt  = _rewrite_symmetric
    rewrite_op_float_le  = _rewrite_symmetric
    rewrite_op_float_gt  = _rewrite_symmetric
    rewrite_op_float_ge  = _rewrite_symmetric

    def rewrite_op_int_add_ovf(self, op):
        op0 = self._rewrite_symmetric(op)
        op1 = SpaceOperation('-live-', [], None)
        return [op1, op0]

    rewrite_op_int_mul_ovf = rewrite_op_int_add_ovf

    def rewrite_op_int_sub_ovf(self, op):
        op1 = SpaceOperation('-live-', [], None)
        return [op1, op]

    def _noop_rewrite(self, op):
        return op

    rewrite_op_convert_float_bytes_to_longlong = _noop_rewrite
    rewrite_op_convert_longlong_bytes_to_float = _noop_rewrite
    cast_ptr_to_weakrefptr = _noop_rewrite
    cast_weakrefptr_to_ptr = _noop_rewrite

    # ----------
    # Various kinds of calls

    def rewrite_op_direct_call(self, op):
        kind = self.callcontrol.guess_call_kind(op)
        return getattr(self, 'handle_%s_call' % kind)(op)

    def rewrite_op_indirect_call(self, op):
        kind = self.callcontrol.guess_call_kind(op)
        return getattr(self, 'handle_%s_indirect_call' % kind)(op)

    def rewrite_call(self, op, namebase, initialargs, args=None,
                     calldescr=None, force_ir=False):
        """Turn 'i0 = direct_call(fn, i1, i2, ref1, ref2)'
           into 'i0 = xxx_call_ir_i(fn, descr, [i1,i2], [ref1,ref2])'.
           The name is one of '{residual,direct}_call_{r,ir,irf}_{i,r,f,v}'."""
        if args is None:
            args = op.args[1:]
        self._check_no_vable_array(args)
        lst_i, lst_r, lst_f = self.make_three_lists(args)
        reskind = getkind(op.result.concretetype)[0]
        if lst_f or reskind == 'f': kinds = 'irf'
        elif lst_i or force_ir: kinds = 'ir'
        else: kinds = 'r'
        if force_ir: assert kinds == 'ir'    # no 'f'
        sublists = []
        if 'i' in kinds: sublists.append(lst_i)
        if 'r' in kinds: sublists.append(lst_r)
        if 'f' in kinds: sublists.append(lst_f)
        if calldescr is not None:
            sublists.append(calldescr)
        return SpaceOperation('%s_%s_%s' % (namebase, kinds, reskind),
                              initialargs + sublists, op.result)

    def make_three_lists(self, vars):
        args_i = []
        args_r = []
        args_f = []
        for v in vars:
            self.add_in_correct_list(v, args_i, args_r, args_f)
        return [ListOfKind('int', args_i),
                ListOfKind('ref', args_r),
                ListOfKind('float', args_f)]

    def add_in_correct_list(self, v, lst_i, lst_r, lst_f):
        kind = getkind(v.concretetype)
        if kind == 'void': return
        elif kind == 'int': lst = lst_i
        elif kind == 'ref': lst = lst_r
        elif kind == 'float': lst = lst_f
        else: raise AssertionError(kind)
        lst.append(v)

    def handle_residual_call(self, op, extraargs=[], may_call_jitcodes=False,
                             oopspecindex=EffectInfo.OS_NONE,
                             extraeffect=None,
                             extradescr=None):
        """A direct_call turns into the operation 'residual_call_xxx' if it
        is calling a function that we don't want to JIT.  The initial args
        of 'residual_call_xxx' are the function to call, and its calldescr."""
        calldescr = self.callcontrol.getcalldescr(op, oopspecindex=oopspecindex,
                                                  extraeffect=extraeffect,
                                                  extradescr=extradescr)
        op1 = self.rewrite_call(op, 'residual_call',
                                [op.args[0]] + extraargs, calldescr=calldescr)
        if may_call_jitcodes or self.callcontrol.calldescr_canraise(calldescr):
            op1 = [op1, SpaceOperation('-live-', [], None)]
        return op1

    def handle_regular_call(self, op):
        """A direct_call turns into the operation 'inline_call_xxx' if it
        is calling a function that we want to JIT.  The initial arg of
        'inline_call_xxx' is the JitCode of the called function."""
        [targetgraph] = self.callcontrol.graphs_from(op)
        jitcode = self.callcontrol.get_jitcode(targetgraph,
                                               called_from=self.graph)
        op0 = self.rewrite_call(op, 'inline_call', [jitcode])
        op1 = SpaceOperation('-live-', [], None)
        return [op0, op1]

    def handle_builtin_call(self, op):
        oopspec_name, args = support.decode_builtin_call(op)
        # dispatch to various implementations depending on the oopspec_name
        if oopspec_name.startswith('list.') or oopspec_name.startswith('newlist'):
            prepare = self._handle_list_call
        elif oopspec_name.startswith('int.'):
            prepare = self._handle_int_special
        elif oopspec_name.startswith('stroruni.'):
            prepare = self._handle_stroruni_call
        elif oopspec_name == 'str.str2unicode':
            prepare = self._handle_str2unicode_call
        elif oopspec_name.startswith('virtual_ref'):
            prepare = self._handle_virtual_ref_call
        elif oopspec_name.startswith('jit.'):
            prepare = self._handle_jit_call
        elif oopspec_name.startswith('libffi_'):
            prepare = self._handle_libffi_call
        elif oopspec_name.startswith('math.sqrt'):
            prepare = self._handle_math_sqrt_call
        elif oopspec_name.startswith('rgc.'):
            prepare = self._handle_rgc_call
        elif oopspec_name.startswith('rvmprof.'):
            prepare = self._handle_rvmprof_call
        elif oopspec_name.endswith('dict.lookup'):
            # also ordereddict.lookup
            prepare = self._handle_dict_lookup_call
        else:
            prepare = self.prepare_builtin_call
        try:
            op1 = prepare(op, oopspec_name, args)
        except NotSupported:
            op1 = op
        # If the resulting op1 is still a direct_call, turn it into a
        # residual_call.
        if isinstance(op1, SpaceOperation) and op1.opname == 'direct_call':
            op1 = self.handle_residual_call(op1)
        return op1

    def handle_recursive_call(self, op):
        jitdriver_sd = self.callcontrol.jitdriver_sd_from_portal_runner_ptr(
            op.args[0].value)
        assert jitdriver_sd is not None
        ops = self.promote_greens(op.args[1:], jitdriver_sd.jitdriver)
        num_green_args = len(jitdriver_sd.jitdriver.greens)
        args = ([Constant(jitdriver_sd.index, lltype.Signed)] +
                self.make_three_lists(op.args[1:1+num_green_args]) +
                self.make_three_lists(op.args[1+num_green_args:]))
        kind = getkind(op.result.concretetype)[0]
        op0 = SpaceOperation('recursive_call_%s' % kind, args, op.result)
        op1 = SpaceOperation('-live-', [], None)
        return ops + [op0, op1]

    handle_residual_indirect_call = handle_residual_call

    def handle_regular_indirect_call(self, op):
        """An indirect call where at least one target has a JitCode."""
        lst = []
        for targetgraph in self.callcontrol.graphs_from(op):
            jitcode = self.callcontrol.get_jitcode(targetgraph,
                                                   called_from=self.graph)
            lst.append(jitcode)
        op0 = SpaceOperation('-live-', [], None)
        op1 = SpaceOperation('int_guard_value', [op.args[0]], None)
        op2 = self.handle_residual_call(op, [IndirectCallTargets(lst)], True)
        result = [op0, op1]
        if isinstance(op2, list):
            result += op2
        else:
            result.append(op2)
        return result

    def prepare_builtin_call(self, op, oopspec_name, args,
                              extra=None, extrakey=None):
        argtypes = [v.concretetype for v in args]
        resulttype = op.result.concretetype
        c_func, TP = support.builtin_func_for_spec(self.cpu.rtyper,
                                                   oopspec_name, argtypes,
                                                   resulttype, extra, extrakey)
        return SpaceOperation('direct_call', [c_func] + args, op.result)


    def _do_builtin_call(self, op, oopspec_name=None, args=None,
                         extra=None, extrakey=None):
        if oopspec_name is None: oopspec_name = op.opname
        if args is None: args = op.args
        op1 = self.prepare_builtin_call(op, oopspec_name, args,
                                        extra, extrakey)
        return self.rewrite_op_direct_call(op1)

    # XXX some of the following functions should not become residual calls
    # but be really compiled
    rewrite_op_int_abs                = _do_builtin_call
    rewrite_op_int_floordiv           = _do_builtin_call
    rewrite_op_int_mod                = _do_builtin_call
    rewrite_op_llong_abs              = _do_builtin_call
    rewrite_op_llong_floordiv         = _do_builtin_call
    rewrite_op_llong_mod              = _do_builtin_call
    rewrite_op_ullong_floordiv        = _do_builtin_call
    rewrite_op_ullong_mod             = _do_builtin_call
    rewrite_op_gc_identityhash        = _do_builtin_call
    rewrite_op_gc_id                  = _do_builtin_call
    rewrite_op_gc_pin                 = _do_builtin_call
    rewrite_op_gc_unpin               = _do_builtin_call
    rewrite_op_cast_float_to_uint     = _do_builtin_call
    rewrite_op_cast_uint_to_float     = _do_builtin_call
    rewrite_op_weakref_create         = _do_builtin_call
    rewrite_op_weakref_deref          = _do_builtin_call
    rewrite_op_gc_add_memory_pressure = _do_builtin_call

    # ----------
    # getfield/setfield/mallocs etc.

    def rewrite_op_hint(self, op):
        hints = op.args[1].value

        # hack: if there are both 'promote' and 'promote_string', kill
        # one of them based on the type of the value
        if hints.get('promote_string') and hints.get('promote'):
            hints = hints.copy()
            if op.args[0].concretetype == lltype.Ptr(rstr.STR):
                del hints['promote']
            else:
                del hints['promote_string']

        if hints.get('promote') and op.args[0].concretetype is not lltype.Void:
            assert op.args[0].concretetype != lltype.Ptr(rstr.STR)
            kind = getkind(op.args[0].concretetype)
            op0 = SpaceOperation('-live-', [], None)
            op1 = SpaceOperation('%s_guard_value' % kind, [op.args[0]], None)
            # the special return value None forces op.result to be considered
            # equal to op.args[0]
            return [op0, op1, None]
        if (hints.get('promote_string') and
            op.args[0].concretetype is not lltype.Void):
            S = lltype.Ptr(rstr.STR)
            assert op.args[0].concretetype == S
            self._register_extra_helper(EffectInfo.OS_STREQ_NONNULL,
                                        "str.eq_nonnull",
                                        [S, S],
                                        lltype.Signed,
                                        EffectInfo.EF_ELIDABLE_CANNOT_RAISE)
            descr, p = self.callcontrol.callinfocollection.callinfo_for_oopspec(
                EffectInfo.OS_STREQ_NONNULL)
            # XXX this is fairly ugly way of creating a constant,
            #     however, callinfocollection has no better interface
            c = Constant(p.adr.ptr, lltype.typeOf(p.adr.ptr))
            op1 = SpaceOperation('str_guard_value', [op.args[0], c, descr],
                                 op.result)
            return [SpaceOperation('-live-', [], None), op1, None]
        if hints.get('force_virtualizable'):
            return SpaceOperation('hint_force_virtualizable', [op.args[0]], None)
        if hints.get('force_no_const'):   # for tests only
            assert getkind(op.args[0].concretetype) == 'int'
            return SpaceOperation('int_same_as', [op.args[0]], op.result)
        log.WARNING('ignoring hint %r at %r' % (hints, self.graph))

    def _rewrite_raw_malloc(self, op, name, args):
        # NB. the operation 'raw_malloc' is not supported; this is for
        # the operation 'malloc'/'malloc_varsize' with {flavor: 'gc'}
        d = op.args[1].value.copy()
        d.pop('flavor')
        add_memory_pressure = d.pop('add_memory_pressure', False)
        zero = d.pop('zero', False)
        track_allocation = d.pop('track_allocation', True)
        if d:
            raise UnsupportedMallocFlags(d)
        if zero:
            name += '_zero'
        if add_memory_pressure:
            name += '_add_memory_pressure'
        if not track_allocation:
            name += '_no_track_allocation'
        TYPE = op.args[0].value
        op1 = self.prepare_builtin_call(op, name, args, (TYPE,), TYPE)
        if name.startswith('raw_malloc_varsize') and TYPE.OF == lltype.Char:
            return self._handle_oopspec_call(op1, args,
                                             EffectInfo.OS_RAW_MALLOC_VARSIZE_CHAR,
                                             EffectInfo.EF_CAN_RAISE)
        return self.rewrite_op_direct_call(op1)

    def rewrite_op_malloc_varsize(self, op):
        if op.args[1].value['flavor'] == 'raw':
            return self._rewrite_raw_malloc(op, 'raw_malloc_varsize',
                                            [op.args[2]])
        if op.args[0].value == rstr.STR:
            return SpaceOperation('newstr', [op.args[2]], op.result)
        elif op.args[0].value == rstr.UNICODE:
            return SpaceOperation('newunicode', [op.args[2]], op.result)
        else:
            # XXX only strings or simple arrays for now
            ARRAY = op.args[0].value
            arraydescr = self.cpu.arraydescrof(ARRAY)
            if op.args[1].value.get('zero', False):
                opname = 'new_array_clear'
            elif ((isinstance(ARRAY.OF, lltype.Ptr) and ARRAY.OF._needsgc()) or
                  isinstance(ARRAY.OF, lltype.Struct)):
                opname = 'new_array_clear'
            else:
                opname = 'new_array'
            return SpaceOperation(opname, [op.args[2], arraydescr], op.result)

    def zero_contents(self, ops, v, TYPE):
        if isinstance(TYPE, lltype.Struct):
            for name, FIELD in TYPE._flds.iteritems():
                if isinstance(FIELD, lltype.Struct):
                    # substruct
                    self.zero_contents(ops, v, FIELD)
                else:
                    c_name = Constant(name, lltype.Void)
                    c_null = Constant(FIELD._defl(), FIELD)
                    op = SpaceOperation('setfield', [v, c_name, c_null],
                                        None)
                    self.extend_with(ops, self.rewrite_op_setfield(op,
                                          override_type=TYPE))
        elif isinstance(TYPE, lltype.Array):
            assert False # this operation disappeared
        else:
            raise TypeError("Expected struct or array, got '%r'", (TYPE,))
        if len(ops) == 1:
            return ops[0]
        return ops

    def extend_with(self, l, ops):
        if ops is None:
            return
        if isinstance(ops, list):
            l.extend(ops)
        else:
            l.append(ops)

    def rewrite_op_free(self, op):
        d = op.args[1].value.copy()
        assert d['flavor'] == 'raw'
        d.pop('flavor')
        track_allocation = d.pop('track_allocation', True)
        if d:
            raise UnsupportedMallocFlags(d)
        STRUCT = op.args[0].concretetype.TO
        name = 'raw_free'
        if not track_allocation:
            name += '_no_track_allocation'
        op1 = self.prepare_builtin_call(op, name, [op.args[0]], (STRUCT,),
                                        STRUCT)
        if name.startswith('raw_free'):
            return self._handle_oopspec_call(op1, [op.args[0]],
                                             EffectInfo.OS_RAW_FREE,
                                             EffectInfo.EF_CANNOT_RAISE)
        return self.rewrite_op_direct_call(op1)

    def rewrite_op_getarrayitem(self, op):
        ARRAY = op.args[0].concretetype.TO
        if self._array_of_voids(ARRAY):
            return []
        if isinstance(ARRAY, lltype.FixedSizeArray):
            raise NotImplementedError(
                "%r uses %r, which is not supported by the JIT codewriter"
                % (self.graph, ARRAY))
        if op.args[0] in self.vable_array_vars:     # for virtualizables
            vars = self.vable_array_vars[op.args[0]]
            (v_base, arrayfielddescr, arraydescr) = vars
            kind = getkind(op.result.concretetype)
            return [SpaceOperation('-live-', [], None),
                    SpaceOperation('getarrayitem_vable_%s' % kind[0],
                                   [v_base, op.args[1], arrayfielddescr,
                                    arraydescr], op.result)]
        # normal case follows
        pure = ''
        immut = ARRAY._immutable_field(None)
        if immut:
            pure = '_pure'
        arraydescr = self.cpu.arraydescrof(ARRAY)
        kind = getkind(op.result.concretetype)
        if ARRAY._gckind != 'gc':
            assert ARRAY._gckind == 'raw'
            if kind == 'r':
                raise Exception("getarrayitem_raw_r not supported")
            pure = ''   # always redetected from pyjitpl.py: we don't need
                        # a '_pure' version of getarrayitem_raw
        return SpaceOperation('getarrayitem_%s_%s%s' % (ARRAY._gckind,
                                                        kind[0], pure),
                              [op.args[0], op.args[1], arraydescr],
                              op.result)

    def rewrite_op_setarrayitem(self, op):
        ARRAY = op.args[0].concretetype.TO
        if self._array_of_voids(ARRAY):
            return []
        if isinstance(ARRAY, lltype.FixedSizeArray):
            raise NotImplementedError(
                "%r uses %r, which is not supported by the JIT codewriter"
                % (self.graph, ARRAY))
        if op.args[0] in self.vable_array_vars:     # for virtualizables
            vars = self.vable_array_vars[op.args[0]]
            (v_base, arrayfielddescr, arraydescr) = vars
            kind = getkind(op.args[2].concretetype)
            return [SpaceOperation('-live-', [], None),
                    SpaceOperation('setarrayitem_vable_%s' % kind[0],
                                   [v_base, op.args[1], op.args[2],
                                    arrayfielddescr, arraydescr], None)]
        arraydescr = self.cpu.arraydescrof(ARRAY)
        kind = getkind(op.args[2].concretetype)
        return SpaceOperation('setarrayitem_%s_%s' % (ARRAY._gckind, kind[0]),
                              [op.args[0], op.args[1], op.args[2], arraydescr],
                              None)

    def rewrite_op_getarraysize(self, op):
        ARRAY = op.args[0].concretetype.TO
        assert ARRAY._gckind == 'gc'
        if op.args[0] in self.vable_array_vars:     # for virtualizables
            vars = self.vable_array_vars[op.args[0]]
            (v_base, arrayfielddescr, arraydescr) = vars
            return [SpaceOperation('-live-', [], None),
                    SpaceOperation('arraylen_vable',
                                   [v_base, arrayfielddescr, arraydescr],
                                   op.result)]
        # normal case follows
        arraydescr = self.cpu.arraydescrof(ARRAY)
        return SpaceOperation('arraylen_gc', [op.args[0], arraydescr],
                              op.result)

    def rewrite_op_getarraysubstruct(self, op):
        ARRAY = op.args[0].concretetype.TO
        assert ARRAY._gckind == 'raw'
        assert ARRAY._hints.get('nolength') is True
        return self.rewrite_op_direct_ptradd(op)

    def _array_of_voids(self, ARRAY):
        return ARRAY.OF == lltype.Void

    def rewrite_op_getfield(self, op):
        if self.is_typeptr_getset(op):
            return self.handle_getfield_typeptr(op)
        # turn the flow graph 'getfield' operation into our own version
        [v_inst, c_fieldname] = op.args
        RESULT = op.result.concretetype
        if RESULT is lltype.Void:
            return
        # check for virtualizable
        try:
            if self.is_virtualizable_getset(op):
                descr = self.get_virtualizable_field_descr(op)
                kind = getkind(RESULT)[0]
                return [SpaceOperation('-live-', [], None),
                        SpaceOperation('getfield_vable_%s' % kind,
                                       [v_inst, descr], op.result)]
        except VirtualizableArrayField as e:
            # xxx hack hack hack
            vinfo = e.args[1]
            arrayindex = vinfo.array_field_counter[op.args[1].value]
            arrayfielddescr = vinfo.array_field_descrs[arrayindex]
            arraydescr = vinfo.array_descrs[arrayindex]
            self.vable_array_vars[op.result] = (op.args[0],
                                                arrayfielddescr,
                                                arraydescr)
            return []
        # check for the string or unicode hash field
        STRUCT = v_inst.concretetype.TO
        if STRUCT == rstr.STR:
            assert c_fieldname.value == 'hash'
            return SpaceOperation('strhash', [v_inst], op.result)
        elif STRUCT == rstr.UNICODE:
            assert c_fieldname.value == 'hash'
            return SpaceOperation('unicodehash', [v_inst], op.result)
        # check for _immutable_fields_ hints
        immut = STRUCT._immutable_field(c_fieldname.value)
        need_live = False
        if immut:
            if (self.callcontrol is not None and
                self.callcontrol.could_be_green_field(STRUCT,
                                                      c_fieldname.value)):
                pure = '_greenfield'
                need_live = True
            else:
                pure = '_pure'
        else:
            pure = ''
        self.check_field_access(STRUCT)
        argname = getattr(STRUCT, '_gckind', 'gc')
        descr = self.cpu.fielddescrof(STRUCT, c_fieldname.value)
        kind = getkind(RESULT)[0]
        if argname != 'gc':
            assert argname == 'raw'
            if (kind, pure) == ('r', ''):
                # note: a pure 'getfield_raw_r' is used e.g. to load class
                # attributes that are GC objects, so that one is supported.
                raise Exception("getfield_raw_r (without _pure) not supported")
            pure = ''   # always redetected from pyjitpl.py: we don't need
                        # a '_pure' version of getfield_raw
        #
        op1 = SpaceOperation('getfield_%s_%s%s' % (argname, kind, pure),
                             [v_inst, descr], op.result)
        #
        if immut in (IR_QUASIIMMUTABLE, IR_QUASIIMMUTABLE_ARRAY):
            op1.opname += "_pure"
            descr1 = self.cpu.fielddescrof(
                STRUCT,
                quasiimmut.get_mutate_field_name(c_fieldname.value))
            return [SpaceOperation('-live-', [], None),
                   SpaceOperation('record_quasiimmut_field',
                                  [v_inst, descr, descr1], None),
                   op1]
        if need_live:
            return [SpaceOperation('-live-', [], None), op1]
        return op1

    def rewrite_op_setfield(self, op, override_type=None):
        if self.is_typeptr_getset(op):
            # ignore the operation completely -- instead, it's done by 'new'
            return
        self._check_no_vable_array(op.args)
        # turn the flow graph 'setfield' operation into our own version
        [v_inst, c_fieldname, v_value] = op.args
        RESULT = v_value.concretetype
        if override_type is not None:
            TYPE = override_type
        else:
            TYPE = v_inst.concretetype.TO
        if RESULT is lltype.Void:
            return
        # check for virtualizable
        if self.is_virtualizable_getset(op):
            descr = self.get_virtualizable_field_descr(op)
            kind = getkind(RESULT)[0]
            return [SpaceOperation('-live-', [], None),
                    SpaceOperation('setfield_vable_%s' % kind,
                                   [v_inst, v_value, descr], None)]
        self.check_field_access(TYPE)
        if override_type:
            argname = 'gc'
        else:
            argname = getattr(TYPE, '_gckind', 'gc')
        descr = self.cpu.fielddescrof(TYPE, c_fieldname.value)
        kind = getkind(RESULT)[0]
        if argname == 'raw' and kind == 'r':
            raise Exception("setfield_raw_r not supported")
        return SpaceOperation('setfield_%s_%s' % (argname, kind),
                              [v_inst, v_value, descr],
                              None)

    def rewrite_op_getsubstruct(self, op):
        STRUCT = op.args[0].concretetype.TO
        argname = getattr(STRUCT, '_gckind', 'gc')
        if argname != 'raw':
            raise Exception("%r: only supported for gckind=raw" % (op,))
        ofs = llmemory.offsetof(STRUCT, op.args[1].value)
        return SpaceOperation('int_add',
                              [op.args[0], Constant(ofs, lltype.Signed)],
                              op.result)

    def is_typeptr_getset(self, op):
        return (op.args[1].value == 'typeptr' and
                op.args[0].concretetype.TO._hints.get('typeptr'))

    def check_field_access(self, STRUCT):
        # check against a GcStruct with a nested GcStruct as a first argument
        # but which is not an object at all; see metainterp/test/test_loop,
        # test_regular_pointers_in_short_preamble.
        if not isinstance(STRUCT, lltype.GcStruct):
            return
        if STRUCT._first_struct() == (None, None):
            return
        PARENT = STRUCT
        while not PARENT._hints.get('typeptr'):
            _, PARENT = PARENT._first_struct()
            if PARENT is None:
                raise NotImplementedError("%r is a GcStruct using nesting but "
                                          "not inheriting from object" %
                                          (STRUCT,))

    def get_vinfo(self, v_virtualizable):
        if self.callcontrol is None:      # for tests
            return None
        return self.callcontrol.get_vinfo(v_virtualizable.concretetype)

    def is_virtualizable_getset(self, op):
        # every access of an object of exactly the type VTYPEPTR is
        # likely to be a virtualizable access, but we still have to
        # check it in pyjitpl.py.
        vinfo = self.get_vinfo(op.args[0])
        if vinfo is None:
            return False
        res = False
        if op.args[1].value in vinfo.static_field_to_extra_box:
            res = True
        if op.args[1].value in vinfo.array_fields:
            res = VirtualizableArrayField(self.graph, vinfo)

        if res:
            flags = self.vable_flags[op.args[0]]
            if 'fresh_virtualizable' in flags:
                return False
        if isinstance(res, Exception):
            raise res
        return res

    def get_virtualizable_field_descr(self, op):
        fieldname = op.args[1].value
        vinfo = self.get_vinfo(op.args[0])
        index = vinfo.static_field_to_extra_box[fieldname]
        return vinfo.static_field_descrs[index]

    def handle_getfield_typeptr(self, op):
        if isinstance(op.args[0], Constant):
            cls = op.args[0].value.typeptr
            return Constant(cls, concretetype=rclass.CLASSTYPE)
        op0 = SpaceOperation('-live-', [], None)
        op1 = SpaceOperation('guard_class', [op.args[0]], op.result)
        return [op0, op1]

    def rewrite_op_malloc(self, op):
        d = op.args[1].value
        if d.get('nonmovable', False):
            raise UnsupportedMallocFlags(d)
        if d['flavor'] == 'raw':
            return self._rewrite_raw_malloc(op, 'raw_malloc_fixedsize', [])
        #
        if d.get('zero', False):
            zero = True
        else:
            zero = False
        STRUCT = op.args[0].value
        vtable = heaptracker.get_vtable_for_gcstruct(self.cpu, STRUCT)
        if vtable:
            # do we have a __del__?
            try:
                rtti = lltype.getRuntimeTypeInfo(STRUCT)
            except ValueError:
                pass
            else:
                if hasattr(rtti._obj, 'destructor_funcptr'):
                    RESULT = lltype.Ptr(STRUCT)
                    assert RESULT == op.result.concretetype
                    return self._do_builtin_call(op, 'alloc_with_del', [],
                                                 extra=(RESULT, vtable),
                                                 extrakey=STRUCT)
            opname = 'new_with_vtable'
        else:
            opname = 'new'
            vtable = lltype.nullptr(rclass.OBJECT_VTABLE)
        sizedescr = self.cpu.sizeof(STRUCT, vtable)
        op1 = SpaceOperation(opname, [sizedescr], op.result)
        if zero:
            return self.zero_contents([op1], op.result, STRUCT)
        return op1

    def _has_gcptrs_in(self, STRUCT):
        if isinstance(STRUCT, lltype.Array):
            ITEM = STRUCT.OF
            if isinstance(ITEM, lltype.Struct):
                STRUCT = ITEM
            else:
                return isinstance(ITEM, lltype.Ptr) and ITEM._needsgc()
        for FIELD in STRUCT._flds.values():
            if isinstance(FIELD, lltype.Ptr) and FIELD._needsgc():
                return True
            elif isinstance(FIELD, lltype.Struct):