duktape/tools/genbuiltins.py

#!/usr/bin/env python2
#
#  Generate initialization data for built-in strings and objects.
#
#  Supports two different initialization approaches:
#
#    1. Bit-packed format for unpacking strings and objects during
#       heap or thread init into RAM-based structures.  This is the
#       default behavior.
#
#    2. Embedding strings and/or objects into a read-only data section
#       at compile time.  This is useful for low memory targets to reduce
#       memory usage.  Objects in data section will be immutable.
#
#  Both of these have practical complications like endianness differences,
#  pointer compression variants, object property table layout variants,
#  and so on.  Multiple #if defined()'d initializer sections are emitted
#  to cover all supported alternatives.
#

import logging
import sys
logging.basicConfig(level=logging.INFO, stream=sys.stdout, format='%(name)-21s %(levelname)-7s %(message)s')
logger = logging.getLogger('genbuiltins.py')
logger.setLevel(logging.INFO)

import os
import re
import traceback
import json
import yaml
import math
import struct
import optparse
import copy
import logging

import dukutil

# Fixed seed for ROM strings, must match src-input/duk_heap_alloc.c.
DUK__FIXED_HASH_SEED = 0xabcd1234

# Base value for compressed ROM pointers, used range is [ROMPTR_FIRST,0xffff].
# Must match DUK_USE_ROM_PTRCOMP_FIRST (generated header checks).
ROMPTR_FIRST = 0xf800  # 2048 should be enough; now around ~1000 used

# ROM string table size
ROMSTR_LOOKUP_SIZE = 256

#
#  Miscellaneous helpers
#

# Convert Unicode to bytes, identifying Unicode U+0000 to U+00FF as bytes.
# This representation is used in YAML metadata and allows invalid UTF-8 to
# be represented exactly (which is necessary).
def unicode_to_bytes(x):
    if isinstance(x, str):
        return x
    tmp = ''
    for c in x:
        if ord(c) > 0xff:
            raise Exception('invalid codepoint: %r' % x)
        tmp += chr(ord(c))
    assert(isinstance(tmp, str))
    return tmp

# Convert bytes to Unicode, identifying bytes as U+0000 to U+00FF.
def bytes_to_unicode(x):
    if isinstance(x, unicode):
        return x
    tmp = u''
    for c in x:
        tmp += unichr(ord(c))
    assert(isinstance(tmp, unicode))
    return tmp

# Convert all strings in an object to bytes recursively.  Useful for
# normalizing all strings in a YAML document.
def recursive_strings_to_bytes(doc):
    def f(x):
        if isinstance(x, unicode):
            return unicode_to_bytes(x)
        if isinstance(x, dict):
            res = {}
            for k in x.keys():
                res[f(k)] = f(x[k])
            return res
        if isinstance(x, list):
            res = []
            for e in x:
                res.append(f(e))
            return res
        return x

    return f(doc)

# Convert all strings in an object to from bytes to Unicode recursively.
# Useful for writing back JSON/YAML dumps.
def recursive_bytes_to_strings(doc):
    def f(x):
        if isinstance(x, str):
            return bytes_to_unicode(x)
        if isinstance(x, dict):
            res = {}
            for k in x.keys():
                res[f(k)] = f(x[k])
            return res
        if isinstance(x, list):
            res = []
            for e in x:
                res.append(f(e))
            return res
        return x

    return f(doc)

# Check if string is an "array index" in Ecmascript terms.
def string_is_arridx(v):
    is_arridx = False
    try:
        ival = int(v)
        if ival >= 0 and ival <= 0xfffffffe and ('%d' % ival == v):
            is_arridx = True
    except ValueError:
        pass

    return is_arridx

#
#  Metadata loading, merging, and other preprocessing
#
#  Final metadata object contains merged and normalized objects and strings.
#  Keys added include (see more below):
#
#    strings_stridx: string objects which have a stridx, matches stridx index order
#    objects_bidx: objects which have a bidx, matches bidx index order
#    objects_ram_toplevel: objects which are top level for RAM init
#
#  Various helper keys are also added, containing auxiliary object/string
#  lists, lookup maps, etc.  See code below for details of these.
#

def metadata_lookup_object(meta, obj_id):
    return meta['_objid_to_object'][obj_id]

def metadata_lookup_object_and_index(meta, obj_id):
    for i,t in enumerate(meta['objects']):
        if t['id'] == obj_id:
            return t, i
    return None, None

def metadata_lookup_property(obj, key):
    for p in obj['properties']:
        if p['key'] == key:
            return p
    return None

def metadata_lookup_property_and_index(obj, key):
    for i,t in enumerate(obj['properties']):
        if t['key'] == key:
            return t, i
    return None, None

# Remove disabled objects and properties.
def metadata_remove_disabled(meta, active_opts):
    objlist = []

    count_disabled_object = 0
    count_notneeded_object = 0
    count_disabled_property = 0
    count_notneeded_property = 0

    def present_if_check(v):
        pi = v.get('present_if', None)
        if pi is None:
            return True
        if isinstance(pi, (str, unicode)):
            pi = [ pi ]
        if not isinstance(pi, list):
            raise Exception('invalid present_if syntax: %r' % pi)
        # Present if all listed options are true or unknown.
        # Absent if any option is known to be false.
        for opt in pi:
            if active_opts.get(opt, None) == False:
                return False
        return True

    for o in meta['objects']:
        if o.get('disable', False):
            logger.debug('Remove disabled object: %s' % o['id'])
            count_disabled_object += 1
        elif not present_if_check(o):
            logger.debug('Removed object not needed in active configuration: %s' % o['id'])
            count_notneeded_object += 1
        else:
            objlist.append(o)

        props = []
        for p in o['properties']:
            if p.get('disable', False):
                logger.debug('Remove disabled property: %s, object: %s' % (p['key'], o['id']))
                count_disabled_property += 1
            elif not present_if_check(p):
                logger.debug('Removed property not needed in active configuration: %s, object: %s' % (p['key'], o['id']))
                count_notneeded_property += 1
            else:
                props.append(p)

        o['properties'] = props

    meta['objects'] = objlist

    if count_disabled_object + count_notneeded_object + count_disabled_property + count_notneeded_property > 0:
        logger.info('Removed %d objects (%d disabled, %d not needed by config), %d properties (%d disabled, %d not needed by config)' % (count_disabled_object + count_notneeded_object, count_disabled_object, count_notneeded_object, count_disabled_property + count_notneeded_property, count_disabled_property, count_notneeded_property))

# Delete dangling references to removed/missing objects.
def metadata_delete_dangling_references_to_object(meta, obj_id):
    for o in meta['objects']:
        new_p = []
        for p in o['properties']:
            v = p['value']
            ptype = None
            if isinstance(v, dict):
                ptype = p['value']['type']
            delprop = False
            if ptype == 'object' and v['id'] == obj_id:
                delprop = True
            if ptype == 'accessor' and v.get('getter_id') == obj_id:
                p['getter_id'] = None
            if ptype == 'accessor' and v.get('setter_id') == obj_id:
                p['setter_id'] = None
            # XXX: Should empty accessor (= no getter, no setter) be deleted?
            # If so, beware of shorthand.
            if delprop:
                logger.debug('Deleted property %s of object %s, points to deleted object %s' % \
                             (p['key'], o['id'], obj_id))
            else:
                new_p.append(p)
        o['properties'] = new_p

# Merge a user YAML file into current metadata.
def metadata_merge_user_objects(meta, user_meta):
    if user_meta.has_key('add_objects'):
        raise Exception('"add_objects" removed, use "objects" with "add: True"')
    if user_meta.has_key('replace_objects'):
        raise Exception('"replace_objects" removed, use "objects" with "replace: True"')
    if user_meta.has_key('modify_objects'):
        raise Exception('"modify_objects" removed, use "objects" with "modify: True"')

    for o in user_meta.get('objects', []):
        if o.get('disable', False):
            logger.debug('Skip disabled object: %s' % o['id'])
            continue
        targ, targ_idx = metadata_lookup_object_and_index(meta, o['id'])

        if o.get('delete', False):
            logger.debug('Delete object: %s' % targ['id'])
            if targ is None:
                raise Exception('Cannot delete object %s which doesn\'t exist' % o['id'])
            meta['objects'].pop(targ_idx)
            metadata_delete_dangling_references_to_object(meta, targ['id'])
            continue

        if o.get('replace', False):
            logger.debug('Replace object %s' % o['id'])
            if targ is None:
                logger.warning('object to be replaced doesn\'t exist, append new object')
                meta['objects'].append(o)
            else:
                meta['objects'][targ_idx] = o
            continue

        if o.get('add', False) or not o.get('modify', False):  # 'add' is the default
            logger.debug('Add object %s' % o['id'])
            if targ is not None:
                raise Exception('Cannot add object %s which already exists' % o['id'])
            meta['objects'].append(o)
            continue

        assert(o.get('modify', False))  # modify handling
        if targ is None:
            raise Exception('Cannot modify object %s which doesn\'t exist' % o['id'])

        for k in sorted(o.keys()):
            # Merge top level keys by copying over, except 'properties'
            if k == 'properties':
                continue
            targ[k] = o[k]
        for p in o.get('properties', []):
            if p.get('disable', False):
                logger.debug('Skip disabled property: %s' % p['key'])
                continue
            prop = None
            prop_idx = None
            prop, prop_idx = metadata_lookup_property_and_index(targ, p['key'])
            if prop is not None:
                if p.get('delete', False):
                    logger.debug('Delete property %s of %s' % (p['key'], o['id']))
                    targ['properties'].pop(prop_idx)
                else:
                    logger.debug('Replace property %s of %s' % (p['key'], o['id']))
                    targ['properties'][prop_idx] = p
            else:
                if p.get('delete', False):
                    logger.debug('Deleting property %s of %s: doesn\'t exist, nop' % (p['key'], o['id']))
                else:
                    logger.debug('Add property %s of %s' % (p['key'], o['id']))
                    targ['properties'].append(p)

# Replace 'symbol' keys and values with encoded strings.
def format_symbol(sym):
    #print(repr(sym))
    assert(isinstance(sym, dict))
    assert(sym.get('type', None) == 'symbol')
    variant = sym['variant']
    if variant == 'global':
        return '\x80' + sym['string']
    elif variant == 'wellknown':
        # Well known symbols use an empty suffix which never occurs for
        # runtime local symbols.
        return '\x81' + sym['string'] + '\xff'
    elif variant == 'userhidden':
        return '\xff' + sym['string']
    elif variant == 'hidden':  # hidden == Duktape hidden Symbol
        return '\x82' + sym['string']
    raise Exception('invalid symbol variant %r' % variant)

def metadata_normalize_symbol_strings(meta):
    for o in meta['strings']:
        if isinstance(o['str'], dict) and o['str'].get('type') == 'symbol':
            o['str'] = format_symbol(o['str'])
            #print('normalized symbol as string list element: %r', o)

    for o in meta['objects']:
        for p in o['properties']:
            if isinstance(p['key'], dict) and p['key'].get('type') == 'symbol':
                p['key'] = format_symbol(p['key'])
                #print('normalized symbol as property key: %r', p)
            if isinstance(p['value'], dict) and p['value'].get('type') == 'symbol':
                p['value'] = format_symbol(p['value'])
                #print('normalized symbol as property value: %r', p)

# Normalize nargs for top level functions by defaulting 'nargs' from 'length'.
def metadata_normalize_nargs_length(meta):
    # Default 'nargs' from 'length' for top level function objects.
    for o in meta['objects']:
        if o.has_key('nargs'):
            continue
        if not o.get('callable', False):
            continue
        for p in o['properties']:
            if p['key'] != 'length':
                continue
            logger.debug('Default nargs for top level: %r' % p)
            assert(isinstance(p['value'], int))
            o['nargs'] = p['value']
            break
        assert(o.has_key('nargs'))

    # Default 'nargs' from 'length' for function property shorthand.
    for o in meta['objects']:
        for p in o['properties']:
            if not (isinstance(p['value'], dict) and p['value']['type'] == 'function'):
                continue
            pval = p['value']
            if not pval.has_key('length'):
                logger.debug('Default length for function shorthand: %r' % p)
                pval['length'] = 0
            if not pval.has_key('nargs'):
                logger.debug('Default nargs for function shorthand: %r' % p)
                pval['nargs'] = pval['length']

# Prepare a list of built-in objects which need a runtime 'bidx'.
def metadata_prepare_objects_bidx(meta):
    objlist = meta['objects']
    meta['objects'] = []
    meta['objects_bidx'] = []

    # Objects have a 'bidx: true' if they need a DUK_BIDX_xxx constant
    # and need to be present in thr->builtins[].  The list is already
    # stripped of built-in objects which are not needed based on config.
    # Ideally we'd scan the actually needed indices from the source
    # but since some usage is inside #if defined()s that's not trivial.
    for obj in objlist:
        if obj.get('bidx', False):
            obj['bidx_used'] = True
            meta['objects'].append(obj)
            meta['objects_bidx'].append(obj)

    # Append remaining objects.
    for obj in objlist:
        if obj.get('bidx_used', False):
            # Already in meta['objects'].
            pass
        else:
            meta['objects'].append(obj)

# Normalize metadata property shorthand.  For example, if a property value
# is a shorthand function, create a function object and change the property
# to point to that function object.
def metadata_normalize_shorthand(meta):
    # Gather objects through the top level built-ins list.
    objs = []
    subobjs = []

    def getSubObject():
        obj = {}
        obj['id'] = 'subobj_%d' % len(subobjs)  # synthetic ID
        obj['properties'] = []
        obj['auto_generated'] = True  # mark as autogenerated (just FYI)
        subobjs.append(obj)
        return obj

    def decodeFunctionShorthand(funprop):
        # Convert the built-in function property "shorthand" into an actual
        # object for ROM built-ins.
        assert(funprop['value']['type'] == 'function')
        val = funprop['value']
        obj = getSubObject()
        props = obj['properties']
        obj['native'] = val['native']
        obj['nargs'] = val.get('nargs', val['length'])
        obj['varargs'] = val.get('varargs', False)
        obj['magic'] = val.get('magic', 0)
        obj['internal_prototype'] = 'bi_function_prototype'
        obj['class'] = 'Function'
        obj['callable'] = val.get('callable', True)
        obj['constructable'] = val.get('constructable', False)
        obj['special_call'] = val.get('special_call', False)
        fun_name = val.get('name', funprop['key'])
        props.append({ 'key': 'length', 'value': val['length'], 'attributes': 'c' })  # Configurable in ES2015
        props.append({ 'key': 'name', 'value': fun_name, 'attributes': 'c' })   # Configurable in ES2015
        return obj

    def addAccessor(funprop, magic, nargs, length, name, native_func):
        assert(funprop['value']['type'] == 'accessor')
        obj = getSubObject()
        props = obj['properties']
        obj['native'] = native_func
        obj['nargs'] = nargs
        obj['varargs'] = False
        obj['magic'] = magic
        obj['internal_prototype'] = 'bi_function_prototype'
        obj['class'] = 'Function'
        obj['callable'] = val.get('callable', True)
        obj['constructable'] = val.get('constructable', False)
        assert(obj.get('special_call', False) == False)
        # Shorthand accessors are minimal and have no .length or .name
        # right now.  Use longhand if these matter.
        #props.append({ 'key': 'length', 'value': length, 'attributes': 'c' })
        #props.append({ 'key': 'name', 'value': name, 'attributes': 'c' })
        return obj

    def decodeGetterShorthand(key, funprop):
        assert(funprop['value']['type'] == 'accessor')
        val = funprop['value']
        if not val.has_key('getter'):
            return None
        return addAccessor(funprop,
                           val['getter_magic'],
                           val['getter_nargs'],
                           val.get('getter_length', 0),
                           key,
                           val['getter'])

    def decodeSetterShorthand(key, funprop):
        assert(funprop['value']['type'] == 'accessor')
        val = funprop['value']
        if not val.has_key('setter'):
            return None
        return addAccessor(funprop,
                           val['setter_magic'],
                           val['setter_nargs'],
                           val.get('setter_length', 0),
                           key,
                           val['setter'])

    def decodeStructuredValue(val):
        logger.debug('Decode structured value: %r' % val)
        if isinstance(val, (int, long, float, str)):
            return val  # as is
        elif isinstance(val, (dict)):
            # Object: decode recursively
            obj = decodeStructuredObject(val)
            return { 'type': 'object', 'id': obj['id'] }
        elif isinstance(val, (list)):
            raise Exception('structured shorthand does not yet support array literals')
        else:
            raise Exception('unsupported value in structured shorthand: %r' % v)

    def decodeStructuredObject(val):
        # XXX: We'd like to preserve dict order from YAML source but
        # Python doesn't do that.  Use sorted order to make the result
        # deterministic.  User can always use longhand for exact
        # property control.

        logger.debug('Decode structured object: %r' % val)
        obj = getSubObject()
        obj['class'] = 'Object'
        obj['internal_prototype'] = 'bi_object_prototype'

        props = obj['properties']
        keys = sorted(val.keys())
        for k in keys:
            logger.debug('Decode property %s' % k)
            prop = { 'key': k, 'value': decodeStructuredValue(val[k]), 'attributes': 'wec' }
            props.append(prop)

        return obj

    def decodeStructuredShorthand(structprop):
        assert(structprop['value']['type'] == 'structured')
        val = structprop['value']['value']
        return decodeStructuredValue(val)

    def clonePropShared(prop):
        res = {}
        for k in [ 'key', 'attributes', 'auto_lightfunc' ]:
            if prop.has_key(k):
                res[k] = prop[k]
        return res

    for idx,obj in enumerate(meta['objects']):
        props = []
        repl_props = []

        for val in obj['properties']:
            # Date.prototype.toGMTString must point to the same Function object
            # as Date.prototype.toUTCString, so special case hack it here.
            if obj['id'] == 'bi_date_prototype' and val['key'] == 'toGMTString':
                logger.debug('Skip Date.prototype.toGMTString')
                continue

            if isinstance(val['value'], dict) and val['value']['type'] == 'function':
                # Function shorthand.
                subfun = decodeFunctionShorthand(val)
                prop = clonePropShared(val)
                prop['value'] = { 'type': 'object', 'id': subfun['id'] }
                repl_props.append(prop)
            elif isinstance(val['value'], dict) and val['value']['type'] == 'accessor' and \
                 (val['value'].has_key('getter') or val['value'].has_key('setter')):
                # Accessor normal and shorthand forms both use the type 'accessor',
                # but are differentiated by properties.
                sub_getter = decodeGetterShorthand(val['key'], val)
                sub_setter = decodeSetterShorthand(val['key'], val)
                prop = clonePropShared(val)
                prop['value'] = { 'type': 'accessor' }
                if sub_getter is not None:
                    prop['value']['getter_id'] = sub_getter['id']
                if sub_setter is not None:
                    prop['value']['setter_id'] = sub_setter['id']
                assert('a' in prop['attributes'])  # If missing, weird things happen runtime
                logger.debug('Expand accessor shorthand: %r -> %r' % (val, prop))
                repl_props.append(prop)
            elif isinstance(val['value'], dict) and val['value']['type'] == 'structured':
                # Structured shorthand.
                subval = decodeStructuredShorthand(val)
                prop = clonePropShared(val)
                prop['value'] = subval
                repl_props.append(prop)
                logger.debug('Decoded structured shorthand for object %s, property %s' % (obj['id'], val['key']))
            elif isinstance(val['value'], dict) and val['value']['type'] == 'buffer':
                # Duktape buffer type not yet supported.
                raise Exception('Buffer type not yet supported for builtins: %r' % val)
            elif isinstance(val['value'], dict) and val['value']['type'] == 'pointer':
                # Duktape pointer type not yet supported.
                raise Exception('Pointer type not yet supported for builtins: %r' % val)
            else:
                # Property already in normalized form.
                repl_props.append(val)

            if obj['id'] == 'bi_date_prototype' and val['key'] == 'toUTCString':
                logger.debug('Clone Date.prototype.toUTCString to Date.prototype.toGMTString')
                prop2 = copy.deepcopy(repl_props[-1])
                prop2['key'] = 'toGMTString'
                repl_props.append(prop2)

        # Replace properties with a variant where function properties
        # point to built-ins rather than using an inline syntax.
        obj['properties'] = repl_props

    len_before = len(meta['objects'])
    meta['objects'] += subobjs
    len_after = len(meta['objects'])

    logger.debug('Normalized metadata shorthand, %d objects -> %d final objects' % (len_before, len_after))

# Normalize property attribute order, default attributes, etc.
def metadata_normalize_property_attributes(meta):
    for o in meta['objects']:
        for p in o['properties']:
            orig_attrs = p.get('attributes', None)
            is_accessor = (isinstance(p['value'], dict) and p['value']['type'] == 'accessor')

            # If missing, set default attributes.
            attrs = orig_attrs
            if attrs is None:
                if is_accessor:
                    attrs = 'ca'  # accessor default is configurable
                else:
                    attrs = 'wc'  # default is writable, configurable
                logger.debug('Defaulted attributes of %s/%s to %s' % (o['id'], p['key'], attrs))

            # Decode flags to normalize their order in the end.
            writable = 'w' in attrs
            enumerable = 'e' in attrs
            configurable = 'c' in attrs
            accessor = 'a' in attrs

            # Force 'accessor' attribute for accessors.
            if is_accessor and not accessor:
                logger.debug('Property %s is accessor but has no "a" attribute, add attribute' % p['key'])
                accessor = True

            # Normalize order and write back.
            attrs = ''
            if writable:
                attrs += 'w'
            if enumerable:
                attrs += 'e'
            if configurable:
                attrs += 'c'
            if accessor:
                attrs += 'a'
            p['attributes'] = attrs

            if orig_attrs != attrs:
                logger.debug('Updated attributes of %s/%s from %r to %r' % (o['id'], p['key'], orig_attrs, attrs))
                pass

# Normalize ROM property attributes.
def metadata_normalize_rom_property_attributes(meta):
    for o in meta['objects']:
        for p in o['properties']:
            # ROM properties must not be configurable (runtime code
            # depends on this).  Writability is kept so that instance
            # objects can override parent properties.
            p['attributes'] = p['attributes'].replace('c', '')

# Add a 'name' property for all top level functions; expected by RAM
# initialization code.
def metadata_normalize_ram_function_names(meta):
    num_added = 0
    for o in meta['objects']:
        if not o.get('callable', False):
            continue
        name_prop = None
        for p in o['properties']:
            if p['key'] == 'name':
                name_prop = p
                break
        if name_prop is None:
            num_added += 1
            logger.debug('Adding missing "name" property for function %s' % o['id'])
            o['properties'].append({ 'key': 'name', 'value': '', 'attributes': 'c' })

    if num_added > 0:
        logger.debug('Added missing "name" property for %d functions' % num_added)

# Add a built-in objects list for RAM initialization.
def metadata_add_ram_filtered_object_list(meta):
    # For RAM init data to support user objects, we need to prepare a
    # filtered top level object list, containing only those objects which
    # need a value stack index during duk_hthread_builtins.c init process.
    #
    # Objects in meta['objects'] which are covered by inline property
    # notation in the init data (this includes e.g. member functions like
    # Math.cos) must not be present.

    objlist = []
    for o in meta['objects']:
        keep = o.get('bidx_used', False)
        if o.has_key('native') and not o.has_key('bidx'):
            # Handled inline by run-time init code
            pass
        else:
            # Top level object
            keep = True
        if keep:
            objlist.append(o)

    logger.debug('Filtered RAM object list: %d objects with bidx, %d total top level objects' % \
          (len(meta['objects_bidx']), len(objlist)))

    meta['objects_ram_toplevel'] = objlist

# Add missing strings into strings metadata.  For example, if an object
# property key is not part of the strings list, append it there.  This
# is critical for ROM builtins because all property keys etc must also
# be in ROM.
def metadata_normalize_missing_strings(meta, user_meta):
    # We just need plain strings here.
    strs_have = {}
    for s in meta['strings']:
        strs_have[s['str']] = True

    # For ROM builtins all the strings must be in the strings list,
    # so scan objects for any strings not explicitly listed in metadata.
    for idx, obj in enumerate(meta['objects']):
        for prop in obj['properties']:
            key = prop['key']
            if not strs_have.get(key):
                logger.debug('Add missing string: %r' % key)
                meta['strings'].append({ 'str': key, '_auto_add_ref': True })
                strs_have[key] = True
            if prop.has_key('value') and isinstance(prop['value'], (str, unicode)):
                val = unicode_to_bytes(prop['value'])  # should already be, just in case
                if not strs_have.get(val):
                    logger.debug('Add missing string: %r' % val)
                    meta['strings'].append({ 'str': val, '_auto_add_ref': True })
                    strs_have[val] = True

    # Force user strings to be in ROM data.
    for s in user_meta.get('add_forced_strings', []):
        if not strs_have.get(s['str']):
            logger.debug('Add user string: %r' % s['str'])
            s['_auto_add_user'] = True
            meta['strings'].append(s)

# Convert built-in function properties into lightfuncs where applicable.
def metadata_convert_lightfuncs(meta):
    num_converted = 0
    num_skipped = 0

    for o in meta['objects']:
        for p in o['properties']:
            v = p['value']
            ptype = None
            if isinstance(v, dict):
                ptype = p['value']['type']
            if ptype != 'object':
                continue
            targ, targ_idx = metadata_lookup_object_and_index(meta, p['value']['id'])

            reasons = []
            if not targ.get('callable', False):
                reasons.append('not-callable')
            #if targ.get('constructable', False):
            #    reasons.append('constructable')

            lf_len = 0
            for p2 in targ['properties']:
                # Don't convert if function has more properties than
                # we're willing to sacrifice.
                logger.debug('   - Check %r . %s' % (o.get('id', None), p2['key']))
                if p2['key'] == 'length' and isinstance(p2['value'], (int, long)):
                    lf_len = p2['value']
                if p2['key'] not in [ 'length', 'name' ]:
                    reasons.append('nonallowed-property')

            if not p.get('auto_lightfunc', True):
                logger.debug('Automatic lightfunc conversion rejected for key %s, explicitly requested in metadata' % p['key'])
                reasons.append('no-auto-lightfunc')

            # lf_len comes from actual property table (after normalization)
            if targ.has_key('magic'):
                try:
                    # Magic values which resolve to 'bidx' indices cannot
                    # be resolved here yet, because the bidx map is not
                    # yet ready.  If so, reject the lightfunc conversion
                    # for now.  In practice this doesn't matter.
                    lf_magic = resolve_magic(targ.get('magic'), {})  # empty map is a "fake" bidx map
                    logger.debug('resolved magic ok -> %r' % lf_magic)
                except Exception, e:
                    logger.debug('Failed to resolve magic for %r: %r' % (p['key'], e))
                    reasons.append('magic-resolve-failed')
                    lf_magic = 0xffffffff  # dummy, will be out of bounds
            else:
                lf_magic = 0
            if targ.get('varargs', True):
                lf_nargs = None
                lf_varargs = True
            else:
                lf_nargs = targ['nargs']
                lf_varargs = False

            if lf_len < 0 or lf_len > 15:
                logger.debug('lf_len out of bounds: %r' % lf_len)
                reasons.append('len-bounds')
            if lf_magic < -0x80 or lf_magic > 0x7f:
                logger.debug('lf_magic out of bounds: %r' % lf_magic)
                reasons.append('magic-bounds')
            if not lf_varargs and (lf_nargs < 0 or lf_nargs > 14):
                logger.debug('lf_nargs out of bounds: %r' % lf_nargs)
                reasons.append('nargs-bounds')

            if len(reasons) > 0:
                logger.debug('Don\'t convert to lightfunc: %r %r (%r): %r' % (o.get('id', None), p.get('key', None), p['value']['id'], reasons))
                num_skipped += 1
                continue

            p_id = p['value']['id']
            p['value'] = {
                'type': 'lightfunc',
                'native': targ['native'],
                'length': lf_len,
                'magic': lf_magic,
                'nargs': lf_nargs,
                'varargs': lf_varargs
            }
            logger.debug(' - Convert to lightfunc: %r %r (%r) -> %r' % (o.get('id', None), p.get('key', None), p_id, p['value']))

            num_converted += 1

    logger.debug('Converted %d built-in function properties to lightfuncs, %d skipped as non-eligible' % (num_converted, num_skipped))

# Detect objects not reachable from any object with a 'bidx'.  This is usually
# a user error because such objects can't be reached at runtime so they're
# useless in RAM or ROM init data.
def metadata_remove_orphan_objects(meta):
    reachable = {}

    for o in meta['objects']:
        if o.get('bidx_used', False):
            reachable[o['id']] = True

    while True:
        reachable_count = len(reachable.keys())

        def _markId(obj_id):
            if obj_id is None:
                return
            reachable[obj_id] = True

        for o in meta['objects']:
            if not reachable.has_key(o['id']):
                continue
            _markId(o.get('internal_prototype', None))
            for p in o['properties']:
                # Shorthand has been normalized so no need
                # to support it here.
                v = p['value']
                ptype = None
                if isinstance(v, dict):
                    ptype = p['value']['type']
                if ptype == 'object':
                    _markId(v['id'])
                if ptype == 'accessor':
                    _markId(v.get('getter_id'))
                    _markId(v.get('setter_id'))

        logger.debug('Mark reachable: reachable count initially %d, now %d' % \
                     (reachable_count, len(reachable.keys())))
        if reachable_count == len(reachable.keys()):
            break

    num_deleted = 0
    deleted = True
    while deleted:
        deleted = False
        for i,o in enumerate(meta['objects']):
            if not reachable.has_key(o['id']):
                logger.debug('object %s not reachable, dropping' % o['id'])
                meta['objects'].pop(i)
                deleted = True
                num_deleted += 1
                break

    if num_deleted > 0:
        logger.debug('Deleted %d unreachable objects' % num_deleted)

# Add C define names for builtin strings.  These defines are added to all
# strings, even when they won't get a stridx because the define names are
# used to autodetect referenced strings.
def metadata_add_string_define_names(strlist, special_defs):
    for s in strlist:
        v = s['str']

        if special_defs.has_key(v):
            s['define'] = 'DUK_STRIDX_' + special_defs[v]
            continue

        if len(v) >= 1 and v[0] == '\x82':
            pfx = 'DUK_STRIDX_INT_'
            v = v[1:]
        else:
            pfx = 'DUK_STRIDX_'

        t = re.sub(r'([a-z0-9])([A-Z])', r'\1_\2', v)  # add underscores: aB -> a_B
        s['define'] = pfx + t.upper()

# Add a 'stridx_used' flag for strings which need a stridx.
def metadata_add_string_used_stridx(strlist, used_stridx_meta):
    defs_needed = {}
    defs_found = {}
    for s in used_stridx_meta['used_stridx_defines']:
        defs_needed[s] = True

    # strings whose define is referenced
    for s in strlist:
        if s.has_key('define') and defs_needed.has_key(s['define']):
            s['stridx_used'] = True
            defs_found[s['define']] = True

    # duk_lexer.h needs all reserved words
    for s in strlist:
        if s.get('reserved_word', False):
            s['stridx_used'] = True

    # ensure all needed defines are provided
    defs_found['DUK_STRIDX_START_RESERVED'] = True  # special defines provided automatically
    defs_found['DUK_STRIDX_START_STRICT_RESERVED'] = True
    defs_found['DUK_STRIDX_END_RESERVED'] = True
    defs_found['DUK_STRIDX_TO_TOK'] = True
    for k in sorted(defs_needed.keys()):
        if not defs_found.has_key(k):
            raise Exception('source code needs define %s not provided by strings' % repr(k))

# Merge duplicate strings in string metadata.
def metadata_merge_string_entries(strlist):
    # The raw string list may contain duplicates so merge entries.
    # The list is processed in reverse because the last entry should
    # "win" and keep its place (this matters for reserved words).

    strs = []
    str_map = {}   # plain string -> object in strs[]
    tmp = copy.deepcopy(strlist)
    tmp.reverse()
    for s in tmp:
        prev = str_map.get(s['str'])
        if prev is not None:
            for k in s.keys():
                if prev.has_key(k) and prev[k] != s[k]:
                    raise Exception('fail to merge string entry, conflicting keys: %r <-> %r' % (prev, s))
                prev[k] = s[k]
        else:
            strs.append(s)
            str_map[s['str']] = s
    strs.reverse()
    return strs

# Order builtin strings (strings with a stridx) into an order satisfying
# multiple constraints.
def metadata_order_builtin_strings(input_strlist, keyword_list, strip_unused_stridx=False):
    # Strings are ordered in the result as follows:
    #   1. Non-reserved words requiring 8-bit indices
    #   2. Non-reserved words not requiring 8-bit indices
    #   3. Reserved words in non-strict mode only
    #   4. Reserved words in strict mode
    #
    # Reserved words must follow an exact order because they are
    # translated to/from token numbers by addition/subtraction.
    # Some strings require an 8-bit index and must be in the
    # beginning.

    tmp_strs = []
    for s in copy.deepcopy(input_strlist):
        if not s.get('stridx_used', False):
            # Drop strings which are not actually needed by src-input/*.(c|h).
            # Such strings won't be in heap->strs[] or ROM legacy list.
            pass
        else:
            tmp_strs.append(s)

    # The reserved word list must match token order in duk_lexer.h
    # exactly, so pluck them out first.

    str_index = {}
    kw_index = {}
    keywords = []
    strs = []
    for idx,s in enumerate(tmp_strs):
        str_index[s['str']] = s
    for idx,s in enumerate(keyword_list):
        keywords.append(str_index[s])
        kw_index[s] = True
    for idx,s in enumerate(tmp_strs):
        if not kw_index.has_key(s['str']):
            strs.append(s)

    # Sort the strings by category number; within category keep
    # previous order.

    for idx,s in enumerate(strs):
        s['_idx'] = idx  # for ensuring stable sort

    def req8Bit(s):
        return s.get('class_name', False)   # currently just class names

    def getCat(s):
        req8 = req8Bit(s)
        if s.get('reserved_word', False):
            # XXX: unused path now, because keywords are "plucked out"
            # explicitly.
            assert(not req8)
            if s.get('future_reserved_word_strict', False):
                return 4
            else:
                return 3
        elif req8:
            return 1
        else:
            return 2

    def sortCmp(a,b):
        return cmp( (getCat(a),a['_idx']), (getCat(b),b['_idx']) )

    strs.sort(cmp=sortCmp)

    for idx,s in enumerate(strs):
        # Remove temporary _idx properties
        del s['_idx']

    for idx,s in enumerate(strs):
        if req8Bit(s) and i >= 256:
            raise Exception('8-bit string index not satisfied: ' + repr(s))

    return strs + keywords

# Dump metadata into a JSON file.
def dump_metadata(meta, fn):
    tmp = json.dumps(recursive_bytes_to_strings(meta), indent=4)
    with open(fn, 'wb') as f:
        f.write(tmp)
    logger.debug('Wrote metadata dump to %s' % fn)

# Main metadata loading function: load metadata from multiple sources,
# merge and normalize, prepare various indexes etc.
def load_metadata(opts, rom=False, build_info=None, active_opts=None):
    # Load built-in strings and objects.
    with open(opts.strings_metadata, 'rb') as f:
        strings_metadata = recursive_strings_to_bytes(yaml.load(f))
    with open(opts.objects_metadata, 'rb') as f:
        objects_metadata = recursive_strings_to_bytes(yaml.load(f))

    # Merge strings and objects metadata as simple top level key merge.
    meta = {}
    for k in objects_metadata.keys():
        meta[k] = objects_metadata[k]
    for k in strings_metadata.keys():
        meta[k] = strings_metadata[k]

    # Add user objects.
    user_meta = {}
    for fn in opts.builtin_files:
        logger.debug('Merging user builtin metadata file %s' % fn)
        with open(fn, 'rb') as f:
            user_meta = recursive_strings_to_bytes(yaml.load(f))
        metadata_merge_user_objects(meta, user_meta)

    # Remove disabled objects and properties.  Also remove objects and
    # properties which are disabled in (known) active duk_config.h.
    metadata_remove_disabled(meta, active_opts)

    # Replace Symbol keys and property values with plain (encoded) strings.
    metadata_normalize_symbol_strings(meta)

    # Normalize 'nargs' and 'length' defaults.
    metadata_normalize_nargs_length(meta)

    # Normalize property attributes.
    metadata_normalize_property_attributes(meta)

    # Normalize property shorthand into full objects.
    metadata_normalize_shorthand(meta)

    # RAM top-level functions must have a 'name'.
    if not rom:
        metadata_normalize_ram_function_names(meta)

    # Add Duktape.version and (Duktape.env for ROM case).
    for o in meta['objects']:
        if o['id'] == 'bi_duktape':
            o['properties'].insert(0, { 'key': 'version', 'value': int(build_info['duk_version']), 'attributes': '' })
            if rom:
                # Use a fixed (quite dummy for now) Duktape.env
                # when ROM builtins are in use.  In the RAM case
                # this is added during global object initialization
                # based on config options in use.
                o['properties'].insert(0, { 'key': 'env', 'value': 'ROM', 'attributes': '' })

    # Normalize property attributes (just in case shorthand handling
    # didn't add attributes to all properties).
    metadata_normalize_property_attributes(meta)

    # For ROM objects, mark all properties non-configurable.
    if rom:
        metadata_normalize_rom_property_attributes(meta)

    # Convert built-in function properties automatically into
    # lightfuncs if requested and function is eligible.
    if rom and opts.rom_auto_lightfunc:
        metadata_convert_lightfuncs(meta)

    # Create a list of objects needing a 'bidx'.  Ensure 'objects' and
    # 'objects_bidx' match in order for shared length.
    metadata_prepare_objects_bidx(meta)

    # Merge duplicate strings.
    meta['strings'] = metadata_merge_string_entries(meta['strings'])

    # Prepare an ordered list of strings with 'stridx':
    #   - Add a 'stridx_used' flag for strings which need an index in current code base
    #   - Add a C define (DUK_STRIDX_xxx) for such strings
    #   - Compute a stridx string order satisfying current runtime constraints
    #
    # The meta['strings_stridx'] result will be in proper order and stripped of
    # any strings which don't need a stridx.
    metadata_add_string_define_names(meta['strings'], meta['special_define_names'])
    with open(opts.used_stridx_metadata, 'rb') as f:
        metadata_add_string_used_stridx(meta['strings'], json.loads(f.read()))
    meta['strings_stridx'] = metadata_order_builtin_strings(meta['strings'], meta['reserved_word_token_order'])

    # For the ROM build: add any strings referenced by built-in objects
    # into the string list (not the 'stridx' list though): all strings
    # referenced by ROM objects must also be in ROM.
    if rom:
        for fn in opts.builtin_files:
            # XXX: awkward second pass
            with open(fn, 'rb') as f:
                user_meta = recursive_strings_to_bytes(yaml.load(f))
                metadata_normalize_missing_strings(meta, user_meta)
        metadata_normalize_missing_strings(meta, {})  # in case no files

    # Check for orphan objects and remove them.
    metadata_remove_orphan_objects(meta)

    # Add final stridx and bidx indices to metadata objects and strings.
    idx = 0
    for o in meta['objects']:
        if o.get('bidx_used', False):
            o['bidx'] = idx
            idx += 1
    idx = 0
    for s in meta['strings']:
        if s.get('stridx_used', False):
            s['stridx'] = idx
            idx += 1

    # Prepare a filtered RAM top level object list, needed for technical
    # reasons during RAM init handling.
    if not rom:
        metadata_add_ram_filtered_object_list(meta)

    # Sanity check: object index must match 'bidx' for all objects
    # which have a runtime 'bidx'.  This is assumed by e.g. RAM
    # thread init.
    for i,o in enumerate(meta['objects']):
        if i < len(meta['objects_bidx']):
            assert(meta['objects_bidx'][i] == meta['objects'][i])
        if o.get('bidx', False):
            assert(o['bidx'] == i)

    # Create a set of helper lists and maps now that the metadata is
    # in its final form.
    meta['_strings_plain'] = []
    meta['_strings_stridx_plain'] = []
    meta['_stridx_to_string'] = {}
    meta['_idx_to_string'] = {}
    meta['_stridx_to_plain'] = {}
    meta['_idx_to_plain'] = {}
    meta['_string_to_stridx'] = {}
    meta['_plain_to_stridx'] = {}
    meta['_string_to_idx'] = {}
    meta['_plain_to_idx'] = {}
    meta['_define_to_stridx'] = {}
    meta['_stridx_to_define'] = {}
    meta['_is_plain_reserved_word'] = {}
    meta['_is_plain_strict_reserved_word'] = {}
    meta['_objid_to_object'] = {}
    meta['_objid_to_bidx'] = {}
    meta['_objid_to_idx'] = {}
    meta['_objid_to_ramidx'] = {}
    meta['_bidx_to_objid'] = {}
    meta['_idx_to_objid'] = {}
    meta['_bidx_to_object'] = {}
    meta['_idx_to_object'] = {}

    for i,s in enumerate(meta['strings']):
        assert(s['str'] not in meta['_strings_plain'])
        meta['_strings_plain'].append(s['str'])
        if s.get('reserved_word', False):
            meta['_is_plain_reserved_word'][s['str']] = True  # includes also strict reserved words
        if s.get('future_reserved_word_strict', False):
            meta['_is_plain_strict_reserved_word'][s['str']] = True
        meta['_idx_to_string'][i] = s
        meta['_idx_to_plain'][i] = s['str']
        meta['_plain_to_idx'][s['str']] = i
        #meta['_string_to_idx'][s] = i
    for i,s in enumerate(meta['strings_stridx']):
        assert(s.get('stridx_used', False) == True)
        meta['_strings_stridx_plain'].append(s['str'])
        meta['_stridx_to_string'][i] = s
        meta['_stridx_to_plain'][i] = s['str']
        #meta['_string_to_stridx'][s] = i
        meta['_plain_to_stridx'][s['str']] = i
        meta['_define_to_stridx'][s['define']] = i
        meta['_stridx_to_define'][i] = s['define']
    for i,o in enumerate(meta['objects']):
        meta['_objid_to_object'][o['id']] = o
        meta['_objid_to_idx'][o['id']] = i
        meta['_idx_to_objid'][i] = o['id']
        meta['_idx_to_object'][i] = o
    for i,o in enumerate(meta['objects_bidx']):
        assert(o.get('bidx_used', False) == True)
        meta['_objid_to_bidx'][o['id']] = i
        assert(meta['_objid_to_bidx'][o['id']] == meta['_objid_to_idx'][o['id']])
        meta['_bidx_to_objid'][i] = o['id']
        meta['_bidx_to_object'][i] = o
    if meta.has_key('objects_ram_toplevel'):
        for i,o in enumerate(meta['objects_ram_toplevel']):
            meta['_objid_to_ramidx'][o['id']] = i

    # Dump stats.

    if rom:
        meta_name = 'ROM'
    else:
        meta_name = 'RAM'

    count_add_ref = 0
    count_add_user = 0
    for s in meta['strings']:
        if s.get('_auto_add_ref', False):
            count_add_ref += 1
        if s.get('_auto_add_user', False):
            count_add_user += 1
    count_add = count_add_ref + count_add_user

    logger.debug(('Prepared %s metadata: %d objects, %d objects with bidx, ' + \
                  '%d strings, %d strings with stridx, %d strings added ' + \
                  '(%d property key references, %d user strings)') % \
                 (meta_name, len(meta['objects']), len(meta['objects_bidx']), \
                  len(meta['strings']), len(meta['strings_stridx']), \
                  count_add, count_add_ref, count_add_user))

    return meta

#
#  Metadata helpers
#

# Magic values for Math built-in.
math_onearg_magic = {
    'fabs': 0,    # BI_MATH_FABS_IDX
    'acos': 1,    # BI_MATH_ACOS_IDX
    'asin': 2,    # BI_MATH_ASIN_IDX
    'atan': 3,    # BI_MATH_ATAN_IDX
    'ceil': 4,    # BI_MATH_CEIL_IDX
    'cos': 5,     # BI_MATH_COS_IDX
    'exp': 6,     # BI_MATH_EXP_IDX
    'floor': 7,   # BI_MATH_FLOOR_IDX
    'log': 8,     # BI_MATH_LOG_IDX
    'round': 9,   # BI_MATH_ROUND_IDX
    'sin': 10,    # BI_MATH_SIN_IDX
    'sqrt': 11,   # BI_MATH_SQRT_IDX
    'tan': 12,    # BI_MATH_TAN_IDX
    'cbrt': 13,   # BI_MATH_CBRT_IDX
    'log2': 14,   # BI_MATH_LOG2_IDX
    'log10': 15,  # BI_MATH_LOG10_IDX
    'trunc': 16,  # BI_MATH_TRUNC_IDX
}
math_twoarg_magic = {
    'atan2': 0,  # BI_MATH_ATAN2_IDX
    'pow': 1     # BI_MATH_POW_IDX
}

# Magic values for Array built-in.
array_iter_magic = {
    'every': 0,    # BI_ARRAY_ITER_EVERY
    'some': 1,     # BI_ARRAY_ITER_SOME
    'forEach': 2,  # BI_ARRAY_ITER_FOREACH
    'map': 3,      # BI_ARRAY_ITER_MAP
    'filter': 4    # BI_ARRAY_ITER_FILTER
}

# Magic value for typedarray/node.js buffer read field operations.
def magic_readfield(elem, signed=None, bigendian=None, typedarray=None):
    # Must match duk__FLD_xxx in duk_bi_buffer.c
    elemnum = {
        '8bit': 0,
        '16bit': 1,
        '32bit': 2,
        'float': 3,
        'double': 4,
        'varint': 5
    }[elem]
    if signed == True:
        signednum = 1
    elif signed == False:
        signednum = 0
    else:
        raise Exception('missing "signed"')
    if bigendian == True:
        bigendiannum = 1
    elif bigendian == False:
        bigendiannum = 0
    else:
        raise Exception('missing "bigendian"')
    if typedarray == True:
        typedarraynum = 1
    elif typedarray == False:
        typedarraynum = 0
    else:
        raise Exception('missing "typedarray"')
    return elemnum + (signednum << 4) + (bigendiannum << 3) + (typedarraynum << 5)

# Magic value for typedarray/node.js buffer write field operations.
def magic_writefield(elem, signed=None, bigendian=None, typedarray=None):
    return magic_readfield(elem, signed=signed, bigendian=bigendian, typedarray=typedarray)

# Magic value for typedarray constructors.
def magic_typedarray_constructor(elem, shift):
    # Must match duk_hbufobj.h header
    elemnum = {
        'uint8': 0,
        'uint8clamped': 1,
        'int8': 2,
        'uint16': 3,
        'int16': 4,
        'uint32': 5,
        'int32': 6,
        'float32': 7,
        'float64': 8
    }[elem]
    return (elemnum << 2) + shift

# Resolve a magic value from a YAML metadata element into an integer.
def resolve_magic(elem, objid_to_bidx):
    if elem is None:
        return 0
    if isinstance(elem, (int, long)):
        v = int(elem)
        if not (v >= -0x8000 and v <= 0x7fff):
            raise Exception('invalid plain value for magic: %s' % repr(v))
        return v
    if not isinstance(elem, dict):
        raise Exception('invalid magic: %r' % elem)

    assert(elem.has_key('type'))
    if elem['type'] == 'bidx':
        # Maps to thr->builtins[].
        v = elem['id']
        return objid_to_bidx[v]
    elif elem['type'] == 'plain':
        v = elem['value']
        if not (v >= -0x8000 and v <= 0x7fff):
            raise Exception('invalid plain value for magic: %s' % repr(v))
        return v
    elif elem['type'] == 'math_onearg':
        return math_onearg_magic[elem['funcname']]
    elif elem['type'] == 'math_twoarg':
        return math_twoarg_magic[elem['funcname']]
    elif elem['type'] == 'array_iter':
        return array_iter_magic[elem['funcname']]
    elif elem['type'] == 'typedarray_constructor':
        return magic_typedarray_constructor(elem['elem'], elem['shift'])
    elif elem['type'] == 'buffer_readfield':
        return magic_readfield(elem['elem'], elem['signed'], elem['bigendian'], elem['typedarray'])
    elif elem['type'] == 'buffer_writefield':
        return magic_writefield(elem['elem'], elem['signed'], elem['bigendian'], elem['typedarray'])
    else:
        raise Exception('invalid magic type: %r' % elem)

# Helper to find a property from a property list, remove it from the
# property list, and return the removed property.
def steal_prop(props, key, allow_accessor=True):
    for idx,prop in enumerate(props):
        if prop['key'] == key:
            if not (isinstance(prop['value'], dict) and prop['value']['type'] == 'accessor') or allow_accessor:
                return props.pop(idx)
    return None

#
#  RAM initialization data
#
#  Init data for built-in strings and objects.  The init data for both
#  strings and objects is a bit-packed stream tailored to match the decoders
#  in duk_heap_alloc.c (strings) and duk_hthread_builtins.c (objects).
#  Various bitfield sizes are used to minimize the bitstream size without
#  resorting to actual, expensive compression.  The goal is to minimize the
#  overall size of the init code and the init data.
#
#  The built-in data created here is used to set up initial RAM versions
#  of the strings and objects.  References to these objects are tracked in
#  heap->strs[] and thr->builtins[] which allows Duktape internals to refer
#  to built-ins e.g. as thr->builtins[DUK_BIDX_STRING_PROTOTYPE].
#
#  Not all strings and objects need to be reachable through heap->strs[]
#  or thr->builtins[]: the strings/objects that need to be in these arrays
#  is determined based on metadata and source code scanning.
#

# XXX: Reserved word stridxs could be made to match token numbers
#      directly so that a duk_stridx2token[] would not be needed.

# Default property attributes.
LENGTH_PROPERTY_ATTRIBUTES = 'c'
ACCESSOR_PROPERTY_ATTRIBUTES = 'c'
DEFAULT_DATA_PROPERTY_ATTRIBUTES = 'wc'

# Encoding constants (must match duk_hthread_builtins.c).
PROP_FLAGS_BITS = 3
LENGTH_PROP_BITS = 3
NARGS_BITS = 3
PROP_TYPE_BITS = 3

NARGS_VARARGS_MARKER = 0x07

PROP_TYPE_DOUBLE = 0
PROP_TYPE_STRING = 1
PROP_TYPE_STRIDX = 2
PROP_TYPE_BUILTIN = 3
PROP_TYPE_UNDEFINED = 4
PROP_TYPE_BOOLEAN_TRUE = 5
PROP_TYPE_BOOLEAN_FALSE = 6
PROP_TYPE_ACCESSOR = 7

# must match duk_hobject.h
PROPDESC_FLAG_WRITABLE =     (1 << 0)
PROPDESC_FLAG_ENUMERABLE =   (1 << 1)
PROPDESC_FLAG_CONFIGURABLE = (1 << 2)
PROPDESC_FLAG_ACCESSOR =     (1 << 3)  # unused now

# Class names, numeric indices must match duk_hobject.h class numbers.
class_names = [
    'Unused',
    'Object',
    'Array',
    'Function',
    'Arguments',
    'Boolean',
    'Date',
    'Error',
    'JSON',
    'Math',
    'Number',
    'RegExp',
    'String',
    'global',
    'Symbol',
    'ObjEnv',
    'DecEnv',
    'Pointer',
    'Thread'
    # Remaining class names are not currently needed.
]
class2num = {}
for i,v in enumerate(class_names):
    class2num[v] = i

# Map class name to a class number.
def class_to_number(x):
    return class2num[x]

# Bitpack a string into a format shared by heap and thread init data.
def bitpack_string(be, s, stats=None):
    # Strings are encoded as follows: a string begins in lowercase
    # mode and recognizes the following 5-bit symbols:
    #
    #    0-25    'a' ... 'z' or 'A' ... 'Z' depending on uppercase mode
    #    26-31   special controls, see code below

    LOOKUP1 = 26
    LOOKUP2 = 27
    SWITCH1 = 28
    SWITCH = 29
    UNUSED1 = 30
    EIGHTBIT = 31
    LOOKUP = '0123456789_ \x82\x80"{'  # special characters
    assert(len(LOOKUP) == 16)

    # support up to 256 byte strings now, cases above ~30 bytes are very
    # rare, so favor short strings in encoding
    if len(s) <= 30:
        be.bits(len(s), 5)
    else:
        be.bits(31, 5)
        be.bits(len(s), 8)

    # 5-bit character, mode specific
    mode = 'lowercase'

    for idx, c in enumerate(s):
        # This encoder is not that optimal, but good enough for now.

        islower = (ord(c) >= ord('a') and ord(c) <= ord('z'))
        isupper = (ord(c) >= ord('A') and ord(c) <= ord('Z'))
        islast = (idx == len(s) - 1)
        isnextlower = False
        isnextupper = False
        if not islast:
            c2 = s[idx+1]
            isnextlower = (ord(c2) >= ord('a') and ord(c2) <= ord('z'))
            isnextupper = (ord(c2) >= ord('A') and ord(c2) <= ord('Z'))

        # XXX: Add back special handling for hidden or other symbols?

        if islower and mode == 'lowercase':
            be.bits(ord(c) - ord('a'), 5)
            if stats is not None: stats['n_optimal'] += 1
        elif isupper and mode == 'uppercase':
            be.bits(ord(c) - ord('A'), 5)
            if stats is not None: stats['n_optimal'] += 1
        elif islower and mode == 'uppercase':
            if isnextlower:
                be.bits(SWITCH, 5)
                be.bits(ord(c) - ord('a'), 5)
                mode = 'lowercase'
                if stats is not None: stats['n_switch'] += 1
            else:
                be.bits(SWITCH1, 5)
                be.bits(ord(c) - ord('a'), 5)
                if stats is not None: stats['n_switch1'] += 1
        elif isupper and mode == 'lowercase':
            if isnextupper:
                be.bits(SWITCH, 5)
                be.bits(ord(c) - ord('A'), 5)
                mode = 'uppercase'
                if stats is not None: stats['n_switch'] += 1
            else:
                be.bits(SWITCH1, 5)
                be.bits(ord(c) - ord('A'), 5)
                if stats is not None: stats['n_switch1'] += 1
        elif c in LOOKUP:
            idx = LOOKUP.find(c)
            if idx >= 8:
                be.bits(LOOKUP2, 5)
                be.bits(idx - 8, 3)
                if stats is not None: stats['n_lookup2'] += 1
            else:
                be.bits(LOOKUP1, 5)
                be.bits(idx, 3)
                if stats is not None: stats['n_lookup1'] += 1
        elif ord(c) >= 0 and ord(c) <= 255:
            logger.debug('eightbit encoding for %d (%s)' % (ord(c), c))
            be.bits(EIGHTBIT, 5)
            be.bits(ord(c), 8)
            if stats is not None: stats['n_eightbit'] += 1
        else:
            raise Exception('internal error in bitpacking a string')

# Generate bit-packed RAM string init data.
def gen_ramstr_initdata_bitpacked(meta):
    be = dukutil.BitEncoder()

    maxlen = 0
    stats = {
        'n_optimal': 0,
        'n_lookup1': 0,
        'n_lookup2': 0,
        'n_switch1': 0,
        'n_switch': 0,
        'n_eightbit': 0
    }

    for s_obj in meta['strings_stridx']:
        s = s_obj['str']
        if len(s) > maxlen:
            maxlen = len(s)
        bitpack_string(be, s, stats)

    # end marker not necessary, C code knows length from define

    if be._varuint_count > 0:
        logger.debug('Varuint distribution:')
        logger.debug(json.dumps(be._varuint_dist[0:1024]))
        logger.debug('Varuint encoding categories: %r' % be._varuint_cats)
        logger.debug('Varuint efficiency: %f bits/value' % (float(be._varuint_bits) / float(be._varuint_count)))
    res = be.getByteString()

    logger.debug(('%d ram strings, %d bytes of string init data, %d maximum string length, ' + \
                 'encoding: optimal=%d,lookup1=%d,lookup2=%d,switch1=%d,switch=%d,eightbit=%d') % \
                 (len(meta['strings_stridx']), len(res), maxlen, \
                 stats['n_optimal'],
                 stats['n_lookup1'], stats['n_lookup2'],
                 stats['n_switch1'], stats['n_switch'],
                 stats['n_eightbit']))

    return res, maxlen

# Functions to emit string-related source/header parts.

def emit_ramstr_source_strinit_data(genc, strdata):
    genc.emitArray(strdata, 'duk_strings_data', visibility='DUK_INTERNAL', typename='duk_uint8_t', intvalues=True, const=True, size=len(strdata))

def emit_ramstr_header_strinit_defines(genc, meta, strdata, strmaxlen):
    genc.emitLine('#if !defined(DUK_SINGLE_FILE)')
    genc.emitLine('DUK_INTERNAL_DECL const duk_uint8_t duk_strings_data[%d];' % len(strdata))
    genc.emitLine('#endif  /* !DUK_SINGLE_FILE */')
    genc.emitDefine('DUK_STRDATA_MAX_STRLEN', strmaxlen)
    genc.emitDefine('DUK_STRDATA_DATA_LENGTH', len(strdata))

# This is used for both RAM and ROM strings.
def emit_header_stridx_defines(genc, meta):
    strlist = meta['strings_stridx']

    for idx,s in enumerate(strlist):
        genc.emitDefine(s['define'], idx, repr(s['str']))
        defname = s['define'].replace('_STRIDX','_HEAP_STRING')
        genc.emitDefine(defname + '(heap)', 'DUK_HEAP_GET_STRING((heap),%s)' % s['define'])
        defname = s['define'].replace('_STRIDX', '_HTHREAD_STRING')
        genc.emitDefine(defname + '(thr)', 'DUK_HTHREAD_GET_STRING((thr),%s)' % s['define'])

    idx_start_reserved = None
    idx_start_strict_reserved = None
    for idx,s in enumerate(strlist):
        if idx_start_reserved is None and s.get('reserved_word', False):
            idx_start_reserved = idx
        if idx_start_strict_reserved is None and s.get('future_reserved_word_strict', False):
            idx_start_strict_reserved = idx
    assert(idx_start_reserved is not None)
    assert(idx_start_strict_reserved is not None)

    genc.emitLine('')
    genc.emitDefine('DUK_HEAP_NUM_STRINGS', len(strlist))
    genc.emitDefine('DUK_STRIDX_START_RESERVED', idx_start_reserved)
    genc.emitDefine('DUK_STRIDX_START_STRICT_RESERVED', idx_start_strict_reserved)
    genc.emitDefine('DUK_STRIDX_END_RESERVED', len(strlist), comment='exclusive endpoint')
    genc.emitLine('')
    genc.emitLine('/* To convert a heap stridx to a token number, subtract')
    genc.emitLine(' * DUK_STRIDX_START_RESERVED and add DUK_TOK_START_RESERVED.')
    genc.emitLine(' */')

# Encode property flags for RAM initializers.
def encode_property_flags(flags):
    # Note: must match duk_hobject.h

    res = 0
    nflags = 0
    if 'w' in flags:
        nflags += 1
        res = res | PROPDESC_FLAG_WRITABLE
    if 'e' in flags:
        nflags += 1
        res = res | PROPDESC_FLAG_ENUMERABLE
    if 'c' in flags:
        nflags += 1
        res = res | PROPDESC_FLAG_CONFIGURABLE
    if 'a' in flags:
        nflags += 1
        res = res | PROPDESC_FLAG_ACCESSOR

    if nflags != len(flags):
        raise Exception('unsupported flags: %s' % repr(flags))

    return res

# Generate RAM object initdata for an object (but not its properties).
def gen_ramobj_initdata_for_object(meta, be, bi, string_to_stridx, natfunc_name_to_natidx, objid_to_bidx):
    def _stridx(strval):
        stridx = string_to_stridx[strval]
        be.varuint(stridx)
    def _stridx_or_string(strval):
        stridx = string_to_stridx.get(strval)
        if stridx is not None:
            be.varuint(stridx + 1)
        else:
            be.varuint(0)
            bitpack_string(be, strval)
    def _natidx(native_name):
        natidx = natfunc_name_to_natidx[native_name]
        be.varuint(natidx)

    class_num = class_to_number(bi['class'])
    be.varuint(class_num)

    props = [x for x in bi['properties']]  # clone

    prop_proto = steal_prop(props, 'prototype')
    prop_constr = steal_prop(props, 'constructor')
    prop_name = steal_prop(props, 'name', allow_accessor=False)
    prop_length = steal_prop(props, 'length', allow_accessor=False)

    length = -1  # default value -1 signifies varargs
    if prop_length is not None:
        assert(isinstance(prop_length['value'], int))
        length = prop_length['value']
        be.bits(1, 1)  # flag: have length
        be.bits(length, LENGTH_PROP_BITS)
    else:
        be.bits(0, 1)  # flag: no length

    # The attributes for 'length' are standard ("none") except for
    # Array.prototype.length which must be writable (this is handled
    # separately in duk_hthread_builtins.c).

    len_attrs = LENGTH_PROPERTY_ATTRIBUTES
    if prop_length is not None:
        len_attrs = prop_length['attributes']

    if len_attrs != LENGTH_PROPERTY_ATTRIBUTES:
        # Attributes are assumed to be the same, except for Array.prototype.
        if bi['class'] != 'Array':  # Array.prototype is the only one with this class
            raise Exception('non-default length attribute for unexpected object')

    # For 'Function' classed objects, emit the native function stuff.
    # Unfortunately this is more or less a copy of what we do for
    # function properties now.  This should be addressed if a rework
    # on the init format is done.

    if bi['class'] == 'Function':
        _natidx(bi['native'])

        if bi.get('varargs', False):
            be.bits(1, 1)  # flag: non-default nargs
            be.bits(NARGS_VARARGS_MARKER, NARGS_BITS)
        elif bi.has_key('nargs') and bi['nargs'] != length:
            be.bits(1, 1)  # flag: non-default nargs
            be.bits(bi['nargs'], NARGS_BITS)
        else:
            assert(length is not None)
            be.bits(0, 1)  # flag: default nargs OK

        # All Function-classed global level objects are callable
        # (have [[Call]]) but not all are constructable (have
        # [[Construct]]).  Flag that.

        assert(bi.has_key('callable'))
        assert(bi['callable'] == True)

        assert(prop_name is not None)
        assert(isinstance(prop_name['value'], str))
        _stridx_or_string(prop_name['value'])

        if bi.get('constructable', False):
            be.bits(1, 1)    # flag: constructable
        else:
            be.bits(0, 1)    # flag: not constructable
        # DUK_HOBJECT_FLAG_SPECIAL_CALL is handled at runtime without init data.

        # Convert signed magic to 16-bit unsigned for encoding
        magic = resolve_magic(bi.get('magic'), objid_to_bidx) & 0xffff
        assert(magic >= 0)
        assert(magic <= 0xffff)
        be.varuint(magic)

# Generate RAM object initdata for an object's properties.
def gen_ramobj_initdata_for_props(meta, be, bi, string_to_stridx, natfunc_name_to_natidx, objid_to_bidx, double_byte_order):
    count_normal_props = 0
    count_function_props = 0

    def _bidx(bi_id):
        be.varuint(objid_to_bidx[bi_id])
    def _bidx_or_none(bi_id):
        if bi_id is None:
            be.varuint(0)
        else:
            be.varuint(objid_to_bidx[bi_id] + 1)
    def _stridx(strval):
        stridx = string_to_stridx[strval]
        be.varuint(stridx)
    def _stridx_or_string(strval):
        stridx = string_to_stridx.get(strval)
        if stridx is not None:
            be.varuint(stridx + 1)
        else:
            be.varuint(0)
            bitpack_string(be, strval)
    def _natidx(native_name):
        if native_name is None:
            natidx = 0  # 0 is NULL in the native functions table, denotes missing function
        else:
            natidx = natfunc_name_to_natidx[native_name]
        be.varuint(natidx)
    props = [x for x in bi['properties']]  # clone

    # internal prototype: not an actual property so not in property list
    if bi.has_key('internal_prototype'):
        _bidx_or_none(bi['internal_prototype'])
    else:
        _bidx_or_none(None)

    # external prototype: encoded specially, steal from property list
    prop_proto = steal_prop(props, 'prototype')
    if prop_proto is not None:
        assert(prop_proto['value']['type'] == 'object')
        assert(prop_proto['attributes'] == '')
        _bidx_or_none(prop_proto['value']['id'])
    else:
        _bidx_or_none(None)

    # external constructor: encoded specially, steal from property list
    prop_constr = steal_prop(props, 'constructor')
    if prop_constr is not None:
        assert(prop_constr['value']['type'] == 'object')
        assert(prop_constr['attributes'] == 'wc')
        _bidx_or_none(prop_constr['value']['id'])
    else:
        _bidx_or_none(None)

    # name: encoded specially for function objects, so steal and ignore here
    if bi['class'] == 'Function':
        prop_name = steal_prop(props, 'name', allow_accessor=False)
        assert(prop_name is not None)
        assert(isinstance(prop_name['value'], str))
        assert(prop_name['attributes'] == 'c')

    # length: encoded specially, so steal and ignore
    prop_proto = steal_prop(props, 'length', allow_accessor=False)

    # Date.prototype.toGMTString needs special handling and is handled
    # directly in duk_hthread_builtins.c; so steal and ignore here.
    if bi['id'] == 'bi_date_prototype':
        prop_togmtstring = steal_prop(props, 'toGMTString')
        assert(prop_togmtstring is not None)
        logger.debug('Stole Date.prototype.toGMTString')

    # Split properties into non-builtin functions and other properties.
    # This split is a bit arbitrary, but is used to reduce flag bits in
    # the bit stream.
    values = []
    functions = []
    for prop in props:
        if isinstance(prop['value'], dict) and \
           prop['value']['type'] == 'object' and \
           metadata_lookup_object(meta, prop['value']['id']).has_key('native') and \
           not metadata_lookup_object(meta, prop['value']['id']).has_key('bidx'):
            functions.append(prop)
        else:
            values.append(prop)

    be.varuint(len(values))

    for valspec in values:
        count_normal_props += 1

        val = valspec['value']

        _stridx_or_string(valspec['key'])

        # Attribute check doesn't check for accessor flag; that is now
        # automatically set by C code when value is an accessor type.
        # Accessors must not have 'writable', so they'll always have
        # non-default attributes (less footprint than adding a different
        # default).
        default_attrs = DEFAULT_DATA_PROPERTY_ATTRIBUTES

        attrs = valspec.get('attributes', default_attrs)
        attrs = attrs.replace('a', '')  # ram bitstream doesn't encode 'accessor' attribute
        if attrs != default_attrs:
            logger.debug('non-default attributes: %s -> %r (default %r)' % (valspec['key'], attrs, default_attrs))
            be.bits(1, 1)  # flag: have custom attributes
            be.bits(encode_property_flags(attrs), PROP_FLAGS_BITS)
        else:
            be.bits(0, 1)  # flag: no custom attributes

        if val is None:
            logger.warning('RAM init data format doesn\'t support "null" now, value replaced with "undefined": %r' % valspec)
            #raise Exception('RAM init format doesn\'t support a "null" value now')
            be.bits(PROP_TYPE_UNDEFINED, PROP_TYPE_BITS)
        elif isinstance(val, bool):
            if val == True:
                be.bits(PROP_TYPE_BOOLEAN_TRUE, PROP_TYPE_BITS)
            else:
                be.bits(PROP_TYPE_BOOLEAN_FALSE, PROP_TYPE_BITS)
        elif isinstance(val, (float, int)) or isinstance(val, dict) and val['type'] == 'double':
            # Avoid converting a manually specified NaN temporarily into
            # a float to avoid risk of e.g. NaN being replaced by another.
            if isinstance(val, dict):
                val = val['bytes'].decode('hex')
                assert(len(val) == 8)
            else:
                val = struct.pack('>d', float(val))

            be.bits(PROP_TYPE_DOUBLE, PROP_TYPE_BITS)

            # encoding of double must match target architecture byte order
            indexlist = {
                'big':    [ 0, 1, 2, 3, 4, 5, 6, 7 ],
                'little': [ 7, 6, 5, 4, 3, 2, 1, 0 ],
                'mixed':  [ 3, 2, 1, 0, 7, 6, 5, 4 ]    # some arm platforms
            }[double_byte_order]

            data = ''.join([ val[indexlist[idx]] for idx in xrange(8) ])

            logger.debug('DOUBLE: %s -> %s' % (val.encode('hex'), data.encode('hex')))

            if len(data) != 8:
                raise Exception('internal error')
            be.string(data)
        elif isinstance(val, str) or isinstance(val, unicode):
            if isinstance(val, unicode):
                # Note: non-ASCII characters will not currently work,
                # because bits/char is too low.
                val = val.encode('utf-8')

            if string_to_stridx.has_key(val):
                # String value is in built-in string table -> encode
                # using a string index.  This saves some space,
                # especially for the 'name' property of errors
                # ('EvalError' etc).

                be.bits(PROP_TYPE_STRIDX, PROP_TYPE_BITS)
                _stridx(val)
            else:
                # Not in string table, bitpack string value as is.
                be.bits(PROP_TYPE_STRING, PROP_TYPE_BITS)
                bitpack_string(be, val)
        elif isinstance(val, dict):
            if val['type'] == 'object':
                be.bits(PROP_TYPE_BUILTIN, PROP_TYPE_BITS)
                _bidx(val['id'])
            elif val['type'] == 'undefined':
                be.bits(PROP_TYPE_UNDEFINED, PROP_TYPE_BITS)
            elif val['type'] == 'accessor':
                be.bits(PROP_TYPE_ACCESSOR, PROP_TYPE_BITS)
                getter_natfun = None
                setter_natfun = None
                getter_magic = 0
                setter_magic = 0
                if val.has_key('getter_id'):
                    getter_fn = metadata_lookup_object(meta, val['getter_id'])
                    getter_natfun = getter_fn['native']
                    assert(getter_fn['nargs'] == 0)
                    getter_magic = getter_fn['magic']
                if val.has_key('setter_id'):
                    setter_fn = metadata_lookup_object(meta, val['setter_id'])
                    setter_natfun = setter_fn['native']
                    assert(setter_fn['nargs'] == 1)
                    setter_magic = setter_fn['magic']
                if getter_natfun is not None and setter_natfun is not None:
                    assert(getter_magic == setter_magic)
                _natidx(getter_natfun)
                _natidx(setter_natfun)
                be.varuint(getter_magic)
            elif val['type'] == 'lightfunc':
                logger.warning('RAM init data format doesn\'t support "lightfunc" now, value replaced with "undefined": %r' % valspec)
                be.bits(PROP_TYPE_UNDEFINED, PROP_TYPE_BITS)
            else:
                raise Exception('unsupported value: %s' % repr(val))
        else:
            raise Exception('unsupported value: %s' % repr(val))

    be.varuint(len(functions))

    for funprop in functions:
        count_function_props += 1

        funobj = metadata_lookup_object(meta, funprop['value']['id'])
        prop_len = metadata_lookup_property(funobj, 'length')
        assert(prop_len is not None)
        assert(isinstance(prop_len['value'], (int)))
        length = prop_len['value']

        # XXX: this doesn't currently handle a function .name != its key
        # At least warn about it here.  Or maybe generate the correct name
        # at run time (it's systematic at the moment, @@toPrimitive has the
        # name "[Symbol.toPrimitive]" which can be computed from the symbol
        # internal representation.

        _stridx_or_string(funprop['key'])
        _natidx(funobj['native'])
        be.bits(length, LENGTH_PROP_BITS)

        if funobj.get('varargs', False):
            be.bits(1, 1)  # flag: non-default nargs
            be.bits(NARGS_VARARGS_MARKER, NARGS_BITS)
        elif funobj.has_key('nargs') and funobj['nargs'] != length:
            be.bits(1, 1)  # flag: non-default nargs
            be.bits(funobj['nargs'], NARGS_BITS)
        else:
            be.bits(0, 1)  # flag: default nargs OK

        # XXX: make this check conditional to minimize bit count
        # (there are quite a lot of function properties)
        # Convert signed magic to 16-bit unsigned for encoding
        magic = resolve_magic(funobj.get('magic'), objid_to_bidx) & 0xffff
        assert(magic >= 0)
        assert(magic <= 0xffff)
        be.varuint(magic)

    return count_normal_props, count_function_props

# Get helper maps for RAM objects.
def get_ramobj_native_func_maps(meta):
    # Native function list and index
    native_funcs_found = {}
    native_funcs = []
    natfunc_name_to_natidx = {}

    native_funcs.append(None)  # natidx 0 is reserved for NULL

    for o in meta['objects']:
        if o.has_key('native'):
            native_funcs_found[o['native']] = True
        for v in o['properties']:
            val = v['value']
            if isinstance(val, dict):
                if val['type'] == 'accessor':
                    if val.has_key('getter_id'):
                        getter = metadata_lookup_object(meta, val['getter_id'])
                        native_funcs_found[getter['native']] = True
                    if val.has_key('setter_id'):
                        setter = metadata_lookup_object(meta, val['setter_id'])
                        native_funcs_found[setter['native']] = True
                if val['type'] == 'object':
                    target = metadata_lookup_object(meta, val['id'])
                    if target.has_key('native'):
                        native_funcs_found[target['native']] = True
                if val['type'] == 'lightfunc':
                    # No lightfunc support for RAM initializer now.
                    pass

    for idx,k in enumerate(sorted(native_funcs_found.keys())):
        natfunc_name_to_natidx[k] = len(native_funcs)
        native_funcs.append(k)  # native func names

    return native_funcs, natfunc_name_to_natidx

# Generate bit-packed RAM object init data.
def gen_ramobj_initdata_bitpacked(meta, native_funcs, natfunc_name_to_natidx, double_byte_order):
    # RAM initialization is based on a specially filtered list of top
    # level objects which includes objects with 'bidx' and objects
    # which aren't handled as inline values in the init bitstream.
    objlist = meta['objects_ram_toplevel']
    objid_to_idx = meta['_objid_to_ramidx']
    objid_to_object = meta['_objid_to_object']  # This index is valid even for filtered object list
    string_index = meta['_plain_to_stridx']

    # Generate bitstream
    be = dukutil.BitEncoder()
    count_builtins = 0
    count_normal_props = 0
    count_function_props = 0
    for o in objlist:
        count_builtins += 1
        gen_ramobj_initdata_for_object(meta, be, o, string_index, natfunc_name_to_natidx, objid_to_idx)
    for o in objlist:
        count_obj_normal, count_obj_func = gen_ramobj_initdata_for_props(meta, be, o, string_index, natfunc_name_to_natidx, objid_to_idx, double_byte_order)
        count_normal_props += count_obj_normal
        count_function_props += count_obj_func

    if be._varuint_count > 0:
        logger.debug('varuint distribution:')
        logger.debug(json.dumps(be._varuint_dist[0:1024]))
        logger.debug('Varuint encoding categories: %r' % be._varuint_cats)
        logger.debug('Varuint efficiency: %f bits/value' % (float(be._varuint_bits) / float(be._varuint_count)))
    romobj_init_data = be.getByteString()
    #logger.debug(repr(romobj_init_data))
    #logger.debug(len(romobj_init_data))

    logger.debug('%d ram builtins, %d normal properties, %d function properties, %d bytes of object init data' % \
          (count_builtins, count_normal_props, count_function_props, len(romobj_init_data)))

    return romobj_init_data

# Functions to emit object-related source/header parts.

def emit_ramobj_source_nativefunc_array(genc, native_func_list):
    genc.emitLine('/* native functions: %d */' % len(native_func_list))
    genc.emitLine('DUK_INTERNAL const duk_c_function duk_bi_native_functions[%d] = {' % len(native_func_list))
    for i in native_func_list:
        # The function pointer cast here makes BCC complain about
        # "initializer too complicated", so omit the cast.
        #genc.emitLine('\t(duk_c_function) %s,' % i)
        if i is None:
            genc.emitLine('\tNULL,')
        else:
            genc.emitLine('\t%s,' % i)
    genc.emitLine('};')

def emit_ramobj_source_objinit_data(genc, init_data):
    genc.emitArray(init_data, 'duk_builtins_data', visibility='DUK_INTERNAL', typename='duk_uint8_t', intvalues=True, const=True, size=len(init_data))

def emit_ramobj_header_nativefunc_array(genc, native_func_list):
    genc.emitLine('#if !defined(DUK_SINGLE_FILE)')
    genc.emitLine('DUK_INTERNAL_DECL const duk_c_function duk_bi_native_functions[%d];' % len(native_func_list))
    genc.emitLine('#endif  /* !DUK_SINGLE_FILE */')

def emit_ramobj_header_objects(genc, meta):
    objlist = meta['objects_bidx']
    for idx,o in enumerate(objlist):
        defname = 'DUK_BIDX_' + '_'.join(o['id'].upper().split('_')[1:])  # bi_foo_bar -> FOO_BAR
        genc.emitDefine(defname, idx)
    genc.emitDefine('DUK_NUM_BUILTINS', len(objlist))
    genc.emitDefine('DUK_NUM_BIDX_BUILTINS', len(objlist))                      # Objects with 'bidx'
    genc.emitDefine('DUK_NUM_ALL_BUILTINS', len(meta['objects_ram_toplevel']))  # Objects with 'bidx' + temps needed in init

def emit_ramobj_header_initdata(genc, init_data):
    genc.emitLine('#if !defined(DUK_SINGLE_FILE)')
    genc.emitLine('DUK_INTERNAL_DECL const duk_uint8_t duk_builtins_data[%d];' % len(init_data))
    genc.emitLine('#endif  /* !DUK_SINGLE_FILE */')
    genc.emitDefine('DUK_BUILTINS_DATA_LENGTH', len(init_data))

#
#  ROM init data
#
#  Compile-time initializers for ROM strings and ROM objects.  This involves
#  a lot of small details:
#
#    - Several variants are needed for different options: unpacked vs.
#      packed duk_tval, endianness, string hash in use, etc).
#
#    - Static initializers must represent objects of different size.  For
#      example, separate structs are needed for property tables of different
#      size or value typing.
#
#    - Union initializers cannot be used portable because they're only
#      available in C99 and above.
#
#    - Initializers must use 'const' correctly to ensure that the entire
#      initialization data will go into ROM (read-only data section).
#      Const pointers etc will need to be cast into non-const pointers at
#      some point to properly mix with non-const RAM pointers, so a portable
#      const losing cast is needed.
#
#    - C++ doesn't allow forward declaration of "static const" structures
#      which is problematic because there are cyclical const structures.
#

# Get string hash initializers; need to compute possible string hash variants
# which will match runtime values.
def rom_get_strhash16_macro(val):
    hash16le = dukutil.duk_heap_hashstring_dense(val, DUK__FIXED_HASH_SEED, big_endian=False, strhash16=True)
    hash16be = dukutil.duk_heap_hashstring_dense(val, DUK__FIXED_HASH_SEED, big_endian=True, strhash16=True)
    hash16sparse = dukutil.duk_heap_hashstring_sparse(val, DUK__FIXED_HASH_SEED, strhash16=True)
    return 'DUK__STRHASH16(%dU,%dU,%dU)' % (hash16le, hash16be, hash16sparse)
def rom_get_strhash32_macro(val):
    hash32le = dukutil.duk_heap_hashstring_dense(val, DUK__FIXED_HASH_SEED, big_endian=False, strhash16=False)
    hash32be = dukutil.duk_heap_hashstring_dense(val, DUK__FIXED_HASH_SEED, big_endian=True, strhash16=False)
    hash32sparse = dukutil.duk_heap_hashstring_sparse(val, DUK__FIXED_HASH_SEED, strhash16=False)
    return 'DUK__STRHASH32(%dUL,%dUL,%dUL)' % (hash32le, hash32be, hash32sparse)

# Get string character .length; must match runtime .length computation.
def rom_charlen(x):
    return dukutil.duk_unicode_unvalidated_utf8_length(x)

# Get an initializer type and initializer literal for a specified value
# (expressed in YAML metadata format).  The types and initializers depend
# on declarations emitted before the initializers, and in several cases
# use a macro to hide the selection between several initializer variants.
def rom_get_value_initializer(meta, val, bi_str_map, bi_obj_map):
    def double_bytes_initializer(val):
        # Portable and exact float initializer.
        assert(isinstance(val, str) and len(val) == 16)  # hex encoded bytes
        val = val.decode('hex')
        tmp = []
        for i in xrange(8):
            t = ord(val[i])
            if t >= 128:
                tmp.append('%dU' % t)
            else:
                tmp.append('%d' % t)
        return 'DUK__DBLBYTES(' + ','.join(tmp) + ')'

    def tval_number_initializer(val):
        return 'DUK__TVAL_NUMBER(%s)' % double_bytes_initializer(val)

    v = val['value']
    if v is None:
        init_type = 'duk_rom_tval_null'
        init_lit = 'DUK__TVAL_NULL()'
    elif isinstance(v, (bool)):
        init_type = 'duk_rom_tval_boolean'
        bval = 0
        if v:
            bval = 1
        init_lit = 'DUK__TVAL_BOOLEAN(%d)' % bval
    elif isinstance(v, (int, float)):
        fval = struct.pack('>d', float(v)).encode('hex')
        init_type = 'duk_rom_tval_number'
        init_lit = tval_number_initializer(fval)
    elif isinstance(v, (str, unicode)):
        init_type = 'duk_rom_tval_string'
        init_lit = 'DUK__TVAL_STRING(&%s)' % bi_str_map[v]
    elif isinstance(v, (dict)):
        if v['type'] == 'double':
            init_type = 'duk_rom_tval_number'
            init_lit = tval_number_initializer(v['bytes'])
        elif v['type'] == 'undefined':
            init_type = 'duk_rom_tval_undefined'
            init_lit = 'DUK__TVAL_UNDEFINED()'
        elif v['type'] == 'null':
            init_type = 'duk_rom_tval_null'
            init_lit = 'DUK__TVAL_UNDEFINED()'
        elif v['type'] == 'object':
            init_type = 'duk_rom_tval_object'
            init_lit = 'DUK__TVAL_OBJECT(&%s)' % bi_obj_map[v['id']]
        elif v['type'] == 'accessor':
            getter_ref = 'NULL'
            setter_ref = 'NULL'
            if v.has_key('getter_id'):
                getter_object = metadata_lookup_object(meta, v['getter_id'])
                getter_ref = '&%s' % bi_obj_map[getter_object['id']]
            if v.has_key('setter_id'):
                setter_object = metadata_lookup_object(meta, v['setter_id'])
                setter_ref = '&%s' % bi_obj_map[setter_object['id']]
            init_type = 'duk_rom_tval_accessor'
            init_lit = 'DUK__TVAL_ACCESSOR(%s, %s)' % (getter_ref, setter_ref)
        elif v['type'] == 'lightfunc':
            # Match DUK_LFUNC_FLAGS_PACK() in duk_tval.h.
            if v.has_key('length'):
                assert(v['length'] >= 0 and v['length'] <= 15)
                lf_length = v['length']
            else:
                lf_length = 0
            if v.get('varargs', True):
                lf_nargs = 15  # varargs marker
            else:
                assert(v['nargs'] >= 0 and v['nargs'] <= 14)
                lf_nargs = v['nargs']
            if v.has_key('magic'):
                assert(v['magic'] >= -0x80 and v['magic'] <= 0x7f)
                lf_magic = v['magic'] & 0xff
            else:
                lf_magic = 0
            lf_flags = (lf_magic << 8) + (lf_length << 4) + lf_nargs
            init_type = 'duk_rom_tval_lightfunc'
            init_lit = 'DUK__TVAL_LIGHTFUNC(%s, %dL)' % (v['native'], lf_flags)
        else:
            raise Exception('unhandled value: %r' % val)
    else:
        raise Exception('internal error: %r' % val)
    return init_type, init_lit

# Helpers to get either initializer type or value only (not both).
def rom_get_value_initializer_type(meta, val, bi_str_map, bi_obj_map):
    init_type, init_lit = rom_get_value_initializer(meta, val, bi_str_map, bi_obj_map)
    return init_type
def rom_get_value_initializer_literal(meta, val, bi_str_map, bi_obj_map):
    init_type, init_lit = rom_get_value_initializer(meta, val, bi_str_map, bi_obj_map)
    return init_lit

# Emit ROM strings source: structs/typedefs and their initializers.
# Separate initialization structs are needed for strings of different
# length.
def rom_emit_strings_source(genc, meta):
    # Write built-in strings as code section initializers.

    strs = meta['_strings_plain']  # all strings, plain versions
    reserved_words = meta['_is_plain_reserved_word']
    strict_reserved_words = meta['_is_plain_strict_reserved_word']
    strs_needing_stridx = meta['strings_stridx']

    # Sort used lengths and declare per-length initializers.
    lens = []
    for v in strs:
        strlen = len(v)
        if strlen not in lens:
            lens.append(strlen)
    lens.sort()
    for strlen in lens:
        genc.emitLine('typedef struct duk_romstr_%d duk_romstr_%d; ' % (strlen, strlen) +
                      'struct duk_romstr_%d { duk_hstring hdr; duk_uint8_t data[%d]; };' % (strlen, strlen + 1))
    genc.emitLine('')

    # String hash values depend on endianness and other factors,
    # use an initializer macro to select the appropriate hash.
    genc.emitLine('/* When unaligned access possible, 32-bit values are fetched using host order.')
    genc.emitLine(' * When unaligned access not possible, always simulate little endian order.')
    genc.emitLine(' * See: src-input/duk_util_hashbytes.c:duk_util_hashbytes().')
    genc.emitLine(' */')
    genc.emitLine('#if defined(DUK_USE_STRHASH_DENSE)')
    genc.emitLine('#if defined(DUK_USE_HASHBYTES_UNALIGNED_U32_ACCESS)')  # XXX: config option to be reworked
    genc.emitLine('#if defined(DUK_USE_INTEGER_BE)')
    genc.emitLine('#define DUK__STRHASH16(hash16le,hash16be,hash16sparse) (hash16be)')
    genc.emitLine('#define DUK__STRHASH32(hash32le,hash32be,hash32sparse) (hash32be)')
    genc.emitLine('#else')
    genc.emitLine('#define DUK__STRHASH16(hash16le,hash16be,hash16sparse) (hash16le)')
    genc.emitLine('#define DUK__STRHASH32(hash32le,hash32be,hash32sparse) (hash32le)')
    genc.emitLine('#endif')
    genc.emitLine('#else')
    genc.emitLine('#define DUK__STRHASH16(hash16le,hash16be,hash16sparse) (hash16le)')
    genc.emitLine('#define DUK__STRHASH32(hash32le,hash32be,hash32sparse) (hash32le)')
    genc.emitLine('#endif')
    genc.emitLine('#else  /* DUK_USE_STRHASH_DENSE */')
    genc.emitLine('#define DUK__STRHASH16(hash16le,hash16be,hash16sparse) (hash16sparse)')
    genc.emitLine('#define DUK__STRHASH32(hash32le,hash32be,hash32sparse) (hash32sparse)')
    genc.emitLine('#endif  /* DUK_USE_STRHASH_DENSE */')

    # String header initializer macro, takes into account lowmem etc.
    genc.emitLine('#if defined(DUK_USE_HEAPPTR16)')
    genc.emitLine('#if !defined(DUK_USE_REFCOUNT16)')
    genc.emitLine('#error currently assumes DUK_USE_HEAPPTR16 and DUK_USE_REFCOUNT16 are both defined')
    genc.emitLine('#endif')
    genc.emitLine('#if defined(DUK_USE_HSTRING_CLEN)')
    genc.emitLine('#define DUK__STRINIT(heaphdr_flags,refcount,hash32,hash16,blen,clen,next) \\')
    genc.emitLine('\t{ { (heaphdr_flags) | ((hash16) << 16), DUK__REFCINIT((refcount)), (blen), (duk_hstring *) DUK_LOSE_CONST((next)) }, (clen) }')
    genc.emitLine('#else  /* DUK_USE_HSTRING_CLEN */')
    genc.emitLine('#define DUK__STRINIT(heaphdr_flags,refcount,hash32,hash16,blen,clen,next) \\')
    genc.emitLine('\t{ { (heaphdr_flags) | ((hash16) << 16), DUK__REFCINIT((refcount)), (blen), (duk_hstring *) DUK_LOSE_CONST((next)) } }')
    genc.emitLine('#endif  /* DUK_USE_HSTRING_CLEN */')
    genc.emitLine('#else  /* DUK_USE_HEAPPTR16 */')
    genc.emitLine('#define DUK__STRINIT(heaphdr_flags,refcount,hash32,hash16,blen,clen,next) \\')
    genc.emitLine('\t{ { (heaphdr_flags), DUK__REFCINIT((refcount)), DUK_LOSE_CONST((next)) }, (hash32), (blen), (clen) }')
    genc.emitLine('#endif  /* DUK_USE_HEAPPTR16 */')

    # Organize ROM strings into a chained ROM string table.  The ROM string
    # h_next link pointer is used for chaining just like for RAM strings but
    # in a separate string table.
    #
    # To avoid dealing with the different possible string hash algorithms,
    # use a much more trivial lookup key for ROM strings for now.
    romstr_hash = []
    while len(romstr_hash) < ROMSTR_LOOKUP_SIZE:
        romstr_hash.append([])
    for str_index,v in enumerate(strs):
        if len(v) > 0:
            rom_lookup_hash = ord(v[0]) + (len(v) << 4)
        else:
            rom_lookup_hash = 0 + (len(v) << 4)
        rom_lookup_hash = rom_lookup_hash & 0xff
        romstr_hash[rom_lookup_hash].append(v)

    romstr_next = {}   # string -> the string's 'next' link
    for lst in romstr_hash:
        prev = None
        #print(repr(lst))
        for v in lst:
            if prev is not None:
                romstr_next[prev] = v
            prev = v

    chain_lens = {}
    for lst in romstr_hash:
        chainlen = len(lst)
        if not chain_lens.has_key(chainlen):
            chain_lens[chainlen] = 0
        chain_lens[chainlen] += 1
    tmp = []
    for k in sorted(chain_lens.keys()):
        tmp.append('%d: %d' % (k, chain_lens[k]))
    logger.info('ROM string table chain lengths: %s' % ', '.join(tmp))

    bi_str_map = {}   # string -> initializer variable name
    for str_index,v in enumerate(strs):
        bi_str_map[v] = 'duk_str_%d' % str_index

    # Emit string initializers.  Emit the strings in an order which avoids
    # forward declarations for the h_next link pointers; const forward
    # declarations are a problem in C++.
    genc.emitLine('')
    for lst in romstr_hash:
        for v in reversed(lst):
            tmp = 'DUK_INTERNAL const duk_romstr_%d %s = {' % (len(v), bi_str_map[v])
            flags = [ 'DUK_HTYPE_STRING',
                      'DUK_HEAPHDR_FLAG_READONLY',
                      'DUK_HEAPHDR_FLAG_REACHABLE',
                      'DUK_HSTRING_FLAG_PINNED_LITERAL' ]
            is_arridx = string_is_arridx(v)

            blen = len(v)
            clen = rom_charlen(v)

            if blen == clen:
                flags.append('DUK_HSTRING_FLAG_ASCII')
            if is_arridx:
                flags.append('DUK_HSTRING_FLAG_ARRIDX')
            if len(v) >= 1 and v[0] in [ '\x80', '\x81', '\x82', '\xff' ]:
                flags.append('DUK_HSTRING_FLAG_SYMBOL')
            if len(v) >= 1 and v[0] in [ '\x82', '\xff' ]:
                flags.append('DUK_HSTRING_FLAG_HIDDEN')
            if v in [ 'eval', 'arguments' ]:
                flags.append('DUK_HSTRING_FLAG_EVAL_OR_ARGUMENTS')
            if reserved_words.has_key(v):
                flags.append('DUK_HSTRING_FLAG_RESERVED_WORD')
            if strict_reserved_words.has_key(v):
                flags.append('DUK_HSTRING_FLAG_STRICT_RESERVED_WORD')

            h_next = 'NULL'
            if romstr_next.has_key(v):
                h_next = '&' + bi_str_map[romstr_next[v]]

            tmp += 'DUK__STRINIT(%s,%d,%s,%s,%d,%d,%s),' % \
                ('|'.join(flags), 1, rom_get_strhash32_macro(v), \
                 rom_get_strhash16_macro(v), blen, clen, h_next)

            tmpbytes = []
            for c in v:
                if ord(c) < 128:
                    tmpbytes.append('%d' % ord(c))
                else:
                    tmpbytes.append('%dU' % ord(c))
            tmpbytes.append('%d' % 0)  # NUL term
            tmp += '{' + ','.join(tmpbytes) + '}'
            tmp += '};'
            genc.emitLine(tmp)

    # Emit the ROM string lookup table used by string interning.
    #
    # cdecl> explain const int * const foo;
    # declare foo as const pointer to const int
    genc.emitLine('')
    genc.emitLine('DUK_INTERNAL const duk_hstring * const duk_rom_strings_lookup[%d] = {'% len(romstr_hash))
    tmp = []
    linecount = 0
    for lst in romstr_hash:
        if len(lst) == 0:
            genc.emitLine('\tNULL,')
        else:
            genc.emitLine('\t(const duk_hstring *) &%s,' % bi_str_map[lst[0]])
    genc.emitLine('};')

    # Emit an array of duk_hstring pointers indexed using DUK_STRIDX_xxx.
    # This will back e.g. DUK_HTHREAD_STRING_XYZ(thr) directly, without
    # needing an explicit array in thr/heap->strs[].
    #
    # cdecl > explain const int * const foo;
    # declare foo as const pointer to const int
    genc.emitLine('')
    genc.emitLine('DUK_INTERNAL const duk_hstring * const duk_rom_strings_stridx[%d] = {' % len(strs_needing_stridx))
    for s in strs_needing_stridx:
        genc.emitLine('\t(const duk_hstring *) &%s,' % bi_str_map[s['str']])  # strs_needing_stridx is a list of objects, not plain strings
    genc.emitLine('};')

    return bi_str_map

# Emit ROM strings header.
def rom_emit_strings_header(genc, meta):
    genc.emitLine('#if !defined(DUK_SINGLE_FILE)')  # C++ static const workaround
    genc.emitLine('DUK_INTERNAL_DECL const duk_hstring * const duk_rom_strings_lookup[%d];' % ROMSTR_LOOKUP_SIZE)
    genc.emitLine('DUK_INTERNAL_DECL const duk_hstring * const duk_rom_strings_stridx[%d];' % len(meta['strings_stridx']))
    genc.emitLine('#endif')

# Emit ROM objects initialized types and macros.
def rom_emit_object_initializer_types_and_macros(genc):
    # Objects and functions are straightforward because they just use the
    # RAM structure which has no dynamic or variable size parts.
    genc.emitLine('typedef struct duk_romobj duk_romobj; ' + \
                  'struct duk_romobj { duk_hobject hdr; };')
    genc.emitLine('typedef struct duk_romarr duk_romarr; ' + \
                  'struct duk_romarr { duk_harray hdr; };')
    genc.emitLine('typedef struct duk_romfun duk_romfun; ' + \
                  'struct duk_romfun { duk_hnatfunc hdr; };')
    genc.emitLine('typedef struct duk_romobjenv duk_romobjenv; ' + \
                  'struct duk_romobjenv { duk_hobjenv hdr; };')

    # For ROM pointer compression we'd need a -compile time- variant.
    # The current portable solution is to just assign running numbers
    # to ROM compressed pointers, and provide the table for user pointer
    # compression function.  Much better solutions would be possible,
    # but such solutions are often compiler/platform specific.

    # Emit object/function initializer which is aware of options affecting
    # the header.  Heap next/prev pointers are always NULL.
    genc.emitLine('#if defined(DUK_USE_HEAPPTR16)')
    genc.emitLine('#if !defined(DUK_USE_REFCOUNT16) || defined(DUK_USE_HOBJECT_HASH_PART)')
    genc.emitLine('#error currently assumes DUK_USE_HEAPPTR16 and DUK_USE_REFCOUNT16 are both defined and DUK_USE_HOBJECT_HASH_PART is undefined')
    genc.emitLine('#endif')
    #genc.emitLine('#if !defined(DUK_USE_HEAPPTR_ENC16_STATIC)')
    #genc.emitLine('#error need DUK_USE_HEAPPTR_ENC16_STATIC which provides compile-time pointer compression')
    #genc.emitLine('#endif')
    genc.emitLine('#define DUK__ROMOBJ_INIT(heaphdr_flags,refcount,props,props_enc16,iproto,iproto_enc16,esize,enext,asize,hsize) \\')
    genc.emitLine('\t{ { { (heaphdr_flags), DUK__REFCINIT((refcount)), 0, 0, (props_enc16) }, (iproto_enc16), (esize), (enext), (asize) } }')
    genc.emitLine('#define DUK__ROMARR_INIT(heaphdr_flags,refcount,props,props_enc16,iproto,iproto_enc16,esize,enext,asize,hsize,length) \\')
    genc.emitLine('\t{ { { { (heaphdr_flags), DUK__REFCINIT((refcount)), 0, 0, (props_enc16) }, (iproto_enc16), (esize), (enext), (asize) }, (length), 0 /*length_nonwritable*/ } }')
    genc.emitLine('#define DUK__ROMFUN_INIT(heaphdr_flags,refcount,props,props_enc16,iproto,iproto_enc16,esize,enext,asize,hsize,nativefunc,nargs,magic) \\')
    genc.emitLine('\t{ { { { (heaphdr_flags), DUK__REFCINIT((refcount)), 0, 0, (props_enc16) }, (iproto_enc16), (esize), (enext), (asize) }, (nativefunc), (duk_int16_t) (nargs), (duk_int16_t) (magic) } }')
    genc.emitLine('#define DUK__ROMOBJENV_INIT(heaphdr_flags,refcount,props,props_enc16,iproto,iproto_enc16,esize,enext,asize,hsize,target,has_this) \\')
    genc.emitLine('\t{ { { { (heaphdr_flags), DUK__REFCINIT((refcount)), 0, 0, (props_enc16) }, (iproto_enc16), (esize), (enext), (asize) }, (duk_hobject *) DUK_LOSE_CONST(target), (has_this) } }')
    genc.emitLine('#else  /* DUK_USE_HEAPPTR16 */')
    genc.emitLine('#define DUK__ROMOBJ_INIT(heaphdr_flags,refcount,props,props_enc16,iproto,iproto_enc16,esize,enext,asize,hsize) \\')
    genc.emitLine('\t{ { { (heaphdr_flags), DUK__REFCINIT((refcount)), NULL, NULL }, (duk_uint8_t *) DUK_LOSE_CONST(props), (duk_hobject *) DUK_LOSE_CONST(iproto), (esize), (enext), (asize), (hsize) } }')
    genc.emitLine('#define DUK__ROMARR_INIT(heaphdr_flags,refcount,props,props_enc16,iproto,iproto_enc16,esize,enext,asize,hsize,length) \\')
    genc.emitLine('\t{ { { { (heaphdr_flags), DUK__REFCINIT((refcount)), NULL, NULL }, (duk_uint8_t *) DUK_LOSE_CONST(props), (duk_hobject *) DUK_LOSE_CONST(iproto), (esize), (enext), (asize), (hsize) }, (length), 0 /*length_nonwritable*/ } }')
    genc.emitLine('#define DUK__ROMFUN_INIT(heaphdr_flags,refcount,props,props_enc16,iproto,iproto_enc16,esize,enext,asize,hsize,nativefunc,nargs,magic) \\')
    genc.emitLine('\t{ { { { (heaphdr_flags), DUK__REFCINIT((refcount)), NULL, NULL }, (duk_uint8_t *) DUK_LOSE_CONST(props), (duk_hobject *) DUK_LOSE_CONST(iproto), (esize), (enext), (asize), (hsize) }, (nativefunc), (duk_int16_t) (nargs), (duk_int16_t) (magic) } }')
    genc.emitLine('#define DUK__ROMOBJENV_INIT(heaphdr_flags,refcount,props,props_enc16,iproto,iproto_enc16,esize,enext,asize,hsize,target,has_this) \\')
    genc.emitLine('\t{ { { { (heaphdr_flags), DUK__REFCINIT((refcount)), NULL, NULL }, (duk_uint8_t *) DUK_LOSE_CONST(props), (duk_hobject *) DUK_LOSE_CONST(iproto), (esize), (enext), (asize), (hsize) }, (duk_hobject *) DUK_LOSE_CONST(target), (has_this) } }')
    genc.emitLine('#endif  /* DUK_USE_HEAPPTR16 */')

    # Initializer typedef for a dummy function pointer.  ROM support assumes
    # function pointers are 32 bits.  Using a dummy function pointer type
    # avoids function pointer to normal pointer cast which emits warnings.
    genc.emitLine('typedef void (*duk_rom_funcptr)(void);')

    # Emit duk_tval structs.  This gets a bit messier with packed/unpacked
    # duk_tval, endianness variants, pointer sizes, etc.
    genc.emitLine('#if defined(DUK_USE_PACKED_TVAL)')
    genc.emitLine('typedef struct duk_rom_tval_undefined duk_rom_tval_undefined;')
    genc.emitLine('typedef struct duk_rom_tval_null duk_rom_tval_null;')
    genc.emitLine('typedef struct duk_rom_tval_lightfunc duk_rom_tval_lightfunc;')
    genc.emitLine('typedef struct duk_rom_tval_boolean duk_rom_tval_boolean;')
    genc.emitLine('typedef struct duk_rom_tval_number duk_rom_tval_number;')
    genc.emitLine('typedef struct duk_rom_tval_object duk_rom_tval_object;')
    genc.emitLine('typedef struct duk_rom_tval_string duk_rom_tval_string;')
    genc.emitLine('typedef struct duk_rom_tval_accessor duk_rom_tval_accessor;')
    genc.emitLine('struct duk_rom_tval_number { duk_uint8_t bytes[8]; };')
    genc.emitLine('struct duk_rom_tval_accessor { const duk_hobject *get; const duk_hobject *set; };')
    genc.emitLine('#if defined(DUK_USE_DOUBLE_LE)')
    genc.emitLine('struct duk_rom_tval_object { const void *ptr; duk_uint32_t hiword; };')
    genc.emitLine('struct duk_rom_tval_string { const void *ptr; duk_uint32_t hiword; };')
    genc.emitLine('struct duk_rom_tval_undefined { const void *ptr; duk_uint32_t hiword; };')
    genc.emitLine('struct duk_rom_tval_null { const void *ptr; duk_uint32_t hiword; };')
    genc.emitLine('struct duk_rom_tval_lightfunc { duk_rom_funcptr ptr; duk_uint32_t hiword; };')
    genc.emitLine('struct duk_rom_tval_boolean { duk_uint32_t dummy; duk_uint32_t hiword; };')
    genc.emitLine('#elif defined(DUK_USE_DOUBLE_BE)')
    genc.emitLine('struct duk_rom_tval_object { duk_uint32_t hiword; const void *ptr; };')
    genc.emitLine('struct duk_rom_tval_string { duk_uint32_t hiword; const void *ptr; };')
    genc.emitLine('struct duk_rom_tval_undefined { duk_uint32_t hiword; const void *ptr; };')
    genc.emitLine('struct duk_rom_tval_null { duk_uint32_t hiword; const void *ptr; };')
    genc.emitLine('struct duk_rom_tval_lightfunc { duk_uint32_t hiword; duk_rom_funcptr ptr; };')
    genc.emitLine('struct duk_rom_tval_boolean { duk_uint32_t hiword; duk_uint32_t dummy; };')
    genc.emitLine('#elif defined(DUK_USE_DOUBLE_ME)')
    genc.emitLine('struct duk_rom_tval_object { duk_uint32_t hiword; const void *ptr; };')
    genc.emitLine('struct duk_rom_tval_string { duk_uint32_t hiword; const void *ptr; };')
    genc.emitLine('struct duk_rom_tval_undefined { duk_uint32_t hiword; const void *ptr; };')
    genc.emitLine('struct duk_rom_tval_null { duk_uint32_t hiword; const void *ptr; };')
    genc.emitLine('struct duk_rom_tval_lightfunc { duk_uint32_t hiword; duk_rom_funcptr ptr; };')
    genc.emitLine('struct duk_rom_tval_boolean { duk_uint32_t hiword; duk_uint32_t dummy; };')
    genc.emitLine('#else')
    genc.emitLine('#error invalid endianness defines')
    genc.emitLine('#endif')
    genc.emitLine('#else  /* DUK_USE_PACKED_TVAL */')
    # Unpacked initializers are written assuming normal struct alignment
    # rules so that sizeof(duk_tval) == 16.  32-bit pointers need special
    # handling to ensure the individual initializers pad to 16 bytes as
    # necessary.
    # XXX: 32-bit unpacked duk_tval is not yet supported.
    genc.emitLine('#if defined(DUK_UINTPTR_MAX)')
    genc.emitLine('#if (DUK_UINTPTR_MAX <= 0xffffffffUL)')
    genc.emitLine('#error ROM initializer with unpacked duk_tval does not currently work on 32-bit targets')
    genc.emitLine('#endif')
    genc.emitLine('#endif')
    genc.emitLine('typedef struct duk_rom_tval_undefined duk_rom_tval_undefined;')
    genc.emitLine('struct duk_rom_tval_undefined { duk_small_uint_t tag; duk_small_uint_t extra; duk_uint8_t bytes[8]; };')
    genc.emitLine('typedef struct duk_rom_tval_null duk_rom_tval_null;')
    genc.emitLine('struct duk_rom_tval_null { duk_small_uint_t tag; duk_small_uint_t extra; duk_uint8_t bytes[8]; };')
    genc.emitLine('typedef struct duk_rom_tval_boolean duk_rom_tval_boolean;')
    genc.emitLine('struct duk_rom_tval_boolean { duk_small_uint_t tag; duk_small_uint_t extra; duk_uint32_t val; duk_uint32_t unused; };')
    genc.emitLine('typedef struct duk_rom_tval_number duk_rom_tval_number;')
    genc.emitLine('struct duk_rom_tval_number { duk_small_uint_t tag; duk_small_uint_t extra; duk_uint8_t bytes[8]; };')
    genc.emitLine('typedef struct duk_rom_tval_object duk_rom_tval_object;')
    genc.emitLine('struct duk_rom_tval_object { duk_small_uint_t tag; duk_small_uint_t extra; const duk_heaphdr *val; };')
    genc.emitLine('typedef struct duk_rom_tval_string duk_rom_tval_string;')
    genc.emitLine('struct duk_rom_tval_string { duk_small_uint_t tag; duk_small_uint_t extra; const duk_heaphdr *val; };')
    genc.emitLine('typedef struct duk_rom_tval_lightfunc duk_rom_tval_lightfunc;')
    genc.emitLine('struct duk_rom_tval_lightfunc { duk_small_uint_t tag; duk_small_uint_t extra; duk_rom_funcptr ptr; };')
    genc.emitLine('typedef struct duk_rom_tval_accessor duk_rom_tval_accessor;')
    genc.emitLine('struct duk_rom_tval_accessor { const duk_hobject *get; const duk_hobject *set; };')
    genc.emitLine('#endif  /* DUK_USE_PACKED_TVAL */')
    genc.emitLine('')

    # Double initializer byte shuffle macro to handle byte orders
    # without duplicating the entire initializers.
    genc.emitLine('#if defined(DUK_USE_DOUBLE_LE)')
    genc.emitLine('#define DUK__DBLBYTES(a,b,c,d,e,f,g,h) { (h), (g), (f), (e), (d), (c), (b), (a) }')
    genc.emitLine('#elif defined(DUK_USE_DOUBLE_BE)')
    genc.emitLine('#define DUK__DBLBYTES(a,b,c,d,e,f,g,h) { (a), (b), (c), (d), (e), (f), (g), (h) }')
    genc.emitLine('#elif defined(DUK_USE_DOUBLE_ME)')
    genc.emitLine('#define DUK__DBLBYTES(a,b,c,d,e,f,g,h) { (d), (c), (b), (a), (h), (g), (f), (e) }')
    genc.emitLine('#else')
    genc.emitLine('#error invalid endianness defines')
    genc.emitLine('#endif')
    genc.emitLine('')

    # Emit duk_tval initializer literal macros.
    genc.emitLine('#if defined(DUK_USE_PACKED_TVAL)')
    genc.emitLine('#define DUK__TVAL_NUMBER(hostbytes) { hostbytes }')  # bytes already in host order
    genc.emitLine('#if defined(DUK_USE_DOUBLE_LE)')
    genc.emitLine('#define DUK__TVAL_UNDEFINED() { (const void *) NULL, (DUK_TAG_UNDEFINED << 16) }')
    genc.emitLine('#define DUK__TVAL_NULL() { (const void *) NULL, (DUK_TAG_NULL << 16) }')
    genc.emitLine('#define DUK__TVAL_LIGHTFUNC(func,flags) { (duk_rom_funcptr) (func), (DUK_TAG_LIGHTFUNC << 16) + (flags) }')
    genc.emitLine('#define DUK__TVAL_BOOLEAN(bval) { 0, (DUK_TAG_BOOLEAN << 16) + (bval) }')
    genc.emitLine('#define DUK__TVAL_OBJECT(ptr) { (const void *) (ptr), (DUK_TAG_OBJECT << 16) }')
    genc.emitLine('#define DUK__TVAL_STRING(ptr) { (const void *) (ptr), (DUK_TAG_STRING << 16) }')
    genc.emitLine('#elif defined(DUK_USE_DOUBLE_BE)')
    genc.emitLine('#define DUK__TVAL_UNDEFINED() { (DUK_TAG_UNDEFINED << 16), (const void *) NULL }')
    genc.emitLine('#define DUK__TVAL_NULL() { (DUK_TAG_NULL << 16), (const void *) NULL }')
    genc.emitLine('#define DUK__TVAL_LIGHTFUNC(func,flags) { (DUK_TAG_LIGHTFUNC << 16) + (flags), (duk_rom_funcptr) (func) }')
    genc.emitLine('#define DUK__TVAL_BOOLEAN(bval) { (DUK_TAG_BOOLEAN << 16) + (bval), 0 }')
    genc.emitLine('#define DUK__TVAL_OBJECT(ptr) { (DUK_TAG_OBJECT << 16), (const void *) (ptr) }')
    genc.emitLine('#define DUK__TVAL_STRING(ptr) { (DUK_TAG_STRING << 16), (const void *) (ptr) }')
    genc.emitLine('#elif defined(DUK_USE_DOUBLE_ME)')
    genc.emitLine('#define DUK__TVAL_UNDEFINED() { (DUK_TAG_UNDEFINED << 16), (const void *) NULL }')
    genc.emitLine('#define DUK__TVAL_NULL() { (DUK_TAG_NULL << 16), (const void *) NULL }')
    genc.emitLine('#define DUK__TVAL_LIGHTFUNC(func,flags) { (DUK_TAG_LIGHTFUNC << 16) + (flags), (duk_rom_funcptr) (func) }')
    genc.emitLine('#define DUK__TVAL_BOOLEAN(bval) { (DUK_TAG_BOOLEAN << 16) + (bval), 0 }')
    genc.emitLine('#define DUK__TVAL_OBJECT(ptr) { (DUK_TAG_OBJECT << 16), (const void *) (ptr) }')
    genc.emitLine('#define DUK__TVAL_STRING(ptr) { (DUK_TAG_STRING << 16), (const void *) (ptr) }')
    genc.emitLine('#else')
    genc.emitLine('#error invalid endianness defines')
    genc.emitLine('#endif')
    genc.emitLine('#else  /* DUK_USE_PACKED_TVAL */')
    genc.emitLine('#define DUK__TVAL_NUMBER(hostbytes) { DUK_TAG_NUMBER, 0, hostbytes }')  # bytes already in host order
    genc.emitLine('#define DUK__TVAL_UNDEFINED() { DUK_TAG_UNDEFINED, 0, {0,0,0,0,0,0,0,0} }')
    genc.emitLine('#define DUK__TVAL_NULL() { DUK_TAG_NULL, 0, {0,0,0,0,0,0,0,0} }')
    genc.emitLine('#define DUK__TVAL_BOOLEAN(bval) { DUK_TAG_BOOLEAN, 0, (bval), 0 }')
    genc.emitLine('#define DUK__TVAL_OBJECT(ptr) { DUK_TAG_OBJECT, 0, (const duk_heaphdr *) (ptr) }')
    genc.emitLine('#define DUK__TVAL_STRING(ptr) { DUK_TAG_STRING, 0, (const duk_heaphdr *) (ptr) }')
    genc.emitLine('#define DUK__TVAL_LIGHTFUNC(func,flags) { DUK_TAG_LIGHTFUNC, (flags), (duk_rom_funcptr) (func) }')
    genc.emitLine('#endif  /* DUK_USE_PACKED_TVAL */')
    genc.emitLine('#define DUK__TVAL_ACCESSOR(getter,setter) { (const duk_hobject *) (getter), (const duk_hobject *) (setter) }')

# Emit ROM objects source: the object/function headers themselves, property
# table structs for different property table sizes/types, and property table
# initializers.
def rom_emit_objects(genc, meta, bi_str_map):
    objs = meta['objects']
    id_to_bidx = meta['_objid_to_bidx']

    # Table for compressed ROM pointers; reserve high range of compressed pointer
    # values for this purpose.  This must contain all ROM pointers that might be
    # referenced (all objects, strings, and property tables at least).
    romptr_compress_list = []
    def compress_rom_ptr(x):
        if x == 'NULL':
            return 0
        try:
            idx = romptr_compress_list.index(x)
            res = ROMPTR_FIRST + idx
        except ValueError:
            romptr_compress_list.append(x)
            res = ROMPTR_FIRST + len(romptr_compress_list) - 1
        assert(res <= 0xffff)
        return res

    # Need string and object maps (id -> C symbol name) early.
    bi_obj_map = {}   # object id -> initializer variable name
    for idx,obj in enumerate(objs):
        bi_obj_map[obj['id']] = 'duk_obj_%d' % idx

    # Add built-in strings and objects to compressed ROM pointers first.
    for k in sorted(bi_str_map.keys()):
        compress_rom_ptr('&%s' % bi_str_map[k])
    for k in sorted(bi_obj_map.keys()):
        compress_rom_ptr('&%s' % bi_obj_map[k])

    # Property attributes lookup, map metadata attribute string into a
    # C initializer.
    attr_lookup = {
        '':    'DUK_PROPDESC_FLAGS_NONE',
        'w':    'DUK_PROPDESC_FLAGS_W',
        'e':    'DUK_PROPDESC_FLAGS_E',
        'c':    'DUK_PROPDESC_FLAGS_C',
        'we':    'DUK_PROPDESC_FLAGS_WE',
        'wc':    'DUK_PROPDESC_FLAGS_WC',
        'ec':    'DUK_PROPDESC_FLAGS_EC',
        'wec':    'DUK_PROPDESC_FLAGS_WEC',
        'a':    'DUK_PROPDESC_FLAGS_NONE|DUK_PROPDESC_FLAG_ACCESSOR',
        'ea':    'DUK_PROPDESC_FLAGS_E|DUK_PROPDESC_FLAG_ACCESSOR',
        'ca':    'DUK_PROPDESC_FLAGS_C|DUK_PROPDESC_FLAG_ACCESSOR',
        'eca':    'DUK_PROPDESC_FLAGS_EC|DUK_PROPDESC_FLAG_ACCESSOR',
    }

    # Emit property table structs.  These are very complex because
    # property count *and* individual property type affect the fields
    # in the initializer, properties can be data properties or accessor
    # properties or different duk_tval types.  There are also several
    # property table memory layouts, each with a different ordering of
    # keys, values, etc.  Union initializers would make things a bit
    # easier but they're not very portable (being C99).
    #
    # The easy solution is to use a separate initializer type for each
    # property type.  Could also cache and reuse identical initializers
    # but there'd be very few of them so it's more straightforward to
    # not reuse the structs.
    #
    # NOTE: naming is a bit inconsistent here, duk_tval is used also
    # to refer to property value initializers like a getter/setter pair.

    genc.emitLine('#if defined(DUK_USE_HOBJECT_LAYOUT_1)')
    for idx,obj in enumerate(objs):
        numprops = len(obj['properties'])
        if numprops == 0:
            continue
        tmp = 'typedef struct duk_romprops_%d duk_romprops_%d; ' % (idx, idx)
        tmp += 'struct duk_romprops_%d { ' % idx
        for idx,val in enumerate(obj['properties']):
            tmp += 'const duk_hstring *key%d; ' % idx
        for idx,val in enumerate(obj['properties']):
            # XXX: fastint support
            tmp += '%s val%d; ' % (rom_get_value_initializer_type(meta, val, bi_str_map, bi_obj_map), idx)
        for idx,val in enumerate(obj['properties']):
            tmp += 'duk_uint8_t flags%d; ' % idx
        tmp += '};'
        genc.emitLine(tmp)
    genc.emitLine('#elif defined(DUK_USE_HOBJECT_LAYOUT_2)')
    for idx,obj in enumerate(objs):
        numprops = len(obj['properties'])
        if numprops == 0:
            continue
        tmp = 'typedef struct duk_romprops_%d duk_romprops_%d; ' % (idx, idx)
        tmp += 'struct duk_romprops_%d { ' % idx
        for idx,val in enumerate(obj['properties']):
            # XXX: fastint support
            tmp += '%s val%d; ' % (rom_get_value_initializer_type(meta, val, bi_str_map, bi_obj_map), idx)
        for idx,val in enumerate(obj['properties']):
            tmp += 'const duk_hstring *key%d; ' % idx
        for idx,val in enumerate(obj['properties']):
            tmp += 'duk_uint8_t flags%d; ' % idx
        # Padding follows for flags, but we don't need to emit it
        # (at the moment there is never an array or hash part).
        tmp += '};'
        genc.emitLine(tmp)
    genc.emitLine('#elif defined(DUK_USE_HOBJECT_LAYOUT_3)')
    for idx,obj in enumerate(objs):
        numprops = len(obj['properties'])
        if numprops == 0:
            continue
        tmp = 'typedef struct duk_romprops_%d duk_romprops_%d; ' % (idx, idx)
        tmp += 'struct duk_romprops_%d { ' % idx
        for idx,val in enumerate(obj['properties']):
            # XXX: fastint support
            tmp += '%s val%d; ' % (rom_get_value_initializer_type(meta, val, bi_str_map, bi_obj_map), idx)
        # No array values
        for idx,val in enumerate(obj['properties']):
            tmp += 'const duk_hstring *key%d; ' % idx
        # No hash index
        for idx,val in enumerate(obj['properties']):
            tmp += 'duk_uint8_t flags%d; ' % idx
        tmp += '};'
        genc.emitLine(tmp)
    genc.emitLine('#else')
    genc.emitLine('#error invalid object layout')
    genc.emitLine('#endif')
    genc.emitLine('')

    # Forward declare all property tables so that objects can reference them.
    # Also pointer compress them.

    for idx,obj in enumerate(objs):
        numprops = len(obj['properties'])
        if numprops == 0:
            continue

        # We would like to use DUK_INTERNAL_DECL here, but that maps
        # to "static const" in a single file build which has C++
        # portability issues: you can't forward declare a static const.
        # We can't reorder the property tables to avoid this because
        # there are cyclic references.  So, as the current workaround,
        # declare as external.
        genc.emitLine('DUK_EXTERNAL_DECL const duk_romprops_%d duk_prop_%d;' % (idx, idx))

        # Add property tables to ROM compressed pointers too.
        compress_rom_ptr('&duk_prop_%d' % idx)
    genc.emitLine('')

    # Forward declare all objects so that objects can reference them,
    # e.g. internal prototype reference.

    for idx,obj in enumerate(objs):
        # Careful with C++: must avoid redefining a non-extern const.
        # See commentary above for duk_prop_%d forward declarations.
        if obj.get('callable', False):
            genc.emitLine('DUK_EXTERNAL_DECL const duk_romfun duk_obj_%d;' % idx)
        elif obj.get('class') == 'Array':
            genc.emitLine('DUK_EXTERNAL_DECL const duk_romarr duk_obj_%d;' % idx)
        elif obj.get('class') == 'ObjEnv':
            genc.emitLine('DUK_EXTERNAL_DECL const duk_romobjenv duk_obj_%d;' % idx)
        else:
            genc.emitLine('DUK_EXTERNAL_DECL const duk_romobj duk_obj_%d;' % idx)
    genc.emitLine('')

    # Define objects, reference property tables.  Objects will be
    # logically non-extensible so also leave their extensible flag
    # cleared despite what metadata requests; the runtime code expects
    # ROM objects to be non-extensible.
    for idx,obj in enumerate(objs):
        numprops = len(obj['properties'])

        isfunc = obj.get('callable', False)

        if isfunc:
            tmp = 'DUK_EXTERNAL const duk_romfun duk_obj_%d = ' % idx
        elif obj.get('class') == 'Array':
            tmp = 'DUK_EXTERNAL const duk_romarr duk_obj_%d = ' % idx
        elif obj.get('class') == 'ObjEnv':
            tmp = 'DUK_EXTERNAL const duk_romobjenv duk_obj_%d = ' % idx
        else:
            tmp = 'DUK_EXTERNAL const duk_romobj duk_obj_%d = ' % idx

        flags = [ 'DUK_HTYPE_OBJECT', 'DUK_HEAPHDR_FLAG_READONLY', 'DUK_HEAPHDR_FLAG_REACHABLE' ]
        if isfunc:
            flags.append('DUK_HOBJECT_FLAG_NATFUNC')
            flags.append('DUK_HOBJECT_FLAG_STRICT')
            flags.append('DUK_HOBJECT_FLAG_NEWENV')
        if obj.get('callable', False):
            flags.append('DUK_HOBJECT_FLAG_CALLABLE')
        if obj.get('constructable', False):
            flags.append('DUK_HOBJECT_FLAG_CONSTRUCTABLE')
        if obj.get('class') == 'Array':
            flags.append('DUK_HOBJECT_FLAG_EXOTIC_ARRAY')
        if obj.get('special_call', False):
            flags.append('DUK_HOBJECT_FLAG_SPECIAL_CALL')
        flags.append('DUK_HOBJECT_CLASS_AS_FLAGS(%d)' % class_to_number(obj['class']))  # XXX: use constant, not number

        refcount = 1  # refcount is faked to be always 1
        if numprops == 0:
            props = 'NULL'
        else:
            props = '&duk_prop_%d' % idx
        props_enc16 = compress_rom_ptr(props)

        if obj.has_key('internal_prototype'):
            iproto = '&%s' % bi_obj_map[obj['internal_prototype']]
        else:
            iproto = 'NULL'
        iproto_enc16 = compress_rom_ptr(iproto)

        e_size = numprops
        e_next = e_size
        a_size = 0  # never an array part for now
        h_size = 0  # never a hash for now; not appropriate for perf relevant builds

        if isfunc:
            nativefunc = obj['native']
            if obj.get('varargs', False):
                nargs = 'DUK_VARARGS'
            elif obj.has_key('nargs'):
                nargs = '%d' % obj['nargs']
            else:
                assert(False)  # 'nargs' should be defaulted from 'length' at metadata load
            magic = '%d' % resolve_magic(obj.get('magic', None), id_to_bidx)
        else:
            nativefunc = 'dummy'
            nargs = '0'
            magic = '0'

        assert(a_size == 0)
        assert(h_size == 0)
        if isfunc:
            tmp += 'DUK__ROMFUN_INIT(%s,%d,%s,%d,%s,%d,%d,%d,%d,%d,%s,%s,%s);' % \
                ('|'.join(flags), refcount, props, props_enc16, \
                 iproto, iproto_enc16, e_size, e_next, a_size, h_size, \
                 nativefunc, nargs, magic)
        elif obj.get('class') == 'Array':
            arrlen = 0
            tmp += 'DUK__ROMARR_INIT(%s,%d,%s,%d,%s,%d,%d,%d,%d,%d,%d);' % \
                ('|'.join(flags), refcount, props, props_enc16, \
                 iproto, iproto_enc16, e_size, e_next, a_size, h_size, arrlen)
        elif obj.get('class') == 'ObjEnv':
            objenv_target = '&%s' % bi_obj_map[obj['objenv_target']]
            objenv_has_this = obj['objenv_has_this']
            tmp += 'DUK__ROMOBJENV_INIT(%s,%d,%s,%d,%s,%d,%d,%d,%d,%d,%s,%d);' % \
                ('|'.join(flags), refcount, props, props_enc16, \
                 iproto, iproto_enc16, e_size, e_next, a_size, h_size, objenv_target, objenv_has_this)
        else:
            tmp += 'DUK__ROMOBJ_INIT(%s,%d,%s,%d,%s,%d,%d,%d,%d,%d);' % \
                ('|'.join(flags), refcount, props, props_enc16, \
                 iproto, iproto_enc16, e_size, e_next, a_size, h_size)

        genc.emitLine(tmp)

    # Property tables.  Can reference arbitrary strings and objects as
    # they're defined before them.

    # Properties will be non-configurable, but must be writable so that
    # standard property semantics allow shadowing properties to be
    # established in inherited objects (e.g. "var obj={}; obj.toString
    # = myToString").  Enumerable can also be kept.

    def _prepAttrs(val):
        attrs = val['attributes']
        assert('c' not in attrs)
        return attr_lookup[attrs]

    def _emitPropTableInitializer(idx, obj, layout):
        init_vals = []
        init_keys = []
        init_flags = []

        numprops = len(obj['properties'])
        for val in obj['properties']:
            init_keys.append('(const duk_hstring *)&%s' % bi_str_map[val['key']])
        for val in obj['properties']:
            # XXX: fastint support
            init_vals.append('%s' % rom_get_value_initializer_literal(meta, val, bi_str_map, bi_obj_map))
        for val in obj['properties']:
            init_flags.append('%s' % _prepAttrs(val))

        if layout == 1:
            initlist = init_keys + init_vals + init_flags
        elif layout == 2:
            initlist = init_vals + init_keys + init_flags
        elif layout == 3:
            # Same as layout 2 now, no hash/array
            initlist = init_vals + init_keys + init_flags

        if len(initlist) > 0:
            genc.emitLine('DUK_EXTERNAL const duk_romprops_%d duk_prop_%d = {%s};' % (idx, idx, ','.join(initlist)))

    genc.emitLine('#if defined(DUK_USE_HOBJECT_LAYOUT_1)')
    for idx,obj in enumerate(objs):
        _emitPropTableInitializer(idx, obj, 1)
    genc.emitLine('#elif defined(DUK_USE_HOBJECT_LAYOUT_2)')
    for idx,obj in enumerate(objs):
        _emitPropTableInitializer(idx, obj, 2)
    genc.emitLine('#elif defined(DUK_USE_HOBJECT_LAYOUT_3)')
    for idx,obj in enumerate(objs):
        _emitPropTableInitializer(idx, obj, 3)
    genc.emitLine('#else')
    genc.emitLine('#error invalid object layout')
    genc.emitLine('#endif')
    genc.emitLine('')

    # Emit a list of ROM builtins (those objects needing a bidx).
    #
    # cdecl > explain const int * const foo;
    # declare foo as const pointer to const int

    count_bidx = 0
    for bi in objs:
        if bi.get('bidx_used', False):
            count_bidx += 1
    genc.emitLine('DUK_INTERNAL const duk_hobject * const duk_rom_builtins_bidx[%d] = {' % count_bidx)
    for bi in objs:
        if not bi.get('bidx_used', False):
            continue  # for this we want the toplevel objects only
        genc.emitLine('\t(const duk_hobject *) &%s,' % bi_obj_map[bi['id']])
    genc.emitLine('};')

    # Emit a table of compressed ROM pointers.  We must be able to
    # compress ROM pointers at compile time so we assign running
    # indices to them.  User pointer compression macros must use this
    # array to encode/decode ROM pointers.

    genc.emitLine('')
    genc.emitLine('#if defined(DUK_USE_ROM_OBJECTS) && defined(DUK_USE_HEAPPTR16)')
    genc.emitLine('DUK_EXTERNAL const void * const duk_rom_compressed_pointers[%d] = {' % (len(romptr_compress_list) + 1))
    for idx,ptr in enumerate(romptr_compress_list):
        genc.emitLine('\t(const void *) %s,  /* 0x%04x */' % (ptr, ROMPTR_FIRST + idx))
    romptr_highest = ROMPTR_FIRST + len(romptr_compress_list) - 1
    genc.emitLine('\tNULL')  # for convenience
    genc.emitLine('};')
    genc.emitLine('#endif')

    logger.debug('%d compressed rom pointers (used range is [0x%04x,0x%04x], %d space left)' % \
                 (len(romptr_compress_list), ROMPTR_FIRST, romptr_highest, 0xffff - romptr_highest))

    # Undefine helpers.
    genc.emitLine('')
    for i in [
        'DUK__STRHASH16',
        'DUK__STRHASH32',
        'DUK__DBLBYTES',
        'DUK__TVAL_NUMBER',
        'DUK__TVAL_UNDEFINED',
        'DUK__TVAL_NULL',
        'DUK__TVAL_BOOLEAN',
        'DUK__TVAL_OBJECT',
        'DUK__TVAL_STRING',
        'DUK__STRINIT',
        'DUK__ROMOBJ_INIT',
        'DUK__ROMFUN_INIT'
    ]:
        genc.emitLine('#undef ' + i)

    return romptr_compress_list

# Emit ROM objects header.
def rom_emit_objects_header(genc, meta):
    bidx = 0
    for bi in meta['objects']:
        if not bi.get('bidx_used', False):
            continue  # for this we want the toplevel objects only
        genc.emitDefine('DUK_BIDX_' + '_'.join(bi['id'].upper().split('_')[1:]), bidx)  # bi_foo_bar -> FOO_BAR
        bidx += 1
    count_bidx = bidx
    genc.emitDefine('DUK_NUM_BUILTINS', count_bidx)
    genc.emitDefine('DUK_NUM_BIDX_BUILTINS', count_bidx)
    genc.emitDefine('DUK_NUM_ALL_BUILTINS', len(meta['objects']))
    genc.emitLine('')
    genc.emitLine('#if !defined(DUK_SINGLE_FILE)')  # C++ static const workaround
    genc.emitLine('DUK_INTERNAL_DECL const duk_hobject * const duk_rom_builtins_bidx[%d];' % count_bidx)
    genc.emitLine('#endif')

    # XXX: missing declarations here, not an issue for single source build.
    # Add missing declarations.
    # XXX: For example, 'DUK_EXTERNAL_DECL ... duk_rom_compressed_pointers[]' is missing.

#
#  Shared for both RAM and ROM
#

def emit_header_native_function_declarations(genc, meta):
    emitted = {}  # To suppress duplicates
    funclist = []
    def _emit(fname):
        if not emitted.has_key(fname):
            emitted[fname] = True
            funclist.append(fname)

    for o in meta['objects']:
        if o.has_key('native'):
            _emit(o['native'])

        for p in o['properties']:
            v = p['value']
            if isinstance(v, dict) and v['type'] == 'lightfunc':
                assert(v.has_key('native'))
                _emit(v['native'])
                logger.debug('Lightfunc function declaration: %r' % v['native'])

    for fname in funclist:
        # Visibility depends on whether the function is Duktape internal or user.
        # Use a simple prefix for now.
        if fname[:4] == 'duk_':
            genc.emitLine('DUK_INTERNAL_DECL duk_ret_t %s(duk_context *ctx);' % fname)
        else:
            genc.emitLine('extern duk_ret_t %s(duk_context *ctx);' % fname)

#
#  Main
#

def main():
    parser = optparse.OptionParser()
    parser.add_option('--git-commit', dest='git_commit', default=None, help='Git commit hash')
    parser.add_option('--git-describe', dest='git_describe', default=None, help='Git describe')
    parser.add_option('--git-branch', dest='git_branch', default=None, help='Git branch name')
    parser.add_option('--duk-version', dest='duk_version', default=None, help='Duktape version (e.g. 10203)')
    parser.add_option('--quiet', dest='quiet', action='store_true', default=False, help='Suppress info messages (show warnings)')
    parser.add_option('--verbose', dest='verbose', action='store_true', default=False, help='Show verbose debug messages')
    parser.add_option('--used-stridx-metadata', dest='used_stridx_metadata', help='DUK_STRIDX_xxx used by source/headers, JSON format')
    parser.add_option('--strings-metadata', dest='strings_metadata', help='Default built-in strings metadata file, YAML format')
    parser.add_option('--objects-metadata', dest='objects_metadata', help='Default built-in objects metadata file, YAML format')
    parser.add_option('--active-options', dest='active_options', help='Active config options from genconfig.py, JSON format')
    parser.add_option('--user-builtin-metadata', dest='obsolete_builtin_metadata', default=None, help=optparse.SUPPRESS_HELP)
    parser.add_option('--builtin-file', dest='builtin_files', metavar='FILENAME', action='append', default=[], help='Built-in string/object YAML metadata to be applied over default built-ins (multiple files may be given, applied in sequence)')
    parser.add_option('--ram-support', dest='ram_support', action='store_true', default=False, help='Support RAM strings/objects')
    parser.add_option('--rom-support', dest='rom_support', action='store_true', default=False, help='Support ROM strings/objects (increases output size considerably)')
    parser.add_option('--rom-auto-lightfunc', dest='rom_auto_lightfunc', action='store_true', default=False, help='Convert ROM built-in function properties into lightfuncs automatically whenever possible')
    parser.add_option('--out-header', dest='out_header', help='Output header file')
    parser.add_option('--out-source', dest='out_source', help='Output source file')
    parser.add_option('--out-metadata-json', dest='out_metadata_json', help='Output metadata file')
    parser.add_option('--dev-dump-final-ram-metadata', dest='dev_dump_final_ram_metadata', help='Development option')
    parser.add_option('--dev-dump-final-rom-metadata', dest='dev_dump_final_rom_metadata', help='Development option')
    (opts, args) = parser.parse_args()

    if opts.obsolete_builtin_metadata is not None:
        raise Exception('--user-builtin-metadata has been removed, use --builtin-file instead')

    # Log level.
    if opts.quiet:
        logger.setLevel(logging.WARNING)
    elif opts.verbose:
        logger.setLevel(logging.DEBUG)

    # Options processing.

    build_info = {
        'git_commit': opts.git_commit,
        'git_branch': opts.git_branch,
        'git_describe': opts.git_describe,
        'duk_version': int(opts.duk_version),
    }

    desc = []
    if opts.ram_support:
        desc += [ 'ram built-in support' ]
    if opts.rom_support:
        desc += [ 'rom built-in support' ]
    if opts.rom_auto_lightfunc:
        desc += [ 'rom auto lightfunc' ]
    logger.info('Creating built-in initialization data: ' + ', '.join(desc))

    # Read in metadata files, normalizing and merging as necessary.

    active_opts = {}
    if opts.active_options is not None:
        with open(opts.active_options, 'rb') as f:
            active_opts = json.loads(f.read())

    ram_meta = load_metadata(opts, rom=False, build_info=build_info, active_opts=active_opts)
    rom_meta = load_metadata(opts, rom=True, build_info=build_info, active_opts=active_opts)
    if opts.dev_dump_final_ram_metadata is not None:
        dump_metadata(ram_meta, opts.dev_dump_final_ram_metadata)
    if opts.dev_dump_final_rom_metadata is not None:
        dump_metadata(rom_meta, opts.dev_dump_final_rom_metadata)

    # Create RAM init data bitstreams.

    ramstr_data, ramstr_maxlen = gen_ramstr_initdata_bitpacked(ram_meta)
    ram_native_funcs, ram_natfunc_name_to_natidx = get_ramobj_native_func_maps(ram_meta)

    if opts.ram_support:
        ramobj_data_le = gen_ramobj_initdata_bitpacked(ram_meta, ram_native_funcs, ram_natfunc_name_to_natidx, 'little')
        ramobj_data_be = gen_ramobj_initdata_bitpacked(ram_meta, ram_native_funcs, ram_natfunc_name_to_natidx, 'big')
        ramobj_data_me = gen_ramobj_initdata_bitpacked(ram_meta, ram_native_funcs, ram_natfunc_name_to_natidx, 'mixed')

    # Write source and header files.

    gc_src = dukutil.GenerateC()
    gc_src.emitHeader('genbuiltins.py')
    gc_src.emitLine('#include "duk_internal.h"')
    gc_src.emitLine('')
    gc_src.emitLine('#if defined(DUK_USE_ASSERTIONS)')
    gc_src.emitLine('#define DUK__REFCINIT(refc) 0 /*h_assert_refcount*/, (refc) /*actual*/')
    gc_src.emitLine('#else')
    gc_src.emitLine('#define DUK__REFCINIT(refc) (refc) /*actual*/')
    gc_src.emitLine('#endif')
    gc_src.emitLine('')
    gc_src.emitLine('#if defined(DUK_USE_ROM_STRINGS)')
    if opts.rom_support:
        rom_bi_str_map = rom_emit_strings_source(gc_src, rom_meta)
        rom_emit_object_initializer_types_and_macros(gc_src)
        rom_emit_objects(gc_src, rom_meta, rom_bi_str_map)
    else:
        gc_src.emitLine('#error ROM support not enabled, rerun configure.py with --rom-support')
    gc_src.emitLine('#else  /* DUK_USE_ROM_STRINGS */')
    emit_ramstr_source_strinit_data(gc_src, ramstr_data)
    gc_src.emitLine('#endif  /* DUK_USE_ROM_STRINGS */')
    gc_src.emitLine('')
    gc_src.emitLine('#if defined(DUK_USE_ROM_OBJECTS)')
    if opts.rom_support:
        gc_src.emitLine('#if !defined(DUK_USE_ROM_STRINGS)')
        gc_src.emitLine('#error DUK_USE_ROM_OBJECTS requires DUK_USE_ROM_STRINGS')
        gc_src.emitLine('#endif')
        gc_src.emitLine('#if defined(DUK_USE_HSTRING_ARRIDX)')
        gc_src.emitLine('#error DUK_USE_HSTRING_ARRIDX is currently incompatible with ROM built-ins')
        gc_src.emitLine('#endif')
    else:
        gc_src.emitLine('#error ROM support not enabled, rerun configure.py with --rom-support')
    gc_src.emitLine('#else  /* DUK_USE_ROM_OBJECTS */')
    if opts.ram_support:
        emit_ramobj_source_nativefunc_array(gc_src, ram_native_funcs)  # endian independent
        gc_src.emitLine('#if defined(DUK_USE_DOUBLE_LE)')
        emit_ramobj_source_objinit_data(gc_src, ramobj_data_le)
        gc_src.emitLine('#elif defined(DUK_USE_DOUBLE_BE)')
        emit_ramobj_source_objinit_data(gc_src, ramobj_data_be)
        gc_src.emitLine('#elif defined(DUK_USE_DOUBLE_ME)')
        emit_ramobj_source_objinit_data(gc_src, ramobj_data_me)
        gc_src.emitLine('#else')
        gc_src.emitLine('#error invalid endianness defines')
        gc_src.emitLine('#endif')
    else:
        gc_src.emitLine('#error RAM support not enabled, rerun configure.py with --ram-support')
    gc_src.emitLine('#endif  /* DUK_USE_ROM_OBJECTS */')

    gc_hdr = dukutil.GenerateC()
    gc_hdr.emitHeader('genbuiltins.py')
    gc_hdr.emitLine('#if !defined(DUK_BUILTINS_H_INCLUDED)')
    gc_hdr.emitLine('#define DUK_BUILTINS_H_INCLUDED')
    gc_hdr.emitLine('')
    gc_hdr.emitLine('#if defined(DUK_USE_ROM_STRINGS)')
    if opts.rom_support:
        emit_header_stridx_defines(gc_hdr, rom_meta)
        rom_emit_strings_header(gc_hdr, rom_meta)
    else:
        gc_hdr.emitLine('#error ROM support not enabled, rerun configure.py with --rom-support')
    gc_hdr.emitLine('#else  /* DUK_USE_ROM_STRINGS */')
    if opts.ram_support:
        emit_header_stridx_defines(gc_hdr, ram_meta)
        emit_ramstr_header_strinit_defines(gc_hdr, ram_meta, ramstr_data, ramstr_maxlen)
    else:
        gc_hdr.emitLine('#error RAM support not enabled, rerun configure.py with --ram-support')
    gc_hdr.emitLine('#endif  /* DUK_USE_ROM_STRINGS */')
    gc_hdr.emitLine('')
    gc_hdr.emitLine('#if defined(DUK_USE_ROM_OBJECTS)')
    if opts.rom_support:
        # Currently DUK_USE_ROM_PTRCOMP_FIRST must match our fixed
        # define, and the two must be updated in sync.  Catch any
        # mismatch to avoid difficult to diagnose errors.
        gc_hdr.emitLine('#if !defined(DUK_USE_ROM_PTRCOMP_FIRST)')
        gc_hdr.emitLine('#error missing DUK_USE_ROM_PTRCOMP_FIRST define')
        gc_hdr.emitLine('#endif')
        gc_hdr.emitLine('#if (DUK_USE_ROM_PTRCOMP_FIRST != %dL)' % ROMPTR_FIRST)
        gc_hdr.emitLine('#error DUK_USE_ROM_PTRCOMP_FIRST must match ROMPTR_FIRST in genbuiltins.py (%d), update manually and re-dist' % ROMPTR_FIRST)
        gc_hdr.emitLine('#endif')
        emit_header_native_function_declarations(gc_hdr, rom_meta)
        rom_emit_objects_header(gc_hdr, rom_meta)
    else:
        gc_hdr.emitLine('#error RAM support not enabled, rerun configure.py with --ram-support')
    gc_hdr.emitLine('#else  /* DUK_USE_ROM_OBJECTS */')
    if opts.ram_support:
        emit_header_native_function_declarations(gc_hdr, ram_meta)
        emit_ramobj_header_nativefunc_array(gc_hdr, ram_native_funcs)
        emit_ramobj_header_objects(gc_hdr, ram_meta)
        gc_hdr.emitLine('#if defined(DUK_USE_DOUBLE_LE)')
        emit_ramobj_header_initdata(gc_hdr, ramobj_data_le)
        gc_hdr.emitLine('#elif defined(DUK_USE_DOUBLE_BE)')
        emit_ramobj_header_initdata(gc_hdr, ramobj_data_be)
        gc_hdr.emitLine('#elif defined(DUK_USE_DOUBLE_ME)')
        emit_ramobj_header_initdata(gc_hdr, ramobj_data_me)
        gc_hdr.emitLine('#else')
        gc_hdr.emitLine('#error invalid endianness defines')
        gc_hdr.emitLine('#endif')
    else:
        gc_hdr.emitLine('#error RAM support not enabled, rerun configure.py with --ram-support')
    gc_hdr.emitLine('#endif  /* DUK_USE_ROM_OBJECTS */')
    gc_hdr.emitLine('#endif  /* DUK_BUILTINS_H_INCLUDED */')

    with open(opts.out_source, 'wb') as f:
        f.write(gc_src.getString())
    logger.debug('Wrote built-ins source to ' + opts.out_source)

    with open(opts.out_header, 'wb') as f:
        f.write(gc_hdr.getString())
    logger.debug('Wrote built-ins header to ' + opts.out_header)

    # Write a JSON file with build metadata, e.g. built-in strings.

    ver = long(build_info['duk_version'])
    plain_strs = []
    base64_strs = []
    str_objs = []
    for s in ram_meta['strings_stridx']:  # XXX: provide all lists?
        t1 = bytes_to_unicode(s['str'])
        t2 = unicode_to_bytes(s['str']).encode('base64').strip()
        plain_strs.append(t1)
        base64_strs.append(t2)
        str_objs.append({
            'plain': t1, 'base64': t2, 'define': s['define']
        })
    meta = {
        'comment': 'Metadata for Duktape sources',
        'duk_version': ver,
        'duk_version_string': '%d.%d.%d' % (ver / 10000, (ver / 100) % 100, ver % 100),
        'git_commit': build_info['git_commit'],
        'git_branch': build_info['git_branch'],
        'git_describe': build_info['git_describe'],
        'builtin_strings': plain_strs,
        'builtin_strings_base64': base64_strs,
        'builtin_strings_info': str_objs
    }

    with open(opts.out_metadata_json, 'wb') as f:
        f.write(json.dumps(meta, indent=4, sort_keys=True, ensure_ascii=True))
    logger.debug('Wrote built-ins metadata to ' + opts.out_metadata_json)

if __name__ == '__main__':
    main()