# orm/strategies.py
# Copyright (C) 2005-2020 the SQLAlchemy authors and contributors
# <see AUTHORS file>
#
# This module is part of SQLAlchemy and is released under
# the MIT License: http://www.opensource.org/licenses/mit-license.php

"""sqlalchemy.orm.interfaces.LoaderStrategy
   implementations, and related MapperOptions."""
from __future__ import absolute_import

import collections
import itertools

from . import attributes
from . import exc as orm_exc
from . import interfaces
from . import loading
from . import properties
from . import query
from . import relationships
from . import unitofwork
from . import util as orm_util
from .base import _DEFER_FOR_STATE
from .base import _RAISE_FOR_STATE
from .base import _SET_DEFERRED_EXPIRED
from .context import _column_descriptions
from .context import ORMCompileState
from .interfaces import LoaderStrategy
from .interfaces import StrategizedProperty
from .session import _state_session
from .state import InstanceState
from .util import _none_set
from .util import aliased
from .. import event
from .. import exc as sa_exc
from .. import future
from .. import inspect
from .. import log
from .. import sql
from .. import util
from ..sql import util as sql_util
from ..sql import visitors


def _register_attribute(
    prop,
    mapper,
    useobject,
    compare_function=None,
    typecallable=None,
    callable_=None,
    proxy_property=None,
    active_history=False,
    impl_class=None,
    **kw
):

    listen_hooks = []

    uselist = useobject and prop.uselist

    if useobject and prop.single_parent:
        listen_hooks.append(single_parent_validator)

    if prop.key in prop.parent.validators:
        fn, opts = prop.parent.validators[prop.key]
        listen_hooks.append(
            lambda desc, prop: orm_util._validator_events(
                desc, prop.key, fn, **opts
            )
        )

    if useobject:
        listen_hooks.append(unitofwork.track_cascade_events)

    # need to assemble backref listeners
    # after the singleparentvalidator, mapper validator
    if useobject:
        backref = prop.back_populates
        if backref and prop._effective_sync_backref:
            listen_hooks.append(
                lambda desc, prop: attributes.backref_listeners(
                    desc, backref, uselist
                )
            )

    # a single MapperProperty is shared down a class inheritance
    # hierarchy, so we set up attribute instrumentation and backref event
    # for each mapper down the hierarchy.

    # typically, "mapper" is the same as prop.parent, due to the way
    # the configure_mappers() process runs, however this is not strongly
    # enforced, and in the case of a second configure_mappers() run the
    # mapper here might not be prop.parent; also, a subclass mapper may
    # be called here before a superclass mapper.  That is, can't depend
    # on mappers not already being set up so we have to check each one.

    for m in mapper.self_and_descendants:
        if prop is m._props.get(
            prop.key
        ) and not m.class_manager._attr_has_impl(prop.key):

            desc = attributes.register_attribute_impl(
                m.class_,
                prop.key,
                parent_token=prop,
                uselist=uselist,
                compare_function=compare_function,
                useobject=useobject,
                trackparent=useobject
                and (
                    prop.single_parent
                    or prop.direction is interfaces.ONETOMANY
                ),
                typecallable=typecallable,
                callable_=callable_,
                active_history=active_history,
                impl_class=impl_class,
                send_modified_events=not useobject or not prop.viewonly,
                doc=prop.doc,
                **kw
            )

            for hook in listen_hooks:
                hook(desc, prop)


@properties.ColumnProperty.strategy_for(instrument=False, deferred=False)
class UninstrumentedColumnLoader(LoaderStrategy):
    """Represent a non-instrumented MapperProperty.

    The polymorphic_on argument of mapper() often results in this,
    if the argument is against the with_polymorphic selectable.

    """

    __slots__ = ("columns",)

    def __init__(self, parent, strategy_key):
        super(UninstrumentedColumnLoader, self).__init__(parent, strategy_key)
        self.columns = self.parent_property.columns

    def setup_query(
        self,
        compile_state,
        query_entity,
        path,
        loadopt,
        adapter,
        column_collection=None,
        **kwargs
    ):
        for c in self.columns:
            if adapter:
                c = adapter.columns[c]
            column_collection.append(c)

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        pass


@log.class_logger
@properties.ColumnProperty.strategy_for(instrument=True, deferred=False)
class ColumnLoader(LoaderStrategy):
    """Provide loading behavior for a :class:`.ColumnProperty`."""

    __slots__ = "columns", "is_composite"

    def __init__(self, parent, strategy_key):
        super(ColumnLoader, self).__init__(parent, strategy_key)
        self.columns = self.parent_property.columns
        self.is_composite = hasattr(self.parent_property, "composite_class")

    def setup_query(
        self,
        compile_state,
        query_entity,
        path,
        loadopt,
        adapter,
        column_collection,
        memoized_populators,
        check_for_adapt=False,
        **kwargs
    ):
        for c in self.columns:
            if adapter:
                if check_for_adapt:
                    c = adapter.adapt_check_present(c)
                    if c is None:
                        return
                else:
                    c = adapter.columns[c]

            column_collection.append(c)

        fetch = self.columns[0]
        if adapter:
            fetch = adapter.columns[fetch]
        memoized_populators[self.parent_property] = fetch

    def init_class_attribute(self, mapper):
        self.is_class_level = True
        coltype = self.columns[0].type
        # TODO: check all columns ?  check for foreign key as well?
        active_history = (
            self.parent_property.active_history
            or self.columns[0].primary_key
            or (
                mapper.version_id_col is not None
                and mapper._columntoproperty.get(mapper.version_id_col, None)
                is self.parent_property
            )
        )

        _register_attribute(
            self.parent_property,
            mapper,
            useobject=False,
            compare_function=coltype.compare_values,
            active_history=active_history,
        )

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        # look through list of columns represented here
        # to see which, if any, is present in the row.
        for col in self.columns:
            if adapter:
                col = adapter.columns[col]
            getter = result._getter(col, False)
            if getter:
                populators["quick"].append((self.key, getter))
                break
        else:
            populators["expire"].append((self.key, True))


@log.class_logger
@properties.ColumnProperty.strategy_for(query_expression=True)
class ExpressionColumnLoader(ColumnLoader):
    def __init__(self, parent, strategy_key):
        super(ExpressionColumnLoader, self).__init__(parent, strategy_key)

        null = sql.null()
        self._have_default_expression = any(
            not c.compare(null) for c in self.parent_property.columns
        )

    def setup_query(
        self,
        compile_state,
        query_entity,
        path,
        loadopt,
        adapter,
        column_collection,
        memoized_populators,
        **kwargs
    ):
        columns = None
        if loadopt and "expression" in loadopt.local_opts:
            columns = [loadopt.local_opts["expression"]]
        elif self._have_default_expression:
            columns = self.parent_property.columns

        if columns is None:
            return

        for c in columns:
            if adapter:
                c = adapter.columns[c]
            column_collection.append(c)

        fetch = columns[0]
        if adapter:
            fetch = adapter.columns[fetch]
        memoized_populators[self.parent_property] = fetch

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        # look through list of columns represented here
        # to see which, if any, is present in the row.
        if loadopt and "expression" in loadopt.local_opts:
            columns = [loadopt.local_opts["expression"]]

            for col in columns:
                if adapter:
                    col = adapter.columns[col]
                getter = result._getter(col, False)
                if getter:
                    populators["quick"].append((self.key, getter))
                    break
            else:
                populators["expire"].append((self.key, True))

    def init_class_attribute(self, mapper):
        self.is_class_level = True

        _register_attribute(
            self.parent_property,
            mapper,
            useobject=False,
            compare_function=self.columns[0].type.compare_values,
            accepts_scalar_loader=False,
        )


@log.class_logger
@properties.ColumnProperty.strategy_for(deferred=True, instrument=True)
@properties.ColumnProperty.strategy_for(
    deferred=True, instrument=True, raiseload=True
)
@properties.ColumnProperty.strategy_for(do_nothing=True)
class DeferredColumnLoader(LoaderStrategy):
    """Provide loading behavior for a deferred :class:`.ColumnProperty`."""

    __slots__ = "columns", "group", "raiseload"

    def __init__(self, parent, strategy_key):
        super(DeferredColumnLoader, self).__init__(parent, strategy_key)
        if hasattr(self.parent_property, "composite_class"):
            raise NotImplementedError(
                "Deferred loading for composite " "types not implemented yet"
            )
        self.raiseload = self.strategy_opts.get("raiseload", False)
        self.columns = self.parent_property.columns
        self.group = self.parent_property.group

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):

        # for a DeferredColumnLoader, this method is only used during a
        # "row processor only" query; see test_deferred.py ->
        # tests with "rowproc_only" in their name.  As of the 1.0 series,
        # loading._instance_processor doesn't use a "row processing" function
        # to populate columns, instead it uses data in the "populators"
        # dictionary.  Normally, the DeferredColumnLoader.setup_query()
        # sets up that data in the "memoized_populators" dictionary
        # and "create_row_processor()" here is never invoked.
        if not self.is_class_level:
            if self.raiseload:
                set_deferred_for_local_state = (
                    self.parent_property._raise_column_loader
                )
            else:
                set_deferred_for_local_state = (
                    self.parent_property._deferred_column_loader
                )
            populators["new"].append((self.key, set_deferred_for_local_state))
        else:
            populators["expire"].append((self.key, False))

    def init_class_attribute(self, mapper):
        self.is_class_level = True

        _register_attribute(
            self.parent_property,
            mapper,
            useobject=False,
            compare_function=self.columns[0].type.compare_values,
            callable_=self._load_for_state,
            load_on_unexpire=False,
        )

    def setup_query(
        self,
        compile_state,
        query_entity,
        path,
        loadopt,
        adapter,
        column_collection,
        memoized_populators,
        only_load_props=None,
        **kw
    ):

        if (
            (
                loadopt
                and "undefer_pks" in loadopt.local_opts
                and set(self.columns).intersection(
                    self.parent._should_undefer_in_wildcard
                )
            )
            or (
                loadopt
                and self.group
                and loadopt.local_opts.get(
                    "undefer_group_%s" % self.group, False
                )
            )
            or (only_load_props and self.key in only_load_props)
        ):
            self.parent_property._get_strategy(
                (("deferred", False), ("instrument", True))
            ).setup_query(
                compile_state,
                query_entity,
                path,
                loadopt,
                adapter,
                column_collection,
                memoized_populators,
                **kw
            )
        elif self.is_class_level:
            memoized_populators[self.parent_property] = _SET_DEFERRED_EXPIRED
        elif not self.raiseload:
            memoized_populators[self.parent_property] = _DEFER_FOR_STATE
        else:
            memoized_populators[self.parent_property] = _RAISE_FOR_STATE

    def _load_for_state(self, state, passive):
        if not state.key:
            return attributes.ATTR_EMPTY

        if not passive & attributes.SQL_OK:
            return attributes.PASSIVE_NO_RESULT

        localparent = state.manager.mapper

        if self.group:
            toload = [
                p.key
                for p in localparent.iterate_properties
                if isinstance(p, StrategizedProperty)
                and isinstance(p.strategy, DeferredColumnLoader)
                and p.group == self.group
            ]
        else:
            toload = [self.key]

        # narrow the keys down to just those which have no history
        group = [k for k in toload if k in state.unmodified]

        session = _state_session(state)
        if session is None:
            raise orm_exc.DetachedInstanceError(
                "Parent instance %s is not bound to a Session; "
                "deferred load operation of attribute '%s' cannot proceed"
                % (orm_util.state_str(state), self.key)
            )

        if self.raiseload:
            self._invoke_raise_load(state, passive, "raise")

        if (
            loading.load_on_ident(
                session,
                future.select(localparent).apply_labels(),
                state.key,
                only_load_props=group,
                refresh_state=state,
            )
            is None
        ):
            raise orm_exc.ObjectDeletedError(state)

        return attributes.ATTR_WAS_SET

    def _invoke_raise_load(self, state, passive, lazy):
        raise sa_exc.InvalidRequestError(
            "'%s' is not available due to raiseload=True" % (self,)
        )


class LoadDeferredColumns(object):
    """serializable loader object used by DeferredColumnLoader"""

    def __init__(self, key, raiseload=False):
        self.key = key
        self.raiseload = raiseload

    def __call__(self, state, passive=attributes.PASSIVE_OFF):
        key = self.key

        localparent = state.manager.mapper
        prop = localparent._props[key]
        if self.raiseload:
            strategy_key = (
                ("deferred", True),
                ("instrument", True),
                ("raiseload", True),
            )
        else:
            strategy_key = (("deferred", True), ("instrument", True))
        strategy = prop._get_strategy(strategy_key)
        return strategy._load_for_state(state, passive)


class AbstractRelationshipLoader(LoaderStrategy):
    """LoaderStratgies which deal with related objects."""

    __slots__ = "mapper", "target", "uselist", "entity"

    def __init__(self, parent, strategy_key):
        super(AbstractRelationshipLoader, self).__init__(parent, strategy_key)
        self.mapper = self.parent_property.mapper
        self.entity = self.parent_property.entity
        self.target = self.parent_property.target
        self.uselist = self.parent_property.uselist


@log.class_logger
@relationships.RelationshipProperty.strategy_for(do_nothing=True)
class DoNothingLoader(LoaderStrategy):
    """Relationship loader that makes no change to the object's state.

    Compared to NoLoader, this loader does not initialize the
    collection/attribute to empty/none; the usual default LazyLoader will
    take effect.

    """


@log.class_logger
@relationships.RelationshipProperty.strategy_for(lazy="noload")
@relationships.RelationshipProperty.strategy_for(lazy=None)
class NoLoader(AbstractRelationshipLoader):
    """Provide loading behavior for a :class:`.RelationshipProperty`
    with "lazy=None".

    """

    __slots__ = ()

    def init_class_attribute(self, mapper):
        self.is_class_level = True

        _register_attribute(
            self.parent_property,
            mapper,
            useobject=True,
            typecallable=self.parent_property.collection_class,
        )

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        def invoke_no_load(state, dict_, row):
            if self.uselist:
                attributes.init_state_collection(state, dict_, self.key)
            else:
                dict_[self.key] = None

        populators["new"].append((self.key, invoke_no_load))


@log.class_logger
@relationships.RelationshipProperty.strategy_for(lazy=True)
@relationships.RelationshipProperty.strategy_for(lazy="select")
@relationships.RelationshipProperty.strategy_for(lazy="raise")
@relationships.RelationshipProperty.strategy_for(lazy="raise_on_sql")
@relationships.RelationshipProperty.strategy_for(lazy="baked_select")
class LazyLoader(AbstractRelationshipLoader, util.MemoizedSlots):
    """Provide loading behavior for a :class:`.RelationshipProperty`
    with "lazy=True", that is loads when first accessed.

    """

    __slots__ = (
        "_lazywhere",
        "_rev_lazywhere",
        "_lazyload_reverse_option",
        "_order_by",
        "use_get",
        "is_aliased_class",
        "_bind_to_col",
        "_equated_columns",
        "_rev_bind_to_col",
        "_rev_equated_columns",
        "_simple_lazy_clause",
        "_raise_always",
        "_raise_on_sql",
        "_bakery",
    )

    def __init__(self, parent, strategy_key):
        super(LazyLoader, self).__init__(parent, strategy_key)
        self._raise_always = self.strategy_opts["lazy"] == "raise"
        self._raise_on_sql = self.strategy_opts["lazy"] == "raise_on_sql"

        self.is_aliased_class = inspect(self.entity).is_aliased_class

        join_condition = self.parent_property._join_condition
        (
            self._lazywhere,
            self._bind_to_col,
            self._equated_columns,
        ) = join_condition.create_lazy_clause()

        (
            self._rev_lazywhere,
            self._rev_bind_to_col,
            self._rev_equated_columns,
        ) = join_condition.create_lazy_clause(reverse_direction=True)

        if self.parent_property.order_by:
            self._order_by = [
                sql_util._deep_annotate(elem, {"_orm_adapt": True})
                for elem in util.to_list(self.parent_property.order_by)
            ]
        else:
            self._order_by = None

        self.logger.info("%s lazy loading clause %s", self, self._lazywhere)

        # determine if our "lazywhere" clause is the same as the mapper's
        # get() clause.  then we can just use mapper.get()
        #
        # TODO: the "not self.uselist" can be taken out entirely; a m2o
        # load that populates for a list (very unusual, but is possible with
        # the API) can still set for "None" and the attribute system will
        # populate as an empty list.
        self.use_get = (
            not self.is_aliased_class
            and not self.uselist
            and self.entity._get_clause[0].compare(
                self._lazywhere,
                use_proxies=True,
                equivalents=self.mapper._equivalent_columns,
            )
        )

        if self.use_get:
            for col in list(self._equated_columns):
                if col in self.mapper._equivalent_columns:
                    for c in self.mapper._equivalent_columns[col]:
                        self._equated_columns[c] = self._equated_columns[col]

            self.logger.info(
                "%s will use Session.get() to " "optimize instance loads", self
            )

    def init_class_attribute(self, mapper):
        self.is_class_level = True

        active_history = (
            self.parent_property.active_history
            or self.parent_property.direction is not interfaces.MANYTOONE
            or not self.use_get
        )

        # MANYTOONE currently only needs the
        # "old" value for delete-orphan
        # cascades.  the required _SingleParentValidator
        # will enable active_history
        # in that case.  otherwise we don't need the
        # "old" value during backref operations.
        _register_attribute(
            self.parent_property,
            mapper,
            useobject=True,
            callable_=self._load_for_state,
            typecallable=self.parent_property.collection_class,
            active_history=active_history,
        )

    def _memoized_attr__simple_lazy_clause(self):

        lazywhere = sql_util._deep_annotate(
            self._lazywhere, {"_orm_adapt": True}
        )

        criterion, bind_to_col = (lazywhere, self._bind_to_col)

        params = []

        def visit_bindparam(bindparam):
            bindparam.unique = False

        visitors.traverse(criterion, {}, {"bindparam": visit_bindparam})

        def visit_bindparam(bindparam):
            if bindparam._identifying_key in bind_to_col:
                params.append(
                    (
                        bindparam.key,
                        bind_to_col[bindparam._identifying_key],
                        None,
                    )
                )
            elif bindparam.callable is None:
                params.append((bindparam.key, None, bindparam.value))

        criterion = visitors.cloned_traverse(
            criterion, {}, {"bindparam": visit_bindparam}
        )

        return criterion, params

    def _generate_lazy_clause(self, state, passive):
        criterion, param_keys = self._simple_lazy_clause

        if state is None:
            return sql_util.adapt_criterion_to_null(
                criterion, [key for key, ident, value in param_keys]
            )

        mapper = self.parent_property.parent

        o = state.obj()  # strong ref
        dict_ = attributes.instance_dict(o)

        if passive & attributes.INIT_OK:
            passive ^= attributes.INIT_OK

        params = {}
        for key, ident, value in param_keys:
            if ident is not None:
                if passive and passive & attributes.LOAD_AGAINST_COMMITTED:
                    value = mapper._get_committed_state_attr_by_column(
                        state, dict_, ident, passive
                    )
                else:
                    value = mapper._get_state_attr_by_column(
                        state, dict_, ident, passive
                    )

            params[key] = value

        return criterion, params

    def _invoke_raise_load(self, state, passive, lazy):
        raise sa_exc.InvalidRequestError(
            "'%s' is not available due to lazy='%s'" % (self, lazy)
        )

    def _load_for_state(self, state, passive):

        if not state.key and (
            (
                not self.parent_property.load_on_pending
                and not state._load_pending
            )
            or not state.session_id
        ):
            return attributes.ATTR_EMPTY

        pending = not state.key
        primary_key_identity = None

        if (not passive & attributes.SQL_OK and not self.use_get) or (
            not passive & attributes.NON_PERSISTENT_OK and pending
        ):
            return attributes.PASSIVE_NO_RESULT

        if (
            # we were given lazy="raise"
            self._raise_always
            # the no_raise history-related flag was not passed
            and not passive & attributes.NO_RAISE
            and (
                # if we are use_get and related_object_ok is disabled,
                # which means we are at most looking in the identity map
                # for history purposes or otherwise returning
                # PASSIVE_NO_RESULT, don't raise.  This is also a
                # history-related flag
                not self.use_get
                or passive & attributes.RELATED_OBJECT_OK
            )
        ):

            self._invoke_raise_load(state, passive, "raise")

        session = _state_session(state)
        if not session:
            if passive & attributes.NO_RAISE:
                return attributes.PASSIVE_NO_RESULT

            raise orm_exc.DetachedInstanceError(
                "Parent instance %s is not bound to a Session; "
                "lazy load operation of attribute '%s' cannot proceed"
                % (orm_util.state_str(state), self.key)
            )

        # if we have a simple primary key load, check the
        # identity map without generating a Query at all
        if self.use_get:
            primary_key_identity = self._get_ident_for_use_get(
                session, state, passive
            )
            if attributes.PASSIVE_NO_RESULT in primary_key_identity:
                return attributes.PASSIVE_NO_RESULT
            elif attributes.NEVER_SET in primary_key_identity:
                return attributes.NEVER_SET

            if _none_set.issuperset(primary_key_identity):
                return None

            if self.key in state.dict:
                return attributes.ATTR_WAS_SET

            # look for this identity in the identity map.  Delegate to the
            # Query class in use, as it may have special rules for how it
            # does this, including how it decides what the correct
            # identity_token would be for this identity.

            instance = session._identity_lookup(
                self.entity,
                primary_key_identity,
                passive=passive,
                lazy_loaded_from=state,
            )

            if instance is not None:
                if instance is attributes.PASSIVE_CLASS_MISMATCH:
                    return None
                else:
                    return instance
            elif (
                not passive & attributes.SQL_OK
                or not passive & attributes.RELATED_OBJECT_OK
            ):
                return attributes.PASSIVE_NO_RESULT

        return self._emit_lazyload(
            session, state, primary_key_identity, passive
        )

    def _get_ident_for_use_get(self, session, state, passive):
        instance_mapper = state.manager.mapper

        if passive & attributes.LOAD_AGAINST_COMMITTED:
            get_attr = instance_mapper._get_committed_state_attr_by_column
        else:
            get_attr = instance_mapper._get_state_attr_by_column

        dict_ = state.dict

        return [
            get_attr(state, dict_, self._equated_columns[pk], passive=passive)
            for pk in self.mapper.primary_key
        ]

    @util.preload_module("sqlalchemy.ext.baked")
    def _memoized_attr__bakery(self):
        return util.preloaded.ext_baked.bakery(size=50)

    @util.preload_module("sqlalchemy.orm.strategy_options")
    def _emit_lazyload(self, session, state, primary_key_identity, passive):
        # emit lazy load now using BakedQuery, to cut way down on the overhead
        # of generating queries.
        # there are two big things we are trying to guard against here:
        #
        # 1. two different lazy loads that need to have a different result,
        #    being cached on the same key.  The results between two lazy loads
        #    can be different due to the options passed to the query, which
        #    take effect for descendant objects.  Therefore we have to make
        #    sure paths and load options generate good cache keys, and if they
        #    don't, we don't cache.
        # 2. a lazy load that gets cached on a key that includes some
        #    "throwaway" object, like a per-query AliasedClass, meaning
        #    the cache key will never be seen again and the cache itself
        #    will fill up.   (the cache is an LRU cache, so while we won't
        #    run out of memory, it will perform terribly when it's full.  A
        #    warning is emitted if this occurs.)   We must prevent the
        #    generation of a cache key that is including a throwaway object
        #    in the key.

        strategy_options = util.preloaded.orm_strategy_options

        # note that "lazy='select'" and "lazy=True" make two separate
        # lazy loaders.   Currently the LRU cache is local to the LazyLoader,
        # however add ourselves to the initial cache key just to future
        # proof in case it moves
        q = self._bakery(lambda session: session.query(self.entity), self)

        q.add_criteria(
            lambda q: q._with_invoke_all_eagers(False), self.parent_property,
        )

        if not self.parent_property.bake_queries:
            q.spoil(full=True)

        if self.parent_property.secondary is not None:
            q.add_criteria(
                lambda q: q.select_from(
                    self.mapper, self.parent_property.secondary
                )
            )

        pending = not state.key

        # don't autoflush on pending
        if pending or passive & attributes.NO_AUTOFLUSH:
            q.add_criteria(lambda q: q.autoflush(False))

        if state.load_options:
            # here, if any of the options cannot return a cache key,
            # the BakedQuery "spoils" and caching will not occur.  a path
            # that features Cls.attribute.of_type(some_alias) will cancel
            # caching, for example, since "some_alias" is user-defined and
            # is usually a throwaway object.
            effective_path = state.load_path[self.parent_property]

            q._add_lazyload_options(state.load_options, effective_path)

        if self.use_get:
            if self._raise_on_sql:
                self._invoke_raise_load(state, passive, "raise_on_sql")

            return (
                q(session)
                .with_post_criteria(lambda q: q._set_lazyload_from(state))
                ._load_on_pk_identity(
                    session, session.query(self.mapper), primary_key_identity
                )
            )

        if self._order_by:
            q.add_criteria(lambda q: q.order_by(*self._order_by))

        def _lazyload_reverse(compile_context):
            for rev in self.parent_property._reverse_property:
                # reverse props that are MANYTOONE are loading *this*
                # object from get(), so don't need to eager out to those.
                if (
                    rev.direction is interfaces.MANYTOONE
                    and rev._use_get
                    and not isinstance(rev.strategy, LazyLoader)
                ):
                    strategy_options.Load.for_existing_path(
                        compile_context.compile_options._current_path[
                            rev.parent
                        ]
                    ).lazyload(rev.key).process_compile_state(compile_context)

        q.add_criteria(
            lambda q: q._add_context_option(
                _lazyload_reverse, self.parent_property
            )
        )

        lazy_clause, params = self._generate_lazy_clause(state, passive)
        if self.key in state.dict:
            return attributes.ATTR_WAS_SET

        if pending:
            if util.has_intersection(orm_util._none_set, params.values()):
                return None

        elif util.has_intersection(orm_util._never_set, params.values()):
            return None

        if self._raise_on_sql:
            self._invoke_raise_load(state, passive, "raise_on_sql")

        q.add_criteria(lambda q: q.filter(lazy_clause))

        # set parameters in the query such that we don't overwrite
        # parameters that are already set within it
        def set_default_params(q):
            params.update(q.load_options._params)
            q.load_options += {"_params": params}
            return q

        result = (
            q(session)
            .with_post_criteria(lambda q: q._set_lazyload_from(state))
            .with_post_criteria(set_default_params)
            .all()
        )
        if self.uselist:
            return result
        else:
            l = len(result)
            if l:
                if l > 1:
                    util.warn(
                        "Multiple rows returned with "
                        "uselist=False for lazily-loaded attribute '%s' "
                        % self.parent_property
                    )

                return result[0]
            else:
                return None

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        key = self.key

        if not self.is_class_level:
            # we are not the primary manager for this attribute
            # on this class - set up a
            # per-instance lazyloader, which will override the
            # class-level behavior.
            # this currently only happens when using a
            # "lazyload" option on a "no load"
            # attribute - "eager" attributes always have a
            # class-level lazyloader installed.
            set_lazy_callable = (
                InstanceState._instance_level_callable_processor
            )(mapper.class_manager, LoadLazyAttribute(key, self), key)

            populators["new"].append((self.key, set_lazy_callable))
        elif context.populate_existing or mapper.always_refresh:

            def reset_for_lazy_callable(state, dict_, row):
                # we are the primary manager for this attribute on
                # this class - reset its
                # per-instance attribute state, so that the class-level
                # lazy loader is
                # executed when next referenced on this instance.
                # this is needed in
                # populate_existing() types of scenarios to reset
                # any existing state.
                state._reset(dict_, key)

            populators["new"].append((self.key, reset_for_lazy_callable))


class LoadLazyAttribute(object):
    """serializable loader object used by LazyLoader"""

    def __init__(self, key, initiating_strategy):
        self.key = key
        self.strategy_key = initiating_strategy.strategy_key

    def __call__(self, state, passive=attributes.PASSIVE_OFF):
        key = self.key
        instance_mapper = state.manager.mapper
        prop = instance_mapper._props[key]
        strategy = prop._strategies[self.strategy_key]

        return strategy._load_for_state(state, passive)


class PostLoader(AbstractRelationshipLoader):
    """A relationship loader that emits a second SELECT statement."""

    def _immediateload_create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        return self.parent_property._get_strategy(
            (("lazy", "immediate"),)
        ).create_row_processor(
            context,
            query_entity,
            path,
            loadopt,
            mapper,
            result,
            adapter,
            populators,
        )


@relationships.RelationshipProperty.strategy_for(lazy="immediate")
class ImmediateLoader(PostLoader):
    __slots__ = ()

    def init_class_attribute(self, mapper):
        self.parent_property._get_strategy(
            (("lazy", "select"),)
        ).init_class_attribute(mapper)

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        def load_immediate(state, dict_, row):
            state.get_impl(self.key).get(state, dict_)

        populators["delayed"].append((self.key, load_immediate))


@log.class_logger
@relationships.RelationshipProperty.strategy_for(lazy="subquery")
class SubqueryLoader(PostLoader):
    __slots__ = ("join_depth",)

    def __init__(self, parent, strategy_key):
        super(SubqueryLoader, self).__init__(parent, strategy_key)
        self.join_depth = self.parent_property.join_depth

    def init_class_attribute(self, mapper):
        self.parent_property._get_strategy(
            (("lazy", "select"),)
        ).init_class_attribute(mapper)

    def _get_leftmost(self, subq_path):
        subq_path = subq_path.path
        subq_mapper = orm_util._class_to_mapper(subq_path[0])

        # determine attributes of the leftmost mapper
        if (
            self.parent.isa(subq_mapper)
            and self.parent_property is subq_path[1]
        ):
            leftmost_mapper, leftmost_prop = self.parent, self.parent_property
        else:
            leftmost_mapper, leftmost_prop = subq_mapper, subq_path[1]

        leftmost_cols = leftmost_prop.local_columns

        leftmost_attr = [
            getattr(
                subq_path[0].entity, leftmost_mapper._columntoproperty[c].key
            )
            for c in leftmost_cols
        ]

        return leftmost_mapper, leftmost_attr, leftmost_prop

    def _generate_from_original_query(
        self,
        orig_compile_state,
        orig_query,
        leftmost_mapper,
        leftmost_attr,
        leftmost_relationship,
        orig_entity,
    ):
        # reformat the original query
        # to look only for significant columns
        q = orig_query._clone().correlate(None)

        # LEGACY: make a Query back from the select() !!
        # This suits at least two legacy cases:
        # 1. applications which expect before_compile() to be called
        #    below when we run .subquery() on this query (Keystone)
        # 2. applications which are doing subqueryload with complex
        #    from_self() queries, as query.subquery() / .statement
        #    has to do the full compile context for multiply-nested
        #    from_self() (Neutron) - see test_subqload_from_self
        #    for demo.
        q2 = query.Query.__new__(query.Query)
        q2.__dict__.update(q.__dict__)
        q = q2

        # set the query's "FROM" list explicitly to what the
        # FROM list would be in any case, as we will be limiting
        # the columns in the SELECT list which may no longer include
        # all entities mentioned in things like WHERE, JOIN, etc.
        if not q._from_obj:
            q._enable_assertions = False
            q.select_from.non_generative(
                q,
                *{
                    ent["entity"]
                    for ent in _column_descriptions(
                        orig_query, compile_state=orig_compile_state
                    )
                    if ent["entity"] is not None
                }
            )

        # select from the identity columns of the outer (specifically, these
        # are the 'local_cols' of the property).  This will remove other
        # columns from the query that might suggest the right entity which is
        # why we do set select_from above.   The attributes we have are
        # coerced and adapted using the original query's adapter, which is
        # needed only for the case of adapting a subclass column to
        # that of a polymorphic selectable, e.g. we have
        # Engineer.primary_language and the entity is Person.  All other
        # adaptations, e.g. from_self, select_entity_from(), will occur
        # within the new query when it compiles, as the compile_state we are
        # using here is only a partial one.  If the subqueryload is from a
        # with_polymorphic() or other aliased() object, left_attr will already
        # be the correct attributes so no adaptation is needed.
        target_cols = orig_compile_state._adapt_col_list(
            [
                sql.coercions.expect(sql.roles.ColumnsClauseRole, o)
                for o in leftmost_attr
            ],
            orig_compile_state._get_current_adapter(),
        )
        q._raw_columns = target_cols

        distinct_target_key = leftmost_relationship.distinct_target_key

        if distinct_target_key is True:
            q._distinct = True
        elif distinct_target_key is None:
            # if target_cols refer to a non-primary key or only
            # part of a composite primary key, set the q as distinct
            for t in set(c.table for c in target_cols):
                if not set(target_cols).issuperset(t.primary_key):
                    q._distinct = True
                    break

        # don't need ORDER BY if no limit/offset
        if q._limit_clause is None and q._offset_clause is None:
            q._order_by_clauses = ()

        if q._distinct is True and q._order_by_clauses:
            # the logic to automatically add the order by columns to the query
            # when distinct is True is deprecated in the query
            to_add = sql_util.expand_column_list_from_order_by(
                target_cols, q._order_by_clauses
            )
            if to_add:
                q._set_entities(target_cols + to_add)

        # the original query now becomes a subquery
        # which we'll join onto.
        # LEGACY: as "q" is a Query, the before_compile() event is invoked
        # here.
        embed_q = q.apply_labels().subquery()
        left_alias = orm_util.AliasedClass(
            leftmost_mapper, embed_q, use_mapper_path=True
        )
        return left_alias

    def _prep_for_joins(self, left_alias, subq_path):
        # figure out what's being joined.  a.k.a. the fun part
        to_join = []
        pairs = list(subq_path.pairs())

        for i, (mapper, prop) in enumerate(pairs):
            if i > 0:
                # look at the previous mapper in the chain -
                # if it is as or more specific than this prop's
                # mapper, use that instead.
                # note we have an assumption here that
                # the non-first element is always going to be a mapper,
                # not an AliasedClass

                prev_mapper = pairs[i - 1][1].mapper
                to_append = prev_mapper if prev_mapper.isa(mapper) else mapper
            else:
                to_append = mapper

            to_join.append((to_append, prop.key))

        # determine the immediate parent class we are joining from,
        # which needs to be aliased.

        if len(to_join) < 2:
            # in the case of a one level eager load, this is the
            # leftmost "left_alias".
            parent_alias = left_alias
        else:
            info = inspect(to_join[-1][0])
            if info.is_aliased_class:
                parent_alias = info.entity
            else:
                # alias a plain mapper as we may be
                # joining multiple times
                parent_alias = orm_util.AliasedClass(
                    info.entity, use_mapper_path=True
                )

        local_cols = self.parent_property.local_columns

        local_attr = [
            getattr(parent_alias, self.parent._columntoproperty[c].key)
            for c in local_cols
        ]
        return to_join, local_attr, parent_alias

    def _apply_joins(
        self, q, to_join, left_alias, parent_alias, effective_entity
    ):

        ltj = len(to_join)
        if ltj == 1:
            to_join = [
                getattr(left_alias, to_join[0][1]).of_type(effective_entity)
            ]
        elif ltj == 2:
            to_join = [
                getattr(left_alias, to_join[0][1]).of_type(parent_alias),
                getattr(parent_alias, to_join[-1][1]).of_type(
                    effective_entity
                ),
            ]
        elif ltj > 2:
            middle = [
                (
                    orm_util.AliasedClass(item[0])
                    if not inspect(item[0]).is_aliased_class
                    else item[0].entity,
                    item[1],
                )
                for item in to_join[1:-1]
            ]
            inner = []

            while middle:
                item = middle.pop(0)
                attr = getattr(item[0], item[1])
                if middle:
                    attr = attr.of_type(middle[0][0])
                else:
                    attr = attr.of_type(parent_alias)

                inner.append(attr)

            to_join = (
                [getattr(left_alias, to_join[0][1]).of_type(inner[0].parent)]
                + inner
                + [
                    getattr(parent_alias, to_join[-1][1]).of_type(
                        effective_entity
                    )
                ]
            )

        for attr in to_join:
            q = q.join(attr)

        return q

    def _setup_options(self, q, subq_path, orig_query, effective_entity):
        # propagate loader options etc. to the new query.
        # these will fire relative to subq_path.
        q = q._with_current_path(subq_path)
        q = q.options(*orig_query._with_options)

        return q

    def _setup_outermost_orderby(self, q):
        if self.parent_property.order_by:

            def _setup_outermost_orderby(compile_context):
                compile_context.eager_order_by += tuple(
                    util.to_list(self.parent_property.order_by)
                )

            q = q._add_context_option(
                _setup_outermost_orderby, self.parent_property
            )

        return q

    class _SubqCollections(object):
        """Given a :class:`_query.Query` used to emit the "subquery load",
        provide a load interface that executes the query at the
        first moment a value is needed.

        """

        __slots__ = (
            "session",
            "execution_options",
            "load_options",
            "subq",
            "_data",
        )

        def __init__(self, context, subq):
            # avoid creating a cycle by storing context
            # even though that's preferable
            self.session = context.session
            self.execution_options = context.execution_options
            self.load_options = context.load_options
            self.subq = subq
            self._data = None

        def get(self, key, default):
            if self._data is None:
                self._load()
            return self._data.get(key, default)

        def _load(self):
            self._data = collections.defaultdict(list)

            q = self.subq
            assert q.session is None
            if "compiled_cache" in self.execution_options:
                q = q.execution_options(
                    compiled_cache=self.execution_options["compiled_cache"]
                )
            q = q.with_session(self.session)

            if self.load_options._populate_existing:
                q = q.populate_existing()
            # to work with baked query, the parameters may have been
            # updated since this query was created, so take these into account

            rows = list(q.params(self.load_options._params))
            for k, v in itertools.groupby(rows, lambda x: x[1:]):
                self._data[k].extend(vv[0] for vv in v)

        def loader(self, state, dict_, row):
            if self._data is None:
                self._load()

    def _setup_query_from_rowproc(
        self, context, path, entity, loadopt, adapter,
    ):
        compile_state = context.compile_state
        if (
            not compile_state.compile_options._enable_eagerloads
            or compile_state.compile_options._for_refresh_state
        ):
            return

        context.loaders_require_buffering = True

        path = path[self.parent_property]

        # build up a path indicating the path from the leftmost
        # entity to the thing we're subquery loading.
        with_poly_entity = path.get(
            compile_state.attributes, "path_with_polymorphic", None
        )
        if with_poly_entity is not None:
            effective_entity = with_poly_entity
        else:
            effective_entity = self.entity

        subq_path = context.query._execution_options.get(
            ("subquery_path", None), orm_util.PathRegistry.root
        )

        subq_path = subq_path + path

        # if not via query option, check for
        # a cycle
        if not path.contains(compile_state.attributes, "loader"):
            if self.join_depth:
                if (
                    (
                        compile_state.current_path.length
                        if compile_state.current_path
                        else 0
                    )
                    + path.length
                ) / 2 > self.join_depth:
                    return
            elif subq_path.contains_mapper(self.mapper):
                return

        (
            leftmost_mapper,
            leftmost_attr,
            leftmost_relationship,
        ) = self._get_leftmost(subq_path)

        # use the current query being invoked, not the compile state
        # one.  this is so that we get the current parameters.  however,
        # it means we can't use the existing compile state, we have to make
        # a new one.    other approaches include possibly using the
        # compiled query but swapping the params, seems only marginally
        # less time spent but more complicated
        orig_query = context.query._execution_options.get(
            ("orig_query", SubqueryLoader), context.query
        )

        # make a new compile_state for the query that's probably cached, but
        # we're sort of undoing a bit of that caching :(
        compile_state_cls = ORMCompileState._get_plugin_class_for_plugin(
            orig_query, "orm"
        )

        # this would create the full blown compile state, which we don't
        # need
        # orig_compile_state = compile_state_cls.create_for_statement(
        # orig_query, None)

        if orig_query._is_lambda_element:
            util.warn(
                'subqueryloader for "%s" must invoke lambda callable at %r in '
                "order to produce a new query, decreasing the efficiency "
                "of caching for this statement.  Consider using "
                "selectinload() for more effective full-lambda caching"
                % (self, orig_query)
            )
            orig_query = orig_query._resolved

        # this is the more "quick" version, however it's not clear how
        # much of this we need.    in particular I can't get a test to
        # fail if the "set_base_alias" is missing and not sure why that is.
        orig_compile_state = compile_state_cls._create_entities_collection(
            orig_query
        )

        # generate a new Query from the original, then
        # produce a subquery from it.
        left_alias = self._generate_from_original_query(
            orig_compile_state,
            orig_query,
            leftmost_mapper,
            leftmost_attr,
            leftmost_relationship,
            entity,
        )

        # generate another Query that will join the
        # left alias to the target relationships.
        # basically doing a longhand
        # "from_self()".  (from_self() itself not quite industrial
        # strength enough for all contingencies...but very close)

        q = query.Query(effective_entity)

        q._execution_options = q._execution_options.union(
            {
                ("orig_query", SubqueryLoader): orig_query,
                ("subquery_path", None): subq_path,
            }
        )

        q = q._set_enable_single_crit(False)
        to_join, local_attr, parent_alias = self._prep_for_joins(
            left_alias, subq_path
        )

        q = q.add_columns(*local_attr)
        q = self._apply_joins(
            q, to_join, left_alias, parent_alias, effective_entity
        )

        q = self._setup_options(q, subq_path, orig_query, effective_entity)
        q = self._setup_outermost_orderby(q)

        return q

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        if context.refresh_state:
            return self._immediateload_create_row_processor(
                context,
                query_entity,
                path,
                loadopt,
                mapper,
                result,
                adapter,
                populators,
            )

        if not self.parent.class_manager[self.key].impl.supports_population:
            raise sa_exc.InvalidRequestError(
                "'%s' does not support object "
                "population - eager loading cannot be applied." % self
            )

        # a little dance here as the "path" is still something that only
        # semi-tracks the exact series of things we are loading, still not
        # telling us about with_polymorphic() and stuff like that when it's at
        # the root..  the initial MapperEntity is more accurate for this case.
        if len(path) == 1:
            if not orm_util._entity_isa(query_entity.entity_zero, self.parent):
                return
        elif not orm_util._entity_isa(path[-1], self.parent):
            return

        subq = self._setup_query_from_rowproc(
            context, path, path[-1], loadopt, adapter,
        )

        if subq is None:
            return

        assert subq.session is None

        path = path[self.parent_property]

        local_cols = self.parent_property.local_columns

        # cache the loaded collections in the context
        # so that inheriting mappers don't re-load when they
        # call upon create_row_processor again
        collections = path.get(context.attributes, "collections")
        if collections is None:
            collections = self._SubqCollections(context, subq)
            path.set(context.attributes, "collections", collections)

        if adapter:
            local_cols = [adapter.columns[c] for c in local_cols]

        if self.uselist:
            self._create_collection_loader(
                context, result, collections, local_cols, populators
            )
        else:
            self._create_scalar_loader(
                context, result, collections, local_cols, populators
            )

    def _create_collection_loader(
        self, context, result, collections, local_cols, populators
    ):
        tuple_getter = result._tuple_getter(local_cols)

        def load_collection_from_subq(state, dict_, row):
            collection = collections.get(tuple_getter(row), ())
            state.get_impl(self.key).set_committed_value(
                state, dict_, collection
            )

        def load_collection_from_subq_existing_row(state, dict_, row):
            if self.key not in dict_:
                load_collection_from_subq(state, dict_, row)

        populators["new"].append((self.key, load_collection_from_subq))
        populators["existing"].append(
            (self.key, load_collection_from_subq_existing_row)
        )

        if context.invoke_all_eagers:
            populators["eager"].append((self.key, collections.loader))

    def _create_scalar_loader(
        self, context, result, collections, local_cols, populators
    ):
        tuple_getter = result._tuple_getter(local_cols)

        def load_scalar_from_subq(state, dict_, row):
            collection = collections.get(tuple_getter(row), (None,))
            if len(collection) > 1:
                util.warn(
                    "Multiple rows returned with "
                    "uselist=False for eagerly-loaded attribute '%s' " % self
                )

            scalar = collection[0]
            state.get_impl(self.key).set_committed_value(state, dict_, scalar)

        def load_scalar_from_subq_existing_row(state, dict_, row):
            if self.key not in dict_:
                load_scalar_from_subq(state, dict_, row)

        populators["new"].append((self.key, load_scalar_from_subq))
        populators["existing"].append(
            (self.key, load_scalar_from_subq_existing_row)
        )
        if context.invoke_all_eagers:
            populators["eager"].append((self.key, collections.loader))


@log.class_logger
@relationships.RelationshipProperty.strategy_for(lazy="joined")
@relationships.RelationshipProperty.strategy_for(lazy=False)
class JoinedLoader(AbstractRelationshipLoader):
    """Provide loading behavior for a :class:`.RelationshipProperty`
    using joined eager loading.

    """

    __slots__ = "join_depth", "_aliased_class_pool"

    def __init__(self, parent, strategy_key):
        super(JoinedLoader, self).__init__(parent, strategy_key)
        self.join_depth = self.parent_property.join_depth
        self._aliased_class_pool = []

    def init_class_attribute(self, mapper):
        self.parent_property._get_strategy(
            (("lazy", "select"),)
        ).init_class_attribute(mapper)

    def setup_query(
        self,
        compile_state,
        query_entity,
        path,
        loadopt,
        adapter,
        column_collection=None,
        parentmapper=None,
        chained_from_outerjoin=False,
        **kwargs
    ):
        """Add a left outer join to the statement that's being constructed."""

        if not compile_state.compile_options._enable_eagerloads:
            return
        elif self.uselist:
            compile_state.multi_row_eager_loaders = True

        path = path[self.parent_property]

        with_polymorphic = None

        user_defined_adapter = (
            self._init_user_defined_eager_proc(
                loadopt, compile_state, compile_state.attributes
            )
            if loadopt
            else False
        )

        if user_defined_adapter is not False:
            (
                clauses,
                adapter,
                add_to_collection,
            ) = self._setup_query_on_user_defined_adapter(
                compile_state,
                query_entity,
                path,
                adapter,
                user_defined_adapter,
            )
        else:
            # if not via query option, check for
            # a cycle
            if not path.contains(compile_state.attributes, "loader"):
                if self.join_depth:
                    if path.length / 2 > self.join_depth:
                        return
                elif path.contains_mapper(self.mapper):
                    return

            (
                clauses,
                adapter,
                add_to_collection,
                chained_from_outerjoin,
            ) = self._generate_row_adapter(
                compile_state,
                query_entity,
                path,
                loadopt,
                adapter,
                column_collection,
                parentmapper,
                chained_from_outerjoin,
            )

        with_poly_entity = path.get(
            compile_state.attributes, "path_with_polymorphic", None
        )
        if with_poly_entity is not None:
            with_polymorphic = inspect(
                with_poly_entity
            ).with_polymorphic_mappers
        else:
            with_polymorphic = None

        path = path[self.entity]

        loading._setup_entity_query(
            compile_state,
            self.mapper,
            query_entity,
            path,
            clauses,
            add_to_collection,
            with_polymorphic=with_polymorphic,
            parentmapper=self.mapper,
            chained_from_outerjoin=chained_from_outerjoin,
        )

        if with_poly_entity is not None and None in set(
            compile_state.secondary_columns
        ):
            raise sa_exc.InvalidRequestError(
                "Detected unaliased columns when generating joined "
                "load.  Make sure to use aliased=True or flat=True "
                "when using joined loading with with_polymorphic()."
            )

    def _init_user_defined_eager_proc(
        self, loadopt, compile_state, target_attributes
    ):

        # check if the opt applies at all
        if "eager_from_alias" not in loadopt.local_opts:
            # nope
            return False

        path = loadopt.path.parent

        # the option applies.  check if the "user_defined_eager_row_processor"
        # has been built up.
        adapter = path.get(
            compile_state.attributes, "user_defined_eager_row_processor", False
        )
        if adapter is not False:
            # just return it
            return adapter

        # otherwise figure it out.
        alias = loadopt.local_opts["eager_from_alias"]
        root_mapper, prop = path[-2:]

        if alias is not None:
            if isinstance(alias, str):
                alias = prop.target.alias(alias)
            adapter = sql_util.ColumnAdapter(
                alias, equivalents=prop.mapper._equivalent_columns
            )
        else:
            if path.contains(
                compile_state.attributes, "path_with_polymorphic"
            ):
                with_poly_entity = path.get(
                    compile_state.attributes, "path_with_polymorphic"
                )
                adapter = orm_util.ORMAdapter(
                    with_poly_entity,
                    equivalents=prop.mapper._equivalent_columns,
                )
            else:
                adapter = compile_state._polymorphic_adapters.get(
                    prop.mapper, None
                )
        path.set(
            target_attributes, "user_defined_eager_row_processor", adapter,
        )

        return adapter

    def _setup_query_on_user_defined_adapter(
        self, context, entity, path, adapter, user_defined_adapter
    ):

        # apply some more wrapping to the "user defined adapter"
        # if we are setting up the query for SQL render.
        adapter = entity._get_entity_clauses(context)

        if adapter and user_defined_adapter:
            user_defined_adapter = user_defined_adapter.wrap(adapter)
            path.set(
                context.attributes,
                "user_defined_eager_row_processor",
                user_defined_adapter,
            )
        elif adapter:
            user_defined_adapter = adapter
            path.set(
                context.attributes,
                "user_defined_eager_row_processor",
                user_defined_adapter,
            )

        add_to_collection = context.primary_columns
        return user_defined_adapter, adapter, add_to_collection

    def _gen_pooled_aliased_class(self, context):
        # keep a local pool of AliasedClass objects that get re-used.
        # we need one unique AliasedClass per query per appearance of our
        # entity in the query.

        if inspect(self.entity).is_aliased_class:
            alt_selectable = inspect(self.entity).selectable
        else:
            alt_selectable = None

        key = ("joinedloader_ac", self)
        if key not in context.attributes:
            context.attributes[key] = idx = 0
        else:
            context.attributes[key] = idx = context.attributes[key] + 1

        if idx >= len(self._aliased_class_pool):
            to_adapt = orm_util.AliasedClass(
                self.mapper,
                alias=alt_selectable.alias(flat=True)
                if alt_selectable is not None
                else None,
                flat=True,
                use_mapper_path=True,
            )

            # load up the .columns collection on the Alias() before
            # the object becomes shared among threads.  this prevents
            # races for column identities.
            inspect(to_adapt).selectable.c
            self._aliased_class_pool.append(to_adapt)

        return self._aliased_class_pool[idx]

    def _generate_row_adapter(
        self,
        compile_state,
        entity,
        path,
        loadopt,
        adapter,
        column_collection,
        parentmapper,
        chained_from_outerjoin,
    ):
        with_poly_entity = path.get(
            compile_state.attributes, "path_with_polymorphic", None
        )
        if with_poly_entity:
            to_adapt = with_poly_entity
        else:
            to_adapt = self._gen_pooled_aliased_class(compile_state)

        clauses = inspect(to_adapt)._memo(
            ("joinedloader_ormadapter", self),
            orm_util.ORMAdapter,
            to_adapt,
            equivalents=self.mapper._equivalent_columns,
            adapt_required=True,
            allow_label_resolve=False,
            anonymize_labels=True,
        )

        assert clauses.aliased_class is not None

        innerjoin = (
            loadopt.local_opts.get("innerjoin", self.parent_property.innerjoin)
            if loadopt is not None
            else self.parent_property.innerjoin
        )

        if not innerjoin:
            # if this is an outer join, all non-nested eager joins from
            # this path must also be outer joins
            chained_from_outerjoin = True

        compile_state.create_eager_joins.append(
            (
                self._create_eager_join,
                entity,
                path,
                adapter,
                parentmapper,
                clauses,
                innerjoin,
                chained_from_outerjoin,
            )
        )

        add_to_collection = compile_state.secondary_columns
        path.set(compile_state.attributes, "eager_row_processor", clauses)

        return clauses, adapter, add_to_collection, chained_from_outerjoin

    def _create_eager_join(
        self,
        compile_state,
        query_entity,
        path,
        adapter,
        parentmapper,
        clauses,
        innerjoin,
        chained_from_outerjoin,
    ):
        if parentmapper is None:
            localparent = query_entity.mapper
        else:
            localparent = parentmapper

        # whether or not the Query will wrap the selectable in a subquery,
        # and then attach eager load joins to that (i.e., in the case of
        # LIMIT/OFFSET etc.)
        should_nest_selectable = (
            compile_state.multi_row_eager_loaders
            and compile_state._should_nest_selectable
        )

        query_entity_key = None

        if (
            query_entity not in compile_state.eager_joins
            and not should_nest_selectable
            and compile_state.from_clauses
        ):
            indexes = sql_util.find_left_clause_that_matches_given(
                compile_state.from_clauses, query_entity.selectable
            )

            if len(indexes) > 1:
                # for the eager load case, I can't reproduce this right
                # now.   For query.join() I can.
                raise sa_exc.InvalidRequestError(
                    "Can't identify which query entity in which to joined "
                    "eager load from.   Please use an exact match when "
                    "specifying the join path."
                )

            if indexes:
                clause = compile_state.from_clauses[indexes[0]]
                # join to an existing FROM clause on the query.
                # key it to its list index in the eager_joins dict.
                # Query._compile_context will adapt as needed and
                # append to the FROM clause of the select().
                query_entity_key, default_towrap = indexes[0], clause

        if query_entity_key is None:
            query_entity_key, default_towrap = (
                query_entity,
                query_entity.selectable,
            )

        towrap = compile_state.eager_joins.setdefault(
            query_entity_key, default_towrap
        )

        if adapter:
            if getattr(adapter, "aliased_class", None):
                # joining from an adapted entity.  The adapted entity
                # might be a "with_polymorphic", so resolve that to our
                # specific mapper's entity before looking for our attribute
                # name on it.
                efm = inspect(adapter.aliased_class)._entity_for_mapper(
                    localparent
                    if localparent.isa(self.parent)
                    else self.parent
                )

                # look for our attribute on the adapted entity, else fall back
                # to our straight property
                onclause = getattr(efm.entity, self.key, self.parent_property)
            else:
                onclause = getattr(
                    orm_util.AliasedClass(
                        self.parent, adapter.selectable, use_mapper_path=True
                    ),
                    self.key,
                    self.parent_property,
                )

        else:
            onclause = self.parent_property

        assert clauses.aliased_class is not None

        attach_on_outside = (
            not chained_from_outerjoin
            or not innerjoin
            or innerjoin == "unnested"
            or query_entity.entity_zero.represents_outer_join
        )

        if attach_on_outside:
            # this is the "classic" eager join case.
            eagerjoin = orm_util._ORMJoin(
                towrap,
                clauses.aliased_class,
                onclause,
                isouter=not innerjoin
                or query_entity.entity_zero.represents_outer_join
                or (chained_from_outerjoin and isinstance(towrap, sql.Join)),
                _left_memo=self.parent,
                _right_memo=self.mapper,
            )
        else:
            # all other cases are innerjoin=='nested' approach
            eagerjoin = self._splice_nested_inner_join(
                path, towrap, clauses, onclause
            )

        compile_state.eager_joins[query_entity_key] = eagerjoin

        # send a hint to the Query as to where it may "splice" this join
        eagerjoin.stop_on = query_entity.selectable

        if not parentmapper:
            # for parentclause that is the non-eager end of the join,
            # ensure all the parent cols in the primaryjoin are actually
            # in the
            # columns clause (i.e. are not deferred), so that aliasing applied
            # by the Query propagates those columns outward.
            # This has the effect
            # of "undefering" those columns.
            for col in sql_util._find_columns(
                self.parent_property.primaryjoin
            ):
                if localparent.persist_selectable.c.contains_column(col):
                    if adapter:
                        col = adapter.columns[col]
                    compile_state.primary_columns.append(col)

        if self.parent_property.order_by:
            compile_state.eager_order_by += tuple(
                (eagerjoin._target_adapter.copy_and_process)(
                    util.to_list(self.parent_property.order_by)
                )
            )

    def _splice_nested_inner_join(
        self, path, join_obj, clauses, onclause, splicing=False
    ):

        if splicing is False:
            # first call is always handed a join object
            # from the outside
            assert isinstance(join_obj, orm_util._ORMJoin)
        elif isinstance(join_obj, sql.selectable.FromGrouping):
            return self._splice_nested_inner_join(
                path, join_obj.element, clauses, onclause, splicing
            )
        elif not isinstance(join_obj, orm_util._ORMJoin):
            if path[-2] is splicing:
                return orm_util._ORMJoin(
                    join_obj,
                    clauses.aliased_class,
                    onclause,
                    isouter=False,
                    _left_memo=splicing,
                    _right_memo=path[-1].mapper,
                )
            else:
                # only here if splicing == True
                return None

        target_join = self._splice_nested_inner_join(
            path, join_obj.right, clauses, onclause, join_obj._right_memo
        )
        if target_join is None:
            right_splice = False
            target_join = self._splice_nested_inner_join(
                path, join_obj.left, clauses, onclause, join_obj._left_memo
            )
            if target_join is None:
                # should only return None when recursively called,
                # e.g. splicing==True
                assert (
                    splicing is not False
                ), "assertion failed attempting to produce joined eager loads"
                return None
        else:
            right_splice = True

        if right_splice:
            # for a right splice, attempt to flatten out
            # a JOIN b JOIN c JOIN .. to avoid needless
            # parenthesis nesting
            if not join_obj.isouter and not target_join.isouter:
                eagerjoin = join_obj._splice_into_center(target_join)
            else:
                eagerjoin = orm_util._ORMJoin(
                    join_obj.left,
                    target_join,
                    join_obj.onclause,
                    isouter=join_obj.isouter,
                    _left_memo=join_obj._left_memo,
                )
        else:
            eagerjoin = orm_util._ORMJoin(
                target_join,
                join_obj.right,
                join_obj.onclause,
                isouter=join_obj.isouter,
                _right_memo=join_obj._right_memo,
            )

        eagerjoin._target_adapter = target_join._target_adapter
        return eagerjoin

    def _create_eager_adapter(self, context, result, adapter, path, loadopt):
        compile_state = context.compile_state

        user_defined_adapter = (
            self._init_user_defined_eager_proc(
                loadopt, compile_state, context.attributes
            )
            if loadopt
            else False
        )

        if user_defined_adapter is not False:
            decorator = user_defined_adapter
            # user defined eagerloads are part of the "primary"
            # portion of the load.
            # the adapters applied to the Query should be honored.
            if compile_state.compound_eager_adapter and decorator:
                decorator = decorator.wrap(
                    compile_state.compound_eager_adapter
                )
            elif compile_state.compound_eager_adapter:
                decorator = compile_state.compound_eager_adapter
        else:
            decorator = path.get(
                compile_state.attributes, "eager_row_processor"
            )
            if decorator is None:
                return False

        if self.mapper._result_has_identity_key(result, decorator):
            return decorator
        else:
            # no identity key - don't return a row
            # processor, will cause a degrade to lazy
            return False

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        if not self.parent.class_manager[self.key].impl.supports_population:
            raise sa_exc.InvalidRequestError(
                "'%s' does not support object "
                "population - eager loading cannot be applied." % self
            )

        if self.uselist:
            context.loaders_require_uniquing = True

        our_path = path[self.parent_property]

        eager_adapter = self._create_eager_adapter(
            context, result, adapter, our_path, loadopt
        )

        if eager_adapter is not False:
            key = self.key

            _instance = loading._instance_processor(
                query_entity,
                self.mapper,
                context,
                result,
                our_path[self.entity],
                eager_adapter,
            )

            if not self.uselist:
                self._create_scalar_loader(context, key, _instance, populators)
            else:
                self._create_collection_loader(
                    context, key, _instance, populators
                )
        else:
            self.parent_property._get_strategy(
                (("lazy", "select"),)
            ).create_row_processor(
                context,
                query_entity,
                path,
                loadopt,
                mapper,
                result,
                adapter,
                populators,
            )

    def _create_collection_loader(self, context, key, _instance, populators):
        def load_collection_from_joined_new_row(state, dict_, row):
            # note this must unconditionally clear out any existing collection.
            # an existing collection would be present only in the case of
            # populate_existing().
            collection = attributes.init_state_collection(state, dict_, key)
            result_list = util.UniqueAppender(
                collection, "append_without_event"
            )
            context.attributes[(state, key)] = result_list
            inst = _instance(row)
            if inst is not None:
                result_list.append(inst)

        def load_collection_from_joined_existing_row(state, dict_, row):
            if (state, key) in context.attributes:
                result_list = context.attributes[(state, key)]
            else:
                # appender_key can be absent from context.attributes
                # with isnew=False when self-referential eager loading
                # is used; the same instance may be present in two
                # distinct sets of result columns
                collection = attributes.init_state_collection(
                    state, dict_, key
                )
                result_list = util.UniqueAppender(
                    collection, "append_without_event"
                )
                context.attributes[(state, key)] = result_list
            inst = _instance(row)
            if inst is not None:
                result_list.append(inst)

        def load_collection_from_joined_exec(state, dict_, row):
            _instance(row)

        populators["new"].append(
            (self.key, load_collection_from_joined_new_row)
        )
        populators["existing"].append(
            (self.key, load_collection_from_joined_existing_row)
        )
        if context.invoke_all_eagers:
            populators["eager"].append(
                (self.key, load_collection_from_joined_exec)
            )

    def _create_scalar_loader(self, context, key, _instance, populators):
        def load_scalar_from_joined_new_row(state, dict_, row):
            # set a scalar object instance directly on the parent
            # object, bypassing InstrumentedAttribute event handlers.
            dict_[key] = _instance(row)

        def load_scalar_from_joined_existing_row(state, dict_, row):
            # call _instance on the row, even though the object has
            # been created, so that we further descend into properties
            existing = _instance(row)

            # conflicting value already loaded, this shouldn't happen
            if key in dict_:
                if existing is not dict_[key]:
                    util.warn(
                        "Multiple rows returned with "
                        "uselist=False for eagerly-loaded attribute '%s' "
                        % self
                    )
            else:
                # this case is when one row has multiple loads of the
                # same entity (e.g. via aliasing), one has an attribute
                # that the other doesn't.
                dict_[key] = existing

        def load_scalar_from_joined_exec(state, dict_, row):
            _instance(row)

        populators["new"].append((self.key, load_scalar_from_joined_new_row))
        populators["existing"].append(
            (self.key, load_scalar_from_joined_existing_row)
        )
        if context.invoke_all_eagers:
            populators["eager"].append(
                (self.key, load_scalar_from_joined_exec)
            )


@log.class_logger
@relationships.RelationshipProperty.strategy_for(lazy="selectin")
class SelectInLoader(PostLoader, util.MemoizedSlots):
    __slots__ = (
        "join_depth",
        "omit_join",
        "_parent_alias",
        "_query_info",
        "_fallback_query_info",
        "_bakery",
    )

    query_info = collections.namedtuple(
        "queryinfo",
        [
            "load_only_child",
            "load_with_join",
            "in_expr",
            "pk_cols",
            "zero_idx",
            "child_lookup_cols",
        ],
    )

    _chunksize = 500

    def __init__(self, parent, strategy_key):
        super(SelectInLoader, self).__init__(parent, strategy_key)
        self.join_depth = self.parent_property.join_depth
        is_m2o = self.parent_property.direction is interfaces.MANYTOONE

        if self.parent_property.omit_join is not None:
            self.omit_join = self.parent_property.omit_join
        else:
            lazyloader = self.parent_property._get_strategy(
                (("lazy", "select"),)
            )
            if is_m2o:
                self.omit_join = lazyloader.use_get
            else:
                self.omit_join = self.parent._get_clause[0].compare(
                    lazyloader._rev_lazywhere,
                    use_proxies=True,
                    equivalents=self.parent._equivalent_columns,
                )

        if self.omit_join:
            if is_m2o:
                self._query_info = self._init_for_omit_join_m2o()
                self._fallback_query_info = self._init_for_join()
            else:
                self._query_info = self._init_for_omit_join()
        else:
            self._query_info = self._init_for_join()

    def _init_for_omit_join(self):
        pk_to_fk = dict(
            self.parent_property._join_condition.local_remote_pairs
        )
        pk_to_fk.update(
            (equiv, pk_to_fk[k])
            for k in list(pk_to_fk)
            for equiv in self.parent._equivalent_columns.get(k, ())
        )

        pk_cols = fk_cols = [
            pk_to_fk[col] for col in self.parent.primary_key if col in pk_to_fk
        ]
        if len(fk_cols) > 1:
            in_expr = sql.tuple_(*fk_cols)
            zero_idx = False
        else:
            in_expr = fk_cols[0]
            zero_idx = True

        return self.query_info(False, False, in_expr, pk_cols, zero_idx, None)

    def _init_for_omit_join_m2o(self):
        pk_cols = self.mapper.primary_key
        if len(pk_cols) > 1:
            in_expr = sql.tuple_(*pk_cols)
            zero_idx = False
        else:
            in_expr = pk_cols[0]
            zero_idx = True

        lazyloader = self.parent_property._get_strategy((("lazy", "select"),))
        lookup_cols = [lazyloader._equated_columns[pk] for pk in pk_cols]

        return self.query_info(
            True, False, in_expr, pk_cols, zero_idx, lookup_cols
        )

    def _init_for_join(self):
        self._parent_alias = aliased(self.parent.class_)
        pa_insp = inspect(self._parent_alias)
        pk_cols = [
            pa_insp._adapt_element(col) for col in self.parent.primary_key
        ]
        if len(pk_cols) > 1:
            in_expr = sql.tuple_(*pk_cols)
            zero_idx = False
        else:
            in_expr = pk_cols[0]
            zero_idx = True
        return self.query_info(False, True, in_expr, pk_cols, zero_idx, None)

    def init_class_attribute(self, mapper):
        self.parent_property._get_strategy(
            (("lazy", "select"),)
        ).init_class_attribute(mapper)

    @util.preload_module("sqlalchemy.ext.baked")
    def _memoized_attr__bakery(self):
        return util.preloaded.ext_baked.bakery(size=50)

    def create_row_processor(
        self,
        context,
        query_entity,
        path,
        loadopt,
        mapper,
        result,
        adapter,
        populators,
    ):
        if context.refresh_state:
            return self._immediateload_create_row_processor(
                context,
                query_entity,
                path,
                loadopt,
                mapper,
                result,
                adapter,
                populators,
            )

        if not self.parent.class_manager[self.key].impl.supports_population:
            raise sa_exc.InvalidRequestError(
                "'%s' does not support object "
                "population - eager loading cannot be applied." % self
            )

        # a little dance here as the "path" is still something that only
        # semi-tracks the exact series of things we are loading, still not
        # telling us about with_polymorphic() and stuff like that when it's at
        # the root..  the initial MapperEntity is more accurate for this case.
        if len(path) == 1:
            if not orm_util._entity_isa(query_entity.entity_zero, self.parent):
                return
        elif not orm_util._entity_isa(path[-1], self.parent):
            return

        selectin_path = (
            context.compile_state.current_path or orm_util.PathRegistry.root
        ) + path

        if loading.PostLoad.path_exists(
            context, selectin_path, self.parent_property
        ):
            return

        path_w_prop = path[self.parent_property]
        selectin_path_w_prop = selectin_path[self.parent_property]

        # build up a path indicating the path from the leftmost
        # entity to the thing we're subquery loading.
        with_poly_entity = path_w_prop.get(
            context.attributes, "path_with_polymorphic", None
        )

        if with_poly_entity is not None:
            effective_entity = with_poly_entity
        else:
            effective_entity = self.entity

        if not path_w_prop.contains(context.attributes, "loader"):
            if self.join_depth:
                if selectin_path_w_prop.length / 2 > self.join_depth:
                    return
            elif selectin_path_w_prop.contains_mapper(self.mapper):
                return
        loading.PostLoad.callable_for_path(
            context,
            selectin_path,
            self.parent,
            self.parent_property,
            self._load_for_path,
            effective_entity,
        )

    def _load_for_path(
        self, context, path, states, load_only, effective_entity
    ):
        if load_only and self.key not in load_only:
            return

        query_info = self._query_info

        if query_info.load_only_child:
            our_states = collections.defaultdict(list)
            none_states = []

            mapper = self.parent

            for state, overwrite in states:
                state_dict = state.dict
                related_ident = tuple(
                    mapper._get_state_attr_by_column(
                        state,
                        state_dict,
                        lk,
                        passive=attributes.PASSIVE_NO_FETCH,
                    )
                    for lk in query_info.child_lookup_cols
                )
                # if the loaded parent objects do not have the foreign key
                # to the related item loaded, then degrade into the joined
                # version of selectinload
                if attributes.PASSIVE_NO_RESULT in related_ident:
                    query_info = self._fallback_query_info
                    break

                # organize states into lists keyed to particular foreign
                # key values.
                if None not in related_ident:
                    our_states[related_ident].append(
                        (state, state_dict, overwrite)
                    )
                else:
                    # For FK values that have None, add them to a
                    # separate collection that will be populated separately
                    none_states.append((state, state_dict, overwrite))

        # note the above conditional may have changed query_info
        if not query_info.load_only_child:
            our_states = [
                (state.key[1], state, state.dict, overwrite)
                for state, overwrite in states
            ]

        pk_cols = query_info.pk_cols
        in_expr = query_info.in_expr

        if not query_info.load_with_join:
            # in "omit join" mode, the primary key column and the
            # "in" expression are in terms of the related entity.  So
            # if the related entity is polymorphic or otherwise aliased,
            # we need to adapt our "pk_cols" and "in_expr" to that
            # entity.   in non-"omit join" mode, these are against the
            # parent entity and do not need adaption.
            insp = inspect(effective_entity)
            if insp.is_aliased_class:
                pk_cols = [insp._adapt_element(col) for col in pk_cols]
                in_expr = insp._adapt_element(in_expr)
                pk_cols = [insp._adapt_element(col) for col in pk_cols]

        q = self._bakery(
            lambda session: session.query(
                orm_util.Bundle("pk", *pk_cols), effective_entity
            ),
            self,
        )

        if not query_info.load_with_join:
            # the Bundle we have in the "omit_join" case is against raw, non
            # annotated columns, so to ensure the Query knows its primary
            # entity, we add it explicitly.  If we made the Bundle against
            # annotated columns, we hit a performance issue in this specific
            # case, which is detailed in issue #4347.
            q.add_criteria(lambda q: q.select_from(effective_entity))
        else:
            # in the non-omit_join case, the Bundle is against the annotated/
            # mapped column of the parent entity, but the #4347 issue does not
            # occur in this case.
            pa = self._parent_alias
            q.add_criteria(
                lambda q: q.select_from(pa).join(
                    getattr(pa, self.parent_property.key).of_type(
                        effective_entity
                    )
                )
            )

        if query_info.load_only_child:
            q.add_criteria(
                lambda q: q.filter(
                    in_expr.in_(sql.bindparam("primary_keys", expanding=True))
                )
            )
        else:
            q.add_criteria(
                lambda q: q.filter(
                    in_expr.in_(sql.bindparam("primary_keys", expanding=True))
                )
            )

        # a test which exercises what these comments talk about is
        # test_selectin_relations.py -> test_twolevel_selectin_w_polymorphic
        #
        # effective_entity above is given to us in terms of the cached
        # statement, namely this one:
        orig_query = context.compile_state.select_statement

        # the actual statement that was requested is this one:
        #  context_query = context.query
        #
        # that's not the cached one, however.  So while it is of the identical
        # structure, if it has entities like AliasedInsp, which we get from
        # aliased() or with_polymorphic(), the AliasedInsp will likely be a
        # different object identity each time, and will not match up
        # hashing-wise to the corresponding AliasedInsp that's in the
        # cached query, meaning it won't match on paths and loader lookups
        # and loaders like this one will be skipped if it is used in options.
        #
        # Now we want to transfer loader options from the parent query to the
        # "selectinload" query we're about to run.   Which query do we transfer
        # the options from?  We use the cached query, because the options in
        # that query will be in terms of the effective entity we were just
        # handed.
        #
        # But now the selectinload/ baked query we are running is *also*
        # cached.  What if it's cached and running from some previous iteration
        # of that AliasedInsp?  Well in that case it will also use the previous
        # iteration of the loader options.   If the baked query expires and
        # gets generated again, it will be handed the current effective_entity
        # and the current _with_options, again in terms of whatever
        # compile_state.select_statement happens to be right now, so the
        # query will still be internally consistent and loader callables
        # will be correctly invoked.

        q._add_lazyload_options(
            orig_query._with_options, path[self.parent_property]
        )

        if context.populate_existing:
            q.add_criteria(lambda q: q.populate_existing())

        if self.parent_property.order_by:
            if not query_info.load_with_join:
                eager_order_by = self.parent_property.order_by
                if insp.is_aliased_class:
                    eager_order_by = [
                        insp._adapt_element(elem) for elem in eager_order_by
                    ]
                q.add_criteria(lambda q: q.order_by(*eager_order_by))
            else:

                def _setup_outermost_orderby(compile_context):
                    compile_context.eager_order_by += tuple(
                        util.to_list(self.parent_property.order_by)
                    )

                q.add_criteria(
                    lambda q: q._add_context_option(
                        _setup_outermost_orderby, self.parent_property
                    )
                )

        if query_info.load_only_child:
            self._load_via_child(
                our_states, none_states, query_info, q, context
            )
        else:
            self._load_via_parent(our_states, query_info, q, context)

    def _load_via_child(self, our_states, none_states, query_info, q, context):
        uselist = self.uselist

        # this sort is really for the benefit of the unit tests
        our_keys = sorted(our_states)
        while our_keys:
            chunk = our_keys[0 : self._chunksize]
            our_keys = our_keys[self._chunksize :]
            data = {
                k: v
                for k, v in q(context.session).params(
                    primary_keys=[
                        key[0] if query_info.zero_idx else key for key in chunk
                    ]
                )
            }

            for key in chunk:
                # for a real foreign key and no concurrent changes to the
                # DB while running this method, "key" is always present in
                # data.  However, for primaryjoins without real foreign keys
                # a non-None primaryjoin condition may still refer to no
                # related object.
                related_obj = data.get(key, None)
                for state, dict_, overwrite in our_states[key]:
                    if not overwrite and self.key in dict_:
                        continue

                    state.get_impl(self.key).set_committed_value(
                        state,
                        dict_,
                        related_obj if not uselist else [related_obj],
                    )
        # populate none states with empty value / collection
        for state, dict_, overwrite in none_states:
            if not overwrite and self.key in dict_:
                continue

            # note it's OK if this is a uselist=True attribute, the empty
            # collection will be populated
            state.get_impl(self.key).set_committed_value(state, dict_, None)

    def _load_via_parent(self, our_states, query_info, q, context):
        uselist = self.uselist
        _empty_result = () if uselist else None

        while our_states:
            chunk = our_states[0 : self._chunksize]
            our_states = our_states[self._chunksize :]

            primary_keys = [
                key[0] if query_info.zero_idx else key
                for key, state, state_dict, overwrite in chunk
            ]

            data = collections.defaultdict(list)
            for k, v in itertools.groupby(
                q(context.session).params(primary_keys=primary_keys),
                lambda x: x[0],
            ):
                data[k].extend(vv[1] for vv in v)

            for key, state, state_dict, overwrite in chunk:

                if not overwrite and self.key in state_dict:
                    continue

                collection = data.get(key, _empty_result)

                if not uselist and collection:
                    if len(collection) > 1:
                        util.warn(
                            "Multiple rows returned with "
                            "uselist=False for eagerly-loaded "
                            "attribute '%s' " % self
                        )
                    state.get_impl(self.key).set_committed_value(
                        state, state_dict, collection[0]
                    )
                else:
                    # note that empty tuple set on uselist=False sets the
                    # value to None
                    state.get_impl(self.key).set_committed_value(
                        state, state_dict, collection
                    )


def single_parent_validator(desc, prop):
    def _do_check(state, value, oldvalue, initiator):
        if value is not None and initiator.key == prop.key:
            hasparent = initiator.hasparent(attributes.instance_state(value))
            if hasparent and oldvalue is not value:
                raise sa_exc.InvalidRequestError(
                    "Instance %s is already associated with an instance "
                    "of %s via its %s attribute, and is only allowed a "
                    "single parent."
                    % (orm_util.instance_str(value), state.class_, prop),
                    code="bbf1",
                )
        return value

    def append(state, value, initiator):
        return _do_check(state, value, None, initiator)

    def set_(state, value, oldvalue, initiator):
        return _do_check(state, value, oldvalue, initiator)

    event.listen(
        desc, "append", append, raw=True, retval=True, active_history=True
    )
    event.listen(desc, "set", set_, raw=True, retval=True, active_history=True)