query.py

"""
Create SQL statements for QuerySets.

The code in here encapsulates all of the SQL construction so that QuerySets
themselves do not have to (and could be backed by things other than SQL
databases). The abstraction barrier only works one way: this module has to know
all about the internals of models in order to get the information it needs.
"""
from collections import Counter, Iterator, Mapping, OrderedDict, namedtuple
from itertools import chain, count, product
from string import ascii_uppercase

from django.core.exceptions import (
    EmptyResultSet, FieldDoesNotExist, FieldError,
)
from django.db import DEFAULT_DB_ALIAS, NotSupportedError, connections
from django.db.models.aggregates import Count
from django.db.models.constants import LOOKUP_SEP
from django.db.models.expressions import Col, Ref
from django.db.models.fields.related_lookups import MultiColSource
from django.db.models.lookups import Lookup
from django.db.models.query_utils import (
    Q, check_rel_lookup_compatibility, refs_expression,
)
from django.db.models.sql.constants import (
    INNER, LOUTER, ORDER_DIR, ORDER_PATTERN, SINGLE,
)
from django.db.models.sql.datastructures import (
    BaseTable, Empty, Join, MultiJoin,
)
from django.db.models.sql.where import (
    AND, OR, ExtraWhere, NothingNode, WhereNode,
)
from django.utils.encoding import force_text
from django.utils.functional import cached_property
from django.utils.tree import Node

__all__ = ['Query', 'RawQuery']


def get_field_names_from_opts(opts):
    return set(chain.from_iterable(
        (f.name, f.attname) if f.concrete else (f.name,)
        for f in opts.get_fields()
    ))


def get_children_from_q(q):
    for child in q.children:
        if isinstance(child, Node):
            yield from get_children_from_q(child)
        else:
            yield child


JoinInfo = namedtuple(
    'JoinInfo',
    ('final_field', 'targets', 'opts', 'joins', 'path')
)


class RawQuery:
    """A single raw SQL query."""

    def __init__(self, sql, using, params=None):
        self.params = params or ()
        self.sql = sql
        self.using = using
        self.cursor = None

        # Mirror some properties of a normal query so that
        # the compiler can be used to process results.
        self.low_mark, self.high_mark = 0, None  # Used for offset/limit
        self.extra_select = {}
        self.annotation_select = {}

    def chain(self, using):
        return self.clone(using)

    def clone(self, using):
        return RawQuery(self.sql, using, params=self.params)

    def get_columns(self):
        if self.cursor is None:
            self._execute_query()
        converter = connections[self.using].introspection.column_name_converter
        return [converter(column_meta[0])
                for column_meta in self.cursor.description]

    def __iter__(self):
        # Always execute a new query for a new iterator.
        # This could be optimized with a cache at the expense of RAM.
        self._execute_query()
        if not connections[self.using].features.can_use_chunked_reads:
            # If the database can't use chunked reads we need to make sure we
            # evaluate the entire query up front.
            result = list(self.cursor)
        else:
            result = self.cursor
        return iter(result)

    def __repr__(self):
        return "<%s: %s>" % (self.__class__.__name__, self)

    @property
    def params_type(self):
        return dict if isinstance(self.params, Mapping) else tuple

    def __str__(self):
        return self.sql % self.params_type(self.params)

    def _execute_query(self):
        connection = connections[self.using]

        # Adapt parameters to the database, as much as possible considering
        # that the target type isn't known. See #17755.
        params_type = self.params_type
        adapter = connection.ops.adapt_unknown_value
        if params_type is tuple:
            params = tuple(adapter(val) for val in self.params)
        elif params_type is dict:
            params = {key: adapter(val) for key, val in self.params.items()}
        else:
            raise RuntimeError("Unexpected params type: %s" % params_type)

        self.cursor = connection.cursor()
        self.cursor.execute(self.sql, params)


class Query:
    """A single SQL query."""

    alias_prefix = 'T'
    subq_aliases = frozenset([alias_prefix])

    compiler = 'SQLCompiler'

    def __init__(self, model, where=WhereNode):
        self.model = model
        self.alias_refcount = {}
        # alias_map is the most important data structure regarding joins.
        # It's used for recording which joins exist in the query and what
        # types they are. The key is the alias of the joined table (possibly
        # the table name) and the value is a Join-like object (see
        # sql.datastructures.Join for more information).
        self.alias_map = OrderedDict()
        # Sometimes the query contains references to aliases in outer queries (as
        # a result of split_exclude). Correct alias quoting needs to know these
        # aliases too.
        self.external_aliases = set()
        self.table_map = {}     # Maps table names to list of aliases.
        self.default_cols = True
        self.default_ordering = True
        self.standard_ordering = True
        self.used_aliases = set()
        self.filter_is_sticky = False
        self.subquery = False

        # SQL-related attributes
        # Select and related select clauses are expressions to use in the
        # SELECT clause of the query.
        # The select is used for cases where we want to set up the select
        # clause to contain other than default fields (values(), subqueries...)
        # Note that annotations go to annotations dictionary.
        self.select = ()
        self.where = where()
        self.where_class = where
        # The group_by attribute can have one of the following forms:
        #  - None: no group by at all in the query
        #  - A tuple of expressions: group by (at least) those expressions.
        #    String refs are also allowed for now.
        #  - True: group by all select fields of the model
        # See compiler.get_group_by() for details.
        self.group_by = None
        self.order_by = ()
        self.low_mark, self.high_mark = 0, None  # Used for offset/limit
        self.distinct = False
        self.distinct_fields = ()
        self.select_for_update = False
        self.select_for_update_nowait = False
        self.select_for_update_skip_locked = False
        self.select_for_update_of = ()

        self.select_related = False
        # Arbitrary limit for select_related to prevents infinite recursion.
        self.max_depth = 5

        # Holds the selects defined by a call to values() or values_list()
        # excluding annotation_select and extra_select.
        self.values_select = ()

        # SQL annotation-related attributes
        # The _annotations will be an OrderedDict when used. Due to the cost
        # of creating OrderedDict this attribute is created lazily (in
        # self.annotations property).
        self._annotations = None  # Maps alias -> Annotation Expression
        self.annotation_select_mask = None
        self._annotation_select_cache = None

        # Set combination attributes
        self.combinator = None
        self.combinator_all = False
        self.combined_queries = ()

        # These are for extensions. The contents are more or less appended
        # verbatim to the appropriate clause.
        # The _extra attribute is an OrderedDict, lazily created similarly to
        # .annotations
        self._extra = None  # Maps col_alias -> (col_sql, params).
        self.extra_select_mask = None
        self._extra_select_cache = None

        self.extra_tables = ()
        self.extra_order_by = ()

        # A tuple that is a set of model field names and either True, if these
        # are the fields to defer, or False if these are the only fields to
        # load.
        self.deferred_loading = (frozenset(), True)

        self._filtered_relations = {}

    @property
    def extra(self):
        if self._extra is None:
            self._extra = OrderedDict()
        return self._extra

    @property
    def annotations(self):
        if self._annotations is None:
            self._annotations = OrderedDict()
        return self._annotations

    @property
    def has_select_fields(self):
        return bool(self.select or self.annotation_select_mask or self.extra_select_mask)

    @cached_property
    def base_table(self):
        for alias in self.alias_map:
            return alias

    def __str__(self):
        """
        Return the query as a string of SQL with the parameter values
        substituted in (use sql_with_params() to see the unsubstituted string).

        Parameter values won't necessarily be quoted correctly, since that is
        done by the database interface at execution time.
        """
        sql, params = self.sql_with_params()
        return sql % params

    def sql_with_params(self):
        """
        Return the query as an SQL string and the parameters that will be
        substituted into the query.
        """
        return self.get_compiler(DEFAULT_DB_ALIAS).as_sql()

    def __deepcopy__(self, memo):
        """Limit the amount of work when a Query is deepcopied."""
        result = self.clone()
        memo[id(self)] = result
        return result

    def _prepare(self, field):
        return self

    def get_compiler(self, using=None, connection=None):
        if using is None and connection is None:
            raise ValueError("Need either using or connection")
        if using:
            connection = connections[using]
        return connection.ops.compiler(self.compiler)(self, connection, using)

    def get_meta(self):
        """
        Return the Options instance (the model._meta) from which to start
        processing. Normally, this is self.model._meta, but it can be changed
        by subclasses.
        """
        return self.model._meta

    def clone(self):
        """
        Return a copy of the current Query. A lightweight alternative to
        to deepcopy().
        """
        obj = Empty()
        obj.__class__ = self.__class__
        # Copy references to everything.
        obj.__dict__ = self.__dict__.copy()
        # Clone attributes that can't use shallow copy.
        obj.alias_refcount = self.alias_refcount.copy()
        obj.alias_map = self.alias_map.copy()
        obj.external_aliases = self.external_aliases.copy()
        obj.table_map = self.table_map.copy()
        obj.where = self.where.clone()
        obj._annotations = self._annotations.copy() if self._annotations is not None else None
        if self.annotation_select_mask is None:
            obj.annotation_select_mask = None
        else:
            obj.annotation_select_mask = self.annotation_select_mask.copy()
        # _annotation_select_cache cannot be copied, as doing so breaks the
        # (necessary) state in which both annotations and
        # _annotation_select_cache point to the same underlying objects.
        # It will get re-populated in the cloned queryset the next time it's
        # used.
        obj._annotation_select_cache = None
        obj._extra = self._extra.copy() if self._extra is not None else None
        if self.extra_select_mask is None:
            obj.extra_select_mask = None
        else:
            obj.extra_select_mask = self.extra_select_mask.copy()
        if self._extra_select_cache is None:
            obj._extra_select_cache = None
        else:
            obj._extra_select_cache = self._extra_select_cache.copy()
        if 'subq_aliases' in self.__dict__:
            obj.subq_aliases = self.subq_aliases.copy()
        obj.used_aliases = self.used_aliases.copy()
        obj._filtered_relations = self._filtered_relations.copy()
        # Clear the cached_property
        try:
            del obj.base_table
        except AttributeError:
            pass
        return obj

    def chain(self, klass=None):
        """
        Return a copy of the current Query that's ready for another operation.
        The klass argument changes the type of the Query, e.g. UpdateQuery.
        """
        obj = self.clone()
        if klass and obj.__class__ != klass:
            obj.__class__ = klass
        if not obj.filter_is_sticky:
            obj.used_aliases = set()
        obj.filter_is_sticky = False
        if hasattr(obj, '_setup_query'):
            obj._setup_query()
        return obj

    def relabeled_clone(self, change_map):
        clone = self.clone()
        clone.change_aliases(change_map)
        return clone

    def rewrite_cols(self, annotation, col_cnt):
        # We must make sure the inner query has the referred columns in it.
        # If we are aggregating over an annotation, then Django uses Ref()
        # instances to note this. However, if we are annotating over a column
        # of a related model, then it might be that column isn't part of the
        # SELECT clause of the inner query, and we must manually make sure
        # the column is selected. An example case is:
        #    .aggregate(Sum('author__awards'))
        # Resolving this expression results in a join to author, but there
        # is no guarantee the awards column of author is in the select clause
        # of the query. Thus we must manually add the column to the inner
        # query.
        orig_exprs = annotation.get_source_expressions()
        new_exprs = []
        for expr in orig_exprs:
            # FIXME: These conditions are fairly arbitrary. Identify a better
            # method of having expressions decide which code path they should
            # take.
            if isinstance(expr, Ref):
                # Its already a Ref to subquery (see resolve_ref() for
                # details)
                new_exprs.append(expr)
            elif isinstance(expr, (WhereNode, Lookup)):
                # Decompose the subexpressions further. The code here is
                # copied from the else clause, but this condition must appear
                # before the contains_aggregate/is_summary condition below.
                new_expr, col_cnt = self.rewrite_cols(expr, col_cnt)
                new_exprs.append(new_expr)
            elif isinstance(expr, Col) or (expr.contains_aggregate and not expr.is_summary):
                # Reference to column. Make sure the referenced column
                # is selected.
                col_cnt += 1
                col_alias = '__col%d' % col_cnt
                self.annotations[col_alias] = expr
                self.append_annotation_mask([col_alias])
                new_exprs.append(Ref(col_alias, expr))
            else:
                # Some other expression not referencing database values
                # directly. Its subexpression might contain Cols.
                new_expr, col_cnt = self.rewrite_cols(expr, col_cnt)
                new_exprs.append(new_expr)
        annotation.set_source_expressions(new_exprs)
        return annotation, col_cnt

    def get_aggregation(self, using, added_aggregate_names):
        """
        Return the dictionary with the values of the existing aggregations.
        """
        if not self.annotation_select:
            return {}
        has_limit = self.low_mark != 0 or self.high_mark is not None
        has_existing_annotations = any(
            annotation for alias, annotation
            in self.annotations.items()
            if alias not in added_aggregate_names
        )
        # Decide if we need to use a subquery.
        #
        # Existing annotations would cause incorrect results as get_aggregation()
        # must produce just one result and thus must not use GROUP BY. But we
        # aren't smart enough to remove the existing annotations from the
        # query, so those would force us to use GROUP BY.
        #
        # If the query has limit or distinct, or uses set operations, then
        # those operations must be done in a subquery so that the query
        # aggregates on the limit and/or distinct results instead of applying
        # the distinct and limit after the aggregation.
        if (isinstance(self.group_by, tuple) or has_limit or has_existing_annotations or
                self.distinct or self.combinator):
            from django.db.models.sql.subqueries import AggregateQuery
            outer_query = AggregateQuery(self.model)
            inner_query = self.clone()
            inner_query.select_for_update = False
            inner_query.select_related = False
            if not has_limit and not self.distinct_fields:
                # Queries with distinct_fields need ordering and when a limit
                # is applied we must take the slice from the ordered query.
                # Otherwise no need for ordering.
                inner_query.clear_ordering(True)
            if not inner_query.distinct:
                # If the inner query uses default select and it has some
                # aggregate annotations, then we must make sure the inner
                # query is grouped by the main model's primary key. However,
                # clearing the select clause can alter results if distinct is
                # used.
                if inner_query.default_cols and has_existing_annotations:
                    inner_query.group_by = (self.model._meta.pk.get_col(inner_query.get_initial_alias()),)
                inner_query.default_cols = False

            relabels = {t: 'subquery' for t in inner_query.alias_map}
            relabels[None] = 'subquery'
            # Remove any aggregates marked for reduction from the subquery
            # and move them to the outer AggregateQuery.
            col_cnt = 0
            for alias, expression in list(inner_query.annotation_select.items()):
                if expression.is_summary:
                    expression, col_cnt = inner_query.rewrite_cols(expression, col_cnt)
                    outer_query.annotations[alias] = expression.relabeled_clone(relabels)
                    del inner_query.annotations[alias]
                # Make sure the annotation_select wont use cached results.
                inner_query.set_annotation_mask(inner_query.annotation_select_mask)
            if inner_query.select == () and not inner_query.default_cols and not inner_query.annotation_select_mask:
                # In case of Model.objects[0:3].count(), there would be no
                # field selected in the inner query, yet we must use a subquery.
                # So, make sure at least one field is selected.
                inner_query.select = (self.model._meta.pk.get_col(inner_query.get_initial_alias()),)
            try:
                outer_query.add_subquery(inner_query, using)
            except EmptyResultSet:
                return {
                    alias: None
                    for alias in outer_query.annotation_select
                }
        else:
            outer_query = self
            self.select = ()
            self.default_cols = False
            self._extra = {}

        outer_query.clear_ordering(True)
        outer_query.clear_limits()
        outer_query.select_for_update = False
        outer_query.select_related = False
        compiler = outer_query.get_compiler(using)
        result = compiler.execute_sql(SINGLE)
        if result is None:
            result = [None] * len(outer_query.annotation_select)

        converters = compiler.get_converters(outer_query.annotation_select.values())
        result = next(compiler.apply_converters((result,), converters))

        return dict(zip(outer_query.annotation_select, result))

    def get_count(self, using):
        """
        Perform a COUNT() query using the current filter constraints.
        """
        obj = self.clone()
        obj.add_annotation(Count('*'), alias='__count', is_summary=True)
        number = obj.get_aggregation(using, ['__count'])['__count']
        if number is None:
            number = 0
        return number

    def has_filters(self):
        return self.where

    def has_results(self, using):
        q = self.clone()
        if not q.distinct:
            if q.group_by is True:
                q.add_fields((f.attname for f in self.model._meta.concrete_fields), False)
                q.set_group_by()
            q.clear_select_clause()
        q.clear_ordering(True)
        q.set_limits(high=1)
        compiler = q.get_compiler(using=using)
        return compiler.has_results()

    def combine(self, rhs, connector):
        """
        Merge the 'rhs' query into the current one (with any 'rhs' effects
        being applied *after* (that is, "to the right of") anything in the
        current query. 'rhs' is not modified during a call to this function.

        The 'connector' parameter describes how to connect filters from the
        'rhs' query.
        """
        assert self.model == rhs.model, \
            "Cannot combine queries on two different base models."
        assert self.can_filter(), \
            "Cannot combine queries once a slice has been taken."
        assert self.distinct == rhs.distinct, \
            "Cannot combine a unique query with a non-unique query."
        assert self.distinct_fields == rhs.distinct_fields, \
            "Cannot combine queries with different distinct fields."

        # Work out how to relabel the rhs aliases, if necessary.
        change_map = {}
        conjunction = (connector == AND)

        # Determine which existing joins can be reused. When combining the
        # query with AND we must recreate all joins for m2m filters. When
        # combining with OR we can reuse joins. The reason is that in AND
        # case a single row can't fulfill a condition like:
        #     revrel__col=1 & revrel__col=2
        # But, there might be two different related rows matching this
        # condition. In OR case a single True is enough, so single row is
        # enough, too.
        #
        # Note that we will be creating duplicate joins for non-m2m joins in
        # the AND case. The results will be correct but this creates too many
        # joins. This is something that could be fixed later on.
        reuse = set() if conjunction else set(self.alias_map)
        # Base table must be present in the query - this is the same
        # table on both sides.
        self.get_initial_alias()
        joinpromoter = JoinPromoter(connector, 2, False)
        joinpromoter.add_votes(
            j for j in self.alias_map if self.alias_map[j].join_type == INNER)
        rhs_votes = set()
        # Now, add the joins from rhs query into the new query (skipping base
        # table).
        rhs_tables = list(rhs.alias_map)[1:]
        for alias in rhs_tables:
            join = rhs.alias_map[alias]
            # If the left side of the join was already relabeled, use the
            # updated alias.
            join = join.relabeled_clone(change_map)
            new_alias = self.join(join, reuse=reuse)
            if join.join_type == INNER:
                rhs_votes.add(new_alias)
            # We can't reuse the same join again in the query. If we have two
            # distinct joins for the same connection in rhs query, then the
            # combined query must have two joins, too.
            reuse.discard(new_alias)
            if alias != new_alias:
                change_map[alias] = new_alias
            if not rhs.alias_refcount[alias]:
                # The alias was unused in the rhs query. Unref it so that it
                # will be unused in the new query, too. We have to add and
                # unref the alias so that join promotion has information of
                # the join type for the unused alias.
                self.unref_alias(new_alias)
        joinpromoter.add_votes(rhs_votes)
        joinpromoter.update_join_types(self)

        # Now relabel a copy of the rhs where-clause and add it to the current
        # one.
        w = rhs.where.clone()
        w.relabel_aliases(change_map)
        self.where.add(w, connector)

        # Selection columns and extra extensions are those provided by 'rhs'.
        if rhs.select:
            self.set_select([col.relabeled_clone(change_map) for col in rhs.select])
        else:
            self.select = ()

        if connector == OR:
            # It would be nice to be able to handle this, but the queries don't
            # really make sense (or return consistent value sets). Not worth
            # the extra complexity when you can write a real query instead.
            if self._extra and rhs._extra:
                raise ValueError("When merging querysets using 'or', you cannot have extra(select=...) on both sides.")
        self.extra.update(rhs.extra)
        extra_select_mask = set()
        if self.extra_select_mask is not None:
            extra_select_mask.update(self.extra_select_mask)
        if rhs.extra_select_mask is not None:
            extra_select_mask.update(rhs.extra_select_mask)
        if extra_select_mask:
            self.set_extra_mask(extra_select_mask)
        self.extra_tables += rhs.extra_tables

        # Ordering uses the 'rhs' ordering, unless it has none, in which case
        # the current ordering is used.
        self.order_by = rhs.order_by if rhs.order_by else self.order_by
        self.extra_order_by = rhs.extra_order_by or self.extra_order_by

    def deferred_to_data(self, target, callback):
        """
        Convert the self.deferred_loading data structure to an alternate data
        structure, describing the field that *will* be loaded. This is used to
        compute the columns to select from the database and also by the
        QuerySet class to work out which fields are being initialized on each
        model. Models that have all their fields included aren't mentioned in
        the result, only those that have field restrictions in place.

        The "target" parameter is the instance that is populated (in place).
        The "callback" is a function that is called whenever a (model, field)
        pair need to be added to "target". It accepts three parameters:
        "target", and the model and list of fields being added for that model.
        """
        field_names, defer = self.deferred_loading
        if not field_names:
            return
        orig_opts = self.get_meta()
        seen = {}
        must_include = {orig_opts.concrete_model: {orig_opts.pk}}
        for field_name in field_names:
            parts = field_name.split(LOOKUP_SEP)
            cur_model = self.model._meta.concrete_model
            opts = orig_opts
            for name in parts[:-1]:
                old_model = cur_model
                if name in self._filtered_relations:
                    name = self._filtered_relations[name].relation_name
                source = opts.get_field(name)
                if is_reverse_o2o(source):
                    cur_model = source.related_model
                else:
                    cur_model = source.remote_field.model
                opts = cur_model._meta
                # Even if we're "just passing through" this model, we must add
                # both the current model's pk and the related reference field
                # (if it's not a reverse relation) to the things we select.
                if not is_reverse_o2o(source):
                    must_include[old_model].add(source)
                add_to_dict(must_include, cur_model, opts.pk)
            field = opts.get_field(parts[-1])
            is_reverse_object = field.auto_created and not field.concrete
            model = field.related_model if is_reverse_object else field.model
            model = model._meta.concrete_model
            if model == opts.model:
                model = cur_model
            if not is_reverse_o2o(field):
                add_to_dict(seen, model, field)

        if defer:
            # We need to load all fields for each model, except those that
            # appear in "seen" (for all models that appear in "seen"). The only
            # slight complexity here is handling fields that exist on parent
            # models.
            workset = {}
            for model, values in seen.items():
                for field in model._meta.local_fields:
                    if field in values:
                        continue
                    m = field.model._meta.concrete_model
                    add_to_dict(workset, m, field)
            for model, values in must_include.items():
                # If we haven't included a model in workset, we don't add the
                # corresponding must_include fields for that model, since an
                # empty set means "include all fields". That's why there's no
                # "else" branch here.
                if model in workset:
                    workset[model].update(values)
            for model, values in workset.items():
                callback(target, model, values)
        else:
            for model, values in must_include.items():
                if model in seen:
                    seen[model].update(values)
                else:
                    # As we've passed through this model, but not explicitly
                    # included any fields, we have to make sure it's mentioned
                    # so that only the "must include" fields are pulled in.
                    seen[model] = values
            # Now ensure that every model in the inheritance chain is mentioned
            # in the parent list. Again, it must be mentioned to ensure that
            # only "must include" fields are pulled in.
            for model in orig_opts.get_parent_list():
                if model not in seen:
                    seen[model] = set()
            for model, values in seen.items():
                callback(target, model, values)

    def table_alias(self, table_name, create=False, filtered_relation=None):
        """
        Return a table alias for the given table_name and whether this is a
        new alias or not.

        If 'create' is true, a new alias is always created. Otherwise, the
        most recently created alias for the table (if one exists) is reused.
        """
        alias_list = self.table_map.get(table_name)
        if not create and alias_list:
            alias = alias_list[0]
            self.alias_refcount[alias] += 1
            return alias, False

        # Create a new alias for this table.
        if alias_list:
            alias = '%s%d' % (self.alias_prefix, len(self.alias_map) + 1)
            alias_list.append(alias)
        else:
            # The first occurrence of a table uses the table name directly.
            alias = filtered_relation.alias if filtered_relation is not None else table_name
            self.table_map[table_name] = [alias]
        self.alias_refcount[alias] = 1
        return alias, True

    def ref_alias(self, alias):
        """Increases the reference count for this alias."""
        self.alias_refcount[alias] += 1

    def unref_alias(self, alias, amount=1):
        """Decreases the reference count for this alias."""
        self.alias_refcount[alias] -= amount

    def promote_joins(self, aliases):
        """
        Promote recursively the join type of given aliases and its children to
        an outer join. If 'unconditional' is False, only promote the join if
        it is nullable or the parent join is an outer join.

        The children promotion is done to avoid join chains that contain a LOUTER
        b INNER c. So, if we have currently a INNER b INNER c and a->b is promoted,
        then we must also promote b->c automatically, or otherwise the promotion
        of a->b doesn't actually change anything in the query results.
        """
        aliases = list(aliases)
        while aliases:
            alias = aliases.pop(0)
            if self.alias_map[alias].join_type is None:
                # This is the base table (first FROM entry) - this table
                # isn't really joined at all in the query, so we should not
                # alter its join type.
                continue
            # Only the first alias (skipped above) should have None join_type
            assert self.alias_map[alias].join_type is not None
            parent_alias = self.alias_map[alias].parent_alias
            parent_louter = parent_alias and self.alias_map[parent_alias].join_type == LOUTER
            already_louter = self.alias_map[alias].join_type == LOUTER
            if ((self.alias_map[alias].nullable or parent_louter) and
                    not already_louter):
                self.alias_map[alias] = self.alias_map[alias].promote()
                # Join type of 'alias' changed, so re-examine all aliases that
                # refer to this one.
                aliases.extend(
                    join for join in self.alias_map
                    if self.alias_map[join].parent_alias == alias and join not in aliases
                )

    def demote_joins(self, aliases):
        """
        Change join type from LOUTER to INNER for all joins in aliases.

        Similarly to promote_joins(), this method must ensure no join chains
        containing first an outer, then an inner join are generated. If we
        are demoting b->c join in chain a LOUTER b LOUTER c then we must
        demote a->b automatically, or otherwise the demotion of b->c doesn't
        actually change anything in the query results. .
        """
        aliases = list(aliases)
        while aliases:
            alias = aliases.pop(0)
            if self.alias_map[alias].join_type == LOUTER:
                self.alias_map[alias] = self.alias_map[alias].demote()
                parent_alias = self.alias_map[alias].parent_alias
                if self.alias_map[parent_alias].join_type == INNER:
                    aliases.append(parent_alias)

    def reset_refcounts(self, to_counts):
        """
        Reset reference counts for aliases so that they match the value passed
        in `to_counts`.
        """
        for alias, cur_refcount in self.alias_refcount.copy().items():
            unref_amount = cur_refcount - to_counts.get(alias, 0)
            self.unref_alias(alias, unref_amount)

    def change_aliases(self, change_map):
        """
        Change the aliases in change_map (which maps old-alias -> new-alias),
        relabelling any references to them in select columns and the where
        clause.
        """
        assert set(change_map).isdisjoint(change_map.values())

        # 1. Update references in "select" (normal columns plus aliases),
        # "group by" and "where".
        self.where.relabel_aliases(change_map)
        if isinstance(self.group_by, tuple):
            self.group_by = tuple([col.relabeled_clone(change_map) for col in self.group_by])
        self.select = tuple([col.relabeled_clone(change_map) for col in self.select])
        if self._annotations:
            self._annotations = OrderedDict(
                (key, col.relabeled_clone(change_map)) for key, col in self._annotations.items())

        # 2. Rename the alias in the internal table/alias datastructures.
        for old_alias, new_alias in change_map.items():
            if old_alias not in self.alias_map:
                continue
            alias_data = self.alias_map[old_alias].relabeled_clone(change_map)
            self.alias_map[new_alias] = alias_data
            self.alias_refcount[new_alias] = self.alias_refcount[old_alias]
            del self.alias_refcount[old_alias]
            del self.alias_map[old_alias]

            table_aliases = self.table_map[alias_data.table_name]
            for pos, alias in enumerate(table_aliases):
                if alias == old_alias:
                    table_aliases[pos] = new_alias
                    break
        self.external_aliases = {change_map.get(alias, alias)
                                 for alias in self.external_aliases}

    def bump_prefix(self, outer_query):
        """
        Change the alias prefix to the next letter in the alphabet in a way
        that the outer query's aliases and this query's aliases will not
        conflict. Even tables that previously had no alias will get an alias
        after this call.
        """
        def prefix_gen():
            """
            Generate a sequence of characters in alphabetical order:
                -> 'A', 'B', 'C', ...

            When the alphabet is finished, the sequence will continue with the
            Cartesian product:
                -> 'AA', 'AB', 'AC', ...
            """
            alphabet = ascii_uppercase
            prefix = chr(ord(self.alias_prefix) + 1)
            yield prefix
            for n in count(1):
                seq = alphabet[alphabet.index(prefix):] if prefix else alphabet
                for s in product(seq, repeat=n):
                    yield ''.join(s)
                prefix = None

        if self.alias_prefix != outer_query.alias_prefix:
            # No clashes between self and outer query should be possible.
            return

        local_recursion_limit = 127  # explicitly avoid infinite loop
        for pos, prefix in enumerate(prefix_gen()):
            if prefix not in self.subq_aliases:
                self.alias_prefix = prefix
                break
            if pos > local_recursion_limit:
                raise RuntimeError(
                    'Maximum recursion depth exceeded: too many subqueries.'
                )
        self.subq_aliases = self.subq_aliases.union([self.alias_prefix])
        outer_query.subq_aliases = outer_query.subq_aliases.union(self.subq_aliases)
        change_map = OrderedDict()
        for pos, alias in enumerate(self.alias_map):
            new_alias = '%s%d' % (self.alias_prefix, pos)
            change_map[alias] = new_alias
        self.change_aliases(change_map)

    def get_initial_alias(self):
        """
        Return the first alias for this query, after increasing its reference
        count.
        """
        if self.alias_map:
            alias = self.base_table
            self.ref_alias(alias)
        else:
            alias = self.join(BaseTable(self.get_meta().db_table, None))
        return alias

    def count_active_tables(self):
        """
        Return the number of tables in this query with a non-zero reference
        count. After execution, the reference counts are zeroed, so tables
        added in compiler will not be seen by this method.
        """
        return len([1 for count in self.alias_refcount.values() if count])

    def join(self, join, reuse=None, reuse_with_filtered_relation=False):
        """
        Return an alias for the 'join', either reusing an existing alias for
        that join or creating a new one. 'join' is either a
        sql.datastructures.BaseTable or Join.

        The 'reuse' parameter can be either None which means all joins are
        reusable, or it can be a set containing the aliases that can be reused.

        The 'reuse_with_filtered_relation' parameter is used when computing
        FilteredRelation instances.

        A join is always created as LOUTER if the lhs alias is LOUTER to make
        sure chains like t1 LOUTER t2 INNER t3 aren't generated. All new
        joins are created as LOUTER if the join is nullable.
        """
        if reuse_with_filtered_relation and reuse:
            reuse_aliases = [
                a for a, j in self.alias_map.items()
                if a in reuse and j.equals(join, with_filtered_relation=False)
            ]
        else:
            reuse_aliases = [
                a for a, j in self.alias_map.items()
                if (reuse is None or a in reuse) and j == join
            ]
        if reuse_aliases:
            self.ref_alias(reuse_aliases[0])
            return reuse_aliases[0]

        # No reuse is possible, so we need a new alias.
        alias, _ = self.table_alias(join.table_name, create=True, filtered_relation=join.filtered_relation)
        if join.join_type:
            if self.alias_map[join.parent_alias].join_type == LOUTER or join.nullable:
                join_type = LOUTER
            else:
                join_type = INNER
            join.join_type = join_type
        join.table_alias = alias
        self.alias_map[alias] = join
        return alias

    def join_parent_model(self, opts, model, alias, seen):
        """
        Make sure the given 'model' is joined in the query. If 'model' isn't
        a parent of 'opts' or if it is None this method is a no-op.

        The 'alias' is the root alias for starting the join, 'seen' is a dict
        of model -> alias of existing joins. It must also contain a mapping
        of None -> some alias. This will be returned in the no-op case.
        """
        if model in seen:
            return seen[model]
        chain = opts.get_base_chain(model)
        if not chain:
            return alias
        curr_opts = opts
        for int_model in chain:
            if int_model in seen:
                curr_opts = int_model._meta
                alias = seen[int_model]
                continue
            # Proxy model have elements in base chain
            # with no parents, assign the new options
            # object and skip to the next base in that
            # case
            if not curr_opts.parents[int_model]:
                curr_opts = int_model._meta
                continue
            link_field = curr_opts.get_ancestor_link(int_model)
            join_info = self.setup_joins([link_field.name], curr_opts, alias)
            curr_opts = int_model._meta
            alias = seen[int_model] = join_info.joins[-1]
        return alias or seen[None]

    def add_annotation(self, annotation, alias, is_summary=False):
        """Add a single annotation expression to the Query."""
        annotation = annotation.resolve_expression(self, allow_joins=True, reuse=None,
                                                   summarize=is_summary)
        self.append_annotation_mask([alias])
        self.annotations[alias] = annotation

    def resolve_expression(self, query, *args, **kwargs):
        clone = self.clone()
        # Subqueries need to use a different set of aliases than the outer query.
        clone.bump_prefix(query)
        clone.subquery = True
        # It's safe to drop ordering if the queryset isn't using slicing,
        # distinct(*fields) or select_for_update().
        if (self.low_mark == 0 and self.high_mark is None and
                not self.distinct_fields and
                not self.select_for_update):
            clone.clear_ordering(True)
        return clone

    def as_sql(self, compiler, connection):
        return self.get_compiler(connection=connection).as_sql()

    def resolve_lookup_value(self, value, can_reuse, allow_joins):
        if hasattr(value, 'resolve_expression'):
            value = value.resolve_expression(self, reuse=can_reuse, allow_joins=allow_joins)
        elif isinstance(value, (list, tuple)):
            # The items of the iterable may be expressions and therefore need
            # to be resolved independently.
            for sub_value in value:
                if hasattr(sub_value, 'resolve_expression'):
                    sub_value.resolve_expression(self, reuse=can_reuse, allow_joins=allow_joins)
        return value

    def solve_lookup_type(self, lookup):
        """
        Solve the lookup type from the lookup (e.g.: 'foobar__id__icontains').
        """
        lookup_splitted = lookup.split(LOOKUP_SEP)
        if self._annotations:
            expression, expression_lookups = refs_expression(lookup_splitted, self.annotations)
            if expression:
                return expression_lookups, (), expression
        _, field, _, lookup_parts = self.names_to_path(lookup_splitted, self.get_meta())
        field_parts = lookup_splitted[0:len(lookup_splitted) - len(lookup_parts)]
        if len(lookup_parts) > 1 and not field_parts:
            raise FieldError(
                'Invalid lookup "%s" for model %s".' %
                (lookup, self.get_meta().model.__name__)
            )
        return lookup_parts, field_parts, False

    def check_query_object_type(self, value, opts, field):
        """
        Check whether the object passed while querying is of the correct type.
        If not, raise a ValueError specifying the wrong object.
        """
        if hasattr(value, '_meta'):
            if not check_rel_lookup_compatibility(value._meta.model, opts, field):
                raise ValueError(
                    'Cannot query "%s": Must be "%s" instance.' %
                    (value, opts.object_name))

    def check_related_objects(self, field, value, opts):
        """Check the type of object passed to query relations."""
        if field.is_relation:
            # Check that the field and the queryset use the same model in a
            # query like .filter(author=Author.objects.all()). For example, the
            # opts would be Author's (from the author field) and value.model
            # would be Author.objects.all() queryset's .model (Author also).
            # The field is the related field on the lhs side.
            if (isinstance(value, Query) and not value.has_select_fields and
                    not check_rel_lookup_compatibility(value.model, opts, field)):
                raise ValueError(
                    'Cannot use QuerySet for "%s": Use a QuerySet for "%s".' %
                    (value.model._meta.object_name, opts.object_name)
                )
            elif hasattr(value, '_meta'):
                self.check_query_object_type(value, opts, field)
            elif hasattr(value, '__iter__'):
                for v in value:
                    self.check_query_object_type(v, opts, field)

    def build_lookup(self, lookups, lhs, rhs):
        """
        Try to extract transforms and lookup from given lhs.

        The lhs value is something that works like SQLExpression.
        The rhs value is what the lookup is going to compare against.
        The lookups is a list of names to extract using get_lookup()
        and get_transform().
        """
        # __exact is the default lookup if one isn't given.
        if len(lookups) == 0:
            lookups = ['exact']

        for name in lookups[:-1]:
            lhs = self.try_transform(lhs, name)
        # First try get_lookup() so that the lookup takes precedence if the lhs
        # supports both transform and lookup for the name.
        lookup_class = lhs.get_lookup(lookups[-1])
        if not lookup_class:
            if lhs.field.is_relation:
                raise FieldError('Related Field got invalid lookup: {}'.format(lookups[-1]))
            # A lookup wasn't found. Try to interpret the name as a transform
            # and do an Exact lookup against it.
            lhs = self.try_transform(lhs, lookups[-1])
            lookup_class = lhs.get_lookup('exact')

        if not lookup_class:
            return

        lookup = lookup_class(lhs, rhs)
        # Interpret '__exact=None' as the sql 'is NULL'; otherwise, reject all
        # uses of None as a query value.
        if lookup.rhs is None:
            if lookup.lookup_name not in ('exact', 'iexact'):
                raise ValueError("Cannot use None as a query value")
            return lhs.get_lookup('isnull')(lhs, True)

        # For Oracle '' is equivalent to null. The check must be done at this
        # stage because join promotion can't be done in the compiler. Using
        # DEFAULT_DB_ALIAS isn't nice but it's the best that can be done here.
        # A similar thing is done in is_nullable(), too.
        if (connections[DEFAULT_DB_ALIAS].features.interprets_empty_strings_as_nulls and
                lookup.lookup_name == 'exact' and lookup.rhs == ''):
            return lhs.get_lookup('isnull')(lhs, True)

        return lookup

    def try_transform(self, lhs, name):
        """
        Helper method for build_lookup(). Try to fetch and initialize
        a transform for name parameter from lhs.
        """
        transform_class = lhs.get_transform(name)
        if transform_class:
            return transform_class(lhs)
        else:
            raise FieldError(
                "Unsupported lookup '%s' for %s or join on the field not "
                "permitted." %
                (name, lhs.output_field.__class__.__name__))

    def build_filter(self, filter_expr, branch_negated=False, current_negated=False,
                     can_reuse=None, allow_joins=True, split_subq=True,
                     reuse_with_filtered_relation=False):
        """
        Build a WhereNode for a single filter clause but don't add it
        to this Query. Query.add_q() will then add this filter to the where
        Node.

        The 'branch_negated' tells us if the current branch contains any
        negations. This will be used to determine if subqueries are needed.

        The 'current_negated' is used to determine if the current filter is
        negated or not and this will be used to determine if IS NULL filtering
        is needed.

        The difference between current_negated and branch_negated is that
        branch_negated is set on first negation, but current_negated is
        flipped for each negation.

        Note that add_filter will not do any negating itself, that is done
        upper in the code by add_q().

        The 'can_reuse' is a set of reusable joins for multijoins.

        If 'reuse_with_filtered_relation' is True, then only joins in can_reuse
        will be reused.

        The method will create a filter clause that can be added to the current
        query. However, if the filter isn't added to the query then the caller
        is responsible for unreffing the joins used.
        """
        if isinstance(filter_expr, dict):
            raise FieldError("Cannot parse keyword query as dict")
        arg, value = filter_expr
        if not arg:
            raise FieldError("Cannot parse keyword query %r" % arg)
        lookups, parts, reffed_expression = self.solve_lookup_type(arg)

        if not getattr(reffed_expression, 'filterable', True):
            raise NotSupportedError(
                reffed_expression.__class__.__name__ + ' is disallowed in '
                'the filter clause.'
            )

        if not allow_joins and len(parts) > 1:
            raise FieldError("Joined field references are not permitted in this query")

        pre_joins = self.alias_refcount.copy()
        value = self.resolve_lookup_value(value, can_reuse, allow_joins)
        used_joins = {k for k, v in self.alias_refcount.items() if v > pre_joins.get(k, 0)}

        clause = self.where_class()
        if reffed_expression:
            condition = self.build_lookup(lookups, reffed_expression, value)
            clause.add(condition, AND)
            return clause, []

        opts = self.get_meta()
        alias = self.get_initial_alias()
        allow_many = not branch_negated or not split_subq

        try:
            join_info = self.setup_joins(
                parts, opts, alias, can_reuse=can_reuse, allow_many=allow_many,
                reuse_with_filtered_relation=reuse_with_filtered_relation,
            )

            # Prevent iterator from being consumed by check_related_objects()
            if isinstance(value, Iterator):
                value = list(value)
            self.check_related_objects(join_info.final_field, value, join_info.opts)

            # split_exclude() needs to know which joins were generated for the
            # lookup parts
            self._lookup_joins = join_info.joins
        except MultiJoin as e:
            return self.split_exclude(filter_expr, LOOKUP_SEP.join(parts[:e.level]),
                                      can_reuse, e.names_with_path)

        # Update used_joins before trimming since they are reused to determine
        # which joins could be later promoted to INNER.
        used_joins.update(join_info.joins)
        targets, alias, join_list = self.trim_joins(join_info.targets, join_info.joins, join_info.path)
        if can_reuse is not None:
            can_reuse.update(join_list)

        if join_info.final_field.is_relation:
            # No support for transforms for relational fields
            num_lookups = len(lookups)
            if num_lookups > 1:
                raise FieldError('Related Field got invalid lookup: {}'.format(lookups[0]))
            if len(targets) == 1:
                col = targets[0].get_col(alias, join_info.final_field)
            else:
                col = MultiColSource(alias, targets, join_info.targets, join_info.final_field)
        else:
            col = targets[0].get_col(alias, join_info.final_field)

        condition = self.build_lookup(lookups, col, value)
        lookup_type = condition.lookup_name
        clause.add(condition, AND)

        require_outer = lookup_type == 'isnull' and condition.rhs is True and not current_negated
        if current_negated and (lookup_type != 'isnull' or condition.rhs is False):
            require_outer = True
            if (lookup_type != 'isnull' and (
                    self.is_nullable(targets[0]) or
                    self.alias_map[join_list[-1]].join_type == LOUTER)):
                # The condition added here will be SQL like this:
                # NOT (col IS NOT NULL), where the first NOT is added in
                # upper layers of code. The reason for addition is that if col
                # is null, then col != someval will result in SQL "unknown"
                # which isn't the same as in Python. The Python None handling
                # is wanted, and it can be gotten by
                # (col IS NULL OR col != someval)
                #   <=>
                # NOT (col IS NOT NULL AND col = someval).
                lookup_class = targets[0].get_lookup('isnull')
                clause.add(lookup_class(targets[0].get_col(alias, join_info.targets[0]), False), AND)
        return clause, used_joins if not require_outer else ()

    def add_filter(self, filter_clause):
        self.add_q(Q(**{filter_clause[0]: filter_clause[1]}))

    def add_q(self, q_object):
        """
        A preprocessor for the internal _add_q(). Responsible for doing final
        join promotion.
        """
        # For join promotion this case is doing an AND for the added q_object
        # and existing conditions. So, any existing inner join forces the join
        # type to remain inner. Existing outer joins can however be demoted.
        # (Consider case where rel_a is LOUTER and rel_a__col=1 is added - if
        # rel_a doesn't produce any rows, then the whole condition must fail.
        # So, demotion is OK.
        existing_inner = {a for a in self.alias_map if self.alias_map[a].join_type == INNER}
        clause, _ = self._add_q(q_object, self.used_aliases)
        if clause:
            self.where.add(clause, AND)
        self.demote_joins(existing_inner)

    def _add_q(self, q_object, used_aliases, branch_negated=False,
               current_negated=False, allow_joins=True, split_subq=True):
        """Add a Q-object to the current filter."""
        connector = q_object.connector
        current_negated = current_negated ^ q_object.negated
        branch_negated = branch_negated or q_object.negated
        target_clause = self.where_class(connector=connector,
                                         negated=q_object.negated)
        joinpromoter = JoinPromoter(q_object.connector, len(q_object.children), current_negated)
        for child in q_object.children:
            if isinstance(child, Node):
                child_clause, needed_inner = self._add_q(
                    child, used_aliases, branch_negated,
                    current_negated, allow_joins, split_subq)
                joinpromoter.add_votes(needed_inner)
            else:
                child_clause, needed_inner = self.build_filter(
                    child, can_reuse=used_aliases, branch_negated=branch_negated,
                    current_negated=current_negated, allow_joins=allow_joins,
                    split_subq=split_subq,
                )
                joinpromoter.add_votes(needed_inner)
            if child_clause:
                target_clause.add(child_clause, connector)
        needed_inner = joinpromoter.update_join_types(self)
        return target_clause, needed_inner

    def build_filtered_relation_q(self, q_object, reuse, branch_negated=False, current_negated=False):
        """Add a FilteredRelation object to the current filter."""
        connector = q_object.connector
        current_negated ^= q_object.negated
        branch_negated = branch_negated or q_object.negated
        target_clause = self.where_class(connector=connector, negated=q_object.negated)
        for child in q_object.children:
            if isinstance(child, Node):
                child_clause = self.build_filtered_relation_q(
                    child, reuse=reuse, branch_negated=branch_negated,
                    current_negated=current_negated,
                )
            else:
                child_clause, _ = self.build_filter(
                    child, can_reuse=reuse, branch_negated=branch_negated,
                    current_negated=current_negated,
                    allow_joins=True, split_subq=False,
                    reuse_with_filtered_relation=True,
                )
            target_clause.add(child_clause, connector)
        return target_clause

    def add_filtered_relation(self, filtered_relation, alias):
        filtered_relation.alias = alias
        lookups = dict(get_children_from_q(filtered_relation.condition))
        for lookup in chain((filtered_relation.relation_name,), lookups):
            lookup_parts, field_parts, _ = self.solve_lookup_type(lookup)
            shift = 2 if not lookup_parts else 1
            if len(field_parts) > (shift + len(lookup_parts)):
                raise ValueError(
                    "FilteredRelation's condition doesn't support nested "
                    "relations (got %r)." % lookup
                )
        self._filtered_relations[filtered_relation.alias] = filtered_relation

    def names_to_path(self, names, opts, allow_many=True, fail_on_missing=False):
        """
        Walk the list of names and turns them into PathInfo tuples. A single
        name in 'names' can generate multiple PathInfos (m2m, for example).

        'names' is the path of names to travel, 'opts' is the model Options we
        start the name resolving from, 'allow_many' is as for setup_joins().
        If fail_on_missing is set to True, then a name that can't be resolved
        will generate a FieldError.

        Return a list of PathInfo tuples. In addition return the final field
        (the last used join field) and target (which is a field guaranteed to
        contain the same value as the final field). Finally, return those names
        that weren't found (which are likely transforms and the final lookup).
        """
        path, names_with_path = [], []
        for pos, name in enumerate(names):
            cur_names_with_path = (name, [])
            if name == 'pk':
                name = opts.pk.name

            field = None
            filtered_relation = None
            try:
                field = opts.get_field(name)
            except FieldDoesNotExist:
                if name in self.annotation_select:
                    field = self.annotation_select[name].output_field
                elif name in self._filtered_relations and pos == 0:
                    filtered_relation = self._filtered_relations[name]
                    field = opts.get_field(filtered_relation.relation_name)
            if field is not None:
                # Fields that contain one-to-many relations with a generic
                # model (like a GenericForeignKey) cannot generate reverse
                # relations and therefore cannot be used for reverse querying.
                if field.is_relation and not field.related_model:
                    raise FieldError(
                        "Field %r does not generate an automatic reverse "
                        "relation and therefore cannot be used for reverse "
                        "querying. If it is a GenericForeignKey, consider "
                        "adding a GenericRelation." % name
                    )
                try:
                    model = field.model._meta.concrete_model
                except AttributeError:
                    # QuerySet.annotate() may introduce fields that aren't
                    # attached to a model.
                    model = None
            else:
                # We didn't find the current field, so move position back
                # one step.
                pos -= 1
                if pos == -1 or fail_on_missing:
                    field_names = list(get_field_names_from_opts(opts))
                    available = sorted(
                        field_names + list(self.annotation_select) +
                        list(self._filtered_relations)
                    )
                    raise FieldError("Cannot resolve keyword '%s' into field. "
                                     "Choices are: %s" % (name, ", ".join(available)))
                break
            # Check if we need any joins for concrete inheritance cases (the
            # field lives in parent, but we are currently in one of its
            # children)
            if model is not opts.model:
                path_to_parent = opts.get_path_to_parent(model)
                if path_to_parent:
                    path.extend(path_to_parent)
                    cur_names_with_path[1].extend(path_to_parent)
                    opts = path_to_parent[-1].to_opts
            if hasattr(field, 'get_path_info'):
                pathinfos = field.get_path_info(filtered_relation)
                if not allow_many:
                    for inner_pos, p in enumerate(pathinfos):
                        if p.m2m:
                            cur_names_with_path[1].extend(pathinfos[0:inner_pos + 1])
                            names_with_path.append(cur_names_with_path)
                            raise MultiJoin(pos + 1, names_with_path)
                last = pathinfos[-1]
                path.extend(pathinfos)
                final_field = last.join_field
                opts = last.to_opts
                targets = last.target_fields
                cur_names_with_path[1].extend(pathinfos)
                names_with_path.append(cur_names_with_path)
            else:
                # Local non-relational field.
                final_field = field
                targets = (field,)
                if fail_on_missing and pos + 1 != len(names):
                    raise FieldError(
                        "Cannot resolve keyword %r into field. Join on '%s'"
                        " not permitted." % (names[pos + 1], name))
                break
        return path, final_field, targets, names[pos + 1:]

    def setup_joins(self, names, opts, alias, can_reuse=None, allow_many=True,
                    reuse_with_filtered_relation=False):
        """
        Compute the necessary table joins for the passage through the fields
        given in 'names'. 'opts' is the Options class for the current model
        (which gives the table we are starting from), 'alias' is the alias for
        the table to start the joining from.

        The 'can_reuse' defines the reverse foreign key joins we can reuse. It
        can be None in which case all joins are reusable or a set of aliases
        that can be reused. Note that non-reverse foreign keys are always
        reusable when using setup_joins().

        The 'reuse_with_filtered_relation' can be used to force 'can_reuse'
        parameter and force the relation on the given connections.

        If 'allow_many' is False, then any reverse foreign key seen will
        generate a MultiJoin exception.

        Return the final field involved in the joins, the target field (used
        for any 'where' constraint), the final 'opts' value, the joins and the
        field path travelled to generate the joins.

        The target field is the field containing the concrete value. Final
        field can be something different, for example foreign key pointing to
        that value. Final field is needed for example in some value
        conversions (convert 'obj' in fk__id=obj to pk val using the foreign
        key field for example).
        """
        joins = [alias]
        # First, generate the path for the names
        path, final_field, targets, rest = self.names_to_path(
            names, opts, allow_many, fail_on_missing=True)

        # Then, add the path to the query's joins. Note that we can't trim
        # joins at this stage - we will need the information about join type
        # of the trimmed joins.
        for join in path:
            if join.filtered_relation:
                filtered_relation = join.filtered_relation.clone()
                table_alias = filtered_relation.alias
            else:
                filtered_relation = None
                table_alias = None
            opts = join.to_opts
            if join.direct:
                nullable = self.is_nullable(join.join_field)
            else:
                nullable = True
            connection = Join(
                opts.db_table, alias, table_alias, INNER, join.join_field,
                nullable, filtered_relation=filtered_relation,
            )
            reuse = can_reuse if join.m2m or reuse_with_filtered_relation else None
            alias = self.join(
                connection, reuse=reuse,
                reuse_with_filtered_relation=reuse_with_filtered_relation,
            )
            joins.append(alias)
            if filtered_relation:
                filtered_relation.path = joins[:]
        return JoinInfo(final_field, targets, opts, joins, path)

    def trim_joins(self, targets, joins, path):
        """
        The 'target' parameter is the final field being joined to, 'joins'
        is the full list of join aliases. The 'path' contain the PathInfos
        used to create the joins.

        Return the final target field and table alias and the new active
        joins.

        Always trim any direct join if the target column is already in the
        previous table. Can't trim reverse joins as it's unknown if there's
        anything on the other side of the join.
        """
        joins = joins[:]
        for pos, info in enumerate(reversed(path)):
            if len(joins) == 1 or not info.direct:
                break
            if info.filtered_relation:
                break
            join_targets = {t.column for t in info.join_field.foreign_related_fields}
            cur_targets = {t.column for t in targets}
            if not cur_targets.issubset(join_targets):
                break
            targets_dict = {r[1].column: r[0] for r in info.join_field.related_fields if r[1].column in cur_targets}
            targets = tuple(targets_dict[t.column] for t in targets)
            self.unref_alias(joins.pop())
        return targets, joins[-1], joins

    def resolve_ref(self, name, allow_joins=True, reuse=None, summarize=False):
        if not allow_joins and LOOKUP_SEP in name:
            raise FieldError("Joined field references are not permitted in this query")
        if name in self.annotations:
            if summarize:
                # Summarize currently means we are doing an aggregate() query
                # which is executed as a wrapped subquery if any of the
                # aggregate() elements reference an existing annotation. In
                # that case we need to return a Ref to the subquery's annotation.
                return Ref(name, self.annotation_select[name])
            else:
                return self.annotation_select[name]
        else:
            field_list = name.split(LOOKUP_SEP)
            join_info = self.setup_joins(field_list, self.get_meta(), self.get_initial_alias(), can_reuse=reuse)
            targets, _, join_list = self.trim_joins(join_info.targets, join_info.joins, join_info.path)
            if len(targets) > 1:
                raise FieldError("Referencing multicolumn fields with F() objects "
                                 "isn't supported")
            if reuse is not None:
                reuse.update(join_list)
            col = targets[0].get_col(join_list[-1], join_info.targets[0])
            return col

    def split_exclude(self, filter_expr, prefix, can_reuse, names_with_path):
        """
        When doing an exclude against any kind of N-to-many relation, we need
        to use a subquery. This method constructs the nested query, given the
        original exclude filter (filter_expr) and the portion up to the first
        N-to-many relation field.

        As an example we could have original filter ~Q(child__name='foo').
        We would get here with filter_expr = child__name, prefix = child and
        can_reuse is a set of joins usable for filters in the original query.

        We will turn this into equivalent of:
            WHERE NOT (pk IN (SELECT parent_id FROM thetable
                              WHERE name = 'foo' AND parent_id IS NOT NULL))

        It might be worth it to consider using WHERE NOT EXISTS as that has
        saner null handling, and is easier for the backend's optimizer to
        handle.
        """
        # Generate the inner query.
        query = Query(self.model)
        query.add_filter(filter_expr)
        query.clear_ordering(True)
        # Try to have as simple as possible subquery -> trim leading joins from
        # the subquery.
        trimmed_prefix, contains_louter = query.trim_start(names_with_path)

        # Add extra check to make sure the selected field will not be null
        # since we are adding an IN <subquery> clause. This prevents the
        # database from tripping over IN (...,NULL,...) selects and returning
        # nothing
        col = query.select[0]
        select_field = col.target
        alias = col.alias
        if self.is_nullable(select_field):
            lookup_class = select_field.get_lookup('isnull')
            lookup = lookup_class(select_field.get_col(alias), False)
            query.where.add(lookup, AND)
        if alias in can_reuse:
            pk = select_field.model._meta.pk
            # Need to add a restriction so that outer query's filters are in effect for
            # the subquery, too.
            query.bump_prefix(self)
            lookup_class = select_field.get_lookup('exact')
            # Note that the query.select[0].alias is different from alias
            # due to bump_prefix above.
            lookup = lookup_class(pk.get_col(query.select[0].alias),
                                  pk.get_col(alias))
            query.where.add(lookup, AND)
            query.external_aliases.add(alias)

        condition, needed_inner = self.build_filter(
            ('%s__in' % trimmed_prefix, query),
            current_negated=True, branch_negated=True, can_reuse=can_reuse)
        if contains_louter:
            or_null_condition, _ = self.build_filter(
                ('%s__isnull' % trimmed_prefix, True),
                current_negated=True, branch_negated=True, can_reuse=can_reuse)
            condition.add(or_null_condition, OR)
            # Note that the end result will be:
            # (outercol NOT IN innerq AND outercol IS NOT NULL) OR outercol IS NULL.
            # This might look crazy but due to how IN works, this seems to be
            # correct. If the IS NOT NULL check is removed then outercol NOT
            # IN will return UNKNOWN. If the IS NULL check is removed, then if
            # outercol IS NULL we will not match the row.
        return condition, needed_inner

    def set_empty(self):
        self.where.add(NothingNode(), AND)

    def is_empty(self):
        return any(isinstance(c, NothingNode) for c in self.where.children)

    def set_limits(self, low=None, high=None):
        """
        Adjust the limits on the rows retrieved. Use low/high to set these,
        as it makes it more Pythonic to read and write. When the SQL query is
        created, convert them to the appropriate offset and limit values.

        Apply any limits passed in here to the existing constraints. Add low
        to the current low value and clamp both to any existing high value.
        """
        if high is not None:
            if self.high_mark is not None:
                self.high_mark = min(self.high_mark, self.low_mark + high)
            else:
                self.high_mark = self.low_mark + high
        if low is not None:
            if self.high_mark is not None:
                self.low_mark = min(self.high_mark, self.low_mark + low)
            else:
                self.low_mark = self.low_mark + low

        if self.low_mark == self.high_mark:
            self.set_empty()

    def clear_limits(self):
        """Clear any existing limits."""
        self.low_mark, self.high_mark = 0, None

    def has_limit_one(self):
        return self.high_mark is not None and (self.high_mark - self.low_mark) == 1

    def can_filter(self):
        """
        Return True if adding filters to this instance is still possible.

        Typically, this means no limits or offsets have been put on the results.
        """
        return not self.low_mark and self.high_mark is None

    def clear_select_clause(self):
        """Remove all fields from SELECT clause."""
        self.select = ()
        self.default_cols = False
        self.select_related = False
        self.set_extra_mask(())
        self.set_annotation_mask(())

    def clear_select_fields(self):
        """
        Clear the list of fields to select (but not extra_select columns).
        Some queryset types completely replace any existing list of select
        columns.
        """
        self.select = ()
        self.values_select = ()

    def set_select(self, cols):
        self.default_cols = False
        self.select = tuple(cols)

    def add_distinct_fields(self, *field_names):
        """
        Add and resolve the given fields to the query's "distinct on" clause.
        """
        self.distinct_fields = field_names
        self.distinct = True

    def add_fields(self, field_names, allow_m2m=True):
        """
        Add the given (model) fields to the select set. Add the field names in
        the order specified.
        """
        alias = self.get_initial_alias()
        opts = self.get_meta()

        try:
            cols = []
            for name in field_names:
                # Join promotion note - we must not remove any rows here, so
                # if there is no existing joins, use outer join.
                join_info = self.setup_joins(name.split(LOOKUP_SEP), opts, alias, allow_many=allow_m2m)
                targets, final_alias, joins = self.trim_joins(
                    join_info.targets,
                    join_info.joins,
                    join_info.path,
                )
                for target in targets:
                    cols.append(target.get_col(final_alias))
            if cols:
                self.set_select(cols)
        except MultiJoin:
            raise FieldError("Invalid field name: '%s'" % name)
        except FieldError:
            if LOOKUP_SEP in name:
                # For lookups spanning over relationships, show the error
                # from the model on which the lookup failed.
                raise
            else:
                names = sorted(
                    list(get_field_names_from_opts(opts)) + list(self.extra) +
                    list(self.annotation_select) + list(self._filtered_relations)
                )
                raise FieldError("Cannot resolve keyword %r into field. "
                                 "Choices are: %s" % (name, ", ".join(names)))

    def add_ordering(self, *ordering):
        """
        Add items from the 'ordering' sequence to the query's "order by"
        clause. These items are either field names (not column names) --
        possibly with a direction prefix ('-' or '?') -- or OrderBy
        expressions.

        If 'ordering' is empty, clear all ordering from the query.
        """
        errors = []
        for item in ordering:
            if not hasattr(item, 'resolve_expression') and not ORDER_PATTERN.match(item):
                errors.append(item)
            if getattr(item, 'contains_aggregate', False):
                raise FieldError(
                    'Using an aggregate in order_by() without also including '
                    'it in annotate() is not allowed: %s' % item
                )
        if errors:
            raise FieldError('Invalid order_by arguments: %s' % errors)
        if ordering:
            self.order_by += ordering
        else:
            self.default_ordering = False

    def clear_ordering(self, force_empty):
        """
        Remove any ordering settings. If 'force_empty' is True, there will be
        no ordering in the resulting query (not even the model's default).
        """
        self.order_by = ()
        self.extra_order_by = ()
        if force_empty:
            self.default_ordering = False

    def set_group_by(self):
        """
        Expand the GROUP BY clause required by the query.

        This will usually be the set of all non-aggregate fields in the
        return data. If the database backend supports grouping by the
        primary key, and the query would be equivalent, the optimization
        will be made automatically.
        """
        group_by = list(self.select)
        if self.annotation_select:
            for alias, annotation in self.annotation_select.items():
                for col in annotation.get_group_by_cols():
                    group_by.append(col)
        self.group_by = tuple(group_by)

    def add_select_related(self, fields):
        """
        Set up the select_related data structure so that we only select
        certain related models (as opposed to all models, when
        self.select_related=True).
        """
        if isinstance(self.select_related, bool):
            field_dict = {}
        else:
            field_dict = self.select_related
        for field in fields:
            d = field_dict
            for part in field.split(LOOKUP_SEP):
                d = d.setdefault(part, {})
        self.select_related = field_dict

    def add_extra(self, select, select_params, where, params, tables, order_by):
        """
        Add data to the various extra_* attributes for user-created additions
        to the query.
        """
        if select:
            # We need to pair any placeholder markers in the 'select'
            # dictionary with their parameters in 'select_params' so that
            # subsequent updates to the select dictionary also adjust the
            # parameters appropriately.
            select_pairs = OrderedDict()
            if select_params:
                param_iter = iter(select_params)
            else:
                param_iter = iter([])
            for name, entry in select.items():
                entry = force_text(entry)
                entry_params = []
                pos = entry.find("%s")
                while pos != -1:
                    if pos == 0 or entry[pos - 1] != '%':
                        entry_params.append(next(param_iter))
                    pos = entry.find("%s", pos + 2)
                select_pairs[name] = (entry, entry_params)
            # This is order preserving, since self.extra_select is an OrderedDict.
            self.extra.update(select_pairs)
        if where or params:
            self.where.add(ExtraWhere(where, params), AND)
        if tables:
            self.extra_tables += tuple(tables)
        if order_by:
            self.extra_order_by = order_by

    def clear_deferred_loading(self):
        """Remove any fields from the deferred loading set."""
        self.deferred_loading = (frozenset(), True)

    def add_deferred_loading(self, field_names):
        """
        Add the given list of model field names to the set of fields to
        exclude from loading from the database when automatic column selection
        is done. Add the new field names to any existing field names that
        are deferred (or removed from any existing field names that are marked
        as the only ones for immediate loading).
        """
        # Fields on related models are stored in the literal double-underscore
        # format, so that we can use a set datastructure. We do the foo__bar
        # splitting and handling when computing the SQL column names (as part of
        # get_columns()).
        existing, defer = self.deferred_loading
        if defer:
            # Add to existing deferred names.
            self.deferred_loading = existing.union(field_names), True
        else:
            # Remove names from the set of any existing "immediate load" names.
            self.deferred_loading = existing.difference(field_names), False

    def add_immediate_loading(self, field_names):
        """
        Add the given list of model field names to the set of fields to
        retrieve when the SQL is executed ("immediate loading" fields). The
        field names replace any existing immediate loading field names. If
        there are field names already specified for deferred loading, remove
        those names from the new field_names before storing the new names
        for immediate loading. (That is, immediate loading overrides any
        existing immediate values, but respects existing deferrals.)
        """
        existing, defer = self.deferred_loading
        field_names = set(field_names)
        if 'pk' in field_names:
            field_names.remove('pk')
            field_names.add(self.get_meta().pk.name)

        if defer:
            # Remove any existing deferred names from the current set before
            # setting the new names.
            self.deferred_loading = field_names.difference(existing), False
        else:
            # Replace any existing "immediate load" field names.
            self.deferred_loading = frozenset(field_names), False

    def get_loaded_field_names(self):
        """
        If any fields are marked to be deferred, return a dictionary mapping
        models to a set of names in those fields that will be loaded. If a
        model is not in the returned dictionary, none of its fields are
        deferred.

        If no fields are marked for deferral, return an empty dictionary.
        """
        # We cache this because we call this function multiple times
        # (compiler.fill_related_selections, query.iterator)
        try:
            return self._loaded_field_names_cache
        except AttributeError:
            collection = {}
            self.deferred_to_data(collection, self.get_loaded_field_names_cb)
            self._loaded_field_names_cache = collection
            return collection

    def get_loaded_field_names_cb(self, target, model, fields):
        """Callback used by get_deferred_field_names()."""
        target[model] = {f.attname for f in fields}

    def set_annotation_mask(self, names):
        """Set the mask of annotations that will be returned by the SELECT."""
        if names is None:
            self.annotation_select_mask = None
        else:
            self.annotation_select_mask = set(names)
        self._annotation_select_cache = None

    def append_annotation_mask(self, names):
        if self.annotation_select_mask is not None:
            self.set_annotation_mask(self.annotation_select_mask.union(names))

    def set_extra_mask(self, names):
        """
        Set the mask of extra select items that will be returned by SELECT.
        Don't remove them from the Query since they might be used later.
        """
        if names is None:
            self.extra_select_mask = None
        else:
            self.extra_select_mask = set(names)
        self._extra_select_cache = None

    def set_values(self, fields):
        self.select_related = False
        self.clear_deferred_loading()
        self.clear_select_fields()

        if self.group_by is True:
            self.add_fields((f.attname for f in self.model._meta.concrete_fields), False)
            self.set_group_by()
            self.clear_select_fields()

        if fields:
            field_names = []
            extra_names = []
            annotation_names = []
            if not self._extra and not self._annotations:
                # Shortcut - if there are no extra or annotations, then
                # the values() clause must be just field names.
                field_names = list(fields)
            else:
                self.default_cols = False
                for f in fields:
                    if f in self.extra_select:
                        extra_names.append(f)
                    elif f in self.annotation_select:
                        annotation_names.append(f)
                    else:
                        field_names.append(f)
            self.set_extra_mask(extra_names)
            self.set_annotation_mask(annotation_names)
        else:
            field_names = [f.attname for f in self.model._meta.concrete_fields]

        self.values_select = tuple(field_names)
        self.add_fields(field_names, True)

    @property
    def annotation_select(self):
        """
        Return the OrderedDict of aggregate columns that are not masked and
        should be used in the SELECT clause. Cache this result for performance.
        """
        if self._annotation_select_cache is not None:
            return self._annotation_select_cache
        elif not self._annotations:
            return {}
        elif self.annotation_select_mask is not None:
            self._annotation_select_cache = OrderedDict(
                (k, v) for k, v in self.annotations.items()
                if k in self.annotation_select_mask
            )
            return self._annotation_select_cache
        else:
            return self.annotations

    @property
    def extra_select(self):
        if self._extra_select_cache is not None:
            return self._extra_select_cache
        if not self._extra:
            return {}
        elif self.extra_select_mask is not None:
            self._extra_select_cache = OrderedDict(
                (k, v) for k, v in self.extra.items()
                if k in self.extra_select_mask
            )
            return self._extra_select_cache
        else:
            return self.extra

    def trim_start(self, names_with_path):
        """
        Trim joins from the start of the join path. The candidates for trim
        are the PathInfos in names_with_path structure that are m2m joins.

        Also set the select column so the start matches the join.

        This method is meant to be used for generating the subquery joins &
        cols in split_exclude().

        Return a lookup usable for doing outerq.filter(lookup=self) and a
        boolean indicating if the joins in the prefix contain a LEFT OUTER join.
        _"""
        all_paths = []
        for _, paths in names_with_path:
            all_paths.extend(paths)
        contains_louter = False
        # Trim and operate only on tables that were generated for
        # the lookup part of the query. That is, avoid trimming
        # joins generated for F() expressions.
        lookup_tables = [
            t for t in self.alias_map
            if t in self._lookup_joins or t == self.base_table
        ]
        for trimmed_paths, path in enumerate(all_paths):
            if path.m2m:
                break
            if self.alias_map[lookup_tables[trimmed_paths + 1]].join_type == LOUTER:
                contains_louter = True
            alias = lookup_tables[trimmed_paths]
            self.unref_alias(alias)
        # The path.join_field is a Rel, lets get the other side's field
        join_field = path.join_field.field
        # Build the filter prefix.
        paths_in_prefix = trimmed_paths
        trimmed_prefix = []
        for name, path in names_with_path:
            if paths_in_prefix - len(path) < 0:
                break
            trimmed_prefix.append(name)
            paths_in_prefix -= len(path)
        trimmed_prefix.append(
            join_field.foreign_related_fields[0].name)
        trimmed_prefix = LOOKUP_SEP.join(trimmed_prefix)
        # Lets still see if we can trim the first join from the inner query
        # (that is, self). We can't do this for LEFT JOINs because we would
        # miss those rows that have nothing on the outer side.
        if self.alias_map[lookup_tables[trimmed_paths + 1]].join_type != LOUTER:
            select_fields = [r[0] for r in join_field.related_fields]
            select_alias = lookup_tables[trimmed_paths + 1]
            self.unref_alias(lookup_tables[trimmed_paths])
            extra_restriction = join_field.get_extra_restriction(
                self.where_class, None, lookup_tables[trimmed_paths + 1])
            if extra_restriction:
                self.where.add(extra_restriction, AND)
        else:
            # TODO: It might be possible to trim more joins from the start of the
            # inner query if it happens to have a longer join chain containing the
            # values in select_fields. Lets punt this one for now.
            select_fields = [r[1] for r in join_field.related_fields]
            select_alias = lookup_tables[trimmed_paths]
        # The found starting point is likely a Join instead of a BaseTable reference.
        # But the first entry in the query's FROM clause must not be a JOIN.
        for table in self.alias_map:
            if self.alias_refcount[table] > 0:
                self.alias_map[table] = BaseTable(self.alias_map[table].table_name, table)
                break
        self.set_select([f.get_col(select_alias) for f in select_fields])
        return trimmed_prefix, contains_louter

    def is_nullable(self, field):
        """
        Check if the given field should be treated as nullable.

        Some backends treat '' as null and Django treats such fields as
        nullable for those backends. In such situations field.null can be
        False even if we should treat the field as nullable.
        """
        # We need to use DEFAULT_DB_ALIAS here, as QuerySet does not have
        # (nor should it have) knowledge of which connection is going to be
        # used. The proper fix would be to defer all decisions where
        # is_nullable() is needed to the compiler stage, but that is not easy
        # to do currently.
        if connections[DEFAULT_DB_ALIAS].features.interprets_empty_strings_as_nulls and field.empty_strings_allowed:
            return True
        else:
            return field.null


def get_order_dir(field, default='ASC'):
    """
    Return the field name and direction for an order specification. For
    example, '-foo' is returned as ('foo', 'DESC').

    The 'default' param is used to indicate which way no prefix (or a '+'
    prefix) should sort. The '-' prefix always sorts the opposite way.
    """
    dirn = ORDER_DIR[default]
    if field[0] == '-':
        return field[1:], dirn[1]
    return field, dirn[0]


def add_to_dict(data, key, value):
    """
    Add "value" to the set of values for "key", whether or not "key" already
    exists.
    """
    if key in data:
        data[key].add(value)
    else:
        data[key] = {value}


def is_reverse_o2o(field):
    """
    Check if the given field is reverse-o2o. The field is expected to be some
    sort of relation field or related object.
    """
    return field.is_relation and field.one_to_one and not field.concrete


class JoinPromoter:
    """
    A class to abstract away join promotion problems for complex filter
    conditions.
    """

    def __init__(self, connector, num_children, negated):
        self.connector = connector
        self.negated = negated
        if self.negated:
            if connector == AND:
                self.effective_connector = OR
            else:
                self.effective_connector = AND
        else:
            self.effective_connector = self.connector
        self.num_children = num_children
        # Maps of table alias to how many times it is seen as required for
        # inner and/or outer joins.
        self.votes = Counter()

    def add_votes(self, votes):
        """
        Add single vote per item to self.votes. Parameter can be any
        iterable.
        """
        self.votes.update(votes)

    def update_join_types(self, query):
        """
        Change join types so that the generated query is as efficient as
        possible, but still correct. So, change as many joins as possible
        to INNER, but don't make OUTER joins INNER if that could remove
        results from the query.
        """
        to_promote = set()
        to_demote = set()
        # The effective_connector is used so that NOT (a AND b) is treated
        # similarly to (a OR b) for join promotion.
        for table, votes in self.votes.items():
            # We must use outer joins in OR case when the join isn't contained
            # in all of the joins. Otherwise the INNER JOIN itself could remove
            # valid results. Consider the case where a model with rel_a and
            # rel_b relations is queried with rel_a__col=1 | rel_b__col=2. Now,
            # if rel_a join doesn't produce any results is null (for example
            # reverse foreign key or null value in direct foreign key), and
            # there is a matching row in rel_b with col=2, then an INNER join
            # to rel_a would remove a valid match from the query. So, we need
            # to promote any existing INNER to LOUTER (it is possible this
            # promotion in turn will be demoted later on).
            if self.effective_connector == 'OR' and votes < self.num_children:
                to_promote.add(table)
            # If connector is AND and there is a filter that can match only
            # when there is a joinable row, then use INNER. For example, in
            # rel_a__col=1 & rel_b__col=2, if either of the rels produce NULL
            # as join output, then the col=1 or col=2 can't match (as
            # NULL=anything is always false).
            # For the OR case, if all children voted for a join to be inner,
            # then we can use INNER for the join. For example:
            #     (rel_a__col__icontains=Alex | rel_a__col__icontains=Russell)
            # then if rel_a doesn't produce any rows, the whole condition
            # can't match. Hence we can safely use INNER join.
            if self.effective_connector == 'AND' or (
                    self.effective_connector == 'OR' and votes == self.num_children):
                to_demote.add(table)
            # Finally, what happens in cases where we have:
            #    (rel_a__col=1|rel_b__col=2) & rel_a__col__gte=0
            # Now, we first generate the OR clause, and promote joins for it
            # in the first if branch above. Both rel_a and rel_b are promoted
            # to LOUTER joins. After that we do the AND case. The OR case
            # voted no inner joins but the rel_a__col__gte=0 votes inner join
            # for rel_a. We demote it back to INNER join (in AND case a single
            # vote is enough). The demotion is OK, if rel_a doesn't produce
            # rows, then the rel_a__col__gte=0 clause can't be true, and thus
            # the whole clause must be false. So, it is safe to use INNER
            # join.
            # Note that in this example we could just as well have the __gte
            # clause and the OR clause swapped. Or we could replace the __gte
            # clause with an OR clause containing rel_a__col=1|rel_a__col=2,
            # and again we could safely demote to INNER.
        query.promote_joins(to_promote)
        query.demote_joins(to_demote)
        return to_demote