View on GitHub


1 wk
Test Coverage
import collections
import re
from functools import partial
from itertools import chain

from django.core.exceptions import EmptyResultSet, FieldError
from django.db import DatabaseError, NotSupportedError
from django.db.models.constants import LOOKUP_SEP
from django.db.models.expressions import F, OrderBy, Random, RawSQL, Ref, Value
from django.db.models.functions import Cast
from django.db.models.query_utils import Q, select_related_descend
from django.db.models.sql.constants import (
from django.db.models.sql.query import Query, get_order_dir
from django.db.transaction import TransactionManagementError
from django.utils.functional import cached_property
from django.utils.hashable import make_hashable
from django.utils.regex_helper import _lazy_re_compile

class SQLCompiler:
    # Multiline ordering SQL clause may appear from RawSQL.
    ordering_parts = _lazy_re_compile(
        re.MULTILINE | re.DOTALL,

    def __init__(self, query, connection, using):
        self.query = query
        self.connection = connection
        self.using = using
        self.quote_cache = {'*': '*'}
        # The select, klass_info, and annotations are needed by QuerySet.iterator()
        # these are set as a side-effect of executing the query. Note that we calculate
        # separately a list of extra select columns needed for grammatical correctness
        # of the query, but these columns are not included in = None
        self.annotation_col_map = None
        self.klass_info = None
        self._meta_ordering = None

    def setup_query(self):
        if all(self.query.alias_refcount[a] == 0 for a in self.query.alias_map):
            self.query.get_initial_alias(), self.klass_info, self.annotation_col_map = self.get_select()
        self.col_count = len(

    def pre_sql_setup(self):
        Do any necessary class setup immediately prior to producing SQL. This
        is for things that can't necessarily be done in __init__ because we
        might not have all the pieces in place at that time.
        order_by = self.get_order_by()
        self.where, self.having = self.query.where.split_having()
        extra_select = self.get_extra_select(order_by,
        self.has_extra_select = bool(extra_select)
        group_by = self.get_group_by( + extra_select, order_by)
        return extra_select, order_by, group_by

    def get_group_by(self, select, order_by):
        Return a list of 2-tuples of form (sql, params).

        The logic of what exactly the GROUP BY clause contains is hard
        to describe in other words than "if it passes the test suite,
        then it is correct".
        # Some examples:
        #     SomeModel.objects.annotate(Count('somecol'))
        #     GROUP BY: all fields of the model
        #    SomeModel.objects.values('name').annotate(Count('somecol'))
        #    GROUP BY: name
        #    SomeModel.objects.annotate(Count('somecol')).values('name')
        #    GROUP BY: all cols of the model
        #    SomeModel.objects.values('name', 'pk').annotate(Count('somecol')).values('pk')
        #    GROUP BY: name, pk
        #    SomeModel.objects.values('name').annotate(Count('somecol')).values('pk')
        #    GROUP BY: name, pk
        # In fact, the self.query.group_by is the minimal set to GROUP BY. It
        # can't be ever restricted to a smaller set, but additional columns in
        # HAVING, ORDER BY, and SELECT clauses are added to it. Unfortunately
        # the end result is that it is impossible to force the query to have
        # a chosen GROUP BY clause - you can almost do this by using the form:
        #     .values(*wanted_cols).annotate(AnAggregate())
        # but any later annotations, extra selects, values calls that
        # refer some column outside of the wanted_cols, order_by, or even
        # filter calls can alter the GROUP BY clause.

        # The query.group_by is either None (no GROUP BY at all), True
        # (group by select fields), or a list of expressions to be added
        # to the group by.
        if self.query.group_by is None:
            return []
        expressions = []
        if self.query.group_by is not True:
            # If the group by is set to a list (by .values() call most likely),
            # then we need to add everything in it to the GROUP BY clause.
            # Backwards compatibility hack for setting query.group_by. Remove
            # when  we have public API way of forcing the GROUP BY clause.
            # Converts string references to expressions.
            for expr in self.query.group_by:
                if not hasattr(expr, 'as_sql'):
        # Note that even if the group_by is set, it is only the minimal
        # set to group by. So, we need to add cols in select, order_by, and
        # having into the select in any case.
        ref_sources = {
            expr.source for expr in expressions if isinstance(expr, Ref)
        for expr, _, _ in select:
            # Skip members of the select clause that are already included
            # by reference.
            if expr in ref_sources:
            cols = expr.get_group_by_cols()
            for col in cols:
        for expr, (sql, params, is_ref) in order_by:
            # Skip References to the select clause, as all expressions in the
            # select clause are already part of the group by.
            if not is_ref:
        having_group_by = self.having.get_group_by_cols() if self.having else ()
        for expr in having_group_by:
        result = []
        seen = set()
        expressions = self.collapse_group_by(expressions, having_group_by)

        for expr in expressions:
            sql, params = self.compile(expr)
            sql, params = expr.select_format(self, sql, params)
            params_hash = make_hashable(params)
            if (sql, params_hash) not in seen:
                result.append((sql, params))
                seen.add((sql, params_hash))
        return result

    def collapse_group_by(self, expressions, having):
        # If the DB can group by primary key, then group by the primary key of
        # query's main model. Note that for PostgreSQL the GROUP BY clause must
        # include the primary key of every table, but for MySQL it is enough to
        # have the main table's primary key.
        if self.connection.features.allows_group_by_pk:
            # Determine if the main model's primary key is in the query.
            pk = None
            for expr in expressions:
                # Is this a reference to query's base table primary key? If the
                # expression isn't a Col-like, then skip the expression.
                if (getattr(expr, 'target', None) == and
                        getattr(expr, 'alias', None) == self.query.base_table):
                    pk = expr
            # If the main model's primary key is in the query, group by that
            # field, HAVING expressions, and expressions associated with tables
            # that don't have a primary key included in the grouped columns.
            if pk:
                pk_aliases = {
                    expr.alias for expr in expressions
                    if hasattr(expr, 'target') and
                expressions = [pk] + [
                    expr for expr in expressions
                    if expr in having or (
                        getattr(expr, 'alias', None) is not None and expr.alias not in pk_aliases
        elif self.connection.features.allows_group_by_selected_pks:
            # Filter out all expressions associated with a table's primary key
            # present in the grouped columns. This is done by identifying all
            # tables that have their primary key included in the grouped
            # columns and removing non-primary key columns referring to them.
            # Unmanaged models are excluded because they could be representing
            # database views on which the optimization might not be allowed.
            pks = {
                expr for expr in expressions
                if (
                    hasattr(expr, 'target') and
            aliases = {expr.alias for expr in pks}
            expressions = [
                expr for expr in expressions if expr in pks or getattr(expr, 'alias', None) not in aliases
        return expressions

    def get_select(self):
        Return three values:
        - a list of 3-tuples of (expression, (sql, params), alias)
        - a klass_info structure,
        - a dictionary of annotations

        The (sql, params) is what the expression will produce, and alias is the
        "AS alias" for the column (possibly None).

        The klass_info structure contains the following information:
        - The base model of the query.
        - Which columns for that model are present in the query (by
          position of the select clause).
        - related_klass_infos: [f, klass_info] to descent into

        The annotations is a dictionary of {'attname': column position} values.
        select = []
        klass_info = None
        annotations = {}
        select_idx = 0
        for alias, (sql, params) in self.query.extra_select.items():
            annotations[alias] = select_idx
            select.append((RawSQL(sql, params), alias))
            select_idx += 1
        assert not ( and self.query.default_cols)
        if self.query.default_cols:
            cols = self.get_default_columns()
            # is a special case. These columns never go to
            # any model.
            cols =
        if cols:
            select_list = []
            for col in cols:
                select.append((col, None))
                select_idx += 1
            klass_info = {
                'model': self.query.model,
                'select_fields': select_list,
        for alias, annotation in self.query.annotation_select.items():
            annotations[alias] = select_idx
            select.append((annotation, alias))
            select_idx += 1

        if self.query.select_related:
            related_klass_infos = self.get_related_selections(select)
            klass_info['related_klass_infos'] = related_klass_infos

            def get_select_from_parent(klass_info):
                for ki in klass_info['related_klass_infos']:
                    if ki['from_parent']:
                        ki['select_fields'] = (klass_info['select_fields'] +

        ret = []
        for col, alias in select:
                sql, params = self.compile(col)
            except EmptyResultSet:
                # Select a predicate that's always False.
                sql, params = '0', ()
                sql, params = col.select_format(self, sql, params)
            ret.append((col, (sql, params), alias))
        return ret, klass_info, annotations

    def get_order_by(self):
        Return a list of 2-tuples of form (expr, (sql, params, is_ref)) for the
        ORDER BY clause.

        The order_by clause can alter the select clause (for example it
        can add aliases to clauses that do not yet have one, or it can
        add totally new select clauses).
        if self.query.extra_order_by:
            ordering = self.query.extra_order_by
        elif not self.query.default_ordering:
            ordering = self.query.order_by
        elif self.query.order_by:
            ordering = self.query.order_by
        elif self.query.get_meta().ordering:
            ordering = self.query.get_meta().ordering
            self._meta_ordering = ordering
            ordering = []
        if self.query.standard_ordering:
            asc, desc = ORDER_DIR['ASC']
            asc, desc = ORDER_DIR['DESC']

        order_by = []
        for field in ordering:
            if hasattr(field, 'resolve_expression'):
                if isinstance(field, Value):
                    # output_field must be resolved for constants.
                    field = Cast(field, field.output_field)
                if not isinstance(field, OrderBy):
                    field = field.asc()
                if not self.query.standard_ordering:
                    field = field.copy()
                order_by.append((field, False))
            if field == '?':  # random
                order_by.append((OrderBy(Random()), False))

            col, order = get_order_dir(field, asc)
            descending = order == 'DESC'

            if col in self.query.annotation_select:
                # Reference to expression in SELECT clause
                    OrderBy(Ref(col, self.query.annotation_select[col]), descending=descending),
            if col in self.query.annotations:
                # References to an expression which is masked out of the SELECT
                # clause.
                expr = self.query.annotations[col]
                if isinstance(expr, Value):
                    # output_field must be resolved for constants.
                    expr = Cast(expr, expr.output_field)
                order_by.append((OrderBy(expr, descending=descending), False))

            if '.' in field:
                # This came in through an extra(order_by=...) addition. Pass it
                # on verbatim.
                table, col = col.split('.', 1)
                        RawSQL('%s.%s' % (self.quote_name_unless_alias(table), col), []),
                    ), False))

            if not self.query.extra or col not in self.query.extra:
                # 'col' is of the form 'field' or 'field1__field2' or
                # '-field1__field2__field', etc.
                    field, self.query.get_meta(), default_order=asc))
                if col not in self.query.extra_select:
                        OrderBy(RawSQL(*self.query.extra[col]), descending=descending),
                        OrderBy(Ref(col, RawSQL(*self.query.extra[col])), descending=descending),
        result = []
        seen = set()

        for expr, is_ref in order_by:
            resolved = expr.resolve_expression(self.query, allow_joins=True, reuse=None)
            if self.query.combinator:
                src = resolved.get_source_expressions()[0]
                expr_src = expr.get_source_expressions()[0]
                # Relabel order by columns to raw numbers if this is a combined
                # query; necessary since the columns can't be referenced by the
                # fully qualified name and the simple column names may collide.
                for idx, (sel_expr, _, col_alias) in enumerate(
                    if is_ref and col_alias == src.refs:
                        src = src.source
                    elif col_alias and not (
                        isinstance(expr_src, F) and col_alias ==
                    if src == sel_expr:
                        resolved.set_source_expressions([RawSQL('%d' % (idx + 1), ())])
                    if col_alias:
                        raise DatabaseError('ORDER BY term does not match any column in the result set.')
                    # Add column used in ORDER BY clause without an alias to
                    # the selected columns.
                    resolved.set_source_expressions([RawSQL('%d' % len(, ())])
            sql, params = self.compile(resolved)
            # Don't add the same column twice, but the order direction is
            # not taken into account so we strip it. When this entire method
            # is refactored into expressions, then we can check each part as we
            # generate it.
            without_ordering =[1]
            params_hash = make_hashable(params)
            if (without_ordering, params_hash) in seen:
            seen.add((without_ordering, params_hash))
            result.append((resolved, (sql, params, is_ref)))
        return result

    def get_extra_select(self, order_by, select):
        extra_select = []
        if self.query.distinct and not self.query.distinct_fields:
            select_sql = [t[1] for t in select]
            for expr, (sql, params, is_ref) in order_by:
                without_ordering =[1]
                if not is_ref and (without_ordering, params) not in select_sql:
                    extra_select.append((expr, (without_ordering, params), None))
        return extra_select

    def quote_name_unless_alias(self, name):
        A wrapper around connection.ops.quote_name that doesn't quote aliases
        for table names. This avoids problems with some SQL dialects that treat
        quoted strings specially (e.g. PostgreSQL).
        if name in self.quote_cache:
            return self.quote_cache[name]
        if ((name in self.query.alias_map and name not in self.query.table_map) or
                name in self.query.extra_select or (
                    self.query.external_aliases.get(name) and name not in self.query.table_map)):
            self.quote_cache[name] = name
            return name
        r = self.connection.ops.quote_name(name)
        self.quote_cache[name] = r
        return r

    def compile(self, node):
        vendor_impl = getattr(node, 'as_' + self.connection.vendor, None)
        if vendor_impl:
            sql, params = vendor_impl(self, self.connection)
            sql, params = node.as_sql(self, self.connection)
        return sql, params

    def get_combinator_sql(self, combinator, all):
        features = self.connection.features
        compilers = [
            query.get_compiler(self.using, self.connection)
            for query in self.query.combined_queries if not query.is_empty()
        if not features.supports_slicing_ordering_in_compound:
            for query, compiler in zip(self.query.combined_queries, compilers):
                if query.low_mark or query.high_mark:
                    raise DatabaseError('LIMIT/OFFSET not allowed in subqueries of compound statements.')
                if compiler.get_order_by():
                    raise DatabaseError('ORDER BY not allowed in subqueries of compound statements.')
        parts = ()
        for compiler in compilers:
                # If the columns list is limited, then all combined queries
                # must have the same columns list. Set the selects defined on
                # the query on all combined queries, if not already set.
                if not compiler.query.values_select and self.query.values_select:
                    compiler.query = compiler.query.clone()
                part_sql, part_args = compiler.as_sql()
                if compiler.query.combinator:
                    # Wrap in a subquery if wrapping in parentheses isn't
                    # supported.
                    if not features.supports_parentheses_in_compound:
                        part_sql = 'SELECT * FROM ({})'.format(part_sql)
                    # Add parentheses when combining with compound query if not
                    # already added for all compound queries.
                    elif not features.supports_slicing_ordering_in_compound:
                        part_sql = '({})'.format(part_sql)
                parts += ((part_sql, part_args),)
            except EmptyResultSet:
                # Omit the empty queryset with UNION and with DIFFERENCE if the
                # first queryset is nonempty.
                if combinator == 'union' or (combinator == 'difference' and parts):
        if not parts:
            raise EmptyResultSet
        combinator_sql = self.connection.ops.set_operators[combinator]
        if all and combinator == 'union':
            combinator_sql += ' ALL'
        braces = '({})' if features.supports_slicing_ordering_in_compound else '{}'
        sql_parts, args_parts = zip(*((braces.format(sql), args) for sql, args in parts))
        result = [' {} '.format(combinator_sql).join(sql_parts)]
        params = []
        for part in args_parts:
        return result, params

    def as_sql(self, with_limits=True, with_col_aliases=False):
        Create the SQL for this query. Return the SQL string and list of

        If 'with_limits' is False, any limit/offset information is not included
        in the query.
        refcounts_before = self.query.alias_refcount.copy()
            extra_select, order_by, group_by = self.pre_sql_setup()
            for_update_part = None
            # Is a LIMIT/OFFSET clause needed?
            with_limit_offset = with_limits and (self.query.high_mark is not None or self.query.low_mark)
            combinator = self.query.combinator
            features = self.connection.features
            if combinator:
                if not getattr(features, 'supports_select_{}'.format(combinator)):
                    raise NotSupportedError('{} is not supported on this database backend.'.format(combinator))
                result, params = self.get_combinator_sql(combinator, self.query.combinator_all)
                distinct_fields, distinct_params = self.get_distinct()
                # This must come after 'select', 'ordering', and 'distinct'
                # (see docstring of get_from_clause() for details).
                from_, f_params = self.get_from_clause()
                where, w_params = self.compile(self.where) if self.where is not None else ("", [])
                having, h_params = self.compile(self.having) if self.having is not None else ("", [])
                result = ['SELECT']
                params = []

                if self.query.distinct:
                    distinct_result, distinct_params = self.connection.ops.distinct_sql(
                    result += distinct_result
                    params += distinct_params

                out_cols = []
                col_idx = 1
                for _, (s_sql, s_params), alias in + extra_select:
                    if alias:
                        s_sql = '%s AS %s' % (s_sql, self.connection.ops.quote_name(alias))
                    elif with_col_aliases:
                        s_sql = '%s AS %s' % (s_sql, 'Col%d' % col_idx)
                        col_idx += 1

                result += [', '.join(out_cols), 'FROM', *from_]

                if self.query.select_for_update and self.connection.features.has_select_for_update:
                    if self.connection.get_autocommit():
                        raise TransactionManagementError('select_for_update cannot be used outside of a transaction.')

                    if with_limit_offset and not self.connection.features.supports_select_for_update_with_limit:
                        raise NotSupportedError(
                            'LIMIT/OFFSET is not supported with '
                            'select_for_update on this database backend.'
                    nowait = self.query.select_for_update_nowait
                    skip_locked = self.query.select_for_update_skip_locked
                    of = self.query.select_for_update_of
                    no_key = self.query.select_for_no_key_update
                    # If it's a NOWAIT/SKIP LOCKED/OF/NO KEY query but the
                    # backend doesn't support it, raise NotSupportedError to
                    # prevent a possible deadlock.
                    if nowait and not self.connection.features.has_select_for_update_nowait:
                        raise NotSupportedError('NOWAIT is not supported on this database backend.')
                    elif skip_locked and not self.connection.features.has_select_for_update_skip_locked:
                        raise NotSupportedError('SKIP LOCKED is not supported on this database backend.')
                    elif of and not self.connection.features.has_select_for_update_of:
                        raise NotSupportedError('FOR UPDATE OF is not supported on this database backend.')
                    elif no_key and not self.connection.features.has_select_for_no_key_update:
                        raise NotSupportedError(
                            'FOR NO KEY UPDATE is not supported on this '
                            'database backend.'
                    for_update_part = self.connection.ops.for_update_sql(

                if for_update_part and self.connection.features.for_update_after_from:

                if where:
                    result.append('WHERE %s' % where)

                grouping = []
                for g_sql, g_params in group_by:
                if grouping:
                    if distinct_fields:
                        raise NotImplementedError('annotate() + distinct(fields) is not implemented.')
                    order_by = order_by or self.connection.ops.force_no_ordering()
                    result.append('GROUP BY %s' % ', '.join(grouping))
                    if self._meta_ordering:
                        order_by = None
                if having:
                    result.append('HAVING %s' % having)

            if self.query.explain_query:
                result.insert(0, self.connection.ops.explain_query_prefix(

            if order_by:
                ordering = []
                for _, (o_sql, o_params, _) in order_by:
                result.append('ORDER BY %s' % ', '.join(ordering))

            if with_limit_offset:
                result.append(self.connection.ops.limit_offset_sql(self.query.low_mark, self.query.high_mark))

            if for_update_part and not self.connection.features.for_update_after_from:

            if self.query.subquery and extra_select:
                # If the query is used as a subquery, the extra selects would
                # result in more columns than the left-hand side expression is
                # expecting. This can happen when a subquery uses a combination
                # of order_by() and distinct(), forcing the ordering expressions
                # to be selected as well. Wrap the query in another subquery
                # to exclude extraneous selects.
                sub_selects = []
                sub_params = []
                for index, (select, _, alias) in enumerate(, start=1):
                    if not alias and with_col_aliases:
                        alias = 'col%d' % index
                    if alias:
                        sub_selects.append("%s.%s" % (
                        select_clone = select.relabeled_clone({select.alias: 'subquery'})
                        subselect, subparams = select_clone.as_sql(self, self.connection)
                return 'SELECT %s FROM (%s) subquery' % (
                    ', '.join(sub_selects),
                    ' '.join(result),
                ), tuple(sub_params + params)

            return ' '.join(result), tuple(params)
            # Finally do cleanup - get rid of the joins we created above.

    def get_default_columns(self, start_alias=None, opts=None, from_parent=None):
        Compute the default columns for selecting every field in the base
        model. Will sometimes be called to pull in related models (e.g. via
        select_related), in which case "opts" and "start_alias" will be given
        to provide a starting point for the traversal.

        Return a list of strings, quoted appropriately for use in SQL
        directly, as well as a set of aliases used in the select statement (if
        'as_pairs' is True, return a list of (alias, col_name) pairs instead
        of strings as the first component and None as the second component).
        result = []
        if opts is None:
            opts = self.query.get_meta()
        only_load = self.deferred_to_columns()
        start_alias = start_alias or self.query.get_initial_alias()
        # The 'seen_models' is used to optimize checking the needed parent
        # alias for a given field. This also includes None -> start_alias to
        # be used by local fields.
        seen_models = {None: start_alias}

        for field in opts.concrete_fields:
            model = field.model._meta.concrete_model
            # A proxy model will have a different model and concrete_model. We
            # will assign None if the field belongs to this model.
            if model == opts.model:
                model = None
            if from_parent and model is not None and issubclass(
                    from_parent._meta.concrete_model, model._meta.concrete_model):
                # Avoid loading data for already loaded parents.
                # We end up here in the case select_related() resolution
                # proceeds from parent model to child model. In that case the
                # parent model data is already present in the SELECT clause,
                # and we want to avoid reloading the same data again.
            if field.model in only_load and field.attname not in only_load[field.model]:
            alias = self.query.join_parent_model(opts, model, start_alias,
            column = field.get_col(alias)
        return result

    def get_distinct(self):
        Return a quoted list of fields to use in DISTINCT ON part of the query.

        This method can alter the tables in the query, and thus it must be
        called before get_from_clause().
        result = []
        params = []
        opts = self.query.get_meta()

        for name in self.query.distinct_fields:
            parts = name.split(LOOKUP_SEP)
            _, targets, alias, joins, path, _, transform_function = self._setup_joins(parts, opts, None)
            targets, alias, _ = self.query.trim_joins(targets, joins, path)
            for target in targets:
                if name in self.query.annotation_select:
                    r, p = self.compile(transform_function(target, alias))
        return result, params

    def find_ordering_name(self, name, opts, alias=None, default_order='ASC',
        Return the table alias (the name might be ambiguous, the alias will
        not be) and column name for ordering by the given 'name' parameter.
        The 'name' is of the form 'field1__field2__...__fieldN'.
        name, order = get_order_dir(name, default_order)
        descending = order == 'DESC'
        pieces = name.split(LOOKUP_SEP)
        field, targets, alias, joins, path, opts, transform_function = self._setup_joins(pieces, opts, alias)

        # If we get to this point and the field is a relation to another model,
        # append the default ordering for that model unless it is the pk
        # shortcut or the attribute name of the field that is specified.
        if (
            field.is_relation and
            opts.ordering and
            getattr(field, 'attname', None) != pieces[-1] and
            name != 'pk'
            # Firstly, avoid infinite loops.
            already_seen = already_seen or set()
            join_tuple = tuple(getattr(self.query.alias_map[j], 'join_cols', None) for j in joins)
            if join_tuple in already_seen:
                raise FieldError('Infinite loop caused by ordering.')

            results = []
            for item in opts.ordering:
                if hasattr(item, 'resolve_expression') and not isinstance(item, OrderBy):
                    item = item.desc() if descending else item.asc()
                if isinstance(item, OrderBy):
                    results.append((item, False))
                results.extend(self.find_ordering_name(item, opts, alias,
                                                       order, already_seen))
            return results
        targets, alias, _ = self.query.trim_joins(targets, joins, path)
        return [(OrderBy(transform_function(t, alias), descending=descending), False) for t in targets]

    def _setup_joins(self, pieces, opts, alias):
        Helper method for get_order_by() and get_distinct().

        get_ordering() and get_distinct() must produce same target columns on
        same input, as the prefixes of get_ordering() and get_distinct() must
        match. Executing SQL where this is not true is an error.
        alias = alias or self.query.get_initial_alias()
        field, targets, opts, joins, path, transform_function = self.query.setup_joins(pieces, opts, alias)
        alias = joins[-1]
        return field, targets, alias, joins, path, opts, transform_function

    def get_from_clause(self):
        Return a list of strings that are joined together to go after the
        "FROM" part of the query, as well as a list any extra parameters that
        need to be included. Subclasses, can override this to create a
        from-clause via a "select".

        This should only be called after any SQL construction methods that
        might change the tables that are needed. This means the select columns,
        ordering, and distinct must be done first.
        result = []
        params = []
        for alias in tuple(self.query.alias_map):
            if not self.query.alias_refcount[alias]:
                from_clause = self.query.alias_map[alias]
            except KeyError:
                # Extra tables can end up in self.tables, but not in the
                # alias_map if they aren't in a join. That's OK. We skip them.
            clause_sql, clause_params = self.compile(from_clause)
        for t in self.query.extra_tables:
            alias, _ = self.query.table_alias(t)
            # Only add the alias if it's not already present (the table_alias()
            # call increments the refcount, so an alias refcount of one means
            # this is the only reference).
            if alias not in self.query.alias_map or self.query.alias_refcount[alias] == 1:
                result.append(', %s' % self.quote_name_unless_alias(alias))
        return result, params

    def get_related_selections(self, select, opts=None, root_alias=None, cur_depth=1,
                               requested=None, restricted=None):
        Fill in the information needed for a select_related query. The current
        depth is measured as the number of connections away from the root model
        (for example, cur_depth=1 means we are looking at models with direct
        connections to the root model).
        def _get_field_choices():
            direct_choices = ( for f in opts.fields if f.is_relation)
            reverse_choices = (
                for f in opts.related_objects if f.field.unique
            return chain(direct_choices, reverse_choices, self.query._filtered_relations)

        related_klass_infos = []
        if not restricted and cur_depth > self.query.max_depth:
            # We've recursed far enough; bail out.
            return related_klass_infos

        if not opts:
            opts = self.query.get_meta()
            root_alias = self.query.get_initial_alias()
        only_load = self.query.get_loaded_field_names()

        # Setup for the case when only particular related fields should be
        # included in the related selection.
        fields_found = set()
        if requested is None:
            restricted = isinstance(self.query.select_related, dict)
            if restricted:
                requested = self.query.select_related

        def get_related_klass_infos(klass_info, related_klass_infos):
            klass_info['related_klass_infos'] = related_klass_infos

        for f in opts.fields:
            field_model = f.model._meta.concrete_model

            if restricted:
                next = requested.get(, {})
                if not f.is_relation:
                    # If a non-related field is used like a relation,
                    # or if a single non-relational field is given.
                    if next or in requested:
                        raise FieldError(
                            "Non-relational field given in select_related: '%s'. "
                            "Choices are: %s" % (
                                ", ".join(_get_field_choices()) or '(none)',
                next = False

            if not select_related_descend(f, restricted, requested,
            klass_info = {
                'model': f.remote_field.model,
                'field': f,
                'reverse': False,
                'local_setter': f.set_cached_value,
                'remote_setter': f.remote_field.set_cached_value if f.unique else lambda x, y: None,
                'from_parent': False,
            select_fields = []
            _, _, _, joins, _, _ = self.query.setup_joins(
                [], opts, root_alias)
            alias = joins[-1]
            columns = self.get_default_columns(start_alias=alias, opts=f.remote_field.model._meta)
            for col in columns:
                select.append((col, None))
            klass_info['select_fields'] = select_fields
            next_klass_infos = self.get_related_selections(
                select, f.remote_field.model._meta, alias, cur_depth + 1, next, restricted)
            get_related_klass_infos(klass_info, next_klass_infos)

        if restricted:
            related_fields = [
                (o.field, o.related_model)
                for o in opts.related_objects
                if o.field.unique and not o.many_to_many
            for f, model in related_fields:
                if not select_related_descend(f, restricted, requested,
                                              only_load.get(model), reverse=True):

                related_field_name = f.related_query_name()

                join_info = self.query.setup_joins([related_field_name], opts, root_alias)
                alias = join_info.joins[-1]
                from_parent = issubclass(model, opts.model) and model is not opts.model
                klass_info = {
                    'model': model,
                    'field': f,
                    'reverse': True,
                    'local_setter': f.remote_field.set_cached_value,
                    'remote_setter': f.set_cached_value,
                    'from_parent': from_parent,
                select_fields = []
                columns = self.get_default_columns(
                    start_alias=alias, opts=model._meta, from_parent=opts.model)
                for col in columns:
                    select.append((col, None))
                klass_info['select_fields'] = select_fields
                next = requested.get(f.related_query_name(), {})
                next_klass_infos = self.get_related_selections(
                    select, model._meta, alias, cur_depth + 1,
                    next, restricted)
                get_related_klass_infos(klass_info, next_klass_infos)

            def local_setter(obj, from_obj):
                # Set a reverse fk object when relation is non-empty.
                if from_obj:
                    f.remote_field.set_cached_value(from_obj, obj)

            def remote_setter(name, obj, from_obj):
                setattr(from_obj, name, obj)

            for name in list(requested):
                # Filtered relations work only on the topmost level.
                if cur_depth > 1:
                if name in self.query._filtered_relations:
                    f, _, join_opts, joins, _, _ = self.query.setup_joins([name], opts, root_alias)
                    model = join_opts.model
                    alias = joins[-1]
                    from_parent = issubclass(model, opts.model) and model is not opts.model
                    klass_info = {
                        'model': model,
                        'field': f,
                        'reverse': True,
                        'local_setter': local_setter,
                        'remote_setter': partial(remote_setter, name),
                        'from_parent': from_parent,
                    select_fields = []
                    columns = self.get_default_columns(
                        start_alias=alias, opts=model._meta,
                    for col in columns:
                        select.append((col, None))
                    klass_info['select_fields'] = select_fields
                    next_requested = requested.get(name, {})
                    next_klass_infos = self.get_related_selections(
                        select, opts=model._meta, root_alias=alias,
                        cur_depth=cur_depth + 1, requested=next_requested,
                    get_related_klass_infos(klass_info, next_klass_infos)
            fields_not_found = set(requested).difference(fields_found)
            if fields_not_found:
                invalid_fields = ("'%s'" % s for s in fields_not_found)
                raise FieldError(
                    'Invalid field name(s) given in select_related: %s. '
                    'Choices are: %s' % (
                        ', '.join(invalid_fields),
                        ', '.join(_get_field_choices()) or '(none)',
        return related_klass_infos

    def get_select_for_update_of_arguments(self):
        Return a quoted list of arguments for the SELECT FOR UPDATE OF part of
        the query.
        def _get_parent_klass_info(klass_info):
            for parent_model, parent_link in klass_info['model']._meta.parents.items():
                parent_list = parent_model._meta.get_parent_list()
                yield {
                    'model': parent_model,
                    'field': parent_link,
                    'reverse': False,
                    'select_fields': [
                        for select_index in klass_info['select_fields']
                        # Selected columns from a model or its parents.
                        if (
                  [select_index][0].target.model == parent_model or
                  [select_index][0].target.model in parent_list

        def _get_first_selected_col_from_model(klass_info):
            Find the first selected column from a model. If it doesn't exist,
            don't lock a model.

            select_fields is filled recursively, so it also contains fields
            from the parent models.
            for select_index in klass_info['select_fields']:
                if[select_index][0].target.model == klass_info['model']:

        def _get_field_choices():
            """Yield all allowed field paths in breadth-first search order."""
            queue = collections.deque([(None, self.klass_info)])
            while queue:
                parent_path, klass_info = queue.popleft()
                if parent_path is None:
                    path = []
                    yield 'self'
                    field = klass_info['field']
                    if klass_info['reverse']:
                        field = field.remote_field
                    path = parent_path + []
                    yield LOOKUP_SEP.join(path)
                    (path, klass_info)
                    for klass_info in _get_parent_klass_info(klass_info)
                    (path, klass_info)
                    for klass_info in klass_info.get('related_klass_infos', [])
        result = []
        invalid_names = []
        for name in self.query.select_for_update_of:
            klass_info = self.klass_info
            if name == 'self':
                col = _get_first_selected_col_from_model(klass_info)
                for part in name.split(LOOKUP_SEP):
                    klass_infos = (
                        *klass_info.get('related_klass_infos', []),
                    for related_klass_info in klass_infos:
                        field = related_klass_info['field']
                        if related_klass_info['reverse']:
                            field = field.remote_field
                        if == part:
                            klass_info = related_klass_info
                        klass_info = None
                if klass_info is None:
                col = _get_first_selected_col_from_model(klass_info)
            if col is not None:
                if self.connection.features.select_for_update_of_column:
        if invalid_names:
            raise FieldError(
                'Invalid field name(s) given in select_for_update(of=(...)): %s. '
                'Only relational fields followed in the query are allowed. '
                'Choices are: %s.' % (
                    ', '.join(invalid_names),
                    ', '.join(_get_field_choices()),
        return result

    def deferred_to_columns(self):
        Convert the self.deferred_loading data structure to mapping of table
        names to sets of column names which are to be loaded. Return the
        columns = {}
        self.query.deferred_to_data(columns, self.query.get_loaded_field_names_cb)
        return columns

    def get_converters(self, expressions):
        converters = {}
        for i, expression in enumerate(expressions):
            if expression:
                backend_converters = self.connection.ops.get_db_converters(expression)
                field_converters = expression.get_db_converters(self.connection)
                if backend_converters or field_converters:
                    converters[i] = (backend_converters + field_converters, expression)
        return converters

    def apply_converters(self, rows, converters):
        connection = self.connection
        converters = list(converters.items())
        for row in map(list, rows):
            for pos, (convs, expression) in converters:
                value = row[pos]
                for converter in convs:
                    value = converter(value, expression, connection)
                row[pos] = value
            yield row

    def results_iter(self, results=None, tuple_expected=False, chunked_fetch=False,
        """Return an iterator over the results from executing this query."""
        if results is None:
            results = self.execute_sql(MULTI, chunked_fetch=chunked_fetch, chunk_size=chunk_size)
        fields = [s[0] for s in[0:self.col_count]]
        converters = self.get_converters(fields)
        rows = chain.from_iterable(results)
        if converters:
            rows = self.apply_converters(rows, converters)
            if tuple_expected:
                rows = map(tuple, rows)
        return rows

    def has_results(self):
        Backends (e.g. NoSQL) can override this in order to use optimized
        versions of "query has any results."
        # This is always executed on a query clone, so we can modify self.query
        self.query.add_extra({'a': 1}, None, None, None, None, None)
        return bool(self.execute_sql(SINGLE))

    def execute_sql(self, result_type=MULTI, chunked_fetch=False, chunk_size=GET_ITERATOR_CHUNK_SIZE):
        Run the query against the database and return the result(s). The
        return value is a single data item if result_type is SINGLE, or an
        iterator over the results if the result_type is MULTI.

        result_type is either MULTI (use fetchmany() to retrieve all rows),
        SINGLE (only retrieve a single row), or None. In this last case, the
        cursor is returned if any query is executed, since it's used by
        subclasses such as InsertQuery). It's possible, however, that no query
        is needed, as the filters describe an empty set. In that case, None is
        returned, to avoid any unnecessary database interaction.
        result_type = result_type or NO_RESULTS
            sql, params = self.as_sql()
            if not sql:
                raise EmptyResultSet
        except EmptyResultSet:
            if result_type == MULTI:
                return iter([])
        if chunked_fetch:
            cursor = self.connection.chunked_cursor()
            cursor = self.connection.cursor()
            cursor.execute(sql, params)
        except Exception:
            # Might fail for server-side cursors (e.g. connection closed)

        if result_type == CURSOR:
            # Give the caller the cursor to process and close.
            return cursor
        if result_type == SINGLE:
                val = cursor.fetchone()
                if val:
                    return val[0:self.col_count]
                return val
                # done with the cursor
        if result_type == NO_RESULTS:

        result = cursor_iter(
            cursor, self.connection.features.empty_fetchmany_value,
            self.col_count if self.has_extra_select else None,
        if not chunked_fetch or not self.connection.features.can_use_chunked_reads:
                # If we are using non-chunked reads, we return the same data
                # structure as normally, but ensure it is all read into memory
                # before going any further. Use chunked_fetch if requested,
                # unless the database doesn't support it.
                return list(result)
                # done with the cursor
        return result

    def as_subquery_condition(self, alias, columns, compiler):
        qn = compiler.quote_name_unless_alias
        qn2 = self.connection.ops.quote_name

        for index, select_col in enumerate(
            lhs_sql, lhs_params = self.compile(select_col)
            rhs = '%s.%s' % (qn(alias), qn2(columns[index]))
                RawSQL('%s = %s' % (lhs_sql, rhs), lhs_params), 'AND')

        sql, params = self.as_sql()
        return 'EXISTS (%s)' % sql, params

    def explain_query(self):
        result = list(self.execute_sql())
        # Some backends return 1 item tuples with strings, and others return
        # tuples with integers and strings. Flatten them out into strings.
        for row in result[0]:
            if not isinstance(row, str):
                yield ' '.join(str(c) for c in row)
                yield row

class SQLInsertCompiler(SQLCompiler):
    returning_fields = None
    returning_params = tuple()

    def field_as_sql(self, field, val):
        Take a field and a value intended to be saved on that field, and
        return placeholder SQL and accompanying params. Check for raw values,
        expressions, and fields with get_placeholder() defined in that order.

        When field is None, consider the value raw and use it as the
        placeholder, with no corresponding parameters returned.
        if field is None:
            # A field value of None means the value is raw.
            sql, params = val, []
        elif hasattr(val, 'as_sql'):
            # This is an expression, let's compile it.
            sql, params = self.compile(val)
        elif hasattr(field, 'get_placeholder'):
            # Some fields (e.g. geo fields) need special munging before
            # they can be inserted.
            sql, params = field.get_placeholder(val, self, self.connection), [val]
            # Return the common case for the placeholder
            sql, params = '%s', [val]

        # The following hook is only used by Oracle Spatial, which sometimes
        # needs to yield 'NULL' and [] as its placeholder and params instead
        # of '%s' and [None]. The 'NULL' placeholder is produced earlier by
        # OracleOperations.get_geom_placeholder(). The following line removes
        # the corresponding None parameter. See ticket #10888.
        params = self.connection.ops.modify_insert_params(sql, params)

        return sql, params

    def prepare_value(self, field, value):
        Prepare a value to be used in a query by resolving it if it is an
        expression and otherwise calling the field's get_db_prep_save().
        if hasattr(value, 'resolve_expression'):
            value = value.resolve_expression(self.query, allow_joins=False, for_save=True)
            # Don't allow values containing Col expressions. They refer to
            # existing columns on a row, but in the case of insert the row
            # doesn't exist yet.
            if value.contains_column_references:
                raise ValueError(
                    'Failed to insert expression "%s" on %s. F() expressions '
                    'can only be used to update, not to insert.' % (value, field)
            if value.contains_aggregate:
                raise FieldError(
                    'Aggregate functions are not allowed in this query '
                    '(%s=%r).' % (, value)
            if value.contains_over_clause:
                raise FieldError(
                    'Window expressions are not allowed in this query (%s=%r).'
                    % (, value)
            value = field.get_db_prep_save(value, connection=self.connection)
        return value

    def pre_save_val(self, field, obj):
        Get the given field's value off the given obj. pre_save() is used for
        things like auto_now on DateTimeField. Skip it if this is a raw query.
        if self.query.raw:
            return getattr(obj, field.attname)
        return field.pre_save(obj, add=True)

    def assemble_as_sql(self, fields, value_rows):
        Take a sequence of N fields and a sequence of M rows of values, and
        generate placeholder SQL and parameters for each field and value.
        Return a pair containing:
         * a sequence of M rows of N SQL placeholder strings, and
         * a sequence of M rows of corresponding parameter values.

        Each placeholder string may contain any number of '%s' interpolation
        strings, and each parameter row will contain exactly as many params
        as the total number of '%s's in the corresponding placeholder row.
        if not value_rows:
            return [], []

        # list of (sql, [params]) tuples for each object to be saved
        # Shape: [n_objs][n_fields][2]
        rows_of_fields_as_sql = (
            (self.field_as_sql(field, v) for field, v in zip(fields, row))
            for row in value_rows

        # tuple like ([sqls], [[params]s]) for each object to be saved
        # Shape: [n_objs][2][n_fields]
        sql_and_param_pair_rows = (zip(*row) for row in rows_of_fields_as_sql)

        # Extract separate lists for placeholders and params.
        # Each of these has shape [n_objs][n_fields]
        placeholder_rows, param_rows = zip(*sql_and_param_pair_rows)

        # Params for each field are still lists, and need to be flattened.
        param_rows = [[p for ps in row for p in ps] for row in param_rows]

        return placeholder_rows, param_rows

    def as_sql(self):
        # We don't need quote_name_unless_alias() here, since these are all
        # going to be column names (so we can avoid the extra overhead).
        qn = self.connection.ops.quote_name
        opts = self.query.get_meta()
        insert_statement = self.connection.ops.insert_statement(ignore_conflicts=self.query.ignore_conflicts)
        result = ['%s %s' % (insert_statement, qn(opts.db_table))]
        fields = self.query.fields or []
        result.append('(%s)' % ', '.join(qn(f.column) for f in fields))

        if self.query.fields:
            value_rows = [
                [self.prepare_value(field, self.pre_save_val(field, obj)) for field in fields]
                for obj in self.query.objs
            # An empty object.
            value_rows = [[self.connection.ops.pk_default_value()] for _ in self.query.objs]
            fields = [None]

        # Currently the backends just accept values when generating bulk
        # queries and generate their own placeholders. Doing that isn't
        # necessary and it should be possible to use placeholders and
        # expressions in bulk inserts too.
        can_bulk = (not self.returning_fields and self.connection.features.has_bulk_insert)

        placeholder_rows, param_rows = self.assemble_as_sql(fields, value_rows)

        ignore_conflicts_suffix_sql = self.connection.ops.ignore_conflicts_suffix_sql(
        if self.returning_fields and self.connection.features.can_return_columns_from_insert:
            if self.connection.features.can_return_rows_from_bulk_insert:
                result.append(self.connection.ops.bulk_insert_sql(fields, placeholder_rows))
                params = param_rows
                result.append("VALUES (%s)" % ", ".join(placeholder_rows[0]))
                params = [param_rows[0]]
            if ignore_conflicts_suffix_sql:
            # Skip empty r_sql to allow subclasses to customize behavior for
            # 3rd party backends. Refs #19096.
            r_sql, self.returning_params = self.connection.ops.return_insert_columns(self.returning_fields)
            if r_sql:
                params += [self.returning_params]
            return [(" ".join(result), tuple(chain.from_iterable(params)))]

        if can_bulk:
            result.append(self.connection.ops.bulk_insert_sql(fields, placeholder_rows))
            if ignore_conflicts_suffix_sql:
            return [(" ".join(result), tuple(p for ps in param_rows for p in ps))]
            if ignore_conflicts_suffix_sql:
            return [
                (" ".join(result + ["VALUES (%s)" % ", ".join(p)]), vals)
                for p, vals in zip(placeholder_rows, param_rows)

    def execute_sql(self, returning_fields=None):
        assert not (
            returning_fields and len(self.query.objs) != 1 and
            not self.connection.features.can_return_rows_from_bulk_insert
        self.returning_fields = returning_fields
        with self.connection.cursor() as cursor:
            for sql, params in self.as_sql():
                cursor.execute(sql, params)
            if not self.returning_fields:
                return []
            if self.connection.features.can_return_rows_from_bulk_insert and len(self.query.objs) > 1:
                return self.connection.ops.fetch_returned_insert_rows(cursor)
            if self.connection.features.can_return_columns_from_insert:
                assert len(self.query.objs) == 1
                return [self.connection.ops.fetch_returned_insert_columns(cursor, self.returning_params)]
            return [(self.connection.ops.last_insert_id(
                cursor, self.query.get_meta().db_table, self.query.get_meta().pk.column

class SQLDeleteCompiler(SQLCompiler):
    def single_alias(self):
        # Ensure base table is in aliases.
        return sum(self.query.alias_refcount[t] > 0 for t in self.query.alias_map) == 1

    def _as_sql(self, query):
        result = [
            'DELETE FROM %s' % self.quote_name_unless_alias(query.base_table)
        where, params = self.compile(query.where)
        if where:
            result.append('WHERE %s' % where)
        return ' '.join(result), tuple(params)

    def as_sql(self):
        Create the SQL for this query. Return the SQL string and list of
        if self.single_alias:
            return self._as_sql(self.query)
        innerq = self.query.clone()
        innerq.__class__ = Query
        pk = = [
        outerq = Query(self.query.model)
        outerq.where = self.query.where_class()
        return self._as_sql(outerq)

class SQLUpdateCompiler(SQLCompiler):
    def as_sql(self):
        Create the SQL for this query. Return the SQL string and list of
        if not self.query.values:
            return '', ()
        qn = self.quote_name_unless_alias
        values, update_params = [], []
        for field, model, val in self.query.values:
            if hasattr(val, 'resolve_expression'):
                val = val.resolve_expression(self.query, allow_joins=False, for_save=True)
                if val.contains_aggregate:
                    raise FieldError(
                        'Aggregate functions are not allowed in this query '
                        '(%s=%r).' % (, val)
                if val.contains_over_clause:
                    raise FieldError(
                        'Window expressions are not allowed in this query '
                        '(%s=%r).' % (, val)
            elif hasattr(val, 'prepare_database_save'):
                if field.remote_field:
                    val = field.get_db_prep_save(
                    raise TypeError(
                        "Tried to update field %s with a model instance, %r. "
                        "Use a value compatible with %s."
                        % (field, val, field.__class__.__name__)
                val = field.get_db_prep_save(val, connection=self.connection)

            # Getting the placeholder for the field.
            if hasattr(field, 'get_placeholder'):
                placeholder = field.get_placeholder(val, self, self.connection)
                placeholder = '%s'
            name = field.column
            if hasattr(val, 'as_sql'):
                sql, params = self.compile(val)
                values.append('%s = %s' % (qn(name), placeholder % sql))
            elif val is not None:
                values.append('%s = %s' % (qn(name), placeholder))
                values.append('%s = NULL' % qn(name))
        table = self.query.base_table
        result = [
            'UPDATE %s SET' % qn(table),
            ', '.join(values),
        where, params = self.compile(self.query.where)
        if where:
            result.append('WHERE %s' % where)
        return ' '.join(result), tuple(update_params + params)

    def execute_sql(self, result_type):
        Execute the specified update. Return the number of rows affected by
        the primary update query. The "primary update query" is the first
        non-empty query that is executed. Row counts for any subsequent,
        related queries are not available.
        cursor = super().execute_sql(result_type)
            rows = cursor.rowcount if cursor else 0
            is_empty = cursor is None
            if cursor:
        for query in self.query.get_related_updates():
            aux_rows = query.get_compiler(self.using).execute_sql(result_type)
            if is_empty and aux_rows:
                rows = aux_rows
                is_empty = False
        return rows

    def pre_sql_setup(self):
        If the update depends on results from other tables, munge the "where"
        conditions to match the format required for (portable) SQL updates.

        If multiple updates are required, pull out the id values to update at
        this point so that they don't change as a result of the progressive
        refcounts_before = self.query.alias_refcount.copy()
        # Ensure base table is in the query
        count = self.query.count_active_tables()
        if not self.query.related_updates and count == 1:
        query = self.query.chain(klass=Query)
        query.select_related = False
        query.extra = {} = []

        must_pre_select = count > 1 and not self.connection.features.update_can_self_select

        # Now we adjust the current query: reset the where clause and get rid
        # of all the tables we don't need (since they're in the sub-select).
        self.query.where = self.query.where_class()
        if self.query.related_updates or must_pre_select:
            # Either we're using the idents in multiple update queries (so
            # don't want them to change), or the db backend doesn't support
            # selecting from the updating table (e.g. MySQL).
            idents = []
            for rows in query.get_compiler(self.using).execute_sql(MULTI):
                idents.extend(r[0] for r in rows)
            self.query.add_filter(('pk__in', idents))
            self.query.related_ids = idents
            # The fast path. Filters and updates in one query.
            self.query.add_filter(('pk__in', query))

class SQLAggregateCompiler(SQLCompiler):
    def as_sql(self):
        Create the SQL for this query. Return the SQL string and list of
        sql, params = [], []
        for annotation in self.query.annotation_select.values():
            ann_sql, ann_params = self.compile(annotation)
            ann_sql, ann_params = annotation.select_format(self, ann_sql, ann_params)
        self.col_count = len(self.query.annotation_select)
        sql = ', '.join(sql)
        params = tuple(params)

        sql = 'SELECT %s FROM (%s) subquery' % (sql, self.query.subquery)
        params = params + self.query.sub_params
        return sql, params

def cursor_iter(cursor, sentinel, col_count, itersize):
    Yield blocks of rows from a cursor and ensure the cursor is closed when
        for rows in iter((lambda: cursor.fetchmany(itersize)), sentinel):
            yield rows if col_count is None else [r[:col_count] for r in rows]