server/sources/rql2sql.py
author Sylvain Thénault <sylvain.thenault@logilab.fr>
Mon, 06 Jun 2016 21:19:04 +0200
changeset 11293 63c589e6b076
parent 10677 59ec0aaae08b
permissions -rw-r--r--
[rql2sql] Test and fix some case of sql generation when an unstable variable (ie whose type change from one solution to another) is used in an EXISTS node but belongs to an outer scope. Closes #13518646.

# copyright 2003-2013 LOGILAB S.A. (Paris, FRANCE), all rights reserved.
# contact http://www.logilab.fr/ -- mailto:contact@logilab.fr
#
# This file is part of CubicWeb.
#
# CubicWeb is free software: you can redistribute it and/or modify it under the
# terms of the GNU Lesser General Public License as published by the Free
# Software Foundation, either version 2.1 of the License, or (at your option)
# any later version.
#
# CubicWeb is distributed in the hope that it will be useful, but WITHOUT
# ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
# FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for more
# details.
#
# You should have received a copy of the GNU Lesser General Public License along
# with CubicWeb.  If not, see <http://www.gnu.org/licenses/>.
"""RQL to SQL generator for native sources.


SQL queries optimization
~~~~~~~~~~~~~~~~~~~~~~~~
1. CWUser X WHERE X in_group G, G name 'users':

   CWUser is the only subject entity type for the in_group relation,
   which allow us to do ::

     SELECT eid_from FROM in_group, CWGroup
     WHERE in_group.eid_to = CWGroup.eid_from
     AND CWGroup.name = 'users'


2. Any X WHERE X nonfinal1 Y, Y nonfinal2 Z

   -> direct join between nonfinal1 and nonfinal2, whatever X,Y, Z (unless
      inlined...)

      NOT IMPLEMENTED (and quite hard to implement)

Potential optimization information is collected by the querier, sql generation
is done according to this information

cross RDMS note : read `Comparison of different SQL implementations`_
by Troels Arvin. Features SQL ISO Standard, PG, mysql, Oracle, MS SQL, DB2
and Informix.

.. _Comparison of different SQL implementations: http://www.troels.arvin.dk/db/rdbms
"""

__docformat__ = "restructuredtext en"

import threading

from six import PY2
from six.moves import range

from logilab.database import FunctionDescr, SQL_FUNCTIONS_REGISTRY

from rql import BadRQLQuery, CoercionError
from rql.utils import common_parent
from rql.stmts import Union, Select
from rql.nodes import (VariableRef, Constant, Function, Variable, Or,
                       Not, Comparison, ColumnAlias, Relation, SubQuery)

from cubicweb import QueryError
from cubicweb.rqlrewrite import cleanup_solutions
from cubicweb.server.sqlutils import SQL_PREFIX

ColumnAlias._q_invariant = False # avoid to check for ColumnAlias / Variable

FunctionDescr.source_execute = None

def default_update_cb_stack(self, stack):
    stack.append(self.source_execute)
FunctionDescr.update_cb_stack = default_update_cb_stack

get_func_descr = SQL_FUNCTIONS_REGISTRY.get_function

LENGTH = get_func_descr('LENGTH')
def length_source_execute(source, session, value):
    return len(value.getvalue())
LENGTH.source_execute = length_source_execute

def _new_var(select, varname):
    newvar = select.get_variable(varname)
    if not 'relations' in newvar.stinfo:
        # not yet initialized
        newvar.prepare_annotation()
        newvar.stinfo['scope'] = select
        newvar._q_invariant = False
        select.selection.append(VariableRef(newvar))
    return newvar

def _fill_to_wrap_rel(var, newselect, towrap, schema):
    for rel in var.stinfo['relations'] - var.stinfo['rhsrelations']:
        rschema = schema.rschema(rel.r_type)
        if rschema.inlined:
            towrap.add( (var, rel) )
            for vref in rel.children[1].iget_nodes(VariableRef):
                newivar = _new_var(newselect, vref.name)
                _fill_to_wrap_rel(vref.variable, newselect, towrap, schema)
        elif rschema.final:
            towrap.add( (var, rel) )
            for vref in rel.children[1].iget_nodes(VariableRef):
                newivar = _new_var(newselect, vref.name)
                newivar.stinfo['attrvar'] = (var, rel.r_type)

def rewrite_unstable_outer_join(select, solutions, unstable, schema):
    """if some optional variables are unstable, they should be selected in a
    subquery. This function check this and rewrite the rql syntax tree if
    necessary (in place). Return a boolean telling if the tree has been modified
    """
    modified = False
    for varname in tuple(unstable):
        var = select.defined_vars[varname]
        if not var.stinfo.get('optrelations'):
            continue
        unstable.remove(varname)
        newselect = Select()
        myunion = Union()
        myunion.append(newselect)
        # extract aliases / selection
        newvar = _new_var(newselect, var.name)
        newselect.selection = [VariableRef(newvar)]
        towrap_rels = set()
        _fill_to_wrap_rel(var, newselect, towrap_rels, schema)
        # extract relations
        for var, rel in towrap_rels:
            newrel = rel.copy(newselect)
            newselect.add_restriction(newrel)
            select.remove_node(rel)
            var.stinfo['relations'].remove(rel)
            newvar.stinfo['relations'].add(newrel)
            if rel.optional in ('left', 'both'):
                newvar.add_optional_relation(newrel)
            for vref in newrel.children[1].iget_nodes(VariableRef):
                var = vref.variable
                var.stinfo['relations'].add(newrel)
                var.stinfo['rhsrelations'].add(newrel)
                if rel.optional in ('right', 'both'):
                    var.add_optional_relation(newrel)
        if not select.where and not modified:
            # oops, generated the same thing as the original select....
            # restore original query, else we'll indefinitly loop
            for var, rel in towrap_rels:
                select.add_restriction(rel)
            continue
        modified = True
        # extract subquery solutions
        mysolutions = [sol.copy() for sol in solutions]
        cleanup_solutions(newselect, mysolutions)
        newselect.set_possible_types(mysolutions)
        # full sub-query
        aliases = [VariableRef(select.get_variable(avar.name, i))
                   for i, avar in enumerate(newselect.selection)]
        select.add_subquery(SubQuery(aliases, myunion), check=False)
    return modified

def _new_solutions(rqlst, solutions):
    """first filter out subqueries variables from solutions"""
    newsolutions = []
    for origsol in solutions:
        asol = {}
        for vname in rqlst.defined_vars:
            asol[vname] = origsol[vname]
        if not asol in newsolutions:
            newsolutions.append(asol)
    return newsolutions

def remove_unused_solutions(rqlst, solutions, varmap, schema):
    """cleanup solutions: remove solutions where invariant variables are taking
    different types
    """
    newsols = _new_solutions(rqlst, solutions)
    existssols = {}
    unstable = set()
    invariants = {}
    for vname, var in rqlst.defined_vars.items():
        vtype = newsols[0][vname]
        if var._q_invariant or vname in varmap:
            # remove invariant variable from solutions to remove duplicates
            # later, then reinserting a type for the variable even later
            for sol in newsols:
                invariants.setdefault(id(sol), {})[vname] = sol.pop(vname)
        elif var.scope is not rqlst:
            # move apart variables which are in a EXISTS scope and are variating
            try:
                thisexistssols, thisexistsvars = existssols[var.scope]
            except KeyError:
                # copy to avoid shared dict in newsols and exists sols
                thisexistssols = [newsols[0].copy()]
                thisexistsvars = set()
                existssols[var.scope] = thisexistssols, thisexistsvars
            for i in range(len(newsols)-1, 0, -1):
                if vtype != newsols[i][vname]:
                    thisexistssols.append(newsols.pop(i))
                    thisexistsvars.add(vname)
        else:
            # remember unstable variables
            for i in range(1, len(newsols)):
                if vtype != newsols[i][vname]:
                    unstable.add(vname)
    # remove unstable variables from exists solutions: the possible types of these variables are not
    # properly represented in exists solutions, so we have to remove and reinject them later
    # according to the outer solution (see `iter_exists_sols`)
    for sols, _ in existssols.values():
        for vname in unstable:
            for sol in sols:
                sol.pop(vname, None)
    if invariants:
        # filter out duplicates
        newsols_ = []
        for sol in newsols:
            if not sol in newsols_:
                newsols_.append(sol)
        newsols = newsols_
        # reinsert solutions for invariants
        for sol in newsols:
            for invvar, vartype in invariants[id(sol)].items():
                sol[invvar] = vartype
        for sol in existssols:
            try:
                for invvar, vartype in invariants[id(sol)].items():
                    sol[invvar] = vartype
            except KeyError:
                continue
    if len(newsols) > 1:
        if rewrite_unstable_outer_join(rqlst, newsols, unstable, schema):
            # remove variables extracted to subqueries from solutions
            newsols = _new_solutions(rqlst, newsols)
    return newsols, existssols, unstable

def relation_info(relation):
    lhs, rhs = relation.get_variable_parts()
    try:
        lhs = lhs.variable
        lhsconst = lhs.stinfo['constnode']
    except AttributeError:
        lhsconst = lhs
        lhs = None
    except KeyError:
        lhsconst = None # ColumnAlias
    try:
        rhs = rhs.variable
        rhsconst = rhs.stinfo['constnode']
    except AttributeError:
        rhsconst = rhs
        rhs = None
    except KeyError:
        rhsconst = None # ColumnAlias
    return lhs, lhsconst, rhs, rhsconst

def sort_term_selection(sorts, rqlst, groups):
    # XXX beurk
    if isinstance(rqlst, list):
        def append(term):
            rqlst.append(term)
        selectionidx = set(str(term) for term in rqlst)
    else:
        def append(term):
            rqlst.selection.append(term.copy(rqlst))
        selectionidx = set(str(term) for term in rqlst.selection)

    for sortterm in sorts:
        term = sortterm.term
        if not isinstance(term, Constant) and not str(term) in selectionidx:
            selectionidx.add(str(term))
            append(term)
            if groups:
                for vref in term.iget_nodes(VariableRef):
                    if not any(vref.is_equivalent(g) for g in groups):
                        groups.append(vref)

def fix_selection_and_group(rqlst, needwrap, selectsortterms,
                            sorts, groups, having):
    if selectsortterms and sorts:
        sort_term_selection(sorts, rqlst, not needwrap and groups)
    groupvrefs = [vref for term in groups for vref in term.iget_nodes(VariableRef)]
    if sorts and groups:
        # when a query is grouped, ensure sort terms are grouped as well
        for sortterm in sorts:
            term = sortterm.term
            if not (isinstance(term, Constant) or \
                    (isinstance(term, Function) and
                     get_func_descr(term.name).aggregat)):
                for vref in term.iget_nodes(VariableRef):
                    if not any(vref.is_equivalent(group) for group in groupvrefs):
                        groups.append(vref)
                        groupvrefs.append(vref)
    if needwrap and (groups or having):
        selectedidx = set(vref.name for term in rqlst.selection
                          for vref in term.get_nodes(VariableRef))
        if groups:
            for vref in groupvrefs:
                if vref.name not in selectedidx:
                    selectedidx.add(vref.name)
                    rqlst.selection.append(vref)
        if having:
            for term in having:
                for vref in term.iget_nodes(VariableRef):
                    if vref.name not in selectedidx:
                        selectedidx.add(vref.name)
                        rqlst.selection.append(vref)

def iter_mapped_var_sels(stmt, variable):
    # variable is a Variable or ColumnAlias node mapped to a source side
    # callback
    if not (len(variable.stinfo['rhsrelations']) <= 1 and # < 1 on column alias
            variable.stinfo['selected']):
        raise QueryError("can't use %s as a restriction variable"
                         % variable.name)
    for selectidx in variable.stinfo['selected']:
        vrefs = stmt.selection[selectidx].get_nodes(VariableRef)
        if len(vrefs) != 1:
            raise QueryError()
        yield selectidx, vrefs[0]

def update_source_cb_stack(state, stmt, node, stack):
    while True:
        node = node.parent
        if node is stmt:
            break
        if not isinstance(node, Function):
            raise QueryError()
        funcd = get_func_descr(node.name)
        if funcd.source_execute is None:
            raise QueryError('%s can not be called on mapped attribute'
                             % node.name)
        state.source_cb_funcs.add(node)
        funcd.update_cb_stack(stack)


# IGenerator implementation for RQL->SQL #######################################

class StateInfo(object):
    """this class stores data accumulated during the RQL syntax tree visit
    for later SQL generation.

    Attributes related to OUTER JOIN handling
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    * `outer_chains`, list of list of strings. Each list represent a tables
      that have to be outer joined together.

    * `outer_tables`, dictionary used as index of tables used in outer join ::

        'table alias': (outertype, [conditions], [chain])

      where:

      * `outertype` is one of None, 'LEFT', 'RIGHT', 'FULL'
      * `conditions` is a list of join conditions (string)
      * `chain` is a list of table alias (the *outer chain*) in which the key
        alias appears

    * `outer_pending` is a dictionary containing some conditions that will have
      to be added to the outer join when the table will be turned into an
      outerjoin ::

       'table alias': [conditions]
    """
    def __init__(self, select, existssols, unstablevars):
        self.existssols = existssols
        self.unstablevars = unstablevars
        self.subtables = {}
        self.needs_source_cb = None
        self.subquery_source_cb = None
        self.source_cb_funcs = set()
        self.scopes = {select: 0}
        self.scope_nodes = []

    def reset(self, solution):
        """reset some visit variables"""
        self.solution = solution
        self.count = 0
        self.done = set()
        self.tables = self.subtables.copy()
        self.actual_tables = [[]]
        for _, tsql in self.tables.values():
            self.actual_tables[-1].append(tsql)
        self.outer_chains = []
        self.outer_tables = {}
        self.outer_pending = {}
        self.duplicate_switches = []
        self.aliases = {}
        self.restrictions = []
        self._restr_stack = []
        self.ignore_varmap = False
        self._needs_source_cb = {}

    def merge_source_cbs(self, needs_source_cb):
        if self.needs_source_cb is None:
            self.needs_source_cb = needs_source_cb
        elif needs_source_cb != self.needs_source_cb:
            raise QueryError('query fetch some source mapped attribute, some not')

    def finalize_source_cbs(self):
        if self.subquery_source_cb is not None:
            self.needs_source_cb.update(self.subquery_source_cb)

    def add_restriction(self, restr):
        if restr:
            self.restrictions.append(restr)

    def iter_exists_sols(self, exists):
        if not exists in self.existssols:
            yield 1
            return
        thisexistssols, thisexistsvars = self.existssols[exists]
        notdone_outside_vars = set()
        # when iterating other solutions inner to an EXISTS subquery, we should
        # reset variables which have this EXISTS node as scope at each iteration
        for var in exists.stmt.defined_vars.values():
            if var.scope is exists:
                thisexistsvars.add(var.name)
            elif var.name not in self.done:
                notdone_outside_vars.add(var)
        # make a copy of the outer statement's solution for later restore
        origsol = self.solution.copy()
        origtables = self.tables
        done = self.done
        for thisexistssol in thisexistssols:
            for vname in self.unstablevars:
                # check first if variable belong to the EXISTS's scope, else it may be missing from
                # `thisexistssol`
                if vname in thisexistsvars and thisexistssol[vname] != origsol[vname]:
                    break
            else:
                self.tables = origtables.copy()
                # overwrite current outer solution by EXISTS solution (the later will be missing
                # unstable outer variables)
                self.solution.update(thisexistssol)
                yield 1
                # cleanup self.done from stuff specific to EXISTS, so they will be reconsidered in
                # the next round
                for var in thisexistsvars:
                    if var in done:
                        done.remove(var)
                for var in list(notdone_outside_vars):
                    if var.name in done and var._q_sqltable in self.tables:
                        origtables[var._q_sqltable] = self.tables[var._q_sqltable]
                        notdone_outside_vars.remove(var)
                for rel in exists.iget_nodes(Relation):
                    if rel in done:
                        done.remove(rel)
        # restore original solution
        self.solution = origsol
        self.tables = origtables

    def push_scope(self, scope_node):
        self.scope_nodes.append(scope_node)
        self.scopes[scope_node] = len(self.actual_tables)
        self.actual_tables.append([])
        self._restr_stack.append(self.restrictions)
        self.restrictions = []

    def pop_scope(self):
        del self.scopes[self.scope_nodes[-1]]
        self.scope_nodes.pop()
        restrictions = self.restrictions
        self.restrictions = self._restr_stack.pop()
        scope = len(self.actual_tables) - 1
        # check if we have some outer chain for this scope
        matching_chains = []
        for chain in self.outer_chains:
            for tablealias in chain:
                if self.tables[tablealias][0] < scope:
                    # chain belongs to outer scope
                    break
            else:
                # chain match current scope
                matching_chains.append(chain)
        # call to `tables_sql` will pop actual_tables
        tables = self.tables_sql(matching_chains)
        # cleanup outer join related structure for tables in matching chains
        for chain in matching_chains:
            self.outer_chains.remove(chain)
            for alias in chain:
                del self.outer_tables[alias]
        return restrictions, tables

    # tables handling #########################################################

    def add_table(self, table, key=None, scope=-1):
        if key is None:
            key = table
        if key in self.tables:
            return
        if scope < 0:
            scope = len(self.actual_tables) + scope
        self.tables[key] = (scope, table)
        self.actual_tables[scope].append(table)

    def alias_and_add_table(self, tablename, scope=-1):
        alias = '%s%s' % (tablename, self.count)
        self.count += 1
        self.add_table('%s AS %s' % (tablename, alias), alias, scope)
        return alias

    def relation_table(self, relation):
        """return the table alias used by the given relation"""
        if relation in self.done:
            return relation._q_sqltable
        rid = 'rel_%s%s' % (relation.r_type, self.count)
        # relation's table is belonging to the root scope if it is the principal
        # table of one of its variable and that variable belong's to parent
        # scope
        for varref in relation.iget_nodes(VariableRef):
            var = varref.variable
            # XXX may have a principal without being invariant for this generation,
            #     not sure this is a pb or not
            if var.stinfo.get('principal') is relation and var.scope is var.stmt:
                scope = 0
                break
        else:
            scope = -1
        self.count += 1
        self.add_table('%s_relation AS %s' % (relation.r_type, rid), rid, scope=scope)
        relation._q_sqltable = rid
        self.done.add(relation)
        return rid

    def fti_table(self, relation, fti_table):
        """return the table alias used by the given has_text relation,
        `fti_table` being the table name for the plain text index
        """
        if relation in self.done:
            try:
                return relation._q_sqltable
            except AttributeError:
                pass
        self.done.add(relation)
        scope = self.scopes[relation.scope]
        alias = self.alias_and_add_table(fti_table, scope=scope)
        relation._q_sqltable = alias
        return alias

    # outer join handling ######################################################

    def mark_as_used_in_outer_join(self, tablealias, addpending=True):
        """Mark table of given alias as used in outer join. This must be called
        after `outer_tables[tablealias]` has been initialized.
        """
        # remove a table from actual_table because it's used in an outer join
        # chain
        scope, tabledef = self.tables[tablealias]
        self.actual_tables[scope].remove(tabledef)
        # check if there are some pending outer join condition for this table
        if addpending:
            try:
                pending_conditions = self.outer_pending.pop(tablealias)
            except KeyError:
                pass
            else:
                self.outer_tables[tablealias][1].extend(pending_conditions)
        else:
            assert not tablealias in self.outer_pending

    def add_outer_join_condition(self, tablealias, condition):
        try:
            outer, conditions, chain = self.outer_tables[tablealias]
            conditions.append(condition)
        except KeyError:
            self.outer_pending.setdefault(tablealias, []).append(condition)

    def replace_tables_by_outer_join(self, leftalias, rightalias,
                                     outertype, condition):
        """tell we need <leftalias> <outertype> JOIN <rightalias> ON <condition>
        """
        assert leftalias != rightalias, leftalias
        outer_tables = self.outer_tables
        louter, lconditions, lchain = outer_tables.get(leftalias,
                                                      (None, None, None))
        router, rconditions, rchain = outer_tables.get(rightalias,
                                                      (None, None, None))
        if lchain is None and rchain is None:
            # create a new outer chaine
            chain = [leftalias, rightalias]
            outer_tables[leftalias] = (None, [], chain)
            outer_tables[rightalias] = (outertype, [condition], chain)
            self.outer_chains.append(chain)
            self.mark_as_used_in_outer_join(leftalias, addpending=False)
            self.mark_as_used_in_outer_join(rightalias)
        elif lchain is None:
            # [A > B > C] + [D > A] -> [D > A > B > C]
            if rightalias == rchain[0]:
                outer_tables[leftalias] = (None, [], rchain)
                conditions = outer_tables[rightalias][1] + [condition]
                outer_tables[rightalias] = (outertype, conditions, rchain)
                rchain.insert(0, leftalias)
            else:
                # [A > B > C] + [D > B] -> [A > B > C < D]
                if outertype == 'LEFT':
                    outertype = 'RIGHT'
                outer_tables[leftalias] = (outertype, [condition], rchain)
                rchain.append(leftalias)
            self.mark_as_used_in_outer_join(leftalias)
        elif rchain is None:
            # [A > B > C] + [B > D] -> [A > B > C > D]
            outer_tables[rightalias] = (outertype, [condition], lchain)
            lchain.append(rightalias)
            self.mark_as_used_in_outer_join(rightalias)
        elif lchain is rchain:
            # already in the same chain, simply check compatibility and append
            # the condition if it's ok
            lidx = lchain.index(leftalias)
            ridx = lchain.index(rightalias)
            if (outertype == 'FULL' and router != 'FULL') \
                   or (lidx < ridx and router != 'LEFT') \
                   or (ridx < lidx and louter != 'RIGHT'):
                raise BadRQLQuery()
            # merge conditions
            if lidx < ridx:
                rconditions.append(condition)
            else:
                lconditions.append(condition)
        elif louter is None:
            # merge chains
            self.outer_chains.remove(lchain)
            rchain += lchain
            self.mark_as_used_in_outer_join(leftalias)
            for alias, (aouter, aconditions, achain) in outer_tables.items():
                if achain is lchain:
                    outer_tables[alias] = (aouter, aconditions, rchain)
        else:
            raise BadRQLQuery()

    # sql generation helpers ###################################################

    def tables_sql(self, outer_chains=None):
        """generate SQL for FROM clause"""
        # sort for test predictability
        tables = sorted(self.actual_tables.pop())
        # process outer joins
        if outer_chains is None:
            assert not self.actual_tables, self.actual_tables
            assert not self.outer_pending
            outer_chains = self.outer_chains
        for chain in sorted(outer_chains):
            tablealias = chain[0]
            outertype, conditions, _ = self.outer_tables[tablealias]
            assert _ is chain, (chain, _)
            assert outertype is None, (chain, self.outer_chains)
            assert not conditions, (chain, self.outer_chains)
            assert len(chain) > 1
            tabledef = self.tables[tablealias][1]
            outerjoin = [tabledef]
            for tablealias in chain[1:]:
                outertype, conditions, _ = self.outer_tables[tablealias]
                assert _ is chain, (chain, self.outer_chains)
                assert outertype in ('LEFT', 'RIGHT', 'FULL'), (
                    tablealias, outertype, conditions)
                assert isinstance(conditions, (list)), (
                    tablealias, outertype, conditions)
                tabledef = self.tables[tablealias][1]
                outerjoin.append('%s OUTER JOIN %s ON (%s)' % (
                    outertype, tabledef, ' AND '.join(conditions)))
            tables.append(' '.join(outerjoin))
        return ', '.join(tables)


def extract_fake_having_terms(having):
    """RQL's HAVING may be used to contains stuff that should go in the WHERE
    clause of the SQL query, due to RQL grammar limitation. Split them...

    Return a list nodes that can be ANDed with query's WHERE clause. Having
    subtrees updated in place.
    """
    fakehaving = []
    for subtree in having:
        ors, tocheck = set(), []
        for compnode in subtree.get_nodes(Comparison):
            for fnode in compnode.get_nodes(Function):
                if fnode.descr().aggregat:
                    p = compnode.parent
                    oor = None
                    while not isinstance(p, Select):
                        if isinstance(p, (Or, Not)):
                            oor = p
                        p = p.parent
                    if oor is not None:
                        ors.add(oor)
                    break
            else:
                tocheck.append(compnode)
        # tocheck hold a set of comparison not implying an aggregat function
        # put them in fakehaving if they don't share an Or node as ancestor
        # with another comparison containing an aggregat function
        for compnode in tocheck:
            parents = set()
            p = compnode.parent
            oor = None
            while not isinstance(p, Select):
                if p in ors or p is None: # p is None for nodes already in fakehaving
                    break
                if isinstance(p, (Or, Not)):
                    oor = p
                p = p.parent
            else:
                node = oor or compnode
                fakehaving.append(node)
                node.parent.remove(node)
    return fakehaving


class SQLGenerator(object):
    """
    generation of SQL from the fully expanded RQL syntax tree
    SQL is designed to be used with a CubicWeb SQL schema

    Groups and sort are not handled here since they should not be handled at
    this level (see cubicweb.server.querier)

    we should not have errors here!

    WARNING: a CubicWebSQLGenerator instance is not thread safe, but generate is
    protected by a lock
    """

    def __init__(self, schema, dbhelper, attrmap=None):
        self.schema = schema
        self.dbhelper = dbhelper
        self.dbencoding = dbhelper.dbencoding
        self.keyword_map = {'NOW' : self.dbhelper.sql_current_timestamp,
                            'TODAY': self.dbhelper.sql_current_date,
                            }
        if not self.dbhelper.union_parentheses_support:
            self.union_sql = self.noparen_union_sql
        self._lock = threading.Lock()
        if attrmap is None:
            attrmap = {}
        self.attr_map = attrmap

    def generate(self, union, args=None, varmap=None):
        """return SQL queries and a variable dictionary from a RQL syntax tree

        :partrqls: a list of couple (rqlst, solutions)
        :args: optional dictionary with values of substitutions used in the query
        :varmap: optional dictionary mapping variable name to a special table
          name, in case the query as to fetch data from temporary tables

        return an sql string and a dictionary with substitutions values
        """
        if args is None:
            args = {}
        if varmap is None:
            varmap =  {}
        self._lock.acquire()
        self._args = args
        self._varmap = varmap
        self._query_attrs = {}
        self._state = None
        # self._not_scope_offset = 0
        try:
            # union query for each rqlst / solution
            sql = self.union_sql(union)
            # we are done
            return sql, self._query_attrs, self._state.needs_source_cb
        finally:
            self._lock.release()

    def union_sql(self, union, needalias=False): # pylint: disable=E0202
        if len(union.children) == 1:
            return self.select_sql(union.children[0], needalias)
        sqls = ('(%s)' % self.select_sql(select, needalias)
                for select in union.children)
        return '\nUNION ALL\n'.join(sqls)

    def noparen_union_sql(self, union, needalias=False):
        # needed for sqlite backend which doesn't like parentheses around union
        # query. This may cause bug in some condition (sort in one of the
        # subquery) but will work in most case
        #
        # see http://www.sqlite.org/cvstrac/tktview?tn=3074
        sqls = (self.select_sql(select, needalias)
                for i, select in enumerate(union.children))
        return '\nUNION ALL\n'.join(sqls)

    def select_sql(self, select, needalias=False):
        """return SQL queries and a variable dictionary from a RQL syntax tree

        :select: a selection statement of the syntax tree (`rql.stmts.Select`)
        :solution: a dictionary containing variables binding.
          A solution's dictionary has variable's names as key and variable's
          types as values
        :needwrap: boolean telling if the query will be wrapped in an outer
          query (to deal with aggregat and/or grouping)
        """
        if select.distinct:
            distinct = True
        elif self.dbhelper.fti_need_distinct:
            distinct = getattr(select.parent, 'has_text_query', False)
        else:
            distinct = False
        sorts = select.orderby
        groups = select.groupby
        having = select.having
        for restr in extract_fake_having_terms(having):
            scope = None
            for vref in restr.get_nodes(VariableRef):
                vscope = vref.variable.scope
                if vscope is select:
                    continue # ignore select scope, so restriction is added to
                             # the inner most scope possible
                if scope is None:
                    scope = vscope
                elif vscope is not scope:
                    scope = common_parent(scope, vscope).scope
            if scope is None:
                scope = select
            scope.add_restriction(restr)
        # remember selection, it may be changed and have to be restored
        origselection = select.selection[:]
        # check if the query will have union subquery, if it need sort term
        # selection (union or distinct query) and wrapping (union with groups)
        needwrap = False
        sols = select.solutions
        selectsortterms = distinct
        if len(sols) > 1:
            # remove invariant from solutions
            sols, existssols, unstable = remove_unused_solutions(
                select, sols, self._varmap, self.schema)
            if len(sols) > 1:
                # if there is still more than one solution, a UNION will be
                # generated and so sort terms have to be selected
                selectsortterms = True
                # and if select is using group by or aggregat, a wrapping
                # query will be necessary
                if groups or select.has_aggregat:
                    select.select_only_variables()
                    needwrap = True
        else:
            existssols, unstable = {}, ()
        state = StateInfo(select, existssols, unstable)
        if self._state is not None:
            # state from a previous unioned select
            state.merge_source_cbs(self._state.needs_source_cb)
        # treat subqueries
        self._subqueries_sql(select, state)
        # generate sql for this select node
        if needwrap:
            outerselection = origselection[:]
            if sorts and selectsortterms:
                if distinct:
                    sort_term_selection(sorts, outerselection, groups)
        fix_selection_and_group(select, needwrap, selectsortterms,
                                sorts, groups, having)
        if needwrap:
            fneedwrap = len(outerselection) != len(origselection)
        else:
            fneedwrap = len(select.selection) != len(origselection)
        if fneedwrap:
            needalias = True
        self._in_wrapping_query = False
        self._state = state
        try:
            sql = self._solutions_sql(select, sols, distinct,
                                      needalias or needwrap)
            # generate groups / having before wrapping query selection to get
            # correct column aliases
            self._in_wrapping_query = needwrap
            if groups:
                # no constant should be inserted in GROUP BY else the backend
                # will interpret it as a positional index in the selection
                groups = ','.join(vref.accept(self) for vref in groups
                                  if not isinstance(vref, Constant))
            if having:
                # filter out constants as for GROUP BY
                having = ' AND '.join(term.accept(self) for term in having
                                      if not isinstance(term, Constant))
            if needwrap:
                sql = '%s FROM (%s) AS T1' % (
                    self._selection_sql(outerselection, distinct,needalias),
                    sql)
            if groups:
                sql += '\nGROUP BY %s' % groups
            if having:
                sql += '\nHAVING %s' % having
            # sort
            if sorts:
                sqlsortterms = []
                if needwrap:
                    selectidx = [str(term) for term in outerselection]
                else:
                    selectidx = [str(term) for term in select.selection]
                for sortterm in sorts:
                    _term = self._sortterm_sql(sortterm, selectidx)
                    if _term is not None:
                        sqlsortterms.append(_term)
                if sqlsortterms:
                    sql = self.dbhelper.sql_add_order_by(
                        sql, sqlsortterms, origselection, fneedwrap,
                        select.limit or select.offset)
            else:
                sqlsortterms = None
            state.finalize_source_cbs()
        finally:
            select.selection = origselection
        # limit / offset
        sql = self.dbhelper.sql_add_limit_offset(sql,
                                                 select.limit,
                                                 select.offset,
                                                 sqlsortterms)
        return sql

    def _subqueries_sql(self, select, state):
        for i, subquery in enumerate(select.with_):
            sql = self.union_sql(subquery.query, needalias=True)
            tablealias = '_T%s' % i # XXX nested subqueries
            sql = '(%s) AS %s' % (sql, tablealias)
            state.subtables[tablealias] = (0, sql)
            latest_state = self._state
            for vref in subquery.aliases:
                alias = vref.variable
                alias._q_sqltable = tablealias
                alias._q_sql = '%s.C%s' % (tablealias, alias.colnum)
                try:
                    stack = latest_state.needs_source_cb[alias.colnum]
                    if state.subquery_source_cb is None:
                        state.subquery_source_cb = {}
                    for selectidx, vref in iter_mapped_var_sels(select, alias):
                        stack = stack[:]
                        update_source_cb_stack(state, select, vref, stack)
                        state.subquery_source_cb[selectidx] = stack
                except KeyError:
                    continue

    def _solutions_sql(self, select, solutions, distinct, needalias):
        sqls = []
        for solution in solutions:
            self._state.reset(solution)
            # visit restriction subtree
            if select.where is not None:
                self._state.add_restriction(select.where.accept(self))
            sql = [self._selection_sql(select.selection, distinct, needalias)]
            if self._state.restrictions:
                sql.append('WHERE %s' % ' AND '.join(self._state.restrictions))
            self._state.merge_source_cbs(self._state._needs_source_cb)
            # add required tables
            assert len(self._state.actual_tables) == 1, self._state.actual_tables
            tables = self._state.tables_sql()
            if tables:
                sql.insert(1, 'FROM %s' % tables)
            elif self._state.restrictions and self.dbhelper.needs_from_clause:
                sql.insert(1, 'FROM (SELECT 1) AS _T')
            sqls.append('\n'.join(sql))
        if distinct:
            return '\nUNION\n'.join(sqls)
        else:
            return '\nUNION ALL\n'.join(sqls)

    def _selection_sql(self, selected, distinct, needaliasing=False):
        clause = []
        for term in selected:
            sql = term.accept(self)
            if needaliasing:
                colalias = 'C%s' % len(clause)
                clause.append('%s AS %s' % (sql, colalias))
                if isinstance(term, VariableRef):
                    self._state.aliases[term.name] = colalias
            else:
                clause.append(sql)
        if distinct:
            return 'SELECT DISTINCT %s' % ', '.join(clause)
        return 'SELECT %s' % ', '.join(clause)

    def _sortterm_sql(self, sortterm, selectidx):
        term = sortterm.term
        try:
            sqlterm = selectidx.index(str(term)) + 1
        except ValueError:
            # Constant node or non selected term
            sqlterm = term.accept(self)
            if sqlterm is None:
                return None
        if sortterm.asc:
            return str(sqlterm)
        else:
            return '%s DESC' % sqlterm

    def visit_and(self, et):
        """generate SQL for a AND subtree"""
        res = []
        for c in et.children:
            part = c.accept(self)
            if part:
                res.append(part)
        return ' AND '.join(res)

    def visit_or(self, ou):
        """generate SQL for a OR subtree"""
        res = []
        for c in ou.children:
            part = c.accept(self)
            if part:
                res.append('(%s)' % part)
        if res:
            if len(res) > 1:
                return '(%s)' % ' OR '.join(res)
            return res[0]
        return ''

    def visit_not(self, node):
        csql = node.children[0].accept(self)
        if node in self._state.done or not csql:
            # already processed or no sql generated by children
            return csql
        return 'NOT (%s)' % csql

    def visit_exists(self, exists):
        """generate SQL name for a exists subquery"""
        sqls = []
        for dummy in self._state.iter_exists_sols(exists):
            sql = self._visit_exists(exists)
            if sql:
                sqls.append(sql)
        if not sqls:
            return ''
        return 'EXISTS(%s)' % ' UNION '.join(sqls)

    def _visit_exists(self, exists):
        self._state.push_scope(exists)
        restriction = exists.children[0].accept(self)
        restrictions, tables = self._state.pop_scope()
        if restriction:
            restrictions.append(restriction)
        restriction = ' AND '.join(restrictions)
        if not restriction:
            if tables:
                return 'SELECT 1 FROM %s' % tables
            return ''
        if not tables:
            # XXX could leave surrounding EXISTS() in this case no?
            sql = 'SELECT 1 WHERE %s' % restriction
        else:
            sql = 'SELECT 1 FROM %s WHERE %s' % (tables, restriction)
        return sql


    def visit_relation(self, relation):
        """generate SQL for a relation"""
        rtype = relation.r_type
        # don't care of type constraint statement (i.e. relation_type = 'is')
        if relation.is_types_restriction():
            return ''
        lhs, rhs = relation.get_parts()
        rschema = self.schema.rschema(rtype)
        if rschema.final:
            if rtype == 'eid' and lhs.variable._q_invariant and \
                   lhs.variable.stinfo['constnode']:
                # special case where this restriction is already generated by
                # some other relation
                return ''
            # attribute relation
            if rtype == 'has_text':
                sql = self._visit_has_text_relation(relation)
            else:
                rhs_vars = rhs.get_nodes(VariableRef)
                if rhs_vars:
                    # if variable(s) in the RHS
                    sql = self._visit_var_attr_relation(relation, rhs_vars)
                else:
                    # no variables in the RHS
                    sql = self._visit_attribute_relation(relation)
        elif (rtype == 'is' and isinstance(rhs.children[0], Constant)
              and rhs.children[0].eval(self._args) is None):
            # special case "C is NULL"
            if lhs.name in self._varmap:
                lhssql = self._varmap[lhs.name]
            else:
                lhssql = lhs.accept(self)
            return '%s%s' % (lhssql, rhs.accept(self))
        elif '%s.%s' % (lhs, relation.r_type) in self._varmap:
            # relation has already been processed by a previous step
            return ''
        elif relation.optional:
            # OPTIONAL relation, generate a left|right outer join
            if rtype == 'identity' or rschema.inlined:
                sql = self._visit_outer_join_inlined_relation(relation, rschema)
            else:
                sql = self._visit_outer_join_relation(relation, rschema)
        elif rschema.inlined:
            sql = self._visit_inlined_relation(relation)
        else:
            # regular (non final) relation
            sql = self._visit_relation(relation, rschema)
        return sql

    def _visit_inlined_relation(self, relation):
        lhsvar, _, rhsvar, rhsconst = relation_info(relation)
        # we are sure lhsvar is not None
        lhssql = self._inlined_var_sql(lhsvar, relation.r_type)
        if rhsvar is None:
            moresql = None
        else:
            moresql = self._extra_join_sql(relation, lhssql, rhsvar)
        if isinstance(relation.parent, Not):
            self._state.done.add(relation.parent)
            if rhsvar is not None and rhsvar._q_invariant:
                sql = '%s IS NULL' % lhssql
            else:
                # column != 1234 may not get back rows where column is NULL...
                sql = '(%s IS NULL OR %s!=%s)' % (
                    lhssql, lhssql, (rhsvar or rhsconst).accept(self))
        elif rhsconst is not None:
            sql = '%s=%s' % (lhssql, rhsconst.accept(self))
        elif isinstance(rhsvar, Variable) and rhsvar._q_invariant and \
                 not rhsvar.name in self._varmap:
            # if the rhs variable is only linked to this relation, this mean we
            # only want the relation to exists, eg NOT NULL in case of inlined
            # relation
            if moresql is not None:
                return moresql
            return '%s IS NOT NULL' % lhssql
        else:
            sql = '%s=%s' % (lhssql, rhsvar.accept(self))
        if moresql is None:
            return sql
        return '%s AND %s' % (sql, moresql)

    def _process_relation_term(self, relation, rid, termvar, termconst, relfield):
        if termconst or not termvar._q_invariant:
            termsql = termconst and termconst.accept(self) or termvar.accept(self)
            yield '%s.%s=%s' % (rid, relfield, termsql)
        elif termvar._q_invariant:
            # if the variable is mapped, generate restriction anyway
            if termvar.name in self._varmap:
                termsql = termvar.accept(self)
                yield '%s.%s=%s' % (rid, relfield, termsql)
            extrajoin = self._extra_join_sql(relation, '%s.%s' % (rid, relfield), termvar)
            if extrajoin is not None:
                yield extrajoin

    def _visit_relation(self, relation, rschema):
        """generate SQL for a relation

        implements optimization 1.
        """
        if relation.r_type == 'identity':
            # special case "X identity Y"
            lhs, rhs = relation.get_parts()
            return '%s%s' % (lhs.accept(self), rhs.accept(self))
        lhsvar, lhsconst, rhsvar, rhsconst = relation_info(relation)
        rid = self._state.relation_table(relation)
        sqls = []
        sqls += self._process_relation_term(relation, rid, lhsvar, lhsconst, 'eid_from')
        sqls += self._process_relation_term(relation, rid, rhsvar, rhsconst, 'eid_to')
        sql = ' AND '.join(sqls)
        return sql

    def _visit_outer_join_relation(self, relation, rschema):
        """
        left outer join syntax (optional=='right'):
          X relation Y?

        right outer join syntax (optional=='left'):
          X? relation Y

        full outer join syntaxes (optional=='both'):
          X? relation Y?

        if relation is inlined:
           if it's a left outer join:
           -> X LEFT OUTER JOIN Y ON (X.relation=Y.eid)
           elif it's a right outer join:
           -> Y LEFT OUTER JOIN X ON (X.relation=Y.eid)
           elif it's a full outer join:
           -> X FULL OUTER JOIN Y ON (X.relation=Y.eid)
        else:
           if it's a left outer join:
           -> X LEFT OUTER JOIN relation ON (relation.eid_from=X.eid)
              LEFT OUTER JOIN Y ON (relation.eid_to=Y.eid)
           elif it's a right outer join:
           -> Y LEFT OUTER JOIN relation ON (relation.eid_to=Y.eid)
              LEFT OUTER JOIN X ON (relation.eid_from=X.eid)
           elif it's a full outer join:
           -> X FULL OUTER JOIN Y ON (X.relation=Y.eid)
        """
        leftvar, leftconst, rightvar, rightconst = relation_info(relation)
        assert not (leftconst and rightconst), "doesn't make sense"
        if relation.optional == 'left':
            leftvar, rightvar = rightvar, leftvar
            leftconst, rightconst = rightconst, leftconst
            joinattr, restrattr = 'eid_to', 'eid_from'
        else:
            joinattr, restrattr = 'eid_from', 'eid_to'
        # search table for this variable, to use as left table of the outer join
        leftalias = None
        if leftvar:
            # take care, may return None for invariant variable
            leftalias = self._var_table(leftvar)
        if leftalias is None:
            if leftvar.stinfo['principal'] is not relation:
                # use variable's principal relation
                leftalias = leftvar.stinfo['principal']._q_sqltable
            else:
                # search for relation on which we should join
                for orelation in leftvar.stinfo['relations']:
                    if (orelation is not relation and
                        not self.schema.rschema(orelation.r_type).final):
                        break
                else:
                    for orelation in rightvar.stinfo['relations']:
                        if (orelation is not relation and
                            not self.schema.rschema(orelation.r_type).final
                            and orelation.optional):
                            break
                    else:
                        # unexpected
                        assert False, leftvar
                leftalias = self._state.relation_table(orelation)
        # right table of the outer join
        rightalias = self._state.relation_table(relation)
        # compute join condition
        if not leftconst or (leftvar and not leftvar._q_invariant):
            leftsql = leftvar.accept(self)
        else:
            leftsql = leftconst.accept(self)
        condition = '%s.%s=%s' % (rightalias, joinattr, leftsql)
        if rightconst:
            condition += ' AND %s.%s=%s' % (rightalias, restrattr, rightconst.accept(self))
        # record outer join
        outertype = 'FULL' if relation.optional == 'both' else 'LEFT'
        self._state.replace_tables_by_outer_join(leftalias, rightalias,
                                                 outertype, condition)
        # need another join?
        if rightconst is None:
            # we need another outer join for the other side of the relation (e.g.
            # for "X relation Y?" in RQL, we treated earlier the (cw_X.eid /
            # relation.eid_from) join, now we've to do (relation.eid_to /
            # cw_Y.eid)
            leftalias = rightalias
            rightsql = rightvar.accept(self) # accept before using var_table
            rightalias = self._var_table(rightvar)
            if rightalias is None:
                if rightvar.stinfo['principal'] is not relation:
                    self._state.replace_tables_by_outer_join(
                        leftalias, rightvar.stinfo['principal']._q_sqltable,
                        outertype, '%s.%s=%s' % (leftalias, restrattr, rightvar.accept(self)))
            else:
                self._state.replace_tables_by_outer_join(
                    leftalias, rightalias, outertype,
                    '%s.%s=%s' % (leftalias, restrattr, rightvar.accept(self)))
        # this relation will hence be expressed in FROM clause, return nothing
        # here
        return ''


    def _visit_outer_join_inlined_relation(self, relation, rschema):
        lhsvar, lhsconst, rhsvar, rhsconst = relation_info(relation)
        assert not (lhsconst and rhsconst), "doesn't make sense"
        attr = 'eid' if relation.r_type == 'identity' else relation.r_type
        lhsalias = self._var_table(lhsvar)
        rhsalias = rhsvar and self._var_table(rhsvar)
        try:
            lhssql = self._varmap['%s.%s' % (lhsvar.name, attr)]
        except KeyError:
            if lhsalias is None:
                lhssql = lhsconst.accept(self)
            elif attr == 'eid':
                lhssql = lhsvar.accept(self)
            else:
                lhssql = '%s.%s%s' % (lhsalias, SQL_PREFIX, attr)
        condition = '%s=%s' % (lhssql, (rhsconst or rhsvar).accept(self))
        # this is not a typo, rhs optional variable means lhs outer join and vice-versa
        if relation.optional == 'left':
            lhsvar, rhsvar = rhsvar, lhsvar
            lhsconst, rhsconst = rhsconst, lhsconst
            lhsalias, rhsalias = rhsalias, lhsalias
            outertype = 'LEFT'
        elif relation.optional == 'both':
            outertype = 'FULL'
        else:
            outertype = 'LEFT'
        if rhsalias is None:
            if rhsconst is not None:
                # inlined relation with invariant as rhs
                if relation.r_type != 'identity':
                    condition = '(%s OR %s IS NULL)' % (condition, lhssql)
                if not lhsvar.stinfo.get('optrelations'):
                    return condition
                self._state.add_outer_join_condition(lhsalias, condition)
            return
        if lhsalias is None:
            if lhsconst is not None and not rhsvar.stinfo.get('optrelations'):
                return condition
            lhsalias = lhsvar._q_sql.split('.', 1)[0]
        if lhsalias == rhsalias:
            self._state.add_outer_join_condition(lhsalias, condition)
        else:
            self._state.replace_tables_by_outer_join(
                lhsalias, rhsalias, outertype, condition)
        return ''

    def _visit_var_attr_relation(self, relation, rhs_vars):
        """visit an attribute relation with variable(s) in the RHS

        attribute variables are used either in the selection or for unification
        (eg X attr1 A, Y attr2 A). In case of selection, nothing to do here.
        """
        ored = relation.ored()
        for vref in rhs_vars:
            var = vref.variable
            if var.name in self._varmap:
                # ensure table is added
                self._var_info(var)
            if isinstance(var, ColumnAlias):
                # force sql generation whatever the computed principal
                principal = 1
            else:
                principal = var.stinfo.get('principal')
            # we've to return some sql if:
            # 1. visited relation is ored
            # 2. variable's principal is not this relation and not 1.
            if ored or (principal is not None and principal is not relation
                        and not getattr(principal, 'ored', lambda : 0)()):
                # we have to generate unification expression
                if principal is relation:
                    # take care if ored case and principal is the relation to
                    # use the right relation in the unification term
                    _rel = [rel for rel in var.stinfo['rhsrelations']
                            if not rel is principal][0]
                else:
                    _rel = relation
                lhssql = self._inlined_var_sql(_rel.children[0].variable,
                                               _rel.r_type)
                try:
                    self._state.ignore_varmap = True
                    sql = lhssql + relation.children[1].accept(self)
                finally:
                    self._state.ignore_varmap = False
                if relation.optional == 'right':
                    leftalias = self._var_table(principal.children[0].variable)
                    rightalias = self._var_table(relation.children[0].variable)
                    self._state.replace_tables_by_outer_join(
                        leftalias, rightalias, 'LEFT', sql)
                    return ''
                return sql
        return ''

    def _visit_attribute_relation(self, rel):
        """generate SQL for an attribute relation"""
        lhs, rhs = rel.get_parts()
        rhssql = rhs.accept(self)
        table = self._var_table(lhs.variable)
        if table is None:
            assert rel.r_type == 'eid'
            lhssql = lhs.accept(self)
        else:
            try:
                lhssql = self._varmap['%s.%s' % (lhs.name, rel.r_type)]
            except KeyError:
                mapkey = '%s.%s' % (self._state.solution[lhs.name], rel.r_type)
                if mapkey in self.attr_map:
                    cb, sourcecb = self.attr_map[mapkey]
                    if sourcecb:
                        # callback is a source callback, we can't use this
                        # attribute in restriction
                        raise QueryError("can't use %s (%s) in restriction"
                                         % (mapkey, rel.as_string()))
                    lhssql = cb(self, lhs.variable, rel)
                elif rel.r_type == 'eid':
                    lhssql = lhs.variable._q_sql
                else:
                    lhssql = '%s.%s%s' % (table, SQL_PREFIX, rel.r_type)
        try:
            if rel._q_needcast == 'TODAY':
                sql = 'DATE(%s)%s' % (lhssql, rhssql)
            # XXX which cast function should be used
            #elif rel._q_needcast == 'NOW':
            #    sql = 'TIMESTAMP(%s)%s' % (lhssql, rhssql)
            else:
                sql = '%s%s' % (lhssql, rhssql)
        except AttributeError:
            sql = '%s%s' % (lhssql, rhssql)
        if lhs.variable.stinfo.get('optrelations'):
            self._state.add_outer_join_condition(table, sql)
        else:
            return sql

    def _visit_has_text_relation(self, rel):
        """generate SQL for a has_text relation"""
        lhs, rhs = rel.get_parts()
        const = rhs.children[0]
        alias = self._state.fti_table(rel, self.dbhelper.fti_table)
        jointo = lhs.accept(self)
        restriction = ''
        lhsvar = lhs.variable
        me_is_principal = lhsvar.stinfo.get('principal') is rel
        if me_is_principal:
            if lhsvar.stinfo['typerel'] is None:
                # the variable is using the fti table, no join needed
                jointo = None
            elif not lhsvar.name in self._varmap:
                # join on entities instead of etype's table to get result for
                # external entities on multisources configurations
                ealias = lhsvar._q_sqltable = '_' + lhsvar.name
                jointo = lhsvar._q_sql = '%s.eid' % ealias
                self._state.add_table('entities AS %s' % ealias, ealias)
                if not lhsvar._q_invariant or len(lhsvar.stinfo['possibletypes']) == 1:
                    restriction = " AND %s.type='%s'" % (ealias, self._state.solution[lhs.name])
                else:
                    etypes = ','.join("'%s'" % etype for etype in lhsvar.stinfo['possibletypes'])
                    restriction = " AND %s.type IN (%s)" % (ealias, etypes)
        if isinstance(rel.parent, Not):
            self._state.done.add(rel.parent)
            not_ = True
        else:
            not_ = False
        query = const.eval(self._args)
        return self.dbhelper.fti_restriction_sql(alias, query,
                                                 jointo, not_) + restriction

    def visit_comparison(self, cmp):
        """generate SQL for a comparison"""
        optional = getattr(cmp, 'optional', None) # rql < 0.30
        if len(cmp.children) == 2:
            # simplified expression from HAVING clause
            lhs, rhs = cmp.children
        else:
            lhs = None
            rhs = cmp.children[0]
            assert not optional
        sql = None
        operator = cmp.operator
        if operator in ('LIKE', 'ILIKE'):
            if operator == 'ILIKE' and not self.dbhelper.ilike_support:
                operator = ' LIKE '
            else:
                operator = ' %s ' % operator
        elif operator == 'REGEXP':
            sql = ' %s' % self.dbhelper.sql_regexp_match_expression(rhs.accept(self))
        elif (operator == '=' and isinstance(rhs, Constant)
              and rhs.eval(self._args) is None):
            if lhs is None:
                sql = ' IS NULL'
            else:
                sql = '%s IS NULL' % lhs.accept(self)
        elif isinstance(rhs, Function) and rhs.name == 'IN':
            assert operator == '='
            operator = ' '
        if sql is None:
            if lhs is None:
                sql = '%s%s'% (operator, rhs.accept(self))
            else:
                sql = '%s%s%s'% (lhs.accept(self), operator, rhs.accept(self))
        if optional is None:
            return sql
        leftvars = cmp.children[0].get_nodes(VariableRef)
        assert len(leftvars) == 1
        if leftvars[0].variable.stinfo['attrvar'] is None:
            assert isinstance(leftvars[0].variable, ColumnAlias)
            leftalias = leftvars[0].variable._q_sqltable
        else:
            leftalias = self._var_table(leftvars[0].variable.stinfo['attrvar'])
        rightvars = cmp.children[1].get_nodes(VariableRef)
        assert len(rightvars) == 1
        if rightvars[0].variable.stinfo['attrvar'] is None:
            assert isinstance(rightvars[0].variable, ColumnAlias)
            rightalias = rightvars[0].variable._q_sqltable
        else:
            rightalias = self._var_table(rightvars[0].variable.stinfo['attrvar'])
        if optional == 'right':
            self._state.replace_tables_by_outer_join(
                leftalias, rightalias, 'LEFT', sql)
        elif optional == 'left':
            self._state.replace_tables_by_outer_join(
                rightalias, leftalias, 'LEFT', sql)
        else:
            self._state.replace_tables_by_outer_join(
                leftalias, rightalias, 'FULL', sql)
        return ''

    def visit_mathexpression(self, mexpr):
        """generate SQL for a mathematic expression"""
        lhs, rhs = mexpr.get_parts()
        # check for string concatenation
        operator = mexpr.operator
        if operator == '%':
            operator = '%%'
        try:
            if mexpr.operator == '+' and mexpr.get_type(self._state.solution, self._args) == 'String':
                return '(%s)' % self.dbhelper.sql_concat_string(lhs.accept(self),
                                                                rhs.accept(self))
        except CoercionError:
            pass
        return '(%s %s %s)'% (lhs.accept(self), operator, rhs.accept(self))

    def visit_unaryexpression(self, uexpr):
        """generate SQL for a unary expression"""
        return '%s%s'% (uexpr.operator, uexpr.children[0].accept(self))

    def visit_function(self, func):
        """generate SQL name for a function"""
        if func.name == 'FTIRANK':
            try:
                rel = next(iter(func.children[0].variable.stinfo['ftirels']))
            except KeyError:
                raise BadRQLQuery("can't use FTIRANK on variable not used in an"
                                  " 'has_text' relation (eg full-text search)")
            const = rel.get_parts()[1].children[0]
            return self.dbhelper.fti_rank_order(
                self._state.fti_table(rel, self.dbhelper.fti_table),
                const.eval(self._args))
        args = [c.accept(self) for c in func.children]
        if func in self._state.source_cb_funcs:
            # function executed as a callback on the source
            assert len(args) == 1
            return args[0]
        # func_as_sql will check function is supported by the backend
        return self.dbhelper.func_as_sql(func.name, args)

    def visit_constant(self, constant):
        """generate SQL name for a constant"""
        if constant.type is None:
            return 'NULL'
        value = constant.value
        if constant.type == 'etype':
            return value
        # don't substitute int, causes pb when used as sorting column number
        if constant.type == 'Int':
            return str(value)
        if constant.type in ('Date', 'Datetime'):
            rel = constant.relation()
            if rel is not None:
                rel._q_needcast = value
            return self.keyword_map[value]()
        if constant.type == 'Substitute':
            try:
                # we may found constant from simplified var in varmap
                return self._mapped_term(constant, '%%(%s)s' % value)[0]
            except KeyError:
                _id = value
                if PY2 and isinstance(_id, unicode):
                    _id = _id.encode()
        else:
            _id = str(id(constant)).replace('-', '', 1)
            self._query_attrs[_id] = value
        return '%%(%s)s' % _id

    def visit_variableref(self, variableref):
        """get the sql name for a variable reference"""
        # use accept, .variable may be a variable or a columnalias
        return variableref.variable.accept(self)

    def visit_columnalias(self, colalias):
        """get the sql name for a subquery column alias"""
        if colalias.name in self._varmap:
            sql = self._varmap[colalias.name]
            table = sql.split('.', 1)[0]
            colalias._q_sqltable = table
            colalias._q_sql = sql
            self._state.add_table(table)
            return sql
        return colalias._q_sql

    def visit_variable(self, variable):
        """get the table name and sql string for a variable"""
        #if contextrels is None and variable.name in self._state.done:
        if variable.name in self._state.done:
            if self._in_wrapping_query:
                return 'T1.%s' % self._state.aliases[variable.name]
            return variable._q_sql
        self._state.done.add(variable.name)
        vtablename = None
        if not self._state.ignore_varmap and variable.name in self._varmap:
            sql, vtablename = self._var_info(variable)
        elif variable.stinfo['attrvar']:
            # attribute variable (systematically used in rhs of final
            # relation(s)), get table name and sql from any rhs relation
            sql = self._linked_var_sql(variable)
        elif variable._q_invariant:
            # since variable is invariant, we know we won't found final relation
            principal = variable.stinfo['principal']
            if principal is None:
                vtablename = '_' + variable.name
                self._state.add_table('entities AS %s' % vtablename, vtablename)
                sql = '%s.eid' % vtablename
                if variable.stinfo['typerel'] is not None:
                    # add additional restriction on entities.type column
                    pts = variable.stinfo['possibletypes']
                    if len(pts) == 1:
                        etype = next(iter(variable.stinfo['possibletypes']))
                        restr = "%s.type='%s'" % (vtablename, etype)
                    else:
                        etypes = ','.join("'%s'" % et for et in pts)
                        restr = '%s.type IN (%s)' % (vtablename, etypes)
                    self._state.add_restriction(restr)
            elif principal.r_type == 'has_text':
                sql = '%s.%s' % (self._state.fti_table(principal,
                                                       self.dbhelper.fti_table),
                                 self.dbhelper.fti_uid_attr)
            elif principal in variable.stinfo['rhsrelations']:
                if self.schema.rschema(principal.r_type).inlined:
                    sql = self._linked_var_sql(variable)
                else:
                    sql = '%s.eid_to' % self._state.relation_table(principal)
            else:
                sql = '%s.eid_from' % self._state.relation_table(principal)
        else:
            # standard variable: get table name according to etype and use .eid
            # attribute
            sql, vtablename = self._var_info(variable)
        variable._q_sqltable = vtablename
        variable._q_sql = sql
        return sql

    # various utilities #######################################################

    def _extra_join_sql(self, relation, sql, var):
        # if rhs var is invariant, and this relation is not its principal,
        # generate extra join
        try:
            if not var.stinfo['principal'] is relation:
                op = relation.operator()
                if op == '=':
                    # need a predicable result for tests
                    args = sorted( (sql, var.accept(self)) )
                    args.insert(1, op)
                else:
                    args = (sql, op, var.accept(self))
                return '%s%s%s' % tuple(args)
        except KeyError:
            # no principal defined, relation is necessarily the principal and
            # so nothing to return here
            pass
        return None

    def _temp_table_scope(self, select, table):
        scope = 9999
        for var, sql in self._varmap.items():
            # skip "attribute variable" in varmap (such 'T.login')
            if not '.' in var and table == sql.split('.', 1)[0]:
                try:
                    scope = min(scope, self._state.scopes[select.defined_vars[var].scope])
                except KeyError:
                    scope = 0 # XXX
                if scope == 0:
                    break
        return scope

    def _mapped_term(self, term, key):
        """return sql and table alias to the `term`, mapped as `key` or raise
        KeyError when the key is not found in the varmap
        """
        sql = self._varmap[key]
        tablealias = sql.split('.', 1)[0]
        scope = self._temp_table_scope(term.stmt, tablealias)
        self._state.add_table(tablealias, scope=scope)
        return sql, tablealias

    def _var_info(self, var):
        try:
            return self._mapped_term(var, var.name)
        except KeyError:
            scope = self._state.scopes[var.scope]
            etype = self._state.solution[var.name]
            # XXX this check should be moved in rql.stcheck
            if self.schema.eschema(etype).final:
                raise BadRQLQuery(var.stmt.root)
            tablealias = '_' + var.name
            sql = '%s.%seid' % (tablealias, SQL_PREFIX)
            self._state.add_table('%s%s AS %s' % (SQL_PREFIX, etype, tablealias),
                           tablealias, scope=scope)
        return sql, tablealias

    def _inlined_var_sql(self, var, rtype):
        try:
            sql = self._varmap['%s.%s' % (var.name, rtype)]
            scope = self._state.scopes[var.scope]
            self._state.add_table(sql.split('.', 1)[0], scope=scope)
        except KeyError:
            # rtype may be an attribute relation when called from
            # _visit_var_attr_relation.  take care about 'eid' rtype, since in
            # some case we may use the `entities` table, so in that case we've
            # to properly use variable'sql
            if rtype == 'eid':
                sql = var.accept(self)
            else:
                sql = '%s.%s%s' % (self._var_table(var), SQL_PREFIX, rtype)
        return sql

    def _linked_var_sql(self, variable):
        if not self._state.ignore_varmap:
            try:
                return self._varmap[variable.name]
            except KeyError:
                pass
        rel = (variable.stinfo.get('principal') or
               next(iter(variable.stinfo['rhsrelations'])))
        linkedvar = rel.children[0].variable
        if rel.r_type == 'eid':
            return linkedvar.accept(self)
        if isinstance(linkedvar, ColumnAlias):
            raise BadRQLQuery('variable %s should be selected by the subquery'
                              % variable.name)
        try:
            sql = self._varmap['%s.%s' % (linkedvar.name, rel.r_type)]
        except KeyError:
            mapkey = '%s.%s' % (self._state.solution[linkedvar.name], rel.r_type)
            if mapkey in self.attr_map:
                cb, sourcecb = self.attr_map[mapkey]
                if not sourcecb:
                    return cb(self, linkedvar, rel)
                # attribute mapped at the source level (bfss for instance)
                stmt = rel.stmt
                for selectidx, vref in iter_mapped_var_sels(stmt, variable):
                    stack = [cb]
                    update_source_cb_stack(self._state, stmt, vref, stack)
                    self._state._needs_source_cb[selectidx] = stack
            linkedvar.accept(self)
            sql = '%s.%s%s' % (linkedvar._q_sqltable, SQL_PREFIX, rel.r_type)
        return sql

    # tables handling #########################################################

    def _var_table(self, var):
        var.accept(self)#.visit_variable(var)
        return var._q_sqltable