Google Git
Sign in
apache / hive / 4f4cbeda00d5ebb7d0b8cedee5daa2c03df4a755 / . / ql / src / java / org / apache / hadoop / hive / ql / optimizer / calcite / rules / HiveRelDecorrelator.java
blob: e43848720913be86ea2baae66d2282689367532b [file] [log] [blame]
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.hadoop.hive.ql.optimizer.calcite.rules;
import java.math.BigDecimal;
import java.util.ArrayList;
import java.util.Collection;
import java.util.Collections;
import java.util.HashMap;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.NavigableMap;
import java.util.Objects;
import java.util.Set;
import java.util.SortedMap;
import java.util.TreeMap;
import java.util.TreeSet;
import javax.annotation.Nonnull;
import org.apache.calcite.linq4j.Ord;
import org.apache.calcite.linq4j.function.Function2;
import org.apache.calcite.plan.Context;
import org.apache.calcite.plan.RelOptCluster;
import org.apache.calcite.plan.RelOptCostImpl;
import org.apache.calcite.plan.RelOptRule;
import org.apache.calcite.plan.RelOptRuleCall;
import org.apache.calcite.plan.RelOptUtil;
import org.apache.calcite.plan.hep.HepPlanner;
import org.apache.calcite.plan.hep.HepProgram;
import org.apache.calcite.plan.hep.HepRelVertex;
import org.apache.calcite.rel.BiRel;
import org.apache.calcite.rel.RelCollation;
import org.apache.calcite.rel.RelNode;
import org.apache.calcite.rel.core.Aggregate;
import org.apache.calcite.rel.core.AggregateCall;
import org.apache.calcite.rel.core.Correlate;
import org.apache.calcite.rel.core.CorrelationId;
import org.apache.calcite.rel.core.Filter;
import org.apache.calcite.rel.core.Join;
import org.apache.calcite.rel.core.JoinRelType;
import org.apache.calcite.rel.core.Project;
import org.apache.calcite.rel.core.Sort;
import org.apache.calcite.rel.core.Values;
import org.apache.calcite.rel.logical.LogicalCorrelate;
import org.apache.calcite.rel.metadata.RelMdUtil;
import org.apache.calcite.rel.metadata.RelMetadataQuery;
import org.apache.calcite.rel.type.RelDataType;
import org.apache.calcite.rel.type.RelDataTypeFactory;
import org.apache.calcite.rel.type.RelDataTypeField;
import org.apache.calcite.rex.RexBuilder;
import org.apache.calcite.rex.RexCall;
import org.apache.calcite.rex.RexCorrelVariable;
import org.apache.calcite.rex.RexFieldAccess;
import org.apache.calcite.rex.RexInputRef;
import org.apache.calcite.rex.RexLiteral;
import org.apache.calcite.rex.RexNode;
import org.apache.calcite.rex.RexOver;
import org.apache.calcite.rex.RexShuttle;
import org.apache.calcite.rex.RexSubQuery;
import org.apache.calcite.rex.RexUtil;
import org.apache.calcite.rex.RexVisitorImpl;
import org.apache.calcite.sql.SqlFunction;
import org.apache.calcite.sql.SqlKind;
import org.apache.calcite.sql.SqlOperator;
import org.apache.calcite.sql.fun.SqlCountAggFunction;
import org.apache.calcite.sql.fun.SqlSingleValueAggFunction;
import org.apache.calcite.sql.fun.SqlStdOperatorTable;
import org.apache.calcite.sql.validate.SqlValidatorUtil;
import org.apache.calcite.tools.RelBuilder;
import org.apache.calcite.util.Holder;
import org.apache.calcite.util.ImmutableBitSet;
import org.apache.calcite.util.Litmus;
import org.apache.calcite.util.Pair;
import org.apache.calcite.util.ReflectUtil;
import org.apache.calcite.util.ReflectiveVisitor;
import org.apache.calcite.util.Stacks;
import org.apache.calcite.util.Util;
import org.apache.calcite.util.mapping.Mappings;
import org.apache.hadoop.hive.ql.optimizer.calcite.Bug;
import org.apache.hadoop.hive.ql.optimizer.calcite.HiveRelFactories;
import org.apache.hadoop.hive.ql.optimizer.calcite.HiveRelOptUtil;
import org.apache.hadoop.hive.ql.optimizer.calcite.HiveRelShuttleImpl;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveAggregate;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveAntiJoin;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveFilter;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveIntersect;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveJoin;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveProject;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveRelNode;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveSemiJoin;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveSortLimit;
import org.apache.hadoop.hive.ql.optimizer.calcite.reloperators.HiveUnion;
import org.apache.hadoop.hive.ql.parse.SemanticException;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import com.google.common.base.Preconditions;
import com.google.common.base.Supplier;
import com.google.common.collect.ImmutableList;
import com.google.common.collect.ImmutableMap;
import com.google.common.collect.ImmutableSet;
import com.google.common.collect.ImmutableSortedMap;
import com.google.common.collect.Lists;
import com.google.common.collect.Maps;
import com.google.common.collect.Multimap;
import com.google.common.collect.Multimaps;
import com.google.common.collect.Sets;
import com.google.common.collect.SortedSetMultimap;
import java.util.Stack;
/**
* NOTE: this whole logic is replicated from Calcite's RelDecorrelator
* and is exteneded to make it suitable for HIVE
* We should get rid of this and replace it with Calcite's RelDecorrelator
* once that works with Join, Project etc instead of Join, Project.
* At this point this has differed from Calcite's version significantly so cannot
* get rid of this.
*
* RelDecorrelator replaces all correlated expressions (corExp) in a relational
* expression (RelNode) tree with non-correlated expressions that are produced
* from joining the RelNode that produces the corExp with the RelNode that
* references it.
*
* <p>TODO:</p>
* <ul>
* <li>replace {@code CorelMap} constructor parameter with a RelNode
* <li>make {@link #currentRel} immutable (would require a fresh
* RelDecorrelator for each node being decorrelated)</li>
* <li>make fields of {@code CorelMap} immutable</li>
* <li>make sub-class rules static, and have them create their own
* de-correlator</li>
* </ul>
*/
public final class HiveRelDecorrelator implements ReflectiveVisitor {
//~ Static fields/initializers ---------------------------------------------
protected static final Logger LOG = LoggerFactory.getLogger(
HiveRelDecorrelator.class);
private static final FilterFlattenCorrelatedConditionRule FLATTEN_CORRELATED_CONDITION_RULE =
new FilterFlattenCorrelatedConditionRule(HiveFilter.class, HiveRelFactories.HIVE_BUILDER,
"HiveFilterFlattenCorrelatedConditionRule");
//~ Instance fields --------------------------------------------------------
private final RelBuilder relBuilder;
// map built during translation
private CorelMap cm;
private final ReflectUtil.MethodDispatcher<Frame> dispatcher =
ReflectUtil.createMethodDispatcher(Frame.class, this, "decorrelateRel",
RelNode.class);
private final RexBuilder rexBuilder;
// The rel which is being visited
private RelNode currentRel;
private final Context context;
/** Built during decorrelation, of rel to all the newly created correlated
* variables in its output, and to map old input positions to new input
* positions. This is from the view point of the parent rel of a new rel. */
private final Map<RelNode, Frame> map = new HashMap<>();
private final HashSet<LogicalCorrelate> generatedCorRels = Sets.newHashSet();
private Stack valueGen = new Stack();
//~ Constructors -----------------------------------------------------------
private HiveRelDecorrelator(
RelOptCluster cluster,
CorelMap cm,
Context context) {
this.cm = cm;
this.rexBuilder = cluster.getRexBuilder();
this.context = context;
relBuilder = HiveRelFactories.HIVE_BUILDER.create(cluster, null);
}
//~ Methods ----------------------------------------------------------------
/** Decorrelates a query.
*
* <p>This is the main entry point to {@code RelDecorrelator}.
*
* @param rootRel Root node of the query
*
* @return Equivalent query with all
* {@link org.apache.calcite.rel.logical.LogicalCorrelate} instances removed
*/
public static RelNode decorrelateQuery(RelNode rootRel) {
final CorelMap corelMap = new CorelMapBuilder().build(rootRel);
if (!corelMap.hasCorrelation()) {
return rootRel;
}
final RelOptCluster cluster = rootRel.getCluster();
final HiveRelDecorrelator decorrelator =
new HiveRelDecorrelator(cluster, corelMap,
cluster.getPlanner().getContext());
RelNode newRootRel = decorrelator.removeCorrelationViaRule(rootRel);
if (!decorrelator.cm.mapCorToCorRel.isEmpty()) {
newRootRel = decorrelator.decorrelate(newRootRel);
}
return newRootRel;
}
private void setCurrent(RelNode root, LogicalCorrelate corRel) {
currentRel = corRel;
if (corRel != null) {
cm = new CorelMapBuilder().build(Util.first(root, corRel));
}
}
private RelNode decorrelate(RelNode root) {
// first adjust count() expression if any
HepProgram program = HepProgram.builder()
.addRuleInstance(new AdjustProjectForCountAggregateRule(false))
.addRuleInstance(new AdjustProjectForCountAggregateRule(true))
.addRuleInstance(HiveFilterJoinRule.FILTER_ON_JOIN)
.addRuleInstance(HiveFilterProjectTransposeRule.INSTANCE)
.addRuleInstance(FLATTEN_CORRELATED_CONDITION_RULE)
// FilterCorrelateRule rule mistakenly pushes a FILTER, consiting of correlated vars,
// on top of LogicalCorrelate to within left input for scalar corr queries
// which causes exception during decorrelation. This has been disabled for now.
//.addRuleInstance(FilterCorrelateRule.INSTANCE)
.build();
HepPlanner planner = createPlanner(program);
planner.setRoot(root);
root = planner.findBestExp();
if (LOG.isDebugEnabled()) {
LOG.debug("Plan before extracting correlated computations:\n" + RelOptUtil.toString(root));
}
root = root.accept(new CorrelateProjectExtractor(HiveRelFactories.HIVE_BUILDER));
// Necessary to update cm (CorrelMap) since CorrelateProjectExtractor above may modify the plan
this.cm = new CorelMapBuilder().build(root);
if (LOG.isDebugEnabled()) {
LOG.debug("Plan after extracting correlated computations:\n" + RelOptUtil.toString(root));
}
// Perform decorrelation.
map.clear();
final Frame frame = getInvoke(root, null);
if (frame != null) {
// has been rewritten; apply rules post-decorrelation
final HepProgram program2 = HepProgram.builder()
.addRuleInstance(HiveFilterJoinRule.FILTER_ON_JOIN)
.addRuleInstance(HiveFilterJoinRule.JOIN)
.build();
final HepPlanner planner2 = createPlanner(program2);
final RelNode newRoot = frame.r;
planner2.setRoot(newRoot);
return planner2.findBestExp();
}
assert(valueGen.isEmpty());
return root;
}
private Function2<RelNode, RelNode, Void> createCopyHook() {
return new Function2<RelNode, RelNode, Void>() {
@Override
public Void apply(RelNode oldNode, RelNode newNode) {
if (cm.mapRefRelToCorRef.containsKey(oldNode)) {
final CorelMap corelMap = new CorelMapBuilder().build(newNode);
// since after various rules original relnode could have different
// corref (or might not have at all) we need to traverse the new node
// to figure out new cor refs and put that into map
if(!corelMap.mapRefRelToCorRef.isEmpty()){
cm.mapRefRelToCorRef.putAll(newNode,
corelMap.mapRefRelToCorRef.get(newNode));
}
}
if (oldNode instanceof LogicalCorrelate
&& newNode instanceof LogicalCorrelate) {
LogicalCorrelate oldCor = (LogicalCorrelate) oldNode;
CorrelationId c = oldCor.getCorrelationId();
if (cm.mapCorToCorRel.get(c) == oldNode) {
cm.mapCorToCorRel.put(c, newNode);
}
if (generatedCorRels.contains(oldNode)) {
generatedCorRels.add((LogicalCorrelate) newNode);
}
}
return null;
}
};
}
private HepPlanner createPlanner(HepProgram program) {
// Create a planner with a hook to update the mapping tables when a
// node is copied when it is registered.
return new HepPlanner(
program,
context,
true,
createCopyHook(),
RelOptCostImpl.FACTORY);
}
public RelNode removeCorrelationViaRule(RelNode root) {
HepProgram program = HepProgram.builder()
.addRuleInstance(new RemoveSingleAggregateRule())
.addRuleInstance(new RemoveCorrelationForScalarProjectRule())
.addRuleInstance(new RemoveCorrelationForScalarAggregateRule())
.build();
HepPlanner planner = createPlanner(program);
planner.setRoot(root);
return planner.findBestExp();
}
protected RexNode decorrelateExpr(RexNode exp) {
return decorrelateExpr(exp, true);
}
protected RexNode decorrelateExpr(RexNode exp, boolean valueGenerator) {
DecorrelateRexShuttle shuttle = new DecorrelateRexShuttle();
shuttle.setValueGenerator(valueGenerator);
return exp.accept(shuttle);
}
protected RexNode removeCorrelationExpr(
RexNode exp,
boolean projectPulledAboveLeftCorrelator) {
return removeCorrelationExpr(exp, projectPulledAboveLeftCorrelator, ImmutableSet.of());
}
protected RexNode removeCorrelationExpr(
RexNode exp,
boolean projectPulledAboveLeftCorrelator,
RexInputRef nullIndicator) {
return removeCorrelationExpr(exp, projectPulledAboveLeftCorrelator, nullIndicator, ImmutableSet.of());
}
protected RexNode removeCorrelationExpr(
RexNode exp,
boolean projectPulledAboveLeftCorrelator,
Set<Integer> isCount) {
return removeCorrelationExpr(exp, projectPulledAboveLeftCorrelator, null, isCount);
}
protected RexNode removeCorrelationExpr(
RexNode exp,
boolean projectPulledAboveLeftCorrelator,
RexInputRef nullIndicator,
Set<Integer> isCount) {
RemoveCorrelationRexShuttle shuttle =
new RemoveCorrelationRexShuttle(rexBuilder, projectPulledAboveLeftCorrelator, nullIndicator, isCount);
return exp.accept(shuttle);
}
/** Fallback if none of the other {@code decorrelateRel} methods match. */
public Frame decorrelateRel(RelNode rel) {
RelNode newRel = rel.copy(rel.getTraitSet(), rel.getInputs());
if (rel.getInputs().size() > 0) {
List<RelNode> oldInputs = rel.getInputs();
List<RelNode> newInputs = Lists.newArrayList();
for (int i = 0; i < oldInputs.size(); ++i) {
final Frame frame = getInvoke(oldInputs.get(i), rel);
if (frame == null || !frame.corDefOutputs.isEmpty()) {
// if input is not rewritten, or if it produces correlated
// variables, terminate rewrite
return null;
}
newInputs.add(frame.r);
newRel.replaceInput(i, frame.r);
}
if (!Util.equalShallow(oldInputs, newInputs)) {
newRel = rel.copy(rel.getTraitSet(), newInputs);
}
}
// the output position should not change since there are no corVars
// coming from below.
return register(rel, newRel, identityMap(rel.getRowType().getFieldCount()),
ImmutableSortedMap.<CorDef, Integer>of());
}
/**
* Rewrite Sort.
*
* @param rel Sort to be rewritten
*/
public Frame decorrelateRel(HiveSortLimit rel) {
return decorrelateRel((Sort) rel);
}
/**
* Rewrite Sort.
*
* @param rel Sort to be rewritten
*/
public Frame decorrelateRel(Sort rel) {
//
// Rewrite logic:
//
// 1. change the collations field to reference the new input.
//
// Sort itself should not reference cor vars.
assert !cm.mapRefRelToCorRef.containsKey(rel);
// Sort only references field positions in collations field.
// The collations field in the newRel now need to refer to the
// new output positions in its input.
// Its output does not change the input ordering, so there's no
// need to call propagateExpr.
final RelNode oldInput = rel.getInput();
final Frame frame = getInvoke(oldInput, rel);
if (frame == null) {
// If input has not been rewritten, do not rewrite this rel.
return null;
}
final RelNode newInput = frame.r;
Mappings.TargetMapping mapping =
Mappings.target(
frame.oldToNewOutputs,
oldInput.getRowType().getFieldCount(),
newInput.getRowType().getFieldCount());
RelCollation oldCollation = rel.getCollation();
RelCollation newCollation = RexUtil.apply(mapping, oldCollation);
final RelNode newSort = HiveSortLimit.create(newInput, newCollation, rel.offset, rel.fetch);
// Sort does not change input ordering
return register(rel, newSort, frame.oldToNewOutputs,
frame.corDefOutputs);
}
/**
* Rewrites a {@link Values}.
*
* @param rel Values to be rewritten
*/
public Frame decorrelateRel(Values rel) {
// There are no inputs, so rel does not need to be changed.
return null;
}
/**
* Rewrites a {@link Aggregate}.
*
* @param rel Aggregate to rewrite
*/
public Frame decorrelateRel(Aggregate rel) throws SemanticException {
//
// Rewrite logic:
//
// 1. Permute the group by keys to the front.
// 2. If the input of an aggregate produces correlated variables,
// add them to the group list.
// 3. Change aggCalls to reference the new project.
//
// Aggregate itself should not reference cor vars.
assert !cm.mapRefRelToCorRef.containsKey(rel);
final RelNode oldInput = rel.getInput();
final Frame frame = getInvoke(oldInput, rel);
if (frame == null) {
// If input has not been rewritten, do not rewrite this rel.
return null;
}
final RelNode newInput = frame.r;
// map from newInput
Map<Integer, Integer> mapNewInputToProjOutputs = new HashMap<>();
final int oldGroupKeyCount = rel.getGroupSet().cardinality();
// Project projects the original expressions,
// plus any correlated variables the input wants to pass along.
final List<Pair<RexNode, String>> projects = Lists.newArrayList();
List<RelDataTypeField> newInputOutput =
newInput.getRowType().getFieldList();
int newPos = 0;
// oldInput has the original group by keys in the front.
final NavigableMap<Integer, RexLiteral> omittedConstants = new TreeMap<>();
for (int i = 0; i < oldGroupKeyCount; i++) {
final RexLiteral constant = projectedLiteral(newInput, i);
if (constant != null) {
// Exclude constants. Aggregate({true}) occurs because Aggregate({})
// would generate 1 row even when applied to an empty table.
omittedConstants.put(i, constant);
continue;
}
int newInputPos = frame.oldToNewOutputs.get(i);
projects.add(RexInputRef.of2(newInputPos, newInputOutput));
mapNewInputToProjOutputs.put(newInputPos, newPos);
newPos++;
}
final SortedMap<CorDef, Integer> corDefOutputs = new TreeMap<>();
if (!frame.corDefOutputs.isEmpty()) {
// If input produces correlated variables, move them to the front,
// right after any existing GROUP BY fields.
// Now add the corVars from the input, starting from
// position oldGroupKeyCount.
for (Map.Entry<CorDef, Integer> entry
: frame.corDefOutputs.entrySet()) {
projects.add(RexInputRef.of2(entry.getValue(), newInputOutput));
corDefOutputs.put(entry.getKey(), newPos);
mapNewInputToProjOutputs.put(entry.getValue(), newPos);
newPos++;
}
}
// add the remaining fields
final int newGroupKeyCount = newPos;
for (int i = 0; i < newInputOutput.size(); i++) {
if (!mapNewInputToProjOutputs.containsKey(i)) {
projects.add(RexInputRef.of2(i, newInputOutput));
mapNewInputToProjOutputs.put(i, newPos);
newPos++;
}
}
assert newPos == newInputOutput.size();
// This Project will be what the old input maps to,
// replacing any previous mapping from old input).
RelNode newProject = HiveProject.create(newInput, Pair.left(projects), Pair.right(projects));
// update mappings:
// oldInput ----> newInput
//
// newProject
// |
// oldInput ----> newInput
//
// is transformed to
//
// oldInput ----> newProject
// |
// newInput
Map<Integer, Integer> combinedMap = Maps.newHashMap();
for (Integer oldInputPos : frame.oldToNewOutputs.keySet()) {
combinedMap.put(oldInputPos,
mapNewInputToProjOutputs.get(
frame.oldToNewOutputs.get(oldInputPos)));
}
register(oldInput, newProject, combinedMap, corDefOutputs);
// now it's time to rewrite the Aggregate
final ImmutableBitSet newGroupSet = ImmutableBitSet.range(newGroupKeyCount);
List<AggregateCall> newAggCalls = Lists.newArrayList();
List<AggregateCall> oldAggCalls = rel.getAggCallList();
int oldInputOutputFieldCount = rel.getGroupSet().cardinality();
int newInputOutputFieldCount = newGroupSet.cardinality();
int i = -1;
for (AggregateCall oldAggCall : oldAggCalls) {
++i;
List<Integer> oldAggArgs = oldAggCall.getArgList();
List<Integer> aggArgs = Lists.newArrayList();
// Adjust the aggregator argument positions.
// Note aggregator does not change input ordering, so the input
// output position mapping can be used to derive the new positions
// for the argument.
for (int oldPos : oldAggArgs) {
aggArgs.add(combinedMap.get(oldPos));
}
final int filterArg = oldAggCall.filterArg < 0 ? oldAggCall.filterArg
: combinedMap.get(oldAggCall.filterArg);
newAggCalls.add(
oldAggCall.adaptTo(newProject, aggArgs, filterArg,
oldGroupKeyCount, newGroupKeyCount));
// The old to new output position mapping will be the same as that
// of newProject, plus any aggregates that the oldAgg produces.
combinedMap.put(
oldInputOutputFieldCount + i,
newInputOutputFieldCount + i);
}
relBuilder.push(newProject)
.aggregate(relBuilder.groupKey(newGroupSet, null), newAggCalls);
if (!omittedConstants.isEmpty()) {
final List<RexNode> postProjects = new ArrayList<>(relBuilder.fields());
for (Map.Entry<Integer, RexLiteral> entry
: omittedConstants.descendingMap().entrySet()) {
postProjects.add(entry.getKey() + frame.corDefOutputs.size(),
entry.getValue());
}
relBuilder.project(postProjects);
}
// Aggregate does not change input ordering so corVars will be
// located at the same position as the input newProject.
return register(rel, relBuilder.build(), combinedMap, corDefOutputs);
}
public Frame getInvoke(RelNode r, RelNode parent) {
final Frame frame = dispatcher.invoke(r);
if (frame != null) {
map.put(r, frame);
}
currentRel = parent;
return frame;
}
/** Returns a literal output field, or null if it is not literal. */
private static RexLiteral projectedLiteral(RelNode rel, int i) {
if (rel instanceof Project) {
final Project project = (Project) rel;
final RexNode node = project.getProjects().get(i);
if (node instanceof RexLiteral) {
return (RexLiteral) node;
}
}
return null;
}
public Frame decorrelateRel(HiveAggregate rel) throws SemanticException {
//
// Rewrite logic:
//
// 1. Permute the group by keys to the front.
// 2. If the input of an aggregate produces correlated variables,
// add them to the group list.
// 3. Change aggCalls to reference the new project.
//
// Aggregate itself should not reference cor vars.
assert !cm.mapRefRelToCorRef.containsKey(rel);
final RelNode oldInput = rel.getInput();
final Frame frame = getInvoke(oldInput, rel);
if (frame == null) {
// If input has not been rewritten, do not rewrite this rel.
return null;
}
//assert !frame.corVarOutputPos.isEmpty();
final RelNode newInput = frame.r;
// map from newInput
Map<Integer, Integer> mapNewInputToProjOutputs = new HashMap<>();
final int oldGroupKeyCount = rel.getGroupSet().cardinality();
// Project projects the original expressions,
// plus any correlated variables the input wants to pass along.
final List<Pair<RexNode, String>> projects = Lists.newArrayList();
List<RelDataTypeField> newInputOutput =
newInput.getRowType().getFieldList();
int newPos = 0;
// oldInput has the original group by keys in the front.
final NavigableMap<Integer, RexLiteral> omittedConstants = new TreeMap<>();
for (int i = 0; i < oldGroupKeyCount; i++) {
final RexLiteral constant = projectedLiteral(newInput, i);
int newInputPos = frame.oldToNewOutputs.get(i);
projects.add(RexInputRef.of2(newInputPos, newInputOutput));
mapNewInputToProjOutputs.put(newInputPos, newPos);
newPos++;
}
final SortedMap<CorDef, Integer> corDefOutputs = new TreeMap<>();
if (!frame.corDefOutputs.isEmpty()) {
// If input produces correlated variables, move them to the front,
// right after any existing GROUP BY fields.
// Now add the corVars from the input, starting from
// position oldGroupKeyCount.
for (Map.Entry<CorDef, Integer> entry
: frame.corDefOutputs.entrySet()) {
projects.add(RexInputRef.of2(entry.getValue(), newInputOutput));
corDefOutputs.put(entry.getKey(), newPos);
mapNewInputToProjOutputs.put(entry.getValue(), newPos);
newPos++;
}
}
// add the remaining fields
final int newGroupKeyCount = newPos;
for (int i = 0; i < newInputOutput.size(); i++) {
if (!mapNewInputToProjOutputs.containsKey(i)) {
projects.add(RexInputRef.of2(i, newInputOutput));
mapNewInputToProjOutputs.put(i, newPos);
newPos++;
}
}
assert newPos == newInputOutput.size();
// This Project will be what the old input maps to,
// replacing any previous mapping from old input).
RelNode newProject = HiveProject.create(newInput, Pair.left(projects), Pair.right(projects));
// update mappings:
// oldInput ----> newInput
//
// newProject
// |
// oldInput ----> newInput
//
// is transformed to
//
// oldInput ----> newProject
// |
// newInput
Map<Integer, Integer> combinedMap = Maps.newHashMap();
for (Integer oldInputPos : frame.oldToNewOutputs.keySet()) {
combinedMap.put(oldInputPos,
mapNewInputToProjOutputs.get(
frame.oldToNewOutputs.get(oldInputPos)));
}
register(oldInput, newProject, combinedMap, corDefOutputs);
// now it's time to rewrite the Aggregate
final ImmutableBitSet newGroupSet = ImmutableBitSet.range(newGroupKeyCount);
List<AggregateCall> newAggCalls = Lists.newArrayList();
List<AggregateCall> oldAggCalls = rel.getAggCallList();
int oldInputOutputFieldCount = rel.getGroupSet().cardinality();
int newInputOutputFieldCount = newGroupSet.cardinality();
int i = -1;
for (AggregateCall oldAggCall : oldAggCalls) {
++i;
List<Integer> oldAggArgs = oldAggCall.getArgList();
List<Integer> aggArgs = Lists.newArrayList();
// Adjust the aggregator argument positions.
// Note aggregator does not change input ordering, so the input
// output position mapping can be used to derive the new positions
// for the argument.
for (int oldPos : oldAggArgs) {
aggArgs.add(combinedMap.get(oldPos));
}
final int filterArg = oldAggCall.filterArg < 0 ? oldAggCall.filterArg
: combinedMap.get(oldAggCall.filterArg);
newAggCalls.add(
oldAggCall.adaptTo(newProject, aggArgs, filterArg,
oldGroupKeyCount, newGroupKeyCount));
// The old to new output position mapping will be the same as that
// of newProject, plus any aggregates that the oldAgg produces.
combinedMap.put(
oldInputOutputFieldCount + i,
newInputOutputFieldCount + i);
}
relBuilder.push(
new HiveAggregate(rel.getCluster(), rel.getTraitSet(), newProject,
newGroupSet, null, newAggCalls));
if (!omittedConstants.isEmpty()) {
final List<RexNode> postProjects = new ArrayList<>(relBuilder.fields());
for (Map.Entry<Integer, RexLiteral> entry
: omittedConstants.descendingMap().entrySet()) {
postProjects.add(entry.getKey() + frame.corDefOutputs.size(),
entry.getValue());
}
relBuilder.project(postProjects);
}
// Aggregate does not change input ordering so corVars will be
// located at the same position as the input newProject.
return register(rel, relBuilder.build(), combinedMap, corDefOutputs);
}
public Frame decorrelateRel(HiveProject rel) throws SemanticException {
return decorrelateRel((Project) rel);
}
public Frame decorrelateRel(Project rel) throws SemanticException {
//
// Rewrite logic:
//
// 1. Pass along any correlated variables coming from the input.
//
final RelNode oldInput = rel.getInput();
Frame frame = getInvoke(oldInput, rel);
if (frame == null) {
// If input has not been rewritten, do not rewrite this rel.
return null;
}
final List<RexNode> oldProjects = rel.getProjects();
final List<RelDataTypeField> relOutput = rel.getRowType().getFieldList();
// Project projects the original expressions,
// plus any correlated variables the input wants to pass along.
final List<Pair<RexNode, String>> projects = Lists.newArrayList();
// If this Project has correlated reference, create value generator
// and produce the correlated variables in the new output.
if (cm.mapRefRelToCorRef.containsKey(rel)) {
frame = decorrelateInputWithValueGenerator(rel);
}
// Project projects the original expressions
final Map<Integer, Integer> mapOldToNewOutputs = new HashMap<>();
int newPos;
for (newPos = 0; newPos < oldProjects.size(); newPos++) {
projects.add(
newPos,
Pair.of(
decorrelateExpr(oldProjects.get(newPos)),
relOutput.get(newPos).getName()));
mapOldToNewOutputs.put(newPos, newPos);
}
// Project any correlated variables the input wants to pass along.
final SortedMap<CorDef, Integer> corDefOutputs = new TreeMap<>();
for (Map.Entry<CorDef, Integer> entry : frame.corDefOutputs.entrySet()) {
projects.add(
RexInputRef.of2(entry.getValue(),
frame.r.getRowType().getFieldList()));
corDefOutputs.put(entry.getKey(), newPos);
newPos++;
}
RelNode newProject = HiveProject.create(frame.r, Pair.left(projects),
SqlValidatorUtil.uniquify(Pair.right(projects)));
return register(rel, newProject, mapOldToNewOutputs,
corDefOutputs);
}
/**
* Create RelNode tree that produces a list of correlated variables.
*/
private RelNode createValueGenerator(
Iterable<CorRef> correlations,
int valueGenFieldOffset,
SortedMap<CorDef, Integer> corDefOutputs) {
final Map<RelNode, List<Integer>> mapNewInputToOutputs =
new HashMap<>();
final Map<RelNode, Integer> mapNewInputToNewOffset = new HashMap<>();
// Input provides the definition of a correlated variable.
// Add to map all the referenced positions (relative to each input rel).
for (CorRef corVar : correlations) {
final int oldCorVarOffset = corVar.field;
final RelNode oldInput = getCorRel(corVar);
assert oldInput != null;
final Frame frame = map.get(oldInput);
assert frame != null;
final RelNode newInput = frame.r;
final List<Integer> newLocalOutputs;
if (!mapNewInputToOutputs.containsKey(newInput)) {
newLocalOutputs = new ArrayList<>();
} else {
newLocalOutputs =
mapNewInputToOutputs.get(newInput);
}
final int newCorVarOffset = frame.oldToNewOutputs.get(oldCorVarOffset);
// Add all unique positions referenced.
if (!newLocalOutputs.contains(newCorVarOffset)) {
newLocalOutputs.add(newCorVarOffset);
}
mapNewInputToOutputs.put(newInput, newLocalOutputs);
}
int offset = 0;
// Project only the correlated fields out of each inputRel
// and join the projectRel together.
// To make sure the plan does not change in terms of join order,
// join these rels based on their occurrence in cor var list which
// is sorted.
final Set<RelNode> joinedInputs = new HashSet<>();
RelNode r = null;
for (CorRef corVar : correlations) {
final RelNode oldInput = getCorRel(corVar);
assert oldInput != null;
final RelNode newInput = map.get(oldInput).r;
assert newInput != null;
if (!joinedInputs.contains(newInput)) {
RelNode project = RelOptUtil.createProject(
HiveRelFactories.HIVE_PROJECT_FACTORY, newInput, mapNewInputToOutputs.get(newInput));
RelNode distinct = relBuilder.push(project).distinct().build();
RelOptCluster cluster = distinct.getCluster();
joinedInputs.add(newInput);
mapNewInputToNewOffset.put(newInput, offset);
offset += distinct.getRowType().getFieldCount();
if (r == null) {
r = distinct;
} else {
r = relBuilder.push(r).push(distinct)
.join(JoinRelType.INNER, cluster.getRexBuilder().makeLiteral(true)).build();
}
}
}
// Translate the positions of correlated variables to be relative to
// the join output, leaving room for valueGenFieldOffset because
// valueGenerators are joined with the original left input of the rel
// referencing correlated variables.
for (CorRef corRef : correlations) {
// The first input of a Correlator is always the rel defining
// the correlated variables.
final RelNode oldInput = getCorRel(corRef);
assert oldInput != null;
final Frame frame = map.get(oldInput);
final RelNode newInput = frame.r;
assert newInput != null;
final List<Integer> newLocalOutputs =
mapNewInputToOutputs.get(newInput);
final int newLocalOutput = frame.oldToNewOutputs.get(corRef.field);
// newOutput is the index of the cor var in the referenced
// position list plus the offset of referenced position list of
// each newInput.
final int newOutput =
newLocalOutputs.indexOf(newLocalOutput)
+ mapNewInputToNewOffset.get(newInput)
+ valueGenFieldOffset;
corDefOutputs.put(corRef.def(), newOutput);
}
return r;
}
//this returns the source of corVar i.e. Rel which produces cor var
// value. Therefore it is always LogicalCorrelate's left input which is outer query
private RelNode getCorRel(CorRef corVar) {
final RelNode r = cm.mapCorToCorRel.get(corVar.corr);
RelNode ret = r.getInput(0);
return ret;
}
private Frame decorrelateInputWithValueGenerator(RelNode rel) {
// currently only handles one input input
assert rel.getInputs().size() == 1;
RelNode oldInput = rel.getInput(0);
final Frame frame = map.get(oldInput);
final SortedMap<CorDef, Integer> corDefOutputs =
new TreeMap<>(frame.corDefOutputs);
final Collection<CorRef> corVarList = cm.mapRefRelToCorRef.get(rel);
// Try to populate correlation variables using local fields.
// This means that we do not need a value generator.
if (rel instanceof Filter) {
SortedMap<CorDef, Integer> coreMap = new TreeMap<>();
for (CorRef correlation : corVarList) {
final CorDef def = correlation.def();
// If the correlation/variable is already provided by the input
// then we don't have to look for equivalent expressions and
// we don't need to create value generator for them.
if (corDefOutputs.containsKey(def)) {
coreMap.put(def, corDefOutputs.get(def));
continue;
}
// If the correlation/variable is in the map then we already
// seen this before in this loop so we don't need to treat it again.
if (coreMap.containsKey(def)) {
continue;
}
try {
findCorrelationEquivalent(correlation, ((Filter) rel).getCondition());
} catch (Util.FoundOne e) {
// we need to keep predicate kind e.g. EQUAL or NOT EQUAL
// so that later while decorrelating LogicalCorrelate appropriate join predicate
// is generated
def.setPredicateKind((SqlOperator) ((Pair)((Pair)e.getNode()).getValue()).getKey());
def.setIsLeft((boolean)((Pair)((Pair) e.getNode()).getValue()).getValue());
final Integer oldInputRef = (Integer) ((Pair) e.getNode()).getKey();
final Integer newInputRef = frame.oldToNewOutputs.get(oldInputRef);
coreMap.put(def, newInputRef);
}
}
// If all correlation variables are now satisfied, skip creating a value
// generator.
if (coreMap.size() == corVarList.size()) {
coreMap.putAll(frame.corDefOutputs);
return register(oldInput, frame.r,
frame.oldToNewOutputs, coreMap);
}
}
int leftInputOutputCount = frame.r.getRowType().getFieldCount();
// can directly add positions into corDefOutputs since join
// does not change the output ordering from the inputs.
RelNode valueGenRel =
createValueGenerator(
corVarList,
leftInputOutputCount,
corDefOutputs);
RelNode join = relBuilder.push(frame.r).push(valueGenRel)
.join(JoinRelType.INNER, rexBuilder.makeLiteral(true)).build();
// Join or Filter does not change the old input ordering. All
// input fields from newLeftInput(i.e. the original input to the old
// Filter) are in the output and in the same position.
return register(oldInput, join, frame.oldToNewOutputs, corDefOutputs);
}
/** Finds a {@link RexInputRef} that is equivalent to a {@link CorRef},
* and if found, throws a {@link Util.FoundOne}. */
private void findCorrelationEquivalent(CorRef correlation, RexNode e)
throws Util.FoundOne {
if(e instanceof RexCall){
switch (e.getKind()) {
case AND:
for (RexNode operand : ((RexCall) e).getOperands()) {
findCorrelationEquivalent(correlation, operand);
}
default:
final RexCall call = (RexCall) e;
final List<RexNode> operands = call.getOperands();
if(operands.size() == 2) {
if (references(operands.get(0), correlation)
&& operands.get(1) instanceof RexInputRef) {
// if call isn't EQUAL type and it has been determined that value generate might be
// required we should rather generate value generator
if(e.getKind() != SqlKind.EQUALS && (boolean)valueGen.peek()) {
return;
}
throw new Util.FoundOne(Pair.of(((RexInputRef) operands.get(1)).getIndex(),
Pair.of(((RexCall) e).getOperator(), true)));
}
if (references(operands.get(1), correlation)
&& operands.get(0) instanceof RexInputRef) {
// if call isn't EQUAL type and it has been determined that value generate might be
// required we should rather generate value generator
if(e.getKind() != SqlKind.EQUALS && (boolean)valueGen.peek()) {
return;
}
throw new Util.FoundOne(Pair.of(((RexInputRef) operands.get(0)).getIndex(),
Pair.of(((RexCall) e).getOperator(), false)));
}
break;
}
}
}
}
private boolean references(RexNode e, CorRef correlation) {
switch (e.getKind()) {
case CAST:
final RexNode operand = ((RexCall) e).getOperands().get(0);
if (isWidening(e.getType(), operand.getType())) {
return references(operand, correlation);
}
return false;
case FIELD_ACCESS:
final RexFieldAccess f = (RexFieldAccess) e;
if (f.getField().getIndex() == correlation.field
&& f.getReferenceExpr() instanceof RexCorrelVariable) {
if (((RexCorrelVariable) f.getReferenceExpr()).id.equals(correlation.corr)) {
return true;
}
}
// fall through
default:
return false;
}
}
/** Returns whether one type is just a widening of another.
*
* <p>For example:<ul>
* <li>{@code VARCHAR(10)} is a widening of {@code VARCHAR(5)}.
* <li>{@code VARCHAR(10)} is a widening of {@code VARCHAR(10) NOT NULL}.
* </ul>
*/
private boolean isWidening(RelDataType type, RelDataType type1) {
return type.getSqlTypeName() == type1.getSqlTypeName()
&& type.getPrecision() >= type1.getPrecision();
}
public Frame decorrelateRel(HiveFilter rel) {
return decorrelateRel((Filter) rel);
}
public Frame decorrelateRel(Filter rel) {
//
// Rewrite logic:
//
// 1. If a Filter references a correlated field in its filter
// condition, rewrite the Filter to be
// Filter
// Join(cross product)
// OriginalFilterInput
// ValueGenerator(produces distinct sets of correlated variables)
// and rewrite the correlated fieldAccess in the filter condition to
// reference the Join output.
//
// 2. If Filter does not reference correlated variables, simply
// rewrite the filter condition using new input.
//
final RelNode oldInput = rel.getInput();
Frame frame = getInvoke(oldInput, rel);
if (frame == null) {
// If input has not been rewritten, do not rewrite this rel.
return null;
}
Frame oldInputFrame = frame;
// If this Filter has correlated reference, create value generator
// and produce the correlated variables in the new output.
if (cm.mapRefRelToCorRef.containsKey(rel)) {
frame = decorrelateInputWithValueGenerator(rel);
}
boolean valueGenerator = true;
if(frame.r == oldInputFrame.r) {
// this means correated value generator wasn't generated
valueGenerator = false;
}
boolean isSemiJoin = false;
if (oldInput instanceof LogicalCorrelate) {
isSemiJoin = ((LogicalCorrelate) oldInput).getJoinType() == JoinRelType.SEMI ||
((LogicalCorrelate) oldInput).getJoinType() == JoinRelType.ANTI;
}
if(isSemiJoin && !cm.mapRefRelToCorRef.containsKey(rel)) {
// this conditions need to be pushed into semi-join since this condition
// corresponds to IN
Join join = ((Join)frame.r);
final List<RexNode> conditions = new ArrayList<>();
RexNode joinCond = join.getCondition();
conditions.add(joinCond);
conditions.add(decorrelateExpr(rel.getCondition(), valueGenerator));
final RexNode condition =
RexUtil.composeConjunction(rexBuilder, conditions, false);
RelNode newRel;
if (((LogicalCorrelate) oldInput).getJoinType() == JoinRelType.SEMI) {
newRel = HiveSemiJoin.getSemiJoin(frame.r.getCluster(), frame.r.getTraitSet(),
join.getLeft(), join.getRight(), condition);
} else {
newRel = HiveAntiJoin.getAntiJoin(frame.r.getCluster(), frame.r.getTraitSet(),
join.getLeft(), join.getRight(), condition);
}
return register(rel, newRel, frame.oldToNewOutputs, frame.corDefOutputs);
}
// Replace the filter expression to reference output of the join
// Map filter to the new filter over join
relBuilder.push(frame.r).filter(
(decorrelateExpr(rel.getCondition(), valueGenerator)));
// Filter does not change the input ordering.
// Filter rel does not permute the input.
// All corvars produced by filter will have the same output positions in the
// input rel.
return register(rel, relBuilder.build(), frame.oldToNewOutputs,
frame.corDefOutputs);
}
/**
* Rewrite Correlator into a left outer join.
*
* @param rel Correlator
*/
public Frame decorrelateRel(LogicalCorrelate rel) {
//
// Rewrite logic:
//
// The original left input will be joined with the new right input that
// has generated correlated variables propagated up. For any generated
// cor vars that are not used in the join key, pass them along to be
// joined later with the CorrelatorRels that produce them.
//
// the right input to Correlator should produce correlated variables
final RelNode oldLeft = rel.getInput(0);
final RelNode oldRight = rel.getInput(1);
boolean mightRequireValueGen = new findIfValueGenRequired().traverse(oldRight);
valueGen.push(mightRequireValueGen);
final Frame leftFrame = getInvoke(oldLeft, rel);
final Frame rightFrame = getInvoke(oldRight, rel);
if (leftFrame == null || rightFrame == null) {
// If any input has not been rewritten, do not rewrite this rel.
valueGen.pop();
return null;
}
if (rightFrame.corDefOutputs.isEmpty()) {
valueGen.pop();
return null;
}
assert rel.getRequiredColumns().cardinality()
<= rightFrame.corDefOutputs.keySet().size();
// Change correlator rel into a join.
// Join all the correlated variables produced by this correlator rel
// with the values generated and propagated from the right input
final SortedMap<CorDef, Integer> corDefOutputs =
new TreeMap<>(rightFrame.corDefOutputs);
final List<RexNode> conditions = new ArrayList<>();
final List<RelDataTypeField> newLeftOutput =
leftFrame.r.getRowType().getFieldList();
int newLeftFieldCount = newLeftOutput.size();
final List<RelDataTypeField> newRightOutput =
rightFrame.r.getRowType().getFieldList();
for (Map.Entry<CorDef, Integer> rightOutput
: new ArrayList<>(corDefOutputs.entrySet())) {
final CorDef corDef = rightOutput.getKey();
if (!corDef.corr.equals(rel.getCorrelationId())) {
continue;
}
final int newLeftPos = leftFrame.oldToNewOutputs.get(corDef.field);
final int newRightPos = rightOutput.getValue();
SqlOperator callOp = corDef.getPredicateKind() == null ?
SqlStdOperatorTable.EQUALS: corDef.getPredicateKind();
if(corDef.isLeft) {
conditions.add(
rexBuilder.makeCall(callOp,
RexInputRef.of(newLeftPos, newLeftOutput),
new RexInputRef(newLeftFieldCount + newRightPos,
newRightOutput.get(newRightPos).getType())));
} else {
conditions.add(
rexBuilder.makeCall(callOp,
new RexInputRef(newLeftFieldCount + newRightPos,
newRightOutput.get(newRightPos).getType()),
RexInputRef.of(newLeftPos, newLeftOutput)));
}
// remove this cor var from output position mapping
corDefOutputs.remove(corDef);
}
// Update the output position for the cor vars: only pass on the cor
// vars that are not used in the join key.
for (CorDef corDef : corDefOutputs.keySet()) {
int newPos = corDefOutputs.get(corDef) + newLeftFieldCount;
corDefOutputs.put(corDef, newPos);
}
// then add any cor var from the left input. Do not need to change
// output positions.
corDefOutputs.putAll(leftFrame.corDefOutputs);
// Create the mapping between the output of the old correlation rel
// and the new join rel
final Map<Integer, Integer> mapOldToNewOutputs = new HashMap<>();
int oldLeftFieldCount = oldLeft.getRowType().getFieldCount();
int oldRightFieldCount = oldRight.getRowType().getFieldCount();
// Left input positions are not changed.
mapOldToNewOutputs.putAll(leftFrame.oldToNewOutputs);
final RexNode condition =
RexUtil.composeConjunction(rexBuilder, conditions, false);
RelNode newJoin = null;
// this indicates original query was either correlated EXISTS or IN
if(rel.getJoinType() == JoinRelType.SEMI || rel.getJoinType() == JoinRelType.ANTI) {
final List<Integer> leftKeys = new ArrayList<Integer>();
final List<Integer> rightKeys = new ArrayList<Integer>();
RelNode[] inputRels = new RelNode[] {leftFrame.r, rightFrame.r};
if (rel.getJoinType() == JoinRelType.ANTI) {
newJoin = HiveAntiJoin.getAntiJoin(rel.getCluster(),
rel.getCluster().traitSetOf(HiveRelNode.CONVENTION), leftFrame.r, rightFrame.r, condition);
} else {
newJoin = HiveSemiJoin.getSemiJoin(rel.getCluster(),
rel.getCluster().traitSetOf(HiveRelNode.CONVENTION), leftFrame.r, rightFrame.r, condition);
}
} else {
// Right input positions are shifted by newLeftFieldCount.
for (int i = 0; i < oldRightFieldCount; i++) {
mapOldToNewOutputs.put(
i + oldLeftFieldCount,
rightFrame.oldToNewOutputs.get(i) + newLeftFieldCount);
}
newJoin = relBuilder.push(leftFrame.r).push(rightFrame.r)
.join(rel.getJoinType(), condition).build();
}
valueGen.pop();
return register(rel, newJoin, mapOldToNewOutputs, corDefOutputs);
}
public Frame decorrelateRel(HiveJoin rel) {
return decorrelateRel((Join) rel);
}
public Frame decorrelateRel(Join rel) {
// For SEMI/ANTI join decorrelate it's input directly,
// because the correlate variables can only be propagated from
// the left side, which is not supported yet.
if (!rel.getJoinType().projectsRight()) {
return decorrelateRel((RelNode) rel);
}
//
// Rewrite logic:
//
// 1. rewrite join condition.
// 2. map output positions and produce cor vars if any.
//
final RelNode oldLeft = rel.getInput(0);
final RelNode oldRight = rel.getInput(1);
final Frame leftFrame = getInvoke(oldLeft, rel);
final Frame rightFrame = getInvoke(oldRight, rel);
if (leftFrame == null || rightFrame == null) {
// If any input has not been rewritten, do not rewrite this rel.
return null;
}
final RelNode newJoin = HiveJoin.getJoin(rel.getCluster(), leftFrame.r, rightFrame.r,
decorrelateExpr(rel.getCondition()), rel.getJoinType());
// Create the mapping between the output of the old correlation rel
// and the new join rel
Map<Integer, Integer> mapOldToNewOutputs = Maps.newHashMap();
int oldLeftFieldCount = oldLeft.getRowType().getFieldCount();
int newLeftFieldCount = leftFrame.r.getRowType().getFieldCount();
int oldRightFieldCount = oldRight.getRowType().getFieldCount();
assert rel.getRowType().getFieldCount()
== oldLeftFieldCount + oldRightFieldCount;
// Left input positions are not changed.
mapOldToNewOutputs.putAll(leftFrame.oldToNewOutputs);
// Right input positions are shifted by newLeftFieldCount.
for (int i = 0; i < oldRightFieldCount; i++) {
mapOldToNewOutputs.put(i + oldLeftFieldCount,
rightFrame.oldToNewOutputs.get(i) + newLeftFieldCount);
}
final SortedMap<CorDef, Integer> corDefOutputs =
new TreeMap<>(leftFrame.corDefOutputs);
// Right input positions are shifted by newLeftFieldCount.
for (Map.Entry<CorDef, Integer> entry
: rightFrame.corDefOutputs.entrySet()) {
corDefOutputs.put(entry.getKey(),
entry.getValue() + newLeftFieldCount);
}
return register(rel, newJoin, mapOldToNewOutputs, corDefOutputs);
}
private RexInputRef getNewForOldInputRef(RexInputRef oldInputRef) {
assert currentRel != null;
int oldOrdinal = oldInputRef.getIndex();
int oldOrdinalNo = oldOrdinal;
int newOrdinal = 0;
// determine which input rel oldOrdinal references, and adjust
// oldOrdinal to be relative to that input rel
RelNode oldInput = null;
for (RelNode oldInput0 : currentRel.getInputs()) {
RelDataType oldInputType = oldInput0.getRowType();
int n = oldInputType.getFieldCount();
if (oldOrdinal < n) {
oldInput = oldInput0;
break;
}
RelNode newInput = map.get(oldInput0).r;
newOrdinal += newInput.getRowType().getFieldCount();
oldOrdinal -= n;
}
assert oldInput != null;
final Frame frame = map.get(oldInput);
assert frame != null;
// now oldOrdinal is relative to oldInput
int oldLocalOrdinal = oldOrdinal;
// figure out the newLocalOrdinal, relative to the newInput.
int newLocalOrdinal = oldLocalOrdinal;
if (!frame.oldToNewOutputs.isEmpty()) {
newLocalOrdinal = frame.oldToNewOutputs.get(oldLocalOrdinal);
}
newOrdinal += newLocalOrdinal;
return new RexInputRef(newOrdinal,
frame.r.getRowType().getFieldList().get(newLocalOrdinal).getType());
}
/**
* Pulls project above the join from its RHS input. Enforces nullability
* for join output.
*
* @param join Join
* @param project Original project as the right-hand input of the join
* @param nullIndicatorPos Position of null indicator
* @return the subtree with the new Project at the root
*/
private RelNode projectJoinOutputWithNullability(
Join join,
Project project,
int nullIndicatorPos) {
final RelDataTypeFactory typeFactory = join.getCluster().getTypeFactory();
final RelNode left = join.getLeft();
final JoinRelType joinType = join.getJoinType();
RexInputRef nullIndicator =
new RexInputRef(
nullIndicatorPos,
typeFactory.createTypeWithNullability(
join.getRowType().getFieldList().get(nullIndicatorPos)
.getType(),
true));
// now create the new project
List<Pair<RexNode, String>> newProjExprs = Lists.newArrayList();
// project everything from the LHS and then those from the original
// projRel
List<RelDataTypeField> leftInputFields =
left.getRowType().getFieldList();
for (int i = 0; i < leftInputFields.size(); i++) {
newProjExprs.add(RexInputRef.of2(i, leftInputFields));
}
// Marked where the projected expr is coming from so that the types will
// become nullable for the original projections which are now coming out
// of the nullable side of the OJ.
boolean projectPulledAboveLeftCorrelator =
joinType.generatesNullsOnRight();
for (Pair<RexNode, String> pair : project.getNamedProjects()) {
RexNode newProjExpr =
removeCorrelationExpr(
pair.left,
projectPulledAboveLeftCorrelator,
nullIndicator);
newProjExprs.add(Pair.of(newProjExpr, pair.right));
}
return RelOptUtil.createProject(join, Pair.left(newProjExprs), Pair.right(newProjExprs),
false, relBuilder);
}
/**
* Pulls a {@link Project} above a {@link Correlate} from its RHS input.
* Enforces nullability for join output.
*
* @param correlate Correlate
* @param project the original project as the RHS input of the join
* @param isCount Positions which are calls to the <code>COUNT</code>
* aggregation function
* @return the subtree with the new Project at the root
*/
private RelNode aggregateCorrelatorOutput(
Correlate correlate,
Project project,
Set<Integer> isCount) {
final RelNode left = correlate.getLeft();
final JoinRelType joinType = correlate.getJoinType();
// now create the new project
final List<Pair<RexNode, String>> newProjects = Lists.newArrayList();
// Project everything from the LHS and then those from the original
// project
final List<RelDataTypeField> leftInputFields =
left.getRowType().getFieldList();
for (int i = 0; i < leftInputFields.size(); i++) {
newProjects.add(RexInputRef.of2(i, leftInputFields));
}
// Marked where the projected expr is coming from so that the types will
// become nullable for the original projections which are now coming out
// of the nullable side of the OJ.
boolean projectPulledAboveLeftCorrelator =
joinType.generatesNullsOnRight();
for (Pair<RexNode, String> pair : project.getNamedProjects()) {
RexNode newProjExpr =
removeCorrelationExpr(
pair.left,
projectPulledAboveLeftCorrelator,
isCount);
newProjects.add(Pair.of(newProjExpr, pair.right));
}
return RelOptUtil.createProject(correlate, Pair.left(newProjects), Pair.right(newProjects),
false, relBuilder);
}
/**
* Checks whether the correlations in projRel and filter are related to
* the correlated variables provided by corRel.
*
* @param correlate Correlate
* @param project The original Project as the RHS input of the join
* @param filter Filter
* @param correlatedJoinKeys Correlated join keys
* @return true if filter and proj only references corVar provided by corRel
*/
private boolean checkCorVars(
LogicalCorrelate correlate,
Project project,
Filter filter,
List<RexFieldAccess> correlatedJoinKeys) {
if (filter != null) {
assert correlatedJoinKeys != null;
// check that all correlated refs in the filter condition are
// used in the join(as field access).
Set<CorRef> corVarInFilter =
Sets.newHashSet(cm.mapRefRelToCorRef.get(filter));
for (RexFieldAccess correlatedJoinKey : correlatedJoinKeys) {
corVarInFilter.remove(cm.mapFieldAccessToCorRef.get(correlatedJoinKey));
}
if (!corVarInFilter.isEmpty()) {
return false;
}
// Check that the correlated variables referenced in these
// comparisons do come from the correlatorRel.
corVarInFilter.addAll(cm.mapRefRelToCorRef.get(filter));
for (CorRef corVar : corVarInFilter) {
if (cm.mapCorToCorRel.get(corVar.corr) != correlate) {
return false;
}
}
}
// if project has any correlated reference, make sure they are also
// provided by the current correlate. They will be projected out of the LHS
// of the correlate.
if ((project != null) && cm.mapRefRelToCorRef.containsKey(project)) {
for (CorRef corVar : cm.mapRefRelToCorRef.get(project)) {
if (cm.mapCorToCorRel.get(corVar.corr) != correlate) {
return false;
}
}
}
return true;
}
/**
* Remove correlated variables from the tree at root corRel.
*
* @param correlate Correlator
*/
private void removeCorVarFromTree(LogicalCorrelate correlate) {
if (cm.mapCorToCorRel.get(correlate.getCorrelationId()) == correlate) {
cm.mapCorToCorRel.remove(correlate.getCorrelationId());
}
}
/**
* Projects all {@code input} output fields plus the additional expressions.
*
* @param input Input relational expression
* @param additionalExprs Additional expressions and names
* @return the new Project
*/
private RelNode createProjectWithAdditionalExprs(
RelNode input,
List<Pair<RexNode, String>> additionalExprs) {
final List<RelDataTypeField> fieldList =
input.getRowType().getFieldList();
List<Pair<RexNode, String>> projects = Lists.newArrayList();
for (Ord<RelDataTypeField> field : Ord.zip(fieldList)) {
projects.add(
Pair.of(
(RexNode) rexBuilder.makeInputRef(
field.e.getType(), field.i),
field.e.getName()));
}
projects.addAll(additionalExprs);
return RelOptUtil.createProject(input, Pair.left(projects), Pair.right(projects),
false, relBuilder);
}
/* Returns an immutable map with the identity [0: 0, .., count-1: count-1]. */
static Map<Integer, Integer> identityMap(int count) {
ImmutableMap.Builder<Integer, Integer> builder = ImmutableMap.builder();
for (int i = 0; i < count; i++) {
builder.put(i, i);
}
return builder.build();
}
/** Registers a relational expression and the relational expression it became
* after decorrelation. */
Frame register(RelNode rel, RelNode newRel,
Map<Integer, Integer> oldToNewOutputs,
SortedMap<CorDef, Integer> corDefOutputs) {
final Frame frame = new Frame(rel, newRel, corDefOutputs, oldToNewOutputs);
map.put(rel, frame);
return frame;
}
static boolean allLessThan(Collection<Integer> integers, int limit,
Litmus ret) {
for (int value : integers) {
if (value >= limit) {
return ret.fail("out of range; value: " + value + ", limit: " + limit);
}
}
return ret.succeed();
}
private static RelNode stripHep(RelNode rel) {
if (rel instanceof HepRelVertex) {
HepRelVertex hepRelVertex = (HepRelVertex) rel;
rel = hepRelVertex.getCurrentRel();
}
return rel;
}
//~ Inner Classes ----------------------------------------------------------
/** Shuttle that decorrelates. */
private class DecorrelateRexShuttle extends RexShuttle {
private boolean valueGenerator;
public void setValueGenerator(boolean valueGenerator) {
this.valueGenerator = valueGenerator;
}
@Override public RexNode visitCall(final RexCall call) {
if (Bug.CALCITE_1851_FIXED) {
throw new AssertionError("Overriding should be removed");
}
if (valueGenerator) {
return super.visitCall(call);
}
switch (call.getKind()) {
case EQUALS:
case NOT_EQUALS:
case GREATER_THAN:
case GREATER_THAN_OR_EQUAL:
case LESS_THAN:
case LESS_THAN_OR_EQUAL:
// Expressions of the form OP(cor.col1, $4) are transformed to OP($4, $4) where OP in (<,<=,=,>,>=,<>).
// OP($4, $4) can be further simplified to TRUE/FALSE (e.g., $4 <> $4) that is wrong and messes up the tree.
// To prevent this simplifications we rewrite OP($4, $4) to IS_NOT_NULL($4).
// This should be done only for correlated predicates and not user specified explicit predicates
if (hasCorrelation(call)) {
List<RexNode> ops = visitList(call.operands, new boolean[] { false });
assert ops.size() == 2;
RexNode o0 = ops.get(0);
RexNode o1 = ops.get(1);
if (o0.equals(o1)) {
return relBuilder.call(SqlStdOperatorTable.IS_NOT_NULL, o0);
} else {
return relBuilder.call(call.getOperator(), ops);
}
}
default:
return super.visitCall(call);
}
}
private boolean hasCorrelation(RexCall call) {
for (RexNode n : call.operands) {
if (n instanceof RexFieldAccess && cm.mapFieldAccessToCorRef.containsKey(n)) {
return true;
}
}
return false;
}
@Override public RexNode visitFieldAccess(RexFieldAccess fieldAccess) {
return decorrFieldAccess(fieldAccess);
}
@Override public RexNode visitInputRef(RexInputRef inputRef) {
final RexInputRef ref = getNewForOldInputRef(inputRef);
if (ref.getIndex() == inputRef.getIndex()
&& ref.getType() == inputRef.getType()) {
return inputRef; // re-use old object, to prevent needless expr cloning
}
return ref;
}
private RexNode decorrFieldAccess(RexFieldAccess fieldAccess) {
int newInputOutputOffset = 0;
for (RelNode input : currentRel.getInputs()) {
final Frame frame = map.get(input);
if (frame != null) {
// try to find in this input rel the position of cor var
final CorRef corRef = cm.mapFieldAccessToCorRef.get(fieldAccess);
if (corRef != null) {
Integer newInputPos = frame.corDefOutputs.get(corRef.def());
if (newInputPos != null) {
// This input rel does produce the cor var referenced.
// Assume fieldAccess has the correct type info.
return new RexInputRef(newInputPos + newInputOutputOffset,
frame.r.getRowType().getFieldList().get(newInputPos)
.getType());
}
}
// this input rel does not produce the cor var needed
newInputOutputOffset += frame.r.getRowType().getFieldCount();
} else {
// this input rel is not rewritten
newInputOutputOffset += input.getRowType().getFieldCount();
}
}
return fieldAccess;
}
}
/** Shuttle that removes correlations. */
private class RemoveCorrelationRexShuttle extends RexShuttle {
private final RexBuilder rexBuilder;
private final RelDataTypeFactory typeFactory;
private final boolean projectPulledAboveLeftCorrelator;
private final RexInputRef nullIndicator;
private final ImmutableSet<Integer> isCount;
RemoveCorrelationRexShuttle(
RexBuilder rexBuilder,
boolean projectPulledAboveLeftCorrelator,
RexInputRef nullIndicator,
Set<Integer> isCount) {
this.projectPulledAboveLeftCorrelator =
projectPulledAboveLeftCorrelator;
this.nullIndicator = nullIndicator; // may be null
this.isCount = ImmutableSet.copyOf(isCount);
this.rexBuilder = rexBuilder;
this.typeFactory = rexBuilder.getTypeFactory();
}
private RexNode createCaseExpression(
RexInputRef nullInputRef,
RexLiteral lit,
RexNode rexNode) {
RexNode[] caseOperands = new RexNode[3];
// Construct a CASE expression to handle the null indicator.
//
// This also covers the case where a left correlated subquery
// projects fields from outer relation. Since LOJ cannot produce
// nulls on the LHS, the projection now need to make a nullable LHS
// reference using a nullability indicator. If this this indicator
// is null, it means the subquery does not produce any value. As a
// result, any RHS ref by this usbquery needs to produce null value.
// WHEN indicator IS NULL
caseOperands[0] =
rexBuilder.makeCall(
SqlStdOperatorTable.IS_NULL,
new RexInputRef(
nullInputRef.getIndex(),
typeFactory.createTypeWithNullability(
nullInputRef.getType(),
true)));
// THEN CAST(NULL AS newInputTypeNullable)
caseOperands[1] =
rexBuilder.makeCast(
typeFactory.createTypeWithNullability(
rexNode.getType(),
true),
lit);
// ELSE cast (newInput AS newInputTypeNullable) END
caseOperands[2] =
rexBuilder.makeCast(
typeFactory.createTypeWithNullability(
rexNode.getType(),
true),
rexNode);
return rexBuilder.makeCall(
SqlStdOperatorTable.CASE,
caseOperands);
}
@Override public RexNode visitFieldAccess(RexFieldAccess fieldAccess) {
if (cm.mapFieldAccessToCorRef.containsKey(fieldAccess)) {
// if it is a corVar, change it to be input ref.
CorRef corVar = cm.mapFieldAccessToCorRef.get(fieldAccess);
// corVar offset should point to the leftInput of currentRel,
// which is the Correlator.
RexNode newRexNode =
new RexInputRef(corVar.field, fieldAccess.getType());
if (projectPulledAboveLeftCorrelator
&& (nullIndicator != null)) {
// need to enforce nullability by applying an additional
// cast operator over the transformed expression.
newRexNode =
createCaseExpression(
nullIndicator,
rexBuilder.constantNull(),
newRexNode);
}
return newRexNode;
}
return fieldAccess;
}
@Override public RexNode visitInputRef(RexInputRef inputRef) {
if (currentRel instanceof LogicalCorrelate) {
// if this rel references corVar
// and now it needs to be rewritten
// it must have been pulled above the Correlator
// replace the input ref to account for the LHS of the
// Correlator
final int leftInputFieldCount =
((LogicalCorrelate) currentRel).getLeft().getRowType()
.getFieldCount();
RelDataType newType = inputRef.getType();
if (projectPulledAboveLeftCorrelator) {
newType =
typeFactory.createTypeWithNullability(newType, true);
}
int pos = inputRef.getIndex();
RexInputRef newInputRef =
new RexInputRef(leftInputFieldCount + pos, newType);
if ((isCount != null) && isCount.contains(pos)) {
return createCaseExpression(
newInputRef,
rexBuilder.makeExactLiteral(BigDecimal.ZERO),
newInputRef);
} else {
return newInputRef;
}
}
return inputRef;
}
@Override public RexNode visitLiteral(RexLiteral literal) {
// Use nullIndicator to decide whether to project null.
// Do nothing if the literal is null.
if (!RexUtil.isNull(literal)
&& projectPulledAboveLeftCorrelator
&& (nullIndicator != null)) {
return createCaseExpression(
nullIndicator,
rexBuilder.constantNull(),
literal);
}
return literal;
}
@Override public RexNode visitCall(final RexCall call) {
RexNode newCall;
boolean[] update = {false};
List<RexNode> clonedOperands = visitList(call.operands, update);
if (update[0]) {
SqlOperator operator = call.getOperator();
boolean isSpecialCast = false;
if (operator instanceof SqlFunction) {
SqlFunction function = (SqlFunction) operator;
if (function.getKind() == SqlKind.CAST) {
if (call.operands.size() < 2) {
isSpecialCast = true;
}
}
}
if (!isSpecialCast) {
// TODO: ideally this only needs to be called if the result
// type will also change. However, since that requires
// support from type inference rules to tell whether a rule
// decides return type based on input types, for now all
// operators will be recreated with new type if any operand
// changed, unless the operator has "built-in" type.
newCall = rexBuilder.ensureType(
call.getType(),
rexBuilder.makeCall(
rexBuilder.deriveReturnType(operator, clonedOperands), operator, clonedOperands),
true);
} else {
// Use the current return type when creating a new call, for
// operators with return type built into the operator
// definition, and with no type inference rules, such as
// cast function with less than 2 operands.
// TODO: Comments in RexShuttle.visitCall() mention other
// types in this category. Need to resolve those together
// and preferably in the base class RexShuttle.
newCall =
rexBuilder.makeCall(
call.getType(),
operator,
clonedOperands);
}
} else {
newCall = call;
}
if (projectPulledAboveLeftCorrelator && (nullIndicator != null)) {
return createCaseExpression(
nullIndicator,
rexBuilder.constantNull(),
newCall);
}
return newCall;
}
}
/**
* Rule to remove single_value rel. For cases like
*
* <blockquote>AggRel single_value proj/filter/agg/ join on unique LHS key
* AggRel single group</blockquote>
*/
private final class RemoveSingleAggregateRule extends RelOptRule {
RemoveSingleAggregateRule() {
super(
operand(
Aggregate.class,
operand(
Project.class,
operand(Aggregate.class, any()))));
}
@Override
public void onMatch(RelOptRuleCall call) {
Aggregate singleAggregate = call.rel(0);
Project project = call.rel(1);
Aggregate aggregate = call.rel(2);
// check singleAggRel is single_value agg
if ((!singleAggregate.getGroupSet().isEmpty())
|| (singleAggregate.getAggCallList().size() != 1)
|| !(singleAggregate.getAggCallList().get(0).getAggregation()
instanceof SqlSingleValueAggFunction)) {
return;
}
// check projRel only projects one expression
// check this project only projects one expression, i.e. scalar
// subqueries.
List<RexNode> projExprs = project.getProjects();
if (projExprs.size() != 1) {
return;
}
// check the input to projRel is an aggregate on the entire input
if (!aggregate.getGroupSet().isEmpty()) {
return;
}
// singleAggRel produces a nullable type, so create the new
// projection that casts proj expr to a nullable type.
final RelOptCluster cluster = project.getCluster();
RelNode newProject = RelOptUtil.createProject(aggregate,
ImmutableList.of(rexBuilder.makeCast(
cluster.getTypeFactory().createTypeWithNullability(projExprs.get(0).getType(), true),
projExprs.get(0))),
null, false, relBuilder);
call.transformTo(newProject);
}
}
/** Planner rule that removes correlations for scalar projects. */
private final class RemoveCorrelationForScalarProjectRule extends RelOptRule {
RemoveCorrelationForScalarProjectRule() {
super(
operand(LogicalCorrelate.class,
operand(RelNode.class, any()),
operand(Aggregate.class,
operand(Project.class,
operand(RelNode.class, any())))));
}
@Override
public void onMatch(RelOptRuleCall call) {
final LogicalCorrelate correlate = call.rel(0);
final RelNode left = call.rel(1);
final Aggregate aggregate = call.rel(2);
final Project project = call.rel(3);
RelNode right = call.rel(4);
final RelOptCluster cluster = correlate.getCluster();
setCurrent(call.getPlanner().getRoot(), correlate);
// Check for this pattern.
// The pattern matching could be simplified if rules can be applied
// during decorrelation.
//
// CorrelateRel(left correlation, condition = true)
// LeftInputRel
// Aggregate (groupby (0) single_value())
// Project-A (may reference coVar)
// RightInputRel
if(correlate.getJoinType() != JoinRelType.LEFT) {
return;
}
final JoinRelType joinType = correlate.getJoinType();
// corRel.getCondition was here, however Correlate was updated so it
// never includes a join condition. The code was not modified for brevity.
RexNode joinCond = rexBuilder.makeLiteral(true);
if ((joinType != JoinRelType.LEFT)
|| (joinCond != rexBuilder.makeLiteral(true))) {
return;
}
// check that the agg is of the following type:
// doing a single_value() on the entire input
if ((!aggregate.getGroupSet().isEmpty())
|| (aggregate.getAggCallList().size() != 1)
|| !(aggregate.getAggCallList().get(0).getAggregation()
instanceof SqlSingleValueAggFunction)) {
return;
}
// check this project only projects one expression, i.e. scalar
// subqueries.
if (project.getProjects().size() != 1) {
return;
}
int nullIndicatorPos;
if ((right instanceof Filter)
&& cm.mapRefRelToCorRef.containsKey(right)) {
// rightInputRel has this shape:
//
// Filter (references corvar)
// FilterInputRel
// If rightInputRel is a filter and contains correlated
// reference, make sure the correlated keys in the filter
// condition forms a unique key of the RHS.
Filter filter = (Filter) right;
right = filter.getInput();
assert right instanceof HepRelVertex;
right = ((HepRelVertex) right).getCurrentRel();
// check filter input contains no correlation
if (RelOptUtil.getVariablesUsed(right).size() > 0) {
return;
}
// extract the correlation out of the filter
// First breaking up the filter conditions into equality
// comparisons between rightJoinKeys(from the original
// filterInputRel) and correlatedJoinKeys. correlatedJoinKeys
// can be expressions, while rightJoinKeys need to be input
// refs. These comparisons are AND'ed together.
List<RexNode> tmpRightJoinKeys = Lists.newArrayList();
List<RexNode> correlatedJoinKeys = Lists.newArrayList();
HiveRelOptUtil.splitCorrelatedFilterCondition(
filter,
tmpRightJoinKeys,
correlatedJoinKeys,
false);
// check that the columns referenced in these comparisons form
// an unique key of the filterInputRel
final List<RexInputRef> rightJoinKeys = new ArrayList<>();
for (RexNode key : tmpRightJoinKeys) {
assert key instanceof RexInputRef;
rightJoinKeys.add((RexInputRef) key);
}
// check that the columns referenced in rightJoinKeys form an
// unique key of the filterInputRel
if (rightJoinKeys.isEmpty()) {
return;
}
// The join filters out the nulls. So, it's ok if there are
// nulls in the join keys.
final RelMetadataQuery mq = call.getMetadataQuery();
if (!RelMdUtil.areColumnsDefinitelyUniqueWhenNullsFiltered(mq, right,
rightJoinKeys)) {
//SQL2REL_LOGGER.fine(rightJoinKeys.toString()
// + "are not unique keys for "
// + right.toString());
return;
}
RexUtil.FieldAccessFinder visitor =
new RexUtil.FieldAccessFinder();
RexUtil.apply(visitor, correlatedJoinKeys, null);
List<RexFieldAccess> correlatedKeyList =
visitor.getFieldAccessList();
if (!checkCorVars(correlate, project, filter, correlatedKeyList)) {
return;
}
// Change the plan to this structure.
// Note that the aggregateRel is removed.
//
// Project-A' (replace corvar to input ref from the Join)
// Join (replace corvar to input ref from LeftInputRel)
// LeftInputRel
// RightInputRel(oreviously FilterInputRel)
// Change the filter condition into a join condition
joinCond =
removeCorrelationExpr(filter.getCondition(), false);
nullIndicatorPos =
left.getRowType().getFieldCount()
+ rightJoinKeys.get(0).getIndex();
} else if (cm.mapRefRelToCorRef.containsKey(project)) {
// check filter input contains no correlation
if (RelOptUtil.getVariablesUsed(right).size() > 0) {
return;
}
if (!checkCorVars(correlate, project, null, null)) {
return;
}
// Change the plan to this structure.
//
// Project-A' (replace corvar to input ref from Join)
// Join (left, condition = true)
// LeftInputRel
// Aggregate(groupby(0), single_value(0), s_v(1)....)
// Project-B (everything from input plus literal true)
// ProjInputRel
// make the new projRel to provide a null indicator
right =
createProjectWithAdditionalExprs(right,
ImmutableList.of(
Pair.<RexNode, String>of(
rexBuilder.makeLiteral(true), "nullIndicator")));
// make the new aggRel
right =
HiveRelOptUtil.createSingleValueAggRel(cluster, right, HiveRelFactories.HIVE_AGGREGATE_FACTORY);
// The last field:
// single_value(true)
// is the nullIndicator
nullIndicatorPos =
left.getRowType().getFieldCount()
+ right.getRowType().getFieldCount() - 1;
} else {
return;
}
// make the new join rel
Join join = (Join) relBuilder.push(left).push(right)
.join(joinType, joinCond).build();
RelNode newProject =
projectJoinOutputWithNullability(join, project, nullIndicatorPos);
call.transformTo(newProject);
removeCorVarFromTree(correlate);
}
}
/** Planner rule that removes correlations for scalar aggregates. */
private final class RemoveCorrelationForScalarAggregateRule
extends RelOptRule {
RemoveCorrelationForScalarAggregateRule() {
super(
operand(LogicalCorrelate.class,
operand(RelNode.class, any()),
operand(Project.class,
operandJ(Aggregate.class, null, Aggregate::isSimple,
operand(Project.class,
operand(RelNode.class, any()))))));
}
@Override
public void onMatch(RelOptRuleCall call) {
final LogicalCorrelate correlate = call.rel(0);
final RelNode left = call.rel(1);
final Project aggOutputProject = call.rel(2);
final Aggregate aggregate = call.rel(3);
final Project aggInputProject = call.rel(4);
RelNode right = call.rel(5);
final RelOptCluster cluster = correlate.getCluster();
setCurrent(call.getPlanner().getRoot(), correlate);
// check for this pattern
// The pattern matching could be simplified if rules can be applied
// during decorrelation,
//
// CorrelateRel(left correlation, condition = true)
// LeftInputRel
// Project-A (a RexNode)
// Aggregate (groupby (0), agg0(), agg1()...)
// Project-B (references coVar)
// rightInputRel
// check aggOutputProject projects only one expression
final List<RexNode> aggOutputProjects = aggOutputProject.getProjects();
if (aggOutputProjects.size() != 1) {
return;
}
if(correlate.getJoinType() != JoinRelType.LEFT) {
return;
}
final JoinRelType joinType = correlate.getJoinType();
// corRel.getCondition was here, however Correlate was updated so it
// never includes a join condition. The code was not modified for brevity.
RexNode joinCond = rexBuilder.makeLiteral(true);
if ((joinType != JoinRelType.LEFT)
|| (joinCond != rexBuilder.makeLiteral(true))) {
return;
}
// check that the agg is on the entire input
if (!aggregate.getGroupSet().isEmpty()) {
return;
}
final List<RexNode> aggInputProjects = aggInputProject.getProjects();
final List<AggregateCall> aggCalls = aggregate.getAggCallList();
final Set<Integer> isCountStar = Sets.newHashSet();
// mark if agg produces count(*) which needs to reference the
// nullIndicator after the transformation.
int k = -1;
for (AggregateCall aggCall : aggCalls) {
++k;
if ((aggCall.getAggregation() instanceof SqlCountAggFunction)
&& (aggCall.getArgList().size() == 0)) {
isCountStar.add(k);
}
}
if ((right instanceof Filter)
&& cm.mapRefRelToCorRef.containsKey(right)) {
// rightInputRel has this shape:
//
// Filter (references corvar)
// FilterInputRel
Filter filter = (Filter) right;
right = filter.getInput();
assert right instanceof HepRelVertex;
right = ((HepRelVertex) right).getCurrentRel();
// check filter input contains no correlation
if (RelOptUtil.getVariablesUsed(right).size() > 0) {
return;
}
// check filter condition type First extract the correlation out
// of the filter
// First breaking up the filter conditions into equality
// comparisons between rightJoinKeys(from the original
// filterInputRel) and correlatedJoinKeys. correlatedJoinKeys
// can only be RexFieldAccess, while rightJoinKeys can be
// expressions. These comparisons are AND'ed together.
List<RexNode> rightJoinKeys = Lists.newArrayList();
List<RexNode> tmpCorrelatedJoinKeys = Lists.newArrayList();
HiveRelOptUtil.splitCorrelatedFilterCondition(
filter,
rightJoinKeys,
tmpCorrelatedJoinKeys,
true);
// make sure the correlated reference forms a unique key check
// that the columns referenced in these comparisons form an
// unique key of the leftInputRel
List<RexFieldAccess> correlatedJoinKeys = Lists.newArrayList();
List<RexInputRef> correlatedInputRefJoinKeys = Lists.newArrayList();
for (RexNode joinKey : tmpCorrelatedJoinKeys) {
assert joinKey instanceof RexFieldAccess;
correlatedJoinKeys.add((RexFieldAccess) joinKey);
RexNode correlatedInputRef =
removeCorrelationExpr(joinKey, false);
assert correlatedInputRef instanceof RexInputRef;
correlatedInputRefJoinKeys.add(
(RexInputRef) correlatedInputRef);
}
// check that the columns referenced in rightJoinKeys form an
// unique key of the filterInputRel
if (correlatedInputRefJoinKeys.isEmpty()) {
return;
}
// The join filters out the nulls. So, it's ok if there are
// nulls in the join keys.
final RelMetadataQuery mq = call.getMetadataQuery();
if (!RelMdUtil.areColumnsDefinitelyUniqueWhenNullsFiltered(mq, left,
correlatedInputRefJoinKeys)) {
//SQL2REL_LOGGER.fine(correlatedJoinKeys.toString()
// + "are not unique keys for "
// + left.toString());
return;
}
// check cor var references are valid
if (!checkCorVars(correlate,
aggInputProject,
filter,
correlatedJoinKeys)) {
return;
}
// Rewrite the above plan:
//
// CorrelateRel(left correlation, condition = true)
// LeftInputRel
// Project-A (a RexNode)
// Aggregate (groupby(0), agg0(),agg1()...)
// Project-B (may reference coVar)
// Filter (references corVar)
// RightInputRel (no correlated reference)
//
// to this plan:
//
// Project-A' (all gby keys + rewritten nullable ProjExpr)
// Aggregate (groupby(all left input refs)
// agg0(rewritten expression),
// agg1()...)
// Project-B' (rewriten original projected exprs)
// Join(replace corvar w/ input ref from LeftInputRel)
// LeftInputRel
// RightInputRel
//
// In the case where agg is count(*) or count($corVar), it is
// changed to count(nullIndicator).
// Note: any non-nullable field from the RHS can be used as
// the indicator however a "true" field is added to the
// projection list from the RHS for simplicity to avoid
// searching for non-null fields.
//
// Project-A' (all gby keys + rewritten nullable ProjExpr)
// Aggregate (groupby(all left input refs),
// count(nullIndicator), other aggs...)
// Project-B' (all left input refs plus
// the rewritten original projected exprs)
// Join(replace corvar to input ref from LeftInputRel)
// LeftInputRel
// Project (everything from RightInputRel plus
// the nullIndicator "true")
// RightInputRel
//
// first change the filter condition into a join condition
joinCond =
removeCorrelationExpr(filter.getCondition(), false);
} else if (cm.mapRefRelToCorRef.containsKey(aggInputProject)) {
// check rightInputRel contains no correlation
if (RelOptUtil.getVariablesUsed(right).size() > 0) {
return;
}
// check cor var references are valid
if (!checkCorVars(correlate, aggInputProject, null, null)) {
return;
}
int nFields = left.getRowType().getFieldCount();
ImmutableBitSet allCols = ImmutableBitSet.range(nFields);
// leftInputRel contains unique keys
// i.e. each row is distinct and can group by on all the left
// fields
final RelMetadataQuery mq = call.getMetadataQuery();
if (!RelMdUtil.areColumnsDefinitelyUnique(mq, left, allCols)) {
//SQL2REL_LOGGER.fine("There are no unique keys for " + left);
return;
}
//
// Rewrite the above plan:
//
// CorrelateRel(left correlation, condition = true)
// LeftInputRel
// Project-A (a RexNode)
// Aggregate (groupby(0), agg0(), agg1()...)
// Project-B (references coVar)
// RightInputRel (no correlated reference)
//
// to this plan:
//
// Project-A' (all gby keys + rewritten nullable ProjExpr)
// Aggregate (groupby(all left input refs)
// agg0(rewritten expression),
// agg1()...)
// Project-B' (rewriten original projected exprs)
// Join (LOJ cond = true)
// LeftInputRel
// RightInputRel
//
// In the case where agg is count($corVar), it is changed to
// count(nullIndicator).
// Note: any non-nullable field from the RHS can be used as
// the indicator however a "true" field is added to the
// projection list from the RHS for simplicity to avoid
// searching for non-null fields.
//
// Project-A' (all gby keys + rewritten nullable ProjExpr)
// Aggregate (groupby(all left input refs),
// count(nullIndicator), other aggs...)
// Project-B' (all left input refs plus
// the rewritten original projected exprs)
// Join(replace corvar to input ref from LeftInputRel)
// LeftInputRel
// Project (everything from RightInputRel plus
// the nullIndicator "true")
// RightInputRel
} else {
return;
}
RelDataType leftInputFieldType = left.getRowType();
int leftInputFieldCount = leftInputFieldType.getFieldCount();
int joinOutputProjExprCount =
leftInputFieldCount + aggInputProjects.size() + 1;
right =
createProjectWithAdditionalExprs(right,
ImmutableList.of(
Pair.<RexNode, String>of(rexBuilder.makeLiteral(true),
"nullIndicator")));
Join join = (Join) relBuilder.push(left).push(right)
.join(joinType, joinCond).build();
// To the consumer of joinOutputProjRel, nullIndicator is located
// at the end
int nullIndicatorPos = join.getRowType().getFieldCount() - 1;
RexInputRef nullIndicator =
new RexInputRef(
nullIndicatorPos,
cluster.getTypeFactory().createTypeWithNullability(
join.getRowType().getFieldList()
.get(nullIndicatorPos).getType(),
true));
// first project all group-by keys plus the transformed agg input
List<RexNode> joinOutputProjects = Lists.newArrayList();
// LOJ Join preserves LHS types
for (int i = 0; i < leftInputFieldCount; i++) {
joinOutputProjects.add(
rexBuilder.makeInputRef(
leftInputFieldType.getFieldList().get(i).getType(), i));
}
for (RexNode aggInputProjExpr : aggInputProjects) {
joinOutputProjects.add(
removeCorrelationExpr(aggInputProjExpr,
joinType.generatesNullsOnRight(),
nullIndicator));
}
joinOutputProjects.add(
rexBuilder.makeInputRef(join, nullIndicatorPos));
RelNode joinOutputProject = RelOptUtil.createProject(
join, joinOutputProjects, null, false, relBuilder);
// nullIndicator is now at a different location in the output of
// the join
nullIndicatorPos = joinOutputProjExprCount - 1;
final int groupCount = leftInputFieldCount;
List<AggregateCall> newAggCalls = Lists.newArrayList();
k = -1;
for (AggregateCall aggCall : aggCalls) {
++k;
final List<Integer> argList;
if (isCountStar.contains(k)) {
// this is a count(*), transform it to count(nullIndicator)
// the null indicator is located at the end
argList = Collections.singletonList(nullIndicatorPos);
} else {
argList = Lists.newArrayList();
for (int aggArg : aggCall.getArgList()) {
argList.add(aggArg + groupCount);
}
}
int filterArg = aggCall.filterArg < 0 ? aggCall.filterArg
: aggCall.filterArg + groupCount;
newAggCalls.add(
aggCall.adaptTo(joinOutputProject, argList, filterArg,
aggregate.getGroupCount(), groupCount));
}
ImmutableBitSet groupSet =
ImmutableBitSet.range(groupCount);
Aggregate newAggregate =
(Aggregate) relBuilder.push(joinOutputProject)
.aggregate(relBuilder.groupKey(groupSet, null), newAggCalls).build();
List<RexNode> newAggOutputProjectList = Lists.newArrayList();
for (int i : groupSet) {
newAggOutputProjectList.add(
rexBuilder.makeInputRef(newAggregate, i));
}
RexNode newAggOutputProjects =
removeCorrelationExpr(aggOutputProjects.get(0), false);
newAggOutputProjectList.add(
rexBuilder.makeCast(
cluster.getTypeFactory().createTypeWithNullability(
newAggOutputProjects.getType(),
true),
newAggOutputProjects));
RelNode newAggOutputProject = RelOptUtil.createProject(
newAggregate, newAggOutputProjectList, null, false, relBuilder);
call.transformTo(newAggOutputProject);
removeCorVarFromTree(correlate);
}
}
// REVIEW jhyde 29-Oct-2007: This rule is non-static, depends on the state
// of members in RelDecorrelator, and has side-effects in the decorrelator.
// This breaks the contract of a planner rule, and the rule will not be
// reusable in other planners.
// REVIEW jvs 29-Oct-2007: Shouldn't it also be incorporating
// the flavor attribute into the description?
/** Planner rule that adjusts projects when counts are added. */
private final class AdjustProjectForCountAggregateRule extends RelOptRule {
private final boolean flavor;
AdjustProjectForCountAggregateRule(boolean flavor) {
super(
flavor
? operand(LogicalCorrelate.class,
operand(RelNode.class, any()),
operand(Project.class,
operand(Aggregate.class, any())))
: operand(LogicalCorrelate.class,
operand(RelNode.class, any()),
operand(Aggregate.class, any())));
this.flavor = flavor;
}
@Override
public void onMatch(RelOptRuleCall call) {
final LogicalCorrelate correlate = call.rel(0);
final RelNode left = call.rel(1);
final Project aggOutputProject;
final Aggregate aggregate;
if (flavor) {
aggOutputProject = call.rel(2);
aggregate = call.rel(3);
} else {
aggregate = call.rel(2);
// Create identity projection
final List<Pair<RexNode, String>> projects = Lists.newArrayList();
final List<RelDataTypeField> fields =
aggregate.getRowType().getFieldList();
for (int i = 0; i < fields.size(); i++) {
projects.add(RexInputRef.of2(projects.size(), fields));
}
aggOutputProject = (Project) RelOptUtil.createProject(
aggregate, Pair.left(projects), Pair.right(projects), false, relBuilder);
}
onMatch2(call, correlate, left, aggOutputProject, aggregate);
}
private void onMatch2(
RelOptRuleCall call,
LogicalCorrelate correlate,
RelNode leftInput,
Project aggOutputProject,
Aggregate aggregate) {
if (generatedCorRels.contains(correlate)) {
// This correlator was generated by a previous invocation of
// this rule. No further work to do.
return;
}
setCurrent(call.getPlanner().getRoot(), correlate);
// check for this pattern
// The pattern matching could be simplified if rules can be applied
// during decorrelation,
//
// CorrelateRel(left correlation, condition = true)
// LeftInputRel
// Project-A (a RexNode)
// Aggregate (groupby (0), agg0(), agg1()...)
// check aggOutputProj projects only one expression
List<RexNode> aggOutputProjExprs = aggOutputProject.getProjects();
if (aggOutputProjExprs.size() != 1) {
return;
}
if(correlate.getJoinType() != JoinRelType.LEFT) {
return;
}
JoinRelType joinType = correlate.getJoinType();
// corRel.getCondition was here, however Correlate was updated so it
// never includes a join condition. The code was not modified for brevity.
RexNode joinCond = rexBuilder.makeLiteral(true);
if ((joinType != JoinRelType.LEFT)
|| (joinCond != rexBuilder.makeLiteral(true))) {
return;
}
// check that the agg is on the entire input
if (!aggregate.getGroupSet().isEmpty()) {
return;
}
List<AggregateCall> aggCalls = aggregate.getAggCallList();
Set<Integer> isCount = Sets.newHashSet();
// remember the count() positions
int i = -1;
for (AggregateCall aggCall : aggCalls) {
++i;
if (aggCall.getAggregation() instanceof SqlCountAggFunction) {
isCount.add(i);
}
}
// now rewrite the plan to
//
// Project-A' (all LHS plus transformed original projections,
// replacing references to count() with case statement)
// Correlator(left correlation, condition = true)
// LeftInputRel
// Aggregate (groupby (0), agg0(), agg1()...)
//
LogicalCorrelate newCorrelate =
LogicalCorrelate.create(leftInput, aggregate,
correlate.getCorrelationId(), correlate.getRequiredColumns(),
correlate.getJoinType());
// remember this rel so we don't fire rule on it again
// REVIEW jhyde 29-Oct-2007: rules should not save state; rule
// should recognize patterns where it does or does not need to do
// work
generatedCorRels.add(newCorrelate);
// need to update the mapCorVarToCorRel Update the output position
// for the cor vars: only pass on the cor vars that are not used in
// the join key.
if (cm.mapCorToCorRel.get(correlate.getCorrelationId()) == correlate) {
cm.mapCorToCorRel.put(correlate.getCorrelationId(), newCorrelate);
}
RelNode newOutput =
aggregateCorrelatorOutput(newCorrelate, aggOutputProject, isCount);
call.transformTo(newOutput);
}
}
/**
* A unique reference to a correlation field.
*
* <p>For instance, if a RelNode references emp.name multiple times, it would
* result in multiple {@code CorRef} objects that differ just in
* {@link CorRef#uniqueKey}.
*/
static class CorRef implements Comparable<CorRef> {
private final int uniqueKey;
private final CorrelationId corr;
private final int field;
CorRef(CorrelationId corr, int field, int uniqueKey) {
this.corr = corr;
this.field = field;
this.uniqueKey = uniqueKey;
}
@Override public String toString() {
return corr.getName() + '.' + field;
}
@Override public int hashCode() {
return Objects.hash(uniqueKey, corr, field);
}
@Override public boolean equals(Object o) {
return this == o
|| o instanceof CorRef
&& uniqueKey == ((CorRef) o).uniqueKey
&& corr == ((CorRef) o).corr
&& field == ((CorRef) o).field;
}
@Override
public int compareTo(@Nonnull CorRef o) {
int c = corr.compareTo(o.corr);
if (c != 0) {
return c;
}
c = Integer.compare(field, o.field);
if (c != 0) {
return c;
}
return Integer.compare(uniqueKey, o.uniqueKey);
}
public CorDef def() {
return new CorDef(corr, field);
}
}
/** A correlation and a field. */
static class CorDef implements Comparable<CorDef> {
private final CorrelationId corr;
private final int field;
private SqlOperator predicateKind;
// this indicates if corr var is left operand of rex call or not
// this is used in decorrelate(logical correlate) to appropriately
// create Rex node expression
private boolean isLeft;
CorDef(CorrelationId corr, int field) {
this.corr = corr;
this.field = field;
this.predicateKind = null;
this.isLeft=false;
}
@Override public String toString() {
return corr.getName() + '.' + field;
}
@Override public int hashCode() {
return Objects.hash(corr, field);
}
@Override public boolean equals(Object o) {
return this == o
|| o instanceof CorDef
&& corr == ((CorDef) o).corr
&& field == ((CorDef) o).field;
}
@Override
public int compareTo(@Nonnull CorDef o) {
int c = corr.compareTo(o.corr);
if (c != 0) {
return c;
}
return Integer.compare(field, o.field);
}
public SqlOperator getPredicateKind() {
return predicateKind;
}
public void setPredicateKind(SqlOperator predKind) {
this.predicateKind = predKind;
}
public boolean getIsLeft() {
return this.isLeft;
}
public void setIsLeft(boolean isLeft) {
this.isLeft = isLeft;
}
}
/** A map of the locations of
* {@link org.apache.calcite.rel.logical.LogicalCorrelate}
* in a tree of {@link RelNode}s.
*
* <p>It is used to drive the decorrelation process.
* Treat it as immutable; rebuild if you modify the tree.
*
* <p>There are three maps:<ol>
*
* <li>{@link #mapRefRelToCorRef} maps a {@link RelNode} to the correlated
* variables it references;
*
* <li>{@link #mapCorToCorRel} maps a correlated variable to the
* {@link Correlate} providing it;
*
* <li>{@link #mapFieldAccessToCorRef} maps a rex field access to
* the corVar it represents. Because typeFlattener does not clone or
* modify a correlated field access this map does not need to be
* updated.
*
* </ol> */
private static final class CorelMap {
private final Multimap<RelNode, CorRef> mapRefRelToCorRef;
private final SortedMap<CorrelationId, RelNode> mapCorToCorRel;
private final Map<RexFieldAccess, CorRef> mapFieldAccessToCorRef;
// TODO: create immutable copies of all maps
private CorelMap(Multimap<RelNode, CorRef> mapRefRelToCorRef,
SortedMap<CorrelationId, RelNode> mapCorToCorRel,
Map<RexFieldAccess, CorRef> mapFieldAccessToCorRef) {
this.mapRefRelToCorRef = mapRefRelToCorRef;
this.mapCorToCorRel = mapCorToCorRel;
this.mapFieldAccessToCorRef = ImmutableMap.copyOf(mapFieldAccessToCorRef);
}
@Override public String toString() {
return "mapRefRelToCorRef=" + mapRefRelToCorRef
+ "\nmapCorToCorRel=" + mapCorToCorRel
+ "\nmapFieldAccessToCorRef=" + mapFieldAccessToCorRef
+ "\n";
}
@Override public boolean equals(Object obj) {
return obj == this
|| obj instanceof CorelMap
&& mapRefRelToCorRef.equals(((CorelMap) obj).mapRefRelToCorRef)
&& mapCorToCorRel.equals(((CorelMap) obj).mapCorToCorRel)
&& mapFieldAccessToCorRef.equals(
((CorelMap) obj).mapFieldAccessToCorRef);
}
@Override public int hashCode() {
return Objects.hash(mapRefRelToCorRef, mapCorToCorRel,
mapFieldAccessToCorRef);
}
/** Creates a CorelMap with given contents. */
public static CorelMap of(
SortedSetMultimap<RelNode, CorRef> mapRefRelToCorVar,
SortedMap<CorrelationId, RelNode> mapCorToCorRel,
Map<RexFieldAccess, CorRef> mapFieldAccessToCorVar) {
return new CorelMap(mapRefRelToCorVar, mapCorToCorRel,
mapFieldAccessToCorVar);
}
/**
* Returns whether there are any correlating variables in this statement.
*
* @return whether there are any correlating variables
*/
public boolean hasCorrelation() {
return !mapCorToCorRel.isEmpty();
}
}
private static class findIfValueGenRequired extends HiveRelShuttleImpl {
private boolean mightRequireValueGen;
findIfValueGenRequired() {
this.mightRequireValueGen = true;
}
private boolean hasRexOver(List<RexNode> projects) {
for(RexNode expr : projects) {
if(expr instanceof RexOver) {
return true;
}
}
return false;
}
@Override public RelNode visit(HiveJoin rel) {
mightRequireValueGen = true;
return rel;
}
public RelNode visit(HiveSortLimit rel) {
mightRequireValueGen = true;
return rel;
}
public RelNode visit(HiveUnion rel) {
mightRequireValueGen = true;
return rel;
}
public RelNode visit(HiveIntersect rel) {
mightRequireValueGen = true;
return rel;
}
@Override public RelNode visit(HiveProject rel) {
if(!(hasRexOver(rel.getProjects()))) {
mightRequireValueGen = false;
return super.visit(rel);
} else {
mightRequireValueGen = true;
return rel;
}
}
@Override public RelNode visit(HiveAggregate rel) {
// if there are aggregate functions or grouping sets we will need
// value generator
if(rel.getAggCallList().isEmpty() && !rel.indicator) {
this.mightRequireValueGen = false;
return super.visit(rel);
} else {
// need to reset to true in case previous aggregate/project
// has set it to false
this.mightRequireValueGen = true;
return rel;
}
}
@Override public RelNode visit(LogicalCorrelate rel) {
// this means we are hitting nested subquery so don't
// need to go further
return rel;
}
public boolean traverse(RelNode root) {
root.accept(this);
return mightRequireValueGen;
}
}
/** Builds a {@link org.apache.calcite.sql2rel.RelDecorrelator.CorelMap}. */
private static class CorelMapBuilder extends HiveRelShuttleImpl {
private final SortedMap<CorrelationId, RelNode> mapCorToCorRel =
new TreeMap<>();
private final SortedSetMultimap<RelNode, CorRef> mapRefRelToCorRef =
Multimaps.newSortedSetMultimap(
new HashMap<RelNode, Collection<CorRef>>(),
new Supplier<TreeSet<CorRef>>() {
@Override
public TreeSet<CorRef> get() {
return Sets.newTreeSet();
}
});
private final Map<RexFieldAccess, CorRef> mapFieldAccessToCorVar = new HashMap<>();
private final Holder<Integer> offset = Holder.of(0);
private int corrIdGenerator = 0;
private final List<RelNode> stack = new ArrayList<>();
/** Creates a CorelMap by iterating over a {@link RelNode} tree. */
CorelMap build(RelNode rel) {
stripHep(rel).accept(this);
return new CorelMap(mapRefRelToCorRef, mapCorToCorRel,
mapFieldAccessToCorVar);
}
@Override
public RelNode visit(HiveJoin join) {
try {
Stacks.push(stack, join);
join.getCondition().accept(rexVisitor(join));
} finally {
Stacks.pop(stack, join);
}
return visitJoin(join);
}
@Override protected RelNode visitChild(RelNode parent, int i,
RelNode input) {
return super.visitChild(parent, i, stripHep(input));
}
@Override public RelNode visit(LogicalCorrelate correlate) {
mapCorToCorRel.put(correlate.getCorrelationId(), correlate);
return visitJoin(correlate);
}
private RelNode visitJoin(BiRel join) {
final int x = offset.get();
visitChild(join, 0, join.getLeft());
offset.set(x + join.getLeft().getRowType().getFieldCount());
visitChild(join, 1, join.getRight());
offset.set(x);
return join;
}
@Override
public RelNode visit(final HiveProject project) {
try {
Stacks.push(stack, project);
for (RexNode node : project.getProjects()) {
node.accept(rexVisitor(project));
}
} finally {
Stacks.pop(stack, project);
}
return super.visit(project);
}
@Override
public RelNode visit(final HiveFilter filter) {
try {
Stacks.push(stack, filter);
filter.getCondition().accept(rexVisitor(filter));
} finally {
Stacks.pop(stack, filter);
}
return super.visit(filter);
}
private RexVisitorImpl<Void> rexVisitor(final RelNode rel) {
return new RexVisitorImpl<Void>(true) {
@Override public Void visitFieldAccess(RexFieldAccess fieldAccess) {
final RexNode ref = fieldAccess.getReferenceExpr();
if (ref instanceof RexCorrelVariable) {
final RexCorrelVariable var = (RexCorrelVariable) ref;
if (mapFieldAccessToCorVar.containsKey(fieldAccess)) {
//for cases where different Rel nodes are referring to
// same correlation var (e.g. in case of NOT IN)
// avoid generating another correlation var
// and record the 'rel' is using the same correlation
mapRefRelToCorRef.put(rel,
mapFieldAccessToCorVar.get(fieldAccess));
} else {
final CorRef correlation =
new CorRef(var.id, fieldAccess.getField().getIndex(),
corrIdGenerator++);
mapFieldAccessToCorVar.put(fieldAccess, correlation);
mapRefRelToCorRef.put(rel, correlation);
}
}
return super.visitFieldAccess(fieldAccess);
}
@Override public Void visitSubQuery(RexSubQuery subQuery) {
subQuery.rel.accept(CorelMapBuilder.this);
return super.visitSubQuery(subQuery);
}
};
}
}
/** Frame describing the relational expression after decorrelation
* and where to find the output fields and correlation variables
* among its output fields. */
static class Frame {
private final RelNode r;
private final ImmutableSortedMap<CorDef, Integer> corDefOutputs;
private final ImmutableSortedMap<Integer, Integer> oldToNewOutputs;
Frame(RelNode oldRel, RelNode r, SortedMap<CorDef, Integer> corDefOutputs,
Map<Integer, Integer> oldToNewOutputs) {
this.r = Preconditions.checkNotNull(r);
this.corDefOutputs = ImmutableSortedMap.copyOf(corDefOutputs);
this.oldToNewOutputs = ImmutableSortedMap.copyOf(oldToNewOutputs);
assert allLessThan(corDefOutputs.values(),
r.getRowType().getFieldCount(), Litmus.THROW);
assert allLessThan(oldToNewOutputs.keySet(),
oldRel.getRowType().getFieldCount(), Litmus.THROW);
assert allLessThan(oldToNewOutputs.values(),
r.getRowType().getFieldCount(), Litmus.THROW);
}
}
}
// End RelDecorrelator.java