Examples with ImmutableBitSet org.apache.calcite.util.ImmutableBitSet used on opensource projects

Example 21 with ImmutableBitSet

use of org.apache.calcite.util.ImmutableBitSet in project calcite by apache.

the class AggregateJoinTransposeRule method onMatch.

public void onMatch(RelOptRuleCall call) {
    final Aggregate aggregate = call.rel(0);
    final Join join = call.rel(1);
    final RexBuilder rexBuilder = aggregate.getCluster().getRexBuilder();
    final RelBuilder relBuilder = call.builder();
    // If any aggregate call has a filter, bail out
    for (AggregateCall aggregateCall : aggregate.getAggCallList()) {
        if (aggregateCall.getAggregation().unwrap(SqlSplittableAggFunction.class) == null) {
            return;
        }
        if (aggregateCall.filterArg >= 0) {
            return;
        }
    }
    // aggregate operator
    if (join.getJoinType() != JoinRelType.INNER) {
        return;
    }
    if (!allowFunctions && !aggregate.getAggCallList().isEmpty()) {
        return;
    }
    // Do the columns used by the join appear in the output of the aggregate?
    final ImmutableBitSet aggregateColumns = aggregate.getGroupSet();
    final RelMetadataQuery mq = call.getMetadataQuery();
    final ImmutableBitSet keyColumns = keyColumns(aggregateColumns, mq.getPulledUpPredicates(join).pulledUpPredicates);
    final ImmutableBitSet joinColumns = RelOptUtil.InputFinder.bits(join.getCondition());
    final boolean allColumnsInAggregate = keyColumns.contains(joinColumns);
    final ImmutableBitSet belowAggregateColumns = aggregateColumns.union(joinColumns);
    // Split join condition
    final List<Integer> leftKeys = Lists.newArrayList();
    final List<Integer> rightKeys = Lists.newArrayList();
    final List<Boolean> filterNulls = Lists.newArrayList();
    RexNode nonEquiConj = RelOptUtil.splitJoinCondition(join.getLeft(), join.getRight(), join.getCondition(), leftKeys, rightKeys, filterNulls);
    // If it contains non-equi join conditions, we bail out
    if (!nonEquiConj.isAlwaysTrue()) {
        return;
    }
    // Push each aggregate function down to each side that contains all of its
    // arguments. Note that COUNT(*), because it has no arguments, can go to
    // both sides.
    final Map<Integer, Integer> map = new HashMap<>();
    final List<Side> sides = new ArrayList<>();
    int uniqueCount = 0;
    int offset = 0;
    int belowOffset = 0;
    for (int s = 0; s < 2; s++) {
        final Side side = new Side();
        final RelNode joinInput = join.getInput(s);
        int fieldCount = joinInput.getRowType().getFieldCount();
        final ImmutableBitSet fieldSet = ImmutableBitSet.range(offset, offset + fieldCount);
        final ImmutableBitSet belowAggregateKeyNotShifted = belowAggregateColumns.intersect(fieldSet);
        for (Ord<Integer> c : Ord.zip(belowAggregateKeyNotShifted)) {
            map.put(c.e, belowOffset + c.i);
        }
        final Mappings.TargetMapping mapping = s == 0 ? Mappings.createIdentity(fieldCount) : Mappings.createShiftMapping(fieldCount + offset, 0, offset, fieldCount);
        final ImmutableBitSet belowAggregateKey = belowAggregateKeyNotShifted.shift(-offset);
        final boolean unique;
        if (!allowFunctions) {
            assert aggregate.getAggCallList().isEmpty();
            // If there are no functions, it doesn't matter as much whether we
            // aggregate the inputs before the join, because there will not be
            // any functions experiencing a cartesian product effect.
            // 
            // But finding out whether the input is already unique requires a call
            // to areColumnsUnique that currently (until [CALCITE-1048] "Make
            // metadata more robust" is fixed) places a heavy load on
            // the metadata system.
            // 
            // So we choose to imagine the the input is already unique, which is
            // untrue but harmless.
            // 
            Util.discard(Bug.CALCITE_1048_FIXED);
            unique = true;
        } else {
            final Boolean unique0 = mq.areColumnsUnique(joinInput, belowAggregateKey);
            unique = unique0 != null && unique0;
        }
        if (unique) {
            ++uniqueCount;
            side.aggregate = false;
            relBuilder.push(joinInput);
            final List<RexNode> projects = new ArrayList<>();
            for (Integer i : belowAggregateKey) {
                projects.add(relBuilder.field(i));
            }
            for (Ord<AggregateCall> aggCall : Ord.zip(aggregate.getAggCallList())) {
                final SqlAggFunction aggregation = aggCall.e.getAggregation();
                final SqlSplittableAggFunction splitter = Preconditions.checkNotNull(aggregation.unwrap(SqlSplittableAggFunction.class));
                if (!aggCall.e.getArgList().isEmpty() && fieldSet.contains(ImmutableBitSet.of(aggCall.e.getArgList()))) {
                    final RexNode singleton = splitter.singleton(rexBuilder, joinInput.getRowType(), aggCall.e.transform(mapping));
                    if (singleton instanceof RexInputRef) {
                        side.split.put(aggCall.i, ((RexInputRef) singleton).getIndex());
                    } else {
                        projects.add(singleton);
                        side.split.put(aggCall.i, projects.size() - 1);
                    }
                }
            }
            relBuilder.project(projects);
            side.newInput = relBuilder.build();
        } else {
            side.aggregate = true;
            List<AggregateCall> belowAggCalls = new ArrayList<>();
            final SqlSplittableAggFunction.Registry<AggregateCall> belowAggCallRegistry = registry(belowAggCalls);
            final int oldGroupKeyCount = aggregate.getGroupCount();
            final int newGroupKeyCount = belowAggregateKey.cardinality();
            for (Ord<AggregateCall> aggCall : Ord.zip(aggregate.getAggCallList())) {
                final SqlAggFunction aggregation = aggCall.e.getAggregation();
                final SqlSplittableAggFunction splitter = Preconditions.checkNotNull(aggregation.unwrap(SqlSplittableAggFunction.class));
                final AggregateCall call1;
                if (fieldSet.contains(ImmutableBitSet.of(aggCall.e.getArgList()))) {
                    final AggregateCall splitCall = splitter.split(aggCall.e, mapping);
                    call1 = splitCall.adaptTo(joinInput, splitCall.getArgList(), splitCall.filterArg, oldGroupKeyCount, newGroupKeyCount);
                } else {
                    call1 = splitter.other(rexBuilder.getTypeFactory(), aggCall.e);
                }
                if (call1 != null) {
                    side.split.put(aggCall.i, belowAggregateKey.cardinality() + belowAggCallRegistry.register(call1));
                }
            }
            side.newInput = relBuilder.push(joinInput).aggregate(relBuilder.groupKey(belowAggregateKey, null), belowAggCalls).build();
        }
        offset += fieldCount;
        belowOffset += side.newInput.getRowType().getFieldCount();
        sides.add(side);
    }
    if (uniqueCount == 2) {
        // invocation of this rule; if we continue we might loop forever.
        return;
    }
    // Update condition
    final Mapping mapping = (Mapping) Mappings.target(new Function<Integer, Integer>() {

        public Integer apply(Integer a0) {
            return map.get(a0);
        }
    }, join.getRowType().getFieldCount(), belowOffset);
    final RexNode newCondition = RexUtil.apply(mapping, join.getCondition());
    // Create new join
    relBuilder.push(sides.get(0).newInput).push(sides.get(1).newInput).join(join.getJoinType(), newCondition);
    // Aggregate above to sum up the sub-totals
    final List<AggregateCall> newAggCalls = new ArrayList<>();
    final int groupIndicatorCount = aggregate.getGroupCount() + aggregate.getIndicatorCount();
    final int newLeftWidth = sides.get(0).newInput.getRowType().getFieldCount();
    final List<RexNode> projects = new ArrayList<>(rexBuilder.identityProjects(relBuilder.peek().getRowType()));
    for (Ord<AggregateCall> aggCall : Ord.zip(aggregate.getAggCallList())) {
        final SqlAggFunction aggregation = aggCall.e.getAggregation();
        final SqlSplittableAggFunction splitter = Preconditions.checkNotNull(aggregation.unwrap(SqlSplittableAggFunction.class));
        final Integer leftSubTotal = sides.get(0).split.get(aggCall.i);
        final Integer rightSubTotal = sides.get(1).split.get(aggCall.i);
        newAggCalls.add(splitter.topSplit(rexBuilder, registry(projects), groupIndicatorCount, relBuilder.peek().getRowType(), aggCall.e, leftSubTotal == null ? -1 : leftSubTotal, rightSubTotal == null ? -1 : rightSubTotal + newLeftWidth));
    }
    relBuilder.project(projects);
    boolean aggConvertedToProjects = false;
    if (allColumnsInAggregate) {
        // let's see if we can convert aggregate into projects
        List<RexNode> projects2 = new ArrayList<>();
        for (int key : Mappings.apply(mapping, aggregate.getGroupSet())) {
            projects2.add(relBuilder.field(key));
        }
        for (AggregateCall newAggCall : newAggCalls) {
            final SqlSplittableAggFunction splitter = newAggCall.getAggregation().unwrap(SqlSplittableAggFunction.class);
            if (splitter != null) {
                final RelDataType rowType = relBuilder.peek().getRowType();
                projects2.add(splitter.singleton(rexBuilder, rowType, newAggCall));
            }
        }
        if (projects2.size() == aggregate.getGroupSet().cardinality() + newAggCalls.size()) {
            // We successfully converted agg calls into projects.
            relBuilder.project(projects2);
            aggConvertedToProjects = true;
        }
    }
    if (!aggConvertedToProjects) {
        relBuilder.aggregate(relBuilder.groupKey(Mappings.apply(mapping, aggregate.getGroupSet()), Mappings.apply2(mapping, aggregate.getGroupSets())), newAggCalls);
    }
    call.transformTo(relBuilder.build());
}

Example 22 with ImmutableBitSet

use of org.apache.calcite.util.ImmutableBitSet in project calcite by apache.

the class AggregateJoinTransposeRule method keyColumns.

/**
 * Computes the closure of a set of columns according to a given list of
 * constraints. Each 'x = y' constraint causes bit y to be set if bit x is
 * set, and vice versa.
 */
private static ImmutableBitSet keyColumns(ImmutableBitSet aggregateColumns, ImmutableList<RexNode> predicates) {
    SortedMap<Integer, BitSet> equivalence = new TreeMap<>();
    for (RexNode predicate : predicates) {
        populateEquivalences(equivalence, predicate);
    }
    ImmutableBitSet keyColumns = aggregateColumns;
    for (Integer aggregateColumn : aggregateColumns) {
        final BitSet bitSet = equivalence.get(aggregateColumn);
        if (bitSet != null) {
            keyColumns = keyColumns.union(bitSet);
        }
    }
    return keyColumns;
}

Example 23 with ImmutableBitSet

use of org.apache.calcite.util.ImmutableBitSet in project calcite by apache.

the class AggregateProjectMergeRule method apply.

public static RelNode apply(RelOptRuleCall call, Aggregate aggregate, Project project) {
    final List<Integer> newKeys = Lists.newArrayList();
    final Map<Integer, Integer> map = new HashMap<>();
    for (int key : aggregate.getGroupSet()) {
        final RexNode rex = project.getProjects().get(key);
        if (rex instanceof RexInputRef) {
            final int newKey = ((RexInputRef) rex).getIndex();
            newKeys.add(newKey);
            map.put(key, newKey);
        } else {
            // Cannot handle "GROUP BY expression"
            return null;
        }
    }
    final ImmutableBitSet newGroupSet = aggregate.getGroupSet().permute(map);
    ImmutableList<ImmutableBitSet> newGroupingSets = null;
    if (aggregate.getGroupType() != Group.SIMPLE) {
        newGroupingSets = ImmutableBitSet.ORDERING.immutableSortedCopy(ImmutableBitSet.permute(aggregate.getGroupSets(), map));
    }
    final ImmutableList.Builder<AggregateCall> aggCalls = ImmutableList.builder();
    for (AggregateCall aggregateCall : aggregate.getAggCallList()) {
        final ImmutableList.Builder<Integer> newArgs = ImmutableList.builder();
        for (int arg : aggregateCall.getArgList()) {
            final RexNode rex = project.getProjects().get(arg);
            if (rex instanceof RexInputRef) {
                newArgs.add(((RexInputRef) rex).getIndex());
            } else {
                // Cannot handle "AGG(expression)"
                return null;
            }
        }
        final int newFilterArg;
        if (aggregateCall.filterArg >= 0) {
            final RexNode rex = project.getProjects().get(aggregateCall.filterArg);
            if (!(rex instanceof RexInputRef)) {
                return null;
            }
            newFilterArg = ((RexInputRef) rex).getIndex();
        } else {
            newFilterArg = -1;
        }
        aggCalls.add(aggregateCall.copy(newArgs.build(), newFilterArg));
    }
    final Aggregate newAggregate = aggregate.copy(aggregate.getTraitSet(), project.getInput(), aggregate.indicator, newGroupSet, newGroupingSets, aggCalls.build());
    // Add a project if the group set is not in the same order or
    // contains duplicates.
    final RelBuilder relBuilder = call.builder();
    relBuilder.push(newAggregate);
    if (!newKeys.equals(newGroupSet.asList())) {
        final List<Integer> posList = Lists.newArrayList();
        for (int newKey : newKeys) {
            posList.add(newGroupSet.indexOf(newKey));
        }
        if (aggregate.indicator) {
            for (int newKey : newKeys) {
                posList.add(aggregate.getGroupCount() + newGroupSet.indexOf(newKey));
            }
        }
        for (int i = newAggregate.getGroupCount() + newAggregate.getIndicatorCount(); i < newAggregate.getRowType().getFieldCount(); i++) {
            posList.add(i);
        }
        relBuilder.project(relBuilder.fields(posList));
    }
    return relBuilder.build();
}

Example 24 with ImmutableBitSet

use of org.apache.calcite.util.ImmutableBitSet in project calcite by apache.

the class JoinToMultiJoinRule method onMatch.

// ~ Methods ----------------------------------------------------------------
public void onMatch(RelOptRuleCall call) {
    final Join origJoin = call.rel(0);
    final RelNode left = call.rel(1);
    final RelNode right = call.rel(2);
    // combine the children MultiJoin inputs into an array of inputs
    // for the new MultiJoin
    final List<ImmutableBitSet> projFieldsList = Lists.newArrayList();
    final List<int[]> joinFieldRefCountsList = Lists.newArrayList();
    final List<RelNode> newInputs = combineInputs(origJoin, left, right, projFieldsList, joinFieldRefCountsList);
    // combine the outer join information from the left and right
    // inputs, and include the outer join information from the current
    // join, if it's a left/right outer join
    final List<Pair<JoinRelType, RexNode>> joinSpecs = Lists.newArrayList();
    combineOuterJoins(origJoin, newInputs, left, right, joinSpecs);
    // pull up the join filters from the children MultiJoinRels and
    // combine them with the join filter associated with this LogicalJoin to
    // form the join filter for the new MultiJoin
    List<RexNode> newJoinFilters = combineJoinFilters(origJoin, left, right);
    // add on the join field reference counts for the join condition
    // associated with this LogicalJoin
    final ImmutableMap<Integer, ImmutableIntList> newJoinFieldRefCountsMap = addOnJoinFieldRefCounts(newInputs, origJoin.getRowType().getFieldCount(), origJoin.getCondition(), joinFieldRefCountsList);
    List<RexNode> newPostJoinFilters = combinePostJoinFilters(origJoin, left, right);
    final RexBuilder rexBuilder = origJoin.getCluster().getRexBuilder();
    RelNode multiJoin = new MultiJoin(origJoin.getCluster(), newInputs, RexUtil.composeConjunction(rexBuilder, newJoinFilters, false), origJoin.getRowType(), origJoin.getJoinType() == JoinRelType.FULL, Pair.right(joinSpecs), Pair.left(joinSpecs), projFieldsList, newJoinFieldRefCountsMap, RexUtil.composeConjunction(rexBuilder, newPostJoinFilters, true));
    call.transformTo(multiJoin);
}

Example 25 with ImmutableBitSet

use of org.apache.calcite.util.ImmutableBitSet in project calcite by apache.

the class LoptOptimizeJoinRule method computeJoinCardinality.

/**
 * Computes the cardinality of the join columns from a particular factor,
 * when that factor is joined with another join tree.
 *
 * @param multiJoin join factors being optimized
 * @param semiJoinOpt optimal semijoins chosen for each factor
 * @param joinTree the join tree that the factor is being joined with
 * @param filters possible join filters to select from
 * @param factor the factor being added
 *
 * @return computed cardinality
 */
private Double computeJoinCardinality(RelMetadataQuery mq, LoptMultiJoin multiJoin, LoptSemiJoinOptimizer semiJoinOpt, LoptJoinTree joinTree, List<RexNode> filters, int factor) {
    final ImmutableBitSet childFactors = ImmutableBitSet.builder().addAll(joinTree.getTreeOrder()).set(factor).build();
    int factorStart = multiJoin.getJoinStart(factor);
    int nFields = multiJoin.getNumFieldsInJoinFactor(factor);
    final ImmutableBitSet.Builder joinKeys = ImmutableBitSet.builder();
    // first loop through the inner join filters, picking out the ones
    // that reference only the factors in either the join tree or the
    // factor that will be added
    setFactorJoinKeys(multiJoin, filters, childFactors, factorStart, nFields, joinKeys);
    // then loop through the outer join filters where the factor being
    // added is the null generating factor in the outer join
    setFactorJoinKeys(multiJoin, RelOptUtil.conjunctions(multiJoin.getOuterJoinCond(factor)), childFactors, factorStart, nFields, joinKeys);
    // null in that case
    if (joinKeys.isEmpty()) {
        return null;
    } else {
        return mq.getDistinctRowCount(semiJoinOpt.getChosenSemiJoin(factor), joinKeys.build(), null);
    }
}