HashJoin.cpp File Reference

#include "QueryEngine/JoinHashTable/HashJoin.h"
#include "QueryEngine/ColumnFetcher.h"
#include "QueryEngine/EquiJoinCondition.h"
#include "QueryEngine/Execute.h"
#include "QueryEngine/JoinHashTable/BaselineJoinHashTable.h"
#include "QueryEngine/JoinHashTable/OverlapsJoinHashTable.h"
#include "QueryEngine/JoinHashTable/PerfectJoinHashTable.h"
#include "QueryEngine/RangeTableIndexVisitor.h"
#include "QueryEngine/RuntimeFunctions.h"
#include "QueryEngine/ScalarExprVisitor.h"

Include dependency graph for HashJoin.cpp:

Go to the source code of this file.

Classes
class	AllColumnVarsVisitor

Namespaces
	anonymous_namespace{HashJoin.cpp}

Functions
template<typename T >
std::string	anonymous_namespace{HashJoin.cpp}::toStringFlat (const HashJoin *hash_table, const ExecutorDeviceType device_type, const int device_id)
std::ostream &	operator<< (std::ostream &os, const DecodedJoinHashBufferEntry &e)
std::ostream &	operator<< (std::ostream &os, const DecodedJoinHashBufferSet &s)
std::shared_ptr < Analyzer::ColumnVar >	getSyntheticColumnVar (std::string_view table, std::string_view column, int rte_idx, Executor *executor)
void	setupSyntheticCaching (std::set< const Analyzer::ColumnVar * > cvs, Executor *executor)
std::vector< InputTableInfo >	getSyntheticInputTableInfo (std::set< const Analyzer::ColumnVar * > cvs, Executor *executor)
InnerOuter	anonymous_namespace{HashJoin.cpp}::get_cols (const Analyzer::BinOper *qual_bin_oper, const Catalog_Namespace::Catalog &cat, const TemporaryTables *temporary_tables)
size_t	get_shard_count (const Analyzer::BinOper *join_condition, const Executor *executor)
InnerOuter	normalize_column_pair (const Analyzer::Expr *lhs, const Analyzer::Expr *rhs, const Catalog_Namespace::Catalog &cat, const TemporaryTables *temporary_tables, const bool is_overlaps_join)
std::vector< InnerOuter >	normalize_column_pairs (const Analyzer::BinOper *condition, const Catalog_Namespace::Catalog &cat, const TemporaryTables *temporary_tables)

Variables
bool	g_enable_overlaps_hashjoin

Function Documentation

size_t get_shard_count	(	const Analyzer::BinOper *	join_condition,
		const Executor *	executor
	)

Definition at line 553 of file HashJoin.cpp.

References anonymous_namespace{HashJoin.cpp}::get_cols(), and get_shard_count().

Referenced by get_shard_count(), BaselineJoinHashTable::getShardCountForCondition(), PerfectJoinHashTable::reify(), PerfectJoinHashTable::shardCount(), and Executor::skipFragmentPair().

                                                 {
  const Analyzer::ColumnVar* inner_col{nullptr};
  const Analyzer::Expr* outer_col{nullptr};
  std::shared_ptr<Analyzer::BinOper> redirected_bin_oper;
  try {
    std::tie(inner_col, outer_col) =
        get_cols(join_condition, *executor->getCatalog(), executor->getTemporaryTables());
  } catch (...) {
    return 0;
  }
  if (!inner_col || !outer_col) {
    return 0;
  }
  return get_shard_count({inner_col, outer_col}, executor);
}

Here is the call graph for this function:

Here is the caller graph for this function:

std::shared_ptr<Analyzer::ColumnVar> getSyntheticColumnVar	(	std::string_view	table,
		std::string_view	column,
		int	rte_idx,
		Executor *	executor
	)

Definition at line 356 of file HashJoin.cpp.

References CHECK, kLINESTRING, kMULTIPOLYGON, kPOINT, and kPOLYGON.

Referenced by HashJoin::getSyntheticInstance().

                                                                               {
  auto catalog = executor->getCatalog();
  CHECK(catalog);

  auto tmeta = catalog->getMetadataForTable(std::string(table));
  CHECK(tmeta);

  auto cmeta = catalog->getMetadataForColumn(tmeta->tableId, std::string(column));
  CHECK(cmeta);

  auto ti = cmeta->columnType;

  if (ti.is_geometry() && ti.get_type() != kPOINT) {
    int geoColumnId{0};
    switch (ti.get_type()) {
      case kLINESTRING: {
        geoColumnId = cmeta->columnId + 2;
        break;
      }
      case kPOLYGON: {
        geoColumnId = cmeta->columnId + 3;
        break;
      }
      case kMULTIPOLYGON: {
        geoColumnId = cmeta->columnId + 4;
        break;
      }
      default:
        CHECK(false);
    }
    cmeta = catalog->getMetadataForColumn(tmeta->tableId, geoColumnId);
    CHECK(cmeta);
    ti = cmeta->columnType;
  }

  auto cv =
      std::make_shared<Analyzer::ColumnVar>(ti, tmeta->tableId, cmeta->columnId, rte_idx);
  return cv;
}

Here is the caller graph for this function:

std::vector<InputTableInfo> getSyntheticInputTableInfo	(	std::set< const Analyzer::ColumnVar * >	cvs,
		Executor *	executor
	)

Definition at line 441 of file HashJoin.cpp.

References CHECK, and i.

Referenced by HashJoin::getSyntheticInstance().

                        {
  auto catalog = executor->getCatalog();
  CHECK(catalog);

  std::unordered_set<int> phys_table_ids;
  for (auto cv : cvs) {
    phys_table_ids.insert(cv->get_table_id());
  }

  // NOTE(sy): This vector ordering seems to work for now, but maybe we need to
  // review how rte_idx is assigned for ColumnVars. See for example Analyzer.h
  // and RelAlgExecutor.cpp and rte_idx there.
  std::vector<InputTableInfo> query_infos(phys_table_ids.size());
  size_t i = 0;
  for (auto id : phys_table_ids) {
    auto tmeta = catalog->getMetadataForTable(id);
    query_infos[i].table_id = id;
    query_infos[i].info = tmeta->fragmenter->getFragmentsForQuery();
    ++i;
  }

  return query_infos;
}

Here is the caller graph for this function:

InnerOuter normalize_column_pair	(	const Analyzer::Expr *	lhs,
		const Analyzer::Expr *	rhs,
		const Catalog_Namespace::Catalog &	cat,
		const TemporaryTables *	temporary_tables,
		const bool	is_overlaps_join
	)

Definition at line 570 of file HashJoin.cpp.

References cat(), get_column_descriptor_maybe(), get_column_type(), Analyzer::Expr::get_type_info(), is_constructed_point(), kCAST, kENCODING_DICT, kPOINT, gpu_enabled::swap(), and ScalarExprVisitor< T >::visit().

Referenced by anonymous_namespace{PerfectJoinHashTable.cpp}::get_cols(), anonymous_namespace{HashJoin.cpp}::get_cols(), and normalize_column_pairs().

                                                              {
  const auto& lhs_ti = lhs->get_type_info();
  const auto& rhs_ti = rhs->get_type_info();
  if (!is_overlaps_join) {
    if (lhs_ti.get_type() != rhs_ti.get_type()) {
      throw HashJoinFail("Equijoin types must be identical, found: " +
                         lhs_ti.get_type_name() + ", " + rhs_ti.get_type_name());
    }
    if (!lhs_ti.is_integer() && !lhs_ti.is_time() && !lhs_ti.is_string() &&
        !lhs_ti.is_decimal()) {
      throw HashJoinFail("Cannot apply hash join to inner column type " +
                         lhs_ti.get_type_name());
    }
    // Decimal types should be identical.
    if (lhs_ti.is_decimal() && (lhs_ti.get_scale() != rhs_ti.get_scale() ||
                                lhs_ti.get_precision() != rhs_ti.get_precision())) {
      throw HashJoinFail("Equijoin with different decimal types");
    }
  }

  const auto lhs_cast = dynamic_cast<const Analyzer::UOper*>(lhs);
  const auto rhs_cast = dynamic_cast<const Analyzer::UOper*>(rhs);
  if (lhs_ti.is_string() && (static_cast<bool>(lhs_cast) != static_cast<bool>(rhs_cast) ||
                             (lhs_cast && lhs_cast->get_optype() != kCAST) ||
                             (rhs_cast && rhs_cast->get_optype() != kCAST))) {
    throw HashJoinFail("Cannot use hash join for given expression");
  }
  // Casts to decimal are not suported.
  if (lhs_ti.is_decimal() && (lhs_cast || rhs_cast)) {
    throw HashJoinFail("Cannot use hash join for given expression");
  }
  const auto lhs_col =
      lhs_cast ? dynamic_cast<const Analyzer::ColumnVar*>(lhs_cast->get_operand())
               : dynamic_cast<const Analyzer::ColumnVar*>(lhs);
  const auto rhs_col =
      rhs_cast ? dynamic_cast<const Analyzer::ColumnVar*>(rhs_cast->get_operand())
               : dynamic_cast<const Analyzer::ColumnVar*>(rhs);
  if (!lhs_col && !rhs_col) {
    throw HashJoinFail("Cannot use hash join for given expression");
  }
  const Analyzer::ColumnVar* inner_col{nullptr};
  const Analyzer::ColumnVar* outer_col{nullptr};
  auto outer_ti = lhs_ti;
  auto inner_ti = rhs_ti;
  const Analyzer::Expr* outer_expr{lhs};
  if ((!lhs_col || (rhs_col && lhs_col->get_rte_idx() < rhs_col->get_rte_idx())) &&
      (!rhs_col || (!lhs_col || lhs_col->get_rte_idx() < rhs_col->get_rte_idx()))) {
    inner_col = rhs_col;
    outer_col = lhs_col;
  } else {
    if (lhs_col && lhs_col->get_rte_idx() == 0) {
      throw HashJoinFail("Cannot use hash join for given expression");
    }
    inner_col = lhs_col;
    outer_col = rhs_col;
    std::swap(outer_ti, inner_ti);
    outer_expr = rhs;
  }
  if (!inner_col) {
    throw HashJoinFail("Cannot use hash join for given expression");
  }
  if (!outer_col) {
    MaxRangeTableIndexVisitor rte_idx_visitor;
    int outer_rte_idx = rte_idx_visitor.visit(outer_expr);
    // The inner column candidate is not actually inner; the outer
    // expression contains columns which are at least as deep.
    if (inner_col->get_rte_idx() <= outer_rte_idx) {
      throw HashJoinFail("Cannot use hash join for given expression");
    }
  }
  // We need to fetch the actual type information from the catalog since Analyzer
  // always reports nullable as true for inner table columns in left joins.
  const auto inner_col_cd = get_column_descriptor_maybe(
      inner_col->get_column_id(), inner_col->get_table_id(), cat);
  const auto inner_col_real_ti = get_column_type(inner_col->get_column_id(),
                                                 inner_col->get_table_id(),
                                                 inner_col_cd,
                                                 temporary_tables);
  const auto& outer_col_ti =
      !(dynamic_cast<const Analyzer::FunctionOper*>(lhs)) && outer_col
          ? outer_col->get_type_info()
          : outer_ti;
  // Casts from decimal are not supported.
  if ((inner_col_real_ti.is_decimal() || outer_col_ti.is_decimal()) &&
      (lhs_cast || rhs_cast)) {
    throw HashJoinFail("Cannot use hash join for given expression");
  }
  if (is_overlaps_join) {
    if (!inner_col_real_ti.is_array()) {
      throw HashJoinFail(
          "Overlaps join only supported for inner columns with array type");
    }
    auto is_bounds_array = [](const auto ti) {
      return ti.is_fixlen_array() && ti.get_size() == 32;
    };
    if (!is_bounds_array(inner_col_real_ti)) {
      throw HashJoinFail(
          "Overlaps join only supported for 4-element double fixed length arrays");
    }
    if (!(outer_col_ti.get_type() == kPOINT || is_bounds_array(outer_col_ti) ||
          is_constructed_point(outer_expr))) {
      throw HashJoinFail(
          "Overlaps join only supported for geometry outer columns of type point, "
          "geometry columns with bounds or constructed points");
    }
  } else {
    if (!(inner_col_real_ti.is_integer() || inner_col_real_ti.is_time() ||
          inner_col_real_ti.is_decimal() ||
          (inner_col_real_ti.is_string() &&
           inner_col_real_ti.get_compression() == kENCODING_DICT))) {
      throw HashJoinFail(
          "Can only apply hash join to integer-like types and dictionary encoded "
          "strings");
    }
  }

  auto normalized_inner_col = inner_col;
  auto normalized_outer_col = outer_col ? outer_col : outer_expr;

  const auto& normalized_inner_ti = normalized_inner_col->get_type_info();
  const auto& normalized_outer_ti = normalized_outer_col->get_type_info();

  if (normalized_inner_ti.is_string() != normalized_outer_ti.is_string()) {
    throw HashJoinFail(std::string("Could not build hash tables for incompatible types " +
                                   normalized_inner_ti.get_type_name() + " and " +
                                   normalized_outer_ti.get_type_name()));
  }

  return {normalized_inner_col, normalized_outer_col};
}

Here is the call graph for this function:

Here is the caller graph for this function:

std::vector<InnerOuter> normalize_column_pairs	(	const Analyzer::BinOper *	condition,
		const Catalog_Namespace::Catalog &	cat,
		const TemporaryTables *	temporary_tables
	)

Definition at line 705 of file HashJoin.cpp.

References cat(), CHECK, CHECK_EQ, Analyzer::BinOper::get_left_operand(), Analyzer::BinOper::get_right_operand(), i, Analyzer::BinOper::is_overlaps_oper(), normalize_column_pair(), and run_benchmark_import::result.

Referenced by anonymous_namespace{FromTableReordering.cpp}::get_join_qual_cost(), BaselineJoinHashTable::getInstance(), OverlapsJoinHashTable::getInstance(), and Executor::skipFragmentPair().

                                                                                        {
  std::vector<InnerOuter> result;
  const auto lhs_tuple_expr =
      dynamic_cast<const Analyzer::ExpressionTuple*>(condition->get_left_operand());
  const auto rhs_tuple_expr =
      dynamic_cast<const Analyzer::ExpressionTuple*>(condition->get_right_operand());

  CHECK_EQ(static_cast<bool>(lhs_tuple_expr), static_cast<bool>(rhs_tuple_expr));
  if (lhs_tuple_expr) {
    const auto& lhs_tuple = lhs_tuple_expr->getTuple();
    const auto& rhs_tuple = rhs_tuple_expr->getTuple();
    CHECK_EQ(lhs_tuple.size(), rhs_tuple.size());
    for (size_t i = 0; i < lhs_tuple.size(); ++i) {
      result.push_back(normalize_column_pair(lhs_tuple[i].get(),
                                             rhs_tuple[i].get(),
                                             cat,
                                             temporary_tables,
                                             condition->is_overlaps_oper()));
    }
  } else {
    CHECK(!lhs_tuple_expr && !rhs_tuple_expr);
    result.push_back(normalize_column_pair(condition->get_left_operand(),
                                           condition->get_right_operand(),
                                           cat,
                                           temporary_tables,
                                           condition->is_overlaps_oper()));
  }

  return result;
}

Here is the call graph for this function:

Here is the caller graph for this function:

std::ostream& operator<<	(	std::ostream &	os,
		const DecodedJoinHashBufferEntry &	e
	)

Definition at line 122 of file HashJoin.cpp.

References generate_TableFunctionsFactory_init::k, DecodedJoinHashBufferEntry::key, and DecodedJoinHashBufferEntry::payload.

                                                                            {
  os << "  {{";
  bool first = true;
  for (auto k : e.key) {
    if (!first) {
      os << ",";
    } else {
      first = false;
    }
    os << k;
  }
  os << "}, ";
  os << "{";
  first = true;
  for (auto p : e.payload) {
    if (!first) {
      os << ", ";
    } else {
      first = false;
    }
    os << p;
  }
  os << "}}";
  return os;
}

std::ostream& operator<<	(	std::ostream &	os,
		const DecodedJoinHashBufferSet &	s
	)

Definition at line 148 of file HashJoin.cpp.

                                                                          {
  os << "{\n";
  bool first = true;
  for (auto e : s) {
    if (!first) {
      os << ",\n";
    } else {
      first = false;
    }
    os << e;
  }
  if (!s.empty()) {
    os << "\n";
  }
  os << "}\n";
  return os;
}

void setupSyntheticCaching	(	std::set< const Analyzer::ColumnVar * >	cvs,
		Executor *	executor
	)

Definition at line 427 of file HashJoin.cpp.

Referenced by HashJoin::getSyntheticInstance().

                                                                                     {
  std::unordered_set<int> phys_table_ids;
  for (auto cv : cvs) {
    phys_table_ids.insert(cv->get_table_id());
  }

  std::unordered_set<PhysicalInput> phys_inputs;
  for (auto cv : cvs) {
    phys_inputs.emplace(PhysicalInput{cv->get_column_id(), cv->get_table_id()});
  }

  executor->setupCaching(phys_inputs, phys_table_ids);
}

Here is the caller graph for this function:

Variable Documentation

bool g_enable_overlaps_hashjoin

Definition at line 96 of file Execute.cpp.