Executor Class Reference

#include <Execute.h>

Collaboration diagram for Executor:

Classes
class	CgenStateManager
struct	ExecutorMutexHolder
class	FetchCacheAnchor
struct	GroupColLLVMValue
struct	JoinHashTableOrError

Public Types
enum	ExtModuleKinds {   ExtModuleKinds::template_module, ExtModuleKinds::udf_cpu_module, ExtModuleKinds::udf_gpu_module, ExtModuleKinds::rt_udf_cpu_module,   ExtModuleKinds::rt_udf_gpu_module, ExtModuleKinds::rt_geos_module, ExtModuleKinds::rt_libdevice_module }
using	ExecutorId = size_t
using	CachedCardinality = std::pair< bool, size_t >

Public Member Functions
	Executor (const ExecutorId id, Data_Namespace::DataMgr *data_mgr, const size_t block_size_x, const size_t grid_size_x, const size_t max_gpu_slab_size, const std::string &debug_dir, const std::string &debug_file)
void	clearCaches (bool runtime_only=false)
std::string	dumpCache () const
void	reset (bool discard_runtime_modules_only=false)
const std::unique_ptr < llvm::Module > &	get_rt_module () const
const std::unique_ptr < llvm::Module > &	get_udf_module (bool is_gpu=false) const
const std::unique_ptr < llvm::Module > &	get_rt_udf_module (bool is_gpu=false) const
const std::unique_ptr < llvm::Module > &	get_geos_module () const
const std::unique_ptr < llvm::Module > &	get_libdevice_module () const
bool	has_rt_module () const
bool	has_udf_module (bool is_gpu=false) const
bool	has_rt_udf_module (bool is_gpu=false) const
bool	has_geos_module () const
bool	has_libdevice_module () const
const TemporaryTables *	getTemporaryTables ()
StringDictionaryProxy *	getStringDictionaryProxy (const shared::StringDictKey &dict_key, const bool with_generation) const
StringDictionaryProxy *	getStringDictionaryProxy (const shared::StringDictKey &dict_key, const std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner, const bool with_generation) const
const StringDictionaryProxy::IdMap *	getStringProxyTranslationMap (const shared::StringDictKey &source_dict_key, const shared::StringDictKey &dest_dict_key, const RowSetMemoryOwner::StringTranslationType translation_type, const std::vector< StringOps_Namespace::StringOpInfo > &string_op_infos, std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner, const bool with_generation) const
const StringDictionaryProxy::IdMap *	getJoinIntersectionStringProxyTranslationMap (const StringDictionaryProxy *source_proxy, StringDictionaryProxy *dest_proxy, const std::vector< StringOps_Namespace::StringOpInfo > &source_string_op_infos, const std::vector< StringOps_Namespace::StringOpInfo > &dest_source_string_op_infos, std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner) const
const StringDictionaryProxy::TranslationMap < Datum > *	getStringProxyNumericTranslationMap (const shared::StringDictKey &source_dict_key, const std::vector< StringOps_Namespace::StringOpInfo > &string_op_infos, std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner, const bool with_generation) const
bool	isCPUOnly () const
bool	isArchMaxwell (const ExecutorDeviceType dt) const
bool	containsLeftDeepOuterJoin () const
const ColumnDescriptor *	getColumnDescriptor (const Analyzer::ColumnVar *) const
const ColumnDescriptor *	getPhysicalColumnDescriptor (const Analyzer::ColumnVar *, int) const
Data_Namespace::DataMgr *	getDataMgr () const
const std::shared_ptr < RowSetMemoryOwner >	getRowSetMemoryOwner () const
const TemporaryTables *	getTemporaryTables () const
Fragmenter_Namespace::TableInfo	getTableInfo (const shared::TableKey &table_key) const
const TableGeneration &	getTableGeneration (const shared::TableKey &table_key) const
ExpressionRange	getColRange (const PhysicalInput &) const
size_t	getNumBytesForFetchedRow (const std::set< shared::TableKey > &table_keys_to_fetch) const
bool	hasLazyFetchColumns (const std::vector< Analyzer::Expr * > &target_exprs) const
std::vector< ColumnLazyFetchInfo >	getColLazyFetchInfo (const std::vector< Analyzer::Expr * > &target_exprs) const
void	interrupt (const QuerySessionId &query_session="", const QuerySessionId &interrupt_session="")
void	resetInterrupt ()
void	enableRuntimeQueryInterrupt (const double runtime_query_check_freq, const unsigned pending_query_check_freq) const
int8_t	warpSize () const
unsigned	gridSize () const
void	setGridSize (unsigned grid_size)
void	resetGridSize ()
unsigned	numBlocksPerMP () const
unsigned	blockSize () const
void	setBlockSize (unsigned block_size)
void	resetBlockSize ()
size_t	maxGpuSlabSize () const
ResultSetPtr	executeWorkUnit (size_t &max_groups_buffer_entry_guess, const bool is_agg, const std::vector< InputTableInfo > &, const RelAlgExecutionUnit &, const CompilationOptions &, const ExecutionOptions &options, RenderInfo *render_info, const bool has_cardinality_estimation, ColumnCacheMap &column_cache)
TableUpdateMetadata	executeUpdate (const RelAlgExecutionUnit &ra_exe_unit, const std::vector< InputTableInfo > &table_infos, const TableDescriptor *updated_table_desc, const CompilationOptions &co, const ExecutionOptions &eo, const Catalog_Namespace::Catalog &cat, std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner, const UpdateLogForFragment::Callback &cb, const bool is_agg)
void	addTransientStringLiterals (const RelAlgExecutionUnit &ra_exe_unit, const std::shared_ptr< RowSetMemoryOwner > &row_set_mem_owner)
int	deviceCount (const ExecutorDeviceType) const
void	setupCaching (const std::unordered_set< PhysicalInput > &phys_inputs, const std::unordered_set< shared::TableKey > &phys_table_keys)
void	setColRangeCache (const AggregatedColRange &aggregated_col_range)
ExecutorId	getExecutorId () const
QuerySessionId &	getCurrentQuerySession (heavyai::shared_lock< heavyai::shared_mutex > &read_lock)
QuerySessionStatus::QueryStatus	getQuerySessionStatus (const QuerySessionId &candidate_query_session, heavyai::shared_lock< heavyai::shared_mutex > &read_lock)
bool	checkCurrentQuerySession (const std::string &candidate_query_session, heavyai::shared_lock< heavyai::shared_mutex > &read_lock)
void	invalidateRunningQuerySession (heavyai::unique_lock< heavyai::shared_mutex > &write_lock)
bool	addToQuerySessionList (const QuerySessionId &query_session, const std::string &query_str, const std::string &submitted, const size_t executor_id, const QuerySessionStatus::QueryStatus query_status, heavyai::unique_lock< heavyai::shared_mutex > &write_lock)
bool	removeFromQuerySessionList (const QuerySessionId &query_session, const std::string &submitted_time_str, heavyai::unique_lock< heavyai::shared_mutex > &write_lock)
void	setQuerySessionAsInterrupted (const QuerySessionId &query_session, heavyai::unique_lock< heavyai::shared_mutex > &write_lock)
bool	checkIsQuerySessionInterrupted (const std::string &query_session, heavyai::shared_lock< heavyai::shared_mutex > &read_lock)
bool	checkIsQuerySessionEnrolled (const QuerySessionId &query_session, heavyai::shared_lock< heavyai::shared_mutex > &read_lock)
bool	updateQuerySessionStatusWithLock (const QuerySessionId &query_session, const std::string &submitted_time_str, const QuerySessionStatus::QueryStatus updated_query_status, heavyai::unique_lock< heavyai::shared_mutex > &write_lock)
bool	updateQuerySessionExecutorAssignment (const QuerySessionId &query_session, const std::string &submitted_time_str, const size_t executor_id, heavyai::unique_lock< heavyai::shared_mutex > &write_lock)
std::vector< QuerySessionStatus >	getQuerySessionInfo (const QuerySessionId &query_session, heavyai::shared_lock< heavyai::shared_mutex > &read_lock)
heavyai::shared_mutex &	getSessionLock ()
CurrentQueryStatus	attachExecutorToQuerySession (const QuerySessionId &query_session_id, const std::string &query_str, const std::string &query_submitted_time)
void	checkPendingQueryStatus (const QuerySessionId &query_session)
void	clearQuerySessionStatus (const QuerySessionId &query_session, const std::string &submitted_time_str)
void	updateQuerySessionStatus (const QuerySessionId &query_session, const std::string &submitted_time_str, const QuerySessionStatus::QueryStatus new_query_status)
void	enrollQuerySession (const QuerySessionId &query_session, const std::string &query_str, const std::string &submitted_time_str, const size_t executor_id, const QuerySessionStatus::QueryStatus query_session_status)
size_t	getNumCurentSessionsEnrolled () const
const std::vector< size_t >	getExecutorIdsRunningQuery (const QuerySessionId &interrupt_session) const
bool	checkNonKernelTimeInterrupted () const
void	registerExtractedQueryPlanDag (const QueryPlanDAG &query_plan_dag)
const QueryPlanDAG	getLatestQueryPlanDagExtracted () const
void	addToCardinalityCache (const std::string &cache_key, const size_t cache_value)
CachedCardinality	getCachedCardinality (const std::string &cache_key)
heavyai::shared_mutex &	getDataRecyclerLock ()
QueryPlanDagCache &	getQueryPlanDagCache ()
ResultSetRecyclerHolder &	getRecultSetRecyclerHolder ()
CgenState *	getCgenStatePtr () const
PlanState *	getPlanStatePtr () const
llvm::LLVMContext &	getContext ()
void	update_extension_modules (bool update_runtime_modules_only=false)

Static Public Member Functions
static void	clearExternalCaches (bool for_update, const TableDescriptor *td, const int current_db_id)
template<typename F >
static void	registerExtensionFunctions (F register_extension_functions)
static std::shared_ptr< Executor >	getExecutor (const ExecutorId id, const std::string &debug_dir="", const std::string &debug_file="", const SystemParameters &system_parameters=SystemParameters())
static void	nukeCacheOfExecutors ()
static void	clearMemory (const Data_Namespace::MemoryLevel memory_level)
static size_t	getArenaBlockSize ()
static void	addUdfIrToModule (const std::string &udf_ir_filename, const bool is_cuda_ir)
static void	initialize_extension_module_sources ()
static void	registerActiveModule (void *module, const int device_id)
static void	unregisterActiveModule (const int device_id)
static std::pair< int64_t, int32_t >	reduceResults (const SQLAgg agg, const SQLTypeInfo &ti, const int64_t agg_init_val, const int8_t out_byte_width, const int64_t *out_vec, const size_t out_vec_sz, const bool is_group_by, const bool float_argument_input)
static void	update_after_registration (bool update_runtime_modules_only=false)
static size_t	getBaselineThreshold (bool for_count_distinct, ExecutorDeviceType device_type)

Public Attributes
std::mutex	compilation_mutex_

Static Public Attributes
static const ExecutorId	UNITARY_EXECUTOR_ID = 0
static const ExecutorId	INVALID_EXECUTOR_ID = SIZE_MAX
static std::map < ExtModuleKinds, std::string >	extension_module_sources
static const size_t	high_scan_limit
static const int32_t	ERR_DIV_BY_ZERO {1}
static const int32_t	ERR_OUT_OF_GPU_MEM {2}
static const int32_t	ERR_OUT_OF_SLOTS {3}
static const int32_t	ERR_UNSUPPORTED_SELF_JOIN {4}
static const int32_t	ERR_OUT_OF_RENDER_MEM {5}
static const int32_t	ERR_OUT_OF_CPU_MEM {6}
static const int32_t	ERR_OVERFLOW_OR_UNDERFLOW {7}
static const int32_t	ERR_OUT_OF_TIME {9}
static const int32_t	ERR_INTERRUPTED {10}
static const int32_t	ERR_COLUMNAR_CONVERSION_NOT_SUPPORTED {11}
static const int32_t	ERR_TOO_MANY_LITERALS {12}
static const int32_t	ERR_STRING_CONST_IN_RESULTSET {13}
static const int32_t	ERR_STREAMING_TOP_N_NOT_SUPPORTED_IN_RENDER_QUERY {14}
static const int32_t	ERR_SINGLE_VALUE_FOUND_MULTIPLE_VALUES {15}
static const int32_t	ERR_GEOS {16}
static const int32_t	ERR_WIDTH_BUCKET_INVALID_ARGUMENT {17}
static std::mutex	register_runtime_extension_functions_mutex_
static std::mutex	kernel_mutex_

Private Types
using	PerFragmentCallBack = std::function< void(ResultSetPtr, const Fragmenter_Namespace::FragmentInfo &)>

Private Member Functions
void	clearMetaInfoCache ()
int	deviceCountForMemoryLevel (const Data_Namespace::MemoryLevel memory_level) const
llvm::Value *	codegenWindowFunction (const size_t target_index, const CompilationOptions &co)
llvm::Value *	codegenConditionalAggregateCondValSelector (llvm::Value *cond_lv, SQLAgg const aggKind, CompilationOptions const &co) const
llvm::Value *	codegenWindowFunctionAggregate (const CompilationOptions &co)
llvm::BasicBlock *	codegenWindowResetStateControlFlow ()
void	codegenWindowFunctionStateInit (llvm::Value *aggregate_state)
llvm::Value *	codegenWindowFunctionAggregateCalls (llvm::Value *aggregate_state, const CompilationOptions &co)
llvm::Value *	codegenWindowNavigationFunctionOnFrame (const CompilationOptions &co)
llvm::Value *	codegenCurrentPartitionIndex (const WindowFunctionContext *window_func_context, llvm::Value *current_row_pos_lv)
llvm::Value *	codegenFrameBoundExpr (const Analyzer::WindowFunction *window_func, const Analyzer::WindowFrame *frame_bound, CodeGenerator &code_generator, const CompilationOptions &co)
llvm::Value *	codegenFrameBound (bool for_start_bound, bool for_range_mode, bool for_window_frame_naviation, const Analyzer::WindowFrame *frame_bound, bool is_timestamp_type_frame, llvm::Value *order_key_null_val, const WindowFrameBoundFuncArgs &args)
std::pair< std::string, llvm::Value * >	codegenLoadOrderKeyBufPtr (WindowFunctionContext *window_func_context) const
std::pair< llvm::Value *, llvm::Value * >	codegenFrameNullRange (WindowFunctionContext *window_func_context, llvm::Value *partition_index_lv) const
WindowPartitionBufferPtrs	codegenLoadPartitionBuffers (WindowFunctionContext *window_func_context, llvm::Value *partition_index_lv) const
std::pair< llvm::Value *, llvm::Value * >	codegenWindowFrameBounds (WindowFunctionContext *window_func_context, const Analyzer::WindowFrame *frame_start_bound, const Analyzer::WindowFrame *frame_end_bound, llvm::Value *order_key_col_null_val_lv, WindowFrameBoundFuncArgs &args, CodeGenerator &code_generator)
std::pair< llvm::Value *, llvm::Value * >	codegenFrameBoundRange (const Analyzer::WindowFunction *window_func, CodeGenerator &code_generator, const CompilationOptions &co)
std::vector< llvm::Value * >	prepareRowModeFuncArgs (bool for_start_bound, SqlWindowFrameBoundType bound_type, const WindowFrameBoundFuncArgs &args) const
std::vector< llvm::Value * >	prepareRangeModeFuncArgs (bool for_start_bound, const Analyzer::WindowFrame *frame_bound, bool is_timestamp_type_frame, llvm::Value *order_key_null_val, const WindowFrameBoundFuncArgs &frame_args) const
const std::string	getOrderKeyTypeName (WindowFunctionContext *window_func_context) const
llvm::Value *	codegenLoadCurrentValueFromColBuf (WindowFunctionContext *window_func_context, CodeGenerator &code_generator, WindowFrameBoundFuncArgs &args) const
size_t	getOrderKeySize (WindowFunctionContext *window_func_context) const
const SQLTypeInfo	getFirstOrderColTypeInfo (WindowFunctionContext *window_func_context) const
std::string	getFramingFuncName (const std::string &bound_type, const std::string &order_col_type, const std::string &op_type, bool for_timestamp_type) const
void	codegenWindowAvgEpilogue (llvm::Value *crt_val, llvm::Value *window_func_null_val)
llvm::Value *	codegenAggregateWindowState ()
llvm::Value *	aggregateWindowStatePtr ()
CudaMgr_Namespace::CudaMgr *	cudaMgr () const
bool	isArchPascalOrLater (const ExecutorDeviceType dt) const
bool	needFetchAllFragments (const InputColDescriptor &col_desc, const RelAlgExecutionUnit &ra_exe_unit, const FragmentsList &selected_fragments) const
bool	needLinearizeAllFragments (const ColumnDescriptor *cd, const InputColDescriptor &inner_col_desc, const RelAlgExecutionUnit &ra_exe_unit, const FragmentsList &selected_fragments, const Data_Namespace::MemoryLevel memory_level) const
void	executeWorkUnitPerFragment (const RelAlgExecutionUnit &ra_exe_unit, const InputTableInfo &table_info, const CompilationOptions &co, const ExecutionOptions &eo, const Catalog_Namespace::Catalog &cat, PerFragmentCallBack &cb, const std::set< size_t > &fragment_indexes_param)
	Compiles and dispatches a work unit per fragment processing results with the per fragment callback. Currently used for computing metrics over fragments (metadata). More...
ResultSetPtr	executeExplain (const QueryCompilationDescriptor &)
ResultSetPtr	executeTableFunction (const TableFunctionExecutionUnit exe_unit, const std::vector< InputTableInfo > &table_infos, const CompilationOptions &co, const ExecutionOptions &eo)
	Compiles and dispatches a table function; that is, a function that takes as input one or more columns and returns a ResultSet, which can be parsed by subsequent execution steps. More...
ExecutorDeviceType	getDeviceTypeForTargets (const RelAlgExecutionUnit &ra_exe_unit, const ExecutorDeviceType requested_device_type)
ResultSetPtr	collectAllDeviceResults (SharedKernelContext &shared_context, const RelAlgExecutionUnit &ra_exe_unit, const QueryMemoryDescriptor &query_mem_desc, const ExecutorDeviceType device_type, std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner)
ResultSetPtr	collectAllDeviceShardedTopResults (SharedKernelContext &shared_context, const RelAlgExecutionUnit &ra_exe_unit) const
std::unordered_map < shared::TableKey, const Analyzer::BinOper * >	getInnerTabIdToJoinCond () const
std::vector< std::unique_ptr < ExecutionKernel > >	createKernels (SharedKernelContext &shared_context, const RelAlgExecutionUnit &ra_exe_unit, ColumnFetcher &column_fetcher, const std::vector< InputTableInfo > &table_infos, const ExecutionOptions &eo, const bool is_agg, const bool allow_single_frag_table_opt, const size_t context_count, const QueryCompilationDescriptor &query_comp_desc, const QueryMemoryDescriptor &query_mem_desc, RenderInfo *render_info, std::unordered_set< int > &available_gpus, int &available_cpus)
void	launchKernels (SharedKernelContext &shared_context, std::vector< std::unique_ptr< ExecutionKernel >> &&kernels, const ExecutorDeviceType device_type)
std::vector< size_t >	getTableFragmentIndices (const RelAlgExecutionUnit &ra_exe_unit, const ExecutorDeviceType device_type, const size_t table_idx, const size_t outer_frag_idx, std::map< shared::TableKey, const TableFragments * > &selected_tables_fragments, const std::unordered_map< shared::TableKey, const Analyzer::BinOper * > &inner_table_id_to_join_condition)
bool	skipFragmentPair (const Fragmenter_Namespace::FragmentInfo &outer_fragment_info, const Fragmenter_Namespace::FragmentInfo &inner_fragment_info, const int inner_table_id, const std::unordered_map< shared::TableKey, const Analyzer::BinOper * > &inner_table_id_to_join_condition, const RelAlgExecutionUnit &ra_exe_unit, const ExecutorDeviceType device_type)
FetchResult	fetchChunks (const ColumnFetcher &, const RelAlgExecutionUnit &ra_exe_unit, const int device_id, const Data_Namespace::MemoryLevel, const std::map< shared::TableKey, const TableFragments * > &, const FragmentsList &selected_fragments, std::list< ChunkIter > &, std::list< std::shared_ptr< Chunk_NS::Chunk >> &, DeviceAllocator *device_allocator, const size_t thread_idx, const bool allow_runtime_interrupt)
FetchResult	fetchUnionChunks (const ColumnFetcher &, const RelAlgExecutionUnit &ra_exe_unit, const int device_id, const Data_Namespace::MemoryLevel, const std::map< shared::TableKey, const TableFragments * > &, const FragmentsList &selected_fragments, std::list< ChunkIter > &, std::list< std::shared_ptr< Chunk_NS::Chunk >> &, DeviceAllocator *device_allocator, const size_t thread_idx, const bool allow_runtime_interrupt)
std::pair< std::vector < std::vector< int64_t > >, std::vector< std::vector < uint64_t > > >	getRowCountAndOffsetForAllFrags (const RelAlgExecutionUnit &ra_exe_unit, const CartesianProduct< std::vector< std::vector< size_t >>> &frag_ids_crossjoin, const std::vector< InputDescriptor > &input_descs, const std::map< shared::TableKey, const TableFragments * > &all_tables_fragments)
void	buildSelectedFragsMapping (std::vector< std::vector< size_t >> &selected_fragments_crossjoin, std::vector< size_t > &local_col_to_frag_pos, const std::list< std::shared_ptr< const InputColDescriptor >> &col_global_ids, const FragmentsList &selected_fragments, const RelAlgExecutionUnit &ra_exe_unit)
void	buildSelectedFragsMappingForUnion (std::vector< std::vector< size_t >> &selected_fragments_crossjoin, const FragmentsList &selected_fragments, const RelAlgExecutionUnit &ra_exe_unit)
std::vector< size_t >	getFragmentCount (const FragmentsList &selected_fragments, const size_t scan_idx, const RelAlgExecutionUnit &ra_exe_unit)
int32_t	executePlanWithGroupBy (const RelAlgExecutionUnit &ra_exe_unit, const CompilationResult &, const bool hoist_literals, ResultSetPtr *results, const ExecutorDeviceType device_type, std::vector< std::vector< const int8_t * >> &col_buffers, const std::vector< size_t > outer_tab_frag_ids, QueryExecutionContext *, const std::vector< std::vector< int64_t >> &num_rows, const std::vector< std::vector< uint64_t >> &frag_offsets, Data_Namespace::DataMgr *, const int device_id, const shared::TableKey &outer_table_key, const int64_t limit, const uint32_t start_rowid, const uint32_t num_tables, const bool allow_runtime_interrupt, RenderInfo *render_info, const bool optimize_cuda_block_and_grid_sizes, const int64_t rows_to_process=-1)
int32_t	executePlanWithoutGroupBy (const RelAlgExecutionUnit &ra_exe_unit, const CompilationResult &, const bool hoist_literals, ResultSetPtr *results, const std::vector< Analyzer::Expr * > &target_exprs, const ExecutorDeviceType device_type, std::vector< std::vector< const int8_t * >> &col_buffers, QueryExecutionContext *query_exe_context, const std::vector< std::vector< int64_t >> &num_rows, const std::vector< std::vector< uint64_t >> &frag_offsets, Data_Namespace::DataMgr *data_mgr, const int device_id, const uint32_t start_rowid, const uint32_t num_tables, const bool allow_runtime_interrupt, RenderInfo *render_info, const bool optimize_cuda_block_and_grid_sizes, const int64_t rows_to_process=-1)
ResultSetPtr	resultsUnion (SharedKernelContext &shared_context, const RelAlgExecutionUnit &ra_exe_unit)
std::vector< int8_t * >	getJoinHashTablePtrs (const ExecutorDeviceType device_type, const int device_id)
ResultSetPtr	reduceMultiDeviceResults (const RelAlgExecutionUnit &, std::vector< std::pair< ResultSetPtr, std::vector< size_t >>> &all_fragment_results, std::shared_ptr< RowSetMemoryOwner >, const QueryMemoryDescriptor &) const
ResultSetPtr	reduceMultiDeviceResultSets (std::vector< std::pair< ResultSetPtr, std::vector< size_t >>> &all_fragment_results, std::shared_ptr< RowSetMemoryOwner >, const QueryMemoryDescriptor &) const
ResultSetPtr	reduceSpeculativeTopN (const RelAlgExecutionUnit &, std::vector< std::pair< ResultSetPtr, std::vector< size_t >>> &all_fragment_results, std::shared_ptr< RowSetMemoryOwner >, const QueryMemoryDescriptor &) const
ResultSetPtr	executeWorkUnitImpl (size_t &max_groups_buffer_entry_guess, const bool is_agg, const bool allow_single_frag_table_opt, const std::vector< InputTableInfo > &, const RelAlgExecutionUnit &, const CompilationOptions &, const ExecutionOptions &options, std::shared_ptr< RowSetMemoryOwner >, RenderInfo *render_info, const bool has_cardinality_estimation, ColumnCacheMap &column_cache)
std::vector< llvm::Value * >	inlineHoistedLiterals ()
void	AutoTrackBuffersInRuntimeIR ()
std::tuple< CompilationResult, std::unique_ptr < QueryMemoryDescriptor > >	compileWorkUnit (const std::vector< InputTableInfo > &query_infos, const PlanState::DeletedColumnsMap &deleted_cols_map, const RelAlgExecutionUnit &ra_exe_unit, const CompilationOptions &co, const ExecutionOptions &eo, const CudaMgr_Namespace::CudaMgr *cuda_mgr, const bool allow_lazy_fetch, std::shared_ptr< RowSetMemoryOwner >, const size_t max_groups_buffer_entry_count, const int8_t crt_min_byte_width, const bool has_cardinality_estimation, ColumnCacheMap &column_cache, RenderInfo *render_info=nullptr)
llvm::BasicBlock *	codegenSkipDeletedOuterTableRow (const RelAlgExecutionUnit &ra_exe_unit, const CompilationOptions &co)
std::vector< JoinLoop >	buildJoinLoops (RelAlgExecutionUnit &ra_exe_unit, const CompilationOptions &co, const ExecutionOptions &eo, const std::vector< InputTableInfo > &query_infos, ColumnCacheMap &column_cache)
JoinLoop::HoistedFiltersCallback	buildHoistLeftHandSideFiltersCb (const RelAlgExecutionUnit &ra_exe_unit, const size_t level_idx, const shared::TableKey &inner_table_key, const CompilationOptions &co)
std::function< llvm::Value *(const std::vector < llvm::Value * > &, llvm::Value *)>	buildIsDeletedCb (const RelAlgExecutionUnit &ra_exe_unit, const size_t level_idx, const CompilationOptions &co)
std::shared_ptr< HashJoin >	buildCurrentLevelHashTable (const JoinCondition &current_level_join_conditions, size_t level_idx, RelAlgExecutionUnit &ra_exe_unit, const CompilationOptions &co, const std::vector< InputTableInfo > &query_infos, ColumnCacheMap &column_cache, std::vector< std::string > &fail_reasons)
void	redeclareFilterFunction ()
llvm::Value *	addJoinLoopIterator (const std::vector< llvm::Value * > &prev_iters, const size_t level_idx)
void	codegenJoinLoops (const std::vector< JoinLoop > &join_loops, const RelAlgExecutionUnit &ra_exe_unit, GroupByAndAggregate &group_by_and_aggregate, llvm::Function *query_func, llvm::BasicBlock *entry_bb, QueryMemoryDescriptor &query_mem_desc, const CompilationOptions &co, const ExecutionOptions &eo)
bool	compileBody (const RelAlgExecutionUnit &ra_exe_unit, GroupByAndAggregate &group_by_and_aggregate, QueryMemoryDescriptor &query_mem_desc, const CompilationOptions &co, const GpuSharedMemoryContext &gpu_smem_context={})
void	createErrorCheckControlFlow (llvm::Function *query_func, bool run_with_dynamic_watchdog, bool run_with_allowing_runtime_interrupt, const std::vector< JoinLoop > &join_loops, ExecutorDeviceType device_type, const std::vector< InputTableInfo > &input_table_infos)
void	insertErrorCodeChecker (llvm::Function *query_func, bool hoist_literals, bool allow_runtime_query_interrupt)
void	preloadFragOffsets (const std::vector< InputDescriptor > &input_descs, const std::vector< InputTableInfo > &query_infos)
JoinHashTableOrError	buildHashTableForQualifier (const std::shared_ptr< Analyzer::BinOper > &qual_bin_oper, const std::vector< InputTableInfo > &query_infos, const MemoryLevel memory_level, const JoinType join_type, const HashType preferred_hash_type, ColumnCacheMap &column_cache, const HashTableBuildDagMap &hashtable_build_dag_map, const RegisteredQueryHint &query_hint, const TableIdToNodeMap &table_id_to_node_map)
void	nukeOldState (const bool allow_lazy_fetch, const std::vector< InputTableInfo > &query_infos, const PlanState::DeletedColumnsMap &deleted_cols_map, const RelAlgExecutionUnit *ra_exe_unit)
std::shared_ptr < CompilationContext >	optimizeAndCodegenCPU (llvm::Function *, llvm::Function *, const std::unordered_set< llvm::Function * > &, const CompilationOptions &)
std::shared_ptr < CompilationContext >	optimizeAndCodegenGPU (llvm::Function *, llvm::Function *, std::unordered_set< llvm::Function * > &, const bool no_inline, const CudaMgr_Namespace::CudaMgr *cuda_mgr, const bool is_gpu_smem_used, const CompilationOptions &)
std::string	generatePTX (const std::string &) const
void	initializeNVPTXBackend () const
int64_t	deviceCycles (int milliseconds) const
GroupColLLVMValue	groupByColumnCodegen (Analyzer::Expr *group_by_col, const size_t col_width, const CompilationOptions &, const bool translate_null_val, const int64_t translated_null_val, DiamondCodegen &, std::stack< llvm::BasicBlock * > &, const bool thread_mem_shared)
llvm::Value *	castToFP (llvm::Value *, SQLTypeInfo const &from_ti, SQLTypeInfo const &to_ti)
llvm::Value *	castToIntPtrTyIn (llvm::Value *val, const size_t bit_width)
std::tuple < RelAlgExecutionUnit, PlanState::DeletedColumnsMap >	addDeletedColumn (const RelAlgExecutionUnit &ra_exe_unit, const CompilationOptions &co)
bool	isFragmentFullyDeleted (const InputDescriptor &table_desc, const Fragmenter_Namespace::FragmentInfo &fragment)
FragmentSkipStatus	canSkipFragmentForFpQual (const Analyzer::BinOper *comp_expr, const Analyzer::ColumnVar *lhs_col, const Fragmenter_Namespace::FragmentInfo &fragment, const Analyzer::Constant *rhs_const) const
std::pair< bool, int64_t >	skipFragment (const InputDescriptor &table_desc, const Fragmenter_Namespace::FragmentInfo &frag_info, const std::list< std::shared_ptr< Analyzer::Expr >> &simple_quals, const std::vector< uint64_t > &frag_offsets, const size_t frag_idx)
std::pair< bool, int64_t >	skipFragmentInnerJoins (const InputDescriptor &table_desc, const RelAlgExecutionUnit &ra_exe_unit, const Fragmenter_Namespace::FragmentInfo &fragment, const std::vector< uint64_t > &frag_offsets, const size_t frag_idx)
AggregatedColRange	computeColRangesCache (const std::unordered_set< PhysicalInput > &phys_inputs)
StringDictionaryGenerations	computeStringDictionaryGenerations (const std::unordered_set< PhysicalInput > &phys_inputs)
TableGenerations	computeTableGenerations (const std::unordered_set< shared::TableKey > &phys_table_keys)
std::vector< int8_t >	serializeLiterals (const std::unordered_map< int, CgenState::LiteralValues > &literals, const int device_id)
const std::unique_ptr < llvm::Module > &	get_extension_module (ExtModuleKinds kind) const
bool	has_extension_module (ExtModuleKinds kind) const
llvm::Value *	spillDoubleElement (llvm::Value *elem_val, llvm::Type *elem_ty)
ExecutorMutexHolder	acquireExecuteMutex ()

Static Private Member Functions
static size_t	align (const size_t off_in, const size_t alignment)

Private Attributes
const ExecutorId	executor_id_
std::unique_ptr < llvm::LLVMContext >	context_
std::unique_ptr< CgenState >	cgen_state_
std::map< ExtModuleKinds, std::unique_ptr< llvm::Module > >	extension_modules_
std::unique_ptr< PlanState >	plan_state_
std::shared_ptr < RowSetMemoryOwner >	row_set_mem_owner_
std::mutex	gpu_exec_mutex_ [max_gpu_count]
std::atomic< bool >	interrupted_
std::mutex	str_dict_mutex_
std::unique_ptr < llvm::TargetMachine >	nvptx_target_machine_
unsigned	block_size_x_
unsigned	grid_size_x_
const size_t	max_gpu_slab_size_
const std::string	debug_dir_
const std::string	debug_file_
Data_Namespace::DataMgr *	data_mgr_
const TemporaryTables *	temporary_tables_
TableIdToNodeMap	table_id_to_node_map_
int64_t	kernel_queue_time_ms_ = 0
int64_t	compilation_queue_time_ms_ = 0
std::unique_ptr < WindowProjectNodeContext >	window_project_node_context_owned_
WindowFunctionContext *	active_window_function_ {nullptr}
InputTableInfoCache	input_table_info_cache_
AggregatedColRange	agg_col_range_cache_
TableGenerations	table_generations_
QuerySessionId	current_query_session_

Static Private Attributes
static const int	max_gpu_count {16}
static std::mutex	gpu_active_modules_mutex_
static uint32_t	gpu_active_modules_device_mask_ {0x0}
static void *	gpu_active_modules_ [max_gpu_count]
static const size_t	baseline_threshold
static heavyai::shared_mutex	executor_session_mutex_
static InterruptFlagMap	queries_interrupt_flag_
static QuerySessionMap	queries_session_map_
static std::map< int, std::shared_ptr< Executor > >	executors_
static heavyai::shared_mutex	execute_mutex_
static heavyai::shared_mutex	executors_cache_mutex_
static QueryPlanDagCache	query_plan_dag_cache_
static heavyai::shared_mutex	recycler_mutex_
static std::unordered_map < std::string, size_t >	cardinality_cache_
static ResultSetRecyclerHolder	resultset_recycler_holder_
static QueryPlanDAG	latest_query_plan_extracted_ {EMPTY_QUERY_PLAN}

Friends
class	BaselineJoinHashTable
class	CodeGenerator
class	ColumnFetcher
struct	DiamondCodegen
class	ExecutionKernel
class	KernelSubtask
class	HashJoin
class	OverlapsJoinHashTable
class	RangeJoinHashTable
class	GroupByAndAggregate
class	QueryCompilationDescriptor
class	QueryMemoryDescriptor
class	QueryMemoryInitializer
class	QueryFragmentDescriptor
class	QueryExecutionContext
class	ResultSet
class	InValuesBitmap
class	StringDictionaryTranslationMgr
class	LeafAggregator
class	PerfectJoinHashTable
class	QueryRewriter
class	PendingExecutionClosure
class	RelAlgExecutor
class	TableOptimizer
class	TableFunctionCompilationContext
class	TableFunctionExecutionContext
struct	TargetExprCodegenBuilder
struct	TargetExprCodegen
class	WindowProjectNodeContext

Detailed Description

Definition at line 365 of file Execute.h.

Member Typedef Documentation

using Executor::CachedCardinality = std::pair<bool, size_t>

Definition at line 1250 of file Execute.h.

using Executor::ExecutorId = size_t

Definition at line 372 of file Execute.h.

using Executor::PerFragmentCallBack = std::function<void(ResultSetPtr, const Fragmenter_Namespace::FragmentInfo&)>

private

Definition at line 779 of file Execute.h.

Member Enumeration Documentation

enum Executor::ExtModuleKinds

strong

Enumerator
template_module
udf_cpu_module
udf_gpu_module
rt_udf_cpu_module
rt_udf_gpu_module
rt_geos_module
rt_libdevice_module

Definition at line 466 of file Execute.h.

                            {
    template_module,     // RuntimeFunctions.bc
    udf_cpu_module,      // Load-time UDFs for CPU execution
    udf_gpu_module,      // Load-time UDFs for GPU execution
    rt_udf_cpu_module,   // Run-time UDF/UDTFs for CPU execution
    rt_udf_gpu_module,   // Run-time UDF/UDTFs for GPU execution
    rt_geos_module,      // geos functions
    rt_libdevice_module  // math library functions for GPU execution
  };

Constructor & Destructor Documentation

Executor::Executor	(	const ExecutorId	id,
		Data_Namespace::DataMgr *	data_mgr,
		const size_t	block_size_x,
		const size_t	grid_size_x,
		const size_t	max_gpu_slab_size,
		const std::string &	debug_dir,
		const std::string &	debug_file
	)

Definition at line 244 of file Execute.cpp.

    : executor_id_(executor_id)
    , context_(new llvm::LLVMContext())
    , cgen_state_(new CgenState({}, false, this))

Member Function Documentation

ExecutorMutexHolder Executor::acquireExecuteMutex ( )

inlineprivate

Definition at line 1405 of file Execute.h.

References execute_mutex_, executor_id_, Executor::ExecutorMutexHolder::shared_lock, Executor::ExecutorMutexHolder::unique_lock, and UNITARY_EXECUTOR_ID.

                                                   {
    ExecutorMutexHolder ret;
    if (executor_id_ == Executor::UNITARY_EXECUTOR_ID) {
      // Only one unitary executor can run at a time
      ret.unique_lock = heavyai::unique_lock<heavyai::shared_mutex>(execute_mutex_);
    } else {
      ret.shared_lock = heavyai::shared_lock<heavyai::shared_mutex>(execute_mutex_);
    }
    return ret;
  }

std::tuple< RelAlgExecutionUnit, PlanState::DeletedColumnsMap > Executor::addDeletedColumn ( const RelAlgExecutionUnit &  ra_exe_unit, const CompilationOptions &  co  )

private

Definition at line 3930 of file Execute.cpp.

References anonymous_namespace{Execute.cpp}::add_deleted_col_to_map(), CHECK, CompilationOptions::filter_on_deleted_column, Catalog_Namespace::SysCatalog::getCatalog(), Catalog_Namespace::SysCatalog::instance(), and TABLE.

Referenced by executeWorkUnitImpl(), and executeWorkUnitPerFragment().

                                  {
  if (!co.filter_on_deleted_column) {
    return std::make_tuple(ra_exe_unit, PlanState::DeletedColumnsMap{});
  }
  auto ra_exe_unit_with_deleted = ra_exe_unit;
  PlanState::DeletedColumnsMap deleted_cols_map;
  for (const auto& input_table : ra_exe_unit_with_deleted.input_descs) {
    if (input_table.getSourceType() != InputSourceType::TABLE) {
      continue;
    }
    const auto& table_key = input_table.getTableKey();
    const auto catalog =
        Catalog_Namespace::SysCatalog::instance().getCatalog(table_key.db_id);
    CHECK(catalog);
    const auto td = catalog->getMetadataForTable(table_key.table_id);
    CHECK(td);
    const auto deleted_cd = catalog->getDeletedColumnIfRowsDeleted(td);
    if (!deleted_cd) {
      continue;
    }
    CHECK(deleted_cd->columnType.is_boolean());
    // check deleted column is not already present
    bool found = false;
    for (const auto& input_col : ra_exe_unit_with_deleted.input_col_descs) {
      if (input_col.get()->getColId() == deleted_cd->columnId &&
          input_col.get()->getScanDesc().getTableKey() == table_key &&
          input_col.get()->getScanDesc().getNestLevel() == input_table.getNestLevel()) {
        found = true;
        add_deleted_col_to_map(deleted_cols_map, deleted_cd, table_key);
        break;
      }
    }
    if (!found) {
      // add deleted column
      ra_exe_unit_with_deleted.input_col_descs.emplace_back(
          new InputColDescriptor(deleted_cd->columnId,
                                 deleted_cd->tableId,
                                 table_key.db_id,
                                 input_table.getNestLevel()));
      add_deleted_col_to_map(deleted_cols_map, deleted_cd, table_key);
    }
  }
  return std::make_tuple(ra_exe_unit_with_deleted, deleted_cols_map);
}

Here is the call graph for this function:

Here is the caller graph for this function:

llvm::Value * Executor::addJoinLoopIterator ( const std::vector< llvm::Value * > &  prev_iters, const size_t  level_idx  )

private

Definition at line 1119 of file IRCodegen.cpp.

References AUTOMATIC_IR_METADATA, and CHECK.

                                                                   {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  // Iterators are added for loop-outer joins when the head of the loop is generated,
  // then once again when the body if generated. Allow this instead of special handling
  // of call sites.
  const auto it = cgen_state_->scan_idx_to_hash_pos_.find(level_idx);
  if (it != cgen_state_->scan_idx_to_hash_pos_.end()) {
    return it->second;
  }
  CHECK(!prev_iters.empty());
  llvm::Value* matching_row_index = prev_iters.back();
  const auto it_ok =
      cgen_state_->scan_idx_to_hash_pos_.emplace(level_idx, matching_row_index);
  CHECK(it_ok.second);
  return matching_row_index;
}

void Executor::addToCardinalityCache	(	const std::string &	cache_key,
		const size_t	cache_value
	)

Definition at line 4744 of file Execute.cpp.

References cardinality_cache_, g_use_estimator_result_cache, recycler_mutex_, and VLOG.

                                                               {
  if (g_use_estimator_result_cache) {
    heavyai::unique_lock<heavyai::shared_mutex> lock(recycler_mutex_);
    cardinality_cache_[cache_key] = cache_value;
    VLOG(1) << "Put estimated cardinality to the cache";
  }
}

bool Executor::addToQuerySessionList	(	const QuerySessionId &	query_session,
		const std::string &	query_str,
		const std::string &	submitted,
		const size_t	executor_id,
		const QuerySessionStatus::QueryStatus	query_status,
		heavyai::unique_lock< heavyai::shared_mutex > &	write_lock
	)

Definition at line 4575 of file Execute.cpp.

References queries_interrupt_flag_, and queries_session_map_.

Referenced by enrollQuerySession().

                                                         {
  // an internal API that enrolls the query session into the Executor's session map
  if (queries_session_map_.count(query_session)) {
    if (queries_session_map_.at(query_session).count(submitted_time_str)) {
      queries_session_map_.at(query_session).erase(submitted_time_str);
      queries_session_map_.at(query_session)
          .emplace(submitted_time_str,
                   QuerySessionStatus(query_session,
                                      executor_id,
                                      query_str,
                                      submitted_time_str,
                                      query_status));
    } else {
      queries_session_map_.at(query_session)
          .emplace(submitted_time_str,
                   QuerySessionStatus(query_session,
                                      executor_id,
                                      query_str,
                                      submitted_time_str,
                                      query_status));
    }
  } else {
    std::map<std::string, QuerySessionStatus> executor_per_query_map;
    executor_per_query_map.emplace(
        submitted_time_str,
        QuerySessionStatus(
            query_session, executor_id, query_str, submitted_time_str, query_status));
    queries_session_map_.emplace(query_session, executor_per_query_map);
  }
  return queries_interrupt_flag_.emplace(query_session, false).second;
}

Here is the caller graph for this function:

void Executor::addTransientStringLiterals	(	const RelAlgExecutionUnit &	ra_exe_unit,
		const std::shared_ptr< RowSetMemoryOwner > &	row_set_mem_owner
	)

Definition at line 2132 of file Execute.cpp.

References CHECK, getStringDictionaryProxy(), RelAlgExecutionUnit::groupby_exprs, kENCODING_DICT, kMODE, kSAMPLE, kSINGLE_VALUE, RelAlgExecutionUnit::quals, RelAlgExecutionUnit::simple_quals, RelAlgExecutionUnit::target_exprs, RelAlgExecutionUnit::target_exprs_union, and ScalarExprVisitor< T >::visit().

                                                               {
  TransientDictIdVisitor dict_id_visitor;

  auto visit_expr =
      [this, &dict_id_visitor, &row_set_mem_owner](const Analyzer::Expr* expr) {
        if (!expr) {
          return;
        }
        const auto& dict_key = dict_id_visitor.visit(expr);
        if (dict_key.dict_id >= 0) {
          auto sdp = getStringDictionaryProxy(dict_key, row_set_mem_owner, true);
          CHECK(sdp);
          TransientStringLiteralsVisitor visitor(sdp, this);
          visitor.visit(expr);
        }
      };

  for (const auto& group_expr : ra_exe_unit.groupby_exprs) {
    visit_expr(group_expr.get());
  }

  for (const auto& group_expr : ra_exe_unit.quals) {
    visit_expr(group_expr.get());
  }

  for (const auto& group_expr : ra_exe_unit.simple_quals) {
    visit_expr(group_expr.get());
  }

  const auto visit_target_expr = [&](const Analyzer::Expr* target_expr) {
    const auto& target_type = target_expr->get_type_info();
    if (!target_type.is_string() || target_type.get_compression() == kENCODING_DICT) {
      const auto agg_expr = dynamic_cast<const Analyzer::AggExpr*>(target_expr);
      if (agg_expr) {
        // The following agg types require taking into account transient string values
        if (agg_expr->get_is_distinct() || agg_expr->get_aggtype() == kSINGLE_VALUE ||
            agg_expr->get_aggtype() == kSAMPLE || agg_expr->get_aggtype() == kMODE) {
          visit_expr(agg_expr->get_arg());
        }
      } else {
        visit_expr(target_expr);
      }
    }
  };
  const auto& target_exprs = ra_exe_unit.target_exprs;
  std::for_each(target_exprs.begin(), target_exprs.end(), visit_target_expr);
  const auto& target_exprs_union = ra_exe_unit.target_exprs_union;
  std::for_each(target_exprs_union.begin(), target_exprs_union.end(), visit_target_expr);
}

Here is the call graph for this function:

void Executor::addUdfIrToModule ( const std::string &  udf_ir_filename, const bool  is_cuda_ir  )

static

Definition at line 1897 of file NativeCodegen.cpp.

Referenced by DBHandler::initialize().

                                                       {
  Executor::extension_module_sources[is_cuda_ir
                                         ? Executor::ExtModuleKinds::udf_gpu_module
                                         : Executor::ExtModuleKinds::udf_cpu_module] =
      udf_ir_filename;
}

Here is the caller graph for this function:

llvm::Value * Executor::aggregateWindowStatePtr ( )

private

Definition at line 226 of file WindowFunctionIR.cpp.

References AUTOMATIC_IR_METADATA, anonymous_namespace{WindowFunctionIR.cpp}::get_adjusted_window_type_info(), get_int_type(), WindowProjectNodeContext::getActiveWindowFunctionContext(), and kFLOAT.

                                             {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto window_func_context =
      WindowProjectNodeContext::getActiveWindowFunctionContext(this);
  const auto window_func = window_func_context->getWindowFunction();
  const auto arg_ti = get_adjusted_window_type_info(window_func);
  llvm::Type* aggregate_state_type =
      arg_ti.get_type() == kFLOAT
          ? llvm::PointerType::get(get_int_type(32, cgen_state_->context_), 0)
          : llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0);
  const auto aggregate_state_i64 = cgen_state_->llInt(
      reinterpret_cast<const int64_t>(window_func_context->aggregateState()));
  return cgen_state_->ir_builder_.CreateIntToPtr(aggregate_state_i64,
                                                 aggregate_state_type);
}

Here is the call graph for this function:

static size_t Executor::align ( const size_t  off_in, const size_t  alignment  )

inlinestaticprivate

Definition at line 1283 of file Execute.h.

Referenced by serializeLiterals().

                                                                   {
    size_t off = off_in;
    if (off % alignment != 0) {
      off += (alignment - off % alignment);
    }
    return off;
  }

Here is the caller graph for this function:

CurrentQueryStatus Executor::attachExecutorToQuerySession	(	const QuerySessionId &	query_session_id,
		const std::string &	query_str,
		const std::string &	query_submitted_time
	)

Definition at line 4473 of file Execute.cpp.

References executor_id_, executor_session_mutex_, updateQuerySessionExecutorAssignment(), and updateQuerySessionStatusWithLock().

                                           {
  if (!query_session_id.empty()) {
    // if session is valid, do update 1) the exact executor id and 2) query status
    heavyai::unique_lock<heavyai::shared_mutex> write_lock(executor_session_mutex_);
    updateQuerySessionExecutorAssignment(
        query_session_id, query_submitted_time, executor_id_, write_lock);
    updateQuerySessionStatusWithLock(query_session_id,
                                     query_submitted_time,
                                     QuerySessionStatus::QueryStatus::PENDING_EXECUTOR,
                                     write_lock);
  }
  return {query_session_id, query_str};
}

Here is the call graph for this function:

void Executor::AutoTrackBuffersInRuntimeIR ( )

private

Definition at line 2240 of file NativeCodegen.cpp.

                                           {
  llvm::Module* M = cgen_state_->module_;
  if (M->getFunction("allocate_varlen_buffer") == nullptr)
    return;

  // read metadata
  bool should_track = false;
  auto* flag = M->getModuleFlag("manage_memory_buffer");
  if (auto* cnt = llvm::mdconst::extract_or_null<llvm::ConstantInt>(flag)) {
    if (cnt->getZExtValue() == 1) {
      should_track = true;
    }
  }

  if (!should_track) {
    // metadata is not present
    return;
  }

  LOG(INFO) << "Found 'manage_memory_buffer' metadata.";
  llvm::SmallVector<llvm::CallInst*, 4> calls_to_analyze;

  for (llvm::Function& F : *M) {
    for (llvm::BasicBlock& BB : F) {
      for (llvm::Instruction& I : BB) {
        if (llvm::CallInst* CI = llvm::dyn_cast<llvm::CallInst>(&I)) {
          // Keep track of calls to "allocate_varlen_buffer" for later processing
          llvm::Function* called = CI->getCalledFunction();
          if (called) {
            if (called->getName() == "allocate_varlen_buffer") {
              calls_to_analyze.push_back(CI);
            }
          }
        }
      }
    }
  }

  // for each call to "allocate_varlen_buffer", check if there's a corresponding
  // call to "register_buffer_with_executor_rsm". If not, add a call to it
  llvm::IRBuilder<> Builder(cgen_state_->context_);
  auto i64 = get_int_type(64, cgen_state_->context_);
  auto i8p = get_int_ptr_type(8, cgen_state_->context_);
  auto void_ = llvm::Type::getVoidTy(cgen_state_->context_);
  llvm::FunctionType* fnty = llvm::FunctionType::get(void_, {i64, i8p}, false);
  llvm::FunctionCallee register_buffer_fn =
      M->getOrInsertFunction("register_buffer_with_executor_rsm", fnty, {});

  int64_t executor_addr = reinterpret_cast<int64_t>(this);
  for (llvm::CallInst* CI : calls_to_analyze) {
    bool found = false;
    // for each user of the function, check if its a callinst
    // and if the callinst is calling "register_buffer_with_executor_rsm"
    // if no such instruction exist, add one registering the buffer
    for (llvm::User* U : CI->users()) {
      if (llvm::CallInst* call = llvm::dyn_cast<llvm::CallInst>(U)) {
        if (call->getCalledFunction() and
            call->getCalledFunction()->getName() == "register_buffer_with_executor_rsm") {
          found = true;
          break;
        }
      }
    }
    if (!found) {
      Builder.SetInsertPoint(CI->getNextNode());
      Builder.CreateCall(register_buffer_fn,
                         {ll_int(executor_addr, cgen_state_->context_), CI});
    }
  }
}

unsigned Executor::blockSize

(

)

const

Definition at line 3821 of file Execute.cpp.

References block_size_x_, CHECK, data_mgr_, CudaMgr_Namespace::CudaMgr::getAllDeviceProperties(), and Data_Namespace::DataMgr::getCudaMgr().

Referenced by collectAllDeviceShardedTopResults(), executePlanWithGroupBy(), executePlanWithoutGroupBy(), executeTableFunction(), executeWorkUnitImpl(), reduceMultiDeviceResults(), reduceMultiDeviceResultSets(), and resultsUnion().

                                   {
  CHECK(data_mgr_);
  const auto cuda_mgr = data_mgr_->getCudaMgr();
  if (!cuda_mgr) {
    return 0;
  }
  const auto& dev_props = cuda_mgr->getAllDeviceProperties();
  return block_size_x_ ? block_size_x_ : dev_props.front().maxThreadsPerBlock;
}

Here is the call graph for this function:

Here is the caller graph for this function:

std::shared_ptr< HashJoin > Executor::buildCurrentLevelHashTable ( const JoinCondition &  current_level_join_conditions, size_t  level_idx, RelAlgExecutionUnit &  ra_exe_unit, const CompilationOptions &  co, const std::vector< InputTableInfo > &  query_infos, ColumnCacheMap &  column_cache, std::vector< std::string > &  fail_reasons  )

private

Definition at line 960 of file IRCodegen.cpp.

References anonymous_namespace{IRCodegen.cpp}::add_qualifier_to_execution_unit(), AUTOMATIC_IR_METADATA, anonymous_namespace{IRCodegen.cpp}::check_valid_join_qual(), Data_Namespace::CPU_LEVEL, CompilationOptions::device_type, Executor::JoinHashTableOrError::fail_reason, GPU, Data_Namespace::GPU_LEVEL, Executor::JoinHashTableOrError::hash_table, RelAlgExecutionUnit::hash_table_build_plan_dag, IS_EQUIVALENCE, LEFT, OneToOne, JoinCondition::quals, RelAlgExecutionUnit::query_hint, RelAlgExecutionUnit::table_id_to_node_map, JoinCondition::type, and VLOG.

                                        {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  std::shared_ptr<HashJoin> current_level_hash_table;
  auto handleNonHashtableQual = [&ra_exe_unit, &level_idx, this](
                                    JoinType join_type,
                                    std::shared_ptr<Analyzer::Expr> qual) {
    if (join_type == JoinType::LEFT) {
      plan_state_->addNonHashtableQualForLeftJoin(level_idx, qual);
    } else {
      add_qualifier_to_execution_unit(ra_exe_unit, qual);
    }
  };
  for (const auto& join_qual : current_level_join_conditions.quals) {
    auto qual_bin_oper = std::dynamic_pointer_cast<Analyzer::BinOper>(join_qual);
    if (current_level_hash_table || !qual_bin_oper ||
        !IS_EQUIVALENCE(qual_bin_oper->get_optype())) {
      handleNonHashtableQual(current_level_join_conditions.type, join_qual);
      if (!current_level_hash_table) {
        fail_reasons.emplace_back("No equijoin expression found");
      }
      continue;
    }
    check_valid_join_qual(qual_bin_oper);
    JoinHashTableOrError hash_table_or_error;
    if (!current_level_hash_table) {
      hash_table_or_error = buildHashTableForQualifier(
          qual_bin_oper,
          query_infos,
          co.device_type == ExecutorDeviceType::GPU ? MemoryLevel::GPU_LEVEL
                                                    : MemoryLevel::CPU_LEVEL,
          current_level_join_conditions.type,
          HashType::OneToOne,
          column_cache,
          ra_exe_unit.hash_table_build_plan_dag,
          ra_exe_unit.query_hint,
          ra_exe_unit.table_id_to_node_map);
      current_level_hash_table = hash_table_or_error.hash_table;
    }
    if (hash_table_or_error.hash_table) {
      plan_state_->join_info_.join_hash_tables_.push_back(hash_table_or_error.hash_table);
      plan_state_->join_info_.equi_join_tautologies_.push_back(qual_bin_oper);
    } else {
      fail_reasons.push_back(hash_table_or_error.fail_reason);
      if (!current_level_hash_table) {
        VLOG(2) << "Building a hashtable based on a qual " << qual_bin_oper->toString()
                << " fails: " << hash_table_or_error.fail_reason;
      }
      handleNonHashtableQual(current_level_join_conditions.type, qual_bin_oper);
    }
  }
  return current_level_hash_table;
}

Here is the call graph for this function:

Executor::JoinHashTableOrError Executor::buildHashTableForQualifier ( const std::shared_ptr< Analyzer::BinOper > &  qual_bin_oper, const std::vector< InputTableInfo > &  query_infos, const MemoryLevel  memory_level, const JoinType  join_type, const HashType  preferred_hash_type, ColumnCacheMap &  column_cache, const HashTableBuildDagMap &  hashtable_build_dag_map, const RegisteredQueryHint &  query_hint, const TableIdToNodeMap &  table_id_to_node_map  )

private

Definition at line 3765 of file Execute.cpp.

References deviceCountForMemoryLevel(), ERR_INTERRUPTED, g_enable_dynamic_watchdog, g_enable_overlaps_hashjoin, HashJoin::getInstance(), and interrupted_.

                                                  {
  if (!g_enable_overlaps_hashjoin && qual_bin_oper->is_overlaps_oper()) {
    return {nullptr, "Overlaps hash join disabled, attempting to fall back to loop join"};
  }
  if (g_enable_dynamic_watchdog && interrupted_.load()) {
    throw QueryExecutionError(ERR_INTERRUPTED);
  }
  try {
    auto tbl = HashJoin::getInstance(qual_bin_oper,
                                     query_infos,
                                     memory_level,
                                     join_type,
                                     preferred_hash_type,
                                     deviceCountForMemoryLevel(memory_level),
                                     column_cache,
                                     this,
                                     hashtable_build_dag_map,
                                     query_hint,
                                     table_id_to_node_map);
    return {tbl, ""};
  } catch (const HashJoinFail& e) {
    return {nullptr, e.what()};
  }
}

Here is the call graph for this function:

JoinLoop::HoistedFiltersCallback Executor::buildHoistLeftHandSideFiltersCb ( const RelAlgExecutionUnit &  ra_exe_unit, const size_t  level_idx, const shared::TableKey &  inner_table_key, const CompilationOptions &  co  )

private

Definition at line 792 of file IRCodegen.cpp.

References AUTOMATIC_IR_METADATA, CHECK, CodeGenerator::codegen(), g_enable_left_join_filter_hoisting, RelAlgExecutionUnit::join_quals, LEFT, RelAlgExecutionUnit::quals, RelAlgExecutionUnit::simple_quals, CodeGenerator::toBool(), and VLOG.

                                  {
  if (!g_enable_left_join_filter_hoisting) {
    return nullptr;
  }

  const auto& current_level_join_conditions = ra_exe_unit.join_quals[level_idx];
  if (level_idx == 0 && current_level_join_conditions.type == JoinType::LEFT) {
    const auto& condition = current_level_join_conditions.quals.front();
    const auto bin_oper = dynamic_cast<const Analyzer::BinOper*>(condition.get());
    CHECK(bin_oper) << condition->toString();
    const auto rhs =
        dynamic_cast<const Analyzer::ColumnVar*>(bin_oper->get_right_operand());
    const auto lhs =
        dynamic_cast<const Analyzer::ColumnVar*>(bin_oper->get_left_operand());
    if (lhs && rhs && lhs->getTableKey() != rhs->getTableKey()) {
      const Analyzer::ColumnVar* selected_lhs{nullptr};
      // grab the left hand side column -- this is somewhat similar to normalize column
      // pair, and a better solution may be to hoist that function out of the join
      // framework and normalize columns at the top of build join loops
      if (lhs->getTableKey() == inner_table_id) {
        selected_lhs = rhs;
      } else if (rhs->getTableKey() == inner_table_id) {
        selected_lhs = lhs;
      }
      if (selected_lhs) {
        std::list<std::shared_ptr<Analyzer::Expr>> hoisted_quals;
        // get all LHS-only filters
        auto should_hoist_qual = [&hoisted_quals](const auto& qual,
                                                  const shared::TableKey& table_key) {
          CHECK(qual);

          ExprTableIdVisitor visitor;
          const auto table_keys = visitor.visit(qual.get());
          if (table_keys.size() == 1 && table_keys.find(table_key) != table_keys.end()) {
            hoisted_quals.push_back(qual);
          }
        };
        for (const auto& qual : ra_exe_unit.simple_quals) {
          should_hoist_qual(qual, selected_lhs->getTableKey());
        }
        for (const auto& qual : ra_exe_unit.quals) {
          should_hoist_qual(qual, selected_lhs->getTableKey());
        }

        // build the filters callback and return it
        if (!hoisted_quals.empty()) {
          return [this, hoisted_quals, co](llvm::BasicBlock* true_bb,
                                           llvm::BasicBlock* exit_bb,
                                           const std::string& loop_name,
                                           llvm::Function* parent_func,
                                           CgenState* cgen_state) -> llvm::BasicBlock* {
            // make sure we have quals to hoist
            bool has_quals_to_hoist = false;
            for (const auto& qual : hoisted_quals) {
              // check to see if the filter was previously hoisted. if all filters were
              // previously hoisted, this callback becomes a noop
              if (plan_state_->hoisted_filters_.count(qual) == 0) {
                has_quals_to_hoist = true;
                break;
              }
            }

            if (!has_quals_to_hoist) {
              return nullptr;
            }

            AUTOMATIC_IR_METADATA(cgen_state);

            llvm::IRBuilder<>& builder = cgen_state->ir_builder_;
            auto& context = builder.getContext();

            const auto filter_bb =
                llvm::BasicBlock::Create(context,
                                         "hoisted_left_join_filters_" + loop_name,
                                         parent_func,
                                         /*insert_before=*/true_bb);
            builder.SetInsertPoint(filter_bb);

            llvm::Value* filter_lv = cgen_state_->llBool(true);
            CodeGenerator code_generator(this);
            CHECK(plan_state_);
            for (const auto& qual : hoisted_quals) {
              if (plan_state_->hoisted_filters_.insert(qual).second) {
                // qual was inserted into the hoisted filters map, which means we have not
                // seen this qual before. Generate filter.
                VLOG(1) << "Generating code for hoisted left hand side qualifier "
                        << qual->toString();
                auto cond = code_generator.toBool(
                    code_generator.codegen(qual.get(), true, co).front());
                filter_lv = builder.CreateAnd(filter_lv, cond);
              }
            }
            CHECK(filter_lv->getType()->isIntegerTy(1));

            builder.CreateCondBr(filter_lv, true_bb, exit_bb);
            return filter_bb;
          };
        }
      }
    }
  }
  return nullptr;
}

Here is the call graph for this function:

std::function< llvm::Value *(const std::vector< llvm::Value * > &, llvm::Value *)> Executor::buildIsDeletedCb ( const RelAlgExecutionUnit &  ra_exe_unit, const size_t  level_idx, const CompilationOptions &  co  )

private

Definition at line 901 of file IRCodegen.cpp.

References AUTOMATIC_IR_METADATA, CHECK, CHECK_LT, CodeGenerator::codegen(), CompilationOptions::filter_on_deleted_column, RelAlgExecutionUnit::input_descs, TABLE, and CodeGenerator::toBool().

                                                         {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  if (!co.filter_on_deleted_column) {
    return nullptr;
  }
  CHECK_LT(level_idx + 1, ra_exe_unit.input_descs.size());
  const auto input_desc = ra_exe_unit.input_descs[level_idx + 1];
  if (input_desc.getSourceType() != InputSourceType::TABLE) {
    return nullptr;
  }

  const auto deleted_cd = plan_state_->getDeletedColForTable(input_desc.getTableKey());
  if (!deleted_cd) {
    return nullptr;
  }
  CHECK(deleted_cd->columnType.is_boolean());
  const auto deleted_expr = makeExpr<Analyzer::ColumnVar>(
      deleted_cd->columnType,
      shared::ColumnKey{input_desc.getTableKey(), deleted_cd->columnId},
      input_desc.getNestLevel());
  return [this, deleted_expr, level_idx, &co](const std::vector<llvm::Value*>& prev_iters,
                                              llvm::Value* have_more_inner_rows) {
    const auto matching_row_index = addJoinLoopIterator(prev_iters, level_idx + 1);
    // Avoid fetching the deleted column from a position which is not valid.
    // An invalid position can be returned by a one to one hash lookup (negative)
    // or at the end of iteration over a set of matching values.
    llvm::Value* is_valid_it{nullptr};
    if (have_more_inner_rows) {
      is_valid_it = have_more_inner_rows;
    } else {
      is_valid_it = cgen_state_->ir_builder_.CreateICmp(
          llvm::ICmpInst::ICMP_SGE, matching_row_index, cgen_state_->llInt<int64_t>(0));
    }
    const auto it_valid_bb = llvm::BasicBlock::Create(
        cgen_state_->context_, "it_valid", cgen_state_->current_func_);
    const auto it_not_valid_bb = llvm::BasicBlock::Create(
        cgen_state_->context_, "it_not_valid", cgen_state_->current_func_);
    cgen_state_->ir_builder_.CreateCondBr(is_valid_it, it_valid_bb, it_not_valid_bb);
    const auto row_is_deleted_bb = llvm::BasicBlock::Create(
        cgen_state_->context_, "row_is_deleted", cgen_state_->current_func_);
    cgen_state_->ir_builder_.SetInsertPoint(it_valid_bb);
    CodeGenerator code_generator(this);
    const auto row_is_deleted = code_generator.toBool(
        code_generator.codegen(deleted_expr.get(), true, co).front());
    cgen_state_->ir_builder_.CreateBr(row_is_deleted_bb);
    cgen_state_->ir_builder_.SetInsertPoint(it_not_valid_bb);
    const auto row_is_deleted_default = cgen_state_->llBool(false);
    cgen_state_->ir_builder_.CreateBr(row_is_deleted_bb);
    cgen_state_->ir_builder_.SetInsertPoint(row_is_deleted_bb);
    auto row_is_deleted_or_default =
        cgen_state_->ir_builder_.CreatePHI(row_is_deleted->getType(), 2);
    row_is_deleted_or_default->addIncoming(row_is_deleted, it_valid_bb);
    row_is_deleted_or_default->addIncoming(row_is_deleted_default, it_not_valid_bb);
    return row_is_deleted_or_default;
  };
}

Here is the call graph for this function:

std::vector< JoinLoop > Executor::buildJoinLoops ( RelAlgExecutionUnit &  ra_exe_unit, const CompilationOptions &  co, const ExecutionOptions &  eo, const std::vector< InputTableInfo > &  query_infos, ColumnCacheMap &  column_cache  )

private

Definition at line 544 of file IRCodegen.cpp.

References AUTOMATIC_IR_METADATA, CodeGenerator::cgen_state_, CHECK, CHECK_LT, CodeGenerator::codegen(), INJECT_TIMER, CgenState::ir_builder_, RelAlgExecutionUnit::join_quals, LEFT, PlanState::left_join_non_hashtable_quals_, CgenState::llBool(), MultiSet, OneToOne, CgenState::outer_join_match_found_per_level_, CodeGenerator::plan_state_, Set, Singleton, JoinLoopDomain::slot_lookup_result, CodeGenerator::toBool(), and JoinLoopDomain::values_buffer.

                                  {
  INJECT_TIMER(buildJoinLoops);
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  std::vector<JoinLoop> join_loops;
  for (size_t level_idx = 0, current_hash_table_idx = 0;
       level_idx < ra_exe_unit.join_quals.size();
       ++level_idx) {
    const auto& current_level_join_conditions = ra_exe_unit.join_quals[level_idx];
    std::vector<std::string> fail_reasons;
    const auto current_level_hash_table =
        buildCurrentLevelHashTable(current_level_join_conditions,
                                   level_idx,
                                   ra_exe_unit,
                                   co,
                                   query_infos,
                                   column_cache,
                                   fail_reasons);
    const auto found_outer_join_matches_cb =
        [this, level_idx](llvm::Value* found_outer_join_matches) {
          CHECK_LT(level_idx, cgen_state_->outer_join_match_found_per_level_.size());
          CHECK(!cgen_state_->outer_join_match_found_per_level_[level_idx]);
          cgen_state_->outer_join_match_found_per_level_[level_idx] =
              found_outer_join_matches;
        };
    const auto is_deleted_cb = buildIsDeletedCb(ra_exe_unit, level_idx, co);
    auto rem_left_join_quals_it =
        plan_state_->left_join_non_hashtable_quals_.find(level_idx);
    bool has_remaining_left_join_quals =
        rem_left_join_quals_it != plan_state_->left_join_non_hashtable_quals_.end() &&
        !rem_left_join_quals_it->second.empty();
    const auto outer_join_condition_remaining_quals_cb =
        [this, level_idx, &co](const std::vector<llvm::Value*>& prev_iters) {
          // when we have multiple quals for the left join in the current join level
          // we first try to build a hashtable by using one of the possible qual,
          // and deal with remaining quals as extra join conditions
          FetchCacheAnchor anchor(cgen_state_.get());
          addJoinLoopIterator(prev_iters, level_idx + 1);
          llvm::Value* left_join_cond = cgen_state_->llBool(true);
          CodeGenerator code_generator(this);
          auto it = plan_state_->left_join_non_hashtable_quals_.find(level_idx);
          if (it != plan_state_->left_join_non_hashtable_quals_.end()) {
            for (auto expr : it->second) {
              left_join_cond = cgen_state_->ir_builder_.CreateAnd(
                  left_join_cond,
                  code_generator.toBool(
                      code_generator.codegen(expr.get(), true, co).front()));
            }
          }
          return left_join_cond;
        };
    if (current_level_hash_table) {
      const auto hoisted_filters_cb = buildHoistLeftHandSideFiltersCb(
          ra_exe_unit, level_idx, current_level_hash_table->getInnerTableId(), co);
      if (current_level_hash_table->getHashType() == HashType::OneToOne) {
        join_loops.emplace_back(
            /*kind=*/JoinLoopKind::Singleton,
            /*type=*/current_level_join_conditions.type,
            /*iteration_domain_codegen=*/
            [this, current_hash_table_idx, level_idx, current_level_hash_table, &co](
                const std::vector<llvm::Value*>& prev_iters) {
              addJoinLoopIterator(prev_iters, level_idx);
              JoinLoopDomain domain{{0}};
              domain.slot_lookup_result =
                  current_level_hash_table->codegenSlot(co, current_hash_table_idx);
              return domain;
            },
            /*outer_condition_match=*/
            current_level_join_conditions.type == JoinType::LEFT &&
                    has_remaining_left_join_quals
                ? std::function<llvm::Value*(const std::vector<llvm::Value*>&)>(
                      outer_join_condition_remaining_quals_cb)
                : nullptr,
            /*found_outer_matches=*/current_level_join_conditions.type == JoinType::LEFT
                ? std::function<void(llvm::Value*)>(found_outer_join_matches_cb)
                : nullptr,
            /*hoisted_filters=*/hoisted_filters_cb,
            /*is_deleted=*/is_deleted_cb,
            /*nested_loop_join=*/false);
      } else if (auto range_join_table =
                     dynamic_cast<RangeJoinHashTable*>(current_level_hash_table.get())) {
        join_loops.emplace_back(
            /* kind= */ JoinLoopKind::MultiSet,
            /* type= */ current_level_join_conditions.type,
            /* iteration_domain_codegen= */
            [this,
             range_join_table,
             current_hash_table_idx,
             level_idx,
             current_level_hash_table,
             &co](const std::vector<llvm::Value*>& prev_iters) {
              addJoinLoopIterator(prev_iters, level_idx);
              JoinLoopDomain domain{{0}};
              CHECK(!prev_iters.empty());
              const auto matching_set = range_join_table->codegenMatchingSetWithOffset(
                  co, current_hash_table_idx, prev_iters.back());
              domain.values_buffer = matching_set.elements;
              domain.element_count = matching_set.count;
              return domain;
            },
            /* outer_condition_match= */
            current_level_join_conditions.type == JoinType::LEFT
                ? std::function<llvm::Value*(const std::vector<llvm::Value*>&)>(
                      outer_join_condition_remaining_quals_cb)
                : nullptr,
            /* found_outer_matches= */
            current_level_join_conditions.type == JoinType::LEFT
                ? std::function<void(llvm::Value*)>(found_outer_join_matches_cb)
                : nullptr,
            /* hoisted_filters= */ nullptr,  // <<! TODO
            /* is_deleted= */ is_deleted_cb,
            /*nested_loop_join=*/false);
      } else {
        join_loops.emplace_back(
            /*kind=*/JoinLoopKind::Set,
            /*type=*/current_level_join_conditions.type,
            /*iteration_domain_codegen=*/
            [this, current_hash_table_idx, level_idx, current_level_hash_table, &co](
                const std::vector<llvm::Value*>& prev_iters) {
              addJoinLoopIterator(prev_iters, level_idx);
              JoinLoopDomain domain{{0}};
              const auto matching_set = current_level_hash_table->codegenMatchingSet(
                  co, current_hash_table_idx);
              domain.values_buffer = matching_set.elements;
              domain.element_count = matching_set.count;
              return domain;
            },
            /*outer_condition_match=*/
            current_level_join_conditions.type == JoinType::LEFT
                ? std::function<llvm::Value*(const std::vector<llvm::Value*>&)>(
                      outer_join_condition_remaining_quals_cb)
                : nullptr,
            /*found_outer_matches=*/current_level_join_conditions.type == JoinType::LEFT
                ? std::function<void(llvm::Value*)>(found_outer_join_matches_cb)
                : nullptr,
            /*hoisted_filters=*/hoisted_filters_cb,
            /*is_deleted=*/is_deleted_cb,
            /*nested_loop_join=*/false);
      }
      ++current_hash_table_idx;
    } else {
      const auto fail_reasons_str = current_level_join_conditions.quals.empty()
                                        ? "No equijoin expression found"
                                        : boost::algorithm::join(fail_reasons, " | ");
      check_if_loop_join_is_allowed(
          ra_exe_unit, eo, query_infos, level_idx, fail_reasons_str);
      // Callback provided to the `JoinLoop` framework to evaluate the (outer) join
      // condition.
      VLOG(1) << "Unable to build hash table, falling back to loop join: "
              << fail_reasons_str;
      const auto outer_join_condition_cb =
          [this, level_idx, &co, &current_level_join_conditions](
              const std::vector<llvm::Value*>& prev_iters) {
            // The values generated for the match path don't dominate all uses
            // since on the non-match path nulls are generated. Reset the cache
            // once the condition is generated to avoid incorrect reuse.
            FetchCacheAnchor anchor(cgen_state_.get());
            addJoinLoopIterator(prev_iters, level_idx + 1);
            llvm::Value* left_join_cond = cgen_state_->llBool(true);
            CodeGenerator code_generator(this);
            for (auto expr : current_level_join_conditions.quals) {
              left_join_cond = cgen_state_->ir_builder_.CreateAnd(
                  left_join_cond,
                  code_generator.toBool(
                      code_generator.codegen(expr.get(), true, co).front()));
            }
            return left_join_cond;
          };
      join_loops.emplace_back(
          /*kind=*/JoinLoopKind::UpperBound,
          /*type=*/current_level_join_conditions.type,
          /*iteration_domain_codegen=*/
          [this, level_idx](const std::vector<llvm::Value*>& prev_iters) {
            addJoinLoopIterator(prev_iters, level_idx);
            JoinLoopDomain domain{{0}};
            auto* arg = get_arg_by_name(cgen_state_->row_func_, "num_rows_per_scan");
            const auto rows_per_scan_ptr = cgen_state_->ir_builder_.CreateGEP(
                arg->getType()->getScalarType()->getPointerElementType(),
                arg,
                cgen_state_->llInt(int32_t(level_idx + 1)));
            domain.upper_bound = cgen_state_->ir_builder_.CreateLoad(
                rows_per_scan_ptr->getType()->getPointerElementType(),
                rows_per_scan_ptr,
                "num_rows_per_scan");
            return domain;
          },
          /*outer_condition_match=*/
          current_level_join_conditions.type == JoinType::LEFT
              ? std::function<llvm::Value*(const std::vector<llvm::Value*>&)>(
                    outer_join_condition_cb)
              : nullptr,
          /*found_outer_matches=*/
          current_level_join_conditions.type == JoinType::LEFT
              ? std::function<void(llvm::Value*)>(found_outer_join_matches_cb)
              : nullptr,
          /*hoisted_filters=*/nullptr,
          /*is_deleted=*/is_deleted_cb,
          /*nested_loop_join=*/true);
    }
  }
  return join_loops;
}

Here is the call graph for this function:

void Executor::buildSelectedFragsMapping ( std::vector< std::vector< size_t >> &  selected_fragments_crossjoin, std::vector< size_t > &  local_col_to_frag_pos, const std::list< std::shared_ptr< const InputColDescriptor >> &  col_global_ids, const FragmentsList &  selected_fragments, const RelAlgExecutionUnit &  ra_exe_unit  )

private

Definition at line 3234 of file Execute.cpp.

References CHECK, CHECK_EQ, CHECK_LT, getFragmentCount(), RelAlgExecutionUnit::input_descs, and plan_state_.

Referenced by fetchChunks().

                                            {
  local_col_to_frag_pos.resize(plan_state_->global_to_local_col_ids_.size());
  size_t frag_pos{0};
  const auto& input_descs = ra_exe_unit.input_descs;
  for (size_t scan_idx = 0; scan_idx < input_descs.size(); ++scan_idx) {
    const auto& table_key = input_descs[scan_idx].getTableKey();
    CHECK_EQ(selected_fragments[scan_idx].table_key, table_key);
    selected_fragments_crossjoin.push_back(
        getFragmentCount(selected_fragments, scan_idx, ra_exe_unit));
    for (const auto& col_id : col_global_ids) {
      CHECK(col_id);
      const auto& input_desc = col_id->getScanDesc();
      if (input_desc.getTableKey() != table_key ||
          input_desc.getNestLevel() != static_cast<int>(scan_idx)) {
        continue;
      }
      auto it = plan_state_->global_to_local_col_ids_.find(*col_id);
      CHECK(it != plan_state_->global_to_local_col_ids_.end());
      CHECK_LT(static_cast<size_t>(it->second),
               plan_state_->global_to_local_col_ids_.size());
      local_col_to_frag_pos[it->second] = frag_pos;
    }
    ++frag_pos;
  }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Executor::buildSelectedFragsMappingForUnion ( std::vector< std::vector< size_t >> &  selected_fragments_crossjoin, const FragmentsList &  selected_fragments, const RelAlgExecutionUnit &  ra_exe_unit  )

private

Definition at line 3265 of file Execute.cpp.

References RelAlgExecutionUnit::input_descs.

Referenced by fetchUnionChunks().

                                            {
  const auto& input_descs = ra_exe_unit.input_descs;
  for (size_t scan_idx = 0; scan_idx < input_descs.size(); ++scan_idx) {
    // selected_fragments is set in assignFragsToKernelDispatch execution_kernel.fragments
    if (selected_fragments[0].table_key == input_descs[scan_idx].getTableKey()) {
      selected_fragments_crossjoin.push_back({size_t(1)});
    }
  }
}

Here is the caller graph for this function:

FragmentSkipStatus Executor::canSkipFragmentForFpQual ( const Analyzer::BinOper *  comp_expr, const Analyzer::ColumnVar *  lhs_col, const Fragmenter_Namespace::FragmentInfo &  fragment, const Analyzer::Constant *  rhs_const  ) const

private

Definition at line 4053 of file Execute.cpp.

References CHECK, shared::ColumnKey::column_id, extract_max_stat_fp_type(), extract_min_stat_fp_type(), Analyzer::Constant::get_constval(), Analyzer::BinOper::get_optype(), SQLTypeInfo::get_type(), Analyzer::Expr::get_type_info(), Fragmenter_Namespace::FragmentInfo::getChunkMetadataMap(), Analyzer::ColumnVar::getColumnKey(), INVALID, kDOUBLE, kEQ, kFLOAT, kGE, kGT, kLE, kLT, NOT_SKIPPABLE, and SKIPPABLE.

Referenced by skipFragment().

                                             {
  auto col_id = lhs_col->getColumnKey().column_id;
  auto chunk_meta_it = fragment.getChunkMetadataMap().find(col_id);
  if (chunk_meta_it == fragment.getChunkMetadataMap().end()) {
    return FragmentSkipStatus::NOT_SKIPPABLE;
  }
  double chunk_min{0.};
  double chunk_max{0.};
  const auto& chunk_type = lhs_col->get_type_info();
  chunk_min = extract_min_stat_fp_type(chunk_meta_it->second->chunkStats, chunk_type);
  chunk_max = extract_max_stat_fp_type(chunk_meta_it->second->chunkStats, chunk_type);
  if (chunk_min > chunk_max) {
    return FragmentSkipStatus::INVALID;
  }

  const auto datum_fp = rhs_const->get_constval();
  const auto rhs_type = rhs_const->get_type_info().get_type();
  CHECK(rhs_type == kFLOAT || rhs_type == kDOUBLE);

  // Do we need to codegen the constant like the integer path does?
  const auto rhs_val = rhs_type == kFLOAT ? datum_fp.floatval : datum_fp.doubleval;

  // Todo: dedup the following comparison code with the integer/timestamp path, it is
  // slightly tricky due to do cleanly as we do not have rowid on this path
  switch (comp_expr->get_optype()) {
    case kGE:
      if (chunk_max < rhs_val) {
        return FragmentSkipStatus::SKIPPABLE;
      }
      break;
    case kGT:
      if (chunk_max <= rhs_val) {
        return FragmentSkipStatus::SKIPPABLE;
      }
      break;
    case kLE:
      if (chunk_min > rhs_val) {
        return FragmentSkipStatus::SKIPPABLE;
      }
      break;
    case kLT:
      if (chunk_min >= rhs_val) {
        return FragmentSkipStatus::SKIPPABLE;
      }
      break;
    case kEQ:
      if (chunk_min > rhs_val || chunk_max < rhs_val) {
        return FragmentSkipStatus::SKIPPABLE;
      }
      break;
    default:
      break;
  }
  return FragmentSkipStatus::NOT_SKIPPABLE;
}

Here is the call graph for this function:

Here is the caller graph for this function:

llvm::Value * Executor::castToFP ( llvm::Value *  value, SQLTypeInfo const &  from_ti, SQLTypeInfo const &  to_ti  )

private

Definition at line 3856 of file Execute.cpp.

References AUTOMATIC_IR_METADATA, cgen_state_, exp_to_scale(), logger::FATAL, SQLTypeInfo::get_scale(), SQLTypeInfo::get_size(), SQLTypeInfo::is_fp(), SQLTypeInfo::is_number(), and LOG.

                                                          {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  if (value->getType()->isIntegerTy() && from_ti.is_number() && to_ti.is_fp() &&
      (!from_ti.is_fp() || from_ti.get_size() != to_ti.get_size())) {
    llvm::Type* fp_type{nullptr};
    switch (to_ti.get_size()) {
      case 4:
        fp_type = llvm::Type::getFloatTy(cgen_state_->context_);
        break;
      case 8:
        fp_type = llvm::Type::getDoubleTy(cgen_state_->context_);
        break;
      default:
        LOG(FATAL) << "Unsupported FP size: " << to_ti.get_size();
    }
    value = cgen_state_->ir_builder_.CreateSIToFP(value, fp_type);
    if (from_ti.get_scale()) {
      value = cgen_state_->ir_builder_.CreateFDiv(
          value,
          llvm::ConstantFP::get(value->getType(), exp_to_scale(from_ti.get_scale())));
    }
  }
  return value;
}

Here is the call graph for this function:

llvm::Value * Executor::castToIntPtrTyIn ( llvm::Value *  val, const size_t  bit_width  )

private

Definition at line 3883 of file Execute.cpp.

References AUTOMATIC_IR_METADATA, cgen_state_, CHECK, CHECK_LT, and get_int_type().

                                                                           {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  CHECK(val->getType()->isPointerTy());

  const auto val_ptr_type = static_cast<llvm::PointerType*>(val->getType());
  const auto val_type = val_ptr_type->getPointerElementType();
  size_t val_width = 0;
  if (val_type->isIntegerTy()) {
    val_width = val_type->getIntegerBitWidth();
  } else {
    if (val_type->isFloatTy()) {
      val_width = 32;
    } else {
      CHECK(val_type->isDoubleTy());
      val_width = 64;
    }
  }
  CHECK_LT(size_t(0), val_width);
  if (bitWidth == val_width) {
    return val;
  }
  return cgen_state_->ir_builder_.CreateBitCast(
      val, llvm::PointerType::get(get_int_type(bitWidth, cgen_state_->context_), 0));
}

Here is the call graph for this function:

bool Executor::checkCurrentQuerySession	(	const std::string &	candidate_query_session,
		heavyai::shared_lock< heavyai::shared_mutex > &	read_lock
	)

Definition at line 4446 of file Execute.cpp.

References current_query_session_.

                                                        {
  // if current_query_session is equal to the candidate_query_session,
  // or it is empty session we consider
  return !candidate_query_session.empty() &&
         (current_query_session_ == candidate_query_session);
}

bool Executor::checkIsQuerySessionEnrolled	(	const QuerySessionId &	query_session,
		heavyai::shared_lock< heavyai::shared_mutex > &	read_lock
	)

Definition at line 4720 of file Execute.cpp.

References queries_session_map_.

Referenced by executeWorkUnitImpl().

                                                        {
  if (query_session.empty()) {
    return false;
  }
  return !query_session.empty() && queries_session_map_.count(query_session);
}

Here is the caller graph for this function:

bool Executor::checkIsQuerySessionInterrupted	(	const std::string &	query_session,
		heavyai::shared_lock< heavyai::shared_mutex > &	read_lock
	)

Definition at line 4709 of file Execute.cpp.

References queries_interrupt_flag_.

Referenced by executePlanWithGroupBy(), executePlanWithoutGroupBy(), fetchChunks(), and fetchUnionChunks().

                                                        {
  if (query_session.empty()) {
    return false;
  }
  auto flag_it = queries_interrupt_flag_.find(query_session);
  return !query_session.empty() && flag_it != queries_interrupt_flag_.end() &&
         flag_it->second;
}

Here is the caller graph for this function:

bool Executor::checkNonKernelTimeInterrupted

(

)

const

Definition at line 4797 of file Execute.cpp.

References current_query_session_, executor_id_, executor_session_mutex_, queries_interrupt_flag_, and UNITARY_EXECUTOR_ID.

                                                   {
  // this function should be called within an executor which is assigned
  // to the specific query thread (that indicates we already enroll the session)
  // check whether this is called from non unitary executor
  if (executor_id_ == UNITARY_EXECUTOR_ID) {
    return false;
  };
  heavyai::shared_lock<heavyai::shared_mutex> session_read_lock(executor_session_mutex_);
  auto flag_it = queries_interrupt_flag_.find(current_query_session_);
  return !current_query_session_.empty() && flag_it != queries_interrupt_flag_.end() &&
         flag_it->second;
}

void Executor::checkPendingQueryStatus

(

const QuerySessionId &

query_session

)

Definition at line 4490 of file Execute.cpp.

References ERR_INTERRUPTED, executor_session_mutex_, queries_interrupt_flag_, queries_session_map_, and VLOG.

                                                                          {
  // check whether we are okay to execute the "pending" query
  // i.e., before running the query check if this query session is "ALREADY" interrupted
  heavyai::shared_lock<heavyai::shared_mutex> session_read_lock(executor_session_mutex_);
  if (query_session.empty()) {
    return;
  }
  if (queries_interrupt_flag_.find(query_session) == queries_interrupt_flag_.end()) {
    // something goes wrong since we assume this is caller's responsibility
    // (call this function only for enrolled query session)
    if (!queries_session_map_.count(query_session)) {
      VLOG(1) << "Interrupting pending query is not available since the query session is "
                 "not enrolled";
    } else {
      // here the query session is enrolled but the interrupt flag is not registered
      VLOG(1)
          << "Interrupting pending query is not available since its interrupt flag is "
             "not registered";
    }
    return;
  }
  if (queries_interrupt_flag_[query_session]) {
    throw QueryExecutionError(Executor::ERR_INTERRUPTED);
  }
}

void Executor::clearCaches

(

bool

runtime_only = false

)

static void Executor::clearExternalCaches ( bool  for_update, const TableDescriptor *  td, const int  current_db_id  )

inlinestatic

Definition at line 388 of file Execute.h.

References TableDescriptor::getTableChunkKey(), CacheInvalidator< CACHE_HOLDING_TYPES >::invalidateCaches(), and CacheInvalidator< CACHE_HOLDING_TYPES >::invalidateCachesByTable().

Referenced by Parser::InsertIntoTableAsSelectStmt::execute(), Parser::DropTableStmt::execute(), Parser::TruncateTableStmt::execute(), Parser::DropColumnStmt::execute(), Parser::CopyTableStmt::execute(), RelAlgExecutor::executeDelete(), and RelAlgExecutor::executeUpdate().

                                                           {
    bool clearEntireCache = true;
    if (td) {
      const auto& table_chunk_key_prefix = td->getTableChunkKey(current_db_id);
      if (!table_chunk_key_prefix.empty()) {
        auto table_key = boost::hash_value(table_chunk_key_prefix);
        ResultSetCacheInvalidator::invalidateCachesByTable(table_key);
        if (for_update) {
          UpdateTriggeredCacheInvalidator::invalidateCachesByTable(table_key);
        } else {
          DeleteTriggeredCacheInvalidator::invalidateCachesByTable(table_key);
        }
        clearEntireCache = false;
      }
    }
    if (clearEntireCache) {
      ResultSetCacheInvalidator::invalidateCaches();
      if (for_update) {
        UpdateTriggeredCacheInvalidator::invalidateCaches();
      } else {
        DeleteTriggeredCacheInvalidator::invalidateCaches();
      }
    }
  }

Here is the call graph for this function:

Here is the caller graph for this function:

void Executor::clearMemory ( const Data_Namespace::MemoryLevel  memory_level )

static

Definition at line 497 of file Execute.cpp.

References Data_Namespace::DataMgr::clearMemory(), Data_Namespace::CPU_LEVEL, execute_mutex_, Catalog_Namespace::SysCatalog::getDataMgr(), Data_Namespace::GPU_LEVEL, Catalog_Namespace::SysCatalog::instance(), and CacheInvalidator< CACHE_HOLDING_TYPES >::invalidateCaches().

Referenced by DBHandler::clear_cpu_memory(), DBHandler::clear_gpu_memory(), QueryRunner::QueryRunner::clearCpuMemory(), and QueryRunner::QueryRunner::clearGpuMemory().

                                                                       {
  switch (memory_level) {
    case Data_Namespace::MemoryLevel::CPU_LEVEL:
    case Data_Namespace::MemoryLevel::GPU_LEVEL: {
      heavyai::unique_lock<heavyai::shared_mutex> flush_lock(
          execute_mutex_);  // Don't flush memory while queries are running

      if (memory_level == Data_Namespace::MemoryLevel::CPU_LEVEL) {
        // The hash table cache uses CPU memory not managed by the buffer manager. In the
        // future, we should manage these allocations with the buffer manager directly.
        // For now, assume the user wants to purge the hash table cache when they clear
        // CPU memory (currently used in ExecuteTest to lower memory pressure)
        JoinHashTableCacheInvalidator::invalidateCaches();
      }
      ResultSetCacheInvalidator::invalidateCaches();
      Catalog_Namespace::SysCatalog::instance().getDataMgr().clearMemory(memory_level);
      break;
    }
    default: {
      throw std::runtime_error(
          "Clearing memory levels other than the CPU level or GPU level is not "
          "supported.");
    }
  }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Executor::clearMetaInfoCache ( )

private

Definition at line 771 of file Execute.cpp.

References agg_col_range_cache_, TableGenerations::clear(), AggregatedColRange::clear(), InputTableInfoCache::clear(), input_table_info_cache_, and table_generations_.

                                  {
  input_table_info_cache_.clear();
  agg_col_range_cache_.clear();
  table_generations_.clear();
}

Here is the call graph for this function:

void Executor::clearQuerySessionStatus	(	const QuerySessionId &	query_session,
		const std::string &	submitted_time_str
	)

Definition at line 4516 of file Execute.cpp.

References current_query_session_, executor_session_mutex_, invalidateRunningQuerySession(), removeFromQuerySessionList(), and resetInterrupt().

                                                                            {
  heavyai::unique_lock<heavyai::shared_mutex> session_write_lock(executor_session_mutex_);
  // clear the interrupt-related info for a finished query
  if (query_session.empty()) {
    return;
  }
  removeFromQuerySessionList(query_session, submitted_time_str, session_write_lock);
  if (query_session.compare(current_query_session_) == 0) {
    invalidateRunningQuerySession(session_write_lock);
    resetInterrupt();
  }
}

Here is the call graph for this function:

llvm::Value * Executor::codegenAggregateWindowState ( )

private

Definition at line 1335 of file WindowFunctionIR.cpp.

References AUTOMATIC_IR_METADATA, AVG, COUNT, anonymous_namespace{WindowFunctionIR.cpp}::get_adjusted_window_type_info(), get_int_type(), WindowProjectNodeContext::getActiveWindowFunctionContext(), Analyzer::WindowFunction::getKind(), kDECIMAL, kDOUBLE, and kFLOAT.

                                                 {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto pi32_type =
      llvm::PointerType::get(get_int_type(32, cgen_state_->context_), 0);
  const auto pi64_type =
      llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0);
  const auto window_func_context =
      WindowProjectNodeContext::getActiveWindowFunctionContext(this);
  const Analyzer::WindowFunction* window_func = window_func_context->getWindowFunction();
  const auto window_func_ti = get_adjusted_window_type_info(window_func);
  const auto aggregate_state_type =
      window_func_ti.get_type() == kFLOAT ? pi32_type : pi64_type;
  auto aggregate_state = aggregateWindowStatePtr();
  if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
    const auto aggregate_state_count_i64 = cgen_state_->llInt(
        reinterpret_cast<const int64_t>(window_func_context->aggregateStateCount()));
    auto aggregate_state_count = cgen_state_->ir_builder_.CreateIntToPtr(
        aggregate_state_count_i64, aggregate_state_type);
    const auto double_null_lv = cgen_state_->inlineFpNull(SQLTypeInfo(kDOUBLE));
    switch (window_func_ti.get_type()) {
      case kFLOAT: {
        return cgen_state_->emitCall(
            "load_avg_float", {aggregate_state, aggregate_state_count, double_null_lv});
      }
      case kDOUBLE: {
        return cgen_state_->emitCall(
            "load_avg_double", {aggregate_state, aggregate_state_count, double_null_lv});
      }
      case kDECIMAL: {
        return cgen_state_->emitCall(
            "load_avg_decimal",
            {aggregate_state,
             aggregate_state_count,
             double_null_lv,
             cgen_state_->llInt<int32_t>(window_func_ti.get_scale())});
      }
      default: {
        return cgen_state_->emitCall(
            "load_avg_int", {aggregate_state, aggregate_state_count, double_null_lv});
      }
    }
  }
  if (window_func->getKind() == SqlWindowFunctionKind::COUNT) {
    return cgen_state_->ir_builder_.CreateLoad(
        aggregate_state->getType()->getPointerElementType(), aggregate_state);
  }
  switch (window_func_ti.get_type()) {
    case kFLOAT: {
      return cgen_state_->emitCall("load_float", {aggregate_state});
    }
    case kDOUBLE: {
      return cgen_state_->emitCall("load_double", {aggregate_state});
    }
    default: {
      return cgen_state_->ir_builder_.CreateLoad(
          aggregate_state->getType()->getPointerElementType(), aggregate_state);
    }
  }
}

Here is the call graph for this function:

llvm::Value * Executor::codegenConditionalAggregateCondValSelector ( llvm::Value *  cond_lv, SQLAgg const  aggKind, CompilationOptions const &  co  ) const

private

Definition at line 1395 of file WindowFunctionIR.cpp.

References CHECK, and kSUM_IF.

                                        {
  llvm::Value* res_cond_lv{nullptr};
  switch (aggKind) {
    case kSUM_IF:
      if (cond_lv->getType()->isIntegerTy(1)) {
        // cond_expr returns i1 type val, just need to cast to i8 type
        // i.e., cond_expr IS NULL
        res_cond_lv = cgen_state_->castToTypeIn(cond_lv, 8);
      } else {
        CHECK(cond_lv->getType()->isIntegerTy(8));
        // cond_expr may have null value instead of upcasted bool (i1-type) value
        // so we have to correctly set true condition
        // i.e., i8 @gt_int32_t_nullable_lhs(..., i64 -2147483648, i8 -128)
        // has one of the following i8-type values: 1, 0, -128
        auto true_cond_lv =
            cgen_state_->ir_builder_.CreateICmpEQ(cond_lv, cgen_state_->llInt((int8_t)1));
        res_cond_lv = cgen_state_->ir_builder_.CreateSelect(
            true_cond_lv, cgen_state_->llInt((int8_t)1), cgen_state_->llInt((int8_t)0));
      }
      break;
    default:
      break;
  }
  return res_cond_lv;
}

llvm::Value * Executor::codegenCurrentPartitionIndex ( const WindowFunctionContext *  window_func_context, llvm::Value *  current_row_pos_lv  )

private

Definition at line 699 of file WindowFunctionIR.cpp.

References WindowFunctionContext::elementCount(), get_int_type(), WindowFunctionContext::getWindowFunction(), Analyzer::WindowFunction::isFrameNavigateWindowFunction(), WindowFunctionContext::partitionCount(), WindowFunctionContext::partitionNumCountBuf(), and WindowFunctionContext::payload().

                                   {
  const auto pi64_type =
      llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0);
  const auto pi32_type =
      llvm::PointerType::get(get_int_type(32, cgen_state_->context_), 0);
  auto row_pos_lv = current_row_pos_lv;
  if (window_func_context->getWindowFunction()->isFrameNavigateWindowFunction()) {
    // `current_row_pos_lv` indicates the index of the current row, but to figure out it's
    // index of window partition it belongs to, we need a special approach especially for
    // window framing navigation function  for instance, when we have five rows having two
    // columns pc and val such as (2,1), (2,2), (2,3), (1,1), (1,2), we build a OneToMany
    // Perfect Hash Table as: offset: 0 2 / count: 2 3 / payload: i1, i2, i3, i4, i5 where
    // i1 ~ i3 and i4 ~ i5 are rows for partition 1 (i.e., pc = 1) and 2 (i.e., prc = 2),
    // respectively. But when processing the first row (2, 1), the original
    // `current_row_pos_lv` stands for zero so computing which partitions it belongs to is
    // hard unless hashing the value at runtime. Even if we do hash, we cannot know the
    // exact hash slot unless we do binary + linear searches multiple times (via payload
    // buffer and the ordered payload buffer) i.e., when the row (1,2) is assigned to the
    // partition[4], we cannot find the hash slot index '4' by using `current_row_pos_lv`
    // unless doing a costly operation like a linear search over the entire window
    // partition Instead, we collect a hash slot that each row is assigned to and keep
    // this info at the payload buffer `hash_slot_idx_ptr_lv` and use it for computing
    // window frame navigation functions
    auto* const hash_slot_idx_ptr =
        window_func_context->payload() + window_func_context->elementCount();
    auto hash_slot_idx_buf_lv =
        cgen_state_->llInt(reinterpret_cast<int64_t>(hash_slot_idx_ptr));
    auto hash_slot_idx_ptr_lv =
        cgen_state_->ir_builder_.CreateIntToPtr(hash_slot_idx_buf_lv, pi32_type);
    auto hash_slot_idx_load_lv = cgen_state_->ir_builder_.CreateGEP(
        hash_slot_idx_ptr_lv->getType()->getPointerElementType(),
        hash_slot_idx_ptr_lv,
        current_row_pos_lv);
    row_pos_lv = cgen_state_->castToTypeIn(
        cgen_state_->ir_builder_.CreateLoad(
            hash_slot_idx_load_lv->getType()->getPointerElementType(),
            hash_slot_idx_load_lv,
            "cur_row_hash_slot_idx"),
        64);
  }
  auto partition_count_lv = cgen_state_->llInt(window_func_context->partitionCount());
  auto partition_num_count_buf_lv = cgen_state_->llInt(
      reinterpret_cast<int64_t>(window_func_context->partitionNumCountBuf()));
  auto partition_num_count_ptr_lv =
      cgen_state_->ir_builder_.CreateIntToPtr(partition_num_count_buf_lv, pi64_type);
  return cgen_state_->emitCall(
      "compute_int64_t_lower_bound",
      {partition_count_lv, row_pos_lv, partition_num_count_ptr_lv});
}

Here is the call graph for this function:

llvm::Value * Executor::codegenFrameBound ( bool  for_start_bound, bool  for_range_mode, bool  for_window_frame_naviation, const Analyzer::WindowFrame *  frame_bound, bool  is_timestamp_type_frame, llvm::Value *  order_key_null_val, const WindowFrameBoundFuncArgs &  args  )

private

Definition at line 567 of file WindowFunctionIR.cpp.

References CHECK, CURRENT_ROW, WindowFrameBoundFuncArgs::current_row_pos_lv, EXPR_FOLLOWING, EXPR_PRECEDING, WindowFrameBoundFuncArgs::frame_end_bound_expr_lv, WindowFrameBoundFuncArgs::frame_start_bound_expr_lv, Analyzer::WindowFrame::getBoundType(), WindowFrameBoundFuncArgs::int64_t_one_val_lv, WindowFrameBoundFuncArgs::int64_t_zero_val_lv, WindowFrameBoundFuncArgs::num_elem_current_partition_lv, WindowFrameBoundFuncArgs::order_type_col_name, UNBOUNDED_FOLLOWING, and UNBOUNDED_PRECEDING.

                                                                               {
  const auto bound_type = frame_bound->getBoundType();
  auto adjust_frame_end_bound = [&](llvm::Value* target_bound_lv) {
    return cgen_state_->ir_builder_.CreateSub(target_bound_lv, args.int64_t_one_val_lv);
  };
  if (bound_type == SqlWindowFrameBoundType::UNBOUNDED_PRECEDING) {
    CHECK(for_start_bound) << "frame end cannot be UNBOUNDED PRECEDING";
    return args.int64_t_zero_val_lv;
  } else if (bound_type == SqlWindowFrameBoundType::UNBOUNDED_FOLLOWING) {
    CHECK(!for_start_bound) << "frame start cannot be UNBOUNDED FOLLOWING";
    // adjust frame bound w.r.t the open frame interval if necessary
    return for_window_frame_naviation
               ? adjust_frame_end_bound(args.num_elem_current_partition_lv)
               : args.num_elem_current_partition_lv;
  }
  std::vector<llvm::Value*> func_args;
  std::string op_name =
      bound_type == SqlWindowFrameBoundType::EXPR_FOLLOWING ? "add" : "sub";
  if (!for_range_mode) {
    llvm::Value* current_row_bound_expr_lv{nullptr};
    if (for_window_frame_naviation) {
      // we already know a current row's index in (ordered) window frame in this case
      auto bound_expr =
          for_start_bound ? args.frame_start_bound_expr_lv : args.frame_end_bound_expr_lv;
      if (bound_type == SqlWindowFrameBoundType::EXPR_FOLLOWING) {
        current_row_bound_expr_lv =
            cgen_state_->ir_builder_.CreateAdd(args.current_row_pos_lv, bound_expr);
      } else if (bound_type == SqlWindowFrameBoundType::EXPR_PRECEDING) {
        current_row_bound_expr_lv =
            cgen_state_->ir_builder_.CreateSub(args.current_row_pos_lv, bound_expr);
      } else {
        CHECK(bound_type == SqlWindowFrameBoundType::CURRENT_ROW);
        current_row_bound_expr_lv = args.current_row_pos_lv;
      }
      // adjust frame bound w.r.t the open frame interval
      if (for_start_bound) {
        return cgen_state_->ir_builder_.CreateSelect(
            cgen_state_->ir_builder_.CreateICmpSLT(current_row_bound_expr_lv,
                                                   args.int64_t_zero_val_lv),
            args.int64_t_zero_val_lv,
            current_row_bound_expr_lv);
      } else {
        return cgen_state_->ir_builder_.CreateSelect(
            cgen_state_->ir_builder_.CreateICmpSGE(current_row_bound_expr_lv,
                                                   args.num_elem_current_partition_lv),
            adjust_frame_end_bound(args.num_elem_current_partition_lv),
            current_row_bound_expr_lv);
      }
    } else {
      std::string func_class = for_start_bound ? "start" : "end";
      auto const func_name = "compute_row_mode_" + func_class + "_index_" + op_name;
      func_args = prepareRowModeFuncArgs(for_start_bound, bound_type, args);
      current_row_bound_expr_lv = cgen_state_->emitCall(func_name, func_args);
    }
    return current_row_bound_expr_lv;
  } else {
    std::string func_class = for_start_bound ? "lower" : "upper";
    auto const func_name = getFramingFuncName(
        func_class,
        args.order_type_col_name,
        op_name,
        bound_type != SqlWindowFrameBoundType::CURRENT_ROW && is_timestamp_type_frame);
    func_args = prepareRangeModeFuncArgs(
        for_start_bound, frame_bound, is_timestamp_type_frame, order_key_null_val, args);
    auto frame_bound_lv = cgen_state_->emitCall(func_name, func_args);
    if (!for_start_bound && for_window_frame_naviation) {
      // adjust frame end bound w.r.t the open frame interval
      frame_bound_lv = cgen_state_->ir_builder_.CreateSelect(
          cgen_state_->ir_builder_.CreateICmpSGE(frame_bound_lv,
                                                 args.num_elem_current_partition_lv),
          adjust_frame_end_bound(args.num_elem_current_partition_lv),
          frame_bound_lv);
    }
    return frame_bound_lv;
  }
}

Here is the call graph for this function:

llvm::Value * Executor::codegenFrameBoundExpr ( const Analyzer::WindowFunction *  window_func, const Analyzer::WindowFrame *  frame_bound, CodeGenerator &  code_generator, const CompilationOptions &  co  )

private

Definition at line 516 of file WindowFunctionIR.cpp.

References CHECK, CodeGenerator::codegen(), EXPR_FOLLOWING, EXPR_PRECEDING, g_cluster, SQLTypeInfo::get_size(), Analyzer::Expr::get_type_info(), Analyzer::WindowFrame::getBoundExpr(), Analyzer::WindowFunction::getOrderKeys(), Analyzer::WindowFunction::hasRangeModeFraming(), kBIGINT, kINT, and kSMALLINT.

                                                                           {
  auto needs_bound_expr_codegen = [](const Analyzer::WindowFrame* window_frame) {
    return window_frame->getBoundType() == SqlWindowFrameBoundType::EXPR_FOLLOWING ||
           window_frame->getBoundType() == SqlWindowFrameBoundType::EXPR_PRECEDING;
  };
  const auto order_col_ti = window_func->getOrderKeys().front()->get_type_info();
  auto encode_date_col_val = [&order_col_ti, this](llvm::Value* bound_expr_lv) {
    if (order_col_ti.get_comp_param() == 16) {
      return cgen_state_->emitCall(
          "fixed_width_date_encode_noinline",
          {bound_expr_lv,
           cgen_state_->castToTypeIn(cgen_state_->inlineIntNull(SQLTypeInfo(kSMALLINT)),
                                     32),
           cgen_state_->inlineIntNull(SQLTypeInfo(kBIGINT))});
    } else {
      return cgen_state_->emitCall("fixed_width_date_encode_noinline",
                                   {bound_expr_lv,
                                    cgen_state_->inlineIntNull(SQLTypeInfo(kINT)),
                                    cgen_state_->inlineIntNull(SQLTypeInfo(kBIGINT))});
    }
  };
  llvm::Value* bound_expr_lv{nullptr};
  if (needs_bound_expr_codegen(frame_bound)) {
    auto bound_expr = frame_bound->getBoundExpr();
    if (auto dateadd_expr = dynamic_cast<const Analyzer::DateaddExpr*>(bound_expr)) {
      if (dateadd_expr->get_datetime_expr()->get_type_info().is_encoded_timestamp()) {
        dateadd_expr->set_fixed_encoding_null_val();
      }
    }
    auto bound_expr_lvs = code_generator.codegen(bound_expr, true, co);
    bound_expr_lv = bound_expr_lvs.front();
    if (order_col_ti.is_date() && window_func->hasRangeModeFraming()) {
      if (g_cluster) {
        throw std::runtime_error(
            "Range mode with date type ordering column is not supported yet.");
      }
      bound_expr_lv = encode_date_col_val(bound_expr_lv);
    }
    if (frame_bound->getBoundExpr()->get_type_info().get_size() != 8) {
      bound_expr_lv = cgen_state_->castToTypeIn(bound_expr_lv, 64);
    }
  } else {
    bound_expr_lv = cgen_state_->llInt((int64_t)-1);
  }
  CHECK(bound_expr_lv);
  return bound_expr_lv;
}

Here is the call graph for this function:

std::pair< llvm::Value *, llvm::Value * > Executor::codegenFrameBoundRange ( const Analyzer::WindowFunction *  window_func, CodeGenerator &  code_generator, const CompilationOptions &  co  )

private

Definition at line 921 of file WindowFunctionIR.cpp.

References CHECK, Analyzer::WindowFunction::getFrameEndBound(), and Analyzer::WindowFunction::getFrameStartBound().

                                  {
  const auto frame_start_bound = window_func->getFrameStartBound();
  const auto frame_end_bound = window_func->getFrameEndBound();
  auto frame_start_bound_expr_lv =
      codegenFrameBoundExpr(window_func, frame_start_bound, code_generator, co);
  auto frame_end_bound_expr_lv =
      codegenFrameBoundExpr(window_func, frame_end_bound, code_generator, co);
  CHECK(frame_start_bound_expr_lv);
  CHECK(frame_end_bound_expr_lv);
  return std::make_pair(frame_start_bound_expr_lv, frame_end_bound_expr_lv);
}

Here is the call graph for this function:

std::pair< llvm::Value *, llvm::Value * > Executor::codegenFrameNullRange ( WindowFunctionContext *  window_func_context, llvm::Value *  partition_index_lv  ) const

private

Definition at line 808 of file WindowFunctionIR.cpp.

References get_int_type(), WindowFunctionContext::getNullValueEndPos(), and WindowFunctionContext::getNullValueStartPos().

                                         {
  const auto pi64_type =
      llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0);
  const auto null_start_pos_buf = cgen_state_->llInt(
      reinterpret_cast<int64_t>(window_func_context->getNullValueStartPos()));
  const auto null_start_pos_buf_ptr =
      cgen_state_->ir_builder_.CreateIntToPtr(null_start_pos_buf, pi64_type);
  const auto null_start_pos_ptr =
      cgen_state_->ir_builder_.CreateGEP(get_int_type(64, cgen_state_->context_),
                                         null_start_pos_buf_ptr,
                                         partition_index_lv);
  auto null_start_pos_lv = cgen_state_->ir_builder_.CreateLoad(
      null_start_pos_ptr->getType()->getPointerElementType(),
      null_start_pos_ptr,
      "null_start_pos");
  const auto null_end_pos_buf = cgen_state_->llInt(
      reinterpret_cast<int64_t>(window_func_context->getNullValueEndPos()));
  const auto null_end_pos_buf_ptr =
      cgen_state_->ir_builder_.CreateIntToPtr(null_end_pos_buf, pi64_type);
  const auto null_end_pos_ptr = cgen_state_->ir_builder_.CreateGEP(
      get_int_type(64, cgen_state_->context_), null_end_pos_buf_ptr, partition_index_lv);
  auto null_end_pos_lv = cgen_state_->ir_builder_.CreateLoad(
      null_end_pos_ptr->getType()->getPointerElementType(),
      null_end_pos_ptr,
      "null_end_pos");
  return std::make_pair(null_start_pos_lv, null_end_pos_lv);
}

Here is the call graph for this function:

void Executor::codegenJoinLoops ( const std::vector< JoinLoop > &  join_loops, const RelAlgExecutionUnit &  ra_exe_unit, GroupByAndAggregate &  group_by_and_aggregate, llvm::Function *  query_func, llvm::BasicBlock *  entry_bb, QueryMemoryDescriptor &  query_mem_desc, const CompilationOptions &  co, const ExecutionOptions &  eo  )

private

Definition at line 1137 of file IRCodegen.cpp.

References ExecutionOptions::allow_runtime_query_interrupt, anonymous_namespace{QueryMemoryDescriptor.cpp}::any_of(), AUTOMATIC_IR_METADATA, CHECK, CHECK_EQ, JoinLoop::codegen(), CompilationOptions::device_type, JoinLoopDomain::element_count, get_int_array_type(), get_int_type(), INNER, MultiSet, CodeGenerator::posArg(), GroupByAndAggregate::query_infos_, query_mem_desc, Set, and ExecutionOptions::with_dynamic_watchdog.

                                                            {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto exit_bb =
      llvm::BasicBlock::Create(cgen_state_->context_, "exit", cgen_state_->current_func_);
  cgen_state_->ir_builder_.SetInsertPoint(exit_bb);
  cgen_state_->ir_builder_.CreateRet(cgen_state_->llInt<int32_t>(0));
  cgen_state_->ir_builder_.SetInsertPoint(entry_bb);
  CodeGenerator code_generator(this);

  llvm::BasicBlock* loops_entry_bb{nullptr};
  auto has_range_join =
      std::any_of(join_loops.begin(), join_loops.end(), [](const auto& join_loop) {
        return join_loop.kind() == JoinLoopKind::MultiSet;
      });
  if (has_range_join) {
    CHECK_EQ(join_loops.size(), size_t(1));
    const auto element_count =
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_), 9);

    auto compute_packed_offset = [](const int32_t x, const int32_t y) -> uint64_t {
      const uint64_t y_shifted = static_cast<uint64_t>(y) << 32;
      return y_shifted | static_cast<uint32_t>(x);
    };

    const auto values_arr = std::vector<llvm::Constant*>{
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_), 0),
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_),
                               compute_packed_offset(0, 1)),
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_),
                               compute_packed_offset(0, -1)),
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_),
                               compute_packed_offset(1, 0)),
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_),
                               compute_packed_offset(1, 1)),
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_),
                               compute_packed_offset(1, -1)),
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_),
                               compute_packed_offset(-1, 0)),
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_),
                               compute_packed_offset(-1, 1)),
        llvm::ConstantInt::get(get_int_type(64, cgen_state_->context_),
                               compute_packed_offset(-1, -1))};

    const auto constant_values_array = llvm::ConstantArray::get(
        get_int_array_type(64, 9, cgen_state_->context_), values_arr);
    CHECK(cgen_state_->module_);
    const auto values =
        new llvm::GlobalVariable(*cgen_state_->module_,
                                 get_int_array_type(64, 9, cgen_state_->context_),
                                 true,
                                 llvm::GlobalValue::LinkageTypes::InternalLinkage,
                                 constant_values_array);
    JoinLoop join_loop(
        JoinLoopKind::Set,
        JoinType::INNER,
        [element_count, values](const std::vector<llvm::Value*>& v) {
          JoinLoopDomain domain{{0}};
          domain.element_count = element_count;
          domain.values_buffer = values;
          return domain;
        },
        nullptr,
        nullptr,
        nullptr,
        nullptr,
        "range_key_loop");

    loops_entry_bb = JoinLoop::codegen(
        {join_loop},
        [this,
         query_func,
         &query_mem_desc,
         &co,
         &eo,
         &group_by_and_aggregate,
         &join_loops,
         &ra_exe_unit](const std::vector<llvm::Value*>& prev_iters) {
          auto& builder = cgen_state_->ir_builder_;

          auto body_exit_bb =
              llvm::BasicBlock::Create(cgen_state_->context_,
                                       "range_key_inner_body_exit",
                                       builder.GetInsertBlock()->getParent());

          auto range_key_body_bb =
              llvm::BasicBlock::Create(cgen_state_->context_,
                                       "range_key_loop_body",
                                       builder.GetInsertBlock()->getParent());
          builder.SetInsertPoint(range_key_body_bb);

          const auto body_loops_entry_bb = JoinLoop::codegen(
              join_loops,
              [this,
               query_func,
               &query_mem_desc,
               &co,
               &eo,
               &group_by_and_aggregate,
               &join_loops,
               &ra_exe_unit](const std::vector<llvm::Value*>& prev_iters) {
                addJoinLoopIterator(prev_iters, join_loops.size());
                auto& builder = cgen_state_->ir_builder_;
                const auto loop_body_bb =
                    llvm::BasicBlock::Create(builder.getContext(),
                                             "loop_body",
                                             builder.GetInsertBlock()->getParent());
                builder.SetInsertPoint(loop_body_bb);
                const bool can_return_error =
                    compileBody(ra_exe_unit, group_by_and_aggregate, query_mem_desc, co);
                if (can_return_error || cgen_state_->needs_error_check_ ||
                    eo.with_dynamic_watchdog || eo.allow_runtime_query_interrupt) {
                  createErrorCheckControlFlow(query_func,
                                              eo.with_dynamic_watchdog,
                                              eo.allow_runtime_query_interrupt,
                                              join_loops,
                                              co.device_type,
                                              group_by_and_aggregate.query_infos_);
                }
                return loop_body_bb;
              },
              prev_iters.back(),
              body_exit_bb,
              cgen_state_.get());

          builder.SetInsertPoint(range_key_body_bb);
          cgen_state_->ir_builder_.CreateBr(body_loops_entry_bb);

          builder.SetInsertPoint(body_exit_bb);
          return range_key_body_bb;
        },
        code_generator.posArg(nullptr),
        exit_bb,
        cgen_state_.get());
  } else {
    loops_entry_bb = JoinLoop::codegen(
        join_loops,
        /*body_codegen=*/
        [this,
         query_func,
         &query_mem_desc,
         &co,
         &eo,
         &group_by_and_aggregate,
         &join_loops,
         &ra_exe_unit](const std::vector<llvm::Value*>& prev_iters) {
          AUTOMATIC_IR_METADATA(cgen_state_.get());
          addJoinLoopIterator(prev_iters, join_loops.size());
          auto& builder = cgen_state_->ir_builder_;
          const auto loop_body_bb = llvm::BasicBlock::Create(
              builder.getContext(), "loop_body", builder.GetInsertBlock()->getParent());
          builder.SetInsertPoint(loop_body_bb);
          const bool can_return_error =
              compileBody(ra_exe_unit, group_by_and_aggregate, query_mem_desc, co);
          if (can_return_error || cgen_state_->needs_error_check_ ||
              eo.with_dynamic_watchdog || eo.allow_runtime_query_interrupt) {
            createErrorCheckControlFlow(query_func,
                                        eo.with_dynamic_watchdog,
                                        eo.allow_runtime_query_interrupt,
                                        join_loops,
                                        co.device_type,
                                        group_by_and_aggregate.query_infos_);
          }
          return loop_body_bb;
        },
        /*outer_iter=*/code_generator.posArg(nullptr),
        exit_bb,
        cgen_state_.get());
  }
  CHECK(loops_entry_bb);
  cgen_state_->ir_builder_.SetInsertPoint(entry_bb);
  cgen_state_->ir_builder_.CreateBr(loops_entry_bb);
}

Here is the call graph for this function:

llvm::Value * Executor::codegenLoadCurrentValueFromColBuf ( WindowFunctionContext *  window_func_context, CodeGenerator &  code_generator, WindowFrameBoundFuncArgs &  args  ) const

private

Definition at line 671 of file WindowFunctionIR.cpp.

References CHECK, CodeGenerator::codegenWindowPosition(), WindowFrameBoundFuncArgs::current_row_pos_lv, get_fp_type(), get_int_type(), Analyzer::WindowFunction::getOrderKeys(), WindowFunctionContext::getWindowFunction(), Analyzer::WindowFunction::isFrameNavigateWindowFunction(), and WindowFrameBoundFuncArgs::order_key_buf_ptr_lv.

                                          {
  llvm::Value* current_col_value_ptr_lv{nullptr};
  const auto order_key_size_in_byte = getOrderKeySize(window_func_context) * 8;
  auto const order_key_ptr =
      window_func_context->getWindowFunction()->getOrderKeys().front();
  CHECK(order_key_ptr);
  auto const order_col_ti = order_key_ptr->get_type_info();
  auto const order_col_llvm_type =
      order_col_ti.is_fp() ? get_fp_type(order_key_size_in_byte, cgen_state_->context_)
                           : get_int_type(order_key_size_in_byte, cgen_state_->context_);
  if (!window_func_context->getWindowFunction()->isFrameNavigateWindowFunction()) {
    auto rowid_in_partition_lv = code_generator.codegenWindowPosition(
        window_func_context, args.current_row_pos_lv);
    current_col_value_ptr_lv = cgen_state_->ir_builder_.CreateGEP(
        order_col_llvm_type, args.order_key_buf_ptr_lv, rowid_in_partition_lv);
  } else {
    current_col_value_ptr_lv = cgen_state_->ir_builder_.CreateGEP(
        order_col_llvm_type, args.order_key_buf_ptr_lv, args.current_row_pos_lv);
  }
  return cgen_state_->ir_builder_.CreateLoad(
      current_col_value_ptr_lv->getType()->getPointerElementType(),
      current_col_value_ptr_lv,
      "current_col_value");
}

Here is the call graph for this function:

std::pair< std::string, llvm::Value * > Executor::codegenLoadOrderKeyBufPtr ( WindowFunctionContext *  window_func_context ) const

private

Definition at line 838 of file WindowFunctionIR.cpp.

References anonymous_namespace{WindowFunctionIR.cpp}::get_col_type_name_by_size(), get_fp_type(), get_int_type(), WindowFunctionContext::getOrderKeyColumnBuffers(), WindowFunctionContext::getOrderKeyColumnBufferTypes(), Analyzer::WindowFunction::getOrderKeys(), and WindowFunctionContext::getWindowFunction().

                                                      {
  auto const order_key_ti =
      window_func_context->getWindowFunction()->getOrderKeys().front()->get_type_info();
  auto const order_key_size = order_key_ti.get_size();
  auto const order_col_type_name = get_col_type_name_by_size(
      order_key_size,
      window_func_context->getOrderKeyColumnBufferTypes().front().is_fp());
  size_t order_key_size_in_byte = order_key_size * 8;
  auto const order_key_type =
      order_key_ti.is_fp() ? get_fp_type(order_key_size_in_byte, cgen_state_->context_)
                           : get_int_type(order_key_size_in_byte, cgen_state_->context_);
  auto const order_key_buf_type = llvm::PointerType::get(order_key_type, 0);
  auto const order_key_buf = cgen_state_->llInt(
      reinterpret_cast<int64_t>(window_func_context->getOrderKeyColumnBuffers().front()));
  auto const order_key_buf_ptr_lv =
      cgen_state_->ir_builder_.CreateIntToPtr(order_key_buf, order_key_buf_type);
  return std::make_pair(order_col_type_name, order_key_buf_ptr_lv);
}

Here is the call graph for this function:

WindowPartitionBufferPtrs Executor::codegenLoadPartitionBuffers ( WindowFunctionContext *  window_func_context, llvm::Value *  partition_index_lv  ) const

private

Definition at line 858 of file WindowFunctionIR.cpp.

References WindowFunctionContext::counts(), WindowPartitionBufferPtrs::current_partition_start_offset_lv, get_int_type(), WindowPartitionBufferPtrs::num_elem_current_partition_lv, WindowFunctionContext::partitionStartOffset(), WindowFunctionContext::payload(), WindowFunctionContext::sortedPartition(), WindowPartitionBufferPtrs::target_partition_rowid_ptr_lv, and WindowPartitionBufferPtrs::target_partition_sorted_rowid_ptr_lv.

                                         {
  WindowPartitionBufferPtrs bufferPtrs;
  const auto pi64_type =
      llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0);
  const auto pi32_type =
      llvm::PointerType::get(get_int_type(32, cgen_state_->context_), 0);

  // partial sum of # elems of partitions
  auto partition_start_offset_buf_lv = cgen_state_->llInt(
      reinterpret_cast<int64_t>(window_func_context->partitionStartOffset()));
  auto partition_start_offset_ptr_lv =
      cgen_state_->ir_builder_.CreateIntToPtr(partition_start_offset_buf_lv, pi64_type);

  // get start offset of the current partition
  auto current_partition_start_offset_ptr_lv =
      cgen_state_->ir_builder_.CreateGEP(get_int_type(64, cgen_state_->context_),
                                         partition_start_offset_ptr_lv,
                                         partition_index_lv);
  bufferPtrs.current_partition_start_offset_lv = cgen_state_->ir_builder_.CreateLoad(
      current_partition_start_offset_ptr_lv->getType()->getPointerElementType(),
      current_partition_start_offset_ptr_lv);

  // row_id buf of the current partition
  const auto partition_rowid_buf_lv =
      cgen_state_->llInt(reinterpret_cast<int64_t>(window_func_context->payload()));
  const auto partition_rowid_ptr_lv =
      cgen_state_->ir_builder_.CreateIntToPtr(partition_rowid_buf_lv, pi32_type);
  bufferPtrs.target_partition_rowid_ptr_lv =
      cgen_state_->ir_builder_.CreateGEP(get_int_type(32, cgen_state_->context_),
                                         partition_rowid_ptr_lv,
                                         bufferPtrs.current_partition_start_offset_lv);

  // row_id buf of ordered current partition
  const auto sorted_rowid_lv = cgen_state_->llInt(
      reinterpret_cast<int64_t>(window_func_context->sortedPartition()));
  const auto sorted_rowid_ptr_lv =
      cgen_state_->ir_builder_.CreateIntToPtr(sorted_rowid_lv, pi64_type);
  bufferPtrs.target_partition_sorted_rowid_ptr_lv =
      cgen_state_->ir_builder_.CreateGEP(get_int_type(64, cgen_state_->context_),
                                         sorted_rowid_ptr_lv,
                                         bufferPtrs.current_partition_start_offset_lv);

  // # elems per partition
  const auto partition_count_buf =
      cgen_state_->llInt(reinterpret_cast<int64_t>(window_func_context->counts()));
  auto partition_count_buf_ptr_lv =
      cgen_state_->ir_builder_.CreateIntToPtr(partition_count_buf, pi32_type);

  // # elems of the given partition
  const auto num_elem_current_partition_ptr =
      cgen_state_->ir_builder_.CreateGEP(get_int_type(32, cgen_state_->context_),
                                         partition_count_buf_ptr_lv,
                                         partition_index_lv);
  bufferPtrs.num_elem_current_partition_lv = cgen_state_->castToTypeIn(
      cgen_state_->ir_builder_.CreateLoad(
          num_elem_current_partition_ptr->getType()->getPointerElementType(),
          num_elem_current_partition_ptr),
      64);
  return bufferPtrs;
}

Here is the call graph for this function:

llvm::BasicBlock * Executor::codegenSkipDeletedOuterTableRow ( const RelAlgExecutionUnit &  ra_exe_unit, const CompilationOptions &  co  )

private

Definition at line 3218 of file NativeCodegen.cpp.

                                  {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  if (!co.filter_on_deleted_column) {
    return nullptr;
  }
  CHECK(!ra_exe_unit.input_descs.empty());
  const auto& outer_input_desc = ra_exe_unit.input_descs[0];
  if (outer_input_desc.getSourceType() != InputSourceType::TABLE) {
    return nullptr;
  }
  const auto& table_key = outer_input_desc.getTableKey();
  const auto deleted_cd = plan_state_->getDeletedColForTable(table_key);
  if (!deleted_cd) {
    return nullptr;
  }
  CHECK(deleted_cd->columnType.is_boolean());
  const auto deleted_expr =
      makeExpr<Analyzer::ColumnVar>(deleted_cd->columnType,
                                    shared::ColumnKey{table_key, deleted_cd->columnId},
                                    outer_input_desc.getNestLevel());
  CodeGenerator code_generator(this);
  const auto is_deleted =
      code_generator.toBool(code_generator.codegen(deleted_expr.get(), true, co).front());
  const auto is_deleted_bb = llvm::BasicBlock::Create(
      cgen_state_->context_, "is_deleted", cgen_state_->row_func_);
  llvm::BasicBlock* bb = llvm::BasicBlock::Create(
      cgen_state_->context_, "is_not_deleted", cgen_state_->row_func_);
  cgen_state_->ir_builder_.CreateCondBr(is_deleted, is_deleted_bb, bb);
  cgen_state_->ir_builder_.SetInsertPoint(is_deleted_bb);
  cgen_state_->ir_builder_.CreateRet(cgen_state_->llInt<int32_t>(0));
  cgen_state_->ir_builder_.SetInsertPoint(bb);
  return bb;
}

void Executor::codegenWindowAvgEpilogue ( llvm::Value *  crt_val, llvm::Value *  window_func_null_val  )

private

Definition at line 1299 of file WindowFunctionIR.cpp.

References AUTOMATIC_IR_METADATA, anonymous_namespace{WindowFunctionIR.cpp}::get_adjusted_window_type_info(), get_int_type(), WindowProjectNodeContext::getActiveWindowFunctionContext(), kDOUBLE, and kFLOAT.

                                                                         {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto window_func_context =
      WindowProjectNodeContext::getActiveWindowFunctionContext(this);
  const auto window_func = window_func_context->getWindowFunction();
  const auto window_func_ti = get_adjusted_window_type_info(window_func);
  const auto pi32_type =
      llvm::PointerType::get(get_int_type(32, cgen_state_->context_), 0);
  const auto pi64_type =
      llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0);
  const auto aggregate_state_type =
      window_func_ti.get_type() == kFLOAT ? pi32_type : pi64_type;
  const auto aggregate_state_count_i64 = cgen_state_->llInt(
      reinterpret_cast<const int64_t>(window_func_context->aggregateStateCount()));
  auto aggregate_state_count = cgen_state_->ir_builder_.CreateIntToPtr(
      aggregate_state_count_i64, aggregate_state_type);
  std::string agg_count_func_name = "agg_count";
  switch (window_func_ti.get_type()) {
    case kFLOAT: {
      agg_count_func_name += "_float";
      break;
    }
    case kDOUBLE: {
      agg_count_func_name += "_double";
      break;
    }
    default: {
      break;
    }
  }
  agg_count_func_name += "_skip_val";
  cgen_state_->emitCall(agg_count_func_name,
                        {aggregate_state_count, crt_val, window_func_null_val});
}

Here is the call graph for this function:

std::pair< llvm::Value *, llvm::Value * > Executor::codegenWindowFrameBounds ( WindowFunctionContext *  window_func_context, const Analyzer::WindowFrame *  frame_start_bound, const Analyzer::WindowFrame *  frame_end_bound, llvm::Value *  order_key_col_null_val_lv, WindowFrameBoundFuncArgs &  args, CodeGenerator &  code_generator  )

private

Definition at line 936 of file WindowFunctionIR.cpp.

References run_benchmark_import::args, CHECK, WindowFrameBoundFuncArgs::current_col_value_lv, WindowFunctionContext::getOrderKeyColumnBuffers(), WindowFunctionContext::getWindowFunction(), Analyzer::WindowFrame::hasTimestampTypeFrameBound(), and WindowFrameBoundFuncArgs::order_type_col_name.

                                   {
  const auto window_func = window_func_context->getWindowFunction();
  CHECK(window_func);
  const auto is_timestamp_type_frame = frame_start_bound->hasTimestampTypeFrameBound() ||
                                       frame_end_bound->hasTimestampTypeFrameBound();

  if (window_func->hasRangeModeFraming()) {
    CHECK(window_func_context->getOrderKeyColumnBuffers().size() == 1);
    CHECK(window_func->getOrderKeys().size() == 1UL);
    CHECK(window_func_context->getOrderKeyColumnBuffers().size() == 1UL);
    args.order_type_col_name = getOrderKeyTypeName(window_func_context);
    args.current_col_value_lv =
        codegenLoadCurrentValueFromColBuf(window_func_context, code_generator, args);
  }

  auto get_order_key_null_val = [is_timestamp_type_frame,
                                 &order_key_col_null_val_lv,
                                 this](const Analyzer::WindowFrame* frame_bound) {
    return is_timestamp_type_frame && !frame_bound->isCurrentRowBound()
               ? cgen_state_->castToTypeIn(order_key_col_null_val_lv, 64)
               : order_key_col_null_val_lv;
  };
  auto frame_start_bound_lv =
      codegenFrameBound(true,
                        window_func->hasRangeModeFraming(),
                        window_func->isFrameNavigateWindowFunction(),
                        frame_start_bound,
                        is_timestamp_type_frame,
                        get_order_key_null_val(frame_start_bound),
                        args);
  auto frame_end_bound_lv =
      codegenFrameBound(false,
                        window_func->hasRangeModeFraming(),
                        window_func->isFrameNavigateWindowFunction(),
                        frame_end_bound,
                        is_timestamp_type_frame,
                        get_order_key_null_val(frame_end_bound),
                        args);
  CHECK(frame_start_bound_lv);
  CHECK(frame_end_bound_lv);
  return std::make_pair(frame_start_bound_lv, frame_end_bound_lv);
}

Here is the call graph for this function:

llvm::Value * Executor::codegenWindowFunction ( const size_t  target_index, const CompilationOptions &  co  )

private

Definition at line 21 of file WindowFunctionIR.cpp.

References WindowProjectNodeContext::activateWindowFunctionContext(), run_benchmark_import::args, AUTOMATIC_IR_METADATA, AVG, CHECK, CHECK_EQ, COUNT, COUNT_IF, CUME_DIST, DENSE_RANK, logger::FATAL, FIRST_VALUE, WindowProjectNodeContext::get(), WindowFunctionContext::getWindowFunction(), LAG, LAG_IN_FRAME, LAST_VALUE, LEAD, LEAD_IN_FRAME, LOG, MAX, MIN, NTH_VALUE, NTH_VALUE_IN_FRAME, NTILE, PERCENT_RANK, RANK, ROW_NUMBER, SUM, and SUM_IF.

                                                                           {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  CodeGenerator code_generator(this);

  const auto window_func_context =
      WindowProjectNodeContext::get(this)->activateWindowFunctionContext(this,
                                                                         target_index);
  const auto window_func = window_func_context->getWindowFunction();
  switch (window_func->getKind()) {
    case SqlWindowFunctionKind::ROW_NUMBER:
    case SqlWindowFunctionKind::RANK:
    case SqlWindowFunctionKind::DENSE_RANK:
    case SqlWindowFunctionKind::NTILE: {
      // they are always evaluated on the entire partition
      return code_generator.codegenWindowPosition(window_func_context,
                                                  code_generator.posArg(nullptr));
    }
    case SqlWindowFunctionKind::PERCENT_RANK:
    case SqlWindowFunctionKind::CUME_DIST: {
      // they are always evaluated on the entire partition
      return cgen_state_->emitCall("percent_window_func",
                                   {cgen_state_->llInt(reinterpret_cast<const int64_t>(
                                        window_func_context->output())),
                                    code_generator.posArg(nullptr)});
    }
    case SqlWindowFunctionKind::LAG:
    case SqlWindowFunctionKind::LEAD:
    case SqlWindowFunctionKind::FIRST_VALUE:
    case SqlWindowFunctionKind::LAST_VALUE:
    case SqlWindowFunctionKind::NTH_VALUE: {
      // they are always evaluated on the current frame
      CHECK(WindowProjectNodeContext::get(this));
      const auto& args = window_func->getArgs();
      CHECK(!args.empty());
      const auto arg_lvs = code_generator.codegen(args.front().get(), true, co);
      CHECK_EQ(arg_lvs.size(), size_t(1));
      return arg_lvs.front();
    }
    case SqlWindowFunctionKind::AVG:
    case SqlWindowFunctionKind::MIN:
    case SqlWindowFunctionKind::MAX:
    case SqlWindowFunctionKind::SUM:
    case SqlWindowFunctionKind::SUM_IF:
    case SqlWindowFunctionKind::COUNT:
    case SqlWindowFunctionKind::COUNT_IF: {
      // they are always evaluated on the current frame
      return codegenWindowFunctionAggregate(co);
    }
    case SqlWindowFunctionKind::LEAD_IN_FRAME:
    case SqlWindowFunctionKind::LAG_IN_FRAME:
    case SqlWindowFunctionKind::NTH_VALUE_IN_FRAME: {
      return codegenWindowNavigationFunctionOnFrame(co);
    }
    default: {
      LOG(FATAL) << "Invalid window function kind";
    }
  }
  return nullptr;
}

Here is the call graph for this function:

llvm::Value * Executor::codegenWindowFunctionAggregate ( const CompilationOptions &  co )

private

Definition at line 242 of file WindowFunctionIR.cpp.

References AUTOMATIC_IR_METADATA, AVG, CHECK, WindowProjectNodeContext::get(), get_int_type(), and WindowProjectNodeContext::getActiveWindowFunctionContext().

                                                                                {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto reset_state_false_bb = codegenWindowResetStateControlFlow();
  auto aggregate_state = aggregateWindowStatePtr();
  llvm::Value* aggregate_state_count = nullptr;
  const auto window_func_context =
      WindowProjectNodeContext::getActiveWindowFunctionContext(this);
  const auto window_func = window_func_context->getWindowFunction();
  if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
    const auto aggregate_state_count_i64 = cgen_state_->llInt(
        reinterpret_cast<const int64_t>(window_func_context->aggregateStateCount()));
    const auto pi64_type =
        llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0);
    aggregate_state_count =
        cgen_state_->ir_builder_.CreateIntToPtr(aggregate_state_count_i64, pi64_type);
  }
  codegenWindowFunctionStateInit(aggregate_state);
  if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
    const auto count_zero = cgen_state_->llInt(int64_t(0));
    cgen_state_->emitCall("agg_id", {aggregate_state_count, count_zero});
  }
  cgen_state_->ir_builder_.CreateBr(reset_state_false_bb);
  cgen_state_->ir_builder_.SetInsertPoint(reset_state_false_bb);
  CHECK(WindowProjectNodeContext::get(this));
  return codegenWindowFunctionAggregateCalls(aggregate_state, co);
}

Here is the call graph for this function:

llvm::Value * Executor::codegenWindowFunctionAggregateCalls ( llvm::Value *  aggregate_state, const CompilationOptions &  co  )

private

Definition at line 985 of file WindowFunctionIR.cpp.

References run_benchmark_import::args, AUTOMATIC_IR_METADATA, AVG, CHECK, CHECK_EQ, CodeGenerator::codegen(), CodeGenerator::codegenCastBetweenIntTypes(), COUNT, COUNT_IF, WindowFrameBoundFuncArgs::current_partition_start_offset_lv, anonymous_namespace{WindowFunctionIR.cpp}::get_adjusted_window_type_info(), anonymous_namespace{WindowFunctionIR.cpp}::get_col_type_name_by_size(), get_int_type(), anonymous_namespace{WindowFunctionIR.cpp}::get_window_agg_name(), WindowProjectNodeContext::getActiveWindowFunctionContext(), inline_fixed_encoding_null_val(), kDATE, kDOUBLE, kENCODING_DATE_IN_DAYS, kENCODING_FIXED, kFLOAT, kSUM_IF, kTIME, kTIMESTAMP, kTINYINT, MAX, MIN, CodeGenerator::posArg(), SUM, SUM_IF, and window_function_conditional_aggregate().

                                                                                         {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto window_func_context =
      WindowProjectNodeContext::getActiveWindowFunctionContext(this);
  const auto window_func = window_func_context->getWindowFunction();
  const auto window_func_ti = get_adjusted_window_type_info(window_func);
  const auto window_func_null_val =
      window_func_ti.is_fp()
          ? cgen_state_->inlineFpNull(window_func_ti)
          : cgen_state_->castToTypeIn(cgen_state_->inlineIntNull(window_func_ti), 64);
  if (window_func_context->elementCount() == 0) {
    // we do not need to generate a code for an empty input table
    return window_func->getKind() == SqlWindowFunctionKind::AVG
               ? cgen_state_->inlineFpNull(SQLTypeInfo(SQLTypes::kDOUBLE))
               : window_func_null_val;
  }
  const auto& args = window_func->getArgs();
  CodeGenerator code_generator(this);
  if (window_func_context->needsToBuildAggregateTree()) {
    // compute an aggregated value for each row of the window frame by using segment tree
    // when constructing a window context, we build a necessary segment tree (so called
    // `aggregate tree`) to query the aggregated value of the specific window frame
    const auto pi64_type =
        llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0);
    const auto ppi64_type = llvm::PointerType::get(
        llvm::PointerType::get(get_int_type(64, cgen_state_->context_), 0), 0);

    auto [frame_start_bound_expr_lv, frame_end_bound_expr_lv] =
        codegenFrameBoundRange(window_func, code_generator, co);

    // compute aggregated value over the computed frame range
    auto current_row_pos_lv = code_generator.posArg(nullptr);
    auto partition_index_lv =
        codegenCurrentPartitionIndex(window_func_context, current_row_pos_lv);

    // ordering column buffer
    const auto target_col_ti = args.front()->get_type_info();
    const auto target_col_size = target_col_ti.get_size();
    const auto col_type_name =
        get_col_type_name_by_size(target_col_size, target_col_ti.is_fp());

    const auto partition_buf_ptrs =
        codegenLoadPartitionBuffers(window_func_context, partition_index_lv);

    auto [order_col_type_name, order_key_buf_ptr_lv] =
        codegenLoadOrderKeyBufPtr(window_func_context);

    // null value of the ordering column
    const auto order_key_buf_ti =
        window_func_context->getOrderKeyColumnBufferTypes().front();
    auto const ordering_spec = window_func->getCollation().front();
    llvm::Value* order_key_col_null_val_lv{nullptr};
    switch (order_key_buf_ti.get_type()) {
      case kDATE:
      case kTIMESTAMP:
      case kTIME: {
        if (order_key_buf_ti.get_compression() == kENCODING_FIXED ||
            order_key_buf_ti.get_compression() == kENCODING_DATE_IN_DAYS) {
          auto null_val = inline_fixed_encoding_null_val(order_key_buf_ti);
          order_key_col_null_val_lv = cgen_state_->llInt((int32_t)null_val);
          break;
        }
      }
      default: {
        order_key_col_null_val_lv = cgen_state_->inlineNull(order_key_buf_ti);
        break;
      }
    }

    auto [null_start_pos_lv, null_end_pos_lv] =
        codegenFrameNullRange(window_func_context, partition_index_lv);
    auto nulls_first_lv = cgen_state_->llBool(ordering_spec.nulls_first);

    WindowFrameBoundFuncArgs WindowFrameBoundFuncArgs{
        frame_start_bound_expr_lv,
        frame_end_bound_expr_lv,
        current_row_pos_lv,
        nullptr,
        partition_buf_ptrs.current_partition_start_offset_lv,
        cgen_state_->llInt((int64_t)0),
        cgen_state_->llInt((int64_t)1),
        partition_buf_ptrs.num_elem_current_partition_lv,
        order_key_buf_ptr_lv,
        "",
        partition_buf_ptrs.target_partition_rowid_ptr_lv,
        partition_buf_ptrs.target_partition_sorted_rowid_ptr_lv,
        nulls_first_lv,
        null_start_pos_lv,
        null_end_pos_lv};
    auto [frame_start_bound_lv, frame_end_bound_lv] =
        codegenWindowFrameBounds(window_func_context,
                                 window_func->getFrameStartBound(),
                                 window_func->getFrameEndBound(),
                                 order_key_col_null_val_lv,
                                 WindowFrameBoundFuncArgs,
                                 code_generator);

    // codegen to send a query with frame bound to aggregate tree searcher
    llvm::Value* aggregation_trees_lv{nullptr};
    llvm::Value* invalid_val_lv{nullptr};
    llvm::Value* null_val_lv{nullptr};
    std::string aggregation_tree_search_func_name{"search_"};
    std::string aggregation_tree_getter_func_name{"get_"};

    // prepare null values and aggregate_tree getter and searcher depending on
    // a type of the ordering column
    auto agg_expr_ti = args.front()->get_type_info();
    if (agg_expr_ti.is_fp()) {
      if (window_func->getKind() == SqlWindowFunctionKind::MIN) {
        invalid_val_lv = cgen_state_->llFp(std::numeric_limits<double>::max());
      } else if (window_func->getKind() == SqlWindowFunctionKind::MAX) {
        invalid_val_lv = cgen_state_->llFp(std::numeric_limits<double>::lowest());
      } else {
        invalid_val_lv = cgen_state_->llFp((double)0);
      }
      null_val_lv = cgen_state_->inlineFpNull(SQLTypeInfo(kDOUBLE));
      aggregation_tree_search_func_name += "double";
      aggregation_tree_getter_func_name += "double";
    } else {
      if (window_func->getKind() == SqlWindowFunctionKind::MIN) {
        invalid_val_lv = cgen_state_->llInt(std::numeric_limits<int64_t>::max());
      } else if (window_func->getKind() == SqlWindowFunctionKind::MAX) {
        invalid_val_lv = cgen_state_->llInt(std::numeric_limits<int64_t>::lowest());
      } else {
        invalid_val_lv = cgen_state_->llInt((int64_t)0);
      }
      null_val_lv = cgen_state_->llInt(inline_int_null_value<int64_t>());
      aggregation_tree_search_func_name += "int64_t";
      aggregation_tree_getter_func_name += "integer";
    }

    // derived aggregation has a different code path
    if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
      aggregation_tree_search_func_name += "_derived";
      aggregation_tree_getter_func_name += "_derived";
    }

    // get a buffer holding aggregate trees for each partition
    if (agg_expr_ti.is_fp()) {
      if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
        aggregation_trees_lv = cgen_state_->llInt(reinterpret_cast<int64_t>(
            window_func_context->getDerivedAggregationTreesForDoubleTypeWindowExpr()));
      } else {
        aggregation_trees_lv = cgen_state_->llInt(reinterpret_cast<int64_t>(
            window_func_context->getAggregationTreesForDoubleTypeWindowExpr()));
      }
    } else {
      if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
        aggregation_trees_lv = cgen_state_->llInt(reinterpret_cast<int64_t>(
            window_func_context->getDerivedAggregationTreesForIntegerTypeWindowExpr()));
      } else {
        aggregation_trees_lv = cgen_state_->llInt(reinterpret_cast<int64_t>(
            window_func_context->getAggregationTreesForIntegerTypeWindowExpr()));
      }
    }

    CHECK(aggregation_trees_lv);
    CHECK(invalid_val_lv);
    aggregation_tree_search_func_name += "_aggregation_tree";
    aggregation_tree_getter_func_name += "_aggregation_tree";

    // get the aggregate tree of the current partition from a window context
    auto aggregation_trees_ptr =
        cgen_state_->ir_builder_.CreateIntToPtr(aggregation_trees_lv, ppi64_type);
    auto target_aggregation_tree_lv = cgen_state_->emitCall(
        aggregation_tree_getter_func_name, {aggregation_trees_ptr, partition_index_lv});

    // a depth of segment tree
    const auto tree_depth_buf = cgen_state_->llInt(
        reinterpret_cast<int64_t>(window_func_context->getAggregateTreeDepth()));
    const auto tree_depth_buf_ptr =
        cgen_state_->ir_builder_.CreateIntToPtr(tree_depth_buf, pi64_type);
    const auto current_partition_tree_depth_buf_ptr = cgen_state_->ir_builder_.CreateGEP(
        get_int_type(64, cgen_state_->context_), tree_depth_buf_ptr, partition_index_lv);
    const auto current_partition_tree_depth_lv = cgen_state_->ir_builder_.CreateLoad(
        current_partition_tree_depth_buf_ptr->getType()->getPointerElementType(),
        current_partition_tree_depth_buf_ptr);

    // a fanout of the current partition's segment tree
    const auto aggregation_tree_fanout_lv = cgen_state_->llInt(
        static_cast<int64_t>(window_func_context->getAggregateTreeFanout()));

    // agg_type
    const auto agg_type_lv =
        cgen_state_->llInt(static_cast<int32_t>(window_func->getKind()));

    // send a query to the aggregate tree with the frame range:
    // `frame_start_bound_lv` ~ `frame_end_bound_lv`
    auto res_lv =
        cgen_state_->emitCall(aggregation_tree_search_func_name,
                              {target_aggregation_tree_lv,
                               frame_start_bound_lv,
                               frame_end_bound_lv,
                               current_partition_tree_depth_lv,
                               aggregation_tree_fanout_lv,
                               cgen_state_->llBool(agg_expr_ti.is_decimal()),
                               cgen_state_->llInt((int64_t)agg_expr_ti.get_scale()),
                               invalid_val_lv,
                               null_val_lv,
                               agg_type_lv});

    // handling returned null value if exists
    std::string null_handler_func_name{"handle_null_val_"};
    std::vector<llvm::Value*> null_handler_args{res_lv, null_val_lv};

    // determine null_handling function's name
    if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
      // average aggregate function returns a value as a double
      // (and our search* function also returns a double)
      if (agg_expr_ti.is_fp()) {
        // fp type: double null value
        null_handler_func_name += "double_double";
      } else {
        // non-fp type: int64_t null type
        null_handler_func_name += "double_int64_t";
      }
    } else if (agg_expr_ti.is_fp()) {
      // fp type: double null value
      null_handler_func_name += "double_double";
    } else {
      // non-fp type: int64_t null type
      null_handler_func_name += "int64_t_int64_t";
    }
    null_handler_func_name += "_window_framing_agg";

    // prepare null_val
    if (window_func->getKind() == SqlWindowFunctionKind::COUNT) {
      if (agg_expr_ti.is_fp()) {
        null_handler_args.push_back(cgen_state_->llFp((double)0));
      } else {
        null_handler_args.push_back(cgen_state_->llInt((int64_t)0));
      }
    } else if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
      null_handler_args.push_back(cgen_state_->inlineFpNull(SQLTypeInfo(kDOUBLE)));
    } else {
      null_handler_args.push_back(cgen_state_->castToTypeIn(window_func_null_val, 64));
    }
    res_lv = cgen_state_->emitCall(null_handler_func_name, null_handler_args);

    // when AGG_TYPE is double, we get a double type return value we expect an integer
    // type value for the count aggregation
    if (window_func->getKind() == SqlWindowFunctionKind::COUNT && agg_expr_ti.is_fp()) {
      return cgen_state_->ir_builder_.CreateFPToSI(
          res_lv, get_int_type(64, cgen_state_->context_));
    }
    return res_lv;
  } else {
    auto agg_name = get_window_agg_name(window_func->getKind(), window_func_ti);
    Analyzer::Expr* arg_target_expr;
    std::vector<llvm::Value*> agg_func_args{aggregate_state};
    auto modified_window_func_null_val = window_func_null_val;
    if (args.empty() ||
        (window_func->getKind() == SqlWindowFunctionKind::COUNT &&
         dynamic_cast<Analyzer::Constant*>(args.front().get()) != nullptr)) {
      // a count aggregation without an expression: COUNT(1) or COUNT(*)
      agg_func_args.push_back(cgen_state_->llInt(int64_t(1)));
    } else {
      // we use #base_agg_func_name##_skip_val agg function
      // i.e.,int64_t agg_sum_skip_val(int64_t* agg, int64_t val, int64_t skip_val)
      arg_target_expr = args.front().get();
      const auto arg_lvs = code_generator.codegen(arg_target_expr, true, co);
      CHECK_EQ(arg_lvs.size(), size_t(1));
      // handling current row's value
      auto crt_val = arg_lvs.front();
      if ((window_func->getKind() == SqlWindowFunctionKind::SUM ||
           window_func->getKind() == SqlWindowFunctionKind::SUM_IF) &&
          !window_func_ti.is_fp()) {
        crt_val = code_generator.codegenCastBetweenIntTypes(
            arg_lvs.front(), args.front()->get_type_info(), window_func_ti, false);
      }
      agg_func_args.push_back(window_func_ti.get_type() == kFLOAT
                                  ? crt_val
                                  : cgen_state_->castToTypeIn(crt_val, 64));
      // handle null value and conditional value for conditional aggregates if necessary
      llvm::Value* cond_lv{nullptr};
      if (window_function_conditional_aggregate(window_func->getKind())) {
        switch (window_func->getKind()) {
          case SqlWindowFunctionKind::COUNT_IF:
            // COUNT_IF has a single condition expr which is always bool type
            modified_window_func_null_val = cgen_state_->castToTypeIn(
                cgen_state_->inlineNull(SQLTypeInfo(kTINYINT)), 64);
            break;
          case SqlWindowFunctionKind::SUM_IF: {
            // FP type input col uses its own null value depending on the type
            // otherwise (integer type input col), we use 8-byte type
            if (args.front()->get_type_info().is_integer()) {
              agg_func_args[1] = cgen_state_->castToTypeIn(agg_func_args[1], 64);
              // keep the null value but casting its type to 8-byte
              modified_window_func_null_val =
                  cgen_state_->castToTypeIn(window_func_null_val, 64);
            }
            auto cond_expr_lv = code_generator.codegen(args[1].get(), true, co).front();
            cond_lv =
                codegenConditionalAggregateCondValSelector(cond_expr_lv, kSUM_IF, co);
          }
          default:
            break;
        }
      }
      agg_name += "_skip_val";
      agg_func_args.push_back(modified_window_func_null_val);
      if (cond_lv) {
        agg_func_args.push_back(cond_lv);
      }
    }
    cgen_state_->emitCall(agg_name, agg_func_args);
    if (window_func->getKind() == SqlWindowFunctionKind::AVG) {
      codegenWindowAvgEpilogue(agg_func_args[1], window_func_null_val);
    }
    return codegenAggregateWindowState();
  }
}

Here is the call graph for this function:

void Executor::codegenWindowFunctionStateInit ( llvm::Value *  aggregate_state )

private

Definition at line 298 of file WindowFunctionIR.cpp.

References AUTOMATIC_IR_METADATA, COUNT, COUNT_IF, anonymous_namespace{WindowFunctionIR.cpp}::get_adjusted_window_type_info(), get_int_type(), WindowProjectNodeContext::getActiveWindowFunctionContext(), kDOUBLE, and kFLOAT.

                                                                        {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto window_func_context =
      WindowProjectNodeContext::getActiveWindowFunctionContext(this);
  const auto window_func = window_func_context->getWindowFunction();
  const auto window_func_ti = get_adjusted_window_type_info(window_func);
  const auto window_func_null_val =
      window_func_ti.is_fp()
          ? cgen_state_->inlineFpNull(window_func_ti)
          : cgen_state_->castToTypeIn(cgen_state_->inlineIntNull(window_func_ti), 64);
  llvm::Value* window_func_init_val;
  const auto window_func_kind = window_func_context->getWindowFunction()->getKind();
  if (window_func_kind == SqlWindowFunctionKind::COUNT ||
      window_func_kind == SqlWindowFunctionKind::COUNT_IF) {
    switch (window_func_ti.get_type()) {
      case kFLOAT: {
        window_func_init_val = cgen_state_->llFp(float(0));
        break;
      }
      case kDOUBLE: {
        window_func_init_val = cgen_state_->llFp(double(0));
        break;
      }
      default: {
        window_func_init_val = cgen_state_->llInt(int64_t(0));
        break;
      }
    }
  } else {
    window_func_init_val = window_func_null_val;
  }
  const auto pi32_type =
      llvm::PointerType::get(get_int_type(32, cgen_state_->context_), 0);
  switch (window_func_ti.get_type()) {
    case kDOUBLE: {
      cgen_state_->emitCall("agg_id_double", {aggregate_state, window_func_init_val});
      break;
    }
    case kFLOAT: {
      aggregate_state =
          cgen_state_->ir_builder_.CreateBitCast(aggregate_state, pi32_type);
      cgen_state_->emitCall("agg_id_float", {aggregate_state, window_func_init_val});
      break;
    }
    default: {
      cgen_state_->emitCall("agg_id", {aggregate_state, window_func_init_val});
      break;
    }
  }
}

Here is the call graph for this function:

llvm::Value * Executor::codegenWindowNavigationFunctionOnFrame ( const CompilationOptions &  co )

private

Definition at line 349 of file WindowFunctionIR.cpp.

References run_benchmark_import::args, AUTOMATIC_IR_METADATA, CHECK, WindowFrameBoundFuncArgs::current_partition_start_offset_lv, anonymous_namespace{WindowFunctionIR.cpp}::get_col_type_name_by_size(), get_fp_type(), get_int_type(), anonymous_namespace{WindowFunctionIR.cpp}::get_null_value_by_size(), anonymous_namespace{WindowFunctionIR.cpp}::get_null_value_by_size_with_encoding(), WindowProjectNodeContext::getActiveWindowFunctionContext(), kENCODING_DATE_IN_DAYS, kSecsPerDay, LAG_IN_FRAME, LEAD_IN_FRAME, NTH_VALUE_IN_FRAME, and UNREACHABLE.

                                  {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto window_func_context =
      WindowProjectNodeContext::getActiveWindowFunctionContext(this);
  const auto window_func = window_func_context->getWindowFunction();
  const auto window_func_kind = window_func->getKind();
  const auto& args = window_func->getArgs();
  CHECK(args.size() >= 1 && args.size() <= 3);
  CodeGenerator code_generator(this);

  const auto target_col_ti = args.front()->get_type_info();
  const auto target_col_size = target_col_ti.get_size();
  const auto target_col_type_name =
      get_col_type_name_by_size(target_col_size, target_col_ti.is_fp());
  const auto target_col_logical_type_name = get_col_type_name_by_size(
      window_func->get_type_info().get_size(), window_func->get_type_info().is_fp());

  // when target_column is fixed encoded, we store the actual column value by
  // considering it, but our resultset analyzer only considers the type without encoding
  // scheme so we handle them separately
  auto logical_null_val_lv =
      get_null_value_by_size(cgen_state_.get(), window_func->get_type_info());
  auto target_col_null_val_lv =
      get_null_value_by_size_with_encoding(cgen_state_.get(), target_col_ti);
  if (window_func_context->elementCount() == 0) {
    // we do not need to generate a code for an empty input table
    return target_col_null_val_lv;
  }

  auto [frame_start_bound_expr_lv, frame_end_bound_expr_lv] =
      codegenFrameBoundRange(window_func, code_generator, co);

  auto current_row_pos_lv = code_generator.posArg(nullptr);
  auto partition_index_lv =
      codegenCurrentPartitionIndex(window_func_context, current_row_pos_lv);

  // load window function input expression; target_column
  size_t target_col_size_in_byte = target_col_size * 8;
  llvm::Type* col_buf_ptr_type =
      target_col_ti.is_fp()
          ? get_fp_type(target_col_size_in_byte, cgen_state_->context_)
          : get_int_type(target_col_size_in_byte, cgen_state_->context_);
  auto col_buf_type = llvm::PointerType::get(col_buf_ptr_type, 0);
  auto target_col_buf_ptr_lv = cgen_state_->llInt(reinterpret_cast<int64_t>(
      window_func_context->getColumnBufferForWindowFunctionExpressions().front()));
  auto target_col_buf_lv =
      cgen_state_->ir_builder_.CreateIntToPtr(target_col_buf_ptr_lv, col_buf_type);

  // prepare various buffer ptrs related to the window partition
  auto partition_buf_ptrs =
      codegenLoadPartitionBuffers(window_func_context, partition_index_lv);

  // null value of the ordering column
  const auto order_key_buf_ti =
      window_func_context->getOrderKeyColumnBufferTypes().front();
  auto const ordering_spec = window_func->getCollation().front();
  auto order_key_col_null_val_lv =
      get_null_value_by_size_with_encoding(cgen_state_.get(), order_key_buf_ti);

  // load ordering column
  auto [order_col_type_name, order_key_buf_ptr_lv] =
      codegenLoadOrderKeyBufPtr(window_func_context);

  // null range
  auto [null_start_pos_lv, null_end_pos_lv] =
      codegenFrameNullRange(window_func_context, partition_index_lv);

  // compute a row index of the current row w.r.t the window frame it belongs to
  std::string row_idx_on_frame_func = "compute_";
  row_idx_on_frame_func += order_col_type_name;
  row_idx_on_frame_func += ordering_spec.is_desc ? "_greater_equal" : "_less_equal";
  row_idx_on_frame_func += "_current_row_idx_in_frame";
  auto int64_t_one_val_lv = cgen_state_->llInt((int64_t)1);
  auto nulls_first_lv = cgen_state_->llBool(ordering_spec.nulls_first);
  auto cur_row_idx_in_frame_lv =
      cgen_state_->emitCall(row_idx_on_frame_func,
                            {partition_buf_ptrs.num_elem_current_partition_lv,
                             current_row_pos_lv,
                             order_key_buf_ptr_lv,
                             partition_buf_ptrs.target_partition_rowid_ptr_lv,
                             partition_buf_ptrs.target_partition_sorted_rowid_ptr_lv,
                             order_key_col_null_val_lv,
                             nulls_first_lv,
                             null_start_pos_lv,
                             null_end_pos_lv});

  // compute frame bound for the current row
  auto const int64_t_zero_val_lv = cgen_state_->llInt((int64_t)0);
  WindowFrameBoundFuncArgs WindowFrameBoundFuncArgs{
      frame_start_bound_expr_lv,
      frame_end_bound_expr_lv,
      window_func->hasRangeModeFraming() ? current_row_pos_lv : cur_row_idx_in_frame_lv,
      nullptr,
      window_func->hasRangeModeFraming()
          ? int64_t_zero_val_lv
          : partition_buf_ptrs.current_partition_start_offset_lv,
      int64_t_zero_val_lv,
      int64_t_one_val_lv,
      partition_buf_ptrs.num_elem_current_partition_lv,
      order_key_buf_ptr_lv,
      "",
      partition_buf_ptrs.target_partition_rowid_ptr_lv,
      partition_buf_ptrs.target_partition_sorted_rowid_ptr_lv,
      nulls_first_lv,
      null_start_pos_lv,
      null_end_pos_lv};
  auto [frame_start_bound_lv, frame_end_bound_lv] =
      codegenWindowFrameBounds(window_func_context,
                               window_func->getFrameStartBound(),
                               window_func->getFrameEndBound(),
                               order_key_col_null_val_lv,
                               WindowFrameBoundFuncArgs,
                               code_generator);

  // apply offset
  llvm::Value* modified_cur_row_idx_in_frame_lv{nullptr};
  auto const offset_lv =
      cgen_state_->castToTypeIn(code_generator.codegen(args[1].get(), true, co)[0], 64);
  switch (window_func_kind) {
    case SqlWindowFunctionKind::LAG_IN_FRAME:
      modified_cur_row_idx_in_frame_lv =
          cgen_state_->ir_builder_.CreateSub(cur_row_idx_in_frame_lv, offset_lv);
      break;
    case SqlWindowFunctionKind::LEAD_IN_FRAME:
      modified_cur_row_idx_in_frame_lv =
          cgen_state_->ir_builder_.CreateAdd(cur_row_idx_in_frame_lv, offset_lv);
      break;
    case SqlWindowFunctionKind::NTH_VALUE_IN_FRAME: {
      auto candidate_row_idx =
          cgen_state_->ir_builder_.CreateAdd(frame_start_bound_lv, offset_lv);
      auto out_of_frame_bound_lv =
          cgen_state_->ir_builder_.CreateICmpSGT(candidate_row_idx, frame_end_bound_lv);
      // return null if the candidate_row_idx is out of frame bounds
      modified_cur_row_idx_in_frame_lv = cgen_state_->ir_builder_.CreateSelect(
          out_of_frame_bound_lv, cgen_state_->llInt((int64_t)-1), candidate_row_idx);
      break;
    }
    default:
      UNREACHABLE() << "Unsupported window function to navigate a window frame.";
  }
  CHECK(modified_cur_row_idx_in_frame_lv);

  // get the target column value in the frame w.r.t the offset
  std::string target_func_name = "get_";
  target_func_name += target_col_type_name + "_value_";
  target_func_name += target_col_logical_type_name + "_type_";
  target_func_name += "in_frame";
  auto res_lv =
      cgen_state_->emitCall(target_func_name,
                            {modified_cur_row_idx_in_frame_lv,
                             frame_start_bound_lv,
                             frame_end_bound_lv,
                             target_col_buf_lv,
                             partition_buf_ptrs.target_partition_rowid_ptr_lv,
                             partition_buf_ptrs.target_partition_sorted_rowid_ptr_lv,
                             logical_null_val_lv,
                             target_col_null_val_lv});
  if (target_col_ti.get_compression() == kENCODING_DATE_IN_DAYS) {
    res_lv = cgen_state_->emitCall(
        "encode_date",
        {res_lv, logical_null_val_lv, cgen_state_->llInt((int64_t)kSecsPerDay)});
  }
  CHECK(res_lv);
  return res_lv;
}

Here is the call graph for this function:

llvm::BasicBlock * Executor::codegenWindowResetStateControlFlow ( )

private

Definition at line 269 of file WindowFunctionIR.cpp.

References AUTOMATIC_IR_METADATA, WindowProjectNodeContext::getActiveWindowFunctionContext(), CodeGenerator::posArg(), and CodeGenerator::toBool().

                                                             {
  AUTOMATIC_IR_METADATA(cgen_state_.get());
  const auto window_func_context =
      WindowProjectNodeContext::getActiveWindowFunctionContext(this);
  const auto bitset = cgen_state_->llInt(
      reinterpret_cast<const int64_t>(window_func_context->partitionStart()));
  const auto min_val = cgen_state_->llInt(int64_t(0));
  const auto max_val = cgen_state_->llInt(window_func_context->elementCount() - 1);
  const auto null_val = cgen_state_->llInt(inline_int_null_value<int64_t>());
  const auto null_bool_val = cgen_state_->llInt<int8_t>(inline_int_null_value<int8_t>());
  CodeGenerator code_generator(this);
  const auto reset_state =
      code_generator.toBool(cgen_state_->emitCall("bit_is_set",
                                                  {bitset,
                                                   code_generator.posArg(nullptr),
                                                   min_val,
                                                   max_val,
                                                   null_val,
                                                   null_bool_val}));
  const auto reset_state_true_bb = llvm::BasicBlock::Create(
      cgen_state_->context_, "reset_state.true", cgen_state_->current_func_);
  const auto reset_state_false_bb = llvm::BasicBlock::Create(
      cgen_state_->context_, "reset_state.false", cgen_state_->current_func_);
  cgen_state_->ir_builder_.CreateCondBr(
      reset_state, reset_state_true_bb, reset_state_false_bb);
  cgen_state_->ir_builder_.SetInsertPoint(reset_state_true_bb);
  return reset_state_false_bb;
}

Here is the call graph for this function:

ResultSetPtr Executor::collectAllDeviceResults ( SharedKernelContext &  shared_context, const RelAlgExecutionUnit &  ra_exe_unit, const QueryMemoryDescriptor &  query_mem_desc, const ExecutorDeviceType  device_type, std::shared_ptr< RowSetMemoryOwner >  row_set_mem_owner  )

private

Definition at line 2317 of file Execute.cpp.

References anonymous_namespace{Execute.cpp}::build_row_for_empty_input(), collectAllDeviceShardedTopResults(), DEBUG_TIMER, SharedKernelContext::getFragmentResults(), QueryMemoryDescriptor::getQueryDescriptionType(), GPU, NonGroupedAggregate, reduceMultiDeviceResults(), reduceSpeculativeTopN(), GroupByAndAggregate::shard_count_for_top_groups(), RelAlgExecutionUnit::target_exprs, and use_speculative_top_n().

Referenced by executeWorkUnitImpl().

                                                        {
  auto timer = DEBUG_TIMER(__func__);
  auto& result_per_device = shared_context.getFragmentResults();
  if (result_per_device.empty() && query_mem_desc.getQueryDescriptionType() ==
                                       QueryDescriptionType::NonGroupedAggregate) {
    return build_row_for_empty_input(
        ra_exe_unit.target_exprs, query_mem_desc, device_type);
  }
  if (use_speculative_top_n(ra_exe_unit, query_mem_desc)) {
    try {
      return reduceSpeculativeTopN(
          ra_exe_unit, result_per_device, row_set_mem_owner, query_mem_desc);
    } catch (const std::bad_alloc&) {
      throw SpeculativeTopNFailed("Failed during multi-device reduction.");
    }
  }
  const auto shard_count =
      device_type == ExecutorDeviceType::GPU
          ? GroupByAndAggregate::shard_count_for_top_groups(ra_exe_unit)
          : 0;

  if (shard_count && !result_per_device.empty()) {
    return collectAllDeviceShardedTopResults(shared_context, ra_exe_unit);
  }
  return reduceMultiDeviceResults(
      ra_exe_unit, result_per_device, row_set_mem_owner, query_mem_desc);
}

Here is the call graph for this function:

Here is the caller graph for this function:

ResultSetPtr Executor::collectAllDeviceShardedTopResults ( SharedKernelContext &  shared_context, const RelAlgExecutionUnit &  ra_exe_unit  ) const

private

Definition at line 2432 of file Execute.cpp.

References blockSize(), CHECK, CHECK_EQ, CHECK_LE, SharedKernelContext::getFragmentResults(), gridSize(), SortInfo::limit, SortInfo::offset, SortInfo::order_entries, anonymous_namespace{Execute.cpp}::permute_storage_columnar(), anonymous_namespace{Execute.cpp}::permute_storage_row_wise(), run_benchmark_import::result, and RelAlgExecutionUnit::sort_info.

Referenced by collectAllDeviceResults().

                                                  {
  auto& result_per_device = shared_context.getFragmentResults();
  const auto first_result_set = result_per_device.front().first;
  CHECK(first_result_set);
  auto top_query_mem_desc = first_result_set->getQueryMemDesc();
  CHECK(!top_query_mem_desc.hasInterleavedBinsOnGpu());
  const auto top_n = ra_exe_unit.sort_info.limit + ra_exe_unit.sort_info.offset;
  top_query_mem_desc.setEntryCount(0);
  for (auto& result : result_per_device) {
    const auto result_set = result.first;
    CHECK(result_set);
    result_set->sort(ra_exe_unit.sort_info.order_entries, top_n, this);
    size_t new_entry_cnt = top_query_mem_desc.getEntryCount() + result_set->rowCount();
    top_query_mem_desc.setEntryCount(new_entry_cnt);
  }
  auto top_result_set = std::make_shared<ResultSet>(first_result_set->getTargetInfos(),
                                                    first_result_set->getDeviceType(),
                                                    top_query_mem_desc,
                                                    first_result_set->getRowSetMemOwner(),
                                                    blockSize(),
                                                    gridSize());
  auto top_storage = top_result_set->allocateStorage();
  size_t top_output_row_idx{0};
  for (auto& result : result_per_device) {
    const auto result_set = result.first;
    CHECK(result_set);
    const auto& top_permutation = result_set->getPermutationBuffer();
    CHECK_LE(top_permutation.size(), top_n);
    if (top_query_mem_desc.didOutputColumnar()) {
      top_output_row_idx = permute_storage_columnar(result_set->getStorage(),
                                                    result_set->getQueryMemDesc(),
                                                    top_storage,
                                                    top_output_row_idx,
                                                    top_query_mem_desc,
                                                    top_permutation);
    } else {
      top_output_row_idx = permute_storage_row_wise(result_set->getStorage(),
                                                    top_storage,
                                                    top_output_row_idx,
                                                    top_query_mem_desc,
                                                    top_permutation);
    }
  }
  CHECK_EQ(top_output_row_idx, top_query_mem_desc.getEntryCount());
  return top_result_set;
}

Here is the call graph for this function:

Here is the caller graph for this function:

bool Executor::compileBody ( const RelAlgExecutionUnit &  ra_exe_unit, GroupByAndAggregate &  group_by_and_aggregate, QueryMemoryDescriptor &  query_mem_desc, const CompilationOptions &  co, const GpuSharedMemoryContext &  gpu_smem_context = {}  )

private

Definition at line 3254 of file NativeCodegen.cpp.

                                                                           {
  AUTOMATIC_IR_METADATA(cgen_state_.get());

  // Switch the code generation into a separate filter function if enabled.
  // Note that accesses to function arguments are still codegenned from the
  // row function's arguments, then later automatically forwarded and
  // remapped into filter function arguments by redeclareFilterFunction().
  cgen_state_->row_func_bb_ = cgen_state_->ir_builder_.GetInsertBlock();
  llvm::Value* loop_done{nullptr};
  std::unique_ptr<Executor::FetchCacheAnchor> fetch_cache_anchor;
  if (cgen_state_->filter_func_) {
    if (cgen_state_->row_func_bb_->getName() == "loop_body") {
      auto row_func_entry_bb = &cgen_state_->row_func_->getEntryBlock();
      cgen_state_->ir_builder_.SetInsertPoint(row_func_entry_bb,
                                              row_func_entry_bb->begin());
      loop_done = cgen_state_->ir_builder_.CreateAlloca(
          get_int_type(1, cgen_state_->context_), nullptr, "loop_done");
      cgen_state_->ir_builder_.SetInsertPoint(cgen_state_->row_func_bb_);
      cgen_state_->ir_builder_.CreateStore(cgen_state_->llBool(true), loop_done);
    }
    cgen_state_->ir_builder_.SetInsertPoint(cgen_state_->filter_func_bb_);
    cgen_state_->current_func_ = cgen_state_->filter_func_;
    fetch_cache_anchor = std::make_unique<Executor::FetchCacheAnchor>(cgen_state_.get());
  }

  // generate the code for the filter
  std::vector<Analyzer::Expr*> primary_quals;
  std::vector<Analyzer::Expr*> deferred_quals;
  bool short_circuited = CodeGenerator::prioritizeQuals(
      ra_exe_unit, primary_quals, deferred_quals, plan_state_->hoisted_filters_);
  if (short_circuited) {
    VLOG(1) << "Prioritized " << std::to_string(primary_quals.size()) << " quals, "
            << "short-circuited and deferred " << std::to_string(deferred_quals.size())
            << " quals";
  }
  llvm::Value* filter_lv = cgen_state_->llBool(true);
  CodeGenerator code_generator(this);
  for (auto expr : primary_quals) {
    // Generate the filter for primary quals
    auto cond = code_generator.toBool(code_generator.codegen(expr, true, co).front());
    filter_lv = cgen_state_->ir_builder_.CreateAnd(filter_lv, cond);
  }
  CHECK(filter_lv->getType()->isIntegerTy(1));
  llvm::BasicBlock* sc_false{nullptr};
  if (!deferred_quals.empty()) {
    auto sc_true = llvm::BasicBlock::Create(
        cgen_state_->context_, "sc_true", cgen_state_->current_func_);
    sc_false = llvm::BasicBlock::Create(
        cgen_state_->context_, "sc_false", cgen_state_->current_func_);
    cgen_state_->ir_builder_.CreateCondBr(filter_lv, sc_true, sc_false);
    cgen_state_->ir_builder_.SetInsertPoint(sc_false);
    if (ra_exe_unit.join_quals.empty()) {
      cgen_state_->ir_builder_.CreateRet(cgen_state_->llInt(int32_t(0)));
    }
    cgen_state_->ir_builder_.SetInsertPoint(sc_true);
    filter_lv = cgen_state_->llBool(true);
  }
  for (auto expr : deferred_quals) {
    filter_lv = cgen_state_->ir_builder_.CreateAnd(
        filter_lv, code_generator.toBool(code_generator.codegen(expr, true, co).front()));
  }

  CHECK(filter_lv->getType()->isIntegerTy(1));
  auto ret = group_by_and_aggregate.codegen(
      filter_lv, sc_false, query_mem_desc, co, gpu_smem_context);

  // Switch the code generation back to the row function if a filter
  // function was enabled.
  if (cgen_state_->filter_func_) {
    if (cgen_state_->row_func_bb_->getName() == "loop_body") {
      cgen_state_->ir_builder_.CreateStore(cgen_state_->llBool(false), loop_done);
      cgen_state_->ir_builder_.CreateRet(cgen_state_->llInt<int32_t>(0));
    }

    cgen_state_->ir_builder_.SetInsertPoint(cgen_state_->row_func_bb_);
    cgen_state_->current_func_ = cgen_state_->row_func_;
    cgen_state_->filter_func_call_ =
        cgen_state_->ir_builder_.CreateCall(cgen_state_->filter_func_, {});

    // Create real filter function declaration after placeholder call
    // is emitted.
    redeclareFilterFunction();

    if (cgen_state_->row_func_bb_->getName() == "loop_body") {
      auto loop_done_true = llvm::BasicBlock::Create(
          cgen_state_->context_, "loop_done_true", cgen_state_->row_func_);
      auto loop_done_false = llvm::BasicBlock::Create(
          cgen_state_->context_, "loop_done_false", cgen_state_->row_func_);
      auto loop_done_flag = cgen_state_->ir_builder_.CreateLoad(
          loop_done->getType()->getPointerElementType(), loop_done);
      cgen_state_->ir_builder_.CreateCondBr(
          loop_done_flag, loop_done_true, loop_done_false);
      cgen_state_->ir_builder_.SetInsertPoint(loop_done_true);
      cgen_state_->ir_builder_.CreateRet(cgen_state_->filter_func_call_);
      cgen_state_->ir_builder_.SetInsertPoint(loop_done_false);
    } else {
      cgen_state_->ir_builder_.CreateRet(cgen_state_->filter_func_call_);
    }
  }
  return ret;
}

std::tuple< CompilationResult, std::unique_ptr< QueryMemoryDescriptor > > Executor::compileWorkUnit ( const std::vector< InputTableInfo > &  query_infos, const PlanState::DeletedColumnsMap &  deleted_cols_map, const RelAlgExecutionUnit &  ra_exe_unit, const CompilationOptions &  co, const ExecutionOptions &  eo, const CudaMgr_Namespace::CudaMgr *  cuda_mgr, const bool  allow_lazy_fetch, std::shared_ptr< RowSetMemoryOwner >  row_set_mem_owner, const size_t  max_groups_buffer_entry_count, const int8_t  crt_min_byte_width, const bool  has_cardinality_estimation, ColumnCacheMap &  column_cache, RenderInfo *  render_info = nullptr  )

private

Definition at line 2708 of file NativeCodegen.cpp.

                                                   {
  auto timer = DEBUG_TIMER(__func__);

  if (co.device_type == ExecutorDeviceType::GPU) {
    if (!cuda_mgr) {
      throw QueryMustRunOnCpu();
    }
  }

#ifndef NDEBUG
  static std::uint64_t counter = 0;
  ++counter;
  VLOG(1) << "CODEGEN #" << counter << ":";
  LOG(IR) << "CODEGEN #" << counter << ":";
  LOG(PTX) << "CODEGEN #" << counter << ":";
  LOG(ASM) << "CODEGEN #" << counter << ":";
#endif

  // cgenstate_manager uses RAII pattern to manage the live time of
  // CgenState instances.
  Executor::CgenStateManager cgenstate_manager(*this,
                                               allow_lazy_fetch,
                                               query_infos,
                                               deleted_cols_map,
                                               &ra_exe_unit);  // locks compilation_mutex

  addTransientStringLiterals(ra_exe_unit, row_set_mem_owner);

  GroupByAndAggregate group_by_and_aggregate(
      this,
      co.device_type,
      ra_exe_unit,
      query_infos,
      row_set_mem_owner,
      has_cardinality_estimation ? std::optional<int64_t>(max_groups_buffer_entry_guess)
                                 : std::nullopt);
  auto query_mem_desc =
      group_by_and_aggregate.initQueryMemoryDescriptor(eo.allow_multifrag,
                                                       max_groups_buffer_entry_guess,
                                                       crt_min_byte_width,
                                                       render_info,
                                                       eo.output_columnar_hint);

  if (query_mem_desc->getQueryDescriptionType() ==
          QueryDescriptionType::GroupByBaselineHash &&
      !has_cardinality_estimation && (!render_info || !render_info->isInSitu()) &&
      !eo.just_explain) {
    const auto col_range_info = group_by_and_aggregate.getColRangeInfo();
    throw CardinalityEstimationRequired(col_range_info.max - col_range_info.min);
  }

  const bool output_columnar = query_mem_desc->didOutputColumnar();
  const bool gpu_shared_mem_optimization =
      is_gpu_shared_mem_supported(query_mem_desc.get(),
                                  ra_exe_unit,
                                  cuda_mgr,
                                  co.device_type,
                                  cuda_mgr ? this->blockSize() : 1,
                                  cuda_mgr ? this->numBlocksPerMP() : 1);
  if (gpu_shared_mem_optimization) {
    // disable interleaved bins optimization on the GPU
    query_mem_desc->setHasInterleavedBinsOnGpu(false);
    LOG(DEBUG1) << "GPU shared memory is used for the " +
                       query_mem_desc->queryDescTypeToString() + " query(" +
                       std::to_string(get_shared_memory_size(gpu_shared_mem_optimization,
                                                             query_mem_desc.get())) +
                       " out of " + std::to_string(g_gpu_smem_threshold) + " bytes).";
  }

  const GpuSharedMemoryContext gpu_smem_context(
      get_shared_memory_size(gpu_shared_mem_optimization, query_mem_desc.get()));

  if (co.device_type == ExecutorDeviceType::GPU) {
    const size_t num_count_distinct_descs =
        query_mem_desc->getCountDistinctDescriptorsSize();
    for (size_t i = 0; i < num_count_distinct_descs; i++) {
      const auto& count_distinct_descriptor =
          query_mem_desc->getCountDistinctDescriptor(i);
      if (count_distinct_descriptor.impl_type_ == CountDistinctImplType::UnorderedSet ||
          (count_distinct_descriptor.impl_type_ != CountDistinctImplType::Invalid &&
           !co.hoist_literals)) {
        throw QueryMustRunOnCpu();
      }
    }

    // we currently do not support varlen projection based on baseline groupby when
    // 1) target table is multi-fragmented and 2) multiple gpus are involved for query
    // processing in this case, we punt the query to cpu to avoid server crash
    for (const auto expr : ra_exe_unit.target_exprs) {
      if (auto gby_expr = dynamic_cast<Analyzer::AggExpr*>(expr)) {
        bool has_multiple_gpus = cuda_mgr ? cuda_mgr->getDeviceCount() > 1 : false;
        if (gby_expr->get_aggtype() == SQLAgg::kSAMPLE && has_multiple_gpus &&
            !g_leaf_count) {
          std::set<const Analyzer::ColumnVar*,
                   bool (*)(const Analyzer::ColumnVar*, const Analyzer::ColumnVar*)>
              colvar_set(