You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@quickstep.apache.org by ji...@apache.org on 2016/11/01 06:16:45 UTC
[1/5] incubator-quickstep git commit: Cleanup FastHashTable + Avoid
unnecessary copying of data before aggregation. [Forced Update!]
Repository: incubator-quickstep
Updated Branches:
refs/heads/collision-free-agg 17067ca27 -> 7048fcc1f (forced update)
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/PackedPayloadAggregationStateHashTable.cpp
----------------------------------------------------------------------
diff --git a/storage/PackedPayloadAggregationStateHashTable.cpp b/storage/PackedPayloadAggregationStateHashTable.cpp
new file mode 100644
index 0000000..34c4177
--- /dev/null
+++ b/storage/PackedPayloadAggregationStateHashTable.cpp
@@ -0,0 +1,434 @@
+/**
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ **/
+
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
+
+namespace quickstep {
+
+PackedPayloadSeparateChainingAggregationStateHashTable
+ ::PackedPayloadSeparateChainingAggregationStateHashTable(
+ const std::vector<const Type *> &key_types,
+ const std::size_t num_entries,
+ const std::vector<AggregationHandle *> &handles,
+ StorageManager *storage_manager)
+ : key_types_(key_types),
+ num_handles_(handles.size()),
+ handles_(handles),
+ total_payload_size_(ComputeTotalPayloadSize(handles)),
+ storage_manager_(storage_manager),
+ kBucketAlignment(alignof(std::atomic<std::size_t>)),
+ kValueOffset(sizeof(std::atomic<std::size_t>) + sizeof(std::size_t)),
+ key_manager_(key_types_, kValueOffset + total_payload_size_),
+ bucket_size_(ComputeBucketSize(key_manager_.getFixedKeySize())) {
+ std::size_t payload_offset_running_sum = sizeof(SpinMutex);
+ for (const auto *handle : handles) {
+ payload_offsets_.emplace_back(payload_offset_running_sum);
+ payload_offset_running_sum += handle->getPayloadSize();
+ }
+
+ // NOTE(jianqiao): Potential memory leak / double freeing by copying from
+ // init_payload to buckets if payload contains out of line data.
+ init_payload_ =
+ static_cast<std::uint8_t *>(calloc(this->total_payload_size_, 1));
+ DCHECK(init_payload_ != nullptr);
+
+ for (std::size_t i = 0; i < num_handles_; ++i) {
+ handles_[i]->initPayload(init_payload_ + payload_offsets_[i]);
+ }
+
+ // Bucket size always rounds up to the alignment requirement of the atomic
+ // size_t "next" pointer at the front or a ValueT, whichever is larger.
+ //
+ // Give base HashTable information about what key components are stored
+ // inline from 'key_manager_'.
+ setKeyInline(key_manager_.getKeyInline());
+
+ // Pick out a prime number of slots and calculate storage requirements.
+ std::size_t num_slots_tmp =
+ get_next_prime_number(num_entries * kHashTableLoadFactor);
+ std::size_t required_memory =
+ sizeof(Header) + num_slots_tmp * sizeof(std::atomic<std::size_t>) +
+ (num_slots_tmp / kHashTableLoadFactor) *
+ (bucket_size_ + key_manager_.getEstimatedVariableKeySize());
+ std::size_t num_storage_slots =
+ this->storage_manager_->SlotsNeededForBytes(required_memory);
+ if (num_storage_slots == 0) {
+ FATAL_ERROR(
+ "Storage requirement for SeparateChainingHashTable "
+ "exceeds maximum allocation size.");
+ }
+
+ // Get a StorageBlob to hold the hash table.
+ const block_id blob_id =
+ this->storage_manager_->createBlob(num_storage_slots);
+ this->blob_ = this->storage_manager_->getBlobMutable(blob_id);
+
+ void *aligned_memory_start = this->blob_->getMemoryMutable();
+ std::size_t available_memory = num_storage_slots * kSlotSizeBytes;
+ if (align(alignof(Header),
+ sizeof(Header),
+ aligned_memory_start,
+ available_memory) == nullptr) {
+ // With current values from StorageConstants.hpp, this should be
+ // impossible. A blob is at least 1 MB, while a Header has alignment
+ // requirement of just kCacheLineBytes (64 bytes).
+ FATAL_ERROR(
+ "StorageBlob used to hold resizable "
+ "SeparateChainingHashTable is too small to meet alignment "
+ "requirements of SeparateChainingHashTable::Header.");
+ } else if (aligned_memory_start != this->blob_->getMemoryMutable()) {
+ // This should also be impossible, since the StorageManager allocates slots
+ // aligned to kCacheLineBytes.
+ DEV_WARNING("StorageBlob memory adjusted by "
+ << (num_storage_slots * kSlotSizeBytes - available_memory)
+ << " bytes to meet alignment requirement for "
+ << "SeparateChainingHashTable::Header.");
+ }
+
+ // Locate the header.
+ header_ = static_cast<Header *>(aligned_memory_start);
+ aligned_memory_start =
+ static_cast<char *>(aligned_memory_start) + sizeof(Header);
+ available_memory -= sizeof(Header);
+
+ // Recompute the number of slots & buckets using the actual available memory.
+ // Most likely, we got some extra free bucket space due to "rounding up" to
+ // the storage blob's size. It's also possible (though very unlikely) that we
+ // will wind up with fewer buckets than we initially wanted because of screwy
+ // alignment requirements for ValueT.
+ std::size_t num_buckets_tmp =
+ available_memory /
+ (kHashTableLoadFactor * sizeof(std::atomic<std::size_t>) + bucket_size_ +
+ key_manager_.getEstimatedVariableKeySize());
+ num_slots_tmp =
+ get_previous_prime_number(num_buckets_tmp * kHashTableLoadFactor);
+ num_buckets_tmp = num_slots_tmp / kHashTableLoadFactor;
+ DEBUG_ASSERT(num_slots_tmp > 0);
+ DEBUG_ASSERT(num_buckets_tmp > 0);
+
+ // Locate the slot array.
+ slots_ = static_cast<std::atomic<std::size_t> *>(aligned_memory_start);
+ aligned_memory_start = static_cast<char *>(aligned_memory_start) +
+ sizeof(std::atomic<std::size_t>) * num_slots_tmp;
+ available_memory -= sizeof(std::atomic<std::size_t>) * num_slots_tmp;
+
+ // Locate the buckets.
+ buckets_ = aligned_memory_start;
+ // Extra-paranoid: If ValueT has an alignment requirement greater than that
+ // of std::atomic<std::size_t>, we may need to adjust the start of the bucket
+ // array.
+ if (align(kBucketAlignment, bucket_size_, buckets_, available_memory) ==
+ nullptr) {
+ FATAL_ERROR(
+ "StorageBlob used to hold resizable "
+ "SeparateChainingHashTable is too small to meet "
+ "alignment requirements of buckets.");
+ } else if (buckets_ != aligned_memory_start) {
+ DEV_WARNING(
+ "Bucket array start position adjusted to meet alignment "
+ "requirement for SeparateChainingHashTable's value type.");
+ if (num_buckets_tmp * bucket_size_ > available_memory) {
+ --num_buckets_tmp;
+ }
+ }
+
+ // Fill in the header.
+ header_->num_slots = num_slots_tmp;
+ header_->num_buckets = num_buckets_tmp;
+ header_->buckets_allocated.store(0, std::memory_order_relaxed);
+ header_->variable_length_bytes_allocated.store(0, std::memory_order_relaxed);
+ available_memory -= bucket_size_ * (header_->num_buckets);
+
+ // Locate variable-length key storage region, and give it all the remaining
+ // bytes in the blob.
+ key_manager_.setVariableLengthStorageInfo(
+ static_cast<char *>(buckets_) + header_->num_buckets * bucket_size_,
+ available_memory,
+ &(header_->variable_length_bytes_allocated));
+}
+
+PackedPayloadSeparateChainingAggregationStateHashTable
+ ::~PackedPayloadSeparateChainingAggregationStateHashTable() {
+ if (blob_.valid()) {
+ const block_id blob_id = blob_->getID();
+ blob_.release();
+ storage_manager_->deleteBlockOrBlobFile(blob_id);
+ }
+ std::free(init_payload_);
+}
+
+void PackedPayloadSeparateChainingAggregationStateHashTable::clear() {
+ const std::size_t used_buckets =
+ header_->buckets_allocated.load(std::memory_order_relaxed);
+ // Destroy existing values, if necessary.
+ destroyPayload();
+
+ // Zero-out slot array.
+ std::memset(
+ slots_, 0x0, sizeof(std::atomic<std::size_t>) * header_->num_slots);
+
+ // Zero-out used buckets.
+ std::memset(buckets_, 0x0, used_buckets * bucket_size_);
+
+ header_->buckets_allocated.store(0, std::memory_order_relaxed);
+ header_->variable_length_bytes_allocated.store(0, std::memory_order_relaxed);
+ key_manager_.zeroNextVariableLengthKeyOffset();
+}
+
+void PackedPayloadSeparateChainingAggregationStateHashTable::destroyPayload() {
+ const std::size_t num_buckets =
+ header_->buckets_allocated.load(std::memory_order_relaxed);
+ void *bucket_ptr = static_cast<char *>(buckets_) + kValueOffset;
+ for (std::size_t bucket_num = 0; bucket_num < num_buckets; ++bucket_num) {
+ for (std::size_t handle_id = 0; handle_id < num_handles_; ++handle_id) {
+ void *value_internal_ptr =
+ static_cast<char *>(bucket_ptr) + this->payload_offsets_[handle_id];
+ handles_[handle_id]->destroyPayload(static_cast<std::uint8_t *>(value_internal_ptr));
+ }
+ bucket_ptr = static_cast<char *>(bucket_ptr) + bucket_size_;
+ }
+}
+
+bool PackedPayloadSeparateChainingAggregationStateHashTable::upsertValueAccessor(
+ const std::vector<std::vector<attribute_id>> &argument_ids,
+ const std::vector<attribute_id> &key_attr_ids,
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor) {
+ if (aux_accessor == nullptr) {
+ return upsertValueAccessorCompositeKeyInternal<false>(argument_ids,
+ key_attr_ids,
+ accessor,
+ aux_accessor);
+ } else {
+ return upsertValueAccessorCompositeKeyInternal<true>(argument_ids,
+ key_attr_ids,
+ accessor,
+ aux_accessor);
+ }
+}
+
+void PackedPayloadSeparateChainingAggregationStateHashTable
+ ::resize(const std::size_t extra_buckets,
+ const std::size_t extra_variable_storage,
+ const std::size_t retry_num) {
+ // A retry should never be necessary with this implementation of HashTable.
+ // Separate chaining ensures that any resized hash table with more buckets
+ // than the original table will be able to hold more entries than the
+ // original.
+ DEBUG_ASSERT(retry_num == 0);
+
+ SpinSharedMutexExclusiveLock<true> write_lock(this->resize_shared_mutex_);
+
+ // Recheck whether the hash table is still full. Note that multiple threads
+ // might wait to rebuild this hash table simultaneously. Only the first one
+ // should do the rebuild.
+ if (!isFull(extra_variable_storage)) {
+ return;
+ }
+
+ // Approximately double the number of buckets and slots.
+ //
+ // TODO(chasseur): It may be worth it to more than double the number of
+ // buckets here so that we can maintain a good, sparse fill factor for a
+ // longer time as more values are inserted. Such behavior should take into
+ // account kHashTableLoadFactor.
+ std::size_t resized_num_slots = get_next_prime_number(
+ (header_->num_buckets + extra_buckets / 2) * kHashTableLoadFactor * 2);
+ std::size_t variable_storage_required =
+ (resized_num_slots / kHashTableLoadFactor) *
+ key_manager_.getEstimatedVariableKeySize();
+ const std::size_t original_variable_storage_used =
+ header_->variable_length_bytes_allocated.load(std::memory_order_relaxed);
+ // If this resize was triggered by a too-large variable-length key, bump up
+ // the variable-length storage requirement.
+ if ((extra_variable_storage > 0) &&
+ (extra_variable_storage + original_variable_storage_used >
+ key_manager_.getVariableLengthKeyStorageSize())) {
+ variable_storage_required += extra_variable_storage;
+ }
+
+ const std::size_t resized_memory_required =
+ sizeof(Header) + resized_num_slots * sizeof(std::atomic<std::size_t>) +
+ (resized_num_slots / kHashTableLoadFactor) * bucket_size_ +
+ variable_storage_required;
+ const std::size_t resized_storage_slots =
+ this->storage_manager_->SlotsNeededForBytes(resized_memory_required);
+ if (resized_storage_slots == 0) {
+ FATAL_ERROR(
+ "Storage requirement for resized SeparateChainingHashTable "
+ "exceeds maximum allocation size.");
+ }
+
+ // Get a new StorageBlob to hold the resized hash table.
+ const block_id resized_blob_id =
+ this->storage_manager_->createBlob(resized_storage_slots);
+ MutableBlobReference resized_blob =
+ this->storage_manager_->getBlobMutable(resized_blob_id);
+
+ // Locate data structures inside the new StorageBlob.
+ void *aligned_memory_start = resized_blob->getMemoryMutable();
+ std::size_t available_memory = resized_storage_slots * kSlotSizeBytes;
+ if (align(alignof(Header),
+ sizeof(Header),
+ aligned_memory_start,
+ available_memory) == nullptr) {
+ // Should be impossible, as noted in constructor.
+ FATAL_ERROR(
+ "StorageBlob used to hold resized SeparateChainingHashTable "
+ "is too small to meet alignment requirements of "
+ "LinearOpenAddressingHashTable::Header.");
+ } else if (aligned_memory_start != resized_blob->getMemoryMutable()) {
+ // Again, should be impossible.
+ DEV_WARNING("In SeparateChainingHashTable::resize(), StorageBlob "
+ << "memory adjusted by "
+ << (resized_num_slots * kSlotSizeBytes - available_memory)
+ << " bytes to meet alignment requirement for "
+ << "LinearOpenAddressingHashTable::Header.");
+ }
+
+ Header *resized_header = static_cast<Header *>(aligned_memory_start);
+ aligned_memory_start =
+ static_cast<char *>(aligned_memory_start) + sizeof(Header);
+ available_memory -= sizeof(Header);
+
+ // As in constructor, recompute the number of slots and buckets using the
+ // actual available memory.
+ std::size_t resized_num_buckets =
+ (available_memory - extra_variable_storage) /
+ (kHashTableLoadFactor * sizeof(std::atomic<std::size_t>) + bucket_size_ +
+ key_manager_.getEstimatedVariableKeySize());
+ resized_num_slots =
+ get_previous_prime_number(resized_num_buckets * kHashTableLoadFactor);
+ resized_num_buckets = resized_num_slots / kHashTableLoadFactor;
+
+ // Locate slot array.
+ std::atomic<std::size_t> *resized_slots =
+ static_cast<std::atomic<std::size_t> *>(aligned_memory_start);
+ aligned_memory_start = static_cast<char *>(aligned_memory_start) +
+ sizeof(std::atomic<std::size_t>) * resized_num_slots;
+ available_memory -= sizeof(std::atomic<std::size_t>) * resized_num_slots;
+
+ // As in constructor, we will be extra paranoid and use align() to locate the
+ // start of the array of buckets, as well.
+ void *resized_buckets = aligned_memory_start;
+ if (align(
+ kBucketAlignment, bucket_size_, resized_buckets, available_memory) ==
+ nullptr) {
+ FATAL_ERROR(
+ "StorageBlob used to hold resized SeparateChainingHashTable "
+ "is too small to meet alignment requirements of buckets.");
+ } else if (resized_buckets != aligned_memory_start) {
+ DEV_WARNING(
+ "Bucket array start position adjusted to meet alignment "
+ "requirement for SeparateChainingHashTable's value type.");
+ if (resized_num_buckets * bucket_size_ + variable_storage_required >
+ available_memory) {
+ --resized_num_buckets;
+ }
+ }
+ aligned_memory_start = static_cast<char *>(aligned_memory_start) +
+ resized_num_buckets * bucket_size_;
+ available_memory -= resized_num_buckets * bucket_size_;
+
+ void *resized_variable_length_key_storage = aligned_memory_start;
+ const std::size_t resized_variable_length_key_storage_size = available_memory;
+
+ const std::size_t original_buckets_used =
+ header_->buckets_allocated.load(std::memory_order_relaxed);
+
+ // Initialize the header.
+ resized_header->num_slots = resized_num_slots;
+ resized_header->num_buckets = resized_num_buckets;
+ resized_header->buckets_allocated.store(original_buckets_used,
+ std::memory_order_relaxed);
+ resized_header->variable_length_bytes_allocated.store(
+ original_variable_storage_used, std::memory_order_relaxed);
+
+ // Bulk-copy buckets. This is safe because:
+ // 1. The "next" pointers will be adjusted when rebuilding chains below.
+ // 2. The hash codes will stay the same.
+ // 3. For key components:
+ // a. Inline keys will stay exactly the same.
+ // b. Offsets into variable-length storage will remain valid, because
+ // we also do a byte-for-byte copy of variable-length storage below.
+ // c. Absolute external pointers will still point to the same address.
+ // d. Relative pointers are not used with resizable hash tables.
+ // 4. If values are not trivially copyable, then we invoke ValueT's copy
+ // or move constructor with placement new.
+ // NOTE(harshad) - Regarding point 4 above, as this is a specialized
+ // hash table implemented for aggregation, the values are trivially copyable,
+ // therefore we don't need to invoke payload values' copy/move constructors.
+ std::memcpy(resized_buckets, buckets_, original_buckets_used * bucket_size_);
+
+ // Copy over variable-length key components, if any.
+ if (original_variable_storage_used > 0) {
+ DEBUG_ASSERT(original_variable_storage_used ==
+ key_manager_.getNextVariableLengthKeyOffset());
+ DEBUG_ASSERT(original_variable_storage_used <=
+ resized_variable_length_key_storage_size);
+ std::memcpy(resized_variable_length_key_storage,
+ key_manager_.getVariableLengthKeyStorage(),
+ original_variable_storage_used);
+ }
+
+ destroyPayload();
+
+ // Make resized structures active.
+ std::swap(this->blob_, resized_blob);
+ header_ = resized_header;
+ slots_ = resized_slots;
+ buckets_ = resized_buckets;
+ key_manager_.setVariableLengthStorageInfo(
+ resized_variable_length_key_storage,
+ resized_variable_length_key_storage_size,
+ &(resized_header->variable_length_bytes_allocated));
+
+ // Drop the old blob.
+ const block_id old_blob_id = resized_blob->getID();
+ resized_blob.release();
+ this->storage_manager_->deleteBlockOrBlobFile(old_blob_id);
+
+ // Rebuild chains.
+ void *current_bucket = buckets_;
+ for (std::size_t bucket_num = 0; bucket_num < original_buckets_used;
+ ++bucket_num) {
+ std::atomic<std::size_t> *next_ptr =
+ static_cast<std::atomic<std::size_t> *>(current_bucket);
+ const std::size_t hash_code = *reinterpret_cast<const std::size_t *>(
+ static_cast<const char *>(current_bucket) +
+ sizeof(std::atomic<std::size_t>));
+
+ const std::size_t slot_number = hash_code % header_->num_slots;
+ std::size_t slot_ptr_value = 0;
+ if (slots_[slot_number].compare_exchange_strong(
+ slot_ptr_value, bucket_num + 1, std::memory_order_relaxed)) {
+ // This bucket is the first in the chain for this block, so reset its
+ // next pointer to 0.
+ next_ptr->store(0, std::memory_order_relaxed);
+ } else {
+ // A chain already exists starting from this slot, so put this bucket at
+ // the head.
+ next_ptr->store(slot_ptr_value, std::memory_order_relaxed);
+ slots_[slot_number].store(bucket_num + 1, std::memory_order_relaxed);
+ }
+ current_bucket = static_cast<char *>(current_bucket) + bucket_size_;
+ }
+}
+
+} // namespace quickstep
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/PackedPayloadAggregationStateHashTable.hpp
----------------------------------------------------------------------
diff --git a/storage/PackedPayloadAggregationStateHashTable.hpp b/storage/PackedPayloadAggregationStateHashTable.hpp
new file mode 100644
index 0000000..70152e7
--- /dev/null
+++ b/storage/PackedPayloadAggregationStateHashTable.hpp
@@ -0,0 +1,721 @@
+/**
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ **/
+
+#ifndef QUICKSTEP_STORAGE_PACKED_PAYLOAD_AGGREGATION_STATE_HASH_TABLE_HPP_
+#define QUICKSTEP_STORAGE_PACKED_PAYLOAD_AGGREGATION_STATE_HASH_TABLE_HPP_
+
+#include <algorithm>
+#include <atomic>
+#include <cstddef>
+#include <cstdlib>
+#include <limits>
+#include <memory>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "catalog/CatalogTypedefs.hpp"
+#include "expressions/aggregation/AggregationHandle.hpp"
+#include "storage/HashTableBase.hpp"
+#include "storage/HashTableKeyManager.hpp"
+#include "storage/StorageBlob.hpp"
+#include "storage/StorageBlockInfo.hpp"
+#include "storage/StorageConstants.hpp"
+#include "storage/StorageManager.hpp"
+#include "storage/TupleReference.hpp"
+#include "storage/ValueAccessor.hpp"
+#include "storage/ValueAccessorUtil.hpp"
+#include "threading/SpinMutex.hpp"
+#include "threading/SpinSharedMutex.hpp"
+#include "types/Type.hpp"
+#include "types/TypedValue.hpp"
+#include "types/containers/ColumnVectorsValueAccessor.hpp"
+#include "utility/Alignment.hpp"
+#include "utility/HashPair.hpp"
+#include "utility/Macros.hpp"
+#include "utility/PrimeNumber.hpp"
+
+namespace quickstep {
+
+/** \addtogroup Storage
+ * @{
+ */
+
+class PackedPayloadSeparateChainingAggregationStateHashTable
+ : public AggregationStateHashTableBase {
+ public:
+ PackedPayloadSeparateChainingAggregationStateHashTable(
+ const std::vector<const Type *> &key_types,
+ const std::size_t num_entries,
+ const std::vector<AggregationHandle *> &handles,
+ StorageManager *storage_manager);
+
+ ~PackedPayloadSeparateChainingAggregationStateHashTable() override;
+
+ void clear();
+
+ void destroyPayload() override;
+
+ bool upsertValueAccessor(
+ const std::vector<std::vector<attribute_id>> &argument_ids,
+ const std::vector<attribute_id> &key_attr_ids,
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor = nullptr) override;
+
+ inline block_id getBlobId() const {
+ return blob_->getID();
+ }
+
+ inline std::size_t numEntries() const {
+ return header_->buckets_allocated.load(std::memory_order_relaxed);
+ }
+
+ inline bool upsertCompositeKey(const std::vector<TypedValue> &key,
+ const std::uint8_t *source_state);
+
+ inline const std::uint8_t* getSingleCompositeKey(
+ const std::vector<TypedValue> &key) const;
+
+ inline const std::uint8_t* getSingleCompositeKey(
+ const std::vector<TypedValue> &key,
+ const int index) const;
+
+ template <typename FunctorT>
+ inline std::size_t forEach(FunctorT *functor) const;
+
+ template <typename FunctorT>
+ inline std::size_t forEach(FunctorT *functor, const int index) const;
+
+ private:
+ void resize(const std::size_t extra_buckets,
+ const std::size_t extra_variable_storage,
+ const std::size_t retry_num = 0);
+
+ inline std::size_t calculateVariableLengthCompositeKeyCopySize(
+ const std::vector<TypedValue> &key) const {
+ std::size_t total = 0;
+ for (std::vector<TypedValue>::size_type idx = 0; idx < key.size(); ++idx) {
+ if (!(*key_inline_)[idx]) {
+ total += key[idx].getDataSize();
+ }
+ }
+ return total;
+ }
+
+ inline bool getNextEntryCompositeKey(std::vector<TypedValue> *key,
+ const std::uint8_t **value,
+ std::size_t *entry_num) const;
+
+ inline std::uint8_t* upsertCompositeKeyInternal(
+ const std::vector<TypedValue> &key,
+ const std::size_t variable_key_size);
+
+ template <bool has_aux_accessor>
+ inline bool upsertValueAccessorCompositeKeyInternal(
+ const std::vector<std::vector<attribute_id>> &argument_ids,
+ const std::vector<attribute_id> &key_attr_ids,
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor);
+
+ // Generate a hash for a composite key by hashing each component of 'key' and
+ // mixing their bits with CombineHashes().
+ inline std::size_t hashCompositeKey(const std::vector<TypedValue> &key) const;
+
+ // Set information about which key components are stored inline. This usually
+ // comes from a HashTableKeyManager, and is set by the constructor of a
+ // subclass of HashTable.
+ inline void setKeyInline(const std::vector<bool> *key_inline) {
+ scalar_key_inline_ = key_inline->front();
+ key_inline_ = key_inline;
+ }
+
+ inline static std::size_t ComputeTotalPayloadSize(
+ const std::vector<AggregationHandle *> &handles) {
+ std::size_t total_payload_size = sizeof(SpinMutex);
+ for (const auto *handle : handles) {
+ total_payload_size += handle->getPayloadSize();
+ }
+ return total_payload_size;
+ }
+
+ // Assign '*key_vector' with the attribute values specified by 'key_attr_ids'
+ // at the current position of 'accessor'. If 'check_for_null_keys' is true,
+ // stops and returns true if any of the values is null, otherwise returns
+ // false.
+ template <typename ValueAccessorT>
+ inline static bool GetCompositeKeyFromValueAccessor(
+ const ValueAccessorT &accessor,
+ const std::vector<attribute_id> &key_attr_ids,
+ const bool check_for_null_keys,
+ std::vector<TypedValue> *key_vector) {
+ for (std::vector<attribute_id>::size_type key_idx = 0;
+ key_idx < key_attr_ids.size();
+ ++key_idx) {
+ (*key_vector)[key_idx] = accessor.getTypedValue(key_attr_ids[key_idx]);
+ if (check_for_null_keys && (*key_vector)[key_idx].isNull()) {
+ return true;
+ }
+ }
+ return false;
+ }
+
+ struct Header {
+ std::size_t num_slots;
+ std::size_t num_buckets;
+ alignas(kCacheLineBytes) std::atomic<std::size_t> buckets_allocated;
+ alignas(kCacheLineBytes)
+ std::atomic<std::size_t> variable_length_bytes_allocated;
+ };
+
+ // Type(s) of keys.
+ const std::vector<const Type *> key_types_;
+
+ // Information about whether key components are stored inline or in a
+ // separate variable-length storage region. This is usually determined by a
+ // HashTableKeyManager and set by calling setKeyInline().
+ bool scalar_key_inline_;
+ const std::vector<bool> *key_inline_;
+
+ const std::size_t num_handles_;
+ const std::vector<AggregationHandle *> handles_;
+
+ std::size_t total_payload_size_;
+ std::vector<std::size_t> payload_offsets_;
+ std::uint8_t *init_payload_;
+
+ StorageManager *storage_manager_;
+ MutableBlobReference blob_;
+
+ // Locked in shared mode for most operations, exclusive mode during resize.
+ // Not locked at all for non-resizable HashTables.
+ alignas(kCacheLineBytes) SpinSharedMutex<true> resize_shared_mutex_;
+
+ std::size_t kBucketAlignment;
+
+ // Value's offset in a bucket is the first alignof(ValueT) boundary after the
+ // next pointer and hash code.
+ std::size_t kValueOffset;
+
+ // Round bucket size up to a multiple of kBucketAlignment.
+ constexpr std::size_t ComputeBucketSize(const std::size_t fixed_key_size) {
+ return (((kValueOffset + this->total_payload_size_ + fixed_key_size - 1) /
+ kBucketAlignment) +
+ 1) *
+ kBucketAlignment;
+ }
+
+ // Attempt to find an empty bucket to insert 'hash_code' into, starting after
+ // '*bucket' in the chain (or, if '*bucket' is NULL, starting from the slot
+ // array). Returns true and stores SIZE_T_MAX in '*pending_chain_ptr' if an
+ // empty bucket is found. Returns false if 'allow_duplicate_keys' is false
+ // and a hash collision is found (caller should then check whether there is a
+ // genuine key collision or the hash collision is spurious). Returns false
+ // and sets '*bucket' to NULL if there are no more empty buckets in the hash
+ // table. If 'variable_key_allocation_required' is nonzero, this method will
+ // attempt to allocate storage for a variable-length key BEFORE allocating a
+ // bucket, so that no bucket number below 'header_->num_buckets' is ever
+ // deallocated after being allocated.
+ inline bool locateBucketForInsertion(
+ const std::size_t hash_code,
+ const std::size_t variable_key_allocation_required,
+ void **bucket,
+ std::atomic<std::size_t> **pending_chain_ptr,
+ std::size_t *pending_chain_ptr_finish_value);
+
+ // Write a scalar 'key' and its 'hash_code' into the '*bucket', which was
+ // found by locateBucketForInsertion(). Assumes that storage for a
+ // variable-length key copy (if any) was already allocated by a successful
+ // call to allocateVariableLengthKeyStorage().
+ inline void writeScalarKeyToBucket(
+ const TypedValue &key,
+ const std::size_t hash_code,
+ void *bucket);
+
+ // Write a composite 'key' and its 'hash_code' into the '*bucket', which was
+ // found by locateBucketForInsertion(). Assumes that storage for
+ // variable-length key copies (if any) was already allocated by a successful
+ // call to allocateVariableLengthKeyStorage().
+ inline void writeCompositeKeyToBucket(
+ const std::vector<TypedValue> &key,
+ const std::size_t hash_code,
+ void *bucket);
+
+ // Determine whether it is actually necessary to resize this hash table.
+ // Checks that there is at least one unallocated bucket, and that there is
+ // at least 'extra_variable_storage' bytes of variable-length storage free.
+ inline bool isFull(const std::size_t extra_variable_storage) const;
+
+ // Helper object to manage key storage.
+ HashTableKeyManager<false, true> key_manager_;
+
+ // In-memory structure is as follows:
+ // - SeparateChainingHashTable::Header
+ // - Array of slots, interpreted as follows:
+ // - 0 = Points to nothing (empty)
+ // - SIZE_T_MAX = Pending (some thread is starting a chain from this
+ // slot and will overwrite it soon)
+ // - Anything else = The number of the first bucket in the chain for
+ // this slot PLUS ONE (i.e. subtract one to get the actual bucket
+ // number).
+ // - Array of buckets, each of which is:
+ // - atomic size_t "next" pointer, interpreted the same as slots above.
+ // - size_t hash value
+ // - possibly some unused bytes as needed so that ValueT's alignment
+ // requirement is met
+ // - ValueT value slot
+ // - fixed-length key storage (which may include pointers to external
+ // memory or offsets of variable length keys stored within this hash
+ // table)
+ // - possibly some additional unused bytes so that bucket size is a
+ // multiple of both alignof(std::atomic<std::size_t>) and
+ // alignof(ValueT)
+ // - Variable-length key storage region (referenced by offsets stored in
+ // fixed-length keys).
+ Header *header_;
+
+ std::atomic<std::size_t> *slots_;
+ void *buckets_;
+ const std::size_t bucket_size_;
+
+ DISALLOW_COPY_AND_ASSIGN(PackedPayloadSeparateChainingAggregationStateHashTable);
+};
+
+/** @} */
+
+// ----------------------------------------------------------------------------
+// Implementations of template class methods follow.
+
+class HashTableMergerFast {
+ public:
+ /**
+ * @brief Constructor
+ *
+ * @param handle The Aggregation handle being used.
+ * @param destination_hash_table The destination hash table to which other
+ * hash tables will be merged.
+ **/
+ explicit HashTableMergerFast(
+ AggregationStateHashTableBase *destination_hash_table)
+ : destination_hash_table_(
+ static_cast<PackedPayloadSeparateChainingAggregationStateHashTable *>(
+ destination_hash_table)) {}
+
+ /**
+ * @brief The operator for the functor.
+ *
+ * @param group_by_key The group by key being merged.
+ * @param source_state The aggregation state for the given key in the source
+ * aggregation hash table.
+ **/
+ inline void operator()(const std::vector<TypedValue> &group_by_key,
+ const std::uint8_t *source_state) {
+ destination_hash_table_->upsertCompositeKey(group_by_key, source_state);
+ }
+
+ private:
+ PackedPayloadSeparateChainingAggregationStateHashTable *destination_hash_table_;
+
+ DISALLOW_COPY_AND_ASSIGN(HashTableMergerFast);
+};
+
+inline std::size_t PackedPayloadSeparateChainingAggregationStateHashTable
+ ::hashCompositeKey(const std::vector<TypedValue> &key) const {
+ DEBUG_ASSERT(!key.empty());
+ DEBUG_ASSERT(key.size() == key_types_.size());
+ std::size_t hash = key.front().getHash();
+ for (std::vector<TypedValue>::const_iterator key_it = key.begin() + 1;
+ key_it != key.end();
+ ++key_it) {
+ hash = CombineHashes(hash, key_it->getHash());
+ }
+ return hash;
+}
+
+inline bool PackedPayloadSeparateChainingAggregationStateHashTable
+ ::getNextEntryCompositeKey(std::vector<TypedValue> *key,
+ const std::uint8_t **value,
+ std::size_t *entry_num) const {
+ if (*entry_num < header_->buckets_allocated.load(std::memory_order_relaxed)) {
+ const char *bucket =
+ static_cast<const char *>(buckets_) + (*entry_num) * bucket_size_;
+ for (std::vector<const Type *>::size_type key_idx = 0;
+ key_idx < this->key_types_.size();
+ ++key_idx) {
+ key->emplace_back(key_manager_.getKeyComponentTyped(bucket, key_idx));
+ }
+ *value = reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset);
+ ++(*entry_num);
+ return true;
+ } else {
+ return false;
+ }
+}
+
+
+inline bool PackedPayloadSeparateChainingAggregationStateHashTable
+ ::locateBucketForInsertion(const std::size_t hash_code,
+ const std::size_t variable_key_allocation_required,
+ void **bucket,
+ std::atomic<std::size_t> **pending_chain_ptr,
+ std::size_t *pending_chain_ptr_finish_value) {
+ if (*bucket == nullptr) {
+ *pending_chain_ptr = &(slots_[hash_code % header_->num_slots]);
+ } else {
+ *pending_chain_ptr = static_cast<std::atomic<std::size_t> *>(*bucket);
+ }
+ for (;;) {
+ std::size_t existing_chain_ptr = 0;
+ if ((*pending_chain_ptr)
+ ->compare_exchange_strong(existing_chain_ptr,
+ std::numeric_limits<std::size_t>::max(),
+ std::memory_order_acq_rel)) {
+ // Got to the end of the chain. Allocate a new bucket.
+
+ // First, allocate variable-length key storage, if needed (i.e. if this
+ // is an upsert and we didn't allocate up-front).
+ if (!key_manager_.allocateVariableLengthKeyStorage(
+ variable_key_allocation_required)) {
+ // Ran out of variable-length storage.
+ (*pending_chain_ptr)->store(0, std::memory_order_release);
+ *bucket = nullptr;
+ return false;
+ }
+
+ const std::size_t allocated_bucket_num =
+ header_->buckets_allocated.fetch_add(1, std::memory_order_relaxed);
+ if (allocated_bucket_num >= header_->num_buckets) {
+ // Ran out of buckets.
+ header_->buckets_allocated.fetch_sub(1, std::memory_order_relaxed);
+ (*pending_chain_ptr)->store(0, std::memory_order_release);
+ *bucket = nullptr;
+ return false;
+ } else {
+ *bucket =
+ static_cast<char *>(buckets_) + allocated_bucket_num * bucket_size_;
+ *pending_chain_ptr_finish_value = allocated_bucket_num + 1;
+ return true;
+ }
+ }
+ // Spin until the real "next" pointer is available.
+ while (existing_chain_ptr == std::numeric_limits<std::size_t>::max()) {
+ existing_chain_ptr =
+ (*pending_chain_ptr)->load(std::memory_order_acquire);
+ }
+ if (existing_chain_ptr == 0) {
+ // Other thread had to roll back, so try again.
+ continue;
+ }
+ // Chase the next pointer.
+ *bucket =
+ static_cast<char *>(buckets_) + (existing_chain_ptr - 1) * bucket_size_;
+ *pending_chain_ptr = static_cast<std::atomic<std::size_t> *>(*bucket);
+ const std::size_t hash_in_bucket = *reinterpret_cast<const std::size_t *>(
+ static_cast<const char *>(*bucket) +
+ sizeof(std::atomic<std::size_t>));
+ if (hash_in_bucket == hash_code) {
+ return false;
+ }
+ }
+}
+
+inline const std::uint8_t* PackedPayloadSeparateChainingAggregationStateHashTable
+ ::getSingleCompositeKey(const std::vector<TypedValue> &key) const {
+ DEBUG_ASSERT(this->key_types_.size() == key.size());
+
+ const std::size_t hash_code = this->hashCompositeKey(key);
+ std::size_t bucket_ref =
+ slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
+ while (bucket_ref != 0) {
+ DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
+ const char *bucket =
+ static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
+ const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
+ bucket + sizeof(std::atomic<std::size_t>));
+ if ((bucket_hash == hash_code) &&
+ key_manager_.compositeKeyCollisionCheck(key, bucket)) {
+ // Match located.
+ return reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset);
+ }
+ bucket_ref =
+ reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
+ std::memory_order_relaxed);
+ }
+
+ // Reached the end of the chain and didn't find a match.
+ return nullptr;
+}
+
+inline const std::uint8_t* PackedPayloadSeparateChainingAggregationStateHashTable
+ ::getSingleCompositeKey(const std::vector<TypedValue> &key,
+ const int index) const {
+ DEBUG_ASSERT(this->key_types_.size() == key.size());
+
+ const std::size_t hash_code = this->hashCompositeKey(key);
+ std::size_t bucket_ref =
+ slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
+ while (bucket_ref != 0) {
+ DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
+ const char *bucket =
+ static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
+ const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
+ bucket + sizeof(std::atomic<std::size_t>));
+ if ((bucket_hash == hash_code) &&
+ key_manager_.compositeKeyCollisionCheck(key, bucket)) {
+ // Match located.
+ return reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset) +
+ this->payload_offsets_[index];
+ }
+ bucket_ref =
+ reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
+ std::memory_order_relaxed);
+ }
+
+ // Reached the end of the chain and didn't find a match.
+ return nullptr;
+}
+
+inline bool PackedPayloadSeparateChainingAggregationStateHashTable
+ ::upsertCompositeKey(const std::vector<TypedValue> &key,
+ const std::uint8_t *source_state) {
+ const std::size_t variable_size =
+ calculateVariableLengthCompositeKeyCopySize(key);
+ for (;;) {
+ {
+ SpinSharedMutexSharedLock<true> resize_lock(resize_shared_mutex_);
+ std::uint8_t *value =
+ upsertCompositeKeyInternal(key, variable_size);
+ if (value != nullptr) {
+ SpinMutexLock lock(*(reinterpret_cast<SpinMutex *>(value)));
+ for (unsigned int k = 0; k < num_handles_; ++k) {
+ handles_[k]->mergeStates(source_state + payload_offsets_[k],
+ value + payload_offsets_[k]);
+ }
+ return true;
+ }
+ }
+ resize(0, variable_size);
+ }
+}
+
+inline std::uint8_t* PackedPayloadSeparateChainingAggregationStateHashTable
+ ::upsertCompositeKeyInternal(const std::vector<TypedValue> &key,
+ const std::size_t variable_key_size) {
+ if (variable_key_size > 0) {
+ // Don't allocate yet, since the key may already be present. However, we
+ // do check if either the allocated variable storage space OR the free
+ // space is big enough to hold the key (at least one must be true: either
+ // the key is already present and allocated, or we need to be able to
+ // allocate enough space for it).
+ std::size_t allocated_bytes = header_->variable_length_bytes_allocated.load(
+ std::memory_order_relaxed);
+ if ((allocated_bytes < variable_key_size) &&
+ (allocated_bytes + variable_key_size >
+ key_manager_.getVariableLengthKeyStorageSize())) {
+ return nullptr;
+ }
+ }
+
+ const std::size_t hash_code = this->hashCompositeKey(key);
+ void *bucket = nullptr;
+ std::atomic<std::size_t> *pending_chain_ptr;
+ std::size_t pending_chain_ptr_finish_value;
+ for (;;) {
+ if (locateBucketForInsertion(hash_code,
+ variable_key_size,
+ &bucket,
+ &pending_chain_ptr,
+ &pending_chain_ptr_finish_value)) {
+ // Found an empty bucket.
+ break;
+ } else if (bucket == nullptr) {
+ // Ran out of buckets or variable-key space.
+ return nullptr;
+ } else if (key_manager_.compositeKeyCollisionCheck(key, bucket)) {
+ // Found an already-existing entry for this key.
+ return reinterpret_cast<std::uint8_t *>(static_cast<char *>(bucket) +
+ kValueOffset);
+ }
+ }
+
+ // We are now writing to an empty bucket.
+ // Write the key and hash.
+ writeCompositeKeyToBucket(key, hash_code, bucket);
+
+ std::uint8_t *value = static_cast<unsigned char *>(bucket) + kValueOffset;
+ std::memcpy(value, init_payload_, this->total_payload_size_);
+
+ // Update the previous chaing pointer to point to the new bucket.
+ pending_chain_ptr->store(pending_chain_ptr_finish_value,
+ std::memory_order_release);
+
+ // Return the value.
+ return value;
+}
+
+template <bool has_aux_accessor>
+inline bool PackedPayloadSeparateChainingAggregationStateHashTable
+ ::upsertValueAccessorCompositeKeyInternal(
+ const std::vector<std::vector<attribute_id>> &argument_ids,
+ const std::vector<attribute_id> &key_attr_ids,
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor) {
+ std::size_t variable_size;
+ std::vector<TypedValue> key_vector;
+ key_vector.resize(key_attr_ids.size());
+
+ // TODO(jianqiao): determine this bool value
+ const bool check_for_null_keys = true;
+
+ return InvokeOnAnyValueAccessor(
+ accessor,
+ [&](auto *accessor) -> bool { // NOLINT(build/c++11)
+ bool continuing = true;
+ while (continuing) {
+ {
+ continuing = false;
+ SpinSharedMutexSharedLock<true> lock(resize_shared_mutex_);
+ while (accessor->next()) {
+ if (has_aux_accessor) {
+ aux_accessor->next();
+ }
+ // TODO(jianqiao): templatize to involve aux_accessor
+ if (this->GetCompositeKeyFromValueAccessor(*accessor,
+ key_attr_ids,
+ check_for_null_keys,
+ &key_vector)) {
+ continue;
+ }
+ variable_size = this->calculateVariableLengthCompositeKeyCopySize(key_vector);
+ std::uint8_t *value = this->upsertCompositeKeyInternal(
+ key_vector, variable_size);
+ if (value == nullptr) {
+ continuing = true;
+ break;
+ } else {
+ SpinMutexLock lock(*(reinterpret_cast<SpinMutex *>(value)));
+ for (unsigned int k = 0; k < num_handles_; ++k) {
+ const auto &ids = argument_ids[k];
+ if (ids.empty()) {
+ handles_[k]->updateStateNullary(
+ value + payload_offsets_[k]);
+ } else {
+ const attribute_id argument_id = ids.front();
+ if (has_aux_accessor && argument_id < 0) {
+ DCHECK_NE(argument_id, kInvalidAttributeID);
+ handles_[k]->updateStateUnary(aux_accessor->getTypedValue(-(argument_id+2)),
+ value + payload_offsets_[k]);
+ } else {
+ handles_[k]->updateStateUnary(accessor->getTypedValue(argument_id),
+ value + payload_offsets_[k]);
+ }
+ }
+ }
+ }
+ }
+ }
+ if (continuing) {
+ this->resize(0, variable_size);
+ accessor->previous();
+ if (has_aux_accessor) {
+ aux_accessor->previous();
+ }
+ }
+ }
+ return true;
+ });
+}
+
+inline void PackedPayloadSeparateChainingAggregationStateHashTable
+ ::writeScalarKeyToBucket(const TypedValue &key,
+ const std::size_t hash_code,
+ void *bucket) {
+ *reinterpret_cast<std::size_t *>(static_cast<char *>(bucket) +
+ sizeof(std::atomic<std::size_t>)) =
+ hash_code;
+ key_manager_.writeKeyComponentToBucket(key, 0, bucket, nullptr);
+}
+
+inline void PackedPayloadSeparateChainingAggregationStateHashTable
+ ::writeCompositeKeyToBucket(const std::vector<TypedValue> &key,
+ const std::size_t hash_code,
+ void *bucket) {
+ DEBUG_ASSERT(key.size() == this->key_types_.size());
+ *reinterpret_cast<std::size_t *>(static_cast<char *>(bucket) +
+ sizeof(std::atomic<std::size_t>)) =
+ hash_code;
+ for (std::size_t idx = 0; idx < this->key_types_.size(); ++idx) {
+ key_manager_.writeKeyComponentToBucket(key[idx], idx, bucket, nullptr);
+ }
+}
+
+inline bool PackedPayloadSeparateChainingAggregationStateHashTable::isFull(
+ const std::size_t extra_variable_storage) const {
+ if (header_->buckets_allocated.load(std::memory_order_relaxed) >=
+ header_->num_buckets) {
+ // All buckets are allocated.
+ return true;
+ }
+
+ if (extra_variable_storage > 0) {
+ if (extra_variable_storage +
+ header_->variable_length_bytes_allocated.load(
+ std::memory_order_relaxed) >
+ key_manager_.getVariableLengthKeyStorageSize()) {
+ // Not enough variable-length key storage space.
+ return true;
+ }
+ }
+
+ return false;
+}
+
+template <typename FunctorT>
+inline std::size_t PackedPayloadSeparateChainingAggregationStateHashTable
+ ::forEach(FunctorT *functor) const {
+ std::size_t entries_visited = 0;
+ std::size_t entry_num = 0;
+ std::vector<TypedValue> key;
+ const std::uint8_t *value_ptr;
+ while (getNextEntryCompositeKey(&key, &value_ptr, &entry_num)) {
+ ++entries_visited;
+ (*functor)(key, value_ptr);
+ key.clear();
+ }
+ return entries_visited;
+}
+
+template <typename FunctorT>
+inline std::size_t PackedPayloadSeparateChainingAggregationStateHashTable
+ ::forEach(FunctorT *functor, const int index) const {
+ std::size_t entries_visited = 0;
+ std::size_t entry_num = 0;
+ std::vector<TypedValue> key;
+ const std::uint8_t *value_ptr;
+ while (getNextEntryCompositeKey(&key, &value_ptr, &entry_num)) {
+ ++entries_visited;
+ (*functor)(key, value_ptr + payload_offsets_[index]);
+ key.clear();
+ }
+ return entries_visited;
+}
+
+
+} // namespace quickstep
+
+#endif // QUICKSTEP_STORAGE_PACKED_PAYLOAD_AGGREGATION_STATE_HASH_TABLE_HPP_
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/PartitionedHashTablePool.hpp
----------------------------------------------------------------------
diff --git a/storage/PartitionedHashTablePool.hpp b/storage/PartitionedHashTablePool.hpp
index 95d1810..e9ca022 100644
--- a/storage/PartitionedHashTablePool.hpp
+++ b/storage/PartitionedHashTablePool.hpp
@@ -28,8 +28,7 @@
#include "expressions/aggregation/AggregationHandle.hpp"
#include "storage/HashTableBase.hpp"
-#include "storage/FastHashTable.hpp"
-#include "storage/FastHashTableFactory.hpp"
+#include "storage/HashTableFactory.hpp"
#include "utility/Macros.hpp"
#include "utility/StringUtil.hpp"
@@ -54,33 +53,6 @@ class PartitionedHashTablePool {
/**
* @brief Constructor.
*
- * @param estimated_num_entries The maximum number of entries in a hash table.
- * @param num_partitions The number of partitions (i.e. number of HashTables)
- * @param hash_table_impl_type The type of hash table implementation.
- * @param group_by_types A vector of pointer of types which form the group by
- * key.
- * @param agg_handle The aggregation handle.
- * @param storage_manager A pointer to the storage manager.
- **/
- PartitionedHashTablePool(const std::size_t estimated_num_entries,
- const std::size_t num_partitions,
- const HashTableImplType hash_table_impl_type,
- const std::vector<const Type *> &group_by_types,
- AggregationHandle *agg_handle,
- StorageManager *storage_manager)
- : estimated_num_entries_(
- setHashTableSize(estimated_num_entries, num_partitions)),
- num_partitions_(num_partitions),
- hash_table_impl_type_(hash_table_impl_type),
- group_by_types_(group_by_types),
- agg_handle_(DCHECK_NOTNULL(agg_handle)),
- storage_manager_(DCHECK_NOTNULL(storage_manager)) {
- initializeAllHashTables();
- }
-
- /**
- * @brief Constructor.
- *
* @note This constructor is relevant for the HashTable specialized for
* aggregation.
*
@@ -89,8 +61,6 @@ class PartitionedHashTablePool {
* @param hash_table_impl_type The type of hash table implementation.
* @param group_by_types A vector of pointer of types which form the group by
* key.
- * @param payload_sizes The sizes of the payload elements (i.e.
- * AggregationStates).
* @param handles The aggregation handles.
* @param storage_manager A pointer to the storage manager.
**/
@@ -98,7 +68,6 @@ class PartitionedHashTablePool {
const std::size_t num_partitions,
const HashTableImplType hash_table_impl_type,
const std::vector<const Type *> &group_by_types,
- const std::vector<std::size_t> &payload_sizes,
const std::vector<AggregationHandle *> &handles,
StorageManager *storage_manager)
: estimated_num_entries_(
@@ -106,7 +75,6 @@ class PartitionedHashTablePool {
num_partitions_(num_partitions),
hash_table_impl_type_(hash_table_impl_type),
group_by_types_(group_by_types),
- payload_sizes_(payload_sizes),
handles_(handles),
storage_manager_(DCHECK_NOTNULL(storage_manager)) {
initializeAllHashTables();
@@ -150,25 +118,17 @@ class PartitionedHashTablePool {
private:
void initializeAllHashTables() {
for (std::size_t part_num = 0; part_num < num_partitions_; ++part_num) {
- AggregationStateHashTableBase *part_hash_table = createNewHashTableFast();
+ AggregationStateHashTableBase *part_hash_table = createNewHashTable();
hash_tables_.push_back(
std::unique_ptr<AggregationStateHashTableBase>(part_hash_table));
}
}
AggregationStateHashTableBase* createNewHashTable() {
- return agg_handle_->createGroupByHashTable(hash_table_impl_type_,
- group_by_types_,
- estimated_num_entries_,
- storage_manager_);
- }
-
- AggregationStateHashTableBase* createNewHashTableFast() {
- return AggregationStateFastHashTableFactory::CreateResizable(
+ return AggregationStateHashTableFactory::CreateResizable(
hash_table_impl_type_,
group_by_types_,
estimated_num_entries_,
- payload_sizes_,
handles_,
storage_manager_);
}
@@ -189,10 +149,6 @@ class PartitionedHashTablePool {
const HashTableImplType hash_table_impl_type_;
const std::vector<const Type *> group_by_types_;
-
- std::vector<std::size_t> payload_sizes_;
-
- AggregationHandle *agg_handle_;
const std::vector<AggregationHandle *> handles_;
StorageManager *storage_manager_;
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/StorageBlock.cpp
----------------------------------------------------------------------
diff --git a/storage/StorageBlock.cpp b/storage/StorageBlock.cpp
index ea74ee6..01a6681 100644
--- a/storage/StorageBlock.cpp
+++ b/storage/StorageBlock.cpp
@@ -373,166 +373,6 @@ void StorageBlock::selectSimple(const std::vector<attribute_id> &selection,
accessor.get());
}
-AggregationState* StorageBlock::aggregate(
- const AggregationHandle &handle,
- const std::vector<std::unique_ptr<const Scalar>> &arguments,
- const std::vector<attribute_id> *arguments_as_attributes,
- const TupleIdSequence *filter) const {
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- // If all the arguments to this aggregate are plain relation attributes,
- // aggregate directly on a ValueAccessor from this block to avoid a copy.
- if ((arguments_as_attributes != nullptr) && (!arguments_as_attributes->empty())) {
- DCHECK_EQ(arguments.size(), arguments_as_attributes->size())
- << "Mismatch between number of arguments and number of attribute_ids";
- return aggregateHelperValueAccessor(handle, *arguments_as_attributes, filter);
- }
- // TODO(shoban): We may want to optimize for ScalarLiteral here.
-#endif
-
- // Call aggregateHelperColumnVector() to materialize each argument as a
- // ColumnVector, then aggregate over those.
- return aggregateHelperColumnVector(handle, arguments, filter);
-}
-
-void StorageBlock::aggregateGroupBy(
- const std::vector<std::vector<std::unique_ptr<const Scalar>>> &arguments,
- const std::vector<std::unique_ptr<const Scalar>> &group_by,
- const TupleIdSequence *filter,
- AggregationStateHashTableBase *hash_table,
- std::vector<std::unique_ptr<ColumnVector>> *reuse_group_by_vectors) const {
- DCHECK_GT(group_by.size(), 0u)
- << "Called aggregateGroupBy() with zero GROUP BY expressions";
-
- SubBlocksReference sub_blocks_ref(*tuple_store_,
- indices_,
- indices_consistent_);
-
- // IDs of 'arguments' as attributes in the ValueAccessor we create below.
- std::vector<attribute_id> argument_ids;
-
- // IDs of GROUP BY key element(s) in the ValueAccessor we create below.
- std::vector<attribute_id> key_ids;
-
- // An intermediate ValueAccessor that stores the materialized 'arguments' for
- // this aggregate, as well as the GROUP BY expression values.
- ColumnVectorsValueAccessor temp_result;
- {
- std::unique_ptr<ValueAccessor> accessor(tuple_store_->createValueAccessor(filter));
- attribute_id attr_id = 0;
-
- // First, put GROUP BY keys into 'temp_result'.
- if (reuse_group_by_vectors->empty()) {
- // Compute GROUP BY values from group_by Scalars, and store them in
- // reuse_group_by_vectors for reuse by other aggregates on this same
- // block.
- reuse_group_by_vectors->reserve(group_by.size());
- for (const std::unique_ptr<const Scalar> &group_by_element : group_by) {
- reuse_group_by_vectors->emplace_back(
- group_by_element->getAllValues(accessor.get(), &sub_blocks_ref));
- temp_result.addColumn(reuse_group_by_vectors->back().get(), false);
- key_ids.push_back(attr_id++);
- }
- } else {
- // Reuse precomputed GROUP BY values from reuse_group_by_vectors.
- DCHECK_EQ(group_by.size(), reuse_group_by_vectors->size())
- << "Wrong number of reuse_group_by_vectors";
- for (const std::unique_ptr<ColumnVector> &reuse_cv : *reuse_group_by_vectors) {
- temp_result.addColumn(reuse_cv.get(), false);
- key_ids.push_back(attr_id++);
- }
- }
-
- // Compute argument vectors and add them to 'temp_result'.
- for (const std::vector<std::unique_ptr<const Scalar>> &argument : arguments) {
- for (const std::unique_ptr<const Scalar> &args : argument) {
- temp_result.addColumn(args->getAllValues(accessor.get(), &sub_blocks_ref));
- argument_ids.push_back(attr_id++);
- }
- if (argument.empty()) {
- argument_ids.push_back(kInvalidAttributeID);
- }
- }
- }
-
- hash_table->upsertValueAccessorCompositeKeyFast(argument_ids,
- &temp_result,
- key_ids,
- true);
-}
-
-
-void StorageBlock::aggregateDistinct(
- const AggregationHandle &handle,
- const std::vector<std::unique_ptr<const Scalar>> &arguments,
- const std::vector<attribute_id> *arguments_as_attributes,
- const std::vector<std::unique_ptr<const Scalar>> &group_by,
- const TupleIdSequence *filter,
- AggregationStateHashTableBase *distinctify_hash_table,
- std::vector<std::unique_ptr<ColumnVector>> *reuse_group_by_vectors) const {
- DCHECK_GT(arguments.size(), 0u)
- << "Called aggregateDistinct() with zero argument expressions";
- DCHECK((group_by.size() == 0 || reuse_group_by_vectors != nullptr));
-
- std::vector<attribute_id> key_ids;
-
- // An intermediate ValueAccessor that stores the materialized 'arguments' for
- // this aggregate, as well as the GROUP BY expression values.
- ColumnVectorsValueAccessor temp_result;
- {
- std::unique_ptr<ValueAccessor> accessor(tuple_store_->createValueAccessor(filter));
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- // If all the arguments to this aggregate are plain relation attributes,
- // aggregate directly on a ValueAccessor from this block to avoid a copy.
- if ((arguments_as_attributes != nullptr) && (!arguments_as_attributes->empty())) {
- DCHECK_EQ(arguments.size(), arguments_as_attributes->size())
- << "Mismatch between number of arguments and number of attribute_ids";
- DCHECK_EQ(group_by.size(), 0u);
- handle.insertValueAccessorIntoDistinctifyHashTable(
- accessor.get(), *arguments_as_attributes, distinctify_hash_table);
- return;
- }
-#endif
-
- SubBlocksReference sub_blocks_ref(*tuple_store_,
- indices_,
- indices_consistent_);
- attribute_id attr_id = 0;
-
- if (!group_by.empty()) {
- // Put GROUP BY keys into 'temp_result'.
- if (reuse_group_by_vectors->empty()) {
- // Compute GROUP BY values from group_by Scalars, and store them in
- // reuse_group_by_vectors for reuse by other aggregates on this same
- // block.
- reuse_group_by_vectors->reserve(group_by.size());
- for (const std::unique_ptr<const Scalar> &group_by_element : group_by) {
- reuse_group_by_vectors->emplace_back(
- group_by_element->getAllValues(accessor.get(), &sub_blocks_ref));
- temp_result.addColumn(reuse_group_by_vectors->back().get(), false);
- key_ids.push_back(attr_id++);
- }
- } else {
- // Reuse precomputed GROUP BY values from reuse_group_by_vectors.
- DCHECK_EQ(group_by.size(), reuse_group_by_vectors->size())
- << "Wrong number of reuse_group_by_vectors";
- for (const std::unique_ptr<ColumnVector> &reuse_cv : *reuse_group_by_vectors) {
- temp_result.addColumn(reuse_cv.get(), false);
- key_ids.push_back(attr_id++);
- }
- }
- }
- // Compute argument vectors and add them to 'temp_result'.
- for (const std::unique_ptr<const Scalar> &argument : arguments) {
- temp_result.addColumn(argument->getAllValues(accessor.get(), &sub_blocks_ref));
- key_ids.push_back(attr_id++);
- }
- }
-
- handle.insertValueAccessorIntoDistinctifyHashTable(
- &temp_result, key_ids, distinctify_hash_table);
-}
-
// TODO(chasseur): Vectorization for updates.
StorageBlock::UpdateResult StorageBlock::update(
const unordered_map<attribute_id, unique_ptr<const Scalar>> &assignments,
@@ -1245,61 +1085,6 @@ std::unordered_map<attribute_id, TypedValue>* StorageBlock::generateUpdatedValue
return update_map;
}
-AggregationState* StorageBlock::aggregateHelperColumnVector(
- const AggregationHandle &handle,
- const std::vector<std::unique_ptr<const Scalar>> &arguments,
- const TupleIdSequence *matches) const {
- if (arguments.empty()) {
- // Special case. This is a nullary aggregate (i.e. COUNT(*)).
- return handle.accumulateNullary(matches == nullptr ? tuple_store_->numTuples()
- : matches->size());
- } else {
- // Set up a ValueAccessor that will be used when materializing argument
- // values below (possibly filtered based on the '*matches' to a filter
- // predicate).
- std::unique_ptr<ValueAccessor> accessor;
- if (matches == nullptr) {
- accessor.reset(tuple_store_->createValueAccessor());
- } else {
- accessor.reset(tuple_store_->createValueAccessor(matches));
- }
-
- SubBlocksReference sub_blocks_ref(*tuple_store_,
- indices_,
- indices_consistent_);
-
- // Materialize each argument's values for this block as a ColumnVector.
- std::vector<std::unique_ptr<ColumnVector>> column_vectors;
- for (const std::unique_ptr<const Scalar> &argument : arguments) {
- column_vectors.emplace_back(argument->getAllValues(accessor.get(), &sub_blocks_ref));
- }
-
- // Have the AggregationHandle actually do the aggregation.
- return handle.accumulateColumnVectors(column_vectors);
- }
-}
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-AggregationState* StorageBlock::aggregateHelperValueAccessor(
- const AggregationHandle &handle,
- const std::vector<attribute_id> &argument_ids,
- const TupleIdSequence *matches) const {
- // Set up a ValueAccessor to aggregate over (possibly filtered based on the
- // '*matches' to a filter predicate).
- std::unique_ptr<ValueAccessor> accessor;
- if (matches == nullptr) {
- accessor.reset(tuple_store_->createValueAccessor());
- } else {
- accessor.reset(tuple_store_->createValueAccessor(matches));
- }
-
- // Have the AggregationHandle actually do the aggregation.
- return handle.accumulateValueAccessor(
- accessor.get(),
- argument_ids);
-}
-#endif // QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-
void StorageBlock::updateHeader() {
DEBUG_ASSERT(*static_cast<const int*>(block_memory_) == block_header_.ByteSize());
@@ -1329,59 +1114,4 @@ const std::size_t StorageBlock::getNumTuples() const {
return tuple_store_->numTuples();
}
-void StorageBlock::aggregateGroupByPartitioned(
- const std::vector<std::vector<std::unique_ptr<const Scalar>>> &arguments,
- const std::vector<std::unique_ptr<const Scalar>> &group_by,
- const TupleIdSequence *filter,
- const std::size_t num_partitions,
- ColumnVectorsValueAccessor *temp_result,
- std::vector<attribute_id> *argument_ids,
- std::vector<attribute_id> *key_ids,
- std::vector<std::unique_ptr<ColumnVector>> *reuse_group_by_vectors) const {
- DCHECK(!group_by.empty())
- << "Called aggregateGroupByPartitioned() with zero GROUP BY expressions";
-
- SubBlocksReference sub_blocks_ref(*tuple_store_,
- indices_,
- indices_consistent_);
-
- std::unique_ptr<ValueAccessor> accessor(
- tuple_store_->createValueAccessor(filter));
-
- attribute_id attr_id = 0;
-
- // First, put GROUP BY keys into 'temp_result'.
- if (reuse_group_by_vectors->empty()) {
- // Compute GROUP BY values from group_by Scalars, and store them in
- // reuse_group_by_vectors for reuse by other aggregates on this same
- // block.
- reuse_group_by_vectors->reserve(group_by.size());
- for (const std::unique_ptr<const Scalar> &group_by_element : group_by) {
- reuse_group_by_vectors->emplace_back(
- group_by_element->getAllValues(accessor.get(), &sub_blocks_ref));
- temp_result->addColumn(reuse_group_by_vectors->back().get(), false);
- key_ids->push_back(attr_id++);
- }
- } else {
- // Reuse precomputed GROUP BY values from reuse_group_by_vectors.
- DCHECK_EQ(group_by.size(), reuse_group_by_vectors->size())
- << "Wrong number of reuse_group_by_vectors";
- for (const std::unique_ptr<ColumnVector> &reuse_cv : *reuse_group_by_vectors) {
- temp_result->addColumn(reuse_cv.get(), false);
- key_ids->push_back(attr_id++);
- }
- }
-
- // Compute argument vectors and add them to 'temp_result'.
- for (const std::vector<std::unique_ptr<const Scalar>> &argument : arguments) {
- for (const std::unique_ptr<const Scalar> &args : argument) {
- temp_result->addColumn(args->getAllValues(accessor.get(), &sub_blocks_ref));
- argument_ids->push_back(attr_id++);
- }
- if (argument.empty()) {
- argument_ids->push_back(kInvalidAttributeID);
- }
- }
-}
-
} // namespace quickstep
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/StorageBlock.hpp
----------------------------------------------------------------------
diff --git a/storage/StorageBlock.hpp b/storage/StorageBlock.hpp
index 56b3bdc..31c27e2 100644
--- a/storage/StorageBlock.hpp
+++ b/storage/StorageBlock.hpp
@@ -387,156 +387,6 @@ class StorageBlock : public StorageBlockBase {
InsertDestinationInterface *destination) const;
/**
- * @brief Perform non GROUP BY aggregation on the tuples in the this storage
- * block, returning the aggregated result (for this block) in an
- * AggregationState.
- *
- * @param handle Aggregation handle that will be used to compute aggregate.
- * @param arguments The arguments of the aggregate function as expressions.
- * @param arguments_as_attributes If non-NULL, indicates a valid attribute_id
- * for each of the elements in arguments, and is used to elide a copy.
- * Has no effect if NULL, or if VECTOR_COPY_ELISION_LEVEL is NONE.
- * @param filter If non-NULL, then only tuple IDs which are set in the
- * filter will be checked (all others will be assumed to be false).
- *
- * @return Aggregated state for this block in the form of an
- * AggregationState. AggregationHandle::mergeStates() can be called
- * to merge with states from other blocks, and
- * AggregationHandle::finalize() can be used to generate a final
- * result.
- **/
- AggregationState* aggregate(
- const AggregationHandle &handle,
- const std::vector<std::unique_ptr<const Scalar>> &arguments,
- const std::vector<attribute_id> *arguments_as_attributes,
- const TupleIdSequence *filter) const;
-
- /**
- * @brief Perform GROUP BY aggregation on the tuples in the this storage
- * block.
- *
- * @param arguments The arguments to the aggregation function as Scalars.
- * @param group_by The list of GROUP BY attributes/expressions. The tuples in
- * this storage block are grouped by these attributes before
- * aggregation.
- * @param filter If non-NULL, then only tuple IDs which are set in the
- * filter will be checked (all others will be assumed to be false).
- * @param hash_table Hash table to store aggregation state mapped based on
- * GROUP BY value list (defined by \c group_by).
- * @param reuse_group_by_vectors This parameter is used to store and reuse
- * GROUP BY attribute vectors pre-computed in an earlier invocation of
- * aggregateGroupBy(). \c reuse_group_by_vectors is never \c nullptr
- * for ease of use. Current invocation of aggregateGroupBy() will reuse
- * ColumnVectors if non-empty, otherwise computes ColumnVectors based
- * on \c group_by and stores them in \c reuse_group_by_vectors.
- *
- * For sample usage of aggregateGroupBy, see this relevant pseudo-C++ code:
- * \code
- * std::vector<std::unique_ptr<ColumnVector>> group_by_vectors;
- * for each aggregate {
- * block.aggregateGroupBy(..., &group_by_vectors);
- * }
- * \endcode
- **/
- /*
- * TODO(shoban): Currently, we use ColumnVectorsValueAccessor to compute
- * temporary result for Scalars of aggregation attributes and GROUP BY
- * attributes. We will have to support specifying aggregation and GROUP BY
- * attributes as std::vector<attribute_id> (like in selectSimple()) for fast
- * path when there are no expressions specified in the query.
- */
- void aggregateGroupBy(
- const std::vector<std::vector<std::unique_ptr<const Scalar>>> &arguments,
- const std::vector<std::unique_ptr<const Scalar>> &group_by,
- const TupleIdSequence *filter,
- AggregationStateHashTableBase *hash_table,
- std::vector<std::unique_ptr<ColumnVector>> *reuse_group_by_vectors) const;
-
-
- /**
- * @brief Perform the GROUP BY aggregation for the case when aggregation is
- * partitioned.
- *
- * TODO(harshad) - Refactor this class to use only one function
- * aggregateGroupBy.
- * @note The difference between this method and the aggregateGroupBy method
- * is that in this method, the tuples are routed to different HashTables
- * based on the partition to which they belong to. The partition is
- * determined by the GROUP BY attributes. Right now hash based
- * partitioning is performed.
- *
- * @note This function only creates the ColumnVectorsValueAccessor needed for
- * the insertion in the hash table. The actual insertion in respective
- * hash tables should be handled by the caller. See
- * AggregationOperationState::aggregateHashTable() for one such
- * implementation.
- *
- * @param arguments The arguments to the aggregation function as Scalars.
- * @param group_by The list of GROUP BY attributes/expressions. The tuples in
- * this storage block are grouped by these attributes before
- * aggregation.
- * @param filter If non-NULL, then only tuple IDs which are set in the
- * filter will be checked (all others will be assumed to be false).
- * @param num_partitions The number of partitions used for the aggregation.
- * @param temp_result The ColumnVectorsValueAccessor used for collecting
- * the attribute values from this StorageBlock.
- * @param arguments_ids The attribute IDs used for the aggregation, which
- * come from the arguments vector. If arguments is empty, this vector
- * is filled with invalid attribute IDs.
- * @param key_ids The attribute IDs of the group by attributes.
- * @param reuse_group_by_vectors This parameter is used to store and reuse
- * GROUP BY attribute vectors pre-computed in an earlier invocation of
- * aggregateGroupBy(). \c reuse_group_by_vectors is never \c nullptr
- * for ease of use. Current invocation of aggregateGroupBy() will reuse
- * ColumnVectors if non-empty, otherwise computes ColumnVectors based
- * on \c group_by and stores them in \c reuse_group_by_vectors.
- **/
- void aggregateGroupByPartitioned(
- const std::vector<std::vector<std::unique_ptr<const Scalar>>> &arguments,
- const std::vector<std::unique_ptr<const Scalar>> &group_by,
- const TupleIdSequence *filter,
- const std::size_t num_partitions,
- ColumnVectorsValueAccessor *temp_result,
- std::vector<attribute_id> *argument_ids,
- std::vector<attribute_id> *key_ids,
- std::vector<std::unique_ptr<ColumnVector>> *reuse_group_by_vectors) const;
-
- /**
- * @brief Inserts the GROUP BY expressions and aggregation arguments together
- * as keys into the distinctify hash table.
- *
- * This is the first step for DISTINCT aggregation. It populates the distinctify
- * hash table so that arguments are distinctified within each GROUP BY group.
- * Later, a second-round aggregation on the distinctify hash table will be
- * performed to actually compute the aggregated result for each GROUP BY group.
- *
- * @param handle Aggregation handle to compute aggregates with.
- * @param arguments The arguments to the aggregation function as Scalars.
- * @param arguments_as_attributes If non-NULL, indicates a valid attribute_id
- * for each of the elements in arguments, and is used to elide a copy.
- * Has no effect if NULL, or if VECTOR_COPY_ELISION_LEVEL is NONE.
- * @param group_by The list of GROUP BY attributes/expressions.
- * @param filter If non-NULL, then only tuple IDs which are set in the
- * filter will be checked (all others will be assumed to be false).
- * @param distinctify_hash_table Hash table to store the arguments and GROUP
- * BY expressions together as hash table key and a bool constant \c true
- * as hash table value. (So the hash table actually serves as a hash set.)
- * @param reuse_group_by_vectors This parameter is used to store and reuse
- * GROUP BY attribute vectors pre-computed in an earlier invocation of
- * aggregateGroupBy(). \c reuse_group_by_vectors is never \c nullptr
- * for ease of use. Current invocation of aggregateGroupBy() will reuse
- * ColumnVectors if non-empty, otherwise computes ColumnVectors based
- * on \c group_by and stores them in \c reuse_group_by_vectors.
- */
- void aggregateDistinct(const AggregationHandle &handle,
- const std::vector<std::unique_ptr<const Scalar>> &arguments,
- const std::vector<attribute_id> *arguments_as_attributes,
- const std::vector<std::unique_ptr<const Scalar>> &group_by,
- const TupleIdSequence *filter,
- AggregationStateHashTableBase *distinctify_hash_table,
- std::vector<std::unique_ptr<ColumnVector>> *reuse_group_by_vectors) const;
-
- /**
* @brief Perform an UPDATE query over the tuples in this StorageBlock.
* @warning In some edge cases, calling this method may cause IndexSubBlocks
* in this block to become inconsistent (the TupleStorageSubBlock
[4/5] incubator-quickstep git commit: Cleanup FastHashTable + Avoid
unnecessary copying of data before aggregation.
Posted by ji...@apache.org.
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleSum.hpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleSum.hpp b/expressions/aggregation/AggregationHandleSum.hpp
index f0d23e1..60d8416 100644
--- a/expressions/aggregation/AggregationHandleSum.hpp
+++ b/expressions/aggregation/AggregationHandleSum.hpp
@@ -28,7 +28,6 @@
#include "catalog/CatalogTypedefs.hpp"
#include "expressions/aggregation/AggregationConcreteHandle.hpp"
#include "expressions/aggregation/AggregationHandle.hpp"
-#include "storage/FastHashTable.hpp"
#include "storage/HashTableBase.hpp"
#include "threading/SpinMutex.hpp"
#include "types/Type.hpp"
@@ -41,6 +40,7 @@
namespace quickstep {
class ColumnVector;
+class ColumnVectorsValueAccessor;
class StorageManager;
class ValueAccessor;
@@ -105,12 +105,6 @@ class AggregationHandleSum : public AggregationConcreteHandle {
return new AggregationStateSum(blank_state_);
}
- AggregationStateHashTableBase* createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const override;
-
inline void iterateUnaryInl(AggregationStateSum *state,
const TypedValue &value) const {
DCHECK(value.isPlausibleInstanceOf(argument_type_.getSignature()));
@@ -121,28 +115,19 @@ class AggregationHandleSum : public AggregationConcreteHandle {
state->null_ = false;
}
- inline void iterateUnaryInlFast(const TypedValue &value,
- std::uint8_t *byte_ptr) const {
- DCHECK(value.isPlausibleInstanceOf(argument_type_.getSignature()));
- if (value.isNull()) return;
- TypedValue *sum_ptr =
- reinterpret_cast<TypedValue *>(byte_ptr + blank_state_.sum_offset_);
- bool *null_ptr =
- reinterpret_cast<bool *>(byte_ptr + blank_state_.null_offset_);
- *sum_ptr = fast_operator_->applyToTypedValues(*sum_ptr, value);
- *null_ptr = false;
- }
+ AggregationState* accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const override;
- inline void updateStateUnary(const TypedValue &argument,
- std::uint8_t *byte_ptr) const override {
- if (!block_update_) {
- iterateUnaryInlFast(argument, byte_ptr);
- }
- }
+ void mergeStates(const AggregationState &source,
+ AggregationState *destination) const override;
- void blockUpdate() override { block_update_ = true; }
+ TypedValue finalize(const AggregationState &state) const override;
- void allowUpdate() override { block_update_ = false; }
+ std::size_t getPayloadSize() const override {
+ return blank_state_.getPayloadSize();
+ }
void initPayload(std::uint8_t *byte_ptr) const override {
TypedValue *sum_ptr =
@@ -161,41 +146,23 @@ class AggregationHandleSum : public AggregationConcreteHandle {
}
}
- AggregationState* accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors)
- const override;
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- AggregationState* accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_id) const override;
-#endif
-
- void aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const override;
-
- void mergeStates(const AggregationState &source,
- AggregationState *destination) const override;
-
- void mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const override;
-
- TypedValue finalize(const AggregationState &state) const override;
-
- inline TypedValue finalizeHashTableEntry(
- const AggregationState &state) const {
- return static_cast<const AggregationStateSum &>(state).sum_;
+ inline void updateStateUnary(const TypedValue &argument,
+ std::uint8_t *byte_ptr) const override {
+ DCHECK(argument.isPlausibleInstanceOf(argument_type_.getSignature()));
+ if (argument.isNull()) return;
+ TypedValue *sum_ptr =
+ reinterpret_cast<TypedValue *>(byte_ptr + blank_state_.sum_offset_);
+ bool *null_ptr =
+ reinterpret_cast<bool *>(byte_ptr + blank_state_.null_offset_);
+ *sum_ptr = fast_operator_->applyToTypedValues(*sum_ptr, argument);
+ *null_ptr = false;
}
- inline TypedValue finalizeHashTableEntryFast(
- const std::uint8_t *byte_ptr) const {
- std::uint8_t *value_ptr = const_cast<std::uint8_t *>(byte_ptr);
- TypedValue *sum_ptr =
- reinterpret_cast<TypedValue *>(value_ptr + blank_state_.sum_offset_);
- return *sum_ptr;
+ void mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const override;
+
+ inline TypedValue finalizeHashTableEntry(const std::uint8_t *byte_ptr) const {
+ return *reinterpret_cast<const TypedValue *>(byte_ptr + blank_state_.sum_offset_);
}
ColumnVector* finalizeHashTable(
@@ -203,29 +170,6 @@ class AggregationHandleSum : public AggregationConcreteHandle {
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const override;
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForSingle()
- * for SUM aggregation.
- */
- AggregationState* aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table)
- const override;
-
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForGroupBy()
- * for SUM aggregation.
- */
- void aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const override;
-
- std::size_t getPayloadSize() const override {
- return blank_state_.getPayloadSize();
- }
-
private:
friend class AggregateFunctionSum;
@@ -242,8 +186,6 @@ class AggregationHandleSum : public AggregationConcreteHandle {
std::unique_ptr<UncheckedBinaryOperator> fast_operator_;
std::unique_ptr<UncheckedBinaryOperator> merge_operator_;
- bool block_update_;
-
DISALLOW_COPY_AND_ASSIGN(AggregationHandleSum);
};
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/CMakeLists.txt
----------------------------------------------------------------------
diff --git a/expressions/aggregation/CMakeLists.txt b/expressions/aggregation/CMakeLists.txt
index e9503f7..a31ee36 100644
--- a/expressions/aggregation/CMakeLists.txt
+++ b/expressions/aggregation/CMakeLists.txt
@@ -146,8 +146,6 @@ target_link_libraries(quickstep_expressions_aggregation_AggregationConcreteHandl
glog
quickstep_catalog_CatalogTypedefs
quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_HashTable
quickstep_storage_HashTableBase
quickstep_storage_HashTableFactory
quickstep_threading_SpinMutex
@@ -165,8 +163,6 @@ target_link_libraries(quickstep_expressions_aggregation_AggregationHandleAvg
quickstep_catalog_CatalogTypedefs
quickstep_expressions_aggregation_AggregationConcreteHandle
quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_HashTable
quickstep_storage_HashTableBase
quickstep_storage_HashTableFactory
quickstep_threading_SpinMutex
@@ -183,8 +179,6 @@ target_link_libraries(quickstep_expressions_aggregation_AggregationHandleCount
quickstep_catalog_CatalogTypedefs
quickstep_expressions_aggregation_AggregationConcreteHandle
quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_HashTable
quickstep_storage_HashTableBase
quickstep_storage_HashTableFactory
quickstep_storage_ValueAccessor
@@ -208,8 +202,6 @@ target_link_libraries(quickstep_expressions_aggregation_AggregationHandleMax
quickstep_catalog_CatalogTypedefs
quickstep_expressions_aggregation_AggregationConcreteHandle
quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_HashTable
quickstep_storage_HashTableBase
quickstep_storage_HashTableFactory
quickstep_threading_SpinMutex
@@ -225,8 +217,6 @@ target_link_libraries(quickstep_expressions_aggregation_AggregationHandleMin
quickstep_catalog_CatalogTypedefs
quickstep_expressions_aggregation_AggregationConcreteHandle
quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_HashTable
quickstep_storage_HashTableBase
quickstep_storage_HashTableFactory
quickstep_threading_SpinMutex
@@ -242,8 +232,6 @@ target_link_libraries(quickstep_expressions_aggregation_AggregationHandleSum
quickstep_catalog_CatalogTypedefs
quickstep_expressions_aggregation_AggregationConcreteHandle
quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_HashTable
quickstep_storage_HashTableBase
quickstep_storage_HashTableFactory
quickstep_threading_SpinMutex
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/query_execution/QueryContext.hpp
----------------------------------------------------------------------
diff --git a/query_execution/QueryContext.hpp b/query_execution/QueryContext.hpp
index 7ad8fa1..8733093 100644
--- a/query_execution/QueryContext.hpp
+++ b/query_execution/QueryContext.hpp
@@ -200,20 +200,6 @@ class QueryContext {
}
/**
- * @brief Destroy the payloads from the aggregation hash tables.
- *
- * @warning After calling these methods, the hash table will be in an invalid
- * state. No other operation should be performed on them.
- *
- * @param id The ID of the AggregationOperationState.
- **/
- inline void destroyAggregationHashTablePayload(const aggregation_state_id id) {
- DCHECK_LT(id, aggregation_states_.size());
- DCHECK(aggregation_states_[id]);
- aggregation_states_[id]->destroyAggregationHashTablePayload();
- }
-
- /**
* @brief Whether the given GeneratorFunctionHandle id is valid.
*
* @param id The GeneratorFunctionHandle id.
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/relational_operators/DestroyAggregationStateOperator.cpp
----------------------------------------------------------------------
diff --git a/relational_operators/DestroyAggregationStateOperator.cpp b/relational_operators/DestroyAggregationStateOperator.cpp
index 49be43d..62ca9e7 100644
--- a/relational_operators/DestroyAggregationStateOperator.cpp
+++ b/relational_operators/DestroyAggregationStateOperator.cpp
@@ -58,13 +58,6 @@ bool DestroyAggregationStateOperator::getAllWorkOrderProtos(WorkOrderProtosConta
}
void DestroyAggregationStateWorkOrder::execute() {
- // NOTE(harshad) : The destroyAggregationHashTablePayload call is separate
- // from the destroyAggregationState call. The reason is that the aggregation
- // hash tables don't own the AggregationHandle objects. However the hash table
- // class requires the handles for destroying the payload (see the
- // destroyPayload methods in AggregationHandle classes). Therefore, we first
- // destroy the payloads in the hash table and then destroy the hash table.
- query_context_->destroyAggregationHashTablePayload(aggr_state_index_);
query_context_->destroyAggregationState(aggr_state_index_);
}
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/AggregationOperationState.cpp
----------------------------------------------------------------------
diff --git a/storage/AggregationOperationState.cpp b/storage/AggregationOperationState.cpp
index b942c1b..8faf099 100644
--- a/storage/AggregationOperationState.cpp
+++ b/storage/AggregationOperationState.cpp
@@ -39,13 +39,13 @@
#include "expressions/predicate/Predicate.hpp"
#include "expressions/scalar/Scalar.hpp"
#include "storage/AggregationOperationState.pb.h"
-#include "storage/HashTable.hpp"
#include "storage/HashTableBase.hpp"
-#include "storage/HashTableFactory.hpp"
#include "storage/InsertDestination.hpp"
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
#include "storage/StorageBlock.hpp"
#include "storage/StorageBlockInfo.hpp"
#include "storage/StorageManager.hpp"
+#include "storage/SubBlocksReference.hpp"
#include "storage/TupleIdSequence.hpp"
#include "storage/ValueAccessor.hpp"
#include "types/TypedValue.hpp"
@@ -83,47 +83,51 @@ AggregationOperationState::AggregationOperationState(
is_aggregate_partitioned_(checkAggregatePartitioned(
estimated_num_entries, is_distinct, group_by, aggregate_functions)),
predicate_(predicate),
- group_by_list_(std::move(group_by)),
- arguments_(std::move(arguments)),
is_distinct_(std::move(is_distinct)),
storage_manager_(storage_manager) {
// Sanity checks: each aggregate has a corresponding list of arguments.
- DCHECK(aggregate_functions.size() == arguments_.size());
+ DCHECK(aggregate_functions.size() == arguments.size());
// Get the types of GROUP BY expressions for creating HashTables below.
- std::vector<const Type *> group_by_types;
- for (const std::unique_ptr<const Scalar> &group_by_element : group_by_list_) {
- group_by_types.emplace_back(&group_by_element->getType());
+ for (const std::unique_ptr<const Scalar> &group_by_element : group_by) {
+ group_by_types_.emplace_back(&group_by_element->getType());
+ }
+
+ // Prepare group-by element attribute ids and non-trivial expressions.
+ for (std::unique_ptr<const Scalar> &group_by_element : group_by) {
+ const attribute_id attr_id =
+ group_by_element->getAttributeIdForValueAccessor();
+ if (attr_id == kInvalidAttributeID) {
+ const attribute_id non_trivial_attr_id =
+ -(static_cast<attribute_id>(non_trivial_expressions_.size()) + 2);
+ non_trivial_expressions_.emplace_back(group_by_element.release());
+ group_by_key_ids_.emplace_back(non_trivial_attr_id);
+ } else {
+ group_by_key_ids_.emplace_back(attr_id);
+ }
}
std::vector<AggregationHandle *> group_by_handles;
- group_by_handles.clear();
if (aggregate_functions.size() == 0) {
// If there is no aggregation function, then it is a distinctify operation
// on the group-by expressions.
- DCHECK_GT(group_by_list_.size(), 0u);
+ DCHECK_GT(group_by_key_ids_.size(), 0u);
handles_.emplace_back(new AggregationHandleDistinct());
- arguments_.push_back({});
is_distinct_.emplace_back(false);
group_by_hashtable_pool_.reset(new HashTablePool(estimated_num_entries,
hash_table_impl_type,
- group_by_types,
- {1},
+ group_by_types_,
handles_,
storage_manager));
} else {
// Set up each individual aggregate in this operation.
std::vector<const AggregateFunction *>::const_iterator agg_func_it =
aggregate_functions.begin();
- std::vector<std::vector<std::unique_ptr<const Scalar>>>::const_iterator
- args_it = arguments_.begin();
+ std::vector<std::vector<std::unique_ptr<const Scalar>>>::iterator
+ args_it = arguments.begin();
std::vector<bool>::const_iterator is_distinct_it = is_distinct_.begin();
- std::vector<HashTableImplType>::const_iterator
- distinctify_hash_table_impl_types_it =
- distinctify_hash_table_impl_types.begin();
- std::vector<std::size_t> payload_sizes;
for (; agg_func_it != aggregate_functions.end();
++agg_func_it, ++args_it, ++is_distinct_it) {
// Get the Types of this aggregate's arguments so that we can create an
@@ -133,6 +137,22 @@ AggregationOperationState::AggregationOperationState(
argument_types.emplace_back(&argument->getType());
}
+ // Prepare argument attribute ids and non-trivial expressions.
+ std::vector<attribute_id> argument_ids;
+ for (std::unique_ptr<const Scalar> &argument : *args_it) {
+ const attribute_id attr_id =
+ argument->getAttributeIdForValueAccessor();
+ if (attr_id == kInvalidAttributeID) {
+ const attribute_id non_trivial_attr_id =
+ -(static_cast<attribute_id>(non_trivial_expressions_.size()) + 2);
+ non_trivial_expressions_.emplace_back(argument.release());
+ argument_ids.emplace_back(non_trivial_attr_id);
+ } else {
+ argument_ids.emplace_back(attr_id);
+ }
+ }
+ argument_ids_.emplace_back(std::move(argument_ids));
+
// Sanity checks: aggregate function exists and can apply to the specified
// arguments.
DCHECK(*agg_func_it != nullptr);
@@ -142,74 +162,25 @@ AggregationOperationState::AggregationOperationState(
// to do actual aggregate computation.
handles_.emplace_back((*agg_func_it)->createHandle(argument_types));
- if (!group_by_list_.empty()) {
+ if (!group_by_key_ids_.empty()) {
// Aggregation with GROUP BY: combined payload is partially updated in
// the presence of DISTINCT.
if (*is_distinct_it) {
- handles_.back()->blockUpdate();
+ LOG(FATAL) << "Distinct aggregation not supported";
}
group_by_handles.emplace_back(handles_.back());
- payload_sizes.emplace_back(group_by_handles.back()->getPayloadSize());
} else {
// Aggregation without GROUP BY: create a single global state.
single_states_.emplace_back(handles_.back()->createInitialState());
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- // See if all of this aggregate's arguments are attributes in the input
- // relation. If so, remember the attribute IDs so that we can do copy
- // elision when actually performing the aggregation.
- std::vector<attribute_id> local_arguments_as_attributes;
- local_arguments_as_attributes.reserve(args_it->size());
- for (const std::unique_ptr<const Scalar> &argument : *args_it) {
- const attribute_id argument_id =
- argument->getAttributeIdForValueAccessor();
- if (argument_id == -1) {
- local_arguments_as_attributes.clear();
- break;
- } else {
- DCHECK_EQ(input_relation_.getID(),
- argument->getRelationIdForValueAccessor());
- local_arguments_as_attributes.push_back(argument_id);
- }
- }
-
- arguments_as_attributes_.emplace_back(
- std::move(local_arguments_as_attributes));
-#endif
- }
-
- // Initialize the corresponding distinctify hash table if this is a
- // DISTINCT aggregation.
- if (*is_distinct_it) {
- std::vector<const Type *> key_types(group_by_types);
- key_types.insert(
- key_types.end(), argument_types.begin(), argument_types.end());
- // TODO(jianqiao): estimated_num_entries is quite inaccurate for
- // estimating the number of entries in the distinctify hash table.
- // We may estimate for each distinct aggregation an
- // estimated_num_distinct_keys value during query optimization, if it's
- // worth.
- distinctify_hashtables_.emplace_back(
- AggregationStateFastHashTableFactory::CreateResizable(
- *distinctify_hash_table_impl_types_it,
- key_types,
- estimated_num_entries,
- {0},
- {},
- storage_manager));
- ++distinctify_hash_table_impl_types_it;
- } else {
- distinctify_hashtables_.emplace_back(nullptr);
}
}
- if (!group_by_handles.empty()) {
- // Aggregation with GROUP BY: create a HashTable pool.
+ // Aggregation with GROUP BY: create a HashTable pool.
+ if (!group_by_key_ids_.empty()) {
if (!is_aggregate_partitioned_) {
group_by_hashtable_pool_.reset(new HashTablePool(estimated_num_entries,
hash_table_impl_type,
- group_by_types,
- payload_sizes,
+ group_by_types_,
group_by_handles,
storage_manager));
} else {
@@ -217,8 +188,7 @@ AggregationOperationState::AggregationOperationState(
new PartitionedHashTablePool(estimated_num_entries,
FLAGS_num_aggregation_partitions,
hash_table_impl_type,
- group_by_types,
- payload_sizes,
+ group_by_types_,
group_by_handles,
storage_manager));
}
@@ -269,7 +239,7 @@ AggregationOperationState* AggregationOperationState::ReconstructFromProto(
proto.group_by_expressions(group_by_idx), database));
}
- unique_ptr<Predicate> predicate;
+ std::unique_ptr<Predicate> predicate;
if (proto.has_predicate()) {
predicate.reset(
PredicateFactory::ReconstructFromProto(proto.predicate(), database));
@@ -355,7 +325,7 @@ bool AggregationOperationState::ProtoIsValid(
void AggregationOperationState::aggregateBlock(const block_id input_block,
LIPFilterAdaptiveProber *lip_filter_adaptive_prober) {
- if (group_by_list_.empty()) {
+ if (group_by_key_ids_.empty()) {
aggregateBlockSingleState(input_block);
} else {
aggregateBlockHashTable(input_block, lip_filter_adaptive_prober);
@@ -364,7 +334,7 @@ void AggregationOperationState::aggregateBlock(const block_id input_block,
void AggregationOperationState::finalizeAggregate(
InsertDestination *output_destination) {
- if (group_by_list_.empty()) {
+ if (group_by_key_ids_.empty()) {
finalizeSingleState(output_destination);
} else {
finalizeHashTable(output_destination);
@@ -392,38 +362,38 @@ void AggregationOperationState::aggregateBlockSingleState(
std::unique_ptr<TupleIdSequence> matches;
if (predicate_ != nullptr) {
- std::unique_ptr<ValueAccessor> accessor(
- block->getTupleStorageSubBlock().createValueAccessor());
matches.reset(block->getMatchesForPredicate(predicate_.get(), matches.get()));
}
- for (std::size_t agg_idx = 0; agg_idx < handles_.size(); ++agg_idx) {
- const std::vector<attribute_id> *local_arguments_as_attributes = nullptr;
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- // If all arguments are attributes of the input relation, elide a copy.
- if (!arguments_as_attributes_[agg_idx].empty()) {
- local_arguments_as_attributes = &(arguments_as_attributes_[agg_idx]);
+ const auto &tuple_store = block->getTupleStorageSubBlock();
+ std::unique_ptr<ValueAccessor> accessor(
+ tuple_store.createValueAccessor(matches.get()));
+
+ ColumnVectorsValueAccessor non_trivial_results;
+ if (!non_trivial_expressions_.empty()) {
+ SubBlocksReference sub_blocks_ref(tuple_store,
+ block->getIndices(),
+ block->getIndicesConsistent());
+ for (const auto &expression : non_trivial_expressions_) {
+ non_trivial_results.addColumn(
+ expression->getAllValues(accessor.get(), &sub_blocks_ref));
}
-#endif
- if (is_distinct_[agg_idx]) {
- // Call StorageBlock::aggregateDistinct() to put the arguments as keys
- // directly into the (threadsafe) shared global distinctify HashTable
- // for this aggregate.
- block->aggregateDistinct(*handles_[agg_idx],
- arguments_[agg_idx],
- local_arguments_as_attributes,
- {}, /* group_by */
- matches.get(),
- distinctify_hashtables_[agg_idx].get(),
- nullptr /* reuse_group_by_vectors */);
- local_state.emplace_back(nullptr);
+ }
+
+ for (std::size_t agg_idx = 0; agg_idx < handles_.size(); ++agg_idx) {
+ const auto &argument_ids = argument_ids_[agg_idx];
+ const auto *handle = handles_[agg_idx];
+
+ AggregationState *state;
+ if (argument_ids.empty()) {
+ // Special case. This is a nullary aggregate (i.e. COUNT(*)).
+ state = handle->accumulateNullary(matches == nullptr ? tuple_store.numTuples()
+ : matches->size());
} else {
- // Call StorageBlock::aggregate() to actually do the aggregation.
- local_state.emplace_back(block->aggregate(*handles_[agg_idx],
- arguments_[agg_idx],
- local_arguments_as_attributes,
- matches.get()));
+ // Have the AggregationHandle actually do the aggregation.
+ state = handle->accumulate(accessor.get(), &non_trivial_results, argument_ids);
}
+ local_state.emplace_back(state);
}
// Merge per-block aggregation states back with global state.
@@ -435,104 +405,93 @@ void AggregationOperationState::aggregateBlockHashTable(
LIPFilterAdaptiveProber *lip_filter_adaptive_prober) {
BlockReference block(
storage_manager_->getBlock(input_block, input_relation_));
+ const auto &tuple_store = block->getTupleStorageSubBlock();
+ std::unique_ptr<ValueAccessor> base_accessor(tuple_store.createValueAccessor());
+ std::unique_ptr<ValueAccessor> shared_accessor;
+ ValueAccessor *accessor = base_accessor.get();
// Apply the predicate first, then the LIPFilters, to generate a TupleIdSequence
// as the existence map for the tuples.
std::unique_ptr<TupleIdSequence> matches;
if (predicate_ != nullptr) {
matches.reset(block->getMatchesForPredicate(predicate_.get()));
+ shared_accessor.reset(
+ base_accessor->createSharedTupleIdSequenceAdapterVirtual(*matches));
+ accessor = shared_accessor.get();
}
if (lip_filter_adaptive_prober != nullptr) {
- std::unique_ptr<ValueAccessor> accessor(
- block->getTupleStorageSubBlock().createValueAccessor(matches.get()));
- matches.reset(lip_filter_adaptive_prober->filterValueAccessor(accessor.get()));
+ matches.reset(lip_filter_adaptive_prober->filterValueAccessor(accessor));
+ shared_accessor.reset(
+ base_accessor->createSharedTupleIdSequenceAdapterVirtual(*matches));
+ accessor = shared_accessor.get();
}
- // This holds values of all the GROUP BY attributes so that the can be reused
- // across multiple aggregates (i.e. we only pay the cost of evaluatin the
- // GROUP BY expressions once).
- std::vector<std::unique_ptr<ColumnVector>> reuse_group_by_vectors;
-
- for (std::size_t agg_idx = 0; agg_idx < handles_.size(); ++agg_idx) {
- if (is_distinct_[agg_idx]) {
- // Call StorageBlock::aggregateDistinct() to insert the GROUP BY
- // expression
- // values and the aggregation arguments together as keys directly into the
- // (threadsafe) shared global distinctify HashTable for this aggregate.
- block->aggregateDistinct(*handles_[agg_idx],
- arguments_[agg_idx],
- nullptr, /* arguments_as_attributes */
- group_by_list_,
- matches.get(),
- distinctify_hashtables_[agg_idx].get(),
- &reuse_group_by_vectors);
+ std::unique_ptr<ColumnVectorsValueAccessor> non_trivial_results;
+ if (!non_trivial_expressions_.empty()) {
+ non_trivial_results.reset(new ColumnVectorsValueAccessor());
+ SubBlocksReference sub_blocks_ref(tuple_store,
+ block->getIndices(),
+ block->getIndicesConsistent());
+ for (const auto &expression : non_trivial_expressions_) {
+ non_trivial_results->addColumn(
+ expression->getAllValues(accessor, &sub_blocks_ref));
}
}
if (!is_aggregate_partitioned_) {
- // Call StorageBlock::aggregateGroupBy() to aggregate this block's values
- // directly into the (threadsafe) shared global HashTable for this
- // aggregate.
DCHECK(group_by_hashtable_pool_ != nullptr);
+
AggregationStateHashTableBase *agg_hash_table =
- group_by_hashtable_pool_->getHashTableFast();
+ group_by_hashtable_pool_->getHashTable();
DCHECK(agg_hash_table != nullptr);
- block->aggregateGroupBy(arguments_,
- group_by_list_,
- matches.get(),
- agg_hash_table,
- &reuse_group_by_vectors);
+
+ std::cout << "Upsert VA start\n";
+ std::flush(std::cout);
+ accessor->beginIterationVirtual();
+ agg_hash_table->upsertValueAccessor(argument_ids_,
+ group_by_key_ids_,
+ accessor,
+ non_trivial_results.get());
+ std::cout << "Aggregate done\n";
+ std::flush(std::cout);
group_by_hashtable_pool_->returnHashTable(agg_hash_table);
} else {
- ColumnVectorsValueAccessor temp_result;
- // IDs of 'arguments' as attributes in the ValueAccessor we create below.
- std::vector<attribute_id> argument_ids;
-
- // IDs of GROUP BY key element(s) in the ValueAccessor we create below.
- std::vector<attribute_id> key_ids;
- const std::size_t num_partitions = partitioned_group_by_hashtable_pool_->getNumPartitions();
- block->aggregateGroupByPartitioned(
- arguments_,
- group_by_list_,
- matches.get(),
- num_partitions,
- &temp_result,
- &argument_ids,
- &key_ids,
- &reuse_group_by_vectors);
- // Compute the partitions for the tuple formed by group by values.
- std::vector<std::unique_ptr<TupleIdSequence>> partition_membership;
- partition_membership.resize(num_partitions);
-
- // Create a tuple-id sequence for each partition.
- for (std::size_t partition = 0; partition < num_partitions; ++partition) {
- partition_membership[partition].reset(
- new TupleIdSequence(temp_result.getEndPosition()));
- }
-
- // Iterate over ValueAccessor for each tuple,
- // set a bit in the appropriate TupleIdSequence.
- temp_result.beginIteration();
- while (temp_result.next()) {
- // We need a unique_ptr because getTupleWithAttributes() uses "new".
- std::unique_ptr<Tuple> curr_tuple(temp_result.getTupleWithAttributes(key_ids));
- const std::size_t curr_tuple_partition_id =
- curr_tuple->getTupleHash() % num_partitions;
- partition_membership[curr_tuple_partition_id]->set(
- temp_result.getCurrentPosition(), true);
- }
- // For each partition, create an adapter around Value Accessor and
- // TupleIdSequence.
- std::vector<std::unique_ptr<
- TupleIdSequenceAdapterValueAccessor<ColumnVectorsValueAccessor>>> adapter;
- adapter.resize(num_partitions);
- for (std::size_t partition = 0; partition < num_partitions; ++partition) {
- adapter[partition].reset(temp_result.createSharedTupleIdSequenceAdapter(
- *(partition_membership)[partition]));
- partitioned_group_by_hashtable_pool_->getHashTable(partition)
- ->upsertValueAccessorCompositeKeyFast(
- argument_ids, adapter[partition].get(), key_ids, true);
- }
+ LOG(FATAL) << "Partitioned aggregation not supported";
+// const std::size_t num_partitions = partitioned_group_by_hashtable_pool_->getNumPartitions();
+//
+// // Compute the partitions for the tuple formed by group by values.
+// std::vector<std::unique_ptr<TupleIdSequence>> partition_membership;
+// partition_membership.resize(num_partitions);
+//
+// // Create a tuple-id sequence for each partition.
+// for (std::size_t partition = 0; partition < num_partitions; ++partition) {
+// partition_membership[partition].reset(
+// new TupleIdSequence(temp_result.getEndPosition()));
+// }
+//
+// // Iterate over ValueAccessor for each tuple,
+// // set a bit in the appropriate TupleIdSequence.
+// temp_result.beginIteration();
+// while (temp_result.next()) {
+// // We need a unique_ptr because getTupleWithAttributes() uses "new".
+// std::unique_ptr<Tuple> curr_tuple(temp_result.getTupleWithAttributes(key_ids));
+// const std::size_t curr_tuple_partition_id =
+// curr_tuple->getTupleHash() % num_partitions;
+// partition_membership[curr_tuple_partition_id]->set(
+// temp_result.getCurrentPosition(), true);
+// }
+// // For each partition, create an adapter around Value Accessor and
+// // TupleIdSequence.
+// std::vector<std::unique_ptr<
+// TupleIdSequenceAdapterValueAccessor<ColumnVectorsValueAccessor>>> adapter;
+// adapter.resize(num_partitions);
+// for (std::size_t partition = 0; partition < num_partitions; ++partition) {
+// adapter[partition].reset(temp_result.createSharedTupleIdSequenceAdapter(
+// *(partition_membership)[partition]));
+// partitioned_group_by_hashtable_pool_->getHashTable(partition)
+// ->upsertValueAccessorCompositeKeyFast(
+// argument_ids, adapter[partition].get(), key_ids, true);
+// }
}
}
@@ -543,12 +502,6 @@ void AggregationOperationState::finalizeSingleState(
std::vector<TypedValue> attribute_values;
for (std::size_t agg_idx = 0; agg_idx < handles_.size(); ++agg_idx) {
- if (is_distinct_[agg_idx]) {
- single_states_[agg_idx].reset(
- handles_[agg_idx]->aggregateOnDistinctifyHashTableForSingle(
- *distinctify_hashtables_[agg_idx]));
- }
-
attribute_values.emplace_back(
handles_[agg_idx]->finalize(*single_states_[agg_idx]));
}
@@ -559,8 +512,8 @@ void AggregationOperationState::finalizeSingleState(
void AggregationOperationState::mergeGroupByHashTables(
AggregationStateHashTableBase *src, AggregationStateHashTableBase *dst) {
HashTableMergerFast merger(dst);
- (static_cast<FastHashTable<true, false, true, false> *>(src))
- ->forEachCompositeKeyFast(&merger);
+ static_cast<PackedPayloadSeparateChainingAggregationStateHashTable *>(src)
+ ->forEach(&merger);
}
void AggregationOperationState::finalizeHashTable(
@@ -577,59 +530,30 @@ void AggregationOperationState::finalizeHashTable(
// e.g. Keep merging entries from smaller hash tables to larger.
auto *hash_tables = group_by_hashtable_pool_->getAllHashTables();
- if (hash_tables->size() > 1) {
- for (int hash_table_index = 0;
- hash_table_index < static_cast<int>(hash_tables->size() - 1);
- ++hash_table_index) {
- // Merge each hash table to the last hash table.
- mergeGroupByHashTables((*hash_tables)[hash_table_index].get(),
- hash_tables->back().get());
- }
+ DCHECK(hash_tables != nullptr);
+ if (hash_tables->empty()) {
+ return;
+ }
+
+ std::unique_ptr<AggregationStateHashTableBase> final_hash_table(
+ hash_tables->back().release());
+ for (std::size_t i = 0; i < hash_tables->size() - 1; ++i) {
+ std::unique_ptr<AggregationStateHashTableBase> hash_table(
+ hash_tables->at(i).release());
+ mergeGroupByHashTables(hash_table.get(), final_hash_table.get());
+ hash_table->destroyPayload();
}
// Collect per-aggregate finalized values.
std::vector<std::unique_ptr<ColumnVector>> final_values;
for (std::size_t agg_idx = 0; agg_idx < handles_.size(); ++agg_idx) {
- if (is_distinct_[agg_idx]) {
- DCHECK(group_by_hashtable_pool_ != nullptr);
- auto *hash_tables = group_by_hashtable_pool_->getAllHashTables();
- DCHECK(hash_tables != nullptr);
- if (hash_tables->empty()) {
- // We may have a case where hash_tables is empty, e.g. no input blocks.
- // However for aggregateOnDistinctifyHashTableForGroupBy to work
- // correctly, we should create an empty group by hash table.
- AggregationStateHashTableBase *new_hash_table =
- group_by_hashtable_pool_->getHashTableFast();
- group_by_hashtable_pool_->returnHashTable(new_hash_table);
- hash_tables = group_by_hashtable_pool_->getAllHashTables();
- }
- DCHECK(hash_tables->back() != nullptr);
- AggregationStateHashTableBase *agg_hash_table = hash_tables->back().get();
- DCHECK(agg_hash_table != nullptr);
- handles_[agg_idx]->allowUpdate();
- handles_[agg_idx]->aggregateOnDistinctifyHashTableForGroupBy(
- *distinctify_hashtables_[agg_idx], agg_hash_table, agg_idx);
- }
-
- auto *hash_tables = group_by_hashtable_pool_->getAllHashTables();
- DCHECK(hash_tables != nullptr);
- if (hash_tables->empty()) {
- // We may have a case where hash_tables is empty, e.g. no input blocks.
- // However for aggregateOnDistinctifyHashTableForGroupBy to work
- // correctly, we should create an empty group by hash table.
- AggregationStateHashTableBase *new_hash_table =
- group_by_hashtable_pool_->getHashTableFast();
- group_by_hashtable_pool_->returnHashTable(new_hash_table);
- hash_tables = group_by_hashtable_pool_->getAllHashTables();
- }
- AggregationStateHashTableBase *agg_hash_table = hash_tables->back().get();
- DCHECK(agg_hash_table != nullptr);
ColumnVector *agg_result_col = handles_[agg_idx]->finalizeHashTable(
- *agg_hash_table, &group_by_keys, agg_idx);
+ *final_hash_table, &group_by_keys, agg_idx);
if (agg_result_col != nullptr) {
final_values.emplace_back(agg_result_col);
}
}
+ final_hash_table->destroyPayload();
// Reorganize 'group_by_keys' in column-major order so that we can make a
// ColumnVectorsValueAccessor to bulk-insert results.
@@ -640,11 +564,10 @@ void AggregationOperationState::finalizeHashTable(
// in a single HashTable.
std::vector<std::unique_ptr<ColumnVector>> group_by_cvs;
std::size_t group_by_element_idx = 0;
- for (const std::unique_ptr<const Scalar> &group_by_element : group_by_list_) {
- const Type &group_by_type = group_by_element->getType();
- if (NativeColumnVector::UsableForType(group_by_type)) {
+ for (const Type *group_by_type : group_by_types_) {
+ if (NativeColumnVector::UsableForType(*group_by_type)) {
NativeColumnVector *element_cv =
- new NativeColumnVector(group_by_type, group_by_keys.size());
+ new NativeColumnVector(*group_by_type, group_by_keys.size());
group_by_cvs.emplace_back(element_cv);
for (std::vector<TypedValue> &group_key : group_by_keys) {
element_cv->appendTypedValue(
@@ -652,7 +575,7 @@ void AggregationOperationState::finalizeHashTable(
}
} else {
IndirectColumnVector *element_cv =
- new IndirectColumnVector(group_by_type, group_by_keys.size());
+ new IndirectColumnVector(*group_by_type, group_by_keys.size());
group_by_cvs.emplace_back(element_cv);
for (std::vector<TypedValue> &group_key : group_by_keys) {
element_cv->appendTypedValue(
@@ -676,25 +599,6 @@ void AggregationOperationState::finalizeHashTable(
output_destination->bulkInsertTuples(&complete_result);
}
-void AggregationOperationState::destroyAggregationHashTablePayload() {
- std::vector<std::unique_ptr<AggregationStateHashTableBase>> *all_hash_tables =
- nullptr;
- if (!is_aggregate_partitioned_) {
- if (group_by_hashtable_pool_ != nullptr) {
- all_hash_tables = group_by_hashtable_pool_->getAllHashTables();
- }
- } else {
- if (partitioned_group_by_hashtable_pool_ != nullptr) {
- all_hash_tables = partitioned_group_by_hashtable_pool_->getAllHashTables();
- }
- }
- if (all_hash_tables != nullptr) {
- for (std::size_t ht_index = 0; ht_index < all_hash_tables->size(); ++ht_index) {
- (*all_hash_tables)[ht_index]->destroyPayload();
- }
- }
-}
-
void AggregationOperationState::finalizeAggregatePartitioned(
const std::size_t partition_id, InsertDestination *output_destination) {
// Each element of 'group_by_keys' is a vector of values for a particular
@@ -703,15 +607,16 @@ void AggregationOperationState::finalizeAggregatePartitioned(
// Collect per-aggregate finalized values.
std::vector<std::unique_ptr<ColumnVector>> final_values;
+ AggregationStateHashTableBase *hash_table =
+ partitioned_group_by_hashtable_pool_->getHashTable(partition_id);
for (std::size_t agg_idx = 0; agg_idx < handles_.size(); ++agg_idx) {
- AggregationStateHashTableBase *hash_table =
- partitioned_group_by_hashtable_pool_->getHashTable(partition_id);
ColumnVector *agg_result_col = handles_[agg_idx]->finalizeHashTable(
*hash_table, &group_by_keys, agg_idx);
if (agg_result_col != nullptr) {
final_values.emplace_back(agg_result_col);
}
}
+ hash_table->destroyPayload();
// Reorganize 'group_by_keys' in column-major order so that we can make a
// ColumnVectorsValueAccessor to bulk-insert results.
@@ -722,16 +627,17 @@ void AggregationOperationState::finalizeAggregatePartitioned(
// in a single HashTable.
std::vector<std::unique_ptr<ColumnVector>> group_by_cvs;
std::size_t group_by_element_idx = 0;
- for (const std::unique_ptr<const Scalar> &group_by_element : group_by_list_) {
- const Type &group_by_type = group_by_element->getType();
- if (NativeColumnVector::UsableForType(group_by_type)) {
- NativeColumnVector *element_cv = new NativeColumnVector(group_by_type, group_by_keys.size());
+ for (const Type *group_by_type : group_by_types_) {
+ if (NativeColumnVector::UsableForType(*group_by_type)) {
+ NativeColumnVector *element_cv =
+ new NativeColumnVector(*group_by_type, group_by_keys.size());
group_by_cvs.emplace_back(element_cv);
for (std::vector<TypedValue> &group_key : group_by_keys) {
element_cv->appendTypedValue(std::move(group_key[group_by_element_idx]));
}
} else {
- IndirectColumnVector *element_cv = new IndirectColumnVector(group_by_type, group_by_keys.size());
+ IndirectColumnVector *element_cv =
+ new IndirectColumnVector(*group_by_type, group_by_keys.size());
group_by_cvs.emplace_back(element_cv);
for (std::vector<TypedValue> &group_key : group_by_keys) {
element_cv->appendTypedValue(std::move(group_key[group_by_element_idx]));
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/AggregationOperationState.hpp
----------------------------------------------------------------------
diff --git a/storage/AggregationOperationState.hpp b/storage/AggregationOperationState.hpp
index e0826b0..0f3807e 100644
--- a/storage/AggregationOperationState.hpp
+++ b/storage/AggregationOperationState.hpp
@@ -33,6 +33,7 @@
#include "storage/HashTableBase.hpp"
#include "storage/HashTablePool.hpp"
#include "storage/PartitionedHashTablePool.hpp"
+#include "storage/StorageBlock.hpp"
#include "storage/StorageBlockInfo.hpp"
#include "utility/Macros.hpp"
@@ -46,6 +47,7 @@ class CatalogRelationSchema;
class InsertDestination;
class LIPFilterAdaptiveProber;
class StorageManager;
+class TupleIdSequence;
DECLARE_int32(num_aggregation_partitions);
DECLARE_int32(partition_aggregation_num_groups_threshold);
@@ -178,11 +180,6 @@ class AggregationOperationState {
void finalizeAggregate(InsertDestination *output_destination);
/**
- * @brief Destroy the payloads in the aggregation hash tables.
- **/
- void destroyAggregationHashTablePayload();
-
- /**
* @brief Generate the final results for the aggregates managed by this
* AggregationOperationState and write them out to StorageBlock(s).
* In this implementation, each thread picks a hash table belonging to
@@ -236,27 +233,28 @@ class AggregationOperationState {
const std::vector<bool> &is_distinct,
const std::vector<std::unique_ptr<const Scalar>> &group_by,
const std::vector<const AggregateFunction *> &aggregate_functions) const {
- // If there's no aggregation, return false.
- if (aggregate_functions.empty()) {
- return false;
- }
- // Check if there's a distinct operation involved in any aggregate, if so
- // the aggregate can't be partitioned.
- for (auto distinct : is_distinct) {
- if (distinct) {
- return false;
- }
- }
- // There's no distinct aggregation involved, Check if there's at least one
- // GROUP BY operation.
- if (group_by.empty()) {
- return false;
- }
- // There are GROUP BYs without DISTINCT. Check if the estimated number of
- // groups is large enough to warrant a partitioned aggregation.
- return estimated_num_groups >
- static_cast<std::size_t>(
- FLAGS_partition_aggregation_num_groups_threshold);
+// // If there's no aggregation, return false.
+// if (aggregate_functions.empty()) {
+// return false;
+// }
+// // Check if there's a distinct operation involved in any aggregate, if so
+// // the aggregate can't be partitioned.
+// for (auto distinct : is_distinct) {
+// if (distinct) {
+// return false;
+// }
+// }
+// // There's no distinct aggregation involved, Check if there's at least one
+// // GROUP BY operation.
+// if (group_by.empty()) {
+// return false;
+// }
+// // There are GROUP BYs without DISTINCT. Check if the estimated number of
+// // groups is large enough to warrant a partitioned aggregation.
+// return estimated_num_groups >
+// static_cast<std::size_t>(
+// FLAGS_partition_aggregation_num_groups_threshold);
+ return false;
}
// Common state for all aggregates in this operation: the input relation, the
@@ -267,26 +265,26 @@ class AggregationOperationState {
const bool is_aggregate_partitioned_;
std::unique_ptr<const Predicate> predicate_;
- std::vector<std::unique_ptr<const Scalar>> group_by_list_;
// Each individual aggregate in this operation has an AggregationHandle and
- // some number of Scalar arguments.
+ // zero (indicated by -1) or one argument.
std::vector<AggregationHandle *> handles_;
- std::vector<std::vector<std::unique_ptr<const Scalar>>> arguments_;
// For each aggregate, whether DISTINCT should be applied to the aggregate's
// arguments.
std::vector<bool> is_distinct_;
- // Hash table for obtaining distinct (i.e. unique) arguments.
- std::vector<std::unique_ptr<AggregationStateHashTableBase>>
- distinctify_hashtables_;
+ // Non-trivial group-by/argument expressions that need to be evaluated.
+ std::vector<std::unique_ptr<const Scalar>> non_trivial_expressions_;
+
+ std::vector<attribute_id> group_by_key_ids_;
+ std::vector<std::vector<attribute_id>> argument_ids_;
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- // If all an aggregate's argument expressions are simply attributes in
- // 'input_relation_', then this caches the attribute IDs of those arguments.
- std::vector<std::vector<attribute_id>> arguments_as_attributes_;
-#endif
+ std::vector<const Type *> group_by_types_;
+
+ // Hash table for obtaining distinct (i.e. unique) arguments.
+// std::vector<std::unique_ptr<AggregationStateHashTableBase>>
+// distinctify_hashtables_;
// Per-aggregate global states for aggregation without GROUP BY.
std::vector<std::unique_ptr<AggregationState>> single_states_;
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/CMakeLists.txt
----------------------------------------------------------------------
diff --git a/storage/CMakeLists.txt b/storage/CMakeLists.txt
index be60662..2709d9d 100644
--- a/storage/CMakeLists.txt
+++ b/storage/CMakeLists.txt
@@ -198,9 +198,6 @@ if (ENABLE_DISTRIBUTED)
endif()
add_library(quickstep_storage_EvictionPolicy EvictionPolicy.cpp EvictionPolicy.hpp)
-add_library(quickstep_storage_FastHashTable ../empty_src.cpp FastHashTable.hpp)
-add_library(quickstep_storage_FastHashTableFactory ../empty_src.cpp FastHashTableFactory.hpp)
-add_library(quickstep_storage_FastSeparateChainingHashTable ../empty_src.cpp FastSeparateChainingHashTable.hpp)
add_library(quickstep_storage_FileManager ../empty_src.cpp FileManager.hpp)
if (QUICKSTEP_HAVE_FILE_MANAGER_HDFS)
add_library(quickstep_storage_FileManagerHdfs FileManagerHdfs.cpp FileManagerHdfs.hpp)
@@ -229,6 +226,9 @@ add_library(quickstep_storage_InsertDestination_proto
add_library(quickstep_storage_LinearOpenAddressingHashTable
../empty_src.cpp
LinearOpenAddressingHashTable.hpp)
+add_library(quickstep_storage_PackedPayloadAggregationStateHashTable
+ PackedPayloadAggregationStateHashTable.cpp
+ PackedPayloadAggregationStateHashTable.hpp)
add_library(quickstep_storage_PackedRowStoreTupleStorageSubBlock
PackedRowStoreTupleStorageSubBlock.cpp
PackedRowStoreTupleStorageSubBlock.hpp)
@@ -286,15 +286,15 @@ target_link_libraries(quickstep_storage_AggregationOperationState
quickstep_expressions_predicate_Predicate
quickstep_expressions_scalar_Scalar
quickstep_storage_AggregationOperationState_proto
- quickstep_storage_HashTable
quickstep_storage_HashTableBase
- quickstep_storage_HashTableFactory
quickstep_storage_HashTablePool
quickstep_storage_InsertDestination
+ quickstep_storage_PackedPayloadAggregationStateHashTable
quickstep_storage_PartitionedHashTablePool
quickstep_storage_StorageBlock
quickstep_storage_StorageBlockInfo
quickstep_storage_StorageManager
+ quickstep_storage_SubBlocksReference
quickstep_storage_TupleIdSequence
quickstep_storage_ValueAccessor
quickstep_types_TypedValue
@@ -635,52 +635,6 @@ target_link_libraries(quickstep_storage_EvictionPolicy
quickstep_threading_SpinMutex
quickstep_threading_SpinSharedMutex
quickstep_utility_Macros)
-target_link_libraries(quickstep_storage_FastHashTable
- quickstep_catalog_CatalogTypedefs
- quickstep_storage_HashTableBase
- quickstep_storage_StorageBlob
- quickstep_storage_StorageBlockInfo
- quickstep_storage_StorageConstants
- quickstep_storage_StorageManager
- quickstep_storage_TupleReference
- quickstep_storage_ValueAccessor
- quickstep_storage_ValueAccessorUtil
- quickstep_threading_SpinMutex
- quickstep_threading_SpinSharedMutex
- quickstep_types_Type
- quickstep_types_TypedValue
- quickstep_utility_HashPair
- quickstep_utility_Macros)
-target_link_libraries(quickstep_storage_FastHashTableFactory
- glog
- quickstep_storage_FastHashTable
- quickstep_storage_FastSeparateChainingHashTable
- quickstep_storage_HashTable
- quickstep_storage_HashTable_proto
- quickstep_storage_HashTableBase
- quickstep_storage_HashTableFactory
- quickstep_storage_LinearOpenAddressingHashTable
- quickstep_storage_SeparateChainingHashTable
- quickstep_storage_SimpleScalarSeparateChainingHashTable
- quickstep_storage_TupleReference
- quickstep_types_TypeFactory
- quickstep_utility_Macros)
-target_link_libraries(quickstep_storage_FastSeparateChainingHashTable
- quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_HashTable
- quickstep_storage_HashTableBase
- quickstep_storage_HashTableKeyManager
- quickstep_storage_StorageBlob
- quickstep_storage_StorageBlockInfo
- quickstep_storage_StorageConstants
- quickstep_storage_StorageManager
- quickstep_threading_SpinSharedMutex
- quickstep_types_Type
- quickstep_types_TypedValue
- quickstep_utility_Alignment
- quickstep_utility_Macros
- quickstep_utility_PrimeNumber)
target_link_libraries(quickstep_storage_FileManager
quickstep_storage_StorageBlockInfo
quickstep_utility_Macros
@@ -747,6 +701,7 @@ target_link_libraries(quickstep_storage_HashTableFactory
quickstep_storage_HashTable_proto
quickstep_storage_HashTableBase
quickstep_storage_LinearOpenAddressingHashTable
+ quickstep_storage_PackedPayloadAggregationStateHashTable
quickstep_storage_SeparateChainingHashTable
quickstep_storage_SimpleScalarSeparateChainingHashTable
quickstep_storage_TupleReference
@@ -766,9 +721,8 @@ target_link_libraries(quickstep_storage_HashTableKeyManager
target_link_libraries(quickstep_storage_HashTablePool
glog
quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_FastHashTableFactory
quickstep_storage_HashTableBase
+ quickstep_storage_HashTableFactory
quickstep_threading_SpinMutex
quickstep_utility_Macros
quickstep_utility_StringUtil)
@@ -826,6 +780,27 @@ target_link_libraries(quickstep_storage_LinearOpenAddressingHashTable
quickstep_utility_Alignment
quickstep_utility_Macros
quickstep_utility_PrimeNumber)
+target_link_libraries(quickstep_storage_PackedPayloadAggregationStateHashTable
+ quickstep_catalog_CatalogTypedefs
+ quickstep_expressions_aggregation_AggregationHandle
+ quickstep_storage_HashTableBase
+ quickstep_storage_HashTableKeyManager
+ quickstep_storage_StorageBlob
+ quickstep_storage_StorageBlockInfo
+ quickstep_storage_StorageConstants
+ quickstep_storage_StorageManager
+ quickstep_storage_TupleReference
+ quickstep_storage_ValueAccessor
+ quickstep_storage_ValueAccessorUtil
+ quickstep_threading_SpinMutex
+ quickstep_threading_SpinSharedMutex
+ quickstep_types_Type
+ quickstep_types_TypedValue
+ quickstep_types_containers_ColumnVectorsValueAccessor
+ quickstep_utility_Alignment
+ quickstep_utility_HashPair
+ quickstep_utility_Macros
+ quickstep_utility_PrimeNumber)
target_link_libraries(quickstep_storage_PackedRowStoreTupleStorageSubBlock
quickstep_catalog_CatalogAttribute
quickstep_catalog_CatalogRelationSchema
@@ -858,8 +833,7 @@ target_link_libraries(quickstep_storage_PackedRowStoreValueAccessor
target_link_libraries(quickstep_storage_PartitionedHashTablePool
glog
quickstep_expressions_aggregation_AggregationHandle
- quickstep_storage_FastHashTable
- quickstep_storage_FastHashTableFactory
+ quickstep_storage_HashTableFactory
quickstep_storage_HashTableBase
quickstep_utility_Macros
quickstep_utility_StringUtil)
@@ -1163,9 +1137,6 @@ target_link_libraries(quickstep_storage
quickstep_storage_EvictionPolicy
quickstep_storage_FileManager
quickstep_storage_FileManagerLocal
- quickstep_storage_FastHashTable
- quickstep_storage_FastHashTableFactory
- quickstep_storage_FastSeparateChainingHashTable
quickstep_storage_HashTable
quickstep_storage_HashTable_proto
quickstep_storage_HashTableBase
@@ -1180,6 +1151,7 @@ target_link_libraries(quickstep_storage
quickstep_storage_LinearOpenAddressingHashTable
quickstep_storage_PackedRowStoreTupleStorageSubBlock
quickstep_storage_PackedRowStoreValueAccessor
+ quickstep_storage_PackedPayloadAggregationStateHashTable
quickstep_storage_PartitionedHashTablePool
quickstep_storage_PreloaderThread
quickstep_storage_SMAIndexSubBlock
[5/5] incubator-quickstep git commit: Cleanup FastHashTable + Avoid
unnecessary copying of data before aggregation.
Posted by ji...@apache.org.
Cleanup FastHashTable + Avoid unnecessary copying of data before aggregation.
Project: http://git-wip-us.apache.org/repos/asf/incubator-quickstep/repo
Commit: http://git-wip-us.apache.org/repos/asf/incubator-quickstep/commit/7048fcc1
Tree: http://git-wip-us.apache.org/repos/asf/incubator-quickstep/tree/7048fcc1
Diff: http://git-wip-us.apache.org/repos/asf/incubator-quickstep/diff/7048fcc1
Branch: refs/heads/collision-free-agg
Commit: 7048fcc1f7a0e152f06e84a549196655718af64a
Parents: bc3d8a5
Author: Jianqiao Zhu <ji...@cs.wisc.edu>
Authored: Thu Oct 27 12:23:03 2016 -0500
Committer: Jianqiao Zhu <ji...@cs.wisc.edu>
Committed: Tue Nov 1 01:16:29 2016 -0500
----------------------------------------------------------------------
.../aggregation/AggregationConcreteHandle.cpp | 44 -
.../aggregation/AggregationConcreteHandle.hpp | 127 +-
expressions/aggregation/AggregationHandle.hpp | 188 +-
.../aggregation/AggregationHandleAvg.cpp | 82 +-
.../aggregation/AggregationHandleAvg.hpp | 108 +-
.../aggregation/AggregationHandleCount.cpp | 127 +-
.../aggregation/AggregationHandleCount.hpp | 120 +-
.../aggregation/AggregationHandleDistinct.cpp | 34 +-
.../aggregation/AggregationHandleDistinct.hpp | 43 +-
.../aggregation/AggregationHandleMax.cpp | 82 +-
.../aggregation/AggregationHandleMax.hpp | 93 +-
.../aggregation/AggregationHandleMin.cpp | 82 +-
.../aggregation/AggregationHandleMin.hpp | 103 +-
.../aggregation/AggregationHandleSum.cpp | 79 +-
.../aggregation/AggregationHandleSum.hpp | 110 +-
expressions/aggregation/CMakeLists.txt | 12 -
query_execution/QueryContext.hpp | 14 -
.../DestroyAggregationStateOperator.cpp | 7 -
storage/AggregationOperationState.cpp | 436 ++--
storage/AggregationOperationState.hpp | 72 +-
storage/CMakeLists.txt | 88 +-
storage/FastHashTable.hpp | 2403 ------------------
storage/FastHashTableFactory.hpp | 224 --
storage/FastSeparateChainingHashTable.hpp | 1551 -----------
storage/HashTableBase.hpp | 43 +-
storage/HashTableFactory.hpp | 37 +-
storage/HashTablePool.hpp | 74 +-
.../PackedPayloadAggregationStateHashTable.cpp | 434 ++++
.../PackedPayloadAggregationStateHashTable.hpp | 721 ++++++
storage/PartitionedHashTablePool.hpp | 50 +-
storage/StorageBlock.cpp | 270 --
storage/StorageBlock.hpp | 150 --
32 files changed, 1707 insertions(+), 6301 deletions(-)
----------------------------------------------------------------------
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationConcreteHandle.cpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationConcreteHandle.cpp b/expressions/aggregation/AggregationConcreteHandle.cpp
index e3fb520..3151a91 100644
--- a/expressions/aggregation/AggregationConcreteHandle.cpp
+++ b/expressions/aggregation/AggregationConcreteHandle.cpp
@@ -19,50 +19,6 @@
#include "expressions/aggregation/AggregationConcreteHandle.hpp"
-#include <cstddef>
-#include <vector>
-
-#include "catalog/CatalogTypedefs.hpp"
-#include "storage/FastHashTable.hpp"
-#include "storage/HashTable.hpp"
-#include "storage/HashTableFactory.hpp"
-
namespace quickstep {
-class StorageManager;
-class Type;
-class ValueAccessor;
-
-AggregationStateHashTableBase* AggregationConcreteHandle::createDistinctifyHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type*> &key_types,
- const std::size_t estimated_num_distinct_keys,
- StorageManager *storage_manager) const {
- // Create a hash table with key types as key_types and value type as bool.
- return AggregationStateHashTableFactory<bool>::CreateResizable(
- hash_table_impl,
- key_types,
- estimated_num_distinct_keys,
- storage_manager);
-}
-
-void AggregationConcreteHandle::insertValueAccessorIntoDistinctifyHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_ids,
- AggregationStateHashTableBase *distinctify_hash_table) const {
- // If the key-value pair is already there, we don't need to update the value,
- // which should always be "true". I.e. the value is just a placeholder.
-
- AggregationStateFastHashTable *hash_table =
- static_cast<AggregationStateFastHashTable *>(distinctify_hash_table);
- if (key_ids.size() == 1) {
- hash_table->upsertValueAccessorFast(
- key_ids, accessor, key_ids[0], true /* check_for_null_keys */);
- } else {
- std::vector<attribute_id> empty_args {kInvalidAttributeID};
- hash_table->upsertValueAccessorCompositeKeyFast(
- empty_args, accessor, key_ids, true /* check_for_null_keys */);
- }
-}
-
} // namespace quickstep
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationConcreteHandle.hpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationConcreteHandle.hpp b/expressions/aggregation/AggregationConcreteHandle.hpp
index 398a032..f1259c0 100644
--- a/expressions/aggregation/AggregationConcreteHandle.hpp
+++ b/expressions/aggregation/AggregationConcreteHandle.hpp
@@ -26,8 +26,6 @@
#include "catalog/CatalogTypedefs.hpp"
#include "expressions/aggregation/AggregationHandle.hpp"
-#include "storage/FastHashTable.hpp"
-#include "storage/HashTable.hpp"
#include "storage/HashTableBase.hpp"
#include "threading/SpinMutex.hpp"
#include "types/TypedValue.hpp"
@@ -51,7 +49,7 @@ class ValueAccessor;
* merging two group by hash tables.
**/
template <typename HandleT>
-class HashTableStateUpserterFast {
+class HashTableStateUpserter {
public:
/**
* @brief Constructor.
@@ -61,7 +59,7 @@ class HashTableStateUpserterFast {
* table. The corresponding state (for the same key) in the destination
* hash table will be upserted.
**/
- HashTableStateUpserterFast(const HandleT &handle,
+ HashTableStateUpserter(const HandleT &handle,
const std::uint8_t *source_state)
: handle_(handle), source_state_(source_state) {}
@@ -72,14 +70,14 @@ class HashTableStateUpserterFast {
* table that is being upserted.
**/
void operator()(std::uint8_t *destination_state) {
- handle_.mergeStatesFast(source_state_, destination_state);
+ handle_.mergeStates(source_state_, destination_state);
}
private:
const HandleT &handle_;
const std::uint8_t *source_state_;
- DISALLOW_COPY_AND_ASSIGN(HashTableStateUpserterFast);
+ DISALLOW_COPY_AND_ASSIGN(HashTableStateUpserter);
};
/**
@@ -102,50 +100,18 @@ class AggregationConcreteHandle : public AggregationHandle {
<< "takes at least one argument.";
}
- /**
- * @brief Implementaion for AggregationHandle::createDistinctifyHashTable()
- * that creates a new HashTable for the distinctify step for
- * DISTINCT aggregation.
- */
- AggregationStateHashTableBase* createDistinctifyHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &key_types,
- const std::size_t estimated_num_distinct_keys,
- StorageManager *storage_manager) const override;
-
- /**
- * @brief Implementaion for
- * AggregationHandle::insertValueAccessorIntoDistinctifyHashTable()
- * that inserts the GROUP BY expressions and aggregation arguments together
- * as keys into the distinctify hash table.
- */
- void insertValueAccessorIntoDistinctifyHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_ids,
- AggregationStateHashTableBase *distinctify_hash_table) const override;
-
protected:
AggregationConcreteHandle() {}
- template <typename HandleT, typename StateT>
- StateT* aggregateOnDistinctifyHashTableForSingleUnaryHelperFast(
- const AggregationStateHashTableBase &distinctify_hash_table) const;
-
template <typename HandleT, typename HashTableT>
- void aggregateOnDistinctifyHashTableForGroupByUnaryHelperFast(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *hash_table,
- std::size_t index) const;
-
- template <typename HandleT, typename HashTableT>
- ColumnVector* finalizeHashTableHelperFast(
+ ColumnVector* finalizeHashTableHelper(
const Type &result_type,
const AggregationStateHashTableBase &hash_table,
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const;
template <typename HandleT, typename HashTableT>
- inline TypedValue finalizeGroupInHashTableFast(
+ inline TypedValue finalizeGroupInHashTable(
const AggregationStateHashTableBase &hash_table,
const std::vector<TypedValue> &group_key,
int index) const {
@@ -153,15 +119,10 @@ class AggregationConcreteHandle : public AggregationHandle {
static_cast<const HashTableT &>(hash_table).getSingleCompositeKey(group_key, index);
DCHECK(group_state != nullptr)
<< "Could not find entry for specified group_key in HashTable";
- return static_cast<const HandleT *>(this)->finalizeHashTableEntryFast(
+ return static_cast<const HandleT *>(this)->finalizeHashTableEntry(
group_state);
}
- template <typename HandleT, typename HashTableT>
- void mergeGroupByHashTablesHelperFast(
- const AggregationStateHashTableBase &source_hash_table,
- AggregationStateHashTableBase *destination_hash_table) const;
-
private:
DISALLOW_COPY_AND_ASSIGN(AggregationConcreteHandle);
};
@@ -195,7 +156,7 @@ class HashTableAggregateFinalizer {
const unsigned char *byte_ptr) {
group_by_keys_->emplace_back(group_by_key);
output_column_vector_->appendTypedValue(
- handle_.finalizeHashTableEntryFast(byte_ptr));
+ handle_.finalizeHashTableEntry(byte_ptr));
}
private:
@@ -209,70 +170,8 @@ class HashTableAggregateFinalizer {
// ----------------------------------------------------------------------------
// Implementations of templated methods follow:
-template <typename HandleT, typename StateT>
-StateT* AggregationConcreteHandle::
- aggregateOnDistinctifyHashTableForSingleUnaryHelperFast(
- const AggregationStateHashTableBase &distinctify_hash_table) const {
- const HandleT &handle = static_cast<const HandleT &>(*this);
- StateT *state = static_cast<StateT *>(createInitialState());
-
- // A lambda function which will be called on each key from the distinctify
- // hash table.
- const auto aggregate_functor = [&handle, &state](
- const TypedValue &key, const std::uint8_t &dumb_placeholder) {
- // For each (unary) key in the distinctify hash table, aggregate the key
- // into "state".
- handle.iterateUnaryInl(state, key);
- };
-
- const AggregationStateFastHashTable &hash_table =
- static_cast<const AggregationStateFastHashTable &>(
- distinctify_hash_table);
- // Invoke the lambda function "aggregate_functor" on each key from the
- // distinctify hash table.
- hash_table.forEach(&aggregate_functor);
-
- return state;
-}
-
-template <typename HandleT, typename HashTableT>
-void AggregationConcreteHandle::
- aggregateOnDistinctifyHashTableForGroupByUnaryHelperFast(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const {
- const HandleT &handle = static_cast<const HandleT &>(*this);
- HashTableT *target_hash_table =
- static_cast<HashTableT *>(aggregation_hash_table);
-
- // A lambda function which will be called on each key-value pair from the
- // distinctify hash table.
- const auto aggregate_functor = [&handle, &target_hash_table, &index](
- std::vector<TypedValue> &key, const bool &dumb_placeholder) {
- // For each (composite) key vector in the distinctify hash table with size N.
- // The first N-1 entries are GROUP BY columns and the last entry is the
- // argument to be aggregated on.
- const TypedValue argument(std::move(key.back()));
- key.pop_back();
-
- // An upserter as lambda function for aggregating the argument into its
- // GROUP BY group's entry inside aggregation_hash_table.
- const auto upserter = [&handle, &argument](std::uint8_t *state) {
- handle.iterateUnaryInlFast(argument, state);
- };
-
- target_hash_table->upsertCompositeKeyFast(key, nullptr, &upserter, index);
- };
-
- const HashTableT &source_hash_table =
- static_cast<const HashTableT &>(distinctify_hash_table);
- // Invoke the lambda function "aggregate_functor" on each composite key vector
- // from the distinctify hash table.
- source_hash_table.forEachCompositeKeyFast(&aggregate_functor);
-}
-
template <typename HandleT, typename HashTableT>
-ColumnVector* AggregationConcreteHandle::finalizeHashTableHelperFast(
+ColumnVector* AggregationConcreteHandle::finalizeHashTableHelper(
const Type &result_type,
const AggregationStateHashTableBase &hash_table,
std::vector<std::vector<TypedValue>> *group_by_keys,
@@ -287,14 +186,14 @@ ColumnVector* AggregationConcreteHandle::finalizeHashTableHelperFast(
new NativeColumnVector(result_type, hash_table_concrete.numEntries());
HashTableAggregateFinalizer<HandleT, NativeColumnVector> finalizer(
handle, group_by_keys, result);
- hash_table_concrete.forEachCompositeKeyFast(&finalizer, index);
+ hash_table_concrete.forEach(&finalizer, index);
return result;
} else {
IndirectColumnVector *result = new IndirectColumnVector(
result_type, hash_table_concrete.numEntries());
HashTableAggregateFinalizer<HandleT, IndirectColumnVector> finalizer(
handle, group_by_keys, result);
- hash_table_concrete.forEachCompositeKeyFast(&finalizer, index);
+ hash_table_concrete.forEach(&finalizer, index);
return result;
}
} else {
@@ -303,7 +202,7 @@ ColumnVector* AggregationConcreteHandle::finalizeHashTableHelperFast(
new NativeColumnVector(result_type, group_by_keys->size());
for (const std::vector<TypedValue> &group_by_key : *group_by_keys) {
result->appendTypedValue(
- finalizeGroupInHashTableFast<HandleT, HashTableT>(
+ finalizeGroupInHashTable<HandleT, HashTableT>(
hash_table, group_by_key, index));
}
return result;
@@ -312,7 +211,7 @@ ColumnVector* AggregationConcreteHandle::finalizeHashTableHelperFast(
result_type, hash_table_concrete.numEntries());
for (const std::vector<TypedValue> &group_by_key : *group_by_keys) {
result->appendTypedValue(
- finalizeGroupInHashTableFast<HandleT, HashTableT>(
+ finalizeGroupInHashTable<HandleT, HashTableT>(
hash_table, group_by_key, index));
}
return result;
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandle.hpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandle.hpp b/expressions/aggregation/AggregationHandle.hpp
index 4b51179..e004511 100644
--- a/expressions/aggregation/AggregationHandle.hpp
+++ b/expressions/aggregation/AggregationHandle.hpp
@@ -32,6 +32,7 @@
namespace quickstep {
class ColumnVector;
+class ColumnVectorsValueAccessor;
class StorageManager;
class Type;
class ValueAccessor;
@@ -116,29 +117,6 @@ class AggregationHandle {
virtual AggregationState* createInitialState() const = 0;
/**
- * @brief Create a new HashTable for aggregation with GROUP BY.
- *
- * @param hash_table_impl The choice of which concrete HashTable
- * implementation to use.
- * @param group_by_types The types of the GROUP BY columns/expressions. These
- * correspond to the (composite) key type for the HashTable.
- * @param estimated_num_groups The estimated number of distinct groups for
- * the GROUP BY aggregation. This is used to size the initial
- * HashTable. This is an estimate only, and the HashTable will be
- * resized if it becomes over-full.
- * @param storage_manager The StorageManager to use to create the HashTable.
- * A StorageBlob will be allocated to serve as the HashTable's
- * in-memory storage.
- * @return A new HashTable instance with the appropriate state type for this
- * aggregate.
- **/
- virtual AggregationStateHashTableBase* createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const = 0;
-
- /**
* @brief Accumulate over tuples for a nullary aggregate function (one that
* has zero arguments, i.e. COUNT(*)).
*
@@ -153,63 +131,16 @@ class AggregationHandle {
const std::size_t num_tuples) const = 0;
/**
- * @brief Accumulate (iterate over) all values in one or more ColumnVectors
- * and return a new AggregationState which can be merged with other
- * states or finalized.
+ * @brief TODO
*
- * @param column_vectors One or more ColumnVectors that the aggregate will be
- * applied to. These correspond to the aggregate function's arguments,
- * in order.
* @return A new AggregationState which contains the accumulated results from
* applying the aggregate to column_vectors. Caller is responsible
* for deleting the returned AggregationState.
**/
- virtual AggregationState* accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors) const = 0;
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- /**
- * @brief Accumulate (iterate over) all values in columns accessible through
- * a ValueAccessor and return a new AggregationState which can be
- * merged with other states or finalized.
- *
- * @param accessor A ValueAccessor that the columns to be aggregated can be
- * accessed through.
- * @param accessor_ids The attribute_ids that correspond to the columns in
- * accessor to aggeregate. These correspond to the aggregate
- * function's arguments, in order.
- * @return A new AggregationState which contains the accumulated results from
- * applying the aggregate to the specified columns in accessor.
- * Caller is responsible for deleting the returned AggregationState.
- **/
- virtual AggregationState* accumulateValueAccessor(
+ virtual AggregationState* accumulate(
ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const = 0;
-#endif
-
- /**
- * @brief Perform an aggregation with GROUP BY over all the tuples accessible
- * through a ValueAccessor, upserting states in a HashTable.
- *
- * @note Implementations of this method are threadsafe with respect to
- * hash_table, and can be called concurrently from multiple threads
- * with the same HashTable object.
- *
- * @param accessor The ValueAccessor that will be iterated over to read
- * tuples.
- * @param argument_ids The attribute_ids of the arguments to this aggregate
- * in accessor, in order.
- * @param group_by_key_ids The attribute_ids of the group-by
- * columns/expressions in accessor.
- * @param hash_table The HashTable to upsert AggregationStates in. This
- * should have been created by calling createGroupByHashTable() on
- * this same AggregationHandle.
- **/
- virtual void aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const = 0;
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const = 0;
/**
* @brief Merge two AggregationStates, updating one in-place. This computes a
@@ -269,99 +200,12 @@ class AggregationHandle {
int index) const = 0;
/**
- * @brief Create a new HashTable for the distinctify step for DISTINCT
- * aggregation.
- *
- * Distinctify is the first step for DISTINCT aggregation. This step inserts
- * the GROUP BY expression values and aggregation arguments together as keys
- * into the distinctify hash table, so that arguments are distinctified within
- * each GROUP BY group. Later, a second-round aggregation on the distinctify
- * hash table will be performed to actually compute the aggregated result for
- * each GROUP BY group.
- *
- * In the case of single aggregation where there is no GROUP BY expressions,
- * we simply treat it as a special GROUP BY case that the GROUP BY expression
- * vector is empty.
- *
- * @param hash_table_impl The choice of which concrete HashTable
- * implementation to use.
- * @param key_types The types of the GROUP BY expressions together with the
- * types of the aggregation arguments.
- * @param estimated_num_distinct_keys The estimated number of distinct keys
- * (i.e. GROUP BY expressions together with aggregation arguments) for
- * the distinctify step. This is used to size the initial HashTable.
- * This is an estimate only, and the HashTable will be resized if it
- * becomes over-full.
- * @param storage_manager The StorageManager to use to create the HashTable.
- * A StorageBlob will be allocated to serve as the HashTable's
- * in-memory storage.
- *
- * @return A new HashTable instance with the appropriate state type for this
- * aggregate.
- */
- virtual AggregationStateHashTableBase* createDistinctifyHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &key_types,
- const std::size_t estimated_num_distinct_keys,
- StorageManager *storage_manager) const = 0;
-
- /**
- * @brief Inserts the GROUP BY expressions and aggregation arguments together
- * as keys into the distinctify hash table.
- *
- * @param accessor The ValueAccessor that will be iterated over to read
- * tuples.
- * @param key_ids The attribute_ids of the GROUP BY expressions in accessor
- * together with the attribute_ids of the arguments to this aggregate
- * in accessor, in order.
- * @param distinctify_hash_table The HashTable to store the GROUP BY
- * expressions and the aggregation arguments together as hash table
- * keys and a bool constant \c true as hash table value (So the hash
- * table actually serves as a hash set). This should have been created
- * by calling createDistinctifyHashTable();
- */
- virtual void insertValueAccessorIntoDistinctifyHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_ids,
- AggregationStateHashTableBase *distinctify_hash_table) const = 0;
-
- /**
- * @brief Perform single (i.e. without GROUP BY) aggregation on the keys from
- * the distinctify hash table to actually compute the aggregated results.
- *
- * @param distinctify_hash_table Hash table which stores the distinctified
- * aggregation arguments as hash table keys. This should have been
- * created by calling createDistinctifyHashTable();
- * @return A new AggregationState which contains the aggregated results from
- * applying the aggregate to the distinctify hash table.
- * Caller is responsible for deleting the returned AggregationState.
- */
- virtual AggregationState* aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table) const = 0;
-
- /**
- * @brief Perform GROUP BY aggregation on the keys from the distinctify hash
- * table and upserts states into the aggregation hash table.
- *
- * @param distinctify_hash_table Hash table which stores the GROUP BY
- * expression values and aggregation arguments together as hash table
- * keys.
- * @param aggregation_hash_table The HashTable to upsert AggregationStates in.
- * This should have been created by calling createGroupByHashTable() on
- * this same AggregationHandle.
- * @param index The index of the distinctify hash table for which we perform
- * the DISTINCT aggregation.
- */
- virtual void aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const = 0;
-
- /**
* @brief Get the number of bytes needed to store the aggregation handle's
* state.
**/
- virtual std::size_t getPayloadSize() const { return 1; }
+ virtual std::size_t getPayloadSize() const {
+ return 1u;
+ }
/**
* @brief Update the aggregation state for nullary aggregation function e.g.
@@ -394,8 +238,8 @@ class AggregationHandle {
* @param src A pointer to the source aggregation state.
* @param dst A pointer to the destination aggregation state.
**/
- virtual void mergeStatesFast(const std::uint8_t *src,
- std::uint8_t *dst) const {}
+ virtual void mergeStates(const std::uint8_t *src,
+ std::uint8_t *dst) const {}
/**
* @brief Initialize the payload (in the aggregation hash table) for the given
@@ -413,18 +257,6 @@ class AggregationHandle {
**/
virtual void destroyPayload(std::uint8_t *byte_ptr) const {}
- /**
- * @brief Inform the aggregation handle to block (prohibit) updates on the
- * aggregation state.
- **/
- virtual void blockUpdate() {}
-
- /**
- * @brief Inform the aggregation handle to allow updates on the
- * aggregation state.
- **/
- virtual void allowUpdate() {}
-
protected:
AggregationHandle() {}
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleAvg.cpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleAvg.cpp b/expressions/aggregation/AggregationHandleAvg.cpp
index 2481092..3d2db8c 100644
--- a/expressions/aggregation/AggregationHandleAvg.cpp
+++ b/expressions/aggregation/AggregationHandleAvg.cpp
@@ -24,8 +24,7 @@
#include <vector>
#include "catalog/CatalogTypedefs.hpp"
-#include "storage/HashTable.hpp"
-#include "storage/HashTableFactory.hpp"
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
#include "threading/SpinMutex.hpp"
#include "types/Type.hpp"
#include "types/TypeFactory.hpp"
@@ -42,7 +41,7 @@ namespace quickstep {
class StorageManager;
AggregationHandleAvg::AggregationHandleAvg(const Type &type)
- : argument_type_(type), block_update_(false) {
+ : argument_type_(type) {
// We sum Int as Long and Float as Double so that we have more headroom when
// adding many values.
TypeID type_precision_id;
@@ -87,52 +86,28 @@ AggregationHandleAvg::AggregationHandleAvg(const Type &type)
->getNullableVersion());
}
-AggregationStateHashTableBase* AggregationHandleAvg::createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const {
- return AggregationStateHashTableFactory<AggregationStateAvg>::CreateResizable(
- hash_table_impl, group_by_types, estimated_num_groups, storage_manager);
-}
-
-AggregationState* AggregationHandleAvg::accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors) const {
- DCHECK_EQ(1u, column_vectors.size())
- << "Got wrong number of ColumnVectors for AVG: " << column_vectors.size();
+AggregationState* AggregationHandleAvg::accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const {
+ DCHECK_EQ(1u, argument_ids.size())
+ << "Got wrong number of attributes for AVG: " << argument_ids.size();
- AggregationStateAvg *state = new AggregationStateAvg(blank_state_);
- std::size_t count = 0;
- state->sum_ = fast_add_operator_->accumulateColumnVector(
- state->sum_, *column_vectors.front(), &count);
- state->count_ = count;
- return state;
-}
+ const attribute_id argument_id = argument_ids.front();
+ DCHECK_NE(argument_id, kInvalidAttributeID);
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-AggregationState* AggregationHandleAvg::accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const {
- DCHECK_EQ(1u, accessor_ids.size())
- << "Got wrong number of attributes for AVG: " << accessor_ids.size();
+ ValueAccessor *target_accessor =
+ argument_id >= 0 ? accessor : aux_accessor;
+ const attribute_id target_argument_id =
+ argument_id >= 0 ? argument_id : -(argument_id+2);
AggregationStateAvg *state = new AggregationStateAvg(blank_state_);
std::size_t count = 0;
state->sum_ = fast_add_operator_->accumulateValueAccessor(
- state->sum_, accessor, accessor_ids.front(), &count);
+ state->sum_, target_accessor, target_argument_id, &count);
state->count_ = count;
return state;
}
-#endif
-
-void AggregationHandleAvg::aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const {
- DCHECK_EQ(1u, argument_ids.size())
- << "Got wrong number of arguments for AVG: " << argument_ids.size();
-}
void AggregationHandleAvg::mergeStates(const AggregationState &source,
AggregationState *destination) const {
@@ -147,8 +122,8 @@ void AggregationHandleAvg::mergeStates(const AggregationState &source,
avg_destination->sum_, avg_source.sum_);
}
-void AggregationHandleAvg::mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const {
+void AggregationHandleAvg::mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const {
const TypedValue *src_sum_ptr =
reinterpret_cast<const TypedValue *>(source + blank_state_.sum_offset_);
const std::int64_t *src_count_ptr = reinterpret_cast<const std::int64_t *>(
@@ -179,27 +154,10 @@ ColumnVector* AggregationHandleAvg::finalizeHashTable(
const AggregationStateHashTableBase &hash_table,
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const {
- return finalizeHashTableHelperFast<AggregationHandleAvg,
- AggregationStateFastHashTable>(
- *result_type_, hash_table, group_by_keys, index);
-}
-
-AggregationState*
-AggregationHandleAvg::aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table) const {
- return aggregateOnDistinctifyHashTableForSingleUnaryHelperFast<
- AggregationHandleAvg,
- AggregationStateAvg>(distinctify_hash_table);
-}
-
-void AggregationHandleAvg::aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const {
- aggregateOnDistinctifyHashTableForGroupByUnaryHelperFast<
+ return finalizeHashTableHelper<
AggregationHandleAvg,
- AggregationStateFastHashTable>(
- distinctify_hash_table, aggregation_hash_table, index);
+ PackedPayloadSeparateChainingAggregationStateHashTable>(
+ *result_type_, hash_table, group_by_keys, index);
}
} // namespace quickstep
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleAvg.hpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleAvg.hpp b/expressions/aggregation/AggregationHandleAvg.hpp
index 47132c6..b40305c 100644
--- a/expressions/aggregation/AggregationHandleAvg.hpp
+++ b/expressions/aggregation/AggregationHandleAvg.hpp
@@ -28,7 +28,6 @@
#include "catalog/CatalogTypedefs.hpp"
#include "expressions/aggregation/AggregationConcreteHandle.hpp"
#include "expressions/aggregation/AggregationHandle.hpp"
-#include "storage/FastHashTable.hpp"
#include "storage/HashTableBase.hpp"
#include "threading/SpinMutex.hpp"
#include "types/Type.hpp"
@@ -110,12 +109,6 @@ class AggregationHandleAvg : public AggregationConcreteHandle {
return new AggregationStateAvg(blank_state_);
}
- AggregationStateHashTableBase* createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const override;
-
/**
* @brief Iterate method with average aggregation state.
**/
@@ -129,28 +122,19 @@ class AggregationHandleAvg : public AggregationConcreteHandle {
++state->count_;
}
- inline void iterateUnaryInlFast(const TypedValue &value,
- std::uint8_t *byte_ptr) const {
- DCHECK(value.isPlausibleInstanceOf(argument_type_.getSignature()));
- if (value.isNull()) return;
- TypedValue *sum_ptr =
- reinterpret_cast<TypedValue *>(byte_ptr + blank_state_.sum_offset_);
- std::int64_t *count_ptr =
- reinterpret_cast<std::int64_t *>(byte_ptr + blank_state_.count_offset_);
- *sum_ptr = fast_add_operator_->applyToTypedValues(*sum_ptr, value);
- ++(*count_ptr);
- }
+ AggregationState* accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const override;
- inline void updateStateUnary(const TypedValue &argument,
- std::uint8_t *byte_ptr) const override {
- if (!block_update_) {
- iterateUnaryInlFast(argument, byte_ptr);
- }
- }
+ void mergeStates(const AggregationState &source,
+ AggregationState *destination) const override;
- void blockUpdate() override { block_update_ = true; }
+ TypedValue finalize(const AggregationState &state) const override;
- void allowUpdate() override { block_update_ = false; }
+ std::size_t getPayloadSize() const override {
+ return blank_state_.getPayloadSize();
+ }
void initPayload(std::uint8_t *byte_ptr) const override {
TypedValue *sum_ptr =
@@ -169,43 +153,22 @@ class AggregationHandleAvg : public AggregationConcreteHandle {
}
}
- AggregationState* accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors)
- const override;
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- AggregationState* accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_id) const override;
-#endif
-
- void aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const override;
-
- void mergeStates(const AggregationState &source,
- AggregationState *destination) const override;
-
- void mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const override;
+ inline void updateStateUnary(const TypedValue &argument,
+ std::uint8_t *byte_ptr) const override {
+ DCHECK(argument.isPlausibleInstanceOf(argument_type_.getSignature()));
+ if (argument.isNull()) return;
+ TypedValue *sum_ptr =
+ reinterpret_cast<TypedValue *>(byte_ptr + blank_state_.sum_offset_);
+ std::int64_t *count_ptr =
+ reinterpret_cast<std::int64_t *>(byte_ptr + blank_state_.count_offset_);
+ *sum_ptr = fast_add_operator_->applyToTypedValues(*sum_ptr, argument);
+ ++(*count_ptr);
+ }
- TypedValue finalize(const AggregationState &state) const override;
+ void mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const override;
inline TypedValue finalizeHashTableEntry(
- const AggregationState &state) const {
- const AggregationStateAvg &agg_state =
- static_cast<const AggregationStateAvg &>(state);
- // TODO(chasseur): Could improve performance further if we made a special
- // version of finalizeHashTable() that collects all the sums into one
- // ColumnVector and all the counts into another and then applies
- // '*divide_operator_' to them in bulk.
- return divide_operator_->applyToTypedValues(
- agg_state.sum_, TypedValue(static_cast<double>(agg_state.count_)));
- }
-
- inline TypedValue finalizeHashTableEntryFast(
const std::uint8_t *byte_ptr) const {
std::uint8_t *value_ptr = const_cast<std::uint8_t *>(byte_ptr);
TypedValue *sum_ptr =
@@ -221,29 +184,6 @@ class AggregationHandleAvg : public AggregationConcreteHandle {
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const override;
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForSingle()
- * for AVG aggregation.
- */
- AggregationState* aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table)
- const override;
-
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForGroupBy()
- * for AVG aggregation.
- */
- void aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const override;
-
- std::size_t getPayloadSize() const override {
- return blank_state_.getPayloadSize();
- }
-
private:
friend class AggregateFunctionAvg;
@@ -261,8 +201,6 @@ class AggregationHandleAvg : public AggregationConcreteHandle {
std::unique_ptr<UncheckedBinaryOperator> merge_add_operator_;
std::unique_ptr<UncheckedBinaryOperator> divide_operator_;
- bool block_update_;
-
DISALLOW_COPY_AND_ASSIGN(AggregationHandleAvg);
};
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleCount.cpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleCount.cpp b/expressions/aggregation/AggregationHandleCount.cpp
index 034c942..a5c9fd8 100644
--- a/expressions/aggregation/AggregationHandleCount.cpp
+++ b/expressions/aggregation/AggregationHandleCount.cpp
@@ -25,14 +25,9 @@
#include <vector>
#include "catalog/CatalogTypedefs.hpp"
-#include "storage/HashTable.hpp"
-#include "storage/HashTableFactory.hpp"
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
#include "storage/ValueAccessor.hpp"
#include "storage/ValueAccessorUtil.hpp"
-#endif
-
#include "types/TypeFactory.hpp"
#include "types/TypeID.hpp"
#include "types/TypedValue.hpp"
@@ -48,73 +43,32 @@ class Type;
class ValueAccessor;
template <bool count_star, bool nullable_type>
-AggregationStateHashTableBase*
-AggregationHandleCount<count_star, nullable_type>::createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const {
- return AggregationStateHashTableFactory<
- AggregationStateCount>::CreateResizable(hash_table_impl,
- group_by_types,
- estimated_num_groups,
- storage_manager);
-}
-
-template <bool count_star, bool nullable_type>
-AggregationState*
-AggregationHandleCount<count_star, nullable_type>::accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors) const {
+AggregationState* AggregationHandleCount<count_star, nullable_type>::accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const {
DCHECK(!count_star)
<< "Called non-nullary accumulation method on an AggregationHandleCount "
<< "set up for nullary COUNT(*)";
- DCHECK_EQ(1u, column_vectors.size())
- << "Got wrong number of ColumnVectors for COUNT: "
- << column_vectors.size();
+ DCHECK_EQ(1u, argument_ids.size())
+ << "Got wrong number of attributes for COUNT: " << argument_ids.size();
- std::size_t count = 0;
- InvokeOnColumnVector(
- *column_vectors.front(),
- [&](const auto &column_vector) -> void { // NOLINT(build/c++11)
- if (nullable_type) {
- // TODO(shoban): Iterating over the ColumnVector is a rather slow way
- // to do this. We should look at extending the ColumnVector interface
- // to do a quick count of the non-null values (i.e. the length minus
- // the population count of the null bitmap). We should do something
- // similar for ValueAccessor too.
- for (std::size_t pos = 0; pos < column_vector.size(); ++pos) {
- count += !column_vector.getTypedValue(pos).isNull();
- }
- } else {
- count = column_vector.size();
- }
- });
+ const attribute_id argument_id = argument_ids.front();
+ DCHECK_NE(argument_id, kInvalidAttributeID);
- return new AggregationStateCount(count);
-}
+ ValueAccessor *target_accessor =
+ argument_id >= 0 ? accessor : aux_accessor;
+ const attribute_id target_argument_id =
+ argument_id >= 0 ? argument_id : -(argument_id+2);
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-template <bool count_star, bool nullable_type>
-AggregationState*
-AggregationHandleCount<count_star, nullable_type>::accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const {
- DCHECK(!count_star)
- << "Called non-nullary accumulation method on an AggregationHandleCount "
- << "set up for nullary COUNT(*)";
-
- DCHECK_EQ(1u, accessor_ids.size())
- << "Got wrong number of attributes for COUNT: " << accessor_ids.size();
-
- const attribute_id accessor_id = accessor_ids.front();
std::size_t count = 0;
InvokeOnValueAccessorMaybeTupleIdSequenceAdapter(
- accessor,
- [&accessor_id, &count](auto *accessor) -> void { // NOLINT(build/c++11)
+ target_accessor,
+ [&target_argument_id, &count](auto *accessor) -> void { // NOLINT(build/c++11)
if (nullable_type) {
while (accessor->next()) {
- count += !accessor->getTypedValue(accessor_id).isNull();
+ count += !accessor->getTypedValue(target_argument_id).isNull();
}
} else {
count = accessor->getNumTuples();
@@ -123,24 +77,6 @@ AggregationHandleCount<count_star, nullable_type>::accumulateValueAccessor(
return new AggregationStateCount(count);
}
-#endif
-
-template <bool count_star, bool nullable_type>
-void AggregationHandleCount<count_star, nullable_type>::
- aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const {
- if (count_star) {
- DCHECK_EQ(0u, argument_ids.size())
- << "Got wrong number of arguments for COUNT(*): "
- << argument_ids.size();
- } else {
- DCHECK_EQ(1u, argument_ids.size())
- << "Got wrong number of arguments for COUNT: " << argument_ids.size();
- }
-}
template <bool count_star, bool nullable_type>
void AggregationHandleCount<count_star, nullable_type>::mergeStates(
@@ -156,7 +92,7 @@ void AggregationHandleCount<count_star, nullable_type>::mergeStates(
}
template <bool count_star, bool nullable_type>
-void AggregationHandleCount<count_star, nullable_type>::mergeStatesFast(
+void AggregationHandleCount<count_star, nullable_type>::mergeStates(
const std::uint8_t *source, std::uint8_t *destination) const {
const std::int64_t *src_count_ptr =
reinterpret_cast<const std::int64_t *>(source);
@@ -170,33 +106,10 @@ AggregationHandleCount<count_star, nullable_type>::finalizeHashTable(
const AggregationStateHashTableBase &hash_table,
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const {
- return finalizeHashTableHelperFast<
- AggregationHandleCount<count_star, nullable_type>,
- AggregationStateFastHashTable>(
- TypeFactory::GetType(kLong), hash_table, group_by_keys, index);
-}
-
-template <bool count_star, bool nullable_type>
-AggregationState* AggregationHandleCount<count_star, nullable_type>::
- aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table) const {
- DCHECK_EQ(count_star, false);
- return aggregateOnDistinctifyHashTableForSingleUnaryHelperFast<
- AggregationHandleCount<count_star, nullable_type>,
- AggregationStateCount>(distinctify_hash_table);
-}
-
-template <bool count_star, bool nullable_type>
-void AggregationHandleCount<count_star, nullable_type>::
- aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const {
- DCHECK_EQ(count_star, false);
- aggregateOnDistinctifyHashTableForGroupByUnaryHelperFast<
+ return finalizeHashTableHelper<
AggregationHandleCount<count_star, nullable_type>,
- AggregationStateFastHashTable>(
- distinctify_hash_table, aggregation_hash_table, index);
+ PackedPayloadSeparateChainingAggregationStateHashTable>(
+ TypeFactory::GetType(kLong), hash_table, group_by_keys, index);
}
// Explicitly instantiate and compile in the different versions of
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleCount.hpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleCount.hpp b/expressions/aggregation/AggregationHandleCount.hpp
index 6aab0cd..86da2a8 100644
--- a/expressions/aggregation/AggregationHandleCount.hpp
+++ b/expressions/aggregation/AggregationHandleCount.hpp
@@ -29,7 +29,6 @@
#include "catalog/CatalogTypedefs.hpp"
#include "expressions/aggregation/AggregationConcreteHandle.hpp"
#include "expressions/aggregation/AggregationHandle.hpp"
-#include "storage/FastHashTable.hpp"
#include "storage/HashTableBase.hpp"
#include "types/TypedValue.hpp"
#include "utility/Macros.hpp"
@@ -102,21 +101,10 @@ class AggregationHandleCount : public AggregationConcreteHandle {
return new AggregationStateCount();
}
- AggregationStateHashTableBase* createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const override;
-
inline void iterateNullaryInl(AggregationStateCount *state) const {
state->count_.fetch_add(1, std::memory_order_relaxed);
}
- inline void iterateNullaryInlFast(std::uint8_t *byte_ptr) const {
- std::int64_t *count_ptr = reinterpret_cast<std::int64_t *>(byte_ptr);
- (*count_ptr)++;
- }
-
/**
* @brief Iterate with count aggregation state.
*/
@@ -127,81 +115,50 @@ class AggregationHandleCount : public AggregationConcreteHandle {
}
}
- inline void iterateUnaryInlFast(const TypedValue &value,
- std::uint8_t *byte_ptr) const {
- if ((!nullable_type) || (!value.isNull())) {
- std::int64_t *count_ptr = reinterpret_cast<std::int64_t *>(byte_ptr);
- (*count_ptr)++;
- }
- }
-
- inline void updateStateUnary(const TypedValue &argument,
- std::uint8_t *byte_ptr) const override {
- if (!block_update_) {
- iterateUnaryInlFast(argument, byte_ptr);
- }
- }
-
- inline void updateStateNullary(std::uint8_t *byte_ptr) const override {
- if (!block_update_) {
- iterateNullaryInlFast(byte_ptr);
- }
- }
-
- void blockUpdate() override { block_update_ = true; }
-
- void allowUpdate() override { block_update_ = false; }
-
- void initPayload(std::uint8_t *byte_ptr) const override {
- std::int64_t *count_ptr = reinterpret_cast<std::int64_t *>(byte_ptr);
- *count_ptr = 0;
- }
-
AggregationState* accumulateNullary(
const std::size_t num_tuples) const override {
return new AggregationStateCount(num_tuples);
}
- AggregationState* accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors)
- const override;
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- AggregationState* accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_id) const override;
-#endif
-
- void aggregateValueAccessorIntoHashTable(
+ AggregationState* accumulate(
ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const override;
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const override;
void mergeStates(const AggregationState &source,
AggregationState *destination) const override;
- void mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const override;
-
TypedValue finalize(const AggregationState &state) const override {
return TypedValue(
static_cast<const AggregationStateCount &>(state).count_.load(
std::memory_order_relaxed));
}
- inline TypedValue finalizeHashTableEntry(
- const AggregationState &state) const {
- return TypedValue(
- static_cast<const AggregationStateCount &>(state).count_.load(
- std::memory_order_relaxed));
+ std::size_t getPayloadSize() const override {
+ return sizeof(std::int64_t);
+ }
+
+ void initPayload(std::uint8_t *byte_ptr) const override {
+ std::int64_t *count_ptr = reinterpret_cast<std::int64_t *>(byte_ptr);
+ *count_ptr = 0;
+ }
+
+ inline void updateStateNullary(std::uint8_t *byte_ptr) const override {
+ ++(*reinterpret_cast<std::int64_t *>(byte_ptr));
+ }
+
+ inline void updateStateUnary(const TypedValue &argument,
+ std::uint8_t *byte_ptr) const override {
+ if ((!nullable_type) || (!argument.isNull())) {
+ ++(*reinterpret_cast<std::int64_t *>(byte_ptr));
+ }
}
- inline TypedValue finalizeHashTableEntryFast(
- const std::uint8_t *byte_ptr) const {
- const std::int64_t *count_ptr =
- reinterpret_cast<const std::int64_t *>(byte_ptr);
- return TypedValue(*count_ptr);
+ void mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const override;
+
+ inline TypedValue finalizeHashTableEntry(const std::uint8_t *byte_ptr) const {
+ return TypedValue(*reinterpret_cast<const std::int64_t *>(byte_ptr));
}
ColumnVector* finalizeHashTable(
@@ -209,36 +166,13 @@ class AggregationHandleCount : public AggregationConcreteHandle {
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const override;
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForSingle()
- * for SUM aggregation.
- */
- AggregationState* aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table)
- const override;
-
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForGroupBy()
- * for SUM aggregation.
- */
- void aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const override;
-
- std::size_t getPayloadSize() const override { return sizeof(std::int64_t); }
-
private:
friend class AggregateFunctionCount;
/**
* @brief Constructor.
**/
- AggregationHandleCount() : block_update_(false) {}
-
- bool block_update_;
+ AggregationHandleCount() {}
DISALLOW_COPY_AND_ASSIGN(AggregationHandleCount);
};
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleDistinct.cpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleDistinct.cpp b/expressions/aggregation/AggregationHandleDistinct.cpp
index 0dc8b56..c6c47c7 100644
--- a/expressions/aggregation/AggregationHandleDistinct.cpp
+++ b/expressions/aggregation/AggregationHandleDistinct.cpp
@@ -22,10 +22,9 @@
#include <cstddef>
#include <memory>
#include <vector>
-#include <utility>
#include "catalog/CatalogTypedefs.hpp"
-#include "storage/HashTable.hpp"
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
#include "types/TypedValue.hpp"
@@ -34,34 +33,6 @@
namespace quickstep {
class ColumnVector;
-class StorageManager;
-class Type;
-class ValueAccessor;
-
-AggregationStateHashTableBase* AggregationHandleDistinct::createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type*> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const {
- return createDistinctifyHashTable(
- hash_table_impl,
- group_by_types,
- estimated_num_groups,
- storage_manager);
-}
-
-void AggregationHandleDistinct::aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const {
- DCHECK_EQ(argument_ids.size(), 0u);
-
- insertValueAccessorIntoDistinctifyHashTable(
- accessor,
- group_by_key_ids,
- hash_table);
-}
ColumnVector* AggregationHandleDistinct::finalizeHashTable(
const AggregationStateHashTableBase &hash_table,
@@ -73,7 +44,8 @@ ColumnVector* AggregationHandleDistinct::finalizeHashTable(
const bool &dumb_placeholder) -> void {
group_by_keys->emplace_back(std::move(group_by_key));
};
- static_cast<const AggregationStateFastHashTable&>(hash_table).forEachCompositeKeyFast(&keys_retriever);
+ static_cast<const PackedPayloadSeparateChainingAggregationStateHashTable &>(
+ hash_table).forEach(&keys_retriever);
return nullptr;
}
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleDistinct.hpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleDistinct.hpp b/expressions/aggregation/AggregationHandleDistinct.hpp
index 838bfdd..deb928a 100644
--- a/expressions/aggregation/AggregationHandleDistinct.hpp
+++ b/expressions/aggregation/AggregationHandleDistinct.hpp
@@ -62,21 +62,13 @@ class AggregationHandleDistinct : public AggregationConcreteHandle {
<< "AggregationHandleDistinct does not support accumulateNullary().";
}
- AggregationState* accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors)
- const override {
- LOG(FATAL) << "AggregationHandleDistinct does not support "
- "accumulateColumnVectors().";
- }
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- AggregationState* accumulateValueAccessor(
+ AggregationState* accumulate(
ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const override {
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const override {
LOG(FATAL) << "AggregationHandleDistinct does not support "
- "accumulateValueAccessor().";
+ "accumulate().";
}
-#endif
void mergeStates(const AggregationState &source,
AggregationState *destination) const override {
@@ -87,33 +79,6 @@ class AggregationHandleDistinct : public AggregationConcreteHandle {
LOG(FATAL) << "AggregationHandleDistinct does not support finalize().";
}
- AggregationState* aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table)
- const override {
- LOG(FATAL) << "AggregationHandleDistinct does not support "
- << "aggregateOnDistinctifyHashTableForSingle().";
- }
-
- void aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *groupby_hash_table,
- std::size_t index) const override {
- LOG(FATAL) << "AggregationHandleDistinct does not support "
- << "aggregateOnDistinctifyHashTableForGroupBy().";
- }
-
- AggregationStateHashTableBase* createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const override;
-
- void aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const override;
-
ColumnVector* finalizeHashTable(
const AggregationStateHashTableBase &hash_table,
std::vector<std::vector<TypedValue>> *group_by_keys,
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleMax.cpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleMax.cpp b/expressions/aggregation/AggregationHandleMax.cpp
index c2d571b..6ffca0a 100644
--- a/expressions/aggregation/AggregationHandleMax.cpp
+++ b/expressions/aggregation/AggregationHandleMax.cpp
@@ -23,8 +23,7 @@
#include <vector>
#include "catalog/CatalogTypedefs.hpp"
-#include "storage/HashTable.hpp"
-#include "storage/HashTableFactory.hpp"
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
#include "types/Type.hpp"
#include "types/TypedValue.hpp"
#include "types/containers/ColumnVector.hpp"
@@ -39,51 +38,31 @@ namespace quickstep {
class StorageManager;
AggregationHandleMax::AggregationHandleMax(const Type &type)
- : type_(type), block_update_(false) {
+ : type_(type) {
fast_comparator_.reset(
ComparisonFactory::GetComparison(ComparisonID::kGreater)
.makeUncheckedComparatorForTypes(type, type.getNonNullableVersion()));
}
-AggregationStateHashTableBase* AggregationHandleMax::createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const {
- return AggregationStateHashTableFactory<AggregationStateMax>::CreateResizable(
- hash_table_impl, group_by_types, estimated_num_groups, storage_manager);
-}
-
-AggregationState* AggregationHandleMax::accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors) const {
- DCHECK_EQ(1u, column_vectors.size())
- << "Got wrong number of ColumnVectors for MAX: " << column_vectors.size();
+AggregationState* AggregationHandleMax::accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const {
+ DCHECK_EQ(1u, argument_ids.size())
+ << "Got wrong number of attributes for MAX: " << argument_ids.size();
- return new AggregationStateMax(fast_comparator_->accumulateColumnVector(
- type_.getNullableVersion().makeNullValue(), *column_vectors.front()));
-}
+ const attribute_id argument_id = argument_ids.front();
+ DCHECK_NE(argument_id, kInvalidAttributeID);
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-AggregationState* AggregationHandleMax::accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const {
- DCHECK_EQ(1u, accessor_ids.size())
- << "Got wrong number of attributes for MAX: " << accessor_ids.size();
+ ValueAccessor *target_accessor =
+ argument_id >= 0 ? accessor : aux_accessor;
+ const attribute_id target_argument_id =
+ argument_id >= 0 ? argument_id : -(argument_id+2);
return new AggregationStateMax(fast_comparator_->accumulateValueAccessor(
type_.getNullableVersion().makeNullValue(),
- accessor,
- accessor_ids.front()));
-}
-#endif // QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-
-void AggregationHandleMax::aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const {
- DCHECK_EQ(1u, argument_ids.size())
- << "Got wrong number of arguments for MAX: " << argument_ids.size();
+ target_accessor,
+ target_argument_id));
}
void AggregationHandleMax::mergeStates(const AggregationState &source,
@@ -98,12 +77,12 @@ void AggregationHandleMax::mergeStates(const AggregationState &source,
}
}
-void AggregationHandleMax::mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const {
+void AggregationHandleMax::mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const {
const TypedValue *src_max_ptr = reinterpret_cast<const TypedValue *>(source);
TypedValue *dst_max_ptr = reinterpret_cast<TypedValue *>(destination);
if (!(src_max_ptr->isNull())) {
- compareAndUpdateFast(dst_max_ptr, *src_max_ptr);
+ compareAndUpdate(dst_max_ptr, *src_max_ptr);
}
}
@@ -111,27 +90,10 @@ ColumnVector* AggregationHandleMax::finalizeHashTable(
const AggregationStateHashTableBase &hash_table,
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const {
- return finalizeHashTableHelperFast<AggregationHandleMax,
- AggregationStateFastHashTable>(
- type_.getNullableVersion(), hash_table, group_by_keys, index);
-}
-
-AggregationState*
-AggregationHandleMax::aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table) const {
- return aggregateOnDistinctifyHashTableForSingleUnaryHelperFast<
- AggregationHandleMax,
- AggregationStateMax>(distinctify_hash_table);
-}
-
-void AggregationHandleMax::aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const {
- aggregateOnDistinctifyHashTableForGroupByUnaryHelperFast<
+ return finalizeHashTableHelper<
AggregationHandleMax,
- AggregationStateFastHashTable>(
- distinctify_hash_table, aggregation_hash_table, index);
+ PackedPayloadSeparateChainingAggregationStateHashTable>(
+ type_.getNullableVersion(), hash_table, group_by_keys, index);
}
} // namespace quickstep
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleMax.hpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleMax.hpp b/expressions/aggregation/AggregationHandleMax.hpp
index d851a0c..8eddd7c 100644
--- a/expressions/aggregation/AggregationHandleMax.hpp
+++ b/expressions/aggregation/AggregationHandleMax.hpp
@@ -28,7 +28,6 @@
#include "catalog/CatalogTypedefs.hpp"
#include "expressions/aggregation/AggregationConcreteHandle.hpp"
#include "expressions/aggregation/AggregationHandle.hpp"
-#include "storage/FastHashTable.hpp"
#include "storage/HashTableBase.hpp"
#include "threading/SpinMutex.hpp"
#include "types/Type.hpp"
@@ -90,12 +89,6 @@ class AggregationHandleMax : public AggregationConcreteHandle {
return new AggregationStateMax(type_);
}
- AggregationStateHashTableBase* createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const override;
-
/**
* @brief Iterate with max aggregation state.
*/
@@ -105,23 +98,17 @@ class AggregationHandleMax : public AggregationConcreteHandle {
compareAndUpdate(static_cast<AggregationStateMax *>(state), value);
}
- inline void iterateUnaryInlFast(const TypedValue &value,
- std::uint8_t *byte_ptr) const {
- DCHECK(value.isPlausibleInstanceOf(type_.getSignature()));
- TypedValue *max_ptr = reinterpret_cast<TypedValue *>(byte_ptr);
- compareAndUpdateFast(max_ptr, value);
- }
-
- inline void updateStateUnary(const TypedValue &argument,
- std::uint8_t *byte_ptr) const override {
- if (!block_update_) {
- iterateUnaryInlFast(argument, byte_ptr);
- }
- }
+ AggregationState* accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const override;
- void blockUpdate() override { block_update_ = true; }
+ void mergeStates(const AggregationState &source,
+ AggregationState *destination) const override;
- void allowUpdate() override { block_update_ = false; }
+ std::size_t getPayloadSize() const override {
+ return sizeof(TypedValue);
+ }
void initPayload(std::uint8_t *byte_ptr) const override {
TypedValue *max_ptr = reinterpret_cast<TypedValue *>(byte_ptr);
@@ -136,38 +123,21 @@ class AggregationHandleMax : public AggregationConcreteHandle {
}
}
- AggregationState* accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors)
- const override;
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- AggregationState* accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const override;
-#endif
-
- void aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const override;
-
- void mergeStates(const AggregationState &source,
- AggregationState *destination) const override;
-
- void mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const override;
+ inline void updateStateUnary(const TypedValue &argument,
+ std::uint8_t *byte_ptr) const override {
+ DCHECK(argument.isPlausibleInstanceOf(type_.getSignature()));
+ TypedValue *max_ptr = reinterpret_cast<TypedValue *>(byte_ptr);
+ compareAndUpdate(max_ptr, argument);
+ }
TypedValue finalize(const AggregationState &state) const override {
return TypedValue(static_cast<const AggregationStateMax &>(state).max_);
}
- inline TypedValue finalizeHashTableEntry(
- const AggregationState &state) const {
- return TypedValue(static_cast<const AggregationStateMax &>(state).max_);
- }
+ void mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const override;
- inline TypedValue finalizeHashTableEntryFast(
+ inline TypedValue finalizeHashTableEntry(
const std::uint8_t *byte_ptr) const {
const TypedValue *max_ptr = reinterpret_cast<const TypedValue *>(byte_ptr);
return TypedValue(*max_ptr);
@@ -178,27 +148,6 @@ class AggregationHandleMax : public AggregationConcreteHandle {
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const override;
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForSingle()
- * for MAX aggregation.
- */
- AggregationState* aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table)
- const override;
-
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForGroupBy()
- * for MAX aggregation.
- */
- void aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const override;
-
- std::size_t getPayloadSize() const override { return sizeof(TypedValue); }
-
private:
friend class AggregateFunctionMax;
@@ -227,8 +176,8 @@ class AggregationHandleMax : public AggregationConcreteHandle {
}
}
- inline void compareAndUpdateFast(TypedValue *max_ptr,
- const TypedValue &value) const {
+ inline void compareAndUpdate(TypedValue *max_ptr,
+ const TypedValue &value) const {
if (value.isNull()) return;
if (max_ptr->isNull() ||
fast_comparator_->compareTypedValues(value, *max_ptr)) {
@@ -239,8 +188,6 @@ class AggregationHandleMax : public AggregationConcreteHandle {
const Type &type_;
std::unique_ptr<UncheckedComparator> fast_comparator_;
- bool block_update_;
-
DISALLOW_COPY_AND_ASSIGN(AggregationHandleMax);
};
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleMin.cpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleMin.cpp b/expressions/aggregation/AggregationHandleMin.cpp
index a07f299..f73fedb 100644
--- a/expressions/aggregation/AggregationHandleMin.cpp
+++ b/expressions/aggregation/AggregationHandleMin.cpp
@@ -23,8 +23,7 @@
#include <vector>
#include "catalog/CatalogTypedefs.hpp"
-#include "storage/HashTable.hpp"
-#include "storage/HashTableFactory.hpp"
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
#include "types/Type.hpp"
#include "types/TypedValue.hpp"
#include "types/containers/ColumnVector.hpp"
@@ -39,51 +38,31 @@ namespace quickstep {
class StorageManager;
AggregationHandleMin::AggregationHandleMin(const Type &type)
- : type_(type), block_update_(false) {
+ : type_(type) {
fast_comparator_.reset(
ComparisonFactory::GetComparison(ComparisonID::kLess)
.makeUncheckedComparatorForTypes(type, type.getNonNullableVersion()));
}
-AggregationStateHashTableBase* AggregationHandleMin::createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const {
- return AggregationStateHashTableFactory<AggregationStateMin>::CreateResizable(
- hash_table_impl, group_by_types, estimated_num_groups, storage_manager);
-}
-
-AggregationState* AggregationHandleMin::accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors) const {
- DCHECK_EQ(1u, column_vectors.size())
- << "Got wrong number of ColumnVectors for MIN: " << column_vectors.size();
+AggregationState* AggregationHandleMin::accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const {
+ DCHECK_EQ(1u, argument_ids.size())
+ << "Got wrong number of attributes for MIN: " << argument_ids.size();
- return new AggregationStateMin(fast_comparator_->accumulateColumnVector(
- type_.getNullableVersion().makeNullValue(), *column_vectors.front()));
-}
+ const attribute_id argument_id = argument_ids.front();
+ DCHECK_NE(argument_id, kInvalidAttributeID);
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-AggregationState* AggregationHandleMin::accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const {
- DCHECK_EQ(1u, accessor_ids.size())
- << "Got wrong number of attributes for MIN: " << accessor_ids.size();
+ ValueAccessor *target_accessor =
+ argument_id >= 0 ? accessor : aux_accessor;
+ const attribute_id target_argument_id =
+ argument_id >= 0 ? argument_id : -(argument_id+2);
return new AggregationStateMin(fast_comparator_->accumulateValueAccessor(
type_.getNullableVersion().makeNullValue(),
- accessor,
- accessor_ids.front()));
-}
-#endif // QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-
-void AggregationHandleMin::aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const {
- DCHECK_EQ(1u, argument_ids.size())
- << "Got wrong number of arguments for MIN: " << argument_ids.size();
+ target_accessor,
+ target_argument_id));
}
void AggregationHandleMin::mergeStates(const AggregationState &source,
@@ -98,13 +77,13 @@ void AggregationHandleMin::mergeStates(const AggregationState &source,
}
}
-void AggregationHandleMin::mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const {
+void AggregationHandleMin::mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const {
const TypedValue *src_min_ptr = reinterpret_cast<const TypedValue *>(source);
TypedValue *dst_min_ptr = reinterpret_cast<TypedValue *>(destination);
if (!(src_min_ptr->isNull())) {
- compareAndUpdateFast(dst_min_ptr, *src_min_ptr);
+ compareAndUpdate(dst_min_ptr, *src_min_ptr);
}
}
@@ -112,27 +91,10 @@ ColumnVector* AggregationHandleMin::finalizeHashTable(
const AggregationStateHashTableBase &hash_table,
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const {
- return finalizeHashTableHelperFast<AggregationHandleMin,
- AggregationStateFastHashTable>(
- type_.getNonNullableVersion(), hash_table, group_by_keys, index);
-}
-
-AggregationState*
-AggregationHandleMin::aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table) const {
- return aggregateOnDistinctifyHashTableForSingleUnaryHelperFast<
- AggregationHandleMin,
- AggregationStateMin>(distinctify_hash_table);
-}
-
-void AggregationHandleMin::aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const {
- aggregateOnDistinctifyHashTableForGroupByUnaryHelperFast<
+ return finalizeHashTableHelper<
AggregationHandleMin,
- AggregationStateFastHashTable>(
- distinctify_hash_table, aggregation_hash_table, index);
+ PackedPayloadSeparateChainingAggregationStateHashTable>(
+ type_.getNonNullableVersion(), hash_table, group_by_keys, index);
}
} // namespace quickstep
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleMin.hpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleMin.hpp b/expressions/aggregation/AggregationHandleMin.hpp
index e3472ec..7e129e6 100644
--- a/expressions/aggregation/AggregationHandleMin.hpp
+++ b/expressions/aggregation/AggregationHandleMin.hpp
@@ -28,7 +28,6 @@
#include "catalog/CatalogTypedefs.hpp"
#include "expressions/aggregation/AggregationConcreteHandle.hpp"
#include "expressions/aggregation/AggregationHandle.hpp"
-#include "storage/FastHashTable.hpp"
#include "storage/HashTableBase.hpp"
#include "threading/SpinMutex.hpp"
#include "types/Type.hpp"
@@ -92,38 +91,27 @@ class AggregationHandleMin : public AggregationConcreteHandle {
return new AggregationStateMin(type_);
}
- AggregationStateHashTableBase* createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const override;
-
- /**
- * @brief Iterate with min aggregation state.
- */
inline void iterateUnaryInl(AggregationStateMin *state,
const TypedValue &value) const {
DCHECK(value.isPlausibleInstanceOf(type_.getSignature()));
compareAndUpdate(state, value);
}
- inline void iterateUnaryInlFast(const TypedValue &value,
- std::uint8_t *byte_ptr) const {
- DCHECK(value.isPlausibleInstanceOf(type_.getSignature()));
- TypedValue *min_ptr = reinterpret_cast<TypedValue *>(byte_ptr);
- compareAndUpdateFast(min_ptr, value);
- }
+ AggregationState* accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const override;
- inline void updateStateUnary(const TypedValue &argument,
- std::uint8_t *byte_ptr) const override {
- if (!block_update_) {
- iterateUnaryInlFast(argument, byte_ptr);
- }
- }
+ void mergeStates(const AggregationState &source,
+ AggregationState *destination) const override;
- void blockUpdate() override { block_update_ = true; }
+ TypedValue finalize(const AggregationState &state) const override {
+ return static_cast<const AggregationStateMin &>(state).min_;
+ }
- void allowUpdate() override { block_update_ = false; }
+ std::size_t getPayloadSize() const override {
+ return sizeof(TypedValue);
+ }
void initPayload(std::uint8_t *byte_ptr) const override {
TypedValue *min_ptr = reinterpret_cast<TypedValue *>(byte_ptr);
@@ -138,41 +126,19 @@ class AggregationHandleMin : public AggregationConcreteHandle {
}
}
- AggregationState* accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors)
- const override;
-
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
- AggregationState* accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const override;
-#endif
-
- void aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const override;
-
- void mergeStates(const AggregationState &source,
- AggregationState *destination) const override;
-
- void mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const override;
-
- TypedValue finalize(const AggregationState &state) const override {
- return static_cast<const AggregationStateMin &>(state).min_;
+ inline void updateStateUnary(const TypedValue &argument,
+ std::uint8_t *byte_ptr) const override {
+ DCHECK(argument.isPlausibleInstanceOf(type_.getSignature()));
+ TypedValue *min_ptr = reinterpret_cast<TypedValue *>(byte_ptr);
+ compareAndUpdate(min_ptr, argument);
}
- inline TypedValue finalizeHashTableEntry(
- const AggregationState &state) const {
- return static_cast<const AggregationStateMin &>(state).min_;
- }
+ void mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const override;
- inline TypedValue finalizeHashTableEntryFast(
+ inline TypedValue finalizeHashTableEntry(
const std::uint8_t *byte_ptr) const {
- const TypedValue *min_ptr = reinterpret_cast<const TypedValue *>(byte_ptr);
- return TypedValue(*min_ptr);
+ return *reinterpret_cast<const TypedValue *>(byte_ptr);
}
ColumnVector* finalizeHashTable(
@@ -180,27 +146,6 @@ class AggregationHandleMin : public AggregationConcreteHandle {
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const override;
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForSingle()
- * for MIN aggregation.
- */
- AggregationState* aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table)
- const override;
-
- /**
- * @brief Implementation of
- * AggregationHandle::aggregateOnDistinctifyHashTableForGroupBy()
- * for MIN aggregation.
- */
- void aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const override;
-
- std::size_t getPayloadSize() const override { return sizeof(TypedValue); }
-
private:
friend class AggregateFunctionMin;
@@ -228,8 +173,8 @@ class AggregationHandleMin : public AggregationConcreteHandle {
}
}
- inline void compareAndUpdateFast(TypedValue *min_ptr,
- const TypedValue &value) const {
+ inline void compareAndUpdate(TypedValue *min_ptr,
+ const TypedValue &value) const {
if (value.isNull()) return;
if (min_ptr->isNull() ||
fast_comparator_->compareTypedValues(value, *min_ptr)) {
@@ -240,8 +185,6 @@ class AggregationHandleMin : public AggregationConcreteHandle {
const Type &type_;
std::unique_ptr<UncheckedComparator> fast_comparator_;
- bool block_update_;
-
DISALLOW_COPY_AND_ASSIGN(AggregationHandleMin);
};
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/expressions/aggregation/AggregationHandleSum.cpp
----------------------------------------------------------------------
diff --git a/expressions/aggregation/AggregationHandleSum.cpp b/expressions/aggregation/AggregationHandleSum.cpp
index 642d88d..9ddeb99 100644
--- a/expressions/aggregation/AggregationHandleSum.cpp
+++ b/expressions/aggregation/AggregationHandleSum.cpp
@@ -25,8 +25,7 @@
#include <vector>
#include "catalog/CatalogTypedefs.hpp"
-#include "storage/HashTable.hpp"
-#include "storage/HashTableFactory.hpp"
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
#include "threading/SpinMutex.hpp"
#include "types/Type.hpp"
#include "types/TypeFactory.hpp"
@@ -43,7 +42,7 @@ namespace quickstep {
class StorageManager;
AggregationHandleSum::AggregationHandleSum(const Type &type)
- : argument_type_(type), block_update_(false) {
+ : argument_type_(type) {
// We sum Int as Long and Float as Double so that we have more headroom when
// adding many values.
TypeID type_precision_id;
@@ -79,47 +78,26 @@ AggregationHandleSum::AggregationHandleSum(const Type &type)
result_type_ = &sum_type.getNullableVersion();
}
-AggregationStateHashTableBase* AggregationHandleSum::createGroupByHashTable(
- const HashTableImplType hash_table_impl,
- const std::vector<const Type *> &group_by_types,
- const std::size_t estimated_num_groups,
- StorageManager *storage_manager) const {
- return AggregationStateHashTableFactory<AggregationStateSum>::CreateResizable(
- hash_table_impl, group_by_types, estimated_num_groups, storage_manager);
-}
+AggregationState* AggregationHandleSum::accumulate(
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor,
+ const std::vector<attribute_id> &argument_ids) const {
+ DCHECK_EQ(1u, argument_ids.size())
+ << "Got wrong number of attributes for SUM: " << argument_ids.size();
-AggregationState* AggregationHandleSum::accumulateColumnVectors(
- const std::vector<std::unique_ptr<ColumnVector>> &column_vectors) const {
- DCHECK_EQ(1u, column_vectors.size())
- << "Got wrong number of ColumnVectors for SUM: " << column_vectors.size();
- std::size_t num_tuples = 0;
- TypedValue cv_sum = fast_operator_->accumulateColumnVector(
- blank_state_.sum_, *column_vectors.front(), &num_tuples);
- return new AggregationStateSum(std::move(cv_sum), num_tuples == 0);
-}
+ const attribute_id argument_id = argument_ids.front();
+ DCHECK_NE(argument_id, kInvalidAttributeID);
-#ifdef QUICKSTEP_ENABLE_VECTOR_COPY_ELISION_SELECTION
-AggregationState* AggregationHandleSum::accumulateValueAccessor(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &accessor_ids) const {
- DCHECK_EQ(1u, accessor_ids.size())
- << "Got wrong number of attributes for SUM: " << accessor_ids.size();
+ ValueAccessor *target_accessor =
+ argument_id >= 0 ? accessor : aux_accessor;
+ const attribute_id target_argument_id =
+ argument_id >= 0 ? argument_id : -(argument_id+2);
std::size_t num_tuples = 0;
TypedValue va_sum = fast_operator_->accumulateValueAccessor(
- blank_state_.sum_, accessor, accessor_ids.front(), &num_tuples);
+ blank_state_.sum_, target_accessor, target_argument_id, &num_tuples);
return new AggregationStateSum(std::move(va_sum), num_tuples == 0);
}
-#endif
-
-void AggregationHandleSum::aggregateValueAccessorIntoHashTable(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &argument_ids,
- const std::vector<attribute_id> &group_by_key_ids,
- AggregationStateHashTableBase *hash_table) const {
- DCHECK_EQ(1u, argument_ids.size())
- << "Got wrong number of arguments for SUM: " << argument_ids.size();
-}
void AggregationHandleSum::mergeStates(const AggregationState &source,
AggregationState *destination) const {
@@ -134,8 +112,8 @@ void AggregationHandleSum::mergeStates(const AggregationState &source,
sum_destination->null_ = sum_destination->null_ && sum_source.null_;
}
-void AggregationHandleSum::mergeStatesFast(const std::uint8_t *source,
- std::uint8_t *destination) const {
+void AggregationHandleSum::mergeStates(const std::uint8_t *source,
+ std::uint8_t *destination) const {
const TypedValue *src_sum_ptr =
reinterpret_cast<const TypedValue *>(source + blank_state_.sum_offset_);
const bool *src_null_ptr =
@@ -164,27 +142,10 @@ ColumnVector* AggregationHandleSum::finalizeHashTable(
const AggregationStateHashTableBase &hash_table,
std::vector<std::vector<TypedValue>> *group_by_keys,
int index) const {
- return finalizeHashTableHelperFast<AggregationHandleSum,
- AggregationStateFastHashTable>(
- *result_type_, hash_table, group_by_keys, index);
-}
-
-AggregationState*
-AggregationHandleSum::aggregateOnDistinctifyHashTableForSingle(
- const AggregationStateHashTableBase &distinctify_hash_table) const {
- return aggregateOnDistinctifyHashTableForSingleUnaryHelperFast<
- AggregationHandleSum,
- AggregationStateSum>(distinctify_hash_table);
-}
-
-void AggregationHandleSum::aggregateOnDistinctifyHashTableForGroupBy(
- const AggregationStateHashTableBase &distinctify_hash_table,
- AggregationStateHashTableBase *aggregation_hash_table,
- std::size_t index) const {
- aggregateOnDistinctifyHashTableForGroupByUnaryHelperFast<
+ return finalizeHashTableHelper<
AggregationHandleSum,
- AggregationStateFastHashTable>(
- distinctify_hash_table, aggregation_hash_table, index);
+ PackedPayloadSeparateChainingAggregationStateHashTable>(
+ *result_type_, hash_table, group_by_keys, index);
}
} // namespace quickstep
[3/5] incubator-quickstep git commit: Cleanup FastHashTable + Avoid
unnecessary copying of data before aggregation.
Posted by ji...@apache.org.
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/FastHashTable.hpp
----------------------------------------------------------------------
diff --git a/storage/FastHashTable.hpp b/storage/FastHashTable.hpp
deleted file mode 100644
index 4a82a62..0000000
--- a/storage/FastHashTable.hpp
+++ /dev/null
@@ -1,2403 +0,0 @@
-/**
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements. See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership. The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * "License"); you may not use this file except in compliance
- * with the License. You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations
- * under the License.
- **/
-
-#ifndef QUICKSTEP_STORAGE_FAST_HASH_TABLE_HPP_
-#define QUICKSTEP_STORAGE_FAST_HASH_TABLE_HPP_
-
-#include <atomic>
-#include <cstddef>
-#include <cstdlib>
-#include <type_traits>
-#include <vector>
-
-#include "catalog/CatalogTypedefs.hpp"
-#include "storage/HashTableBase.hpp"
-#include "storage/StorageBlob.hpp"
-#include "storage/StorageBlockInfo.hpp"
-#include "storage/StorageConstants.hpp"
-#include "storage/StorageManager.hpp"
-#include "storage/TupleReference.hpp"
-#include "storage/ValueAccessor.hpp"
-#include "storage/ValueAccessorUtil.hpp"
-#include "threading/SpinMutex.hpp"
-#include "threading/SpinSharedMutex.hpp"
-#include "types/Type.hpp"
-#include "types/TypedValue.hpp"
-#include "utility/HashPair.hpp"
-#include "utility/Macros.hpp"
-
-namespace quickstep {
-
-/** \addtogroup Storage
- * @{
- */
-
-/**
- * @brief Base class for the hash table implementation in which the payload can
- * be just a bunch of bytes. This implementation is suitable for
- * aggregation hash table with multiple aggregation handles (e.g. SUM,
- * MAX, MIN etc).
- *
- * At present there is one implementation for this base class.
- * 1. SeparateChainingHashTable - Keys/values are stored in a separate
- * region of memory from the base hash table slot array. Every bucket
- * has a "next" pointer so that entries that collide (i.e. map to the
- * same base slot) form chains of pointers with each other. Although
- * this implementation has some extra indirection compared to
- * LinearOpenAddressingHashTable, it does not have the same
- * vulnerabilities to key skew, and it additionally supports a very
- * efficient bucket-preallocation mechanism that minimizes cache
- * coherency overhead when multiple threads are building a HashTable.
- *
- * @note If you need to create a HashTable and not just use it as a client, see
- * HashTableFactory, which simplifies the process of creating a
- * HashTable.
- *
- * @param resizable Whether this hash table is resizable (using memory from a
- * StorageManager) or not (using a private, fixed memory allocation).
- * @param serializable If true, this hash table can safely be saved to and
- * loaded from disk. If false, some out of band memory may be used (e.g.
- * to store variable length keys).
- * @param force_key_copy If true, inserted keys are always copied into this
- * HashTable's memory. If false, pointers to external values may be
- * stored instead. force_key_copy should be true if the hash table will
- * outlive the external key values which are inserted into it. Note that
- * if serializable is true and force_key_copy is false, then relative
- * offsets will be used instead of absolute pointers to keys, meaning
- * that the pointed-to keys must be serialized and deserialized in
- * exactly the same relative byte order (e.g. as part of the same
- * StorageBlock), and keys must not change position relative to this
- * HashTable (beware TupleStorageSubBlocks that may self-reorganize when
- * modified). If serializable and resizable are both true, then
- * force_key_copy must also be true.
- * @param allow_duplicate_keys If true, multiple values can be mapped to the
- * same key. If false, one and only one value may be mapped.
- **/
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-class FastHashTable : public HashTableBase<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys> {
- static_assert(!(serializable && resizable && !force_key_copy),
- "A HashTable must have force_key_copy=true when serializable "
- "and resizable are both true.");
-
- public:
- // Shadow template parameters. This is useful for shared test harnesses.
- static constexpr bool template_resizable = resizable;
- static constexpr bool template_serializable = serializable;
- static constexpr bool template_force_key_copy = force_key_copy;
- static constexpr bool template_allow_duplicate_keys = allow_duplicate_keys;
-
- // Some HashTable implementations (notably LinearOpenAddressingHashTable)
- // use a special hash code to represent an empty bucket, and another special
- // code to indicate that a bucket is currently being inserted into. For those
- // HashTables, this is a surrogate hash value for empty buckets. Keys which
- // actually hash to this value should have their hashes mutated (e.g. by
- // adding 1). We use zero, since we will often be using memory which is
- // already zeroed-out and this saves us the trouble of a memset. This has
- // some downside, as the hash function we use is the identity hash for
- // integers, and the integer 0 is common in many data sets and must be
- // adjusted (and will then spuriously collide with 1). Nevertheless, this
- // expense is outweighed by no longer having to memset large regions of
- // memory when initializing a HashTable.
- static constexpr unsigned char kEmptyHashByte = 0x0;
- static constexpr std::size_t kEmptyHash = 0x0;
-
- // A surrogate hash value for a bucket which is currently being inserted
- // into. As with kEmptyHash, keys which actually hash to this value should
- // have their hashes adjusted.
- static constexpr std::size_t kPendingHash = ~kEmptyHash;
-
- /**
- * @brief Virtual destructor.
- **/
- virtual ~FastHashTable() {
- if (resizable) {
- if (blob_.valid()) {
- if (serializable) {
- DEV_WARNING(
- "Destroying a resizable serializable HashTable's underlying "
- "StorageBlob.");
- }
- const block_id blob_id = blob_->getID();
- blob_.release();
- storage_manager_->deleteBlockOrBlobFile(blob_id);
- }
- }
- }
-
- /**
- * @brief Get the ID of the StorageBlob used to store a resizable HashTable.
- *
- * @warning This method must not be used for a non-resizable HashTable.
- *
- * @return The ID of the StorageBlob used to store this HashTable.
- **/
- inline block_id getBlobId() const {
- DEBUG_ASSERT(resizable);
- return blob_->getID();
- }
-
- /**
- * @brief Erase all entries in this hash table.
- *
- * @warning This method is not guaranteed to be threadsafe.
- **/
- virtual void clear() = 0;
-
- /**
- * @brief Add a new entry into the hash table.
- *
- * @warning The key must not be null.
- * @warning This method is threadsafe with regard to other calls to put(),
- * putCompositeKey(), putValueAccessor(), and
- * putValueAccessorCompositeKey(), but should not be used
- * simultaneously with upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey().
- * @note This version is for single scalar keys, see also putCompositeKey().
- * @note If the hash table is (close to) full and resizable is true, this
- * routine might result in rebuilding the entire hash table.
- *
- * @param key The key.
- * @param value The value payload.
- * @return HashTablePutResult::kOK if an entry was successfully inserted,
- * HashTablePutResult::kDuplicateKey if allow_duplicate_keys is false
- * and key was a duplicate, or HashTablePutResult::kOutOfSpace if
- * resizable is false and storage space for the hash table has been
- * exhausted.
- **/
- HashTablePutResult put(const TypedValue &key, const std::uint8_t &value);
-
- /**
- * @brief Add a new entry into the hash table (composite key version).
- *
- * @warning No component of the key may be null.
- * @warning This method is threadsafe with regard to other calls to put(),
- * putCompositeKey(), putValueAccessor(), and
- * putValueAccessorCompositeKey(), but should not be used
- * simultaneously with upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey().
- * @note This version is for composite keys, see also put().
- * @note If the hash table is (close to) full and resizable is true, this
- * routine might result in rebuilding the entire hash table.
- *
- * @param key The components of the key.
- * @param value The value payload.
- * @return HashTablePutResult::kOK if an entry was successfully inserted,
- * HashTablePutResult::kDuplicateKey if allow_duplicate_keys is false
- * and key was a duplicate, or HashTablePutResult::kOutOfSpace if
- * resizable is false and storage space for the hash table has been
- * exhausted.
- **/
-
- HashTablePutResult putCompositeKey(const std::vector<TypedValue> &key,
- const std::uint8_t *value_ptr);
-
- /**
- * @brief Add (multiple) new entries into the hash table from a
- * ValueAccessor.
- *
- * @warning This method is threadsafe with regard to other calls to put(),
- * putCompositeKey(), putValueAccessor(), and
- * putValueAccessorCompositeKey(), but should not be used
- * simultaneously with upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey().
- * @note This version is for scalar keys, see also
- * putValueAccessorCompositeKey().
- * @note If the hash table fills up while this call is in progress and
- * resizable is true, this might result in rebuilding the entire hash
- * table.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before inserting it (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor, which should provide a call
- * operator that takes const ValueAccessor& as an argument (or better
- * yet, a templated call operator which takes a const reference to
- * some subclass of ValueAccessor as an argument) and returns either
- * a ValueT or a reference to a ValueT. The functor should generate
- * the appropriate mapped value for the current tuple the accessor is
- * iterating on.
- * @return HashTablePutResult::kOK if all keys and generated values from
- * accessor were successfully inserted.
- * HashTablePutResult::kOutOfSpace is returned if this hash-table is
- * non-resizable and ran out of space (note that some entries may
- * still have been inserted, and accessor's iteration will be left on
- * the first tuple which could not be inserted).
- * HashTablePutResult::kDuplicateKey is returned if
- * allow_duplicate_keys is false and a duplicate key is encountered
- * (as with HashTablePutResult::kOutOfSpace, some entries may have
- * been inserted, and accessor will be left on the tuple with a
- * duplicate key).
- **/
- template <typename FunctorT>
- HashTablePutResult putValueAccessor(ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys,
- FunctorT *functor);
-
- /**
- * @brief Add (multiple) new entries into the hash table from a
- * ValueAccessor (composite key version).
- *
- * @warning This method is threadsafe with regard to other calls to put(),
- * putCompositeKey(), putValueAccessor(), and
- * putValueAccessorCompositeKey(), but should not be used
- * simultaneously with upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey().
- * @note This version is for composite keys, see also putValueAccessor().
- * @note If the hash table fills up while this call is in progress and
- * resizable is true, this might result in rebuilding the entire hash
- * table.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_ids The attribute IDs of each key component to be read
- * from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * has a null component before inserting it (null keys are skipped).
- * This must be set to true if some of the keys that will be read from
- * accessor may be null.
- * @param functor A pointer to a functor, which should provide a call
- * operator that takes const ValueAccessor& as an argument (or better
- * yet, a templated call operator which takes a const reference to
- * some subclass of ValueAccessor as an argument) and returns either
- * a ValueT or a reference to a ValueT. The functor should generate
- * the appropriate mapped value for the current tuple the accessor is
- * iterating on.
- * @return HashTablePutResult::kOK if all keys and generated values from
- * accessor were successfully inserted.
- * HashTablePutResult::kOutOfSpace is returned if this hash-table is
- * non-resizable and ran out of space (note that some entries may
- * still have been inserted, and accessor's iteration will be left on
- * the first tuple which could not be inserted).
- * HashTablePutResult::kDuplicateKey is returned if
- * allow_duplicate_keys is false and a duplicate key is encountered
- * (as with HashTablePutResult::kOutOfSpace, some entries may have
- * been inserted, and accessor will be left on the tuple with a
- * duplicate key).
- **/
- template <typename FunctorT>
- HashTablePutResult putValueAccessorCompositeKey(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys,
- FunctorT *functor);
-
- /**
- * @brief Apply a functor to the value mapped to a key, first inserting a new
- * value if one is not already present.
- *
- * @warning The key must not be null.
- * @warning This method is only usable if allow_duplicate_keys is false.
- * @warning This method is threadsafe with regard to other calls to upsert(),
- * upsertCompositeKey(), upsertValueAccessor(), and
- * upsertValueAccessorCompositeKey(), but should not be used
- * simultaneously with put(), putCompositeKey(), putValueAccessor(),
- * or putValueAccessorCompositeKey().
- * @warning The ValueT* pointer passed to functor's call operator is only
- * guaranteed to be valid for the duration of the call. The functor
- * should not store a copy of the pointer and assume that it remains
- * valid.
- * @warning Although this method itself is threadsafe, the ValueT object
- * accessed by functor is not guaranteed to be (although it is
- * guaranteed that its initial insertion will be atomic). If it is
- * possible for multiple threads to call upsert() with the same key
- * at the same time, then their access to ValueT should be made
- * threadsafe (e.g. with the use of atomic types, mutexes, or some
- * other external synchronization).
- * @note This version is for single scalar keys, see also
- * upsertCompositeKey().
- * @note If the hash table is (close to) full and resizable is true, this
- * routine might result in rebuilding the entire hash table.
- *
- * @param key The key.
- * @param initial_value If there was not already a preexisting entry in this
- * HashTable for the specified key, then the value will be initialized
- * with a copy of initial_value. This parameter is ignored if a value
- * is already present for key.
- * @param functor A pointer to a functor, which should provide a call
- * operator which takes ValueT* as an argument. The call operator will
- * be invoked once on the value corresponding to key (which may be
- * newly inserted and default-constructed).
- * @return True on success, false if upsert failed because there was not
- * enough space to insert a new entry in this HashTable.
- **/
- template <typename FunctorT>
- bool upsert(const TypedValue &key,
- const std::uint8_t *initial_value_ptr,
- FunctorT *functor);
-
- /**
- * @brief Apply a functor to the value mapped to a key, first inserting a new
- * value if one is not already present.
- *
- * @warning The key must not be null.
- * @warning This method is only usable if allow_duplicate_keys is false.
- * @warning This method is threadsafe with regard to other calls to upsert(),
- * upsertCompositeKey(), upsertValueAccessor(), and
- * upsertValueAccessorCompositeKey(), but should not be used
- * simultaneously with put(), putCompositeKey(), putValueAccessor(),
- * or putValueAccessorCompositeKey().
- * @warning The ValueT* pointer passed to functor's call operator is only
- * guaranteed to be valid for the duration of the call. The functor
- * should not store a copy of the pointer and assume that it remains
- * valid.
- * @warning Although this method itself is threadsafe, the ValueT object
- * accessed by functor is not guaranteed to be (although it is
- * guaranteed that its initial insertion will be atomic). If it is
- * possible for multiple threads to call upsertCompositeKey() with
- * the same key at the same time, then their access to ValueT should
- * be made threadsafe (e.g. with the use of atomic types, mutexes,
- * or some other external synchronization).
- * @note This version is for composite keys, see also upsert().
- * @note If the hash table is (close to) full and resizable is true, this
- * routine might result in rebuilding the entire hash table.
- *
- * @param key The key.
- * @param initial_value If there was not already a preexisting entry in this
- * HashTable for the specified key, then the value will be initialized
- * with a copy of initial_value. This parameter is ignored if a value
- * is already present for key.
- * @param functor A pointer to a functor, which should provide a call
- * operator which takes ValueT* as an argument. The call operator will
- * be invoked once on the value corresponding to key (which may be
- * newly inserted and default-constructed).
- * @return True on success, false if upsert failed because there was not
- * enough space to insert a new entry in this HashTable.
- **/
- template <typename FunctorT>
- bool upsertCompositeKeyFast(const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- FunctorT *functor);
-
- template <typename FunctorT>
- bool upsertCompositeKeyFast(const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- FunctorT *functor,
- int index);
-
- bool upsertCompositeKeyFast(const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- const std::uint8_t *source_state);
-
- /**
- * @brief Apply a functor to (multiple) entries in this hash table, with keys
- * drawn from a ValueAccessor. New values are first inserted if not
- * already present.
- *
- * @warning This method is only usable if allow_duplicate_keys is false.
- * @warning This method is threadsafe with regard to other calls to upsert(),
- * upsertCompositeKey(), upsertValueAccessor(), and
- * upsertValueAccessorCompositeKey(), but should not be used
- * simultaneously with put(), putCompositeKey(), putValueAccessor(),
- * or putValueAccessorCompositeKey().
- * @warning The ValueAccessor reference and ValueT* pointer passed to
- * functor's call operator are only guaranteed to be valid for the
- * duration of the call. The functor should not store a copy of
- * these pointers and assume that they remain valid.
- * @warning Although this method itself is threadsafe, the ValueT object
- * accessed by functor is not guaranteed to be (although it is
- * guaranteed that its initial insertion will be atomic). If it is
- * possible for multiple threads to call upsertValueAccessor() with
- * the same key at the same time, then their access to ValueT should
- * be made threadsafe (e.g. with the use of atomic types, mutexes,
- * or some other external synchronization).
- * @note This version is for single scalar keys, see also
- * upsertValueAccessorCompositeKey().
- * @note If the hash table is (close to) full and resizable is true, this
- * routine might result in rebuilding the entire hash table.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before upserting it (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor, which should provide a call
- * operator that takes two arguments: const ValueAccessor& (or better
- * yet, a templated call operator which takes a const reference to
- * some subclass of ValueAccessor as its first argument) and ValueT*.
- * The call operator will be invoked once for every tuple with a
- * non-null key in accessor.
- * @return True on success, false if upsert failed because there was not
- * enough space to insert new entries for all the keys in accessor
- * (note that some entries may still have been upserted, and
- * accessor's iteration will be left on the first tuple which could
- * not be inserted).
- **/
- bool upsertValueAccessorFast(
- const std::vector<attribute_id> &argument_ids,
- ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys);
-
- /**
- * @brief Apply a functor to (multiple) entries in this hash table, with keys
- * drawn from a ValueAccessor. New values are first inserted if not
- * already present. Composite key version.
- *
- * @warning This method is only usable if allow_duplicate_keys is false.
- * @warning This method is threadsafe with regard to other calls to upsert(),
- * upsertCompositeKey(), upsertValueAccessor(), and
- * upsertValueAccessorCompositeKey(), but should not be used
- * simultaneously with put(), putCompositeKey(), putValueAccessor(),
- * or putValueAccessorCompositeKey().
- * @warning The ValueAccessor reference and ValueT* pointer passed to
- * functor's call operator are only guaranteed to be valid for the
- * duration of the call. The functor should not store a copy of
- * these pointers and assume that they remain valid.
- * @warning Although this method itself is threadsafe, the ValueT object
- * accessed by functor is not guaranteed to be (although it is
- * guaranteed that its initial insertion will be atomic). If it is
- * possible for multiple threads to call upsertValueAccessor() with
- * the same key at the same time, then their access to ValueT should
- * be made threadsafe (e.g. with the use of atomic types, mutexes,
- * or some other external synchronization).
- * @note This version is for composite keys, see also upsertValueAccessor().
- * @note If the hash table is (close to) full and resizable is true, this
- * routine might result in rebuilding the entire hash table.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_ids The attribute IDs of each key component to be read
- * from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before upserting it (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor, which should provide a call
- * operator that takes two arguments: const ValueAccessor& (or better
- * yet, a templated call operator which takes a const reference to
- * some subclass of ValueAccessor as its first argument) and ValueT*.
- * The call operator will be invoked once for every tuple with a
- * non-null key in accessor.
- * @return True on success, false if upsert failed because there was not
- * enough space to insert new entries for all the keys in accessor
- * (note that some entries may still have been upserted, and
- * accessor's iteration will be left on the first tuple which could
- * not be inserted).
- **/
- bool upsertValueAccessorCompositeKeyFast(
- const std::vector<attribute_id> &argument,
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys) override;
-
- /**
- * @brief Determine the number of entries (key-value pairs) contained in this
- * HashTable.
- * @note For some HashTable implementations, this is O(1), but for others it
- * may be O(n) where n is the number of buckets.
- *
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- *
- * @return The number of entries in this HashTable.
- **/
- virtual std::size_t numEntries() const = 0;
-
- /**
- * @brief Lookup a key against this hash table to find a matching entry.
- *
- * @warning Only usable with the hash table that does not allow duplicate
- * keys.
- * @warning The key must not be null.
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note This version is for single scalar keys. See also
- * getSingleCompositeKey().
- *
- * @param key The key to look up.
- * @return The value of a matched entry if a matching key is found.
- * Otherwise, return NULL.
- **/
- virtual const std::uint8_t* getSingle(const TypedValue &key) const = 0;
-
- /**
- * @brief Lookup a composite key against this hash table to find a matching
- * entry.
- *
- * @warning Only usable with the hash table that does not allow duplicate
- * keys.
- * @warning The key must not be null.
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note This version is for composite keys. See also getSingle().
- *
- * @param key The key to look up.
- * @return The value of a matched entry if a matching key is found.
- * Otherwise, return NULL.
- **/
- virtual const std::uint8_t* getSingleCompositeKey(
- const std::vector<TypedValue> &key) const = 0;
- virtual const std::uint8_t *getSingleCompositeKey(
- const std::vector<TypedValue> &key, int index) const = 0;
-
- /**
- * @brief Lookup a key against this hash table to find matching entries.
- *
- * @warning The key must not be null.
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note It is more efficient to call getSingle() if the hash table does not
- * allow duplicate keys.
- * @note This version is for single scalar keys. See also
- * getAllCompositeKey().
- *
- * @param key The key to look up.
- * @param values A vector to hold values of all matching entries. Matches
- * will be appended to the vector.
- **/
- virtual void getAll(const TypedValue &key,
- std::vector<const std::uint8_t *> *values) const = 0;
-
- /**
- * @brief Lookup a composite key against this hash table to find matching
- * entries.
- *
- * @warning The key must not be null.
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note It is more efficient to call getSingleCompositeKey() if the hash
- * table does not allow duplicate keys.
- * @note This version is for composite keys. See also getAll().
- *
- * @param key The key to look up.
- * @param values A vector to hold values of all matching entries. Matches
- * will be appended to the vector.
- **/
- virtual void getAllCompositeKey(
- const std::vector<TypedValue> &key,
- std::vector<const std::uint8_t *> *values) const = 0;
-
- /**
- * @brief Lookup (multiple) keys from a ValueAccessor and apply a functor to
- * the matching values.
- *
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note This version is for single scalar keys. See also
- * getAllFromValueAccessorCompositeKey().
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before looking it up (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor, which should provide a call
- * operator that takes 2 arguments: const ValueAccessor& (or better
- * yet, a templated call operator which takes a const reference to
- * some subclass of ValueAccessor as its first argument) and
- * const ValueT&. The functor will be invoked once for each pair of a
- * key taken from accessor and matching value.
- **/
- template <typename FunctorT>
- void getAllFromValueAccessor(ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys,
- FunctorT *functor) const;
-
- /**
- * @brief Lookup (multiple) keys from a ValueAccessor, apply a functor to the
- * matching values and additionally call a recordMatch() function of
- * the functor when the first match for a key is found.
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note This version is for single scalar keys. See also
- * getAllFromValueAccessorCompositeKeyWithExtraWorkForFirstMatch().
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before looking it up (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor, which should provide two functions:
- * 1) An operator that takes 2 arguments: const ValueAccessor& (or
- * better
- * yet, a templated call operator which takes a const reference to
- * some subclass of ValueAccessor as its first argument) and
- * const ValueT&. The operator will be invoked once for each pair of a
- * key taken from accessor and matching value.
- * 2) A function hasMatch that takes 1 argument: const ValueAccessor&.
- * The function will be called only once for a key from accessor when
- * the first match is found.
- */
- template <typename FunctorT>
- void getAllFromValueAccessorWithExtraWorkForFirstMatch(
- ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys,
- FunctorT *functor) const;
-
- /**
- * @brief Lookup (multiple) keys from a ValueAccessor, apply a functor to the
- * matching values and additionally call a recordMatch() function of
- * the functor when the first match for a key is found. Composite key
- * version.
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before looking it up (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor, which should provide two functions:
- * 1) An operator that takes 2 arguments: const ValueAccessor& (or
- * better
- * yet, a templated call operator which takes a const reference to
- * some subclass of ValueAccessor as its first argument) and
- * const ValueT&. The operator will be invoked once for each pair of a
- * key taken from accessor and matching value.
- * 2) A function hasMatch that takes 1 argument: const ValueAccessor&.
- * The function will be called only once for a key from accessor when
- * the first match is found.
- */
- template <typename FunctorT>
- void getAllFromValueAccessorCompositeKeyWithExtraWorkForFirstMatch(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys,
- FunctorT *functor) const;
-
- /**
- * @brief Lookup (multiple) keys from a ValueAccessor and apply a functor to
- * the matching values. Composite key version.
- *
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note This version is for composite keys. See also
- * getAllFromValueAccessor().
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_ids The attribute IDs of each key component to be read
- * from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * has a null component before inserting it (null keys are skipped).
- * This must be set to true if some of the keys that will be read from
- * accessor may be null.
- * @param functor A pointer to a functor, which should provide a call
- * operator that takes 2 arguments: const ValueAccessor& (or better
- * yet, a templated call operator which takes a const reference to
- * some subclass of ValueAccessor as its first argument) and
- * const ValueT&. The functor will be invoked once for each pair of a
- * key taken from accessor and matching value.
- **/
- template <typename FunctorT>
- void getAllFromValueAccessorCompositeKey(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys,
- FunctorT *functor) const;
-
- /**
- * @brief Apply the functor to each key with a match in the hash table.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before looking it up (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor which should provide an operator that
- * takes 1 argument: const ValueAccessor&. The operator will be called
- * only once for a key from accessor if there is a match.
- */
- template <typename FunctorT>
- void runOverKeysFromValueAccessorIfMatchFound(ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys,
- FunctorT *functor) const {
- return runOverKeysFromValueAccessor<true>(
- accessor, key_attr_id, check_for_null_keys, functor);
- }
-
- /**
- * @brief Apply the functor to each key with a match in the hash table.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before looking it up (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor which should provide an operator that
- * takes 1 argument: const ValueAccessor&. The operator will be called
- * only once for a key from accessor if there is a match.
- */
- template <typename FunctorT>
- void runOverKeysFromValueAccessorIfMatchFoundCompositeKey(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys,
- FunctorT *functor) const {
- return runOverKeysFromValueAccessorCompositeKey<true>(
- accessor, key_attr_ids, check_for_null_keys, functor);
- }
-
- /**
- * @brief Apply the functor to each key without a match in the hash table.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before looking it up (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor which should provide an operator that
- * takes 1 argument: const ValueAccessor&. The operator will be called
- * only once for a key from accessor if there is no match.
- */
- template <typename FunctorT>
- void runOverKeysFromValueAccessorIfMatchNotFound(
- ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys,
- FunctorT *functor) const {
- return runOverKeysFromValueAccessor<false>(
- accessor, key_attr_id, check_for_null_keys, functor);
- }
-
- /**
- * @brief Apply the functor to each key without a match in the hash table.
- *
- * @param accessor A ValueAccessor which will be used to access keys.
- * beginIteration() should be called on accessor before calling this
- * method.
- * @param key_attr_id The attribute ID of the keys to be read from accessor.
- * @param check_for_null_keys If true, each key will be checked to see if it
- * is null before looking it up (null keys are skipped). This must be
- * set to true if some of the keys that will be read from accessor may
- * be null.
- * @param functor A pointer to a functor which should provide an operator that
- * takes 1 argument: const ValueAccessor&. The operator will be called
- * only once for a key from accessor if there is no match.
- */
- template <typename FunctorT>
- void runOverKeysFromValueAccessorIfMatchNotFoundCompositeKey(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys,
- FunctorT *functor) const {
- return runOverKeysFromValueAccessorCompositeKey<false>(
- accessor, key_attr_ids, check_for_null_keys, functor);
- }
-
- /**
- * @brief Apply a functor to each key, value pair in this hash table.
- *
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note This version is for single scalar keys. See also
- * forEachCompositeKey().
- *
- * @param functor A pointer to a functor, which should provide a call
- * operator which takes 2 arguments: const TypedValue&, const ValueT&.
- * The call operator will be invoked once on each key, value pair in
- * this hash table (note that if allow_duplicate_keys is true,
- * the call may occur multiple times for the same key with different
- * values).
- * @return The number of key-value pairs visited.
- **/
- template <typename FunctorT>
- std::size_t forEach(FunctorT *functor) const;
-
- /**
- * @brief Apply a functor to each key, value pair in this hash table.
- *
- * @warning This method assumes that no concurrent calls to put(),
- * putCompositeKey(), putValueAccessor(),
- * putValueAccessorCompositeKey(), upsert(), upsertCompositeKey(),
- * upsertValueAccessor(), or upsertValueAccessorCompositeKey() are
- * taking place (i.e. that this HashTable is immutable for the
- * duration of the call and as long as the returned pointer may be
- * dereferenced). Concurrent calls to getSingle(),
- * getSingleCompositeKey(), getAll(), getAllCompositeKey(),
- * getAllFromValueAccessor(), getAllFromValueAccessorCompositeKey(),
- * forEach(), and forEachCompositeKey() are safe.
- * @note This version is for composite keys. See also forEach().
- *
- * @param functor A pointer to a functor, which should provide a call
- * operator which takes 2 arguments: const std::vector<TypedValue>&,
- * const ValueT&. The call operator will be invoked once on each key,
- * value pair in this hash table (note that if allow_duplicate_keys is
- * true, the call may occur multiple times for the same key with
- * different values).
- * @return The number of key-value pairs visited.
- **/
- template <typename FunctorT>
- std::size_t forEachCompositeKeyFast(FunctorT *functor) const;
-
- template <typename FunctorT>
- std::size_t forEachCompositeKeyFast(FunctorT *functor, int index) const;
-
- protected:
- /**
- * @brief Constructor for new resizable hash table.
- *
- * @param key_types A vector of one or more types (>1 indicates a composite
- * key).
- * @param num_entries The estimated number of entries this hash table will
- * hold.
- * @param storage_manager The StorageManager to use (a StorageBlob will be
- * allocated to hold this hash table's contents).
- * @param adjust_hashes If true, the hash of a key should be modified by
- * applying AdjustHash() so that it does not collide with one of the
- * special values kEmptyHash or kPendingHash. If false, the hash is
- * used as-is.
- * @param use_scalar_literal_hash If true, the key is a single scalar literal
- * (non-composite) that it is safe to use the simplified hash function
- * TypedValue::getHashScalarLiteral() on. If false, the generic
- * TypedValue::getHash() method will be used.
- * @param preallocate_supported If true, this HashTable overrides
- * preallocateForBulkInsert() to allow bulk-allocation of resources
- * (i.e. buckets and variable-length key storage) in a single up-front
- * pass when bulk-inserting entries. If false, resources are allocated
- * on the fly for each entry.
- **/
- FastHashTable(const std::vector<const Type *> &key_types,
- const std::size_t num_entries,
- const std::vector<AggregationHandle *> &handles,
- const std::vector<std::size_t> &payload_sizes,
- StorageManager *storage_manager,
- const bool adjust_hashes,
- const bool use_scalar_literal_hash,
- const bool preallocate_supported)
- : key_types_(key_types),
- scalar_key_inline_(true),
- key_inline_(nullptr),
- adjust_hashes_(adjust_hashes),
- use_scalar_literal_hash_(use_scalar_literal_hash),
- preallocate_supported_(preallocate_supported),
- handles_(handles),
- num_handles_(handles.size()),
- total_payload_size_(std::accumulate(
- payload_sizes.begin(), payload_sizes.end(), sizeof(SpinMutex))),
- storage_manager_(storage_manager),
- hash_table_memory_(nullptr),
- hash_table_memory_size_(0) {
- DEBUG_ASSERT(resizable);
- std::size_t running_sum = sizeof(SpinMutex);
- for (auto size : payload_sizes) {
- payload_offsets_.emplace_back(running_sum);
- running_sum += size;
- }
- }
-
- /**
- * @brief Constructor for non-resizable hash table.
- *
- * @param key_types A vector of one or more types (>1 indicates a composite
- * key).
- * @param hash_table_memory A pointer to memory to use for this hash table.
- * @param hash_table_memory_size The size of hash_table_memory in bytes.
- * @param new_hash_table If true, this hash table is being constructed for
- * the first time and hash_table_memory will be cleared. If false,
- * reload a pre-existing hash table.
- * @param hash_table_memory_zeroed If new_hash_table is true, setting this to
- * true means that this HashTable will assume that hash_table_memory
- * has already been zeroed-out (any newly-allocated block or blob
- * memory from StorageManager is zeroed-out). If false, this HashTable
- * will explicitly zero-fill its memory as neccessary. This parameter
- * has no effect when new_hash_table is false.
- * @param adjust_hashes If true, the hash of a key should be modified by
- * applying AdjustHash() so that it does not collide with one of the
- * special values kEmptyHash or kPendingHash. If false, the hash is
- * used as-is.
- * @param use_scalar_literal_hash If true, the key is a single scalar literal
- * (non-composite) that it is safe to use the simplified hash function
- * TypedValue::getHashScalarLiteral() on. If false, the generic
- * TypedValue::getHash() method will be used.
- * @param preallocate_supported If true, this HashTable overrides
- * preallocateForBulkInsert() to allow bulk-allocation of resources
- * (i.e. buckets and variable-length key storage) in a single up-front
- * pass when bulk-inserting entries. If false, resources are allocated
- * on the fly for each entry.
- **/
- FastHashTable(const std::vector<const Type *> &key_types,
- void *hash_table_memory,
- const std::size_t hash_table_memory_size,
- const bool new_hash_table,
- const bool hash_table_memory_zeroed,
- const bool adjust_hashes,
- const bool use_scalar_literal_hash,
- const bool preallocate_supported)
- : key_types_(key_types),
- scalar_key_inline_(true),
- key_inline_(nullptr),
- adjust_hashes_(adjust_hashes),
- use_scalar_literal_hash_(use_scalar_literal_hash),
- preallocate_supported_(preallocate_supported),
- storage_manager_(nullptr),
- hash_table_memory_(hash_table_memory),
- hash_table_memory_size_(hash_table_memory_size) {
- DEBUG_ASSERT(!resizable);
- }
-
- // Adjust 'hash' so that it is not exactly equal to either of the special
- // values kEmptyHash or kPendingHash.
- inline constexpr static std::size_t AdjustHash(const std::size_t hash) {
- return hash + (hash == kEmptyHash) - (hash == kPendingHash);
- }
-
- // Set information about which key components are stored inline. This usually
- // comes from a HashTableKeyManager, and is set by the constructor of a
- // subclass of HashTable.
- inline void setKeyInline(const std::vector<bool> *key_inline) {
- scalar_key_inline_ = key_inline->front();
- key_inline_ = key_inline;
- }
-
- // Generate a hash for a composite key by hashing each component of 'key' and
- // mixing their bits with CombineHashes().
- inline std::size_t hashCompositeKey(const std::vector<TypedValue> &key) const;
-
- // If 'force_key_copy' is true and some part of a composite key is
- // variable-length, calculate the total number of bytes for variable-length
- // key components that need to be copied. Otherwise, return 0 to indicate
- // that no variable-length copy is required.
- inline std::size_t calculateVariableLengthCompositeKeyCopySize(
- const std::vector<TypedValue> &key) const;
-
- // Helpers for put. If this HashTable is resizable, 'resize_shared_mutex_'
- // should be locked in shared mode before calling either of these methods.
- virtual HashTablePutResult putInternal(
- const TypedValue &key,
- const std::size_t variable_key_size,
- const std::uint8_t &value,
- HashTablePreallocationState *prealloc_state) = 0;
-
- virtual HashTablePutResult putCompositeKeyInternalFast(
- const std::vector<TypedValue> &key,
- const std::size_t variable_key_size,
- const std::uint8_t *init_value_ptr,
- HashTablePreallocationState *prealloc_state) = 0;
-
- // Helpers for upsert. Both return a pointer to the value corresponding to
- // 'key'. If this HashTable is resizable, 'resize_shared_mutex_' should be
- // locked in shared mode while calling and using the returned pointer. May
- // return NULL if there is not enough space to insert a new key, in which
- // case a resizable HashTable should release the 'resize_shared_mutex_' and
- // call resize(), then try again.
- virtual std::uint8_t *upsertInternalFast(
- const TypedValue &key,
- const std::size_t variable_key_size,
- const std::uint8_t *init_value_ptr) = 0;
-
- virtual std::uint8_t *upsertCompositeKeyInternalFast(
- const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- const std::size_t variable_key_size) = 0;
-
- // Helpers for forEach. Each return true on success, false if no more entries
- // exist to iterate over. After a successful call, '*key' is overwritten with
- // the key of the next entry, '*value' points to the associated value, and
- // '*entry_num' is incremented to the next (implementation defined) entry to
- // check ('*entry_num' should initially be set to zero).
- virtual bool getNextEntry(TypedValue *key,
- const std::uint8_t **value,
- std::size_t *entry_num) const = 0;
- virtual bool getNextEntryCompositeKey(std::vector<TypedValue> *key,
- const std::uint8_t **value,
- std::size_t *entry_num) const = 0;
-
- // Helpers for getAllFromValueAccessor. Each return true on success, false if
- // no more entries exist for the specified key. After a successful call,
- // '*value' points to the associated value, and '*entry_num' is incremented
- // to the next (implementation defined) entry to check ('*entry_num' should
- // initially be set to zero).
- virtual bool getNextEntryForKey(const TypedValue &key,
- const std::size_t hash_code,
- const std::uint8_t **value,
- std::size_t *entry_num) const = 0;
- virtual bool getNextEntryForCompositeKey(const std::vector<TypedValue> &key,
- const std::size_t hash_code,
- const std::uint8_t **value,
- std::size_t *entry_num) const = 0;
-
- // Return true if key exists in the hash table.
- virtual bool hasKey(const TypedValue &key) const = 0;
- virtual bool hasCompositeKey(const std::vector<TypedValue> &key) const = 0;
-
- // For a resizable HashTable, grow to accomodate more entries. If
- // 'extra_buckets' is not zero, it may serve as a "hint" to implementations
- // that at least the requested number of extra buckets are required when
- // resizing (mainly used in putValueAccessor() and
- // putValueAccessorCompositeKey() when 'preallocate_supported_' is true).
- // Implementations are free to ignore 'extra_buckets'. If
- // 'extra_variable_storage' is not zero, implementations will attempt to
- // allocate at least enough additional variable-key storage space to
- // accomodate the number of bytes specified. 'retry_num' is intended ONLY for
- // when resize() recursively calls itself and should not be set to nonzero by
- // any other caller.
- virtual void resize(const std::size_t extra_buckets,
- const std::size_t extra_variable_storage,
- const std::size_t retry_num = 0) = 0;
-
- // In the case where 'allow_duplicate_keys' is true, it is possible to
- // pre-calculate the number of key-value entries and the amount of
- // variable-length key storage that will be needed to insert all the
- // entries from a ValueAccessor in putValueAccessor() or
- // putValueAccessorCompositeKey() before actually inserting anything. Some
- // HashTable implemetations (notably SeparateChainingHashTable) can achieve
- // better performance by ammortizing the cost of allocating certain resources
- // (buckets and variable-length key storage) in one up-front allocation. This
- // method is intended to support that. Returns true and fills in
- // '*prealloc_state' if pre-allocation was successful. Returns false if a
- // resize() is needed.
- virtual bool preallocateForBulkInsert(
- const std::size_t total_entries,
- const std::size_t total_variable_key_size,
- HashTablePreallocationState *prealloc_state) {
- FATAL_ERROR(
- "Called HashTable::preallocateForBulkInsert() on a HashTable "
- "implementation that does not support preallocation.");
- }
-
- // Type(s) of keys.
- const std::vector<const Type *> key_types_;
-
- // Information about whether key components are stored inline or in a
- // separate variable-length storage region. This is usually determined by a
- // HashTableKeyManager and set by calling setKeyInline().
- bool scalar_key_inline_;
- const std::vector<bool> *key_inline_;
-
- // Whether hashes should be adjusted by AdjustHash() before being used.
- const bool adjust_hashes_;
- // Whether it is safe to use the simplified TypedValue::getHashScalarLiteral()
- // method instead of the generic TypedValue::getHash() method.
- const bool use_scalar_literal_hash_;
- // Whether preallocateForBulkInsert() is supported by this HashTable.
- const bool preallocate_supported_;
-
- const std::vector<AggregationHandle *> handles_;
- const unsigned int num_handles_;
- const std::size_t total_payload_size_;
- std::vector<std::size_t> payload_offsets_;
-
- // Used only when resizable is true:
- StorageManager *storage_manager_;
- MutableBlobReference blob_;
- // Locked in shared mode for most operations, exclusive mode during resize.
- // Not locked at all for non-resizable HashTables.
- alignas(kCacheLineBytes) SpinSharedMutex<true> resize_shared_mutex_;
-
- // Used only when resizable is false:
- void *hash_table_memory_;
- const std::size_t hash_table_memory_size_;
-
- private:
- // Assign '*key_vector' with the attribute values specified by 'key_attr_ids'
- // at the current position of 'accessor'. If 'check_for_null_keys' is true,
- // stops and returns true if any of the values is null, otherwise returns
- // false.
- template <typename ValueAccessorT>
- inline static bool GetCompositeKeyFromValueAccessor(
- const ValueAccessorT &accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys,
- std::vector<TypedValue> *key_vector) {
- for (std::vector<attribute_id>::size_type key_idx = 0;
- key_idx < key_attr_ids.size();
- ++key_idx) {
- (*key_vector)[key_idx] = accessor.getTypedValue(key_attr_ids[key_idx]);
- if (check_for_null_keys && (*key_vector)[key_idx].isNull()) {
- return true;
- }
- }
- return false;
- }
-
- // If run_if_match_found is true, apply the functor to each key if a match is
- // found; otherwise, apply the functor if no match is found.
- template <bool run_if_match_found, typename FunctorT>
- void runOverKeysFromValueAccessor(ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys,
- FunctorT *functor) const;
-
- template <bool run_if_match_found, typename FunctorT>
- void runOverKeysFromValueAccessorCompositeKey(
- ValueAccessor *accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys,
- FunctorT *functor) const;
-
- // Method containing the actual logic implementing getAllFromValueAccessor().
- // Has extra template parameters that control behavior to avoid some
- // inner-loop branching.
- template <typename FunctorT,
- bool check_for_null_keys,
- bool adjust_hashes_template,
- bool use_scalar_literal_hash_template>
- void getAllFromValueAccessorImpl(ValueAccessor *accessor,
- const attribute_id key_attr_id,
- FunctorT *functor) const;
-
- DISALLOW_COPY_AND_ASSIGN(FastHashTable);
-};
-
-/**
- * @brief An instantiation of the HashTable template for use in aggregations.
- * @note This has force_key_copy = true, so that we don't have dangling pointers
- * to blocks that are evicted.
- **/
-using AggregationStateFastHashTable = FastHashTable<true, false, true, false>;
-
-/** @} */
-
-// ----------------------------------------------------------------------------
-// Implementations of template class methods follow.
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-HashTablePutResult
-FastHashTable<resizable, serializable, force_key_copy, allow_duplicate_keys>::
- put(const TypedValue &key, const std::uint8_t &value) {
- const std::size_t variable_size =
- (force_key_copy && !scalar_key_inline_) ? key.getDataSize() : 0;
- if (resizable) {
- HashTablePutResult result = HashTablePutResult::kOutOfSpace;
- while (result == HashTablePutResult::kOutOfSpace) {
- {
- SpinSharedMutexSharedLock<true> lock(resize_shared_mutex_);
- result = putInternal(key, variable_size, value, nullptr);
- }
- if (result == HashTablePutResult::kOutOfSpace) {
- resize(0, variable_size);
- }
- }
- return result;
- } else {
- return putInternal(key, variable_size, value, nullptr);
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-HashTablePutResult
-FastHashTable<resizable, serializable, force_key_copy, allow_duplicate_keys>::
- putCompositeKey(const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr) {
- const std::size_t variable_size =
- calculateVariableLengthCompositeKeyCopySize(key);
- if (resizable) {
- HashTablePutResult result = HashTablePutResult::kOutOfSpace;
- while (result == HashTablePutResult::kOutOfSpace) {
- {
- SpinSharedMutexSharedLock<true> lock(resize_shared_mutex_);
- result = putCompositeKeyInternalFast(
- key, variable_size, init_value_ptr, nullptr);
- }
- if (result == HashTablePutResult::kOutOfSpace) {
- resize(0, variable_size);
- }
- }
- return result;
- } else {
- return putCompositeKeyInternalFast(
- key, variable_size, init_value_ptr, nullptr);
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-template <typename FunctorT>
-HashTablePutResult
-FastHashTable<resizable, serializable, force_key_copy, allow_duplicate_keys>::
- putValueAccessor(ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys,
- FunctorT *functor) {
- HashTablePutResult result = HashTablePutResult::kOutOfSpace;
- std::size_t variable_size;
- HashTablePreallocationState prealloc_state;
- bool using_prealloc = allow_duplicate_keys && preallocate_supported_;
- return InvokeOnAnyValueAccessor(
- accessor,
- [&](auto *accessor) -> HashTablePutResult { // NOLINT(build/c++11)
- if (using_prealloc) {
- std::size_t total_entries = 0;
- std::size_t total_variable_key_size = 0;
- if (check_for_null_keys || (force_key_copy && !scalar_key_inline_)) {
- // If we need to filter out nulls OR make variable copies, make a
- // prepass over the ValueAccessor.
- while (accessor->next()) {
- TypedValue key = accessor->getTypedValue(key_attr_id);
- if (check_for_null_keys && key.isNull()) {
- continue;
- }
- ++total_entries;
- total_variable_key_size += (force_key_copy && !scalar_key_inline_)
- ? key.getDataSize()
- : 0;
- }
- accessor->beginIteration();
- } else {
- total_entries = accessor->getNumTuples();
- }
- if (resizable) {
- bool prealloc_succeeded = false;
- while (!prealloc_succeeded) {
- {
- SpinSharedMutexSharedLock<true> lock(resize_shared_mutex_);
- prealloc_succeeded = this->preallocateForBulkInsert(
- total_entries, total_variable_key_size, &prealloc_state);
- }
- if (!prealloc_succeeded) {
- this->resize(total_entries, total_variable_key_size);
- }
- }
- } else {
- using_prealloc = this->preallocateForBulkInsert(
- total_entries, total_variable_key_size, &prealloc_state);
- }
- }
- if (resizable) {
- while (result == HashTablePutResult::kOutOfSpace) {
- {
- result = HashTablePutResult::kOK;
- SpinSharedMutexSharedLock<true> lock(resize_shared_mutex_);
- while (accessor->next()) {
- TypedValue key = accessor->getTypedValue(key_attr_id);
- if (check_for_null_keys && key.isNull()) {
- continue;
- }
- variable_size = (force_key_copy && !scalar_key_inline_)
- ? key.getDataSize()
- : 0;
- result = this->putInternal(
- key,
- variable_size,
- (*functor)(*accessor),
- using_prealloc ? &prealloc_state : nullptr);
- if (result == HashTablePutResult::kDuplicateKey) {
- DEBUG_ASSERT(!using_prealloc);
- return result;
- } else if (result == HashTablePutResult::kOutOfSpace) {
- DEBUG_ASSERT(!using_prealloc);
- break;
- }
- }
- }
- if (result == HashTablePutResult::kOutOfSpace) {
- this->resize(0, variable_size);
- accessor->previous();
- }
- }
- } else {
- while (accessor->next()) {
- TypedValue key = accessor->getTypedValue(key_attr_id);
- if (check_for_null_keys && key.isNull()) {
- continue;
- }
- variable_size =
- (force_key_copy && !scalar_key_inline_) ? key.getDataSize() : 0;
- result =
- this->putInternal(key,
- variable_size,
- (*functor)(*accessor),
- using_prealloc ? &prealloc_state : nullptr);
- if (result != HashTablePutResult::kOK) {
- return result;
- }
- }
- }
-
- return HashTablePutResult::kOK;
- });
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-template <typename FunctorT>
-HashTablePutResult
-FastHashTable<resizable, serializable, force_key_copy, allow_duplicate_keys>::
- putValueAccessorCompositeKey(ValueAccessor *accessor,
- const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys,
- FunctorT *functor) {
- DEBUG_ASSERT(key_types_.size() == key_attr_ids.size());
- HashTablePutResult result = HashTablePutResult::kOutOfSpace;
- std::size_t variable_size;
- HashTablePreallocationState prealloc_state;
- bool using_prealloc = allow_duplicate_keys && preallocate_supported_;
- std::vector<TypedValue> key_vector;
- key_vector.resize(key_attr_ids.size());
- return InvokeOnAnyValueAccessor(
- accessor,
- [&](auto *accessor) -> HashTablePutResult { // NOLINT(build/c++11)
- if (using_prealloc) {
- std::size_t total_entries = 0;
- std::size_t total_variable_key_size = 0;
- if (check_for_null_keys || force_key_copy) {
- // If we need to filter out nulls OR make variable copies, make a
- // prepass over the ValueAccessor.
- while (accessor->next()) {
- if (this->GetCompositeKeyFromValueAccessor(*accessor,
- key_attr_ids,
- check_for_null_keys,
- &key_vector)) {
- continue;
- }
- ++total_entries;
- total_variable_key_size +=
- this->calculateVariableLengthCompositeKeyCopySize(key_vector);
- }
- accessor->beginIteration();
- } else {
- total_entries = accessor->getNumTuples();
- }
- if (resizable) {
- bool prealloc_succeeded = false;
- while (!prealloc_succeeded) {
- {
- SpinSharedMutexSharedLock<true> lock(resize_shared_mutex_);
- prealloc_succeeded = this->preallocateForBulkInsert(
- total_entries, total_variable_key_size, &prealloc_state);
- }
- if (!prealloc_succeeded) {
- this->resize(total_entries, total_variable_key_size);
- }
- }
- } else {
- using_prealloc = this->preallocateForBulkInsert(
- total_entries, total_variable_key_size, &prealloc_state);
- }
- }
- if (resizable) {
- while (result == HashTablePutResult::kOutOfSpace) {
- {
- result = HashTablePutResult::kOK;
- SpinSharedMutexSharedLock<true> lock(resize_shared_mutex_);
- while (accessor->next()) {
- if (this->GetCompositeKeyFromValueAccessor(*accessor,
- key_attr_ids,
- check_for_null_keys,
- &key_vector)) {
- continue;
- }
- variable_size =
- this->calculateVariableLengthCompositeKeyCopySize(
- key_vector);
- result = this->putCompositeKeyInternal(
- key_vector,
- variable_size,
- (*functor)(*accessor),
- using_prealloc ? &prealloc_state : nullptr);
- if (result == HashTablePutResult::kDuplicateKey) {
- DEBUG_ASSERT(!using_prealloc);
- return result;
- } else if (result == HashTablePutResult::kOutOfSpace) {
- DEBUG_ASSERT(!using_prealloc);
- break;
- }
- }
- }
- if (result == HashTablePutResult::kOutOfSpace) {
- this->resize(0, variable_size);
- accessor->previous();
- }
- }
- } else {
- while (accessor->next()) {
- if (this->GetCompositeKeyFromValueAccessor(*accessor,
- key_attr_ids,
- check_for_null_keys,
- &key_vector)) {
- continue;
- }
- variable_size =
- this->calculateVariableLengthCompositeKeyCopySize(key_vector);
- result = this->putCompositeKeyInternal(
- key_vector,
- variable_size,
- (*functor)(*accessor),
- using_prealloc ? &prealloc_state : nullptr);
- if (result != HashTablePutResult::kOK) {
- return result;
- }
- }
- }
-
- return HashTablePutResult::kOK;
- });
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-template <typename FunctorT>
-bool FastHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::upsert(const TypedValue &key,
- const std::uint8_t
- *initial_value_ptr,
- FunctorT *functor) {
- DEBUG_ASSERT(!allow_duplicate_keys);
- const std::size_t variable_size =
- (force_key_copy && !scalar_key_inline_) ? key.getDataSize() : 0;
- if (resizable) {
- for (;;) {
- {
- SpinSharedMutexSharedLock<true> resize_lock(resize_shared_mutex_);
- std::uint8_t *value =
- upsertInternalFast(key, variable_size, initial_value_ptr);
- if (value != nullptr) {
- (*functor)(value);
- return true;
- }
- }
- resize(0, force_key_copy && !scalar_key_inline_ ? key.getDataSize() : 0);
- }
- } else {
- std::uint8_t *value =
- upsertInternalFast(key, variable_size, initial_value_ptr);
- if (value == nullptr) {
- return false;
- } else {
- (*functor)(value);
- return true;
- }
- }
-}
-
-class HashTableMergerFast {
- public:
- /**
- * @brief Constructor
- *
- * @param handle The Aggregation handle being used.
- * @param destination_hash_table The destination hash table to which other
- * hash tables will be merged.
- **/
- explicit HashTableMergerFast(
- AggregationStateHashTableBase *destination_hash_table)
- : destination_hash_table_(
- static_cast<FastHashTable<true, false, true, false> *>(
- destination_hash_table)) {}
-
- /**
- * @brief The operator for the functor.
- *
- * @param group_by_key The group by key being merged.
- * @param source_state The aggregation state for the given key in the source
- * aggregation hash table.
- **/
- inline void operator()(const std::vector<TypedValue> &group_by_key,
- const std::uint8_t *source_state) {
- const std::uint8_t *original_state =
- destination_hash_table_->getSingleCompositeKey(group_by_key);
- if (original_state != nullptr) {
- // The CHECK is required as upsertCompositeKey can return false if the
- // hash table runs out of space during the upsert process. The ideal
- // solution will be to retry again if the upsert fails.
- CHECK(destination_hash_table_->upsertCompositeKeyFast(
- group_by_key, original_state, source_state));
- } else {
- destination_hash_table_->putCompositeKey(group_by_key, source_state);
- }
- }
-
- private:
- FastHashTable<true, false, true, false> *destination_hash_table_;
-
- DISALLOW_COPY_AND_ASSIGN(HashTableMergerFast);
-};
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-template <typename FunctorT>
-bool FastHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- upsertCompositeKeyFast(const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- FunctorT *functor) {
- DEBUG_ASSERT(!allow_duplicate_keys);
- const std::size_t variable_size =
- calculateVariableLengthCompositeKeyCopySize(key);
- if (resizable) {
- for (;;) {
- {
- SpinSharedMutexSharedLock<true> resize_lock(resize_shared_mutex_);
- std::uint8_t *value =
- upsertCompositeKeyInternalFast(key, init_value_ptr, variable_size);
- if (value != nullptr) {
- (*functor)(value);
- return true;
- }
- }
- resize(0, variable_size);
- }
- } else {
- std::uint8_t *value =
- upsertCompositeKeyInternalFast(key, init_value_ptr, variable_size);
- if (value == nullptr) {
- return false;
- } else {
- (*functor)(value);
- return true;
- }
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-template <typename FunctorT>
-bool FastHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- upsertCompositeKeyFast(const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- FunctorT *functor,
- int index) {
- DEBUG_ASSERT(!allow_duplicate_keys);
- const std::size_t variable_size =
- calculateVariableLengthCompositeKeyCopySize(key);
- if (resizable) {
- for (;;) {
- {
- SpinSharedMutexSharedLock<true> resize_lock(resize_shared_mutex_);
- std::uint8_t *value =
- upsertCompositeKeyInternalFast(key, init_value_ptr, variable_size);
- if (value != nullptr) {
- (*functor)(value + payload_offsets_[index]);
- return true;
- }
- }
- resize(0, variable_size);
- }
- } else {
- std::uint8_t *value =
- upsertCompositeKeyInternalFast(key, init_value_ptr, variable_size);
- if (value == nullptr) {
- return false;
- } else {
- (*functor)(value + payload_offsets_[index]);
- return true;
- }
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- upsertCompositeKeyFast(const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- const std::uint8_t *source_state) {
- DEBUG_ASSERT(!allow_duplicate_keys);
- const std::size_t variable_size =
- calculateVariableLengthCompositeKeyCopySize(key);
- if (resizable) {
- for (;;) {
- {
- SpinSharedMutexSharedLock<true> resize_lock(resize_shared_mutex_);
- std::uint8_t *value =
- upsertCompositeKeyInternalFast(key, init_value_ptr, variable_size);
- if (value != nullptr) {
- SpinMutexLock lock(*(reinterpret_cast<SpinMutex *>(value)));
- for (unsigned int k = 0; k < num_handles_; ++k) {
- handles_[k]->mergeStatesFast(source_state + payload_offsets_[k],
- value + payload_offsets_[k]);
- }
- return true;
- }
- }
- resize(0, variable_size);
- }
- } else {
- std::uint8_t *value =
- upsertCompositeKeyInternalFast(key, init_value_ptr, variable_size);
- if (value == nullptr) {
- return false;
- } else {
- SpinMutexLock lock(*(reinterpret_cast<SpinMutex *>(value)));
- for (unsigned int k = 0; k < num_handles_; ++k) {
- handles_[k]->mergeStatesFast(source_state + payload_offsets_[k],
- value + payload_offsets_[k]);
- }
- return true;
- }
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- upsertValueAccessorFast(
- const std::vector<attribute_id> &argument_ids,
- ValueAccessor *accessor,
- const attribute_id key_attr_id,
- const bool check_for_null_keys) {
- DEBUG_ASSERT(!allow_duplicate_keys);
- std::size_t variable_size;
- return InvokeOnAnyValueAccessor(
- accessor,
- [&](auto *accessor) -> bool { // NOLINT(build/c++11)
- if (resizable) {
- bool continuing = true;
- while (continuing) {
- {
- continuing = false;
- SpinSharedMutexSharedLock<true> lock(resize_shared_mutex_);
- while (accessor->next()) {
- TypedValue key = accessor->getTypedValue(key_attr_id);
- if (check_for_null_keys && key.isNull()) {
- continue;
- }
- variable_size = (force_key_copy && !scalar_key_inline_)
- ? key.getDataSize()
- : 0;
- std::uint8_t *value =
- this->upsertInternalFast(key, variable_size, nullptr);
- if (value == nullptr) {
- continuing = true;
- break;
- } else {
- SpinMutexLock lock(*(reinterpret_cast<SpinMutex *>(value)));
- for (unsigned int k = 0; k < num_handles_; ++k) {
- if (argument_ids[k] != kInvalidAttributeID) {
- handles_[k]->updateStateUnary(
- accessor->getTypedValue(argument_ids[k]),
- value + payload_offsets_[k]);
- } else {
- handles_[k]->updateStateNullary(value +
- payload_offsets_[k]);
- }
- }
- }
- }
- }
- if (continuing) {
- this->resize(0, variable_size);
- accessor->previous();
- }
- }
- } else {
- while (accessor->next()) {
- TypedValue key = accessor->getTypedValue(key_attr_id);
- if (check_for_null_keys && key.isNull()) {
- continue;
- }
- variable_size =
- (force_key_copy && !scalar_key_inline_) ? key.getDataSize() : 0;
- std::uint8_t *value =
- this->upsertInternalFast(key, variable_size, nullptr);
- if (value == nullptr) {
- return false;
- } else {
- SpinMutexLock lock(*(reinterpret_cast<SpinMutex *>(value)));
- for (unsigned int k = 0; k < num_handles_; ++k) {
- if (argument_ids[k] != kInvalidAttributeID) {
- handles_[k]->updateStateUnary(
- accessor->getTypedValue(argument_ids[k]),
- value + payload_offsets_[k]);
- } else {
- handles_[k]->updateStateNullary(value +
-
<TRUNCATED>
[2/5] incubator-quickstep git commit: Cleanup FastHashTable + Avoid
unnecessary copying of data before aggregation.
Posted by ji...@apache.org.
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/FastHashTableFactory.hpp
----------------------------------------------------------------------
diff --git a/storage/FastHashTableFactory.hpp b/storage/FastHashTableFactory.hpp
deleted file mode 100644
index 682cc2a..0000000
--- a/storage/FastHashTableFactory.hpp
+++ /dev/null
@@ -1,224 +0,0 @@
-/**
- * Copyright 2015-2016 Pivotal Software, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- **/
-
-#ifndef QUICKSTEP_STORAGE_FAST_HASH_TABLE_FACTORY_HPP_
-#define QUICKSTEP_STORAGE_FAST_HASH_TABLE_FACTORY_HPP_
-
-#include <cstddef>
-#include <string>
-#include <vector>
-
-#include "storage/HashTable.hpp"
-#include "storage/FastHashTable.hpp"
-#include "storage/HashTableBase.hpp"
-#include "storage/HashTableFactory.hpp"
-#include "storage/HashTable.pb.h"
-#include "storage/LinearOpenAddressingHashTable.hpp"
-#include "storage/SeparateChainingHashTable.hpp"
-#include "storage/FastSeparateChainingHashTable.hpp"
-#include "storage/SimpleScalarSeparateChainingHashTable.hpp"
-#include "storage/TupleReference.hpp"
-#include "types/TypeFactory.hpp"
-#include "utility/Macros.hpp"
-
-#include "glog/logging.h"
-
-namespace quickstep {
-
-class StorageManager;
-class Type;
-
-/** \addtogroup Storage
- * @{
- */
-
-/**
- * @brief Templated all-static factory class that makes it easier to
- * instantiate HashTables with the particular HashTable implementation
- * chosen at runtime. All template parameters are exactly the same as
- * those of HashTable.
- **/
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-class FastHashTableFactory {
- public:
- /**
- * @brief Create a new resizable HashTable, with the type selected by
- * hash_table_type. Other parameters are forwarded to the HashTable's
- * constructor.
- *
- * @param hash_table_type The specific HashTable implementation that should
- * be used.
- * @param key_types A vector of one or more types (>1 indicates a composite
- * key). Forwarded as-is to the HashTable's constructor.
- * @param num_entries The estimated number of entries the HashTable will
- * hold. Forwarded as-is to the HashTable's constructor.
- * @param payload_sizes The sizes in bytes for the AggregationStates for the
- * respective AggregationHandles.
- * @param handles The AggregationHandles used in this HashTable.
- * @param storage_manager The StorageManager to use (a StorageBlob will be
- * allocated to hold the HashTable's contents). Forwarded as-is to the
- * HashTable's constructor.
- * @return A new resizable HashTable.
- **/
- static FastHashTable<resizable, serializable, force_key_copy, allow_duplicate_keys>*
- CreateResizable(const HashTableImplType hash_table_type,
- const std::vector<const Type*> &key_types,
- const std::size_t num_entries,
- const std::vector<std::size_t> &payload_sizes,
- const std::vector<AggregationHandle *> &handles,
- StorageManager *storage_manager) {
- DCHECK(resizable);
-
- switch (hash_table_type) {
- case HashTableImplType::kSeparateChaining:
- return new FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>(key_types, num_entries, payload_sizes, handles, storage_manager);
- default: {
- LOG(FATAL) << "Unrecognized HashTableImplType in HashTableFactory::createResizable()\n";
- }
- }
- }
-
- /**
- * @brief Create a new fixed-sized HashTable, with the type selected by
- * hash_table_type. Other parameters are forwarded to the HashTables's
- * constructor.
- *
- * @param hash_table_type The specific HashTable implementation that should
- * be used.
- * @param key_types A vector of one or more types (>1 indicates a composite
- * key). Forwarded as-is to the HashTable's constructor.
- * @param hash_table_memory A pointer to memory to use for the HashTable.
- * Forwarded as-is to the HashTable's constructor.
- * @param hash_table_memory_size The size of hash_table_memory in bytes.
- * Forwarded as-is to the HashTable's constructor.
- * @param new_hash_table If true, the HashTable is being constructed for the
- * first time and hash_table_memory will be cleared. If false, reload
- * a pre-existing HashTable. Forwarded as-is to the HashTable's
- * constructor.
- * @param hash_table_memory_zeroed If new_hash_table is true, setting this to
- * true means that the HashTable will assume that hash_table_memory
- * has already been zeroed-out (any newly-allocated block or blob
- * memory from StorageManager is zeroed-out). If false, the HashTable
- * will explicitly zero-fill its memory as neccessary. This parameter
- * has no effect when new_hash_table is false. Forwarded as-is to the
- * HashTable's constructor.
- * @return A new (or reloaded) fixed-size HashTable.
- **/
- static FastHashTable<resizable, serializable, force_key_copy, allow_duplicate_keys>*
- CreateFixedSize(const HashTableImplType hash_table_type,
- const std::vector<const Type*> &key_types,
- void *hash_table_memory,
- const std::size_t hash_table_memory_size,
- const bool new_hash_table,
- const bool hash_table_memory_zeroed) {
- DCHECK(!resizable);
-
- switch (hash_table_type) {
- case HashTableImplType::kSeparateChaining:
- return new SeparateChainingHashTable<
- int,
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>(key_types,
- hash_table_memory,
- hash_table_memory_size,
- new_hash_table,
- hash_table_memory_zeroed);
- default: {
- LOG(FATAL) << "Unrecognized HashTableImplType\n";
- }
- }
- }
-
- /**
- * @brief Check whether a serialization::HashTable describing a resizable
- * HashTable is fully-formed and all parts are valid.
- *
- * @param proto A serialized Protocol Buffer description of a HashTable,
- * originally generated by the optimizer.
- * @return Whether proto is fully-formed and valid.
- **/
- static bool ProtoIsValid(const serialization::HashTable &proto) {
- if (!proto.IsInitialized() ||
- !serialization::HashTableImplType_IsValid(
- proto.hash_table_impl_type())) {
- return false;
- }
-
- for (int i = 0; i < proto.key_types_size(); ++i) {
- if (!TypeFactory::ProtoIsValid(proto.key_types(i))) {
- return false;
- }
- }
-
- return true;
- }
-
- /**
- * @brief Create a new resizable HashTable according to a protobuf
- * description.
- *
- * @param proto A protobuf description of a resizable HashTable.
- * @param storage_manager The StorageManager to use (a StorageBlob will be
- * allocated to hold the HashTable's contents).
- * @return A new resizable HashTable with parameters specified by proto.
- **/
- static FastHashTable<resizable, serializable, force_key_copy, allow_duplicate_keys>*
- CreateResizableFromProto(const serialization::HashTable &proto,
- StorageManager *storage_manager) {
- DCHECK(ProtoIsValid(proto))
- << "Attempted to create HashTable from invalid proto description:\n"
- << proto.DebugString();
-
- std::vector<const Type*> key_types;
- for (int i = 0; i < proto.key_types_size(); ++i) {
- key_types.emplace_back(&TypeFactory::ReconstructFromProto(proto.key_types(i)));
- }
-
- auto hash_table = CreateResizable(HashTableImplTypeFromProto(proto.hash_table_impl_type()),
- key_types,
- proto.estimated_num_entries(),
- storage_manager);
- return hash_table;
- }
-
- private:
- // Class is all-static and should not be instantiated.
- FastHashTableFactory();
-
- DISALLOW_COPY_AND_ASSIGN(FastHashTableFactory);
-};
-
-/**
- * @brief Convenient alias that provides a HashTableFactory whose only template
- * parameter is the aggregate state type.
- **/
-using AggregationStateFastHashTableFactory
- = FastHashTableFactory<true, false, true, false>;
-
-/** @} */
-
-} // namespace quickstep
-
-#endif // QUICKSTEP_STORAGE_HASH_TABLE_FACTORY_HPP_
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/FastSeparateChainingHashTable.hpp
----------------------------------------------------------------------
diff --git a/storage/FastSeparateChainingHashTable.hpp b/storage/FastSeparateChainingHashTable.hpp
deleted file mode 100644
index 2435d45..0000000
--- a/storage/FastSeparateChainingHashTable.hpp
+++ /dev/null
@@ -1,1551 +0,0 @@
-/**
- * Copyright 2011-2015 Quickstep Technologies LLC.
- * Copyright 2015-2016 Pivotal Software, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- **/
-
-#ifndef QUICKSTEP_STORAGE_FAST_SEPARATE_CHAINING_HASH_TABLE_HPP_
-#define QUICKSTEP_STORAGE_FAST_SEPARATE_CHAINING_HASH_TABLE_HPP_
-
-#include <algorithm>
-#include <atomic>
-#include <cstddef>
-#include <cstring>
-#include <limits>
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "expressions/aggregation/AggregationHandle.hpp"
-#include "storage/FastHashTable.hpp"
-#include "storage/HashTable.hpp"
-#include "storage/HashTableBase.hpp"
-#include "storage/HashTableKeyManager.hpp"
-#include "storage/StorageBlob.hpp"
-#include "storage/StorageBlockInfo.hpp"
-#include "storage/StorageConstants.hpp"
-#include "storage/StorageManager.hpp"
-#include "threading/SpinSharedMutex.hpp"
-#include "types/Type.hpp"
-#include "types/TypedValue.hpp"
-#include "utility/Alignment.hpp"
-#include "utility/Macros.hpp"
-#include "utility/PrimeNumber.hpp"
-
-namespace quickstep {
-
-/** \addtogroup Storage
- * @{
- */
-
-/**
- * @brief A hash table implementation which uses separate chaining for buckets.
- **/
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-class FastSeparateChainingHashTable
- : public FastHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys> {
- public:
- FastSeparateChainingHashTable(const std::vector<const Type *> &key_types,
- const std::size_t num_entries,
- const std::vector<std::size_t> &payload_sizes,
- const std::vector<AggregationHandle *> &handles,
- StorageManager *storage_manager);
-
- ~FastSeparateChainingHashTable() override {
- std::free(init_payload_);
- }
-
- void clear() override;
-
- std::size_t numEntries() const override {
- return header_->buckets_allocated.load(std::memory_order_relaxed);
- }
-
- const std::uint8_t* getSingle(const TypedValue &key) const override;
- const std::uint8_t* getSingleCompositeKey(
- const std::vector<TypedValue> &key) const override;
- const std::uint8_t* getSingleCompositeKey(const std::vector<TypedValue> &key,
- int index) const override;
-
- void getAll(const TypedValue &key,
- std::vector<const std::uint8_t *> *values) const override;
- void getAllCompositeKey(
- const std::vector<TypedValue> &key,
- std::vector<const std::uint8_t *> *values) const override;
-
- protected:
- HashTablePutResult putInternal(
- const TypedValue &key,
- const std::size_t variable_key_size,
- const std::uint8_t &value,
- HashTablePreallocationState *prealloc_state) override;
-
- HashTablePutResult putCompositeKeyInternalFast(
- const std::vector<TypedValue> &key,
- const std::size_t variable_key_size,
- const std::uint8_t *init_value_ptr,
- HashTablePreallocationState *prealloc_state) override;
-
- std::uint8_t* upsertInternalFast(const TypedValue &key,
- const std::size_t variable_key_size,
- const std::uint8_t *init_value_ptr) override;
-
- std::uint8_t* upsertCompositeKeyInternalFast(
- const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- const std::size_t variable_key_size) override;
-
- bool getNextEntry(TypedValue *key,
- const std::uint8_t **value,
- std::size_t *entry_num) const override;
- bool getNextEntryCompositeKey(std::vector<TypedValue> *key,
- const std::uint8_t **value,
- std::size_t *entry_num) const override;
-
- bool getNextEntryForKey(const TypedValue &key,
- const std::size_t hash_code,
- const std::uint8_t **value,
- std::size_t *entry_num) const override;
- bool getNextEntryForCompositeKey(const std::vector<TypedValue> &key,
- const std::size_t hash_code,
- const std::uint8_t **value,
- std::size_t *entry_num) const override;
-
- bool hasKey(const TypedValue &key) const override;
- bool hasCompositeKey(const std::vector<TypedValue> &key) const override;
-
- void resize(const std::size_t extra_buckets,
- const std::size_t extra_variable_storage,
- const std::size_t retry_num = 0) override;
-
- bool preallocateForBulkInsert(
- const std::size_t total_entries,
- const std::size_t total_variable_key_size,
- HashTablePreallocationState *prealloc_state) override;
-
- void destroyPayload() override {
- const std::size_t num_buckets =
- header_->buckets_allocated.load(std::memory_order_relaxed);
- void *bucket_ptr = static_cast<char *>(buckets_) + kValueOffset;
- for (std::size_t bucket_num = 0; bucket_num < num_buckets; ++bucket_num) {
- for (std::size_t handle_id = 0; handle_id < num_handles_; ++handle_id) {
- void *value_internal_ptr =
- static_cast<char *>(bucket_ptr) + this->payload_offsets_[handle_id];
- handles_[handle_id]->destroyPayload(static_cast<std::uint8_t *>(value_internal_ptr));
- }
- bucket_ptr = static_cast<char *>(bucket_ptr) + bucket_size_;
- }
- }
-
- private:
- struct Header {
- std::size_t num_slots;
- std::size_t num_buckets;
- alignas(kCacheLineBytes) std::atomic<std::size_t> buckets_allocated;
- alignas(kCacheLineBytes)
- std::atomic<std::size_t> variable_length_bytes_allocated;
- };
-
- std::uint8_t *init_payload_;
- std::size_t kBucketAlignment;
-
- // Value's offset in a bucket is the first alignof(ValueT) boundary after the
- // next pointer and hash code.
- std::size_t kValueOffset;
-
- // Round bucket size up to a multiple of kBucketAlignment.
- constexpr std::size_t ComputeBucketSize(const std::size_t fixed_key_size) {
- return (((kValueOffset + this->total_payload_size_ + fixed_key_size - 1) /
- kBucketAlignment) +
- 1) *
- kBucketAlignment;
- }
-
- // Attempt to find an empty bucket to insert 'hash_code' into, starting after
- // '*bucket' in the chain (or, if '*bucket' is NULL, starting from the slot
- // array). Returns true and stores SIZE_T_MAX in '*pending_chain_ptr' if an
- // empty bucket is found. Returns false if 'allow_duplicate_keys' is false
- // and a hash collision is found (caller should then check whether there is a
- // genuine key collision or the hash collision is spurious). Returns false
- // and sets '*bucket' to NULL if there are no more empty buckets in the hash
- // table. If 'variable_key_allocation_required' is nonzero, this method will
- // attempt to allocate storage for a variable-length key BEFORE allocating a
- // bucket, so that no bucket number below 'header_->num_buckets' is ever
- // deallocated after being allocated.
- inline bool locateBucketForInsertion(
- const std::size_t hash_code,
- const std::size_t variable_key_allocation_required,
- void **bucket,
- std::atomic<std::size_t> **pending_chain_ptr,
- std::size_t *pending_chain_ptr_finish_value,
- HashTablePreallocationState *prealloc_state);
-
- // Write a scalar 'key' and its 'hash_code' into the '*bucket', which was
- // found by locateBucketForInsertion(). Assumes that storage for a
- // variable-length key copy (if any) was already allocated by a successful
- // call to allocateVariableLengthKeyStorage().
- inline void writeScalarKeyToBucket(
- const TypedValue &key,
- const std::size_t hash_code,
- void *bucket,
- HashTablePreallocationState *prealloc_state);
-
- // Write a composite 'key' and its 'hash_code' into the '*bucket', which was
- // found by locateBucketForInsertion(). Assumes that storage for
- // variable-length key copies (if any) was already allocated by a successful
- // call to allocateVariableLengthKeyStorage().
- inline void writeCompositeKeyToBucket(
- const std::vector<TypedValue> &key,
- const std::size_t hash_code,
- void *bucket,
- HashTablePreallocationState *prealloc_state);
-
- // Determine whether it is actually necessary to resize this hash table.
- // Checks that there is at least one unallocated bucket, and that there is
- // at least 'extra_variable_storage' bytes of variable-length storage free.
- bool isFull(const std::size_t extra_variable_storage) const;
-
- const std::vector<AggregationHandle *> &handles_;
- const std::size_t num_handles_;
-
- // Helper object to manage key storage.
- HashTableKeyManager<serializable, force_key_copy> key_manager_;
-
- // In-memory structure is as follows:
- // - SeparateChainingHashTable::Header
- // - Array of slots, interpreted as follows:
- // - 0 = Points to nothing (empty)
- // - SIZE_T_MAX = Pending (some thread is starting a chain from this
- // slot and will overwrite it soon)
- // - Anything else = The number of the first bucket in the chain for
- // this slot PLUS ONE (i.e. subtract one to get the actual bucket
- // number).
- // - Array of buckets, each of which is:
- // - atomic size_t "next" pointer, interpreted the same as slots above.
- // - size_t hash value
- // - possibly some unused bytes as needed so that ValueT's alignment
- // requirement is met
- // - ValueT value slot
- // - fixed-length key storage (which may include pointers to external
- // memory or offsets of variable length keys stored within this hash
- // table)
- // - possibly some additional unused bytes so that bucket size is a
- // multiple of both alignof(std::atomic<std::size_t>) and
- // alignof(ValueT)
- // - Variable-length key storage region (referenced by offsets stored in
- // fixed-length keys).
- Header *header_;
-
- std::atomic<std::size_t> *slots_;
- void *buckets_;
- const std::size_t bucket_size_;
-
- DISALLOW_COPY_AND_ASSIGN(FastSeparateChainingHashTable);
-};
-
-/** @} */
-
-// ----------------------------------------------------------------------------
-// Implementations of template class methods follow.
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- FastSeparateChainingHashTable(
- const std::vector<const Type *> &key_types,
- const std::size_t num_entries,
- const std::vector<std::size_t> &payload_sizes,
- const std::vector<AggregationHandle *> &handles,
- StorageManager *storage_manager)
- : FastHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>(key_types,
- num_entries,
- handles,
- payload_sizes,
- storage_manager,
- false,
- false,
- true),
- kBucketAlignment(alignof(std::atomic<std::size_t>)),
- kValueOffset(sizeof(std::atomic<std::size_t>) + sizeof(std::size_t)),
- handles_(handles),
- num_handles_(handles.size()),
- key_manager_(this->key_types_, kValueOffset + this->total_payload_size_),
- bucket_size_(ComputeBucketSize(key_manager_.getFixedKeySize())) {
- init_payload_ =
- static_cast<std::uint8_t *>(calloc(this->total_payload_size_, 1));
- DCHECK(init_payload_ != nullptr);
- int k = 0;
- for (auto handle : this->handles_) {
- handle->initPayload(init_payload_ + this->payload_offsets_[k]);
- k++;
- }
- // Bucket size always rounds up to the alignment requirement of the atomic
- // size_t "next" pointer at the front or a ValueT, whichever is larger.
- //
- // Give base HashTable information about what key components are stored
- // inline from 'key_manager_'.
- this->setKeyInline(key_manager_.getKeyInline());
-
- // Pick out a prime number of slots and calculate storage requirements.
- std::size_t num_slots_tmp =
- get_next_prime_number(num_entries * kHashTableLoadFactor);
- std::size_t required_memory =
- sizeof(Header) + num_slots_tmp * sizeof(std::atomic<std::size_t>) +
- (num_slots_tmp / kHashTableLoadFactor) *
- (bucket_size_ + key_manager_.getEstimatedVariableKeySize());
- std::size_t num_storage_slots =
- this->storage_manager_->SlotsNeededForBytes(required_memory);
- if (num_storage_slots == 0) {
- FATAL_ERROR(
- "Storage requirement for SeparateChainingHashTable "
- "exceeds maximum allocation size.");
- }
-
- // Get a StorageBlob to hold the hash table.
- const block_id blob_id =
- this->storage_manager_->createBlob(num_storage_slots);
- this->blob_ = this->storage_manager_->getBlobMutable(blob_id);
-
- void *aligned_memory_start = this->blob_->getMemoryMutable();
- std::size_t available_memory = num_storage_slots * kSlotSizeBytes;
- if (align(alignof(Header),
- sizeof(Header),
- aligned_memory_start,
- available_memory) == nullptr) {
- // With current values from StorageConstants.hpp, this should be
- // impossible. A blob is at least 1 MB, while a Header has alignment
- // requirement of just kCacheLineBytes (64 bytes).
- FATAL_ERROR(
- "StorageBlob used to hold resizable "
- "SeparateChainingHashTable is too small to meet alignment "
- "requirements of SeparateChainingHashTable::Header.");
- } else if (aligned_memory_start != this->blob_->getMemoryMutable()) {
- // This should also be impossible, since the StorageManager allocates slots
- // aligned to kCacheLineBytes.
- DEV_WARNING("StorageBlob memory adjusted by "
- << (num_storage_slots * kSlotSizeBytes - available_memory)
- << " bytes to meet alignment requirement for "
- << "SeparateChainingHashTable::Header.");
- }
-
- // Locate the header.
- header_ = static_cast<Header *>(aligned_memory_start);
- aligned_memory_start =
- static_cast<char *>(aligned_memory_start) + sizeof(Header);
- available_memory -= sizeof(Header);
-
- // Recompute the number of slots & buckets using the actual available memory.
- // Most likely, we got some extra free bucket space due to "rounding up" to
- // the storage blob's size. It's also possible (though very unlikely) that we
- // will wind up with fewer buckets than we initially wanted because of screwy
- // alignment requirements for ValueT.
- std::size_t num_buckets_tmp =
- available_memory /
- (kHashTableLoadFactor * sizeof(std::atomic<std::size_t>) + bucket_size_ +
- key_manager_.getEstimatedVariableKeySize());
- num_slots_tmp =
- get_previous_prime_number(num_buckets_tmp * kHashTableLoadFactor);
- num_buckets_tmp = num_slots_tmp / kHashTableLoadFactor;
- DEBUG_ASSERT(num_slots_tmp > 0);
- DEBUG_ASSERT(num_buckets_tmp > 0);
-
- // Locate the slot array.
- slots_ = static_cast<std::atomic<std::size_t> *>(aligned_memory_start);
- aligned_memory_start = static_cast<char *>(aligned_memory_start) +
- sizeof(std::atomic<std::size_t>) * num_slots_tmp;
- available_memory -= sizeof(std::atomic<std::size_t>) * num_slots_tmp;
-
- // Locate the buckets.
- buckets_ = aligned_memory_start;
- // Extra-paranoid: If ValueT has an alignment requirement greater than that
- // of std::atomic<std::size_t>, we may need to adjust the start of the bucket
- // array.
- if (align(kBucketAlignment, bucket_size_, buckets_, available_memory) ==
- nullptr) {
- FATAL_ERROR(
- "StorageBlob used to hold resizable "
- "SeparateChainingHashTable is too small to meet "
- "alignment requirements of buckets.");
- } else if (buckets_ != aligned_memory_start) {
- DEV_WARNING(
- "Bucket array start position adjusted to meet alignment "
- "requirement for SeparateChainingHashTable's value type.");
- if (num_buckets_tmp * bucket_size_ > available_memory) {
- --num_buckets_tmp;
- }
- }
-
- // Fill in the header.
- header_->num_slots = num_slots_tmp;
- header_->num_buckets = num_buckets_tmp;
- header_->buckets_allocated.store(0, std::memory_order_relaxed);
- header_->variable_length_bytes_allocated.store(0, std::memory_order_relaxed);
- available_memory -= bucket_size_ * (header_->num_buckets);
-
- // Locate variable-length key storage region, and give it all the remaining
- // bytes in the blob.
- key_manager_.setVariableLengthStorageInfo(
- static_cast<char *>(buckets_) + header_->num_buckets * bucket_size_,
- available_memory,
- &(header_->variable_length_bytes_allocated));
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-void FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::clear() {
- const std::size_t used_buckets =
- header_->buckets_allocated.load(std::memory_order_relaxed);
- // Destroy existing values, if necessary.
- destroyPayload();
-
- // Zero-out slot array.
- std::memset(
- slots_, 0x0, sizeof(std::atomic<std::size_t>) * header_->num_slots);
-
- // Zero-out used buckets.
- std::memset(buckets_, 0x0, used_buckets * bucket_size_);
-
- header_->buckets_allocated.store(0, std::memory_order_relaxed);
- header_->variable_length_bytes_allocated.store(0, std::memory_order_relaxed);
- key_manager_.zeroNextVariableLengthKeyOffset();
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-const std::uint8_t* FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::getSingle(const TypedValue &key) const {
- DEBUG_ASSERT(!allow_duplicate_keys);
- DEBUG_ASSERT(this->key_types_.size() == 1);
- DEBUG_ASSERT(
- key.isPlausibleInstanceOf(this->key_types_.front()->getSignature()));
-
- const std::size_t hash_code = key.getHash();
- std::size_t bucket_ref =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- while (bucket_ref != 0) {
- DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.scalarKeyCollisionCheck(key, bucket)) {
- // Match located.
- return reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset);
- }
- bucket_ref =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- }
-
- // Reached the end of the chain and didn't find a match.
- return nullptr;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-const std::uint8_t* FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- getSingleCompositeKey(const std::vector<TypedValue> &key) const {
- DEBUG_ASSERT(!allow_duplicate_keys);
- DEBUG_ASSERT(this->key_types_.size() == key.size());
-
- const std::size_t hash_code = this->hashCompositeKey(key);
- std::size_t bucket_ref =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- while (bucket_ref != 0) {
- DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.compositeKeyCollisionCheck(key, bucket)) {
- // Match located.
- return reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset);
- }
- bucket_ref =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- }
-
- // Reached the end of the chain and didn't find a match.
- return nullptr;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-const std::uint8_t* FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- getSingleCompositeKey(const std::vector<TypedValue> &key, int index) const {
- DEBUG_ASSERT(!allow_duplicate_keys);
- DEBUG_ASSERT(this->key_types_.size() == key.size());
-
- const std::size_t hash_code = this->hashCompositeKey(key);
- std::size_t bucket_ref =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- while (bucket_ref != 0) {
- DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.compositeKeyCollisionCheck(key, bucket)) {
- // Match located.
- return reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset) +
- this->payload_offsets_[index];
- }
- bucket_ref =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- }
-
- // Reached the end of the chain and didn't find a match.
- return nullptr;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-void FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::getAll(const TypedValue &key,
- std::vector<const std::uint8_t *> *values)
- const {
- DEBUG_ASSERT(this->key_types_.size() == 1);
- DEBUG_ASSERT(
- key.isPlausibleInstanceOf(this->key_types_.front()->getSignature()));
-
- const std::size_t hash_code = key.getHash();
- std::size_t bucket_ref =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- while (bucket_ref != 0) {
- DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.scalarKeyCollisionCheck(key, bucket)) {
- // Match located.
- values->push_back(
- reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset));
- if (!allow_duplicate_keys) {
- return;
- }
- }
- bucket_ref =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-void FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- getAllCompositeKey(const std::vector<TypedValue> &key,
- std::vector<const std::uint8_t *> *values) const {
- DEBUG_ASSERT(this->key_types_.size() == key.size());
-
- const std::size_t hash_code = this->hashCompositeKey(key);
- std::size_t bucket_ref =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- while (bucket_ref != 0) {
- DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.compositeKeyCollisionCheck(key, bucket)) {
- // Match located.
- values->push_back(
- reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset));
- if (!allow_duplicate_keys) {
- return;
- }
- }
- bucket_ref =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-HashTablePutResult FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- putInternal(const TypedValue &key,
- const std::size_t variable_key_size,
- const std::uint8_t &value,
- HashTablePreallocationState *prealloc_state) {
- DEBUG_ASSERT(this->key_types_.size() == 1);
- DEBUG_ASSERT(
- key.isPlausibleInstanceOf(this->key_types_.front()->getSignature()));
-
- if (prealloc_state == nullptr) {
- // Early check for a free bucket.
- if (header_->buckets_allocated.load(std::memory_order_relaxed) >=
- header_->num_buckets) {
- return HashTablePutResult::kOutOfSpace;
- }
-
- // TODO(chasseur): If allow_duplicate_keys is true, avoid storing more than
- // one copy of the same variable-length key.
- if (!key_manager_.allocateVariableLengthKeyStorage(variable_key_size)) {
- // Ran out of variable-length key storage space.
- return HashTablePutResult::kOutOfSpace;
- }
- }
-
- const std::size_t hash_code = key.getHash();
- void *bucket = nullptr;
- std::atomic<std::size_t> *pending_chain_ptr;
- std::size_t pending_chain_ptr_finish_value;
- for (;;) {
- if (locateBucketForInsertion(hash_code,
- 0,
- &bucket,
- &pending_chain_ptr,
- &pending_chain_ptr_finish_value,
- prealloc_state)) {
- // Found an empty bucket.
- break;
- } else if (bucket == nullptr) {
- // Ran out of buckets. Deallocate any variable space that we were unable
- // to use.
- DEBUG_ASSERT(prealloc_state == nullptr);
- key_manager_.deallocateVariableLengthKeyStorage(variable_key_size);
- return HashTablePutResult::kOutOfSpace;
- } else {
- // Hash collision found, and duplicates aren't allowed.
- DEBUG_ASSERT(!allow_duplicate_keys);
- DEBUG_ASSERT(prealloc_state == nullptr);
- if (key_manager_.scalarKeyCollisionCheck(key, bucket)) {
- // Duplicate key. Deallocate any variable storage space and return.
- key_manager_.deallocateVariableLengthKeyStorage(variable_key_size);
- return HashTablePutResult::kDuplicateKey;
- }
- }
- }
-
- // Write the key and hash.
- writeScalarKeyToBucket(key, hash_code, bucket, prealloc_state);
-
- // Store the value by using placement new with ValueT's copy constructor.
- new (static_cast<char *>(bucket) + kValueOffset) std::uint8_t(value);
-
- // Update the previous chain pointer to point to the new bucket.
- pending_chain_ptr->store(pending_chain_ptr_finish_value,
- std::memory_order_release);
-
- // We're all done.
- return HashTablePutResult::kOK;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-HashTablePutResult FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- putCompositeKeyInternalFast(const std::vector<TypedValue> &key,
- const std::size_t variable_key_size,
- const std::uint8_t *init_value_ptr,
- HashTablePreallocationState *prealloc_state) {
- DEBUG_ASSERT(this->key_types_.size() == key.size());
-
- if (prealloc_state == nullptr) {
- // Early check for a free bucket.
- if (header_->buckets_allocated.load(std::memory_order_relaxed) >=
- header_->num_buckets) {
- return HashTablePutResult::kOutOfSpace;
- }
-
- // TODO(chasseur): If allow_duplicate_keys is true, avoid storing more than
- // one copy of the same variable-length key.
- if (!key_manager_.allocateVariableLengthKeyStorage(variable_key_size)) {
- // Ran out of variable-length key storage space.
- return HashTablePutResult::kOutOfSpace;
- }
- }
-
- const std::size_t hash_code = this->hashCompositeKey(key);
- void *bucket = nullptr;
- std::atomic<std::size_t> *pending_chain_ptr;
- std::size_t pending_chain_ptr_finish_value;
- for (;;) {
- if (locateBucketForInsertion(hash_code,
- 0,
- &bucket,
- &pending_chain_ptr,
- &pending_chain_ptr_finish_value,
- prealloc_state)) {
- // Found an empty bucket.
- break;
- } else if (bucket == nullptr) {
- // Ran out of buckets. Deallocate any variable space that we were unable
- // to use.
- DEBUG_ASSERT(prealloc_state == nullptr);
- key_manager_.deallocateVariableLengthKeyStorage(variable_key_size);
- return HashTablePutResult::kOutOfSpace;
- } else {
- // Hash collision found, and duplicates aren't allowed.
- DEBUG_ASSERT(!allow_duplicate_keys);
- DEBUG_ASSERT(prealloc_state == nullptr);
- if (key_manager_.compositeKeyCollisionCheck(key, bucket)) {
- // Duplicate key. Deallocate any variable storage space and return.
- key_manager_.deallocateVariableLengthKeyStorage(variable_key_size);
- return HashTablePutResult::kDuplicateKey;
- }
- }
- }
-
- // Write the key and hash.
- writeCompositeKeyToBucket(key, hash_code, bucket, prealloc_state);
-
- std::uint8_t *value = static_cast<std::uint8_t *>(bucket) + kValueOffset;
- memcpy(value, init_value_ptr, this->total_payload_size_);
- // Update the previous chain pointer to point to the new bucket.
- pending_chain_ptr->store(pending_chain_ptr_finish_value,
- std::memory_order_release);
-
- // We're all done.
- return HashTablePutResult::kOK;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-std::uint8_t* FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- upsertInternalFast(const TypedValue &key,
- const std::size_t variable_key_size,
- const std::uint8_t *init_value_ptr) {
- DEBUG_ASSERT(!allow_duplicate_keys);
- DEBUG_ASSERT(this->key_types_.size() == 1);
- DEBUG_ASSERT(
- key.isPlausibleInstanceOf(this->key_types_.front()->getSignature()));
-
- if (variable_key_size > 0) {
- // Don't allocate yet, since the key may already be present. However, we
- // do check if either the allocated variable storage space OR the free
- // space is big enough to hold the key (at least one must be true: either
- // the key is already present and allocated, or we need to be able to
- // allocate enough space for it).
- std::size_t allocated_bytes = header_->variable_length_bytes_allocated.load(
- std::memory_order_relaxed);
- if ((allocated_bytes < variable_key_size) &&
- (allocated_bytes + variable_key_size >
- key_manager_.getVariableLengthKeyStorageSize())) {
- return nullptr;
- }
- }
-
- const std::size_t hash_code = key.getHash();
- void *bucket = nullptr;
- std::atomic<std::size_t> *pending_chain_ptr;
- std::size_t pending_chain_ptr_finish_value;
- for (;;) {
- if (locateBucketForInsertion(hash_code,
- variable_key_size,
- &bucket,
- &pending_chain_ptr,
- &pending_chain_ptr_finish_value,
- nullptr)) {
- // Found an empty bucket.
- break;
- } else if (bucket == nullptr) {
- // Ran out of buckets or variable-key space.
- return nullptr;
- } else if (key_manager_.scalarKeyCollisionCheck(key, bucket)) {
- // Found an already-existing entry for this key.
- return reinterpret_cast<std::uint8_t *>(static_cast<char *>(bucket) +
- kValueOffset);
- }
- }
-
- // We are now writing to an empty bucket.
- // Write the key and hash.
- writeScalarKeyToBucket(key, hash_code, bucket, nullptr);
-
- // Copy the supplied 'initial_value' into place.
- std::uint8_t *value = static_cast<unsigned char *>(bucket) + kValueOffset;
- if (init_value_ptr == nullptr) {
- memcpy(value, init_payload_, this->total_payload_size_);
- } else {
- memcpy(value, init_value_ptr, this->total_payload_size_);
- }
-
- // Update the previous chain pointer to point to the new bucket.
- pending_chain_ptr->store(pending_chain_ptr_finish_value,
- std::memory_order_release);
-
- // Return the value.
- return value;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-std::uint8_t* FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- upsertCompositeKeyInternalFast(const std::vector<TypedValue> &key,
- const std::uint8_t *init_value_ptr,
- const std::size_t variable_key_size) {
- DEBUG_ASSERT(!allow_duplicate_keys);
- DEBUG_ASSERT(this->key_types_.size() == key.size());
-
- if (variable_key_size > 0) {
- // Don't allocate yet, since the key may already be present. However, we
- // do check if either the allocated variable storage space OR the free
- // space is big enough to hold the key (at least one must be true: either
- // the key is already present and allocated, or we need to be able to
- // allocate enough space for it).
- std::size_t allocated_bytes = header_->variable_length_bytes_allocated.load(
- std::memory_order_relaxed);
- if ((allocated_bytes < variable_key_size) &&
- (allocated_bytes + variable_key_size >
- key_manager_.getVariableLengthKeyStorageSize())) {
- return nullptr;
- }
- }
-
- const std::size_t hash_code = this->hashCompositeKey(key);
- void *bucket = nullptr;
- std::atomic<std::size_t> *pending_chain_ptr;
- std::size_t pending_chain_ptr_finish_value;
- for (;;) {
- if (locateBucketForInsertion(hash_code,
- variable_key_size,
- &bucket,
- &pending_chain_ptr,
- &pending_chain_ptr_finish_value,
- nullptr)) {
- // Found an empty bucket.
- break;
- } else if (bucket == nullptr) {
- // Ran out of buckets or variable-key space.
- return nullptr;
- } else if (key_manager_.compositeKeyCollisionCheck(key, bucket)) {
- // Found an already-existing entry for this key.
- return reinterpret_cast<std::uint8_t *>(static_cast<char *>(bucket) +
- kValueOffset);
- }
- }
-
- // We are now writing to an empty bucket.
- // Write the key and hash.
- writeCompositeKeyToBucket(key, hash_code, bucket, nullptr);
-
- std::uint8_t *value = static_cast<unsigned char *>(bucket) + kValueOffset;
- if (init_value_ptr == nullptr) {
- memcpy(value, init_payload_, this->total_payload_size_);
- } else {
- memcpy(value, init_value_ptr, this->total_payload_size_);
- }
-
- // Update the previous chaing pointer to point to the new bucket.
- pending_chain_ptr->store(pending_chain_ptr_finish_value,
- std::memory_order_release);
-
- // Return the value.
- return value;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::getNextEntry(TypedValue *key,
- const std::uint8_t **value,
- std::size_t *entry_num) const {
- DEBUG_ASSERT(this->key_types_.size() == 1);
- if (*entry_num < header_->buckets_allocated.load(std::memory_order_relaxed)) {
- const char *bucket =
- static_cast<const char *>(buckets_) + (*entry_num) * bucket_size_;
- *key = key_manager_.getKeyComponentTyped(bucket, 0);
- *value = reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset);
- ++(*entry_num);
- return true;
- } else {
- return false;
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- getNextEntryCompositeKey(std::vector<TypedValue> *key,
- const std::uint8_t **value,
- std::size_t *entry_num) const {
- if (*entry_num < header_->buckets_allocated.load(std::memory_order_relaxed)) {
- const char *bucket =
- static_cast<const char *>(buckets_) + (*entry_num) * bucket_size_;
- for (std::vector<const Type *>::size_type key_idx = 0;
- key_idx < this->key_types_.size();
- ++key_idx) {
- key->emplace_back(key_manager_.getKeyComponentTyped(bucket, key_idx));
- }
- *value = reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset);
- ++(*entry_num);
- return true;
- } else {
- return false;
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::getNextEntryForKey(const TypedValue &key,
- const std::size_t hash_code,
- const std::uint8_t **value,
- std::size_t *entry_num) const {
- DEBUG_ASSERT(this->key_types_.size() == 1);
- DEBUG_ASSERT(
- key.isPlausibleInstanceOf(this->key_types_.front()->getSignature()));
-
- if (*entry_num == 0) {
- *entry_num =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- } else if (*entry_num == std::numeric_limits<std::size_t>::max()) {
- return false;
- }
-
- while (*entry_num != 0) {
- DEBUG_ASSERT(*entry_num != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (*entry_num - 1) * bucket_size_;
- *entry_num =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.scalarKeyCollisionCheck(key, bucket)) {
- // Match located.
- *value = reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset);
- if (*entry_num == 0) {
- // If this is the last bucket in the chain, prevent the next call from
- // starting over again.
- *entry_num = std::numeric_limits<std::size_t>::max();
- }
- return true;
- }
- }
-
- // Reached the end of the chain.
- return false;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- getNextEntryForCompositeKey(const std::vector<TypedValue> &key,
- const std::size_t hash_code,
- const std::uint8_t **value,
- std::size_t *entry_num) const {
- DEBUG_ASSERT(this->key_types_.size() == key.size());
-
- if (*entry_num == 0) {
- *entry_num =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- } else if (*entry_num == std::numeric_limits<std::size_t>::max()) {
- return false;
- }
-
- while (*entry_num != 0) {
- DEBUG_ASSERT(*entry_num != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (*entry_num - 1) * bucket_size_;
- *entry_num =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.compositeKeyCollisionCheck(key, bucket)) {
- // Match located.
- *value = reinterpret_cast<const std::uint8_t *>(bucket + kValueOffset);
- if (*entry_num == 0) {
- // If this is the last bucket in the chain, prevent the next call from
- // starting over again.
- *entry_num = std::numeric_limits<std::size_t>::max();
- }
- return true;
- }
- }
-
- // Reached the end of the chain.
- return false;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::hasKey(const TypedValue &key) const {
- DEBUG_ASSERT(this->key_types_.size() == 1);
- DEBUG_ASSERT(
- key.isPlausibleInstanceOf(this->key_types_.front()->getSignature()));
-
- const std::size_t hash_code = key.getHash();
- std::size_t bucket_ref =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- while (bucket_ref != 0) {
- DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.scalarKeyCollisionCheck(key, bucket)) {
- // Find a match.
- return true;
- }
- bucket_ref =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- }
- return false;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::hasCompositeKey(const std::vector<TypedValue> &key)
- const {
- DEBUG_ASSERT(this->key_types_.size() == key.size());
-
- const std::size_t hash_code = this->hashCompositeKey(key);
- std::size_t bucket_ref =
- slots_[hash_code % header_->num_slots].load(std::memory_order_relaxed);
- while (bucket_ref != 0) {
- DEBUG_ASSERT(bucket_ref != std::numeric_limits<std::size_t>::max());
- const char *bucket =
- static_cast<const char *>(buckets_) + (bucket_ref - 1) * bucket_size_;
- const std::size_t bucket_hash = *reinterpret_cast<const std::size_t *>(
- bucket + sizeof(std::atomic<std::size_t>));
- if ((bucket_hash == hash_code) &&
- key_manager_.compositeKeyCollisionCheck(key, bucket)) {
- // Find a match.
- return true;
- }
- bucket_ref =
- reinterpret_cast<const std::atomic<std::size_t> *>(bucket)->load(
- std::memory_order_relaxed);
- }
- return false;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-void FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::resize(const std::size_t extra_buckets,
- const std::size_t extra_variable_storage,
- const std::size_t retry_num) {
- DEBUG_ASSERT(resizable);
-
- // A retry should never be necessary with this implementation of HashTable.
- // Separate chaining ensures that any resized hash table with more buckets
- // than the original table will be able to hold more entries than the
- // original.
- DEBUG_ASSERT(retry_num == 0);
-
- SpinSharedMutexExclusiveLock<true> write_lock(this->resize_shared_mutex_);
-
- // Recheck whether the hash table is still full. Note that multiple threads
- // might wait to rebuild this hash table simultaneously. Only the first one
- // should do the rebuild.
- if (!isFull(extra_variable_storage)) {
- return;
- }
-
- // Approximately double the number of buckets and slots.
- //
- // TODO(chasseur): It may be worth it to more than double the number of
- // buckets here so that we can maintain a good, sparse fill factor for a
- // longer time as more values are inserted. Such behavior should take into
- // account kHashTableLoadFactor.
- std::size_t resized_num_slots = get_next_prime_number(
- (header_->num_buckets + extra_buckets / 2) * kHashTableLoadFactor * 2);
- std::size_t variable_storage_required =
- (resized_num_slots / kHashTableLoadFactor) *
- key_manager_.getEstimatedVariableKeySize();
- const std::size_t original_variable_storage_used =
- header_->variable_length_bytes_allocated.load(std::memory_order_relaxed);
- // If this resize was triggered by a too-large variable-length key, bump up
- // the variable-length storage requirement.
- if ((extra_variable_storage > 0) &&
- (extra_variable_storage + original_variable_storage_used >
- key_manager_.getVariableLengthKeyStorageSize())) {
- variable_storage_required += extra_variable_storage;
- }
-
- const std::size_t resized_memory_required =
- sizeof(Header) + resized_num_slots * sizeof(std::atomic<std::size_t>) +
- (resized_num_slots / kHashTableLoadFactor) * bucket_size_ +
- variable_storage_required;
- const std::size_t resized_storage_slots =
- this->storage_manager_->SlotsNeededForBytes(resized_memory_required);
- if (resized_storage_slots == 0) {
- FATAL_ERROR(
- "Storage requirement for resized SeparateChainingHashTable "
- "exceeds maximum allocation size.");
- }
-
- // Get a new StorageBlob to hold the resized hash table.
- const block_id resized_blob_id =
- this->storage_manager_->createBlob(resized_storage_slots);
- MutableBlobReference resized_blob =
- this->storage_manager_->getBlobMutable(resized_blob_id);
-
- // Locate data structures inside the new StorageBlob.
- void *aligned_memory_start = resized_blob->getMemoryMutable();
- std::size_t available_memory = resized_storage_slots * kSlotSizeBytes;
- if (align(alignof(Header),
- sizeof(Header),
- aligned_memory_start,
- available_memory) == nullptr) {
- // Should be impossible, as noted in constructor.
- FATAL_ERROR(
- "StorageBlob used to hold resized SeparateChainingHashTable "
- "is too small to meet alignment requirements of "
- "LinearOpenAddressingHashTable::Header.");
- } else if (aligned_memory_start != resized_blob->getMemoryMutable()) {
- // Again, should be impossible.
- DEV_WARNING("In SeparateChainingHashTable::resize(), StorageBlob "
- << "memory adjusted by "
- << (resized_num_slots * kSlotSizeBytes - available_memory)
- << " bytes to meet alignment requirement for "
- << "LinearOpenAddressingHashTable::Header.");
- }
-
- Header *resized_header = static_cast<Header *>(aligned_memory_start);
- aligned_memory_start =
- static_cast<char *>(aligned_memory_start) + sizeof(Header);
- available_memory -= sizeof(Header);
-
- // As in constructor, recompute the number of slots and buckets using the
- // actual available memory.
- std::size_t resized_num_buckets =
- (available_memory - extra_variable_storage) /
- (kHashTableLoadFactor * sizeof(std::atomic<std::size_t>) + bucket_size_ +
- key_manager_.getEstimatedVariableKeySize());
- resized_num_slots =
- get_previous_prime_number(resized_num_buckets * kHashTableLoadFactor);
- resized_num_buckets = resized_num_slots / kHashTableLoadFactor;
-
- // Locate slot array.
- std::atomic<std::size_t> *resized_slots =
- static_cast<std::atomic<std::size_t> *>(aligned_memory_start);
- aligned_memory_start = static_cast<char *>(aligned_memory_start) +
- sizeof(std::atomic<std::size_t>) * resized_num_slots;
- available_memory -= sizeof(std::atomic<std::size_t>) * resized_num_slots;
-
- // As in constructor, we will be extra paranoid and use align() to locate the
- // start of the array of buckets, as well.
- void *resized_buckets = aligned_memory_start;
- if (align(
- kBucketAlignment, bucket_size_, resized_buckets, available_memory) ==
- nullptr) {
- FATAL_ERROR(
- "StorageBlob used to hold resized SeparateChainingHashTable "
- "is too small to meet alignment requirements of buckets.");
- } else if (resized_buckets != aligned_memory_start) {
- DEV_WARNING(
- "Bucket array start position adjusted to meet alignment "
- "requirement for SeparateChainingHashTable's value type.");
- if (resized_num_buckets * bucket_size_ + variable_storage_required >
- available_memory) {
- --resized_num_buckets;
- }
- }
- aligned_memory_start = static_cast<char *>(aligned_memory_start) +
- resized_num_buckets * bucket_size_;
- available_memory -= resized_num_buckets * bucket_size_;
-
- void *resized_variable_length_key_storage = aligned_memory_start;
- const std::size_t resized_variable_length_key_storage_size = available_memory;
-
- const std::size_t original_buckets_used =
- header_->buckets_allocated.load(std::memory_order_relaxed);
-
- // Initialize the header.
- resized_header->num_slots = resized_num_slots;
- resized_header->num_buckets = resized_num_buckets;
- resized_header->buckets_allocated.store(original_buckets_used,
- std::memory_order_relaxed);
- resized_header->variable_length_bytes_allocated.store(
- original_variable_storage_used, std::memory_order_relaxed);
-
- // Bulk-copy buckets. This is safe because:
- // 1. The "next" pointers will be adjusted when rebuilding chains below.
- // 2. The hash codes will stay the same.
- // 3. For key components:
- // a. Inline keys will stay exactly the same.
- // b. Offsets into variable-length storage will remain valid, because
- // we also do a byte-for-byte copy of variable-length storage below.
- // c. Absolute external pointers will still point to the same address.
- // d. Relative pointers are not used with resizable hash tables.
- // 4. If values are not trivially copyable, then we invoke ValueT's copy
- // or move constructor with placement new.
- // NOTE(harshad) - Regarding point 4 above, as this is a specialized
- // hash table implemented for aggregation, the values are trivially copyable,
- // therefore we don't need to invoke payload values' copy/move constructors.
- std::memcpy(resized_buckets, buckets_, original_buckets_used * bucket_size_);
-
- // Copy over variable-length key components, if any.
- if (original_variable_storage_used > 0) {
- DEBUG_ASSERT(original_variable_storage_used ==
- key_manager_.getNextVariableLengthKeyOffset());
- DEBUG_ASSERT(original_variable_storage_used <=
- resized_variable_length_key_storage_size);
- std::memcpy(resized_variable_length_key_storage,
- key_manager_.getVariableLengthKeyStorage(),
- original_variable_storage_used);
- }
-
- destroyPayload();
-
- // Make resized structures active.
- std::swap(this->blob_, resized_blob);
- header_ = resized_header;
- slots_ = resized_slots;
- buckets_ = resized_buckets;
- key_manager_.setVariableLengthStorageInfo(
- resized_variable_length_key_storage,
- resized_variable_length_key_storage_size,
- &(resized_header->variable_length_bytes_allocated));
-
- // Drop the old blob.
- const block_id old_blob_id = resized_blob->getID();
- resized_blob.release();
- this->storage_manager_->deleteBlockOrBlobFile(old_blob_id);
-
- // Rebuild chains.
- void *current_bucket = buckets_;
- for (std::size_t bucket_num = 0; bucket_num < original_buckets_used;
- ++bucket_num) {
- std::atomic<std::size_t> *next_ptr =
- static_cast<std::atomic<std::size_t> *>(current_bucket);
- const std::size_t hash_code = *reinterpret_cast<const std::size_t *>(
- static_cast<const char *>(current_bucket) +
- sizeof(std::atomic<std::size_t>));
-
- const std::size_t slot_number = hash_code % header_->num_slots;
- std::size_t slot_ptr_value = 0;
- if (slots_[slot_number].compare_exchange_strong(
- slot_ptr_value, bucket_num + 1, std::memory_order_relaxed)) {
- // This bucket is the first in the chain for this block, so reset its
- // next pointer to 0.
- next_ptr->store(0, std::memory_order_relaxed);
- } else {
- // A chain already exists starting from this slot, so put this bucket at
- // the head.
- next_ptr->store(slot_ptr_value, std::memory_order_relaxed);
- slots_[slot_number].store(bucket_num + 1, std::memory_order_relaxed);
- }
- current_bucket = static_cast<char *>(current_bucket) + bucket_size_;
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- preallocateForBulkInsert(const std::size_t total_entries,
- const std::size_t total_variable_key_size,
- HashTablePreallocationState *prealloc_state) {
- DEBUG_ASSERT(allow_duplicate_keys);
- if (!key_manager_.allocateVariableLengthKeyStorage(total_variable_key_size)) {
- return false;
- }
-
- // We use load then compare-exchange here instead of simply fetch-add,
- // because if multiple threads are simultaneously trying to allocate more
- // than one bucket and exceed 'header_->num_buckets', their respective
- // rollbacks might happen in such an order that some bucket ranges get
- // skipped, while others might get double-allocated later.
- std::size_t original_buckets_allocated =
- header_->buckets_allocated.load(std::memory_order_relaxed);
- std::size_t buckets_post_allocation =
- original_buckets_allocated + total_entries;
- while ((buckets_post_allocation <= header_->num_buckets) &&
- !header_->buckets_allocated.compare_exchange_weak(
- original_buckets_allocated,
- buckets_post_allocation,
- std::memory_order_relaxed)) {
- buckets_post_allocation = original_buckets_allocated + total_entries;
- }
-
- if (buckets_post_allocation > header_->num_buckets) {
- key_manager_.deallocateVariableLengthKeyStorage(total_variable_key_size);
- return false;
- }
-
- prealloc_state->bucket_position = original_buckets_allocated;
- if (total_variable_key_size != 0) {
- prealloc_state->variable_length_key_position =
- key_manager_.incrementNextVariableLengthKeyOffset(
- total_variable_key_size);
- }
- return true;
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-inline bool FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- locateBucketForInsertion(const std::size_t hash_code,
- const std::size_t variable_key_allocation_required,
- void **bucket,
- std::atomic<std::size_t> **pending_chain_ptr,
- std::size_t *pending_chain_ptr_finish_value,
- HashTablePreallocationState *prealloc_state) {
- DEBUG_ASSERT((prealloc_state == nullptr) || allow_duplicate_keys);
- if (*bucket == nullptr) {
- *pending_chain_ptr = &(slots_[hash_code % header_->num_slots]);
- } else {
- *pending_chain_ptr = static_cast<std::atomic<std::size_t> *>(*bucket);
- }
- for (;;) {
- std::size_t existing_chain_ptr = 0;
- if ((*pending_chain_ptr)
- ->compare_exchange_strong(existing_chain_ptr,
- std::numeric_limits<std::size_t>::max(),
- std::memory_order_acq_rel)) {
- // Got to the end of the chain. Allocate a new bucket.
-
- // First, allocate variable-length key storage, if needed (i.e. if this
- // is an upsert and we didn't allocate up-front).
- if ((prealloc_state == nullptr) &&
- !key_manager_.allocateVariableLengthKeyStorage(
- variable_key_allocation_required)) {
- // Ran out of variable-length storage.
- (*pending_chain_ptr)->store(0, std::memory_order_release);
- *bucket = nullptr;
- return false;
- }
-
- const std::size_t allocated_bucket_num =
- (prealloc_state == nullptr)
- ? header_->buckets_allocated.fetch_add(1,
- std::memory_order_relaxed)
- : (prealloc_state->bucket_position)++;
- if (allocated_bucket_num >= header_->num_buckets) {
- // Ran out of buckets.
- DEBUG_ASSERT(prealloc_state == nullptr);
- header_->buckets_allocated.fetch_sub(1, std::memory_order_relaxed);
- (*pending_chain_ptr)->store(0, std::memory_order_release);
- *bucket = nullptr;
- return false;
- } else {
- *bucket =
- static_cast<char *>(buckets_) + allocated_bucket_num * bucket_size_;
- *pending_chain_ptr_finish_value = allocated_bucket_num + 1;
- return true;
- }
- }
- // Spin until the real "next" pointer is available.
- while (existing_chain_ptr == std::numeric_limits<std::size_t>::max()) {
- existing_chain_ptr =
- (*pending_chain_ptr)->load(std::memory_order_acquire);
- }
- if (existing_chain_ptr == 0) {
- // Other thread had to roll back, so try again.
- continue;
- }
- // Chase the next pointer.
- *bucket =
- static_cast<char *>(buckets_) + (existing_chain_ptr - 1) * bucket_size_;
- *pending_chain_ptr = static_cast<std::atomic<std::size_t> *>(*bucket);
- if (!allow_duplicate_keys) {
- const std::size_t hash_in_bucket = *reinterpret_cast<const std::size_t *>(
- static_cast<const char *>(*bucket) +
- sizeof(std::atomic<std::size_t>));
- if (hash_in_bucket == hash_code) {
- return false;
- }
- }
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-inline void FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- writeScalarKeyToBucket(const TypedValue &key,
- const std::size_t hash_code,
- void *bucket,
- HashTablePreallocationState *prealloc_state) {
- *reinterpret_cast<std::size_t *>(static_cast<char *>(bucket) +
- sizeof(std::atomic<std::size_t>)) =
- hash_code;
- key_manager_.writeKeyComponentToBucket(key, 0, bucket, prealloc_state);
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-inline void FastSeparateChainingHashTable<resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::
- writeCompositeKeyToBucket(const std::vector<TypedValue> &key,
- const std::size_t hash_code,
- void *bucket,
- HashTablePreallocationState *prealloc_state) {
- DEBUG_ASSERT(key.size() == this->key_types_.size());
- *reinterpret_cast<std::size_t *>(static_cast<char *>(bucket) +
- sizeof(std::atomic<std::size_t>)) =
- hash_code;
- for (std::size_t idx = 0; idx < this->key_types_.size(); ++idx) {
- key_manager_.writeKeyComponentToBucket(
- key[idx], idx, bucket, prealloc_state);
- }
-}
-
-template <bool resizable,
- bool serializable,
- bool force_key_copy,
- bool allow_duplicate_keys>
-bool FastSeparateChainingHashTable<
- resizable,
- serializable,
- force_key_copy,
- allow_duplicate_keys>::isFull(const std::size_t extra_variable_storage)
- const {
- if (header_->buckets_allocated.load(std::memory_order_relaxed) >=
- header_->num_buckets) {
- // All buckets are allocated.
- return true;
- }
-
- if (extra_variable_storage > 0) {
- if (extra_variable_storage +
- header_->variable_length_bytes_allocated.load(
- std::memory_order_relaxed) >
- key_manager_.getVariableLengthKeyStorageSize()) {
- // Not enough variable-length key storage space.
- return true;
- }
- }
-
- return false;
-}
-
-} // namespace quickstep
-
-#endif // QUICKSTEP_STORAGE_SEPARATE_CHAINING_HASH_TABLE_HPP_
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/HashTableBase.hpp
----------------------------------------------------------------------
diff --git a/storage/HashTableBase.hpp b/storage/HashTableBase.hpp
index a3180bb..edea63b 100644
--- a/storage/HashTableBase.hpp
+++ b/storage/HashTableBase.hpp
@@ -23,11 +23,13 @@
#include <cstddef>
#include <vector>
-#include "ValueAccessor.hpp"
#include "utility/Macros.hpp"
namespace quickstep {
+class ColumnVectorsValueAccessor;
+class ValueAccessor;
+
/** \addtogroup Storage
* @{
*/
@@ -75,6 +77,23 @@ class HashTableBase {
virtual ~HashTableBase() {}
/**
+ * @brief Destroy the payload stored in the hash table.
+ **/
+ virtual void destroyPayload() {
+ }
+
+ protected:
+ HashTableBase() {}
+
+ private:
+ DISALLOW_COPY_AND_ASSIGN(HashTableBase);
+};
+
+class AggregationStateHashTableBase {
+ public:
+ virtual ~AggregationStateHashTableBase() {}
+
+ /**
* TODO(harshad) We should get rid of this function from here. We are
* postponing it because of the amount of work to be done is significant.
* The steps are as follows:
@@ -91,29 +110,21 @@ class HashTableBase {
* Optionally, we can also remove the AggregationStateHashTableBase
* specialization from this file.
**/
- virtual bool upsertValueAccessorCompositeKeyFast(
- const std::vector<attribute_id> &argument,
- ValueAccessor *accessor,
+ virtual bool upsertValueAccessor(
+ const std::vector<std::vector<attribute_id>> &argument_ids,
const std::vector<attribute_id> &key_attr_ids,
- const bool check_for_null_keys) {
- return false;
- }
+ ValueAccessor *accessor,
+ ColumnVectorsValueAccessor *aux_accessor = nullptr) = 0;
- /**
- * @brief Destroy the payload stored in the hash table.
- **/
- virtual void destroyPayload() {
- }
+ virtual void destroyPayload() = 0;
protected:
- HashTableBase() {}
+ AggregationStateHashTableBase() {}
private:
- DISALLOW_COPY_AND_ASSIGN(HashTableBase);
+ DISALLOW_COPY_AND_ASSIGN(AggregationStateHashTableBase);
};
-typedef HashTableBase<true, false, true, false> AggregationStateHashTableBase;
-
/** @} */
} // namespace quickstep
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/HashTableFactory.hpp
----------------------------------------------------------------------
diff --git a/storage/HashTableFactory.hpp b/storage/HashTableFactory.hpp
index d690557..16d13c8 100644
--- a/storage/HashTableFactory.hpp
+++ b/storage/HashTableFactory.hpp
@@ -28,6 +28,7 @@
#include "storage/HashTableBase.hpp"
#include "storage/HashTable.pb.h"
#include "storage/LinearOpenAddressingHashTable.hpp"
+#include "storage/PackedPayloadAggregationStateHashTable.hpp"
#include "storage/SeparateChainingHashTable.hpp"
#include "storage/SimpleScalarSeparateChainingHashTable.hpp"
#include "storage/TupleReference.hpp"
@@ -324,19 +325,39 @@ class HashTableFactory {
};
/**
- * @brief Convenient alias that provides a HashTableFactory whose only template
- * parameter is the aggregate state type.
- **/
-template <typename ValueT>
-using AggregationStateHashTableFactory
- = HashTableFactory<ValueT, true, false, true, false>;
-
-/**
* @brief Convenient alias for a HashTableFactory that makes JoinHashTables.
**/
typedef HashTableFactory<TupleReference, true, false, false, true>
JoinHashTableFactory;
+/**
+ * @brief TODO
+ */
+class AggregationStateHashTableFactory {
+ public:
+ static AggregationStateHashTableBase* CreateResizable(
+ const HashTableImplType hash_table_type,
+ const std::vector<const Type*> &key_types,
+ const std::size_t num_entries,
+ const std::vector<AggregationHandle *> &handles,
+ StorageManager *storage_manager) {
+ switch (hash_table_type) {
+ case HashTableImplType::kSeparateChaining:
+ return new PackedPayloadSeparateChainingAggregationStateHashTable(
+ key_types, num_entries, handles, storage_manager);
+ default: {
+ LOG(FATAL) << "Unrecognized HashTableImplType in HashTableFactory::createResizable()\n";
+ }
+ }
+ }
+
+ private:
+ // Class is all-static and should not be instantiated.
+ AggregationStateHashTableFactory();
+
+ DISALLOW_COPY_AND_ASSIGN(AggregationStateHashTableFactory);
+};
+
/** @} */
} // namespace quickstep
http://git-wip-us.apache.org/repos/asf/incubator-quickstep/blob/7048fcc1/storage/HashTablePool.hpp
----------------------------------------------------------------------
diff --git a/storage/HashTablePool.hpp b/storage/HashTablePool.hpp
index 96cf849..5ba703b 100644
--- a/storage/HashTablePool.hpp
+++ b/storage/HashTablePool.hpp
@@ -27,8 +27,7 @@
#include "expressions/aggregation/AggregationHandle.hpp"
#include "storage/HashTableBase.hpp"
-#include "storage/FastHashTable.hpp"
-#include "storage/FastHashTableFactory.hpp"
+#include "storage/HashTableFactory.hpp"
#include "threading/SpinMutex.hpp"
#include "utility/Macros.hpp"
#include "utility/StringUtil.hpp"
@@ -56,36 +55,6 @@ class HashTablePool {
/**
* @brief Constructor.
*
- * @param estimated_num_entries The maximum number of entries in a hash table.
- * @param hash_table_impl_type The type of hash table implementation.
- * @param group_by_types A vector of pointer of types which form the group by
- * key.
- * @param agg_handle The aggregation handle.
- * @param storage_manager A pointer to the storage manager.
- *
- * @note The estimate of number of entries is quite inaccurate at this time.
- * If we go by the current estimate, each hash table demands much
- * larger space than it actually needs, which causes the system to
- * either trigger evictions or worse - run out of memory. To fix this
- * issue, we divide the estimate by 100. The division will not affect
- * correctness, however it may allocate some hash tables smaller space
- * than their requirement, causing them to be resized during build
- * phase, which has a performance penalty.
- **/
- HashTablePool(const std::size_t estimated_num_entries,
- const HashTableImplType hash_table_impl_type,
- const std::vector<const Type *> &group_by_types,
- AggregationHandle *agg_handle,
- StorageManager *storage_manager)
- : estimated_num_entries_(reduceEstimatedCardinality(estimated_num_entries)),
- hash_table_impl_type_(hash_table_impl_type),
- group_by_types_(group_by_types),
- agg_handle_(DCHECK_NOTNULL(agg_handle)),
- storage_manager_(DCHECK_NOTNULL(storage_manager)) {}
-
- /**
- * @brief Constructor.
- *
* @note This constructor is relevant for HashTables specialized for
* aggregation.
*
@@ -93,52 +62,29 @@ class HashTablePool {
* @param hash_table_impl_type The type of hash table implementation.
* @param group_by_types A vector of pointer of types which form the group by
* key.
- * @param payload_sizes The sizes in bytes for the AggregationStates for the
- * respective AggregationHandles.
* @param handles The AggregationHandles in this query.
* @param storage_manager A pointer to the storage manager.
**/
HashTablePool(const std::size_t estimated_num_entries,
const HashTableImplType hash_table_impl_type,
const std::vector<const Type *> &group_by_types,
- const std::vector<std::size_t> &payload_sizes,
const std::vector<AggregationHandle *> &handles,
StorageManager *storage_manager)
: estimated_num_entries_(reduceEstimatedCardinality(estimated_num_entries)),
hash_table_impl_type_(hash_table_impl_type),
group_by_types_(group_by_types),
- payload_sizes_(payload_sizes),
handles_(handles),
storage_manager_(DCHECK_NOTNULL(storage_manager)) {}
/**
* @brief Check out a hash table for insertion.
*
- * @return A hash table pointer.
- **/
- AggregationStateHashTableBase* getHashTable() {
- {
- SpinMutexLock lock(mutex_);
- if (!hash_tables_.empty()) {
- std::unique_ptr<AggregationStateHashTableBase> ret_hash_table(
- std::move(hash_tables_.back()));
- hash_tables_.pop_back();
- DCHECK(ret_hash_table != nullptr);
- return ret_hash_table.release();
- }
- }
- return createNewHashTable();
- }
-
- /**
- * @brief Check out a hash table for insertion.
- *
* @note This method is relevant for specialized (for aggregation)
* hash table implementation.
*
* @return A hash table pointer.
**/
- AggregationStateHashTableBase* getHashTableFast() {
+ AggregationStateHashTableBase* getHashTable() {
{
SpinMutexLock lock(mutex_);
if (!hash_tables_.empty()) {
@@ -149,7 +95,7 @@ class HashTablePool {
return ret_hash_table.release();
}
}
- return createNewHashTableFast();
+ return createNewHashTable();
}
/**
@@ -180,18 +126,10 @@ class HashTablePool {
private:
AggregationStateHashTableBase* createNewHashTable() {
- return agg_handle_->createGroupByHashTable(hash_table_impl_type_,
- group_by_types_,
- estimated_num_entries_,
- storage_manager_);
- }
-
- AggregationStateHashTableBase* createNewHashTableFast() {
- return AggregationStateFastHashTableFactory::CreateResizable(
+ return AggregationStateHashTableFactory::CreateResizable(
hash_table_impl_type_,
group_by_types_,
estimated_num_entries_,
- payload_sizes_,
handles_,
storage_manager_);
}
@@ -214,10 +152,6 @@ class HashTablePool {
const HashTableImplType hash_table_impl_type_;
const std::vector<const Type *> group_by_types_;
-
- std::vector<std::size_t> payload_sizes_;
-
- AggregationHandle *agg_handle_;
const std::vector<AggregationHandle *> handles_;
StorageManager *storage_manager_;