Enable merging of LocalVocabs

src/engine/AddCombinedRowToTable.h

@@ -218,21 +218,23 @@ class AddCombinedRowToIdTable {
     AD_CORRECTNESS_CHECK(inputLeftAndRight_.has_value());
     result.resize(oldSize + nextIndex_);
 
-    // Sometimes columns are combined where one value is UNDEF and the other one
-    // is not. This function very efficiently returns the not-UNDEF value in
-    // this case.
-    // TODO<joka921> If we keep track of the information that one of the
-    // involved columns contains no UNDEF values at all, we can omit this step
-    // and simply copy the values from this column without looking at the other
-    // input.
-    auto mergeWithUndefined = [](const ValueId a, const ValueId b) {
-      static_assert(ValueId::makeUndefined().getBits() == 0u);
-      return ValueId::fromBits(a.getBits() | b.getBits());
+    // Precondition: `a` and `b` compare equal or at least one of them is UNDEF
+    // If exactly one of them is UNDEF, return the other one, else return any of
+    // them (they are equal anyway).
+    auto getJoinValue = [](const ValueId a, const ValueId b) {
+      // NOTE: For localVocabIndices we might have different pointers that
+      // compare equal because they point to the same word. Therefore we cannot
+      // use a simple bitwise operation to handle the "one of them is UNDEF"
+      // case as we previously did.
+      if (a.isUndefined()) {
+        return b;
+      }
+      return a;
     };
 
     // A lambda that writes the join column with the given `colIdx` to the
     // `nextResultColIdx`-th column of the result.
-    auto writeJoinColumn = [&result, &mergeWithUndefined, oldSize, this](
+    auto writeJoinColumn = [&result, &getJoinValue, oldSize, this](
                                size_t colIdx, size_t resultColIdx) {
       const auto& colLeft = inputLeft().getColumn(colIdx);
       const auto& colRight = inputRight().getColumn(colIdx);
@@ -242,7 +244,8 @@ class AddCombinedRowToIdTable {
 
       // Write the matching rows.
       for (const auto& [targetIndex, sourceIndices] : indexBuffer_) {
-        auto resultId = mergeWithUndefined(colLeft[sourceIndices[0]],
+        auto resultId =
+            getJoinValue(colLeft[sourceIndices[0]],
                                            colRight[sourceIndices[1]]);
         numUndef += static_cast<size_t>(resultId.isUndefined());
         resultCol[oldSize + targetIndex] = resultId;

src/engine/CartesianProductJoin.cpp

@@ -202,7 +202,7 @@ ResultTable CartesianProductJoin::computeResult() {
   auto subResultsDeref = std::views::transform(
       subResults, [](auto& x) -> decltype(auto) { return *x; });
   return {std::move(result), resultSortedOn(),
-          ResultTable::getSharedLocalVocabFromNonEmptyOf(subResultsDeref)};
+          ResultTable::getMergedLocalVocab(subResultsDeref)};
 }
 
 // ____________________________________________________________________________

src/engine/Join.cpp

@@ -187,7 +187,7 @@ ResultTable Join::computeResult() {
   // If only one of the two operands has a non-empty local vocabulary, share
   // with that one (otherwise, throws an exception).
   return {std::move(idTable), resultSortedOn(),
-          ResultTable::getSharedLocalVocabFromNonEmptyOf(*leftRes, *rightRes)};
+          ResultTable::getMergedLocalVocab(*leftRes, *rightRes)};
 }
 
 // _____________________________________________________________________________

src/engine/LocalVocab.cpp

@@ -11,56 +11,36 @@
 // _____________________________________________________________________________
 LocalVocab LocalVocab::clone() const {
   LocalVocab localVocabClone;
-  // First, make a deep copy of the `absl::node_hash_map` holding the actual
-  // map of strings to indexes.
-  localVocabClone.wordsToIndexesMap_ = this->wordsToIndexesMap_;
-  // The next free index should be the same.
-  localVocabClone.nextFreeIndex_ = this->nextFreeIndex_;
-  // The map from local ids to strings stores pointers to strings. So we cannot
-  // just copy these from `this->indexesToWordsMap_` to `localVocabClone`, but
-  // we need to make sure to store the pointers to the strings of the new map
-  // `localVocabClone.wordsToIndexesMap_`.
-  //
-  // NOTE: An alternative algorithm would be to sort the word-index pairs in
-  // `wordsToIndexesMap_` by index and then fill `indexesToWordsMap_` in order.
-  // This would have better locality, but the sorting takes non-linear time plus
-  // the sorting has to handle pairs of `LocalVocabIndex` and `std::string`. So
-  // for very large local vocabularies (and only then is this operation
-  // performance-criticial at all), the simpler approach below is probably
-  // better.
-  const size_t localVocabSize = this->size();
-  localVocabClone.indexesToWordsMap_.resize(localVocabSize);
-  for (const auto& [wordInMap, index] : localVocabClone.wordsToIndexesMap_) {
-    AD_CONTRACT_CHECK(index.get() < localVocabSize);
-    localVocabClone.indexesToWordsMap_[index.get()] = std::addressof(wordInMap);
-  }
+  localVocabClone.otherWordSets_ = otherWordSets_;
+  localVocabClone.otherWordSets_.push_back(primaryWordSet_);
   // Return the clone.
   return localVocabClone;
 }
 
+// _____________________________________________________________________________
+LocalVocab LocalVocab::merge(const LocalVocab& a, const LocalVocab& b) {
+  LocalVocab res;
+  auto inserter = std::back_inserter(res.otherWordSets_);
+  std::ranges::copy(a.otherWordSets_, inserter);
+  std::ranges::copy(b.otherWordSets_, inserter);
+  *inserter = a.primaryWordSet_;
+  *inserter = b.primaryWordSet_;
+  return res;
+}
+
 // _____________________________________________________________________________
 template <typename WordT>
 LocalVocabIndex LocalVocab::getIndexAndAddIfNotContainedImpl(WordT&& word) {
-  // The following code contains two subtle, but important optimizations:
+  // The following code contains a subtle, but important optimizations:
   //
   // 1. The variant of `insert` used covers the case that `word` was already
   // contained in the map as well as the case that it is newly inserted. This
   // avoids computing the hash for `word` twice in case we see it for the first
   // time (note that hashing a string is not cheap).
-  //
-  // 2. The fact that we have a member variable `nextFreeIndex_` avoids that we
-  // tentatively have to compute the next free ID every time this function is
-  // called (even when the ID is not actually needed because the word is already
-  // contained in the map).
-  //
-  auto [wordInMapAndIndex, isNewWord] =
-      wordsToIndexesMap_.insert({std::forward<WordT>(word), nextFreeIndex_});
-  const auto& [wordInMap, index] = *wordInMapAndIndex;
-  if (isNewWord) {
-    indexesToWordsMap_.push_back(&wordInMap);
-    nextFreeIndex_ = LocalVocabIndex::make(indexesToWordsMap_.size());
-  }
-  return index;
+  // TODO<joka921> Make this work with transparent hashing and use try_emplace.
+  auto [wordIterator, isNewWord] =
+      primaryWordSet().insert(StringAligned16{std::forward<WordT>(word)});
+  return std::addressof(*wordIterator);
 }
 
 // _____________________________________________________________________________
@@ -77,21 +57,27 @@ LocalVocabIndex LocalVocab::getIndexAndAddIfNotContained(std::string&& word) {
 // _____________________________________________________________________________
 std::optional<LocalVocabIndex> LocalVocab::getIndexOrNullopt(
     const std::string& word) const {
-  auto localVocabIndex = wordsToIndexesMap_.find(word);
-  if (localVocabIndex != wordsToIndexesMap_.end()) {
-    return localVocabIndex->second;
+  // TODO<joka921> Maybe we can make this work with transparent hashing,
+  // but if this is only a testing API this is probably not worth the hassle.
+  auto localVocabIndex = primaryWordSet().find(StringAligned16{word});
+  if (localVocabIndex != primaryWordSet().end()) {
+    return &(*localVocabIndex);
   } else {
     return std::nullopt;
   }
 }
 
 // _____________________________________________________________________________
 const std::string& LocalVocab::getWord(LocalVocabIndex localVocabIndex) const {
-  if (localVocabIndex.get() >= indexesToWordsMap_.size()) {
-    throw std::runtime_error(absl::StrCat(
-        "LocalVocab error: request for word with local vocab index ",
-        localVocabIndex.get(), ", but size of local vocab is only ",
-        indexesToWordsMap_.size(), ", please contact the developers"));
+  return *localVocabIndex;
+}
+
+// _____________________________________________________________________________
+std::vector<std::string> LocalVocab::getAllWordsForTesting() const {
+  std::vector<std::string> result;
+  std::ranges::copy(primaryWordSet(), std::back_inserter(result));
+  for (const auto& previous : otherWordSets_) {
+    std::ranges::copy(*previous, std::back_inserter(result));
   }
-  return *(indexesToWordsMap_.at(localVocabIndex.get()));
+  return result;
 }

src/engine/LocalVocab.h

@@ -9,7 +9,7 @@
 #include <string>
 #include <vector>
 
-#include "absl/container/node_hash_map.h"
+#include "absl/container/node_hash_set.h"
 #include "global/Id.h"
 
 // A class for maintaing a local vocabulary with contiguous (local) IDs. This is
@@ -24,18 +24,17 @@ class LocalVocab {
   // A map of the words in the local vocabulary to their local IDs. This is a
   // node hash map because we need the addresses of the words (which are of type
   // `std::string`) to remain stable over their lifetime in the hash map because
-  // we refer to them in `wordsToIdsMap_` below.
-  absl::node_hash_map<std::string, LocalVocabIndex> wordsToIndexesMap_;
+  // we hand out pointers to them.
+  using Set = absl::node_hash_set<StringAligned16>;
+  std::shared_ptr<Set> primaryWordSet_ = std::make_shared<Set>();
 
-  // A map of the local IDs to the words. Since the IDs are contiguous, we can
-  // use a `std::vector`. We store pointers to the actual words in
-  // `wordsToIdsMap_` to avoid storing every word twice. This saves space, but
-  // costs us an indirection when looking up a word by its ID.
-  std::vector<const std::string*> indexesToWordsMap_;
+  // Local vocabularies from child operations that were merged into this
+  // vocabulary s.t. the pointers are kept alive. They have to be `const`
+  // because they are possibly shared concurrently (for example via the cache).
+  std::vector<std::shared_ptr<const Set>> otherWordSets_;
 
-  // The next free local ID (will be incremented by one each time we add a new
-  // word).
-  LocalVocabIndex nextFreeIndex_ = LocalVocabIndex::make(0);
+  auto& primaryWordSet() { return *primaryWordSet_; }
+  const auto& primaryWordSet() const { return *primaryWordSet_; }
 
  public:
   // Create a new, empty local vocabulary.
@@ -45,7 +44,9 @@ class LocalVocab {
   LocalVocab(const LocalVocab&) = delete;
   LocalVocab& operator=(const LocalVocab&) = delete;
 
-  // Make a deep copy explicitly.
+  // Make a logical copy. The clone will have an empty primary set so it can
+  // safely be modified. The contents are copied as shared pointers to const, so
+  // the function runs in linear time in the number of word sets.
   LocalVocab clone() const;
 
   // Moving a local vocabulary is not problematic (though the typical use case
@@ -65,14 +66,28 @@ class LocalVocab {
       const std::string& word) const;
 
   // The number of words in the vocabulary.
-  size_t size() const { return indexesToWordsMap_.size(); }
+  // Note: This is not constant time, but linear in the number of word sets.
+  size_t size() const {
+    auto result = primaryWordSet().size();
+    for (const auto& previous : otherWordSets_) {
+      result += previous->size();
+    }
+    return result;
+  }
 
   // Return true if and only if the local vocabulary is empty.
-  bool empty() const { return indexesToWordsMap_.empty(); }
+  bool empty() const { return size() == 0; }
 
   // Return a const reference to the word.
   const std::string& getWord(LocalVocabIndex localVocabIndex) const;
 
+  // Create a local vocab that contains and keeps alive all the words from `a`
+  // as well as `b`.
+  static LocalVocab merge(const LocalVocab& a, const LocalVocab& b);
+
+  // Return all the words from all the word sets as a vector.
+  std::vector<std::string> getAllWordsForTesting() const;
+
  private:
   // Common implementation for the two variants of
   // `getIndexAndAddIfNotContainedImpl` above.

src/engine/Minus.cpp

@@ -56,8 +56,7 @@ ResultTable Minus::computeResult() {
   // If only one of the two operands has a non-empty local vocabulary, share
   // with that one (otherwise, throws an exception).
   return {std::move(idTable), resultSortedOn(),
-          ResultTable::getSharedLocalVocabFromNonEmptyOf(*leftResult,
-                                                         *rightResult)};
+          ResultTable::getMergedLocalVocab(*leftResult, *rightResult)};
 }
 
 // _____________________________________________________________________________

src/engine/MultiColumnJoin.cpp

@@ -86,8 +86,7 @@ ResultTable MultiColumnJoin::computeResult() {
   // If only one of the two operands has a non-empty local vocabulary, share
   // with that one (otherwise, throws an exception).
   return {std::move(idTable), resultSortedOn(),
-          ResultTable::getSharedLocalVocabFromNonEmptyOf(*leftResult,
-                                                         *rightResult)};
+          ResultTable::getMergedLocalVocab(*leftResult, *rightResult)};
 }
 
 // _____________________________________________________________________________

src/engine/OptionalJoin.cpp

@@ -115,8 +115,7 @@ ResultTable OptionalJoin::computeResult() {
   // If only one of the two operands has a non-empty local vocabulary, share
   // with that one (otherwise, throws an exception).
   return {std::move(idTable), resultSortedOn(),
-          ResultTable::getSharedLocalVocabFromNonEmptyOf(*leftResult,
-                                                         *rightResult)};
+          ResultTable::getMergedLocalVocab(*leftResult, *rightResult)};
 }
 
 // _____________________________________________________________________________

src/engine/ResultTable.cpp

@@ -23,10 +23,10 @@ string ResultTable::asDebugString() const {
 }
 
 // _____________________________________________________________________________
-auto ResultTable::getSharedLocalVocabFromNonEmptyOf(
-    const ResultTable& resultTable1, const ResultTable& resultTable2)
+auto ResultTable::getMergedLocalVocab(const ResultTable& resultTable1,
+                                      const ResultTable& resultTable2)
     -> SharedLocalVocabWrapper {
-  return getSharedLocalVocabFromNonEmptyOf(
+  return getMergedLocalVocab(
       std::array{std::cref(resultTable1), std::cref(resultTable2)});
 }
 

src/engine/ResultTable.h

@@ -133,46 +133,25 @@ class ResultTable {
     return SharedLocalVocabWrapper{localVocab_};
   }
 
-  // Like `getSharedLocalVocabFrom`, but takes more than one result and assumes
-  // that exactly one of the local vocabularies is empty and gets the shared
-  // local vocab from the non-empty one (if all are empty, arbitrarily share
-  // with the first one).
-  //
-  // TODO: Eventually, we want to be able to merge two non-empty local
-  // vocabularies, but that requires more work since we have to rewrite IDs then
-  // (from the previous separate local vocabularies to the new merged one).
-  static SharedLocalVocabWrapper getSharedLocalVocabFromNonEmptyOf(
+  // Like `getSharedLocalVocabFrom`, but takes more than one result and merges
+  // all the corresponding local vocabs.
+  static SharedLocalVocabWrapper getMergedLocalVocab(
       const ResultTable& resultTable1, const ResultTable& resultTable2);
 
   // Overload for more than two `ResultTables`
   template <std::ranges::forward_range R>
   requires std::convertible_to<std::ranges::range_value_t<R>,
                                const ResultTable&>
-  static SharedLocalVocabWrapper getSharedLocalVocabFromNonEmptyOf(
-      R&& subResults) {
-    AD_CONTRACT_CHECK(!std::ranges::empty(subResults));
-    auto hasNonEmptyVocab = [](const ResultTable& tbl) {
-      return !tbl.localVocab_->empty();
-    };
-    auto numNonEmptyVocabs =
-        std::ranges::count_if(subResults, hasNonEmptyVocab);
-    if (numNonEmptyVocabs > 1) {
-      throw std::runtime_error(
-          "Merging of more than one non-empty local vocabularies is currently "
-          "not supported, please contact the developers");
-    }
-    // The static casts in the following are needed to make this code work for
-    // types that are implicitly convertible to `const ResultTable&`, in
-    // particular `std::reference_wrapper<const ResultTable>`.
-    if (numNonEmptyVocabs == 0) {
-      return SharedLocalVocabWrapper{
-          static_cast<const ResultTable&>(*subResults.begin()).localVocab_};
-    } else {
-      return SharedLocalVocabWrapper{
-          static_cast<const ResultTable&>(
-              *std::ranges::find_if(subResults, hasNonEmptyVocab))
-              .localVocab_};
+  static SharedLocalVocabWrapper getMergedLocalVocab(R&& subResults) {
+    // We could implement the k-way merge much smarter, but in practice we
+    // rarely have more than two inputs, so the performance should be fine (we
+    // are only copying a few shared pointers after all).
+    // TODO<joka921> implement `LocalVocab::merge` in an n-way fashion`.
+    LocalVocab res;
+    for (const ResultTable& table : subResults) {
+      res = LocalVocab::merge(res, *table.localVocab_);
     }
+    return SharedLocalVocabWrapper{std::move(res)};
   }
 
   // Get a (deep) copy of the local vocabulary from the given result. Use this

src/engine/TransitivePath.cpp

@@ -256,7 +256,7 @@ ResultTable TransitivePath::computeResult() {
                     sideRes->idTable());
 
     return {std::move(idTable), resultSortedOn(),
-            ResultTable::getSharedLocalVocabFromNonEmptyOf(*sideRes, *subRes)};
+            ResultTable::getMergedLocalVocab(*sideRes, *subRes)};
   };
 
   if (lhs_.isBoundVariable()) {

src/engine/Union.cpp

@@ -178,9 +178,8 @@ ResultTable Union::computeResult() {
   LOG(DEBUG) << "Union result computation done" << std::endl;
   // If only one of the two operands has a non-empty local vocabulary, share
   // with that one (otherwise, throws an exception).
-  return ResultTable{
-      std::move(idTable), resultSortedOn(),
-      ResultTable::getSharedLocalVocabFromNonEmptyOf(*subRes1, *subRes2)};
+  return ResultTable{std::move(idTable), resultSortedOn(),
+                     ResultTable::getMergedLocalVocab(*subRes1, *subRes2)};
 }
 
 void Union::computeUnion(

src/global/IndexTypes.h

@@ -13,7 +13,15 @@
 // this (very intrusive) renaming doesn't interfere with too many open pull
 // requests.
 using VocabIndex = ad_utility::TypedIndex<uint64_t, "VocabIndex">;
-using LocalVocabIndex = ad_utility::TypedIndex<uint64_t, "LocalVocabIndex">;
+
+// A `std::string` that is aligned to 16 bytes s.t. pointers always end with 4
+// bits that are zero and that are reused for payloads in the `ValueId` class.
+struct alignas(16) StringAligned16 : public std::string {
+  using std::string::basic_string;
+  explicit StringAligned16(std::string s) : std::string{std::move(s)} {}
+};
+// using LocalVocabIndex = ad_utility::TypedIndex<uint64_t, "LocalVocabIndex">;
+using LocalVocabIndex = const StringAligned16*;
 using TextRecordIndex = ad_utility::TypedIndex<uint64_t, "TextRecordIndex">;
 using WordVocabIndex = ad_utility::TypedIndex<uint64_t, "WordVocabIndex">;
 using BlankNodeIndex = ad_utility::TypedIndex<uint64_t, "BlankNodeIndex">;