Create the optimizer framework (#219)

Create the optimizer framework

libursa/CMakeLists.txt

@@ -50,6 +50,8 @@ add_library(
  QueryParser.h
  QueryResult.cpp
  QueryResult.h
+ QueryOptimizer.cpp
+ QueryOptimizer.h
  RawFile.cpp
  RawFile.h
  Responses.cpp

libursa/OnDiskDataset.cpp

@@ -8,6 +8,7 @@
 #include "DatabaseName.h"
 #include "Json.h"
 #include "Query.h"
+#include "QueryOptimizer.h"
 #include "spdlog/fmt/ostr.h"
 #include "spdlog/spdlog.h"
 
@@ -91,7 +92,8 @@ void OnDiskDataset::execute(const Query &query, ResultWriter *out,
  for (const auto &ndx : get_indexes()) {
  types_to_query.emplace(ndx.index_type());
  }
- const Query plan = query.plan(types_to_query);
+ Query plan = query.plan(types_to_query);
+ plan = q_optimize(std::move(plan));
  spdlog::debug("PLAN: {}", plan);
 
  QueryResult result = this->query(plan, counters);

libursa/Query.cpp

@@ -18,7 +18,14 @@ const std::vector<Query> &Query::as_queries() const {
  type != QueryType::MIN_OF) {
  throw std::runtime_error("This query doesn\'t contain subqueries.");
  }
+ return queries;
+}
 
+std::vector<Query> &Query::as_queries() {
+ if (type != QueryType::AND && type != QueryType::OR &&
+ type != QueryType::MIN_OF) {
+ throw std::runtime_error("This query doesn\'t contain subqueries.");
+ }
  return queries;
 }
 

libursa/Query.h

@@ -22,8 +22,8 @@ class PrimitiveQuery {
  PrimitiveQuery(IndexType itype, TriGram trigram)
  : itype(itype), trigram(trigram) {}
 
- const IndexType itype;
- const TriGram trigram;
+ IndexType itype;
+ TriGram trigram;
 
  // We want to use PrimitiveQuery in STL containers, and this means they
  // must be comparable using <. Specific order doesn't matter.
@@ -51,6 +51,7 @@ class Query {
  Query &operator=(Query &&) = default;
 
  const std::vector<Query> &as_queries() const;
+ std::vector<Query> &as_queries();
  const QString &as_value() const;
  uint32_t as_count() const;
  std::string as_string_repr() const;

libursa/QueryOptimizer.cpp

@@ -0,0 +1,31 @@
+#include "QueryOptimizer.h"
+
+#include <vector>
+
+// Run the optimization pases on subqueries.
+// After this step, every subquery should be maximally optimized,
+// So I believe there's no need to run this in a loop.
+Query simplify_subqueries(Query &&q) {
+ // q_optimize ensures QueryType is not PRIMITIVE already
+ std::vector<Query> newqueries;
+ for (auto &&query : q.as_queries()) {
+ newqueries.emplace_back(q_optimize(std::move(query)));
+ }
+ if (q.get_type() == QueryType::MIN_OF) {
+ return q_min_of(q.as_count(), std::move(newqueries));
+ }
+ return std::move(Query(q.get_type(), std::move(newqueries)));
+}
+
+Query q_optimize(Query &&q) {
+ if (q.get_type() == QueryType::PRIMITIVE) {
+ // Nothing to improve here.
+ return std::move(q);
+ }
+
+ q = simplify_subqueries(std::move(q));
+
+ // Optimization passes will be added here later.
+
+ return std::move(q);
+}

libursa/QueryOptimizer.h

@@ -0,0 +1,8 @@
+#pragma once
+
+#include "Query.h"
+
+// Optimizes a query, and returns the optimized version.
+// Optimizations try to simplify the expression in various ways to make the
+// execution faster - for example by enabling short-circuiting in some places.
+Query q_optimize(Query &&query);

libursa/Version.h.in

@@ -8,5 +8,6 @@ constexpr std::string_view ursadb_format_version = "1.5.0";
 
 // Project version.
 // Consider updating the version tag when doing PRs.
-constexpr std::string_view ursadb_version_string =
- "@PROJECT_VERSION@+debuglogs";
+// clang-format off
+constexpr std::string_view ursadb_version_string = "@PROJECT_VERSION@+opt0";
+// clang-format on