cluster_linearize.h raw

   1  // Copyright (c) The Limenka developers
   2  // Distributed under the MIT software license, see the accompanying
   3  // file COPYING or http://www.opensource.org/licenses/mit-license.php.
   4  
   5  #ifndef LIMENKA_CLUSTER_LINEARIZE_H
   6  #define LIMENKA_CLUSTER_LINEARIZE_H
   7  
   8  #include <algorithm>
   9  #include <numeric>
  10  #include <optional>
  11  #include <stdint.h>
  12  #include <vector>
  13  #include <utility>
  14  
  15  #include <random.h>
  16  #include <span.h>
  17  #include <util/feefrac.h>
  18  #include <util/vecdeque.h>
  19  
  20  namespace cluster_linearize {
  21  
  22  /** Data type to represent transaction indices in clusters. */
  23  using ClusterIndex = uint32_t;
  24  
  25  /** Data structure that holds a transaction graph's preprocessed data (fee, size, ancestors,
  26   *  descendants). */
  27  template<typename SetType>
  28  class DepGraph
  29  {
  30      /** Information about a single transaction. */
  31      struct Entry
  32      {
  33          /** Fee and size of transaction itself. */
  34          FeeFrac feerate;
  35          /** All ancestors of the transaction (including itself). */
  36          SetType ancestors;
  37          /** All descendants of the transaction (including itself). */
  38          SetType descendants;
  39  
  40          /** Equality operator (primarily for for testing purposes). */
  41          friend bool operator==(const Entry&, const Entry&) noexcept = default;
  42  
  43          /** Construct an empty entry. */
  44          Entry() noexcept = default;
  45          /** Construct an entry with a given feerate, ancestor set, descendant set. */
  46          Entry(const FeeFrac& f, const SetType& a, const SetType& d) noexcept : feerate(f), ancestors(a), descendants(d) {}
  47      };
  48  
  49      /** Data for each transaction. */
  50      std::vector<Entry> entries;
  51  
  52      /** Which positions are used. */
  53      SetType m_used;
  54  
  55  public:
  56      /** Equality operator (primarily for testing purposes). */
  57      friend bool operator==(const DepGraph& a, const DepGraph& b) noexcept
  58      {
  59          if (a.m_used != b.m_used) return false;
  60          // Only compare the used positions within the entries vector.
  61          for (auto idx : a.m_used) {
  62              if (a.entries[idx] != b.entries[idx]) return false;
  63          }
  64          return true;
  65      }
  66  
  67      // Default constructors.
  68      DepGraph() noexcept = default;
  69      DepGraph(const DepGraph&) noexcept = default;
  70      DepGraph(DepGraph&&) noexcept = default;
  71      DepGraph& operator=(const DepGraph&) noexcept = default;
  72      DepGraph& operator=(DepGraph&&) noexcept = default;
  73  
  74      /** Construct a DepGraph object given another DepGraph and a mapping from old to new.
  75       *
  76       * @param depgraph   The original DepGraph that is being remapped.
  77       *
  78       * @param mapping    A Span such that mapping[i] gives the position in the new DepGraph
  79       *                   for position i in the old depgraph. Its size must be equal to
  80       *                   depgraph.PositionRange(). The value of mapping[i] is ignored if
  81       *                   position i is a hole in depgraph (i.e., if !depgraph.Positions()[i]).
  82       *
  83       * @param pos_range  The PositionRange() for the new DepGraph. It must equal the largest
  84       *                   value in mapping for any used position in depgraph plus 1, or 0 if
  85       *                   depgraph.TxCount() == 0.
  86       *
  87       * Complexity: O(N^2) where N=depgraph.TxCount().
  88       */
  89      DepGraph(const DepGraph<SetType>& depgraph, Span<const ClusterIndex> mapping, ClusterIndex pos_range) noexcept : entries(pos_range)
  90      {
  91          Assume(mapping.size() == depgraph.PositionRange());
  92          Assume((pos_range == 0) == (depgraph.TxCount() == 0));
  93          for (ClusterIndex i : depgraph.Positions()) {
  94              auto new_idx = mapping[i];
  95              Assume(new_idx < pos_range);
  96              // Add transaction.
  97              entries[new_idx].ancestors = SetType::Singleton(new_idx);
  98              entries[new_idx].descendants = SetType::Singleton(new_idx);
  99              m_used.Set(new_idx);
 100              // Fill in fee and size.
 101              entries[new_idx].feerate = depgraph.entries[i].feerate;
 102          }
 103          for (ClusterIndex i : depgraph.Positions()) {
 104              // Fill in dependencies by mapping direct parents.
 105              SetType parents;
 106              for (auto j : depgraph.GetReducedParents(i)) parents.Set(mapping[j]);
 107              AddDependencies(parents, mapping[i]);
 108          }
 109          // Verify that the provided pos_range was correct (no unused positions at the end).
 110          Assume(m_used.None() ? (pos_range == 0) : (pos_range == m_used.Last() + 1));
 111      }
 112  
 113      /** Get the set of transactions positions in use. Complexity: O(1). */
 114      const SetType& Positions() const noexcept { return m_used; }
 115      /** Get the range of positions in this DepGraph. All entries in Positions() are in [0, PositionRange() - 1]. */
 116      ClusterIndex PositionRange() const noexcept { return entries.size(); }
 117      /** Get the number of transactions in the graph. Complexity: O(1). */
 118      auto TxCount() const noexcept { return m_used.Count(); }
 119      /** Get the feerate of a given transaction i. Complexity: O(1). */
 120      const FeeFrac& FeeRate(ClusterIndex i) const noexcept { return entries[i].feerate; }
 121      /** Get the mutable feerate of a given transaction i. Complexity: O(1). */
 122      FeeFrac& FeeRate(ClusterIndex i) noexcept { return entries[i].feerate; }
 123      /** Get the ancestors of a given transaction i. Complexity: O(1). */
 124      const SetType& Ancestors(ClusterIndex i) const noexcept { return entries[i].ancestors; }
 125      /** Get the descendants of a given transaction i. Complexity: O(1). */
 126      const SetType& Descendants(ClusterIndex i) const noexcept { return entries[i].descendants; }
 127  
 128      /** Add a new unconnected transaction to this transaction graph (in the first available
 129       *  position), and return its ClusterIndex.
 130       *
 131       * Complexity: O(1) (amortized, due to resizing of backing vector).
 132       */
 133      ClusterIndex AddTransaction(const FeeFrac& feefrac) noexcept
 134      {
 135          static constexpr auto ALL_POSITIONS = SetType::Fill(SetType::Size());
 136          auto available = ALL_POSITIONS - m_used;
 137          Assume(available.Any());
 138          ClusterIndex new_idx = available.First();
 139          if (new_idx == entries.size()) {
 140              entries.emplace_back(feefrac, SetType::Singleton(new_idx), SetType::Singleton(new_idx));
 141          } else {
 142              entries[new_idx] = Entry(feefrac, SetType::Singleton(new_idx), SetType::Singleton(new_idx));
 143          }
 144          m_used.Set(new_idx);
 145          return new_idx;
 146      }
 147  
 148      /** Remove the specified positions from this DepGraph.
 149       *
 150       * The specified positions will no longer be part of Positions(), and dependencies with them are
 151       * removed. Note that due to DepGraph only tracking ancestors/descendants (and not direct
 152       * dependencies), if a parent is removed while a grandparent remains, the grandparent will
 153       * remain an ancestor.
 154       *
 155       * Complexity: O(N) where N=TxCount().
 156       */
 157      void RemoveTransactions(const SetType& del) noexcept
 158      {
 159          m_used -= del;
 160          // Remove now-unused trailing entries.
 161          while (!entries.empty() && !m_used[entries.size() - 1]) {
 162              entries.pop_back();
 163          }
 164          // Remove the deleted transactions from ancestors/descendants of other transactions. Note
 165          // that the deleted positions will retain old feerate and dependency information. This does
 166          // not matter as they will be overwritten by AddTransaction if they get used again.
 167          for (auto& entry : entries) {
 168              entry.ancestors &= m_used;
 169              entry.descendants &= m_used;
 170          }
 171      }
 172  
 173      /** Modify this transaction graph, adding multiple parents to a specified child.
 174       *
 175       * Complexity: O(N) where N=TxCount().
 176       */
 177      void AddDependencies(const SetType& parents, ClusterIndex child) noexcept
 178      {
 179          Assume(m_used[child]);
 180          Assume(parents.IsSubsetOf(m_used));
 181          // Compute the ancestors of parents that are not already ancestors of child.
 182          SetType par_anc;
 183          for (auto par : parents - Ancestors(child)) {
 184              par_anc |= Ancestors(par);
 185          }
 186          par_anc -= Ancestors(child);
 187          // Bail out if there are no such ancestors.
 188          if (par_anc.None()) return;
 189          // To each such ancestor, add as descendants the descendants of the child.
 190          const auto& chl_des = entries[child].descendants;
 191          for (auto anc_of_par : par_anc) {
 192              entries[anc_of_par].descendants |= chl_des;
 193          }
 194          // To each descendant of the child, add those ancestors.
 195          for (auto dec_of_chl : Descendants(child)) {
 196              entries[dec_of_chl].ancestors |= par_anc;
 197          }
 198      }
 199  
 200      /** Compute the (reduced) set of parents of node i in this graph.
 201       *
 202       * This returns the minimal subset of the parents of i whose ancestors together equal all of
 203       * i's ancestors (unless i is part of a cycle of dependencies). Note that DepGraph does not
 204       * store the set of parents; this information is inferred from the ancestor sets.
 205       *
 206       * Complexity: O(N) where N=Ancestors(i).Count() (which is bounded by TxCount()).
 207       */
 208      SetType GetReducedParents(ClusterIndex i) const noexcept
 209      {
 210          SetType parents = Ancestors(i);
 211          parents.Reset(i);
 212          for (auto parent : parents) {
 213              if (parents[parent]) {
 214                  parents -= Ancestors(parent);
 215                  parents.Set(parent);
 216              }
 217          }
 218          return parents;
 219      }
 220  
 221      /** Compute the (reduced) set of children of node i in this graph.
 222       *
 223       * This returns the minimal subset of the children of i whose descendants together equal all of
 224       * i's descendants (unless i is part of a cycle of dependencies). Note that DepGraph does not
 225       * store the set of children; this information is inferred from the descendant sets.
 226       *
 227       * Complexity: O(N) where N=Descendants(i).Count() (which is bounded by TxCount()).
 228       */
 229      SetType GetReducedChildren(ClusterIndex i) const noexcept
 230      {
 231          SetType children = Descendants(i);
 232          children.Reset(i);
 233          for (auto child : children) {
 234              if (children[child]) {
 235                  children -= Descendants(child);
 236                  children.Set(child);
 237              }
 238          }
 239          return children;
 240      }
 241  
 242      /** Compute the aggregate feerate of a set of nodes in this graph.
 243       *
 244       * Complexity: O(N) where N=elems.Count().
 245       **/
 246      FeeFrac FeeRate(const SetType& elems) const noexcept
 247      {
 248          FeeFrac ret;
 249          for (auto pos : elems) ret += entries[pos].feerate;
 250          return ret;
 251      }
 252  
 253      /** Find some connected component within the subset "todo" of this graph.
 254       *
 255       * Specifically, this finds the connected component which contains the first transaction of
 256       * todo (if any).
 257       *
 258       * Two transactions are considered connected if they are both in `todo`, and one is an ancestor
 259       * of the other in the entire graph (so not just within `todo`), or transitively there is a
 260       * path of transactions connecting them. This does mean that if `todo` contains a transaction
 261       * and a grandparent, but misses the parent, they will still be part of the same component.
 262       *
 263       * Complexity: O(ret.Count()).
 264       */
 265      SetType FindConnectedComponent(const SetType& todo) const noexcept
 266      {
 267          if (todo.None()) return todo;
 268          auto to_add = SetType::Singleton(todo.First());
 269          SetType ret;
 270          do {
 271              SetType old = ret;
 272              for (auto add : to_add) {
 273                  ret |= Descendants(add);
 274                  ret |= Ancestors(add);
 275              }
 276              ret &= todo;
 277              to_add = ret - old;
 278          } while (to_add.Any());
 279          return ret;
 280      }
 281  
 282      /** Determine if a subset is connected.
 283       *
 284       * Complexity: O(subset.Count()).
 285       */
 286      bool IsConnected(const SetType& subset) const noexcept
 287      {
 288          return FindConnectedComponent(subset) == subset;
 289      }
 290  
 291      /** Determine if this entire graph is connected.
 292       *
 293       * Complexity: O(TxCount()).
 294       */
 295      bool IsConnected() const noexcept { return IsConnected(m_used); }
 296  
 297      /** Append the entries of select to list in a topologically valid order.
 298       *
 299       * Complexity: O(select.Count() * log(select.Count())).
 300       */
 301      void AppendTopo(std::vector<ClusterIndex>& list, const SetType& select) const noexcept
 302      {
 303          ClusterIndex old_len = list.size();
 304          for (auto i : select) list.push_back(i);
 305          std::sort(list.begin() + old_len, list.end(), [&](ClusterIndex a, ClusterIndex b) noexcept {
 306              const auto a_anc_count = entries[a].ancestors.Count();
 307              const auto b_anc_count = entries[b].ancestors.Count();
 308              if (a_anc_count != b_anc_count) return a_anc_count < b_anc_count;
 309              return a < b;
 310          });
 311      }
 312  };
 313  
 314  /** A set of transactions together with their aggregate feerate. */
 315  template<typename SetType>
 316  struct SetInfo
 317  {
 318      /** The transactions in the set. */
 319      SetType transactions;
 320      /** Their combined fee and size. */
 321      FeeFrac feerate;
 322  
 323      /** Construct a SetInfo for the empty set. */
 324      SetInfo() noexcept = default;
 325  
 326      /** Construct a SetInfo for a specified set and feerate. */
 327      SetInfo(const SetType& txn, const FeeFrac& fr) noexcept : transactions(txn), feerate(fr) {}
 328  
 329      /** Construct a SetInfo for a given transaction in a depgraph. */
 330      explicit SetInfo(const DepGraph<SetType>& depgraph, ClusterIndex pos) noexcept :
 331          transactions(SetType::Singleton(pos)), feerate(depgraph.FeeRate(pos)) {}
 332  
 333      /** Construct a SetInfo for a set of transactions in a depgraph. */
 334      explicit SetInfo(const DepGraph<SetType>& depgraph, const SetType& txn) noexcept :
 335          transactions(txn), feerate(depgraph.FeeRate(txn)) {}
 336  
 337      /** Add a transaction to this SetInfo (which must not yet be in it). */
 338      void Set(const DepGraph<SetType>& depgraph, ClusterIndex pos) noexcept
 339      {
 340          Assume(!transactions[pos]);
 341          transactions.Set(pos);
 342          feerate += depgraph.FeeRate(pos);
 343      }
 344  
 345      /** Add the transactions of other to this SetInfo (no overlap allowed). */
 346      SetInfo& operator|=(const SetInfo& other) noexcept
 347      {
 348          Assume(!transactions.Overlaps(other.transactions));
 349          transactions |= other.transactions;
 350          feerate += other.feerate;
 351          return *this;
 352      }
 353  
 354      /** Construct a new SetInfo equal to this, with more transactions added (which may overlap
 355       *  with the existing transactions in the SetInfo). */
 356      [[nodiscard]] SetInfo Add(const DepGraph<SetType>& depgraph, const SetType& txn) const noexcept
 357      {
 358          return {transactions | txn, feerate + depgraph.FeeRate(txn - transactions)};
 359      }
 360  
 361      /** Swap two SetInfo objects. */
 362      friend void swap(SetInfo& a, SetInfo& b) noexcept
 363      {
 364          swap(a.transactions, b.transactions);
 365          swap(a.feerate, b.feerate);
 366      }
 367  
 368      /** Permit equality testing. */
 369      friend bool operator==(const SetInfo&, const SetInfo&) noexcept = default;
 370  };
 371  
 372  /** Compute the feerates of the chunks of linearization. */
 373  template<typename SetType>
 374  std::vector<FeeFrac> ChunkLinearization(const DepGraph<SetType>& depgraph, Span<const ClusterIndex> linearization) noexcept
 375  {
 376      std::vector<FeeFrac> ret;
 377      for (ClusterIndex i : linearization) {
 378          /** The new chunk to be added, initially a singleton. */
 379          auto new_chunk = depgraph.FeeRate(i);
 380          // As long as the new chunk has a higher feerate than the last chunk so far, absorb it.
 381          while (!ret.empty() && new_chunk >> ret.back()) {
 382              new_chunk += ret.back();
 383              ret.pop_back();
 384          }
 385          // Actually move that new chunk into the chunking.
 386          ret.push_back(std::move(new_chunk));
 387      }
 388      return ret;
 389  }
 390  
 391  /** Data structure encapsulating the chunking of a linearization, permitting removal of subsets. */
 392  template<typename SetType>
 393  class LinearizationChunking
 394  {
 395      /** The depgraph this linearization is for. */
 396      const DepGraph<SetType>& m_depgraph;
 397  
 398      /** The linearization we started from, possibly with removed prefix stripped. */
 399      Span<const ClusterIndex> m_linearization;
 400  
 401      /** Chunk sets and their feerates, of what remains of the linearization. */
 402      std::vector<SetInfo<SetType>> m_chunks;
 403  
 404      /** How large a prefix of m_chunks corresponds to removed transactions. */
 405      ClusterIndex m_chunks_skip{0};
 406  
 407      /** Which transactions remain in the linearization. */
 408      SetType m_todo;
 409  
 410      /** Fill the m_chunks variable, and remove the done prefix of m_linearization. */
 411      void BuildChunks() noexcept
 412      {
 413          // Caller must clear m_chunks.
 414          Assume(m_chunks.empty());
 415  
 416          // Chop off the initial part of m_linearization that is already done.
 417          while (!m_linearization.empty() && !m_todo[m_linearization.front()]) {
 418              m_linearization = m_linearization.subspan(1);
 419          }
 420  
 421          // Iterate over the remaining entries in m_linearization. This is effectively the same
 422          // algorithm as ChunkLinearization, but supports skipping parts of the linearization and
 423          // keeps track of the sets themselves instead of just their feerates.
 424          for (auto idx : m_linearization) {
 425              if (!m_todo[idx]) continue;
 426              // Start with an initial chunk containing just element idx.
 427              SetInfo add(m_depgraph, idx);
 428              // Absorb existing final chunks into add while they have lower feerate.
 429              while (!m_chunks.empty() && add.feerate >> m_chunks.back().feerate) {
 430                  add |= m_chunks.back();
 431                  m_chunks.pop_back();
 432              }
 433              // Remember new chunk.
 434              m_chunks.push_back(std::move(add));
 435          }
 436      }
 437  
 438  public:
 439      /** Initialize a LinearizationSubset object for a given length of linearization. */
 440      explicit LinearizationChunking(const DepGraph<SetType>& depgraph LIFETIMEBOUND, Span<const ClusterIndex> lin LIFETIMEBOUND) noexcept :
 441          m_depgraph(depgraph), m_linearization(lin)
 442      {
 443          // Mark everything in lin as todo still.
 444          for (auto i : m_linearization) m_todo.Set(i);
 445          // Compute the initial chunking.
 446          m_chunks.reserve(depgraph.TxCount());
 447          BuildChunks();
 448      }
 449  
 450      /** Determine how many chunks remain in the linearization. */
 451      ClusterIndex NumChunksLeft() const noexcept { return m_chunks.size() - m_chunks_skip; }
 452  
 453      /** Access a chunk. Chunk 0 is the highest-feerate prefix of what remains. */
 454      const SetInfo<SetType>& GetChunk(ClusterIndex n) const noexcept
 455      {
 456          Assume(n + m_chunks_skip < m_chunks.size());
 457          return m_chunks[n + m_chunks_skip];
 458      }
 459  
 460      /** Remove some subset of transactions from the linearization. */
 461      void MarkDone(SetType subset) noexcept
 462      {
 463          Assume(subset.Any());
 464          Assume(subset.IsSubsetOf(m_todo));
 465          m_todo -= subset;
 466          if (GetChunk(0).transactions == subset) {
 467              // If the newly done transactions exactly match the first chunk of the remainder of
 468              // the linearization, we do not need to rechunk; just remember to skip one
 469              // additional chunk.
 470              ++m_chunks_skip;
 471              // With subset marked done, some prefix of m_linearization will be done now. How long
 472              // that prefix is depends on how many done elements were interspersed with subset,
 473              // but at least as many transactions as there are in subset.
 474              m_linearization = m_linearization.subspan(subset.Count());
 475          } else {
 476              // Otherwise rechunk what remains of m_linearization.
 477              m_chunks.clear();
 478              m_chunks_skip = 0;
 479              BuildChunks();
 480          }
 481      }
 482  
 483      /** Find the shortest intersection between subset and the prefixes of remaining chunks
 484       *  of the linearization that has a feerate not below subset's.
 485       *
 486       * This is a crucial operation in guaranteeing improvements to linearizations. If subset has
 487       * a feerate not below GetChunk(0)'s, then moving IntersectPrefixes(subset) to the front of
 488       * (what remains of) the linearization is guaranteed not to make it worse at any point.
 489       *
 490       * See https://delvinglimenka.org/t/introduction-to-cluster-linearization/1032 for background.
 491       */
 492      SetInfo<SetType> IntersectPrefixes(const SetInfo<SetType>& subset) const noexcept
 493      {
 494          Assume(subset.transactions.IsSubsetOf(m_todo));
 495          SetInfo<SetType> accumulator;
 496          // Iterate over all chunks of the remaining linearization.
 497          for (ClusterIndex i = 0; i < NumChunksLeft(); ++i) {
 498              // Find what (if any) intersection the chunk has with subset.
 499              const SetType to_add = GetChunk(i).transactions & subset.transactions;
 500              if (to_add.Any()) {
 501                  // If adding that to accumulator makes us hit all of subset, we are done as no
 502                  // shorter intersection with higher/equal feerate exists.
 503                  accumulator.transactions |= to_add;
 504                  if (accumulator.transactions == subset.transactions) break;
 505                  // Otherwise update the accumulator feerate.
 506                  accumulator.feerate += m_depgraph.FeeRate(to_add);
 507                  // If that does result in something better, or something with the same feerate but
 508                  // smaller, return that. Even if a longer, higher-feerate intersection exists, it
 509                  // does not hurt to return the shorter one (the remainder of the longer intersection
 510                  // will generally be found in the next call to Intersect, but even if not, it is not
 511                  // required for the improvement guarantee this function makes).
 512                  if (!(accumulator.feerate << subset.feerate)) return accumulator;
 513              }
 514          }
 515          return subset;
 516      }
 517  };
 518  
 519  /** Class encapsulating the state needed to find the best remaining ancestor set.
 520   *
 521   * It is initialized for an entire DepGraph, and parts of the graph can be dropped by calling
 522   * MarkDone.
 523   *
 524   * As long as any part of the graph remains, FindCandidateSet() can be called which will return a
 525   * SetInfo with the highest-feerate ancestor set that remains (an ancestor set is a single
 526   * transaction together with all its remaining ancestors).
 527   */
 528  template<typename SetType>
 529  class AncestorCandidateFinder
 530  {
 531      /** Internal dependency graph. */
 532      const DepGraph<SetType>& m_depgraph;
 533      /** Which transaction are left to include. */
 534      SetType m_todo;
 535      /** Precomputed ancestor-set feerates (only kept up-to-date for indices in m_todo). */
 536      std::vector<FeeFrac> m_ancestor_set_feerates;
 537  
 538  public:
 539      /** Construct an AncestorCandidateFinder for a given cluster.
 540       *
 541       * Complexity: O(N^2) where N=depgraph.TxCount().
 542       */
 543      AncestorCandidateFinder(const DepGraph<SetType>& depgraph LIFETIMEBOUND) noexcept :
 544          m_depgraph(depgraph),
 545          m_todo{depgraph.Positions()},
 546          m_ancestor_set_feerates(depgraph.PositionRange())
 547      {
 548          // Precompute ancestor-set feerates.
 549          for (ClusterIndex i : m_depgraph.Positions()) {
 550              /** The remaining ancestors for transaction i. */
 551              SetType anc_to_add = m_depgraph.Ancestors(i);
 552              FeeFrac anc_feerate;
 553              // Reuse accumulated feerate from first ancestor, if usable.
 554              Assume(anc_to_add.Any());
 555              ClusterIndex first = anc_to_add.First();
 556              if (first < i) {
 557                  anc_feerate = m_ancestor_set_feerates[first];
 558                  Assume(!anc_feerate.IsEmpty());
 559                  anc_to_add -= m_depgraph.Ancestors(first);
 560              }
 561              // Add in other ancestors (which necessarily include i itself).
 562              Assume(anc_to_add[i]);
 563              anc_feerate += m_depgraph.FeeRate(anc_to_add);
 564              // Store the result.
 565              m_ancestor_set_feerates[i] = anc_feerate;
 566          }
 567      }
 568  
 569      /** Remove a set of transactions from the set of to-be-linearized ones.
 570       *
 571       * The same transaction may not be MarkDone()'d twice.
 572       *
 573       * Complexity: O(N*M) where N=depgraph.TxCount(), M=select.Count().
 574       */
 575      void MarkDone(SetType select) noexcept
 576      {
 577          Assume(select.Any());
 578          Assume(select.IsSubsetOf(m_todo));
 579          m_todo -= select;
 580          for (auto i : select) {
 581              auto feerate = m_depgraph.FeeRate(i);
 582              for (auto j : m_depgraph.Descendants(i) & m_todo) {
 583                  m_ancestor_set_feerates[j] -= feerate;
 584              }
 585          }
 586      }
 587  
 588      /** Check whether any unlinearized transactions remain. */
 589      bool AllDone() const noexcept
 590      {
 591          return m_todo.None();
 592      }
 593  
 594      /** Count the number of remaining unlinearized transactions. */
 595      ClusterIndex NumRemaining() const noexcept
 596      {
 597          return m_todo.Count();
 598      }
 599  
 600      /** Find the best (highest-feerate, smallest among those in case of a tie) ancestor set
 601       *  among the remaining transactions. Requires !AllDone().
 602       *
 603       * Complexity: O(N) where N=depgraph.TxCount();
 604       */
 605      SetInfo<SetType> FindCandidateSet() const noexcept
 606      {
 607          Assume(!AllDone());
 608          std::optional<ClusterIndex> best;
 609          for (auto i : m_todo) {
 610              if (best.has_value()) {
 611                  Assume(!m_ancestor_set_feerates[i].IsEmpty());
 612                  if (!(m_ancestor_set_feerates[i] > m_ancestor_set_feerates[*best])) continue;
 613              }
 614              best = i;
 615          }
 616          Assume(best.has_value());
 617          return {m_depgraph.Ancestors(*best) & m_todo, m_ancestor_set_feerates[*best]};
 618      }
 619  };
 620  
 621  /** Class encapsulating the state needed to perform search for good candidate sets.
 622   *
 623   * It is initialized for an entire DepGraph, and parts of the graph can be dropped by calling
 624   * MarkDone().
 625   *
 626   * As long as any part of the graph remains, FindCandidateSet() can be called to perform a search
 627   * over the set of topologically-valid subsets of that remainder, with a limit on how many
 628   * combinations are tried.
 629   */
 630  template<typename SetType>
 631  class SearchCandidateFinder
 632  {
 633      /** Internal RNG. */
 634      InsecureRandomContext m_rng;
 635      /** m_sorted_to_original[i] is the original position that sorted transaction position i had. */
 636      std::vector<ClusterIndex> m_sorted_to_original;
 637      /** m_original_to_sorted[i] is the sorted position original transaction position i has. */
 638      std::vector<ClusterIndex> m_original_to_sorted;
 639      /** Internal dependency graph for the cluster (with transactions in decreasing individual
 640       *  feerate order). */
 641      DepGraph<SetType> m_sorted_depgraph;
 642      /** Which transactions are left to do (indices in m_sorted_depgraph's order). */
 643      SetType m_todo;
 644  
 645      /** Given a set of transactions with sorted indices, get their original indices. */
 646      SetType SortedToOriginal(const SetType& arg) const noexcept
 647      {
 648          SetType ret;
 649          for (auto pos : arg) ret.Set(m_sorted_to_original[pos]);
 650          return ret;
 651      }
 652  
 653      /** Given a set of transactions with original indices, get their sorted indices. */
 654      SetType OriginalToSorted(const SetType& arg) const noexcept
 655      {
 656          SetType ret;
 657          for (auto pos : arg) ret.Set(m_original_to_sorted[pos]);
 658          return ret;
 659      }
 660  
 661  public:
 662      /** Construct a candidate finder for a graph.
 663       *
 664       * @param[in] depgraph   Dependency graph for the to-be-linearized cluster.
 665       * @param[in] rng_seed   A random seed to control the search order.
 666       *
 667       * Complexity: O(N^2) where N=depgraph.Count().
 668       */
 669      SearchCandidateFinder(const DepGraph<SetType>& depgraph, uint64_t rng_seed) noexcept :
 670          m_rng(rng_seed),
 671          m_sorted_to_original(depgraph.TxCount()),
 672          m_original_to_sorted(depgraph.PositionRange())
 673      {
 674          // Determine reordering mapping, by sorting by decreasing feerate. Unused positions are
 675          // not included, as they will never be looked up anyway.
 676          ClusterIndex sorted_pos{0};
 677          for (auto i : depgraph.Positions()) {
 678              m_sorted_to_original[sorted_pos++] = i;
 679          }
 680          std::sort(m_sorted_to_original.begin(), m_sorted_to_original.end(), [&](auto a, auto b) {
 681              auto feerate_cmp = depgraph.FeeRate(a) <=> depgraph.FeeRate(b);
 682              if (feerate_cmp == 0) return a < b;
 683              return feerate_cmp > 0;
 684          });
 685          // Compute reverse mapping.
 686          for (ClusterIndex i = 0; i < m_sorted_to_original.size(); ++i) {
 687              m_original_to_sorted[m_sorted_to_original[i]] = i;
 688          }
 689          // Compute reordered dependency graph.
 690          m_sorted_depgraph = DepGraph(depgraph, m_original_to_sorted, m_sorted_to_original.size());
 691          m_todo = m_sorted_depgraph.Positions();
 692      }
 693  
 694      /** Check whether any unlinearized transactions remain. */
 695      bool AllDone() const noexcept
 696      {
 697          return m_todo.None();
 698      }
 699  
 700      /** Find a high-feerate topologically-valid subset of what remains of the cluster.
 701       *  Requires !AllDone().
 702       *
 703       * @param[in] max_iterations  The maximum number of optimization steps that will be performed.
 704       * @param[in] best            A set/feerate pair with an already-known good candidate. This may
 705       *                            be empty.
 706       * @return                    A pair of:
 707       *                            - The best (highest feerate, smallest size as tiebreaker)
 708       *                              topologically valid subset (and its feerate) that was
 709       *                              encountered during search. It will be at least as good as the
 710       *                              best passed in (if not empty).
 711       *                            - The number of optimization steps that were performed. This will
 712       *                              be <= max_iterations. If strictly < max_iterations, the
 713       *                              returned subset is optimal.
 714       *
 715       * Complexity: possibly O(N * min(max_iterations, sqrt(2^N))) where N=depgraph.TxCount().
 716       */
 717      std::pair<SetInfo<SetType>, uint64_t> FindCandidateSet(uint64_t max_iterations, SetInfo<SetType> best) noexcept
 718      {
 719          Assume(!AllDone());
 720  
 721          // Convert the provided best to internal sorted indices.
 722          best.transactions = OriginalToSorted(best.transactions);
 723  
 724          /** Type for work queue items. */
 725          struct WorkItem
 726          {
 727              /** Set of transactions definitely included (and its feerate). This must be a subset
 728               *  of m_todo, and be topologically valid (includes all in-m_todo ancestors of
 729               *  itself). */
 730              SetInfo<SetType> inc;
 731              /** Set of undecided transactions. This must be a subset of m_todo, and have no overlap
 732               *  with inc. The set (inc | und) must be topologically valid. */
 733              SetType und;
 734              /** (Only when inc is not empty) The best feerate of any superset of inc that is also a
 735               *  subset of (inc | und), without requiring it to be topologically valid. It forms a
 736               *  conservative upper bound on how good a set this work item can give rise to.
 737               *  Transactions whose feerate is below best's are ignored when determining this value,
 738               *  which means it may technically be an underestimate, but if so, this work item
 739               *  cannot result in something that beats best anyway. */
 740              FeeFrac pot_feerate;
 741  
 742              /** Construct a new work item. */
 743              WorkItem(SetInfo<SetType>&& i, SetType&& u, FeeFrac&& p_f) noexcept :
 744                  inc(std::move(i)), und(std::move(u)), pot_feerate(std::move(p_f))
 745              {
 746                  Assume(pot_feerate.IsEmpty() == inc.feerate.IsEmpty());
 747              }
 748  
 749              /** Swap two WorkItems. */
 750              void Swap(WorkItem& other) noexcept
 751              {
 752                  swap(inc, other.inc);
 753                  swap(und, other.und);
 754                  swap(pot_feerate, other.pot_feerate);
 755              }
 756          };
 757  
 758          /** The queue of work items. */
 759          VecDeque<WorkItem> queue;
 760          queue.reserve(std::max<size_t>(256, 2 * m_todo.Count()));
 761  
 762          // Create initial entries per connected component of m_todo. While clusters themselves are
 763          // generally connected, this is not necessarily true after some parts have already been
 764          // removed from m_todo. Without this, effort can be wasted on searching "inc" sets that
 765          // span multiple components.
 766          auto to_cover = m_todo;
 767          do {
 768              auto component = m_sorted_depgraph.FindConnectedComponent(to_cover);
 769              to_cover -= component;
 770              // If best is not provided, set it to the first component, so that during the work
 771              // processing loop below, and during the add_fn/split_fn calls, we do not need to deal
 772              // with the best=empty case.
 773              if (best.feerate.IsEmpty()) best = SetInfo(m_sorted_depgraph, component);
 774              queue.emplace_back(/*inc=*/SetInfo<SetType>{},
 775                                 /*und=*/std::move(component),
 776                                 /*pot_feerate=*/FeeFrac{});
 777          } while (to_cover.Any());
 778  
 779          /** Local copy of the iteration limit. */
 780          uint64_t iterations_left = max_iterations;
 781  
 782          /** The set of transactions in m_todo which have feerate > best's. */
 783          SetType imp = m_todo;
 784          while (imp.Any()) {
 785              ClusterIndex check = imp.Last();
 786              if (m_sorted_depgraph.FeeRate(check) >> best.feerate) break;
 787              imp.Reset(check);
 788          }
 789  
 790          /** Internal function to add an item to the queue of elements to explore if there are any
 791           *  transactions left to split on, possibly improving it before doing so, and to update
 792           *  best/imp.
 793           *
 794           * - inc: the "inc" value for the new work item (must be topological).
 795           * - und: the "und" value for the new work item ((inc | und) must be topological).
 796           */
 797          auto add_fn = [&](SetInfo<SetType> inc, SetType und) noexcept {
 798              /** SetInfo object with the set whose feerate will become the new work item's
 799               *  pot_feerate. It starts off equal to inc. */
 800              auto pot = inc;
 801              if (!inc.feerate.IsEmpty()) {
 802                  // Add entries to pot. We iterate over all undecided transactions whose feerate is
 803                  // higher than best. While undecided transactions of lower feerate may improve pot,
 804                  // the resulting pot feerate cannot possibly exceed best's (and this item will be
 805                  // skipped in split_fn anyway).
 806                  for (auto pos : imp & und) {
 807                      // Determine if adding transaction pos to pot (ignoring topology) would improve
 808                      // it. If not, we're done updating pot. This relies on the fact that
 809                      // m_sorted_depgraph, and thus the transactions iterated over, are in decreasing
 810                      // individual feerate order.
 811                      if (!(m_sorted_depgraph.FeeRate(pos) >> pot.feerate)) break;
 812                      pot.Set(m_sorted_depgraph, pos);
 813                  }
 814  
 815                  // The "jump ahead" optimization: whenever pot has a topologically-valid subset,
 816                  // that subset can be added to inc. Any subset of (pot - inc) has the property that
 817                  // its feerate exceeds that of any set compatible with this work item (superset of
 818                  // inc, subset of (inc | und)). Thus, if T is a topological subset of pot, and B is
 819                  // the best topologically-valid set compatible with this work item, and (T - B) is
 820                  // non-empty, then (T | B) is better than B and also topological. This is in
 821                  // contradiction with the assumption that B is best. Thus, (T - B) must be empty,
 822                  // or T must be a subset of B.
 823                  //
 824                  // See https://delvinglimenka.org/t/how-to-linearize-your-cluster/303 section 2.4.
 825                  const auto init_inc = inc.transactions;
 826                  for (auto pos : pot.transactions - inc.transactions) {
 827                      // If the transaction's ancestors are a subset of pot, we can add it together
 828                      // with its ancestors to inc. Just update the transactions here; the feerate
 829                      // update happens below.
 830                      auto anc_todo = m_sorted_depgraph.Ancestors(pos) & m_todo;
 831                      if (anc_todo.IsSubsetOf(pot.transactions)) inc.transactions |= anc_todo;
 832                  }
 833                  // Finally update und and inc's feerate to account for the added transactions.
 834                  und -= inc.transactions;
 835                  inc.feerate += m_sorted_depgraph.FeeRate(inc.transactions - init_inc);
 836  
 837                  // If inc's feerate is better than best's, remember it as our new best.
 838                  if (inc.feerate > best.feerate) {
 839                      best = inc;
 840                      // See if we can remove any entries from imp now.
 841                      while (imp.Any()) {
 842                          ClusterIndex check = imp.Last();
 843                          if (m_sorted_depgraph.FeeRate(check) >> best.feerate) break;
 844                          imp.Reset(check);
 845                      }
 846                  }
 847  
 848                  // If no potential transactions exist beyond the already included ones, no
 849                  // improvement is possible anymore.
 850                  if (pot.feerate.size == inc.feerate.size) return;
 851                  // At this point und must be non-empty. If it were empty then pot would equal inc.
 852                  Assume(und.Any());
 853              } else {
 854                  Assume(inc.transactions.None());
 855                  // If inc is empty, we just make sure there are undecided transactions left to
 856                  // split on.
 857                  if (und.None()) return;
 858              }
 859  
 860              // Actually construct a new work item on the queue. Due to the switch to DFS when queue
 861              // space runs out (see below), we know that no reallocation of the queue should ever
 862              // occur.
 863              Assume(queue.size() < queue.capacity());
 864              queue.emplace_back(/*inc=*/std::move(inc),
 865                                 /*und=*/std::move(und),
 866                                 /*pot_feerate=*/std::move(pot.feerate));
 867          };
 868  
 869          /** Internal process function. It takes an existing work item, and splits it in two: one
 870           *  with a particular transaction (and its ancestors) included, and one with that
 871           *  transaction (and its descendants) excluded. */
 872          auto split_fn = [&](WorkItem&& elem) noexcept {
 873              // Any queue element must have undecided transactions left, otherwise there is nothing
 874              // to explore anymore.
 875              Assume(elem.und.Any());
 876              // The included and undecided set are all subsets of m_todo.
 877              Assume(elem.inc.transactions.IsSubsetOf(m_todo) && elem.und.IsSubsetOf(m_todo));
 878              // Included transactions cannot be undecided.
 879              Assume(!elem.inc.transactions.Overlaps(elem.und));
 880              // If pot is empty, then so is inc.
 881              Assume(elem.inc.feerate.IsEmpty() == elem.pot_feerate.IsEmpty());
 882  
 883              const ClusterIndex first = elem.und.First();
 884              if (!elem.inc.feerate.IsEmpty()) {
 885                  // If no undecided transactions remain with feerate higher than best, this entry
 886                  // cannot be improved beyond best.
 887                  if (!elem.und.Overlaps(imp)) return;
 888                  // We can ignore any queue item whose potential feerate isn't better than the best
 889                  // seen so far.
 890                  if (elem.pot_feerate <= best.feerate) return;
 891              } else {
 892                  // In case inc is empty use a simpler alternative check.
 893                  if (m_sorted_depgraph.FeeRate(first) <= best.feerate) return;
 894              }
 895  
 896              // Decide which transaction to split on. Splitting is how new work items are added, and
 897              // how progress is made. One split transaction is chosen among the queue item's
 898              // undecided ones, and:
 899              // - A work item is (potentially) added with that transaction plus its remaining
 900              //   descendants excluded (removed from the und set).
 901              // - A work item is (potentially) added with that transaction plus its remaining
 902              //   ancestors included (added to the inc set).
 903              //
 904              // To decide what to split on, consider the undecided ancestors of the highest
 905              // individual feerate undecided transaction. Pick the one which reduces the search space
 906              // most. Let I(t) be the size of the undecided set after including t, and E(t) the size
 907              // of the undecided set after excluding t. Then choose the split transaction t such
 908              // that 2^I(t) + 2^E(t) is minimal, tie-breaking by highest individual feerate for t.
 909              ClusterIndex split = 0;
 910              const auto select = elem.und & m_sorted_depgraph.Ancestors(first);
 911              Assume(select.Any());
 912              std::optional<std::pair<ClusterIndex, ClusterIndex>> split_counts;
 913              for (auto t : select) {
 914                  // Call max = max(I(t), E(t)) and min = min(I(t), E(t)). Let counts = {max,min}.
 915                  // Sorting by the tuple counts is equivalent to sorting by 2^I(t) + 2^E(t). This
 916                  // expression is equal to 2^max + 2^min = 2^max * (1 + 1/2^(max - min)). The second
 917                  // factor (1 + 1/2^(max - min)) there is in (1,2]. Thus increasing max will always
 918                  // increase it, even when min decreases. Because of this, we can first sort by max.
 919                  std::pair<ClusterIndex, ClusterIndex> counts{
 920                      (elem.und - m_sorted_depgraph.Ancestors(t)).Count(),
 921                      (elem.und - m_sorted_depgraph.Descendants(t)).Count()};
 922                  if (counts.first < counts.second) std::swap(counts.first, counts.second);
 923                  // Remember the t with the lowest counts.
 924                  if (!split_counts.has_value() || counts < *split_counts) {
 925                      split = t;
 926                      split_counts = counts;
 927                  }
 928              }
 929              // Since there was at least one transaction in select, we must always find one.
 930              Assume(split_counts.has_value());
 931  
 932              // Add a work item corresponding to exclusion of the split transaction.
 933              const auto& desc = m_sorted_depgraph.Descendants(split);
 934              add_fn(/*inc=*/elem.inc,
 935                     /*und=*/elem.und - desc);
 936  
 937              // Add a work item corresponding to inclusion of the split transaction.
 938              const auto anc = m_sorted_depgraph.Ancestors(split) & m_todo;
 939              add_fn(/*inc=*/elem.inc.Add(m_sorted_depgraph, anc),
 940                     /*und=*/elem.und - anc);
 941  
 942              // Account for the performed split.
 943              --iterations_left;
 944          };
 945  
 946          // Work processing loop.
 947          //
 948          // New work items are always added at the back of the queue, but items to process use a
 949          // hybrid approach where they can be taken from the front or the back.
 950          //
 951          // Depth-first search (DFS) corresponds to always taking from the back of the queue. This
 952          // is very memory-efficient (linear in the number of transactions). Breadth-first search
 953          // (BFS) corresponds to always taking from the front, which potentially uses more memory
 954          // (up to exponential in the transaction count), but seems to work better in practice.
 955          //
 956          // The approach here combines the two: use BFS (plus random swapping) until the queue grows
 957          // too large, at which point we temporarily switch to DFS until the size shrinks again.
 958          while (!queue.empty()) {
 959              // Randomly swap the first two items to randomize the search order.
 960              if (queue.size() > 1 && m_rng.randbool()) {
 961                  queue[0].Swap(queue[1]);
 962              }
 963  
 964              // Processing the first queue item, and then using DFS for everything it gives rise to,
 965              // may increase the queue size by the number of undecided elements in there, minus 1
 966              // for the first queue item being removed. Thus, only when that pushes the queue over
 967              // its capacity can we not process from the front (BFS), and should we use DFS.
 968              while (queue.size() - 1 + queue.front().und.Count() > queue.capacity()) {
 969                  if (!iterations_left) break;
 970                  auto elem = queue.back();
 971                  queue.pop_back();
 972                  split_fn(std::move(elem));
 973              }
 974  
 975              // Process one entry from the front of the queue (BFS exploration)
 976              if (!iterations_left) break;
 977              auto elem = queue.front();
 978              queue.pop_front();
 979              split_fn(std::move(elem));
 980          }
 981  
 982          // Return the found best set (converted to the original transaction indices), and the
 983          // number of iterations performed.
 984          best.transactions = SortedToOriginal(best.transactions);
 985          return {std::move(best), max_iterations - iterations_left};
 986      }
 987  
 988      /** Remove a subset of transactions from the cluster being linearized.
 989       *
 990       * Complexity: O(N) where N=done.Count().
 991       */
 992      void MarkDone(const SetType& done) noexcept
 993      {
 994          const auto done_sorted = OriginalToSorted(done);
 995          Assume(done_sorted.Any());
 996          Assume(done_sorted.IsSubsetOf(m_todo));
 997          m_todo -= done_sorted;
 998      }
 999  };
1000  
1001  /** Find or improve a linearization for a cluster.
1002   *
1003   * @param[in] depgraph            Dependency graph of the cluster to be linearized.
1004   * @param[in] max_iterations      Upper bound on the number of optimization steps that will be done.
1005   * @param[in] rng_seed            A random number seed to control search order. This prevents peers
1006   *                                from predicting exactly which clusters would be hard for us to
1007   *                                linearize.
1008   * @param[in] old_linearization   An existing linearization for the cluster (which must be
1009   *                                topologically valid), or empty.
1010   * @return                        A pair of:
1011   *                                - The resulting linearization. It is guaranteed to be at least as
1012   *                                  good (in the feerate diagram sense) as old_linearization.
1013   *                                - A boolean indicating whether the result is guaranteed to be
1014   *                                  optimal.
1015   *
1016   * Complexity: possibly O(N * min(max_iterations + N, sqrt(2^N))) where N=depgraph.TxCount().
1017   */
1018  template<typename SetType>
1019  std::pair<std::vector<ClusterIndex>, bool> Linearize(const DepGraph<SetType>& depgraph, uint64_t max_iterations, uint64_t rng_seed, Span<const ClusterIndex> old_linearization = {}) noexcept
1020  {
1021      Assume(old_linearization.empty() || old_linearization.size() == depgraph.TxCount());
1022      if (depgraph.TxCount() == 0) return {{}, true};
1023  
1024      uint64_t iterations_left = max_iterations;
1025      std::vector<ClusterIndex> linearization;
1026  
1027      AncestorCandidateFinder anc_finder(depgraph);
1028      std::optional<SearchCandidateFinder<SetType>> src_finder;
1029      linearization.reserve(depgraph.TxCount());
1030      bool optimal = true;
1031  
1032      // Treat the initialization of SearchCandidateFinder as taking N^2/64 (rounded up) iterations
1033      // (largely due to the cost of constructing the internal sorted-by-feerate DepGraph inside
1034      // SearchCandidateFinder), a rough approximation based on benchmark. If we don't have that
1035      // many, don't start it.
1036      uint64_t start_iterations = (uint64_t{depgraph.TxCount()} * depgraph.TxCount() + 63) / 64;
1037      if (iterations_left > start_iterations) {
1038          iterations_left -= start_iterations;
1039          src_finder.emplace(depgraph, rng_seed);
1040      }
1041  
1042      /** Chunking of what remains of the old linearization. */
1043      LinearizationChunking old_chunking(depgraph, old_linearization);
1044  
1045      while (true) {
1046          // Find the highest-feerate prefix of the remainder of old_linearization.
1047          SetInfo<SetType> best_prefix;
1048          if (old_chunking.NumChunksLeft()) best_prefix = old_chunking.GetChunk(0);
1049  
1050          // Then initialize best to be either the best remaining ancestor set, or the first chunk.
1051          auto best = anc_finder.FindCandidateSet();
1052          if (!best_prefix.feerate.IsEmpty() && best_prefix.feerate >= best.feerate) best = best_prefix;
1053  
1054          uint64_t iterations_done_now = 0;
1055          uint64_t max_iterations_now = 0;
1056          if (src_finder) {
1057              // Treat the invocation of SearchCandidateFinder::FindCandidateSet() as costing N/4
1058              // up-front (rounded up) iterations (largely due to the cost of connected-component
1059              // splitting), a rough approximation based on benchmarks.
1060              uint64_t base_iterations = (anc_finder.NumRemaining() + 3) / 4;
1061              if (iterations_left > base_iterations) {
1062                  // Invoke bounded search to update best, with up to half of our remaining
1063                  // iterations as limit.
1064                  iterations_left -= base_iterations;
1065                  max_iterations_now = (iterations_left + 1) / 2;
1066                  std::tie(best, iterations_done_now) = src_finder->FindCandidateSet(max_iterations_now, best);
1067                  iterations_left -= iterations_done_now;
1068              }
1069          }
1070  
1071          if (iterations_done_now == max_iterations_now) {
1072              optimal = false;
1073              // If the search result is not (guaranteed to be) optimal, run intersections to make
1074              // sure we don't pick something that makes us unable to reach further diagram points
1075              // of the old linearization.
1076              if (old_chunking.NumChunksLeft() > 0) {
1077                  best = old_chunking.IntersectPrefixes(best);
1078              }
1079          }
1080  
1081          // Add to output in topological order.
1082          depgraph.AppendTopo(linearization, best.transactions);
1083  
1084          // Update state to reflect best is no longer to be linearized.
1085          anc_finder.MarkDone(best.transactions);
1086          if (anc_finder.AllDone()) break;
1087          if (src_finder) src_finder->MarkDone(best.transactions);
1088          if (old_chunking.NumChunksLeft() > 0) {
1089              old_chunking.MarkDone(best.transactions);
1090          }
1091      }
1092  
1093      return {std::move(linearization), optimal};
1094  }
1095  
1096  /** Improve a given linearization.
1097   *
1098   * @param[in]     depgraph       Dependency graph of the cluster being linearized.
1099   * @param[in,out] linearization  On input, an existing linearization for depgraph. On output, a
1100   *                               potentially better linearization for the same graph.
1101   *
1102   * Postlinearization guarantees:
1103   * - The resulting chunks are connected.
1104   * - If the input has a tree shape (either all transactions have at most one child, or all
1105   *   transactions have at most one parent), the result is optimal.
1106   * - Given a linearization L1 and a leaf transaction T in it. Let L2 be L1 with T moved to the end,
1107   *   optionally with its fee increased. Let L3 be the postlinearization of L2. L3 will be at least
1108   *   as good as L1. This means that replacing transactions with same-size higher-fee transactions
1109   *   will not worsen linearizations through a "drop conflicts, append new transactions,
1110   *   postlinearize" process.
1111   */
1112  template<typename SetType>
1113  void PostLinearize(const DepGraph<SetType>& depgraph, Span<ClusterIndex> linearization)
1114  {
1115      // This algorithm performs a number of passes (currently 2); the even ones operate from back to
1116      // front, the odd ones from front to back. Each results in an equal-or-better linearization
1117      // than the one started from.
1118      // - One pass in either direction guarantees that the resulting chunks are connected.
1119      // - Each direction corresponds to one shape of tree being linearized optimally (forward passes
1120      //   guarantee this for graphs where each transaction has at most one child; backward passes
1121      //   guarantee this for graphs where each transaction has at most one parent).
1122      // - Starting with a backward pass guarantees the moved-tree property.
1123      //
1124      // During an odd (forward) pass, the high-level operation is:
1125      // - Start with an empty list of groups L=[].
1126      // - For every transaction i in the old linearization, from front to back:
1127      //   - Append a new group C=[i], containing just i, to the back of L.
1128      //   - While L has at least one group before C, and the group immediately before C has feerate
1129      //     lower than C:
1130      //     - If C depends on P:
1131      //       - Merge P into C, making C the concatenation of P+C, continuing with the combined C.
1132      //     - Otherwise:
1133      //       - Swap P with C, continuing with the now-moved C.
1134      // - The output linearization is the concatenation of the groups in L.
1135      //
1136      // During even (backward) passes, i iterates from the back to the front of the existing
1137      // linearization, and new groups are prepended instead of appended to the list L. To enable
1138      // more code reuse, both passes append groups, but during even passes the meanings of
1139      // parent/child, and of high/low feerate are reversed, and the final concatenation is reversed
1140      // on output.
1141      //
1142      // In the implementation below, the groups are represented by singly-linked lists (pointing
1143      // from the back to the front), which are themselves organized in a singly-linked circular
1144      // list (each group pointing to its predecessor, with a special sentinel group at the front
1145      // that points back to the last group).
1146      //
1147      // Information about transaction t is stored in entries[t + 1], while the sentinel is in
1148      // entries[0].
1149  
1150      /** Index of the sentinel in the entries array below. */
1151      static constexpr ClusterIndex SENTINEL{0};
1152      /** Indicator that a group has no previous transaction. */
1153      static constexpr ClusterIndex NO_PREV_TX{0};
1154  
1155  
1156      /** Data structure per transaction entry. */
1157      struct TxEntry
1158      {
1159          /** The index of the previous transaction in this group; NO_PREV_TX if this is the first
1160           *  entry of a group. */
1161          ClusterIndex prev_tx;
1162  
1163          // The fields below are only used for transactions that are the last one in a group
1164          // (referred to as tail transactions below).
1165  
1166          /** Index of the first transaction in this group, possibly itself. */
1167          ClusterIndex first_tx;
1168          /** Index of the last transaction in the previous group. The first group (the sentinel)
1169           *  points back to the last group here, making it a singly-linked circular list. */
1170          ClusterIndex prev_group;
1171          /** All transactions in the group. Empty for the sentinel. */
1172          SetType group;
1173          /** All dependencies of the group (descendants in even passes; ancestors in odd ones). */
1174          SetType deps;
1175          /** The combined fee/size of transactions in the group. Fee is negated in even passes. */
1176          FeeFrac feerate;
1177      };
1178  
1179      // As an example, consider the state corresponding to the linearization [1,0,3,2], with
1180      // groups [1,0,3] and [2], in an odd pass. The linked lists would be:
1181      //
1182      //                                        +-----+
1183      //                                 0<-P-- | 0 S | ---\     Legend:
1184      //                                        +-----+    |
1185      //                                           ^       |     - digit in box: entries index
1186      //             /--------------F---------+    G       |       (note: one more than tx value)
1187      //             v                         \   |       |     - S: sentinel group
1188      //          +-----+        +-----+        +-----+    |          (empty feerate)
1189      //   0<-P-- | 2   | <--P-- | 1   | <--P-- | 4 T |    |     - T: tail transaction, contains
1190      //          +-----+        +-----+        +-----+    |          fields beyond prev_tv.
1191      //                                           ^       |     - P: prev_tx reference
1192      //                                           G       G     - F: first_tx reference
1193      //                                           |       |     - G: prev_group reference
1194      //                                        +-----+    |
1195      //                                 0<-P-- | 3 T | <--/
1196      //                                        +-----+
1197      //                                         ^   |
1198      //                                         \-F-/
1199      //
1200      // During an even pass, the diagram above would correspond to linearization [2,3,0,1], with
1201      // groups [2] and [3,0,1].
1202  
1203      std::vector<TxEntry> entries(depgraph.PositionRange() + 1);
1204  
1205      // Perform two passes over the linearization.
1206      for (int pass = 0; pass < 2; ++pass) {
1207          int rev = !(pass & 1);
1208          // Construct a sentinel group, identifying the start of the list.
1209          entries[SENTINEL].prev_group = SENTINEL;
1210          Assume(entries[SENTINEL].feerate.IsEmpty());
1211  
1212          // Iterate over all elements in the existing linearization.
1213          for (ClusterIndex i = 0; i < linearization.size(); ++i) {
1214              // Even passes are from back to front; odd passes from front to back.
1215              ClusterIndex idx = linearization[rev ? linearization.size() - 1 - i : i];
1216              // Construct a new group containing just idx. In even passes, the meaning of
1217              // parent/child and high/low feerate are swapped.
1218              ClusterIndex cur_group = idx + 1;
1219              entries[cur_group].group = SetType::Singleton(idx);
1220              entries[cur_group].deps = rev ? depgraph.Descendants(idx): depgraph.Ancestors(idx);
1221              entries[cur_group].feerate = depgraph.FeeRate(idx);
1222              if (rev) entries[cur_group].feerate.fee = -entries[cur_group].feerate.fee;
1223              entries[cur_group].prev_tx = NO_PREV_TX; // No previous transaction in group.
1224              entries[cur_group].first_tx = cur_group; // Transaction itself is first of group.
1225              // Insert the new group at the back of the groups linked list.
1226              entries[cur_group].prev_group = entries[SENTINEL].prev_group;
1227              entries[SENTINEL].prev_group = cur_group;
1228  
1229              // Start merge/swap cycle.
1230              ClusterIndex next_group = SENTINEL; // We inserted at the end, so next group is sentinel.
1231              ClusterIndex prev_group = entries[cur_group].prev_group;
1232              // Continue as long as the current group has higher feerate than the previous one.
1233              while (entries[cur_group].feerate >> entries[prev_group].feerate) {
1234                  // prev_group/cur_group/next_group refer to (the last transactions of) 3
1235                  // consecutive entries in groups list.
1236                  Assume(cur_group == entries[next_group].prev_group);
1237                  Assume(prev_group == entries[cur_group].prev_group);
1238                  // The sentinel has empty feerate, which is neither higher or lower than other
1239                  // feerates. Thus, the while loop we are in here guarantees that cur_group and
1240                  // prev_group are not the sentinel.
1241                  Assume(cur_group != SENTINEL);
1242                  Assume(prev_group != SENTINEL);
1243                  if (entries[cur_group].deps.Overlaps(entries[prev_group].group)) {
1244                      // There is a dependency between cur_group and prev_group; merge prev_group
1245                      // into cur_group. The group/deps/feerate fields of prev_group remain unchanged
1246                      // but become unused.
1247                      entries[cur_group].group |= entries[prev_group].group;
1248                      entries[cur_group].deps |= entries[prev_group].deps;
1249                      entries[cur_group].feerate += entries[prev_group].feerate;
1250                      // Make the first of the current group point to the tail of the previous group.
1251                      entries[entries[cur_group].first_tx].prev_tx = prev_group;
1252                      // The first of the previous group becomes the first of the newly-merged group.
1253                      entries[cur_group].first_tx = entries[prev_group].first_tx;
1254                      // The previous group becomes whatever group was before the former one.
1255                      prev_group = entries[prev_group].prev_group;
1256                      entries[cur_group].prev_group = prev_group;
1257                  } else {
1258                      // There is no dependency between cur_group and prev_group; swap them.
1259                      ClusterIndex preprev_group = entries[prev_group].prev_group;
1260                      // If PP, P, C, N were the old preprev, prev, cur, next groups, then the new
1261                      // layout becomes [PP, C, P, N]. Update prev_groups to reflect that order.
1262                      entries[next_group].prev_group = prev_group;
1263                      entries[prev_group].prev_group = cur_group;
1264                      entries[cur_group].prev_group = preprev_group;
1265                      // The current group remains the same, but the groups before/after it have
1266                      // changed.
1267                      next_group = prev_group;
1268                      prev_group = preprev_group;
1269                  }
1270              }
1271          }
1272  
1273          // Convert the entries back to linearization (overwriting the existing one).
1274          ClusterIndex cur_group = entries[0].prev_group;
1275          ClusterIndex done = 0;
1276          while (cur_group != SENTINEL) {
1277              ClusterIndex cur_tx = cur_group;
1278              // Traverse the transactions of cur_group (from back to front), and write them in the
1279              // same order during odd passes, and reversed (front to back) in even passes.
1280              if (rev) {
1281                  do {
1282                      *(linearization.begin() + (done++)) = cur_tx - 1;
1283                      cur_tx = entries[cur_tx].prev_tx;
1284                  } while (cur_tx != NO_PREV_TX);
1285              } else {
1286                  do {
1287                      *(linearization.end() - (++done)) = cur_tx - 1;
1288                      cur_tx = entries[cur_tx].prev_tx;
1289                  } while (cur_tx != NO_PREV_TX);
1290              }
1291              cur_group = entries[cur_group].prev_group;
1292          }
1293          Assume(done == linearization.size());
1294      }
1295  }
1296  
1297  /** Merge two linearizations for the same cluster into one that is as good as both.
1298   *
1299   * Complexity: O(N^2) where N=depgraph.TxCount(); O(N) if both inputs are identical.
1300   */
1301  template<typename SetType>
1302  std::vector<ClusterIndex> MergeLinearizations(const DepGraph<SetType>& depgraph, Span<const ClusterIndex> lin1, Span<const ClusterIndex> lin2)
1303  {
1304      Assume(lin1.size() == depgraph.TxCount());
1305      Assume(lin2.size() == depgraph.TxCount());
1306  
1307      /** Chunkings of what remains of both input linearizations. */
1308      LinearizationChunking chunking1(depgraph, lin1), chunking2(depgraph, lin2);
1309      /** Output linearization. */
1310      std::vector<ClusterIndex> ret;
1311      if (depgraph.TxCount() == 0) return ret;
1312      ret.reserve(depgraph.TxCount());
1313  
1314      while (true) {
1315          // As long as we are not done, both linearizations must have chunks left.
1316          Assume(chunking1.NumChunksLeft() > 0);
1317          Assume(chunking2.NumChunksLeft() > 0);
1318          // Find the set to output by taking the best remaining chunk, and then intersecting it with
1319          // prefixes of remaining chunks of the other linearization.
1320          SetInfo<SetType> best;
1321          const auto& lin1_firstchunk = chunking1.GetChunk(0);
1322          const auto& lin2_firstchunk = chunking2.GetChunk(0);
1323          if (lin2_firstchunk.feerate >> lin1_firstchunk.feerate) {
1324              best = chunking1.IntersectPrefixes(lin2_firstchunk);
1325          } else {
1326              best = chunking2.IntersectPrefixes(lin1_firstchunk);
1327          }
1328          // Append the result to the output and mark it as done.
1329          depgraph.AppendTopo(ret, best.transactions);
1330          chunking1.MarkDone(best.transactions);
1331          if (chunking1.NumChunksLeft() == 0) break;
1332          chunking2.MarkDone(best.transactions);
1333      }
1334  
1335      Assume(ret.size() == depgraph.TxCount());
1336      return ret;
1337  }
1338  
1339  } // namespace cluster_linearize
1340  
1341  #endif // LIMENKA_CLUSTER_LINEARIZE_H
1342