RegAllocPBQP.cpp 33.1 KB
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942
//===- RegAllocPBQP.cpp ---- PBQP Register Allocator ----------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file contains a Partitioned Boolean Quadratic Programming (PBQP) based
// register allocator for LLVM. This allocator works by constructing a PBQP
// problem representing the register allocation problem under consideration,
// solving this using a PBQP solver, and mapping the solution back to a
// register assignment. If any variables are selected for spilling then spill
// code is inserted and the process repeated.
//
// The PBQP solver (pbqp.c) provided for this allocator uses a heuristic tuned
// for register allocation. For more information on PBQP for register
// allocation, see the following papers:
//
//   (1) Hames, L. and Scholz, B. 2006. Nearly optimal register allocation with
//   PBQP. In Proceedings of the 7th Joint Modular Languages Conference
//   (JMLC'06). LNCS, vol. 4228. Springer, New York, NY, USA. 346-361.
//
//   (2) Scholz, B., Eckstein, E. 2002. Register allocation for irregular
//   architectures. In Proceedings of the Joint Conference on Languages,
//   Compilers and Tools for Embedded Systems (LCTES'02), ACM Press, New York,
//   NY, USA, 139-148.
//
//===----------------------------------------------------------------------===//

#include "llvm/CodeGen/RegAllocPBQP.h"
#include "RegisterCoalescer.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/CalcSpillWeights.h"
#include "llvm/CodeGen/LiveInterval.h"
#include "llvm/CodeGen/LiveIntervals.h"
#include "llvm/CodeGen/LiveRangeEdit.h"
#include "llvm/CodeGen/LiveStacks.h"
#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PBQP/Graph.h"
#include "llvm/CodeGen/PBQP/Math.h"
#include "llvm/CodeGen/PBQP/Solution.h"
#include "llvm/CodeGen/PBQPRAConstraint.h"
#include "llvm/CodeGen/RegAllocRegistry.h"
#include "llvm/CodeGen/SlotIndexes.h"
#include "llvm/CodeGen/Spiller.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/CodeGen/VirtRegMap.h"
#include "llvm/Config/llvm-config.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/Module.h"
#include "llvm/MC/MCRegisterInfo.h"
#include "llvm/Pass.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/Printable.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <cassert>
#include <cstddef>
#include <limits>
#include <map>
#include <memory>
#include <queue>
#include <set>
#include <sstream>
#include <string>
#include <system_error>
#include <tuple>
#include <utility>
#include <vector>

using namespace llvm;

#define DEBUG_TYPE "regalloc"

static RegisterRegAlloc
RegisterPBQPRepAlloc("pbqp", "PBQP register allocator",
                       createDefaultPBQPRegisterAllocator);

static cl::opt<bool>
PBQPCoalescing("pbqp-coalescing",
                cl::desc("Attempt coalescing during PBQP register allocation."),
                cl::init(false), cl::Hidden);

#ifndef NDEBUG
static cl::opt<bool>
PBQPDumpGraphs("pbqp-dump-graphs",
               cl::desc("Dump graphs for each function/round in the compilation unit."),
               cl::init(false), cl::Hidden);
#endif

namespace {

///
/// PBQP based allocators solve the register allocation problem by mapping
/// register allocation problems to Partitioned Boolean Quadratic
/// Programming problems.
class RegAllocPBQP : public MachineFunctionPass {
public:
  static char ID;

  /// Construct a PBQP register allocator.
  RegAllocPBQP(char *cPassID = nullptr)
      : MachineFunctionPass(ID), customPassID(cPassID) {
    initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
    initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
    initializeLiveStacksPass(*PassRegistry::getPassRegistry());
    initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
  }

  /// Return the pass name.
  StringRef getPassName() const override { return "PBQP Register Allocator"; }

  /// PBQP analysis usage.
  void getAnalysisUsage(AnalysisUsage &au) const override;

  /// Perform register allocation
  bool runOnMachineFunction(MachineFunction &MF) override;

  MachineFunctionProperties getRequiredProperties() const override {
    return MachineFunctionProperties().set(
        MachineFunctionProperties::Property::NoPHIs);
  }

private:
  using LI2NodeMap = std::map<const LiveInterval *, unsigned>;
  using Node2LIMap = std::vector<const LiveInterval *>;
  using AllowedSet = std::vector<unsigned>;
  using AllowedSetMap = std::vector<AllowedSet>;
  using RegPair = std::pair<unsigned, unsigned>;
  using CoalesceMap = std::map<RegPair, PBQP::PBQPNum>;
  using RegSet = std::set<unsigned>;

  char *customPassID;

  RegSet VRegsToAlloc, EmptyIntervalVRegs;

  /// Inst which is a def of an original reg and whose defs are already all
  /// dead after remat is saved in DeadRemats. The deletion of such inst is
  /// postponed till all the allocations are done, so its remat expr is
  /// always available for the remat of all the siblings of the original reg.
  SmallPtrSet<MachineInstr *, 32> DeadRemats;

  /// Finds the initial set of vreg intervals to allocate.
  void findVRegIntervalsToAlloc(const MachineFunction &MF, LiveIntervals &LIS);

  /// Constructs an initial graph.
  void initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM, Spiller &VRegSpiller);

  /// Spill the given VReg.
  void spillVReg(Register VReg, SmallVectorImpl<Register> &NewIntervals,
                 MachineFunction &MF, LiveIntervals &LIS, VirtRegMap &VRM,
                 Spiller &VRegSpiller);

  /// Given a solved PBQP problem maps this solution back to a register
  /// assignment.
  bool mapPBQPToRegAlloc(const PBQPRAGraph &G,
                         const PBQP::Solution &Solution,
                         VirtRegMap &VRM,
                         Spiller &VRegSpiller);

  /// Postprocessing before final spilling. Sets basic block "live in"
  /// variables.
  void finalizeAlloc(MachineFunction &MF, LiveIntervals &LIS,
                     VirtRegMap &VRM) const;

  void postOptimization(Spiller &VRegSpiller, LiveIntervals &LIS);
};

char RegAllocPBQP::ID = 0;

/// Set spill costs for each node in the PBQP reg-alloc graph.
class SpillCosts : public PBQPRAConstraint {
public:
  void apply(PBQPRAGraph &G) override {
    LiveIntervals &LIS = G.getMetadata().LIS;

    // A minimum spill costs, so that register constraints can can be set
    // without normalization in the [0.0:MinSpillCost( interval.
    const PBQP::PBQPNum MinSpillCost = 10.0;

    for (auto NId : G.nodeIds()) {
      PBQP::PBQPNum SpillCost =
          LIS.getInterval(G.getNodeMetadata(NId).getVReg()).weight();
      if (SpillCost == 0.0)
        SpillCost = std::numeric_limits<PBQP::PBQPNum>::min();
      else
        SpillCost += MinSpillCost;
      PBQPRAGraph::RawVector NodeCosts(G.getNodeCosts(NId));
      NodeCosts[PBQP::RegAlloc::getSpillOptionIdx()] = SpillCost;
      G.setNodeCosts(NId, std::move(NodeCosts));
    }
  }
};

/// Add interference edges between overlapping vregs.
class Interference : public PBQPRAConstraint {
private:
  using AllowedRegVecPtr = const PBQP::RegAlloc::AllowedRegVector *;
  using IKey = std::pair<AllowedRegVecPtr, AllowedRegVecPtr>;
  using IMatrixCache = DenseMap<IKey, PBQPRAGraph::MatrixPtr>;
  using DisjointAllowedRegsCache = DenseSet<IKey>;
  using IEdgeKey = std::pair<PBQP::GraphBase::NodeId, PBQP::GraphBase::NodeId>;
  using IEdgeCache = DenseSet<IEdgeKey>;

  bool haveDisjointAllowedRegs(const PBQPRAGraph &G, PBQPRAGraph::NodeId NId,
                               PBQPRAGraph::NodeId MId,
                               const DisjointAllowedRegsCache &D) const {
    const auto *NRegs = &G.getNodeMetadata(NId).getAllowedRegs();
    const auto *MRegs = &G.getNodeMetadata(MId).getAllowedRegs();

    if (NRegs == MRegs)
      return false;

    if (NRegs < MRegs)
      return D.count(IKey(NRegs, MRegs)) > 0;

    return D.count(IKey(MRegs, NRegs)) > 0;
  }

  void setDisjointAllowedRegs(const PBQPRAGraph &G, PBQPRAGraph::NodeId NId,
                              PBQPRAGraph::NodeId MId,
                              DisjointAllowedRegsCache &D) {
    const auto *NRegs = &G.getNodeMetadata(NId).getAllowedRegs();
    const auto *MRegs = &G.getNodeMetadata(MId).getAllowedRegs();

    assert(NRegs != MRegs && "AllowedRegs can not be disjoint with itself");

    if (NRegs < MRegs)
      D.insert(IKey(NRegs, MRegs));
    else
      D.insert(IKey(MRegs, NRegs));
  }

  // Holds (Interval, CurrentSegmentID, and NodeId). The first two are required
  // for the fast interference graph construction algorithm. The last is there
  // to save us from looking up node ids via the VRegToNode map in the graph
  // metadata.
  using IntervalInfo =
      std::tuple<LiveInterval*, size_t, PBQP::GraphBase::NodeId>;

  static SlotIndex getStartPoint(const IntervalInfo &I) {
    return std::get<0>(I)->segments[std::get<1>(I)].start;
  }

  static SlotIndex getEndPoint(const IntervalInfo &I) {
    return std::get<0>(I)->segments[std::get<1>(I)].end;
  }

  static PBQP::GraphBase::NodeId getNodeId(const IntervalInfo &I) {
    return std::get<2>(I);
  }

  static bool lowestStartPoint(const IntervalInfo &I1,
                               const IntervalInfo &I2) {
    // Condition reversed because priority queue has the *highest* element at
    // the front, rather than the lowest.
    return getStartPoint(I1) > getStartPoint(I2);
  }

  static bool lowestEndPoint(const IntervalInfo &I1,
                             const IntervalInfo &I2) {
    SlotIndex E1 = getEndPoint(I1);
    SlotIndex E2 = getEndPoint(I2);

    if (E1 < E2)
      return true;

    if (E1 > E2)
      return false;

    // If two intervals end at the same point, we need a way to break the tie or
    // the set will assume they're actually equal and refuse to insert a
    // "duplicate". Just compare the vregs - fast and guaranteed unique.
    return std::get<0>(I1)->reg() < std::get<0>(I2)->reg();
  }

  static bool isAtLastSegment(const IntervalInfo &I) {
    return std::get<1>(I) == std::get<0>(I)->size() - 1;
  }

  static IntervalInfo nextSegment(const IntervalInfo &I) {
    return std::make_tuple(std::get<0>(I), std::get<1>(I) + 1, std::get<2>(I));
  }

public:
  void apply(PBQPRAGraph &G) override {
    // The following is loosely based on the linear scan algorithm introduced in
    // "Linear Scan Register Allocation" by Poletto and Sarkar. This version
    // isn't linear, because the size of the active set isn't bound by the
    // number of registers, but rather the size of the largest clique in the
    // graph. Still, we expect this to be better than N^2.
    LiveIntervals &LIS = G.getMetadata().LIS;

    // Interferenc matrices are incredibly regular - they're only a function of
    // the allowed sets, so we cache them to avoid the overhead of constructing
    // and uniquing them.
    IMatrixCache C;

    // Finding an edge is expensive in the worst case (O(max_clique(G))). So
    // cache locally edges we have already seen.
    IEdgeCache EC;

    // Cache known disjoint allowed registers pairs
    DisjointAllowedRegsCache D;

    using IntervalSet = std::set<IntervalInfo, decltype(&lowestEndPoint)>;
    using IntervalQueue =
        std::priority_queue<IntervalInfo, std::vector<IntervalInfo>,
                            decltype(&lowestStartPoint)>;
    IntervalSet Active(lowestEndPoint);
    IntervalQueue Inactive(lowestStartPoint);

    // Start by building the inactive set.
    for (auto NId : G.nodeIds()) {
      unsigned VReg = G.getNodeMetadata(NId).getVReg();
      LiveInterval &LI = LIS.getInterval(VReg);
      assert(!LI.empty() && "PBQP graph contains node for empty interval");
      Inactive.push(std::make_tuple(&LI, 0, NId));
    }

    while (!Inactive.empty()) {
      // Tentatively grab the "next" interval - this choice may be overriden
      // below.
      IntervalInfo Cur = Inactive.top();

      // Retire any active intervals that end before Cur starts.
      IntervalSet::iterator RetireItr = Active.begin();
      while (RetireItr != Active.end() &&
             (getEndPoint(*RetireItr) <= getStartPoint(Cur))) {
        // If this interval has subsequent segments, add the next one to the
        // inactive list.
        if (!isAtLastSegment(*RetireItr))
          Inactive.push(nextSegment(*RetireItr));

        ++RetireItr;
      }
      Active.erase(Active.begin(), RetireItr);

      // One of the newly retired segments may actually start before the
      // Cur segment, so re-grab the front of the inactive list.
      Cur = Inactive.top();
      Inactive.pop();

      // At this point we know that Cur overlaps all active intervals. Add the
      // interference edges.
      PBQP::GraphBase::NodeId NId = getNodeId(Cur);
      for (const auto &A : Active) {
        PBQP::GraphBase::NodeId MId = getNodeId(A);

        // Do not add an edge when the nodes' allowed registers do not
        // intersect: there is obviously no interference.
        if (haveDisjointAllowedRegs(G, NId, MId, D))
          continue;

        // Check that we haven't already added this edge
        IEdgeKey EK(std::min(NId, MId), std::max(NId, MId));
        if (EC.count(EK))
          continue;

        // This is a new edge - add it to the graph.
        if (!createInterferenceEdge(G, NId, MId, C))
          setDisjointAllowedRegs(G, NId, MId, D);
        else
          EC.insert(EK);
      }

      // Finally, add Cur to the Active set.
      Active.insert(Cur);
    }
  }

private:
  // Create an Interference edge and add it to the graph, unless it is
  // a null matrix, meaning the nodes' allowed registers do not have any
  // interference. This case occurs frequently between integer and floating
  // point registers for example.
  // return true iff both nodes interferes.
  bool createInterferenceEdge(PBQPRAGraph &G,
                              PBQPRAGraph::NodeId NId, PBQPRAGraph::NodeId MId,
                              IMatrixCache &C) {
    const TargetRegisterInfo &TRI =
        *G.getMetadata().MF.getSubtarget().getRegisterInfo();
    const auto &NRegs = G.getNodeMetadata(NId).getAllowedRegs();
    const auto &MRegs = G.getNodeMetadata(MId).getAllowedRegs();

    // Try looking the edge costs up in the IMatrixCache first.
    IKey K(&NRegs, &MRegs);
    IMatrixCache::iterator I = C.find(K);
    if (I != C.end()) {
      G.addEdgeBypassingCostAllocator(NId, MId, I->second);
      return true;
    }

    PBQPRAGraph::RawMatrix M(NRegs.size() + 1, MRegs.size() + 1, 0);
    bool NodesInterfere = false;
    for (unsigned I = 0; I != NRegs.size(); ++I) {
      unsigned PRegN = NRegs[I];
      for (unsigned J = 0; J != MRegs.size(); ++J) {
        unsigned PRegM = MRegs[J];
        if (TRI.regsOverlap(PRegN, PRegM)) {
          M[I + 1][J + 1] = std::numeric_limits<PBQP::PBQPNum>::infinity();
          NodesInterfere = true;
        }
      }
    }

    if (!NodesInterfere)
      return false;

    PBQPRAGraph::EdgeId EId = G.addEdge(NId, MId, std::move(M));
    C[K] = G.getEdgeCostsPtr(EId);

    return true;
  }
};

class Coalescing : public PBQPRAConstraint {
public:
  void apply(PBQPRAGraph &G) override {
    MachineFunction &MF = G.getMetadata().MF;
    MachineBlockFrequencyInfo &MBFI = G.getMetadata().MBFI;
    CoalescerPair CP(*MF.getSubtarget().getRegisterInfo());

    // Scan the machine function and add a coalescing cost whenever CoalescerPair
    // gives the Ok.
    for (const auto &MBB : MF) {
      for (const auto &MI : MBB) {
        // Skip not-coalescable or already coalesced copies.
        if (!CP.setRegisters(&MI) || CP.getSrcReg() == CP.getDstReg())
          continue;

        unsigned DstReg = CP.getDstReg();
        unsigned SrcReg = CP.getSrcReg();

        PBQP::PBQPNum CBenefit = MBFI.getBlockFreqRelativeToEntryBlock(&MBB);

        if (CP.isPhys()) {
          if (!MF.getRegInfo().isAllocatable(DstReg))
            continue;

          PBQPRAGraph::NodeId NId = G.getMetadata().getNodeIdForVReg(SrcReg);

          const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed =
            G.getNodeMetadata(NId).getAllowedRegs();

          unsigned PRegOpt = 0;
          while (PRegOpt < Allowed.size() && Allowed[PRegOpt] != DstReg)
            ++PRegOpt;

          if (PRegOpt < Allowed.size()) {
            PBQPRAGraph::RawVector NewCosts(G.getNodeCosts(NId));
            NewCosts[PRegOpt + 1] -= CBenefit;
            G.setNodeCosts(NId, std::move(NewCosts));
          }
        } else {
          PBQPRAGraph::NodeId N1Id = G.getMetadata().getNodeIdForVReg(DstReg);
          PBQPRAGraph::NodeId N2Id = G.getMetadata().getNodeIdForVReg(SrcReg);
          const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed1 =
            &G.getNodeMetadata(N1Id).getAllowedRegs();
          const PBQPRAGraph::NodeMetadata::AllowedRegVector *Allowed2 =
            &G.getNodeMetadata(N2Id).getAllowedRegs();

          PBQPRAGraph::EdgeId EId = G.findEdge(N1Id, N2Id);
          if (EId == G.invalidEdgeId()) {
            PBQPRAGraph::RawMatrix Costs(Allowed1->size() + 1,
                                         Allowed2->size() + 1, 0);
            addVirtRegCoalesce(Costs, *Allowed1, *Allowed2, CBenefit);
            G.addEdge(N1Id, N2Id, std::move(Costs));
          } else {
            if (G.getEdgeNode1Id(EId) == N2Id) {
              std::swap(N1Id, N2Id);
              std::swap(Allowed1, Allowed2);
            }
            PBQPRAGraph::RawMatrix Costs(G.getEdgeCosts(EId));
            addVirtRegCoalesce(Costs, *Allowed1, *Allowed2, CBenefit);
            G.updateEdgeCosts(EId, std::move(Costs));
          }
        }
      }
    }
  }

private:
  void addVirtRegCoalesce(
                    PBQPRAGraph::RawMatrix &CostMat,
                    const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed1,
                    const PBQPRAGraph::NodeMetadata::AllowedRegVector &Allowed2,
                    PBQP::PBQPNum Benefit) {
    assert(CostMat.getRows() == Allowed1.size() + 1 && "Size mismatch.");
    assert(CostMat.getCols() == Allowed2.size() + 1 && "Size mismatch.");
    for (unsigned I = 0; I != Allowed1.size(); ++I) {
      unsigned PReg1 = Allowed1[I];
      for (unsigned J = 0; J != Allowed2.size(); ++J) {
        unsigned PReg2 = Allowed2[J];
        if (PReg1 == PReg2)
          CostMat[I + 1][J + 1] -= Benefit;
      }
    }
  }
};

} // end anonymous namespace

// Out-of-line destructor/anchor for PBQPRAConstraint.
PBQPRAConstraint::~PBQPRAConstraint() = default;

void PBQPRAConstraint::anchor() {}

void PBQPRAConstraintList::anchor() {}

void RegAllocPBQP::getAnalysisUsage(AnalysisUsage &au) const {
  au.setPreservesCFG();
  au.addRequired<AAResultsWrapperPass>();
  au.addPreserved<AAResultsWrapperPass>();
  au.addRequired<SlotIndexes>();
  au.addPreserved<SlotIndexes>();
  au.addRequired<LiveIntervals>();
  au.addPreserved<LiveIntervals>();
  //au.addRequiredID(SplitCriticalEdgesID);
  if (customPassID)
    au.addRequiredID(*customPassID);
  au.addRequired<LiveStacks>();
  au.addPreserved<LiveStacks>();
  au.addRequired<MachineBlockFrequencyInfo>();
  au.addPreserved<MachineBlockFrequencyInfo>();
  au.addRequired<MachineLoopInfo>();
  au.addPreserved<MachineLoopInfo>();
  au.addRequired<MachineDominatorTree>();
  au.addPreserved<MachineDominatorTree>();
  au.addRequired<VirtRegMap>();
  au.addPreserved<VirtRegMap>();
  MachineFunctionPass::getAnalysisUsage(au);
}

void RegAllocPBQP::findVRegIntervalsToAlloc(const MachineFunction &MF,
                                            LiveIntervals &LIS) {
  const MachineRegisterInfo &MRI = MF.getRegInfo();

  // Iterate over all live ranges.
  for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) {
    unsigned Reg = Register::index2VirtReg(I);
    if (MRI.reg_nodbg_empty(Reg))
      continue;
    VRegsToAlloc.insert(Reg);
  }
}

static bool isACalleeSavedRegister(unsigned reg, const TargetRegisterInfo &TRI,
                                   const MachineFunction &MF) {
  const MCPhysReg *CSR = MF.getRegInfo().getCalleeSavedRegs();
  for (unsigned i = 0; CSR[i] != 0; ++i)
    if (TRI.regsOverlap(reg, CSR[i]))
      return true;
  return false;
}

void RegAllocPBQP::initializeGraph(PBQPRAGraph &G, VirtRegMap &VRM,
                                   Spiller &VRegSpiller) {
  MachineFunction &MF = G.getMetadata().MF;

  LiveIntervals &LIS = G.getMetadata().LIS;
  const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
  const TargetRegisterInfo &TRI =
      *G.getMetadata().MF.getSubtarget().getRegisterInfo();

  std::vector<unsigned> Worklist(VRegsToAlloc.begin(), VRegsToAlloc.end());

  std::map<unsigned, std::vector<unsigned>> VRegAllowedMap;

  while (!Worklist.empty()) {
    unsigned VReg = Worklist.back();
    Worklist.pop_back();

    LiveInterval &VRegLI = LIS.getInterval(VReg);

    // If this is an empty interval move it to the EmptyIntervalVRegs set then
    // continue.
    if (VRegLI.empty()) {
      EmptyIntervalVRegs.insert(VRegLI.reg());
      VRegsToAlloc.erase(VRegLI.reg());
      continue;
    }

    const TargetRegisterClass *TRC = MRI.getRegClass(VReg);

    // Record any overlaps with regmask operands.
    BitVector RegMaskOverlaps;
    LIS.checkRegMaskInterference(VRegLI, RegMaskOverlaps);

    // Compute an initial allowed set for the current vreg.
    std::vector<unsigned> VRegAllowed;
    ArrayRef<MCPhysReg> RawPRegOrder = TRC->getRawAllocationOrder(MF);
    for (unsigned I = 0; I != RawPRegOrder.size(); ++I) {
      unsigned PReg = RawPRegOrder[I];
      if (MRI.isReserved(PReg))
        continue;

      // vregLI crosses a regmask operand that clobbers preg.
      if (!RegMaskOverlaps.empty() && !RegMaskOverlaps.test(PReg))
        continue;

      // vregLI overlaps fixed regunit interference.
      bool Interference = false;
      for (MCRegUnitIterator Units(PReg, &TRI); Units.isValid(); ++Units) {
        if (VRegLI.overlaps(LIS.getRegUnit(*Units))) {
          Interference = true;
          break;
        }
      }
      if (Interference)
        continue;

      // preg is usable for this virtual register.
      VRegAllowed.push_back(PReg);
    }

    // Check for vregs that have no allowed registers. These should be
    // pre-spilled and the new vregs added to the worklist.
    if (VRegAllowed.empty()) {
      SmallVector<Register, 8> NewVRegs;
      spillVReg(VReg, NewVRegs, MF, LIS, VRM, VRegSpiller);
      Worklist.insert(Worklist.end(), NewVRegs.begin(), NewVRegs.end());
      continue;
    } else
      VRegAllowedMap[VReg] = std::move(VRegAllowed);
  }

  for (auto &KV : VRegAllowedMap) {
    auto VReg = KV.first;

    // Move empty intervals to the EmptyIntervalVReg set.
    if (LIS.getInterval(VReg).empty()) {
      EmptyIntervalVRegs.insert(VReg);
      VRegsToAlloc.erase(VReg);
      continue;
    }

    auto &VRegAllowed = KV.second;

    PBQPRAGraph::RawVector NodeCosts(VRegAllowed.size() + 1, 0);

    // Tweak cost of callee saved registers, as using then force spilling and
    // restoring them. This would only happen in the prologue / epilogue though.
    for (unsigned i = 0; i != VRegAllowed.size(); ++i)
      if (isACalleeSavedRegister(VRegAllowed[i], TRI, MF))
        NodeCosts[1 + i] += 1.0;

    PBQPRAGraph::NodeId NId = G.addNode(std::move(NodeCosts));
    G.getNodeMetadata(NId).setVReg(VReg);
    G.getNodeMetadata(NId).setAllowedRegs(
      G.getMetadata().getAllowedRegs(std::move(VRegAllowed)));
    G.getMetadata().setNodeIdForVReg(VReg, NId);
  }
}

void RegAllocPBQP::spillVReg(Register VReg,
                             SmallVectorImpl<Register> &NewIntervals,
                             MachineFunction &MF, LiveIntervals &LIS,
                             VirtRegMap &VRM, Spiller &VRegSpiller) {
  VRegsToAlloc.erase(VReg);
  LiveRangeEdit LRE(&LIS.getInterval(VReg), NewIntervals, MF, LIS, &VRM,
                    nullptr, &DeadRemats);
  VRegSpiller.spill(LRE);

  const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
  (void)TRI;
  LLVM_DEBUG(dbgs() << "VREG " << printReg(VReg, &TRI) << " -> SPILLED (Cost: "
                    << LRE.getParent().weight() << ", New vregs: ");

  // Copy any newly inserted live intervals into the list of regs to
  // allocate.
  for (LiveRangeEdit::iterator I = LRE.begin(), E = LRE.end();
       I != E; ++I) {
    const LiveInterval &LI = LIS.getInterval(*I);
    assert(!LI.empty() && "Empty spill range.");
    LLVM_DEBUG(dbgs() << printReg(LI.reg(), &TRI) << " ");
    VRegsToAlloc.insert(LI.reg());
  }

  LLVM_DEBUG(dbgs() << ")\n");
}

bool RegAllocPBQP::mapPBQPToRegAlloc(const PBQPRAGraph &G,
                                     const PBQP::Solution &Solution,
                                     VirtRegMap &VRM,
                                     Spiller &VRegSpiller) {
  MachineFunction &MF = G.getMetadata().MF;
  LiveIntervals &LIS = G.getMetadata().LIS;
  const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
  (void)TRI;

  // Set to true if we have any spills
  bool AnotherRoundNeeded = false;

  // Clear the existing allocation.
  VRM.clearAllVirt();

  // Iterate over the nodes mapping the PBQP solution to a register
  // assignment.
  for (auto NId : G.nodeIds()) {
    unsigned VReg = G.getNodeMetadata(NId).getVReg();
    unsigned AllocOption = Solution.getSelection(NId);

    if (AllocOption != PBQP::RegAlloc::getSpillOptionIdx()) {
      unsigned PReg = G.getNodeMetadata(NId).getAllowedRegs()[AllocOption - 1];
      LLVM_DEBUG(dbgs() << "VREG " << printReg(VReg, &TRI) << " -> "
                        << TRI.getName(PReg) << "\n");
      assert(PReg != 0 && "Invalid preg selected.");
      VRM.assignVirt2Phys(VReg, PReg);
    } else {
      // Spill VReg. If this introduces new intervals we'll need another round
      // of allocation.
      SmallVector<Register, 8> NewVRegs;
      spillVReg(VReg, NewVRegs, MF, LIS, VRM, VRegSpiller);
      AnotherRoundNeeded |= !NewVRegs.empty();
    }
  }

  return !AnotherRoundNeeded;
}

void RegAllocPBQP::finalizeAlloc(MachineFunction &MF,
                                 LiveIntervals &LIS,
                                 VirtRegMap &VRM) const {
  MachineRegisterInfo &MRI = MF.getRegInfo();

  // First allocate registers for the empty intervals.
  for (RegSet::const_iterator
         I = EmptyIntervalVRegs.begin(), E = EmptyIntervalVRegs.end();
         I != E; ++I) {
    LiveInterval &LI = LIS.getInterval(*I);

    unsigned PReg = MRI.getSimpleHint(LI.reg());

    if (PReg == 0) {
      const TargetRegisterClass &RC = *MRI.getRegClass(LI.reg());
      const ArrayRef<MCPhysReg> RawPRegOrder = RC.getRawAllocationOrder(MF);
      for (unsigned CandidateReg : RawPRegOrder) {
        if (!VRM.getRegInfo().isReserved(CandidateReg)) {
          PReg = CandidateReg;
          break;
        }
      }
      assert(PReg &&
             "No un-reserved physical registers in this register class");
    }

    VRM.assignVirt2Phys(LI.reg(), PReg);
  }
}

void RegAllocPBQP::postOptimization(Spiller &VRegSpiller, LiveIntervals &LIS) {
  VRegSpiller.postOptimization();
  /// Remove dead defs because of rematerialization.
  for (auto DeadInst : DeadRemats) {
    LIS.RemoveMachineInstrFromMaps(*DeadInst);
    DeadInst->eraseFromParent();
  }
  DeadRemats.clear();
}

static inline float normalizePBQPSpillWeight(float UseDefFreq, unsigned Size,
                                         unsigned NumInstr) {
  // All intervals have a spill weight that is mostly proportional to the number
  // of uses, with uses in loops having a bigger weight.
  return NumInstr * normalizeSpillWeight(UseDefFreq, Size, 1);
}

bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
  LiveIntervals &LIS = getAnalysis<LiveIntervals>();
  MachineBlockFrequencyInfo &MBFI =
    getAnalysis<MachineBlockFrequencyInfo>();

  VirtRegMap &VRM = getAnalysis<VirtRegMap>();

  VirtRegAuxInfo VRAI(MF, LIS, &VRM, getAnalysis<MachineLoopInfo>(), MBFI,
                      normalizePBQPSpillWeight);
  VRAI.calculateSpillWeightsAndHints();

  std::unique_ptr<Spiller> VRegSpiller(createInlineSpiller(*this, MF, VRM));

  MF.getRegInfo().freezeReservedRegs(MF);

  LLVM_DEBUG(dbgs() << "PBQP Register Allocating for " << MF.getName() << "\n");

  // Allocator main loop:
  //
  // * Map current regalloc problem to a PBQP problem
  // * Solve the PBQP problem
  // * Map the solution back to a register allocation
  // * Spill if necessary
  //
  // This process is continued till no more spills are generated.

  // Find the vreg intervals in need of allocation.
  findVRegIntervalsToAlloc(MF, LIS);

#ifndef NDEBUG
  const Function &F = MF.getFunction();
  std::string FullyQualifiedName =
    F.getParent()->getModuleIdentifier() + "." + F.getName().str();
#endif

  // If there are non-empty intervals allocate them using pbqp.
  if (!VRegsToAlloc.empty()) {
    const TargetSubtargetInfo &Subtarget = MF.getSubtarget();
    std::unique_ptr<PBQPRAConstraintList> ConstraintsRoot =
      std::make_unique<PBQPRAConstraintList>();
    ConstraintsRoot->addConstraint(std::make_unique<SpillCosts>());
    ConstraintsRoot->addConstraint(std::make_unique<Interference>());
    if (PBQPCoalescing)
      ConstraintsRoot->addConstraint(std::make_unique<Coalescing>());
    ConstraintsRoot->addConstraint(Subtarget.getCustomPBQPConstraints());

    bool PBQPAllocComplete = false;
    unsigned Round = 0;

    while (!PBQPAllocComplete) {
      LLVM_DEBUG(dbgs() << "  PBQP Regalloc round " << Round << ":\n");

      PBQPRAGraph G(PBQPRAGraph::GraphMetadata(MF, LIS, MBFI));
      initializeGraph(G, VRM, *VRegSpiller);
      ConstraintsRoot->apply(G);

#ifndef NDEBUG
      if (PBQPDumpGraphs) {
        std::ostringstream RS;
        RS << Round;
        std::string GraphFileName = FullyQualifiedName + "." + RS.str() +
                                    ".pbqpgraph";
        std::error_code EC;
        raw_fd_ostream OS(GraphFileName, EC, sys::fs::OF_Text);
        LLVM_DEBUG(dbgs() << "Dumping graph for round " << Round << " to \""
                          << GraphFileName << "\"\n");
        G.dump(OS);
      }
#endif

      PBQP::Solution Solution = PBQP::RegAlloc::solve(G);
      PBQPAllocComplete = mapPBQPToRegAlloc(G, Solution, VRM, *VRegSpiller);
      ++Round;
    }
  }

  // Finalise allocation, allocate empty ranges.
  finalizeAlloc(MF, LIS, VRM);
  postOptimization(*VRegSpiller, LIS);
  VRegsToAlloc.clear();
  EmptyIntervalVRegs.clear();

  LLVM_DEBUG(dbgs() << "Post alloc VirtRegMap:\n" << VRM << "\n");

  return true;
}

/// Create Printable object for node and register info.
static Printable PrintNodeInfo(PBQP::RegAlloc::PBQPRAGraph::NodeId NId,
                               const PBQP::RegAlloc::PBQPRAGraph &G) {
  return Printable([NId, &G](raw_ostream &OS) {
    const MachineRegisterInfo &MRI = G.getMetadata().MF.getRegInfo();
    const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo();
    unsigned VReg = G.getNodeMetadata(NId).getVReg();
    const char *RegClassName = TRI->getRegClassName(MRI.getRegClass(VReg));
    OS << NId << " (" << RegClassName << ':' << printReg(VReg, TRI) << ')';
  });
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
LLVM_DUMP_METHOD void PBQP::RegAlloc::PBQPRAGraph::dump(raw_ostream &OS) const {
  for (auto NId : nodeIds()) {
    const Vector &Costs = getNodeCosts(NId);
    assert(Costs.getLength() != 0 && "Empty vector in graph.");
    OS << PrintNodeInfo(NId, *this) << ": " << Costs << '\n';
  }
  OS << '\n';

  for (auto EId : edgeIds()) {
    NodeId N1Id = getEdgeNode1Id(EId);
    NodeId N2Id = getEdgeNode2Id(EId);
    assert(N1Id != N2Id && "PBQP graphs should not have self-edges.");
    const Matrix &M = getEdgeCosts(EId);
    assert(M.getRows() != 0 && "No rows in matrix.");
    assert(M.getCols() != 0 && "No cols in matrix.");
    OS << PrintNodeInfo(N1Id, *this) << ' ' << M.getRows() << " rows / ";
    OS << PrintNodeInfo(N2Id, *this) << ' ' << M.getCols() << " cols:\n";
    OS << M << '\n';
  }
}

LLVM_DUMP_METHOD void PBQP::RegAlloc::PBQPRAGraph::dump() const {
  dump(dbgs());
}
#endif

void PBQP::RegAlloc::PBQPRAGraph::printDot(raw_ostream &OS) const {
  OS << "graph {\n";
  for (auto NId : nodeIds()) {
    OS << "  node" << NId << " [ label=\""
       << PrintNodeInfo(NId, *this) << "\\n"
       << getNodeCosts(NId) << "\" ]\n";
  }

  OS << "  edge [ len=" << nodeIds().size() << " ]\n";
  for (auto EId : edgeIds()) {
    OS << "  node" << getEdgeNode1Id(EId)
       << " -- node" << getEdgeNode2Id(EId)
       << " [ label=\"";
    const Matrix &EdgeCosts = getEdgeCosts(EId);
    for (unsigned i = 0; i < EdgeCosts.getRows(); ++i) {
      OS << EdgeCosts.getRowAsVector(i) << "\\n";
    }
    OS << "\" ]\n";
  }
  OS << "}\n";
}

FunctionPass *llvm::createPBQPRegisterAllocator(char *customPassID) {
  return new RegAllocPBQP(customPassID);
}

FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
  return createPBQPRegisterAllocator();
}