ARMParallelDSP.cpp 26.9 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
//===- ARMParallelDSP.cpp - Parallel DSP Pass -----------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
/// \file
/// Armv6 introduced instructions to perform 32-bit SIMD operations. The
/// purpose of this pass is do some IR pattern matching to create ACLE
/// DSP intrinsics, which map on these 32-bit SIMD operations.
/// This pass runs only when unaligned accesses is supported/enabled.
//
//===----------------------------------------------------------------------===//

#include "ARM.h"
#include "ARMSubtarget.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/LoopAccessAnalysis.h"
#include "llvm/Analysis/OrderedBasicBlock.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicsARM.h"
#include "llvm/IR/NoFolder.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Pass.h"
#include "llvm/PassRegistry.h"
#include "llvm/PassSupport.h"
#include "llvm/Support/Debug.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"

using namespace llvm;
using namespace PatternMatch;

#define DEBUG_TYPE "arm-parallel-dsp"

STATISTIC(NumSMLAD , "Number of smlad instructions generated");

static cl::opt<bool>
DisableParallelDSP("disable-arm-parallel-dsp", cl::Hidden, cl::init(false),
                   cl::desc("Disable the ARM Parallel DSP pass"));

static cl::opt<unsigned>
NumLoadLimit("arm-parallel-dsp-load-limit", cl::Hidden, cl::init(16),
             cl::desc("Limit the number of loads analysed"));

namespace {
  struct MulCandidate;
  class Reduction;

  using MulCandList = SmallVector<std::unique_ptr<MulCandidate>, 8>;
  using MemInstList = SmallVectorImpl<LoadInst*>;
  using MulPairList = SmallVector<std::pair<MulCandidate*, MulCandidate*>, 8>;

  // 'MulCandidate' holds the multiplication instructions that are candidates
  // for parallel execution.
  struct MulCandidate {
    Instruction   *Root;
    Value*        LHS;
    Value*        RHS;
    bool          Exchange = false;
    bool          ReadOnly = true;
    bool          Paired = false;
    SmallVector<LoadInst*, 2> VecLd;    // Container for loads to widen.

    MulCandidate(Instruction *I, Value *lhs, Value *rhs) :
      Root(I), LHS(lhs), RHS(rhs) { }

    bool HasTwoLoadInputs() const {
      return isa<LoadInst>(LHS) && isa<LoadInst>(RHS);
    }

    LoadInst *getBaseLoad() const {
      return VecLd.front();
    }
  };

  /// Represent a sequence of multiply-accumulate operations with the aim to
  /// perform the multiplications in parallel.
  class Reduction {
    Instruction     *Root = nullptr;
    Value           *Acc = nullptr;
    MulCandList     Muls;
    MulPairList        MulPairs;
    SetVector<Instruction*> Adds;

  public:
    Reduction() = delete;

    Reduction (Instruction *Add) : Root(Add) { }

    /// Record an Add instruction that is a part of the this reduction.
    void InsertAdd(Instruction *I) { Adds.insert(I); }

    /// Create MulCandidates, each rooted at a Mul instruction, that is a part
    /// of this reduction.
    void InsertMuls() {
      auto GetMulOperand = [](Value *V) -> Instruction* {
        if (auto *SExt = dyn_cast<SExtInst>(V)) {
          if (auto *I = dyn_cast<Instruction>(SExt->getOperand(0)))
            if (I->getOpcode() == Instruction::Mul)
              return I;
        } else if (auto *I = dyn_cast<Instruction>(V)) {
          if (I->getOpcode() == Instruction::Mul)
            return I;
        }
        return nullptr;
      };

      auto InsertMul = [this](Instruction *I) {
        Value *LHS = cast<Instruction>(I->getOperand(0))->getOperand(0);
        Value *RHS = cast<Instruction>(I->getOperand(1))->getOperand(0);
        Muls.push_back(std::make_unique<MulCandidate>(I, LHS, RHS));
      };

      for (auto *Add : Adds) {
        if (Add == Acc)
          continue;
        if (auto *Mul = GetMulOperand(Add->getOperand(0)))
          InsertMul(Mul);
        if (auto *Mul = GetMulOperand(Add->getOperand(1)))
          InsertMul(Mul);
      }
    }

    /// Add the incoming accumulator value, returns true if a value had not
    /// already been added. Returning false signals to the user that this
    /// reduction already has a value to initialise the accumulator.
    bool InsertAcc(Value *V) {
      if (Acc)
        return false;
      Acc = V;
      return true;
    }

    /// Set two MulCandidates, rooted at muls, that can be executed as a single
    /// parallel operation.
    void AddMulPair(MulCandidate *Mul0, MulCandidate *Mul1,
                    bool Exchange = false) {
      LLVM_DEBUG(dbgs() << "Pairing:\n"
                 << *Mul0->Root << "\n"
                 << *Mul1->Root << "\n");
      Mul0->Paired = true;
      Mul1->Paired = true;
      if (Exchange)
        Mul1->Exchange = true;
      MulPairs.push_back(std::make_pair(Mul0, Mul1));
    }

    /// Return true if enough mul operations are found that can be executed in
    /// parallel.
    bool CreateParallelPairs();

    /// Return the add instruction which is the root of the reduction.
    Instruction *getRoot() { return Root; }

    bool is64Bit() const { return Root->getType()->isIntegerTy(64); }

    Type *getType() const { return Root->getType(); }

    /// Return the incoming value to be accumulated. This maybe null.
    Value *getAccumulator() { return Acc; }

    /// Return the set of adds that comprise the reduction.
    SetVector<Instruction*> &getAdds() { return Adds; }

    /// Return the MulCandidate, rooted at mul instruction, that comprise the
    /// the reduction.
    MulCandList &getMuls() { return Muls; }

    /// Return the MulCandidate, rooted at mul instructions, that have been
    /// paired for parallel execution.
    MulPairList &getMulPairs() { return MulPairs; }

    /// To finalise, replace the uses of the root with the intrinsic call.
    void UpdateRoot(Instruction *SMLAD) {
      Root->replaceAllUsesWith(SMLAD);
    }

    void dump() {
      LLVM_DEBUG(dbgs() << "Reduction:\n";
        for (auto *Add : Adds)
          LLVM_DEBUG(dbgs() << *Add << "\n");
        for (auto &Mul : Muls)
          LLVM_DEBUG(dbgs() << *Mul->Root << "\n"
                     << "  " << *Mul->LHS << "\n"
                     << "  " << *Mul->RHS << "\n");
        LLVM_DEBUG(if (Acc) dbgs() << "Acc in: " << *Acc << "\n")
      );
    }
  };

  class WidenedLoad {
    LoadInst *NewLd = nullptr;
    SmallVector<LoadInst*, 4> Loads;

  public:
    WidenedLoad(SmallVectorImpl<LoadInst*> &Lds, LoadInst *Wide)
      : NewLd(Wide) {
      for (auto *I : Lds)
        Loads.push_back(I);
    }
    LoadInst *getLoad() {
      return NewLd;
    }
  };

  class ARMParallelDSP : public FunctionPass {
    ScalarEvolution   *SE;
    AliasAnalysis     *AA;
    TargetLibraryInfo *TLI;
    DominatorTree     *DT;
    const DataLayout  *DL;
    Module            *M;
    std::map<LoadInst*, LoadInst*> LoadPairs;
    SmallPtrSet<LoadInst*, 4> OffsetLoads;
    std::map<LoadInst*, std::unique_ptr<WidenedLoad>> WideLoads;

    template<unsigned>
    bool IsNarrowSequence(Value *V);
    bool Search(Value *V, BasicBlock *BB, Reduction &R);
    bool RecordMemoryOps(BasicBlock *BB);
    void InsertParallelMACs(Reduction &Reduction);
    bool AreSequentialLoads(LoadInst *Ld0, LoadInst *Ld1, MemInstList &VecMem);
    LoadInst* CreateWideLoad(MemInstList &Loads, IntegerType *LoadTy);
    bool CreateParallelPairs(Reduction &R);

    /// Try to match and generate: SMLAD, SMLADX - Signed Multiply Accumulate
    /// Dual performs two signed 16x16-bit multiplications. It adds the
    /// products to a 32-bit accumulate operand. Optionally, the instruction can
    /// exchange the halfwords of the second operand before performing the
    /// arithmetic.
    bool MatchSMLAD(Function &F);

  public:
    static char ID;

    ARMParallelDSP() : FunctionPass(ID) { }

    void getAnalysisUsage(AnalysisUsage &AU) const override {
      FunctionPass::getAnalysisUsage(AU);
      AU.addRequired<AssumptionCacheTracker>();
      AU.addRequired<ScalarEvolutionWrapperPass>();
      AU.addRequired<AAResultsWrapperPass>();
      AU.addRequired<TargetLibraryInfoWrapperPass>();
      AU.addRequired<DominatorTreeWrapperPass>();
      AU.addRequired<TargetPassConfig>();
      AU.addPreserved<ScalarEvolutionWrapperPass>();
      AU.addPreserved<GlobalsAAWrapperPass>();
      AU.setPreservesCFG();
    }

    bool runOnFunction(Function &F) override {
      if (DisableParallelDSP)
        return false;
      if (skipFunction(F))
        return false;

      SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
      AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
      TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
      DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
      auto &TPC = getAnalysis<TargetPassConfig>();

      M = F.getParent();
      DL = &M->getDataLayout();

      auto &TM = TPC.getTM<TargetMachine>();
      auto *ST = &TM.getSubtarget<ARMSubtarget>(F);

      if (!ST->allowsUnalignedMem()) {
        LLVM_DEBUG(dbgs() << "Unaligned memory access not supported: not "
                             "running pass ARMParallelDSP\n");
        return false;
      }

      if (!ST->hasDSP()) {
        LLVM_DEBUG(dbgs() << "DSP extension not enabled: not running pass "
                             "ARMParallelDSP\n");
        return false;
      }

      if (!ST->isLittle()) {
        LLVM_DEBUG(dbgs() << "Only supporting little endian: not running pass "
                          << "ARMParallelDSP\n");
        return false;
      }

      LLVM_DEBUG(dbgs() << "\n== Parallel DSP pass ==\n");
      LLVM_DEBUG(dbgs() << " - " << F.getName() << "\n\n");

      bool Changes = MatchSMLAD(F);
      return Changes;
    }
  };
}

template<typename MemInst>
static bool AreSequentialAccesses(MemInst *MemOp0, MemInst *MemOp1,
                                  const DataLayout &DL, ScalarEvolution &SE) {
  if (isConsecutiveAccess(MemOp0, MemOp1, DL, SE))
    return true;
  return false;
}

bool ARMParallelDSP::AreSequentialLoads(LoadInst *Ld0, LoadInst *Ld1,
                                        MemInstList &VecMem) {
  if (!Ld0 || !Ld1)
    return false;

  if (!LoadPairs.count(Ld0) || LoadPairs[Ld0] != Ld1)
    return false;

  LLVM_DEBUG(dbgs() << "Loads are sequential and valid:\n";
    dbgs() << "Ld0:"; Ld0->dump();
    dbgs() << "Ld1:"; Ld1->dump();
  );

  VecMem.clear();
  VecMem.push_back(Ld0);
  VecMem.push_back(Ld1);
  return true;
}

// MaxBitwidth: the maximum supported bitwidth of the elements in the DSP
// instructions, which is set to 16. So here we should collect all i8 and i16
// narrow operations.
// TODO: we currently only collect i16, and will support i8 later, so that's
// why we check that types are equal to MaxBitWidth, and not <= MaxBitWidth.
template<unsigned MaxBitWidth>
bool ARMParallelDSP::IsNarrowSequence(Value *V) {
  if (auto *SExt = dyn_cast<SExtInst>(V)) {
    if (SExt->getSrcTy()->getIntegerBitWidth() != MaxBitWidth)
      return false;

    if (auto *Ld = dyn_cast<LoadInst>(SExt->getOperand(0))) {
      // Check that this load could be paired.
      return LoadPairs.count(Ld) || OffsetLoads.count(Ld);
    }
  }
  return false;
}

/// Iterate through the block and record base, offset pairs of loads which can
/// be widened into a single load.
bool ARMParallelDSP::RecordMemoryOps(BasicBlock *BB) {
  SmallVector<LoadInst*, 8> Loads;
  SmallVector<Instruction*, 8> Writes;
  LoadPairs.clear();
  WideLoads.clear();
  OrderedBasicBlock OrderedBB(BB);

  // Collect loads and instruction that may write to memory. For now we only
  // record loads which are simple, sign-extended and have a single user.
  // TODO: Allow zero-extended loads.
  for (auto &I : *BB) {
    if (I.mayWriteToMemory())
      Writes.push_back(&I);
    auto *Ld = dyn_cast<LoadInst>(&I);
    if (!Ld || !Ld->isSimple() ||
        !Ld->hasOneUse() || !isa<SExtInst>(Ld->user_back()))
      continue;
    Loads.push_back(Ld);
  }

  if (Loads.empty() || Loads.size() > NumLoadLimit)
    return false;

  using InstSet = std::set<Instruction*>;
  using DepMap = std::map<Instruction*, InstSet>;
  DepMap RAWDeps;

  // Record any writes that may alias a load.
  const auto Size = LocationSize::unknown();
  for (auto Write : Writes) {
    for (auto Read : Loads) {
      MemoryLocation ReadLoc =
        MemoryLocation(Read->getPointerOperand(), Size);

      if (!isModOrRefSet(intersectModRef(AA->getModRefInfo(Write, ReadLoc),
          ModRefInfo::ModRef)))
        continue;
      if (OrderedBB.dominates(Write, Read))
        RAWDeps[Read].insert(Write);
    }
  }

  // Check whether there's not a write between the two loads which would
  // prevent them from being safely merged.
  auto SafeToPair = [&](LoadInst *Base, LoadInst *Offset) {
    LoadInst *Dominator = OrderedBB.dominates(Base, Offset) ? Base : Offset;
    LoadInst *Dominated = OrderedBB.dominates(Base, Offset) ? Offset : Base;

    if (RAWDeps.count(Dominated)) {
      InstSet &WritesBefore = RAWDeps[Dominated];

      for (auto Before : WritesBefore) {
        // We can't move the second load backward, past a write, to merge
        // with the first load.
        if (OrderedBB.dominates(Dominator, Before))
          return false;
      }
    }
    return true;
  };

  // Record base, offset load pairs.
  for (auto *Base : Loads) {
    for (auto *Offset : Loads) {
      if (Base == Offset || OffsetLoads.count(Offset))
        continue;

      if (AreSequentialAccesses<LoadInst>(Base, Offset, *DL, *SE) &&
          SafeToPair(Base, Offset)) {
        LoadPairs[Base] = Offset;
        OffsetLoads.insert(Offset);
        break;
      }
    }
  }

  LLVM_DEBUG(if (!LoadPairs.empty()) {
               dbgs() << "Consecutive load pairs:\n";
               for (auto &MapIt : LoadPairs) {
                 LLVM_DEBUG(dbgs() << *MapIt.first << ", "
                            << *MapIt.second << "\n");
               }
             });
  return LoadPairs.size() > 1;
}

// Search recursively back through the operands to find a tree of values that
// form a multiply-accumulate chain. The search records the Add and Mul
// instructions that form the reduction and allows us to find a single value
// to be used as the initial input to the accumlator.
bool ARMParallelDSP::Search(Value *V, BasicBlock *BB, Reduction &R) {
  // If we find a non-instruction, try to use it as the initial accumulator
  // value. This may have already been found during the search in which case
  // this function will return false, signaling a search fail.
  auto *I = dyn_cast<Instruction>(V);
  if (!I)
    return R.InsertAcc(V);

  if (I->getParent() != BB)
    return false;

  switch (I->getOpcode()) {
  default:
    break;
  case Instruction::PHI:
    // Could be the accumulator value.
    return R.InsertAcc(V);
  case Instruction::Add: {
    // Adds should be adding together two muls, or another add and a mul to
    // be within the mac chain. One of the operands may also be the
    // accumulator value at which point we should stop searching.
    R.InsertAdd(I);
    Value *LHS = I->getOperand(0);
    Value *RHS = I->getOperand(1);
    bool ValidLHS = Search(LHS, BB, R);
    bool ValidRHS = Search(RHS, BB, R);

    if (ValidLHS && ValidRHS)
      return true;

    return R.InsertAcc(I);
  }
  case Instruction::Mul: {
    Value *MulOp0 = I->getOperand(0);
    Value *MulOp1 = I->getOperand(1);
    return IsNarrowSequence<16>(MulOp0) && IsNarrowSequence<16>(MulOp1);
  }
  case Instruction::SExt:
    return Search(I->getOperand(0), BB, R);
  }
  return false;
}

// The pass needs to identify integer add/sub reductions of 16-bit vector
// multiplications.
// To use SMLAD:
// 1) we first need to find integer add then look for this pattern:
//
// acc0 = ...
// ld0 = load i16
// sext0 = sext i16 %ld0 to i32
// ld1 = load i16
// sext1 = sext i16 %ld1 to i32
// mul0 = mul %sext0, %sext1
// ld2 = load i16
// sext2 = sext i16 %ld2 to i32
// ld3 = load i16
// sext3 = sext i16 %ld3 to i32
// mul1 = mul i32 %sext2, %sext3
// add0 = add i32 %mul0, %acc0
// acc1 = add i32 %add0, %mul1
//
// Which can be selected to:
//
// ldr r0
// ldr r1
// smlad r2, r0, r1, r2
//
// If constants are used instead of loads, these will need to be hoisted
// out and into a register.
//
// If loop invariants are used instead of loads, these need to be packed
// before the loop begins.
//
bool ARMParallelDSP::MatchSMLAD(Function &F) {
  bool Changed = false;

  for (auto &BB : F) {
    SmallPtrSet<Instruction*, 4> AllAdds;
    if (!RecordMemoryOps(&BB))
      continue;

    for (Instruction &I : reverse(BB)) {
      if (I.getOpcode() != Instruction::Add)
        continue;

      if (AllAdds.count(&I))
        continue;

      const auto *Ty = I.getType();
      if (!Ty->isIntegerTy(32) && !Ty->isIntegerTy(64))
        continue;

      Reduction R(&I);
      if (!Search(&I, &BB, R))
        continue;

      R.InsertMuls();
      LLVM_DEBUG(dbgs() << "After search, Reduction:\n"; R.dump());

      if (!CreateParallelPairs(R))
        continue;

      InsertParallelMACs(R);
      Changed = true;
      AllAdds.insert(R.getAdds().begin(), R.getAdds().end());
    }
  }

  return Changed;
}

bool ARMParallelDSP::CreateParallelPairs(Reduction &R) {

  // Not enough mul operations to make a pair.
  if (R.getMuls().size() < 2)
    return false;

  // Check that the muls operate directly upon sign extended loads.
  for (auto &MulCand : R.getMuls()) {
    if (!MulCand->HasTwoLoadInputs())
      return false;
  }

  auto CanPair = [&](Reduction &R, MulCandidate *PMul0, MulCandidate *PMul1) {
    // The first elements of each vector should be loads with sexts. If we
    // find that its two pairs of consecutive loads, then these can be
    // transformed into two wider loads and the users can be replaced with
    // DSP intrinsics.
    auto Ld0 = static_cast<LoadInst*>(PMul0->LHS);
    auto Ld1 = static_cast<LoadInst*>(PMul1->LHS);
    auto Ld2 = static_cast<LoadInst*>(PMul0->RHS);
    auto Ld3 = static_cast<LoadInst*>(PMul1->RHS);

    if (AreSequentialLoads(Ld0, Ld1, PMul0->VecLd)) {
      if (AreSequentialLoads(Ld2, Ld3, PMul1->VecLd)) {
        LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
        R.AddMulPair(PMul0, PMul1);
        return true;
      } else if (AreSequentialLoads(Ld3, Ld2, PMul1->VecLd)) {
        LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
        LLVM_DEBUG(dbgs() << "    exchanging Ld2 and Ld3\n");
        R.AddMulPair(PMul0, PMul1, true);
        return true;
      }
    } else if (AreSequentialLoads(Ld1, Ld0, PMul0->VecLd) &&
               AreSequentialLoads(Ld2, Ld3, PMul1->VecLd)) {
      LLVM_DEBUG(dbgs() << "OK: found two pairs of parallel loads!\n");
      LLVM_DEBUG(dbgs() << "    exchanging Ld0 and Ld1\n");
      LLVM_DEBUG(dbgs() << "    and swapping muls\n");
      // Only the second operand can be exchanged, so swap the muls.
      R.AddMulPair(PMul1, PMul0, true);
      return true;
    }
    return false;
  };

  MulCandList &Muls = R.getMuls();
  const unsigned Elems = Muls.size();
  for (unsigned i = 0; i < Elems; ++i) {
    MulCandidate *PMul0 = static_cast<MulCandidate*>(Muls[i].get());
    if (PMul0->Paired)
      continue;

    for (unsigned j = 0; j < Elems; ++j) {
      if (i == j)
        continue;

      MulCandidate *PMul1 = static_cast<MulCandidate*>(Muls[j].get());
      if (PMul1->Paired)
        continue;

      const Instruction *Mul0 = PMul0->Root;
      const Instruction *Mul1 = PMul1->Root;
      if (Mul0 == Mul1)
        continue;

      assert(PMul0 != PMul1 && "expected different chains");

      if (CanPair(R, PMul0, PMul1))
        break;
    }
  }
  return !R.getMulPairs().empty();
}

void ARMParallelDSP::InsertParallelMACs(Reduction &R) {

  auto CreateSMLAD = [&](LoadInst* WideLd0, LoadInst *WideLd1,
                         Value *Acc, bool Exchange,
                         Instruction *InsertAfter) {
    // Replace the reduction chain with an intrinsic call

    Value* Args[] = { WideLd0, WideLd1, Acc };
    Function *SMLAD = nullptr;
    if (Exchange)
      SMLAD = Acc->getType()->isIntegerTy(32) ?
        Intrinsic::getDeclaration(M, Intrinsic::arm_smladx) :
        Intrinsic::getDeclaration(M, Intrinsic::arm_smlaldx);
    else
      SMLAD = Acc->getType()->isIntegerTy(32) ?
        Intrinsic::getDeclaration(M, Intrinsic::arm_smlad) :
        Intrinsic::getDeclaration(M, Intrinsic::arm_smlald);

    IRBuilder<NoFolder> Builder(InsertAfter->getParent(),
                                BasicBlock::iterator(InsertAfter));
    Instruction *Call = Builder.CreateCall(SMLAD, Args);
    NumSMLAD++;
    return Call;
  };

  // Return the instruction after the dominated instruction.
  auto GetInsertPoint = [this](Value *A, Value *B) {
    assert((isa<Instruction>(A) || isa<Instruction>(B)) &&
           "expected at least one instruction");

    Value *V = nullptr;
    if (!isa<Instruction>(A))
      V = B;
    else if (!isa<Instruction>(B))
      V = A;
    else
      V = DT->dominates(cast<Instruction>(A), cast<Instruction>(B)) ? B : A;

    return &*++BasicBlock::iterator(cast<Instruction>(V));
  };

  Value *Acc = R.getAccumulator();

  // For any muls that were discovered but not paired, accumulate their values
  // as before.
  IRBuilder<NoFolder> Builder(R.getRoot()->getParent());
  MulCandList &MulCands = R.getMuls();
  for (auto &MulCand : MulCands) {
    if (MulCand->Paired)
      continue;

    Instruction *Mul = cast<Instruction>(MulCand->Root);
    LLVM_DEBUG(dbgs() << "Accumulating unpaired mul: " << *Mul << "\n");

    if (R.getType() != Mul->getType()) {
      assert(R.is64Bit() && "expected 64-bit result");
      Builder.SetInsertPoint(&*++BasicBlock::iterator(Mul));
      Mul = cast<Instruction>(Builder.CreateSExt(Mul, R.getRoot()->getType()));
    }

    if (!Acc) {
      Acc = Mul;
      continue;
    }

    // If Acc is the original incoming value to the reduction, it could be a
    // phi. But the phi will dominate Mul, meaning that Mul will be the
    // insertion point.
    Builder.SetInsertPoint(GetInsertPoint(Mul, Acc));
    Acc = Builder.CreateAdd(Mul, Acc);
  }

  if (!Acc) {
    Acc = R.is64Bit() ?
      ConstantInt::get(IntegerType::get(M->getContext(), 64), 0) :
      ConstantInt::get(IntegerType::get(M->getContext(), 32), 0);
  } else if (Acc->getType() != R.getType()) {
    Builder.SetInsertPoint(R.getRoot());
    Acc = Builder.CreateSExt(Acc, R.getType());
  }

  // Roughly sort the mul pairs in their program order.
  OrderedBasicBlock OrderedBB(R.getRoot()->getParent());
  llvm::sort(R.getMulPairs(), [&OrderedBB](auto &PairA, auto &PairB) {
               const Instruction *A = PairA.first->Root;
               const Instruction *B = PairB.first->Root;
               return OrderedBB.dominates(A, B);
             });

  IntegerType *Ty = IntegerType::get(M->getContext(), 32);
  for (auto &Pair : R.getMulPairs()) {
    MulCandidate *LHSMul = Pair.first;
    MulCandidate *RHSMul = Pair.second;
    LoadInst *BaseLHS = LHSMul->getBaseLoad();
    LoadInst *BaseRHS = RHSMul->getBaseLoad();
    LoadInst *WideLHS = WideLoads.count(BaseLHS) ?
      WideLoads[BaseLHS]->getLoad() : CreateWideLoad(LHSMul->VecLd, Ty);
    LoadInst *WideRHS = WideLoads.count(BaseRHS) ?
      WideLoads[BaseRHS]->getLoad() : CreateWideLoad(RHSMul->VecLd, Ty);

    Instruction *InsertAfter = GetInsertPoint(WideLHS, WideRHS);
    InsertAfter = GetInsertPoint(InsertAfter, Acc);
    Acc = CreateSMLAD(WideLHS, WideRHS, Acc, RHSMul->Exchange, InsertAfter);
  }
  R.UpdateRoot(cast<Instruction>(Acc));
}

LoadInst* ARMParallelDSP::CreateWideLoad(MemInstList &Loads,
                                         IntegerType *LoadTy) {
  assert(Loads.size() == 2 && "currently only support widening two loads");

  LoadInst *Base = Loads[0];
  LoadInst *Offset = Loads[1];

  Instruction *BaseSExt = dyn_cast<SExtInst>(Base->user_back());
  Instruction *OffsetSExt = dyn_cast<SExtInst>(Offset->user_back());

  assert((BaseSExt && OffsetSExt)
         && "Loads should have a single, extending, user");

  std::function<void(Value*, Value*)> MoveBefore =
    [&](Value *A, Value *B) -> void {
      if (!isa<Instruction>(A) || !isa<Instruction>(B))
        return;

      auto *Source = cast<Instruction>(A);
      auto *Sink = cast<Instruction>(B);

      if (DT->dominates(Source, Sink) ||
          Source->getParent() != Sink->getParent() ||
          isa<PHINode>(Source) || isa<PHINode>(Sink))
        return;

      Source->moveBefore(Sink);
      for (auto &Op : Source->operands())
        MoveBefore(Op, Source);
    };

  // Insert the load at the point of the original dominating load.
  LoadInst *DomLoad = DT->dominates(Base, Offset) ? Base : Offset;
  IRBuilder<NoFolder> IRB(DomLoad->getParent(),
                          ++BasicBlock::iterator(DomLoad));

  // Bitcast the pointer to a wider type and create the wide load, while making
  // sure to maintain the original alignment as this prevents ldrd from being
  // generated when it could be illegal due to memory alignment.
  const unsigned AddrSpace = DomLoad->getPointerAddressSpace();
  Value *VecPtr = IRB.CreateBitCast(Base->getPointerOperand(),
                                    LoadTy->getPointerTo(AddrSpace));
  LoadInst *WideLoad = IRB.CreateAlignedLoad(LoadTy, VecPtr,
                                             Base->getAlignment());

  // Make sure everything is in the correct order in the basic block.
  MoveBefore(Base->getPointerOperand(), VecPtr);
  MoveBefore(VecPtr, WideLoad);

  // From the wide load, create two values that equal the original two loads.
  // Loads[0] needs trunc while Loads[1] needs a lshr and trunc.
  // TODO: Support big-endian as well.
  Value *Bottom = IRB.CreateTrunc(WideLoad, Base->getType());
  Value *NewBaseSExt = IRB.CreateSExt(Bottom, BaseSExt->getType());
  BaseSExt->replaceAllUsesWith(NewBaseSExt);

  IntegerType *OffsetTy = cast<IntegerType>(Offset->getType());
  Value *ShiftVal = ConstantInt::get(LoadTy, OffsetTy->getBitWidth());
  Value *Top = IRB.CreateLShr(WideLoad, ShiftVal);
  Value *Trunc = IRB.CreateTrunc(Top, OffsetTy);
  Value *NewOffsetSExt = IRB.CreateSExt(Trunc, OffsetSExt->getType());
  OffsetSExt->replaceAllUsesWith(NewOffsetSExt);

  LLVM_DEBUG(dbgs() << "From Base and Offset:\n"
             << *Base << "\n" << *Offset << "\n"
             << "Created Wide Load:\n"
             << *WideLoad << "\n"
             << *Bottom << "\n"
             << *NewBaseSExt << "\n"
             << *Top << "\n"
             << *Trunc << "\n"
             << *NewOffsetSExt << "\n");
  WideLoads.emplace(std::make_pair(Base,
                                   std::make_unique<WidenedLoad>(Loads, WideLoad)));
  return WideLoad;
}

Pass *llvm::createARMParallelDSPPass() {
  return new ARMParallelDSP();
}

char ARMParallelDSP::ID = 0;

INITIALIZE_PASS_BEGIN(ARMParallelDSP, "arm-parallel-dsp",
                "Transform functions to use DSP intrinsics", false, false)
INITIALIZE_PASS_END(ARMParallelDSP, "arm-parallel-dsp",
                "Transform functions to use DSP intrinsics", false, false)