Training courses

Kernel and Embedded Linux

Bootlin training courses

Embedded Linux, kernel,
Yocto Project, Buildroot, real-time,
graphics, boot time, debugging...

Bootlin logo

Elixir Cross Referencer

  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
//===- HexagonVectorLoopCarriedReuse.cpp ----------------------------------===//
//
// 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 pass removes the computation of provably redundant expressions that have
// been computed earlier in a previous iteration. It relies on the use of PHIs
// to identify loop carried dependences. This is scalar replacement for vector
// types.
//
//-----------------------------------------------------------------------------
// Motivation: Consider the case where we have the following loop structure.
//
// Loop:
//  t0 = a[i];
//  t1 = f(t0);
//  t2 = g(t1);
//  ...
//  t3 = a[i+1];
//  t4 = f(t3);
//  t5 = g(t4);
//  t6 = op(t2, t5)
//  cond_branch <Loop>
//
// This can be converted to
//  t00 = a[0];
//  t10 = f(t00);
//  t20 = g(t10);
// Loop:
//  t2 = t20;
//  t3 = a[i+1];
//  t4 = f(t3);
//  t5 = g(t4);
//  t6 = op(t2, t5)
//  t20 = t5
//  cond_branch <Loop>
//
// SROA does a good job of reusing a[i+1] as a[i] in the next iteration.
// Such a loop comes to this pass in the following form.
//
// LoopPreheader:
//  X0 = a[0];
// Loop:
//  X2 = PHI<(X0, LoopPreheader), (X1, Loop)>
//  t1 = f(X2)   <-- I1
//  t2 = g(t1)
//  ...
//  X1 = a[i+1]
//  t4 = f(X1)   <-- I2
//  t5 = g(t4)
//  t6 = op(t2, t5)
//  cond_branch <Loop>
//
// In this pass, we look for PHIs such as X2 whose incoming values come only
// from the Loop Preheader and over the backedge and additionaly, both these
// values are the results of the same operation in terms of opcode. We call such
// a PHI node a dependence chain or DepChain. In this case, the dependence of X2
// over X1 is carried over only one iteration and so the DepChain is only one
// PHI node long.
//
// Then, we traverse the uses of the PHI (X2) and the uses of the value of the
// PHI coming  over the backedge (X1). We stop at the first pair of such users
// I1 (of X2) and I2 (of X1) that meet the following conditions.
// 1. I1 and I2 are the same operation, but with different operands.
// 2. X2 and X1 are used at the same operand number in the two instructions.
// 3. All other operands Op1 of I1 and Op2 of I2 are also such that there is a
//    a DepChain from Op1 to Op2 of the same length as that between X2 and X1.
//
// We then make the following transformation
// LoopPreheader:
//  X0 = a[0];
//  Y0 = f(X0);
// Loop:
//  X2 = PHI<(X0, LoopPreheader), (X1, Loop)>
//  Y2 = PHI<(Y0, LoopPreheader), (t4, Loop)>
//  t1 = f(X2)   <-- Will be removed by DCE.
//  t2 = g(Y2)
//  ...
//  X1 = a[i+1]
//  t4 = f(X1)
//  t5 = g(t4)
//  t6 = op(t2, t5)
//  cond_branch <Loop>
//
// We proceed until we cannot find any more such instructions I1 and I2.
//
// --- DepChains & Loop carried dependences ---
// Consider a single basic block loop such as
//
// LoopPreheader:
//  X0 = ...
//  Y0 = ...
// Loop:
//  X2 = PHI<(X0, LoopPreheader), (X1, Loop)>
//  Y2 = PHI<(Y0, LoopPreheader), (X2, Loop)>
//  ...
//  X1 = ...
//  ...
//  cond_branch <Loop>
//
// Then there is a dependence between X2 and X1 that goes back one iteration,
// i.e. X1 is used as X2 in the very next iteration. We represent this as a
// DepChain from X2 to X1 (X2->X1).
// Similarly, there is a dependence between Y2 and X1 that goes back two
// iterations. X1 is used as Y2 two iterations after it is computed. This is
// represented by a DepChain as (Y2->X2->X1).
//
// A DepChain has the following properties.
// 1. Num of edges in DepChain = Number of Instructions in DepChain = Number of
//    iterations of carried dependence + 1.
// 2. All instructions in the DepChain except the last are PHIs.
//
//===----------------------------------------------------------------------===//

#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsHexagon.h"
#include "llvm/IR/Use.h"
#include "llvm/IR/User.h"
#include "llvm/IR/Value.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils.h"
#include <algorithm>
#include <cassert>
#include <cstddef>
#include <map>
#include <memory>
#include <set>

using namespace llvm;

#define DEBUG_TYPE "hexagon-vlcr"

STATISTIC(HexagonNumVectorLoopCarriedReuse,
          "Number of values that were reused from a previous iteration.");

static cl::opt<int> HexagonVLCRIterationLim("hexagon-vlcr-iteration-lim",
    cl::Hidden,
    cl::desc("Maximum distance of loop carried dependences that are handled"),
    cl::init(2), cl::ZeroOrMore);

namespace llvm {

void initializeHexagonVectorLoopCarriedReusePass(PassRegistry&);
Pass *createHexagonVectorLoopCarriedReusePass();

} // end namespace llvm

namespace {

  // See info about DepChain in the comments at the top of this file.
  using ChainOfDependences = SmallVector<Instruction *, 4>;

  class DepChain {
    ChainOfDependences Chain;

  public:
    bool isIdentical(DepChain &Other) const {
      if (Other.size() != size())
        return false;
      ChainOfDependences &OtherChain = Other.getChain();
      for (int i = 0; i < size(); ++i) {
        if (Chain[i] != OtherChain[i])
          return false;
      }
      return true;
    }

    ChainOfDependences &getChain() {
      return Chain;
    }

    int size() const {
      return Chain.size();
    }

    void clear() {
      Chain.clear();
    }

    void push_back(Instruction *I) {
      Chain.push_back(I);
    }

    int iterations() const {
      return size() - 1;
    }

    Instruction *front() const {
      return Chain.front();
    }

    Instruction *back() const {
      return Chain.back();
    }

    Instruction *&operator[](const int index) {
      return Chain[index];
    }

   friend raw_ostream &operator<< (raw_ostream &OS, const DepChain &D);
  };

  LLVM_ATTRIBUTE_UNUSED
  raw_ostream &operator<<(raw_ostream &OS, const DepChain &D) {
    const ChainOfDependences &CD = D.Chain;
    int ChainSize = CD.size();
    OS << "**DepChain Start::**\n";
    for (int i = 0; i < ChainSize -1; ++i) {
      OS << *(CD[i]) << " -->\n";
    }
    OS << *CD[ChainSize-1] << "\n";
    return OS;
  }

  struct ReuseValue {
    Instruction *Inst2Replace = nullptr;

    // In the new PHI node that we'll construct this is the value that'll be
    // used over the backedge. This is the value that gets reused from a
    // previous iteration.
    Instruction *BackedgeInst = nullptr;
    std::map<Instruction *, DepChain *> DepChains;
    int Iterations = -1;

    ReuseValue() = default;

    void reset() {
      Inst2Replace = nullptr;
      BackedgeInst = nullptr;
      DepChains.clear();
      Iterations = -1;
    }
    bool isDefined() { return Inst2Replace != nullptr; }
  };

  LLVM_ATTRIBUTE_UNUSED
  raw_ostream &operator<<(raw_ostream &OS, const ReuseValue &RU) {
    OS << "** ReuseValue ***\n";
    OS << "Instruction to Replace: " << *(RU.Inst2Replace) << "\n";
    OS << "Backedge Instruction: " << *(RU.BackedgeInst) << "\n";
    return OS;
  }

  class HexagonVectorLoopCarriedReuse : public LoopPass {
  public:
    static char ID;

    explicit HexagonVectorLoopCarriedReuse() : LoopPass(ID) {
      PassRegistry *PR = PassRegistry::getPassRegistry();
      initializeHexagonVectorLoopCarriedReusePass(*PR);
    }

    StringRef getPassName() const override {
      return "Hexagon-specific loop carried reuse for HVX vectors";
    }

    void getAnalysisUsage(AnalysisUsage &AU) const override {
      AU.addRequired<LoopInfoWrapperPass>();
      AU.addRequiredID(LoopSimplifyID);
      AU.addRequiredID(LCSSAID);
      AU.addPreservedID(LCSSAID);
      AU.setPreservesCFG();
    }

    bool runOnLoop(Loop *L, LPPassManager &LPM) override;

  private:
    SetVector<DepChain *> Dependences;
    std::set<Instruction *> ReplacedInsts;
    Loop *CurLoop;
    ReuseValue ReuseCandidate;

    bool doVLCR();
    void findLoopCarriedDeps();
    void findValueToReuse();
    void findDepChainFromPHI(Instruction *I, DepChain &D);
    void reuseValue();
    Value *findValueInBlock(Value *Op, BasicBlock *BB);
    DepChain *getDepChainBtwn(Instruction *I1, Instruction *I2, int Iters);
    bool isEquivalentOperation(Instruction *I1, Instruction *I2);
    bool canReplace(Instruction *I);
    bool isCallInstCommutative(CallInst *C);
  };

} // end anonymous namespace

char HexagonVectorLoopCarriedReuse::ID = 0;

INITIALIZE_PASS_BEGIN(HexagonVectorLoopCarriedReuse, "hexagon-vlcr",
    "Hexagon-specific predictive commoning for HVX vectors", false, false)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
INITIALIZE_PASS_END(HexagonVectorLoopCarriedReuse, "hexagon-vlcr",
    "Hexagon-specific predictive commoning for HVX vectors", false, false)

bool HexagonVectorLoopCarriedReuse::runOnLoop(Loop *L, LPPassManager &LPM) {
  if (skipLoop(L))
    return false;

  if (!L->getLoopPreheader())
    return false;

  // Work only on innermost loops.
  if (!L->getSubLoops().empty())
    return false;

  // Work only on single basic blocks loops.
  if (L->getNumBlocks() != 1)
    return false;

  CurLoop = L;

  return doVLCR();
}

bool HexagonVectorLoopCarriedReuse::isCallInstCommutative(CallInst *C) {
  switch (C->getCalledFunction()->getIntrinsicID()) {
    case Intrinsic::hexagon_V6_vaddb:
    case Intrinsic::hexagon_V6_vaddb_128B:
    case Intrinsic::hexagon_V6_vaddh:
    case Intrinsic::hexagon_V6_vaddh_128B:
    case Intrinsic::hexagon_V6_vaddw:
    case Intrinsic::hexagon_V6_vaddw_128B:
    case Intrinsic::hexagon_V6_vaddubh:
    case Intrinsic::hexagon_V6_vaddubh_128B:
    case Intrinsic::hexagon_V6_vadduhw:
    case Intrinsic::hexagon_V6_vadduhw_128B:
    case Intrinsic::hexagon_V6_vaddhw:
    case Intrinsic::hexagon_V6_vaddhw_128B:
    case Intrinsic::hexagon_V6_vmaxb:
    case Intrinsic::hexagon_V6_vmaxb_128B:
    case Intrinsic::hexagon_V6_vmaxh:
    case Intrinsic::hexagon_V6_vmaxh_128B:
    case Intrinsic::hexagon_V6_vmaxw:
    case Intrinsic::hexagon_V6_vmaxw_128B:
    case Intrinsic::hexagon_V6_vmaxub:
    case Intrinsic::hexagon_V6_vmaxub_128B:
    case Intrinsic::hexagon_V6_vmaxuh:
    case Intrinsic::hexagon_V6_vmaxuh_128B:
    case Intrinsic::hexagon_V6_vminub:
    case Intrinsic::hexagon_V6_vminub_128B:
    case Intrinsic::hexagon_V6_vminuh:
    case Intrinsic::hexagon_V6_vminuh_128B:
    case Intrinsic::hexagon_V6_vminb:
    case Intrinsic::hexagon_V6_vminb_128B:
    case Intrinsic::hexagon_V6_vminh:
    case Intrinsic::hexagon_V6_vminh_128B:
    case Intrinsic::hexagon_V6_vminw:
    case Intrinsic::hexagon_V6_vminw_128B:
    case Intrinsic::hexagon_V6_vmpyub:
    case Intrinsic::hexagon_V6_vmpyub_128B:
    case Intrinsic::hexagon_V6_vmpyuh:
    case Intrinsic::hexagon_V6_vmpyuh_128B:
    case Intrinsic::hexagon_V6_vavgub:
    case Intrinsic::hexagon_V6_vavgub_128B:
    case Intrinsic::hexagon_V6_vavgh:
    case Intrinsic::hexagon_V6_vavgh_128B:
    case Intrinsic::hexagon_V6_vavguh:
    case Intrinsic::hexagon_V6_vavguh_128B:
    case Intrinsic::hexagon_V6_vavgw:
    case Intrinsic::hexagon_V6_vavgw_128B:
    case Intrinsic::hexagon_V6_vavgb:
    case Intrinsic::hexagon_V6_vavgb_128B:
    case Intrinsic::hexagon_V6_vavguw:
    case Intrinsic::hexagon_V6_vavguw_128B:
    case Intrinsic::hexagon_V6_vabsdiffh:
    case Intrinsic::hexagon_V6_vabsdiffh_128B:
    case Intrinsic::hexagon_V6_vabsdiffub:
    case Intrinsic::hexagon_V6_vabsdiffub_128B:
    case Intrinsic::hexagon_V6_vabsdiffuh:
    case Intrinsic::hexagon_V6_vabsdiffuh_128B:
    case Intrinsic::hexagon_V6_vabsdiffw:
    case Intrinsic::hexagon_V6_vabsdiffw_128B:
      return true;
    default:
      return false;
  }
}

bool HexagonVectorLoopCarriedReuse::isEquivalentOperation(Instruction *I1,
                                                          Instruction *I2) {
  if (!I1->isSameOperationAs(I2))
    return false;
  // This check is in place specifically for intrinsics. isSameOperationAs will
  // return two for any two hexagon intrinsics because they are essentially the
  // same instruciton (CallInst). We need to scratch the surface to see if they
  // are calls to the same function.
  if (CallInst *C1 = dyn_cast<CallInst>(I1)) {
    if (CallInst *C2 = dyn_cast<CallInst>(I2)) {
      if (C1->getCalledFunction() != C2->getCalledFunction())
        return false;
    }
  }

  // If both the Instructions are of Vector Type and any of the element
  // is integer constant, check their values too for equivalence.
  if (I1->getType()->isVectorTy() && I2->getType()->isVectorTy()) {
    unsigned NumOperands = I1->getNumOperands();
    for (unsigned i = 0; i < NumOperands; ++i) {
      ConstantInt *C1 = dyn_cast<ConstantInt>(I1->getOperand(i));
      ConstantInt *C2 = dyn_cast<ConstantInt>(I2->getOperand(i));
      if(!C1) continue;
      assert(C2);
      if (C1->getSExtValue() != C2->getSExtValue())
        return false;
    }
  }

  return true;
}

bool HexagonVectorLoopCarriedReuse::canReplace(Instruction *I) {
  const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
  if (!II)
    return true;

  switch (II->getIntrinsicID()) {
  case Intrinsic::hexagon_V6_hi:
  case Intrinsic::hexagon_V6_lo:
  case Intrinsic::hexagon_V6_hi_128B:
  case Intrinsic::hexagon_V6_lo_128B:
    LLVM_DEBUG(dbgs() << "Not considering for reuse: " << *II << "\n");
    return false;
  default:
    return true;
  }
}
void HexagonVectorLoopCarriedReuse::findValueToReuse() {
  for (auto *D : Dependences) {
    LLVM_DEBUG(dbgs() << "Processing dependence " << *(D->front()) << "\n");
    if (D->iterations() > HexagonVLCRIterationLim) {
      LLVM_DEBUG(
          dbgs()
          << ".. Skipping because number of iterations > than the limit\n");
      continue;
    }

    PHINode *PN = cast<PHINode>(D->front());
    Instruction *BEInst = D->back();
    int Iters = D->iterations();
    BasicBlock *BB = PN->getParent();
    LLVM_DEBUG(dbgs() << "Checking if any uses of " << *PN
                      << " can be reused\n");

    SmallVector<Instruction *, 4> PNUsers;
    for (auto UI = PN->use_begin(), E = PN->use_end(); UI != E; ++UI) {
      Use &U = *UI;
      Instruction *User = cast<Instruction>(U.getUser());

      if (User->getParent() != BB)
        continue;
      if (ReplacedInsts.count(User)) {
        LLVM_DEBUG(dbgs() << *User
                          << " has already been replaced. Skipping...\n");
        continue;
      }
      if (isa<PHINode>(User))
        continue;
      if (User->mayHaveSideEffects())
        continue;
      if (!canReplace(User))
        continue;

      PNUsers.push_back(User);
    }
    LLVM_DEBUG(dbgs() << PNUsers.size() << " use(s) of the PHI in the block\n");

    // For each interesting use I of PN, find an Instruction BEUser that
    // performs the same operation as I on BEInst and whose other operands,
    // if any, can also be rematerialized in OtherBB. We stop when we find the
    // first such Instruction BEUser. This is because once BEUser is
    // rematerialized in OtherBB, we may find more such "fixup" opportunities
    // in this block. So, we'll start over again.
    for (Instruction *I : PNUsers) {
      for (auto UI = BEInst->use_begin(), E = BEInst->use_end(); UI != E;
           ++UI) {
        Use &U = *UI;
        Instruction *BEUser = cast<Instruction>(U.getUser());

        if (BEUser->getParent() != BB)
          continue;
        if (!isEquivalentOperation(I, BEUser))
          continue;

        int NumOperands = I->getNumOperands();

        // Take operands of each PNUser one by one and try to find DepChain
        // with every operand of the BEUser. If any of the operands of BEUser
        // has DepChain with current operand of the PNUser, break the matcher
        // loop. Keep doing this for Every PNUser operand. If PNUser operand
        // does not have DepChain with any of the BEUser operand, break the
        // outer matcher loop, mark the BEUser as null and reset the ReuseCandidate.
        // This ensures that DepChain exist for all the PNUser operand with
        // BEUser operand. This also ensures that DepChains are independent of
        // the positions in PNUser and BEUser.
        std::map<Instruction *, DepChain *> DepChains;
        CallInst *C1 = dyn_cast<CallInst>(I);
        if ((I && I->isCommutative()) || (C1 && isCallInstCommutative(C1))) {
          bool Found = false;
          for (int OpNo = 0; OpNo < NumOperands; ++OpNo) {
            Value *Op = I->getOperand(OpNo);
            Instruction *OpInst = dyn_cast<Instruction>(Op);
            Found = false;
            for (int T = 0; T < NumOperands; ++T) {
              Value *BEOp = BEUser->getOperand(T);
              Instruction *BEOpInst = dyn_cast<Instruction>(BEOp);
              if (!OpInst && !BEOpInst) {
                if (Op == BEOp) {
                  Found = true;
                  break;
                }
              }

              if ((OpInst && !BEOpInst) || (!OpInst && BEOpInst))
                continue;

              DepChain *D = getDepChainBtwn(OpInst, BEOpInst, Iters);

              if (D) {
                Found = true;
                DepChains[OpInst] = D;
                break;
              }
            }
            if (!Found) {
              BEUser = nullptr;
              break;
            }
          }
        } else {

          for (int OpNo = 0; OpNo < NumOperands; ++OpNo) {
            Value *Op = I->getOperand(OpNo);
            Value *BEOp = BEUser->getOperand(OpNo);

            Instruction *OpInst = dyn_cast<Instruction>(Op);
            if (!OpInst) {
              if (Op == BEOp)
                continue;
              // Do not allow reuse to occur when the operands may be different
              // values.
              BEUser = nullptr;
              break;
            }

            Instruction *BEOpInst = dyn_cast<Instruction>(BEOp);
            DepChain *D = getDepChainBtwn(OpInst, BEOpInst, Iters);

            if (D) {
              DepChains[OpInst] = D;
            } else {
              BEUser = nullptr;
              break;
            }
          }
        }
        if (BEUser) {
          LLVM_DEBUG(dbgs() << "Found Value for reuse.\n");
          ReuseCandidate.Inst2Replace = I;
          ReuseCandidate.BackedgeInst = BEUser;
          ReuseCandidate.DepChains = DepChains;
          ReuseCandidate.Iterations = Iters;
          return;
        }
        ReuseCandidate.reset();
      }
    }
  }
  ReuseCandidate.reset();
}

Value *HexagonVectorLoopCarriedReuse::findValueInBlock(Value *Op,
                                                       BasicBlock *BB) {
  PHINode *PN = dyn_cast<PHINode>(Op);
  assert(PN);
  Value *ValueInBlock = PN->getIncomingValueForBlock(BB);
  return ValueInBlock;
}

void HexagonVectorLoopCarriedReuse::reuseValue() {
  LLVM_DEBUG(dbgs() << ReuseCandidate);
  Instruction *Inst2Replace = ReuseCandidate.Inst2Replace;
  Instruction *BEInst = ReuseCandidate.BackedgeInst;
  int NumOperands = Inst2Replace->getNumOperands();
  std::map<Instruction *, DepChain *> &DepChains = ReuseCandidate.DepChains;
  int Iterations = ReuseCandidate.Iterations;
  BasicBlock *LoopPH = CurLoop->getLoopPreheader();
  assert(!DepChains.empty() && "No DepChains");
  LLVM_DEBUG(dbgs() << "reuseValue is making the following changes\n");

  SmallVector<Instruction *, 4> InstsInPreheader;
  for (int i = 0; i < Iterations; ++i) {
    Instruction *InstInPreheader = Inst2Replace->clone();
    SmallVector<Value *, 4> Ops;
    for (int j = 0; j < NumOperands; ++j) {
      Instruction *I = dyn_cast<Instruction>(Inst2Replace->getOperand(j));
      if (!I)
        continue;
      // Get the DepChain corresponding to this operand.
      DepChain &D = *DepChains[I];
      // Get the PHI for the iteration number and find
      // the incoming value from the Loop Preheader for
      // that PHI.
      Value *ValInPreheader = findValueInBlock(D[i], LoopPH);
      InstInPreheader->setOperand(j, ValInPreheader);
    }
    InstsInPreheader.push_back(InstInPreheader);
    InstInPreheader->setName(Inst2Replace->getName() + ".hexagon.vlcr");
    InstInPreheader->insertBefore(LoopPH->getTerminator());
    LLVM_DEBUG(dbgs() << "Added " << *InstInPreheader << " to "
                      << LoopPH->getName() << "\n");
  }
  BasicBlock *BB = BEInst->getParent();
  IRBuilder<> IRB(BB);
  IRB.SetInsertPoint(BB->getFirstNonPHI());
  Value *BEVal = BEInst;
  PHINode *NewPhi;
  for (int i = Iterations-1; i >=0 ; --i) {
    Instruction *InstInPreheader = InstsInPreheader[i];
    NewPhi = IRB.CreatePHI(InstInPreheader->getType(), 2);
    NewPhi->addIncoming(InstInPreheader, LoopPH);
    NewPhi->addIncoming(BEVal, BB);
    LLVM_DEBUG(dbgs() << "Adding " << *NewPhi << " to " << BB->getName()
                      << "\n");
    BEVal = NewPhi;
  }
  // We are in LCSSA form. So, a value defined inside the Loop is used only
  // inside the loop. So, the following is safe.
  Inst2Replace->replaceAllUsesWith(NewPhi);
  ReplacedInsts.insert(Inst2Replace);
  ++HexagonNumVectorLoopCarriedReuse;
}

bool HexagonVectorLoopCarriedReuse::doVLCR() {
  assert(CurLoop->getSubLoops().empty() &&
         "Can do VLCR on the innermost loop only");
  assert((CurLoop->getNumBlocks() == 1) &&
         "Can do VLCR only on single block loops");

  bool Changed = false;
  bool Continue;

  LLVM_DEBUG(dbgs() << "Working on Loop: " << *CurLoop->getHeader() << "\n");
  do {
    // Reset datastructures.
    Dependences.clear();
    Continue = false;

    findLoopCarriedDeps();
    findValueToReuse();
    if (ReuseCandidate.isDefined()) {
      reuseValue();
      Changed = true;
      Continue = true;
    }
    llvm::for_each(Dependences, std::default_delete<DepChain>());
  } while (Continue);
  return Changed;
}

void HexagonVectorLoopCarriedReuse::findDepChainFromPHI(Instruction *I,
                                                        DepChain &D) {
  PHINode *PN = dyn_cast<PHINode>(I);
  if (!PN) {
    D.push_back(I);
    return;
  } else {
    auto NumIncomingValues = PN->getNumIncomingValues();
    if (NumIncomingValues != 2) {
      D.clear();
      return;
    }

    BasicBlock *BB = PN->getParent();
    if (BB != CurLoop->getHeader()) {
      D.clear();
      return;
    }

    Value *BEVal = PN->getIncomingValueForBlock(BB);
    Instruction *BEInst = dyn_cast<Instruction>(BEVal);
    // This is a single block loop with a preheader, so at least
    // one value should come over the backedge.
    assert(BEInst && "There should be a value over the backedge");

    Value *PreHdrVal =
      PN->getIncomingValueForBlock(CurLoop->getLoopPreheader());
    if(!PreHdrVal || !isa<Instruction>(PreHdrVal)) {
      D.clear();
      return;
    }
    D.push_back(PN);
    findDepChainFromPHI(BEInst, D);
  }
}

DepChain *HexagonVectorLoopCarriedReuse::getDepChainBtwn(Instruction *I1,
                                                         Instruction *I2,
                                                         int Iters) {
  for (auto *D : Dependences) {
    if (D->front() == I1 && D->back() == I2 && D->iterations() == Iters)
      return D;
  }
  return nullptr;
}

void HexagonVectorLoopCarriedReuse::findLoopCarriedDeps() {
  BasicBlock *BB = CurLoop->getHeader();
  for (auto I = BB->begin(), E = BB->end(); I != E && isa<PHINode>(I); ++I) {
    auto *PN = cast<PHINode>(I);
    if (!isa<VectorType>(PN->getType()))
      continue;

    DepChain *D = new DepChain();
    findDepChainFromPHI(PN, *D);
    if (D->size() != 0)
      Dependences.insert(D);
    else
      delete D;
  }
  LLVM_DEBUG(dbgs() << "Found " << Dependences.size() << " dependences\n");
  LLVM_DEBUG(for (size_t i = 0; i < Dependences.size();
                  ++i) { dbgs() << *Dependences[i] << "\n"; });
}

Pass *llvm::createHexagonVectorLoopCarriedReusePass() {
  return new HexagonVectorLoopCarriedReuse();
}