llvm.org GIT mirror llvm / 046ca04 lib / Analysis / BlockFrequencyInfoImpl.cpp
046ca04

Tree @046ca04 (Download .tar.gz)

BlockFrequencyInfoImpl.cpp @046ca04raw · history · blame

  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
//===- BlockFrequencyImplInfo.cpp - Block Frequency Info Implementation ---===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Loops should be simplified before this analysis.
//
//===----------------------------------------------------------------------===//

#include "llvm/Analysis/BlockFrequencyInfoImpl.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/GraphTraits.h"
#include "llvm/ADT/None.h"
#include "llvm/ADT/SCCIterator.h"
#include "llvm/IR/Function.h"
#include "llvm/Support/BlockFrequency.h"
#include "llvm/Support/BranchProbability.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ScaledNumber.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <cassert>
#include <cstddef>
#include <cstdint>
#include <iterator>
#include <list>
#include <numeric>
#include <utility>
#include <vector>

using namespace llvm;
using namespace llvm::bfi_detail;

#define DEBUG_TYPE "block-freq"

ScaledNumber<uint64_t> BlockMass::toScaled() const {
  if (isFull())
    return ScaledNumber<uint64_t>(1, 0);
  return ScaledNumber<uint64_t>(getMass() + 1, -64);
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
LLVM_DUMP_METHOD void BlockMass::dump() const { print(dbgs()); }
#endif

static char getHexDigit(int N) {
  assert(N < 16);
  if (N < 10)
    return '0' + N;
  return 'a' + N - 10;
}

raw_ostream &BlockMass::print(raw_ostream &OS) const {
  for (int Digits = 0; Digits < 16; ++Digits)
    OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
  return OS;
}

namespace {

using BlockNode = BlockFrequencyInfoImplBase::BlockNode;
using Distribution = BlockFrequencyInfoImplBase::Distribution;
using WeightList = BlockFrequencyInfoImplBase::Distribution::WeightList;
using Scaled64 = BlockFrequencyInfoImplBase::Scaled64;
using LoopData = BlockFrequencyInfoImplBase::LoopData;
using Weight = BlockFrequencyInfoImplBase::Weight;
using FrequencyData = BlockFrequencyInfoImplBase::FrequencyData;

/// \brief Dithering mass distributer.
///
/// This class splits up a single mass into portions by weight, dithering to
/// spread out error.  No mass is lost.  The dithering precision depends on the
/// precision of the product of \a BlockMass and \a BranchProbability.
///
/// The distribution algorithm follows.
///
///  1. Initialize by saving the sum of the weights in \a RemWeight and the
///     mass to distribute in \a RemMass.
///
///  2. For each portion:
///
///      1. Construct a branch probability, P, as the portion's weight divided
///         by the current value of \a RemWeight.
///      2. Calculate the portion's mass as \a RemMass times P.
///      3. Update \a RemWeight and \a RemMass at each portion by subtracting
///         the current portion's weight and mass.
struct DitheringDistributer {
  uint32_t RemWeight;
  BlockMass RemMass;

  DitheringDistributer(Distribution &Dist, const BlockMass &Mass);

  BlockMass takeMass(uint32_t Weight);
};

} // end anonymous namespace

DitheringDistributer::DitheringDistributer(Distribution &Dist,
                                           const BlockMass &Mass) {
  Dist.normalize();
  RemWeight = Dist.Total;
  RemMass = Mass;
}

BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
  assert(Weight && "invalid weight");
  assert(Weight <= RemWeight);
  BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);

  // Decrement totals (dither).
  RemWeight -= Weight;
  RemMass -= Mass;
  return Mass;
}

void Distribution::add(const BlockNode &Node, uint64_t Amount,
                       Weight::DistType Type) {
  assert(Amount && "invalid weight of 0");
  uint64_t NewTotal = Total + Amount;

  // Check for overflow.  It should be impossible to overflow twice.
  bool IsOverflow = NewTotal < Total;
  assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
  DidOverflow |= IsOverflow;

  // Update the total.
  Total = NewTotal;

  // Save the weight.
  Weights.push_back(Weight(Type, Node, Amount));
}

static void combineWeight(Weight &W, const Weight &OtherW) {
  assert(OtherW.TargetNode.isValid());
  if (!W.Amount) {
    W = OtherW;
    return;
  }
  assert(W.Type == OtherW.Type);
  assert(W.TargetNode == OtherW.TargetNode);
  assert(OtherW.Amount && "Expected non-zero weight");
  if (W.Amount > W.Amount + OtherW.Amount)
    // Saturate on overflow.
    W.Amount = UINT64_MAX;
  else
    W.Amount += OtherW.Amount;
}

static void combineWeightsBySorting(WeightList &Weights) {
  // Sort so edges to the same node are adjacent.
  std::sort(Weights.begin(), Weights.end(),
            [](const Weight &L,
               const Weight &R) { return L.TargetNode < R.TargetNode; });

  // Combine adjacent edges.
  WeightList::iterator O = Weights.begin();
  for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
       ++O, (I = L)) {
    *O = *I;

    // Find the adjacent weights to the same node.
    for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
      combineWeight(*O, *L);
  }

  // Erase extra entries.
  Weights.erase(O, Weights.end());
}

static void combineWeightsByHashing(WeightList &Weights) {
  // Collect weights into a DenseMap.
  using HashTable = DenseMap<BlockNode::IndexType, Weight>;

  HashTable Combined(NextPowerOf2(2 * Weights.size()));
  for (const Weight &W : Weights)
    combineWeight(Combined[W.TargetNode.Index], W);

  // Check whether anything changed.
  if (Weights.size() == Combined.size())
    return;

  // Fill in the new weights.
  Weights.clear();
  Weights.reserve(Combined.size());
  for (const auto &I : Combined)
    Weights.push_back(I.second);
}

static void combineWeights(WeightList &Weights) {
  // Use a hash table for many successors to keep this linear.
  if (Weights.size() > 128) {
    combineWeightsByHashing(Weights);
    return;
  }

  combineWeightsBySorting(Weights);
}

static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
  assert(Shift >= 0);
  assert(Shift < 64);
  if (!Shift)
    return N;
  return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
}

void Distribution::normalize() {
  // Early exit for termination nodes.
  if (Weights.empty())
    return;

  // Only bother if there are multiple successors.
  if (Weights.size() > 1)
    combineWeights(Weights);

  // Early exit when combined into a single successor.
  if (Weights.size() == 1) {
    Total = 1;
    Weights.front().Amount = 1;
    return;
  }

  // Determine how much to shift right so that the total fits into 32-bits.
  //
  // If we shift at all, shift by 1 extra.  Otherwise, the lower limit of 1
  // for each weight can cause a 32-bit overflow.
  int Shift = 0;
  if (DidOverflow)
    Shift = 33;
  else if (Total > UINT32_MAX)
    Shift = 33 - countLeadingZeros(Total);

  // Early exit if nothing needs to be scaled.
  if (!Shift) {
    // If we didn't overflow then combineWeights() shouldn't have changed the
    // sum of the weights, but let's double-check.
    assert(Total == std::accumulate(Weights.begin(), Weights.end(), UINT64_C(0),
                                    [](uint64_t Sum, const Weight &W) {
                      return Sum + W.Amount;
                    }) &&
           "Expected total to be correct");
    return;
  }

  // Recompute the total through accumulation (rather than shifting it) so that
  // it's accurate after shifting and any changes combineWeights() made above.
  Total = 0;

  // Sum the weights to each node and shift right if necessary.
  for (Weight &W : Weights) {
    // Scale down below UINT32_MAX.  Since Shift is larger than necessary, we
    // can round here without concern about overflow.
    assert(W.TargetNode.isValid());
    W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
    assert(W.Amount <= UINT32_MAX);

    // Update the total.
    Total += W.Amount;
  }
  assert(Total <= UINT32_MAX);
}

void BlockFrequencyInfoImplBase::clear() {
  // Swap with a default-constructed std::vector, since std::vector<>::clear()
  // does not actually clear heap storage.
  std::vector<FrequencyData>().swap(Freqs);
  std::vector<WorkingData>().swap(Working);
  Loops.clear();
}

/// \brief Clear all memory not needed downstream.
///
/// Releases all memory not used downstream.  In particular, saves Freqs.
static void cleanup(BlockFrequencyInfoImplBase &BFI) {
  std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
  BFI.clear();
  BFI.Freqs = std::move(SavedFreqs);
}

bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
                                           const LoopData *OuterLoop,
                                           const BlockNode &Pred,
                                           const BlockNode &Succ,
                                           uint64_t Weight) {
  if (!Weight)
    Weight = 1;

  auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
    return OuterLoop && OuterLoop->isHeader(Node);
  };

  BlockNode Resolved = Working[Succ.Index].getResolvedNode();

#ifndef NDEBUG
  auto debugSuccessor = [&](const char *Type) {
    dbgs() << "  =>"
           << " [" << Type << "] weight = " << Weight;
    if (!isLoopHeader(Resolved))
      dbgs() << ", succ = " << getBlockName(Succ);
    if (Resolved != Succ)
      dbgs() << ", resolved = " << getBlockName(Resolved);
    dbgs() << "\n";
  };
  (void)debugSuccessor;
#endif

  if (isLoopHeader(Resolved)) {
    DEBUG(debugSuccessor("backedge"));
    Dist.addBackedge(Resolved, Weight);
    return true;
  }

  if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
    DEBUG(debugSuccessor("  exit  "));
    Dist.addExit(Resolved, Weight);
    return true;
  }

  if (Resolved < Pred) {
    if (!isLoopHeader(Pred)) {
      // If OuterLoop is an irreducible loop, we can't actually handle this.
      assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
             "unhandled irreducible control flow");

      // Irreducible backedge.  Abort.
      DEBUG(debugSuccessor("abort!!!"));
      return false;
    }

    // If "Pred" is a loop header, then this isn't really a backedge; rather,
    // OuterLoop must be irreducible.  These false backedges can come only from
    // secondary loop headers.
    assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
           "unhandled irreducible control flow");
  }

  DEBUG(debugSuccessor(" local  "));
  Dist.addLocal(Resolved, Weight);
  return true;
}

bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
    const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
  // Copy the exit map into Dist.
  for (const auto &I : Loop.Exits)
    if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
                   I.second.getMass()))
      // Irreducible backedge.
      return false;

  return true;
}

/// \brief Compute the loop scale for a loop.
void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
  // Compute loop scale.
  DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");

  // Infinite loops need special handling. If we give the back edge an infinite
  // mass, they may saturate all the other scales in the function down to 1,
  // making all the other region temperatures look exactly the same. Choose an
  // arbitrary scale to avoid these issues.
  //
  // FIXME: An alternate way would be to select a symbolic scale which is later
  // replaced to be the maximum of all computed scales plus 1. This would
  // appropriately describe the loop as having a large scale, without skewing
  // the final frequency computation.
  const Scaled64 InfiniteLoopScale(1, 12);

  // LoopScale == 1 / ExitMass
  // ExitMass == HeadMass - BackedgeMass
  BlockMass TotalBackedgeMass;
  for (auto &Mass : Loop.BackedgeMass)
    TotalBackedgeMass += Mass;
  BlockMass ExitMass = BlockMass::getFull() - TotalBackedgeMass;

  // Block scale stores the inverse of the scale. If this is an infinite loop,
  // its exit mass will be zero. In this case, use an arbitrary scale for the
  // loop scale.
  Loop.Scale =
      ExitMass.isEmpty() ? InfiniteLoopScale : ExitMass.toScaled().inverse();

  DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
               << " - " << TotalBackedgeMass << ")\n"
               << " - scale = " << Loop.Scale << "\n");
}

/// \brief Package up a loop.
void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
  DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");

  // Clear the subloop exits to prevent quadratic memory usage.
  for (const BlockNode &M : Loop.Nodes) {
    if (auto *Loop = Working[M.Index].getPackagedLoop())
      Loop->Exits.clear();
    DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
  }
  Loop.IsPackaged = true;
}

#ifndef NDEBUG
static void debugAssign(const BlockFrequencyInfoImplBase &BFI,
                        const DitheringDistributer &D, const BlockNode &T,
                        const BlockMass &M, const char *Desc) {
  dbgs() << "  => assign " << M << " (" << D.RemMass << ")";
  if (Desc)
    dbgs() << " [" << Desc << "]";
  if (T.isValid())
    dbgs() << " to " << BFI.getBlockName(T);
  dbgs() << "\n";
}
#endif

void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
                                                LoopData *OuterLoop,
                                                Distribution &Dist) {
  BlockMass Mass = Working[Source.Index].getMass();
  DEBUG(dbgs() << "  => mass:  " << Mass << "\n");

  // Distribute mass to successors as laid out in Dist.
  DitheringDistributer D(Dist, Mass);

  for (const Weight &W : Dist.Weights) {
    // Check for a local edge (non-backedge and non-exit).
    BlockMass Taken = D.takeMass(W.Amount);
    if (W.Type == Weight::Local) {
      Working[W.TargetNode.Index].getMass() += Taken;
      DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
      continue;
    }

    // Backedges and exits only make sense if we're processing a loop.
    assert(OuterLoop && "backedge or exit outside of loop");

    // Check for a backedge.
    if (W.Type == Weight::Backedge) {
      OuterLoop->BackedgeMass[OuterLoop->getHeaderIndex(W.TargetNode)] += Taken;
      DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "back"));
      continue;
    }

    // This must be an exit.
    assert(W.Type == Weight::Exit);
    OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
    DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "exit"));
  }
}

static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
                                     const Scaled64 &Min, const Scaled64 &Max) {
  // Scale the Factor to a size that creates integers.  Ideally, integers would
  // be scaled so that Max == UINT64_MAX so that they can be best
  // differentiated.  However, in the presence of large frequency values, small
  // frequencies are scaled down to 1, making it impossible to differentiate
  // small, unequal numbers. When the spread between Min and Max frequencies
  // fits well within MaxBits, we make the scale be at least 8.
  const unsigned MaxBits = 64;
  const unsigned SpreadBits = (Max / Min).lg();
  Scaled64 ScalingFactor;
  if (SpreadBits <= MaxBits - 3) {
    // If the values are small enough, make the scaling factor at least 8 to
    // allow distinguishing small values.
    ScalingFactor = Min.inverse();
    ScalingFactor <<= 3;
  } else {
    // If the values need more than MaxBits to be represented, saturate small
    // frequency values down to 1 by using a scaling factor that benefits large
    // frequency values.
    ScalingFactor = Scaled64(1, MaxBits) / Max;
  }

  // Translate the floats to integers.
  DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
               << ", factor = " << ScalingFactor << "\n");
  for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
    Scaled64 Scaled = BFI.Freqs[Index].Scaled * ScalingFactor;
    BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
    DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
                 << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
                 << ", int = " << BFI.Freqs[Index].Integer << "\n");
  }
}

/// \brief Unwrap a loop package.
///
/// Visits all the members of a loop, adjusting their BlockData according to
/// the loop's pseudo-node.
static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
  DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
               << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
               << "\n");
  Loop.Scale *= Loop.Mass.toScaled();
  Loop.IsPackaged = false;
  DEBUG(dbgs() << "  => combined-scale = " << Loop.Scale << "\n");

  // Propagate the head scale through the loop.  Since members are visited in
  // RPO, the head scale will be updated by the loop scale first, and then the
  // final head scale will be used for updated the rest of the members.
  for (const BlockNode &N : Loop.Nodes) {
    const auto &Working = BFI.Working[N.Index];
    Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
                                       : BFI.Freqs[N.Index].Scaled;
    Scaled64 New = Loop.Scale * F;
    DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => " << New
                 << "\n");
    F = New;
  }
}

void BlockFrequencyInfoImplBase::unwrapLoops() {
  // Set initial frequencies from loop-local masses.
  for (size_t Index = 0; Index < Working.size(); ++Index)
    Freqs[Index].Scaled = Working[Index].Mass.toScaled();

  for (LoopData &Loop : Loops)
    unwrapLoop(*this, Loop);
}

void BlockFrequencyInfoImplBase::finalizeMetrics() {
  // Unwrap loop packages in reverse post-order, tracking min and max
  // frequencies.
  auto Min = Scaled64::getLargest();
  auto Max = Scaled64::getZero();
  for (size_t Index = 0; Index < Working.size(); ++Index) {
    // Update min/max scale.
    Min = std::min(Min, Freqs[Index].Scaled);
    Max = std::max(Max, Freqs[Index].Scaled);
  }

  // Convert to integers.
  convertFloatingToInteger(*this, Min, Max);

  // Clean up data structures.
  cleanup(*this);

  // Print out the final stats.
  DEBUG(dump());
}

BlockFrequency
BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
  if (!Node.isValid())
    return 0;
  return Freqs[Node.Index].Integer;
}

Optional<uint64_t>
BlockFrequencyInfoImplBase::getBlockProfileCount(const Function &F,
                                                 const BlockNode &Node) const {
  return getProfileCountFromFreq(F, getBlockFreq(Node).getFrequency());
}

Optional<uint64_t>
BlockFrequencyInfoImplBase::getProfileCountFromFreq(const Function &F,
                                                    uint64_t Freq) const {
  auto EntryCount = F.getEntryCount();
  if (!EntryCount)
    return None;
  // Use 128 bit APInt to do the arithmetic to avoid overflow.
  APInt BlockCount(128, EntryCount.getValue());
  APInt BlockFreq(128, Freq);
  APInt EntryFreq(128, getEntryFreq());
  BlockCount *= BlockFreq;
  BlockCount = BlockCount.udiv(EntryFreq);
  return BlockCount.getLimitedValue();
}

Scaled64
BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
  if (!Node.isValid())
    return Scaled64::getZero();
  return Freqs[Node.Index].Scaled;
}

void BlockFrequencyInfoImplBase::setBlockFreq(const BlockNode &Node,
                                              uint64_t Freq) {
  assert(Node.isValid() && "Expected valid node");
  assert(Node.Index < Freqs.size() && "Expected legal index");
  Freqs[Node.Index].Integer = Freq;
}

std::string
BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
  return {};
}

std::string
BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
  return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
}

raw_ostream &
BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
                                           const BlockNode &Node) const {
  return OS << getFloatingBlockFreq(Node);
}

raw_ostream &
BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
                                           const BlockFrequency &Freq) const {
  Scaled64 Block(Freq.getFrequency(), 0);
  Scaled64 Entry(getEntryFreq(), 0);

  return OS << Block / Entry;
}

void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
  Start = OuterLoop.getHeader();
  Nodes.reserve(OuterLoop.Nodes.size());
  for (auto N : OuterLoop.Nodes)
    addNode(N);
  indexNodes();
}

void IrreducibleGraph::addNodesInFunction() {
  Start = 0;
  for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
    if (!BFI.Working[Index].isPackaged())
      addNode(Index);
  indexNodes();
}

void IrreducibleGraph::indexNodes() {
  for (auto &I : Nodes)
    Lookup[I.Node.Index] = &I;
}

void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
                               const BFIBase::LoopData *OuterLoop) {
  if (OuterLoop && OuterLoop->isHeader(Succ))
    return;
  auto L = Lookup.find(Succ.Index);
  if (L == Lookup.end())
    return;
  IrrNode &SuccIrr = *L->second;
  Irr.Edges.push_back(&SuccIrr);
  SuccIrr.Edges.push_front(&Irr);
  ++SuccIrr.NumIn;
}

namespace llvm {

template <> struct GraphTraits<IrreducibleGraph> {
  using GraphT = bfi_detail::IrreducibleGraph;
  using NodeRef = const GraphT::IrrNode *;
  using ChildIteratorType = GraphT::IrrNode::iterator;

  static NodeRef getEntryNode(const GraphT &G) { return G.StartIrr; }
  static ChildIteratorType child_begin(NodeRef N) { return N->succ_begin(); }
  static ChildIteratorType child_end(NodeRef N) { return N->succ_end(); }
};

} // end namespace llvm

/// \brief Find extra irreducible headers.
///
/// Find entry blocks and other blocks with backedges, which exist when \c G
/// contains irreducible sub-SCCs.
static void findIrreducibleHeaders(
    const BlockFrequencyInfoImplBase &BFI,
    const IrreducibleGraph &G,
    const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
    LoopData::NodeList &Headers, LoopData::NodeList &Others) {
  // Map from nodes in the SCC to whether it's an entry block.
  SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;

  // InSCC also acts the set of nodes in the graph.  Seed it.
  for (const auto *I : SCC)
    InSCC[I] = false;

  for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
    auto &Irr = *I->first;
    for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
      if (InSCC.count(P))
        continue;

      // This is an entry block.
      I->second = true;
      Headers.push_back(Irr.Node);
      DEBUG(dbgs() << "  => entry = " << BFI.getBlockName(Irr.Node) << "\n");
      break;
    }
  }
  assert(Headers.size() >= 2 &&
         "Expected irreducible CFG; -loop-info is likely invalid");
  if (Headers.size() == InSCC.size()) {
    // Every block is a header.
    std::sort(Headers.begin(), Headers.end());
    return;
  }

  // Look for extra headers from irreducible sub-SCCs.
  for (const auto &I : InSCC) {
    // Entry blocks are already headers.
    if (I.second)
      continue;

    auto &Irr = *I.first;
    for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
      // Skip forward edges.
      if (P->Node < Irr.Node)
        continue;

      // Skip predecessors from entry blocks.  These can have inverted
      // ordering.
      if (InSCC.lookup(P))
        continue;

      // Store the extra header.
      Headers.push_back(Irr.Node);
      DEBUG(dbgs() << "  => extra = " << BFI.getBlockName(Irr.Node) << "\n");
      break;
    }
    if (Headers.back() == Irr.Node)
      // Added this as a header.
      continue;

    // This is not a header.
    Others.push_back(Irr.Node);
    DEBUG(dbgs() << "  => other = " << BFI.getBlockName(Irr.Node) << "\n");
  }
  std::sort(Headers.begin(), Headers.end());
  std::sort(Others.begin(), Others.end());
}

static void createIrreducibleLoop(
    BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
    LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
    const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
  // Translate the SCC into RPO.
  DEBUG(dbgs() << " - found-scc\n");

  LoopData::NodeList Headers;
  LoopData::NodeList Others;
  findIrreducibleHeaders(BFI, G, SCC, Headers, Others);

  auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
                                Headers.end(), Others.begin(), Others.end());

  // Update loop hierarchy.
  for (const auto &N : Loop->Nodes)
    if (BFI.Working[N.Index].isLoopHeader())
      BFI.Working[N.Index].Loop->Parent = &*Loop;
    else
      BFI.Working[N.Index].Loop = &*Loop;
}

iterator_range<std::list<LoopData>::iterator>
BlockFrequencyInfoImplBase::analyzeIrreducible(
    const IrreducibleGraph &G, LoopData *OuterLoop,
    std::list<LoopData>::iterator Insert) {
  assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
  auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();

  for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
    if (I->size() < 2)
      continue;

    // Translate the SCC into RPO.
    createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
  }

  if (OuterLoop)
    return make_range(std::next(Prev), Insert);
  return make_range(Loops.begin(), Insert);
}

void
BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
  OuterLoop.Exits.clear();
  for (auto &Mass : OuterLoop.BackedgeMass)
    Mass = BlockMass::getEmpty();
  auto O = OuterLoop.Nodes.begin() + 1;
  for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
    if (!Working[I->Index].isPackaged())
      *O++ = *I;
  OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
}

void BlockFrequencyInfoImplBase::adjustLoopHeaderMass(LoopData &Loop) {
  assert(Loop.isIrreducible() && "this only makes sense on irreducible loops");

  // Since the loop has more than one header block, the mass flowing back into
  // each header will be different. Adjust the mass in each header loop to
  // reflect the masses flowing through back edges.
  //
  // To do this, we distribute the initial mass using the backedge masses
  // as weights for the distribution.
  BlockMass LoopMass = BlockMass::getFull();
  Distribution Dist;

  DEBUG(dbgs() << "adjust-loop-header-mass:\n");
  for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
    auto &HeaderNode = Loop.Nodes[H];
    auto &BackedgeMass = Loop.BackedgeMass[Loop.getHeaderIndex(HeaderNode)];
    DEBUG(dbgs() << " - Add back edge mass for node "
                 << getBlockName(HeaderNode) << ": " << BackedgeMass << "\n");
    if (BackedgeMass.getMass() > 0)
      Dist.addLocal(HeaderNode, BackedgeMass.getMass());
    else
      DEBUG(dbgs() << "   Nothing added. Back edge mass is zero\n");
  }

  DitheringDistributer D(Dist, LoopMass);

  DEBUG(dbgs() << " Distribute loop mass " << LoopMass
               << " to headers using above weights\n");
  for (const Weight &W : Dist.Weights) {
    BlockMass Taken = D.takeMass(W.Amount);
    assert(W.Type == Weight::Local && "all weights should be local");
    Working[W.TargetNode.Index].getMass() = Taken;
    DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
  }
}