lib/Analysis/BlockFrequencyInfoImpl.cpp

   1 //===- BlockFrequencyImplInfo.cpp - Block Frequency Info Implementation ---===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // Loops should be simplified before this analysis.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
  15 #include "llvm/ADT/APInt.h"
  16 #include "llvm/ADT/DenseMap.h"
  17 #include "llvm/ADT/GraphTraits.h"
  18 #include "llvm/ADT/None.h"
  19 #include "llvm/ADT/SCCIterator.h"
  20 #include "llvm/Config/llvm-config.h"
  21 #include "llvm/IR/Function.h"
  22 #include "llvm/Support/BlockFrequency.h"
  23 #include "llvm/Support/BranchProbability.h"
  24 #include "llvm/Support/Compiler.h"
  25 #include "llvm/Support/Debug.h"
  26 #include "llvm/Support/ScaledNumber.h"
  27 #include "llvm/Support/MathExtras.h"
  28 #include "llvm/Support/raw_ostream.h"
  29 #include <algorithm>
  30 #include <cassert>
  31 #include <cstddef>
  32 #include <cstdint>
  33 #include <iterator>
  34 #include <list>
  35 #include <numeric>
  36 #include <utility>
  37 #include <vector>
  38
  39 using namespace llvm;
  40 using namespace llvm::bfi_detail;
  41
  42 #define DEBUG_TYPE "block-freq"
  43
  44 ScaledNumber<uint64_t> BlockMass::toScaled() const {
  45   if (isFull())
  46     return ScaledNumber<uint64_t>(1, 0);
  47   return ScaledNumber<uint64_t>(getMass() + 1, -64);
  48 }
  49
  50 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
  51 LLVM_DUMP_METHOD void BlockMass::dump() const { print(dbgs()); }
  52 #endif
  53
  54 static char getHexDigit(int N) {
  55   assert(N < 16);
  56   if (N < 10)
  57     return '0' + N;
  58   return 'a' + N - 10;
  59 }
  60
  61 raw_ostream &BlockMass::print(raw_ostream &OS) const {
  62   for (int Digits = 0; Digits < 16; ++Digits)
  63     OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
  64   return OS;
  65 }
  66
  67 namespace {
  68
  69 using BlockNode = BlockFrequencyInfoImplBase::BlockNode;
  70 using Distribution = BlockFrequencyInfoImplBase::Distribution;
  71 using WeightList = BlockFrequencyInfoImplBase::Distribution::WeightList;
  72 using Scaled64 = BlockFrequencyInfoImplBase::Scaled64;
  73 using LoopData = BlockFrequencyInfoImplBase::LoopData;
  74 using Weight = BlockFrequencyInfoImplBase::Weight;
  75 using FrequencyData = BlockFrequencyInfoImplBase::FrequencyData;
  76
  77 /// Dithering mass distributer.
  78 ///
  79 /// This class splits up a single mass into portions by weight, dithering to
  80 /// spread out error.  No mass is lost.  The dithering precision depends on the
  81 /// precision of the product of \a BlockMass and \a BranchProbability.
  82 ///
  83 /// The distribution algorithm follows.
  84 ///
  85 ///  1. Initialize by saving the sum of the weights in \a RemWeight and the
  86 ///     mass to distribute in \a RemMass.
  87 ///
  88 ///  2. For each portion:
  89 ///
  90 ///      1. Construct a branch probability, P, as the portion's weight divided
  91 ///         by the current value of \a RemWeight.
  92 ///      2. Calculate the portion's mass as \a RemMass times P.
  93 ///      3. Update \a RemWeight and \a RemMass at each portion by subtracting
  94 ///         the current portion's weight and mass.
  95 struct DitheringDistributer {
  96   uint32_t RemWeight;
  97   BlockMass RemMass;
  98
  99   DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
 100
 101   BlockMass takeMass(uint32_t Weight);
 102 };
 103
 104 } // end anonymous namespace
 105
 106 DitheringDistributer::DitheringDistributer(Distribution &Dist,
 107                                            const BlockMass &Mass) {
 108   Dist.normalize();
 109   RemWeight = Dist.Total;
 110   RemMass = Mass;
 111 }
 112
 113 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
 114   assert(Weight && "invalid weight");
 115   assert(Weight <= RemWeight);
 116   BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
 117
 118   // Decrement totals (dither).
 119   RemWeight -= Weight;
 120   RemMass -= Mass;
 121   return Mass;
 122 }
 123
 124 void Distribution::add(const BlockNode &Node, uint64_t Amount,
 125                        Weight::DistType Type) {
 126   assert(Amount && "invalid weight of 0");
 127   uint64_t NewTotal = Total + Amount;
 128
 129   // Check for overflow.  It should be impossible to overflow twice.
 130   bool IsOverflow = NewTotal < Total;
 131   assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
 132   DidOverflow |= IsOverflow;
 133
 134   // Update the total.
 135   Total = NewTotal;
 136
 137   // Save the weight.
 138   Weights.push_back(Weight(Type, Node, Amount));
 139 }
 140
 141 static void combineWeight(Weight &W, const Weight &OtherW) {
 142   assert(OtherW.TargetNode.isValid());
 143   if (!W.Amount) {
 144     W = OtherW;
 145     return;
 146   }
 147   assert(W.Type == OtherW.Type);
 148   assert(W.TargetNode == OtherW.TargetNode);
 149   assert(OtherW.Amount && "Expected non-zero weight");
 150   if (W.Amount > W.Amount + OtherW.Amount)
 151     // Saturate on overflow.
 152     W.Amount = UINT64_MAX;
 153   else
 154     W.Amount += OtherW.Amount;
 155 }
 156
 157 static void combineWeightsBySorting(WeightList &Weights) {
 158   // Sort so edges to the same node are adjacent.
 159   llvm::sort(Weights, [](const Weight &L, const Weight &R) {
 160     return L.TargetNode < R.TargetNode;
 161   });
 162
 163   // Combine adjacent edges.
 164   WeightList::iterator O = Weights.begin();
 165   for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
 166        ++O, (I = L)) {
 167     *O = *I;
 168
 169     // Find the adjacent weights to the same node.
 170     for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
 171       combineWeight(*O, *L);
 172   }
 173
 174   // Erase extra entries.
 175   Weights.erase(O, Weights.end());
 176 }
 177
 178 static void combineWeightsByHashing(WeightList &Weights) {
 179   // Collect weights into a DenseMap.
 180   using HashTable = DenseMap<BlockNode::IndexType, Weight>;
 181
 182   HashTable Combined(NextPowerOf2(2 * Weights.size()));
 183   for (const Weight &W : Weights)
 184     combineWeight(Combined[W.TargetNode.Index], W);
 185
 186   // Check whether anything changed.
 187   if (Weights.size() == Combined.size())
 188     return;
 189
 190   // Fill in the new weights.
 191   Weights.clear();
 192   Weights.reserve(Combined.size());
 193   for (const auto &I : Combined)
 194     Weights.push_back(I.second);
 195 }
 196
 197 static void combineWeights(WeightList &Weights) {
 198   // Use a hash table for many successors to keep this linear.
 199   if (Weights.size() > 128) {
 200     combineWeightsByHashing(Weights);
 201     return;
 202   }
 203
 204   combineWeightsBySorting(Weights);
 205 }
 206
 207 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
 208   assert(Shift >= 0);
 209   assert(Shift < 64);
 210   if (!Shift)
 211     return N;
 212   return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
 213 }
 214
 215 void Distribution::normalize() {
 216   // Early exit for termination nodes.
 217   if (Weights.empty())
 218     return;
 219
 220   // Only bother if there are multiple successors.
 221   if (Weights.size() > 1)
 222     combineWeights(Weights);
 223
 224   // Early exit when combined into a single successor.
 225   if (Weights.size() == 1) {
 226     Total = 1;
 227     Weights.front().Amount = 1;
 228     return;
 229   }
 230
 231   // Determine how much to shift right so that the total fits into 32-bits.
 232   //
 233   // If we shift at all, shift by 1 extra.  Otherwise, the lower limit of 1
 234   // for each weight can cause a 32-bit overflow.
 235   int Shift = 0;
 236   if (DidOverflow)
 237     Shift = 33;
 238   else if (Total > UINT32_MAX)
 239     Shift = 33 - countLeadingZeros(Total);
 240
 241   // Early exit if nothing needs to be scaled.
 242   if (!Shift) {
 243     // If we didn't overflow then combineWeights() shouldn't have changed the
 244     // sum of the weights, but let's double-check.
 245     assert(Total == std::accumulate(Weights.begin(), Weights.end(), UINT64_C(0),
 246                                     [](uint64_t Sum, const Weight &W) {
 247                       return Sum + W.Amount;
 248                     }) &&
 249            "Expected total to be correct");
 250     return;
 251   }
 252
 253   // Recompute the total through accumulation (rather than shifting it) so that
 254   // it's accurate after shifting and any changes combineWeights() made above.
 255   Total = 0;
 256
 257   // Sum the weights to each node and shift right if necessary.
 258   for (Weight &W : Weights) {
 259     // Scale down below UINT32_MAX.  Since Shift is larger than necessary, we
 260     // can round here without concern about overflow.
 261     assert(W.TargetNode.isValid());
 262     W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
 263     assert(W.Amount <= UINT32_MAX);
 264
 265     // Update the total.
 266     Total += W.Amount;
 267   }
 268   assert(Total <= UINT32_MAX);
 269 }
 270
 271 void BlockFrequencyInfoImplBase::clear() {
 272   // Swap with a default-constructed std::vector, since std::vector<>::clear()
 273   // does not actually clear heap storage.
 274   std::vector<FrequencyData>().swap(Freqs);
 275   IsIrrLoopHeader.clear();
 276   std::vector<WorkingData>().swap(Working);
 277   Loops.clear();
 278 }
 279
 280 /// Clear all memory not needed downstream.
 281 ///
 282 /// Releases all memory not used downstream.  In particular, saves Freqs.
 283 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
 284   std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
 285   SparseBitVector<> SavedIsIrrLoopHeader(std::move(BFI.IsIrrLoopHeader));
 286   BFI.clear();
 287   BFI.Freqs = std::move(SavedFreqs);
 288   BFI.IsIrrLoopHeader = std::move(SavedIsIrrLoopHeader);
 289 }
 290
 291 bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
 292                                            const LoopData *OuterLoop,
 293                                            const BlockNode &Pred,
 294                                            const BlockNode &Succ,
 295                                            uint64_t Weight) {
 296   if (!Weight)
 297     Weight = 1;
 298
 299   auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
 300     return OuterLoop && OuterLoop->isHeader(Node);
 301   };
 302
 303   BlockNode Resolved = Working[Succ.Index].getResolvedNode();
 304
 305 #ifndef NDEBUG
 306   auto debugSuccessor = [&](const char *Type) {
 307     dbgs() << "  =>"
 308            << " [" << Type << "] weight = " << Weight;
 309     if (!isLoopHeader(Resolved))
 310       dbgs() << ", succ = " << getBlockName(Succ);
 311     if (Resolved != Succ)
 312       dbgs() << ", resolved = " << getBlockName(Resolved);
 313     dbgs() << "\n";
 314   };
 315   (void)debugSuccessor;
 316 #endif
 317
 318   if (isLoopHeader(Resolved)) {
 319     LLVM_DEBUG(debugSuccessor("backedge"));
 320     Dist.addBackedge(Resolved, Weight);
 321     return true;
 322   }
 323
 324   if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
 325     LLVM_DEBUG(debugSuccessor("  exit  "));
 326     Dist.addExit(Resolved, Weight);
 327     return true;
 328   }
 329
 330   if (Resolved < Pred) {
 331     if (!isLoopHeader(Pred)) {
 332       // If OuterLoop is an irreducible loop, we can't actually handle this.
 333       assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
 334              "unhandled irreducible control flow");
 335
 336       // Irreducible backedge.  Abort.
 337       LLVM_DEBUG(debugSuccessor("abort!!!"));
 338       return false;
 339     }
 340
 341     // If "Pred" is a loop header, then this isn't really a backedge; rather,
 342     // OuterLoop must be irreducible.  These false backedges can come only from
 343     // secondary loop headers.
 344     assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
 345            "unhandled irreducible control flow");
 346   }
 347
 348   LLVM_DEBUG(debugSuccessor(" local  "));
 349   Dist.addLocal(Resolved, Weight);
 350   return true;
 351 }
 352
 353 bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
 354     const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
 355   // Copy the exit map into Dist.
 356   for (const auto &I : Loop.Exits)
 357     if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
 358                    I.second.getMass()))
 359       // Irreducible backedge.
 360       return false;
 361
 362   return true;
 363 }
 364
 365 /// Compute the loop scale for a loop.
 366 void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
 367   // Compute loop scale.
 368   LLVM_DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
 369
 370   // Infinite loops need special handling. If we give the back edge an infinite
 371   // mass, they may saturate all the other scales in the function down to 1,
 372   // making all the other region temperatures look exactly the same. Choose an
 373   // arbitrary scale to avoid these issues.
 374   //
 375   // FIXME: An alternate way would be to select a symbolic scale which is later
 376   // replaced to be the maximum of all computed scales plus 1. This would
 377   // appropriately describe the loop as having a large scale, without skewing
 378   // the final frequency computation.
 379   const Scaled64 InfiniteLoopScale(1, 12);
 380
 381   // LoopScale == 1 / ExitMass
 382   // ExitMass == HeadMass - BackedgeMass
 383   BlockMass TotalBackedgeMass;
 384   for (auto &Mass : Loop.BackedgeMass)
 385     TotalBackedgeMass += Mass;
 386   BlockMass ExitMass = BlockMass::getFull() - TotalBackedgeMass;
 387
 388   // Block scale stores the inverse of the scale. If this is an infinite loop,
 389   // its exit mass will be zero. In this case, use an arbitrary scale for the
 390   // loop scale.
 391   Loop.Scale =
 392       ExitMass.isEmpty() ? InfiniteLoopScale : ExitMass.toScaled().inverse();
 393
 394   LLVM_DEBUG(dbgs() << " - exit-mass = " << ExitMass << " ("
 395                     << BlockMass::getFull() << " - " << TotalBackedgeMass
 396                     << ")\n"
 397                     << " - scale = " << Loop.Scale << "\n");
 398 }
 399
 400 /// Package up a loop.
 401 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
 402   LLVM_DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
 403
 404   // Clear the subloop exits to prevent quadratic memory usage.
 405   for (const BlockNode &M : Loop.Nodes) {
 406     if (auto *Loop = Working[M.Index].getPackagedLoop())
 407       Loop->Exits.clear();
 408     LLVM_DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
 409   }
 410   Loop.IsPackaged = true;
 411 }
 412
 413 #ifndef NDEBUG
 414 static void debugAssign(const BlockFrequencyInfoImplBase &BFI,
 415                         const DitheringDistributer &D, const BlockNode &T,
 416                         const BlockMass &M, const char *Desc) {
 417   dbgs() << "  => assign " << M << " (" << D.RemMass << ")";
 418   if (Desc)
 419     dbgs() << " [" << Desc << "]";
 420   if (T.isValid())
 421     dbgs() << " to " << BFI.getBlockName(T);
 422   dbgs() << "\n";
 423 }
 424 #endif
 425
 426 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
 427                                                 LoopData *OuterLoop,
 428                                                 Distribution &Dist) {
 429   BlockMass Mass = Working[Source.Index].getMass();
 430   LLVM_DEBUG(dbgs() << "  => mass:  " << Mass << "\n");
 431
 432   // Distribute mass to successors as laid out in Dist.
 433   DitheringDistributer D(Dist, Mass);
 434
 435   for (const Weight &W : Dist.Weights) {
 436     // Check for a local edge (non-backedge and non-exit).
 437     BlockMass Taken = D.takeMass(W.Amount);
 438     if (W.Type == Weight::Local) {
 439       Working[W.TargetNode.Index].getMass() += Taken;
 440       LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
 441       continue;
 442     }
 443
 444     // Backedges and exits only make sense if we're processing a loop.
 445     assert(OuterLoop && "backedge or exit outside of loop");
 446
 447     // Check for a backedge.
 448     if (W.Type == Weight::Backedge) {
 449       OuterLoop->BackedgeMass[OuterLoop->getHeaderIndex(W.TargetNode)] += Taken;
 450       LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "back"));
 451       continue;
 452     }
 453
 454     // This must be an exit.
 455     assert(W.Type == Weight::Exit);
 456     OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
 457     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "exit"));
 458   }
 459 }
 460
 461 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
 462                                      const Scaled64 &Min, const Scaled64 &Max) {
 463   // Scale the Factor to a size that creates integers.  Ideally, integers would
 464   // be scaled so that Max == UINT64_MAX so that they can be best
 465   // differentiated.  However, in the presence of large frequency values, small
 466   // frequencies are scaled down to 1, making it impossible to differentiate
 467   // small, unequal numbers. When the spread between Min and Max frequencies
 468   // fits well within MaxBits, we make the scale be at least 8.
 469   const unsigned MaxBits = 64;
 470   const unsigned SpreadBits = (Max / Min).lg();
 471   Scaled64 ScalingFactor;
 472   if (SpreadBits <= MaxBits - 3) {
 473     // If the values are small enough, make the scaling factor at least 8 to
 474     // allow distinguishing small values.
 475     ScalingFactor = Min.inverse();
 476     ScalingFactor <<= 3;
 477   } else {
 478     // If the values need more than MaxBits to be represented, saturate small
 479     // frequency values down to 1 by using a scaling factor that benefits large
 480     // frequency values.
 481     ScalingFactor = Scaled64(1, MaxBits) / Max;
 482   }
 483
 484   // Translate the floats to integers.
 485   LLVM_DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
 486                     << ", factor = " << ScalingFactor << "\n");
 487   for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
 488     Scaled64 Scaled = BFI.Freqs[Index].Scaled * ScalingFactor;
 489     BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
 490     LLVM_DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
 491                       << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
 492                       << ", int = " << BFI.Freqs[Index].Integer << "\n");
 493   }
 494 }
 495
 496 /// Unwrap a loop package.
 497 ///
 498 /// Visits all the members of a loop, adjusting their BlockData according to
 499 /// the loop's pseudo-node.
 500 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
 501   LLVM_DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
 502                     << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
 503                     << "\n");
 504   Loop.Scale *= Loop.Mass.toScaled();
 505   Loop.IsPackaged = false;
 506   LLVM_DEBUG(dbgs() << "  => combined-scale = " << Loop.Scale << "\n");
 507
 508   // Propagate the head scale through the loop.  Since members are visited in
 509   // RPO, the head scale will be updated by the loop scale first, and then the
 510   // final head scale will be used for updated the rest of the members.
 511   for (const BlockNode &N : Loop.Nodes) {
 512     const auto &Working = BFI.Working[N.Index];
 513     Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
 514                                        : BFI.Freqs[N.Index].Scaled;
 515     Scaled64 New = Loop.Scale * F;
 516     LLVM_DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => "
 517                       << New << "\n");
 518     F = New;
 519   }
 520 }
 521
 522 void BlockFrequencyInfoImplBase::unwrapLoops() {
 523   // Set initial frequencies from loop-local masses.
 524   for (size_t Index = 0; Index < Working.size(); ++Index)
 525     Freqs[Index].Scaled = Working[Index].Mass.toScaled();
 526
 527   for (LoopData &Loop : Loops)
 528     unwrapLoop(*this, Loop);
 529 }
 530
 531 void BlockFrequencyInfoImplBase::finalizeMetrics() {
 532   // Unwrap loop packages in reverse post-order, tracking min and max
 533   // frequencies.
 534   auto Min = Scaled64::getLargest();
 535   auto Max = Scaled64::getZero();
 536   for (size_t Index = 0; Index < Working.size(); ++Index) {
 537     // Update min/max scale.
 538     Min = std::min(Min, Freqs[Index].Scaled);
 539     Max = std::max(Max, Freqs[Index].Scaled);
 540   }
 541
 542   // Convert to integers.
 543   convertFloatingToInteger(*this, Min, Max);
 544
 545   // Clean up data structures.
 546   cleanup(*this);
 547
 548   // Print out the final stats.
 549   LLVM_DEBUG(dump());
 550 }
 551
 552 BlockFrequency
 553 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
 554   if (!Node.isValid())
 555     return 0;
 556   return Freqs[Node.Index].Integer;
 557 }
 558
 559 Optional<uint64_t>
 560 BlockFrequencyInfoImplBase::getBlockProfileCount(const Function &F,
 561                                                  const BlockNode &Node) const {
 562   return getProfileCountFromFreq(F, getBlockFreq(Node).getFrequency());
 563 }
 564
 565 Optional<uint64_t>
 566 BlockFrequencyInfoImplBase::getProfileCountFromFreq(const Function &F,
 567                                                     uint64_t Freq) const {
 568   auto EntryCount = F.getEntryCount();
 569   if (!EntryCount)
 570     return None;
 571   // Use 128 bit APInt to do the arithmetic to avoid overflow.
 572   APInt BlockCount(128, EntryCount.getCount());
 573   APInt BlockFreq(128, Freq);
 574   APInt EntryFreq(128, getEntryFreq());
 575   BlockCount *= BlockFreq;
 576   // Rounded division of BlockCount by EntryFreq. Since EntryFreq is unsigned
 577   // lshr by 1 gives EntryFreq/2.
 578   BlockCount = (BlockCount + EntryFreq.lshr(1)).udiv(EntryFreq);
 579   return BlockCount.getLimitedValue();
 580 }
 581
 582 bool
 583 BlockFrequencyInfoImplBase::isIrrLoopHeader(const BlockNode &Node) {
 584   if (!Node.isValid())
 585     return false;
 586   return IsIrrLoopHeader.test(Node.Index);
 587 }
 588
 589 Scaled64
 590 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
 591   if (!Node.isValid())
 592     return Scaled64::getZero();
 593   return Freqs[Node.Index].Scaled;
 594 }
 595
 596 void BlockFrequencyInfoImplBase::setBlockFreq(const BlockNode &Node,
 597                                               uint64_t Freq) {
 598   assert(Node.isValid() && "Expected valid node");
 599   assert(Node.Index < Freqs.size() && "Expected legal index");
 600   Freqs[Node.Index].Integer = Freq;
 601 }
 602
 603 std::string
 604 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
 605   return {};
 606 }
 607
 608 std::string
 609 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
 610   return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
 611 }
 612
 613 raw_ostream &
 614 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
 615                                            const BlockNode &Node) const {
 616   return OS << getFloatingBlockFreq(Node);
 617 }
 618
 619 raw_ostream &
 620 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
 621                                            const BlockFrequency &Freq) const {
 622   Scaled64 Block(Freq.getFrequency(), 0);
 623   Scaled64 Entry(getEntryFreq(), 0);
 624
 625   return OS << Block / Entry;
 626 }
 627
 628 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
 629   Start = OuterLoop.getHeader();
 630   Nodes.reserve(OuterLoop.Nodes.size());
 631   for (auto N : OuterLoop.Nodes)
 632     addNode(N);
 633   indexNodes();
 634 }
 635
 636 void IrreducibleGraph::addNodesInFunction() {
 637   Start = 0;
 638   for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
 639     if (!BFI.Working[Index].isPackaged())
 640       addNode(Index);
 641   indexNodes();
 642 }
 643
 644 void IrreducibleGraph::indexNodes() {
 645   for (auto &I : Nodes)
 646     Lookup[I.Node.Index] = &I;
 647 }
 648
 649 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
 650                                const BFIBase::LoopData *OuterLoop) {
 651   if (OuterLoop && OuterLoop->isHeader(Succ))
 652     return;
 653   auto L = Lookup.find(Succ.Index);
 654   if (L == Lookup.end())
 655     return;
 656   IrrNode &SuccIrr = *L->second;
 657   Irr.Edges.push_back(&SuccIrr);
 658   SuccIrr.Edges.push_front(&Irr);
 659   ++SuccIrr.NumIn;
 660 }
 661
 662 namespace llvm {
 663
 664 template <> struct GraphTraits<IrreducibleGraph> {
 665   using GraphT = bfi_detail::IrreducibleGraph;
 666   using NodeRef = const GraphT::IrrNode *;
 667   using ChildIteratorType = GraphT::IrrNode::iterator;
 668
 669   static NodeRef getEntryNode(const GraphT &G) { return G.StartIrr; }
 670   static ChildIteratorType child_begin(NodeRef N) { return N->succ_begin(); }
 671   static ChildIteratorType child_end(NodeRef N) { return N->succ_end(); }
 672 };
 673
 674 } // end namespace llvm
 675
 676 /// Find extra irreducible headers.
 677 ///
 678 /// Find entry blocks and other blocks with backedges, which exist when \c G
 679 /// contains irreducible sub-SCCs.
 680 static void findIrreducibleHeaders(
 681     const BlockFrequencyInfoImplBase &BFI,
 682     const IrreducibleGraph &G,
 683     const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
 684     LoopData::NodeList &Headers, LoopData::NodeList &Others) {
 685   // Map from nodes in the SCC to whether it's an entry block.
 686   SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
 687
 688   // InSCC also acts the set of nodes in the graph.  Seed it.
 689   for (const auto *I : SCC)
 690     InSCC[I] = false;
 691
 692   for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
 693     auto &Irr = *I->first;
 694     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
 695       if (InSCC.count(P))
 696         continue;
 697
 698       // This is an entry block.
 699       I->second = true;
 700       Headers.push_back(Irr.Node);
 701       LLVM_DEBUG(dbgs() << "  => entry = " << BFI.getBlockName(Irr.Node)
 702                         << "\n");
 703       break;
 704     }
 705   }
 706   assert(Headers.size() >= 2 &&
 707          "Expected irreducible CFG; -loop-info is likely invalid");
 708   if (Headers.size() == InSCC.size()) {
 709     // Every block is a header.
 710     llvm::sort(Headers);
 711     return;
 712   }
 713
 714   // Look for extra headers from irreducible sub-SCCs.
 715   for (const auto &I : InSCC) {
 716     // Entry blocks are already headers.
 717     if (I.second)
 718       continue;
 719
 720     auto &Irr = *I.first;
 721     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
 722       // Skip forward edges.
 723       if (P->Node < Irr.Node)
 724         continue;
 725
 726       // Skip predecessors from entry blocks.  These can have inverted
 727       // ordering.
 728       if (InSCC.lookup(P))
 729         continue;
 730
 731       // Store the extra header.
 732       Headers.push_back(Irr.Node);
 733       LLVM_DEBUG(dbgs() << "  => extra = " << BFI.getBlockName(Irr.Node)
 734                         << "\n");
 735       break;
 736     }
 737     if (Headers.back() == Irr.Node)
 738       // Added this as a header.
 739       continue;
 740
 741     // This is not a header.
 742     Others.push_back(Irr.Node);
 743     LLVM_DEBUG(dbgs() << "  => other = " << BFI.getBlockName(Irr.Node) << "\n");
 744   }
 745   llvm::sort(Headers);
 746   llvm::sort(Others);
 747 }
 748
 749 static void createIrreducibleLoop(
 750     BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
 751     LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
 752     const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
 753   // Translate the SCC into RPO.
 754   LLVM_DEBUG(dbgs() << " - found-scc\n");
 755
 756   LoopData::NodeList Headers;
 757   LoopData::NodeList Others;
 758   findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
 759
 760   auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
 761                                 Headers.end(), Others.begin(), Others.end());
 762
 763   // Update loop hierarchy.
 764   for (const auto &N : Loop->Nodes)
 765     if (BFI.Working[N.Index].isLoopHeader())
 766       BFI.Working[N.Index].Loop->Parent = &*Loop;
 767     else
 768       BFI.Working[N.Index].Loop = &*Loop;
 769 }
 770
 771 iterator_range<std::list<LoopData>::iterator>
 772 BlockFrequencyInfoImplBase::analyzeIrreducible(
 773     const IrreducibleGraph &G, LoopData *OuterLoop,
 774     std::list<LoopData>::iterator Insert) {
 775   assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
 776   auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
 777
 778   for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
 779     if (I->size() < 2)
 780       continue;
 781
 782     // Translate the SCC into RPO.
 783     createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
 784   }
 785
 786   if (OuterLoop)
 787     return make_range(std::next(Prev), Insert);
 788   return make_range(Loops.begin(), Insert);
 789 }
 790
 791 void
 792 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
 793   OuterLoop.Exits.clear();
 794   for (auto &Mass : OuterLoop.BackedgeMass)
 795     Mass = BlockMass::getEmpty();
 796   auto O = OuterLoop.Nodes.begin() + 1;
 797   for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
 798     if (!Working[I->Index].isPackaged())
 799       *O++ = *I;
 800   OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
 801 }
 802
 803 void BlockFrequencyInfoImplBase::adjustLoopHeaderMass(LoopData &Loop) {
 804   assert(Loop.isIrreducible() && "this only makes sense on irreducible loops");
 805
 806   // Since the loop has more than one header block, the mass flowing back into
 807   // each header will be different. Adjust the mass in each header loop to
 808   // reflect the masses flowing through back edges.
 809   //
 810   // To do this, we distribute the initial mass using the backedge masses
 811   // as weights for the distribution.
 812   BlockMass LoopMass = BlockMass::getFull();
 813   Distribution Dist;
 814
 815   LLVM_DEBUG(dbgs() << "adjust-loop-header-mass:\n");
 816   for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
 817     auto &HeaderNode = Loop.Nodes[H];
 818     auto &BackedgeMass = Loop.BackedgeMass[Loop.getHeaderIndex(HeaderNode)];
 819     LLVM_DEBUG(dbgs() << " - Add back edge mass for node "
 820                       << getBlockName(HeaderNode) << ": " << BackedgeMass
 821                       << "\n");
 822     if (BackedgeMass.getMass() > 0)
 823       Dist.addLocal(HeaderNode, BackedgeMass.getMass());
 824     else
 825       LLVM_DEBUG(dbgs() << "   Nothing added. Back edge mass is zero\n");
 826   }
 827
 828   DitheringDistributer D(Dist, LoopMass);
 829
 830   LLVM_DEBUG(dbgs() << " Distribute loop mass " << LoopMass
 831                     << " to headers using above weights\n");
 832   for (const Weight &W : Dist.Weights) {
 833     BlockMass Taken = D.takeMass(W.Amount);
 834     assert(W.Type == Weight::Local && "all weights should be local");
 835     Working[W.TargetNode.Index].getMass() = Taken;
 836     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
 837   }
 838 }
 839
 840 void BlockFrequencyInfoImplBase::distributeIrrLoopHeaderMass(Distribution &Dist) {
 841   BlockMass LoopMass = BlockMass::getFull();
 842   DitheringDistributer D(Dist, LoopMass);
 843   for (const Weight &W : Dist.Weights) {
 844     BlockMass Taken = D.takeMass(W.Amount);
 845     assert(W.Type == Weight::Local && "all weights should be local");
 846     Working[W.TargetNode.Index].getMass() = Taken;
 847     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
 848   }
 849 }