lib/Analysis/BlockFrequencyInfoImpl.cpp

   1 //===- BlockFrequencyImplInfo.cpp - Block Frequency Info Implementation ---===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // Loops should be simplified before this analysis.
  10 //
  11 //===----------------------------------------------------------------------===//
  12
  13 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
  14 #include "llvm/ADT/APInt.h"
  15 #include "llvm/ADT/DenseMap.h"
  16 #include "llvm/ADT/GraphTraits.h"
  17 #include "llvm/ADT/None.h"
  18 #include "llvm/ADT/SCCIterator.h"
  19 #include "llvm/Config/llvm-config.h"
  20 #include "llvm/IR/Function.h"
  21 #include "llvm/Support/BlockFrequency.h"
  22 #include "llvm/Support/BranchProbability.h"
  23 #include "llvm/Support/Compiler.h"
  24 #include "llvm/Support/Debug.h"
  25 #include "llvm/Support/ScaledNumber.h"
  26 #include "llvm/Support/MathExtras.h"
  27 #include "llvm/Support/raw_ostream.h"
  28 #include <algorithm>
  29 #include <cassert>
  30 #include <cstddef>
  31 #include <cstdint>
  32 #include <iterator>
  33 #include <list>
  34 #include <numeric>
  35 #include <utility>
  36 #include <vector>
  37
  38 using namespace llvm;
  39 using namespace llvm::bfi_detail;
  40
  41 #define DEBUG_TYPE "block-freq"
  42
  43 ScaledNumber<uint64_t> BlockMass::toScaled() const {
  44   if (isFull())
  45     return ScaledNumber<uint64_t>(1, 0);
  46   return ScaledNumber<uint64_t>(getMass() + 1, -64);
  47 }
  48
  49 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
  50 LLVM_DUMP_METHOD void BlockMass::dump() const { print(dbgs()); }
  51 #endif
  52
  53 static char getHexDigit(int N) {
  54   assert(N < 16);
  55   if (N < 10)
  56     return '0' + N;
  57   return 'a' + N - 10;
  58 }
  59
  60 raw_ostream &BlockMass::print(raw_ostream &OS) const {
  61   for (int Digits = 0; Digits < 16; ++Digits)
  62     OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
  63   return OS;
  64 }
  65
  66 namespace {
  67
  68 using BlockNode = BlockFrequencyInfoImplBase::BlockNode;
  69 using Distribution = BlockFrequencyInfoImplBase::Distribution;
  70 using WeightList = BlockFrequencyInfoImplBase::Distribution::WeightList;
  71 using Scaled64 = BlockFrequencyInfoImplBase::Scaled64;
  72 using LoopData = BlockFrequencyInfoImplBase::LoopData;
  73 using Weight = BlockFrequencyInfoImplBase::Weight;
  74 using FrequencyData = BlockFrequencyInfoImplBase::FrequencyData;
  75
  76 /// Dithering mass distributer.
  77 ///
  78 /// This class splits up a single mass into portions by weight, dithering to
  79 /// spread out error.  No mass is lost.  The dithering precision depends on the
  80 /// precision of the product of \a BlockMass and \a BranchProbability.
  81 ///
  82 /// The distribution algorithm follows.
  83 ///
  84 ///  1. Initialize by saving the sum of the weights in \a RemWeight and the
  85 ///     mass to distribute in \a RemMass.
  86 ///
  87 ///  2. For each portion:
  88 ///
  89 ///      1. Construct a branch probability, P, as the portion's weight divided
  90 ///         by the current value of \a RemWeight.
  91 ///      2. Calculate the portion's mass as \a RemMass times P.
  92 ///      3. Update \a RemWeight and \a RemMass at each portion by subtracting
  93 ///         the current portion's weight and mass.
  94 struct DitheringDistributer {
  95   uint32_t RemWeight;
  96   BlockMass RemMass;
  97
  98   DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
  99
 100   BlockMass takeMass(uint32_t Weight);
 101 };
 102
 103 } // end anonymous namespace
 104
 105 DitheringDistributer::DitheringDistributer(Distribution &Dist,
 106                                            const BlockMass &Mass) {
 107   Dist.normalize();
 108   RemWeight = Dist.Total;
 109   RemMass = Mass;
 110 }
 111
 112 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
 113   assert(Weight && "invalid weight");
 114   assert(Weight <= RemWeight);
 115   BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
 116
 117   // Decrement totals (dither).
 118   RemWeight -= Weight;
 119   RemMass -= Mass;
 120   return Mass;
 121 }
 122
 123 void Distribution::add(const BlockNode &Node, uint64_t Amount,
 124                        Weight::DistType Type) {
 125   assert(Amount && "invalid weight of 0");
 126   uint64_t NewTotal = Total + Amount;
 127
 128   // Check for overflow.  It should be impossible to overflow twice.
 129   bool IsOverflow = NewTotal < Total;
 130   assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
 131   DidOverflow |= IsOverflow;
 132
 133   // Update the total.
 134   Total = NewTotal;
 135
 136   // Save the weight.
 137   Weights.push_back(Weight(Type, Node, Amount));
 138 }
 139
 140 static void combineWeight(Weight &W, const Weight &OtherW) {
 141   assert(OtherW.TargetNode.isValid());
 142   if (!W.Amount) {
 143     W = OtherW;
 144     return;
 145   }
 146   assert(W.Type == OtherW.Type);
 147   assert(W.TargetNode == OtherW.TargetNode);
 148   assert(OtherW.Amount && "Expected non-zero weight");
 149   if (W.Amount > W.Amount + OtherW.Amount)
 150     // Saturate on overflow.
 151     W.Amount = UINT64_MAX;
 152   else
 153     W.Amount += OtherW.Amount;
 154 }
 155
 156 static void combineWeightsBySorting(WeightList &Weights) {
 157   // Sort so edges to the same node are adjacent.
 158   llvm::sort(Weights, [](const Weight &L, const Weight &R) {
 159     return L.TargetNode < R.TargetNode;
 160   });
 161
 162   // Combine adjacent edges.
 163   WeightList::iterator O = Weights.begin();
 164   for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
 165        ++O, (I = L)) {
 166     *O = *I;
 167
 168     // Find the adjacent weights to the same node.
 169     for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
 170       combineWeight(*O, *L);
 171   }
 172
 173   // Erase extra entries.
 174   Weights.erase(O, Weights.end());
 175 }
 176
 177 static void combineWeightsByHashing(WeightList &Weights) {
 178   // Collect weights into a DenseMap.
 179   using HashTable = DenseMap<BlockNode::IndexType, Weight>;
 180
 181   HashTable Combined(NextPowerOf2(2 * Weights.size()));
 182   for (const Weight &W : Weights)
 183     combineWeight(Combined[W.TargetNode.Index], W);
 184
 185   // Check whether anything changed.
 186   if (Weights.size() == Combined.size())
 187     return;
 188
 189   // Fill in the new weights.
 190   Weights.clear();
 191   Weights.reserve(Combined.size());
 192   for (const auto &I : Combined)
 193     Weights.push_back(I.second);
 194 }
 195
 196 static void combineWeights(WeightList &Weights) {
 197   // Use a hash table for many successors to keep this linear.
 198   if (Weights.size() > 128) {
 199     combineWeightsByHashing(Weights);
 200     return;
 201   }
 202
 203   combineWeightsBySorting(Weights);
 204 }
 205
 206 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
 207   assert(Shift >= 0);
 208   assert(Shift < 64);
 209   if (!Shift)
 210     return N;
 211   return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
 212 }
 213
 214 void Distribution::normalize() {
 215   // Early exit for termination nodes.
 216   if (Weights.empty())
 217     return;
 218
 219   // Only bother if there are multiple successors.
 220   if (Weights.size() > 1)
 221     combineWeights(Weights);
 222
 223   // Early exit when combined into a single successor.
 224   if (Weights.size() == 1) {
 225     Total = 1;
 226     Weights.front().Amount = 1;
 227     return;
 228   }
 229
 230   // Determine how much to shift right so that the total fits into 32-bits.
 231   //
 232   // If we shift at all, shift by 1 extra.  Otherwise, the lower limit of 1
 233   // for each weight can cause a 32-bit overflow.
 234   int Shift = 0;
 235   if (DidOverflow)
 236     Shift = 33;
 237   else if (Total > UINT32_MAX)
 238     Shift = 33 - countLeadingZeros(Total);
 239
 240   // Early exit if nothing needs to be scaled.
 241   if (!Shift) {
 242     // If we didn't overflow then combineWeights() shouldn't have changed the
 243     // sum of the weights, but let's double-check.
 244     assert(Total == std::accumulate(Weights.begin(), Weights.end(), UINT64_C(0),
 245                                     [](uint64_t Sum, const Weight &W) {
 246                       return Sum + W.Amount;
 247                     }) &&
 248            "Expected total to be correct");
 249     return;
 250   }
 251
 252   // Recompute the total through accumulation (rather than shifting it) so that
 253   // it's accurate after shifting and any changes combineWeights() made above.
 254   Total = 0;
 255
 256   // Sum the weights to each node and shift right if necessary.
 257   for (Weight &W : Weights) {
 258     // Scale down below UINT32_MAX.  Since Shift is larger than necessary, we
 259     // can round here without concern about overflow.
 260     assert(W.TargetNode.isValid());
 261     W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
 262     assert(W.Amount <= UINT32_MAX);
 263
 264     // Update the total.
 265     Total += W.Amount;
 266   }
 267   assert(Total <= UINT32_MAX);
 268 }
 269
 270 void BlockFrequencyInfoImplBase::clear() {
 271   // Swap with a default-constructed std::vector, since std::vector<>::clear()
 272   // does not actually clear heap storage.
 273   std::vector<FrequencyData>().swap(Freqs);
 274   IsIrrLoopHeader.clear();
 275   std::vector<WorkingData>().swap(Working);
 276   Loops.clear();
 277 }
 278
 279 /// Clear all memory not needed downstream.
 280 ///
 281 /// Releases all memory not used downstream.  In particular, saves Freqs.
 282 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
 283   std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
 284   SparseBitVector<> SavedIsIrrLoopHeader(std::move(BFI.IsIrrLoopHeader));
 285   BFI.clear();
 286   BFI.Freqs = std::move(SavedFreqs);
 287   BFI.IsIrrLoopHeader = std::move(SavedIsIrrLoopHeader);
 288 }
 289
 290 bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
 291                                            const LoopData *OuterLoop,
 292                                            const BlockNode &Pred,
 293                                            const BlockNode &Succ,
 294                                            uint64_t Weight) {
 295   if (!Weight)
 296     Weight = 1;
 297
 298   auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
 299     return OuterLoop && OuterLoop->isHeader(Node);
 300   };
 301
 302   BlockNode Resolved = Working[Succ.Index].getResolvedNode();
 303
 304 #ifndef NDEBUG
 305   auto debugSuccessor = [&](const char *Type) {
 306     dbgs() << "  =>"
 307            << " [" << Type << "] weight = " << Weight;
 308     if (!isLoopHeader(Resolved))
 309       dbgs() << ", succ = " << getBlockName(Succ);
 310     if (Resolved != Succ)
 311       dbgs() << ", resolved = " << getBlockName(Resolved);
 312     dbgs() << "\n";
 313   };
 314   (void)debugSuccessor;
 315 #endif
 316
 317   if (isLoopHeader(Resolved)) {
 318     LLVM_DEBUG(debugSuccessor("backedge"));
 319     Dist.addBackedge(Resolved, Weight);
 320     return true;
 321   }
 322
 323   if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
 324     LLVM_DEBUG(debugSuccessor("  exit  "));
 325     Dist.addExit(Resolved, Weight);
 326     return true;
 327   }
 328
 329   if (Resolved < Pred) {
 330     if (!isLoopHeader(Pred)) {
 331       // If OuterLoop is an irreducible loop, we can't actually handle this.
 332       assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
 333              "unhandled irreducible control flow");
 334
 335       // Irreducible backedge.  Abort.
 336       LLVM_DEBUG(debugSuccessor("abort!!!"));
 337       return false;
 338     }
 339
 340     // If "Pred" is a loop header, then this isn't really a backedge; rather,
 341     // OuterLoop must be irreducible.  These false backedges can come only from
 342     // secondary loop headers.
 343     assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
 344            "unhandled irreducible control flow");
 345   }
 346
 347   LLVM_DEBUG(debugSuccessor(" local  "));
 348   Dist.addLocal(Resolved, Weight);
 349   return true;
 350 }
 351
 352 bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
 353     const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
 354   // Copy the exit map into Dist.
 355   for (const auto &I : Loop.Exits)
 356     if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
 357                    I.second.getMass()))
 358       // Irreducible backedge.
 359       return false;
 360
 361   return true;
 362 }
 363
 364 /// Compute the loop scale for a loop.
 365 void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
 366   // Compute loop scale.
 367   LLVM_DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
 368
 369   // Infinite loops need special handling. If we give the back edge an infinite
 370   // mass, they may saturate all the other scales in the function down to 1,
 371   // making all the other region temperatures look exactly the same. Choose an
 372   // arbitrary scale to avoid these issues.
 373   //
 374   // FIXME: An alternate way would be to select a symbolic scale which is later
 375   // replaced to be the maximum of all computed scales plus 1. This would
 376   // appropriately describe the loop as having a large scale, without skewing
 377   // the final frequency computation.
 378   const Scaled64 InfiniteLoopScale(1, 12);
 379
 380   // LoopScale == 1 / ExitMass
 381   // ExitMass == HeadMass - BackedgeMass
 382   BlockMass TotalBackedgeMass;
 383   for (auto &Mass : Loop.BackedgeMass)
 384     TotalBackedgeMass += Mass;
 385   BlockMass ExitMass = BlockMass::getFull() - TotalBackedgeMass;
 386
 387   // Block scale stores the inverse of the scale. If this is an infinite loop,
 388   // its exit mass will be zero. In this case, use an arbitrary scale for the
 389   // loop scale.
 390   Loop.Scale =
 391       ExitMass.isEmpty() ? InfiniteLoopScale : ExitMass.toScaled().inverse();
 392
 393   LLVM_DEBUG(dbgs() << " - exit-mass = " << ExitMass << " ("
 394                     << BlockMass::getFull() << " - " << TotalBackedgeMass
 395                     << ")\n"
 396                     << " - scale = " << Loop.Scale << "\n");
 397 }
 398
 399 /// Package up a loop.
 400 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
 401   LLVM_DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
 402
 403   // Clear the subloop exits to prevent quadratic memory usage.
 404   for (const BlockNode &M : Loop.Nodes) {
 405     if (auto *Loop = Working[M.Index].getPackagedLoop())
 406       Loop->Exits.clear();
 407     LLVM_DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
 408   }
 409   Loop.IsPackaged = true;
 410 }
 411
 412 #ifndef NDEBUG
 413 static void debugAssign(const BlockFrequencyInfoImplBase &BFI,
 414                         const DitheringDistributer &D, const BlockNode &T,
 415                         const BlockMass &M, const char *Desc) {
 416   dbgs() << "  => assign " << M << " (" << D.RemMass << ")";
 417   if (Desc)
 418     dbgs() << " [" << Desc << "]";
 419   if (T.isValid())
 420     dbgs() << " to " << BFI.getBlockName(T);
 421   dbgs() << "\n";
 422 }
 423 #endif
 424
 425 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
 426                                                 LoopData *OuterLoop,
 427                                                 Distribution &Dist) {
 428   BlockMass Mass = Working[Source.Index].getMass();
 429   LLVM_DEBUG(dbgs() << "  => mass:  " << Mass << "\n");
 430
 431   // Distribute mass to successors as laid out in Dist.
 432   DitheringDistributer D(Dist, Mass);
 433
 434   for (const Weight &W : Dist.Weights) {
 435     // Check for a local edge (non-backedge and non-exit).
 436     BlockMass Taken = D.takeMass(W.Amount);
 437     if (W.Type == Weight::Local) {
 438       Working[W.TargetNode.Index].getMass() += Taken;
 439       LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
 440       continue;
 441     }
 442
 443     // Backedges and exits only make sense if we're processing a loop.
 444     assert(OuterLoop && "backedge or exit outside of loop");
 445
 446     // Check for a backedge.
 447     if (W.Type == Weight::Backedge) {
 448       OuterLoop->BackedgeMass[OuterLoop->getHeaderIndex(W.TargetNode)] += Taken;
 449       LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "back"));
 450       continue;
 451     }
 452
 453     // This must be an exit.
 454     assert(W.Type == Weight::Exit);
 455     OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
 456     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "exit"));
 457   }
 458 }
 459
 460 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
 461                                      const Scaled64 &Min, const Scaled64 &Max) {
 462   // Scale the Factor to a size that creates integers.  Ideally, integers would
 463   // be scaled so that Max == UINT64_MAX so that they can be best
 464   // differentiated.  However, in the presence of large frequency values, small
 465   // frequencies are scaled down to 1, making it impossible to differentiate
 466   // small, unequal numbers. When the spread between Min and Max frequencies
 467   // fits well within MaxBits, we make the scale be at least 8.
 468   const unsigned MaxBits = 64;
 469   const unsigned SpreadBits = (Max / Min).lg();
 470   Scaled64 ScalingFactor;
 471   if (SpreadBits <= MaxBits - 3) {
 472     // If the values are small enough, make the scaling factor at least 8 to
 473     // allow distinguishing small values.
 474     ScalingFactor = Min.inverse();
 475     ScalingFactor <<= 3;
 476   } else {
 477     // If the values need more than MaxBits to be represented, saturate small
 478     // frequency values down to 1 by using a scaling factor that benefits large
 479     // frequency values.
 480     ScalingFactor = Scaled64(1, MaxBits) / Max;
 481   }
 482
 483   // Translate the floats to integers.
 484   LLVM_DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
 485                     << ", factor = " << ScalingFactor << "\n");
 486   for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
 487     Scaled64 Scaled = BFI.Freqs[Index].Scaled * ScalingFactor;
 488     BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
 489     LLVM_DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
 490                       << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
 491                       << ", int = " << BFI.Freqs[Index].Integer << "\n");
 492   }
 493 }
 494
 495 /// Unwrap a loop package.
 496 ///
 497 /// Visits all the members of a loop, adjusting their BlockData according to
 498 /// the loop's pseudo-node.
 499 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
 500   LLVM_DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
 501                     << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
 502                     << "\n");
 503   Loop.Scale *= Loop.Mass.toScaled();
 504   Loop.IsPackaged = false;
 505   LLVM_DEBUG(dbgs() << "  => combined-scale = " << Loop.Scale << "\n");
 506
 507   // Propagate the head scale through the loop.  Since members are visited in
 508   // RPO, the head scale will be updated by the loop scale first, and then the
 509   // final head scale will be used for updated the rest of the members.
 510   for (const BlockNode &N : Loop.Nodes) {
 511     const auto &Working = BFI.Working[N.Index];
 512     Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
 513                                        : BFI.Freqs[N.Index].Scaled;
 514     Scaled64 New = Loop.Scale * F;
 515     LLVM_DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => "
 516                       << New << "\n");
 517     F = New;
 518   }
 519 }
 520
 521 void BlockFrequencyInfoImplBase::unwrapLoops() {
 522   // Set initial frequencies from loop-local masses.
 523   for (size_t Index = 0; Index < Working.size(); ++Index)
 524     Freqs[Index].Scaled = Working[Index].Mass.toScaled();
 525
 526   for (LoopData &Loop : Loops)
 527     unwrapLoop(*this, Loop);
 528 }
 529
 530 void BlockFrequencyInfoImplBase::finalizeMetrics() {
 531   // Unwrap loop packages in reverse post-order, tracking min and max
 532   // frequencies.
 533   auto Min = Scaled64::getLargest();
 534   auto Max = Scaled64::getZero();
 535   for (size_t Index = 0; Index < Working.size(); ++Index) {
 536     // Update min/max scale.
 537     Min = std::min(Min, Freqs[Index].Scaled);
 538     Max = std::max(Max, Freqs[Index].Scaled);
 539   }
 540
 541   // Convert to integers.
 542   convertFloatingToInteger(*this, Min, Max);
 543
 544   // Clean up data structures.
 545   cleanup(*this);
 546
 547   // Print out the final stats.
 548   LLVM_DEBUG(dump());
 549 }
 550
 551 BlockFrequency
 552 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
 553   if (!Node.isValid())
 554     return 0;
 555   return Freqs[Node.Index].Integer;
 556 }
 557
 558 Optional<uint64_t>
 559 BlockFrequencyInfoImplBase::getBlockProfileCount(const Function &F,
 560                                                  const BlockNode &Node,
 561                                                  bool AllowSynthetic) const {
 562   return getProfileCountFromFreq(F, getBlockFreq(Node).getFrequency(),
 563                                  AllowSynthetic);
 564 }
 565
 566 Optional<uint64_t>
 567 BlockFrequencyInfoImplBase::getProfileCountFromFreq(const Function &F,
 568                                                     uint64_t Freq,
 569                                                     bool AllowSynthetic) const {
 570   auto EntryCount = F.getEntryCount(AllowSynthetic);
 571   if (!EntryCount)
 572     return None;
 573   // Use 128 bit APInt to do the arithmetic to avoid overflow.
 574   APInt BlockCount(128, EntryCount.getCount());
 575   APInt BlockFreq(128, Freq);
 576   APInt EntryFreq(128, getEntryFreq());
 577   BlockCount *= BlockFreq;
 578   // Rounded division of BlockCount by EntryFreq. Since EntryFreq is unsigned
 579   // lshr by 1 gives EntryFreq/2.
 580   BlockCount = (BlockCount + EntryFreq.lshr(1)).udiv(EntryFreq);
 581   return BlockCount.getLimitedValue();
 582 }
 583
 584 bool
 585 BlockFrequencyInfoImplBase::isIrrLoopHeader(const BlockNode &Node) {
 586   if (!Node.isValid())
 587     return false;
 588   return IsIrrLoopHeader.test(Node.Index);
 589 }
 590
 591 Scaled64
 592 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
 593   if (!Node.isValid())
 594     return Scaled64::getZero();
 595   return Freqs[Node.Index].Scaled;
 596 }
 597
 598 void BlockFrequencyInfoImplBase::setBlockFreq(const BlockNode &Node,
 599                                               uint64_t Freq) {
 600   assert(Node.isValid() && "Expected valid node");
 601   assert(Node.Index < Freqs.size() && "Expected legal index");
 602   Freqs[Node.Index].Integer = Freq;
 603 }
 604
 605 std::string
 606 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
 607   return {};
 608 }
 609
 610 std::string
 611 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
 612   return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
 613 }
 614
 615 raw_ostream &
 616 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
 617                                            const BlockNode &Node) const {
 618   return OS << getFloatingBlockFreq(Node);
 619 }
 620
 621 raw_ostream &
 622 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
 623                                            const BlockFrequency &Freq) const {
 624   Scaled64 Block(Freq.getFrequency(), 0);
 625   Scaled64 Entry(getEntryFreq(), 0);
 626
 627   return OS << Block / Entry;
 628 }
 629
 630 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
 631   Start = OuterLoop.getHeader();
 632   Nodes.reserve(OuterLoop.Nodes.size());
 633   for (auto N : OuterLoop.Nodes)
 634     addNode(N);
 635   indexNodes();
 636 }
 637
 638 void IrreducibleGraph::addNodesInFunction() {
 639   Start = 0;
 640   for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
 641     if (!BFI.Working[Index].isPackaged())
 642       addNode(Index);
 643   indexNodes();
 644 }
 645
 646 void IrreducibleGraph::indexNodes() {
 647   for (auto &I : Nodes)
 648     Lookup[I.Node.Index] = &I;
 649 }
 650
 651 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
 652                                const BFIBase::LoopData *OuterLoop) {
 653   if (OuterLoop && OuterLoop->isHeader(Succ))
 654     return;
 655   auto L = Lookup.find(Succ.Index);
 656   if (L == Lookup.end())
 657     return;
 658   IrrNode &SuccIrr = *L->second;
 659   Irr.Edges.push_back(&SuccIrr);
 660   SuccIrr.Edges.push_front(&Irr);
 661   ++SuccIrr.NumIn;
 662 }
 663
 664 namespace llvm {
 665
 666 template <> struct GraphTraits<IrreducibleGraph> {
 667   using GraphT = bfi_detail::IrreducibleGraph;
 668   using NodeRef = const GraphT::IrrNode *;
 669   using ChildIteratorType = GraphT::IrrNode::iterator;
 670
 671   static NodeRef getEntryNode(const GraphT &G) { return G.StartIrr; }
 672   static ChildIteratorType child_begin(NodeRef N) { return N->succ_begin(); }
 673   static ChildIteratorType child_end(NodeRef N) { return N->succ_end(); }
 674 };
 675
 676 } // end namespace llvm
 677
 678 /// Find extra irreducible headers.
 679 ///
 680 /// Find entry blocks and other blocks with backedges, which exist when \c G
 681 /// contains irreducible sub-SCCs.
 682 static void findIrreducibleHeaders(
 683     const BlockFrequencyInfoImplBase &BFI,
 684     const IrreducibleGraph &G,
 685     const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
 686     LoopData::NodeList &Headers, LoopData::NodeList &Others) {
 687   // Map from nodes in the SCC to whether it's an entry block.
 688   SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
 689
 690   // InSCC also acts the set of nodes in the graph.  Seed it.
 691   for (const auto *I : SCC)
 692     InSCC[I] = false;
 693
 694   for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
 695     auto &Irr = *I->first;
 696     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
 697       if (InSCC.count(P))
 698         continue;
 699
 700       // This is an entry block.
 701       I->second = true;
 702       Headers.push_back(Irr.Node);
 703       LLVM_DEBUG(dbgs() << "  => entry = " << BFI.getBlockName(Irr.Node)
 704                         << "\n");
 705       break;
 706     }
 707   }
 708   assert(Headers.size() >= 2 &&
 709          "Expected irreducible CFG; -loop-info is likely invalid");
 710   if (Headers.size() == InSCC.size()) {
 711     // Every block is a header.
 712     llvm::sort(Headers);
 713     return;
 714   }
 715
 716   // Look for extra headers from irreducible sub-SCCs.
 717   for (const auto &I : InSCC) {
 718     // Entry blocks are already headers.
 719     if (I.second)
 720       continue;
 721
 722     auto &Irr = *I.first;
 723     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
 724       // Skip forward edges.
 725       if (P->Node < Irr.Node)
 726         continue;
 727
 728       // Skip predecessors from entry blocks.  These can have inverted
 729       // ordering.
 730       if (InSCC.lookup(P))
 731         continue;
 732
 733       // Store the extra header.
 734       Headers.push_back(Irr.Node);
 735       LLVM_DEBUG(dbgs() << "  => extra = " << BFI.getBlockName(Irr.Node)
 736                         << "\n");
 737       break;
 738     }
 739     if (Headers.back() == Irr.Node)
 740       // Added this as a header.
 741       continue;
 742
 743     // This is not a header.
 744     Others.push_back(Irr.Node);
 745     LLVM_DEBUG(dbgs() << "  => other = " << BFI.getBlockName(Irr.Node) << "\n");
 746   }
 747   llvm::sort(Headers);
 748   llvm::sort(Others);
 749 }
 750
 751 static void createIrreducibleLoop(
 752     BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
 753     LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
 754     const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
 755   // Translate the SCC into RPO.
 756   LLVM_DEBUG(dbgs() << " - found-scc\n");
 757
 758   LoopData::NodeList Headers;
 759   LoopData::NodeList Others;
 760   findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
 761
 762   auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
 763                                 Headers.end(), Others.begin(), Others.end());
 764
 765   // Update loop hierarchy.
 766   for (const auto &N : Loop->Nodes)
 767     if (BFI.Working[N.Index].isLoopHeader())
 768       BFI.Working[N.Index].Loop->Parent = &*Loop;
 769     else
 770       BFI.Working[N.Index].Loop = &*Loop;
 771 }
 772
 773 iterator_range<std::list<LoopData>::iterator>
 774 BlockFrequencyInfoImplBase::analyzeIrreducible(
 775     const IrreducibleGraph &G, LoopData *OuterLoop,
 776     std::list<LoopData>::iterator Insert) {
 777   assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
 778   auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
 779
 780   for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
 781     if (I->size() < 2)
 782       continue;
 783
 784     // Translate the SCC into RPO.
 785     createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
 786   }
 787
 788   if (OuterLoop)
 789     return make_range(std::next(Prev), Insert);
 790   return make_range(Loops.begin(), Insert);
 791 }
 792
 793 void
 794 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
 795   OuterLoop.Exits.clear();
 796   for (auto &Mass : OuterLoop.BackedgeMass)
 797     Mass = BlockMass::getEmpty();
 798   auto O = OuterLoop.Nodes.begin() + 1;
 799   for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
 800     if (!Working[I->Index].isPackaged())
 801       *O++ = *I;
 802   OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
 803 }
 804
 805 void BlockFrequencyInfoImplBase::adjustLoopHeaderMass(LoopData &Loop) {
 806   assert(Loop.isIrreducible() && "this only makes sense on irreducible loops");
 807
 808   // Since the loop has more than one header block, the mass flowing back into
 809   // each header will be different. Adjust the mass in each header loop to
 810   // reflect the masses flowing through back edges.
 811   //
 812   // To do this, we distribute the initial mass using the backedge masses
 813   // as weights for the distribution.
 814   BlockMass LoopMass = BlockMass::getFull();
 815   Distribution Dist;
 816
 817   LLVM_DEBUG(dbgs() << "adjust-loop-header-mass:\n");
 818   for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
 819     auto &HeaderNode = Loop.Nodes[H];
 820     auto &BackedgeMass = Loop.BackedgeMass[Loop.getHeaderIndex(HeaderNode)];
 821     LLVM_DEBUG(dbgs() << " - Add back edge mass for node "
 822                       << getBlockName(HeaderNode) << ": " << BackedgeMass
 823                       << "\n");
 824     if (BackedgeMass.getMass() > 0)
 825       Dist.addLocal(HeaderNode, BackedgeMass.getMass());
 826     else
 827       LLVM_DEBUG(dbgs() << "   Nothing added. Back edge mass is zero\n");
 828   }
 829
 830   DitheringDistributer D(Dist, LoopMass);
 831
 832   LLVM_DEBUG(dbgs() << " Distribute loop mass " << LoopMass
 833                     << " to headers using above weights\n");
 834   for (const Weight &W : Dist.Weights) {
 835     BlockMass Taken = D.takeMass(W.Amount);
 836     assert(W.Type == Weight::Local && "all weights should be local");
 837     Working[W.TargetNode.Index].getMass() = Taken;
 838     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
 839   }
 840 }
 841
 842 void BlockFrequencyInfoImplBase::distributeIrrLoopHeaderMass(Distribution &Dist) {
 843   BlockMass LoopMass = BlockMass::getFull();
 844   DitheringDistributer D(Dist, LoopMass);
 845   for (const Weight &W : Dist.Weights) {
 846     BlockMass Taken = D.takeMass(W.Amount);
 847     assert(W.Type == Weight::Local && "all weights should be local");
 848     Working[W.TargetNode.Index].getMass() = Taken;
 849     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
 850   }
 851 }