llvm/lib/Transforms/Scalar/SampleProfile.cpp

//===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===//
//
//                      The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements the SampleProfileLoader transformation. This pass
// reads a profile file generated by a sampling profiler (e.g. Linux Perf -
// http://perf.wiki.kernel.org/) and generates IR metadata to reflect the
// profile information in the given profile.
//
// This pass generates branch weight annotations on the IR:
//
// - prof: Represents branch weights. This annotation is added to branches
//      to indicate the weights of each edge coming out of the branch.
//      The weight of each edge is the weight of the target block for
//      that edge. The weight of a block B is computed as the maximum
//      number of samples found in B.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "sample-profile"

#include "llvm/Transforms/Scalar.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/StringMap.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/PostDominators.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DebugInfo.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/MDBuilder.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/Module.h"
#include "llvm/Pass.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/LineIterator.h"
#include "llvm/Support/MemoryBuffer.h"
#include "llvm/Support/Regex.h"
#include "llvm/Support/raw_ostream.h"
#include <cctype>

using namespace llvm;

// Command line option to specify the file to read samples from. This is
// mainly used for debugging.
static cl::opt<std::string> SampleProfileFile(
    "sample-profile-file", cl::init(""), cl::value_desc("filename"),
    cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
    "sample-profile-max-propagate-iterations", cl::init(100),
    cl::desc("Maximum number of iterations to go through when propagating "
             "sample block/edge weights through the CFG."));

namespace {
/// \brief Represents the relative location of an instruction.
///
/// Instruction locations are specified by the line offset from the
/// beginning of the function (marked by the line where the function
/// header is) and the discriminator value within that line.
///
/// The discriminator value is useful to distinguish instructions
/// that are on the same line but belong to different basic blocks
/// (e.g., the two post-increment instructions in "if (p) x++; else y++;").
struct InstructionLocation {
  InstructionLocation(int L, unsigned D) : LineOffset(L), Discriminator(D) {}
  int LineOffset;
  unsigned Discriminator;
};
}

namespace llvm {
template <> struct DenseMapInfo<InstructionLocation> {
  typedef DenseMapInfo<int> OffsetInfo;
  typedef DenseMapInfo<unsigned> DiscriminatorInfo;
  static inline InstructionLocation getEmptyKey() {
    return InstructionLocation(OffsetInfo::getEmptyKey(),
                               DiscriminatorInfo::getEmptyKey());
  }
  static inline InstructionLocation getTombstoneKey() {
    return InstructionLocation(OffsetInfo::getTombstoneKey(),
                               DiscriminatorInfo::getTombstoneKey());
  }
  static inline unsigned getHashValue(InstructionLocation Val) {
    return DenseMapInfo<std::pair<int, unsigned> >::getHashValue(
        std::pair<int, unsigned>(Val.LineOffset, Val.Discriminator));
  }
  static inline bool isEqual(InstructionLocation LHS, InstructionLocation RHS) {
    return LHS.LineOffset == RHS.LineOffset &&
           LHS.Discriminator == RHS.Discriminator;
  }
};
}

namespace {
typedef DenseMap<InstructionLocation, unsigned> BodySampleMap;
typedef DenseMap<BasicBlock *, unsigned> BlockWeightMap;
typedef DenseMap<BasicBlock *, BasicBlock *> EquivalenceClassMap;
typedef std::pair<BasicBlock *, BasicBlock *> Edge;
typedef DenseMap<Edge, unsigned> EdgeWeightMap;
typedef DenseMap<BasicBlock *, SmallVector<BasicBlock *, 8> > BlockEdgeMap;

/// \brief Representation of the runtime profile for a function.
///
/// This data structure contains the runtime profile for a given
/// function. It contains the total number of samples collected
/// in the function and a map of samples collected in every statement.
class SampleFunctionProfile {
public:
  SampleFunctionProfile()
      : TotalSamples(0), TotalHeadSamples(0), HeaderLineno(0), DT(0), PDT(0),
        LI(0), Ctx(0) {}

  unsigned getFunctionLoc(Function &F);
  bool emitAnnotations(Function &F, DominatorTree *DomTree,
                       PostDominatorTree *PostDomTree, LoopInfo *Loops);
  unsigned getInstWeight(Instruction &I);
  unsigned getBlockWeight(BasicBlock *B);
  void addTotalSamples(unsigned Num) { TotalSamples += Num; }
  void addHeadSamples(unsigned Num) { TotalHeadSamples += Num; }
  void addBodySamples(int LineOffset, unsigned Discriminator, unsigned Num) {
    assert(LineOffset >= 0);
    BodySamples[InstructionLocation(LineOffset, Discriminator)] += Num;
  }
  void print(raw_ostream &OS);
  void printEdgeWeight(raw_ostream &OS, Edge E);
  void printBlockWeight(raw_ostream &OS, BasicBlock *BB);
  void printBlockEquivalence(raw_ostream &OS, BasicBlock *BB);
  bool computeBlockWeights(Function &F);
  void findEquivalenceClasses(Function &F);
  void findEquivalencesFor(BasicBlock *BB1,
                           SmallVector<BasicBlock *, 8> Descendants,
                           DominatorTreeBase<BasicBlock> *DomTree);
  void propagateWeights(Function &F);
  unsigned visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
  void buildEdges(Function &F);
  bool propagateThroughEdges(Function &F);
  bool empty() { return BodySamples.empty(); }

protected:
  /// \brief Total number of samples collected inside this function.
  ///
  /// Samples are cumulative, they include all the samples collected
  /// inside this function and all its inlined callees.
  unsigned TotalSamples;

  /// \brief Total number of samples collected at the head of the function.
  /// FIXME: Use head samples to estimate a cold/hot attribute for the function.
  unsigned TotalHeadSamples;

  /// \brief Line number for the function header. Used to compute relative
  /// line numbers from the absolute line LOCs found in instruction locations.
  /// The relative line numbers are needed to address the samples from the
  /// profile file.
  unsigned HeaderLineno;

  /// \brief Map line offsets to collected samples.
  ///
  /// Each entry in this map contains the number of samples
  /// collected at the corresponding line offset. All line locations
  /// are an offset from the start of the function.
  BodySampleMap BodySamples;

  /// \brief Map basic blocks to their computed weights.
  ///
  /// The weight of a basic block is defined to be the maximum
  /// of all the instruction weights in that block.
  BlockWeightMap BlockWeights;

  /// \brief Map edges to their computed weights.
  ///
  /// Edge weights are computed by propagating basic block weights in
  /// SampleProfile::propagateWeights.
  EdgeWeightMap EdgeWeights;

  /// \brief Set of visited blocks during propagation.
  SmallPtrSet<BasicBlock *, 128> VisitedBlocks;

  /// \brief Set of visited edges during propagation.
  SmallSet<Edge, 128> VisitedEdges;

  /// \brief Equivalence classes for block weights.
  ///
  /// Two blocks BB1 and BB2 are in the same equivalence class if they
  /// dominate and post-dominate each other, and they are in the same loop
  /// nest. When this happens, the two blocks are guaranteed to execute
  /// the same number of times.
  EquivalenceClassMap EquivalenceClass;

  /// \brief Dominance, post-dominance and loop information.
  DominatorTree *DT;
  PostDominatorTree *PDT;
  LoopInfo *LI;

  /// \brief Predecessors for each basic block in the CFG.
  BlockEdgeMap Predecessors;

  /// \brief Successors for each basic block in the CFG.
  BlockEdgeMap Successors;

  /// \brief LLVM context holding the debug data we need.
  LLVMContext *Ctx;
};

/// \brief Sample-based profile reader.
///
/// Each profile contains sample counts for all the functions
/// executed. Inside each function, statements are annotated with the
/// collected samples on all the instructions associated with that
/// statement.
///
/// For this to produce meaningful data, the program needs to be
/// compiled with some debug information (at minimum, line numbers:
/// -gline-tables-only). Otherwise, it will be impossible to match IR
/// instructions to the line numbers collected by the profiler.
///
/// From the profile file, we are interested in collecting the
/// following information:
///
/// * A list of functions included in the profile (mangled names).
///
/// * For each function F:
///   1. The total number of samples collected in F.
///
///   2. The samples collected at each line in F. To provide some
///      protection against source code shuffling, line numbers should
///      be relative to the start of the function.
class SampleModuleProfile {
public:
  SampleModuleProfile(StringRef F) : Profiles(0), Filename(F) {}

  void dump();
  void loadText();
  void loadNative() { llvm_unreachable("not implemented"); }
  void printFunctionProfile(raw_ostream &OS, StringRef FName);
  void dumpFunctionProfile(StringRef FName);
  SampleFunctionProfile &getProfile(const Function &F) {
    return Profiles[F.getName()];
  }

  /// \brief Report a parse error message and stop compilation.
  void reportParseError(int64_t LineNumber, Twine Msg) const {
    report_fatal_error(Filename + ":" + Twine(LineNumber) + ": " + Msg + "\n");
  }

protected:
  /// \brief Map every function to its associated profile.
  ///
  /// The profile of every function executed at runtime is collected
  /// in the structure SampleFunctionProfile. This maps function objects
  /// to their corresponding profiles.
  StringMap<SampleFunctionProfile> Profiles;

  /// \brief Path name to the file holding the profile data.
  ///
  /// The format of this file is defined by each profiler
  /// independently. If possible, the profiler should have a text
  /// version of the profile format to be used in constructing test
  /// cases and debugging.
  StringRef Filename;
};

/// \brief Sample profile pass.
///
/// This pass reads profile data from the file specified by
/// -sample-profile-file and annotates every affected function with the
/// profile information found in that file.
class SampleProfileLoader : public FunctionPass {
public:
  // Class identification, replacement for typeinfo
  static char ID;

  SampleProfileLoader(StringRef Name = SampleProfileFile)
      : FunctionPass(ID), Profiler(), Filename(Name) {
    initializeSampleProfileLoaderPass(*PassRegistry::getPassRegistry());
  }

  bool doInitialization(Module &M) override;

  void dump() { Profiler->dump(); }

  const char *getPassName() const override { return "Sample profile pass"; }

  bool runOnFunction(Function &F) override;

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.setPreservesCFG();
    AU.addRequired<LoopInfo>();
    AU.addRequired<DominatorTreeWrapperPass>();
    AU.addRequired<PostDominatorTree>();
  }

protected:
  /// \brief Profile reader object.
  std::unique_ptr<SampleModuleProfile> Profiler;

  /// \brief Name of the profile file to load.
  StringRef Filename;
};
}

/// \brief Print this function profile on stream \p OS.
///
/// \param OS Stream to emit the output to.
void SampleFunctionProfile::print(raw_ostream &OS) {
  OS << TotalSamples << ", " << TotalHeadSamples << ", " << BodySamples.size()
     << " sampled lines\n";
  for (BodySampleMap::const_iterator SI = BodySamples.begin(),
                                     SE = BodySamples.end();
       SI != SE; ++SI)
    OS << "\tline offset: " << SI->first.LineOffset
       << ", discriminator: " << SI->first.Discriminator
       << ", number of samples: " << SI->second << "\n";
  OS << "\n";
}

/// \brief Print the weight of edge \p E on stream \p OS.
///
/// \param OS  Stream to emit the output to.
/// \param E  Edge to print.
void SampleFunctionProfile::printEdgeWeight(raw_ostream &OS, Edge E) {
  OS << "weight[" << E.first->getName() << "->" << E.second->getName()
     << "]: " << EdgeWeights[E] << "\n";
}

/// \brief Print the equivalence class of block \p BB on stream \p OS.
///
/// \param OS  Stream to emit the output to.
/// \param BB  Block to print.
void SampleFunctionProfile::printBlockEquivalence(raw_ostream &OS,
                                                  BasicBlock *BB) {
  BasicBlock *Equiv = EquivalenceClass[BB];
  OS << "equivalence[" << BB->getName()
     << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
}

/// \brief Print the weight of block \p BB on stream \p OS.
///
/// \param OS  Stream to emit the output to.
/// \param BB  Block to print.
void SampleFunctionProfile::printBlockWeight(raw_ostream &OS, BasicBlock *BB) {
  OS << "weight[" << BB->getName() << "]: " << BlockWeights[BB] << "\n";
}

/// \brief Print the function profile for \p FName on stream \p OS.
///
/// \param OS Stream to emit the output to.
/// \param FName Name of the function to print.
void SampleModuleProfile::printFunctionProfile(raw_ostream &OS,
                                               StringRef FName) {
  OS << "Function: " << FName << ":\n";
  Profiles[FName].print(OS);
}

/// \brief Dump the function profile for \p FName.
///
/// \param FName Name of the function to print.
void SampleModuleProfile::dumpFunctionProfile(StringRef FName) {
  printFunctionProfile(dbgs(), FName);
}

/// \brief Dump all the function profiles found.
void SampleModuleProfile::dump() {
  for (StringMap<SampleFunctionProfile>::const_iterator I = Profiles.begin(),
                                                        E = Profiles.end();
       I != E; ++I)
    dumpFunctionProfile(I->getKey());
}

/// \brief Load samples from a text file.
///
/// The file contains a list of samples for every function executed at
/// runtime. Each function profile has the following format:
///
///    function1:total_samples:total_head_samples
///    offset1[.discriminator]: number_of_samples [fn1:num fn2:num ... ]
///    offset2[.discriminator]: number_of_samples [fn3:num fn4:num ... ]
///    ...
///    offsetN[.discriminator]: number_of_samples [fn5:num fn6:num ... ]
///
/// Function names must be mangled in order for the profile loader to
/// match them in the current translation unit. The two numbers in the
/// function header specify how many total samples were accumulated in
/// the function (first number), and the total number of samples accumulated
/// at the prologue of the function (second number). This head sample
/// count provides an indicator of how frequent is the function invoked.
///
/// Each sampled line may contain several items. Some are optional
/// (marked below):
///
/// a- Source line offset. This number represents the line number
///    in the function where the sample was collected. The line number
///    is always relative to the line where symbol of the function
///    is defined. So, if the function has its header at line 280,
///    the offset 13 is at line 293 in the file.
///
/// b- [OPTIONAL] Discriminator. This is used if the sampled program
///    was compiled with DWARF discriminator support
///    (http://wiki.dwarfstd.org/index.php?title=Path_Discriminators)
///
/// c- Number of samples. This is the number of samples collected by
///    the profiler at this source location.
///
/// d- [OPTIONAL] Potential call targets and samples. If present, this
///    line contains a call instruction. This models both direct and
///    indirect calls. Each called target is listed together with the
///    number of samples. For example,
///
///    130: 7  foo:3  bar:2  baz:7
///
///    The above means that at relative line offset 130 there is a
///    call instruction that calls one of foo(), bar() and baz(). With
///    baz() being the relatively more frequent call target.
///
///    FIXME: This is currently unhandled, but it has a lot of
///           potential for aiding the inliner.
///
///
/// Since this is a flat profile, a function that shows up more than
/// once gets all its samples aggregated across all its instances.
///
/// FIXME: flat profiles are too imprecise to provide good optimization
///        opportunities. Convert them to context-sensitive profile.
///
/// This textual representation is useful to generate unit tests and
/// for debugging purposes, but it should not be used to generate
/// profiles for large programs, as the representation is extremely
/// inefficient.
void SampleModuleProfile::loadText() {
  std::unique_ptr<MemoryBuffer> Buffer;
  error_code EC = MemoryBuffer::getFile(Filename, Buffer);
  if (EC)
    report_fatal_error("Could not open file " + Filename + ": " + EC.message());
  line_iterator LineIt(*Buffer, '#');

  // Read the profile of each function. Since each function may be
  // mentioned more than once, and we are collecting flat profiles,
  // accumulate samples as we parse them.
  Regex HeadRE("^([^:]+):([0-9]+):([0-9]+)$");
  Regex LineSample("^([0-9]+)\\.?([0-9]+)?: ([0-9]+)(.*)$");
  while (!LineIt.is_at_eof()) {
    // Read the header of each function. The function header should
    // have this format:
    //
    //        function_name:total_samples:total_head_samples
    //
    // See above for an explanation of each field.
    SmallVector<StringRef, 3> Matches;
    if (!HeadRE.match(*LineIt, &Matches))
      reportParseError(LineIt.line_number(),
                       "Expected 'mangled_name:NUM:NUM', found " + *LineIt);
    assert(Matches.size() == 4);
    StringRef FName = Matches[1];
    unsigned NumSamples, NumHeadSamples;
    Matches[2].getAsInteger(10, NumSamples);
    Matches[3].getAsInteger(10, NumHeadSamples);
    Profiles[FName] = SampleFunctionProfile();
    SampleFunctionProfile &FProfile = Profiles[FName];
    FProfile.addTotalSamples(NumSamples);
    FProfile.addHeadSamples(NumHeadSamples);
    ++LineIt;

    // Now read the body. The body of the function ends when we reach
    // EOF or when we see the start of the next function.
    while (!LineIt.is_at_eof() && isdigit((*LineIt)[0])) {
      if (!LineSample.match(*LineIt, &Matches))
        reportParseError(
            LineIt.line_number(),
            "Expected 'NUM[.NUM]: NUM[ mangled_name:NUM]*', found " + *LineIt);
      assert(Matches.size() == 5);
      unsigned LineOffset, NumSamples, Discriminator = 0;
      Matches[1].getAsInteger(10, LineOffset);
      if (Matches[2] != "")
        Matches[2].getAsInteger(10, Discriminator);
      Matches[3].getAsInteger(10, NumSamples);

      // FIXME: Handle called targets (in Matches[4]).

      // When dealing with instruction weights, we use the value
      // zero to indicate the absence of a sample. If we read an
      // actual zero from the profile file, return it as 1 to
      // avoid the confusion later on.
      if (NumSamples == 0)
        NumSamples = 1;
      FProfile.addBodySamples(LineOffset, Discriminator, NumSamples);
      ++LineIt;
    }
  }
}

/// \brief Get the weight for an instruction.
///
/// The "weight" of an instruction \p Inst is the number of samples
/// collected on that instruction at runtime. To retrieve it, we
/// need to compute the line number of \p Inst relative to the start of its
/// function. We use HeaderLineno to compute the offset. We then
/// look up the samples collected for \p Inst using BodySamples.
///
/// \param Inst Instruction to query.
///
/// \returns The profiled weight of I.
unsigned SampleFunctionProfile::getInstWeight(Instruction &Inst) {
  DebugLoc DLoc = Inst.getDebugLoc();
  unsigned Lineno = DLoc.getLine();
  if (Lineno < HeaderLineno)
    return 0;

  DILocation DIL(DLoc.getAsMDNode(*Ctx));
  int LOffset = Lineno - HeaderLineno;
  unsigned Discriminator = DIL.getDiscriminator();
  unsigned Weight =
      BodySamples.lookup(InstructionLocation(LOffset, Discriminator));
  DEBUG(dbgs() << "    " << Lineno << "." << Discriminator << ":" << Inst
               << " (line offset: " << LOffset << "." << Discriminator
               << " - weight: " << Weight << ")\n");
  return Weight;
}

/// \brief Compute the weight of a basic block.
///
/// The weight of basic block \p B is the maximum weight of all the
/// instructions in B. The weight of \p B is computed and cached in
/// the BlockWeights map.
///
/// \param B The basic block to query.
///
/// \returns The computed weight of B.
unsigned SampleFunctionProfile::getBlockWeight(BasicBlock *B) {
  // If we've computed B's weight before, return it.
  std::pair<BlockWeightMap::iterator, bool> Entry =
      BlockWeights.insert(std::make_pair(B, 0));
  if (!Entry.second)
    return Entry.first->second;

  // Otherwise, compute and cache B's weight.
  unsigned Weight = 0;
  for (BasicBlock::iterator I = B->begin(), E = B->end(); I != E; ++I) {
    unsigned InstWeight = getInstWeight(*I);
    if (InstWeight > Weight)
      Weight = InstWeight;
  }
  Entry.first->second = Weight;
  return Weight;
}

/// \brief Compute and store the weights of every basic block.
///
/// This populates the BlockWeights map by computing
/// the weights of every basic block in the CFG.
///
/// \param F The function to query.
bool SampleFunctionProfile::computeBlockWeights(Function &F) {
  bool Changed = false;
  DEBUG(dbgs() << "Block weights\n");
  for (Function::iterator B = F.begin(), E = F.end(); B != E; ++B) {
    unsigned Weight = getBlockWeight(B);
    Changed |= (Weight > 0);
    DEBUG(printBlockWeight(dbgs(), B));
  }

  return Changed;
}

/// \brief Find equivalence classes for the given block.
///
/// This finds all the blocks that are guaranteed to execute the same
/// number of times as \p BB1. To do this, it traverses all the the
/// descendants of \p BB1 in the dominator or post-dominator tree.
///
/// A block BB2 will be in the same equivalence class as \p BB1 if
/// the following holds:
///
/// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
///    is a descendant of \p BB1 in the dominator tree, then BB2 should
///    dominate BB1 in the post-dominator tree.
///
/// 2- Both BB2 and \p BB1 must be in the same loop.
///
/// For every block BB2 that meets those two requirements, we set BB2's
/// equivalence class to \p BB1.
///
/// \param BB1  Block to check.
/// \param Descendants  Descendants of \p BB1 in either the dom or pdom tree.
/// \param DomTree  Opposite dominator tree. If \p Descendants is filled
///                 with blocks from \p BB1's dominator tree, then
///                 this is the post-dominator tree, and vice versa.
void SampleFunctionProfile::findEquivalencesFor(
    BasicBlock *BB1, SmallVector<BasicBlock *, 8> Descendants,
    DominatorTreeBase<BasicBlock> *DomTree) {
  for (SmallVectorImpl<BasicBlock *>::iterator I = Descendants.begin(),
                                               E = Descendants.end();
       I != E; ++I) {
    BasicBlock *BB2 = *I;
    bool IsDomParent = DomTree->dominates(BB2, BB1);
    bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
    if (BB1 != BB2 && VisitedBlocks.insert(BB2) && IsDomParent &&
        IsInSameLoop) {
      EquivalenceClass[BB2] = BB1;

      // If BB2 is heavier than BB1, make BB2 have the same weight
      // as BB1.
      //
      // Note that we don't worry about the opposite situation here
      // (when BB2 is lighter than BB1). We will deal with this
      // during the propagation phase. Right now, we just want to
      // make sure that BB1 has the largest weight of all the
      // members of its equivalence set.
      unsigned &BB1Weight = BlockWeights[BB1];
      unsigned &BB2Weight = BlockWeights[BB2];
      BB1Weight = std::max(BB1Weight, BB2Weight);
    }
  }
}

/// \brief Find equivalence classes.
///
/// Since samples may be missing from blocks, we can fill in the gaps by setting
/// the weights of all the blocks in the same equivalence class to the same
/// weight. To compute the concept of equivalence, we use dominance and loop
/// information. Two blocks B1 and B2 are in the same equivalence class if B1
/// dominates B2, B2 post-dominates B1 and both are in the same loop.
///
/// \param F The function to query.
void SampleFunctionProfile::findEquivalenceClasses(Function &F) {
  SmallVector<BasicBlock *, 8> DominatedBBs;
  DEBUG(dbgs() << "\nBlock equivalence classes\n");
  // Find equivalence sets based on dominance and post-dominance information.
  for (Function::iterator B = F.begin(), E = F.end(); B != E; ++B) {
    BasicBlock *BB1 = B;

    // Compute BB1's equivalence class once.
    if (EquivalenceClass.count(BB1)) {
      DEBUG(printBlockEquivalence(dbgs(), BB1));
      continue;
    }

    // By default, blocks are in their own equivalence class.
    EquivalenceClass[BB1] = BB1;

    // Traverse all the blocks dominated by BB1. We are looking for
    // every basic block BB2 such that:
    //
    // 1- BB1 dominates BB2.
    // 2- BB2 post-dominates BB1.
    // 3- BB1 and BB2 are in the same loop nest.
    //
    // If all those conditions hold, it means that BB2 is executed
    // as many times as BB1, so they are placed in the same equivalence
    // class by making BB2's equivalence class be BB1.
    DominatedBBs.clear();
    DT->getDescendants(BB1, DominatedBBs);
    findEquivalencesFor(BB1, DominatedBBs, PDT->DT);

    // Repeat the same logic for all the blocks post-dominated by BB1.
    // We are looking for every basic block BB2 such that:
    //
    // 1- BB1 post-dominates BB2.
    // 2- BB2 dominates BB1.
    // 3- BB1 and BB2 are in the same loop nest.
    //
    // If all those conditions hold, BB2's equivalence class is BB1.
    DominatedBBs.clear();
    PDT->getDescendants(BB1, DominatedBBs);
    findEquivalencesFor(BB1, DominatedBBs, DT);

    DEBUG(printBlockEquivalence(dbgs(), BB1));
  }

  // Assign weights to equivalence classes.
  //
  // All the basic blocks in the same equivalence class will execute
  // the same number of times. Since we know that the head block in
  // each equivalence class has the largest weight, assign that weight
  // to all the blocks in that equivalence class.
  DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
  for (Function::iterator B = F.begin(), E = F.end(); B != E; ++B) {
    BasicBlock *BB = B;
    BasicBlock *EquivBB = EquivalenceClass[BB];
    if (BB != EquivBB)
      BlockWeights[BB] = BlockWeights[EquivBB];
    DEBUG(printBlockWeight(dbgs(), BB));
  }
}

/// \brief Visit the given edge to decide if it has a valid weight.
///
/// If \p E has not been visited before, we copy to \p UnknownEdge
/// and increment the count of unknown edges.
///
/// \param E  Edge to visit.
/// \param NumUnknownEdges  Current number of unknown edges.
/// \param UnknownEdge  Set if E has not been visited before.
///
/// \returns E's weight, if known. Otherwise, return 0.
unsigned SampleFunctionProfile::visitEdge(Edge E, unsigned *NumUnknownEdges,
                                          Edge *UnknownEdge) {
  if (!VisitedEdges.count(E)) {
    (*NumUnknownEdges)++;
    *UnknownEdge = E;
    return 0;
  }

  return EdgeWeights[E];
}

/// \brief Propagate weights through incoming/outgoing edges.
///
/// If the weight of a basic block is known, and there is only one edge
/// with an unknown weight, we can calculate the weight of that edge.
///
/// Similarly, if all the edges have a known count, we can calculate the
/// count of the basic block, if needed.
///
/// \param F  Function to process.
///
/// \returns  True if new weights were assigned to edges or blocks.
bool SampleFunctionProfile::propagateThroughEdges(Function &F) {
  bool Changed = false;
  DEBUG(dbgs() << "\nPropagation through edges\n");
  for (Function::iterator BI = F.begin(), EI = F.end(); BI != EI; ++BI) {
    BasicBlock *BB = BI;

    // Visit all the predecessor and successor edges to determine
    // which ones have a weight assigned already. Note that it doesn't
    // matter that we only keep track of a single unknown edge. The
    // only case we are interested in handling is when only a single
    // edge is unknown (see setEdgeOrBlockWeight).
    for (unsigned i = 0; i < 2; i++) {
      unsigned TotalWeight = 0;
      unsigned NumUnknownEdges = 0;
      Edge UnknownEdge, SelfReferentialEdge;

      if (i == 0) {
        // First, visit all predecessor edges.
        for (size_t I = 0; I < Predecessors[BB].size(); I++) {
          Edge E = std::make_pair(Predecessors[BB][I], BB);
          TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
          if (E.first == E.second)
            SelfReferentialEdge = E;
        }
      } else {
        // On the second round, visit all successor edges.
        for (size_t I = 0; I < Successors[BB].size(); I++) {
          Edge E = std::make_pair(BB, Successors[BB][I]);
          TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
        }
      }

      // After visiting all the edges, there are three cases that we
      // can handle immediately:
      //
      // - All the edge weights are known (i.e., NumUnknownEdges == 0).
      //   In this case, we simply check that the sum of all the edges
      //   is the same as BB's weight. If not, we change BB's weight
      //   to match. Additionally, if BB had not been visited before,
      //   we mark it visited.
      //
      // - Only one edge is unknown and BB has already been visited.
      //   In this case, we can compute the weight of the edge by
      //   subtracting the total block weight from all the known
      //   edge weights. If the edges weight more than BB, then the
      //   edge of the last remaining edge is set to zero.
      //
      // - There exists a self-referential edge and the weight of BB is
      //   known. In this case, this edge can be based on BB's weight.
      //   We add up all the other known edges and set the weight on
      //   the self-referential edge as we did in the previous case.
      //
      // In any other case, we must continue iterating. Eventually,
      // all edges will get a weight, or iteration will stop when
      // it reaches SampleProfileMaxPropagateIterations.
      if (NumUnknownEdges <= 1) {
        unsigned &BBWeight = BlockWeights[BB];
        if (NumUnknownEdges == 0) {
          // If we already know the weight of all edges, the weight of the
          // basic block can be computed. It should be no larger than the sum
          // of all edge weights.
          if (TotalWeight > BBWeight) {
            BBWeight = TotalWeight;
            Changed = true;
            DEBUG(dbgs() << "All edge weights for " << BB->getName()
                         << " known. Set weight for block: ";
                  printBlockWeight(dbgs(), BB););
          }
          if (VisitedBlocks.insert(BB))
            Changed = true;
        } else if (NumUnknownEdges == 1 && VisitedBlocks.count(BB)) {
          // If there is a single unknown edge and the block has been
          // visited, then we can compute E's weight.
          if (BBWeight >= TotalWeight)
            EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
          else
            EdgeWeights[UnknownEdge] = 0;
          VisitedEdges.insert(UnknownEdge);
          Changed = true;
          DEBUG(dbgs() << "Set weight for edge: ";
                printEdgeWeight(dbgs(), UnknownEdge));
        }
      } else if (SelfReferentialEdge.first && VisitedBlocks.count(BB)) {
        unsigned &BBWeight = BlockWeights[BB];
        // We have a self-referential edge and the weight of BB is known.
        if (BBWeight >= TotalWeight)
          EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
        else
          EdgeWeights[SelfReferentialEdge] = 0;
        VisitedEdges.insert(SelfReferentialEdge);
        Changed = true;
        DEBUG(dbgs() << "Set self-referential edge weight to: ";
              printEdgeWeight(dbgs(), SelfReferentialEdge));
      }
    }
  }

  return Changed;
}

/// \brief Build in/out edge lists for each basic block in the CFG.
///
/// We are interested in unique edges. If a block B1 has multiple
/// edges to another block B2, we only add a single B1->B2 edge.
void SampleFunctionProfile::buildEdges(Function &F) {
  for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
    BasicBlock *B1 = I;

    // Add predecessors for B1.
    SmallPtrSet<BasicBlock *, 16> Visited;
    if (!Predecessors[B1].empty())
      llvm_unreachable("Found a stale predecessors list in a basic block.");
    for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
      BasicBlock *B2 = *PI;
      if (Visited.insert(B2))
        Predecessors[B1].push_back(B2);
    }

    // Add successors for B1.
    Visited.clear();
    if (!Successors[B1].empty())
      llvm_unreachable("Found a stale successors list in a basic block.");
    for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
      BasicBlock *B2 = *SI;
      if (Visited.insert(B2))
        Successors[B1].push_back(B2);
    }
  }
}

/// \brief Propagate weights into edges
///
/// The following rules are applied to every block B in the CFG:
///
/// - If B has a single predecessor/successor, then the weight
///   of that edge is the weight of the block.
///
/// - If all incoming or outgoing edges are known except one, and the
///   weight of the block is already known, the weight of the unknown
///   edge will be the weight of the block minus the sum of all the known
///   edges. If the sum of all the known edges is larger than B's weight,
///   we set the unknown edge weight to zero.
///
/// - If there is a self-referential edge, and the weight of the block is
///   known, the weight for that edge is set to the weight of the block
///   minus the weight of the other incoming edges to that block (if
///   known).
void SampleFunctionProfile::propagateWeights(Function &F) {
  bool Changed = true;
  unsigned i = 0;

  // Before propagation starts, build, for each block, a list of
  // unique predecessors and successors. This is necessary to handle
  // identical edges in multiway branches. Since we visit all blocks and all
  // edges of the CFG, it is cleaner to build these lists once at the start
  // of the pass.
  buildEdges(F);

  // Propagate until we converge or we go past the iteration limit.
  while (Changed && i++ < SampleProfileMaxPropagateIterations) {
    Changed = propagateThroughEdges(F);
  }

  // Generate MD_prof metadata for every branch instruction using the
  // edge weights computed during propagation.
  DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
  MDBuilder MDB(F.getContext());
  for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
    BasicBlock *B = I;
    TerminatorInst *TI = B->getTerminator();
    if (TI->getNumSuccessors() == 1)
      continue;
    if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
      continue;

    DEBUG(dbgs() << "\nGetting weights for branch at line "
                 << TI->getDebugLoc().getLine() << ".\n");
    SmallVector<unsigned, 4> Weights;
    bool AllWeightsZero = true;
    for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
      BasicBlock *Succ = TI->getSuccessor(I);
      Edge E = std::make_pair(B, Succ);
      unsigned Weight = EdgeWeights[E];
      DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
      Weights.push_back(Weight);
      if (Weight != 0)
        AllWeightsZero = false;
    }

    // Only set weights if there is at least one non-zero weight.
    // In any other case, let the analyzer set weights.
    if (!AllWeightsZero) {
      DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
      TI->setMetadata(llvm::LLVMContext::MD_prof,
                      MDB.createBranchWeights(Weights));
    } else {
      DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
    }
  }
}

/// \brief Get the line number for the function header.
///
/// This looks up function \p F in the current compilation unit and
/// retrieves the line number where the function is defined. This is
/// line 0 for all the samples read from the profile file. Every line
/// number is relative to this line.
///
/// \param F  Function object to query.
///
/// \returns the line number where \p F is defined.
unsigned SampleFunctionProfile::getFunctionLoc(Function &F) {
  NamedMDNode *CUNodes = F.getParent()->getNamedMetadata("llvm.dbg.cu");
  if (CUNodes) {
    for (unsigned I = 0, E1 = CUNodes->getNumOperands(); I != E1; ++I) {
      DICompileUnit CU(CUNodes->getOperand(I));
      DIArray Subprograms = CU.getSubprograms();
      for (unsigned J = 0, E2 = Subprograms.getNumElements(); J != E2; ++J) {
        DISubprogram Subprogram(Subprograms.getElement(J));
        if (Subprogram.describes(&F))
          return Subprogram.getLineNumber();
      }
    }
  }

  report_fatal_error("No debug information found in function " + F.getName() +
                     "\n");
}

/// \brief Generate branch weight metadata for all branches in \p F.
///
/// Branch weights are computed out of instruction samples using a
/// propagation heuristic. Propagation proceeds in 3 phases:
///
/// 1- Assignment of block weights. All the basic blocks in the function
///    are initial assigned the same weight as their most frequently
///    executed instruction.
///
/// 2- Creation of equivalence classes. Since samples may be missing from
///    blocks, we can fill in the gaps by setting the weights of all the
///    blocks in the same equivalence class to the same weight. To compute
///    the concept of equivalence, we use dominance and loop information.
///    Two blocks B1 and B2 are in the same equivalence class if B1
///    dominates B2, B2 post-dominates B1 and both are in the same loop.
///
/// 3- Propagation of block weights into edges. This uses a simple
///    propagation heuristic. The following rules are applied to every
///    block B in the CFG:
///
///    - If B has a single predecessor/successor, then the weight
///      of that edge is the weight of the block.
///
///    - If all the edges are known except one, and the weight of the
///      block is already known, the weight of the unknown edge will
///      be the weight of the block minus the sum of all the known
///      edges. If the sum of all the known edges is larger than B's weight,
///      we set the unknown edge weight to zero.
///
///    - If there is a self-referential edge, and the weight of the block is
///      known, the weight for that edge is set to the weight of the block
///      minus the weight of the other incoming edges to that block (if
///      known).
///
/// Since this propagation is not guaranteed to finalize for every CFG, we
/// only allow it to proceed for a limited number of iterations (controlled
/// by -sample-profile-max-propagate-iterations).
///
/// FIXME: Try to replace this propagation heuristic with a scheme
/// that is guaranteed to finalize. A work-list approach similar to
/// the standard value propagation algorithm used by SSA-CCP might
/// work here.
///
/// Once all the branch weights are computed, we emit the MD_prof
/// metadata on B using the computed values for each of its branches.
///
/// \param F The function to query.
bool SampleFunctionProfile::emitAnnotations(Function &F, DominatorTree *DomTree,
                                            PostDominatorTree *PostDomTree,
                                            LoopInfo *Loops) {
  bool Changed = false;

  // Initialize invariants used during computation and propagation.
  HeaderLineno = getFunctionLoc(F);
  DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
               << ": " << HeaderLineno << "\n");
  DT = DomTree;
  PDT = PostDomTree;
  LI = Loops;
  Ctx = &F.getParent()->getContext();

  // Compute basic block weights.
  Changed |= computeBlockWeights(F);

  if (Changed) {
    // Find equivalence classes.
    findEquivalenceClasses(F);

    // Propagate weights to all edges.
    propagateWeights(F);
  }

  return Changed;
}

char SampleProfileLoader::ID = 0;
INITIALIZE_PASS_BEGIN(SampleProfileLoader, "sample-profile",
                      "Sample Profile loader", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(PostDominatorTree)
INITIALIZE_PASS_DEPENDENCY(LoopInfo)
INITIALIZE_PASS_DEPENDENCY(AddDiscriminators)
INITIALIZE_PASS_END(SampleProfileLoader, "sample-profile",
                    "Sample Profile loader", false, false)

bool SampleProfileLoader::doInitialization(Module &M) {
  Profiler.reset(new SampleModuleProfile(Filename));
  Profiler->loadText();
  return true;
}

FunctionPass *llvm::createSampleProfileLoaderPass() {
  return new SampleProfileLoader(SampleProfileFile);
}

FunctionPass *llvm::createSampleProfileLoaderPass(StringRef Name) {
  return new SampleProfileLoader(Name);
}

bool SampleProfileLoader::runOnFunction(Function &F) {
  DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
  PostDominatorTree *PDT = &getAnalysis<PostDominatorTree>();
  LoopInfo *LI = &getAnalysis<LoopInfo>();
  SampleFunctionProfile &FunctionProfile = Profiler->getProfile(F);
  if (!FunctionProfile.empty())
    return FunctionProfile.emitAnnotations(F, DT, PDT, LI);
  return false;
}