llvm/lib/Transforms/IPO/PartialInlining.cpp

//===- PartialInlining.cpp - Inline parts of functions --------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass performs partial inlining, typically by inlining an if statement
// that surrounds the body of the function.
//
//===----------------------------------------------------------------------===//

#include "llvm/Transforms/IPO/PartialInlining.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/None.h"
#include "llvm/ADT/Optional.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/BlockFrequencyInfo.h"
#include "llvm/Analysis/BranchProbabilityInfo.h"
#include "llvm/Analysis/InlineCost.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/ProfileSummaryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/IR/Attributes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/CallSite.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/User.h"
#include "llvm/Pass.h"
#include "llvm/Support/BlockFrequency.h"
#include "llvm/Support/BranchProbability.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Transforms/IPO.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Transforms/Utils/CodeExtractor.h"
#include "llvm/Transforms/Utils/ValueMapper.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <functional>
#include <iterator>
#include <memory>
#include <tuple>
#include <vector>

using namespace llvm;

#define DEBUG_TYPE "partial-inlining"

STATISTIC(NumPartialInlined,
          "Number of callsites functions partially inlined into.");

// Command line option to disable partial-inlining. The default is false:
static cl::opt<bool>
    DisablePartialInlining("disable-partial-inlining", cl::init(false),
                           cl::Hidden, cl::desc("Disable partial ininling"));

// This is an option used by testing:
static cl::opt<bool> SkipCostAnalysis("skip-partial-inlining-cost-analysis",
                                      cl::init(false), cl::ZeroOrMore,
                                      cl::ReallyHidden,
                                      cl::desc("Skip Cost Analysis"));

static cl::opt<unsigned> MaxNumInlineBlocks(
    "max-num-inline-blocks", cl::init(5), cl::Hidden,
    cl::desc("Max number of blocks to be partially inlined"));

// Command line option to set the maximum number of partial inlining allowed
// for the module. The default value of -1 means no limit.
static cl::opt<int> MaxNumPartialInlining(
    "max-partial-inlining", cl::init(-1), cl::Hidden, cl::ZeroOrMore,
    cl::desc("Max number of partial inlining. The default is unlimited"));

// Used only when PGO or user annotated branch data is absent. It is
// the least value that is used to weigh the outline region. If BFI
// produces larger value, the BFI value will be used.
static cl::opt<int>
    OutlineRegionFreqPercent("outline-region-freq-percent", cl::init(75),
                             cl::Hidden, cl::ZeroOrMore,
                             cl::desc("Relative frequency of outline region to "
                                      "the entry block"));

static cl::opt<unsigned> ExtraOutliningPenalty(
    "partial-inlining-extra-penalty", cl::init(0), cl::Hidden,
    cl::desc("A debug option to add additional penalty to the computed one."));

namespace {

struct FunctionOutliningInfo {
  FunctionOutliningInfo() = default;

  // Returns the number of blocks to be inlined including all blocks
  // in Entries and one return block.
  unsigned GetNumInlinedBlocks() const { return Entries.size() + 1; }

  // A set of blocks including the function entry that guard
  // the region to be outlined.
  SmallVector<BasicBlock *, 4> Entries;

  // The return block that is not included in the outlined region.
  BasicBlock *ReturnBlock = nullptr;

  // The dominating block of the region to be outlined.
  BasicBlock *NonReturnBlock = nullptr;

  // The set of blocks in Entries that that are predecessors to ReturnBlock
  SmallVector<BasicBlock *, 4> ReturnBlockPreds;
};

struct PartialInlinerImpl {
  PartialInlinerImpl(
      std::function<AssumptionCache &(Function &)> *GetAC,
      std::function<TargetTransformInfo &(Function &)> *GTTI,
      Optional<function_ref<BlockFrequencyInfo &(Function &)>> GBFI,
      ProfileSummaryInfo *ProfSI)
      : GetAssumptionCache(GetAC), GetTTI(GTTI), GetBFI(GBFI), PSI(ProfSI) {}

  bool run(Module &M);
  Function *unswitchFunction(Function *F);

  // This class speculatively clones the the function to be partial inlined.
  // At the end of partial inlining, the remaining callsites to the cloned
  // function that are not partially inlined will be fixed up to reference
  // the original function, and the cloned function will be erased.
  struct FunctionCloner {
    FunctionCloner(Function *F, FunctionOutliningInfo *OI);
    ~FunctionCloner();

    // Prepare for function outlining: making sure there is only
    // one incoming edge from the extracted/outlined region to
    // the return block.
    void NormalizeReturnBlock();

    // Do function outlining:
    Function *doFunctionOutlining();

    Function *OrigFunc = nullptr;
    Function *ClonedFunc = nullptr;
    Function *OutlinedFunc = nullptr;
    BasicBlock *OutliningCallBB = nullptr;
    // ClonedFunc is inlined in one of its callers after function
    // outlining.
    bool IsFunctionInlined = false;
    // The cost of the region to be outlined.
    int OutlinedRegionCost = 0;
    std::unique_ptr<FunctionOutliningInfo> ClonedOI = nullptr;
    std::unique_ptr<BlockFrequencyInfo> ClonedFuncBFI = nullptr;
  };

private:
  int NumPartialInlining = 0;
  std::function<AssumptionCache &(Function &)> *GetAssumptionCache;
  std::function<TargetTransformInfo &(Function &)> *GetTTI;
  Optional<function_ref<BlockFrequencyInfo &(Function &)>> GetBFI;
  ProfileSummaryInfo *PSI;

  // Return the frequency of the OutlininingBB relative to F's entry point.
  // The result is no larger than 1 and is represented using BP.
  // (Note that the outlined region's 'head' block can only have incoming
  // edges from the guarding entry blocks).
  BranchProbability getOutliningCallBBRelativeFreq(FunctionCloner &Cloner);

  // Return true if the callee of CS should be partially inlined with
  // profit.
  bool shouldPartialInline(CallSite CS, FunctionCloner &Cloner,
                           BlockFrequency WeightedOutliningRcost,
                           OptimizationRemarkEmitter &ORE);

  // Try to inline DuplicateFunction (cloned from F with call to
  // the OutlinedFunction into its callers. Return true
  // if there is any successful inlining.
  bool tryPartialInline(FunctionCloner &Cloner);

  // Compute the mapping from use site of DuplicationFunction to the enclosing
  // BB's profile count.
  void computeCallsiteToProfCountMap(Function *DuplicateFunction,
                                     DenseMap<User *, uint64_t> &SiteCountMap);

  bool IsLimitReached() {
    return (MaxNumPartialInlining != -1 &&
            NumPartialInlining >= MaxNumPartialInlining);
  }

  static CallSite getCallSite(User *U) {
    CallSite CS;
    if (CallInst *CI = dyn_cast<CallInst>(U))
      CS = CallSite(CI);
    else if (InvokeInst *II = dyn_cast<InvokeInst>(U))
      CS = CallSite(II);
    else
      llvm_unreachable("All uses must be calls");
    return CS;
  }

  static CallSite getOneCallSiteTo(Function *F) {
    User *User = *F->user_begin();
    return getCallSite(User);
  }

  std::tuple<DebugLoc, BasicBlock *> getOneDebugLoc(Function *F) {
    CallSite CS = getOneCallSiteTo(F);
    DebugLoc DLoc = CS.getInstruction()->getDebugLoc();
    BasicBlock *Block = CS.getParent();
    return std::make_tuple(DLoc, Block);
  }

  // Returns the costs associated with function outlining:
  // - The first value is the non-weighted runtime cost for making the call
  //   to the outlined function, including the addtional  setup cost in the
  //    outlined function itself;
  // - The second value is the estimated size of the new call sequence in
  //   basic block Cloner.OutliningCallBB;
  std::tuple<int, int> computeOutliningCosts(FunctionCloner &Cloner);

  // Compute the 'InlineCost' of block BB. InlineCost is a proxy used to
  // approximate both the size and runtime cost (Note that in the current
  // inline cost analysis, there is no clear distinction there either).
  static int computeBBInlineCost(BasicBlock *BB);

  std::unique_ptr<FunctionOutliningInfo> computeOutliningInfo(Function *F);
};

struct PartialInlinerLegacyPass : public ModulePass {
  static char ID; // Pass identification, replacement for typeid

  PartialInlinerLegacyPass() : ModulePass(ID) {
    initializePartialInlinerLegacyPassPass(*PassRegistry::getPassRegistry());
  }

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addRequired<AssumptionCacheTracker>();
    AU.addRequired<ProfileSummaryInfoWrapperPass>();
    AU.addRequired<TargetTransformInfoWrapperPass>();
  }

  bool runOnModule(Module &M) override {
    if (skipModule(M))
      return false;

    AssumptionCacheTracker *ACT = &getAnalysis<AssumptionCacheTracker>();
    TargetTransformInfoWrapperPass *TTIWP =
        &getAnalysis<TargetTransformInfoWrapperPass>();
    ProfileSummaryInfo *PSI =
        getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();

    std::function<AssumptionCache &(Function &)> GetAssumptionCache =
        [&ACT](Function &F) -> AssumptionCache & {
      return ACT->getAssumptionCache(F);
    };

    std::function<TargetTransformInfo &(Function &)> GetTTI =
        [&TTIWP](Function &F) -> TargetTransformInfo & {
      return TTIWP->getTTI(F);
    };

    return PartialInlinerImpl(&GetAssumptionCache, &GetTTI, None, PSI).run(M);
  }
};

} // end anonymous namespace

std::unique_ptr<FunctionOutliningInfo>
PartialInlinerImpl::computeOutliningInfo(Function *F) {
  BasicBlock *EntryBlock = &F->front();
  BranchInst *BR = dyn_cast<BranchInst>(EntryBlock->getTerminator());
  if (!BR || BR->isUnconditional())
    return std::unique_ptr<FunctionOutliningInfo>();

  // Returns true if Succ is BB's successor
  auto IsSuccessor = [](BasicBlock *Succ, BasicBlock *BB) {
    return is_contained(successors(BB), Succ);
  };

  auto SuccSize = [](BasicBlock *BB) {
    return std::distance(succ_begin(BB), succ_end(BB));
  };

  auto IsReturnBlock = [](BasicBlock *BB) {
    TerminatorInst *TI = BB->getTerminator();
    return isa<ReturnInst>(TI);
  };

  auto GetReturnBlock = [&](BasicBlock *Succ1, BasicBlock *Succ2) {
    if (IsReturnBlock(Succ1))
      return std::make_tuple(Succ1, Succ2);
    if (IsReturnBlock(Succ2))
      return std::make_tuple(Succ2, Succ1);

    return std::make_tuple<BasicBlock *, BasicBlock *>(nullptr, nullptr);
  };

  // Detect a triangular shape:
  auto GetCommonSucc = [&](BasicBlock *Succ1, BasicBlock *Succ2) {
    if (IsSuccessor(Succ1, Succ2))
      return std::make_tuple(Succ1, Succ2);
    if (IsSuccessor(Succ2, Succ1))
      return std::make_tuple(Succ2, Succ1);

    return std::make_tuple<BasicBlock *, BasicBlock *>(nullptr, nullptr);
  };

  std::unique_ptr<FunctionOutliningInfo> OutliningInfo =
      llvm::make_unique<FunctionOutliningInfo>();

  BasicBlock *CurrEntry = EntryBlock;
  bool CandidateFound = false;
  do {
    // The number of blocks to be inlined has already reached
    // the limit. When MaxNumInlineBlocks is set to 0 or 1, this
    // disables partial inlining for the function.
    if (OutliningInfo->GetNumInlinedBlocks() >= MaxNumInlineBlocks)
      break;

    if (SuccSize(CurrEntry) != 2)
      break;

    BasicBlock *Succ1 = *succ_begin(CurrEntry);
    BasicBlock *Succ2 = *(succ_begin(CurrEntry) + 1);

    BasicBlock *ReturnBlock, *NonReturnBlock;
    std::tie(ReturnBlock, NonReturnBlock) = GetReturnBlock(Succ1, Succ2);

    if (ReturnBlock) {
      OutliningInfo->Entries.push_back(CurrEntry);
      OutliningInfo->ReturnBlock = ReturnBlock;
      OutliningInfo->NonReturnBlock = NonReturnBlock;
      CandidateFound = true;
      break;
    }

    BasicBlock *CommSucc;
    BasicBlock *OtherSucc;
    std::tie(CommSucc, OtherSucc) = GetCommonSucc(Succ1, Succ2);

    if (!CommSucc)
      break;

    OutliningInfo->Entries.push_back(CurrEntry);
    CurrEntry = OtherSucc;
  } while (true);

  if (!CandidateFound)
    return std::unique_ptr<FunctionOutliningInfo>();

  // Do sanity check of the entries: threre should not
  // be any successors (not in the entry set) other than
  // {ReturnBlock, NonReturnBlock}
  assert(OutliningInfo->Entries[0] == &F->front() &&
         "Function Entry must be the first in Entries vector");
  DenseSet<BasicBlock *> Entries;
  for (BasicBlock *E : OutliningInfo->Entries)
    Entries.insert(E);

  // Returns true of BB has Predecessor which is not
  // in Entries set.
  auto HasNonEntryPred = [Entries](BasicBlock *BB) {
    for (auto Pred : predecessors(BB)) {
      if (!Entries.count(Pred))
        return true;
    }
    return false;
  };
  auto CheckAndNormalizeCandidate =
      [Entries, HasNonEntryPred](FunctionOutliningInfo *OutliningInfo) {
        for (BasicBlock *E : OutliningInfo->Entries) {
          for (auto Succ : successors(E)) {
            if (Entries.count(Succ))
              continue;
            if (Succ == OutliningInfo->ReturnBlock)
              OutliningInfo->ReturnBlockPreds.push_back(E);
            else if (Succ != OutliningInfo->NonReturnBlock)
              return false;
          }
          // There should not be any outside incoming edges either:
          if (HasNonEntryPred(E))
            return false;
        }
        return true;
      };

  if (!CheckAndNormalizeCandidate(OutliningInfo.get()))
    return std::unique_ptr<FunctionOutliningInfo>();

  // Now further growing the candidate's inlining region by
  // peeling off dominating blocks from the outlining region:
  while (OutliningInfo->GetNumInlinedBlocks() < MaxNumInlineBlocks) {
    BasicBlock *Cand = OutliningInfo->NonReturnBlock;
    if (SuccSize(Cand) != 2)
      break;

    if (HasNonEntryPred(Cand))
      break;

    BasicBlock *Succ1 = *succ_begin(Cand);
    BasicBlock *Succ2 = *(succ_begin(Cand) + 1);

    BasicBlock *ReturnBlock, *NonReturnBlock;
    std::tie(ReturnBlock, NonReturnBlock) = GetReturnBlock(Succ1, Succ2);
    if (!ReturnBlock || ReturnBlock != OutliningInfo->ReturnBlock)
      break;

    if (NonReturnBlock->getSinglePredecessor() != Cand)
      break;

    // Now grow and update OutlininigInfo:
    OutliningInfo->Entries.push_back(Cand);
    OutliningInfo->NonReturnBlock = NonReturnBlock;
    OutliningInfo->ReturnBlockPreds.push_back(Cand);
    Entries.insert(Cand);
  }

  return OutliningInfo;
}

// Check if there is PGO data or user annoated branch data:
static bool hasProfileData(Function *F, FunctionOutliningInfo *OI) {
  if (F->getEntryCount())
    return true;
  // Now check if any of the entry block has MD_prof data:
  for (auto *E : OI->Entries) {
    BranchInst *BR = dyn_cast<BranchInst>(E->getTerminator());
    if (!BR || BR->isUnconditional())
      continue;
    uint64_t T, F;
    if (BR->extractProfMetadata(T, F))
      return true;
  }
  return false;
}

BranchProbability
PartialInlinerImpl::getOutliningCallBBRelativeFreq(FunctionCloner &Cloner) {
  auto EntryFreq =
      Cloner.ClonedFuncBFI->getBlockFreq(&Cloner.ClonedFunc->getEntryBlock());
  auto OutliningCallFreq =
      Cloner.ClonedFuncBFI->getBlockFreq(Cloner.OutliningCallBB);

  auto OutlineRegionRelFreq =
      BranchProbability::getBranchProbability(OutliningCallFreq.getFrequency(),
                                              EntryFreq.getFrequency());

  if (hasProfileData(Cloner.OrigFunc, Cloner.ClonedOI.get()))
    return OutlineRegionRelFreq;

  // When profile data is not available, we need to be conservative in
  // estimating the overall savings. Static branch prediction can usually
  // guess the branch direction right (taken/non-taken), but the guessed
  // branch probability is usually not biased enough. In case when the
  // outlined region is predicted to be likely, its probability needs
  // to be made higher (more biased) to not under-estimate the cost of
  // function outlining. On the other hand, if the outlined region
  // is predicted to be less likely, the predicted probablity is usually
  // higher than the actual. For instance, the actual probability of the
  // less likely target is only 5%, but the guessed probablity can be
  // 40%. In the latter case, there is no need for further adjustement.
  // FIXME: add an option for this.
  if (OutlineRegionRelFreq < BranchProbability(45, 100))
    return OutlineRegionRelFreq;

  OutlineRegionRelFreq = std::max(
      OutlineRegionRelFreq, BranchProbability(OutlineRegionFreqPercent, 100));

  return OutlineRegionRelFreq;
}

bool PartialInlinerImpl::shouldPartialInline(
    CallSite CS, FunctionCloner &Cloner, BlockFrequency WeightedOutliningRcost,
    OptimizationRemarkEmitter &ORE) {
  using namespace ore;

  if (SkipCostAnalysis)
    return true;

  Instruction *Call = CS.getInstruction();
  Function *Callee = CS.getCalledFunction();
  assert(Callee == Cloner.ClonedFunc);

  Function *Caller = CS.getCaller();
  auto &CalleeTTI = (*GetTTI)(*Callee);
  InlineCost IC = getInlineCost(CS, getInlineParams(), CalleeTTI,
                                *GetAssumptionCache, GetBFI, PSI, &ORE);

  if (IC.isAlways()) {
    ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, "AlwaysInline", Call)
             << NV("Callee", Cloner.OrigFunc)
             << " should always be fully inlined, not partially");
    return false;
  }

  if (IC.isNever()) {
    ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline", Call)
             << NV("Callee", Cloner.OrigFunc) << " not partially inlined into "
             << NV("Caller", Caller)
             << " because it should never be inlined (cost=never)");
    return false;
  }

  if (!IC) {
    ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, "TooCostly", Call)
             << NV("Callee", Cloner.OrigFunc) << " not partially inlined into "
             << NV("Caller", Caller) << " because too costly to inline (cost="
             << NV("Cost", IC.getCost()) << ", threshold="
             << NV("Threshold", IC.getCostDelta() + IC.getCost()) << ")");
    return false;
  }
  const DataLayout &DL = Caller->getParent()->getDataLayout();

  // The savings of eliminating the call:
  int NonWeightedSavings = getCallsiteCost(CS, DL);
  BlockFrequency NormWeightedSavings(NonWeightedSavings);

  // Weighted saving is smaller than weighted cost, return false
  if (NormWeightedSavings < WeightedOutliningRcost) {
    ORE.emit(
        OptimizationRemarkAnalysis(DEBUG_TYPE, "OutliningCallcostTooHigh", Call)
        << NV("Callee", Cloner.OrigFunc) << " not partially inlined into "
        << NV("Caller", Caller) << " runtime overhead (overhead="
        << NV("Overhead", (unsigned)WeightedOutliningRcost.getFrequency())
        << ", savings="
        << NV("Savings", (unsigned)NormWeightedSavings.getFrequency()) << ")"
        << " of making the outlined call is too high");

    return false;
  }

  ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, "CanBePartiallyInlined", Call)
           << NV("Callee", Cloner.OrigFunc) << " can be partially inlined into "
           << NV("Caller", Caller) << " with cost=" << NV("Cost", IC.getCost())
           << " (threshold="
           << NV("Threshold", IC.getCostDelta() + IC.getCost()) << ")");
  return true;
}

// TODO: Ideally  we should share Inliner's InlineCost Analysis code.
// For now use a simplified version. The returned 'InlineCost' will be used
// to esimate the size cost as well as runtime cost of the BB.
int PartialInlinerImpl::computeBBInlineCost(BasicBlock *BB) {
  int InlineCost = 0;
  const DataLayout &DL = BB->getParent()->getParent()->getDataLayout();
  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
    if (isa<DbgInfoIntrinsic>(I))
      continue;

    switch (I->getOpcode()) {
    case Instruction::BitCast:
    case Instruction::PtrToInt:
    case Instruction::IntToPtr:
    case Instruction::Alloca:
      continue;
    case Instruction::GetElementPtr:
      if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
        continue;
    default:
      break;
    }

    IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(I);
    if (IntrInst) {
      if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_start ||
          IntrInst->getIntrinsicID() == Intrinsic::lifetime_end)
        continue;
    }

    if (CallInst *CI = dyn_cast<CallInst>(I)) {
      InlineCost += getCallsiteCost(CallSite(CI), DL);
      continue;
    }

    if (InvokeInst *II = dyn_cast<InvokeInst>(I)) {
      InlineCost += getCallsiteCost(CallSite(II), DL);
      continue;
    }

    if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
      InlineCost += (SI->getNumCases() + 1) * InlineConstants::InstrCost;
      continue;
    }
    InlineCost += InlineConstants::InstrCost;
  }
  return InlineCost;
}

std::tuple<int, int>
PartialInlinerImpl::computeOutliningCosts(FunctionCloner &Cloner) {
  // Now compute the cost of the call sequence to the outlined function
  // 'OutlinedFunction' in BB 'OutliningCallBB':
  int OutliningFuncCallCost = computeBBInlineCost(Cloner.OutliningCallBB);

  // Now compute the cost of the extracted/outlined function itself:
  int OutlinedFunctionCost = 0;
  for (BasicBlock &BB : *Cloner.OutlinedFunc) {
    OutlinedFunctionCost += computeBBInlineCost(&BB);
  }

  assert(OutlinedFunctionCost >= Cloner.OutlinedRegionCost &&
         "Outlined function cost should be no less than the outlined region");
  // The code extractor introduces a new root and exit stub blocks with
  // additional unconditional branches. Those branches will be eliminated
  // later with bb layout. The cost should be adjusted accordingly:
  OutlinedFunctionCost -= 2 * InlineConstants::InstrCost;

  int OutliningRuntimeOverhead =
      OutliningFuncCallCost +
      (OutlinedFunctionCost - Cloner.OutlinedRegionCost) +
      ExtraOutliningPenalty;

  return std::make_tuple(OutliningFuncCallCost, OutliningRuntimeOverhead);
}

// Create the callsite to profile count map which is
// used to update the original function's entry count,
// after the function is partially inlined into the callsite.
void PartialInlinerImpl::computeCallsiteToProfCountMap(
    Function *DuplicateFunction,
    DenseMap<User *, uint64_t> &CallSiteToProfCountMap) {
  std::vector<User *> Users(DuplicateFunction->user_begin(),
                            DuplicateFunction->user_end());
  Function *CurrentCaller = nullptr;
  std::unique_ptr<BlockFrequencyInfo> TempBFI;
  BlockFrequencyInfo *CurrentCallerBFI = nullptr;

  auto ComputeCurrBFI = [&,this](Function *Caller) {
      // For the old pass manager:
      if (!GetBFI) {
        DominatorTree DT(*Caller);
        LoopInfo LI(DT);
        BranchProbabilityInfo BPI(*Caller, LI);
        TempBFI.reset(new BlockFrequencyInfo(*Caller, BPI, LI));
        CurrentCallerBFI = TempBFI.get();
      } else {
        // New pass manager:
        CurrentCallerBFI = &(*GetBFI)(*Caller);
      }
  };

  for (User *User : Users) {
    CallSite CS = getCallSite(User);
    Function *Caller = CS.getCaller();
    if (CurrentCaller != Caller) {
      CurrentCaller = Caller;
      ComputeCurrBFI(Caller);
    } else {
      assert(CurrentCallerBFI && "CallerBFI is not set");
    }
    BasicBlock *CallBB = CS.getInstruction()->getParent();
    auto Count = CurrentCallerBFI->getBlockProfileCount(CallBB);
    if (Count)
      CallSiteToProfCountMap[User] = *Count;
    else
      CallSiteToProfCountMap[User] = 0;
  }
}

PartialInlinerImpl::FunctionCloner::FunctionCloner(Function *F,
                                                   FunctionOutliningInfo *OI)
    : OrigFunc(F) {
  ClonedOI = llvm::make_unique<FunctionOutliningInfo>();

  // Clone the function, so that we can hack away on it.
  ValueToValueMapTy VMap;
  ClonedFunc = CloneFunction(F, VMap);

  ClonedOI->ReturnBlock = cast<BasicBlock>(VMap[OI->ReturnBlock]);
  ClonedOI->NonReturnBlock = cast<BasicBlock>(VMap[OI->NonReturnBlock]);
  for (BasicBlock *BB : OI->Entries) {
    ClonedOI->Entries.push_back(cast<BasicBlock>(VMap[BB]));
  }
  for (BasicBlock *E : OI->ReturnBlockPreds) {
    BasicBlock *NewE = cast<BasicBlock>(VMap[E]);
    ClonedOI->ReturnBlockPreds.push_back(NewE);
  }
  // Go ahead and update all uses to the duplicate, so that we can just
  // use the inliner functionality when we're done hacking.
  F->replaceAllUsesWith(ClonedFunc);
}

void PartialInlinerImpl::FunctionCloner::NormalizeReturnBlock() {
  auto getFirstPHI = [](BasicBlock *BB) {
    BasicBlock::iterator I = BB->begin();
    PHINode *FirstPhi = nullptr;
    while (I != BB->end()) {
      PHINode *Phi = dyn_cast<PHINode>(I);
      if (!Phi)
        break;
      if (!FirstPhi) {
        FirstPhi = Phi;
        break;
      }
    }
    return FirstPhi;
  };

  // Special hackery is needed with PHI nodes that have inputs from more than
  // one extracted block.  For simplicity, just split the PHIs into a two-level
  // sequence of PHIs, some of which will go in the extracted region, and some
  // of which will go outside.
  BasicBlock *PreReturn = ClonedOI->ReturnBlock;
  // only split block when necessary:
  PHINode *FirstPhi = getFirstPHI(PreReturn);
  unsigned NumPredsFromEntries = ClonedOI->ReturnBlockPreds.size();

  if (!FirstPhi || FirstPhi->getNumIncomingValues() <= NumPredsFromEntries + 1)
    return;

  auto IsTrivialPhi = [](PHINode *PN) -> Value * {
    Value *CommonValue = PN->getIncomingValue(0);
    if (all_of(PN->incoming_values(),
               [&](Value *V) { return V == CommonValue; }))
      return CommonValue;
    return nullptr;
  };

  ClonedOI->ReturnBlock = ClonedOI->ReturnBlock->splitBasicBlock(
      ClonedOI->ReturnBlock->getFirstNonPHI()->getIterator());
  BasicBlock::iterator I = PreReturn->begin();
  Instruction *Ins = &ClonedOI->ReturnBlock->front();
  SmallVector<Instruction *, 4> DeadPhis;
  while (I != PreReturn->end()) {
    PHINode *OldPhi = dyn_cast<PHINode>(I);
    if (!OldPhi)
      break;

    PHINode *RetPhi =
        PHINode::Create(OldPhi->getType(), NumPredsFromEntries + 1, "", Ins);
    OldPhi->replaceAllUsesWith(RetPhi);
    Ins = ClonedOI->ReturnBlock->getFirstNonPHI();

    RetPhi->addIncoming(&*I, PreReturn);
    for (BasicBlock *E : ClonedOI->ReturnBlockPreds) {
      RetPhi->addIncoming(OldPhi->getIncomingValueForBlock(E), E);
      OldPhi->removeIncomingValue(E);
    }

    // After incoming values splitting, the old phi may become trivial.
    // Keeping the trivial phi can introduce definition inside the outline
    // region which is live-out, causing necessary overhead (load, store
    // arg passing etc).
    if (auto *OldPhiVal = IsTrivialPhi(OldPhi)) {
      OldPhi->replaceAllUsesWith(OldPhiVal);
      DeadPhis.push_back(OldPhi);
    }
    ++I;
    }
    for (auto *DP : DeadPhis)
      DP->eraseFromParent();

    for (auto E : ClonedOI->ReturnBlockPreds) {
      E->getTerminator()->replaceUsesOfWith(PreReturn, ClonedOI->ReturnBlock);
    }
}

Function *PartialInlinerImpl::FunctionCloner::doFunctionOutlining() {
  // Returns true if the block is to be partial inlined into the caller
  // (i.e. not to be extracted to the out of line function)
  auto ToBeInlined = [&, this](BasicBlock *BB) {
    return BB == ClonedOI->ReturnBlock ||
           (std::find(ClonedOI->Entries.begin(), ClonedOI->Entries.end(), BB) !=
            ClonedOI->Entries.end());
  };

  // Gather up the blocks that we're going to extract.
  std::vector<BasicBlock *> ToExtract;
  ToExtract.push_back(ClonedOI->NonReturnBlock);
  OutlinedRegionCost +=
      PartialInlinerImpl::computeBBInlineCost(ClonedOI->NonReturnBlock);
  for (BasicBlock &BB : *ClonedFunc)
    if (!ToBeInlined(&BB) && &BB != ClonedOI->NonReturnBlock) {
      ToExtract.push_back(&BB);
      // FIXME: the code extractor may hoist/sink more code
      // into the outlined function which may make the outlining
      // overhead (the difference of the outlined function cost
      // and OutliningRegionCost) look larger.
      OutlinedRegionCost += computeBBInlineCost(&BB);
    }

  // The CodeExtractor needs a dominator tree.
  DominatorTree DT;
  DT.recalculate(*ClonedFunc);

  // Manually calculate a BlockFrequencyInfo and BranchProbabilityInfo.
  LoopInfo LI(DT);
  BranchProbabilityInfo BPI(*ClonedFunc, LI);
  ClonedFuncBFI.reset(new BlockFrequencyInfo(*ClonedFunc, BPI, LI));

  // Extract the body of the if.
  OutlinedFunc = CodeExtractor(ToExtract, &DT, /*AggregateArgs*/ false,
                               ClonedFuncBFI.get(), &BPI)
                     .extractCodeRegion();

  if (OutlinedFunc) {
    OutliningCallBB = PartialInlinerImpl::getOneCallSiteTo(OutlinedFunc)
        .getInstruction()
        ->getParent();
    assert(OutliningCallBB->getParent() == ClonedFunc);
  }

  return OutlinedFunc;
}

PartialInlinerImpl::FunctionCloner::~FunctionCloner() {
  // Ditch the duplicate, since we're done with it, and rewrite all remaining
  // users (function pointers, etc.) back to the original function.
  ClonedFunc->replaceAllUsesWith(OrigFunc);
  ClonedFunc->eraseFromParent();
  if (!IsFunctionInlined) {
    // Remove the function that is speculatively created if there is no
    // reference.
    if (OutlinedFunc)
      OutlinedFunc->eraseFromParent();
  }
}

Function *PartialInlinerImpl::unswitchFunction(Function *F) {
  if (F->hasAddressTaken())
    return nullptr;

  // Let inliner handle it
  if (F->hasFnAttribute(Attribute::AlwaysInline))
    return nullptr;

  if (F->hasFnAttribute(Attribute::NoInline))
    return nullptr;

  if (PSI->isFunctionEntryCold(F))
    return nullptr;

  if (F->user_begin() == F->user_end())
    return nullptr;

  std::unique_ptr<FunctionOutliningInfo> OI = computeOutliningInfo(F);

  if (!OI)
    return nullptr;

  FunctionCloner Cloner(F, OI.get());
  Cloner.NormalizeReturnBlock();
  Function *OutlinedFunction = Cloner.doFunctionOutlining();

  bool AnyInline = tryPartialInline(Cloner);

  if (AnyInline)
    return OutlinedFunction;

  return nullptr;
}

bool PartialInlinerImpl::tryPartialInline(FunctionCloner &Cloner) {
  int NonWeightedRcost;
  int SizeCost;

  if (Cloner.OutlinedFunc == nullptr)
    return false;

  std::tie(SizeCost, NonWeightedRcost) = computeOutliningCosts(Cloner);

  auto RelativeToEntryFreq = getOutliningCallBBRelativeFreq(Cloner);
  auto WeightedRcost = BlockFrequency(NonWeightedRcost) * RelativeToEntryFreq;

  // The call sequence to the outlined function is larger than the original
  // outlined region size, it does not increase the chances of inlining
  // the function with outlining (The inliner uses the size increase to
  // model the cost of inlining a callee).
  if (!SkipCostAnalysis && Cloner.OutlinedRegionCost < SizeCost) {
    OptimizationRemarkEmitter ORE(Cloner.OrigFunc);
    DebugLoc DLoc;
    BasicBlock *Block;
    std::tie(DLoc, Block) = getOneDebugLoc(Cloner.ClonedFunc);
    ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, "OutlineRegionTooSmall",
                                        DLoc, Block)
             << ore::NV("Function", Cloner.OrigFunc)
             << " not partially inlined into callers (Original Size = "
             << ore::NV("OutlinedRegionOriginalSize", Cloner.OutlinedRegionCost)
             << ", Size of call sequence to outlined function = "
             << ore::NV("NewSize", SizeCost) << ")");
    return false;
  }

  assert(Cloner.OrigFunc->user_begin() == Cloner.OrigFunc->user_end() &&
         "F's users should all be replaced!");

  std::vector<User *> Users(Cloner.ClonedFunc->user_begin(),
                            Cloner.ClonedFunc->user_end());

  DenseMap<User *, uint64_t> CallSiteToProfCountMap;
  if (Cloner.OrigFunc->getEntryCount())
    computeCallsiteToProfCountMap(Cloner.ClonedFunc, CallSiteToProfCountMap);

  auto CalleeEntryCount = Cloner.OrigFunc->getEntryCount();
  uint64_t CalleeEntryCountV = (CalleeEntryCount ? *CalleeEntryCount : 0);

  bool AnyInline = false;
  for (User *User : Users) {
    CallSite CS = getCallSite(User);

    if (IsLimitReached())
      continue;

    OptimizationRemarkEmitter ORE(CS.getCaller());

    if (!shouldPartialInline(CS, Cloner, WeightedRcost, ORE))
      continue;

    ORE.emit(
        OptimizationRemark(DEBUG_TYPE, "PartiallyInlined", CS.getInstruction())
        << ore::NV("Callee", Cloner.OrigFunc) << " partially inlined into "
        << ore::NV("Caller", CS.getCaller()));

    InlineFunctionInfo IFI(nullptr, GetAssumptionCache, PSI);
    InlineFunction(CS, IFI);

    // Now update the entry count:
    if (CalleeEntryCountV && CallSiteToProfCountMap.count(User)) {
      uint64_t CallSiteCount = CallSiteToProfCountMap[User];
      CalleeEntryCountV -= std::min(CalleeEntryCountV, CallSiteCount);
    }

    AnyInline = true;
    NumPartialInlining++;
    // Update the stats
    NumPartialInlined++;
  }

  if (AnyInline) {
    Cloner.IsFunctionInlined = true;
    if (CalleeEntryCount)
      Cloner.OrigFunc->setEntryCount(CalleeEntryCountV);
  }

  return AnyInline;
}

bool PartialInlinerImpl::run(Module &M) {
  if (DisablePartialInlining)
    return false;

  std::vector<Function *> Worklist;
  Worklist.reserve(M.size());
  for (Function &F : M)
    if (!F.use_empty() && !F.isDeclaration())
      Worklist.push_back(&F);

  bool Changed = false;
  while (!Worklist.empty()) {
    Function *CurrFunc = Worklist.back();
    Worklist.pop_back();

    if (CurrFunc->use_empty())
      continue;

    bool Recursive = false;
    for (User *U : CurrFunc->users())
      if (Instruction *I = dyn_cast<Instruction>(U))
        if (I->getParent()->getParent() == CurrFunc) {
          Recursive = true;
          break;
        }
    if (Recursive)
      continue;

    if (Function *NewFunc = unswitchFunction(CurrFunc)) {
      Worklist.push_back(NewFunc);
      Changed = true;
    }
  }

  return Changed;
}

char PartialInlinerLegacyPass::ID = 0;

INITIALIZE_PASS_BEGIN(PartialInlinerLegacyPass, "partial-inliner",
                      "Partial Inliner", false, false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
INITIALIZE_PASS_END(PartialInlinerLegacyPass, "partial-inliner",
                    "Partial Inliner", false, false)

ModulePass *llvm::createPartialInliningPass() {
  return new PartialInlinerLegacyPass();
}

PreservedAnalyses PartialInlinerPass::run(Module &M,
                                          ModuleAnalysisManager &AM) {
  auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();

  std::function<AssumptionCache &(Function &)> GetAssumptionCache =
      [&FAM](Function &F) -> AssumptionCache & {
    return FAM.getResult<AssumptionAnalysis>(F);
  };

  std::function<BlockFrequencyInfo &(Function &)> GetBFI =
      [&FAM](Function &F) -> BlockFrequencyInfo & {
    return FAM.getResult<BlockFrequencyAnalysis>(F);
  };

  std::function<TargetTransformInfo &(Function &)> GetTTI =
      [&FAM](Function &F) -> TargetTransformInfo & {
    return FAM.getResult<TargetIRAnalysis>(F);
  };

  ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(M);

  if (PartialInlinerImpl(&GetAssumptionCache, &GetTTI, {GetBFI}, PSI).run(M))
    return PreservedAnalyses::none();
  return PreservedAnalyses::all();
}