//===--- ClosureSpecializer.cpp - Performs Closure Specialization ---------===//
//
// This source file is part of the Swift.org open source project
//
// Copyright (c) 2014 - 2017 Apple Inc. and the Swift project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
// See https://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
///
/// \file
///
/// Closure Specialization
/// ----------------------
///
/// The purpose of the algorithm in this file is to perform the following
/// transformation: given a closure passed into a function which the closure is
/// then invoked in, clone the function and create a copy of the closure inside
/// the function. This closure will be able to be eliminated easily and the
/// overhead is gone. We then try to remove the original closure.
///
/// There are some complications. They are listed below and how we work around
/// them:
///
/// 1. If we support the specialization of closures with multiple user callsites
///    that can be specialized, we need to ensure that any captured values have
///    their reference counts adjusted properly. This implies for every
///    specialized call site, we insert an additional retain for each captured
///    argument with reference semantics. We will pass them in as extra @owned
///    to the specialized function. This @owned will be consumed by the "copy"
///    partial apply that is in the specialized function. Now the partial apply
///    will own those ref counts. This is unapplicable to thin_to_thick_function
///    since they do not have any captured args.
///
/// 2. If the closure was passed in @owned vs if the closure was passed in
///    @guaranteed. If the original closure was passed in @owned, then we know
///    that there is a balancing release for the new "copy" partial apply. But
///    since the original partial apply no longer will have that corresponding
///    -1, we need to insert a release for the old partial apply. We do this
///    right after the old call site where the original partial apply was
///    called. This ensures we do not shrink the lifetime of the old partial
///    apply. In the case where the old partial_apply was passed in at +0, we
///    know that the old partial_apply does not need to have any ref count
///    adjustments. On the other hand, the new "copy" partial apply in the
///    specialized function now needs to be balanced lest we leak. Thus we
///    insert a release right before any exit from the function. This ensures
///    that the release occurs in the epilog after any retains associated with
///    @owned return values.
///
/// 3. In !useLoweredAddresses mode, we do not support specialization of closures
///    with arguments passed using any indirect calling conventions besides
///    @inout and @inout_aliasable.  This is a temporary limitation that goes
///    away with sil-opaque-values.
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "closure-specialization"
#include "swift/Basic/Range.h"
#include "swift/Demangling/Demangler.h"
#include "swift/SIL/InstructionUtils.h"
#include "swift/SIL/SILCloner.h"
#include "swift/SIL/SILFunction.h"
#include "swift/SIL/SILInstruction.h"
#include "swift/SIL/SILModule.h"
#include "swift/SILOptimizer/Analysis/BasicCalleeAnalysis.h"
#include "swift/SILOptimizer/Analysis/FunctionOrder.h"
#include "swift/SILOptimizer/Analysis/ValueTracking.h"
#include "swift/SILOptimizer/PassManager/Passes.h"
#include "swift/SILOptimizer/PassManager/Transforms.h"
#include "swift/SILOptimizer/Utils/CFGOptUtils.h"
#include "swift/SILOptimizer/Utils/InstOptUtils.h"
#include "swift/SILOptimizer/Utils/SILInliner.h"
#include "swift/SILOptimizer/Utils/SILOptFunctionBuilder.h"
#include "swift/SILOptimizer/Utils/SpecializationMangler.h"
#include "swift/SILOptimizer/Utils/StackNesting.h"
#include "swift/SILOptimizer/Utils/ValueLifetime.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"

using namespace swift;

STATISTIC(NumClosureSpecialized,
          "Number of functions with closures specialized");
STATISTIC(NumPropagatedClosuresEliminated,
          "Number of closures propagated and then eliminated");
STATISTIC(NumPropagatedClosuresNotEliminated,
          "Number of closures propagated but not eliminated");

llvm::cl::opt<bool> EliminateDeadClosures(
    "closure-specialize-eliminate-dead-closures", llvm::cl::init(true),
    llvm::cl::desc("Do not eliminate dead closures after closure "
                   "specialization. This is meant ot be used when testing."));

//===----------------------------------------------------------------------===//
//                                  Utility
//===----------------------------------------------------------------------===//

static bool isSupportedClosureKind(const SILInstruction *I) {
  return isa<ThinToThickFunctionInst>(I) || isa<PartialApplyInst>(I);
}

//===----------------------------------------------------------------------===//
//                       Closure Spec Cloner Interface
//===----------------------------------------------------------------------===//

namespace {

class CallSiteDescriptor;

/// A SILCloner subclass which clones a function that takes a closure
/// argument. We update the parameter list to remove the parameter for the
/// closure argument and to append the variables captured in the closure.
/// We also need to replace the closure parameter with the partial apply
/// on the closure. We need to update the callsite to pass in the correct
/// arguments.
class ClosureSpecCloner : public SILClonerWithScopes<ClosureSpecCloner> {
public:
  using SuperTy = SILClonerWithScopes<ClosureSpecCloner>;
  friend class SILInstructionVisitor<ClosureSpecCloner>;
  friend class SILCloner<ClosureSpecCloner>;

  ClosureSpecCloner(SILOptFunctionBuilder &FunctionBuilder,
                    const CallSiteDescriptor &CallSiteDesc,
                    StringRef ClonedName)
      : SuperTy(*initCloned(FunctionBuilder, CallSiteDesc, ClonedName)),
        CallSiteDesc(CallSiteDesc) {}

  void populateCloned();

  SILValue
  cloneCalleeConversion(SILValue calleeValue, SILValue NewClosure,
                        SILBuilder &Builder,
                        SmallVectorImpl<PartialApplyInst *> &NeedsRelease,
                        llvm::DenseMap<SILValue, SILValue> &CapturedMap);

  SILFunction *getCloned() { return &getBuilder().getFunction(); }
  static SILFunction *cloneFunction(SILOptFunctionBuilder &FunctionBuilder,
                                    const CallSiteDescriptor &CallSiteDesc,
                                    StringRef NewName) {
    ClosureSpecCloner C(FunctionBuilder, CallSiteDesc, NewName);
    C.populateCloned();
    ++NumClosureSpecialized;
    return C.getCloned();
  };

private:
  static SILFunction *initCloned(SILOptFunctionBuilder &FunctionBuilder,
                                 const CallSiteDescriptor &CallSiteDesc,
                                 StringRef ClonedName);
  const CallSiteDescriptor &CallSiteDesc;
};

} // end anonymous namespace

//===----------------------------------------------------------------------===//
//                            Call Site Descriptor
//===----------------------------------------------------------------------===//

namespace {
struct ClosureInfo;

static SILFunction *getClosureCallee(SILInstruction *inst) {
  if (auto *PAI = dyn_cast<PartialApplyInst>(inst))
    return cast<FunctionRefInst>(PAI->getCallee())->getReferencedFunction();

  auto *TTTFI = cast<ThinToThickFunctionInst>(inst);
  return cast<FunctionRefInst>(TTTFI->getCallee())->getReferencedFunction();
}

class CallSiteDescriptor {
  ClosureInfo *CInfo;
  FullApplySite AI;
  unsigned ClosureIndex;
  SILParameterInfo ClosureParamInfo;

  // This is only needed if we have guaranteed parameters. In most cases it will
  // have only one element, a return inst.
  llvm::TinyPtrVector<SILBasicBlock *> NonFailureExitBBs;

public:
  CallSiteDescriptor(ClosureInfo *CInfo, FullApplySite AI,
                     unsigned ClosureIndex, SILParameterInfo ClosureParamInfo,
                     llvm::TinyPtrVector<SILBasicBlock *> &&NonFailureExitBBs)
    : CInfo(CInfo), AI(AI), ClosureIndex(ClosureIndex),
      ClosureParamInfo(ClosureParamInfo),
      NonFailureExitBBs(NonFailureExitBBs) {}

  CallSiteDescriptor(CallSiteDescriptor&&) =default;
  CallSiteDescriptor &operator=(CallSiteDescriptor &&) =default;

  SILFunction *getApplyCallee() const {
    return cast<FunctionRefInst>(AI.getCallee())->getReferencedFunction();
  }

  SILFunction *getClosureCallee() const {
    return ::getClosureCallee(getClosure());
  }

  bool closureHasRefSemanticContext() const {
    return isa<PartialApplyInst>(getClosure()) &&
           !cast<PartialApplyInst>(getClosure())->isOnStack();
  }

  bool destroyIfPartialApplyStack(SILBuilder &B,
                                  SingleValueInstruction *newClosure) const {
    auto *PA = dyn_cast<PartialApplyInst>(newClosure);
    if (!PA || !PA->isOnStack())
      return false;

    if (B.getFunction().hasOwnership()) {
      // Under OSSA, the closure acts as an owned value whose lifetime is a
      // borrow scope for the captures, so we need to end the borrow scope
      // before ending the lifetimes of the captures themselves.
      B.createDestroyValue(getClosure()->getLoc(), PA);
      insertDestroyOfCapturedArguments(PA, B);
      // The stack slot for the partial_apply doesn't get reified until after
      // OSSA.
      return false;
    } else {
      insertDestroyOfCapturedArguments(PA, B);
      B.createDeallocStack(getClosure()->getLoc(), PA);
      return true;
    }
  }

  unsigned getClosureIndex() const { return ClosureIndex; }

  // Get the closure value passed to the apply (on the caller side).
  SILValue getClosureCallerArg() const {
    return getApplyInst().getArgument(ClosureIndex);
  }

  SILParameterInfo getClosureParameterInfo() const { return ClosureParamInfo; }

  SingleValueInstruction *
  createNewClosure(SILBuilder &B, SILValue V,
                   llvm::SmallVectorImpl<SILValue> &Args) const {
    if (auto *PA = dyn_cast<PartialApplyInst>(getClosure()))
      return B.createPartialApply(getClosure()->getLoc(), V, {}, Args,
                                  PA->getCalleeConvention(),
                                  PA->getResultIsolation(),
                                  PA->isOnStack());

    assert(isa<ThinToThickFunctionInst>(getClosure()) &&
           "We only support partial_apply and thin_to_thick_function");
    return B.createThinToThickFunction(getClosure()->getLoc(), V,
                                       getClosure()->getType());
  }

  FullApplySite getApplyInst() const { return AI; }

  bool isSerialized() const;
  SerializedKind_t getSerializedKind() const;

  std::string createName() const;

  OperandValueArrayRef getArguments() const {
    if (auto *PAI = dyn_cast<PartialApplyInst>(getClosure()))
      return PAI->getArguments();

    // Thin to thick function has no non-callee arguments.
    assert(isa<ThinToThickFunctionInst>(getClosure()) &&
           "We only support partial_apply and thin_to_thick_function");
    return OperandValueArrayRef(ArrayRef<Operand>());
  }

  inline SingleValueInstruction *getClosure() const;

  unsigned getNumArguments() const {
    if (auto *PAI = dyn_cast<PartialApplyInst>(getClosure()))
      return PAI->getNumArguments();

    // Thin to thick function has no non-callee arguments.
    assert(isa<ThinToThickFunctionInst>(getClosure()) &&
           "We only support partial_apply and thin_to_thick_function");
    return 0;
  }

  bool isClosureGuaranteed() const {
    return getClosureParameterInfo().isGuaranteed();
  }

  bool isClosureConsumed() const {
    return getClosureParameterInfo().isConsumed();
  }

  bool isClosureOnStack() const {
    auto *PA = dyn_cast<PartialApplyInst>(getClosure());
    if (!PA)
      return false;
    return PA->isOnStack();
  }

  bool isTrivialNoEscapeParameter() const {
    auto ClosureParmFnTy =
        getClosureParameterInfo().getInterfaceType()->getAs<SILFunctionType>();
    return ClosureParmFnTy->isTrivialNoEscape();
  }

  SILLocation getLoc() const { return getClosure()->getLoc(); }

  SILModule &getModule() const { return AI.getModule(); }

  ArrayRef<SILBasicBlock *> getNonFailureExitBBs() const {
    return NonFailureExitBBs;
  }

  /// Extend the lifetime of 'Arg' to the lifetime of the closure.
  void extendArgumentLifetime(SILValue Arg,
                              SILArgumentConvention ArgConvention) const;
};
} // end anonymous namespace

namespace {
struct ClosureInfo {
  SingleValueInstruction *Closure;
  ValueLifetimeAnalysis::Frontier LifetimeFrontier;
  llvm::SmallVector<CallSiteDescriptor, 8> CallSites;

  ClosureInfo(SingleValueInstruction *Closure) : Closure(Closure) {}

  ClosureInfo(ClosureInfo &&) =default;
  ClosureInfo &operator=(ClosureInfo &&) =default;
};
} // end anonymous namespace

SingleValueInstruction *CallSiteDescriptor::getClosure() const {
  return CInfo->Closure;
}

static bool isNonInoutIndirectSILArgument(SILValue Arg,
                                          SILArgumentConvention ArgConvention) {
  return !Arg->getType().isObject() && ArgConvention.isIndirectConvention() &&
         ArgConvention != SILArgumentConvention::Indirect_Inout &&
         ArgConvention != SILArgumentConvention::Indirect_InoutAliasable;
}

/// Update the callsite to pass in the correct arguments.
static void rewriteApplyInst(const CallSiteDescriptor &CSDesc,
                             SILFunction *NewF) {
  FullApplySite AI = CSDesc.getApplyInst();
  SingleValueInstruction *Closure = CSDesc.getClosure();
  SILBuilderWithScope Builder(Closure);
  FunctionRefInst *FRI = Builder.createFunctionRef(AI.getLoc(), NewF);

  // Create the args for the new apply by removing the closure argument...
  llvm::SmallVector<SILValue, 8> NewArgs;
  unsigned Index = 0;
  for (auto Arg : AI.getArguments()) {
    if (Index != CSDesc.getClosureIndex())
      NewArgs.push_back(Arg);
    ++Index;
  }

  // ... and appending the captured arguments. We also insert retains here at
  // the location of the original closure. This is needed to balance the
  // implicit release of all captured arguments that occurs when the partial
  // apply is destroyed.
  auto ClosureCalleeConv = CSDesc.getClosureCallee()->getConventions();
  unsigned ClosureArgIdx =
      ClosureCalleeConv.getNumSILArguments() - CSDesc.getNumArguments();
  for (auto Arg : CSDesc.getArguments()) {
    SILType ArgTy = Arg->getType();

    // If our argument is of trivial type, continue...
    if (ArgTy.isTrivial(*NewF)) {
      NewArgs.push_back(Arg);
      ++ClosureArgIdx;
      continue;
    }

    auto ArgConvention =
        ClosureCalleeConv.getSILArgumentConvention(ClosureArgIdx);

    // Non-inout indirect arguments are not supported yet.
    assert(ArgTy.isObject() ||
           !isNonInoutIndirectSILArgument(Arg, ArgConvention));

    // If argument is not an object and it is an inout parameter,
    // continue...
    if (!ArgTy.isObject() &&
        !isNonInoutIndirectSILArgument(Arg, ArgConvention)) {
      NewArgs.push_back(Arg);
      ++ClosureArgIdx;
      continue;
    }

    // TODO: When we support address types, this code path will need to be
    // updated.

    // We need to balance the consumed argument of the new partial_apply in the
    // specialized callee by a retain. If both the original partial_apply and
    // the apply of the callee are in the same basic block we can assume they
    // are executed the same number of times. Therefore it is sufficient to just
    // retain the argument at the site of the original partial_apply.
    //
    //    %closure = partial_apply (%arg)
    //             = apply %callee(%closure)
    //  =>
    //             retain %arg
    //    %closure = partial_apply (%arg)
    //               apply %specialized_callee(..., %arg)
    //
    // However, if they are not in the same basic block the callee might be
    // executed more frequently than the closure (for example, if the closure is
    // created in a loop preheader and the callee taking the closure is executed
    // in the loop). In such a case we must keep the argument live across the
    // call site of the callee and emit a matching retain for every invocation
    // of the callee.
    //
    //    %closure = partial_apply (%arg)
    //
    //    while () {
    //             = %callee(%closure)
    //    }
    // =>
    //               retain %arg
    //    %closure = partial_apply (%arg)
    //
    //    while () {
    //               retain %arg
    //               apply %specialized_callee(.., %arg)
    //    }
    //               release %arg
    //
    if (AI.getParent() != Closure->getParent()) {
      // Emit the retain and release that keeps the argument life across the
      // callee using the closure.
      CSDesc.extendArgumentLifetime(Arg, ArgConvention);

      // Emit the retain that matches the captured argument by the
      // partial_apply
      // in the callee that is consumed by the partial_apply.
      Builder.setInsertionPoint(AI.getInstruction());
      Builder.createRetainValue(Closure->getLoc(), Arg,
                                Builder.getDefaultAtomicity());
    } else {
      Builder.createRetainValue(Closure->getLoc(), Arg,
                                Builder.getDefaultAtomicity());
    }

    NewArgs.push_back(Arg);
    ++ClosureArgIdx;
  }

  Builder.setInsertionPoint(AI.getInstruction());
  FullApplySite NewAI;
  switch (AI.getKind()) {
  case FullApplySiteKind::TryApplyInst: {
    auto *TAI = cast<TryApplyInst>(AI);
    NewAI = Builder.createTryApply(AI.getLoc(), FRI,
                                   SubstitutionMap(), NewArgs,
                                   TAI->getNormalBB(), TAI->getErrorBB(),
                                   TAI->getApplyOptions());
    // If we passed in the original closure as @owned, then insert a release
    // right after NewAI. This is to balance the +1 from being an @owned
    // argument to AI.
    if (!CSDesc.isClosureConsumed() || CSDesc.isTrivialNoEscapeParameter() ||
        !CSDesc.closureHasRefSemanticContext()) {
      break;
    }

    Builder.setInsertionPoint(TAI->getNormalBB()->begin());
    Builder.createReleaseValue(Closure->getLoc(), Closure,
                               Builder.getDefaultAtomicity());
    Builder.setInsertionPoint(TAI->getErrorBB()->begin());
    Builder.createReleaseValue(Closure->getLoc(), Closure,
                               Builder.getDefaultAtomicity());
    Builder.setInsertionPoint(AI.getInstruction());
    break;
  }
  case FullApplySiteKind::ApplyInst: {
    auto oldApply = cast<ApplyInst>(AI);
    auto newApply = Builder.createApply(oldApply->getLoc(), FRI,
                                        SubstitutionMap(), NewArgs,
                                        oldApply->getApplyOptions());
    // If we passed in the original closure as @owned, then insert a release
    // right after NewAI. This is to balance the +1 from being an @owned
    // argument to AI.
    if (CSDesc.isClosureConsumed() && !CSDesc.isTrivialNoEscapeParameter() &&
        CSDesc.closureHasRefSemanticContext())
      Builder.createReleaseValue(Closure->getLoc(), Closure,
                                 Builder.getDefaultAtomicity());

    // Replace all uses of the old apply with the new apply.
    oldApply->replaceAllUsesWith(newApply);
    break;
  }
  case FullApplySiteKind::BeginApplyInst:
    llvm_unreachable("Unhandled case");
  }
    
  // Erase the old apply.
  AI.getInstruction()->eraseFromParent();

  // TODO: Maybe include invalidation code for CallSiteDescriptor after we erase
  // AI from parent?
}

bool CallSiteDescriptor::isSerialized() const {
  return getClosure()->getFunction()->getSerializedKind() == IsSerialized;
}
SerializedKind_t CallSiteDescriptor::getSerializedKind() const {
  return getClosure()->getFunction()->getSerializedKind();
}

std::string CallSiteDescriptor::createName() const {
  auto P = Demangle::SpecializationPass::ClosureSpecializer;
  Mangle::FunctionSignatureSpecializationMangler Mangler(P, getSerializedKind(),
                                                         getApplyCallee());

  if (auto *PAI = dyn_cast<PartialApplyInst>(getClosure())) {
    Mangler.setArgumentClosureProp(getClosureIndex(), PAI);
  } else {
    auto *TTTFI = cast<ThinToThickFunctionInst>(getClosure());
    Mangler.setArgumentClosureProp(getClosureIndex(), TTTFI);
  }
  return Mangler.mangle();
}

void CallSiteDescriptor::extendArgumentLifetime(
    SILValue Arg, SILArgumentConvention ArgConvention) const {
  assert(!CInfo->LifetimeFrontier.empty() &&
         "Need a post-dominating release(s)");

  auto ArgTy = Arg->getType();

  // Extend the lifetime of a captured argument to cover the callee.
  SILBuilderWithScope Builder(getClosure());

  // Indirect non-inout arguments are not supported yet.
  assert(!isNonInoutIndirectSILArgument(Arg, ArgConvention));

  if (ArgTy.isObject()) {
    Builder.createRetainValue(getClosure()->getLoc(), Arg,
                              Builder.getDefaultAtomicity());
    for (auto *I : CInfo->LifetimeFrontier) {
      Builder.setInsertionPoint(I);
      Builder.createReleaseValue(getClosure()->getLoc(), Arg,
                                 Builder.getDefaultAtomicity());
    }
  }
}

static bool isSupportedClosure(const SILInstruction *Closure) {
  if (!isSupportedClosureKind(Closure))
    return false;

  // We only support simple closures where a partial_apply or
  // thin_to_thick_function is passed a function_ref. This will be stored here
  // so the checking of the Callee can use the same code in both cases.
  SILValue Callee;

  // If Closure is a partial apply...
  if (auto *PAI = dyn_cast<PartialApplyInst>(Closure)) {
    // And it has substitutions, return false.
    if (PAI->hasSubstitutions())
      return false;

    // Ok, it is a closure we support, set Callee.
    Callee = PAI->getCallee();

  } else {
    // Otherwise closure must be a thin_to_thick_function.
    Callee = cast<ThinToThickFunctionInst>(Closure)->getCallee();
  }

  // Make sure that it is a simple partial apply (i.e. its callee is a
  // function_ref).
  //
  // TODO: We can probably handle other partial applies here.
  auto *FRI = dyn_cast_or_null<FunctionRefInst>(Callee);
  if (!FRI)
    return false;

  if (auto *PAI = dyn_cast<PartialApplyInst>(Closure)) {
    // Check whether each argument is supported.
    auto ClosureCallee = FRI->getReferencedFunction();
    auto ClosureCalleeConv = ClosureCallee->getConventions();
    unsigned ClosureArgIdxBase =
        ClosureCalleeConv.getNumSILArguments() - PAI->getNumArguments();
    for (auto pair : llvm::enumerate(PAI->getArguments())) {
      auto Arg = pair.value();
      auto ClosureArgIdx = pair.index() + ClosureArgIdxBase;
      auto ArgConvention =
          ClosureCalleeConv.getSILArgumentConvention(ClosureArgIdx);

      SILType ArgTy = Arg->getType();
      // Specializing (currently) always produces a retain in the caller.
      // That's not allowed for values of move-only type.
      if (ArgTy.isMoveOnly()) {
        return false;
      }

      // Only @inout/@inout_aliasable addresses are (currently) supported.
      // If our argument is an object, continue...
      if (ArgTy.isObject()) {
        ++ClosureArgIdx;
        continue;
      }
      if (ArgConvention != SILArgumentConvention::Indirect_Inout &&
          ArgConvention != SILArgumentConvention::Indirect_InoutAliasable)
        return false;
      ++ClosureArgIdx;
    }
  }

  // Otherwise, we do support specializing this closure.
  return true;
}

//===----------------------------------------------------------------------===//
//                     Closure Spec Cloner Implementation
//===----------------------------------------------------------------------===//

/// In this function we create the actual cloned function and its proper cloned
/// type. But we do not create any body. This implies that the creation of the
/// actual arguments in the function is in populateCloned.
///
/// \arg PAUser The function that is being passed the partial apply.
/// \arg PAI The partial apply that is being passed to PAUser.
/// \arg ClosureIndex The index of the partial apply in PAUser's function
///                   signature.
/// \arg ClonedName The name of the cloned function that we will create.
SILFunction *
ClosureSpecCloner::initCloned(SILOptFunctionBuilder &FunctionBuilder,
                              const CallSiteDescriptor &CallSiteDesc,
                              StringRef ClonedName) {
  SILFunction *ClosureUser = CallSiteDesc.getApplyCallee();

  // This is the list of new interface parameters of the cloned function.
  llvm::SmallVector<SILParameterInfo, 4> NewParameterInfoList;

  // First add to NewParameterInfoList all of the SILParameterInfo in the
  // original function except for the closure.
  CanSILFunctionType ClosureUserFunTy = ClosureUser->getLoweredFunctionType();
  auto ClosureUserConv = ClosureUser->getConventions();
  unsigned Index = ClosureUserConv.getSILArgIndexOfFirstParam();
  for (auto &param : ClosureUserConv.getParameters()) {
    if (Index != CallSiteDesc.getClosureIndex())
      NewParameterInfoList.push_back(param);
    ++Index;
  }

  // Then add any arguments that are captured in the closure to the function's
  // argument type. Since they are captured, we need to pass them directly into
  // the new specialized function.
  SILFunction *ClosedOverFun = CallSiteDesc.getClosureCallee();
  auto ClosedOverFunConv = ClosedOverFun->getConventions();
  SILModule &M = ClosureUser->getModule();

  // Captured parameters are always appended to the function signature. If the
  // type of the captured argument is:
  // - direct and trivial, pass the argument as Direct_Unowned.
  // - direct and non-trivial, pass the argument as Direct_Owned.
  // - indirect, pass the argument using the same parameter convention as in the
  // original closure.
  //
  // We use the type of the closure here since we allow for the closure to be an
  // external declaration.
  unsigned NumTotalParams = ClosedOverFunConv.getNumParameters();
  unsigned NumNotCaptured = NumTotalParams - CallSiteDesc.getNumArguments();
  for (auto &PInfo : ClosedOverFunConv.getParameters().slice(NumNotCaptured)) {
    ParameterConvention ParamConv;
    if (PInfo.isFormalIndirect()) {
      ParamConv = PInfo.getConvention();
      assert(!SILModuleConventions(M).useLoweredAddresses()
             || ParamConv == ParameterConvention::Indirect_Inout
             || ParamConv == ParameterConvention::Indirect_InoutAliasable);
    } else {
      ParamConv = ClosedOverFunConv
                          .getSILType(PInfo, CallSiteDesc.getApplyInst()
                                                 .getFunction()
                                                 ->getTypeExpansionContext())
                          .isTrivial(*ClosureUser)
                      ? ParameterConvention::Direct_Unowned
                      : ParameterConvention::Direct_Owned;
    }

    SILParameterInfo NewPInfo(PInfo.getInterfaceType(), ParamConv);
    NewParameterInfoList.push_back(NewPInfo);
  }

  // The specialized function is always a thin function. This is important
  // because we may add additional parameters after the Self parameter of
  // witness methods. In this case the new function is not a method anymore.
  auto ExtInfo = ClosureUserFunTy->getExtInfo();
  ExtInfo = ExtInfo.withRepresentation(SILFunctionTypeRepresentation::Thin);

  auto ClonedTy = SILFunctionType::get(
      ClosureUserFunTy->getInvocationGenericSignature(), ExtInfo,
      ClosureUserFunTy->getCoroutineKind(),
      ClosureUserFunTy->getCalleeConvention(), NewParameterInfoList,
      ClosureUserFunTy->getYields(), ClosureUserFunTy->getResults(),
      ClosureUserFunTy->getOptionalErrorResult(),
      ClosureUserFunTy->getPatternSubstitutions(),
      ClosureUserFunTy->getInvocationSubstitutions(),
      M.getASTContext());

  // We make this function bare so we don't have to worry about decls in the
  // SILArgument.
  auto *Fn = FunctionBuilder.createFunction(
      // It's important to use a shared linkage for the specialized function
      // and not the original linkage.
      // Otherwise the new function could have an external linkage (in case the
      // original function was de-serialized) and would not be code-gen'd.
      // It's also important to disconnect this specialized function from any
      // classes (the classSubclassScope), because that may incorrectly
      // influence the linkage.
      getSpecializedLinkage(ClosureUser, ClosureUser->getLinkage()), ClonedName,
      ClonedTy, ClosureUser->getGenericEnvironment(),
      ClosureUser->getLocation(), IsBare, ClosureUser->isTransparent(),
      CallSiteDesc.getSerializedKind(), IsNotDynamic, IsNotDistributed,
      IsNotRuntimeAccessible, ClosureUser->getEntryCount(),
      ClosureUser->isThunk(),
      /*classSubclassScope=*/SubclassScope::NotApplicable,
      ClosureUser->getInlineStrategy(), ClosureUser->getEffectsKind(),
      ClosureUser, ClosureUser->getDebugScope());
  if (!ClosureUser->hasOwnership()) {
    Fn->setOwnershipEliminated();
  }
  for (auto &Attr : ClosureUser->getSemanticsAttrs())
    Fn->addSemanticsAttr(Attr);
  return Fn;
}

// Clone a chain of ConvertFunctionInsts.
SILValue ClosureSpecCloner::cloneCalleeConversion(
    SILValue calleeValue, SILValue NewClosure, SILBuilder &Builder,
    SmallVectorImpl<PartialApplyInst *> &NeedsRelease,
    llvm::DenseMap<SILValue, SILValue> &CapturedMap) {

  // There might be a mark dependence on a previous closure value. Therefore, we
  // add all closure values to the map.
  auto addToOldToNewClosureMap = [&](SILValue origValue,
                                     SILValue newValue) -> SILValue {
    assert(!CapturedMap.count(origValue));
    CapturedMap[origValue] = newValue;
    return newValue;
  };

  if (calleeValue == CallSiteDesc.getClosure())
    return addToOldToNewClosureMap(calleeValue, NewClosure);

  if (auto *CFI = dyn_cast<ConvertFunctionInst>(calleeValue)) {
    SILValue origCalleeValue = calleeValue;
    calleeValue = cloneCalleeConversion(CFI->getOperand(), NewClosure, Builder,
                                        NeedsRelease, CapturedMap);
    return addToOldToNewClosureMap(
        origCalleeValue, Builder.createConvertFunction(
                             CallSiteDesc.getLoc(), calleeValue, CFI->getType(),
                             CFI->withoutActuallyEscaping()));
  }

  if (auto *PAI = dyn_cast<PartialApplyInst>(calleeValue)) {
    assert(isPartialApplyOfReabstractionThunk(PAI) && isSupportedClosure(PAI) &&
           PAI->getArgument(0)
               ->getType()
               .getAs<SILFunctionType>()
               ->isTrivialNoEscape());
    SILValue origCalleeValue = calleeValue;
    calleeValue = cloneCalleeConversion(PAI->getArgument(0), NewClosure,
                                        Builder, NeedsRelease, CapturedMap);
    auto origRef = PAI->getReferencedFunctionOrNull();
    assert(origRef);
    auto FunRef = Builder.createFunctionRef(CallSiteDesc.getLoc(), origRef);
    auto NewPA = Builder.createPartialApply(
        CallSiteDesc.getLoc(), FunRef, {}, {calleeValue},
        PAI->getCalleeConvention(), PAI->getResultIsolation(),
        PAI->isOnStack());
    // If the partial_apply is on stack we will emit a dealloc_stack in the
    // epilog.
    NeedsRelease.push_back(NewPA);
    return addToOldToNewClosureMap(origCalleeValue, NewPA);
  }

  if (auto *MD = dyn_cast<MarkDependenceInst>(calleeValue)) {
    SILValue origCalleeValue = calleeValue;
    calleeValue = cloneCalleeConversion(MD->getValue(), NewClosure, Builder,
                                        NeedsRelease, CapturedMap);
    if (!CapturedMap.count(MD->getBase())) {
      CallSiteDesc.getClosure()->dump();
      MD->dump();
      MD->getFunction()->dump();
    }
    assert(CapturedMap.count(MD->getBase()));
    return addToOldToNewClosureMap(
        origCalleeValue,
        Builder.createMarkDependence(CallSiteDesc.getLoc(), calleeValue,
                                     CapturedMap[MD->getBase()],
                                     MarkDependenceKind::Escaping));
  }


  auto *Cvt = cast<ConvertEscapeToNoEscapeInst>(calleeValue);
  SILValue origCalleeValue = calleeValue;
  calleeValue = cloneCalleeConversion(Cvt->getOperand(), NewClosure, Builder,
                                      NeedsRelease, CapturedMap);
  return addToOldToNewClosureMap(
      origCalleeValue,
      Builder.createConvertEscapeToNoEscape(CallSiteDesc.getLoc(), calleeValue,
                                            Cvt->getType(), true));
}

/// Populate the body of the cloned closure, modifying instructions as
/// necessary. This is where we create the actual specialized BB Arguments
void ClosureSpecCloner::populateCloned() {
  bool invalidatedStackNesting = false;
  SILFunction *Cloned = getCloned();
  SILFunction *ClosureUser = CallSiteDesc.getApplyCallee();

  // Create arguments for the entry block.
  SILBasicBlock *ClosureUserEntryBB = &*ClosureUser->begin();
  SILBasicBlock *ClonedEntryBB = Cloned->createBasicBlock();

  SmallVector<SILValue, 4> entryArgs;
  entryArgs.reserve(ClosureUserEntryBB->getArguments().size());

  // Remove the closure argument.
  for (size_t i = 0, e = ClosureUserEntryBB->args_size(); i != e; ++i) {
    SILArgument *Arg = ClosureUserEntryBB->getArgument(i);
    if (i == CallSiteDesc.getClosureIndex()) {
      entryArgs.push_back(SILValue());
      continue;
    }

    // Otherwise, create a new argument which copies the original argument
    auto typeInContext = Cloned->getLoweredType(Arg->getType());
    auto *MappedValue =
        ClonedEntryBB->createFunctionArgument(typeInContext, Arg->getDecl());
    MappedValue->copyFlags(cast<SILFunctionArgument>(Arg));
    entryArgs.push_back(MappedValue);
  }

  // Next we need to add in any arguments that are not captured as arguments to
  // the cloned function.
  //
  // We do not insert the new mapped arguments into the value map since there by
  // definition is nothing in the partial apply user function that references
  // such arguments. After this pass is done the only thing that will reference
  // the arguments is the partial apply that we will create.
  SILFunction *ClosedOverFun = CallSiteDesc.getClosureCallee();
  SILBuilder &Builder = getBuilder();
  auto ClosedOverFunConv = ClosedOverFun->getConventions();
  unsigned NumTotalParams = ClosedOverFunConv.getNumParameters();
  unsigned NumNotCaptured = NumTotalParams - CallSiteDesc.getNumArguments();
  llvm::SmallVector<SILValue, 4> NewPAIArgs;
  llvm::DenseMap<SILValue, SILValue> CapturedMap;
  unsigned idx = 0;
  for (auto &PInfo : ClosedOverFunConv.getParameters().slice(NumNotCaptured)) {
    auto paramTy =
        ClosedOverFunConv.getSILType(PInfo, Builder.getTypeExpansionContext());
    // Get the type in context of the new function.
    paramTy = Cloned->getLoweredType(paramTy);
    SILValue MappedValue = ClonedEntryBB->createFunctionArgument(paramTy);
    NewPAIArgs.push_back(MappedValue);
    auto CapturedVal =
        cast<PartialApplyInst>(CallSiteDesc.getClosure())->getArgument(idx++);
    CapturedMap[CapturedVal] = MappedValue;
  }

  Builder.setInsertionPoint(ClonedEntryBB);

  // Clone FRI and PAI, and replace usage of the removed closure argument
  // with result of cloned PAI.
  SILValue FnVal =
      Builder.createFunctionRef(CallSiteDesc.getLoc(), ClosedOverFun);
  auto *NewClosure = CallSiteDesc.createNewClosure(Builder, FnVal, NewPAIArgs);

  // Clone a chain of ConvertFunctionInsts. This can create further
  // reabstraction partial_apply instructions.
  SmallVector<PartialApplyInst*, 4> NeedsRelease;
  SILValue ConvertedCallee =
      cloneCalleeConversion(CallSiteDesc.getClosureCallerArg(), NewClosure,
                            Builder, NeedsRelease, CapturedMap);

  // Make sure that we actually emit the releases for reabstraction thunks. We
  // have guaranteed earlier that we only allow reabstraction thunks if the
  // closure was passed trivial.
  assert(NeedsRelease.empty() || CallSiteDesc.isTrivialNoEscapeParameter());

  entryArgs[CallSiteDesc.getClosureIndex()] = ConvertedCallee;

  // Visit original BBs in depth-first preorder, starting with the
  // entry block, cloning all instructions and terminators.
  cloneFunctionBody(ClosureUser, ClonedEntryBB, entryArgs);

  // Then insert a release in all non failure exit BBs if our partial apply was
  // guaranteed. This is b/c it was passed at +0 originally and we need to
  // balance the initial increment of the newly created closure(s).
  bool ClosureHasRefSemantics = CallSiteDesc.closureHasRefSemanticContext();
  if ((CallSiteDesc.isClosureGuaranteed() ||
       CallSiteDesc.isTrivialNoEscapeParameter()) &&
      (ClosureHasRefSemantics || !NeedsRelease.empty() ||
       CallSiteDesc.isClosureOnStack())) {
    for (SILBasicBlock *BB : CallSiteDesc.getNonFailureExitBBs()) {
      SILBasicBlock *OpBB = getOpBasicBlock(BB);

      TermInst *TI = OpBB->getTerminator();
      auto Loc = CleanupLocation(NewClosure->getLoc());

      // If we have an exit, we place the release right before it so we know
      // that it will be executed at the end of the epilogue.
      if (TI->isFunctionExiting()) {
        Builder.setInsertionPoint(TI);
        if (ClosureHasRefSemantics)
          Builder.createReleaseValue(Loc, SILValue(NewClosure),
                                     Builder.getDefaultAtomicity());
        else
          invalidatedStackNesting |=
              CallSiteDesc.destroyIfPartialApplyStack(Builder, NewClosure);
        for (auto PAI : NeedsRelease) {
          if (PAI->isOnStack())
            invalidatedStackNesting |=
                CallSiteDesc.destroyIfPartialApplyStack(Builder, PAI);
          else
            Builder.createReleaseValue(Loc, SILValue(PAI),
                                       Builder.getDefaultAtomicity());
        }
        continue;
      }

      // We use casts where findAllNonFailureExitBBs should have made sure that
      // this is true. This will ensure that the code is updated when we hit the
      // cast failure in debug builds.
      auto *Unreachable = cast<UnreachableInst>(TI);
      auto PrevIter = std::prev(SILBasicBlock::iterator(Unreachable));
      auto NoReturnApply = FullApplySite::isa(&*PrevIter);

      // We insert the release value right before the no return apply so that if
      // the partial apply is passed into the no-return function as an @owned
      // value, we will retain the partial apply before we release it and
      // potentially eliminate it.
      Builder.setInsertionPoint(NoReturnApply.getInstruction());
      if (ClosureHasRefSemantics)
        Builder.createReleaseValue(Loc, SILValue(NewClosure),
                                   Builder.getDefaultAtomicity());
      else
        invalidatedStackNesting |=
            CallSiteDesc.destroyIfPartialApplyStack(Builder, NewClosure);
      for (auto PAI : NeedsRelease) {
        if (PAI->isOnStack())
          invalidatedStackNesting |=
              CallSiteDesc.destroyIfPartialApplyStack(Builder, PAI);
        else
          Builder.createReleaseValue(Loc, SILValue(PAI),
                                     Builder.getDefaultAtomicity());
      }
    }
  }
  if (invalidatedStackNesting) {
    StackNesting::fixNesting(Cloned);
  }
}

//===----------------------------------------------------------------------===//
//                            Closure Specializer
//===----------------------------------------------------------------------===//

namespace {

class SILClosureSpecializerTransform : public SILFunctionTransform {
  bool gatherCallSites(
      SILFunction *Caller,
      llvm::SmallVectorImpl<std::unique_ptr<ClosureInfo>> &ClosureCandidates,
      llvm::DenseSet<FullApplySite> &MultipleClosureAI);
  bool specialize(SILFunction *Caller,
                  std::vector<SingleValueInstruction *> &PropagatedClosures);

public:
  SILClosureSpecializerTransform() {}

  void run() override;

};

void SILClosureSpecializerTransform::run() {
  SILFunction *F = getFunction();

  // Don't optimize functions that are marked with the opt.never
  // attribute.
  if (!F->shouldOptimize())
    return;

  // If F is an external declaration, there is nothing to specialize.
  if (F->isExternalDeclaration())
    return;

  std::vector<SingleValueInstruction *> PropagatedClosures;

  if (!specialize(F, PropagatedClosures))
    return;

  // If for testing purposes we were asked to not eliminate dead closures,
  // return.
  if (EliminateDeadClosures) {
    // Otherwise, remove any local dead closures that are now dead since we
    // specialized all of their uses.
    LLVM_DEBUG(llvm::dbgs() << "Trying to remove dead closures!\n");
    sortUnique(PropagatedClosures);
    bool invalidatedStackNesting = false;

    for (auto *Closure : PropagatedClosures) {
      LLVM_DEBUG(llvm::dbgs() << "    Visiting: " << *Closure);
      if (!tryDeleteDeadClosure(Closure)) {
        LLVM_DEBUG(llvm::dbgs() << "        Failed to delete closure!\n");
        ++NumPropagatedClosuresNotEliminated;
        continue;
      }

      LLVM_DEBUG(llvm::dbgs() << "        Deleted closure!\n");
      ++NumPropagatedClosuresEliminated;
      invalidatedStackNesting = true;
    }

    if (invalidatedStackNesting) {
      StackNesting::fixNesting(F);
    }
  }

  // Invalidate everything since we delete calls as well as add new
  // calls and branches.
  invalidateAnalysis(SILAnalysis::InvalidationKind::FunctionBody);
}

static void markReabstractionPartialApplyAsUsed(
    SILValue FirstClosure, SILValue Current,
    llvm::DenseSet<SingleValueInstruction *> &UsedReabstractionClosure) {
  if (Current == FirstClosure)
    return;
  if (auto PA = dyn_cast<PartialApplyInst>(Current)) {
    UsedReabstractionClosure.insert(PA);
    return markReabstractionPartialApplyAsUsed(FirstClosure, PA->getArgument(0),
                                        UsedReabstractionClosure);
  }
  if (auto Cvt = dyn_cast<ConvertFunctionInst>(Current)) {
    return markReabstractionPartialApplyAsUsed(FirstClosure, Cvt->getOperand(),
                                               UsedReabstractionClosure);
  }
  if (auto Cvt = dyn_cast<ConvertEscapeToNoEscapeInst>(Current)) {
    return markReabstractionPartialApplyAsUsed(FirstClosure, Cvt->getOperand(),
                                               UsedReabstractionClosure);
  }
  if (auto MD = dyn_cast<MarkDependenceInst>(Current)) {
    return markReabstractionPartialApplyAsUsed(FirstClosure, MD->getValue(),
                                               UsedReabstractionClosure);
  }
  llvm_unreachable("Unexpect instruction");
}

/// Returns true if the \p closureArgIdx argument of \p callee is called in
/// \p callee or any function called by callee.
static bool isClosureAppliedIn(SILFunction *Callee, unsigned closureArgIdx,
                               SmallPtrSetImpl<SILFunction *> &HandledFuncs) {
  // Limit the number of recursive calls to not go into exponential behavior in
  // corner cases.
  const int RecursionBudget = 8;

  SILValue Arg = Callee->getArgument(closureArgIdx);
  for (Operand *ArgUse : Arg->getUses()) {
    if (auto UserAI = FullApplySite::isa(ArgUse->getUser())) {
      if (UserAI.getCallee() == Arg)
        return true;

      assert(UserAI.isArgumentOperand(*ArgUse) &&
             "any other non-argument operands than the callee?");

      SILFunction *ApplyCallee = UserAI.getReferencedFunctionOrNull();
      if (ApplyCallee && !ApplyCallee->isExternalDeclaration() &&
          HandledFuncs.count(ApplyCallee) == 0 &&
          HandledFuncs.size() < RecursionBudget) {
        HandledFuncs.insert(ApplyCallee);
        if (isClosureAppliedIn(UserAI.getReferencedFunctionOrNull(),
                               UserAI.getCalleeArgIndex(*ArgUse), HandledFuncs))
          return true;
      }
    }
  }
  return false;
}

static bool canSpecializeFullApplySite(FullApplySiteKind kind) {
  switch (kind) {
  case FullApplySiteKind::TryApplyInst:
  case FullApplySiteKind::ApplyInst:
    return true;
  case FullApplySiteKind::BeginApplyInst:
    return false;
  }
  llvm_unreachable("covered switch");
}

const int SpecializationLevelLimit = 2;

static int getSpecializationLevelRecursive(StringRef funcName, Demangler &parent) {
  using namespace Demangle;

  Demangler demangler;
  demangler.providePreallocatedMemory(parent);

  // Check for this kind of node tree:
  //
  // kind=Global
  //   kind=FunctionSignatureSpecialization
  //     kind=SpecializationPassID, index=1
  //     kind=FunctionSignatureSpecializationParam
  //       kind=FunctionSignatureSpecializationParamKind, index=5
  //       kind=FunctionSignatureSpecializationParamPayload, text="..."
  //
  Node *root = demangler.demangleSymbol(funcName);
  if (!root)
    return 0;
  if (root->getKind() != Node::Kind::Global)
    return 0;
  Node *funcSpec = root->getFirstChild();
  if (!funcSpec || funcSpec->getNumChildren() < 2)
    return 0;
  if (funcSpec->getKind() != Node::Kind::FunctionSignatureSpecialization)
    return 0;

  // Match any function specialization. We check for constant propagation at the
  // parameter level.
  Node *param = funcSpec->getChild(0);
  if (param->getKind() != Node::Kind::SpecializationPassID)
    return SpecializationLevelLimit + 1; // unrecognized format

  unsigned maxParamLevel = 0;
  for (unsigned paramIdx = 1; paramIdx < funcSpec->getNumChildren();
       ++paramIdx) {
    Node *param = funcSpec->getChild(paramIdx);
    if (param->getKind() != Node::Kind::FunctionSignatureSpecializationParam)
      return SpecializationLevelLimit + 1; // unrecognized format

    // A parameter is recursive if it has a kind with index and type payload
    if (param->getNumChildren() < 2)
      continue;

    Node *kindNd = param->getChild(0);
    if (kindNd->getKind()
        != Node::Kind::FunctionSignatureSpecializationParamKind) {
      return SpecializationLevelLimit + 1; // unrecognized format
    }
    auto kind = FunctionSigSpecializationParamKind(kindNd->getIndex());
    if (kind != FunctionSigSpecializationParamKind::ConstantPropFunction)
      continue;
    Node *payload = param->getChild(1);
    if (payload->getKind()
        != Node::Kind::FunctionSignatureSpecializationParamPayload) {
      return SpecializationLevelLimit + 1; // unrecognized format
    }
    // Check if the specialized function is a specialization itself.
    unsigned paramLevel =
      1 + getSpecializationLevelRecursive(payload->getText(), demangler);
    if (paramLevel > maxParamLevel)
      maxParamLevel = paramLevel;
  }
  return maxParamLevel;
}

/// If \p function is a function-signature specialization for a constant-
/// propagated function argument, returns 1.
/// If \p function is a specialization of such a specialization, returns 2.
/// And so on.
static int getSpecializationLevel(SILFunction *f) {
  Demangle::StackAllocatedDemangler<1024> demangler;
  return getSpecializationLevelRecursive(f->getName(), demangler);
}

bool SILClosureSpecializerTransform::gatherCallSites(
    SILFunction *Caller,
    llvm::SmallVectorImpl<std::unique_ptr<ClosureInfo>> &ClosureCandidates,
    llvm::DenseSet<FullApplySite> &MultipleClosureAI) {

  // A set of apply inst that we have associated with a closure. We use this to
  // make sure that we do not handle call sites with multiple closure arguments.
  llvm::DenseSet<FullApplySite> VisitedAI;

  // We should not look at reabstraction closure twice who we ultimately ended
  // up using as an argument that we specialize on.
  llvm::DenseSet<SingleValueInstruction *> UsedReabstractionClosure;

  bool CFGChanged = false;

  // For each basic block BB in Caller...
  for (auto &BB : *Caller) {

    // For each instruction II in BB...
    for (auto &II : BB) {
      // If II is not a closure that we support specializing, skip it...
      if (!isSupportedClosure(&II))
        continue;
      auto ClosureInst = cast<SingleValueInstruction>(&II);
      if (UsedReabstractionClosure.count(ClosureInst))
        continue;

      std::unique_ptr<ClosureInfo> CInfo;

      // Go through all uses of our closure.

      // Worklist of operands.
      SmallVector<Operand *, 8> Uses(ClosureInst->getUses());

      // Live range end points.
      SmallVector<SILInstruction *, 8> UsePoints;

      // Set of possible arguments for mark_dependence. The base of a
      // mark_dependence we copy must be available in the specialized function.
      llvm::SmallSet<SILValue, 16> PossibleMarkDependenceBases;
      if (auto *PA = dyn_cast<PartialApplyInst>(ClosureInst)) {
        for (auto Opd : PA->getArguments())
          PossibleMarkDependenceBases.insert(Opd);
      }

      bool HaveUsedReabstraction = false;
      // Uses may grow in this loop.
      for (size_t UseIndex = 0; UseIndex < Uses.size(); ++UseIndex) {
        auto *Use = Uses[UseIndex];
        UsePoints.push_back(Use->getUser());

        // Recurse through conversions.
        if (auto *CFI = dyn_cast<ConvertFunctionInst>(Use->getUser())) {
          // Push Uses in reverse order so they are visited in forward order.
          Uses.append(CFI->getUses().begin(), CFI->getUses().end());
          PossibleMarkDependenceBases.insert(CFI);
          continue;
        }
        if (auto *Cvt = dyn_cast<ConvertEscapeToNoEscapeInst>(Use->getUser())) {
          Uses.append(Cvt->getUses().begin(), Cvt->getUses().end());
          PossibleMarkDependenceBases.insert(Cvt);
          continue;
        }

        // Look through reabstraction thunks.
        if (auto *PA = dyn_cast<PartialApplyInst>(Use->getUser())) {
          // Reabstraction can cause series of partial_apply to be emitted. It
          // is okay to treat these like conversion instructions. Current
          // restriction: if the partial_apply does not take ownership of its
          // argument we don't need to analyze which partial_apply to emit
          // release for (its all of them).
          if (isPartialApplyOfReabstractionThunk(PA) &&
              isSupportedClosure(PA) &&
              PA->getArgument(0)
                  ->getType()
                  .getAs<SILFunctionType>()
                  ->isTrivialNoEscape()) {
            Uses.append(PA->getUses().begin(), PA->getUses().end());
            PossibleMarkDependenceBases.insert(PA);
            HaveUsedReabstraction = true;
          }
          continue;
        }

        // Look through mark_dependence on partial_apply [stack].
        if (auto *MD = dyn_cast<MarkDependenceInst>(Use->getUser())) {
          // We can't copy a closure if the mark_dependence base is not
          // available in the specialized function.
          if (!PossibleMarkDependenceBases.count(MD->getBase()))
            continue;
          if (MD->getValue() == Use->get() &&
              MD->getValue()->getType().is<SILFunctionType>() &&
              MD->getValue()
                  ->getType()
                  .castTo<SILFunctionType>()
                  ->isTrivialNoEscape()) {
            Uses.append(MD->getUses().begin(), MD->getUses().end());
            continue;
          }
        }

        // If this use is not a full apply site that we can process or an apply
        // inst with substitutions, there is nothing interesting for us to do,
        // so continue...
        auto AI = FullApplySite::isa(Use->getUser());
        if (!AI || AI.hasSubstitutions() ||
            !canSpecializeFullApplySite(AI.getKind()) ||
            !AI.canOptimize())
          continue;

        // Check if we have already associated this apply inst with a closure to
        // be specialized. We do not handle applies that take in multiple
        // closures at this time.
        if (!VisitedAI.insert(AI).second) {
          MultipleClosureAI.insert(AI);
          continue;
        }

        // If AI does not have a function_ref definition as its callee, we can
        // not do anything here... so continue...
        SILFunction *ApplyCallee = AI.getReferencedFunctionOrNull();
        if (!ApplyCallee || ApplyCallee->isExternalDeclaration())
          continue;

        // Don't specialize non-fragile callees if the caller is fragile;
        // the specialized callee will have shared linkage, and thus cannot
        // be referenced from the fragile caller.
        if (!ApplyCallee->canBeInlinedIntoCaller(Caller->getSerializedKind()))
          continue;

        // If the callee uses a dynamic Self, we cannot specialize it,
        // since the resulting specialization might longer has 'self' as the
        // last parameter.
        //
        // We could fix this by inserting new arguments more carefully, or
        // changing how we model dynamic Self altogether.
        if (mayBindDynamicSelf(ApplyCallee))
          return CFGChanged;

        // Check if the closure is passed as an argument (and not called).
        if (!AI.isArgumentOperand(*Use))
          continue;

        unsigned ClosureIndex = AI.getCalleeArgIndex(*Use);

        // Ok, we know that we can perform the optimization but not whether or
        // not the optimization is profitable. Check if the closure is actually
        // called in the callee (or in a function called by the callee).
        SmallPtrSet<SILFunction *, 8> HandledFuncs;
        if (!isClosureAppliedIn(ApplyCallee, ClosureIndex, HandledFuncs))
          continue;

        unsigned firstParamArgIdx =
            AI.getSubstCalleeConv().getSILArgIndexOfFirstParam();
        assert(ClosureIndex >= firstParamArgIdx);
        auto ClosureParamIndex = ClosureIndex - firstParamArgIdx;

        auto ParamInfo = AI.getSubstCalleeType()->getParameters();
        SILParameterInfo ClosureParamInfo = ParamInfo[ClosureParamIndex];

        // We currently only support copying intermediate reabstraction
        // closures if the closure is ultimately passed trivially.
        bool IsClosurePassedTrivially = ClosureParamInfo.getInterfaceType()
                                            ->castTo<SILFunctionType>()
                                            ->isTrivialNoEscape();
        if (HaveUsedReabstraction &&  !IsClosurePassedTrivially)
          continue;

        // Get all non-failure exit BBs in the Apply Callee if our partial apply
        // is guaranteed. If we do not understand one of the exit BBs, bail.
        //
        // We need this to make sure that we insert a release in the appropriate
        // locations to balance the +1 from the creation of the partial apply.
        //
        // However, thin_to_thick_function closures don't have a context and
        // don't need to be released.
        bool OnlyHaveThinToThickClosure =
            isa<ThinToThickFunctionInst>(ClosureInst) && !HaveUsedReabstraction;

        llvm::TinyPtrVector<SILBasicBlock *> NonFailureExitBBs;
        if ((ClosureParamInfo.isGuaranteed() || IsClosurePassedTrivially) &&
            !OnlyHaveThinToThickClosure &&
            !findAllNonFailureExitBBs(ApplyCallee, NonFailureExitBBs)) {
          continue;
        }

        // Specializing a readnone, readonly, releasenone function with a
        // nontrivial context is illegal. Inserting a release in such a function
        // results in miscompilation after other optimizations.
        // For now, the specialization is disabled.
        //
        // TODO: A @noescape closure should never be converted to an @owned
        // argument regardless of the function attribute.
        if (!OnlyHaveThinToThickClosure
            && ApplyCallee->getEffectsKind() <= EffectsKind::ReleaseNone) {
          continue;
        }

        // Avoid an infinite specialization loop caused by repeated runs of
        // ClosureSpecializer and CapturePropagation.
        // CapturePropagation propagates constant function-literals. Such
        // function specializations can then be optimized again by the
        // ClosureSpecializer and so on.
        // This happens if a closure argument is called _and_ referenced in
        // another closure, which is passed to a recursive call. E.g.
        //
        // func foo(_ c: @escaping () -> ()) {
        //   c()
        //   foo({ c() })
        // }
        //
        // A limit of 2 is good enough and will not be exceeded in "regular"
        // optimization scenarios.
        if (getSpecializationLevel(getClosureCallee(ClosureInst))
            > SpecializationLevelLimit) {
          continue;
        }
        // Compute the final release points of the closure. We will insert
        // release of the captured arguments here.
        if (!CInfo)
          CInfo.reset(new ClosureInfo(ClosureInst));

        // Mark the reabstraction closures as used.
        if (HaveUsedReabstraction)
          markReabstractionPartialApplyAsUsed(ClosureInst, Use->get(),
                                              UsedReabstractionClosure);
        // Now we know that CSDesc is profitable to specialize. Add it to our
        // call site list.
        CInfo->CallSites.push_back(
          CallSiteDescriptor(CInfo.get(), AI, ClosureIndex,
                             ClosureParamInfo, std::move(NonFailureExitBBs)));
      }
      if (CInfo) {
        ValueLifetimeAnalysis VLA(CInfo->Closure, UsePoints);
        if (!VLA.computeFrontier(CInfo->LifetimeFrontier,
                                 ValueLifetimeAnalysis::AllowToModifyCFG)) {
          CFGChanged = true;
        }
        ClosureCandidates.push_back(std::move(CInfo));
      }
    }
  }
  return CFGChanged;
}

bool SILClosureSpecializerTransform::specialize(SILFunction *Caller,
                    std::vector<SingleValueInstruction *> &PropagatedClosures) {
  LLVM_DEBUG(llvm::dbgs() << "Optimizing callsites that take closure "
                             "argument in "
                          << Caller->getName() << '\n');

  // Collect all of the PartialApplyInsts that are used as arguments to
  // ApplyInsts. Check the profitability of specializing the closure argument.
  llvm::SmallVector<std::unique_ptr<ClosureInfo>, 8> ClosureCandidates;
  llvm::DenseSet<FullApplySite> MultipleClosureAI;
  if (gatherCallSites(Caller, ClosureCandidates, MultipleClosureAI)) {
    invalidateAnalysis(SILAnalysis::InvalidationKind::Branches);
  }

  SILOptFunctionBuilder FuncBuilder(*this);
  bool Changed = false;
  for (const auto &CInfo : ClosureCandidates) {
    for (auto &CSDesc : CInfo->CallSites) {
      // Do not specialize apply insts that take in multiple closures. This pass
      // does not know how to do this yet.
      if (MultipleClosureAI.count(CSDesc.getApplyInst()))
        continue;

      auto NewFName = CSDesc.createName();
      LLVM_DEBUG(llvm::dbgs() << "    Perform optimizations with new name "
                              << NewFName << '\n');

      // Then see if we already have a specialized version of this function in
      // our module.
      SILFunction *NewF = CInfo->Closure->getModule().lookUpFunction(NewFName);

      // If not, create a specialized version of ApplyCallee calling the closure
      // directly.
      if (!NewF) {
        NewF = ClosureSpecCloner::cloneFunction(FuncBuilder, CSDesc, NewFName);
        addFunctionToPassManagerWorklist(NewF, CSDesc.getApplyCallee());
        LLVM_DEBUG(llvm::dbgs() << "\nThe rewritten callee is:\n";
                   NewF->dump());
      }

      // Rewrite the call
      rewriteApplyInst(CSDesc, NewF);

      PropagatedClosures.push_back(CSDesc.getClosure());
      Changed = true;
    }
  }
  LLVM_DEBUG(if (Changed) {
      llvm::dbgs() << "\nThe rewritten caller is:\n";
      Caller->dump();
    });
  return Changed;
}

} // end anonymous namespace

SILTransform *swift::createClosureSpecializer() {
  return new SILClosureSpecializerTransform();
}