OMPIRBuilder.cpp

Go to the documentation of this file.

//===- OpenMPIRBuilder.cpp - Builder for LLVM-IR for OpenMP directives ----===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
/// \file
///
/// This file implements the OpenMPIRBuilder class, which is used as a
/// convenient way to create LLVM instructions for OpenMP directives.
///
//===----------------------------------------------------------------------===//
 
#include "llvm/Frontend/OpenMP/OMPIRBuilder.h"
#include "llvm/ADT/SmallBitVector.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVectorExtras.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/CodeMetrics.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/PostDominators.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Bitcode/BitcodeReader.h"
#include "llvm/Frontend/Offloading/Utility.h"
#include "llvm/Frontend/OpenMP/OMPGridValues.h"
#include "llvm/IR/Attributes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/Constant.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DIBuilder.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/MDBuilder.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/PassInstrumentation.h"
#include "llvm/IR/PassManager.h"
#include "llvm/IR/ReplaceConstant.h"
#include "llvm/IR/Value.h"
#include "llvm/MC/TargetRegistry.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Error.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/NVVMAttributes.h"
#include "llvm/Support/VirtualFileSystem.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Transforms/Utils/CodeExtractor.h"
#include "llvm/Transforms/Utils/LoopPeel.h"
#include "llvm/Transforms/Utils/UnrollLoop.h"
 
#include <cstdint>
#include <optional>
 
#define DEBUG_TYPE "openmp-ir-builder"
 
using namespace llvm;
using namespace omp;
 
static cl::opt<bool>
    OptimisticAttributes("openmp-ir-builder-optimistic-attributes", cl::Hidden,
                         cl::desc("Use optimistic attributes describing "
                                  "'as-if' properties of runtime calls."),
                         cl::init(false));
 
static cl::opt<double> UnrollThresholdFactor(
    "openmp-ir-builder-unroll-threshold-factor", cl::Hidden,
    cl::desc("Factor for the unroll threshold to account for code "
             "simplifications still taking place"),
    cl::init(1.5));
 
static cl::opt<bool> UseDefaultMaxThreads(
    "openmp-ir-builder-use-default-max-threads", cl::Hidden,
    cl::desc("Use a default max threads if none is provided."), cl::init(true));
 
#ifndef NDEBUG
/// Return whether IP1 and IP2 are ambiguous, i.e. that inserting instructions
/// at position IP1 may change the meaning of IP2 or vice-versa. This is because
/// an InsertPoint stores the instruction before something is inserted. For
/// instance, if both point to the same instruction, two IRBuilders alternating
/// creating instruction will cause the instructions to be interleaved.
static bool isConflictIP(IRBuilder<>::InsertPoint IP1,
                         IRBuilder<>::InsertPoint IP2) {
  if (!IP1.isSet() || !IP2.isSet())
    return false;
  return IP1.getBlock() == IP2.getBlock() && IP1.getPoint() == IP2.getPoint();
}
 
static bool isValidWorkshareLoopScheduleType(OMPScheduleType SchedType) {
  // Valid ordered/unordered and base algorithm combinations.
  switch (SchedType & ~OMPScheduleType::MonotonicityMask) {
  case OMPScheduleType::UnorderedStaticChunked:
  case OMPScheduleType::UnorderedStatic:
  case OMPScheduleType::UnorderedDynamicChunked:
  case OMPScheduleType::UnorderedGuidedChunked:
  case OMPScheduleType::UnorderedRuntime:
  case OMPScheduleType::UnorderedAuto:
  case OMPScheduleType::UnorderedTrapezoidal:
  case OMPScheduleType::UnorderedGreedy:
  case OMPScheduleType::UnorderedBalanced:
  case OMPScheduleType::UnorderedGuidedIterativeChunked:
  case OMPScheduleType::UnorderedGuidedAnalyticalChunked:
  case OMPScheduleType::UnorderedSteal:
  case OMPScheduleType::UnorderedStaticBalancedChunked:
  case OMPScheduleType::UnorderedGuidedSimd:
  case OMPScheduleType::UnorderedRuntimeSimd:
  case OMPScheduleType::OrderedStaticChunked:
  case OMPScheduleType::OrderedStatic:
  case OMPScheduleType::OrderedDynamicChunked:
  case OMPScheduleType::OrderedGuidedChunked:
  case OMPScheduleType::OrderedRuntime:
  case OMPScheduleType::OrderedAuto:
  case OMPScheduleType::OrderdTrapezoidal:
  case OMPScheduleType::NomergeUnorderedStaticChunked:
  case OMPScheduleType::NomergeUnorderedStatic:
  case OMPScheduleType::NomergeUnorderedDynamicChunked:
  case OMPScheduleType::NomergeUnorderedGuidedChunked:
  case OMPScheduleType::NomergeUnorderedRuntime:
  case OMPScheduleType::NomergeUnorderedAuto:
  case OMPScheduleType::NomergeUnorderedTrapezoidal:
  case OMPScheduleType::NomergeUnorderedGreedy:
  case OMPScheduleType::NomergeUnorderedBalanced:
  case OMPScheduleType::NomergeUnorderedGuidedIterativeChunked:
  case OMPScheduleType::NomergeUnorderedGuidedAnalyticalChunked:
  case OMPScheduleType::NomergeUnorderedSteal:
  case OMPScheduleType::NomergeOrderedStaticChunked:
  case OMPScheduleType::NomergeOrderedStatic:
  case OMPScheduleType::NomergeOrderedDynamicChunked:
  case OMPScheduleType::NomergeOrderedGuidedChunked:
  case OMPScheduleType::NomergeOrderedRuntime:
  case OMPScheduleType::NomergeOrderedAuto:
  case OMPScheduleType::NomergeOrderedTrapezoidal:
  case OMPScheduleType::OrderedDistributeChunked:
  case OMPScheduleType::OrderedDistribute:
    break;
  default:
    return false;
  }
 
  // Must not set both monotonicity modifiers at the same time.
  OMPScheduleType MonotonicityFlags =
      SchedType & OMPScheduleType::MonotonicityMask;
  if (MonotonicityFlags == OMPScheduleType::MonotonicityMask)
    return false;
 
  return true;
}
#endif
 
/// This is a wrapper over IRBuilderBase::restoreIP that also restores a current
/// debug location when the insert point is at the end of a block. It picks a
/// location scoped to the current function: the block's last instruction
/// location if the block is non-empty, otherwise a location synthesized from
///  the function's subprogram (when the function has debug info).
static void restoreIPandDebugLoc(llvm::IRBuilderBase &Builder,
                                 llvm::IRBuilderBase::InsertPoint IP) {
  Builder.restoreIP(IP);
  // When IP points at a real instruction, restoreIP (SetInsertPoint) already
  // set the debug location from that instruction, so leave it alone.
  llvm::BasicBlock *BB = Builder.GetInsertBlock();
  if (Builder.GetInsertPoint() != BB->end())
    return;
 
  // At the end of a block, pick a location guaranteed to belong to the current
  // insertion function's subprogram. Prefer the block's own last instruction;
  // otherwise synthesize a location from the function's subprogram.
  if (!BB->empty())
    Builder.SetCurrentDebugLocation(BB->back().getStableDebugLoc());
  else if (llvm::DISubprogram *FSP =
               BB->getParent() ? BB->getParent()->getSubprogram() : nullptr) {
    unsigned Line = FSP->getScopeLine() ? FSP->getScopeLine() : FSP->getLine();
    Builder.SetCurrentDebugLocation(
        llvm::DILocation::get(FSP->getContext(), Line, /*Column=*/0, FSP));
  }
}
 
static bool hasGridValue(const Triple &T) {
  return T.isAMDGPU() || T.isNVPTX() || T.isSPIRV();
}
 
static const omp::GV &getGridValue(const Triple &T, Function *Kernel) {
  if (T.isAMDGPU()) {
    StringRef Features =
        Kernel->getFnAttribute("target-features").getValueAsString();
    if (Features.count("+wavefrontsize64"))
      return omp::getAMDGPUGridValues<64>();
    return omp::getAMDGPUGridValues<32>();
  }
  if (T.isNVPTX())
    return omp::NVPTXGridValues;
  if (T.isSPIRV())
    return omp::SPIRVGridValues;
  llvm_unreachable("No grid value available for this architecture!");
}
 
/// Determine which scheduling algorithm to use, determined from schedule clause
/// arguments.
static OMPScheduleType
getOpenMPBaseScheduleType(llvm::omp::ScheduleKind ClauseKind, bool HasChunks,
                          bool HasSimdModifier, bool HasDistScheduleChunks) {
  // Currently, the default schedule it static.
  switch (ClauseKind) {
  case OMP_SCHEDULE_Default:
  case OMP_SCHEDULE_Static:
    return HasChunks ? OMPScheduleType::BaseStaticChunked
                     : OMPScheduleType::BaseStatic;
  case OMP_SCHEDULE_Dynamic:
    return OMPScheduleType::BaseDynamicChunked;
  case OMP_SCHEDULE_Guided:
    return HasSimdModifier ? OMPScheduleType::BaseGuidedSimd
                           : OMPScheduleType::BaseGuidedChunked;
  case OMP_SCHEDULE_Auto:
    return llvm::omp::OMPScheduleType::BaseAuto;
  case OMP_SCHEDULE_Runtime:
    return HasSimdModifier ? OMPScheduleType::BaseRuntimeSimd
                           : OMPScheduleType::BaseRuntime;
  case OMP_SCHEDULE_Distribute:
    return HasDistScheduleChunks ? OMPScheduleType::BaseDistributeChunked
                                 : OMPScheduleType::BaseDistribute;
  }
  llvm_unreachable("unhandled schedule clause argument");
}
 
/// Adds ordering modifier flags to schedule type.
static OMPScheduleType
getOpenMPOrderingScheduleType(OMPScheduleType BaseScheduleType,
                              bool HasOrderedClause) {
  assert((BaseScheduleType & OMPScheduleType::ModifierMask) ==
             OMPScheduleType::None &&
         "Must not have ordering nor monotonicity flags already set");
 
  OMPScheduleType OrderingModifier = HasOrderedClause
                                         ? OMPScheduleType::ModifierOrdered
                                         : OMPScheduleType::ModifierUnordered;
  OMPScheduleType OrderingScheduleType = BaseScheduleType | OrderingModifier;
 
  // Unsupported combinations
  if (OrderingScheduleType ==
      (OMPScheduleType::BaseGuidedSimd | OMPScheduleType::ModifierOrdered))
    return OMPScheduleType::OrderedGuidedChunked;
  else if (OrderingScheduleType == (OMPScheduleType::BaseRuntimeSimd |
                                    OMPScheduleType::ModifierOrdered))
    return OMPScheduleType::OrderedRuntime;
 
  return OrderingScheduleType;
}
 
/// Adds monotonicity modifier flags to schedule type.
static OMPScheduleType
getOpenMPMonotonicityScheduleType(OMPScheduleType ScheduleType,
                                  bool HasSimdModifier, bool HasMonotonic,
                                  bool HasNonmonotonic, bool HasOrderedClause) {
  assert((ScheduleType & OMPScheduleType::MonotonicityMask) ==
             OMPScheduleType::None &&
         "Must not have monotonicity flags already set");
  assert((!HasMonotonic || !HasNonmonotonic) &&
         "Monotonic and Nonmonotonic are contradicting each other");
 
  if (HasMonotonic) {
    return ScheduleType | OMPScheduleType::ModifierMonotonic;
  } else if (HasNonmonotonic) {
    return ScheduleType | OMPScheduleType::ModifierNonmonotonic;
  } else {
    // OpenMP 5.1, 2.11.4 Worksharing-Loop Construct, Description.
    // If the static schedule kind is specified or if the ordered clause is
    // specified, and if the nonmonotonic modifier is not specified, the
    // effect is as if the monotonic modifier is specified. Otherwise, unless
    // the monotonic modifier is specified, the effect is as if the
    // nonmonotonic modifier is specified.
    OMPScheduleType BaseScheduleType =
        ScheduleType & ~OMPScheduleType::ModifierMask;
    if ((BaseScheduleType == OMPScheduleType::BaseStatic) ||
        (BaseScheduleType == OMPScheduleType::BaseStaticChunked) ||
        HasOrderedClause) {
      // The monotonic is used by default in openmp runtime library, so no need
      // to set it.
      return ScheduleType;
    } else {
      return ScheduleType | OMPScheduleType::ModifierNonmonotonic;
    }
  }
}
 
/// Determine the schedule type using schedule and ordering clause arguments.
static OMPScheduleType
computeOpenMPScheduleType(ScheduleKind ClauseKind, bool HasChunks,
                          bool HasSimdModifier, bool HasMonotonicModifier,
                          bool HasNonmonotonicModifier, bool HasOrderedClause,
                          bool HasDistScheduleChunks) {
  OMPScheduleType BaseSchedule = getOpenMPBaseScheduleType(
      ClauseKind, HasChunks, HasSimdModifier, HasDistScheduleChunks);
  OMPScheduleType OrderedSchedule =
      getOpenMPOrderingScheduleType(BaseSchedule, HasOrderedClause);
  OMPScheduleType Result = getOpenMPMonotonicityScheduleType(
      OrderedSchedule, HasSimdModifier, HasMonotonicModifier,
      HasNonmonotonicModifier, HasOrderedClause);
 
  assert(isValidWorkshareLoopScheduleType(Result));
  return Result;
}
 
/// Given a function, if it represents the entry point of a target kernel, this
/// returns the execution mode flags associated with that kernel.
static std::optional<omp::OMPTgtExecModeFlags>
getTargetKernelExecMode(Function &Kernel) {
  CallInst *TargetInitCall = nullptr;
  for (Instruction &Inst : Kernel.getEntryBlock()) {
    if (auto *Call = dyn_cast<CallInst>(&Inst)) {
      if (Call->getCalledFunction()->getName() == "__kmpc_target_init") {
        TargetInitCall = Call;
        break;
      }
    }
  }
 
  if (!TargetInitCall)
    return std::nullopt;
 
  // Get the kernel mode information from the global variable associated to the
  // first argument to the call to __kmpc_target_init. Refer to
  // createTargetInit() to see how this is initialized.
  Value *InitOperand = TargetInitCall->getArgOperand(0);
  GlobalVariable *KernelEnv = nullptr;
  if (auto *Cast = dyn_cast<ConstantExpr>(InitOperand))
    KernelEnv = cast<GlobalVariable>(Cast->getOperand(0));
  else
    KernelEnv = cast<GlobalVariable>(InitOperand);
  auto *KernelEnvInit = cast<ConstantStruct>(KernelEnv->getInitializer());
  auto *ConfigEnv = cast<ConstantStruct>(KernelEnvInit->getOperand(0));
  auto *KernelMode = cast<ConstantInt>(ConfigEnv->getOperand(2));
  return static_cast<OMPTgtExecModeFlags>(KernelMode->getZExtValue());
}
 
static bool isGenericKernel(Function &Fn) {
  std::optional<omp::OMPTgtExecModeFlags> ExecMode =
      getTargetKernelExecMode(Fn);
  return !ExecMode || (*ExecMode & OMP_TGT_EXEC_MODE_GENERIC);
}
 
/// Make \p Source branch to \p Target.
///
/// Handles two situations:
/// * \p Source already has an unconditional branch.
/// * \p Source is a degenerate block (no terminator because the BB is
///             the current head of the IR construction).
static void redirectTo(BasicBlock *Source, BasicBlock *Target, DebugLoc DL) {
  if (Instruction *Term = Source->getTerminatorOrNull()) {
    auto *Br = cast<UncondBrInst>(Term);
    BasicBlock *Succ = Br->getSuccessor();
    Succ->removePredecessor(Source, /*KeepOneInputPHIs=*/true);
    Br->setSuccessor(Target);
    return;
  }
 
  auto *NewBr = UncondBrInst::Create(Target, Source);
  NewBr->setDebugLoc(DL);
}
 
void llvm::spliceBB(IRBuilderBase::InsertPoint IP, BasicBlock *New,
                    bool CreateBranch, DebugLoc DL) {
  assert(New->getFirstInsertionPt() == New->begin() &&
         "Target BB must not have PHI nodes");
 
  // Move instructions to new block.
  BasicBlock *Old = IP.getBlock();
  // If the `Old` block is empty then there are no instructions to move. But in
  // the new debug scheme, it could have trailing debug records which will be
  // moved to `New` in `spliceDebugInfoEmptyBlock`. We dont want that for 2
  // reasons:
  // 1. If `New` is also empty, `BasicBlock::splice` crashes.
  // 2. Even if `New` is not empty, the rationale to move those records to `New`
  // (in `spliceDebugInfoEmptyBlock`) does not apply here. That function
  // assumes that `Old` is optimized out and is going away. This is not the case
  // here. The `Old` block is still being used e.g. a branch instruction is
  // added to it later in this function.
  // So we call `BasicBlock::splice` only when `Old` is not empty.
  if (!Old->empty())
    New->splice(New->begin(), Old, IP.getPoint(), Old->end());
 
  if (CreateBranch) {
    auto *NewBr = UncondBrInst::Create(New, Old);
    NewBr->setDebugLoc(DL);
  }
}
 
void llvm::spliceBB(IRBuilder<> &Builder, BasicBlock *New, bool CreateBranch) {
  DebugLoc DebugLoc = Builder.getCurrentDebugLocation();
  BasicBlock *Old = Builder.GetInsertBlock();
 
  spliceBB(Builder.saveIP(), New, CreateBranch, DebugLoc);
  if (CreateBranch)
    Builder.SetInsertPoint(Old->getTerminator());
  else
    Builder.SetInsertPoint(Old);
 
  // SetInsertPoint also updates the Builder's debug location, but we want to
  // keep the one the Builder was configured to use.
  Builder.SetCurrentDebugLocation(DebugLoc);
}
 
BasicBlock *llvm::splitBB(IRBuilderBase::InsertPoint IP, bool CreateBranch,
                          DebugLoc DL, llvm::Twine Name) {
  BasicBlock *Old = IP.getBlock();
  BasicBlock *New = BasicBlock::Create(
      Old->getContext(), Name.isTriviallyEmpty() ? Old->getName() : Name,
      Old->getParent(), Old->getNextNode());
  spliceBB(IP, New, CreateBranch, DL);
  New->replaceSuccessorsPhiUsesWith(Old, New);
  return New;
}
 
BasicBlock *llvm::splitBB(IRBuilderBase &Builder, bool CreateBranch,
                          llvm::Twine Name) {
  DebugLoc DebugLoc = Builder.getCurrentDebugLocation();
  BasicBlock *New = splitBB(Builder.saveIP(), CreateBranch, DebugLoc, Name);
  if (CreateBranch)
    Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator());
  else
    Builder.SetInsertPoint(Builder.GetInsertBlock());
  // SetInsertPoint also updates the Builder's debug location, but we want to
  // keep the one the Builder was configured to use.
  Builder.SetCurrentDebugLocation(DebugLoc);
  return New;
}
 
BasicBlock *llvm::splitBB(IRBuilder<> &Builder, bool CreateBranch,
                          llvm::Twine Name) {
  DebugLoc DebugLoc = Builder.getCurrentDebugLocation();
  BasicBlock *New = splitBB(Builder.saveIP(), CreateBranch, DebugLoc, Name);
  if (CreateBranch)
    Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator());
  else
    Builder.SetInsertPoint(Builder.GetInsertBlock());
  // SetInsertPoint also updates the Builder's debug location, but we want to
  // keep the one the Builder was configured to use.
  Builder.SetCurrentDebugLocation(DebugLoc);
  return New;
}
 
BasicBlock *llvm::splitBBWithSuffix(IRBuilderBase &Builder, bool CreateBranch,
                                    llvm::Twine Suffix) {
  BasicBlock *Old = Builder.GetInsertBlock();
  return splitBB(Builder, CreateBranch, Old->getName() + Suffix);
}
 
// This function creates a fake integer value and a fake use for the integer
// value. It returns the fake value created. This is useful in modeling the
// extra arguments to the outlined functions.
Value *createFakeIntVal(IRBuilderBase &Builder,
                        OpenMPIRBuilder::InsertPointTy OuterAllocaIP,
                        llvm::SmallVectorImpl<Instruction *> &ToBeDeleted,
                        OpenMPIRBuilder::InsertPointTy InnerAllocaIP,
                        const Twine &Name = "", bool AsPtr = true,
                        bool Is64Bit = false) {
  Builder.restoreIP(OuterAllocaIP);
  IntegerType *IntTy = Is64Bit ? Builder.getInt64Ty() : Builder.getInt32Ty();
  Instruction *FakeVal;
  AllocaInst *FakeValAddr =
      Builder.CreateAlloca(IntTy, nullptr, Name + ".addr");
  ToBeDeleted.push_back(FakeValAddr);
 
  if (AsPtr) {
    FakeVal = FakeValAddr;
  } else {
    FakeVal = Builder.CreateLoad(IntTy, FakeValAddr, Name + ".val");
    ToBeDeleted.push_back(FakeVal);
  }
 
  // Generate a fake use of this value
  Builder.restoreIP(InnerAllocaIP);
  Instruction *UseFakeVal;
  if (AsPtr) {
    UseFakeVal = Builder.CreateLoad(IntTy, FakeVal, Name + ".use");
  } else {
    UseFakeVal = cast<BinaryOperator>(Builder.CreateAdd(
        FakeVal, Is64Bit ? Builder.getInt64(10) : Builder.getInt32(10)));
  }
  ToBeDeleted.push_back(UseFakeVal);
  return FakeVal;
}
 
//===----------------------------------------------------------------------===//
// OpenMPIRBuilderConfig
//===----------------------------------------------------------------------===//
 
namespace {
LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE();
/// Values for bit flags for marking which requires clauses have been used.
enum OpenMPOffloadingRequiresDirFlags {
  /// flag undefined.
  OMP_REQ_UNDEFINED = 0x000,
  /// no requires directive present.
  OMP_REQ_NONE = 0x001,
  /// reverse_offload clause.
  OMP_REQ_REVERSE_OFFLOAD = 0x002,
  /// unified_address clause.
  OMP_REQ_UNIFIED_ADDRESS = 0x004,
  /// unified_shared_memory clause.
  OMP_REQ_UNIFIED_SHARED_MEMORY = 0x008,
  /// dynamic_allocators clause.
  OMP_REQ_DYNAMIC_ALLOCATORS = 0x010,
  LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_REQ_DYNAMIC_ALLOCATORS)
};
 
class OMPCodeExtractor : public CodeExtractor {
public:
  OMPCodeExtractor(OpenMPIRBuilder &OMPBuilder, ArrayRef<BasicBlock *> BBs,
                   DominatorTree *DT = nullptr, bool AggregateArgs = false,
                   BlockFrequencyInfo *BFI = nullptr,
                   BranchProbabilityInfo *BPI = nullptr,
                   AssumptionCache *AC = nullptr, bool AllowVarArgs = false,
                   bool AllowAlloca = false,
                   BasicBlock *AllocationBlock = nullptr,
                   ArrayRef<BasicBlock *> DeallocationBlocks = {},
                   std::string Suffix = "", bool ArgsInZeroAddressSpace = false)
      : CodeExtractor(BBs, DT, AggregateArgs, BFI, BPI, AC, AllowVarArgs,
                      AllowAlloca, AllocationBlock, DeallocationBlocks, Suffix,
                      ArgsInZeroAddressSpace),
        OMPBuilder(OMPBuilder) {}
 
  virtual ~OMPCodeExtractor() = default;
 
protected:
  OpenMPIRBuilder &OMPBuilder;
};
 
class DeviceSharedMemCodeExtractor : public OMPCodeExtractor {
public:
  using OMPCodeExtractor::OMPCodeExtractor;
  virtual ~DeviceSharedMemCodeExtractor() = default;
 
protected:
  virtual Instruction *
  allocateVar(IRBuilder<>::InsertPoint AllocaIP, Type *VarType,
              const Twine &Name = Twine(""),
              AddrSpaceCastInst **CastedAlloc = nullptr) override {
    return OMPBuilder.createOMPAllocShared(AllocaIP, VarType, Name);
  }
 
  virtual Instruction *deallocateVar(IRBuilder<>::InsertPoint DeallocIP,
                                     Value *Var, Type *VarType) override {
    return OMPBuilder.createOMPFreeShared(DeallocIP, Var, VarType);
  }
};
 
/// Helper storing information about regions to outline using device shared
/// memory for intermediate allocations.
struct DeviceSharedMemOutlineInfo : public OpenMPIRBuilder::OutlineInfo {
  OpenMPIRBuilder &OMPBuilder;
 
  DeviceSharedMemOutlineInfo(OpenMPIRBuilder &OMPBuilder)
      : OMPBuilder(OMPBuilder) {}
  virtual ~DeviceSharedMemOutlineInfo() = default;
 
  virtual std::unique_ptr<CodeExtractor>
  createCodeExtractor(ArrayRef<BasicBlock *> Blocks,
                      bool ArgsInZeroAddressSpace,
                      Twine Suffix = Twine("")) override;
};
 
} // anonymous namespace
 
OpenMPIRBuilderConfig::OpenMPIRBuilderConfig()
    : RequiresFlags(OMP_REQ_UNDEFINED) {}
 
OpenMPIRBuilderConfig::OpenMPIRBuilderConfig(
    bool IsTargetDevice, bool IsGPU, bool OpenMPOffloadMandatory,
    bool HasRequiresReverseOffload, bool HasRequiresUnifiedAddress,
    bool HasRequiresUnifiedSharedMemory, bool HasRequiresDynamicAllocators)
    : IsTargetDevice(IsTargetDevice), IsGPU(IsGPU),
      OpenMPOffloadMandatory(OpenMPOffloadMandatory),
      RequiresFlags(OMP_REQ_UNDEFINED) {
  if (HasRequiresReverseOffload)
    RequiresFlags |= OMP_REQ_REVERSE_OFFLOAD;
  if (HasRequiresUnifiedAddress)
    RequiresFlags |= OMP_REQ_UNIFIED_ADDRESS;
  if (HasRequiresUnifiedSharedMemory)
    RequiresFlags |= OMP_REQ_UNIFIED_SHARED_MEMORY;
  if (HasRequiresDynamicAllocators)
    RequiresFlags |= OMP_REQ_DYNAMIC_ALLOCATORS;
}
 
bool OpenMPIRBuilderConfig::hasRequiresReverseOffload() const {
  return RequiresFlags & OMP_REQ_REVERSE_OFFLOAD;
}
 
bool OpenMPIRBuilderConfig::hasRequiresUnifiedAddress() const {
  return RequiresFlags & OMP_REQ_UNIFIED_ADDRESS;
}
 
bool OpenMPIRBuilderConfig::hasRequiresUnifiedSharedMemory() const {
  return RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY;
}
 
bool OpenMPIRBuilderConfig::hasRequiresDynamicAllocators() const {
  return RequiresFlags & OMP_REQ_DYNAMIC_ALLOCATORS;
}
 
int64_t OpenMPIRBuilderConfig::getRequiresFlags() const {
  return hasRequiresFlags() ? RequiresFlags
                            : static_cast<int64_t>(OMP_REQ_NONE);
}
 
void OpenMPIRBuilderConfig::setHasRequiresReverseOffload(bool Value) {
  if (Value)
    RequiresFlags |= OMP_REQ_REVERSE_OFFLOAD;
  else
    RequiresFlags &= ~OMP_REQ_REVERSE_OFFLOAD;
}
 
void OpenMPIRBuilderConfig::setHasRequiresUnifiedAddress(bool Value) {
  if (Value)
    RequiresFlags |= OMP_REQ_UNIFIED_ADDRESS;
  else
    RequiresFlags &= ~OMP_REQ_UNIFIED_ADDRESS;
}
 
void OpenMPIRBuilderConfig::setHasRequiresUnifiedSharedMemory(bool Value) {
  if (Value)
    RequiresFlags |= OMP_REQ_UNIFIED_SHARED_MEMORY;
  else
    RequiresFlags &= ~OMP_REQ_UNIFIED_SHARED_MEMORY;
}
 
void OpenMPIRBuilderConfig::setHasRequiresDynamicAllocators(bool Value) {
  if (Value)
    RequiresFlags |= OMP_REQ_DYNAMIC_ALLOCATORS;
  else
    RequiresFlags &= ~OMP_REQ_DYNAMIC_ALLOCATORS;
}
 
//===----------------------------------------------------------------------===//
// OpenMPIRBuilder
//===----------------------------------------------------------------------===//
 
void OpenMPIRBuilder::getKernelArgsVector(TargetKernelArgs &KernelArgs,
                                          IRBuilderBase &Builder,
                                          SmallVector<Value *> &ArgsVector) {
  Value *Version = Builder.getInt32(OMP_KERNEL_ARG_VERSION);
  Value *PointerNum = Builder.getInt32(KernelArgs.NumTargetItems);
  auto Int32Ty = Type::getInt32Ty(Builder.getContext());
  constexpr size_t MaxDim = 3;
  Value *ZeroArray = Constant::getNullValue(ArrayType::get(Int32Ty, MaxDim));
 
  Value *HasNoWaitFlag = Builder.getInt64(KernelArgs.HasNoWait);
 
  Value *DynCGroupMemFallbackFlag =
      Builder.getInt64(static_cast<uint64_t>(KernelArgs.DynCGroupMemFallback));
  DynCGroupMemFallbackFlag = Builder.CreateShl(DynCGroupMemFallbackFlag, 2);
 
  Value *StrictBlocksFlag = Builder.getInt64(KernelArgs.StrictBlocks);
  Value *StrictThreadsFlag = Builder.getInt64(KernelArgs.StrictThreads);
 
  StrictBlocksFlag = Builder.CreateShl(StrictBlocksFlag, 6);
  StrictThreadsFlag = Builder.CreateShl(StrictThreadsFlag, 7);
 
  Value *Flags = Builder.CreateOr(HasNoWaitFlag, DynCGroupMemFallbackFlag);
  Flags = Builder.CreateOr(Flags, StrictBlocksFlag);
  Flags = Builder.CreateOr(Flags, StrictThreadsFlag);
 
  assert(!KernelArgs.NumTeams.empty() && !KernelArgs.NumThreads.empty());
 
  Value *NumTeams3D =
      Builder.CreateInsertValue(ZeroArray, KernelArgs.NumTeams[0], {0});
  Value *NumThreads3D =
      Builder.CreateInsertValue(ZeroArray, KernelArgs.NumThreads[0], {0});
  for (unsigned I :
       seq<unsigned>(1, std::min(KernelArgs.NumTeams.size(), MaxDim)))
    NumTeams3D =
        Builder.CreateInsertValue(NumTeams3D, KernelArgs.NumTeams[I], {I});
  for (unsigned I :
       seq<unsigned>(1, std::min(KernelArgs.NumThreads.size(), MaxDim)))
    NumThreads3D =
        Builder.CreateInsertValue(NumThreads3D, KernelArgs.NumThreads[I], {I});
 
  ArgsVector = {Version,
                PointerNum,
                KernelArgs.RTArgs.BasePointersArray,
                KernelArgs.RTArgs.PointersArray,
                KernelArgs.RTArgs.SizesArray,
                KernelArgs.RTArgs.MapTypesArray,
                KernelArgs.RTArgs.MapNamesArray,
                KernelArgs.RTArgs.MappersArray,
                KernelArgs.NumIterations,
                Flags,
                NumTeams3D,
                NumThreads3D,
                KernelArgs.DynCGroupMem};
}
 
void OpenMPIRBuilder::addAttributes(omp::RuntimeFunction FnID, Function &Fn) {
  LLVMContext &Ctx = Fn.getContext();
 
  // Get the function's current attributes.
  auto Attrs = Fn.getAttributes();
  auto FnAttrs = Attrs.getFnAttrs();
  auto RetAttrs = Attrs.getRetAttrs();
  SmallVector<AttributeSet, 4> ArgAttrs;
  for (size_t ArgNo = 0; ArgNo < Fn.arg_size(); ++ArgNo)
    ArgAttrs.emplace_back(Attrs.getParamAttrs(ArgNo));
 
  // Add AS to FnAS while taking special care with integer extensions.
  auto addAttrSet = [&](AttributeSet &FnAS, const AttributeSet &AS,
                        bool Param = true) -> void {
    bool HasSignExt = AS.hasAttribute(Attribute::SExt);
    bool HasZeroExt = AS.hasAttribute(Attribute::ZExt);
    if (HasSignExt || HasZeroExt) {
      assert(AS.getNumAttributes() == 1 &&
             "Currently not handling extension attr combined with others.");
      if (Param) {
        if (auto AK = TargetLibraryInfo::getExtAttrForI32Param(T, HasSignExt))
          FnAS = FnAS.addAttribute(Ctx, AK);
      } else if (auto AK =
                     TargetLibraryInfo::getExtAttrForI32Return(T, HasSignExt))
        FnAS = FnAS.addAttribute(Ctx, AK);
    } else {
      FnAS = FnAS.addAttributes(Ctx, AS);
    }
  };
 
#define OMP_ATTRS_SET(VarName, AttrSet) AttributeSet VarName = AttrSet;
#include "llvm/Frontend/OpenMP/OMPKinds.def"
 
  // Add attributes to the function declaration.
  switch (FnID) {
#define OMP_RTL_ATTRS(Enum, FnAttrSet, RetAttrSet, ArgAttrSets)                \
  case Enum:                                                                   \
    FnAttrs = FnAttrs.addAttributes(Ctx, FnAttrSet);                           \
    addAttrSet(RetAttrs, RetAttrSet, /*Param*/ false);                         \
    for (size_t ArgNo = 0; ArgNo < ArgAttrSets.size(); ++ArgNo)                \
      addAttrSet(ArgAttrs[ArgNo], ArgAttrSets[ArgNo]);                         \
    Fn.setAttributes(AttributeList::get(Ctx, FnAttrs, RetAttrs, ArgAttrs));    \
    break;
#include "llvm/Frontend/OpenMP/OMPKinds.def"
  default:
    // Attributes are optional.
    break;
  }
}
 
FunctionCallee
OpenMPIRBuilder::getOrCreateRuntimeFunction(Module &M, RuntimeFunction FnID) {
  FunctionType *FnTy = nullptr;
  Function *Fn = nullptr;
 
  // Try to find the declation in the module first.
  switch (FnID) {
#define OMP_RTL(Enum, Str, IsVarArg, ReturnType, ...)                          \
  case Enum:                                                                   \
    FnTy = FunctionType::get(ReturnType, ArrayRef<Type *>{__VA_ARGS__},        \
                             IsVarArg);                                        \
    Fn = M.getFunction(Str);                                                   \
    break;
#include "llvm/Frontend/OpenMP/OMPKinds.def"
  }
 
  if (!Fn) {
    // Create a new declaration if we need one.
    switch (FnID) {
#define OMP_RTL(Enum, Str, ...)                                                \
  case Enum:                                                                   \
    Fn = Function::Create(FnTy, GlobalValue::ExternalLinkage, Str, M);         \
    break;
#include "llvm/Frontend/OpenMP/OMPKinds.def"
    }
    Fn->setCallingConv(Config.getRuntimeCC());
    // Add information if the runtime function takes a callback function
    if (FnID == OMPRTL___kmpc_fork_call || FnID == OMPRTL___kmpc_fork_teams) {
      if (!Fn->hasMetadata(LLVMContext::MD_callback)) {
        LLVMContext &Ctx = Fn->getContext();
        MDBuilder MDB(Ctx);
        // Annotate the callback behavior of the runtime function:
        //  - The callback callee is argument number 2 (microtask).
        //  - The first two arguments of the callback callee are unknown (-1).
        //  - All variadic arguments to the runtime function are passed to the
        //    callback callee.
        Fn->addMetadata(
            LLVMContext::MD_callback,
            *MDNode::get(Ctx, {MDB.createCallbackEncoding(
                                  2, {-1, -1}, /* VarArgsArePassed */ true)}));
      }
    }
 
    LLVM_DEBUG(dbgs() << "Created OpenMP runtime function " << Fn->getName()
                      << " with type " << *Fn->getFunctionType() << "\n");
    addAttributes(FnID, *Fn);
 
  } else {
    LLVM_DEBUG(dbgs() << "Found OpenMP runtime function " << Fn->getName()
                      << " with type " << *Fn->getFunctionType() << "\n");
  }
 
  assert(Fn && "Failed to create OpenMP runtime function");
 
  return {FnTy, Fn};
}
 
Expected<BasicBlock *>
OpenMPIRBuilder::FinalizationInfo::getFiniBB(IRBuilderBase &Builder) {
  if (!FiniBB) {
    Function *ParentFunc = Builder.GetInsertBlock()->getParent();
    IRBuilderBase::InsertPointGuard Guard(Builder);
    FiniBB = BasicBlock::Create(Builder.getContext(), ".fini", ParentFunc);
    Builder.SetInsertPoint(FiniBB);
    // FiniCB adds the branch to the exit stub.
    if (Error Err = FiniCB(Builder.saveIP()))
      return Err;
  }
  return FiniBB;
}
 
Error OpenMPIRBuilder::FinalizationInfo::mergeFiniBB(IRBuilderBase &Builder,
                                                     BasicBlock *OtherFiniBB) {
  // Simple case: FiniBB does not exist yet: re-use OtherFiniBB.
  if (!FiniBB) {
    FiniBB = OtherFiniBB;
 
    Builder.SetInsertPoint(FiniBB->getFirstNonPHIIt());
    if (Error Err = FiniCB(Builder.saveIP()))
      return Err;
 
    return Error::success();
  }
 
  // Move instructions from FiniBB to the start of OtherFiniBB.
  auto EndIt = FiniBB->end();
  if (FiniBB->size() >= 1)
    if (auto Prev = std::prev(EndIt); Prev->isTerminator())
      EndIt = Prev;
  OtherFiniBB->splice(OtherFiniBB->getFirstNonPHIIt(), FiniBB, FiniBB->begin(),
                      EndIt);
 
  FiniBB->replaceAllUsesWith(OtherFiniBB);
  FiniBB->eraseFromParent();
  FiniBB = OtherFiniBB;
  return Error::success();
}
 
Function *OpenMPIRBuilder::getOrCreateRuntimeFunctionPtr(RuntimeFunction FnID) {
  FunctionCallee RTLFn = getOrCreateRuntimeFunction(M, FnID);
  auto *Fn = dyn_cast<llvm::Function>(RTLFn.getCallee());
  assert(Fn && "Failed to create OpenMP runtime function pointer");
  return Fn;
}
 
CallInst *OpenMPIRBuilder::createRuntimeFunctionCall(FunctionCallee Callee,
                                                     ArrayRef<Value *> Args,
                                                     StringRef Name) {
  CallInst *Call = Builder.CreateCall(Callee, Args, Name);
  Call->setCallingConv(Config.getRuntimeCC());
  return Call;
}
 
void OpenMPIRBuilder::initialize() { initializeTypes(M); }
 
static void raiseUserConstantDataAllocasToEntryBlock(IRBuilderBase &Builder,
                                                     Function *Function) {
  BasicBlock &EntryBlock = Function->getEntryBlock();
  BasicBlock::iterator MoveLocInst = EntryBlock.getFirstNonPHIIt();
 
  // Loop over blocks looking for constant allocas, skipping the entry block
  // as any allocas there are already in the desired location.
  for (auto Block = std::next(Function->begin(), 1); Block != Function->end();
       Block++) {
    for (auto Inst = Block->getReverseIterator()->begin();
         Inst != Block->getReverseIterator()->end();) {
      if (auto *AllocaInst = dyn_cast_if_present<llvm::AllocaInst>(Inst)) {
        Inst++;
        if (!isa<ConstantData>(AllocaInst->getArraySize()))
          continue;
        AllocaInst->moveBeforePreserving(MoveLocInst);
      } else {
        Inst++;
      }
    }
  }
}
 
static void hoistNonEntryAllocasToEntryBlock(llvm::BasicBlock &Block) {
  llvm::SmallVector<llvm::Instruction *> AllocasToMove;
 
  auto ShouldHoistAlloca = [](const llvm::AllocaInst &AllocaInst) {
    // TODO: For now, we support simple static allocations, we might need to
    // move non-static ones as well. However, this will need further analysis to
    // move the lenght arguments as well.
    return !AllocaInst.isArrayAllocation();
  };
 
  for (llvm::Instruction &Inst : Block)
    if (auto *AllocaInst = llvm::dyn_cast<llvm::AllocaInst>(&Inst))
      if (ShouldHoistAlloca(*AllocaInst))
        AllocasToMove.push_back(AllocaInst);
 
  auto InsertPoint =
      Block.getParent()->getEntryBlock().getTerminator()->getIterator();
 
  for (llvm::Instruction *AllocaInst : AllocasToMove)
    AllocaInst->moveBefore(InsertPoint);
}
 
static void hoistNonEntryAllocasToEntryBlock(llvm::Function *Func) {
  PostDominatorTree PostDomTree(*Func);
  for (llvm::BasicBlock &BB : *Func)
    if (PostDomTree.properlyDominates(&BB, &Func->getEntryBlock()))
      hoistNonEntryAllocasToEntryBlock(BB);
}
 
void OpenMPIRBuilder::finalize(Function *Fn) {
  SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet;
  SmallVector<BasicBlock *, 32> Blocks;
  SmallVector<std::unique_ptr<OutlineInfo>, 16> DeferredOutlines;
  for (std::unique_ptr<OutlineInfo> &OI : OutlineInfos) {
    // Skip functions that have not finalized yet; may happen with nested
    // function generation.
    if (Fn && OI->getFunction() != Fn) {
      DeferredOutlines.push_back(std::move(OI));
      continue;
    }
 
    ParallelRegionBlockSet.clear();
    Blocks.clear();
    OI->collectBlocks(ParallelRegionBlockSet, Blocks);
 
    Function *OuterFn = OI->getFunction();
    CodeExtractorAnalysisCache CEAC(*OuterFn);
    // If we generate code for the target device, we need to allocate
    // struct for aggregate params in the device default alloca address space.
    // OpenMP runtime requires that the params of the extracted functions are
    // passed as zero address space pointers. This flag ensures that
    // CodeExtractor generates correct code for extracted functions
    // which are used by OpenMP runtime.
    bool ArgsInZeroAddressSpace = Config.isTargetDevice();
    std::unique_ptr<CodeExtractor> Extractor =
        OI->createCodeExtractor(Blocks, ArgsInZeroAddressSpace, ".omp_par");
 
    LLVM_DEBUG(dbgs() << "Before     outlining: " << *OuterFn << "\n");
    LLVM_DEBUG(dbgs() << "Entry " << OI->EntryBB->getName()
                      << " Exit: " << OI->ExitBB->getName() << "\n");
    assert(Extractor->isEligible() &&
           "Expected OpenMP outlining to be possible!");
 
    for (auto *V : OI->ExcludeArgsFromAggregate)
      Extractor->excludeArgFromAggregate(V);
 
    Function *OutlinedFn =
        Extractor->extractCodeRegion(CEAC, OI->Inputs, OI->Outputs);
 
    // Forward target-cpu, target-features attributes to the outlined function.
    auto TargetCpuAttr = OuterFn->getFnAttribute("target-cpu");
    if (TargetCpuAttr.isStringAttribute())
      OutlinedFn->addFnAttr(TargetCpuAttr);
 
    auto TargetFeaturesAttr = OuterFn->getFnAttribute("target-features");
    if (TargetFeaturesAttr.isStringAttribute())
      OutlinedFn->addFnAttr(TargetFeaturesAttr);
 
    LLVM_DEBUG(dbgs() << "After      outlining: " << *OuterFn << "\n");
    LLVM_DEBUG(dbgs() << "   Outlined function: " << *OutlinedFn << "\n");
    assert(OutlinedFn->getReturnType()->isVoidTy() &&
           "OpenMP outlined functions should not return a value!");
 
    // For compability with the clang CG we move the outlined function after the
    // one with the parallel region.
    OutlinedFn->removeFromParent();
    M.getFunctionList().insertAfter(OuterFn->getIterator(), OutlinedFn);
 
    // Remove the artificial entry introduced by the extractor right away, we
    // made our own entry block after all.
    {
      BasicBlock &ArtificialEntry = OutlinedFn->getEntryBlock();
      assert(ArtificialEntry.getUniqueSuccessor() == OI->EntryBB);
      assert(OI->EntryBB->getUniquePredecessor() == &ArtificialEntry);
      // Move instructions from the to-be-deleted ArtificialEntry to the entry
      // basic block of the parallel region. CodeExtractor generates
      // instructions to unwrap the aggregate argument and may sink
      // allocas/bitcasts for values that are solely used in the outlined region
      // and do not escape.
      assert(!ArtificialEntry.empty() &&
             "Expected instructions to add in the outlined region entry");
      for (BasicBlock::reverse_iterator It = ArtificialEntry.rbegin(),
                                        End = ArtificialEntry.rend();
           It != End;) {
        Instruction &I = *It;
        It++;
 
        if (I.isTerminator()) {
          // Absorb any debug value that terminator may have
          if (Instruction *TI = OI->EntryBB->getTerminatorOrNull())
            TI->adoptDbgRecords(&ArtificialEntry, I.getIterator(), false);
          continue;
        }
 
        I.moveBeforePreserving(*OI->EntryBB,
                               OI->EntryBB->getFirstInsertionPt());
      }
 
      OI->EntryBB->moveBefore(&ArtificialEntry);
      ArtificialEntry.eraseFromParent();
    }
    assert(&OutlinedFn->getEntryBlock() == OI->EntryBB);
    assert(OutlinedFn && OutlinedFn->hasNUses(1));
 
    // Run a user callback, e.g. to add attributes.
    if (OI->PostOutlineCB)
      OI->PostOutlineCB(*OutlinedFn);
 
    if (OI->FixUpNonEntryAllocas)
      hoistNonEntryAllocasToEntryBlock(OutlinedFn);
  }
 
  // Remove work items that have been completed.
  OutlineInfos = std::move(DeferredOutlines);
 
  // The createTarget functions embeds user written code into
  // the target region which may inject allocas which need to
  // be moved to the entry block of our target or risk malformed
  // optimisations by later passes, this is only relevant for
  // the device pass which appears to be a little more delicate
  // when it comes to optimisations (however, we do not block on
  // that here, it's up to the inserter to the list to do so).
  // This notbaly has to occur after the OutlinedInfo candidates
  // have been extracted so we have an end product that will not
  // be implicitly adversely affected by any raises unless
  // intentionally appended to the list.
  // NOTE: This only does so for ConstantData, it could be extended
  // to ConstantExpr's with further effort, however, they should
  // largely be folded when they get here. Extending it to runtime
  // defined/read+writeable allocation sizes would be non-trivial
  // (need to factor in movement of any stores to variables the
  // allocation size depends on, as well as the usual loads,
  // otherwise it'll yield the wrong result after movement) and
  // likely be more suitable as an LLVM optimisation pass.
  for (Function *F : ConstantAllocaRaiseCandidates)
    raiseUserConstantDataAllocasToEntryBlock(Builder, F);
 
  EmitMetadataErrorReportFunctionTy &&ErrorReportFn =
      [](EmitMetadataErrorKind Kind,
         const TargetRegionEntryInfo &EntryInfo) -> void {
    errs() << "Error of kind: " << Kind
           << " when emitting offload entries and metadata during "
              "OMPIRBuilder finalization \n";
  };
 
  if (!OffloadInfoManager.empty())
    createOffloadEntriesAndInfoMetadata(ErrorReportFn);
 
  // Rewrite uses of globals to their replacement declare target globals if
  // we are processing a device module.
  if (Config.isTargetDevice())
    applyDeclareTargetGlobalReplacements();
 
  if (Config.EmitLLVMUsedMetaInfo.value_or(false)) {
    std::vector<WeakTrackingVH> LLVMCompilerUsed = {
        M.getGlobalVariable("__openmp_nvptx_data_transfer_temporary_storage")};
    emitUsed("llvm.compiler.used", LLVMCompilerUsed);
  }
 
  IsFinalized = true;
}
 
bool OpenMPIRBuilder::isFinalized() { return IsFinalized; }
 
void OpenMPIRBuilder::registerDeclareTargetGlobalReplacement(
    GlobalValue *Original, GlobalValue *Replacement) {
  assert(Original && Replacement &&
         "Null values provided to registerDeclareTargetGlobalReplacement");
  DeclareTargetGlobalReplacements.push_back({Original, Replacement});
}
 
void OpenMPIRBuilder::applyDeclareTargetGlobalReplacements() {
  for (DeclareTargetGlobalReplacement &R : DeclareTargetGlobalReplacements) {
    GlobalValue *OldGV = R.Original;
    GlobalValue *NewGV = R.Replacement;
 
    assert(OldGV && NewGV &&
           "A null value was inserted into DeclareTargetGlobalReplacements");
 
    // The assert above should catch this case, but this is kept to attempt
    // to proceed without issue when asserts are off.
    if (!OldGV || !NewGV)
      continue;
 
    // The replacement global is a reference pointer that holds the
    // address of the device-resident storage. Every use must load the
    // reference pointer first and use the loaded address.
    //
    // Constant expression users (e.g. a constant GEP embedded in another
    // global's initializer or in an instruction) cannot have a load inserted
    // in place, so first expand any constant-expression users that live inside
    // functions into instructions. Any remaining constant users are handled
    // via a direct constant rewrite below as we cannot materialize a load
    // there.
    //
    // NOTE: We extend the constant rewrite to module scope, as we replace all
    // usages.
    if (auto *OldConst = dyn_cast<Constant>(OldGV))
      convertUsersOfConstantsToInstructions(OldConst,
                                            /*RestrictToFunc=*/nullptr,
                                            /*RemoveDeadConstants=*/false);
 
    IRBuilderBase::InsertPointGuard Guard(Builder);
    SmallVector<User *, 16> Users(OldGV->users());
    for (User *U : Users) {
      auto *Insn = dyn_cast<Instruction>(U);
      if (!Insn)
        continue;
 
      // A PHI node cannot have a load inserted immediately before it, as PHIs
      // must remain grouped at the top of their basic block. So we need to
      // make sure any loads we emit are generated in the preceding edge, a
      // PHI may reference the global on more than one edge, so every matching
      // slot must be handled.
      if (auto *PHI = dyn_cast<PHINode>(Insn)) {
        for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
          if (PHI->getIncomingValue(I) != OldGV)
            continue;
 
          BasicBlock *IncomingBB = PHI->getIncomingBlock(I);
          Builder.SetInsertPoint(IncomingBB->getTerminator());
          Builder.SetCurrentDebugLocation(PHI->getDebugLoc());
          LoadInst *EdgeLoad = Builder.CreateLoad(NewGV->getType(), NewGV);
          PHI->setIncomingValue(I, EdgeLoad);
        }
        continue;
      }
 
      Builder.SetInsertPoint(Insn);
      Builder.SetCurrentDebugLocation(Insn->getDebugLoc());
      LoadInst *Load = Builder.CreateLoad(NewGV->getType(), NewGV);
 
      // The replacement declare target global lives in the default address
      // space, whereas the original global may reside in a non-default
      // address space. In that case the initial lowering may have
      // emitted an addrspacecast that is no longer valid.  Replace the
      // whole addrspacecast with the load and erase it rather than
      // feeding the load back into the (now pointless) cast.
      // NOTE: If we end up with replacement declare target globals in
      // non-zero AS's the below will need some minor extensions to have the
      // option to alter the address space cast to the new address space where
      // required rather than just replacing it.
      if (auto *ASC = dyn_cast<AddrSpaceCastInst>(Insn)) {
        unsigned NewGVAS = NewGV->getType()->getPointerAddressSpace();
        assert(NewGVAS == 0 &&
               "Non-default address space declare target global");
        unsigned OldGVAS = OldGV->getType()->getPointerAddressSpace();
        unsigned DestAS = ASC->getType()->getPointerAddressSpace();
        if (DestAS == 0 && NewGVAS != OldGVAS) {
          ASC->replaceAllUsesWith(Load);
          ASC->eraseFromParent();
          continue;
        }
      }
 
      Insn->replaceUsesOfWith(OldGV, Load);
    }
  }
 
  DeclareTargetGlobalReplacements.clear();
}
 
OpenMPIRBuilder::~OpenMPIRBuilder() {
  assert(OutlineInfos.empty() && "There must be no outstanding outlinings");
}
 
GlobalValue *OpenMPIRBuilder::createGlobalFlag(unsigned Value, StringRef Name) {
  IntegerType *I32Ty = Type::getInt32Ty(M.getContext());
  auto *GV =
      new GlobalVariable(M, I32Ty,
                         /* isConstant = */ true, GlobalValue::WeakODRLinkage,
                         ConstantInt::get(I32Ty, Value), Name);
  GV->setVisibility(GlobalValue::HiddenVisibility);
 
  return GV;
}
 
void OpenMPIRBuilder::emitUsed(StringRef Name, ArrayRef<WeakTrackingVH> List) {
  if (List.empty())
    return;
 
  // Convert List to what ConstantArray needs.
  SmallVector<Constant *, 8> UsedArray;
  UsedArray.resize(List.size());
  for (unsigned I = 0, E = List.size(); I != E; ++I)
    UsedArray[I] = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
        cast<Constant>(&*List[I]), Builder.getPtrTy());
 
  if (UsedArray.empty())
    return;
  ArrayType *ATy = ArrayType::get(Builder.getPtrTy(), UsedArray.size());
 
  auto *GV = new GlobalVariable(M, ATy, false, GlobalValue::AppendingLinkage,
                                ConstantArray::get(ATy, UsedArray), Name);
 
  GV->setSection("llvm.metadata");
}
 
GlobalVariable *
OpenMPIRBuilder::emitKernelExecutionMode(StringRef KernelName,
                                         OMPTgtExecModeFlags Mode) {
  auto *Int8Ty = Builder.getInt8Ty();
  auto *GVMode = new GlobalVariable(
      M, Int8Ty, /*isConstant=*/true, GlobalValue::WeakAnyLinkage,
      ConstantInt::get(Int8Ty, Mode), Twine(KernelName, "_exec_mode"));
  GVMode->setVisibility(GlobalVariable::ProtectedVisibility);
  return GVMode;
}
 
Constant *OpenMPIRBuilder::getOrCreateIdent(Constant *SrcLocStr,
                                            uint32_t SrcLocStrSize,
                                            IdentFlag LocFlags,
                                            unsigned Reserve2Flags) {
  // Enable "C-mode".
  LocFlags |= OMP_IDENT_FLAG_KMPC;
 
  Constant *&Ident =
      IdentMap[{SrcLocStr, uint64_t(LocFlags) << 31 | Reserve2Flags}];
  if (!Ident) {
    Constant *I32Null = ConstantInt::getNullValue(Int32);
    Constant *IdentData[] = {I32Null,
                             ConstantInt::get(Int32, uint32_t(LocFlags)),
                             ConstantInt::get(Int32, Reserve2Flags),
                             ConstantInt::get(Int32, SrcLocStrSize), SrcLocStr};
 
    size_t SrcLocStrArgIdx = 4;
    if (OpenMPIRBuilder::Ident->getElementType(SrcLocStrArgIdx)
            ->getPointerAddressSpace() !=
        IdentData[SrcLocStrArgIdx]->getType()->getPointerAddressSpace())
      IdentData[SrcLocStrArgIdx] = ConstantExpr::getAddrSpaceCast(
          SrcLocStr, OpenMPIRBuilder::Ident->getElementType(SrcLocStrArgIdx));
    Constant *Initializer =
        ConstantStruct::get(OpenMPIRBuilder::Ident, IdentData);
 
    // Look for existing encoding of the location + flags, not needed but
    // minimizes the difference to the existing solution while we transition.
    for (GlobalVariable &GV : M.globals())
      if (GV.getValueType() == OpenMPIRBuilder::Ident && GV.hasInitializer())
        if (GV.getInitializer() == Initializer)
          Ident = &GV;
 
    if (!Ident) {
      auto *GV = new GlobalVariable(
          M, OpenMPIRBuilder::Ident,
          /* isConstant = */ true, GlobalValue::PrivateLinkage, Initializer, "",
          nullptr, GlobalValue::NotThreadLocal,
          M.getDataLayout().getDefaultGlobalsAddressSpace());
      GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
      GV->setAlignment(Align(8));
      Ident = GV;
    }
  }
 
  return ConstantExpr::getPointerBitCastOrAddrSpaceCast(Ident, IdentPtr);
}
 
Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef LocStr,
                                                uint32_t &SrcLocStrSize) {
  SrcLocStrSize = LocStr.size();
  Constant *&SrcLocStr = SrcLocStrMap[LocStr];
  if (!SrcLocStr) {
    Constant *Initializer =
        ConstantDataArray::getString(M.getContext(), LocStr);
 
    // Look for existing encoding of the location, not needed but minimizes the
    // difference to the existing solution while we transition.
    for (GlobalVariable &GV : M.globals())
      if (GV.isConstant() && GV.hasInitializer() &&
          GV.getInitializer() == Initializer)
        return SrcLocStr = ConstantExpr::getPointerCast(&GV, Int8Ptr);
 
    SrcLocStr = Builder.CreateGlobalString(
        LocStr, /*Name=*/"", M.getDataLayout().getDefaultGlobalsAddressSpace(),
        &M);
  }
  return SrcLocStr;
}
 
Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef FunctionName,
                                                StringRef FileName,
                                                unsigned Line, unsigned Column,
                                                uint32_t &SrcLocStrSize) {
  SmallString<128> Buffer;
  Buffer.push_back(';');
  Buffer.append(FileName);
  Buffer.push_back(';');
  Buffer.append(FunctionName);
  Buffer.push_back(';');
  Buffer.append(std::to_string(Line));
  Buffer.push_back(';');
  Buffer.append(std::to_string(Column));
  Buffer.push_back(';');
  Buffer.push_back(';');
  return getOrCreateSrcLocStr(Buffer.str(), SrcLocStrSize);
}
 
Constant *
OpenMPIRBuilder::getOrCreateDefaultSrcLocStr(uint32_t &SrcLocStrSize) {
  StringRef UnknownLoc = ";unknown;unknown;0;0;;";
  return getOrCreateSrcLocStr(UnknownLoc, SrcLocStrSize);
}
 
Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(DebugLoc DL,
                                                uint32_t &SrcLocStrSize,
                                                Function *F) {
  DILocation *DIL = DL.get();
  if (!DIL)
    return getOrCreateDefaultSrcLocStr(SrcLocStrSize);
  StringRef FileName =
      !DIL->getFilename().empty() ? DIL->getFilename() : M.getName();
  StringRef Function = DIL->getScope()->getSubprogram()->getName();
  if (Function.empty() && F)
    Function = F->getName();
  return getOrCreateSrcLocStr(Function, FileName, DIL->getLine(),
                              DIL->getColumn(), SrcLocStrSize);
}
 
Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(const LocationDescription &Loc,
                                                uint32_t &SrcLocStrSize) {
  return getOrCreateSrcLocStr(Loc.DL, SrcLocStrSize,
                              Loc.IP.getBlock()->getParent());
}
 
Value *OpenMPIRBuilder::getOrCreateThreadID(Value *Ident) {
  return createRuntimeFunctionCall(
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_global_thread_num), Ident,
      "omp_global_thread_num");
}
 
OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createTargetInReduction(
    const LocationDescription &Loc, ArrayRef<Value *> OrigPtrs,
    ArrayRef<Type *> ResultPtrTys,
    function_ref<void(unsigned, Value *)> MapPrivateCB) {
  assert(OrigPtrs.size() == ResultPtrTys.size() &&
         "expected one result pointer type per in_reduction item");
  if (!updateToLocation(Loc))
    return Loc.IP;
  if (OrigPtrs.empty())
    return Builder.saveIP();
 
  // Compute the executing thread's gtid once for the whole target body and
  // reuse it for every in_reduction lookup, so a target with several
  // in_reduction items does not emit a redundant __kmpc_global_thread_num per
  // item.
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *Gtid = getOrCreateThreadID(Ident);
 
  // The runtime entry point takes (and returns) a generic, default-address-
  // space `ptr`. A NULL descriptor makes the runtime walk the enclosing
  // taskgroups to find the matching task_reduction registration for the item.
  Type *PtrTy = PointerType::getUnqual(M.getContext());
  Value *NullDesc = ConstantPointerNull::get(PtrTy);
  FunctionCallee GetThData =
      getOrCreateRuntimeFunction(M, OMPRTL___kmpc_task_reduction_get_th_data);
 
  for (unsigned Idx = 0; Idx < OrigPtrs.size(); ++Idx) {
    // Normalize a non-default-address-space original pointer to the generic
    // address space before the call.
    Value *OrigPtr = OrigPtrs[Idx];
    if (auto *OrigPtrTy = dyn_cast<PointerType>(OrigPtr->getType());
        OrigPtrTy && OrigPtrTy->getAddressSpace() != 0)
      OrigPtr = Builder.CreateAddrSpaceCast(OrigPtr, PtrTy);
 
    Value *Priv = Builder.CreateCall(GetThData, {Gtid, NullDesc, OrigPtr},
                                     "omp.inred.priv");
 
    // Cast the returned private pointer back to the requested address space
    // when it differs.
    if (auto *ResPtrTy = dyn_cast<PointerType>(ResultPtrTys[Idx]);
        ResPtrTy && ResPtrTy->getAddressSpace() != 0)
      Priv = Builder.CreateAddrSpaceCast(Priv, ResultPtrTys[Idx]);
 
    MapPrivateCB(Idx, Priv);
  }
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::createBarrier(const LocationDescription &Loc, Directive Kind,
                               bool ForceSimpleCall, bool CheckCancelFlag) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  // Build call __kmpc_cancel_barrier(loc, thread_id) or
  //            __kmpc_barrier(loc, thread_id);
 
  IdentFlag BarrierLocFlags;
  switch (Kind) {
  case OMPD_for:
    BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_FOR;
    break;
  case OMPD_sections:
    BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SECTIONS;
    break;
  case OMPD_single:
    BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SINGLE;
    break;
  case OMPD_barrier:
    BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_EXPL;
    break;
  default:
    BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL;
    break;
  }
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Args[] = {
      getOrCreateIdent(SrcLocStr, SrcLocStrSize, BarrierLocFlags),
      getOrCreateThreadID(getOrCreateIdent(SrcLocStr, SrcLocStrSize))};
 
  // If we are in a cancellable parallel region, barriers are cancellation
  // points.
  // TODO: Check why we would force simple calls or to ignore the cancel flag.
  bool UseCancelBarrier =
      !ForceSimpleCall && isLastFinalizationInfoCancellable(OMPD_parallel);
 
  Value *Result = createRuntimeFunctionCall(
      getOrCreateRuntimeFunctionPtr(UseCancelBarrier
                                        ? OMPRTL___kmpc_cancel_barrier
                                        : OMPRTL___kmpc_barrier),
      Args);
 
  if (UseCancelBarrier && CheckCancelFlag)
    if (Error Err = emitCancelationCheckImpl(Result, OMPD_parallel))
      return Err;
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::createCancel(const LocationDescription &Loc,
                              Value *IfCondition,
                              omp::Directive CanceledDirective) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  // LLVM utilities like blocks with terminators.
  auto *UI = Builder.CreateUnreachable();
 
  Instruction *ThenTI = UI, *ElseTI = nullptr;
  if (IfCondition) {
    SplitBlockAndInsertIfThenElse(IfCondition, UI, &ThenTI, &ElseTI);
 
    // Even if the if condition evaluates to false, this should count as a
    // cancellation point
    Builder.SetInsertPoint(ElseTI);
    auto ElseIP = Builder.saveIP();
 
    InsertPointOrErrorTy IPOrErr = createCancellationPoint(
        LocationDescription{ElseIP, Loc.DL}, CanceledDirective);
    if (!IPOrErr)
      return IPOrErr;
  }
 
  Builder.SetInsertPoint(ThenTI);
 
  Value *CancelKind = nullptr;
  switch (CanceledDirective) {
#define OMP_CANCEL_KIND(Enum, Str, DirectiveEnum, Value)                       \
  case DirectiveEnum:                                                          \
    CancelKind = Builder.getInt32(Value);                                      \
    break;
#include "llvm/Frontend/OpenMP/OMPKinds.def"
  default:
    llvm_unreachable("Unknown cancel kind!");
  }
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *Args[] = {Ident, getOrCreateThreadID(Ident), CancelKind};
  Value *Result = createRuntimeFunctionCall(
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_cancel), Args);
 
  // The actual cancel logic is shared with others, e.g., cancel_barriers.
  if (Error Err = emitCancelationCheckImpl(Result, CanceledDirective))
    return Err;
 
  // Update the insertion point and remove the terminator we introduced.
  Builder.SetInsertPoint(UI->getParent());
  UI->eraseFromParent();
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::createCancellationPoint(const LocationDescription &Loc,
                                         omp::Directive CanceledDirective) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  // LLVM utilities like blocks with terminators.
  auto *UI = Builder.CreateUnreachable();
  Builder.SetInsertPoint(UI);
 
  Value *CancelKind = nullptr;
  switch (CanceledDirective) {
#define OMP_CANCEL_KIND(Enum, Str, DirectiveEnum, Value)                       \
  case DirectiveEnum:                                                          \
    CancelKind = Builder.getInt32(Value);                                      \
    break;
#include "llvm/Frontend/OpenMP/OMPKinds.def"
  default:
    llvm_unreachable("Unknown cancel kind!");
  }
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *Args[] = {Ident, getOrCreateThreadID(Ident), CancelKind};
  Value *Result = createRuntimeFunctionCall(
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_cancellationpoint), Args);
 
  // The actual cancel logic is shared with others, e.g., cancel_barriers.
  if (Error Err = emitCancelationCheckImpl(Result, CanceledDirective))
    return Err;
 
  // Update the insertion point and remove the terminator we introduced.
  Builder.SetInsertPoint(UI->getParent());
  UI->eraseFromParent();
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitTargetKernel(
    const LocationDescription &Loc, InsertPointTy AllocaIP, Value *&Return,
    Value *Ident, Value *DeviceID, Value *NumTeams, Value *NumThreads,
    Value *HostPtr, ArrayRef<Value *> KernelArgs) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  Builder.restoreIP(AllocaIP);
  auto *KernelArgsPtr =
      Builder.CreateAlloca(OpenMPIRBuilder::KernelArgs, nullptr, "kernel_args");
  updateToLocation(Loc);
 
  for (unsigned I = 0, Size = KernelArgs.size(); I != Size; ++I) {
    llvm::Value *Arg =
        Builder.CreateStructGEP(OpenMPIRBuilder::KernelArgs, KernelArgsPtr, I);
    Builder.CreateAlignedStore(
        KernelArgs[I], Arg,
        M.getDataLayout().getPrefTypeAlign(KernelArgs[I]->getType()));
  }
 
  SmallVector<Value *> OffloadingArgs{Ident,      DeviceID, NumTeams,
                                      NumThreads, HostPtr,  KernelArgsPtr};
 
  Return = createRuntimeFunctionCall(
      getOrCreateRuntimeFunction(M, OMPRTL___tgt_target_kernel),
      OffloadingArgs);
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::emitKernelLaunch(
    const LocationDescription &Loc, Value *OutlinedFnID,
    EmitFallbackCallbackTy EmitTargetCallFallbackCB, TargetKernelArgs &Args,
    Value *DeviceID, Value *RTLoc, InsertPointTy AllocaIP) {
 
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  // On top of the arrays that were filled up, the target offloading call
  // takes as arguments the device id as well as the host pointer. The host
  // pointer is used by the runtime library to identify the current target
  // region, so it only has to be unique and not necessarily point to
  // anything. It could be the pointer to the outlined function that
  // implements the target region, but we aren't using that so that the
  // compiler doesn't need to keep that, and could therefore inline the host
  // function if proven worthwhile during optimization.
 
  // From this point on, we need to have an ID of the target region defined.
  assert(OutlinedFnID && "Invalid outlined function ID!");
  (void)OutlinedFnID;
 
  // Return value of the runtime offloading call.
  Value *Return = nullptr;
 
  // Arguments for the target kernel.
  SmallVector<Value *> ArgsVector;
  getKernelArgsVector(Args, Builder, ArgsVector);
 
  // The target region is an outlined function launched by the runtime
  // via calls to __tgt_target_kernel().
  //
  // Note that on the host and CPU targets, the runtime implementation of
  // these calls simply call the outlined function without forking threads.
  // The outlined functions themselves have runtime calls to
  // __kmpc_fork_teams() and __kmpc_fork() for this purpose, codegen'd by
  // the compiler in emitTeamsCall() and emitParallelCall().
  //
  // In contrast, on the NVPTX target, the implementation of
  // __tgt_target_teams() launches a GPU kernel with the requested number
  // of teams and threads so no additional calls to the runtime are required.
  // Check the error code and execute the host version if required.
  Builder.restoreIP(emitTargetKernel(
      Builder, AllocaIP, Return, RTLoc, DeviceID, Args.NumTeams.front(),
      Args.NumThreads.front(), OutlinedFnID, ArgsVector));
 
  BasicBlock *OffloadFailedBlock =
      BasicBlock::Create(Builder.getContext(), "omp_offload.failed");
  BasicBlock *OffloadContBlock =
      BasicBlock::Create(Builder.getContext(), "omp_offload.cont");
  Value *Failed = Builder.CreateIsNotNull(Return);
  Builder.CreateCondBr(Failed, OffloadFailedBlock, OffloadContBlock);
 
  auto CurFn = Builder.GetInsertBlock()->getParent();
  emitBlock(OffloadFailedBlock, CurFn);
  InsertPointOrErrorTy AfterIP = EmitTargetCallFallbackCB(Builder.saveIP());
  if (!AfterIP)
    return AfterIP.takeError();
  Builder.restoreIP(*AfterIP);
  emitBranch(OffloadContBlock);
  emitBlock(OffloadContBlock, CurFn, /*IsFinished=*/true);
  return Builder.saveIP();
}
 
Error OpenMPIRBuilder::emitCancelationCheckImpl(
    Value *CancelFlag, omp::Directive CanceledDirective) {
  assert(isLastFinalizationInfoCancellable(CanceledDirective) &&
         "Unexpected cancellation!");
 
  // For a cancel barrier we create two new blocks.
  BasicBlock *BB = Builder.GetInsertBlock();
  BasicBlock *NonCancellationBlock;
  if (Builder.GetInsertPoint() == BB->end()) {
    // TODO: This branch will not be needed once we moved to the
    // OpenMPIRBuilder codegen completely.
    NonCancellationBlock = BasicBlock::Create(
        BB->getContext(), BB->getName() + ".cont", BB->getParent());
  } else {
    NonCancellationBlock = SplitBlock(BB, &*Builder.GetInsertPoint());
    BB->getTerminator()->eraseFromParent();
    Builder.SetInsertPoint(BB);
  }
  BasicBlock *CancellationBlock = BasicBlock::Create(
      BB->getContext(), BB->getName() + ".cncl", BB->getParent());
 
  // Jump to them based on the return value.
  Value *Cmp = Builder.CreateIsNull(CancelFlag);
  Builder.CreateCondBr(Cmp, NonCancellationBlock, CancellationBlock,
                       /* TODO weight */ nullptr, nullptr);
 
  // From the cancellation block we finalize all variables and go to the
  // post finalization block that is known to the FiniCB callback.
  auto &FI = FinalizationStack.back();
  Expected<BasicBlock *> FiniBBOrErr = FI.getFiniBB(Builder);
  if (!FiniBBOrErr)
    return FiniBBOrErr.takeError();
  Builder.SetInsertPoint(CancellationBlock);
  Builder.CreateBr(*FiniBBOrErr);
 
  // The continuation block is where code generation continues.
  Builder.SetInsertPoint(NonCancellationBlock, NonCancellationBlock->begin());
  return Error::success();
}
 
/// Create wrapper function used to gather the outlined function's argument
/// structure from a shared buffer and to forward them to it when running in
/// Generic mode.
///
/// The outlined function is expected to receive 2 integer arguments followed by
/// an optional pointer argument to an argument structure holding the rest.
static Function *createTargetParallelWrapper(OpenMPIRBuilder *OMPIRBuilder,
                                             Function &OutlinedFn) {
  size_t NumArgs = OutlinedFn.arg_size();
  assert((NumArgs == 2 || NumArgs == 3) &&
         "expected a 2-3 argument parallel outlined function");
  bool UseArgStruct = NumArgs == 3;
 
  IRBuilder<> &Builder = OMPIRBuilder->Builder;
  IRBuilder<>::InsertPointGuard IPG(Builder);
  auto *FnTy = FunctionType::get(Builder.getVoidTy(),
                                 {Builder.getInt16Ty(), Builder.getInt32Ty()},
                                 /*isVarArg=*/false);
  auto *WrapperFn =
      Function::Create(FnTy, GlobalValue::InternalLinkage,
                       OutlinedFn.getName() + ".wrapper", OMPIRBuilder->M);
 
  WrapperFn->addParamAttr(0, Attribute::NoUndef);
  WrapperFn->addParamAttr(0, Attribute::ZExt);
  WrapperFn->addParamAttr(1, Attribute::NoUndef);
 
  BasicBlock *EntryBB =
      BasicBlock::Create(OMPIRBuilder->M.getContext(), "entry", WrapperFn);
  Builder.SetInsertPoint(EntryBB);
 
  // Allocation.
  Value *AddrAlloca = Builder.CreateAlloca(Builder.getInt32Ty(),
                                           /*ArraySize=*/nullptr, "addr");
  AddrAlloca = Builder.CreatePointerBitCastOrAddrSpaceCast(
      AddrAlloca, Builder.getPtrTy(/*AddrSpace=*/0),
      AddrAlloca->getName() + ".ascast");
 
  Value *ZeroAlloca = Builder.CreateAlloca(Builder.getInt32Ty(),
                                           /*ArraySize=*/nullptr, "zero");
  ZeroAlloca = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ZeroAlloca, Builder.getPtrTy(/*AddrSpace=*/0),
      ZeroAlloca->getName() + ".ascast");
 
  Value *ArgsAlloca = nullptr;
  if (UseArgStruct) {
    ArgsAlloca = Builder.CreateAlloca(Builder.getPtrTy(),
                                      /*ArraySize=*/nullptr, "global_args");
    ArgsAlloca = Builder.CreatePointerBitCastOrAddrSpaceCast(
        ArgsAlloca, Builder.getPtrTy(/*AddrSpace=*/0),
        ArgsAlloca->getName() + ".ascast");
  }
 
  // Initialization.
  Builder.CreateStore(WrapperFn->getArg(1), AddrAlloca);
  Builder.CreateStore(Builder.getInt32(0), ZeroAlloca);
  if (UseArgStruct) {
    Builder.CreateCall(
        OMPIRBuilder->getOrCreateRuntimeFunctionPtr(
            llvm::omp::RuntimeFunction::OMPRTL___kmpc_get_shared_variables),
        {ArgsAlloca});
  }
 
  SmallVector<Value *, 3> Args{AddrAlloca, ZeroAlloca};
 
  // Load structArg from global_args.
  if (UseArgStruct) {
    Value *StructArg = Builder.CreateLoad(Builder.getPtrTy(), ArgsAlloca);
    StructArg = Builder.CreateInBoundsGEP(Builder.getPtrTy(), StructArg,
                                          {Builder.getInt64(0)});
    StructArg = Builder.CreateLoad(Builder.getPtrTy(), StructArg, "structArg");
    Args.push_back(StructArg);
  }
 
  // Call the outlined function holding the parallel body.
  Builder.CreateCall(&OutlinedFn, Args);
  Builder.CreateRetVoid();
 
  return WrapperFn;
}
 
// Callback used to create OpenMP runtime calls to support
// omp parallel clause for the device.
// We need to use this callback to replace call to the OutlinedFn in OuterFn
// by the call to the OpenMP DeviceRTL runtime function (kmpc_parallel_60)
static void targetParallelCallback(
    OpenMPIRBuilder *OMPIRBuilder, Function &OutlinedFn, Function *OuterFn,
    BasicBlock *OuterAllocaBB, Value *Ident, Value *IfCondition,
    Value *NumThreads, Instruction *PrivTID, AllocaInst *PrivTIDAddr,
    Value *ThreadID, const SmallVector<Instruction *, 4> &ToBeDeleted) {
  assert(OutlinedFn.arg_size() >= 2 &&
         "Expected at least tid and bounded tid as arguments");
  unsigned NumCapturedVars = OutlinedFn.arg_size() - /* tid & bounded tid */ 2;
 
  // Add some known attributes.
  IRBuilder<> &Builder = OMPIRBuilder->Builder;
  OutlinedFn.addParamAttr(0, Attribute::NoAlias);
  OutlinedFn.addParamAttr(1, Attribute::NoAlias);
  OutlinedFn.addParamAttr(0, Attribute::NoUndef);
  OutlinedFn.addParamAttr(1, Attribute::NoUndef);
  OutlinedFn.addFnAttr(Attribute::NoUnwind);
 
  CallInst *CI = cast<CallInst>(OutlinedFn.user_back());
  assert(CI && "Expected call instruction to outlined function");
  CI->getParent()->setName("omp_parallel");
 
  Builder.SetInsertPoint(CI);
  Type *PtrTy = OMPIRBuilder->VoidPtr;
 
  // Add alloca for kernel args
  OpenMPIRBuilder ::InsertPointTy CurrentIP = Builder.saveIP();
  Builder.SetInsertPoint(OuterAllocaBB, OuterAllocaBB->getFirstInsertionPt());
  AllocaInst *ArgsAlloca =
      Builder.CreateAlloca(ArrayType::get(PtrTy, NumCapturedVars));
  Value *Args = ArgsAlloca;
  // Add address space cast if array for storing arguments is not allocated
  // in address space 0
  if (ArgsAlloca->getAddressSpace())
    Args = Builder.CreatePointerCast(ArgsAlloca, PtrTy);
  Builder.restoreIP(CurrentIP);
 
  // Store captured vars which are used by kmpc_parallel_60
  for (unsigned Idx = 0; Idx < NumCapturedVars; Idx++) {
    Value *V = *(CI->arg_begin() + 2 + Idx);
    Value *StoreAddress = Builder.CreateConstInBoundsGEP2_64(
        ArrayType::get(PtrTy, NumCapturedVars), Args, 0, Idx);
    Builder.CreateStore(V, StoreAddress);
  }
 
  Value *Cond =
      IfCondition ? Builder.CreateSExtOrTrunc(IfCondition, OMPIRBuilder->Int32)
                  : Builder.getInt32(1);
  Value *NumThreadsArg =
      NumThreads ? Builder.CreateZExtOrTrunc(NumThreads, OMPIRBuilder->Int32)
                 : Builder.getInt32(-1);
 
  // If this is not a Generic kernel, we can skip generating the wrapper.
  Value *WrapperFn;
  if (isGenericKernel(*OuterFn))
    WrapperFn = createTargetParallelWrapper(OMPIRBuilder, OutlinedFn);
  else
    WrapperFn = Constant::getNullValue(PtrTy);
 
  // Build kmpc_parallel_60 call
  Value *Parallel60CallArgs[] = {
      /* identifier*/ Ident,
      /* global thread num*/ ThreadID,
      /* if expression */ Cond,
      /* number of threads */ NumThreadsArg,
      /* Proc bind */ Builder.getInt32(-1),
      /* outlined function */ &OutlinedFn,
      /* wrapper function */ WrapperFn,
      /* arguments of the outlined funciton*/ Args,
      /* number of arguments */ Builder.getInt64(NumCapturedVars),
      /* strict for number of threads */ Builder.getInt32(0)};
 
  FunctionCallee RTLFn =
      OMPIRBuilder->getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_parallel_60);
 
  OMPIRBuilder->createRuntimeFunctionCall(RTLFn, Parallel60CallArgs);
 
  LLVM_DEBUG(dbgs() << "With kmpc_parallel_60 placed: "
                    << *Builder.GetInsertBlock()->getParent() << "\n");
 
  // Initialize the local TID stack location with the argument value.
  Builder.SetInsertPoint(PrivTID);
  Function::arg_iterator OutlinedAI = OutlinedFn.arg_begin();
  Builder.CreateStore(Builder.CreateLoad(OMPIRBuilder->Int32, OutlinedAI),
                      PrivTIDAddr);
 
  // Remove redundant call to the outlined function.
  CI->eraseFromParent();
 
  for (Instruction *I : ToBeDeleted) {
    I->eraseFromParent();
  }
}
 
// Callback used to create OpenMP runtime calls to support
// omp parallel clause for the host.
// We need to use this callback to replace call to the OutlinedFn in OuterFn
// by the call to the OpenMP host runtime function ( __kmpc_fork_call[_if])
static void
hostParallelCallback(OpenMPIRBuilder *OMPIRBuilder, Function &OutlinedFn,
                     Function *OuterFn, Value *Ident, Value *IfCondition,
                     Instruction *PrivTID, AllocaInst *PrivTIDAddr,
                     const SmallVector<Instruction *, 4> &ToBeDeleted) {
  IRBuilder<> &Builder = OMPIRBuilder->Builder;
  FunctionCallee RTLFn;
  if (IfCondition) {
    RTLFn =
        OMPIRBuilder->getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_fork_call_if);
  } else {
    RTLFn =
        OMPIRBuilder->getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_fork_call);
  }
  if (auto *F = dyn_cast<Function>(RTLFn.getCallee())) {
    if (!F->hasMetadata(LLVMContext::MD_callback)) {
      LLVMContext &Ctx = F->getContext();
      MDBuilder MDB(Ctx);
      // Annotate the callback behavior of the __kmpc_fork_call:
      //  - The callback callee is argument number 2 (microtask).
      //  - The first two arguments of the callback callee are unknown (-1).
      //  - All variadic arguments to the __kmpc_fork_call are passed to the
      //    callback callee.
      F->addMetadata(LLVMContext::MD_callback,
                     *MDNode::get(Ctx, {MDB.createCallbackEncoding(
                                           2, {-1, -1},
                                           /* VarArgsArePassed */ true)}));
    }
  }
  // Add some known attributes.
  OutlinedFn.addParamAttr(0, Attribute::NoAlias);
  OutlinedFn.addParamAttr(1, Attribute::NoAlias);
  OutlinedFn.addFnAttr(Attribute::NoUnwind);
 
  assert(OutlinedFn.arg_size() >= 2 &&
         "Expected at least tid and bounded tid as arguments");
  unsigned NumCapturedVars = OutlinedFn.arg_size() - /* tid & bounded tid */ 2;
 
  CallInst *CI = cast<CallInst>(OutlinedFn.user_back());
  CI->getParent()->setName("omp_parallel");
  Builder.SetInsertPoint(CI);
 
  // Build call __kmpc_fork_call[_if](Ident, n, microtask, var1, .., varn);
  Value *ForkCallArgs[] = {Ident, Builder.getInt32(NumCapturedVars),
                           &OutlinedFn};
 
  SmallVector<Value *, 16> RealArgs;
  RealArgs.append(std::begin(ForkCallArgs), std::end(ForkCallArgs));
  if (IfCondition) {
    Value *Cond = Builder.CreateSExtOrTrunc(IfCondition, OMPIRBuilder->Int32);
    RealArgs.push_back(Cond);
  }
  RealArgs.append(CI->arg_begin() + /* tid & bound tid */ 2, CI->arg_end());
 
  // __kmpc_fork_call_if always expects a void ptr as the last argument
  // If there are no arguments, pass a null pointer.
  auto PtrTy = OMPIRBuilder->VoidPtr;
  if (IfCondition && NumCapturedVars == 0) {
    Value *NullPtrValue = Constant::getNullValue(PtrTy);
    RealArgs.push_back(NullPtrValue);
  }
 
  OMPIRBuilder->createRuntimeFunctionCall(RTLFn, RealArgs);
 
  LLVM_DEBUG(dbgs() << "With fork_call placed: "
                    << *Builder.GetInsertBlock()->getParent() << "\n");
 
  // Initialize the local TID stack location with the argument value.
  Builder.SetInsertPoint(PrivTID);
  Function::arg_iterator OutlinedAI = OutlinedFn.arg_begin();
  Builder.CreateStore(Builder.CreateLoad(OMPIRBuilder->Int32, OutlinedAI),
                      PrivTIDAddr);
 
  // Remove redundant call to the outlined function.
  CI->eraseFromParent();
 
  for (Instruction *I : ToBeDeleted) {
    I->eraseFromParent();
  }
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createParallel(
    const LocationDescription &Loc, InsertPointTy OuterAllocIP,
    ArrayRef<BasicBlock *> OuterDeallocBlocks, BodyGenCallbackTy BodyGenCB,
    PrivatizeCallbackTy PrivCB, FinalizeCallbackTy FiniCB, Value *IfCondition,
    Value *NumThreads, omp::ProcBindKind ProcBind, bool IsCancellable) {
  assert(!isConflictIP(Loc.IP, OuterAllocIP) && "IPs must not be ambiguous");
 
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  const bool NeedThreadID = NumThreads || Config.isTargetDevice() ||
                            (ProcBind != OMP_PROC_BIND_default);
  Value *ThreadID = NeedThreadID ? getOrCreateThreadID(Ident) : nullptr;
  // If we generate code for the target device, we need to allocate
  // struct for aggregate params in the device default alloca address space.
  // OpenMP runtime requires that the params of the extracted functions are
  // passed as zero address space pointers. This flag ensures that extracted
  // function arguments are declared in zero address space
  bool ArgsInZeroAddressSpace = Config.isTargetDevice();
 
  // Build call __kmpc_push_num_threads(&Ident, global_tid, num_threads)
  // only if we compile for host side.
  if (NumThreads && !Config.isTargetDevice()) {
    Value *Args[] = {
        Ident, ThreadID,
        Builder.CreateIntCast(NumThreads, Int32, /*isSigned*/ false)};
    createRuntimeFunctionCall(
        getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_push_num_threads), Args);
  }
 
  if (ProcBind != OMP_PROC_BIND_default) {
    // Build call __kmpc_push_proc_bind(&Ident, global_tid, proc_bind)
    Value *Args[] = {
        Ident, ThreadID,
        ConstantInt::get(Int32, unsigned(ProcBind), /*isSigned=*/true)};
    createRuntimeFunctionCall(
        getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_push_proc_bind), Args);
  }
 
  BasicBlock *InsertBB = Builder.GetInsertBlock();
  Function *OuterFn = InsertBB->getParent();
 
  // Save the outer alloca block because the insertion iterator may get
  // invalidated and we still need this later.
  BasicBlock *OuterAllocaBlock = OuterAllocIP.getBlock();
 
  // Vector to remember instructions we used only during the modeling but which
  // we want to delete at the end.
  SmallVector<Instruction *, 4> ToBeDeleted;
 
  // Change the location to the outer alloca insertion point to create and
  // initialize the allocas we pass into the parallel region.
  InsertPointTy NewOuter(OuterAllocaBlock, OuterAllocaBlock->begin());
  Builder.restoreIP(NewOuter);
  AllocaInst *TIDAddrAlloca = Builder.CreateAlloca(Int32, nullptr, "tid.addr");
  AllocaInst *ZeroAddrAlloca =
      Builder.CreateAlloca(Int32, nullptr, "zero.addr");
  Instruction *TIDAddr = TIDAddrAlloca;
  Instruction *ZeroAddr = ZeroAddrAlloca;
  if (ArgsInZeroAddressSpace && M.getDataLayout().getAllocaAddrSpace() != 0) {
    // Add additional casts to enforce pointers in zero address space
    TIDAddr = new AddrSpaceCastInst(
        TIDAddrAlloca, PointerType ::get(M.getContext(), 0), "tid.addr.ascast");
    TIDAddr->insertAfter(TIDAddrAlloca->getIterator());
    ToBeDeleted.push_back(TIDAddr);
    ZeroAddr = new AddrSpaceCastInst(ZeroAddrAlloca,
                                     PointerType ::get(M.getContext(), 0),
                                     "zero.addr.ascast");
    ZeroAddr->insertAfter(ZeroAddrAlloca->getIterator());
    ToBeDeleted.push_back(ZeroAddr);
  }
 
  // We only need TIDAddr and ZeroAddr for modeling purposes to get the
  // associated arguments in the outlined function, so we delete them later.
  ToBeDeleted.push_back(TIDAddrAlloca);
  ToBeDeleted.push_back(ZeroAddrAlloca);
 
  // Create an artificial insertion point that will also ensure the blocks we
  // are about to split are not degenerated.
  auto *UI = new UnreachableInst(Builder.getContext(), InsertBB);
 
  BasicBlock *EntryBB = UI->getParent();
  BasicBlock *PRegEntryBB = EntryBB->splitBasicBlock(UI, "omp.par.entry");
  BasicBlock *PRegBodyBB = PRegEntryBB->splitBasicBlock(UI, "omp.par.region");
  BasicBlock *PRegPreFiniBB =
      PRegBodyBB->splitBasicBlock(UI, "omp.par.pre_finalize");
  BasicBlock *PRegExitBB = PRegPreFiniBB->splitBasicBlock(UI, "omp.par.exit");
 
  auto FiniCBWrapper = [&](InsertPointTy IP) {
    // Hide "open-ended" blocks from the given FiniCB by setting the right jump
    // target to the region exit block.
    if (IP.getBlock()->end() == IP.getPoint()) {
      IRBuilder<>::InsertPointGuard IPG(Builder);
      Builder.restoreIP(IP);
      Instruction *I = Builder.CreateBr(PRegExitBB);
      IP = InsertPointTy(I->getParent(), I->getIterator());
    }
    assert(IP.getBlock()->getTerminator()->getNumSuccessors() == 1 &&
           IP.getBlock()->getTerminator()->getSuccessor(0) == PRegExitBB &&
           "Unexpected insertion point for finalization call!");
    return FiniCB(IP);
  };
 
  FinalizationStack.push_back({FiniCBWrapper, OMPD_parallel, IsCancellable});
 
  // Generate the privatization allocas in the block that will become the entry
  // of the outlined function.
  Builder.SetInsertPoint(PRegEntryBB->getTerminator());
  InsertPointTy InnerAllocaIP = Builder.saveIP();
 
  AllocaInst *PrivTIDAddr =
      Builder.CreateAlloca(Int32, nullptr, "tid.addr.local");
  Instruction *PrivTID = Builder.CreateLoad(Int32, PrivTIDAddr, "tid");
 
  // Add some fake uses for OpenMP provided arguments.
  ToBeDeleted.push_back(Builder.CreateLoad(Int32, TIDAddr, "tid.addr.use"));
  Instruction *ZeroAddrUse =
      Builder.CreateLoad(Int32, ZeroAddr, "zero.addr.use");
  ToBeDeleted.push_back(ZeroAddrUse);
 
  // EntryBB
  //   |
  //   V
  // PRegionEntryBB         <- Privatization allocas are placed here.
  //   |
  //   V
  // PRegionBodyBB          <- BodeGen is invoked here.
  //   |
  //   V
  // PRegPreFiniBB          <- The block we will start finalization from.
  //   |
  //   V
  // PRegionExitBB          <- A common exit to simplify block collection.
  //
 
  LLVM_DEBUG(dbgs() << "Before body codegen: " << *OuterFn << "\n");
 
  // Let the caller create the body.
  assert(BodyGenCB && "Expected body generation callback!");
  InsertPointTy CodeGenIP(PRegBodyBB, PRegBodyBB->begin());
  if (Error Err = BodyGenCB(InnerAllocaIP, CodeGenIP, PRegExitBB))
    return Err;
 
  LLVM_DEBUG(dbgs() << "After  body codegen: " << *OuterFn << "\n");
 
  // If OuterFn is a Generic kernel, we need to use device shared memory to
  // allocate argument structures. Otherwise, we use stack allocations as usual.
  bool UsesDeviceSharedMemory =
      Config.isTargetDevice() && isGenericKernel(*OuterFn);
  std::unique_ptr<OutlineInfo> OI =
      UsesDeviceSharedMemory
          ? std::make_unique<DeviceSharedMemOutlineInfo>(*this)
          : std::make_unique<OutlineInfo>();
 
  if (Config.isTargetDevice()) {
    // Generate OpenMP target specific runtime call
    OI->PostOutlineCB = [=, ToBeDeletedVec =
                                std::move(ToBeDeleted)](Function &OutlinedFn) {
      targetParallelCallback(this, OutlinedFn, OuterFn, OuterAllocaBlock, Ident,
                             IfCondition, NumThreads, PrivTID, PrivTIDAddr,
                             ThreadID, ToBeDeletedVec);
    };
  } else {
    // Generate OpenMP host runtime call
    OI->PostOutlineCB = [=, ToBeDeletedVec =
                                std::move(ToBeDeleted)](Function &OutlinedFn) {
      hostParallelCallback(this, OutlinedFn, OuterFn, Ident, IfCondition,
                           PrivTID, PrivTIDAddr, ToBeDeletedVec);
    };
  }
 
  OI->FixUpNonEntryAllocas = true;
  OI->OuterAllocBB = OuterAllocaBlock;
  OI->EntryBB = PRegEntryBB;
  OI->ExitBB = PRegExitBB;
  OI->OuterDeallocBBs.reserve(OuterDeallocBlocks.size());
  copy(OuterDeallocBlocks, OI->OuterDeallocBBs.end());
 
  SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet;
  SmallVector<BasicBlock *, 32> Blocks;
  OI->collectBlocks(ParallelRegionBlockSet, Blocks);
 
  CodeExtractorAnalysisCache CEAC(*OuterFn);
  CodeExtractor Extractor(Blocks, /* DominatorTree */ nullptr,
                          /* AggregateArgs */ false,
                          /* BlockFrequencyInfo */ nullptr,
                          /* BranchProbabilityInfo */ nullptr,
                          /* AssumptionCache */ nullptr,
                          /* AllowVarArgs */ true,
                          /* AllowAlloca */ true,
                          /* AllocationBlock */ OuterAllocaBlock,
                          /* DeallocationBlocks */ {},
                          /* Suffix */ ".omp_par", ArgsInZeroAddressSpace);
 
  // Find inputs to, outputs from the code region.
  BasicBlock *CommonExit = nullptr;
  SetVector<Value *> Inputs, Outputs, SinkingCands, HoistingCands;
  Extractor.findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit);
 
  Extractor.findInputsOutputs(Inputs, Outputs, SinkingCands,
                              /*CollectGlobalInputs=*/true);
 
  Inputs.remove_if([&](Value *I) {
    if (auto *GV = dyn_cast_if_present<GlobalVariable>(I))
      return GV->getValueType() == OpenMPIRBuilder::Ident;
 
    return false;
  });
 
  LLVM_DEBUG(dbgs() << "Before privatization: " << *OuterFn << "\n");
 
  FunctionCallee TIDRTLFn =
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_global_thread_num);
 
  auto PrivHelper = [&](Value &V) -> Error {
    if (&V == TIDAddr || &V == ZeroAddr) {
      OI->ExcludeArgsFromAggregate.push_back(&V);
      return Error::success();
    }
 
    SetVector<Use *> Uses;
    for (Use &U : V.uses())
      if (auto *UserI = dyn_cast<Instruction>(U.getUser()))
        if (ParallelRegionBlockSet.count(UserI->getParent()))
          Uses.insert(&U);
 
    // __kmpc_fork_call expects extra arguments as pointers. If the input
    // already has a pointer type, everything is fine. Otherwise, store the
    // value onto stack and load it back inside the to-be-outlined region. This
    // will ensure only the pointer will be passed to the function.
    // FIXME: if there are more than 15 trailing arguments, they must be
    // additionally packed in a struct.
    Value *Inner = &V;
    if (!V.getType()->isPointerTy()) {
      IRBuilder<>::InsertPointGuard Guard(Builder);
      LLVM_DEBUG(llvm::dbgs() << "Forwarding input as pointer: " << V << "\n");
 
      Builder.restoreIP(OuterAllocIP);
      Value *Ptr;
      if (UsesDeviceSharedMemory) {
        // Use device shared memory instead, if needed.
        Ptr = createOMPAllocShared(OuterAllocIP, V.getType(),
                                   V.getName() + ".reloaded");
        for (BasicBlock *DeallocBlock : OuterDeallocBlocks)
          createOMPFreeShared(
              InsertPointTy(DeallocBlock, DeallocBlock->getFirstInsertionPt()),
              Ptr, V.getType());
      } else {
        Ptr = Builder.CreateAlloca(V.getType(), nullptr,
                                   V.getName() + ".reloaded");
      }
 
      // Store to stack at end of the block that currently branches to the entry
      // block of the to-be-outlined region.
      Builder.SetInsertPoint(InsertBB,
                             InsertBB->getTerminator()->getIterator());
      Builder.CreateStore(&V, Ptr);
 
      // Load back next to allocations in the to-be-outlined region.
      Builder.restoreIP(InnerAllocaIP);
      Inner = Builder.CreateLoad(V.getType(), Ptr);
    }
 
    Value *ReplacementValue = nullptr;
    CallInst *CI = dyn_cast<CallInst>(&V);
    if (CI && CI->getCalledFunction() == TIDRTLFn.getCallee()) {
      ReplacementValue = PrivTID;
    } else {
      InsertPointOrErrorTy AfterIP =
          PrivCB(InnerAllocaIP, Builder.saveIP(), V, *Inner, ReplacementValue);
      if (!AfterIP)
        return AfterIP.takeError();
      Builder.restoreIP(*AfterIP);
      InnerAllocaIP = {
          InnerAllocaIP.getBlock(),
          InnerAllocaIP.getBlock()->getTerminator()->getIterator()};
 
      assert(ReplacementValue &&
             "Expected copy/create callback to set replacement value!");
      if (ReplacementValue == &V)
        return Error::success();
    }
 
    for (Use *UPtr : Uses)
      UPtr->set(ReplacementValue);
 
    return Error::success();
  };
 
  // Reset the inner alloca insertion as it will be used for loading the values
  // wrapped into pointers before passing them into the to-be-outlined region.
  // Configure it to insert immediately after the fake use of zero address so
  // that they are available in the generated body and so that the
  // OpenMP-related values (thread ID and zero address pointers) remain leading
  // in the argument list.
  InnerAllocaIP = IRBuilder<>::InsertPoint(
      ZeroAddrUse->getParent(), ZeroAddrUse->getNextNode()->getIterator());
 
  // Reset the outer alloca insertion point to the entry of the relevant block
  // in case it was invalidated.
  OuterAllocIP = IRBuilder<>::InsertPoint(
      OuterAllocaBlock, OuterAllocaBlock->getFirstInsertionPt());
 
  for (Value *Input : Inputs) {
    LLVM_DEBUG(dbgs() << "Captured input: " << *Input << "\n");
    if (Error Err = PrivHelper(*Input))
      return Err;
  }
  LLVM_DEBUG({
    for (Value *Output : Outputs)
      LLVM_DEBUG(dbgs() << "Captured output: " << *Output << "\n");
  });
  assert(Outputs.empty() &&
         "OpenMP outlining should not produce live-out values!");
 
  LLVM_DEBUG(dbgs() << "After  privatization: " << *OuterFn << "\n");
  LLVM_DEBUG({
    for (auto *BB : Blocks)
      dbgs() << " PBR: " << BB->getName() << "\n";
  });
 
  // Adjust the finalization stack, verify the adjustment, and call the
  // finalize function a last time to finalize values between the pre-fini
  // block and the exit block if we left the parallel "the normal way".
  auto FiniInfo = FinalizationStack.pop_back_val();
  (void)FiniInfo;
  assert(FiniInfo.DK == OMPD_parallel &&
         "Unexpected finalization stack state!");
 
  Instruction *PRegPreFiniTI = PRegPreFiniBB->getTerminator();
 
  InsertPointTy PreFiniIP(PRegPreFiniBB, PRegPreFiniTI->getIterator());
  Expected<BasicBlock *> FiniBBOrErr = FiniInfo.getFiniBB(Builder);
  if (!FiniBBOrErr)
    return FiniBBOrErr.takeError();
  {
    IRBuilderBase::InsertPointGuard Guard(Builder);
    Builder.restoreIP(PreFiniIP);
    Builder.CreateBr(*FiniBBOrErr);
    // There's currently a branch to omp.par.exit. Delete it. We will get there
    // via the fini block
    if (Instruction *Term = Builder.GetInsertBlock()->getTerminator())
      Term->eraseFromParent();
  }
 
  // Register the outlined info.
  addOutlineInfo(std::move(OI));
 
  InsertPointTy AfterIP(UI->getParent(), UI->getParent()->end());
  UI->eraseFromParent();
 
  return AfterIP;
}
 
void OpenMPIRBuilder::emitFlush(const LocationDescription &Loc) {
  // Build call void __kmpc_flush(ident_t *loc)
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Args[] = {getOrCreateIdent(SrcLocStr, SrcLocStrSize)};
 
  createRuntimeFunctionCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_flush),
                            Args);
}
 
void OpenMPIRBuilder::createFlush(const LocationDescription &Loc) {
  if (!updateToLocation(Loc))
    return;
  emitFlush(Loc);
}
 
void OpenMPIRBuilder::createError(const LocationDescription &Loc, bool IsFatal,
                                  Value *Message) {
  if (!updateToLocation(Loc))
    return;
 
  // Build call void __kmpc_error(ident_t *loc, int severity,
  //                              const char *message)
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  // Severity: 1 = warning, 2 = fatal.
  Value *Severity = ConstantInt::get(Int32, IsFatal ? 2 : 1);
  Value *MessageArg = Message ? Message : ConstantPointerNull::get(Int8Ptr);
  Value *Args[] = {Ident, Severity, MessageArg};
 
  createRuntimeFunctionCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_error),
                            Args);
}
 
void OpenMPIRBuilder::emitTaskyieldImpl(const LocationDescription &Loc) {
  // Build call __kmpc_omp_taskyield(loc, thread_id, 0);
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Constant *I32Null = ConstantInt::getNullValue(Int32);
  Value *Args[] = {Ident, getOrCreateThreadID(Ident), I32Null};
 
  createRuntimeFunctionCall(
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_taskyield), Args);
}
 
void OpenMPIRBuilder::createTaskyield(const LocationDescription &Loc) {
  if (!updateToLocation(Loc))
    return;
  emitTaskyieldImpl(Loc);
}
 
void OpenMPIRBuilder::emitTaskDependency(IRBuilderBase &Builder, Value *Entry,
                                         const DependData &Dep) {
  // Store the pointer to the variable
  Value *Addr = Builder.CreateStructGEP(
      DependInfo, Entry,
      static_cast<unsigned int>(RTLDependInfoFields::BaseAddr));
  Value *DepValPtr = Builder.CreatePtrToInt(Dep.DepVal, SizeTy);
  Builder.CreateStore(DepValPtr, Addr);
  // Store the size of the variable
  Value *Size = Builder.CreateStructGEP(
      DependInfo, Entry, static_cast<unsigned int>(RTLDependInfoFields::Len));
  Builder.CreateStore(
      ConstantInt::get(SizeTy,
                       M.getDataLayout().getTypeStoreSize(Dep.DepValueType)),
      Size);
  // Store the dependency kind
  Value *Flags = Builder.CreateStructGEP(
      DependInfo, Entry, static_cast<unsigned int>(RTLDependInfoFields::Flags));
  Builder.CreateStore(ConstantInt::get(Builder.getInt8Ty(),
                                       static_cast<unsigned int>(Dep.DepKind)),
                      Flags);
}
 
// Processes the dependencies in Dependencies and does the following
// - Allocates space on the stack of an array of DependInfo objects
// - Populates each DependInfo object with relevant information of
//   the corresponding dependence.
// - All code is inserted in the entry block of the current function.
static Value *emitTaskDependencies(
    OpenMPIRBuilder &OMPBuilder,
    const SmallVectorImpl<OpenMPIRBuilder::DependData> &Dependencies) {
  // Early return if we have no dependencies to process
  if (Dependencies.empty())
    return nullptr;
 
  // Given a vector of DependData objects, in this function we create an
  // array on the stack that holds kmp_depend_info objects corresponding
  // to each dependency. This is then passed to the OpenMP runtime.
  // For example, if there are 'n' dependencies then the following psedo
  // code is generated. Assume the first dependence is on a variable 'a'
  //
  // \code{c}
  // DepArray = alloc(n x sizeof(kmp_depend_info);
  // idx = 0;
  // DepArray[idx].base_addr = ptrtoint(&a);
  // DepArray[idx].len = 8;
  // DepArray[idx].flags = Dep.DepKind; /*(See OMPContants.h for DepKind)*/
  // ++idx;
  // DepArray[idx].base_addr = ...;
  // \endcode
 
  IRBuilderBase &Builder = OMPBuilder.Builder;
  Type *DependInfo = OMPBuilder.DependInfo;
 
  Value *DepArray = nullptr;
  OpenMPIRBuilder::InsertPointTy OldIP = Builder.saveIP();
  Builder.SetInsertPoint(
      OldIP.getBlock()->getParent()->getEntryBlock().getTerminator());
 
  Type *DepArrayTy = ArrayType::get(DependInfo, Dependencies.size());
  DepArray = Builder.CreateAlloca(DepArrayTy, nullptr, ".dep.arr.addr");
 
  Builder.restoreIP(OldIP);
 
  for (const auto &[DepIdx, Dep] : enumerate(Dependencies)) {
    Value *Base =
        Builder.CreateConstInBoundsGEP2_64(DepArrayTy, DepArray, 0, DepIdx);
    OMPBuilder.emitTaskDependency(Builder, Base, Dep);
  }
  return DepArray;
}
 
void OpenMPIRBuilder::emitTaskwaitImpl(const LocationDescription &Loc) {
  // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32
  // global_tid);
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *Args[] = {Ident, getOrCreateThreadID(Ident)};
 
  // Ignore return result until untied tasks are supported.
  createRuntimeFunctionCall(
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_taskwait), Args);
}
 
void OpenMPIRBuilder::createTaskwait(const LocationDescription &Loc,
                                     DependenciesInfo Dependencies) {
  if (!updateToLocation(Loc))
    return;
 
  Value *DepArray = nullptr;
  Type *DepArrayTy = nullptr;
  Value *NumDeps = nullptr;
  if (Dependencies.DepArray) {
    DepArray = Dependencies.DepArray;
    NumDeps = Dependencies.NumDeps;
  } else if (!Dependencies.Deps.empty()) {
    InsertPointTy OldIP = Builder.saveIP();
    BasicBlock &entryBB =
        Builder.GetInsertBlock()->getParent()->getEntryBlock();
    Builder.SetInsertPoint(&entryBB, entryBB.getFirstInsertionPt());
 
    DepArrayTy = ArrayType::get(DependInfo, Dependencies.Deps.size());
    DepArray = Builder.CreateAlloca(DepArrayTy, nullptr, ".dep.arr.addr");
    NumDeps = Builder.getInt32(Dependencies.Deps.size());
 
    Builder.restoreIP(OldIP);
    for (const auto &[DepIdx, Dep] : enumerate(Dependencies.Deps)) {
      Value *Base =
          Builder.CreateConstInBoundsGEP2_64(DepArrayTy, DepArray, 0, DepIdx);
      this->emitTaskDependency(Builder, Base, Dep);
    }
  }
 
  if (DepArray) {
    uint32_t SrcLocStrSize;
    Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
    Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
    Value *Args[] = {
        Ident,
        getOrCreateThreadID(Ident),
        NumDeps,
        DepArray,
        ConstantInt::get(Builder.getInt32Ty(), 0),
        ConstantPointerNull::get(PointerType::getUnqual(M.getContext())),
        ConstantInt::get(Builder.getInt32Ty(), false)};
    createRuntimeFunctionCall(
        getOrCreateRuntimeFunctionPtr(
            omp::RuntimeFunction::OMPRTL___kmpc_omp_taskwait_deps_51),
        Args);
  } else {
    emitTaskwaitImpl(Loc);
  }
}
 
/// Create the task duplication function passed to kmpc_taskloop.
Expected<Value *> OpenMPIRBuilder::createTaskDuplicationFunction(
    Type *PrivatesTy, int32_t PrivatesIndex, TaskDupCallbackTy DupCB) {
  unsigned ProgramAddressSpace = M.getDataLayout().getProgramAddressSpace();
  if (!DupCB)
    return Constant::getNullValue(
        PointerType::get(Builder.getContext(), ProgramAddressSpace));
 
  // From OpenMP Runtime p_task_dup_t:
  // Routine optionally generated by the compiler for setting the lastprivate
  // flag and calling needed constructors for private/firstprivate objects (used
  // to form taskloop tasks from pattern task) Parameters: dest task, src task,
  // lastprivate flag.
  // typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32);
 
  auto *VoidPtrTy = PointerType::get(Builder.getContext(), ProgramAddressSpace);
 
  FunctionType *DupFuncTy = FunctionType::get(
      Builder.getVoidTy(), {VoidPtrTy, VoidPtrTy, Builder.getInt32Ty()},
      /*isVarArg=*/false);
 
  Function *DupFunction = Function::Create(DupFuncTy, Function::InternalLinkage,
                                           "omp_taskloop_dup", M);
  Value *DestTaskArg = DupFunction->getArg(0);
  Value *SrcTaskArg = DupFunction->getArg(1);
  Value *LastprivateFlagArg = DupFunction->getArg(2);
  DestTaskArg->setName("dest_task");
  SrcTaskArg->setName("src_task");
  LastprivateFlagArg->setName("lastprivate_flag");
 
  IRBuilderBase::InsertPointGuard Guard(Builder);
  Builder.SetInsertPoint(
      BasicBlock::Create(Builder.getContext(), "entry", DupFunction));
 
  auto GetTaskContextPtrFromArg = [&](Value *Arg) -> Value * {
    Type *TaskWithPrivatesTy =
        StructType::get(Builder.getContext(), {Task, PrivatesTy});
    Value *TaskPrivates = Builder.CreateGEP(
        TaskWithPrivatesTy, Arg, {Builder.getInt32(0), Builder.getInt32(1)});
    Value *ContextPtr = Builder.CreateGEP(
        PrivatesTy, TaskPrivates,
        {Builder.getInt32(0), Builder.getInt32(PrivatesIndex)});
    return ContextPtr;
  };
 
  Value *DestTaskContextPtr = GetTaskContextPtrFromArg(DestTaskArg);
  Value *SrcTaskContextPtr = GetTaskContextPtrFromArg(SrcTaskArg);
 
  DestTaskContextPtr->setName("destPtr");
  SrcTaskContextPtr->setName("srcPtr");
 
  InsertPointTy AllocaIP(&DupFunction->getEntryBlock(),
                         DupFunction->getEntryBlock().begin());
  InsertPointTy CodeGenIP = Builder.saveIP();
  Expected<IRBuilderBase::InsertPoint> AfterIPOrError =
      DupCB(AllocaIP, CodeGenIP, DestTaskContextPtr, SrcTaskContextPtr);
  if (!AfterIPOrError)
    return AfterIPOrError.takeError();
  Builder.restoreIP(*AfterIPOrError);
 
  Builder.CreateRetVoid();
 
  return DupFunction;
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createTaskloop(
    const LocationDescription &Loc, InsertPointTy AllocaIP,
    ArrayRef<BasicBlock *> DeallocBlocks, BodyGenCallbackTy BodyGenCB,
    llvm::function_ref<llvm::Expected<llvm::CanonicalLoopInfo *>()> LoopInfo,
    Value *LBVal, Value *UBVal, Value *StepVal, bool Untied, Value *IfCond,
    Value *GrainSize, bool NoGroup, int Sched, Value *Final, bool Mergeable,
    Value *Priority, uint64_t NumOfCollapseLoops, TaskDupCallbackTy DupCB,
    Value *TaskContextStructPtrVal) {
 
  if (!updateToLocation(Loc))
    return InsertPointTy();
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
 
  BasicBlock *TaskloopExitBB =
      splitBB(Builder, /*CreateBranch=*/true, "taskloop.exit");
  BasicBlock *TaskloopBodyBB =
      splitBB(Builder, /*CreateBranch=*/true, "taskloop.body");
  BasicBlock *TaskloopAllocaBB =
      splitBB(Builder, /*CreateBranch=*/true, "taskloop.alloca");
 
  InsertPointTy TaskloopAllocaIP =
      InsertPointTy(TaskloopAllocaBB, TaskloopAllocaBB->begin());
  InsertPointTy TaskloopBodyIP =
      InsertPointTy(TaskloopBodyBB, TaskloopBodyBB->begin());
 
  if (Error Err = BodyGenCB(TaskloopAllocaIP, TaskloopBodyIP, TaskloopExitBB))
    return Err;
 
  llvm::Expected<llvm::CanonicalLoopInfo *> result = LoopInfo();
  if (!result) {
    return result.takeError();
  }
 
  llvm::CanonicalLoopInfo *CLI = result.get();
  auto OI = std::make_unique<OutlineInfo>();
  OI->EntryBB = TaskloopAllocaBB;
  OI->OuterAllocBB = AllocaIP.getBlock();
  OI->ExitBB = TaskloopExitBB;
  OI->OuterDeallocBBs.reserve(DeallocBlocks.size());
  copy(DeallocBlocks, OI->OuterDeallocBBs.end());
 
  // Add the thread ID argument.
  SmallVector<Instruction *> ToBeDeleted;
  // dummy instruction to be used as a fake argument
  OI->ExcludeArgsFromAggregate.push_back(createFakeIntVal(
      Builder, AllocaIP, ToBeDeleted, TaskloopAllocaIP, "global.tid", false));
  Value *FakeLB = createFakeIntVal(Builder, AllocaIP, ToBeDeleted,
                                   TaskloopAllocaIP, "lb", false, true);
  Value *FakeUB = createFakeIntVal(Builder, AllocaIP, ToBeDeleted,
                                   TaskloopAllocaIP, "ub", false, true);
  Value *FakeStep = createFakeIntVal(Builder, AllocaIP, ToBeDeleted,
                                     TaskloopAllocaIP, "step", false, true);
  // For Taskloop, we want to force the bounds being the first 3 inputs in the
  // aggregate struct
  OI->Inputs.insert(FakeLB);
  OI->Inputs.insert(FakeUB);
  OI->Inputs.insert(FakeStep);
  if (TaskContextStructPtrVal)
    OI->Inputs.insert(TaskContextStructPtrVal);
  assert(((TaskContextStructPtrVal && DupCB) ||
          (!TaskContextStructPtrVal && !DupCB)) &&
         "Task context struct ptr and duplication callback must be both set "
         "or both null");
 
  // It isn't safe to run the duplication bodygen callback inside the post
  // outlining callback so this has to be run now before we know the real task
  // shareds structure type.
  unsigned ProgramAddressSpace = M.getDataLayout().getProgramAddressSpace();
  Type *PointerTy = PointerType::get(Builder.getContext(), ProgramAddressSpace);
  Type *FakeSharedsTy = StructType::get(
      Builder.getContext(),
      {FakeLB->getType(), FakeUB->getType(), FakeStep->getType(), PointerTy});
  Expected<Value *> TaskDupFnOrErr = createTaskDuplicationFunction(
      FakeSharedsTy,
      /*PrivatesIndex: the pointer after the three indices above*/ 3, DupCB);
  if (!TaskDupFnOrErr) {
    return TaskDupFnOrErr.takeError();
  }
  Value *TaskDupFn = *TaskDupFnOrErr;
 
  OI->PostOutlineCB = [this, Ident, LBVal, UBVal, StepVal, Untied,
                       TaskloopAllocaBB, CLI, TaskDupFn, ToBeDeleted, IfCond,
                       GrainSize, NoGroup, Sched, FakeLB, FakeUB, FakeStep,
                       FakeSharedsTy, Final, Mergeable, Priority,
                       NumOfCollapseLoops](Function &OutlinedFn) mutable {
    // Replace the Stale CI by appropriate RTL function call.
    assert(OutlinedFn.hasOneUse() &&
           "there must be a single user for the outlined function");
    CallInst *StaleCI = cast<CallInst>(OutlinedFn.user_back());
 
    /* Create the casting for the Bounds Values that can be used when outlining
     * to replace the uses of the fakes with real values */
    BasicBlock *CodeReplBB = StaleCI->getParent();
    Builder.SetInsertPoint(CodeReplBB->getFirstInsertionPt());
    Value *CastedLBVal =
        Builder.CreateIntCast(LBVal, Builder.getInt64Ty(), true, "lb64");
    Value *CastedUBVal =
        Builder.CreateIntCast(UBVal, Builder.getInt64Ty(), true, "ub64");
    Value *CastedStepVal =
        Builder.CreateIntCast(StepVal, Builder.getInt64Ty(), true, "step64");
 
    Builder.SetInsertPoint(StaleCI);
 
    // Gather the arguments for emitting the runtime call for
    // @__kmpc_omp_task_alloc
    Function *TaskAllocFn =
        getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_alloc);
 
    Value *ThreadID = getOrCreateThreadID(Ident);
 
    if (!NoGroup) {
      // Emit runtime call for @__kmpc_taskgroup
      Function *TaskgroupFn =
          getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_taskgroup);
      Builder.CreateCall(TaskgroupFn, {Ident, ThreadID});
    }
 
    // `flags` Argument Configuration
    // Task is tied if (Flags & 1) == 1.
    // Task is untied if (Flags & 1) == 0.
    // Task is final if (Flags & 2) == 2.
    // Task is not final if (Flags & 2) == 0.
    // Task is mergeable if (Flags & 4) == 4.
    // Task is not mergeable if (Flags & 4) == 0.
    // Task is priority if (Flags & 32) == 32.
    // Task is not priority if (Flags & 32) == 0.
    Value *Flags = Builder.getInt32(Untied ? 0 : 1);
    if (Final)
      Flags = Builder.CreateOr(Builder.getInt32(2), Flags);
    if (Mergeable)
      Flags = Builder.CreateOr(Builder.getInt32(4), Flags);
    if (Priority)
      Flags = Builder.CreateOr(Builder.getInt32(32), Flags);
 
    Value *TaskSize = Builder.getInt64(
        divideCeil(M.getDataLayout().getTypeSizeInBits(Task), 8));
 
    AllocaInst *ArgStructAlloca =
        dyn_cast<AllocaInst>(StaleCI->getArgOperand(1));
    assert(ArgStructAlloca &&
           "Unable to find the alloca instruction corresponding to arguments "
           "for extracted function");
    std::optional<TypeSize> ArgAllocSize =
        ArgStructAlloca->getAllocationSize(M.getDataLayout());
    assert(ArgAllocSize &&
           "Unable to determine size of arguments for extracted function");
    Value *SharedsSize = Builder.getInt64(ArgAllocSize->getFixedValue());
 
    // Emit the @__kmpc_omp_task_alloc runtime call
    // The runtime call returns a pointer to an area where the task captured
    // variables must be copied before the task is run (TaskData)
    CallInst *TaskData = Builder.CreateCall(
        TaskAllocFn, {/*loc_ref=*/Ident, /*gtid=*/ThreadID, /*flags=*/Flags,
                      /*sizeof_task=*/TaskSize, /*sizeof_shared=*/SharedsSize,
                      /*task_func=*/&OutlinedFn});
 
    Value *Shareds = StaleCI->getArgOperand(1);
    Align Alignment = TaskData->getPointerAlignment(M.getDataLayout());
    Value *TaskShareds = Builder.CreateLoad(VoidPtr, TaskData);
    Builder.CreateMemCpy(TaskShareds, Alignment, Shareds, Alignment,
                         SharedsSize);
    // Get the pointer to loop lb, ub, step from task ptr
    // and set up the lowerbound,upperbound and step values
    llvm::Value *Lb = Builder.CreateGEP(
        FakeSharedsTy, TaskShareds, {Builder.getInt32(0), Builder.getInt32(0)});
 
    llvm::Value *Ub = Builder.CreateGEP(
        FakeSharedsTy, TaskShareds, {Builder.getInt32(0), Builder.getInt32(1)});
 
    llvm::Value *Step = Builder.CreateGEP(
        FakeSharedsTy, TaskShareds, {Builder.getInt32(0), Builder.getInt32(2)});
    llvm::Value *Loadstep = Builder.CreateLoad(Builder.getInt64Ty(), Step);
 
    // set up the arguments for emitting kmpc_taskloop runtime call
    // setting values for ifval, nogroup, sched, grainsize, task_dup
    Value *IfCondVal =
        IfCond ? Builder.CreateIntCast(IfCond, Builder.getInt32Ty(), true)
               : Builder.getInt32(1);
    // As __kmpc_taskgroup is called manually in OMPIRBuilder, NoGroupVal should
    // always be 1 when calling __kmpc_taskloop to ensure it is not called again
    Value *NoGroupVal = Builder.getInt32(1);
    Value *SchedVal = Builder.getInt32(Sched);
    Value *GrainSizeVal =
        GrainSize ? Builder.CreateIntCast(GrainSize, Builder.getInt64Ty(), true)
                  : Builder.getInt64(0);
    Value *TaskDup = TaskDupFn;
 
    Value *Args[] = {Ident,    ThreadID,   TaskData, IfCondVal,    Lb,     Ub,
                     Loadstep, NoGroupVal, SchedVal, GrainSizeVal, TaskDup};
 
    // taskloop runtime call
    Function *TaskloopFn =
        getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_taskloop);
    Builder.CreateCall(TaskloopFn, Args);
 
    // Emit the @__kmpc_end_taskgroup runtime call to end the taskgroup if
    // nogroup is not defined
    if (!NoGroup) {
      Function *EndTaskgroupFn =
          getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_taskgroup);
      Builder.CreateCall(EndTaskgroupFn, {Ident, ThreadID});
    }
 
    StaleCI->eraseFromParent();
 
    Builder.SetInsertPoint(TaskloopAllocaBB, TaskloopAllocaBB->begin());
 
    LoadInst *SharedsOutlined =
        Builder.CreateLoad(VoidPtr, OutlinedFn.getArg(1));
    OutlinedFn.getArg(1)->replaceUsesWithIf(
        SharedsOutlined,
        [SharedsOutlined](Use &U) { return U.getUser() != SharedsOutlined; });
 
    Value *IV = CLI->getIndVar();
    Type *IVTy = IV->getType();
    Constant *One = ConstantInt::get(Builder.getInt64Ty(), 1);
 
    // When outlining, CodeExtractor will create GEP's to the LowerBound and
    // UpperBound. These GEP's can be reused for loading the tasks respective
    // bounds.
    Value *TaskLB = nullptr;
    Value *TaskUB = nullptr;
    Value *TaskStep = nullptr;
    Value *LoadTaskLB = nullptr;
    Value *LoadTaskUB = nullptr;
    Value *LoadTaskStep = nullptr;
    for (Instruction &I : *TaskloopAllocaBB) {
      if (I.getOpcode() == Instruction::GetElementPtr) {
        GetElementPtrInst &Gep = cast<GetElementPtrInst>(I);
        if (ConstantInt *CI = dyn_cast<ConstantInt>(Gep.getOperand(2))) {
          switch (CI->getZExtValue()) {
          case 0:
            TaskLB = &I;
            break;
          case 1:
            TaskUB = &I;
            break;
          case 2:
            TaskStep = &I;
            break;
          }
        }
      } else if (I.getOpcode() == Instruction::Load) {
        LoadInst &Load = cast<LoadInst>(I);
        if (Load.getPointerOperand() == TaskLB) {
          assert(TaskLB != nullptr && "Expected value for TaskLB");
          LoadTaskLB = &I;
        } else if (Load.getPointerOperand() == TaskUB) {
          assert(TaskUB != nullptr && "Expected value for TaskUB");
          LoadTaskUB = &I;
        } else if (Load.getPointerOperand() == TaskStep) {
          assert(TaskStep != nullptr && "Expected value for TaskStep");
          LoadTaskStep = &I;
        }
      }
    }
 
    Builder.SetInsertPoint(CLI->getPreheader()->getTerminator());
 
    assert(LoadTaskLB != nullptr && "Expected value for LoadTaskLB");
    assert(LoadTaskUB != nullptr && "Expected value for LoadTaskUB");
    assert(LoadTaskStep != nullptr && "Expected value for LoadTaskStep");
    Value *TripCountMinusOne = Builder.CreateSDiv(
        Builder.CreateSub(LoadTaskUB, LoadTaskLB), LoadTaskStep);
    Value *TripCount = Builder.CreateAdd(TripCountMinusOne, One, "trip_cnt");
    Value *CastedTripCount = Builder.CreateIntCast(TripCount, IVTy, true);
    Value *CastedTaskLB = Builder.CreateIntCast(LoadTaskLB, IVTy, true);
    // set the trip count in the CLI
    CLI->setTripCount(CastedTripCount);
 
    Builder.SetInsertPoint(CLI->getBody(),
                           CLI->getBody()->getFirstInsertionPt());
 
    if (NumOfCollapseLoops > 1) {
      llvm::SmallVector<User *> UsersToReplace;
      // When using the collapse clause, the bounds of the loop have to be
      // adjusted to properly represent the iterator of the outer loop.
      Value *IVPlusTaskLB = Builder.CreateAdd(
          CLI->getIndVar(),
          Builder.CreateSub(CastedTaskLB, ConstantInt::get(IVTy, 1)));
      // To ensure every Use is correctly captured, we first want to record
      // which users to replace the value in, and then replace the value.
      for (auto IVUse = CLI->getIndVar()->uses().begin();
           IVUse != CLI->getIndVar()->uses().end(); IVUse++) {
        User *IVUser = IVUse->getUser();
        if (auto *Op = dyn_cast<BinaryOperator>(IVUser)) {
          if (Op->getOpcode() == Instruction::URem ||
              Op->getOpcode() == Instruction::UDiv) {
            UsersToReplace.push_back(IVUser);
          }
        }
      }
      for (User *User : UsersToReplace) {
        User->replaceUsesOfWith(CLI->getIndVar(), IVPlusTaskLB);
      }
    } else {
      // The canonical loop is generated with a fixed lower bound. We need to
      // update the index calculation code to use the task's lower bound. The
      // generated code looks like this:
      // %omp_loop.iv = phi ...
      // ...
      // %tmp = mul [type] %omp_loop.iv, step
      // %user_index = add [type] tmp, lb
      // OpenMPIRBuilder constructs canonical loops to have exactly three uses
      // of the normalised induction variable:
      // 1. This one: converting the normalised IV to the user IV
      // 2. The increment (add)
      // 3. The comparison against the trip count (icmp)
      // (1) is the only use that is a mul followed by an add so this cannot
      // match other IR.
      assert(CLI->getIndVar()->getNumUses() == 3 &&
             "Canonical loop should have exactly three uses of the ind var");
      for (User *IVUser : CLI->getIndVar()->users()) {
        if (auto *Mul = dyn_cast<BinaryOperator>(IVUser)) {
          if (Mul->getOpcode() == Instruction::Mul) {
            for (User *MulUser : Mul->users()) {
              if (auto *Add = dyn_cast<BinaryOperator>(MulUser)) {
                if (Add->getOpcode() == Instruction::Add) {
                  Add->setOperand(1, CastedTaskLB);
                }
              }
            }
          }
        }
      }
    }
 
    FakeLB->replaceAllUsesWith(CastedLBVal);
    FakeUB->replaceAllUsesWith(CastedUBVal);
    FakeStep->replaceAllUsesWith(CastedStepVal);
    for (Instruction *I : llvm::reverse(ToBeDeleted)) {
      I->eraseFromParent();
    }
  };
 
  addOutlineInfo(std::move(OI));
  Builder.SetInsertPoint(TaskloopExitBB, TaskloopExitBB->begin());
  return Builder.saveIP();
}
 
llvm::StructType *OpenMPIRBuilder::getKmpTaskAffinityInfoTy() {
  llvm::Type *IntPtrTy = llvm::Type::getIntNTy(
      M.getContext(), M.getDataLayout().getPointerSizeInBits());
  return llvm::StructType::get(IntPtrTy, IntPtrTy,
                               llvm::Type::getInt32Ty(M.getContext()));
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createTask(
    const LocationDescription &Loc, InsertPointTy AllocaIP,
    ArrayRef<BasicBlock *> DeallocBlocks, BodyGenCallbackTy BodyGenCB,
    bool Tied, Value *Final, Value *IfCondition,
    const DependenciesInfo &Dependencies, const AffinityData &Affinities,
    bool Mergeable, Value *EventHandle, Value *Priority) {
 
  if (!updateToLocation(Loc))
    return InsertPointTy();
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  // The current basic block is split into four basic blocks. After outlining,
  // they will be mapped as follows:
  // ```
  // def current_fn() {
  //   current_basic_block:
  //     br label %task.exit
  //   task.exit:
  //     ; instructions after task
  // }
  // def outlined_fn() {
  //   task.alloca:
  //     br label %task.body
  //   task.body:
  //     ret void
  // }
  // ```
  BasicBlock *TaskExitBB = splitBB(Builder, /*CreateBranch=*/true, "task.exit");
  BasicBlock *TaskBodyBB = splitBB(Builder, /*CreateBranch=*/true, "task.body");
  BasicBlock *TaskAllocaBB =
      splitBB(Builder, /*CreateBranch=*/true, "task.alloca");
 
  InsertPointTy TaskAllocaIP =
      InsertPointTy(TaskAllocaBB, TaskAllocaBB->begin());
  InsertPointTy TaskBodyIP = InsertPointTy(TaskBodyBB, TaskBodyBB->begin());
  if (Error Err = BodyGenCB(TaskAllocaIP, TaskBodyIP, TaskExitBB))
    return Err;
 
  auto OI = std::make_unique<OutlineInfo>();
  OI->EntryBB = TaskAllocaBB;
  OI->OuterAllocBB = AllocaIP.getBlock();
  OI->ExitBB = TaskExitBB;
  OI->OuterDeallocBBs.reserve(DeallocBlocks.size());
  copy(DeallocBlocks, OI->OuterDeallocBBs.end());
 
  // Add the thread ID argument.
  SmallVector<Instruction *, 4> ToBeDeleted;
  OI->ExcludeArgsFromAggregate.push_back(createFakeIntVal(
      Builder, AllocaIP, ToBeDeleted, TaskAllocaIP, "global.tid", false));
 
  OI->PostOutlineCB = [this, Ident, Tied, Final, IfCondition, Dependencies,
                       Affinities, Mergeable, Priority, EventHandle,
                       TaskAllocaBB,
                       ToBeDeleted](Function &OutlinedFn) mutable {
    // Replace the Stale CI by appropriate RTL function call.
    assert(OutlinedFn.hasOneUse() &&
           "there must be a single user for the outlined function");
    CallInst *StaleCI = cast<CallInst>(OutlinedFn.user_back());
 
    // HasShareds is true if any variables are captured in the outlined region,
    // false otherwise.
    bool HasShareds = StaleCI->arg_size() > 1;
    Builder.SetInsertPoint(StaleCI);
 
    // Gather the arguments for emitting the runtime call for
    // @__kmpc_omp_task_alloc
    Function *TaskAllocFn =
        getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_alloc);
 
    // Arguments - `loc_ref` (Ident) and `gtid` (ThreadID)
    // call.
    Value *ThreadID = getOrCreateThreadID(Ident);
 
    // Argument - `flags`
    // Task is tied iff (Flags & 1) == 1.
    // Task is untied iff (Flags & 1) == 0.
    // Task is final iff (Flags & 2) == 2.
    // Task is not final iff (Flags & 2) == 0.
    // Task is mergeable or merged-if0 iff (Flags & 4) == 4.
    // Task is neither mergeable nor merged-if0 iff (Flags & 4) == 0.
    // Task is detachable iff (Flags & 64) == 64.
    // Task is not detachable iff (Flags & 64) == 0.
    // Task is priority iff (Flags & 32) == 32.
    // Task is not priority iff (Flags & 32) == 0.
    // TODO: Handle the other flags.
    Value *Flags = Builder.getInt32(Tied);
    auto *ConstIfCondition = dyn_cast_or_null<ConstantInt>(IfCondition);
    bool UseMergedIf0Path = ConstIfCondition && ConstIfCondition->isZero();
    if (Final) {
      Value *FinalFlag =
          Builder.CreateSelect(Final, Builder.getInt32(2), Builder.getInt32(0));
      Flags = Builder.CreateOr(FinalFlag, Flags);
    }
 
    if (Mergeable || UseMergedIf0Path)
      Flags = Builder.CreateOr(Builder.getInt32(4), Flags);
    if (EventHandle)
      Flags = Builder.CreateOr(Builder.getInt32(64), Flags);
    if (Priority)
      Flags = Builder.CreateOr(Builder.getInt32(32), Flags);
 
    // Argument - `sizeof_kmp_task_t` (TaskSize)
    // Tasksize refers to the size in bytes of kmp_task_t data structure
    // including private vars accessed in task.
    // TODO: add kmp_task_t_with_privates (privates)
    Value *TaskSize = Builder.getInt64(
        divideCeil(M.getDataLayout().getTypeSizeInBits(Task), 8));
 
    // Argument - `sizeof_shareds` (SharedsSize)
    // SharedsSize refers to the shareds array size in the kmp_task_t data
    // structure.
    Value *SharedsSize = Builder.getInt64(0);
    if (HasShareds) {
      AllocaInst *ArgStructAlloca =
          dyn_cast<AllocaInst>(StaleCI->getArgOperand(1));
      assert(ArgStructAlloca &&
             "Unable to find the alloca instruction corresponding to arguments "
             "for extracted function");
      std::optional<TypeSize> ArgAllocSize =
          ArgStructAlloca->getAllocationSize(M.getDataLayout());
      assert(ArgAllocSize &&
             "Unable to determine size of arguments for extracted function");
      SharedsSize = Builder.getInt64(ArgAllocSize->getFixedValue());
    }
    // Emit the @__kmpc_omp_task_alloc runtime call
    // The runtime call returns a pointer to an area where the task captured
    // variables must be copied before the task is run (TaskData)
    CallInst *TaskData = createRuntimeFunctionCall(
        TaskAllocFn, {/*loc_ref=*/Ident, /*gtid=*/ThreadID, /*flags=*/Flags,
                      /*sizeof_task=*/TaskSize, /*sizeof_shared=*/SharedsSize,
                      /*task_func=*/&OutlinedFn});
 
    if (Affinities.Count && Affinities.Info) {
      Function *RegAffFn = getOrCreateRuntimeFunctionPtr(
          OMPRTL___kmpc_omp_reg_task_with_affinity);
 
      createRuntimeFunctionCall(RegAffFn, {Ident, ThreadID, TaskData,
                                           Affinities.Count, Affinities.Info});
    }
 
    // Emit detach clause initialization.
    // evt = (typeof(evt))__kmpc_task_allow_completion_event(loc, tid,
    // task_descriptor);
    if (EventHandle) {
      Function *TaskDetachFn = getOrCreateRuntimeFunctionPtr(
          OMPRTL___kmpc_task_allow_completion_event);
      llvm::Value *EventVal =
          createRuntimeFunctionCall(TaskDetachFn, {Ident, ThreadID, TaskData});
      llvm::Value *EventHandleAddr =
          Builder.CreatePointerBitCastOrAddrSpaceCast(EventHandle,
                                                      Builder.getPtrTy(0));
      EventVal = Builder.CreatePtrToInt(EventVal, Builder.getInt64Ty());
      Builder.CreateStore(EventVal, EventHandleAddr);
    }
    // Copy the arguments for outlined function
    if (HasShareds) {
      Value *Shareds = StaleCI->getArgOperand(1);
      Align Alignment = TaskData->getPointerAlignment(M.getDataLayout());
      Value *TaskShareds = Builder.CreateLoad(VoidPtr, TaskData);
      Builder.CreateMemCpy(TaskShareds, Alignment, Shareds, Alignment,
                           SharedsSize);
    }
 
    if (Priority) {
      //
      // The return type of "__kmpc_omp_task_alloc" is "kmp_task_t *",
      // we populate the priority information into the "kmp_task_t" here
      //
      // The struct "kmp_task_t" definition is available in kmp.h
      // kmp_task_t = { shareds, routine, part_id, data1, data2 }
      // data2 is used for priority
      //
      Type *Int32Ty = Builder.getInt32Ty();
      Constant *Zero = ConstantInt::get(Int32Ty, 0);
      // kmp_task_t* => { ptr }
      Type *TaskPtr = StructType::get(VoidPtr);
      Value *TaskGEP =
          Builder.CreateInBoundsGEP(TaskPtr, TaskData, {Zero, Zero});
      // kmp_task_t => { ptr, ptr, i32, ptr, ptr }
      Type *TaskStructType = StructType::get(
          VoidPtr, VoidPtr, Builder.getInt32Ty(), VoidPtr, VoidPtr);
      Value *PriorityData = Builder.CreateInBoundsGEP(
          TaskStructType, TaskGEP, {Zero, ConstantInt::get(Int32Ty, 4)});
      // kmp_cmplrdata_t => { ptr, ptr }
      Type *CmplrStructType = StructType::get(VoidPtr, VoidPtr);
      Value *CmplrData = Builder.CreateInBoundsGEP(CmplrStructType,
                                                   PriorityData, {Zero, Zero});
      Builder.CreateStore(Priority, CmplrData);
    }
 
    Value *DepArray = nullptr;
    Value *NumDeps = nullptr;
    if (Dependencies.DepArray) {
      DepArray = Dependencies.DepArray;
      NumDeps = Dependencies.NumDeps;
    } else if (!Dependencies.Deps.empty()) {
      DepArray = emitTaskDependencies(*this, Dependencies.Deps);
      NumDeps = Builder.getInt32(Dependencies.Deps.size());
    }
 
    // In the presence of the `if` clause, the following IR is generated:
    //    ...
    //    %data = call @__kmpc_omp_task_alloc(...)
    //    br i1 %if_condition, label %then, label %else
    //  then:
    //    call @__kmpc_omp_task(...)
    //    br label %exit
    //  else:
    //    ;; Wait for resolution of dependencies, if any, before
    //    ;; beginning the task
    //    call @__kmpc_omp_wait_deps(...)
    //    call @__kmpc_omp_task_begin_if0(...)
    //    call @outlined_fn(...)
    //    call @__kmpc_omp_task_complete_if0(...)
    //    br label %exit
    //  exit:
    //    ...
    if (IfCondition && !UseMergedIf0Path) {
      // `SplitBlockAndInsertIfThenElse` requires the block to have a
      // terminator.
      splitBB(Builder, /*CreateBranch=*/true, "if.end");
      Instruction *IfTerminator =
          Builder.GetInsertPoint()->getParent()->getTerminator();
      Instruction *ThenTI = IfTerminator, *ElseTI = nullptr;
      Builder.SetInsertPoint(IfTerminator);
      SplitBlockAndInsertIfThenElse(IfCondition, IfTerminator, &ThenTI,
                                    &ElseTI);
      Builder.SetInsertPoint(ElseTI);
 
      if (DepArray) {
        Function *TaskWaitFn =
            getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_wait_deps);
        createRuntimeFunctionCall(
            TaskWaitFn,
            {Ident, ThreadID, NumDeps, DepArray,
             ConstantInt::get(Builder.getInt32Ty(), 0),
             ConstantPointerNull::get(PointerType::getUnqual(M.getContext()))});
      }
      Function *TaskBeginFn =
          getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_begin_if0);
      Function *TaskCompleteFn =
          getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_complete_if0);
      createRuntimeFunctionCall(TaskBeginFn, {Ident, ThreadID, TaskData});
      CallInst *CI = nullptr;
      if (HasShareds)
        CI = createRuntimeFunctionCall(&OutlinedFn, {ThreadID, TaskData});
      else
        CI = createRuntimeFunctionCall(&OutlinedFn, {ThreadID});
      CI->setDebugLoc(StaleCI->getDebugLoc());
      createRuntimeFunctionCall(TaskCompleteFn, {Ident, ThreadID, TaskData});
      Builder.SetInsertPoint(ThenTI);
    }
 
    if (DepArray) {
      Function *TaskFn =
          getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_with_deps);
      createRuntimeFunctionCall(
          TaskFn,
          {Ident, ThreadID, TaskData, NumDeps, DepArray,
           ConstantInt::get(Builder.getInt32Ty(), 0),
           ConstantPointerNull::get(PointerType::getUnqual(M.getContext()))});
 
    } else {
      // Emit the @__kmpc_omp_task runtime call to spawn the task
      Function *TaskFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task);
      createRuntimeFunctionCall(TaskFn, {Ident, ThreadID, TaskData});
    }
 
    StaleCI->eraseFromParent();
 
    Builder.SetInsertPoint(TaskAllocaBB, TaskAllocaBB->begin());
    if (HasShareds) {
      LoadInst *Shareds = Builder.CreateLoad(VoidPtr, OutlinedFn.getArg(1));
      OutlinedFn.getArg(1)->replaceUsesWithIf(
          Shareds, [Shareds](Use &U) { return U.getUser() != Shareds; });
    }
 
    // The insert point may refer to one of the instructions about to be
    // deleted. It is not needed anymore so clear it instead of leaving it
    // dangling.
    Builder.ClearInsertionPoint();
    for (Instruction *I : llvm::reverse(ToBeDeleted))
      I->eraseFromParent();
  };
 
  addOutlineInfo(std::move(OI));
  Builder.SetInsertPoint(TaskExitBB, TaskExitBB->begin());
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createTaskgroup(
    const LocationDescription &Loc, InsertPointTy AllocaIP,
    ArrayRef<BasicBlock *> DeallocBlocks, BodyGenCallbackTy BodyGenCB) {
  if (!updateToLocation(Loc))
    return InsertPointTy();
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadID = getOrCreateThreadID(Ident);
 
  // Emit the @__kmpc_taskgroup runtime call to start the taskgroup
  Function *TaskgroupFn =
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_taskgroup);
  createRuntimeFunctionCall(TaskgroupFn, {Ident, ThreadID});
 
  BasicBlock *TaskgroupExitBB = splitBB(Builder, true, "taskgroup.exit");
  if (Error Err = BodyGenCB(AllocaIP, Builder.saveIP(), DeallocBlocks))
    return Err;
 
  Builder.SetInsertPoint(TaskgroupExitBB);
  // Emit the @__kmpc_end_taskgroup runtime call to end the taskgroup
  Function *EndTaskgroupFn =
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_taskgroup);
  createRuntimeFunctionCall(EndTaskgroupFn, {Ident, ThreadID});
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createSections(
    const LocationDescription &Loc, InsertPointTy AllocaIP,
    ArrayRef<StorableBodyGenCallbackTy> SectionCBs, PrivatizeCallbackTy PrivCB,
    FinalizeCallbackTy FiniCB, bool IsCancellable, bool IsNowait) {
  assert(!isConflictIP(AllocaIP, Loc.IP) && "Dedicated IP allocas required");
 
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  FinalizationStack.push_back({FiniCB, OMPD_sections, IsCancellable});
 
  // Each section is emitted as a switch case
  // Each finalization callback is handled from clang.EmitOMPSectionDirective()
  // -> OMP.createSection() which generates the IR for each section
  // Iterate through all sections and emit a switch construct:
  // switch (IV) {
  //   case 0:
  //     <SectionStmt[0]>;
  //     break;
  // ...
  //   case <NumSection> - 1:
  //     <SectionStmt[<NumSection> - 1]>;
  //     break;
  // }
  // ...
  // section_loop.after:
  // <FiniCB>;
  auto LoopBodyGenCB = [&](InsertPointTy CodeGenIP, Value *IndVar) -> Error {
    Builder.restoreIP(CodeGenIP);
    BasicBlock *Continue =
        splitBBWithSuffix(Builder, /*CreateBranch=*/false, ".sections.after");
    Function *CurFn = Continue->getParent();
    SwitchInst *SwitchStmt = Builder.CreateSwitch(IndVar, Continue);
 
    unsigned CaseNumber = 0;
    for (auto SectionCB : SectionCBs) {
      BasicBlock *CaseBB = BasicBlock::Create(
          M.getContext(), "omp_section_loop.body.case", CurFn, Continue);
      SwitchStmt->addCase(Builder.getInt32(CaseNumber), CaseBB);
      Builder.SetInsertPoint(CaseBB);
      UncondBrInst *CaseEndBr = Builder.CreateBr(Continue);
      if (Error Err =
              SectionCB(InsertPointTy(),
                        {CaseEndBr->getParent(), CaseEndBr->getIterator()}, {}))
        return Err;
      CaseNumber++;
    }
    // remove the existing terminator from body BB since there can be no
    // terminators after switch/case
    return Error::success();
  };
  // Loop body ends here
  // LowerBound, UpperBound, and STride for createCanonicalLoop
  Type *I32Ty = Type::getInt32Ty(M.getContext());
  Value *LB = ConstantInt::get(I32Ty, 0);
  Value *UB = ConstantInt::get(I32Ty, SectionCBs.size());
  Value *ST = ConstantInt::get(I32Ty, 1);
  Expected<CanonicalLoopInfo *> LoopInfo = createCanonicalLoop(
      Loc, LoopBodyGenCB, LB, UB, ST, true, false, AllocaIP, "section_loop");
  if (!LoopInfo)
    return LoopInfo.takeError();
 
  InsertPointOrErrorTy WsloopIP =
      applyStaticWorkshareLoop(Loc.DL, *LoopInfo, AllocaIP,
                               WorksharingLoopType::ForStaticLoop, !IsNowait);
  if (!WsloopIP)
    return WsloopIP.takeError();
  InsertPointTy AfterIP = *WsloopIP;
 
  BasicBlock *LoopFini = AfterIP.getBlock()->getSinglePredecessor();
  assert(LoopFini && "Bad structure of static workshare loop finalization");
 
  // Apply the finalization callback in LoopAfterBB
  auto FiniInfo = FinalizationStack.pop_back_val();
  assert(FiniInfo.DK == OMPD_sections &&
         "Unexpected finalization stack state!");
  if (Error Err = FiniInfo.mergeFiniBB(Builder, LoopFini))
    return Err;
 
  return AfterIP;
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::createSection(const LocationDescription &Loc,
                               BodyGenCallbackTy BodyGenCB,
                               FinalizeCallbackTy FiniCB) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  auto FiniCBWrapper = [&](InsertPointTy IP) {
    if (IP.getBlock()->end() != IP.getPoint())
      return FiniCB(IP);
    // This must be done otherwise any nested constructs using FinalizeOMPRegion
    // will fail because that function requires the Finalization Basic Block to
    // have a terminator, which is already removed by EmitOMPRegionBody.
    // IP is currently at cancelation block.
    // We need to backtrack to the condition block to fetch
    // the exit block and create a branch from cancelation
    // to exit block.
    IRBuilder<>::InsertPointGuard IPG(Builder);
    Builder.restoreIP(IP);
    auto *CaseBB = Loc.IP.getBlock();
    auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor();
    auto *ExitBB = CondBB->getTerminator()->getSuccessor(1);
    Instruction *I = Builder.CreateBr(ExitBB);
    IP = InsertPointTy(I->getParent(), I->getIterator());
    return FiniCB(IP);
  };
 
  Directive OMPD = Directive::OMPD_sections;
  // Since we are using Finalization Callback here, HasFinalize
  // and IsCancellable have to be true
  return EmitOMPInlinedRegion(OMPD, nullptr, nullptr, BodyGenCB, FiniCBWrapper,
                              /*Conditional*/ false, /*hasFinalize*/ true,
                              /*IsCancellable*/ true);
}
 
static OpenMPIRBuilder::InsertPointTy getInsertPointAfterInstr(Instruction *I) {
  BasicBlock::iterator IT(I);
  IT++;
  return OpenMPIRBuilder::InsertPointTy(I->getParent(), IT);
}
 
Value *OpenMPIRBuilder::getGPUThreadID() {
  return createRuntimeFunctionCall(
      getOrCreateRuntimeFunction(M,
                                 OMPRTL___kmpc_get_hardware_thread_id_in_block),
      {});
}
 
Value *OpenMPIRBuilder::getGPUWarpSize() {
  return createRuntimeFunctionCall(
      getOrCreateRuntimeFunction(M, OMPRTL___kmpc_get_warp_size), {});
}
 
Value *OpenMPIRBuilder::getNVPTXWarpID() {
  unsigned LaneIDBits = Log2_32(Config.getGridValue().GV_Warp_Size);
  return Builder.CreateAShr(getGPUThreadID(), LaneIDBits, "nvptx_warp_id");
}
 
Value *OpenMPIRBuilder::getNVPTXLaneID() {
  unsigned LaneIDBits = Log2_32(Config.getGridValue().GV_Warp_Size);
  assert(LaneIDBits < 32 && "Invalid LaneIDBits size in NVPTX device.");
  unsigned LaneIDMask = ~0u >> (32u - LaneIDBits);
  return Builder.CreateAnd(getGPUThreadID(), Builder.getInt32(LaneIDMask),
                           "nvptx_lane_id");
}
 
Value *OpenMPIRBuilder::castValueToType(InsertPointTy AllocaIP, Value *From,
                                        Type *ToType) {
  Type *FromType = From->getType();
  uint64_t FromSize = M.getDataLayout().getTypeStoreSize(FromType);
  uint64_t ToSize = M.getDataLayout().getTypeStoreSize(ToType);
  assert(FromSize > 0 && "From size must be greater than zero");
  assert(ToSize > 0 && "To size must be greater than zero");
  if (FromType == ToType)
    return From;
  if (FromSize == ToSize)
    return Builder.CreateBitCast(From, ToType);
  if (ToType->isIntegerTy() && FromType->isIntegerTy())
    return Builder.CreateIntCast(From, ToType, /*isSigned*/ true);
  InsertPointTy SaveIP = Builder.saveIP();
  Builder.restoreIP(AllocaIP);
  Value *CastItem = Builder.CreateAlloca(ToType);
  Builder.restoreIP(SaveIP);
 
  Value *ValCastItem = Builder.CreatePointerBitCastOrAddrSpaceCast(
      CastItem, Builder.getPtrTy(0));
  Builder.CreateStore(From, ValCastItem);
  return Builder.CreateLoad(ToType, CastItem);
}
 
Value *OpenMPIRBuilder::createRuntimeShuffleFunction(InsertPointTy AllocaIP,
                                                     Value *Element,
                                                     Type *ElementType,
                                                     Value *Offset) {
  uint64_t Size = M.getDataLayout().getTypeStoreSize(ElementType);
  assert(Size <= 8 && "Unsupported bitwidth in shuffle instruction");
 
  // Cast all types to 32- or 64-bit values before calling shuffle routines.
  Type *CastTy = Builder.getIntNTy(Size <= 4 ? 32 : 64);
  Value *ElemCast = castValueToType(AllocaIP, Element, CastTy);
  Value *WarpSize =
      Builder.CreateIntCast(getGPUWarpSize(), Builder.getInt16Ty(), true);
  Function *ShuffleFunc = getOrCreateRuntimeFunctionPtr(
      Size <= 4 ? RuntimeFunction::OMPRTL___kmpc_shuffle_int32
                : RuntimeFunction::OMPRTL___kmpc_shuffle_int64);
  Value *WarpSizeCast =
      Builder.CreateIntCast(WarpSize, Builder.getInt16Ty(), /*isSigned=*/true);
  Value *ShuffleCall =
      createRuntimeFunctionCall(ShuffleFunc, {ElemCast, Offset, WarpSizeCast});
  // The shuffle runtime functions return a 32- or 64-bit value. Cast it back
  // down to the requested element type, otherwise storing the result would
  // write past the end of an element narrower than the shuffle width.
  return castValueToType(AllocaIP, ShuffleCall, ElementType);
}
 
void OpenMPIRBuilder::shuffleAndStore(InsertPointTy AllocaIP, Value *SrcAddr,
                                      Value *DstAddr, Type *ElemType,
                                      Value *Offset, Type *ReductionArrayTy,
                                      bool IsByRefElem) {
  uint64_t Size = M.getDataLayout().getTypeStoreSize(ElemType);
  // Create the loop over the big sized data.
  // ptr = (void*)Elem;
  // ptrEnd = (void*) Elem + 1;
  // Step = 8;
  // while (ptr + Step < ptrEnd)
  //   shuffle((int64_t)*ptr);
  // Step = 4;
  // while (ptr + Step < ptrEnd)
  //   shuffle((int32_t)*ptr);
  // ...
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  Value *ElemPtr = DstAddr;
  Value *Ptr = SrcAddr;
  for (unsigned IntSize = 8; IntSize >= 1; IntSize /= 2) {
    if (Size < IntSize)
      continue;
    Type *IntType = Builder.getIntNTy(IntSize * 8);
    Ptr = Builder.CreatePointerBitCastOrAddrSpaceCast(
        Ptr, Builder.getPtrTy(0), Ptr->getName() + ".ascast");
    Value *SrcAddrGEP =
        Builder.CreateGEP(ElemType, SrcAddr, {ConstantInt::get(IndexTy, 1)});
    ElemPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
        ElemPtr, Builder.getPtrTy(0), ElemPtr->getName() + ".ascast");
 
    Function *CurFunc = Builder.GetInsertBlock()->getParent();
    if ((Size / IntSize) > 1) {
      Value *PtrEnd = Builder.CreatePointerBitCastOrAddrSpaceCast(
          SrcAddrGEP, Builder.getPtrTy());
      BasicBlock *PreCondBB =
          BasicBlock::Create(M.getContext(), ".shuffle.pre_cond");
      BasicBlock *ThenBB = BasicBlock::Create(M.getContext(), ".shuffle.then");
      BasicBlock *ExitBB = BasicBlock::Create(M.getContext(), ".shuffle.exit");
      BasicBlock *CurrentBB = Builder.GetInsertBlock();
      emitBlock(PreCondBB, CurFunc);
      PHINode *PhiSrc =
          Builder.CreatePHI(Ptr->getType(), /*NumReservedValues=*/2);
      PhiSrc->addIncoming(Ptr, CurrentBB);
      PHINode *PhiDest =
          Builder.CreatePHI(ElemPtr->getType(), /*NumReservedValues=*/2);
      PhiDest->addIncoming(ElemPtr, CurrentBB);
      Ptr = PhiSrc;
      ElemPtr = PhiDest;
      Value *PtrDiff = Builder.CreatePtrDiff(
          Builder.getInt8Ty(), PtrEnd,
          Builder.CreatePointerBitCastOrAddrSpaceCast(Ptr, Builder.getPtrTy()));
      Builder.CreateCondBr(
          Builder.CreateICmpSGT(PtrDiff, Builder.getInt64(IntSize - 1)), ThenBB,
          ExitBB);
      emitBlock(ThenBB, CurFunc);
      Value *Res = createRuntimeShuffleFunction(
          AllocaIP,
          Builder.CreateAlignedLoad(
              IntType, Ptr, M.getDataLayout().getPrefTypeAlign(ElemType)),
          IntType, Offset);
      Builder.CreateAlignedStore(Res, ElemPtr,
                                 M.getDataLayout().getPrefTypeAlign(ElemType));
      Value *LocalPtr =
          Builder.CreateGEP(IntType, Ptr, {ConstantInt::get(IndexTy, 1)});
      Value *LocalElemPtr =
          Builder.CreateGEP(IntType, ElemPtr, {ConstantInt::get(IndexTy, 1)});
      PhiSrc->addIncoming(LocalPtr, ThenBB);
      PhiDest->addIncoming(LocalElemPtr, ThenBB);
      emitBranch(PreCondBB);
      emitBlock(ExitBB, CurFunc);
    } else {
      // The shuffled value comes back as the chunk's integer type, so the
      // store covers exactly this chunk regardless of what ElemType is.
      Value *Res = createRuntimeShuffleFunction(
          AllocaIP, Builder.CreateLoad(IntType, Ptr), IntType, Offset);
      Builder.CreateStore(Res, ElemPtr);
      Ptr = Builder.CreateGEP(IntType, Ptr, {ConstantInt::get(IndexTy, 1)});
      ElemPtr =
          Builder.CreateGEP(IntType, ElemPtr, {ConstantInt::get(IndexTy, 1)});
    }
    Size = Size % IntSize;
  }
}
 
Error OpenMPIRBuilder::emitReductionListCopy(
    InsertPointTy AllocaIP, CopyAction Action, Type *ReductionArrayTy,
    ArrayRef<ReductionInfo> ReductionInfos, Value *SrcBase, Value *DestBase,
    ArrayRef<bool> IsByRef, CopyOptionsTy CopyOptions) {
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset;
 
  // Iterates, element-by-element, through the source Reduce list and
  // make a copy.
  for (auto En : enumerate(ReductionInfos)) {
    const ReductionInfo &RI = En.value();
    Value *SrcElementAddr = nullptr;
    AllocaInst *DestAlloca = nullptr;
    Value *DestElementAddr = nullptr;
    Value *DestElementPtrAddr = nullptr;
    // Should we shuffle in an element from a remote lane?
    bool ShuffleInElement = false;
    // Set to true to update the pointer in the dest Reduce list to a
    // newly created element.
    bool UpdateDestListPtr = false;
 
    // Step 1.1: Get the address for the src element in the Reduce list.
    Value *SrcElementPtrAddr = Builder.CreateInBoundsGEP(
        ReductionArrayTy, SrcBase,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
    SrcElementAddr = Builder.CreateLoad(Builder.getPtrTy(), SrcElementPtrAddr);
 
    // Step 1.2: Create a temporary to store the element in the destination
    // Reduce list.
    DestElementPtrAddr = Builder.CreateInBoundsGEP(
        ReductionArrayTy, DestBase,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
    bool IsByRefElem = (!IsByRef.empty() && IsByRef[En.index()]);
    switch (Action) {
    case CopyAction::RemoteLaneToThread: {
      InsertPointTy CurIP = Builder.saveIP();
      Builder.restoreIP(AllocaIP);
 
      Type *DestAllocaType =
          IsByRefElem ? RI.ByRefAllocatedType : RI.ElementType;
      DestAlloca = Builder.CreateAlloca(DestAllocaType, nullptr,
                                        ".omp.reduction.element");
      DestAlloca->setAlignment(
          M.getDataLayout().getPrefTypeAlign(DestAllocaType));
      DestElementAddr = DestAlloca;
      DestElementAddr =
          Builder.CreateAddrSpaceCast(DestElementAddr, Builder.getPtrTy(),
                                      DestElementAddr->getName() + ".ascast");
      Builder.restoreIP(CurIP);
      ShuffleInElement = true;
      UpdateDestListPtr = true;
      break;
    }
    case CopyAction::ThreadCopy: {
      DestElementAddr =
          Builder.CreateLoad(Builder.getPtrTy(), DestElementPtrAddr);
      break;
    }
    }
 
    // Now that all active lanes have read the element in the
    // Reduce list, shuffle over the value from the remote lane.
    if (ShuffleInElement) {
      Type *ShuffleType = RI.ElementType;
      Value *ShuffleSrcAddr = SrcElementAddr;
      Value *ShuffleDestAddr = DestElementAddr;
      AllocaInst *LocalStorage = nullptr;
 
      if (IsByRefElem) {
        assert(RI.ByRefElementType && "Expected by-ref element type to be set");
        assert(RI.ByRefAllocatedType &&
               "Expected by-ref allocated type to be set");
        // For by-ref reductions, we need to copy from the remote lane the
        // actual value of the partial reduction computed by that remote lane;
        // rather than, for example, a pointer to that data or, even worse, a
        // pointer to the descriptor of the by-ref reduction element.
        ShuffleType = RI.ByRefElementType;
 
        if (RI.DataPtrPtrGen) {
          // Descriptor-based by-ref: extract data pointer from descriptor.
          InsertPointOrErrorTy GenResult = RI.DataPtrPtrGen(
              Builder.saveIP(), ShuffleSrcAddr, ShuffleSrcAddr);
 
          if (!GenResult)
            return GenResult.takeError();
 
          ShuffleSrcAddr =
              Builder.CreateLoad(Builder.getPtrTy(), ShuffleSrcAddr);
 
          {
            InsertPointTy OldIP = Builder.saveIP();
            Builder.restoreIP(AllocaIP);
 
            LocalStorage = Builder.CreateAlloca(ShuffleType);
            Builder.restoreIP(OldIP);
            ShuffleDestAddr = LocalStorage;
          }
        } else {
          // Non-descriptor by-ref: the pointer already references data
          // directly. Shuffle into the destination alloca.
          ShuffleDestAddr = DestElementAddr;
        }
      }
 
      shuffleAndStore(AllocaIP, ShuffleSrcAddr, ShuffleDestAddr, ShuffleType,
                      RemoteLaneOffset, ReductionArrayTy, IsByRefElem);
 
      if (IsByRefElem && RI.DataPtrPtrGen) {
        // Copy descriptor from source and update base_ptr to shuffled data
        Value *DestDescriptorAddr = Builder.CreatePointerBitCastOrAddrSpaceCast(
            DestAlloca, Builder.getPtrTy(), ".ascast");
 
        InsertPointOrErrorTy GenResult = generateReductionDescriptor(
            DestDescriptorAddr, LocalStorage, SrcElementAddr,
            RI.ByRefAllocatedType, RI.DataPtrPtrGen);
 
        if (!GenResult)
          return GenResult.takeError();
      }
    } else {
      switch (RI.EvaluationKind) {
      case EvalKind::Scalar: {
        Value *Elem = Builder.CreateLoad(RI.ElementType, SrcElementAddr);
        // Store the source element value to the dest element address.
        Builder.CreateStore(Elem, DestElementAddr);
        break;
      }
      case EvalKind::Complex: {
        Value *SrcRealPtr = Builder.CreateConstInBoundsGEP2_32(
            RI.ElementType, SrcElementAddr, 0, 0, ".realp");
        Value *SrcReal = Builder.CreateLoad(
            RI.ElementType->getStructElementType(0), SrcRealPtr, ".real");
        Value *SrcImgPtr = Builder.CreateConstInBoundsGEP2_32(
            RI.ElementType, SrcElementAddr, 0, 1, ".imagp");
        Value *SrcImg = Builder.CreateLoad(
            RI.ElementType->getStructElementType(1), SrcImgPtr, ".imag");
 
        Value *DestRealPtr = Builder.CreateConstInBoundsGEP2_32(
            RI.ElementType, DestElementAddr, 0, 0, ".realp");
        Value *DestImgPtr = Builder.CreateConstInBoundsGEP2_32(
            RI.ElementType, DestElementAddr, 0, 1, ".imagp");
        Builder.CreateStore(SrcReal, DestRealPtr);
        Builder.CreateStore(SrcImg, DestImgPtr);
        break;
      }
      case EvalKind::Aggregate: {
        Value *SizeVal = Builder.getInt64(
            M.getDataLayout().getTypeStoreSize(RI.ElementType));
        Builder.CreateMemCpy(
            DestElementAddr, M.getDataLayout().getPrefTypeAlign(RI.ElementType),
            SrcElementAddr, M.getDataLayout().getPrefTypeAlign(RI.ElementType),
            SizeVal, false);
        break;
      }
      };
    }
 
    // Step 3.1: Modify reference in dest Reduce list as needed.
    // Modifying the reference in Reduce list to point to the newly
    // created element.  The element is live in the current function
    // scope and that of functions it invokes (i.e., reduce_function).
    // RemoteReduceData[i] = (void*)&RemoteElem
    if (UpdateDestListPtr) {
      Value *CastDestAddr = Builder.CreatePointerBitCastOrAddrSpaceCast(
          DestElementAddr, Builder.getPtrTy(),
          DestElementAddr->getName() + ".ascast");
      Builder.CreateStore(CastDestAddr, DestElementPtrAddr);
    }
  }
 
  return Error::success();
}
 
Expected<Function *> OpenMPIRBuilder::emitInterWarpCopyFunction(
    const LocationDescription &Loc, ArrayRef<ReductionInfo> ReductionInfos,
    AttributeList FuncAttrs, ArrayRef<bool> IsByRef) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  LLVMContext &Ctx = M.getContext();
  FunctionType *FuncTy = FunctionType::get(
      Builder.getVoidTy(), {Builder.getPtrTy(), Builder.getInt32Ty()},
      /* IsVarArg */ false);
  Function *WcFunc =
      Function::Create(FuncTy, GlobalVariable::InternalLinkage,
                       "_omp_reduction_inter_warp_copy_func", &M);
  WcFunc->setCallingConv(Config.getRuntimeCC());
  WcFunc->setAttributes(FuncAttrs);
  WcFunc->addParamAttr(0, Attribute::NoUndef);
  WcFunc->addParamAttr(1, Attribute::NoUndef);
  BasicBlock *EntryBB = BasicBlock::Create(M.getContext(), "entry", WcFunc);
  Builder.SetInsertPoint(EntryBB);
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());
 
  // ReduceList: thread local Reduce list.
  // At the stage of the computation when this function is called, partially
  // aggregated values reside in the first lane of every active warp.
  Argument *ReduceListArg = WcFunc->getArg(0);
  // NumWarps: number of warps active in the parallel region.  This could
  // be smaller than 32 (max warps in a CTA) for partial block reduction.
  Argument *NumWarpsArg = WcFunc->getArg(1);
 
  // This array is used as a medium to transfer, one reduce element at a time,
  // the data from the first lane of every warp to lanes in the first warp
  // in order to perform the final step of a reduction in a parallel region
  // (reduction across warps).  The array is placed in NVPTX __shared__ memory
  // for reduced latency, as well as to have a distinct copy for concurrently
  // executing target regions.  The array is declared with common linkage so
  // as to be shared across compilation units.
  StringRef TransferMediumName =
      "__openmp_nvptx_data_transfer_temporary_storage";
  GlobalVariable *TransferMedium = M.getGlobalVariable(TransferMediumName);
  unsigned WarpSize = Config.getGridValue().GV_Warp_Size;
  ArrayType *ArrayTy = ArrayType::get(Builder.getInt32Ty(), WarpSize);
  if (!TransferMedium) {
    TransferMedium = new GlobalVariable(
        M, ArrayTy, /*isConstant=*/false, GlobalVariable::WeakAnyLinkage,
        UndefValue::get(ArrayTy), TransferMediumName,
        /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal,
        /*AddressSpace=*/3);
  }
 
  // Get the CUDA thread id of the current OpenMP thread on the GPU.
  Value *GPUThreadID = getGPUThreadID();
  // nvptx_lane_id = nvptx_id % warpsize
  Value *LaneID = getNVPTXLaneID();
  // nvptx_warp_id = nvptx_id / warpsize
  Value *WarpID = getNVPTXWarpID();
 
  InsertPointTy AllocaIP =
      InsertPointTy(Builder.GetInsertBlock(),
                    Builder.GetInsertBlock()->getFirstInsertionPt());
  Type *Arg0Type = ReduceListArg->getType();
  Type *Arg1Type = NumWarpsArg->getType();
  Builder.restoreIP(AllocaIP);
  AllocaInst *ReduceListAlloca = Builder.CreateAlloca(
      Arg0Type, nullptr, ReduceListArg->getName() + ".addr");
  AllocaInst *NumWarpsAlloca =
      Builder.CreateAlloca(Arg1Type, nullptr, NumWarpsArg->getName() + ".addr");
  Value *ReduceListAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ReduceListAlloca, Arg0Type, ReduceListAlloca->getName() + ".ascast");
  Value *NumWarpsAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      NumWarpsAlloca, Builder.getPtrTy(0),
      NumWarpsAlloca->getName() + ".ascast");
  Builder.CreateStore(ReduceListArg, ReduceListAddrCast);
  Builder.CreateStore(NumWarpsArg, NumWarpsAddrCast);
  AllocaIP = getInsertPointAfterInstr(NumWarpsAlloca);
  InsertPointTy CodeGenIP =
      getInsertPointAfterInstr(&Builder.GetInsertBlock()->back());
  Builder.restoreIP(CodeGenIP);
 
  Value *ReduceList =
      Builder.CreateLoad(Builder.getPtrTy(), ReduceListAddrCast);
 
  for (auto En : enumerate(ReductionInfos)) {
    //
    // Warp master copies reduce element to transfer medium in __shared__
    // memory.
    //
    const ReductionInfo &RI = En.value();
    bool IsByRefElem = !IsByRef.empty() && IsByRef[En.index()];
    unsigned RealTySize = M.getDataLayout().getTypeAllocSize(
        IsByRefElem ? RI.ByRefElementType : RI.ElementType);
    for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /= 2) {
      Type *CType = Builder.getIntNTy(TySize * 8);
 
      unsigned NumIters = RealTySize / TySize;
      if (NumIters == 0)
        continue;
      Value *Cnt = nullptr;
      Value *CntAddr = nullptr;
      BasicBlock *PrecondBB = nullptr;
      BasicBlock *ExitBB = nullptr;
      if (NumIters > 1) {
        CodeGenIP = Builder.saveIP();
        Builder.restoreIP(AllocaIP);
        CntAddr =
            Builder.CreateAlloca(Builder.getInt32Ty(), nullptr, ".cnt.addr");
 
        CntAddr = Builder.CreateAddrSpaceCast(CntAddr, Builder.getPtrTy(),
                                              CntAddr->getName() + ".ascast");
        Builder.restoreIP(CodeGenIP);
        Builder.CreateStore(Constant::getNullValue(Builder.getInt32Ty()),
                            CntAddr,
                            /*Volatile=*/false);
        PrecondBB = BasicBlock::Create(Ctx, "precond");
        ExitBB = BasicBlock::Create(Ctx, "exit");
        BasicBlock *BodyBB = BasicBlock::Create(Ctx, "body");
        emitBlock(PrecondBB, Builder.GetInsertBlock()->getParent());
        Cnt = Builder.CreateLoad(Builder.getInt32Ty(), CntAddr,
                                 /*Volatile=*/false);
        Value *Cmp = Builder.CreateICmpULT(
            Cnt, ConstantInt::get(Builder.getInt32Ty(), NumIters));
        Builder.CreateCondBr(Cmp, BodyBB, ExitBB);
        emitBlock(BodyBB, Builder.GetInsertBlock()->getParent());
      }
 
      // kmpc_barrier.
      InsertPointOrErrorTy BarrierIP1 =
          createBarrier(LocationDescription(Builder.saveIP(), DebugLoc()),
                        omp::Directive::OMPD_unknown,
                        /* ForceSimpleCall */ false,
                        /* CheckCancelFlag */ true);
      if (!BarrierIP1)
        return BarrierIP1.takeError();
      BasicBlock *ThenBB = BasicBlock::Create(Ctx, "then");
      BasicBlock *ElseBB = BasicBlock::Create(Ctx, "else");
      BasicBlock *MergeBB = BasicBlock::Create(Ctx, "ifcont");
 
      // if (lane_id  == 0)
      Value *IsWarpMaster = Builder.CreateIsNull(LaneID, "warp_master");
      Builder.CreateCondBr(IsWarpMaster, ThenBB, ElseBB);
      emitBlock(ThenBB, Builder.GetInsertBlock()->getParent());
 
      // Reduce element = LocalReduceList[i]
      auto *RedListArrayTy =
          ArrayType::get(Builder.getPtrTy(), ReductionInfos.size());
      Type *IndexTy = Builder.getIndexTy(
          M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
      Value *ElemPtrPtr =
          Builder.CreateInBoundsGEP(RedListArrayTy, ReduceList,
                                    {ConstantInt::get(IndexTy, 0),
                                     ConstantInt::get(IndexTy, En.index())});
      // elemptr = ((CopyType*)(elemptrptr)) + I
      Value *ElemPtr = Builder.CreateLoad(Builder.getPtrTy(), ElemPtrPtr);
 
      if (IsByRefElem && RI.DataPtrPtrGen) {
        InsertPointOrErrorTy GenRes =
            RI.DataPtrPtrGen(Builder.saveIP(), ElemPtr, ElemPtr);
 
        if (!GenRes)
          return GenRes.takeError();
 
        ElemPtr = Builder.CreateLoad(Builder.getPtrTy(), ElemPtr);
      }
 
      if (NumIters > 1)
        ElemPtr = Builder.CreateGEP(Builder.getInt32Ty(), ElemPtr, Cnt);
 
      // Get pointer to location in transfer medium.
      // MediumPtr = &medium[warp_id]
      Value *MediumPtr = Builder.CreateInBoundsGEP(
          ArrayTy, TransferMedium, {Builder.getInt64(0), WarpID});
      // elem = *elemptr
      //*MediumPtr = elem
      Value *Elem = Builder.CreateLoad(CType, ElemPtr);
      // Store the source element value to the dest element address.
      Builder.CreateStore(Elem, MediumPtr,
                          /*IsVolatile*/ true);
      Builder.CreateBr(MergeBB);
 
      // else
      emitBlock(ElseBB, Builder.GetInsertBlock()->getParent());
      Builder.CreateBr(MergeBB);
 
      // endif
      emitBlock(MergeBB, Builder.GetInsertBlock()->getParent());
      InsertPointOrErrorTy BarrierIP2 =
          createBarrier(LocationDescription(Builder.saveIP(), DebugLoc()),
                        omp::Directive::OMPD_unknown,
                        /* ForceSimpleCall */ false,
                        /* CheckCancelFlag */ true);
      if (!BarrierIP2)
        return BarrierIP2.takeError();
 
      // Warp 0 copies reduce element from transfer medium
      BasicBlock *W0ThenBB = BasicBlock::Create(Ctx, "then");
      BasicBlock *W0ElseBB = BasicBlock::Create(Ctx, "else");
      BasicBlock *W0MergeBB = BasicBlock::Create(Ctx, "ifcont");
 
      Value *NumWarpsVal =
          Builder.CreateLoad(Builder.getInt32Ty(), NumWarpsAddrCast);
      // Up to 32 threads in warp 0 are active.
      Value *IsActiveThread =
          Builder.CreateICmpULT(GPUThreadID, NumWarpsVal, "is_active_thread");
      Builder.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB);
 
      emitBlock(W0ThenBB, Builder.GetInsertBlock()->getParent());
 
      // SecMediumPtr = &medium[tid]
      // SrcMediumVal = *SrcMediumPtr
      Value *SrcMediumPtrVal = Builder.CreateInBoundsGEP(
          ArrayTy, TransferMedium, {Builder.getInt64(0), GPUThreadID});
      // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I
      Value *TargetElemPtrPtr =
          Builder.CreateInBoundsGEP(RedListArrayTy, ReduceList,
                                    {ConstantInt::get(IndexTy, 0),
                                     ConstantInt::get(IndexTy, En.index())});
      Value *TargetElemPtrVal =
          Builder.CreateLoad(Builder.getPtrTy(), TargetElemPtrPtr);
      Value *TargetElemPtr = TargetElemPtrVal;
 
      if (IsByRefElem && RI.DataPtrPtrGen) {
        InsertPointOrErrorTy GenRes =
            RI.DataPtrPtrGen(Builder.saveIP(), TargetElemPtr, TargetElemPtr);
 
        if (!GenRes)
          return GenRes.takeError();
 
        TargetElemPtr = Builder.CreateLoad(Builder.getPtrTy(), TargetElemPtr);
      }
 
      if (NumIters > 1)
        TargetElemPtr =
            Builder.CreateGEP(Builder.getInt32Ty(), TargetElemPtr, Cnt);
 
      // *TargetElemPtr = SrcMediumVal;
      Value *SrcMediumValue =
          Builder.CreateLoad(CType, SrcMediumPtrVal, /*IsVolatile*/ true);
      Builder.CreateStore(SrcMediumValue, TargetElemPtr);
      Builder.CreateBr(W0MergeBB);
 
      emitBlock(W0ElseBB, Builder.GetInsertBlock()->getParent());
      Builder.CreateBr(W0MergeBB);
 
      emitBlock(W0MergeBB, Builder.GetInsertBlock()->getParent());
 
      if (NumIters > 1) {
        Cnt = Builder.CreateNSWAdd(
            Cnt, ConstantInt::get(Builder.getInt32Ty(), /*V=*/1));
        Builder.CreateStore(Cnt, CntAddr, /*Volatile=*/false);
 
        auto *CurFn = Builder.GetInsertBlock()->getParent();
        emitBranch(PrecondBB);
        emitBlock(ExitBB, CurFn);
      }
      RealTySize %= TySize;
    }
  }
 
  Builder.CreateRetVoid();
 
  return WcFunc;
}
 
Expected<Function *> OpenMPIRBuilder::emitShuffleAndReduceFunction(
    ArrayRef<ReductionInfo> ReductionInfos, Function *ReduceFn,
    AttributeList FuncAttrs, ArrayRef<bool> IsByRef) {
  LLVMContext &Ctx = M.getContext();
  IRBuilder<>::InsertPointGuard IPG(Builder);
  FunctionType *FuncTy =
      FunctionType::get(Builder.getVoidTy(),
                        {Builder.getPtrTy(), Builder.getInt16Ty(),
                         Builder.getInt16Ty(), Builder.getInt16Ty()},
                        /* IsVarArg */ false);
  Function *SarFunc =
      Function::Create(FuncTy, GlobalVariable::InternalLinkage,
                       "_omp_reduction_shuffle_and_reduce_func", &M);
  SarFunc->setCallingConv(Config.getRuntimeCC());
  SarFunc->setAttributes(FuncAttrs);
  SarFunc->addParamAttr(0, Attribute::NoUndef);
  SarFunc->addParamAttr(1, Attribute::NoUndef);
  SarFunc->addParamAttr(2, Attribute::NoUndef);
  SarFunc->addParamAttr(3, Attribute::NoUndef);
  SarFunc->addParamAttr(1, Attribute::SExt);
  SarFunc->addParamAttr(2, Attribute::SExt);
  SarFunc->addParamAttr(3, Attribute::SExt);
  BasicBlock *EntryBB = BasicBlock::Create(M.getContext(), "entry", SarFunc);
  Builder.SetInsertPoint(EntryBB);
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());
 
  // Thread local Reduce list used to host the values of data to be reduced.
  Argument *ReduceListArg = SarFunc->getArg(0);
  // Current lane id; could be logical.
  Argument *LaneIDArg = SarFunc->getArg(1);
  // Offset of the remote source lane relative to the current lane.
  Argument *RemoteLaneOffsetArg = SarFunc->getArg(2);
  // Algorithm version.  This is expected to be known at compile time.
  Argument *AlgoVerArg = SarFunc->getArg(3);
 
  Type *ReduceListArgType = ReduceListArg->getType();
  Type *LaneIDArgType = LaneIDArg->getType();
  Type *LaneIDArgPtrType = Builder.getPtrTy(0);
  Value *ReduceListAlloca = Builder.CreateAlloca(
      ReduceListArgType, nullptr, ReduceListArg->getName() + ".addr");
  Value *LaneIdAlloca = Builder.CreateAlloca(LaneIDArgType, nullptr,
                                             LaneIDArg->getName() + ".addr");
  Value *RemoteLaneOffsetAlloca = Builder.CreateAlloca(
      LaneIDArgType, nullptr, RemoteLaneOffsetArg->getName() + ".addr");
  Value *AlgoVerAlloca = Builder.CreateAlloca(LaneIDArgType, nullptr,
                                              AlgoVerArg->getName() + ".addr");
  ArrayType *RedListArrayTy =
      ArrayType::get(Builder.getPtrTy(), ReductionInfos.size());
 
  // Create a local thread-private variable to host the Reduce list
  // from a remote lane.
  Instruction *RemoteReductionListAlloca = Builder.CreateAlloca(
      RedListArrayTy, nullptr, ".omp.reduction.remote_reduce_list");
 
  Value *ReduceListAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ReduceListAlloca, ReduceListArgType,
      ReduceListAlloca->getName() + ".ascast");
  Value *LaneIdAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      LaneIdAlloca, LaneIDArgPtrType, LaneIdAlloca->getName() + ".ascast");
  Value *RemoteLaneOffsetAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      RemoteLaneOffsetAlloca, LaneIDArgPtrType,
      RemoteLaneOffsetAlloca->getName() + ".ascast");
  Value *AlgoVerAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      AlgoVerAlloca, LaneIDArgPtrType, AlgoVerAlloca->getName() + ".ascast");
  Value *RemoteListAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      RemoteReductionListAlloca, Builder.getPtrTy(),
      RemoteReductionListAlloca->getName() + ".ascast");
 
  Builder.CreateStore(ReduceListArg, ReduceListAddrCast);
  Builder.CreateStore(LaneIDArg, LaneIdAddrCast);
  Builder.CreateStore(RemoteLaneOffsetArg, RemoteLaneOffsetAddrCast);
  Builder.CreateStore(AlgoVerArg, AlgoVerAddrCast);
 
  Value *ReduceList = Builder.CreateLoad(ReduceListArgType, ReduceListAddrCast);
  Value *LaneId = Builder.CreateLoad(LaneIDArgType, LaneIdAddrCast);
  Value *RemoteLaneOffset =
      Builder.CreateLoad(LaneIDArgType, RemoteLaneOffsetAddrCast);
  Value *AlgoVer = Builder.CreateLoad(LaneIDArgType, AlgoVerAddrCast);
 
  InsertPointTy AllocaIP = getInsertPointAfterInstr(RemoteReductionListAlloca);
 
  // This loop iterates through the list of reduce elements and copies,
  // element by element, from a remote lane in the warp to RemoteReduceList,
  // hosted on the thread's stack.
  Error EmitRedLsCpRes = emitReductionListCopy(
      AllocaIP, CopyAction::RemoteLaneToThread, RedListArrayTy, ReductionInfos,
      ReduceList, RemoteListAddrCast, IsByRef,
      {RemoteLaneOffset, nullptr, nullptr});
 
  if (EmitRedLsCpRes)
    return EmitRedLsCpRes;
 
  // The actions to be performed on the Remote Reduce list is dependent
  // on the algorithm version.
  //
  //  if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 &&
  //  LaneId % 2 == 0 && Offset > 0):
  //    do the reduction value aggregation
  //
  //  The thread local variable Reduce list is mutated in place to host the
  //  reduced data, which is the aggregated value produced from local and
  //  remote lanes.
  //
  //  Note that AlgoVer is expected to be a constant integer known at compile
  //  time.
  //  When AlgoVer==0, the first conjunction evaluates to true, making
  //    the entire predicate true during compile time.
  //  When AlgoVer==1, the second conjunction has only the second part to be
  //    evaluated during runtime.  Other conjunctions evaluates to false
  //    during compile time.
  //  When AlgoVer==2, the third conjunction has only the second part to be
  //    evaluated during runtime.  Other conjunctions evaluates to false
  //    during compile time.
  Value *CondAlgo0 = Builder.CreateIsNull(AlgoVer);
  Value *Algo1 = Builder.CreateICmpEQ(AlgoVer, Builder.getInt16(1));
  Value *LaneComp = Builder.CreateICmpULT(LaneId, RemoteLaneOffset);
  Value *CondAlgo1 = Builder.CreateAnd(Algo1, LaneComp);
  Value *Algo2 = Builder.CreateICmpEQ(AlgoVer, Builder.getInt16(2));
  Value *LaneIdAnd1 = Builder.CreateAnd(LaneId, Builder.getInt16(1));
  Value *LaneIdComp = Builder.CreateIsNull(LaneIdAnd1);
  Value *Algo2AndLaneIdComp = Builder.CreateAnd(Algo2, LaneIdComp);
  Value *RemoteOffsetComp =
      Builder.CreateICmpSGT(RemoteLaneOffset, Builder.getInt16(0));
  Value *CondAlgo2 = Builder.CreateAnd(Algo2AndLaneIdComp, RemoteOffsetComp);
  Value *CA0OrCA1 = Builder.CreateOr(CondAlgo0, CondAlgo1);
  Value *CondReduce = Builder.CreateOr(CA0OrCA1, CondAlgo2);
 
  BasicBlock *ThenBB = BasicBlock::Create(Ctx, "then");
  BasicBlock *ElseBB = BasicBlock::Create(Ctx, "else");
  BasicBlock *MergeBB = BasicBlock::Create(Ctx, "ifcont");
 
  Builder.CreateCondBr(CondReduce, ThenBB, ElseBB);
  emitBlock(ThenBB, Builder.GetInsertBlock()->getParent());
  Value *LocalReduceListPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ReduceList, Builder.getPtrTy());
  Value *RemoteReduceListPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
      RemoteListAddrCast, Builder.getPtrTy());
  createRuntimeFunctionCall(ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr})
      ->addFnAttr(Attribute::NoUnwind);
  Builder.CreateBr(MergeBB);
 
  emitBlock(ElseBB, Builder.GetInsertBlock()->getParent());
  Builder.CreateBr(MergeBB);
 
  emitBlock(MergeBB, Builder.GetInsertBlock()->getParent());
 
  // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local
  // Reduce list.
  Algo1 = Builder.CreateICmpEQ(AlgoVer, Builder.getInt16(1));
  Value *LaneIdGtOffset = Builder.CreateICmpUGE(LaneId, RemoteLaneOffset);
  Value *CondCopy = Builder.CreateAnd(Algo1, LaneIdGtOffset);
 
  BasicBlock *CpyThenBB = BasicBlock::Create(Ctx, "then");
  BasicBlock *CpyElseBB = BasicBlock::Create(Ctx, "else");
  BasicBlock *CpyMergeBB = BasicBlock::Create(Ctx, "ifcont");
  Builder.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB);
 
  emitBlock(CpyThenBB, Builder.GetInsertBlock()->getParent());
 
  EmitRedLsCpRes = emitReductionListCopy(
      AllocaIP, CopyAction::ThreadCopy, RedListArrayTy, ReductionInfos,
      RemoteListAddrCast, ReduceList, IsByRef);
 
  if (EmitRedLsCpRes)
    return EmitRedLsCpRes;
 
  Builder.CreateBr(CpyMergeBB);
 
  emitBlock(CpyElseBB, Builder.GetInsertBlock()->getParent());
  Builder.CreateBr(CpyMergeBB);
 
  emitBlock(CpyMergeBB, Builder.GetInsertBlock()->getParent());
 
  Builder.CreateRetVoid();
 
  return SarFunc;
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::generateReductionDescriptor(
    Value *DescriptorAddr, Value *DataPtr, Value *SrcDescriptorAddr,
    Type *DescriptorType,
    function_ref<InsertPointOrErrorTy(InsertPointTy, Value *, Value *&)>
        DataPtrPtrGen) {
 
  // Copy the source descriptor to preserve all metadata (rank, extents,
  // strides, etc.)
  Value *DescriptorSize =
      Builder.getInt64(M.getDataLayout().getTypeStoreSize(DescriptorType));
  Builder.CreateMemCpy(
      DescriptorAddr, M.getDataLayout().getPrefTypeAlign(DescriptorType),
      SrcDescriptorAddr, M.getDataLayout().getPrefTypeAlign(DescriptorType),
      DescriptorSize);
 
  // Update the base pointer field to point to the local shuffled data
  Value *DataPtrField;
  InsertPointOrErrorTy GenResult =
      DataPtrPtrGen(Builder.saveIP(), DescriptorAddr, DataPtrField);
 
  if (!GenResult)
    return GenResult.takeError();
 
  Builder.CreateStore(Builder.CreatePointerBitCastOrAddrSpaceCast(
                          DataPtr, Builder.getPtrTy(), ".ascast"),
                      DataPtrField);
 
  return Builder.saveIP();
}
 
Expected<Value *> OpenMPIRBuilder::createReductionDescriptorCopy(
    InsertPointTy AllocaIP, const ReductionInfo &RI, Value *DataPtr,
    Value *SrcDescriptorAddr, Type *DescriptorPtrTy, const Twine &Name) {
  InsertPointTy OldIP = Builder.saveIP();
  Builder.restoreIP(AllocaIP);
 
  AllocaInst *DescriptorAlloca =
      Builder.CreateAlloca(RI.ByRefAllocatedType, nullptr, Name);
  DescriptorAlloca->setAlignment(
      M.getDataLayout().getPrefTypeAlign(RI.ByRefAllocatedType));
  Value *DescriptorAddr = Builder.CreatePointerBitCastOrAddrSpaceCast(
      DescriptorAlloca, DescriptorPtrTy,
      DescriptorAlloca->getName() + ".ascast");
 
  Builder.restoreIP(OldIP);
 
  InsertPointOrErrorTy GenResult =
      generateReductionDescriptor(DescriptorAddr, DataPtr, SrcDescriptorAddr,
                                  RI.ByRefAllocatedType, RI.DataPtrPtrGen);
  if (!GenResult)
    return GenResult.takeError();
 
  return DescriptorAddr;
}
 
Expected<Function *> OpenMPIRBuilder::emitListToGlobalCopyFunction(
    ArrayRef<ReductionInfo> ReductionInfos, Type *ReductionsBufferTy,
    AttributeList FuncAttrs, ArrayRef<bool> IsByRef) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  LLVMContext &Ctx = M.getContext();
  FunctionType *FuncTy = FunctionType::get(
      Builder.getVoidTy(),
      {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()},
      /* IsVarArg */ false);
  Function *LtGCFunc =
      Function::Create(FuncTy, GlobalVariable::InternalLinkage,
                       "_omp_reduction_list_to_global_copy_func", &M);
  LtGCFunc->setAttributes(FuncAttrs);
  LtGCFunc->addParamAttr(0, Attribute::NoUndef);
  LtGCFunc->addParamAttr(1, Attribute::NoUndef);
  LtGCFunc->addParamAttr(2, Attribute::NoUndef);
 
  BasicBlock *EntryBlock = BasicBlock::Create(Ctx, "entry", LtGCFunc);
  Builder.SetInsertPoint(EntryBlock);
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());
 
  // Buffer: global reduction buffer.
  Argument *BufferArg = LtGCFunc->getArg(0);
  // Idx: index of the buffer.
  Argument *IdxArg = LtGCFunc->getArg(1);
  // ReduceList: thread local Reduce list.
  Argument *ReduceListArg = LtGCFunc->getArg(2);
 
  Value *BufferArgAlloca = Builder.CreateAlloca(Builder.getPtrTy(), nullptr,
                                                BufferArg->getName() + ".addr");
  Value *IdxArgAlloca = Builder.CreateAlloca(Builder.getInt32Ty(), nullptr,
                                             IdxArg->getName() + ".addr");
  Value *ReduceListArgAlloca = Builder.CreateAlloca(
      Builder.getPtrTy(), nullptr, ReduceListArg->getName() + ".addr");
  Value *BufferArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      BufferArgAlloca, Builder.getPtrTy(),
      BufferArgAlloca->getName() + ".ascast");
  Value *IdxArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      IdxArgAlloca, Builder.getPtrTy(), IdxArgAlloca->getName() + ".ascast");
  Value *ReduceListArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ReduceListArgAlloca, Builder.getPtrTy(),
      ReduceListArgAlloca->getName() + ".ascast");
 
  Builder.CreateStore(BufferArg, BufferArgAddrCast);
  Builder.CreateStore(IdxArg, IdxArgAddrCast);
  Builder.CreateStore(ReduceListArg, ReduceListArgAddrCast);
 
  Value *LocalReduceList =
      Builder.CreateLoad(Builder.getPtrTy(), ReduceListArgAddrCast);
  Value *BufferArgVal =
      Builder.CreateLoad(Builder.getPtrTy(), BufferArgAddrCast);
  Value *Idxs[] = {Builder.CreateLoad(Builder.getInt32Ty(), IdxArgAddrCast)};
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  for (auto En : enumerate(ReductionInfos)) {
    const ReductionInfo &RI = En.value();
    auto *RedListArrayTy =
        ArrayType::get(Builder.getPtrTy(), ReductionInfos.size());
    // Reduce element = LocalReduceList[i]
    Value *ElemPtrPtr = Builder.CreateInBoundsGEP(
        RedListArrayTy, LocalReduceList,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
    // elemptr = ((CopyType*)(elemptrptr)) + I
    Value *ElemPtr = Builder.CreateLoad(Builder.getPtrTy(), ElemPtrPtr);
 
    // Global = Buffer.VD[Idx];
    Value *BufferVD =
        Builder.CreateInBoundsGEP(ReductionsBufferTy, BufferArgVal, Idxs);
    Value *GlobVal = Builder.CreateConstInBoundsGEP2_32(
        ReductionsBufferTy, BufferVD, 0, En.index());
 
    switch (RI.EvaluationKind) {
    case EvalKind::Scalar: {
      Value *TargetElement;
 
      if (IsByRef.empty() || !IsByRef[En.index()]) {
        TargetElement = Builder.CreateLoad(RI.ElementType, ElemPtr);
      } else {
        if (RI.DataPtrPtrGen) {
          InsertPointOrErrorTy GenResult =
              RI.DataPtrPtrGen(Builder.saveIP(), ElemPtr, ElemPtr);
 
          if (!GenResult)
            return GenResult.takeError();
 
          ElemPtr = Builder.CreateLoad(Builder.getPtrTy(), ElemPtr);
        }
        TargetElement = Builder.CreateLoad(RI.ByRefElementType, ElemPtr);
      }
 
      Builder.CreateStore(TargetElement, GlobVal);
      break;
    }
    case EvalKind::Complex: {
      Value *SrcRealPtr = Builder.CreateConstInBoundsGEP2_32(
          RI.ElementType, ElemPtr, 0, 0, ".realp");
      Value *SrcReal = Builder.CreateLoad(
          RI.ElementType->getStructElementType(0), SrcRealPtr, ".real");
      Value *SrcImgPtr = Builder.CreateConstInBoundsGEP2_32(
          RI.ElementType, ElemPtr, 0, 1, ".imagp");
      Value *SrcImg = Builder.CreateLoad(
          RI.ElementType->getStructElementType(1), SrcImgPtr, ".imag");
 
      Value *DestRealPtr = Builder.CreateConstInBoundsGEP2_32(
          RI.ElementType, GlobVal, 0, 0, ".realp");
      Value *DestImgPtr = Builder.CreateConstInBoundsGEP2_32(
          RI.ElementType, GlobVal, 0, 1, ".imagp");
      Builder.CreateStore(SrcReal, DestRealPtr);
      Builder.CreateStore(SrcImg, DestImgPtr);
      break;
    }
    case EvalKind::Aggregate: {
      Value *SizeVal =
          Builder.getInt64(M.getDataLayout().getTypeStoreSize(RI.ElementType));
      Builder.CreateMemCpy(
          GlobVal, M.getDataLayout().getPrefTypeAlign(RI.ElementType), ElemPtr,
          M.getDataLayout().getPrefTypeAlign(RI.ElementType), SizeVal, false);
      break;
    }
    }
  }
 
  Builder.CreateRetVoid();
  return LtGCFunc;
}
 
Expected<Function *> OpenMPIRBuilder::emitListToGlobalReduceFunction(
    ArrayRef<ReductionInfo> ReductionInfos, Function *ReduceFn,
    Type *ReductionsBufferTy, AttributeList FuncAttrs, ArrayRef<bool> IsByRef) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  LLVMContext &Ctx = M.getContext();
  FunctionType *FuncTy = FunctionType::get(
      Builder.getVoidTy(),
      {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()},
      /* IsVarArg */ false);
  Function *LtGRFunc =
      Function::Create(FuncTy, GlobalVariable::InternalLinkage,
                       "_omp_reduction_list_to_global_reduce_func", &M);
  LtGRFunc->setAttributes(FuncAttrs);
  LtGRFunc->addParamAttr(0, Attribute::NoUndef);
  LtGRFunc->addParamAttr(1, Attribute::NoUndef);
  LtGRFunc->addParamAttr(2, Attribute::NoUndef);
 
  BasicBlock *EntryBlock = BasicBlock::Create(Ctx, "entry", LtGRFunc);
  Builder.SetInsertPoint(EntryBlock);
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());
 
  // Buffer: global reduction buffer.
  Argument *BufferArg = LtGRFunc->getArg(0);
  // Idx: index of the buffer.
  Argument *IdxArg = LtGRFunc->getArg(1);
  // ReduceList: thread local Reduce list.
  Argument *ReduceListArg = LtGRFunc->getArg(2);
 
  Value *BufferArgAlloca = Builder.CreateAlloca(Builder.getPtrTy(), nullptr,
                                                BufferArg->getName() + ".addr");
  Value *IdxArgAlloca = Builder.CreateAlloca(Builder.getInt32Ty(), nullptr,
                                             IdxArg->getName() + ".addr");
  Value *ReduceListArgAlloca = Builder.CreateAlloca(
      Builder.getPtrTy(), nullptr, ReduceListArg->getName() + ".addr");
  auto *RedListArrayTy =
      ArrayType::get(Builder.getPtrTy(), ReductionInfos.size());
 
  // 1. Build a list of reduction variables.
  // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
  Value *LocalReduceList =
      Builder.CreateAlloca(RedListArrayTy, nullptr, ".omp.reduction.red_list");
 
  InsertPointTy AllocaIP{EntryBlock, EntryBlock->begin()};
 
  Value *BufferArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      BufferArgAlloca, Builder.getPtrTy(),
      BufferArgAlloca->getName() + ".ascast");
  Value *IdxArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      IdxArgAlloca, Builder.getPtrTy(), IdxArgAlloca->getName() + ".ascast");
  Value *ReduceListArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ReduceListArgAlloca, Builder.getPtrTy(),
      ReduceListArgAlloca->getName() + ".ascast");
  Value *LocalReduceListAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      LocalReduceList, Builder.getPtrTy(),
      LocalReduceList->getName() + ".ascast");
 
  Builder.CreateStore(BufferArg, BufferArgAddrCast);
  Builder.CreateStore(IdxArg, IdxArgAddrCast);
  Builder.CreateStore(ReduceListArg, ReduceListArgAddrCast);
 
  Value *BufferVal = Builder.CreateLoad(Builder.getPtrTy(), BufferArgAddrCast);
  Value *Idxs[] = {Builder.CreateLoad(Builder.getInt32Ty(), IdxArgAddrCast)};
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  for (auto En : enumerate(ReductionInfos)) {
    const ReductionInfo &RI = En.value();
 
    Value *TargetElementPtrPtr = Builder.CreateInBoundsGEP(
        RedListArrayTy, LocalReduceListAddrCast,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
    Value *BufferVD =
        Builder.CreateInBoundsGEP(ReductionsBufferTy, BufferVal, Idxs);
    // Global = Buffer.VD[Idx];
    Value *GlobValPtr = Builder.CreateConstInBoundsGEP2_32(
        ReductionsBufferTy, BufferVD, 0, En.index());
 
    if (!IsByRef.empty() && IsByRef[En.index()] && RI.DataPtrPtrGen) {
      // Get source descriptor from the reduce list argument
      Value *ReduceList =
          Builder.CreateLoad(Builder.getPtrTy(), ReduceListArgAddrCast);
      Value *SrcElementPtrPtr =
          Builder.CreateInBoundsGEP(RedListArrayTy, ReduceList,
                                    {ConstantInt::get(IndexTy, 0),
                                     ConstantInt::get(IndexTy, En.index())});
      Value *SrcDescriptorAddr =
          Builder.CreateLoad(Builder.getPtrTy(), SrcElementPtrPtr);
 
      // Copy descriptor from source and update base_ptr to global buffer data
      Expected<Value *> ByRefAlloc = createReductionDescriptorCopy(
          AllocaIP, RI, GlobValPtr, SrcDescriptorAddr, Builder.getPtrTy());
      if (!ByRefAlloc)
        return ByRefAlloc.takeError();
 
      Builder.CreateStore(*ByRefAlloc, TargetElementPtrPtr);
    } else {
      Builder.CreateStore(GlobValPtr, TargetElementPtrPtr);
    }
  }
 
  // Call reduce_function(GlobalReduceList, ReduceList)
  Value *ReduceList =
      Builder.CreateLoad(Builder.getPtrTy(), ReduceListArgAddrCast);
  createRuntimeFunctionCall(ReduceFn, {LocalReduceListAddrCast, ReduceList})
      ->addFnAttr(Attribute::NoUnwind);
  Builder.CreateRetVoid();
  return LtGRFunc;
}
 
Expected<Function *> OpenMPIRBuilder::emitGlobalToListCopyFunction(
    ArrayRef<ReductionInfo> ReductionInfos, Type *ReductionsBufferTy,
    AttributeList FuncAttrs, ArrayRef<bool> IsByRef) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  LLVMContext &Ctx = M.getContext();
  FunctionType *FuncTy = FunctionType::get(
      Builder.getVoidTy(),
      {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()},
      /* IsVarArg */ false);
  Function *GtLCFunc =
      Function::Create(FuncTy, GlobalVariable::InternalLinkage,
                       "_omp_reduction_global_to_list_copy_func", &M);
  GtLCFunc->setAttributes(FuncAttrs);
  GtLCFunc->addParamAttr(0, Attribute::NoUndef);
  GtLCFunc->addParamAttr(1, Attribute::NoUndef);
  GtLCFunc->addParamAttr(2, Attribute::NoUndef);
 
  BasicBlock *EntryBlock = BasicBlock::Create(Ctx, "entry", GtLCFunc);
  Builder.SetInsertPoint(EntryBlock);
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());
 
  // Buffer: global reduction buffer.
  Argument *BufferArg = GtLCFunc->getArg(0);
  // Idx: index of the buffer.
  Argument *IdxArg = GtLCFunc->getArg(1);
  // ReduceList: thread local Reduce list.
  Argument *ReduceListArg = GtLCFunc->getArg(2);
 
  Value *BufferArgAlloca = Builder.CreateAlloca(Builder.getPtrTy(), nullptr,
                                                BufferArg->getName() + ".addr");
  Value *IdxArgAlloca = Builder.CreateAlloca(Builder.getInt32Ty(), nullptr,
                                             IdxArg->getName() + ".addr");
  Value *ReduceListArgAlloca = Builder.CreateAlloca(
      Builder.getPtrTy(), nullptr, ReduceListArg->getName() + ".addr");
  Value *BufferArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      BufferArgAlloca, Builder.getPtrTy(),
      BufferArgAlloca->getName() + ".ascast");
  Value *IdxArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      IdxArgAlloca, Builder.getPtrTy(), IdxArgAlloca->getName() + ".ascast");
  Value *ReduceListArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ReduceListArgAlloca, Builder.getPtrTy(),
      ReduceListArgAlloca->getName() + ".ascast");
  Builder.CreateStore(BufferArg, BufferArgAddrCast);
  Builder.CreateStore(IdxArg, IdxArgAddrCast);
  Builder.CreateStore(ReduceListArg, ReduceListArgAddrCast);
 
  Value *LocalReduceList =
      Builder.CreateLoad(Builder.getPtrTy(), ReduceListArgAddrCast);
  Value *BufferVal = Builder.CreateLoad(Builder.getPtrTy(), BufferArgAddrCast);
  Value *Idxs[] = {Builder.CreateLoad(Builder.getInt32Ty(), IdxArgAddrCast)};
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  for (auto En : enumerate(ReductionInfos)) {
    const OpenMPIRBuilder::ReductionInfo &RI = En.value();
    auto *RedListArrayTy =
        ArrayType::get(Builder.getPtrTy(), ReductionInfos.size());
    // Reduce element = LocalReduceList[i]
    Value *ElemPtrPtr = Builder.CreateInBoundsGEP(
        RedListArrayTy, LocalReduceList,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
    // elemptr = ((CopyType*)(elemptrptr)) + I
    Value *ElemPtr = Builder.CreateLoad(Builder.getPtrTy(), ElemPtrPtr);
    // Global = Buffer.VD[Idx];
    Value *BufferVD =
        Builder.CreateInBoundsGEP(ReductionsBufferTy, BufferVal, Idxs);
    Value *GlobValPtr = Builder.CreateConstInBoundsGEP2_32(
        ReductionsBufferTy, BufferVD, 0, En.index());
 
    switch (RI.EvaluationKind) {
    case EvalKind::Scalar: {
      Type *ElemType = RI.ElementType;
 
      if (!IsByRef.empty() && IsByRef[En.index()]) {
        ElemType = RI.ByRefElementType;
        if (RI.DataPtrPtrGen) {
          InsertPointOrErrorTy GenResult =
              RI.DataPtrPtrGen(Builder.saveIP(), ElemPtr, ElemPtr);
 
          if (!GenResult)
            return GenResult.takeError();
 
          ElemPtr = Builder.CreateLoad(Builder.getPtrTy(), ElemPtr);
        }
      }
 
      Value *TargetElement = Builder.CreateLoad(ElemType, GlobValPtr);
      Builder.CreateStore(TargetElement, ElemPtr);
      break;
    }
    case EvalKind::Complex: {
      Value *SrcRealPtr = Builder.CreateConstInBoundsGEP2_32(
          RI.ElementType, GlobValPtr, 0, 0, ".realp");
      Value *SrcReal = Builder.CreateLoad(
          RI.ElementType->getStructElementType(0), SrcRealPtr, ".real");
      Value *SrcImgPtr = Builder.CreateConstInBoundsGEP2_32(
          RI.ElementType, GlobValPtr, 0, 1, ".imagp");
      Value *SrcImg = Builder.CreateLoad(
          RI.ElementType->getStructElementType(1), SrcImgPtr, ".imag");
 
      Value *DestRealPtr = Builder.CreateConstInBoundsGEP2_32(
          RI.ElementType, ElemPtr, 0, 0, ".realp");
      Value *DestImgPtr = Builder.CreateConstInBoundsGEP2_32(
          RI.ElementType, ElemPtr, 0, 1, ".imagp");
      Builder.CreateStore(SrcReal, DestRealPtr);
      Builder.CreateStore(SrcImg, DestImgPtr);
      break;
    }
    case EvalKind::Aggregate: {
      Value *SizeVal =
          Builder.getInt64(M.getDataLayout().getTypeStoreSize(RI.ElementType));
      Builder.CreateMemCpy(
          ElemPtr, M.getDataLayout().getPrefTypeAlign(RI.ElementType),
          GlobValPtr, M.getDataLayout().getPrefTypeAlign(RI.ElementType),
          SizeVal, false);
      break;
    }
    }
  }
 
  Builder.CreateRetVoid();
  return GtLCFunc;
}
 
Expected<Function *> OpenMPIRBuilder::emitGlobalToListReduceFunction(
    ArrayRef<ReductionInfo> ReductionInfos, Function *ReduceFn,
    Type *ReductionsBufferTy, AttributeList FuncAttrs, ArrayRef<bool> IsByRef) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  LLVMContext &Ctx = M.getContext();
  auto *FuncTy = FunctionType::get(
      Builder.getVoidTy(),
      {Builder.getPtrTy(), Builder.getInt32Ty(), Builder.getPtrTy()},
      /* IsVarArg */ false);
  Function *GtLRFunc =
      Function::Create(FuncTy, GlobalVariable::InternalLinkage,
                       "_omp_reduction_global_to_list_reduce_func", &M);
  GtLRFunc->setAttributes(FuncAttrs);
  GtLRFunc->addParamAttr(0, Attribute::NoUndef);
  GtLRFunc->addParamAttr(1, Attribute::NoUndef);
  GtLRFunc->addParamAttr(2, Attribute::NoUndef);
 
  BasicBlock *EntryBlock = BasicBlock::Create(Ctx, "entry", GtLRFunc);
  Builder.SetInsertPoint(EntryBlock);
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());
 
  // Buffer: global reduction buffer.
  Argument *BufferArg = GtLRFunc->getArg(0);
  // Idx: index of the buffer.
  Argument *IdxArg = GtLRFunc->getArg(1);
  // ReduceList: thread local Reduce list.
  Argument *ReduceListArg = GtLRFunc->getArg(2);
 
  Value *BufferArgAlloca = Builder.CreateAlloca(Builder.getPtrTy(), nullptr,
                                                BufferArg->getName() + ".addr");
  Value *IdxArgAlloca = Builder.CreateAlloca(Builder.getInt32Ty(), nullptr,
                                             IdxArg->getName() + ".addr");
  Value *ReduceListArgAlloca = Builder.CreateAlloca(
      Builder.getPtrTy(), nullptr, ReduceListArg->getName() + ".addr");
  ArrayType *RedListArrayTy =
      ArrayType::get(Builder.getPtrTy(), ReductionInfos.size());
 
  // 1. Build a list of reduction variables.
  // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
  Value *LocalReduceList =
      Builder.CreateAlloca(RedListArrayTy, nullptr, ".omp.reduction.red_list");
 
  InsertPointTy AllocaIP{EntryBlock, EntryBlock->begin()};
 
  Value *BufferArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      BufferArgAlloca, Builder.getPtrTy(),
      BufferArgAlloca->getName() + ".ascast");
  Value *IdxArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      IdxArgAlloca, Builder.getPtrTy(), IdxArgAlloca->getName() + ".ascast");
  Value *ReduceListArgAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ReduceListArgAlloca, Builder.getPtrTy(),
      ReduceListArgAlloca->getName() + ".ascast");
  Value *ReductionList = Builder.CreatePointerBitCastOrAddrSpaceCast(
      LocalReduceList, Builder.getPtrTy(),
      LocalReduceList->getName() + ".ascast");
 
  Builder.CreateStore(BufferArg, BufferArgAddrCast);
  Builder.CreateStore(IdxArg, IdxArgAddrCast);
  Builder.CreateStore(ReduceListArg, ReduceListArgAddrCast);
 
  Value *BufferVal = Builder.CreateLoad(Builder.getPtrTy(), BufferArgAddrCast);
  Value *Idxs[] = {Builder.CreateLoad(Builder.getInt32Ty(), IdxArgAddrCast)};
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  for (auto En : enumerate(ReductionInfos)) {
    const ReductionInfo &RI = En.value();
 
    Value *TargetElementPtrPtr = Builder.CreateInBoundsGEP(
        RedListArrayTy, ReductionList,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
    // Global = Buffer.VD[Idx];
    Value *BufferVD =
        Builder.CreateInBoundsGEP(ReductionsBufferTy, BufferVal, Idxs);
    Value *GlobValPtr = Builder.CreateConstInBoundsGEP2_32(
        ReductionsBufferTy, BufferVD, 0, En.index());
 
    if (!IsByRef.empty() && IsByRef[En.index()] && RI.DataPtrPtrGen) {
      // Get source descriptor from the reduce list
      Value *ReduceListVal =
          Builder.CreateLoad(Builder.getPtrTy(), ReduceListArgAddrCast);
      Value *SrcElementPtrPtr =
          Builder.CreateInBoundsGEP(RedListArrayTy, ReduceListVal,
                                    {ConstantInt::get(IndexTy, 0),
                                     ConstantInt::get(IndexTy, En.index())});
      Value *SrcDescriptorAddr =
          Builder.CreateLoad(Builder.getPtrTy(), SrcElementPtrPtr);
 
      // Copy descriptor from source and update base_ptr to global buffer data
      Expected<Value *> ByRefAlloc = createReductionDescriptorCopy(
          AllocaIP, RI, GlobValPtr, SrcDescriptorAddr, Builder.getPtrTy());
      if (!ByRefAlloc)
        return ByRefAlloc.takeError();
 
      Builder.CreateStore(*ByRefAlloc, TargetElementPtrPtr);
    } else {
      Builder.CreateStore(GlobValPtr, TargetElementPtrPtr);
    }
  }
 
  // Call reduce_function(ReduceList, GlobalReduceList)
  Value *ReduceList =
      Builder.CreateLoad(Builder.getPtrTy(), ReduceListArgAddrCast);
  createRuntimeFunctionCall(ReduceFn, {ReduceList, ReductionList})
      ->addFnAttr(Attribute::NoUnwind);
  Builder.CreateRetVoid();
  return GtLRFunc;
}
 
std::string OpenMPIRBuilder::getReductionFuncName(StringRef Name) const {
  std::string Suffix =
      createPlatformSpecificName({"omp", "reduction", "reduction_func"});
  return (Name + Suffix).str();
}
 
Expected<Function *> OpenMPIRBuilder::createReductionFunction(
    StringRef ReducerName, ArrayRef<ReductionInfo> ReductionInfos,
    ArrayRef<bool> IsByRef, ReductionGenCBKind ReductionGenCBKind,
    AttributeList FuncAttrs) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  auto *FuncTy = FunctionType::get(Builder.getVoidTy(),
                                   {Builder.getPtrTy(), Builder.getPtrTy()},
                                   /* IsVarArg */ false);
  std::string Name = getReductionFuncName(ReducerName);
  Function *ReductionFunc =
      Function::Create(FuncTy, GlobalVariable::InternalLinkage, Name, &M);
  ReductionFunc->setCallingConv(Config.getRuntimeCC());
  ReductionFunc->setAttributes(FuncAttrs);
  ReductionFunc->addParamAttr(0, Attribute::NoUndef);
  ReductionFunc->addParamAttr(1, Attribute::NoUndef);
  BasicBlock *EntryBB =
      BasicBlock::Create(M.getContext(), "entry", ReductionFunc);
  Builder.SetInsertPoint(EntryBB);
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());
 
  // Need to alloca memory here and deal with the pointers before getting
  // LHS/RHS pointers out
  Value *LHSArrayPtr = nullptr;
  Value *RHSArrayPtr = nullptr;
  Argument *Arg0 = ReductionFunc->getArg(0);
  Argument *Arg1 = ReductionFunc->getArg(1);
  Type *Arg0Type = Arg0->getType();
  Type *Arg1Type = Arg1->getType();
 
  Value *LHSAlloca =
      Builder.CreateAlloca(Arg0Type, nullptr, Arg0->getName() + ".addr");
  Value *RHSAlloca =
      Builder.CreateAlloca(Arg1Type, nullptr, Arg1->getName() + ".addr");
  Value *LHSAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      LHSAlloca, Arg0Type, LHSAlloca->getName() + ".ascast");
  Value *RHSAddrCast = Builder.CreatePointerBitCastOrAddrSpaceCast(
      RHSAlloca, Arg1Type, RHSAlloca->getName() + ".ascast");
  Builder.CreateStore(Arg0, LHSAddrCast);
  Builder.CreateStore(Arg1, RHSAddrCast);
  LHSArrayPtr = Builder.CreateLoad(Arg0Type, LHSAddrCast);
  RHSArrayPtr = Builder.CreateLoad(Arg1Type, RHSAddrCast);
 
  Type *RedArrayTy = ArrayType::get(Builder.getPtrTy(), ReductionInfos.size());
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  SmallVector<Value *> LHSPtrs, RHSPtrs;
  for (auto En : enumerate(ReductionInfos)) {
    const ReductionInfo &RI = En.value();
    Value *RHSI8PtrPtr = Builder.CreateInBoundsGEP(
        RedArrayTy, RHSArrayPtr,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
    Value *RHSI8Ptr = Builder.CreateLoad(Builder.getPtrTy(), RHSI8PtrPtr);
    Value *RHSPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
        RHSI8Ptr, RI.PrivateVariable->getType(),
        RHSI8Ptr->getName() + ".ascast");
 
    Value *LHSI8PtrPtr = Builder.CreateInBoundsGEP(
        RedArrayTy, LHSArrayPtr,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
    Value *LHSI8Ptr = Builder.CreateLoad(Builder.getPtrTy(), LHSI8PtrPtr);
    Value *LHSPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
        LHSI8Ptr, RI.Variable->getType(), LHSI8Ptr->getName() + ".ascast");
 
    if (ReductionGenCBKind == ReductionGenCBKind::Clang) {
      LHSPtrs.emplace_back(LHSPtr);
      RHSPtrs.emplace_back(RHSPtr);
    } else {
      Value *LHS = LHSPtr;
      Value *RHS = RHSPtr;
 
      if (!IsByRef.empty() && !IsByRef[En.index()]) {
        LHS = Builder.CreateLoad(RI.ElementType, LHSPtr);
        RHS = Builder.CreateLoad(RI.ElementType, RHSPtr);
      }
 
      Value *Reduced;
      InsertPointOrErrorTy AfterIP =
          RI.ReductionGen(Builder.saveIP(), LHS, RHS, Reduced);
      if (!AfterIP)
        return AfterIP.takeError();
      if (!Builder.GetInsertBlock())
        return ReductionFunc;
 
      Builder.restoreIP(*AfterIP);
 
      if (!IsByRef.empty() && !IsByRef[En.index()])
        Builder.CreateStore(Reduced, LHSPtr);
    }
  }
 
  if (ReductionGenCBKind == ReductionGenCBKind::Clang)
    for (auto En : enumerate(ReductionInfos)) {
      unsigned Index = En.index();
      const ReductionInfo &RI = En.value();
      Value *LHSFixupPtr, *RHSFixupPtr;
      Builder.restoreIP(RI.ReductionGenClang(
          Builder.saveIP(), Index, &LHSFixupPtr, &RHSFixupPtr, ReductionFunc));
 
      // Fix the CallBack code genereated to use the correct Values for the LHS
      // and RHS
      LHSFixupPtr->replaceUsesWithIf(
          LHSPtrs[Index], [ReductionFunc](const Use &U) {
            return cast<Instruction>(U.getUser())->getParent()->getParent() ==
                   ReductionFunc;
          });
      RHSFixupPtr->replaceUsesWithIf(
          RHSPtrs[Index], [ReductionFunc](const Use &U) {
            return cast<Instruction>(U.getUser())->getParent()->getParent() ==
                   ReductionFunc;
          });
    }
 
  Builder.CreateRetVoid();
  // Compiling with `-O0`, `alloca`s emitted in non-entry blocks are not hoisted
  // to the entry block (this is dones for higher opt levels by later passes in
  // the pipeline). This has caused issues because non-entry `alloca`s force the
  // function to use dynamic stack allocations and we might run out of scratch
  // memory.
  hoistNonEntryAllocasToEntryBlock(ReductionFunc);
 
  return ReductionFunc;
}
 
static void
checkReductionInfos(ArrayRef<OpenMPIRBuilder::ReductionInfo> ReductionInfos,
                    bool IsGPU) {
  for (const OpenMPIRBuilder::ReductionInfo &RI : ReductionInfos) {
    (void)RI;
    assert(RI.Variable && "expected non-null variable");
    assert(RI.PrivateVariable && "expected non-null private variable");
    assert((RI.ReductionGen || RI.ReductionGenClang) &&
           "expected non-null reduction generator callback");
    if (!IsGPU) {
      assert(
          RI.Variable->getType() == RI.PrivateVariable->getType() &&
          "expected variables and their private equivalents to have the same "
          "type");
    }
    assert(RI.Variable->getType()->isPointerTy() &&
           "expected variables to be pointers");
  }
}
 
// The atomic cross-team reduction fast path applies when every reduction in the
// set can be represented by an atomicrmw. Clang only populates it for scalar
// reductions with a supported atomic operator.
static bool isAtomicableReductionSet(
    ArrayRef<OpenMPIRBuilder::ReductionInfo> ReductionInfos) {
  return all_of(ReductionInfos, [](const OpenMPIRBuilder::ReductionInfo &RI) {
    return static_cast<bool>(RI.AtomicReductionGen);
  });
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createReductionsGPU(
    const LocationDescription &Loc, InsertPointTy AllocaIP,
    InsertPointTy CodeGenIP, ArrayRef<ReductionInfo> ReductionInfos,
    ArrayRef<bool> IsByRef, bool IsNoWait, bool IsTeamsReduction, bool IsSPMD,
    ReductionGenCBKind ReductionGenCBKind, std::optional<omp::GV> GridValue,
    Value *SrcLocInfo) {
  if (!updateToLocation(Loc))
    return InsertPointTy();
  Builder.restoreIP(CodeGenIP);
  checkReductionInfos(ReductionInfos, /*IsGPU*/ true);
  LLVMContext &Ctx = M.getContext();
 
  // Source location for the ident struct
  if (!SrcLocInfo) {
    uint32_t SrcLocStrSize;
    Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
    SrcLocInfo = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  }
 
  if (ReductionInfos.size() == 0)
    return Builder.saveIP();
 
  BasicBlock *ContinuationBlock = nullptr;
  if (ReductionGenCBKind != ReductionGenCBKind::Clang) {
    // Copied code from createReductions
    BasicBlock *InsertBlock = Loc.IP.getBlock();
    ContinuationBlock =
        InsertBlock->splitBasicBlock(Loc.IP.getPoint(), "reduce.finalize");
    InsertBlock->getTerminator()->eraseFromParent();
    Builder.SetInsertPoint(InsertBlock, InsertBlock->end());
  }
 
  Function *CurFunc = Builder.GetInsertBlock()->getParent();
  AttributeList FuncAttrs;
  AttrBuilder AttrBldr(Ctx);
  for (auto Attr : CurFunc->getAttributes().getFnAttrs())
    AttrBldr.addAttribute(Attr);
  AttrBldr.removeAttribute(Attribute::OptimizeNone);
  FuncAttrs = FuncAttrs.addFnAttributes(Ctx, AttrBldr);
 
  CodeGenIP = Builder.saveIP();
  Expected<Function *> ReductionResult = createReductionFunction(
      Builder.GetInsertBlock()->getParent()->getName(), ReductionInfos, IsByRef,
      ReductionGenCBKind, FuncAttrs);
  if (!ReductionResult)
    return ReductionResult.takeError();
  Function *ReductionFunc = *ReductionResult;
  Builder.restoreIP(CodeGenIP);
 
  // Set the grid value in the config needed for lowering later on
  if (GridValue.has_value())
    Config.setGridValue(GridValue.value());
  else
    Config.setGridValue(getGridValue(T, ReductionFunc));
 
  // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList),
  // RedList, shuffle_reduce_func, interwarp_copy_func);
  // or
  // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>);
  Value *Res;
 
  // 1. Build a list of reduction variables.
  // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
  auto Size = ReductionInfos.size();
  Type *PtrTy = PointerType::get(Ctx, Config.getDefaultTargetAS());
  Type *FuncPtrTy =
      Builder.getPtrTy(M.getDataLayout().getProgramAddressSpace());
  Type *RedArrayTy = ArrayType::get(PtrTy, Size);
  CodeGenIP = Builder.saveIP();
  Builder.restoreIP(AllocaIP);
  Value *ReductionListAlloca =
      Builder.CreateAlloca(RedArrayTy, nullptr, ".omp.reduction.red_list");
  Value *ReductionList = Builder.CreatePointerBitCastOrAddrSpaceCast(
      ReductionListAlloca, PtrTy, ReductionListAlloca->getName() + ".ascast");
  Builder.restoreIP(CodeGenIP);
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  for (auto En : enumerate(ReductionInfos)) {
    const ReductionInfo &RI = En.value();
    Value *ElemPtr = Builder.CreateInBoundsGEP(
        RedArrayTy, ReductionList,
        {ConstantInt::get(IndexTy, 0), ConstantInt::get(IndexTy, En.index())});
 
    Value *PrivateVar = RI.PrivateVariable;
    bool IsByRefElem = !IsByRef.empty() && IsByRef[En.index()];
    if (IsByRefElem)
      PrivateVar = Builder.CreateLoad(RI.ElementType, PrivateVar);
 
    Value *CastElem =
        Builder.CreatePointerBitCastOrAddrSpaceCast(PrivateVar, PtrTy);
    Builder.CreateStore(CastElem, ElemPtr);
  }
  CodeGenIP = Builder.saveIP();
  Expected<Function *> SarFunc = emitShuffleAndReduceFunction(
      ReductionInfos, ReductionFunc, FuncAttrs, IsByRef);
 
  if (!SarFunc)
    return SarFunc.takeError();
 
  Expected<Function *> CopyResult =
      emitInterWarpCopyFunction(Loc, ReductionInfos, FuncAttrs, IsByRef);
  if (!CopyResult)
    return CopyResult.takeError();
  Function *WcFunc = *CopyResult;
  Builder.restoreIP(CodeGenIP);
 
  Value *RL = Builder.CreatePointerBitCastOrAddrSpaceCast(ReductionList, PtrTy);
 
  // NOTE: ReductionDataSize is passed as the reduce_data_size argument to
  // __kmpc_nvptx_parallel_reduce_nowait_v2, but the runtime implementations do
  // not currently use it.  It is computed here conservatively as max(element
  // sizes) * N rather than the exact sum, which over-calculates the size for
  // mixed reduction types but is harmless given the argument is unused.
  // TODO: Consider dropping this computation if the runtime API is ever revised
  // to remove the unused parameter.
  unsigned MaxDataSize = 0;
  SmallVector<Type *> ReductionTypeArgs;
  for (auto En : enumerate(ReductionInfos)) {
    // Use ByRefElementType for by-ref reductions so that MaxDataSize matches
    // the actual data size stored in the global reduction buffer, consistent
    // with the ReductionsBufferTy struct used for GEP offsets below.
    Type *RedTypeArg = (!IsByRef.empty() && IsByRef[En.index()])
                           ? En.value().ByRefElementType
                           : En.value().ElementType;
    auto Size = M.getDataLayout().getTypeStoreSize(RedTypeArg);
    if (Size > MaxDataSize)
      MaxDataSize = Size;
    ReductionTypeArgs.emplace_back(RedTypeArg);
  }
  Value *ReductionDataSize =
      Builder.getInt64(MaxDataSize * ReductionInfos.size());
 
  // Helper function to copy thread-local data back to the original reduction
  // list.
  Function *CopyScratchToListFunc = nullptr;
  // Thread-local storage for the reduction variables.
  Value *ScratchForCopyBack = nullptr;
  // RL pointer to which the final value from the per-thread scratch should be
  // copied back. (Basically RL, appropriately casted if necessary.)
  Value *RLForCopyBack = RL;
 
  bool IsAtomicReduction =
      IsTeamsReduction && isAtomicableReductionSet(ReductionInfos);
 
  if (!IsTeamsReduction) {
    Value *SarFuncCast =
        Builder.CreatePointerBitCastOrAddrSpaceCast(*SarFunc, FuncPtrTy);
    Value *WcFuncCast =
        Builder.CreatePointerBitCastOrAddrSpaceCast(WcFunc, FuncPtrTy);
    Value *Args[] = {SrcLocInfo, ReductionDataSize, RL, SarFuncCast,
                     WcFuncCast};
    Function *Pv2Ptr = getOrCreateRuntimeFunctionPtr(
        RuntimeFunction::OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2);
    Res = createRuntimeFunctionCall(Pv2Ptr, Args);
  } else if (IsAtomicReduction) {
    // Atomic cross-team reduction fast path: determine the team's main thread
    // that is later to fold its value atomically into the mapped variable.
    Function *IsMainThreadFn = getOrCreateRuntimeFunctionPtr(
        RuntimeFunction::OMPRTL___kmpc_is_team_main_thread);
    Res = createRuntimeFunctionCall(IsMainThreadFn, {});
  } else {
    CodeGenIP = Builder.saveIP();
    StructType *ReductionsBufferTy = StructType::create(
        Ctx, ReductionTypeArgs, "struct._globalized_locals_ty");
 
    Expected<Function *> LtGCFunc = emitListToGlobalCopyFunction(
        ReductionInfos, ReductionsBufferTy, FuncAttrs, IsByRef);
    if (!LtGCFunc)
      return LtGCFunc.takeError();
 
    Expected<Function *> GtLCFunc = emitGlobalToListCopyFunction(
        ReductionInfos, ReductionsBufferTy, FuncAttrs, IsByRef);
    if (!GtLCFunc)
      return GtLCFunc.takeError();
 
    Expected<Function *> GtLRFunc = emitGlobalToListReduceFunction(
        ReductionInfos, ReductionFunc, ReductionsBufferTy, FuncAttrs, IsByRef);
    if (!GtLRFunc)
      return GtLRFunc.takeError();
 
    Builder.restoreIP(CodeGenIP);
 
    // The runtime's cross-team final aggregate uses the storage pointed at by
    // its reduce-list argument as per-thread scratch.  When the surrounding
    // kernel is already in SPMD execution mode, clang emitted each reduction
    // private as a per-thread `alloca addrspace(5)`, so the original red_list
    // (RL) is already per-thread and nothing else is needed.
    //
    // When the kernel is in Non-SPMD execution mode at codegen time, clang's
    // Generic-mode globalization put the reduction private into team-shared
    // LDS.  OpenMPOpt may later upgrade the kernel to Generic-SPMD, at which
    // point all threads of the last team would race on the shared LDS slot.
    // Emit a per-thread scratch buffer and a per-thread RL, copy the team-local
    // value in, and hand the per-thread RL to the runtime instead. The writer
    // thread copies the final value from that per-thread scratch back to RL
    // before running the existing combine path below.
 
    // Thread-local RL (might need localization below before being passed to the
    // runtime).
    Value *RuntimeRL = RL;
 
    if (!IsSPMD) {
      CodeGenIP = Builder.saveIP();
      Builder.restoreIP(AllocaIP);
      // Allocate thread-local buffer for the reduction variables.
      Value *PerThreadScratchAlloca = Builder.CreateAlloca(
          ReductionsBufferTy, /*ArraySize=*/nullptr, ".omp.reduction.scratch");
      Value *PerThreadScratch = Builder.CreatePointerBitCastOrAddrSpaceCast(
          PerThreadScratchAlloca, PtrTy,
          PerThreadScratchAlloca->getName() + ".ascast");
      // Allocate thread-local buffer for the pointers to the reduction
      // variables.
      Value *PerThreadRedListAlloca =
          Builder.CreateAlloca(RedArrayTy, /*ArraySize=*/nullptr,
                               ".omp.reduction.per_thread_red_list");
      RuntimeRL = Builder.CreatePointerBitCastOrAddrSpaceCast(
          PerThreadRedListAlloca, PtrTy,
          PerThreadRedListAlloca->getName() + ".ascast");
      Builder.restoreIP(CodeGenIP);
 
      // Iterate over the reduction variables and copy the team-local value to
      // the thread-local buffer.
      for (auto En : enumerate(ReductionInfos)) {
        const ReductionInfo &RI = En.value();
        bool IsByRefElem = !IsByRef.empty() && IsByRef[En.index()];
 
        Value *FieldPtr = Builder.CreateConstInBoundsGEP2_32(
            ReductionsBufferTy, PerThreadScratch, 0, En.index());
        Value *Slot = Builder.CreateConstInBoundsGEP2_32(RedArrayTy, RuntimeRL,
                                                         0, En.index());
 
        Value *RuntimeListEntry = FieldPtr;
        if (IsByRefElem && RI.DataPtrPtrGen) {
          Value *SrcDescriptor =
              Builder.CreateLoad(RI.ElementType, RI.PrivateVariable);
          Expected<Value *> Descriptor = createReductionDescriptorCopy(
              AllocaIP, RI, FieldPtr, SrcDescriptor, PtrTy);
          if (!Descriptor)
            return Descriptor.takeError();
          RuntimeListEntry = *Descriptor;
        }
        Builder.CreateStore(RuntimeListEntry, Slot);
      }
      // The copy helpers were emitted with default-AS (AS 0) pointer params
      // (see emitListToGlobalCopyFunction / emitGlobalToListCopyFunction),
      // but PerThreadScratch and RL live in the target's default AS, which
      // is non-zero on e.g. SPIRV. (See Config.getDefaultTargetAS().)
      Type *CopyArg0Ty = (*LtGCFunc)->getFunctionType()->getParamType(0);
      Type *CopyArg2Ty = (*LtGCFunc)->getFunctionType()->getParamType(2);
      ScratchForCopyBack = Builder.CreatePointerBitCastOrAddrSpaceCast(
          PerThreadScratch, CopyArg0Ty);
      RLForCopyBack =
          Builder.CreatePointerBitCastOrAddrSpaceCast(RL, CopyArg2Ty);
      // Use index 0 because there is no array of target values to index into,
      // there is only one thread-local memory slot.
      // restoreIP above left a stale/empty debug location; this inlinable call
      // to a debug-info-bearing helper needs one or the verifier rejects the
      // module ("!dbg attachment points at wrong subprogram") after inlining.
      Builder.SetCurrentDebugLocation(Loc.DL);
      Builder.CreateCall(
          *LtGCFunc, {ScratchForCopyBack, Builder.getInt32(0), RLForCopyBack});
      CopyScratchToListFunc = *GtLCFunc;
    }
 
    Value *Args3[] = {SrcLocInfo, RuntimeRL, *SarFunc, WcFunc,
                      *LtGCFunc,  *GtLCFunc, *GtLRFunc};
 
    Function *TeamsReduceFn = getOrCreateRuntimeFunctionPtr(
        RuntimeFunction::OMPRTL___kmpc_gpu_xteam_reduce_nowait);
    Res = createRuntimeFunctionCall(TeamsReduceFn, Args3);
  }
 
  // 5. Build if (res == 1)
  BasicBlock *ExitBB = BasicBlock::Create(Ctx, ".omp.reduction.done");
  BasicBlock *ThenBB = BasicBlock::Create(Ctx, ".omp.reduction.then");
  Value *Cond = Builder.CreateICmpEQ(Res, Builder.getInt32(1));
  Builder.CreateCondBr(Cond, ThenBB, ExitBB);
 
  // 6. Build then branch: where we have reduced values in the master
  //    thread in each team.
  //    __kmpc_end_reduce{_nowait}(<gtid>);
  //    break;
  emitBlock(ThenBB, CurFunc);
 
  // Copy the writer thread's per-thread scratch result back into the original
  // red-list storage before the existing combine path reads RI.PrivateVariable.
  // Set a debug location: this inlinable call to a debug-info-bearing helper
  // needs one or the verifier rejects the module after inlining.
  if (ScratchForCopyBack) {
    Builder.SetCurrentDebugLocation(Loc.DL);
    Builder.CreateCall(
        CopyScratchToListFunc,
        {ScratchForCopyBack, Builder.getInt32(0), RLForCopyBack});
  }
 
  // Add emission of __kmpc_end_reduce{_nowait}(<gtid>);
  for (auto En : enumerate(ReductionInfos)) {
    const ReductionInfo &RI = En.value();
 
    // Atomic cross-team fast path: each team's main thread folds its
    // team-reduced value directly into the mapped reduction variable with a
    // single atomicrmw.
    if (IsAtomicReduction) {
      InsertPointOrErrorTy AfterIP = RI.AtomicReductionGen(
          Builder.saveIP(), RI.ElementType, RI.Variable, RI.PrivateVariable);
      if (!AfterIP)
        return AfterIP.takeError();
      Builder.restoreIP(*AfterIP);
      continue;
    }
 
    Type *ValueType = RI.ElementType;
    Value *RedValue = RI.Variable;
 
    Value *RHS =
        Builder.CreatePointerBitCastOrAddrSpaceCast(RI.PrivateVariable, PtrTy);
 
    if (ReductionGenCBKind == ReductionGenCBKind::Clang) {
      Value *LHSPtr, *RHSPtr;
      Builder.restoreIP(RI.ReductionGenClang(Builder.saveIP(), En.index(),
                                             &LHSPtr, &RHSPtr, CurFunc));
 
      // Fix the CallBack code genereated to use the correct Values for the LHS
      // and RHS. Cast to match types before replacing (necessary to handle
      // different address spaces).
      if (LHSPtr->getType() != RedValue->getType())
        RedValue = Builder.CreatePointerBitCastOrAddrSpaceCast(
            RedValue, LHSPtr->getType());
      if (RHSPtr->getType() != RHS->getType())
        RHS =
            Builder.CreatePointerBitCastOrAddrSpaceCast(RHS, RHSPtr->getType());
 
      LHSPtr->replaceUsesWithIf(RedValue, [ReductionFunc](const Use &U) {
        return cast<Instruction>(U.getUser())->getParent()->getParent() ==
               ReductionFunc;
      });
      RHSPtr->replaceUsesWithIf(RHS, [ReductionFunc](const Use &U) {
        return cast<Instruction>(U.getUser())->getParent()->getParent() ==
               ReductionFunc;
      });
    } else {
      if (IsByRef.empty() || !IsByRef[En.index()]) {
        RedValue = Builder.CreateLoad(ValueType, RI.Variable,
                                      "red.value." + Twine(En.index()));
      }
      Value *PrivateRedValue = Builder.CreateLoad(
          ValueType, RHS, "red.private.value" + Twine(En.index()));
      Value *Reduced;
      InsertPointOrErrorTy AfterIP =
          RI.ReductionGen(Builder.saveIP(), RedValue, PrivateRedValue, Reduced);
      if (!AfterIP)
        return AfterIP.takeError();
      Builder.restoreIP(*AfterIP);
 
      if (!IsByRef.empty() && !IsByRef[En.index()])
        Builder.CreateStore(Reduced, RI.Variable);
    }
  }
  emitBlock(ExitBB, CurFunc);
  if (ContinuationBlock) {
    Builder.CreateBr(ContinuationBlock);
    Builder.SetInsertPoint(ContinuationBlock);
  }
  Config.setEmitLLVMUsed();
 
  return Builder.saveIP();
}
 
static Function *getFreshReductionFunc(Module &M) {
  Type *VoidTy = Type::getVoidTy(M.getContext());
  Type *Int8PtrTy = PointerType::getUnqual(M.getContext());
  auto *FuncTy =
      FunctionType::get(VoidTy, {Int8PtrTy, Int8PtrTy}, /* IsVarArg */ false);
  return Function::Create(FuncTy, GlobalVariable::InternalLinkage,
                          ".omp.reduction.func", &M);
}
 
static Error populateReductionFunction(
    Function *ReductionFunc,
    ArrayRef<OpenMPIRBuilder::ReductionInfo> ReductionInfos,
    IRBuilder<> &Builder, ArrayRef<bool> IsByRef, bool IsGPU) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  Module *Module = ReductionFunc->getParent();
  BasicBlock *ReductionFuncBlock =
      BasicBlock::Create(Module->getContext(), "", ReductionFunc);
  Builder.SetInsertPoint(ReductionFuncBlock);
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());
  Value *LHSArrayPtr = nullptr;
  Value *RHSArrayPtr = nullptr;
  if (IsGPU) {
    // Need to alloca memory here and deal with the pointers before getting
    // LHS/RHS pointers out
    //
    Argument *Arg0 = ReductionFunc->getArg(0);
    Argument *Arg1 = ReductionFunc->getArg(1);
    Type *Arg0Type = Arg0->getType();
    Type *Arg1Type = Arg1->getType();
 
    Value *LHSAlloca =
        Builder.CreateAlloca(Arg0Type, nullptr, Arg0->getName() + ".addr");
    Value *RHSAlloca =
        Builder.CreateAlloca(Arg1Type, nullptr, Arg1->getName() + ".addr");
    Value *LHSAddrCast =
        Builder.CreatePointerBitCastOrAddrSpaceCast(LHSAlloca, Arg0Type);
    Value *RHSAddrCast =
        Builder.CreatePointerBitCastOrAddrSpaceCast(RHSAlloca, Arg1Type);
    Builder.CreateStore(Arg0, LHSAddrCast);
    Builder.CreateStore(Arg1, RHSAddrCast);
    LHSArrayPtr = Builder.CreateLoad(Arg0Type, LHSAddrCast);
    RHSArrayPtr = Builder.CreateLoad(Arg1Type, RHSAddrCast);
  } else {
    LHSArrayPtr = ReductionFunc->getArg(0);
    RHSArrayPtr = ReductionFunc->getArg(1);
  }
 
  unsigned NumReductions = ReductionInfos.size();
  Type *RedArrayTy = ArrayType::get(Builder.getPtrTy(), NumReductions);
 
  for (auto En : enumerate(ReductionInfos)) {
    const OpenMPIRBuilder::ReductionInfo &RI = En.value();
    Value *LHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64(
        RedArrayTy, LHSArrayPtr, 0, En.index());
    Value *LHSI8Ptr = Builder.CreateLoad(Builder.getPtrTy(), LHSI8PtrPtr);
    Value *LHSPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
        LHSI8Ptr, RI.Variable->getType());
    Value *LHS = Builder.CreateLoad(RI.ElementType, LHSPtr);
    Value *RHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64(
        RedArrayTy, RHSArrayPtr, 0, En.index());
    Value *RHSI8Ptr = Builder.CreateLoad(Builder.getPtrTy(), RHSI8PtrPtr);
    Value *RHSPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
        RHSI8Ptr, RI.PrivateVariable->getType());
    Value *RHS = Builder.CreateLoad(RI.ElementType, RHSPtr);
    Value *Reduced;
    OpenMPIRBuilder::InsertPointOrErrorTy AfterIP =
        RI.ReductionGen(Builder.saveIP(), LHS, RHS, Reduced);
    if (!AfterIP)
      return AfterIP.takeError();
 
    Builder.restoreIP(*AfterIP);
    // TODO: Consider flagging an error.
    if (!Builder.GetInsertBlock())
      return Error::success();
 
    // store is inside of the reduction region when using by-ref
    if (!IsByRef[En.index()])
      Builder.CreateStore(Reduced, LHSPtr);
  }
  Builder.CreateRetVoid();
  return Error::success();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createReductions(
    const LocationDescription &Loc, InsertPointTy AllocaIP,
    ArrayRef<ReductionInfo> ReductionInfos, ArrayRef<bool> IsByRef,
    bool IsNoWait, bool IsTeamsReduction) {
  assert(ReductionInfos.size() == IsByRef.size());
  if (Config.isGPU())
    return createReductionsGPU(Loc, AllocaIP, Builder.saveIP(), ReductionInfos,
                               IsByRef, IsNoWait, IsTeamsReduction);
 
  checkReductionInfos(ReductionInfos, /*IsGPU*/ false);
 
  if (!updateToLocation(Loc))
    return InsertPointTy();
 
  if (ReductionInfos.size() == 0)
    return Builder.saveIP();
 
  BasicBlock *InsertBlock = Loc.IP.getBlock();
  BasicBlock *ContinuationBlock =
      InsertBlock->splitBasicBlock(Loc.IP.getPoint(), "reduce.finalize");
  InsertBlock->getTerminator()->eraseFromParent();
 
  // Create and populate array of type-erased pointers to private reduction
  // values.
  unsigned NumReductions = ReductionInfos.size();
  Type *RedArrayTy = ArrayType::get(Builder.getPtrTy(), NumReductions);
  Builder.SetInsertPoint(AllocaIP.getBlock()->getTerminator());
  Value *RedArray = Builder.CreateAlloca(RedArrayTy, nullptr, "red.array");
 
  Builder.SetInsertPoint(InsertBlock, InsertBlock->end());
 
  for (auto En : enumerate(ReductionInfos)) {
    unsigned Index = En.index();
    const ReductionInfo &RI = En.value();
    Value *RedArrayElemPtr = Builder.CreateConstInBoundsGEP2_64(
        RedArrayTy, RedArray, 0, Index, "red.array.elem." + Twine(Index));
    Builder.CreateStore(RI.PrivateVariable, RedArrayElemPtr);
  }
 
  // Emit a call to the runtime function that orchestrates the reduction.
  // Declare the reduction function in the process.
  Type *IndexTy = Builder.getIndexTy(
      M.getDataLayout(), M.getDataLayout().getDefaultGlobalsAddressSpace());
  Function *Func = Builder.GetInsertBlock()->getParent();
  Module *Module = Func->getParent();
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  bool CanGenerateAtomic = all_of(ReductionInfos, [](const ReductionInfo &RI) {
    return RI.AtomicReductionGen;
  });
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize,
                                  CanGenerateAtomic
                                      ? IdentFlag::OMP_IDENT_FLAG_ATOMIC_REDUCE
                                      : IdentFlag(0));
  Value *ThreadId = getOrCreateThreadID(Ident);
  Constant *NumVariables = Builder.getInt32(NumReductions);
  const DataLayout &DL = Module->getDataLayout();
  unsigned RedArrayByteSize = DL.getTypeStoreSize(RedArrayTy);
  Constant *RedArraySize = ConstantInt::get(IndexTy, RedArrayByteSize);
  Function *ReductionFunc = getFreshReductionFunc(*Module);
  Value *Lock = getOMPCriticalRegionLock(".reduction");
  Function *ReduceFunc = getOrCreateRuntimeFunctionPtr(
      IsNoWait ? RuntimeFunction::OMPRTL___kmpc_reduce_nowait
               : RuntimeFunction::OMPRTL___kmpc_reduce);
  CallInst *ReduceCall =
      createRuntimeFunctionCall(ReduceFunc,
                                {Ident, ThreadId, NumVariables, RedArraySize,
                                 RedArray, ReductionFunc, Lock},
                                "reduce");
 
  // Create final reduction entry blocks for the atomic and non-atomic case.
  // Emit IR that dispatches control flow to one of the blocks based on the
  // reduction supporting the atomic mode.
  BasicBlock *NonAtomicRedBlock =
      BasicBlock::Create(Module->getContext(), "reduce.switch.nonatomic", Func);
  BasicBlock *AtomicRedBlock =
      BasicBlock::Create(Module->getContext(), "reduce.switch.atomic", Func);
  SwitchInst *Switch =
      Builder.CreateSwitch(ReduceCall, ContinuationBlock, /* NumCases */ 2);
  Switch->addCase(Builder.getInt32(1), NonAtomicRedBlock);
  Switch->addCase(Builder.getInt32(2), AtomicRedBlock);
 
  // Populate the non-atomic reduction using the elementwise reduction function.
  // This loads the elements from the global and private variables and reduces
  // them before storing back the result to the global variable.
  Builder.SetInsertPoint(NonAtomicRedBlock);
  for (auto En : enumerate(ReductionInfos)) {
    const ReductionInfo &RI = En.value();
    Type *ValueType = RI.ElementType;
    // We have one less load for by-ref case because that load is now inside of
    // the reduction region
    Value *RedValue = RI.Variable;
    if (!IsByRef[En.index()]) {
      RedValue = Builder.CreateLoad(ValueType, RI.Variable,
                                    "red.value." + Twine(En.index()));
    }
    Value *PrivateRedValue =
        Builder.CreateLoad(ValueType, RI.PrivateVariable,
                           "red.private.value." + Twine(En.index()));
    Value *Reduced;
    InsertPointOrErrorTy AfterIP =
        RI.ReductionGen(Builder.saveIP(), RedValue, PrivateRedValue, Reduced);
    if (!AfterIP)
      return AfterIP.takeError();
    Builder.restoreIP(*AfterIP);
 
    if (!Builder.GetInsertBlock())
      return InsertPointTy();
    // for by-ref case, the load is inside of the reduction region
    if (!IsByRef[En.index()])
      Builder.CreateStore(Reduced, RI.Variable);
  }
  Function *EndReduceFunc = getOrCreateRuntimeFunctionPtr(
      IsNoWait ? RuntimeFunction::OMPRTL___kmpc_end_reduce_nowait
               : RuntimeFunction::OMPRTL___kmpc_end_reduce);
  createRuntimeFunctionCall(EndReduceFunc, {Ident, ThreadId, Lock});
  Builder.CreateBr(ContinuationBlock);
 
  // Populate the atomic reduction using the atomic elementwise reduction
  // function. There are no loads/stores here because they will be happening
  // inside the atomic elementwise reduction.
  Builder.SetInsertPoint(AtomicRedBlock);
  if (CanGenerateAtomic && llvm::none_of(IsByRef, [](bool P) { return P; })) {
    for (const ReductionInfo &RI : ReductionInfos) {
      InsertPointOrErrorTy AfterIP = RI.AtomicReductionGen(
          Builder.saveIP(), RI.ElementType, RI.Variable, RI.PrivateVariable);
      if (!AfterIP)
        return AfterIP.takeError();
      Builder.restoreIP(*AfterIP);
      if (!Builder.GetInsertBlock())
        return InsertPointTy();
    }
    Builder.CreateBr(ContinuationBlock);
  } else {
    Builder.CreateUnreachable();
  }
 
  // Populate the outlined reduction function using the elementwise reduction
  // function. Partial values are extracted from the type-erased array of
  // pointers to private variables.
  Error Err = populateReductionFunction(ReductionFunc, ReductionInfos, Builder,
                                        IsByRef, /*isGPU=*/false);
  if (Err)
    return Err;
 
  if (!Builder.GetInsertBlock())
    return InsertPointTy();
 
  Builder.SetInsertPoint(ContinuationBlock);
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::createMaster(const LocationDescription &Loc,
                              BodyGenCallbackTy BodyGenCB,
                              FinalizeCallbackTy FiniCB) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  Directive OMPD = Directive::OMPD_master;
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Value *Args[] = {Ident, ThreadId};
 
  Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_master);
  Instruction *EntryCall = createRuntimeFunctionCall(EntryRTLFn, Args);
 
  Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_master);
  Instruction *ExitCall = createRuntimeFunctionCall(ExitRTLFn, Args);
 
  return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
                              /*Conditional*/ true, /*hasFinalize*/ true);
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::createMasked(const LocationDescription &Loc,
                              BodyGenCallbackTy BodyGenCB,
                              FinalizeCallbackTy FiniCB, Value *Filter) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  Directive OMPD = Directive::OMPD_masked;
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Value *Args[] = {Ident, ThreadId, Filter};
  Value *ArgsEnd[] = {Ident, ThreadId};
 
  Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_masked);
  Instruction *EntryCall = createRuntimeFunctionCall(EntryRTLFn, Args);
 
  Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_masked);
  Instruction *ExitCall = createRuntimeFunctionCall(ExitRTLFn, ArgsEnd);
 
  return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
                              /*Conditional*/ true, /*hasFinalize*/ true);
}
 
static llvm::CallInst *emitNoUnwindRuntimeCall(IRBuilder<> &Builder,
                                               llvm::FunctionCallee Callee,
                                               ArrayRef<llvm::Value *> Args,
                                               const llvm::Twine &Name) {
  llvm::CallInst *Call = Builder.CreateCall(
      Callee, Args, SmallVector<llvm::OperandBundleDef, 1>(), Name);
  Call->setDoesNotThrow();
  return Call;
}
 
// Expects input basic block is dominated by BeforeScanBB.
// Once Scan directive is encountered, the code after scan directive should be
// dominated by AfterScanBB. Scan directive splits the code sequence to
// scan and input phase. Based on whether inclusive or exclusive
// clause is used in the scan directive and whether input loop or scan loop
// is lowered, it adds jumps to input and scan phase. First Scan loop is the
// input loop and second is the scan loop. The code generated handles only
// inclusive scans now.
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createScan(
    const LocationDescription &Loc, InsertPointTy AllocaIP,
    ArrayRef<llvm::Value *> ScanVars, ArrayRef<llvm::Type *> ScanVarsType,
    bool IsInclusive, ScanInfo *ScanRedInfo) {
  if (ScanRedInfo->OMPFirstScanLoop) {
    llvm::Error Err = emitScanBasedDirectiveDeclsIR(AllocaIP, ScanVars,
                                                    ScanVarsType, ScanRedInfo);
    if (Err)
      return Err;
  }
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  llvm::Value *IV = ScanRedInfo->IV;
 
  if (ScanRedInfo->OMPFirstScanLoop) {
    // Emit buffer[i] = red; at the end of the input phase.
    for (size_t i = 0; i < ScanVars.size(); i++) {
      Value *BuffPtr = (*(ScanRedInfo->ScanBuffPtrs))[ScanVars[i]];
      Value *Buff = Builder.CreateLoad(Builder.getPtrTy(), BuffPtr);
      Type *DestTy = ScanVarsType[i];
      Value *Val = Builder.CreateInBoundsGEP(DestTy, Buff, IV, "arrayOffset");
      Value *Src = Builder.CreateLoad(DestTy, ScanVars[i]);
 
      Builder.CreateStore(Src, Val);
    }
  }
  Builder.CreateBr(ScanRedInfo->OMPScanLoopExit);
  emitBlock(ScanRedInfo->OMPScanDispatch,
            Builder.GetInsertBlock()->getParent());
 
  if (!ScanRedInfo->OMPFirstScanLoop) {
    IV = ScanRedInfo->IV;
    // Emit red = buffer[i]; at the entrance to the scan phase.
    // TODO: if exclusive scan, the red = buffer[i-1] needs to be updated.
    for (size_t i = 0; i < ScanVars.size(); i++) {
      Value *BuffPtr = (*(ScanRedInfo->ScanBuffPtrs))[ScanVars[i]];
      Value *Buff = Builder.CreateLoad(Builder.getPtrTy(), BuffPtr);
      Type *DestTy = ScanVarsType[i];
      Value *SrcPtr =
          Builder.CreateInBoundsGEP(DestTy, Buff, IV, "arrayOffset");
      Value *Src = Builder.CreateLoad(DestTy, SrcPtr);
      Builder.CreateStore(Src, ScanVars[i]);
    }
  }
 
  // TODO: Update it to CreateBr and remove dead blocks
  llvm::Value *CmpI = Builder.getInt1(true);
  if (ScanRedInfo->OMPFirstScanLoop == IsInclusive) {
    Builder.CreateCondBr(CmpI, ScanRedInfo->OMPBeforeScanBlock,
                         ScanRedInfo->OMPAfterScanBlock);
  } else {
    Builder.CreateCondBr(CmpI, ScanRedInfo->OMPAfterScanBlock,
                         ScanRedInfo->OMPBeforeScanBlock);
  }
  emitBlock(ScanRedInfo->OMPAfterScanBlock,
            Builder.GetInsertBlock()->getParent());
  Builder.SetInsertPoint(ScanRedInfo->OMPAfterScanBlock);
  return Builder.saveIP();
}
 
Error OpenMPIRBuilder::emitScanBasedDirectiveDeclsIR(
    InsertPointTy AllocaIP, ArrayRef<Value *> ScanVars,
    ArrayRef<Type *> ScanVarsType, ScanInfo *ScanRedInfo) {
 
  Builder.restoreIP(AllocaIP);
  // Create the shared pointer at alloca IP.
  for (size_t i = 0; i < ScanVars.size(); i++) {
    llvm::Value *BuffPtr =
        Builder.CreateAlloca(Builder.getPtrTy(), nullptr, "vla");
    (*(ScanRedInfo->ScanBuffPtrs))[ScanVars[i]] = BuffPtr;
  }
 
  // Allocate temporary buffer by master thread
  auto BodyGenCB = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
                       ArrayRef<BasicBlock *> DeallocBlocks) -> Error {
    Builder.restoreIP(CodeGenIP);
    Value *AllocSpan =
        Builder.CreateAdd(ScanRedInfo->Span, Builder.getInt32(1));
    for (size_t i = 0; i < ScanVars.size(); i++) {
      Type *IntPtrTy = Builder.getInt32Ty();
      Constant *Allocsize = ConstantExpr::getSizeOf(ScanVarsType[i]);
      Allocsize = ConstantExpr::getTruncOrBitCast(Allocsize, IntPtrTy);
      Value *Buff = Builder.CreateMalloc(IntPtrTy, ScanVarsType[i], Allocsize,
                                         AllocSpan, nullptr, "arr");
      Builder.CreateStore(Buff, (*(ScanRedInfo->ScanBuffPtrs))[ScanVars[i]]);
    }
    return Error::success();
  };
  // TODO: Perform finalization actions for variables. This has to be
  // called for variables which have destructors/finalizers.
  auto FiniCB = [&](InsertPointTy CodeGenIP) { return llvm::Error::success(); };
 
  Builder.SetInsertPoint(ScanRedInfo->OMPScanInit->getTerminator());
  llvm::Value *FilterVal = Builder.getInt32(0);
  llvm::OpenMPIRBuilder::InsertPointOrErrorTy AfterIP =
      createMasked(Builder.saveIP(), BodyGenCB, FiniCB, FilterVal);
 
  if (!AfterIP)
    return AfterIP.takeError();
  Builder.restoreIP(*AfterIP);
  BasicBlock *InputBB = Builder.GetInsertBlock();
  if (InputBB->hasTerminator())
    Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator());
  AfterIP = createBarrier(Builder.saveIP(), llvm::omp::OMPD_barrier);
  if (!AfterIP)
    return AfterIP.takeError();
  Builder.restoreIP(*AfterIP);
 
  return Error::success();
}
 
Error OpenMPIRBuilder::emitScanBasedDirectiveFinalsIR(
    ArrayRef<ReductionInfo> ReductionInfos, ScanInfo *ScanRedInfo) {
  auto BodyGenCB = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
                       ArrayRef<BasicBlock *> DeallocBlocks) -> Error {
    Builder.restoreIP(CodeGenIP);
    for (ReductionInfo RedInfo : ReductionInfos) {
      Value *PrivateVar = RedInfo.PrivateVariable;
      Value *OrigVar = RedInfo.Variable;
      Value *BuffPtr = (*(ScanRedInfo->ScanBuffPtrs))[PrivateVar];
      Value *Buff = Builder.CreateLoad(Builder.getPtrTy(), BuffPtr);
 
      Type *SrcTy = RedInfo.ElementType;
      Value *Val = Builder.CreateInBoundsGEP(SrcTy, Buff, ScanRedInfo->Span,
                                             "arrayOffset");
      Value *Src = Builder.CreateLoad(SrcTy, Val);
 
      Builder.CreateStore(Src, OrigVar);
      Builder.CreateFree(Buff);
    }
    return Error::success();
  };
  // TODO: Perform finalization actions for variables. This has to be
  // called for variables which have destructors/finalizers.
  auto FiniCB = [&](InsertPointTy CodeGenIP) { return llvm::Error::success(); };
 
  if (Instruction *TI = ScanRedInfo->OMPScanFinish->getTerminatorOrNull())
    Builder.SetInsertPoint(TI);
  else
    Builder.SetInsertPoint(ScanRedInfo->OMPScanFinish);
 
  llvm::Value *FilterVal = Builder.getInt32(0);
  llvm::OpenMPIRBuilder::InsertPointOrErrorTy AfterIP =
      createMasked(Builder.saveIP(), BodyGenCB, FiniCB, FilterVal);
 
  if (!AfterIP)
    return AfterIP.takeError();
  Builder.restoreIP(*AfterIP);
  BasicBlock *InputBB = Builder.GetInsertBlock();
  if (InputBB->hasTerminator())
    Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator());
  AfterIP = createBarrier(Builder.saveIP(), llvm::omp::OMPD_barrier);
  if (!AfterIP)
    return AfterIP.takeError();
  Builder.restoreIP(*AfterIP);
  return Error::success();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::emitScanReduction(
    const LocationDescription &Loc,
    ArrayRef<llvm::OpenMPIRBuilder::ReductionInfo> ReductionInfos,
    ScanInfo *ScanRedInfo) {
 
  if (!updateToLocation(Loc))
    return Loc.IP;
  auto BodyGenCB = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
                       ArrayRef<BasicBlock *> DeallocBlocks) -> Error {
    Builder.restoreIP(CodeGenIP);
    Function *CurFn = Builder.GetInsertBlock()->getParent();
    // for (int k = 0; k <= ceil(log2(n)); ++k)
    llvm::BasicBlock *LoopBB =
        BasicBlock::Create(CurFn->getContext(), "omp.outer.log.scan.body");
    llvm::BasicBlock *ExitBB =
        splitBB(Builder, false, "omp.outer.log.scan.exit");
    llvm::Function *F = llvm::Intrinsic::getOrInsertDeclaration(
        Builder.GetInsertBlock()->getModule(),
        (llvm::Intrinsic::ID)llvm::Intrinsic::log2, Builder.getDoubleTy());
    llvm::BasicBlock *InputBB = Builder.GetInsertBlock();
    llvm::Value *Arg =
        Builder.CreateUIToFP(ScanRedInfo->Span, Builder.getDoubleTy());
    llvm::Value *LogVal = emitNoUnwindRuntimeCall(Builder, F, Arg, "");
    F = llvm::Intrinsic::getOrInsertDeclaration(
        Builder.GetInsertBlock()->getModule(),
        (llvm::Intrinsic::ID)llvm::Intrinsic::ceil, Builder.getDoubleTy());
    LogVal = emitNoUnwindRuntimeCall(Builder, F, LogVal, "");
    LogVal = Builder.CreateFPToUI(LogVal, Builder.getInt32Ty());
    llvm::Value *NMin1 = Builder.CreateNUWSub(
        ScanRedInfo->Span,
        llvm::ConstantInt::get(ScanRedInfo->Span->getType(), 1));
    Builder.SetInsertPoint(InputBB);
    Builder.CreateBr(LoopBB);
    emitBlock(LoopBB, CurFn);
    Builder.SetInsertPoint(LoopBB);
 
    PHINode *Counter = Builder.CreatePHI(Builder.getInt32Ty(), 2);
    // size pow2k = 1;
    PHINode *Pow2K = Builder.CreatePHI(Builder.getInt32Ty(), 2);
    Counter->addIncoming(llvm::ConstantInt::get(Builder.getInt32Ty(), 0),
                         InputBB);
    Pow2K->addIncoming(llvm::ConstantInt::get(Builder.getInt32Ty(), 1),
                       InputBB);
    // for (size i = n - 1; i >= 2 ^ k; --i)
    //   tmp[i] op= tmp[i-pow2k];
    llvm::BasicBlock *InnerLoopBB =
        BasicBlock::Create(CurFn->getContext(), "omp.inner.log.scan.body");
    llvm::BasicBlock *InnerExitBB =
        BasicBlock::Create(CurFn->getContext(), "omp.inner.log.scan.exit");
    llvm::Value *CmpI = Builder.CreateICmpUGE(NMin1, Pow2K);
    Builder.CreateCondBr(CmpI, InnerLoopBB, InnerExitBB);
    emitBlock(InnerLoopBB, CurFn);
    Builder.SetInsertPoint(InnerLoopBB);
    PHINode *IVal = Builder.CreatePHI(Builder.getInt32Ty(), 2);
    IVal->addIncoming(NMin1, LoopBB);
    for (ReductionInfo RedInfo : ReductionInfos) {
      Value *ReductionVal = RedInfo.PrivateVariable;
      Value *BuffPtr = (*(ScanRedInfo->ScanBuffPtrs))[ReductionVal];
      Value *Buff = Builder.CreateLoad(Builder.getPtrTy(), BuffPtr);
      Type *DestTy = RedInfo.ElementType;
      Value *IV = Builder.CreateAdd(IVal, Builder.getInt32(1));
      Value *LHSPtr =
          Builder.CreateInBoundsGEP(DestTy, Buff, IV, "arrayOffset");
      Value *OffsetIval = Builder.CreateNUWSub(IV, Pow2K);
      Value *RHSPtr =
          Builder.CreateInBoundsGEP(DestTy, Buff, OffsetIval, "arrayOffset");
      Value *LHS = Builder.CreateLoad(DestTy, LHSPtr);
      Value *RHS = Builder.CreateLoad(DestTy, RHSPtr);
      llvm::Value *Result;
      InsertPointOrErrorTy AfterIP =
          RedInfo.ReductionGen(Builder.saveIP(), LHS, RHS, Result);
      if (!AfterIP)
        return AfterIP.takeError();
      Builder.CreateStore(Result, LHSPtr);
    }
    llvm::Value *NextIVal = Builder.CreateNUWSub(
        IVal, llvm::ConstantInt::get(Builder.getInt32Ty(), 1));
    IVal->addIncoming(NextIVal, Builder.GetInsertBlock());
    CmpI = Builder.CreateICmpUGE(NextIVal, Pow2K);
    Builder.CreateCondBr(CmpI, InnerLoopBB, InnerExitBB);
    emitBlock(InnerExitBB, CurFn);
    llvm::Value *Next = Builder.CreateNUWAdd(
        Counter, llvm::ConstantInt::get(Counter->getType(), 1));
    Counter->addIncoming(Next, Builder.GetInsertBlock());
    // pow2k <<= 1;
    llvm::Value *NextPow2K = Builder.CreateShl(Pow2K, 1, "", /*HasNUW=*/true);
    Pow2K->addIncoming(NextPow2K, Builder.GetInsertBlock());
    llvm::Value *Cmp = Builder.CreateICmpNE(Next, LogVal);
    Builder.CreateCondBr(Cmp, LoopBB, ExitBB);
    Builder.SetInsertPoint(ExitBB->getFirstInsertionPt());
    return Error::success();
  };
 
  // TODO: Perform finalization actions for variables. This has to be
  // called for variables which have destructors/finalizers.
  auto FiniCB = [&](InsertPointTy CodeGenIP) { return llvm::Error::success(); };
 
  llvm::Value *FilterVal = Builder.getInt32(0);
  llvm::OpenMPIRBuilder::InsertPointOrErrorTy AfterIP =
      createMasked(Builder.saveIP(), BodyGenCB, FiniCB, FilterVal);
 
  if (!AfterIP)
    return AfterIP.takeError();
  Builder.restoreIP(*AfterIP);
  AfterIP = createBarrier(Builder.saveIP(), llvm::omp::OMPD_barrier);
 
  if (!AfterIP)
    return AfterIP.takeError();
  Builder.restoreIP(*AfterIP);
  Error Err = emitScanBasedDirectiveFinalsIR(ReductionInfos, ScanRedInfo);
  if (Err)
    return Err;
 
  return AfterIP;
}
 
Error OpenMPIRBuilder::emitScanBasedDirectiveIR(
    llvm::function_ref<Error()> InputLoopGen,
    llvm::function_ref<Error(LocationDescription Loc)> ScanLoopGen,
    ScanInfo *ScanRedInfo) {
 
  {
    // Emit loop with input phase:
    // for (i: 0..<num_iters>) {
    //   <input phase>;
    //   buffer[i] = red;
    // }
    ScanRedInfo->OMPFirstScanLoop = true;
    Error Err = InputLoopGen();
    if (Err)
      return Err;
  }
  {
    // Emit loop with scan phase:
    // for (i: 0..<num_iters>) {
    //   red = buffer[i];
    //   <scan phase>;
    // }
    ScanRedInfo->OMPFirstScanLoop = false;
    Error Err = ScanLoopGen(Builder.saveIP());
    if (Err)
      return Err;
  }
  return Error::success();
}
 
void OpenMPIRBuilder::createScanBBs(ScanInfo *ScanRedInfo) {
  Function *Fun = Builder.GetInsertBlock()->getParent();
  ScanRedInfo->OMPScanDispatch =
      BasicBlock::Create(Fun->getContext(), "omp.inscan.dispatch");
  ScanRedInfo->OMPAfterScanBlock =
      BasicBlock::Create(Fun->getContext(), "omp.after.scan.bb");
  ScanRedInfo->OMPBeforeScanBlock =
      BasicBlock::Create(Fun->getContext(), "omp.before.scan.bb");
  ScanRedInfo->OMPScanLoopExit =
      BasicBlock::Create(Fun->getContext(), "omp.scan.loop.exit");
}
CanonicalLoopInfo *OpenMPIRBuilder::createLoopSkeleton(
    DebugLoc DL, Value *TripCount, Function *F, BasicBlock *PreInsertBefore,
    BasicBlock *PostInsertBefore, const Twine &Name, bool IsCollapsed) {
  Module *M = F->getParent();
  LLVMContext &Ctx = M->getContext();
  Type *IndVarTy = TripCount->getType();
 
  // Create the basic block structure.
  BasicBlock *Preheader =
      BasicBlock::Create(Ctx, "omp_" + Name + ".preheader", F, PreInsertBefore);
  BasicBlock *Header =
      BasicBlock::Create(Ctx, "omp_" + Name + ".header", F, PreInsertBefore);
  BasicBlock *Cond =
      BasicBlock::Create(Ctx, "omp_" + Name + ".cond", F, PreInsertBefore);
  BasicBlock *Body =
      BasicBlock::Create(Ctx, "omp_" + Name + ".body", F, PreInsertBefore);
  BasicBlock *Latch =
      BasicBlock::Create(Ctx, "omp_" + Name + ".inc", F, PostInsertBefore);
  BasicBlock *Exit =
      BasicBlock::Create(Ctx, "omp_" + Name + ".exit", F, PostInsertBefore);
  BasicBlock *After =
      BasicBlock::Create(Ctx, "omp_" + Name + ".after", F, PostInsertBefore);
 
  // Use specified DebugLoc for new instructions.
  Builder.SetCurrentDebugLocation(DL);
 
  Builder.SetInsertPoint(Preheader);
  Builder.CreateBr(Header);
 
  Builder.SetInsertPoint(Header);
  PHINode *IndVarPHI = Builder.CreatePHI(IndVarTy, 2, "omp_" + Name + ".iv");
  IndVarPHI->addIncoming(ConstantInt::get(IndVarTy, 0), Preheader);
  Builder.CreateBr(Cond);
 
  Builder.SetInsertPoint(Cond);
  Value *Cmp =
      Builder.CreateICmpULT(IndVarPHI, TripCount, "omp_" + Name + ".cmp");
  Builder.CreateCondBr(Cmp, Body, Exit);
 
  Builder.SetInsertPoint(Body);
  Builder.CreateBr(Latch);
 
  Builder.SetInsertPoint(Latch);
  // Decide whether the induction variable increment can carry nsw.
  //
  // Single loops: nsw is always kept (matching Clang). Any Fortran program
  // whose trip count overflows i32 is non-conforming per F2018 11.1.7.4.1, so
  // for valid programs 0 <= count <= INT_MAX always holds.
  //
  // Collapsed loops: the trip count is a product that can overflow i32 even for
  // a conforming program, so nsw is kept only when the product is a constant
  // that provably fits, dropped otherwise.
  bool HasNSW = Config.hasNoSignedWrap();
  if (HasNSW) {
    if (auto *CI = dyn_cast<ConstantInt>(TripCount)) {
      unsigned BitWidth = CI->getType()->getIntegerBitWidth();
      APInt SignedMax = APInt::getSignedMaxValue(BitWidth);
      if (CI->getValue().ugt(SignedMax))
        HasNSW = false;
    } else if (IsCollapsed) {
      HasNSW = false;
    }
  }
  Value *Next =
      Builder.CreateAdd(IndVarPHI, ConstantInt::get(IndVarTy, 1),
                        "omp_" + Name + ".next", /*HasNUW=*/true, HasNSW);
  Builder.CreateBr(Header);
  IndVarPHI->addIncoming(Next, Latch);
 
  Builder.SetInsertPoint(Exit);
  Builder.CreateBr(After);
 
  // Remember and return the canonical control flow.
  LoopInfos.emplace_front();
  CanonicalLoopInfo *CL = &LoopInfos.front();
 
  CL->Header = Header;
  CL->Cond = Cond;
  CL->Latch = Latch;
  CL->Exit = Exit;
 
#ifndef NDEBUG
  CL->assertOK();
#endif
  return CL;
}
 
Expected<CanonicalLoopInfo *>
OpenMPIRBuilder::createCanonicalLoop(const LocationDescription &Loc,
                                     LoopBodyGenCallbackTy BodyGenCB,
                                     Value *TripCount, const Twine &Name) {
  BasicBlock *BB = Loc.IP.getBlock();
  BasicBlock *NextBB = BB->getNextNode();
 
  CanonicalLoopInfo *CL = createLoopSkeleton(Loc.DL, TripCount, BB->getParent(),
                                             NextBB, NextBB, Name);
  BasicBlock *After = CL->getAfter();
 
  // If location is not set, don't connect the loop.
  if (updateToLocation(Loc)) {
    // Split the loop at the insertion point: Branch to the preheader and move
    // every following instruction to after the loop (the After BB). Also, the
    // new successor is the loop's after block.
    spliceBB(Builder, After, /*CreateBranch=*/false);
    Builder.CreateBr(CL->getPreheader());
  }
 
  // Emit the body content. We do it after connecting the loop to the CFG to
  // avoid that the callback encounters degenerate BBs.
  if (Error Err = BodyGenCB(CL->getBodyIP(), CL->getIndVar()))
    return Err;
 
#ifndef NDEBUG
  CL->assertOK();
#endif
  return CL;
}
 
Expected<ScanInfo *> OpenMPIRBuilder::scanInfoInitialize() {
  ScanInfos.emplace_front();
  ScanInfo *Result = &ScanInfos.front();
  return Result;
}
 
Expected<SmallVector<llvm::CanonicalLoopInfo *>>
OpenMPIRBuilder::createCanonicalScanLoops(
    const LocationDescription &Loc, LoopBodyGenCallbackTy BodyGenCB,
    Value *Start, Value *Stop, Value *Step, bool IsSigned, bool InclusiveStop,
    InsertPointTy ComputeIP, const Twine &Name, ScanInfo *ScanRedInfo) {
  LocationDescription ComputeLoc =
      ComputeIP.isSet() ? LocationDescription(ComputeIP, Loc.DL) : Loc;
  updateToLocation(ComputeLoc);
 
  SmallVector<CanonicalLoopInfo *> Result;
 
  Value *TripCount = calculateCanonicalLoopTripCount(
      ComputeLoc, Start, Stop, Step, IsSigned, InclusiveStop, Name);
  ScanRedInfo->Span = TripCount;
  ScanRedInfo->OMPScanInit = splitBB(Builder, true, "scan.init");
  Builder.SetInsertPoint(ScanRedInfo->OMPScanInit);
 
  auto BodyGen = [=](InsertPointTy CodeGenIP, Value *IV) {
    Builder.restoreIP(CodeGenIP);
    ScanRedInfo->IV = IV;
    createScanBBs(ScanRedInfo);
    BasicBlock *InputBlock = Builder.GetInsertBlock();
    Instruction *Terminator = InputBlock->getTerminator();
    assert(Terminator->getNumSuccessors() == 1);
    BasicBlock *ContinueBlock = Terminator->getSuccessor(0);
    Terminator->setSuccessor(0, ScanRedInfo->OMPScanDispatch);
    emitBlock(ScanRedInfo->OMPBeforeScanBlock,
              Builder.GetInsertBlock()->getParent());
    Builder.CreateBr(ScanRedInfo->OMPScanLoopExit);
    emitBlock(ScanRedInfo->OMPScanLoopExit,
              Builder.GetInsertBlock()->getParent());
    Builder.CreateBr(ContinueBlock);
    Builder.SetInsertPoint(
        ScanRedInfo->OMPBeforeScanBlock->getFirstInsertionPt());
    return BodyGenCB(Builder.saveIP(), IV);
  };
 
  const auto &&InputLoopGen = [&]() -> Error {
    Expected<CanonicalLoopInfo *> LoopInfo = createCanonicalLoop(
        Builder.saveIP(), BodyGen, Start, Stop, Step, IsSigned, InclusiveStop,
        ComputeIP, Name, true, ScanRedInfo);
    if (!LoopInfo)
      return LoopInfo.takeError();
    Result.push_back(*LoopInfo);
    Builder.restoreIP((*LoopInfo)->getAfterIP());
    return Error::success();
  };
  const auto &&ScanLoopGen = [&](LocationDescription Loc) -> Error {
    Expected<CanonicalLoopInfo *> LoopInfo =
        createCanonicalLoop(Loc, BodyGen, Start, Stop, Step, IsSigned,
                            InclusiveStop, ComputeIP, Name, true, ScanRedInfo);
    if (!LoopInfo)
      return LoopInfo.takeError();
    Result.push_back(*LoopInfo);
    Builder.restoreIP((*LoopInfo)->getAfterIP());
    ScanRedInfo->OMPScanFinish = Builder.GetInsertBlock();
    return Error::success();
  };
  Error Err = emitScanBasedDirectiveIR(InputLoopGen, ScanLoopGen, ScanRedInfo);
  if (Err)
    return Err;
  return Result;
}
 
Value *OpenMPIRBuilder::calculateCanonicalLoopTripCount(
    const LocationDescription &Loc, Value *Start, Value *Stop, Value *Step,
    bool IsSigned, bool InclusiveStop, const Twine &Name) {
 
  // Consider the following difficulties (assuming 8-bit signed integers):
  //  * Adding \p Step to the loop counter which passes \p Stop may overflow:
  //      DO I = 1, 100, 50
  ///  * A \p Step of INT_MIN cannot not be normalized to a positive direction:
  //      DO I = 100, 0, -128
 
  // Start, Stop and Step must be of the same integer type.
  auto *IndVarTy = cast<IntegerType>(Start->getType());
  assert(IndVarTy == Stop->getType() && "Stop type mismatch");
  assert(IndVarTy == Step->getType() && "Step type mismatch");
 
  updateToLocation(Loc);
 
  ConstantInt *Zero = ConstantInt::get(IndVarTy, 0);
  ConstantInt *One = ConstantInt::get(IndVarTy, 1);
 
  // Like Step, but always positive.
  Value *Incr = Step;
 
  // Distance between Start and Stop; always positive.
  Value *Span;
 
  // Condition whether there are no iterations are executed at all, e.g. because
  // UB < LB.
  Value *ZeroCmp;
 
  if (IsSigned) {
    // Ensure that increment is positive. If not, negate and invert LB and UB.
    Value *IsNeg = Builder.CreateICmpSLT(Step, Zero);
    Incr = Builder.CreateSelect(IsNeg, Builder.CreateNeg(Step), Step);
    Value *LB = Builder.CreateSelect(IsNeg, Stop, Start);
    Value *UB = Builder.CreateSelect(IsNeg, Start, Stop);
    Span = Builder.CreateSub(UB, LB, "", false, true);
    ZeroCmp = Builder.CreateICmp(
        InclusiveStop ? CmpInst::ICMP_SLT : CmpInst::ICMP_SLE, UB, LB);
  } else {
    Span = Builder.CreateSub(Stop, Start, "", true);
    ZeroCmp = Builder.CreateICmp(
        InclusiveStop ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE, Stop, Start);
  }
 
  Value *CountIfLooping;
  if (InclusiveStop) {
    CountIfLooping = Builder.CreateAdd(Builder.CreateUDiv(Span, Incr), One);
  } else {
    // Avoid incrementing past stop since it could overflow.
    Value *CountIfTwo = Builder.CreateAdd(
        Builder.CreateUDiv(Builder.CreateSub(Span, One), Incr), One);
    Value *OneCmp = Builder.CreateICmp(CmpInst::ICMP_ULE, Span, Incr);
    CountIfLooping = Builder.CreateSelect(OneCmp, One, CountIfTwo);
  }
 
  return Builder.CreateSelect(ZeroCmp, Zero, CountIfLooping,
                              "omp_" + Name + ".tripcount");
}
 
Expected<CanonicalLoopInfo *> OpenMPIRBuilder::createCanonicalLoop(
    const LocationDescription &Loc, LoopBodyGenCallbackTy BodyGenCB,
    Value *Start, Value *Stop, Value *Step, bool IsSigned, bool InclusiveStop,
    InsertPointTy ComputeIP, const Twine &Name, bool InScan,
    ScanInfo *ScanRedInfo) {
  LocationDescription ComputeLoc =
      ComputeIP.isSet() ? LocationDescription(ComputeIP, Loc.DL) : Loc;
 
  Value *TripCount = calculateCanonicalLoopTripCount(
      ComputeLoc, Start, Stop, Step, IsSigned, InclusiveStop, Name);
 
  auto BodyGen = [=](InsertPointTy CodeGenIP, Value *IV) {
    Builder.restoreIP(CodeGenIP);
    Value *Span = Builder.CreateMul(IV, Step, "", /*HasNUW=*/false,
                                    /*HasNSW=*/Config.hasNoSignedWrap());
    Value *IndVar = Builder.CreateAdd(Span, Start, "", /*HasNUW=*/false,
                                      /*HasNSW=*/Config.hasNoSignedWrap());
    if (InScan)
      ScanRedInfo->IV = IndVar;
    return BodyGenCB(Builder.saveIP(), IndVar);
  };
  LocationDescription LoopLoc =
      ComputeIP.isSet()
          ? Loc
          : LocationDescription(Builder.saveIP(),
                                Builder.getCurrentDebugLocation());
  return createCanonicalLoop(LoopLoc, BodyGen, TripCount, Name);
}
 
// Returns an LLVM function to call for initializing loop bounds using OpenMP
// static scheduling for composite `distribute parallel for` depending on
// `type`. Only i32 and i64 are supported by the runtime. Always interpret
// integers as unsigned similarly to CanonicalLoopInfo.
static FunctionCallee
getKmpcDistForStaticInitForType(Type *Ty, Module &M,
                                OpenMPIRBuilder &OMPBuilder) {
  unsigned Bitwidth = Ty->getIntegerBitWidth();
  if (Bitwidth == 32)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_dist_for_static_init_4u);
  if (Bitwidth == 64)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_dist_for_static_init_8u);
  llvm_unreachable("unknown OpenMP loop iterator bitwidth");
}
 
// Returns an LLVM function to call for initializing loop bounds using OpenMP
// static scheduling depending on `type`. Only i32 and i64 are supported by the
// runtime. Always interpret integers as unsigned similarly to
// CanonicalLoopInfo.
static FunctionCallee getKmpcForStaticInitForType(Type *Ty, Module &M,
                                                  OpenMPIRBuilder &OMPBuilder) {
  unsigned Bitwidth = Ty->getIntegerBitWidth();
  if (Bitwidth == 32)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_4u);
  if (Bitwidth == 64)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_8u);
  llvm_unreachable("unknown OpenMP loop iterator bitwidth");
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::applyStaticWorkshareLoop(
    DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP,
    WorksharingLoopType LoopType, bool NeedsBarrier, bool HasDistSchedule,
    OMPScheduleType DistScheduleSchedType) {
  assert(CLI->isValid() && "Requires a valid canonical loop");
  assert(!isConflictIP(AllocaIP, CLI->getPreheaderIP()) &&
         "Require dedicated allocate IP");
 
  // Set up the source location value for OpenMP runtime.
  Builder.restoreIP(CLI->getPreheaderIP());
  Builder.SetCurrentDebugLocation(DL);
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize);
  IdentFlag Flag = IdentFlag(0);
  switch (LoopType) {
  case WorksharingLoopType::ForStaticLoop:
    Flag = OMP_IDENT_FLAG_WORK_LOOP;
    break;
  case WorksharingLoopType::DistributeStaticLoop:
    Flag = OMP_IDENT_FLAG_WORK_DISTRIBUTE;
    break;
  case WorksharingLoopType::DistributeForStaticLoop:
    Flag = OMP_IDENT_FLAG_WORK_DISTRIBUTE | OMP_IDENT_FLAG_WORK_LOOP;
    break;
  }
  Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize, Flag);
 
  // Declare useful OpenMP runtime functions.
  Value *IV = CLI->getIndVar();
  Type *IVTy = IV->getType();
  FunctionCallee StaticInit =
      LoopType == WorksharingLoopType::DistributeForStaticLoop
          ? getKmpcDistForStaticInitForType(IVTy, M, *this)
          : getKmpcForStaticInitForType(IVTy, M, *this);
  FunctionCallee StaticFini =
      getOrCreateRuntimeFunction(M, omp::OMPRTL___kmpc_for_static_fini);
 
  // Allocate space for computed loop bounds as expected by the "init" function.
  Builder.SetInsertPoint(AllocaIP.getBlock()->getFirstNonPHIOrDbgOrAlloca());
 
  Type *I32Type = Type::getInt32Ty(M.getContext());
  Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter");
  Value *PLowerBound = Builder.CreateAlloca(IVTy, nullptr, "p.lowerbound");
  Value *PUpperBound = Builder.CreateAlloca(IVTy, nullptr, "p.upperbound");
  Value *PStride = Builder.CreateAlloca(IVTy, nullptr, "p.stride");
  CLI->setLastIter(PLastIter);
 
  // At the end of the preheader, prepare for calling the "init" function by
  // storing the current loop bounds into the allocated space. A canonical loop
  // always iterates from 0 to trip-count with step 1. Note that "init" expects
  // and produces an inclusive upper bound.
  Builder.SetInsertPoint(CLI->getPreheader()->getTerminator());
  Constant *Zero = ConstantInt::get(IVTy, 0);
  Constant *One = ConstantInt::get(IVTy, 1);
  Builder.CreateStore(Zero, PLowerBound);
  Value *UpperBound = Builder.CreateSub(CLI->getTripCount(), One);
  Builder.CreateStore(UpperBound, PUpperBound);
  Builder.CreateStore(One, PStride);
 
  Value *ThreadNum =
      getOrCreateThreadID(getOrCreateIdent(SrcLocStr, SrcLocStrSize));
 
  OMPScheduleType SchedType =
      (LoopType == WorksharingLoopType::DistributeStaticLoop)
          ? OMPScheduleType::OrderedDistribute
          : OMPScheduleType::UnorderedStatic;
  Constant *SchedulingType =
      ConstantInt::get(I32Type, static_cast<int>(SchedType));
 
  // Call the "init" function and update the trip count of the loop with the
  // value it produced.
  auto BuildInitCall = [LoopType, SrcLoc, ThreadNum, PLastIter, PLowerBound,
                        PUpperBound, IVTy, PStride, One, Zero, StaticInit,
                        this](Value *SchedulingType, auto &Builder) {
    SmallVector<Value *, 10> Args({SrcLoc, ThreadNum, SchedulingType, PLastIter,
                                   PLowerBound, PUpperBound});
    if (LoopType == WorksharingLoopType::DistributeForStaticLoop) {
      Value *PDistUpperBound =
          Builder.CreateAlloca(IVTy, nullptr, "p.distupperbound");
      Args.push_back(PDistUpperBound);
    }
    Args.append({PStride, One, Zero});
    createRuntimeFunctionCall(StaticInit, Args);
  };
  BuildInitCall(SchedulingType, Builder);
  if (HasDistSchedule &&
      LoopType != WorksharingLoopType::DistributeStaticLoop) {
    Constant *DistScheduleSchedType = ConstantInt::get(
        I32Type, static_cast<int>(omp::OMPScheduleType::OrderedDistribute));
    // We want to emit a second init function call for the dist_schedule clause
    // to the Distribute construct. This should only be done however if a
    // Workshare Loop is nested within a Distribute Construct
    BuildInitCall(DistScheduleSchedType, Builder);
  }
  Value *LowerBound = Builder.CreateLoad(IVTy, PLowerBound);
  Value *InclusiveUpperBound = Builder.CreateLoad(IVTy, PUpperBound);
  Value *TripCountMinusOne = Builder.CreateSub(InclusiveUpperBound, LowerBound);
  Value *TripCount = Builder.CreateAdd(TripCountMinusOne, One);
  CLI->setTripCount(TripCount);
 
  // Update all uses of the induction variable except the one in the condition
  // block that compares it with the actual upper bound, and the increment in
  // the latch block.
 
  CLI->mapIndVar([&](Instruction *OldIV) -> Value * {
    Builder.SetInsertPoint(CLI->getBody(),
                           CLI->getBody()->getFirstInsertionPt());
    Builder.SetCurrentDebugLocation(DL);
    return Builder.CreateAdd(OldIV, LowerBound, "", /*HasNUW=*/false,
                             /*HasNSW=*/Config.hasNoSignedWrap());
  });
 
  // In the "exit" block, call the "fini" function.
  Builder.SetInsertPoint(CLI->getExit(),
                         CLI->getExit()->getTerminator()->getIterator());
  createRuntimeFunctionCall(StaticFini, {SrcLoc, ThreadNum});
 
  // Add the barrier if requested.
  if (NeedsBarrier) {
    InsertPointOrErrorTy BarrierIP =
        createBarrier(LocationDescription(Builder.saveIP(), DL),
                      omp::Directive::OMPD_for, /* ForceSimpleCall */ false,
                      /* CheckCancelFlag */ false);
    if (!BarrierIP)
      return BarrierIP.takeError();
  }
 
  InsertPointTy AfterIP = CLI->getAfterIP();
  CLI->invalidate();
 
  return AfterIP;
}
 
static void addAccessGroupMetadata(BasicBlock *Block, MDNode *AccessGroup,
                                   LoopInfo &LI);
static void addLoopMetadata(CanonicalLoopInfo *Loop,
                            ArrayRef<Metadata *> Properties);
 
static void applyParallelAccessesMetadata(CanonicalLoopInfo *CLI,
                                          LLVMContext &Ctx, Loop *Loop,
                                          LoopInfo &LoopInfo,
                                          SmallVector<Metadata *> &LoopMDList) {
  SmallSet<BasicBlock *, 8> Reachable;
 
  // Get the basic blocks from the loop in which memref instructions
  // can be found.
  // TODO: Generalize getting all blocks inside a CanonicalizeLoopInfo,
  // preferably without running any passes.
  for (BasicBlock *Block : Loop->getBlocks()) {
    if (Block == CLI->getCond() || Block == CLI->getHeader())
      continue;
    Reachable.insert(Block);
  }
 
  // Add access group metadata to memory-access instructions.
  MDNode *AccessGroup = MDNode::getDistinct(Ctx, {});
  for (BasicBlock *BB : Reachable)
    addAccessGroupMetadata(BB, AccessGroup, LoopInfo);
  // TODO:  If the loop has existing parallel access metadata, have
  // to combine two lists.
  LoopMDList.push_back(MDNode::get(
      Ctx, {MDString::get(Ctx, "llvm.loop.parallel_accesses"), AccessGroup}));
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::applyStaticChunkedWorkshareLoop(
    DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP,
    bool NeedsBarrier, Value *ChunkSize, OMPScheduleType SchedType,
    Value *DistScheduleChunkSize, OMPScheduleType DistScheduleSchedType) {
  assert(CLI->isValid() && "Requires a valid canonical loop");
  assert((ChunkSize || DistScheduleChunkSize) && "Chunk size is required");
 
  LLVMContext &Ctx = CLI->getFunction()->getContext();
  Value *IV = CLI->getIndVar();
  Value *OrigTripCount = CLI->getTripCount();
  Type *IVTy = IV->getType();
  assert(IVTy->getIntegerBitWidth() <= 64 &&
         "Max supported tripcount bitwidth is 64 bits");
  Type *InternalIVTy = IVTy->getIntegerBitWidth() <= 32 ? Type::getInt32Ty(Ctx)
                                                        : Type::getInt64Ty(Ctx);
  Type *I32Type = Type::getInt32Ty(M.getContext());
  Constant *Zero = ConstantInt::get(InternalIVTy, 0);
  Constant *One = ConstantInt::get(InternalIVTy, 1);
 
  Function *F = CLI->getFunction();
  // Blocks must have terminators.
  // FIXME: Don't run analyses on incomplete/invalid IR.
  SmallVector<Instruction *> UIs;
  for (BasicBlock &BB : *F)
    if (!BB.hasTerminator())
      UIs.push_back(new UnreachableInst(F->getContext(), &BB));
  FunctionAnalysisManager FAM;
  FAM.registerPass([]() { return DominatorTreeAnalysis(); });
  FAM.registerPass([]() { return PassInstrumentationAnalysis(); });
  LoopAnalysis LIA;
  LoopInfo &&LI = LIA.run(*F, FAM);
  for (Instruction *I : UIs)
    I->eraseFromParent();
  Loop *L = LI.getLoopFor(CLI->getHeader());
  SmallVector<Metadata *> LoopMDList;
  if (ChunkSize || DistScheduleChunkSize)
    applyParallelAccessesMetadata(CLI, Ctx, L, LI, LoopMDList);
  addLoopMetadata(CLI, LoopMDList);
 
  // Declare useful OpenMP runtime functions.
  FunctionCallee StaticInit =
      getKmpcForStaticInitForType(InternalIVTy, M, *this);
  FunctionCallee StaticFini =
      getOrCreateRuntimeFunction(M, omp::OMPRTL___kmpc_for_static_fini);
 
  // Allocate space for computed loop bounds as expected by the "init" function.
  Builder.restoreIP(AllocaIP);
  Builder.SetCurrentDebugLocation(DL);
  Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter");
  Value *PLowerBound =
      Builder.CreateAlloca(InternalIVTy, nullptr, "p.lowerbound");
  Value *PUpperBound =
      Builder.CreateAlloca(InternalIVTy, nullptr, "p.upperbound");
  Value *PStride = Builder.CreateAlloca(InternalIVTy, nullptr, "p.stride");
  CLI->setLastIter(PLastIter);
 
  // Set up the source location value for the OpenMP runtime.
  Builder.restoreIP(CLI->getPreheaderIP());
  Builder.SetCurrentDebugLocation(DL);
 
  // TODO: Detect overflow in ubsan or max-out with current tripcount.
  Value *CastedChunkSize = Builder.CreateZExtOrTrunc(
      ChunkSize ? ChunkSize : Zero, InternalIVTy, "chunksize");
  Value *CastedDistScheduleChunkSize = Builder.CreateZExtOrTrunc(
      DistScheduleChunkSize ? DistScheduleChunkSize : Zero, InternalIVTy,
      "distschedulechunksize");
  Value *CastedTripCount =
      Builder.CreateZExt(OrigTripCount, InternalIVTy, "tripcount");
 
  Constant *SchedulingType =
      ConstantInt::get(I32Type, static_cast<int>(SchedType));
  Constant *DistSchedulingType =
      ConstantInt::get(I32Type, static_cast<int>(DistScheduleSchedType));
  Builder.CreateStore(Zero, PLowerBound);
  Value *OrigUpperBound = Builder.CreateSub(CastedTripCount, One);
  Value *IsTripCountZero = Builder.CreateICmpEQ(CastedTripCount, Zero);
  Value *UpperBound =
      Builder.CreateSelect(IsTripCountZero, Zero, OrigUpperBound);
  Builder.CreateStore(UpperBound, PUpperBound);
  Builder.CreateStore(One, PStride);
 
  // Call the "init" function and update the trip count of the loop with the
  // value it produced.
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize);
  IdentFlag Flag = OMP_IDENT_FLAG_WORK_LOOP;
  if (DistScheduleSchedType != OMPScheduleType::None) {
    Flag |= OMP_IDENT_FLAG_WORK_DISTRIBUTE;
  }
  Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize, Flag);
  Value *ThreadNum =
      getOrCreateThreadID(getOrCreateIdent(SrcLocStr, SrcLocStrSize));
  auto BuildInitCall = [StaticInit, SrcLoc, ThreadNum, PLastIter, PLowerBound,
                        PUpperBound, PStride, One,
                        this](Value *SchedulingType, Value *ChunkSize,
                              auto &Builder) {
    createRuntimeFunctionCall(
        StaticInit, {/*loc=*/SrcLoc, /*global_tid=*/ThreadNum,
                     /*schedtype=*/SchedulingType, /*plastiter=*/PLastIter,
                     /*plower=*/PLowerBound, /*pupper=*/PUpperBound,
                     /*pstride=*/PStride, /*incr=*/One,
                     /*chunk=*/ChunkSize});
  };
  BuildInitCall(SchedulingType, CastedChunkSize, Builder);
  if (DistScheduleSchedType != OMPScheduleType::None &&
      SchedType != OMPScheduleType::OrderedDistributeChunked &&
      SchedType != OMPScheduleType::OrderedDistribute) {
    // We want to emit a second init function call for the dist_schedule clause
    // to the Distribute construct. This should only be done however if a
    // Workshare Loop is nested within a Distribute Construct
    BuildInitCall(DistSchedulingType, CastedDistScheduleChunkSize, Builder);
  }
 
  // Load values written by the "init" function.
  Value *FirstChunkStart =
      Builder.CreateLoad(InternalIVTy, PLowerBound, "omp_firstchunk.lb");
  Value *FirstChunkStop =
      Builder.CreateLoad(InternalIVTy, PUpperBound, "omp_firstchunk.ub");
  Value *FirstChunkEnd = Builder.CreateAdd(FirstChunkStop, One);
  Value *ChunkRange =
      Builder.CreateSub(FirstChunkEnd, FirstChunkStart, "omp_chunk.range");
  Value *NextChunkStride =
      Builder.CreateLoad(InternalIVTy, PStride, "omp_dispatch.stride");
 
  // Create outer "dispatch" loop for enumerating the chunks.
  BasicBlock *DispatchEnter = splitBB(Builder, true);
  Value *DispatchCounter;
 
  // It is safe to assume this didn't return an error because the callback
  // passed into createCanonicalLoop is the only possible error source, and it
  // always returns success.
  CanonicalLoopInfo *DispatchCLI = cantFail(createCanonicalLoop(
      {Builder.saveIP(), DL},
      [&](InsertPointTy BodyIP, Value *Counter) {
        DispatchCounter = Counter;
        return Error::success();
      },
      FirstChunkStart, CastedTripCount, NextChunkStride,
      /*IsSigned=*/false, /*InclusiveStop=*/false, /*ComputeIP=*/{},
      "dispatch"));
 
  // Remember the BasicBlocks of the dispatch loop we need, then invalidate to
  // not have to preserve the canonical invariant.
  BasicBlock *DispatchBody = DispatchCLI->getBody();
  BasicBlock *DispatchLatch = DispatchCLI->getLatch();
  BasicBlock *DispatchExit = DispatchCLI->getExit();
  BasicBlock *DispatchAfter = DispatchCLI->getAfter();
  DispatchCLI->invalidate();
 
  // Rewire the original loop to become the chunk loop inside the dispatch loop.
  redirectTo(DispatchAfter, CLI->getAfter(), DL);
  redirectTo(CLI->getExit(), DispatchLatch, DL);
  redirectTo(DispatchBody, DispatchEnter, DL);
 
  // Prepare the prolog of the chunk loop.
  Builder.restoreIP(CLI->getPreheaderIP());
  Builder.SetCurrentDebugLocation(DL);
 
  // Compute the number of iterations of the chunk loop.
  Builder.SetInsertPoint(CLI->getPreheader()->getTerminator());
  Value *ChunkEnd = Builder.CreateAdd(DispatchCounter, ChunkRange);
  Value *IsLastChunk =
      Builder.CreateICmpUGE(ChunkEnd, CastedTripCount, "omp_chunk.is_last");
  Value *CountUntilOrigTripCount =
      Builder.CreateSub(CastedTripCount, DispatchCounter);
  Value *ChunkTripCount = Builder.CreateSelect(
      IsLastChunk, CountUntilOrigTripCount, ChunkRange, "omp_chunk.tripcount");
  Value *BackcastedChunkTC =
      Builder.CreateTrunc(ChunkTripCount, IVTy, "omp_chunk.tripcount.trunc");
  CLI->setTripCount(BackcastedChunkTC);
 
  // Update all uses of the induction variable except the one in the condition
  // block that compares it with the actual upper bound, and the increment in
  // the latch block.
  Value *BackcastedDispatchCounter =
      Builder.CreateTrunc(DispatchCounter, IVTy, "omp_dispatch.iv.trunc");
  CLI->mapIndVar([&](Instruction *) -> Value * {
    Builder.restoreIP(CLI->getBodyIP());
    return Builder.CreateAdd(IV, BackcastedDispatchCounter);
  });
 
  // In the "exit" block, call the "fini" function.
  Builder.SetInsertPoint(DispatchExit, DispatchExit->getFirstInsertionPt());
  createRuntimeFunctionCall(StaticFini, {SrcLoc, ThreadNum});
 
  // Add the barrier if requested.
  if (NeedsBarrier) {
    InsertPointOrErrorTy AfterIP =
        createBarrier(LocationDescription(Builder.saveIP(), DL), OMPD_for,
                      /*ForceSimpleCall=*/false, /*CheckCancelFlag=*/false);
    if (!AfterIP)
      return AfterIP.takeError();
  }
 
#ifndef NDEBUG
  // Even though we currently do not support applying additional methods to it,
  // the chunk loop should remain a canonical loop.
  CLI->assertOK();
#endif
 
  return InsertPointTy(DispatchAfter, DispatchAfter->getFirstInsertionPt());
}
 
// Returns an LLVM function to call for executing an OpenMP static worksharing
// for loop depending on `type`. Only i32 and i64 are supported by the runtime.
// Always interpret integers as unsigned similarly to CanonicalLoopInfo.
static FunctionCallee
getKmpcForStaticLoopForType(Type *Ty, OpenMPIRBuilder *OMPBuilder,
                            WorksharingLoopType LoopType) {
  unsigned Bitwidth = Ty->getIntegerBitWidth();
  Module &M = OMPBuilder->M;
  switch (LoopType) {
  case WorksharingLoopType::ForStaticLoop:
    if (Bitwidth == 32)
      return OMPBuilder->getOrCreateRuntimeFunction(
          M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_loop_4u);
    if (Bitwidth == 64)
      return OMPBuilder->getOrCreateRuntimeFunction(
          M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_loop_8u);
    break;
  case WorksharingLoopType::DistributeStaticLoop:
    if (Bitwidth == 32)
      return OMPBuilder->getOrCreateRuntimeFunction(
          M, omp::RuntimeFunction::OMPRTL___kmpc_distribute_static_loop_4u);
    if (Bitwidth == 64)
      return OMPBuilder->getOrCreateRuntimeFunction(
          M, omp::RuntimeFunction::OMPRTL___kmpc_distribute_static_loop_8u);
    break;
  case WorksharingLoopType::DistributeForStaticLoop:
    if (Bitwidth == 32)
      return OMPBuilder->getOrCreateRuntimeFunction(
          M, omp::RuntimeFunction::OMPRTL___kmpc_distribute_for_static_loop_4u);
    if (Bitwidth == 64)
      return OMPBuilder->getOrCreateRuntimeFunction(
          M, omp::RuntimeFunction::OMPRTL___kmpc_distribute_for_static_loop_8u);
    break;
  }
  if (Bitwidth != 32 && Bitwidth != 64) {
    llvm_unreachable("Unknown OpenMP loop iterator bitwidth");
  }
  llvm_unreachable("Unknown type of OpenMP worksharing loop");
}
 
// Inserts a call to proper OpenMP Device RTL function which handles
// loop worksharing.
static void createTargetLoopWorkshareCall(OpenMPIRBuilder *OMPBuilder,
                                          WorksharingLoopType LoopType,
                                          BasicBlock *InsertBlock, Value *Ident,
                                          Value *LoopBodyArg, Value *TripCount,
                                          Function &LoopBodyFn, bool NoLoop) {
  Type *TripCountTy = TripCount->getType();
  Module &M = OMPBuilder->M;
  IRBuilder<> &Builder = OMPBuilder->Builder;
  FunctionCallee RTLFn =
      getKmpcForStaticLoopForType(TripCountTy, OMPBuilder, LoopType);
  SmallVector<Value *, 8> RealArgs;
  RealArgs.push_back(Ident);
  RealArgs.push_back(&LoopBodyFn);
  RealArgs.push_back(LoopBodyArg);
  RealArgs.push_back(TripCount);
  if (LoopType == WorksharingLoopType::DistributeStaticLoop) {
    RealArgs.push_back(ConstantInt::get(TripCountTy, 0));
    RealArgs.push_back(ConstantInt::get(Builder.getInt8Ty(), 0));
    Builder.restoreIP({InsertBlock, std::prev(InsertBlock->end())});
    OMPBuilder->createRuntimeFunctionCall(RTLFn, RealArgs);
    return;
  }
  FunctionCallee RTLNumThreads = OMPBuilder->getOrCreateRuntimeFunction(
      M, omp::RuntimeFunction::OMPRTL_omp_get_num_threads);
  Builder.restoreIP({InsertBlock, std::prev(InsertBlock->end())});
  Value *NumThreads = OMPBuilder->createRuntimeFunctionCall(RTLNumThreads, {});
 
  RealArgs.push_back(
      Builder.CreateZExtOrTrunc(NumThreads, TripCountTy, "num.threads.cast"));
  RealArgs.push_back(ConstantInt::get(TripCountTy, 0));
  if (LoopType == WorksharingLoopType::DistributeForStaticLoop) {
    RealArgs.push_back(ConstantInt::get(TripCountTy, 0));
    RealArgs.push_back(ConstantInt::get(Builder.getInt8Ty(), NoLoop));
  } else {
    RealArgs.push_back(ConstantInt::get(Builder.getInt8Ty(), 0));
  }
 
  OMPBuilder->createRuntimeFunctionCall(RTLFn, RealArgs);
}
 
static void workshareLoopTargetCallback(
    OpenMPIRBuilder *OMPIRBuilder, CanonicalLoopInfo *CLI, Value *Ident,
    Function &OutlinedFn, const SmallVector<Instruction *, 4> &ToBeDeleted,
    WorksharingLoopType LoopType, bool NoLoop) {
  IRBuilder<> &Builder = OMPIRBuilder->Builder;
  BasicBlock *Preheader = CLI->getPreheader();
  Value *TripCount = CLI->getTripCount();
 
  // After loop body outling, the loop body contains only set up
  // of loop body argument structure and the call to the outlined
  // loop body function. Firstly, we need to move setup of loop body args
  // into loop preheader.
  Preheader->splice(std::prev(Preheader->end()), CLI->getBody(),
                    CLI->getBody()->begin(), std::prev(CLI->getBody()->end()));
 
  // The next step is to remove the whole loop. We do not it need anymore.
  // That's why make an unconditional branch from loop preheader to loop
  // exit block
  Builder.restoreIP({Preheader, Preheader->end()});
  Builder.SetCurrentDebugLocation(Preheader->getTerminator()->getDebugLoc());
  Preheader->getTerminator()->eraseFromParent();
  Builder.CreateBr(CLI->getExit());
 
  // Delete dead loop blocks
  OpenMPIRBuilder::OutlineInfo CleanUpInfo;
  SmallPtrSet<BasicBlock *, 32> RegionBlockSet;
  SmallVector<BasicBlock *, 32> BlocksToBeRemoved;
  CleanUpInfo.EntryBB = CLI->getHeader();
  CleanUpInfo.ExitBB = CLI->getExit();
  CleanUpInfo.collectBlocks(RegionBlockSet, BlocksToBeRemoved);
  DeleteDeadBlocks(BlocksToBeRemoved);
 
  // Find the instruction which corresponds to loop body argument structure
  // and remove the call to loop body function instruction.
  Value *LoopBodyArg;
  User *OutlinedFnUser = OutlinedFn.getUniqueUndroppableUser();
  assert(OutlinedFnUser &&
         "Expected unique undroppable user of outlined function");
  CallInst *OutlinedFnCallInstruction = dyn_cast<CallInst>(OutlinedFnUser);
  assert(OutlinedFnCallInstruction && "Expected outlined function call");
  assert((OutlinedFnCallInstruction->getParent() == Preheader) &&
         "Expected outlined function call to be located in loop preheader");
  // Check in case no argument structure has been passed.
  if (OutlinedFnCallInstruction->arg_size() > 1)
    LoopBodyArg = OutlinedFnCallInstruction->getArgOperand(1);
  else
    LoopBodyArg = Constant::getNullValue(Builder.getPtrTy());
  OutlinedFnCallInstruction->eraseFromParent();
 
  createTargetLoopWorkshareCall(OMPIRBuilder, LoopType, Preheader, Ident,
                                LoopBodyArg, TripCount, OutlinedFn, NoLoop);
 
  for (auto &ToBeDeletedItem : ToBeDeleted)
    ToBeDeletedItem->eraseFromParent();
  CLI->invalidate();
}
 
OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyWorkshareLoopTarget(
    DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP,
    WorksharingLoopType LoopType, bool NoLoop) {
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize);
  IdentFlag Flag = IdentFlag(0);
  switch (LoopType) {
  case WorksharingLoopType::ForStaticLoop:
    Flag = OMP_IDENT_FLAG_WORK_LOOP;
    break;
  case WorksharingLoopType::DistributeStaticLoop:
    Flag = OMP_IDENT_FLAG_WORK_DISTRIBUTE;
    break;
  case WorksharingLoopType::DistributeForStaticLoop:
    Flag = OMP_IDENT_FLAG_WORK_DISTRIBUTE | OMP_IDENT_FLAG_WORK_LOOP;
    break;
  }
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize, Flag);
 
  auto OI = std::make_unique<OutlineInfo>();
  OI->OuterAllocBB = CLI->getPreheader();
  Function *OuterFn = CLI->getPreheader()->getParent();
 
  // Instructions which need to be deleted at the end of code generation
  SmallVector<Instruction *, 4> ToBeDeleted;
 
  OI->OuterAllocBB = AllocaIP.getBlock();
 
  // Mark the body loop as region which needs to be extracted
  OI->EntryBB = CLI->getBody();
  OI->ExitBB = CLI->getLatch()->splitBasicBlockBefore(CLI->getLatch()->begin(),
                                                      "omp.prelatch");
 
  // Prepare loop body for extraction
  Builder.restoreIP({CLI->getPreheader(), CLI->getPreheader()->begin()});
 
  // Insert new loop counter variable which will be used only in loop
  // body.
  AllocaInst *NewLoopCnt = Builder.CreateAlloca(CLI->getIndVarType(), 0, "");
  Instruction *NewLoopCntLoad =
      Builder.CreateLoad(CLI->getIndVarType(), NewLoopCnt);
  // New loop counter instructions are redundant in the loop preheader when
  // code generation for workshare loop is finshed. That's why mark them as
  // ready for deletion.
  ToBeDeleted.push_back(NewLoopCntLoad);
  ToBeDeleted.push_back(NewLoopCnt);
 
  // Analyse loop body region. Find all input variables which are used inside
  // loop body region.
  SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet;
  SmallVector<BasicBlock *, 32> Blocks;
  OI->collectBlocks(ParallelRegionBlockSet, Blocks);
 
  CodeExtractorAnalysisCache CEAC(*OuterFn);
  CodeExtractor Extractor(Blocks,
                          /* DominatorTree */ nullptr,
                          /* AggregateArgs */ true,
                          /* BlockFrequencyInfo */ nullptr,
                          /* BranchProbabilityInfo */ nullptr,
                          /* AssumptionCache */ nullptr,
                          /* AllowVarArgs */ true,
                          /* AllowAlloca */ true,
                          /* AllocationBlock */ CLI->getPreheader(),
                          /* DeallocationBlocks */ {},
                          /* Suffix */ ".omp_wsloop",
                          /* AggrArgsIn0AddrSpace */ true);
 
  BasicBlock *CommonExit = nullptr;
  SetVector<Value *> SinkingCands, HoistingCands;
 
  // Find allocas outside the loop body region which are used inside loop
  // body
  Extractor.findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit);
 
  // We need to model loop body region as the function f(cnt, loop_arg).
  // That's why we replace loop induction variable by the new counter
  // which will be one of loop body function argument
  SmallVector<User *> Users(CLI->getIndVar()->user_begin(),
                            CLI->getIndVar()->user_end());
  for (auto Use : Users) {
    if (Instruction *Inst = dyn_cast<Instruction>(Use)) {
      if (ParallelRegionBlockSet.count(Inst->getParent())) {
        Inst->replaceUsesOfWith(CLI->getIndVar(), NewLoopCntLoad);
      }
    }
  }
  // Make sure that loop counter variable is not merged into loop body
  // function argument structure and it is passed as separate variable
  OI->ExcludeArgsFromAggregate.push_back(NewLoopCntLoad);
 
  // PostOutline CB is invoked when loop body function is outlined and
  // loop body is replaced by call to outlined function. We need to add
  // call to OpenMP device rtl inside loop preheader. OpenMP device rtl
  // function will handle loop control logic.
  //
  OI->PostOutlineCB = [=, ToBeDeletedVec =
                              std::move(ToBeDeleted)](Function &OutlinedFn) {
    workshareLoopTargetCallback(this, CLI, Ident, OutlinedFn, ToBeDeletedVec,
                                LoopType, NoLoop);
  };
  addOutlineInfo(std::move(OI));
  return CLI->getAfterIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::applyWorkshareLoop(
    DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP,
    bool NeedsBarrier, omp::ScheduleKind SchedKind, Value *ChunkSize,
    bool HasSimdModifier, bool HasMonotonicModifier,
    bool HasNonmonotonicModifier, bool HasOrderedClause,
    WorksharingLoopType LoopType, bool NoLoop, bool HasDistSchedule,
    Value *DistScheduleChunkSize) {
  if (Config.isTargetDevice())
    return applyWorkshareLoopTarget(DL, CLI, AllocaIP, LoopType, NoLoop);
  OMPScheduleType EffectiveScheduleType = computeOpenMPScheduleType(
      SchedKind, ChunkSize, HasSimdModifier, HasMonotonicModifier,
      HasNonmonotonicModifier, HasOrderedClause, DistScheduleChunkSize);
 
  bool IsOrdered = (EffectiveScheduleType & OMPScheduleType::ModifierOrdered) ==
                   OMPScheduleType::ModifierOrdered;
  OMPScheduleType DistScheduleSchedType = OMPScheduleType::None;
  if (HasDistSchedule) {
    DistScheduleSchedType = DistScheduleChunkSize
                                ? OMPScheduleType::OrderedDistributeChunked
                                : OMPScheduleType::OrderedDistribute;
  }
  switch (EffectiveScheduleType & ~OMPScheduleType::ModifierMask) {
  case OMPScheduleType::BaseStatic:
  case OMPScheduleType::BaseDistribute:
    assert((!ChunkSize || !DistScheduleChunkSize) &&
           "No chunk size with static-chunked schedule");
    if (IsOrdered && !HasDistSchedule)
      return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType,
                                       NeedsBarrier, ChunkSize);
    // FIXME: Monotonicity ignored?
    if (DistScheduleChunkSize)
      return applyStaticChunkedWorkshareLoop(
          DL, CLI, AllocaIP, NeedsBarrier, ChunkSize, EffectiveScheduleType,
          DistScheduleChunkSize, DistScheduleSchedType);
    return applyStaticWorkshareLoop(DL, CLI, AllocaIP, LoopType, NeedsBarrier,
                                    HasDistSchedule);
 
  case OMPScheduleType::BaseStaticChunked:
  case OMPScheduleType::BaseDistributeChunked:
    if (IsOrdered && !HasDistSchedule)
      return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType,
                                       NeedsBarrier, ChunkSize);
    // FIXME: Monotonicity ignored?
    return applyStaticChunkedWorkshareLoop(
        DL, CLI, AllocaIP, NeedsBarrier, ChunkSize, EffectiveScheduleType,
        DistScheduleChunkSize, DistScheduleSchedType);
 
  case OMPScheduleType::BaseRuntime:
  case OMPScheduleType::BaseAuto:
  case OMPScheduleType::BaseGreedy:
  case OMPScheduleType::BaseBalanced:
  case OMPScheduleType::BaseSteal:
  case OMPScheduleType::BaseRuntimeSimd:
    assert(!ChunkSize &&
           "schedule type does not support user-defined chunk sizes");
    [[fallthrough]];
  case OMPScheduleType::BaseGuidedSimd:
  case OMPScheduleType::BaseDynamicChunked:
  case OMPScheduleType::BaseGuidedChunked:
  case OMPScheduleType::BaseGuidedIterativeChunked:
  case OMPScheduleType::BaseGuidedAnalyticalChunked:
  case OMPScheduleType::BaseStaticBalancedChunked:
    return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType,
                                     NeedsBarrier, ChunkSize);
 
  default:
    llvm_unreachable("Unknown/unimplemented schedule kind");
  }
}
 
/// Returns an LLVM function to call for initializing loop bounds using OpenMP
/// dynamic scheduling depending on `type`. Only i32 and i64 are supported by
/// the runtime. Always interpret integers as unsigned similarly to
/// CanonicalLoopInfo.
static FunctionCallee
getKmpcForDynamicInitForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) {
  unsigned Bitwidth = Ty->getIntegerBitWidth();
  if (Bitwidth == 32)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_4u);
  if (Bitwidth == 64)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_8u);
  llvm_unreachable("unknown OpenMP loop iterator bitwidth");
}
 
/// Returns an LLVM function to call for updating the next loop using OpenMP
/// dynamic scheduling depending on `type`. Only i32 and i64 are supported by
/// the runtime. Always interpret integers as unsigned similarly to
/// CanonicalLoopInfo.
static FunctionCallee
getKmpcForDynamicNextForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) {
  unsigned Bitwidth = Ty->getIntegerBitWidth();
  if (Bitwidth == 32)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_4u);
  if (Bitwidth == 64)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_8u);
  llvm_unreachable("unknown OpenMP loop iterator bitwidth");
}
 
/// Returns an LLVM function to call for finalizing the dynamic loop using
/// depending on `type`. Only i32 and i64 are supported by the runtime. Always
/// interpret integers as unsigned similarly to CanonicalLoopInfo.
static FunctionCallee
getKmpcForDynamicFiniForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) {
  unsigned Bitwidth = Ty->getIntegerBitWidth();
  if (Bitwidth == 32)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_4u);
  if (Bitwidth == 64)
    return OMPBuilder.getOrCreateRuntimeFunction(
        M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_8u);
  llvm_unreachable("unknown OpenMP loop iterator bitwidth");
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::applyDynamicWorkshareLoop(DebugLoc DL, CanonicalLoopInfo *CLI,
                                           InsertPointTy AllocaIP,
                                           OMPScheduleType SchedType,
                                           bool NeedsBarrier, Value *Chunk) {
  assert(CLI->isValid() && "Requires a valid canonical loop");
  assert(!isConflictIP(AllocaIP, CLI->getPreheaderIP()) &&
         "Require dedicated allocate IP");
  assert(isValidWorkshareLoopScheduleType(SchedType) &&
         "Require valid schedule type");
 
  bool Ordered = (SchedType & OMPScheduleType::ModifierOrdered) ==
                 OMPScheduleType::ModifierOrdered;
 
  // Set up the source location value for OpenMP runtime.
  Builder.SetCurrentDebugLocation(DL);
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize);
  Value *SrcLoc =
      getOrCreateIdent(SrcLocStr, SrcLocStrSize, OMP_IDENT_FLAG_WORK_LOOP);
 
  // Declare useful OpenMP runtime functions.
  Value *IV = CLI->getIndVar();
  Type *IVTy = IV->getType();
  FunctionCallee DynamicInit = getKmpcForDynamicInitForType(IVTy, M, *this);
  FunctionCallee DynamicNext = getKmpcForDynamicNextForType(IVTy, M, *this);
 
  // Allocate space for computed loop bounds as expected by the "init" function.
  Builder.SetInsertPoint(AllocaIP.getBlock()->getFirstNonPHIOrDbgOrAlloca());
  Type *I32Type = Type::getInt32Ty(M.getContext());
  Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter");
  Value *PLowerBound = Builder.CreateAlloca(IVTy, nullptr, "p.lowerbound");
  Value *PUpperBound = Builder.CreateAlloca(IVTy, nullptr, "p.upperbound");
  Value *PStride = Builder.CreateAlloca(IVTy, nullptr, "p.stride");
  CLI->setLastIter(PLastIter);
 
  // At the end of the preheader, prepare for calling the "init" function by
  // storing the current loop bounds into the allocated space. A canonical loop
  // always iterates from 0 to trip-count with step 1. Note that "init" expects
  // and produces an inclusive upper bound.
  BasicBlock *PreHeader = CLI->getPreheader();
  Builder.SetInsertPoint(PreHeader->getTerminator());
  Constant *One = ConstantInt::get(IVTy, 1);
  Builder.CreateStore(One, PLowerBound);
  Value *UpperBound = CLI->getTripCount();
  Builder.CreateStore(UpperBound, PUpperBound);
  Builder.CreateStore(One, PStride);
 
  BasicBlock *Header = CLI->getHeader();
  BasicBlock *Exit = CLI->getExit();
  BasicBlock *Cond = CLI->getCond();
  BasicBlock *Latch = CLI->getLatch();
  InsertPointTy AfterIP = CLI->getAfterIP();
 
  // The CLI will be "broken" in the code below, as the loop is no longer
  // a valid canonical loop.
 
  if (!Chunk)
    Chunk = One;
 
  Value *ThreadNum =
      getOrCreateThreadID(getOrCreateIdent(SrcLocStr, SrcLocStrSize));
 
  Constant *SchedulingType =
      ConstantInt::get(I32Type, static_cast<int>(SchedType));
 
  // Call the "init" function.
  createRuntimeFunctionCall(DynamicInit, {SrcLoc, ThreadNum, SchedulingType,
                                          /* LowerBound */ One, UpperBound,
                                          /* step */ One, Chunk});
 
  // An outer loop around the existing one.
  BasicBlock *OuterCond = BasicBlock::Create(
      PreHeader->getContext(), Twine(PreHeader->getName()) + ".outer.cond",
      PreHeader->getParent());
  // This needs to be 32-bit always, so can't use the IVTy Zero above.
  Builder.SetInsertPoint(OuterCond, OuterCond->getFirstInsertionPt());
  Value *Res = createRuntimeFunctionCall(
      DynamicNext,
      {SrcLoc, ThreadNum, PLastIter, PLowerBound, PUpperBound, PStride});
  Constant *Zero32 = ConstantInt::get(I32Type, 0);
  Value *MoreWork = Builder.CreateCmp(CmpInst::ICMP_NE, Res, Zero32);
  Value *LowerBound =
      Builder.CreateSub(Builder.CreateLoad(IVTy, PLowerBound), One, "lb");
  Builder.CreateCondBr(MoreWork, Header, Exit);
 
  // Change PHI-node in loop header to use outer cond rather than preheader,
  // and set IV to the LowerBound.
  Instruction *Phi = &Header->front();
  auto *PI = cast<PHINode>(Phi);
  PI->setIncomingBlock(0, OuterCond);
  PI->setIncomingValue(0, LowerBound);
 
  // Then set the pre-header to jump to the OuterCond
  Instruction *Term = PreHeader->getTerminator();
  auto *Br = cast<UncondBrInst>(Term);
  Br->setSuccessor(OuterCond);
 
  // Modify the inner condition:
  // * Use the UpperBound returned from the DynamicNext call.
  // * jump to the loop outer loop when done with one of the inner loops.
  Builder.SetInsertPoint(Cond, Cond->getFirstInsertionPt());
  UpperBound = Builder.CreateLoad(IVTy, PUpperBound, "ub");
  Instruction *Comp = &*Builder.GetInsertPoint();
  auto *CI = cast<CmpInst>(Comp);
  CI->setOperand(1, UpperBound);
  // Redirect the inner exit to branch to outer condition.
  Instruction *Branch = &Cond->back();
  auto *BI = cast<CondBrInst>(Branch);
  assert(BI->getSuccessor(1) == Exit);
  BI->setSuccessor(1, OuterCond);
 
  // Call the "fini" function if "ordered" is present in wsloop directive.
  if (Ordered) {
    Builder.SetInsertPoint(&Latch->back());
    FunctionCallee DynamicFini = getKmpcForDynamicFiniForType(IVTy, M, *this);
    createRuntimeFunctionCall(DynamicFini, {SrcLoc, ThreadNum});
  }
 
  // Add the barrier if requested.
  if (NeedsBarrier) {
    Builder.SetInsertPoint(&Exit->back());
    InsertPointOrErrorTy BarrierIP =
        createBarrier(LocationDescription(Builder.saveIP(), DL),
                      omp::Directive::OMPD_for, /* ForceSimpleCall */ false,
                      /* CheckCancelFlag */ false);
    if (!BarrierIP)
      return BarrierIP.takeError();
  }
 
  CLI->invalidate();
  return AfterIP;
}
 
/// Redirect all edges that branch to \p OldTarget to \p NewTarget. That is,
/// after this \p OldTarget will be orphaned.
static void redirectAllPredecessorsTo(BasicBlock *OldTarget,
                                      BasicBlock *NewTarget, DebugLoc DL) {
  for (BasicBlock *Pred : make_early_inc_range(predecessors(OldTarget)))
    redirectTo(Pred, NewTarget, DL);
}
 
static void removeUnusedBlocksFromParent(ArrayRef<BasicBlock *> BBs) {
  SmallPtrSet<BasicBlock *, 8> InternalBBs(from_range, BBs);
  // We add a block to BBsToKeep iff we have proven it has an external use.
  SmallPtrSet<BasicBlock *, 8> BBsToKeep;
 
  while (true) {
    bool Changed = false;
 
    for (BasicBlock *BB : BBs) {
      if (BBsToKeep.contains(BB))
        continue;
 
      for (Use &U : BB->uses()) {
        auto *UseInst = dyn_cast<Instruction>(U.getUser());
        if (!UseInst)
          continue;
        BasicBlock *UseBB = UseInst->getParent();
        if (!InternalBBs.contains(UseBB) || BBsToKeep.contains(UseBB)) {
          BBsToKeep.insert(BB);
          Changed = true;
          break;
        }
      }
    }
 
    if (!Changed)
      break;
  }
 
  SmallVector<BasicBlock *> BBsToDelete = filter_to_vector(
      BBs, [&BBsToKeep](BasicBlock *BB) { return !BBsToKeep.contains(BB); });
  DeleteDeadBlocks(BBsToDelete);
}
 
CanonicalLoopInfo *
OpenMPIRBuilder::collapseLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops,
                               InsertPointTy ComputeIP) {
  assert(Loops.size() >= 1 && "At least one loop required");
  size_t NumLoops = Loops.size();
 
  // Nothing to do if there is already just one loop.
  if (NumLoops == 1)
    return Loops.front();
 
  CanonicalLoopInfo *Outermost = Loops.front();
  CanonicalLoopInfo *Innermost = Loops.back();
  BasicBlock *OrigPreheader = Outermost->getPreheader();
  BasicBlock *OrigAfter = Outermost->getAfter();
  Function *F = OrigPreheader->getParent();
 
  // Loop control blocks that may become orphaned later.
  SmallVector<BasicBlock *, 12> OldControlBBs;
  OldControlBBs.reserve(6 * Loops.size());
  for (CanonicalLoopInfo *Loop : Loops)
    Loop->collectControlBlocks(OldControlBBs);
 
  // Setup the IRBuilder for inserting the trip count computation.
  Builder.SetCurrentDebugLocation(DL);
  if (ComputeIP.isSet())
    Builder.restoreIP(ComputeIP);
  else
    Builder.restoreIP(Outermost->getPreheaderIP());
 
  // Derive the collapsed' loop trip count.
  // TODO: Find common/largest indvar type.
  Value *CollapsedTripCount = nullptr;
  for (CanonicalLoopInfo *L : Loops) {
    assert(L->isValid() &&
           "All loops to collapse must be valid canonical loops");
    Value *OrigTripCount = L->getTripCount();
    if (!CollapsedTripCount) {
      CollapsedTripCount = OrigTripCount;
      continue;
    }
 
    // TODO: Enable UndefinedSanitizer to diagnose an overflow here.
    CollapsedTripCount =
        Builder.CreateNUWMul(CollapsedTripCount, OrigTripCount);
  }
 
  // Create the collapsed loop control flow.
  CanonicalLoopInfo *Result =
      createLoopSkeleton(DL, CollapsedTripCount, F,
                         OrigPreheader->getNextNode(), OrigAfter, "collapsed",
                         /*IsCollapsed=*/true);
 
  // Build the collapsed loop body code.
  // Start with deriving the input loop induction variables from the collapsed
  // one, using a divmod scheme. To preserve the original loops' order, the
  // innermost loop use the least significant bits.
  Builder.restoreIP(Result->getBodyIP());
 
  Value *Leftover = Result->getIndVar();
  SmallVector<Value *> NewIndVars;
  NewIndVars.resize(NumLoops);
  for (int i = NumLoops - 1; i >= 1; --i) {
    Value *OrigTripCount = Loops[i]->getTripCount();
 
    Value *NewIndVar = Builder.CreateURem(Leftover, OrigTripCount);
    NewIndVars[i] = NewIndVar;
 
    Leftover = Builder.CreateUDiv(Leftover, OrigTripCount);
  }
  // Outermost loop gets all the remaining bits.
  NewIndVars[0] = Leftover;
 
  // Construct the loop body control flow.
  // We progressively construct the branch structure following in direction of
  // the control flow, from the leading in-between code, the loop nest body, the
  // trailing in-between code, and rejoining the collapsed loop's latch.
  // ContinueBlock and ContinuePred keep track of the source(s) of next edge. If
  // the ContinueBlock is set, continue with that block. If ContinuePred, use
  // its predecessors as sources.
  BasicBlock *ContinueBlock = Result->getBody();
  BasicBlock *ContinuePred = nullptr;
  auto ContinueWith = [&ContinueBlock, &ContinuePred, DL](BasicBlock *Dest,
                                                          BasicBlock *NextSrc) {
    if (ContinueBlock)
      redirectTo(ContinueBlock, Dest, DL);
    else
      redirectAllPredecessorsTo(ContinuePred, Dest, DL);
 
    ContinueBlock = nullptr;
    ContinuePred = NextSrc;
  };
 
  // The code before the nested loop of each level.
  // Because we are sinking it into the nest, it will be executed more often
  // that the original loop. More sophisticated schemes could keep track of what
  // the in-between code is and instantiate it only once per thread.
  for (size_t i = 0; i < NumLoops - 1; ++i)
    ContinueWith(Loops[i]->getBody(), Loops[i + 1]->getHeader());
 
  // Connect the loop nest body.
  ContinueWith(Innermost->getBody(), Innermost->getLatch());
 
  // The code after the nested loop at each level.
  for (size_t i = NumLoops - 1; i > 0; --i)
    ContinueWith(Loops[i]->getAfter(), Loops[i - 1]->getLatch());
 
  // Connect the finished loop to the collapsed loop latch.
  ContinueWith(Result->getLatch(), nullptr);
 
  // Replace the input loops with the new collapsed loop.
  redirectTo(Outermost->getPreheader(), Result->getPreheader(), DL);
  redirectTo(Result->getAfter(), Outermost->getAfter(), DL);
 
  // Replace the input loop indvars with the derived ones.
  for (size_t i = 0; i < NumLoops; ++i)
    Loops[i]->getIndVar()->replaceAllUsesWith(NewIndVars[i]);
 
  // Remove unused parts of the input loops.
  removeUnusedBlocksFromParent(OldControlBBs);
 
  for (CanonicalLoopInfo *L : Loops)
    L->invalidate();
 
#ifndef NDEBUG
  Result->assertOK();
#endif
  return Result;
}
 
std::vector<CanonicalLoopInfo *>
OpenMPIRBuilder::tileLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops,
                           ArrayRef<Value *> TileSizes) {
  assert(TileSizes.size() == Loops.size() &&
         "Must pass as many tile sizes as there are loops");
  int NumLoops = Loops.size();
  assert(NumLoops >= 1 && "At least one loop to tile required");
 
  CanonicalLoopInfo *OutermostLoop = Loops.front();
  CanonicalLoopInfo *InnermostLoop = Loops.back();
  Function *F = OutermostLoop->getBody()->getParent();
  BasicBlock *InnerEnter = InnermostLoop->getBody();
  BasicBlock *InnerLatch = InnermostLoop->getLatch();
 
  // Loop control blocks that may become orphaned later.
  SmallVector<BasicBlock *, 12> OldControlBBs;
  OldControlBBs.reserve(6 * Loops.size());
  for (CanonicalLoopInfo *Loop : Loops)
    Loop->collectControlBlocks(OldControlBBs);
 
  // Collect original trip counts and induction variable to be accessible by
  // index. Also, the structure of the original loops is not preserved during
  // the construction of the tiled loops, so do it before we scavenge the BBs of
  // any original CanonicalLoopInfo.
  SmallVector<Value *, 4> OrigTripCounts, OrigIndVars;
  for (CanonicalLoopInfo *L : Loops) {
    assert(L->isValid() && "All input loops must be valid canonical loops");
    OrigTripCounts.push_back(L->getTripCount());
    OrigIndVars.push_back(L->getIndVar());
  }
 
  // Collect the code between loop headers. These may contain SSA definitions
  // that are used in the loop nest body. To be usable with in the innermost
  // body, these BasicBlocks will be sunk into the loop nest body. That is,
  // these instructions may be executed more often than before the tiling.
  // TODO: It would be sufficient to only sink them into body of the
  // corresponding tile loop.
  SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> InbetweenCode;
  for (int i = 0; i < NumLoops - 1; ++i) {
    CanonicalLoopInfo *Surrounding = Loops[i];
    CanonicalLoopInfo *Nested = Loops[i + 1];
 
    BasicBlock *EnterBB = Surrounding->getBody();
    BasicBlock *ExitBB = Nested->getHeader();
    InbetweenCode.emplace_back(EnterBB, ExitBB);
  }
 
  // Compute the trip counts of the floor loops.
  Builder.SetCurrentDebugLocation(DL);
  Builder.restoreIP(OutermostLoop->getPreheaderIP());
  SmallVector<Value *, 4> FloorCompleteCount, FloorCount, FloorRems;
  for (int i = 0; i < NumLoops; ++i) {
    Value *TileSize = TileSizes[i];
    Value *OrigTripCount = OrigTripCounts[i];
    Type *IVType = OrigTripCount->getType();
 
    Value *FloorCompleteTripCount = Builder.CreateUDiv(OrigTripCount, TileSize);
    Value *FloorTripRem = Builder.CreateURem(OrigTripCount, TileSize);
 
    // 0 if tripcount divides the tilesize, 1 otherwise.
    // 1 means we need an additional iteration for a partial tile.
    //
    // Unfortunately we cannot just use the roundup-formula
    //   (tripcount + tilesize - 1)/tilesize
    // because the summation might overflow. We do not want introduce undefined
    // behavior when the untiled loop nest did not.
    Value *FloorTripOverflow =
        Builder.CreateICmpNE(FloorTripRem, ConstantInt::get(IVType, 0));
 
    FloorTripOverflow = Builder.CreateZExt(FloorTripOverflow, IVType);
    Value *FloorTripCount =
        Builder.CreateAdd(FloorCompleteTripCount, FloorTripOverflow,
                          "omp_floor" + Twine(i) + ".tripcount", true);
 
    // Remember some values for later use.
    FloorCompleteCount.push_back(FloorCompleteTripCount);
    FloorCount.push_back(FloorTripCount);
    FloorRems.push_back(FloorTripRem);
  }
 
  // Generate the new loop nest, from the outermost to the innermost.
  std::vector<CanonicalLoopInfo *> Result;
  Result.reserve(NumLoops * 2);
 
  // The basic block of the surrounding loop that enters the nest generated
  // loop.
  BasicBlock *Enter = OutermostLoop->getPreheader();
 
  // The basic block of the surrounding loop where the inner code should
  // continue.
  BasicBlock *Continue = OutermostLoop->getAfter();
 
  // Where the next loop basic block should be inserted.
  BasicBlock *OutroInsertBefore = InnermostLoop->getExit();
 
  auto EmbeddNewLoop =
      [this, DL, F, InnerEnter, &Enter, &Continue, &OutroInsertBefore](
          Value *TripCount, const Twine &Name) -> CanonicalLoopInfo * {
    CanonicalLoopInfo *EmbeddedLoop = createLoopSkeleton(
        DL, TripCount, F, InnerEnter, OutroInsertBefore, Name);
    redirectTo(Enter, EmbeddedLoop->getPreheader(), DL);
    redirectTo(EmbeddedLoop->getAfter(), Continue, DL);
 
    // Setup the position where the next embedded loop connects to this loop.
    Enter = EmbeddedLoop->getBody();
    Continue = EmbeddedLoop->getLatch();
    OutroInsertBefore = EmbeddedLoop->getLatch();
    return EmbeddedLoop;
  };
 
  auto EmbeddNewLoops = [&Result, &EmbeddNewLoop](ArrayRef<Value *> TripCounts,
                                                  const Twine &NameBase) {
    for (auto P : enumerate(TripCounts)) {
      CanonicalLoopInfo *EmbeddedLoop =
          EmbeddNewLoop(P.value(), NameBase + Twine(P.index()));
      Result.push_back(EmbeddedLoop);
    }
  };
 
  EmbeddNewLoops(FloorCount, "floor");
 
  // Within the innermost floor loop, emit the code that computes the tile
  // sizes.
  Builder.SetInsertPoint(Enter->getTerminator());
  SmallVector<Value *, 4> TileCounts;
  for (int i = 0; i < NumLoops; ++i) {
    CanonicalLoopInfo *FloorLoop = Result[i];
    Value *TileSize = TileSizes[i];
 
    Value *FloorIsEpilogue =
        Builder.CreateICmpEQ(FloorLoop->getIndVar(), FloorCompleteCount[i]);
    Value *TileTripCount =
        Builder.CreateSelect(FloorIsEpilogue, FloorRems[i], TileSize);
 
    TileCounts.push_back(TileTripCount);
  }
 
  // Create the tile loops.
  EmbeddNewLoops(TileCounts, "tile");
 
  // Insert the inbetween code into the body.
  BasicBlock *BodyEnter = Enter;
  BasicBlock *BodyEntered = nullptr;
  for (std::pair<BasicBlock *, BasicBlock *> P : InbetweenCode) {
    BasicBlock *EnterBB = P.first;
    BasicBlock *ExitBB = P.second;
 
    if (BodyEnter)
      redirectTo(BodyEnter, EnterBB, DL);
    else
      redirectAllPredecessorsTo(BodyEntered, EnterBB, DL);
 
    BodyEnter = nullptr;
    BodyEntered = ExitBB;
  }
 
  // Append the original loop nest body into the generated loop nest body.
  if (BodyEnter)
    redirectTo(BodyEnter, InnerEnter, DL);
  else
    redirectAllPredecessorsTo(BodyEntered, InnerEnter, DL);
  redirectAllPredecessorsTo(InnerLatch, Continue, DL);
 
  // Replace the original induction variable with an induction variable computed
  // from the tile and floor induction variables.
  Builder.restoreIP(Result.back()->getBodyIP());
  for (int i = 0; i < NumLoops; ++i) {
    CanonicalLoopInfo *FloorLoop = Result[i];
    CanonicalLoopInfo *TileLoop = Result[NumLoops + i];
    Value *OrigIndVar = OrigIndVars[i];
    Value *Size = TileSizes[i];
 
    Value *Scale =
        Builder.CreateMul(Size, FloorLoop->getIndVar(), {}, /*HasNUW=*/true);
    Value *Shift =
        Builder.CreateAdd(Scale, TileLoop->getIndVar(), {}, /*HasNUW=*/true);
    OrigIndVar->replaceAllUsesWith(Shift);
  }
 
  // Remove unused parts of the original loops.
  removeUnusedBlocksFromParent(OldControlBBs);
 
  for (CanonicalLoopInfo *L : Loops)
    L->invalidate();
 
#ifndef NDEBUG
  for (CanonicalLoopInfo *GenL : Result)
    GenL->assertOK();
#endif
  return Result;
}
 
/// Attach metadata \p Properties to the basic block described by \p BB. If the
/// basic block already has metadata, the basic block properties are appended.
static void addBasicBlockMetadata(BasicBlock *BB,
                                  ArrayRef<Metadata *> Properties) {
  // Nothing to do if no property to attach.
  if (Properties.empty())
    return;
 
  LLVMContext &Ctx = BB->getContext();
  SmallVector<Metadata *> NewProperties;
  NewProperties.push_back(nullptr);
 
  // If the basic block already has metadata, prepend it to the new metadata.
  MDNode *Existing = BB->getTerminator()->getMetadata(LLVMContext::MD_loop);
  if (Existing)
    append_range(NewProperties, drop_begin(Existing->operands(), 1));
 
  append_range(NewProperties, Properties);
  MDNode *BasicBlockID = MDNode::getDistinct(Ctx, NewProperties);
  BasicBlockID->replaceOperandWith(0, BasicBlockID);
 
  BB->getTerminator()->setMetadata(LLVMContext::MD_loop, BasicBlockID);
}
 
/// Attach loop metadata \p Properties to the loop described by \p Loop. If the
/// loop already has metadata, the loop properties are appended.
static void addLoopMetadata(CanonicalLoopInfo *Loop,
                            ArrayRef<Metadata *> Properties) {
  assert(Loop->isValid() && "Expecting a valid CanonicalLoopInfo");
 
  // Attach metadata to the loop's latch
  BasicBlock *Latch = Loop->getLatch();
  assert(Latch && "A valid CanonicalLoopInfo must have a unique latch");
  addBasicBlockMetadata(Latch, Properties);
}
 
/// Attach llvm.access.group metadata to the memref instructions of \p Block
static void addAccessGroupMetadata(BasicBlock *Block, MDNode *AccessGroup,
                                   LoopInfo &LI) {
  for (Instruction &I : *Block) {
    if (I.mayReadOrWriteMemory()) {
      // TODO: This instruction may already have access group from
      // other pragmas e.g. #pragma clang loop vectorize.  Append
      // so that the existing metadata is not overwritten.
      I.setMetadata(LLVMContext::MD_access_group, AccessGroup);
    }
  }
}
 
CanonicalLoopInfo *
OpenMPIRBuilder::fuseLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops) {
  CanonicalLoopInfo *firstLoop = Loops.front();
  CanonicalLoopInfo *lastLoop = Loops.back();
  Function *F = firstLoop->getPreheader()->getParent();
 
  // Loop control blocks that will become orphaned later
  SmallVector<BasicBlock *> oldControlBBs;
  for (CanonicalLoopInfo *Loop : Loops)
    Loop->collectControlBlocks(oldControlBBs);
 
  // Collect original trip counts
  SmallVector<Value *> origTripCounts;
  for (CanonicalLoopInfo *L : Loops) {
    assert(L->isValid() && "All input loops must be valid canonical loops");
    origTripCounts.push_back(L->getTripCount());
  }
 
  Builder.SetCurrentDebugLocation(DL);
 
  // Compute max trip count.
  // The fused loop will be from 0 to max(origTripCounts)
  BasicBlock *TCBlock = BasicBlock::Create(F->getContext(), "omp.fuse.comp.tc",
                                           F, firstLoop->getHeader());
  Builder.SetInsertPoint(TCBlock);
  Value *fusedTripCount = nullptr;
  for (CanonicalLoopInfo *L : Loops) {
    assert(L->isValid() && "All loops to fuse must be valid canonical loops");
    Value *origTripCount = L->getTripCount();
    if (!fusedTripCount) {
      fusedTripCount = origTripCount;
      continue;
    }
    Value *condTP = Builder.CreateICmpSGT(fusedTripCount, origTripCount);
    fusedTripCount = Builder.CreateSelect(condTP, fusedTripCount, origTripCount,
                                          ".omp.fuse.tc");
  }
 
  // Generate new loop
  CanonicalLoopInfo *fused =
      createLoopSkeleton(DL, fusedTripCount, F, firstLoop->getBody(),
                         lastLoop->getLatch(), "fused");
 
  // Replace original loops with the fused loop
  // Preheader and After are not considered inside the CLI.
  // These are used to compute the individual TCs of the loops
  // so they have to be put before the resulting fused loop.
  // Moving them up for readability.
  for (size_t i = 0; i < Loops.size() - 1; ++i) {
    Loops[i]->getPreheader()->moveBefore(TCBlock);
    Loops[i]->getAfter()->moveBefore(TCBlock);
  }
  lastLoop->getPreheader()->moveBefore(TCBlock);
 
  for (size_t i = 0; i < Loops.size() - 1; ++i) {
    redirectTo(Loops[i]->getPreheader(), Loops[i]->getAfter(), DL);
    redirectTo(Loops[i]->getAfter(), Loops[i + 1]->getPreheader(), DL);
  }
  redirectTo(lastLoop->getPreheader(), TCBlock, DL);
  redirectTo(TCBlock, fused->getPreheader(), DL);
  redirectTo(fused->getAfter(), lastLoop->getAfter(), DL);
 
  // Build the fused body
  // Create new Blocks with conditions that jump to the original loop bodies
  SmallVector<BasicBlock *> condBBs;
  SmallVector<Value *> condValues;
  for (size_t i = 0; i < Loops.size(); ++i) {
    BasicBlock *condBlock = BasicBlock::Create(
        F->getContext(), "omp.fused.inner.cond", F, Loops[i]->getBody());
    Builder.SetInsertPoint(condBlock);
    Value *condValue =
        Builder.CreateICmpSLT(fused->getIndVar(), origTripCounts[i]);
    condBBs.push_back(condBlock);
    condValues.push_back(condValue);
  }
  // Join the condition blocks with the bodies of the original loops
  redirectTo(fused->getBody(), condBBs[0], DL);
  for (size_t i = 0; i < Loops.size() - 1; ++i) {
    Builder.SetInsertPoint(condBBs[i]);
    Builder.CreateCondBr(condValues[i], Loops[i]->getBody(), condBBs[i + 1]);
    redirectAllPredecessorsTo(Loops[i]->getLatch(), condBBs[i + 1], DL);
    // Replace the IV with the fused IV
    Loops[i]->getIndVar()->replaceAllUsesWith(fused->getIndVar());
  }
  // Last body jumps to the created end body block
  Builder.SetInsertPoint(condBBs.back());
  Builder.CreateCondBr(condValues.back(), lastLoop->getBody(),
                       fused->getLatch());
  redirectAllPredecessorsTo(lastLoop->getLatch(), fused->getLatch(), DL);
  // Replace the IV with the fused IV
  lastLoop->getIndVar()->replaceAllUsesWith(fused->getIndVar());
 
  // The loop latch must have only one predecessor. Currently it is branched to
  // from both the last condition block and the last loop body
  fused->getLatch()->splitBasicBlockBefore(fused->getLatch()->begin(),
                                           "omp.fused.pre_latch");
 
  // Remove unused parts
  removeUnusedBlocksFromParent(oldControlBBs);
 
  // Invalidate old CLIs
  for (CanonicalLoopInfo *L : Loops)
    L->invalidate();
 
#ifndef NDEBUG
  fused->assertOK();
#endif
  return fused;
}
 
void OpenMPIRBuilder::unrollLoopFull(DebugLoc, CanonicalLoopInfo *Loop) {
  LLVMContext &Ctx = Builder.getContext();
  addLoopMetadata(
      Loop, {MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
             MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.full"))});
}
 
void OpenMPIRBuilder::unrollLoopHeuristic(DebugLoc, CanonicalLoopInfo *Loop) {
  LLVMContext &Ctx = Builder.getContext();
  addLoopMetadata(
      Loop, {
                MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
            });
}
 
void OpenMPIRBuilder::createIfVersion(CanonicalLoopInfo *CanonicalLoop,
                                      Value *IfCond, ValueToValueMapTy &VMap,
                                      LoopAnalysis &LIA, LoopInfo &LI, Loop *L,
                                      const Twine &NamePrefix) {
  Function *F = CanonicalLoop->getFunction();
 
  // We can't do
  // if (cond) {
  //   simd_loop;
  // } else {
  //   non_simd_loop;
  // }
  // because then the CanonicalLoopInfo would only point to one of the loops:
  // leading to other constructs operating on the same loop to malfunction.
  // Instead generate
  // while (...) {
  //   if (cond) {
  //     simd_body;
  //   } else {
  //     not_simd_body;
  //   }
  // }
  // At least for simple loops, LLVM seems able to hoist the if out of the loop
  // body at -O3
 
  // Define where if branch should be inserted
  auto SplitBeforeIt = CanonicalLoop->getBody()->getFirstNonPHIIt();
 
  // Create additional blocks for the if statement
  BasicBlock *Cond = SplitBeforeIt->getParent();
  llvm::LLVMContext &C = Cond->getContext();
  llvm::BasicBlock *ThenBlock = llvm::BasicBlock::Create(
      C, NamePrefix + ".if.then", Cond->getParent(), Cond->getNextNode());
  llvm::BasicBlock *ElseBlock = llvm::BasicBlock::Create(
      C, NamePrefix + ".if.else", Cond->getParent(), CanonicalLoop->getExit());
 
  // Create if condition branch.
  Builder.SetInsertPoint(SplitBeforeIt);
  Instruction *BrInstr =
      Builder.CreateCondBr(IfCond, ThenBlock, /*ifFalse*/ ElseBlock);
  InsertPointTy IP{BrInstr->getParent(), ++BrInstr->getIterator()};
  // Then block contains branch to omp loop body which needs to be vectorized
  spliceBB(IP, ThenBlock, false, Builder.getCurrentDebugLocation());
  ThenBlock->replaceSuccessorsPhiUsesWith(Cond, ThenBlock);
 
  Builder.SetInsertPoint(ElseBlock);
 
  // Clone loop for the else branch
  SmallVector<BasicBlock *, 8> NewBlocks;
 
  SmallVector<BasicBlock *, 8> ExistingBlocks;
  ExistingBlocks.reserve(L->getNumBlocks() + 1);
  ExistingBlocks.push_back(ThenBlock);
  ExistingBlocks.append(L->block_begin(), L->block_end());
  // Cond is the block that has the if clause condition
  // LoopCond is omp_loop.cond
  // LoopHeader is omp_loop.header
  BasicBlock *LoopCond = Cond->getUniquePredecessor();
  BasicBlock *LoopHeader = LoopCond->getUniquePredecessor();
  assert(LoopCond && LoopHeader && "Invalid loop structure");
  for (BasicBlock *Block : ExistingBlocks) {
    if (Block == L->getLoopPreheader() || Block == L->getLoopLatch() ||
        Block == LoopHeader || Block == LoopCond || Block == Cond) {
      continue;
    }
    BasicBlock *NewBB = CloneBasicBlock(Block, VMap, "", F);
 
    // fix name not to be omp.if.then
    if (Block == ThenBlock)
      NewBB->setName(NamePrefix + ".if.else");
 
    NewBB->moveBefore(CanonicalLoop->getExit());
    VMap[Block] = NewBB;
    NewBlocks.push_back(NewBB);
  }
  remapInstructionsInBlocks(NewBlocks, VMap);
  Builder.CreateBr(NewBlocks.front());
 
  // The loop latch must have only one predecessor. Currently it is branched to
  // from both the 'then' and 'else' branches.
  L->getLoopLatch()->splitBasicBlockBefore(L->getLoopLatch()->begin(),
                                           NamePrefix + ".pre_latch");
 
  // Ensure that the then block is added to the loop so we add the attributes in
  // the next step
  L->addBasicBlockToLoop(ThenBlock, LI);
}
 
unsigned
OpenMPIRBuilder::getOpenMPDefaultSimdAlign(const Triple &TargetTriple,
                                           const StringMap<bool> &Features) {
  if (TargetTriple.isX86()) {
    if (Features.lookup("avx512f"))
      return 512;
    else if (Features.lookup("avx"))
      return 256;
    return 128;
  }
  if (TargetTriple.isPPC())
    return 128;
  if (TargetTriple.isWasm())
    return 128;
  return 0;
}
 
void OpenMPIRBuilder::applySimd(CanonicalLoopInfo *CanonicalLoop,
                                MapVector<Value *, Value *> AlignedVars,
                                Value *IfCond, OrderKind Order,
                                ConstantInt *Simdlen, ConstantInt *Safelen) {
  LLVMContext &Ctx = Builder.getContext();
 
  Function *F = CanonicalLoop->getFunction();
 
  // Blocks must have terminators.
  // FIXME: Don't run analyses on incomplete/invalid IR.
  SmallVector<Instruction *> UIs;
  for (BasicBlock &BB : *F)
    if (!BB.hasTerminator())
      UIs.push_back(new UnreachableInst(F->getContext(), &BB));
 
  // TODO: We should not rely on pass manager. Currently we use pass manager
  // only for getting llvm::Loop which corresponds to given CanonicalLoopInfo
  // object. We should have a method  which returns all blocks between
  // CanonicalLoopInfo::getHeader() and CanonicalLoopInfo::getAfter()
  FunctionAnalysisManager FAM;
  FAM.registerPass([]() { return DominatorTreeAnalysis(); });
  FAM.registerPass([]() { return LoopAnalysis(); });
  FAM.registerPass([]() { return PassInstrumentationAnalysis(); });
 
  LoopAnalysis LIA;
  LoopInfo &&LI = LIA.run(*F, FAM);
 
  for (Instruction *I : UIs)
    I->eraseFromParent();
 
  Loop *L = LI.getLoopFor(CanonicalLoop->getHeader());
  if (AlignedVars.size()) {
    InsertPointTy IP = Builder.saveIP();
    for (auto &AlignedItem : AlignedVars) {
      Value *AlignedPtr = AlignedItem.first;
      Value *Alignment = AlignedItem.second;
      Instruction *loadInst = dyn_cast<Instruction>(AlignedPtr);
      Builder.SetInsertPoint(loadInst->getNextNode());
      Builder.CreateAlignmentAssumption(F->getDataLayout(), AlignedPtr,
                                        Alignment);
    }
    Builder.restoreIP(IP);
  }
 
  if (IfCond) {
    ValueToValueMapTy VMap;
    createIfVersion(CanonicalLoop, IfCond, VMap, LIA, LI, L, "simd");
  }
 
  SmallPtrSet<BasicBlock *, 8> Reachable;
 
  // Get the basic blocks from the loop in which memref instructions
  // can be found.
  // TODO: Generalize getting all blocks inside a CanonicalizeLoopInfo,
  // preferably without running any passes.
  for (BasicBlock *Block : L->getBlocks()) {
    if (Block == CanonicalLoop->getCond() ||
        Block == CanonicalLoop->getHeader())
      continue;
    Reachable.insert(Block);
  }
 
  SmallVector<Metadata *> LoopMDList;
 
  // In presence of finite 'safelen', it may be unsafe to mark all
  // the memory instructions parallel, because loop-carried
  // dependences of 'safelen' iterations are possible.
  // If clause order(concurrent) is specified then the memory instructions
  // are marked parallel even if 'safelen' is finite.
  if ((Safelen == nullptr) || (Order == OrderKind::OMP_ORDER_concurrent))
    applyParallelAccessesMetadata(CanonicalLoop, Ctx, L, LI, LoopMDList);
 
  // FIXME: the IF clause shares a loop backedge for the SIMD and non-SIMD
  // versions so we can't add the loop attributes in that case.
  if (IfCond) {
    // we can still add llvm.loop.parallel_access
    addLoopMetadata(CanonicalLoop, LoopMDList);
    return;
  }
 
  // Use the above access group metadata to create loop level
  // metadata, which should be distinct for each loop.
  LoopMDList.push_back(
      MDNode::get(Ctx, {MDString::get(Ctx, "llvm.loop.vectorize.enable")}));
 
  if (Simdlen || Safelen) {
    // If both simdlen and safelen clauses are specified, the value of the
    // simdlen parameter must be less than or equal to the value of the safelen
    // parameter. Therefore, use safelen only in the absence of simdlen.
    ConstantInt *VectorizeWidth = Simdlen == nullptr ? Safelen : Simdlen;
    LoopMDList.push_back(
        MDNode::get(Ctx, {MDString::get(Ctx, "llvm.loop.vectorize.width"),
                          ConstantAsMetadata::get(VectorizeWidth)}));
  }
 
  addLoopMetadata(CanonicalLoop, LoopMDList);
}
 
/// Create the TargetMachine object to query the backend for optimization
/// preferences.
///
/// Ideally, this would be passed from the front-end to the OpenMPBuilder, but
/// e.g. Clang does not pass it to its CodeGen layer and creates it only when
/// needed for the LLVM pass pipline. We use some default options to avoid
/// having to pass too many settings from the frontend that probably do not
/// matter.
///
/// Currently, TargetMachine is only used sometimes by the unrollLoopPartial
/// method. If we are going to use TargetMachine for more purposes, especially
/// those that are sensitive to TargetOptions, RelocModel and CodeModel, it
/// might become be worth requiring front-ends to pass on their TargetMachine,
/// or at least cache it between methods. Note that while fontends such as Clang
/// have just a single main TargetMachine per translation unit, "target-cpu" and
/// "target-features" that determine the TargetMachine are per-function and can
/// be overrided using __attribute__((target("OPTIONS"))).
static std::unique_ptr<TargetMachine>
createTargetMachine(Function *F, CodeGenOptLevel OptLevel) {
  Module *M = F->getParent();
 
  StringRef CPU = F->getFnAttribute("target-cpu").getValueAsString();
  StringRef Features = F->getFnAttribute("target-features").getValueAsString();
  const llvm::Triple &Triple = M->getTargetTriple();
 
  std::string Error;
  const llvm::Target *TheTarget = TargetRegistry::lookupTarget(Triple, Error);
  if (!TheTarget)
    return {};
 
  llvm::TargetOptions Options;
  return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
      Triple, CPU, Features, Options, /*RelocModel=*/std::nullopt,
      /*CodeModel=*/std::nullopt, OptLevel));
}
 
/// Heuristically determine the best-performant unroll factor for \p CLI. This
/// depends on the target processor. We are re-using the same heuristics as the
/// LoopUnrollPass.
static int32_t computeHeuristicUnrollFactor(CanonicalLoopInfo *CLI) {
  Function *F = CLI->getFunction();
 
  // Assume the user requests the most aggressive unrolling, even if the rest of
  // the code is optimized using a lower setting.
  CodeGenOptLevel OptLevel = CodeGenOptLevel::Aggressive;
  std::unique_ptr<TargetMachine> TM = createTargetMachine(F, OptLevel);
 
  // Blocks must have terminators.
  // FIXME: Don't run analyses on incomplete/invalid IR.
  SmallVector<Instruction *> UIs;
  for (BasicBlock &BB : *F)
    if (!BB.hasTerminator())
      UIs.push_back(new UnreachableInst(F->getContext(), &BB));
 
  FunctionAnalysisManager FAM;
  FAM.registerPass([]() { return TargetLibraryAnalysis(); });
  FAM.registerPass([]() { return AssumptionAnalysis(); });
  FAM.registerPass([]() { return DominatorTreeAnalysis(); });
  FAM.registerPass([]() { return LoopAnalysis(); });
  FAM.registerPass([]() { return ScalarEvolutionAnalysis(); });
  FAM.registerPass([]() { return PassInstrumentationAnalysis(); });
  TargetIRAnalysis TIRA;
  if (TM)
    TIRA = TargetIRAnalysis(
        [&](const Function &F) { return TM->getTargetTransformInfo(F); });
  FAM.registerPass([&]() { return TIRA; });
 
  TargetIRAnalysis::Result &&TTI = TIRA.run(*F, FAM);
  ScalarEvolutionAnalysis SEA;
  ScalarEvolution &&SE = SEA.run(*F, FAM);
  DominatorTreeAnalysis DTA;
  DominatorTree &&DT = DTA.run(*F, FAM);
  LoopAnalysis LIA;
  LoopInfo &&LI = LIA.run(*F, FAM);
  AssumptionAnalysis ACT;
  AssumptionCache &&AC = ACT.run(*F, FAM);
  OptimizationRemarkEmitter ORE{F};
 
  for (Instruction *I : UIs)
    I->eraseFromParent();
 
  Loop *L = LI.getLoopFor(CLI->getHeader());
  assert(L && "Expecting CanonicalLoopInfo to be recognized as a loop");
 
  TargetTransformInfo::UnrollingPreferences UP = gatherUnrollingPreferences(
      L, SE, TTI,
      /*BlockFrequencyInfo=*/nullptr,
      /*ProfileSummaryInfo=*/nullptr, ORE, static_cast<int>(OptLevel),
      /*UserThreshold=*/std::nullopt,
      /*UserAllowPartial=*/true,
      /*UserAllowRuntime=*/true,
      /*UserUpperBound=*/std::nullopt,
      /*UserFullUnrollMaxCount=*/std::nullopt);
 
  UP.Force = true;
 
  // Account for additional optimizations taking place before the LoopUnrollPass
  // would unroll the loop.
  UP.Threshold *= UnrollThresholdFactor;
  UP.PartialThreshold *= UnrollThresholdFactor;
 
  // Use normal unroll factors even if the rest of the code is optimized for
  // size.
  UP.OptSizeThreshold = UP.Threshold;
  UP.PartialOptSizeThreshold = UP.PartialThreshold;
 
  LLVM_DEBUG(dbgs() << "Unroll heuristic thresholds:\n"
                    << "  Threshold=" << UP.Threshold << "\n"
                    << "  PartialThreshold=" << UP.PartialThreshold << "\n"
                    << "  OptSizeThreshold=" << UP.OptSizeThreshold << "\n"
                    << "  PartialOptSizeThreshold="
                    << UP.PartialOptSizeThreshold << "\n");
 
  // Disable peeling.
  TargetTransformInfo::PeelingPreferences PP =
      gatherPeelingPreferences(L, SE, TTI,
                               /*UserAllowPeeling=*/false,
                               /*UserAllowProfileBasedPeeling=*/false,
                               /*UnrollingSpecficValues=*/false);
 
  SmallPtrSet<const Value *, 32> EphValues;
  CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
 
  // Assume that reads and writes to stack variables can be eliminated by
  // Mem2Reg, SROA or LICM. That is, don't count them towards the loop body's
  // size.
  for (BasicBlock *BB : L->blocks()) {
    for (Instruction &I : *BB) {
      Value *Ptr;
      if (auto *Load = dyn_cast<LoadInst>(&I)) {
        Ptr = Load->getPointerOperand();
      } else if (auto *Store = dyn_cast<StoreInst>(&I)) {
        Ptr = Store->getPointerOperand();
      } else
        continue;
 
      Ptr = Ptr->stripPointerCasts();
 
      if (auto *Alloca = dyn_cast<AllocaInst>(Ptr)) {
        if (Alloca->getParent() == &F->getEntryBlock())
          EphValues.insert(&I);
      }
    }
  }
 
  UnrollCostEstimator UCE(L, TTI, EphValues, UP.BEInsns);
 
  // Loop is not unrollable if the loop contains certain instructions.
  if (!UCE.canUnroll()) {
    LLVM_DEBUG(dbgs() << "Loop not considered unrollable\n");
    return 1;
  }
 
  LLVM_DEBUG(dbgs() << "Estimated loop size is " << UCE.getRolledLoopSize()
                    << "\n");
 
  // TODO: Determine trip count of \p CLI if constant, computeUnrollCount might
  // be able to use it.
  int TripCount = 0;
  int MaxTripCount = 0;
  bool MaxOrZero = false;
  unsigned TripMultiple = 0;
 
  unsigned Factor =
      computeUnrollCount(L, TTI, DT, &LI, &AC, SE, EphValues, &ORE, TripCount,
                         MaxTripCount, MaxOrZero, TripMultiple, UCE, UP, PP);
  LLVM_DEBUG(dbgs() << "Suggesting unroll factor of " << Factor << "\n");
 
  // This function returns 1 to signal to not unroll a loop.
  if (Factor == 0)
    return 1;
  return Factor;
}
 
void OpenMPIRBuilder::unrollLoopPartial(DebugLoc DL, CanonicalLoopInfo *Loop,
                                        int32_t Factor,
                                        CanonicalLoopInfo **UnrolledCLI) {
  assert(Factor >= 0 && "Unroll factor must not be negative");
 
  Function *F = Loop->getFunction();
  LLVMContext &Ctx = F->getContext();
 
  // If the unrolled loop is not used for another loop-associated directive, it
  // is sufficient to add metadata for the LoopUnrollPass.
  if (!UnrolledCLI) {
    SmallVector<Metadata *, 2> LoopMetadata;
    LoopMetadata.push_back(
        MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")));
 
    if (Factor >= 1) {
      ConstantAsMetadata *FactorConst = ConstantAsMetadata::get(
          ConstantInt::get(Type::getInt32Ty(Ctx), APInt(32, Factor)));
      LoopMetadata.push_back(MDNode::get(
          Ctx, {MDString::get(Ctx, "llvm.loop.unroll.count"), FactorConst}));
    }
 
    addLoopMetadata(Loop, LoopMetadata);
    return;
  }
 
  // Heuristically determine the unroll factor.
  if (Factor == 0)
    Factor = computeHeuristicUnrollFactor(Loop);
 
  // No change required with unroll factor 1.
  if (Factor == 1) {
    *UnrolledCLI = Loop;
    return;
  }
 
  assert(Factor >= 2 &&
         "unrolling only makes sense with a factor of 2 or larger");
 
  Type *IndVarTy = Loop->getIndVarType();
 
  // Apply partial unrolling by tiling the loop by the unroll-factor, then fully
  // unroll the inner loop.
  Value *FactorVal =
      ConstantInt::get(IndVarTy, APInt(IndVarTy->getIntegerBitWidth(), Factor,
                                       /*isSigned=*/false));
  std::vector<CanonicalLoopInfo *> LoopNest =
      tileLoops(DL, {Loop}, {FactorVal});
  assert(LoopNest.size() == 2 && "Expect 2 loops after tiling");
  *UnrolledCLI = LoopNest[0];
  CanonicalLoopInfo *InnerLoop = LoopNest[1];
 
  // LoopUnrollPass can only fully unroll loops with constant trip count.
  // Unroll by the unroll factor with a fallback epilog for the remainder
  // iterations if necessary.
  ConstantAsMetadata *FactorConst = ConstantAsMetadata::get(
      ConstantInt::get(Type::getInt32Ty(Ctx), APInt(32, Factor)));
  addLoopMetadata(
      InnerLoop,
      {MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
       MDNode::get(
           Ctx, {MDString::get(Ctx, "llvm.loop.unroll.count"), FactorConst})});
 
#ifndef NDEBUG
  (*UnrolledCLI)->assertOK();
#endif
}
 
OpenMPIRBuilder::InsertPointTy
OpenMPIRBuilder::createCopyPrivate(const LocationDescription &Loc,
                                   llvm::Value *BufSize, llvm::Value *CpyBuf,
                                   llvm::Value *CpyFn, llvm::Value *DidIt) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
 
  llvm::Value *DidItLD = Builder.CreateLoad(Builder.getInt32Ty(), DidIt);
 
  Value *Args[] = {Ident, ThreadId, BufSize, CpyBuf, CpyFn, DidItLD};
 
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_copyprivate);
  createRuntimeFunctionCall(Fn, Args);
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createSingle(
    const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB,
    FinalizeCallbackTy FiniCB, bool IsNowait, ArrayRef<llvm::Value *> CPVars,
    ArrayRef<llvm::Function *> CPFuncs) {
 
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  // If needed allocate and initialize `DidIt` with 0.
  // DidIt: flag variable: 1=single thread; 0=not single thread.
  llvm::Value *DidIt = nullptr;
  if (!CPVars.empty()) {
    DidIt = Builder.CreateAlloca(llvm::Type::getInt32Ty(Builder.getContext()));
    Builder.CreateStore(Builder.getInt32(0), DidIt);
  }
 
  Directive OMPD = Directive::OMPD_single;
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Value *Args[] = {Ident, ThreadId};
 
  Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_single);
  Instruction *EntryCall = createRuntimeFunctionCall(EntryRTLFn, Args);
 
  Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_single);
  Instruction *ExitCall = createRuntimeFunctionCall(ExitRTLFn, Args);
 
  auto FiniCBWrapper = [&](InsertPointTy IP) -> Error {
    if (Error Err = FiniCB(IP))
      return Err;
 
    // The thread that executes the single region must set `DidIt` to 1.
    // This is used by __kmpc_copyprivate, to know if the caller is the
    // single thread or not.
    if (DidIt)
      Builder.CreateStore(Builder.getInt32(1), DidIt);
 
    return Error::success();
  };
 
  // generates the following:
  // if (__kmpc_single()) {
  //    .... single region ...
  //    __kmpc_end_single
  // }
  // __kmpc_copyprivate
  // __kmpc_barrier
 
  InsertPointOrErrorTy AfterIP =
      EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCBWrapper,
                           /*Conditional*/ true,
                           /*hasFinalize*/ true);
  if (!AfterIP)
    return AfterIP.takeError();
 
  if (DidIt) {
    for (size_t I = 0, E = CPVars.size(); I < E; ++I)
      // NOTE BufSize is currently unused, so just pass 0.
      createCopyPrivate(LocationDescription(Builder.saveIP(), Loc.DL),
                        /*BufSize=*/ConstantInt::get(Int64, 0), CPVars[I],
                        CPFuncs[I], DidIt);
    // NOTE __kmpc_copyprivate already inserts a barrier
  } else if (!IsNowait) {
    InsertPointOrErrorTy AfterIP =
        createBarrier(LocationDescription(Builder.saveIP(), Loc.DL),
                      omp::Directive::OMPD_unknown, /* ForceSimpleCall */ false,
                      /* CheckCancelFlag */ false);
    if (!AfterIP)
      return AfterIP.takeError();
  }
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy
OpenMPIRBuilder::createScope(const LocationDescription &Loc,
                             BodyGenCallbackTy BodyGenCB,
                             FinalizeCallbackTy FiniCB, bool IsNowait) {
 
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  // All threads execute the scope body — no conditional entry.
  InsertPointOrErrorTy AfterIP = EmitOMPInlinedRegion(
      Directive::OMPD_scope, /*EntryCall=*/nullptr, /*ExitCall=*/nullptr,
      BodyGenCB, FiniCB, /*Conditional=*/false, /*HasFinalize=*/true,
      /*IsCancellable=*/false);
  if (!AfterIP)
    return AfterIP.takeError();
 
  Builder.restoreIP(*AfterIP);
  if (!IsNowait) {
    AfterIP = createBarrier(LocationDescription(Builder.saveIP(), Loc.DL),
                            omp::Directive::OMPD_unknown,
                            /*ForceSimpleCall=*/false,
                            /*CheckCancelFlag=*/false);
    if (!AfterIP)
      return AfterIP.takeError();
  }
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createCritical(
    const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB,
    FinalizeCallbackTy FiniCB, StringRef CriticalName, Value *HintInst) {
 
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  Directive OMPD = Directive::OMPD_critical;
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Value *LockVar = getOMPCriticalRegionLock(CriticalName);
  Value *Args[] = {Ident, ThreadId, LockVar};
 
  SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args), std::end(Args));
  Function *RTFn = nullptr;
  if (HintInst) {
    // Add Hint to entry Args and create call
    EnterArgs.push_back(HintInst);
    RTFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_critical_with_hint);
  } else {
    RTFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_critical);
  }
  Instruction *EntryCall = createRuntimeFunctionCall(RTFn, EnterArgs);
 
  Function *ExitRTLFn =
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_critical);
  Instruction *ExitCall = createRuntimeFunctionCall(ExitRTLFn, Args);
 
  return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
                              /*Conditional*/ false, /*hasFinalize*/ true);
}
 
OpenMPIRBuilder::InsertPointTy
OpenMPIRBuilder::createOrderedDepend(const LocationDescription &Loc,
                                     InsertPointTy AllocaIP, unsigned NumLoops,
                                     ArrayRef<llvm::Value *> StoreValues,
                                     const Twine &Name, bool IsDependSource) {
  assert(
      llvm::all_of(StoreValues,
                   [](Value *SV) { return SV->getType()->isIntegerTy(64); }) &&
      "OpenMP runtime requires depend vec with i64 type");
 
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  // Allocate space for vector and generate alloc instruction.
  auto *ArrI64Ty = ArrayType::get(Int64, NumLoops);
  Builder.restoreIP(AllocaIP);
  AllocaInst *ArgsBase = Builder.CreateAlloca(ArrI64Ty, nullptr, Name);
  ArgsBase->setAlignment(Align(8));
  updateToLocation(Loc);
 
  // Store the index value with offset in depend vector.
  for (unsigned I = 0; I < NumLoops; ++I) {
    Value *DependAddrGEPIter = Builder.CreateInBoundsGEP(
        ArrI64Ty, ArgsBase, {Builder.getInt64(0), Builder.getInt64(I)});
    StoreInst *STInst = Builder.CreateStore(StoreValues[I], DependAddrGEPIter);
    STInst->setAlignment(Align(8));
  }
 
  Value *DependBaseAddrGEP = Builder.CreateInBoundsGEP(
      ArrI64Ty, ArgsBase, {Builder.getInt64(0), Builder.getInt64(0)});
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Value *Args[] = {Ident, ThreadId, DependBaseAddrGEP};
 
  Function *RTLFn = nullptr;
  if (IsDependSource)
    RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_doacross_post);
  else
    RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_doacross_wait);
  createRuntimeFunctionCall(RTLFn, Args);
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createOrderedThreadsSimd(
    const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB,
    FinalizeCallbackTy FiniCB, bool IsThreads) {
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  Directive OMPD = Directive::OMPD_ordered_blockassoc;
  Instruction *EntryCall = nullptr;
  Instruction *ExitCall = nullptr;
 
  if (IsThreads) {
    uint32_t SrcLocStrSize;
    Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
    Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
    Value *ThreadId = getOrCreateThreadID(Ident);
    Value *Args[] = {Ident, ThreadId};
 
    Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_ordered);
    EntryCall = createRuntimeFunctionCall(EntryRTLFn, Args);
 
    Function *ExitRTLFn =
        getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_ordered);
    ExitCall = createRuntimeFunctionCall(ExitRTLFn, Args);
  }
 
  return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
                              /*Conditional*/ false, /*hasFinalize*/ true);
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::EmitOMPInlinedRegion(
    Directive OMPD, Instruction *EntryCall, Instruction *ExitCall,
    BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, bool Conditional,
    bool HasFinalize, bool IsCancellable) {
 
  if (HasFinalize)
    FinalizationStack.push_back({FiniCB, OMPD, IsCancellable});
 
  // Create inlined region's entry and body blocks, in preparation
  // for conditional creation
  BasicBlock *EntryBB = Builder.GetInsertBlock();
  Instruction *SplitPos = EntryBB->getTerminatorOrNull();
  if (!isa_and_nonnull<UncondBrInst, CondBrInst>(SplitPos))
    SplitPos = new UnreachableInst(Builder.getContext(), EntryBB);
  BasicBlock *ExitBB = EntryBB->splitBasicBlock(SplitPos, "omp_region.end");
  BasicBlock *FiniBB =
      EntryBB->splitBasicBlock(EntryBB->getTerminator(), "omp_region.finalize");
 
  Builder.SetInsertPoint(EntryBB->getTerminator());
  emitCommonDirectiveEntry(OMPD, EntryCall, ExitBB, Conditional);
 
  // generate body
  if (Error Err =
          BodyGenCB(/* AllocaIP */ InsertPointTy(),
                    /* CodeGenIP */ Builder.saveIP(), /* DeallocBlocks */ {}))
    return Err;
 
  // emit exit call and do any needed finalization.
  auto FinIP = InsertPointTy(FiniBB, FiniBB->getFirstInsertionPt());
  assert(FiniBB->getTerminator()->getNumSuccessors() == 1 &&
         FiniBB->getTerminator()->getSuccessor(0) == ExitBB &&
         "Unexpected control flow graph state!!");
  InsertPointOrErrorTy AfterIP =
      emitCommonDirectiveExit(OMPD, FinIP, ExitCall, HasFinalize);
  if (!AfterIP)
    return AfterIP.takeError();
 
  // If we are skipping the region of a non conditional, remove the exit
  // block, and clear the builder's insertion point.
  assert(SplitPos->getParent() == ExitBB &&
         "Unexpected Insertion point location!");
  auto merged = MergeBlockIntoPredecessor(ExitBB);
  BasicBlock *ExitPredBB = SplitPos->getParent();
  auto InsertBB = merged ? ExitPredBB : ExitBB;
  if (!isa_and_nonnull<UncondBrInst, CondBrInst>(SplitPos))
    SplitPos->eraseFromParent();
  Builder.SetInsertPoint(InsertBB);
 
  return Builder.saveIP();
}
 
OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveEntry(
    Directive OMPD, Value *EntryCall, BasicBlock *ExitBB, bool Conditional) {
  // if nothing to do, Return current insertion point.
  if (!Conditional || !EntryCall)
    return Builder.saveIP();
 
  BasicBlock *EntryBB = Builder.GetInsertBlock();
  Value *CallBool = Builder.CreateIsNotNull(EntryCall);
  auto *ThenBB = BasicBlock::Create(M.getContext(), "omp_region.body");
  auto *UI = new UnreachableInst(Builder.getContext(), ThenBB);
 
  // Emit thenBB and set the Builder's insertion point there for
  // body generation next. Place the block after the current block.
  Function *CurFn = EntryBB->getParent();
  CurFn->insert(std::next(EntryBB->getIterator()), ThenBB);
 
  // Move Entry branch to end of ThenBB, and replace with conditional
  // branch (If-stmt)
  Instruction *EntryBBTI = EntryBB->getTerminator();
  Builder.CreateCondBr(CallBool, ThenBB, ExitBB);
  EntryBBTI->removeFromParent();
  Builder.SetInsertPoint(UI);
  Builder.Insert(EntryBBTI);
  UI->eraseFromParent();
  Builder.SetInsertPoint(ThenBB->getTerminator());
 
  // return an insertion point to ExitBB.
  return IRBuilder<>::InsertPoint(ExitBB, ExitBB->getFirstInsertionPt());
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::emitCommonDirectiveExit(
    omp::Directive OMPD, InsertPointTy FinIP, Instruction *ExitCall,
    bool HasFinalize) {
 
  Builder.restoreIP(FinIP);
 
  // If there is finalization to do, emit it before the exit call
  if (HasFinalize) {
    assert(!FinalizationStack.empty() &&
           "Unexpected finalization stack state!");
 
    FinalizationInfo Fi = FinalizationStack.pop_back_val();
    assert(Fi.DK == OMPD && "Unexpected Directive for Finalization call!");
 
    if (Error Err = Fi.mergeFiniBB(Builder, FinIP.getBlock()))
      return std::move(Err);
 
    // Exit condition: insertion point is before the terminator of the new Fini
    // block
    Builder.SetInsertPoint(FinIP.getBlock()->getTerminator());
  }
 
  if (!ExitCall)
    return Builder.saveIP();
 
  // place the Exitcall as last instruction before Finalization block terminator
  ExitCall->removeFromParent();
  Builder.Insert(ExitCall);
 
  return IRBuilder<>::InsertPoint(ExitCall->getParent(),
                                  ExitCall->getIterator());
}
 
OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCopyinClauseBlocks(
    InsertPointTy IP, Value *MasterAddr, Value *PrivateAddr,
    llvm::IntegerType *IntPtrTy, bool BranchtoEnd) {
  if (!IP.isSet())
    return IP;
 
  IRBuilder<>::InsertPointGuard IPG(Builder);
 
  // creates the following CFG structure
  //     OMP_Entry : (MasterAddr != PrivateAddr)?
  //       F     T
  //       |      \
  //       |     copin.not.master
  //       |      /
  //       v     /
  //   copyin.not.master.end
  //         |
  //         v
  //   OMP.Entry.Next
 
  BasicBlock *OMP_Entry = IP.getBlock();
  Function *CurFn = OMP_Entry->getParent();
  BasicBlock *CopyBegin =
      BasicBlock::Create(M.getContext(), "copyin.not.master", CurFn);
  BasicBlock *CopyEnd = nullptr;
 
  // If entry block is terminated, split to preserve the branch to following
  // basic block (i.e. OMP.Entry.Next), otherwise, leave everything as is.
  if (isa_and_nonnull<CondBrInst>(OMP_Entry->getTerminatorOrNull())) {
    CopyEnd = OMP_Entry->splitBasicBlock(OMP_Entry->getTerminator(),
                                         "copyin.not.master.end");
    OMP_Entry->getTerminator()->eraseFromParent();
  } else {
    CopyEnd =
        BasicBlock::Create(M.getContext(), "copyin.not.master.end", CurFn);
  }
 
  Builder.SetInsertPoint(OMP_Entry);
  Value *MasterPtr = Builder.CreatePtrToInt(MasterAddr, IntPtrTy);
  Value *PrivatePtr = Builder.CreatePtrToInt(PrivateAddr, IntPtrTy);
  Value *cmp = Builder.CreateICmpNE(MasterPtr, PrivatePtr);
  Builder.CreateCondBr(cmp, CopyBegin, CopyEnd);
 
  Builder.SetInsertPoint(CopyBegin);
  if (BranchtoEnd)
    Builder.SetInsertPoint(Builder.CreateBr(CopyEnd));
 
  return Builder.saveIP();
}
 
CallInst *OpenMPIRBuilder::createOMPAlloc(const LocationDescription &Loc,
                                          Value *Size, Value *Allocator,
                                          std::string Name) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  if (!updateToLocation(Loc))
    return nullptr;
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Value *Args[] = {ThreadId, Size, Allocator};
 
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_alloc);
 
  return createRuntimeFunctionCall(Fn, Args, Name);
}
 
CallInst *OpenMPIRBuilder::createOMPAlignedAlloc(const LocationDescription &Loc,
                                                 Value *Align, Value *Size,
                                                 Value *Allocator,
                                                 std::string Name) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  if (!updateToLocation(Loc))
    return nullptr;
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Value *Args[] = {ThreadId, Align, Size, Allocator};
 
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_aligned_alloc);
 
  return Builder.CreateCall(Fn, Args, Name);
}
 
CallInst *OpenMPIRBuilder::createOMPFree(const LocationDescription &Loc,
                                         Value *Addr, Value *Allocator,
                                         std::string Name) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  if (!updateToLocation(Loc))
    return nullptr;
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Value *Args[] = {ThreadId, Addr, Allocator};
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_free);
  return createRuntimeFunctionCall(Fn, Args, Name);
}
 
CallInst *OpenMPIRBuilder::createOMPAllocShared(const LocationDescription &Loc,
                                                Value *Size,
                                                const Twine &Name) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  updateToLocation(Loc);
 
  Value *Args[] = {Size};
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_alloc_shared);
  CallInst *Call = Builder.CreateCall(Fn, Args, Name);
  Call->addRetAttr(Attribute::getWithAlignment(
      M.getContext(), M.getDataLayout().getPrefTypeAlign(Int64)));
  return Call;
}
 
CallInst *OpenMPIRBuilder::createOMPAllocShared(const LocationDescription &Loc,
                                                Type *VarType,
                                                const Twine &Name) {
  return createOMPAllocShared(
      Loc, Builder.getInt64(M.getDataLayout().getTypeAllocSize(VarType)), Name);
}
 
CallInst *OpenMPIRBuilder::createOMPFreeShared(const LocationDescription &Loc,
                                               Value *Addr, Value *Size,
                                               const Twine &Name) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  updateToLocation(Loc);
 
  Value *Args[] = {Addr, Size};
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_free_shared);
  return Builder.CreateCall(Fn, Args, Name);
}
 
CallInst *OpenMPIRBuilder::createOMPFreeShared(const LocationDescription &Loc,
                                               Value *Addr, Type *VarType,
                                               const Twine &Name) {
  return createOMPFreeShared(
      Loc, Addr, Builder.getInt64(M.getDataLayout().getTypeAllocSize(VarType)),
      Name);
}
 
CallInst *OpenMPIRBuilder::createOMPInteropInit(
    const LocationDescription &Loc, Value *InteropVar,
    omp::OMPInteropType InteropType, Value *Device, Value *NumDependences,
    Value *DependenceAddress, bool HaveNowaitClause) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  updateToLocation(Loc);
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  if (Device == nullptr)
    Device = Constant::getAllOnesValue(Int32);
  else if (Device->getType() != Int32)
    Device = Builder.CreateIntCast(Device, Int32, /*isSigned=*/true);
  Constant *InteropTypeVal = ConstantInt::get(Int32, (int)InteropType);
  if (NumDependences == nullptr) {
    NumDependences = ConstantInt::get(Int32, 0);
    PointerType *PointerTypeVar = PointerType::getUnqual(M.getContext());
    DependenceAddress = ConstantPointerNull::get(PointerTypeVar);
  }
  Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause);
  Value *Args[] = {
      Ident,  ThreadId,       InteropVar,        InteropTypeVal,
      Device, NumDependences, DependenceAddress, HaveNowaitClauseVal};
 
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_init);
 
  return createRuntimeFunctionCall(Fn, Args);
}
 
CallInst *OpenMPIRBuilder::createOMPInteropDestroy(
    const LocationDescription &Loc, Value *InteropVar, Value *Device,
    Value *NumDependences, Value *DependenceAddress, bool HaveNowaitClause) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  updateToLocation(Loc);
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  if (Device == nullptr)
    Device = Constant::getAllOnesValue(Int32);
  else if (Device->getType() != Int32)
    Device = Builder.CreateIntCast(Device, Int32, /*isSigned=*/true);
  if (NumDependences == nullptr) {
    NumDependences = ConstantInt::get(Int32, 0);
    PointerType *PointerTypeVar = PointerType::getUnqual(M.getContext());
    DependenceAddress = ConstantPointerNull::get(PointerTypeVar);
  }
  Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause);
  Value *Args[] = {
      Ident,          ThreadId,          InteropVar,         Device,
      NumDependences, DependenceAddress, HaveNowaitClauseVal};
 
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_destroy);
 
  return createRuntimeFunctionCall(Fn, Args);
}
 
CallInst *OpenMPIRBuilder::createOMPInteropUse(const LocationDescription &Loc,
                                               Value *InteropVar, Value *Device,
                                               Value *NumDependences,
                                               Value *DependenceAddress,
                                               bool HaveNowaitClause) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  updateToLocation(Loc);
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  if (Device == nullptr)
    Device = Constant::getAllOnesValue(Int32);
  else if (Device->getType() != Int32)
    Device = Builder.CreateIntCast(Device, Int32, /*isSigned=*/true);
  if (NumDependences == nullptr) {
    NumDependences = ConstantInt::get(Int32, 0);
    PointerType *PointerTypeVar = PointerType::getUnqual(M.getContext());
    DependenceAddress = ConstantPointerNull::get(PointerTypeVar);
  }
  Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause);
  Value *Args[] = {
      Ident,          ThreadId,          InteropVar,         Device,
      NumDependences, DependenceAddress, HaveNowaitClauseVal};
 
  Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_use);
 
  return createRuntimeFunctionCall(Fn, Args);
}
 
CallInst *OpenMPIRBuilder::createCachedThreadPrivate(
    const LocationDescription &Loc, llvm::Value *Pointer,
    llvm::ConstantInt *Size, const llvm::Twine &Name) {
  IRBuilder<>::InsertPointGuard IPG(Builder);
  updateToLocation(Loc);
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Value *ThreadId = getOrCreateThreadID(Ident);
  Constant *ThreadPrivateCache =
      getOrCreateInternalVariable(Int8PtrPtr, Name.str());
  llvm::Value *Args[] = {Ident, ThreadId, Pointer, Size, ThreadPrivateCache};
 
  Function *Fn =
      getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_threadprivate_cached);
 
  return createRuntimeFunctionCall(Fn, Args);
}
 
OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createTargetInit(
    const LocationDescription &Loc,
    const llvm::OpenMPIRBuilder::TargetKernelDefaultAttrs &Attrs) {
  assert(!Attrs.MaxThreads.empty() && !Attrs.MaxTeams.empty() &&
         "expected num_threads and num_teams to be specified");
 
  if (!updateToLocation(Loc))
    return Loc.IP;
 
  uint32_t SrcLocStrSize;
  Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
  Constant *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
  Constant *IsSPMDVal = ConstantInt::getSigned(Int8, Attrs.ExecFlags);
  Constant *UseGenericStateMachineVal = ConstantInt::getSigned(
      Int8, Attrs.ExecFlags != omp::OMP_TGT_EXEC_MODE_SPMD &&
                Attrs.ExecFlags != omp::OMP_TGT_EXEC_MODE_SPMD_NO_LOOP);
  Constant *MayUseNestedParallelismVal = ConstantInt::getSigned(Int8, true);
  Constant *DebugIndentionLevelVal = ConstantInt::getSigned(Int16, 0);
 
  Function *DebugKernelWrapper = Builder.GetInsertBlock()->getParent();
  Function *Kernel = DebugKernelWrapper;
 
  // We need to strip the debug prefix to get the correct kernel name.
  StringRef KernelName = Kernel->getName();
  const std::string DebugPrefix = "_debug__";
  if (KernelName.ends_with(DebugPrefix)) {
    KernelName = KernelName.drop_back(DebugPrefix.length());
    Kernel = M.getFunction(KernelName);
    assert(Kernel && "Expected the real kernel to exist");
  }
 
  // Manifest the launch configuration in the metadata matching the kernel
  // environment.
  if (Attrs.MinTeams.front() > 1 || Attrs.MaxTeams.front() > 0)
    writeTeamsForKernel(T, *Kernel, Attrs.MinTeams.front(),
                        Attrs.MaxTeams.front());
 
  // If MaxThreads is not set and needs adjustment, select the maximum between
  // the default workgroup size and the MinThreads value.
  int32_t MaxThreadsVal = Attrs.MaxThreads.front();
  if (MaxThreadsVal < 0 && UseDefaultMaxThreads) {
    if (hasGridValue(T)) {
      MaxThreadsVal =
          std::max(int32_t(getGridValue(T, Kernel).GV_Default_WG_Size),
                   Attrs.MinThreads.front());
    } else {
      MaxThreadsVal = Attrs.MinThreads.front();
    }
  }
 
  if (MaxThreadsVal > 0)
    writeThreadBoundsForKernel(T, *Kernel, Attrs.MinThreads.front(),
                               MaxThreadsVal);
 
  Constant *MinThreads =
      ConstantInt::getSigned(Int32, Attrs.MinThreads.front());
  Constant *MaxThreads = ConstantInt::getSigned(Int32, MaxThreadsVal);
  Constant *MinTeams = ConstantInt::getSigned(Int32, Attrs.MinTeams.front());
  Constant *MaxTeams = ConstantInt::getSigned(Int32, Attrs.MaxTeams.front());
  Constant *ReductionDataSize =
      ConstantInt::getSigned(Int32, Attrs.ReductionDataSize);
 
  Function *Fn = getOrCreateRuntimeFunctionPtr(
      omp::RuntimeFunction::OMPRTL___kmpc_target_init);
  const DataLayout &DL = Fn->getDataLayout();
 
  Twine DynamicEnvironmentName = KernelName + "_dynamic_environment";
  Constant *DynamicEnvironmentInitializer =
      ConstantStruct::get(DynamicEnvironment, {DebugIndentionLevelVal});
  GlobalVariable *DynamicEnvironmentGV = new GlobalVariable(
      M, DynamicEnvironment, /*IsConstant=*/false, GlobalValue::WeakODRLinkage,
      DynamicEnvironmentInitializer, DynamicEnvironmentName,
      /*InsertBefore=*/nullptr, GlobalValue::NotThreadLocal,
      DL.getDefaultGlobalsAddressSpace());
  DynamicEnvironmentGV->setVisibility(GlobalValue::ProtectedVisibility);
 
  Constant *DynamicEnvironment =
      DynamicEnvironmentGV->getType() == DynamicEnvironmentPtr
          ? DynamicEnvironmentGV
          : ConstantExpr::getAddrSpaceCast(DynamicEnvironmentGV,
                                           DynamicEnvironmentPtr);
 
  Constant *ConfigurationEnvironmentInitializer = ConstantStruct::get(
      ConfigurationEnvironment, {
                                    UseGenericStateMachineVal,
                                    MayUseNestedParallelismVal,
                                    IsSPMDVal,
                                    MinThreads,
                                    MaxThreads,
                                    MinTeams,
                                    MaxTeams,
                                    ReductionDataSize,
                                });
  Constant *KernelEnvironmentInitializer = ConstantStruct::get(
      KernelEnvironment, {
                             ConfigurationEnvironmentInitializer,
                             Ident,
                             DynamicEnvironment,
                         });
  std::string KernelEnvironmentName =
      (KernelName + "_kernel_environment").str();
  GlobalVariable *KernelEnvironmentGV = new GlobalVariable(
      M, KernelEnvironment, /*IsConstant=*/true, GlobalValue::WeakODRLinkage,
      KernelEnvironmentInitializer, KernelEnvironmentName,
      /*InsertBefore=*/nullptr, GlobalValue::NotThreadLocal,
      DL.getDefaultGlobalsAddressSpace());
  KernelEnvironmentGV->setVisibility(GlobalValue::ProtectedVisibility);
 
  Constant *KernelEnvironment =
      KernelEnvironmentGV->getType() == KernelEnvironmentPtr
          ? KernelEnvironmentGV
          : ConstantExpr::getAddrSpaceCast(KernelEnvironmentGV,
                                           KernelEnvironmentPtr);
  Value *KernelLaunchEnvironment =
      DebugKernelWrapper->getArg(DebugKernelWrapper->arg_size() - 1);
  Type *KernelLaunchEnvParamTy = Fn->getFunctionType()->getParamType(1);
  KernelLaunchEnvironment =
      KernelLaunchEnvironment->getType() == KernelLaunchEnvParamTy
          ? KernelLaunchEnvironment
          : Builder.CreateAddrSpaceCast(KernelLaunchEnvironment,
                                        KernelLaunchEnvParamTy);
  CallInst *ThreadKind = createRuntimeFunctionCall(
      Fn, {KernelEnvironment, KernelLaunchEnvironment});
 
  Value *ExecUserCode = Builder.CreateICmpEQ(
      ThreadKind, Constant::getAllOnesValue(ThreadKind->getType()),
      "exec_user_code");
 
  // ThreadKind = __kmpc_target_init(...)
  // if (ThreadKind == -1)
  //   user_code
  // else
  //   return;
 
  auto *UI = Builder.CreateUnreachable();
  BasicBlock *CheckBB = UI->getParent();
  BasicBlock *UserCodeEntryBB = CheckBB->splitBasicBlock(UI, "user_code.entry");
 
  BasicBlock *WorkerExitBB = BasicBlock::Create(
      CheckBB->getContext(), "worker.exit", CheckBB->getParent());
  Builder.SetInsertPoint(WorkerExitBB);
  Builder.CreateRetVoid();
 
  auto *CheckBBTI = CheckBB->getTerminator();
  Builder.SetInsertPoint(CheckBBTI);
  Builder.CreateCondBr(ExecUserCode, UI->getParent(), WorkerExitBB);
 
  CheckBBTI->eraseFromParent();
  UI->eraseFromParent();
 
  // Continue in the "user_code" block, see diagram above and in
  // openmp/libomptarget/deviceRTLs/common/include/target.h .
  return InsertPointTy(UserCodeEntryBB, UserCodeEntryBB->getFirstInsertionPt());
}
 
void OpenMPIRBuilder::createTargetDeinit(const LocationDescription &Loc,
                                         int32_t TeamsReductionDataSize) {
  if (!updateToLocation(Loc))
    return;
 
  Function *Fn = getOrCreateRuntimeFunctionPtr(
      omp::RuntimeFunction::OMPRTL___kmpc_target_deinit);
 
  createRuntimeFunctionCall(Fn, {});
 
  if (!TeamsReductionDataSize)
    return;
 
  Function *Kernel = Builder.GetInsertBlock()->getParent();
  // We need to strip the debug prefix to get the correct kernel name.
  StringRef KernelName = Kernel->getName();
  const std::string DebugPrefix = "_debug__";
  if (KernelName.ends_with(DebugPrefix))
    KernelName = KernelName.drop_back(DebugPrefix.length());
  auto *KernelEnvironmentGV =
      M.getNamedGlobal((KernelName + "_kernel_environment").str());
  assert(KernelEnvironmentGV && "Expected kernel environment global\n");
  auto *KernelEnvironmentInitializer = KernelEnvironmentGV->getInitializer();
  auto *NewInitializer = ConstantFoldInsertValueInstruction(
      KernelEnvironmentInitializer,
      ConstantInt::get(Int32, TeamsReductionDataSize), {0, 7});
  KernelEnvironmentGV->setInitializer(NewInitializer);
}
 
static void updateNVPTXAttr(Function &Kernel, StringRef Name, int32_t Value,
                            bool Min) {
  if (Kernel.hasFnAttribute(Name)) {
    int32_t OldLimit = Kernel.getFnAttributeAsParsedInteger(Name);
    Value = Min ? std::min(OldLimit, Value) : std::max(OldLimit, Value);
  }
  Kernel.addFnAttr(Name, llvm::utostr(Value));
}
 
std::pair<int32_t, int32_t>
OpenMPIRBuilder::readThreadBoundsForKernel(const Triple &T, Function &Kernel) {
  int32_t ThreadLimit =
      Kernel.getFnAttributeAsParsedInteger("omp_target_thread_limit");
 
  if (T.isAMDGPU()) {
    const auto &Attr = Kernel.getFnAttribute("amdgpu-flat-work-group-size");
    if (!Attr.isValid() || !Attr.isStringAttribute())
      return {0, ThreadLimit};
    auto [LBStr, UBStr] = Attr.getValueAsString().split(',');
    int32_t LB, UB;
    if (!llvm::to_integer(UBStr, UB, 10))
      return {0, ThreadLimit};
    UB = ThreadLimit ? std::min(ThreadLimit, UB) : UB;
    if (!llvm::to_integer(LBStr, LB, 10))
      return {0, UB};
    return {LB, UB};
  }
 
  if (Kernel.hasFnAttribute(NVVMAttr::MaxNTID)) {
    int32_t UB = Kernel.getFnAttributeAsParsedInteger(NVVMAttr::MaxNTID);
    return {0, ThreadLimit ? std::min(ThreadLimit, UB) : UB};
  }
  return {0, ThreadLimit};
}
 
void OpenMPIRBuilder::writeThreadBoundsForKernel(const Triple &T,
                                                 Function &Kernel, int32_t LB,
                                                 int32_t UB) {
  Kernel.addFnAttr("omp_target_thread_limit", std::to_string(UB));
 
  if (T.isAMDGPU()) {
    Kernel.addFnAttr("amdgpu-flat-work-group-size",
                     llvm::utostr(LB) + "," + llvm::utostr(UB));
    return;
  }
 
  updateNVPTXAttr(Kernel, NVVMAttr::MaxNTID, UB, true);
}
 
std::pair<int32_t, int32_t>
OpenMPIRBuilder::readTeamBoundsForKernel(const Triple &, Function &Kernel) {
  // TODO: Read from backend annotations if available.
  return {0, Kernel.getFnAttributeAsParsedInteger("omp_target_num_teams")};
}
 
void OpenMPIRBuilder::writeTeamsForKernel(const Triple &T, Function &Kernel,
                                          int32_t LB, int32_t UB) {
  if (UB > 0) {
    if (T.isNVPTX())
      Kernel.addFnAttr(NVVMAttr::MaxClusterRank, llvm::utostr(UB));
    if (T.isAMDGPU())
      Kernel.addFnAttr("amdgpu-max-num-workgroups", llvm::utostr(UB) + ",1,1");
  }
 
  Kernel.addFnAttr("omp_target_num_teams", std::to_string(LB));
}
 
void OpenMPIRBuilder::setOutlinedTargetRegionFunctionAttributes(
    Function *OutlinedFn) {
  if (Config.isTargetDevice()) {
    OutlinedFn->setLinkage(GlobalValue::WeakODRLinkage);
    // TODO: Determine if DSO local can be set to true.
    OutlinedFn->setDSOLocal(false);
    OutlinedFn->setVisibility(GlobalValue::ProtectedVisibility);
    if (T.isAMDGCN())
      OutlinedFn->setCallingConv(CallingConv::AMDGPU_KERNEL);
    else if (T.isNVPTX())
      OutlinedFn->setCallingConv(CallingConv::PTX_Kernel);
    else if (T.isSPIRV())
      OutlinedFn->setCallingConv(CallingConv::SPIR_KERNEL);
  }
}
 
Constant *OpenMPIRBuilder::createOutlinedFunctionID(Function *OutlinedFn,
                                                    StringRef EntryFnIDName) {
  if (Config.isTargetDevice()) {
    assert(OutlinedFn && "The outlined function must exist if embedded");
    return OutlinedFn;
  }
 
  return new GlobalVariable(
      M, Builder.getInt8Ty(), /*isConstant=*/true, GlobalValue::WeakAnyLinkage,
      Constant::getNullValue(Builder.getInt8Ty()), EntryFnIDName);
}
 
Constant *OpenMPIRBuilder::createTargetRegionEntryAddr(Function *OutlinedFn,
                                                       StringRef EntryFnName) {
  if (OutlinedFn)
    return OutlinedFn;
 
  assert(!M.getGlobalVariable(EntryFnName, true) &&
         "Named kernel already exists?");
  return new GlobalVariable(
      M, Builder.getInt8Ty(), /*isConstant=*/true, GlobalValue::InternalLinkage,
      Constant::getNullValue(Builder.getInt8Ty()), EntryFnName);
}
 
Error OpenMPIRBuilder::emitTargetRegionFunction(
    TargetRegionEntryInfo &EntryInfo,
    FunctionGenCallback &GenerateFunctionCallback, bool IsOffloadEntry,
    Function *&OutlinedFn, Constant *&OutlinedFnID) {
 
  SmallString<64> EntryFnName;
  OffloadInfoManager.getTargetRegionEntryFnName(EntryFnName, EntryInfo);
 
  if (Config.isTargetDevice() || !Config.openMPOffloadMandatory()) {
    Expected<Function *> CBResult = GenerateFunctionCallback(EntryFnName);
    if (!CBResult)
      return CBResult.takeError();
    OutlinedFn = *CBResult;
  } else {
    OutlinedFn = nullptr;
  }
 
  // If this target outline function is not an offload entry, we don't need to
  // register it. This may be in the case of a false if clause, or if there are
  // no OpenMP targets.
  if (!IsOffloadEntry)
    return Error::success();
 
  std::string EntryFnIDName =
      Config.isTargetDevice()
          ? std::string(EntryFnName)
          : createPlatformSpecificName({EntryFnName, "region_id"});
 
  OutlinedFnID = registerTargetRegionFunction(EntryInfo, OutlinedFn,
                                              EntryFnName, EntryFnIDName);
  return Error::success();
}
 
Constant *OpenMPIRBuilder::registerTargetRegionFunction(
    TargetRegionEntryInfo &EntryInfo, Function *OutlinedFn,
    StringRef EntryFnName, StringRef EntryFnIDName) {
  if (OutlinedFn)
    setOutlinedTargetRegionFunctionAttributes(OutlinedFn);
  auto OutlinedFnID = createOutlinedFunctionID(OutlinedFn, EntryFnIDName);
  auto EntryAddr = createTargetRegionEntryAddr(OutlinedFn, EntryFnName);
  OffloadInfoManager.registerTargetRegionEntryInfo(
      EntryInfo, EntryAddr, OutlinedFnID,
      OffloadEntriesInfoManager::OMPTargetRegionEntryTargetRegion);
  return OutlinedFnID;
}
 
OpenMPIRBuilder::InsertPointOrErrorTy OpenMPIRBuilder::createTargetData(
    const LocationDescription &Loc, InsertPointTy AllocaIP,
    InsertPointTy CodeGenIP, ArrayRef<BasicBlock *> DeallocBlocks,
    Value *DeviceID, Value *IfCond, TargetDataInfo &Info,
    GenMapInfoCallbackTy GenMapInfoCB, CustomMapperCallbackTy CustomMapperCB,
    omp::RuntimeFunction *MapperFunc,
    function_ref<InsertPointOrErrorTy(InsertPointTy CodeGenIP,
                                      BodyGenTy BodyGenType)>
        BodyGenCB,
    function_ref<void(unsigned int, Value *)> DeviceAddrCB, Value *SrcLocInfo) {
  if (!updateToLocation(Loc))
    return InsertPointTy();
 
  Builder.restoreIP(CodeGenIP);
 
  bool IsStandAlone = !BodyGenCB;
  MapInfosTy *MapInfo;
  // Generate the code for the opening of the data environment. Capture all the
  // arguments of the runtime call by reference because they are used in the
  // closing of the region.
  auto BeginThenGen = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
                          ArrayRef<BasicBlock *> DeallocBlocks) -> Error {
    MapInfo = &GenMapInfoCB(Builder.saveIP());
    if (Error Err = emitOffloadingArrays(
            AllocaIP, Builder.saveIP(), *MapInfo, Info, CustomMapperCB,
            /*IsNonContiguous=*/true, DeviceAddrCB))
      return Err;
 
    TargetDataRTArgs RTArgs;
    emitOffloadingArraysArgument(Builder, RTArgs, Info);
 
    // Emit the number of elements in the offloading arrays.
    Value *PointerNum = Builder.getInt32(Info.NumberOfPtrs);
 
    // Source location for the ident struct
    if (!SrcLocInfo) {
      uint32_t SrcLocStrSize;
      Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
      SrcLocInfo = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
    }
 
    SmallVector<llvm::Value *, 13> OffloadingArgs = {
        SrcLocInfo,           DeviceID,
        PointerNum,           RTArgs.BasePointersArray,
        RTArgs.PointersArray, RTArgs.SizesArray,
        RTArgs.MapTypesArray, RTArgs.MapNamesArray,
        RTArgs.MappersArray};
 
    if (IsStandAlone) {
      assert(MapperFunc && "MapperFunc missing for standalone target data");
 
      auto TaskBodyCB = [&](Value *, Value *,
                            IRBuilderBase::InsertPoint) -> Error {
        if (Info.HasNoWait) {
          OffloadingArgs.append({llvm::Constant::getNullValue(Int32),
                                 llvm::Constant::getNullValue(VoidPtr),
                                 llvm::Constant::getNullValue(Int32),
                                 llvm::Constant::getNullValue(VoidPtr)});
        }
 
        createRuntimeFunctionCall(getOrCreateRuntimeFunctionPtr(*MapperFunc),
                                  OffloadingArgs);
 
        if (Info.HasNoWait) {
          BasicBlock *OffloadContBlock =
              BasicBlock::Create(Builder.getContext(), "omp_offload.cont");
          Function *CurFn = Builder.GetInsertBlock()->getParent();
          emitBlock(OffloadContBlock, CurFn, /*IsFinished=*/true);
          Builder.restoreIP(Builder.saveIP());
        }
        return Error::success();
      };
 
      bool RequiresOuterTargetTask = Info.HasNoWait;
      if (!RequiresOuterTargetTask)
        cantFail(TaskBodyCB(/*DeviceID=*/nullptr, /*RTLoc=*/nullptr,
                            /*TargetTaskAllocaIP=*/{}));
      else
        cantFail(emitTargetTask(TaskBodyCB, DeviceID, SrcLocInfo, AllocaIP,
                                /*Dependencies=*/{}, RTArgs, Info.HasNoWait));
    } else {
      Function *BeginMapperFunc = getOrCreateRuntimeFunctionPtr(
          omp::OMPRTL___tgt_target_data_begin_mapper);
 
      createRuntimeFunctionCall(BeginMapperFunc, OffloadingArgs);
 
      for (auto DeviceMap : Info.DevicePtrInfoMap) {
        if (isa<AllocaInst>(DeviceMap.second.second)) {
          auto *LI =
              Builder.CreateLoad(Builder.getPtrTy(), DeviceMap.second.first);
          Builder.CreateStore(LI, DeviceMap.second.second);
        }
      }
 
      // If device pointer privatization is required, emit the body of the
      // region here. It will have to be duplicated: with and without
      // privatization.
      InsertPointOrErrorTy AfterIP =
          BodyGenCB(Builder.saveIP(), BodyGenTy::Priv);
      if (!AfterIP)
        return AfterIP.takeError();
      Builder.restoreIP(*AfterIP);
    }
    return Error::success();
  };
 
  // If we need device pointer privatization, we need to emit the body of the
  // region with no privatization in the 'else' branch of the conditional.
  // Otherwise, we don't have to do anything.
  auto BeginElseGen = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
                          ArrayRef<BasicBlock *> DeallocBlocks) -> Error {
    InsertPointOrErrorTy AfterIP =
        BodyGenCB(Builder.saveIP(), BodyGenTy::DupNoPriv);
    if (!AfterIP)
      return AfterIP.takeError();
    Builder.restoreIP(*AfterIP);
    return Error::success();
  };
 
  // Generate code for the closing of the data region.
  auto EndThenGen = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
                        ArrayRef<BasicBlock *> DeallocBlocks) {
    TargetDataRTArgs RTArgs;
    Info.EmitDebug = !MapInfo->Names.empty();
    emitOffloadingArraysArgument(Builder, RTArgs, Info, /*ForEndCall=*/true);
 
    // Emit the number of elements in the offloading arrays.
    Value *PointerNum = Builder.getInt32(Info.NumberOfPtrs);
 
    // Source location for the ident struct
    if (!SrcLocInfo) {
      uint32_t SrcLocStrSize;
      Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
      SrcLocInfo = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
    }
 
    Value *OffloadingArgs[] = {SrcLocInfo,           DeviceID,
                               PointerNum,           RTArgs.BasePointersArray,
                               RTArgs.PointersArray, RTArgs.SizesArray,
                               RTArgs.MapTypesArray, RTArgs.MapNamesArray,
                               RTArgs.MappersArray};
    Function *EndMapperFunc =
        getOrCreateRuntimeFunctionPtr(omp::OMPRTL___tgt_target_data_end_mapper);
 
    createRuntimeFunctionCall(EndMapperFunc, OffloadingArgs);
    return Error::success();
  };
 
  // We don't have to do anything to close the region if the if clause evaluates
  // to false.
  auto EndElseGen = [&](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
                        ArrayRef<BasicBlock *> DeallocBlocks) {
    return Error::success();
  };
 
  Error Err = [&]() -> Error {
    if (BodyGenCB) {
      Error Err = [&]() {
        if (IfCond)
          return emitIfClause(IfCond, BeginThenGen, BeginElseGen, AllocaIP);
        return BeginThenGen(AllocaIP, Builder.saveIP(), DeallocBlocks);
      }();
 
      if (Err)
        return Err;
 
      // If we don't require privatization of device pointers, we emit the body
      // in between the runtime calls. This avoids duplicating the body code.
      InsertPointOrErrorTy AfterIP =
          BodyGenCB(Builder.saveIP(), BodyGenTy::NoPriv);
      if (!AfterIP)
        return AfterIP.takeError();
      restoreIPandDebugLoc(Builder, *AfterIP);
 
      if (IfCond)
        return emitIfClause(IfCond, EndThenGen, EndElseGen, AllocaIP);
      return EndThenGen(AllocaIP, Builder.saveIP(), DeallocBlocks);
    }
    if (IfCond)
      return emitIfClause(IfCond, BeginThenGen, EndElseGen, AllocaIP);
    return BeginThenGen(AllocaIP, Builder.saveIP(), DeallocBlocks);
  }();
 
  if (Err)
    return Err;
 
  return Builder.saveIP();
}
 
FunctionCallee
OpenMPIRBuilder::createForStaticInitFunction(unsigned IVSize, bool IVSigned,
                                             bool IsGPUDistribute) {
  assert((IVSize == 32 || IVSize == 64) &&
         "IV size is not compatible with the omp runtime");
  RuntimeFunction Name;
  if (IsGPUDistribute)
    Name = IVSize == 32
               ? (IVSigned ? omp::OMPRTL___kmpc_distribute_static_init_4
                           : omp::OMPRTL___kmpc_distribute_static_init_4u)
               : (IVSigned ? omp::OMPRTL___kmpc_distribute_static_init_8
                           : omp::OMPRTL___kmpc_distribute_static_init_8u);
  else
    Name = IVSize == 32 ? (IVSigned ? omp::OMPRTL___kmpc_for_static_init_4
                                    : omp::OMPRTL___kmpc_for_static_init_4u)
                        : (IVSigned ? omp::OMPRTL___kmpc_for_static_init_8
                                    : omp::OMPRTL___kmpc_for_static_init_8u);
 
  return getOrCreateRuntimeFunction(M, Name);
}
 
FunctionCallee OpenMPIRBuilder::createDispatchInitFunction(unsigned IVSize,
                                                           bool IVSigned) {
  assert((IVSize == 32 || IVSize == 64) &&
         "IV size is not compatible with the omp runtime");
  RuntimeFunction Name = IVSize == 32
                             ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_init_4
                                         : omp::OMPRTL___kmpc_dispatch_init_4u)
                             : (IVSigned ? omp::OMPRTL___kmpc_dispatch_init_8
                                         : omp::OMPRTL___kmpc_dispatch_init_8u);
 
  return getOrCreateRuntimeFunction(M, Name);
}
 
FunctionCallee OpenMPIRBuilder::createDispatchNextFunction(unsigned IVSize,
                                                           bool IVSigned) {
  assert((IVSize == 32 || IVSize == 64) &&
         "IV size is not compatible with the omp runtime");
  RuntimeFunction Name = IVSize == 32
                             ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_next_4
                                         : omp::OMPRTL___kmpc_dispatch_next_4u)
                             : (IVSigned ? omp::OMPRTL___kmpc_dispatch_next_8
                                         : omp::OMPRTL___kmpc_dispatch_next_8u);
 
  return getOrCreateRuntimeFunction(M, Name);
}
 
FunctionCallee OpenMPIRBuilder::createDispatchFiniFunction(unsigned IVSize,
                                                           bool IVSigned) {
  assert((IVSize == 32 || IVSize == 64) &&
         "IV size is not compatible with the omp runtime");
  RuntimeFunction Name = IVSize == 32
                             ? (IVSigned ? omp::OMPRTL___kmpc_dispatch_fini_4
                                         : omp::OMPRTL___kmpc_dispatch_fini_4u)
                             : (IVSigned ? omp::OMPRTL___kmpc_dispatch_fini_8
                                         : omp::OMPRTL___kmpc_dispatch_fini_8u);
 
  return getOrCreateRuntimeFunction(M, Name);
}
 
FunctionCallee OpenMPIRBuilder::createDispatchDeinitFunction() {
  return getOrCreateRuntimeFunction(M, omp::OMPRTL___kmpc_dispatch_deinit);
}
 
static void FixupDebugInfoForOutlinedFunction(
    OpenMPIRBuilder &OMPBuilder, IRBuilderBase &Builder, Function *Func,
    DenseMap<Value *, std::tuple<Value *, unsigned>> &ValueReplacementMap) {
 
  DISubprogram *NewSP = Func->getSubprogram();
  if (!NewSP)
    return;
 
  SmallDenseMap<DILocalVariable *, DILocalVariable *> RemappedVariables;
 
  auto GetUpdatedDIVariable = [&](DILocalVariable *OldVar, unsigned arg) {
    DILocalVariable *&NewVar = RemappedVariables[OldVar];
    // Only use cached variable if the arg number matches. This is important
    // so that DIVariable created for privatized variables are not discarded.
    if (NewVar && (arg == NewVar->getArg()))
      return NewVar;
 
    NewVar = llvm::DILocalVariable::get(
        Builder.getContext(), OldVar->getScope(), OldVar->getName(),
        OldVar->getFile(), OldVar->getLine(), OldVar->getType(), arg,
        OldVar->getFlags(), OldVar->getAlignInBits(), OldVar->getAnnotations());
    return NewVar;
  };
 
  auto UpdateDebugRecord = [&](auto *DR) {
    DILocalVariable *OldVar = DR->getVariable();
    unsigned ArgNo = 0;
    for (auto Loc : DR->location_ops()) {
      auto Iter = ValueReplacementMap.find(Loc);
      if (Iter != ValueReplacementMap.end()) {
        DR->replaceVariableLocationOp(Loc, std::get<0>(Iter->second));
        ArgNo = std::get<1>(Iter->second) + 1;
      }
    }
    if (ArgNo != 0)
      DR->setVariable(GetUpdatedDIVariable(OldVar, ArgNo));
  };
 
  SmallVector<DbgVariableRecord *, 4> DVRsToDelete;
  auto MoveDebugRecordToCorrectBlock = [&](DbgVariableRecord *DVR) {
    if (DVR->getNumVariableLocationOps() != 1u) {
      DVR->setKillLocation();
      return;
    }
    Value *Loc = DVR->getVariableLocationOp(0u);
    BasicBlock *CurBB = DVR->getParent();
    BasicBlock *RequiredBB = nullptr;
 
    if (Instruction *LocInst = dyn_cast<Instruction>(Loc))
      RequiredBB = LocInst->getParent();
    else if (isa<llvm::Argument>(Loc))
      RequiredBB = &DVR->getFunction()->getEntryBlock();
 
    if (RequiredBB && RequiredBB != CurBB) {
      assert(!RequiredBB->empty());
      RequiredBB->insertDbgRecordBefore(DVR->clone(),
                                        RequiredBB->back().getIterator());
      DVRsToDelete.push_back(DVR);
    }
  };
 
  // The location and scope of variable intrinsics and records still point to
  // the parent function of the target region. Update them.
  for (Instruction &I : instructions(Func)) {
    assert(!isa<llvm::DbgVariableIntrinsic>(&I) &&
           "Unexpected debug intrinsic");
    for (DbgVariableRecord &DVR : filterDbgVars(I.getDbgRecordRange())) {
      UpdateDebugRecord(&DVR);
      MoveDebugRecordToCorrectBlock(&DVR);
    }
  }
  for (auto *DVR : DVRsToDelete)
    DVR->getMarker()->MarkedInstr->dropOneDbgRecord(DVR);
  // An extra argument is passed to the device. Create the debug data for it.
  if (OMPBuilder.Config.isTargetDevice()) {
    DICompileUnit *CU = NewSP->getUnit();
    Module *M = Func->getParent();
    DIBuilder DB(*M, true, CU);
    DIType *VoidPtrTy =
        DB.createQualifiedType(dwarf::DW_TAG_pointer_type, nullptr);
    unsigned ArgNo = Func->arg_size();
    DILocalVariable *Var = DB.createParameterVariable(
        NewSP, "dyn_ptr", ArgNo, NewSP->getFile(), /*LineNo=*/0, VoidPtrTy,
        /*AlwaysPreserve=*/false, DINode::DIFlags::FlagArtificial);
    auto Loc = DILocation::get(Func->getContext(), 0, 0, NewSP, 0);
    Argument *LastArg = Func->getArg(Func->arg_size() - 1);
    DB.insertDeclare(LastArg, Var, DB.createExpression(), Loc,
                     &(*Func->begin()));
  }
}
 
static Value *removeASCastIfPresent(Value *V) {
  if (Operator::getOpcode(V) == Instruction::AddrSpaceCast)
    return cast<Operator>(V)->getOperand(0);
  return V;
}
 
static Expected<Function *> createOutlinedFunction(
    OpenMPIRBuilder &OMPBuilder, IRBuilderBase &Builder,
    const OpenMPIRBuilder::TargetKernelDefaultAttrs &DefaultAttrs,
    StringRef FuncName, SmallVectorImpl<Value *> &Inputs,
    OpenMPIRBuilder::TargetBodyGenCallbackTy &CBFunc,
    OpenMPIRBuilder::TargetGenArgAccessorsCallbackTy &ArgAccessorFuncCB) {
  SmallVector<Type *> ParameterTypes;
  if (OMPBuilder.Config.isTargetDevice()) {
    // All parameters to target devices are passed as pointers
    // or i64. This assumes 64-bit address spaces/pointers.
    for (auto &Arg : Inputs)
      ParameterTypes.push_back(Arg->getType()->isPointerTy()
                                   ? Arg->getType()
                                   : Type::getInt64Ty(Builder.getContext()));
  } else {
    for (auto &Arg : Inputs)
      ParameterTypes.push_back(Arg->getType());
  }
 
  // The implicit dyn_ptr argument is always the last parameter on both host
  // and device so the argument counts match without runtime manipulation.
  auto *PtrTy = PointerType::getUnqual(Builder.getContext());
  ParameterTypes.push_back(PtrTy);
 
  auto BB = Builder.GetInsertBlock();
  auto M = BB->getModule();
  auto FuncType = FunctionType::get(Builder.getVoidTy(), ParameterTypes,
                                    /*isVarArg*/ false);
  auto Func =
      Function::Create(FuncType, GlobalValue::InternalLinkage, FuncName, M);
 
  // Forward target-cpu and target-features function attributes from the
  // original function to the new outlined function.
  Function *ParentFn = Builder.GetInsertBlock()->getParent();
 
  auto TargetCpuAttr = ParentFn->getFnAttribute("target-cpu");
  if (TargetCpuAttr.isStringAttribute())
    Func->addFnAttr(TargetCpuAttr);
 
  auto TargetFeaturesAttr = ParentFn->getFnAttribute("target-features");
  if (TargetFeaturesAttr.isStringAttribute())
    Func->addFnAttr(TargetFeaturesAttr);
 
  if (OMPBuilder.Config.isTargetDevice()) {
    Value *ExecMode =
        OMPBuilder.emitKernelExecutionMode(FuncName, DefaultAttrs.ExecFlags);
    OMPBuilder.emitUsed("llvm.compiler.used", {ExecMode});
  }
 
  // Save insert point.
  IRBuilder<>::InsertPointGuard IPG(Builder);
  // We will generate the entries in the outlined function but the debug
  // location may still be pointing to the parent function. Reset it now.
  Builder.SetCurrentDebugLocation(llvm::DebugLoc());