/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp

Bug Summary

File:	clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
Warning:	line 4437, column 8 Called C++ object pointer is null
Annotated Source Code

Press '?' to see keyboard shortcuts
Show analyzer invocation
clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name CGOpenMPRuntimeGPU.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -fhalf-no-semantic-interposition -mframe-pointer=none -relaxed-aliasing -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-12/lib/clang/12.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/tools/clang/lib/CodeGen -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/include -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/tools/clang/include -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/include -I /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-12/lib/clang/12.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/build-llvm/tools/clang/lib/CodeGen -fdebug-prefix-map=/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2021-01-24-223304-31662-1 -x c++ /build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
1//===---- CGOpenMPRuntimeGPU.cpp - Interface to OpenMP GPU Runtimes ----===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This provides a generalized class for OpenMP runtime code generation
10// specialized by GPU targets NVPTX and AMDGCN.
11//
12//===----------------------------------------------------------------------===//

14#include "CGOpenMPRuntimeGPU.h"
15#include "CGOpenMPRuntimeNVPTX.h"
16#include "CodeGenFunction.h"
17#include "clang/AST/Attr.h"
18#include "clang/AST/DeclOpenMP.h"
19#include "clang/AST/StmtOpenMP.h"
20#include "clang/AST/StmtVisitor.h"
21#include "clang/Basic/Cuda.h"
22#include "llvm/ADT/SmallPtrSet.h"
23#include "llvm/Frontend/OpenMP/OMPGridValues.h"
24#include "llvm/IR/IntrinsicsNVPTX.h"

26using namespace clang;
27using namespace CodeGen;
28using namespace llvm::omp;

30namespace {
31/// Pre(post)-action for different OpenMP constructs specialized for NVPTX.
32class NVPTXActionTy final : public PrePostActionTy {
llvm::FunctionCallee EnterCallee = nullptr;
ArrayRef<llvm::Value *> EnterArgs;
llvm::FunctionCallee ExitCallee = nullptr;
ArrayRef<llvm::Value *> ExitArgs;
bool Conditional = false;
llvm::BasicBlock *ContBlock = nullptr;

40public:
NVPTXActionTy(llvm::FunctionCallee EnterCallee,
              ArrayRef<llvm::Value *> EnterArgs,
              llvm::FunctionCallee ExitCallee,
              ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false)
    : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee),
      ExitArgs(ExitArgs), Conditional(Conditional) {}
void Enter(CodeGenFunction &CGF) override {
  llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs);
  if (Conditional) {
    llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes);
    auto *ThenBlock = CGF.createBasicBlock("omp_if.then");
    ContBlock = CGF.createBasicBlock("omp_if.end");
    // Generate the branch (If-stmt)
    CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock);
    CGF.EmitBlock(ThenBlock);
  }
}
void Done(CodeGenFunction &CGF) {
  // Emit the rest of blocks/branches
  CGF.EmitBranch(ContBlock);
  CGF.EmitBlock(ContBlock, true);
}
void Exit(CodeGenFunction &CGF) override {
  CGF.EmitRuntimeCall(ExitCallee, ExitArgs);
}
66};

68/// A class to track the execution mode when codegening directives within
69/// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry
70/// to the target region and used by containing directives such as 'parallel'
71/// to emit optimized code.
72class ExecutionRuntimeModesRAII {
73private:
CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode =
    CGOpenMPRuntimeGPU::EM_Unknown;
CGOpenMPRuntimeGPU::ExecutionMode &ExecMode;
bool SavedRuntimeMode = false;
bool *RuntimeMode = nullptr;

80public:
/// Constructor for Non-SPMD mode.
ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode)
    : ExecMode(ExecMode) {
  SavedExecMode = ExecMode;
  ExecMode = CGOpenMPRuntimeGPU::EM_NonSPMD;
}
/// Constructor for SPMD mode.
ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode,
                          bool &RuntimeMode, bool FullRuntimeMode)
    : ExecMode(ExecMode), RuntimeMode(&RuntimeMode) {
  SavedExecMode = ExecMode;
  SavedRuntimeMode = RuntimeMode;
  ExecMode = CGOpenMPRuntimeGPU::EM_SPMD;
  RuntimeMode = FullRuntimeMode;
}
~ExecutionRuntimeModesRAII() {
  ExecMode = SavedExecMode;
  if (RuntimeMode)
    *RuntimeMode = SavedRuntimeMode;
}
101};

103/// GPU Configuration:  This information can be derived from cuda registers,
104/// however, providing compile time constants helps generate more efficient
105/// code.  For all practical purposes this is fine because the configuration
106/// is the same for all known NVPTX architectures.
107enum MachineConfiguration : unsigned {
/// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target
/// specific Grid Values like GV_Warp_Size, GV_Warp_Size_Log2,
/// and GV_Warp_Size_Log2_Mask.

/// Global memory alignment for performance.
GlobalMemoryAlignment = 128,

/// Maximal size of the shared memory buffer.
SharedMemorySize = 128,
117};

119static const ValueDecl *getPrivateItem(const Expr *RefExpr) {
RefExpr = RefExpr->IgnoreParens();
if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) {
  const Expr *Base = ASE->getBase()->IgnoreParenImpCasts();
  while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
    Base = TempASE->getBase()->IgnoreParenImpCasts();
  RefExpr = Base;
} else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) {
  const Expr *Base = OASE->getBase()->IgnoreParenImpCasts();
  while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base))
    Base = TempOASE->getBase()->IgnoreParenImpCasts();
  while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
    Base = TempASE->getBase()->IgnoreParenImpCasts();
  RefExpr = Base;
}
RefExpr = RefExpr->IgnoreParenImpCasts();
if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr))
  return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl());
const auto *ME = cast<MemberExpr>(RefExpr);
return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl());
139}


142static RecordDecl *buildRecordForGlobalizedVars(
  ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls,
  ArrayRef<const ValueDecl *> EscapedDeclsForTeams,
  llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
      &MappedDeclsFields, int BufSize) {
using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>;
if (EscapedDecls.empty() && EscapedDeclsForTeams.empty())
  return nullptr;
SmallVector<VarsDataTy, 4> GlobalizedVars;
for (const ValueDecl *D : EscapedDecls)
  GlobalizedVars.emplace_back(
      CharUnits::fromQuantity(std::max(
          C.getDeclAlign(D).getQuantity(),
          static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))),
      D);
for (const ValueDecl *D : EscapedDeclsForTeams)
  GlobalizedVars.emplace_back(C.getDeclAlign(D), D);
llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) {
  return L.first > R.first;
});

// Build struct _globalized_locals_ty {
//         /*  globalized vars  */[WarSize] align (max(decl_align,
//         GlobalMemoryAlignment))
//         /*  globalized vars  */ for EscapedDeclsForTeams
//       };
RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty");
GlobalizedRD->startDefinition();
llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped(
    EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end());
for (const auto &Pair : GlobalizedVars) {
  const ValueDecl *VD = Pair.second;
  QualType Type = VD->getType();
  if (Type->isLValueReferenceType())
    Type = C.getPointerType(Type.getNonReferenceType());
  else
    Type = Type.getNonReferenceType();
  SourceLocation Loc = VD->getLocation();
  FieldDecl *Field;
  if (SingleEscaped.count(VD)) {
    Field = FieldDecl::Create(
        C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
        C.getTrivialTypeSourceInfo(Type, SourceLocation()),
        /*BW=*/nullptr, /*Mutable=*/false,
        /*InitStyle=*/ICIS_NoInit);
    Field->setAccess(AS_public);
    if (VD->hasAttrs()) {
      for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
           E(VD->getAttrs().end());
           I != E; ++I)
        Field->addAttr(*I);
    }
  } else {
    llvm::APInt ArraySize(32, BufSize);
    Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal,
                                  0);
    Field = FieldDecl::Create(
        C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
        C.getTrivialTypeSourceInfo(Type, SourceLocation()),
        /*BW=*/nullptr, /*Mutable=*/false,
        /*InitStyle=*/ICIS_NoInit);
    Field->setAccess(AS_public);
    llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(),
                                   static_cast<CharUnits::QuantityType>(
                                       GlobalMemoryAlignment)));
    Field->addAttr(AlignedAttr::CreateImplicit(
        C, /*IsAlignmentExpr=*/true,
        IntegerLiteral::Create(C, Align,
                               C.getIntTypeForBitwidth(32, /*Signed=*/0),
                               SourceLocation()),
        {}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned));
  }
  GlobalizedRD->addDecl(Field);
  MappedDeclsFields.try_emplace(VD, Field);
}
GlobalizedRD->completeDefinition();
return GlobalizedRD;
219}

221/// Get the list of variables that can escape their declaration context.
222class CheckVarsEscapingDeclContext final
  : public ConstStmtVisitor<CheckVarsEscapingDeclContext> {
CodeGenFunction &CGF;
llvm::SetVector<const ValueDecl *> EscapedDecls;
llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls;
llvm::SmallPtrSet<const Decl *, 4> EscapedParameters;
RecordDecl *GlobalizedRD = nullptr;
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
bool AllEscaped = false;
bool IsForCombinedParallelRegion = false;

void markAsEscaped(const ValueDecl *VD) {
  // Do not globalize declare target variables.
  if (!isa<VarDecl>(VD) ||
      OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD))
    return;
  VD = cast<ValueDecl>(VD->getCanonicalDecl());
  // Use user-specified allocation.
  if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>())
    return;
  // Variables captured by value must be globalized.
  if (auto *CSI = CGF.CapturedStmtInfo) {
    if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) {
      // Check if need to capture the variable that was already captured by
      // value in the outer region.
      if (!IsForCombinedParallelRegion) {
        if (!FD->hasAttrs())
          return;
        const auto *Attr = FD->getAttr<OMPCaptureKindAttr>();
        if (!Attr)
          return;
        if (((Attr->getCaptureKind() != OMPC_map) &&
             !isOpenMPPrivate(Attr->getCaptureKind())) ||
            ((Attr->getCaptureKind() == OMPC_map) &&
             !FD->getType()->isAnyPointerType()))
          return;
      }
      if (!FD->getType()->isReferenceType()) {
        assert(!VD->getType()->isVariablyModifiedType() &&((!VD->getType()->isVariablyModifiedType() && "Parameter captured by value with variably modified type"
) ? static_cast<void> (0) : __assert_fail ("!VD->getType()->isVariablyModifiedType() && \"Parameter captured by value with variably modified type\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 261, __PRETTY_FUNCTION__))
               "Parameter captured by value with variably modified type")((!VD->getType()->isVariablyModifiedType() && "Parameter captured by value with variably modified type"
) ? static_cast<void> (0) : __assert_fail ("!VD->getType()->isVariablyModifiedType() && \"Parameter captured by value with variably modified type\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 261, __PRETTY_FUNCTION__));
        EscapedParameters.insert(VD);
      } else if (!IsForCombinedParallelRegion) {
        return;
      }
    }
  }
  if ((!CGF.CapturedStmtInfo ||
       (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) &&
      VD->getType()->isReferenceType())
    // Do not globalize variables with reference type.
    return;
  if (VD->getType()->isVariablyModifiedType())
    EscapedVariableLengthDecls.insert(VD);
  else
    EscapedDecls.insert(VD);
}

void VisitValueDecl(const ValueDecl *VD) {
  if (VD->getType()->isLValueReferenceType())
    markAsEscaped(VD);
  if (const auto *VarD = dyn_cast<VarDecl>(VD)) {
    if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) {
      const bool SavedAllEscaped = AllEscaped;
      AllEscaped = VD->getType()->isLValueReferenceType();
      Visit(VarD->getInit());
      AllEscaped = SavedAllEscaped;
    }
  }
}
void VisitOpenMPCapturedStmt(const CapturedStmt *S,
                             ArrayRef<OMPClause *> Clauses,
                             bool IsCombinedParallelRegion) {
  if (!S)
    return;
  for (const CapturedStmt::Capture &C : S->captures()) {
    if (C.capturesVariable() && !C.capturesVariableByCopy()) {
      const ValueDecl *VD = C.getCapturedVar();
      bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion;
      if (IsCombinedParallelRegion) {
        // Check if the variable is privatized in the combined construct and
        // those private copies must be shared in the inner parallel
        // directive.
        IsForCombinedParallelRegion = false;
        for (const OMPClause *C : Clauses) {
          if (!isOpenMPPrivate(C->getClauseKind()) ||
              C->getClauseKind() == OMPC_reduction ||
              C->getClauseKind() == OMPC_linear ||
              C->getClauseKind() == OMPC_private)
            continue;
          ArrayRef<const Expr *> Vars;
          if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C))
            Vars = PC->getVarRefs();
          else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C))
            Vars = PC->getVarRefs();
          else
            llvm_unreachable("Unexpected clause.")::llvm::llvm_unreachable_internal("Unexpected clause.", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 317);
          for (const auto *E : Vars) {
            const Decl *D =
                cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl();
            if (D == VD->getCanonicalDecl()) {
              IsForCombinedParallelRegion = true;
              break;
            }
          }
          if (IsForCombinedParallelRegion)
            break;
        }
      }
      markAsEscaped(VD);
      if (isa<OMPCapturedExprDecl>(VD))
        VisitValueDecl(VD);
      IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion;
    }
  }
}

void buildRecordForGlobalizedVars(bool IsInTTDRegion) {
  assert(!GlobalizedRD &&((!GlobalizedRD && "Record for globalized variables is built already."
) ? static_cast<void> (0) : __assert_fail ("!GlobalizedRD && \"Record for globalized variables is built already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 340, __PRETTY_FUNCTION__))
         "Record for globalized variables is built already.")((!GlobalizedRD && "Record for globalized variables is built already."
) ? static_cast<void> (0) : __assert_fail ("!GlobalizedRD && \"Record for globalized variables is built already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 340, __PRETTY_FUNCTION__));
  ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams;
  unsigned WarpSize = CGF.getTarget().getGridValue(llvm::omp::GV_Warp_Size);
  if (IsInTTDRegion)
    EscapedDeclsForTeams = EscapedDecls.getArrayRef();
  else
    EscapedDeclsForParallel = EscapedDecls.getArrayRef();
  GlobalizedRD = ::buildRecordForGlobalizedVars(
      CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams,
      MappedDeclsFields, WarpSize);
}

352public:
CheckVarsEscapingDeclContext(CodeGenFunction &CGF,
                             ArrayRef<const ValueDecl *> TeamsReductions)
    : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) {
}
virtual ~CheckVarsEscapingDeclContext() = default;
void VisitDeclStmt(const DeclStmt *S) {
  if (!S)
    return;
  for (const Decl *D : S->decls())
    if (const auto *VD = dyn_cast_or_null<ValueDecl>(D))
      VisitValueDecl(VD);
}
void VisitOMPExecutableDirective(const OMPExecutableDirective *D) {
  if (!D)
    return;
  if (!D->hasAssociatedStmt())
    return;
  if (const auto *S =
          dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) {
    // Do not analyze directives that do not actually require capturing,
    // like `omp for` or `omp simd` directives.
    llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
    getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind());
    if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) {
      VisitStmt(S->getCapturedStmt());
      return;
    }
    VisitOpenMPCapturedStmt(
        S, D->clauses(),
        CaptureRegions.back() == OMPD_parallel &&
            isOpenMPDistributeDirective(D->getDirectiveKind()));
  }
}
void VisitCapturedStmt(const CapturedStmt *S) {
  if (!S)
    return;
  for (const CapturedStmt::Capture &C : S->captures()) {
    if (C.capturesVariable() && !C.capturesVariableByCopy()) {
      const ValueDecl *VD = C.getCapturedVar();
      markAsEscaped(VD);
      if (isa<OMPCapturedExprDecl>(VD))
        VisitValueDecl(VD);
    }
  }
}
void VisitLambdaExpr(const LambdaExpr *E) {
  if (!E)
    return;
  for (const LambdaCapture &C : E->captures()) {
    if (C.capturesVariable()) {
      if (C.getCaptureKind() == LCK_ByRef) {
        const ValueDecl *VD = C.getCapturedVar();
        markAsEscaped(VD);
        if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD))
          VisitValueDecl(VD);
      }
    }
  }
}
void VisitBlockExpr(const BlockExpr *E) {
  if (!E)
    return;
  for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) {
    if (C.isByRef()) {
      const VarDecl *VD = C.getVariable();
      markAsEscaped(VD);
      if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture())
        VisitValueDecl(VD);
    }
  }
}
void VisitCallExpr(const CallExpr *E) {
  if (!E)
    return;
  for (const Expr *Arg : E->arguments()) {
    if (!Arg)
      continue;
    if (Arg->isLValue()) {
      const bool SavedAllEscaped = AllEscaped;
      AllEscaped = true;
      Visit(Arg);
      AllEscaped = SavedAllEscaped;
    } else {
      Visit(Arg);
    }
  }
  Visit(E->getCallee());
}
void VisitDeclRefExpr(const DeclRefExpr *E) {
  if (!E)
    return;
  const ValueDecl *VD = E->getDecl();
  if (AllEscaped)
    markAsEscaped(VD);
  if (isa<OMPCapturedExprDecl>(VD))
    VisitValueDecl(VD);
  else if (const auto *VarD = dyn_cast<VarDecl>(VD))
    if (VarD->isInitCapture())
      VisitValueDecl(VD);
}
void VisitUnaryOperator(const UnaryOperator *E) {
  if (!E)
    return;
  if (E->getOpcode() == UO_AddrOf) {
    const bool SavedAllEscaped = AllEscaped;
    AllEscaped = true;
    Visit(E->getSubExpr());
    AllEscaped = SavedAllEscaped;
  } else {
    Visit(E->getSubExpr());
  }
}
void VisitImplicitCastExpr(const ImplicitCastExpr *E) {
  if (!E)
    return;
  if (E->getCastKind() == CK_ArrayToPointerDecay) {
    const bool SavedAllEscaped = AllEscaped;
    AllEscaped = true;
    Visit(E->getSubExpr());
    AllEscaped = SavedAllEscaped;
  } else {
    Visit(E->getSubExpr());
  }
}
void VisitExpr(const Expr *E) {
  if (!E)
    return;
  bool SavedAllEscaped = AllEscaped;
  if (!E->isLValue())
    AllEscaped = false;
  for (const Stmt *Child : E->children())
    if (Child)
      Visit(Child);
  AllEscaped = SavedAllEscaped;
}
void VisitStmt(const Stmt *S) {
  if (!S)
    return;
  for (const Stmt *Child : S->children())
    if (Child)
      Visit(Child);
}

/// Returns the record that handles all the escaped local variables and used
/// instead of their original storage.
const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) {
  if (!GlobalizedRD)
    buildRecordForGlobalizedVars(IsInTTDRegion);
  return GlobalizedRD;
}

/// Returns the field in the globalized record for the escaped variable.
const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const {
  assert(GlobalizedRD &&((GlobalizedRD && "Record for globalized variables must be generated already."
) ? static_cast<void> (0) : __assert_fail ("GlobalizedRD && \"Record for globalized variables must be generated already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 507, __PRETTY_FUNCTION__))
         "Record for globalized variables must be generated already.")((GlobalizedRD && "Record for globalized variables must be generated already."
) ? static_cast<void> (0) : __assert_fail ("GlobalizedRD && \"Record for globalized variables must be generated already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 507, __PRETTY_FUNCTION__));
  auto I = MappedDeclsFields.find(VD);
  if (I == MappedDeclsFields.end())
    return nullptr;
  return I->getSecond();
}

/// Returns the list of the escaped local variables/parameters.
ArrayRef<const ValueDecl *> getEscapedDecls() const {
  return EscapedDecls.getArrayRef();
}

/// Checks if the escaped local variable is actually a parameter passed by
/// value.
const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const {
  return EscapedParameters;
}

/// Returns the list of the escaped variables with the variably modified
/// types.
ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const {
  return EscapedVariableLengthDecls.getArrayRef();
}
530};
531} // anonymous namespace

533/// Get the id of the warp in the block.
534/// We assume that the warp size is 32, which is always the case
535/// on the NVPTX device, to generate more efficient code.
536static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
unsigned LaneIDBits =
    CGF.getTarget().getGridValue(llvm::omp::GV_Warp_Size_Log2);
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id");
542}

544/// Get the id of the current lane in the Warp.
545/// We assume that the warp size is 32, which is always the case
546/// on the NVPTX device, to generate more efficient code.
547static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
unsigned LaneIDMask = CGF.getContext().getTargetInfo().getGridValue(
    llvm::omp::GV_Warp_Size_Log2_Mask);
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask),
                     "nvptx_lane_id");
554}

556/// Get the value of the thread_limit clause in the teams directive.
557/// For the 'generic' execution mode, the runtime encodes thread_limit in
558/// the launch parameters, always starting thread_limit+warpSize threads per
559/// CTA. The threads in the last warp are reserved for master execution.
560/// For the 'spmd' execution mode, all threads in a CTA are part of the team.
561static llvm::Value *getThreadLimit(CodeGenFunction &CGF,
                                 bool IsInSPMDExecutionMode = false) {
CGBuilderTy &Bld = CGF.Builder;
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
return IsInSPMDExecutionMode
           ? RT.getGPUNumThreads(CGF)
           : Bld.CreateNUWSub(RT.getGPUNumThreads(CGF),
                              RT.getGPUWarpSize(CGF), "thread_limit");
569}

571/// Get the thread id of the OMP master thread.
572/// The master thread id is the first thread (lane) of the last warp in the
573/// GPU block.  Warp size is assumed to be some power of 2.
574/// Thread id is 0 indexed.
575/// E.g: If NumThreads is 33, master id is 32.
576///      If NumThreads is 64, master id is 32.
577///      If NumThreads is 1024, master id is 992.
578static llvm::Value *getMasterThreadID(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
llvm::Value *NumThreads = RT.getGPUNumThreads(CGF);
// We assume that the warp size is a power of 2.
llvm::Value *Mask = Bld.CreateNUWSub(RT.getGPUWarpSize(CGF), Bld.getInt32(1));

return Bld.CreateAnd(Bld.CreateNUWSub(NumThreads, Bld.getInt32(1)),
                     Bld.CreateNot(Mask), "master_tid");
587}

589CGOpenMPRuntimeGPU::WorkerFunctionState::WorkerFunctionState(
  CodeGenModule &CGM, SourceLocation Loc)
  : WorkerFn(nullptr), CGFI(CGM.getTypes().arrangeNullaryFunction()),
    Loc(Loc) {
createWorkerFunction(CGM);
594}

596void CGOpenMPRuntimeGPU::WorkerFunctionState::createWorkerFunction(
  CodeGenModule &CGM) {
// Create an worker function with no arguments.

WorkerFn = llvm::Function::Create(
    CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
    /*placeholder=*/"_worker", &CGM.getModule());
CGM.SetInternalFunctionAttributes(GlobalDecl(), WorkerFn, CGFI);
WorkerFn->setDoesNotRecurse();
605}

607CGOpenMPRuntimeGPU::ExecutionMode
608CGOpenMPRuntimeGPU::getExecutionMode() const {
return CurrentExecutionMode;
610}

612static CGOpenMPRuntimeGPU::DataSharingMode
613getDataSharingMode(CodeGenModule &CGM) {
return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA
                                        : CGOpenMPRuntimeGPU::Generic;
616}

618/// Check for inner (nested) SPMD construct, if any
619static bool hasNestedSPMDDirective(ASTContext &Ctx,
                                 const OMPExecutableDirective &D) {
const auto *CS = D.getInnermostCapturedStmt();
const auto *Body =
    CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);

if (const auto *NestedDir =
        dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
  OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
  switch (D.getDirectiveKind()) {
  case OMPD_target:
    if (isOpenMPParallelDirective(DKind))
      return true;
    if (DKind == OMPD_teams) {
      Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
          /*IgnoreCaptured=*/true);
      if (!Body)
        return false;
      ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
      if (const auto *NND =
              dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
        DKind = NND->getDirectiveKind();
        if (isOpenMPParallelDirective(DKind))
          return true;
      }
    }
    return false;
  case OMPD_target_teams:
    return isOpenMPParallelDirective(DKind);
  case OMPD_target_simd:
  case OMPD_target_parallel:
  case OMPD_target_parallel_for:
  case OMPD_target_parallel_for_simd:
  case OMPD_target_teams_distribute:
  case OMPD_target_teams_distribute_simd:
  case OMPD_target_teams_distribute_parallel_for:
  case OMPD_target_teams_distribute_parallel_for_simd:
  case OMPD_parallel:
  case OMPD_for:
  case OMPD_parallel_for:
  case OMPD_parallel_master:
  case OMPD_parallel_sections:
  case OMPD_for_simd:
  case OMPD_parallel_for_simd:
  case OMPD_cancel:
  case OMPD_cancellation_point:
  case OMPD_ordered:
  case OMPD_threadprivate:
  case OMPD_allocate:
  case OMPD_task:
  case OMPD_simd:
  case OMPD_sections:
  case OMPD_section:
  case OMPD_single:
  case OMPD_master:
  case OMPD_critical:
  case OMPD_taskyield:
  case OMPD_barrier:
  case OMPD_taskwait:
  case OMPD_taskgroup:
  case OMPD_atomic:
  case OMPD_flush:
  case OMPD_depobj:
  case OMPD_scan:
  case OMPD_teams:
  case OMPD_target_data:
  case OMPD_target_exit_data:
  case OMPD_target_enter_data:
  case OMPD_distribute:
  case OMPD_distribute_simd:
  case OMPD_distribute_parallel_for:
  case OMPD_distribute_parallel_for_simd:
  case OMPD_teams_distribute:
  case OMPD_teams_distribute_simd:
  case OMPD_teams_distribute_parallel_for:
  case OMPD_teams_distribute_parallel_for_simd:
  case OMPD_target_update:
  case OMPD_declare_simd:
  case OMPD_declare_variant:
  case OMPD_begin_declare_variant:
  case OMPD_end_declare_variant:
  case OMPD_declare_target:
  case OMPD_end_declare_target:
  case OMPD_declare_reduction:
  case OMPD_declare_mapper:
  case OMPD_taskloop:
  case OMPD_taskloop_simd:
  case OMPD_master_taskloop:
  case OMPD_master_taskloop_simd:
  case OMPD_parallel_master_taskloop:
  case OMPD_parallel_master_taskloop_simd:
  case OMPD_requires:
  case OMPD_unknown:
  default:
    llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 714);
  }
}

return false;
719}

721static bool supportsSPMDExecutionMode(ASTContext &Ctx,
                                    const OMPExecutableDirective &D) {
OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
switch (DirectiveKind) {
case OMPD_target:
case OMPD_target_teams:
  return hasNestedSPMDDirective(Ctx, D);
case OMPD_target_parallel:
case OMPD_target_parallel_for:
case OMPD_target_parallel_for_simd:
case OMPD_target_teams_distribute_parallel_for:
case OMPD_target_teams_distribute_parallel_for_simd:
case OMPD_target_simd:
case OMPD_target_teams_distribute_simd:
  return true;
case OMPD_target_teams_distribute:
  return false;
case OMPD_parallel:
case OMPD_for:
case OMPD_parallel_for:
case OMPD_parallel_master:
case OMPD_parallel_sections:
case OMPD_for_simd:
case OMPD_parallel_for_simd:
case OMPD_cancel:
case OMPD_cancellation_point:
case OMPD_ordered:
case OMPD_threadprivate:
case OMPD_allocate:
case OMPD_task:
case OMPD_simd:
case OMPD_sections:
case OMPD_section:
case OMPD_single:
case OMPD_master:
case OMPD_critical:
case OMPD_taskyield:
case OMPD_barrier:
case OMPD_taskwait:
case OMPD_taskgroup:
case OMPD_atomic:
case OMPD_flush:
case OMPD_depobj:
case OMPD_scan:
case OMPD_teams:
case OMPD_target_data:
case OMPD_target_exit_data:
case OMPD_target_enter_data:
case OMPD_distribute:
case OMPD_distribute_simd:
case OMPD_distribute_parallel_for:
case OMPD_distribute_parallel_for_simd:
case OMPD_teams_distribute:
case OMPD_teams_distribute_simd:
case OMPD_teams_distribute_parallel_for:
case OMPD_teams_distribute_parallel_for_simd:
case OMPD_target_update:
case OMPD_declare_simd:
case OMPD_declare_variant:
case OMPD_begin_declare_variant:
case OMPD_end_declare_variant:
case OMPD_declare_target:
case OMPD_end_declare_target:
case OMPD_declare_reduction:
case OMPD_declare_mapper:
case OMPD_taskloop:
case OMPD_taskloop_simd:
case OMPD_master_taskloop:
case OMPD_master_taskloop_simd:
case OMPD_parallel_master_taskloop:
case OMPD_parallel_master_taskloop_simd:
case OMPD_requires:
case OMPD_unknown:
default:
  break;
}
llvm_unreachable(::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target."
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 798)
    "Unknown programming model for OpenMP directive on NVPTX target.")::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target."
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 798);
799}

801/// Check if the directive is loops based and has schedule clause at all or has
802/// static scheduling.
803static bool hasStaticScheduling(const OMPExecutableDirective &D) {
assert(isOpenMPWorksharingDirective(D.getDirectiveKind()) &&((isOpenMPWorksharingDirective(D.getDirectiveKind()) &&
 isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive."
) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 806, __PRETTY_FUNCTION__))
       isOpenMPLoopDirective(D.getDirectiveKind()) &&((isOpenMPWorksharingDirective(D.getDirectiveKind()) &&
 isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive."
) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 806, __PRETTY_FUNCTION__))
       "Expected loop-based directive.")((isOpenMPWorksharingDirective(D.getDirectiveKind()) &&
 isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive."
) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 806, __PRETTY_FUNCTION__));
return !D.hasClausesOfKind<OMPOrderedClause>() &&
       (!D.hasClausesOfKind<OMPScheduleClause>() ||
        llvm::any_of(D.getClausesOfKind<OMPScheduleClause>(),
                     [](const OMPScheduleClause *C) {
                       return C->getScheduleKind() == OMPC_SCHEDULE_static;
                     }));
813}

815/// Check for inner (nested) lightweight runtime construct, if any
816static bool hasNestedLightweightDirective(ASTContext &Ctx,
                                        const OMPExecutableDirective &D) {
assert(supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive.")((supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive."
) ? static_cast<void> (0) : __assert_fail ("supportsSPMDExecutionMode(Ctx, D) && \"Expected SPMD mode directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 818, __PRETTY_FUNCTION__));
const auto *CS = D.getInnermostCapturedStmt();
const auto *Body =
    CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);

if (const auto *NestedDir =
        dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
  OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
  switch (D.getDirectiveKind()) {
  case OMPD_target:
    if (isOpenMPParallelDirective(DKind) &&
        isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) &&
        hasStaticScheduling(*NestedDir))
      return true;
    if (DKind == OMPD_teams_distribute_simd || DKind == OMPD_simd)
      return true;
    if (DKind == OMPD_parallel) {
      Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
          /*IgnoreCaptured=*/true);
      if (!Body)
        return false;
      ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
      if (const auto *NND =
              dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
        DKind = NND->getDirectiveKind();
        if (isOpenMPWorksharingDirective(DKind) &&
            isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
          return true;
      }
    } else if (DKind == OMPD_teams) {
      Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
          /*IgnoreCaptured=*/true);
      if (!Body)
        return false;
      ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
      if (const auto *NND =
              dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
        DKind = NND->getDirectiveKind();
        if (isOpenMPParallelDirective(DKind) &&
            isOpenMPWorksharingDirective(DKind) &&
            isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
          return true;
        if (DKind == OMPD_parallel) {
          Body = NND->getInnermostCapturedStmt()->IgnoreContainers(
              /*IgnoreCaptured=*/true);
          if (!Body)
            return false;
          ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
          if (const auto *NND =
                  dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
            DKind = NND->getDirectiveKind();
            if (isOpenMPWorksharingDirective(DKind) &&
                isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
              return true;
          }
        }
      }
    }
    return false;
  case OMPD_target_teams:
    if (isOpenMPParallelDirective(DKind) &&
        isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) &&
        hasStaticScheduling(*NestedDir))
      return true;
    if (DKind == OMPD_distribute_simd || DKind == OMPD_simd)
      return true;
    if (DKind == OMPD_parallel) {
      Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
          /*IgnoreCaptured=*/true);
      if (!Body)
        return false;
      ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
      if (const auto *NND =
              dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
        DKind = NND->getDirectiveKind();
        if (isOpenMPWorksharingDirective(DKind) &&
            isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
          return true;
      }
    }
    return false;
  case OMPD_target_parallel:
    if (DKind == OMPD_simd)
      return true;
    return isOpenMPWorksharingDirective(DKind) &&
           isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NestedDir);
  case OMPD_target_teams_distribute:
  case OMPD_target_simd:
  case OMPD_target_parallel_for:
  case OMPD_target_parallel_for_simd:
  case OMPD_target_teams_distribute_simd:
  case OMPD_target_teams_distribute_parallel_for:
  case OMPD_target_teams_distribute_parallel_for_simd:
  case OMPD_parallel:
  case OMPD_for:
  case OMPD_parallel_for:
  case OMPD_parallel_master:
  case OMPD_parallel_sections:
  case OMPD_for_simd:
  case OMPD_parallel_for_simd:
  case OMPD_cancel:
  case OMPD_cancellation_point:
  case OMPD_ordered:
  case OMPD_threadprivate:
  case OMPD_allocate:
  case OMPD_task:
  case OMPD_simd:
  case OMPD_sections:
  case OMPD_section:
  case OMPD_single:
  case OMPD_master:
  case OMPD_critical:
  case OMPD_taskyield:
  case OMPD_barrier:
  case OMPD_taskwait:
  case OMPD_taskgroup:
  case OMPD_atomic:
  case OMPD_flush:
  case OMPD_depobj:
  case OMPD_scan:
  case OMPD_teams:
  case OMPD_target_data:
  case OMPD_target_exit_data:
  case OMPD_target_enter_data:
  case OMPD_distribute:
  case OMPD_distribute_simd:
  case OMPD_distribute_parallel_for:
  case OMPD_distribute_parallel_for_simd:
  case OMPD_teams_distribute:
  case OMPD_teams_distribute_simd:
  case OMPD_teams_distribute_parallel_for:
  case OMPD_teams_distribute_parallel_for_simd:
  case OMPD_target_update:
  case OMPD_declare_simd:
  case OMPD_declare_variant:
  case OMPD_begin_declare_variant:
  case OMPD_end_declare_variant:
  case OMPD_declare_target:
  case OMPD_end_declare_target:
  case OMPD_declare_reduction:
  case OMPD_declare_mapper:
  case OMPD_taskloop:
  case OMPD_taskloop_simd:
  case OMPD_master_taskloop:
  case OMPD_master_taskloop_simd:
  case OMPD_parallel_master_taskloop:
  case OMPD_parallel_master_taskloop_simd:
  case OMPD_requires:
  case OMPD_unknown:
  default:
    llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 969);
  }
}

return false;
974}

976/// Checks if the construct supports lightweight runtime. It must be SPMD
977/// construct + inner loop-based construct with static scheduling.
978static bool supportsLightweightRuntime(ASTContext &Ctx,
                                     const OMPExecutableDirective &D) {
if (!supportsSPMDExecutionMode(Ctx, D))
  return false;
OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
switch (DirectiveKind) {
case OMPD_target:
case OMPD_target_teams:
case OMPD_target_parallel:
  return hasNestedLightweightDirective(Ctx, D);
case OMPD_target_parallel_for:
case OMPD_target_parallel_for_simd:
case OMPD_target_teams_distribute_parallel_for:
case OMPD_target_teams_distribute_parallel_for_simd:
  // (Last|First)-privates must be shared in parallel region.
  return hasStaticScheduling(D);
case OMPD_target_simd:
case OMPD_target_teams_distribute_simd:
  return true;
case OMPD_target_teams_distribute:
  return false;
case OMPD_parallel:
case OMPD_for:
case OMPD_parallel_for:
case OMPD_parallel_master:
case OMPD_parallel_sections:
case OMPD_for_simd:
case OMPD_parallel_for_simd:
case OMPD_cancel:
case OMPD_cancellation_point:
case OMPD_ordered:
case OMPD_threadprivate:
case OMPD_allocate:
case OMPD_task:
case OMPD_simd:
case OMPD_sections:
case OMPD_section:
case OMPD_single:
case OMPD_master:
case OMPD_critical:
case OMPD_taskyield:
case OMPD_barrier:
case OMPD_taskwait:
case OMPD_taskgroup:
case OMPD_atomic:
case OMPD_flush:
case OMPD_depobj:
case OMPD_scan:
case OMPD_teams:
case OMPD_target_data:
case OMPD_target_exit_data:
case OMPD_target_enter_data:
case OMPD_distribute:
case OMPD_distribute_simd:
case OMPD_distribute_parallel_for:
case OMPD_distribute_parallel_for_simd:
case OMPD_teams_distribute:
case OMPD_teams_distribute_simd:
case OMPD_teams_distribute_parallel_for:
case OMPD_teams_distribute_parallel_for_simd:
case OMPD_target_update:
case OMPD_declare_simd:
case OMPD_declare_variant:
case OMPD_begin_declare_variant:
case OMPD_end_declare_variant:
case OMPD_declare_target:
case OMPD_end_declare_target:
case OMPD_declare_reduction:
case OMPD_declare_mapper:
case OMPD_taskloop:
case OMPD_taskloop_simd:
case OMPD_master_taskloop:
case OMPD_master_taskloop_simd:
case OMPD_parallel_master_taskloop:
case OMPD_parallel_master_taskloop_simd:
case OMPD_requires:
case OMPD_unknown:
default:
  break;
}
llvm_unreachable(::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target."
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1059)
    "Unknown programming model for OpenMP directive on NVPTX target.")::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target."
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1059);
1060}

1062void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D,
                                           StringRef ParentName,
                                           llvm::Function *&OutlinedFn,
                                           llvm::Constant *&OutlinedFnID,
                                           bool IsOffloadEntry,
                                           const RegionCodeGenTy &CodeGen) {
ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode);
EntryFunctionState EST;
WorkerFunctionState WST(CGM, D.getBeginLoc());
Work.clear();
WrapperFunctionsMap.clear();

// Emit target region as a standalone region.
class NVPTXPrePostActionTy : public PrePostActionTy {
  CGOpenMPRuntimeGPU::EntryFunctionState &EST;
  CGOpenMPRuntimeGPU::WorkerFunctionState &WST;

public:
  NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST,
                       CGOpenMPRuntimeGPU::WorkerFunctionState &WST)
      : EST(EST), WST(WST) {}
  void Enter(CodeGenFunction &CGF) override {
    auto &RT =
        static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
    RT.emitNonSPMDEntryHeader(CGF, EST, WST);
    // Skip target region initialization.
    RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
  }
  void Exit(CodeGenFunction &CGF) override {
    auto &RT =
        static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
    RT.clearLocThreadIdInsertPt(CGF);
    RT.emitNonSPMDEntryFooter(CGF, EST);
  }
} Action(EST, WST);
CodeGen.setAction(Action);
IsInTTDRegion = true;
// Reserve place for the globalized memory.
GlobalizedRecords.emplace_back();
if (!KernelStaticGlobalized) {
  KernelStaticGlobalized = new llvm::GlobalVariable(
      CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/false,
      llvm::GlobalValue::InternalLinkage,
      llvm::UndefValue::get(CGM.VoidPtrTy),
      "_openmp_kernel_static_glob_rd$ptr", /*InsertBefore=*/nullptr,
      llvm::GlobalValue::NotThreadLocal,
      CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
}
emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
                                 IsOffloadEntry, CodeGen);
IsInTTDRegion = false;

// Now change the name of the worker function to correspond to this target
// region's entry function.
WST.WorkerFn->setName(Twine(OutlinedFn->getName(), "_worker"));

// Create the worker function
emitWorkerFunction(WST);
1120}

1122// Setup NVPTX threads for master-worker OpenMP scheme.
1123void CGOpenMPRuntimeGPU::emitNonSPMDEntryHeader(CodeGenFunction &CGF,
                                                EntryFunctionState &EST,
                                                WorkerFunctionState &WST) {
CGBuilderTy &Bld = CGF.Builder;

llvm::BasicBlock *WorkerBB = CGF.createBasicBlock(".worker");
llvm::BasicBlock *MasterCheckBB = CGF.createBasicBlock(".mastercheck");
llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master");
EST.ExitBB = CGF.createBasicBlock(".exit");

auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
llvm::Value *IsWorker =
    Bld.CreateICmpULT(RT.getGPUThreadID(CGF), getThreadLimit(CGF));
Bld.CreateCondBr(IsWorker, WorkerBB, MasterCheckBB);

CGF.EmitBlock(WorkerBB);
emitCall(CGF, WST.Loc, WST.WorkerFn);
CGF.EmitBranch(EST.ExitBB);

CGF.EmitBlock(MasterCheckBB);
llvm::Value *IsMaster =
    Bld.CreateICmpEQ(RT.getGPUThreadID(CGF), getMasterThreadID(CGF));
Bld.CreateCondBr(IsMaster, MasterBB, EST.ExitBB);

CGF.EmitBlock(MasterBB);
IsInTargetMasterThreadRegion = true;
// SEQUENTIAL (MASTER) REGION START
// First action in sequential region:
// Initialize the state of the OpenMP runtime library on the GPU.
// TODO: Optimize runtime initialization and pass in correct value.
llvm::Value *Args[] = {getThreadLimit(CGF),
                       Bld.getInt16(/*RequiresOMPRuntime=*/1)};
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                        CGM.getModule(), OMPRTL___kmpc_kernel_init),
                    Args);

// For data sharing, we need to initialize the stack.
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
    CGM.getModule(), OMPRTL___kmpc_data_sharing_init_stack));

emitGenericVarsProlog(CGF, WST.Loc);
1164}

1166void CGOpenMPRuntimeGPU::emitNonSPMDEntryFooter(CodeGenFunction &CGF,
                                                EntryFunctionState &EST) {
IsInTargetMasterThreadRegion = false;
if (!CGF.HaveInsertPoint())
  return;

emitGenericVarsEpilog(CGF);

if (!EST.ExitBB)
  EST.ExitBB = CGF.createBasicBlock(".exit");

llvm::BasicBlock *TerminateBB = CGF.createBasicBlock(".termination.notifier");
CGF.EmitBranch(TerminateBB);

CGF.EmitBlock(TerminateBB);
// Signal termination condition.
// TODO: Optimize runtime initialization and pass in correct value.
llvm::Value *Args[] = {CGF.Builder.getInt16(/*IsOMPRuntimeInitialized=*/1)};
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                        CGM.getModule(), OMPRTL___kmpc_kernel_deinit),
                    Args);
// Barrier to terminate worker threads.
syncCTAThreads(CGF);
// Master thread jumps to exit point.
CGF.EmitBranch(EST.ExitBB);

CGF.EmitBlock(EST.ExitBB);
EST.ExitBB = nullptr;
1194}

1196void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D,
                                        StringRef ParentName,
                                        llvm::Function *&OutlinedFn,
                                        llvm::Constant *&OutlinedFnID,
                                        bool IsOffloadEntry,
                                        const RegionCodeGenTy &CodeGen) {
ExecutionRuntimeModesRAII ModeRAII(
    CurrentExecutionMode, RequiresFullRuntime,
    CGM.getLangOpts().OpenMPCUDAForceFullRuntime ||
        !supportsLightweightRuntime(CGM.getContext(), D));
EntryFunctionState EST;

// Emit target region as a standalone region.
class NVPTXPrePostActionTy : public PrePostActionTy {
  CGOpenMPRuntimeGPU &RT;
  CGOpenMPRuntimeGPU::EntryFunctionState &EST;
  const OMPExecutableDirective &D;

public:
  NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT,
                       CGOpenMPRuntimeGPU::EntryFunctionState &EST,
                       const OMPExecutableDirective &D)
      : RT(RT), EST(EST), D(D) {}
  void Enter(CodeGenFunction &CGF) override {
    RT.emitSPMDEntryHeader(CGF, EST, D);
    // Skip target region initialization.
    RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
  }
  void Exit(CodeGenFunction &CGF) override {
    RT.clearLocThreadIdInsertPt(CGF);
    RT.emitSPMDEntryFooter(CGF, EST);
  }
} Action(*this, EST, D);
CodeGen.setAction(Action);
IsInTTDRegion = true;
// Reserve place for the globalized memory.
GlobalizedRecords.emplace_back();
if (!KernelStaticGlobalized) {
  KernelStaticGlobalized = new llvm::GlobalVariable(
      CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/false,
      llvm::GlobalValue::InternalLinkage,
      llvm::UndefValue::get(CGM.VoidPtrTy),
      "_openmp_kernel_static_glob_rd$ptr", /*InsertBefore=*/nullptr,
      llvm::GlobalValue::NotThreadLocal,
      CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
}
emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
                                 IsOffloadEntry, CodeGen);
IsInTTDRegion = false;
1245}

1247void CGOpenMPRuntimeGPU::emitSPMDEntryHeader(
  CodeGenFunction &CGF, EntryFunctionState &EST,
  const OMPExecutableDirective &D) {
CGBuilderTy &Bld = CGF.Builder;

// Setup BBs in entry function.
llvm::BasicBlock *ExecuteBB = CGF.createBasicBlock(".execute");
EST.ExitBB = CGF.createBasicBlock(".exit");

llvm::Value *Args[] = {getThreadLimit(CGF, /*IsInSPMDExecutionMode=*/true),
                       /*RequiresOMPRuntime=*/
                       Bld.getInt16(RequiresFullRuntime ? 1 : 0)};
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                        CGM.getModule(), OMPRTL___kmpc_spmd_kernel_init),
                    Args);

if (RequiresFullRuntime) {
  // For data sharing, we need to initialize the stack.
  CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
      CGM.getModule(), OMPRTL___kmpc_data_sharing_init_stack_spmd));
}

CGF.EmitBranch(ExecuteBB);

CGF.EmitBlock(ExecuteBB);

IsInTargetMasterThreadRegion = true;
1274}

1276void CGOpenMPRuntimeGPU::emitSPMDEntryFooter(CodeGenFunction &CGF,
                                             EntryFunctionState &EST) {
IsInTargetMasterThreadRegion = false;
if (!CGF.HaveInsertPoint())
  return;

if (!EST.ExitBB)
  EST.ExitBB = CGF.createBasicBlock(".exit");

llvm::BasicBlock *OMPDeInitBB = CGF.createBasicBlock(".omp.deinit");
CGF.EmitBranch(OMPDeInitBB);

CGF.EmitBlock(OMPDeInitBB);
// DeInitialize the OMP state in the runtime; called by all active threads.
llvm::Value *Args[] = {/*RequiresOMPRuntime=*/
                       CGF.Builder.getInt16(RequiresFullRuntime ? 1 : 0)};
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                        CGM.getModule(), OMPRTL___kmpc_spmd_kernel_deinit_v2),
                    Args);
CGF.EmitBranch(EST.ExitBB);

CGF.EmitBlock(EST.ExitBB);
EST.ExitBB = nullptr;
1299}

1301// Create a unique global variable to indicate the execution mode of this target
1302// region. The execution mode is either 'generic', or 'spmd' depending on the
1303// target directive. This variable is picked up by the offload library to setup
1304// the device appropriately before kernel launch. If the execution mode is
1305// 'generic', the runtime reserves one warp for the master, otherwise, all
1306// warps participate in parallel work.
1307static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name,
                                   bool Mode) {
auto *GVMode =
    new llvm::GlobalVariable(CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
                             llvm::GlobalValue::WeakAnyLinkage,
                             llvm::ConstantInt::get(CGM.Int8Ty, Mode ? 0 : 1),
                             Twine(Name, "_exec_mode"));
CGM.addCompilerUsedGlobal(GVMode);
1315}

1317void CGOpenMPRuntimeGPU::emitWorkerFunction(WorkerFunctionState &WST) {
ASTContext &Ctx = CGM.getContext();

CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, WST.WorkerFn, WST.CGFI, {},
                  WST.Loc, WST.Loc);
emitWorkerLoop(CGF, WST);
CGF.FinishFunction();
1325}

1327void CGOpenMPRuntimeGPU::emitWorkerLoop(CodeGenFunction &CGF,
                                      WorkerFunctionState &WST) {
//
// The workers enter this loop and wait for parallel work from the master.
// When the master encounters a parallel region it sets up the work + variable
// arguments, and wakes up the workers.  The workers first check to see if
// they are required for the parallel region, i.e., within the # of requested
// parallel threads.  The activated workers load the variable arguments and
// execute the parallel work.
//

CGBuilderTy &Bld = CGF.Builder;

llvm::BasicBlock *AwaitBB = CGF.createBasicBlock(".await.work");
llvm::BasicBlock *SelectWorkersBB = CGF.createBasicBlock(".select.workers");
llvm::BasicBlock *ExecuteBB = CGF.createBasicBlock(".execute.parallel");
llvm::BasicBlock *TerminateBB = CGF.createBasicBlock(".terminate.parallel");
llvm::BasicBlock *BarrierBB = CGF.createBasicBlock(".barrier.parallel");
llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");

CGF.EmitBranch(AwaitBB);

// Workers wait for work from master.
CGF.EmitBlock(AwaitBB);
// Wait for parallel work
syncCTAThreads(CGF);

Address WorkFn =
    CGF.CreateDefaultAlignTempAlloca(CGF.Int8PtrTy, /*Name=*/"work_fn");
Address ExecStatus =
    CGF.CreateDefaultAlignTempAlloca(CGF.Int8Ty, /*Name=*/"exec_status");
CGF.InitTempAlloca(ExecStatus, Bld.getInt8(/*C=*/0));
CGF.InitTempAlloca(WorkFn, llvm::Constant::getNullValue(CGF.Int8PtrTy));

// TODO: Optimize runtime initialization and pass in correct value.
llvm::Value *Args[] = {WorkFn.getPointer()};
llvm::Value *Ret =
    CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                            CGM.getModule(), OMPRTL___kmpc_kernel_parallel),
                        Args);
Bld.CreateStore(Bld.CreateZExt(Ret, CGF.Int8Ty), ExecStatus);

// On termination condition (workid == 0), exit loop.
llvm::Value *WorkID = Bld.CreateLoad(WorkFn);
llvm::Value *ShouldTerminate = Bld.CreateIsNull(WorkID, "should_terminate");
Bld.CreateCondBr(ShouldTerminate, ExitBB, SelectWorkersBB);

// Activate requested workers.
CGF.EmitBlock(SelectWorkersBB);
llvm::Value *IsActive =
    Bld.CreateIsNotNull(Bld.CreateLoad(ExecStatus), "is_active");
Bld.CreateCondBr(IsActive, ExecuteBB, BarrierBB);

// Signal start of parallel region.
CGF.EmitBlock(ExecuteBB);
// Skip initialization.
setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);

// Process work items: outlined parallel functions.
for (llvm::Function *W : Work) {
  // Try to match this outlined function.
  llvm::Value *ID = Bld.CreatePointerBitCastOrAddrSpaceCast(W, CGM.Int8PtrTy);

  llvm::Value *WorkFnMatch =
      Bld.CreateICmpEQ(Bld.CreateLoad(WorkFn), ID, "work_match");

  llvm::BasicBlock *ExecuteFNBB = CGF.createBasicBlock(".execute.fn");
  llvm::BasicBlock *CheckNextBB = CGF.createBasicBlock(".check.next");
  Bld.CreateCondBr(WorkFnMatch, ExecuteFNBB, CheckNextBB);

  // Execute this outlined function.
  CGF.EmitBlock(ExecuteFNBB);

  // Insert call to work function via shared wrapper. The shared
  // wrapper takes two arguments:
  //   - the parallelism level;
  //   - the thread ID;
  emitCall(CGF, WST.Loc, W,
           {Bld.getInt16(/*ParallelLevel=*/0), getThreadID(CGF, WST.Loc)});

  // Go to end of parallel region.
  CGF.EmitBranch(TerminateBB);

  CGF.EmitBlock(CheckNextBB);
}
// Default case: call to outlined function through pointer if the target
// region makes a declare target call that may contain an orphaned parallel
// directive.
auto *ParallelFnTy =
    llvm::FunctionType::get(CGM.VoidTy, {CGM.Int16Ty, CGM.Int32Ty},
                            /*isVarArg=*/false);
llvm::Value *WorkFnCast =
    Bld.CreateBitCast(WorkID, ParallelFnTy->getPointerTo());
// Insert call to work function via shared wrapper. The shared
// wrapper takes two arguments:
//   - the parallelism level;
//   - the thread ID;
emitCall(CGF, WST.Loc, {ParallelFnTy, WorkFnCast},
         {Bld.getInt16(/*ParallelLevel=*/0), getThreadID(CGF, WST.Loc)});
// Go to end of parallel region.
CGF.EmitBranch(TerminateBB);

// Signal end of parallel region.
CGF.EmitBlock(TerminateBB);
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                        CGM.getModule(), OMPRTL___kmpc_kernel_end_parallel),
                    llvm::None);
CGF.EmitBranch(BarrierBB);

// All active and inactive workers wait at a barrier after parallel region.
CGF.EmitBlock(BarrierBB);
// Barrier after parallel region.
syncCTAThreads(CGF);
CGF.EmitBranch(AwaitBB);

// Exit target region.
CGF.EmitBlock(ExitBB);
// Skip initialization.
clearLocThreadIdInsertPt(CGF);
1446}

1448void CGOpenMPRuntimeGPU::createOffloadEntry(llvm::Constant *ID,
                                            llvm::Constant *Addr,
                                            uint64_t Size, int32_t,
                                            llvm::GlobalValue::LinkageTypes) {
// TODO: Add support for global variables on the device after declare target
// support.
if (!isa<llvm::Function>(Addr))
  return;
llvm::Module &M = CGM.getModule();
llvm::LLVMContext &Ctx = CGM.getLLVMContext();

// Get "nvvm.annotations" metadata node
llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations");

llvm::Metadata *MDVals[] = {
    llvm::ConstantAsMetadata::get(Addr), llvm::MDString::get(Ctx, "kernel"),
    llvm::ConstantAsMetadata::get(
        llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
// Append metadata to nvvm.annotations
MD->addOperand(llvm::MDNode::get(Ctx, MDVals));
1468}

1470void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction(
  const OMPExecutableDirective &D, StringRef ParentName,
  llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
  bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
if (!IsOffloadEntry) // Nothing to do.
  return;

assert(!ParentName.empty() && "Invalid target region parent name!")((!ParentName.empty() && "Invalid target region parent name!"
) ? static_cast<void> (0) : __assert_fail ("!ParentName.empty() && \"Invalid target region parent name!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1477, __PRETTY_FUNCTION__));

bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D);
if (Mode)
  emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
                 CodeGen);
else
  emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
                    CodeGen);

setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode);
1488}

1490namespace {
1491LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE()using ::llvm::BitmaskEnumDetail::operator~; using ::llvm::BitmaskEnumDetail
::operator|; using ::llvm::BitmaskEnumDetail::operator&; using
 ::llvm::BitmaskEnumDetail::operator^; using ::llvm::BitmaskEnumDetail
::operator|=; using ::llvm::BitmaskEnumDetail::operator&=
; using ::llvm::BitmaskEnumDetail::operator^=;
1492/// Enum for accesseing the reserved_2 field of the ident_t struct.
1493enum ModeFlagsTy : unsigned {
/// Bit set to 1 when in SPMD mode.
KMP_IDENT_SPMD_MODE = 0x01,
/// Bit set to 1 when a simplified runtime is used.
KMP_IDENT_SIMPLE_RT_MODE = 0x02,
LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/KMP_IDENT_SIMPLE_RT_MODE)LLVM_BITMASK_LARGEST_ENUMERATOR = KMP_IDENT_SIMPLE_RT_MODE
1499};

1501/// Special mode Undefined. Is the combination of Non-SPMD mode + SimpleRuntime.
1502static const ModeFlagsTy UndefinedMode =
  (~KMP_IDENT_SPMD_MODE) & KMP_IDENT_SIMPLE_RT_MODE;
1504} // anonymous namespace

1506unsigned CGOpenMPRuntimeGPU::getDefaultLocationReserved2Flags() const {
switch (getExecutionMode()) {
case EM_SPMD:
  if (requiresFullRuntime())
    return KMP_IDENT_SPMD_MODE & (~KMP_IDENT_SIMPLE_RT_MODE);
  return KMP_IDENT_SPMD_MODE | KMP_IDENT_SIMPLE_RT_MODE;
case EM_NonSPMD:
  assert(requiresFullRuntime() && "Expected full runtime.")((requiresFullRuntime() && "Expected full runtime.") ?
 static_cast<void> (0) : __assert_fail ("requiresFullRuntime() && \"Expected full runtime.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1513, __PRETTY_FUNCTION__));
  return (~KMP_IDENT_SPMD_MODE) & (~KMP_IDENT_SIMPLE_RT_MODE);
case EM_Unknown:
  return UndefinedMode;
}
llvm_unreachable("Unknown flags are requested.")::llvm::llvm_unreachable_internal("Unknown flags are requested."
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1518);
1519}

1521CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM)
  : CGOpenMPRuntime(CGM, "_", "$") {
if (!CGM.getLangOpts().OpenMPIsDevice)
  llvm_unreachable("OpenMP NVPTX can only handle device code.")::llvm::llvm_unreachable_internal("OpenMP NVPTX can only handle device code."
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1524);
1525}

1527void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF,
                                            ProcBindKind ProcBind,
                                            SourceLocation Loc) {
// Do nothing in case of SPMD mode and L0 parallel.
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
  return;

CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc);
1535}

1537void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF,
                                              llvm::Value *NumThreads,
                                              SourceLocation Loc) {
// Do nothing in case of SPMD mode and L0 parallel.
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
  return;

CGOpenMPRuntime::emitNumThreadsClause(CGF, NumThreads, Loc);
1545}

1547void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF,
                                            const Expr *NumTeams,
                                            const Expr *ThreadLimit,
                                            SourceLocation Loc) {}

1552llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction(
  const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
  OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
// Emit target region as a standalone region.
class NVPTXPrePostActionTy : public PrePostActionTy {
  bool &IsInParallelRegion;
  bool PrevIsInParallelRegion;

public:
  NVPTXPrePostActionTy(bool &IsInParallelRegion)
      : IsInParallelRegion(IsInParallelRegion) {}
  void Enter(CodeGenFunction &CGF) override {
    PrevIsInParallelRegion = IsInParallelRegion;
    IsInParallelRegion = true;
  }
  void Exit(CodeGenFunction &CGF) override {
    IsInParallelRegion = PrevIsInParallelRegion;
  }
} Action(IsInParallelRegion);
CodeGen.setAction(Action);
bool PrevIsInTTDRegion = IsInTTDRegion;
IsInTTDRegion = false;
bool PrevIsInTargetMasterThreadRegion = IsInTargetMasterThreadRegion;
IsInTargetMasterThreadRegion = false;
auto *OutlinedFun =
    cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction(
        D, ThreadIDVar, InnermostKind, CodeGen));
if (CGM.getLangOpts().Optimize) {
  OutlinedFun->removeFnAttr(llvm::Attribute::NoInline);
  OutlinedFun->removeFnAttr(llvm::Attribute::OptimizeNone);
  OutlinedFun->addFnAttr(llvm::Attribute::AlwaysInline);
}
IsInTargetMasterThreadRegion = PrevIsInTargetMasterThreadRegion;
IsInTTDRegion = PrevIsInTTDRegion;
if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD &&
    !IsInParallelRegion) {
  llvm::Function *WrapperFun =
      createParallelDataSharingWrapper(OutlinedFun, D);
  WrapperFunctionsMap[OutlinedFun] = WrapperFun;
}

return OutlinedFun;
1594}

1596/// Get list of lastprivate variables from the teams distribute ... or
1597/// teams {distribute ...} directives.
1598static void
1599getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D,
                           llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&((isOpenMPTeamsDirective(D.getDirectiveKind()) && "expected teams directive."
) ? static_cast<void> (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1602, __PRETTY_FUNCTION__))
       "expected teams directive.")((isOpenMPTeamsDirective(D.getDirectiveKind()) && "expected teams directive."
) ? static_cast<void> (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1602, __PRETTY_FUNCTION__));
const OMPExecutableDirective *Dir = &D;
if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
  if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild(
          Ctx,
          D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
              /*IgnoreCaptured=*/true))) {
    Dir = dyn_cast_or_null<OMPExecutableDirective>(S);
    if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind()))
      Dir = nullptr;
  }
}
if (!Dir)
  return;
for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) {
  for (const Expr *E : C->getVarRefs())
    Vars.push_back(getPrivateItem(E));
}
1620}

1622/// Get list of reduction variables from the teams ... directives.
1623static void
1624getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D,
                    llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&((isOpenMPTeamsDirective(D.getDirectiveKind()) && "expected teams directive."
) ? static_cast<void> (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1627, __PRETTY_FUNCTION__))
       "expected teams directive.")((isOpenMPTeamsDirective(D.getDirectiveKind()) && "expected teams directive."
) ? static_cast<void> (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1627, __PRETTY_FUNCTION__));
for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
  for (const Expr *E : C->privates())
    Vars.push_back(getPrivateItem(E));
}
1632}

1634llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction(
  const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
  OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
SourceLocation Loc = D.getBeginLoc();

const RecordDecl *GlobalizedRD = nullptr;
llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions;
llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
unsigned WarpSize = CGM.getTarget().getGridValue(llvm::omp::GV_Warp_Size);
// Globalize team reductions variable unconditionally in all modes.
if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
  getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions);
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
  getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions);
  if (!LastPrivatesReductions.empty()) {
    GlobalizedRD = ::buildRecordForGlobalizedVars(
        CGM.getContext(), llvm::None, LastPrivatesReductions,
        MappedDeclsFields, WarpSize);
  }
} else if (!LastPrivatesReductions.empty()) {
  assert(!TeamAndReductions.first &&((!TeamAndReductions.first && "Previous team declaration is not expected."
) ? static_cast<void> (0) : __assert_fail ("!TeamAndReductions.first && \"Previous team declaration is not expected.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1655, __PRETTY_FUNCTION__))
         "Previous team declaration is not expected.")((!TeamAndReductions.first && "Previous team declaration is not expected."
) ? static_cast<void> (0) : __assert_fail ("!TeamAndReductions.first && \"Previous team declaration is not expected.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1655, __PRETTY_FUNCTION__));
  TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl();
  std::swap(TeamAndReductions.second, LastPrivatesReductions);
}

// Emit target region as a standalone region.
class NVPTXPrePostActionTy : public PrePostActionTy {
  SourceLocation &Loc;
  const RecordDecl *GlobalizedRD;
  llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
      &MappedDeclsFields;

public:
  NVPTXPrePostActionTy(
      SourceLocation &Loc, const RecordDecl *GlobalizedRD,
      llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
          &MappedDeclsFields)
      : Loc(Loc), GlobalizedRD(GlobalizedRD),
        MappedDeclsFields(MappedDeclsFields) {}
  void Enter(CodeGenFunction &CGF) override {
    auto &Rt =
        static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
    if (GlobalizedRD) {
      auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
      I->getSecond().GlobalRecord = GlobalizedRD;
      I->getSecond().MappedParams =
          std::make_unique<CodeGenFunction::OMPMapVars>();
      DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
      for (const auto &Pair : MappedDeclsFields) {
        assert(Pair.getFirst()->isCanonicalDecl() &&((Pair.getFirst()->isCanonicalDecl() && "Expected canonical declaration"
) ? static_cast<void> (0) : __assert_fail ("Pair.getFirst()->isCanonicalDecl() && \"Expected canonical declaration\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1685, __PRETTY_FUNCTION__))
               "Expected canonical declaration")((Pair.getFirst()->isCanonicalDecl() && "Expected canonical declaration"
) ? static_cast<void> (0) : __assert_fail ("Pair.getFirst()->isCanonicalDecl() && \"Expected canonical declaration\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1685, __PRETTY_FUNCTION__));
        Data.insert(std::make_pair(Pair.getFirst(),
                                   MappedVarData(Pair.getSecond(),
                                                 /*IsOnePerTeam=*/true)));
      }
    }
    Rt.emitGenericVarsProlog(CGF, Loc);
  }
  void Exit(CodeGenFunction &CGF) override {
    static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
        .emitGenericVarsEpilog(CGF);
  }
} Action(Loc, GlobalizedRD, MappedDeclsFields);
CodeGen.setAction(Action);
llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction(
    D, ThreadIDVar, InnermostKind, CodeGen);
if (CGM.getLangOpts().Optimize) {
  OutlinedFun->removeFnAttr(llvm::Attribute::NoInline);
  OutlinedFun->removeFnAttr(llvm::Attribute::OptimizeNone);
  OutlinedFun->addFnAttr(llvm::Attribute::AlwaysInline);
}

return OutlinedFun;
1708}

1710void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF,
                                               SourceLocation Loc,
                                               bool WithSPMDCheck) {
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
    getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
  return;

CGBuilderTy &Bld = CGF.Builder;

const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
if (I == FunctionGlobalizedDecls.end())
  return;
if (const RecordDecl *GlobalizedVarsRecord = I->getSecond().GlobalRecord) {
  QualType GlobalRecTy = CGM.getContext().getRecordType(GlobalizedVarsRecord);
  QualType SecGlobalRecTy;

  // Recover pointer to this function's global record. The runtime will
  // handle the specifics of the allocation of the memory.
  // Use actual memory size of the record including the padding
  // for alignment purposes.
  unsigned Alignment =
      CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity();
  unsigned GlobalRecordSize =
      CGM.getContext().getTypeSizeInChars(GlobalRecTy).getQuantity();
  GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment);

  llvm::PointerType *GlobalRecPtrTy =
      CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo();
  llvm::Value *GlobalRecCastAddr;
  llvm::Value *IsTTD = nullptr;
  if (!IsInTTDRegion &&
      (WithSPMDCheck ||
       getExecutionMode() == CGOpenMPRuntimeGPU::EM_Unknown)) {
    llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
    llvm::BasicBlock *SPMDBB = CGF.createBasicBlock(".spmd");
    llvm::BasicBlock *NonSPMDBB = CGF.createBasicBlock(".non-spmd");
    if (I->getSecond().SecondaryGlobalRecord.hasValue()) {
      llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
      llvm::Value *ThreadID = getThreadID(CGF, Loc);
      llvm::Value *PL = CGF.EmitRuntimeCall(
          OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
                                                OMPRTL___kmpc_parallel_level),
          {RTLoc, ThreadID});
      IsTTD = Bld.CreateIsNull(PL);
    }
    llvm::Value *IsSPMD = Bld.CreateIsNotNull(
        CGF.EmitNounwindRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
            CGM.getModule(), OMPRTL___kmpc_is_spmd_exec_mode)));
    Bld.CreateCondBr(IsSPMD, SPMDBB, NonSPMDBB);
    // There is no need to emit line number for unconditional branch.
    (void)ApplyDebugLocation::CreateEmpty(CGF);
    CGF.EmitBlock(SPMDBB);
    Address RecPtr = Address(llvm::ConstantPointerNull::get(GlobalRecPtrTy),
                             CharUnits::fromQuantity(Alignment));
    CGF.EmitBranch(ExitBB);
    // There is no need to emit line number for unconditional branch.
    (void)ApplyDebugLocation::CreateEmpty(CGF);
    CGF.EmitBlock(NonSPMDBB);
    llvm::Value *Size = llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize);
    if (const RecordDecl *SecGlobalizedVarsRecord =
            I->getSecond().SecondaryGlobalRecord.getValueOr(nullptr)) {
      SecGlobalRecTy =
          CGM.getContext().getRecordType(SecGlobalizedVarsRecord);

      // Recover pointer to this function's global record. The runtime will
      // handle the specifics of the allocation of the memory.
      // Use actual memory size of the record including the padding
      // for alignment purposes.
      unsigned Alignment =
          CGM.getContext().getTypeAlignInChars(SecGlobalRecTy).getQuantity();
      unsigned GlobalRecordSize =
          CGM.getContext().getTypeSizeInChars(SecGlobalRecTy).getQuantity();
      GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment);
      Size = Bld.CreateSelect(
          IsTTD, llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize), Size);
    }
    // TODO: allow the usage of shared memory to be controlled by
    // the user, for now, default to global.
    llvm::Value *GlobalRecordSizeArg[] = {
        Size, CGF.Builder.getInt16(/*UseSharedMemory=*/0)};
    llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
        OMPBuilder.getOrCreateRuntimeFunction(
            CGM.getModule(), OMPRTL___kmpc_data_sharing_coalesced_push_stack),
        GlobalRecordSizeArg);
    GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
        GlobalRecValue, GlobalRecPtrTy);
    CGF.EmitBlock(ExitBB);
    auto *Phi = Bld.CreatePHI(GlobalRecPtrTy,
                              /*NumReservedValues=*/2, "_select_stack");
    Phi->addIncoming(RecPtr.getPointer(), SPMDBB);
    Phi->addIncoming(GlobalRecCastAddr, NonSPMDBB);
    GlobalRecCastAddr = Phi;
    I->getSecond().GlobalRecordAddr = Phi;
    I->getSecond().IsInSPMDModeFlag = IsSPMD;
  } else if (!CGM.getLangOpts().OpenMPCUDATargetParallel && IsInTTDRegion) {
    assert(GlobalizedRecords.back().Records.size() < 2 &&((GlobalizedRecords.back().Records.size() < 2 && "Expected less than 2 globalized records: one for target and one "
 "for teams.") ? static_cast<void> (0) : __assert_fail (
"GlobalizedRecords.back().Records.size() < 2 && \"Expected less than 2 globalized records: one for target and one \" \"for teams.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1807, __PRETTY_FUNCTION__))
           "Expected less than 2 globalized records: one for target and one "((GlobalizedRecords.back().Records.size() < 2 && "Expected less than 2 globalized records: one for target and one "
 "for teams.") ? static_cast<void> (0) : __assert_fail (
"GlobalizedRecords.back().Records.size() < 2 && \"Expected less than 2 globalized records: one for target and one \" \"for teams.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1807, __PRETTY_FUNCTION__))
           "for teams.")((GlobalizedRecords.back().Records.size() < 2 && "Expected less than 2 globalized records: one for target and one "
 "for teams.") ? static_cast<void> (0) : __assert_fail (
"GlobalizedRecords.back().Records.size() < 2 && \"Expected less than 2 globalized records: one for target and one \" \"for teams.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1807, __PRETTY_FUNCTION__));
    unsigned Offset = 0;
    for (const RecordDecl *RD : GlobalizedRecords.back().Records) {
      QualType RDTy = CGM.getContext().getRecordType(RD);
      unsigned Alignment =
          CGM.getContext().getTypeAlignInChars(RDTy).getQuantity();
      unsigned Size = CGM.getContext().getTypeSizeInChars(RDTy).getQuantity();
      Offset =
          llvm::alignTo(llvm::alignTo(Offset, Alignment) + Size, Alignment);
    }
    unsigned Alignment =
        CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity();
    Offset = llvm::alignTo(Offset, Alignment);
    GlobalizedRecords.back().Records.push_back(GlobalizedVarsRecord);
    ++GlobalizedRecords.back().RegionCounter;
    if (GlobalizedRecords.back().Records.size() == 1) {
      assert(KernelStaticGlobalized &&((KernelStaticGlobalized && "Kernel static pointer must be initialized already."
) ? static_cast<void> (0) : __assert_fail ("KernelStaticGlobalized && \"Kernel static pointer must be initialized already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1824, __PRETTY_FUNCTION__))
             "Kernel static pointer must be initialized already.")((KernelStaticGlobalized && "Kernel static pointer must be initialized already."
) ? static_cast<void> (0) : __assert_fail ("KernelStaticGlobalized && \"Kernel static pointer must be initialized already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1824, __PRETTY_FUNCTION__));
      auto *UseSharedMemory = new llvm::GlobalVariable(
          CGM.getModule(), CGM.Int16Ty, /*isConstant=*/true,
          llvm::GlobalValue::InternalLinkage, nullptr,
          "_openmp_static_kernel$is_shared");
      UseSharedMemory->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
      QualType Int16Ty = CGM.getContext().getIntTypeForBitwidth(
          /*DestWidth=*/16, /*Signed=*/0);
      llvm::Value *IsInSharedMemory = CGF.EmitLoadOfScalar(
          Address(UseSharedMemory,
                  CGM.getContext().getTypeAlignInChars(Int16Ty)),
          /*Volatile=*/false, Int16Ty, Loc);
      auto *StaticGlobalized = new llvm::GlobalVariable(
          CGM.getModule(), CGM.Int8Ty, /*isConstant=*/false,
          llvm::GlobalValue::CommonLinkage, nullptr);
      auto *RecSize = new llvm::GlobalVariable(
          CGM.getModule(), CGM.SizeTy, /*isConstant=*/true,
          llvm::GlobalValue::InternalLinkage, nullptr,
          "_openmp_static_kernel$size");
      RecSize->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
      llvm::Value *Ld = CGF.EmitLoadOfScalar(
          Address(RecSize, CGM.getSizeAlign()), /*Volatile=*/false,
          CGM.getContext().getSizeType(), Loc);
      llvm::Value *ResAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
          KernelStaticGlobalized, CGM.VoidPtrPtrTy);
      llvm::Value *GlobalRecordSizeArg[] = {
          llvm::ConstantInt::get(
              CGM.Int16Ty,
              getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD ? 1 : 0),
          StaticGlobalized, Ld, IsInSharedMemory, ResAddr};
      CGF.EmitRuntimeCall(
          OMPBuilder.getOrCreateRuntimeFunction(
              CGM.getModule(), OMPRTL___kmpc_get_team_static_memory),
          GlobalRecordSizeArg);
      GlobalizedRecords.back().Buffer = StaticGlobalized;
      GlobalizedRecords.back().RecSize = RecSize;
      GlobalizedRecords.back().UseSharedMemory = UseSharedMemory;
      GlobalizedRecords.back().Loc = Loc;
    }
    assert(KernelStaticGlobalized && "Global address must be set already.")((KernelStaticGlobalized && "Global address must be set already."
) ? static_cast<void> (0) : __assert_fail ("KernelStaticGlobalized && \"Global address must be set already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1863, __PRETTY_FUNCTION__));
    Address FrameAddr = CGF.EmitLoadOfPointer(
        Address(KernelStaticGlobalized, CGM.getPointerAlign()),
        CGM.getContext()
            .getPointerType(CGM.getContext().VoidPtrTy)
            .castAs<PointerType>());
    llvm::Value *GlobalRecValue =
        Bld.CreateConstInBoundsGEP(FrameAddr, Offset).getPointer();
    I->getSecond().GlobalRecordAddr = GlobalRecValue;
    I->getSecond().IsInSPMDModeFlag = nullptr;
    GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
        GlobalRecValue, CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo());
  } else {
    // TODO: allow the usage of shared memory to be controlled by
    // the user, for now, default to global.
    bool UseSharedMemory =
        IsInTTDRegion && GlobalRecordSize <= SharedMemorySize;
    llvm::Value *GlobalRecordSizeArg[] = {
        llvm::ConstantInt::get(CGM.SizeTy, GlobalRecordSize),
        CGF.Builder.getInt16(UseSharedMemory ? 1 : 0)};
    llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
        OMPBuilder.getOrCreateRuntimeFunction(
            CGM.getModule(),
            IsInTTDRegion ? OMPRTL___kmpc_data_sharing_push_stack
                          : OMPRTL___kmpc_data_sharing_coalesced_push_stack),
        GlobalRecordSizeArg);
    GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
        GlobalRecValue, GlobalRecPtrTy);
    I->getSecond().GlobalRecordAddr = GlobalRecValue;
    I->getSecond().IsInSPMDModeFlag = nullptr;
  }
  LValue Base =
      CGF.MakeNaturalAlignPointeeAddrLValue(GlobalRecCastAddr, GlobalRecTy);

  // Emit the "global alloca" which is a GEP from the global declaration
  // record using the pointer returned by the runtime.
  LValue SecBase;
  decltype(I->getSecond().LocalVarData)::const_iterator SecIt;
  if (IsTTD) {
    SecIt = I->getSecond().SecondaryLocalVarData->begin();
    llvm::PointerType *SecGlobalRecPtrTy =
        CGF.ConvertTypeForMem(SecGlobalRecTy)->getPointerTo();
    SecBase = CGF.MakeNaturalAlignPointeeAddrLValue(
        Bld.CreatePointerBitCastOrAddrSpaceCast(
            I->getSecond().GlobalRecordAddr, SecGlobalRecPtrTy),
        SecGlobalRecTy);
  }
  for (auto &Rec : I->getSecond().LocalVarData) {
    bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first);
    llvm::Value *ParValue;
    if (EscapedParam) {
      const auto *VD = cast<VarDecl>(Rec.first);
      LValue ParLVal =
          CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
      ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
    }
    LValue VarAddr = CGF.EmitLValueForField(Base, Rec.second.FD);
    // Emit VarAddr basing on lane-id if required.
    QualType VarTy;
    if (Rec.second.IsOnePerTeam) {
      VarTy = Rec.second.FD->getType();
    } else {
      llvm::Value *Ptr = CGF.Builder.CreateInBoundsGEP(
          VarAddr.getAddress(CGF).getPointer(),
          {Bld.getInt32(0), getNVPTXLaneID(CGF)});
      VarTy =
          Rec.second.FD->getType()->castAsArrayTypeUnsafe()->getElementType();
      VarAddr = CGF.MakeAddrLValue(
          Address(Ptr, CGM.getContext().getDeclAlign(Rec.first)), VarTy,
          AlignmentSource::Decl);
    }
    Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
    if (!IsInTTDRegion &&
        (WithSPMDCheck ||
         getExecutionMode() == CGOpenMPRuntimeGPU::EM_Unknown)) {
      assert(I->getSecond().IsInSPMDModeFlag &&((I->getSecond().IsInSPMDModeFlag && "Expected unknown execution mode or required SPMD check."
) ? static_cast<void> (0) : __assert_fail ("I->getSecond().IsInSPMDModeFlag && \"Expected unknown execution mode or required SPMD check.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1939, __PRETTY_FUNCTION__))
             "Expected unknown execution mode or required SPMD check.")((I->getSecond().IsInSPMDModeFlag && "Expected unknown execution mode or required SPMD check."
) ? static_cast<void> (0) : __assert_fail ("I->getSecond().IsInSPMDModeFlag && \"Expected unknown execution mode or required SPMD check.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1939, __PRETTY_FUNCTION__));
      if (IsTTD) {
        assert(SecIt->second.IsOnePerTeam &&((SecIt->second.IsOnePerTeam && "Secondary glob data must be one per team."
) ? static_cast<void> (0) : __assert_fail ("SecIt->second.IsOnePerTeam && \"Secondary glob data must be one per team.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1942, __PRETTY_FUNCTION__))
               "Secondary glob data must be one per team.")((SecIt->second.IsOnePerTeam && "Secondary glob data must be one per team."
) ? static_cast<void> (0) : __assert_fail ("SecIt->second.IsOnePerTeam && \"Secondary glob data must be one per team.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 1942, __PRETTY_FUNCTION__));
        LValue SecVarAddr = CGF.EmitLValueForField(SecBase, SecIt->second.FD);
        VarAddr.setAddress(
            Address(Bld.CreateSelect(IsTTD, SecVarAddr.getPointer(CGF),
                                     VarAddr.getPointer(CGF)),
                    VarAddr.getAlignment()));
        Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
      }
      Address GlobalPtr = Rec.second.PrivateAddr;
      Address LocalAddr = CGF.CreateMemTemp(VarTy, Rec.second.FD->getName());
      Rec.second.PrivateAddr = Address(
          Bld.CreateSelect(I->getSecond().IsInSPMDModeFlag,
                           LocalAddr.getPointer(), GlobalPtr.getPointer()),
          LocalAddr.getAlignment());
    }
    if (EscapedParam) {
      const auto *VD = cast<VarDecl>(Rec.first);
      CGF.EmitStoreOfScalar(ParValue, VarAddr);
      I->getSecond().MappedParams->setVarAddr(CGF, VD,
                                              VarAddr.getAddress(CGF));
    }
    if (IsTTD)
      ++SecIt;
  }
}
for (const ValueDecl *VD : I->getSecond().EscapedVariableLengthDecls) {
  // Recover pointer to this function's global record. The runtime will
  // handle the specifics of the allocation of the memory.
  // Use actual memory size of the record including the padding
  // for alignment purposes.
  CGBuilderTy &Bld = CGF.Builder;
  llvm::Value *Size = CGF.getTypeSize(VD->getType());
  CharUnits Align = CGM.getContext().getDeclAlign(VD);
  Size = Bld.CreateNUWAdd(
      Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1));
  llvm::Value *AlignVal =
      llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity());
  Size = Bld.CreateUDiv(Size, AlignVal);
  Size = Bld.CreateNUWMul(Size, AlignVal);
  // TODO: allow the usage of shared memory to be controlled by
  // the user, for now, default to global.
  llvm::Value *GlobalRecordSizeArg[] = {
      Size, CGF.Builder.getInt16(/*UseSharedMemory=*/0)};
  llvm::Value *GlobalRecValue = CGF.EmitRuntimeCall(
      OMPBuilder.getOrCreateRuntimeFunction(
          CGM.getModule(), OMPRTL___kmpc_data_sharing_coalesced_push_stack),
      GlobalRecordSizeArg);
  llvm::Value *GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
      GlobalRecValue, CGF.ConvertTypeForMem(VD->getType())->getPointerTo());
  LValue Base = CGF.MakeAddrLValue(GlobalRecCastAddr, VD->getType(),
                                   CGM.getContext().getDeclAlign(VD),
                                   AlignmentSource::Decl);
  I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD),
                                          Base.getAddress(CGF));
  I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(GlobalRecValue);
}
I->getSecond().MappedParams->apply(CGF);
1999}

2001void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF,
                                               bool WithSPMDCheck) {
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
    getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
  return;

const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
if (I != FunctionGlobalizedDecls.end()) {
  I->getSecond().MappedParams->restore(CGF);
  if (!CGF.HaveInsertPoint())
    return;
  for (llvm::Value *Addr :
       llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) {
    CGF.EmitRuntimeCall(
        OMPBuilder.getOrCreateRuntimeFunction(
            CGM.getModule(), OMPRTL___kmpc_data_sharing_pop_stack),
        Addr);
  }
  if (I->getSecond().GlobalRecordAddr) {
    if (!IsInTTDRegion &&
        (WithSPMDCheck ||
         getExecutionMode() == CGOpenMPRuntimeGPU::EM_Unknown)) {
      CGBuilderTy &Bld = CGF.Builder;
      llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
      llvm::BasicBlock *NonSPMDBB = CGF.createBasicBlock(".non-spmd");
      Bld.CreateCondBr(I->getSecond().IsInSPMDModeFlag, ExitBB, NonSPMDBB);
      // There is no need to emit line number for unconditional branch.
      (void)ApplyDebugLocation::CreateEmpty(CGF);
      CGF.EmitBlock(NonSPMDBB);
      CGF.EmitRuntimeCall(
          OMPBuilder.getOrCreateRuntimeFunction(
              CGM.getModule(), OMPRTL___kmpc_data_sharing_pop_stack),
          CGF.EmitCastToVoidPtr(I->getSecond().GlobalRecordAddr));
      CGF.EmitBlock(ExitBB);
    } else if (!CGM.getLangOpts().OpenMPCUDATargetParallel && IsInTTDRegion) {
      assert(GlobalizedRecords.back().RegionCounter > 0 &&((GlobalizedRecords.back().RegionCounter > 0 && "region counter must be > 0."
) ? static_cast<void> (0) : __assert_fail ("GlobalizedRecords.back().RegionCounter > 0 && \"region counter must be > 0.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2037, __PRETTY_FUNCTION__))
             "region counter must be > 0.")((GlobalizedRecords.back().RegionCounter > 0 && "region counter must be > 0."
) ? static_cast<void> (0) : __assert_fail ("GlobalizedRecords.back().RegionCounter > 0 && \"region counter must be > 0.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2037, __PRETTY_FUNCTION__));
      --GlobalizedRecords.back().RegionCounter;
      // Emit the restore function only in the target region.
      if (GlobalizedRecords.back().RegionCounter == 0) {
        QualType Int16Ty = CGM.getContext().getIntTypeForBitwidth(
            /*DestWidth=*/16, /*Signed=*/0);
        llvm::Value *IsInSharedMemory = CGF.EmitLoadOfScalar(
            Address(GlobalizedRecords.back().UseSharedMemory,
                    CGM.getContext().getTypeAlignInChars(Int16Ty)),
            /*Volatile=*/false, Int16Ty, GlobalizedRecords.back().Loc);
        llvm::Value *Args[] = {
            llvm::ConstantInt::get(
                CGM.Int16Ty,
                getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD ? 1 : 0),
            IsInSharedMemory};
        CGF.EmitRuntimeCall(
            OMPBuilder.getOrCreateRuntimeFunction(
                CGM.getModule(), OMPRTL___kmpc_restore_team_static_memory),
            Args);
      }
    } else {
      CGF.EmitRuntimeCall(
          OMPBuilder.getOrCreateRuntimeFunction(
              CGM.getModule(), OMPRTL___kmpc_data_sharing_pop_stack),
          I->getSecond().GlobalRecordAddr);
    }
  }
}
2065}

2067void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF,
                                       const OMPExecutableDirective &D,
                                       SourceLocation Loc,
                                       llvm::Function *OutlinedFn,
                                       ArrayRef<llvm::Value *> CapturedVars) {
if (!CGF.HaveInsertPoint())
  return;

Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
                                                    /*Name=*/".zero.addr");
CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer());
OutlinedFnArgs.push_back(ZeroAddr.getPointer());
OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
2083}

2085void CGOpenMPRuntimeGPU::emitParallelCall(
  CodeGenFunction &CGF, SourceLocation Loc, llvm::Function *OutlinedFn,
  ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
if (!CGF.HaveInsertPoint())
  return;

if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
  emitSPMDParallelCall(CGF, Loc, OutlinedFn, CapturedVars, IfCond);
else
  emitNonSPMDParallelCall(CGF, Loc, OutlinedFn, CapturedVars, IfCond);
2095}

2097void CGOpenMPRuntimeGPU::emitNonSPMDParallelCall(
  CodeGenFunction &CGF, SourceLocation Loc, llvm::Value *OutlinedFn,
  ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
llvm::Function *Fn = cast<llvm::Function>(OutlinedFn);

// Force inline this outlined function at its call site.
Fn->setLinkage(llvm::GlobalValue::InternalLinkage);

Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
                                                    /*Name=*/".zero.addr");
CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
// ThreadId for serialized parallels is 0.
Address ThreadIDAddr = ZeroAddr;
auto &&CodeGen = [this, Fn, CapturedVars, Loc, &ThreadIDAddr](
                     CodeGenFunction &CGF, PrePostActionTy &Action) {
  Action.Enter(CGF);

  Address ZeroAddr =
      CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
                                       /*Name=*/".bound.zero.addr");
  CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
  llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
  OutlinedFnArgs.push_back(ThreadIDAddr.getPointer());
  OutlinedFnArgs.push_back(ZeroAddr.getPointer());
  OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
  emitOutlinedFunctionCall(CGF, Loc, Fn, OutlinedFnArgs);
};
auto &&SeqGen = [this, &CodeGen, Loc](CodeGenFunction &CGF,
                                      PrePostActionTy &) {

  RegionCodeGenTy RCG(CodeGen);
  llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
  llvm::Value *ThreadID = getThreadID(CGF, Loc);
  llvm::Value *Args[] = {RTLoc, ThreadID};

  NVPTXActionTy Action(
      OMPBuilder.getOrCreateRuntimeFunction(
          CGM.getModule(), OMPRTL___kmpc_serialized_parallel),
      Args,
      OMPBuilder.getOrCreateRuntimeFunction(
          CGM.getModule(), OMPRTL___kmpc_end_serialized_parallel),
      Args);
  RCG.setAction(Action);
  RCG(CGF);
};

auto &&L0ParallelGen = [this, CapturedVars, Fn](CodeGenFunction &CGF,
                                                PrePostActionTy &Action) {
  CGBuilderTy &Bld = CGF.Builder;
  llvm::Function *WFn = WrapperFunctionsMap[Fn];
  assert(WFn && "Wrapper function does not exist!")((WFn && "Wrapper function does not exist!") ? static_cast
<void> (0) : __assert_fail ("WFn && \"Wrapper function does not exist!\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2147, __PRETTY_FUNCTION__));
  llvm::Value *ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy);

  // Prepare for parallel region. Indicate the outlined function.
  llvm::Value *Args[] = {ID};
  CGF.EmitRuntimeCall(
      OMPBuilder.getOrCreateRuntimeFunction(
          CGM.getModule(), OMPRTL___kmpc_kernel_prepare_parallel),
      Args);

  // Create a private scope that will globalize the arguments
  // passed from the outside of the target region.
  CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF);

  // There's something to share.
  if (!CapturedVars.empty()) {
    // Prepare for parallel region. Indicate the outlined function.
    Address SharedArgs =
        CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "shared_arg_refs");
    llvm::Value *SharedArgsPtr = SharedArgs.getPointer();

    llvm::Value *DataSharingArgs[] = {
        SharedArgsPtr,
        llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())};
    CGF.EmitRuntimeCall(
        OMPBuilder.getOrCreateRuntimeFunction(
            CGM.getModule(), OMPRTL___kmpc_begin_sharing_variables),
        DataSharingArgs);

    // Store variable address in a list of references to pass to workers.
    unsigned Idx = 0;
    ASTContext &Ctx = CGF.getContext();
    Address SharedArgListAddress = CGF.EmitLoadOfPointer(
        SharedArgs, Ctx.getPointerType(Ctx.getPointerType(Ctx.VoidPtrTy))
                        .castAs<PointerType>());
    for (llvm::Value *V : CapturedVars) {
      Address Dst = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
      llvm::Value *PtrV;
      if (V->getType()->isIntegerTy())
        PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy);
      else
        PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy);
      CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false,
                            Ctx.getPointerType(Ctx.VoidPtrTy));
      ++Idx;
    }
  }

  // Activate workers. This barrier is used by the master to signal
  // work for the workers.
  syncCTAThreads(CGF);

  // OpenMP [2.5, Parallel Construct, p.49]
  // There is an implied barrier at the end of a parallel region. After the
  // end of a parallel region, only the master thread of the team resumes
  // execution of the enclosing task region.
  //
  // The master waits at this barrier until all workers are done.
  syncCTAThreads(CGF);

  if (!CapturedVars.empty())
    CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
        CGM.getModule(), OMPRTL___kmpc_end_sharing_variables));

  // Remember for post-processing in worker loop.
  Work.emplace_back(WFn);
};

auto &&LNParallelGen = [this, Loc, &SeqGen, &L0ParallelGen](
                           CodeGenFunction &CGF, PrePostActionTy &Action) {
  if (IsInParallelRegion) {
    SeqGen(CGF, Action);
  } else if (IsInTargetMasterThreadRegion) {
    L0ParallelGen(CGF, Action);
  } else {
    // Check for master and then parallelism:
    // if (__kmpc_is_spmd_exec_mode() || __kmpc_parallel_level(loc, gtid)) {
    //   Serialized execution.
    // } else {
    //   Worker call.
    // }
    CGBuilderTy &Bld = CGF.Builder;
    llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
    llvm::BasicBlock *SeqBB = CGF.createBasicBlock(".sequential");
    llvm::BasicBlock *ParallelCheckBB = CGF.createBasicBlock(".parcheck");
    llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master");
    llvm::Value *IsSPMD = Bld.CreateIsNotNull(
        CGF.EmitNounwindRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
            CGM.getModule(), OMPRTL___kmpc_is_spmd_exec_mode)));
    Bld.CreateCondBr(IsSPMD, SeqBB, ParallelCheckBB);
    // There is no need to emit line number for unconditional branch.
    (void)ApplyDebugLocation::CreateEmpty(CGF);
    CGF.EmitBlock(ParallelCheckBB);
    llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
    llvm::Value *ThreadID = getThreadID(CGF, Loc);
    llvm::Value *PL = CGF.EmitRuntimeCall(
        OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(),
                                              OMPRTL___kmpc_parallel_level),
        {RTLoc, ThreadID});
    llvm::Value *Res = Bld.CreateIsNotNull(PL);
    Bld.CreateCondBr(Res, SeqBB, MasterBB);
    CGF.EmitBlock(SeqBB);
    SeqGen(CGF, Action);
    CGF.EmitBranch(ExitBB);
    // There is no need to emit line number for unconditional branch.
    (void)ApplyDebugLocation::CreateEmpty(CGF);
    CGF.EmitBlock(MasterBB);
    L0ParallelGen(CGF, Action);
    CGF.EmitBranch(ExitBB);
    // There is no need to emit line number for unconditional branch.
    (void)ApplyDebugLocation::CreateEmpty(CGF);
    // Emit the continuation block for code after the if.
    CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
  }
};

if (IfCond) {
  emitIfClause(CGF, IfCond, LNParallelGen, SeqGen);
} else {
  CodeGenFunction::RunCleanupsScope Scope(CGF);
  RegionCodeGenTy ThenRCG(LNParallelGen);
  ThenRCG(CGF);
}
2270}

2272void CGOpenMPRuntimeGPU::emitSPMDParallelCall(
  CodeGenFunction &CGF, SourceLocation Loc, llvm::Function *OutlinedFn,
  ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond) {
// Just call the outlined function to execute the parallel region.
// OutlinedFn(&GTid, &zero, CapturedStruct);
//
llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;

Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
                                                    /*Name=*/".zero.addr");
CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
// ThreadId for serialized parallels is 0.
Address ThreadIDAddr = ZeroAddr;
auto &&CodeGen = [this, OutlinedFn, CapturedVars, Loc, &ThreadIDAddr](
                     CodeGenFunction &CGF, PrePostActionTy &Action) {
  Action.Enter(CGF);

  Address ZeroAddr =
      CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
                                       /*Name=*/".bound.zero.addr");
  CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
  llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
  OutlinedFnArgs.push_back(ThreadIDAddr.getPointer());
  OutlinedFnArgs.push_back(ZeroAddr.getPointer());
  OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
  emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
};
auto &&SeqGen = [this, &CodeGen, Loc](CodeGenFunction &CGF,
                                      PrePostActionTy &) {

  RegionCodeGenTy RCG(CodeGen);
  llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
  llvm::Value *ThreadID = getThreadID(CGF, Loc);
  llvm::Value *Args[] = {RTLoc, ThreadID};

  NVPTXActionTy Action(
      OMPBuilder.getOrCreateRuntimeFunction(
          CGM.getModule(), OMPRTL___kmpc_serialized_parallel),
      Args,
      OMPBuilder.getOrCreateRuntimeFunction(
          CGM.getModule(), OMPRTL___kmpc_end_serialized_parallel),
      Args);
  RCG.setAction(Action);
  RCG(CGF);
};

if (IsInTargetMasterThreadRegion) {
  // In the worker need to use the real thread id.
  ThreadIDAddr = emitThreadIDAddress(CGF, Loc);
  RegionCodeGenTy RCG(CodeGen);
  RCG(CGF);
} else {
  // If we are not in the target region, it is definitely L2 parallelism or
  // more, because for SPMD mode we always has L1 parallel level, sowe don't
  // need to check for orphaned directives.
  RegionCodeGenTy RCG(SeqGen);
  RCG(CGF);
}
2330}

2332void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) {
// Always emit simple barriers!
if (!CGF.HaveInsertPoint())
  return;
// Build call __kmpc_barrier_simple_spmd(nullptr, 0);
// This function does not use parameters, so we can emit just default values.
llvm::Value *Args[] = {
    llvm::ConstantPointerNull::get(
        cast<llvm::PointerType>(getIdentTyPointerTy())),
    llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)};
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                        CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd),
                    Args);
2345}

2347void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF,
                                         SourceLocation Loc,
                                         OpenMPDirectiveKind Kind, bool,
                                         bool) {
// Always emit simple barriers!
if (!CGF.HaveInsertPoint())
  return;
// Build call __kmpc_cancel_barrier(loc, thread_id);
unsigned Flags = getDefaultFlagsForBarriers(Kind);
llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags),
                       getThreadID(CGF, Loc)};

CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                        CGM.getModule(), OMPRTL___kmpc_barrier),
                    Args);
2362}

2364void CGOpenMPRuntimeGPU::emitCriticalRegion(
  CodeGenFunction &CGF, StringRef CriticalName,
  const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc,
  const Expr *Hint) {
llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop");
llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test");
llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync");
llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body");
llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit");

auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());

// Get the mask of active threads in the warp.
llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
    CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask));
// Fetch team-local id of the thread.
llvm::Value *ThreadID = RT.getGPUThreadID(CGF);

// Get the width of the team.
llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF);

// Initialize the counter variable for the loop.
QualType Int32Ty =
    CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0);
Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter");
LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty);
CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal,
                      /*isInit=*/true);

// Block checks if loop counter exceeds upper bound.
CGF.EmitBlock(LoopBB);
llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth);
CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB);

// Block tests which single thread should execute region, and which threads
// should go straight to synchronisation point.
CGF.EmitBlock(TestBB);
CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
llvm::Value *CmpThreadToCounter =
    CGF.Builder.CreateICmpEQ(ThreadID, CounterVal);
CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB);

// Block emits the body of the critical region.
CGF.EmitBlock(BodyBB);

// Output the critical statement.
CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc,
                                    Hint);

// After the body surrounded by the critical region, the single executing
// thread will jump to the synchronisation point.
// Block waits for all threads in current team to finish then increments the
// counter variable and returns to the loop.
CGF.EmitBlock(SyncBB);
// Reconverge active threads in the warp.
(void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                              CGM.getModule(), OMPRTL___kmpc_syncwarp),
                          Mask);

llvm::Value *IncCounterVal =
    CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1));
CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal);
CGF.EmitBranch(LoopBB);

// Block that is reached when  all threads in the team complete the region.
CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
2431}

2433/// Cast value to the specified type.
2434static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val,
                                  QualType ValTy, QualType CastTy,
                                  SourceLocation Loc) {
assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&((!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&
 "Cast type must sized.") ? static_cast<void> (0) : __assert_fail
 ("!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && \"Cast type must sized.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2438, __PRETTY_FUNCTION__))
       "Cast type must sized.")((!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&
 "Cast type must sized.") ? static_cast<void> (0) : __assert_fail
 ("!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && \"Cast type must sized.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2438, __PRETTY_FUNCTION__));
assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&((!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&
 "Val type must sized.") ? static_cast<void> (0) : __assert_fail
 ("!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && \"Val type must sized.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2440, __PRETTY_FUNCTION__))
       "Val type must sized.")((!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&
 "Val type must sized.") ? static_cast<void> (0) : __assert_fail
 ("!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && \"Val type must sized.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2440, __PRETTY_FUNCTION__));
llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy);
if (ValTy == CastTy)
  return Val;
if (CGF.getContext().getTypeSizeInChars(ValTy) ==
    CGF.getContext().getTypeSizeInChars(CastTy))
  return CGF.Builder.CreateBitCast(Val, LLVMCastTy);
if (CastTy->isIntegerType() && ValTy->isIntegerType())
  return CGF.Builder.CreateIntCast(Val, LLVMCastTy,
                                   CastTy->hasSignedIntegerRepresentation());
Address CastItem = CGF.CreateMemTemp(CastTy);
Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
    CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()));
CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy,
                      LValueBaseInfo(AlignmentSource::Type),
                      TBAAAccessInfo());
return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc,
                            LValueBaseInfo(AlignmentSource::Type),
                            TBAAAccessInfo());
2459}

2461/// This function creates calls to one of two shuffle functions to copy
2462/// variables between lanes in a warp.
2463static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF,
                                               llvm::Value *Elem,
                                               QualType ElemType,
                                               llvm::Value *Offset,
                                               SourceLocation Loc) {
CodeGenModule &CGM = CGF.CGM;
CGBuilderTy &Bld = CGF.Builder;
CGOpenMPRuntimeGPU &RT =
    *(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime()));
llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder();

CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
assert(Size.getQuantity() <= 8 &&((Size.getQuantity() <= 8 && "Unsupported bitwidth in shuffle instruction."
) ? static_cast<void> (0) : __assert_fail ("Size.getQuantity() <= 8 && \"Unsupported bitwidth in shuffle instruction.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2476, __PRETTY_FUNCTION__))
       "Unsupported bitwidth in shuffle instruction.")((Size.getQuantity() <= 8 && "Unsupported bitwidth in shuffle instruction."
) ? static_cast<void> (0) : __assert_fail ("Size.getQuantity() <= 8 && \"Unsupported bitwidth in shuffle instruction.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 2476, __PRETTY_FUNCTION__));

RuntimeFunction ShuffleFn = Size.getQuantity() <= 4
                                ? OMPRTL___kmpc_shuffle_int32
                                : OMPRTL___kmpc_shuffle_int64;

// Cast all types to 32- or 64-bit values before calling shuffle routines.
QualType CastTy = CGF.getContext().getIntTypeForBitwidth(
    Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1);
llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc);
llvm::Value *WarpSize =
    Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true);

llvm::Value *ShuffledVal = CGF.EmitRuntimeCall(
    OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn),
    {ElemCast, Offset, WarpSize});

return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc);
2494}

2496static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr,
                          Address DestAddr, QualType ElemType,
                          llvm::Value *Offset, SourceLocation Loc) {
CGBuilderTy &Bld = CGF.Builder;

CharUnits Size = CGF.getContext().getTypeSizeInChars(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);
// ...
Address ElemPtr = DestAddr;
Address Ptr = SrcAddr;
Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast(
    Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy);
for (int IntSize = 8; IntSize >= 1; IntSize /= 2) {
  if (Size < CharUnits::fromQuantity(IntSize))
    continue;
  QualType IntType = CGF.getContext().getIntTypeForBitwidth(
      CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)),
      /*Signed=*/1);
  llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType);
  Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo());
  ElemPtr =
      Bld.CreatePointerBitCastOrAddrSpaceCast(ElemPtr, IntTy->getPointerTo());
  if (Size.getQuantity() / IntSize > 1) {
    llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond");
    llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then");
    llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit");
    llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock();
    CGF.EmitBlock(PreCondBB);
    llvm::PHINode *PhiSrc =
        Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2);
    PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB);
    llvm::PHINode *PhiDest =
        Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2);
    PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB);
    Ptr = Address(PhiSrc, Ptr.getAlignment());
    ElemPtr = Address(PhiDest, ElemPtr.getAlignment());
    llvm::Value *PtrDiff = Bld.CreatePtrDiff(
        PtrEnd.getPointer(), Bld.CreatePointerBitCastOrAddrSpaceCast(
                                 Ptr.getPointer(), CGF.VoidPtrTy));
    Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)),
                     ThenBB, ExitBB);
    CGF.EmitBlock(ThenBB);
    llvm::Value *Res = createRuntimeShuffleFunction(
        CGF,
        CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc,
                             LValueBaseInfo(AlignmentSource::Type),
                             TBAAAccessInfo()),
        IntType, Offset, Loc);
    CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType,
                          LValueBaseInfo(AlignmentSource::Type),
                          TBAAAccessInfo());
    Address LocalPtr = Bld.CreateConstGEP(Ptr, 1);
    Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
    PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB);
    PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB);
    CGF.EmitBranch(PreCondBB);
    CGF.EmitBlock(ExitBB);
  } else {
    llvm::Value *Res = createRuntimeShuffleFunction(
        CGF,
        CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc,
                             LValueBaseInfo(AlignmentSource::Type),
                             TBAAAccessInfo()),
        IntType, Offset, Loc);
    CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType,
                          LValueBaseInfo(AlignmentSource::Type),
                          TBAAAccessInfo());
    Ptr = Bld.CreateConstGEP(Ptr, 1);
    ElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
  }
  Size = Size % IntSize;
}
2576}

2578namespace {
2579enum CopyAction : unsigned {
// RemoteLaneToThread: Copy over a Reduce list from a remote lane in
// the warp using shuffle instructions.
RemoteLaneToThread,
// ThreadCopy: Make a copy of a Reduce list on the thread's stack.
ThreadCopy,
// ThreadToScratchpad: Copy a team-reduced array to the scratchpad.
ThreadToScratchpad,
// ScratchpadToThread: Copy from a scratchpad array in global memory
// containing team-reduced data to a thread's stack.
ScratchpadToThread,
2590};
2591} // namespace

2593struct CopyOptionsTy {
llvm::Value *RemoteLaneOffset;
llvm::Value *ScratchpadIndex;
llvm::Value *ScratchpadWidth;
2597};

2599/// Emit instructions to copy a Reduce list, which contains partially
2600/// aggregated values, in the specified direction.
2601static void emitReductionListCopy(
  CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy,
  ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase,
  CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) {

CodeGenModule &CGM = CGF.CGM;
ASTContext &C = CGM.getContext();
CGBuilderTy &Bld = CGF.Builder;

llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset;
llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex;
llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth;

// Iterates, element-by-element, through the source Reduce list and
// make a copy.
unsigned Idx = 0;
unsigned Size = Privates.size();
for (const Expr *Private : Privates) {
  Address SrcElementAddr = Address::invalid();
  Address DestElementAddr = Address::invalid();
  Address DestElementPtrAddr = Address::invalid();
  // 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;
  // Increment the src or dest pointer to the scratchpad, for each
  // new element.
  bool IncrScratchpadSrc = false;
  bool IncrScratchpadDest = false;

  switch (Action) {
  case RemoteLaneToThread: {
    // Step 1.1: Get the address for the src element in the Reduce list.
    Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
    SrcElementAddr = CGF.EmitLoadOfPointer(
        SrcElementPtrAddr,
        C.getPointerType(Private->getType())->castAs<PointerType>());

    // Step 1.2: Create a temporary to store the element in the destination
    // Reduce list.
    DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
    DestElementAddr =
        CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
    ShuffleInElement = true;
    UpdateDestListPtr = true;
    break;
  }
  case ThreadCopy: {
    // Step 1.1: Get the address for the src element in the Reduce list.
    Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
    SrcElementAddr = CGF.EmitLoadOfPointer(
        SrcElementPtrAddr,
        C.getPointerType(Private->getType())->castAs<PointerType>());

    // Step 1.2: Get the address for dest element.  The destination
    // element has already been created on the thread's stack.
    DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
    DestElementAddr = CGF.EmitLoadOfPointer(
        DestElementPtrAddr,
        C.getPointerType(Private->getType())->castAs<PointerType>());
    break;
  }
  case ThreadToScratchpad: {
    // Step 1.1: Get the address for the src element in the Reduce list.
    Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
    SrcElementAddr = CGF.EmitLoadOfPointer(
        SrcElementPtrAddr,
        C.getPointerType(Private->getType())->castAs<PointerType>());

    // Step 1.2: Get the address for dest element:
    // address = base + index * ElementSizeInChars.
    llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
    llvm::Value *CurrentOffset =
        Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
    llvm::Value *ScratchPadElemAbsolutePtrVal =
        Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset);
    ScratchPadElemAbsolutePtrVal =
        Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
    DestElementAddr = Address(ScratchPadElemAbsolutePtrVal,
                              C.getTypeAlignInChars(Private->getType()));
    IncrScratchpadDest = true;
    break;
  }
  case ScratchpadToThread: {
    // Step 1.1: Get the address for the src element in the scratchpad.
    // address = base + index * ElementSizeInChars.
    llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
    llvm::Value *CurrentOffset =
        Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
    llvm::Value *ScratchPadElemAbsolutePtrVal =
        Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset);
    ScratchPadElemAbsolutePtrVal =
        Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
    SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal,
                             C.getTypeAlignInChars(Private->getType()));
    IncrScratchpadSrc = true;

    // Step 1.2: Create a temporary to store the element in the destination
    // Reduce list.
    DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
    DestElementAddr =
        CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
    UpdateDestListPtr = true;
    break;
  }
  }

  // Regardless of src and dest of copy, we emit the load of src
  // element as this is required in all directions
  SrcElementAddr = Bld.CreateElementBitCast(
      SrcElementAddr, CGF.ConvertTypeForMem(Private->getType()));
  DestElementAddr = Bld.CreateElementBitCast(DestElementAddr,
                                             SrcElementAddr.getElementType());

  // Now that all active lanes have read the element in the
  // Reduce list, shuffle over the value from the remote lane.
  if (ShuffleInElement) {
    shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(),
                    RemoteLaneOffset, Private->getExprLoc());
  } else {
    switch (CGF.getEvaluationKind(Private->getType())) {
    case TEK_Scalar: {
      llvm::Value *Elem = CGF.EmitLoadOfScalar(
          SrcElementAddr, /*Volatile=*/false, Private->getType(),
          Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type),
          TBAAAccessInfo());
      // Store the source element value to the dest element address.
      CGF.EmitStoreOfScalar(
          Elem, DestElementAddr, /*Volatile=*/false, Private->getType(),
          LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
      break;
    }
    case TEK_Complex: {
      CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex(
          CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
          Private->getExprLoc());
      CGF.EmitStoreOfComplex(
          Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
          /*isInit=*/false);
      break;
    }
    case TEK_Aggregate:
      CGF.EmitAggregateCopy(
          CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
          CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
          Private->getType(), AggValueSlot::DoesNotOverlap);
      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) {
    CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast(
                              DestElementAddr.getPointer(), CGF.VoidPtrTy),
                          DestElementPtrAddr, /*Volatile=*/false,
                          C.VoidPtrTy);
  }

  // Step 4.1: Increment SrcBase/DestBase so that it points to the starting
  // address of the next element in scratchpad memory, unless we're currently
  // processing the last one.  Memory alignment is also taken care of here.
  if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) {
    llvm::Value *ScratchpadBasePtr =
        IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer();
    llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
    ScratchpadBasePtr = Bld.CreateNUWAdd(
        ScratchpadBasePtr,
        Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars));

    // Take care of global memory alignment for performance
    ScratchpadBasePtr = Bld.CreateNUWSub(
        ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
    ScratchpadBasePtr = Bld.CreateUDiv(
        ScratchpadBasePtr,
        llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
    ScratchpadBasePtr = Bld.CreateNUWAdd(
        ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
    ScratchpadBasePtr = Bld.CreateNUWMul(
        ScratchpadBasePtr,
        llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));

    if (IncrScratchpadDest)
      DestBase = Address(ScratchpadBasePtr, CGF.getPointerAlign());
    else /* IncrScratchpadSrc = true */
      SrcBase = Address(ScratchpadBasePtr, CGF.getPointerAlign());
  }

  ++Idx;
}
2795}

2797/// This function emits a helper that gathers Reduce lists from the first
2798/// lane of every active warp to lanes in the first warp.
2799///
2800/// void inter_warp_copy_func(void* reduce_data, num_warps)
2801///   shared smem[warp_size];
2802///   For all data entries D in reduce_data:
2803///     sync
2804///     If (I am the first lane in each warp)
2805///       Copy my local D to smem[warp_id]
2806///     sync
2807///     if (I am the first warp)
2808///       Copy smem[thread_id] to my local D
2809static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM,
                                            ArrayRef<const Expr *> Privates,
                                            QualType ReductionArrayTy,
                                            SourceLocation Loc) {
ASTContext &C = CGM.getContext();
llvm::Module &M = CGM.getModule();

// 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.
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                                C.VoidPtrTy, ImplicitParamDecl::Other);
// NumWarps: number of warps active in the parallel region.  This could
// be smaller than 32 (max warps in a CTA) for partial block reduction.
ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                              C.getIntTypeForBitwidth(32, /* Signed */ true),
                              ImplicitParamDecl::Other);
FunctionArgList Args;
Args.push_back(&ReduceListArg);
Args.push_back(&NumWarpsArg);

const CGFunctionInfo &CGFI =
    CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI),
                                  llvm::GlobalValue::InternalLinkage,
                                  "_omp_reduction_inter_warp_copy_func", &M);
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
Fn->setDoesNotRecurse();
CodeGenFunction CGF(CGM);
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);

CGBuilderTy &Bld = CGF.Builder;

// 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";
llvm::GlobalVariable *TransferMedium =
    M.getGlobalVariable(TransferMediumName);
unsigned WarpSize = CGF.getTarget().getGridValue(llvm::omp::GV_Warp_Size);
if (!TransferMedium) {
  auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize);
  unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared);
  TransferMedium = new llvm::GlobalVariable(
      M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage,
      llvm::UndefValue::get(Ty), TransferMediumName,
      /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal,
      SharedAddressSpace);
  CGM.addCompilerUsedGlobal(TransferMedium);
}

auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
// Get the CUDA thread id of the current OpenMP thread on the GPU.
llvm::Value *ThreadID = RT.getGPUThreadID(CGF);
// nvptx_lane_id = nvptx_id % warpsize
llvm::Value *LaneID = getNVPTXLaneID(CGF);
// nvptx_warp_id = nvptx_id / warpsize
llvm::Value *WarpID = getNVPTXWarpID(CGF);

Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
Address LocalReduceList(
    Bld.CreatePointerBitCastOrAddrSpaceCast(
        CGF.EmitLoadOfScalar(
            AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc,
            LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()),
        CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
    CGF.getPointerAlign());

unsigned Idx = 0;
for (const Expr *Private : Privates) {
  //
  // Warp master copies reduce element to transfer medium in __shared__
  // memory.
  //
  unsigned RealTySize =
      C.getTypeSizeInChars(Private->getType())
          .alignTo(C.getTypeAlignInChars(Private->getType()))
          .getQuantity();
  for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) {
    unsigned NumIters = RealTySize / TySize;
    if (NumIters == 0)
      continue;
    QualType CType = C.getIntTypeForBitwidth(
        C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1);
    llvm::Type *CopyType = CGF.ConvertTypeForMem(CType);
    CharUnits Align = CharUnits::fromQuantity(TySize);
    llvm::Value *Cnt = nullptr;
    Address CntAddr = Address::invalid();
    llvm::BasicBlock *PrecondBB = nullptr;
    llvm::BasicBlock *ExitBB = nullptr;
    if (NumIters > 1) {
      CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr");
      CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr,
                            /*Volatile=*/false, C.IntTy);
      PrecondBB = CGF.createBasicBlock("precond");
      ExitBB = CGF.createBasicBlock("exit");
      llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body");
      // There is no need to emit line number for unconditional branch.
      (void)ApplyDebugLocation::CreateEmpty(CGF);
      CGF.EmitBlock(PrecondBB);
      Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc);
      llvm::Value *Cmp =
          Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters));
      Bld.CreateCondBr(Cmp, BodyBB, ExitBB);
      CGF.EmitBlock(BodyBB);
    }
    // kmpc_barrier.
    CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
                                           /*EmitChecks=*/false,
                                           /*ForceSimpleCall=*/true);
    llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
    llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
    llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");

    // if (lane_id == 0)
    llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master");
    Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB);
    CGF.EmitBlock(ThenBB);

    // Reduce element = LocalReduceList[i]
    Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
    llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
        ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
    // elemptr = ((CopyType*)(elemptrptr)) + I
    Address ElemPtr = Address(ElemPtrPtr, Align);
    ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType);
    if (NumIters > 1) {
      ElemPtr = Address(Bld.CreateGEP(ElemPtr.getPointer(), Cnt),
                        ElemPtr.getAlignment());
    }

    // Get pointer to location in transfer medium.
    // MediumPtr = &medium[warp_id]
    llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP(
        TransferMedium, {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID});
    Address MediumPtr(MediumPtrVal, Align);
    // Casting to actual data type.
    // MediumPtr = (CopyType*)MediumPtrAddr;
    MediumPtr = Bld.CreateElementBitCast(MediumPtr, CopyType);

    // elem = *elemptr
    //*MediumPtr = elem
    llvm::Value *Elem = CGF.EmitLoadOfScalar(
        ElemPtr, /*Volatile=*/false, CType, Loc,
        LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
    // Store the source element value to the dest element address.
    CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType,
                          LValueBaseInfo(AlignmentSource::Type),
                          TBAAAccessInfo());

    Bld.CreateBr(MergeBB);

    CGF.EmitBlock(ElseBB);
    Bld.CreateBr(MergeBB);

    CGF.EmitBlock(MergeBB);

    // kmpc_barrier.
    CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
                                           /*EmitChecks=*/false,
                                           /*ForceSimpleCall=*/true);

    //
    // Warp 0 copies reduce element from transfer medium.
    //
    llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then");
    llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else");
    llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont");

    Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg);
    llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar(
        AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc);

    // Up to 32 threads in warp 0 are active.
    llvm::Value *IsActiveThread =
        Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread");
    Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB);

    CGF.EmitBlock(W0ThenBB);

    // SrcMediumPtr = &medium[tid]
    llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP(
        TransferMedium,
        {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID});
    Address SrcMediumPtr(SrcMediumPtrVal, Align);
    // SrcMediumVal = *SrcMediumPtr;
    SrcMediumPtr = Bld.CreateElementBitCast(SrcMediumPtr, CopyType);

    // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I
    Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
    llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar(
        TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc);
    Address TargetElemPtr = Address(TargetElemPtrVal, Align);
    TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType);
    if (NumIters > 1) {
      TargetElemPtr = Address(Bld.CreateGEP(TargetElemPtr.getPointer(), Cnt),
                              TargetElemPtr.getAlignment());
    }

    // *TargetElemPtr = SrcMediumVal;
    llvm::Value *SrcMediumValue =
        CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc);
    CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false,
                          CType);
    Bld.CreateBr(W0MergeBB);

    CGF.EmitBlock(W0ElseBB);
    Bld.CreateBr(W0MergeBB);

    CGF.EmitBlock(W0MergeBB);

    if (NumIters > 1) {
      Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1));
      CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy);
      CGF.EmitBranch(PrecondBB);
      (void)ApplyDebugLocation::CreateEmpty(CGF);
      CGF.EmitBlock(ExitBB);
    }
    RealTySize %= TySize;
  }
  ++Idx;
}

CGF.FinishFunction();
return Fn;
3039}

3041/// Emit a helper that reduces data across two OpenMP threads (lanes)
3042/// in the same warp.  It uses shuffle instructions to copy over data from
3043/// a remote lane's stack.  The reduction algorithm performed is specified
3044/// by the fourth parameter.
3045///
3046/// Algorithm Versions.
3047/// Full Warp Reduce (argument value 0):
3048///   This algorithm assumes that all 32 lanes are active and gathers
3049///   data from these 32 lanes, producing a single resultant value.
3050/// Contiguous Partial Warp Reduce (argument value 1):
3051///   This algorithm assumes that only a *contiguous* subset of lanes
3052///   are active.  This happens for the last warp in a parallel region
3053///   when the user specified num_threads is not an integer multiple of
3054///   32.  This contiguous subset always starts with the zeroth lane.
3055/// Partial Warp Reduce (argument value 2):
3056///   This algorithm gathers data from any number of lanes at any position.
3057/// All reduced values are stored in the lowest possible lane.  The set
3058/// of problems every algorithm addresses is a super set of those
3059/// addressable by algorithms with a lower version number.  Overhead
3060/// increases as algorithm version increases.
3061///
3062/// Terminology
3063/// Reduce element:
3064///   Reduce element refers to the individual data field with primitive
3065///   data types to be combined and reduced across threads.
3066/// Reduce list:
3067///   Reduce list refers to a collection of local, thread-private
3068///   reduce elements.
3069/// Remote Reduce list:
3070///   Remote Reduce list refers to a collection of remote (relative to
3071///   the current thread) reduce elements.
3072///
3073/// We distinguish between three states of threads that are important to
3074/// the implementation of this function.
3075/// Alive threads:
3076///   Threads in a warp executing the SIMT instruction, as distinguished from
3077///   threads that are inactive due to divergent control flow.
3078/// Active threads:
3079///   The minimal set of threads that has to be alive upon entry to this
3080///   function.  The computation is correct iff active threads are alive.
3081///   Some threads are alive but they are not active because they do not
3082///   contribute to the computation in any useful manner.  Turning them off
3083///   may introduce control flow overheads without any tangible benefits.
3084/// Effective threads:
3085///   In order to comply with the argument requirements of the shuffle
3086///   function, we must keep all lanes holding data alive.  But at most
3087///   half of them perform value aggregation; we refer to this half of
3088///   threads as effective. The other half is simply handing off their
3089///   data.
3090///
3091/// Procedure
3092/// Value shuffle:
3093///   In this step active threads transfer data from higher lane positions
3094///   in the warp to lower lane positions, creating Remote Reduce list.
3095/// Value aggregation:
3096///   In this step, effective threads combine their thread local Reduce list
3097///   with Remote Reduce list and store the result in the thread local
3098///   Reduce list.
3099/// Value copy:
3100///   In this step, we deal with the assumption made by algorithm 2
3101///   (i.e. contiguity assumption).  When we have an odd number of lanes
3102///   active, say 2k+1, only k threads will be effective and therefore k
3103///   new values will be produced.  However, the Reduce list owned by the
3104///   (2k+1)th thread is ignored in the value aggregation.  Therefore
3105///   we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so
3106///   that the contiguity assumption still holds.
3107static llvm::Function *emitShuffleAndReduceFunction(
  CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
  QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) {
ASTContext &C = CGM.getContext();

// Thread local Reduce list used to host the values of data to be reduced.
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                                C.VoidPtrTy, ImplicitParamDecl::Other);
// Current lane id; could be logical.
ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy,
                            ImplicitParamDecl::Other);
// Offset of the remote source lane relative to the current lane.
ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                                      C.ShortTy, ImplicitParamDecl::Other);
// Algorithm version.  This is expected to be known at compile time.
ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                             C.ShortTy, ImplicitParamDecl::Other);
FunctionArgList Args;
Args.push_back(&ReduceListArg);
Args.push_back(&LaneIDArg);
Args.push_back(&RemoteLaneOffsetArg);
Args.push_back(&AlgoVerArg);

const CGFunctionInfo &CGFI =
    CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
auto *Fn = llvm::Function::Create(
    CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
    "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule());
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
Fn->setDoesNotRecurse();
if (CGM.getLangOpts().Optimize) {
  Fn->removeFnAttr(llvm::Attribute::NoInline);
  Fn->removeFnAttr(llvm::Attribute::OptimizeNone);
  Fn->addFnAttr(llvm::Attribute::AlwaysInline);
}

CodeGenFunction CGF(CGM);
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);

CGBuilderTy &Bld = CGF.Builder;

Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
Address LocalReduceList(
    Bld.CreatePointerBitCastOrAddrSpaceCast(
        CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
                             C.VoidPtrTy, SourceLocation()),
        CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
    CGF.getPointerAlign());

Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg);
llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar(
    AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation());

Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg);
llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar(
    AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation());

Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg);
llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar(
    AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation());

// Create a local thread-private variable to host the Reduce list
// from a remote lane.
Address RemoteReduceList =
    CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list");

// 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.
emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates,
                      LocalReduceList, RemoteReduceList,
                      {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal,
                       /*ScratchpadIndex=*/nullptr,
                       /*ScratchpadWidth=*/nullptr});

// 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.
llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal);

llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
llvm::Value *CondAlgo1 = Bld.CreateAnd(
    Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal));

llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2));
llvm::Value *CondAlgo2 = Bld.CreateAnd(
    Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1))));
CondAlgo2 = Bld.CreateAnd(
    CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0)));

llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1);
CondReduce = Bld.CreateOr(CondReduce, CondAlgo2);

llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
Bld.CreateCondBr(CondReduce, ThenBB, ElseBB);

CGF.EmitBlock(ThenBB);
// reduce_function(LocalReduceList, RemoteReduceList)
llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
    LocalReduceList.getPointer(), CGF.VoidPtrTy);
llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
    RemoteReduceList.getPointer(), CGF.VoidPtrTy);
CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
    CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr});
Bld.CreateBr(MergeBB);

CGF.EmitBlock(ElseBB);
Bld.CreateBr(MergeBB);

CGF.EmitBlock(MergeBB);

// if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local
// Reduce list.
Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
llvm::Value *CondCopy = Bld.CreateAnd(
    Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal));

llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then");
llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else");
llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont");
Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB);

CGF.EmitBlock(CpyThenBB);
emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates,
                      RemoteReduceList, LocalReduceList);
Bld.CreateBr(CpyMergeBB);

CGF.EmitBlock(CpyElseBB);
Bld.CreateBr(CpyMergeBB);

CGF.EmitBlock(CpyMergeBB);

CGF.FinishFunction();
return Fn;
3261}

3263/// This function emits a helper that copies all the reduction variables from
3264/// the team into the provided global buffer for the reduction variables.
3265///
3266/// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
3267///   For all data entries D in reduce_data:
3268///     Copy local D to buffer.D[Idx]
3269static llvm::Value *emitListToGlobalCopyFunction(
  CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
  QualType ReductionArrayTy, SourceLocation Loc,
  const RecordDecl *TeamReductionRec,
  const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
      &VarFieldMap) {
ASTContext &C = CGM.getContext();

// Buffer: global reduction buffer.
ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                            C.VoidPtrTy, ImplicitParamDecl::Other);
// Idx: index of the buffer.
ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
                         ImplicitParamDecl::Other);
// ReduceList: thread local Reduce list.
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                                C.VoidPtrTy, ImplicitParamDecl::Other);
FunctionArgList Args;
Args.push_back(&BufferArg);
Args.push_back(&IdxArg);
Args.push_back(&ReduceListArg);

const CGFunctionInfo &CGFI =
    CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
auto *Fn = llvm::Function::Create(
    CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
    "_omp_reduction_list_to_global_copy_func", &CGM.getModule());
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
Fn->setDoesNotRecurse();
CodeGenFunction CGF(CGM);
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);

CGBuilderTy &Bld = CGF.Builder;

Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
Address LocalReduceList(
    Bld.CreatePointerBitCastOrAddrSpaceCast(
        CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
                             C.VoidPtrTy, Loc),
        CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
    CGF.getPointerAlign());
QualType StaticTy = C.getRecordType(TeamReductionRec);
llvm::Type *LLVMReductionsBufferTy =
    CGM.getTypes().ConvertTypeForMem(StaticTy);
llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
    CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
    LLVMReductionsBufferTy->getPointerTo());
llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
                       CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
                                            /*Volatile=*/false, C.IntTy,
                                            Loc)};
unsigned Idx = 0;
for (const Expr *Private : Privates) {
  // Reduce element = LocalReduceList[i]
  Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
  llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
      ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
  // elemptr = ((CopyType*)(elemptrptr)) + I
  ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
      ElemPtrPtr, CGF.ConvertTypeForMem(Private->getType())->getPointerTo());
  Address ElemPtr =
      Address(ElemPtrPtr, C.getTypeAlignInChars(Private->getType()));
  const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
  // Global = Buffer.VD[Idx];
  const FieldDecl *FD = VarFieldMap.lookup(VD);
  LValue GlobLVal = CGF.EmitLValueForField(
      CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
  llvm::Value *BufferPtr =
      Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs);
  GlobLVal.setAddress(Address(BufferPtr, GlobLVal.getAlignment()));
  switch (CGF.getEvaluationKind(Private->getType())) {
  case TEK_Scalar: {
    llvm::Value *V = CGF.EmitLoadOfScalar(
        ElemPtr, /*Volatile=*/false, Private->getType(), Loc,
        LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
    CGF.EmitStoreOfScalar(V, GlobLVal);
    break;
  }
  case TEK_Complex: {
    CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(
        CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc);
    CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false);
    break;
  }
  case TEK_Aggregate:
    CGF.EmitAggregateCopy(GlobLVal,
                          CGF.MakeAddrLValue(ElemPtr, Private->getType()),
                          Private->getType(), AggValueSlot::DoesNotOverlap);
    break;
  }
  ++Idx;
}

CGF.FinishFunction();
return Fn;
3365}

3367/// This function emits a helper that reduces all the reduction variables from
3368/// the team into the provided global buffer for the reduction variables.
3369///
3370/// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data)
3371///  void *GlobPtrs[];
3372///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
3373///  ...
3374///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
3375///  reduce_function(GlobPtrs, reduce_data);
3376static llvm::Value *emitListToGlobalReduceFunction(
  CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
  QualType ReductionArrayTy, SourceLocation Loc,
  const RecordDecl *TeamReductionRec,
  const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
      &VarFieldMap,
  llvm::Function *ReduceFn) {
ASTContext &C = CGM.getContext();

// Buffer: global reduction buffer.
ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                            C.VoidPtrTy, ImplicitParamDecl::Other);
// Idx: index of the buffer.
ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
                         ImplicitParamDecl::Other);
// ReduceList: thread local Reduce list.
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                                C.VoidPtrTy, ImplicitParamDecl::Other);
FunctionArgList Args;
Args.push_back(&BufferArg);
Args.push_back(&IdxArg);
Args.push_back(&ReduceListArg);

const CGFunctionInfo &CGFI =
    CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
auto *Fn = llvm::Function::Create(
    CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
    "_omp_reduction_list_to_global_reduce_func", &CGM.getModule());
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
Fn->setDoesNotRecurse();
CodeGenFunction CGF(CGM);
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);

CGBuilderTy &Bld = CGF.Builder;

Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
QualType StaticTy = C.getRecordType(TeamReductionRec);
llvm::Type *LLVMReductionsBufferTy =
    CGM.getTypes().ConvertTypeForMem(StaticTy);
llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
    CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
    LLVMReductionsBufferTy->getPointerTo());

// 1. Build a list of reduction variables.
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
Address ReductionList =
    CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
auto IPriv = Privates.begin();
llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
                       CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
                                            /*Volatile=*/false, C.IntTy,
                                            Loc)};
unsigned Idx = 0;
for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
  Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
  // Global = Buffer.VD[Idx];
  const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
  const FieldDecl *FD = VarFieldMap.lookup(VD);
  LValue GlobLVal = CGF.EmitLValueForField(
      CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
  llvm::Value *BufferPtr =
      Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs);
  llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr);
  CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy);
  if ((*IPriv)->getType()->isVariablyModifiedType()) {
    // Store array size.
    ++Idx;
    Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
    llvm::Value *Size = CGF.Builder.CreateIntCast(
        CGF.getVLASize(
               CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
            .NumElts,
        CGF.SizeTy, /*isSigned=*/false);
    CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
                            Elem);
  }
}

// Call reduce_function(GlobalReduceList, ReduceList)
llvm::Value *GlobalReduceList =
    CGF.EmitCastToVoidPtr(ReductionList.getPointer());
Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
    AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
    CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr});
CGF.FinishFunction();
return Fn;
3464}

3466/// This function emits a helper that copies all the reduction variables from
3467/// the team into the provided global buffer for the reduction variables.
3468///
3469/// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
3470///   For all data entries D in reduce_data:
3471///     Copy buffer.D[Idx] to local D;
3472static llvm::Value *emitGlobalToListCopyFunction(
  CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
  QualType ReductionArrayTy, SourceLocation Loc,
  const RecordDecl *TeamReductionRec,
  const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
      &VarFieldMap) {
ASTContext &C = CGM.getContext();

// Buffer: global reduction buffer.
ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                            C.VoidPtrTy, ImplicitParamDecl::Other);
// Idx: index of the buffer.
ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
                         ImplicitParamDecl::Other);
// ReduceList: thread local Reduce list.
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                                C.VoidPtrTy, ImplicitParamDecl::Other);
FunctionArgList Args;
Args.push_back(&BufferArg);
Args.push_back(&IdxArg);
Args.push_back(&ReduceListArg);

const CGFunctionInfo &CGFI =
    CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
auto *Fn = llvm::Function::Create(
    CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
    "_omp_reduction_global_to_list_copy_func", &CGM.getModule());
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
Fn->setDoesNotRecurse();
CodeGenFunction CGF(CGM);
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);

CGBuilderTy &Bld = CGF.Builder;

Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
Address LocalReduceList(
    Bld.CreatePointerBitCastOrAddrSpaceCast(
        CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
                             C.VoidPtrTy, Loc),
        CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
    CGF.getPointerAlign());
QualType StaticTy = C.getRecordType(TeamReductionRec);
llvm::Type *LLVMReductionsBufferTy =
    CGM.getTypes().ConvertTypeForMem(StaticTy);
llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
    CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
    LLVMReductionsBufferTy->getPointerTo());

llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
                       CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
                                            /*Volatile=*/false, C.IntTy,
                                            Loc)};
unsigned Idx = 0;
for (const Expr *Private : Privates) {
  // Reduce element = LocalReduceList[i]
  Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
  llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
      ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
  // elemptr = ((CopyType*)(elemptrptr)) + I
  ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
      ElemPtrPtr, CGF.ConvertTypeForMem(Private->getType())->getPointerTo());
  Address ElemPtr =
      Address(ElemPtrPtr, C.getTypeAlignInChars(Private->getType()));
  const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
  // Global = Buffer.VD[Idx];
  const FieldDecl *FD = VarFieldMap.lookup(VD);
  LValue GlobLVal = CGF.EmitLValueForField(
      CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
  llvm::Value *BufferPtr =
      Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs);
  GlobLVal.setAddress(Address(BufferPtr, GlobLVal.getAlignment()));
  switch (CGF.getEvaluationKind(Private->getType())) {
  case TEK_Scalar: {
    llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc);
    CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(),
                          LValueBaseInfo(AlignmentSource::Type),
                          TBAAAccessInfo());
    break;
  }
  case TEK_Complex: {
    CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc);
    CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()),
                           /*isInit=*/false);
    break;
  }
  case TEK_Aggregate:
    CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()),
                          GlobLVal, Private->getType(),
                          AggValueSlot::DoesNotOverlap);
    break;
  }
  ++Idx;
}

CGF.FinishFunction();
return Fn;
3569}

3571/// This function emits a helper that reduces all the reduction variables from
3572/// the team into the provided global buffer for the reduction variables.
3573///
3574/// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data)
3575///  void *GlobPtrs[];
3576///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
3577///  ...
3578///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
3579///  reduce_function(reduce_data, GlobPtrs);
3580static llvm::Value *emitGlobalToListReduceFunction(
  CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
  QualType ReductionArrayTy, SourceLocation Loc,
  const RecordDecl *TeamReductionRec,
  const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
      &VarFieldMap,
  llvm::Function *ReduceFn) {
ASTContext &C = CGM.getContext();

// Buffer: global reduction buffer.
ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                            C.VoidPtrTy, ImplicitParamDecl::Other);
// Idx: index of the buffer.
ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
                         ImplicitParamDecl::Other);
// ReduceList: thread local Reduce list.
ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
                                C.VoidPtrTy, ImplicitParamDecl::Other);
FunctionArgList Args;
Args.push_back(&BufferArg);
Args.push_back(&IdxArg);
Args.push_back(&ReduceListArg);

const CGFunctionInfo &CGFI =
    CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
auto *Fn = llvm::Function::Create(
    CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
    "_omp_reduction_global_to_list_reduce_func", &CGM.getModule());
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
Fn->setDoesNotRecurse();
CodeGenFunction CGF(CGM);
CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);

CGBuilderTy &Bld = CGF.Builder;

Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
QualType StaticTy = C.getRecordType(TeamReductionRec);
llvm::Type *LLVMReductionsBufferTy =
    CGM.getTypes().ConvertTypeForMem(StaticTy);
llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
    CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
    LLVMReductionsBufferTy->getPointerTo());

// 1. Build a list of reduction variables.
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
Address ReductionList =
    CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
auto IPriv = Privates.begin();
llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
                       CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
                                            /*Volatile=*/false, C.IntTy,
                                            Loc)};
unsigned Idx = 0;
for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
  Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
  // Global = Buffer.VD[Idx];
  const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
  const FieldDecl *FD = VarFieldMap.lookup(VD);
  LValue GlobLVal = CGF.EmitLValueForField(
      CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
  llvm::Value *BufferPtr =
      Bld.CreateInBoundsGEP(GlobLVal.getPointer(CGF), Idxs);
  llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr);
  CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy);
  if ((*IPriv)->getType()->isVariablyModifiedType()) {
    // Store array size.
    ++Idx;
    Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
    llvm::Value *Size = CGF.Builder.CreateIntCast(
        CGF.getVLASize(
               CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
            .NumElts,
        CGF.SizeTy, /*isSigned=*/false);
    CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
                            Elem);
  }
}

// Call reduce_function(ReduceList, GlobalReduceList)
llvm::Value *GlobalReduceList =
    CGF.EmitCastToVoidPtr(ReductionList.getPointer());
Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
    AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
    CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList});
CGF.FinishFunction();
return Fn;
3668}

3670///
3671/// Design of OpenMP reductions on the GPU
3672///
3673/// Consider a typical OpenMP program with one or more reduction
3674/// clauses:
3675///
3676/// float foo;
3677/// double bar;
3678/// #pragma omp target teams distribute parallel for \
3679///             reduction(+:foo) reduction(*:bar)
3680/// for (int i = 0; i < N; i++) {
3681///   foo += A[i]; bar *= B[i];
3682/// }
3683///
3684/// where 'foo' and 'bar' are reduced across all OpenMP threads in
3685/// all teams.  In our OpenMP implementation on the NVPTX device an
3686/// OpenMP team is mapped to a CUDA threadblock and OpenMP threads
3687/// within a team are mapped to CUDA threads within a threadblock.
3688/// Our goal is to efficiently aggregate values across all OpenMP
3689/// threads such that:
3690///
3691///   - the compiler and runtime are logically concise, and
3692///   - the reduction is performed efficiently in a hierarchical
3693///     manner as follows: within OpenMP threads in the same warp,
3694///     across warps in a threadblock, and finally across teams on
3695///     the NVPTX device.
3696///
3697/// Introduction to Decoupling
3698///
3699/// We would like to decouple the compiler and the runtime so that the
3700/// latter is ignorant of the reduction variables (number, data types)
3701/// and the reduction operators.  This allows a simpler interface
3702/// and implementation while still attaining good performance.
3703///
3704/// Pseudocode for the aforementioned OpenMP program generated by the
3705/// compiler is as follows:
3706///
3707/// 1. Create private copies of reduction variables on each OpenMP
3708///    thread: 'foo_private', 'bar_private'
3709/// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned
3710///    to it and writes the result in 'foo_private' and 'bar_private'
3711///    respectively.
3712/// 3. Call the OpenMP runtime on the GPU to reduce within a team
3713///    and store the result on the team master:
3714///
3715///     __kmpc_nvptx_parallel_reduce_nowait_v2(...,
3716///        reduceData, shuffleReduceFn, interWarpCpyFn)
3717///
3718///     where:
3719///       struct ReduceData {
3720///         double *foo;
3721///         double *bar;
3722///       } reduceData
3723///       reduceData.foo = &foo_private
3724///       reduceData.bar = &bar_private
3725///
3726///     'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two
3727///     auxiliary functions generated by the compiler that operate on
3728///     variables of type 'ReduceData'.  They aid the runtime perform
3729///     algorithmic steps in a data agnostic manner.
3730///
3731///     'shuffleReduceFn' is a pointer to a function that reduces data
3732///     of type 'ReduceData' across two OpenMP threads (lanes) in the
3733///     same warp.  It takes the following arguments as input:
3734///
3735///     a. variable of type 'ReduceData' on the calling lane,
3736///     b. its lane_id,
3737///     c. an offset relative to the current lane_id to generate a
3738///        remote_lane_id.  The remote lane contains the second
3739///        variable of type 'ReduceData' that is to be reduced.
3740///     d. an algorithm version parameter determining which reduction
3741///        algorithm to use.
3742///
3743///     'shuffleReduceFn' retrieves data from the remote lane using
3744///     efficient GPU shuffle intrinsics and reduces, using the
3745///     algorithm specified by the 4th parameter, the two operands
3746///     element-wise.  The result is written to the first operand.
3747///
3748///     Different reduction algorithms are implemented in different
3749///     runtime functions, all calling 'shuffleReduceFn' to perform
3750///     the essential reduction step.  Therefore, based on the 4th
3751///     parameter, this function behaves slightly differently to
3752///     cooperate with the runtime to ensure correctness under
3753///     different circumstances.
3754///
3755///     'InterWarpCpyFn' is a pointer to a function that transfers
3756///     reduced variables across warps.  It tunnels, through CUDA
3757///     shared memory, the thread-private data of type 'ReduceData'
3758///     from lane 0 of each warp to a lane in the first warp.
3759/// 4. Call the OpenMP runtime on the GPU to reduce across teams.
3760///    The last team writes the global reduced value to memory.
3761///
3762///     ret = __kmpc_nvptx_teams_reduce_nowait(...,
3763///             reduceData, shuffleReduceFn, interWarpCpyFn,
3764///             scratchpadCopyFn, loadAndReduceFn)
3765///
3766///     'scratchpadCopyFn' is a helper that stores reduced
3767///     data from the team master to a scratchpad array in
3768///     global memory.
3769///
3770///     'loadAndReduceFn' is a helper that loads data from
3771///     the scratchpad array and reduces it with the input
3772///     operand.
3773///
3774///     These compiler generated functions hide address
3775///     calculation and alignment information from the runtime.
3776/// 5. if ret == 1:
3777///     The team master of the last team stores the reduced
3778///     result to the globals in memory.
3779///     foo += reduceData.foo; bar *= reduceData.bar
3780///
3781///
3782/// Warp Reduction Algorithms
3783///
3784/// On the warp level, we have three algorithms implemented in the
3785/// OpenMP runtime depending on the number of active lanes:
3786///
3787/// Full Warp Reduction
3788///
3789/// The reduce algorithm within a warp where all lanes are active
3790/// is implemented in the runtime as follows:
3791///
3792/// full_warp_reduce(void *reduce_data,
3793///                  kmp_ShuffleReductFctPtr ShuffleReduceFn) {
3794///   for (int offset = WARPSIZE/2; offset > 0; offset /= 2)
3795///     ShuffleReduceFn(reduce_data, 0, offset, 0);
3796/// }
3797///
3798/// The algorithm completes in log(2, WARPSIZE) steps.
3799///
3800/// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is
3801/// not used therefore we save instructions by not retrieving lane_id
3802/// from the corresponding special registers.  The 4th parameter, which
3803/// represents the version of the algorithm being used, is set to 0 to
3804/// signify full warp reduction.
3805///
3806/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
3807///
3808/// #reduce_elem refers to an element in the local lane's data structure
3809/// #remote_elem is retrieved from a remote lane
3810/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
3811/// reduce_elem = reduce_elem REDUCE_OP remote_elem;
3812///
3813/// Contiguous Partial Warp Reduction
3814///
3815/// This reduce algorithm is used within a warp where only the first
3816/// 'n' (n <= WARPSIZE) lanes are active.  It is typically used when the
3817/// number of OpenMP threads in a parallel region is not a multiple of
3818/// WARPSIZE.  The algorithm is implemented in the runtime as follows:
3819///
3820/// void
3821/// contiguous_partial_reduce(void *reduce_data,
3822///                           kmp_ShuffleReductFctPtr ShuffleReduceFn,
3823///                           int size, int lane_id) {
3824///   int curr_size;
3825///   int offset;
3826///   curr_size = size;
3827///   mask = curr_size/2;
3828///   while (offset>0) {
3829///     ShuffleReduceFn(reduce_data, lane_id, offset, 1);
3830///     curr_size = (curr_size+1)/2;
3831///     offset = curr_size/2;
3832///   }
3833/// }
3834///
3835/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
3836///
3837/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
3838/// if (lane_id < offset)
3839///     reduce_elem = reduce_elem REDUCE_OP remote_elem
3840/// else
3841///     reduce_elem = remote_elem
3842///
3843/// This algorithm assumes that the data to be reduced are located in a
3844/// contiguous subset of lanes starting from the first.  When there is
3845/// an odd number of active lanes, the data in the last lane is not
3846/// aggregated with any other lane's dat but is instead copied over.
3847///
3848/// Dispersed Partial Warp Reduction
3849///
3850/// This algorithm is used within a warp when any discontiguous subset of
3851/// lanes are active.  It is used to implement the reduction operation
3852/// across lanes in an OpenMP simd region or in a nested parallel region.
3853///
3854/// void
3855/// dispersed_partial_reduce(void *reduce_data,
3856///                          kmp_ShuffleReductFctPtr ShuffleReduceFn) {
3857///   int size, remote_id;
3858///   int logical_lane_id = number_of_active_lanes_before_me() * 2;
3859///   do {
3860///       remote_id = next_active_lane_id_right_after_me();
3861///       # the above function returns 0 of no active lane
3862///       # is present right after the current lane.
3863///       size = number_of_active_lanes_in_this_warp();
3864///       logical_lane_id /= 2;
3865///       ShuffleReduceFn(reduce_data, logical_lane_id,
3866///                       remote_id-1-threadIdx.x, 2);
3867///   } while (logical_lane_id % 2 == 0 && size > 1);
3868/// }
3869///
3870/// There is no assumption made about the initial state of the reduction.
3871/// Any number of lanes (>=1) could be active at any position.  The reduction
3872/// result is returned in the first active lane.
3873///
3874/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
3875///
3876/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
3877/// if (lane_id % 2 == 0 && offset > 0)
3878///     reduce_elem = reduce_elem REDUCE_OP remote_elem
3879/// else
3880///     reduce_elem = remote_elem
3881///
3882///
3883/// Intra-Team Reduction
3884///
3885/// This function, as implemented in the runtime call
3886/// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP
3887/// threads in a team.  It first reduces within a warp using the
3888/// aforementioned algorithms.  We then proceed to gather all such
3889/// reduced values at the first warp.
3890///
3891/// The runtime makes use of the function 'InterWarpCpyFn', which copies
3892/// data from each of the "warp master" (zeroth lane of each warp, where
3893/// warp-reduced data is held) to the zeroth warp.  This step reduces (in
3894/// a mathematical sense) the problem of reduction across warp masters in
3895/// a block to the problem of warp reduction.
3896///
3897///
3898/// Inter-Team Reduction
3899///
3900/// Once a team has reduced its data to a single value, it is stored in
3901/// a global scratchpad array.  Since each team has a distinct slot, this
3902/// can be done without locking.
3903///
3904/// The last team to write to the scratchpad array proceeds to reduce the
3905/// scratchpad array.  One or more workers in the last team use the helper
3906/// 'loadAndReduceDataFn' to load and reduce values from the array, i.e.,
3907/// the k'th worker reduces every k'th element.
3908///
3909/// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to
3910/// reduce across workers and compute a globally reduced value.
3911///
3912void CGOpenMPRuntimeGPU::emitReduction(
  CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates,
  ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
  ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) {
if (!CGF.HaveInsertPoint())
  return;

bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind);
3920#ifndef NDEBUG
bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind);
3922#endif

if (Options.SimpleReduction) {
  assert(!TeamsReduction && !ParallelReduction &&((!TeamsReduction && !ParallelReduction && "Invalid reduction selection in emitReduction."
) ? static_cast<void> (0) : __assert_fail ("!TeamsReduction && !ParallelReduction && \"Invalid reduction selection in emitReduction.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 3926, __PRETTY_FUNCTION__))
         "Invalid reduction selection in emitReduction.")((!TeamsReduction && !ParallelReduction && "Invalid reduction selection in emitReduction."
) ? static_cast<void> (0) : __assert_fail ("!TeamsReduction && !ParallelReduction && \"Invalid reduction selection in emitReduction.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 3926, __PRETTY_FUNCTION__));
  CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs,
                                 ReductionOps, Options);
  return;
}

assert((TeamsReduction || ParallelReduction) &&(((TeamsReduction || ParallelReduction) && "Invalid reduction selection in emitReduction."
) ? static_cast<void> (0) : __assert_fail ("(TeamsReduction || ParallelReduction) && \"Invalid reduction selection in emitReduction.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 3933, __PRETTY_FUNCTION__))
       "Invalid reduction selection in emitReduction.")(((TeamsReduction || ParallelReduction) && "Invalid reduction selection in emitReduction."
) ? static_cast<void> (0) : __assert_fail ("(TeamsReduction || ParallelReduction) && \"Invalid reduction selection in emitReduction.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 3933, __PRETTY_FUNCTION__));

// 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>);
llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
llvm::Value *ThreadId = getThreadID(CGF, Loc);

llvm::Value *Res;
ASTContext &C = CGM.getContext();
// 1. Build a list of reduction variables.
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
auto Size = RHSExprs.size();
for (const Expr *E : Privates) {
  if (E->getType()->isVariablyModifiedType())
    // Reserve place for array size.
    ++Size;
}
llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size);
QualType ReductionArrayTy =
    C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal,
                           /*IndexTypeQuals=*/0);
Address ReductionList =
    CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
auto IPriv = Privates.begin();
unsigned Idx = 0;
for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) {
  Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
  CGF.Builder.CreateStore(
      CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
          CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy),
      Elem);
  if ((*IPriv)->getType()->isVariablyModifiedType()) {
    // Store array size.
    ++Idx;
    Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
    llvm::Value *Size = CGF.Builder.CreateIntCast(
        CGF.getVLASize(
               CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
            .NumElts,
        CGF.SizeTy, /*isSigned=*/false);
    CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
                            Elem);
  }
}

llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
    ReductionList.getPointer(), CGF.VoidPtrTy);
llvm::Function *ReductionFn = emitReductionFunction(
    Loc, CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo(), Privates,
    LHSExprs, RHSExprs, ReductionOps);
llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy);
llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction(
    CGM, Privates, ReductionArrayTy, ReductionFn, Loc);
llvm::Value *InterWarpCopyFn =
    emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc);

if (ParallelReduction) {
  llvm::Value *Args[] = {RTLoc,
                         ThreadId,
                         CGF.Builder.getInt32(RHSExprs.size()),
                         ReductionArrayTySize,
                         RL,
                         ShuffleAndReduceFn,
                         InterWarpCopyFn};

  Res = CGF.EmitRuntimeCall(
      OMPBuilder.getOrCreateRuntimeFunction(
          CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2),
      Args);
} else {
  assert(TeamsReduction && "expected teams reduction.")((TeamsReduction && "expected teams reduction.") ? static_cast
<void> (0) : __assert_fail ("TeamsReduction && \"expected teams reduction.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4005, __PRETTY_FUNCTION__));
  llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap;
  llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size());
  int Cnt = 0;
  for (const Expr *DRE : Privates) {
    PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl();
    ++Cnt;
  }
  const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars(
      CGM.getContext(), PrivatesReductions, llvm::None, VarFieldMap,
      C.getLangOpts().OpenMPCUDAReductionBufNum);
  TeamsReductions.push_back(TeamReductionRec);
  if (!KernelTeamsReductionPtr) {
    KernelTeamsReductionPtr = new llvm::GlobalVariable(
        CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true,
        llvm::GlobalValue::InternalLinkage, nullptr,
        "_openmp_teams_reductions_buffer_$_$ptr");
  }
  llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar(
      Address(KernelTeamsReductionPtr, CGM.getPointerAlign()),
      /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
  llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction(
      CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
  llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction(
      CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
      ReductionFn);
  llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction(
      CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
  llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction(
      CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
      ReductionFn);

  llvm::Value *Args[] = {
      RTLoc,
      ThreadId,
      GlobalBufferPtr,
      CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum),
      RL,
      ShuffleAndReduceFn,
      InterWarpCopyFn,
      GlobalToBufferCpyFn,
      GlobalToBufferRedFn,
      BufferToGlobalCpyFn,
      BufferToGlobalRedFn};

  Res = CGF.EmitRuntimeCall(
      OMPBuilder.getOrCreateRuntimeFunction(
          CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2),
      Args);
}

// 5. Build if (res == 1)
llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done");
llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then");
llvm::Value *Cond = CGF.Builder.CreateICmpEQ(
    Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1));
CGF.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;
CGF.EmitBlock(ThenBB);

// Add emission of __kmpc_end_reduce{_nowait}(<gtid>);
auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps,
                  this](CodeGenFunction &CGF, PrePostActionTy &Action) {
  auto IPriv = Privates.begin();
  auto ILHS = LHSExprs.begin();
  auto IRHS = RHSExprs.begin();
  for (const Expr *E : ReductionOps) {
    emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
                                cast<DeclRefExpr>(*IRHS));
    ++IPriv;
    ++ILHS;
    ++IRHS;
  }
};
llvm::Value *EndArgs[] = {ThreadId};
RegionCodeGenTy RCG(CodeGen);
NVPTXActionTy Action(
    nullptr, llvm::None,
    OMPBuilder.getOrCreateRuntimeFunction(
        CGM.getModule(), OMPRTL___kmpc_nvptx_end_reduce_nowait),
    EndArgs);
RCG.setAction(Action);
RCG(CGF);
// There is no need to emit line number for unconditional branch.
(void)ApplyDebugLocation::CreateEmpty(CGF);
CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
4095}

4097const VarDecl *
4098CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD,
                                     const VarDecl *NativeParam) const {
if (!NativeParam->getType()->isReferenceType())
  return NativeParam;
QualType ArgType = NativeParam->getType();
QualifierCollector QC;
const Type *NonQualTy = QC.strip(ArgType);
QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) {
  if (Attr->getCaptureKind() == OMPC_map) {
    PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy,
                                                      LangAS::opencl_global);
  } else if (Attr->getCaptureKind() == OMPC_firstprivate &&
             PointeeTy.isConstant(CGM.getContext())) {
    PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy,
                                                      LangAS::opencl_generic);
  }
}
ArgType = CGM.getContext().getPointerType(PointeeTy);
QC.addRestrict();
enum { NVPTX_local_addr = 5 };
QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr));
ArgType = QC.apply(CGM.getContext(), ArgType);
if (isa<ImplicitParamDecl>(NativeParam))
  return ImplicitParamDecl::Create(
      CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(),
      NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other);
return ParmVarDecl::Create(
    CGM.getContext(),
    const_cast<DeclContext *>(NativeParam->getDeclContext()),
    NativeParam->getBeginLoc(), NativeParam->getLocation(),
    NativeParam->getIdentifier(), ArgType,
    /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
4131}

4133Address
4134CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF,
                                        const VarDecl *NativeParam,
                                        const VarDecl *TargetParam) const {
assert(NativeParam != TargetParam &&((NativeParam != TargetParam && NativeParam->getType
()->isReferenceType() && "Native arg must not be the same as target arg."
) ? static_cast<void> (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4139, __PRETTY_FUNCTION__))
       NativeParam->getType()->isReferenceType() &&((NativeParam != TargetParam && NativeParam->getType
()->isReferenceType() && "Native arg must not be the same as target arg."
) ? static_cast<void> (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4139, __PRETTY_FUNCTION__))
       "Native arg must not be the same as target arg.")((NativeParam != TargetParam && NativeParam->getType
()->isReferenceType() && "Native arg must not be the same as target arg."
) ? static_cast<void> (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4139, __PRETTY_FUNCTION__));
Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam);
QualType NativeParamType = NativeParam->getType();
QualifierCollector QC;
const Type *NonQualTy = QC.strip(NativeParamType);
QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
unsigned NativePointeeAddrSpace =
    CGF.getContext().getTargetAddressSpace(NativePointeeTy);
QualType TargetTy = TargetParam->getType();
llvm::Value *TargetAddr = CGF.EmitLoadOfScalar(
    LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation());
// First cast to generic.
TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
    TargetAddr, TargetAddr->getType()->getPointerElementType()->getPointerTo(
                    /*AddrSpace=*/0));
// Cast from generic to native address space.
TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
    TargetAddr, TargetAddr->getType()->getPointerElementType()->getPointerTo(
                    NativePointeeAddrSpace));
Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType);
CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false,
                      NativeParamType);
return NativeParamAddr;
4162}

4164void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall(
  CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn,
  ArrayRef<llvm::Value *> Args) const {
SmallVector<llvm::Value *, 4> TargetArgs;
TargetArgs.reserve(Args.size());
auto *FnType = OutlinedFn.getFunctionType();
for (unsigned I = 0, E = Args.size(); I < E; ++I) {
  if (FnType->isVarArg() && FnType->getNumParams() <= I) {
    TargetArgs.append(std::next(Args.begin(), I), Args.end());
    break;
  }
  llvm::Type *TargetType = FnType->getParamType(I);
  llvm::Value *NativeArg = Args[I];
  if (!TargetType->isPointerTy()) {
    TargetArgs.emplace_back(NativeArg);
    continue;
  }
  llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
      NativeArg,
      NativeArg->getType()->getPointerElementType()->getPointerTo());
  TargetArgs.emplace_back(
      CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType));
}
CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs);
4188}

4190/// Emit function which wraps the outline parallel region
4191/// and controls the arguments which are passed to this function.
4192/// The wrapper ensures that the outlined function is called
4193/// with the correct arguments when data is shared.
4194llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper(
  llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) {
ASTContext &Ctx = CGM.getContext();
const auto &CS = *D.getCapturedStmt(OMPD_parallel);

// Create a function that takes as argument the source thread.
FunctionArgList WrapperArgs;
QualType Int16QTy =
    Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false);
QualType Int32QTy =
    Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false);
ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
                                   /*Id=*/nullptr, Int16QTy,
                                   ImplicitParamDecl::Other);
ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
                             /*Id=*/nullptr, Int32QTy,
                             ImplicitParamDecl::Other);
WrapperArgs.emplace_back(&ParallelLevelArg);
WrapperArgs.emplace_back(&WrapperArg);

const CGFunctionInfo &CGFI =
    CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs);

auto *Fn = llvm::Function::Create(
    CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
    Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule());
CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
Fn->setDoesNotRecurse();

CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs,
                  D.getBeginLoc(), D.getBeginLoc());

const auto *RD = CS.getCapturedRecordDecl();
auto CurField = RD->field_begin();

Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
                                                    /*Name=*/".zero.addr");
CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0));
// Get the array of arguments.
SmallVector<llvm::Value *, 8> Args;

Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer());
Args.emplace_back(ZeroAddr.getPointer());

CGBuilderTy &Bld = CGF.Builder;
auto CI = CS.capture_begin();

// Use global memory for data sharing.
// Handle passing of global args to workers.
Address GlobalArgs =
    CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args");
llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer();
llvm::Value *DataSharingArgs[] = {GlobalArgsPtr};
CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                        CGM.getModule(), OMPRTL___kmpc_get_shared_variables),
                    DataSharingArgs);

// Retrieve the shared variables from the list of references returned
// by the runtime. Pass the variables to the outlined function.
Address SharedArgListAddress = Address::invalid();
if (CS.capture_size() > 0 ||
    isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
  SharedArgListAddress = CGF.EmitLoadOfPointer(
      GlobalArgs, CGF.getContext()
                      .getPointerType(CGF.getContext().getPointerType(
                          CGF.getContext().VoidPtrTy))
                      .castAs<PointerType>());
}
unsigned Idx = 0;
if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
  Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
  Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
      Src, CGF.SizeTy->getPointerTo());
  llvm::Value *LB = CGF.EmitLoadOfScalar(
      TypedAddress,
      /*Volatile=*/false,
      CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
      cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc());
  Args.emplace_back(LB);
  ++Idx;
  Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
  TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
      Src, CGF.SizeTy->getPointerTo());
  llvm::Value *UB = CGF.EmitLoadOfScalar(
      TypedAddress,
      /*Volatile=*/false,
      CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
      cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc());
  Args.emplace_back(UB);
  ++Idx;
}
if (CS.capture_size() > 0) {
  ASTContext &CGFContext = CGF.getContext();
  for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) {
    QualType ElemTy = CurField->getType();
    Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx);
    Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
        Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)));
    llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress,
                                            /*Volatile=*/false,
                                            CGFContext.getPointerType(ElemTy),
                                            CI->getLocation());
    if (CI->capturesVariableByCopy() &&
        !CI->getCapturedVar()->getType()->isAnyPointerType()) {
      Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(),
                            CI->getLocation());
    }
    Args.emplace_back(Arg);
  }
}

emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args);
CGF.FinishFunction();
return Fn;
4310}

4312void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF,
                                            const Decl *D) {
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
  return;

assert(D && "Expected function or captured|block decl.")((D && "Expected function or captured|block decl.") ?
 static_cast<void> (0) : __assert_fail ("D && \"Expected function or captured|block decl.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4317, __PRETTY_FUNCTION__));
assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 &&((FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && "Function is registered already."
) ? static_cast<void> (0) : __assert_fail ("FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && \"Function is registered already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4319, __PRETTY_FUNCTION__))
       "Function is registered already.")((FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && "Function is registered already."
) ? static_cast<void> (0) : __assert_fail ("FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && \"Function is registered already.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4319, __PRETTY_FUNCTION__));
assert((!TeamAndReductions.first || TeamAndReductions.first == D) &&(((!TeamAndReductions.first || TeamAndReductions.first == D) &&
 "Team is set but not processed.") ? static_cast<void> (
0) : __assert_fail ("(!TeamAndReductions.first || TeamAndReductions.first == D) && \"Team is set but not processed.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4321, __PRETTY_FUNCTION__))
       "Team is set but not processed.")(((!TeamAndReductions.first || TeamAndReductions.first == D) &&
 "Team is set but not processed.") ? static_cast<void> (
0) : __assert_fail ("(!TeamAndReductions.first || TeamAndReductions.first == D) && \"Team is set but not processed.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4321, __PRETTY_FUNCTION__));
const Stmt *Body = nullptr;
bool NeedToDelayGlobalization = false;
if (const auto *FD = dyn_cast<FunctionDecl>(D)) {
  Body = FD->getBody();
} else if (const auto *BD = dyn_cast<BlockDecl>(D)) {
  Body = BD->getBody();
} else if (const auto *CD = dyn_cast<CapturedDecl>(D)) {
  Body = CD->getBody();
  NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP;
  if (NeedToDelayGlobalization &&
      getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
    return;
}
if (!Body)
  return;
CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second);
VarChecker.Visit(Body);
const RecordDecl *GlobalizedVarsRecord =
    VarChecker.getGlobalizedRecord(IsInTTDRegion);
TeamAndReductions.first = nullptr;
TeamAndReductions.second.clear();
ArrayRef<const ValueDecl *> EscapedVariableLengthDecls =
    VarChecker.getEscapedVariableLengthDecls();
if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty())
  return;
auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
I->getSecond().MappedParams =
    std::make_unique<CodeGenFunction::OMPMapVars>();
I->getSecond().GlobalRecord = GlobalizedVarsRecord;
I->getSecond().EscapedParameters.insert(
    VarChecker.getEscapedParameters().begin(),
    VarChecker.getEscapedParameters().end());
I->getSecond().EscapedVariableLengthDecls.append(
    EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end());
DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
  assert(VD->isCanonicalDecl() && "Expected canonical declaration")((VD->isCanonicalDecl() && "Expected canonical declaration"
) ? static_cast<void> (0) : __assert_fail ("VD->isCanonicalDecl() && \"Expected canonical declaration\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4358, __PRETTY_FUNCTION__));
  const FieldDecl *FD = VarChecker.getFieldForGlobalizedVar(VD);
  Data.insert(std::make_pair(VD, MappedVarData(FD, IsInTTDRegion)));
}
if (!IsInTTDRegion && !NeedToDelayGlobalization && !IsInParallelRegion) {
  CheckVarsEscapingDeclContext VarChecker(CGF, llvm::None);
  VarChecker.Visit(Body);
  I->getSecond().SecondaryGlobalRecord =
      VarChecker.getGlobalizedRecord(/*IsInTTDRegion=*/true);
  I->getSecond().SecondaryLocalVarData.emplace();
  DeclToAddrMapTy &Data = I->getSecond().SecondaryLocalVarData.getValue();
  for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
    assert(VD->isCanonicalDecl() && "Expected canonical declaration")((VD->isCanonicalDecl() && "Expected canonical declaration"
) ? static_cast<void> (0) : __assert_fail ("VD->isCanonicalDecl() && \"Expected canonical declaration\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4370, __PRETTY_FUNCTION__));
    const FieldDecl *FD = VarChecker.getFieldForGlobalizedVar(VD);
    Data.insert(
        std::make_pair(VD, MappedVarData(FD, /*IsInTTDRegion=*/true)));
  }
}
if (!NeedToDelayGlobalization) {
  emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true);
  struct GlobalizationScope final : EHScopeStack::Cleanup {
    GlobalizationScope() = default;

    void Emit(CodeGenFunction &CGF, Flags flags) override {
      static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
          .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true);
    }
  };
  CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup);
}
4388}

4390Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF,
                                                      const VarDecl *VD) {
if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
1
Assuming 'VD' is null→
  const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
  auto AS = LangAS::Default;
  switch (A->getAllocatorType()) {
    // Use the default allocator here as by default local vars are
    // threadlocal.
  case OMPAllocateDeclAttr::OMPNullMemAlloc:
  case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
  case OMPAllocateDeclAttr::OMPThreadMemAlloc:
  case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
  case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
    // Follow the user decision - use default allocation.
    return Address::invalid();
  case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
    // TODO: implement aupport for user-defined allocators.
    return Address::invalid();
  case OMPAllocateDeclAttr::OMPConstMemAlloc:
    AS = LangAS::cuda_constant;
    break;
  case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
    AS = LangAS::cuda_shared;
    break;
  case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
  case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
    break;
  }
  llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
  auto *GV = new llvm::GlobalVariable(
      CGM.getModule(), VarTy, /*isConstant=*/false,
      llvm::GlobalValue::InternalLinkage, llvm::Constant::getNullValue(VarTy),
      VD->getName(),
      /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
      CGM.getContext().getTargetAddressSpace(AS));
  CharUnits Align = CGM.getContext().getDeclAlign(VD);
  GV->setAlignment(Align.getAsAlign());
  return Address(
      CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
          GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace(
                  VD->getType().getAddressSpace()))),
      Align);
}

if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
2
←
Taking false branch→
  return Address::invalid();

VD = VD->getCanonicalDecl();
3
←
Called C++ object pointer is null
auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
if (I == FunctionGlobalizedDecls.end())
  return Address::invalid();
auto VDI = I->getSecond().LocalVarData.find(VD);
if (VDI != I->getSecond().LocalVarData.end())
  return VDI->second.PrivateAddr;
if (VD->hasAttrs()) {
  for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()),
       E(VD->attr_end());
       IT != E; ++IT) {
    auto VDI = I->getSecond().LocalVarData.find(
        cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl())
            ->getCanonicalDecl());
    if (VDI != I->getSecond().LocalVarData.end())
      return VDI->second.PrivateAddr;
  }
}

return Address::invalid();
4457}

4459void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) {
FunctionGlobalizedDecls.erase(CGF.CurFn);
CGOpenMPRuntime::functionFinished(CGF);
4462}

4464void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk(
  CodeGenFunction &CGF, const OMPLoopDirective &S,
  OpenMPDistScheduleClauseKind &ScheduleKind,
  llvm::Value *&Chunk) const {
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
  ScheduleKind = OMPC_DIST_SCHEDULE_static;
  Chunk = CGF.EmitScalarConversion(
      RT.getGPUNumThreads(CGF),
      CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
      S.getIterationVariable()->getType(), S.getBeginLoc());
  return;
}
CGOpenMPRuntime::getDefaultDistScheduleAndChunk(
    CGF, S, ScheduleKind, Chunk);
4479}

4481void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk(
  CodeGenFunction &CGF, const OMPLoopDirective &S,
  OpenMPScheduleClauseKind &ScheduleKind,
  const Expr *&ChunkExpr) const {
ScheduleKind = OMPC_SCHEDULE_static;
// Chunk size is 1 in this case.
llvm::APInt ChunkSize(32, 1);
ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize,
    CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
    SourceLocation());
4491}

4493void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas(
  CodeGenFunction &CGF, const OMPExecutableDirective &D) const {
assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&((isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&
 " Expected target-based directive.") ? static_cast<void>
 (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4496, __PRETTY_FUNCTION__))
       " Expected target-based directive.")((isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&
 " Expected target-based directive.") ? static_cast<void>
 (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4496, __PRETTY_FUNCTION__));
const CapturedStmt *CS = D.getCapturedStmt(OMPD_target);
for (const CapturedStmt::Capture &C : CS->captures()) {
  // Capture variables captured by reference in lambdas for target-based
  // directives.
  if (!C.capturesVariable())
    continue;
  const VarDecl *VD = C.getCapturedVar();
  const auto *RD = VD->getType()
                       .getCanonicalType()
                       .getNonReferenceType()
                       ->getAsCXXRecordDecl();
  if (!RD || !RD->isLambda())
    continue;
  Address VDAddr = CGF.GetAddrOfLocalVar(VD);
  LValue VDLVal;
  if (VD->getType().getCanonicalType()->isReferenceType())
    VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType());
  else
    VDLVal = CGF.MakeAddrLValue(
        VDAddr, VD->getType().getCanonicalType().getNonReferenceType());
  llvm::DenseMap<const VarDecl *, FieldDecl *> Captures;
  FieldDecl *ThisCapture = nullptr;
  RD->getCaptureFields(Captures, ThisCapture);
  if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) {
    LValue ThisLVal =
        CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture);
    llvm::Value *CXXThis = CGF.LoadCXXThis();
    CGF.EmitStoreOfScalar(CXXThis, ThisLVal);
  }
  for (const LambdaCapture &LC : RD->captures()) {
    if (LC.getCaptureKind() != LCK_ByRef)
      continue;
    const VarDecl *VD = LC.getCapturedVar();
    if (!CS->capturesVariable(VD))
      continue;
    auto It = Captures.find(VD);
    assert(It != Captures.end() && "Found lambda capture without field.")((It != Captures.end() && "Found lambda capture without field."
) ? static_cast<void> (0) : __assert_fail ("It != Captures.end() && \"Found lambda capture without field.\""
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4533, __PRETTY_FUNCTION__));
    LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second);
    Address VDAddr = CGF.GetAddrOfLocalVar(VD);
    if (VD->getType().getCanonicalType()->isReferenceType())
      VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr,
                                             VD->getType().getCanonicalType())
                   .getAddress(CGF);
    CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal);
  }
}
4543}

4545unsigned CGOpenMPRuntimeGPU::getDefaultFirstprivateAddressSpace() const {
return CGM.getContext().getTargetAddressSpace(LangAS::cuda_constant);
4547}

4549bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD,
                                                          LangAS &AS) {
if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())
  return false;
const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
switch(A->getAllocatorType()) {
case OMPAllocateDeclAttr::OMPNullMemAlloc:
case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
// Not supported, fallback to the default mem space.
case OMPAllocateDeclAttr::OMPThreadMemAlloc:
case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
  AS = LangAS::Default;
  return true;
case OMPAllocateDeclAttr::OMPConstMemAlloc:
  AS = LangAS::cuda_constant;
  return true;
case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
  AS = LangAS::cuda_shared;
  return true;
case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
  llvm_unreachable("Expected predefined allocator for the variables with the "::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the "
 "static storage.", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4573)
                   "static storage.")::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the "
 "static storage.", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4573);
}
return false;
4576}

4578// Get current CudaArch and ignore any unknown values
4579static CudaArch getCudaArch(CodeGenModule &CGM) {
if (!CGM.getTarget().hasFeature("ptx"))
  return CudaArch::UNKNOWN;
for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) {
  if (Feature.getValue()) {
    CudaArch Arch = StringToCudaArch(Feature.getKey());
    if (Arch != CudaArch::UNKNOWN)
      return Arch;
  }
}
return CudaArch::UNKNOWN;
4590}

4592/// Check to see if target architecture supports unified addressing which is
4593/// a restriction for OpenMP requires clause "unified_shared_memory".
4594void CGOpenMPRuntimeGPU::processRequiresDirective(
  const OMPRequiresDecl *D) {
for (const OMPClause *Clause : D->clauselists()) {
  if (Clause->getClauseKind() == OMPC_unified_shared_memory) {
    CudaArch Arch = getCudaArch(CGM);
    switch (Arch) {
    case CudaArch::SM_20:
    case CudaArch::SM_21:
    case CudaArch::SM_30:
    case CudaArch::SM_32:
    case CudaArch::SM_35:
    case CudaArch::SM_37:
    case CudaArch::SM_50:
    case CudaArch::SM_52:
    case CudaArch::SM_53:
    case CudaArch::SM_60:
    case CudaArch::SM_61:
    case CudaArch::SM_62: {
      SmallString<256> Buffer;
      llvm::raw_svector_ostream Out(Buffer);
      Out << "Target architecture " << CudaArchToString(Arch)
          << " does not support unified addressing";
      CGM.Error(Clause->getBeginLoc(), Out.str());
      return;
    }
    case CudaArch::SM_70:
    case CudaArch::SM_72:
    case CudaArch::SM_75:
    case CudaArch::SM_80:
    case CudaArch::GFX600:
    case CudaArch::GFX601:
    case CudaArch::GFX602:
    case CudaArch::GFX700:
    case CudaArch::GFX701:
    case CudaArch::GFX702:
    case CudaArch::GFX703:
    case CudaArch::GFX704:
    case CudaArch::GFX705:
    case CudaArch::GFX801:
    case CudaArch::GFX802:
    case CudaArch::GFX803:
    case CudaArch::GFX805:
    case CudaArch::GFX810:
    case CudaArch::GFX900:
    case CudaArch::GFX902:
    case CudaArch::GFX904:
    case CudaArch::GFX906:
    case CudaArch::GFX908:
    case CudaArch::GFX909:
    case CudaArch::GFX90c:
    case CudaArch::GFX1010:
    case CudaArch::GFX1011:
    case CudaArch::GFX1012:
    case CudaArch::GFX1030:
    case CudaArch::GFX1031:
    case CudaArch::GFX1032:
    case CudaArch::GFX1033:
    case CudaArch::UNUSED:
    case CudaArch::UNKNOWN:
      break;
    case CudaArch::LAST:
      llvm_unreachable("Unexpected Cuda arch.")::llvm::llvm_unreachable_internal("Unexpected Cuda arch.", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4655);
    }
  }
}
CGOpenMPRuntime::processRequiresDirective(D);
4660}

4662/// Get number of SMs and number of blocks per SM.
4663static std::pair<unsigned, unsigned> getSMsBlocksPerSM(CodeGenModule &CGM) {
std::pair<unsigned, unsigned> Data;
if (CGM.getLangOpts().OpenMPCUDANumSMs)
  Data.first = CGM.getLangOpts().OpenMPCUDANumSMs;
if (CGM.getLangOpts().OpenMPCUDABlocksPerSM)
  Data.second = CGM.getLangOpts().OpenMPCUDABlocksPerSM;
if (Data.first && Data.second)
  return Data;
switch (getCudaArch(CGM)) {
case CudaArch::SM_20:
case CudaArch::SM_21:
case CudaArch::SM_30:
case CudaArch::SM_32:
case CudaArch::SM_35:
case CudaArch::SM_37:
case CudaArch::SM_50:
case CudaArch::SM_52:
case CudaArch::SM_53:
  return {16, 16};
case CudaArch::SM_60:
case CudaArch::SM_61:
case CudaArch::SM_62:
  return {56, 32};
case CudaArch::SM_70:
case CudaArch::SM_72:
case CudaArch::SM_75:
case CudaArch::SM_80:
  return {84, 32};
case CudaArch::GFX600:
case CudaArch::GFX601:
case CudaArch::GFX602:
case CudaArch::GFX700:
case CudaArch::GFX701:
case CudaArch::GFX702:
case CudaArch::GFX703:
case CudaArch::GFX704:
case CudaArch::GFX705:
case CudaArch::GFX801:
case CudaArch::GFX802:
case CudaArch::GFX803:
case CudaArch::GFX805:
case CudaArch::GFX810:
case CudaArch::GFX900:
case CudaArch::GFX902:
case CudaArch::GFX904:
case CudaArch::GFX906:
case CudaArch::GFX908:
case CudaArch::GFX909:
case CudaArch::GFX90c:
case CudaArch::GFX1010:
case CudaArch::GFX1011:
case CudaArch::GFX1012:
case CudaArch::GFX1030:
case CudaArch::GFX1031:
case CudaArch::GFX1032:
case CudaArch::GFX1033:
case CudaArch::UNUSED:
case CudaArch::UNKNOWN:
  break;
case CudaArch::LAST:
  llvm_unreachable("Unexpected Cuda arch.")::llvm::llvm_unreachable_internal("Unexpected Cuda arch.", "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4723);
}
llvm_unreachable("Unexpected NVPTX target without ptx feature.")::llvm::llvm_unreachable_internal("Unexpected NVPTX target without ptx feature."
, "/build/llvm-toolchain-snapshot-12~++20210124100612+2afaf072f5c1/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 4725);
4726}

4728void CGOpenMPRuntimeGPU::clear() {
if (!GlobalizedRecords.empty() &&
    !CGM.getLangOpts().OpenMPCUDATargetParallel) {
  ASTContext &C = CGM.getContext();
  llvm::SmallVector<const GlobalPtrSizeRecsTy *, 4> GlobalRecs;
  llvm::SmallVector<const GlobalPtrSizeRecsTy *, 4> SharedRecs;
  RecordDecl *StaticRD = C.buildImplicitRecord(
      "_openmp_static_memory_type_$_", RecordDecl::TagKind::TTK_Union);
  StaticRD->startDefinition();
  RecordDecl *SharedStaticRD = C.buildImplicitRecord(
      "_shared_openmp_static_memory_type_$_", RecordDecl::TagKind::TTK_Union);
  SharedStaticRD->startDefinition();
  for (const GlobalPtrSizeRecsTy &Records : GlobalizedRecords) {
    if (Records.Records.empty())
      continue;
    unsigned Size = 0;
    unsigned RecAlignment = 0;
    for (const RecordDecl *RD : Records.Records) {
      QualType RDTy = C.getRecordType(RD);
      unsigned Alignment = C.getTypeAlignInChars(RDTy).getQuantity();
      RecAlignment = std::max(RecAlignment, Alignment);
      unsigned RecSize = C.getTypeSizeInChars(RDTy).getQuantity();
      Size =
          llvm::alignTo(llvm::alignTo(Size, Alignment) + RecSize, Alignment);
    }
    Size = llvm::alignTo(Size, RecAlignment);
    llvm::APInt ArySize(/*numBits=*/64, Size);
    QualType SubTy = C.getConstantArrayType(
        C.CharTy, ArySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0);
    const bool UseSharedMemory = Size <= SharedMemorySize;
    auto *Field =
        FieldDecl::Create(C, UseSharedMemory ? SharedStaticRD : StaticRD,
                          SourceLocation(), SourceLocation(), nullptr, SubTy,
                          C.getTrivialTypeSourceInfo(SubTy, SourceLocation()),
                          /*BW=*/nullptr, /*Mutable=*/false,
                          /*InitStyle=*/ICIS_NoInit);
    Field->setAccess(AS_public);
    if (UseSharedMemory) {
      SharedStaticRD->addDecl(Field);
      SharedRecs.push_back(&Records);
    } else {
      StaticRD->addDecl(Field);
      GlobalRecs.push_back(&Records);
    }
    Records.RecSize->setInitializer(llvm::ConstantInt::get(CGM.SizeTy, Size));
    Records.UseSharedMemory->setInitializer(
        llvm::ConstantInt::get(CGM.Int16Ty, UseSharedMemory ? 1 : 0));
  }
  // Allocate SharedMemorySize buffer for the shared memory.
  // FIXME: nvlink does not handle weak linkage correctly (object with the
  // different size are reported as erroneous).
  // Restore this code as sson as nvlink is fixed.
  if (!SharedStaticRD->field_empty()) {
    llvm::APInt ArySize(/*numBits=*/64, SharedMemorySize);
    QualType SubTy = C.getConstantArrayType(
        C.CharTy, ArySize, nullptr, ArrayType::Normal, /*IndexTypeQuals=*/0);
    auto *Field = FieldDecl::Create(
        C, SharedStaticRD, SourceLocation(), SourceLocation(), nullptr, SubTy,
        C.getTrivialTypeSourceInfo(SubTy, SourceLocation()),
        /*BW=*/nullptr, /*Mutable=*/false,
        /*InitStyle=*/ICIS_NoInit);
    Field->setAccess(AS_public);
    SharedStaticRD->addDecl(Field);
  }
  SharedStaticRD->completeDefinition();
  if (!SharedStaticRD->field_empty()) {
    QualType StaticTy = C.getRecordType(SharedStaticRD);
    llvm::Type *LLVMStaticTy = CGM.getTypes().ConvertTypeForMem(StaticTy);
    auto *GV = new llvm::GlobalVariable(
        CGM.getModule(), LLVMStaticTy,
        /*isConstant=*/false, llvm::GlobalValue::WeakAnyLinkage,
        llvm::UndefValue::get(LLVMStaticTy),
        "_openmp_shared_static_glob_rd_$_", /*InsertBefore=*/nullptr,
        llvm::GlobalValue::NotThreadLocal,
        C.getTargetAddressSpace(LangAS::cuda_shared));
    auto *Replacement = llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(
        GV, CGM.VoidPtrTy);
    for (const GlobalPtrSizeRecsTy *Rec : SharedRecs) {
      Rec->Buffer->replaceAllUsesWith(Replacement);
      Rec->Buffer->eraseFromParent();
    }
  }
  StaticRD->completeDefinition();
  if (!StaticRD->field_empty()) {
    QualType StaticTy = C.getRecordType(StaticRD);
    std::pair<unsigned, unsigned> SMsBlockPerSM = getSMsBlocksPerSM(CGM);
    llvm::APInt Size1(32, SMsBlockPerSM.second);
    QualType Arr1Ty =
        C.getConstantArrayType(StaticTy, Size1, nullptr, ArrayType::Normal,
                               /*IndexTypeQuals=*/0);
    llvm::APInt Size2(32, SMsBlockPerSM.first);
    QualType Arr2Ty =
        C.getConstantArrayType(Arr1Ty, Size2, nullptr, ArrayType::Normal,
                               /*IndexTypeQuals=*/0);
    llvm::Type *LLVMArr2Ty = CGM.getTypes().ConvertTypeForMem(Arr2Ty);
    // FIXME: nvlink does not handle weak linkage correctly (object with the
    // different size are reported as erroneous).
    // Restore CommonLinkage as soon as nvlink is fixed.
    auto *GV = new llvm::GlobalVariable(
        CGM.getModule(), LLVMArr2Ty,
        /*isConstant=*/false, llvm::GlobalValue::InternalLinkage,
        llvm::Constant::getNullValue(LLVMArr2Ty),
        "_openmp_static_glob_rd_$_");
    auto *Replacement = llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(
        GV, CGM.VoidPtrTy);
    for (const GlobalPtrSizeRecsTy *Rec : GlobalRecs) {
      Rec->Buffer->replaceAllUsesWith(Replacement);
      Rec->Buffer->eraseFromParent();
    }
  }
}
if (!TeamsReductions.empty()) {
  ASTContext &C = CGM.getContext();
  RecordDecl *StaticRD = C.buildImplicitRecord(
      "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union);
  StaticRD->startDefinition();
  for (const RecordDecl *TeamReductionRec : TeamsReductions) {
    QualType RecTy = C.getRecordType(TeamReductionRec);
    auto *Field = FieldDecl::Create(
        C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy,
        C.getTrivialTypeSourceInfo(RecTy, SourceLocation()),
        /*BW=*/nullptr, /*Mutable=*/false,
        /*InitStyle=*/ICIS_NoInit);
    Field->setAccess(AS_public);
    StaticRD->addDecl(Field);
  }
  StaticRD->completeDefinition();
  QualType StaticTy = C.getRecordType(StaticRD);
  llvm::Type *LLVMReductionsBufferTy =
      CGM.getTypes().ConvertTypeForMem(StaticTy);
  // FIXME: nvlink does not handle weak linkage correctly (object with the
  // different size are reported as erroneous).
  // Restore CommonLinkage as soon as nvlink is fixed.
  auto *GV = new llvm::GlobalVariable(
      CGM.getModule(), LLVMReductionsBufferTy,
      /*isConstant=*/false, llvm::GlobalValue::InternalLinkage,
      llvm::Constant::getNullValue(LLVMReductionsBufferTy),
      "_openmp_teams_reductions_buffer_$_");
  KernelTeamsReductionPtr->setInitializer(
      llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV,
                                                           CGM.VoidPtrTy));
}
CGOpenMPRuntime::clear();
4871}