/build/source/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp

Bug Summary

File:	build/source/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
Warning:	line 531, column 29 The result of the right shift is undefined due to shifting by '33', which is greater or equal to the width of type 'unsigned int'
Annotated Source Code

Press '?' to see keyboard shortcuts
Show analyzer invocation
clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name CGOpenMPRuntimeGPU.cpp -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 -mframe-pointer=none -relaxed-aliasing -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/source/build-llvm -resource-dir /usr/lib/llvm-17/lib/clang/17 -D _DEBUG -D _GLIBCXX_ASSERTIONS -D _GNU_SOURCE -D _LIBCPP_ENABLE_ASSERTIONS -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I tools/clang/lib/CodeGen -I /build/source/clang/lib/CodeGen -I /build/source/clang/include -I tools/clang/include -I include -I /build/source/llvm/include -D _FORTIFY_SOURCE=2 -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-17/lib/clang/17/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/source/build-llvm=build-llvm -fmacro-prefix-map=/build/source/= -fcoverage-prefix-map=/build/source/build-llvm=build-llvm -fcoverage-prefix-map=/build/source/= -source-date-epoch 1679915782 -O3 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -Wno-misleading-indentation -std=c++17 -fdeprecated-macro -fdebug-compilation-dir=/build/source/build-llvm -fdebug-prefix-map=/build/source/build-llvm=build-llvm -fdebug-prefix-map=/build/source/= -fdebug-prefix-map=/build/source/build-llvm=build-llvm -fdebug-prefix-map=/build/source/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2023-03-27-130437-16335-1 -x c++ /build/source/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 "CodeGenFunction.h"
16#include "clang/AST/Attr.h"
17#include "clang/AST/DeclOpenMP.h"
18#include "clang/AST/OpenMPClause.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/Support/MathExtras.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;

78public:
ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode,
                          CGOpenMPRuntimeGPU::ExecutionMode EntryMode)
    : ExecMode(ExecMode) {
  SavedExecMode = ExecMode;
  ExecMode = EntryMode;
}
~ExecutionRuntimeModesRAII() { ExecMode = SavedExecMode; }
86};

88/// GPU Configuration:  This information can be derived from cuda registers,
89/// however, providing compile time constants helps generate more efficient
90/// code.  For all practical purposes this is fine because the configuration
91/// is the same for all known NVPTX architectures.
92enum MachineConfiguration : unsigned {
/// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target
/// specific Grid Values like GV_Warp_Size, GV_Slot_Size

/// Global memory alignment for performance.
GlobalMemoryAlignment = 128,
98};

100static 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());
120}


123static 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;
200}

202/// Get the list of variables that can escape their declaration context.
203class 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() &&(static_cast <bool> (!VD->getType()->isVariablyModifiedType
() && "Parameter captured by value with variably modified type"
) ? void (0) : __assert_fail ("!VD->getType()->isVariablyModifiedType() && \"Parameter captured by value with variably modified type\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 242, __extension__
 __PRETTY_FUNCTION__))
               "Parameter captured by value with variably modified type")(static_cast <bool> (!VD->getType()->isVariablyModifiedType
() && "Parameter captured by value with variably modified type"
) ? void (0) : __assert_fail ("!VD->getType()->isVariablyModifiedType() && \"Parameter captured by value with variably modified type\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 242, __extension__
 __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.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 298);
          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 &&(static_cast <bool> (!GlobalizedRD && "Record for globalized variables is built already."
) ? void (0) : __assert_fail ("!GlobalizedRD && \"Record for globalized variables is built already.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 321, __extension__
 __PRETTY_FUNCTION__))
         "Record for globalized variables is built already.")(static_cast <bool> (!GlobalizedRD && "Record for globalized variables is built already."
) ? void (0) : __assert_fail ("!GlobalizedRD && \"Record for globalized variables is built already.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 321, __extension__
 __PRETTY_FUNCTION__));
  ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams;
  unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size;
  if (IsInTTDRegion)
    EscapedDeclsForTeams = EscapedDecls.getArrayRef();
  else
    EscapedDeclsForParallel = EscapedDecls.getArrayRef();
  GlobalizedRD = ::buildRecordForGlobalizedVars(
      CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams,
      MappedDeclsFields, WarpSize);
}

333public:
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 (VD->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 &&(static_cast <bool> (GlobalizedRD && "Record for globalized variables must be generated already."
) ? void (0) : __assert_fail ("GlobalizedRD && \"Record for globalized variables must be generated already.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 487, __extension__
 __PRETTY_FUNCTION__))
         "Record for globalized variables must be generated already.")(static_cast <bool> (GlobalizedRD && "Record for globalized variables must be generated already."
) ? void (0) : __assert_fail ("GlobalizedRD && \"Record for globalized variables must be generated already.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 487, __extension__
 __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();
}
510};
511} // anonymous namespace

513/// Get the id of the warp in the block.
514/// We assume that the warp size is 32, which is always the case
515/// on the NVPTX device, to generate more efficient code.
516static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
unsigned LaneIDBits =
    llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size);
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id");
522}

524/// Get the id of the current lane in the Warp.
525/// We assume that the warp size is 32, which is always the case
526/// on the NVPTX device, to generate more efficient code.
527static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
unsigned LaneIDBits =
    llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size);
unsigned LaneIDMask = ~0u >> (32u - LaneIDBits);
14
←
The result of the right shift is undefined due to shifting by '33', which is greater or equal to the width of type 'unsigned int'
auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask),
                     "nvptx_lane_id");
535}

537CGOpenMPRuntimeGPU::ExecutionMode
538CGOpenMPRuntimeGPU::getExecutionMode() const {
return CurrentExecutionMode;
540}

542static CGOpenMPRuntimeGPU::DataSharingMode
543getDataSharingMode(CodeGenModule &CGM) {
return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA
                                        : CGOpenMPRuntimeGPU::Generic;
546}

548/// Check for inner (nested) SPMD construct, if any
549static 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.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 644);
  }
}

return false;
649}

651static 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."
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 728)
    "Unknown programming model for OpenMP directive on NVPTX target.")::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target."
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 728);
729}

731void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D,
                                           StringRef ParentName,
                                           llvm::Function *&OutlinedFn,
                                           llvm::Constant *&OutlinedFnID,
                                           bool IsOffloadEntry,
                                           const RegionCodeGenTy &CodeGen) {
ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_NonSPMD);
EntryFunctionState EST;
WrapperFunctionsMap.clear();

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

public:
  NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST)
      : EST(EST) {}
  void Enter(CodeGenFunction &CGF) override {
    auto &RT =
        static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
    RT.emitKernelInit(CGF, EST, /* IsSPMD */ false);
    // 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.emitKernelDeinit(CGF, EST, /* IsSPMD */ false);
  }
} Action(EST);
CodeGen.setAction(Action);
IsInTTDRegion = true;
emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
                                 IsOffloadEntry, CodeGen);
IsInTTDRegion = false;
767}

769void CGOpenMPRuntimeGPU::emitKernelInit(CodeGenFunction &CGF,
                                      EntryFunctionState &EST, bool IsSPMD) {
CGBuilderTy &Bld = CGF.Builder;
Bld.restoreIP(OMPBuilder.createTargetInit(Bld, IsSPMD));
if (!IsSPMD)
  emitGenericVarsProlog(CGF, EST.Loc);
775}

777void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF,
                                        EntryFunctionState &EST,
                                        bool IsSPMD) {
if (!IsSPMD)
  emitGenericVarsEpilog(CGF);

CGBuilderTy &Bld = CGF.Builder;
OMPBuilder.createTargetDeinit(Bld, IsSPMD);
785}

787void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D,
                                        StringRef ParentName,
                                        llvm::Function *&OutlinedFn,
                                        llvm::Constant *&OutlinedFnID,
                                        bool IsOffloadEntry,
                                        const RegionCodeGenTy &CodeGen) {
ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_SPMD);
EntryFunctionState EST;

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

public:
  NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT,
                       CGOpenMPRuntimeGPU::EntryFunctionState &EST)
      : RT(RT), EST(EST) {}
  void Enter(CodeGenFunction &CGF) override {
    RT.emitKernelInit(CGF, EST, /* IsSPMD */ true);
    // Skip target region initialization.
    RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
  }
  void Exit(CodeGenFunction &CGF) override {
    RT.clearLocThreadIdInsertPt(CGF);
    RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true);
  }
} Action(*this, EST);
CodeGen.setAction(Action);
IsInTTDRegion = true;
emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
                                 IsOffloadEntry, CodeGen);
IsInTTDRegion = false;
820}

822// Create a unique global variable to indicate the execution mode of this target
823// region. The execution mode is either 'generic', or 'spmd' depending on the
824// target directive. This variable is picked up by the offload library to setup
825// the device appropriately before kernel launch. If the execution mode is
826// 'generic', the runtime reserves one warp for the master, otherwise, all
827// warps participate in parallel work.
828static 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 ? OMP_TGT_EXEC_MODE_SPMD
                                            : OMP_TGT_EXEC_MODE_GENERIC),
    Twine(Name, "_exec_mode"));
GVMode->setVisibility(llvm::GlobalVariable::ProtectedVisibility);
CGM.addCompilerUsedGlobal(GVMode);
838}

840void 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!")(static_cast <bool> (!ParentName.empty() && "Invalid target region parent name!"
) ? void (0) : __assert_fail ("!ParentName.empty() && \"Invalid target region parent name!\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 847, __extension__
 __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);
858}

860CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM)
  : CGOpenMPRuntime(CGM) {
llvm::OpenMPIRBuilderConfig Config(CGM.getLangOpts().OpenMPIsDevice, true,
                                   hasRequiresUnifiedSharedMemory(),
                                   CGM.getLangOpts().OpenMPOffloadMandatory);
OMPBuilder.setConfig(Config);

if (!CGM.getLangOpts().OpenMPIsDevice)
  llvm_unreachable("OpenMP can only handle device code.")::llvm::llvm_unreachable_internal("OpenMP can only handle device code."
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 868);

llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder();
if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty())
  return;

OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug,
                            "__omp_rtl_debug_kind");
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription,
                            "__omp_rtl_assume_teams_oversubscription");
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription,
                            "__omp_rtl_assume_threads_oversubscription");
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState,
                            "__omp_rtl_assume_no_thread_state");
OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoNestedParallelism,
                            "__omp_rtl_assume_no_nested_parallelism");
884}

886void 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);
894}

896void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF,
                                              llvm::Value *NumThreads,
                                              SourceLocation Loc) {
// Nothing to do.
900}

902void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF,
                                            const Expr *NumTeams,
                                            const Expr *ThreadLimit,
                                            SourceLocation Loc) {}

907llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction(
  const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
  OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
// Emit target region as a standalone region.
bool PrevIsInTTDRegion = IsInTTDRegion;
IsInTTDRegion = false;
auto *OutlinedFun =
    cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction(
        D, ThreadIDVar, InnermostKind, CodeGen));
IsInTTDRegion = PrevIsInTTDRegion;
if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) {
  llvm::Function *WrapperFun =
      createParallelDataSharingWrapper(OutlinedFun, D);
  WrapperFunctionsMap[OutlinedFun] = WrapperFun;
}

return OutlinedFun;
924}

926/// Get list of lastprivate variables from the teams distribute ... or
927/// teams {distribute ...} directives.
928static void
929getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D,
                           llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPTeamsDirective(D.getDirectiveKind
()) && "expected teams directive.") ? void (0) : __assert_fail
 ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 932, __extension__
 __PRETTY_FUNCTION__))
       "expected teams directive.")(static_cast <bool> (isOpenMPTeamsDirective(D.getDirectiveKind
()) && "expected teams directive.") ? void (0) : __assert_fail
 ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 932, __extension__
 __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));
}
950}

952/// Get list of reduction variables from the teams ... directives.
953static void
954getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D,
                    llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPTeamsDirective(D.getDirectiveKind
()) && "expected teams directive.") ? void (0) : __assert_fail
 ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 957, __extension__
 __PRETTY_FUNCTION__))
       "expected teams directive.")(static_cast <bool> (isOpenMPTeamsDirective(D.getDirectiveKind
()) && "expected teams directive.") ? void (0) : __assert_fail
 ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 957, __extension__
 __PRETTY_FUNCTION__));
for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
  for (const Expr *E : C->privates())
    Vars.push_back(getPrivateItem(E));
}
962}

964llvm::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().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(), std::nullopt, LastPrivatesReductions,
        MappedDeclsFields, WarpSize);
  }
} else if (!LastPrivatesReductions.empty()) {
  assert(!TeamAndReductions.first &&(static_cast <bool> (!TeamAndReductions.first &&
 "Previous team declaration is not expected.") ? void (0) : __assert_fail
 ("!TeamAndReductions.first && \"Previous team declaration is not expected.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 985, __extension__
 __PRETTY_FUNCTION__))
         "Previous team declaration is not expected.")(static_cast <bool> (!TeamAndReductions.first &&
 "Previous team declaration is not expected.") ? void (0) : __assert_fail
 ("!TeamAndReductions.first && \"Previous team declaration is not expected.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 985, __extension__
 __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().MappedParams =
          std::make_unique<CodeGenFunction::OMPMapVars>();
      DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
      for (const auto &Pair : MappedDeclsFields) {
        assert(Pair.getFirst()->isCanonicalDecl() &&(static_cast <bool> (Pair.getFirst()->isCanonicalDecl
() && "Expected canonical declaration") ? void (0) : __assert_fail
 ("Pair.getFirst()->isCanonicalDecl() && \"Expected canonical declaration\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1014, __extension__
 __PRETTY_FUNCTION__))
               "Expected canonical declaration")(static_cast <bool> (Pair.getFirst()->isCanonicalDecl
() && "Expected canonical declaration") ? void (0) : __assert_fail
 ("Pair.getFirst()->isCanonicalDecl() && \"Expected canonical declaration\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1014, __extension__
 __PRETTY_FUNCTION__));
        Data.insert(std::make_pair(Pair.getFirst(), MappedVarData()));
      }
    }
    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);

return OutlinedFun;
1030}

1032void 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;

for (auto &Rec : I->getSecond().LocalVarData) {
  const auto *VD = cast<VarDecl>(Rec.first);
  bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first);
  QualType VarTy = VD->getType();

  // Get the local allocation of a firstprivate variable before sharing
  llvm::Value *ParValue;
  if (EscapedParam) {
    LValue ParLVal =
        CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
    ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
  }

  // Allocate space for the variable to be globalized
  llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())};
  llvm::CallBase *VoidPtr =
      CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                              CGM.getModule(), OMPRTL___kmpc_alloc_shared),
                          AllocArgs, VD->getName());
  // FIXME: We should use the variables actual alignment as an argument.
  VoidPtr->addRetAttr(llvm::Attribute::get(
      CGM.getLLVMContext(), llvm::Attribute::Alignment,
      CGM.getContext().getTargetInfo().getNewAlign() / 8));

  // Cast the void pointer and get the address of the globalized variable.
  llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo();
  llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
      VoidPtr, VarPtrTy, VD->getName() + "_on_stack");
  LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy);
  Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
  Rec.second.GlobalizedVal = VoidPtr;

  // Assign the local allocation to the newly globalized location.
  if (EscapedParam) {
    CGF.EmitStoreOfScalar(ParValue, VarAddr);
    I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF));
  }
  if (auto *DI = CGF.getDebugInfo())
    VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation()));
}
for (const auto *VD : I->getSecond().EscapedVariableLengthDecls) {
  // Use actual memory size of the VLA object including the padding
  // for alignment purposes.
  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);

  // Allocate space for this VLA object to be globalized.
  llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())};
  llvm::CallBase *VoidPtr =
      CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                              CGM.getModule(), OMPRTL___kmpc_alloc_shared),
                          AllocArgs, VD->getName());
  VoidPtr->addRetAttr(
      llvm::Attribute::get(CGM.getLLVMContext(), llvm::Attribute::Alignment,
                           CGM.getContext().getTargetInfo().getNewAlign()));

  I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(
      std::pair<llvm::Value *, llvm::Value *>(
          {VoidPtr, CGF.getTypeSize(VD->getType())}));
  LValue Base = CGF.MakeAddrLValue(VoidPtr, VD->getType(),
                                   CGM.getContext().getDeclAlign(VD),
                                   AlignmentSource::Decl);
  I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD),
                                          Base.getAddress(CGF));
}
I->getSecond().MappedParams->apply(CGF);
1118}

1120void 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()) {
  // Deallocate the memory for each globalized VLA object
  for (auto AddrSizePair :
       llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) {
    CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                            CGM.getModule(), OMPRTL___kmpc_free_shared),
                        {AddrSizePair.first, AddrSizePair.second});
  }
  // Deallocate the memory for each globalized value
  for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) {
    const auto *VD = cast<VarDecl>(Rec.first);
    I->getSecond().MappedParams->restore(CGF);

    llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal,
                               CGF.getTypeSize(VD->getType())};
    CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                            CGM.getModule(), OMPRTL___kmpc_free_shared),
                        FreeArgs);
  }
}
1147}

1149void 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.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
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);
1165}

1167void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF,
                                        SourceLocation Loc,
                                        llvm::Function *OutlinedFn,
                                        ArrayRef<llvm::Value *> CapturedVars,
                                        const Expr *IfCond,
                                        llvm::Value *NumThreads) {
if (!CGF.HaveInsertPoint())
  return;

auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond,
                      NumThreads](CodeGenFunction &CGF,
                                  PrePostActionTy &Action) {
  CGBuilderTy &Bld = CGF.Builder;
  llvm::Value *NumThreadsVal = NumThreads;
  llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn];
  llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy);
  if (WFn)
    ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy);
  llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy);

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

  Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca(
      llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()),
      "captured_vars_addrs");
  // There's something to share.
  if (!CapturedVars.empty()) {
    // Prepare for parallel region. Indicate the outlined function.
    ASTContext &Ctx = CGF.getContext();
    unsigned Idx = 0;
    for (llvm::Value *V : CapturedVars) {
      Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, 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;
    }
  }

  llvm::Value *IfCondVal = nullptr;
  if (IfCond)
    IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty,
                                  /* isSigned */ false);
  else
    IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1);

  if (!NumThreadsVal)
    NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1);
  else
    NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty),

    assert(IfCondVal && "Expected a value")(static_cast <bool> (IfCondVal && "Expected a value"
) ? void (0) : __assert_fail ("IfCondVal && \"Expected a value\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1225, __extension__
 __PRETTY_FUNCTION__));
  llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
  llvm::Value *Args[] = {
      RTLoc,
      getThreadID(CGF, Loc),
      IfCondVal,
      NumThreadsVal,
      llvm::ConstantInt::get(CGF.Int32Ty, -1),
      FnPtr,
      ID,
      Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(),
                                 CGF.VoidPtrPtrTy),
      llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())};
  CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                          CGM.getModule(), OMPRTL___kmpc_parallel_51),
                      Args);
};

RegionCodeGenTy RCG(ParallelGen);
RCG(CGF);
1245}

1247void 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);
1260}

1262void 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);
1277}

1279void 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);
1346}

1348/// Cast value to the specified type.
1349static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val,
                                  QualType ValTy, QualType CastTy,
                                  SourceLocation Loc) {
assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&(static_cast <bool> (!CGF.getContext().getTypeSizeInChars
(CastTy).isZero() && "Cast type must sized.") ? void (
0) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && \"Cast type must sized.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1353, __extension__
 __PRETTY_FUNCTION__))
       "Cast type must sized.")(static_cast <bool> (!CGF.getContext().getTypeSizeInChars
(CastTy).isZero() && "Cast type must sized.") ? void (
0) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && \"Cast type must sized.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1353, __extension__
 __PRETTY_FUNCTION__));
assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&(static_cast <bool> (!CGF.getContext().getTypeSizeInChars
(ValTy).isZero() && "Val type must sized.") ? void (0
) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && \"Val type must sized.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1355, __extension__
 __PRETTY_FUNCTION__))
       "Val type must sized.")(static_cast <bool> (!CGF.getContext().getTypeSizeInChars
(ValTy).isZero() && "Val type must sized.") ? void (0
) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && \"Val type must sized.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1355, __extension__
 __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()),
    Val->getType());
CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy,
                      LValueBaseInfo(AlignmentSource::Type),
                      TBAAAccessInfo());
return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc,
                            LValueBaseInfo(AlignmentSource::Type),
                            TBAAAccessInfo());
1375}

1377/// This function creates calls to one of two shuffle functions to copy
1378/// variables between lanes in a warp.
1379static 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 &&(static_cast <bool> (Size.getQuantity() <= 8 &&
 "Unsupported bitwidth in shuffle instruction.") ? void (0) :
 __assert_fail ("Size.getQuantity() <= 8 && \"Unsupported bitwidth in shuffle instruction.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1392, __extension__
 __PRETTY_FUNCTION__))
       "Unsupported bitwidth in shuffle instruction.")(static_cast <bool> (Size.getQuantity() <= 8 &&
 "Unsupported bitwidth in shuffle instruction.") ? void (0) :
 __assert_fail ("Size.getQuantity() <= 8 && \"Unsupported bitwidth in shuffle instruction.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1392, __extension__
 __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);
1410}

1412static 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, CGF.Int8Ty);
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(),
                                                IntTy);
  ElemPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
      ElemPtr, IntTy->getPointerTo(), IntTy);
  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.getElementType(), Ptr.getAlignment());
    ElemPtr =
        Address(PhiDest, ElemPtr.getElementType(), ElemPtr.getAlignment());
    llvm::Value *PtrDiff = Bld.CreatePtrDiff(
        CGF.Int8Ty, 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;
}
1495}

1497namespace {
1498enum 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,
1509};
1510} // namespace

1512struct CopyOptionsTy {
llvm::Value *RemoteLaneOffset;
llvm::Value *ScratchpadIndex;
llvm::Value *ScratchpadWidth;
1516};

1518/// Emit instructions to copy a Reduce list, which contains partially
1519/// aggregated values, in the specified direction.
1520static 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;
  QualType PrivatePtrType = C.getPointerType(Private->getType());
  llvm::Type *PrivateLlvmPtrType = CGF.ConvertType(PrivatePtrType);

  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(CGF.Builder.CreateElementBitCast(
                                  SrcElementPtrAddr, PrivateLlvmPtrType),
                              PrivatePtrType->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(CGF.Builder.CreateElementBitCast(
                                  SrcElementPtrAddr, PrivateLlvmPtrType),
                              PrivatePtrType->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(CGF.Builder.CreateElementBitCast(
                                  DestElementPtrAddr, PrivateLlvmPtrType),
                              PrivatePtrType->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(CGF.Builder.CreateElementBitCast(
                                  SrcElementPtrAddr, PrivateLlvmPtrType),
                              PrivatePtrType->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, CGF.Int8Ty,
                              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, CGF.Int8Ty,
                             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)) {
    // FIXME: This code doesn't make any sense, it's trying to perform
    // integer arithmetic on pointers.
    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.VoidPtrTy, CGF.getPointerAlign());
    else /* IncrScratchpadSrc = true */
      SrcBase =
          Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign());
  }

  ++Idx;
}
1724}

1726/// This function emits a helper that gathers Reduce lists from the first
1727/// lane of every active warp to lanes in the first warp.
1728///
1729/// void inter_warp_copy_func(void* reduce_data, num_warps)
1730///   shared smem[warp_size];
1731///   For all data entries D in reduce_data:
1732///     sync
1733///     If (I am the first lane in each warp)
1734///       Copy my local D to smem[warp_id]
1735///     sync
1736///     if (I am the first warp)
1737///       Copy smem[thread_id] to my local D
1738static 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().GV_Warp_Size;
if (!TransferMedium) {
11
←
Assuming 'TransferMedium' is non-null→
12
←
Taking false branch→
  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);
13
←
Calling 'getNVPTXLaneID'→
// nvptx_warp_id = nvptx_id / warpsize
llvm::Value *WarpID = getNVPTXWarpID(CGF);

Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
Address LocalReduceList(
    Bld.CreatePointerBitCastOrAddrSpaceCast(
        CGF.EmitLoadOfScalar(
            AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc,
            LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()),
        ElemTy->getPointerTo()),
    ElemTy, 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(ElemPtrPtr, CGF.Int8Ty, Align);
    ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType);
    if (NumIters > 1)
      ElemPtr = Bld.CreateGEP(ElemPtr, Cnt);

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

    // 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->getValueType(), TransferMedium,
        {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID});
    // SrcMediumVal = *SrcMediumPtr;
    Address SrcMediumPtr(
        Bld.CreateBitCast(
            SrcMediumPtrVal,
            CopyType->getPointerTo(
                SrcMediumPtrVal->getType()->getPointerAddressSpace())),
        CopyType, Align);

    // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I
    Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
    llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar(
        TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc);
    Address TargetElemPtr(TargetElemPtrVal, CGF.Int8Ty, Align);
    TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType);
    if (NumIters > 1)
      TargetElemPtr = Bld.CreateGEP(TargetElemPtr, Cnt);

    // *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;
1974}

1976/// Emit a helper that reduces data across two OpenMP threads (lanes)
1977/// in the same warp.  It uses shuffle instructions to copy over data from
1978/// a remote lane's stack.  The reduction algorithm performed is specified
1979/// by the fourth parameter.
1980///
1981/// Algorithm Versions.
1982/// Full Warp Reduce (argument value 0):
1983///   This algorithm assumes that all 32 lanes are active and gathers
1984///   data from these 32 lanes, producing a single resultant value.
1985/// Contiguous Partial Warp Reduce (argument value 1):
1986///   This algorithm assumes that only a *contiguous* subset of lanes
1987///   are active.  This happens for the last warp in a parallel region
1988///   when the user specified num_threads is not an integer multiple of
1989///   32.  This contiguous subset always starts with the zeroth lane.
1990/// Partial Warp Reduce (argument value 2):
1991///   This algorithm gathers data from any number of lanes at any position.
1992/// All reduced values are stored in the lowest possible lane.  The set
1993/// of problems every algorithm addresses is a super set of those
1994/// addressable by algorithms with a lower version number.  Overhead
1995/// increases as algorithm version increases.
1996///
1997/// Terminology
1998/// Reduce element:
1999///   Reduce element refers to the individual data field with primitive
2000///   data types to be combined and reduced across threads.
2001/// Reduce list:
2002///   Reduce list refers to a collection of local, thread-private
2003///   reduce elements.
2004/// Remote Reduce list:
2005///   Remote Reduce list refers to a collection of remote (relative to
2006///   the current thread) reduce elements.
2007///
2008/// We distinguish between three states of threads that are important to
2009/// the implementation of this function.
2010/// Alive threads:
2011///   Threads in a warp executing the SIMT instruction, as distinguished from
2012///   threads that are inactive due to divergent control flow.
2013/// Active threads:
2014///   The minimal set of threads that has to be alive upon entry to this
2015///   function.  The computation is correct iff active threads are alive.
2016///   Some threads are alive but they are not active because they do not
2017///   contribute to the computation in any useful manner.  Turning them off
2018///   may introduce control flow overheads without any tangible benefits.
2019/// Effective threads:
2020///   In order to comply with the argument requirements of the shuffle
2021///   function, we must keep all lanes holding data alive.  But at most
2022///   half of them perform value aggregation; we refer to this half of
2023///   threads as effective. The other half is simply handing off their
2024///   data.
2025///
2026/// Procedure
2027/// Value shuffle:
2028///   In this step active threads transfer data from higher lane positions
2029///   in the warp to lower lane positions, creating Remote Reduce list.
2030/// Value aggregation:
2031///   In this step, effective threads combine their thread local Reduce list
2032///   with Remote Reduce list and store the result in the thread local
2033///   Reduce list.
2034/// Value copy:
2035///   In this step, we deal with the assumption made by algorithm 2
2036///   (i.e. contiguity assumption).  When we have an odd number of lanes
2037///   active, say 2k+1, only k threads will be effective and therefore k
2038///   new values will be produced.  However, the Reduce list owned by the
2039///   (2k+1)th thread is ignored in the value aggregation.  Therefore
2040///   we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so
2041///   that the contiguity assumption still holds.
2042static 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();

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

CGBuilderTy &Bld = CGF.Builder;

Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
Address LocalReduceList(
    Bld.CreatePointerBitCastOrAddrSpaceCast(
        CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
                             C.VoidPtrTy, SourceLocation()),
        ElemTy->getPointerTo()),
    ElemTy, 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;
2192}

2194/// This function emits a helper that copies all the reduction variables from
2195/// the team into the provided global buffer for the reduction variables.
2196///
2197/// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
2198///   For all data entries D in reduce_data:
2199///     Copy local D to buffer.D[Idx]
2200static 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);
llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
Address LocalReduceList(
    Bld.CreatePointerBitCastOrAddrSpaceCast(
        CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
                             C.VoidPtrTy, Loc),
        ElemTy->getPointerTo()),
    ElemTy, 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
  ElemTy = CGF.ConvertTypeForMem(Private->getType());
  ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
      ElemPtrPtr, ElemTy->getPointerTo());
  Address ElemPtr =
      Address(ElemPtrPtr, ElemTy, 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);
  Address GlobAddr = GlobLVal.getAddress(CGF);
  llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(),
                                                 GlobAddr.getPointer(), Idxs);
  GlobLVal.setAddress(Address(BufferPtr,
                              CGF.ConvertTypeForMem(Private->getType()),
                              GlobAddr.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;
2301}

2303/// This function emits a helper that reduces all the reduction variables from
2304/// the team into the provided global buffer for the reduction variables.
2305///
2306/// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data)
2307///  void *GlobPtrs[];
2308///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
2309///  ...
2310///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
2311///  reduce_function(GlobPtrs, reduce_data);
2312static 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);
  Address GlobAddr = GlobLVal.getAddress(CGF);
  llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
      GlobAddr.getElementType(), GlobAddr.getPointer(), 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;
2401}

2403/// This function emits a helper that copies all the reduction variables from
2404/// the team into the provided global buffer for the reduction variables.
2405///
2406/// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
2407///   For all data entries D in reduce_data:
2408///     Copy buffer.D[Idx] to local D;
2409static 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);
llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
Address LocalReduceList(
    Bld.CreatePointerBitCastOrAddrSpaceCast(
        CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
                             C.VoidPtrTy, Loc),
        ElemTy->getPointerTo()),
    ElemTy, 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
  ElemTy = CGF.ConvertTypeForMem(Private->getType());
  ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
      ElemPtrPtr, ElemTy->getPointerTo());
  Address ElemPtr =
      Address(ElemPtrPtr, ElemTy, 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);
  Address GlobAddr = GlobLVal.getAddress(CGF);
  llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(),
                                                 GlobAddr.getPointer(), Idxs);
  GlobLVal.setAddress(Address(BufferPtr,
                              CGF.ConvertTypeForMem(Private->getType()),
                              GlobAddr.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;
2511}

2513/// This function emits a helper that reduces all the reduction variables from
2514/// the team into the provided global buffer for the reduction variables.
2515///
2516/// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data)
2517///  void *GlobPtrs[];
2518///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
2519///  ...
2520///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
2521///  reduce_function(reduce_data, GlobPtrs);
2522static 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);
  Address GlobAddr = GlobLVal.getAddress(CGF);
  llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
      GlobAddr.getElementType(), GlobAddr.getPointer(), 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;
2611}

2613///
2614/// Design of OpenMP reductions on the GPU
2615///
2616/// Consider a typical OpenMP program with one or more reduction
2617/// clauses:
2618///
2619/// float foo;
2620/// double bar;
2621/// #pragma omp target teams distribute parallel for \
2622///             reduction(+:foo) reduction(*:bar)
2623/// for (int i = 0; i < N; i++) {
2624///   foo += A[i]; bar *= B[i];
2625/// }
2626///
2627/// where 'foo' and 'bar' are reduced across all OpenMP threads in
2628/// all teams.  In our OpenMP implementation on the NVPTX device an
2629/// OpenMP team is mapped to a CUDA threadblock and OpenMP threads
2630/// within a team are mapped to CUDA threads within a threadblock.
2631/// Our goal is to efficiently aggregate values across all OpenMP
2632/// threads such that:
2633///
2634///   - the compiler and runtime are logically concise, and
2635///   - the reduction is performed efficiently in a hierarchical
2636///     manner as follows: within OpenMP threads in the same warp,
2637///     across warps in a threadblock, and finally across teams on
2638///     the NVPTX device.
2639///
2640/// Introduction to Decoupling
2641///
2642/// We would like to decouple the compiler and the runtime so that the
2643/// latter is ignorant of the reduction variables (number, data types)
2644/// and the reduction operators.  This allows a simpler interface
2645/// and implementation while still attaining good performance.
2646///
2647/// Pseudocode for the aforementioned OpenMP program generated by the
2648/// compiler is as follows:
2649///
2650/// 1. Create private copies of reduction variables on each OpenMP
2651///    thread: 'foo_private', 'bar_private'
2652/// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned
2653///    to it and writes the result in 'foo_private' and 'bar_private'
2654///    respectively.
2655/// 3. Call the OpenMP runtime on the GPU to reduce within a team
2656///    and store the result on the team master:
2657///
2658///     __kmpc_nvptx_parallel_reduce_nowait_v2(...,
2659///        reduceData, shuffleReduceFn, interWarpCpyFn)
2660///
2661///     where:
2662///       struct ReduceData {
2663///         double *foo;
2664///         double *bar;
2665///       } reduceData
2666///       reduceData.foo = &foo_private
2667///       reduceData.bar = &bar_private
2668///
2669///     'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two
2670///     auxiliary functions generated by the compiler that operate on
2671///     variables of type 'ReduceData'.  They aid the runtime perform
2672///     algorithmic steps in a data agnostic manner.
2673///
2674///     'shuffleReduceFn' is a pointer to a function that reduces data
2675///     of type 'ReduceData' across two OpenMP threads (lanes) in the
2676///     same warp.  It takes the following arguments as input:
2677///
2678///     a. variable of type 'ReduceData' on the calling lane,
2679///     b. its lane_id,
2680///     c. an offset relative to the current lane_id to generate a
2681///        remote_lane_id.  The remote lane contains the second
2682///        variable of type 'ReduceData' that is to be reduced.
2683///     d. an algorithm version parameter determining which reduction
2684///        algorithm to use.
2685///
2686///     'shuffleReduceFn' retrieves data from the remote lane using
2687///     efficient GPU shuffle intrinsics and reduces, using the
2688///     algorithm specified by the 4th parameter, the two operands
2689///     element-wise.  The result is written to the first operand.
2690///
2691///     Different reduction algorithms are implemented in different
2692///     runtime functions, all calling 'shuffleReduceFn' to perform
2693///     the essential reduction step.  Therefore, based on the 4th
2694///     parameter, this function behaves slightly differently to
2695///     cooperate with the runtime to ensure correctness under
2696///     different circumstances.
2697///
2698///     'InterWarpCpyFn' is a pointer to a function that transfers
2699///     reduced variables across warps.  It tunnels, through CUDA
2700///     shared memory, the thread-private data of type 'ReduceData'
2701///     from lane 0 of each warp to a lane in the first warp.
2702/// 4. Call the OpenMP runtime on the GPU to reduce across teams.
2703///    The last team writes the global reduced value to memory.
2704///
2705///     ret = __kmpc_nvptx_teams_reduce_nowait(...,
2706///             reduceData, shuffleReduceFn, interWarpCpyFn,
2707///             scratchpadCopyFn, loadAndReduceFn)
2708///
2709///     'scratchpadCopyFn' is a helper that stores reduced
2710///     data from the team master to a scratchpad array in
2711///     global memory.
2712///
2713///     'loadAndReduceFn' is a helper that loads data from
2714///     the scratchpad array and reduces it with the input
2715///     operand.
2716///
2717///     These compiler generated functions hide address
2718///     calculation and alignment information from the runtime.
2719/// 5. if ret == 1:
2720///     The team master of the last team stores the reduced
2721///     result to the globals in memory.
2722///     foo += reduceData.foo; bar *= reduceData.bar
2723///
2724///
2725/// Warp Reduction Algorithms
2726///
2727/// On the warp level, we have three algorithms implemented in the
2728/// OpenMP runtime depending on the number of active lanes:
2729///
2730/// Full Warp Reduction
2731///
2732/// The reduce algorithm within a warp where all lanes are active
2733/// is implemented in the runtime as follows:
2734///
2735/// full_warp_reduce(void *reduce_data,
2736///                  kmp_ShuffleReductFctPtr ShuffleReduceFn) {
2737///   for (int offset = WARPSIZE/2; offset > 0; offset /= 2)
2738///     ShuffleReduceFn(reduce_data, 0, offset, 0);
2739/// }
2740///
2741/// The algorithm completes in log(2, WARPSIZE) steps.
2742///
2743/// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is
2744/// not used therefore we save instructions by not retrieving lane_id
2745/// from the corresponding special registers.  The 4th parameter, which
2746/// represents the version of the algorithm being used, is set to 0 to
2747/// signify full warp reduction.
2748///
2749/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2750///
2751/// #reduce_elem refers to an element in the local lane's data structure
2752/// #remote_elem is retrieved from a remote lane
2753/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2754/// reduce_elem = reduce_elem REDUCE_OP remote_elem;
2755///
2756/// Contiguous Partial Warp Reduction
2757///
2758/// This reduce algorithm is used within a warp where only the first
2759/// 'n' (n <= WARPSIZE) lanes are active.  It is typically used when the
2760/// number of OpenMP threads in a parallel region is not a multiple of
2761/// WARPSIZE.  The algorithm is implemented in the runtime as follows:
2762///
2763/// void
2764/// contiguous_partial_reduce(void *reduce_data,
2765///                           kmp_ShuffleReductFctPtr ShuffleReduceFn,
2766///                           int size, int lane_id) {
2767///   int curr_size;
2768///   int offset;
2769///   curr_size = size;
2770///   mask = curr_size/2;
2771///   while (offset>0) {
2772///     ShuffleReduceFn(reduce_data, lane_id, offset, 1);
2773///     curr_size = (curr_size+1)/2;
2774///     offset = curr_size/2;
2775///   }
2776/// }
2777///
2778/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2779///
2780/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2781/// if (lane_id < offset)
2782///     reduce_elem = reduce_elem REDUCE_OP remote_elem
2783/// else
2784///     reduce_elem = remote_elem
2785///
2786/// This algorithm assumes that the data to be reduced are located in a
2787/// contiguous subset of lanes starting from the first.  When there is
2788/// an odd number of active lanes, the data in the last lane is not
2789/// aggregated with any other lane's dat but is instead copied over.
2790///
2791/// Dispersed Partial Warp Reduction
2792///
2793/// This algorithm is used within a warp when any discontiguous subset of
2794/// lanes are active.  It is used to implement the reduction operation
2795/// across lanes in an OpenMP simd region or in a nested parallel region.
2796///
2797/// void
2798/// dispersed_partial_reduce(void *reduce_data,
2799///                          kmp_ShuffleReductFctPtr ShuffleReduceFn) {
2800///   int size, remote_id;
2801///   int logical_lane_id = number_of_active_lanes_before_me() * 2;
2802///   do {
2803///       remote_id = next_active_lane_id_right_after_me();
2804///       # the above function returns 0 of no active lane
2805///       # is present right after the current lane.
2806///       size = number_of_active_lanes_in_this_warp();
2807///       logical_lane_id /= 2;
2808///       ShuffleReduceFn(reduce_data, logical_lane_id,
2809///                       remote_id-1-threadIdx.x, 2);
2810///   } while (logical_lane_id % 2 == 0 && size > 1);
2811/// }
2812///
2813/// There is no assumption made about the initial state of the reduction.
2814/// Any number of lanes (>=1) could be active at any position.  The reduction
2815/// result is returned in the first active lane.
2816///
2817/// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2818///
2819/// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2820/// if (lane_id % 2 == 0 && offset > 0)
2821///     reduce_elem = reduce_elem REDUCE_OP remote_elem
2822/// else
2823///     reduce_elem = remote_elem
2824///
2825///
2826/// Intra-Team Reduction
2827///
2828/// This function, as implemented in the runtime call
2829/// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP
2830/// threads in a team.  It first reduces within a warp using the
2831/// aforementioned algorithms.  We then proceed to gather all such
2832/// reduced values at the first warp.
2833///
2834/// The runtime makes use of the function 'InterWarpCpyFn', which copies
2835/// data from each of the "warp master" (zeroth lane of each warp, where
2836/// warp-reduced data is held) to the zeroth warp.  This step reduces (in
2837/// a mathematical sense) the problem of reduction across warp masters in
2838/// a block to the problem of warp reduction.
2839///
2840///
2841/// Inter-Team Reduction
2842///
2843/// Once a team has reduced its data to a single value, it is stored in
2844/// a global scratchpad array.  Since each team has a distinct slot, this
2845/// can be done without locking.
2846///
2847/// The last team to write to the scratchpad array proceeds to reduce the
2848/// scratchpad array.  One or more workers in the last team use the helper
2849/// 'loadAndReduceDataFn' to load and reduce values from the array, i.e.,
2850/// the k'th worker reduces every k'th element.
2851///
2852/// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to
2853/// reduce across workers and compute a globally reduced value.
2854///
2855void 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())
1
Taking false branch→
  return;

bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind);
2863#ifndef NDEBUG
bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind);
2865#endif

if (Options.SimpleReduction) {
2
←
Assuming field 'SimpleReduction' is false→
  assert(!TeamsReduction && !ParallelReduction &&(static_cast <bool> (!TeamsReduction && !ParallelReduction
 && "Invalid reduction selection in emitReduction.") ?
 void (0) : __assert_fail ("!TeamsReduction && !ParallelReduction && \"Invalid reduction selection in emitReduction.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 2869, __extension__
 __PRETTY_FUNCTION__))
         "Invalid reduction selection in emitReduction.")(static_cast <bool> (!TeamsReduction && !ParallelReduction
 && "Invalid reduction selection in emitReduction.") ?
 void (0) : __assert_fail ("!TeamsReduction && !ParallelReduction && \"Invalid reduction selection in emitReduction.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 2869, __extension__
 __PRETTY_FUNCTION__));
  CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs,
                                 ReductionOps, Options);
  return;
}

assert((TeamsReduction || ParallelReduction) &&(static_cast <bool> ((TeamsReduction || ParallelReduction
) && "Invalid reduction selection in emitReduction.")
 ? void (0) : __assert_fail ("(TeamsReduction || ParallelReduction) && \"Invalid reduction selection in emitReduction.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 2876, __extension__
 __PRETTY_FUNCTION__))
3
←
Taking false branch→
4
←
Assuming 'TeamsReduction' is false→
5
←
Assuming 'ParallelReduction' is true→
6
←
'?' condition is true→
       "Invalid reduction selection in emitReduction.")(static_cast <bool> ((TeamsReduction || ParallelReduction
) && "Invalid reduction selection in emitReduction.")
 ? void (0) : __assert_fail ("(TeamsReduction || ParallelReduction) && \"Invalid reduction selection in emitReduction.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 2876, __extension__
 __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) {
7
←
Assuming '__begin1' is equal to '__end1'→
  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) {
8
←
Assuming 'I' is >= 'E'→
9
←
Loop condition is false. Execution continues on line 2923→
  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),
                          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);
10
←
Calling 'emitInterWarpCopyFunction'→

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.")(static_cast <bool> (TeamsReduction && "expected teams reduction."
) ? void (0) : __assert_fail ("TeamsReduction && \"expected teams reduction.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 2948, __extension__
 __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, std::nullopt, 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, CGF.VoidPtrTy, 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, std::nullopt,
    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);
3038}

3040const VarDecl *
3041CGOpenMPRuntimeGPU::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);
  }
}
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);
3070}

3072Address
3073CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF,
                                        const VarDecl *NativeParam,
                                        const VarDecl *TargetParam) const {
assert(NativeParam != TargetParam &&(static_cast <bool> (NativeParam != TargetParam &&
 NativeParam->getType()->isReferenceType() && "Native arg must not be the same as target arg."
) ? void (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3078, __extension__
 __PRETTY_FUNCTION__))
       NativeParam->getType()->isReferenceType() &&(static_cast <bool> (NativeParam != TargetParam &&
 NativeParam->getType()->isReferenceType() && "Native arg must not be the same as target arg."
) ? void (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3078, __extension__
 __PRETTY_FUNCTION__))
       "Native arg must not be the same as target arg.")(static_cast <bool> (NativeParam != TargetParam &&
 NativeParam->getType()->isReferenceType() && "Native arg must not be the same as target arg."
) ? void (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3078, __extension__
 __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.getTypes().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, llvm::PointerType::getWithSamePointeeType(
        cast<llvm::PointerType>(TargetAddr->getType()), /*AddrSpace=*/0));
// Cast from generic to native address space.
TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
    TargetAddr, llvm::PointerType::getWithSamePointeeType(
        cast<llvm::PointerType>(TargetAddr->getType()),
                                NativePointeeAddrSpace));
Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType);
CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false,
                      NativeParamType);
return NativeParamAddr;
3102}

3104void 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, llvm::PointerType::getWithSamePointeeType(
          cast<llvm::PointerType>(NativeArg->getType()), /*AddrSpace*/ 0));
  TargetArgs.emplace_back(
      CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType));
}
CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs);
3128}

3130/// Emit function which wraps the outline parallel region
3131/// and controls the arguments which are passed to this function.
3132/// The wrapper ensures that the outlined function is called
3133/// with the correct arguments when data is shared.
3134llvm::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());

// Ensure we do not inline the function. This is trivially true for the ones
// passed to __kmpc_fork_call but the ones calles in serialized regions
// could be inlined. This is not a perfect but it is closer to the invariant
// we want, namely, every data environment starts with a new function.
// TODO: We should pass the if condition to the runtime function and do the
//       handling there. Much cleaner code.
Fn->addFnAttr(llvm::Attribute::NoInline);

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.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
// 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().VoidPtrTy)
                      .castAs<PointerType>());
}
unsigned Idx = 0;
if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
  Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
  Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
      Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy);
  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(), CGF.SizeTy);
  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)),
        CGF.ConvertTypeForMem(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;
3259}

3261void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF,
                                            const Decl *D) {
if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
  return;

assert(D && "Expected function or captured|block decl.")(static_cast <bool> (D && "Expected function or captured|block decl."
) ? void (0) : __assert_fail ("D && \"Expected function or captured|block decl.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3266, __extension__
 __PRETTY_FUNCTION__));
assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 &&(static_cast <bool> (FunctionGlobalizedDecls.count(CGF.
CurFn) == 0 && "Function is registered already.") ? void
 (0) : __assert_fail ("FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && \"Function is registered already.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3268, __extension__
 __PRETTY_FUNCTION__))
       "Function is registered already.")(static_cast <bool> (FunctionGlobalizedDecls.count(CGF.
CurFn) == 0 && "Function is registered already.") ? void
 (0) : __assert_fail ("FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && \"Function is registered already.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3268, __extension__
 __PRETTY_FUNCTION__));
assert((!TeamAndReductions.first || TeamAndReductions.first == D) &&(static_cast <bool> ((!TeamAndReductions.first || TeamAndReductions
.first == D) && "Team is set but not processed.") ? void
 (0) : __assert_fail ("(!TeamAndReductions.first || TeamAndReductions.first == D) && \"Team is set but not processed.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3270, __extension__
 __PRETTY_FUNCTION__))
       "Team is set but not processed.")(static_cast <bool> ((!TeamAndReductions.first || TeamAndReductions
.first == D) && "Team is set but not processed.") ? void
 (0) : __assert_fail ("(!TeamAndReductions.first || TeamAndReductions.first == D) && \"Team is set but not processed.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3270, __extension__
 __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().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")(static_cast <bool> (VD->isCanonicalDecl() &&
 "Expected canonical declaration") ? void (0) : __assert_fail
 ("VD->isCanonicalDecl() && \"Expected canonical declaration\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3306, __extension__
 __PRETTY_FUNCTION__));
  Data.insert(std::make_pair(VD, MappedVarData()));
}
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);
}
3321}

3323Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF,
                                                      const VarDecl *VD) {
if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
  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()))),
      VarTy, Align);
}

if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
  return Address::invalid();

VD = VD->getCanonicalDecl();
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();
3390}

3392void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) {
FunctionGlobalizedDecls.erase(CGF.CurFn);
CGOpenMPRuntime::functionFinished(CGF);
3395}

3397void 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);
3412}

3414void 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());
3424}

3426void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas(
  CodeGenFunction &CGF, const OMPExecutableDirective &D) const {
assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPTargetExecutionDirective(D
.getDirectiveKind()) && " Expected target-based directive."
) ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3429, __extension__
 __PRETTY_FUNCTION__))
       " Expected target-based directive.")(static_cast <bool> (isOpenMPTargetExecutionDirective(D
.getDirectiveKind()) && " Expected target-based directive."
) ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3429, __extension__
 __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 ValueDecl *, 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 ValueDecl *VD = LC.getCapturedVar();
    // FIXME: For now VD is always a VarDecl because OpenMP does not support
    //  capturing structured bindings in lambdas yet.
    if (!CS->capturesVariable(cast<VarDecl>(VD)))
      continue;
    auto It = Captures.find(VD);
    assert(It != Captures.end() && "Found lambda capture without field.")(static_cast <bool> (It != Captures.end() && "Found lambda capture without field."
) ? void (0) : __assert_fail ("It != Captures.end() && \"Found lambda capture without field.\""
, "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3468, __extension__
 __PRETTY_FUNCTION__));
    LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second);
    Address VDAddr = CGF.GetAddrOfLocalVar(cast<VarDecl>(VD));
    if (VD->getType().getCanonicalType()->isReferenceType())
      VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr,
                                             VD->getType().getCanonicalType())
                   .getAddress(CGF);
    CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal);
  }
}
3478}

3480bool 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.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 3504)
                   "static storage.")::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the "
 "static storage.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 3504);
}
return false;
3507}

3509// Get current CudaArch and ignore any unknown values
3510static 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;
3521}

3523/// Check to see if target architecture supports unified addressing which is
3524/// a restriction for OpenMP requires clause "unified_shared_memory".
3525void 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: {
      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_60:
    case CudaArch::SM_61:
    case CudaArch::SM_62:
    case CudaArch::SM_70:
    case CudaArch::SM_72:
    case CudaArch::SM_75:
    case CudaArch::SM_80:
    case CudaArch::SM_86:
    case CudaArch::SM_87:
    case CudaArch::SM_89:
    case CudaArch::SM_90:
    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::GFX90a:
    case CudaArch::GFX90c:
    case CudaArch::GFX940:
    case CudaArch::GFX1010:
    case CudaArch::GFX1011:
    case CudaArch::GFX1012:
    case CudaArch::GFX1013:
    case CudaArch::GFX1030:
    case CudaArch::GFX1031:
    case CudaArch::GFX1032:
    case CudaArch::GFX1033:
    case CudaArch::GFX1034:
    case CudaArch::GFX1035:
    case CudaArch::GFX1036:
    case CudaArch::GFX1100:
    case CudaArch::GFX1101:
    case CudaArch::GFX1102:
    case CudaArch::GFX1103:
    case CudaArch::Generic:
    case CudaArch::UNUSED:
    case CudaArch::UNKNOWN:
      break;
    case CudaArch::LAST:
      llvm_unreachable("Unexpected Cuda arch.")::llvm::llvm_unreachable_internal("Unexpected Cuda arch.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp"
, 3601);
    }
  }
}
CGOpenMPRuntime::processRequiresDirective(D);
3606}

3608void CGOpenMPRuntimeGPU::clear() {

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();
3642}

3644llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) {
CGBuilderTy &Bld = CGF.Builder;
llvm::Module *M = &CGF.CGM.getModule();
const char *LocSize = "__kmpc_get_hardware_num_threads_in_block";
llvm::Function *F = M->getFunction(LocSize);
if (!F) {
  F = llvm::Function::Create(
      llvm::FunctionType::get(CGF.Int32Ty, std::nullopt, false),
      llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule());
}
return Bld.CreateCall(F, std::nullopt, "nvptx_num_threads");
3655}

3657llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) {
ArrayRef<llvm::Value *> Args{};
return CGF.EmitRuntimeCall(
    OMPBuilder.getOrCreateRuntimeFunction(
        CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block),
    Args);
3663}

3665llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) {
ArrayRef<llvm::Value *> Args{};
return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
                               CGM.getModule(), OMPRTL___kmpc_get_warp_size),
                           Args);
3670}